Datasets:

math-ai

/

AutoMathText

like 125

https://arxiv.org/abs/2009.09834

Stochastic Time-Periodic Tonelli Lagrangian on Compact Manifold

In this paper, we study a class of time-periodic stochastic Tonelli Lagrangians on compact manifolds. Precisely, we discuss the stochastic Mane Critical Value, prove the existence of stochastic Weak KAM solutions of the related Hamilton-Jacobi equation. Furthermore, we survey the global minimizer.

\section{Introduction} \quad\quad Time-periodic Tonelli Lagrangians have been extensively studied in recent years by John Mather\cite{mather1991action}, Ricardo Mane \cite{mane1996generic} \cite{mane1992minimizing}, Patrick Bernard \cite{bernard2008dynamics} \cite{bernard2002connecting}, Daniel Massart \cite{massart2006subsolutions}, etc. It is closely related to calculus of variations, the weak KAM theory \cite{fathi2008weak}, Aubry-Mather theory and Optimal Transport(\cite{villani2008optimal}). In this paper, we study the time-periodic Tonelli Lagrangian by adding a stochastic variable. We discuss the measurability of Mane-Critical Value, prove the existence of stochastic viscosity solution for a system of Hamilton-Jacobi equations, and describe the global minimizer. Bacis Setting: Let $M$ be a compact, connected Riemannian manifold, $TM$ its tangent bundle. $(\Omega,\mathscr{F},\mathbb{P})$ is a probability space, $\mathscr{F}$ is the $\sigma-$algebra of $\Omega$. $\theta: R \times \Omega \to \Omega$ is a measurable function with skew-product structure. Let $\mathscr{B}$ be the Borel algebra of $\mathbb{R}$. Then the $\sigma-$algebra of $\mathbb{R}\times\Omega$ is given by $\mathscr{B}\otimes \mathscr{F}$. The stochastic time-periodic lagrangian $L:TM \times T \times \Omega \to R$, is a measurable function. For each $\omega \in \Omega$, $L(\cdot,\cdot,\cdot,\omega):TM \times T \to R$ is a Tonelli Lagrangian, and for each $(x,v,t)\in TM \times T$, $L(x,v,t,\cdot) : \Omega \to R$ is a measurable function. For each $(x,v,t,\omega) \in TM \times T \times \Omega$, $s \in R$, we assume \[ L(x,v,t+s,\omega)=L(x,v,t,\theta(s,\omega)) \] Let $-\infty < s < t < \infty$, $x,y \in M$, $\omega \in \Omega$. $AC(s,x;t,y)$ is the space of absolutely continuous curves in $R \times M$, connecting $(s,x)$ and $(t,y)$ The stochastic Lagrangian action $A^{\omega}(s,x;t,y)$ is \[ A^{\omega}(s,x;t,y)=\inf_{\gamma \in AC(s,x;t,y)}\int^{t}_{s}L(\gamma(\sigma),\dot{\gamma}(\sigma),\sigma,\omega)d\sigma \] Here, we study the following topics: 1. Mane-Critical Value 2. The existence of stochastic viscosity solution for a stochastic system of Hamilton-Jacobi equations. 3. global minimizer. For each $\omega \in \Omega$, $\Phi^{\omega}$ is the set of invariant measure of Hamiltonian flow in $TM \times T$ for the Tonelli Lagrangian $L(\cdot,\cdot,\cdot,\omega)$. The Mane-Critical Value for $L(\cdot,\cdot,\cdot,\omega)$, is \begin{equation}\nonumber \alpha(\omega)=\inf_{\mu \in \Phi^{\omega}} \int_{TM \times T}L(x,v,t,\omega) d \mu \end{equation} The first main result is: \begin{theorem} The function $\alpha(\omega)$ is a measurable function on $\Omega$. In particular, if $\{\theta(t): \Omega \to \Omega|t \in R \}$ is ergodic, $\alpha(\omega)$ is constant almost everywhere on $\Omega$. \end{theorem} Now, we define the Lax-Oleinik Operator, for $u: M \to R$, $(x,t) \in M \times T$, $\lambda \in R$, $\omega \in \Omega$, \[ T^{\omega}_{\lambda}u(x,t)=\min_{y \in M}\{u(y)+A^{\omega}(t-\lambda,y;t,x)+ \lambda \alpha(\omega)\} \] The Hamiltonian, $H: T^{*}M \times T \times \Omega \to R$, is defined as, \[ H(x,p,t,\omega)=\sup_{v \in T_x M}\{p\cdot v - L(x,v,t,\omega)\} \] The second main result is a weak KAM-type theorem, \begin{theorem} For each $u:M \to R$, $(x,t) \in M \times R$, define: \[ u^{\omega}(x,t)=\liminf_{\lambda \to +\infty}T^{\omega}_{\lambda}(u)(x,t) \] Then, the following holds true: (i)For all $\omega \in \Omega$, $u^{\omega}(x,t)$ is a viscosity solution of the Hamilton-Jacobi equation: \[ \partial_t u(x,t) + H(x,\partial_x u(x,t),t,\omega)= \alpha(\omega) \] (ii)For all $(x,t) \in M \times R$, $u^{\omega}(x,t)$ are measurable functions on $\Omega$. (iii)For all $x \in M$, $s,t \in R$, $\omega \in \Omega$, we have \qquad $u^{\omega}(x,t+1)=u^{\omega}(x,t)$ \qquad and \qquad $u^{\theta(s)\omega}(x,t)=u^{\omega}(x,t+s)$ \end{theorem} In this paper, we define the global minimizer in this way: $\gamma^{\omega}:(-\infty,\infty) \to M$ is a global minimizer if and only if for any $s<t$, $\gamma^{\omega}|[s,t]$ is minimizer for the Action $A^{\omega}(x,s;y,t)$. We prove that the set of global minimzer is nonempty for all $\omega \in \Omega$. Besides, under some situations, we can know the structure of ergodic invariant measures on the space of global minimizers. $\mathbf{Acknowledgements}$ I thank Piermarco Cannarsa and Alfonso Sorrentino for their generous help. They spent much time talking with me for this project and they provided some insightful advice and helpful references. \section{Stochastic Time-Periodic Tonelli Lagrangian} In this section, we recall some basic notions, give the definiton of our investigated objects and examples. Throughout this paper, $(M,g)$ is a compact connected Riemannian manifold and $TM$ is its tangent bundle. John Mather originally considered the Time-Periodic Tonelli Lagrangian. \begin{definition}[Time-Periodic Tonelli Lagrangian] On $M$, a $C^2$ map $L:TM \times R \to R$ is a Time-Periodic Tonelli Lagrangian if $L$ satisfies: (I)Periodicity: $L(x,v,t+1)=L(x,v,t)$, $\forall x \in M$,$v \in T_{x}M$ $t \in R$. (II)Convexity: For all $x \in M$,$v \in T_xM$, the Hessian matrix $\frac{\partial^2 L}{\partial v_i \partial v_j}(x,v,t)$ is positive definite. (III)Superlinearity: \[ \lim_{||v|| \to +\infty}\frac{L(x,v,t)}{||v||}=+\infty \] uniformly on $x \in M$,$t \in R$. (IV)Completeness: The maximal solutions of the Euler-Lagrangian, that in local coordinates is: \[ \frac{d}{dt}\frac{\partial L}{\partial v}(x,\dot{x},t)=\frac{\partial L}{\partial x}(x,\dot{x},t) \] are defined on all $R$. \end{definition} Following this definition from John Mather, we introduce the definition of Stochastic Time-Periodic Tonelli Lagrangian. $(\Omega,\mathscr{F},\mathbb{P})$ is a probability space, $\mathscr{F}$ is the $\sigma-$algebra of $\Omega$. \begin{definition}[Stochastic Time-Periodic Tonelli Lagrangian] On $M$,a map $L:TM \times R \times \Omega \to R$ is a measurable function. It is called Stochastic Time-Periodic Tonelli Lagrangian if $L$ satisfies: (I)Fix each $\omega \in \Omega$ , $L(\cdot,\cdot,\cdot,\Omega)$ is a Time-Periodic Tonelli Lagrangian on $TM \times T$. (II)Fix each $(x,v,t) \in TM \times T$, $L(x,v,t,\cdot)$ is a measurable function on $\Omega$. \end{definition} Next, we introduce a skew-product dynamical system. Let $\mathscr{B}$ be the Borel algebra of $\mathbb{R}$.Then the $\sigma-$algebra of $\mathbb{R}\times\Omega$ is given by $\mathscr{B}\otimes \mathscr{F}$. \begin{definition} A skew-product dynamical system is a measurable map $\theta$: $R \times \Omega \to \Omega$,satisfying (I) $\forall x \in \Omega$,$\forall s,t \in R$, $\theta(0,x)=x$, $\theta(s,\theta(t,x))=\theta(s+t,x)$. (II)Fix $t \in R$, $\theta(t): \Omega \to \Omega$ defined as $\theta(t)(x)=\theta(t,x)$, is measure-preserving, i.e $\forall E \in \mathscr{F}$, $\mathbb{P}({\theta(t)}^{-1}(E))=\mathbb{P}(E)$. \end{definition} From Definition 2.3, $ { \{\theta(t)\} }_{t \in \mathbb{R}} $ is a group of measure-preserving endomorphism of $\Omega$. In this paper, we consider the a class of Stochastic Time-Periodic Lagrangians which match the skew-product dynamical system, i.e. satisfying the following assumption: \begin{assumption} $L:TM \times T \times \Omega \to R$ is a Stochastic Time-Periodic Tonelli Lagrangian, we assume that $L$ matches the skew-product dynamical system, if for $(x,v) \in TM$, $s \in R$, $\omega \in \Omega$, we have \[ L(x,v,t+s,\omega)=L(x,v,t,\theta(s)\omega) \] \end{assumption} The next two examples are constructed to show existence of Lagrangians which satisfy Assumption 2.1 \begin{example} $\Omega=[0,1]$ is a probability space with Lebesgue measure. $f:[0,1] \to [0,1]$ is a one-to-one measurable function. Precisely, \[ f(x) = \begin{cases} x & \text{if $0 \le x < \frac{1}{3}$,}\\ x+\frac{1}{3} & \text{if $\frac{1}{3} \le x < \frac{2}{3}$,}\\ x-\frac{1}{3} & \text{if $\frac{2}{3} \le x <1$.} \end{cases} \] $\theta: R \times \Omega \to \Omega$, $t \in R$, is defined by \[ \theta(t,\omega)=f^{-1}(t+f(\omega)) \] We see that $\theta:R \times \Omega \to \Omega$ is a skew-product dynamical system. Then define $L:TM \times T \times \Omega \to R$ by \[ L(x,v,t,\omega)=\frac{1}{2}g_{x}(v,v)+h(t+f(\omega)) \] where $g_x$ is a Riemannian metric over $M$, $h$ is a periodic function from $R$ to $R$, with period $1$, $C^2$ regularity. $L$ is a Stochastic Time-Periodic Tonelli Lagrangian which satisfies Assumption 2.1 \end{example} \begin{example} $\Omega=T^d$ is a probability space with Lebesgue Measure. $T^d$ can be decomposed as \[ T^d=\prod_{k=1}^{d}([0,\frac{1}{n})\bigcup [\frac{1}{n},\frac{2}{n})\cdot\cdot\cdot\bigcup[\frac{n-1}{n},1)) \] Assume that $f$ is a permutation of these $n^d$ cubics by linear maps, so $f$ is a one-to-one measurable map from $T^d$ to $T^d$. Choose $(1,\alpha_2,..\alpha_d) \in R^d$ as a rationally independent vectors, i.e, there is no nonzero vector $(k_1,k_2,...k_d) \in Z^d$ such that $k_1+k_2\alpha_2+k_3\alpha_3+\cdot \cdot\cdot +k_d\alpha_d=0$ Define $\phi(t):\Omega \to \Omega$ for $\forall x\in T^d$, $x=(x_1,...,x_d) \in R^d$,$t\in R$, by \[ \phi(t)(x_1,...,x_d)=(x_1+t,x_2+\alpha_2 t,..,x_d+\alpha_d t)(mod Z^d) \] For $t \in R$, $\phi(t)$ is ergodic over $\Omega$ with stationary property: $\phi(t)\cdot\phi(s)=\phi(t+s)$ Define the skew-product dynamical system $\theta: R \times \Omega \to \Omega$, for each $t \in R$, each $\omega \in \Omega$, \[ \theta(t,x)=f^{-1}(\phi(t)\cdot (f(x)))) \] In particular, $\{\theta(t)|t \in R\}$ is ergodic on $\Omega$. Define $L:TM \times T \times \Omega$ as follows: \[ L(x,v,t,\omega)=\frac{1}{2}g_x(v,v)+h(t+\pi(\omega)) \] where $\pi:T^d \to R$ defined as $\pi(x_1,x_2,...x_d)=x_1$, $h$ is a 1-periodic function on $R$ and has regularity $C^2$, $g_x$ is the Riemannian Metric over $M$. Hence, $L$ is a Stochastic Time-Periodic Tonelli Lagrangian and satisfies the Assumption 2.1 For the translation on Torus, see \cite{katok1997introduction}. \end{example} \begin{remark} In Example 1, Example 2, the one-to-one measurable map $f$ is not unique. \end{remark} \section{Probability Measures on Metric Spaces} In this section, we introduce some tools from Probability Theory. $(X,d)$ is a metric space.Consider the functional space. $C_b(X)=\{ f: X \to R: f $ is continuous and bounded $\}$. Each $f \in C_b(X)$ is integrable with respect to any finite Borel measure on $X$. We introduce the notion of weak convergence of probability measure on $X$. \begin{definition} $\mu,\mu_1,\mu_2,...$ are finite Borel measures on $X$. We say that $(\mu_i)_i$ converges weakly to $\mu$, if for all $f \in C_b(X)$, we have \[ \lim_{i \to \infty} \int_{X}f d\mu_i \to \int_{X}f d\mu \] we denote $\mu_i \rightharpoonup \mu$. \end{definition} Next, we discuss the Prokhorov metric on $(X,d)$. Denote $\mathcal{P}=\{\mu| \mu$ is a probability measure on $X \}$. $\mathcal{B}(X)$ is the set of all Borel algebra generated by open sets on $X$. The Prokhorov metric arises from the distance on $\mathcal{P}$, defined as \begin{definition} For $\mu,\nu \in \mathcal{P}$, $d_p(\mu,\nu)$ is defined by \[ d_P(\mu,\nu)=\inf\{\alpha>0: \mu(A) \le \nu(A_{\alpha}) +\alpha, \nu(A) \le \mu(A_{\alpha})+\alpha, \forall A \in \mathcal{B}(X) \} \] where $A_{\alpha}=\{x \in X: d(x,A)< \alpha \}$ if $A \neq \emptyset$, $\emptyset_{\alpha}:=\emptyset$ for all $\alpha > 0$. \end{definition} Then we have: \begin{lemma} $(X,d)$ is the metric space, $d_{P}$ defined above, (1)$d_P$ is a metric on $\mathcal{P}=\mathcal{P}(X)$. (2)If $\mu, \mu_1, \mu_2,...\in \mathcal{P}$, $\lim_{i \to \infty}d_P(\mu_i, \mu)=0$ implies $\mu_i \rightharpoonup \mu$. \end{lemma} \begin{lemma} If $(X,d)$ is a separable metric space, then for any $\mu,\mu_1,\mu_2,...\in \mathcal{P}(X)$ one has $\mu_i \rightharpoonup \mu$ if and only if $\lim_{i \to \infty}d_P(\mu,\mu_i)=0$. \end{lemma} \begin{lemma} $(X,d)$ is a separable metric space, then $\mathcal{P}=\mathcal{P}(X)$ with the Prokhorov metric $d_P$ is separable. \end{lemma} \begin{lemma} $(X,d)$ is a separable complete metric space, then $\mathcal{P}=\mathcal{P}(X)$ with the Prokhorov metric $d_P$ is complete. \end{lemma} The proof from Lemma 3.1 to Lemma 3.4 can be checked in \cite{billingsley2013convergence} \begin{theorem} Let $X$ and $Y$ be two Polish spaces and $\lambda$ be a Borel probability measure on $X\times Y$.Let us set $\mu=\pi_{X}\lambda$, where $\pi_{X}$ is the standard projection from $X \times Y$ onto $X$. Then there exists a $\mu-$almost everywhere uniquely determined family of Borel probability measures $(\lambda_x)$ on $Y$ such that 1.The function $x \to \lambda_x$ is Borel measurable, in the sense that $x \to \lambda_x(B)$ is a Borel-measurable function for each Borel-measurable set $B \in Y$. 2. For every Borel-measurable function $f:X \times Y \to [0,\infty)$, \[ \int_{X\times Y}f(x,y)d\lambda(x,y)=\int_{X}\int_{Y}f(x,y)d\lambda_x(y)d\nu(x) \] \end{theorem} The disintegration of measure can be seen in \cite{ambrosio2008transport} \section{ Mane-Critical Value} Given a Time-Periodic Tonelli Lagrangian $L:TM \times T \to R$, by Euler-Lagrange equation,i.e, \[ \frac{d}{dt}\frac{dL}{d\dot{x}}(x,\dot{x},t)=\frac{dL}{dx}(x,\dot{x},t) \] we can define a time-dependent Lagrangian flow $\Phi_{s,t}:TM \times \{s\} \to TM \times \{t\}$.see \cite{mather1991action} Denote: $\mathcal{M}_{inv}=\{\mu: \mu$ is a Borel probability over $TM \times T$, $\mu $ is invariant under the flow $\Phi_{s,t}, \forall s,t \in R \}$, the Mane-critical value of $L$ is defined as \begin{equation} -c[0]=\min_{\mu \in \mathcal{M}_{inv} }\int_{TM \times T}L(x,v,t)\,d\mu \end{equation} The corresponding Hamiltonian $H: T^{*}M \times T \to R$ is defined by \[ H(x,p,t)=\sup_{v \in TM}p(v)-L(x,v,t) \] Considering the Hamilton-Jacobi equation, $u: M \times T \to R$. \begin{equation} \partial_t u(x,t)+ H(x,\partial_x u(x,t), t)=c[0] \end{equation} If $u$ is a subsolution of (2), that means for each $(x_0,t_0) \in M \times T $, there exists a $C^1$ function $\phi:M \times T \to R$, $\phi \ge u$, $\phi(x_0,t_0)=u(x_0,t_0)$, we have \[ \partial_t \phi(x_0,t_0)+ H(x_0,\partial_x \phi(x_0 ,t_0),t_0) \le c[0] \] For all $n \in N$, $h_n:(M \times T) \times (M \times T) \to R$ is defined as : \[ h_n((x,t),(y,s))=\min_{\gamma \in \Sigma(x,t;y,s+n)}\int^{s+n}_{t}L(\gamma,\dot{\gamma},t)dt+nc[0] \] The Peierls barrier is then defined as: for $(x,t)\times (y,s) \in (M\times T) \times (M \times T)$: \[ h((x,t),(y,s))=\liminf_{n \to \infty} h_n((x,t),(y,s)) \] The Projected Aubry set is \[ \mathcal{A}_{0}:=\{(x,t)\in M \times T: h((x,t),(x,t)=0)\} \] In \cite{massart2006subsolutions}, Daniel Massart proved a useful theorem that we desire, we present it here: \begin{theorem} There exists a $C^{1}$ critical subsolution of Hamilton-Jacobi equation which is strict at every point of $\mathcal{A}^{c}_{0}$. \end{theorem} To prove theorem 1.1, we introduce the notion of closed measure. \begin{definition} A probability measure $\mu$ on $TM \times T$ is called closed if \[ \int_{TM \times T}|v| \, d\mu(x,v,t) < \infty \] and for every smooth function $f$ on $TM \times T$, we have \[ \int_{TM \times T}df(x,t)(v,1)d\mu(x,v,t)=0 \] We denote the set of closed measures on $TM \times T$ as $\mathcal{M}_{c}$ \end{definition} In \cite{massart2006subsolutions}, Daniel Massart gives a desired theorem as follows: \begin{theorem} For a Time-Periodic Tonelli Lagrangian $L: TM \times T \to R$, its Mane-Critical Value can be formulated as \begin{equation} -c[0]=\min_{\mu \in \mathcal{M}_{f}}\int_{TM \times T}L(x,v,t)\,d\mu \end{equation} \end{theorem} \begin{lemma} For a Time-Periodic Tonelli Lagrangian $L:TM \times T \to R$, if one measure $\mu \in P(TM)$ satisfies that \[ -c[0]=\int_{TM \times T} L(x,v,t)d\mu \] , $supp(\mu)$ is a compact subset in $TM \times T$. \end{lemma} \begin{proof} For each closed measure $\mu$, if $f:M \times T \to R$ is a smooth function, we have \[ \int_{TM \times T}df(x,t)(v,1)d\mu(x,v,t)=0 \] If $g:M \times T \to R$ is $C^1$, we can approximate it in the uniform $C^1$ topology by a sequence of $C^{\infty}$ functions $f_n: M\times T \to R$. In particular, there is a constant $K < \infty$, for each $x \in M$, and $n \in N$, we have $|df(x,t) \cdot(v,1)| \le K(||v||+1)$. Since $\int_{TM \times T}||v||d\mu(x.v.t)<\infty $, $df_n(x,t)(v,1) \to dg(x,t)(v,1)$ by the dominated convergence theorem, we obtain that $\int_{TM \times T}dg(x,t)\cdot(v,1) d\mu(x,v,t)=0$. Suppose that a closed measure $\mu$ does satisfy that $\int_{TM \times T}L(x,v,t)d\mu=-c[0]$, From Theorem 4.1, we know that there exists $u: M \times T \to R$ as a $C^1$ critical subsolution, such that for $x \in \mathcal{A}^{c}_{0}$, we have \[ \partial_t u(x,t)+ H(x,\partial_x u(x,t),t)< c[0] \] We integrate the following equation: \[ \partial_x u(x,t)\cdot v + \partial_t u(x,t) \le L(x,v,t)+ H(x,\partial_x u(x,t),t)+ \partial_tu(x,t) \le L(x,v,t)+c[0] \] Then we get \[ 0 \le \int_{TM \times T}L(x,v,t)+H(x,d_x u,t) d\mu \le 0 \] So we know that if $(x,v)\in supp(\mu)$, we have \begin{equation}\nonumber \begin{aligned} &\partial_t u(x,t)+H(x,\partial_x u(x,t),t)=c[0]\\ &\partial_x u(x,t)(v)=L(x,v,t)+H(x,\partial_x u(x,t) ,t) \end{aligned} \end{equation} So we know that $x \in \mathcal{A}_{0}$, $\partial_x u(x,t)=\frac{\partial L}{\partial v}(x,v,t)$, and $v=\frac{\partial H}{\partial p}(x,\partial_x u(x,t),t)$. Hence we conclude that $supp(\mu)$ is compact. \end{proof} Coming back to a Stochastic Time-Periodic Tonelli Lagrangian $L: TM \times T \times \Omega \to R$, for each $\omega \to \Omega$, $L(\cdot,\cdot,\cdot,\omega)$ is a Time-Periodic Tonelli Lagrangian. So $L(\cdot,\cdot,\cdot,\omega)$ has a Mane-Critical Value, we denote it as $\alpha(\omega)$. \begin{lemma} Given a compact connected Riemannian manifold $(M,g)$, the manifold $M$, its tangent bundle $TM$ and $TM \times T$ are separable, complete metric spaces. \end{lemma} \begin{proof}{First proof of Theorem 1.1} Since $M$ is a compact manifold, we can find a finite number of charts $\{(U_i,\phi_i)| 1 \le i \le N\}$ to cover $M$.For $1 \le i \le N$, $U_i$ is isomorphic to a open subset of $R^n$, so $U_i$ is a separable open set for $1 \le i \le N$. Hence, $M=\cup_{1 \le i \le N}U_i$ is also separable. $TU_i$ is isomorphic to $\phi_i(U_i) \times R^n$, which is separable. So $TU_i$ is separable. Then $TM \subset \cup^{N}_{i=1}TU_i$. Hence $TM$ is separable. To prove $M$ is a metric space, we define a distance on $M$. For any $x,y \in M$, $\Sigma(x,y)=\{\gamma|\gamma :[0,1] \to M, \gamma$ is absolutely continuous on $M\}$. We denote the distance between $x$ and $y$ as $d(x,y)$, defined by \[ d(x,y)=\inf_{\gamma \in \Sigma(x,y)}\int^{1}_{0}g_{\gamma(t)}(\dot{\gamma}(t),\dot{\gamma}(t))dt \] $M$ is connected, any two different points can be connected by an absolutely continuous curve. $d_{M}(x,y)$ defines a metric on $M$.(see \cite{petersen2006riemannian}). Since any compact metric space is complete, we can know that $M$ is complete. To prove $TM$ is a metric space, firstly we prove $TM$ is a Riemannian manifold. We construct the Riemannian metric locally on $TM$. $\{(U_i,\phi_i)|1 \le i \le N \}$ are charts which cover $M$, then $\{(TU_i,d\phi_i)|1 \le i \le N\}$ are charts which cover $TM$, and $d {\phi}_i( TU_i)={\phi}_j(U_i)\times R^n, 1 \le i \le N$. For some $j$, $1 \le j \le N$, $\xi \in TU_j$, $(x,v)=d \phi(\xi) \in \phi(U_j) \times R^n $, for $V,W \in T_{\xi}(TU_j)$, we write $d\phi_j(V)=(V_1,V_2) \in R^{n+n}$, $d\phi_j(W)=(W_1,W_2) \in R^{n+n}$. $\pi:TM \to M$ is the canonical projection. We define the Riemannian Metric on $TM$ as: \[ G_{\xi}(V,W)=g_{\pi(\xi)}((d\phi_j)^{-1}V_1,(d\phi_j)^{-1}W_1)+g_{\pi(\xi)}((d\phi_j)^{-1}V_2,(d\phi_j)^{-1}W_2) \] This Riemannian metric on $TM$ is well defined. Then we can prove than $\mathbb{TM}$ is a metric space by viewing it as a connected Riemannian Manifold, we denote the metric by $d_{TM}(\cdot,\cdot)$. To prove $TM$ is complete under this metric.Let $\xi_1,\xi_2,..,\xi_n,...$ be a cauchy sequnce in $TM$ under the metric $d_{TM}(\cdot,\cdot)$. Then $\pi(\xi_1),\pi(\xi_2),...\pi(\xi_n),..$ are cauchy sequences in $M$ under the metric $d(\cdot,\cdot)$, since $d(\pi(p),\pi(q)) \le d_{TM}(p,q)$ for all $p,q \in TM$.Then we can find $x_0 \in TM$ such that $\lim_{i \to \infty}d(\pi(\xi_i),x_0)=0$.Then there is a local chart $(U,\phi)$ on $M$ such that $x_0 \in U$.There exists an integer $N$, such that for $i \ge N$,$\pi(\xi_i) \in U$. If we write $d\phi(\xi_i)=(x_i,v_i) \in \phi(U_i)$.We have $\lim_{i \to \infty}x_i=x_0$.Since $(x_i,v_i) \in d\phi(U_i)\times R^n$ is cauchy sequnce,then $v_i \in R^n, i \ge N$ is a cauchy sequence. $R^n$ is complete, so there exists $v_0 \in R^n$, such that $\lim_{i \to \infty}v_i=v_0$. Then $\lim_{i \to \infty}(x_i,v_i)=(x_0,v_0)$. So $TM$ is complete. Both $TM$ and $T$ are separable, complete metric space, so we can define the metric $d_{TM \times T}$ in this way , if $(\xi_1,t_1),(\xi_2,t_2) \in TM \times T$, $d_{TM \times T}((\xi_1,t_1),(\xi_2,t_2))=d_{TM}(\xi_1,\xi_2)+d_{T}(t_1,t_2)$. It is clear that $TM \times T$ is also a complete separable metric space. \end{proof} We turn to the first version of proof of Theorem 1.1 \begin{proof} From Lemma 4.1, $TM \times T$ is a complete separable metric space. With Lemma 3.3, Lemma 3.4, $\mathcal{P}(TM \times T)$ with the Prokhorov metric $d_{P}$ is a complete separable metric space. The closed measure is a subset of $\mathcal{P}(TM \times T)$, so we can pick up a countable dense subset of closed measure $\{\mu_k|k=1,2,...,n,...\}$. Fix $\omega$, $L(\cdot,\cdot,\cdot,\omega)$ is a Time-Periodic Tonelli Lagrangian, From the lemma 4.2 and lemma 3.4, we know there is a $\mu \in P(TM \times T)$ making (1) holds, and $\mu$ is indeed a closed measure, (see \cite{massart2006subsolutions},\cite{fathi2004existence}). From Lemma 6.1, we know that the support of $\mu$ is compact,There exists $R(\omega)>0$ such that if $(x,v) \in supp(\mu)$,we have $g_x(v) \le R(\omega)$. We can find a smooth function $\chi_{\omega}: [0,\infty) \to [0,1] $ such that $\chi(x)=1$ when $ 0 \le x \le R(\omega)$; $\chi(x)=0$ when $x \ge R(\omega)+1$. We have \[ -\alpha(\omega)=\inf_{\mu \in {TM \times T}}\int_{\mu \in \mathcal M}L(x,v,t,\omega)\chi_{\omega}(g_x(v))d\mu(x,v,t) \] Since $L(x,v,t,\omega)\chi_{\omega}(g_x(v))$ is a continuos bounded function, so we can apply weak convergence of probability measure on $L(x,v,t,\omega)\chi_{\omega}(g_x(v))$.Since $\{\mu_k|k=1,2,...\}$ is dense on $\mathcal{P}(TM \times T)$, we know that \[ -\alpha(\omega)=\inf_{k \in N} \int_{TM \times T}L(x,v,t,\omega)\chi_{\omega}(g_x(v))d\mu_{k}=\inf_{k \in N} \int_{TM \times T}L(x,v,t,\omega)d\mu_k \] By Fubini Theorem (see \cite{cannarsa2015introduction}), for each closed measure $\mu$, we know that $\int_{TM \times T}L(x,v,t,\omega)d\mu$ is measurable function on $\Omega$. Since the infimum of a countable measurable funtion is measurable on $\Omega$, $\alpha(\omega)$ is measurable. Since $L(x,v,t,\theta(s)\omega)=L(x,v,t+s,\omega)$,for $s \in R$, we know that $\alpha(\theta(s)\omega)=\alpha(\omega)$. If $\{\theta(s)$,$s \in R\}$ is ergodic on $\Omega$, we know that $\alpha(\omega)$ is constant almost everywhere.This finishes the proof of Theorem 1.1 \end{proof} The author has an another simple proof, we take advantage of a useful result from \cite{contreras2013weak} \begin{proposition} If $L:TM \times T \to R$ is a Tonelli Lagrangian, the Mane-Critical Value has another interpretation: $\alpha(0)=\min\{k: \int L+k \ge 0$ for all closed curves $\gamma\}$ A curve $\gamma:[a,b] \to M$ is called closed if $\gamma(a)=\gamma(b)$ and $b-a$ is an integer. \end{proposition} \begin{proof}{second proof of Theorem 1.1} For integers $m<n$, Let $C^{0}([m,n], TM)$ denote the space of continuous functions from $[m,n]$ to $TM$ with the bounded uniform norm. $M$ is a compact manifold, by whitney's theorem, there exists $m \in N$, such that $M$ can be embedded isomorphically into $R^m$. Therefore, $TM$ can be embedded as a submanifold of $R^{2m}$. By Stone-Weierstrass theorem, the space of continuous functions from $[m,n]$ to $R$ is separable. Therefore, the space of continuous functions from $[m,n]$ to $R^{2m}$ is separable. As its subset, $C^{0}([m,n],TM)$ is separable. So $\bigcup_{m<n}C^{0}([m,n],TM)$ is separable. For a Tonelli Lagrangian, the closed curve which integral of Lagrangian attains the minimum along it, satisfies the Euler-Lagrangian equation, and is $C^2$ and compact. So the space of curves, along which Lagrangian Action achieves the minimum, is a subspace of $\bigcup_{m<n}C^{0}([m,n],TM)$.So we can find a sequence of closed curves $\{\gamma_{i}\}$ for $i \in N$, such that \[ \alpha(\omega)=\inf_{i \in N}\int_{\gamma_i}L(\gamma_i(t),\dot{\gamma_i}(t),t,\omega)dt \] By Fubini theorem, for each $i \in N$, $\int_{\gamma_i}L(\gamma_i(t),\dot{\gamma_i}(t),t,\omega)dt$ is a measurable function. As the infinimun of countable measurable functions, we know that $\alpha(\omega)$ is a measurable function. Since $L(x,v,t,\theta(s)\omega)=L(x,v,t+s,\omega)$, for $s \in R$, we get $\alpha(\theta(s)\omega)=\alpha(\omega)$. If $\{\theta(s),s \in R\}$ is ergodic on $\Omega$, we know that $\alpha(\omega)$ is constant almost everywhere. This finishes the proof of Theorem 1.1. \end{proof} \section{Semiconcave Estimates on Lagrangain Action} In this section we have a revision of some basic properties of a class of nonsmooth functions, the so-called semiconcave functions. The good properties of semiconcave functions provide fundamental technical tools for the analysis of singularities of Lagrangian Action and Weak KAM Solutions. It is well known that a real-valued function $u$ is semiconcave in an open demain $U \in R^n$ if, for any compact set $K \subset U$, there exists a constant $C \in R$ such that \[ tu(x_1)+(1-t)u(x_0)- u(tx_1+(1-t)x_0) \le Ct(1-t){|x_1-x_0|}^2 \] for all $t \in [0,1]$ and for all $x_0,x_1 \in K$ satisfying $[x_0,x_1] \subset K$. We refer to such a constant $C$ as a semiconcavity constant for $u$ on $K$. We denote by $SC(U)$ the class of all semiconcave functions in $U$. We review some differentiability properties of semiconcave functions. To begin, let us recall that any $u \in SC(U)$ is locally Lipschitz continuous.(see \cite{cannarsa2004semiconcave}).Hence,by Rademacher's Theorem, $u$ is differentiable a.e in $U$ and the gradient of $u$ is locally bounded. Then, the set \[ D^{*}_{x}=\{p \in R^n: U \ni x_i \to x, Du(x_i) \to p\} \] is nonempty for any $x \in U$. The elements of $D^{*}u(x)$ are called reachable gradients. The superdifferential of any function $u:U \to R$ at a point $x \in U$ is defined as \[ D^{+}_{x}u=\{p \in R^n: \limsup_{h \to 0}\frac{u(x+h)-u(x)-<p,h>}{|h|} \le 0\} \] Similarly, the subdifferential of $u$ at $x$ is given by \[ D^{-}_{x}u=\{p \in R^n: \liminf_{h \to 0}\frac{u(x+h)-u(x)-<p,h>}{|h|} \ge 0\} \] Next, we list some properties: \begin{proposition} Let $u:A \to R$ and $x \in A$. Then the following properties hold true. (1)$D^{+}_{x}u$ and $D^{-}_{x}u$ are closed convex sets. (2)$D^{+}_{x}u$ and $D^{-}_{x}u$ are both nonempty if and only if $u$ is differentiable at $x$; in this case we have that \[ D^{+}_{x}u=D^{-}_{x}u=\{D_{x}u\} \] Furthermore, when $u: U \to R$ be a semiconcave function, we have (3)$D^{+}_{x}u=cov D^{*}_{x}u$ (4)$D^{+}_{x}u \neq \emptyset $ (5)When $D^{+}_{x}u$ is a singleton, $u$ is differentiable at $x$. \end{proposition} The proof of Proposition 5.1 can be seen in \cite{cannarsa2004semiconcave} Given $L: TM \times T \to R $ a Time-Periodic Tonelli Lagrangian. For an absolutely continuous curves $\gamma:[s,t] \to M$, the action of $L$ along $\gamma$ is defined as $A(\gamma)=\int^{t}_{s}L(\gamma(u),\dot{\gamma}(u),u)du$. $\Sigma(s,y;t,x)=\{\gamma: \gamma:[s,t] \to M $ is absolutely continuous , and $\gamma(s)=y,\gamma(t)=x\}$. The Lagrangian action $A(s,y;t,x)$ is defined as \[ A(s,y;t,x)=\min_{\gamma \in \Sigma(s,y;t,x)}\int^{t}_{s}L(\gamma(u),\dot{\gamma}(u),u)du \] If $\gamma:[s,t] \to M$ attains the minimum of $A(s,y;t,x)$, then from variational methods, we know that $\gamma$ satisfies Euler-Lagrange equation and $\gamma$ is $C^2$. \begin{lemma} Fix $s_1<t_1$, for any $s\le s_1,t \ge t_1$, the lagrangian Action $A(s,\cdot;t,\cdot)$ is equi-semiconcave on $M \times M$, therefore equi-Lipschitz. \end{lemma} \begin{proof} To give a proof, we use the variational methods.Fix $\forall x \in M$, we can find a chart such that $x \in U \subset M$. Without loss of generality, we assume that $U$ is a open ball of $R^n$. If $\gamma \in \sum_{m}(s,y;t,x+v)$, we can find $h>0$,such that $\gamma([t-h,t]) \in U$, we can find a Ball $B(0,r) \in R^n$, such that for $v \in B(0,r)$, $s \in [0,h]$, we have $\gamma(t-h+s)+\frac{s}{h}v \in U$. Fix $v \in B(0,r)$, we define $\gamma_h \in \sum(s,y;t,x)$ in the following way: when $s \le u \le t-h$,$\gamma_h(u)=\gamma(u)$; when $t-h \le u \le t$,$\gamma_h(u)=\gamma(u)+\frac{u+h-t}{h}v$. Assume $F_k=\max_{||v||_{U} \le k}||\partial_{vv}L(x,v,t)||_{U}<\infty$, $E_k=\max_{||v||_{U} \le k}||\partial_{xv}L(x,v,t)||_{U}<\infty$. From Tonelli theorem, there exists $K(s_1,t_1)$ such that $|\dot{\gamma}(u)| \le K(s_1,t_1)$ for $s \le u \le t$ where $\gamma \in \Sigma(s,y;t,x)$ is a minimizer. We have the following estimates: \begin{equation}\nonumber \begin{aligned} &A(\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2});t,x+v)-A(\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2});t,x)\\ &\le \int^{t}_{\frac{s_1+t_1}{2}}L(\gamma_h(u),\dot{\gamma_h}(u),u)du-\int^{t}_{\frac{t_1+s_1}{2}}L(\gamma(u),\dot{\gamma}(u),u)\,du\\ &=\int^{t}_{t-h}L(\gamma_h(u),\dot{\gamma_h}(u),u)-L(\gamma(u),\dot{\gamma}(u),u)\,du\\ &\le \int^{t}_{t-h}L(\gamma_h(u),\dot{\gamma_h}(u),u)-L(\gamma_h(u),\dot{\gamma}(u),u)+ L(\gamma_h(u),\dot{\gamma}(u),u)-L(\gamma(u),\dot{\gamma}(u),u)\,du\\ &\le \int^{t}_{t-h}\frac{\partial L}{\partial v}(\gamma_h(u),\dot{\gamma}(u),u)\cdot \frac{1}{h}v+\frac{\partial L}{\partial x}(\gamma(u),\dot{\gamma}(u),u)\cdot \frac{u+h-t}{h}\cdot v\\ &+ \frac{1}{2h^2}F_{K(s_1,t_1)+r}{||v||}^2_{U}+ \frac{1}{2}E_{K(s_1,t_1)}{||v||}^2_{U}\,du \\ &\le \int^{t}_{t-h}\frac{\partial L}{\partial x}(\gamma(u),\dot{\gamma}(u),u)\cdot \frac{u+h-t}{h}\cdot v+ \frac{\partial L}{\partial v}(\gamma(u),\dot{\gamma}(u),u)\cdot \frac{1}{h}v du + \frac{h}{2} E_{K(s_1,t_1)}{||v||}^2_{U}+\\ &+\frac{1}{2 h}F_{K(s_1,t_1)+r}{||v||}^2_{U} + E_{K(s_1,t_1)}{||v||}^2_{U}\\ &=\frac{\partial L}{\partial v}(\gamma(u),\dot{\gamma}(u),u)\frac{u+h-t}{h}\cdot v {\mid}^{t}_{t-h} +E_{K(s_1,t_1)}{||v||}^2_U+ \frac{1}{2h}F_{K(s_1,t_1)+r}{||v||}^2_U + \frac{h}{2}E_{K(s_1,t_1)}{||v||}^2_U\\ &+\int^{t}_{t-h}\{-\frac{d}{dt}\frac{\partial L}{\partial x}(\gamma(u),\dot{\gamma}(u),u)+\frac{\partial L}{\partial x}(\gamma(u),\dot{\gamma}(u),u) \} \cdot \frac{u+h-t}{h} \cdot v \,du \\ &=\frac{\partial L}{\partial v}(\gamma(t),\dot{\gamma}(t),t)\cdot v + \{\frac{h}{2}E_{K(s_1,t_1)}+\frac{1}{2h}F_{K(s_1,t_1)+r}+ E_{K(s_1,t_1)}\}{||v||}^2_U \end{aligned} \end{equation} Similarly, we have \begin{equation}{\nonumber} \begin{aligned} &A(\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2});t,x-v)-A(\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2});t,x)\\ &\le -\frac{\partial L}{\partial v}(\gamma(t),\dot{\gamma}(t),t)\cdot v + \{\frac{h}{2}E_{K(s_1,t_1)}+\frac{1}{2h}F_{K(s_1,t_1)+r}+ E_{K(s_1,t_1)}\}{||v||_U}^2 \end{aligned} \end{equation} So we have \begin{equation}{\nonumber} \begin{aligned} &A(\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2});t,x+v)+A(\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2});t,x-v)\\ &-2A(\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2});t,x) \le \{E_{K(s_1,t_1)}+\frac{1}{h}F_{K(s_1,t_1)+r}+ 2E_{K(s_1,t_1)}\}{||v||_U}^2 \end{aligned} \end{equation} Therefore, there exists a constant $C(U,s_1,t_1)$ such that \begin{equation}{\nonumber} \begin{aligned} &A(\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2});t,x+v)+A(\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2});t,x-v)\\ &-2A(\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2});t,x) \le C(U){||v||_U}^2 \end{aligned} \end{equation} On the other hand, there exists a local chart $V$ such that $y \in V$, there exists $r_{1} > 0$, such that $B(y,r_{1}) \subset V$, for any $w \in B(y,r_{1})$,we have \begin{equation}{\nonumber} \begin{aligned} &A(s,y+w;\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2}))+A(s,y-w;\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2}))\\ &-2 A(s,y;\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2})) \le C(V){||w||_V}^2 \end{aligned} \end{equation} Finally, we integrate the results above, \begin{equation}{\nonumber} \begin{aligned} &A(s,y+w;t,x+v)+A(s,y-w;t,x-v)-2A(s,y;t,x)\\ &\le A(\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2});t,x+v)+A(\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2});t,x-v)\\ &-2 A(\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2});t,x)+ A(s,y+w;\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2}))\\ &A(s,y-w;\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2}))-2 A(s,y;\frac{s_1+t_1}{2},\gamma(\frac{s_1+t_1}{2}))\\ &\le C(U){||v||}^2_{U}+C(V){||w||}^2_{V} \end{aligned} \end{equation} So $A(s,\cdot;t,\cdot)$ is locally equi-semiconcave for $s \le s_1$,$t \ge t_1$. Since $M$ is compact, $A(s,\cdot,t,\cdot)$ is globally equi-semiconcave for $s \le s_1$,$ t \ge t_1$. \end{proof} \begin{lemma} For each minimizing curve $\gamma \in \Sigma(s,y;t,x)$ attaining the minimum of the action $A(s,y;t,x)$, we have \[ p(t)=\partial_v L(x,\dot{\gamma}(t),t,\omega) \in D^{+}_{x}A(s,y;t,x) \] and \[ -p(s)=-\partial_v L(\gamma(s),\dot{\gamma}(s),s) \in D^{+}_{y}A(s,y;t,x) \] \end{lemma} \begin{proof} we find a local chart $U \subset M$, such that $x \in U$. Without loss of generality, we assume that $U$ is a open ball of $R^N$. If $\gamma \in \Sigma(s,y;t,x)$, we can find $h>0$, such that $\gamma([t-h,t]) \in U$, we can find a Ball $B(0,r)$, for $v \in B(0,r)$,we define $\gamma_h \in \sigma(s,y;t,x+v)$ such that $\gamma_h(u)=\gamma(u)$ for $s \le u \le t-h$,and $\gamma_h(u)=\gamma(u)+\frac{u+h-t}{h}v \in U$ when $r$ is small enough. Assume $F_k=\max_{||v||_U \le k }{||\partial_{vv}L(x,v,t)||}_{U}<\infty$, $E_k=\max_{||v||_{U} \le k}{||\partial_{xv}L(x,v,t)||}_{U}<\infty$. From Tonelli Theroem, there exists $K(s,t)$ such that $|\dot{\gamma}(u)| \le K(s,t)$ for $s \le u \le t$. \begin{equation}\nonumber \begin{aligned} &A(s,y,t,x+v)-A(s,y;t,x) \\ &=\int^{t}_{t-h}L(\gamma_h(u),\dot{\gamma_h}(u),u)-L(\gamma(u),\dot{\gamma}(u),u)\,du\\ &\le \int^{t}_{t-h}L(\gamma_h(u),\dot{\gamma_h}(u),u)-L(\gamma_h(u),\dot{\gamma}(u),u)+ L(\gamma_h(u),\dot{\gamma}(u),u)-L(\gamma(u),\dot{\gamma}(u),u)\,du\\ &\le \int^{t}_{t-h}\frac{\partial L}{\partial v}(\gamma_h(u),\dot{\gamma}(u),u)\cdot \frac{1}{h}v+\frac{\partial L}{\partial x}(\gamma(u),\dot{\gamma}(u),u)\cdot \frac{u+h-t}{h}\cdot v\\ &+ \frac{1}{2h^2}F_{K(s_1,t_1)+r}{||v||}^2_{U}+ \frac{1}{2}E_{K(s_1,t_1)}{||v||}^2_{U}\,du \\ &\le \int^{t}_{t-h}\frac{\partial L}{\partial x}(\gamma(u),\dot{\gamma}(u),u)\cdot \frac{u+h-t}{h}\cdot v+ \frac{\partial L}{\partial v}(\gamma(u),\dot{\gamma}(u),u)\cdot \frac{1}{h}v du + \frac{h}{2} E_{K(s_1,t_1)}{||v||_{U}}^2\\ &+ \frac{1}{2 h}F_{K(s_1,t_1)+r}{||v||_{U}}^2 + E_{K(s_1,t_1)}{||v||_{U}}^2 \\ &=\frac{\partial L}{\partial v}(\gamma(u),\dot{\gamma}(u),u)\frac{u+h-t}{h}\cdot v {\mid}^{t}_{t-h} +E_{K(s_1,t_1)}{||v||_U}^2+ \frac{1}{2h}F_{K(s_1,t_1)+r}{||v||_U}^2 + \frac{h}{2}E_{K(s_1,t_1)}{||v||_U}^2\\ &+\int^{t}_{t-h}\{-\frac{d}{dt}\frac{\partial L}{\partial v}(\gamma(u),\dot{\gamma}(u),u)+\frac{\partial L}{\partial x}(\gamma(u),\dot{\gamma}(u),u) \} \cdot \frac{u+h-t}{h} \cdot v \,du \\ &=\frac{\partial L}{\partial v}(\gamma(t),\dot{\gamma}(t),t)\cdot v + \{\frac{h}{2}E_{K(s_1,t_1)}+\frac{1}{2h}F_{K(s_1,t_1)+r}+ E_{K(s_1,t_1)}\}{||v||_U}^2 \end{aligned} \end{equation} So we have \[ \lim_{||v|| \to 0}\frac{A(s,y;t,x+v)-A(s,y;t,x)-\frac{\partial L}{\partial v}(\gamma(t),\dot{\gamma}(t),t)\cdot v}{||v||} \le 0 \] This proves that $\frac{\partial L}{\partial v}(\gamma(t),\dot{\gamma}(t),t) \in D^{+}_{x}A(s,y;t,x)$. In a similar way, we can prove that $-\frac{\partial L}{\partial v}(\gamma(s),\dot{\gamma}(s),s) \in D^{+}_{y}A(s,y;t,x)$ \end{proof} \begin{lemma} Fix $s <t $, $\forall p \in D^{*}_{x}A(s,y;t,x)$, there exists a minimizer $\gamma \in \Sigma(s,y;t,x)$ of $A(s,y;t,x)$ such that $ p=-\frac{\partial L}{\partial v}(\gamma(t),\dot{\gamma}(t),t)$. \end{lemma} \begin{proof} First step, when $A(s,y;t,x)$ is differentiable at $x$, $D^{+}_{x}A(s,y;t,x)=\{D_{x}A(s,y;t,x)\}$. For any minimizer $\gamma \in \Sigma(s,y;t,x)$ of Action $A(s,y;t,x)$, $-\frac{\partial L}{\partial v}(\gamma(t),\dot{\gamma}(t),t)= D_{x}A(s,y;t,x)$. By Euler-Lagrangian equation, the minimizer of $A(s,y,t,x)$ is unique. Second step, when $A(s,y;t,x)$ is not differentiable at $x$, for any $p \in D^{*}_{x}A(s,y;t,x)$, there exists $\{x_n | n =1,2,...\} \subset M$, $A(s,y;t,x)$ is differentiable at $\{x_n, n=1,2...\}$, .and $\lim_{n \to \infty} x_n=x$, $\lim_{n \to \infty} D_{x}A(s,y,t,x_n)=p$. From the first step, we know that there exists $\gamma_n \in \Sigma(s,y,t,x_n)$, such that $-\frac{\partial L}{\partial v}(\gamma_n(t),\dot{\gamma_n}(t),t)=D_{x}A(s,y;t,x_n)$.From Euler-Lagrangian equation, there exists a Lagrangian flow $(\gamma,p):[s,t] \to TM$, such that $\gamma(t)=x$, $p(t)=p$, By continuity of dependence on the initial values of Ordinary Differential Equation, and $\lim_{n \to \infty}(\gamma_n(t),-\frac{\partial L}{\partial v}(\gamma_n(t),\dot{\gamma_n}(t),t))=(x,p)$. so $y=\gamma(s)=\lim_{n \to \infty}\gamma_n(s)$. Therefore, $\gamma \in \Sigma(s,y;t,x)$. By lower semi-continuity of the Action, we have \[\int^{t}_{s}L(\gamma(u),\dot{\gamma}(u),u)du \le \lim_{n \to \infty}\int^{t}_{s}L(\gamma_n(u),\dot{\gamma_n}(u),u)du \] Since $A(s,y,t,x)$ is semi-concave on $M$ and semi-concave functions are lipschitz continuous. We know that $\lim_{n \to \infty}A(s,y;t,x_n)=A(s,y,t,x)$. So $\lim_{n \to \infty}\int^{t}_{s}L(\gamma_n(u),\dot{\gamma_n}(u),u)du =A(s,y,t,x)$. So $\gamma:[s,t] \to M$ is a minimizer of $A(s,y,t,x)$. This finishes the proof. \end{proof} \begin{corollary} The three following conditions are equivalent: (1) $A(s,y;t,x)$ has only one minimizer in $\Sigma(s,y;t,x)$ (2) $A(s,y;t,x)$ is differentiable at $x$. (3) $A(s,y;t,x)$ is differentiable at $y$. \end{corollary} \begin{proof} $(2),(3) \rightarrow (1)$ is the direct consequence of Lemma 5.3. we prove $(1) \rightarrow (2)$. If $A(s,y,t,x)$ is not differentiable at $x$, $D^{+}_{x}A(s,y,t,x)$ contains more than one point. Since $A(s,y,t,x)$ is semi-concave, we know that $D^{+}_{x}A(s,y,t,x)$ is a convex compact set, and is the convex hull of $D^{*}_{x}A(s,y,t,x)$. So $D^{*}_{x}A(s,y,t,x)$ contains more than one point, by lemma 5.3, $A(s,y,t,x)$ has more than one minimizer. $(1) \to (3)$ is similar. \end{proof} \section{Weak KAM Solution} In this section, we use Lax-Oleinik operator to construct a class of Weak KAM Solutions and prove that they are measurable over $\Omega$. \begin{notation} $\Sigma(s,y;t,x)$ is the set of absolutely continuous curves $\gamma:[s,t] \to M $ such that $\gamma(s)=y$ and $\gamma(t)=x$. $\Sigma^{\omega}_{m} (s,y;t,x)$ denotes the set of the minimizers for the Action $A^{\omega}(s,y;t,x)$ $D=C_{0}(M,R)$ is the real-valued continuous function space over $M$ with the uniform topology. $\mathscr{D}$ is the Borel algebra generated by open sets of $D$. $C_{0}(M \times R,R)$ is the real-valued continuous function space over $M \times R$. $f(\omega)=\sup_{x \in M, t \in [0,1]}|L(x,0,s,\omega)|$ is finite since $M$ is compact. $C(\omega)=\sup\{|L(x,v,t)|$ for $x \in M, |v| \le dist(M),t \in [0,1]\}$. \end{notation} \begin{definition} 1. For $\lambda \in R$, $ u \in C_{0}(M,R)$, the operator $T^{\omega}_{\lambda}: C_{0}(M,R) \to C_{0}(M \times R,R)$ is defined as \begin{equation}\nonumber T^{\omega}_{\lambda}(u)(x,t)=\min_{y\in M}\{u(y)+ A^{\omega}(t-\lambda,y;t,x)+\lambda\alpha(\omega)\} \end{equation} 2.For $u \in C_{0}(M,R)$, $u^{\omega}(x,t)$ is defined as \begin{equation}\nonumber u^{\omega}(x,t)=\liminf_{\lambda\to +\infty}{T^{\omega}_{\lambda}(u)(x,t)} \end{equation} \end{definition} 3.For $\lambda \in R$,$ u \in C_{0}(M,R)$, the operator $T^{\omega}_{\lambda,+}: C_{0}(M,R) \to C_{0}(M \times R,R)$ is defined as \begin{equation}\nonumber T^{\omega}_{\lambda,+}(u)(x,t)=\min_{y\in M}\{u(y)+ A^{\omega}(t,x;t+\lambda,y)+\lambda\alpha(\omega)\} \end{equation} 4.For $u \in C_{0}(M,R)$, $u^{\omega}_{+}(x,t)$ is defined as \begin{equation}\nonumber u^{\omega}_{+}(x,t)=\liminf_{\lambda\to +\infty}{T^{\omega}_{\lambda,+}(u)(x,t)} \end{equation} \begin{lemma} When $\omega \in \Omega$. Fix $t \in R$, $x \in M$, $u \in C(M,R)$, the Lax-Oleinik Operator $T^{\omega}_{\lambda}u(x,t)$ is Lipschitz when $\lambda \ge 0$ with Lipschitz constant $|f(\omega)|+|\alpha(\omega)|$. \end{lemma} \begin{proof} For $\lambda_1>\lambda_2\ge 0$, $\forall \epsilon>0$, we can find $y_1,y_2 \in M$, such that \begin{align}{\nonumber} T^{\omega}_{\lambda_1}(u)(x,t)+\epsilon=u(y_1)+A^{\omega}(t-\lambda_1,y_1;t,x)+ \lambda_1\alpha(\omega)\\{\nonumber} T^{\omega}_{\lambda_2}(u)(x,t)+\epsilon=u(y_2)+A^{\omega}(t-\lambda_2,y_1;t,x)+\lambda_2\alpha(\omega){\nonumber} \end{align} then \begin{equation}\nonumber \begin{aligned} &T^{\omega}_{\lambda_1}(u)(x,t)\le u(y_2)+A^{\omega}(t-\lambda_2,y_2;t,x)+\lambda_2\alpha(\omega)+ \int^{t-\lambda_2}_{t-\lambda_1}L(y_2,0,s,\omega)ds\\ &+ (\lambda_1-\lambda_2)\alpha(\omega) \le T^{\omega}_{\lambda_2}(u)(x,t)+\epsilon+(\lambda_1-\lambda_2)(|\alpha(\omega)|+|f(\omega)|) \end{aligned} \end{equation} and \begin{equation}\nonumber \begin{aligned} &T^{\omega}_{\lambda_2}(u)(x,t)\le u(y_1)+A^{\omega}(t-\lambda_1,y_1;t,x)+\lambda_1\alpha(\omega)+ \int^{t-\lambda_1}_{t-\lambda_2}L(y_1,0,s,\omega)ds\\ &+ (\lambda_2-\lambda_1)\alpha(\omega) \le T^{\omega}_{\lambda_1}(u)(x,t)+\epsilon+(\lambda_1-\lambda_2)(|\alpha(\omega)|+|f(\omega)|) \end{aligned} \end{equation} Since $\epsilon>0$ is arbitrary, we have \[ |T^{\omega}_{\lambda_1}u(x,t)-T^{\omega}_{\lambda_2}u(x,t)| \le (\lambda_1-\lambda_2)(|\alpha(\omega)|+|f(\omega)|) \] \end{proof} \begin{lemma} When $\omega$ is fixed, Lax-Oleinik operator is unifromly bounded for continuous function when $\lambda>0$. \end{lemma} \begin{proof} Step 1. Fix $\omega \in \Omega$, $ t\in R$. Define the sequences $M_n(\omega)=\max_{x \in M}T^{\omega}_{n}(0)(x,t)$ and $m_n(\omega)=\min_{x \in M}T^{\omega}_{n}(0)(x,t)$ where $0$ is the zero function on $M$. From Lemma 5.1, the function $T^{\omega}_{n}(0)$,$n\ge 1$, are equi-semi-concave, there exists a constant $K(\omega)$ such that \begin{equation}{\nonumber} 0 \le M_n(\omega) - m_n(\omega) \le K(\omega). \end{equation} for $n \ge 1$. We claim that $M_{n+m}(\omega) \le M_{n}(\omega) + M_{m}(\omega)$. This follows from the inequalities \begin{equation}{\nonumber} T^{\omega}_{n+m}(0)(x,t)=T^{\omega}_{m}(T^{\omega}_{n}(0)(x,t) \le T^{\omega}_{m}(M_n(\omega))(x,t)\le M_n(\omega) + T^{\omega}_{m}(0)(x,t) \end{equation} Hence by a classical result on subadditive sequences, we have $\lim \frac{M_n(\omega)}{n}=\inf\frac{ M_n(\omega)}{n}$. We denote by $-\beta(\omega)$ this limit. In the same way, the sequence $-m_{n}(\omega)$ is subadditive, hence $\frac{m_n(\omega)}{n} \to \sup \frac{m_n(\omega)}{n}$.This limit is also $-\beta(\omega)$ since $0\le M_n(\omega)-m_n(\omega) \le K$.Note that $m_1(\omega) \le -\beta(\omega) \le M_1(\omega)$, so that $\beta(\omega)$ is indeed a finite number. We have, for all $n \le 1$, \begin{equation}{\nonumber} -K(\omega)-n\beta(\omega) \le m_n(\omega) \le -n\beta(\omega) \le M_n(\omega) \le K(\omega)-n\beta(\omega) \end{equation} Now for all $u \in C(M,R)$,$n\in N$ and $x \in M$, we have \[ \min_{M}u-K(\omega) \le \min_{M}u + m_n(\omega) + n\beta(\omega) \le T^{\omega}_{n}u(x,t) + n\beta(\omega) \le \max_{M}u + M_{n}(\omega) +n \beta(\omega) \le \max_{M}u+K(\omega) \] Hence, for all $u \in C(M,R)$,$n \in N$,$x \in M$, we have \begin{equation}{\nonumber} \frac{\min_{M}u-K(\omega)}{n} \le \frac{T^{\omega}_{n}u(x,t)}{n} +\beta(\omega) \le \frac{\max_{M}u+K(\omega)}{n} \end{equation} Step 2.On one hand, from Proposition 4.1 , we can find a sequence of measure $\frac{1}{n_k}[\gamma_{n_k}]$ where $n_k$ is a sequence of increasing intergers towards $+\infty$, $\gamma_{n_k}$ is a closed absolutely continuous curve from $[t-n_k,t]$ to $M$, such that $\forall \epsilon>0$, we can find an interger $N$, such that, when $k \ge N$, we have \begin{equation}{\nonumber} -\alpha(\omega)\le \frac{1}{n_k}\int_{t-n_k}^{t}L(\gamma_{n_k}(s),\dot{\gamma}_{n_k}(s),s,\omega)ds \le -\alpha(\omega) + \epsilon \end{equation} Hence, we have \begin{equation}{\nonumber} T^{\omega}_{n_k}(0)(\gamma_{n_k}(t),t) \le \int_{-n_k+t}^{t}L(\gamma_{n_k}(s),\dot{\gamma}_{n_k}(s),s,\omega)dt+n_k\alpha(\omega) \le n_k \epsilon \end{equation} Therefore, we have \begin{equation}{\nonumber} \frac{1}{n_k}T^{\omega}_{n_k}(0)(\gamma_{n_k}(t),t) \le \epsilon \end{equation} On the other hand, there is $x \in M$ such that \[ T^{\omega}_{n_k}(0)(\gamma_{n_k}(t),t)=A^{\omega}(x,t-n_k;\gamma_{n_k}(t),t)+n_k\alpha(\omega) \] We assume that $\gamma_1 \in \Sigma^{\omega}_{m}(x,t-n_k;\gamma_{n_{k}}(t),t)$. there is an absolutely continuous path $\gamma:[t-n_k-1,t-n_k] \to M$ such that $|\dot{\gamma}(s)|\le dist(M)$, and $\gamma(t-n_k)=x$, $\gamma(t-n_k-1)=\gamma_{n_k}(t)$. Then we know that \[ \int^{t-n_k}_{t-n_k-1}|L(\gamma(s),\dot{\gamma}(s),s,\omega)|ds \le C(\omega) \] Construct a new closed curve $\tilde{\gamma}:[t-n_k-1,t] \to M$ in the following way: when $s \in [t-n_k,t]$, $\tilde{\gamma}(s)=\gamma_1(s)$; when $s \in [t-n_k-1,t-n_k]$, $\tilde{\gamma}(s)=\gamma(s)$. From proposition 4.1, we know that \[ \int^{t}_{t-n_k-1}L(\tilde{\gamma}(s),\dot{\tilde{\gamma}}(s),s,\omega)ds \ge -(n_k+1)\alpha(\omega) \] Hence, we have \begin{equation}\nonumber \begin{aligned} &T^{\omega}_{t-n_k}(0)(\gamma_{n_k}(t),t)=n\alpha(\omega)+ \int^{t}_{t-n_k-1}L(\tilde{\gamma}(s),\dot{\tilde{\gamma}}(s),s,\omega)-\int^{t-n_k}_{t-n_k-1}L(\gamma(s),\dot{\gamma}(s),s,\omega)ds \\ &\ge -\alpha(\omega)-C(\omega)\\ \end{aligned} \end{equation} So we have \[ -\frac{\alpha(\omega)+C(\omega)}{n_k} \le \frac{T^{\omega}_{t-n_k}(0)(\gamma_{n_k}(t),t)}{n_k} \le \epsilon \] So combine the result with step 2, let $k$ tends to $\infty$, we know that $\beta(\omega)=0$. Hence, \begin{equation}{\nonumber} \min_{M}u-K(\omega) \le T^{\omega}_{n}u(x,t)\le \max_{M}u+K(\omega) \end{equation} Use Lemma 6.1, $T^{\omega}_{\lambda}u(x,t)$ is lipschitz for $\lambda$, with Lipschitz constant $|\alpha(\omega)|+|f(\omega)|$, so we have \[ \min_{M}u-K(\omega)-|\alpha(\omega)|-|f(\omega)| \le T^{\omega}_{\lambda}u(x,t) \le \max_{M}u + K(\omega) + |\alpha(\omega)| + |f(\omega)| \] \end{proof} \begin{lemma}If the variables $t,s \in R$, $x,y \in M$, $n \in \mathbb{N}$, $u \in C_{0}(M,R)$ are fixed, $A^{\omega}(s,y;t,x)$, $T^{\omega}_{\lambda}(u)(x,t)$, $T^{\omega}_{\lambda,+}(u)(x,t)$ and $u^{\omega}(x,t)$, $u^{\omega}_{+}(x,t)$ are random variables over the probability space $\Omega$. \end{lemma} \begin{proof} The extreme curves of the Action are $C^2$ since they satisfy the lagrangian equations. Let $C^{0}([s,t],TM)$ denote the space of continuous function from $[s,t]$ to $TM$ with uniform norm in $TM$. $M$ is a manifold, by whitney' embedding theorem, there exists $m \in N$, such that $M$ can be embedded isomorphically into $R^m$. Therefore, $TM$ can be embedded as a submanifold of $R^{2m}$. By Stone Weierstrass theorem, the space of continuous functions from $[s,t]$ into $R$ is separable. Therefore, the space of continuous functions from $[s,t]$ to $R^{2m}$ is separable, and as its subset, $C^{0}([s,t],TM)$ is also separable. Since the extreme curves of the Action is a subset of $C^{0}([s,t],TM)$, So we can pick up a countable dense subset $\{\gamma_i\}_{i \in N}$,such that \begin{equation}\nonumber A^{\omega}(s,y;t,x)=\min_{i \in N}\int^{t}_{s}L(\gamma_i(\sigma),\dot{\gamma_i}(\sigma),\sigma,\omega)d\sigma \end{equation} By Fubini Thoerem, $\int^{t}_{s}L(\gamma_i(\sigma),\dot{\gamma_i}(\sigma),\sigma,\omega)d\sigma$ is measurable function on $\Omega$. As a minimun of countable measurable functions, $A^{\omega}(s,y;t,x)$ is measurable. $M$ is a separable metric space, $u(y)$, $A^{\omega}(t-\lambda,y;t,x)$ are continuous functions with $y$, so we can find a countable dense set $\{y_i\}_{i\in N}$ in $M$ such that \begin{equation}\nonumber T^{\omega}_{\lambda}(u)(x,t)=\min_{i \in N}\{u(y_i)+ A^{\omega}(t-\lambda,y_i;t,x)+\lambda\alpha(\omega)\} \end{equation} As a minimum of countable measurable functions, $T^{\omega}_{\lambda}(u)(x,t)$ is a random variable. Since $T^{\omega}_{\lambda}(u)(x,t)$ is uniformly continuous with $\lambda$, so we can a pick a sequnce $\lambda_n=\sum_{k=1}^{n}\frac{1}{k} \to \infty$. Then, we have \begin{equation}\nonumber u^{\omega}(x,t)=\liminf_{n \to \infty}T^{\omega}_{\lambda_n}(u)(x,t) \end{equation} So as the infimum limit of a countable sequence of measurable function, $u^{\omega}(x,t)$ is measurable over $\Omega$. The measurability of $T^{\omega}_{\lambda,+}(u)(x,t)$,$u^{\omega}_{+}(x,t)$ can be proved in a similar way. \end{proof} \begin{lemma} Fix $u\in C_{0}(M,R)$, for $\lambda>0 ,t,s \in R$, $x \in M $, we have the following formula: 1.\begin{equation}\nonumber T^{\theta(s)\omega}_{\lambda}(u)(x,t)=T^{\omega}_{\lambda}(u)(x,t+s) \end{equation} 2.\begin{equation}\nonumber u^{\omega}(x,t)=T^{\omega}_{t-s}(u^{\omega}(\;,s))(x,t)=\min_{y\in M}\{u^{\omega}(y,s)+A^{\omega}(s,y;t,x)+(t-s)\alpha(\omega)\} \end{equation} 3.\begin{equation}\nonumber u^{\theta(s)\omega}(x,t)=u^{\omega}(x,t+s) \end{equation} 4.\begin{equation}\nonumber u^{\omega}(x,t)=u^{\omega}(x,t+1) \end{equation} These formula have similar versions for $T^{\omega}_{\lambda,+}(u)(x,t)$,$u^{\omega}_{+}(x,t)$. \end{lemma} \begin{proof} First of all, the formula 1 can be derived directly from $L(x,v,t,\theta(s)\omega)=L(x,v,t+s,\omega)$. Secondly,by definition, $\forall \epsilon>0$,there exists $y\in M$,such that \begin{equation}\nonumber T^{\omega}_{t-s}u^{\omega}( ,s)(x,t)+\epsilon=u^{\omega}(y,s)+ A^{\omega}(s,y;t,x)+(t-s)\alpha(\omega) \end{equation} By defnition, there exists a sequence $\lambda_i \to +\infty$ as $i \to +\infty$, such that \begin{align*} u^{\omega}(y,s)=\liminf_{i \to +\infty}T^{\omega}_{\lambda_i}(u)(y,s) \end{align*} Therefore, we have \begin{equation}\nonumber \begin{aligned} &T^{\omega}_{t-s}u^{\omega}( ,s)(x,t)+\epsilon =\liminf_{i \to +\infty}T^{\omega}_{\lambda_i}(u)(y,s)+A^{\omega}(s,y;t,x)+(t-s)\alpha(\omega) \\ &\ge \liminf_{i\to +\infty}T^{\omega}_{\lambda_i+t-s}(u)(x,t) \ge \liminf_{\lambda \to +\infty}T^{\omega}_{\lambda}(u)(x,t)=u^{\omega}(x,t) \end{aligned} \end{equation} Since $\epsilon>0$ is arbitrary, we can get that \begin{equation}\nonumber T^{\omega}_{t-s}u^{\omega}( ,s)(x,t)\ge u^{\omega}(x,t) \end{equation} Conversely, there exists $\lambda_k \to +\infty$ as $k \to +\infty$, such that \begin{equation}\nonumber u^{\omega}(x,t)=\lim_{k \to +\infty}T^{\omega}_{\lambda_k}(u)(x,t)=\lim_{k \to +\infty}T^{\omega}_{t-s}(T^{\omega}_{\lambda_k-t+s}(u))(x,t) \end{equation} By definition, $\forall \epsilon > 0$, there exists $\{q_k\}_{k\in N} \in M$,such that \begin{equation}\nonumber T^{\omega}_{\lambda_k}(u)(x,t)\ge T^{\omega}_{\lambda_k-t+s}(u)(q_k,s)+A^{\omega}(q_k,s;x,t)+(t-s) \alpha(\omega)-\epsilon \end{equation} Since $M$ is a compact manifold, without loss of generality, we can assume that $q_k \to q$ as $k \to \infty$ for some $q\in M$. so \begin{equation}\nonumber \begin{aligned} &u^{\omega}(x,t)\ge \lim_{k\to \infty}T^{\omega}_{\lambda_k-t+s}(u)(q_k,s)+A^{\omega}(q_k,s;x,t)+(t-s) \alpha(\omega)-\epsilon \\ &\ge \liminf_{k \to \infty}T^{\omega}_{\lambda_k-t+s}(u)(q,s)+ A^{\omega}(q,s;x,t)+ (t-s)\alpha(\omega)-\epsilon\\ &\ge u^{\omega}(q,s)+A^{\omega}(q,s;x,t)+(t-s)\alpha(\omega)-\epsilon\\ &\ge T^{\omega}_{t-s}(u( ,s))(x,t)-\epsilon \end{aligned} \end{equation} Since $\epsilon>0$ is arbitrary, we have that \begin{equation}\nonumber u^{\omega}(x,t) \ge T^{\omega}_{t-s}u^{\omega}( ,s)(x,t) \end{equation} Consequently, the formula 2 holds. Thirdly, the formula 3 arises directly from formula 1. Finally, we can get the formula 4 from $L(x,v,t+1,\omega)=L(x,v,t,\omega)$. \end{proof} \begin{theorem} Fix $u \in C_{0}(M,R)$, $u^{\omega}(x,t)$ is a viscosity solution of the Hamilton-Jacobi Equation \begin{equation}\nonumber \partial_t u(x,t) + H(x,\partial_x u(x,t),t,\omega)=\alpha(\omega) \end{equation} \end{theorem} \begin{proof} Fix $\omega \in \Omega$, for any $(x_0,t_0) \in M$. We can find a local chart $(U,\phi)$, $U \in M$,$\phi:U \to \phi(U) \in R^n$ is a diffeomorphism, $(x_0,t_0) \in U$. Without loss of generality, we assume that $U \in R^n$. Step 1.To show $u^{\omega}(x,t)$ is a subsolution of the Hamilton-Jacobi equation, we need to prove that if $(p_x,p_t) \in D^{+}u^{\omega}(x_0,t_0)$, $p_t+H(x,p_x,t,\omega) \le \alpha(\omega)$. $\forall v \in T_{x_0}M$, since $(p_x,p_t) \in D^{+}u^{\omega}(x_0,t_0)$, we have \begin{equation}{\nonumber} \limsup_{h \to 0+} \frac{u^{\omega}(x_0-hv,t_0-h)-u^{\omega}(x_0,t_0)+h(p_t+p_x \cdot v)}{h\sqrt{1+{|v|}^2}}\le 0 \end{equation} Which is equivalent to \begin{equation}{\nonumber} \limsup_{h \to 0+} \frac{u^{\omega}(x_0-hv,t_0-h)-u^{\omega}(x_0,t_0)}{h}\le -p_t - p_x \cdot v \end{equation} Since $U$ is open, there exists $\sigma>0$, such that$\{\gamma(t)=x-s\cdot v| 0 \le s \le \sigma \}\in U$, from lemma 4, we know that when $0<h\le\sigma$, we have \begin{equation}{\nonumber} u^{\omega}(x_0,t_0) \le u^{\omega}(x_0-h\cdot v,t_0-h)+\int^{t_0}_{t_0-h}L(\gamma(s),\dot{\gamma}(s),s,,\omega)ds + h\cdot \alpha(\omega) \end{equation} Then, \begin{equation}{\nonumber} -\alpha(\omega)-\liminf_{h \to 0+}\frac{1}{h}\int^{t_0}_{t_0-h}L(\gamma(s),\dot{\gamma}(s),s,\omega)ds \le \limsup_{h \to 0+}\frac{u(x_0-hv,t_0-h)-u(x_0,t_0)}{h} \le -p_t - p_x \cdot v \end{equation} Hence, we have \begin{equation}{\nonumber} -\alpha(\omega)-L(x_0,v,t_0,\omega) + p_x \cdot v+p_t\le 0 \end{equation} Therefore, \begin{equation}{\nonumber} p_t+H(x_0,v,t_0,\omega)-\alpha(\omega) =p_t+ \sup_{v \in TM}\{p_x \cdot v -L(x_0,v,t_0,\omega)\}-\alpha(\omega) \le 0 \end{equation} Step 2.To show $u^{\omega}(x,t)$ is a subsolution of the Hamilton-Jacobi equation, we need to prove that if $(p_x,p_t) \in D^{+}u^{\omega}(x_0,t_0)$, $p_t + H(x,p_x,t,\omega) \ge \alpha(\omega)$. From Lemma 6.4, we know that there exists $y \in M$, an absolute curve $\{\gamma(t)| t_0-1 \le t \le t_0,\gamma(t_0-1)=y,\gamma(t_0)=x_0\}$,such that \begin{equation}{\nonumber} u^{\omega}(x_0,t_0)=u^{\omega}(y,t_0-1)+\int^{t_0}_{t_0-1}L(\gamma(t),\dot{\gamma}(t),t,\omega)dt+(t_0-t)\alpha(\omega) \end{equation} Since $U$ is open, there exists $\sigma>0$, such that $\{\gamma(t)|t_0-\sigma \le t \le t_0 \}\in U$. Then we have \begin{equation}{\nonumber} u^{\omega}(x_0,t_0)=u^{\omega}(\gamma(t_0-h),t_0-h)+\int^{t_0}_{t_0-h}L(\gamma(t),\dot{\gamma}(t),t,\omega)dt + h\alpha(\omega) \end{equation} Let $w=\dot{\gamma}(t_0)$,since $(p_t,p_x) \in D^{-}u^{\omega}(x_0,t_0)$, we know that \begin{equation}{\nonumber} \liminf_{h \to 0+}\frac{u^{\omega}(\gamma(t_0-h),t_0-h)-u^{\omega}(\gamma(t_0),t_0)}{h}\ge -p_x\cdot w -p_t \end{equation} Hence, we have \begin{equation}{\nonumber} -p_x\cdot w-p_t \le -\limsup_{h \to 0+} \frac{1}{h}\int^{t_0}_{t_0-h}L(\gamma(t),\dot{\gamma}(t),t,\omega)dt -\alpha(\omega) \le -L(x_0,w,t_0,\omega) -\alpha(\omega) \end{equation} Therefore,we have \begin{equation}{\nonumber} H(x,p_x,t,\omega)=\sup_{v \in T_{x_0}M}\{p_x\cdot v-L(x_0,v,t_0,\omega)\}\ge p_x\cdot w-L(x_0,w,t_0,\omega) \ge \alpha(\omega)-p_t \end{equation} From Step 1 and Step 2, we know that $u^{\omega}(x,t)$ is a viscosity solution of Hamilton-Jacobi equation. \end{proof} \section{Global Minimizer} In this section, we discuss the global minimizer and invariant measure under the skew-product dynamics system. \begin{definition} When $\omega$ is fixed, 1.an absolutely continuous orbit $\{\gamma^{\omega}(t), t\in R \}$ is called a global minimizer of the Lagrangian if for any fixed time interval $[s,t]$, we have \begin{equation}\nonumber A^{\omega}(s,\gamma^{\omega}(s);t,\gamma^{\omega}(t))=\int_{s}^{t}L(\gamma^{\omega}(\sigma),\dot{\gamma^{\omega}}(\sigma),\sigma,\omega)d\sigma \end{equation} 2. fix $t_0 \in R$,an absolutely continuous orbit $\{\gamma^{\omega}(t), t \le t_0 \}$ is called a calibrated curve of $u^{\omega}(x,t)$,if for any $s<t \le t_0$, we have \begin{equation}\nonumber u^{\omega}(\gamma^{\omega}(t),t)=u^{\omega}(\gamma^{\omega}(s),s)+ A^{\omega}(s,\gamma^{\omega}(s);t,\gamma^{\omega}(t))+(t-s)\alpha(\omega) \end{equation} 3. fix $t_0 \in R$,an absolutely continuous orbit $\{\gamma^{\omega}(t), t \ge t_0 \}$ is called a calibrated curve of $u^{\omega}_{+}(x,t)$,if for any $s > t \ge t_0$, we have \begin{equation}\nonumber u^{\omega}_{+}(\gamma^{\omega}(t),t)=u^{\omega}_{+}(\gamma^{\omega}(s),s)+ A^{\omega}(t,\gamma^{\omega}(t);s,\gamma^{\omega}(s))+(s-t)\alpha(\omega) \end{equation} \end{definition} \begin{lemma} We fix $\omega$ and $t_0$, if $u^{\omega}(x,t_0)$ is differentiable at $x_0 \in M$, there is a unique calibrated curve with the end point $(x_0,t_0)$, for $u^{\omega}(x,t)$, $t \le t_0$; denoted as $\gamma^{\omega,-}_{x,t_0}$. Similarly,if $u^{\omega}_{+}(x,t_0)$ is differentiable at $x_0 \in M$, there is a unique calibrated curve with the end point $(x_0,t_0)$, for $u^{\omega}_{+}(x,t)$, $t \ge t_0$; denoted as $\gamma^{\omega,+}_{x,t_0}$. \end{lemma} \begin{proof} By lemma 3.3, for any $s<t_0$, $x_0 \in B^{\omega}_{t_0}$, we have \begin{equation}\nonumber u^{\omega}(x_0,t)=\min_{y\in M}\{u^{\omega}(y,s)+A^{\omega}(s,y;t,x)+(t-s)\alpha(\omega)\} \end{equation} Since $u^{\omega}(y,s)$ and $A^{\omega}(s,y;t,x)$ is semiconcave with respect to $y$, so there is $x(s) \in M$ such that \begin{equation}\nonumber u^{\omega}(x_0,t)=u^{\omega}(x(s),s)+A^{\omega}(s,x(s);t,x_0)+(t-s)\alpha(\omega) \end{equation} and $\partial_x u^{\omega}(x(s),s)+\partial_y A^{\omega}(s,x(s);t,x_0)=0$. By Corollary 5.1 , $ \Sigma^{\omega}_{m}(s,x(s);t_0,x_0)$ contains only one lagrangian trajectory $(x(t),p(t))$, $p(t)=\partial_v L(x(s),\dot{x}(s),s,\omega)$ with $s \le t \le t_0$, with $x(t_0)=x_0$, $p(s)=\partial_x u^{\omega}(x(s),s)=-\partial_y A^{\omega}(s,x(s);t,x_0)$ where $p(t)=\partial_x u^{\omega}(x(t_0),t_0)=\partial_x A^{\omega}(s,x(s);t,x_0)$. Therefore, it is obvious that $x(s)$ is unique. For any $s_1 < s_2 < t_0$, the above lagrangian trajectory coincides for $s_2 \le t \le t_0$. So we can extend the trajectory $x(s)$ for time $-\infty < s \le t_0$, with $x(t_0)=x_0$. This is a unique calibrated curve for $u^{\omega}(x,t)$, $t \le t_0$ with the end point a $x(t_0)=x_0$. \end{proof} \begin{lemma} When $\omega$, $t_0$ is fixed, the set $B^{\omega}_{t_0}$ is defined as the subsets of $M$ where the function $u^{\omega}(y,t_0)+u^{\omega}_{+}(y,t_0)$ attains the minimum. We have the following results: 1. $B^{\omega}_{t_0}$ is a closed nonempty subset of $M$. $ u^{\omega}(x,t)$, $ u^{\omega}_{+}(x,t)$ are differentiable over $B^{\omega}_{t_0}$; 2. for any $x_0 \in B^{\omega}_{t_0}$, $\partial_{x}u^{\omega}(x_0,t)+\partial_{x}u^{\omega}_{+}(x_0,t)=0$. 3. $B^{\theta(s)\omega}_{t}=B^{\omega}_{t+s}$, $B^{\omega}_{t+1}=B^{\omega}_{t}$ \end{lemma} \begin{proof} It is obvious that $B^{\omega}_{t_0}$ is closed nonempty subset. Since $u^{\omega}(x,t_0)$,$u^{\omega}_{+}(x,t_0)$ are both semiconcave function over $M$, they are differentiable over $B^{\omega}_{t_0}$ and $\partial_{x}u^{\omega}(x_0,t_0)+\partial_{x}u^{\omega}_{+}(x_0,t_0)=0$. By lemma 3.3, $B^{\theta(s)\omega}_{t}=B^{\omega}_{t+s}$, $B^{\omega}_{t+1}=B^{\omega}_{t}$ are obvious. \end{proof} \begin{proposition} When $\omega$ is fixed, for any $t_0 \in R$, if the following hypothesis holds: $u^{\omega}(\gamma^{\omega}(t_0),t_0)+u^{\omega}_{+}(\gamma^{\omega}(t_0),t_0)=\min_{x \in M}\{u^{\omega}(x,t_0)+u^{\omega}_{+}(x,t_0) \}$ $\{\gamma^{\omega}(t),t \le t_0\}$ is a calibrated curve of $u^{\omega}(x,t)$, $\{\gamma^{\omega}(t),t \ge t_0\}$ is a calibrated curve of $u^{\omega}_{+}(x,t)$, then $\{\gamma^{\omega}(t), t\in R \}$ is a global minimizer of the lagrangian. \end{proposition} \begin{proof} If the conclusion falses, there exists time $t_1<t_2$, and an absolutely continuous curve $\{\gamma_1(t),t_1\le t \le t_2 \}$ with the terminal points $\gamma_1(t_1)=\gamma^{\omega}(t_1)$,$\gamma_1(t_2)=\gamma^{\omega}(t_2)$, such that \begin{equation}\nonumber \int_{t_1}^{t_2}L(\gamma_1(\sigma),\dot{\gamma_1}(\sigma),\sigma,\omega)d\sigma <\int_{t_1}^{t_2}L(\gamma^{\omega}(\sigma),\dot{\gamma^{\omega}}(\sigma),\sigma,\omega)d\sigma \end{equation} then there exist $t_0 \in (t_1,t_2)$ with $\gamma_1(t_0)\ne\gamma^{\omega}(t_0)$.We have the following: \begin{equation}\nonumber \begin{aligned} &u^{\omega}(\gamma_1(t_0),t_0)+u^{\omega}_{+}(\gamma_1(t_0),t_0) \le u^{\omega}(\gamma_1(t_1),t_1)+\int_{t_1}^{t_0}L(\gamma_1(\sigma),\dot{\gamma_1}(\sigma),\sigma,\omega)d\sigma \\ & + (t_0-t_1)\alpha(\omega)+u^{\omega}_{+}(\gamma_1(t_2),t_2)+ \int_{t_0}^{t_2}L(\gamma_1(\sigma),\dot{\gamma_1}(\sigma),\sigma,\omega)d\omega + (t_2-t_0)\alpha(\omega) \\ &=u^{\omega}(\gamma_1(t_1),t_1)+ u^{\omega}_{+}(\gamma_1(t_2),t_2)+\int_{t_1}^{t_2}L(\gamma_1(\sigma),\dot{\gamma_1}(\sigma),\sigma,\omega)d\omega+ (t_2-t_1)\alpha(\omega) \end{aligned} \end{equation} and \begin{equation}\nonumber \begin{aligned} &u^{\omega}(\gamma^{\omega}(t_0),t_0)+u^{\omega}_{+}(\gamma^{\omega}(t_0),t_0)=u^{\omega}(\gamma^{\omega}(t_1),t_1)+ \int_{t_1}^{t_0}L(\gamma^{\omega}(\sigma),\dot{\gamma^{\omega}}(\sigma),\sigma,\omega)d\omega\\ &+ (t_0-t_1)\alpha(\omega)+u^{\omega}_{+}(\gamma^{\omega}(t_2),t_2) + \int_{t_0}^{t_2}L(\gamma^{\omega}(\sigma),\dot{\gamma^{\omega}}(\sigma),\sigma,\omega)d\sigma + (t_2-t_0)\alpha(\omega)\\ &=u^{\omega}(\gamma^{\omega}(t_1),t_1)+u^{\omega}_{+}(\gamma^{\omega}(t_2),t_2)+\int_{t_1}^{t_2}L(\gamma^{\omega}(\sigma),\dot{\gamma^{\omega}}(\sigma),\sigma,\omega)d\sigma+(t_2-t_1)\alpha(\omega) \end{aligned} \end{equation} So, $u^{\omega}(\gamma_1(t_0),t_0)+u^{\omega}_{+}(\gamma_1(t_0),t_0)<u^{\omega}(\gamma^{\omega}(t_0),t_0)+u^{\omega}_{+}(\gamma^{\omega}(t_0),t_0)$, this contradicts the hypothesis. Hence we can complete the proof. \end{proof} \begin{corollary} For fixed $\omega$, $t_0$; for any $x_0 \in B^{\omega}_{t_0}$, a global minimizer $\gamma^{\omega}_{x_0,t_0}$ exists, $\gamma^{\omega}_{x_0,t_0}(t_0)=x_0$.In addition, the set of global minimizer is nonempty for each $\omega \in \Omega$. \end{corollary} \begin{assumption} $(\Omega,\mathscr{F},\mathbb{P})$ is a Polish Space. \end{assumption} Starting from here, we assume Assumption 7.1 holds. From the analysis above, fix $\omega$, the set of global minimizer is nonempty. We denote the set of global minimizer as $H_{\omega}=\{\gamma^{\omega}:R \to M | \gamma^{\omega} $is a global minimizer of action $ A^{\omega}\}$. If $\gamma^{\omega}$ is a global minimizer, $\gamma^{\omega}$ satisfies the Euler-Lagrange equation. $\gamma^{\omega}$ is decided by $(\gamma^{\omega}(0),{\dot{\gamma}}^{\omega}(0))$. So we can define the set $G_{\omega}=\{(\gamma^{\omega}(0),{\dot{\gamma}}^{\omega}(0)|\gamma^{\omega}$ is a global minimizer of the action $A^{\omega} \}$. $G_{\omega}$ is a compact subset of $TM$, due to Tonelli Theorem and continuity of Lagrangian Action. $G_{\omega}$ and $H_{\omega}$ corresponds one to one, $L(x,v,t+s,\omega)=L(x,v,t,\theta(s)\omega)$. if $\gamma^{\omega}(t)$ is a global minimizer of the action $A^{\omega}$, then $\gamma^{\theta(s)\omega}(t)=\gamma^{\omega}(t+s)$ is the global minimizer of $A^{\theta(s)\omega}$, we define the one-to-one map $\Theta(s): G_{\omega} \to G_{\theta(s)\omega}$ as follows, if $(\gamma^{\omega}(0),\dot{\gamma}^{\omega}(0)) \in G_{\omega}$, let $ \Theta(s)(\gamma^{\omega}(0),{\dot{\gamma}}^{\omega}(0))=(\gamma^{\theta(s)\omega}(0),{\dot\gamma}^{\theta(s)\omega}(0))=(\gamma^{\omega}(-s),{\dot{\gamma}}^{\omega}(-s))$.When $\omega$ is fixed, $\Theta(-s)$ is the Lagrangian flow the global minimizer on $TM$. $\Theta(s+t)=\Theta(s)\Theta(t)$.$G_{\omega}=G_{\theta(n)\omega}$. Let $G=\bigcup_{\omega \in \Omega}G_{\omega}\times \{\omega\} \in TM \times \Omega$. Define $\Gamma(s):G \to G$, if $(x,v,\omega) \in G$, $\Gamma(s)(x,v,\omega)=({\Theta(s)|}_{G_{\omega}}(x,v),\theta(s)\omega)$.It is clear that $\Gamma(s+t)=\Gamma(s)\Gamma(t)$. We define the set $\Lambda=\{\mu|\mu$ is a probability measure with support in $G \} $. $\Lambda$ is a convex set. \begin{lemma} Fix $\omega \in \Omega$, $ P(G_{\omega})$ is the convex hull spanned by $\{\delta_{(x,v)}|(x,v) \in G_{\omega}\}$. \end{lemma} \begin{proof} We claim that $F_{\omega}=\{\sum^{n}_{k=1}a_k \delta_{(x_k,v_k)}|n \in N, \sum^{n}_{k=1}a_k=1, 0 \le a_k \le 1,(x_k,v_k) \in G_{\omega} \}$ are dense in $P(G_{\omega})$. $G_{\omega}$ is a compact subset of the complete separable metric space $TM$ with distance $d_{TM}$. The proof that $F_{\omega}$ is dense in $P(G_{\omega})$ is as same as that of Lemma 3.3. See \cite{billingsley2013convergence}. Hence, $P(G_{\omega})$ is the convex hull spanned by $\{\delta_{(x,v)}|(x,v) \in G_{\omega}\}$ \end{proof} \begin{lemma} If $G$ is measurable, For any $\mu \in \Lambda$, for almost every $\omega \in \Omega$, there exists $\mu_{\omega} \in P(G_{\omega})$, such that for any measurable function $f:TM \times \Omega \to R$,we have \[ \int_{G}f(x,v,\omega)d\mu= \int_{\Omega}\int_{G_{\omega}}f(x,v,\omega)d\mu_{\omega}d\omega \] \end{lemma} \begin{proof} If $\mu \in \Lambda$, then $\mu \in P(TM \times \Omega)$, since $TM$ and $\Omega$ are polish spaces, by Theorem 3.1, for almost every $\omega \in \Omega$, there exists $\nu_{\omega} \in P(TM)$, such that for any measurable function $f:TM \times \Omega \to R$, we have \[ \int_{TM \times \Omega}f(x,v,\omega)d\mu=\int_{\Omega}\int_{TM}f(x,v,\omega)d\nu_{\omega}d\omega \] Then, let $\mu_{\omega}=\chi_{G_{\omega}}\nu_{\omega} \in P(G_{\omega})$, we know that \begin{equation}\nonumber \begin{aligned} &\int_{TM \times \Omega}f(x,v,\omega)d\mu=\int_{TM \times \Omega}f(x,v,\omega)\chi_{G} d\mu\\ &=\int_{\Omega}\int_{TM}f(x,v,\omega)\chi_{G_{\omega}}d\nu_{\omega}d\omega=\int_{\Omega}\int_{TM}f(x,v,\omega)d\mu_{\omega}d\omega\\ \end{aligned} \end{equation} \end{proof} Le $\pi_{\Omega}:TM \times \Omega \to \Omega$ as the canonical projection. \begin{lemma} If $\mu \in \Lambda$, and $\mu$ is invariant under the transformation of $\Gamma(s)$, $s \in R$. We decompose $d\mu=d\mu_{\omega}d\omega$, where $\mu_{\omega}$ is a probability measure on $G_{\omega}$ for almost all $\omega \in \Omega$. We have $d\mu_{\omega}d\omega=d\Theta(s)^{*}\mu_{\theta(-s)\omega} d\theta(-s)\omega$. If $\pi_{\Omega}\mu$ is invariant under the transformation $\{\theta(s),s \in R\}$, we have $\mu_{\omega}=\Theta(s)^{*}\mu_{\theta(-s)\omega}$ for almost every $\omega \in \Omega$. \end{lemma} \begin{proof} We assume that $\mu$ is invariant under $\{\Gamma(s),s \in R\}$.For any measurable function $f:TM \times T \times \Omega \to R$ \begin{equation}\nonumber \begin{aligned} &\int_{\Omega}\int_{G_{\omega}}f(x,v,\omega)d\mu_{\omega}d\omega=\int_{G}f(x,v,\omega)d\mu=\int_{G}f(x,v,\omega)d\Gamma^{*}(s)\mu\\ &=\int_{G}f(\Theta(s)|_{G_{\omega}}(x,v),\theta(s)\omega)d\mu=\int_{\Omega}\int_{G_{\omega}}f(\Theta(s)|_{G_{\omega}}(x,v),\theta(s)\omega)d\mu_{\omega}d\omega\\ &=\int_{\Omega}\int_{G_{\theta(-s)\omega}}f(\Theta(s)|_{G_{\theta(-s)(\omega)}}(x,v),\omega)d\mu_{\theta(-s)\omega} d\theta(-s)\omega\\ &=\int_{\Omega}\int_{G_{\theta(-s)\omega}}f(x,v,\omega)d\Theta(s)^{*}\mu_{\theta(-s)\omega} d\theta(-s)\omega\\ &=\int_{G}f(x,v,\omega)d\Theta(s)^{*}\mu_{\theta(-s)\omega} d\theta(-s)\omega\\ \end{aligned} \end{equation} Since $f$ is arbitrary, we have $d\mu_{\omega}d\omega=d\Theta(s)^{*}\mu_{\theta(-s)\omega} d\theta(-s)\omega$. If $\pi_{\Omega}\mu$ is invariant under $\{\theta(s),s\in R\}$, then we have $\mu_{\omega}=\Theta(s)^{*}\mu_{\theta(-s)\omega}$ for almost every $\omega \in \Omega$. \end{proof} \begin{theorem} When $G$ is a measurable set in $TM \times \Omega$. If $\mu$ is an ergodic invariant measure of $\{\Theta(s),s \in R\}$,$\pi_{\Omega}(\mu)$ is invariant under the transformation $\{\theta(s),s \in R\}$. Then we know that for almost $\omega \in \Omega$, there exists $(x_{\omega},v_{\omega})\in TM$, such that $\mu_{\omega}=\delta_{(x_{\omega},v_{\omega})}$. And when $s \in R$, we have $\Theta(s)(x_{\omega},v_{\omega})=(x_{\theta(s)\omega},v_{\theta(s)\omega})$ for almost all $\omega \in \Omega$. \end{theorem} \begin{proof} We assume that $\mu$ is an ergodic invariant measure under the transformation $\{\Theta(s),s \in R\}$. Since the set of invariant measure is a closed convex set. It is well known that the ergodic measure is the extreme point of the convex set, see \cite{katok1997introduction}. By Lemma 7.4, for almost all $\omega \in \Omega$, we know that $\mu_{\omega}$ is an extreme point of $P(G_{\omega})$, there exists $(x_{\omega},v_{\omega}) \in TM$, such that $\mu_{\omega}=\delta_{(x_{\omega},v_{\omega})}$. By Lemma 7.6, when $s \in R$, we know that $(x_{\theta(s)\omega},v_{\theta(s)\omega})=\Theta(s)(x_{\omega},v_{\omega})$ for almost every $\omega \in \Omega$. \end{proof}

https://arxiv.org/abs/0812.4518

The dihedral group $\Dh_5$ as group of symplectic automorphisms on K3 surfaces

We prove that if a K3 surface $X$ admits $\Z/5\Z$ as group of symplectic automorphisms, then it actually admits $\Dh_5$ as group of symplectic automorphisms. The orthogonal complement to the $\Dh_5$-invariants in the second cohomology group of $X$ is a rank 16 lattice, $L$. It is known that $L$ does not depend on $X$: we prove that it is isometric to a lattice recently described by R. L. Griess Jr. and C. H. Lam. We also give an elementary construction of $L$.

\section{Introduction} A finite group of symplectic automorphisms on a K3 surface $X$ has the property that the desingularization of the quotient of $X$ by this group is again a K3 surface. In \cite{Nikulin symplectic} the finite abelian groups of symplectic automorphisms on a K3 surface are classified. The main result of Nikulin in \cite{Nikulin symplectic} is that the isometries induced by finite abelian groups of symplectic automorphisms on the second cohomology group of a K3 surface are essentially unique. The uniqueness of the isometries induced by $G$ on $H^2(X,\mathbb{Z})$ implies that the lattice $\Omega_G:=(H^2(X,\mathbb{Z})^G)^{\perp}$ does not depend on $X$. Thanks to this result it is possible to associate the lattice $\Omega_G$ to each finite abelian group $G$ of symplectic automorphisms on a K3 surface. From this one obtains information on the coarse moduli space of K3 surfaces admitting $G$ as group of symplectic automorphisms (cf. \cite{Nikulin symplectic}, \cite{symplectic prime}, \cite{symplectic not prime}). In \cite{symplectic prime} and \cite{symplectic not prime} the lattices $\Omega_G$ are computed for each finite abelian group $G$ of symplectic automorphisms on a K3 surface.\\ In \cite{mukai} and \cite{Xiao} the finite (not necessary abelian) groups of symplectic automorphisms on a K3 surface are classified. Under some conditions (cf. Section \ref{section: symplectic automorphisms on K3 surfaces}) Nikulin's result, on the uniqueness of the isometries induced by finite groups of symplectic automorphisms on the second cohomology group of the K3 surfaces, can be extended to finite (not necessary abelian) groups (cf. \cite{whitcher}). As a consequence one can attach the lattice $\Omega_G:=(H^2(X,\mathbb{Z})^G)^{\perp}$, which depends only on $G$, also to some finite non abelian groups $G$ of symplectic automorphisms on a K3 surface $X$.\\ Let us consider a pair of finite groups $(G,H)$ such that $G$ acts symplectically on a K3 surface and $H$ is a subgroup of $G$. It is evident that the K3 surfaces admitting $G$ as group of symplectic automorphisms, admits also $H$ as group of symplectic automorphisms. It is more surprising that for certain pairs of groups $(G,H)$ also the viceversa holds indeed for certain pair $(G,H)$ the condition ``a K3 surface $X$ admits $G$ as group of symplectic automorphism" is equivalent to the condition ``$X$ admits $H$ as group of symplectic automorphisms". For these pairs the lattices $\Omega_G$ and $\Omega_H$ coincide. The aim of this paper is to describe this situation and to give explicitely one pair $(G,H)$. We observe that in order to find $(G,H)$ with the described property, one has to consider non abelian groups acting symplectically on K3 surfaces, indeed the lattices $\Omega_K$ associated to abelian groups $K$ are completely described in \cite{symplectic prime}, \cite{symplectic not prime} and one can check that $\Omega_G$ and $\Omega_H$ never coincide if both $G$ and $H$ are abelian and $G\neq H$.\\ In the {\it Section} \ref{section: symplectic automorphisms on K3 surfaces} we describe some known results on symplectic automorphisms over k\"ahlerian K3 surfaces and we prove our main results (Proposition \ref{prop: G as symp group iff H as symp group} and Corollary \ref{cor: Z/5 iff Dh5}). In the Proposition \ref{prop: G as symp group iff H as symp group} we give sufficient conditions on $G$ and $H$ to prove that a K3 surface admits $G$ as group of symplectic automorphisms if and only if it admits $H$ as group of symplectic automorphisms. Applying this proposition we prove (Corollary \ref{cor: Z/5 iff Dh5}) that a K3 surface admits $\mathbb{Z}/5\mathbb{Z}$ as group of symplectic automorphisms if and only if it admits $\Dh_5$ (the dihedral group of order 10) as group of symplectic automorphisms. In particular we prove that $\Omega_{\mathbb{Z}/5\mathbb{Z}}\simeq \Omega_{\Dh_5}$\\ In \cite{symplectic prime} the isometry induced on $\Omega_{\mathbb{Z}/5Z}$ by a symplectic automorphim of order 5, is described. Since we prove that $\Omega_{\mathbb{Z}/5\mathbb{Z}}\simeq\Omega_{\Dh_5}$, there is also an involution acting on this lattice. In order to describe both the isometry of order 5 and the involution generating $\Dh_5$ on $\Omega_{\Dh_5}$, in the {\it Section} \ref{section: construction of omega5} we give a different description of this lattice: it is an overlattice of $A_4(-2)^{\oplus 4}$ (the isometry of order 5 is induced by the natural one on $A_4$). In the proof of the Corollary \ref{cor: omegadh5 is isomteric to DIH10(16)} also the action of the involution is described. Moreover we show that the lattice $\Omega_{\mathbb{Z}/5\mathbb{Z}}$ (computed in \cite{symplectic prime}) is isometric to a lattice describe by Griess and Lam in \cite{EE8- lattices and dihedral}.\\ In the {\it Section} \ref{section: examples} we consider algebraic K3 surfaces and in particular 3-dimensional families of K3 surfaces admitting a symplectic automorphisms of order 5, $\sigma_5$, and a polarization, invariant under $\sigma_5$. It follows from the results of the Section \ref{section: symplectic automorphisms on K3 surfaces} that the K3 surfaces in these families admit also an involution $\iota$, generating together with $\sigma_5$ the dihedral group $\Dh_5$. For each of these families we exhibit the automorphism $\sigma_5$ and we find the automorphism $\iota$. \section{Symplectic automorphisms on K3 surfaces.}\label{section: symplectic automorphisms on K3 surfaces} \begin{definition} Let $X$ be a smooth compact complex surface. The surface $X$ is a K3 surface if the canonical bundle of $X$ is trivial and the irregularity of $X$, $q(X):=h^{1,0}(X)$, is 0. \end{definition} The second cohomology group of a K3 surface, equipped with the cup product, is isometric to a lattice, which is the unique, up to isometries, even unimodular lattice with signature $(3,19)$. This lattice will be denoted by $\Lambda_{K3}$ and is isometric to $U\oplus U\oplus U\oplus E_8(-1)\oplus E_8(-1)$, where $U$ is the unimodular lattice with bilinear form $\left[\begin{array}{ll}0&1\\1&0\end{array}\right]$ and $E_8(-1)$ is the lattice obtained multipying by $-1$ the lattice associated to the Dynking diagram $E_8$.\\ The N\'eron Severi group of a K3 surface $X$, $NS(X)$, coincide with its Picard group. The transcendental lattice of $X$, $T_X$, is the orthogonal to $NS(X)$ in $H^2(X,\mathbb{Z})$. \begin{definition} An isometry $\alpha$ of $H^2(X,\mathbb{Z})$ is an effective isometry if it preserves the K\"ahler cone of $X$. An isometry $\alpha$ of $H^2(X,\mathbb{Z})$ is an Hodge isometry if its $\C$-linear extension to $H^2(X,\C)$ preserves the Hodge decomposition of $H^2(X,\C)$. \end{definition} \begin{theorem}{\rm (\cite{Burns Rapoport torelli theorem K3})} Let $X$ be a K3 surface and $g$ be an automorphism of $X$, then $g^*$ is an effective Hodge isometry of $H^2(X,\mathbb{Z})$. Viceversa, let $f$ be an effective Hodge isometry of $H^2(X,\mathbb{Z})$, then $f$ is induced by an automorphism of $X$.\end{theorem} \begin{definition} An automorphism $\sigma$ on a K3 surface $X$ is symplectic if $\sigma^*$ acts as the identity on $H^{2,0}(X)$, that is $\sigma^*(\omega_X)=\omega_X$, where $\omega_X$ is a nowhere vanishing holomorphic two form on $X$.\\ Equivalently $\sigma$ is symplectic if the isometry induced by $\sigma^*$ on the transcendental lattice is the identity. \end{definition} \begin{remark}{\rm Let $\sigma$ be an automorphism of finite order on a K3 surface. The desingularization of $X/\sigma$ is a K3 surface if and only if $\sigma$ is symplectic.}\end{remark} In \cite{Nikulin symplectic} the finite abelian groups acting symplectically on a k\"ahlerian K3 surface are analyzed. Now it is known that every K3 surface is a K\"ahler variety, \cite{Siu}, so there are no restrictions on the K3 surfaces analyzed in \cite{Nikulin symplectic}.\\ From now on we always assume that $G$ is a finite group of symplectic automorphisms on the K3 surface $X$. \begin{definition} Let the K3 surface $Y$ be the minimal desingularization of the quotient $X/G$. Let $M_j$ be the curves arising form the desingularization of the singularities of $X/G$. \end{definition} The singularities of the quotient $X/G$ are computed by Nikulin (\cite[Sections 6]{Nikulin symplectic}) if $G$ is an abelian group, and by Xiao (\cite[Table 2 ]{Xiao}) for all the other finite groups. If either $G=Q_8$ (binary dihedral group of order 8) or $G=T_{24}$ (binary tetrahedral group of order 24), then there are two possible configurations for the singularities of $X/G$, and hence for the exceptional curve $M_j$ on $Y$. For all the other groups the number and the type of the singularities of $X/G$ are determined by $G$. \begin{definition} Let us assume that $G\neq Q_8$, $G\neq T_{24}$. The minimal primitive sublattice of $NS(Y)$ containing the curves $M_j$ does not depend on $X$ (cf. \cite{Nikulin symplectic}, \cite{Xiao}). It will be denoted by $M_G$.\end{definition} The lattice $M_G$ is computed by Nikulin (\cite[Section 7]{Nikulin symplectic}) for each abelian group $G$, and by Xiao (\cite[Table 2]{Xiao}) for the all the other gorups $G$. \begin{remark}\label{rem: Nikulin lattice} For $G=\mathbb{Z}/2\mathbb{Z}$, the lattice $M_{\mathbb{Z}/2\mathbb{Z}}$ (called Nikulin lattice) is an even overlattice of index 2 of $A_1(-1)^{\oplus 8}$. Its discriminant group is $(\mathbb{Z}/2\mathbb{Z})^6$ (cf. \cite{Nikulin symplectic}) and its discriminant form is the same as $U(2)^{\oplus 3}$ (cf. \cite{morrison}).\end{remark} \begin{definition}{\rm (\cite[Definition 4.6]{Nikulin symplectic})} We say that $G$ has a unique action on $\Lambda_{K3}$ if, given two embeddings $i: G\hookrightarrow Aut(X)$, $i': G\hookrightarrow Aut(X')$ such that $G$ is a group of symplectic automorphisms on the K3 surfaces $X$ and $X'$, there exists an isometry $\phi:H^2(X,\mathbb{Z})\rightarrow H^2(X',\mathbb{Z})$ such that $i'(g)^*=\phi\circ i(g)\circ \phi^{-1}$ for all $g\in G$.\end{definition} \begin{theorem}\label{theorem: unique action}{\rm (\cite[Theorem 4.7]{Nikulin symplectic}, \cite[Corollary 3.0.1]{whitcher})} Let $G$ be a finite group acting symplectically on a K3 surface, $G\neq Q_8$, $G\neq T_{24}$. If $M_G$ admits a unique primitive embedding in $\Lambda_{K3}$, then $G$ has a unique action on $\Lambda_{K3}$.\end{theorem} \begin{definition} Under the assumptions of the Theorem \ref{theorem: unique action} the lattice $(\Lambda_{K3}^G)^{\perp}$ is uniquely determined by $G$, up to isometry. It will be called $\Omega_G$. \end{definition} \begin{remark}\label{rem: rank omega rank m} By \cite[(8,12)]{Nikulin symplectic}, it follows that rank$(\Omega_G)=$rank$(M_G)$.\end{remark} \begin{theorem}\label{theorem: X admits G iff omegaG in NS(X)}{\rm (\cite[Theorem 4.15]{Nikulin symplectic}) } Let $G$ be a finite group acting symplectically on a K3 surface, such that $G$ has a unique action on $\Lambda_{K3}$. A K3 surface $X$ admits $G$ as group of symplectic automorphisms if and only if the lattice $\Omega_G$ is primitively embedded in $NS(X)$.\end{theorem} Nikulin proved that for each abelian group acting symplectically on a K3 surface, the hypothesis of the Theorem \ref{theorem: unique action} (and hence the ones of the Theorem \ref{theorem: X admits G iff omegaG in NS(X)}) are satisfied. \begin{proposition}\label{prop: G as symp group iff H as symp group} Let $G$ be a finite group acting symplectically on a K3 surface and let $H$ be a subgroup of $G$. Let us assume that both $G$ and $H$ are neither $Q_8$ or $T_{24}$. We assume that both $M_H$ and $M_G$ admit a unique primitive embedding in $\Lambda_{K3}$ and rank$(M_G)=$rank$ (M_H)$. Then $\Omega_H\simeq \Omega_G$ and so a K3 surface $X$ admits $G$ as group of symplectic automorphisms if and only if $X$ admits $H$ as group of symplectic automorphisms. \end{proposition} \proof Since $H$ is a subgroup of $G$, $\Omega_H$ is a sublattice of $\Omega_G$. Moreover rank$(\Omega_G)=$rank$(\Omega_H)$, by the Remark \ref{rem: rank omega rank m} and the condition on the rank of the lattices $M_G$ and $M_H$. This implies that $\Omega_H\hookrightarrow \Omega_G$ with a finite index. Let $X$ be a K3 admitting $G$ as group of symplectic automorphisms. Then both $\Omega_G$ and $\Omega_H(\hookrightarrow \Omega_G)$ are primitively embedded in $NS(X)$, hence the index of the inclusion $\Omega_H\hookrightarrow \Omega_G$ is 1, i.e. $\Omega_G\simeq \Omega_H$.\\ The K3 surface $X$ admits $G$ as group of symplectic automorphisms if and only if $\Omega_G$ is primitively embedded in $NS(X)$. By the isometry $\Omega_H\simeq\Omega_G$ this condition is equivalent to require that $\Omega_H$ is primitively embedded in $NS(X)$, which holds if and only if $X$ admits $H$ as group of symplectic automorphisms.\endproof \begin{corollary}\label{cor: Z/5 iff Dh5} A K3 surface admits $\mathbb{Z}/5\mathbb{Z}$ as group of symplectic automorphisms if and only if it admits $\Dh_5$ as group of symplectic automorphisms.\end{corollary} \proof The lattice $M_{\mathbb{Z}/5\mathbb{Z}}$ is computed in \cite{Nikulin symplectic}, where it is proved that it admits a unique primitive embedding in $\Lambda_{K3}$ and that its rank is $16$. The lattice $M_{\Dh_5}$ is described in \cite{Xiao} as an overlattice of index 2 of the lattice $A_4(-1)^{\oplus 2}\oplus A_1(-1)^{\oplus 8}$. In particular rank$(M_{\Dh_5})=16$ and $M_{\Dh_5}\simeq A_4(-1)^{\oplus 2}\oplus M_{\mathbb{Z}/2\mathbb{Z}}$, where $M_{\mathbb{Z}/2\mathbb{Z}}$ is the Nikulin lattice (see Remark \ref{rem: Nikulin lattice}). Thus the discriminant group of $M_{\Dh_5}$ is $(\mathbb{Z}/5\mathbb{Z})^2\oplus (\mathbb{Z}/2\mathbb{Z})^6$. By \cite[Theorem 1.14.4]{Nikulin bilinear}, $M_{\Dh_5}$ admits a unique primitive embedding in $\Lambda_{K3}$. The corollary immediately follows from the Proposition \ref{prop: G as symp group iff H as symp group}.\endproof \begin{remark}{\rm We proved that if a K3 surface admits a symplectic automorphism of order five $\sigma_5$, hence it admits also a symplectic involution generating $\Dh_5$ together with $\sigma_5$. This result can not be improved, i.e.\ it is {\bf not} true that if a K3 surface $X$ admits $\Dh_5=\langle \sigma_5,\iota\rangle$ as group of symplectic automorphisms, then it admits also a symplectic automorphism $\alpha$ such that $J:=\langle \alpha, \sigma_5, \iota\rangle\supset_{_{_{{\!\!\!\!\!\not{=}}}}} \Dh_5$ is a finite group. By contradiction assuming there exists such an $\alpha$, then $\Dh_5\subset_{_{_{{\!\!\!\!\!\not{=}}}}} J$ and $\Omega_{\Dh_5}\simeq \Omega_J$. In particular rank$\Omega_J=$rank$\Omega_{\Dh_5}=16$, but there are no finite groups $J$ of symplectic automorphisms on a K3 surface such that $\Dh_5\subset_{_{_{{\!\!\!\!\!\not{=}}}}} J$ and rank$M_J$(=rank$\Omega_{J})=16$ (cf. \cite[Table 2]{Xiao}). }\end{remark} \section{Construction of $\Omega_{\mathbb{Z}/5\mathbb{Z}}\simeq \Omega_{\Dh_5}$}\label{section: construction of omega5} The aim of this section is to construct the lattice $\Omega_{\mathbb{Z}/5\mathbb{Z}}$ as overlattice of $A_4(-2)^{\oplus 4}$ and to describe the action of $\Dh_5$ on this lattice. The automorphism of order five on $\Omega_{\mathbb{Z}/5\mathbb{Z}}$ will be induced by the automorphism of order five on each copy of $A_4(-2)$. We will add some rational linear combinations of the elements of $A_4(-2)^{\oplus 4}$ to obtain an even overlattice of $A_4(-2)^{\oplus 4}$. The main point is that we would like to extend the automorphism of $A_4(-2)^{\oplus 4}$ to the lattice $\Omega_{\mathbb{Z}/5\mathbb{Z}}$, so if we add an element to $A_4(-2)^{\oplus 4}$ we have to add all elements in its orbit. We recall that the standard basis of $A_4$ is expressed in terms of the standard basis $\{\varepsilon_i\}$ of $\mathbb{R}^5$ in the following way: $\alpha_i=\varepsilon_i-\varepsilon_{i+1}$, hence $\alpha_5=-\alpha_1-\alpha_2-\alpha_3-\alpha_4$. The cyclic permutation of the basis vectors of $\mathbb{R}^5$ induces the automorphism $\gamma$ on $A_4$ ($\gamma(\alpha_i)=\alpha_{i+1}$). \begin{proposition}\label{prop: definition of L and not -2 vectors} Let us consider the lattice $A_4(-2)^{\oplus 4}$ and the automorphism $g:=(\gamma,\gamma,\gamma,\gamma)$ (acting as $\gamma$ on each copy of $A_4(-2)$). Let $a_{j,i}$, $j,i=1,2,3,4$ be the element $\alpha_i$ in the $j$-copy of $A_4(-2)$. Let \begin{equation*}\mu:= \frac{1}{2}(a_{1,1}+a_{2,1}+a_{3,1}+a_{4,1}),\ \ \ \ \nu:=\frac{1}{2}(a_{2,1}+a_{3,3}+a_{3,4}+a_{4,1}+a_{4,3}+a_{4,4}).\end{equation*} Then: \begin{itemize} \item The lattice $$L:=A_4(-2)^{\oplus 4}+\langle g^i(\mu), g^i(\nu)\rangle_{i=0,1,2,3},$$ generated by $A_4(-2)^{\oplus 4}$ and the 8 vectors $g^i(\mu)$, $g^i(\nu)$, for $i=0,1,2,3$, is an even overlattice of $A_4(-2)^{\oplus 4}$ of rank 16. \item The index of $A_4(-2)^{\oplus 4}$ in $L$ is $2^8$. \item There are no vectors of length $-2$ in $L$.\end{itemize} \end{proposition} \begin{proof} Since $\mu$ has an integer intersection with the basis $\{a_{i,j}\}$ of $A_4(-2)^{\oplus 4}$ and has self intersection $-4$, we can add the element $\mu$ to the lattice $A_4(-2)^{\oplus 4}$ obtaining an even overlattice. All the elements $g^i(\mu)$ in the orbit of $\mu$ have an integer intersection with this basis of $A_4(-2)^{\oplus 4}$ and $g^i(\mu)g^i(\mu)\in \mathbb{Z}$ for all $i,j=0,1,2,3$. Thus we can add the four vectors $\mu$, $g(\mu)$, $g^2(\mu)$, $g^3(\mu)$ to the lattice $A_4(-2)^{\oplus 4}$.\\ It is easy to show that the vectors $g^i(\nu)$, $i=0,1,2,3$ have an integer intersection pairing with all the vectors in $A_4(-2)^{\oplus 4}$, and that $g^i(\nu)g^j(\nu)\in \mathbb{Z}$, $(g^i(\nu))^2\in 2\mathbb{Z}$ (indeed these are properties of all the vectors of type $v_{i,j,k,h,l,m}=\frac{1}{2}(0,\varepsilon_i-\varepsilon_{i+1}, \varepsilon_{j}-\varepsilon_{j+2}, \varepsilon_k-\varepsilon_h+\varepsilon_l-\varepsilon_m)$, $k<h<l<m$ and $\nu=v_{1,3,1,2,3,5}$).\\ Moreover $g^i(\nu)g^j(\mu)\in \mathbb{Z}$, $i,j\in\{0,1,2,3\}$ (indeed this is a property of the vectors of type $v_{i,j,k,h,l,m}$ such that $\{i,i+1\}\cap\{j,j+2\}=\emptyset$ and $\{k,h,l,m\}=\{i,i+1,j,j+2\}$).\\ Thus adding the vectors in the orbit of $\nu$ and of $\mu$ to $A_4(-2)^{\oplus 4}$ we construct an even overlattice $L$ of $A_4(-2)^{\oplus 4}$. By the computation of the discriminant of this lattice, it follows that $A_4(-2)^{\oplus 4}$ has index $2^8$ in $L$ (indeed to construct $L$ we add to $A_4(-2)^{\oplus 4}$ exactly eight vectors of type $\frac{1}{2}v$, $v\in A_4(-2)^{\oplus 4}$ and they are independent over $\mathbb{Z}$).\\ Now we prove that there are no vectors with length $-2$ in $L$. Let $y=\sum_{i=0}^3 b_ig^i(\mu)+\sum_{i=0}^3 c_ig^i(\nu)$, $b_i, c_i\in \mathbb{Z}$. In $A_4(-2)^{\oplus 4}\otimes \mathbb{Q}$ we have: $$\begin{array}{cl}y:=&\frac{1}{2}\left(\sum_{i=0}^3b_i\alpha_{i+1},\ \ \ \ \ \sum_{i=0}^3b_i\alpha_{i+1}+\sum_{i=0}^3c_i\alpha_{i+1},\right.\\ &\left.\sum_{i=0}^3b_i\alpha_{i+1}+(-c_1+c_3)\alpha_1+(-c_1-c_2+c_3)\alpha_2+(c_0-c_1-c_2)\alpha_3+(c_0-c_2)\alpha_4,\right.\\&\left.\sum_{i=0}^3b_i\alpha_{i+1}+(c_0-c_1+c_3)\alpha_1+(-c_2+c_3)\alpha_2+(c_0-c_1)\alpha_3+(c_0-c_2+c_3)\alpha_4\right).\end{array}$$ If we require that at least two components of $y$ are equal to zero, we obtain that $b_i=c_i=0$ for all $i$. So if $y\neq 0$, then $y$ has at most one component equal to zero.\\ Each vector $w$ in $L$ is of the form $y+z$ with $y$ as above, $b_i,c_i\in\{0,1\}$, and $z\in A_4(-2)^{\oplus 4}$, moreover such $y$ and $z$ are uniquely determined by $w$.\\ If $y=(0,0,0,0)$, $w\in A_4(-2)^{\oplus 4}$ and hence $w^2\leq -4$.\\ If $y\neq (0,0,0,0)$, then $w=\frac{1}{2}(w_1,w_2,w_3,w_4)$ with $w_i\in A_4(-2)$ and at most one of $w_i=0$. Since $w_i^2\leq -4$ and $w_i\cdot w_j=0$ if $i\neq j$, we get $w^2\leq \frac{3}{4}(-4)$. Hence there are no vectors of length $-2$ in $L$. \end{proof} \begin{proposition} The lattice $L$ is isometric to the lattice $\Omega_{\mathbb{Z}/5\mathbb{Z}}=\Omega_{\Dh_5}$.\end{proposition} \begin{proof} By uniqueness of $\Omega_{\mathbb{Z}/5\mathbb{Z}}$, to prove the proposition it suffices to show that there exists a K3 surface $S$ such that $G=\mathbb{Z}/5\mathbb{Z}$ is a group of symplectic automorphisms on $S$ and $(H^2(S,\mathbb{Z})^G)^{\perp}\simeq L$.\\ By construction $L$ admits an automorphism of order 5, $g$, acting trivially on the discriminant group. Moreover $L$ is negative definite and its discriminant group is $(\mathbb{Z}/5\mathbb{Z})^4$. Hence it admits a primitive embedding in $\Lambda_{K3}$ (\cite[Theorem 1.14.4]{Nikulin bilinear}). Since $g$ acts trivially on the discriminat group, $G:=\langle g\rangle$ extends to a group of isometries on $\Lambda_{K3}$ which acts as the identity on $L^{\perp_{\Lambda_{K3}}}$.\\ Let $S$ be a K3 surface such that $L\subset NS(S)$ (such a K3 surface exists by the surjectivity of the period map). By the Proposition \ref{prop: definition of L and not -2 vectors}, $L$ does not contain elements of length $-2$. This is enough to prove that the isometries of $G$ defined above (if necessary composed with a reflection in the Weil group) are effective isometries for $S$ (the proof of this fact is essentially given in \cite[Theorem 4.3]{Nikulin symplectic}, see also \cite[Step 4, proof of Proposition 5.2]{symplectic prime}). By construction, these are Hodge isometries (cf. \cite[Theorem 4.3]{Nikulin symplectic}), so they are induced by automorphisms on $S$ (by the Torelli theorem, cf. \cite{Burns Rapoport torelli theorem K3}). Since these automorphisms act as the identity on $T_S\subset L^{\perp_{\Lambda_{K3}}}$, they are symplectic. By construction of the isometries of $G$, $L\simeq (H^2(S,\mathbb{Z})^{G})^{\perp}$, and so $L\simeq \Omega_{\mathbb{Z}/5\mathbb{Z}}$. \end{proof} Since $\Omega_{\Dh_5} \simeq L$, on $L$ acts the dihedral group and in particular an involution. This implies that the lattice $\Omega_{\mathbb{Z}/2\mathbb{Z}}\simeq E_8(-2)$ (cf. \cite{morrison}) is a primitively embedded in $L$ and there exists an involution on $L$ acting as $-1$ on this lattice and as the identity on its orthogonal. In the following remark we give an embedding of $E_8(-2)$ in $L$ and in the proof of the Corollary \ref{cor: omegadh5 is isomteric to DIH10(16)} we describe the involution associated to this embedding. \begin{remark}\label{rem: E8 in L}{\rm The vectors $$\begin{array}{llll}e_1:=\mu&e_2:=g^2(\mu)+g^3(\mu)&e_3:=\nu& e_4:=\mu+g^2(\mu)+g^3(\mu)-g^2(\nu)-g^3(\nu)\\ e_5:=a_{1,1}&e_6:=a_{1,3}+a_{1,4}&e_7:=a_{2,1}&e_8:=a_{2,3}+a_{2,4}\end{array}$$ generate of a copy of $E_8(-2)$ embedded in $L$. Indeed the lattice generated by $e_i$ is such that multiplying its bilinear form by $\frac{1}{2}$ one obtains a negative definite even unimodular lattice of rank 8, i.e. a copy of $E_8(-1)$. }\end{remark} \begin{remark}\label{rem: 2 E8 in L} Let $f_{i+8}=g(e_i)$, $i=1,\ldots, 8$ where $\{e_i\}_{i=1,\ldots,8}$ is the basis of $E_8(-2)$ defined in the Remark \ref{rem: E8 in L}. Since $g$ is an isometry of the lattice it is clear that $f_i$, $i=9,\ldots 16$ generate a copy of $E_8(-2)$ embedded in $L$. A direct computation shows that the classes $e_i$, $f_{i+8}$ $i=1,\ldots 8$ generates a lattice of rank 16 and discriminant $5^4$ embedded in $L$ and so they are a $\mathbb{Z}$-basis for $L$.\end{remark} The paper \cite{EE8- lattices and dihedral} classifies positive definite lattices which have dihedral groups $\Dh_n$ (for $n=2,3,4,5,6$) in the group of the isometries and which have the properties: \begin{itemize}\item the lattices are rootless (i.e. there are no elements of length $2$), \item they are the sum of two copies of $E_8(2)$, \item there are two involutions in $\Dh_n$ acting as minus the identity on each copy of $E_8(2)$.\end{itemize} In \cite[Section 7]{EE8- lattices and dihedral} it is proved that there is a unique lattice with all these properties and admitting $\Dh_5$ in the group of isometries, called $DIH_{10}(16)$. \begin{corollary}\label{cor: omegadh5 is isomteric to DIH10(16)} The lattice $DIH_{10}(16)$ described in \cite{EE8- lattices and dihedral} is isometric to the lattice $L(-1)\simeq\Omega_{\Dh_5}(-1)$.\end{corollary} \proof The even lattice $L$ has no vectors of length $-2$ (Proposition \ref{prop: definition of L and not -2 vectors}).\\ On $L$ there is an isometry of order 5, $g$ (Proposition \ref{prop: definition of L and not -2 vectors}).\\ Let us define a map $h$ on the lattice $L$ which acts as $-1$ on the copy of $E_8(-2)$ generated by $e_i$, $i=1,\ldots, 8$ (cf. Remark \ref{rem: E8 in L}) and as the idenitity on the orthogonal complement. Since $h$ acts trivially on the discriminant group of $E_8(-2)\simeq\langle e_i\rangle_{i=1,\ldots, 8}$, $h$ is an isometry of $L$ and in particular an involution. One can directly check that its action on the basis of $A_4(-2)^{\oplus 4}$ is $h(a_{i,1})=-a_{i,1}$, $h(a_{i,2})=-a_{i,5}$, $h(a_{i,3})=-a_{i,4}$, $h(a_{i,4})=-a_{i,3}$, $i=1,2,3,4$. This action extends to a $\mathbb{Z}$-basis of $L$.\\ The group $\langle g,h\rangle $ is $\Dh_5$. The involutions $h$ and $g^2\circ h$ are two involutions generating the group $\Dh_5$. By construction $h$ and $g^2\circ h$ act as minus the identity respectively on the lattice $E_8(-2)\simeq\langle e_i\rangle_{i=1,\ldots 8}$ and on the lattice generated by $E_8(-2)\simeq\langle f_i\rangle_{i=9,\ldots 16}$. These two copies of $E_8(-2)$ generate $L$ (by Remark \ref{rem: 2 E8 in L}). So $L(-1)$ satisfies the conditions which define $DIH_{10}(16)$ and hence $DIH_{10}(16)\simeq L(-1)\simeq \Omega_{\Dh_5}(-1)$. \section{Examples: algebraic K3 surfaces with a polarization of a low degree.}\label{section: examples} Here we give some very explicit examples of families of K3 surfaces admitting $\mathbb{Z}/5\mathbb{Z}$, and hence $\Dh_5$, as group of symplectic automorphisms. In particular in this section we consider algebraic K3 surfaces.\\ We recall that a polarization $L$, with $L^2=2d$, on a K3 surface $X$ defines a map $\phi_L:X\rightarrow \mathbb{P}^{d+1}$. In this section we consider K3 surfaces with a polarization $L$ such that $\phi_L(X)$ is a complete intersection in a certain projective space and K3 surfaces with a polarization of degree 2, which exhibits the K3 surfaces as double covers of the plane.\\ Let $X$ be a general member of a family of K3 surfaces admitting an automorphism of order 5, $\sigma_5$, and a polarization, $L$, invariant under $\sigma_5$. In \cite[Proposition 5.1]{symplectic prime} the possible N\'eron--Severi groups of $X$ are computed. In particular if $L^2=2d<10$, one obtains that $NS(X)\simeq \mathbb{Z} L\oplus \Omega_{\mathbb{Z}/5\mathbb{Z}}=:\mathcal{L}_{2d}$ . This lattice admits a unique primitive embedding in $\Lambda_{K3}$. The family of K3 surfaces with a polarization $L$, of degree $L^2< 10$, invariant under a symplectic automorphism of order 5, is then the family of the $\mathcal{L}_{2d}$-polarized K3 surfaces. In particular for each $d<5$, we find a 3-dimensional family of K3 with such a polarization $L$ and hence we have the following possibilities: \begin{itemize} \item $\phi_L:X\stackrel{2:1}{\rightarrow} \mathbb{P}^2$, so $X$ is a double cover of $\mathbb{P}^2$ branched along a plane sextic curve: in this case $NS(X)\simeq \mathcal{L}_{2};$\item $\phi_L(X)$ is a quartic in $\mathbb{P}^3$: in this case $NS(X)\simeq\mathcal{L}_{4}$; \item $\phi_L(X)$ is the complete intersection of a quadric and a cubic in $\mathbb{P}^4$: in this case $NS(X)\simeq\mathcal{L}_{6}$; \item $\phi_L(X)$ is the complete intersection of three quadrics in $\mathbb{P}^5$: in this case $NS(X)\simeq\mathcal{L}_{8}$. \end{itemize} Now we construct a general member of each of these families and show that it also admits a symplectic involution $\iota$ generating, together with $\sigma_5$, the group $\Dh_5$. Since the automorphisms $\iota$ and $\sigma_5$ leave invariant the polarization, both these automorphisms can be extended to automorphisms of the ambient projective space.\\ We will denote by $\omega$ a primitive $5$-th root of unity. \subsection{$L^2=4$} This polarization gives a map to $\mathbb{P}^3$ where the K3 surfaces are realized as quartic surfaces. Let us consider the automorphism $$\sigma_{\mathbb{P}^3}:(x_0:x_1:x_2:x_3)\rightarrow (x_0:\omega^3 x_1:\omega x_2:\omega^2 x_3).$$ The quartic surfaces in $\mathbb{P}^3$ defined as \begin{equation}\label{formula: equation 5 in P3}V(ax_0^3x_2+bx_0^2x_1^2+cx_0x_3^3+dx_0x_1x_2x_3+ex_1^3x_3+fx_1x_2^3+gx_2^2x_3^2),\end{equation} are invariant for $\sigma_{\mathbb{P}^3}$. Hence the restriction of $\sigma_{\mathbb{P}^3}$ to K3 surfaces with equations \eqref{formula: equation 5 in P3} is an automorphism $\sigma_5$ of the surfaces. To show that this automorphism is symplectic it suffices to apply $\sigma_5$ to the following holomorphic two form in local coordinates $x=x_1/x_0$, $y=x_2/x_0$, $z=x_3/x_0$: $$\left(\frac{\partial f}{\partial z}\right)^{-1}dx\wedge dy,$$ where $f$ denotes the equation of the quartic in the local coordinate $x,y,z$.\\ The equation \eqref{formula: equation 5 in P3} depends on 7 parameters. The automorphisms of $\mathbb{P}^3$ commuting with $\sigma_{\mathbb{P}^3}$ are $diag(\alpha,\beta,\gamma,\delta)$ (which is a four dimensional group), hence this family of $\sigma_{\mathbb{P}^3}$-invariant quartics has $$(7-1)-(4-1)=3$$ moduli. So the family of K3 surfaces given by the equation \eqref{formula: equation 5 in P3} is the family of K3 surfaces admitting an automorphisms of order 5 leaving invariant a polarization of degree 4.\\ Up to a projectivity, commuting with $\sigma_{\mathbb{P}^3}$, the equation \eqref{formula: equation 5 in P3} becomes \begin{equation}\label{formula: second equation 5 in P3}a'x_0^3x_2+b'x_0^2x_1^2+c'x_0x_3^3+d'x_0x_1x_2x_3+a'x_1^3x_3+c'x_1x_2^3+g'x_2^2x_3^2=0.\end{equation} Let us define an involution of $\mathbb{P}^3$: $$\iota_{\mathbb{P}^3}:(x_0:x_1:x_2:x_3)\rightarrow (x_1:x_0:x_3:x_2).$$ The equation \eqref{formula: second equation 5 in P3} is invariant under $\iota_{\mathbb{P}^3}$, hence $\iota_{\mathbb{P}^3}$ induces an automorphism of the quartic surfaces with equation \eqref{formula: second equation 5 in P3}, we call this automorphism $\iota$. The fixed point set of $\iota_{\mathbb{P}^3}$ in $\mathbb{P}^3$ is the union of the lines: $$l_1=\left\{\begin{array}{l} x_0=x_1\\ x_2=x_3\end{array}\right.,\ \ \ l_2=\left\{\begin{array}{l} x_0=-x_1\\ x_2=-x_3\end{array}\right.$$ Hence $\iota$ fixes eight points on the quartics \eqref{formula: second equation 5 in P3}: the intersection of the quartics with $l_1$ and $l_2$. This is enough to show that $\iota$ is a symplectic involution, indeed the involutions which are not symplectic either are fixed point free or fix some curves \cite[Theorem 1]{Zhang}. Hence the quartics given in \eqref{formula: second equation 5 in P3} (and hence, up to a projectivity, in \eqref{formula: equation 5 in P3}) admit both the automorphisms $\sigma_5$ and $\iota$. It is easy to check that $\langle \sigma_{\mathbb{P}^3},\iota_{\mathbb{P}^3}\rangle =\Dh_5$, and hence $\langle \sigma_5,\iota\rangle=\Dh_5$. So the family of smooth quartic surfaces in $\mathbb{P}^3$ admitting the symplectic automorphism $\sigma_5$ admits also a symplectic involution $\iota$ and in fact the group $\Dh_5=\langle \sigma_5,\iota\rangle$. \subsection{$L^2=6$} This polarization gives a map to $\mathbb{P}^4$ where the K3 surfaces are realized as complete intersections of a cubic and a quadric.\\ Let us consider the automorphism $$\sigma_{\mathbb{P}^4}:(x_0:x_1:x_2:x_3:x_4)\rightarrow (x_0:\omega x_1:\omega^2 x_2:\omega^3 x_3:\omega^4 x_4).$$ Let: \begin{eqnarray}\label{formula: equation 5 in P4 quadric}\begin{array}{l} Q:=V(ax_0^2+bx_1x_4+cx_2x_3);\\ C:=V(dx_0^3+ex_0x_1x_4+fx_0x_2x_3+gx_1^2x_3+hx_2x_4^2+lx_1x_2^2+mx_3^2x_4),\end{array}\end{eqnarray} then $Q$ and $C$ are $\sigma_{\mathbb{P}^4}$-invariant hypersurfaces in $\mathbb{P}^4$. We observe that the complete intersection of these two hypersurfaces is generically smooth and thus it is a K3 surface.\\ The complete intersection of $Q$ and $C$ is also the complete intersection of $Q$ and $C+ \lambda x_0Q$. Hence there is a 1-dimensional family of invariant cubics giving the same complete intersection: the cubics giving different complete intersections depend on $7-1=6$ parameters. The automorphisms of $\mathbb{P}^4$ which commute with $\sigma_{\mathbb{P}^4}$ are of the form $diag(\alpha, \beta, \gamma, \delta, \varepsilon)$. So the family of complete intersections of a cubic and a quadric invariant under the automorphism $\sigma_{\mathbb{P}^4}$ has $(3-1)+(6-1)-(5-1)=3$ moduli.\\ Let $X$ be the complete intersection of $Q$ and $C$. The automorphism $\sigma_{\mathbb{P}^4}$ induces a symplectic automorphism on $X$ (this can be shown as in case $L^2=4$ considering the two holomorphic form, in local coordinates $x,y,z,t$, $(dx\wedge dy)/(Q_zC_t-C_tQ_z)$ where $F_x$ is the partial derivative of $F$ w.r.t.\ $x$).\\ Up to the action of the projectivities commuting with $\sigma_{\mathbb{P}^4}$, we can assume that $g=h$ and $l=m$ in the equation of $C$, in \eqref{formula: equation 5 in P4 quadric}. Hence the involution $$\iota_{\mathbb{P}^4}:(x_0:x_1:x_2:x_3:x_4)\rightarrow (x_0:x_4:x_3:x_2:x_1)$$ fixes $Q$ and $C$. So its restriction to $X$ is an involution, $\iota$, of $X$. Moreover $\iota$ has eight fixed points (six on the plane $x_1=x_4, x_2=x_3$ and two on the line $x_0=0$, $x_1+x_4=0$, $x_2+x_3=0$). Thus $\sigma_5$ and $\iota$ are symplectic automorphisms of the K3 surface $X$ and they generate the group $\Dh_5$. \subsection{$L^2=8$} This polarization gives a map to $\mathbb{P}^5$ where the K3 surfaces are realized as complete intersections three quadrics.\\ Let us consider the map $$\sigma_{\mathbb{P}^5}:(x_0:x_1:x_2:x_3:x_4:x_5)\rightarrow (x_0:x_1:\omega x_2:\omega^2 x_3:\omega^3 x_4:\omega^4 x_5)$$ and the complete intersection of the quadrics \begin{eqnarray*}\label{formula: equation 5 in P5 quadric} \begin{array}{lll} Q_1':=&V(ax_0^2+bx_0x_1+cx_1^2+dx_2x_5+ex_3x_4),\\ Q_2':=&V(fx_1x_2+gx_3x_5+hx_4^2),\\ Q_3':=&V(lx_1x_5+mx_2x_4+nx_3^2). \end{array}\end{eqnarray*} The group of automorphisms of $\mathbb{P}^5$ commuting with $\sigma_{\mathbb{P}^5}$ is $(GL(2)\times GL(1)^4)/GL(1)$ which has dimension $7=8-1$. So these complete intersections in $\mathbb{P}^5$ have $(5-1)+(4-1)+(4-1)-(8-1)=3$ moduli. Up to automorphisms of $\mathbb{P}^5$ commuting with $\sigma_{\mathbb{P}^5}$ we can assume that the quadrics have the following equation: \begin{eqnarray*} \begin{array}{lll} Q_1:=&V(x_0^2+bx_0x_1+x_1^2+dx_2x_5+ex_3x_4),\\ Q_2:=&V(x_1x_2+x_3x_5+x_4^2),\\ Q_3:=&V(x_1x_5+x_2x_4+x_3^2), \end{array}\end{eqnarray*} and in fact they depend on 3 parameters.\\ The complete intersection $X$ of the quadrics $Q_1$, $Q_2$, $Q_3$ is smooth for a generic choice of the parameters $b,d,e$ (one can check it directly putting $e=b=0$, $d=1$). Moreover $X$ is invariant under the automorphism $\sigma_{\mathbb{P}^5}$, so $\sigma_{\mathbb{P}^5}$ induces an automorphism $\sigma_5$ on $X$ and $\sigma_5$ is symplectic (this can be shown as in the case $L^2=6$).\\ The involution $$\iota_{\mathbb{P}^5}:(x_0:x_1:x_2:x_3:x_4:x_5)\rightarrow (x_0:x_1:x_5:x_4:x_3:x_2)$$ fixes the quadric $Q_1$ and switches the quadrics $Q_2$ and $Q_3$. So its restriction to the K3 surface $X$ is an involution, $\iota$, of the surface $X$. Moreover $\iota$ has eight fixed points, on the space $x_2=x_5$, $x_3=x_4$. Thus $\sigma_5$ and $\iota$ are symplectic automorphisms of the K3 surface $X$ and they generate the group $\Dh_5$. \subsection{$L^2=2$} This polarization gives a $2:1$ map to $\mathbb{P}^2$ and the K3 surfaces are realized as double cover of $\mathbb{P}^2$ branched along a sextic plane curve.\\ The map $$\sigma_{\mathbb{P}^2}:(x_0:x_1:x_2)\rightarrow (x_0:\omega x_1:\omega^4 x_2)$$ is an automorphism of $\mathbb{P}^2$. Up to projectivity of $\mathbb{P}^2$ commuting with $\sigma_{\mathbb{P}^2}$, the invariant sextic for $\sigma_{\mathbb{P}^2}$ are \begin{eqnarray*}\mathcal{C}_6:= V(x_0^6+x_0x_1^5+x_0x_2^5+ax_0^4x_1^2x_2^2+bx_0^3x_1^3x_2^3+cx_1^3x_2^3) \end{eqnarray*} Let $X$ be the double cover of $\mathbb{P}^2$ branched along $\mathcal{C}_6$, i.e. $X$ is $V(u^2-(x_0^6+x_0x_1^5+x_0x_2^5+ax_0^4x_1^2x_2^2+bx_0^3x_1^3x_2^3+cx_1^3x_2^3))$ in the weighted projective space $\mathbb{PW}(3,1,1,1)$. The automorphism $\sigma_{\mathbb{P}^2}$ lifts to a symplectic automorphism $\sigma_5:(u:x_0:x_1:x_2)\rightarrow (u:x_0:\omega x_1:\omega^4 x_2)$ of $X$. So we constructed the 3-dimensional family of K3 surfaces which are double covers of $\mathbb{P}^2$ and have a symplectic automorphism of order 5 which leavs invarinat the polarization.\\ The involution $$\alpha_{\mathbb{P}^2}:(x_0:x_1:x_2)\rightarrow (x_0:x_2:x_1)$$ leaves the curve $\mathcal{C}_6$ invariant, so it lifts to an involution $\alpha_X:(u:x_0:x_1:x_2)\rightarrow (u:x_0:x_2:x_1)$ of the surface $X$. The involution $\alpha_X$ fixes a curve (the pull back of the line $x_1=x_2$ in $\mathbb{P}^2$), so it is not symplectic. Let $i:(u:x_0:x_1:x_2)\rightarrow (-u:x_0:x_1:x_2)$ be the covering involution on $X$. It is a non symplectic involution (indeed the quotient $X/\iota$ is rational) and it commutes both with $\alpha_X$ and $\sigma_5$. The involution $\iota=\alpha_X\circ i$ is a symplectic involution on $X$ (because it is the composition of two commuting non symplectic involutions). Moreover one has $\iota\circ \sigma_5=\sigma_5^{-1}\circ\iota$, and hence $\Dh_5=\langle \sigma_5, \iota\rangle$ acts symplectically on $X$. \bibliographystyle{amsplain}

https://arxiv.org/abs/0707.3337

On The Capacity of Surfaces in Manifolds with Nonnegative Scalar Curvature

Given a surface in an asymptotically flat 3-manifold with nonnegative scalar curvature, we derive an upper bound for the capacity of the surface in terms of the area of the surface and the Willmore functional of the surface. The capacity of a surface is defined to be the energy of the harmonic function which equals 0 on the surface and goes to 1 at infinity. Even in the special case of Euclidean space, this is a new estimate. More generally, equality holds precisely for a spherically symmetric sphere in a spatial Schwarzschild 3-manifold. As applications, we obtain inequalities relating the capacity of the surface to the Hawking mass of the surface and the total mass of the asymptotically flat manifold.

\section{Introduction} The research in this paper was partly motivated by the following theorem. \vspace{.3cm} \noindent {\bf Theorem} (\cite{Bray_Penrose}) {\em Let $(M^3, g)$ be a complete, asymptotically flat $3$-manifold with boundary with nonnegative scalar curvature. Suppose its boundary $\partial M$ consists of horizons (that is $\partial M$ has zero mean curvature). Let $G(x)$ be a function on $M^3$ which satisfies \begin{equation} \nonumber \left\{ \begin{array}{rcl} \lim_{x \rightarrow \infty} G & = & 1 \\ \triangle G & = & 0 \\ G|_{\partial M} & = & 0 . \end{array} \right. \end{equation} Then \begin{equation} m \geq C, \end{equation} where $m$ is the total mass of $(M^3, g)$ and $C$ is the constant in the asymptotic expansion \begin{equation} \nonumber G(x) = 1 - \frac{C}{|x|} + O \left( \frac{1}{|x|^2} \right) \ \ \mathrm{at} \ \infty. \end{equation} Furthermore, equality holds if and only if the manifold $(M^3, g)$ is isometric to a spatial Schwarzschild manifold outside its horizon. } \vspace{.3cm} This theorem was established to prove the Riemannian Penrose Inequality in \cite{Bray_Penrose}. It was later applied in \cite{Miao_hbdry} for a generalization of Bunting and Masood's rigidity theorem \cite{Bunting_Masood} on static vacuum spacetime with black hole boundary. Considering these applications, it is of interest to know whether a simiar theorem holds on an asymptotically flat $3$-manifold with a boundary that does not necessarily have zero mean curvature. A corollary to our main result, Theorem \ref{mainthm} stated in a moment, is the following: \begin{cor} \label{maincor} Let $(M^3, g)$ be a complete, asymptotically flat $3$-manifold with nonnegative scalar curvature with a connected smooth boundary. Assume $(M^3, g)$ is diffeomorphic to $\mathbb{R}^3 \setminus \Omega$, where $\Omega$ is a bounded domain. Let $\mathcal{G}(x)$ be a function on $(M^3, g)$ which satisfies \begin{equation} \nonumber \left\{ \begin{array}{rcl} \lim_{x \rightarrow \infty} \mathcal{G} & = & 1 \\ \triangle \mathcal{G} & = & 0 \\ \mathcal{G}|_{\partial M} & = & \sqrt{\frac{1}{16 \pi}\int_{\partial M} H^2 d\mu } , \end{array} \right. \end{equation} where $H$ is the mean curvature of $\partial M$ and $d \mu$ is the induced surface measure. If $\partial M$ has nonnegative Hawking mass (that is $\int_{\partial M} H^2 d \mu \leq 16 \pi$), then \begin{equation} m \geq \mathcal{C}, \end{equation} where $m$ is the total mass of $(M^3, g)$ and $\mathcal{C}$ is the constant in the asymptotic expansion \begin{equation} \nonumber \mathcal{G}(x) = 1 - \frac{\mathcal{C}}{|x|} + O \left( \frac{1}{|x|^2} \right) \ \ \mathrm{at} \ \infty. \end{equation} Furthermore, equality holds if and only if the manifold $(M^3, g)$ is isometric to a spatial Schwarzschild manifold $$ (M_{r_0}, g^S_m) = \left([r_0, \infty) \times S^2, \frac{1}{1 - \frac{2m}{r}} dr^2 + d\sigma^2 \right), $$ where $r_0$ is some positive constant satisfying $r_0 \geq 2m$ and $d\sigma^2$ is the standard metric on the unit spere $S^2 \subset \mathbb{R}^3$. \end{cor} Before we state our main result, we first define the capacity of a surface. \begin{df} Let $(M^3, g)$ be a complete, asymptotically flat $3$-manifold with a nonempty boundary $\Sigma$. The {\bf capacity} of $\Sigma$ in $(M^3, g)$, denoted by $C_M(\Sigma, g)$, is defined to be \begin{equation} C_M(\Sigma, g) = \inf \left\{ \frac{1}{4 \pi} \int_{M^3} | \nabla \phi |^2 dg \right\} , \end{equation} where the infimum is taken over all locally Lipschitz $\phi(x)$ which go to $1$ at $\infty$ and equal $0$ on $\Sigma$. \end{df} When $(M^3, g)$ is the complement of a smooth bounded domain $\Omega$ in the Euclidean space $(\mathbb{R}^3, g_0)$, $C_{M}(\partial \Omega, g_0)$ is simply the usual electrostatic capacity of $\partial \Omega$ \cite{Polya_Szego}. In this case, we write $C_{M}(\partial \Omega, g_0)$ as $C(\partial \Omega)$. Our main theorem is the following: \begin{thm} \label{mainthm} Let $(M^3, g)$ be a complete, asymptotically flat $3$-manifold with nonnegative scalar curvature with a connected smooth boundary. Assume $(M^3, g)$ is diffeomorphic to $\mathbb{R}^3 \setminus \Omega$, where $\Omega$ is a bounded domain. Then \begin{equation} \label{estofc} C_M(\partial M, g) \leq \sqrt{\frac{|\partial M|}{16 \pi}} \left( 1 + \sqrt{\frac{1}{16 \pi} \int_{\partial M} H^2 d\mu} \right), \end{equation} where $| \partial M |$ and $H$ are the area and the mean curvature of $\partial M$. Furthermore, equality holds if and only $(M^3, g)$ is isometric to a spatial Schwarzschild manifold $$ (M_{r_0}, g^S_{m}) = \left([r_0, \infty) \times S^2, \frac{1}{1 - \frac{2m}{r}} dr^2 + d\sigma^2 \right), $$ where $r_0$ is some positive constant satisfying $r_0 \geq 2 m$ and $d\sigma^2$ is the standard metric on the unit sphere $S^2 \subset \mathbb{R}^3$. \end{thm} As an immediate corollary, we have a new estimate of the capacity of a surface in $(\mathbb{R}^3, g_0)$. \begin{cor} Let $\Omega \subset (\mathbb{R}^3, g_0)$ be a bounded domain with a connected smooth boundary. Then \begin{equation} \label{flatest} C(\partial \Omega) \leq \sqrt{\frac{|\partial \Omega|}{16 \pi}} \left( 1 + \sqrt{\frac{1}{16 \pi} \int_{\partial \Omega} H^2 d\mu} \right). \end{equation} Furthermore, equality holds if and only if $\Omega$ is a round ball. \end{cor} We note that it is interesting to compare (\ref{flatest}) with the classical isoperimetric inequality in $\mathbb{R}^3$: \begin{equation} \label{classiciso} \left( \frac{3 V}{4 \pi} \right)^\frac{1}{3} \leq \sqrt{ \frac{ | \partial \Omega |}{4 \pi}} = \sqrt{\frac{|\partial \Omega|}{16 \pi}} \left( 1 + 1 \right) , \end{equation} where $V$ is the volume of $\Omega \subset \mathbb{R}^3$. It is known, among all domains $\Omega$ with a fixed amount volume $V > 0$, $C(\partial \Omega)$ is minimized by a round ball \cite{Polya_Szego}, i.e. \begin{equation} \left( \frac{3 V}{4 \pi} \right)^\frac{1}{3} \leq C(\partial \Omega ) . \end{equation} On the other hand, the Willmore functional $\int_{\partial \Omega} H^2 d\mu$ satisfies $$ \int_{\partial \Omega} H^2 d\mu \geq 16 \pi .$$ Hence, (\ref{flatest}) is analogous to the classical isoperimetric inequality (\ref{classiciso}), but where both sides of the inequality are increased. We outline the idea of the proofs of Corollary \ref{maincor} and Theorem \ref{mainthm}. For both results, we apply the technique of weak inverse mean curvature flow as developed by Huisken and Ilmanen in \cite{IMF}. The topological assumptions on $M^3$ ensures that the Hawking mass of the flowing surface $\Sigma_t$ is monotone nondecreasing for positive $t$. A key step in the proof of Theorem \ref{mainthm} is to use the flow to construct a special test function that gives the estimate (\ref{estofc}). Such a construction was first used by Bray and Neves in \cite{Bray_Neves}. It is a convenient feature of Corollary \ref{maincor} and Theorem \ref{mainthm} that, though they are proved by applying the weak inverse mean curvature flow technique, they hold {\em without} assuming the boundary surface is {\em outer minimizing} (see \cite{Bray_Penrose} \cite{IMF}). This paper is organized as follows. In Section \ref{literature}, we recall some classical results on capacity of convex surfaces in $\mathbb{R}^3$ from the work of P\'olya and Szeg\"o in \cite{Polya_Szego} to illustrate that the weak inverse mean curvature flow technique fits naturally with the classical method. In Section \ref{ucap}, we apply the theory of weak inverse mean curvature flow to prove a general theorem on $C_M(\Sigma, g)$, which includes Thereom \ref{mainthm} as a special case. In Section \ref{application}, we relate our estimate on $C_M(\Sigma,g)$ to estimates of the Hawking mass and the total mass, and prove Corollary \ref{maincor}. In Section \ref{discussion}, we give an application of Corollary \ref{maincor} to the study of static metrics in general relativity. \section{Capacity of convex surfaces in $\mathbb{R}^3$} \label{literature} We first give an account of some classical methods and results from $\cite{Polya_Szego}$ in estimating $C(\Sigma)$ for convex surfaces $\Sigma$ in $\mathbb{R}^3$. Let $\Sigma$ be a closed, connected $C^2$ surface bounding some domain $\Omega$ in $\mathbb{R}^3$. One basic idea in estimating $C(\Sigma)$ is to minimize $\int_{\mathbb{R}^3 \setminus \Omega} | \nabla v |^2 dg_0$ over functions $v$ which have {\em given level surfaces} $\{ \Sigma_t \}$. These level surfaces form a one-parameter family. Therefore, after the selection of $\{\Sigma_t\}$, $v$ becomes a function of one variable and the infimum of $\int_{\mathbb{R}^3 \setminus \Omega} | \nabla v|^2 d g_0$ over all such $v$ can be easily evalutated. Precisely, we fix a function $\psi$, defined on $\mathbb{R}^3 \setminus \Omega$, which satisfies $$ \psi \geq 0, \ \ \ \psi|_\Sigma = 0 , \ \ \ \lim_{x \rightarrow \infty} \psi = \infty .$$ Let $ \{ \Sigma_t = \psi^{-1}(t) \ | \ 0 \leq t < \infty \} $ be the family of level surfaces of $\psi$. For any other function $v $ having the same level surfaces $\{ \Sigma_t \}$, $v$ must have the form $ v(x) = f( \psi(x) )$ for some single variable function $f(t)$, which satisfies $f(0) = 0$ and $f(\infty) = 1$. By the co-area formula, we have \begin{eqnarray} \int_{\mathbb{R}^3 \setminus \Omega} | \nabla v |^2 dg_0 & = & \int_0^\infty \left( \int_{ \Sigma_t } f^\prime(t)^2 | \nabla \psi | d \mu \right) dt \nonumber \\ & = & \int_0^\infty f^\prime(t)^2 \left( \int_{ \Sigma_t } | \nabla \psi | d \mu \right) dt . \end{eqnarray} Define \label{defTt} \begin{equation} T(t) = \frac{1}{4 \pi} \int_{ \Sigma_t } | \nabla \psi | d \mu, \end{equation} which is determined solely by the level surfaces $\{\Sigma_t\}$. Then \begin{equation} C(\Sigma) \leq \int_0^\infty f^\prime(t)^2 T(t) dt. \end{equation} Applying the fundamental theorem of calculus and the H\"older inequality, \begin{eqnarray} 1 & = & \left( \int_0^\infty f^\prime(t) dt \right)^2 = \left( \int^\infty_0 f^\prime(t) T(t)^\frac{1}{2} T(t)^{-\frac{1}{2}} dt \right)^2 \nonumber \\ & \leq & \left( \int^\infty_0 f^\prime(t)^2 T(t) dt \right) \left( \int_0^\infty T(t)^{-1} dt \right) . \end{eqnarray} Thus \begin{equation} \left( \int_0^\infty T(t)^{-1} dt \right)^{-1} \leq \int_0^\infty f^\prime(t)^2 T(t) dt \end{equation} for all $f(t)$ with equality if and only if \begin{equation} f(t) = \Lambda \int_0^t \frac{1}{T(s)} ds, \end{equation} where $\Lambda = \left( \int_0^\infty T(t)^{-1} dt \right)^{-1}$. Choosing such a $f(t)$, we show that \begin{equation} C(\Sigma) \leq \left( \int_0^\infty T(t)^{-1} dt \right)^{-1} . \end{equation} Now suppose $\Sigma$ is a {\em convex} surface in $\mathbb{R}^3$. A natural choice for $\{ \Sigma_t \}$ is the family of level surfaces of the distance function to $\Sigma$. We let $\psi(x) = dist(x, \Sigma)$ and define \begin{equation} \Sigma_t = \{ x \ | \ dist(x, \Sigma) = t \}. \end{equation} Then $ | \nabla \psi | = 1$ everywhere and $ T(t) = \frac{ | \Sigma_t | }{4 \pi} $, where $| \Sigma_t |$ is the area of $\Sigma_t$, given by \begin{equation} | \Sigma_t | = | \Sigma | + \left( \int_{\Sigma} H d \mu \right) t + 4 \pi t^2 . \end{equation} This leads to the following theorem of Szeg\"o. \vspace{.3cm} \noindent {\bf Theorem} (\cite{Polya_Szego}) {\em If $\Sigma$ is a convex surface in $(\mathbb{R}^3, g_0)$, then \begin{equation} C(\Sigma) \leq \frac{M}{ 4 \pi} \frac{2 \epsilon}{\log{ \frac{1 + \epsilon}{1 - \epsilon}}}, \end{equation} where \begin{equation} M = \frac{1}{2} \int_\Sigma H d \mu \ \mathrm{and} \ \epsilon^2 = 1 - \frac{4 \pi |\Sigma|}{ M^2}. \end{equation} } \vspace{.3cm} \section{Capacity of surfaces in asymptotically flat manifolds} \label{ucap} Obviously, most of the calculations in Section 2 work on any non-compact Riemannian manifold. The key step is to make a good choice of $\{ \Sigma_t \}$ so that the corresponding $T(t)$ can be efficiently estimated. In this section, we consider {\em asymptotically flat $3$-manifolds}, on which the theory of weak inverse mean curvature flow developed by Huisken and Ilmanen \cite{IMF} gives a nearly canonical foliation. \begin{df} A Riemannian $3$-manifold $(M^3, g)$ is said to be {\bf asymptotically flat} if there is a compact set $K \subset M$ such that $M \setminus K$ is diffeomorphic to $\mathbb{R}^3$ minus a compact set and in the coordinate chart defined by this diffeomorphism, $$ g = \sum_{i,j} g_{ij}(x) dx^i dx^j, $$ where $$ g_{ij} = \delta_{ij} + O(|x|^{-1}), \ \partial_k g_{ij} = O(|x|^{-2}) \ \ \partial_l \partial_{k} g_{ij} = O(|x|^{-3}) .$$ \end{df} \vspace{.3cm} \noindent {\bf Theorem} (\cite{IMF}) {\em Let $(M^3, g)$ be a complete, connected asymptotically flat $3$-manifold with a $C^{1,1}$ boundary $\Sigma$. There exists a proper, locally Lipschitz function $\phi \geq 0$ on $M$, called the solution to the weak inverse mean curvature flow with initial condtion $\Sigma$, which satisfies the following properties: \begin{enumerate} \item $\phi|_\Sigma = 0$, $\lim_{x \rightarrow \infty} \phi = \infty$. For $t > 0$, $\Sigma_t = \partial \{ \phi \geq t\} $ and $\Sigma^\prime_t = \partial \{ \phi > t \}$ define an increasing family of $C^{1, \alpha}$ surfaces. \item The surfaces $\Sigma_t$ ($\Sigma^\prime_t$) minimize (strictly minimize) area among surfaces homologous to $\Sigma_t$ in the region $\{ \phi \geq t \}$. The surface $ \Sigma^\prime = \partial \{ \phi >0 \}$ strictly minimizes area among surfaces homologous to $\Sigma$ in $M$. \item For almost all $t > 0$, the weak mean curvature of $\Sigma_t$ is defined and equals $| \nabla \phi |$, which is positive for almost all $ x \in \Sigma_t $. \item For each $t > 0$, $|\Sigma_t | = e^t | \Sigma^\prime| $, and $| \Sigma_t | = e^t | \Sigma |$ if $\Sigma$ is outer minimizing (that is $\Sigma$ minimizes area among all surfaces homolgous to $\Sigma$ in $M$). \item If $(M^3, g)$ has nonnegative scalar curvature and $\chi(\Sigma_t) \leq 2$ for all $t > 0$, the Hawking mass $$ m_H(\Sigma_t) = \sqrt{ \frac{| \Sigma_t|}{16 \pi} } \left( 1 - \frac{1}{16 \pi} \int_{\Sigma_t} H^2 d \mu \right) $$ is monotone nondecreasing for $t > 0$ and $lim_{t \rightarrow 0+} m_H(\Sigma_t) \geq m_H(\Sigma^\prime)$. Here $\chi(S)$ is the Euler characteristic of a surface $S$. \end{enumerate} } \vspace{.3cm} We note that when $\phi$ is a smooth function with non-vanishing gradient, property $3$ is just saying that the level surfaces $\{ \Sigma_t \}$ move at a speed equal to the inverse of their mean curvature. We are now ready to prove the main result of this section. \begin{thm} \label{gthm} Let $(M^3, g)$ be a complete, connected asymptotically flat $3$-manifold which has a $C^2$ boundary $\Sigma$. Suppose $(M^3, g)$ has nonnegative scalar curvature and the solution to the weak inverse mean curvature flow with initial condition $\Sigma$ satisfies $\chi(\Sigma_t) \leq 2$ for all $t > 0$. Then \begin{equation} \label{estofcL} C_M(\Sigma, g) \leq \sqrt{\frac{|\Sigma|}{16 \pi}} \left( 1 + \sqrt{\frac{1}{16 \pi} \int_\Sigma H^2 d\mu} \right), \end{equation} where $| \Sigma |$ and $H$ denote the area and the mean curvature of $\Sigma$. Furthermore, equality holds if and only $(M^3, g)$ is isometric to a spatial Schwarzschild manifold \begin{equation} \label{schwarzschild} (M_{r_0}, g^S_{m}) = \left([r_0, \infty) \times S^2, \frac{1}{1 - \frac{2m}{r}} dr^2 + d\sigma^2 \right), \end{equation} where $r_0$ is some positive constant satisfying $r_0 \geq 2 m$ and $d\sigma^2$ is the standard metric on the unit sphere $S^2 \subset \mathbb{R}^3$. \end{thm} \noindent {\em{Proof}}: We estimate $\int_M | \nabla v |^2 dg$ for functions $v$ that have the same level surfaces $\{ \Sigma_t \}$ as the function $\phi$, which is the solution to the weak inverse mean curvature flow starting at $\Sigma$. It follows from the calculation in Section 2 that \begin{equation} \label{fT} C_M(\Sigma, g) \leq \inf_f \left\{ \int_0^\infty f^\prime(t)^2 T(t) dt \right\} , \end{equation} where the infimum is taken over all $f(t)$ satisfying $f(0)=0$ and $f(\infty) = 1$ and \begin{equation} T(t) = \frac{1}{ 4 \pi} \int_{\Sigma_t} | \nabla \phi | d\mu . \end{equation} Now, for a.e. $t>0$, \begin{equation} \int_{\Sigma_t} | \nabla \phi | d\mu = \int_{\Sigma_t} H d \mu, \end{equation} where $H$ is the weak mean curvature of $\Sigma_t$. To proceed, we make use of the key property that $m_H(\Sigma_t)$ is monotone nondecreasing for $t>0$ and $m_H(\Sigma^\prime) \leq \lim_{s \rightarrow 0+} m_H(\Sigma_s) $, which are guaranteed by the assumption that $(M^3, g)$ has nonnegative scalar curvature and $\chi(\Sigma_t) \leq 2$. Hence, for each $t > 0$, \begin{equation} m_H(\Sigma^\prime) \leq m_H(\Sigma_t) = \sqrt{\frac{|\Sigma_t|}{16 \pi}} \left( 1 - \frac{1}{16 \pi} \int_{\Sigma_t} H^2 d\mu \right) . \end{equation} This, together with the H\"older inequality, implies \begin{equation} \frac{1}{16 \pi |\Sigma_t| } \left( \int_{\Sigma_t} H d\mu \right)^2 \leq \frac{1}{16 \pi} \int_{\Sigma_t} H^2 d\mu \leq 1 - m_H(\Sigma^\prime) \sqrt{ \frac{16 \pi}{|\Sigma_t|} }. \end{equation} Hence, \begin{eqnarray} \label{b} 4 \pi T(t) = \int_{\Sigma_t} H d\mu & \leq & \left[ 16 \pi | \Sigma_t | \left( 1 - m_H(\Sigma^\prime) \sqrt{ \frac{16 \pi}{|\Sigma_t|} } \right) \right]^{\frac{1}{2}} \nonumber \\ & = & \left[ 16 \pi | \Sigma^\prime | e^t \left( 1 - m_H(\Sigma^\prime) \sqrt{ \frac{16 \pi}{|\Sigma^\prime|} } e^{-\frac{t}{2}} \right) \right]^{\frac{1}{2}} , \end{eqnarray} by the fact $| \Sigma_t | = e^t |\Sigma^\prime|$. Now write $\bar{m}_0 = m_H(\Sigma^\prime) $ and $ \bar{A}_0 = | \Sigma^\prime | $, it follows from (\ref{fT}) and (\ref{b}) that \begin{equation} \label{cc} 4 \pi C_M(\Sigma, g) \leq \inf_f \left\{ \int_0^\infty f^\prime(t)^2 F( \bar{A}_0, \bar{m}_0, t) dt \right\}, \end{equation} where $F(\bar{A}_0, \bar{m}_0, t)$ is an explicit function of $\bar{A}_0$, $\bar{m}_0$ and $t$, given by \begin{equation} F(\bar{A}_0, \bar{m}_0, t) = \left[ 16 \pi \bar{A}_0 e^t \left( 1 - \bar{m}_0 \sqrt{ \frac{16 \pi}{\bar{A}_0} } e^{-\frac{t}{2}} \right) \right]^{\frac{1}{2}}. \end{equation} To calculate $$ \inf_f \left\{ \int_0^\infty f^\prime(t)^2 F( \bar{A}_0, \bar{m}_0, t) dt \right\}, $$ we consider the $3$-dimensional spatial Schwarzschild metric $$ g^S_{\bar{m}_0} = \frac{1}{1 - \frac{2 \bar{m}_0}{r}} dr^2 + r^2 d\sigma^2 . $$ When $\bar{m}_0 < 0$, $g^S_{\bar{m}_0}$ is defined on $(0, \infty) \times S^2$ (the metric has a singularity at $r=0$). When $\bar{m}_0 \geq 0$, $g^S_{\bar{m}_0}$ is defined on $[2 \bar{m}_0, \infty) \times S^2$. In either case, $g^S_{\bar{m}_0}$ is well defined on $[\bar{r}_0, \infty) \times S^2$, where $\bar{r}_0$ satisfies $\bar{A}_0 = 4 \pi \bar{r}_0^2$. For convenience, we let $M_{\bar{r}_0}$ denote $[\bar{r}_0, \infty) \times S^2$, then the spatial Schwarzschild manifold $(M_{\bar{r}_0}, g^S_{\bar{m}_0})$ has a boundary $\{r = \bar{r}_0\}$ whose area is $\bar{A}_0$. A basic fact about $(M_{\bar{r}_0}, g^S_{\bar{m}_0})$ is that the classic inverse mean curvature flow in $(M_{\bar{r}_0}, g^S_{\bar{m}_0})$ with initial condition $\{ r = \bar{r}_0 \}$ is given by the family of coordinate spheres \begin{equation} S_t = \{ r = \bar{r}_0 e^{\frac{1}{2}t} \} , \end{equation} which have constant mean curvature (depending on $t$) and constant Hawking mass $\bar{m}_0$. Therefore, the corresponding function $\phi(x)$ on $(M_{\bar{r}_0}, g^S_{\bar{m}_0})$ is given by \begin{equation} \phi(x) = 2 \log{ \left( \frac{r}{\bar{r}_0} \right) } . \end{equation} Next, we consider the harmonic function $u$ on $(M_{\bar{r}_0}, g^S_{\bar{m}_0})$ that equals $0$ at $\{r = \bar{r}_0\}$ and goes to $1$ at $\infty$. We have two cases: \vspace{.2cm} \noindent {\bf Case 1} $\bar{m}_0 \neq 0$: \vspace{.2cm} \noindent In this case, the function \begin{equation} v = \sqrt{1 - \frac{2 \bar{m}_0 }{r} } = 1 - \frac{\bar{m}_0}{r} + O \left(\frac{1}{r^2}\right) \end{equation} is a non-constant harmonic function on $(M_{\bar{r}_0}, g^S_{\bar{m}_0})$. Hence, \begin{equation} u = \frac{v - v(\bar{r}_0)}{ 1 - v(\bar{r}_0) }. \end{equation} The capacity of the boundary of $(M_{\bar{r}_0}, g^S_{\bar{m}_0})$ is \begin{equation} \label{modelcap} C_{M_{\bar{r}_0}}(\partial M_{\bar{r}_0}, g^S_{\bar{m}_0}) = \frac{1}{4\pi} \int_{M_{\bar{r}_0}} | \nabla {u} |^2 d g^S_{\bar{m}_0} = \frac{\bar{m}_0}{1 - v(\bar{r}_0)}. \end{equation} Note that $u$ can be rewritten as \begin{equation} \label{fphi} u = f_0 \circ \phi , \end{equation} where \begin{equation} \label{fa} f_0(t) = \frac{1}{1 - v(\bar{r}_0) } \left[ \sqrt{ 1 - \frac{2 \bar{m}_0 }{ \bar{r}_0 e^{\frac{t}{2}}}} - v(\bar{r}_0) \right] . \end{equation} It then follows from (\ref{modelcap}), (\ref{fphi}) and the fact that $S_t$ has constant mean curvature and $m_H(S_t) = \bar{m}_0$ for all $t$ that $f_0$ achieves $$ \inf_f \left\{ \int_0^\infty f^\prime(t)^2 F( \bar{A}_0, \bar{m}_0, t) dt \right\} $$ and the infimum is given by \begin{eqnarray} \int_{M_{\bar{r}_0}} | \nabla u |^2 & = & 4 \pi \frac{\bar{m}_0}{1 - v(\bar{r}_0)} \nonumber \\ & = & 4 \pi \sqrt{\frac{|\Sigma^\prime|}{16 \pi}} \left( 1 + \sqrt{\frac{1}{16 \pi} \int_{\Sigma^\prime} H^2 d\mu} \right) . \end{eqnarray} Going back to (\ref{cc}), we have \begin{equation} \label{aa} C_M( \Sigma, g) \leq \sqrt{\frac{|\Sigma^\prime|}{16 \pi}} \left( 1 + \sqrt{\frac{1}{16 \pi} \int_{\Sigma^\prime} H^2 d\mu} \right) . \end{equation} \vspace{.2cm} \noindent {\bf Case 2} $\bar{m}_0 = 0$: \vspace{.2cm} \noindent In this case, our model space $(M_{\bar{r}_0}, g^S_{\bar{m}_0})$ is the Euclidean space $(\mathbb{R}^3, g_0)$ minus a round ball of radius $\bar{r}_0$ centered at the origin. Hence, \begin{equation} u = 1 - \frac{\bar{r}_0}{r}. \end{equation} The capacity of the boundary of $(M_{\bar{r}_0}, g^S_{\bar{m}_0})$ is \begin{equation} C(\partial M_{\bar{r}_0}) = \bar{r}_0 . \end{equation} Defining \begin{equation} \label{fb} f_0 (t) = 1 - e^{- \frac{t}{2}} , \end{equation} we can rewrite $u$ as \begin{equation} u = f_0 \circ \phi . \end{equation} The same argument as in the Case 1 then implies that \begin{equation} \label{bb} C_M( \Sigma, g) \leq \bar{r}_0 = \sqrt{\frac{|\Sigma^\prime|}{4 \pi}} = \sqrt{\frac{|\Sigma^\prime|}{ 16 \pi}} \left( 1 + \sqrt{\frac{1}{16 \pi} \int_{\Sigma^\prime} H^2 d\mu} \right) , \end{equation} where the last equality holds because $\bar{m}_0 = 0 $. \vspace{.2cm} Therefore, in both cases, we have proved that \begin{equation} \label{dd} C_M( \Sigma, g) \leq \sqrt{\frac{|\Sigma^\prime|}{ 16 \pi}} \left( 1 + \sqrt{\frac{1}{16 \pi} \int_{\Sigma^\prime} H^2 d\mu} \right) . \end{equation} To replace $\Sigma^\prime$ by $\Sigma$, we use the property that $\Sigma^\prime$ strictly minimizes area among all surfaces homologous to $\Sigma$. Since $\Sigma$ is $C^2$, $\Sigma^\prime$ is $C^{1,1}$ and $C^\infty$ where $\Sigma^\prime$ does not contact $\Sigma$. Moreoever, the mean curvature $H^\prime$ of $\Sigma^\prime$ satisfies \begin{equation} H^\prime = 0 \ \mathrm{on} \ \Sigma^\prime \setminus \Sigma \ \ \ \mathrm{and} \ \ \ H^\prime = H \geq 0 \ \ \mathcal{H}^2 a.e. \ \mathrm{on} \ \Sigma^\prime \cap \Sigma . \end{equation} In particular, we have \begin{equation} \label{sandsp} |\Sigma^\prime| \leq |\Sigma| \ \ \mathrm{and} \ \ \int_{\Sigma^\prime} {H^\prime}^2 d \mu \leq \int_\Sigma H^2 d\mu . \end{equation} Therefore, it follows from (\ref{dd}) and (\ref{sandsp}) that \begin{equation} \label{kk} C_M(\Sigma, g) \leq \sqrt{\frac{|\Sigma |}{16 \pi}} \left( 1 + \sqrt{\frac{1}{16 \pi} \int_{\Sigma} H^2 d\mu} \right) . \end{equation} \vspace{.2cm} To complete the proof of Theorem \ref{gthm}, we must consider the case of equality. Suppose \begin{equation} \label{equality} C_M(\Sigma, g) = \sqrt{\frac{|\Sigma|}{16 \pi}} \left( 1 + \sqrt{\frac{1}{16 \pi} \int_\Sigma H^2 d\mu} \right) , \end{equation} it follows from the above proof that \begin{equation} \label{outm} |\Sigma^\prime| = |\Sigma|, \ \int_{\Sigma^\prime} {H^\prime}^2 d \mu = \int_\Sigma H^2 d\mu , \end{equation} \begin{equation} \label{hconstant} m_H(\Sigma_t) = m_H(\Sigma^\prime), \ \forall \ t>0 , \end{equation} and \begin{equation} \label{capvalue} C_M(\Sigma, g) = \frac{1}{4\pi} \int_M | \nabla ( f_0 \circ \phi) |^2 dg, \end{equation} where $f_0$ is either given by (\ref{fa}) or (\ref{fb}) and $\phi$ is the solution to the weak inverse mean curvature flow in $(M^3, g)$ with initial condition $\Sigma$. It follows from (\ref{outm}) and (\ref{hconstant}) that $\Sigma$ is outer minimizing and the Hawking mass $m_H(\Sigma_t)$ equals $m_H(\Sigma)$ for every $t$. On the other hand, (\ref{capvalue}) implies that \begin{equation} u_M = f_0 \circ \phi \end{equation} is a smooth harmonic function on $M$. As a result, the surfaces $\Sigma_t$, $\Sigma$ do not ``jump" to $\Sigma_t^\prime$, $\Sigma^\prime$ (as defined in \cite{IMF}, meaning $\Sigma_t = \Sigma_t^\prime, \ \Sigma = \Sigma^\prime$), for otherwise the set $\{ x \in M^3 \ | \ u_M(x) = f_0(t) \} $ would have non-empty interior for some $t \geq 0$, contradicting the maximum principle for harmonic functions. Furthermore, applying the maximum principle to the exterior region of $\Sigma_t$ in $(M^3, g)$ and using the fact that $u_M$ is constant on $\Sigma_t$ and $\Sigma_t$ is at lease $C^1$, we conclude that $\nabla u_M$ never vanishes. Therefore, \begin{equation} \phi = f^{-1}_0 \circ u_M \end{equation} is a smooth function on $M$ with non-vanishing gradient. Hence, $\{\Sigma_t\}$ evolve smoothly at a speed equal to the inverse of their mean curvature. The fact $m_H(\Sigma_t) = m_H(\Sigma)$ for all $t>0$ then readily implies that $(M^3, g)$ is isometric to a spatial Schwarzschild manifold \begin{equation} (M_{r_0}, g^S_{m}) = \left([r_0, \infty) \times S^2, \frac{1}{1 - \frac{2m}{r}} dr^2 + d\sigma^2 \right) \end{equation} with $r_0 \geq 2m$ (see page 423 in \cite{IMF} for detail). Theorem \ref{gthm} is proved. \hfill$\Box$ \vspace{.3cm} Next, we give a topological condition of $M^3$ that is sufficient to guarantee the assumption $\chi(\Sigma_t) \leq 2$ in Theorem \ref{gthm}. \begin{prop} \label{topp} Let $(M^3, g)$ be a complete, connected asymptotically flat $3$-manifold with a nonempty boundary $\Sigma$. If $H_2(M,\Sigma) = 0$ and $\Sigma$ is connected, then $\{\Sigma_t\}_{t>0}$ remains connected. In particular, $\chi(\Sigma_t) \leq 2$ for all $t$. \end{prop} \noindent {\em{Proof}}: Under the assumption that $\Sigma$ is connected, Huisken and Ilmanen proved that the sets $\{ \phi < t \}$ and $\{ \phi > t \}$ are connected \cite{IMF}. They also showed, for each $t >0$, $\Sigma_t$ can be approximated in $C^1$ by earlier surfaces $\Sigma_s$, satisfying $\Sigma_s = \Sigma_s^\prime$. Hence, we may assume $\Sigma_t = \{ \phi = t \}$. Let $\Sigma_1$ be one component of $\Sigma_t$. Since $H_2(M,\Sigma) =0$ and $\Sigma$ is connected, there is a bounded region $D$ in $M$ such that either $\partial D = \Sigma_1$ or $\partial D = \Sigma_1 \cup \Sigma$. As the set $\{ \phi > t \}$ is connected and contains $\infty$, we have $\phi \leq t$ on $D$. It then follows that $\phi < t$ on $D \setminus \Sigma_1$. Hence $D \setminus \Sigma_1$ is a component of the set $\{ \phi < t \}$. Since $\{ \phi < t \}$ is connected, we must have $D \setminus \Sigma_1 = \{ \phi < t \}$. Therefore, $\Sigma_1$ is the only component of $\Sigma_t$ (and $\partial D = \Sigma_1 \cup \Sigma$). We conclude that $\Sigma_t$ is connected. \hfill$\Box$ \vspace{.3cm} Theorem \ref{mainthm} now follows directly from Theorem \ref{gthm}, Proposition \ref{topp} and the fact $H_2(\mathbb{R}^3 \setminus \Omega, \partial \Omega) = 0$ . \section{Estimate of the total mass} \label{application} We prove Corollary \ref{maincor} in this section. First, we point out that Theorem \ref{gthm} translates directly into a statement about the capacity of $\Sigma$ and the Hawking mass of $\Sigma$. \begin{thm} \label{Hmass} Let $(M^3, g)$ be a complete, connected asymptotically flat $3$-manifold with a nonempty boundary $\Sigma$. Suppose $(M^3, g)$ satisfies all the assumptions in Theorem \ref{gthm}. Then \begin{equation} \label{Hmasscap} | m_H(\Sigma) | \geq | 1 - \alpha | C_M(\Sigma, g) , \end{equation} where $\alpha$ is a constant defined by \begin{equation} \alpha = \sqrt{ \frac{1}{16\pi} \int_\Sigma H^2 d \mu }. \end{equation} Furthermore, in the case $ \alpha \neq 1$, equality holds if and only if $(M^3, g)$ is isometric to a spatial Schwarzschild manifold \begin{equation} \label{isometry} (M_{r_0}, g^S_{m}) = \left([r_0, \infty) \times S^2, \frac{1}{1 - \frac{2m}{r}} dr^2 + d\sigma^2 \right), \end{equation} where $r_0$ is some positive constant satisfying $r_0 \geq 2m$ and $d\sigma^2$ is the standard metric on the unit sphere $S^2 \subset \mathbb{R}^3$. \end{thm} \noindent {\em{Proof}}: It follows directly from Theorem \ref{gthm} and the definition of the Hawking mass. \hfill$\Box$ \vspace{.3cm} Next, we recall the definition of the total mass of an asymptotically flat $3$-manifold (see \cite{Bray_Penrose}, \cite{IMF}). \begin{df} Let $(M^3, g)$ be an asymptotically flat $3$-manifold. The total mass of $(M^3, g)$ is defined as the limit \begin{equation} m = \frac{1}{16 \pi} \lim_{r \rightarrow \infty} \int_{S_r} \sum_{ij} ( \partial_i g_{ij} - \partial_j g_{ii} ) \nu^j d \sigma , \end{equation} where $S_r$ is the coordinate sphere $\{ |x| = r\}$, $\nu$ is the coordinate unit normal to $S_r$ and $ d \sigma$ is the area element of $S_r$ in the coordinate chart. \end{df} An important feature of the weak inverse mean curvature flow, proved by Huisken and Ilmanen in \cite{IMF}, is \begin{prop} \label{IMFlimit} Let $(M^3, g)$ be an asymptotically flat $3$-manifold and $\phi$ be a solution to the weak inverse mean curvature flow, then \begin{equation} m \geq \lim_{t \rightarrow \infty} m_H(\Sigma_t) , \end{equation} where $m$ is the total mass of $(M^3, g)$ and $\Sigma_t = \partial \{ \phi \geq t\}$. \end{prop} The next theorem shows that the total mass $m$ is bounded from below by the same quantity $( 1 - \alpha) C_M(\Sigma, g)$ as in (\ref{Hmasscap}). A convenient feature of the theorem is that it does not require $\Sigma$ to be outer minimizing. \begin{thm} \label{totalmass} Let $(M^3, g)$ be a complete, connected asymptotically flat $3$-manifold with a nonempty boundary $\Sigma$. Suppose $(M^3, g)$ satisfies all the assumptions in Theorem \ref{gthm} and $\Sigma$ has nonnegative Hawking mass. Let $\mathcal{G}(x)$ be a function on $(M^3, g)$ which satisfies \begin{equation} \nonumber \left\{ \begin{array}{rcl} \lim_{x \rightarrow \infty} \mathcal{G} & = & 1 \\ \triangle \mathcal{G} & = & 0 \\ \mathcal{G}|_{\Sigma} & = & \alpha , \end{array} \right. \end{equation} where \begin{equation} \alpha = \sqrt{\frac{1}{16 \pi}\int_\Sigma H^2 d\mu } . \end{equation} Then \begin{equation} m \geq \mathcal{C}, \end{equation} where $m$ is the total mass of $(M^3, g)$ and $\mathcal{C}$ is the constant in the asymptotic expansion \begin{equation} \nonumber \mathcal{G}(x) = 1 - \frac{\mathcal{C}}{|x|} + O \left( \frac{1}{|x|^2} \right) \ \ \mathrm{at} \ \infty. \end{equation} Furthermore, the equlity holds if and only if the manifold $(M^3, g)$ is isometric to a spatial Schwarzschild manifold $$ (M_{r_0}, g^S_m) = \left([r_0, \infty) \times S^2, \frac{1}{1 - \frac{2m}{r}} dr^2 + d\sigma^2 \right), $$ where $r_0$ is some positive constant satisfying $r_0 \geq 2m$ and $d\sigma^2$ is the standard metric on the unit spere $S^2 \subset \mathbb{R}^3$. \end{thm} \noindent {\em{Proof}}: We only need to consider the case $m_H(\Sigma)>0$ (that is $\alpha < 1$), as the case $m_H(\Sigma)=0$ is essentially the proof of the Positive Mass Theorem via the inverse mean curvature flow \cite{IMF}. We use the same notations as in the proof of Theorem \ref{gthm}. Applying Proposition \ref{IMFlimit} and using the monotonicity of the Hawking mass, we have \begin{equation} \label{mandsp} m \geq \lim_{t \rightarrow \infty} m_H(\Sigma_t) \geq \lim_{t \rightarrow 0+} m_H(\Sigma_t) \geq m_H(\Sigma^\prime) . \end{equation} The proof of Theorem \ref{gthm} shows \begin{equation} \label{YY} C_M(\Sigma^\prime, g) \leq \sqrt{ \frac{ | \Sigma^\prime|}{16 \pi} } \left( 1+ \sqrt{ \frac{1}{16 \pi} \int_{\Sigma^\prime} H^2 d \mu } \right) . \end{equation} Let $ \alpha^\prime = \sqrt{ \frac{1}{16\pi} \int_{\Sigma^\prime} H^2 d \mu }$, by (\ref{sandsp}) we have \begin{equation} \label{alal} \alpha^\prime \leq \alpha . \end{equation} Thus, $\alpha^\prime < 1$. Hence, (\ref{YY}) is equivalent to \begin{equation} \label{mm} m_H (\Sigma^\prime) \geq ( 1 - \alpha^\prime) C_M(\Sigma^\prime, g) . \end{equation} On the other hand, it follows from the maximum principle that \begin{equation} \label{capcap} \ C_M(\Sigma^\prime, g) \geq C_M(\Sigma, g) \end{equation} with equality if and only if $\Sigma = \Sigma^\prime$. Therefore, it follows from (\ref{mandsp}), (\ref{mm}), (\ref{alal}) and (\ref{capcap}) that \begin{equation} m \geq ( 1 - \alpha ) C_M(\Sigma, g) \end{equation} with equality if and only if $\Sigma = \Sigma^\prime$ and \begin{equation} \label {ZZ} C_M(\Sigma, g) = \sqrt{ \frac{ | \Sigma |}{16 \pi} } \left( 1+ \sqrt{ \frac{1}{16 \pi} \int_{\Sigma} H^2 d \mu } \right) . \end{equation} By Theorem \ref{gthm}, (\ref{ZZ}) holds if and only if $(M^3, g)$ is isometric to a spatial Schwarzschild manifold $(M_{r_0}, g^S_m)$. As $ \mathcal{C} = ( 1 - \alpha ) C_M(\Sigma, g)$, Theorem \ref{totalmass} is proved. \hfill$\Box$ \vspace{.3cm} Corollary \ref{maincor} follows directly from Theorem \ref{totalmass}, Proposition \ref{topp} and the fact $H_2(\mathbb{R}^3 \setminus \Omega, \partial \Omega) = 0 $. \section{Application to static metrics} \label{discussion} In this section, we give a simple application of Corollary \ref{maincor} to the study of static metrics in general relativity. We recall that a $3$-dimensional asymptotically flat manifold $(M^3, g)$ is called {\bf static} \cite{Miao_static} if there is a positive function $N$, called the static potential function of $(M^3, g)$, satisfying $N \rightarrow 1$ at $\infty$ and \begin{equation} \label{staticeq} \left\{ \begin{array}{ccl} N Ric(g) & = & D^2 N \\ \triangle N & = & 0 , \end{array} \right. \end{equation} where $D^2 N$ is the Hessian of $N$ and $Ric(g)$ is the Ricci curvature of $g$. It can be easily checked that $(M^3, g)$ and $N$ satisfy (\ref{staticeq}) if and only if the asymptotically flat spacetime metric $\bar{g} = - N^2 dt^2 + g $ solves the Vaccum Einstein Equation on $ M \times \mathbb{R}$. In particular, (\ref{staticeq}) implies that $(M^3, g)$ has zero scalar curvature. A fundamental result in the study of static, asymptotically flat manifolds with boundary is the following black hole uniqueness theorem, proved by Bunting and Masood-ul-Alam \cite{Bunting_Masood}. \vspace{.3cm} \noindent {\bf Theorem} (\cite{Bunting_Masood}) {\em Let $(M^3, g)$ be a static, asymptotically flat manifold with a nonempty smooth boundary. Let $N$ be the static potential function of $(M^3, g)$. If $N$ satisfies \begin{equation} N |_{\partial M} = 0 , \end{equation} then $(M^3, g)$ is isometric to a spatial Schwarzschild manifold outside its horizon.} \vspace{.3cm} The following theorem is a direct application of Corollary \ref{maincor} and the maximum principle. \begin{thm} \label{gofBM} Let $(M^3, g)$ be a static, asymptotically flat manifold with a connected smooth boundary $\Sigma$. Assume that $(M^3, g)$ is diffeomorphic to $\mathbb{R}^3 \setminus \Omega$, where $\Omega$ is a bounded domain. Let $N$ be the static potential function of $(M^3, g)$. If $\Sigma$ has nonnegative Hawking mass, then \begin{equation} \label{estofN} \min_{\Sigma} N^2 \leq \frac{1}{16 \pi} \int_{\Sigma} H^2 d \mu . \end{equation} Furthermore, equality holds if and only if $(M^3, g)$ is isometric to a spatial Schwarzschild manifold $$ (M_{r_0}, g^S_m) = \left([r_0, \infty) \times S^2, \frac{1}{1 - \frac{2m}{r}} dr^2 + d\sigma^2 \right), $$ where $r_0$ is some positive constant satisfying $r_0 \geq 2m$ and $d\sigma^2$ is the standard metric on the unit spere $S^2 \subset \mathbb{R}^3$. \end{thm} \noindent {\bf Remark}: If $N|_\Sigma = 0$, the static metric system (\ref{staticeq}) implies that $\Sigma$ is totally geodesic \cite{Bunting_Masood}. Hence the equality in (\ref{estofN}) holds automatically. Thus, Theorem \ref{gofBM} can be viewed as a partial generalization of Bunting and Masood-ul-Alam's theorem. \vspace{.3cm} \noindent {\em{Proof}}: Let $\alpha = \sqrt{\frac{1}{16\pi} \int_\Sigma H^2 d \mu}$ and let $\mathcal{G}$ be the function on $M$ defined by \begin{equation} \nonumber \left\{ \begin{array}{rcl} \lim_{x \rightarrow \infty} \mathcal{G} & = & 1 \\ \triangle \mathcal{G} & = & 0 \\ \mathcal{G}|_{\partial M} & = & \alpha . \end{array} \right. \end{equation} Consider the asymptotic expansions of $\mathcal{G}$ and $N$, $$ \mathcal{G} = 1 - \frac{\mathcal{C}}{|x|} + {O} \left( \frac{1}{|x|^2} \right), \ as \ x \rightarrow \infty $$ $$ N = 1 - \frac{A}{|x|} + {O} \left( \frac{1}{|x|^2} \right), \ as \ x \rightarrow \infty . $$ Suppose $\min_{\Sigma} N \geq \alpha$, the strong maximum principle then implies that \begin{equation} \label{AC} A \leq \mathcal{C} \end{equation} with equality if and only if $N = \mathcal{G}$. By analyzing the static metric system (\ref{staticeq}), Bunting and Masood-ul-Alm in \cite{Bunting_Masood} showed that \begin{equation} \label{AM} A = m , \end{equation} where $m$ is the total mass of $(M^3, g)$. On the other hand, as $(M^3, g)$ has zero scalar curvature and $m_H(\Sigma) \geq 0$, Corollary \ref{maincor} shows that \begin{equation} \label{MC} m \geq \mathcal{C}. \end{equation} Therefore, it follows from (\ref{AC}), (\ref{AM}) and (\ref{MC}) that $$ m = \mathcal{C} $$ and $\ N = \mathcal{G} $. In particular, $(M^3, g)$ is isometric to $(M_{r_0}, g^S_m)$ by the rigidity part in Corollary \ref{maincor}. Theorem \ref{gofBM} is proved. \hfill$\Box$ \bibliographystyle{plain}

https://arxiv.org/abs/2112.14999

On weakly coupled systems of partial differential equations with different diffusion terms

We prove maximal Schauder regularity for solutions to elliptic systems and Cauchy problems, in the space $C_b(\mathbb{R}^d;\mathbb{R}^m)$ of bounded and continuous functions, associated to a class of nonautonomous weakly coupled second-order elliptic operators $\bf{\mathcal A}$, with possibly unbounded coefficients and diffusion and drift terms which vary from equation to equation. We also provide estimates of the spatial derivatives up to the third-order and continuity properties both of the evolution operator ${\bf G}(t,s)$ associated to the Cauchy problem $D_t{\bf u}=\bf{\mathcal A}(t){\bf u}$ in $C_b(\mathbb{R}^d;\mathbb{R}^m)$, and, for fixed $\overline t$, of the semigroup ${\bf T}_{\overline t}(\tau)$ associated to the autonomous Cauchy problem $D_{\tau}{\bf u}={\bf{\mathcal A}}(\overline t){\bf u}$ in $C_b(\mathbb{R}^d;\mathbb{R}^m)$. These results allow us to deal with elliptic problems whose coefficients also depend on time.

\section{Introduction} In this paper, we consider weakly coupled nonautonomous elliptic operators $\bm {\mathcal A}$, defined on smooth functions $\boldsymbol\zeta:\mathbb R^d\to\mathbb R^m$ ($m\ge 2$), by \begin{align}\label{picco} (\bm{\mathcal A}(t)\boldsymbol\zeta(x))_k& =\sum_{i,j=1}^dq_{ij}^k(t,x)D_{ij}\zeta_k(x)+ \sum_{j=1}^db^k_j(t,x)D_j\zeta_k(x)+\sum_{h=1}^mc_{kh}(t,x)\zeta_h(x)\notag\\ & = {\rm Tr}(Q^k(t,x)D^2\zeta_k(x))+ \langle {\bm b}^k(t,x), \nabla \zeta_k(x)\rangle+ (C(t,x)\boldsymbol\zeta(x))_k, \end{align} $k=1,\ldots,m$, for every $(t,x)\in I\times \mathbb R^d$, with possible unbounded coefficients, and for every $T>s\in I$ (where $I$ is a right halfline, possibily $I={\mathbb R}$) we prove maximal Schauder regularity for the solutions to the elliptic system \begin{align} \label{intro:ell_eq} \lambda {\bm u}(t,x)-\bm{\mathcal A}(t){\bm u}(t,x)=\bm f(t,x), \quad t\in[s,T], \ x\in\mathbb R^d, \end{align} and to the Cauchy problem \begin{align} \label{intro:cau_pro} \left\{ \begin{array}{lll} D_t{\bm u}(t,x)=\bm{\mathcal A}(t){\bm u}(t,x)+{\bm g}(t,x), & t\in[s,T], & x\in\mathbb R^d, \vspace{1mm} \\ {\bm u}(s,x)=\bm f(x), & & x\in\mathbb R^d. \end{array} \right. \end{align} The study of scalar equations with unbounded coefficients has been widely studied in the last two decades and the theory is well-established; we refer to the monograph \cite{Lo17} and the references therein for a systematic treatment. On the other hand, the study of systems of Kolmogorov equations with unbounded coefficients is at the beginning and just few results are available. One of the first papers dealing with systems of parabolic equations with unbounded coefficients is \cite{HLRS09}, where the authors study the realization of the weakly coupled elliptic operator $\bm{\mathcal A}_p{\bm u}:={\rm div} (Q\nabla {\bm u})+\langle F,\nabla {\bm u}\rangle +V{\bm u}$ in $L^p(\mathbb R^d;\mathbb R^m)$ and characterize its domain under suitable assumptions on the coefficients. Weakly coupled operators in the space of bounded and continuous functions $C_b(\mathbb R^d;\mathbb R^m)$ have been considered in \cite{DeLo11}, where the authors prove existence and uniqueness of a classical solution to the associated Cauchy problem. This allows to define a semigroup of bounded operators $({\bm T}(t))_{t\geq0}$ on $C_b(\mathbb R^d;\mathbb R^m)$, and then to study some of its main properties, such as compactness, uniform estimates of the derivatives and optimal Schauder regularity. A first improvement of the results in \cite{DeLo11} appears in \cite{AALT17}, where nonautonomous second-order operators coupled up to first order are considered. Also in this case, existence and uniqueness of a classical solution to the Cauchy problem associated with the operator $\bm{\mathcal A}(t)$ allow the authors to define an evolution operator $({\bm G}(t,s))_{t\geq s}$ on $C_b(\mathbb R^d;\mathbb R^m)$ and to investigate the main features of the evolution operator. In particular, thanks to weighted gradient estimates the authors prove the existence of Nash equilibria for a class of stochastic differential games. The case of operators coupled up to the first-order has been considered also in \cite{AngLorPal16}, where the authors provide sufficient conditions for the evolution operator $({\bm G}(t,s))_{t\geq s}$ to be extended to the space $L^p(\mathbb R^d;\mathbb R^m)$, and some improving summability properties of the evolution operator are shown. We stress that in all the quoted paper the diffusion coefficients are the same for each equation, i.e., the operator $\bm{\mathcal A}(t)$ is defined on smooth functions $\bm f:\mathbb R^d\rightarrow \mathbb R^m$ by \begin{align*} \left((\bm{\mathcal A}(t)\bm f)(x)\right)_k:=\sum_{i,j=1}^d q_{ij}(t,x) D_{ij}f_k(x)+\sum_{j=1}^d \sum_{h=1}^mb_{jh}^k(t,x)D_jf_h(x)+\sum_{h=1}^m c_{kh}(t,x)f_h(x), \end{align*} for every $k=1,\ldots,m$ and every $(t,x)\in I\times \mathbb R^d$. Under suitable conditions on the coefficients of the operator $\bm{\mathcal A}$, this allows to apply a maximum principle, and so uniqueness of the classical solution to the Cauchy problem associated with $\bm{\mathcal A}$ follows. We notice that this is the main effort when one deals with systems of Kolmogorov equations with unbounded coefficients, since existence of a classical solution usually follows by means of an approximation procedure which mimics the same technique in the case of a single equation (see e.g., \cite{MePaWa02}) and relies on a-priori (interior Schauder) estimates. The first attempt to deal with operators with unbounded coefficients and diffusion terms which may vary from line to line appears in \cite{AngLor20}, where the authors study nonautonomous second-order weakly coupled operators defined as \eqref{picco}. Thanks to a generalization of a result in \cite{Pr67}, the authors get existence and uniqueness of a classical solution to the Cauchy problem associated with $\bm{\mathcal A}$, and define an evolution operator $({\bm G}(t,s))_{t\geq s}$ by means of this solution. Some remarkable properties of the evolution operator, such as its compactness, its action on some functional spaces and the existence of systems of family of invariant measures, are investigated. The approach in \cite{AngLor20} strongly relies on the assumption that the off-diagonal entries of the matrix $C(t,\cdot)$ are bounded from below for every $t\in I$. In this paper we are able to remove this assumption thanks to a comparison principle, which allows us to ``control'' the solution of the Cauchy problem associated to $\bm{\mathcal A}$ with that associated to a suitable nonautonomous operator $\bm{\mathcal A}^P$ which satisfies the assumptions in \cite{AngLor20}. Such an operator $\bm{\mathcal A}^P$ is a fundamental auxiliary tool in the paper: it also appears in Section \ref{section:grad_est}, in order to get both the estimates for spatial derivatives of the evolution operator and the maximal Schauder regularity for problem \eqref{intro:ell_eq}, and in Section \ref{section:special_case} to extend some results in \cite{AdAnLo19} in the autonomous case, showing that the semigroup associated with a class of autonomous systems of ellipic operators extends to a strongly continuous semigroup on the $L^p$-space with respect to a system of {\em signed} invariant measures $\boldsymbol \mu=(\mu_1,\ldots,\mu_m)$ associated to the semigroup. An analogous result has been already proved in \cite{AdAnLo19}, but under the assumptions that the semigroup is nonnegative, and so each component of $\boldsymbol \mu$ is a positive measure, while, in our situation, the semigroup does not preserve positivity and we allow some component of $\boldsymbol \mu$ to be negative. To the best of our knowledge, this is the first time when a semigroup associated to a system of elliptic equations is extended to the $L^p$-spaces with respect to a system of invariant measures, without any positivity preserving condition on it. Also in \cite{AdAnLo18} systems of invariant measures for semigroups associated with a class of autonomous systems of elliptic operators, which do not preserve positivity, are studied, but no results on the possibility to extend the semigroups to $L^p$-spaces related to such systems of measures are established. We finally remark that an abstract approach to asymptotic behaviour of semigroups can be found in \cite{GeGlKu20}, but again under some positivity assumption on the semigroup. The paper is organized as follows. In Section \ref{section:preliminaries}, we provide the main hypotheses on the coefficients of the operator $\bm{\mathcal A}$ and fix some notation which will be useful in the rest of the paper. Further, we prove a comparison principle between the solutions of Cauchy problems on balls, with homogeneous Neumann boundary conditions, associated to $\bm{\mathcal A}$ and $\bm{\mathcal A}^P$, and, based on this result, we show existence, uniqueness and some continuity properties of the evolution operator $({\bm G}(t,s))_{t\geq s}$ (simply denoted by ${\bm G}(t,s)$ from now on) associated to the nonautonomous equation $D_t{\bm u}=\bm{\mathcal A}{\bm u}$ on $C_b(\mathbb R^d;\mathbb R^m)$. Finally, we deduce analogous results for the semigroup $({\bm T}_{\overline t}(\tau))_{\tau\geq0}$ (simply denote by ${\bm T}_{\overline t}(\tau)$ from now on) associated to the autonomous equation $D_\tau{\bm u}=\bm{\mathcal A}(\overline t){\bm u}$ on $C_b(\mathbb R^d;\mathbb R^m)$, where $\overline t\in I$ is fixed. In Section \ref{section:grad_est}, under additional assumptions on the coefficients of the operator $\bm{\mathcal A}$, we prove some estimates for the derivatives of the functions ${\bm G}(t,s)\bm f$ and ${\bm T}_{\overline t}(\tau)\bm f$. To be more precise, we show that for every integer numbers $0\leq h\leq k\leq 3$, every $[s,T]\subset I$, and for every $\overline t\in I$ and every $\mathcal T>0$ there exist positive constants $C$ and $K$ such that \begin{align*} \|{\bm G}(t,s)\bm f\|_{C_b^k(\mathbb R^d;\mathbb R^m)}& \leq \frac{Ce^{K(t-s)}}{(t-s)^{(k-h)/2}}\|\bm f\|_{C^h_b(\mathbb R^d;\mathbb R^m)}, \qquad\;\, \bm f\in C^h_b(\mathbb R^d;\mathbb R^m),\;\, t\in(s,T]. \\ \|{\bm T}_{\overline t}(\tau)\bm f\|_{C_b^k(\mathbb R^d;\mathbb R^m)}& \leq \frac{Ce^{K\tau}}{\tau^{(k-h)/2}}\|\bm f\|_{C^h_b(\mathbb R^d;\mathbb R^m)}, \qquad\;\, \bm f\in C^h_b(\mathbb R^d;\mathbb R^m),\;\, \tau\in (0,\mathcal T]. \end{align*} These estimates are the starting point to prove, in Section \ref{sec:max_reg}, maximal Schauder regularity for problems \eqref{intro:ell_eq} and \eqref{intro:cau_pro}. In Section \ref{section:special_case}, assuming that the diffusion and drift coefficients are the same for each equation and the operator $\bm{\mathcal A}$ is autonomous, we prove the existence of a system of invariant measures for the semigroup ${\bm T}(t)$ associated to the Cauchy problem $D_t{\bm u}=\bm{\mathcal A}{\bm u}$, characterizing them. Moreover, we show that such a semigroup extends to a strongly continuous semigroup on $L^p_{\boldsymbol \mu}(\mathbb R^d;\mathbb R^m)$ and study the asymptotic behaviour of the semigroup ${\bm T}(t)$ as $t$ tends to $\infty$. \medskip \paragraph{\bf Notation} Functions with values in ${\mathbb R}^m$ are displayed in bold style. Given a function $\bm f$ (resp. a sequence $(\bm f_n)$) as above, we denote by $f_i$ (resp. $f_{n,i}$) its $i$-th component (resp. the $i$-th component of the function $\bm f_n$). We denote by $|\bm f|$ the vector-valued function whose components are $|f_1|,\ldots,|f_m|$. By $B_b(\mathbb R^d;\mathbb R^m)$ we denote the set of all the bounded Borel measurable functions $\bm f:\mathbb R^d\to\mathbb R^m$, where $\|\bm f\|_{\infty}^2=\sum_{k=1}^m\sup_{x\in\mathbb R^d}|f_k(x)|^2$. For every $k\ge 0$, $C^k_b({\mathbb R}^d;{\mathbb R}^m)$ is the space of all the functions whose components belong to $C^k_b({\mathbb R}^d)$, where ``$b$'' stays for bounded. Similarly, we use the subscript ``$c$'' for spaces of functions with compact support. When $k\in (0,1)$ we write $C^k_{\rm loc}(\mathbb R^d)$ to denote the space of all the real-valued functions $f\in C(\mathbb R^d)$ which are H\"older continuous in every compact set of $\mathbb R^d$. We assume that the reader is familiar with the parabolic spaces $C^{h+\alpha/2,k+\alpha}(I\times \mathbb R^d)$ ($\alpha\in [0,1)$, $h,k\in{\mathbb N}\cup\{0\}$), and we use the subscript ``loc'' with the same meaning as above. The symbols $D_tf$, $D_i f$, $D_{ij}f$ and $D_{ijh}f$, respectively, denote the time derivative $\frac{\partial f}{\partial t}$ and the spatial derivatives $\frac{\partial f}{\partial x_i}$, $\frac{\partial^2f}{\partial x_i\partial x_j}$ and $\frac{\partial^3f}{\partial x_i\partial x_j\partial x_h}$ for every $i,j,h=1, \ldots,d$. For $k=1,2,3$, $|D_x^k\bm f|^2$ denotes the sum $\sum_{j=1}^m|D^k_xf_j|^2$, where $|D_xf_j|^2=\sum_{i=1}^m|D_if_j|^2$, $|D_x^2f_j|^2=\sum_{i,h=1}^m|D_{ih}f_j|^2$ and $|D_x^3f_j|^2=\sum_{i,h,k=1}^m|D_{ihk}f_j|^2$. If $k=1$ we write $D_x\bm f$ and $J_x\bm f$ indifferently for the Jacobian matrix of $\bm f$ with respect to the spatial variables. If $\bm f$ does not depend on $t$, we omit the subscript $x$ and we simply write $D^k\bm f$. Similarly, we write $D^2_x\bm f$ and $D^3_x\bm f$, when $\bm f$ depends also on $t$, to denote differentiation with respect to the spatial variables. By $e_j$ and $\mathds 1$ we denote, respectively, the $j$-th vector of the Euclidean basis of ${\mathbb R}^d$ and the function identically equal to $1$ in $\mathbb R^m$. The open ball in $\mathbb R^d$ centered at $ 0$ with radius $r>0$ and its closure are denoted by $B_r$ and $\overline B_r$, respectively. \section{Hypotheses and preliminary results} \label{section:preliminaries} Besides the nonautonomous operator $\bm{\mathcal A}$ defined in \eqref{picco}, we introduce the auxiliary nonautonomous operator ${\bm\bm{\mathcal A}}^P$, defined on smooth functions $\bm f:{\mathbb R}^d\rightarrow {\mathbb R}^m$ by \begin{align} \label{auxiliary_operator} (({\bm{\mathcal A}}^P(t)\bm f)(x))_k:=\sum_{i,j=1}^nq^k_{ij}(t,x)D_{ij}f_k(x)+\sum_{j=1}^nb^k_j(t,x)D_jf_k(x)+(C^P(t,x)\bm f(x))_k \end{align} for every $k=1,\ldots,m$ and $(t,x)\in I\times \mathbb R^d$, where the matrix $C^P=(c_{hk}^P)$ is given by \begin{align*} c_{kh}^P:= \begin{cases} c_{kh}, & h=k, \\ |c_{kh}|, & h\neq k, \end{cases}, \quad h,k=1,\ldots,m. \end{align*} Throughout the paper, if not otherwise specified, we assume the following assumptions on the coefficients of the operator $\bm{\mathcal A}$ and ${\bm{\mathcal A}^P}$ in \eqref{picco} and in \eqref{auxiliary_operator}. \begin{hyp} \label{hyp-base} \begin{enumerate}[\rm (i)] \item The coefficients $q^k_{ij}=q^k_{ji}$, $b^k_j$ and the entries $c_{kh}$ of the matrix-valued function $C$ belong to $C^{\alpha/2,\alpha}_{\rm loc}(I\times \mathbb R^d)$ for some $\alpha\in (0,1)$; \item for every bounded interval $J\subset I$, the infimum $\mu_0$ over $J\times \mathbb R^d$ of the minimum eigenvalue $\mu^k(t,x)$ of the matrix $Q^k(t,x)=(q^k_{ij}(t,x))$ is positive for every $k=1,\ldots,m$; \item for every bounded interval $J\subset I$ there exists a componentwise positive function $\boldsymbol\varphi_J\in C^2(\mathbb R^d;\mathbb R^m)$, blowing up componentwise as $|x|$ tends to $\infty$ such that ${\bm{\mathcal A}^P}\boldsymbol\varphi_J\le \lambda_J\boldsymbol\varphi_J$ in $J\times\mathbb R^d$, for some positive constant $\lambda_J$; \item there does not exist a nontrivial set $K\subset \{1, \ldots, m\}$ such that the coefficients $c^P_{ij}$ identically vanish on $I\times \mathbb R^d$ for every $i\in K$ and $j\notin K$; \item for every bounded interval $J\subset I$, the sum of the elements of each row of $C^P$ is a bounded from above function on $J\times \mathbb R^d$. \end{enumerate} \end{hyp} \begin{rmk} {\rm Hypothesis $\ref{hyp-base}(iv)$ is a condition on the entries of the matrix-valued function $C^P$ which guarantees that both the differential systems \eqref{pioggia} and \eqref{cauchy_prob_pos} cannot be (partially) decoupled. It is not hard to see that if $\bm{\mathcal A}$ satisfies \cite[Hypotheses 2.1]{AngLor20}, then the operators $\bm{\mathcal A}$ and $\bm{\mathcal A}^P$ satisfy Hypotheses \ref{hyp-base}} above. \end{rmk} For every bounded interval $J\subset I$ we set \begin{align} \label{def_M_T} M_{J}:=\max_{k=1,\ldots,m}\sup_{(t,x)\in J\times \mathbb R^d}\sum_{j=1}^mc_{kj}^P(t,x)=:\max_{k=1,\ldots,m}\sup_{(t,x)\in J\times \mathbb R^d}M_k(t,x). \end{align} For every $s,t\in I$, every $\alpha\in[0,1)$ and every bounded open set $\Omega\subset \mathbb R^d$ we denote by $\|\bm{\mathcal A}(t)\|_{\alpha,\Omega}$ and by $\|\bm{\mathcal A}(t)-\bm{\mathcal A}(s)\|_{\alpha,\Omega}$ the following quantities: \begin{align*} \|\bm{\mathcal A}(t)\|_{\alpha,\Omega}:=& \max_{ {i,j=1,\ldots,d}\atop{h,k=1,\ldots,m}} \left\{\|q^k_{ij}(t,\cdot)\|_{C_b^\alpha(\Omega)}, \|b^k_{i}(t,\cdot)\|_{C_b^\alpha(\Omega)}, \|c_{kh}(t,\cdot)\|_{C_b^\alpha(\Omega)}\right\}, \\ \|\bm{\mathcal A}(t)-\bm{\mathcal A}(s)\|_{\alpha,\Omega}:=& \max_{{i,j=1,\ldots,d}\atop{ h,k=1,\ldots,m}}\left\{\|q^k_{ij}(t,\cdot)-q_{ij}^k(s,\cdot)\|_{C_b^\alpha(\Omega)}, \|b^k_{i}(t,\cdot)-b_i^k(s,\cdot)\|_{C_b^\alpha(\Omega)}, \right.\\ & \qquad\qquad\;\; \left.\|c_{kh}(t,\cdot)-c_{kh}(s,\cdot)\|_{C_b^\alpha(\Omega)}\right\}, \end{align*} \paragraph{\bf {Further notation}.} Let us fix $R>0$. \begin{enumerate}[(i)] \item We denote by ${\bm G}_{R}^{\mathcal N}(t,s)$ (resp. ${\bm G}_{R}^{\mathcal N,P}(t,s)$) the evolution operator associated to the realization of the nonautonomous operator $\bm{\mathcal A}$ (resp. $\bm{\mathcal A}^p$) in $C(\overline{B_R};\mathbb R^m)$ with homogeneous Neumann boundary conditions. \item For every fixed $\overline t\in I$ we denote by ${\bm T}_{\overline t}^{\mathcal N,R}(\tau)$ (resp. ${\bm T}_{\overline t}^{\mathcal N,R,P}(\tau)$) the semigroup associated to the realization of the operator $\bm{\mathcal A}(\overline t)$ (resp. $\bm{\mathcal A}^P(\overline t)$) in $C(\overline{B_R};\mathbb R^m)$ with homogeneous Neumann boundary conditions. \end{enumerate} We notice that the existence of the above families of operators is guaranteed by \cite[Section IV.2, Theorem 3.4]{Ei69}. Fix $s\in I$ and consider the Cauchy problems \begin{equation} \left\{ \begin{array}{lll} D_t{\bm u}(t,x)=(\bm{\mathcal A}(t){\bm u})(t,x), & t\in(s,\infty), &x\in\mathbb R^d,\\[1mm] {\bm u}(s,x)=\bm f(x), && x \in \mathbb R^d, \end{array} \right. \label{pioggia} \end{equation} and \begin{equation} \left\{ \begin{array}{lll} D_t{\bm u}(t,x)=(\bm{\mathcal A}^P(t){\bm u})(t,x), & t\in(s,\infty), &x\in\mathbb R^d,\\[1mm] {\bm u}(s,x)=\bm f(x), && x \in \mathbb R^d. \end{array} \right. \label{cauchy_prob_pos} \end{equation} By \cite[Theorem 2.4]{AngLor20} problem \eqref{cauchy_prob_pos} admits a unique classical solution ${\bm u}^P$, which is bounded in each strip $[s,T]\times \mathbb R^d\subset I\times\mathbb R^d$, belongs to $ C^{1+\alpha/2,2+\alpha}_{\rm loc}((s,\infty)\times\mathbb R^d;\mathbb R^m)\cap C([s,\infty)\times\mathbb R^d;\mathbb R^m)$ and satisfies the estimate $\|{\bm u}^P(t,\cdot)\|_{\infty}\le e^{K_{[s,T]}(t-s)}\|\bm f\|_{\infty}$ for every $t\in[s,T]$, where $K_{[s,T]}$ is explicitly computed. The aim of this section is to prove that for every $\bm f\in C_b(\mathbb R^d;{\mathbb R}^m)$ the Cauchy problem \eqref{pioggia} admits a unique classical solution which is bounded in the strip $J\times\mathbb R^d$ for every bounded interval $J\subset [s,\infty)$, where by classical solution we mean a function ${\bm u}\in C^{1,2}((s,\infty)\times\mathbb R^d;\mathbb R^m)\cap C([s,\infty)\times\mathbb R^d;\mathbb R^m)$ which solves \eqref{pioggia}. We begin by considering the following proposition. \begin{prop} \label{prop:confronto_1} Let Hypotheses $\ref{hyp-base}(i)$-$(iv)$ be satisfied. Then, for each $n\in{\mathbb N}$, $\bm f\in C_b({\mathbb R}^d;{\mathbb R}^m)$ and $j\in\{1,\ldots,m\}$ it holds that $|({\bm G}_n^{\mathcal N}(t,s)\bm f)_j|\leq ({\bm G}_n^{\mathcal N,P}(t,s)|\bm f|)_j$ in $B_n$ for every $t>s$. \end{prop} \begin{proof} We fix $n\in{\mathbb N}$, $\bm f\in C_b(\mathbb R^d;\mathbb R^m)$ and set ${\bm u}:={\bm G}^{\mathcal N}_n(\cdot,s)\bm f$, $\bm v:={\bm G}^{\mathcal N,P}_n(\cdot,s)|\bm f|$. For every $k=1,\ldots,m$ we denote by $G_{n,k}^{\mathcal N}(t,s)$ the positive evolution operator associated in $C(\overline{B_n})$ to the realization of the scalar operator ${\mathcal A}_{n,k}=\sum_{i,j=1}^d q_{ij}^kD_{ij}+\sum_{j=1}^db_{j}^kD_j+c_{kk}$, with homogeneous Neumann boundary conditions, and note that for every $T>s$ we can estimate $\|G^{\mathcal N}_{n,k}(t,s)\|_{L(C(\overline{B_n}))}\leq e^{c_{T,k,n}}$ for every $k=1,\ldots,m$ and $t\in[s,T]$, where $c_{k,T,n}:=\sup_{(t,x)\in (s,T)\times B_n}c_{kk}(t,x)$. Let us fix $T>s$ and prove the statement for $t\in[s,T]$; the arbitrariness of $T$ will yield the assertion. By the variation-of-constants formula, we can write \begin{align*} u_k(t,x) = & (G^{\mathcal N}_{n,k}(t,s)f_k)(x)+\int_s^t\bigg (G^{\mathcal N}_{n,k}(t,r)\sum_{j\neq k}c_{kj}(r,\cdot)u_j(r,\cdot)\bigg )(x)dr,\\[1mm] v_k(t,x) = & (G^{\mathcal N}_{n,k}(t,s)|f_k|)(x)+\int_s^t\bigg (G^{\mathcal N}_{n,k}(t,r)\sum_{j\neq k}c^P_{kj}(r,\cdot)v_j(r,\cdot)\bigg )(x)dr \end{align*} for every $(t,x)\in[s,T]\times\overline{B_n}$. Moreover, the positivity of the evolution operator $G^{\mathcal N}_{n,k}(t,s)$ implies that \begin{align*} |u_k(t,x)|\leq & (G^{\mathcal N}_{n,k}(t,s)|f_k|)(x)+\int_s^t\bigg (G^{\mathcal N}_{n,k}(t,r)\sum_{j\neq k}c^P_{kj}(r,\cdot)|u_j(r,\cdot)|\bigg )(x)dr \end{align*} for every $(t,x)\in[s,T]\times B_n$, so that \begin{align*} w_k(t,x) \le & \int_s^t\bigg (G^{\mathcal N}_{n,k}(t,r)\sum_{j\neq k}c^P_{kj}(r,\cdot)w_j(r,\cdot)\bigg )(x)dr, \qquad\;\, (t,x)\in[s,T]\times B_n, \end{align*} where ${\bm w}=|{\bm u}|-{\bm v}$ satisfies the inequality ${\bm w}(s,\cdot)\le\bm 0$ in $B_n$. To show that ${\bm w}\le\bm 0$ on $[s,T]\times B_n$, we just need to prove that, if \begin{align} w_k(t,x) \le & \int_{\sigma}^t\bigg (G^{\mathcal N}_{n,k}(t,r)\sum_{j\neq k}c^P_{kj}(r,\cdot)w_j(r,\cdot)\bigg )(x)dr \label{quattro_1} \end{align} for every $(t,x)\in (\sigma,T)\times B_n$, $k=1,\ldots,m$ and $\bm w(\sigma,\cdot)\le\bm 0$ in $B_n$ for some $\sigma\in [s,T)$, then there exists $\delta>0$ such that $\bm w\le\bm 0$ in $[\sigma,\sigma+\delta]\times B_n$. Indeed, once this property is proved, we can introduce the set $E=\{t\in [s,T]:{\bm w}(t,\cdot)\leq\bm 0{\textrm{ in }}B_n\}$, which contains $s$ and is an interval due to \eqref{quattro_1}. Note that $E=[s,T]$. By contradiction, assume that $\tau=\sup E<T$. By continuity, it follows that ${\bm w}(r,\cdot)\leq\bm 0$ in $B_n$ for every $r\in [s,\tau]$. Since $\displaystyle \sum_{j\neq k}c^P_{kj}(r,\cdot)w_j(r,\cdot)\le 0$ in $B_n$ for every $r\in [s,\tau]$, from \eqref{quattro_1} we can write \begin{eqnarray*} w_k(t,x) \le \int_{\tau}^t\bigg (G^{\mathcal N}_{n,k}(t,r)\sum_{j\neq k}c^P_{kj}(r,\cdot)w_j(r,\cdot)\bigg )(x)dr, \qquad\;\, (t,x)\in[\tau,T]\times B_n \end{eqnarray*} and from this inequality we conclude that $\bm w\le\bm 0$ in a right neighborhood of $\tau$, which is a contradiction. Let us prove \eqref{quattro_1} by contradiction. For this purpose, for every $h\in{\mathbb N}$ we denote by $J_h$ the set of all the indexes in $\{1,\ldots,m\}$ such that the condition $w_j\le 0$ is not satisfied in $[\sigma,\sigma+h^{-1}]\times\overline{B_n}$, where $h\in{\mathbb N}$ is such that $h^{-1}<T-\sigma$. Note that there exists at least an index $j_0$ such that $j_0\in J_h$ for infinite values of $h$ (let us say for every $h$ in an increasing sequence of natural numbers $(h_k)$). Finally, denote by $J$ the set of all $j\in\{1,\ldots,m\}$ such that $j\in J_{h_k}$ for every $k\in{\mathbb N}$. For such values of $j$, we can determine a decreasing sequence $(t_{k,j})\subset (\sigma,T]$ converging to $\sigma$ and $(x_{k,j})\subset\overline{B_n}$ such that $w_j(t_{k,j},x_{k,j})>0$ for every $k\in{\mathbb N}$. This implies that $h_j(r):=\sup_{({\sigma},r)\times B_n}w_j>0$ for every $r>{\sigma}$ and $j\in J$. Note that each function $h_j$ is continuous in $[{\sigma},\infty)$ since both $|{\bm u}|$ and ${\bm v}$ are continuous in $[s,T]\times \overline{B_n}$. Let $\delta'>0$ be such that $w_i\le 0$ in $[\sigma,\sigma+\delta']\times B_n$ for every $j\in \{1,\ldots,m\}\setminus J$. If $k\in J$, then we can estimate \begin{align*} w_{k}(t,x) \leq & \int_{\sigma}^t \bigg (G^{\mathcal N}_{n,k}(t,r)\sum_{k\neq j\in J}c_{kj}^P(r,\cdot)w_j(r,\cdot)\bigg )(x)dr \notag\\ \leq & \int_{\sigma}^t \bigg (G^{\mathcal N}_{n,k}(t,r)\sum_{k\neq j\in J}c_{kj}^P(r,\cdot)h_j(r)\bigg )(x)dr \notag\\ \leq & C\int_{\sigma}^t \sum_{j\in J}h_j(r)dr \end{align*} for every $(t,x)\in [\sigma,\sigma+\delta']\times\overline{B_n}$, where we have used the fact that $c_{kj}^P\ge 0$ for every $j\neq k$, $h_j(r)>0$ for every $r\in({\sigma},T]$ and $j\in J$, and \begin{align*} C:=(e^{c_{T,k,n}}\vee 1)\sup_{{j,k\in J}\atop{j\neq k}}\|c_{kj}^P\|_{C([s,T]\times\overline{B_n})}<\infty. \end{align*} Hence, if we fix $t\in[\sigma,\sigma+\delta']$, then for every $\tau\in[\sigma,t]$ we can write \begin{align*} w_k(\tau,x) \leq C\int_{\sigma}^{\tau} \sum_{j\in J}h_j(r)dr\leq C\int_\sigma^t \sum_{j\in J}h_j(r)dr. \end{align*} By taking the supremum of $(\tau,x)$ over $(\sigma,t)\times B_n$ and summing over all $k\in J$, we get \begin{align*} \sum_{k\in J}h_k(t)\leq Cm\int_{\sigma}^t \sum_{k\in J}h_k(r)dr, \qquad\;\, t\in[\sigma,\sigma+\delta']. \end{align*} Gronwall Lemma implies that $\sum_{k\in J}h_k(t)\leq 0$ in $[\sigma,\sigma+\delta']$, which contradicts the fact that $h_k(t)>0$ for every $t>{\sigma}$ and $k\in J$. \end{proof} \begin{prop} \label{prop-2.6} Under Hypotheses $\ref{hyp-base}$, fix $s\in I$, $T>s$ and let $M:=M_{[s,T]}$ be the constant in \eqref{def_M_T}. Then, the following properties are satisfied. \begin{enumerate}[\rm (i)] \item The unique solution ${\bm u}^P$ to problem \eqref{cauchy_prob_pos} satisfies the estimate \begin{align} \max_{k=1,\ldots,d}|u^P_k(t,x)|\leq e^{M(t-s)} \max_{k=1,\ldots,m}\|f_k\|_\infty, \qquad\;\, (t,x)\in [s,T]\times \mathbb R^d. \label{tre} \end{align} \item There exists a unique classical solution ${\bm u}$ to problem \eqref{pioggia}, which is bounded in each strip $[s,T]\times\mathbb R^d$. Moreover, if we denote by $\widetilde {\bm u}^P$ the solution to \eqref{cauchy_prob_pos} with initial datum $|\bm f|$, then $|u_k(t,x)|\leq |\widetilde u_k^P(t,x)|$ for every $k=1,\ldots,m$ and $(t,x)\in[s,T]\times \mathbb R^d$. As a consequence, \begin{align} \label{stima_uni_gen} \max_{k=1,\ldots,d}|u_k(t,x)|\leq e^{M(t-s)} \max_{k=1,\ldots,m}\|f_k\|_\infty, \qquad\;\, (t,x)\in [s,T]\times \mathbb R^d. \end{align} \end{enumerate} \end{prop} \begin{rmk} {\rm Differently from \cite[Theorem 2.4]{AngLor20}, even if the off-diagonal entries of the matrix $C$ assume negative values, the constant $M$ in the above estimates does not depend on the lower bound of the off-diagonal entries of the matrix $C$, but only on the greatest upper bound of the sum of the elements on each row of $C^P$. This is a consequence of Proposition \ref{prop:confronto_1}, as we will see in the subsequent proof.} \end{rmk} \begin{proof}[Proof of Proposition $\ref{prop-2.6}$] (i) Let us set ${\bm v}(t,x):={\bm u}^P(t,x)-e^{M(t-s)}\max_{h=1,\ldots,m}\|f_h\|_\infty \mathds 1$ for each $(t,x)\in [s,T]\times \mathbb R^d$. For every $k\in \{1,\ldots,m\}$, $D_tv_k-(\bm{\mathcal A}^P{\bm v})_k$ is non positive in $(s,T]\times \mathbb R^d$ and ${\bm v}(s,\cdot)\leq \bm 0$ in $\mathbb R^d$. The maximum principle in \cite[Theorem 2.3]{AngLor20} (whose proof holds true also in our situation) implies that $u^P_k(t,x)\leq e^{M(t-s)} \max_{h=1,\ldots,m}\|f_h\|_\infty$ for every $(t,x)\in[s,T]\times \mathbb R^d$ and $k\in\{1,\ldots,m\}$. Replacing ${\bm u}^P$ with the function $-{\bm u}^P$, we obtain the inequality $u^P_k(t,x)\geq -e^{M(t-s)} \max_{k=1,\ldots,m}\|f_k\|_\infty$ for $(t,x)\in[s,T]\times \mathbb R^d$ and $k\in\{1,\ldots,m\}$, and \eqref{tre} follows. \medskip (ii) We split the proof into two parts. In the first one we show that if there exists a classical solution which is bounded in each strip $[s,T]\times\mathbb R^d$, then it is unique. In the second part we construct a classical solution which satisfies \eqref{stima_uni_gen}. {\it Uniqueness}. Let ${\bm u}$ be a classical solution to \eqref{pioggia} with $\bm f=\bm 0$, which is bounded in each strip $[s,T]\times\mathbb R^d$, and let us define the function ${\bm w}=(w_1,\ldots,w_m)$ by setting $w_k=u_k^2$ for $k=1,\ldots,m$. Note that \begin{align} D_tw_k=2u_kD_tu_k={\mathcal A}_{k,0}w_k-2\langle Q_k\nabla u_k,\nabla u_k\rangle+2u_k(Cu)_k,\qquad\;\,k=1,\ldots,m, \label{carboni} \end{align} in $(s,T]\times\mathbb R^d$, where ${\mathcal A}_{k,0}=\sum_{i,j=1}^dq_{ij}^kD^2_{ij}+\sum_{i=1}^d b_i^kD_i$. Let us estimate the last term in the last side of \eqref{carboni}. For this purpose, we observe that \begin{align*} 2u_k(t,x)\sum_{j=1}^m c_{kj}(t,x)u_j(t,x)\leq & 2c_{kk}(t,x)|u_k(t,x)|^2+2\sum_{j\neq k}|c_{kj}(t,x)||u_k(t,x)| |u_j(t,x)| \\ \leq & 2c_{kk}(t,x)|u_k(t,x)|^2+\sum_{j\neq k}|c_{kj}(t,x)|(|u_k(t,x)|^2+ |u_j(t,x)|^2) \\ = & |u_k(t,x)|^2\bigg (c_{kk}(t,x)+\sum_{j\neq k}|c_{kj}(t,x)|\bigg ) \\ &+ \bigg (c_{kk}(t,x)|u_k(t,x)|^2+\sum_{j\neq k} |c_{kj}(t,x)||u_j(t,x)|^2 \bigg )\\ = & |u_k(t,x)|^2\sum_{j=1}^mc^P_{kj}(t,x) +\sum_{j=1}^mc^P_{kj}(t,x)|u_j(t,x)|^2 \\ \leq & Mw_k(t,x)+ \sum_{j=1}^m c^P_{kj}(t,x)w_j(t,x), \end{align*} which, combined with \eqref{carboni}, gives $D_t{\bm w}-(\bm{\mathcal A}^P+M{\rm Id}){\bm w} \leq \bm 0$ in $(s,T]\times\mathbb R^d$. Again the maximum principle in \cite[Theorem 2.3]{AngLor20}, applied to the operator $D_t-(\bm{\mathcal A}^P+M{\rm Id})$, implies that ${\bm w}\leq \bm 0$, i.e., ${\bm u}\equiv\bm 0$. {\it Existence}. We recall that for every $n\in{\mathbb N}$ the function ${\bm G}_n^{\mathcal N}(s,t)\bm f$ is the unique classical solution to the Cauchy problem with homogeneous Neumann boundary conditions \begin{equation*} \left\{ \begin{array}{lll} D_t{\bm u}(t,x)=(\bm{\mathcal A}{\bm u})(t,x), & t\in(s,\infty), &x\in B_n, \vspace{1mm}\\ \displaystyle \frac{\partial {\bm u}}{\partial \nu}(t,x)=\bm 0, & t\in(s,\infty), & x\in\partial B_n, \vspace{1mm} \\ {\bm u}(s,x)=\bm f(x), && x \in B_n, \end{array} \right. \end{equation*} where $\nu$ is the outward unit normal to $B_n$. Arguing as in the proof of \cite[Theorem 2.4]{AngLor20} (see also Remark 2.5 therein and Proposition \ref{prop:app_stime_varie}$(i)$-$(ii)$ in the Appendix) it follows that, for every compact set $E\subset (s,\infty)\times \mathbb R^d$, the sequence $({\bm G}_n^{\mathcal N}(s,t)\bm f)$ converges to a function ${\bm u}$ in $C^{1,2}(E;\mathbb R^m)$, which is a classical solution to \eqref{pioggia}. Proposition \ref{prop:confronto_1} shows that $|({\bm G}_n^{\mathcal N}(\cdot,s)\bm f)_j|\leq ({\bm G}_n^{\mathcal N, P}(\cdot,s)|\bm f|)_j$ in $[s,\infty)\times\mathbb R^d$ for every $j\in\{1,\ldots,m\}$ and $n\in{\mathbb N}$. By letting $n$ tend to $\infty$ we deduce that ${\bm G}_n^{\mathcal N,P}(\cdot,s)|\bm f|$ converges to $\widetilde {\bm u}^P$ in $C^{1,2}(E;\mathbb R^m)$ for every compact set $E\subset(s,\infty)\times \mathbb R^d$. From \eqref{tre} we get \begin{align*} \max_{k=1,\ldots,m}|u_k(t,x)| = & \max_{k=1,\ldots,m}\lim_{n\rightarrow\infty} |(({\bm G}_n^{\mathcal N}(s,t)\bm f)(x))_k| \leq \max_{k=1,\ldots,m}\lim_{n\rightarrow\infty}(({\bm G}_n^{\mathcal N, P}(\cdot,s)|\bm f|)(x))_k \\ =& \max_{k=1,\ldots,m}|\widetilde u_k^P(t,x)|\leq e^{M(t-s)}\max_{k=1,\ldots,m}\|f_k\|_\infty \end{align*} for every $(t,x)\in [s,T]\times \mathbb R^d$. This completes the proof. \end{proof} For further use, we consider the following results. \begin{prop} \label{pro:appr_sol_neumann} In addition to Hypothesis $\ref{hyp-base}(i)$-$(ii)$, assume that the coefficients of the operator $\bm{\mathcal A}$ belong to $C^{\alpha/2,3+\alpha}_{\rm loc}(I\times\mathbb R^d)$. Then, the following properties are satisfied for every $n\in{\mathbb N}$. \begin{enumerate}[\rm (i)] \item for every $\bm f\in C(\overline{B_n};{\mathbb R}^m)$, the function $D^{\beta}_x{\bm G}^{\mathcal N}_n(\cdot,s)\bm f$ belongs to $C_{\rm loc}^{1+\alpha/2,2+\alpha}((s,T)\times B_n;\mathbb R^m)$ for every $\beta\in({\mathbb N}\cup\{0\})^d$, with $|\beta|\le 3$, and $n\in{\mathbb N}$; \item the function $(t,x)\mapsto (t-s)^{(3-j)/2}|(D^3_x{\bm G}^{\mathcal N}_n(\cdot,s)\bm f)(x)|$ is continuous in $[s,T]\times B_n$ for every $n\in{\mathbb N}$, $\bm f\in C^j(\overline{B_n};{\mathbb R}^m)$ and $j=0,1,2,3$, where we extend it at $t=s$ in the trivial way; \item for each $s<t_2<t_1<T$, each pair of bounded sets $\Omega_1\Subset \Omega_2\Subset B_R$ and $n>R$ there exists a positive constant $C$, independent of $n$ and $\bm f\in C(\overline{B_n})$, such that $\|{\bm G}^{\mathcal N}_n(t,s)\bm f\|_{C^{1+\alpha/2,3+\alpha}((t_1,T)\times \Omega_1;\mathbb R^m)}\leq C\|{\bm G}^{\mathcal N}_n(t,s)\bm f\|_{L^\infty((t_2,T)\times \Omega_2;\mathbb R^m)}$. In particular, ${\bm u}\in C^{1+\alpha/2,3+\alpha}_{\rm loc}((s,T)\times \mathbb R^d;\mathbb R^m)$ and ${\bm G}^{\mathcal N}_n(\cdot,s)\bm f$ converges to ${\bm u}$ in $C^{1+\alpha/2,3+\alpha}(E)$ as $n$ tends to $\infty$ for every compact set $E\subset(s,T)\times \mathbb R^d$, where ${\bm u}$ is the solution to the Cauchy problem \eqref{pioggia} provided by Proposition $\ref{prop-2.6}$. \end{enumerate} \end{prop} \begin{proof} (i) This property follows from \cite[Section II.4, Theorem 3.2]{Ei69}. (ii) This property follows arguing as in \cite[Theorem 2.3]{BerLor05} and taking \cite[Chapter IV, Theorem 5.3]{LadSolUra68Lin} into account. (iii) To simplify the notation, we set ${\bm u}_n=G_n^{\mathcal N}(\cdot,s)\bm f$. Let us fix $s<t_2<t_3<t_1<T$ and three bounded open sets $\Omega_1\Subset\Omega_3\Subset\Omega_2$. Further, fix $R>0$ such that $\Omega_2\subset B_R$ and $n>R$. From property (i) we know that $D_{\ell}{\bm u}_n\in C^{1+\alpha/2,2+\alpha}((s,T)\times B_n;\mathbb R^m)$ for every $\ell=1,\ldots,d$. Differentiating along the direction $e_\ell$, $\ell=1,\ldots,d$, the equation $D_t{\bm u}_n=\bm{\mathcal A}{\bm u}_n$, we get $D_tD_\ell{\bm u}_n=\bm{\mathcal A} D_\ell{\bm u}_n+{\bm g}$ in $(s,T)\times B_n$, where \begin{align*} g_k=\sum_{i,j=1}^dD_{\ell}q_{ij}^kD_{ij}u_k+\sum_{i=1}^dD_{\ell}b_i^kD_iu_k+\sum_{j=1}^m D_{\ell}c_{kj}u_j,\qquad\;\,k=1,\ldots,m. \end{align*} By applying \cite[Theorem 7.2]{AngLor20} we infer that there exists a positive constant $C_1$ such that \begin{align*} \|D_{\ell}{\bm u}_n\|_{C^{1+\alpha/2,2+\alpha}((t_1,T)\times \Omega_1;\mathbb R^m)}\leq C_1\big(\|D_{\ell}{\bm u}_n\|_{L^\infty((t_3,T)\times\Omega_3;\mathbb R^m)}+\|{\bm g}\|_{C^{\alpha/2,\alpha}(t_3,T)\times \Omega_3;\mathbb R^m)}\big) \end{align*} and from the definition of ${\bm g}$, the assumptions on the coefficients of the operator $\bm{\mathcal A}$ and applying once again \cite[Theorem 7.2]{AngLor20}, we infer that \begin{align*} \|D_{\ell}{\bm u}_n\|_{L^\infty((t_3,T)\times\Omega_3;\mathbb R^m)}+\|{\bm g}\|_{C^{\alpha/2,\alpha}(t_3,T)\times \Omega_3;\mathbb R^m)}\leq C_2\|{\bm u}_n\|_{L^\infty({t_2},T]\times \Omega_2;\mathbb R^m)} \end{align*} for some positive constant $C_2$, which is independent of $n$ as the constant $C_1$. It follows that there exists a positive constant $C$ such that \begin{align*} \|{\bm u}_n\|_{C^{1+\alpha/2,3+\alpha}((t_1,T)\times \Omega_1;\mathbb R^m)}\leq C_3 \|{\bm u}_n\|_{L^\infty(({t_2},T]\times \Omega_2;\mathbb R^m)} \end{align*} for some constant $C_3>0$, independent of $n$. Since ${\bm u}_n$ converges to ${\bm u}$ in $C^{1+\alpha/2,2+\alpha}(E;\mathbb R^m)$ as $n$ tends to $\infty$ for each compact set $E\subset (s,T]\times \mathbb R^d$, writing this estimate with ${\bm u}_n$ being replaced by ${\bm u}_n-{\bm u}_m$, it follows that $({\bm u}_n)$ is a Cauchy sequence in $C^{1+\alpha/2,3+\alpha}((t_1,T)\times \Omega_1;\mathbb R^m)$ and the arbitrariness of $\Omega_1$ and $t_1$ implies that ${\bm u}\in C_{\rm loc}^{1+\alpha/2,3+\alpha}((s,T)\times \mathbb R^d;\mathbb R^m)$ and ${\bm u}_n$ converges to ${\bm u}$ in $C^{1+\alpha/2,3+\alpha}(E;\mathbb R^m)$ as $n$ tends to $\infty$ for every compact set $E$ as above. \end{proof} In view of Proposition \ref{prop-2.6}, we can define an evolution operator ${\bm G}(t,s)$ on $C_b({\mathbb R}^d;{\mathbb R}^m)$ by setting, for every $\bm f\in C_b({\mathbb R}^d;{\mathbb R}^m)$, ${\bm G}(t,s)\bm f:={\bm u}(t,\cdot)$ if $t>s$ and ${\bm G}(s,s)\bm f:=\bm f$, where ${\bm u}$ is the unique classical solution to \eqref{pioggia} with ${\bm u}(s,\cdot)=\bm f$, which is bounded in each strip $[s,T]\times\mathbb R^d$. The uniqueness part of Proposition \ref{prop-2.6} implies that ${\bm G}(t,r){\bm G}(r,s)\bm f={\bm G}(t,s)\bm f$ for every $t\geq r\geq s\in I$ and $\bm f\in C_b({\mathbb R}^d;{\mathbb R}^m)$. Similarly, we define the evolution operator ${\bm G}^P(t,s)$ on $C_b({\mathbb R}^d;{\mathbb R}^m)$ associated with the parabolic Cauchy problem \eqref{cauchy_prob_pos}. In the following proposition, we prove some basic properties of the evolution operator ${\bm G}(t,s)$. \begin{prop} \label{prop:conv_loc_unif} Fix $s\in I$ and a bounded sequence $(\bm f_n)\subset C_b(\mathbb R^d;\mathbb R^m)$. Then, the following properties are satisfied. \begin{enumerate}[\rm (i)] \item If $\bm f_n$ pointwise converges to $\bm f\in C_b(\mathbb R^d;\mathbb R^m)$, then ${\bm G}(\cdot,s)\bm f_n$ converges to ${\bm G}(\cdot,s)\bm f$ in $C^{1,2}(E;\mathbb R^m)$ for every compact set $E\subset (s,\infty)\times \mathbb R^d$; \item If $\bm f_n$ locally uniformly converges to $\bm f\in C_b(\mathbb R^d;\mathbb R^m)$, then ${\bm G}(\cdot,s)\bm f_n$ converges to ${\bm G}(\cdot,s)\bm f$ locally uniformly in $[s,\infty)\times \mathbb R^d$. \end{enumerate} \end{prop} \begin{proof} The proof follows from \cite[Theorem 2.6]{AngLor20} recalling that \begin{align*} |({\bm G}(t,s)\bm f_n)(x)-({\bm G}(t,s)\bm f)(x)|\le ({\bm G}^P(t,s)|\bm f_n-\bm f|)(x), \qquad\;\, t>s, \;\, x\in\mathbb R^d.\hskip 3cm\qedhere \end{align*} \end{proof} \begin{rmk} \label{rmk-claudio} {\rm We stress that all the results in this section can be extended to the autonomous case when the nonautonomous operator $\bm{\mathcal A}$ is replaced with the autonomous operator $\bm{\mathcal A}(\overline t)$, i.e., when $\overline t\in I$ is frozen. In particular, for every $\overline t\in I$ we can introduce in the natural way the semigroups ${\bm T}_{\overline t}(\tau)$ and ${\bm T}_{\overline t}^P(\tau)$, associated, in $C_b(\mathbb R^d;\mathbb R^m)$, to the operators $\bm{\mathcal A}(\overline t)$ and $\bm{\mathcal A}^P(\overline t)$, respectively. Moreover, \begin{align} &\max\{\|{\bm T}_{\overline{t}}(\tau)\|_{L(C_b(\mathbb R^d;\mathbb R^m))},\|{\bm T}_{\overline{t}}^P(\tau)\|_{L(C_b(\mathbb R^d;\mathbb R^m))}\} \le e^{M_{\overline{t}}\tau}\max_{k=1,\ldots,m}\|f_k\|_{\infty} \label{stima_uni_gen_auto} \end{align} for every $\bm f\in C_b(\mathbb R^d;\mathbb R^m)$ and $\tau>0$, where the constant $M_{\overline{t}}$ is defined by \eqref{def_M_T}, with $J=\{\overline t\}$.} \end{rmk} \section{Uniform estimates of the spatial derivatives} \label{section:grad_est} Throughout this section, we fix $T>s\in I$. We introduce the nonautonomous operator $\widetilde{\bm{\mathcal A}}^P$, defined, componentwise on smooth functions $\bm f:\mathbb R^d\to{\mathbb R}^m$ by \begin{align} \label{operatore_A_tilde} (\widetilde{\bm{\mathcal A}}^P\bm f)_k={\rm Tr}(Q^kD^2f_k)+\langle b^k,\nabla f_k\rangle+\widetilde C^P\bm f)_k \end{align} for every $t\in[s,T]$ and $k=1,\ldots,m$, where $\widetilde c_{kj}^P(t,x):= c_{kj}^P(t,\cdot)+m^{-1}(1+|M_k(t,\cdot)|)$ for every $(t,x)\in [s,T]\times\mathbb R^d$ and $k,j=1,\ldots,m$. Here, $M_k$ is the function in \eqref{def_M_T}. \begin{rmk} {\rm We notice that $\sum_{j=1}^m\widetilde c_{kj}^P(t,x) =1+2(M_k(t,x))^+$ for every $(t,x)\in [s,T]\times\mathbb R^d$, where $(M_k(t,x))^+$ denotes the positive part of $M_k(t,x)$ (see \eqref{def_M_T}). Hence, the potential matrix of the operator $\widetilde {\bm{\mathcal A}}^P$ satisfies Hypothesis {\ref{hyp-base}$(v)$}: indeed, \begin{align} \label{stima_somma_tilde_c} \max _{k=1,\ldots,m}\sup_{(t,x)\in[s,T]\times \mathbb R^d}\sum_{j=1}^m\widetilde c_{kj}^P(t,x)\leq 1+2M^+. \end{align} } \end{rmk} \begin{hyp} \label{hyp:smoothness_coeff} \begin{enumerate}[\rm (i)] \item The coefficients of the operator $\bm{\mathcal A}$ belong to $C^{\alpha/2,3+\alpha}_{\rm loc}(I\times\mathbb R^d)$; \item there exists a function $r^k:[s,T]\times\mathbb R^d\to{\mathbb R}$ such that $\langle ({\rm J}_x\bm b^k)\xi,\xi\rangle\le r^k|\xi|^2$ on $[s,T]\times\mathbb R^d$ for every $\xi\in\mathbb R^d$; \item there exists a positive constant $C$ such that \begin{align*} \max\{|(Q^{k}(t,x)x)_i|, |{\rm Tr}(Q^{k}(t,x))|, \langle \bm b^k(t,x),x\rangle\}\leq C(1+|x|^2)\mu^k(t,x) \end{align*} for every $i=1,\ldots,d$, $k=1,\ldots,m$ and $(t,x)\in [s,T]\times \mathbb R^d$; \item there exist positive functions $b^{k,2}$, $b^{k,3}$ and a positive constant $L_k$ such that \begin{align*} |D^h_{x}b^k_j|\le b^{k,h},\qquad\;\,|\nabla_{x} q^k_{ij}|\le L_k\mu^k,\qquad\;\, |D^h_{x}q^k_{ij}|\le L_k\mu^k \end{align*} on $(s,T]\times\mathbb R^d$ for every $i,j=1,\ldots,d$, $h=2,3$, and $r^k+B_{k,2}b^{k,2}+B_{k,3}b^{k,3}\le {\mathscr M}_k\mu^k$ on $(s,T]\times\mathbb R^d$ for every $k=1,\ldots,m$ and some positive constants $B_{k,2}$, $B_{k,3}$ and ${\mathscr M}_k$; \item there exist positive functions $c^{k,h}$ such that $|D^j_xc_{k\ell}|\le c^{k,j}$ for every $k,\ell=1,\ldots,m$ and $h=1,\ldots,3$ and $c^{k,h}\le \overline{C}(1+|M_k|)$ for every $h,k$ and some positive constant $\overline C$. \end{enumerate} \end{hyp} The following one is the main result of this section. \begin{thm} \label{thm-3.4} Let Hypotheses $\ref{hyp-base}$ and $\ref{hyp:smoothness_coeff}$ be fulfilled. Then, \begin{enumerate}[\rm (i)] \item for each $h,k=0,\ldots,3$, with $h\leq k$ and $T>s\in I$, there exists a constant $C=C_{h,k}(s,T)>0$ such that \begin{align} \label{stima_der_op_ev_totale} \|{\bm G}(t,s)\bm f\|_{C_b^k(\mathbb R^d;\mathbb R^m)}\leq \frac{Ce^{\overline M (t-s)}}{(t-s)^{(k-h)/2}}\|\bm f\|_{C^h_b(\mathbb R^d;\mathbb R^m)} \end{align} for every $\bm f\in C^h_b(\mathbb R^d;\mathbb R^m)$ and $t\in(s,T]$, where $\overline M:=2^{-1}(1+M+2M^+)$; \item for every $\mathcal T>0$, $\overline t\in [s,T]$ and $h,k=0,\ldots,3$, with $h\leq k$, there exists a constant $C=C_{h,k}(\overline t,\mathcal T)$, which is uniform with respect to $\overline t$ belonging to bounded sets, such that \begin{align} \label{stima_der_smgr_totale} \|{\bm T}_{\overline t}(\tau)\bm f\|_{C_b^k(\mathbb R^d;\mathbb R^m)}\leq \frac{Ce^{\overline M \tau}}{\tau^{(k-h)/2}}\|\bm f\|_{C^h_b(\mathbb R^d;\mathbb R^m)}, \end{align} for every $\bm f\in C^h_b(\mathbb R^d;\mathbb R^m)$ and $\tau\in (0,\mathcal T]$, where in this case $\overline M:=2^{-1}(1+M_{\{\overline t\}}+2(M_{\{\overline t\}})^+)$. \end{enumerate} \end{thm} \begin{proof} (i) Let us first consider the case $h=0$ and $k=3$. We fix $\bm f\in C_b(\mathbb R^d;\mathbb R^m)$ and for every $n\in{\mathbb N}$ we denote by $\widetilde {\bm u}_n$ be the unique classical solution to the Neumann-Cauchy problem associated to the operator $\bm{\mathcal A}$ in $(s,T]\times B_n$, such that $\widetilde{\bm u}_n(s,\cdot)=\vartheta_n\bm f$, where $\vartheta_n(x)=\varphi(n^{-1}|x|)$ for every $x\in\mathbb R^d$ and $\varphi\in C^\infty([0,\infty))$ is a nonincreasing function such that $\varphi\equiv1$ on $[0,1/2)$ and $\varphi\equiv 0$ on $[3/4,\infty)$. From Proposition \ref{pro:appr_sol_neumann}(i) it follows that the function $\widetilde {\bm u}_n$ has spatial derivatives up to the third-order which belong to $C^{1+\alpha/2,2+\alpha}_{\rm loc}((s,T]\times B_n;{\mathbb R}^m)$ for every $n\in{\mathbb N}$. In the following, to lighten the notation we do not stress explicitly the dependence on $n$ and we simply write $\widetilde {\bm u}$ and $\vartheta$. Let us introduce the function ${\bm v}=(v_1,\ldots,v_m)$, defined componentwise by \begin{align*} v_k(t,x)=&(\widetilde u_k(t,x))^2+\alpha(t-s)|\nabla_{x} \widetilde u_k(t,x)|^2+\alpha^2(t-s)^2(\vartheta(x))^2|D^2_{x}\widetilde u_k(t,x)|^2\\ &+\alpha^3(t-s)^3(\vartheta(x))^4|D^3_{x}\widetilde u_k(t,s)|^2\\ =&\!: w_k(t,x)+\alpha(t-s)w_k^1(t,x)+\alpha^2(t-s)^2(\vartheta(x))^2w_k^2(t,x)\\ &+\alpha^3(t-s)^3(\vartheta(x))^4w_k^3(t,x), \end{align*} for each $(t,x)\in(s,T]\times B_n$, where $\alpha\in(0,1]$ will be fixed later. Note that $\displaystyle \frac{\partial v_k}{\partial \nu}\leq 0$ on $(s,T]\times \partial B_n$ for every $k\in\{1,\ldots,m\}$. Indeed, $\displaystyle\frac{\partial w_k}{\partial \nu}=2\widetilde u_k\frac{\partial \widetilde u_k}{\partial \nu}=0$ on $(s,T]\times \partial B_n$ and the normal derivative of $w_k^1$ is nonpositive since $B_n$ is convex (see \cite[Lemma 2.4]{BerFor04}). By Proposition \ref{pro:appr_sol_neumann}(ii), ${\bm v}_n$ is continuous on $[s,\infty)\times B_n$. In the following computations, we do not stress the dependence on $t$ and $x$ when it is not really needed, and we stress that the all the constants which appear depend neither on $n$ nor $x$, but depend on the interval $[s,T]$. A long but straightforward computation reveals that \begin{align*} D_tv_k=& \alpha w^1_k+2\vartheta^2\alpha^2(t-s)w^2_k+3\vartheta^4\alpha^3(t-s)^2w^3_k+{\mathscr I}_{\bm{\mathcal A},k}+{\mathscr I}_{Q,1,k}+{\mathscr I}_{Q,2,k}+{\mathscr I}_{Q,3,k}\\ &+{\mathscr I}_{{\bm b},1,k} +{\mathscr I}_{{\bm b},2,k}+{\mathscr I}_{C,1,k} +{\mathscr I}_{C,2,k}+{\mathscr I}_{Q,{\bm b},3} \end{align*} where \begin{align*} {\mathscr I}_{\bm{\mathcal A},k}=&2\widetilde u_k(\bm{\mathcal A}\widetilde{\bm u})_k+2\alpha(t-s)\sum_{h=1}^dD_h\widetilde u_k(\bm{\mathcal A} D_h\widetilde{\bm u})_k+2\vartheta^2\alpha^2(t-s)^2\sum_{h,p=1}^dD_{hp}\widetilde u_k(\bm{\mathcal A} D_{hp}\widetilde{\bm u})_k\\ &+2\vartheta^4\alpha^3(t-s)^3\sum_{h,p,r=1}^dD_{hpr}\widetilde u_k(\bm{\mathcal A} D_{hpr}\widetilde{\bm u})_k,\\ {\mathscr I}_{Q,1,k}=&2\alpha(t-s)\sum_{i,j,h=1}^dD_hq_{ij}^kD_h\widetilde u_kD_{ij}\widetilde u_k +4\vartheta^2\alpha^2(t-s)^2\sum_{i,j,h,p=1}^dD_hq_{ij}^k D_{hp}\widetilde u_k D_{ijp}\widetilde u_k\\ &+6\vartheta^4\alpha^3(t-s)^3\sum_{i,j,h,p,r=1}^dD_rq_{ij}^kD_{hpr}\widetilde u_k D_{ijhp}\widetilde u_k,\\ {\mathscr I}_{Q,2,k}=&2\vartheta^2\alpha^2(t-s)^2\sum_{i,j,h,p=1}^dD_{hp}q_{ij}^kD_{ij}\widetilde u_kD_{hp}\widetilde u_k\\ &+6\vartheta^4\alpha^3(t-s)^3\sum_{i,j,h,p,r=1}^dD_{pr}q_{ij}^kD_{hpr}\widetilde u_kD_{ijh}\widetilde u_k,\\ {\mathscr I}_{{\bm b},1,k}=&2\alpha(t-s)\langle (J_x\bm b^k)\nabla_x \widetilde u_k,\nabla_x\widetilde u_k\rangle +4\vartheta^2\alpha^2(t-s)^2\sum_{p=1}^d\langle (J_x\bm b^k)\nabla_x D_p\widetilde u_k,\nabla_x D_p\widetilde u_k\rangle\\ &+6\vartheta^4\alpha^3(t-s)^3\sum_{p,r=1}^d\langle (J_x\bm b^k)\nabla_x D_{pr}\widetilde u_k,\nabla_x D_{pr}\widetilde u_k\rangle,\\ {\mathscr I}_{{\bm b},2,k}=&2\vartheta^2\alpha^2(t-s)^2 \sum_{j,h,p=1}^dD_{hp}b_j^kD_j\widetilde u_kD_{hp}\widetilde u_k \\ &+6\vartheta^4\alpha^3(t-s)^3\sum_{j,h,p,r=1}^dD_{pr}b_j^k D_{jh}\widetilde u_k D_{hpr}\widetilde u_k,\\ {\mathscr I}_{C,1,k}=&2\alpha(t-s)\sum_{\ell=1}^m\sum_{h=1}^dD_hc_{k\ell} \widetilde u_{\ell} D_h\widetilde u_k +4\vartheta^2\alpha^2(t-s)^2\sum_{\ell=1}^m\sum_{h,p=1}^dD_hc_{k\ell} D_p\widetilde u_{\ell} D_{hp}\widetilde u_k\\ &+6\vartheta^4\alpha^3(t-s)^3\sum_{\ell=1}^m\sum_{h,p,r=1}^dD_hc_{k\ell} D_{pr}\widetilde u_{\ell} D_{hpr}\widetilde u_k,\\ {\mathscr I}_{C,2,k}= & 2\vartheta^2\alpha^2(t-s)^2\sum_{\ell=1}^m\sum_{h,p=1}^d\!\!D_{hp}c_{k\ell} \widetilde u_{\ell} D_{hp}\widetilde u_k \\ & +6\vartheta^4\alpha^3(t-s)^3\sum_{\ell=1}^m\sum_{h,p,r=1}^d\!\!D_{hp}c_{k\ell}D_r\widetilde u_{\ell} D_{hpr}\widetilde u_k,\\ {\mathscr I}_{Q,{\bm b},C,3}=&2\vartheta^4\alpha^3(t-s)^3\bigg (\sum_{i,j,h,p,r=1}^dD_{hpr}q_{ij}^k D_{ij}\widetilde u_k D_{hpr}\widetilde u_k +\sum_{j,h,p,r=1}^dD_{hpr}b_j^k D_j\widetilde u_k D_{hpr}\widetilde u_k\\ &\phantom{+2\vartheta^4\alpha^3(t-s)^3\bigg (\;\;\,} +\sum_{\ell=1}^m\sum_{h,p,r=1}^dD_{hpr}c_{k\ell}\widetilde u_{\ell} D_{hpr}\widetilde u_k\bigg ). \end{align*} Since for every $\bm f\in C^1(B_n;\mathbb R^m)$ it holds that $f_k(\bm{\mathcal A}\bm f)_k\leq (\bm{\mathcal A}^P{\bm g})_k+Mg_k-2\mu^k\|\nabla f_k\|^2$, where $g_k:=f_k^2$ for every $k=1,\ldots,m$ (see the proof of Proposition \ref{prop-2.6}(ii)), applying this formula to $\widetilde u_k$ and its derivatives, and setting $w_k^4=\displaystyle\sum_{i,h,p,r=1}^d|D_{ihpr}\widetilde u_k|^2$, it follows that \begin{align*} {\mathscr I}_{\bm{\mathcal A},k} \le & g_k+(\bm{\mathcal A}^p\bm w)_k+Mw_k-2\mu^kw_k^1 +\alpha(t-s)[(\bm{\mathcal A}^P\bm w^1)_k+Mw_k^1-2\mu^kw^2_k]\\ &+\sum_{i=2}^3\vartheta^{2i-2}\alpha^i(t-s)^i[(\bm{\mathcal A}^P\bm w^i)_k+Mw^i_k-2\mu^kw^{i+1}_k]\\ \le & (\bm{\mathcal A}^P\bm v)_k+M v_k -2\mu^k\sum_{i=1}^4\alpha^{i-1}(t-s)^{i-1}w_k^i, \end{align*} where \begin{align*} g_k= & -2\vartheta\alpha^2(t-s)^2[w_k^2+2\vartheta^2\alpha(t-s)w_k^3]{\rm Tr}(Q^kD^2\vartheta) -4\alpha^2(t-s)^2\vartheta\langle Q^k\nabla_{x} w^2_k,\nabla\vartheta\rangle\\ &-2\vartheta\alpha^2(t-s)^2[w_k^2+2\vartheta^2\alpha(t-s)w_k^3]\langle b^k,\nabla \vartheta\rangle -8\alpha^3(t-s)^3\vartheta^3\langle Q^k\nabla_{x} w^3_k,\nabla\vartheta\rangle. \end{align*} Taking Hypothesis \ref{hyp:smoothness_coeff}$(iii)$ into account, we get $\max\{|{\rm Tr}[Q^kD^2\vartheta]|, |Q^k\nabla \vartheta|, -\langle b^k, \nabla \vartheta\rangle\} \leq C\mu^k$ for some positive constant $C$, which does not depend on $n$. By applying the Young inequality $2ab\leq \varepsilon a^2+\varepsilon ^{-1}b^2$ for every $a,b,\varepsilon>0$, we can estimate \begin{align*} g_k\leq & C\alpha^2(t-s)^2\mu^k(\vartheta+\sqrt{d}\varepsilon^{-1})w_k^2+C\alpha^2(t-s)^2\vartheta^2\mu^k[\sqrt{d}\varepsilon+\alpha(t-s)(\vartheta+\sqrt{d}\varepsilon^{-1})]w_k^3\\ &+C\sqrt d\varepsilon \alpha^3(t-s)^3\vartheta^4\mu^k w_k^4 \end{align*} for some positive constant $C$ and every $\varepsilon>0$. Hence, taking $\varepsilon=\sqrt{\alpha}$, we get \begin{align*} {\mathscr I}_{\bm{\mathcal A},k}\leq & (\bm{\mathcal A}^P\bm v)_k+M v_k-2\mu^k[w^1_k+\alpha(t-s)w_k^2+\alpha^2(t-s)^2\vartheta^2w^3_k+\alpha^3(t-s)^3\vartheta^4w_k^4]\\ &+C\alpha^{\frac{3}{2}}(t-s)^2\mu^k(\sqrt{\alpha}\vartheta+\sqrt{d})w_k^2+C\alpha^{\frac{5}{2}}(t-s)^2\vartheta^2\mu^k[\sqrt{d}+(t-s)(\sqrt{\alpha}\vartheta+\sqrt{d})]w_k^3\\ &+C\sqrt d\alpha^{\frac{7}{2}}(t-s)^3\vartheta^4\mu^k w_k^4. \end{align*} Now, we observe that, since $\langle (J_x\bm b^k)\xi,\xi\rangle\le r^k|\xi|^2$ for every $\xi\in\mathbb R^d$, we can estimate \begin{align*} {\mathscr I}_{{\bm b},1,k} \le &2\alpha(t-s)r^kw_k^1+4\alpha^2(t-s)^2\vartheta^2r^kw_k^2+6\alpha^3(t-s)^3r^k\vartheta^4w_k^3. \end{align*} Using Cauchy-Schwarz inequality and the assumption $\|\nabla_{x} q_{ij}^k\|\le L_k\mu^k$ for every $i,j=1,\ldots,d$, we can estimate \begin{align*} 2\sum_{i,j,h=1}^dD_hq_{ij}^kD_hf_kD_{ij}f_k\le &2\sum_{i,j=1}^d\|\nabla_{x} q_{ij}\|\|\nabla f_k\|\|D_{ij}f_k\|\\ \le &2L_kd\mu^k\|\nabla f_k\|\|D^2_{x} f_k\|\le L_kd\mu^k\Big (\alpha^{-\frac{1}{2}}\|\nabla f_k\|^2+\alpha^{\frac{1}{2}}\|D^2 f_k\|^2\Big ) \end{align*} for every $\bm f\in C^2(B_n;{\mathbb R}^m)$. Applying this estimate to $\widetilde{\bm u}$ and its spatial derivatives, gives \begin{align*} {\mathscr I}_{Q,1,k}\le L_kd\mu^k\sum_{i=1}^3i\vartheta^{2i-2}\alpha^i(t-s)^i \Big(\alpha^{-\frac{1}{2}}w^i_k+\alpha^{\frac{1}{2}}w^{i+1}_k\Big ). \end{align*} Similarly, using the assumption $\|D^2_{x}q_{ij}^k\|\le L_k\mu^k$ and $\|D^2_{x}b_j\|\le b^{k,2}$ for every $i,j=1,\ldots,d$, we can estimate \begin{align*} {\mathscr I}_{Q,2,k}\le &L_kd\mu^k( 2\vartheta^2\alpha^2(t-s)^2w_k^2+6\vartheta^4\alpha^3(t-s)^3w_k^3)\\ {\mathscr I}_{{\bm b},2,k}\le & \sqrt{d}b^{k,2} \left[\vartheta^2\alpha^2(t-s)^2\Big(\alpha^{-\frac12}w^1_k+\alpha^{\frac12}w^2_k\Big)+\vartheta^4\alpha^3(t-s)^3\Big(\alpha^{-\frac12}w^2_k+\alpha^{\frac12}w^3_k\Big)\right]. \end{align*} Next, using the conditions $\|D_x^hc_{k\ell}\|\le c^{k,h}$ for every $k,\ell=1,\ldots,m$ and $h=1,2,3$, we can estimate \begin{align*} 2\sum_{\ell=1}^m\sum_{h=1}^dD_hc_{k\ell} f_{\ell} D_hf_k \le & 2\sqrt{m}c^{k,1}\|\bm f\|{\|\nabla f_k\|} \le \sqrt{m}c^{k,1}\left (\alpha^{-\frac{1}{2}}\|\bm f\|^2+\alpha^{\frac{1}{2}}\|J\bm f\|^2\right ), \end{align*} for every $\bm f\in C^1(B_n;{\mathbb R}^m)$. Applying this estimate to $\widetilde{\bm u}$ and its spatial derivatives we get \begin{align*} {\mathscr I}_{C,1,k} \leq & \sqrt m c^{k,1}\sum_{i=1}^3i\vartheta^{2i-2}\alpha^{i}(t-s)^{i}\bigg(\alpha^{-\frac{1}{2}}\sum_{\ell=1}^m w_\ell^i+\alpha^{\frac{1}{2}}\sum_{\ell=1}^mw_\ell^{i+1}\bigg ). \end{align*} Similarly, we estimate (with the convention that $w_\ell^0=w_\ell$ for every $\ell=1,\ldots,m$) \begin{align*} {\mathscr I}_{C,2,k}\le \sqrt m c^{k,2} \sum_{i=1}^2(2i-1)\vartheta^{2i}\alpha^{i+1}(t-s)^{i+1}\bigg(\alpha^{-\frac{1}{2}}\sum_{\ell=1}^mw_\ell^{i-1}+\alpha^{\frac{1}{2}}\sum_{\ell=1}^mw_\ell^{i+1}\bigg). \end{align*} Finally, using the conditions $\|D^3_{x}q^k_{ij}\|\le L_k\mu^k$, $\|D^3_{x}b^k_j\|\le b^{k,3}$ and $\|D^3_{x}c_{k\ell}\|\le c^{k,3}$ for every $j=1,\ldots,d$ and $\ell=1,\ldots,m$, we get \begin{align*} {\mathscr I}_{Q,{\bm b},C,3}\le \vartheta^4\alpha^3(t-s)^3\bigg [&\sqrt{d}b^{k,3}\left (\alpha^{-\frac{1}{2}}w^1_k+\alpha^{\frac{1}{2}}w^3_k\right ) +L_kd\mu^k\left (\alpha^{-\frac{1}{2}}w_k^2+\alpha^{\frac{1}{2}}w^3_k\right )\\ &\;+\sqrt{m}c^{k,3}\bigg (\alpha^{-\frac{1}{2}}\sum_{\ell=1}^mw_{\ell}+ \alpha^{\frac{1}{2}} \sum_{\ell=1}^mw_{\ell}^3\bigg )\bigg ]. \end{align*} Combining everything together, we conclude that \begin{align*} D_tv_k\leq & ((\bm{\mathcal A}^P+M{\rm Id})\bm v)_k+{\mathscr H}_1w_k^1+{\mathscr H}_2\alpha(t-s)w_k^2+{\mathscr H}_3\alpha^2(t-s)^2\vartheta^2 w_k^3\\ &+{\mathscr H}_4\alpha^3(t-s)^3\vartheta^4w_k^4+{\mathscr H}_5\sum_{j=1}^mw_j +{\mathscr H}_6\alpha(t-s)\sum_{j=1}^mw_j^1 +{\mathscr H}_7\alpha^2(t-s)^2\vartheta^2\sum_{j=1}^mw_j^2\\ &+{\mathscr H}_8\alpha^3(t-s)^3\vartheta^4\sum_{j=1}^mw_j^3, \end{align*} where \begin{align*} {\mathscr H}_1= & \alpha-2\mu^k+\sqrt{\alpha}L_kd\mu^k(t-s)+\alpha(t-s)[2r^k+\sqrt{\alpha}(t-s)\sqrt db^{k,2}+\alpha^{\frac{3}{2}}(t-s)^2\sqrt db^{k,3}], \\[1mm] {\mathscr H}_2= & 2\alpha-2\mu^k+L_kd\sqrt{\alpha}\mu^k+(t-s)[C\sqrt{\alpha}(\sqrt{\alpha}+\sqrt d) +2\sqrt{\alpha}L_kd]\mu^k\\ &+\alpha(t-s)\vartheta^2[4r^k+2L_kd\mu^k+\sqrt{\alpha}\sqrt db^{k,2}+\sqrt{\alpha}(t-s)(3\sqrt db^{k,2}+L_kd\mu^k)]\\[1mm] {\mathscr H}_3= & [3\alpha+(2\sqrt{\alpha}L_kd+C\sqrt d\sqrt{\alpha}-2)\mu^k]+\sqrt{\alpha}(t-s)\mu^k[C\sqrt{\alpha}+C\sqrt{d}+3L_kd]\\ &+\alpha(t-s)\vartheta^2[6r^k+(6L_kd+\sqrt{\alpha} L_kd)\mu^k+3\sqrt{\alpha}\sqrt db^{k,2}+\sqrt{\alpha}\sqrt{d}b^{k,3}]\\[1mm] {\mathscr H}_4 = & [(C\sqrt d+3L_kd)\sqrt{\alpha}-2]\mu^k, \\ {\mathscr H}_5 = & \sqrt{\alpha}(t-s)\sqrt{m}[c^{k,1}+\alpha(t-s)c^{k,2}+\alpha^2(t-s)^2c^{k,3}], \\ {\mathscr H}_6= & \sqrt{\alpha}\sqrt{m}[ c^{k,1}+2(t-s)c^{k,1}+3\alpha(t-s)^2c^{k,2}], \\ {\mathscr H}_7 = & \sqrt{\alpha}\sqrt{m}[2 c^{k,1}+c^{k,2}+3(t-s)c^{k,1}], \\ {\mathscr H}_8 = & \sqrt{\alpha}\sqrt{m}(3c^{k,1}+3c^{k,2}+c^{k,3}). \end{align*} Let us separately estimate these terms. From Hypothesis \ref{hyp:smoothness_coeff} we can take $\alpha$ small enough such that $2r^k+\sqrt{\alpha}(t-s)\sqrt db^{k,2}+\alpha^{\frac{3}{2}}(t-s)^2\sqrt db^{k,3}\le 2{\mathscr M}_k\mu^k$, where $\mathscr M_k$ has been defined in Hypothesis \ref{hyp:smoothness_coeff}$(iv)$, so that \begin{align*} {\mathscr H}_1 \leq & \alpha-2\mu^k+\sqrt{\alpha}L_kd\mu^k(T-s)+2\alpha(T-s){\mathscr M}_k\mu^k. \end{align*} Hence, taking $\alpha$ smaller, if needed, we can make ${\mathscr H}_1$ negative on $(s,T]\times B_n$. Arguing similarly, we can make ${\mathscr H}_j$ ($j=2,3,4$) negative choosing $\alpha$ sufficiently small. As far as the remaining coefficients ${\mathscr H}_j$ are concerned, we observe that we can make each of them smaller than $m^{-1}(1+|M_k|)$, choosing $\alpha$ sufficiently small. This implies that $D_tv_k \leq ((\widetilde {\bm{\mathcal A}}^P+M{\rm Id}){\bm v})_k$ on $(s,T]\times B_n$, where $\widetilde {\bm{\mathcal A}}^P$ is the operator defined in \eqref{operatore_A_tilde}. \begin{comment} Let us notice that \begin{align*} M_k^P(t,x)+1+|M_k^P(t,x)| = \begin{cases} 1, & M_k^P(t,x)\leq 0, \\ 1+2M_k^P(t,x), & M_k^P(t,x)>0, \end{cases} \end{align*} for every $(t,x)\in [s,T]\times B_n$. Hence, \begin{align*} D_tv_k \leq (({\bm{\mathcal A}}^P+(M_k^P+1+|M_k^P|)Id){\bm v})_k\leq (({\bm{\mathcal A}}^P+\overline M Id){\bm v})_k, \end{align*} in $(s,T]\times B_n$, where $\overline M:=1+2M^+$. \end{comment} Now, we consider the function $\widetilde {\bm v}$, defined by $\widetilde {\bm v}(t,x)={\bm v}(t,x)-e^{2\overline M(t-s)}\|\bm f\|_\infty^2\mathds 1$ for every $(t,x)\in[s,T]\times B_n$, and observe that \begin{align*} D_t\widetilde v_k \leq & ((\widetilde{\bm{\mathcal A}}^P+M Id){\bm v})_k -2\overline Me^{2\overline M(t-s)}\|\bm f\|_\infty^2 \\ = & ((\widetilde {\bm{\mathcal A}}^P+M Id)\widetilde {\bm v})_k+\bigg (\sum_{h=1}^m\widetilde c_{kh}^P-(1+2M^+)\bigg )e^{2\overline M(t-s)}\|\bm f\|_\infty^2. \end{align*} From \eqref{stima_somma_tilde_c} it turns out that $\widetilde {\bm v}_n$ satisfies the following system of inequalities \begin{align*} \left\{ \begin{array}{ll} D_t\widetilde {\bm v}_n(t,x)-( \widetilde {\bm{\mathcal A}}^P+M{\rm Id})\widetilde {\bm v}_n(t,x)\leq \bm 0, & (t,x)\in(s,T]\times B_n, \vspace{1mm}\\ \displaystyle \frac{\partial \widetilde {\bm v}_n}{\partial\nu}(t,x)\leq \bm 0, & (t,x)\in(s,T]\times \partial B_n, \vspace{1mm} \\ \widetilde {\bm v}_n(s,x)\leq \bm 0, & x\in \overline B_n. \end{array} \right. \end{align*} By applying \cite[Theorem 3.15]{Pr67}, with $ {\bm{\mathcal A}}$ being replaced by $ \widetilde {\bm{\mathcal A}}^P+ M{\rm Id}$, we conclude that $\widetilde {\bm v}_n\leq \bm 0$ on $(s,T]\times B_n$, i.e., \begin{align*} &(u_{n,k}(t,x))^2+\alpha(t-s)|\nabla_{x} u_{n,k}(t,x)|^2+\alpha^2(t-s)^2(\vartheta_n(x))^2|D^2_{x}u_{n,k}(t,x)|^2\\ &+\alpha^3(t-s)^3(\vartheta_n(x))^4|D^3_{x}u_{n,k}(t,s)|^2\le e^{2\overline M(t-s)}\|\bm f\|_\infty^2 \end{align*} for every $n\in{\mathbb N}$ and $k=1,\ldots,m$, with the constant $\alpha$ being independent of $n$. To conclude the proof, it suffices to let $n$ tend to $\infty$ taking Proposition \ref{pro:appr_sol_neumann}$(iii)$ into account. The other cases can be deduced from similar arguments applied to the function ${\bm v}^h=(v_1^h,\ldots,v_m^h)$ ($h=1,2,3$), defined by \begin{align*} v_k^h(t,x)=&\!: w_k(t,x)+\alpha(t-s)^{(1-h)^+}w_k^1(t,x)+\alpha^2(t-s)^{(2-h)^+}(\vartheta(x))^2w_k^2(t,x)\\ &\,+\alpha^3(t-s)^{(3-h)^+}(\vartheta(x))^4w_k^3(t,x) \end{align*} for $k=1,\ldots,m$ and $(t,x)\in [s,T]\times\mathbb R^d$, where ${\bm w}, {\bm w}^1,{\bm w}^2$ and ${\bm w}^3$ are defined as above and $\alpha\in(0,1]$ is to be properly fixed. {\vspace{2mm}} (ii) Let us fix $\overline t\in[s,T]$ and $\mathcal T>0$. Arguing as in the proof of $(i)$ we get \eqref{stima_der_smgr_totale}. We notice that the constant $C(\overline t,\mathcal T)$ is uniform with respect to $\overline t\in[s,T]$ since the constant $C(s,T)$ in \eqref{stima_der_op_ev_totale} depends on the interval $[s,T]$. \end{proof} From Theorem \ref{thm-3.4} we can estimate the behaviour of some H\"older norms of ${\bm G}(t,s)\bm f$ and of ${\bm T}_{\overline t}(\tau)\bm f$, $\overline t\in I$, when $t-s$ and $\tau$ tend to $0$, respectively. \begin{coro} Under the assumptions of Theorem $\ref{thm-3.4}$, for every $\beta,\theta\in [0,3]$ with $\beta\le\theta$, there exists a positive constant $C=C(\beta,\theta,s,T)$ such that \begin{equation} \|{\bm G}(t,s)\bm f\|_{C^{\theta}_b(\mathbb R^d;\mathbb R^m)}\le C\frac{e^{\overline M(t-s)}}{(t-s)^{(\theta-\beta)/2}}\|\bm f\|_{C^{\beta}_b(\mathbb R^d;\mathbb R^m)},\quad\;\,\bm f\in C_b^\theta(\mathbb R^d;\mathbb R^m), \ t\in(s,T], \label{stima-11} \end{equation} where $\overline M:=2^{-1}(1+M+2M^+)$. Further, for every $\overline t\in [s,T]$, $\beta,\theta\in [0,3]$, with $\beta\le\theta$, and $\omega>0$ there exists a positive constant $K=K(\overline t,\beta,\theta,\omega)$, uniform with respect to $\overline t\in[s,T]$, such that \begin{equation} \|{\bm T}_{\overline t}(\tau)\bm f\|_{C^{\theta}_b(\mathbb R^d;\mathbb R^m)}\le K\frac{e^{(\omega+\overline {M})\tau}}{\tau^{(\theta-\beta)/2}}\|\bm f\|_{C^{\beta}_b(\mathbb R^d;\mathbb R^m)},\quad\;\,\bm f\in C_b^\theta(\mathbb R^d;\mathbb R^m), \ \tau>0, \label{stima-12} \end{equation} where $\overline M:=2^{-1}(1+M_{\{\overline t\}}+2(M_{\{\overline t\}})^+)$. \end{coro} \begin{proof} When at least one between $\theta$ and $\beta$ is not integer, the proof follows from an interpolation argument. We illustrate it in the particular case when $\theta=2$ and $\beta\in (1,2)$. It is well-known that $C^{\beta}_b(\mathbb R^d)$ is the real interpolation space of order $(\beta/2,\infty)$ between $C_b(\mathbb R^d)$ and $C_b^2(\mathbb R^d)$. Clearly, this implies that $C^{\beta}_b(\mathbb R^d;\mathbb R^m)=(C_b(\mathbb R^d;\mathbb R^m);C^{\beta}_b(\mathbb R^d;\mathbb R^m))_{\beta/2,\infty}$. Interpolating the estimates (we notice that $M\leq \overline M$) \begin{align*} &\|{\bm G}(t,s)\|_{L(C_b(\mathbb R^d;\mathbb R^m),C_b^{2k}(\mathbb R^d;\mathbb R^m))}\le C^k\frac{e^{\overline M(t-s)}}{(t-s)^k}, \qquad\;\,k=0,1,\end{align*} we obtain estimate \eqref{stima-11} in this case. Let us prove estimate \eqref{stima-12} with $\beta,\theta\in\{0,1,2,3\}$, $\beta\leq \theta$, the general case follows by interpolation. In particular we consider the case $\beta=0$ and $\theta=3$, since for the other integer values of $\beta$ and $\theta$ we get the assertion by analogous arguments. Let us fix $\omega>0$, and let $T\geq 1$ be such that $e^{\omega \tau}\tau^{-3/2}\geq1$ for every $\tau>T$. For $\tau\in (0,T]$, from \eqref{stima_der_smgr_totale} we get \begin{align*} \|{\bm T}_{\overline t}(\tau)\bm f\|_{C_b^{3}(\mathbb R^d;\mathbb R^m)}\leq \frac{C_{0,3}(\overline t, T)e^{\overline M}}{\tau^{3/2}}\|\bm f\|_{C_b(\mathbb R^d;\mathbb R^m)}, \qquad\;\, \bm f\in C_b(\mathbb R^d;\mathbb R^m),\;\, \tau\in (0,T]. \end{align*} If $\tau>T$, then we use the semigroup property to split ${\bm T}_{\overline t}(\tau)={\bm T}_{\overline t}(T){\bm T}_{\overline t}(\tau-T)$ and so \begin{align*} \|{\bm T}_{\overline t}(\tau)\bm f\|_{C_b^{3}(\mathbb R^d;\mathbb R^m)} \le & \|{\bm T}_{\overline t}(T)\|_{L(C_b(\mathbb R^d;\mathbb R^m);C_b^3(\mathbb R^d;\mathbb R^m))}\|{\bm T}_{\overline t}(\tau-T)\bm f\|_{C_b(\mathbb R^d;\mathbb R^m)} \\ \leq & \frac{C_{0,3}(\overline t, T)e^{\overline M}}{T^{3/2}}e^{\overline M(\tau-T)}\|\bm f\|_{C_b(\mathbb R^d;\mathbb R^m)} \\ \leq & \frac{C_{0,3}(\overline t, T)e^{(\omega+\overline M)\tau}}{\tau^{3/2}}\|\bm f\|_{C_b(\mathbb R^d;\mathbb R^m)}. \end{align*} Estimate \eqref{stima-12} follows with $K(\overline t, 0,3,\omega)=C_{0,3}(\overline t, T)$. \end{proof} \section{Maximal Schauder regularity for problems \eqref{ellipt} and \eqref{parab}} \label{sec:max_reg} \subsection{Continuity results} In this subsection we study the continuity of ${\bm T}_t(\tau)$ also with respect to $t\in I$. \begin{thm} \label{thm:continuita_totale} Let Hypotheses $\ref{hyp-base}$ and $\ref{hyp:smoothness_coeff}$ be satisfied, and fix $\bm f\in C_b(\mathbb R^d;\mathbb R^m)$. Then: \begin{enumerate}[(i)] \item the function $I\times (0,\infty)\times \mathbb R^d\ni (t,\tau,x)\mapsto ({\bm T}_t(\tau)\bm f)(x)$ is continuous; \item for every bounded sequence $(\bm f_n)\subset C_b(\mathbb R^d;\mathbb R^m)$, which converges to $\bm f$ locally uniformly on $\mathbb R^d$, the function $(t,\tau,x)\mapsto ({\bm T}_t(\tau)\bm f_n)(x)$ converges to the function $(t,\tau,x)\mapsto ({\bm T}_t(\tau)\bm f)(x)$ locally uniformly in $I\times (0,\infty)\times \mathbb R^d$ as $n$ tends to $\infty$. \end{enumerate} \end{thm} \begin{proof} (i). For reader's convenience, we split the proof into two steps. In the former, we prove the statement for the semigroup ${\bm T}_t^{\mathcal N,R}(\tau)$ generated by the realization of $\bm{\mathcal A}$ in $C_b(B_R;\mathbb R^m)$ with homogeneous Neumann boundary conditions on $\partial B_R$ for every $R>0$, with $\mathbb R^d$ replaced by $B_R$. In the latter we conclude. {\em Step 1.} We fix $R>0$, $t\in I$ and prove that for every $\tau\in[0,\infty)$ it holds that \begin{align} \label{stima_neumann_smgr} \|{\bm T}_t^{\mathcal N,R}(\tau)\bm f\|_{C_b(B_R;\mathbb R^m)}\leq e^{M_{\{t\}}^P\tau}\|\bm f\|_{C_b(B_R;\mathbb R^m)}. \end{align} For this purpose, we set ${\bm w}:=(w_1,\ldots,w_m)$ with $w_j:=(({\bm T}_t^{\mathcal N,R}(\tau)\bm f)_j)^2-e^{-2M_{\{t\}}\tau}\|\bm f\|_{\infty}$ for $j=1,\ldots,m$. Arguing as in the uniqueness part of the proof of Proposition \ref{prop-2.6}, we deduce that function ${\bm w}$ satisfies the system of inequalities \begin{align*} \left\{ \begin{array}{ll} D_t{\bm w}(\tau,x)-({\bm{\mathcal A}}^P(t)+MId){\bm w}(\tau,x)\leq \bm 0, & (\tau,x)\in(0,\infty)\times B_R, \vspace{1mm}\\ \displaystyle \frac{\partial {\bm w}}{\partial\nu}(\tau,x)= \bm 0, & (\tau,x)\in(0,\infty)\times \partial B_R, \vspace{1mm} \\ {\bm w}(s,x)\leq \bm 0, & x\in \overline B_R. \end{array} \right. \end{align*} From \cite[Theorem 3.15]{Pr67} it follows that ${\bm w}\le \bm 0$, which gives \eqref{stima_neumann_smgr}. Let $(\bm f_n)\subset C_b^{2}(\mathbb R^d;\mathbb R^m)$ converge to $\bm f$ locally uniformly on $\mathbb R^d$, let $\overline t\in I$ and let $J=\{t\in I:|t-\overline t|\leq 1\}$. Then, for every $t\in J$ we can estimate \begin{align} \|({\bm T}_t^{\mathcal N,R}(\tau)\bm f)(x)-({\bm T}_{\overline t}^{\mathcal N,R}(\tau)\bm f)(x)\| \leq & \|({\bm T}_t^{\mathcal N,R}(\tau)\bm f)(x)-({\bm T}_t^{\mathcal N,R}(\tau)\bm f_n)(x)\|\notag \\ & + \|({\bm T}_t^{\mathcal N,R}(\tau)\bm f_n)(x)-({\bm T}_{\overline t}^{\mathcal N,R}(\tau)\bm f_n)(x)\| \notag \\ & + \|({\bm T}_{\overline t}^{\mathcal N,R}(\tau)\bm f_n)(x)-({\bm T}_{\overline t}^{\mathcal N,R}(\tau)\bm f)(x)\| \label{step_1_neumann_1} \end{align} for every $(\tau,x)\in[0,\infty)\times B_R$. From \eqref{stima_neumann_smgr} we infer that \begin{align} & \|({\bm T}_t^{\mathcal N,R}(\tau)\bm f)(x)-({\bm T}_{t}^{\mathcal N,R}(\tau)\bm f_n)(x)\| + \|({\bm T}_{\overline t}^{\mathcal N,R}(\tau)\bm f_n)(x)-({\bm T}_{\overline t}^{\mathcal N,R}(\tau)\bm f)(x)\| \notag \\ \le & 2e^{M_J\tau}\|\bm f_n-\bm f\|_{C_b(B_R;\mathbb R^m)}, \label{step_1_neumann_2} \end{align} which vanishes as $n$ tends to $\infty$, locally uniformly with respect to $\tau$, since $\bm f_n$ converges to $\bm f$ as $n$ tends to $\infty$, locally uniformly on $\mathbb R^d$. Further, if we set ${\bm v}_n={\bm T}_t^{\mathcal N,R}(\cdot)\bm f_n-{\bm T}_{\overline t}^{\mathcal N,R}(\cdot)\bm f_n$ then we get \begin{align*} \left\{ \begin{array}{ll} D_t{\bm v}_n(\tau,x)=\bm{\mathcal A}{\bm v}_n(\tau,x)+(\bm{\mathcal A}-\bm{\mathcal A}(\overline t)){\bm T}_{\overline t}^{\mathcal N,R}(\tau)\bm f_n(x), & (\tau,x)\in(0,\infty)\times B_R, \vspace{1mm}\\ \displaystyle\frac{\partial{\bm v}_n}{\partial\nu}(\tau,x)=\bm 0, & (\tau,x)\in(0,\infty)\times \partial B_R, \vspace{1mm} \\ {\bm v}_n(0,x)=\bm 0, & x\in B_R. \end{array} \right. \end{align*} The variation-of-constants formula gives \begin{align*} {\bm v}_n(\tau,x)=\int_0^\tau {\bm T}_{t}^{\mathcal N,R}(\tau-s)\big ((\bm{\mathcal A}(t)-\bm{\mathcal A}(\overline t)){\bm T}_{\overline t}^{\mathcal N,R}(s)\bm f_n\big )(x)ds, \quad (\tau,x)\in[0,\infty)\times B_R. \end{align*} Hence, for every $\mathcal T>0$ we get \begin{align} \sup_{(\tau,x)\in[0,\mathcal T]\times B_R}\|{\bm v}_n(\tau,x)\| \leq &\sup_{\tau\in[0,\mathcal T]}\int_0^\tau e^{M_{J}(\tau-s)}\|\bm{\mathcal A}(t)-\bm{\mathcal A}(\overline t)\|_{0,B_R}\| \|{\bm T}_{\overline t}^{\mathcal N,R}(s)\bm f_n\|_{C_b^2(B_R;\mathbb R^m)}ds \notag \\ \leq & \widetilde C \|\bm f_n\|_{C^2_b(B_R;\mathbb R^m)}\|\bm{\mathcal A}(t)-\bm{\mathcal A}(\overline t)\|_{0,B_R}, \label{step_1_neumann_3} \end{align} where $\widetilde C$ is a positive constant, uniform with respect to $\tau$ and $t$ on bounded sets (see \cite[Section IV.2, Theorem 3.4]{Ei69}). Let us fix $\varepsilon>0$, $\overline t\in I$ and let $J$ be as above. From \eqref{step_1_neumann_2} there exists $\overline n\in{\mathbb N}$ such that \begin{align} \label{step_1_neumann_4} \|({\bm T}_t^{\mathcal N,R}(\tau)\bm f)(x)-({\bm T}_{t}^{\mathcal N,R}(\tau)\bm f_n)(x)\| + \|({\bm T}_{\overline t}^{\mathcal N,R}(\tau)\bm f_n)(x)-({\bm T}_{\overline t}^{\mathcal N,R}(\tau)\bm f)(x)\|\leq \frac{\varepsilon}{2}, \end{align} for every ${\mathbb N}\ni n\geq \overline n$ and $t\in J$, locally uniformly with respect to $\tau\in[0,\infty)$ and uniformly with respect to $x\in B_R$. Since the coefficients of the operator $\bm{\mathcal A}$ belong to $C^{\alpha/2,\alpha}_{\rm loc}(I\times \mathbb R^d)$, from \eqref{step_1_neumann_3} it follows that there exists $\delta>0$ such that \begin{align} \label{step_1_neumann_5} \|({\bm T}_t^{\mathcal N,R}(\tau)\bm f_{\overline n})(x)-({\bm T}_{\overline t}^{\mathcal N,R}(\tau)\bm f_{\overline n})(x)\| \leq \frac\varepsilon2, \quad {\rm if } \ |t-\overline t|\leq \delta,\ t\in J, \end{align} locally uniformly with respect to $\tau\in [0,\infty)$ and uniformly with respect to $x\in B_R$. Putting together \eqref{step_1_neumann_4} and \eqref{step_1_neumann_5}, from \eqref{step_1_neumann_1} it follows that for every $\varepsilon>0$ there exists $\delta>0$ such that $\|({\bm T}_t^{\mathcal N,R}(\tau)\bm f)(x)-({\bm T}_{\overline t}^{\mathcal N,R}(\tau)\bm f)(x)\|\leq \varepsilon$ for every $t\in J$, such that $|t-\overline t|\leq \delta$, locally uniformly with respect to $\tau\in[0,\infty)$, uniformly with respect to $x\in B_R$. Since $(\tau,x)\mapsto {\bm T}_{t}^{\mathcal N,R}(\tau)\bm f(x)$ is continuous in $(0,\infty)\times B_R$, pointwise with respect to $t$, the assertion follows. {\it Step 2}. Here, we prove that, for every $\bm f\in C_b(\mathbb R^d;\mathbb R^m)$, $({\bm T}_t^{\mathcal N,n}(\tau)\bm f)(x)$ converges to $ ({\bm T}_t(\tau)\bm f)(x)$ as $n$ tends to $\infty$ locally uniformly with respect to $(t,\tau,x)\in I\times (0,\infty)\times \mathbb R^d$. As a byproduct, from Step $1$ we obtain that the function $(t,\tau,x)\mapsto ({\bm T}_t(\tau)\bm f)(x)$ is continuous in $I\times (0,\infty)\times \mathbb R^d$. Let us fix $t\in I$ and observe that from the classical interior Schauder estimates it follows that, for every compact set $E\in(0,\infty)\times \mathbb R^d$, \begin{align} \label{stima_interna_conv_totale} \|{\bm T}_t^{\mathcal N,n}(\cdot)\bm f\|_{C^{1+\alpha/2,2+\alpha}(E;\mathbb R^m)}\leq C, \end{align} where $C$ is a positive constant which depends on $E$ and is locally uniform with respect to $t\in I$ (see Proposition \ref{prop:app_stime_varie}(iv)), but does not depend on $n$. Standard arguments (see e.g., \cite[Remark 2.5]{AngLor20}) imply that for every $t\in I$ the sequence $({\bm T}_t^{\mathcal N,n}(\cdot)\bm f)$ converges to ${\bm T}_t(\cdot)\bm f$ in $C^{1,2}(E;\mathbb R^m)$, as $n$ tends to $\infty$, for every compact subset $E\subset(0,\infty)\times \mathbb R^d$. Hence, $({\bm T}_t^{\mathcal N,n}(\tau)\bm f)(x)$ tends to $({\bm T}_t(\tau)\bm f)(x)$ as $n$ tends to $\infty$, locally uniformly with respect to $(\tau,x)\in (0,\infty)\times \mathbb R^d$, pointwise with respect to $t\in I$. Let us assume by contradiction that such a convergence is not locally uniform with respect to $t$, i.e., there exist $[s,T]\subset I$, $[\varepsilon,\mathcal T]\subset (0,\infty)$ and $B_h\subset \mathbb R^d$ such that \begin{align*} \limsup_{n\rightarrow\infty}\sup_{t\in[s,T]}\|{\bm T}_t^{\mathcal N,n}(\cdot)\bm f-{\bm T}_t(\cdot)\bm f\|_{C_b([\varepsilon,\mathcal T]\times B_h)}\neq0. \end{align*} This implies that there exist $\varepsilon>0$, a sequence $(t_n)\subset [s,T]$ and an increasing sequence $(m_n)\subset {\mathbb N}$ such that \begin{align} \label{stima_contr_conv_totale} \|{\bm T}_{t_n}^{\mathcal N,m_n}(\cdot)\bm f-{\bm T}_{t_n}(\cdot)\bm f\|_{C_b([\varepsilon,\mathcal T]\times B_h;\mathbb R^m)}\geq \varepsilon, \qquad\;\, n\in{\mathbb N}. \end{align} Let us notice that, up to a subsequence, $t_n$ converges to $\widehat t\in [s,T]$ as $n$ tends to $\infty$. Since the constant $C$ in \eqref{stima_interna_conv_totale} is uniform with respect to $t\in[s,T]$, it follows that for every compact set $E\subset (0,\infty)\times \mathbb R^d$ we can find a positive constant $C$ such that \begin{align*} \|{\bm T}_{t_n}^{\mathcal N,m_n}(\cdot)\bm f\|_{C^{1+\alpha/2,2+\alpha}(E;\mathbb R^m)}+\|{\bm T}_{t_n}(\cdot)\bm f\|_{C^{1+\alpha/2,2+\alpha}(E;\mathbb R^m)}\leq C, \qquad\;\, n\in{\mathbb N}. \end{align*} Hence, there exist ${\bm v},{\bm w}\in C^{1+\alpha/2,2+\alpha}_{\rm loc}((0,\infty)\times \mathbb R^d;\mathbb R^m)$ such that, up to a subsequence, ${\bm T}_{t_n}^{\mathcal N,m_n}(\cdot)\bm f$ converges to ${\bm v}$ and ${\bm T}_{t_n}(\cdot)\bm f$ converges to $ {\bm w}$ in $C^{1,2}(E;\mathbb R^m)$, as $n$ tends to $\infty$, and both ${\bm v}$ and ${\bm w}$ solve the equation $D_\tau{\bm u}=\bm{\mathcal A}(\widehat t){\bm u}$ on $(0,\infty)\times \mathbb R^d$. Let us prove that ${\bm v}\equiv{\bm w}\equiv {\bm T}_{\widehat t}(\cdot)\bm f$ by means of a localization argument. Let us fix $M>0$, and let $\vartheta_M\in C_c^\infty(\mathbb R^d)$ satisfy the condition $\chi_{B_M}\leq \vartheta_M\leq \chi_{B_{M+1}}$. Let $\overline n\in{\mathbb N}$ be such that $m_n>M+1$ for every $n\geq \overline n$. For every $n> \overline n$ the function ${\bm v}_n:=\vartheta_M{\bm T}_{t_n}^{\mathcal N,m_n}(\cdot)\bm f$ satisfies the Neumann-Cauchy problem \begin{align*} \left\{ \begin{array}{ll} D_\tau {\bm v}_n(\tau,x)=\bm{\mathcal A}(t_n){\bm v}_n(\tau,x)+{\bm g}_n(\tau,x), & (\tau,x)\in(0,\infty)\times B_{M+1}, \vspace{1mm}, \\ \displaystyle \frac{\partial {\bm v}_n}{\partial\nu}(\tau,x)=\bm 0, & (\tau,x)\in (0,\infty)\times \partial B_{M+1}, \vspace{1mm} \\ {\bm v}_n(0,x)=\vartheta_M(x)\bm f(x), & x\in B_{M+1}, \end{array} \right. \end{align*} with $g_{n,k}=-2\langle Q^k(t_n,\cdot)\nabla \vartheta_M,\nabla_{x} ({\bm T}_{t_n}^{\mathcal N,m_n}(\cdot)\bm f)_k\rangle -({\bm T}_{t_n}^{\mathcal N,m_n}(\cdot)\bm f)_k{\mathcal A}_{k,0}(t_n)\vartheta_M$, where ${\mathcal A}_{k,0}=\sum_{i,j=1}^dq_{ij}^kD^2_{ij}+\sum_{i=1}^d b_i^kD_i$, for every $k=1,\ldots,m$. The variation-of-constants formula gives \begin{align*} {\bm v}_n(\tau,x)={\bm T}^{\mathcal N,M+1}_{t_n}(\tau)(\vartheta_M\bm f)(x)+\int_0^\tau{\bm T}_{t_n}^{\mathcal N,M+1}(\tau-s)({\bm g}_n(s,\cdot))(x)ds \end{align*} for every $(\tau,x)\in(0,\infty)\times B_{M+1}$. Since $\sqrt \tau\|{\bm T}^{\mathcal N,R}_{t}(\tau)\bm f\|_{C^1_b(B_{M+1};\mathbb R^m)}\leq K\|\bm f\|_{C_b(\mathbb R^d;\mathbb R^m)}$ for every $\tau\in(0,\mathcal T]$, some positive constant $K$ which is uniform with respect to $t\in[s,T]$ and $R>M+2$ (see Proposition \ref{prop:app_stime_varie}(iii) and \eqref{stima_neumann_smgr}), we get \begin{align*} \|({\bm T}_{t_n}^{\mathcal N,m_n}(\tau)\bm f)(x)-\bm f(x)\|\leq \|({\bm T}^{\mathcal N,M+1}_{t_n}(\tau)(\vartheta_M\bm f))(x)-\bm f(x)\|+K'\sqrt \tau\|\bm f\|_{C_b(\mathbb R^d;\mathbb R^m)} \end{align*} for every $(\tau,x)\in(0,\mathcal T]\times B_M$ and some positive constant $K'$, uniformly with respect to $(t,\tau)\in[s,T]\times (0,\mathcal T]$. Letting $n$ tend to $\infty$, we obtain that ${\bm T}_{t_n}^{\mathcal N,m_n}(\tau)\bm f(x)$ converges to $ {\bm v}(\tau,x)$ and, from Step $1$, that $ {\bm T}^{\mathcal N,M+1}_{t_n}(\tau)(\vartheta_M\bm f)(x)$ converges to ${\bm T}^{\mathcal N,M+1}_{\widehat t}(\tau)(\vartheta_M\bm f)(x)$ for every $(\tau,x)\in(0,\mathcal T]\times B_M$. Hence, \begin{align*} \|{\bm v}(\tau,x)-\bm f(x)\|\leq \|({\bm T}^{\mathcal N,M+1}_{\widehat t}(\tau)(\vartheta_M\bm f))(x)-\bm f(x)\|+K'\sqrt \tau\|\bm f\|_{C_b(\mathbb R^d;\mathbb R^m)} \end{align*} for every $(\tau,x)\in(0,\mathcal T]\times B_M$, and letting $\tau$ tends to $0^+$ we conclude that ${\bm v}$ is continuous on $\{0\}\times B_M$. This means that ${\bm v}$ is a classical solution to the Cauchy problem \begin{align*} \left\{ \begin{array}{ll} D_t{\bm u}(\tau,x)=\bm{\mathcal A}(\widehat t){\bm u}(\tau,x), & (\tau,x)\in(0,\mathcal T]\times \mathbb R^d, \vspace{1mm} \\ {\bm u}(0,x)=\bm f(x), & x\in\mathbb R^d, \end{array} \right. \end{align*} and so ${\bm v}\equiv {\bm T}_{\widehat t}(\cdot)\bm f$. Similar arguments show that ${\bm w}\equiv {\bm T}_{\widehat t}(\cdot)\bm f$. Letting $n$ tend to $\infty$ in \eqref{stima_contr_conv_totale} the assertion follows. \medskip (ii) We already know that, for every fixed $t\in I$, the sequence ${\bm T}_t(\cdot)\bm f_n$ converges to ${\bm T}_t(\cdot)\bm f$ locally uniformly on $[0,\infty)\times \mathbb R^d$ as $n$ tends to $\infty$ (see Proposition \ref{prop:conv_loc_unif}$(ii)$ and Remark \ref{rmk-claudio}). Let us assume by contradiction that there exist $[s,T]\subset I$, $[\varepsilon,\mathcal T]\subset (0,\infty)$ and $B_h\subset \mathbb R^d$ such that \begin{align*} \limsup_{n\rightarrow\infty}\sup_{t\in[s,T]}\|{\bm T}_t(\cdot)\bm f_n-{\bm T}_t(\cdot)\bm f\|_{C_b([\varepsilon,\mathcal T]\times B_h)}\neq0. \end{align*} This implies that there exist $\varepsilon>0$ , a sequence $(t_n)\subset [s,T]$ and an increasing and divergent sequence $(m_n)\subset {\mathbb N}$ such that \begin{align} \label{stima_contr_conv_totale_2} \|{\bm T}_{t_n}(\cdot)\bm f_{m_n}-{\bm T}_{t_n}(\cdot)\bm f\|_{C_b([\varepsilon,\mathcal T]\times B_h;\mathbb R^m)}\geq \varepsilon, \qquad\;\, n\in{\mathbb N}. \end{align} Let us notice that, up to a subsequence, $t_n$ converges to $\widehat t\in [s,T]$ as $n$ tends to $\infty$. Arguing as in the proof of Step $2$ in (i), it follows that for every compact set $E\subset (0,\infty)\times \mathbb R^d$ we can find a positive constant $C$ such that \begin{align*} \|{\bm T}_{t_n}(\cdot)\bm f_{m_n}\|_{C^{1+\alpha/2,2+\alpha}(E;\mathbb R^m)}+\|{\bm T}_{t_n}(\cdot)\bm f\|_{C^{1+\alpha/2,2+\alpha}(E;\mathbb R^m)}\leq C, \qquad\;\, n\in{\mathbb N}. \end{align*} By applying a diagonal argument, we can find ${\bm v},{\bm w}\in C^{1+\alpha/2,2+\alpha}_{\rm loc}((0,\infty)\times \mathbb R^d;\mathbb R^m)$ such that, up to a subsequence, ${\bm T}_{t_n}(\cdot)\bm f_{m_n}$ and ${\bm T}_{t_n}(\cdot)\bm f$ converge, respectively, to ${\bm v}$ and to $ {\bm w}$ in $C^{1,2}(E;\mathbb R^m)$, as $n$ tends to $\infty$, for every compact set $E\subset(0,\infty)\times\mathbb R^d$. Moreover, both ${\bm v}$ and ${\bm w}$ solve the equation $D_\tau{\bm z}=\bm{\mathcal A}(\widehat t){\bm z}$ on $(0,\infty)\times \mathbb R^d$. From property (i) we already know that ${\bm w}\equiv{\bm T}_{\widehat t}(\cdot)\bm f$. Let us prove that ${\bm v}\equiv {\bm T}_{\widehat t}(\cdot)\bm f$ by means of a localization argument. Let us fix $M>0$, and let $\vartheta_M\in C_c^\infty(\mathbb R^d)$ satisfy the condition $\chi_{B_M}\leq \vartheta\leq \chi_{B_{M+1}}$. For every $n\in{\mathbb N}$ the function ${\bm v}_n:=\vartheta_M{\bm T}_{t_n}(\cdot)\bm f_{m_n}$ satisfies the Neumann-Cauchy problem \begin{align*} \left\{ \begin{array}{ll} D_\tau {\bm v}_n(\tau,x)=\bm{\mathcal A}(t_n){\bm v}_n(\tau,x)+{\bm g}_n(\tau,x), & (\tau,x)\in(0,\infty)\times B_{M+1}, \vspace{1mm}, \\ \displaystyle \frac{\partial {\bm v}_n}{\partial\nu}(\tau,x)=\bm 0, & (\tau,x)\in (0,\infty)\times \partial B_{M+1}, \vspace{1mm} \\ {\bm v}_n(0,x)=\vartheta_M(x)\bm f_{m_n}(x), & x\in B_{M+1}, \end{array} \right. \end{align*} with the function ${\bm g}_n$ being defined as in Step 1, and with ${\bm T}_{t_n}^{\mathcal N, an m_n}(\cdot)\bm f$ being replaced by the function ${\bm T}_{t_n}(\cdot)\bm f_{m_n}$. The variation-of-constants formula gives \begin{align*} {\bm v}_n(\tau,x)={\bm T}^{\mathcal N,M+1}_{t_n}(\tau)(\vartheta_M\bm f_{m_n})(x)+\int_0^\tau{\bm T}_{t_n}^{\mathcal N,M+1}(\tau-s)({\bm g}_n(s,\cdot))(x)ds \end{align*} for each $(\tau,x)\in(0,\infty)\times B_{M+1}$. Since $\sqrt \tau\|{\bm T}_{t}(\tau)\bm f_{m_n}\|_{C^1_b(B_{M+1};\mathbb R^m)}\leq K\sup_{n\in{\mathbb N}}\|\bm f_n\|_{\infty}$ for every $\tau\in(0,\mathcal T]$ and some positive constant $K$, which is uniform with respect to $t\in[s,T]$ (see Proposition \ref{prop:app_stime_varie}(iii)), it follows that \begin{align*} \|({\bm T}_{t_n}(\tau)\bm f)(x)-\bm f(x)\|\leq \|({\bm T}^{\mathcal N,M+1}_{t_n}(\tau)(\vartheta_M\bm f_{m_n}))(x)-\bm f(x)|+K'\sqrt \tau\sup_{n\in{\mathbb N}}\|\bm f_n\|_{C_b(\mathbb R^d;\mathbb R^m)}, \end{align*} for every $(\tau,x)\in(0,\mathcal T]\times B_M$ and a constant $K'>0$, uniformly with respect to $n\in{\mathbb N}$. We claim that ${\bm T}^{\mathcal N,M+1}_{t_n}(\cdot)(\vartheta_M\bm f_{m_n})$ converges to $ {\bm T}^{\mathcal N,M+1}_{\widehat t}(\cdot)(\vartheta_M\bm f)$ as $n$ tends to $\infty$, uniformly in $(0,\mathcal T]\times B_{M+1}$. Once the claim is proved, letting $n$ tend $\infty$ and then $\tau$ tend to $0^+$, we get that ${\bm u}$ is continuous on $\{0\}\times B_M$ for every $M>0$, and so it is a classical solution to \begin{align*} \left\{ \begin{array}{ll} D_t{\bm v}(\tau,x)=\bm{\mathcal A}(\widehat t){\bm v}(\tau,x), & (\tau,x)\in(0,T]\times \mathbb R^d, \vspace{1mm} \\ {\bm v}(0,x)=\bm f(x), & x\in\mathbb R^d. \end{array} \right. \end{align*} This means that ${\bm v}\equiv {\bm T}_{\widehat t}(\cdot)\bm f$. Letting $n$ tend to $\infty$ in \eqref{stima_contr_conv_totale_2} we get a contradiction, and this yields the assertion. It remains to prove the claim. From \eqref{stima_neumann_smgr} it follows that \begin{align*} & \|({\bm T}^{\mathcal N,M+1}_{t_n}(\tau)(\vartheta_M\bm f_{m_n}))(x)- ({\bm T}^{\mathcal N,M+1}_{\widehat t}(\tau)(\vartheta_M\bm f))(x)\|\\ \leq & \|({\bm T}^{\mathcal N,M+1}_{t_n}(\tau)(\vartheta_M\bm f_{m_n}))(x)- ({\bm T}^{\mathcal N,M+1}_{t_n}(\tau)(\vartheta_M\bm f))(x)\| \\ & +\|({\bm T}^{\mathcal N,M+1}_{t_n}(\tau)(\vartheta_M\bm f))(x)- ({\bm T}^{\mathcal N,M+1}_{\widehat t}(\tau)(\vartheta_M\bm f))(x)\| \\ \leq & (e^{M_{[s,T]}^P\mathcal T}\vee 1)\|\vartheta_M \bm f_{m_n}-\vartheta_M\bm f\|_{C_b(B_{M+1};\mathbb R^m)} \\ & +\|({\bm T}^{\mathcal N,M+1}_{t_n}(\tau)(\vartheta_M\bm f))(x)- ({\bm T}^{\mathcal N,M+1}_{\widehat t}(\tau)(\vartheta_M\bm f))(x)\| \\ =:& I_1(n)+I_2(n,\tau,x). \end{align*} Since $\bm f_n$ converges to $\bm f$ locally uniformly as $n$ tends to $\infty$, we infer that $I_1(n)$ converges to $0$. Further, from Step $1$ in the proof of property (i), it follows that $I_2(n,\tau,x)$ vanishes uniformly with respect to $(\tau,x)\in[0,\infty)\times B_{M+1}$. The claim is proved. \end{proof} \subsection{Schauder estimates} In this subsection we prove optimal Schauder estimates for the solutions to the elliptic system \begin{equation} \lambda{\bm u}(t,x)-\bm{\mathcal A}(t){\bm u}(t,x)=\bm f(t,x), \qquad\;\, t\in [s,T],\;\, x\in \mathbb R^d \label{ellipt} \end{equation} and to the Cauchy problem \begin{equation} \left\{ \begin{array}{lll} D_t{\bm u}(t,x)=\bm{\mathcal A}(t){\bm u}(t,x)+{\bm g}(t,x), &t\in [s,T], &x\in\mathbb R^d,\\[1mm] {\bm u}(s,x)=\bm f(x), &&x\in\mathbb R^d, \end{array} \right. \label{parab} \end{equation} At first we characterize the maximal domain of the realization of $\bm{\mathcal A}(t)$ in $C_b(\mathbb R^d;\mathbb R^m)$ for every $t\in I$. To this aim, we fix $t\in I$ and for each $\lambda>M_{\{t\}}^P$ we consider the operator ${\bm R}_t(\lambda):C_b(\mathbb R^d;\mathbb R^m)\to C_b(\mathbb R^d;\mathbb R^m)$ defined by \begin{align} \label{resolvent_operator_t} {\bm R}_t(\lambda)\bm f(x):=\int_0^\infty e^{-\lambda \tau}({\bm T}_t(\tau)\bm f)(x)d\tau, \qquad\;\, \bm f\in C_b(\mathbb R^d;\mathbb R^m). \end{align} Thanks to \eqref{stima_uni_gen_auto}, ${\bm R}_t(\lambda)$ is well defined and $\|{\bm R}_t(\lambda)\|_{L(C_b(\mathbb R^d;\mathbb R^m))}\leq (\lambda-M_{\{t\}}^P)^{-1}$. Since the family $\{{\bm R}_t(\lambda): \lambda>M_{\{t\}}^P\}$ satisfies the resolvent identity, there exists a closed operator ${\mathcal B}(t):D({\mathcal B}(t))\subset C_b(\mathbb R^d;\mathbb R^m)\rightarrow C_b(\mathbb R^d;\mathbb R^m)$ such that ${\bm R}_t(\lambda)=(\lambda-{\mathcal B}(t))^{-1}$ for each $\lambda >M_{\{t\}}^P$. Now, we prove that ${\mathcal B}(t)$ is a suitable realization of $\bm{\mathcal A}(t)$ in $C_b(\mathbb R^d;\mathbb R^m)$, but to this aim we need a maximum principle for Problem \eqref{ellipt}. \begin{prop} \label{prop:ell_max_princ} Fix $t\in I$, $\lambda>M_{\{t\}}^P\vee \lambda_{\{t\}}$ $($see Hypothesis $\ref{hyp-base}(iii))$. Suppose that ${\bm u}\in C_b(\mathbb R^d;\mathbb R^m)\cap \bigcap_{p>1}W^{2,p}_{\rm loc}(\mathbb R^d;\mathbb R^m)$ is such that $\bm{\mathcal A}^P(t){\bm u}\in C(\mathbb R^d;\mathbb R^m)$ and satisfies the inequality $\lambda {\bm u}-\bm{\mathcal A}^P(t){\bm u}\leq \bm 0$. Then, ${\bm u}\leq\bm 0$ on $\mathbb R^d$. \end{prop} \begin{proof} Let $t$ and $\lambda$ be as in the statement and introduce the function ${\bm v}_n={\bm u}-n^{-1}\boldsymbol\varphi_{\{t\}}$, where $\boldsymbol \varphi_{\{t\}}$ is the function introduced in Hypothesis \ref{hyp-base}(iii). The choice of $\lambda$ and Hypothesis \ref{hyp-base}(iii) implies that \begin{align} \label{max_prin_ell_neg} \lambda {\bm v}_n(x)-(\bm{\mathcal A}^P(t){\bm v}_n)(x) <\bm 0,\qquad\;\,x\in\mathbb R^d. \end{align} Fix $n\in{\mathbb N}$. Since ${\bm u}$ is bounded on $\mathbb R^d$ and $\bm\varphi_{\{t\}}(x)$ blows up, componentwise, as $|x|$ tends to $\infty$, for each $k\in \{1,\ldots,m\}$ the function $v_{n,k}$ attains its maximum value at some point $x_{n,k}\in\mathbb R^d$. Let $\overline k$ be such that $v_{n,\overline k}(x_{n,\overline k})=\max_{j=1,\ldots,m}v_{n,j}(x_{n,j})$ and assume that $v_{n,\overline k}(x_{n,\overline k})>0$. Since ${\mathcal A}_{\overline k,0}(t)v_{n,\overline k}=(\bm{\mathcal A}^P(t){\bm v}_n)_{\overline k}-(C^P(t,\cdot){\bm v}_n)_{\overline k}$ is a continuous function in $\mathbb R^d$, we can apply \cite[Theorem 3.1.10]{Lun95} and infer that ${\mathcal A}_{\overline k,0}(t)v_{\overline k}(x_{n,\overline k})\leq 0$. Thus, taking also into account that $(C^P(t,x_{n,\overline k}){\bm v}(x_{n,\overline k}))_{\overline k} \le \sum_{j=1}^mc^P_{\overline k,j}v_{n,\overline k}(x_{n,\overline k})$, we can estimate $\lambda v_{n,\overline k}(x_{n,\overline k})-(\bm{\mathcal A}^P(t){\bm v}_n)_{\overline k}(x_{n,\overline k}) \geq (\lambda-M_{\{t\}}^P)v_{n,\overline k}(x_{n,\overline k})>0$, which contradicts \eqref{max_prin_ell_neg}. It follows that ${\bm v}_n\le\bm 0$ on $\mathbb R^d$ for every $n\in{\mathbb N}$. Letting $n$ tend to $\infty$ we get ${\bm u}\leq \bm 0$ on $\mathbb R^d$. \end{proof} \begin{prop} \label{prop:carat_dom_At} For every $t\in I$ it holds that \begin{align*} D({\mathcal B}(t))=\bigg\{\bm f\in C_b(\mathbb R^d;\mathbb R^m)\cap\bigcap_{p>1}W^{2,p}_{\rm loc}(\mathbb R^d;\mathbb R^m):\bm{\mathcal A}(t)\bm f\in C_b(\mathbb R^d;\mathbb R^m)\bigg\}, \end{align*} and ${\mathcal B}(t)\bm f=\bm{\mathcal A}(t)\bm f$ for every $\bm f\in D({\mathcal B}(t))$. Further, $D({\mathcal B}(t))$ is continuously embedded in $C_b^\theta(\mathbb R^d;\mathbb R^m)$ for every $\theta\in(0,2)$ and there exists a positive constant $C(\theta)$ such that \begin{align*} \|\bm f\|_{C_b^\theta(\mathbb R^d;\mathbb R^m)}\leq C(\theta)\|\bm f\|_\infty^{1-\theta/2}\|\bm f\|_{D({\mathcal B}(t))}^{\theta/2}, \qquad\;\, \bm f\in D({\mathcal B}(t)). \end{align*} \end{prop} \begin{proof} In the proof we separately provide the two inclusions, following the lines of \cite[Proposition 3.1]{DeLo11}. {``$ \subset$''}. Let ${\bm u}\in D({\mathcal B}(t))$ be such that ${\bm u}=R(\lambda,{\mathcal B}(t))\bm f$ for some $\bm f\in C_b(\mathbb R^d;\mathbb R^m)$ and $\lambda>M_{\{t\}}^P$. For every $n\in{\mathbb N}$ we set \begin{align*} {\bm u}_n(x):=\int_{1/n}^ne^{-\lambda \tau}({\bm T}_t(\tau)\bm f)(x) d\tau , \qquad\;\, x\in\mathbb R^d. \end{align*} From \eqref{stima_uni_gen_auto} it follows that ${\bm u}_n$ converges to ${\bm u}$ in $C_b(\mathbb R^d;\mathbb R^m)$ as $n$ tends to $\infty$. Further, the spatial regularity of ${\bm T}_t(\tau)\bm f$ allows us to apply the operator $\bm{\mathcal A}(t)$ to ${\bm u}_n$, which gives $\bm{\mathcal A}(t){\bm u}_n = \lambda {\bm u}_n+e^{-\lambda n}{\bm T}_t(n)\bm f-e^{-\lambda/n}{\bm T}_t(1/n)\bm f$, in $\mathbb R^d$ for every $n\in{\mathbb N}$. This means that the sequence $(\bm{\mathcal A}(t){\bm u}_n)$ is bounded in $C_b(\mathbb R^d;\mathbb R^m)$. By applying the classical $L^p$-interior estimates to each component $u_{n,k}$ of the function ${\bm u}$, it follows that \begin{align*} \|u_{n,k}\|_{L^p(B_R)} \leq & K\left(\|u_{n,k}\|_{L^p(B_{2R})}+\|\mathcal A_{k,0}(t)u_{n,k}\|_{L^p(B_{2R})}\right) \\ \leq & K\left(\|u_{n,k}\|_{L^p(B_{2R})}+\|(\bm{\mathcal A}(t){\bm u}_{n})_k\|_{L^p(B_{2R};\mathbb R^m)}+\|(C{\bm u}_n)_k\|_{L^p(B_{2R};\mathbb R^m)}\right) \end{align*} and the last side of the previous chain of inequalities is independent of $n$. Arguing as in the proof of \cite[Proposition 3.1]{DeLo11} we infer that ${\bm u}\in W^{2,p}_{\rm loc}(\mathbb R^d;\mathbb R^m)$ for every $p\in[1,\infty)$, that $\bm{\mathcal A}(t){\bm u}\in C_b(\mathbb R^d;\mathbb R^m)$ and that ${\mathcal B}(t){\bm u}=\bm{\mathcal A}(t){\bm u}$ a.e. in $\mathbb R^d$. ``$ \supset$''. Let $\lambda\in {\mathbb R}$ be such that $2\lambda-M_{\{t\}}^P>M_{\{t\}}^P\vee \lambda_{\{t\}}$. We claim that if ${\bm u}\in C_b(\mathbb R^d;\mathbb R^m)\cap\bigcap_{p>1}W^{2,p}_{\rm loc}(\mathbb R^d;\mathbb R^m)$ with $\bm{\mathcal A}(t){\bm u}\in C_b(\mathbb R^d;\mathbb R^m)$ satisfies the equation $\lambda {\bm u}-\bm{\mathcal A}(t){\bm u}=\bm 0$ then ${\bm u}=\bm 0$. Assume that the claim is true, and let $\bm f\in C_b(\mathbb R^d;\mathbb R^m)\cap \bigcap _{p>1}W^{2,p}_{\rm loc}(\mathbb R^d;\mathbb R^m)$ be such that $\bm{\mathcal A}(t)\bm f\in C_b(\mathbb R^d;\mathbb R^m)$. Set $\boldsymbol \phi:=\lambda \bm f-\bm{\mathcal A}(t)\bm f$ and ${\bm g}=R(\lambda, {\mathcal B}(t))\boldsymbol \phi\in D({\mathcal B}(t))$. From the proof of the inclusion ``$\subset$'' it follows that $\lambda(\bm f-{\bm g})-\bm{\mathcal A}(t)(\bm f-{\bm g})=\bm 0$, and the claim gives $\bm f={\bm g}$, i.e., $\bm f\in D({\mathcal B}(t))$. To prove the claim, we fix ${\bm u}$ as above and define the function ${\bm w}$ by setting $w_k:=u_k^2$ for every $k\in\{1,\ldots,m\}$. It follows that $2\lambda w_k =2u_k(\bm{\mathcal A}(t){\bm u})_k \leq \mathcal A_{k,0}(t)w_k+2u_k(C(t,\cdot){\bm u})_k$ for $k=1,\ldots,m$. Arguing as in the proof of Proposition \ref{prop-2.6}(ii) (uniqueness part), we infer that $\mathcal A_{k,0}(t)w_k+2u_k(C(t,\cdot){\bm u})_k \leq (\bm{\mathcal A}^P(t){\bm w})_k+M_{\{t\}}^Pw_k$. By combining these two last estimates we infer that $(2\lambda-M_{\{t\}}^P)w_k-(\bm{\mathcal A}^P(t){\bm w})_k\leq 0$ for $k=1,\ldots,m$. Since $2\lambda-M_{\{t\}}^P>M_{\{t\}}^P\vee \lambda_{\{t\}}$ and $\bm{\mathcal A}^P(t){\bm w}\in C(\mathbb R^d;\mathbb R^m)$, from Proposition \ref{prop:ell_max_princ} it follows that ${\bm w}\leq\bm 0$ on $\mathbb R^d$, which implies ${\bm u}=\bm 0$ on $\mathbb R^d$. Let us sketch the proof of the last part of the statement. Fix $\bm 0\neq \bm f\in D({\mathcal B}(t))$, $\lambda> 1+2(M_{\{t\}}^P)^+$, and let $\bm \phi:=\lambda\bm f-\bm{\mathcal A}(t)\bm f$. From \eqref{resolvent_operator_t} it follows that $\bm f=\int_0^\infty e^{-\lambda \tau}{\bm T}_t(\tau){\bm \phi}d\tau$. From estimate \eqref{stima-12}, we get \begin{align*} \|\bm f\|_{C^\theta_b(\mathbb R^d;\mathbb R^m)} \leq & \frac{\Gamma(1-\theta/2)}{(\lambda-M_{\{t\}}^P)^{1-\theta/2}}\|\bm \phi\|_{C_b(\mathbb R^d;\mathbb R^m)} \\ \leq & \frac{\Gamma(1-\theta/2)}{(\lambda-M_{\{t\}}^P)^{1-\theta/2}}\big (\lambda \|\bm f\|_{C_b(\mathbb R^d;\mathbb R^m)}+\|\bm{\mathcal A}(t)\bm f\|_{C_b(\mathbb R^d;\mathbb R^m)}\big ), \end{align*} and the assertion follows by minimizing with respect to $\lambda>1+2(M_{\{t\}}^P)^+$. \end{proof} To state the main result of this subsection. for $\theta\in (0,1)$ we denote by $C^{0,\theta}_b([s,T]\times\mathbb R^d;{\mathbb R}^m)$ the set of functions $\bm f\in C_b([s,T]\times\mathbb R^d;{\mathbb R}^m)$ such that $\sup_{t\in[s,T]}\|\bm f(t,\cdot)\|_{C^{\theta}_b(\mathbb R^d;{\mathbb R}^m)}<\infty$ and by $C^{0,2+\theta}_b([s,T]\times\mathbb R^d;{\mathbb R}^m)$ the set of functions $\bm f\in C_b([s,T]\times\mathbb R^d;{\mathbb R}^m)$ which are twice continuously differentiable in $[s,T]\times\mathbb R^d$ with respect to the spatial variables, with $\sup_{t\in [s,T]}\|D_{ij}\bm f(t,\cdot)\|_{C^{\theta}_b(\mathbb R^d;{\mathbb R}^m)}<\infty$ for every $i,j=1,\ldots,d$. \begin{thm} \label{teo-Schau-est} Fix $T>s\in I $. The following properties are satisfied. \begin{enumerate}[\rm (i)] \item Let {$\bm f\in C_b^{0,\theta}([s,T]\times \mathbb R^d;\mathbb R^m)$} for some $\theta\in(0,1)$. Then, for every $\lambda>M_{[s,T]}^P$, the elliptic equation \eqref{ellipt} admits a unique solution ${\bm u}\in C_b^{0,2+\theta}([s,T]\times \mathbb R^d;\mathbb R^m)$ and there exists a positive constant $C=C(\lambda,\theta)$ such that \begin{equation} \sup_{t\in[s,T]}\|{\bm u}(t,\cdot)\|_{C^{2+\theta}_b(\mathbb R^d;\mathbb R^m)}\le{C\sup_{t\in[s,T]}\|\bm f(t,\cdot)\|_{C_b^\theta(\mathbb R^d;\mathbb R^m)}}. \label{stima-112} \end{equation} \item Let $\bm f\in C^{2+\theta}_b(\mathbb R^d;\mathbb R^m)$ and ${\bm g}\in C^{0,\theta}_b([s,T]\times\mathbb R^d;\mathbb R^m)$ for some $\theta\in (0,1)$. Then, the Cauchy problem \eqref{parab} admits a unique classical solution ${\bm u}\in C^{0,2+\theta}_b([s,T]\times\mathbb R^d;\mathbb R^m)$ and there exists a positive constant $C$ such that \begin{equation*} \sup_{t\in [s,T]}\|{\bm u}(t,\cdot)\|_{C^{2+\theta}_b(\mathbb R^d;\mathbb R^m)}\le C\Big (\|\bm f\|_{C^{2+\theta}_b(\mathbb R^d;\mathbb R^m)}+\sup_{t\in [s,T]}\|{\bm g}(t,\cdot)\|_{C^{\theta}_b(\mathbb R^d;\mathbb R^m)}\Big ). \end{equation*} {Finally, ${\bm u}\in C_b^{\theta/2}([s,T];C^2(\overline{B_R};\mathbb R^m))$ for every $R>0$.} \end{enumerate} \end{thm} \begin{proof} (i) {\it Uniqueness}. Fix $\lambda\in{\mathbb R}$ such that $2\lambda>(\lambda_{[s,T]}\vee M_{[s,T]})+M_{[s,T]}$ and let ${\bm u}$ solve the equation $\lambda {\bm u}(t,\cdot)-\bm{\mathcal A}(t){\bm u}(t,\cdot)=\bm 0$ with ${\bm u}(t,\cdot)\in D({\mathcal B}(t))$ for every $t\in[s,T]$. Arguing as in the second part of the proof of Proposition \ref{prop:carat_dom_At} we infer that ${\bm u}=\bm 0$. This shows that, for every $\bm f:[s,T]\times \mathbb R^d\rightarrow \mathbb R^m$ with $\bm f(t,\cdot)\in C_b(\mathbb R^d;\mathbb R^m)$, for every $\lambda\in{\mathbb R}$, such that $2\lambda>(\lambda_{[s,T]}\vee M_{[s,T]})+M_{[s,T]}$, and for every $t\in[s,T]$, there exists a unique solution ${\bm v}$ to the equation $\lambda {\bm v}(t,\cdot)-\bm{\mathcal A}(t){\bm v}(t,\cdot)=\bm f(t,\cdot)$ given by ${\bm v}(t,\cdot)=R(\lambda,{\mathcal B}(t))\bm f(t,\cdot)$ for $t\in[s,T]$. Based on this result, we extend the previous property to any $\lambda>M_{[s,T]}^P$. So, let ${\bm u}$ be such that ${\bm u}(t,\cdot)\in D({\mathcal B}(t))$ and satisfy the equation $\lambda{\bm u}(t,\cdot)-\bm{\mathcal A}(t){\bm u}(t,\cdot)=\bm0$ for $t\in[s,T]$. Let $K>0$ be such that $\overline \lambda:=\lambda+K$ satisfies $2\overline \lambda>(\lambda_{[s,T]}\vee M_{[s,T]}^P)+M_{[s,T]}^P$. Clearly ${\bm u}$ solves the equation $\overline \lambda{\bm u}(t,\cdot)-\bm{\mathcal A}(t){\bm u}(t,\cdot)=K{\bm u}(t,\cdot)$ for every $t\in[s,T]$, and so ${\bm u}(t,\cdot)=R(\overline \lambda,{\mathcal B}(t))(K{\bm u}(t,\cdot))$ for every $t\in[s,T]$. We fix $t\in[s,T]$ and consider the operator $\Gamma_{t} :\left(C_b(\mathbb R^d;\mathbb R^m),\|\cdot\|_{\widetilde{\infty}}\right)\rightarrow \left(C_b(\mathbb R^d;\mathbb R^m),\|\cdot\|_{\widetilde{\infty}}\right)$ defined by $\Gamma_{t}({\bm g}):=R(\overline \lambda,{\mathcal B}(t))(K{\bm g})$, where $\|\bm f\|_{\widetilde\infty}$ denotes the maximum over $k=1,\ldots,m$ of the sup-norm of the functions $f_k$. We claim that $\Gamma_{t}$ is a contraction. If the claim is true, then $\Gamma_t$ admits as unique fixed point the null function, and since ${\bm u}(t,\cdot)$ is a fixed point for $\Gamma_t$, it follows that ${\bm u}(t,\cdot)\equiv\bm 0$. The arbitrariness of $t$ gives ${\bm u}\equiv\bm0$ in $[s,T]\times \mathbb R^d$. Let us prove the claim. From the definition of $R(\overline \lambda,{\mathcal B}(t))$, \eqref{stima_uni_gen_auto} and recalling that $\overline \lambda=\lambda+K$, for every ${\bm g}\in C_b(\mathbb R^d;\mathbb R^m)$ we get \begin{align*} \|\Gamma_t({\bm g})\|_{\widetilde \infty} \leq K\|{\bm g}\|_{\widetilde \infty} \int_0^\infty e^{(-\overline \lambda+M_{[s,T]}^P)\tau}d\tau = \frac{K}{\overline\lambda-M_{[s,T]}^P}\|{\bm g}\|_{\widetilde \infty} = \frac{\overline \lambda-\lambda}{\overline\lambda-M_{[s,T]}^P}\|{\bm g}\|_{\widetilde \infty}. \end{align*} Since $\lambda>M_{[s,T]}^P$ it follows that $({\overline \lambda-\lambda)}(\overline\lambda-M_{[s,T]}^P)^{-1}<1$. The claim is so proved and it thus follows that ${\bm u}\equiv\bm 0$ in $[s,T]\times \mathbb R^d$. \vspace{2mm} {\it Existence}. Arguing as in \cite[Theorem 1]{Lun98} (see also \cite[Theorem 3.6]{BerLor05}), and taking advantage of estimate \eqref{stima-12} (which is uniform with respect to $t\in[s,T]$ if we replace $M_{\{t\}}^P$ with $M_{[s,T]}^P$), it follows that for every $\lambda>M_{[s,T]}^P$ and $t\in [s,T]$ there exists a unique solution ${\bm v}(t)\in C_b^{2+\theta}(\mathbb R^d;\mathbb R^m)$ to the equation $\lambda{\bm v}(t)-\bm{\mathcal A}(t){\bm v}(t)=\bm f(t,\cdot)$ such that $\|{\bm v}(t)\|_{C_b^{2+\theta}(\mathbb R^d;\mathbb R^m)}\leq C\|\bm f(t,\cdot)\|_{C_b^\theta(\mathbb R^d;\mathbb R^m)}$ where $C$ is a positive constant, which depends on $\lambda$ and $\theta$, and can be chosen uniform with respect to $t\in[s,T]$. If we set ${\bm u}(t,x)={\bm v}(t)(x)$ for every $t\in[s,T]$ and $x\in\mathbb R^d$, then it follows that ${\bm u}$ satisfies \eqref{stima-112}. Let us prove that ${\bm u}\in C_b^{0,2+\theta}([s,T]\times \mathbb R^d;\mathbb R^m)$. For every $t,\overline t\in[s,T]$ we can write \begin{align*} {\bm u}(t,x)-{\bm u}(\overline t,x) = & \int_0^\infty e^{-\lambda \tau}({\bm T}_t(\tau)\bm f(t,\cdot)(x)-{\bm T}_{\overline t}(\tau)\bm f(\overline t,\cdot)(x))d\tau, \qquad\;\, x\in \mathbb R^d. \end{align*} Let us consider the function under the integral sign and split \begin{align*} {\bm T}_t(\tau)\bm f(t,\cdot)(x)-{\bm T}_{\overline t}(\tau)\bm f(\overline t,\cdot)(x) = & \left({\bm T}_t(\tau)\bm f(t,\cdot)(x)-{\bm T}_{t}(\tau)\bm f(\overline t,\cdot)(x) \right)\\ & +\left({\bm T}_t(\tau)\bm f(\overline t,\cdot)(x)-{\bm T}_{\overline t}(\tau)\bm f(\overline t,\cdot)(x) \right)\\ =: & I_1(t,\tau,x)+I_2(t,\tau,x). \end{align*} We estimate $\|I_1(t,\tau,x)\| \leq \sup_{\sigma\in[s,T]} \|{\bm T}_\sigma(\tau)\bm f(t,\cdot)(x)-{\bm T}_{\sigma}(\tau)\bm f(\overline t,\cdot)(x) \|$ for every $(t,\tau,x)\in [s,T]\times (0,\infty)\times \mathbb R^d$. Next, we fix a compact set $E\subset (0,\infty)\times \mathbb R^d$, a sequence $(t_n)\subset [s,T]$ converging to $\overline t$, and notice that $\bm f(t_n,\cdot)$ converges to $\bm f(\overline t,\cdot)$ locally uniformly on $\mathbb R^d$, as $n$ tends to $\infty$. From Theorem \ref{thm:continuita_totale}(ii) it follows that \begin{align*} \lim_{n\rightarrow \infty}\|I_1(t_n,\cdot,\cdot)\|_{C_b(E;\mathbb R^m)} \leq & \lim_{n\rightarrow\infty}\sup_{\sigma\in[s,T]}\|{\bm T}_{\sigma}(\cdot)\bm f(t_n,\cdot)-{\bm T}_{\sigma}(\cdot)\bm f(\overline t,\cdot) \|_{C_b(E;\mathbb R^m)}=0. \end{align*} The arbitrariness of the sequence $(t_n)$ implies that $\lim_{t\rightarrow \overline t}\|I_1(t,\cdot,\cdot)\|_{C_b(E;\mathbb R^m)}=0$. As far as $I_2$ is concerned, from Theorem \ref{thm:continuita_totale}(i) we infer that for every compact set $E\subset (0,\infty)\times \mathbb R^d$ it holds that $\lim_{t\rightarrow \overline t}\|I_2(t,\cdot,\cdot)\|_{C_b(E;\mathbb R^m)}=0$. By combining the previous two estimates we conclude that \begin{align} \label{stima_cont_ell_integr} \lim_{t\to \overline t}\|{\bm T}_t(\cdot)\bm f(t,\cdot)-{\bm T}_{\overline t}(\cdot)\bm f(\overline t,\cdot)\|_{C_b(E;\mathbb R^m)}=0 \end{align} for every compact set $E\subset (0,\infty)\times \mathbb R^d$. Let us fix $R>0$, then \begin{align} \|{\bm u}(t,\cdot)-{\bm u}(\overline t,\cdot)\|_{C_b({B_R};\mathbb R^m)} \leq \int_0^\infty e^{-\lambda \tau}\|{\bm T}_t(\tau)\bm f(t,\cdot)-{\bm T}_{\overline t}(\tau)\bm f(\overline t,\cdot)\|_{C_b({B_R};\mathbb R^m)}d\tau \label{SFV4S} \end{align} and from \eqref{stima_cont_ell_integr} we infer that the function under the integral sign vanishes as $t$ tends to $\overline t$, pointwise with respect to $\tau\in(0,\infty)$. Further, from \eqref{stima_uni_gen_auto} we infer that \begin{align*} \|{\bm T}_t(\tau)\bm f(t,\cdot)-{\bm T}_{\overline t}(\tau)\bm f(\overline t,\cdot)\|_{C_b({B_R};\mathbb R^m)} \leq 2\sqrt m e^{M_{[s,T]}^P\tau}\sup_{\sigma\in[s,T]}\|\bm f(\sigma,\cdot)\|_\infty, \end{align*} for every $\tau\in(0,\infty)$. Since $\lambda>M_{[s,T]}^P$ the dominated convergence theorem implies that the right-hand side of \eqref{SFV4S} vanishes as $t$ tends to $\overline t$, which shows that ${\bm u}(t,\cdot)$ converges to ${\bm u}(\overline t,\cdot)$ in $C_b({B_R};\mathbb R^m)$ as $t$ tends to $\overline t$, for every $R>0$. Finally, by interpolating between $C_b({B_R};\mathbb R^m)$ and $C_b^{2+\theta}({B_R};\mathbb R^m)$ we get \begin{align} \|{\bm u}(t,\cdot)-{\bm u}(\overline t,\cdot)\|_{C_b^2({B_R};\mathbb R^m)} \!\leq &C\|{\bm u}(t,\cdot)-{\bm u}(\overline t,\cdot)\|_{C_b({B_R};\mathbb R^m)}^{\frac{\theta}{2+\theta}} \|{\bm u}(t,\cdot)-{\bm u}(\overline t,\cdot)\|_{C_b^{2+\theta}({B_R};\mathbb R^m)}^{\frac{2}{2+\theta}} \notag \\ \leq & \widetilde C\bigg (\sup_{\sigma\in[s,T]}\|\bm f(\sigma,\cdot)\|_{C_b^\theta(\mathbb R^d;\mathbb R^m)}\bigg )^{\frac{2}{2+\theta}}\|{\bm u}(t,\cdot)-{\bm u}(\overline t,\cdot)\|_{C_b({B_R};\mathbb R^m)}^{\frac{\theta}{2+\theta}}, \label{stima_cont_der_seconde_ell} \end{align} and the last term of \eqref{stima_cont_der_seconde_ell} tends to $0$ as $t$ approaches $\overline t$. Hence, the second-order spatial derivatives of ${\bm u}$ are bounded and continuous on $[s,T]\times\mathbb R^d$. This completes the proof. \vspace{2mm} (ii) The existence and uniqueness part follows arguing as in \cite[Theorem 2]{Lun98} (see also \cite[Theorem 3.7]{BerLor05}). Since $\bm{\mathcal A}(t){\bm u}(t,\cdot)\in C^\theta_b({B_R};\mathbb R^m)$ for every $t\in[s,T]$, by difference we infer that $t\mapsto D_t{\bm u}(t,\cdot)\in C([s,T];C^\theta_b({B_R};\mathbb R^m))$ for every $R>0$ and \begin{align*} \|D_t{\bm u}(t,\cdot)\|_{C^\theta_b({B_R};{\mathbb R}^m)} \leq & \bigg (\sup_{t\in[s,T]}\|\bm{\mathcal A}(t){\bm u}(t,\cdot)\|_{C_b^\theta(B_R;\mathbb R^m)}+\sup_{t\in[s,T]}\|{\bm g}(t,\cdot)\|_{C_b^{\theta}(B_R;\mathbb R^m)}\bigg ) \\ \leq & C\!\!\!\sup_{t\in[s,T]}\|\bm{\mathcal A}(t)\|_{\theta,B_R}\Big (\|\bm f\|_{C^{2+\theta}_b(\mathbb R^d;\mathbb R^m)}+\sup_{t\in [s,T]}\|{\bm g}(t,\cdot)\|_{C^{\theta}_b(\mathbb R^d;\mathbb R^m)}\Big ) \\ & +\sup_{t\in [s,T]}\|{\bm g}(t,\cdot)\|_{C^{\theta}_b(\mathbb R^d;\mathbb R^m)} \end{align*} for every $t\in [s,T]$ and some positive constant $C$, independent of $t$. Thus, we get \begin{align*} \|{\bm u}(t,\cdot)-{\bm u}(\tau,\cdot)\|_{C^2_b({B_R};\mathbb R^m)} \leq & C \|{\bm u}(t,\cdot)-{\bm u}(\tau,\cdot)\|_{C^\theta_b({B_R};\mathbb R^m)}^{\theta/2}\|{\bm u}(t,\cdot)-{\bm u}(\tau,\cdot)\|_{C^{2+\theta}_b({B_R};\mathbb R^m)}^{1-\theta/2} \\ \leq & C_1|t-\tau|^{\theta/2}, \end{align*} which shows that the function $t\mapsto {\bm u}(t,\cdot)$ belongs to $C^{\theta/2}([s,T];C^2_b({B_R};\mathbb R^m))$. \end{proof} \section{A remarkable consequence of Proposition \ref{prop:confronto_1}} \label{section:special_case} In this section, we prove a relevant consequence of Proposition \ref{prop:confronto_1} in the context of systems of invariant measures. For this purpose, we consider a weakly coupled autonomous version of the operator $\bm{\mathcal A}$ in \eqref{picco}, whose associated semigroup in $C_b(\mathbb R^d;\mathbb R^m)$ is not positive (i.e., each operator ${\bm T}(t)$ does not map the cone of componentwise nonnegative functions into itself). For further use, we introduce the operator ${\mathcal A}={\rm Tr}(QD^2)+\langle {\bm b},\nabla\rangle$. Instead of Hypothesis \ref{hyp-base} we assume the following conditions. \begin{hyp} \label{hyp:special_case} \begin{enumerate}[\rm (i)] \item The coefficients $q_{ij}=q_{ji}$, $b_j$ and $c_{kh}$ belong to $C^\alpha_{\rm loc}(\mathbb R^d)$ for every $i,j=1,\ldots,d$ and $h,k=1,\ldots,m$; \item the infimum $\mu_0$ over $\mathbb R^d$ of the minimum eigenvalue $\mu^k(x)$ of the matrix $Q(x)=(q_{ij}(x))$ is positive; \item $\langle C^P(x)y,y\rangle\le 0$ for every $x\in\mathbb R^d$ and $y\in\mathbb R^m$, where $C^P$ is defined in Section $\ref{section:preliminaries}$; \item there exists a positive function $\varphi\in C^2(\mathbb R^d)$, blowing up as $|x|\to\infty$, such that ${\mathcal A}\varphi(x)\le a-c\varphi(x)$ for every $x\in\mathbb R^d$ and some positive constants $a,c$; \item there exist $\eta\in\mathbb R^m\setminus\{0\}$ such that $\eta \in {\rm Ker}(C(x))$ for every $x\in\mathbb R^d$; \item there does not exist a nontrivial set $K\subset \{1, \ldots, m\}$ such that the coefficients $c_{ij}$ identically vanish on $\mathbb R^d$ for every $i\in K$ and $j\notin K$. \end{enumerate} \end{hyp} From \cite[Proposition 2.4 \& Theorem 2.6]{DeLo11}, it follows that we can associate a semigroup of bounded operators ${\bm T}(t)$ (resp. ${\bm T}^P(t)$) to the operator $\bm{\mathcal A}$ (resp. $\bm{\mathcal A}^P$) and \begin{align} \label{spacial_case_SP_1} \max\{\|{\bm T}(t)\bm f(x)\|^2, \|{\bm T}^P(t)\bm f(x)\|^2\}\leq (T(t)\|\bm f\|^2)(x), \qquad\;\, (t,x)\in[0,\infty)\times \mathbb R^d, \end{align} for every $\bm f\in C_b(\mathbb R^d;\mathbb R^m)$, where $T(t)$ is the semigroup associated to the scalar operator $\mathcal A$. In particular, ${\bm T}(\cdot)\bm f$ (resp. ${\bm T}^P(\cdot)\bm f$) is the unique classical solution to the Cauchy problem associated to operator $\bm{\mathcal A}$ (resp. $\bm{\mathcal A}^P$), which is bounded in every strip $[0,T]\times\mathbb R^d$. \begin{rmk} \begin{enumerate}[(i)] {\rm \item Hypothesis \ref{hyp:special_case}(ii) ensures that also the matrix $C(x)$ is nonpositive for every $x\in \mathbb R^d$. Indeed, for every $x_0\in\mathbb R^d$ and $y\in\mathbb R^m$ it holds that $0\ge \langle C^P(x_0)|y|,|y|\rangle\ge C(x_0)y,y\rangle$, where $|y|=(|y_1|,\ldots,|y_m|)$; \item Hypothesis \ref{hyp:special_case}(iv) implies that there exists a unique invariant measure $\mu$ for the scalar semigroup $T(t)$ associated to the scalar operator $\mathcal A$ (see e.g., \cite[Theorem 9.1.20]{Lo17}); \item differently from \cite[Hypothesis 2.1]{AdAnLo19}, we do not require that the off-diagonal entries of $C$ are nonnegative, and so we do not expect the operator ${\bm T}(t)$ ($t\geq0$) to be positive.} \end{enumerate} \end{rmk} \begin{rmk} \label{rmk:controllo_semigruppo} {\rm Let $\bm f\in C_b(\mathbb R^d;\mathbb R^m)$. Arguing as in the proof of \cite[Theorem 2.6]{DeLo11}, we can show that both ${\bm T}(\cdot)\bm f$ and ${\bm T}^P(\cdot)\bm f$ are the limit in $C^{1+\alpha/2,2+\alpha}(D)$ of the solutions ${\bm u}_n$ to the Cauchy problem with Neumann homogeneous boundary conditions \begin{align*} \left\{ \begin{array}{ll} D_t{\bm v}(t,x)={\bm B}{\bm v}(t,x), & (t,x)\in(0,\infty)\times B_n, \vspace{1mm} \\ \displaystyle \frac{\partial {\bm v}}{\partial \nu}(t,x)={\bm 0}, & (t,x)\in(0,\infty)\times \partial B_n, \vspace{1mm} \\ {\bm v}(0,x)=\bm f(x), & x\in B_n, \end{array} \right. \end{align*} with ${\bm B}=\bm{\mathcal A}$ and ${\bm B}=\bm{\mathcal A}^P$, respectively, for every compact subset $D\subset (0,\infty)\times \mathbb R^d$. Hence, By applying Proposition \ref{prop:confronto_1} to ${\bm T}(t)$ and ${\bm T}^P(t)$ (notice that, under Hypothesis \ref{hyp:special_case}, the coefficients of $\bm{\mathcal A}$ satisfy Hypothesis $\ref{hyp-base}$(i)-(iv) and so, in particular, the assumptions of Proposition \ref{prop:confronto_1} are verified) we infer that $|({\bm T}(t)\bm f)_j(x)|\leq ({\bm T}^P(t)|\bm f|)_j(x)$ for every $\bm f\in C_b(\mathbb R^d;\mathbb R^m)$, $j\in\{1,\ldots,m\}$, $t\geq0$ and $x\in\mathbb R^d$.} \end{rmk} We introduce the set $\mathcal E^P:=\{\bm f\in C_b(\mathbb R^d;\mathbb R^m):{\bm T}^P(t)\bm f=\bm f \ \forall t\geq0\}$. From the proof of Step 1 in \cite[Proposition 3.2]{AdAnLo19} it follows that $\bm f\in \mathcal E^P$ if and only if $\bm f$ is constant and $\bm f\in \bigcap_{x\in \mathbb R^d}{\rm Ker}(C^P(x))$. A similar characterization holds true for the fixed point of the semigroup ${\bm T}(t)$, as the following proposition shows. \begin{prop} \label{prop:punti_fissi_T_eta} $\mathcal E:=\{\bm f\in C_b(\mathbb R^d;\mathbb R^m):{\bm T}(t)\bm f=\bm f \ \forall t\geq0\}={\rm span}\{\eta\}$, where $\eta$ is the vector in Hypothesis $\ref{hyp:special_case}(v)$. Further, the vector $|\eta|$ belongs to $\bigcap_{x\in \mathbb R^d}{\rm Ker}(C^P(x))$. \end{prop} \begin{proof} For sake of convenience we split the proof into three steps. {\it Step 1}. Here, we prove that $\bm f$ belongs to ${\mathcal E}$ if and only if $\bm f\equiv\zeta$ for some $\zeta\in\bigcap_{x\in\mathbb R^d}{\rm Ker}(C(x))$. From \eqref{spacial_case_SP_1} it follows that, if $\bm f\in {\mathcal E}$, then $\|\bm f\|^2=\|{\bm T}(t)\bm f\|^2\leq T(t)\|\bm f\|^2$ in $\mathbb R^d$ for every $t\in (0,\infty)$. The above inequality and the invariance property of $\mu$, which yields $\int_{\mathbb R^d}(T(t)\|\bm f\|^2-\|\bm f\|^2)d\mu=0$ for every $t>0$, imply that, for every $t>0$, $T(t)\|\bm f\|^2=\|\bm f\|^2$ $\mu$- almost everywhere. Since $\mu$ is equivalent to the Lebesgue measure, this is enough to conclude that $\|{\bm T}(t)\bm f\|^2=T(t)\|\bm f\|^2=\|\bm f\|^2$ in $\mathbb R^d$ for every $t>0$. Hence, $\|\bm f\|^2$ is a fixed point of the scalar semigroup $T(t)$ and, consequently, is a constant (see \cite[Proposition 9.1.13]{Lo17}). As a byproduct, we can infer that $0 = D_t\|\bm f\|^2= D_t\|{\bm T}(t)\bm f\|^2 = 2\langle {\bm T}(t)\bm f,D_t{\bm T}(t)\bm f\rangle \le -2\mu_0\|J_x{\bm T}(t)\bm f\|^2$ for every $t>0$, which shows that $\bm f\equiv\zeta$ for some $\zeta\in\mathbb R^m$ and ${\bm 0}= D_t\bm f(x)=(D_t{\bm T}(t)\bm f)(x)=(\bm{\mathcal A}{\bm T}(t)\bm f)(x)=\bm{\mathcal A}\bm f(x)=C(x)\zeta$ for every $x\in\mathbb R^d$, so that $\zeta\in\bigcap_{x\in\mathbb R^d}{\rm Ker}(C(x))$. {\em Step 2.} Here, we fix $\zeta=(\zeta_1,\ldots,\zeta_m)$ in $\bigcap_{x\in\mathbb R^d}{\rm Ker}(C(x))$, and prove that $|\zeta|$ belongs to $\mathcal E^P$. By Step 1, $\zeta$ belongs to $\mathcal E$. Moreover, from Remark \ref{rmk:controllo_semigruppo} we infer that $|({\bm T}(\cdot)\zeta)_j|\le ({\bm T}^P(\cdot)|\zeta|)_j$ in $[0,\infty)\times\mathbb R^d$ for every $j=1,\ldots,m$. By taking \eqref{spacial_case_SP_1} into account we get $\|\zeta\|^2=\|{\bm T}(t)\zeta\|^2\leq \|{\bm T}^P(t)|\zeta|\|^2\le T(t)\|\zeta\|^2=\|\zeta\|^2$ for every $t>0$, which gives $\|\zeta\|^2=\|{\bm T}(t)\zeta\|^2=\|{\bm T}^P(t)|\zeta|\|^2$. Hence, $|\zeta_j|=|({\bm T}(t)\zeta)_j|=({\bm T}^P(t)|\zeta|)_j$ for every $t\geq0$, $x\in{\mathbb R}^d$ and $j=1,\ldots,m$, i.e., $|\zeta|$ belongs to $\mathcal E^P$. Since $\eta\in \bigcap_{x\in\mathbb R^d}{\rm Ker}(C(x))$ it follows that $\bm 0\neq |\eta|$ belongs to $\mathcal E^P\equiv \bigcap_{x\in\mathbb R^d}{\rm Ker}(C^P(x))$. From \cite[Proposition 3.2]{AdAnLo19}, we infer that $\mathcal E^P ={\rm span}\{|\eta|\}=\bigcap_{x\in\mathbb R^d}{\rm Ker}(C^P(x))$ and $|\eta_k|> 0$ for every $k=1,\ldots,m$. {\em Step 3.} Assume by contradiction that there exist two linearly independent vectors $\eta,\zeta\in\mathcal E$. Then, there exist $a\in{\mathbb R}\setminus\{0\}$ and $i,j\in\{1,\ldots,m\}$ such that $(\eta-a\zeta)_i=0$ and $(\eta-a\zeta)_j\neq0$. Since $\eta-a\zeta\in \mathcal E=\bigcap_{x\in{\mathbb R}^d}{\rm Ker}(C(x))$, from Step $2$ we infer that $|\eta-a\zeta|\in \mathcal E^P={\rm span}\{|\eta|\}$, i.e., there exists a non-negative constant $w$ such that $|\eta-a\zeta|=w|\eta|$. Since $|\eta_j-a\zeta_j|\neq0$ it follows that $w\neq0$, and $\eta_i-a\zeta_i=0$ implies that $\eta_i=0$, which contradicts the fact that $|\eta_k|>0$ for every $k=1,\ldots,m$. \end{proof} We now recall the definition of systems of invariant measures for vector-valued semigroups. We say that $\boldsymbol\mu=\{\mu_1,\ldots,\mu_m\}$ is a system of invariant measures for ${\bm T}(t)$ if $\mu_k$ is a bounded Borel measure on $\mathbb R^d$ for every $k=1,\ldots,m$ and \begin{align} \label{def_mis_inv} \sum_{k=1}^m\int_{\mathbb R^d}({\bm T}(t)\bm f)_kd\mu_k=\sum_{k=1}^m\int_{\mathbb R^d}f_kd\mu_k, \qquad \bm f\in C_b(\mathbb R^d;\mathbb R^m), \quad t\geq0. \end{align} By Proposition \ref{prop:punti_fissi_T_eta}, there exists a nontrivial vector $\xi\in \bigcap_{x\in\mathbb R^d}{\rm Ker}(C^P(x))$. Hence, the assumptions in \cite{AdAnLo19} are satisfied and we can use the results therein. In particular, in \cite[Proposition 3.2 \& Theorem 3.6]{AdAnLo19} it has been proved that $\mathcal E^P={\rm span}\{\xi\}$, that all the components of the vector $\xi$ have the same sign (here we assume that $\xi_k>0$ for every $k=1,\ldots,m$), and that a family of bounded Borel measures ${\boldsymbol\mu}^P=\{\mu_1^P,\ldots,\mu_m^P\}$ is a system of invariant measures for ${\bm T}^P(t)$, i.e., formula \eqref{def_mis_inv} holds true with ${\bm T}(t)$ and $\bm\mu$ replaced by ${\bm T}^P(t)$ and ${\bm \mu}^P$, respectively, if and only if there exists a constant $c$ such that $\mu_j=c\xi_j\mu$, where $\mu$ is the unique invariant measure associated to $T(t)$. As a byproduct, for every system of invariant measures $\boldsymbol \mu^P$ for ${\bm T}^P(t)$, the semigroup extends to a strongly continuous semigroup (still denoted by ${\bm T}^P(t)$) on the space $L^p_{{\boldsymbol \mu}^P}(\mathbb R^d;\mathbb R^m)$ of functions $\bm f:\mathbb R^d\rightarrow \mathbb R^m$ Borel measurable such that $f_j\in L^p_{\mu_j^P}(\mathbb R^d)$ for every $j=1,\ldots,m$, endowed with the norm $\|\bm f\|^p_{L^p_{{\boldsymbol\mu}^p}(\mathbb R^d;{\mathbb R}^m)}=\displaystyle\sum_{k=1}^m\int_{\mathbb R^d}|f_k|^pd\mu^P_k$. \begin{rmk} {\rm Hereafter, we normalize $\eta$ and $\xi$. From the proof of Proposition \ref{prop:punti_fissi_T_eta} it follows that $\zeta=|\eta|$ for every $k=1,\ldots,m$, and so $\eta_k\neq 0$ for every $k=1,\ldots,m$.} \end{rmk} The following proposition characterizes the systems of invariant measures for ${\bm T}(t)$. Its proof is analogous to that of \cite[Theorem 3.6]{AdAnLo19}, with slight changes. Hence, it is omitted. \begin{prop} $\bm\mu$ is a system of invariant measures for ${\bm T}(t)$ if and only if there exists $c\in{\mathbb R}$ such that $\mu_k=c\eta_k\mu$ for every $k=1,\ldots,m$, where $\mu$ is the invariant measure associated to $T(t)$. \end{prop} The measures $\mu_k$, $k=1,\ldots,m$, may not have the same sign, since the components of $\eta$ may have different signs. Hence, when we talk about $L^p$-spaces with respect to $\bm \mu$ we are considering the space of the measurable functions $\bm f:\mathbb R^d\rightarrow \mathbb R^m$ such that $\sum_{k=1}^m\int_{\mathbb R^d}|f_k|^pd|\mu_k|<\infty$, where $|\mu_k|$ is the total variation of the measure $\mu_k$, for $k=1,\ldots,m$. We stress that if $\mu_k=c\eta_k\mu$ for some $c\in{\mathbb R}$, then $|\mu_k|=c\xi_k\mu$, i.e., the vector $(|\mu_1|,\ldots,|\mu_m|)$ is a system of invariant measures for ${\bm T}^P(t)$. We are able to prove that for every $t\geq0$ the operator ${\bm T}(t)$ extends to a bounded linear operator on $L^p_{\bm \mu}(\mathbb R^d;\mathbb R^m)$. \begin{prop} ${\bm T}(t)$ extends to a strongly continuous semigroup on $L^p_{\bm \mu}(\mathbb R^d;\mathbb R^m)$ with $\|{\bm T}(t)\|_{L(L^p_{{\boldsymbol\mu}})}\leq 2^{(p-1)/p}$ for every $t\geq0$. \end{prop} \begin{proof} Fix $\bm f\in C_b(\mathbb R^d;\mathbb R^m)$. From the computations in \cite[Proposition 3.12]{AdAnLo19} and Remark \ref{rmk:controllo_semigruppo} we infer that $|({\bm T}(t)\bm f(x))_k|^p \leq |({\bm T}^P(t)|\bm f|(x))_k|^p \leq 2^{p-1}({\bm T}^P(t)(|f_1|^p,\ldots,|f_m|^p)(x))_k$ for every $x\in\mathbb R^d$ and $k=1,\ldots,m$. The invariance of $\widetilde {\boldsymbol\mu}:=(|\mu_1|,\ldots,|\mu_m|)$ with respect to ${\bm T}^P(t)$ gives \begin{align*} \|{\bm T}(t)\bm f\|^p_{L^p_{{\boldsymbol\mu}}} \le 2^{p-1}\sum_{k=1}^m\int_{\mathbb R^d}({\bm T}^P(t)(|f_1|^p,\ldots,|f_m|^p)(x))_kd\widetilde \mu_k= 2^{p-1}\|\bm f\|_{L^p_{{\boldsymbol\mu}}}^p. \end{align*} The strong continuity of ${\bm T}(t)$ follows arguing as in \cite[Proposition 3.12]{AdAnLo19}. \end{proof} \subsection{Asymptotic behaviour} Here, we characterize the asymptotic behaviour of the semigroup ${\bm T}(t)$ as $t$ tends to $\infty$. Throughout this subsection, besides Hypotheses \ref{hyp:special_case}, we assume the following additional conditions. \begin{hyp0} \label{hyp:asym_beha} The coefficients of the operator $\bm{\mathcal A}$ belong to $C^{1+\alpha}_{\rm loc}(\mathbb R^d)$. Moreover, there exists a constant $C>0$ such that $\max\{|q_{ij}(x)|,\langle b(x),x\rangle\}\leq C(1+|x|^2)\varphi(x)$ for every $x\in\mathbb R^d$ and $i,j=1,\ldots,d$. \end{hyp0} The proof of the following proposition is analogous to that of \cite[Proposition 4.3]{AdAnLo19} and so we skip it. \begin{prop} \label{prop:conv_L2norm} For every $\bm f\in C_c^{3+\alpha}(\mathbb R^d;\mathbb R^m)$ the $L^2_\mu$-norm of $\|J_x{\bm T}(t)\bm f\|$ vanishes as $t$ tends to $\infty$. \end{prop} In \cite[Proposition 2.3]{AdAnLo19} it has been proved that, for every $x\in\mathbb R^d$, $0$ is the unique eigenvalue of $C^P(x)$ on the imaginary axis. The proof relies on the Perron-Frobenius theorem for matrices whose entries are nonnegative, and so we cannot directly adapt it to our situation. \begin{lemm} \label{lem:eigenvalue_C_x} For every $x\in\mathbb R^d$, the spectrum of the matrix $C(x)$ is contained in the left half-plane, and $0$ is the unique eigenvalue on the imaginary axis. \end{lemm} \begin{proof} At first, we notice that $C(x)$ has not eigenvalues with positive real part, since $C(x)$ is non positive for every $x\in\mathbb R^d$. Next, we fix $x\in\mathbb R^d$ and let $\lambda>0$ be such that $\lambda+\mu>0$ for every real eigenvalue $\mu=\mu(x)$ of $C(x)$ and for every real eigenvalue $\mu=\mu(x)$ of $C^P(x)$, and $\lambda+c_{ii}(x)>0$ for every $i=1,\ldots,m$. It follows that $|C(x)+\lambda{\rm Id}|= C^P(x)+\lambda{\rm Id}$, where for a matrix $A=(a_{ij})_{i,j=1}^{m}$ we set $|A|:=(|a_{ij}|)_{i,j=1}^{m}$. From \cite[Theorem 8.1.18]{HoJo13} we infer that $r(C(x)+\lambda{\rm Id})\leq r(C^P(x)+\lambda{\rm Id})$, where $r(A)$ denotes the spectral radius of the matrix $A$. Since $0$ is an eigenvalue both of $C(x)$ and $C^P(x)$, $\lambda$ is an eigenvalue both of $C(x)+\lambda{\rm Id}$ and $C^P(x)+\lambda {\rm Id}$. This implies that $r(C(x)+\lambda{\rm Id})\geq \lambda$, and from the Perron-Frobenius Theorem, applied to the matrix $C^P(x)+\lambda{\rm Id}$, whose entries are all nonnegative, we get $r(C^P(x)+\lambda{\rm Id})=\lambda$. This means that $\lambda=r(C^P(x)+\lambda{\rm Id})\geq r(C(x)+\lambda{\rm Id})\geq \lambda$, which gives $r(C(x)+\lambda{\rm Id})=\lambda$. Hence, $C(x)+\lambda{\rm Id}$ has no eigenvalue of the form $\lambda+i \gamma$, $\gamma\neq0$, which implies that $C(x)$ has no eigenvalues of the form $i\gamma$, $\gamma\neq0$. \end{proof} Let $\bm \mu:=\eta \mu$. We set $\mathcal M_{\bm f}:=\sum_{i=1}^m\int_{\mathbb R^d}f_i d\mu_i$ for every $\bm f\in B_b(\mathbb R^d;\mathbb R^m)$. We characterize the asymptotic behaviour of ${\bm T}(t)$. \begin{thm} For every $\bm f\in B_b(\mathbb R^d;\mathbb R^m)$ the function ${\bm T}(t)\bm f$ converges to $\mathcal M_{\bm f}\eta$ locally uniformly in $\mathbb R^d$ as $t$ tends to $\infty$. As a byproduct, ${\bm T}(t)\bm f$ converges to $\mathcal M_{\bm f}\eta$ in $L^p_{\bm \mu}(\mathbb R^d;\mathbb R^m)$ as $t$ tends to $\infty$, for every $p\in[1,\infty)$. \end{thm} \begin{proof} Proposition \ref{prop:conv_L2norm}, Lemma \ref{lem:eigenvalue_C_x} and \cite[Proposition 2.6]{AdAnLo19} allow us to repeat the proof of \cite[Theorem 4.4]{AdAnLo19} to get the assertion. \end{proof} \section{Examples} In this section, we provide two examples of operators to which the results of this paper apply. The operator $\bm{\mathcal A}$ in the first example satisfies Hypotheses \ref{hyp-base} and \ref{hyp:smoothness_coeff}, so that we can apply Theorem \ref{teo-Schau-est} to $\bm{\mathcal A}$, while the operator in the latter example enjoys Hypotheses \ref{hyp:special_case} and \ref{hyp:asym_beha}, so that we can apply Theorem \ref{thm-3.4} to this operator. \begin{example} {\rm Let the coefficients of the operator $\bm{\mathcal A}$ in \eqref{picco} be defined as follows: $q_{ij}^k(t,x)=\zeta_{ij}^k(t)(1+|x|^2)^{\alpha_{ij}^k}$, $b_i^k(t,x)=-\eta_i^k(t)x_i(1+|x|^2)^{\beta_i^k}$ and $c_{kh}(t,x)=\theta_{kh}(t)(1+|x|^2)^{\gamma_{kh}}$ for every $(t,x)\in I\times \mathbb R^d$, $i,j=1,\ldots,d$ and $h,k=1,\ldots,m$. We assume the following additional assumptions. \begin{enumerate}[\rm (a)] \item the functions $\zeta_{ij}^k=\zeta_{ji}^k$, $\eta_i^k$ and $\theta_{kh}$ belong to $C^{\alpha/2}_{\rm loc}(I)$ for some $\alpha\in(0,1)$, every $i,j=1,\ldots,d$ and $h,k=1,\ldots,m$. Moreover, each function $\eta_i^k$ is positive in $I$, and $\alpha_{ij}^k=\alpha_{ji}^k$, $\beta_i^k$ and $\gamma_{kh}$ are nonnegative numbers; \item $\theta_{kk}<0$ in $I$ and $\gamma_{kh}<\gamma_{kk}$ for every $h,k=1,\ldots,m$, and there exists no nontrivial sets $K\subset \{1,\ldots,m\}$ such that $\theta_{kh}$ identically vanishes on $I$ for every $k\in K$ and $h\notin K$; \item for every $k=1,\ldots,m$ it holds that $\alpha_{\min}^k:=\min_{i=1,\dots,d} \alpha_{ii}^k\geq \max_{i\neq j}\alpha_{ij}^k$ and \begin{align} \label{exa_cond_ell} \min_{i=1,\ldots,d}\zeta_{ii}^k-\max_{i=1,\ldots,d}\bigg (\sum_{j\neq i}|\zeta_{ij}^k|^2\bigg )^{\frac{1}{2}}>0\quad {\rm in~} I; \end{align} \item for every $k=1\ldots,m$ it holds that $\max_{i=1,\ldots,d}\alpha_{ii}^k\leq 1+\max_{i=1,\ldots,d}\{\gamma_{kk}, \beta_i^k\}$. \end{enumerate} Under these assumptions, Hypotheses $\ref{hyp-base}$ are satisfied. Indeed, Hypothesis \ref{hyp-base}(i) is immediately verified, and, as far as Hypothesis \ref{hyp-base}(ii) is concerned, we fix a bounded set $J\subset I$ and observe that \begin{align*} & \langle Q^k(t,x)\xi,\xi\rangle \geq (1+|x|^2)^{\alpha_{\min}^k}\bigg [\min_{i=1,\ldots,d}\zeta_{ii}^k(t)-\max_{i=1,\ldots,d}\bigg (\sum_{j\neq i}|\zeta_{ij}^k(t)|^2\bigg )^{\frac{1}{2}}\bigg ]|\xi|^2=:\mu^k(t,x)|\xi|^2 \end{align*} for every $(t,x)\in J\times\mathbb R^d$, $k=1,\ldots,m$ and $\xi\in\mathbb R^d$, and condition \eqref{exa_cond_ell} implies that the infimum over $J\times\mathbb R^d$ of $\mu^k$ is positive. Hypothesis \ref{hyp-base}(iii) is satisfied by the function $\boldsymbol\varphi:\mathbb R^d\to{\mathbb R}^m$, defined by $\boldsymbol\varphi(x):=(1+|x|^2)\mathds 1$ for every $x\in\mathbb R^d$, due to conditions (a)-(c). Finally, Hypotheses \ref{hyp-base}(iv)-(v) follow straightforwardly from condition (b). We now introduce the following more restrictive conditions on $\eta_i^k$ and $\beta_i^k$. \begin{enumerate}[\rm (a$^\prime$)] \item $\eta_i^k=\eta^k$ and $\beta_i^k=\beta^k$ for every $i$ and some $\eta^k$ and $\beta^k$; \item $\alpha_{ij}^k\leq \alpha_{\min}^k+1/2$ for every $i,j=1,\ldots,d$. \end{enumerate} Under these additional assumptions, also Hypotheses \ref{hyp:smoothness_coeff} are satisfied. We just need to verify the last three assumptions in Hypotheses \ref{hyp:smoothness_coeff}. To begin with, we observe that $r^k(t,x)=-\eta^k(t)(1+|x|^2)^{\beta^k}$ for every $(t,x)\in [s,T]\times \mathbb R^d$ and $k=1,\ldots,m$. To prove that Hypothesis \ref{hyp:smoothness_coeff}(iii) is satisfied, it is enough to observe that \begin{align*} |{\rm Tr}(Q^{k}(t,x))|+|(Q^{k}(t,x)x)_i|\leq C(1+|x|^2)^{\max_{j=1,\ldots,d}\alpha_{ij}^k+1/2} \end{align*} for some positive constant $C$, every $x\in\mathbb R^d$ and $i=1,\ldots d$, and $\mu^k(t,x)\sim(1+|x|^2)^{\alpha_{\min}^k}$ as $|x|$ tends to $\infty$ for every $k=1,\ldots,m$. Hence, condition (b$^{\prime}$) allows us to conclude. As far Hypothesis \ref{hyp:smoothness_coeff}(iv) is concerned, we observe that for every $k=1,\ldots,m$, $\ell=2,3$ and $s=1,2,3$ it holds that $b^{k,\ell}\sim (1+|x|^2)^{\beta^k+(1-\ell)/2}$ and $q^{k,s}\sim(1+|x|^2)^{\max_{i,j}\alpha_{ij}^k-s/2}$ as $|x|$ tends to $\infty$. Hence, Hypothesis \ref{hyp:smoothness_coeff}(iv) is fulfilled if $\max_{i,j}\alpha_{ij}^k-1/2\leq \alpha_{\min}^k$ and these inequalities are verified under the conditions (a$^{\prime}$) and (b$^{\prime}$). Finally, since $|M_{k}(t,x)|\sim(1+|x|^2)^{\gamma_{kk}}$ and $c^{k,s}\sim(1+|x|^2)^{\gamma_{kk}-s/2}$ as $|x|$ tends to $\infty$, for $s=1,2,3$, it follows that also Hypothesis \ref{hyp:smoothness_coeff}(v) is satisfied.} \end{example} \begin{example} {\rm Let us assume that the coefficients of the autonomous operator $\bm{\mathcal A}$ are given by $q_{ij}(x)=\zeta_{ij}(1+|x|^2)^{\alpha}$, $b_i(x)=-\theta_ix_i(1+|x|^2)^{\beta}$ and \begin{align*} C(x)=(1+|x|^2)^\gamma\left( \begin{matrix} -1 & 0 & - 1\\ 0 & -3 & \sqrt 3 \\ -1 & \sqrt 3 & -2 \end{matrix} \right) \end{align*} for every $x\in\mathbb R^d$, where the matrix $R=(\zeta_{ij})$ is symmetric and positive definite, $\theta_i$ are all positive constants and $\alpha$, $\beta$, $\gamma$ are nonnegative constants, with $\alpha<\beta+1$. If we consider the function $\varphi:\mathbb R^d\to{\mathbb R}$, defined by $\varphi(x)=(1+|x|^2)^k$ for every $x\in\mathbb R^d$ and some $k=\max\{\alpha-1,1\}$, then Hypotheses \ref{hyp:special_case} and \ref{hyp:asym_beha} are satisfied. Finally, we observe that the eigenvalues of $C$ and of $C^P$ are $0, -3+\sqrt 2$ and $-3-\sqrt 2$, $\eta=(-\sqrt 3,1,\sqrt 3)$ and $\xi=(\sqrt 3,1,\sqrt 3)$.} \end{example}

https://arxiv.org/abs/2012.02843

Kolmogorov operator with the vector field in Nash class

We consider divergence-form parabolic equation with measurable uniformly elliptic matrix and the vector field in a large class containing, in particular, the vector fields in $L^p$, $p>d$, as well as some vector fields that are not even in $L_{\rm loc}^{2+\varepsilon}$, $\varepsilon>0$. We establish Hölder continuity of the bounded soutions, sharp two-sided Gaussian bound on the heat kernel, Harnack inequality.

\section{Introduction} A celebrated result of E.\,De Giorgi \cite{DG} and J.\,Nash \cite{N} states that the bounded solutions of the parabolic equation \begin{equation} \label{eq0} (\partial_t+A)u=0, \quad A=-\nabla \cdot a \cdot \nabla \end{equation} on $[0,\infty[ \times \mathbb R^d$, $d \geq 3$, with measurable matrix \begin{equation} \label{H} \tag{$H_{\sigma,\xi}$} \begin{array}{c} a=a^{\ast}:\mathbb R^d \rightarrow \mathbb R^d \otimes \mathbb R^d, \\ \sigma I \leq a(x) \leq \xi I \quad \text{ for a.e. } x \in \mathbb R^d \quad \text{ for constants $0<\sigma<\xi<\infty$} \end{array} \end{equation} are H\"{o}lder continuous, and the heat kernel $e^{-tA}(x,y)$ satisfies two-sided Gaussian bound with constants that depend only on $d$, $\sigma$, $\xi$. The purpose of this paper is to extend their result to the equation \begin{equation} \label{op} (\partial_t + \Lambda)u=0 \end{equation} where $$ \Lambda= -\nabla \cdot a \cdot \nabla + b \cdot \nabla $$ with $b:\mathbb R^d \rightarrow \mathbb R^d$ in a large class of locally unbounded measurable vector fields. \medskip \textbf{1.~}The existence and the precise form of the relationship between the integral characteristics of the coefficients $a$ and $b$ and the regularity properties of solutions to \eqref{eq0} and \eqref{op} is one of the classical and central problems in the theory of elliptic and parabolic PDEs. By a result of D.\,G.\,Aronson \cite{A}, the heat kernel $e^{-t\Lambda}(x,y)$ of equation \eqref{op} satisfies two-sided Gaussian bound. By a result of S.\,D.\,Eidelman-F.\,O.\,Porper \cite{EP}, $t|\partial_te^{-t\Lambda}(x,y)|$ satisfies the Gaussian upper bound. The constants in their bounds depend on $d$, $\sigma$, $\xi$, and the following integral characteristics of $b$: $$\|b_1\|_p+\|b_2\|_\infty, \quad p>d$$ provided that $b_1+b_2=b$. Our first goal is to demonstrate, based on ideas of E.\,De Giorgi and J.\,Nash, that the constants in the two-sided bound on $e^{-t\Lambda}(x,y)$, in the upper bound on $t|\partial_te^{-t\Lambda}(x,y)|$, as well as H\"older continuity of bounded solutions to \eqref{op} (assuming first that the coefficients $a$, $b$ are smooth) depend in fact on a much finer characteristic of the vector field $b$, that is, on its \textit{elliptic Nash norm}: \begin{equation*} n_e(b,h):= \sup_{x \in \mathbb{R}^d} \int_0^h \sqrt{e^{t\Delta}|b|^2(x) } \; \frac{dt}{\sqrt{t}} \quad (h>0), \end{equation*} and only on its elliptic Nash norm (Theorem \ref{apr_GB}). Next, as is well known, the existence of even strong a priori estimates does not always mean that there is a satisfactory a posteriori regularity theory of the corresponding differential operator. Our second goal is to develop an exhaustive a posteriori theory of \eqref{op}, including two-sided Gaussian bound on the heat kernel of $-\nabla \cdot a \cdot \nabla + b \cdot \nabla$, assuming only that $b$ is measurable, $|b| \in L^2_{{\rm loc}}$ and $$\text{$n_e(b,h)$ is sufficiently small}$$ for some $h>0$ (Theorem \ref{nash_apost_thm}). \begin{definition} A measurable vector field $b:\mathbb R^d \rightarrow \mathbb R^d$ such that $|b| \in L^2_{\rm loc}$ is said to be in the Nash class $\mathbf{N}_e$ if $$n_e(b,h)<\infty$$ for some $h>0$. \end{definition} The class $\mathbf{N}_e$ contains the vector fields $b=b_1+b_2$ with $\|b_1\|_p+\|b_2\|_\infty<\infty$, $p>d$. For such $b$ one has $\lim_{h\downarrow 0} n_e(b,h)= 0$. The class $\mathbf{N}_e$ also contains some vector fields $b$ with $|b|$ not even in $L^{2+\varepsilon}_{{\rm loc}}$, $\varepsilon>0$. See more detailed discussion in Section \ref{main_sect}. The elliptic Nash norm $n_e(b,h)$ was introduced in \cite{S2} where the two-sided Gaussian bound on the heat kernel $e^{-t\Lambda}(x,y)$ was obtained under some additional to $b \in \mathbf{N}_e$ assumptions. If $a=I$ or $a$ is H\"{o}lder continuous, then the condition $|b| \in L^1_{\rm loc}$ and $$ \text{$\kappa_{d+1}(b,h)$ is sufficiently small} $$ for some $h>0$, where \[ \kappa_{d+1}(b,h):=\sup_{x\in\mathbb{R}^d}\int_0^h e^{t\Delta}|b|(x)\frac{dt}{\sqrt{t}} \qquad (\text{Kato norm of $b$}), \] provides the upper Gaussian bound \cite{S}, the Harnack inequality and the lower Gaussian bound on the heat kernel $e^{-t\Lambda}(x,y)$ \cite{Z1}, see also \cite{Z2}. The class of the vector fields $b$ such that $|b| \in L^1_{\rm loc}$ and $$\kappa_{d+1}(b,h)<\infty$$ for some $h>0$ is the well known Kato class $\mathbf{K}^{d+1}$. (The results in \cite{Z1,Z2} were obtained, in fact, for $b=b(t,x)$ in the non-autonomous Kato class, itself introduced by Q.\,S.\,Zhang.) Thus, the Nash class $\mathbf{N}_e$ is an analogue of the Kato class $\mathbf{K}^{d+1}$ in case $a=a(x)$ is only measurable. Note that $\mathbf{N}_e \subset \mathbf{K}^{d+1}$ as is immediate from elementary inequality $e^{t\Delta}|b|(x) \leq \sqrt{e^{t\Delta}|b|^2(x)}$. The principal difference between the cases covered by the Nash class $\mathbf{N}_e$ ($a$ is measurable) and the Kato class $\mathbf{K}^{d+1}$ ($a$ is H\"{o}lder continuous) is as follows. For H\"{o}lder continuous $a$ one can appeal, in the proof of the two-sided bound, to the estimate $|\nabla_x e^{-tA}(x,y)| \leq C t^{-\frac{1}{2}}e^{ct\Delta}(x,y)$, which does not hold for merely measurable $a$; for such $a$ the role of the previous estimate is assumed by far-reaching inequalities $$\mathcal{N}(t)\leq \frac{c_0}{t}, \quad \hat{\mathcal{N}}(t)\leq \frac{\hat{c}_0}{t},$$ where $\mathcal{N}(t)$, $\hat{\mathcal{N}}(t)$ are the so-called Nash's functions similar to $$\langle\nabla_x p\cdot\frac{a(x)}{p}\cdot\nabla_x p\rangle, \quad p \equiv p(t,x,y)=e^{-tA}(x,y)$$ employed by J.\,Nash in \cite{N}. See Sections \ref{app_N} and \ref{apr_GB_sect} for details. We comment more on the relationship between the Nash class and the Kato class in Section \ref{sect3} below. \medskip \textbf{2.~}In the context of the semigroup theory of \eqref{op}, the standard assumption on the vector field $b$ used in the literature is the form-boundedness condition: there exist constants $\delta>0$ and $c(\delta) \geq 0$ such that the quadratic inequality \begin{equation*} \|\sqrt{b \cdot a^{-1} \cdot b}\, f \|_2^2\leq \delta\|A^\frac{1}{2}f\|_2^2+c(\delta)\|f\|_2^2, \end{equation*} holds for all $f \in W^{1,2}$. Briefly, $$b\cdot a^{-1}\cdot b \leq \delta A+c(\delta) \quad \text{ (in the sense of quadratic forms)} $$ (written as $b \in \mathbf{F}_\delta(A)$). This is a large class of singular vector fields containing e.g.\,the vector fields $b=b_1+b_2$ with $|b_1|$ in $L^d$ or in the weak $L^d$ class, $|b_2| \in L^\infty$, see discussion below (before Theorem \ref{grad_thm}). If $b \in \mathbf{F}_\delta(A)$ with $\delta<1$, then the corresponding to $\Lambda=-\nabla \cdot a \cdot \nabla + b \cdot \nabla$ quadratic form on $W^{1,2}$ is quasi $m$-accretive, and so it determines an operator $\Lambda_2$ in $L^2$ generating a holomorphic semigroup. The equation \eqref{op} with $\Lambda=\Lambda_2$ possesses a detailed regularity theory in $L^2$ and, moreover, in $L^p$, $p>\frac{2}{2-\sqrt{\delta}}$, but not in $L^1$. See Section \ref{sect3} for more details. If $b \in \mathbf{N}_e$, then the situation is different: the equation \eqref{op} does not seem to admit any $L^p$ theory for $p>1$ beyond the existence of a semigroup. However, it admits a detailed $L^1$ theory. In Theorem \ref{nash_apost_thm} we construct an operator realization $\Lambda_1$ of the formal operator $\Lambda$ in $L^1$ as the \textit{algebraic sum} $$ \Lambda_1=A_1 + (b \cdot \nabla)_1, \quad D(\Lambda_1)=D(A_1), $$ where $A_1$ is the operator realization of $-\nabla \cdot a \cdot \nabla$ in $L^1$ and $(b \cdot \nabla)_1$ is the closure of $b \cdot \nabla$ in the graph norm of $A_1$, and show that $$ e^{-t\Lambda_1}=s\mbox{-}L^1\mbox{-}\lim_{\varepsilon \downarrow 0}e^{-t\Lambda_1^\varepsilon} \quad \text{({\rm loc}.\,uniformly in $t \geq 0$)} $$ where $ \Lambda_1^\varepsilon=-\nabla \cdot a_\varepsilon \cdot \nabla + b_\varepsilon \cdot \nabla$ of domain $ D(\Lambda_1^\varepsilon)=(1-\Delta)^{-1}L^1$ with smooth ($a_\varepsilon$, $b_\varepsilon$) approximating $(a,b)$ and essentially non-increasing the Nash norm: $$n_e(b_\varepsilon,h) \leq n_e(b,h)+\tilde{c}\varepsilon.$$ Armed with the last results and a priori two-sided Gaussian bound on $e^{-t\Lambda^\varepsilon}(x,y)$ of Theorem \ref{apr_GB}, we develop an exhaustive regularity theory of \eqref{op}, including a posteriori two-sided Gaussian bound on the heat kernel $e^{-t\Lambda}(x,y)$, the Harnack inequality, the H\"{o}lder continuity of bounded solutions of \eqref{op}, the strong Feller property, and the Gaussian upper bound on $t|\partial_te^{-t\Lambda}(x,y)|$ with the optimal (up to a strict inequality) exponent in the Gaussian factor. We also establish the bounds $$\|\nabla(\mu+\Lambda_1)^{-\alpha}\|_{1\rightarrow 1}\leq C\mu^{-\frac{2\alpha-1}{2}}$$ for $\frac{1}{2}<\alpha\leq 1$, $\mu>\mu_0>0$ ($\mu_0$ depends on $d, \sigma,\xi, n_e(b,h)$), and $$\|\nabla e^{-t\Lambda_1}\|_{1 \rightarrow 1} \leq ct^{-\frac{1}{2}}e^{\omega t}, \quad t>0,$$ see Theorem \ref{grad_thm}. We conclude this introduction by mentioning that the condition $b \in \mathbf{F}_\delta(A)$, $\delta<\infty$ provides two-sided Gaussian bounds on the heat kernel of $-\nabla \cdot a \cdot \nabla + b \cdot \nabla$ but only as long as ${\rm div\,}b$ satisfies additional integral constraints (that is, ${\rm div\,}b$ is in the Kato class $\mathbf{K}^d$, cf.\,Section \ref{sect3}), see \cite{KiS5}. \bigskip \tableofcontents \section{Preliminaries} We will need the following standard notations and results. \medskip \textbf{1.~}Let $\mathcal B(X,Y)$ denote the space of bounded linear operators between Banach spaces $X \rightarrow Y$, endowed with the operator norm $\|\cdot\|_{X \rightarrow Y}$. $\mathcal B(X):=\mathcal B(X,X)$. We write $T=s\mbox{-} X \mbox{-}\lim_n T_n$ for $T$, $T_n \in \mathcal B(X,Y)$ if $$\lim_n\|Tf- T_nf\|_Y=0 \quad \text{ for every $f \in X$}. $$ Denote by $[L^p]^d$ and $[L^p]^{d \times d}$ the spaces of the $d$-vectors and the $d \times d$-matrices with entries in $L^p \equiv L^p(\mathbb R^d,dx)$. Put $$ \langle f,g\rangle = \langle f\bar{g}\rangle :=\int_{\mathbb R^d}f\bar{g}dx$$ and $ \|\cdot\|_{p \rightarrow q}=\|\cdot\|_{L^p \rightarrow L^q}. $ $C_{\infty}:=\{f \in C(\mathbb R^d)\mid \lim_{|x| \rightarrow \infty}f(x)=0\}$ endowed with the $\sup$-norm. $\mathcal W^{\alpha,1}$, $\alpha>0$, is the Bessel potential space endowed with norm $\|u\|_{1,\alpha}:=\|g\|_1$, $u=(1-\Delta)^{-\frac{\alpha}{2}}g$, $g \in L^1$. Let $E_\varepsilon f := e^{\varepsilon\Delta} f$ ($\varepsilon>0$), the De Giorgi mollifier of $f$. For a vector field $b$ we put $b^2:=|b|^2$ and $b_a^2:=b \cdot a^{-1} \cdot b$. We write $c \neq c(\varepsilon)$ to emphasize that $c$ is independent of $\varepsilon$. Put \begin{align*} k_\mu(t,x,y) \equiv k(\mu t,x,y) :=(4\pi\mu t)^{-\frac{d}{2}}e^{ - \frac{|x-y|^2}{4\mu t}}, \quad \mu>0. \end{align*} \medskip \textbf{2.~}Let $a \in (H_{\sigma,\xi})$, $0<\sigma<\xi<\infty$. Let $p(t,x,y)$ be the heat kernel of $-\nabla \cdot a \cdot \nabla$ (that is, $p(t,x,y)=e^{-tA}(x,y)$ in the notation of the next section). \begin{theorem} \label{thm_p} Fix constants $0<c_2<\sigma$ and $c_4>\xi$. There exist constants $c_1$, $c_3>0$ that depend only on $d,c_2,c_4$ such that, for all $t>0$, $x,y \in \mathbb{R}^d$, \[ p(t,x,y) \leq c_3 k_{c_4} (t,x-y) \tag{${\rm UGB}^p$} \label{UGB_p} \] and \[ c_1 k_{c_2} (t,x-y) \leq p(t,x,y). \tag{${\rm LGB}^p$} \label{LGB_p} \] Also, for a given $c_6 > \xi$ there is a generic constant $c_5$ depending on $c_6$ such that \[ t |\partial_t p(t,x,y)| \leq c_5 k_{c_6} (t, x-y) \label{GB_pt} \tag{${\rm UGB}^{\partial_tp}$} \] for all $t>0$, $x, y \in \mathbb{R}^d$. \end{theorem} The proof of \eqref{UGB_p} and \eqref{LGB_p} with \textit{some} constants $c_2$ and $c_4$ is due to \cite{A}. The proof of \eqref{GB_pt} with some constant $c_6$ is due to \cite{EP}. The proof of \eqref{UGB_p} and \eqref{GB_pt} in the form as stated is due to \cite{KS}, and in a strengthened form, i.e.\,with polynomial factor, can be found in \cite{Da}. The proof of \eqref{LGB_p} as stated is due to \cite{S}. \medskip \textbf{3.~}Recall that if $S$ and $T$ are linear operators in a Banach space $(Y,\|\cdot\|)$, then $S$ is said to be $T$-bounded if $D(S) \supset D(T)$ and there exist constants $\eta$ and $c$ such that $$\|Sy\| \leq \eta\|Ty\|+c\|y\| \quad \text{for all $y \in D(T)$}. $$ By $T \upharpoonright X$ we denote the restriction of $T$ to a subset $X \subset D(T)$. By $ \big(T \upharpoonright X\big)_{Y \rightarrow Y}^{\rm clos} $ we denote the closure of $T \upharpoonright X$ (when it exists). Next, let operator $T$ be closed. A subset $D_T\subset D(T)$ is called a core of $T$ if $$(T\upharpoonright D_T)_{Y \rightarrow Y}^{\rm clos}=T.$$ Let $P$, $Q$ be linear operators in a Banach space $Y$. Assume that $Q$ is closed, $D(P)$ contains a core $D_Q$ of $Q$ and $\|Py\|\leq \eta\|Qy\|+c\|y\|$, $y\in D_Q$ ($\eta, c$ some constants). This inequality extends by continuity to $D(Q)$. An extension of $P$ obtained in this way, say $\tilde{P}$, is $Q$-bounded. \bigskip \section{Main results} \label{main_sect} \textbf{1.~}We first prove \textit{a priori} Gaussian lower and upper bounds on the heat kernel of $-\nabla \cdot a \cdot \nabla + b \cdot \nabla$, $a \in (H_{\sigma,\xi})$. In what follows, $d \geq 3$. \begin{definition} We say that a constant is generic if it depends only on the dimension $d$ and the constants $\sigma$ and $\xi$. \end{definition} \begin{theorem} \label{apr_GB} Let $a \in (H_{\sigma,\xi})$ be smooth, let $b:\mathbb R^d \rightarrow \mathbb R^d$ be smooth and bounded, $\xi_1>\xi$. There exists a generic constant $\tilde{n}>0$ such that if the Nash norm of $b$ $$n_e(b,h) \equiv \sup_{x \in \mathbb{R}^d} \int_0^h \sqrt{e^{t\Delta}|b|^2(x) } \; \frac{dt}{\sqrt{t}}$$ satisfies $$ n_e(b,h) \leq \tilde{n} $$ for some $h>0$, then there exist positive constants $\sigma_1<\sigma$ and $c_{\sigma_1}$, $c_{\xi_1}>0$, $\omega_i \geq 0,$ $i=1,2,$ such that the heat kernel $u(t,x,y)$ of $-\nabla \cdot a \cdot \nabla + b \cdot \nabla$ satisfies the Gaussian lower and upper bounds \[ c_{\sigma_1} e^{-t\omega_1} k_{\sigma_1} (t,x-y) \leq u(t,x,y) \leq c_{\xi_1} e^{t\omega_2} k_{\xi_1} (t,x-y) \tag{${\rm LUGB}^u$} \label{GB_u} \] for all $t>0$ and $x,y \in \mathbb{R}^d$. The constants $\sigma_1$, $c_{\sigma_1}, c_{\xi_1}, \omega_i$ depend only on $d,\xi_1$ and $n_e(b,h)$. \end{theorem} \begin{definition} We say that a constant is generic* if it depends on $d$, $\sigma$, $\xi$ and on the Nash norm $n_e(b,h)$ of the vector field $b$. \end{definition} Thus, the constants in \eqref{GB_u} are generic*. The fact that they do not depend on the smoothness of $a$, $b$, coupled with the next Proposition \ref{prop1} and a careful approximation argument, will allow us to establish the corresponding a posteriori heat kernel bounds (Theorem \ref{nash_apost_thm}). \medskip \textbf{2.~}Recall that a vector field $b \in [L^2_{{\rm loc}}]^d$ is said to be in the Nash class $\mathbf{N}_e$ if $$n_e(b,h)<\infty$$ for some $h>0$. \begin{example} (1) We have $$\text{$|b| \in L^p$, $p>d$ \quad $\Rightarrow$ \quad $b \in \mathbf{N}_e$},$$ as follows easily using $\|e^{t\Delta}\|_{r \rightarrow \infty} \leq Ct^{-\frac{d}{2r}}$ upon taking $r=\frac{p}{2}$: \begin{align*} \sup_{x \in \mathbb R^d}\int_0^h\sqrt{ e^{t\Delta}|b|^2(x)} \frac{dt}{\sqrt{t}} & \leq \int_0^h\sqrt{\|e^{t\Delta}|b|^2\|_\infty}\frac{dt}{\sqrt{t}} \\ & \leq C^{\frac{1}{2}}\int_0^h\sqrt{t^{-\frac{d}{p}}\|b\|^2_p}\frac{dt}{\sqrt{t}} \\ & =C^{\frac{1}{2}}\frac{2p}{p-d}h^\frac{p-d}{2p}\|b\|_p<\infty. \end{align*} \smallskip (2) There exist $b \in \mathbf{N}_e$ such that, for any $\varepsilon>0$, $|b| \not\in L_{{\rm loc}}^{2+\varepsilon}$, e.g.\,consider $$ |b(x)|=\mathbf{1}_{B(0, e^{-1})}(x)|x_1|^{-\frac{1}{2}}|\log |x_1||^{-\alpha}, \quad \alpha>\frac{1}{2}, $$ where $x=(x_1,\dots,x_d)$. \end{example} \medskip \textbf{3.~}Let $A \equiv A_2$ be the self-adjoint operator in $L^2$ associated with the quadratic form $\langle \nabla u, a \cdot \nabla u\rangle$, $u \in W^{1,2}$. A standard application of the Beurling-Deny theory yields that the operator $A$ generates a symmetric Markov semigroup $e^{-tA}$. Then $$ e^{-tA_1}:=\biggl[e^{-tA} \upharpoonright L^1 \cap L^2 \biggr]_{L^1 \rightarrow L^1}^{\rm clos} \in \mathcal B(L^1), \quad t>0. $$ is a $C_0$ semigroup (this is a general fact from the theory of symmetric Markov semigroups). Its generator $-A_1$ is an appropriate operator realization of the formal operator $-\nabla \cdot a \cdot \nabla$ in $L^1$. Given a vector field $b\in [L^1_{{\rm loc}}]^d$, we define in $L^1$ operator $B_{\max} \supset b\cdot \nabla$ of domain $$D(B_{\max}):=\{f\in L^1 \mid f \in W^{1,1}_{{\rm loc}} \text{ and } b\cdot\nabla f\in L^1\}.$$ The following result will allow us to construct an operator realization of the formal Kolmogorov operator $-\nabla \cdot a \cdot \nabla + b \cdot \nabla$, with $a \in (H_{\sigma,\xi})$ measurable and $b \in \mathbf{N}_e$ locally unbounded, in $L^1$. \medskip \begin{proposition} \label{prop1} Let $b \in \mathbf{N}_e$. Then $D(B_{\max})\supset D(A)\cap D(A_1)$ and $B_{\max}\upharpoonright D(A_1) \cap D(A)$ extends by continuity in the graph norm of $A_1$ to $A_1$-bounded operator $(b \cdot \nabla)_1$: $$ \|(b \cdot \nabla)_1 f\|_1 \leq \eta\|A_1 f\|_1+\eta \mu \|f\|_1, \quad f \in D(A_1), $$ with bound $\eta:=\frac{1}{1-e^{-\mu h}}\sqrt{\frac{c_0}{\sigma c_4}}\;n_e(b,h c_4)$, $\mu > 0$. Here and below, $$c_0:=2 c_3c_5 + \frac{d}{2},$$ where $c_i$ ($i=3,4,5$) are generic constants in the Gaussian bounds on the heat kernel $e^{-tA}(x,y)$ and its time derivative in Theorem \ref{thm_p}. \end{proposition} We will also need the following standard result. Since $e^{-tA_1}$ and $e^{-tA}$ have the same integral kernel $e^{-tA}(x,y)$ which satisfies $|\partial_t e^{-tA}(x,y)| \leq c_5 t^{-1}k_{c_6} (t, x-y)$ (Theorem \ref{thm_p}), there exists a generic constant $C>0$ such that $(CtD_te^{-tA_1})^n$ are uniformly (in $0 \leq t \leq 1$ and $n=1,2\dots$) bounded in $\mathcal B(L^1)$, and so, by a classical result \cite[Ch.\,IX, sect.\,10]{Y}, \begin{equation} \label{__A_1_bd} \|(\zeta + A_1)^{-1}\|_{1 \rightarrow 1} \leq \frac{M}{|\zeta|}, \quad {\rm Re} \zeta>0 \end{equation} with generic constant $M$. \begin{theorem} \label{nash_apost_thm} Let $a \in (H_{\sigma,\xi})$, $b \in \mathbf{N}_e$ with the Nash norm $$n_e(b,hc_4) <\sqrt{\frac{\sigma c_4}{c_0}}$$ for some $h>0$ (the constants $c_0$, $c_4$ were introduced above). The following is true: \smallskip {\rm (\textit{i})} The algebraic sum $\Lambda_1:=A_1 + (b \cdot \nabla)_1$, $D(\Lambda_1)=D(A_1)$ generates a quasi bounded holomorphic semigroup $e^{-t\Lambda_1}$ in $L^1$ with the sector of holomorphy $$\{z \in \mathbb C \mid |{\rm arg\,} z|<\frac{\pi}{2} - \theta\}, \quad \text{ where } \tan \theta=\sqrt{2}\biggl(\frac{M}{1-\sqrt{\frac{c_0}{\sigma c_4}}n_e(b,h c_4)}-1\bigg).$$ The operator $\Lambda_1$ is an operator realization of the formal Kolmogorov operator $-\nabla \cdot a \cdot \nabla + b \cdot \nabla$ in $L^1$. \smallskip {\rm (\textit{ii})} $$ e^{-t\Lambda_1} =s{\mbox-}L^1{\mbox-}\lim_{\varepsilon \downarrow 0}e^{-t\Lambda_1^\varepsilon} \quad (\text{loc.\,uniformly in $t \geq 0$}), $$ where $$\Lambda_1^\varepsilon:=-\nabla \cdot a_\varepsilon \cdot \nabla + b_\varepsilon \cdot \nabla, \quad D(\Lambda_1^\varepsilon)=\mathcal W^{2,1}$$ are the approximating operators, with smooth matrices $a_\varepsilon \in (H_{\sigma,\xi})$ and smooth bounded vector fields $b_\varepsilon$ constructed in such a way that $$a_\varepsilon \rightarrow a \quad \text{ strongly in } [L^2_{{\rm loc}}]^{d \times d}, \quad b_\varepsilon\rightarrow b \quad \text{ strongly in } [L^2_{\rm loc}]^d \quad \text{ as } \varepsilon \downarrow 0,$$ and the Nash norm of $b_\varepsilon$ for all small $\varepsilon>0$ is controlled by the Nash norm of $b$: $$n_e(b_\varepsilon,h) \leq n_e(b,h) + \tilde{c}\varepsilon \quad (\tilde{c}\text{ generic constant}).$$ The semigroup $e^{-t\Lambda_1}$ conserves positivity and is a $L^\infty$ contraction (and so the convergence in {\rm(\textit{ii})} holds for $e^{-t\Lambda_r}$ in $L^r$ for all $1<r<\infty$). \medskip Moreover, there exists a generic constant $\tilde{n}>0$ such that if $n_e(b,h c_4) \leq \tilde{n}$, then we further have: \smallskip {\rm (\textit{iii})} For every $t>0$, $e^{-t\Lambda_1}$ is an integral operator. \smallskip {\rm (\textit{iv})} The heat kernel $e^{-t\Lambda}(x,y)$ ($\equiv$ the integral kernel of $e^{-t\Lambda_1}$) satisfies, possibly after redefinition on a measure zero set in $\mathbb R^d \times \mathbb R^d$, the lower and upper Gaussian bounds: For every $\xi_1>\xi$ there exist generic* constants $\sigma_1\in ]0,\sigma[$ and $c_i >0$, $\omega_i \geq 0$, $i=1,2$ such that \[ c_1 e^{-t \omega_1} k_{\sigma_1} (t,x-y) \leq e^{-t\Lambda}(x,y) \leq c_2 e^{t\omega_2} k_{\xi_1} (t,x-y) \] for all $t>0$, $x, y \in \mathbb{R}^d$. \smallskip {\rm (\textit{v})} $e^{-t\Lambda_1}$ conserves probability: $$ \langle e^{-t\Lambda}(x,\cdot)\rangle=1 \quad \text{ for every } x \in \mathbb R^d. $$ \smallskip {\rm (\textit{vi})} For every $f \in L^1$, $u(t,\cdot):=e^{-t\Lambda_1}f(\cdot)$ is H\"{o}lder continuous (possibly after redefinition on a measure zero set in $\mathbb R^d \times \mathbb R^d$), i.e.\,for every $0<\alpha<1$ there exist generic* constants $C<\infty$ and $\beta \in ]0,1[$ such that for all $z \in \mathbb R^d$, $s>R^2$, $0<R \leq 1$ $$ |u(t,x)-u(t',x')| \leq C \|u\|_{L^\infty([s-R^2,s] \times \bar{B}(z,R))} \biggl(\frac{|t-t'|^{\frac{1}{2}} + |x-x'|}{R} \biggr)^\beta $$ for all $(t,x)$, $(t',x') \in [s-(1-\alpha^2)R^2,s] \times \bar{B}(z,(1-\alpha) R)$. Furthermore, $u \geq 0$ satisfies the Harnack inequality: Let $0<\alpha<\beta<1$ and $\gamma \in ]0,1[$, then there exists a constant $K=K(d,\sigma,\xi,\alpha,\beta,\gamma)<\infty$ such that for all $(s,x) \in ]R^2,\infty[ \times \mathbb R^d$, $0<R \leq 1$ one has $$ u(t,y) \leq K u(s,x) $$ for all $(t,y) \in [s-\beta R^2,s-\alpha^2 R^2] \times \bar{B}(x,\delta R)$. \smallskip {\rm(\textit{vii})} $$e^{-t\Lambda_{C_\infty}}:=\bigl[e^{-t\Lambda_1} \upharpoonright C_\infty \cap L^1\bigr]^{\rm clos}_{C_\infty \rightarrow C_\infty}, \quad t>0$$ is a Feller semigroup in $C_\infty$ having the property $ e^{-t\Lambda_{C_\infty}}[L^\infty \cap L^1] \subset C_\infty, $ $t>0$. Moreover, $$e^{-t\Lambda_{C_u}}f(x):=\langle e^{-t\Lambda}(x,\cdot)f(\cdot)\rangle, \quad t>0$$ is a Feller semigroup on $C_u$, the space of bounded uniformly continuous functions on $\mathbb R^d$. \smallskip {\rm(\textit{viii})} For every $c_6>\xi$ there exists a generic* constant $c_5$ such that $$ |\partial_t e^{-t(\omega_2+\Lambda_1)}(x,y)| \leq c_5 t^{-1} k_{c_6}(t,x-y) $$ for all $t>0$, $x$, $y \in \mathbb R^d$. {\rm(\textit{ix})} For every $1<p<\infty$, $$ e^{-t\Lambda_p}:=\biggl[e^{-t\Lambda_1} \upharpoonright L^1 \cap L^p \biggr]^{\rm clos}_{L^p \rightarrow L^p} $$ is a quasi bounded holomorphic semigroup with the same sector of holomorphy as in {\rm(\textit{i})}. \smallskip {\rm(\textit{x})} For every $\frac{1}{2}<\alpha \leq 1$, $$\|\nabla (\zeta+\Lambda_1)^{-\alpha}\|_{1 \rightarrow 1} \leq C({\rm Re} \zeta)^{-\alpha+\frac{1}{2}}.$$ \end{theorem} \bigskip \textbf{4.~}Recall that a vector field $b$ is said to be form-bounded (with respect to $A \equiv A_2$) if there exist finite constants $\delta>0$ and $c(\delta) \geq 0$ such that the quadratic inequality \begin{equation*} \|b_a f \|_2^2\leq \delta\|A^\frac{1}{2}f\|_2^2+c(\delta)\|f\|_2^2 \end{equation*} is valid for all $f \in D(A^\frac{1}{2}) \equiv W^{1,2}$, where $b_a:=\sqrt{b\cdot a^{-1}\cdot b}$. We write $b \in \mathbf{F}_\delta(A)$. It is easily seen that $$ b \in \mathbf{F}_\delta(-\Delta) \quad \Rightarrow \quad b \in \mathbf{F}_{\delta_a}(A) \text{ with } \delta_a=\sigma^{-2}\delta. $$ The class $\mathbf{F}_\delta(A)$ contains, in particular, the vector fields $$b=b_1+b_2, \quad |b_1| \in L^d, \quad |b_2| \in L^\infty,$$ and for every such $b$ the form-bound $\delta$ can be chosen arbitrarily small. The class $\mathbf{F}_\delta(A)$ also contains vector fields having critical-order singularities. For instance, $$ b(x)=\pm \sqrt{\delta} \frac{d-2}{2}|x|^{-2}x \in \mathbf{F}_\delta(-\Delta) \quad \text{ with } c(\delta)=0 $$ (by Hardy's inequality). More generally, $\mathbf{F}_\delta(A)$ contains the vector fields $b=b_1+b_2$ with $|b_1|$ in the weak $L^d$ class or the Campanato-Morrey class, and $|b_2| \in L^\infty$, with $\delta$ depending on the norm of $|b_1|$ in the respective classes. Moreover, for every $\varepsilon>0$ one can find vector fields $b \in \mathbf{F}_\delta(A)$ such that $|b| \not \in L^{2+\varepsilon}_{{\rm loc}}$. We refer to \cite[sect.\,4]{KiS} for details and other examples. \begin{theorem} \label{grad_thm} Let $d \geq 3$, assume that $b \in \mathbf{N}_e$ with the same norm $n_e(b,h)$ as in Theorem \ref{nash_apost_thm}{\rm(\textit{iii})-(\textit{x})} for some $h>0$. Additionally, assume that $ b \in \mathbf{F}_\beta(-\Delta)$ for some $\beta<\infty$. Then \begin{equation} \label{grad_est} \|\nabla e^{-t\Lambda_1}\|_{1 \rightarrow 1} \leq C t^{-\frac{1}{2}} e^{\omega_2 t}, \quad t>0, \end{equation} with constant $C$ depending on $d$, $\sigma$, $\xi$, $n_e(b,h)$, $\beta$ and $c(\beta)$. \end{theorem} \bigskip \begin{remark} It is not clear how to extend \eqref{grad_est} and the bound in Theorem \ref{nash_apost_thm}(\textit{x}) to \begin{equation} \label{p_bd} \tag{$\ast$} \|\nabla e^{-t\Lambda_p}\|_{p \rightarrow p} \leq C_p t^{-\frac{1}{2}}e^{\nu_p t}, \quad \|\nabla (\zeta + \Lambda_p)^{-1}\|_{p \rightarrow p} \leq c_p ({\rm Re} \zeta)^{-\frac{1}{2}} \end{equation} for \textit{some} $p>1$. Of course, if also $b \in \mathbf{F}_\beta(A)$ with $\beta<1$, then by standard theory $\|\nabla e^{-t\Lambda_2}\|_{2 \rightarrow 2} \leq C_2t^{-\frac{1}{2}}e^{\nu_2 t}$, $t>0$ for constants $C_2$, $\nu_2$ depending on $d$, $\xi$, $\sigma$, $\beta$ and $c(\beta)$, and so \eqref{p_bd} follows by interpolation for all $p \in [1,2]$ (similarly for $\nabla (\zeta + \Lambda_p)^{-1}$). \end{remark} \bigskip \section{Nash's function $\mathcal N_\delta$} \label{app_N} Put $p(t,x,y) \equiv p_\varepsilon(t,x,y):=e^{-tA^\varepsilon}(x,y),$ where $A^\varepsilon:=-\nabla \cdot a_\varepsilon \cdot \nabla$, $a_\varepsilon \equiv E_\varepsilon a$ (the De Giorgi mollifier, see above). Below we write for brevity $a \equiv a_\varepsilon$. Define Nash's function \[ \mathcal N_\delta (t,x) := \big\langle \nabla_\cdot p(t,\cdot,x) \cdot \frac{a(\cdot)}{k_\delta(t,x-\cdot)} \cdot \nabla_\cdot p(t,\cdot,x) \big\rangle, \;\;\; \delta > 0. \] In what follows, we use function $\mathcal N_\delta$ (and its counterpart $\hat{\mathcal N}_\delta$, see Section \ref{apr_GB_sect}) with essentially the same purpose as J.\,Nash did himself in \cite{N}. \begin{proposition} \label{TimeIndCorollary2} If $\delta = c_4$ then there exists a generic constant $c_0$ such that \[ \mathcal N_\delta(t,x) \leq \frac{c_0}{t}, \quad (t,x) \in ]0 , \infty[ \times \mathbb{R}^d. \] \end{proposition} \begin{proof} Write $\mathcal N_\delta = \big\langle \nabla p \cdot \frac{a}{k_\delta} \cdot \nabla p \big\rangle$. Integrating by parts and using the equation $\big( \partial_t + A^\varepsilon \big) p(t,\cdot,x) = 0$, we have \[ \mathcal N_\delta = \big\langle -\partial_t p, \frac{p}{k_\delta} \big\rangle + \big\langle \nabla p \cdot \frac{ap}{k_\delta^2} \cdot \nabla k_\delta \big\rangle. \] Let us show that the RHS is finite. By \eqref{UGB_p}, \eqref{GB_pt} and by our choice of $\delta$, \[ \big| \langle -\partial_t p, \frac{p}{k_\delta} \rangle\big| \leq c_3 c_5 t^{-1} \big \langle \frac{k_{c_6} k_{c_4}}{k_\delta} \big \rangle = \frac{c_3c_5}{t}; \] Due to (${\rm UGB}^p$) and a \textit{qualitative} bound $|\nabla_xp(t,x,y)|\leq Ct^{-1/2}k_c(t,x,y)$ (i.e.\,the constants $C$, $c$ depend on $\varepsilon$), we have $|\big\langle \nabla p \cdot \frac{ap}{k_\delta^2} \cdot \nabla k_\delta \big\rangle|<\infty$ and hence $\mathcal N_\delta<\infty$. By quadratic inequalities and \eqref{UGB_p}, \begin{align*} \big |\big \langle \nabla p \cdot \frac{ap}{k_\delta^2} \cdot \nabla k_\delta \big \rangle \big | & \leq c_3 \mathcal N_\delta^\frac{1}{2} \big\langle \nabla k_\delta \cdot \frac{a}{k_\delta} \left( \frac{k_{c_4}}{k_\delta} \right)^2 \cdot \nabla k_\delta \big\rangle^\frac{1}{2}, \end{align*} \begin{align*} \langle \nabla k_\delta \cdot \frac{a k_{c_4}^2}{k_\delta^3} \cdot \nabla k_\delta \rangle & \leq \xi \big \langle \frac{(\nabla k_\delta)^2}{k_\delta} \big \rangle = \frac{\xi d}{2\delta}\frac{1}{t} < \frac{d}{2}\frac{1}{t}. \end{align*} and so \begin{align*} \mathcal N_\delta \leq 2 \big \langle -\partial_t p, \frac{p}{k_\delta} \big \rangle + c_3^2 \big \langle \nabla k_\delta \cdot \frac{a}{k_\delta} \cdot \nabla k_\delta \big \rangle \leq \frac{c_0}{t}, \quad \text{ where }c_0 = 2 c_3c_5 + \frac{d}{2}. \end{align*} \end{proof} \section{Proof of Theorem \ref{apr_GB}} \label{apr_GB_sect} \subsection{Auxiliary estimates} For a given $\lambda>0$, denote \[ k_\lambda := k_\lambda (\tau-s,y-\cdot) \quad \text{ and } \quad \hat{k}_\lambda := k_\lambda (t-\tau,x-\cdot),\qquad s<\tau<t \] and $$\big \langle \frac{(\nabla k_\lambda )^2}{k_\lambda} \big \rangle := \big \langle \frac{(\nabla_\cdot k_\lambda (\tau-s,y-\cdot) )^2}{k_\lambda (\tau-s,y-\cdot)} \big \rangle.$$ The next three facts are evident: \begin{enumerate}[label=($\mathbf{a_1}$)] \item $$ \big \langle \frac{(\nabla k_\lambda )^2}{k_\lambda} \big \rangle = \frac{d}{2\lambda} \frac{1}{\tau-s} =\big \langle \big(\frac{y-\cdot}{2\lambda (\tau-s)}\big )^2 k_\lambda (\tau-s,y-\cdot) \big \rangle, $$ $$ \big \langle \frac{(\nabla \hat{k}_\lambda )^2}{\hat{k}_\lambda} \big \rangle = \frac{d}{2\lambda} \frac{1}{t-\tau}. $$ \end{enumerate} \begin{enumerate}[label=($\mathbf{a_2}$)] \item If $\lambda < \lambda_1$, then $k_\lambda \leq \big(\frac{\lambda_1}{\lambda} \big)^{\frac{d}{2}} k_{\lambda_1}.$ \end{enumerate} \begin{enumerate}[label=($\mathbf{a_3}$)] \item If $2\delta > c_4$, then $$\frac{k_{c_4}^2}{k_\delta} = \big( \frac{\delta^2}{(2\delta-c_4)c_4} \big)^{\frac{d}{2}} k_{\frac{\delta c_4}{2\delta-c_4}}.$$ \end{enumerate} \medskip \begin{enumerate}[label=($\mathbf{a_4^-}$)] \item $\left\{ \begin{array}{l} 0 < 2\delta < \lambda \\ 0 < \varepsilon <1 \\ 0 < \tau-s < (t -s) \varepsilon \end{array} \right. \Rightarrow \quad \left\{ \begin{array}{l} \hat{k}_\lambda^2 k_\delta \leq c_-^2 k_{\frac{\lambda \delta}{\lambda -2 \delta}} \cdot k_\lambda^2 (t-s,x-y), \\ [3mm] \text{where }c_-:= (1-\varepsilon)^{-d/2} \big(\frac{\lambda}{\lambda-2\delta} \big)^{d/4}. \end{array} \right. $ \end{enumerate} \begin{enumerate}[label=($\mathbf{a_4^+}$)] \item $\left\{ \begin{array}{l} 0 < 2\delta < \lambda \\ \frac{\lambda}{2 (\lambda -\delta)} < \varepsilon < 1 \\ (t-s) \varepsilon < \tau-s < t-s \end{array} \right. \Rightarrow \quad \left\{ \begin{array}{l} \hat{k}_\lambda k_{2\delta}^2 \leq c_+^2 \hat{k}_\frac{\lambda}{r} \cdot k_\lambda^2 (t-s,x-y) , \\ [3mm] \text{where }c_+:= \varepsilon^{-d/2} \big(\frac{\lambda}{2 \delta} \big)^{d/2} r^{-d/2} , r = \frac{2(\lambda-\delta)\varepsilon-\lambda}{\lambda-2 \delta \varepsilon} . \end{array} \right. $ \end{enumerate} \begin{proof}[Proof of $(\mathbf{a_4^-})$] Using $a b \leq a^2 + 4^{-1}b^2$ and $t-\tau \geq (1-\varepsilon)(t-s)$ we have, for any $\alpha \in \mathbb{R}^d$, $\alpha \neq 0$, \begin{align*} & e^{\alpha \cdot (x-y)} \hat{k}^2_\lambda k_\delta = e^{\alpha \cdot (x-\cdot)} \hat{k}^2_\lambda e^{\alpha \cdot (\cdot -y)} k_\delta \\ & \leq (1-\varepsilon)^{-d} \big(4 \pi \lambda (t-s)\big)^{-d} e^{\alpha^2 \frac{\lambda}{2} (t-\tau)} \cdot \big(4 \pi \delta (\tau -s)\big)^{-d/2} e^{\alpha^2 \frac{\lambda}{2} (\tau-s)} e^{- \frac{|\cdot - y|^2}{4 (\tau-s)}\big(\frac{1}{\delta} -\frac{2}{\lambda} \big)} \\ & = (1-\varepsilon)^{-d} \big(\lambda/(\lambda-2 \delta) \big)^{d/2} k_{\frac{\lambda \delta}{\lambda - 2 \delta}} \cdot \big( 4\pi \lambda (t-s)\big)^{-d} e^{\alpha^2 \frac{\lambda}{2} (t-s)} ; \end{align*} Therefore, \[ \hat{k}^2_\lambda k_\delta \leq (1-\varepsilon)^{-d} \big(\lambda/(\lambda-2 \delta) \big)^{d/2} k_{\frac{\lambda \delta}{\lambda - 2 \delta}} \cdot \big( 4\pi \lambda (t-s)\big)^{-d} e^{-\alpha \cdot (x-y) + \alpha^2 \frac{\lambda}{2}(t-s)} \] Set $\alpha = \frac{x-y}{\lambda (t-s)}$. \end{proof} \begin{proof}[Proof of $(\mathbf{a_4^+})$] Using $a b \leq a^2 + 4^{-1}b^2$ and $\varepsilon (t-s) \leq \tau -s$ we have, for any $\alpha \in \mathbb{R}^d$, $\alpha \neq 0$ and $r \in ]0 , 1 [ ,$ \begin{align*} & e^{\alpha \cdot (x-y)} \hat{k}_\lambda k_{2 \delta}^2 = e^{\alpha \cdot (\cdot - y)}k_{2 \delta}^2 e^{\alpha \cdot (x-\cdot)} \hat{k}_\lambda \\ & \leq \varepsilon^{-d}(\lambda /(2 \delta))^d \big(4 \pi \lambda (t-s)\big)^{-d} e^{\alpha^2 \delta (\tau -s)} \cdot \big(4\pi \lambda (t-\tau)\big)^{-d/2} e^{\alpha \cdot (x-\cdot) - \frac{|x-\cdot|^2}{4 \lambda(t-\tau)}(1-r+r)} \\ & \leq \varepsilon^{-d} (\lambda/(2 \delta)^d r^{-d/2} \hat{k}_\frac{\lambda}{r} \cdot \big(4 \pi \lambda (t-s)\big)^{-d} e^{\alpha^2 \delta (\tau-s) +\alpha^2 \frac{\lambda}{1-r} (t-\tau)} ; \end{align*} Using $t-\tau \leq (1-\varepsilon)(t-s)$ and taking into account our choice of $r$ and $\varepsilon ,$ we have \begin{align*} \delta (\tau-s) & + \frac{\lambda}{1-r} (t-\tau) = \delta (t-s) + \big(\frac{\lambda}{1-r} -\delta \big)(t-\tau) \\ & \leq \delta (t-s) + \big(\frac{\lambda}{1-r} -\delta \big) (1-\varepsilon)(t-s) = \frac{\lambda}{2} (t-s). \end{align*} Therefore \[ \hat{k}_\lambda k_{2 \delta}^2 \leq \varepsilon^{-d} (\lambda/(2 \delta)^d r^{-d/2} \hat{k}_\frac{\lambda}{r} \cdot \big(4 \pi \lambda (t-s)\big)^{-d} e^{-\alpha \cdot (x-y) + \alpha^2 \frac{\lambda}{2} (t-s)} . \] Set $\alpha = \frac{x-y}{\lambda (t-s)}$. \end{proof} \subsection{Nash's function $\hat{\mathcal N}_\delta$} Let $p(t,x,y)$ denote the heat kernel of $\partial_t + A^\varepsilon$, $A^\varepsilon \equiv -\nabla \cdot a_\varepsilon \cdot \nabla$. Put for brevity $a \equiv a_\varepsilon$. Define $$ \hat{\mathcal{N}}_\delta(t-\tau,\tau-s,x,y):= \bigg \langle \nabla_\cdot p(\tau-s,\cdot,y) \cdot \frac{a(\cdot) k_\lambda(t-\tau,x,\cdot)}{k_{2\delta}^2 (\tau-s,y,\cdot)} \cdot \nabla_\cdot p(\tau-s,\cdot,y) \bigg \rangle, $$ for all $s<\tau<t$, $x,y \in \mathbb R^d$. \begin{proposition} \label{nash_prop2} Let $c_4, c_6 <2\delta<\lambda$, fix $0<\varepsilon<1$. There exists a generic constant $\hat{c}_0$ such that \begin{equation*} \hat{\mathcal N}_\delta(t-\tau,\tau-s,x,y) \leq \frac{\hat{c}_0}{t-\tau} \end{equation*} for all $t>s$, $(t-s)\varepsilon<\tau-s<t-s$, $x,y \in \mathbb R^d$. \end{proposition} \begin{proof} Write $\hat{\mathcal{N}}_\delta = \big\langle \nabla p \cdot \frac{a \hat{k}_\lambda}{k_{2\delta}^2} \cdot \nabla p \big\rangle$. Integrating by parts and using the equation $\big( \partial_\tau + A^\varepsilon \big) p(\tau-s,\cdot,y) = 0$, we obtain $$ \hat{\mathcal N}_\delta = \big\langle -\partial_\tau p , \frac{\hat{k}_\lambda p}{k_{2\delta}^2}\big\rangle - \big\langle \nabla p \cdot \frac{a p}{k_{2\delta}^2} \cdot \nabla \hat{k}_\lambda\big\rangle + 2 \big\langle \nabla p \cdot \frac{a p \hat{k}_\lambda}{k_{2\delta}^3} \cdot \nabla k_{2\delta}\big\rangle. $$ By quadratic inequalities, \begin{align*} |\big \langle \nabla p \cdot \frac{a p}{k_{2\delta}^2} \cdot \nabla \hat{k}_\lambda \big \rangle| & \leq \frac{1}{4} \hat{\mathcal{N}}_\delta +\big \langle \nabla \hat{k}_\lambda \cdot \frac{a p^2}{k_{2\delta}^2 \hat{k}_\lambda} \cdot \nabla \hat{k}_\lambda \big \rangle \\ & \equiv \frac{1}{4} \hat{\mathcal{N}}_\delta +M_1,\\ 2|\big \langle \nabla p \cdot \frac{a p \hat{k}_\lambda}{k_{2\delta}^3} \cdot \nabla k_{2\delta} \big \rangle| &\leq \frac{1}{4} \hat{\mathcal{N}}_\delta + 4\big \langle \nabla k_{2\delta} \cdot \frac{a p^2 \hat{k}_\lambda }{k_{2\delta}^4} \cdot \nabla k_{2\delta} \big \rangle \\ & \equiv \frac{1}{4} \hat{\mathcal{N}}_\delta + 4 M_2. \end{align*} Therefore, \begin{equation} \label{BulEquation} \tag{$\ast$} \hat{\mathcal{N}}_\delta \leq 2\big\langle -\partial_\tau p , \frac{\hat{k}_\lambda p}{k_{2\delta}^2}\big\rangle + 2M_1 + 8M_2. \end{equation} Let us estimate the terms in the RHS of \eqref{BulEquation}. By \eqref{UGB_p}, \eqref{GB_pt} and by our choice of $\delta$, \begin{align*} \big| \big\langle -\partial_\tau p , \frac{\hat{k}_\lambda p}{k_{2\delta}^2}\big\rangle \big| & \leq c_3 c_5 (\tau-s)^{-1} \big \langle \frac{k_{c_6} k_{c_4} \hat{k}_\lambda}{k_{2\delta}^2} \big \rangle \\ & \leq c_3c_5(\tau-s)^{-1}\biggl(\frac{(2\delta)^2}{c_4c_6} \biggr)^{\frac{d}{2}}\langle \hat{k}_\lambda \rangle = c_3c_5(\tau-s)^{-1}\biggl(\frac{(2\delta)^2}{c_4c_6} \biggr)^{\frac{d}{2}}. \end{align*} Taking into account that $\tau-s > \varepsilon (t-s) \Rightarrow \frac{1}{\tau-s} < \frac{1-\varepsilon}{\varepsilon} \frac{1}{t-\tau}$, we thus obtain $$ \big| \big\langle -\partial_\tau p , \frac{\hat{k}_\lambda p}{k_{2\delta}^2}\big\rangle \big| \leq c_3c_5\biggl(\frac{(2\delta)^2}{c_4c_6} \biggr)^{\frac{d}{2}}\frac{1-\varepsilon}{\varepsilon} \frac{1}{t-\tau}. $$ Next, using ($\mathbf{a_1}$)-($\mathbf{a_3}$), we have: \begin{align*} M_1 & \leq \xi c_3^2 \left\langle \left(\frac{k_{c_4}}{k_{2\delta}} \right)^2 \frac{(\nabla \hat{k}_\lambda)^2}{\hat{k}_\lambda} \right\rangle \\ & \leq \xi c_3^2 \left(\frac{2\delta}{c_4} \right)^d \big \langle \frac{(\nabla \hat{k}_\lambda)^2}{\hat{k}_\lambda} \big \rangle \\ & = \xi c_3^2 \left(\frac{2\delta}{c_4} \right)^d \frac{d}{2\lambda} \frac{1}{t-\tau}. \end{align*} \[ M_2 \leq \xi c_3^2 \bigg\langle \left(\frac{k_{c_4}}{k_{2\delta}} \right)^2 \hat{k}_\lambda (\nabla \log k_{2\delta} )^2 \bigg\rangle, \] where \begin{align*} \left(\frac{k_{c_4}}{k_{2\delta}} \right)^2 & = \left(\frac{2\delta}{c_4} \right)^d \exp \bigg[ - \frac{|y-\cdot |^2}{4 (\tau-s)} \bigg(\frac{1}{c_4} - \frac{1}{2\delta} \bigg)2 \bigg] \\ & = \left(\frac{2\delta}{c_4} \right)^d \exp \bigg[ - \frac{|y-\cdot |^2}{4\gamma (\tau-s)} \bigg], \qquad \gamma:= \frac{\delta c_4}{2\delta -c_4}, \end{align*} \begin{align*} (\nabla \log k_{2\delta})^2 &= \bigg(\frac{y-\cdot}{2(2\delta)(\tau-s)} \bigg)^2 =\frac{|y-\cdot|^2}{4\gamma (\tau-s)} \frac{\gamma}{(2\delta)^2} \frac{1}{\tau-s}. \end{align*} Since $0<\eta < e^\eta$, we have therefore \[ \bigg\langle \left(\frac{k_{c_4}}{k_{2\delta}} \right)^2 \hat{k}_\lambda (\nabla \log k_{2\delta} )^2 \bigg\rangle \leq \left(\frac{2\delta}{c_4} \right)^d \frac{\gamma}{(2\delta)^2} \frac{1}{\tau-s} \langle \hat{k}_\lambda \rangle, \] and so $$ M_2 \leq \xi c_3^2 \left(\frac{2\delta}{c_4} \right)^d \frac{c_4}{(2\delta-c_4)4\delta} \frac{1-\varepsilon}{\varepsilon} \frac{1}{t-\tau}. $$ Substituting the previous estimates into (\ref{BulEquation}), we obtain \[ \hat{\mathcal{N}}_\delta \leq 2\,c_3c_5\biggl(\frac{(2\delta)^2}{c_4c_6} \biggr)^{\frac{d}{2}}\frac{1-\varepsilon}{\varepsilon} \frac{1}{t-\tau} +c_3^2 \left(\frac{2\delta}{c_4} \right)^d \left(2 \cdot \frac{\xi d}{2\lambda} + 8\cdot\frac{2\xi}{4\delta} \cdot \frac{c_4}{2\delta-c_4} \cdot \frac{1-\varepsilon}{\varepsilon} \right) \frac{1}{t-\tau}, \] as claimed. \end{proof} \subsection{Proof of the upper bound} For brevity, $b \equiv b_\varepsilon$. We iterate the Duhamel formula \[ u(t-s,x,y) = p(t-s,x,y) - \int_s^t \langle u(t-\tau,x, \cdot) b(\cdot) \cdot \nabla_\cdot p(\tau-s,\cdot,y) \rangle d\tau. \] We obtain the series \[ l(t-s,x,y):=\sum_{n=0}^\infty (-1)^n u_n (t-s,x,y), \] where $u_0 (t-s,x,y):= p(t-s,x,y)$ and, for $n=1,2,\dots,$ \[ u_n(t-s,x,y):=\int_s^t \langle u_{n-1} (t-\tau,x,\cdot) b(\cdot) \cdot \nabla_\cdot p(\tau-s,\cdot,y) \rangle d\tau. \] In particular, \[ u_1 (t-s,x,y)= \int_s^t \langle p(t-\tau,x,\cdot) b(\cdot) \cdot \nabla_\cdot p(\tau-s,\cdot,y) \rangle d\tau, \] and so \[ |u_1(t-s,x,y)| \leq c_3 \int_s^t \langle k_{c_4} (t-\tau,x-\cdot) | b(\cdot) \cdot \nabla_\cdot p(\tau-s,\cdot,y) | \rangle d\tau. \] \textit{Suppose that we are able to find generic* constants $h >0$ and $C_h <1$ such that the bound: \begin{equation} \int_s^t \langle k_{c_4} (t-\tau,x-\cdot) | b(\cdot) \cdot \nabla_\cdot p(\tau-s,\cdot,y) | \rangle d\tau \leq C_h k_{c_4} (t-s,x-y) \label{star_b_star_N} \tag{$\star^b \star^N$} \end{equation} is valid for all $x,y \in \mathbb{R}^d$ and $0<t-s \leq h$.} Then $|u_1 (t-s,x,y)| \leq c_3 C_h k_{c_4} (t-s,x-y)$, and by induction, \[ |u_n (t-s,x,y)| \leq c_3 \big(C_h\big)^n k_{c_4} (t-s,x-y). \] Therefore, for all $0 < t-s \leq h$ and all $x,y \in \mathbb{R}^d$, the series $l(t-s,x,y)$ is well defined and \[ |l(t-s,x,y) | \leq \frac{c_3}{1-C_h} k_{c_4} (t-s,x-y). \] Repeating the standard argument we conclude that $l$ satisfies the Duhamel formula provided that $0<t-s \leq h$. Then the uniqueness of $u(t-s,x,y)$ implies \[ u=l \;\;\; (0<t-s \leq h), \] and the reproduction property of $u$ implies \[ u(t-s,x,y) \leq \frac{c_3}{1-C_h} e^{(t-s) \omega_h} k_{c_4} (t-s,x-y) \] for all $t-s > h$, where $\omega_h = \frac{1}{h} \log \frac{c_3}{1-C_h}$. Thus, we obtain the upper bound in \eqref{GB_u} of Theorem \ref{apr_GB}. It remains to prove \eqref{star_b_star_N}. Without loss of generality, $s=0$. Set $b_a^2:=b\cdot a^{-1} \cdot b$ and denote \[ \langle k_\mu b_a^2 \rangle:=\langle k_\mu(\tau,y-\cdot) b_a^2(\cdot) \rangle, \qquad \langle \hat{k}_\mu b_a^2 \rangle:=\langle k_\mu(t-\tau,x-\cdot) b_a^2(\cdot) \rangle. \] Set \[ I:= \int_0^t \langle k_\lambda (t-\tau,x-\cdot) |b(\cdot) \cdot \nabla_\cdot p(\tau,\cdot,y) | \rangle d\tau. \] \begin{lemma} \label{ELemma} Fix $\lambda > \xi$ and select constants $\delta$, $c_4$ such that \[ \lambda > 2\delta > c_4 > \xi. \] Let $\frac{\lambda}{2 (\lambda -\delta)} < \varepsilon < 1$, $r = \frac{2 (\lambda - \delta) \varepsilon -\lambda}{\lambda - 2 \delta \varepsilon}$, and let $c_\pm$ be the constants defined in ($\mathbf{a_4^{\pm}}$). Then, for all $x,y \in \mathbb{R}^d$ and $t>0$, \[ I \leq (c_- M^- + c_+ M^+ ) k_\lambda (t,x,y), \] where \begin{align*} M^-:= & \int_0^{t\varepsilon} \sqrt{\big\langle k_{\frac{\lambda \delta}{\lambda-2\delta}} b_a^2 \big\rangle } \sqrt{\frac{c_0}{\tau}}\; d\tau, \\ M^+:= & \int_{t \varepsilon}^t \sqrt{\big\langle \hat{k}_\frac{\lambda}{r} b_a^2 \big\rangle } \sqrt{\frac{\hat{c}_0}{t-\tau}}\; d\tau. \end{align*} \end{lemma} \begin{proof} Using quadratic inequality, we bound $\langle \hat{k}_\lambda |b \cdot \nabla p| \rangle^2$ in two ways: \[ \langle \hat{k}_\lambda |b \cdot \nabla p| \rangle^2 \leq \langle \hat{k}_\lambda^2 k_\delta b_a^2 \rangle \langle \nabla p \cdot \frac{a}{k_\delta} \cdot \nabla p \rangle \] and \[ \langle \hat{k}_\lambda |b \cdot \nabla p| \rangle^2 \leq \langle \hat{k}_\lambda k_{2\delta}^2 b_a^2 \rangle \langle \nabla p \cdot \frac{a \hat{k}_\lambda}{k_{2\delta}^2} \cdot \nabla p \rangle, \] and hence \[ I\equiv \int_0^t \langle \hat{k}_\lambda |b \cdot \nabla p| \rangle \; d \tau \leq I_\varepsilon^- + I_\varepsilon^+ , \] where \begin{align*} I_\varepsilon^- := & \int_0^{t \varepsilon} \sqrt{\langle \hat{k}_\lambda^2 k_\delta b_a^2 \rangle} \sqrt{\langle \nabla p \cdot \frac{a}{k_\delta} \cdot \nabla p \rangle} \; d \tau \\ I_\varepsilon^+ := & \int_{t\varepsilon}^t \sqrt{\langle \hat{k}_\lambda k_{2\delta}^2 b_a^2 \rangle} \sqrt{\langle \nabla p \cdot \frac{a \hat{k}_\lambda}{k_{2\delta}^2} \cdot \nabla p \rangle} \; d \tau \end{align*} Now the assertion of Lemma \ref{ELemma} follows directly from ($\mathbf{a_4^\mp}$) and Propositions \ref{TimeIndCorollary2} and \ref{nash_prop2}. (Here we apply Propositions \ref{TimeIndCorollary2} with $\delta$ chosen as in Proposition \ref{nash_prop2}, but it is not difficult to see, using ($\mathbf{a_3}$), that its proof works for all $\delta>\frac{c_4}{2}$ although with different generic constant $c_0$.) \end{proof} It remains to note that both $M_+$, $M_-$ in Lemma \ref{ELemma} are majorated by $c\,n_e(b,h)$ for appropriate multiple $c>0$. Provided that $n_e(b,h)$ is sufficiently small, i.e.\,so that $C_h:=(c_- + c_+ )c n_e(b,h)<1$, we obtain \eqref{star_b_star_N}. \subsection{Proof of the lower bound} The analysis of the previous section and the Gaussian upper bound \eqref{UGB_p} of Theorem \ref{thm_p} yield for $|x-y|^2 \leq t \leq h$ \begin{align} u(t,x,y) & \geq p(t,x,y)- \sum_{n \geq 1} |u_n (t,x,y)| \notag \\ & \geq c_1 k_{c_2} (t,x-y) - \frac{c_3 C_h}{1-C_h} k_{c_4} (t,x-y) \notag \\ & \geq \left(c_1 c_2^{-\frac{d}{2}} e^{-\frac{1}{4c_2}} - \frac{c_3 C_h}{1-C_h} c_4^{-\frac{d}{2}} \right) ( 4\pi t )^{-\frac{d}{2}} \notag \\ & \equiv r t^{-\frac{d}{2}}, \label{lower_diagonal} \tag{$\ast\ast$} \end{align} where $r>0$ provided that $C_h$ is small enough, i.e. $ \frac{C_h}{1-C_h} < \frac{c_1}{c_3} \left( \frac{c_4}{c_2} \right)^{\frac{d}{2}} e^{-\frac{1}{4c_2}}. $ Now the standard argument (``small gains yield large gain'', see e.g.\,\cite[Theorem 3.3.4]{Da}) yields for all $x,y\in\mathbb{R}^d$, $t>0$, \begin{equation*} u(t,x,y) \geq re^{t\nu_h}t^{-\frac{d}{2}} \exp \left(-\frac{|x-y|^2}{4c_2 t} \right),\quad \nu_h=\frac{1}{h}\log r. \end{equation*} The proof of Theorem \ref{apr_GB} is completed. \bigskip \section{Proof of Proposition \ref{prop1}} \label{prop1_sect} \textbf{1.~}Let $\mathbf{1}_\varepsilon$, $\varepsilon>0$ be the indicator of $\{x \in \mathbb R^d \mid |x| \leq \varepsilon^{-1}, |b(x)| \leq \varepsilon^{-1}\}$. Define $$ b_\varepsilon:=E_{\nu_\varepsilon}(\mathbf{1}_\varepsilon b), $$ where, recall, $E_\nu\equiv e^{\nu\Delta}$, and $\varepsilon$, $\nu_\varepsilon>0$. Define also $ (b^2)_\varepsilon=E_{\nu_\varepsilon}(\mathbf{1}_\varepsilon b^2) $ and set $g_{1,\varepsilon}:=b_\varepsilon -\mathbf{1}_\varepsilon b$ and $g_{2,\varepsilon}:=|(b^2)_\varepsilon-\mathbf{1}_\varepsilon b^2|$. In what follows, we select $\{\nu_\varepsilon \}$ so that $\nu_\varepsilon \downarrow 0$ sufficiently rapidly as $\varepsilon\downarrow 0$ so that $\|g_{1,\varepsilon}\|_2\leq \varepsilon$ and $\|g_{2,\varepsilon}\|_q\leq \varepsilon^2$ for some $q\geq d$. Note that $(b^2)_\varepsilon\leq g_{2,\varepsilon}+b^2$. Since $\|\mathbf{1}_{B(0,R)}(b_\varepsilon - b)\|_2\leq \|g_{1,\varepsilon}\|_2+\|\mathbf{1}_{B(0,R)}(\mathbf{1}_\varepsilon b-b)\|_2$, we have $$ b_\varepsilon\rightarrow b \quad \text{ strongly in } [L^2_{\rm loc}]^d. $$ The Nash norm of $b_\varepsilon$ is controlled by the Nash norm of $b$: \begin{lemma} \label{lem1} $n_e(b_\varepsilon,h) \leq n_e(b,h) + c_d h^{\frac{1}{4}}\varepsilon$, $\varepsilon>0$. \end{lemma} \begin{proof} Clearly, $(b_\varepsilon)^2\leq (b^2)_\varepsilon$, and so \begin{align*} n_e(b_\varepsilon,h) & \equiv \sup_{x \in \mathbb R^d}\int_0^h \sqrt{ e^{t\Delta}(b_\varepsilon)^2(x)} \frac{dt}{\sqrt{t}} \\ & \leq n_e(b,h) + \sup_{x \in \mathbb R^d}\int_0^h\sqrt{ e^{t\Delta}g_{2,\varepsilon}(x)} \frac{dt}{\sqrt{t}}, \end{align*} where \begin{align*} \sup_{x \in \mathbb R^d}\int_0^h\sqrt{ e^{t\Delta}g_{2,\varepsilon}(x)} \frac{dt}{\sqrt{t}} & \leq \int_0^h\sqrt{\|e^{t\Delta}g_{2,\varepsilon}\|_\infty}\frac{dt}{\sqrt{t}} \leq C_d\int_0^h\sqrt{t^{-\frac{d}{2q}}\|g_{2,\varepsilon}\|_q}\frac{dt}{\sqrt{t}} \\ & \leq \sqrt{\|g_{2,\varepsilon}\|_q} C_d\frac{2}{1-\frac{d}{2q}}h^{\frac{1}{2}-\frac{d}{4q}} \leq 4C_d h^\frac{1}{4}\varepsilon. \end{align*} \end{proof} \textbf{2.\,}Now we can give \begin{proof}[Proof of Proposition \ref{prop1}] Set $\delta:=c_4$. We will construct $(b \cdot \nabla)_1$ and prove \begin{equation} \label{__bAest} \|(b \cdot \nabla)_1 g\|_1 \leq \eta\|(\zeta + A_1)g\|_1, \quad g \in D(A_1), \end{equation} with $\eta:=\frac{1}{1-e^{-{\rm Re}\zeta h}}\sqrt{\frac{c_0}{\sigma\delta}}\;n_e(b,h\delta),$ for all ${\rm Re}\zeta > 0$, so taking $\zeta:=\mu>0$ we obtain the assertion of the proposition. \smallskip \textit{Step 1.} Put $B_1^\varepsilon:=[b_\varepsilon \cdot\nabla\upharpoonright C_c^1]^{\rm clos}_{L^1\to L^1}$ of domain $\mathcal{W}^{1,1}$, and $$T_1^\varepsilon:=B_1^\varepsilon (\zeta+A_1^\varepsilon)^{-1}\in \mathcal{B}(L^1),$$ where, recall, $A^\varepsilon_1:=-\nabla \cdot a_\varepsilon \cdot \nabla$, $a_\varepsilon \equiv E_\varepsilon a$, $D(A_1^\varepsilon)=\mathcal W^{2,1}.$ Since $B_1^\varepsilon$ is closed, we can write \[ T_1^\varepsilon f(x)=\int_0^\infty e^{-\zeta t}B_1^\varepsilon e^{-tA_1^\varepsilon}f(x)dt=\int_0^\infty e^{-\zeta t}\langle b_\varepsilon(x)\cdot\nabla_x p_\varepsilon(t,x,\cdot)f(\cdot)\rangle dt,\quad f\in \mathcal{W}^{1,1}. \] Denote $\mu:={\rm Re} \zeta$. We have \begin{align*} \|T_1^\varepsilon f\|_1 &\leq \sum_{j=0}^\infty e^{-j\mu h}\int_{jh}^{(j+1)h}\|B_1^\varepsilon e^{-tA_1^\varepsilon}f\|_1dt\\ &= \sum_{j=0}^\infty e^{-j\mu h}\int_0^{ h}\|B_1^\varepsilon e^{-tA_1^\varepsilon} e^{-jhA_1^\varepsilon}f\|_1dt. \end{align*} By the Fubini Theorem and the Cauchy-Bunyakovsky inequality, \begin{align*} \int_0^h\|B_1^\varepsilon e^{-tA^\varepsilon} e^{-jhA_1^\varepsilon}f\|_1dt &\leq \big\langle\int_0^h \langle |b_\varepsilon(x)\cdot\nabla_x p_\varepsilon(t,x,y)|\rangle_x dt|e^{-jhA_1^\varepsilon}f(y)| \big\rangle_y\\ &\leq \sup_{y\in\mathbb{R}^d}\int_0^h \langle |b_\varepsilon(x)\cdot\nabla_x p_\varepsilon(t,x,y)|\rangle_x dt\|f\|_1 \\ &\leq \sup_{y\in\mathbb{R}^d}\int_0^h \sqrt{\langle k_\delta(t,x-y)(b_\varepsilon\cdot a^{-1}_\varepsilon\cdot b_\varepsilon) (x)\rangle_x}\sqrt{\mathcal{N}_\delta(t,y)}dt\|f\|_1, \end{align*} where $\mathcal{N}_\delta(t,y) \equiv \big\langle \nabla_xp_\varepsilon(t,x,y)\cdot\frac{a_\varepsilon(x)}{k_\delta(t,x-y)}\cdot \nabla_xp_\varepsilon(t,x,y)\big\rangle_x\leq \frac{c_0}{t}$ by Proposition \ref{TimeIndCorollary2}. Therefore, \begin{align*} \int_0^h\|B_1^\varepsilon e^{-tA_1^\varepsilon} e^{-jhA_1^\varepsilon}f\|_1dt &\leq \sqrt{\frac{c_0}{\sigma\delta}}\;n_e(b_\varepsilon,h\delta)\|f\|_1 \\ & (\text{we are applying lemma above}) \\ & \leq \sqrt{\frac{c_0}{\sigma\delta}} \bigl(n_e(b,h\delta) + c_d h^\frac{1}{4}\delta^{\frac{1}{4}}\varepsilon\bigr)\|f\|_1. \end{align*} Thus, \[ \|T_1^\varepsilon f\|_{1}\leq \eta_\varepsilon \|f\|_1, \quad f \in L^1, \quad \eta_\varepsilon:=\eta + \tilde{c} \varepsilon, \quad {\rm Re} \zeta> 0. \] \smallskip \textit{Step 2.}~ Set $Tf:=b \cdot \nabla (\zeta+A)^{-1}f$, $f \in L^2$ and note that $\nabla (\zeta+A^\varepsilon)^{-1} \rightarrow \nabla(\zeta+A)^{-1} $ strongly in $[L^2]^d$. [The proof is standard: For $1 \leq i \leq d$, $f \in W^{-1,2}$, $\|\nabla_i (\zeta+A^\varepsilon)^{-1}f - \nabla_i (\zeta+A)^{-1}f\|_{2} =: M_\varepsilon(f)$, \begin{align*} M_\varepsilon(f) & := \|\nabla_i(\zeta+A^\varepsilon)^{-1}\nabla\cdot (a-a_\varepsilon)\cdot\nabla(\zeta+A)^{-1}f\|_{2} \\ & \leq\|\nabla_i(\zeta+A^\varepsilon)^{-1}\nabla\|_{2\rightarrow 2}\|(a-a_\varepsilon)\cdot\nabla(\zeta+A)^{-1}f\|_{2}, \end{align*} where $\|\nabla_i(\zeta+A^\varepsilon)^{-1}\nabla\|_{2\rightarrow 2} \leq \|\nabla (\zeta+A^\varepsilon)^{-\frac{1}{2}} \|^2_{2 \rightarrow 2} \leq C$, $C \neq C(\varepsilon)$ and $\|(a-a_\varepsilon)\cdot\nabla(\zeta+A)^{-1}f\|_{2} \rightarrow 0$ (e.g.\,using the Dominated Convergence Theorem), so $M_\varepsilon(f) \rightarrow 0$ as $\varepsilon \downarrow 0$, in particular, for $f \in L^2$.] Therefore, since $b_\varepsilon \rightarrow b$ strongly in $[L^2_{{\rm loc}}]^d$, \begin{equation} \label{__Tconv} T^\varepsilon f \rightarrow Tf \quad \text{strongly in $L^1_{\rm loc}$ as $\varepsilon\downarrow 0$.} \end{equation} Passing to a subsequence in $\varepsilon$, if necessary, we have $T^\varepsilon f\rightarrow Tf \;\mathcal{L}^d \text{ a.e.}$ Applying Fatou's Lemma, we have by Step 1, for all $f\in L^1\cap L^2$, \begin{equation} \label{__T_liminf} \|T f\|_1\leq \liminf_\varepsilon\|T^\varepsilon f\|_1 \leq \eta\|f\|_1. \end{equation} Let $T_1$ denote the extension of $T\upharpoonright L^1\cap L^2$ by continuity to $L^1$. \smallskip \textit{Step 3.} Since, by Step 2, $\|b\cdot\nabla(\zeta+A)^{-1}f\|_1\leq \eta \|f\|_1$ for all $f\in L^1 \cap L^2$, ${\rm Re} \zeta>0$, the operator $B:=b\cdot\nabla\upharpoonright D(A_1)\cap D(A):L^1\to L^1$, and $$\|b\cdot\nabla h\|_1\leq \eta\|(\zeta+A_1)h\|_1, \quad h\in D(A_1)\cap D(A).$$ Since $D(A_1)\cap D(A)$ ($=(1+A)^{-1}[L^1 \cap L^2]$) is a core of $A_1$, $B$ extends by continuity in the graph norm of $A_1$ to $A_1$-bounded operator $(b\cdot\nabla)_1$. The proof of Proposition \ref{prop1} is completed. \end{proof} \begin{remark} \label{rem_nonlocal} The proof above can be extended to non-local operators of the type $\Lambda=(\mu -\nabla \cdot a \cdot \nabla)^{\frac{\alpha}{2}} + b \cdot \nabla$, $1<\alpha<2$, with $b$ in an appropriate modification of the elliptic Nash class. That is, assume that $b \in [L^2_{{\rm loc}}]^d$ satisfies $$ \tilde{n}^\alpha(b,\mu)=\sup_{y \in \mathbb R^d}\int_0^\infty e^{- \mu t} \sqrt{e^{t \Delta}|b|^2(y)}\frac{dt}{t^\frac{3-\alpha}{2}}<\infty, \quad \mu>0. $$ Put $T_1^\varepsilon:=b_\varepsilon \cdot \nabla (\mu + A_1^\varepsilon)^{-\frac{\alpha}{2}}$. A key bound $\|T^\varepsilon_1 f\|_1 \leq \tilde{\eta}\|f\|_1$, $f \in L^1$ remains valid with $\tilde{\eta}=\delta^\frac{1-\alpha}{2}\sqrt{\frac{c_0}{\sigma}}\tilde{n}^\alpha(b,\mu\delta^{-1})$. Namely, \begin{align*} \|T_1^\varepsilon f\|_1 & \leq \biggl(\sup_y \int_0^\infty e^{-\mu t}t^{\frac{\alpha}{2}-1} \sqrt{\langle k_\delta(t,y-\cdot)b_a^2(\cdot)\rangle}\sqrt{\mathcal N_\delta(t,y)} dt\biggr) \|f\|_1 \qquad (b_a^2=b \cdot a^{-1}\cdot b)\\ & \leq\delta^\frac{1-\alpha}{2} \sqrt{\frac{c_0}{\sigma}}\tilde{n}^\alpha(b,\mu\delta^{-1})\|f\|_1. \end{align*} Above one can replace $\tilde{n}^\alpha(b,\mu)$ by $n^\alpha(b,h):=\sup_{y \in \mathbb R^d}\int_0^h \sqrt{e^{t \Delta}|b|^2(y)}\frac{dt}{t^\frac{3-\alpha}{2}}$. \end{remark} \section{Proof of Theorem \ref{nash_apost_thm}} \label{nash_apost_thm_sect} In the proof of Proposition \ref{prop1} we established: $T_1^\varepsilon:=b_\varepsilon \cdot \nabla(\zeta+A_1^\varepsilon)^{-1}$, $T_1:=(b \cdot \nabla)_1(\zeta +A_1)^{-1}$, ${\rm Re} \zeta>0$ satisfy $T_1 \in \mathcal{B}(L^1)$ and $$\|T^\varepsilon_1\|_{1 \rightarrow 1} \leq \eta+\tilde{c}\varepsilon, \quad \|T_1\|_{1 \rightarrow 1} \leq \eta.$$ \begin{proposition} \label{prop2} $ T_1=s\mbox{-}L^1\mbox{-}\lim_{\varepsilon \downarrow 0} T_1^\varepsilon. $ \end{proposition} \begin{proof}[Proof of Proposition \ref{prop2}] Under the additional assumption $b^2\in L^1+L^\infty$, the assertion of the proposition is evident (use \eqref{__Tconv} in the proof of Proposition \ref{prop1}). In general one has to employ the separation property of $e^{-tA}$, as is done below. Since $\sup_{\varepsilon>0}\|T_1^\varepsilon\|_{1 \rightarrow 1}, \|T_1\|_{1 \rightarrow 1}<\infty$, it suffices to prove the claimed convergence on $C_c^\infty$. Fix $f\in C_c^\infty$ and then $r>0$ by $B(0,r) \supset {\rm sprt\,} f$. Since by \eqref{__Tconv} $T_1^\varepsilon f\rightarrow T_1f$ strongly in $L^1_{{\rm loc}}$, the required convergence in (\textit{ii}) would follow from \eqref{__T_liminf} once we show that, for every $\theta>0$, there exists $R=R(r,\theta)>0$ such that $$ \|\mathbf{1}_{B^c(0,R)}T_1^\varepsilon f\|_1 \leq \theta \|f\|_1 \quad \text{ for all $\varepsilon>0$ sufficiently small.} $$ Here $B^c(0,R):=\mathbb R^d - B(0,R)$. To prove the latter, we write \begin{equation*} \mathbf{1}_{B^c(0,R)}T_1^\varepsilon f(x)=\int_0^\infty e^{-\zeta t}\langle \mathbf{1}_{B^c(0,R)}(x)b_\varepsilon(x)\cdot\nabla_x p_\varepsilon(t,x,\cdot)f(\cdot)\rangle dt, \end{equation*} where $p_\varepsilon(t,x,y)=e^{-tA_1^\varepsilon}(x,y)$. Put $\mu:={\rm Re} \zeta$. Then \begin{align*} \|\mathbf{1}_{B^c(0,R)}T_1^\varepsilon f\|_1 & \leq \sum_{j=0}^\infty e^{-j\mu h}\int_{jh}^{(j+1)h}\|\mathbf{1}_{B^c(0,R)} B_1^\varepsilon e^{-tA_1^\varepsilon}f\|_1dt\\ & = \sum_{j=0}^\infty e^{-j\mu h}\int_0^{ h}\|\mathbf{1}_{B^c(0,R)} B_1^\varepsilon e^{-tA_1^\varepsilon} e^{-jhA_1^\varepsilon}f\|_1dt \\ &= \sum_{j=0}^\infty e^{-j\mu h}\biggl[\int_0^{ h}\|\mathbf{1}_{B^c(0,R)} B_1^\varepsilon e^{-tA_1^\varepsilon} \mathbf{1}_{B(0,mr)} e^{-jhA_1^\varepsilon}f\|_1dt \\ & + \int_0^{ h}\|\mathbf{1}_{B^c(0,R)} B_1^\varepsilon e^{-tA_1^\varepsilon} \mathbf{1}_{B^c(0,mr)} e^{-jhA_1^\varepsilon}f\|_1dt \biggr]=: \sum_{j=0}^\infty e^{-j\mu h}\big[I_j + J_j\big], \end{align*} where constant $m \geq 1$ is to be chosen. Arguing as in the proof of Step 1 of the proof of Proposition \ref{prop1} and putting $\delta:=c_4$, we obtain, for all $j \geq 0$, \begin{align*} I_j & \leq \sqrt{\frac{c_0}{\sigma\delta}}\sup_{y \in B(0,mr)} \int_0^h \sqrt{\langle k_\delta(t,y,\cdot)\mathbf{1}_{B^c(0,R)}(\cdot)|b_\varepsilon(\cdot)|^2\rangle} \frac{dt}{\sqrt{t}}\,\|e^{-khA_1^\varepsilon} f\|_1 \\ & \leq \biggl(\sqrt{\frac{c_0}{\sigma\delta}}M_R + 4C_d(h\delta)^{\frac{1}{4}}\varepsilon\biggr)\|f\|_1, \end{align*} where $M_R:=\sup_{y \in B(0,mr)} \int_0^h \sqrt{\langle k_\delta(t,y,\cdot)\mathbf{1}_{B^c(0,R)}(\cdot)|b(\cdot)|^2\rangle} \frac{dt}{\sqrt{t}}$, $R>mr$. Clearly, $J_0=0$. For all $j \geq 1$ and $\eta_0=\sqrt{\frac{c_0}{\sigma\delta}}n_e(b,h\delta)$, \begin{align*} J_j & \leq \eta_0 \|\mathbf{1}_{B^c(0,mr)} e^{-jhA_1^\varepsilon}f\|_1 \\ & \text{(we are applying \eqref{UGB_p} to $e^{-jhA_1^\varepsilon}(x,y)$)} \\ & \leq \eta_0 c_3 (4\pi c_4jh)^{-\frac{d}{2}}e^{-\frac{(m-1)^2r^2}{4c_4jh}}\|f\|_1. \end{align*} Thus, we have $$ \|\mathbf{1}_{B^c(0,R)}T_1^\varepsilon f\|_1 \leq \theta\|f\|_1,$$ where $$ \theta:=\biggl(\sqrt{\frac{c_0}{\sigma\delta}}M_R + 4C_d(h\delta)^{\frac{1}{4}}\varepsilon\biggr)\frac{1}{1-e^{-\mu h}} + C_g \sum_{j=1}^\infty e^{-\mu j h}(jh)^{-\frac{d}{2}}e^{-\frac{(m-1)^2r^2}{4c_4jh}}. $$ It is clear that selecting $m$ sufficiently large, we can make the second term in the RHS as small as needed. We are left to prove the convergence $M_R \rightarrow 0$ as $R \rightarrow \infty$. $(a_1)$~Fix $n>0$ by $k_\delta(t,z,y) \leq C_n k_\delta(t,z,0)$ for all $t>0$, $z \in B^c(0,(m+n)r)$, $y \in B(0,mr)$. Then \[ M_R\leq C_n \int_0^h \sqrt{\langle k_\delta(t,0,\cdot)\mathbf{1}_{B^c(0,R)}(\cdot)|b(\cdot)|^2\rangle} \frac{dt}{\sqrt{t}} \quad \forall R>(m+n)r. \] $(a_2)$~Due to $b \in \mathbf{N}_e$ the function $$w_R(t):=\sqrt{\langle k_\delta(t,\cdot,0)\mathbf{1}_{B^c(0,R)}(\cdot)|b(\cdot)|^2\rangle} \frac{1}{\sqrt{t}} $$ is in $L^1([0,h])$ for every $R \geq 1$. Moreover, it is seen from the definition of $w_R$ that for every $0<t_1<t_2\leq h$, $ w_R(t_1) \leq C_{t_1,t_2-t_1}w_R(t_2)$, $C_{t_1,t_2-t_1}<\infty. $ Thus, $w_R(t)$ is finite for all $0<t\leq h$. $(a_3)$~$w_R(t) \rightarrow 0$ as $R \rightarrow \infty$ for every $0<t\leq h$. Indeed, fix $t\in ]0,h]$. Set $v_R(x):=k_\delta(t,x,0)\mathbf{1}_{B^c(0,R)}(x)|b(x)|^2$. For a.e.\,$x \in \mathbb R^d$, $v_R(x) \downarrow 0$ as $R \uparrow \infty$, and $v_R \leq v_1$ a.e.\,on $\mathbb R^d$ for all $R \geq 1$, where $v_1$ is summable. Hence by the Dominated Convergence Theorem, $\langle v_R \rangle \rightarrow 0$ as $R \rightarrow \infty$, and so $w_R(t) \rightarrow 0$ as $R \rightarrow \infty$. $(a_4)$~Due to $(a_3)$ and $w_R \leq w_1$ for $R \geq 1$, the Dominated Convergence Theorem yields $$ \int_0^h w_R(t)dt \rightarrow 0 \quad \text{ as } R \rightarrow \infty. $$ Thus, $M_R \rightarrow 0$ as $R \rightarrow \infty.$ The proof of Proposition \ref{prop2} is completed. \end{proof} \smallskip We are in position to complete the proof of Theorem \ref{nash_apost_thm}. Recall $\delta:=c_4$. \smallskip (\textit{i}) By our assumption on $n_e(b,h\delta)$, there exists $\lambda_0>0$ such that $$\eta:=\frac{1}{1-e^{-\lambda_0 h}}\sqrt{\frac{c_0}{\sigma\delta}}n_e(b,h\delta)<1.$$ By Proposition \ref{prop1}, $\Lambda_1$ is a closed densely defined operator. Using \eqref{__bAest}, we obtain that $$(\zeta+\Lambda_1)^{-1}=(\zeta+A_1)^{-1}(1+T_1)^{-1} \in \mathcal B(L^1), \quad {\rm Re} \zeta>\lambda_0.$$ Using \eqref{__A_1_bd}, we obtain \begin{equation} \label{__theta_bd} \|(\zeta+\Lambda_1)^{-1}\|_{1\to 1}\leq \frac{M}{|\zeta|(1-\eta)}, \quad {\rm Re} \zeta > \lambda_0, \end{equation} completing the proof of the first part of assertion (\textit{i}). To prove the second part of (\textit{i}), note that, in view of \eqref{__theta_bd}, the resolvent $\zeta \mapsto (\zeta+\lambda_0+\Lambda_1)^{-1}=\Theta(\zeta+\lambda_0)$ is holomorphic in the right-half plane ${\rm Re}\zeta>0$ and in $|\zeta-\zeta_0|<\sqrt{2}\bigl(\frac{M}{1-\eta}-1\big)|\zeta_0|$ for every $\zeta_0$ with ${\rm Re} \zeta_0=0$ (see, if needed, the argument in \cite[Ch.\,IX, sect.\,10]{Y}). Thus, $e^{-z(\lambda_0+\Lambda_1)}$ is holomorphic in the sector $$\{z \in \mathbb C \mid |{\rm arg\, z}|<\frac{\pi}{2} - \theta_{\lambda_0}\}, \quad \text{ where } \tan \theta_{\lambda_0}=\sqrt{2}\biggl(\frac{M}{1-\eta}-1\bigg).$$ This completes the proof of assertion (\textit{i}). \smallskip (\textit{ii}) The claimed approximation $\{b_\varepsilon\}$ was constructed in the proof of Proposition \ref{prop1}. Let us show that $$ (\lambda+\Lambda_1^\varepsilon)^{-1} \rightarrow (\lambda+\Lambda_1)^{-1} \quad \text{strongly in $L^1$ as $\varepsilon \downarrow 0$}, $$ which, by a standard result, implies the convergence of the semigroups. Since $(\lambda+\Lambda_1^\varepsilon)^{-1}=(\lambda+A_1^\varepsilon)^{-1}(1+T^\varepsilon_1)^{-1}$, $(\lambda+\Lambda_1)^{-1}=(\lambda+A_1)^{-1}(1+T_1)^{-1}$, it suffices to show that 1) $T_1^\varepsilon \rightarrow T_1 $ and 2) $(\lambda+A_1^\varepsilon)^{-1} \rightarrow (\lambda+A_1)^{-1}$ strongly in $L^1$ as $\varepsilon \downarrow 0$. 1) is Proposition \ref{prop2}. 2) follows immediately from $$(\lambda+A^\varepsilon)^{-1} \rightarrow (\lambda+A)^{-1} \quad \text{strongly in $L^2$}$$ and $(\lambda+A^\varepsilon)^{-1}(x,y) \leq C(\lambda-c\Delta)^{-1}(x,y)$ for generic constants $0<c,C<\infty$, an immediate consequence of \eqref{UGB_p}. \medskip (\textit{iii}) The upper bound in \eqref{GB_u} of Theorem \ref{apr_GB} yields $$ \|e^{-t\Lambda_1^\varepsilon}\|_{1 \rightarrow \infty} \leq c_2e^{t\omega_2}t^{-\frac{d}{2}}, \quad t>0, \quad \varepsilon>0 $$ with generic* constants $c_2$, $\omega_2<\infty$. Using Theorem \ref{nash_apost_thm}(\textit{ii}) and applying Fatou's lemma, we obtain $ \|e^{-t\Lambda_1}\|_{1 \rightarrow \infty} \leq c_2e^{t\omega_2}t^{-\frac{d}{2}}$, $t>0. $ Hence $e^{-t\Lambda_1}$ is an integral operator for every $t>0$. \smallskip (\textit{iv}) The a priori bounds \eqref{GB_u} of of Theorem \ref{apr_GB}, and Theorem \ref{nash_apost_thm}(\textit{ii}), yield for every pair of bounded measurable subsets $S_1$, $S_2 \subset \mathbb R^d$: $$ c_1 e^{t\omega_1} \langle \mathbf{1}_{S_1},e^{t\sigma_1\Delta}\mathbf{1}_{S_2} \rangle \leq \langle \mathbf{1}_{S_1},e^{-t\Lambda_1}\mathbf{1}_{S_2} \rangle \leq c_2 e^{t\omega_2} \langle \mathbf{1}_{S_1},e^{t\xi_1\Delta}\mathbf{1}_{S_2} \rangle. $$ Since $e^{-t\Lambda_1}$ is an integral operator for every $t>0$, assertion (\textit{iv}) follows by applying the Lebesgue Differentiation Theorem. \smallskip (\textit{v}) For every $\varepsilon>0$, $\langle e^{-t\Lambda^\varepsilon}(x,\cdot)\rangle=1$, $x \in \mathbb R^d$. Fix $t>0$ and $\Omega \subset \mathbb R^d$, a bounded open set. By the upper bound \eqref{GB_u} of Theorem \ref{apr_GB}, for every $\gamma>0$ there exists $R=R(\gamma,t,\Omega)>0$ such that, for every $x \in \Omega$, $ \langle e^{-t\Lambda^\varepsilon}(x,\cdot)\mathbf{1}_{B^c(0,R)}(\cdot)\rangle<\gamma$, so $ \langle e^{-t\Lambda^\varepsilon}(x,\cdot)\mathbf{1}_{B(0,R)}(\cdot)\rangle \geq 1 - \gamma. $ Hence $$ \langle \mathbf{1}_\Omega e^{-t\Lambda^\varepsilon}\mathbf{1}_{B(0,R)}\rangle \geq (1 - \gamma)|\Omega|. $$ Applying Theorem \ref{nash_apost_thm}(\textit{ii}), we obtain $$ \frac{1}{|\Omega|}\langle \mathbf{1}_\Omega e^{-t\Lambda}1\rangle \geq \frac{1}{|\Omega|}\langle \mathbf{1}_\Omega e^{-t\Lambda}\mathbf{1}_{B(0,R)}\rangle\geq 1 - \gamma. $$ Applying the Lebesgue Differentiation Theorem, we obtain $\langle e^{-t\Lambda}(x,\cdot)\rangle \geq 1-\gamma$ for a.e.\,$x \in \mathbb R^d$. In turn, the opposite inequality $\langle e^{-t\Lambda}(x,\cdot)\rangle \leq 1$ for a.e.\,$x \in \mathbb R^d$ follows easily using Theorem \ref{nash_apost_thm}(\textit{ii}), and hence $1\geq \langle e^{-t\Lambda}(x,\cdot)\rangle \geq 1-\gamma$. The proof of (\textit{v}) is completed. \smallskip (\textit{vi}) Put $u_\varepsilon(t,x):=e^{-t\Lambda^\varepsilon}f(x)$. Repeating the argument in \cite[sect.\,3]{FS} which appeals to the ideas of E.\,De Giorgi, we obtain assertion (\textit{vi}) for $u_\varepsilon$. The result now follows upon applying Theorem \ref{nash_apost_thm}(\textit{ii}) and the Arzel\`{a}-Ascoli Theorem. \smallskip (\textit{vii}) follows from (\textit{iv}), (\textit{v}) and (\textit{vi}) using a standard argument for mollifiers. \smallskip (\textit{viii}) is proved repeating the argument in \cite[sect.\,2]{Da}. \smallskip (\textit{ix}) follows repeating the argument in \cite{Ou}. \smallskip (\textit{x}) In the proof of (\textit{i}) we obtain the resolvent representation as the K.\,Neumann series $$(\zeta+\Lambda_1)^{-1}=(\zeta+A_1)^{-1}(1+T_1)^{-1} \in \mathcal B(L^1), \quad {\rm Re}\zeta \geq \lambda_0,$$ where $\lambda_0=\lambda_0\big(n_e(b,h)\big)>0$, $T_1:=(b \cdot \nabla)_1(\zeta +A_1)^{-1} \in \mathcal{B}(L^1)$. The latter yields $\|\nabla (\zeta + \Lambda_1)^{-1}\|_{1 \rightarrow 1} \leq c ({\rm Re} \zeta)^{-\frac{1}{2}}$. Indeed, $\|\nabla(\zeta+A_1)^{-1}\|_{1\to 1}\leq c({\rm Re} \zeta)^{-\frac{1}{2}}$ (integrating ($\star$) in $t \in [0,\infty[$ in the proof of Theorem \ref{grad_thm}), so the resolvent representation yields the required bound. The latter now easily yields the case $1/2<\alpha<1$. \bigskip \section{Proof of Theorem \ref{grad_thm}} \label{grad_thm_sect} It suffices to carry out the proof on $C^\infty_c$ for smooth bounded $a \in (H_{\sigma,\xi})$, $b$, and then apply Theorem \ref{nash_apost_thm}(\textit{ii}) using the closedness of the gradient. First, let $0<t\leq h$. The Duhamel formula for $\nabla e^{-t\Lambda_1}$ yields: \begin{equation} \label{duhamel} \|\nabla e^{-t\Lambda_1}f\|_{1} \leq \|\nabla e^{-tA_1}f\|_1+\int_0^t \|\nabla e^{-(t-\tau)A_1}\|_{1 \rightarrow 1} \|b \cdot \nabla e^{-\tau \Lambda_1}f\|_1 d\tau, \quad f\in C^\infty_c. \end{equation} We will need (proved below): \[ \|\nabla e^{-tA_1}\|_{1 \rightarrow 1} \leq C/\sqrt{t},\tag{$\star$} \] \[ \int_0^t\frac{C}{\sqrt{t-\tau}}\|b \cdot \nabla e^{-\tau \Lambda_1}f\|_1 d\tau \leq C\sup_{x\in\mathbb{R}^d} \int_0^t \frac{1}{\sqrt{t-\tau}} \sqrt{ e^{\delta \tau \Delta} b_a^2(x)} \sqrt{\mathcal N^u_\delta(\tau,x)}d\tau\, \|f\|_1,\tag{$\star\star$} \] \[ \mathcal N^u_\delta(\tau,x) \leq \frac{C_2}{\tau}, \tag{$\star\star\star$} \] where $\mathcal N^u_\delta(\tau,x) :=\langle\nabla u(\tau,x,\cdot)\cdot\frac{a(\cdot)}{k_\delta(\tau,x,\cdot)}\cdot\nabla u(\tau,x,\cdot)\rangle$, $u(\tau,x,y)=e^{-\tau\Lambda}(x,y)$, $\delta>\xi$, the constants $C_1$, $C_2$, $\omega$ are generic. We estimate the RHS of ($\star\star$): write $\int_0^t=\int_0^{t/2}+\int_{t/2}^t$ and use ($\star\star\star$) to obtain \begin{align*} \sup_{x\in\mathbb{R}^d} \int_0^{t/2} \frac{1}{\sqrt{t-\tau}} \sqrt{ e^{\delta \tau \Delta} b_a^2(x)} \sqrt{\mathcal N^u_\delta(\tau,x)}d\tau & \leq \frac{\sqrt{2 C_2}}{\sqrt{t}}\sup_{x\in\mathbb{R}^d}\int_0^{t/2} \sqrt{ e^{\delta \tau \Delta} b_a^2(x)} \frac{d\tau}{\sqrt{\tau}} \\ & \leq \frac{\sqrt{2 C_2}}{\sqrt{\delta t}} n_e(b,\frac{\delta h}{2}), \end{align*} \begin{align*} \sup_{x\in\mathbb{R}^d} \int_{t/2}^t \frac{1}{\sqrt{t-\tau}} \sqrt{ e^{\delta \tau \Delta} b_a^2(x)} \sqrt{\mathcal N^u_\delta(\tau,x)}d\tau & \leq \sqrt{C_2}\sup_{x\in\mathbb{R}^d}\int_{t/2}^t \frac{1}{\sqrt{t-\tau}}\sqrt{ e^{\delta \tau \Delta} b_a^2(x)} \frac{d\tau}{\sqrt{\tau}} \\ & (\text{we are using $e^{\delta \tau\Delta}b_a^2(x) \leq \frac{\xi d\beta }{8\delta }\frac{1}{\tau}+c(\beta)$ since $b \in \mathbf{F}$}) \\ & \leq \tilde{C}\int_{t/2}^t \frac{1}{\sqrt{t-\tau}}\frac{d\tau}{\tau} \leq \tilde{C}\frac{1}{\sqrt{t}}. \end{align*} Substituting ($\star$), ($\star\star$) and the last two estimates into \eqref{duhamel}, we have $\|\nabla e^{-t\Lambda_1}\|_{1 \rightarrow 1} \leq \frac{c}{\sqrt{t}}$ for $0<t \leq h$. Also, for all $t > h$, $\|\nabla e^{-t\Lambda_1}\|_{1 \rightarrow 1} \leq \|\nabla e^{-h\Lambda_1}\|_{1 \rightarrow 1}\|e^{-(t-h)\Lambda_1}\|_{1 \rightarrow 1} \leq \frac{\tilde{c}}{\sqrt{h}}e^{(t-h)\omega_2}$ (cf.\,Theorem \ref{nash_apost_thm}). The latter yields the assertion of Theorem \ref{grad_thm} for all $t>0$. \medskip It remains to prove ($\star$)-($\star\star\star$). Proof of ($\star$): We have for $\mathsf{h} \in \mathbb R^d$, $\mathsf{h}=(0,\dots,1,\dots,0)$ ($1$ is in the $i$-th coordinate, $1 \leq i \leq d$) \begin{align*} \|\mathsf{h} \cdot \nabla e^{-tA_1}f\|_1 & \leq \sup_{x \in \mathbb R^d}\sqrt{\langle k_\delta(t,x,\cdot) (\mathsf{h} \cdot a^{-1}(\cdot) \cdot \mathsf{h})\rangle}\sqrt{\mathcal N_\delta(t,x)}\|f\|_1 \\ & \leq \sigma^{-\frac{1}{2}}\sup_{x \in \mathbb R^d}\sqrt{ \mathcal N_\delta(t,x)}\|f\|_1 = \sigma^{-\frac{1}{2}}\sqrt{\sup_{x \in \mathbb R^d} \mathcal N_\delta(t,x)}\|f\|_1, \end{align*} and so by Proposition \ref{TimeIndCorollary2} $$ \|\nabla e^{-tA_1}f\|_1 \leq \frac{d \sqrt{\sigma^{-1}c_0}}{\sqrt{t}}\|f\|_1. $$ The estimate ($\star\star$) follows using quadratic inequality. Thus, we are left to prove $(\star\star\star)$. Integrating by parts, using the equation for $u(t,x,y)$ and $(\rm {UGB}^u), ({\rm UGB}^{\partial_tu})$ (see Theorem \ref{nash_apost_thm}\textit{(iv),(viii)}), we obtain for $0<t\leq h$ (below $c$ is a generic constant) \[ \mathcal{N}_\delta^u(t,x)=\langle\nabla u\cdot\frac{a}{k_\delta}\cdot\nabla u\rangle=-\langle k^{-1}_\delta u\partial_t u\rangle-\langle k^{-1}_\delta ub\cdot\nabla u\rangle+\langle uk_\delta^{-2}\nabla k_\delta\cdot a\cdot\nabla u\rangle, \] \[ |\langle k^{-1}_\delta u\partial_t u\rangle|\leq \frac{c}{t}, \quad |\langle uk_\delta^{-2}\nabla k_\delta\cdot a\cdot\nabla u\rangle|\leq c|\langle\nabla k_\delta\cdot\frac{a}{k_\delta}\cdot \nabla u\rangle|. \] Clearly, \[ |\langle \nabla k_\delta\cdot \frac{a}{k_\delta}\cdot\nabla u\rangle|\leq \frac{c}{\sqrt{t}}\sqrt{\mathcal{N}_\delta^u(t,x)}. \] \[ |\langle k^{-1}_\delta ub\cdot\nabla u\rangle|\leq c\sqrt{e^{\delta t\Delta}b_a^2(x)}\sqrt{\mathcal{N}_\delta^u(t,x)}\leq \hat{c}\frac{1}{\sqrt{t}}\sqrt{\mathcal{N}_\delta^u(t,x)} \] (due to $e^{\delta t\Delta}b_a^2(x) \leq \frac{\xi d\beta }{8\delta }\frac{1}{t}+c(\beta)$, see above). Now $(\star\star\star)$ is evident. The proof of Theorem \ref{grad_thm} is completed. \bigskip \section{Comments} \label{sect3} \textbf{1.~}The following result was proved in \cite{KiS} (the reader can compare it with Theorem \ref{nash_apost_thm}). It establishes quantitative dependence of the regularity properties of solutions to $(\partial_t + \Lambda)u=0$ with $b \in \mathbf{F}_\delta(A)$ on the value of $\delta$. \begin{theorem} \label{thm_fb} Let $d \geq 3$. Assume that $b \in \mathbf{F}_\delta(A)$ for some $0 < \delta < 4.$ Set $r_c := \frac{2}{2-\sqrt{\delta}}$ and $b_a^2:=b \cdot a^{-1} \cdot b \in L^2_{{\rm loc}}$. The following is true: \smallskip {\rm (\textit{i})} Let $\mathbf 1_n$ denote the indicator of $\{x \in \mathbb R^d \mid \; b_a(x) \leq n \}$ and set $b_n := \mathbf 1_n b.$ Then the limit \[ s\mbox{-}L^r\mbox{-}\lim_{n\rightarrow \infty} e^{-t \Lambda_r(a,b_n)}, \quad r \in I_c^o := ]r_c, \infty [, \] where $\Lambda_r(a,b_n):=A_r + b_n \cdot \nabla$, exists locally uniformly in $t \geq 0$ and determines a positivity preserving, $L^\infty$ contraction, quasi contraction $C_0$ semigroup on $L^r$, say, $e^{-t \Lambda_r(a,b)}$. \smallskip {\rm (\textit{ii})} Define \[ e^{-t \Lambda_{r_c}(a,b)} := \big[e^{-t \Lambda_r(a,b)} \upharpoonright L^1 \cap L^r \big]^{\rm clos}_{L^{r_c} \rightarrow L^{r_c} }, \quad \quad r \in I_c^o. \] Then $e^{-t \Lambda_{r_c}(a,b)}$ is a $C_0$ semigroup and \[ \|e^{-t \Lambda_r(a,b)} \|_{r \rightarrow r} \leq e^{t \omega_r}, \quad \omega_r = \frac{\lambda \delta}{2(r-1)},\quad r \in I_c:=[r_c,\infty[. \] \smallskip {\rm (\textit{iii})} The interval $I_c$ is the maximal interval of quasi contractive solvability. \smallskip {\rm (\textit{iv})} For each $r \in I_c^o, \; e^{-t\Lambda_r(a,b)}$ is a holomorphic semigroup of quasi contractions in the sector \[ |\arg t | \leq \frac{\pi}{2}-\theta_r, \quad 0 < \theta_r < \frac{\pi}{2}, \; \tan \theta_r \leq \mathcal K(2- r^\prime \sqrt{\delta})^{-1}, \] where $\mathcal K = \frac{|r-2|}{\sqrt{r-1}} +r^\prime \sqrt{\delta}$ if $r\leq 2r_c$ and $\mathcal K=\frac{r-2 +r\sqrt{\delta}}{\sqrt{r-1}}$ if $ r>2r_c.$ \smallskip {\rm (\textit{v})} $e^{-t \Lambda_r(a,b)}$, $r \in I_c$, extends to a positivity preserving, $L^\infty$ contraction, quasi bounded holomorphic semigroup on $L^r$ for every $r \in I_m:= ]\frac{2}{2- \frac{d-2}{d}\sqrt{\delta}}, \infty[$. \smallskip {\rm (\textit{vi})} The interval $I_m$ is the maximal interval of quasi bounded solvability. \smallskip {\rm (\textit{vii})} For every $r \in I_m$ and $q>r$ there exist constants $c_i = c_i(\delta, r, q)$, $i=1,2$ such that the $(L^r, L^q)$ estimate \begin{equation*} \|e^{-t \Lambda_r(a,b)} \|_{r \rightarrow q} \leq c_1 e^{c_2 t} \; t^{-\frac{d}{2} (\frac{1}{r}-\frac{1}{q})} \end{equation*} is valid for all $t > 0.$ \smallskip {\rm (\textit{viii})} Let $\delta < 1$, and let $a_n \in (H_{\sigma,\xi})$, $b_n:\mathbb R^d \rightarrow \mathbb R^d$, $n=1,2,\dots$ be smooth and such that \[ a_n \rightarrow a \text{ strongly in }[L^2_{\rm loc}]^{d\times d}, \quad b_n \rightarrow b \text{ strongly in } [L^2_{\rm loc}]^d \] and $b_n \in \mathbf F_\delta(A^n)$ with $c(\delta)$ independent of $n$, where $A^n \equiv -\nabla \cdot a_n \cdot \nabla$. Then $$e^{-t \Lambda_r(a, b)} = s \mbox{-} L^r \mbox{-} \lim_{n \uparrow \infty} e^{-t \Lambda_r(a_n, b_n)}$$ whenever $r \in I^o_c$, where $\Lambda_r(a_n,b_n)=-\nabla \cdot a_n \cdot \nabla + b_n\cdot\nabla$ of domain $W^{2,r}$. \end{theorem} \begin{remarks} (a) For $\delta<1$, the corresponding to $\Lambda$ quadratic form $t[u]=\langle a \cdot \nabla u,\nabla u\rangle + \langle b \cdot \nabla u,u\rangle$, $D(t)=W^{1,2}$ possesses the Sobolev embedding property $${\rm Re} t[u] \geq c_S\|u\|^2_{2j}, \quad j=\frac{d}{d-2}.$$ \textit{This ceases to be true already for $\delta=1$. The same occurs for $1<\delta<4$ and $r=r_c$.} \smallskip (b) The intervals $I_c$, $I_m$ are maximal already for $a=I$ and $b(x)=\sqrt{\delta}\frac{d-2}{2}|x|^{-2}x$. \smallskip (c) Assertions (\textit{i})-(\textit{iv}) are in fact valid for symmetric $a \in [L^1_{{\rm loc}}]^{d \times d}$ such that $a \geq \sigma I$, $\sigma>0$, and $b_a^2 \in L^1 + L^\infty$, see \cite[Theorem 4.2]{KiS}. \smallskip (d) While for $b \in \mathbf{F}_\delta(A)$, $\delta<1$ one first constructs the semigroup in $L^2$ (using the method of quadratic forms) and then proves the corresponding convergence results, in the case $b \in \mathbf{F}_\delta(A)$, $1 \leq \delta<4$ the convergence result of Theorem \ref{thm_fb}(\textit{i}) becomes the means of construction of the semigroup. \end{remarks} \medskip \textbf{2.~}Note that $\mathbf{N}_e\cap\mathbf{F}\subset \mathbf{K}^d\subset\mathbf{F}$, where $\mathbf{F}:=\cup_{\beta>0}\mathbf F_\beta(-\Delta)$, and \[ \mathbf{K}^d:=\{|b|\in L^2_{\rm loc}\mid \kappa_d(b,h):=\sup_{x\in\mathbb{R}^d}\int_0^h e^{t\Delta}|b|^2(x)dt <\infty \text{ for some } h>0\}. \] Indeed, using $b \in \mathbf{F}$, we have $e^{t\Delta}b^2(x)\equiv\langle k(t,x,\cdot)b^2(\cdot)\rangle\leq \beta\|\nabla\sqrt{k(t,x,\cdot)}\|_2^2+c(\beta)=\frac{\beta d}{8}\frac{1}{t}+c(\beta)$ for some $\beta>0$ and $c(\beta)$. Therefore, for $0<t\leq h$, $$ e^{t\Delta}b^2(x) \leq \sqrt{\frac{\beta d}{8}+c(\beta)h}\,\sqrt{e^{t\Delta}b^2(x)}\frac{1}{\sqrt{t}}, $$ and so the condition $b \in \mathbf{N}_e$ now yields the required. In turn, the inclusion $\mathbf{K}^d\subset\mathbf{F}$ is well known (use the fact that $b \in \mathbf{K}^d$ is equivalent to $\||b|^2(\lambda-\Delta)^{-1}\|_{1 \rightarrow 1}<\infty$, $\lambda>0$). \smallskip \textbf{3.~}Let us fix a continuous function $\phi:[0,\infty[ \rightarrow [0,\infty[$ satisfying the following properties: 1) $\phi(0)=0$, 2) $\phi(t)/t\in L^1[0,1]$. Put $$n_\phi(b,h)=\sup_{x\in \mathbb{R}^d}\int_0^h e^{t\Delta}b^2(x)\frac{dt}{\phi(t)}.$$ If $n_\phi(b,h)<\infty$ for some $h>0$, then we write $b \in \mathbf{N}_\phi$. The class $\mathbf{N}_\phi$ arises as the class providing the two-sided Gaussian on the heat kernel of $-\nabla \cdot a(t,x) \cdot \nabla + b(t,x) \cdot \nabla$, where $a(t,x)$ is a measurable uniformly elliptic matrix, see \cite{S2}, \cite{LS}. Since (for $b=b(x)$) $$ \int_0^h \sqrt{e^{t\Delta}b^2(x)}\frac{dt}{\sqrt{t}} \leq \biggl[\int_0^h e^{t\Delta}b^2(x)\frac{dt}{\phi(t)}\biggr]^{\frac{1}{2}}\biggl[\int_0^h \frac{\phi(t)}{t}dt\biggr]^{\frac{1}{2}}, $$ we have $ \mathbf{N}_\phi \subset \mathbf{N}_e$ for every admissible $\phi$. Moreover, since $\phi$ is continuous and $\phi(0)=0$, it is seen that $n_\phi(b,h) > k_d(b,h)$, and so $\mathbf{N}_\phi \subset \mathbf{K}^{d}$. Thus, $$ \mathbf{N}_\phi \; \subset \; \mathbf{N}_e \cap \mathbf{K}^d \; \subset \; \mathbf{K}^{d+1} \cap \mathbf{K}^d. $$ The need for more restrictive assumption ``$b \in \mathbf{N}_\phi$'' when $a=a(t,x)$ is dictated by the subject matter: in the time-dependent case there are no estimates $\mathcal{N}(t)$, $\hat{\mathcal{N}}(t) \leq c(t)$ for any $c(t)$, cf.\,the previous comment. \smallskip \textbf{4.~}Let us comment more on classes $\mathbf{K}^{d+1}$ and $\mathbf{F}$. Note that $\mathbf{K}^{d+1} \not\subset \mathbf{F}$: There are $b \in \mathbf{K}^{d+1}$ such that, for a given $p>1$, $|b| \not\in L^{p}_{{\rm loc}}$, e.g. consider $$ |b(x)|=\mathbf{1}_{B(0, 1)}(x)|x_1|^{-\alpha_p}, \quad 0<\alpha_p<1. $$ On the other hand, already $[L^d]^d \not\subset \mathbf{K}^{d+1}$, and so $\mathbf{F} \not \subset \mathbf{K}^{d+1}$. [Indeed, let \[ |b(x)|=\mathbf{1}_{B(0,e^{-1})}(x)|x|^{-1}|\log |x||^{-\alpha},\;\alpha>d^{-1}, \;d\geq 3. \] Then $\|b\|_d<\infty$ and $k_{d+1}(b,h)=\infty$.] This dichotomy between the classes $\mathbf{K}^{d+1}$ and $\mathbf{F}$ was resolved in \cite{Ki,KiS} with development of the Sobolev regularity theory of $-\Delta + b \cdot \nabla$ for $b$ in the class $$\mathbf{F}^{\scriptscriptstyle 1/2}=\big\{b \in L^1_{{\rm loc}} \mid \lim_{\lambda \rightarrow \infty}\||b|^{\frac{1}{2}}(\lambda-\Delta)^{-\frac{1}{4}}\|_{2 \rightarrow 2} <\infty \big\}$$ (introduced in \cite{S} as the class responsible for the $(L^p,L^q)$ estimate on the semigroup) that contains $\mathbf{K}^{d+1} + \mathbf{F}:=\{b_1+b_2 \mid b_1 \in \mathbf{K}^{d+1}, b_2 \in \mathbf{F}\}$. By analogy, one can ask if it is possible to extend the convergence results in Theorem \ref{nash_apost_thm} and Theorem \ref{thm_fb}, or ($L^p,L^q$) estimates, to $-\nabla \cdot a \cdot \nabla + b \cdot \nabla$ with a measurable $a \in (H_{\sigma,\xi})$ and $b=b_1+b_2$ with $b_1 \in \mathbf{N}_e$, $b_2 \in \mathbf{F}_{\delta}(A)$. \medskip \textbf{5.~}Theorem \ref{nash_apost_thm}(\textit{iv}), (\textit{viii}) (the two-sided Gaussian bounds on the heat kernel and its time derivative) can be extended to more general operator $$ \Lambda(a,b,\hat{b})=-\nabla \cdot a \cdot \nabla + b \cdot \nabla + \nabla \cdot \hat{b} $$ with $a \in (H_{\sigma,\xi})$, and $(b,\hat{b} \in \mathbf{N}_e, \hat{b} \in \mathbf F)$ or ($b,\hat{b} \in \mathbf{N}_e, b \in \mathbf F$), provided that $n(b,h)$, $n(\hat{b},h)$ are sufficiently small. Note that the above assumptions on $b$ and $\hat{b}$ are non-symmetric, i.e.\,the presence of $b \in \mathbf{N}_e$ forces $\hat{b}$ to be more regular: $\hat{b} \in \mathbf{N}_e\cap \mathbf F$, and vice versa. We also note that here the form-boundedness assumption seems to be justified. The proof follows the argument in the present paper but with the Nash's functions $\mathcal N$, $\hat{\mathcal N}$ defined with respect to $u(t,x,y):=e^{-t\Lambda(a,b)}(x,y)$. We will address this matter in detail elsewhere. \medskip \textbf{6.~}The authors do not know if there is a proof of the Harnack inequality for $\Lambda = -\nabla \cdot a \cdot \nabla + b \cdot \nabla$, $a \in (H_{\sigma,\xi})$, $b \in \mathbf{N}_e$ that does not use the lower bound on $e^{-t\Lambda}(x,y)$. \bigskip

https://arxiv.org/abs/1912.00472

Algorithms in $A_\infty$-algebras

Building on Kadeishvili's original theorem inducing $A_\infty$-algebra structures on the homology of dg-algebras, several directions of algorithmic research in $A_\infty$-algebras have been pursued.In this paper we will survey work done on calculating explicit $A_\infty$-algebra structures from homotopy retractions; in group cohomology; and in persistent homology.

\section{Introduction} In his 1980 paper~\cite{kadeishvili80}, Kadeishvili proved that the homology of any dg-algebra has an induced $A_\infty$-algebra structure. The proof itself is by induction on the arity of the operation, writing out explicitly how to create $m_n(a_1,\dots,a_n)$ in terms of lower order operations. By giving these expressions, the proof is basically a formulation of what an algorithm for the computation of an $A_\infty$ algebra structure on the homology of a dg-algebra would look like. Even though this initial setup has a strongly algorithmic flavor, actual algorithms and computer-supported calculations of $A_\infty$-algebra structures emerged far later. In this paper, we plan to give an overview of work on creating software computing $A_\infty$-algebra structures and the contexts and techniques used for these. Merkulov~\cite{merkulov_sha_kahler99} has given Kadeishvili's construction a more concrete, and combinatorially accessible presentation -- but the focus for this paper is in explicit algorithmic computation with computer implementations, preferably publically accessible. \section{Background} We fix a ring $\kk$. All tensor products are over $\kk$ unless otherwise noted, and tensor powers are denoted by $A^{\otimes n} = A\otimes\dots\otimes A$. A graded $k$-vector space $A$ is an \emph{$A_\infty$-algebra} if one of the following equivalent conditions hold \begin{enumerate} \item There is a family of maps $\mu_i\colon A^{\otimes i}\to A$, called higher multiplications fulfilling the Stasheff identities \[ \operatorname{St}_n\colon \sum_i\sum_j \mu_i\circ_j\mu_{n-i} = 0\quad. \] \item There is a family of chain maps from the cellular chain complex of the associahedra to appropriate higher endomorphisms of $A$ \[ \mu_n\colon C_*(K_n)\to\operatorname{Hom}(A^{\otimes n}, A)\quad. \] \item $A$ is a representation of the free dg-operad resolution $\mathcal Ass_\infty$ of the associative operad. \end{enumerate} An \emph{$A_\infty$-coalgebra} is defined by dualizing this definition. The structure was introduced by Jim Stasheff in \cite{stasheff_ha_hspaces63}. Good introductory surveys have been written by Lu, Palmieri, Wu and Zhang \cite{lu_wu_palmieri_zhang_Aoo_ring04,lu_wu_palmieri_zhang_Aoo_Ext06} as well as by Bernhard Keller \cite{keller_intro01,keller_Aoo_repr00}. A graded $k$-vector space $A$ is a differential graded algebra (dg-algebra) if it is equipped with a differential operator $\partial: A\to A$ of degree $-1$ and an associative multiplication $m_2: A\otimes A\to A$ of degree $0$, such that the Leibniz rule holds: \[ \partial m_2(x, y) = m_2(\partial x, y) + (-1)^{|x|}m_2(x, \partial y) \] A module over a dg-algebra $A$ is a graded vectorspace $M$ with a differential operator $\partial: M\to M$ and an associative multiplication $m_2: M\otimes A\to M$ that obeys the Leibniz rule. Kadeishvili proved in his 1980 paper~\cite{kadeishvili80} that the homology of a dg-algebra has an inherited and quasi-isomorphic $A_\infty$-algebra structure. The proof starts out defining $\mu_1=0$ and $\mu_2([x_1]\otimes [x_2]) = (-1)^{|x_1|+1}[x_1\cdot[x_2]$, as well as starting to define an $A_\infty$-morphism $f$ by $f_1$ simply a cycle-choosing homomorphism. The proof proceeds by induction\footnote{which translates well to recursion: this is how algorithms enter the picture}: if all $m_j$ and $f_j$ have been defined for $j<i$, then let \[ U_n = \sum_{s=1}^n m_2 \circ f_s \otimes f_{n-s}) + \sum_{k=0}^{n-2}\sum_{j=2}^{n-1} f_{n-j+1} \circ \id^{\otimes k}\otimes m_j \otimes \id^{\otimes n-j-k+1} \] The Stasheff axioms can be translated to \[ m_1\circ f_n = (f_1\circ m_n - U_n) \] The right hand side is homological to zero, so we can pick $f_n$ to be a bounding element of this difference. Extend by linearity. The structure of this argument lends itself excellently well to concrete and algorithmic calculations, and there has been a few approaches to algorithmic and computer-aided $A_\infty$-algebra work. There are three main topics that emerge, and we will dedicated a chapter to each of them. First up, in Section~\ref{sec:reductions}, we will review Ainhoa Berciano's work on contractions of dg-algebras to dg-modules, with implementations in the computer algebra system Kenzo. Next, in Section~\ref{sec:group-cohomology} we will go to the realm of group cohomology. Mikael Vejdemo-Johansson worked on algorithms to directly calculate $A_\infty$-algebra structures on the modular group cohomology of $p$-groups, and generated a number of ways to recognize feasibility of the calculation as well as a stopping criterion. Stephan Schmid, later on, used some of Vejdemo-Johansson's results in a concrete calculation of the $A_\infty$-algebra structure on the modular group cohomology of the symmetric group $S_p$. Finally, in Section~\ref{sec:persistent-infinity}, we will describe recent work by Murillo and Belchí on using $A_\infty$-coalgebra structures on persistent homology rings to create new perspectives on bottleneck distances and stability of persistence barcodes. \section{Reductions} \label{sec:reductions} Ainhoa Berciano's work~\cite{berciano-2006, berciano2010computational, berciano2012searching, berciano2004coalgebra, berciano2010coalgebra} starts with perturbation theory. This framework has as its core result the Basic Perturbation Lemma~\cite{brown1967twisted}, that describes how a \emph{contraction} changes under perturbation. A \emph{contraction} connects two dg-modules $M$ and $N$, abstracting the homotopy concepts of deformation retracts. A contraction consists of morphisms $f: N\to M$, $g: M\to N$ and $\phi: N\to N$ such that $f$ and $g$ are almost an isomorphism -- up to a homotopy operation in $N$. In other words, we require \[ fg = \id_M \quad gf + \phi\partial_N + \partial_N\phi = \id_N \quad f\phi = 0 \quad \phi g = 0 \quad \phi\phi = 0 \] A contraction preserves homology: $H(N)$ is canonically isomorphic to $H(M)$, but this isomorphism tends not to transfer algebraic structures from $N$ to $M$. If the differential structure on $N$ is perturbed: instead of boundary operator $\partial_N$, $N$ is equipped with a new boundary operator $\partial_N+\delta$, then the Basic Perturbation Lemma produces a new contraction. The requirement for this construction is that $\phi\delta$ is pointwise nilpotent: for any $x\in N$ there is an $n$ so that $(\phi\delta)^n(x)=0$. Then there is a new contraction $f_\delta, g_\delta, \phi_\delta$ between $N$ equipped with the boundary operator $\partial_N+\delta$ and $M$ equipped with the boundary operator $\partial_M+\partial_\delta$ given by \begin{align*} \partial_\delta &= f\delta\sum_{i\geq 0}(-1)^i(\phi\delta)^i g & f_\delta &= f\left(1-\delta\sum_{i\geq 0}(-1)^i(\phi\delta)^i\phi\right) \\ g_\delta&=\sum_{i\geq 0}(-1)^i(\phi\delta)^ig & \phi_\delta&=\sum_{i\geq 0}(-1)^i(\phi\delta)^i\phi \end{align*} Berciano generates an algorithm for transferring $A_\infty$-coalgebra structures between DG-modules using the \emph{tensor trick}~\cite{gugenheim-lambe-stasheff-91}. The tensor trick starts with a dg-coalgebra $C$, a dg-module $M$ and a contraction from $C$ to $M$. Take the tensor module of the desuspension of all components in this contraction to produce a new contraction. With the cosimplicial differential, we can use the Basic Perturbation Lemma and obtain a new contraction. The tilde cobar differential generates an induced $A_\infty$-coalgebra structure, explicitly given by the comultiplication operations \begin{equation}\label{eqn:Aoo-coalgebra} \Delta_i=(-1)^{[i/2]+i+1}f^{\otimes i}\Delta^{[i]}\phi^{[\otimes(i-1)]}\dots\phi^{[\otimes 2]}\Delta^{[2]}g ; \qquad \Delta^{[k]} = \sum_{i=0}^{k-2}(-1)^i\id^{\otimes i}\otimes\Delta\otimes\id^{k-i-2} \end{equation} This derivation allows Berciano to prove~\cite{berciano2004coalgebra} that in $H_*(K(\pi, n); \mathbb{Z}_p)$ for a finitely generated abelian group $\pi$, the only non-null morphisms in the $A_\infty$-coalgebra structure have to have order $i(p-2)+2$ for some non-negative integer $i$. The final formula in~\ref{eqn:Aoo-coalgebra} is concrete enough that it has been implemented on the computer algebra platform Kenzo in the packages ARAIA (\textbf{A}lgebra \textbf{R}eduction \textbf{A}-\textbf{I}nfinity \textbf{A}lgebra) and CRAIC (\textbf{C}oalgebra \textbf{R}eduction \textbf{A}-\textbf{I}nfinity \textbf{C}oalgebra). \section{Group Cohomology} \label{sec:group-cohomology} Fix a group $G$ and a field $\kk$. The cohomology algebra of the Eilenberg-MacLane space $H^*(K(G, 1))$ is isomorphic to the Ext-algebra $\Ext_{\kk G}(\kk, \kk)$ of the group ring and is called the group cohomology $H^*(G)$. Because of the connection to the Ext-algebra, the group cohomology can be calculated from the composition dg-algebra of $\Hom(F_*, F_*)$ for a free resolution $F_*\to\kk$ in the category of $G$-modules. Several computer algebra systems, including Magma~\cite{magma} and GAP~\cite{GAP4} support calculations with $G$-modules. In such a system, we create a free resolution $F_*$ of $\kk$. Chain maps $F_*\to F_*$ are then represented by a sequence of maps, one for each degree, each determined by lower-dimensional maps through commutativity of the corresponding squares in the chain map diagram. With $\Hom(F_*, F_*)$ represented, we can compute $H^*G$ as $H_*\Hom(F_*, F_*)$. Since the $\Hom(F_*, F_*)$ is a dg-algebra, by Kadeishvili's theorem, $H^*G$ has an induced $A_\infty$-algebra structure. Vejdemo-Johansson~\cite{mikAooCncm07,vejdemo2010blackbox,vejdemo2008computation} studies this $A_\infty$-algebra structure from a strictly algorithmic perspective. \subsection{Blackbox computation of $A_\infty$} \label{sec:blackb-comp-a_infty} A cornerstone of Vejdemo-Johansson's approach to computing $A_\infty$-algebras is the following theorem (\cite[Theorem 3]{vejdemo2010blackbox}): \begin{theorem}\label{thm:mvj:linearity} If $A$ is a dg-algebra and \begin{enumerate} \item There is an element $z\in H_*A$ generating a polynomial subalgebra (ie is not a torsion element) \item $H_*A$ is a free $\kk[z]$-module \item\label{thm:cond:commutativity} $H_*A$ has a $\kk[z]$ linear $A_{n-1}$-algebra structure induced by the dg-algebra structure on $A$, such that $f_1(z)f_k(a_1, \dots, a_k) = f_k(a_1, \dots, a_k)f_1(z)$ \item We have a chosen $\kk[z]$-basis $b_1, \dots$ of $H_*A$ and all $m_k(v_1, \dots, v_k)$ and $f_k(v_1, \dots, v_k)$ are chosen by Kadeishvili's algorithm for all combinations of basis elements $v_j\in\{b_1,\dots\}$ \end{enumerate} Then a choice of $m_n$ and $f_n$ by Kadeishvili's algorithm for all input values taken from this $\kk[z]$-basis extends to a $\kk[z]$-linear $A_n$-algebra structure on $H_*A$ induced by the dg-algebra structure on $A$. \end{theorem} The condition~\ref{thm:cond:commutativity} says that for the $A_\infty$-morphism $H_*A\to A$ produced by Kadeishvili's construction, the cycle chosen for $z$ commutes -- \textbf{on a chain level} -- with each chain map chosen for the higher operations. This is the key condition for the theorem -- and also the one that makes the theorem most fragile. The theorem tells us we can construct an $A_\infty$-algebra structure step by step. If there is one of these non-torsion central elements $z$, we can reduce the complexity of $H_*A$ for the purpose of calculating its higher operations -- if we find a family of central elements $z_1, \dots, z_k$ such that $H_*A$ is a finite module over $\kk[z_1,\dots,z_k]$, then it is enough to study the finitely many basis elements $b_1,\dots,b_m$ in a presentation of $H_*A$ as a $\kk[z_1,\dots,z_k]$-module. This makes each calculation a finite (though large) in terms of the number of input combinations that need to be studied. Using this theorem is easier if -- as is the case for cyclic groups -- the resolution $F_*$ is periodic. Theorem~\ref{thm:mvj:linearity} makes it easier to extend from $A_{n-1}$ to $A_n$, using a condition that can be checked for each extension step. Once the condition -- commutativity of the representative chain maps -- fails, the structure calculated thus far is valid, but further extensions are obstructed. The key to bring computational effort down to a finite time endeavour lies in \cite[Theorem 5]{vejdemo2010blackbox}: \begin{theorem} Let $A$ be a dg-algebra. If in an $A_{2q-2}$-algebra structure on $H_*A$, $f_k=0$ and $m_k=0$ for all $q\leq k\leq 2q-2$ then the $A_{2q-2}$-structure is already an $A_\infty$-structure with all higher $f_n$ and all higher $m_n$ given by zero maps. \end{theorem} Through finding central elements, the infinitely many basis elements of $H_*A$ can be brought down to a finite number of basis elements to check. And by finding a large enough gap, in which all chain representations and all products vanish, the computation can be terminated producing a result. This approach was implemented as a module distributed with Magma~\cite{magma}, and was used both to confirm Madsen's~\cite{madsen_phd} computation of $A_\infty$-algebra structures on the group cohomology of cyclic groups and to conjecture~\cite{vejdemo2008computation} the start of an $A_\infty$-structure on the cohomology on some dihedral groups. \subsection{The Saneblidze-Umble diagonal} \label{sec:sanebl-umble-diag} In~\cite{saneblidze-umble-2004}, Saneblidze and Umble gave an explicit construction for a diagonal on the associahedra. This construction translates directly to a method to combine $A_\infty$-algebra structures on $V$ and $W$ into an $A_\infty$-algebra structure on $V\otimes W$. Vejdemo-Johansson uses this construction in~\cite{mikAooCncm07} to prove non-triviality of some operations on $H^*(C_n\times C_m)$. From results by Berciano and Umble~\cite{berciano2011some}, we know that any non-trivial operation on this group cohomology of arity less than $n+m-1$ has to have arity $2, n, m$ or $n+m-2$. Berciano also shows~\cite{berciano-2006} that any non-zero higher coproduct on $H_*(C_q\times C_q)$ has arity $k(q-2)+2$ for some $k$. In addition to the induced operations in arities $2, n, m$ and $n+m-2$, Vejdemo-Johansson shows that there are non-trivial operations of arity $2n+m-4$ and $n+2m-4$. The original article states a far more generous claim: that all the arities $k(n-2) + k(m-2)+2$, $(k-1)(n-2) + k(m-2) + 2$ and $k(n-2) + (k-1)(m-2) + 2$ have non-zero operations -- this argument turned out to have a subtle flaw, and was retracted. More details are available in~\cite{vejdemo2008computation}. Any practical use of the Saneblidze-Umble diagonal would benefit greatly from a computer-facilitated access to the coefficients of the diagonal construction. In an unpublished preprint~\cite{vejdemo2007enumerating}, Vejdemo-Johansson provides a computer implementation of an algorithm to enumerate the Saneblidze-Umble terms. \subsection{Symmetric groups} \label{sec:symmetric-groups} Schmid~\cite{Schmid2014} studies the group cohomology of the symmetric group $S_p$ on $p$ elements, with coefficients in the finite field $\mathbb{F}_p$ with $p$ elements. For this group cohomology, he presents a basis with which he is able to prove that the only non-trivial $A_\infty$-operations on $H^*S_p$ are of arity $2$ and $p$. To do this, he goes through large and somewhat onerous explicit calculations to show that there is a periodic projective resolution of $\mathbb{F}_p$ over $\mathbb{F_p}S_p$, and that the resolution has a large enough gap to allow the use of Vejdemo-Johansson's theorem. \section{Persistent A-infinity} \label{sec:persistent-infinity} Persistent homology and cohomology form the cornerstone of the fast growing field of Topological Data Analysis. The fundamental idea is to study the homology functor applied to diagrams of topological spaces \[ \VV : \VV_0 \into \VV_1 \into \dots \] These spaces are often generated directly from datasets, by constructions such as the \v{C}ech construction: for data points $\XX = \{x_0, \dots, x_N\}$, an abstract simplical complex $\check{C}_\epsilon$ has as its vertices $\XX$ and includes a simplex $[x_{i_0}, \dots, x_{i_d}]$ precisely if the intersection of balls $\bigcap_{j=0}^d B_\epsilon(x_{i_j})$ is non-empty. If $\epsilon$ increases, no intersections will become empty, and so no simplices will vanish. So the \v{C}ech complexes, as $\epsilon$ sweeps from 0 to $\infty$, generates a nested sequence of topological spaces. The inclusion maps $\iota_i^j:\VV_i\to\VV_j$ lift by functoriality to linear maps on homology: $H_*(\iota_i^j): H_*\VV_i\to H_*\VV_j$. We may define a persistent homology group as the image $PH_*^{i,j}(\VV) = \img H(\iota_i^j)$. For more details on the data analysis side, we recommend the surveys \cite{carlsson2009topology,ghrist2008barcodes,vejdemo2014sketches} \subsection{Barcodes and stability} \label{sec:barcodes-stability} As the homology functor is applied to the diagram of topological spaces, using coefficients from a field $\kk$ for the homology computation, the result is a diagram of vector spaces. By either imbuing the resulting diagram with the structure of a module over the polynomial ring $\kk[t]$, or as representations of a quiver $Q$ of type $A_n$, the corresponding classification theorems produce a decomposition of $H_*(\VV)$ into a direct sum of \emph{interval modules}. These interval modules are $\mathbb N$-graded modules defined by a pair of indices $b, d$, and are defined as $0$-dimensional for degrees $k<b$ and for degrees $k > d$. For degrees from $b$ to $d$, the interval module is one-dimensional, with identity maps connecting each space to the next. Thus, the homology of a diagram $\VV$ of topological spaces with field coefficients can be described by a multiset $\Dgm(\VV) = \{ (b_i, d_i) \}_{i\in I}$, called the \emph{persistence barcode} or \emph{persistence diagram} of the diagram $\VV$. The dimension of the persistent homology group $PH_*^{i,j}(\VV)$ is exactly the number of intervals $(b_k, d_k)$ in $\Dgm(\VV)$ such that $b_k \leq i \leq j \leq d_k$. Between any pair of such diagrams we can create a distance called the \emph{bottleneck distance}. Setting $\Delta = \{(x,x) : x\in\mathbb R\}$, this distance is defined by: \[ d_B(\Dgm(\VV, \WW)) = \inf_{\gamma: \VV\cup\Delta\isoto\WW\cup\Delta} \quad\max_{v\in\VV\cup\Delta} d(v, \gamma v) \] This distance measures the largest displacement needed to change $\Dgm(\VV)$ into $\Dgm(\WW)$, while allowing intervals to disappear into and emerge from the infinite set of possible 0-length intervals. First with the bottleneck distance (and in later research more sophisticated), a range of \emph{stability theorems} have been proven, starting in \cite{cohen2007stability}. Good overviews can be found in \cite{vejdemo2014sketches,chazal2012structure}. These theorems take the shape of \begin{theorem}[Stability meta-theorem] If $d(\VV, \WW) < \epsilon$, then $d'(\Dgm(H(\VV)), \Dgm(H(\WW))) < \epsilon$ for specific choices of distances $d$ and $d'$. \end{theorem} \subsection{$A_\infty$ in persistence} \label{sec:a_infty-persistence} Murillo and Belchí introduced in \cite{belchi2015ainfty,belchi2017optimising} an $A_\infty$-coalgebra approach to barcode distances. If each new cell introduced in the step from $\VV_i$ to $\VV_{i+1}$, with all the $\VV_j$ chosen to be CW-complexes, and working over the rationals $\mathbb{Q}$, then there is a set of compatible choices of $A_\infty$-coalgebras for the entire sequence. They define a $\Delta_n$-persistence group \[ \Delta_nPH_*^{i,j}(\VV) = \img (H_*(\iota_i^j)|_{\bigcap_{k=i}^j\ker(\Delta_n^k\circ\iota_i^k} \] In other words, the $\Delta_n$-persistence group retains from the ordinary persistence groups precisely those elements out of $H_*(\VV_i)$ whose images in each $\VV_k$ vanish under application of the higher coproduct $\Delta_n$. These $\Delta_n$-persistence groups generate $\Delta_n$-persistence barcodes as multisets $\Dgm_{\Delta_n}(\VV)$ of intervals $[b,d]$. From these barcodes, the dimension of $\Delta_nPH_*^{i,j}(\VV)$ equals the number of $[b_k, d_k]\in\Dgm_{\Delta_n}(\VV)$ such that $b_k\leq i\leq j\leq d_k$. As pointed out by the authors, these higher order barcodes may ``flicker'' in a way that classical persistence strictly avoids: the same element can exist over several disjoint intervals. This has been carefully avoided in the greater literature on persistent homology: the flickering behavior invites wild representation theories, where the decomposition that generates barcodes is no longer available. \subsection{$A_\infty$ bottleneck distance} \label{sec:a_infty-bottl-dist} Herscovich \cite{herscovich2014higher} introduces a novel metric on persistent homology. Herscovich constructs a metric on locally finite Adams graded minimal $A_\infty$-algebras, and then quotients by quasi-isomorphism to establish a metric on persistent homology barcodes equipped with an $A_\infty$-algebra structure. The question of stability of this metric is left open by Herscovich, except to note that the 1-ary case coincides with the classical bottleneck distance. \printbibliography \end{document}

https://arxiv.org/abs/1207.6165

Anticipated backward doubly stochastic differential equations

In this paper, we deal with a new type of differential equations called anticipated backward doubly stochastic differential equations (anticipated BDSDEs). The coefficients of these BDSDEs depend on the future value of the solution $(Y, Z)$. We obtain the existence and uniqueness theorem and a comparison theorem for the solutions of these equations. Besides, as an application, we also establish a duality between the anticipated BDSDEs and the delayed doubly stochastic differential equations (delayed DSDEs).

\section{Introduction} Backward stochastic differential equation (BSDE) was considered the general form the first time by Pardoux-Peng \cite{PP1} in 1990. In the last twenty years, the theory of BSDEs has been studied with great interest due to its applications in the pricing/hedging problem (see e.g. \cite{KEM1997, KPQ}), in the stochastic control and game theory (see e.g. \cite{KPQ, HL1}), and in the theory of partial differential equations (see e.g. \cite{BBP1997, BQR, PP2}). In order to give a probabilistic representation for a class of quasilinear stochastic partial differential equations (SPDEs), Pardoux-Peng \cite{PP3} first studied the backward doubly stochastic differential equations (BDSDEs) of the general form \begin{equation}\label{equation:PP BDSDE} Y_t=\xi+\int_t^T f(s, Y_s, Z_s)ds +\int_t^T g(s, Y_s, Z_s)d\overleftarrow{B}_s - \int_t^T Z_s d{W}_s,\ \ \ \ t \in [0, T], \end{equation} where the integral with respect to $\{B_t\}$ is a "backward It\^{o} integral", and the integral with respect to $\{W_t\}$ is a standard forward integral. Note that these two types of integrals are particular cases of the It\^{o}-Skorohod integral, see Nualart-Pardoux \cite{NP}. Pardoux-Peng \cite{PP3} proved that under Lipschitz condition on the coefficients, BDSDE (\ref{equation:PP BDSDE}) has a unique solution. Since then, the theory of BDSDEs has been developed rapidly by many researchers. Bally-Matoussi \cite{BM} gave the probabilistic representation of the solutions in Sobolev space of semilinear SPDEs in terms of BDSDEs. Matoussi-Scheutzow \cite{MS} studied BDSDEs and their applications in SPDEs. Shi et al. \cite{SGL} proved a comparison theorem for BDSDEs with Lipschitz condition on the coefficients. Lin \cite{Lin} obtained a generalized comparison theorem and a generalized existence theorem of BDSDEs. On the other hand, recently, Peng-Yang \cite{PY} (see also \cite{Y}) introduced the socalled anticipated BSDEs (ABSDEs) of the following form: \begin{equation*} \left\{ \begin{tabular}{rlll} $-dY_t$ &=& $f(t, Y_t, Z_t, Y_{t+\delta(t)}, Z_{t+\zeta(t)})dt-Z_tdW_t, $ & $ t\in[0, T];$\\ $Y_t$ &=& $\xi_t, $ & $t\in[T, T+K];$\\ $Z_t$ &=& $\eta_t, $ & $t\in[T, T+K],$ \end{tabular}\right. \end{equation*} where $\delta(\cdot): [0, T]\rightarrow \mathbb{R^+} \setminus \{0\}$ and $\zeta(\cdot): [0, T]\rightarrow \mathbb{R^+} \setminus \{0\}$ are continuous functions satisfying $\mathbf{(a1)}$ there exists a constant $K \geq 0$ such that for each $t\in[0, T],$ $$t+\delta(t) \leq T+K,\quad t+\zeta(t) \leq T+K;$$ $\mathbf{(a2)}$ there exists a constant $M \geq 0$ such that for each $t\in[0, T]$ and each nonnegative integrable function $g(\cdot)$, $$\int_t^T g(s+\delta(s))ds\leq M\int_t^{T+K} g(s)ds,\quad \int_t^T g(s+\zeta(s))ds\leq M\int_t^{T+K} g(s)ds.$$ Peng-Yang \cite{PY} proved the existence and uniqueness of the solution to the above equation, and studied the duality between anticipated BSDEs and delayed SDEs. In this paper, we are interested in the following BDSDEs with coefficients depending on the future value of the solution $(Y, Z)$: \begin{equation}\label{equation:main} \left\{ \begin{tabular}{rlll} $-dY_t$ &=& $f(t, Y_t, Z_t, Y_{t+\delta(t)}, Z_{t+\zeta(t)})dt$ & \vspace{2mm} \\ && $+ g(t, Y_t, Z_t, Y_{t+\delta(t)}, Z_{t+\zeta(t)}) d\overleftarrow{B}_t - Z_t d{W}_t, $ & $ t\in[0, T];$ \vspace{2mm} \\ $Y_t$ &=& $\xi_t, $ & $t\in[T, T+K];$ \vspace{2mm} \\ $Z_t$ &=& $\eta_t, $ & $t\in[T, T+K],$ \end{tabular}\right. \end{equation} where $\delta >0 $ and $\zeta >0$ satisfy $(a1)$-$(a2)$. We prove that under proper assumptions, the solution of the above anticipated BDSDE (ABDSDE) exists uniquely, and a comparison theorem is given for the $1$-dimensional anticipated BDSDEs. It may be mentioned here that, to deal with (\ref{equation:main}), the most important thing for us is to establish the similar conclusions as in \cite{PP3} and \cite{SGL} for BDSDE (\ref{equation:PP BDSDE}) with $\xi$ belonging to a larger space. Besides, as an application, we study a duality between the anticipated BDSDE and delayed DSDE. The paper is organized as follows: in Section $2$, we make some preliminaries. In Section $3$, we mainly study the existence and uniqueness of the solutions of anticipated BDSDEs, and in Section $4$, a comparison result is given. As an application, in Section $5$, we establish a duality between an anticipated BDSDE and a delayed DSDE. Finally in Section $6$, the conclusion and future work are presented. \section{Preliminaries} Let $T > 0$ be fixed throughout this paper. Let $\{W_t\}_{t\in [0, T]}$ and $\{B_t\}_{t\in [0, T]}$ be two mutually independent standard Brownian motion processes, with values respectively in $\mathbb{R}^d$ and $\mathbb{R}^l$, defined on a probability space $(\Omega, \mathcal{F}, P)$. Let $\mathcal{N}$ denote the class of $P-$null sets of $\mathcal{F}$. We define $$\mathcal{F}_t := \mathcal{F}_{0, t}^W \vee \mathcal{F}_{t, T}^B,\ t\in [0, T];\ \ \mathcal{G}_s := \mathcal{F}_{0, s}^W \vee \mathcal{F}_{s, T+K}^B,\ s\in [0, T + K],$$ where for any processes $\{\varphi_t\}$, $\mathcal{F}_{s,t}^\varphi=\sigma\{\varphi_r-\varphi_s, s \leq r \leq t\} \vee \mathcal{N}$. We will use the following notations: \begin{itemize} \item{$L^2(\mathcal{G}_T; \mathbb{R}^m)$ := $\{\xi\in \mathbb{R}^m$ $|$ $\xi$ is a $\mathcal{G}_T$-measurable random variable such that $E|\xi|^2< + \infty\};$} \item{$L_{\mathcal{G}}^2(0, T; \mathbb{R}^m)$ := $\{ \varphi: \Omega\times [0, T]\rightarrow \mathbb{R}^m$ $|$ $\varphi$ is a $\mathcal{G}_t$-progressively measurable process such that $E\int_0^T |\varphi_t|^2dt< + \infty\};$} \item{$S_{\mathcal{G}}^2(0, T; \mathbb{R}^m)$ := $\{\varphi: \Omega\times [0, T]\rightarrow \mathbb{R}^m$ $|$ $\varphi$ is a continuous and $\mathcal{G}_t$-progressively measurable process such that $E[\sup_{0 \leq t \leq T} |\varphi_t|^2]< + \infty\}.$} \end{itemize} \begin{remark} It should be mentioned here that, the existing result about BDSDEs are established almost under the condition that the terminal value $\xi$ is $\mathcal{F}_T$-measurable (see \cite{PP3}, \cite{SGL}, etc.). In this paper, we will first treat the case when $\xi$ is $\mathcal{G}_T$-measurable. \end{remark} \noindent For each $t\in [0, T]$, let $$f(t, \cdot, \cdot, \cdot, \cdot): \Omega \times [0, T]\times \mathbb{R}^m\times \mathbb{R}^{m\times d}\times L_{\mathcal{G}}^2(t, T+K; \mathbb{R}^m)\times L_{\mathcal{G}}^2(t, T+K; \mathbb{R}^{m\times d})\rightarrow L^2 (\mathcal{G}_t; \mathbb{R}^m),$$ $$g(t, \cdot, \cdot, \cdot, \cdot): \Omega \times [0, T]\times \mathbb{R}^m\times \mathbb{R}^{m\times d}\times L_{\mathcal{G}}^2(t, T+K; \mathbb{R}^m)\times L_{\mathcal{G}}^2(t, T+K; \mathbb{R}^{m\times d})\rightarrow L^2 (\mathcal{G}_t; \mathbb{R}^{m\times l}).$$ We make the following hypotheses: ${\bf{(H1)}}$ There exists a constant $c > 0$ such that for any $r, \bar{r} \in [t, T+K]$, $(t, y, z, \theta, \phi)$, $(t, y^\prime, z^\prime, \theta^\prime, \phi^\prime)\in [0, T] \times \mathbb{R}^m \times \mathbb{R}^{m\times d} \times L_{\mathcal{G}}^2(t, T+K; \mathbb{R}^m)\times L_{\mathcal{G}}^2(t, T+K; \mathbb{R}^{m\times d})$, \begin{equation*} |f(t, y, z, \theta_r, \phi_{\bar{r}})-f(t, y^\prime, z^\prime, \theta_r^\prime, \phi_{\bar{r}}^\prime)|^2 \leq c(|y- y^\prime|^2+|z-z^\prime|^2+E^{\mathcal{F}_t}[|\theta_r-\theta_r^\prime|^2+|\phi_{\bar{r}}-\phi_{\bar{r}}^\prime|^2]). \end{equation*} ${\bf{(H2)}}$ $E[\int_0^T |f(s, 0, 0, 0, 0)|^2ds]< + \infty.$ ${\bf{(H3)}}$ There exist constants $c > 0$, $0 < \alpha_1 < 1$, $0 \leq \alpha_2 < \frac{1}{M}$, satisfying $0 < \alpha_1 + \alpha_2 M< 1$, such that for any $r, \bar{r} \in [t, T+K]$, $(t, y, z, \theta, \phi)$, $(t, y^\prime, z^\prime, \theta^\prime, \phi^\prime)\in [0, T] \times \mathbb{R}^m \times \mathbb{R}^{m\times d} \times L_{\mathcal{G}}^2(t, T+K; \mathbb{R}^m)\times L_{\mathcal{G}}^2(t, T+K; \mathbb{R}^{m\times d})$, \begin{equation*} |g(t, y, z, \theta_r, \phi_{\bar{r}})-g(t, y^\prime, z^\prime, \theta_r^\prime, \phi_{\bar{r}}^\prime)|^2 \leq c(|y- y^\prime|^2+E^{\mathcal{F}_t}|\theta_r-\theta_r^\prime|^2)+ \alpha_1 |z-z^\prime|^2+ \alpha_2 E^{\mathcal{F}_t}|\phi_{\bar{r}}-\phi_{\bar{r}}^\prime|^2. \end{equation*} ${\bf{(H4)}}$ $E[\int_0^T |g(s, y, z, \theta, \phi)|^2ds]< + \infty,$ for any $(y, z, \theta, \phi)$. \section{Existence and uniqueness theorem} In this section, we will mainly study the existence and uniqueness of the solution to anticipated BDSDE (\ref{equation:main}). For this purpose, we first consider a simple case when the coefficients $f$ and $g$ do not depend on the value or the future value of $(Y, Z)$: \begin{equation}\label{eq only t} Y_t = \xi_T+ \int_t^T f(s)ds + \int_t^T g(s) d\overleftarrow{B}_s - \int_t^T Z_s d{W}_s,\ \ t\in[0, T], \end{equation} where $f\in L_{\mathcal{G}}^2(0, T; \mathbb{R}^m)$, $g\in L_{\mathcal{G}}^2(0, T; \mathbb{R}^{m\times l})$ and $\xi_T \in L^2(\mathcal{G}_T; \mathbb{R}^m)$. \begin{theorem}\label{thm only t} Given $\xi_T \in L^2(\mathcal{G}_T; \mathbb{R}^m)$, BDSDE (\ref{eq only t}) has a unique solution $(Y, Z) \in L_\mathcal{G}^2(0, T; \mathbb{R}^m)\times L_\mathcal{G}^2(0, T; \mathbb{R}^{m\times d})$. \end{theorem} \begin{proof} To prove the existence, we define a filtration by $$\mathcal{H}_t := \mathcal{F}_{0, t}^W \vee \mathcal{F}_{0, T+K}^B,\ t\in [0, T + K]$$ and a $\mathcal{H}_t$-square integrable martingale $$M_t := E^{\mathcal{H}_t}[\xi_T + \int_0^T f(s)ds +\int_0^T g(s) d\overleftarrow{B}_s],\ \ t\in [0, T].$$ Thanks to It\^{o}'s martingale representation theorem, there exists a process $Z \in L_{\mathcal{H}}^2 (0, T; \mathbb{R}^{m \times d})$ such that $$M_t = M_0 + \int_0^t Z_s dW_s, \ \ t\in [0, T],$$ which implies $$M_t = M_T - \int_t^T Z_s dW_s, \ \ t\in [0, T].$$ Hence $$E^{\mathcal{H}_t}[\xi_T + \int_0^T f(s)ds +\int_0^T g(s) d\overleftarrow{B}_s]=\xi_T + \int_0^T f(s)ds +\int_0^T g(s) d\overleftarrow{B}_s- \int_t^T Z_s dW_s.$$ Subtract $\int_0^t f(s)ds +\int_0^t g(s) d\overleftarrow{B}_s$ from both sides, then we have $$Y_t=\xi_T + \int_t^T f(s)ds +\int_t^T g(s) d\overleftarrow{B}_s- \int_t^T Z_s dW_s,$$ where $$Y_t := E^{\mathcal{H}_t}[\xi_T + \int_t^T f(s)ds +\int_t^T g(s) d\overleftarrow{B}_s].$$ Next we show that $(Y, Z)$ are in fact $\mathcal{G}_t$-adapted. In fact, it is obvious that $$Y_t = E[\Theta|\mathcal{G}_t \vee \mathcal{F}_{0, t}^B],$$ where $\Theta:=\xi_T + \int_t^T f(s)ds +\int_t^T g(s) d\overleftarrow{B}_s$ is $\mathcal{F}_{0, T}^W \vee \mathcal{F}_{t, T+K}^B$ measurable. Note that $\mathcal{F}_{0, t}^B$ is independent of $\mathcal{G}_t \vee \sigma(\Theta)$, then we know $$Y_t = E^{\mathcal{G}_t} [\Theta].$$ Now $$\int_t^T Z_s dW_s=\xi_T + \int_t^T f(s)ds +\int_t^T g(s) d\overleftarrow{B}_s- Y_t,$$ and the right side is $\mathcal{F}_{0, T}^W \vee \mathcal{F}_{t, T+K}^B$ measurable. Then from It\^{o}'s martingale representation theorem, $(Z_s)_{s\in[t, T]}$ is $\mathcal{F}_{0, s}^W \vee \mathcal{F}_{t, T+K}^B$ adapted, which implies $Z_s$ is $\mathcal{F}_{0, s}^W \vee \mathcal{F}_{t, T+K}^B$ measurable for any $t \leq s$. Thus $Z_s$ is $\mathcal{F}_{0, s}^W \vee \mathcal{F}_{s, T+K}^B$ measurable. To show the uniqueness. We suppose that $(\bar{Y}, \bar{Z})$ is the difference of two solutions. Then $$\bar{Y}_t + \int_t^T \bar{Z}_sdW_s=0, \ \ t\in [0, T].$$ Hence $$E |\bar{Y}_t|^2 + E \int_t^T |\bar{Z}_s|^2 ds =0,$$ which implies $\bar{Y}_t \equiv 0,\ a.s.$ and $\bar{Z}_t \equiv 0\ a.s.,\ a.e..$ \end{proof} Now we establish the main result of this part. \begin{theorem}\label{thm existence uniqueness} Assume that $(a1)$-$(a2)$ and $(H1)$-$(H4)$ hold. Then for given $(\xi, \eta) \in S_{\mathcal{G}}^2(T, T+K; \mathbb{R}^m)\times L_\mathcal{G}^2(T, T+K; \mathbb{R}^{m \times d})$, the anticipated BDSDE (\ref{equation:main}) has a unique solution $(Y, Z) \in S_\mathcal{G}^2(0, T+K; \mathbb{R}^m)\times L_\mathcal{G}^2(0, T+K; \mathbb{R}^{m\times d})$. \end{theorem} \begin{proof} Denote by $\mathcal{S}$ the space of $(Y, Z)\in L_{\mathcal{G}}^2(0, T+K; \mathbb{R}^m)\times L_{\mathcal{G}}^2(0, T+K; \mathbb{R}^{m \times d})$ such that $(Y_t, Z_t)_{t\in [T, T+K]}= (\xi_t, \eta_t)_{t\in [T, T+K]}$. Given $(y, z)\in \mathcal{S}$, we consider the following equation: \begin{equation}\label{equation: 3} \left\{ \begin{tabular}{rlll} $-dY_t$ &=& $f(t, y_t, z_t, y_{t+\delta(t)}, z_{t+\zeta(t)})dt$ & \vspace{2mm} \\ && $+ g(t, y_t, z_t, y_{t+\delta(t)}, z_{t+\zeta(t)}) d\overleftarrow{B}_t - Z_t d{W}_t, $ & $ t\in[0, T];$ \vspace{2mm} \\ $Y_t$ &=& $\xi_t, $ & $t\in[T, T+K];$ \vspace{2mm} \\ $Z_t$ &=& $\eta_t, $ & $t\in[T, T+K].$ \end{tabular}\right. \end{equation} It is obvious that the above equation is equivalent to the BDSDE \begin{equation*} \left\{ \begin{tabular}{rlll} $-d\tilde{Y}_t$ &=& $f(t, y_t, z_t, y_{t+\delta(t)}, z_{t+\zeta(t)})dt$ & \vspace{2mm} \\ && $+ g(t, y_t, z_t, y_{t+\delta(t)}, z_{t+\zeta(t)}) d\overleftarrow{B}_t - \tilde{Z}_t d{W}_t, $ & $ t\in[0, T];$ \vspace{2mm} \\ $\tilde{Y}_T$ &=& $\xi_T\in \mathcal{G}_T,$ \end{tabular}\right. \end{equation*} which admits a unique solution in the space $S_{\mathcal{G}}^2(0, T; \mathbb{R}^m)\times L_{\mathcal{G}}^2(0, T; \mathbb{R}^{m \times d})$ according to Theorem \ref{thm only t}. Thus BDSDE (\ref{equation: 3}) has a unique solution in $\mathcal{S}$. Define a mapping $I$ from $\mathcal{S}$ into itself by $(Y, Z)=I(y, z)$, then $(Y, Z)$ is the unique solution of BDSDE (\ref{equation: 3}). Let $(y^\prime, z^\prime)$ be another element of $\mathcal{S}$, and $(Y^\prime, Z^\prime)=I(y^\prime, z^\prime)$. We make the following notations: \begin{align*} & \bar{y}= y-y^\prime,\ \bar{z}=z-z^\prime,\ \bar{Y}=Y-Y^\prime,\ \bar{Z}=Z-Z^\prime, \\& \bar{f}_t = f(t, y_t, z_t, y_{t+\delta(t)}, z_{t+\zeta(t)})-f(t, y_t^\prime, z_t^\prime, y_{t+\delta(t)}^\prime, z_{t+\zeta(t)}^\prime), \\ & \bar{g}_t = g(t, y_t, z_t, y_{t+\delta(t)}, z_{t+\zeta(t)})-g(t, y_t^\prime, z_t^\prime, y_{t+\delta(t)}^\prime, z_{t+\zeta(t)}^\prime). \end{align*} For any $\beta >0$, apply It\^{o}'s formula to $e^{\beta t}|\bar{Y}_t|^2$, \begin{equation*} \begin{tabular}{rll} & & $e^{\beta t} |\bar{Y}_t|^2 + \int_t^T e^{\beta s} [\beta|\bar{Y}_s|^2+|\bar{Z}_s|^2]ds$ \vspace{2mm} \\ && = $2 \int_t^{T} e^{\beta s} \bar{Y}_s \bar{f}_sds + \int_t^T e^{\beta s} |\bar{g}_s|^2 ds + 2\int_t^T e^{\beta s} \bar{Y}_s \bar{g}_s d\overleftarrow{B}_s - 2\int_t^T e^{\beta s} \bar{Y}_s \bar{Z}_s d{W}_s.$ \end{tabular} \end{equation*} Take mathematical expectation on both sides, then we have \begin{equation*} e^{\beta t} E|\bar{Y}_t|^2 + E \int_t^T e^{\beta s} [\beta|\bar{Y}_s|^2+|\bar{Z}_s|^2]ds=2 E \int_t^{T} e^{\beta s} \bar{Y}_s \bar{f}_s ds + E \int_t^T e^{\beta s}|\bar{g}_s|^2 ds. \end{equation*} Hence from $(A1)$, $(A2)$ and the inequality $2ab \leq \lambda a^2 + \frac{1}{\lambda} b^2$, \begin{equation*} \begin{tabular}{rll} & & $e^{\beta t} E|\bar{Y}_t|^2 + E \int_t^T e^{\beta s} [\beta|\bar{Y}_s|^2+|\bar{Z}_s|^2]ds$ \vspace{2mm} \\& & $ \leq E \int_t^{T} e^{\beta s} [\lambda |\bar{Y}_s|^2+ \frac{1}{\lambda} |\bar{f}_s|^2] ds + E \int_t^T e^{\beta s}|\bar{g}_s|^2 ds$ \vspace{2mm} \\ && $\leq E\int_t^T e^{\beta s} [\lambda |\bar{Y}_s|^2+ (\frac{c}{\lambda}+c)(|\bar{y}_s|^2 + |\bar{y}_{s+\delta(s)}|^2)+ (\frac{c}{\lambda}+\alpha_1) |\bar{z}_s|^2 + (\frac{c}{\lambda}+\alpha_2) |\bar{z}_{s+\zeta(s)}|^2 ] ds$ \vspace{2mm} \\ && $\leq E\int_t^{T+K} e^{\beta s} [\lambda |\bar{Y}_s|^2+ (\frac{c}{\lambda}+c)(1 + M)|\bar{y}_s|^2 + (\frac{c}{\lambda}(1+M)+\alpha_1 +\alpha_2 M) |\bar{z}_s|^2] ds,$ \end{tabular} \end{equation*} which implies \begin{equation*} \begin{tabular}{rll} & & $E \int_t^{T+K} e^{\beta s} [(\beta-\lambda)|\bar{Y}_s|^2+|\bar{Z}_s|^2]ds$ \vspace{2mm} \\ && $\leq E\int_t^{T+K} e^{\beta s} [(\frac{c}{\lambda}+c)(1 + M)|\bar{y}_s|^2 + (\frac{c}{\lambda}(1+M)+\alpha_1 +\alpha_2 M) |\bar{z}_s|^2] ds$ \vspace{2mm} \\ && $= (\frac{c}{\lambda}(1+M)+\alpha_1 +\alpha_2 M) E\int_t^{T+K} e^{\beta s} [\frac{c(1+ \lambda)(1+M)}{c(1+ M) + \lambda (\alpha_1+\alpha_2 M)}|\bar{y}_s|^2 + |\bar{z}_s|^2] ds.$ \end{tabular} \end{equation*} Hence if we choose $\lambda=\lambda_0$ satisfying $\bar{c}:=\frac{c}{\lambda_0}(1+M)+\alpha_1 +\alpha_2 M <1 $, choose $\beta=\lambda_0 + \frac{c(1+ \lambda_0)(1+M)}{c(1+ M) + \lambda_0 (\alpha_1+\alpha_2 M)}$, and denote $\gamma:=\frac{c(1+ \lambda_0)(1+M)}{c(1+ M) + \lambda_0 (\alpha_1+\alpha_2 M)}$, then we deduce \begin{equation*} E \int_t^{T+K} e^{\beta s} [\gamma|\bar{Y}_s|^2+|\bar{Z}_s|^2]ds \leq \bar{c} E\int_t^{T+K} e^{\beta s} [\gamma|\bar{y}_s|^2 + |\bar{z}_s|^2] ds. \end{equation*} Thus $I$ is a strict contraction on $\mathcal{S}$ and it has a unique fixed point $(Y, Z)\in \mathcal{S}$. Now due to Burkholder-Davis-Gundy inequality, it is easy to check that $Y \in S_{\mathcal{G}}^2(0, T+K; \mathbb{R}^m)$. The proof is complete. \end{proof} \begin{remark} In the proof of Theorem \ref{thm existence uniqueness}, we use the norm \begin{equation*} |(Y, Z)|_{(\beta, \gamma)} \equiv \{E \int_0^{T+K} e^{\beta_s}(\gamma |Y_s|^2+|Z_s|^2)ds\}^{\frac{1}{2}}, \end{equation*} which is very convenient for us to establish a strict contraction mapping. In fact, it is obvious that this new norm is equivalent to both norms $|(Y, Z)|_{(\beta, 1)}$ and $|(Y, Z)|_{(0, 1)}$, and the latter is just the general norm defined on the space $L_\mathcal{G}^2(0, T+K; \mathbb{R}^m)\times L_\mathcal{G}^2(0, T+K; \mathbb{R}^{m\times d})$. \end{remark} \section{Comparison theorem} In this part, we are concerned with the following $1$-dimensional anticipated BDSDEs: \begin{equation}\label{equation:comparison j} \left\{ \begin{tabular}{rlll} $-dY_t^j$ &=& $f^j(t, Y_t^j, Z_t^j, Y_{t+\delta(t)}^j, Z_{t+\zeta(t)}^j)dt+ g(t, Y_t^j, Z_t^j) d\overleftarrow{B}_t - Z_t^j d{W}_t, $ & $ t\in[0, T];$ \vspace{2mm} \\ $Y_t^j$ &=& $\xi_t^j, $ & $t\in[T, T+K],$ \vspace{2mm} \\ $Z_t^j$ &=& $\eta_t^j, $ & $t\in[T, T+K],$ \end{tabular}\right. \end{equation} where $j=1, 2$, and $(a1)$-$(a2)$, $(H1)$-$(H4)$ hold. Then by Theorem \ref{thm existence uniqueness}, (\ref{equation:comparison j}) has a unique solution. Our objective is to obtain a comparison result for these two equations. For this purpose, we first consider a simple case when the coefficients $f^j$ and $g$ do not depend on the future value of $(Y^j, Z^j)$: \begin{equation}\label{eq j only t} Y_t^j = \xi_T^j+ \int_t^T f^j(s, Y_s^j, Z_s^j)ds + \int_t^T g(s, Y_s^j, Z_s^j) d\overleftarrow{B}_s - \int_t^T Z_s^j d{W}_s,\ \ t\in[0, T]. \end{equation} \begin{theorem}\label{thm comparison only t} Let $(Y^{j}, Z^{j})\in S_\mathcal{G}^2(0, T; \mathbb{R})\times L_\mathcal{G}^2(0, T; \mathbb{R}^d)$ $(j=1, 2)$ be the unique solutions to BDSDEs (\ref{eq j only t}) respectively. If $\xi_T^1 \geq \xi_T^2,\ a.s.,$ and for any $(t, y, z) \in [0, T]\times\mathbb{R}\times \mathbb{R}^d$, $f^1(t, y, z) \geq f^2(t, y, z),\ a.s.$, then $Y_t^1 \geq Y_t^2,\ a.s.,$ for all $t\in [0,T]$. \end{theorem} \begin{proof} Denote $$\bar{Y}_t:=Y_t^2-Y_t^1,\ \ \bar{Z}_t:=Z_t^2-Z_t^1,\ \ \bar{\xi}_T:=\xi_T^2-\xi_T^1,$$ then $(\bar{Y}, \bar{Z})$ satisfies $$\bar{Y}_t=\bar{\xi}_T + \int_t^T [f^2(s, Y_s^2, Z_s^2)-f^1(s, Y_s^1, Z_s^1)]ds + \int_t^T [g(s, Y_s^2, Z_s^2)-g(s, Y_s^1, Z_s^1)]d \overleftarrow{B}_s - \int_t^T \bar{Z}_sdW_s.$$ Applying It\^{o}'s formula to $|\bar{Y}_t^+|^2$, we have \begin{equation*} \begin{tabular}{rll} $|\bar{Y}_t^+|^2$ & = & $|\bar{\xi}_T^+|^2 + 2\int_t^T \bar{Y}_s^+ [f^2(s, Y_s^2, Z_s^2)-f^1(s, Y_s^1, Z_s^1)]ds $ \vspace{5mm} \\ & & $+ 2\int_t^T \bar{Y}_s^+ [g(s, Y_s^2, Z_s^2)-g(s, Y_s^1, Z_s^1)]d \overleftarrow{B}_s- 2\int_t^T \bar{Y}_s^+ \bar{Z}_sdW_s$\vspace{5mm} \\ && $-\int_t^T 1_{\{Y_s^2 > Y_s^1\}}|\bar{Z}_s|^2ds +\int_t^T 1_{\{Y_s^2 > Y_s^1\}} |g(s, Y_s^2, Z_s^2)-g(s, Y_s^1, Z_s^1)|^2ds.$\\ \end{tabular} \end{equation*} Taking expectation on both sides and noting that $\xi_T^1 \geq \xi_T^2$, we get \begin{equation*} \begin{tabular}{rll} $E|\bar{Y}_t^+|^2 + E-\int_t^T 1_{\{Y_s^2 > Y_s^1\}}|\bar{Z}_s|^2ds$ & = & $2E\int_t^T \bar{Y}_s^+ [f^2(s, Y_s^2, Z_s^2)-f^1(s, Y_s^1, Z_s^1)]ds $ \vspace{5mm} \\ & & $+ E\int_t^T 1_{\{Y_s^2 > Y_s^1\}} |g(s, Y_s^2, Z_s^2)-g(s, Y_s^1, Z_s^1)|^2ds.$\\ \end{tabular} \end{equation*} While, \begin{equation*} \begin{tabular}{rll} & $2E\int_t^T \bar{Y}_s^+ [f^2(s, Y_s^2, Z_s^2)-f^1(s, Y_s^1, Z_s^1)]ds$ \vspace{5mm}\\ &=$2 E\int_t^T \bar{Y}_s^+ [f^2(s, Y_s^2, Z_s^2)-f^1(s, Y_s^2, Z_s^2)+f^1(s, Y_s^2, Z_s^2)-f^1(s, Y_s^1, Z_s^1)]ds$\vspace{5mm}\\ & $\leq 2 E\int_t^T \bar{Y}_s^+ |f^1(s, Y_s^2, Z_s^2)-f^1(s, Y_s^1, Z_s^1)|ds \leq 2\sqrt{c}E\int_t^T \bar{Y}_s^+ [|\bar{Y}_s|+|\bar{Z}_s|]ds$ \vspace{5mm}\\ & $\leq (2\sqrt{c}+\frac{c}{1-\alpha_1})E\int_t^T |\bar{Y}_s^+|^2ds+(1-\alpha_1)\int_t^T 1_{\{Y_s^2 > Y_s^1\}}|\bar{Z}_s|^2ds,$ \end{tabular} \end{equation*} and \begin{equation*} \begin{tabular}{rll} & $E\int_t^T 1_{\{Y_s^2 > Y_s^1\}} |g(s, Y_s^2, Z_s^2)-g(s, Y_s^1, Z_s^1)|^2ds$ \vspace{5mm}\\ & $\leq E\int_t^T 1_{\{Y_s^2 > Y_s^1\}}[c|\bar{Y}_s|^2+\alpha_1|\bar{Z}_s|^2]ds$ \vspace{5mm}\\ & $\leq c E\int_t^T |\bar{Y}_s^+|^2ds +\alpha_1 \int_t^T 1_{\{Y_s^2 > Y_s^1\}}|\bar{Z}_s|^2ds.$ \end{tabular} \end{equation*} Then, thanks to the above inequalities, we obtain $$E|\bar{Y}_t^+|^2 \leq (c+2\sqrt{c}+\frac{c}{1-\alpha_1})E\int_t^T |\bar{Y}_s^+|^2ds,$$ which implies $$E|\bar{Y}_t^+|^2=0,\ \ {\rm{for\ all}}\ t\in [0, T].$$ Therefore $Y_t^1 \geq Y_t^2,\ a.s.,$ for all $t\in [0, T]$. \end{proof} From now on, we consider the anticipated BDSDEs (\ref{equation:comparison j}). We give the following result. For the proof, the reader is referred to \cite{Xu}. \begin{proposition} Putting $t_0=T$, we define by iteration \begin{equation*} t_i:=\min\{t \in [0, T]: \min\{s+ \delta(s),\ s+\zeta(s)\}\geq t_{i-1}, {\rm{\ for\ all}}\ s\in [t, T]\},\ \ \ i\geq 1. \end{equation*} Set $N:=\max\{i: t_{i-1}>0\}$. Then $N$ is finite, $t_N=0$ and \begin{equation*} [0, T]=[0, t_{N-1}]\cup [t_{N-1}, t_{N-2}]\cup \cdots \cup [t_2, t_1]\cup [t_1, T]. \end{equation*} \end{proposition} \begin{proposition}\label{prop2} For $j=1, 2$, suppose that $(Y^{j}, Z^{j})$ is the unique solution to the anticipated BDSDE (\ref{equation:comparison j}). Then for fixed $i\in {\{1, 2,\ldots, N\}},$ over time interval $[t_i, t_{i-1}]$, (\ref{equation:comparison j}) is equivalent to \begin{equation}\label{equation:in prop 2} \left\{ \begin{tabular}{rlll} $-d\bar{Y}_t^j$ &=& $f^j(t, \bar{Y}_t^j, \bar{Z}_t^j, \bar{Y}_{t+\delta(t)}^j, \bar{Z}_{t+\zeta(t)}^j)dt+ g(t, \bar{Y}_t^j, \bar{Z}_t^j) d\overleftarrow{B}_t - \bar{Z}_t^j d{W}_t, $ & $ t\in[t_i, t_{i-1}];$ \vspace{2mm} \\ $\bar{Y}_t^j$ &=& $Y_t^j, $ & $t\in[t_{i-1}, T+K],$ \vspace{2mm} \\ $\bar{Z}_t^j$ &=& $Z_t^j, $ & $t\in[t_{i-1}, T+K],$ \end{tabular}\right. \end{equation} which is also equivalent to the following BDSDE with terminal condition $Y_{t_{i-1}}^{j}$: \begin{equation}\label{equation:in prop 3} \tilde{Y}_t^{j}=Y_{t_{i-1}}^{j}+\int_t^{t_{i-1}} f^j(s, \tilde{Y}_s^j, \tilde{Z}_s^j, Y_{s+\delta(s)}^j, Z_{s+\zeta(s)}^j)ds+ \int_t^{t_{i-1}} g(s, \tilde{Y}_s^j, \tilde{Z}_s^j) d\overleftarrow{B}_s - \int_t^{t_{i-1}} \tilde{Z}_s^j d{W}_s. \end{equation} That is to say, $$ Y_t^{j}=\bar{Y}_t^{j}=\tilde{Y}_t^{j},\ Z_t^{j}=\bar{Z}_t^{j}=\tilde{Z}_t^{j}=\frac{d\langle \tilde{Y}^{j}, W \rangle_t}{d t},\ t\in [t_i, {t_{i-1}}],\ j=1,2.,$$ where $\langle \tilde{Y}^{j}, W \rangle$ is the variation process generated by $\tilde{Y}^{j}$ and the Brownian motion $W$. \end{proposition} The main result of this part is \begin{theorem}\label{thm comparison} Let $(Y^{j}, Z^{j})\in S_\mathcal{G}^2(0, T+K; \mathbb{R})\times L_\mathcal{G}^2(0, T+K; \mathbb{R}^d)$ $(j=1, 2)$ be the unique solutions to anticipated BDSDEs (\ref{equation:comparison j}) respectively. If \item{(i)} $\xi_s^{1}\geq \xi_s^{2}, s\in [T, T+K], a.e., a.s.;$ \item{(ii)} for all $t\in [0, T]$, $(y, z)\in \mathbb{R}\times \mathbb{R}^d,$ $\theta^{j}\in S_\mathcal{G}^2(t, T+K; \mathbb{R})$ $(j=1, 2)$ such that $\theta^{1} \geq \theta^{2},$ $\{\theta_{r}^{j}\}_{r\in[t, T]}$ is a continuous semimartingale and $(\theta_{r}^{j})_{r\in [T, T+K]}=(\xi_r^{j})_{r\in [T, T+K]}$, \begin{equation}\label{equation:condition 1} f^1(t, y, z, \theta_{t+\delta(t)}^{1}, \eta_{t+\zeta(t)}^{1}) \geq f^2(t, y, z, \theta_{t+\delta(t)}^{2}, \eta_{t+\zeta(t)}^{2}),\ \ a.e., a.s., \end{equation} \begin{equation}\label{equation:condition 2} f^1(t, y, z, \theta_{t+\delta(t)}^{1}, \frac{d\langle \theta^{1}, W\rangle_r}{d r}|_{r=t+\zeta(t)}) \geq f^2(t, y, z, \theta_{t+\delta(t)}^{2}, \frac{d\langle \theta^{2}, W\rangle_r}{d r}|_{r=t+\zeta(t)}),\ \ a.e., a.s., \end{equation} \begin{equation}\label{equation:condition 3} f^1(t, y, z, \xi_{t+\delta(t)}^{1}, \frac{d\langle \theta^{1}, W\rangle_r}{d r}|_{r=t+\zeta(t)}) \geq f^2(t, y, z, \xi_{t+\delta(t)}^{2}, \frac{d\langle \theta^{2}, W\rangle_r}{d r}|_{r=t+\zeta(t)}),\ \ a.e., a.s., \end{equation} then $Y_t^{1} \geq Y_t^{2}, a.e., a.s..$ \end{theorem} \begin{proof} Consider the anticipated BDSDE (\ref{equation:comparison j}) one time interval by one time interval. For the first step, we consider the case when $t\in [t_1, T]$. According to Proposition \ref{prop2}, we can equivalently consider \begin{equation*} \tilde{Y}_t^{j}=\xi_T^{j}+\int_t^T f^j(s, \tilde{Y}_s^j, \tilde{Z}_s^j, \xi_{s+\delta(s)}^j, \eta_{s+\zeta(s)}^j)ds+ \int_t^T g(s, \tilde{Y}_s^j, \tilde{Z}_s^j) d\overleftarrow{B}_s - \int_t^T \tilde{Z}_s^j d{W}_s, \end{equation*} from which we have \begin{equation}\label{equation:Z t1 T} {Z}_t^{j}=\tilde{Z}_t^{j}=\frac{d\langle \tilde{Y}^{j}, W\rangle_t}{d t},\ \ t\in [t_1, T]. \end{equation} Noticing that $\xi^{j}\in S_\mathcal{G}^2(T, T+K; \mathbb{R})$ $(j=1, 2)$ and $\xi^{1}\geq \xi^{2},$ from (\ref{equation:condition 1}) in (ii), we can get, for $s\in [t_1, T],$ $y\in \mathbb{R}$, $z\in \mathbb{R}^d,$ \begin{center} $f^1(s, y, z, \xi_{s+\delta(s)}^{1}, \eta_{s+\zeta(s)}^{1}) \geq f^2(s, y, z, \xi_{s+\delta(s)}^{2}, \eta_{s+\zeta(s)}^{2}). $ \end{center} According to Theorem \ref{thm comparison only t}, we can get \begin{center} $ \tilde{Y}_t^{1}\geq \tilde{Y}_t^{2},\ \ t\in[t_1, T],\ \ a.e.,a.s., $ \end{center} which implies \begin{equation}\label{equation:Y t1 T} Y_t^{(1)}\geq Y_t^{(2)},\ \ t\in[t_1, T+K],\ \ a.e.,a.s.. \end{equation} For the second step, we consider the case when $t\in [t_2, t_1]$. Similarly, according to Proposition \ref{prop2}, we can consider the following BSDE equivalently: \begin{equation*} \tilde{\tilde{Y}}_t^{j}=Y_{t_1}^{j}+\int_t^{t_1} f^j(s, \tilde{\tilde{Y}}_s^j, \tilde{\tilde{Z}}_s^j, Y_{s+\delta(s)}^j, Z_{s+\zeta(s)}^j)ds+ \int_t^{t_1} g(s, \tilde{\tilde{Y}}_s^j, \tilde{\tilde{Z}}_s^j) d\overleftarrow{B}_s - \int_t^{t_1} \tilde{\tilde{Z}}_s^j d{W}_s, \end{equation*} from which we have ${Z}_t^{j}=\tilde{\tilde{Z}}_t^{j}=\frac{d\langle \tilde{\tilde{Y}}^{j}, W \rangle_t}{d t}$ for $t\in [t_2, t_1]$. Noticing (\ref{equation:Z t1 T}) and (\ref{equation:Y t1 T}), according to (ii), we have, for $s\in [t_2, t_1]$, $y\in \mathbb{R}$, $z\in \mathbb{R}^d,$ \begin{center} $f^1(s, y, z, Y_{s+\delta(s)}^{1}, Z_{s+\zeta(s)}^{1}) \geq f^2(s, y, z, Y_{s+\delta(s)}^{2}, Z_{s+\zeta(s)}^{2}). $ \end{center} Applying Theorem \ref{thm comparison only t} again, we can finally get \begin{center} $ Y_t^{1}\geq Y_t^{2},\ \ t\in[t_2, t_1],\ \ a.e.,a.s.. $ \end{center} Similarly to the above steps, we can give the proofs for the other cases when $t\in [t_3, t_2],$ $[t_4, t_3],$ $\cdots,$ $[t_N, t_{N-1}].$ \end{proof} \begin{example} Now suppose that we are facing with the following two ABDSDEs: \begin{equation*} \left\{ \begin{tabular}{rlll} $-dY_t^1$ &=& $E^{\mathcal{F}_t}[Y_{t+\delta(t)}^{1}+\sin (2Y_{t+\delta(t)}^{1})+|Z_{t+\zeta(t)}^{1}|+2]dt$ \vspace{2mm} \\ && $ + [Y_{t}^{1}+\frac{1}{\sqrt{3}}|Z_{t}^{1}|] d\overleftarrow{B}_t - Z_t^1 d{W}_t, $ & $ t\in[0, T];$ \vspace{2mm} \\ $Y_t^1$ &=& $\xi_t^1, $ & $t\in[T, T+K],$ \vspace{2mm} \\ $Z_t^1$ &=& $\eta_t^1, $ & $t\in[T, T+K],$ \end{tabular}\right. \end{equation*} \begin{equation*} \left\{ \begin{tabular}{rlll} $-dY_t^2$ &=& $E^{\mathcal{F}_t}[Y_{t+\delta(t)}^{2}+2 |\cos Y_{t+\delta(t)}^{2}|+\sin Z_{t+\zeta(t)}^{2}-2]dt$ \vspace{2mm} \\ && $ + [Y_{t}^{2}+\frac{1}{\sqrt{3}}|Z_{t}^{2}|] d\overleftarrow{B}_t - Z_t^2 d{W}_t, $ & $ t\in[0, T];$ \vspace{2mm} \\ $Y_t^2$ &=& $\xi_t^2, $ & $t\in[T, T+K],$ \vspace{2mm} \\ $Z_t^2$ &=& $\eta_t^2, $ & $t\in[T, T+K],$ \end{tabular}\right. \end{equation*} where $\xi_t^{(1)}\geq \xi_t^{(2)}, t\in [T, T+K].$ It is obvious that \begin{center} $x+\sin (2x)+|u|+2 \geq y+2 |\cos y| + \sin v-2,\ for\ all\ x\geq y, x, y\in \mathbb{R}, u, v \in \mathbb{R}^d,$ \end{center} which implies (\ref{equation:condition 1})-(\ref{equation:condition 3}), then according to Theorem \ref{thm comparison}, we get $Y_t^{1}\geq Y_t^{2},\ a.e.,\ a.s..$ \end{example} \begin{remark} By the same way, for the case when $\delta=\zeta$, (\ref{equation:condition 1})-(\ref{equation:condition 3}) can be replaced by (\ref{equation:condition 2}) together with \begin{center} $ f^1(t, y, z, \xi_{t+\delta(t)}^{1}, \eta_{t+\zeta(t)}^{1}) \geq f^2(t, y, z, \xi_{t+\delta(t)}^{2}, \eta_{t+\zeta(t)}^{2}),\ \ a.e., a.s.. $ \end{center} \end{remark} For a special case when $f^1$ and $f^2$ are independent of the anticipated term $Z$, we easily get the following comparison result. \begin{theorem}\label{thm comparison no Z} Let $(Y^{j}, Z^{j})\in S_\mathcal{G}^2(0, T+K; \mathbb{R})\times L_\mathcal{G}^2(0, T+K; \mathbb{R}^d)$ $(j=1, 2)$ be the unique solutions to the following ABDSDEs respectively: \begin{equation*} \left\{ \begin{tabular}{rlll} $-dY_t^j$ &=& $f^j(t, Y_t^j, Z_t^j, Y_{t+\delta(t)}^j)dt+ g(t, Y_t^j, Z_t^j) d\overleftarrow{B}_t - Z_t^j d{W}_t, $ & $ t\in[0, T];$ \vspace{2mm} \\ $Y_t^j$ &=& $\xi_t^j, $ & $t\in[T, T+K].$ \end{tabular}\right. \end{equation*} If \item{(i)} $\xi_s^{1}\geq \xi_s^{2}, s\in [T, T+K], a.e., a.s.;$ \item{(ii)} for all $t\in [0, T]$, $(y, z)\in \mathbb{R}\times \mathbb{R}^d,$ $\theta^{j}\in S_\mathcal{G}^2(t, T+K; \mathbb{R})$ $(j=1, 2)$ such that $\theta^{1} \geq \theta^{2}$ and $(\theta_{r}^{j})_{r\in [T, T+K]}=(\xi_r^{j})_{r\in [T, T+K]}$, \begin{equation}\label{equation:no z} f^1(t, y, z, \theta_{t+\delta(t)}^{1}) \geq f^2(t, y, z, \theta_{t+\delta(t)}^{2}),\ \ a.e., a.s., \end{equation} then $Y_t^{1} \geq Y_t^{2}, a.e., a.s..$ \end{theorem} \begin{remark} The coefficients $f^1$ and $f^2$ will satisfy (\ref{equation:no z}), if for any $(t, y, z)\in [0, T]\times \mathbb{R}\times \mathbb{R}^d,$ $\theta\in L_\mathcal{G}^2(t, T+K; \mathbb{R}),$ $r\in [t, T+K],$ $f^1(t, y, z, \theta_r)\geq f^2(t, y, z, \theta_r),$ together with one of the following: \item{(i)} for any $(t, y, z)\in [0, T]\times \mathbb{R}\times \mathbb{R}^d,$ $f^1(t, y, z, \cdot)$ is increasing, i.e., $f^1(t, y, z, \theta_r)\geq f^1(t, y, z, \theta_r^\prime),$ if $\theta \geq \theta^\prime,$ $\theta, \theta^\prime\in L_\mathcal{G}^2(t, T+K; \mathbb{R}),$ $r\in [t, T+K];$ \item{(ii)} for any $(t, y, z)\in [0, T]\times \mathbb{R}\times \mathbb{R}^d,$ $f^2(t, y, z, \cdot)$ is increasing, i.e., $f^2(t, y, z, \theta_r)\geq f^2(t, y, z, \theta_r^\prime),$ if $\theta \geq \theta^\prime,$ $\theta, \theta^\prime\in L_\mathcal{G}^2(t, T+K; \mathbb{R}),$ $r\in [t, T+K].$ \end{remark} \begin{remark}\label{remark} The coefficients $f^1$ and $f^2$ will satisfy (\ref{equation:no z}), if \begin{center} $ f^1(t, y, z, \theta_r)\geq \tilde{f}(t, y, z, \theta_r)\geq f^2(t, y, z, \theta_r), $ \end{center} for any $(t, y, z)\in [0, T]\times \mathbb{R}\times \mathbb{R}^d,$ $\theta \in L_\mathcal{G}^2(t, T+K; \mathbb{R}), r\in [t, T+K].$ Here the function $\tilde{f}(t, y, z, \cdot)$ is increasing, for any $(t, y, z)\in [0, T]\times \mathbb{R}\times \mathbb{R}^d,$ i.e., $\tilde{f}(t, y, z, \theta_r)\geq \tilde{f}(t, y, z, \theta_r^\prime),$ if $\theta_r\geq \theta_r^\prime,$ $\theta, \theta^\prime\in L_\mathcal{G}^2(t, T+K; \mathbb{R}), r\in [t, T+K].$ \end{remark} \begin{example} The following three functions satisfy the conditions in Remark \ref{remark}: $f^1(t, y, z, \theta_r)=E^{\mathcal{F}_t}[\theta_r- \sin (2\theta_r) + 2]$, $\tilde{f}(t, y, z, \theta_r)=E^{\mathcal{F}_t}[\theta_r+\cos \theta_r]$, $f^2(t, y, z, \theta_r)=E^{\mathcal{F}_t}[\theta_r+2\cos \theta_r-1].$ \end{example} \section{A duality result between delayed DSDEs and anticipated BDSDEs} In this part we will establish a duality between the following anticipated BDSDE \begin{equation}\label{eq duality bdsde} \left\{ \begin{tabular}{rlll} $-dY_t$ &=& $([\mu_t+\kappa_t^2] Y_t+ \bar{\mu}_t E^{\mathcal{F}_{t, T}^B}[Y_{t+\delta}] + \sigma_t Z_t+ \bar{\sigma}_t E^{\mathcal{F}_{t, T}^B}[Z_{t+\delta}] + \rho_t) dt$ & \vspace{2mm} \\ && $+ \kappa_t Y_t d \overleftarrow{B}_t - Z_t d{W}_t, $ & $ t\in[t_0, T];$ \vspace{2mm} \\ $Y_t$ &=& $\xi_t, $ & $t\in[T, T+\delta];$ \vspace{2mm} \\ $Z_t$ &=& $\eta_t, $ & $t\in[T, T+\delta]$ \end{tabular}\right. \end{equation} and the delayed DSDE \begin{equation}\label{eq duality sde} \left\{ \begin{tabular}{rlll} $dX_s$ &=& $(\mu_s X_s +\bar{\mu}_{s-\delta} X_{s-\delta})ds + \kappa_s X_s d\overleftarrow{B}_s + (\sigma_s X_s +\bar{\sigma}_{s-\delta} X_{s-\delta}) d{W}_s, $ & $ s\in[t, T];$ \vspace{2mm} \\ $X_t$ &=& $1, $ & $$ \vspace{2mm} \\ $X_s$ &=& $0, $ & $s\in[t-\delta, t),$ \end{tabular}\right. \end{equation} where we suppose that $t_0 \geq \delta >0$ are fixed constants, $(\xi, \eta) \in S_{\mathcal{G}}^2(T, T+\delta; \mathbb{R})\times L_{\mathcal{G}}^2(T, T+\delta; \mathbb{R}^d)$ with $\xi_T\in L^2({\mathcal{F}_{T, T}^B}; \mathbb{R})$, $\mu_t$, $\bar{\mu}_t\in L_{\mathcal{F}_{t, T}^B}^2(t_0-\delta, T+\delta; \mathbb{R})$, $\sigma_t, \bar{\sigma}_t \in L_{\mathcal{F}_{t, T}^B}^2(t_0-\delta, T+\delta; \mathbb{R}^d)$, $\kappa_t \in L_{\mathcal{F}_{t, T}^B}^2(t_0, T; \mathbb{R}^l)$, $\rho_t \in L_{\mathcal{F}_{t, T}^B}^2(t_0, T; \mathbb{R})$, and $\mu$, $\bar{\mu}$, $\sigma$,$\bar{\sigma}$, $\kappa$ are uniformly bounded. Then by Theorem \ref{thm existence uniqueness}, (\ref{eq duality bdsde}) has a unique solution. \begin{proposition}\label{prop duality} Let $(Y, Z) \in S_\mathcal{G}^2(t_0, T+\delta; \mathbb{R})\times L_\mathcal{G}^2(t_0, T+\delta; \mathbb{R}^d)$ be the unique solution of ABDSDE (\ref{eq duality bdsde}). Then for $t\in[t_0, T]$, $Z_t\equiv 0$, and $Y_t$ is $\mathcal{F}_{t, T}^B$-progressively measurable. \end{proposition} \begin{proof} First we show that $Y_t$ is $\mathcal{F}_{t, T}^B$-progressively measurable. For this we introduce the following auxiliary equation: \begin{equation}\label{eq duality bdsde auxiliary} \left\{ \begin{tabular}{rlll} $-dY_t^\prime$ &=& $E^{\mathcal{F}_{t, T}^B}[(\mu_t+\kappa_t^2) Y_t^\prime+\bar{\mu}_t Y_{t+\delta}^\prime + \sigma_t Z_t^\prime + \bar{\sigma}_t Z_{t+\delta}^\prime + \rho_t] dt$ & \vspace{2mm} \\ && $+ \kappa_t E^{\mathcal{F}_{t, T}^B} [Y_t^\prime] d \overleftarrow{B}_t - Z_t^\prime d{W}_t, $ & $ t\in[t_0, T];$ \vspace{2mm} \\ $Y_t^\prime$ &=& $\xi_t, $ & $t\in[T, T+\delta];$ \vspace{2mm} \\ $Z_t^\prime$ &=& $\eta_t, $ & $t\in[T, T+\delta]$ \end{tabular}\right. \end{equation} which has a unique solution according to Theorem \ref{thm existence uniqueness}. In fact, it is obvious that $$Y_t^\prime = E^{\mathcal{G}_t}[\Theta^\prime],$$ where $$\Theta^\prime:=\xi_T + \int_t^T E^{\mathcal{F}_{s, T}^B}[(\mu_s+\kappa_s^2) Y_s^\prime+\bar{\mu}_s Y_{s+\delta}^\prime + \sigma_s Z_s^\prime + \bar{\sigma}_s Z_{s+\delta}^\prime + \rho_s] ds +\int_t^T \kappa_s E^{\mathcal{F}_{s, T}^B} [Y_s^\prime] d\overleftarrow{B}_s$$ is $\mathcal{F}_{t, T}^B$ measurable thanks to the fact that $\xi_T\in L^2({\mathcal{F}_{T, T}^B}; \mathbb{R})$ and $\mu_t$, $\bar{\mu}_t$, $\sigma_t$, $\bar{\sigma}_t$, $\kappa_t$, $\rho_t$ are all $\mathcal{F}_{t, T}^B$-progressively measurable. Note that $\mathcal{F}_{0, t}^W \vee \mathcal{F}_{T, T+K}^B$ is independent of $\mathcal{F}_{t, T}^B \vee \sigma(\Theta^\prime)$, hence we know $$Y_t^\prime = E^{\mathcal{F}_{t, T}^B} [\Theta^\prime].$$ Thus obviously $Z_t^\prime \equiv 0$, and moreover, $E^{\mathcal{F}_{t, T}^B}[Y_t^\prime]=Y_t^\prime$, $E^{\mathcal{F}_{t, T}^B}[Z_t^\prime]=Z_t^\prime$. Then by Comparing the anticipated BDSDE (\ref{eq duality bdsde}) with (\ref{eq duality bdsde auxiliary}), together with the uniqueness of their solutions, we immediately get the desired conclusion. \end{proof} The next is our main result. \begin{theorem} For any $(\xi, \eta) \in S_{\mathcal{G}}^2(T, T+\delta; \mathbb{R})\times L_{\mathcal{G}}^2(T, T+\delta; \mathbb{R}^d)$ with $\xi_T\in L^2({\mathcal{F}_{T, T}^B}; \mathbb{R})$, the solution $Y_\cdot$ of the anticipated BDSDE (\ref{eq duality bdsde}) can be given by \begin{equation*} Y_t = E^{\mathcal{F}_{t, T}^B} [X_T \xi_T + \int_t^T \rho_s X_s ds] + E^{\mathcal{F}_{t, T}^B} [\int_{T}^{T+\delta} (\bar{\mu}_{s-\delta} X_{s-\delta} E^{\mathcal{F}_{s-\delta, T}^B}[\xi_{s}] +\bar{\sigma}_{s-\delta} X_{s-\delta} E^{\mathcal{F}_{s-\delta, T}^B}[\eta_{s}]) ds], \end{equation*} where $X_\cdot$ is the unique solution of delayed DSDE (\ref{eq duality sde}). \end{theorem} \begin{proof} We first show that DSDE (\ref{eq duality sde}) has a unique solution. In fact, when $s\in [t, t+\delta]$, \begin{equation}\label{eq duality sde 1} \left\{ \begin{tabular}{rlll} $dX_s$ &=& $\mu_s X_s ds + \kappa_s X_s d\overleftarrow{B}_s + \sigma_s X_s d{W}_s, $ & $ s\in[t, t+\delta];$ \vspace{2mm} \\ $X_t$ &=& $1. $ & $$ \end{tabular}\right. \end{equation} Then we can easily obtain a unique solution $\varsigma_\cdot^1$ for (\ref{eq duality sde 1}). When $s\in [t+\delta, t+2\delta]$, \begin{equation}\label{eq duality sde 2} \left\{ \begin{tabular}{rlll} $dX_s$ &=& $ (\mu_s X_s + \bar{\mu}_{s-\delta}\varsigma_{s-\delta}^1)ds + \kappa_s X_s d\overleftarrow{B}_s + (\sigma_s X_s + \bar{\sigma}_{s-\delta}\varsigma_{s-\delta}^1) d{W}_s, $ & $ s\in[t+\delta, t+2\delta];$ \vspace{2mm} \\ $X_{t+\delta}$ &=& $\varsigma_{t+\delta}^1. $ & $$ \end{tabular}\right. \end{equation} Then we can easily obtain a unique solution $\varsigma_\cdot^2$ for (\ref{eq duality sde 2}). Similarly ,we can consider all the other cases when $t\in [t+2\delta, t+3\delta]$, $[t+3\delta, t+4\delta]$, $\cdots$, $[t+ [\frac{T-t}{\delta}]\delta, T]$. Thus DSDE (\ref{eq duality sde}) has a unique solution $X \in \mathcal{S}_{\tilde{\mathcal{G}}}^2(t-\delta, T; \mathbb{R})$ where $\tilde{\mathcal{G}}_t:= \mathcal{F}_t^W \vee \mathcal{F}_t^B$. Applying It\^{o}'s formula to $X_sY_s$, according to Proposition \ref{prop duality}, we have \begin{equation}\label{eq duality 1} \begin{tabular}{rlll} & $X_T Y_T - X_t Y_t - \int_t^T (X_s Z_s + \sigma_s X_s Y_s + \bar{\sigma}_{s-\delta} X_{s-\delta} Y_s) d{W}_s $ \vspace{5mm} \\ = & $\int_t^T (\bar{\mu}_{s-\delta} X_{s-\delta} Y_s -\bar{\mu}_s X_s E^{\mathcal{F}_{s, T}^B}[Y_{s+\delta}] + \bar{\sigma}_{s-\delta} X_{s-\delta} Z_s -\bar{\sigma}_s X_s E^{\mathcal{F}_{s, T}^B}[Z_{s+\delta}]- \rho_s X_s) ds$ \vspace{5mm} \\ = & $\int_t^{T-\delta} (\bar{\mu}_{s-\delta} X_{s-\delta} Y_s -\bar{\mu}_s X_s Y_{s+\delta}) ds$ \vspace{5mm} \\ & $+ \int_{T-\delta}^T (\bar{\mu}_{s-\delta} X_{s-\delta} Y_s -\bar{\mu}_s X_s E^{\mathcal{F}_{s, T}^B}[\xi_{s+\delta}] -\bar{\sigma}_s X_s E^{\mathcal{F}_{s, T}^B}[\eta_{s+\delta}]) ds - \int_t^{T} \rho_s X_s ds.$ \end{tabular} \end{equation} Write $\Delta=\int_t^{T-\delta} (\bar{\mu}_{s-\delta} X_{s-\delta} Y_s -\bar{\mu}_s X_s Y_{s+\delta})ds$, then \begin{equation}\label{eq duality 2} \begin{tabular}{rlll} $\Delta =$ & $\int_t^{T-\delta} (\bar{\mu}_{s-\delta} X_{s-\delta} Y_s -\bar{\mu}_s X_s Y_{s+\delta}) ds= \int_t^{T-\delta} \bar{\mu}_{s-\delta} X_{s-\delta} Y_s ds - \int_{t+\delta}^{T} \bar{\mu}_{s-\delta} X_{s-\delta} Y_s ds$ \vspace{5mm} \\ = & $\int_t^{t+\delta} \bar{\mu}_{s-\delta} X_{s-\delta} Y_s ds - \int_{T-\delta}^{T} \bar{\mu}_{s-\delta} X_{s-\delta} Y_s ds = - \int_{T-\delta}^{T} \bar{\mu}_{s-\delta} X_{s-\delta} Y_s ds,$ \end{tabular} \end{equation} and the last equality is due to the fact that $X_s=0,\ s\in [t-\delta, t)$. \\ Combining (\ref{eq duality 1}) and (\ref{eq duality 2}), we have \begin{align*} & X_T Y_T - X_t Y_t - \int_t^T (X_s Z_s + \sigma_s X_s Y_s + \bar{\sigma}_{s-\delta} X_{s-\delta} Y_s) d{W}_s \\ = & - \int_{T-\delta}^T (\bar{\mu}_s X_s E^{\mathcal{F}_{s, T}^B}[\xi_{s+\delta}] +\bar{\sigma}_s X_s E^{\mathcal{F}_{s, T}^B}[\eta_{s+\delta}]) ds - \int_t^{T} \rho_s X_s ds. \end{align*} Take conditional expectation with respect to $\tilde{\tilde{\mathcal{G}}}_t:=\mathcal{F}_t^W \vee \mathcal{F}_{T}^B$ on both sides, then \begin{align*} X_t Y_t = & E^{\tilde{\tilde{\mathcal{G}}}_t} [X_T Y_T + \int_t^T \rho_s X_s ds] + E^{\tilde{\tilde{\mathcal{G}}}_t} [\int_{T-\delta}^T (\bar{\mu}_s X_s E^{\mathcal{F}_{s, T}^B}[\xi_{s+\delta}] +\bar{\sigma}_s X_s E^{\mathcal{F}_{s, T}^B}[\eta_{s+\delta}]) ds], \end{align*} which implies the desired result when noting that $X_t=1$ and $Y_t\in \mathcal{F}_{t, T}^B$. \end{proof} \section{Conclusion and future work} In this paper, we have established the existence/uniqueness theorem and the comparison theorem for the anticipated BDSDEs. Moreover, as an application, we studied a duality between the anticipated BDSDE and delayed DSDE, where the BDSDE is of a special form, thus the duality is somewhat limited. In fact for the general case, it should be admitted that $\mathcal{G}_t$, which is not a filtration, not increasing non decreasing, brings the main technical difficulty in working with the duality problem. For the future work, I will go on studying this topic and pay more attention to the applications of such equations. \section*{Acknowledgements} The author is grateful to the editor and anonymous referees for their helpful suggestions. This work is supported by the Mathematical Tianyuan Foundation of China (Grant No. 11126050), the Specialized Research Fund for the Doctoral Program of Higher Education of China (Grant No. 20113207120002), and partially supported by the National Natural Science Foundation of China (Grant No. 11101209).

https://arxiv.org/abs/2112.01312

The inverse source problem for the wave equation revisited: A new approach

The inverse problem of reconstructing a source term from boundary measurements, for the wave equation, is revisited. We propose a novel approach to recover the unknown source through measuring the wave fields after injecting small particles, enjoying a high contrast, into the medium. For this purpose, we first derive the asymptotic expansion of the wave field, based on the time-domain Lippmann-Schwinger equation. The dominant term in the asymptotic expansion is expressed as an infinite series in terms of the eigenvalues $\{\lambda_n\}_{n\in \mathbb{N}}$ of the Newtonian operator (for the pure Laplacian). Such expansions are useful under a certain scale between the size of the particles and their contrast. Second, we observe that the relevant eigenvalues appearing in the expansion have non-zero averaged eigenfunctions. We prove that the family $\{\sin(\frac{c_1}{\sqrt{\lambda_n}}\, t),\, \cos(\frac{c_1}{\sqrt{\lambda_n}}\, t)\}$, for those relevant eigenvalues, with $c_1$ as the contrast of the small particle, defines a Riesz basis (contrary to the family corresponding to the whole sequence of eigenvalues). Then, using the Riesz theory, we reconstruct the wave field, generated before injecting the particles, on the center of the particles. Finally, from these (internal values of these) last fields, we reconstruct the source term (by numerical differentiation for instance). A significant advantage of our approach is that we only need the measurements on $\{x\}\times(0, T)$ for a single point $x$ away from $\Omega$, i.e., the support of the source, and large enough $T$.

\section{Introduction}\label{sec_int} Our goal in this work is to propose a different approach to study the well known inverse source problems that might have some advantages as compared to the already known ways of tackling these problems. Let us first describe the mathematical formulation of our inverse source problems. Denote by $A_0(x)$ and $B_0(x)$ two scalar coefficients modeling the background medium. These coefficients can, respectively, model the electric permittivity $\epsilon_0$ and the magnetic permeability $\mu_0$ in the electromagnetism or the inverses of the bulk modulus $\kappa_0$ and the mass density $\rho_0$ in acoustics. Assume that they are positive constants, i.e., $A_0(x)\equiv A_0$ and $B_0(x)\equiv B_0$. Define $c_0:=(A_0B_0)^{-1/2}$. Let $J=J(x,\,t)$ be a source compactly supported in $\Omega$, where $\Omega$ is a bounded Lipschitz domain in $\mathbb R^n$. We also assume that $J=0$ in $\Omega$ for $t<0$. Then the field $V$ generated by the source $J$ in the homogeneous background medium satisfies \begin{equation}\label{model} \begin{cases} c_0^{-2} V_{tt} - \Delta V = J & \mathrm{in}\;\mathbb R^n\times(0,\,T),\\ V|_{t=0}=0,\;V_t|_{t=0}=0 & \mathrm{in}\;\mathbb R^n \end{cases} \end{equation} for $n=2,\, 3$. A typical inverse source problem is to construct the source term $J$ from the measurements of $V$ on a cylinder $\partial \Omega \times (0,\,T)$ for a large time $T$. Such inverse source problems have many significant applications and have been extensively studied; see, for instance, \cite{B-Y2006, C-D1983, D-T2015, J-L-Y2017, Ton2003} and the references therein. We also refer to the monograph \cite{Isakov1990} for uniqueness and stability of various inverse source problems. There are several classical arguments to show the uniqueness and stability for inverse source problems of hyperbolic equations. The approaches based on Carleman estimates and unique continuation have been widely used \cite{B-Y2017, Klibanov1992, Tataru1996}. Based on exact boundary controllability, the stability and a reconstruction formula for the inverse problem of determining space-dependent component in the source term from boundary measurement are shown in \cite{Yama1995}. Under a weak regularity assumption, a uniqueness result for a multidimensional hyperbolic inverse source problem with a single measurement was proved in \cite{Yama1999}. Recently, an approach of using Laplace transform was employed in \cite{Hu2020} to show the uniqueness of some inverse source problems for the wave equation. Here, we propose a different way of looking at this problem. Let $D=z+a B$ be a particle injected into $\Omega$, where $a$ is a small parameter, $B\subset\mathbb R^n$ is a bounded Lipschitz domain containing the origin and $z\in\mathbb R^n$. Define \begin{equation}\label{def_eps} A(x):= \begin{cases} A_0 & \mathrm{in}\; \mathbb R^n\setminus D,\\ A_1 & \mathrm{in}\; D, \end{cases} \end{equation} where $A_1$ is a positive constant with $A_1 \sim a^{-2}$ as $a\ll 1$. Such particles are known to exist in electromagnetism and they are called dielectric nanoparticles (with large values of the permittivity and moderate permeability). In acoustics, bubbles with such small values of the bulk, with moderate mass density, can be designed; see \cite{Z-F2018} for a discussion on this issue. Set $$c(x):=(A(x)B_0)^{-1/2}, \quad c_1:=(A_1B_0)^{-1/2}.$$ Define the contrast as \begin{equation}\label{def_q} q(x):=\frac{c_0^2}{c^2(x)}-1, \quad x\in \mathbb R^n. \end{equation} Then, after injecting the particle $D$ into $\Omega$, the field $U$ generated by $J$ meets \begin{equation}\label{model_U} \begin{cases} c^{-2}(x) U_{tt} - \Delta U = J & \mathrm{in}\;\mathbb R^n\times(0,\,T),\\ U|_{t=0}=0,\;U_t|_{t=0}=0 & \mathrm{in}\;\mathbb R^n. \end{cases} \end{equation} The solvability and stability results for the direct problems \eqref{model} and \eqref{model_U} can be found in \cite[Theorem 8.1]{Isakov2017}.The aim is to reconstruct the unknown source $J$ from the measurement of $U(x,\,t)$ for a fixed $x\in\mathbb R^n\setminus \Omega$ over a large enough time interval, that is, measuring the wave fields after injecting the small particle $D$. The results will be given and proved in the dimensions $n=3$. The corresponding results and proofs for the case $n=2$ can be done similarly at the expense of handling the different type of the singularity of the Green's function. We first analyze the asymptotic behavior of the wave field $U(x,\,t)$ for the wave equation model as $a\to 0$. Asymptotic expansions of physical fields generated by small particles are well developed in the literature for elliptic models, but there are few results on time-domain models. Recently, we studied a heat conduction problem and an acoustic scattering problem with many small holes in \cite{S-W2019} and \cite{S-W-Y2021}, respectively, where we derived the asymptotic analysis of the solutions by using time-domain integral equation methods. In this paper, based on the time-domain Lippmann-Schwinger equation, we shall show the asymptotic expansion of $U(x,\,t)$ for $x\in\mathbb R^3\setminus\overline D$ as $a\to 0$ through the eigensystem of the Newtonian potential operator related to the Laplace equation. We set $W:=U-V$. Then we have \begin{equation}\label{model_W} \begin{cases} c_0^{-2} W_{tt} - \Delta W = -\dfrac{q}{c_0^2} U_{tt} & \mathrm{in}\;\mathbb R^3\times(0,\,T),\\ W|_{t=0}=0,\;W_t|_{t=0}=0 & \mathrm{in}\;\mathbb R^3. \end{cases} \end{equation} The result is as follows. \begin{theorem}\label{Asymptot-Introduction} Under the conditions above, we have the expansion \begin{eqnarray*} W(x,\,t) &=&\sum_{n=1}^{+\infty}\frac{c_1\lambda_n^{-\frac{3}{2}}}{4\pi |x-z|}\,\left(\int_D e_n(y)\,dy\right)^2\,\int_{c_0^{-1}|x-z|}^{t} \sin\left[ \frac{c_1}{\sqrt{\lambda_n}}(t-s)\right]\,V(z,\,s-c_0^{-1}|x-z|)\,ds \nonumber\\ && \qquad\qquad\qquad\qquad\qquad\qquad + O(a^2), \quad (x,\,t) \in (\mathbb R^3\setminus\overline D)\times (0,\,T), \end{eqnarray*} where $\{\lambda_n, e_n\}_{n\in \mathbb{N}}$ is the family of the eigenelements of the Newtonian operator $N:\, L^2(D)\rightarrow L^2(D)$ defined by \begin{equation}\label{op_N} N(f)(x):=\int_{D}\frac{f(y)}{4 \pi\vert x-y\vert}d(y). \end{equation} \end{theorem} Observe that the relevant eigenfunctions $e_n$'s are those with non-vanishing averages. Note that for the Newtonian operator, $1$ is not an eigenvalue. Therefore, the family of eigenelements $\{\lambda_n,\, e_n\}$, with non-vanishing averages, is rich. In addition, in the case that $D$ is a ball, the behavior in terms of the index $n$, i.e., the Weyl law, of these relevant eigenvalues, that we still denote $\{\lambda_n\}_{n \in \mathbb{N}}$, allow us to show that the family $\{\sin(\frac{c_1}{\sqrt{\lambda_n}}\, t),\,\cos(\frac{c_1}{\sqrt{\lambda_n}}\, t)\}$, for those relevant eigenvalues, defines a Riesz basis (contrary to the family corresponding to the whole sequence of eigenvalues). We state this as a second result. \begin{theorem}\label{Riesz-Introduction} Let $D$ be a ball and the set $\{\lambda_n, e_n\}_{n\in \mathbb{N}}$ be the subfamily of the eigenelements of the Newtonian operator $N$ satisfying $\int_B e_n(x)dx\neq 0$. Then the family $\{\sin(\omega_n\, t), \, \cos(\omega_n\, t)\}_{n\in\mathbb N}$ defines a Riesz basis in $L^2[-\frac{\pi}{b},\, \frac{\pi}{b}]$, where $b=\frac{c_1\pi}{a}$ and $\omega_n:=\frac{c_1}{\sqrt{\lambda_n}},\; n \in \mathbb{N}$. \end{theorem} We do believe that this result can be extended to smooth general shapes. \bigskip The dominant term in the expansion is expressed as an infinite series in terms of Riesz basis. Using the Riesz theory, we can approximately obtain the Riesz coefficients of the wave field $V$ at $z$ over a time interval. This is the object of the final result in this introduction. \begin{theorem}\label{Reconstruction-Introduction} Assume that the source $J$ is compactly supported in $\Omega\times[0,\,T]$. Let $x\in \mathbb R^3\setminus \Omega$ and $z \in \Omega$, i.e., the center of the particle, be fixed. We measure both $V(x,\, t)$ and $U(x,\,t)$ for an interval of the form $[\tilde{T},\, \tilde{T}+\frac{2\pi}{b}]$ with $\tilde{T}>T_J+\frac{\vert x-z \vert}{c_0}$ and $T_J:=T+\max_{y\in \Omega}\frac{\vert z-y \vert}{c_0}$. Then \begin{equation}\label{Riesz_h7} V(z,\,t) = \sum_{n=1}^{+\infty}\left[C_n(x,\,z)g_n(t-\pi/b) + D_n(x,\,z) h_n(t-\pi/b)\right], \quad t\in[0,\,2\pi/b] \end{equation} with \begin{eqnarray} C_n(x,\,z)&:=& \int_{0}^{2\pi/b} \cos\left(\omega_n (s-\pi/b)\right) \,V(z,\,s)\,ds,\quad n\in\mathbb N, \label{Fourier-modes-1}\\ D_n(x,\,z)&:=& \int_{0}^{2\pi/b} \sin\left(\omega_n (s-\pi/b)\right) \,V(z,\,s)\,ds,\quad n\in\mathbb N, \label{Fourier-modes-2} \end{eqnarray} where $g_n(t):=(\mathcal T^{-1})^*\cos((n-1) t)$ and $h_n(t):=(\mathcal T^{-1})^*\sin(n t)$. The operator $\mathcal T$ is the one linking the Riesz basis $\{\sin(\omega_n\, t), \,\cos(\omega_n\, t)\}_{n\in\mathbb N}$ to the canonical basis $\{\sin(n\, t), \,\cos((n-1)\, t)\}_{n\in\mathbb N}$; see (\ref{defn_T}). The coefficients $C_n(x, \,z)$ and $D_n(x, \,z)$ can be estimated from the data $W(x,\, t)$ as follows: \begin{eqnarray*} C_n(x,\,z) &=& \cos(\omega_n\tilde T) A_n(x,\,z) - \sin(\omega_n\tilde T) B_n(x,\,z),\\ D_n(x,\,z) &=& \sin(\omega_n\tilde T) A_n(x,\,z) + \cos(\omega_n\tilde T) B_n(x,\,z), \end{eqnarray*} where \begin{eqnarray} A_n(x,\,z)&=& \alpha^{-1}_n \int_{0}^{2\pi/b} h_n(t-\pi/b) \, U(x,\,\tilde T+s)\,ds +O(a),\quad n\in\mathbb N, \label{Fourier-modes-1-computed}\\ B_n(x,\,z)&=& \alpha^{-1}_n \int_{0}^{2\pi/b} g_n(t-\pi/b) \, U(x,\,\tilde T+s)\,ds +O(a),\quad n\in\mathbb N \label{Fourier-modes-2-computed} \end{eqnarray} with \begin{equation}\label{alpha-1} \alpha_n:=\alpha_n(x,\,z):= \frac{\omega_n \left(\int_D e_n(y)\,dy\right)^2}{4\pi |x-z| \lambda_n}. \end{equation} \end{theorem} Varying the location of $z$ in $\Omega$, we get the wave field $V$ in $\Omega$ over a time interval. Then, we can further reconstruct the source term $J$ by numerically differentiating $V$. A significant advantage of the proposed method is that we only need to perform the measurement of $U$ at a fixed point $x\in\mathbb R^3\setminus\Omega$, which is attractive in practical situation. As we reconstruct $V(z,\, t)$ for $ t\in[0,\,2\pi/b]$, then if $2\pi/b > T_J$, we can reconstruct $V$ for any time. This condition on $b=c_1\pi/a$ makes sense since it is related to the used particle which is at our handle. This approach is inspired by the imaging modalities using contrast agents; see \cite{B-B-I2011, S-K-V-H2010, McLeod-Ozcan, Ntziachristos} and the references therein for real world applications. The common motivation there is that the imaging techniques are potentially able to extract features when the physical parameters between healthy tissues and malignant ones have relatively high contrast. However, it is observed that for a benign tissue, the variation of the permittivity is quite low so that conventional imaging modalities are limited to be employed for early detection of diseases. In such cases, it is highly desirable to create such missing contrast. One way to do is to use electromagnetic nanoparticles as contrast agents. The idea can also be used to acoustic imaging modalities using bubbles as contrast agents. The mathematical analysis of such imaging modalities are given in \cite{C-C-S2018, D-G-S2021, G-S2022, G-S2022-2}. The main idea there is that contrasting the fields measured before and after injecting the contrasting particles, one can reconstruct the total fields on the location of those particles. In \cite{D-G-S2021, G-S2022, G-S2022-2}, this is justified in the time harmonic regimes (both for the electromagnetic and acoustic models) when the used incident waves are sent at frequencies close to the resonating frequencies of the used particles. These resonances exist for certain scales of the contrasts of the particles. In our work here, we extend such ideas to the time domain inverse source problem. In this time-domain regime, the same scales of the contrasts allow us to extract the Fourier modes, in time, of the internal field using the eigenvalues of the Newtonian operator; see (\ref{Fourier-modes-1})--(\ref{Fourier-modes-2}). Observe that for these scales the resonances, called the dielectric resonances, used in \cite{D-G-S2021, G-S2022} are related, and actually due, to the eigenvalues of the Newtonian operator as well. We finish this introduction by the following observation. The expansion in Theorem \ref{Asymptot-Introduction} means that the field $U(x, t)$ is composed of the field $V(x, t)$, which we call the first field, generated by solely the source $f$ and the one generated by the interaction of this field and particle $D$ which we call the second field. This second field, which is a local spot, is modeled by (\ref{L-S_3}); see also (\ref{L-S_6}). As mentioned in Theorem \ref{Reconstruction-Introduction}, while measuring $U(x, t)$ for large $t$, the first wave field $V(x, t)$ has already fully passed the point $x$ where we measure. Therefore, we only measure the second field due to the interaction between the source $f$ and the small particle. The remainder of the paper is organized as follows. In Section \ref{sec_asym}, we derive the time-domain Lippmann-Schwinger equation for $W$ and then prove Theorem \ref{Asymptot-Introduction}. In Section \ref{sec-Riesz}, we prove Theorem \ref{Riesz-Introduction} and finally in Section \ref{sec-Reconstruction}, we prove Theorem \ref{Reconstruction-Introduction}. \section{Proof of Theorem \ref{Asymptot-Introduction}}\label{sec_asym} In this section, we show the asymptotic behavior of the field $W(x,\,t)$ for $(x,\,t)\in (\mathbb R^3\setminus\overline D)_T$ as $a\ll 1$, and completes the proof of Theorem \ref{Asymptot-Introduction}. To proceed, we introduce the function space \begin{equation*} H_0^r(0,\,T):=\left\{ g|_{(0,\,T)}:\; g\in H^r(\mathbb R)\, \textrm{ with }\, g\equiv 0 \, \textrm{ in }\, (-\infty,\,0) \right\},\quad r\in\mathbb R, \end{equation*} and generalize it to the $E$-valued function space, denoted by $H_0^r(0,\,T;\, E)$, where $E$ is a Hilbert space. Define \begin{equation*} LT(\sigma,\,E):= \left\{f\in \mathcal D^\prime_+(E):\,e^{-\sigma t}f \in \mathcal S^\prime_+(E) \right\}, \quad \sigma>0, \end{equation*} where $\mathcal D^\prime_+(E)$ and $\mathcal S^\prime_+(E)$ denote the sets of distributions and temperate distributions on $\mathbb R$, respectively, with values in $E$ and supports in $[0,\,+\infty)$. Then we define the space \begin{equation*} H_{0,\sigma}^r(0,\,T;\, E):= \left\{f\in LT(\sigma,\,E) : \, e^{-\sigma t}\partial_t^r f\in L^2_0(0,\,T;\,E) \right\}, \quad r\in\mathbb Z_+. \end{equation*} For nonnegative integer $r$, we introduce the norm \begin{equation}\label{norm} \| f \|_{H_{0,\sigma}^r(0,\,T;\, E)}:=\left(\int_0^T e^{-2\sigma t}\left[\left\| f \right\|^2_E + \sum_{k=1}^r T^{2k} \left\| \frac{\partial^k f}{\partial t^k} \right\|^2_E\right] \,dt\right)^{1/2}. \end{equation} For simplicity of notations, we denote $X\times (0,\,T)$ by $X_T$, where $X$ is a domain in $\mathbb R^3$. We shall also use the notation \lq\lq$\lesssim$\rq\rq{} to denote \lq\lq$\leq$\rq\rq{} with its right-hand side multiplied by a generic positive constant, if we do not emphasize the dependence of the constant on some parameters. For convenience, we set \begin{equation*} q_0=\frac{c_0^2}{c_1^2}-1 \sim a^{-2} \quad \textrm{for}\;a\ll 1. \end{equation*} Define the retarded volume potential $V_D$ by \begin{equation}\label{op_V_D} V_D[f](x,\,t):=\int_D \frac{f(y,\,t-c_0^{-1}|x-y|)}{4\pi|x-y|}\,dy, \quad (x,\,t)\in \mathbb R^3\times(0,\,T). \end{equation} From \eqref{model_W}, we are led to the Lippmann-Schwinger equation \begin{equation}\label{L-S} U(x,\,t) + \frac{q_0}{c_0^2} V_D[U_{tt}](x,\,t)= V(x,\,t), \quad (x,\,t)\in \mathbb R^3\times(0,\,T) \end{equation} or \begin{equation}\label{L-S_0} W(x,\,t) + \frac{q_0}{c_0^2}V_D[W_{tt}](x,\,t) = -\frac{q_0}{c_0^2}V_D[V_{tt}](x,\,t), \quad (x,\,t)\in \mathbb R^3\times(0,\,T). \end{equation} Following the convolution quadrature based argument in \cite{Lubich1994}, we can prove that the equation \eqref{L-S_0} has a unique solution in $H_{0,\sigma}^r(0,\,T;\,L^2(D))$ for $V\in H_{0,\sigma}^{r+2}(0,\,T;\,L^2(D))$. \bigskip In the following, we derive the asymptotic behavior of the field $W(x,\,t)$. We start by scaling both the space and time variables. Set $$T_a := T/a.$$ For any functions $\varphi$ and $\psi$ defined on $D_T$ and $B_{T_a }$, respectively, we use the notations \begin{eqnarray*} &&\hat\varphi(\xi,\,\tau)=\varphi^\wedge(\xi,\,\tau):=\varphi(a \xi+z,\,a \tau), \quad (\xi,\,\tau)\in B_{T_a },\\ &&\check\psi(x,\,t)=\psi^\vee(x,\,t):=\psi\left(\frac{x-z}{a },\,\frac{t}{a }\right), \quad (x,\,t)\in D_T. \end{eqnarray*} Notice that \begin{equation}\label{scaling_t} \frac{\partial^n \hat\varphi(\cdot,\,\tau)}{\partial \tau^n}=\frac{\partial^n \varphi(\cdot,\,a \tau)}{\partial \tau^n} = a ^n \frac{\partial^n \varphi(\cdot,\,t)}{\partial t^n}, \quad n \in\mathbb Z_+. \end{equation} Then we have the following scaling result. \begin{lemma}\label{scalingt} Suppose $0< a \leq 1$. For $\psi \in H^r_{0,\sigma}(0,\,T;\, L^2(D))$ with nonnegative integer $r$, we have \begin{equation}\label{scalingt_11} \|\psi\|_{H^r_{0,\sigma}(0,\,T;\, L^2(D))} = a^2 \|\hat \psi\|_{H^r_{0,a \sigma}(0,\,T_a ;\, L^2(B))}. \end{equation} \end{lemma} {\bf Proof.} From \eqref{norm} and \eqref{scaling_t}, we can easily derive that \begin{eqnarray*} \|\psi\|^2_{H^r_{0,\sigma}(0,\,T;\,L^2(D))} &=& \int_0^T e^{-2\sigma t}\, \sum_{k=0}^r T^{2k} \,\left\|\frac{\partial^k \psi(\cdot,\,t)}{\partial t^k}\right\|^2_{L^2(D)}\,dt\\ & = & a^4 \int_0^{T_a } e^{-2\sigma a \tau}\, \sum_{k=0}^r T_a^{2k} \,\left\|\frac{\partial^k \hat\psi(\cdot,\,\tau)}{\partial \tau^k}\right\|^2_{L^2(B)}\, d\tau \\ & = & a ^4 \, \|\hat \psi\|^2_{H^r_{0,a \sigma}(0,\,T_a ;\,L^2(B))}. \end{eqnarray*} The proof is complete. \hfill $\Box$ \bigskip Let \begin{equation}\label{op_S_D} \mathcal S_D:=-(I + (q_0/c_0^2)\,V_D\,\partial_t^2)^{-1}\, (q_0/c_0^2)\,V_D\,\partial_t^2. \end{equation} Define the retarded volume potential $V_B^a$ by \begin{equation}\label{op_V_B} V_B^a[\psi](\xi,\,\tau):=\int_B \frac{\psi(\eta,\,\tau-c_0^{-1}|\xi-\eta|)}{4\pi|\xi-\eta|}\,d\eta, \quad (\xi,\,\tau)\in \mathbb R^3\times(0,\,T_a), \end{equation} and then introduce the operator $\mathcal S_B^a$ by \begin{equation}\label{op_S_B} \mathcal S_B^a:=-(I + (q_0/c_0^2)\,V_B^a\,\partial_\tau^2)^{-1}\, (q_0/c_0^2)\,V_B^a\,\partial_\tau^2. \end{equation} Then we can easily prove the following result. \begin{lemma}\label{scalingV} Let $\psi\in H^{r+2}_{0,\sigma}(0,\,T;\,L^2(D))$. Then \begin{eqnarray} V_D[\psi_{tt}] &=& (V^a_B[\hat\psi_{\tau\tau}])^\vee, \label{scalingV_1}\\ \|\mathcal S_D\|_{\mathcal L\left(H^{r+2}_{0, \sigma}(0,\,T;\,L^2(D)),\, H^r_{0, \sigma}(0,\,T;\,L^2(D))\right)} &=& \|\mathcal S_B^a\|_{\mathcal L\left(H^{r+2}_{0,a \sigma}(0,\,T_a ;\,L^2(B)),\, H^r_{0,a \sigma}(0,\,T_a ;\,L^2(B))\right)}. \label{scalingS_1} \end{eqnarray} \end{lemma} {\bf Proof.} Let $x=a \xi + z$, $y=a \eta + z$ and $t=a \tau$. Then we have \begin{eqnarray*} V_D[\psi_{tt}](x,\,t) &=& \int_D \frac{\psi_{tt}(y,\,t-c_0^{-1}|x-y|)}{4\pi |x-y|}\,dy\\ &=& \int_B \frac{\psi_{tt}(a \eta+z,\,a \tau-c_0^{-1}a|\xi-\eta|)}{4\pi a |\xi-\eta|}\,a^3\,d\eta\\ &=& \int_B \frac{a^{-2}\hat\psi_{\tau\tau}(\eta,\, \tau-c_0^{-1}|\xi-\eta|)}{4\pi a |\xi-\eta|}\,a^3\,d\eta\\ &=& V^a_B[\hat \psi_{\tau\tau}](\xi,\,\tau), \end{eqnarray*} which completes the proof of \eqref{scalingV_1}. Further, we have $(\mathcal S_D[\psi])^\wedge=\mathcal S_B^a[\hat \psi]$. Then, it can be derived that \begin{eqnarray*} & &\left \| \mathcal S_D \right\|_{\mathcal L \left(H^{r+2}_{0,\sigma}(0,\,T;\,L^2(D)),\,H^r_{0,\sigma}(0,\,T;\,L^2(D))\right)}\\ &:=& \sup\limits_{0\not=\psi\in H^{r+2}_{0,\sigma}(0,\,T;\,L^2(D))} \displaystyle \frac{ \| \mathcal S_D[\psi] \|_{H^r_{0,\sigma}(0,\,T;\,L^2(D))}}{\|\psi\|_{H^{r+2}_{0,\sigma}(0,\,T;\,L^2(D))}}\\ &=& \sup\limits_{0\not=\psi\in H^{r+2}_{0,\sigma}(0,\,T;\,L^2(D))} \displaystyle \frac{ a^2\, \| \left(\mathcal S_D[\psi]\right)^\wedge \|_{H^r_{0,a\sigma}(0,\,T_a;\,L^2(B))}}{a^2 \|\hat\psi\|_{H^{r+2}_{0,a\sigma}(0,\,T_a;\,L^2(B))}}\\ &=& \sup\limits_{0\not=\hat\psi\in H^{r+2}_{0,a\sigma}(0,\,T_a;\,L^2(B))} \displaystyle \frac{ \| \mathcal S^a_B[\hat\psi] \|_{H^r_{0,a\sigma}(0,\,T_a;\,L^2(B))}}{\|\hat\psi\|_{H^{r+2}_{0,a\sigma}(0,\,T_a;\,L^2(B))}}\\ &=& \left \| \mathcal S^a_B \right\|_{\mathcal L \left(H^{r+2}_{0,a\sigma}(0,\,T_a;\,L^2(B)),\,H^r_{0,a\sigma}(0,\,T_a;\,L^2(B))\right)}. \end{eqnarray*} The proof is complete. \hfill $\Box$ \medskip Denote by $\hat V_B$ the volume potential operator for the Helmholtz equation $$-\Delta U + s^2\,U=\hat f.$$ Define $A(s):\,L^2(B)\to L^2(B)$ by \begin{equation}\label{def_A} A(s):=-(I + (q_0s^2/c_0^2)\,\hat V_B)^{-1}\, (q_0s^2/c_0^2)\,\hat V_B, \end{equation} which is the Fourier-Laplace transform of $\mathcal S_B^a$ defined by \eqref{op_S_B}. Then we have the following estimate in \cite[Theorem 4.1]{L-M2015}: \begin{equation}\label{esi_A} \left\| A(s) \right\|_{\mathcal L(L^2(B),\,L^2(B))} \leq \frac{q_0\,|s|^2}{\sigma^2\,c_0^2} \end{equation} for any $s$ with $\mathrm{Re}\, s = \sigma>0$. \begin{lemma}\label{inv_SB} The operator $\mathcal S_B^a :\, H^{r+2}_{0,a \sigma}(0,\,T_a ;\,L^2(B)) \to H^r_{0,a \sigma}(0,\,T_a ;\,L^2(B))$ with $r=0,\,1$ is estimated by $O(q_0)=O(a^{-2})$ for $a \ll 1$. \end{lemma} {\bf Proof.} Note that $\mathcal S_B^a$ is the corresponding time-domain solution operator obtained by the inverse Fourier-Laplace transform of $A(s)$. Using Lubich's notation, we have $\mathcal S_B^a = A(\partial_t)$. Then, for $r=0$ and $g\in H^2_{0,a \sigma}(0,\,T_a ;\,L^2(B))$ compactly supported with respect to $t$ in $[0,\,T_a ]$, we have \begin{eqnarray*}\label{SB_2} \left\| \mathcal S^a_B[g]\right\|^2_{H^0_{0,a \sigma}(0,\,T_a ;\,L^2(B))} &=&\int_0^{T_a } \left(e^{-a \sigma t} \,\|A(\partial_t)g \|_{L^2(B)}\right)^2\,dt\nonumber \\ & \leq & \int_0^{+\infty}\left(\mathcal F \left[e^{-a \sigma t} \|A(\partial_t)g \|_{L^2(B)}\right](\eta)\right)^2\,d\eta \nonumber\\ & = & \int_{a \sigma-i\mathbb R_+}\left(\mathcal L \left[ \|A(\partial_t)g \|_{L^2(B)}\right](s)\right)^2\,ds \nonumber\\ & = & \int_{a \sigma-i\mathbb R_+}\|\mathcal L \left[ A(\partial_t)g\right](s) \|^2_{L^2(B)}\,ds \nonumber\\ & = & \int_{a \sigma-i\mathbb R_+}\| A(s)\,\mathcal L[g](s) \|^2_{L^2(B)}\,ds \nonumber\\ & \leq & \frac{q_0^2}{c_0^4\,(a \sigma)^4}\,\int_{a \sigma-i\mathbb R_+}\| s^2 \mathcal L[g](s) \|^2_{L^2(B)}\,ds \nonumber\\ & = & \frac{q_0^2}{c_0^4\,(a \sigma)^4}\,\int_{a \sigma-i\mathbb R_+} \| \mathcal L[\partial_t^2 g](s) \|^2_{L^2(B)}\,ds \nonumber\\ & = & \frac{q_0^2\,T_a^{-4}}{c_0^4\,(a \sigma)^4}\,\int_0^{+\infty} \left( e^{-a \sigma t}\,T_a^2\,\| \partial_t^2 g \|_{L^2(B)}\right)^2\,dt \nonumber\\ & \leq & \frac{q_0^2\,T^{-4}}{c_0^4\,\sigma^4}\,\left\| g\right\|^2_{H^2_{0,a \sigma}(0,\,T_a ;\,L^2(B))} . \end{eqnarray*} Similarly, we derive for $r=1$ that \begin{eqnarray*}\label{SB_3} \left\| \mathcal S^a_B[g]\right\|^2_{H^1_{0,a \sigma}(0,\,T_a ;\,L^2(B))}&=& \int_0^{T_a }e^{-2a \sigma t} \left( \| A(\partial_t)g \|^2_{L^2(B)} + T_a^2\, \left\|\frac{\partial A(\partial_t)g}{\partial t}\right\|^2_{L^2(B)}\right)\,dt \nonumber\\ &\leq & 2T_a ^2\, \int_0^{T_a }e^{-2a \sigma t}\, \|\partial_t\left( A(\partial_t)g\right)\|^2_{L^2(B)} \,dt\nonumber\\ & \leq & 2T_a ^2\, \int_0^{+\infty} \left(\mathcal F \left[e^{-a \sigma t}\, \| \partial_t\left(A(\partial_t)g\right)\|_{L^2(B)} \right](\eta)\right)^2 \,d\eta \nonumber\\ & \leq & \frac{2T_a ^2\,q_0^2}{c_0^4\,(a \sigma)^4}\,\int_{a \sigma-i\mathbb R_+} \|s^3 \mathcal L[g](s) \|^2_{L^2(B)}\,ds \nonumber\\ & = & \frac{2T_a ^2\,q_0^2}{c_0^4\,(a \sigma)^4}\,\int_{a \sigma-i\mathbb R_+}\| \mathcal L[\partial_t^3 g](s) \|^2_{L^2(B)}\,ds \nonumber\\ & = & \frac{2T_a^{-4}\,q_0^2}{c_0^4\,(a \sigma)^4}\,\int_0^{+\infty} e^{-2a \sigma t}\, T_a^6\,\| \partial_t^3 g \|^2_{L^2(B)}\,dt\nonumber \\ &\leq & \frac{2T^{-4}\,q_0^2}{c_0^4\,\sigma^4}\,\left\|g\right\|^2_{H^3_{0,a \sigma}(0,\,T_a ;\,L^2(B))}. \end{eqnarray*} The proof is complete. \hfill $\Box$ \bigskip Now we estimate $W|_{D_T}$ by \begin{eqnarray}\label{est_us_D} \|W(\cdot,\,t)\|_{L^2(D)} &\lesssim& \|W\|_{H^1_{0,a \sigma}(0,\,T;\,L^2(D))} \nonumber\\ &\lesssim& \|\mathcal S_D\|_{\mathcal L\left(H^3_{0, \sigma}(0,\,T;\,L^2(D)),\, H^1_{0, \sigma}(0,\,T;\,L^2(D))\right)}\, \|V\|_{H^3_{0, \sigma}(0,\,T;\,L^2(D))} \nonumber\\ &\lesssim& a^{3/2}\, \|\mathcal S_B\|_{\mathcal L\left(H^3_{0,a \sigma}(0,\,T_a ;\,L^2(B)),\, H^1_{0,a \sigma}(0,\,T_a ;\,L^2(B))\right)} \nonumber\\ &\lesssim& a^{-1/2}, \end{eqnarray} and hence \begin{equation}\label{est_u_D} \|U(\cdot,\,t)\|_{L^2(D)} \lesssim a^{-1/2} \end{equation} for any $t\in[0,\,T]$. In the same way, we can also estimate $\partial_t^k U|_{D_T}$ by \begin{equation}\label{est_u_D_t} \|\partial_t^k U(\cdot,\,t)\|_{L^2(D)} \lesssim a^{-1/2}, \quad t\in[0,\,T],\;k=1,\,2,\,\cdots. \end{equation} We rewrite \eqref{L-S} as \begin{eqnarray}\label{L-S_2} U(x,\,t) + \frac{q_0}{c_0^2}\int_D \frac{U_{tt}(y,\,t)}{4\pi|x-y|}\,dy &=& V(x,\,t) + \frac{q_0}{c_0^2}\int_D \frac{U_{tt}(y,\,t)-U_{tt}(y,\,t-c_0^{-1}|x-y|)}{4\pi|x-y|}\,dy \nonumber\\ &=:& V(x,\,t) + O(q_0a), \quad (x,\,t)\in \mathbb R^3\times(0,\,T). \end{eqnarray} Consider the following problem: \begin{equation}\label{L-S_3} \begin{cases} \varphi(x,\,t) + \displaystyle\frac{q_0}{c_0^2}\int_D \frac{\varphi_{tt}(y,\,t)}{4\pi|x-y|}\,dy = f(x,\,t) & \mathrm{in}\;D\times(0,\,T),\\ \varphi(x,\,0)=\varphi_t(x,\,0)=0 & \mathrm{in}\;D. \end{cases} \end{equation} Let $\{\lambda_n,\,e_n\}$ be the eigensystem of the Newtonian potential operator $N$ defined by \begin{equation*} N[\varphi](x,\,t):=\int_D \frac{\varphi(y,\,t)}{4\pi |x-y|}\,dy, \quad (x,\,t)\in D_T. \end{equation*} Set \begin{equation*} \varphi_n(t):=\langle \varphi(\cdot,\,t),\, e_n\rangle, \quad f_n(t):=\langle f(\cdot,\,t),\, e_n\rangle. \end{equation*} We have \begin{equation*} \varphi(x,\,t) = \sum_{n=1}^{+\infty} \langle \varphi(\cdot,\,t),\, e_n\rangle\, e_n = \sum_{n=1}^{+\infty} \varphi_n(t)\, e_n. \end{equation*} Then it is derived from \eqref{L-S_3} that \begin{equation}\label{L-S_4} \begin{cases} \frac{q_0}{c_0^2}\lambda_n \varphi_n^{\prime\prime}(t) + \varphi_n(t) = f_n(t) \quad \mathrm{in}\; (0,\,T),\\ \varphi_n(0)=\varphi_n^\prime(0)=0 \end{cases} \end{equation} for $n=1,\,2,\,\cdots$. The solution to \eqref{L-S_4} is given by \begin{equation}\label{L-S_5} \varphi_n(t)=c_0 (q_0\lambda_n)^{-1/2} \int_0^t \sin\left[c_0(q_0\lambda_n)^{-1/2}(t-\tau)\right]\,f_n(\tau)\,d\tau. \end{equation} Moreover, we have the estimate \begin{equation}\label{est_phi} \|\varphi(\cdot,\,t)\|_{L^2(D)}\lesssim \|f(\cdot,\,t)\|_{L^2(D)}. \end{equation} Then we get from \eqref{L-S_2} that \begin{equation}\label{est_u_D2} \|U(\cdot,\,t)\|_{L^2(D)} \lesssim a^{1/2} \end{equation} for any $t\in[0,\,T]$. In the same way, we can also improve the estimate of \eqref{est_u_D_t} as \begin{equation}\label{est_u_D_t2} \|\partial_t^k U(\cdot,\,t)\|_{L^2(D)} \lesssim a^{1/2}, \quad t\in[0,\,T],\;k=1,\,2,\,\cdots. \end{equation} Then the error term in \eqref{L-S_2} can be improved as \begin{eqnarray}\label{L-S_22} U(x,\,t) + \frac{q_0}{c_0^2}\int_D \frac{U_{tt}(y,\,t)}{4\pi|x-y|}\,dy &=& V(x,\,t) + \frac{q_0}{c_0^2}\int_D \frac{U_{tt}(y,\,t)-U_{tt}(y,\,t-c_0^{-1}|x-y|)}{4\pi|x-y|}\,dy \nonumber\\ &=:& V(x,\,t) + O(q_0a^2) , \quad (x,\,t)\in \mathbb R^3\times(0,\,T). \end{eqnarray} Using the estimate \eqref{est_phi} for \eqref{L-S_22}, we get \begin{equation}\label{est_u_D3} \|U(\cdot,\,t)\|_{L^2(D)} \lesssim a^{3/2}. \end{equation} Similarly, we can have \begin{equation}\label{est_u_D_t3} \|\partial_t^k U(\cdot,\,t)\|_{L^2(D)} \lesssim a^{3/2}, \quad t\in[0,\,T],\;k=1,\,2,\,\cdots. \end{equation} Let $\tilde U$ be the solution to \eqref{L-S_3} with $f(x,\,t)=V(z,\,t)$. Then we have \begin{equation*} f_n(t)=V(z,\,t)\int_D e_n(x)\,dx, \end{equation*} and hence \begin{equation}\label{L-S_6} \tilde U(x,\,t)=\sum_{n=1}^{+\infty}c_0 (q_0\lambda_n)^{-1/2}\, e_n(x)\, \int_0^t \sin\left[c_0(q_0\lambda_n)^{-1/2}(t-\tau)\right]\,V(z,\,\tau)\,d\tau\,\int_D e_n(y)\,dy \end{equation} for $(x,\,t)\in D\times (0,\,T)$. Thus, we have \begin{eqnarray}\label{L-S_7} U(x,\,t)&=&\tilde U(x,\,t) + O(a^2) \nonumber\\ &=&\sum_{n=1}^{+\infty}c_0 (q_0\lambda_n)^{-1/2}\, e_n(x)\, \int_0^t \sin\left[c_0(q_0\lambda_n)^{-1/2}(t-\tau)\right]\,V(z,\,\tau)\,d\tau\,\int_D e_n(y)\,dy \nonumber\\ & & \qquad + O(a^2). \end{eqnarray} For $x$ away from $D$ and $y\in D$, since \begin{equation*} \left|\frac{1}{|x-z|} - \frac{1}{|x-y|}\right|=O(a) \end{equation*} and \begin{equation*} \left|U_{tt}(y,\,t-c_0^{-1}|x-z|) - U_{tt}(y,\,t-c_0^{-1}|x-y|)\right|=\left|\partial_t^3 U(y,\,t^*) \right|\, |y-z|, \end{equation*} we have \begin{eqnarray}\label{L-S_1} W(x,\,t)&=& -\frac{q_0}{4\pi c_0^2 |x-z|}\int_D U_{tt}(y,\,t-c_0^{-1}|x-z|)\,dy \nonumber\\ & & + \frac{q_0}{4\pi c_0^2}\int_D \left( \frac{1}{|x-z|} - \frac{1}{|x-y|} \right)\,U_{tt}(y,\,t-c_0^{-1}|x-z|)\,dy \nonumber\\ & & + \frac{q_0}{c_0^2}\int_D \frac{U_{tt}(y,\,t-c_0^{-1}|x-z|) - U_{tt}(y,\,t-c_0^{-1}|x-y|)}{4\pi|x-y|} \,dy \nonumber\\ &=& -\frac{q_0}{4\pi c_0^2 |x-z|}\int_D U_{tt}(y,\,t-c_0^{-1}|x-z|)\,dy + O(q_0a)\int_D U_{tt}(y,\,t-c_0^{-1}|x-z|)\,dy\nonumber\\ &&+ O(q_0a)\int_D \left|\partial_t^3 U(y,\,t^*) \right|\,dy \nonumber\\ &=& -\frac{q_0}{4\pi c_0^2 |x-z|}\int_D U_{tt}(y,\,t-c_0^{-1}|x-z|)\,dy + O(a^2). \end{eqnarray} Inserting \eqref{L-S_7} into \eqref{L-S_1}, we obtain \begin{eqnarray}\label{L-S_8} &&W(x,\,t) \nonumber\\ &=&\sum_{n=1}^{+\infty}\frac{c_0 q_0^{-\frac{1}{2}}\lambda_n^{-\frac{3}{2}}}{4\pi |x-z|}\,\left(\int_D e_n(y)\,dy\right)^2\,\int_0^{t-c_0^{-1}|x-z|} \sin\left[c_0(q_0\lambda_n)^{-\frac{1}{2}}(t-c_0^{-1}|x-z|-\tau)\right]\,V(z,\,\tau)\,d\tau \nonumber\\ && \qquad\qquad\qquad\qquad\qquad\qquad + O(a^2) \nonumber\\ &=&\sum_{n=1}^{+\infty}\frac{c_0 q_0^{-\frac{1}{2}}\lambda_n^{-\frac{3}{2}}}{4\pi |x-z|}\,\left(\int_D e_n(y)\,dy\right)^2\,\int_{c_0^{-1}|x-z|}^{t} \sin\left[c_0(q_0\lambda_n)^{-\frac{1}{2}}(t-s)\right]\,V(z,\,s-c_0^{-1}|x-z|)\,ds \nonumber\\ && \qquad\qquad\qquad\qquad\qquad\qquad + O(a^2)\nonumber\\ &=&\sum_{n=1}^{+\infty}\frac{c_1\lambda_n^{-\frac{3}{2}}}{4\pi |x-z|}\,\left(\int_D e_n(y)\,dy\right)^2\,\int_{c_0^{-1}|x-z|}^{t} \sin\left[ \frac{c_1}{\sqrt{\lambda_n}}(t-s)\right]\,V(z,\,s-c_0^{-1}|x-z|)\,ds \nonumber\\ && \qquad\qquad\qquad\qquad\qquad\qquad + O(a^2), \quad (x,\,t) \in (\mathbb R^3\setminus\overline D)\times (0,\,T), \end{eqnarray} Thus, Theorem \ref{Asymptot-Introduction} is proved. \section{Proof of Theorem \ref{Riesz-Introduction}}\label{sec-Riesz} Based on the expansion \eqref{L-S_8}, we show how to reconstruct $V(z,\,t)$ from the measurement of $W(x,\,t)$ which is obtained by measuring the wave fields before and after injecting the particle $D$. Let $\{{\bar{\lambda}}_n\}_{n \in \mathbb{N}}$ be the sequence of eigenvalues of the Newtonian operator stated on the scaled particle $B$. We start with the following comparison result shown in \cite{Suragan-2012} \begin{equation} {\bar{\lambda}}^N_n \leq {\bar{\lambda}}^{-1}_n \leq {\bar{\lambda}}^D_n,\; ~~ n\in \mathbb{N}, \label{Comparison} \end{equation} where $\{{\bar{\lambda}}^N_n\}_{n \in \mathbb{N}}$ and $\{{\bar{\lambda}}^D\}_{n \in \mathbb{N}}$ are the sequences of eigenvalues of the Laplacian stated on $B$ corresponding to Neumann and Dirichlet boundary conditions, respectively. The following Weyl's expansions, in terms of the order $n$, of the eigenvalues are well known \begin{equation} {\bar{\lambda}}^N_n= \frac{6\pi^2}{\textrm{Vol}(B)}\; n^{2/3} +o(1) \; \mbox{ and } \; {\bar{\lambda}}^D_n= \frac{6\pi^2}{\textrm{Vol}(B)}\; n^{2/3} +o(1) \quad \mbox{ as } n\gg 1. \end{equation} Using (\ref{Comparison}), we deduce the similar properties for $\{{\bar{\lambda}}_n\}_{n \in \mathbb{N}}$ \begin{equation} {\bar{\lambda}}^{-1}_n= \frac{6\pi^2}{\textrm{Vol}(B)}\; n^{2/3} +o(1) \quad \mbox{ as } n\gg 1. \end{equation} Hence, due to the scale $\lambda_n=\bar{\lambda}_n a^2$ and $c_1\sim a$, we deduce that \begin{equation*} \omega_n:= c_1 \lambda^{-\frac{1}{2}}= \sqrt{\frac{6\pi^2}{\textrm{Vol}(B)}}\; n^{1/3} +o(1). \end{equation*} However, we are only interested in eigenvalues $\bar \lambda_n$ for which the eigenfunctions $\bar e_n$ have nonzero average, i.e., $$\int_B \bar e_n(y)\,dy\not=0.$$ To look for those eigenvalues, we first consider the special case that the domain is a ball. \begin{lemma}\label{eigen_N} Let $\lambda_{lj}$ be the eigenvalues of the three-dimensional Newton potential in the ball $\{x:\,|x|<a\}$. Then we have \begin{equation}\label{eigen_1} \lambda_{l j}^{-1}=\frac{\left[\mu_{j}^{\left(l+\frac{1}{2}\right)}\right]^{2}}{a^{2}}, \quad l=0,\,1,\, \ldots, \; j=1,\,2,\, \ldots, \end{equation} where the $\mu_{j}^{\left(l+\frac{1}{2}\right)}$ are the roots of the transcendental equation \begin{equation}\label{eigen_2} (2 l+1) J_{l+\frac{1}{2}}\left(\mu_{j}^{\left(l+\frac{1}{2}\right)}\right)+\frac{\mu_{j}^{\left(l+\frac{1}{2}\right)}}{2}\left[J_{l-\frac{1}{2}}\left(\mu_{j}^{\left(l+\frac{1}{2}\right)}\right)-J_{l+\frac{3}{2}}\left(\mu_{j}^{\left(l+\frac{1}{2}\right)}\right)\right]=0. \end{equation} The eigenfunctions corresponding to each eigenvalue $\lambda_{l j},$ form a complete orthogonal system and can be represented in the form \begin{equation}\label{eigen_3} u_{ljm}=J_{l+\frac{1}{2}}\left(\sqrt{\lambda_{l j}^{-1}} r\right) Y_{l}^{m}(\varphi, \theta), \end{equation} where $J_{l+\frac{1}{2}}$ are the Bessel functions and \begin{eqnarray} &&Y_{l}^{m}(\varphi, \theta)=P_{l}^{m}(\cos \theta) \cos m \varphi, \quad m=0,1, \ldots, l, \label{eigen_31} \\ &&Y_{l}^{m}(\varphi, \theta)=P_{l}^{|m|}(\cos \theta) \sin |m| \varphi, \quad m=-1, \ldots,-l \label{eigen_32} \end{eqnarray} for $l=0,1, \ldots$ are spherical functions. Here $P_{l}^{m}$ are the associated Legendre polynomials and $(r, \theta, \varphi)$ are the spherical coordinates with $\theta\in[0,\,\pi]$ and $\varphi\in[0,\,2\pi]$. \end{lemma} Since \begin{equation*} \int_{-1}^1 P_l^0(x)\,dx =0,\quad l\geq 1, \end{equation*} we can easily observe that the eigenfunctions \eqref{eigen_3} have nonzero average only for $m=0$ and $l=0$. The equation \eqref{eigen_2} with $l=0$ reduces to \begin{equation}\label{eigen_4} J_{\frac{1}{2}}\left(\mu_{j}^{\left(\frac{1}{2}\right)}\right)+\frac{\mu_{j}^{\left(\frac{1}{2}\right)}}{2}\left[J_{-\frac{1}{2}}\left(\mu_{j}^{\left(\frac{1}{2}\right)}\right)- J_{\frac{3}{2}}\left(\mu_{j}^{\left(\frac{1}{2}\right)}\right)\right]=0. \end{equation} Note that \begin{eqnarray} &&J_{-\frac{1}{2}}(x) = \sqrt{\frac{2}{\pi x}}\,\cos x, \; J_{\frac{1}{2}}(x) = \sqrt{\frac{2}{\pi x}}\,\sin x, \quad x>0, \label{Bessel_1}\\ &&J_{\frac{3}{2}}(x) = \frac{1}{x} J_{\frac{1}{2}}(x) - J_{-\frac{1}{2}}(x) = \frac{1}{x} \sqrt{\frac{2}{\pi x}}\,\sin x - \sqrt{\frac{2}{\pi x}}\,\cos x, \quad x>0. \label{Bessel_2} \end{eqnarray} Then we derive from \eqref{eigen_4} that \begin{equation}\label{eigen_41} \sin \left(\mu_{j}^{\left(\frac{1}{2}\right)}\right) + 2 \mu_{j}^{\left(\frac{1}{2}\right)}\, \cos \left(\mu_{j}^{\left(\frac{1}{2}\right)}\right) = 0, \end{equation} that is, \begin{equation}\label{eigen_42} \tan \left(\mu_{j}^{\left(\frac{1}{2}\right)}\right) = - 2 \mu_{j}^{\left(\frac{1}{2}\right)}. \end{equation} Note that $\tan x$ is monotonously increasing in $(j\pi-\pi/2,\,j\pi+\pi/2)$ and goes to $-\infty$ as $x\to j\pi-\pi/2$. So the equation \eqref{eigen_42} has a unique root in each interval $(j\pi-\pi/2,\,j\pi+\pi/2)$ for $ j=1,\,2,\,\ldots$. Moreover, we can also easily see that \begin{equation}\label{eigen_5} \mu_{j}^{\left(\frac{1}{2}\right)}=j\pi-\frac{\pi}{2}+\gamma_j, \quad j=1,\,2,\,\ldots, \end{equation} where $\gamma_j$ monotonously decreases to zero as $j\to +\infty$. Hence, for the spherical domain $\{x:\,|x|<a\}$, the eigenvalues $\bar\lambda_n$ with $\int_B \bar e_n(y)dy\not=0$ have the following asymptotic property: \begin{equation*} \bar \lambda_n^{-1} = a^{-2}(n\pi-\frac{\pi}{2}+\gamma_n)^2,\quad n=1,\,2,\,\ldots. \end{equation*} Consequently, we have \begin{equation*} \omega_n := c_1 \lambda_n^{-\frac{1}{2}} = c_1 \frac{n\pi - \frac{\pi}{2} + \gamma_n}{a}, \quad n=1,\,2,\,\ldots. \end{equation*} For simplicity, we let \begin{equation}\label{Riesz_b} b:=\frac{c_1\pi}{a}=1. \end{equation} Then we have \begin{equation}\label{expansion_omega_n} \omega_n=n-\frac{1}{2}+\frac{\gamma_n}{\pi},\quad n\in\mathbb N. \end{equation} In the following, we show that both $\{\cos(\omega_nt)\}_{n\in\mathbb N}$ and $\{\sin(\omega_nt)\}_{n\in\mathbb N}$ are Riesz basis in $L^2[0,\,\pi]$. To this end, let us state the following results in \cite{Chr2016, X-V2001}. \begin{lemma}\label{Riesz_cond} Let $\mathcal H$ be a Hilbert space. A sequence $\{p_n\}_{n\in\mathbb Z}$ is a Riesz basis in $\mathcal H$ if and only if it is complete in $\mathcal H$ and there exist two positive constants $\mathcal A$ and $\mathcal B$ such that for every finite scalar sequence $\{c_n\}$ \begin{equation}\label{Riesz_ineq} \mathcal A \sum_{n\in\mathbb Z} |c_n|^2 \leq \left\| \sum_{n\in\mathbb Z} c_n p_n \right\|^2 \leq \mathcal B \sum_{n\in\mathbb Z} |c_n|^2. \end{equation} \end{lemma} \begin{lemma}\label{H-V_cos} Let $\{\xi_n\},\,n\in\mathbb N_0:=\{0\}\cup\mathbb N$ be a sequence of nonnegative numbers with the property that $\xi_k\not=\xi_m$ for $k\not=m$ and of the form $\xi_n=n+\gamma+\gamma_n$ with $\gamma_n\in[-l,\,l]$ for sufficiently large $n$, where the constants $\gamma\in[0,\,\frac{1}{2}]$ and $l\in(0,\,\frac{1}{4})$ satisfy the condition \begin{equation}\label{H-V_cond} (1+\sin(2\gamma\pi))^{1/2}\,(1-\cos(l\pi)) + \sin(l\pi)<1. \end{equation} Then $\{\cos(\xi_nt)\}_{n\in\mathbb N_0}$ is a Riesz basis in $L^2[0,\,\pi]$. \end{lemma} \begin{lemma}\label{H-V_sin} Let $\{\xi_n\},\,n\in\mathbb N$ be a sequence of positive numbers with the property that $\xi_k\not=\xi_m$ for $k\not=m$ and of the form $\xi_n=n-\gamma+\gamma_n$ with $\gamma_n\in[-l,\,l]$ for sufficiently large $n$, where the constants $\gamma\in[0,\,\frac{1}{2}]$ and $l\in(0,\,\frac{1}{4})$ satisfy \eqref{H-V_cond}. Then $\{\sin(\xi_nt)\}_{n\in\mathbb N}$ is a Riesz basis in $L^2[0,\,\pi]$. \end{lemma} We are in a position to show the following result. \begin{lemma}\label{Riesz_cs} Both $\{\cos(\omega_nt)\}_{n\in\mathbb N}$ and $\{\sin(\omega_nt)\}_{n\in\mathbb N}$ are Riesz basis in $L^2[0,\,\pi]$. \end{lemma} {\bf Proof.} First, we observe that $\omega_n=n-\frac{1}{2}+\frac{\gamma_n}{\pi},\, n\in\mathbb N$ can be rewritten as $\tilde \omega_n=n+\frac{1}{2}+\frac{\gamma_{n+1}}{\pi},\, n\in\mathbb N_0$. Second, we know from $\lim_{n\to+\infty}\gamma_n=0$ that $\frac{\gamma_n}{\pi}\in[-l,\,l]$ with $l\in(0,\,\frac{1}{4})$ for large enough $n$. In addition, we can easily check that the condition \eqref{H-V_cond} with $\gamma=1/2$ is satisfied. Hence, by Lemma \ref{H-V_cos}, we conclude that $\{\cos(\omega_nt)\}_{n\in\mathbb N}$ is a Riesz basis in $L^2[0,\,\pi]$. Similarly, we can see from Lemma \ref{H-V_sin} that $\{\sin(\omega_nt)\}_{n\in\mathbb N}$ is also a Riesz basis in $L^2[0,\,\pi]$. \hfill $\Box$ \medskip Now define the sequence $\{p_n(t)\}_{n\in \mathbb Z^+\cup\mathbb Z^-}$ by \begin{equation}\label{p_n} p_n(t):= \begin{cases} \sin(\omega_nt), & n\in \mathbb Z^+, \\ \cos(\omega_{-n}t), & n\in\mathbb Z^-, \end{cases} \quad t\in[-\pi,\,\pi]. \end{equation} Then we have the following key result. \begin{theorem}\label{Riesz_pn} The sequence $\{p_n(t)\}_{n\in \mathbb Z^+\cup\mathbb Z^-}$ is a Riesz basis in $L^2[-\pi,\,\pi]$. \end{theorem} {\bf Proof.} For any $f\in L^2[-\pi,\,\pi]$, we define \begin{equation}\label{Riesz_sy_an} f_{sy}(t):=\frac{f(t)+f(-t)}{2},\quad f_{an}(t):=\frac{f(t)-f(-t)}{2}. \end{equation} Note that $f_{sy}(t)$ is symmetric and $f_{an}(t)$ is antisymmetric with respect to $t$. Since $\{\cos(\omega_nt)\}_{n\in\mathbb N}$ is a Riesz basis in $L^2[0,\,\pi]$, we expand $f_{sy}(t)$ as \begin{equation*} f_{sy}(t) = \sum_{n=1}^{+\infty} a_n \cos(\omega_n t), \quad t\in [0,\,\pi]. \end{equation*} Moreover, there exist two positive constants $\mathcal A_1$ and $\mathcal A_2$ such that \begin{equation*} \mathcal A_1 \sum_{n\in\mathbb Z^+} |a_n|^2 \leq \|f_{sy}\|^2_{L^2[0,\,\pi]} \leq \mathcal A_2 \sum_{n\in\mathbb Z^+} |a_n|^2. \end{equation*} By the symmetry of $f_{sy}(t)$ and $\cos(\omega_nt)$, we have \begin{equation}\label{Riesz_p_1} f_{sy}(t) = \sum_{n=1}^{+\infty} a_n \cos(\omega_n t), \quad t\in [-\pi,\,\pi] \end{equation} with \begin{equation} \label{Riesz_p_12} 2\mathcal A_1 \sum_{n\in\mathbb Z^+} |a_n|^2 \leq \|f_{sy}\|^2_{L^2[-\pi,\,\pi]} = 2 \|f_{sy}\|^2_{L^2[0,\,\pi]} \leq 2\mathcal A_2 \sum_{n\in\mathbb Z^+} |a_n|^2. \end{equation} Similarly, using the antisymmetry of $f_{an}(t)$ and $\sin(\omega_nt)$ with respect to $t$, we can expand $f_{an}(t)$ as \begin{equation}\label{Riesz_p_2} f_{an}(t) = \sum_{n=1}^{+\infty} b_n \sin(\omega_n t), \quad t\in [-\pi,\,\pi] \end{equation} with \begin{equation}\label{Riesz_p_22} 2\mathcal B_1 \sum_{n\in\mathbb Z^+} |b_n|^2 \leq \|f_{an}\|^2_{L^2[-\pi,\,\pi]} = 2 \|f_{an}\|^2_{L^2[0,\,\pi]} \leq 2\mathcal B_2 \sum_{n\in\mathbb Z^+} |b_n|^2, \end{equation} where $\mathcal B_1$ and $\mathcal B_2$ are two positive constants. From \eqref{Riesz_p_1} and \eqref{Riesz_p_2}, we obtain \begin{equation}\label{Riesz_p_3} f(t)=f_{sy}(t)+f_{an}(t) = \sum_{n=1}^{+\infty} a_n \cos(\omega_n t) + \sum_{n=1}^{+\infty} b_n \sin(\omega_n t) =: \sum_{n\in\mathbb Z^+\cup\mathbb Z^-} c_n p_n(t), \quad t\in[-\pi,\,\pi], \end{equation} where \begin{equation*} c_n:=\begin{cases} b_n, & n\in\mathbb Z^+,\\ a_{-n}, & n\in\mathbb Z^-. \end{cases} \end{equation*} We observe from the definitions of $f_{sy}$ and $f_{an}$, i.e., \eqref{Riesz_sy_an}, that \begin{equation*} \|f_{sy}\|^2_{L^2[-\pi,\,\pi]} \leq \|f\|^2_{L^2[-\pi,\,\pi]}, \quad \|f_{an}\|^2_{L^2[-\pi,\,\pi]} \leq \|f\|^2_{L^2[-\pi,\,\pi]}. \end{equation*} In addition, we see from $f(t)=f_{sy}(t)+f_{an}(t) $ that \begin{equation*} \|f\|^2_{L^2[-\pi,\,\pi]} \leq 2\left( \|f_{sy}\|^2_{L^2[-\pi,\,\pi]} + \|f_{an}\|^2_{L^2[-\pi,\,\pi]}\right). \end{equation*} Hence, we deduce from \eqref{Riesz_p_12} and \eqref{Riesz_p_22} that \begin{eqnarray*} \min\{\mathcal A_1,\,\mathcal B_1\} \sum_{n\in\mathbb Z^+\cup\mathbb Z^-} |c_n|^2 &=& \min\{\mathcal A_1,\,\mathcal B_1\}\left( \sum_{n\in\mathbb Z^-} |a_{-n}|^2 + \sum_{n\in\mathbb Z^+} |b_n|^2\right) \\ &\leq& \mathcal A_1 \sum_{n\in\mathbb Z^+} |a_n|^2 + \mathcal B_1 \sum_{n\in\mathbb Z^+} |b_n|^2 \\ &\leq&\frac{1}{2}\left( \|f_{sy}\|^2_{L^2[-\pi,\,\pi]} + \|f_{an}\|^2_{L^2[-\pi,\,\pi]}\right)\\ &\leq& \|f\|^2_{L^2[-\pi,\,\pi]}\\ &\leq&2\left( \|f_{sy}\|^2_{L^2[-\pi,\,\pi]} + \|f_{an}\|^2_{L^2[-\pi,\,\pi]}\right)\\ &\leq& 4\left(\mathcal A_2 \sum_{n\in\mathbb Z^+} |a_n|^2 +\mathcal B_2 \sum_{n\in\mathbb Z^+} |b_n|^2\right) \\ &\leq& 4\max\{\mathcal A_2,\,\mathcal B_2\} \left( \sum_{n\in\mathbb Z^-} |a_{-n}|^2 + \sum_{n\in\mathbb Z^+} |b_n|^2 \right)\\ &=& 4\max\{\mathcal A_2,\,\mathcal B_2\} \sum_{n\in\mathbb Z^+\cup\mathbb Z^-} |c_n|^2. \end{eqnarray*} The proof is complete. \hfill $\Box$ \bigskip Let $\mathcal T_c,\,\mathcal T_s:\,L^2[0,\,\pi] \to L^2[0,\,\pi]$ be the operators corresponding to the Riesz basis $\{\cos(\omega_nt)\}_{n\in\mathbb N}$ and $\{\sin(\omega_nt)\}_{n\in\mathbb N}$, respectively. That is, \begin{equation}\label{T_c} \mathcal T_c[\cos((n-1)t)] = \cos(\omega_nt), \quad t\in[0,\,\pi],\; n\in\mathbb N \end{equation} and \begin{equation}\label{T_s} \mathcal T_s[\sin(nt)] = \sin(\omega_nt), \quad t\in[0,\,\pi],\; n\in\mathbb N. \end{equation} Define $L^2_{sy}[-\pi,\,\pi]$ and $L^2_{an}[-\pi,\,\pi]$ by \begin{eqnarray*} L^2_{sy}[-\pi,\,\pi]&:=&\left\{ f\in L^2[-\pi,\,\pi],\, f(-t) = f(t), \; a.e., \,t\in[-\pi,\,\pi] \right\},\\ L^2_{an}[-\pi,\,\pi]&:=&\left\{ f\in L^2[-\pi,\,\pi],\, f(-t) = -f(t), \; a.e., \,t\in[-\pi,\,\pi] \right\}. \end{eqnarray*} We now extend the operators $\mathcal T_c,\,\mathcal T_s$ to $L^2_{sy}[-\pi,\,\pi]$ and $L^2_{an}[-\pi,\,\pi]$. That is, for $f_{sy}\in L^2_{sy}[-\pi,\,\pi]$ and $f_{an}\in L^2_{an}[-\pi,\,\pi]$, we define \begin{equation*} \tilde {\mathcal T}_c [f_{sy}(t)] := \begin{cases} \mathcal T_c [f_{sy}(t)] , & t\in[0,\,\pi],\\ \mathcal T_c [f_{sy}(-t)] , & t\in[-\pi,\,0] \end{cases} \end{equation*} and \begin{equation*} \tilde {\mathcal T}_s [f_{an}(t)] := \begin{cases} \mathcal T_s [f_{an}(t)] , & t\in[0,\,\pi],\\ -\mathcal T_s [f_{an}(-t)] , & t\in[-\pi,\,0], \end{cases} \end{equation*} respectively. Then we easily see that the operators \begin{equation*} \tilde {\mathcal T}_c:\, L^2_{sy}[-\pi,\,\pi] \to L^2_{sy}[-\pi,\,\pi], \quad \tilde {\mathcal T}_s:\, L^2_{an}[-\pi,\,\pi] \to L^2_{an}[-\pi,\,\pi] \end{equation*} are isomorphisms. Finally, noticing $L^2[-\pi,\,\pi]=L^2_{sy}[-\pi,\,\pi]\textcircled+ L^2_{an}[-\pi,\,\pi]$, we define the operator $\mathcal T:\,L^2[-\pi,\,\pi] \to L^2[-\pi,\,\pi]$ by $\mathcal T=\tilde {\mathcal T}_c \textcircled{+} \tilde {\mathcal T}_s$, which is also an isomorphism. Define \begin{equation}\label{xi_n} \xi_n(t):= \begin{cases} \sin(nt), & n\in\mathbb Z^+, \\ \cos((-n-1)t), &n\in\mathbb Z^-, \end{cases} \quad t\in [-\pi,\,\pi]. \end{equation} Then we have \begin{equation}\label{defn_T} \mathcal T[\xi_n(t)] = p_n(t), \quad n\in\mathbb Z^+\cup\mathbb Z^-. \end{equation} \section{Proof of Theorem \ref{Reconstruction-Introduction}}\label{sec-Reconstruction} Based on the property \eqref{defn_T}, we reconstruct the field $V(z,\,t)$ from the measurement $W(x, t)$ for a fixed $x$ outside $\Omega$ using the expansion (\ref{L-S_8}). The procedure is as follows. We rewrite (\ref{L-S_8}) as \begin{equation*} W(x,\,t)=\; \sum_{n=1}^{+\infty}\frac{\omega_n \left(\int_D e_n(y)\,dy\right)^2}{4\pi |x-z| \lambda_n}\int_{c_0^{-1}|x-z|}^{t} \sin(\omega_n (t-s))\, V(z,\,s-c_0^{-1}|x-z|)\,ds + O(a^2), \end{equation*} or \begin{eqnarray*} W(x,\,t)&=& \sum_{n=1}^{+\infty}\frac{\omega_n \left(\int_D e_n(y)\,dy\right)^2}{4\pi |x-z| \lambda_n}\; \sin(\omega_n t) \int_{c_0^{-1}|x-z|}^{t} \cos(\omega_n\, s)\, V(z,\,s-c_0^{-1}|x-z|)\,ds \nonumber\\ && -\sum_{n=1}^{+\infty}\frac{\omega_n \left(\int_D e_n(y)\,dy\right)^2}{4\pi |x-z| \lambda_n}\; \cos(\omega_n t) \int_{c_0^{-1}|x-z|}^{t} \sin(\omega_n\, s)\, V(z,\,s-c_0^{-1}|x-z|)\,ds + O(a^2). \end{eqnarray*} Note that the field $V$ can be expressed by a retarded potential \begin{equation*} V(x,\,t)=\dfrac{1}{4\pi c_0^2} \int_{B_x} \dfrac{J(y,\,t-|x-y|/c_0)}{|x-y|}\, dx, \end{equation*} where $B_x$ is a ball of radius $c_0 t$ centered at $x$. Due to the compactness of the support of $J$, we further suppose that the field $V$ is compactly supported with respect to t in $[0,\,\tilde T]$. Then, for $t>\tilde T+c_0^{-1} |x-z|$, the integrals in the above expansion of $W$ are independent of $t$. Define \begin{equation*} \alpha_n(x,\,z):= \frac{\omega_n \left(\int_D e_n(y)\,dy\right)^2}{4\pi |x-z| \lambda_n}. \end{equation*} Since the Riesz basis is defined on the interval $[-\pi, \pi]$, we need to shift the time interval. Set $$t_*:=\tilde T+c_0^{-1}|x-z|+\pi.$$ Using \begin{eqnarray*} \sin(\omega_nt) &=& \sin[\omega_n(t-t_*)+\omega_nt_*] = \sin(\omega_n(t-t_*)) \, \cos(\omega_nt_*) + \cos(\omega_n(t-t_*)) \, \sin(\omega_nt_*), \\ \cos(\omega_nt) &=& \cos[\omega_n(t-t_*)+\omega_nt_*] = \cos(\omega_n(t-t_*)) \, \cos(\omega_nt_*) - \sin(\omega_n(t-t_*)) \, \sin(\omega_nt_*), \end{eqnarray*} we have \begin{eqnarray}\label{Riesz_h1} W(x,\,t) &=& \sum_{n=1}^{+\infty}\alpha_n(x,\,z)\, \sin(\omega_n (t-t_*)) \int_{c_0^{-1}|x-z|}^{\tilde T +c_0^{-1}|x-z|} \cos(\omega_nt_*)\,\cos(\omega_n s)\, V(z,\,s-c_0^{-1}|x-z|)\,ds \nonumber\\ && + \sum_{n=1}^{+\infty}\alpha_n(x,\,z)\, \cos(\omega_n (t-t_*)) \int_{c_0^{-1}|x-z|}^{\tilde T +c_0^{-1}|x-z|} \sin(\omega_nt_*)\,\cos(\omega_n s)\, V(z,\,s-c_0^{-1}|x-z|)\,ds \nonumber\\ && -\sum_{n=1}^{+\infty}\alpha_n(x,\,z)\, \cos(\omega_n (t-t_*)) \int_{c_0^{-1}|x-z|}^{\tilde T +c_0^{-1}|x-z|} \cos(\omega_nt_*)\sin(\omega_n s)\, V(z,\,s-c_0^{-1}|x-z|)\,ds \nonumber\\ && +\sum_{n=1}^{+\infty}\alpha_n(x,\,z)\, \sin(\omega_n (t-t_*)) \int_{c_0^{-1}|x-z|}^{\tilde T +c_0^{-1}|x-z|} \sin(\omega_nt_*) \sin(\omega_n s)\, V(z,\,s-c_0^{-1}|x-z|)\,ds \nonumber \\ && + O(a^2) \nonumber\\ &=& \sum_{n=1}^{+\infty}\alpha_n(x,\,z)\, \sin(\omega_n (t-t_*)) \int_{c_0^{-1}|x-z|}^{\tilde T +c_0^{-1}|x-z|} \cos(\omega_n(s-t_*))\,V(z,\,s-c_0^{-1}|x-z|)ds \nonumber\\ && -\sum_{n=1}^{+\infty}\alpha_n(x,\,z)\, \cos(\omega_n (t-t_*)) \int_{c_0^{-1}|x-z|}^{\tilde T +c_0^{-1}|x-z|} \sin(\omega_n\, (s-t_*))\, V(z,\,s-c_0^{-1}|x-z|)\,ds \nonumber\\ && + O(a^2). \end{eqnarray} Now we measure $W(x,\,t)$ for $t\in[\tilde T+c_0^{-1}|x-z|,\,\tilde T+c_0^{-1}|x-z|+2\pi]$. Set $\tau:=t-t_*.$ Applying the operator $\mathcal T^{-1}$ to \eqref{Riesz_h1}, we have \begin{eqnarray}\label{Riesz_h2} &&\mathcal T^{-1}[W(x,\,t_*+\tau)] \nonumber\\ &=& \sum_{n=1}^{+\infty}\alpha_n(x,\,z)\, \sin(n \tau) \int_{c_0^{-1}|x-z|}^{\tilde T +c_0^{-1}|x-z|} \cos(\omega_n(s-t_*))\,V(z,\,s-c_0^{-1}|x-z|)\,ds \nonumber\\ && -\sum_{n=1}^{+\infty}\alpha_n(x,\,z)\, \cos((n-1) \tau ) \int_{c_0^{-1}|x-z|}^{\tilde T +c_0^{-1}|x-z|} \sin(\omega_n\, (s-t_*))\, V(z,\,s-c_0^{-1}|x-z|)\,ds \nonumber\\ && + O(a^2). \end{eqnarray} As the sequence $\{\cos((n-1)t),\,\sin(nt) \}_{n=1}^\infty$ is the Fourier basis in $L^2[-\pi,\,\pi]$, we can approximately get the coefficients \begin{equation}\label{Riesz_h3} A_n(x,\,z):=\int_{c_0^{-1}|x-z|}^{\tilde T +c_0^{-1}|x-z|} \cos(\omega_n (s-t_*))\,V(z,\,s-c_0^{-1}|x-z|)\,ds, \quad n\in\mathbb N \end{equation} and \begin{equation}\label{Riesz_h4} B_n(x,\,z):=\int_{c_0^{-1}|x-z|}^{\tilde T +c_0^{-1}|x-z|} \sin(\omega_n\, (s-t_*)) \,V(z,\,s-c_0^{-1}|x-z|)\,ds, \quad n\in\mathbb N. \end{equation} Suppose that \begin{equation}\label{assume_T} 0< \tilde T \leq 2\pi. \end{equation} Set $\tau_*:=s-t_*+\tilde T.$ Then we further have \begin{eqnarray}\label{Riesz_h5} A_n(x,\,z)&=&\int_{c_0^{-1}|x-z|}^{2\pi +c_0^{-1}|x-z|} \cos(\omega_n (s-t_*))\,V(z,\,s-c_0^{-1}|x-z|)\,ds \nonumber\\ &=&\int_{-\pi}^{\pi} \cos(\omega_n (\tau_*- \tilde T)) \,V(z,\,\tau_*+\pi)\,d\tau_* \nonumber\\ &=&\cos(\omega_n\tilde T ) \int_{-\pi}^{\pi} \cos(\omega_n \tau_*) \,V(z,\,\tau_*+\pi)\,d\tau_* \nonumber\\ && + \sin(\omega_n\tilde T ) \int_{-\pi}^{\pi} \sin(\omega_n \tau_*) \,V(z,\,\tau_*+\pi)\,d\tau_*, \quad n\in\mathbb N \end{eqnarray} and \begin{eqnarray}\label{Riesz_h6} B_n(x,\,z)&:=&\int_{c_0^{-1}|x-z|}^{2\pi +c_0^{-1}|x-z|} \sin(\omega_n\, (s-t_*)) \,V(z,\,s-c_0^{-1}|x-z|)\,ds \nonumber\\ &=&\int_{-\pi}^{\pi} \sin(\omega_n (\tau_*- \tilde T)) \,V(z,\,\tau_*+\pi)\,d\tau_* \nonumber\\ &=&\cos(\omega_n\tilde T ) \int_{-\pi}^{\pi} \sin(\omega_n \tau_*) \,V(z,\,\tau_*+\pi)\,d\tau_* \nonumber\\ && - \sin(\omega_n\tilde T ) \int_{-\pi}^{\pi} \cos(\omega_n \tau_*) \,V(z,\,\tau_*+\pi)\,d\tau_*, \quad n\in\mathbb N. \end{eqnarray} Hence, we can easily obtain the following two quantities: \begin{eqnarray*} C_n(x,\,z)&:=& \int_{-\pi}^{\pi} \cos(\omega_n \tau_*) \,V(z,\,\tau_*+\pi)\,d\tau_*,\quad n\in\mathbb N, \\ D_n(x,\,z)&:=& \int_{-\pi}^{\pi} \sin(\omega_n \tau_*) \,V(z,\,\tau_*+\pi)\,d\tau_*,\quad n\in\mathbb N, \end{eqnarray*} since \eqref{Riesz_h5} and \eqref{Riesz_h6} constitute a linear system with respect to $C_n(x,\,z)$ and $D_n(x,\,z)$. By Corollary 3.6.3 in \cite{Chr2016}, we get \begin{equation}\label{Riesz_h70} V(z,\,t+\pi) = \sum_{n=1}^{+\infty}\left[C_n(x,\,z)g_n(t) + D_n(x,\,z) h_n(t)\right], \quad t\in[-\pi,\,\pi], \end{equation} where $g_n(t):=(\mathcal T^{-1})^*\cos((n-1) t)$ and $h_n(t):=(\mathcal T^{-1})^*\sin(n t)$. \medskip By varying the location of $z$ in $\Omega$, we can get the field $V(\cdot,\,t)|_\Omega$ for $t\in[0,\,2\pi]$. Observe that as we measure $W(x, t)$ for the time $t$ larger than $T_J$, we have $V(x,\, t)=0$, and hence we actually measure only $U(x,\, t)$. Then, we can reconstruct the source $J(x,\,t)$ from the measurement data of $U(x,\,t)$ for a fixed $x$ over a time interval, by numerical differentiation applied to $V(\cdot,\,t)|_\Omega$, for $t\in[0,\,2\pi]$, and the equation $c_0^{-2} V_{tt} - \Delta V = J$. Here, we would like to emphasize that the condition \eqref{assume_T} is not critical, and the argument can work for any fixed $\tilde T$. Indeed, in the above derivations, we simply let $b= 1$ such that we could work on the standard interval $[-\pi,\, \pi]$. We can easily see that $2\pi$ is just the length of the time interval, nothing else. For general $b$, we can construct the Riesz basis on the interval $[-\pi/b, \pi/b]$ and do the shift similarly. \bigskip {\bf Acknowledgement:} This work is supported by National Natural Science Foundation of China (Nos. 12071072, 11971104) and the Austrian Science Fund (FWF): P 30756-NBL.

https://arxiv.org/abs/2211.00442

Singularities of Discrete Improper Indefinite Affine Spheres

We consider in this paper discrete improper affine spheres based on asymptotic nets. In this context, we distinguish the discrete edges and vertices that must be considered singular. The singular edges can be considered as discrete cuspidal edges, while some of the singular vertices can be considered as discrete swallowtails. The classification of singularities of discrete nets is a quite difficult task, and our results can be considered as a fisrt step in this direction. We also prove some characterizations of ruled discrete improper affine spheres which are analogous to the smooth case.

\section{Introduction} In this paper we consider asymptotic nets, which are natural nets for the discretization of surfaces parameterized by asymptotic coordinates. Asymptotic nets have been the object of recent and ancient research by many geometers, as one can see in the list of references of this paper (\cite{Bobenko2008}, \cite{Craizer2008}, \cite{Craizer2010}, \cite{Craizer2011}, \cite{Rorig2014}, \cite{Kaferbock2013}, \cite{Matsuura2003}, \cite{Milan2014}) There are many classes of affine surfaces with indefinite metric that have been defined as subclasses of asymptotic nets: Bobenko and Schief have defined discrete affine spheres \cite{Bobenko1999}, Matsuura and Urakawa have considered discrete improper affine spheres \cite{Matsuura2003}, and, generalizing this latter concept, Craizer, Anciaux and Lewiner have defined discrete affine minimal surfaces \cite{Craizer2008}. In this paper we shall consider the singularities of the discrete improper affine spheres (DIAS). We shall also describe some results concerning the ruled case. Smooth improper affine spheres (IAS) with indefinite metric can be obtained from a pair of planar curves by the so called centre-chord construction. The generic singularities of an IAS are cuspidal edges and swallowtails, and the projection of the singularities in the plane of the curves define a new planar curve called midpoint tangent locus (MPTL) (Giblin \cite{Giblin2008}), which consists of the midpoints of the chords connecting points of both curves with parallel tangents. Moreover, the projection of the cuspidal edges of the IAS are smooth points of the MPTL, while the projection of the swallowtails of the IAS are cusps of the MPTL (Craizer \cite{Craizer2011}). In this paper we consider the discrete analog of the centre-chord construction (\cite{Kobayashi2020}) and define the corresponding discrete midpoint tangent locus (DMPTL). The DMPTL is a polygonal line and some special vertices can be considered as discrete cusps. Then we define singular edges and vertices of the DIAS as those who projects in the MPTL. All singular edges of the DIAS will be considered as discrete cuspidal edges, while those vertices which projects into cusps of the MPTL will be considered as discrete swallowtails. With this approach, we can give a simple and easily verifiable definition for discrete cuspidal edges and discrete swallowtails in an arbitrary asymptotic net. The consistency of these definitions is remarkable, since it is quite difficult to have good definitions of singularities in the discrete nets. For a discussion of this question, see Rossman and Yasumoto \cite{Rossman2018}. In the smooth case, a ruled improper affine sphere (IAS) with indefinite Blaschke metric is affinely congruent to the graph of $z(x,y)= xy + \phi(x)$, for some smooth function $\phi$ (\cite{Nomizu1994}). Among these ruled surfaces, we can distinguish the Cayley surface, defined by $\phi(x)=-\frac{x^3}{3}$. It is characterized by the conditions $C\neq 0$ and $\nabla C=0$, where $C$ is the cubic form and $\nabla $ the induced connection. Finally, one can show that the generic singularities of ruled IAS are just cuspidal edges. In this paper we prove analogous results for the discrete improper affine spheres. The paper is organized as follows: In Section 2, we give all definitions and results concerning smooth IAS that are relevant to the discrete setting. In Section 3, we discuss DIAS, emphasizing the centre-chord construction. Section 4 is the core of the paper, where we define discrete cuspidal edges and discrete swallowtails. In Section 5 we discuss ruled DIAS. \section{Preliminaries on smooth affine theory} \subsection{Affine differential structure} Consider a parameterized smooth surface $q:U\subset\mathbb R^2\longrightarrow\mathbb R^3$, where $U$ is an open subset of the plane and, for $(u,v)\in U$, let $$ L(u,v)=[\,q_u,q_v,q_{uu}], \ M(u,v)=[\,q_u,q_v,q_{uv}], \ N(u,v)=[\,q_u,q_v,q_{vv}], $$ where $f_u (f_v)$ denotes the partial derivative of a function $f$ with respect to $u (v)$, and $[\cdot,\cdot,\cdot]$ denotes the determinant. The surface $q$ is said to be non-degenerate if $LN-M^2\neq0$, and in this case, the Berwald-Blaschke metric is defined by $ds^2=\frac{1}{|LN-M^2|^{1/4}}\left(L\,du^2+2M\,du\,dv+N\,dv^2\right).$ If $LN-M^2>0$, the metric is called \emph{definite} and the surface is locally convex. On the contrary, if $LN-M^2<0$, the metric is called \emph{indefinite} and the surface is locally hyperbolic, \emph{i.e.}, the tangent plane crosses the surface. From now on we shall assume that the affine surface has indefinite metric. We may assume that $(u,v)$ are \emph{asymptotic} coordinates, i.e., $L=N=0$. In this case, it is possible to take $M>0$, and the affine Blaschke metric takes the form $ds^2=2\Omega\,du\,dv$, where $\Omega^2=M$. Without loss of generality, we shall take $\Omega>0$. In asymptotic coordinates the structural equations are given by \begin{equation}\label{Gaussequation} \begin{array}{l} q_{uu}= \dfrac{\Omega_u}{\Omega}\,q_u+\dfrac{A}{\Omega}\,q_v \phantom{\dfrac{\frac{1}{1}}{\frac{1}{1}}}\\ q_{vv}= \dfrac{B}{\Omega}\,q_u+\dfrac{\Omega_v}{\Omega}\,q_v \phantom{\dfrac{\frac{1}{1}}{\frac{1}{1}}}\\ q_{uv}= \Omega\,\xi \end{array} \end{equation} where $\Omega=[q_u,q_v,\xi]$, $A=[q_u,q_{uu},\xi]$ and $B=[q_v,q_{vv},\xi]$ are the coefficients of the affine cubic form $A\,du^3+B\,dv^3$, and $\xi$ is the \emph{affine normal} vector field (\cite{Buchin1983}). The surface $q$ is called an {\it improper affine sphere} (IAS) if $\xi$ is constant. From now on we shall be interested only in IAS. In this case, the compatibility equations are given by \begin{equation}\label{Compatibility-eq-smooth} \Omega_{uv}\Omega-\Omega_u\Omega_v=AB,\,\,A_v=0\,\,\textrm{and}\,\,B_u=0. \end{equation} \subsection{The centre-chord construction} Consider two smooth planar curves $\alpha:I\longrightarrow\mathbb R^2$ and $\beta:J\longrightarrow\mathbb R^2$, where $I,J\subset\mathbb R$. Denote $$ x(u,v)=\frac{1}{2}(\alpha(u)+\beta(v)), \ \ y(u,v)=\frac{1}{2}(\beta(v)-\alpha(u)), $$ and define $z(u,v)$ by the relations $z_u=[x_u,y]$ and $z_v=[x_v,y]$. The following proposition is proved in \cite{Craizer2011}. \begin{prop} The map $q:I\times J\longrightarrow\mathbb R^3$ given by $q(s,t)\longrightarrow(x(s,t),z(s,t))$ is an IAS, and conversely, any IAS can be obtained from a pair of smooth planar curves $(\alpha,\beta)$ by the above construction. Moreover, \begin{equation}\label{OmegaCentreChord} \Omega(u,v)=\tfrac{1}{4}\left[ \alpha'(u), \beta'(v) \right], \ A=\tfrac{1}{4}[\alpha'(u),\alpha''(u)],\ B=-\tfrac{1}{4}[\beta'(v),\beta''(v)]. \end{equation} \end{prop} \subsection{Singularities of the IAS and the MPTL}\label{sec:CuspsMPTL} Even if $\Omega$ changes sign, we shall call the surface obtained by the centre-chord construction an IAS. In this context, the IAS may present singularities. The singular set $S$ of $q$ consists of all pairs $(u,v)$ for which $\Omega(u,v)=0$. From Equation \eqref{OmegaCentreChord}, it follows that $(u,v)\in S$ if and only if $\alpha'(u)$ and $\beta'(v)$ are parallel. The set $x(S)$ is called the \emph{midpoint parallel tangent locus} (MPTL) of the pair $(\alpha,\beta)$ (\cite{Giblin2008}). Generically, the MPTL is a smooth regular curve with some cusps (\cite{Craizer2011}). Moreover, a point $(u_0,v_0)\in S$ is a cusp if and only if, in a neighborhood of $(u_0,v_0)$, the MPTL is contained in a half-plane determined by the line supporting the chord $\alpha(u_0)\beta(v_0)$. The following proposition holds for a generic IAS (\cite{Craizer2011}): \begin{prop}\label{prop-germes}$\phantom{.}$ Let $(u_0,v_0)\in S$. \begin{enumerate}[(i)] \item The singular point $(u_0,v_0)$ is a smooth point of the MPTL if and only if it is a cuspidal edge of the IAS. \item The singular point $(u_0,v_0)$ is an ordinary cusp of the MPTL if and only if it is a swallowtail of the IAS. \end{enumerate} \end{prop} \subsection{Ruled improper indefinite affine spheres} A surface $q$ is ruled if either $u$-curves or $v$-curves are all straight lines. From the structural equations \eqref{Gaussequation}, it is clear that an IAS is ruled if and only if $A=0$ or $B=0$. Assume that $B=0$. Then by a change of variable of the form $V=V(v)$ we may also assume that $q_{vv}=0$, which implies that $\Omega$ is in fact a function of $u$, independent of $v$. Now a change of variable $U=\Omega(u)$ implies that we can in fact assume that $\Omega$ is a constant, say $1$. For a ruled IAS, we shall call such a parameterization {\it normalized}. In a normalized parameterization, one can easily verify that the cubic form is given by $C=-2Adu^3$ and so $\nabla C=0$ if and only if $A_u=0$, where $\nabla$ denotes the induced connection. The following result is proved in \cite{Nomizu1994}: \begin{thm}\label{teo-ruled-improper-affine-sphere} If $q$ is a smooth ruled IAS, then it is locally of the form $z=x^1x^2+\varphi(x^1)$ where $\varphi$ is an arbitrary function of $x^1$. \end{thm} One important example of such a surface is the so called Cayley surface, when $\varphi(x^1)=-\frac{(x^1)^3}{3}$. The Cayley surface can be parameterized in asymptotic coordinates by \begin{equation}\label{Cayley-surface-asymptotic} q(u,v)=\left(u,v+\frac{au^2}{2},uv+\frac{au^3}{6}\right) \end{equation} where $(u,v)\in U\subset\mathbb R^2$ and $a\neq 0$ is a constant. Note that this parameterization is normalized and $A=a$, which implies that $\nabla C=0$. Next theorem implies that the conditions $C\neq 0$ and $\nabla C=0$ implies that a ruled IAS is in fact affinely equivalent to the Cayley surface: \begin{thm}\label{Cayley-surface-characterization Let $q(u,v)$ be a normalized parameterization of a ruled IAS. Then $q(u,v)$ is affinely congruent to a Cayley surface if and only if $A\neq 0$ and $A_u=0$. \end{thm} \begin{proof} Assume that $A$ is a non-zero constant $a$. Then Equations \eqref{Gaussequation} imply that \begin{equation}\label{StructuralCayleyNormal} q_{uu}=aq_v,\ \ q_{uv}=(0,0,1), \ \ q_{vv}=0. \end{equation} By an affine transformation, we may assume that $q(0,0)=(0,0,0)$, $q_u(0,0)=(1,0,0)$ and $q_v(0,0)=(0,1,0)$. Now Equations \eqref{StructuralCayleyNormal} imply that $q_v=(0,1,u)$, and $q_u=(1,au,\tfrac{au^2}{2}+v)$, which in turn implies that $q$ is given by Equation \eqref{Cayley-surface-asymptotic}. \end{proof} Any ruled IAS can be otained by the centre-chord construction from a planar curve $\alpha(u)$ and a planar line $\beta(v)$. The singular points of the IAS are the pairs $(u_0,v)$ such that $\alpha'(u_0)$ is parallel to $\beta$ and $v\in\mathbb{R}$. Thus the MPTL is generically a discrete set of lines parallel to $\beta$ and the singularities of the IAS is a discrete set of spatial lines, whose points are all cuspidal edges of the IAS. \section{Basics on asymptotic nets} Given a discrete function $f: D\subset\mathbb Z^2\longrightarrow\mathbb R^3$, we denote the discrete partial derivatives with respect to $u$ and $v$, respectively, by $f_1\left(u+\frac{1}{2},v\right)=f(u+1,v)-f(u,v)$ and $f_2\left(u,v+\frac{1}{2}\right)=f(u,v+1)-f(u,v)$. A net $q:D\subset\mathbb Z^2\longrightarrow\mathbb R^3$ is called asymptotic if the ``crosses are planar'', \emph{i.e.}, $q_1(u+\frac{1}{2},v)$, $q_1(u-\tfrac{1}{2},v)$, $q_2(u,v+\tfrac{1}{2})$ and $q_2(u,v-\tfrac{1}{2})$ are coplanar (see \cite{Bobenko2008}, p.66). From the coplanarity we obtain that $$ \left[\,q_1\left(u\pm\tfrac{1}{2},v\right),\,q_2\left(u,v\pm\tfrac{1}{2}\right),\,q_{11}\left(u,v\right)\,\right] =\left[\,q_1\left(u\pm\tfrac{1}{2},v\right),\,q_2\left(u,v\pm\tfrac{1}{2}\right),\,q_{22}\left(u,v\right)\,\right]=0, $$ and we can assume $$ M\left(u+\tfrac{1}{2},v+\tfrac{1}{2}\right)=\left[\,q_1\left(u+\tfrac{1}{2},v\right),\,q_2\left(u,v+\tfrac{1}{2}\right),\,q_{12}\left(u+\tfrac{1}{2},v+\tfrac{1}{2}\right)\,\right]>0. $$ Similarly to the smooth case, the \emph{affine metric} $\Omega$ at a quadrangle $\left(u+\tfrac{1}{2},v+\tfrac{1}{2}\right)$ is defined by \begin{equation}\label{affine-metric} \Omega\left(u+\tfrac{1}{2},v+\tfrac{1}{2}\right)=\sqrt{M\left(u+\tfrac{1}{2},v+\tfrac{1}{2}\right)}. \end{equation} We also define the affine normal vector by \begin{equation}\label{normal-xie} \xi\left(u+\tfrac{1}{2},v+\tfrac{1}{2}\right)=\frac{q_{12}\left(u+\tfrac{1}{2},v+\tfrac{1}{2}\right)}{\Omega\left(u+\tfrac{1}{2},v+\tfrac{1}{2}\right)},\phantom{aaaa} \end{equation} (see \cite {Craizer2010}). \subsection{Discrete improper affine spheres} An asymptotic net is said to be a \emph{discrete improper affine sphere} (DIAS) if the affine normal $\xi$ is constant. From now on we shall be considering only this case, and we shall denote $\xi(u+\tfrac{1}{2},v+\tfrac{1}{2})$ by $\xi$. Thus we can write $$ q_{12}\left(u+\tfrac{1}{2},v+\tfrac{1}{2}\right)=\Omega\left(u+\tfrac{1}{2},v+\tfrac{1}{2}\right)\xi. $$ We define the coefficients of the cubic form by $$ A(u,v)=\left[q_1\left(u-\tfrac{1}{2},v\right),q_1\left(u+\tfrac{1}{2},v\right),\xi\right],\ \ B(u,v)=\left[q_2\left(u,v-\tfrac{1}{2}\right),q_2\left(u,v+\tfrac{1}{2}\right),\xi\right]. $$ \begin{lem} We have that $A=A(u)$ and $B=B(v)$. \end{lem} \begin{proof} We proof that $A_2(u,v+\tfrac{1}{2})=0$, the case for $B$ being similar. We have that $A_2(u,v+\tfrac{1}{2})=0$ is given by $$ \left[q_1\left(u-\tfrac{1}{2},v+1\right),q_1\left(u+\tfrac{1}{2},v+1\right),\xi\right]-\left[q_1\left(u-\tfrac{1}{2},v+1\right),q_1\left(u+\tfrac{1}{2},v\right),\xi\right]+ $$ $$ +\left[q_1\left(u-\tfrac{1}{2},v+1\right),q_1\left(u+\tfrac{1}{2},v\right),\xi\right]-\left[q_1\left(u-\tfrac{1}{2},v\right),q_1\left(u+\tfrac{1}{2},v\right),\xi\right]. $$ Since $q_1\left(u\pm\tfrac{1}{2},v+1\right)-q_1\left(u\pm\tfrac{1}{2},v\right)=\Omega(u\pm\tfrac{1}{2},v+\tfrac{1}{2})\xi$, the lemma is proved. \end{proof} \begin{prop}\label{prop-eq-structural} The structural equations of the DIAS are given by \begin{equation}\label{eq-structural1} \begin{array}{c} q_{11}(u,v)=\dfrac{\Omega_1\left(u,v+\tfrac{1}{2}\right)}{\Omega\left(u\pm\tfrac{1}{2},v+\tfrac{1}{2}\right)}\,q_1\left(u\pm\tfrac{1}{2},v\right)+ \dfrac{A(u)}{\Omega\left(u\pm\tfrac{1}{2},v+\tfrac{1}{2}\right)}\,q_2\left(u,v+\tfrac{1}{2}\right) \\ q_{11}(u,v)=\dfrac{\Omega_1\left(u,v-\tfrac{1}{2}\right)}{\Omega\left(u\pm\tfrac{1}{2},v-\tfrac{1}{2}\right)}\,q_1\left(u\pm\tfrac{1}{2},v\right)+ \dfrac{A(u)}{\Omega\left(u\pm\tfrac{1}{2},v-\tfrac{1}{2}\right)}\,q_2\left(u,v-\tfrac{1}{2}\right) \end{array} \end{equation} and \begin{equation}\label{eq-structural2} \begin{array}{c} q_{22}(u,v)= \dfrac{B(v)}{\Omega\left(u+\tfrac{1}{2},v\pm\tfrac{1}{2}\right)}\,q_1\left(u+\tfrac{1}{2},v\right)+ \dfrac{\Omega_2\left(u+\tfrac{1}{2},v\right)}{\Omega\left(u+\tfrac{1}{2},v\pm\tfrac{1}{2}\right)}\,q_2\left(u,v\pm\tfrac{1}{2}\right)\\ q_{22}(u,v)= \dfrac{B(v)}{\Omega\left(u-\tfrac{1}{2},v\pm\tfrac{1}{2}\right)}\,q_1\left(u-\tfrac{1}{2},v\right)+ \dfrac{\Omega_2\left(u+\tfrac{1}{2},v\right)}{\Omega\left(u-\tfrac{1}{2},v\pm\tfrac{1}{2}\right)}\,q_2\left(u,v\pm\tfrac{1}{2}\right) \end{array} \end{equation} \end{prop} For a proof of this proposition, see \cite{Craizer2010}. There is also a compatibility equation analogous to the Equation \ref{Compatibility-eq-smooth} in the smooth case: \begin{prop} The following compatibility equation holds: $$ \Omega\left(u+\tfrac{1}{2},v-\tfrac{1}{2}\right)\Omega\left(u-\tfrac{1}{2},v+\tfrac{1}{2}\right)-\Omega\left(u+\tfrac{1}{2},v+\tfrac{1}{2}\right)\Omega\left(u-\tfrac{1}{2},v-\tfrac{1}{2}\right)=A(u)B(v). $$ \end{prop} \begin{proof} We know that $$ q_{121}(u,v+\tfrac{1}{2})=\Omega_1(u,v+\tfrac{1}{2})\xi. $$ On the other hand, Equations \eqref{eq-structural1} imply that $q_{112}(u,v+\tfrac{1}{2})$ is given by $$ \dfrac{\Omega_1\left(u,v+\tfrac{1}{2}\right)}{\Omega\left(u+\tfrac{1}{2},v+\tfrac{1}{2}\right)}\,q_1\left(u+\tfrac{1}{2},v+1\right)-\dfrac{\Omega_1\left(u,v-\tfrac{1}{2}\right)}{\Omega\left(u+\tfrac{1}{2},v-\tfrac{1}{2}\right)}\,q_1\left(u+\tfrac{1}{2},v\right)+ $$ $$ +A(u) \left( \dfrac{q_2\left(u,v+\tfrac{1}{2}\right)}{\Omega\left(u+\tfrac{1}{2},v+\tfrac{1}{2}\right)}- \dfrac{q_2\left(u,v-\tfrac{1}{2}\right)}{\Omega\left(u+\tfrac{1}{2},v-\tfrac{1}{2}\right)}\, \right). $$ These expression can be written as $$ \frac{\Omega_1(u,v+\tfrac{1}{2})}{\Omega(u+\tfrac{1}{2},v+\tfrac{1}{2})}q_{12}(u+\tfrac{1}{2},v+\tfrac{1}{2})+\left( \frac{\Omega_1(u,v+\tfrac{1}{2})}{\Omega(u+\tfrac{1}{2},v+\tfrac{1}{2})}-\frac{\Omega_1(u,v-\tfrac{1}{2})}{\Omega(u+\tfrac{1}{2},v-\tfrac{1}{2})} \right)q_1(u+\tfrac{1}{2},v) $$ $$ +\frac{A(u)}{\Omega(u+\tfrac{1}{2},v+\tfrac{1}{2})}q_{22}(u,v)-\frac{A(u)\Omega_2(u+\tfrac{1}{2},v)}{\Omega(u+\tfrac{1}{2},v+\tfrac{1}{2})\Omega(u+\tfrac{1}{2},v-\tfrac{1}{2})}q_2(u,v-\tfrac{1}{2}). $$ Now take the component $q_1(u+\tfrac{1}{2},v)$ in this expression to obtain $$ \frac{\Omega_1(u,v+\tfrac{1}{2})}{\Omega(u+\tfrac{1}{2},v+\tfrac{1}{2})}-\frac{\Omega_1(u,v-\tfrac{1}{2})}{\Omega(u+\tfrac{1}{2},v-\tfrac{1}{2})}+\frac{A(u)B(v)}{\Omega(u+\tfrac{1}{2},v+\tfrac{1}{2})\Omega(u+\tfrac{1}{2},v-\tfrac{1}{2})}=0, $$ which proves the proposition. \end{proof} \subsection{The $x$-net and the $q$-net} The planar net $x(u,v)$, $(u,v)\in D\subset\mathbb{Z}^2$, will be called the $x$-net. The asymptotic spatial net $q(u,v)$, $(u,v)\in D\subset\mathbb{Z}^2$, will be called the $q$-net. Since $x_{12}(u,v)=0$, the quadrangles of the $x$-net are all paralellograms. \begin{lem}\label{ModelNet} Consider a DIAS with $q(0,0)=(0,0,0)$, $$ q(1,0)=(0,\beta,0),\ q(-1,0)=(\alpha,0,0),\ q(0,1)=(a,b,0),\ q(0,-1)=(c,d,0), $$ and denote $z(1,1)=z_1$, $z(-1,1)=z_2$, $z(-1,-1)=z_3$, $z(1,-1)=z_4$. Then $$ \frac{z_2}{z_1}=-\frac{\alpha\cdot b}{\beta\cdot a}, \ \ \frac{z_3}{z_1}=-\frac{\alpha\cdot d}{\beta\cdot a}, \ \ \frac{z_4}{z_1}=\frac{c}{a}. $$ \end{lem} \begin{proof} Observe first that $$ x(1,1)=(a,b+\beta),\ x(-1,1)=(a+\alpha,b),\ x(-1,-1)=(c+\alpha,d),\ x(1,-1)=(c,d+\beta). $$ Then the planarity of the stars at $(1,0)$, $(0,1)$, $(-1,0)$ and $(0,-1)$ implies the result. \end{proof} \subsection{Bilinear patches} For each quadrangle $(u+\tfrac{1}{2},v+\tfrac{1}{2})$, there exists a unique bilinear patch contained in a hyperbolic paraboloid with affine normal $\xi$ and passing through $q(u,v)$, $q(u+1,v)$, $q(u,v+1)$ and $q(u+1,v+1)$. We shall denote this bilinear patch by $B=B(u+\tfrac{1}{2},v+\tfrac{1}{2})$. A parameterization of this bilinear patch is given by $$ B(s,t)=q(u,v)+sq_1\left(u+\tfrac{1}{2},v\right)+tq_2\left(u,v+\tfrac{1}{2}\right)+st q_{12}\left(u+\tfrac{1}{2},v+\tfrac{1}{2} \right), $$ $0\leq s,t\leq 1$. The tangent plane to $B$ at $(u,v)$ contains $q_1(u+\tfrac{1}{2},v)$ and $q_2(u,v+\tfrac{1}{2})$, thus coinciding with the star plane at $(u,v)$. At a point of the edge $(u+\tfrac{1}{2},v)$, both $B(u+\tfrac{1}{2},v+\tfrac{1}{2})$ and $B(u+\tfrac{1}{2},v-\tfrac{1}{2})$ share the same tangent plane, ``linear interpolators'' of the star planes at $(u,v)$ and $(u+1,v)$. We conclude that the bilinear patches glue at the edges with the same tangent planes (\cite{Craizer2010}, \cite{Rorig2014}, \cite{Kaferbock2013}). In this article, the bilinear patches are used just to visualize the discrete cuspidal edges and discrete swallowtails in Section 4. \subsection{Discrete centre-chord construction} We describe now the centre-chord construction for a general DIAS, which can also be found in \cite{Kobayashi2020}. Let $\alpha:I\longrightarrow\mathbb R^2$ and $\beta:J\longrightarrow\mathbb R^2$, where $I,J\subset\mathbb Z$, and define $$ x(u,v)=\frac{1}{2}(\alpha(u)+\beta(v)),\ \ \ y(u,v)=\frac{1}{2}(\beta(v)-\alpha(u)). $$ Define a function $z:I\times J\longrightarrow\mathbb R$ by the conditions $$ z_1(u+\tfrac{1}{2},v)=[x_1(u+\tfrac{1}{2},v),y(u,v)],\ \ z_2(u,v+\tfrac{1}{2})=[x_2(u,v+\tfrac{1}{2}),y(u,v)]. $$ Observe that $$z_{12}(u+\tfrac{1}{2},v+\tfrac{1}{2})=\frac{1}{4}[\alpha_1(u+\tfrac{1}{2}),\beta_2(v+\tfrac{1}{2})]=z_{21}(u+\tfrac{1}{2},v+\tfrac{1}{2}),$$ which implies that $z$ is well-defined. Define \begin{equation}\label{eq:MetricCenterChord} \Omega(u+\tfrac{1}{2},v+\tfrac{1}{2})=\tfrac{1}{4}\ [\alpha_1(u+\tfrac{1}{2}),\beta_2(v+\tfrac{1}{2})]. \end{equation} \begin{prop} \label{centre-chord-DIAS} Assume $\Omega(u+\tfrac{1}{2},v+\tfrac{1}{2})>0$, for any $(u,v)\in I\times J$. Then the net $q(u,v)=(x(u,v),z(u,v))$ defines a DIAS with cubic form $$ A=\tfrac{1}{4}[\alpha_1(u-\tfrac{1}{2}),\alpha_1(u+\tfrac{1}{2})], \ \ B=-\tfrac{1}{4}[\beta_2(v-\tfrac{1}{2}),\beta_2(v+\tfrac{1}{2})]. $$ \end{prop} \begin{proof} Observe first that $$ \alpha_{11}(u,v)=\frac{\Omega_1(u,v\pm\tfrac{1}{2})}{\Omega(v\pm\tfrac{1}{2},v\pm\tfrac{1}{2})}\alpha_1(u\pm\tfrac{1}{2})+\frac{A(u,v)}{\Omega(v\pm\tfrac{1}{2},v\pm\tfrac{1}{2})}\beta_2(v\pm\tfrac{1}{2}) $$ and $$ \beta_{22}(u,v)=\frac{B(u,v)}{\Omega(v\pm\tfrac{1}{2},v\pm\tfrac{1}{2})}\alpha_1(u\pm\tfrac{1}{2})+\frac{\Omega_2(u\pm\tfrac{1}{2},v)}{\Omega(v\pm\tfrac{1}{2},v\pm\tfrac{1}{2})}\beta_2(v\pm\tfrac{1}{2}). $$ These equations imply that $$ q_{11}(u,v)=\frac{\Omega_1(u,v\pm\tfrac{1}{2})}{\Omega(v\pm\tfrac{1}{2},v\pm\tfrac{1}{2})}q_1(u\pm\tfrac{1}{2})+\frac{A(u,v)}{\Omega(v\pm\tfrac{1}{2},v\pm\tfrac{1}{2})}q_2(v\pm\tfrac{1}{2}) $$ and $$ q_{22}(u,v)=\frac{B(u,v)}{\Omega(v\pm\tfrac{1}{2},v\pm\tfrac{1}{2})}q_1(u\pm\tfrac{1}{2})+\frac{\Omega_2(u\pm\tfrac{1}{2},v)}{\Omega(v\pm\tfrac{1}{2},v\pm\tfrac{1}{2})}q_2(v\pm\tfrac{1}{2}), $$ thus proving that $q$ is an asymptotic net. Moreover $$ q_{12}(u+\tfrac{1}{2},v+\tfrac{1}{2})=\Omega(u+\tfrac{1}{2},v+\tfrac{1}{2})\xi, $$ where $\xi=(0,0,1)$, thus proving that $q$ is a DIAS. \end{proof} The converse of Proposition \ref{centre-chord-DIAS} also holds: \begin{prop}\label{DIAS-centre-chord} Any DIAS $q:I\times J\longrightarrow\mathbb R^3$ can be obtained by the centre-chord construction, that is, $q(u,v)=(x(u,v),z(u,v))$, where $x(u,v)=\frac{1}{2}(\alpha(u)+\beta(v))$, $z_1(u+\tfrac{1}{2},v)=[x_1(u+\tfrac{1}{2},v),y(u,v)]$, $z_2(u,v+\tfrac{1}{2})=[x_2(u,v+\tfrac{1}{2}),y(u,v)]$ and $y(u,v)=\frac{1}{2}(\beta(v)-\alpha(u))$, for some polygonal lines $\alpha:I\longrightarrow\mathbb R^2$ and $\beta:J\longrightarrow\mathbb R^2$, where $I,J\subset\mathbb Z$. \end{prop} \noindent\emph{Proof.} We may assume that $\xi=(0,0,1)$ and so $q(u,v)=(x(u,v),z(u,v))$, where $x(u,v)$ is the projection in the plane $z=0$. Since $$ q_{12}(u+\tfrac{1}{2},v+\tfrac{1}{2})=\Omega(u+\tfrac{1}{2},v+\tfrac{1}{2})\,\xi,$$ we conclude that $x_{12}=0$, which implies that $$x(u,v)=\frac{1}{2}(\alpha(u)+\beta(v)),$$ for some polygonal lines $\alpha:I\longrightarrow\mathbb R^2$ and $\beta:J\longrightarrow\mathbb R^2$, where $I,J\subset\mathbb Z$. Since $$\begin{array}{rcl} q_1(u+\tfrac{1}{2},v) &=& \frac{1}{2}\left(\alpha_1(u+\tfrac{1}{2}),\,z_1(u+\tfrac{1}{2})\right),\\ q_2(u,v+\tfrac{1}{2}) &=& \frac{1}{2}\left(\beta_2(v+\tfrac{1}{2}),\,z_2(v+\tfrac{1}{2})\right), \phantom{\dfrac{1}{1}}\\ q_{12}(u+\tfrac{1}{2},v+\tfrac{1}{2}) &=& \left(0,\,\Omega(u+\tfrac{1}{2},v+\tfrac{1}{2})\right), \end{array}$$ we conclude that $$\Omega^2(u+\tfrac{1}{2},v+\tfrac{1}{2})=[q_1,q_2,q_{12}]=\tfrac{1}{4}\Omega(u+\tfrac{1}{2},v+\tfrac{1}{2})\,\left[\alpha_1(u+\tfrac{1}{2}),\,\beta_2(v+\tfrac{1}{2})\right]$$ and so $$\Omega(u+\tfrac{1}{2},v+\tfrac{1}{2})=\tfrac{1}{4}\left[\alpha_1(u+\tfrac{1}{2}),\,\beta_2(v+\tfrac{1}{2})\right]>0.$$ Thereafter, $$z_{12}(u+\tfrac{1}{2},v+\tfrac{1}{2}))=\tfrac{1}{4}\left[\alpha_1(u+\tfrac{1}{2}),\,\beta_2(v+\tfrac{1}{2})\right],$$ and by discrete integration we obtain $$ z_1(u+\tfrac{1}{2},v)=\left[x_1(u+\tfrac{1}{2}),\,y(u,v)\right],\ \ z_2(u,v+\tfrac{1}{2})=\left[x_2(v+\tfrac{1}{2}),\,y(u,v)\right], $$ where $y(u,v)=\frac{1}{2}(\beta(v)-\alpha(u))$, which completes the proof.\hfill$\square$ \section{Singularities of discrete improper indefinite affine spheres} Even if we do not assume the hypothesis $\Omega>0$, we shall call the asymptotic net obtained by the centre-chord construction a DIAS. In this context, ``singularities'' may appear. Consider two discrete planar polygonal lines $\alpha:I\longrightarrow\mathbb R^2$ and $\beta:J\longrightarrow\mathbb R^2$, where $I,J\subset\mathbb Z$. We shall assume that: \begin{itemize} \item[$\bullet$] For any point $\alpha(u)$ and any triplet $\beta(v-1)$, $\beta(v)$, $\beta(v+1)$, we have that $\beta(v)$ is within the interior of the angle $\beta(v-1)\alpha(u)\beta(v+1)$, supposed less than $180^\circ$. \item[$\bullet$] For any point $\beta(v)$ and any triplet $\alpha(u-1)$, $\alpha(u)$, $\alpha(u+1)$, we have that $\alpha(u)$ is within the interior of the angle $\alpha(u-1)\beta(v)\alpha(u+1)$, supposed less than $180^\circ$. See Figure \ref{Fig-condition-planar-curves}. \end{itemize} \begin{figure}[!htb] \centering \includegraphics[width=.5\linewidth]{Sing-condicao-curvas.eps} \caption{\small Restriction to the pair of planar curves $(\alpha,\beta)$.}\label{Fig-condition-planar-curves} \end{figure} This restriction is made to simplify our first model of singularities, but we think that is something to be explored in future works about the subject. \subsection{Singular (cuspidal) edges of the asymptotic net} The singular set of a smooth IAS is characterized by the condition $\Omega(u,v)=0$. The discrete analog is the following: \begin{defn} An edge $(u+\tfrac{1}{2},v)$ is called {\it singular} if $$ \Omega(u+\tfrac{1}{2},v+\tfrac{1}{2})\cdot\Omega(u+\tfrac{1}{2},v-\tfrac{1}{2})<0. $$ Similarly, An edge $(u,v+\tfrac{1}{2})$ is called {\it singular} if $$ \Omega(u+\tfrac{1}{2},v+\tfrac{1}{2})\cdot\Omega(u-\tfrac{1}{2},v+\tfrac{1}{2})<0. $$ \end{defn} In the smooth case the condition $\Omega(u,v)=0$ is equivalent to $\alpha'(u)$ and $\beta'(v)$ being parallel. In the discrete setting we have the following: \begin{lem}\label{LemmaHalfParallel} Let $\alpha:I\to\mathbb{R}^2$ and $\beta:J\to\mathbb{R}^2$, $I,J\subset\mathbb{Z}$, be polygonal lines. Then: \begin{enumerate} \item\label{HalfParallel1} An edge $(u+\tfrac{1}{2},v)$ is singular if and only if $\beta(v-1)$ and $\beta(v+1)$ are in the same half-plane determined by the straight line given by $\beta(v)+r\alpha_1(u+\tfrac{1}{2})$, for $r\in\mathbb R$. \item\label{HalfParallel2} An edge $(u,v+\tfrac{1}{2})$ is singular if and only if if $\alpha(u-1)$ and $\alpha(u+1)$ are in the same half-plane determined by the straight line given by $\alpha(u)+r\beta_2(v+\tfrac{1}{2})$, for $r\in\mathbb R$. \end{enumerate} \end{lem} \begin{proof} Immediate from Equation \eqref{eq:MetricCenterChord}. \end{proof} When condition (\ref{HalfParallel1}) of Lemma \ref{LemmaHalfParallel} holds, we say that $\alpha_1(u+\tfrac{1}{2})$ is {\it discretely parallel} to $\beta(v)$. When condition (\ref{HalfParallel2}) of Lemma \ref{LemmaHalfParallel} holds, we say that $\beta_2(v+\tfrac{1}{2})$ is {\it discretely parallel} to $\alpha(u)$ (see Figure \ref{Fig-sing-dmptl}). \begin{figure}[!htb] \centering \includegraphics[width=.45\linewidth]{Sing-paralelismo1-pb.eps} \caption{\small Discrete parallelism between $\beta_2(v+\tfrac{1}{2})$ and $\alpha(u)$. The edge $x_2(u,v+\tfrac{1}{2})$ belongs to the DMPTL of the pair of polygonal lines $(\alpha,\beta)$. The line styles indicate the $u$ and $v$ directions.}\label{Fig-sing-dmptl} \end{figure} Observe that the discrete parallelism is associated to a triangle formed by the point of one polygonal line and an edge of the other, as we can see on Figure \ref{Fig-sing-dmptl}. We shall call the set of all midsegments of these triangles the \emph{discrete midpoint parallel tangent locus} (DMPTL) of the pair of polygonal lines $(\alpha,\beta)$. We can characterize an edge of the DMPTL among all edges of the $x$-net by the following proposition: \begin{prop} Consider a segment $x_1(u,v+\tfrac{1}{2})$ of the $x$-net and denote the straight line containing it by $r(u,v+\tfrac{1}{2})$. The following statements are equivalent: \begin{enumerate} \item The segment $x_1(u,v+\tfrac{1}{2})$ is an edge of the DMPTL. \item The straight line $r(u,v+\tfrac{1}{2})$ leaves $x(u-1,v)$ and $x(u+1,v)$ at the same half-plane. \item The straight line $r(u,v+\tfrac{1}{2})$ leaves $x(u-1,v+1)$ and $x(u+1,v+1)$ at the same half-plane. \end{enumerate} \end{prop} \begin{proof} It follows directly from the fact that the quadrangles of the $x$-net are parallelograms with $x_1(u+\frac{1}{2},v)$ parallel to $\alpha_1(u+\tfrac{1}{2})$ \end{proof} \begin{cor}\label{HalfStarPlane} Consider a segment $q_1(u,v+\tfrac{1}{2})$ of the asymptotic net and denote the straight line containing it by $s(u,v+\tfrac{1}{2})$. The following statements are equivalent: \begin{enumerate} \item The segment $q_1(u,v+\tfrac{1}{2})$ is an edge of the DMPTL. \item In the star plane at $q(u,v)$, the straight line $s(u,v+\tfrac{1}{2})$ leaves $q(u-1,v)$ and $q(u+1,v)$ at the same half-plane. \item In the star plane at $q(u,v+1)$, the straight line $s(u,v+\tfrac{1}{2})$ leaves $q(u-1,v+1)$ and $q(u+1,v+1)$ at the same half-plane. \end{enumerate} \end{cor} We say that an edge $q_1(u,v+\tfrac{1}{2})$ of the DIAS is \emph{singular} if it satisfies one, and hence all, the conditions of Corollary \ref{HalfStarPlane}. With this definition, it is clear that an edge of the DIAS is singular if and only if its projection is an edge of the DMPTL, which is a discrete counterpart of the correponding property of the smooth case. Observe that we can check whether or not the edge $q_1(u,v+\tfrac{1}{2})$ is singular by looking at the star plane at $q(u,v)$ or at the star plane at $q(u,v+1)$. Thus the above definition (items (2) and (3) of Corollary \ref{HalfStarPlane}) of a singular edge can be directly extended to any asymptotic net, even if it does not represent a DIAS. The singular edges of a DIAS are the discrete counterpart of the cuspidal edges of the smooth IAS. So, in the discrete setting, the expressions singular edge and {\it cuspidal edge} have the same meaning. \begin{figure}[!htb] \centering \includegraphics[width=.50\linewidth]{CuspidalEdgeInterpolator0.png} \caption{\small Discrete cuspidal edge with bilinear patches to help visualization.}\label{Fig:DiscreteSwallowtailInterpolator0} \end{figure} In Figure \ref{Fig:DiscreteSwallowtailInterpolator0}, one can see how a discrete cuspidal edge with the bilinear interpolators looks like a smooth cuspidal edge. It is easy to verify that the two bilinear patches associated with a cuspidal edge are at the same side of the vertical plane containing this edge, and in fact this property characterizes cuspidal edges. \subsection{Singular polygonal line} Let us construct the DMPTL step by step. Suppose that $\alpha_1(u-\tfrac{1}{2})$ is discretely parallel to $\beta(v)$ for some $u$ and $v$, then we have formed a triangle and its midsegment is part of the DMPTL. The next step is to decide what adjacent triangle we should choose and that is going to be clear after next Proposition. \begin{prop}\label{Prop-dmptl-construction} Let $\alpha$ and $\beta$ be planar polygonal lines such that $\alpha_1(u-\tfrac{1}{2})$ is discretely parallel to $\beta(v)$. Then one and only one of the three follow statements holds: \begin{enumerate} \item $\alpha_1(u+\tfrac{1}{2})$ is discretely parallel to $\beta(v)$, as in Figure \ref{Fig-dmptl-first}(left); \item $\beta_2(v+\tfrac{1}{2})$ is discretely parallel to $\alpha(u)$, as in Figure \ref{Fig-dmptl-first}(centre); \item $\beta_2(v-\tfrac{1}{2})$ is discretely parallel to $\alpha(u)$, as in Figure \ref{Fig-dmptl-first}(right). \end{enumerate} \end{prop} \noindent\emph{Proof.} Let us fix $\alpha(u-1)$, $\alpha(u)$, $\beta(v-1)$, $\beta(v)$ and $\beta(v+1)$ such that the hypothesis keeps valid and see what can happen to the point $\alpha(u+1)$ (see Figure \ref{Fig-sing-proposition-c-possib}). \begin{figure}[!htb] \centering \includegraphics[width=.4\linewidth]{Sing-proposition-0-pb.eps} \caption{\small The three possible configurations for the vertex $\alpha(u+1)$.}\label{Fig-sing-proposition-c-possib} \end{figure} Let $r$ and $s$ be straight lines passing through $\alpha(u)$ and parallel to the edges $\beta_2(v-\tfrac{1}{2})$ and $\beta_2(v+\tfrac{1}{2})$, respectively. They divide the plane in four open regions and the point $\alpha(u+1)$ can be only in three of them. In fact, by the restriction made to the curves $\alpha$ and $\beta$ at the beginning of the section, it can not be at the same region wherein is the point $\alpha(u-1)$. Also, since there is no parallelism between edges of $\alpha$ and $\beta$, $\alpha(u+1)$ can not be neither in the straight line $r$ nor in $s$. If $\alpha(u+1)$ belongs to the region opposed to the region where $\alpha(u-1)$ belongs, then $\alpha_1(u+\tfrac{1}{2})$ is discretely parallel to $\beta(v)$. Thus $x_1(u+\tfrac{1}{2},v)$ belongs to the DMPTL, while $x_2(u,v\pm\tfrac{1}{2})$ does not and we are in case (1). \begin{figure}[!htb] \includegraphics[width=.32\linewidth]{Sing-proposition-1-pb.eps} \hfill \includegraphics[width=.32\linewidth]{Sing-proposition-2-pb.eps} \hfill \includegraphics[width=.32\linewidth]{Sing-proposition-3-pb.eps} \caption{\small Three possibilities for the construction of the DMPTL.}\label{Fig-dmptl-first} \end{figure} If $\alpha(u+1)$ is above $r$ and $s$, then $\beta_2(v+\tfrac{1}{2})$ is discretely parallel to $\alpha(u)$. Thus $x_2(u,v+\tfrac{1}{2})$ belongs to the DMPTL, while $x_1(u+\tfrac{1}{2},v)$ and $x_2(u,v-\tfrac{1}{2})$ does not, and we are in case (2). If $\alpha(u+1)$ is below $r$ and $s$, then $\beta_2(v-\tfrac{1}{2})$ is discretely parallel to $\alpha(u)$. Thus $x_2(u,v-\tfrac{1}{2})$ belongs to the DMPTL, while $x_1(u+\tfrac{1}{2},v)$ and $x_2(u,v+\tfrac{1}{2})$ does not, and we are in case (3). \hfill$\square$ \begin{cor} The DMPTL is a planar polygonal line. As a consequence, the set of singular edges of a DIAS form a spatial polygonal line. \end{cor} \subsection{Configuration of a star} Let us make some notes about possible configurations for star planes in the asymptotic net or in the planar net. A star plane at $q(u,v)$ is called \emph{typical} if the four points $q(u+1,v)$, $q(u,v+1)$, $q(u-1,v)$ and $q(u,v-1)$ appear in this order, clockwise or counter clockwise, with respect to $q(u,v)$, and \emph{atypical} otherwise (see Figure \ref{Fig-typical-at-star}). We can consider similarly typical and atypical vertices in the planar net. It is clear that $q(u,v)$ is typical for the asymptotic net if and only if $x(u,v)$ is typical for the planar net. \begin{figure}[!htb] \begin{minipage}[b]{0.24\linewidth} \centering \includegraphics[width=1\linewidth]{Typical-star-pb.eps} \end{minipage}\hfill \begin{minipage}[b]{0.23\linewidth} \centering \includegraphics[width=1\linewidth]{Typical-star-2-pb.eps} \end{minipage}\hfill \begin{minipage}[b]{0.24\linewidth} \centering \includegraphics[width=1\linewidth]{Atypical-star-2-pb.eps} \end{minipage}\hfill \begin{minipage}[b]{0.23\linewidth} \centering \includegraphics[width=1\linewidth]{Atypical-star-pb.eps} \end{minipage} \caption{\small Both figures on the left show two different possibilities of a typical star, whilst both on the right show two possible configurations for an atypical one. The line styles indicate the $u$ and $v$ directions.}\label{Fig-typical-at-star} \end{figure} We have the following proposition: \begin{prop} Consider a vertex $x(u,v)$ of the planar net. Then one and only one of the following conditions holds: \begin{enumerate} \item[$(0)$] No edge in the star is in the DMPTL. \item[$(1)$] Two consecutive edges with the same label are in the DMPTL. \item[$(2)$] Two adjacent edges with different labels are in the DMPTL and the star is typical. \item[$(3)$] Two adjacent edges with different labels are in the DMPTL and the star is atypical. \end{enumerate} \end{prop} \begin{proof} If no edges of the star at $x(u,v)$ is in the DMPTL, we are in case (0). If at least one is in the DMPTL, we may assume it is $x_1(u-\tfrac{1}{2},v)$. Then proceeding as in the proof of Proposition \ref{Prop-dmptl-construction}, there are three possibilities, cases (1), (2) or (3). In case (1), two consecutive edges with the same label are in the DMPTL. In case (2), two adjacent edges with different labels are in the DMPTL and the star is typical. Finally in case (3), two adjacent edges with different labels are in the DMPTL and the star is atypical. \end{proof} \begin{cor}\label{Configs} Consider a vertex $q(u,v)$ of the asymptotic net. Then one and only one of the following conditions holds: \begin{enumerate} \item[$(0)$] No edge in the star is singular. \item[$(1)$] Two consecutive edges with the same label are singular. \item[$(2)$] Two adjacent edges with different labels are singular and the star is typical. \item[$(3)$] Two adjacent edges with different labels are singular and the star is atypical. \end{enumerate} \end{cor} Figures \ref{Fig-cuspidaledge-config-uu} and \ref{Fig:DiscreteCuspidalEdgesInterpolator} show a neighborhood of $q(u,v)$ satisfying conditions (1) and (2) of Corollary \ref{Configs}. \begin{figure}[!htb] \includegraphics[width=.4\linewidth]{Cuspidaledge-uu-3-pb.eps} \hspace{1cm} \includegraphics[width=.4\linewidth]{cuspidaledge-uv-3-pb.eps} \caption{\small A pair of cuspidal edges satisfying conditions (1) and (2) of Corollary \ref{Configs}, respectively.}\label{Fig-cuspidaledge-config-uu} \end{figure} \begin{figure}[!htb \includegraphics[width=.3\linewidth]{CuspidalEdgeInterpolator1.png} \hfill \includegraphics[width=.5\linewidth]{CuspidalEdgeInterpolator2.png} \caption{\small Same cases as in Figure \ref{Fig-cuspidaledge-config-uu} with bilinear patches to help visualization.}\label{Fig:DiscreteCuspidalEdgesInterpolator} \end{figure} \subsection{Swallowtail vertices of the $q$-net} \begin{prop}\label{CuspMPTL} Consider a vertex $x(u,v)$ of the DMPTL. The following conditions are equivalent: \begin{enumerate} \item Two adjacent edges of the $x$-net with different labels are singular and the star is atypical. \item The two adjacent vertices of the DMPTL are in the same half-plane determined by the line supporting the chord $\alpha(u)\beta(v)$. \end{enumerate} \end{prop} \begin{proof} Straightforward corollary of Proposition \ref{Prop-dmptl-construction}. \end{proof} We say that $x(u,v)$ is a \emph{cusp} of the DMPTL if it satisfies one, and hence both, of the conditions of Proposition \ref{CuspMPTL}. To justify this definition, one should compare condition (2) of this proposition with the condition for a cusp in a smooth MPTL described on Section \ref{sec:CuspsMPTL}. The main definition of this paper is the following: \begin{defn} The vertex $q(u,v)$ of the asymptotic net is called a \emph{swallowtail} if two adjacent edges with different labels are singular and the star is atypical. \end{defn} From this definition, next proposition is immediate: \begin{prop} A vertex $q(u,v)$ of the $q$-net is a swallowtail if and only if the corresponding vertex $x(u,v)$ of the $x$-net is a cusp of the DMPTL. \end{prop} We remark that the definition of a swallowtail vertex can be extended to any asymptotic net, even if it does not correspond to a DIAS. In Figures \ref{Theo-swallowtail} and \ref{Fig:DiscreteSwallowtailInterpolator} we can see a swallowtail vertex $q(u,v)$. Observe the visual similarity of the smooth swallowtail and the discrete swallowtail with bilinear interpolators. \begin{figure}[!htb] \centering \includegraphics[width=.9\linewidth]{Theo-swallowtail-pb.eps} \caption{\small Swallowtail at $q(u,v)$ with cuspidal edges $q_1(u-\tfrac{1}{2},v)$ (strongest dashed segment) and $q_2(u,v-\tfrac{1}{2})$ (strongest solid segment).}\label{Theo-swallowtail} \end{figure} \begin{figure}[!htb \centering \includegraphics[width=.60\linewidth]{SwallowInterpolator.png} \caption{\small Discrete swallowtail with bilinear patches to help visualization.}\label{Fig:DiscreteSwallowtailInterpolator} \end{figure} \subsection{A swallowtail geometrical property} Let us remember that a swallowtail point in the smooth case always imply in self-intersection, so we expect that behavior in discrete case too. \begin{prop} Let $q:I\times J\subset\mathbb Z^2\longrightarrow\mathbb R^3$ be a DIAS. If $q(u,v)$ is a swallowtail point, then there is a pair of quadrangles, with $q(u,v)$ as a vertex, whose corresponding bilinear patches intersect each other. \end{prop} \noindent\emph{Proof.} Assume $q(0,0)=(0,0,0)$ is a swallowtail of the DIAS, with adjacent cuspidal edges $q_1(-\tfrac{1}{2},0)$ and $q_2(0,\tfrac{1}{2})$. Taking into account the notation of Lemma \ref{ModelNet}, we have that $a<0$, $b>0$, $c>0$, $d>0$ which implies that $z_1$ and $z_3$ have the same sign. We shall assume that both are positive, the other case being analogous. Under the above assumptions, the bilinear patches $B_1=B(\tfrac{1}{2},\tfrac{1}{2})$ and $B_{-1}=B(-\tfrac{1}{2},-\tfrac{1}{2})$ are both contained in the half-space $z\geq 0$. Moreover, the segment $q_1(-\tfrac{1}{2},0)\subset B_{-1}$ is contained in the plane $z=0$ and is below $B_1$. Similarly, $q_2(0,\tfrac{1}{2})\subset B_1$ is contained in the plane $z=0$ and is below $B_{-1}$. We conclude that necessarily $B_1\cap B_{-1}\neq \emptyset$. \hfill$\square$ \subsection{Example}\label{Example} Consider the curves $$ \alpha(u)=\left(u,\,5-\frac{(u-2)^2}{8}\right),\ u\in I,\ \ \ \beta(v)=\left(v^2-2,\,v\right),\ v\in J. $$ By considering $I,J\subset\mathbb{R}$, we obtain a smooth IAS by the centre-chord construction, and by considering $I,J\subset\mathbb{Z}$, we obtain a DIAS by the discrete centre-chord construction. \begin{figure}[!htb] \includegraphics[width=.4\linewidth]{Ex1-caustica-suave-pb.eps} \hfill \includegraphics[width=.47\linewidth]{Ex1-caustica-disc-pb.eps} \caption{\small MPTL (left) and DMPTL (right) associated to the pair of polygonal lines $(\alpha,\beta)$ of the example of Section \ref{Example}.}\label{Fig-Ex1-caustica-suave} \end{figure} We show in Figure \ref{Fig-Ex1-caustica-suave} the MPTL in the smooth case and the DMPTL in the discrete case. Note that both of them are formed by two connected components and only one of them presents a cusp, which means that the cuspidal curves of both surfaces (smooth and discrete) generated by the pair $(\alpha,\beta)$ have two connected components and a unique swallowtail point, as it can be seen in Figure \ref{Fig-ex-sing}. \begin{figure}[!htb] \includegraphics[width=.48\linewidth]{Ex-sing.eps} \hfill \includegraphics[width=.48\linewidth]{Ex-sing-vistasuperior.eps} \caption{\small Two views of the DIAS of the example of Section \ref{Example}.}\label{Fig-ex-sing} \end{figure} \section{Ruled nets} Ruled nets are defined in the same way as in smooth case, that is, in at least one of the coordinates direction, $u$-curves or $v$-curves are all straight lines. From Equations \eqref{eq-structural1}, one can easily check that a DIAS is ruled if and only if $A=0$ or $B=0$. \subsection{A characterization of ruled DIAS} Consider a ruled DIAS. We may assume, w.l.o.g., that $B(v)=0$ and $\beta(v)=(0,v)$, $v\in J\subset\mathbb{Z}$. Next proposition is a discrete counterpart of Theorem \ref{teo-ruled-improper-affine-sphere}. \begin{prop} Every ruled DIAS without singularities is of the form $z=x^1x^2+\varphi(x^1)$, for some real function $\varphi$. \end{prop} \noindent \emph{Proof.} Write $\alpha(u)=(\alpha^1(u),\alpha^2(u))$, $u\in I\subset\mathbb{Z}$. The hypothesis of no singularities implies that $\alpha^1_1(u+\tfrac{1}{2})$ does not change sign. Thus $\alpha^1(u)$ is an invertible map. We have $$ x(u,v)=(\alpha^1(u),\alpha^2(u)+v),\ y(u,v)=(-\alpha^1(u),v-\alpha^2(u)), $$ $$ z_1(u+\tfrac{1}{2},v)=v\alpha^1_1(u+\tfrac{1}{2})-[\alpha_1(u+\tfrac{1}{2}),\alpha(u)],\ z_2(u,v+\tfrac{1}{2})=\alpha^1(u). $$ \smallskip\noindent By discrete integration on $u$ we have $z(u,v)=v\alpha^1(u)+g(u)$, where $g_1(u+\tfrac{1}{2})=-[\alpha_1(u+\tfrac{1}{2}),\alpha(u)]$ and so $z=x^1x^2+\varphi(u)$, where $\varphi(u)=g(u)-\alpha^1(u)\alpha^2(u)$. Since $x^1(u)=\alpha^1(u)$ is invertible, the proposition is proved. \hfill$\square$ \subsection{Discrete Cayley surface} Let us now take a look at an example of a ruled DIAS, known as discrete Cayley surface. Its discrete structure equations are $$\begin{array}{l} q_{11}(u,v)=aq_2(u,v+\tfrac{1}{2})=a(0,1,u), \\ q_{22}(u,v)=(0,0,0),\\ q_{12}(u+\tfrac{1}{2},v+\tfrac{1}{2})=(0,0,1). \end{array}$$ We shall assume as initial conditions $q(0,0)=(0,0,0)$, $q(0,1)=(0,1,0)$ and $q(1,0)=(1,0,0)$. So the solution shall be \begin{equation}\label{Cayley-surface-discrete} q(u,v)=\left(u, v+\frac{au(u-1)}{2},uv+\dfrac{au(u^2-1)}{6}\right), \;(u,v)\in\mathbb Z^2. \end{equation} Note that for this example, $\Omega=1$ and $q_{22}=(0,0,0)$. As in the smooth case, we shall call {\it normalized} a DIAS satisfying these conditions. For a normalized DIAS, the structural equations become \begin{equation} q_{11}(u,v)=A(u)q_2(u,v+\tfrac{1}{2}),\ \ q_{12}(u+\tfrac{1}{2},v+\tfrac{1}{2})=(0,0,1)\ \ q_{22}(u,v)=(0,0,0). \end{equation} Thus we have proved the following theorem: \begin{thm}\label{Cayley-surface-characterization-discrete Let $q$ be a normalized DIAS. Then $q$ is affinely congruent to a discrete Cayley surface if and only if $A\neq 0$ and $A_1=0$. \end{thm} This result should be considered as a discrete analog of Theorem \ref{Cayley-surface-characterization}. \subsection{Singularities of ruled DIAS} Any ruled DIAS can be otained by the centre-chord construction from a planar polygonal line $\alpha(u)$ and a planar straight line $\beta(v)$. The singular points of the DIAS are the pairs $(u_0,v)$ such that $\alpha'(u_0)$ is parallel to $\beta$ and $v\in\mathbb{Z}$. Thus the DMPTL is generically a discrete set of lines parallel to $\beta$ and the singularities of the DIAS is a discrete set of spatial lines, whose points are all cuspidal edges of the DIAS. \section*{\textbf{Acknowledgement}} The authors are thankful to CAPES and CNPq for financial support during the preparation of this paper. They also thank Pontifical Catholic University of Rio de Janeiro. \bibliographystyle{amsplain}

https://arxiv.org/abs/2106.12821

Some weighted Hardy-type inequalities and applications

We study the two-weighted estimate\[\bigg\|\sum_{k=0}^na_k(x)\int_0^xt^kf(t)dt|L_{q,v}(0,\infty)\bigg\|\leqc\|f|L_{p,u}(0,\infty)\|,\tag{$*$}\]where the functions $a_k(x)$ are not assumed to be positive. It is shown that for $1<p\leq q\leq\infty$, provided that the weight $u$ satisfies the certain conditions, the estimate $(*)$ holds if and only if the estimate\[\sum_{k=0}^n\bigg\|a_k(x)\int_0^xt^kf(t)dt|L_{q,v}(0,\infty)\bigg\|\leq c\|f|L_{p,u}(0,\infty)\|.\tag{$**$}\]is fulfilled. The necessary and sufficient conditions for $(**)$ to be valid are well-known. The obtained result can be applied to the estimates of differential operators with variable coefficients in some weighted Sobolev spaces.

\section{Introduction} The problem of finding the necessary and sufficient conditions imposed on the functions $u(x)$, $v(x)$, for which the estimate of the kind \begin{equation} \bigg(\int_0^{\infty}\bigg|v(x)\int_0^xA(x,t)f(t)dt\bigg|^qdx\bigg)^ {1/q}\leq c\bigg(\int_0^{\infty}|u(x)f(x)|^pdx\bigg)^{1/p} \end{equation} are valid, attracted great attention for the last decade (see, e.g., a wide bibliography in the work of V.~D.~Stepanov \cite{1}). It is well known that the cases $p>q$ and $p\leq q$ are quite different. Here and everywhere below we assume that $p\leq q$. For $1\leq p\leq q\leq\infty$ and $A(x,t)\equiv1$ the necessary and sufficient condition for validity of the estimate $(1.1)$ has been obtained by J.~S.~Bradley \cite{2} and V.~M.~Kokilashvili \cite{3}. The first substantial progress for $A(x,t)\!\not\equiv\!1$ has been reached in the papers by V.~D.~Stepanov \cite{4}, \cite{5} and F.~J.~Mar\-tin--Reyes and E.~Sawyer \cite{6} where they investigated the case $A(x,t)=(x-t)^{\alpha}$, $\alpha>0$. At present the most general classes of kernels $A(x,t)$ are apparently considered by R.~Oinarov \cite{7}. The kernels of these classes are positive and satisfy some additional different types restrictions of which the most known is ``Oinarov's condition'': $$ c_1A(x,t)\leq A(x,y)+A(y,t)\leq c_2A(x,t),\;\;\;t<y<x,\;\;\;c_1,c_2>0. $$ We should like to emphasize the following fact. As far we know, until recently no work has been available in which the criteria for the validity of (1.1) with the kernels $A(x,t)$ of alternating signs would have been considered. In the recent work \cite{8} the author has, however, presented some class of kernels with alternating signs for which we managed to characterize the admissible weights in (1.1) for $p=q=2$ provided that the weight $u$ satisfies some extra conditions. In the present work we extend this result to all $1<p\leq q\leq\infty$ and weaken the conditions on $u$. As an application we studied in \cite{8} the problem of description of pointwise multipliers in some weighted Sobolev spaces. As an application here, we solve a more general problem of finding the criteria of bounded action for differential operators with variable coefficients from the same weighted Sobolev spaces to the weighted spaces~$L_q$. The paper is organized as follows: the main results are formulated in \S 2. The proofs are given in \S 3 and \S 4. All the functions are assumed to be measurable and finite almost everywhere. \section{Results} Let $\Bbb{R}^+=(0,\infty)$ be a half-line, $1\leq p\leq\infty$, and $u$ be a non-negative function on $\Bbb{R}$ (weight). Denote by $L_{p,u}$ a weighted space of functions $f:\Bbb{R}^+\to\Bbb{C}$ with a norm \begin{align*} &\|f|L_{p,u}\|=\bigg(\int_0^{\infty}|u(x)f(x)|^pdx\bigg)^{1/p},\;\;\; 1\leq p<\infty,\\ &\|f|L_{\infty,u}\|=\operatornamewithlimits{ess\,sup}_{x>0}|u(x)f(x)|. \end{align*} In what follows the latter modification is meant everywhere. We denote by $p'$ a conjugate exponent: $p'=p/(p-1)$. The main result of the present paper is the following \begin{theorem} Given $n\in\Bbb{N}$ and functions $a_k\!:\!\Bbb{R}^+\!\to\!\Bbb{C}$, $k\!=\!0,\dots,n$, let $1<p\leq q\leq\infty$, $u$, $v$ be two non-negative functions on $\Bbb{R}^+$. Assume that there exists a constant $D(u)$ such that \begin{gather} \int_0^ru^{-p'}(x)dx<\infty\;\;\;\forall r>0,\\ \int_0^{2r}x^{(n-1)p'}u^{-p'}(x)dx\leq D(u)\int_0^rx^{(n-1)p'}u^{-p'} (x)dx\;\;\;\forall r>0. \end{gather} Then the following three assertions are equivalent: \rom{(i)} The inequality $$ \bigg\|\sum_{k=0}^na_k(x)\int_0^xt^kf(t)dt|L_{q,v}\bigg\|\leq c \|f|L_{p,u}\| $$ holds; \rom{(ii)} the inequalities $$ \bigg\|a_k(x)\int_0^xt^kf(t)dt|L_{q,v}\bigg\|\leq c_k\|f|L_{p,u}\|, \;\;\;k=0,\dots,n, $$ hold; \rom{(iii)} the values \begin{gather*} S_k=\sup_{r>0}\bigg(\int_r^{\infty}|a_k(x)v(x)|^qdx\bigg)^{1/q}\cdot \bigg(\int_0^rx^{-kp'}u^{-p'}(x)dx\bigg)^{1/p'}<\infty,\\ k=0,\dots,n, \end{gather*} are finite. If $\widetilde{c}$ is the best constant in \rom{(i)}, then $$ K_1(p,q)\widetilde{c}\leq\sum_{k=0}^n S_k\leq K_2(D(u),n,p,q)\widetilde{c}. $$ \end{theorem} \begin{rem} For $p=q=2$ and under more restrictive condition on $u$ of the form $$ \int_{\Delta}u^{-2}(x)dx\leq D(u)\int_{\frac{1}{2}\Delta}u^{-2}(x)dx, $$ where $\Delta$ is any interval in $\Bbb{R}^+$ ($\frac{1}{2}\Delta$ is the twice smaller interval with the same center), such a theorem has been proved by the author in~\cite{8}. \end{rem} \begin{rem} There exist weights $u$ not satisfying $(2.2)$ for which the assertion of Theorem $2.1$ does not hold. For example, for $u(x)=e^{-x}$ the estimate $$ \bigg\|x\int_0^xf(t)dt-\int_0^xtf(t)dt|L_{2,e^{-x}}\bigg\|\leq c\|f| L_{2,e^{-x}}\| $$ is valid, while the corresponding estimates for either summand in the left-hand side do not hold separately. For details we refer to the author's work \cite{8}, Prop.~1.2 and also to A.~Kufner \cite{9}, Example 1. \end{rem} As an application of Theorem 2.1 we consider estimates of differential operators with variable coefficients in weighted Sobolev spaces. For $l\in\Bbb{N}$ we distinguish two types of such spaces: $$ W_{p,u}^l=\bigg\{f:\Bbb{R}^+\to\Bbb{C}\;\big|\; \|f|W_{p,u}^l\|=\sum_{k=0}^{l-1}|f^{(k)}(0)|+\|f^{(l)}|L_{p,u}\| <\infty\bigg\}, $$ and \begin{align*} \overset{\circ}{W}{}_{p,u}^l=&\bigg\{f:\Bbb{R}^+\to\Bbb{C}|f^{(k)}(0)=0,\;\;\; k=0,\dots,l-1,\|f|\overset{\circ}{W}{}_{p,u}^l\|=\|f^{(l)}|L_{p,u}\|<\infty\bigg\}. \end{align*} The use of notation $f^{(l)}$ assumes that the function $f$ has an absolutely continuous derivative $f^{(l-1)}$. Then $f^{(l)}=(f^{(l-1)})'$ exists almost everywhere. In what follows, the condition (2.1) is assumed to be fulfilled. In this case $\|f^{(l)}|L_{p,u}\|<\infty$ implies the existence of $f^{(k)}(0)$, $k=0,\dots,l-1$, which may be understood as the limits $f^{(k)}(0+0)$, and the definitions of the spaces $W_{p,u}^l$ and $\overset{\circ}{W}{}_{p,u}^l$ become meaningful. Note that all polynomials of degree $\leq l-1$ belong to $W_{p,u}^l$. The spaces $W_{p,u}^l$ have been introduced and studied by L.~D.~Kudryavtsev \cite{10}. Let us consider a differential operator of the $l$-th order with variable coefficients $$ P(x,D)=\sum_{m=0}^lb_m(x)D^m, $$ which acts by the rule $$ P(x,D)f(x)=\sum_{m=0}^lb_m(x)f^{(m)}(x). $$ Introduce the notation $$ d_k(x)=\sum_{m=0}^{l-1-k}b_m(x)\frac{x^{l-1-k-m}}{(l-1-k-m)!},\;\;\; k=0,\dots,l-1. $$ We then have the following theorem. \begin{theorem} Let $l\in\Bbb{N}$, $1<p\leq q\leq\infty$, and $u$,$v$ be non-negative on $\Bbb{R}^+$ functions, and let the condition $(2.1)$ be fulfilled. If $l\geq 2$, then let the condition $(2.2)$ with $n=l-1$ be also fulfilled. \rom{(i)} In order for $$ P(x,D):\overset{\circ}{W}{}_{p,u}^l\to L_{q,v}, $$ it is necessary and sufficient that the conditions \begin{gather} \sup_{r>0}\bigg(\int_r^{\infty}|d_k(x)v(x)|^qdx\bigg)^{1/q}\bigg( \int_0^rx^{kp'}u^{-p'}(x)dx\bigg)^{1/p'}<\infty,\\ k=0,\dots,l-1,\notag \end{gather} and \begin{equation} b_l(x)\equiv 0\;\;\;\text{for}\;\;p<q,\;\;\|b_lvu^{-1}|L_{\infty} (\Bbb{R}^+)\|<\infty\;\;\text{for}\;\;p=q \end{equation} be fulfilled. \\[-1pt] \rom{(ii)} In order that $$ P(x,D):W_{p,u}^l\to L_{q,v}, $$ it is necessary and sufficient that the conditions $(2.3)$ and $(2.4)$ be fulfilled and also that \begin{equation} \|P(x,D)x^k|L_{q,v}\|<\infty,\;\;\;k=0,\dots,l-1. \end{equation} \end{theorem} \begin{rem} This theorem generalizes Theorem $2.4$ on pointwise multipliers in $W_{p,u}^l$ from the author's work $\cite{8}$, which considered the case $p=q=2$, and $P(x,D)f=(\varphi f)^{(m)}$, $0\leq m\leq l$. \end{rem} In conclusion let us formulate the following open problem: Find for $l\geq 2$ the necessary and sufficient conditions on the function $\varphi$ under which $$ \|\varphi f|W_{2,e^{-x}}^l\|\leq c\|f|W_{2,e^{-x}}^l\|\;\;\;\;\;\forall f\in W_{2,e^{-x}}^l. $$ In other words, it is required to describe pointwise multipliers in the space $W_{2,e^{-x}}^l$. Theorem 2.4 in \cite{8} does not answer this question. Note that for $l=1$, $1\leq p\leq\infty$ and for arbitrary weights $u,v$ we can show that the estimate $$ \|\varphi f|W_{p,v}^1\|\leq c\|f|W_{p,u}^1\|\;\;\;\;\forall f\in W_{p,u}^1 $$ holds if and only if the function $\varphi$ satisfies the conditions \begin{gather*} \|\varphi vu^{-1}|L_{\infty}(\Bbb{R}^+)\|<\infty,\\ \sup_{r>0}\bigg(\int_r^{\infty}|\varphi'(x)v(x)|^pdx\bigg)^{1/p} \bigg(\int_0^ru^{-p'}(x)dx\bigg)^{1/p'}<\infty. \end{gather*} \section{Proof of Theorem $2.1$} \begin{lem} Let $X$ be a Banach space, $X^*$ be its conjugate and let $Y\subset X$ be a closed subspace. Furthermore, let $e\in X$, $e\not\in Y$ be a fixed vector. Then \rom{(i)} $\sup\{\langle y^*,e\r:y^*|_Y=0$, $\|y^*|X^*\|=1\}=\operatorname{dist}(e,Y)$. \rom{(ii)} If there exists a vector $y_0\in Y$ such that $\operatorname{dist}(e,Y)=\|e-y_0\|$, then there also exists a functional $y_0^*\in X^*$, $y_0^*|_Y=0$, $\|y_0^*|X^*\|=1$ such that $$ \langle y_0^*,e\r=\|e-y_0\|. $$ \end{lem} \begin{rem} The symbol $\langle y^*,e\r$ denotes the value of the functional $y^*\!\in\!X^*$ on the vector $e\in X$. The writing $y^*|_Y=0$ means that $\langle y^*,y\r=0$ $\;\forall y\in Y$. Item (i) of the above lemma is formulated as Exercise $19$ in Chapter $4$ of W.~Rudin's book $\cite{11}$. \end{rem} \begin{proof} Let $y\in Y$, $y^*|_Y=0$, $\|y^*|X^*\|=1$. We see that $$ \langle y^*,e\r=\langle y^*,e-y\r\leq\|e-y\|\leq\operatorname{dist}(e,Y), $$ whence \begin{equation} \sup\{\langle y^*,e\r:y^*|_Y=0,\;\;\;\|y^*\|=1\}\leq\operatorname{dist}(e,Y). \end{equation} Next, since $e\not\in Y$, by the Hahn--Banach theorem (see W.~Rudin \cite{11}, Th.$3.3$, for every $y\in Y$ there exists $y^*\in X^*$ such that $\|y^*\|=1$, $y^*|_Y=0$ and the norm of the functional $y^*$ is attained on the vector $e-y$, i.e., $$ \|e-y\|=\langle y^*,e-y\r=\langle y^*,e\r. $$ This implies that the inverse inequality in (3.1) is valid and hence item (i) is fulfilled. Item (ii) is proved simultaneously. \end{proof} \begin{lem} Let $w$ be a positive function on $\Bbb{R}^+$, such that $$ \int_0^rw(x)dx<\infty\;\;\;\;\;\forall r>0. $$ Let $1\leq s<\infty$, $P_{n,r}(x)$ be a polynomial of the $n$-th degree with a highest degree term $x^n$, satisfying the conditions $$ \int_0^r|P_{n,r}(x)|^{s-1}\operatorname{sign} P_{n,r}(x)x^kw(x)dx=0, \;\;\;k=1,\dots,n. $$ Then \rom{(i)} all roots of the polynomial $P_{n,r}(x)$ are simple, real and belong to the interval $(0,r)$. \rom{(ii)} Let, in addition, the condition \begin{equation} \int_0^{2r}w(x)dx\leq c\int_0^rw(x)dx\;\;\;\;\;\forall r>0 \end{equation} be fulfilled with some constant $c=c(w)$. Let $x_1(r)$ be the least root of the polynomial $P_{n,r}(x)$. Then \begin{equation} \int_0^rw(x)dx\leq K\int_0^{x_1(r)}w(x)dx\;\;\;\;\;\forall r>0, \end{equation} for some constant $K=K(c,n,s)$. \end{lem} \begin{rem} The polynomial $P_{n,r}(x)$ does exist for every $n\in\Bbb{N}$ and $r>0$, and is the solution of the extremal problem $$ \int_0^r|P(x)|^sxw(x)dx\to\underset{P}{\min}, $$ where $P(x)$ runs through all polynomials with the highest degree term $x^n$. See \cite{12}, \S 2.1.1, where also item (i) is proved (for $w\equiv 1$, but the proof works also in the general case). \end{rem} \noindent{\it Proof.} It remains to prove (ii). Let $x_1,\dots,x_n\in(0,r)$ be the roots of $P_{n,r}(x)$ enumerated in the increasing order. Suppose also that $x_{n+1}=r$. We shall prove that \begin{equation} \int_0^{x_{m+1}}wdx\leq K_1\int_0^{x_m}wdx\;\;\;\text{for all}\;\;\; m=1,\dots,n. \end{equation} This will imply the required inequality (3.3) with the constant $K=K_1^n$. For $x_{m+1}\leq 4x_m$ the inequality (3.4) follows from (3.2) with $K_1=c^2$. Let now $x_{m+1}>4x_m$. Taking the polynomial $$ R(x)=\prod(x-x_k), $$ where the product is taken over all $k\in\{1,\dots,n\}\backslash\{m\}$, by the definition of $P_{n,r}$ we have $$ \int_0^r|P_{n,r}|^{s-1}\operatorname{sign}(P_{n,r})Rxwdx=0, $$ which can be written as \begin{equation} \ \hskip-1cm \int_{x_m}^r\!\!\prod_{k=1}^{m-1}|x-x_k|^s|x-x_m|^{s-1} \!\prod_{k=m+1}^n\! |x-x_k|^sxwdx\!=\!\int_0^{x_m}\!\text{(same integrand)} \end{equation} (with obvious modifications for $m=1$ and $m=n$). By virtue of simple estimates \begin{align*} &\left.\begin{array}{l} |x-x_k|\geq x_m,\;\;\;k=1,\dots,m\\ |x-x_k|\geq x_k/2,\;\;\;k=m+1,\dots,n\\ x\geq x_m \end{array}\right\}\;\;\;x\in[2x_m,x_{m+1/2}]\\ &\left.\begin{array}{l} |x-x_k|\leq x_m,\;\;\;k=1,\dots,m\\ |x-x_k|\leq x_k,\;\;\;k=m+1,\dots,n\\ x\leq x_m \end{array}\right\}\;\;\;x\in[0,x_m], \end{align*} from (3.5) we have $$ \int_{2x_m}^{x_{m+1}/2} w\,dx\leq 2^{s(n-m-1)} \int_0^{x_m} w\,dx. $$ Therefore $$ \int_0^{x_{m+1}}w\leq c\int_0^{x_{m+1}/2}w=c\int_0^{2x_m}w+c \int_{2x_m}^{x_{m+1}/2}w\leq(c^2+c 2^{s(n-m-1)}) \int_0^{x_m}w. \;\;\qed $$ \begin{lem} Let $n$, $a_k(x)$, $k=0,\dots,n$, $p$, $q$, $v(x)$ be as in Theorem $2.1$. Let a positive on $\Bbb{R}^+$ function $u(x)$ satisfy the condition $(2.1)$ and $$ \int_0^{2r}u^{-p'}(x)dx\leq D\int_0^r u^{-p'}(x)dx\;\;\;\forall r>0 $$ {\rm(}which is weaker than $(2.2))$. Assume the estimate \rom{(i)} with the constant $c<\infty$ from Theorem $2.1$ holds. Then \begin{gather} \ \hskip-1cm S_0\!=\!\sup_{r>0}\bigg(\int_r^\infty\!|a_0(x) v(x)|^q dx \bigg)^{1/q}\bigg(\int_0^r\! u^{-p'}(x)dx \bigg)^{1/p'}\leq \notag \\ \leq C(D,n,p)c\!<\!\infty. \end{gather} \end{lem} \begin{proof} Consider the extremal problem \begin{equation} \begin{cases} \dfrac{\int\limits_0^r f(t)dt}{\big(\int\limits_0^r|u(t)f(t)|^p dt \big)^{1/p}}\to \max\limits_f, \\[2mm] \displaystyle \int_0^r t^k f(t)dt=0,\;\;\;k=1,\dots,n. \end{cases} \end{equation} By substitution $f(t)=u^{-p'}(t)g(t)$, this problem reduces to a more convenient one: \begin{equation} \begin{cases} \dfrac{\int\limits_0^r g(t)u^{-p'}(t)dt}{\big(\int\limits_0^r|g(t)|^p u^{-p'}(t)dt\big)^{1/p}}\to\max\limits_g, \\ \displaystyle \int_0^r g(t)t^k u^{-p'}(t)dt=0,\;\;\;k=1\dots,n. \end{cases} \end{equation} Let $L_p(A,d\mu)$ denote the space of functions $f:A\to\Bbb{C}$ with the norm $(\int_A|f|^p d\mu)^{1/p}$. Let $X=L_{p'}([0,r],u^{-p'}(t)dt)$, $Y$ be a finite-dimensional subspace $X$ with the basis formed by the functions $t,t^2,\dots,t^n$. Note that $X^*\!=\!L_p([0,r], u^{-p'}(t)dt)$, the pairing between $f\in X$ and $g\in X^*$ being defined by the formula $$ \langle g,f\rangle=\int_0^r g(t)f(t)u^{-p'}(t)dt. $$ If we denote by $e$ the vector of the space $X$ representing the function $e(x)\equiv 1$, then the extremal problem (3.8) can be written in an abstract manner: $$ \begin{cases} \dfrac{\langle g,e\rangle}{\|g|X^*\|}\to \max\limits_{g\in X^*}, \\ g|_{{}_Y}=0. \end{cases} $$ By Lemma 3.1 (i) we have \begin{equation} \sup\bigg\{\frac{\langle g,e\rangle}{\|g|X^*\|}:g|_{{}_Y}=0\bigg\}= \operatorname{dist}(e,Y). \end{equation} Moreover, since the subspace $Y$ is finite-dimensional, $\operatorname{dist}(e,Y)$ is certainly attained at some vector $y_0\in Y$. Hence by Lemma 3.1 (ii), $\sup$ is attained in the right-hand side of (3.9), that is there exists $g_0\in X^*$, $g_0|_{{}_Y}=0$ such that \begin{equation} \frac{\langle g_0,e\rangle}{\|g_0|X^*\|}= \min_{c_1,\dots,c_n}\bigg(\int_0^r |1+c_1t+\cdots c_nt^n|^{p'} u^{-p'}dt\bigg)^{1/p'}. \end{equation} Assume $P(t)=1+c_1t+\cdots+c_nt^n$, $w(t)=u^{-p'}(t)$. The extremum conditions in the right-hand side of (3.10) have the form $$ \int_0^r|P(t)|^{p'-1}\operatorname{sign} P(t)t^k w(t)dt=0, \quad k=1,\dots,n. $$ Therefore it is clear that the extremal polynomial $\widetilde{P}(t)$ is a constant multiple of the polynomial $P_{n,r}(t)$ from Lemma 3.2 (with $s=p'$). According to that lemma (note that the condition (3.2) is fulfilled), all the roots of $\widetilde{P}(t)$ are located on $(0,r)$ and the following inequality holds: $$ \int_0^r u^{-p'}(t)dt\leq K\int_0^{x_1} u^{-p'}(t)dt \quad \forall r>0, $$ where $x_1$ is the smallest root and the constant $K$ does not depend on $r$. Using this fact, we derive from (3.10) the following estimate: \begin{gather*} \frac{\langle g_0,e\rangle}{\|g_0|X^*\|}= \bigg(\int_0^r |\widetilde{P}(t)|^{p'}u^{-p'}(t)dt\bigg)^{1/p'}= \\ =\bigg(\int_0^r \prod_{k=1}^n \Big|1-\frac{x}{x_k}\Big|^{p'} u^{-p'}(t)dt\bigg)^{1/p'}\geq \\ \geq \bigg(\int_0^{x_1} \Big(1-\frac{x}{x_k}\Big)^{np'} u^{-p'}(t)dt\bigg)^{1/p'}\geq \\ \geq \bigg(2^{-np'} \int_0^{x_1/2} u^{-p'}(t)dt\bigg)^{1/p'}\geq \frac{2^{-n}}{(DK)^{1/p'}} \bigg(\int_0^r u^{-p'}(t)dt\bigg)^{1/p'}. \end{gather*} Getting back to the original extremal problem (3.7), we see that for every $r>0$ there exists on $(0,r)$ a function $f_0(t)=u^{-p'}(t)g_0(t)$ such that \begin{gather*} \int_0^r f_0(t)dt\geq \widetilde{K}(D,n,p)\bigg(\int_0^r u^{-p'}(t)dt \bigg)^{1/p'}\bigg(\int_0^r |u(t)f_0(t)|^p dt\bigg)^{1/p}, \\ \int_0^r t^k f_0(t)\,dt=0, \quad k=1,\dots,n. \end{gather*} Extending $f_0$ to $[r,\infty)$ by zero and substituting the obtained function into the estimate (i) of Theorem 2.1, we get (3.6) with $C=1/\widetilde{K}$. \end{proof} \begin{lem} Let $1\leq p\leq q\leq \infty$, $u,v$ be non-negative on $\Bbb{R}^+$ functions. The estimate \begin{equation} \bigg\|\int_0^x f(t)\,dt|L_{q,v}\bigg\|\leq c\|f|L_{p,u}\| \end{equation} holds if and only if $$ S=\sup_{r>0}\bigg(\int_r^\infty v^q(x)\,dx\bigg)^{1/q} \bigg(\int_0^r u^{-p'}(x)\,dx\bigg)^{1/p'}<\infty. $$ Moreover, if $\widetilde{c}$ is the best constant in $(3.11)$, then $S\leq \widetilde{c}\leq (q')^{1/p'}q^{1/q}S$. If $p=1$ or $q=\infty$, then $\widetilde{c}=S$. \end{lem} This is the well-known criterion for the validity of the weighted Hardy's inequality obtained by J.~S.~Bradley \cite{2} and V.~M.~Kokilashvili \cite{3}. The proof can be found in \cite{13}, \S 1.3.1. \vskip+0.2cm \noindent {\it Proof of Theorem $2.1$.} Implication (ii)$\Rightarrow$(i) is obvious. Equivalence (ii)$\Leftrightarrow$(iii) follows from Lemma 3.4. To complete the proof we have to show that (i)$\Rightarrow$(iii). We act by induction in $n$. For $n=0$ the assertion of the theorem follows from Lemma 3.4. Assume that the theorem is already proved for $n=n_0$ and that the following estimate holds: \begin{equation} \bigg\|\sum_{k=0}^{n_0+1} a_k(x)\int_0^x t^k f(t)\,dt|L_{q,v}\bigg\| \leq c\|f|L_{p,u}\|, \end{equation} where the function $u$ satisfies the condition (2.2) with $n=n_0+1$. Lemma 3.3 implies that $S_0<\infty$. By Lemma 3.4 this is equivalent to the fact that the inequalities \begin{equation} \bigg\|a_0(x)\int_0^x f(t)\,dt|L_{q,v}\bigg\|\leq c'\|f|L_{p,u}\| \end{equation} are fulfilled. Inequalities (3.12) and (3.13) yield $$ \bigg\|\sum_{k=1}^{n_0+1} a_k(x)\int_0^x t^k f(t)\,dt|L_{q,v}\bigg\| \leq c''\|f|L_{p,u}\|, $$ which by substitution $\widetilde{f}(t)=tf(t)$ reduces to the form $$ \bigg\|\sum_{k=0}^{n_0} a_{k+1}(x)\int_0^x t^k \widetilde{f}(t)\,dt|L_{q,v}\bigg\| \leq c''\|\widetilde{f}|L_{p,\widetilde{u}}\|, $$ where $\widetilde{u}(x)=x^{-1}u(x)$. Since $\widetilde{u}$ satisfies the condition (2.2) with $n=n_0$, by assumption of the induction we obtain the finiteness of the remaining constants $S_k$, $k=1,\dots,n_0+1$. \;\;\qed \section{Proof of Theorem 2.2} \begin{lem} Let $[a,b]$ be a segment in $\Bbb{R}$. For any set of functions $h_1,\dots,h_l\in L_1([a,b])$ there exists a function $\sigma$ with $|\sigma(x)|=1$ on $[a,b]$ such that $$ \int_a^b h_k(x)\sigma(x)\,dx\-, \quad k=1,\dots,l. $$ \end{lem} The proof of this lemma can be found in \cite{14}, p. 267. \begin{lem}[{\cite{15}}] Let $l\in \Bbb{N}$, $w(x)$ be a non-negative on the segment $[a,b]$ function with $\int_a^b wdx<\infty$. There exists a function $g(x)$, $x\in \Bbb{R}$, such that \rom{(a)} $g(x)=0$ for $x\notin [a,b]$, \rom{(b)} $g,g',\dots,g^{(l-1)}$ are absolutely continuous on $\Bbb{R}$, \rom{(c)} $|g^{(l)}(x)|=w(x)$, $x\in [a,b]$. \end{lem} \begin{proof} By Lemma 4.1 there exists a function $\sigma$ with $|\sigma(x)|=1$ on $[a,b]$ such that $$ \int_a^b x^k w(x)\sigma(x)\,dx=0, \quad k=0,\dots,l-1. $$ Then $$ g(x)= \begin{cases} \displaystyle \frac{1}{(l-1)!}\int_a^x (x-t)^{l-1} w(t)\sigma(t)\,dt, \quad x\in [a,b], \\ 0, \quad x\notin [a,b], \end{cases} $$ is the required function. \end{proof} \vskip+0.2cm \noindent {\it Proof of Theorem $2.2$.} {\em Step $1$.} Let $T_f(x)=\sum\limits_{k=0}^{l-1} f^{(k)}(0) x^k/k!$ be the degree-($l-1$) Taylor polynomial of the function $f$. Then \begin{gather*} P(x,D)f(x)=P(x,D)\bigg( T_f(x)+\int_0^x \frac{(x-t)^{l-1}}{(l-1)!}\,f^{(l)}(t)\,dt\bigg)= \\ =P(x,D)T_f(x)+\sum_{m=0}^l b_m(x) D^m \int_0^x \frac{(x-t)^{l-1}}{(l-1)!}\,f^{(l)}(t)\,dt= \\ =P(x,D)T_f(x)+\sum_{m=0}^{l-1} b_m(x) \int_0^x \frac{(x-t)^{l-m-1}}{(l-m-1)!}\, f^{(l)}(t)\,dt+b_l(x)f^{(l)}(x)= \\ =P(x,D)T_f(x)+\sum_{m=0}^{l-1} \sum_{k=0}^{l-m-1} b_m(x)\, \frac{x^{l-m-1-k}(-1)^k}{(l-m-1-k)!k!} \int_0^x t^kf^{(l)}(t)\,dt+ \\ + b_l(x)f^{(l)}(x). \end{gather*} Changing in the last expression the order of summation with respect to $m$ and $k$, we arrive at the formula \begin{gather} P(x,D)f(x)= P(x,D)T_f(x)+ \notag \\ + \sum_{k=0}^{l-1} \frac{(-1)^k}{k!}\,d_k(x) \int_0^x t^k f^{(l)}(t)dt+b_l(x) f^{(l)}(x). \end{gather} {\em Step $2$ $($Sufficiency$)$.} From the identity (4.1) there follows the estimate \begin{gather} \|P(x,D)f(x)|L_{q,v}\| \leq c\sum_{k=0}^{l-1} |f^{(k)}(0)|\cdot \|P(x,D)x^k|L_{q,v}\|+ \notag \\ +c\sum_{k=0}^{l-1} \bigg\| d_k(x)\int_0^x t^k f^{(l)}(t)\,dt|L_{q,v}\bigg\|+ \|b_lf^{(l)}|L_{q,v}\|. \end{gather} If conditions (2.3) are fulfilled, then by Lemma 3.4 the second term in the right-hand side of (4.2) can be bounded by $c'\|f^{(l)}|L_{p,u}\|$. The third term for $p=q$ can be estimated in an obvious manner: $$ \|b_lf^{(l)}|L_{p,v}\|\leq \|b_l vu^{-1}|L_\infty(\Bbb{R}^+)\| \cdot \|f^{(l)}|L_{p,u}\|. $$ All these arguments imply the sufficiency of conditions in both parts of Theorem 2.2. {\em Step $3$ $($Necessity$)$.} The necessity of the condition (2.5) in item (ii) is obvious. The theorem will be proved if we show that from \begin{equation} P(x,D):\overset{\circ}{W}{}_{p,u}^{(l)}\to L_{q,v} \end{equation} there follows the fulfillment of (2.3) and (2.4). We start with (2.4). Consider the set $A_\alpha=\{x\in \Bbb{R}^+:|b_l vu^{-1}(x)|\geq \alpha\}$. Suppose $\operatorname{mes} A_\alpha>0$. Let $B$ be a bounded subset of $A$ of positive measure on which the functions $v(x)d_k(x)$ are bounded, say $B\subset [0,M]$, $|v(x)d_k(x)|\leq N$, $x\in B$, $k=0,\dots,l-1$. For every $\varepsilon>0$ there exists a segment $\Delta_\varepsilon\subset \Bbb{R}^+$ of length $\varepsilon$ such that $\operatorname{mes} \Delta_\varepsilon\cap B>0$. By Lemma 4.2, there exists an $l$ times differentiable function $g$ supported on $\Delta_\varepsilon$ and such that $$ |g^{(l)}(x)|= \begin{cases} u^{-p'}(x) & \text{if}\;\;x\in \Delta_\varepsilon\cap B, \\ 0 & \text{otherwise}. \end{cases} $$ The inequality $$ \|P(x,D)g|L_{q,v}\|\leq K\|g^{(l)}|L_{p,u}\| $$ and the identity (4.1) yield the estimate \begin{equation} \|b_lf^{(l)}|L_{q,v}\|\leq K\|g^{(l)}|L_{p,u}\|+ c\sum_{k=0}^{l-1}\bigg\|d_k(x) \int_0^x t^k g^{(l)}(t)\,dt |L_{q,v}\bigg\|. \end{equation} Using H\"{o}lder's inequality and taking into account the inclusion \linebreak $\Delta_\varepsilon\cap B\subset A_\alpha$, we can see that \begin{gather} \|b_lf^{(l)}|L_{q,v}\|=\bigg(\int_{\Delta_\varepsilon\cap B} |b_lu^{-p'}v|^qdx\bigg)^{1/q}\geq \notag \\ \geq (\operatorname{mes} \Delta_\varepsilon\cap B)^{1/q-1/p} \bigg(\int_{\Delta_\varepsilon\cap B} |b_lu^{-p'}v|^pdx\bigg)^{1/p}\geq \notag \\ \geq (\operatorname{mes} \Delta_\varepsilon\cap B)^{1/q-1/p} \alpha \bigg(\int_{\Delta_\varepsilon\cap B} u^{-p'}(x)\,dx\bigg)^{1/p}. \end{gather} Further, \begin{gather} \|g^{(l)}|L_{p,u}\|=\bigg(\int_{\Delta_\varepsilon\cap B} u^{-p'}(x)\,dx\bigg)^{1/p}, \\ \ \hskip-1cm \bigg\|d_k(x) \int_0^x\! t^k g^{(l)}(t)\,dt|L_{q,v}\bigg\|\!\leq \! NM^k\bigg(\int_{\Delta_\varepsilon\cap B} u^{-p'}(t)dt\bigg)\times \!(\operatorname{mes} \Delta_\varepsilon\cap B)^{1/q}. \end{gather} Substituting (4.5)--(4.7) into (4.4) we obtain \begin{gather} \alpha\leq (\operatorname{mes} \Delta_\varepsilon\cap B)^{1/p-1/q}\times \notag \\ \times \bigg(K+cNM^{l-1} \bigg(\int_{\Delta_\varepsilon\cap B} u^{-p'}(t)\,dt\bigg)^{1/p'} (\operatorname{mes} \Delta_\varepsilon\cap B)^{1/q} \bigg). \end{gather} We now pass to the limit $\varepsilon \to 0$. Then for $p<q$ Eq.~(4.8) implies that $\alpha=0$, while for $p=q$ it implies $\alpha\leq K$. Thus the proof of (2.4) is complete. Next, by virtue of the identity (4.1) and also from (4.3) and (2.4) it follows that the estimate $$ \bigg\|\sum_{k=0}^{l-1} \frac{(-1)^k}{k!}\,d_k(x) \int_0^x t^k f^{(l)}(t)\,dt|L_{q,v}\bigg\|\leq \|f^{(l)}|L_{p,u}\| $$ holds. Now, Theorem 2.1 implies (2.3). \;\;\qed \section*{Acknowledgement} The work is carried out under the financial support of Russian Fund of Fundamental Investigations, Grant RFFI--96-01-00243. {\bf Added June 2021:} The author is grateful for the hospitality to the Balabanovo bootcamp facility, where this work was carried out in the summer of 1996.

https://arxiv.org/abs/1009.4429

Representation formulas for $L^\infty$ norms of weakly convergent sequences of gradient fields in homogenization

We examine the composition of the $L^{\infty}$ norm with weakly convergent sequences of gradient fields associated with the homogenization of second order divergence form partial differential equations with measurable coefficients. Here the sequences of coefficients are chosen to model heterogeneous media and are piecewise constant and highly oscillatory. We identify local representation formulas that in the fine phase limit provide upper bounds on the limit superior of the $L^{\infty}$ norms of gradient fields. The local representation formulas are expressed in terms of the weak limit of the gradient fields and local corrector problems. The upper bounds may diverge according to the presence of rough interfaces. We also consider the fine phase limits for layered microstructures and for sufficiently smooth periodic microsturctures. For these cases we are able to provide explicit local formulas for the limit of the $L^\infty$ norms of the associated sequence of gradient fields. Local representation formulas for lower bounds are obtained for fields corresponding to continuously graded periodic microstructures as well as for general sequences of oscillatory coefficients. The representation formulas are applied to problems of optimal material design.

\section{Introduction} Understanding the composition of nonlinear functionals with weakly convergent sequences is a central issue in the direct methods of the calculus of variations, homogenization theory and nonlinear partial differential equations. In this paper we discuss a composition motivated by problems of optimal design. To fix ideas consider a domain $\Omega\subset\mathbb{R}^d$, $d=2,3$, partitioned into two measurable subsets $\omega$ and $\Omega/\omega$. Define the piecewise constant coefficient of thermal conductivity taking the values $\alpha I$ for $x\in\omega$ and $\beta I$ for $x\in\Omega/\omega$ by $A(\omega)=(\alpha\chi_\omega+\beta(1-\chi_\omega))I$. Here $\chi_\omega$ is the characteristic function of $\omega$ with $\chi_\omega=1$ for points in $\omega$ and zero otherwise and $I$ is the $d\times d$ identity matrix. Next consider a sequence of sets $\{\omega_n\}_{n=1}^\infty$ with indicator functions $\chi_{\scriptscriptstyle{\omega_n}}$ and the $H^1(\Omega)$ solutions $u_n$ of the boundary value problems $u_n=g$ on $\partial\Omega$ with $g\in H^{1/2}(\partial\Omega)$ and \begin{equation} -{\rm div}\left(A(\omega_n)\nabla u_n\right)=f \label{divequ} \end{equation} for $f\in H^{-1}(\Omega)$. The theory of homogenization \cite{Degeorgi}, \cite{Spagnolo}, \cite{Murattartar} asserts that there is a subsequence of sets, not relabeled, and a matrix valued coefficient $A^H(x)\in L^\infty(\Omega,\mathbb{R}^{d\times d})$ for which the sequence $u_{n}$ converges weakly in $H^1(\Omega)$ to $u^H\in H^1(\Omega)$ with $u^H=g$ for $x\in\partial\Omega$ and \begin{equation} -{\rm div}\left(A^H\nabla u^H\right)=f. \label{divequH} \end{equation} The compositions of interest are given by the $L^\infty$ norm taken over open subsets $S\subset\Omega$ and are of the form \begin{equation} \Vert\nabla u_{n}\Vert_{L^\infty(S)}={\rm esssup}_{x\in S}|\nabla u_{n}(x)|, \label{morminfty} \end{equation} \begin{eqnarray} \Vert\chi_{\scriptscriptstyle{\omega_n}}\nabla u_{n}\Vert_{L^\infty(S)}&\hbox{ and }&\Vert(1-\chi_{\scriptscriptstyle{\omega_n}})\nabla u_{n}\Vert_{L^\infty(S)}, \label{morminftyphase} \end{eqnarray} and we seek to understand the behavior of limits of the kind given by \begin{eqnarray}\label{lims} \liminf_{n\rightarrow\infty}\Vert\chi_{\scriptscriptstyle{\omega_n}}\nabla u_{n}\Vert_{L^\infty(S)}&\hbox{ and }& \limsup_{n\rightarrow\infty}\Vert\chi_{\scriptscriptstyle{\omega_n}}\nabla u_{n}\Vert_{L^\infty(S)}. \end{eqnarray} In this paper we provide examples and identify conditions for which it is possible to represent the limits of these compositions by local formulas expressed in terms of the weak limit $\nabla u^H$. The representation formulas provide a multi-scale description useful for studying the composition. To illustrate the ideas we display local formulas in the context of periodic homogenization. The unit period cell for the microstructure is $Y$ and we partition it into two sets $P$ and $Y/P$. To fix ideas we assume the set $P$ represents a single smooth particle, e.g. an ellipsoid. The union of all particles taken over all periods is denoted by $\omega$. The coefficient $A(\omega)$ is a periodic simple function defined on $\mathbb{R}^d$ taking the value $\alpha I$ in $\omega$ and $\beta I$ in $\mathbb{R}^d/\omega$. On rescaling by $1/n$, $n=1,2,\ldots$ the set given by the union of rescaled particles taken over all rescaled periods is denoted by $\omega_n$ and $\chi_{\omega_n}(x)=\chi_\omega(nx)$. We consider the sequence of coefficients $A(\omega_n)$ restricted to $\Omega$ and the theory of periodic homogenization \cite{BLP}, \cite{SanchezPalencia} delivers a constant matrix $A^{H}$ of effective properties given by the formula \begin{equation} A^{H}_{ij} = \int_{Y}A_{ik}(y)P_{kj}(y)dy \end{equation} where $P_{kj} = \partial _{x_{k}}\phi^{j}(y) + \delta_{kj}$ and $\phi ^{j}$ are $Y$-periodic $H^{1}_{loc}(R^{d})$ solutions of the unit cell problems \begin{equation \text{div}(A(y)(\nabla \phi^{j}(y) + {\bf e}^{j} ))= 0 \quad \text{in $Y$}, \end{equation} where we have written $A(y)=A(\omega)=(\alpha\chi_{\omega}(y)+(1-\chi_\omega(y))\beta)I$ for $y\in Y$. It is well known that the associated energies taken over sets $S\subset\subset\Omega$ converge \cite{Spagnolo}, \cite{Murattartar}, i.e., \begin{eqnarray} \lim_{n\rightarrow\infty}\int_S A^n\nabla u_n\cdot\nabla u_n dx&=&\int_S A^H\nabla u^H\cdot\nabla u^H dx\nonumber\\ &=&\int_{S\times Y} A(y)P(y)\nabla u^H(x)\cdot\nabla u^H(x)\,dydx. \label{energies} \end{eqnarray} In this paper we show that the analogous formulas hold for $L^\infty$ norms and are given by the local representation formulas \begin{eqnarray} \lim_{n\rightarrow\infty}\Vert\chi_{\omega_n}(x)\nabla u_n\Vert_{L^\infty(S)}&=&\Vert\chi_\omega(y)P(y)\nabla u^H(x)\Vert_{L^\infty(S\times Y)},\hbox{ and } \label{peridentinff1}\\ \lim_{n\rightarrow\infty}\Vert(1-\chi_{\omega_n}(x))\nabla u_n\Vert_{L^\infty(S)}&=&\Vert(1-\chi_\omega(y))P(y)\nabla u^H(x)\Vert_{L^\infty(S\times Y)}, \label{peridentinff2} \end{eqnarray} these formulas follow from Theorem \ref{periodicidentity}. For general situations the question of finding local formulas is delicate as the solutions of \eqref{divequ} with measurable coefficients are nominally in $H^1(\Omega)$ with gradients in $L^2(\Omega,\mathbb{R}^d)$. For sufficiently regular $f$, $g$, and $\Omega$, and in the absence of any other hypothesis on the coefficients, the theorems of Bojarski \cite{Bojarski}, for problems in $\mathbb{R}^2$, and Meyers \cite{Meyers}, for problems in $\mathbb{R}^d$, $d\geq 2$, guarantee that gradients belong to $L^{p}(\Omega,\mathbb{R}^d)$ for $2<p<p'$ with $p'$ depending on the aspect ratio $\beta/\alpha$. For the general case one can not expect $p$ to be too large. The recent work of Faraco \cite{Faraco} shows that for $d=2$ and for $\beta=K>1$ and $\alpha=1/K$ that there exist coefficients associated with sequences of layered configurations $\omega_n$ made up of hierarchal laminations for which the sequence of gradients is bounded in $L^p_{loc}(\Omega,\mathbb{R}^d)$ for $p<p*=2K/(K-1)$ and is {\em divergent} in $L^{p}_{loc}(\Omega,\mathbb{R}^d)$ for $p\geq p* $. This precise value for $p*$ was proposed earlier for sequences of laminated structures using physical arguments in the work of Milton \cite{Milton}. For measurable matrix valued coefficients $A(x)\in \mathbb{R}^{2\times 2}$ with eigenvalues in the interval $[1/K,K]$ the same critical exponent $p*=2K/(K-1)$ holds, this result also motivated by \cite{Milton} is shown earlier in the work of Lionetti and Nesi \cite{LionettiNesi}. With these general results in mind we display, in section 2, a set of upper bounds on the limit superior of the compositions \eqref{morminftyphase} that hold with a minimal set of hypothesis on the sequence $\{\omega_{n}\}_{n=1}^\infty$. Here we assume only that the sets $\omega_{n}$ are Lebesgue measurable thus the upper bound may diverge to $\infty$ for cases when these sets have corners or cusps. The upper bound is given by a local representation formula expressed in terms of the weak limit $\nabla u^H$. It is given by the limit superior of a sequence of $L^\infty$ norms of local corrector problems driven by $\nabla u^H$. For periodic microstructures the local correctors reduce to the well known solutions of the periodic cell problems associated with periodic homogenization \cite{BLP}, \cite{SanchezPalencia}. In section 3 we provide a general set of sufficient conditions for which the limits \eqref{lims} agree and are given by a local representation formula see, Theorem \ref{Equality}. As before this formula is given in terms of a limit of a sequence of $L^\infty$ norms for solutions of local corrector problems driven by $\nabla u^H$. From a physical perspective the local formula measures the amplification or diminution of the gradient $\nabla u^H$ by the local microstructure. Formulas of this type have been developed earlier in the context of upper and lower bounds for the linear case \cite{liproy}, \cite{lipjmps}, \cite{lipsima} and lower bounds for the nonlinear case \cite{jimenezlipton}. On the other hand when the boundary of the sets $\omega$ are sufficiently regular one easily constructs examples of coefficients $A(\omega)$ for which the gradients belong to $L^\infty(\Omega,\mathbb{R}^d)$. More systematic treatments developed in the work of Bonnitier and Vogelius \cite{bonnitierVogelius}, Li and Vogelius \cite{LiVogelius}, and Li and Nirenberg \cite{LiNirenberg} describe generic classes of coefficients $A(\omega)$ for which gradients of solutions belong to $L^\infty_{loc}(\Omega,\mathbb{R}^d)$. The earlier work of Chipot, Kinderlehrer and Vergara-Caffarelli \cite{Chipot} establish higher regularity for coefficients $A(\omega)$ associated with laminated configurations. In section \ref{41} we apply the uniform convergence for simple laminates discovered in \cite{Chipot} to show that the sufficient conditions given by Theorem \ref{Equality} hold. We apply this observation to obtain an explicit local formula for the limits of compositions of the $L^{\infty}$ norm with weakly convergent sequences of gradient fields associated with layered microstructures. While in section \ref{42} we use the higher regularity theory for smooth periodic microstructures developed in \cite{LiNirenberg} to recover an explicit representation formula for the upper bound on the limit superior of compositions of the $L^{\infty}$ norm with weakly convergent sequences of gradient fields associated with periodic microstructures. Lower bounds on the limit inferior are developed in section 5 that agree with the upper bounds and we recover explicit local formulas for the limits of compositions of the $L^{\infty}$ norm with weakly convergent sequences of gradient fields associated with periodic microstructures. The $L^\infty$ norm of the field gradient inside each component material \eqref{morminftyphase} is of interest in applications where it is used to describe the strength of a composite structure. Here the strength of a component material is described by a threshold value of the $L^\infty$ norm of the gradient. If the $L^\infty$ norm exceeds the threshold inside $\omega_n$ then failure is initiated in that material and nonlinear phenomena such as plasticity and material degradation occur \cite{kellymac}, \cite{NuismerWhitney}. The design of composite structures to forestall eventual failure initiation is of central interest for aerospace applications \cite{gosschristensen}. For a given set of structural loads one seeks configurations $\omega$ that keep the local gradient field below the failure threshold inside each component material over as much of the structure as possible. As is usual in design problems of this sort the problem is most often ill-posed (see, e.g. \cite{lipnato}) and there is no {\em best} configuration $\omega$. Instead one looks to identify sequences of configurations $\{\omega_n\}_{n=1}^\infty$ from which a {\em nearly} optimal configuration can be chosen. The work of Duysinx and Bendsoe \cite{DesynxBendsoe} presents an insightful engineering approach to the problem of optimal design subject to constraints on the sup norm of the local stress inside a laminated material. The subsequent work of Lipton and Stuebner \cite{Liptstueb1}, \cite{Liptstueb3}, \cite{LiptstuebAIAA} develops the mathematical theory and provides numerical schemes for the design of continuously graded multi-phase elastic composites with constraints on the $L^\infty$ norm of the local stress or strain inside each material. More recent work by Carlos-Bellido, Donoso and Pedregal \cite{donsopedregal} provides the mathematical relaxation of the $L^\infty$ gradient constrained design problem for two-phase heat conducting materials. The feature common to all of these problems is that they involve weakly convergent sequences of gradients and their composition with $L^\infty$ norms of the type given by \eqref{morminfty} and \eqref{morminftyphase}. Motivated by the applications we develop an explicit local representation formula for the lower bound on \eqref{lims} for continuously graded periodic microstructures introduced for optimal design problems in \cite{liproy}, \cite{Liptstueb1}, \cite{Liptstueb3}, see section 5. A similar set of lower bounds have appeared earlier within the context of two-scale homogenization \cite{lipsima}. In section 6 we conclude the paper by outlining the connection between optimal design problems with $L^\infty$ gradient constraints, local representation formulas, and the composition of the $L^\infty$ norm with sequences of gradients. Last it is pointed out that the results presented here can be extended without modification to the system of linear elasticity. \setcounter{theorem}{0} \setcounter{definition}{0} \setcounter{lemma}{0} \setcounter{conjecture}{0} \setcounter{corollary}{0} \setcounter{equation}{0} \section{Mathematical background and upper bounds given by local representation formulas} In this section we present upper bounds on the limit superior of sequences of $L^\infty$ norms of gradient fields associated with G-convergent sequences of coefficient matrices. In what follows the coefficient matrices given by simple functions $A(x)$ taking the finite set of values $A_1,A_2,\ldots, A_N$ in the space of $d\times d$ positive definite symmetric matrices. Here no assumption on the sets $\omega_i$ where $A(x)=A_i$ are made other than that they are Lebesgue measurable subsets of $\Omega$. We consider a sequence of coefficient matrices $A^n(x)=\sum_{i=1}^N\chi^i_n A_i$. Here $A^n(x)=A_i$ on the sets $\omega^i_n$ and the corresponding indicator function $\chi^i_n$ takes the value $\chi^i_n=1$ on $\omega^i_n$ and zero outside for $i=1,2,\dots,N$, with $\sum_{i=1}^N\chi^i_n=1$ on $\Omega$. We suppose that the sequence $\{A^n(x)\}_{n=1}^\infty$ is G-convergent with a G-limit given by the positive definite $d\times d$ coefficient matrix $A^H(x)$. The G-limit is often referred to as the homogenized coefficient matrix. For completeness we recall the definition of G-convergence as presented in \cite{Murattartar}: \begin{definition} \label{def1} The sequence of matrices $\{A^n(x)\}_{n=1}^\infty$ is said to G-converge to $A^H(x)$ iff for every $\omega\subset\Omega$ with closure also contained in $\Omega$ and for every $f\in H^{-1}(\omega)$ the solutions $\varphi_n\in H^{1}_0(\omega)$ of \begin{equation} -{\rm div}\left(A^n\nabla \varphi_n\right)=f \label{sequence} \end{equation} converge weakly in $H^1_0(\omega)$ to the $H^1_0(\omega)$ solution $\varphi^H$ of \begin{equation} -{\rm div}\left(A^H\nabla \varphi^H\right)=f. \label{hlimit} \end{equation} \end{definition} G-convergence is a form of convergence for solution operators and its relation to other notions of operator convergence are provided in \cite{Spagnolo}. From a physical perspective each choice of right hand side $f$ in \eqref{sequence} can be thought of as an experiment with the physical response given by the solution $\varphi_n$ of \eqref{sequence}. The physical response of heterogeneous materials with coefficients belonging to a G-convergent sequence converge in $H_0^1(\omega)$ to that of the G-limit for every choice of sub-domain $\omega$. For sequences of oscillatory periodic and strictly stationary, ergodic random coefficients the G-convergence is described by the more well known notions of homogenization theory \cite{BLP}, \cite{ZOK}, \cite{PapVaradan}, \cite{Spagnolo}, \cite{Murattartar}. We point out that the G-convergence described in Definition \ref{def1} is a specialization of the notion of H-convergence introduced in \cite{Murattartar} which applies to sequences of non-symmetric coefficient matrices subject to suitable coercivity and boundedness conditions. It is known \cite{Murattartar} that if $\{A^n\}_{n=1}^\infty$ G-converges to $A^H$, then for any $g\in H^{1/2}(\partial\Omega)$ and $f\in H^{-1}(\Omega)$, the $H^1(\Omega)$ solutions $u_n$ of \begin{equation} -{\rm div}\left(A^n\nabla u_n\right)=f,\quad\quad u_{n} = g \label{sequenceu} \end{equation} converge weakly in $H^1(\Omega)$ to the $H^1(\Omega)$ solution $u^H$ of \begin{equation} -{\rm div}\left(A^H\nabla u^H\right)=f, \quad\quad u^{H} = g. \end{equation} Last we recall the sequential compactness property of G-convergence \cite{Spagnolo}, \cite{Murattartar} applied to the case at hand. \begin{theorem} \label{compactness} Given any sequence of simple matrix valued functions $\{A^n(x)\}_{n=1}^\infty$ there exists a subsequence $\{A^{n'}(x)\}_{n'=1}^\infty$ and a positive definite $d\times d$ matrix valued function $A^H(x)$ such that the sequence $\{A^{n'}(x)\}_{n'=1}^\infty$ G-converges to $A^H(x)$. \end{theorem} For the remainder of the paper we will suppose that sequence of coefficients $\{A^{n}\}_{n=1}^\infty$ G-converges to $A^H$ and we will investigate the behavior of the gradient fields inside each of the sets $\omega^i_n$. To this end we will consider the limits \begin{eqnarray} \liminf_{n\rightarrow\infty}\Vert\chi^i_n\nabla u_{n}\Vert_{L^\infty(S)} &\hbox{ and }& \limsup_{n\rightarrow\infty}\Vert\chi^i_n\nabla u_{n}\Vert_{L^\infty(S)},\hbox{ for $i=1,2\dots,N$,} \end{eqnarray} where $S\subset\Omega$ is an open set of interest with closure contained inside $\Omega$. In order to proceed we introduce the local corrector functions associated with the sequence of coefficients $\{A^{n}\}_{n=1}^\infty$. Let $Y\subset \mathbb{R}^d$ be the unit cube centered at the origin. For $r>0$ consider $\Omega_r^{int}=\{x\in\Omega:\,dist(x,\partial\Omega)>r\}$ and for $x\in\Omega_r^{int}$ and $z\in Y$ we introduce the $Y$ periodic $H^1(Y)$ solution $w^{r,n}(x,z)$ of \begin{eqnarray} -{\rm div}_z\left(A^{n}(x+rz)(\nabla_z w^{r,n}_{\overline{e}}(x,z)+\overline{e}) \right)=0, \hbox{for $z\in Y$}, \label{correct} \end{eqnarray} where $\overline{E}$ is a constant vector in $\mathbb{R}^d$ with respect to the $z$ variable. Here $x$ appears as a parameter and the differential operators with respect to the $z$ variable are indicated by subscripts. For future reference we note that $w^{r,n}$ depends linearly on $\overline{e}$ and we define the corrector matrix $P^{r,n}(x,z)$ to be given by \begin{eqnarray} P^{r,n}(x,z)\overline{e}=\nabla_z w^{r,n}_{\overline{e}}(x,z)+\overline{e}.\label{corrector} \end{eqnarray} We are interested in the $L^\infty$ norm associated with each phase and introduce the modulation functions ${\mathcal M}^i(\nabla u^H)$ defined for $x\in\Omega$ given by \cite{liproy} \begin{eqnarray} \label{FieldModulation} {\mathcal M}^i(\nabla u^H)(x)=\limsup_{r\rightarrow 0}\limsup_{n\rightarrow\infty}\Vert\chi^i_{n}(x+rz)(P^{r,n}(x,z)\nabla u^H(x))\Vert_{L^\infty(Y)}.\label{f} \end{eqnarray} In what follows we will denote the measure of $\omega\subset\Omega$ by $|\omega|$ and state the following upper bound given by a local representation formula. \begin{theorem} \label{upperboundd1} Let $A^{n}$ G-converge to $A^H$ and consider any open set $S\subset\Omega$ with closure contained inside $\Omega$. There exists a subsequence, not relabeled and a sequence of decreasing measurable sets $E_n\subset S$, with $|E_n|\searrow 0$ such that \begin{eqnarray} \limsup_{n\rightarrow\infty}\Vert\chi^i_{n}\nabla u_n\Vert_{L^\infty(S\setminus E_n)}\leq\Vert {\mathcal M}^i(\nabla u^H)\Vert_{L^\infty(S)},\hbox{ $i=1,2,\ldots,N$.}\label{ubound1} \end{eqnarray} \end{theorem} To proceed we introduce the distribution functions associated with the following sets $S^n_{i,t}$, $i=1,2,\ldots,N$, defined by \begin{eqnarray} S^n_{i,t}=\{x\in S:\,\chi^i_{n}|\nabla u_n|>t\}\label{seti} \end{eqnarray} given by \begin{eqnarray} \lambda_i^n(t)=|S_{i,t}^n|.\label{dist} \end{eqnarray} We state a second upper bound that follows from the homogenization constraint \cite{liproy}. \begin{theorem} \label{upperboundd2} Let $A^{n}$ G-converge to $A^H$ and consider any open set $S\subset\Omega$ with closure contained inside $\Omega$. Suppose for $i=1,2,\ldots,N$ that $\limsup_{n\rightarrow\infty}\Vert\chi^i_{n}\nabla u_n\Vert_{L^\infty(S)}=\ell^i<\infty$ and for every $\delta>0$ sufficiently small there exist positive numbers $\theta^i_\delta>0$ such that \begin{eqnarray} \liminf_{n\rightarrow\infty}\lambda^n_{i}(\ell^i-\delta)>\theta^i_\delta.\label{nonzero} \end{eqnarray} There exists a subsequence, not relabeled, such that \begin{eqnarray} \limsup_{n\rightarrow\infty}\Vert\chi^i_{n}\nabla u_n\Vert_{L^\infty(S)}\leq\Vert {\mathcal M}^i(\nabla u^H)\Vert_{L^\infty(S)}.\label{upbound1} \end{eqnarray} \end{theorem} We provide a proof Theorem \ref{upperboundd1} noting that the proof of Theorem \ref{upperboundd2} is given in \cite{liproy}. \begin{proof} First note that the claim holds trivially if $\Vert {\mathcal M}^i(\nabla u^H)\Vert_{L^\infty(S)}=\infty$. Now suppose otherwise and set $\Vert {\mathcal M}^i(\nabla u^H)\Vert_{L^\infty(S)}=H<\infty$. For this case Corollary 3.3 of \cite{liproy} shows directly that for any $\delta>0$ that the measure of the sets \begin{eqnarray} S^n_{i,H+\delta}=\{x\in S:\,\chi^i_n(x)|\nabla u_n(x)|>H+\delta\}, \label{si} \end{eqnarray} tends to zero as $n$ goes to $\infty$, i.e., \begin{eqnarray} \limsup_{n\rightarrow\infty}\lambda_i^n(H+\delta)=\limsup_{n\rightarrow\infty}|S^n_{i,H+\delta}|= 0. \label{decreasing} \end{eqnarray} We choose a sequence of decreasing positive numbers $\{\delta_\ell\}_{\ell=1}^\infty$, such that $\delta_\ell\searrow 0$ and from \eqref{decreasing} we can pick a subsequence of coefficients $\{A^{n_j(\delta_1)}\}_{j=1}^\infty$ for which \begin{eqnarray} |S^{n_j(\delta_1)}_{i,H+\delta_1}|<2^{-j},\hbox{ $j=1,2,\ldots$}. \label{21} \end{eqnarray} For $\delta_2$ we appeal again to \eqref{decreasing} and pick out a subsequence of $\{A^{n_j(\delta_1)}\}_{j=1}^\infty$ denoted by $\{A^{n_j(\delta_2)}\}_{j=1}^\infty$ for which \begin{eqnarray} |S^{n_j(\delta_2)}_{i,H+\delta_2}|<2^{-j},\hbox{ $j=1,2,\ldots$}. \label{22} \end{eqnarray} We repeat this process for each $\delta_\ell$ to obtain a family of subsequences $\{A^{n_j(\delta_\ell)}\}_{j=1}^\infty$, $\ell=1,2,\ldots$ such that $\{A^{n_j(\delta_{\ell+1})}\}_{j=1}^\infty\subset\{A^{n_j(\delta_{\ell})}\}_{j=1}^\infty$. On choosing the diagonal sequence $\{A^{n_k(\delta_{k})}\}_{k=1}^\infty$ we form the sets \begin{eqnarray} E_K=\cup_{k\geq K}S_{i,H+\delta_k}^{n_k(\delta_k)}=\{x\in S:\,\chi^i_{n_k(\delta_k)}|\nabla u_{n_k(\delta_k)}|>H+\delta_k,\hbox{ for some $k\geq K$}\}, \label{diag} \end{eqnarray} with $E_{K+1}\subset E_K$. Noting that $|S_{i,H+\delta_k}^{n_k(\delta_k)}|<2^{-k}$, we see that $|E_K|<2^{-K+1}$. Since $x\not\in E_K$ implies that \begin{eqnarray} \chi^i_{n_k(\delta_k)}|\nabla u_{n_k(\delta_k)}|<H+\delta_k\hbox{ for all } k\geq K, \label{notinE} \end{eqnarray} we observe that \begin{eqnarray} \Vert\chi^i_{n_k(\delta_k)}\nabla u_{n_k(\delta_k)}\Vert_{L^\infty(S\setminus E_K)}<H+\delta_k\hbox{ for all } k\geq K, \label{linfbound} \end{eqnarray} and we conclude that \begin{eqnarray} \limsup_{K\rightarrow\infty}\Vert\chi^i_{n_K(\delta_K)}\nabla u_{n_K(\delta_k)}\Vert_{L^\infty(S\setminus E_K)}\leq H, \label{linflimbound} \end{eqnarray} with $|E_K|\searrow 0$ and the theorem is proved. \end{proof} \setcounter{theorem}{0} \setcounter{definition}{0} \setcounter{lemma}{0} \setcounter{conjecture}{0} \setcounter{corollary}{0} \setcounter{equation}{0} \section{Lower bounds and sufficient conditions for a local representation formula} We suppose that sequence of coefficients $\{A^{n})\}_{n=1}^\infty$ G-converges to $A^H$ and investigate the behavior of the gradient fields inside each of the sets $\omega^i_n$. Here we consider the limits \begin{eqnarray}\label{limsi} \liminf_{n\rightarrow\infty}\Vert\chi^i_n\nabla u_{n}\Vert_{L^\infty(S)},\hbox{ for $i=1,2\ldots,N$.} \end{eqnarray} and identify a general sufficient condition for obtaining a lower bound on these quantities in terms of ${\mathcal M}^i(\nabla u^H)$. Assume that $u_{n}$, $P^{r,n}$ and $u^{H}$ are defined as in the previous section and we consider an open subset $S\subset\Omega$ with closure contained in $\Omega$. We write $\tau=dist(\partial S,\partial\Omega)>0$ and set \[ S_{\tau} = \{ x\in \Omega: dist(x,S)<\tau\}. \] For $r<\tau$ note that $ S\subset S_{r}\subset S_\tau\subset\Omega$. We next recall for $x\in S$ \begin{eqnarray} {\mathcal M}^i(\nabla u^H)(x)= \limsup_{r \to 0}\limsup_{n \to \infty}\||\chi^i_n(x+ry)P^{r,n}(x,y)\nabla u^{H}(x)|^2\|_{L^{\infty}(Y)}. \label{hyplowerlocal} \end{eqnarray} For this case the sufficient condition is based on the distribution function for the sequence $\{\chi^i_n(x+ry)P^{r,n}(x,y)\nabla u^{H}(x)\}$ and the lower bound is presented in the following theorem. \begin{theorem} \label{lowerbound1} Let $A^{n}$ G-converge to $A^H$ and consider any open set $S\subset\Omega$ with closure contained inside $\Omega$. Suppose that \[\|{\mathcal M}^i(\nabla u^H)\|_{L^\infty(S)}=\ell^i<\infty. \] Assume also that for all $\delta > 0$ small, there exist $\beta_{\delta}>0$ such that \begin{equation}\label{assum} \lim_{r\to 0}\liminf_{n \to \infty}|\{(x,y)\in S\times Y:\,|\chi_n^i(x+ry)P^{r,n}(x,y)\nabla u^{H}(x)|^2> (\ell^i)^2 - \delta \}|\geq \beta_{\delta} > 0. \end{equation} Then there exists a subsequence for which \[ \lim_{r \to 0}\liminf_{n \to \infty}\|\chi^i_n\nabla u_n\|_{L^{\infty}(S_r)} \geq \|{\mathcal M}^i(\nabla u^H)\|_{L^{\infty}(S)}. \] \end{theorem} \begin{proof} Our starting point is Lemma 5.5 of \cite{casadodiazcalvogomez} which is described in the following lemma. \begin{lemma} \begin{eqnarray} \lim_{r\to 0}\limsup_{n \to \infty}\int_{S}\int_{Y}|P^{r,n}(x,y)\nabla u^H(x)-\nabla u_n(x+ry)| ^{2}dydx=0. \label{l2convergence} \end{eqnarray} \end{lemma} On applying the lemma we observe that \begin{equation} \chi^i_n(x+ry)P^{r, n}(x,y)\nabla u^{H}(x)=\chi^i_n(x+ry)\nabla u_n(x+ry) + z^{r, n}(x,y)\quad \forall (x,y)\in S\times Y, \end{equation} where \begin{equation}\label{deco1} \lim_{r\to 0}\limsup_{n \to \infty}\int_{S}\int_{Y}|z^{r,n}(x,y)| ^{2}dydx = 0. \end{equation} It follows that \begin{equation}\label{deco2} |\chi_n^i(x+ry)P^{r, n}(x,y)\nabla u^{H}(x)|^{2}=|\chi_n^i(x+ry)\nabla u_n(x+ry)|^{2} + F^{r,n}(x,y)\quad \forall (x,y)\in S\times Y, \end{equation} where \[ F^{r,n}(x,y) = |z^{r, n}(x,y)|^{2} + (z^{r, n}(x,y),\chi_n^i(x+ry)\nabla u_n(x+ry)). \] We show that $F^{r,n}(x,y) \to 0$ strongly in $L^{1}$ in the sense that \[ \lim_{r\to 0}\limsup_{n \to \infty}\int_{S}\int_{ Y}| F^{r,n}(x,y)|dydx = 0. \] Indeed, from the definition and by Cauchy-Schwarz inequality we have \begin{equation}\label{small1} \begin{split} \int_{S}\int_{ Y}| F^{r,n}(x,y)|dydx&\leq \int_{S_{r}}\int_{Y}|z^{r, n}(x,y)|^{2}dydx \\ &+ \left(\int_{S}\int_{ Y}|z^{r, n}(x,y)|^{2}dydx\right)^{1/2}\left(\int_{S}\int_{ Y}|\nabla u^{n}(x+ry))|^{2}dydx\right)^{1/2} \end{split} \end{equation} Moreover from standard a-priori estimates we know there is a constant $C>0$ independent of $r$ and $n$ for which, \begin{equation}\label{standard1} \int_{S}\int_{ Y}|\nabla u^{n}(x+ry))|^{2}dydx \leq C. \end{equation} The assertion follows on taking the limits in (\ref{small1}) and using estimate (\ref{standard1}) and equation (\ref{deco1}). Now by Chebyshev's inequality, for every $\delta > 0$, we have the inequality \[ |\{ (x,y) \in S\times Y: |F^{r,n} (x, y)| > \delta\}|\leq \frac{1}{\delta }\int_{S\times Y}|F^{r,n}(x,y)|dydx \] and taking the limsup as $n\to \infty$ first and then as $r\to 0$, we see that \begin{equation}\label{small2} \lim_{r\to 0}\limsup_{n \to \infty}|\{ (x,y) \in S\times Y: |F^{r,n} (x, y)| > \delta\}| = 0 \end{equation} From (\ref{deco2}) we see that \[ \begin{split} &\{(x,y)\in S\times Y:|\chi_n^i(x+ry)P^{r,n}(x,y)\nabla u^{H}(x)|^{2}> (\ell^i)^2 - \delta \}\subset\\ &\subset \{(x, y)\in S\times Y: |\chi_n^i(x+ry)\nabla u_n(x+ry)|^{2}> (\ell^i)^2-2\delta\}\cup\{ (x, y)\in S\times Y: |F^{r,n}(x,y)|>\delta\}. \end{split} \] Therefore, applying (\ref{small2}) we obtain \[ \begin{split} &\lim_{r\to 0}\liminf_{n \to \infty}|\{(x,y)\in S\times Y:|\chi_n^i(x+ry)P^{r,n}(x,y)\nabla u^{H}(x)|^{2}> (\ell^i)^2 - \delta \}|\leq\\ &\leq \lim_{r\to 0}\liminf_{n \to \infty}|\{(x, y)\in S\times Y: |\chi_n^i(x+ry)\nabla u^{n}(x+ry)|^{2}> (\ell^i)^2-2\delta\}|. \end{split} \] It follows from the last inequality that \[ \lim_{r\to 0}\liminf_{n \to \infty}|\{(x,y)\in S\times Y: |\chi_n^i(x+ry)\nabla u^{n}(x+ry)|^{2} > (\ell^i)^2-2\delta\}|\geq \beta_{\delta} > 0, \] Here we have used our assumption ( \ref{assum}). Therefore, there exist $R = R(\delta)$ and $N = N(\delta)$ such that \[ |\{(x, y)\in S\times Y: |\chi_n^i(x+ry)\nabla u^{n}(x+ry)|^{2} > (\ell^i)^2-2\delta\}| >0, \quad \forall n\geq N(\delta), r\leq R(\delta). \] From the definition of the $ L^{\infty}$ norm it follows that, \[ \| |\chi_n^i\nabla u^{n}|^{2}\|_{L^{\infty}(S_r)}\geq (\ell^i)^2 - 2\delta\quad \forall n\geq N(\delta), r\leq R(\delta). \] Taking the limit first in $n$ and then in $r$, and using the arbitrariness of of $\delta$, we get \[ \lim_{r \to 0}\liminf_{n \to \infty}\|\chi_n^i\nabla u^{n}\|^2_{L^{\infty}(S_r)}\geq (\ell^i)^2, \] and the theorem follows. \end{proof} Last if we combine the hypotheses of theorems \ref{upperboundd2} and \ref{lowerbound1} we obtain a sufficient condition for a local representation formula for limits of compositions of the $L^\infty$ norm with weakly convergent sequences of gradients associated with homogenization. \begin{theorem} \label{Equality} Let $A^{n}$ G-converge to $A^H$ and consider any open set $S\subset\Omega$ with closure contained inside $\Omega$. Suppose for sufficiently small $r<\tau$, $ S\subset S_{2r}\subset S_\tau\subset\Omega$, for $i=1,2,\ldots,N$ that $\limsup_{r \to 0}\limsup_{n\rightarrow\infty}\Vert\chi^i_{n}\nabla u_n\Vert_{L^\infty(S_r)}=\ell^i<\infty$ and for every $\delta>0$ sufficiently small there exist positive numbers $\theta^i_\delta>0$ such that \begin{eqnarray} \label{nonzero2-1} \limsup_{r \to 0}\limsup_{n\rightarrow\infty}|\{x\in S_r:\,\chi_n^i|\nabla u_n|>\ell^i-\delta)\}|\geq\theta^i_\delta>0, \end{eqnarray} in addition suppose that $\limsup_{r \to 0}\|{\mathcal M}^i(\nabla u^H)\|_{L^\infty(S_r)}=\tilde{\ell}^i<\infty$ and for all $\delta > 0$ small, there exist $\beta_{\delta}>0$ such that \begin{equation}\label{assum2} \lim_{r\to 0}\liminf_{n \to \infty}|\{(x,y)\in S_r\times Y:\,|\chi_n^i(x+ry)P^{r,n}(x,y)\nabla u^{H}(x)|^2> (\tilde{\ell}^i)^2 - \delta \}|\geq \beta_{\delta} > 0. \end{equation} There exists a subsequence, not relabeled, such that \begin{eqnarray} \lim_{r \to 0}\lim_{n\rightarrow\infty}\Vert\chi^i_{n}\nabla u_n\Vert_{L^\infty(S_r)}=\lim_{r \to 0}\Vert {\mathcal M}^i(\nabla u^H)\Vert_{L^\infty(S_r)}.\label{equal1} \end{eqnarray} \end{theorem} \setcounter{theorem}{0} \setcounter{definition}{0} \setcounter{lemma}{0} \setcounter{conjecture}{0} \setcounter{corollary}{0} \setcounter{equation}{0} \section{Local representation formula for layered and periodic microstructures} We now describe sequences of configurations for which one has equality in the spirit of \eqref{equal1}. The first class of configurations are given by sequences of finely layered media. The second class is given by a sequence of progressively finer periodic microstructures comprised of inclusions with smooth boundaries. In what follows the results of \cite{Chipot} provide the sufficient conditions \eqref{nonzero2-1} and \eqref{assum2} for the case of finely layered media. While the higher regularity results of \cite{LiVogelius} and \cite{LiNirenberg} allow for the computation of an upper bound for the periodic case. This upper bound agrees with an explicit lower bound developed in section 5. We note that the lower bound for the periodic case can also be obtained using the earlier results given in \cite{lipsima}. In order to proceed let us recall the fundamental results from homogenization theory for periodic media. We denote a $d$ dimensional cube centered at $x$ and of side length $r$ by $Y(x,r)$. For the unit cube centered at the origin we abbreviate the notation and write $Y$. The coefficient $A(y)$ is a periodic simple function defined on the unit period cell $Y$ taking the $N$ values $A_i$, $i=1,\ldots,N$ in the space of positive symmetric $d\times d$ matrices. We denote the indicator functions of the sets $Y_i$ where $A(y)=A_i$ by $\chi^i$ and write $A(y)=\sum_{i=1}^N A_i\chi^i(y)$. It is well known from the theory of periodic homogenization \cite{BLP} that the sequence of coefficients $A^{n}(x) = A(nx)$ $G-$ converge to the homogenized constant matrix $A^{H}$ given by the formula \begin{equation}\label{effectivematrix} A^{H}_{ij} = \int_{Y}A_{ik}(y)P_{kj}(y)dy \end{equation} where $P_{kj} = \partial _{x_{k}}\phi^{j}(y) + \delta_{kj}$ and $\phi ^{j}$ are $Y$-periodic $H^{1}_{loc}(R^{d})$ solutions of the cell problems \begin{equation}\label{cellproblemperiodic} \text{div}(A(y)(\nabla \phi^{j}(y) + {\bf e}^{j} ))= 0 \quad \text{in $\mathbb{R}^{d}$}, \end{equation} where this equation is understood in the weak sense, i.e., \begin{equation} \label{periodicweakform} \int_{Y} (A(y)(\nabla \phi^{j}(y) + {\bf e}^{j} ), \nabla \psi)dy = 0, \quad \forall \psi \in H^{1}_{per}(Y). \end{equation} For periodic microstructures we define the modulation function by \begin{eqnarray} {\mathcal M}^i(\nabla u^H)(x)=\||\chi^i(\cdot)P(\cdot)\nabla u^{H}(x)| \|_{L^\infty(Y)} \quad i=1,\cdots,N. \end{eqnarray} \subsection{Laminated microstructure} \label{41} The layered configurations as introduced in this section are a special class of periodic configurations. To fix ideas we consider a two dimensional problem and partition the unit period square $Y\subset \mathbb{R}^{2}$ for the layered material as follows: \[ Y_{1} = \{(y_{1},y_{2})\in Y: -\frac{1}{2}\leq y_{1}\leq-\frac{1}{2}+\theta\quad Y_{2} = \{(y_{1},y_{2})\in Y: -\frac{1}{2}+\theta\leq y_{1}\leq \frac{1}{2}\} \] where $\theta$ is a specified value in the interval $(0, 1).$ Let $\chi^1$ and $\chi^2$ denote the indicator functions of $Y_1$ and $Y_2$ respectively and consider the Y-periodic matrix function $A(y)$ given by \[ A(y) = \alpha I \chi^1(y) + \beta I \chi^2(y),\] for positive constants $\alpha<\beta$. $I$ is the $2\times 2$ identity matrix. Let $\Omega\subset\mathbb{R}^2$ and $u_{n}$ be the $H^{1}(\Omega)$ solution to \begin{equation} -{\rm div}\left(A(nx)\nabla u_n\right)=f ~~ \text{in $\Omega$ and} \quad u^{n}=0~~ \text{on $\partial \Omega$}. \end{equation} Then $ u_{n}$ converges weakly in $H^1(\Omega)$ as $n\rightarrow\infty$ to the $H^1(\Omega)$ solution $u^H$ of \begin{equation} -{\rm div}\left(A^H\nabla u^H\right)=f, ~~ \text{in $\Omega$ and} \quad u^{H}=0~~ \text{on $\partial \Omega$}. \label{hlimitu} \end{equation} where $A^{H}$ is determined using the formula \eqref{effectivematrix}. The gradient of solutions of the cell problem \eqref{cellproblemperiodic} for layered materials are given by \[ \begin{split} \nabla \phi^{1}(y) & = \left(\frac{(1-\theta)(\beta -\alpha)}{\theta\beta + (1-\theta)\alpha}\chi^1(y) + \frac{\theta(\beta - \alpha)}{\theta\beta + (1-\theta)\alpha}\chi^2(y)\right){\bf e}^{1} \end{split} \] and \[ \nabla \phi^{2}(y) = {\bf e }^{2}\quad \text{for all $y\in Y$}. \] We define the constants \begin{eqnarray} a_{h}=\frac{\alpha \beta}{\theta\beta + (1-\theta)\alpha}~~\text{ and}~~ a_{m}=\theta\alpha+(1-\theta)\beta, \label{arithhar} \end{eqnarray} and introduce the $Y$ periodic scalar coefficient $a(y)=\alpha\chi^1(y)+\beta\chi^2(y)$. A simple calculation gives \[ P(y)= \begin{bmatrix} p_{11}(y) &0\\ 0&1 \end{bmatrix}\quad\text{where }~~ p_{11}(y)=\frac{a_h}{a(y)} \] The homogenized matrix $A^{H}$ is given by \[ A^{H} = \begin{bmatrix} a_{h}&0\\ 0&a_{m} \end{bmatrix}. \] The modulation function for each phase is given by: \[ \begin{split} {\mathcal M}^{1}(\nabla u^H)(x) &= \sqrt{\left(\frac{\beta}{\theta\beta + (1-\theta)\alpha}\partial_{x_{1}} u^{H}\right)^{2} + (\partial_{x_{2}}u^{H})^{2}}\\ {\mathcal M}^{2}(\nabla u^H)(x) &= \sqrt{\left(\frac{\alpha}{\theta\beta + (1-\theta)\alpha}\partial_{x_{1}} u^{H}\right)^{2} + (\partial_{x_{2}}u^{H})^{2}} \end{split} \] We now apply the regularity and convergence results associated with G-convergent coefficients for sequences of layered materials \cite{Chipot}. For right hand sides $f\in H^{1}(\Omega)$ there exists a $p>2$ such that for any subdomain $\Omega'\Subset \Omega$ \[u^{n} \in H^{1, \infty}(\Omega')\quad \text{and}~~ \partial _{x_{2}}u_n, ~a(nx)\partial_{x_{1}}u_n\in H^{1, p}(\Omega')\] with the estimate that for some $C=C(\alpha, \beta,\Omega', \Omega)$, \begin{equation} \label{Chipotestimate} \|\partial _{x_2}u_n\|_{H^{1,p}(\Omega')} + \|a(nx)\partial_{x_1}u_n\|_{H^{1, p}(\Omega')} \leq C\|f\|_{H^{1}(\Omega)}, \end{equation} see \cite{Chipot}. The Sobolev embedding theorem implies that $\{\partial _{x_{2}}u_n\}_{n=1}^\infty $ and $ \{a(nx)\partial_{x_{1}}u_n\}_{n=1}^\infty$ are equicontinuous families over $\Omega'$ and uniformly bounded in $C(\Omega')$. Then from \eqref{Chipotestimate} and the weak convergence $u_n \rightharpoonup u^{H}$ in $H^{1}(\Omega)$ it follows that for a subsequence \begin{equation} \partial _{x_{2}}u_n \to \partial _{x_{2}} u^{H}, \quad \quad a(nx)\partial _{x_{1}}u_n \to a_{h}\partial _{x_{1}} u^{H}\quad \text{uniformly in $\Omega'$}. \label{uni} \end{equation} We observe that \begin{equation} \alpha|\partial_{x_1}u_n-p_{11}(nx)\partial_{x_1}u^H|\leq a(nx)|\partial_{x_1}u_n-p_{11}(nx)\partial_{x_1}u^H|=|a(nx)\partial_{x_1}u_n-{a_h}\partial_{x_1}u^H|, \label{identuniform} \end{equation} and on applying \eqref{uni} and noting that $P(y)$ is constant inside each phase we see for $i=1,2$ that \begin{eqnarray} |\chi^{i}(nx)\nabla u_n| &=& |\chi^{i}(nx)P(nx)\nabla u^{H}| + m^{i}_{n}(x)\\ \label{uniformremainder} &=&{\mathcal M}^{i}(\nabla u^{H})(x)+ m^{i}_{n}(x) \label{modanduniform} \end{eqnarray} where $m_{n}^{i}(x) \to 0$ uniformly in $\Omega'$. Hence we arrive at the local representation formula for layered microstuctures given by \begin{theorem} \label{layerlocal} \begin{eqnarray} \label{laminateconverge} \lim_{n\to \infty}\|\chi^{i}(nx)\nabla u_n \|_{L^{\infty}(\Omega')} &= & \| {\mathcal M}^{i}(\nabla u^{H})\|_{L^{\infty} (\Omega')}. \end{eqnarray} \end{theorem} It is easily seen that the uniform convergence implies that sequence of the gradients$\{\nabla u_n\}$ satisfy the non-concentrating conditions given by \eqref{nonzero2-1}. Indeed, setting \\$L^{i} = \lim_{n\to \infty}\|\chi^{i}(nx)\nabla u_n\|_{L^{\infty}(\Omega')},$ and for any $\delta>0$ there exists sufficiently large $n$ for which $|m^{i}_{n}(x)|<\frac{\delta}{2}$ for $x\in\Omega'$ and \[|\chi^i(nx)\nabla u_n(x)|>|{\mathcal M}^i\nabla u^H(x)|-\frac{\delta}{2}\] so \begin{eqnarray} \{ x\in \Omega': {\mathcal M}^{i}(\nabla u^{H}(x))> L^{i} -\frac{\delta}{2} \}&\subset & \{x\in \Omega': |\chi^{i}(nx)\nabla u_n|> L^{i} -\delta \}.\nonumber \end{eqnarray} Therefore we conclude that for $L^i>\delta > 0$ \[ \liminf_{n\to \infty}|\{x\in \Omega': |\chi_{i}(nx)\nabla u_n|> L^{i} -\frac{\delta}{2} \} |\geq |\{ x\in \Omega':{\mathcal M}^{i}(\nabla u^{H})(x) > L^{i} -\delta \} | > 0. \] Last the non-concentrating condition \eqref{assum2} follows immediately from the piecewise constant nature of the corrector matrix $P(y)$ for layered materials. \subsection{Periodic microstructure} \label{42} We consider periodic microstructures associated with particle and fiber reinforced composites. As before we divide $Y$ into a union of $N$ disjoint subdomains $Y_{1}\dots Y_{N}$. Instead of proceeding within the general context developed in \cite{LiNirenberg}, \cite{LiVogelius} we fix ideas we suppose that the domains $Y_1,\dots,Y_{N-1}$ denote convex particles with smooth (i.e., $C^2$) boundaries embedded inside a connected phase described by the domain $Y_N$, see Figure \ref{Particles}. As before we denote the indicator function of $Y_i$ by $\chi^i$ and the $Y$ periodic coefficient is written $A(y)=\sum_{i=1}^N\chi^i(y)A_i$ with each $A_i$ being a symmetric $d\times d$ matrix of constants satisfying the coercivity and boundedness conditions given by \[ \lambda |\xi|^{2}\leq (A_i\xi, \xi)\leq \Lambda |\xi|^{2}\quad \forall \xi \in \mathbb{R}^{d},\hbox{ and $i=1,\ldots,N$}. \] For any bounded domain $\Omega \subset R^{d}$ we consider the $H^{1} (\Omega)$ solutions $u_{n}$ of \begin{equation}\label{basicpde} \text{div}(A(nx)\nabla u_{n} )= 0 \quad \text{in $\Omega$} \end{equation} associated with prescribed Neumann or Dirichlet boundary conditions. From the theory of periodic homogenization the solutions converge weakly in $H^{1}$ to the homogenized solution $u^{H}$. In this section we establish the following local representation theorem. \begin{theorem}\label{periodicidentity} Let $A(y)$ and the subdomains $\{Y_{i}\}_{i=1}^N$ be as described above. Suppose $u_{n}$ solves \eqref{basicpde} and $u^{H}$ is the corresponding homogenized solution, then for any subdomain $\Omega'$ compactly contained inside $\Omega$ one has the local representation formula given by \begin{equation} \lim_{n\to \infty} \|\chi^i(nx)\nabla u_{n}\|_{L^{\infty} (\Omega '))} = \|{\mathcal M}^i(\nabla u^H)\|_{L^{\infty}(\Omega ')}. \label{periodicidentityy} \end{equation} \end{theorem} For the proof we will use the $W^{1,\infty}$ estimate for weak solutions of linear equations with oscillatory periodic coefficients obtained in \cite{AvellanedaLin} for smooth coefficients and later extended in \cite{LiNirenberg} to include discontinuous but locally H\"older coefficients. A $W^{1,p}$ estimate for $p<\infty$ is given in \cite{Caffarelliperal}. We point out that we have restricted the discussion to periodic homogenization for particle reinforced configurations of the kind illustrated in Figure \ref{Particles}. However the regularity theory for oscillatory periodic coefficients developed in \cite{LiNirenberg} applies to more general types of domains $Y_1,\ldots,Y_N$ with $C^{1,\alpha}$ boundaries. We note that the proof given here goes through verbatim for period cells with coefficients satisfying the general hypotheses described in \cite{LiNirenberg}. \begin{figure}[tbp] \centerline{\scalebox{0.3}{\includegraphics{Particles.jpg}}} \caption{Particle reinforced geometry for two inclusions $Y_1$ and $Y_2$.} \label{Particles} \end{figure} Theorem 1.9 of \cite{LiNirenberg} and a suitable rescaling shows that for any $r>0$ and $ Y(x_{0}, r)\subset \Omega$ that there exists a positive constant $C$ independent of $r$ and $n$ for which \begin{equation} \|\nabla u_{n}\|_{L^{\infty}(Y(x_{0},r/2))}\leq Cr^{-1}\|u_{n}\|_{L^{\infty}(Y(x_{0}, r))}. \label{linftyest} \end{equation} The local $L^{\infty}$ estimate for weak solutions of elliptic linear problems (Theorem 8.17, \cite{Gilbarg-Trudinger}) gives \begin{equation} \|u_{n}\|_{L^{\infty}(Y(x_{0}, r))} \leq C r^{-d/2} \|u_{n}\|_{L^{2}(Y(x_{0}, 2r))}, \label{linfltwo} \end{equation} where the constant $C$ is independent of $n$ and $r$. Combining the two estimates delivers the following lemma. \begin{lemma}\label{w1inftyestimate} Let $A(y)$ and the subdomains $\{Y_{i}\}$ be as described above. Choose $r\in (0, 1)$ such that $Y(x_{0}, 2r)\subset \Omega$. Then if $u_{n}$ solves \eqref{basicpde}, then there exists $C$, independent of $n$ and $r$ such that \begin{equation} \|\nabla u_{n}\|_{L^{\infty}(Y(x_{0},r/2))}\leq Cr^{\frac{-(d+2)}{2}}\|u_{n}\|_{L^{2}(Y(x_{0}, 2r))}. \label{lemma2} \end{equation} \end{lemma} \begin{proof}[Proof of Theorem \ref{periodicidentity}] To prove the theorem we first show that there is a subsequence, $n_k$, for which \begin{eqnarray} \lim_{k\rightarrow \infty}\Vert\nabla u_{n_k}(x)-P(n_k x)\nabla u^H(x)\Vert_{L^\infty(\Omega')}. \label{seqrep} \end{eqnarray} To begin we choose $x_{0}\in \Omega$ and $r>0$ such that $rn$ is an integer and $Y(x_{0}, r) \subset \Omega$ contains an integral number of periods of diameter $1/n$. Then from \eqref{cellproblemperiodic} we see that $ (1/n) \phi^{j}(n)$ is a $Y(x_{0}, r)$-periodic $H^{1}_{loc}$ function satisfying \begin{equation}\label{basiccellpde} \text{div}(A(nx)(\nabla (\frac{1}{n} \phi^{j}(nx)) + e^{j}))= 0 \quad \text{in $\mathbb{R}^d$} \end{equation} Combining equations (\ref{basicpde}) and (\ref{basiccellpde}) we note that \begin{equation}\label{combinedpde} \text{div}(A(nx)[\nabla u_{n}-(\nabla w_{n}(x, x_{0}) + \nabla u^{H}(x_{0}))] )= 0 \quad \text{in $Y(x_{0}, r)$} \end{equation} where \[ w_{n}(x,x_{0}) = \sum^{d}_{j} \frac{1}{n} \phi^{j}(nx)) \partial_{x_{j}} u^{H}(x_{0}) + u^{H}(x_{0}). \] Observe that for this choice of $Y(x_0,r)$ \[ \nabla w_{n}(x, x_{0}) + \nabla u^{H}(x_{0}) =P(n x) \nabla u^{H}(x_{0}) \] Adding and subtracting $P(n x) \nabla u^{H}(x_0) $ delivers \begin{eqnarray}\label{beginestimate} &&\|\nabla u_{n}(x)-P(nx) \nabla u^{H}(x)\|_{L^{\infty} (Y(x_{0}, r/2))) } \leq \|\nabla u_{n}(x)-P(nx) \nabla u^{H}(x_0)\|_{L^{\infty} (Y(x_{0}, r/2))) }\nonumber\\ &&+ \|P(nx) \nabla u^{H}(x)- P(nx) \nabla u^{H}(x_{0})\|_{L^{2}(Y(x_{0}, r/2)))}. \end{eqnarray} We apply Lemma \ref{w1inftyestimate} to find a constant $C$ independent of $n$ and $r$ such that the following estimate holds true: \[ \begin{split} \|\nabla u_{n}-&P(nx) \nabla u^{H}(x_{0}) \|_{L^{\infty} (Y(x_{0}, r/2))) }\\ &\leq \frac{C}{r^{-(d+2)/2}}\|u_{n} - (w_{n} + \nabla u^{H}(x_{0}) \cdot(x-x_{0}))\|_{L^{2}(Y(x_{0}, r)))} \end{split} \] Combining with \eqref{beginestimate} we obtain \begin{equation}\label{basicestimate} \begin{split} \|\nabla u_{n}-&P(nx) \nabla u^{H}(x)\|_{L^{\infty} (Y(x_{0}, r/2))) }\\ &\leq \frac{C}{r^{-(d+2)/2}}\|u_{n} - (w_{n} + \nabla u^{H}(x_{0}) \cdot(x-x_{0})) \|_{L^{2}(Y(x_{0}, r)))} \\ &+ \|P(nx) \nabla u^{H}(x)- P(nx) \nabla u^{H}(x_{0})\|_{L^{2}(Y(x_{0}, r)))}. \end{split} \end{equation} We bound the the first and second terms on the righthand side of \eqref{basicestimate}. The first term in the right hand side of \eqref{basicestimate} is bounded above by \begin{equation}\label{basicestimate2} \begin{split} \|u_{n} - (w_{n} + \nabla &u^{H}(x_{0}) \cdot(x-x_{0}) )\|_{L^{2}(Y(x_{0}, r)))}\\ &\leq \| u^{H}(x) - (u^{H} (x_{0}) + \nabla u^{H}(x_{0}) \cdot(x-x_{0}))\|_{L^{2}(Y(x_{0}, r))}\\ & +\| u_{n} - u^{H}\|_{L^{2} (Y(x_{0}, r))} + \| \sum^{d}_{j} \frac{1}{n}\phi^{j}(n x) \partial_{x_{j}} u^{H}(x_{0})\|_{L^{2}(Y(x_{0}, r))}\\ \end{split} \end{equation} We apply Lemma \ref{w1inftyestimate} together with a priori elliptic estimates to find that \begin{equation}\label{linftyboundofcorrector} \|\nabla \phi^j(n\cdot) \|_{L^{\infty}(\mathbb{R}^{d})} \leq C, \end{equation} where $C$ is independent of $n$. Moreover, as $u^H$ is a solution of a PDE in divergence form with constant coefficients it satisfies \begin{equation} \label{smoothnessofuh} \begin{split} |u^{H}(x) - (u^{H} (x_{0}) + \nabla u^{H}(x_{0}) \cdot(x-x_{0})) |&\leq M|x-x_{0}|^{2},\hbox{ $x\in\Omega'$}\\ | \nabla u^{H}(x) - \nabla u^{H}(x_{0})| &\leq M |x-x_{0}|, \hbox{ $x\in\Omega'$} \end{split} \end{equation} where $M$ is the supremum of $|D^{2} u^{H}(x)|$ over $\Omega'$. Applying \eqref{linftyboundofcorrector} and \eqref{smoothnessofuh} gives \begin{equation}\label{basicestimate3} \begin{split} \|u_{n} - (w_{n} + \nabla &u^{H}(x_{0}) \cdot(x-x_{0}) )\|_{L^{2}(Y(x_{0}, r)))}\\ &\leq C\left(r^{2+d/2}M + \frac{1}{\sqrt{n}}+\| u_{n} - u^{H}\|_{L^{2} (Y(x_{0}, r))}\right). \end{split} \end{equation} for some constant $C$ independent of $r$ and $n$. From periodicity it follows that $$\Vert \nabla\phi^j(n x)\Vert_{L^\infty(Y(x_0,r))}=\Vert \nabla\phi^j(y)\Vert_{L^\infty(Y)}$$ and applying Lemma \ref{w1inftyestimate} together with \eqref{smoothnessofuh} delivers \begin{equation} \label{linftyboundofcorrector2} \|P(nx) \nabla u^{H}(x)- P(nx) \nabla u^{H}(x_{0})\|_{L^{2}(Y(x_{0}, r))}\leq C r^{(d+2)/2} \end{equation} From the theory of periodic homogenization see, \cite{BLP}, \cite{ZOK}, one has the convergence rate given by \[ \| u_{n} - u^{H}\|_{L^{2} (Y(x_{0}, r))}\leq C\frac{1}{\sqrt{n}} \] and collecting results we have \begin{eqnarray} \|\nabla u_{n}- &P(nx) \nabla u^{H}(x) \|_{L^{\infty} (Y(x_{0}, r/2))) }\leq C\left(Mr + \frac{1}{\sqrt{n r^{d+2}}}\right). \label{rnk} \end{eqnarray} We pass to a subsequence $n_k$, and consider $Y(x_0,r_k)$ such that, $r_{k}\to 0$, $r_{k}n_{k}$ is an integer, and $r_{k}^{(d+2)/2}n_{k}^{1/2} \to \infty$ as $k\to \infty.$ Then consider any subdomain $\Omega '\subset\subset \Omega $, and cover it with cubes $\{ Y(x_{i}, r_{k}/2)\}_{x_{i}\in\Omega ' }$. Using compactness we choose finitely many cubes so that \[ \Omega '\subset \cup_{i=1}^{L} Y(x_{i}, r_{k}/2), \] Now since $\|\nabla u_{n_{k}}-P(n_{k}x) \nabla u^{H}(x) \|_{L^{\infty} (\Omega '))} $ is bounded above by the $L^{\infty}$ norms over a finite collection of cubes, we see that \[ \|\nabla u_{n_{k}}-P(n_{k}x) \nabla u^{H}(x) \|_{L^{\infty} (\Omega '))}\leq C\left(r_{k} + \frac{1}{\sqrt{n_k r_k^{d+2}}}\right) \] for sufficiently large $k$ to conclude \begin{eqnarray} \lim_{k\to \infty}\|\nabla u_{n_{k}}\|_{L^{\infty} (\Omega '))} = \lim_{k \to \infty }\|P(n_{k}x) \nabla u^{H}(x) \|_{L^{\infty} (\Omega '))}, \label{cuniformconvg} \end{eqnarray} so \begin{eqnarray} \lim_{k\to \infty}\|\chi^i(n_k x)\nabla u_{n_{k}}\|_{L^{\infty} (\Omega '))} = \lim_{k \to \infty }\|\chi^i(n_k x)P(n_{k}x) \nabla u^{H}(x) \|_{L^{\infty} (\Omega '))}. \label{cuniformconvgi} \end{eqnarray} Now we bound \eqref{cuniformconvgi} from above and below by $\Vert{\mathcal M}^i(\nabla u^H)\Vert_{L^\infty(\Omega')}$. First note for each $n_k$ and $x\in\Omega'$ that \begin{eqnarray} \label{upbddd} |\chi^i(n_k x)P(n_{k}x) \nabla u^{H}(x)|\leq\Vert\chi^i(\cdot)P(\cdot)\nabla u^{H}(x)\Vert_{L^\infty(Y)} \end{eqnarray} and we conclude that \begin{eqnarray} \lim_{k \to \infty }\|\chi^i(n_k x)P(n_{k}x) \nabla u^{H}(x) \|_{L^{\infty} (\Omega ')}\leq\Vert{\mathcal M}^i(\nabla u^H)\Vert_{L^\infty(\Omega')}. \label{cuniformconvgiubddd} \end{eqnarray} The lower bound \begin{eqnarray} \Vert{\mathcal M}^i(\nabla u^H)\Vert_{L^\infty(\Omega')}\leq\lim_{k\to \infty}\|\chi^i(n_k x)\nabla u_{n_{k}}\|_{L^{\infty} (\Omega '))} \label{cuniformconvgilbddd} \end{eqnarray} follows from a direct application of Corollary \ref{inftyonesidedinequality} proved the next section and we conclude that \begin{eqnarray} \lim_{k\to \infty}\|\chi^i(n_k x)\nabla u_{n_{k}}\|_{L^{\infty} (\Omega '))} =\Vert{\mathcal M}^i(\nabla u^H)\Vert_{L^\infty(\Omega')}. \label{cuniformconvgie} \end{eqnarray} The theorem follows on noting that identical arguments can be applied to every subsequence of $\{\chi^i(nx)\nabla u_n\}_{n=1}^\infty$ to conclude the existence of a further subsequence with limit given by \eqref{cuniformconvgie}. \end{proof} \setcounter{theorem}{0} \setcounter{definition}{0} \setcounter{lemma}{0} \setcounter{conjecture}{0} \setcounter{corollary}{0} \setcounter{equation}{0} \section{Continuously graded microstructures} In this section we consider a class of coefficient matrices associated with {\em continuously graded} composites made from N distinct materials. In order to express the continuous gradation of the microstructure we introduce the characteristic functions $\chi^{i}(x, y)$, $i = 1,\dots, N$ belonging to $L^1(\Omega \times Y)$ such that for each $x$ the function $\chi^{i}(x, \cdot)$ is periodic and represents the characteristic function of the $i^{th} $ material inside the unit period cell $Y$. The characteristic functions are taken to be continuous in the $x$ variable according to the following continuity condition given by \begin{equation}\label{continuity} \lim_{h \to 0}\int_Y |\chi^i(x+h,y)-\chi(x,y)|\,dy=0. \end{equation} The coefficient associated with each material is denoted by $A_i$ and is a constant symmetric matrix satisfying the ellipticity condition \[ \lambda \leq A_{i}\leq\Lambda \] for fixed positive numbers $\lambda < \Lambda$. We define the coefficient matrix \[ A(x,y) = \sum_{i=1}^{N}A_{i}\chi^{i}(x,y). \] This type of coefficient matrix appears in prototypical problems where one seeks to design structural components made from functionally graded materials \cite{markworth} and \cite{ootao}. Here the configuration of the N materials is locally periodic but changes across the domain $\Omega$. The composite is constructed by dividing the domain $\Omega$ into subdomains $\Omega_{k,l}$, $l=1,\ldots,M_k$ of diameter less than or equal to $1/k$, $k=1,2,\ldots$ and $\Omega=\cup_{l=1}^{M_k}\Omega_{k,l}$. Each subdomain contains a periodic configuration of $N$ materials. The following lemma allows us to approximate the ideal continuously graded material by a piecewise periodic functionally graded material that can be manufactured. \begin{lemma} For a given subdivision $\Omega_{k,1},\ldots\Omega_{k,M_k}$ of diameter less than $1/k$ and any $i = 1,\cdots, N$, there exists a sequence $\{\chi^i_k(x,y)\}_{k=1}^\infty$ of approximations to $\chi^i(x,y)$ given by \begin{equation} \chi^i_k(x,y)=\sum_{l}\chi_{\Omega_{k,l}}(x)\chi^i_{k,l}(y) \label{graded} \end{equation} with the property that \begin{equation} \lim_{k\to \infty}\int_{\Omega\times Y}|\chi^i_k(x,y)-\chi^i(x,y)|dydx=0. \label{approx} \end{equation} In \eqref{graded}, $\chi_{\Omega_{k,l}}(x)$ denotes the characteristic function of $\Omega_{k,l}$ and $\chi^i_{k,l}(y) = \chi^{i}(x_{k,l},y)$ is the characteristic function associated with the configuration of the $i^{th}$ phase inside the subdomain $\Omega_{k,l}$ at a sample point $x_{k,l} \in \Omega_{k,l}$. \label{cgraded} \end{lemma} \begin{proof} The definition of the approximating function is given in \eqref{graded}. We verify that \eqref{approx} is satisfied. For each $x\in \Omega,$ define the sequence of functions \[ \Gamma^{i}_{k}(x) = \int_{Y}|\chi^i_k(x,y)-\chi^i(x,y)|dy. \] Then $\Gamma^{i}_{k}(x)\to 0$ for all $x\in \Omega.$ Indeed, for a fixed $x\in \Omega, $ there exists a sequence of subdomains $x\in \Omega_{k, l_{k}}$ and points $x_{k, l_{k}}\in \Omega_{k, l_{k}}$ such that by definition, \[ \Gamma_{k}^{i} (x) = \int_{Y}|\chi^{i}(x_{k, l_{k}}, y) - \chi^{i}(x, y)|dy. \] It is evident that $|x_{k, l_{k}} - x| < 1/k$ since $x_{k, l_{k}}$ and $x$ both belong to $\Omega_{k, l_{k}}.$ Applying the continuity condition \eqref{continuity}, we see that $\Gamma_{k}^{i} (x) \to 0$ as $k\to \infty$ and \eqref{approx} follows from the Lebesgue dominated convergence theorem. \end{proof} Let us define the coefficient matrix of the functionally graded material. Divide the domain $\Omega$ into subdomains $\Omega_{k,l}$, $l=1,\ldots,M_k$ of diameter less than or equal to $1/k$, $k=1,2,\ldots$ and $\Omega=\cup_{l=1}^{M_k}\Omega_{k,l}$. Each subdomain contains a periodic configuration of $N$ materials with period $1/n$ such that $1/k>1/n$. The configuration of the $i^{th}$ phase inside a {\em functionally graded} composite is described by $\chi^i_k(x,nx)$, where $\chi_{k}^{i}(x,y)$ is given by \ref{graded}. The corresponding coefficient matrix is denoted by $A^k(x,nx)$ and is written as \begin{equation} A^k(x, nx) = \sum_{i}^{M_k}\chi^i_{k}(x, nx)A_{i}. \label{coeffk} \end{equation} As seen from the proof of the lemma the continuity condition \eqref{continuity} insures that near by subdomains $\Omega_{k,l}$ and $\Omega_{k,l'}$ have configurations that are nearly the same when $1/k$ is sufficiently small. The fine-scale limit of such composites is obtained by considering a family of partitions indexed by $j=1,2, \dots, $ with subdomains $\Omega_{l}^{k_{j}}$ of diameter less that or equal to $1/k_{j}$. The scale of the microstructure is given by $1/n_{j}$. Both $1/k_{j}$ and $ 1/n_{j} $ approach zero as $j$ goes to infinity and we require that $\lim_{j \to \infty}\frac{1/n_{j}}{1/k_{j}}=0$. For future reference the associated indicator functions and coefficients are written \begin{equation} \chi^i_{k_j}(x, n_j x)~~\text{and} ~~A^{k_j}(x,n_j x). \label{charcoeff} \end{equation} Let \begin{equation}\label{effective} A^{H}(x) = \int_{Y}A(x, y)P(x, y)dy \end{equation} where the matrix $P(x,y)$ is defined by \begin{equation} P(x, y)_{i, j} = \frac{\partial w^{j}}{\partial y_{i}} + \delta_{i j}, \label{pcontinuous} \end{equation} and $w^{i}(x, \cdot)$ is a $Y$ periodic function that solves the PDE \begin{equation}\label{cellgraded} \text{div}_{y}(A(x, y)(\nabla_{y}w^{i}(x, y) + e^{i}) )= 0, \end{equation} where $\{e^{i}\}$, $i=1,\ldots$ is an orthonormal basis for $\mathbb{R}^{d}$. The Sobolev space of square integrable functions with square integrable derivatives periodic on $Y$ is denoted by $H_{\rm{per}}^{1}(Y)$. The functions $w^i(x,y)$ belong to $C(\Omega;H^{1}_{\rm{per}}(Y))$ this follows from \eqref{continuity} and is proved in the Appendix. We present the homogenization theorem for the sequences $A^{k_j}(x,n_j x)$ proved in \cite{liproy}. \begin{lemma}\label{G-convergence} One can construct sequences $\{\chi^i_{k_j}(x,xn_j)\}_{j=1}^\infty$ for which the coefficient matrices $\{ A^{k_{j}}(x, n_{j}x)\}_{j=1}^{\infty}$ $G-$ converge to the effective tensor $A^{H}(x)$ defined by (\ref{effective}). \end{lemma} Let $f\in H^{-1}_{0}(\Omega)$ be given. Then by Lemma \ref{G-convergence} the sequence of solutions$\{u_{j}\}$ of the equation: \[ - \text{div}[A^{k_{j}}(x, n_{j}x)\nabla u_{j} (x)] = f, \quad u_{j} \in H^{1}_{0}(\Omega) \] converge to $u^{H}$ weakly in $H^{1}_{0}$, where $u^{H}$ solves the the equation \[ -\text{div} [A^{E}(x))\nabla u^{H} (x)] = f, \quad u^{H} \in H^{1}_{0}(\Omega). \] We now have the following result. \begin{theorem}\label{representationThm} Let $V \subset L^{1}(Y)\cap L^{\infty}(Y)$ be a countable dense subset of $L^{1}(Y)$. Assume that all elements of $V$ are periodically extended to $\mathbb{R}^{d}$. Suppose that $\phi(x)\in C(\overline{\Omega})$, $\eta(x) \in V$ and $u_{j}$, $P$ and $u^{H}$ be given as above. Then \[ \lim_{j\to \infty}\int_{\Omega}\phi(x)\eta( n_{j}x)\chi^{i}_{k_j}(x, n_{j}x)|\nabla u_{j}|^{2}dx = \int_{\Omega}\int_{Y}\phi(x)\eta(y)\chi^{i}(x,y)|P(x, y)\nabla u^{H}(x)|^{2}dydx. \] \end{theorem} By taking the modulation functions for continuously graded composites (see, \cite{liproy}) to be \begin{eqnarray} {\mathcal M}^i(\nabla u^H)(x)=\|\chi^{i}(x, \cdot)P(x,\cdot)\nabla u^{H}(x) \|_{L^\infty(Y)}. \label{figraded} \end{eqnarray} we obtain the following corollary. \begin{corollary}\label{inftyonesidedinequality} \begin{eqnarray} \|{\mathcal M}^i(\nabla u^H)\|_{L^{\infty}(\Omega)}\leq \limsup_{j\to \infty}\|\chi^{i}_{k_j}(x, n_j x)\nabla u_{j}\|_{L^{\infty}(\Omega)}.\label{lowergraded} \end{eqnarray} \end{corollary} Under an additional asymptotic condition on the distribution functions for the sequence $\{\nabla u_j\}_{j=1}^\infty$, equality can be achieved in the above corollary. Indeed, define \begin{eqnarray} S_{t, i}^{j}=\{ x: \chi^{i}_{k_j}(x, n_j x)|\nabla u_{j}(x)|^{2} > t\}, \quad \chi_{t, i}^{j}(x) := \chi_{S_{t, i}^{j}} \label{setsindicator} \end{eqnarray} and the distribution functions are given by \begin{eqnarray} |S_{t,i}^j|=\int_\Omega\chi_{t, i}^{j}\,dx. \label{distribution} \end{eqnarray} Passing to a subsequence, there exists density functions $\theta_{t, i}$ such that \[ \chi_{t, i}^{j} (x) \stackrel{*}{\rightharpoonup} \theta_{t, i}(x) \quad \text{$L^{\infty} $weak *} \] and for any open subset $S\subset \Omega$ \begin{eqnarray} \lim_{j\to\infty}|S_{t,i}^j\cap S| = \int_S\theta_{t,i}\,dx. \end{eqnarray} We present a sufficient condition on the distribution functions $|S_{t,i}^j|$ associated with $\{\nabla u_j\}_{j=1}^\infty$ for which equality holds in \eqref{lowergraded}. \begin{corollary}\label{inftyEquality} Suppose that $l = \limsup_{j\to \infty}\|\chi^{i}_{k_j}(x, n_j x)|\nabla u_{j}|^{2}\|_{L^{\infty}(\Omega)}< \infty$ and for each $\delta>0$ there exists a positive number $\beta_\delta>0$ for which \begin{eqnarray} |\{x\in\Omega:\theta_{l-\delta, i} > 0\}|>\beta_\delta. \label{suffconddd} \end{eqnarray} Then \[ \limsup_{j\to \infty}\|\chi^{i}_{k_j}(x, n_{j}x)\nabla u_{j}\|_{L^{\infty}(\Omega)} = \| {\mathcal M}^i(\nabla u^H)\|_{L^{\infty}(\Omega)} \] \end{corollary} \begin{proof}(Proof of Corollary \ref{inftyEquality}) The homogenization constraint \cite{liproy} states that for almost every $x\in \Omega$ \[ \theta_{t, i}(x)({\mathcal M}^{i}(\nabla u^H(x)) - t)\geq 0\quad i=1, \dots N. \] It follows that on the set where $\theta_{t, i} >0$, we have ${\mathcal M}^{i}(\nabla u^H(x)) \geq t$. Let \[l_{j} = \|\chi^{i}_{k_j}(x, n_{j}x)|\nabla u^{j}|^{2}\|_{L^{\infty}(\Omega)}.\] For a subsequence $l_{j} \to l$. Then given $\delta > 0$, there exists a natural number $J$ such that \[ l-\delta/2< l_{j} = \|\chi^{i}_{k_j}(x, n_{j}x)|\nabla u^{j}|^{2}\|_{L^{\infty}(\Omega)} < l+\delta/2 \quad \forall j\geq J. \] The measure of the set $S_{l_{j}-\delta/2, i}^{j}$ is positive. Moreover $S_{l_{j}-\delta/2, i}^{j} \subset S_{l-\delta, i}^{j} $ and \[ \chi_{l-\delta, i}^{j}\stackrel{*}{\rightharpoonup}\theta_{l-\delta, i}(x)\quad \text{$L^{\infty}$ weak *~ as $j\to \infty$ } \] From hypothesis the set where $\theta_{l-\delta, i} > 0$ is a set of positive measure for all $\delta > 0$. Therefore, \[ {\mathcal M}^{i}(\nabla u^H(x))\geq l-\delta, \] on a set of positive measure that is $ \|{\mathcal M}^{i}(\nabla u^H(x))\|_{\infty}\geq l-\delta. $ The corollary is proved since $\delta >0$ is arbitrary. \end{proof} \begin{proof}(Proof of corollary \ref{inftyonesidedinequality}) From Theorem \ref{representationThm} it follows that for any $\phi\in C(\overline{\Omega})$ and $\eta \in V$ is $Y-$ periodic, \[ \begin{split} \int_{\Omega}\int_{Y}\chi^{i}(x, y)&\phi(x)\eta( y)|P(x,y)\nabla u^{H}|^{2}dydx \\&\leq \lim_{j\to \infty}\int_{\Omega}|\phi(x)\eta( n_{j}x)|dx \limsup_{j\to \infty}\|\chi^{i}_{k_j}(x, n_{j}x)|\nabla u_{j}|^{2}\|_{L^{\infty}(\Omega)} \end{split} \] By the Riemann-Lebesgue lemma, \[ \lim_{j\to \infty}\int_{\Omega}|\phi(x)\eta( n_{j}x)|dx = \int_{\Omega}\phi(x)dx\int_{Y}|\eta( y)|dy. \] Dividing both sides by the $L^{1}$-norm of $\phi$, we obtain that for every $x\in \Omega \setminus Z$, where $Z$ is a set of measure zero, \[ \int_{Y}\chi^{i}(x, y)\eta( y)|P(x,y)\nabla u^{H}|^{2}dy \leq \int_{Y}|\eta( y)|dy\limsup_{j\to \infty}\|\chi^{i}_{k_j}(x, n_{j}x)|\nabla u_{j}|^{2}\|_{L^{\infty}(\Omega)} \] The set $Z$ depends on the choice of $\eta$. But since $V$ is countable, the union of the sets $Z$ corresponding to elements of $V$ will be of measure zero and the above inequality is true for any $\eta \in V$ and for every $x$ outside this union. Now divide the last inequality by the $L^{1}$ norm of $\eta$ in $Y$. Taking the sup over $V$ and noting that $V$ is dense in $L^{1}(Y)$, proves the corollary. \end{proof} \begin{proof} (Proof of Theorem \ref{representationThm}) For $\beta > 0$ , define \begin{eqnarray} A_{1}(x, y) = A(x, y) + \beta\chi^{i}(x, y)\phi(x)\eta(y)I. \label{coeffpurt} \end{eqnarray} Now let $v_{j}$ solve \[ -\text{div} [A^{k_{j}}_{1}(x, n_{j}x)\nabla v_{j}]=f, \quad v_{j} \in H_{0}^{1}(\Omega) \] Then for any $\varphi\in H_{0}^{1}(\Omega)$, we have \begin{equation}\label{integralform} \begin{split} \int_{\Omega} (A_{1}^{k_{j}}(x, n_{j}x)\nabla v_{j}, \nabla \varphi)dx &= \int_{\Omega} f\varphi dx \quad \text{and }\\ \int_{\Omega} (A^{k_j}(x, n_{j} x)\nabla u_{j}, \nabla \varphi)dx &= \int_{\Omega} f\varphi dx \end{split} \end{equation} Let $\delta u_{j} = v_{j} - u_{j}$. Then subtracting the second equation above from the first, we obtain \[ \int_{\Omega} (A_{1}^{k_{j}}(x, n_{j}x)\nabla\delta u_{j}, \nabla \varphi)dx + \int_{\Omega}((A_{1}^{k_{j}}( x, n_{j}x) - A^{k_{j}}(x, n_{j}x))\nabla u_{j}, \nabla \varphi )dx = 0, \] for all $\varphi\in H_{0}^{1}(\Omega)$. Simplifying the above equation we get \[ \int_{\Omega} (A_{1}^{j}(x, n_{j}x)\nabla\delta u_{j}, \nabla \varphi)dx + \beta\int_{\Omega}\chi^{i}_{k_j}(x, n_{j}x)\phi(x)\eta( n_{j}x)(\nabla u_{j}, \nabla \varphi )dx = 0, \] Plug in $\varphi = u_{j}$ in the above equation to get, \begin{equation}\label{integralform2} \int_{\Omega} (A_{1}^{k_{j}}(x, n_{j}x)\nabla\delta u_{j}, \nabla u_{j})dx +\beta \int_{\Omega}\chi^{i}_{k_{j}}(x, n_{j}x)\phi(x)\eta(n_{j}x)|\nabla u_{j}|^{2}dx = 0, \end{equation} Also plugging in $\varphi = \delta u_{j}$ in (\ref{integralform}) yields \begin{equation}\label{integralform3} \int_{\Omega} (A^{k_j}(x, n_{j}x)\nabla u_{j}, \nabla \delta u_{j})dx = \int_{\Omega} f\delta u_{j} dx. \end{equation} Subtracting (\ref{integralform3}) from (\ref{integralform2}) and noting that the coefficient matrices are symmetric we get \[ \beta\int_{\Omega}\chi^{i}_{k_{j}}(x, n_{j}x)\phi(x)\eta(n_{j}x)|\nabla u_{j}|^{2}dx + T^{j} = -\int_{\Omega} f\delta u_{j} dx \] where \[ T^{j} = \beta \int_{\Omega}\chi_{i}^{k_j}(x, n_{j}x)\phi(x)\eta(n_{j}x)(\nabla u_{j}, \nabla \delta u_{j})dx. \] Let us estimate $T^{j}.$ To begin with, observe that \[ \int_{\Omega} (A_{1}^{k_j}(x, n_{j}x)\nabla\delta u_{j}, \nabla \delta u_{j})dx + \beta\int_{\Omega}\chi^{i}_{k_j}(x, n_{j}x)\phi(x)\eta(n_{j}x)(\nabla u_{j}, \nabla \delta u_{j} )dx = 0, \] Them from ellipticity, we get \[ \begin{split} \alpha \int_{\Omega}|\nabla \delta u_{j}|^{2}dx&\leq \int_{\Omega} (A_{1}^{k_j}(x, n_{j}x)\nabla\delta u_{j}, \nabla \delta u_{j})dx\\ &\leq \beta \int_{\Omega}\chi^{i}_{k_j}(x, n_{j}x)|\phi(x)\eta(n_{j}x)||\nabla u_{j}|| \nabla \delta u_{j} |dx\\ &\leq C\beta \|\nabla \delta u^{j}\|_{L^2}\|\nabla u^{j}\|_{L^2}. \end{split} \] That is \[ \|\nabla \delta u_{j}\|_{L^2} \leq C\beta, \] since the sequence $\nabla u^{j}$ is bounded in $L^2$. From this and the definition on $T^{j}$ we obtain \[ |T^{j}| \leq C\beta^{2} \] From Lemma \ref{G-convergence} we know that $u_{j} \rightharpoonup u^{H}$, and $v_{j} \rightharpoonup v^{H}$, where $u^{H}$ and $v^{H}$ satisfy the following equations, respectively: For any $\varphi\in H_{0}^{1}(\Omega)$ \begin{equation}\label{homoguv} \begin{split} \int_{\Omega} A_{1}^{H}(x)\nabla v^{H}, \nabla \varphi)dx &= \int_{\Omega}f\varphi dx\\ \int_{\Omega} A^{H}(x)\nabla u^{H}, \nabla \varphi)dx &= \int_{\Omega}f\varphi dx \end{split} \end{equation} where $A^{H} (x)$, is the effective matrix given by \eqref{effective} and \begin{eqnarray} A_1^{E}(x) = \int_{Y}A_1(x, y)P_1(x, y)dy \label{effpert} \end{eqnarray} where the matrix $P_1(x,y)$ is defined by \begin{eqnarray} P_1(x, y)_{i, j} = \frac{\partial w_1^{j}}{\partial y_{i}} + \delta_{i j}, \end{eqnarray} and $w_1^{i}(x, \cdot)$ is a $Y$ periodic function that solves the PDE \begin{eqnarray} \text{div}_{y}(A_1(x, y)(\nabla_{y}w_1^{i}(x, y) + e^{i}) )= 0,\label{cellgradedpert} \end{eqnarray} where $\{e^{i}\}$, $i=1,\ldots$ is an orthonormal basis for $\mathbb{R}^{d}$. Writing $\delta u^H=v^H-u^H$ and letting $j\to \infty$, we obtain \begin{equation}\label{firstway} \begin{split} \beta\lim_{j\to\infty}\int_{\Omega}\chi^{i}_{k_j}(x, n_{j}x)\phi(x)\eta(n_{j}x)|\nabla u_{j}|^{2}dx + \lim_{j\to \infty}T^{j} &= -\lim_{j\to \infty}\int_{\Omega} f\delta u_{j} dx\\ &=-\int_{\Omega} f\delta u^H dx. \end{split} \end{equation} One easily verifies that the variational formulations \eqref{homoguv} can be written in terms of the two scale variational principles \cite{Nguetseng}, \cite{Allaire} given by \begin{equation}\label{twohomoguv} \begin{split} \int_{\Omega}\int_Y A_{1}(x,y)(\nabla v^{H}(x)+\nabla_y v_1(x,y)), \nabla \varphi(x)+\nabla_y\varphi_1(x,y)dydx &= \int_{\Omega}f\varphi dx\\ \int_{\Omega}\int_Y A(x,y)(\nabla u^{H}(x)+\nabla_y u_1(x,y)), \nabla \varphi(x)+\nabla_y \varphi_1(x,y)dydx &= \int_{\Omega}f\varphi dx, \end{split} \end{equation} where the solutions $(u^H,u_1)$, $(v^H,v_1)$, and trial fields $(\varphi,\varphi_1)$ belong to the space $H_0^{1}(\Omega)\times L^2[\Omega;W_{\rm{per}}^{1,2}(Y)]$. On writing $\delta u_1=v_1-u_1$, $\delta u^H=v^H-u^H$, $A_1(x,y)=A(x,y)+\beta\chi^i(x,y)\phi(x)\eta(y)I$, substitution into the first equation in \eqref{twohomoguv} and applying the second equation in \eqref{twohomoguv} gives \begin{equation}\label{ident} \begin{split} &\int_{\Omega}\int_Y A_{1}(x,y)(\nabla \delta u^{H}(x)+\nabla_y \delta u_1(x,y)), \nabla \varphi(x)+\nabla_y\varphi_1(x,y)dydx \\ &+\beta\int_{\Omega}\int_Y (\chi^i(x,y)\phi(x)\eta(y)(\nabla u^{H}(x)+\nabla_y u_1(x,y)), \nabla \varphi(x)+\nabla_y \varphi_1(x,y)dydx \\ &=0. \end{split} \end{equation} Next we substitute $(\varphi,\varphi_1)=(\delta u^H,\delta u_1)$ into the second equation of \eqref{twohomoguv} to obtain the identity \begin{equation}\label{secondidentity} \begin{split} &\int_{\Omega}\int_Y A(x,y)(\nabla u^{H}(x)+\nabla_y u_1(x,y)), \nabla \delta u^H(x)+\nabla_y \delta u_1(x,y)dydx \\ &= \int_{\Omega}f(x)\delta u^H(x) dx. \end{split} \end{equation} On choosing $(\varphi,\varphi_1)=(u^H,u_1)$ in \eqref{ident} and applying \eqref{secondidentity} we obtain \begin{equation}\label{Finaltident} \begin{split} &T+\beta\int_{\Omega}\int_Y(\chi^i(x,y)\phi(x)\eta(y)(\nabla u^{H}(x)+\nabla_y u_1(x,y)), \nabla \varphi(x)+\nabla_y \varphi_1(x,y))dydx\\ &=-\int_\Omega f\delta u^H dx, \end{split} \end{equation} where \begin{equation}\label{Tident} T=\beta\int_{\Omega}\int_Y(\chi^i(x,y)\phi(x)\eta(y)(\nabla u^{H}(x)+\nabla_y u_1(x,y)), \nabla \delta u^H(x)+\nabla_y \delta u_1(x,y))dydx. \end{equation} Next we set $(\varphi,\varphi_1)=(\delta u^H,\delta u_1)$ in \eqref{ident} and applying ellipticity delivers the estimate \begin{equation} \Vert\nabla \delta u^H+\nabla_y\delta u_1\Vert_{L^2(\Omega\times Y)}\leq C\beta, \label{deltaestimate} \end{equation} and we find that \begin{equation} \label{Tbound} |T|\leq C\beta^2. \end{equation} Since \eqref{firstway} and \eqref{Finaltident} have the same right hand sides we equate them and the theorem follows on identifying like powers of $\beta$. \end{proof} We conclude this section noting that lower bounds similar to those given here can be obtained in the context of two-scale convergent coefficient matrices \cite{lipsima}. \setcounter{theorem}{0} \setcounter{definition}{0} \setcounter{lemma}{0} \setcounter{conjecture}{0} \setcounter{corollary}{0} \setcounter{equation}{0} \section{Local representation formulas and gradient constrained design for graded materials} In view of applications it is important to identify graded material properties that deliver a desired level of structural performance while at the same time provide a hedge against failure initiation \cite{gosschristensen}. In many applications there is a separation of scales and the material configurations forming up the microstructure exist on length scales significantly smaller than the characteristic length scale of the loading. Under this hypothesis the structural properties are modeled using effective thermophysical properties that depend upon features of the underlying micro-geometry, see \cite{fuji}, \cite{markworth}. In this context overall structural performance measured by resonance frequency and structural stiffness are recovered from the solutions of homogenized equations given in terms of the effective coefficients (G-limits). In order to go further and design against failure initiation we record the effects of $L^\infty$ constraints on the local gradient field inside functionally graded materials. For this we use the local representation formulas given by modulation functions \eqref{figraded}. The multi-scale formulation of the graded material design problem has three features \cite{Liptstueb3}, \cite{LiptstuebAIAA}: \begin{enumerate} \item It admits a convenient local parametrization of microstructural information expressed in terms of a homogenized coefficient matrix \eqref{effective} and local representation formulas given by the modulation functions \eqref{figraded}. \item Is well posed, i.e., an optimal design exists. \item The optimal design is used to identify an explicit ``functionally graded microstructure'' that delivers an acceptable level of structural performance while controlling the local gradient field over a predetermined part of the structural domain. \end{enumerate} \noindent In what follows we describe the multiscale material design problem and focus the discussion on the control of the $L^\infty$ norm of the local gradient field. The admissible set of continuously graded locally periodic microstructures is specified by a vector $\underline{\beta}=(\beta_1,\ldots,\beta_n)$ of local geometric parameters. For example one may consider a periodic array of spheroids described by the orientation of their principle axis and aspect ratio. The periodic microstructure is specified in a unit period cell $Y$ centered at the origin. Points in the cell are denoted by $y$. The characteristic function of the $i^{th}$ phase in the unit cell is denoted by $\chi^i(\underline{\beta},y)$, $i=1,\ldots,N$. The vector $\underline{\beta}$ for the graded microgeometry can change across the structural domain $\Omega$ and we write $\underline{\beta}=\underline{\beta}(x)$ for $x\in\Omega$. The $x$ dependence of $\underline{\beta}$ corresponds to the gradation of material properties through a gradation in microstructure. The design vector $\underline{\beta}(x)$ is a uniformly H\"older continuous function of $x$ in the closure of $\Omega$. We write \begin{eqnarray} \chi^i(x,y)=\chi^i(\underline{\beta}(x),y) \label{identgrade} \end{eqnarray} and since $\underline{\beta}(x)$ is continuous one sees that $\chi^i(x,y)$ is continuous in the sense of \eqref{continuity}. The multi-scale design problem is formulated as follows: The admissible set $Ad$ of design vectors $\underline{\beta}(x)$ is the set of uniformly H\"older continuous functions satisfying the two conditions: \begin{itemize} \item There is a fixed positive constant $C$ such that: \begin{eqnarray} \sup_{x,x'\in\overline{\Omega}}\frac{|\underline{\beta}(x)-\underline{\beta}(x')|}{|x-x'|}<C. \label{holder} \end{eqnarray} \item The design vector $\underline{\beta}(x)$ takes values inside the closed bounded set given by the constraints \begin{eqnarray} \underline{b}_i\leq\beta_i(x)\leq\overline{b}_i,\hbox{ $i=1,\ldots,n$.} \label{box} \end{eqnarray} \end{itemize} The local volume fraction of the $i^{th}$ phase in the composite is given by $\theta_i(x)=\int_Y\chi^i(x,y)\,dy$. A resource constraint is placed on the amount of each phase appearing the design. It is given by \begin{equation} \int_\Omega\,\theta_i(x)\,d x\leq \gamma_i,\,\,i=1,\ldots,N. \label{constr2} \end{equation} The vector of constraints $(\gamma_1,\ldots,\gamma_N)$ is denoted by $\underline{\gamma}$. The set of controls $\underline{\beta}(x)\in Ad$ that satisfy the resource constraints (\ref{constr2}) is denoted by ${\mathcal A}d_{\underline{\gamma}}$. As an example we assume homogeneous Dirichlet conditions on the boundary of the design domain $\Omega$. For a given right hand side $f\in H^{-1}(\Omega)$ the overall structural performance of the graded composite is modeled using the solution $u^H$ of the homogenized equilibrium equation given by the $H_0^1(\Omega)$ solution of \begin{equation} -{\rm div\,}\left(A^H(x)\nabla u^H\right)=f. \label{equlibelastth} \end{equation} Here $A^H$ is given by \eqref{effective} with $\chi_i(x,y)$ given by \eqref{identgrade}. In this example the overall work done against the load is used as the performance measure of the graded material structure. This functional depends nonlinearly on the design $\underline{\beta}$ through the solution $u^H$ and is given by \begin{eqnarray} W(\underline{\beta})=\int_\Omega\,f u^H\,dx, \label{objective} \end{eqnarray} We pick an open subset $S\subset\Omega$ of interest and the gradient constraint for the multi-scale problem is written in terms of the modulation function. We set \begin{eqnarray} C_i(\underline{\beta})=\Vert {\mathcal M}^i(\nabla u^H)\Vert_{L^\infty(S)}, \hbox{ for $i=1,\ldots,N$} \label{gradconstmulti} \end{eqnarray} and the multi-scale optimal design problem is given by \begin{equation} P=\inf_{\underline{\beta}\in {\mathcal A}d_{\underline{\gamma}}} \{W(\underline{\beta}):\, C_i(\underline{\beta})\leq M,\,i=1,\ldots,N\}. \label{basicopth} \end{equation} When the constraint $M$ is chosen such that there exists a control $\underline{\beta}\in{\mathcal A}d_{\underline{\gamma}}$ for which $C_i(\underline{\beta})\leq M$ then an optimal design $\underline{\beta}^*$ exists for the design problem (\ref{basicopth}), this is established in \cite{lipnato}, \cite{Liptstueb1}. The optimal design $\underline{\beta}^*$ specifies characteristic functions $\chi^{i^*}(x,y)=\chi^i(\underline{\beta}^*(x),y)$ from which we recover continuously graded microgeometries $\chi^{i^*}_{k_j}\left(x,n_j x\right)$ and coefficient matrices\\ $A^{*,k_j}(x,n_j x)$ of the form \eqref{charcoeff}. The coefficients $A^{*,k_j}\left(x,n_j x\right)$ G-converge to the effective coefficient $A^{H^*}$ associated with the optimal design $\underline{\beta}^*$, see Lemma \ref{G-convergence}. Here the effective coefficient is given by \eqref{effective} with $\chi^i(x,y)=\chi^{i^*}(x,y)$. For each $j=1,\ldots$ the $H_0^1(\Omega)$ solution $u_j$ of the equilibrium problem inside the graded composite satisfies \begin{eqnarray} -{\rm div}\left(A^{*,k_j}(x,n_j x)\nabla u_j\right)=f \label{eqstar} \end{eqnarray} and the work done against the load is given by $W(u_j)=\int_{\Omega} f u_j dx$. This functional is continuous with respect to G-convergence hence $\lim_{j \to \infty} W(u_j)=W(\underline{\beta}*)$. We now apply Theorem \ref{upperboundd1} to discover that for any open set $S\subset\Omega$ with closure contained inside $\Omega$ there exists a decreasing sequence of sets $E_{k_j}$ for which $|E_{k_j}|\searrow 0$ and \begin{eqnarray} \limsup_{j\rightarrow\infty}\Vert\chi^{i^*}_{k_j}(x,n_j x)\nabla u_j(x)\Vert_{L^\infty(S\setminus E_{k_j})}\leq M,\hbox{ $i=1,2,\ldots,M$.}\label{ubound1applied} \end{eqnarray} Therefore we can choose a graded material design specified by $\chi^{i^*}_{k_j}(x, n_j x)$ with overall structural properties $W(u_j)$ close to the optimal one $W(\beta^*)$ and with $$\Vert\chi^{i^*}_{k_j}(x,n_j x)\nabla u_j(x)\Vert_{L^\infty(S\setminus E_{k_j})}\leq M$$ outside controllably small sets $E_{k_j}$. This is the essence of the design scheme for continuously graded composite structures developed in \cite{Liptstueb1}, \cite{LiptstuebAIAA}. We conclude this section with a conjecture. Numerical simulations \cite{LiptstuebAIAA} show that when the microstructure corresponds to smooth inclusions embedded inside a matrix, such as shafts reinforced with long prismatic fibers with circular cross section, then the design method implies full control of the local gradient over the set $S$ i.e., \begin{eqnarray} \Vert\chi^{i^*}_{k_j}(x,n_j x)\nabla u_j(x)\Vert_{L^\infty(S)}\leq M,\hbox{ $i=1,2,\ldots,M$.}\label{ubound1appliedexample} \end{eqnarray} With this in mind and in view of Theorem \ref{periodicidentity} we are motivated to propose the following conjecture. \begin{conjecture} For continuously graded composites containing inclusions with $C^{1,\alpha}$ boundaries for which $A^{k_j}(x,n_j)$ G-converges to $A^H(x)$ then \begin{eqnarray} \limsup_{j\to \infty}\|\chi^{i}_{k_j}(x, n_{j}x)\nabla u_{j}\|_{L^{\infty}(S)} = \Vert {\mathcal M}^i(\nabla u^H)\Vert_{L^\infty(S)}. \label{eq2conj} \end{eqnarray} \end{conjecture} \setcounter{equation}{0} \section*{Appendix } Here we will show that the solutions $w^{i}$ of the cell problem \eqref{cellgraded} satisfying $\int_{Y} w^{i}(x, y)dy = 0$ are in $C(\Omega, H^{1}_{per}(Y) )$ under the continuity assumption \eqref{continuity}. To that end, it suffices to show that as $h\to 0$ \[ \|\nabla_{y}w^{i}(x+h, \cdot) - \nabla _{y}w^{i}(x, \cdot)\|_{L^{2}(Y)}\to 0. \] Since $w^{i}(x+h, y)$ solves equation \eqref{cellgraded} $A(x,y)$ replaced by $A(x+h, y)$, we have that \[ \text{div}~(A(x+h,y)(\nabla_{y}w^{i}(x+h, y) + e^{i}))= \text{div}~(A(x,y)(\nabla_{y}w^{i}(x, y) + e^{i}))= 0. \] Rewriting the above equation we obtain \[ \text{div}~[(A(x+h,y)-A(x,y))](\nabla_{y}w^{i}(x+h, y) + e^{i}) = \text{div}~(A(x,y)(\nabla_{y}w^{i}(x, y) -\nabla_{y}w^{i}(x+h, y)). \] Define the difference mapping $\delta_{h}(F) = F(x+h, y)-F(x,y)$. Then for any $\psi\in H^{1}_{per}(Y)$, we have \begin{equation}\label{continuityDiff} -\int_{Y}(A(x,y)\nabla_{y}\delta_{h}(w^{i})(x, y), \nabla \psi)dx = \int_{Y}(\delta_{h}(A)(x,y)(\nabla_{y}w^{i}(x+h, y) + e^{i}), \nabla \psi). \end{equation} Then plugging $\psi(x,y) = \delta_{h}(w^{i})(x, y) \in H^{1}_{per}(Y)$ in \eqref{continuityDiff} and using the uniform ellipticity of the coefficients, we have \[ \begin{split} \lambda \|\delta_{h}(w^{i})(x, \cdot)\|^{2}_{L^{2}(Y)}&\leq \int_{Y}(\delta_{h}(A)(x,y)[\nabla_{y}w^{i}(x+h, y) + e^{i}], \nabla \delta_{h}(w^{i})(x,y))dy\\ &\leq \left(\int_{Y}|\delta_{h}(A)(x,y)[\nabla_{y}w^{i}(x+h, y) + e^{i}]|^{2}\right)^{1/2}\|\delta_{h}(w)(x,\cdot)\|_{L^{2}(Y)} \end{split} \] The last inequality implies that \[ \|\delta_{h}(w^{i})(x,\cdot)\|_{L^{2}(Y)}\leq\Lambda/\lambda \sum_{i=1}^{N} \left(\int_{Y}|\chi^{i}(x+h,y)-\chi^{i}(x,y)|^{2}|\nabla_{y}w^{i}(x+h, y) + e^{i}|^{2}dy\right)^{1/2} \] By Meyer's higher regularity result, $\nabla_{y}w^{i}(x+h, \cdot)\in L^{p}(Y)$ for some $p>2$. Moreover, the $L^{p}$ norm is bounded from above by a constant $C$ independent of $x$, and $h$. After applying Holder's inequality we get \[ \|\delta_{h}(w^{i})(x,\cdot)\|_{L^{2}(Y)}\leq\frac{C\Lambda}{\lambda} \sum_{i=1}^{N}\left(\int_{Y}|\chi^{i}(x+h,y)-\chi^{i}(x,y)|^{2}dy\right)^{1/2} \] Applying \eqref{continuity}, the right hand side approaches $0$ as $h\to 0$ and the proof is complete. \section*{Acknowledgment} This work is supported by grants: NSF DMS-0807265 and AFOSR FA9550-05-0008. Tadele Mengesha gratefully acknowledges the support of Coastal Carolina University, where he is an assistant professor and currently on leave of absence.

https://arxiv.org/abs/2212.08103

Arithmetic statistics of families of integer Sn-polynomials and application to class group torsion

We study the distributions of the splitting primes in certain families of number fields. The first and main example is the family Pn,N of integer polynomials monic of degree n with height less or equal then N, and then let N go to infinity. We prove an average version of the Chebotarev Density Theorem for this family. In particular, this gives Central Limit Theorem for the number of primes with given splitting type in some ranges. As an application, we deduce some estimates for the torsion in the class groups and for the average of ramified primes.

\section{Introduction} Let $n\ge2$ and $N$ be positive integers. The set $\mathscr{P}_{n,N}$ consists of all monic polynomials $$f(X)=X^n+a_{n-1}X^{n-1}+\dots+a_0$$in $ \mathbb{Z}[X] $, of degree $ n $ and we view the coefficients $ a_0,\dots,a_{n-1} $ as independent, identically distributed random variables taking values uniformly in $ \{-N,\dots,N \} $ with $ N\rightarrow\infty $. We are interested to a subset of $\mathscr{P}_{n,N}$, namely$$\mathscr{P}_{n,N}^{0}=\{f\in \mathscr{P}_{n,N}:G_f\cong S_n\},$$where $ K_f $ is the splitting field of $ f $ inside a fixed algebraic closure $ \overline{\Q} $ of $ \mathbb{Q} $, and $G_f$ is the Galois group of $K_f/\mathbb{Q}$. It is known that "almost all" polynomials $f$ as above have splitting field $K_f$ over $\mathbb{Q}$ with Galois group isomorphic to the symmetric group $S_n$. On the other hand, all $S_n$-extensions of $\mathbb{Q}$ arise in this way for some $f$. \blfootnote{This paper is an extract from my Ph.D. thesis, in which it is trated a more general case, as mentioned in Section 6.3.} The main theorem that we state here is proven in a more general version in Section 6. It is a "Central Limit Theorem". The limit below shows how $\pi_f(x)$ fluctuates about the central value $\pi(x)/n!$, the one expected by the Chebotarev Density Theorem. Let $f\in \mathscr{P}_{n,N}^0$ and let $\pi_f(x)$ the number of primes less or equal than $x$ splitting completely in $K_f/\mathbb{Q}$. Here $\pi(x)$ denotes the number of rational primes less or equal than $x$. \begin{thm2} For $x=N^{1/\log\log N}$ (in particular $x\ll N^{\varepsilon}$ for every $\varepsilon>0$) and for any real numbers $a<b$ one has$$\frac{1}{|\mathscr{P}_{n,N}^0|}\Big|\Big\{f\in\mathscr{P}_{n,N}^0:a\le\Big(\frac{1}{n!}-\frac{1}{n!^2}\Big)^{-1/2}\Big(\frac{\pi_f(x)-\frac{1}{n!}\pi(x)}{\pi(x)^{1/2}}\Big)\le b\Big\}\Big|$$ $$\underset{N\rightarrow+\infty}{\longrightarrow}\frac{1}{\sqrt{2\pi}}\int_{a}^{b}e^{-t^2/2}dt.$$ \end{thm2} \noindent \textbf{Example.} For $a=-0.9$ and $b=0.9$ the above integral is about $0.6$. If $n=3$ and $N$ is near $10^2$ (so $x$ near $20$), the ratio is approximately $\pi_f(x)-1.11$, which lies in the interval $[-0.9,0.9]$ if and only if $\pi_f(x)$ is in $[0.21,2.01]\sim[0,3]$. This means that the proportion of cubic $S_3$-polynomials with integer coefficients in a box $[-100,100]$ having less than 3 primes below 20 splitting completely in their splitting field, is about 60 percent.\\ This paper is an extract from my Ph.D. thesis, in which it is trated a more general case (see a brief note about in Section 6.3). \subsection*{Application 1} The main application we discuss here concerns some bounds for the $\ell$-torsion part of the class number of the above number fields, which improve on the trivial bound provided by the order of the entire class group. The class number of a number field $K/\mathbb{Q}$, that we denote by $h_K$, is by definition the order of the ideal class group of its ring of integers, that is, the abelian quotient group of the the group of fractional ideals and its subgroup of principal ideals. It measures the "distance" of a ring of integers to be a unique factorization domain; namely, this turns out to be the case if and only if the class number is 1. \begin{coroll2} Let $D_f$ be the discriminant of $K_f/\mathbb{Q}$ for $f\in\mathscr{P}_{n,N}^0$. Then for every positive integer $\ell$, for all $ f \in\mathscr{P}_{n,N}^0$ outside of a set of size $ o(N^n) $, the torsion part of the class group of $K_f/\mathbb{Q}$ is bounded by$$h_{K_f}[\ell]\ll_{n,\ell,\varepsilon}D_f^{\frac{1}{2}-\frac{1}{(2n-2)(n-1)!\log\log d_f}+\varepsilon},$$for every $\varepsilon>0$, as $ N\rightarrow+\infty $. \end{coroll2} This will be proved in Section 7, in which we also state the crucial result of Ellenberg and Venkatesh in order to show the corollary. \subsection*{Application 2} Let $f$ be an $S_n$-polynomial and let $d_f$ be its discriminant. The relation between $d_f$ and the discriminant $D_f$ of its splitting field is still an open problem in many cases, and leads to difficulties when counting ramified primes. Call an irreducible monic integral polynomial $f$ \textit{essential} if the equality between the two discriminant holds. It is well known that this implies that the ring of integers of the splitting field of $f$ is monogenic. Our results can be applied to study this relation, and to bound on average the number of primes dividing the discriminant. \begin{coroll}For almost all $f\in\mathscr{P}_{n,N}^0$, the number of ramified primes is$$\ll_{n}\log\log N,$$as $N\rightarrow+\infty$. \end{coroll} See Section 3.1 for this discussion. \subsection*{Notations and background} For the sake of clarity, we recall that given a normal extension $K$ over $\mathbb{Q}$ of degree $s$, the \textit{Galois group} $G_K=G_{K/\mathbb{Q}}$ of $K$ over $\mathbb{Q}$ is defined to be the group of automorphisms of $K$ that fix $\mathbb{Q}$ pointwise. There is a natural embedding$$G_K\hookrightarrow S_s$$given by the action of the Galois group on the $s$ homomorphisms of $K$ into $\overline{\mathbb{Q}}$. In the following we'll identify $G_K$ with its image via the above morphism. If $p$ is an unramified prime in $K/\mathbb{Q}$, i.e. the inertia group for every prime $\mathfrak{p}$ of $K$ is trivial, there is a canonical isomorphism between the Galois group of the residue fields exension $\mathcal{O}_K/\mathfrak{p}/\mathbb{F}_p$ and the decomposition group $D_{\mathfrak{p}}$ at $\mathfrak{p}$. Now, $G_{\mathcal{O}_K/\mathfrak{p}/\mathbb{F}_p}$ is cyclic with canonical generator the Frobenius at $p$. The corresponding to $\mathfrak{p}$ is $\Frob_{K/\mathbb{Q},\mathfrak{p}|p}\in D_{\mathfrak{p}}$. It is the unique element of $G_K$ such that for all $\alpha\in\mathcal{O}_K$ we have$$\Frob_{K/\mathbb{Q},\mathfrak{p}|p}(\alpha)\equiv \alpha^p\mod\mathfrak{p}.$$If we consider another prime over $p$, that is a conjugated one through an element of the Galois group, the new Frobenius is conjugated to the previuous one via the same automorphism. Hence we denote by $\Frob_{K/\mathbb{Q},p}$ the \textit{Frobenius element} at $p$, namely the conjugacy class of Frobenius automorphisms in $G_K$. \subsubsection*{Counting $S_n$-polynomials} Regarding polynomials $f\in\mathscr{P}_{n,N}$, it is well known that$$\lim_{N\rightarrow\infty}\mathbb{P}(\mbox{f irreducible})=1.$$The error term, given by Kuba \cite{Ku} is $ O(N^{-1}) $ for $ n>2 $. Van der Waerden \cite{Wa} proved that $$\frac{|\mathscr{P}_{n,N}^{0}|}{|\mathscr{P}_{n,N}|}\underset{N\rightarrow+\infty}{\longrightarrow}1.$$with explicit error term $ O(N^{-1/6}) $. It has improved in \cite{Gal} using large sieve to $ O(N^{-1/2}\log N) $, and more recently by Dietmann \cite{Di} using resolvent polynomials to $O(N^{-2+\sqrt{2}+\varepsilon})$. The best estimate can be found in \cite{Bh1}, who proved the following result, conjectured by van der Waerden. \begin{namedthm*}{Theorem}[Bhargava] If either $n=3,4$ or $n\ge6$, one has, $$|\mathscr{P}_{n,N}^{0}(\mathbb{Q})|=(2N)^n+O(N^{n-1}),$$as $ N\rightarrow\infty $. \end{namedthm*} \subsubsection*{The number of splitting primes} Let $ f\in \mathscr{P}_{n,N}^{0} $. We say that $ r=(r_1,r_2,\dots,r_n) $ is the \textbf{splitting type} of $ f$ mod a prime $ p $ if $ f\mod p $ splits into distinct monic irreducible factors (so a square-free factorization), with $ r_1 $ linear factors, $ r_2 $ quadratic factors and so on. For the primes $ p $ that not divide the discriminant $ D_f $ of the extension $ K_f/\mathbb{Q} $, $ r $ corresponds to the cycle structure of the Frobenius element $ \Frob_{K_f/\mathbb{Q},p}=:\Frob_{f,p} $ acting on the roots of $ f $. For each $ r $ we have $$\sum_{i=1}^{n}ir_i=n.$$Let $ \mathscr{C}_r $ be the conjugacy class in $ S_n $ of elements of cycle type $ r $; then the order of $ \mathscr{C}_r $ is $ n!\delta(r) $, where $ \delta(r)=\prod_{i=1}^{n}\frac{1}{i^{r_i}r_i!} $. In general, for a subgroup $G\subseteq S_n$, the elements in a conjugacy class in $G$ necessarily have the same cycle type, but the converse need not to be true. That is, the cycle type of a conjugacy class in $G$ need not determine in uniquely. This uniqueness property does hold for cycle types for the full symmetric group, which implies that the cycle type of an $S_n$-polynomial having a square-free factorization mod $p$ uniquely determines the Frobenius element for an $S_n$-number field obtained by adjoining one root of it. If $p$ is a prime number not dividing the discriminant of $f$, then we can write $f$ modulo $p$ as a product of distinct irreducible factors over the finite field $\mathbb{F}_p$. The degrees of these irreducible factors form the splitting type of $f$ mod $p$; this is also a partition of $n$. Frobenius's theorem asserts, roughly speaking, that the number of primes with a given decomposition type is proportional to the number of permutations in $S_n$ with the same cycle pattern. \begin{namedthm*}{Theorem}[Frobenius, 1880] The density of the set of primes $p$ for which $f$ has a given splitting type $r_1,\dots,r_n$ exists, and it is equal to $1/|G_f|$ times the number of $\sigma\in G_f$ with cycle pattern $r_1,\dots,r_n$. \end{namedthm*} We want to compute the density (on average) of primes $p$ unramified in $K_f/\mathbb{Q}$ for which $f$ has a given splitting type modulo $p$. This is a special case of the Chebotarev density theorem, for which we want an explicit asymptotic, with an effective error term, for "almost all" polynomials in our family. For $ x\ge1 $, let$$\pi_{f,r}(x)=\underset{f\tiny{\mbox{ splitting type }}r\tiny{\mbox{ mod }} p}{\sum_{p\le x}}1=\sum_{p\le x}\mathbbm{1}_{f,r}(p).$$We may view $ \pi_{f,r}(x) $ as a sum of random variables on $ \mathscr{P}_{n,N}^{0} $, seen as a sublattice of $ [-N,N]^n $. For a fixed $ r $, we abbreviate $ \mathbbm{1}_{f,r}(p) $ with $ \mathbbm{1}_{p}(f) $:$$\mathbbm{1}_{p}:\mathscr{P}_{n,N}^{0}\longrightarrow\{0,1\}.$$ For all $ N $, let $ \mathbb{P}_N $ be the probability measure on $ [-N,N]^n $. We'll denote by $\mathbb{E}_N$ and $ \sigma^{2}_N $ the expectation and the variance of a random variable on $ \mathscr{P}_{n,N}^{0} $, respectively. \section{Strategy of the proofs} For every splitting type $ r $, let \begin{align*} X_{n,r,p}=\Big\{&\Big( \prod_{i=1}^{r_1}g_i^{(1)}\Big) \dots\Big( \prod_{i=1}^{r_n}g_i^{(n)}\Big):g_i^{(j)}\in\mathbb{F}_p[X]\ \mbox{irreducile, monic},\\ &\deg(g_i^{(j)})=j,\ g_i^{(j)}\neq g_k^{(j)}\mbox{ if }i\neq k\Big\}. \end{align*} The following key fact is what we'll use to estimate the error term in the asymptotic of the expectation $\mathbb{E}_N(\pi_{f,r}(x))$ of $\pi_{f,r}(x)$ and its powers. The fact that the primes $p$ are small enough with respect to $N$, allows to have "many" $S_n$-polynomials congruent to given polynomials mod $p$. \begin{lemma} Let $k\ge 1$, $p_1,\dots, p_k$ primes and $g_i\in X_{n,r,p_i}$ for all $i=1,\dots,k$. Then if $p_i<N^{1/kn}$ for all $i=1,\dots,k$ $$\mathbb{P}_N(f\in\mathscr{P}_{n,N}^{0}:f\equiv g_i\mod p_i\ \ \forall i=1,\dots,k)=\frac{1}{(p_1\dots p_k)^n}+O_n(N^{-1}),$$as $N\rightarrow+\infty$. \end{lemma} \begin{proof}We prove the case $k=1$. An application of the Chinese Remainder Theorem lead to the result for $k>1$.\\ If $ g=\sum_{i=1}^{n}g_iX^i $ and $ f=\sum_{i=1}^{n}f_iX^i $, then $ f\equiv g\mod p $ if and only if $ f_i\equiv g_i\mod p $ for $ i=0,\dots,n-1 $. This means $ f_i=g_i+pk_i $ for some $ k_i\in\mathbb{Z} $ and $ -N\le f_i\le N $. Therefore$$\frac{-N-g_i}{p}\le k_i\le\frac{N-g_i}{p};$$so for each coefficient $ f_i $ of $ f $ we have$$\Big[\frac{N-g_i}{p}\Big]-\Big[\frac{-N-g_i}{p}\Big]=\frac{2N}{p}+O(1)$$choices. Hence$$|\{f\in\mathscr{P}_{n,N}:f\equiv g\mod p\}|=\frac{(2N)^n}{p^n}+O(N^{n-1}),$$so\begin{align*} |\{f\in\mathscr{P}_{n,N}^0:f\equiv g\mod p\}|&=\underset{f\equiv g\tiny\mbox{ mod }p}{\sum_{f\in\mathscr{P}_{n,N}}}1+O\Big(\sum_{f\notin\mathscr{P}_{n,N}^{0}}1\Big)\\&=\frac{(2N)^n}{p^n}+O(N^{n-\xi+\varepsilon}). \end{align*} As long as $ p^n<N $, we get\begin{align*} \frac{1}{|\mathscr{P}_{n,N}^{0}|}\underset{f\equiv g\tiny\mbox{ mod }p}{\sum_{f\in\mathscr{P}_{n,N}^{0}}}1&=\frac{1}{(2N)^n}(1+O(N^{n-1}))\Big(\frac{(2N)^n}{p^n}+O(N^{n-1})\Big)\\ &=(1+O(N^{-1}))\Big(\frac{1}{p^n}+O(N^{-1})\Big)\\ &=\frac{1}{p^n}+O(N^{-1}). \end{align*} \end{proof} \section{Prime splitting densities} \begin{prop} One has, for all primes $ p<N^{1/(n+1)} $, \begin{enumerate} \item[$\paruno1\pardue$] $ \mathbb{P}_N(\mathbbm{1}_{p}=1)=\mathbb{E}_N(\mathbbm{1}_{p})=\delta(r)+\frac{C_{r}}{p}+O\Big(\frac{1}{p^2}+p^nN^{-1}\Big) $,\\for some constant $ C_{r}$; \item[$\paruno2\pardue$] $ \sigma^2_N(\mathbbm{1}_{p})=(\delta(r)-\delta(r)^2)+\frac{C_{r}(1-2\delta(r))}{p}+O\Big(\frac{1}{p^2}+p^nN^{-1}\Big) $. \end{enumerate}It follows that, for $ x<N^{1/(n+1)} $ \begin{enumerate} \item[$\paruno3\pardue$] $ \mathbb{E}_N(\pi_{f,r}(x))=\delta(r)\pi(x)+C_{r}\log\log x+O_n(1)$ \end{enumerate} \end{prop} Hence, the \textit{normal order} of $ \pi_{f,r}(x) $ is $ \delta(r)\pi(x) $, which means that $ \pi_{f,r}(x)\sim \delta(r)\pi(x)$ for almost all $ f $'s, as $ x\rightarrow+\infty $ and $ N $ large enough. Part (3) of the above proposition is an average version of the Chebotarev Density Theorem. \begin{proof} Once fixed a prime $ p $, \begin{align*} \mathbb{E}_N(\mathbbm{1}_{p})&=\frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}\mathbbm{1}_p(f)\\ &=\frac{1}{|\mathscr{P}_{n,N}^{0}|}\underset{f\tiny{\mbox{ of splitting type }}r\tiny{\mbox{ mod }} p}{\sum_{f\in\mathscr{P}_{n,N}^{0}}}1\\ &=\frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{g\in X_{n,r,p}}\underset{f\equiv g\tiny\mbox{ mod }p}{\sum_{f\in\mathscr{P}_{n,N}^{0}}}1. \end{align*} On the other hand,$$|X_{n,r,p}|=\prod_{k=1}^{n}\binom{A_{p,k}}{r_k},$$where $ A_{p,k} $ is the number of degree-$ k $ irreducible polynomials in $ \mathbb{F}_p[X] $, which, by the M\"{o}bius inversion formula, equals$$\frac{1}{k}\sum_{d|k}\mu(d)p^{k/d}=\frac{p^k}{k}+O(p^{\alpha_k}),$$where $\alpha_k=1$ if $k=2$, and $\alpha_k< k-1$ if $k>2$. One has, for all $ k\ge2$\begin{align*} \binom{A_{p,k}}{r_k}&=\frac{A_{p,k}(A_{p,k}-1)\dots(A_{p,k}-r_k+1)}{r_k!}\\ &=\frac{1}{r_k!}\Big(\frac{p^k}{k}+O(p^{\alpha_k})\Big)\dots\Big(\frac{p^k}{k}-r_k+1+O(p^{\alpha_k})\Big). \end{align*}It turns out that$$ \binom{A_{p,k}}{r_k}=\begin{cases} \frac{1}{r_1!}p(p-1)\dots(p-r_1+1)&\mbox{ if }k=1\\ \frac{1}{r_2!2^{r_2}}p^{2r_2}+C(r_2)p^{2r_2-1}+O(p^{2r_2-2})&\mbox{ if }k=2\\ \frac{1}{r_k!k^{r_k}}p^{kr_k}+O(p^{k(r_k-1)+\alpha_k})&\mbox{ if }k>1. \end{cases} $$Hence\begin{multline*} |X_{n,r,p}|=\frac{1}{r_1!}p(p-1)\dots(p-r_1+1)\frac{1}{r_2!2^{r_2}}(p^{2r_2}+C(r_2)p^{2r_2-1}+O(p^{2r_2-2}))\\\prod_{k=3}^{n}(\frac{1}{r_k!k^{r_k}}p^{kr_k}+O(p^{k(r_k-1)+\alpha_k}))\\ =\delta(r)p^{n}+C_rp^{n-1}+O(p^{n-2}), \end{multline*}where $ C_r:=-\delta(r)C(r_2)\frac{(r_1+1)(r_1+2)}{2r_1!}.$ By Lemma 1, for $ p^{n+1}<N$\begin{align*} \mathbb{E}_N(\mathbbm{1}_{p})&=(\delta(r)p^n+C_{r}p^{n-1}+O(p^{n-2}))\Big(\frac{1}{p^n}+O(N^{-1})\Big)\\ &=\delta(r)+\frac{C_{r}}{p}+O\Big(\frac{1}{p^2}+p^nN^{-1}\Big), \end{align*}which proves (1) and (2) follows by definition. For (3), by linearity, we simply have to sum over all primes $ p\le x $ and use the elementary estimate$$\sum_{p\le x}\frac{1}{p}=\log\log x+O(1)$$to get$$\mathbb{E}_N(\pi_{f,r}(x))=\delta(r)\pi(x)+C_{r}\log\log x+O(1+\pi(x)^{n+1}N^{-1})$$as long as$$\pi(x)^{n+1}N^{-1}=o(\log\log x).$$If moreover $ x<(N^{\xi-\varepsilon})^{1/(n+1)} $, then the term $ \pi(x)^{n+1}N^{-1} $ is negligible. \end{proof} \begin{flushleft} \textbf{Remark.} From (3) of Proposition 1, we have that for every $m\ge2$,$$\pi_{f,r}(x)-\delta(r)\pi(x)=O((\log\log x)^m)$$as $x\rightarrow+\infty$, for all but $O_{n}\Big(x^{(n+1)(n-1)}(\log\log x)^{1-m}\Big)$ $S_n$-polynomials $f$ of height $\ll x^{n+1}$. \end{flushleft} For $f\in\mathscr{P}_{n,N}^{0}$, Let $\varphi:G_f\rightarrow\mathbb{C}$ be a central function, i.e. constant on conjugacy classes. Define$$\pi_{f,\varphi}(x)=\sum_{p\le x,\ p\nmid D_f}\varphi(\Frob_{f,p}).$$Then, if we sum over the conjugacy classes, that is over the splitting types $r=(r_1,\dots,r_n)$ we get$$\pi_{f,\varphi}(x)=\sum_{r}\varphi(g_r)\pi_{f,r}(x),$$where $g_r$ is any element of the conjugacy class $\mathscr{C}_r$ for every $r$. \begin{coroll} If $x<N^{1/(n+1)}$,$$\mathbb{E}_N(\pi_{f,\varphi}(x))=\sum_{r}\delta(r)\varphi(g_r)\pi(x)+\sum_{r}\delta(r)\varphi(g_r)\log\log x+O(||\varphi||)$$where $||\varphi||=\sup_{g\in G_f}|\varphi(g)|$. \end{coroll} \subsection{Higher moments} The following approach is motivated by the proof of the Erd\H{o}s-Kac Theorem in \cite{GS}, where they compute the moments$$\sum_{n\le x}(\omega(n)-\log\log x)^k$$of the prime divisor function, uniformely in a wide range of $ k $. Fix a splitting type $ r $ and a prime $ p $. Consider the independent discrete random variables $ X_p $ defined by$$\mathbb{P}(X_p=1)=\frac{|X_{n,r,p}|}{p^n}.$$So \begin{equation} \mathbb{P}(X_p=1)=\frac{|X_{n,r,p}|}{p^n}=\delta(r)+\frac{C_{r}}{p}+O\Big(\frac{1}{p^2}\Big) \end{equation}For all prime $p$ we define the function$$Y_p(f)=\begin{cases} 1-\frac{|X_{n,r,p}|}{p^n}&\mbox{ if }\mathbbm{1}_p(f)=1\\ -\frac{|X_{n,r,p}|}{p^n}&\mbox{ otherwise }. \end{cases}$$Now, we consider a generalization $Y_a$ of the function $Y_p$ for any natural number $a>0$, whose $ k $-moments are "small" unless $ a $ satisfies the following property ($ \textasteriskcentered $): $ p^\alpha||a\Rightarrow\alpha\ge 2 $ (see the next Lemma). Let $ a=\prod_{i=1}^{s}p_i^{\alpha_i} $, where the $ p_i $ are distinct primes and $ \alpha_i\ge1 $. Let $ A:=\prod_{i=1}^{s}p_i $ be the square-free part of $ a $, and set$$Y_a(f):=\prod_{i=1}^{s}Y_{p_i}(f)^{\alpha_i}.$$ \begin{lemma} Uniformly for \textbf{even} natural numbers $ k $ with $ k\ll_{n,r}\frac{\log N}{\log z} $, one has\begin{multline*} \mathbb{E}_N\Big( \Big( \sum_{p\le z}Y_p(f)\Big) ^k\Big) =C_{k,r}\pi(z)^{k/2}\Big(1+O\Big(\frac{k^3}{(1-\delta(r))^{k/2}}\frac{\log\log z}{\pi(z)}\Big)\Big)\\+O(\pi(z)^{k(n+1)}N^{-1}), \end{multline*}as $ z,N\rightarrow+\infty $. While uniformly for \textbf{odd} natural numbers $ k $ with $k\ll_{n,r}\frac{\log N}{\log z}$, one has $$\mathbb{E}_N\Big( \Big( \sum_{p\le z}Y_p(f)\Big) ^k\Big)\ll C_{k,r}\pi(z)^{k/2}k\frac{\log\log z}{\pi(z)^{1/2}}+\pi(z)^{k(n+1)}N^{-1}$$as $ z,N\rightarrow+\infty $. Here$$C_{k}=\begin{cases} \frac{k!}{2^{k/2}(k/2)!}&\mbox{ for }k\mbox{ even}\\ \frac{k!}{2^{\frac{k-1}{2}}(\frac{k-1}{2})!}&\mbox{ for }k\mbox{ odd}, \end{cases}$$and$$C_{k,r}=\begin{cases} C_k(\delta(r)-\delta(r)^2)^{k/2}&\mbox{ for }k\mbox{ even}\\ C_k\delta(r)^{\frac{k-1}{2}}&\mbox{ for }k\mbox{ odd}. \end{cases}.$$ \end{lemma} Observe that, for any real number $ z<x $, we can write\begin{align*} \pi_{f,r}(x)-\delta(r)\pi(x)&=\sum_{p\le z}Y_p(f)+\sum_{z<p\le x}\mathbbm{1}_p(f)\\&+\Big(\sum_{p\le z}\frac{|X_{n,r,p}|}{p^n}-\delta(r)\pi(x)\Big). \end{align*} Pick $z=x-k$. Since\begin{align*} \pi(z)&\sim\frac{x-k}{\log x\left(1+\frac{\log\left( 1-\frac{k}{x}\right)}{\log x}\right) }\\ &=\frac{x}{\log x}+O\left( \frac{k}{\log x}\right), \end{align*}by the above one has$$\pi_{f,r}(x)-\delta(r)\pi(x)=\sum_{p\le z}Y_p(f)+O_r\left( \frac{k}{\log x}\right).$$ \begin{proof} We may write$$\frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}(\sum_{p\le z}Y_p(f))^k=\sum_{p_1,\dots,p_k\le z}\frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}Y_{p_1\dots p_k}(f).$$Let us then consider more generally $ \frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}Y_a(f) $. By definition, for any prime $ p $, $ Y_p(f)=Y_p(g) $ if $ f\equiv g $ mod $ p $; therefore\begin{align*} \frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}Y_a(f)&=\frac{1}{|\mathscr{P}_{n,N}^{0}|}\underset{i=1,\dots,s}{\sum_{g_i\tiny{\mbox{ mod }}p_i}}\underset{f\equiv g_i\tiny{\mbox{ mod }}p_i\ \forall i}{\sum_{f\in\mathscr{P}_{n,N}^{0}}}Y_{p_1}(g_1)^{\alpha_1}\dots Y_{p_s}(g_s)^{\alpha_s}, \end{align*}where the first sum in the right-hand side is over $ g_i\in\mathbb{Z}[X] $ monic, with coefficients in $ \{0,\dots,p_i-1\} $. As long as $ (p_1\dots p_s)^n<N $ the sum is, by Lemma 1,\begin{multline*} \sum_{g_1,\dots,g_s}Y_{p_1}(g_1)^{\alpha_1}\dots Y_{p_s}(g_s)^{\alpha_s} \frac{1}{|\mathscr{P}_{n,N}^{0}|}\underset{f\equiv g_i\tiny{\mbox{ mod }}p_i\ \forall i}{\sum_{f\in\mathscr{P}_{n,N}^{0}}}1\\ =\sum_{g_1,\dots,g_s}Y_{p_1}(g_1)^{\alpha_1}\dots Y_{p_s}(g_s)^{\alpha_s}\Big(\frac{1}{(p_1\dots p_s)^n}+O(N^{-1})\Big)\\ =\frac{1}{A^n}\sum_{g_1,\dots,g_s}Y_{p_1}(g_1)^{\alpha_1}\dots Y_{p_s}(g_s)^{\alpha_s} + O\Big(N^{-1}\sum_{g_1,\dots,g_s}1\Big)\\ =\frac{1}{A^n}\sum_{g_1,\dots,g_s}Y_{p_1}(g_1)^{\alpha_1}\dots Y_{p_s}(g_s)^{\alpha_s} + O(A^nN^{-1}), \end{multline*}since $ |Y_{p_i}(g_i)^{\alpha_i}|\ll 1 $. Denoting the main term by $ Y(a) $, we have\begin{multline*} Y(a)=\frac{1}{A^n}\sum_{g_1}Y_{p_1}(g_1)^{\alpha_1}\dots \sum_{g_s}Y_{p_s}(g_s)^{\alpha_s}\\ =\frac{1}{A^n}\prod_{i=1}^{s}\Big(\sum_{g_i\in X_{n,r,p_i}}\Big(1-\frac{|X_{n,r,p_i}|}{p_i^n}\Big)^{\alpha_i}+\sum_{g_i\notin X_{n,r,p_i}}\Big(-\frac{|X_{n,r,p_i}|}{p_i^n}\Big)^{\alpha_i}\Big)\\ =\frac{1}{A^n}\prod_{i=1}^{s}\Big(|X_{n,r,p_i}|\Big(1-\frac{|X_{n,r,p_i}|}{p_i^n}\Big)^{\alpha_i}+(p_i^{n}-|X_{n,r,p_i}|)\Big(-\frac{|X_{n,r,p_i}|}{p_i^n}\Big)^{\alpha_i}\Big)\\ =\prod_{p^{\alpha}||a}\Big(\frac{|X_{n,r,p}|}{p^n}\Big(1-\frac{|X_{n,r,p}|}{p^n}\Big)^{\alpha}+\Big(1-\frac{|X_{n,r,p}|}{p^n}\Big)\Big(-\frac{|X_{n,r,p}|}{p^n}\Big)^{\alpha}\Big), \end{multline*}using the inductive formula $$\prod_{i=1}^{\ell}(a_i+b_i)=\underset{k+h=\ell}{\underset{j_1<\dots<j_h\le\in\{1,\dots,\ell\}\setminus\{i_1,\dots,i_k\}}{\sum_{1\le i_1<\dots<i_k\le \ell}}}a_{i_1}\dots a_{i_k}b_{j_1}\dots b_{j_h}.$$ Thus\begin{equation*} \frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}Y_a(f)=Y(a)+O(A^nN^{-1}); \end{equation*}Observe now that $ Y(a)=0 $ unless $ \alpha_i\ge 2 $ for all $ i=1,\dots,s $. It turns out that\begin{multline*} \frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}(\sum_{p\le z}Y_p(f))^k=\underset{p_1\dots p_k\ (\textasteriskcentered)}{\sum_{p_1,\dots,p_k\le z}}Y(p_1\dots p_k)\\+O\Big(\sum_{p_1,\dots,p_k\le z}(p_1\dots p_k)^nN^{-1}\Big)\\ =\underset{p_1\dots p_k\ (\textasteriskcentered)}{\sum_{p_1,\dots,p_k\le z}}Y(p_1\dots p_k)+O(\pi(z)^{k(n+1)}N^{-1}). \end{multline*}Let $ q_1<\dots<q_s $ be the distinct primes in $ p_1\dots p_k $. Since $ p_1\dots p_k $ satisfies $ (\textasteriskcentered) $, we have $ s\le k/2 $. The main term above is\begin{equation} \sum_{s\le k/2}\sum_{q_1<\dots<q_s\le z}\underset{\alpha_1+\dots+\alpha_s=k}{\sum_{\alpha_1,\dots,\alpha_s\ge2}}\binom{k}{\alpha_1,\dots,\alpha_s}Y(q_1^{\alpha_1}\dots q_s^{\alpha_s}). \end{equation}At this point, we have to divide into two cases, since if $ k $ is even there is a term $ s=k/2 $ with all $ \alpha_i=2 $. This main term contributes\begin{multline*} \frac{k!}{2^{k/2}(k/2)!}\underset{q_j\tiny{\mbox{ distinct}}}{\sum_{q_1,\dots,q_{k/2}\le z}}Y(q_1^2\dots q_{k/2}^2)\\ =\frac{k!}{2^{k/2}(k/2)!}\underset{q_j\tiny{\mbox{ distinct}}}{\sum_{q_1,\dots,q_{k/2}\le z}}\prod_{i=1}^{k/2}\frac{|X_{n,r,q_i}|}{q_i^n}\Big(1-\frac{|X_{n,r,q_i}|}{q_i^n}\Big). \end{multline*}Now, clearly$$\underset{q_j\tiny{\mbox{ distinct}}}{\sum_{q_1,\dots,q_{k/2}\le z}}\prod_{i=1}^{k/2}\frac{|X_{n,r,q_i}|}{q_i^n}\Big(1-\frac{|X_{n,r,q_i}|}{q_i^n}\Big)\le\Big(\sum_{p\le z}\frac{|X_{n,r,p}|}{p^n}\Big(1-\frac{|X_{n,r,p}|}{p^n}\Big)\Big)^{k/2}.$$On the other hand, by induction\begin{multline*} \underset{q_j\tiny{\mbox{ distinct}}}{\sum_{q_1,\dots,q_{k/2}\le z}}\prod_{i=1}^{k/2}\frac{|X_{n,r,q_i}|}{q_i^n}\Big(1-\frac{|X_{n,r,q_i}|}{q_i^n}\Big)\\ =\underset{q_j\tiny{\mbox{ distinct}}}{\sum_{q_1,\dots,q_{k/2-1}\le z}}\prod_{i=1}^{k/2-1}\frac{|X_{n,r,q_i}|}{q_i^n}\Big(1-\frac{|X_{n,r,q_i}|}{q_i^n}\Big)\underset{q_{k/2}\neq q_j\ \forall j}{\sum_{q_{k/2}\le z}}\frac{|X_{n,r,q_{k/2}}|}{q_{k/2}^n}\Big(1-\frac{|X_{n,r,q_{k/2}}|}{q_{k/2}^n}\Big)\\ \ge\underset{q_j\tiny{\mbox{ distinct}}}{\sum_{q_1,\dots,q_{k/2-1}\le z}}\prod_{i=1}^{k/2-1}\frac{|X_{n,r,q_i}|}{q_i^n}\Big(1-\frac{|X_{n,r,q_i}|}{q_i^n}\Big)\sum_{\pi_{k/2}\le p\le z}\frac{|X_{n,r,p}|}{p^n}\Big(1-\frac{|X_{n,r,p}|}{p^n}\Big)\\ \ge\dots\ge \sum_{ 2\le p\le z}\frac{|X_{n,r,p}|}{p^n}\Big(1-\frac{|X_{n,r,p}|}{p^n}\Big)\dots\sum_{\pi_{k/2}\le p\le z}\frac{|X_{n,r,p}|}{p^n}\Big(1-\frac{|X_{n,r,p}|}{p^n}\Big)\\ \ge \Big(\sum_{\pi_{k/2}\le p\le z}\frac{|X_{n,r,p}|}{p^n}\Big(1-\frac{|X_{n,r,p}|}{p^n}\Big)\Big)^{k/2}, \end{multline*}where $\pi_n$ denotes the $n$-th smallest prime. By (1)\begin{align*} \sum_{p\le z}\frac{|X_{n,r,p}|}{p^n}\Big(1-\frac{|X_{n,r,p}|}{p^n}\Big)&=(\delta(r)-\delta(r)^2)\pi(z)+O(\log\log z),\\ \sum_{\pi_{k/2}\le p\le z}\frac{|X_{n,r,p}|}{p^n}\Big(1-\frac{|X_{n,r,p}|}{p^n}\Big)&=(\delta(r)-\delta(r)^2)\pi(z)+O(\log\log z+k). \end{align*}The main term in (2) is then\begin{multline*} \frac{k!}{2^{k/2}(k/2)!}((\delta(r)-\delta(r)^2)\pi(z)+O(\log\log z+k))^{k/2}\\ =\frac{k!}{2^{k/2}(k/2)!}(\delta(r)-\delta(r)^2)^{k/2}(\pi(z)^{k/2}+O(k^2\pi(z)^{k/2-1}\log\log z)). \end{multline*}We have now to estimate the error term in (2), for $ s=k/2-1$. Since $ Y(q_1^{\alpha_1}\dots q_s^{\alpha_s})\le\frac{|X_{n,r,q_1}|\dots|X_{n,r,q_s}|}{(q_1\dots q_s)^n} $ one has \begin{multline*} \sum_{q_1<\dots<q_s\le z}\underset{\alpha_1+\dots+\alpha_s=k}{\sum_{\alpha_1,\dots,\alpha_s\ge2}}\binom{k}{\alpha_1,\dots,\alpha_s}Y(q_1^{\alpha_1}\dots q_s^{\alpha_s})\\ \le\frac{k!}{(k/2-1)!}\Big(\sum_{q\le z}\frac{|X_{n,r,q}|}{q^n}\Big)^{k/2-1}\underset{\alpha_1+\dots+\alpha_{k/2-1}=k}{\sum_{\alpha_1,\dots,\alpha_{k/2-1}\ge 2}}\frac{1}{\alpha_1!\dots\alpha_{k/2-1}!}\\ \le\frac{k!}{2^{k/2-1}(k/2-1)!}\binom{k/2}{k/2-2}(\delta(r)\pi(z)+O(\log\log z))^{k/2-1} \\ \ll \frac{k!}{2^{k/2}(k/2)!} k^3\left( \delta(r)^{k/2-1}\pi(z)^{k/2-1}+k\pi(z)^{k/2-2}\log\log z\right)\\ \ll \frac{k!}{2^{k/2}(k/2)!} k^3\delta(r)^{k/2-1}\pi(z)^{k/2-1}. \end{multline*}For the last inequality, we use the fact that the number of sequences of integers $ (\alpha_1,\dots,\alpha_{k/2-1}) $, $ \alpha_i\ge 2 $ such that $ \sum\alpha_i=k $ is the number of sequences $ (\alpha_1',\dots,\alpha_{k/2-1}') $, $ \alpha_i'\ge 1 $ such that $ \sum\alpha_i=k/2+1 $, that is the number of strong compositions of $ k/2+1 $ into $ k/2-1 $ parts, which is $ \binom{k/2}{k/2-2} $. Thus, for $ k $ even,\begin{multline*} \frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}(\sum_{p\le z}Y_p(f))^k\\ = \frac{k!}{2^{k/2}(k/2)!}(\delta(r)-\delta(r)^2)^{k/2}(\pi(z)^{k/2}+O(k^2\pi(z)^{k/2-1}\log\log z\\+\frac{k^3}{(1-\delta(r))^{k/2}}\pi(z)^{k/2-1})) +O(\pi(z)^{k(n+1)}N^{-1})\\ =\frac{k!}{2^{k/2}(k/2)!}(\delta(r)-\delta(r)^2)^{k/2}\pi(z)^{k/2}\Big(1+O\Big(\frac{k^3}{(1-\delta(r))^{k/2}}\frac{\log\log z}{\pi(z)}\Big)\Big)\\+O(\pi(z)^{k(n+1)}N^{-1}). \end{multline*}Finally, for $ k $ odd, we have the estimate for the term with $ s=k/2-1/2 $ as for the previous case, obtaining\begin{multline*} \frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}\left( \sum_{p\le z}Y_p(f)\right) ^k\\ \ll \frac{k!}{2^{\frac{k-1}{2}}(\frac{k-1}{2})!}k\left(\delta(r)^{\frac{k-1}{2}}\pi(z)^{\frac{k-1}{2}}+O(k\pi(z)^{\frac{k-3}{2}}\log\log z)\right) \\+\pi(z)^{k(n+1)}N^{-1}\\ \ll C_{k,r}\pi(z)^{k/2}k\frac{\log\log z}{\pi(z)^{1/2}}+\pi(z)^{k(n+1)}N^{-1}. \end{multline*} \end{proof} \begin{prop} Uniformly for \textbf{even} natural numbers $ k $ with $ k\ll_{n,r}\frac{\log N}{\log x} $, one has\begin{multline*} \mathbb{E}_N((\pi_{f,r}(x)-\delta(r)\pi(x))^k)\\=C_{k,r}\pi(x)^{k/2}\left( 1+O\left( \frac{k^{3/2}}{(1-\delta(r))^{k/2}}\frac{\log\log x}{\pi(x)^{1/2}}\right) \right) +O(\pi(x)^{k(n+1)}N^{-1}), \end{multline*}as $ x,N\rightarrow+\infty $. While uniformly for \textbf{odd} natural numbers $ k $ with $k\ll_{n,r}\frac{\log N}{\log x}$, one has$$\mathbb{E}_N((\pi_{f,r}(x)-\delta(r)\pi(x))^k)\ll C_{k,r}\pi(x)^{k/2}k\frac{\log\log x}{\pi(x)^{1/2}}+\pi(x)^{k(n+1)}N^{-1},$$as $ x,N\rightarrow+\infty $. \end{prop} \begin{proof} For $ z=x-k $ we obtained$$\pi_{f,r}(x)-\delta(r)\pi(x)=\sum_{p\le z}Y_p(f)+O_r\left(\frac{k}{\log x} \right) .$$In particular, $$ (\pi_{f,r}(x)-\delta(r)\pi(x))^k=\left( \sum_{p\le z}Y_p(f)\right) ^k$$\begin{equation}+O\left( \sum_{j=0}^{k-1}\left( \frac{k}{\log x}\right) ^{k-j}\binom{k}{j}\left| \sum_{p\le z}Y_p(f)\right| ^j\right) \end{equation}The dominant term in the error is obtained for $ j=k-1 $. If $k$ is even, we apply Lemma 2 to (3) and we get\begin{multline*} \frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}(\pi_{f,r}(x)-\delta(r)\pi(x))^k\\ =C_{k,r}\pi(x-k)^{k/2}\left( 1+O\left( \frac{k^3}{(1-\delta(r))^{k/2}}\frac{\log\log(x-k)}{\pi(x-k)}+k^3\frac{C_{k-1,r}}{C_{k,r}}\frac{\log\log(x-k)}{\pi(x-k)\log(x-k)}\right) \right)\\+O(\pi(x-k)^{k(n+1)}N^{-1})\\ =C_{k,r}\pi(x)^{k/2}\left( 1+O\left( \frac{k^{3/2}}{(1-\delta(r))^{k/2}}\frac{\log\log x}{\pi(x)^{1/2}}\right) \right) \\+O(\pi(x)^{k(n+1)}N^{-1}), \end{multline*} since$$\pi(x-k)^{k/2}=\pi(x)^{k/2}+O\left(\pi(x)^{k/2-1}\frac{k^2}{\log x} \right)$$and$$\frac{C_{k-1,r}}{C_{k,r}}\ll_r1.$$ If $k$ is odd, we can handle it using the Cauchy-Schwartz inequality:\begin{multline*} \frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}\Big|\sum_{p\le z}Y_p (f)\Big|^{k-1}\\ \le\Big(\frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}\Big|\sum_{p\le z}Y_p(f)\Big|^{k-2}\Big)^{1/2}\Big(\frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}\Big|\sum_{p\le z}Y_p(f)\Big|^{k}\Big)^{1/2}. \end{multline*}Lemma 2 leads to$$\frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}\Big|\sum_{p\le z}Y_p(f)\Big|^{k-1}\ll(C_{k-2,r}C_{k,r})^{1/2}k\pi(z)^{\frac{k}{2}-1}\log\log z.$$Since$$\frac{(C_{k-2,r}C_{k,r})^{1/2}}{C_{k,r}}\binom{k}{k-1}\asymp k^{1/2},$$we obtain from (3) \begin{multline*} \frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}(\pi_{f,r}(x)-\delta(r)\pi(x))^k\ll C_{k,r}\pi(x)^{\frac{k}{2}}\log\log x\left( \frac{k}{\pi(x)^{1/2}}+\frac{k^{3/2}}{\pi(x)\log x}\right)\\ +\pi(x)^{k(n+1)}N^{-1}\\ \ll C_{k,r}\pi(x)^{k/2}k\frac{\log\log x}{\pi(x)^{1/2}}+\pi(x)^{k(n+1)}N^{-1}. \end{multline*} \end{proof} \section{Proof of the main theorem} \begin{thm} For $ x=N^{1/\log\log N} $ and for any $ a<b\in\mathbb{R} $,$$\mathbb{P}_N\left(a\le \frac{\pi_{f,r}(x)-\delta(r)\pi(x)}{(\delta(r)-\delta(r)^2)^{1/2}\pi(x)^{1/2}}\le b\right) \longrightarrow\frac{1}{\sqrt{2\pi}}\int_{a}^{b}e^{-t^2/2}dt,$$as $ N\rightarrow+\infty $. \end{thm} \begin{proof} Firstly, note that it is equivalent to say that$$\mathbb{P}_N\left( \frac{\pi_{f,r}(x)-\delta(r)\pi(x)}{(\delta(r)-\delta(r)^2)^{1/2}\pi(x)^{1/2}}\le b\right) \longrightarrow\Phi(b)$$as $ N\rightarrow+\infty $, where $\Phi(b)=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{b}e^{-t^2/2}dt$. We use the method of moments. Since the function $ \Phi $ is determined by its moments$$\mu_k=\int_{-\infty}^{+\infty}x^kd\Phi(x),$$if a family of distribution functions $ F_n $ satisfies $\int_{-\infty}^{+\infty}x^kdF_n(x)\rightarrow\mu_k$ for all $ k\ge 1 $, then $ F_n(x)\rightarrow\Phi(x) $ pointwise (see \cite{Fel}, p. 262). On the other hand, if $ F_n(x)\rightarrow\Phi(x) $ for each $ x $ and if $ \int_{-\infty}^{+\infty}|x|^{k+\varepsilon}dF_n(x) $ is bounded in $ n $ for some $ \varepsilon>0 $, then $\int_{-\infty}^{+\infty}x^kdF_n(x)\rightarrow\mu_k$ (\cite{Fel}, p. 245). So the theorem will follow by the method of moments if we prove that for $ k\ge1 $,$$\mathbb{E}_N\left( \frac{(\pi_{f,r}(x)-\delta(r)\pi(x))^k}{((\delta(r)-\delta(r)^2)^{1/2}\pi(x)^{1/2})^k}\right)$$converges to $ \mu_k $ as $ N\rightarrow+\infty $. It's well known that$$\mu_k=\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{+\infty}x^ke^{-x^2/2}dx=\begin{cases} \frac{k!}{2^{k/2}(k/2)!}&\mbox{if }k\mbox{ is even}\\ 0&\mbox{if }k\mbox{ is odd}. \end{cases}$$From Proposition 2, if we fix $k\ge1$, we see exactly that$$\frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}\left( \frac{(\pi_{f,r}(x)-\delta(r)\pi(x))^k}{((\delta(r)-\delta(r)^2)^{1/2}\pi(x)^{1/2})^k}\right)\underset{x\rightarrow+\infty}{\longrightarrow}C_k=\mu_k$$if $k$ is even, and$$\frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^{0}}\left( \frac{(\pi_{f,r}(x)-\delta(r)\pi(x))^k}{((\delta(r)-\delta(r)^2)^{1/2}\pi(x)^{1/2})^k}\right)\ll_{k,r}\frac{\log\log x}{\pi(x)^{1/2}}\underset{x\rightarrow+\infty}{\longrightarrow}0=\mu_k$$if $k$ is odd. \end{proof} \section{Applications} \subsection{Upper bounds for the torsion part of the class number} All these bounds represent evidence towards the so-called $\varepsilon$-conjecture. \begin{conj} Let $K/\mathbb{Q}$ be a number field of degree $s$ with discriminant $D_K$. Then for every integer $\ell\ge1$ and every $\varepsilon>0$,$$h_K[\ell]\ll_{s,\ell,\varepsilon}D_K^{\varepsilon},$$where $h_K[\ell]$ is the order of the $\ell$-torsion subgroup of the class group. \end{conj} Using the well-known Minkowski bound ($ r_2 $ is the number of real $ \mathbb{Q} $-embeddings of $ K $)$$h_K\le\frac{s!}{s^s}\frac{4^{r_2}}{\pi^{r_2}}D_K^{1/2}(\log D_K)^{s-1}$$one has$$h_K\ll_{s,\varepsilon} D^{1/2+\varepsilon}_K$$for any $ \varepsilon>0 $. We can of course use the above to bound the $ \ell $-part $ h_K[\ell] $ of $ h_K $. But we'd like to improve the above estimate, and the main point we're going to use is the existence of "many" splitting completely primes, which contributes significantly to the quotient of the class group by its $\ell$-torsion. See Theorem 2 for the precise statement. The GRH guarantees the existence of such primes, but here, we'd like to proceed unconditionally.\\ Let $K/\mathbb{Q}$ be a number field. The presence of "small" primes that split completely in $ K$, give a means to improve the Minkowski lower bound, using the following theorem (\cite{EV}, Lemma 2.3). \begin{namedthm}{Theorem}[Ellenberg, Venkatesh] Let $ K/\mathbb{Q} $ be a field extension of degree $ s $. Set $ \delta<\frac{1}{2\ell(s-1)} $ and suppose that$$|\{p\le D_K^\delta:p\mbox{ splits completely in }K/\mathbb{Q}\}|\ge M.$$Then, for any $ \varepsilon>0 $ $$h_K[\ell]\ll_{s,\ell,\varepsilon}D^{1/2+\varepsilon}_KM^{-1}. $$ \end{namedthm} \begin{coroll} For every positive integer $\ell$, $\varepsilon>0$ and for almost all $ f \in\mathscr{P}_{n,N}^{0}$, outside of a set of size $ o(N^{n}) $, we have$$h_f[\ell]\ll_{n,\ell,\varepsilon}D_f^{\frac{1}{2}-\frac{1}{(2n-2)(n-1)!\log\log d_f}+\varepsilon},$$as $ N\rightarrow+\infty $. \end{coroll} \begin{proof} By Theorem 1, for $ x=N^{1/\log\log N} $, and $ \alpha=\alpha(N)>0 $ for large $ N $, $$\mathbb{P}_N\left( -N^{1/\alpha}\le\frac{\pi_{f,r}(x)-\delta(r)\pi(x)}{(\delta(r)-\delta(r)^2)^{1/2}\pi(x)^{1/2}}1\le N^{1/\alpha}\right)\underset{N\rightarrow+\infty}{\longrightarrow}1.$$In particular,$$\pi_{f,r}(x)\ge\delta(r)\pi(x)-N^{1/\alpha}(\delta(r)-\delta(r)^2)^{1/2}\pi(x)^{1/2}$$for all but $ o(N^n) $ $ f $ in the set $ \mathscr{P}_{n,N}^{0} $. Pick $ \alpha=3\log\log N $; then $ N^{1/\alpha}\pi(x)^{1/2}\ll_r\pi(x) $. By enlarging $N$, we can assume that$$N^{1/\alpha}(\delta(r)-\delta(r)^2)^{1/2}\pi(x)^{1/2}\ll_r\frac{1}{2}\delta(r)\pi(x).$$Then we get\begin{align*} \pi_{f,r}(x) \gg\delta(r)\pi(x). \end{align*} For $ \mathscr{C}_r=\{\id\} $, since $d_f\asymp N^{2n-2}$ for almost all $f$, and by the relation $d_f=D_f^{1/(n-1)!}a_{f}^2N_{\mathbb{Q}(\alpha)/\mathbb{Q}}(\mathfrak{D}_{K_f/\mathbb{Q}(\alpha)})^{1/(n-1)!}$ (see Section 5.2), we are bounding below the primes \begin{align*} p&\ll_n N^{1/\log\log N}\\ &\ll(d_f^{1/(2n-2)})^{1/\log\log d_f}\\ &=(D_f^{1/(2n-2)(n-1)!}a_f^{1/(n-1)}N_{\mathbb{Q}(\alpha)/\mathbb{Q}}(\mathfrak{D}_{K_f/\mathbb{Q}(\alpha)})^{1/(2n-2)(n-1)!})^{1/\log\log d_f}\\ &=\Big(D_f^{\frac{1}{(2n-2)(n-1)!}} D_f^{\frac{\log a_f}{(n-1)\log D_f}}D_f^{\frac{\log N_{\mathbb{Q}(\alpha)/\mathbb{Q}}(\mathfrak{D}_{K_f/\mathbb{Q}(\alpha)})}{(2n-2)(n-1)!\log D_f}}\Big)^{1/\log\log d_f}\\ &\le D_f^{\frac{1}{(2n-2)(n-1)!}\frac{\max(1,\log a_f,\log N_{\mathbb{Q}(\alpha)/\mathbb{Q}}(\mathfrak{D}_{K_f/\mathbb{Q}(\alpha)}))}{\log\log d_f}} \end{align*} splitting completely in $ K_f/\mathbb{Q} $. Now, by enlarging $ d_f $ if necessary (so by excluding a set of size $ O_n(1) $), pick a $\delta>0$ such that$$\frac{1}{(2n-2)(n-1)!}\frac{\max(1,\log a_f,\log N_{\mathbb{Q}(\alpha)/\mathbb{Q}}(\mathfrak{D}_{K_f/\mathbb{Q}(\alpha)}))}{\log\log d_f}<\delta<\frac{1}{2\ell(n!-1)}.$$It turns out that the primes $ p\ll D_f^\delta $ splitting completely in $ K_f/\mathbb{Q} $ are at least$$\gg_{n,\ell}\frac{(d_f)^{\frac{1}{(2n-2)\log\log d_f}}\log\log d_f}{\log d_f}.$$ By Theorem 2,\begin{align*} h_f[\ell]&\ll_{n,\ell,\varepsilon}\frac{D_f^{\frac{1}{2}+\varepsilon}\log d_f}{D_f^{\frac{1}{(2n-2)(n-1)!\log\log d_f}}a_f^{\frac{1}{(n-1)\log\log d_f}}N_{\mathbb{Q}(\alpha)/\mathbb{Q}}(\mathfrak{D}_{K_f/\mathbb{Q}(\alpha)})^{\frac{1}{(2n-2)(n-1)!\log\log d_f}}\log\log d_f}. \end{align*}In particular$$h_f[\ell]\ll_{n,\ell,\varepsilon}D_f^{\frac{1}{2}-\frac{1}{(2n-2)(n-1)!\log\log d_f}+\varepsilon}\cdot\frac{\log N}{\log\log N}$$for any $ \varepsilon>0 $, for almost all $ f $'s as $ N\rightarrow+\infty $, and the claim follows. \end{proof} We can improve this last upper bound by adding an additional hypothesis. \begin{thm} Let $K$ be a number field of degree $s$. There exists $\theta\in\mathcal{O}_K-\mathbb{Z}$ whose minimal polynomial $f_{\theta}$ has height$$\mbox{ht}(f_{\theta})\le 3^{s}\Big(\frac{D_K}{s}\Big)^{\frac{s}{2s-2}}.$$ \end{thm} \begin{proof} See \cite{GJ}, Appendix A. \end{proof} \begin{coroll}Assume that $K_f$ is generated over $\mathbb{Q}$ by an element $ \theta$ of "small height" of Theorem 3. Then for every positive integer $\ell$, $\varepsilon>0$ and for almost all $f\in\mathscr{P}_{n,N}^0$ outside of a set of size $o(N^n)$, we have $$h_{f}[\ell]\ll_{n,\ell,\varepsilon} D_f^{\frac{1}{2}-\frac{n!+1}{n!(2n-2)(n-1)!}\frac{1}{\log\log d_f}+\varepsilon},$$ as $N\rightarrow+\infty$. \end{coroll} \begin{proof}In particular we have $\mbox{ht}(f_{\theta})\ll_n D_f^{\frac{n!}{2n!-2}}$ and $d_{f_{\theta}}=D_f\cdot a^{'2}_{f_{\theta}}\asymp D_f^{n!}$ with "high probability". Since $d_f\asymp N^{2n-2}$ for almost all $f$, and$$D_f\asymp \frac{N^{(2n-2)(n-1)!}}{c^2_{f_{\theta}}N_{\mathbb{Q}(\alpha)/\mathbb{Q}}(\mathfrak{D}_{K_f/\mathbb{Q}(\alpha)})}$$for almost all $f$ (here $c_f=a_f^{(n-1)!}$), we obtain$$D_f\asymp\frac{N^{(2n-2)n!(n-1)!}}{a^{'2}_{f_{\theta}}(c_f^2N_{\mathbb{Q}(\alpha)/\mathbb{Q}}(\mathfrak{D}_{K_f/\mathbb{Q}(\alpha)}))^{n!}}.$$It turns out that$$N\asymp C_n(f,\theta)\cdot D_f^{\frac{1}{n!(2n-2)(n-1)!}}$$for almost all $f\in\mathscr{P}_{n,N}^{0}$. We denoted by $$C_n(f,\theta)=(a^{'2}_{f_{\theta}}(c_f^2N_{\mathbb{Q}(\alpha)/\mathbb{Q}}(\mathfrak{D}_{K_f/\mathbb{Q}(\alpha)}))^{n!})^{\frac{1}{n!(2n-2)(n-1)!}}.$$ As in Corollary 2 we want to count the primes\begin{align*} p&\ll_n N^{1/\log\log N}\\ &\ll \Big(C_n(f,\theta)\cdot D_f^{\frac{1}{n!(2n-2)(n-1)!}}\Big)^{1/\log\log d_f}\\ &=\Big( D_f^{\frac{1}{n!(2n-2)(n-1)!}+\frac{\log C_n(f,\theta)}{\log D_f}}\Big)^{1/\log\log d_f} \end{align*}splitting completely in $K_f/\mathbb{Q}$. If we fix a $\delta>0$ so that$$\left(\frac{1}{n!(2n-2)(n-1)!}+\frac{\log C_n(f,\theta)}{\log D_f}\right)\frac{1}{\log\log d_f}<\delta<\frac{1}{2\ell(n!-1)}$$(by enlarging $d_f$ if necessary), then the number of primes $p\ll D_f^{\delta}$ splitting completely is bounded above by$$\gg_{n,\ell}\frac{\Big(C_n(f,\theta)\cdot D_f^{\frac{1}{n!(2n-2)(n-1)!}}\Big)^{1/\log\log d_f}\log\log d_f}{\log d_f},$$and so, by noting that \begin{align*} C_n(f,\theta)&\gg (N_{\mathbb{Q}(\alpha)/\mathbb{Q}}\mathfrak{D}_{K_f/\mathbb{Q}(\alpha)})^{\frac{1}{(2n-2)(n-1)!}}\\ &=\frac{D_f^{\frac{1}{(2n-2)(n-1)!}}}{D_{\mathbb{Q}(\alpha)}^{\frac{1}{2n-2}}}\\ &=D_f^{\frac{1}{(2n-2)(n-1)!}}\frac{a_f^{n-1}}{d_f^{\frac{1}{2n-2}}}\\ &\gg\frac{D_f^{\frac{1}{(2n-2)(n-1)!}}}{d_f^{\frac{1}{2n-2}}} \end{align*}one gets$$h_{f}[\ell]\ll_{n,\ell,\varepsilon} D_f^{\frac{1}{2}-\frac{n!+1}{n!(2n-2)(n-1)!}\frac{1}{\log\log d_f}+\varepsilon}\cdot\frac{N^{\frac{1}{\log\log N}}\log N}{\log\log N}$$ as $N\rightarrow+\infty$, for every $\varepsilon>0$ and for almost all $f$. The result follows. \end{proof} \subsection{Discriminant and average of ramified primes} Let $f$ be an $S_n$-polynomial and let $d_f$ be its discriminant. We are going to discuss the relation between the number of primes $p$ dividing $d_f$ and the discriminant of its splitting field field $K_f/\mathbb{Q}$ (i.e. the ramified primes in the extension $K_f/\mathbb{Q}$).\\ For a polynomial $ f\in \mathscr{P}_{n,N} $, the bound$$d_f\ll N^{2n-2}$$holds, since $ d_f $ is given by the $ (2n-1) $-dimensional determinant$$d_f=(-1)^{n(n-1)/2}\det\begin{pmatrix} 1&a_{n-1}&a_{n-2}&\cdots&a_0&0&\cdots\\ 0&a_n&a_{n-1}&\cdots&a_1&a_1&\cdots\\ \vdots&&&&&&\vdots\\ 0&\cdots&0&a_n&\cdots&a_1&a_0\\ n&(n-1)a_{n-1}&(n-2)a_{n-2}&\cdots&0&0&\cdots\\ \vdots&&&&&&\vdots\\ 0&\cdots&\cdots&0&na_n&\cdots&a_1 \end{pmatrix}$$with $ a_n=1 $ in our case. However, it turns out (see \cite{GZ}, Corollary 2.2) that$$ d_f\asymp N^{2n-2}$$for almost all $ f $. Indeed, for all $ \varepsilon>0 $ there exists $ \delta=\delta(n) $ s.t. for $ N $ large enough$$ \mathbb{P}_N(|d_f|>\delta N^{2n-2})>1-\varepsilon. $$ By the primitive element theorem, we know there is an integral element $\theta\in\mathcal{O}_{K_f}$ so that $K_f=\mathbb{Q}(\theta)$. Let $f_{\theta}\in \mathbb{Z}[X]$ be the minimal polynomial of $\theta$. Then it holds the following relation between the discriminant of $f_{\theta}$ and the discriminant $D_f$ of the number field extension $K_f/\mathbb{Q}$:$$d_{f_{\theta}}=a_{f_{\theta}}^2\cdot D_f,$$where $a_{f_{\theta}}\in\mathbb{Z}$ (see \cite{La}, Chapter III).\\ Now, let $\alpha$ be a root of $f\in\mathscr{P}_{n,N}^{0}$ and consider the extension generated by $\alpha$ over $\mathbb{Q}$. $$\begin{tikzpicture} \matrix (m) [matrix of math nodes,row sep=1em,column sep=0.001em,minimum width=2em] {K_f\\ \mathbb{Q}(\alpha)\\ \mathbb{Q}\\}; \path[-] (m-1-1) edge node [right] {$\scriptstyle{(n-1)!}$} (m-2-1) (m-2-1) edge node [right] {$\scriptstyle{n}$} (m-3-1) ; \end{tikzpicture}$$By the transitivity of the discriminant in towers of extensions, one has$$D_f=D_{\mathbb{Q}(\alpha)}^{(n-1)!}N_{\mathbb{Q}(\alpha)/\mathbb{Q}}\mathfrak{D}_{K_f/\mathbb{Q}(\alpha)}.$$As above, $d_f=a_{f}^2\cdot D_{\mathbb{Q}(\alpha)}$ with $a_f\in\mathbb{Z} $. It turns out that\begin{equation} d_f=D_f^{1/(n-1)!}a_f^2(N_{\mathbb{Q}(\alpha)/\mathbb{Q}}\mathfrak{D}_{K_f/\mathbb{Q}(\alpha)})^{1/(n-1)!} \end{equation} As in Proposition 6.4 of \cite{ABZ}, we see that the probability that a monic, irreducible, degree $ n $ polynomial with height $ \le N $ has discriminant coprime with $p $ is $ 1-\frac{1}{p} $, hence$$\frac{|\{f\in\mathscr{P}_{n,N}^{\tiny\mbox{irr}}:p|d_f\}|}{|\mathscr{P}_{n,N}^{\tiny\mbox{irr}}|}\underset{N\rightarrow+\infty}{\longrightarrow}\frac{1}{p}.$$ \begin{coroll} The average of the number of ramified primes in $K_f/\mathbb{Q}$ is$$\mathbb{E}_N(|\{p:p|D_f\}|)\ll_{n}\log\log N,$$as $N\rightarrow+\infty$. \end{coroll} \begin{proof} Since almost all polynomials in $ \mathscr{P}_{n,N} $ are irreducible, with error term $ O(N^{-1}) $, and since $|\mathscr{P}_{n,N}^{0}|=(2N)^{n}+O(N^{n-1})$, we also have that$$\frac{|\{f\in\mathscr{P}_{n,N}^{0}:p|d_f\}|}{|\mathscr{P}_{n,N}^{0}|}=\frac{1}{p}+o(1),$$as $ N\rightarrow+\infty $. In particular, for the primes $p<N^{1/n}$, we can also write down explicitely the error term by applying Lemma 1:\begin{align*} \mathbb{P}_N(f\in\mathscr{P}_{n,N}^{0}:p|d_f)&=\frac{1}{|\mathscr{P}_{n,N}^{0}|}\underset{g\tiny\mbox{ double root}}{\underset{\tiny\mbox{monic, }\deg g=n}{\sum_{g\in\mathbb{F}_{p}[X]}}}\ \underset{f\equiv g\tiny\mbox{ mod }p}{\sum_{f\in \mathscr{P}_{n,N}^{0}}}1\\ &=p^{n-1}\Big(\frac{1}{p^n}+O(N^{-1})\Big)\\ &=\frac{1}{p}+O(p^{n-1}N^{-1}), \end{align*}as $N\rightarrow+\infty$, as long as $p<N^{1/n}$. It follows that\begin{align*} \mathbb{E}_N(|\{p:p|d_f\}|)&=\sum_{ p<N^{1/n}}\frac{1}{|\mathscr{P}_{n,N}^{0}|}\underset{p|d_f}{\sum_{f\in \mathscr{P}_{n,N}^{0}}}1+O\Big(\frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in\mathscr{P}_{n,N}^0}\underset{p|d_f}{\sum_{ p\ge N^{1/n}}}1\Big)\\ &=\log\log N+O_{n}(1), \end{align*}as $N\rightarrow+\infty$. From the transitive relation (4), one has the claim. \end{proof} Let $\theta=\theta_1$ be a root of $f$, and let $K_1=\mathbb{Q}(\theta)$.\\ For a prime $p$, the ring $\mathbb{Z}[\theta]$ is called $p$\textbf{-maximal} if $p$ is not a divisor of the index of $\mathbb{Z}[\theta]$ in $\mathcal{O}_{K_1}$. In particular $\mathbb{Z}[\theta]$ is not $p$-maximal if and only if $\wp|(d_f/D_f)$.\\ There is an equivalent condition for $\mathbb{Z}[\theta]$ to be $p$-maximal. \begin{namedthm}{Theorem}[\cite{ABZ}, Corollary 3.2] The ring $\mathbb{Z}[\theta]$ is not $p$-maximal if and only if there exists $u\in\mathbb{Z}[X]$, with $u$ mod $p$ irreducible, such that $f\in (p^2+up+u^2)$ in $\mathbb{Z}[X]$. \end{namedthm} In particular, the $p$-maximality depends just on $f$ mod $p^2$. The probability that such a polynomial modulo $p^2$ is in the above ideal (for a fixed $u$) is given by the following. \begin{namedthm}{Theorem}[\cite{ABZ}, Proposition 3.4] Let $g\in\mathbb{F}_{p}[X]$ monic, of degree $m$; then$$\frac{1}{p^{2n}}\underset{f\in (gp+g^2) }{\underset{\tiny{\mbox{monic,\ }\deg f=n}}{\sum_{f\in(\mathbb{Z}/p^2\mathbb{Z})[X]}}}1=\begin{cases} 0&if\ 2m>n\\ \frac{1}{p^{3m}}&if\ 2m\le n. \end{cases}$$ \end{namedthm} \begin{coroll}In the above notations, the average of the number of primes dividing $a_f$ is$$ \mathbb{E}_N(|\{p:p|a_f\}|)\ll_{n}1,$$as $N\rightarrow+\infty$. \end{coroll} \begin{proof} From theorems 4 and 5, we deduce that if $g\in\mathbb{F}_{p}[X]$ is monic, of degree $m\le n/2$, then \begin{align*} \frac{1}{|\mathscr{P}_{n,N}^{0}|}\underset{f\in (p^2+gp+g^2)}{\sum_{f\in\mathscr{P}_{n,N}^{0}}}1&=\frac{1}{|\mathscr{P}_{n,N}^{0}|}\underset{h\in (gp+g^2)}{\underset{\tiny{\mbox{monic,\ }\deg h=n}}{\sum_{h\tiny\mbox{ mod }p^2}}}\ \underset{f\equiv h\tiny\mbox{ mod }p}{\sum_{f\in \mathscr{P}_{n,N}^{0}}}1\\ &=\underset{h\in (gp+g^2)}{\underset{\tiny{\mbox{monic,\ }\deg h=n}}{\sum_{h\tiny\mbox{ mod }p^2}}}\Big(\frac{1}{p^{2n}}+O(N^{-1})\Big)\\ &=\frac{1}{p^{3m}}+O(p^{2n-3m}N^{-1}), \end{align*}for all primes $p<N^{1/2n}$. The next step is to compute the probability $\mathbb{P}_N(f\in \mathscr{P}_{n,N}^{0}:p|(d_f/D_{K_1})$, which is, by the above\begin{multline*} \frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{m\le n/2}\underset{\deg g=m}{\underset{\tiny{\mbox{monic, irreducible}}}{\sum_{g\in\mathbb{F}_{p}[X]}}}\ \underset{f\in(p^2+gp+g^2)}{\sum_{f\in \mathscr{P}_{n,N}^{0}}}1\\=\sum_{m\le n/2}\underset{\deg g=m}{\underset{\tiny{\mbox{monic, irreducible}}}{\sum_{g\in\mathbb{F}_{p}[X]}}}\Big(\frac{1}{p^{3m}}+O(p^{2n-3m}N^{-1})\Big)\\ =\sum_{m\le n/2}\Big(\frac{p^m}{m}+O\Big(\frac{p^{m-1}}{m}\Big)\Big)\Big(\frac{1}{p^{3m}}+O(p^{2n-3m}N^{-1})\Big)\\ =\sum_{m\le n/2}\Big(\frac{1}{mp^{2m}}+O\Big(\frac{p^{2n}}{mp^{2m}}N^{-1}+\frac{1}{mp^{2m+1}}\Big)\Big)\\ =\frac{1}{p^2}+O\Big(p^{2n-2}N^{-1}+\frac{1}{p^3}\Big), \end{multline*}as $N\rightarrow+\infty$, for $p<N^{1/2n}$. Then, the number of primes (on average) diving $d_f/D_{K_1}$ is\begin{multline*} \mathbb{E}_N(|\{p:p|(d_f/D_{K_1})\}|)\\ =\sum_{p<N^{1/2n}}\mathbb{P}_N(f:p|(d_f/D_{K_1}))+\frac{1}{|\mathscr{P}_{n,N}^{0}|}\sum_{f\in \mathscr{P}_{n,N}^{0}}\underset{p|(d_f/D_{K_1})}{\sum_{p\ge N^{1/2n}}}1\\ =\sum_{p<N^{1/2n}}\frac{1}{p^2}+O(1+\pi(N^{1/2n})^{2n-1}N)\\ \ll_{n} 1, \end{multline*}since$$\underset{p|(d_f/D_{K_1})}{\sum_{ p\ge N^{1/2n}}}1\ll_{n}\frac{\log N}{\log N}\ll_{n} 1.$$ \end{proof} \section{Further results and problems} \subsection{The range of $x$ and $N$} Regarding the range of $x,N$ for the average Chebotarev Theorem, in our proofs, the restriction $x\le N^{1/\log\log N}$ or something similar is essential. It would be interesting to know in what interval of $x$ and $N$ these results actually hold. \subsection{Other Galois groups} Another goal, is to provide similar results for polynomials in having as Galois group over $\mathbb{Q}$, either $S_n$ or a transitive proper subgroup of $S_n$. It would be interesting to exploit the Hilbert Irreducibility Theorem to get results for some group $G\subseteq S_n$. In general, for a subgroup $G\subseteq S_n$, the elements in a conjugacy class in $G$ necessarily have the same cycle type, but the converse need not to be true. That is, the cycle type of a conjugacy class in $G$ need not determine it uniquely. This uniqueness property does hold for cycle types for the full symmetric group, which implies that the cycle type of an $S_n$-polynomial having a square-free factorization mod $p$ uniquely determines the Frobenius element for an $S_n$-number field obtained by adjoining one root of it.\\ Let's consider the case of the alternating group $A_n\subseteq S_n$. A single conjugacy class in $S_n$ that is contained in $A_n$ may split into two distinct classes. Also, note that the fact that conjugacy in $S_n$ is determined by cycle type, means that if $\sigma\in A_n$, then all of its conjugates in $S_n$ also lie in $A_n$. There is a full characterization of the behaviour of conjugacy classes in $A_n$. \begin{lemma} A conjugacy class in $S_n$ splits into two distinct conjugacy classes under the action of $A_n$ if and only if its cycle type consists of distinct odd integers. Otherwise, it remains a single conjugacy class in $A_n$. \end{lemma} \begin{proof} Note that the conjugacy class in $S_n$ of an element $\sigma\in A_n$ splits, if and only if there is no element $\tau\in S_n\setminus A_n$ commuting with $\sigma$. For if there is one, for each $\tau'\in S_n\setminus A_n$ we have$$\tau'\sigma\tau'^{-1}=\tau'\sigma\tau\tau^{-1}\tau'^{-1}=(\tau'\tau)\sigma(\tau'\tau)^{-1},$$and $\tau\tau'\in A_n$. On the other hand, if $\tau\sigma\tau^{-1}$ and $\sigma$, with $\tau\in S_n\setminus A_n$, are conjugated in $A_n$, then for some $\tau'\in A_n$, we have $\tau\sigma\tau^{-1}=\tau'\sigma\tau'^{-1}$, giving$$\tau'^{-1}\tau\sigma=\sigma\tau'^{-1}\tau,$$and hence $\tau'^{-1}\tau\in S_n\setminus A_n$ commutes with $\sigma$. Now suppose, $\sigma$ has a cycle $c_i$ of even length. A cycle of even length is an element of $S_n\setminus A_n$, and as $\sigma$ commutes with its cycles, we are done by the above. If $\sigma$ has two cycles $(a_1\dots a_k)$ and $(b_1\dots b_k)$ of the same odd length $k$, then $(a_1 b_1)\dots(a_kb_k)$ is a product of $k$ permutations (hence odd, so an element of $S_n\setminus A_n$) commuting with $\sigma$. Suppose $\sigma=c_1\dots c_s$ is a product of odd cycles $c_i$ of distinct lengths $d_i$. Let $\tau\in S_n$ be a permutation commuting with $\sigma$. Then $\tau$ must fix each of the $c_i$, that is, $\tau$ must be of the form $\tau=c_1^{a_1}\dots c_s^{a_s}$ for some $a_i\in\mathbb{Z}$. But as the $c_i$ are even permutations (as cycles of odd length), we have $\tau\in A_n$. So no $\tau\in S_n\setminus A_n$ commutes with $\sigma$ and we have the claim. \end{proof} As in the case of $S_n$ polynomials, we want to count the number of $G$-polynomials, where $G$ is a subgroup of the symmetric group $S_n$. We use the following result. \begin{namedthm}{Theorem}[Dietmann] For every $\varepsilon>0$ and positive integer $n$,\begin{multline*} |\{(a_0,\dots,a_{n-1})\in\mathbb{Z}^n:|a_j|\le N\ \forall j,\\ f(X)=X^n+a_{n-1}X^{n-1}+\dots+a_0\mbox{ has }G_f=G\}|\\ \ll_{n,\varepsilon}N^{n-1+1/[S_n:G]+\varepsilon}, \end{multline*}where $[S_n:G]$ is the index of $G$ in $S_n$. \end{namedthm} \begin{proof} See \cite{Di} and \cite{Di2}. \end{proof} Consider the set$$\mathscr{P}_{n,N}^1=\{f\in\mathscr{P}_{n,N}:G_f=S_n\mbox{ or }A_n\}.$$ Now, if $G\subseteq S_n$, $G\neq S_n,A_n$ then its index in $S_n$ is greater or equal then $n$. From Theorem 6 we thus have that$$|\mathscr{P}_{n,N}^1|=(2N)^{n}+O( N^{n-1+\frac{1}{n} +\varepsilon }) .$$ Let $\mathscr{C}_r\in A_n$ be a conjugacy class, with $r=(r_1,\dots,r_n)$ a square-free splitting type such that either $r_i$ is even for some $i$, or all the $r_i$'s are odd but $r_i=r_j$ for some $i\neq j$. By following the same argument as in Lemma 1 and Proposition 1, we get the Chebotarev Theorem on average:$$\frac{1}{|\mathscr{P}_{n,N}^1|}\sum_{f\in \mathscr{P}_{n,N}^{1}}\Big(\underset{\Frob_{f,p}=\mathscr{C}_r}{\sum_{ p\le x}}1 \Big)=\delta(r)\pi(x)+C_r\log\log x+O_{n}(1),$$as $x,N\rightarrow+\infty$, if $x<N^{\frac{1-(1/n)-\varepsilon}{n+1}}$.\\ \noindent\textbf{Remark.} Note that even if one doesn't fully control the conjugacy classes of a subgroup $G\subseteq S_n$ in terms of the cycle type, there is still interesting information to extract from it. Especially, about the number of totally splitting primes (corresponding to the trivial conjugacy class), which was the main tool in the application to class group torsion upper bounds of Section 5.1. \subsection{Generalization for number fields} One can consider the analogous over number fields, that is polynomials with coefficients in the ring of algebraic integers of a fixed finite extension $K/\mathbb{Q}$. It is possible to generalize the work of Bhargava, towards the van der Waerden's conjecture. In this case, for a polynomial$$f(X)=X^n+\alpha_{n-1}X^{n-1}+\dots+\alpha_0\in\mathcal{O}_K[X],$$the \textit{height} is defined in term of the integer coefficients of the $\alpha_i$ in a fixed integral basis of $\mathcal{O}_K$ over $\mathbb{Z}$. In particular, we get that the number of non-$S_n$-polynomials as above is$$\ll_{n,K}N^{d(n-1/2)}\log N,$$for all $d\ge1$, $n\ge2$, as $N\rightarrow+\infty$. The above result, as well as a refined version for some values of $d$ and $n$, are contained in my Ph.D. thesis.

https://arxiv.org/abs/2002.09210

An algorithm for computing the $Υ$-invariant and the $d$-invariants of Dehn surgeries

By using grid homology theory, we give an explicit algorithm for computing Ozsváth-Stipsicz-Szabó's $\Upsilon$-invariant and the $d$-invariant of Dehn surgeries along knots in $S^3$. As its application, we compute the two invariants for all prime knots with up to 11 crossings.

\section{Introduction} \label{sec:intro} The {\it $\Upsilon$-invariant} defined by Ozsv\'ath, Stipsicz and Szab\'o \cite{OSS17} is a group homomorphism \[ \Upsilon \colon \mathcal{C} \to \PL([0,2], \mathbb{R}), \] where $\mathcal{C}$ denotes the smooth knot concordance group and $\PL([0,2], \mathbb{R})$ the vector space of piecewise-linear functions on $[0,2]$. The $\Upsilon$-invariant is known as a very effective tool in the study of smooth knot concordance. Indeed, the slope of $\Upsilon$ gives a surjective homomorphism $\mathcal{C} \to \mathbb{Z}^\infty$ whose image is generated by topologically slice knots, and this proves that the subgroup of $\mathcal{C}$ generated by topologically slice knots contains a direct summand isomorphic to $\mathbb{Z}^\infty$. For a given knot $K$ in $S^3$, we denote the $\Upsilon$-invariant of $K$ by $\Upsilon_K$. The {\it $d$-invariant} defined by Ozsv\'ath and Szab\'o \cite{OS03grading} is a group homomorphism \[ d \colon \theta^c \to \mathbb{Q}, \] where $\theta^c$ denotes the $\Spin^c$ rational homology cobordism group of $\Spin^c$ rational homology 3-spheres (see \cite[Definition 1.1]{OS03grading} for the precise definition of $\theta^c$). The $d$-invariant is also known as a very powerful invariant so that it enables us to reprove Donaldson's diagonalization theorem. Moreover, the $d$-invariants of Dehn surgeries and prime-powered branched covers along knots in $S^3$ yield an infinite family of smooth knot concordance invariants (see \cite{Hom17} for a survey describing these facts and studies of them). For any knot $K$ in $S^3$ and coprime integers $p$ and $q$ with $p \neq 0$ and $q>0$, let $S^3_{p/q}(K)$ denote the $p/q$-surgery along $K$. In this paper, we focus on the $d$-invariant of $S^3_{p/q}(K)$. \footnote{While $0$-surgeries are not rational homology 3-spheres, the $d$-invariants of them are also defined. For the cases of knots in $S^3$, they are determined by $d(S^3_{\pm 1}(K))$. See \cite[Section 4.2]{OS03grading}.} Note that since $S^3_{p/q}(K)$ has $|p|$ different $\Spin^c$ structures, the $d$-invariant of $S^3_{p/q}(K)$ is obtained as a $|p|$-tuple of rational numbers. We denote the $|p|$-tuple by $d(S^3_{p/q}(K))$. The invariants $\Upsilon_K$ and $d(S^3_{p/q}(K))$ were originally defined in different packages of Heegaard Floer theory, but later Livingston \cite{Liv17} and Ni-Wu \cite{NW15} have translated them into the words of the doubly filtered chain complex $CFK^{\infty}(K)$ defined in \cite{OS04knot}. Moreover, $CFK^{\infty}(K)$ has a combinatorial description called {\it grid homology theory} introduced by Manolescu, Ozsv\'ath and Sarkar \cite{MOS09grid1}. Therefore, it is natural to consider computing $\Upsilon_K$ and $d(S^3_{p/q}(K))$ algorithmically by using grid homology theory. However, while $CFK^\infty(K)$ is freely and finitely generated over the Laurent polynomial $\mathbb{F}[U^{\pm 1}]$ (where $\mathbb{F}:=\mathbb{Z}/2\mathbb{Z}$), the combinatorial chain complex $C^\infty(G)$ corresponding to $CFK^{\infty}(K)$ is freely and finitely generated over $\mathbb{F}[U^{\pm 1}, U^{\pm 1}_2, \ldots, U_n^{\pm 1}]$ with $n>1$. In particular, $C^\infty(G)$ has infinite rank over $\mathbb{F}[U^{\pm 1}]$, and hence there is no obvious way to use $C^\infty(G)$ for algorithmic computation of $\Upsilon_K$ and $d(S^3_{p/q}(K))$. In this paper, we provide a method for avoiding the infinite-dimensional problem, and give an explicit algorithm for computing $\Upsilon_K$ and $d(S^3_{p/q}(K))$ of any knot $K$ in $S^3$ and $p/q \in \mathbb{Q}$. Moreover, as its applications, we complete the lists of \begin{itemize} \item the $\Upsilon$-invariants of all prime knots with up to 11 crossings, \item all $d$-invariants of all Dehn surgeries along all prime knots with up to 11 crossings, and \item all prime knots with up to 12 crossings whose $\Upsilon$ and $d$-invariants coincide with those of the unknot. \end{itemize} \subsection{The Key idea} \label{subsec: key ideas} In order to give an algorithm for computing $\Upsilon_K$ and $d(S^3_{p/q}(K))$, we will translate these invariants several times. The key point is to use the second author's invariant $\mathcal{G}_0(K)$ \cite{2019arXiv190709116S}, from which we can easily compute both $\Upsilon_K$ and $d(S^3_{p/q}(K))$. While $\mathcal{G}_0(K)$ was also defined using $CFK^\infty(K)$, we show in this paper that it can be translated into the words of the subcomplex $CFK^-(K)$ over $\mathbb{F}[U]$. The combinatorial chain complex $C^-(G)$ corresponding to $CFK^-(K)$ is finitely and freely generated by $\mathbb{F}[U, U_2, \ldots, U_n]$ and so it has infinite rank over $\mathbb{F}[U]$, but as a graded $\mathbb{F}$-vector space, each degree of $C^-(G)$ has finite rank. (It does not hold for $C^\infty(G)$, i.e.\ each degree of $C^\infty(G)$ has infinite rank over $\mathbb{F}$.) This fact enables us to reduce the problem to finite $\mathbb{F}$-linear systems, which makes $\mathcal{G}_0(K)$ algorithmically computable. \subsection{Observations of computation results} \label{subsec: observations} Here we show several observations of our computation results. We first state the relationship between the {\it $\tau$-invariants} \cite{OS03tau} and the {\it $\nu^+$-equivalence classes} \cite{Hom17, KP18} of knots with up to 12 crossings. Here, the {\it $\tau$-invariant} is a group homomorphism $\mathcal{C} \to \mathbb{Z}$ which is already computed for all knots with up to 11 crossings \cite{BG12}. (Now, the list of the $\tau$-invariants for knots with up to 12 crossings is available in KnotInfo~\cite{Knotinfo}.) The {\it $\nu^+$-equivalence} is an equivalence relation on $\mathcal{C}$ whose quotient set can be regarded as a quotient group $\mathcal{C}_{\nu^+}$ of $\mathcal{C}$. It is proved in $\cite{Hom17}$ that a knot $K$ is $\nu^+$-equivalent to the unknot if and only if $CFK^\infty(K)$ is filtered chain homotopic to $CFK^\infty(O) \oplus A$, where $O$ denotes the unknot and $A$ is an acyclic complex, i.e.\ $H_*(A)=0$. Moreover, all of $\tau$, $\Upsilon_K$ and $d(S^3_{p/q}(K))$ are invariant under $\nu^+$-equivalence. In particular, the triviality of $K$ in $\mathcal{C}_{\nu^+}$ implies $\tau(K)=0$, while the converse does not hold in general. (For instance, the composite knot $T_{2,7}\# (T_{3,4})^*$ is such an example, where $T_{p,q}$ denotes the positive $(p,q)$ torus knot and $K^*$ the mirror image of $K$.) On the other hand, it is also known that the converse holds for all quasi-alternating knots \cite{Pet13} and all genus one knots \cite{2019arXiv190709116S}. As an observation of our computation results, we show that the converse also holds for any prime knot with up to 12 crossings. \begin{theorem} \label{thm: converse} For any prime knot $K$ with up to 12 crossings, $K$ is $\nu^+$-equivalent to the unknot if and only if $\tau(K)=0$. \end{theorem} Here we also mention that the triviality of a knot in $\mathcal{C}_{\nu^+}$ is determined by the diffeomorphism type of its $0$-surgery. (The proof immediately follows from the original definition of $\nu^+$-equivalence in \cite{KP18} with \cite[Proposition 1.6]{NW15} and \cite[Proposition 4.12]{OS03grading}.) \begin{proposition} If two knots $K$ and $K'$ have the same 0-surgeries, then $K$ is $\nu^+$-equivalent to the unknot if and only if $K'$ is $\nu^+$-equivalent to the unknot. \end{proposition} \begin{remark} The $\nu^+$-equivalence class of a knot is not an invariant of its $0$-surgery in general. Indeed, Yasui \cite{Yas15pre} provides an infinite family of pairs of knots with the same 0-surgeries one of whose $\tau$ is greater than the other by 1. \end{remark} It is meaningful to compare the above propositions with Piccirillo's proof of non-sliceness of the Conway knot $11n34$ \cite{Pic18pre}. To prove the non-sliceness of the Conway knot, Piccirillo consider the {\it knot trace}, which is a 4-manifold obtained by attaching a 0-framed 2-handle to $B^4$ along a given knot. It is known that if two knots have the same knot traces, then the sliceness of them coincides. Piccirillo finds a knot $K'$ whose knot trace is diffeomorphic to that of the Conway knot, and she shows $s(K') \neq 0$, where $s$ denotes the Rasmussen's homomorphism $s \colon \mathcal{C} \to 2\mathbb{Z}$ \cite{rasmussen2010}. Here we note that sharing the knot trace implies sharing the 0-surgeries. Moreover, since the Conway knot has 11 crossings and trivial $\tau$, Theorem~\ref{thm: converse} implies that the Conway knot is $\nu^+$-equivalent to the unknot. Therefore, in contrast to the $s$-invariant, we have the following: \begin{theorem} For any knot $K$, if the 0-surgery along $K$ is diffeomorphic to that of the Conway knot, then $K$ is $\nu^+$-equivalent to the unknot. In particular, all of $\tau(K)$, $\Upsilon_K$ and $d(S^3_{p/q}(K))$ coincide with those of the unknot. \end{theorem} In particular, we cannot replace $s(K')$ in Piccirillo's proof with any of $\tau(K')$, $\Upsilon_{K'}$ and $d(S^3_{p/q}(K'))$. \subsection*{Organization} In Section~\ref{sec: review of G_0}, we review the invariant $\mathcal{G}_0(K)$. In Section~\ref{sec: translate G_0}, we translate $\mathcal{G}_0(K)$ into the words of $CFK^-(K)$. In Section~\ref{sec: translating G_0 to C^-(G)}, we review the grid complex $C^-(G)$, and translate $\mathcal{G}_0(K)$ into $C^-(G)$. In Section~\ref{sec: algorithm}, we give an algorithm for computing $\mathcal{G}_0(K)$ from $C^-(G)$. In Section~\ref{sec: results}, we show our computation results. This paper also contains several appendices. In Appendix~\ref{sec: duality}, we give the duality theorem for $\mathcal{G}_0(K)$ with respect to the mirroring of $K$, which proves that $\mathcal{G}_0$ for the mirror $K^*$ is algorithmically determined from $\mathcal{G}_0(K)$. This is used for determining $d(S^3_{p/q}(K))$ where $p/q$ is negative. In Appendix~\ref{subsec:sparse-linear-system}, we discuss methods for handling sparse linear systems. \subsection*{Acknowledgements} The authors thank the first author's supervisor Mikio Furuta for his support. Development of the computer program and its computation on a cloud computing platform is supported by JSPS KAKENHI Grant Number 17H06461. The authors also thank Charles Bouillaguet for many helpful suggestions about sparse linear systems, \textit{omochimetaru} and the members of \textit{swift-developers-japan} discord server for helpful advice about the implementation of the program. The first author thanks members of his \textit{academist fanclub}\footnote{\url{https://taketo1024.jp/supporters}} for their support. The second author was supported by JSPS KAKENHI Grant Number 18J00808. \section{A review of the invariant $\mathcal{G}_0$} \label{sec: review of G_0} In this section, we review the invariant $\mathcal{G}_0(K)$ defined in \cite{2019arXiv190709116S}, which determines both $\Upsilon_{K}$ and $d(S^3_{p/q}(K))$. \subsection{Knot complexes, $CFK^{\infty}$} \label{subsec: formal knot complexes} \subsubsection{Poset filtered chain complexes} \label{poset filtered chain complexes} Let $P$ be a {\it poset}, i.e.\ a set $P$ with partial order $\leq$. Then, a \textit{lower set} $R \subset P$ is a subset such that for any $x \in P$, if there exists an element $y \in R$ satisfying $x \leq y$, then $x \in R$. In this paper, we mainly consider the partial order $\leq$ on $\mathbb{Z}^2$ given by $(i,j)\leq(k,l)$ if $i \leq k$ and $j \leq l$. We specially call a lower set R $\subset$ $\mathbb{Z}^2$ with respect to $\leq$ a {\it closed region}. Let $\mathbb{F} := \mathbb{Z} /2\mathbb{Z}$ and $\Lambda$ be an $\mathbb{F}$-algebra. In this paper, we say that $(C,\partial)$ is a {\it chain complex $C$ over $\Lambda$} if $(C,\partial)$ satisfies the following: \begin{itemize} \item $C$ is a $\Lambda$-module and $\partial \colon C \to C$ is a $\Lambda$-linear map with $\partial \circ \partial = 0$. \item As an $\mathbb{F}$-vector space, $C$ is decomposed into $\bigoplus_{n \in \mathbb{Z}} C_n$, where each $C_n$ satisfies $\partial(C_n) \subset C_{n-1}$. \end{itemize} (Note that the $\Lambda$-action does not preserve the grading in general.) We often abbreviate $(C,\partial)$ to $C$. Moreover, we say that $C$ is \textit{$P$-filtered} if a subcomplex $C_R$ of $C$ over $\mathbb{F}$ is associated to each closed region $R \subset P$ so that if $R \subset R'$ then $C_R \subset C_{R'}$. (Here we remark that $C_R$ is not a $\Lambda$-submodule of $C$ in general.) We call the set $\{C_R\}_{R \in \mathcal{CR}(P)}$ a {\it $P$-filtration} on $C$. For instance, a $\mathbb{Z}$-filtration of $C$ can be identified as an ascending filtration in the usual sense, under the identification of an integer $i$ with a closed region $\{m \in \mathbb{Z} \mid m \leq i\}$. Moreover, given two $\mathbb{Z}$-filtrations $\{\mathcal{F}^1_i\}_{i \in \mathbb{Z}}$ and $\{\mathcal{F}^2_j\}_{j \in \mathbb{Z}}$ on $C$, the set $$ \{C_R\}_{R \in \mathcal{CR}(\mathbb{Z}^2)} := \{\sum_{(i,j)\in R} \mathcal{F}^1_i \cap \mathcal{F}^2_j\}_{R \in \mathcal{CR}(\mathbb{Z}^2)} $$ defines a $\mathbb{Z}^2$-filtration on $C$. We call it {\it the $\mathbb{Z}^2$-filtration induced by the ordered pair} $(\{\mathcal{F}^1_i\}_{i \in \mathbb{Z}}, \{\mathcal{F}^2_j\}_{j \in \mathbb{Z}})$. When $C$ is endowed with such an induced $\mathbb{Z}^2$-filtration, we denote by $C^r$ the $\mathbb{Z}^2$-filtered chain complex with the same underlying complex $C$ endowed with the $\mathbb{Z}^2$-filtration induced by $(\{\mathcal{F}^2_i\}, \{\mathcal{F}^1_j\})$. We call $C^r$ {\it the reflection of $C$}. For any two $P$-filtered chain complexes $C$ and $C'$, a map $f: C \to C'$ is said to be \textit{$P$-filtered} if $f(C_R) \subset C'_R$ for any closed region $R$. Two $P$-filtered chain complexes $C$ and $C'$ are \textit{$P$-filtered homotopy equivalent over $\Lambda$} (and denoted $C \heq{P}{\Lambda} C'$) if there exists a chain homotopy equivalence map $f: C \to C'$ over $\Lambda$, a chain homotopy inverse of $f$ and their chain homotopies such that all of the four maps are $P$-filtered and graded. Such $f$ is called a \textit{$P$-filtered homotopy equivalence map over $\Lambda$}. Particularly, we call the above $f$ a \textit{$P$-filtered chain isomorphism over $\Lambda$} if $f$ is a chain isomorphism. It is obvious that if $C \heq{P}{\Lambda} C'$, then $C \heq{P}{\Lambda'} C'$ for any subalgebra $\Lambda' \subset \Lambda$. Moreover, it is easy to show the following lemma. \begin{proposition} \label{exact} Let $C$ and $C'$ be $P$-filtered chain complexes. If $C \heq{P}{\Lambda} C'$, then for any closed regions $R \subset R'$, we have an isomorphism between the long exact sequences of $\Lambda'$-modules: $$ \begin{CD} \cdots @>{\partial_{*}}>> H_*(C_R) @>{i_{*}}>> H_*(C_{R'}) @>{p_{*}}>> H_*(C_{R'}/C_{R}) @>{\partial_{*}}>> \cdots \\ @. @V{\cong}VV @V{\cong}VV @V{\cong}VV @. \\ \cdots @>{\partial_{*}}>> H_*(C'_R) @>{i_{*}}>> H_*(C'_{R'}) @>{p_{*}}>> H_*(C'_{R'}/C'_{R}) @>{\partial_{*}}>> \cdots \\ \end{CD} $$ Here, $i \colon C_R \to C_{R'}$ (resp.\ $p \colon C_{R'} \to C_{R'}/C_R$) denote the inclusion (resp.\ the projection), and $\Lambda'$ is the maximal subalgebra of $\Lambda$ so that all of $C_R$, $C_{R'}$, $C'_{R}$ and $C'_{R'}$ are $\Lambda'$-submodules of $C$ and $C'$, respectively. Moreover, the above isomorphism induces an isomorphism between the long exact sequences of graded $\mathbb{F}$-vector spaces: $$ \begin{CD} \cdots @>{\partial_{*,n+1}}>> H_n(C_R) @>{i_{*,n}}>> H_n(C_{R'}) @>{p_{*,n}}>> H_n(C_{R'}/C_{R}) @>{\partial_{*,n}}>> \cdots \\ @. @V{\cong}VV @V{\cong}VV @V{\cong}VV @. \\ \cdots @>{\partial_{*,n+1}}>> H_n(C'_R) @>{i_{*,n}}>> H_n(C'_{R'}) @>{p_{*,n}}>> H_n(C'_{R'}/C'_{R}) @>{\partial_{*,n}}>> \cdots \\ \end{CD} $$ \end{proposition} \subsubsection{$CFK^{\infty}$ and $CFK^-$} \label{subsubsec: CFK} To each knot $K$ in $S^3$, Ozsv\'ath and Szab\'o \cite{OS04knot} associate a $\mathbb{Z}^2$-filtered chain complex $CFK^{\infty}(K)$ over $\Lambda := \mathbb{F}[U, U^{-1}]$ such that if two knots $K$ and $J$ are isotopic, then $CFK^{\infty}(K) \heq{\mathbb{Z}^2}{\Lambda} CFK^{\infty}(J)$. Let $C := CFK^{\infty}(K)$. Here we summarize the algrabraic properties: \begin{enumerate} \item $C$ is a chain complex over $\Lambda$ with decomposition $C= \bigoplus_{n \in \mathbb{Z}}C_n$. The grading of a homogeneous element $x$ is denoted $\gr(x)$ and called \textit{the Maslov grading} of $x$. \item $C$ has a $\mathbb{Z}$-filtration $\{\falex{j}\}_{j \in \mathbb{Z}}$ called \textit{the Alexander filtration}. The filtration level of an element $x \in C$ is denoted $\Alex(x)$ (i.e.\ $\Alex(x) := \min \{ j \in \mathbb{Z} \mid x \in \falex{j} \}$). \item $C$ also has a $\mathbb{Z}$-filtration $\{ \falg{i} \}_{i \in \mathbb{Z}}$ called {\it the algebraic filtration}. The filtration level of an element $x$ is denoted $\Alg(x)$. When we regard $C$ as a $\mathbb{Z}^2$-filtered complex, we use the $\mathbb{Z}^2$-filtration induced by the ordered pair $(\{\falg{i}\}_{i \in \mathbb{Z}}, \{\falex{j}\}_{j \in \mathbb{Z}})$. \item The action of $U$ lowers the Maslov grading by $2$, and lowers the Alexander and algebraic filtration levels by $1$. \item $C$ is a free $\Lambda$-module with finite rank, and there exists a basis $\{x_k\}_{1 \leq k \leq r}$ such that \begin{itemize} \item each $x_k$ is homogeneous with respect to the homological grading, \item $\falex{0}$ is a free $\mathbb{F}[U]$-module with a basis $\{ U^{\Alex(x_k)} x_k \}_{1 \leq k \leq r}$, and \item $\falg{0}$ is a free $\mathbb{F}[U]$-module with a basis $\{ U^{\Alg(x_k)} x_k \}_{1 \leq k \leq r}$. \end{itemize} We call such $\{x_k\}_{1 \leq k \leq r}$ a \textit{filtered basis}. \item There exists a $\mathbb{Z}^2$-filtered homotopy equivalence map $\iota: C \to C^r$ over $\Lambda$. \item Regard $\Lambda$ as a chain complex with trivial boundary map, and define the homological grading by $$ \Lambda_n = \left\{ \begin{array}{ll} \{0, U^{-n/2}\} &(n: \text{ even})\\ 0 & (n: \text{ odd}) \end{array} \right. $$ and the Alexander and algebraic filtrations by $$ \falex{i}(\Lambda)=\falg{i}(\Lambda) = U^{-i} \cdot \mathbb{F}[U]. $$ Then there exists a $\mathbb{Z}$-filtered homotopy equivalence map $f_{\Alex}$ (resp.\ $f_{\Alg}$) $:C \to \Lambda$ over $\Lambda$ with respect to the Alexander (resp.\ algebraic) filtrations. \end{enumerate} \begin{remark} The above seven conditions are used as the axioms of {\it formal knot complexes}, to which we can generalize the definitions of the invariants $\tau(K)$, $\Upsilon_K$ and $d(S^3_{p/q}(K))$. For more details, see \cite[Section 2]{2019arXiv190709116S}. \end{remark} Particularly, the subcomplex $\falg{0}$ is denoted by $CFK^-(K)$. \subsubsection{The dual of $CFK^{\infty}$} Next, we discuss the following duality theorem for $CFK^{\infty}$, which will be used to prove the duality theorem for $\mathcal{G}_0$ (\Cref{thm: duality}) and to compute $d(S^3_{p/q}(K))$ with $p/q < 0$. \begin{theorem}[\text{\cite[Proposition~3.7]{OS04knot}}] \label{dual thm} For any knot $K$ and its mirror $K^*$, we have \[ CFK^\infty(K^*)\heq{\mathbb{Z}^2}{\Lambda}(CFK^\infty(K))^*, \] where \[(CFK^\infty(K))^* := \Hom_{\Lambda}(CFK^\infty(K),\Lambda) \] is the dual complex of $CFK^{\infty}(K)$. \end{theorem} \begin{remark} Precisely, \cite[Proposition~3.7]{OS04knot} states the theorem for the knot Floer homology $\widehat{HFK}$, while the same proof can be applied to $CFK^{\infty}$. \end{remark} Here we explain how to write the Maslov graiding and the Alexander and algebraic filtrations on $(CFK^\infty(K))^*$ in the words of $CFK^{\infty}(K)$. (The following arguments follow from \cite[Subsection 2.4]{2019arXiv190709116S}.) Let $C := CFK^{\infty}(K)$. Define a $\mathbb{F}$-linear map $\varepsilon : \Lambda \to \mathbb{F}$ by $\varepsilon (p(U)) = p(0)$ for each $p(U) \in \Lambda$ (i.e. $\varepsilon$ maps a Laurent polynomial to its constant term). Then, the Maslov graiding and the Alexander and algebraic filtrations on $C^*=(CFK^\infty(K))^*$ are given by \[ C^*_n= \left\{ \varphi \in C^* \ \middle|\ \varepsilon \circ \varphi (\bigoplus_{m \neq -n} C_m) = \{ 0\} \right\}, \] \[ \falex{j}(C^*)= \left\{ \varphi \in C^* \ \middle|\ \varepsilon \circ \varphi (\falex{-j-1}) = \{ 0\} \right\}, \] and \[ \falg{i}(C^*)= \left\{ \varphi \in C^* \ \middle|\ \varepsilon \circ \varphi (\falg{-i-1}) = \{ 0\} \right\}. \] In addition, we also note that for any filtered basis $\{ x_k \}_{1\leq k \leq r}$ for $C$, the dual basis $\{ x^*_k \}_{1\leq k \leq r}$ becomes a filtered basis for $C^*$. Here we also mention the following lemma. \begin{lemma}[\text{\cite[Lemma 2.17]{2019arXiv190709116S}}] \label{lem: dual epsilon} Let $C := CFK^\infty(K)$. Then the $\mathbb{F}$-linear map $\varepsilon_n:C^*_{-n} \to \Hom_{\mathbb{F}}(C_n,\mathbb{F})$ defined by $\varphi \mapsto \varepsilon \circ \varphi$ is a cochain isomorphism (where we see $\{C^*_{-n}\}_{n \in \mathbb{Z}}$ as a graded cochain complex over $\mathbb{F}$). In particular, we have $\mathbb{F}$-linear isomorphisms $$ H_{-n}(C^*) \cong H^n(C_*;\mathbb{F}) \cong \Hom_{\mathbb{F}}(H_n(C_*),\mathbb{F}), $$ where the first isomorphism is induced from $\varepsilon_n$. \end{lemma} \subsection{$\nu^+$-equivalence} \label{np equivalence} The {\it $\nu^+$-equivalence} is an equivalence relation on knots and regarded as a $CFK^{\infty}$-version of knot concordance. First, we note that the seventh property of $CFK^\infty$ implies $$ H_*(CFK^\infty(K)) \cong \Lambda $$ and $$ H_n(CFK^\infty(K)) \cong \begin{cases} \mathbb{F} & (n \colon \text{even})\\ 0 & (n \colon \text{odd}) \end{cases} $$ for any knot $K$. (This property is called {\it the global triviality}.) Here we consider morphisms between formal knot complexes in terms of the global triviality. For two knots $K$ and $J$, a chain map $f \colon CFK^\infty(K) \to CFK^\infty(J)$ over $\Lambda$ is a {\it $\mathbb{Z}^2$-filtered quasi-isomorphism} if $f$ is $\mathbb{Z}^2$-filtered, graded, and induces an isomorphism $f_* \colon H_*(CFK^\infty(K)) \to H_*(CFK^\infty(J))$. Now, {\it the $\nu^+$-equivalence} (or {\it local equivalence}) is defined as follows. \begin{definition} Two knots $K$ and $J$ are {\it $\nu^+$-equivalent} (and denoted $K \overset{\np}{\sim} J$) if there exist $\mathbb{Z}^2$-filtered quasi-isomorphisms $f \colon CFK^\infty(K) \to CFK^\infty(J)$ and $g \colon CFK^\infty(J) \to CFK^\infty(K)$. \end{definition} \begin{remark} Originally, the $\nu^+$-equivalence is defined by using {\it Hom-Wu's $\nu^+$-invariant} \cite{HW16}, which is a $\mathbb{Z}_{\geq 0}$-valued knot concordance invariant. Namely, two knots $K$ and $J$ are $\nu^+$-equivalent if and only if the equalities $\nu^+(K \# -J^*) = \nu^+(-K^* \# J)=0$ hold. For more details, see \cite[Section 2]{2019arXiv190709116S}. \end{remark} It is obvious that $\overset{\np}{\sim}$ is an equivalence relation on knots. We call the equivalence class of a knot $K$ under $\overset{\np}{\sim}$ \textit{the $\nu^+$-equivalence class} or \textit{$\nu^+$-class of $K$}, and denote it by $[K]_{\nu^+}$. The quotient set of knots under $\overset{\np}{\sim}$ is denoted by $\mathcal{C}_{\nu^+}$. Then, we have the following theorem. \begin{theorem}[\text{\cite{Hom17}, \cite[Theorem 2.37]{2019arXiv190709116S}}] If two knots are concordant, then they are $\nu^+$-equivalent. Moreover, the operation $[K]_{\nu^+} + [J]_{\nu^+} := [K\# J]_{\nu^+}$ endows $\mathcal{C}_{\nu^+}$ with an abelian group strucutre so that the surjective map \[ \mathcal{C} \to \mathcal{C}_{\nu^+},\ [K] \mapsto [K]_{\nu^+} \] is a group homomorphism. \end{theorem} Here we also mention a partial order on $\mathcal{C}_{\nu^+}$. For two elements $[K]_{\nu^+},[J]_{\nu^+} \in \mathcal{C}_{\nu^+}$, we denote $[K]_{\nu^+} \leq [J]_{\nu^+}$ if there exists a $\mathbb{Z}^2$-filtered quasi-isomorphism $f \colon CFK^\infty(K) \to CFK^\infty(J)$. Then we see that the relation $\leq$ defines a partial order on $\mathcal{C}_{\nu^+}$. For the partial order, we have the following 4-genus bound. \begin{theorem}[\text{\cite[Theorem 1.5]{2019arXiv190709116S}}] \label{thm:g4-bound} Let $g_4$ denote the 4-genus of a knot $K$. Then we have \[ -g_4[T_{2,3}]_{\nu^+} \leq [K]_{\nu^+} \leq g_4[T_{2,3}]_{\nu^+}. \] \end{theorem} \subsection{The invariant $\mathcal{G}_0$} \label{the invariant G_0} Now we recall the invariant $\mathcal{G}_0$, which is given as an invariant of knots under $\nu^+$-equivalence. \subsubsection{The invariants $\widetilde{\mathcal{G}}_0$ and $\mathcal{G}_0$} \label{tG_0 and G_0} For $C:=CFK^\infty(K)$, a cycle $x\in C$ is called a {\it homological generator (of degree 0)} if $x$ is homogeneous with $\gr(x)=0$ and the homology class $[x] \in H_0(C)\cong \mathbb{F}$ is non-zero. Then, we define $$ \widetilde{\mathcal{G}}_0(K) := \{R \in \mathcal{CR}(\mathbb{Z}^2) \mid C_R \text{ contains a homological generator} \}. $$ Any element $R \in \widetilde{\mathcal{G}}_0(K)$ is called a {\it realizable region of $K$}. The set $\widetilde{\mathcal{G}}_0(K)$ behaves naturally with respect to filtered quasi-isomorphism. \begin{theorem}[\text{\cite[Thereom~5.1]{2019arXiv190709116S}}] \label{tG_0 ineq} If $[K]_{\nu^+} \leq [J]_{\nu^+}$, then $\widetilde{\mathcal{G}}_0(K) \supset \widetilde{\mathcal{G}}_0(J)$. \end{theorem} As a corollary, we have the invariance of $\widetilde{\mathcal{G}}_0$ under $\overset{\np}{\sim}$. Here $\mathcal{P}(\mathcal{CR}(\mathbb{Z}^2))$ denotes the power set of $\mathcal{CR}(\mathbb{Z}^2)$. \begin{corollary} \label{tG_0 invariance} $\widetilde{\mathcal{G}}_0(K)$ is invariant under $\nu^+$-equivalence. In particular, \[ \widetilde{\mathcal{G}}_0 \colon [K]_{\nu^+} \mapsto \widetilde{\mathcal{G}}_0(K) \] is a well-defined map $ \mathcal{C}_{\nu^+} \to \mathcal{P}(\mathcal{CR}(\mathbb{Z}^2)). $ \end{corollary} Here we note that $\widetilde{\mathcal{G}}_0(K)$ is an infinite set for any $K$. To extract the essential part of $\widetilde{\mathcal{G}}_0$, we consider the minimalization of $\widetilde{\mathcal{G}}_0$. For a subset $\mathcal{S} \subset \mathcal{CR}(\mathbb{Z}^2)$, an element $R \in \mathcal{S}$ is {\it minimal in $\mathcal{S}$} if it satisfies $$ \text{if } R' \in \mathcal{S} \text{ and } R' \subset R, \text{ then } R'=R. $$ Define the map $$ \min \colon \mathcal{P}(\mathcal{CR}(\mathbb{Z}^2)) \to \mathcal{P}(\mathcal{CR}(\mathbb{Z}^2)) $$ by $$ \mathcal{S} \mapsto \{ R \in \mathcal{S} \mid R \text{ is minimal in } \mathcal{S}\}. $$ Now we define $\mathcal{G}_0(K)$ by $$ \mathcal{G}_0(K) := \min \widetilde{\mathcal{G}}_0(K). $$ The invariance of $\mathcal{G}_0$ under $\overset{\np}{\sim}$ immediately follows from Corollary~\ref{tG_0 invariance}. Moreover, $\mathcal{G}_0(K)$ has the following nice properties. \begin{theorem}[\text{\cite[Theorem 5.7]{2019arXiv190709116S}}] \label{thm: G'_0} $\mathcal{G}_0(K)$ is non-empty and finite. \end{theorem} \begin{theorem}[\text{\cite[Corollary 5.8]{2019arXiv190709116S}}] \label{thm: minimalize} For any closed region $R$, the following holds: $$ R \in \widetilde{\mathcal{G}}_0(K) \Leftrightarrow \exists R' \in \mathcal{G}_0(K), \ R' \subset R. $$ \end{theorem} \begin{theorem}[\text{\cite[Theorem 5.16]{2019arXiv190709116S}}] \label{thm: detect zero} For any knot $K$, the following holds: $$ [K]_{\nu^+} = 0 \Leftrightarrow \mathcal{G}_0(K) = \{ R_{(0,0)}\}, $$ where $R_{(0,0)} := \{(i,j)\in \mathbb{Z}^2 \mid (i, j) \leq (0,0) \}$. \end{theorem} Here we also mention the relationship of $\mathcal{G}_0$ to filtered quasi-isomorphism. \begin{proposition}[\text{\cite[Proposotion 5.9]{2019arXiv190709116S}}] \label{prop:nuplus-ineq} If $[K]_{\nu^+} \leq [J]_{\nu^+}$, then for any $R' \in \mathcal{G}_0(J)$, there exists an element $R \in \mathcal{G}_0(K)$ with $R \subset R'$. \end{proposition} Next, as a new result on the invariant $\mathcal{G}_0$, we state the duality theorem for $\mathcal{G}_0$ which implies that $\mathcal{G}_0(K)$ and $\mathcal{G}_0(K^*)$ can recover each other. Here, for a subset $S \subset \mathbb{Z}^2$, we define the subset $-S \subset \mathbb{Z}^2$ by \[ -S := \{ (i, j) \in \mathbb{Z}^2 \mid (-i, -j) \in S \}. \] \begin{theorem} \label{thm: duality} For any knot $K$, the equalities $$ \widetilde{\mathcal{G}}_0(K^*) = \{ R \in \mathcal{CR}(\mathbb{Z}^2) \mid \forall R' \in \mathcal{G}_0(K),\ R \cap (-R') \neq \varnothing \} $$ and $$ \mathcal{G}_0(K^*) = \min\{ R \in \mathcal{CR}(\mathbb{Z}^2) \mid \forall R' \in \mathcal{G}_0(K),\ R \cap (-R') \neq \varnothing \} $$ hold. \end{theorem} \begin{remark} This equality holds for any formal knot complex. Namely, we can prove the theorem purely algebraically. \end{remark} \begin{remark} Combining with \Cref{thm: minimalize}, we can also prove the equality \[ \widetilde{\mathcal{G}}_0(K^*) = \{ R \in \mathcal{CR}(\mathbb{Z}^2) \mid \forall R' \in \widetilde{\mathcal{G}}_0(K),\ R \cap (-R') \neq \varnothing \}, \] which looks more symmetric with respect to $\widetilde{\mathcal{G}}_0$. \end{remark} \Cref{thm: duality} is proved in Appendix~\ref{sec: duality}, where we also provide a method for computing $\mathcal{G}_0(K^*)$ from $\mathcal{G}_0(K)$ algorithmically. \subsubsection{Relationship to other concordance invariants} Here we state the relationship of $\mathcal{G}_0(K)$ to the invariants $\tau(K)$, $d(S^3_{p/q}(K))$ and $\Upsilon_K$. We first explain the translation of $\{d(S^3_{p/q}(K))\}_{p/q \in \mathbb{Q}_{>0}}$ into {\it Ni-Wu's $V_k$-sequence} $\{V_k(K)\}_{k \in \mathbb{Z}_{\geq 0}}$ \cite{NW15}. Note that there is a canonical identification between the set of $\Spin^c$ structures over $S^3_{p/q}(K)$ and $\{ i \mid 0 \leq i \leq p-1\}$. This identification can be made explicit by the procedure in \cite[Section 4, Section 7]{OS11rational}. Let $d(S^3_{p/q}(K), i)$ denote the correction term of $S^3_{p/q}(K)$ with the $i$-th $\Spin^c$ structure ($0 \leq i \leq p-1$). \begin{proposition}[\text{\cite[Proposition~1.6]{NW15}}] \label{V_k d} For any knot $K$, $p/q > 0$ and $0 \leq i \leq p-1$, the equality $$ d(S^3_{p/q}(K),i) = d(S^3_{p/q}(O),i) - 2 \max \left\{ V_{\lfloor \frac{i}{q} \rfloor}(K), V_{\lfloor \frac{p+q -1-i}{q} \rfloor}(K) \right\} $$ holds, where $O$ denotes the unknot and $\lfloor \cdot \rfloor$ is the floor function. \end{proposition} Since we have the orientation-preserving diffeomorphism \[ S^3_{-p/q}(K) \cong -S^3_{p/q}(K^*), \] Proposition~\ref{V_k d} shows that the two sequences $\{V_k(K)\}$ and $\{V_k(K^*)\}$ determine all values of $\{d(S^3_{p/q}(K))\}_{p/q \in \mathbb{Q}}$. Here we also need to introduce several classes of closed regions. For each coordinate $(k,l) \in \mathbb{Z}^2$, we define the {\it simple region $R_{(k,l)}$ with the corner} $(k,l)$ by \[ R_{(k,l)}:=\{(i,j) \in \mathbb{Z}^2 \mid i \leq k \text{ and } j \leq l \}. \] In addition, for each $t \in [0,2]$ and $s \in \mathbb{R}$, we define the {\it linear region $R^t(s)$ with respect to $(t,s)$} by \[ R^t(s) := \left\{(i,j) \in \mathbb{Z}^2 \ \middle|\ (1-\frac{t}{2})i + \frac{t}{2}j \leq s \right\}. \] Now, the relationship of $\mathcal{G}_0(K)$ to the invariants $\tau(K)$, $V_k(K)$ and $\Upsilon_K$ is stated as follows. \begin{theorem}[\text{\cite[Proposition 5.17]{2019arXiv190709116S}}] \label{thm: G_0 and others} The invariants $\tau(K)$, $V_k(K)$ and $\Upsilon_K$ are determined from $\mathcal{G}_0(K)$ by the formulas: \begin{eqnarray*} \tau(K)&=&\min\{m \in \mathbb{Z} \mid \exists R \in \mathcal{G}_0(K), R \subset (\{ i \leq -1\} \cup R_{(0,m)}) \}\\ V_k(K)&=&\min\{m \in \mathbb{Z}_{\geq 0} \mid \exists R \in \mathcal{G}_0(K), R \subset R_{(m,k+m)} \}\\ \Upsilon_K(t)&=&-2 \left( \min\{s \in \mathbb{R} \mid \exists R \in \mathcal{G}_0(K), R \subset R^t(s)\} \right), \end{eqnarray*} \end{theorem} Figure~\ref{regions for concordance invariants} describes the family of closed regions corresponding to each concordance invariant. \begin{figure}[ht] \centering \input{tikz/tikz_invs-from-G_0.tex} \caption{Closed regions corresponding to $\tau$, $V_k$, and $\Upsilon$.} \label{regions for concordance invariants} \end{figure} Lastly, we give a formula for computing $V_k(K^*)$ from $\mathcal{G}_0(K)$, which is obtained via the duality theorem for $\mathcal{G}_0$. \begin{proposition} \label{prop: G_0 and mirror V_k} The invariant $V_k(K^*)$ is determined from $\mathcal{G}_0(K)$ by the formula \[ V_k(K^*) = \min\{m \in \mathbb{Z}_{\geq 0} \mid \forall R \in \mathcal{G}_0(K),\ (-m, -k-m) \in R\}. \] \end{proposition} \begin{proof} For any $R \in \mathcal{CR}(\mathbb{Z}^2)$, we see that \begin{eqnarray*} (-m,-k-m) \in R &\Leftrightarrow& R \cap (-R_{(m,k+m)}) \neq \varnothing\\ &\Leftrightarrow& R_{(m,k+m)} \cap (-R) \neq \varnothing. \end{eqnarray*} Therefore, it follows from \Cref{thm: duality} that $R_{(m,k+m)}\in \widetilde{\mathcal{G}}_0(K^*)$ if and only if any $R \in \mathcal{G}_0(K)$ contains $(-m,-k-m)$. Now, by \Cref{thm: minimalize} and \Cref{thm: G_0 and others}, we have \begin{eqnarray*} V_k(K^*)&=&\min\{m \in \mathbb{Z}_{\geq 0} \mid \exists R \in \mathcal{G}_0(K^*), R \subset R_{(m,k+m)} \}\\ &=&\min\{m \in \mathbb{Z}_{\geq 0} \mid R_{(m, k+m)} \in \widetilde{\mathcal{G}}_0(K^*) \}\\ &=&\min\{m \in \mathbb{Z}_{\geq 0} \mid \forall R \in \mathcal{G}_0(K), (-m,-k-m) \in R \}. \end{eqnarray*} \end{proof} \subsubsection{On the range of $\mathcal{G}_0$} \label{range} Here we discuss the range of $\mathcal{G}_0$. We first introduce several notions of closed regions. For any subset $S \subset \mathbb{Z}^2$, define the {\it closure} of $S$ by $$ cl(S) := \bigcup_{(i,j) \in S} R_{(i,j)}. $$ Then we also have $cl(S) \in \mathcal{CR}(\mathbb{Z}^2)$. Moreover, the equality $$ cl(S) = \bigcap_{R \in \mathcal{CR}(\mathbb{Z}^2), S \subset R} R $$ holds. We say that a closed region $R \in \mathcal{CR}(\mathbb{Z}^2)$ is a {\it semi-simple region} if there exists a non-empty finite subset $S \subset \mathbb{Z}^2$ such that $R=cl(S)$. As examples of semi-simple regions, we define the {\it closure} of any chain $x = \sum_{1 \leq k \leq r} p_k(U) x_k \in CFK^\infty(K)$ by \[ cl(x) := cl\left\{(\Alg(U^{l(p_k)}x_k), \Alex(U^{l(p_k)}x_k)) \ \middle| \ \begin{array}{ll} 1 \leq k \leq r\\ p_k(U) \neq 0 \end{array} \right\}, \] where $l(p_k)$ denotes the lowest degree of $p_k(U)\in \Lambda$. Note that the equality \[ cl(x) = \bigcap_{R \in \mathcal{CR}, x \in (CFK^\infty(K))_R} R\] holds. (The proof is seen in \cite[Lemma 5.5]{2019arXiv190709116S}.) Let us denote the set of semi-simple regions by $\mathcal{CR}^{ss}(\mathbb{Z}^2)$. Then, the following proposition immediately follows from \cite[Theorem 5.7]{2019arXiv190709116S}. \begin{proposition} \label{prop: realizer} For any $R \in \mathcal{G}_0(K)$, there exists a homological generator $x$ of $CFK^\infty(K)$ whose closure is equal to $R$. In particular, for any knot $K$, we have \[ \mathcal{G}_0(K) \subset \mathcal{CR}^{ss}(\mathbb{Z}^2). \] \end{proposition} As a corollary, we have the following. \begin{corollary} \label{cor: realizer} For any $R \in \mathcal{G}_0(K)$, there exists a non-empty finite subset $\{x_k\}_{k=1}^n \subset CFK^\infty(K)$ of linearly independent 0-chains such that \begin{itemize} \item $\sum_{k=1}^n x_k$ is a homological generator of $CFK^\infty(K)$, and \item $R= cl(\{(\Alg(x_k), \Alex(x_k))\}_{k=1}^n)$. \end{itemize} \end{corollary} Next, we introduce the notion of the {\it corners} of a semi-simple region, which will be used in the coming sections. For a subset $S \subset \mathbb{Z}^2$, an element $s \in S$ is {\it maximal in $S$} if it satisfies $$ \text{if } s' \in S \text{ and } s' \geq s, \text{ then } s'=s. $$ Define the map $$ \max \colon \mathcal{P}(\mathbb{Z}^2) \to \mathcal{P}(\mathbb{Z}^2) $$ by $$ S \mapsto \{ s \in S \mid s \text{ is maximal in } S\}. $$ Then, for any $R \in \mathcal{CR}^{ss}(\mathbb{Z}^2)$, the set of {\it corners} of $R$ is defined by $$ c(R) := \max R. $$ Here we prove that for $R \in \mathcal{CR}^{ss}(\mathbb{Z}^2)$, the set $c(R)$ is non-empty and finite. \begin{lemma} \label{max S = max R} For any $R \in \mathcal{CR}^{ss}(\mathbb{Z}^2)$ and non-empty finite set $S \subset \mathbb{Z}^2$ satisfying $cl(S)=R$, the equality $$ \max S = c(R) $$ holds. In particular, $c(R)$ is non-empty and finite. \end{lemma} \begin{proof} For proving Lemma~\ref{max S = max R}, we use the following lemma, which is obtained as an analogy of \cite[Lemma~5.4]{2019arXiv190709116S}. \begin{lemma} \label{replace to maximal} For any $s \in S$, there exists an element $s' \in \max S$ with $s' \geq s$. \end{lemma} We first prove $\max S \subset \max R$. Let $s \in \max S$, and suppose that $p \in R$ satisfies $s \leq p$. Then, since $cl(S)=R$, there exists an element $s' \in S$ such that $p \leq s'$. Moreover, by Lemma~\ref{replace to maximal}, we have $s'' \in \max S$ with $s' \leq s''$. Now, we have $$ s \leq p \leq s' \leq s''. $$ Here, since $s'' \in S$ and $s \in \max S$, we have $s = p = s' = s''$. This implies $s \in \max R$. Next, we prove $\max S \supset \max R$. Let $p \in \max R$, and suppose that $s \in S$ satisfies $p \leq s$. Here, by Lemma~\ref{replace to maximal}, we may assume that $s \in \max S$. Then, we have $$ s \in S \subset cl(S)=R, $$ and hence the maximality of $p$ gives $p=s \in \max S$. \end{proof} Actually, the equality `$\max S = c(R)$' is a necessary and sufficient condition for a non-empty finite subset $S \subset \mathbb{Z}^2$ to generate $R \in \mathcal{CR}^{ss}(\mathbb{Z}^2)$. Namely, we have the following. \begin{lemma} \label{lem: char ss} For any semi-simple region $R$, a non-empty finite set $S \subset \mathbb{Z}^2$ satisfies $cl(S)=R$ if and only if $\max S = c(R)$. In particular, $cl(c(R))=R$. \end{lemma} \begin{proof} First, suppose that a non-empty finite set $S \subset \mathbb{Z}^2$ satisfies $cl(S)=R$. Then, by Lemma~\ref{max S = max R}, we have $$ \max S = \max R = c(R). $$ Conversely, suppose that a non-empty finite set $S \subset \mathbb{Z}^2$ satisfies $\max S = c(R)$, and take a non-empty finite set $S' \subset \mathbb{Z}^2$ with $cl(S')=R$. Then we have $$ \max S' = \max R = c(R)= \max S. $$ Here, by Lemma~\ref{replace to maximal}, we see that $$ cl(\max F) = cl(F) $$ for any finite subset $F \subset \mathbb{Z}^2$. This implies $$ cl(S) = cl(\max S) = cl(\max S') = cl(S') = R. $$ \end{proof} \section{Translating $\mathcal{G}_0$ into $CFK^-$} \label{sec: translate G_0} In this section, we introduce shifted versions of $\mathcal{G}_0(K)$ which enable us to define $\mathcal{G}_0$-type invariants for $CFK^-(K)$ and $C^-(G)$. Moreover, we prove that after shifting sufficiently large times, such shifted $\mathcal{G}_0$ for $CFK^-(K)$ recovers the original $\mathcal{G}_0(K)$. \subsection{Shifted $\mathcal{G}_0$} \label{generalization of G_0} Let $C$ be a $\mathbb{Z}^2$-filtered chain complex over $\mathbb{F}$ such that $H_n(C) \cong \mathbb{F}$ for given $n \in \mathbb{Z}$. Then, a homogeneous cycle $x \in C_n$ is called a {\it homological generator of degree $n$} if the homology class $[x] \in H_n(C) \cong \mathbb{F}$ is non-zero. Now we define $$ \widetilde{\mathcal{G}}_0^{(n)}(C) := \{R \in \mathcal{CR}(\mathbb{Z}^2) \mid \text{$C_R$ contains a homological generator of degree $n$}\} $$ and $$ \mathcal{G}_0^{(n)}(C) := \min \widetilde{\mathcal{G}}_0^{(n)}(C). $$ Let $C$ and $C'$ be two $\mathbb{Z}^2$-filtered chain complexes over $\mathbb{F}$ with $H_n(C) \cong H_n(C') \cong \mathbb{F}$ for $n \in \mathbb{Z}$. Then, a chain map $f \colon C \to C'$ over $\mathbb{F}$ is a {\it $\mathbb{Z}^2$-filtered quasi-isomorphism at degree $n$} if $f$ is $\mathbb{Z}^2$-filtered, graded and induces an isomorphism $f_{*,n} \colon H_n(C) \to H_n(C')$. \begin{proposition} \label{prop: G_0 deg n} Let $C$ and $C'$ be two $\mathbb{Z}^2$-filtered chain complexes over $\mathbb{F}$ with $H_n(C) \cong H_n(C') \cong \mathbb{F}$ for given $n \in \mathbb{Z}$. If there exists a $\mathbb{Z}^2$-filtered quasi-isomorphism $f \colon C \to C'$ at degree $n$, then we have $$ \widetilde{\mathcal{G}}_0^{(n)}(C) \subset \widetilde{\mathcal{G}}_0^{(n)}(C'). $$ In particular, if $C \heq{\mathbb{Z}^2}{\mathbb{F}} C'$, then we have $$ \widetilde{\mathcal{G}}_0^{(n)}(C) = \widetilde{\mathcal{G}}_0^{(n)}(C') $$ and $$ \mathcal{G}_0^{(n)}(C) = \mathcal{G}_0^{(n)}(C'). $$ \end{proposition} \begin{proof} If $x\in C_R$ is a homological generator of degree $n$, then $f(x)\in C'$ is also a homological generator of degree $n$, and $f(x) \in f(C_R) \subset C'_R$. \end{proof} Here let us consider the case of $CFK^{\infty}$. By the global triviality, we can define $\mathcal{G}^{(n)}_{0}(CFK^\infty(K))$ for any even integer $n$. We show that all $\mathcal{G}^{(n)}_{0}(CFK^\infty(K))$ can recover one another. To show that, we define {\it shifts} of closed regions. Let $R \in \mathcal{CR}(\mathbb{Z}^2)$ and $s \in \mathbb{Z}$. Then we define $$ R[s] := \{(i,j) \in \mathbb{Z}^2 \mid (i+s,j+s) \in R\}. $$ For instance, a simple region $R_{(k,l)}$ is shifted to $R_{(k-s,l-s)}$ by $s \in \mathbb{Z}$, i.e.\ $(R_{(k,l)})[s]=R_{(k-s,l-s)}$. Obviously, for any $s,t \in \mathbb{Z}$, we have $$ (R[s])[t]=R[s+t] $$ and $$ R[0]=R. $$ Next, for $\mathcal{S} \subset \mathcal{CR}(\mathbb{Z}^2)$ and $s \in \mathbb{Z}$, we define $$ \mathcal{S}[s] := \{R[s] \in \mathcal{CR}(\mathbb{Z}^2) \mid R \in \mathcal{S} \}, $$ and then we have $(\mathcal{S}[s])[t]= \mathcal{S}[s+t]$ and $\mathcal{S}[0]=\mathcal{S}$. \begin{proposition} \label{prop: shifted G_0 for CFK} For any knot $K$ and $s \in \mathbb{Z}$, we have $$ \widetilde{\mathcal{G}}_0^{(-2s)}(CFK^\infty(K))= \widetilde{\mathcal{G}}_0(K)[s]. $$ and $$ \mathcal{G}_0^{(-2s)}(CFK^\infty(K))= \mathcal{G}_0(K)[s]. $$ \end{proposition} \begin{proof} By the fourth property of $CFK^\infty$, the following hold: \begin{itemize} \item An element $x \in C$ is a homological generator of degree $-2s$ if and only if $U^{-s}x$ is a homological generator (of degree $0$). \item The equality $U^{-s}\cdot C_R = C_{R[-s]}$ holds. \end{itemize} In particular, $C_R$ contains a homological generator of degree $-2s$ if and only if $C_{R[-s]}$ contains a homological generator (of degree 0). This implies \begin{eqnarray*} R \in \widetilde{\mathcal{G}}^{(-2s)}_0(C) &\Leftrightarrow& R[-s] \in \widetilde{\mathcal{G}}_0(C)\\ &\Leftrightarrow& R \in \widetilde{\mathcal{G}}_0(C)[s]. \end{eqnarray*} This completes the proof. \end{proof} As corollaries of Proposition~\ref{prop: shifted G_0 for CFK}, we see that $\mathcal{G}_0^{(-2s)}(C)$ shares several nice properties such as Theorem~\ref{thm: G'_0} and Theorem~\ref{thm: minimalize} with $\mathcal{G}_0(C)$. \begin{corollary} \label{cor: shifted G'_0} For any $s \in \mathbb{Z}$, the set $\mathcal{G}^{(-2s)}_0(C)$ is non-empty and finite. \end{corollary} \begin{corollary} \label{cor: shifted minimalize} For any $s \in \mathbb{Z}$ and closed region $R$, the following holds: $$ R \in \widetilde{\mathcal{G}}^{(-2s)}_0(C) \Leftrightarrow \exists R' \in \mathcal{G}^{(-2s)}_0(C), \ R' \subset R. $$ \end{corollary} \subsection{Recovering $\mathcal{G}_0(K)$ from $\mathcal{G}_0^{(-2s)}(CFK^-(K))$} Let $C := CFK^\infty(K)$. Here we consider $\mathcal{G}^{(n)}_0(CFK^-(K))= \mathcal{G}^{(n)}_0(\mathcal{F}^{\Alg}_0(C))$. By the seventh property of $CFK^\infty$, we have the isomorphism $$ H_{n}(\mathcal{F}^{\Alg}_0(C))\cong H_n(\mathcal{F}^{\Alg}_0(\Lambda))\cong \begin{cases} \mathbb{F} & (n \leq 0 \text{ and $n \colon$ even})\\ 0 & (\text{otherwise}) \end{cases}. $$ Therefore, we can define $\widetilde{\mathcal{G}}_0^{(-2s)}(\mathcal{F}^{\Alg}_0(C))$ and $\mathcal{G}_0^{(-2s)}(\mathcal{F}^{\Alg}_0(C))$ for any $s \in \mathbb{Z}_{\geq 0}$. Here, $\mathcal{F}^{\Alg}_0(C)$ is regarded as a $\mathbb{Z}^2$-filtered chain complex (over $\mathbb{F}[U]$) by $$ (\mathcal{F}^{\Alg}_0(C))_R := \mathcal{F}^{\Alg}_0(C) \cap C_R $$ for each $R \in \mathcal{CR}(\mathbb{Z}^2)$. In order to state the relation between $\mathcal{G}_0^{(-2s)}(\mathcal{F}^{\Alg}_0(C))$ and $\mathcal{G}_0(C)$, we first introduce the {\it shift number} of a closed region. For any $R \in \mathcal{CR}(\mathbb{Z}^2)$, we define the \textit{shift number} of $R$ by \[ \shift(R) := \max\{ i \in \mathbb{Z} \mid \exists j \in \mathbb{Z} \ \text{s.t.}\ (i, j) \in R \}. \] Here we define $\shift(\varnothing) := -\infty$ and $\shift(R):= \infty$ if there exists no integer $M$ such that $R \subset \{i\leq M\}$. Note that $\shift(R)$ is finite if $R \in \mathcal{CR}^{ss}(\mathbb{Z}^2)$. In addition, if $R \subset R'$, then $\shift(R) \leq \shift(R')$. Now, we can state the relation between $\mathcal{G}^{(-2s)}_0(CFK^-(K))$ and $\mathcal{G}_0(K)$ as follows. \begin{theorem} \label{thm: shifted G_0 for CFK^-} For any knot $K$ and $s \in \mathbb{Z}_{\geq 0}$, we have $$ \mathcal{G}_0^{(-2s)}(CFK^-(K))= \{R[s] \mid R \in \mathcal{G}_0(K), \ \shift(R)\leq s \}. $$ \end{theorem} To prove \Cref{thm: shifted G_0 for CFK^-}, we use the following three lemmas. \begin{lemma} \label{lem: CFK^- 1} For any $R \in \mathcal{CR}(\mathbb{Z}^2)$, we have $$ (\mathcal{F}^{\Alg}_0(C))_R = C_{R \cap \{i \leq 0\}} = (\mathcal{F}^{\Alg}_0(C))_{R \cap \{i \leq 0 \}}. $$ \end{lemma} \begin{proof} By \cite[Lemma 2.9]{2019arXiv190709116S}, there exists a direct decomposition \[ C = \bigoplus_{(i,j) \in \mathbb{Z}} C_{(i,j)} \] as a $\mathbb{F}$-vector space such that for any $R \in \mathcal{CR}(\mathbb{Z}^2)$, the equality $$ C_R = \bigoplus_{(i,j) \in R} C_{(i,j)} $$ holds. Now, since $\mathcal{F}^{\Alg}_0(C)=C_{\{i \leq 0\}}$, it is easy to see that \begin{eqnarray*} (\mathcal{F}^{\Alg}_0(C))_R &=& C_{\{i \leq 0\} } \cap C_R\\ &=& (\bigoplus_{i \leq 0}C_{(i,j)} ) \cap (\bigoplus_{(i,j)\in R}C_{(i,j)} )\\ &=& \bigoplus_{(i,j)\in R \cap \{i \leq 0\}}C_{(i,j)} = C_{R \cap \{ i\leq 0\}}. \end{eqnarray*} Now, the second equality follows from $ C_{R \cap \{i \leq 0\}} = (\mathcal{F}^{\Alg}_0(C))_R \subset \mathcal{F}^{\Alg}_0(C). $ \end{proof} \begin{lemma} \label{lem: CFK^- 2} If $R \in \widetilde{\mathcal{G}}^{(-2s)}_0(\mathcal{F}^{\Alg}_0(C))$, then there exists an element $R' \in \mathcal{G}^{(-2s)}_0(C)$ such that $R' \subset R$ and $\shift(R') \leq 0$. \end{lemma} \begin{proof} Since $(C, \{\mathcal{F}^{\Alg}_i(C)\}) \heq{\mathbb{Z}}{\mathbb{F}} (\Lambda, \{\mathcal{F}^{\Alg}_i(\Lambda)\})$, the inclusion $i \colon \mathcal{F}^{\Alg}_0(C) \to C$ induces the isomorphism $H_{-2s}(\mathcal{F}^{\Alg}_0(C)) \cong H_{-2s}(C)$. Moreover, by the definition of the $\mathbb{Z}^2$-filtration on $\mathcal{F}^{\Alg}_0(C)$, it is obvious that the map $i$ is $\mathbb{Z}^2$-filtered. Therefore, the inclusion $i \colon \mathcal{F}^{\Alg}_0(C) \to C$ is a $\mathbb{Z}^2$-filtered quasi-isomorphism at degree $-2s$, and hence Proposition~\ref{prop: G_0 deg n} gives $$ \widetilde{\mathcal{G}}^{(-2s)}_0(\mathcal{F}^{\Alg}_0(C)) \subset \widetilde{\mathcal{G}}^{(-2s)}_0(C). $$ Now, let $R \in \widetilde{\mathcal{G}}^{(-2s)}_0(\mathcal{F}^{\Alg}_0(C))$. Then, Lemma~\ref{lem: CFK^- 1} and the above arguments imply that $$ R \cap \{i \leq 0\} \in \widetilde{\mathcal{G}}^{(-2s)}_0(\mathcal{F}^1_0(C)) \subset \widetilde{\mathcal{G}}^{(-2s)}_0(C). $$ Therefore, by Corollary~\ref{cor: shifted minimalize}, there exists an element $R' \in \mathcal{G}^{(-2s)}_0(C)$ such that $R' \subset R \cap \{i \leq 0\}$. Obviously, we have $R' \subset R$ and $$ \shift(R') \leq \shift(R \cap \{i \leq 0\}) \leq 0. $$ \end{proof} \begin{lemma} \label{lem: CFK^- 3} If $R \in \mathcal{G}^{(-2s)}_0(C)$ and $\shift(R) \leq 0$, then $R \in \widetilde{\mathcal{G}}^{(-2s)}_0(\mathcal{F}^{\Alg}_0(C))$. \end{lemma} \begin{proof} Note that for any $R \in \mathcal{CR}(\mathbb{Z}^2)$, the inequality $\shift(R) \leq 0$ holds if and only if $R \subset \{i \leq 0\}$. Now, let $R \in \mathcal{G}^{(-2s)}_0(C)$ and $\shift(R) \leq 0$. Then there exists a homological generator $x \in C$ of degree $-2s$ which satisfies $$ x \in C_{R} \subset C_{\{i \leq 0\}} = \mathcal{F}^{\Alg}_0(C). $$ In particular, $x \in \mathcal{F}^{\Alg}_0(C) \cap C_R = (\mathcal{F}^{\Alg}_0(C))_R$, and hence we can regard $x$ as a homological generator of degree $-2s$ for $\mathcal{F}^{\Alg}_0(C)$ which lies in $(\mathcal{F}^{\Alg}_0(C))_R$. This shows $R \in \widetilde{\mathcal{G}}^{(-2s)}_0(\mathcal{F}^{\Alg}_0(C))$. \end{proof} Now, let us prove \Cref{thm: shifted G_0 for CFK^-}. \defProof{Proof of \Cref{thm: shifted G_0 for CFK^-}} \begin{proof} For any $R \in \mathcal{CR}^{ss}(\mathbb{Z}^2)$, we see that $$ \shift(R) \leq s \Leftrightarrow \shift(R[s]) \leq 0. $$ Hence, Proposition~\ref{prop: shifted G_0 for CFK} implies \begin{eqnarray*} \{R[s] \mid R\in \mathcal{G}_0(K), \ \shift(R) \leq s\} &=& \{R \in \mathcal{G}_0(K)[s] \mid \shift(R) \leq 0\}\\ &=& \{R \in \mathcal{G}^{(-2s)}_0(C) \mid \shift(R) \leq 0\}. \end{eqnarray*} Therefore, it suffices to prove that $$ \mathcal{G}_0^{(-2s)}(\mathcal{F}^{\Alg}_0(C)) =\{R \in \mathcal{G}^{(-2s)}_0(C) \mid \shift(R) \leq 0\}. $$ We first suppose that $R \in \mathcal{G}_0^{(-2s)}(\mathcal{F}^{\Alg}_0(C))$. Then $R \in \widetilde{\mathcal{G}}_0^{(-2s)}(\mathcal{F}^{\Alg}_0(C))$, and hence Lemma~\ref{lem: CFK^- 2} gives an element $R' \in \mathcal{G}_0^{(-2s)}(C)$ such that $R' \subset R$ and $\shift(R') \leq 0$. Moreover, Lemma~\ref{lem: CFK^- 3} shows that $R'\in \widetilde{\mathcal{G}}_0^{(-2s)}(\mathcal{F}^{\Alg}_0(C))$. By the minimality of $R$ in $\widetilde{\mathcal{G}}_0^{(-2s)}(\mathcal{F}^{\Alg}_0(C))$, we must have $R=R' \in \mathcal{G}_0^{(-2s)}(C)$. Next, we suppose that $R \in \mathcal{G}_0^{(-2s)}(C)$ and $\shift(R) \leq 0$. Then, Lemma~\ref{lem: CFK^- 3} shows that $R \in \widetilde{\mathcal{G}}_0^{(-2s)}(\mathcal{F}^{\Alg}_0(C))$. Assume that $R' \in \widetilde{\mathcal{G}}_0^{(-2s)}(\mathcal{F}^{\Alg}_0(C))$ satisfies $R' \subset R$. Then, Lemma~\ref{lem: CFK^- 2} gives an element $R'' \in \mathcal{G}_0^{(-2s)}(C) \subset \widetilde{\mathcal{G}}_0^{(-2s)}(C)$ such that $R'' \subset R' \subset R$. By the minimality of $R$ in $\widetilde{\mathcal{G}}_0^{(-2s)}(C)$, we must have $R''=R'=R$. This implies that $R \in \mathcal{G}_0^{(-2s)}(\mathcal{F}^{\Alg}_0(C))$. \end{proof} \defProof{Proof} As a corollary of \Cref{thm: shifted G_0 for CFK^-}, we see that the sequence \[\{\mathcal{G}_0^{(-2s)}(CFK^-(K))\}_{s=0}^\infty \] converges to shifted $\mathcal{G}_0(K)$. \begin{corollary} For any $s \geq \max\{\shift (R) \mid R \in \mathcal{G}_0(K)\}$, we have \[ \mathcal{G}_0^{(-2s)}(CFK^-(K))= \mathcal{G}_0(K)[s]. \] \end{corollary} \section{Translating $\mathcal{G}_0$ into $C^-(G)$} \label{sec: translating G_0 to C^-(G)} In this section, we recall the grid complex $C^-(G)$ and its relationship to $CFK^-$. Combining such arguments with \Cref{thm: shifted G_0 for CFK^-}, we prove that $\mathcal{G}_0(K)$ is determined from $\mathcal{G}^{(-2s)}_0(C^-(G))$ for sufficiently large $s \in \mathbb{Z}_{\geq 0}$. \subsection{A review of grid complexes} \label{subsec:grid-cpx} \textit{Grid homology theory} is a combinatorial description of knot Floer homology theory, introduced by Manolescu, Ozsv\'ath, Sarkar, Szab\'o and Thurston in \cite{MOS09grid1, MOST07grid2}. For any knot $K \subset S^3$, by choosing a \textit{grid diagram} $G$ of $K$ (also called an \textit{arc representation}), one associates the \textit{grid complex} $C^-(G)$ which is filtered homotopy equivalent to $CFK^-(K)$. Here we briefly review the construction. \begin{figure}[t] \centering \input{tikz/tikz_grid.tex} \caption{Grid diagram and the corresponding knot} \end{figure} A \textit{grid diagram} $G$ lies on an $n \times n$ grid of squares, where some squares are decorated either with an $O$ or an $X$ so that \begin{itemize} \item every row contains exactly one $O$ and one $X$; \item every column contains exactly one $O$ and one $X$. \end{itemize} The number $n$ is called the \textit{grid number} of $G$. Given such data, one obtains a planar link diagram by drawing horizontal segments from the $O$'s to the $X$'s in each row, and vertical segments from the $X$'s to the $O$'s in each column, while letting the horizontal segment underpass the vertical segment at every intersection point. It is also true that every link in $S^3$ possesses such representation. We usually place the diagram in the standard plane so that bottom left corner is at the origin, each square has length one, and each $O$ and each $X$ is centered at a half-integer point. We set $\mathbb{O} = \{O_i\}_{1 \leq i \leq n},\ \mathbb{X} = \{X_i\}_{1 \leq i \leq n}$ so that each $O_i$ and $X_i$ has its center in $x = i - \frac{1}{2}$. Given a grid diagram $G$, the chain complex $C^-(G)$ is constructed as follows. First we regard $G$ as a diagram on the torus by gluing the two opposite sides. The generating set $S$ is given by $n$-tuples of intersection points between the horizontal and vertical circles, with the property that each horizontal (or vertical) circle contains exactly one intersection point. By fixing the bottom left corner of the grid, each $\mathbf{x} \in S$ can be identified with the set of permutations of length $n$ under the correspondence \[ \sigma \mapsto \mathbf{x} = \{ (i, \sigma(i)) \}_{0 \leq i < n }. \] $C^-(G)$ is generated by $S$ over the multivariate polynomial ring $\mathbb{F}[U_1, \cdots, U_n]$. The differential $\partial$ is defined as \[ \partial(\mathbf{x}) = \sum_{\mathbf{y} \in S}\sum_{r \in \mathrm{Rect}^\circ (\mathbf{x}, \mathbf{y})} U_1^{\epsilon_1} \cdots U_n^{\epsilon_n} \mathbf{y} \] where $\mathrm{Rect}^\circ(\mathbf{x}, \mathbf{y})$ denotes the set of \textit{empty rectangles} connecting $\mathbf{x}$ to $\mathbf{y}$ (which exist only when $\mathbf{x}$ and $\mathbf{y}$ are related by a single transposition), and for each empty rectangle $r$, the exponent $\epsilon_i \in \{0, 1\}$ is given by the number of intersections of $r$ and $O_i$. \Cref{fig:grid_diffential} depicts an empty rectangle $r$ connecting $\mathbf{x}$ to $\mathbf{y}$, and for this $r$ we have $\epsilon_4 = 1$. See \cite{MOST07grid2} for the precise definition. \begin{figure}[t] \centering \input{tikz/tikz_grid_differential.tex} \caption{An empty rectangle connecting $\mathbf{x}$ to $\mathbf{y}$} \label{fig:grid_diffential} \end{figure} $C^-(G)$ is endowed the \textit{Maslov grading} and the \textit{Alexander grading} as follows. If $P$ and $Q$ are sets of finitely many points in $\mathbb{R}^2$, we define $I(P, Q)$ by the number of pairs $p \in P$ and $q \in Q$ with $p < q$. We symmetrize this function as \[ J(P, Q) := I(P, Q) + I(Q, P). \] Here $J$ is extended bilinearly over formal sums of finite subsets of $\mathbb{R}^2$. For any $\mathbf{x} \in S$, we define \begin{align*} M(\mathbf{x}) &:= J(\mathbf{x} - \mathbb{O}, \mathbf{x} - \mathbb{O}) + 1, \\ A(\mathbf{x}) &:= J(\mathbf{x} - \frac{1}{2}(\mathbb{X} + \mathbb{O}), \mathbb{X} - \mathbb{O}) - \frac{n - 1}{2}. \end{align*}{} We also declare that each factor $U_i$ decreases $M$ by $-2$ and $A$ by $-1$. Then it follows that $\partial$ decreases $M$ by $-1$, while $A$ is non-increasing under $\partial$. Thus $M$ gives a homological grading of $C^-(G)$ and $A$ gives a $\mathbb{Z}$-filtration $\{\falex{j}\}$ on $C^-(G)$. We regard $C^-(G)$ as a $\mathbb{Z}$-filtered complex over $\mathbb{F}[U]$, where the action by $U$ is defined to be multiplication by $U_1$. \begin{theorem}[{\cite[Theorem 3.3]{MOS09grid1}}] \label{thm: Z-filtered C^-(G)} If $G$ is a grid diagram of a knot $K$, then we have \[(C^-(G),\falex{j}) \heq{\mathbb{Z}}{\mathbb{F}[U]} (CFK^-(K),\falex{j}). \] \end{theorem} \subsection{Computing $\mathcal{G}_0(K)$ from $C^-(G)$} \label{subsec:G_0-from-grid-cpx} Here we show that the $\mathbb{Z}$-filtered homotopy equivalence stated in \Cref{thm: Z-filtered C^-(G)} induces the $\mathbb{Z}^2$-filtered homotopy equivalence between $C^-(G)$ and $CFK^-(K)$. To show that, we introduce the {\it algebraic filtration} $\{\falg{i}\}$ on $C^-(G)$ as follows: For any $i \in \mathbb{Z}$, the $i$-th subcomplex $\falg{i}$ is defined to be the $\mathbb{F}[U]$-submodule generated by elements of the form $U_1^{a_1} \cdots U_N^{a_N} \mathbf{x} \in C^-(G)$ with $a_1 \geq -i$. By the definition of the differential on $C^-(G)$, it is obvious that $\{\falg{i}\}$ is a $\mathbb{Z}$-filtration on $C^-(G)$. Moreover, we see that \[ U \cdot \falg{i} = \falg{i-1} \] for any $i \leq 0$, and \[ \falg{i} = C^-(G) \] for any $i \geq 0$. (Note that $CFK^-(K)$ also has the same property.) We use the $\mathbb{Z}^2$-filtration induced by $(\{\falg{i}\},\{\falex{j}\})$ to regard $C^-(G)$ as a $\mathbb{Z}^2$-filtered complex over $\mathbb{F}[U]$. Then, we have the following. \begin{corollary} \label{cor: Z^2-filtered C^-(G)} If $G$ is a grid diagram of a knot $K$, then we have \[ C^-(G) \heq{\mathbb{Z}^2}{\mathbb{F}[U]} CFK^-(K). \] \end{corollary} \begin{proof} Let $C$ and $C'$ be either one of $C^-(G)$ and $CFK^-(K)$, and suppose that $f \colon C \to C'$ is a map over $\mathbb{F}[U]$. Then, obviously we have \[ f(\falg{i}(C)) \subset C' = \falg{i}(C') \] if $i \geq 0$. Moreover, if $i < 0$, then \begin{eqnarray*} f(\falg{i}(C)) &=& f(U^{-i} \cdot\falg{0}(C)) \\ &=& U^{-i} \cdot f(\falg{0}(C)) \subset U^{-i} \cdot \falg{0}(C') = \falg{i}(C'). \end{eqnarray*} Now, it is easy to see that a $\mathbb{Z}$-filtered homotopy equivalence map given by \Cref{thm: Z-filtered C^-(G)} induces $C^-(G) \heq{\mathbb{Z}^2}{\mathbb{F}[U]} CFK^-(K)$. \end{proof} Now, combining \Cref{cor: Z^2-filtered C^-(G)} with \Cref{prop: G_0 deg n} and \Cref{thm: shifted G_0 for CFK^-}, we have the following. \begin{theorem} \label{thm: G_0 and C^-(G)} For any knot $K$, grid diagram $G$ for $K$ and $s \in \mathbb{Z}_{\geq 0}$, we have $$ \mathcal{G}_0^{(-2s)}(C^-(G)) = \{R[s] \mid R \in \mathcal{G}_0(K), \ \shift(R) \leq s\}. $$ \end{theorem} We also have the following convergence theorem. \begin{corollary} For any $s \geq \max\{\shift (R) \mid R \in \mathcal{G}_0(K)\}$, we have \[ \mathcal{G}_0^{(-2s)}(C^-(G))= \mathcal{G}_0(K)[s]. \] \end{corollary} Moreover, we have the following corollary, which enables us to check the realizability of each closed region with finite shift number via $C^-(G)$. \begin{corollary} \label{cor: tG_0 and C^-(G)} For any closed region $R$ with $\shift(R) \leq s$, we have $$ R \in \widetilde{\mathcal{G}}_0(K) \Leftrightarrow R[s]\in \widetilde{\mathcal{G}}_0^{(-2s)}(C^-(G)). $$ \end{corollary} \begin{proof} First, suppose that $R \in \widetilde{\mathcal{G}}_0(K)$. Then, by \Cref{thm: minimalize}, we have $R' \in \mathcal{G}_0(K)$ with $R' \subset R$. In particular, the inequalities $\shift(R') \leq \shift (R) \leq s$ hold, and hence \Cref{thm: G_0 and C^-(G)} gives $R'[s] \in \mathcal{G}_0^{(-2s)}(C^-(G))$. This implies that $R[s]\in \widetilde{\mathcal{G}}_0^{(-2s)}(C^-(G))$. Next, suppose that $R[s] \in \widetilde{\mathcal{G}}_0^{(-2s)}(C^-(G))$. Then \Cref{cor: Z^2-filtered C^-(G)} gives $R[s] \in \widetilde{\mathcal{G}}_0^{(-2s)}(CFK^-(K))$. Moreover, \Cref{lem: CFK^- 2} and \Cref{prop: shifted G_0 for CFK} implies \[ R[s] \in \widetilde{\mathcal{G}}_0^{(-2s)}(CFK^\infty(K))= \widetilde{\mathcal{G}}_0(K)[s]. \] Therefore, we have $R \in \widetilde{\mathcal{G}}_0(K)$. \end{proof} \section{Algorithm} \label{sec: algorithm} Now we describe the algorithm to compute $\mathcal{G}_0(K)$. First we describe the four main procedures which assures that the algorithm exists, and then describe how these procedures are integrated so that the computations are done effectively. Throughout this section we assume that we are given a grid diagram $G = (\mathbb{O}, \mathbb{X})$ of $K$, and denote $C = C^-(G)$. \subsection{Enumerating the candidate regions} First we show that there exists a finite set of closed regions that is assured to include $\mathcal{G}_0(K)$. \begin{proposition} \label{prop:G_0-torus-knot} For any $g \geq 0$, \begin{align*} \mathcal{G}_0(T_{2, 2g + 1}) &= \{ R_{(i, g - i)} \mid i = 0, \cdots, g \}, \\ \mathcal{G}_0((T_{2, 2g + 1})^*) &= \{ R_{(-g, 0)} \cup R_{(-g + 1, -1)} \cup \cdots \cup R_{(0, -g)} \} \end{align*} \end{proposition} \begin{proof} It is known that $CFK^\infty((T_{2, 2r + 1})^*)$ possesses a unique homological generator $a := a_0 + \cdots a_g$ such that $(\Alg(a_i), \Alex(a_i)) = (-g + i, -i)$. This implies the latter equality. Similarly, we can easily compute $\mathcal{G}_0(T_{2,2g+1})$. (We can also determine $\mathcal{G}_0(T_{2,2g+1})$ from $\mathcal{G}_0((T_{2,2g+1})^*)$ using \Cref{char G_0 of mirror}.) \end{proof} \begin{proposition} \label{prop:candidate-cond} Let $g_3, g_4$ denote the 3-, 4- genus of a knot $K$ respectively. Any $R \in \mathcal{G}_0(K)$ satisfies the following three conditions: \begin{enumerate} \item Each corner $(i, j)$ of $R$ satisfies $|i - j| \leq g_3$. \item $R$ includes the region $R_{(-g_4, 0)} \cup R_{(-g_4 + 1, -1)} \cup \cdots \cup R_{(0, -g_4)}$. \item If $R$ contains a point $(i, g_4 - i)$ for some $i \in \{0, \ldots, g_4 \}$, then $R = R_{(i, g_4 - i)}$. \end{enumerate}{} \end{proposition} \begin{proof} By \cite[Theorem 4.5]{2019arXiv190709116S}, up to $\mathbb{Z}^2$-filtered homotopy equivalence, we may assume that $CFK^\infty$ has a filtered basis $\{x_k\}_{1 \leq k \leq r}$ such that \[ |\Alex(U^n x_k) - \Alg(U^n x_k)| \leq g_3 \] for any $1 \leq k \leq r$ and $n \in \mathbb{Z}$. Hence the first statement follows from \Cref{prop: realizer}. Next, from \cite[Section 3.1]{Pet13}, we have $\pm g[T_{2,3}]_{\nu^+} = \pm[T_{2,2g + 1}]_{\nu^+}$. Now the remaining two statements follow from \Cref{thm:g4-bound}, \Cref{prop:nuplus-ineq} and \Cref{prop:G_0-torus-knot}. \end{proof} In particular, any corner $(i, j)$ of $R \in \mathcal{G}_0(K)$ lies in the bounded area \begin{align*} |i + j| \leq g_4 &\quad ( ij \geq 0), \\ |i - j| \leq g_3 &\quad ( ij < 0). \end{align*} Thus the set of all semi-simple regions satisfying the conditions of \Cref{prop:candidate-cond} is finite. Furthermore, another strong condition can be imposed from the bigraded module structure of $\widehat{HFK}(K)$. As mentioned in \cite[Section 4.]{KP18}, by \cite[Lemma 4.5]{rasmussen2003floer}, up to $\mathbb{Z}^2$-filtered homotopy equivalence, $CFK^\infty(K)$ may be regarded as a chain complex generated by $\widehat{HFK}(K)$ over $\mathbb{F}[U, U^{-1}]$. Define \[ \widehat{S}_k := \{\ U^a x \mid M(x) \equiv k \bmod{2},\ a = (M(x) - k) / 2 \ \} \] where $x$ runs over the homogeneous generators of $\widehat{HFK}(K)$. For a closed region $R$, we say two corners $(i, j)$ and $(k, l)$ of $R$ are \textit{adjacent} if there are no corners in between them. \begin{proposition} \label{prop:HFK-hat-reduction} Any $R \in \mathcal{G}_0(K)$ satisfies the following two conditions: \begin{enumerate} \item For each corner $(i, j)$ of $R$, there exists an element $x \in \widehat{S}_0$ such that $(\Alg(x), \Alex(x)) = (i, j)$. \item For each pairs of adjacent corners $(i, j)$ and $(k, l)$ of $R$ with $i < k$, there exists an element $x \in \widehat{S}_{-1}$ such that $(\Alg(x), \Alex(x)) \leq (i, l)$. \end{enumerate} \end{proposition} \begin{proof} The first statement is obvious from \Cref{cor: realizer} and the above arguments. For the second statement, take a homological generator $z$ of degree 0 whose closure gives $R$. Decompose $z$ as $x + y$ so that $x$ consists of terms of $z$ having $\Alg \leq i$. Obviously $x, y \neq 0$ and $\partial x + \partial y = 0$. Also from the minimality of $R$, we have $\partial x \neq 0$. Thus $\partial x = \partial y$ belongs to $cl(x) \cap cl(y) = R_{(i, l)}$. \end{proof}{} If $g_3(K), g_4(K)$ and $\widehat{HFK}(K)$ are known beforehand, we take them as additional inputs. If not, then we can compute $\widehat{HFK}(K)$ (which also gives $g_3(K)$) from the simplified grid complex \[ \widetilde{C}(G) := \frac{C^-(G)}{U_1 = \cdots = U_n = 0}. \] From \cite[Theorem 3.6.]{MOST07grid2} we have \[ H(\widetilde{C}(G)) \cong \widehat{HFK}(K) \otimes W^{\otimes (n - 1)}, \] where $W$ is the two-dimensional bigraded vector space spanned by two generators, one generator in bigrading $(0, 0)$ and another in bigrading $(-1, -1)$. An effective algorithm for computing $\widehat{HFK}$ is described in \cite{BG12}. We call the set of all semi-simple regions satisfying \Cref{prop:candidate-cond} and \ref{prop:HFK-hat-reduction} the set of \textit{candidate regions}. Having enumerated the candidate regions, the remaining task is to determine the realizability for each candidate region. In fact, from the symmetry $CFK^{\infty}(K) \heq{\mathbb{Z}^2}{\Lambda} (CFK^{\infty}(K))^r$, it follows that \[ R \in \mathcal{G}_0(K) \Leftrightarrow R^r \in \mathcal{G}_0(K). \] where $R^r$ is the \textit{reflection} of $R$ \[ R^r := \{ (i, j) \mid (j, i) \in R \}. \] Thus it suffices to consider only one of the reflection pairs. For reasons to be explained in the coming sections, we check the one that has the smaller shift number. \subsection{Inflating the generators}\label{subsec:gens} Recall that the grid complex $C = C^-(G)$ is generated by the set $S$ over $\mathbb{F}[U_1, \cdots, U_n]$, where each $\mathbf{x} \in S$ corresponds one-to-one to a permutation of length $n$. \textit{Heap's algorithm} \cite{heap1963permutations} is well known in the area of computer science, which enumerates all permutations of any fixed length by a sequence of transpositions. As for the two gradings $M$ and $A$, if $\mathbf{x}, \mathbf{y} \in S$ are related by a transposition and $r$ is a rectangle that connects $\mathbf{x}$ to $\mathbf{y}$, it is known that the following relations hold: \begin{align*} M(\mathbf{y}) - M(\mathbf{x}) &= 2 \#(r \cap \mathbb{O}) - 2 \#(\mathbf{x} \cap \mathrm{Int}(r)) - 1, \\ A(\mathbf{y}) - A(\mathbf{x}) &= \#(r \cap \mathbb{O}) - \#(r \cap \mathbb{X}). \end{align*} Note that computing the two gradings sequentially from the above formulas is more effective than computing them directly using the formulas given in \Cref{subsec:grid-cpx}. Thus we obtain the generating set $S$ together with the gradings $(M(\mathbf{x}), A(\mathbf{x}))$ for each $\mathbf{x} \in S$. Next we inflate these generators by multiplying monomials in $U_1, \cdots U_n$ and regard $C$ as an (infinitely generated) chain complex over $\mathbb{F}$. Each factor $U_i$ decreases the homological degree by $2$, so for any $k \in \mathbb{Z}$, the $k$-th chain group $C_k$ can be regarded as a finite dimensional vector space over $\mathbb{F}$ with generators of the form $$ U_1^{a_1} \cdots U_N^{a_N} \mathbf{x}, \quad \deg{\mathbf{x}} - 2\textstyle{\sum_i} a_i = k. $$ We call generators of this form \textit{inflated generators}. Recall from \Cref{subsec:grid-cpx} that the $\mathbb{Z}^2$-filtration is given by \begin{align*} \Alg(U_1^{a_1} \cdots U_N^{a_N} \mathbf{x}) &= -a_1,\\ \Alex(U_1^{a_1} \cdots U_N^{a_N} \mathbf{x}) &= A(\mathbf{x}) - \sum_i a_i. \end{align*} Note that the number of inflated generators increases infinitely as the homological degree $k$ decreases. Nonetheless, from \Cref{prop:candidate-cond} (and reasons to be stated in the coming sections), we only need to consider chain groups of degree between $-2 g_3$ and 1, hence the enumeration is finite. \subsection{Computing a homological generator} Next we describe the algorithm for computing a homological generator of $C(G)$, i.e.\ a representative cycle of the unique generator of $H^-_0(G) \cong \mathbb{F}$. For efficiency, the algorithm avoids direct computation of the homology group. We call the set of points $\{ (i, n - i - 1) \mid i \in \frac{1}{2}\mathbb{Z} \}$ the \textit{anti-diagonal} of $G$. \begin{proposition} Suppose all $O$'s of $G$ lie in the anti-diagonal of $G$. The element $\mathbf{x}_0 := \{ (i, n - i - 1) \}_{0 \leq i < n}$ represents the unique generator of $H^-_0(G) \cong \mathbb{F}$. \end{proposition} \begin{proof} $M(\mathbf{x}_0) = 0$ is proved in \cite[Lemma 4.3.5]{OSS15gridbook}. For any $\mathbf{y}$ that is related to $\mathbf{x}_0$ by a transposition, there are exactly two empty rectangles that connect $\mathbf{x}_0$ to $\mathbf{y}$, and neither of them contains $O$. Thus $\partial \mathbf{x}_0 = 0$. Conversely, rectangles that connect $\mathbf{y}$ to $\mathbf{x}_0$ contains either a point of $\mathbf{x}_0$ or one of $O$, so there is no chain $c$ such that $\partial c$ contains $\mathbf{x}_0$ with coefficient 1 in its terms. Thus $\mathbf{x}_0$ is non-boundary. \end{proof} Obviously $G$ can be transformed into a grid diagram $G_0$, whose $O$'s lie in its anti-diagonal, by applying some sequence of \textit{commutation moves} (moves that swap two adjacent columns). Now reverse the sequence. Following the arguments given in \cite[Section 3.1]{MOST07grid2}, each commutation move $G_i \rightarrow G_{i+1}$ is assigned a chain homotopy equivalence map $\Phi: C(G_i) \rightarrow C(G_{i+1})$. Starting from $\mathbf{x}_0 \in C_0(G_0)$, we obtain one desired cycle in $C_0(G)$ by sequentially applying the corresponding chain maps. Here we explain the explicit computation of $\Phi$. Given two diagrams that are related by a single commutation move, we depict the move by drawing the two diagrams on the same grid, where the intermediate circle $\beta$ is to be replaced with a different circle $\gamma$. By perturbing $\gamma$ so that it intersects $\beta$ transversally at two points that do not lie in the horizontal circles, the chain map $\Phi$ is given by \[ \Phi(\mathbf{x}) = \sum_{\mathbf{y} \in S(G')}\sum_{\Pi \in \mathrm{Pent}^\circ (\mathbf{x}, \mathbf{y})} U_1^{\epsilon_1} \cdots U_n^{\epsilon_n} \mathbf{y} \] where $\mathrm{Pent}^\circ(\mathbf{x}, \mathbf{y})$ denotes the set of empty pentagons connecting $\mathbf{x}$ to $\mathbf{y}$, and for each such pentagon $\Pi$, the exponent $\epsilon_i \in \{0, 1\}$ is given by the number of intersections of $\Pi$ and $O_i$. See \cite{MOST07grid2} for precise definition. The actual computation is easy. Suppose we are swapping the $i$-th and the $(i + 1)$-th columns, so that $O_i$ and $O_{i+1}$ interchanges. Then $\epsilon_i$ is always $0$, and $\epsilon_{i+1} = 1$ if and only if, before the commutation move, $O_i$ is located at the left-upper of $O_{i + 1}$ and its vertical coordinate is between the lower and upper sides of $\Pi$. This can be seen from \Cref{fig:pentagon}, where in each of the four cases $\beta$ is drawn as a straight vertical blue line, $\gamma$ is drawn as a curve close to $\beta$. The intersection of the two curves is drawn as a small gray disk, and the other four small disks form the vertices of the pentagon $\Pi$. Note that we can ignore the $X$'s since they are not involved in both $\partial$ and $\Phi$. For the remaining exponents $\epsilon_j$, it is easy to check for each $O_j$ whether it is contained in $\Pi$. \begin{figure}[t] \centering \input{tikz/tikz_pentagon.tex} \caption{Relative positions of $O_i, O_{i + 1}$ and the pentagon $\Pi$} \label{fig:pentagon} \end{figure} \subsection{Determining the realizability} \label{subsec:realizability-by-linear-system} Having prepared the inflated generators and one homological generator $z$, we are ready to determine the realizability of the candidate regions. From \Cref{cor: tG_0 and C^-(G)}, a closed region $R$ of shift number $s$ is realizable (i.e. $R \in \widetilde{\mathcal{G}}_0(K)$) if and only if there is a cycle that is homologous to $(U_1)^s z$ and is contained in $C_{R[s]}$. This is equivalent to the condition that (the projection of) $(U_1)^s z$ is null homologous in the quotient complex $Q_R := C/C_{R[s]}$. For any $k \in \mathbb{Z}$, we can take a finite basis for $(Q_R)_k$ from the inflated generators contained in $C_k$ and modding out those contained in $(C_{R[s]})_k$. Having fixed such bases in degree $-2s$ and $-2s + 1$, the differential $\partial: (Q_R)_{-2s + 1} \rightarrow (Q_R)_{-2s}$ is represented by a matrix $A$. Thus the realizability of $R$ is equivalent to the existence of a solution $x$ of the linear system \[ A x = b \] where $b$ is the vector corresponding to $(U_1)^s z \in (Q_R)_{-2s}$. Now that we have reduced the problem to linear systems, the remaining task is purely computational. Here we only mention that the matrix $A$ is generally sparse, and becomes extremely large as $n$ increases. More details on sparse linear systems is given in \Cref{subsec:sparse-linear-system}. \begin{remark} This procedure of reducing a filtration-level type problem to linear systems can be applied to computing some other filtration-level type invariants. For example, Rasmussen's $s$-invariant (\cite{rasmussen2010}), which is an integer valued knot invariant defined by a $\mathbb{Z}$-filtration on (a deformed version of) Khovanov homology over $\mathbb{Q}$, can be computed by the same method. \end{remark} \subsection{Integration} The four procedures are integrated into a program that goes through the following six steps: \begin{enumerate} \item Enumerate the candidate regions. \item Setup the inflated generators. \item Compute one homological generator. \item Compute $\tau(K)$. \item Determine the simple regions in $\mathcal{G}_0(K)$. \item Check the remaining candidate regions. \end{enumerate} Assuming that we have gone through the first three steps, we explain how remaining three work, together with some techniques for reducing the computational cost. Before moving on, we remark that determining the realizability of a candidate region may contribute to discarding some other candidates. Namely, if we find a realizable region $R$, then we may discard candidates that include $R$, since they are obviously realizable and non-minimal. Conversely, if we find a non-realizable $R$, then we discard candidates that are included in $R$, since they are obviously non-realizable. The program terminates at any step if the candidates become empty. \subsubsection{Computing $\tau(K)$} From \Cref{thm: G_0 and others}, we know that $\tau(K)$ is given by the minimum integer $j$ such that $R_{(-1, \infty)} \cup R_{(0, j)}$ is realizable and $R_{(-1, \infty)} \cup R_{(0, j - 1)}$ is not. Here if $\tau(K) = 0$, there is a chance of an early exit: if $R_{(0, 0)}$ is realizable for both $K$ and $K^*$, then from \Cref{thm: duality} we immediately have $\mathcal{G}_0(K) = \{ R_{(0, 0)} \}$. In the following we assume that $\tau(K) \geq 0$, by replacing $G$ with its mirror if necessary. \subsubsection{Determining the simple regions in $\mathcal{G}_0(K)$} Next we narrow down the candidates by determining the simple regions in $\mathcal{G}_0(K)$. Starting from $R_{(0, g_4)}$, we check all simple candidate regions. Note that a simple region $R = R_{(i, j)}$ belongs to $\mathcal{G}_0(K)$ if and only if $R$ is realizable and $R \setminus \{(i, j)\}$ is not. This strategy actually works, as we see from the results that for prime knots $K$ with up to $11$ crossings and $\tau(K) \geq 0$, $\mathcal{G}_0(K)$ consist only of regions whose corners lie in the first quadrant, and in many cases they are all simple. \subsubsection{Checking the remaining candidates} Having determined the simple regions in $\mathcal{G}_0(K)$, we expect that the remaining candidates are all non-realizable (we still assume $\tau(K) \geq 0$). Suppose $R$ is the maximal remaining candidate. If we succeed to show that $R$ is non-realizable, then we are done. However such computation tends to be highly expensive, since $R$ has relatively high shift number. To avoid such computation whenever possible, it is useful to consider the mirror simultaneously. For any subset $R \subset \mathbb{Z}^2$, we denote by $-R$ the subset of $\mathbb{Z}^2$ given by \[ -R := \{ (i, j) \in \mathbb{Z}^2 \mid (-i, -j) \in R \}. \] \begin{lemma} For any pair of realizable regions $R$ of $K$ and $R'$ of $K^*$, we have $R \cap (-R') \neq \varnothing$. \end{lemma} \begin{lemma} Suppose $\mathcal{S}$ is a set of closed regions that includes $\mathcal{G}_0(K)$. If a closed region $R'$ intersects with $-R$ for any $R \in \mathcal{S}$, then $R'$ is a realizable region of $K^*$. \end{lemma} Thus the intermediate results obtained for $K$ can be reflected to those of $K^*$ and vice versa. After having determined the simple regions in $\mathcal{G}_0(K)$, we expect that the maximal remaining region of $K$ is non-realizable, and the minimal remaining region of $K^*$ is realizable. We select the one with smaller computational cost (which is simply estimated by the size of the corresponding matrix). This final process is continued until the candidates are empty. \subsubsection{Rejection by a smaller system} As we have seen, there are cases where we expect that a specific region $R$ is non-realizable. Suppose there is a smaller chain complex $C'$ and a $\mathbb{Z}^2$-filtration preserving chain map $f: C \rightarrow C'$. The existence of $c \in C$ such that \[ z - \partial c \in C_R \] implies the existence of $c' \in C'$ such that \[ f(z) - \partial c' \in C'_R. \] Hence by contraposition, the non-realizability of $R$ can possibly be detected using a smaller system. Our program use the chain map induced by the following projection: \[ \pi: \mathbb{F}[U_1, \cdots, U_n] \rightarrow \mathbb{F}[U_1, U_2], \quad U_1 \mapsto U_1, \ U_{\geq 2} \mapsto U_2. \] Further reduction techniques on the level of linear systems is described in \Cref{subsec:sparse-linear-system}. \section{Computational results} \label{sec: results} Here we list the computational results performed for prime knots of crossing number up to 11, that are both non-slice and homologically thick (with respect to $\widehat{HFK}$). The reason for this selection is that (i) if a knot $K$ is slice then $\mathcal{G}_0(K) = \{R_{(0, 0)}\}$, and (ii) if $K$ is homologically thin then $\mathcal{G}_0(K) = \mathcal{G}_0(\tau(K)[T_{2, 3}])$, so the non-obvious ones are those complements. We used Culler's program \texttt{Gridlink} \cite{gridlink} to produce the input grid diagrams. This program contains the preset data for prime knots with up to 12 crossings, which are extracted from Livingston's Knotinfo database \cite{Knotinfo}. It also has the \texttt{simplify} function that randomly modifies a given diagram to reduce its grid number. In addition to the input diagram data, we used values $g_3, g_4$ and $\tau$ which are read from Knotinfo, and also $\widehat{HFK}$ which are taken from \cite{BG12}. We have implemented the algorithm as described in \Cref{sec: algorithm}. With such input data set, the results listed below were obtained within an hour by running the program on a usual laptop computer\footnote{Currently we are preparing to open source the program.}. \Cref{table:G_0-types} presents the \textit{$\mathcal{G}_0$-types} of the target knots. We say that $K$ has the \textit{$\mathcal{G}_0$-type} of a typical knot $T$, when $\mathcal{G}_0(K) = \mathcal{G}_0(T)$. We use torus knots and those cables as typical knots, for example we read that $8_{19}$ has the $\mathcal{G}_0$-type of the $(3, 4)$-torus knot. The listed types are the `positive side' of the knot, that is, either one of $K$ and $K^*$ having $\tau \geq 0$. In this list, we see that there are only six types that appear as those $\mathcal{G}_0$-types. The next \Cref{table:G_0-tau-V_k} presents the values of $\mathcal{G}_0(K)$, $\tau(K)$, $V_k(K)$ and $V_k(K^*)$ for such typical knots $K$. Note that $T_{2,3; 2,1}$ (the $(2, 1)$ cable of $T_{2, 3}$) is the only type in this list that possesses a non-simple minimal region $R_{(0, 1)} \cup R_{(1, 0)}$. The algorithm for computing $\mathcal{G}_0(K^*)$ from $\mathcal{G}_0(K)$ is described in \Cref{char G_0 of mirror}. Finally \Cref{table:Upsilon} presents $\Upsilon_K$ for the typical knots $K$, where $PL[(x_0, y_0), \cdots, (x_n, y_n)]$ denotes the PL-function from $[x_0, x_n]$ to $\mathbb{R}$ whose graph is given by linearly connecting the adjacent coordinates. Computations are performed for another set of knots, that is, prime knots of crossing number 12 that are non-slice, homologically thick, and $\tau = 0$. As mentioned in \Cref{sec:intro}, results have shown that all such knots have $\mathcal{G}_0$-type of the unknot. Hence we obtain \Cref{thm: converse}. Computations for the remaining knots with $12$ crossings are under progress. \vspace{1em} \input{table/table_G0.tex} \section{The duality theorem for $\mathcal{G}_0$} \label{sec: duality} In this appendix, we prove the duality theorem for $\mathcal{G}_0$, stated as \Cref{thm: duality}. To prove the theorem, we first introduce the notion of {\it the dual} of a closed region $R$ (denoted $R^*$). Then, for any $C = CFK^\infty(K)$, it is proved that the subcomplex $C^*_R$ consists of chains which map $C_{R^*}$ to zero. In \Cref{subsec: proof of duality}, we give a proof of \Cref{thm: duality}. In \Cref{subsec: computing G_0(K^*) from G_0(K)}, we provide a method for computing $\mathcal{G}_0(K^*)$ from $\mathcal{G}_0(K)$ algorithmically. \subsection{The dual of a closed region} For any given $R \in \mathcal{CR}(\mathbb{Z}^2)$, {\it the dual region $R^*$ of $R$} is defined by $$ R^* := \mathbb{Z}^2 \setminus (-R) = \{(i,j) \in \mathbb{Z}^2 \mid (-i,-j) \not\in R\}. $$ Here we show several basic properties of dual regions. \begin{lemma} $R^* \in \mathcal{CR}(\mathbb{Z}^2)$. \end{lemma} \begin{proof} Suppose that $(k,l) \in \mathbb{Z}^2$, $(i,j) \in R^*$ and $(k,l) \leq (i,j)$. Under these assumptions, if $(k,l) \neq R^*$, then $(-k,-l) \in R$. On the other hand, since $(-i,-j) \leq (-k,-l)$, we have $(-i,-j) \in R$. This contradicts to $(i,j) \in R^*$, and hence $(k,l) \in R^*$. \end{proof} \begin{lemma} $R^{**}=R$. \end{lemma} \begin{proof} $(i,j) \in R^{**} \Leftrightarrow (-i,-j) \neq R^* \Leftrightarrow (i,j) \in R$. \end{proof} \begin{lemma} $R^* = \bigcap_{(k,l)\in R} \{ i \leq -k-1 \text{ {\rm or }} j \leq -l-1 \}$. \end{lemma} \begin{proof} We first prove $R^* \subset \bigcap_{(k,l)\in R} \{ i \leq -k-1 \text{ {\rm or }} j \leq -l-1 \}$. Let $(i,j) \in R^*$. Then, for any $(k,l) \in R$, either $-i>k$ or $-j > l$ holds. Equivalently, we have $i \leq -k-1$ or $j \leq -l-1$. Next, we prove $R^* \supset \bigcap_{(k,l)\in R} \{ i \leq -k-1 \text{ {\rm or }} j \leq -l-1 \}$. Let $(i,j) \in \cap_{(k,l) \in R}\{ i \leq -k-1 \text{ {\rm or }} j \leq -l-1 \}$. Then, for any $(k,l) \in R$, either $-i > k$ or $-j > l$ holds. In particular, we have $(-i,-j) \neq (k,l)$, and hence $(-i,-j) \not\in R$. \end{proof} Using the above lemmas, we have the following proposition, which plays an essential role in the proof of \Cref{thm: duality}. \begin{proposition} \label{dual complex and region} Let $C := CFK^\infty(K)$. Then, for any $R \in \mathcal{CR}(\mathbb{Z}^2)$, we have \[ C^*_R = \{ \psi \in C^* \mid \varepsilon \circ \psi (C_{R^*}) = 0\}. \] \end{proposition} \begin{proof} We first prove $C^*_R \subset \{ \psi \in C^* \mid \varepsilon \circ \psi (C_{R^*}) = 0\}$. Indeed, this follows from the inequalities \begin{eqnarray*} C^*_R &=& \sum_{(k,l) \in R} \falg{k}(C^*) \cap \falex{l}(C^*)\\ &=& \sum_{(k,l) \in R} \left\{ \psi \in C^* \ \middle| \ \varepsilon \circ \psi (\falg{-k-1} + \falex{-l-1})=0 \right\} \\ &=& \sum_{(k,l) \in R} \left\{ \psi \in C^* \ \middle| \ \varepsilon \circ \psi (C_{\{i \leq -k-1 \text{ {\rm or} } j \leq -l-1\}})=0 \right\} \\ &\subset& \left\{ \psi \in C^* \ \middle| \ \varepsilon \circ \psi (C_{\bigcap_{(k,l) \in R} \{i \leq -k-1 \text{ {\rm or} } j \leq -l-1\}})=0 \right\}\\ &=& \left\{ \psi \in C^* \ \middle| \ \varepsilon \circ \psi (C_{R^*})=0 \right\}. \end{eqnarray*} Next, we prove $C^*_R \supset \{ \psi \in C^* \mid \varepsilon \circ \psi (C_{R^*}) = 0\}$. Take a filtered basis $\{x_s\}_{1 \leq s \leq r}$ for $C$, and then any $\psi \in C^*$ is written as $\psi = \sum_{1 \leq s \leq r, \ t \in \mathbb{Z}} a_{s,t} U^t x^*_s$, where $a_{s,t}=0$ for all but finitely many $(s,t)$. Now, suppose that $\varepsilon \circ \psi (C_{R^*})= \varepsilon \circ \psi (C_{\bigcap_{(k,l) \in R} \{i \leq -k-1 \text{ {\rm or} } j \leq -l-1\}})=0$. Then, we see that $U^{-t}x_s \not\in C_{R^*}$ for any $(s,t)$ with $a_{s,t} \neq 0$. This implies that $R_{U^{-t}x_s}=R_{(\Alg(U^{-t}x_s), \Alex(U^{-t}x_s))} \not\in R^*$, and hence $(\Alg(U^{-t}x_s), \Alex(U^{-t}x_s)) \geq (-k,-l)$ for some $(k,l) \in R$. In particular, we have $$U^t x_s^*(\falg{-k-1}+\falex{-l-1})=0$$ for such $(k,l)$. Consequently, we have $$ \psi = \sum_{a_{s,t}\neq 0} a_{s,t} U^t x^*_s \in \sum_{(k,l) \in R} \left\{ \psi \in C^* \ \middle| \ \varepsilon \circ \psi (\falg{-k-1}+\falex{-l-1})=0 \right\} = C^*_R. $$ \end{proof} \subsection{Proof of \Cref{thm: duality}} \label{subsec: proof of duality} Now we prove \Cref{thm: duality}. \renewcommand{\thesection}{\arabic{section}} \setcounter{section}{2} \setcounter{proposition}{15} \begin{theorem} For any knot $K$, the equalities $$ \widetilde{\mathcal{G}}_0(K^*) = \{ R \in \mathcal{CR}(\mathbb{Z}^2) \mid \forall R' \in \mathcal{G}_0(K),\ R \cap (-R') \neq \varnothing \} $$ and $$ \mathcal{G}_0(K^*) = \min\{ R \in \mathcal{CR}(\mathbb{Z}^2) \mid \forall R' \in \mathcal{G}_0(K),\ R \cap (-R') \neq \varnothing \} $$ hold. \end{theorem} \renewcommand{\thesection}{\Alph{section}} \setcounter{section}{1} \setcounter{proposition}{4} The proof is an analogy of \cite[Lemma 2.28]{2019arXiv190709116S}. \begin{proof} By the definition of $\mathcal{G}_0(K)$, it is sufficient to prove the first equality. We denote by $\mathcal{G}_0(K)^{\perp}$ the right hand side of the desired equality; namely, we set \[ \mathcal{G}_0(K)^{\perp} := \{ R \in \mathcal{CR}(\mathbb{Z}^2) \mid \forall R' \in \mathcal{G}_0(K),\ R \cap (-R') \neq \varnothing \}. \] Then, by the definition of dual regions, it is obvious that \[ \mathcal{G}_0(K)^{\perp} = \{ R \in \mathcal{CR}(\mathbb{Z}^2) \mid \forall R' \in \mathcal{G}_0(K),\ R' \not\subset R^*\}. \] We first prove $\widetilde{\mathcal{G}}_0(K^*) \subset \mathcal{G}_0(K)^{\perp}$. Let $R \in \widetilde{\mathcal{G}}_0(K^*)$ and $C:=CFK^\infty(K)$. Then, by \Cref{dual complex and region}, there exists a homological generator $\varphi \in C^*_0$ lying in $$ C^*_R = \left\{ \psi \in C^* \ \middle| \ \varepsilon \circ \psi (C_{R^*})=0 \right\}. $$ In particular, we have $\varepsilon \circ \varphi(C_{R^*}) =0$, and $\varepsilon \circ \varphi$ is decomposed as $\varepsilon \circ \varphi = \widetilde{\varphi} \circ p$ where $\widetilde{\varphi} \in \Hom_{\mathbb{F}}(C_{R^*}, \mathbb{F})$ is a cocycle and $p : C \to C/C_{R^*}$ is the projection. Now, let $R' \in \mathcal{G}_0(C)$ and $x$ be a homological generator whose closure gives $R'$. Then we have $\widetilde{\varphi}(p (x))= (\varepsilon \circ \varphi)(x)=1$. This implies that $p(x) \neq 0$, and hence $R^* \not\supset R'$. Therefore, we have $R \in \mathcal{G}_0(K)^{\perp}$. Next, we prove $\widetilde{\mathcal{G}}_0(K^*) \supset \mathcal{G}_0(K)^{\perp}$. Suppose that $R \in \mathcal{G}_0(K)^{\perp}$. Then, by Theorem~\ref{thm: minimalize}, we see that $R^* \not\in \widetilde{\mathcal{G}}_0(C)$, and hence $p_{*,0} \colon H_0(C) \to H_0(C/C_{R^*})$ is injective. Let $x \in C_0$ be a homological generator, and then we have $p_{*,0}([x]) \neq 0$. In addition, $\dim_{\mathbb{F}}(C/C_{R^*})_0$ is finite, and hence we can take a finite $\mathbb{F}$-basis for $H_0(C/C_{R^*})$ containing $p_{*,0}([x])$. Thus, by using the identification (given by \Cref{lem: dual epsilon}) $$ \Hom_{\mathbb{F}}(H_0(C/C_{R^*}),\mathbb{F}) \cong H^0(C/C_{R^*} ;\mathbb{F}), $$ we can take a cocycle $\psi \in \Hom_{\mathbb{F}}((C/C_{R^*})_0,\mathbb{F})$ whose cohomology class is the dual $(p_{*,0}([x]))^*$. Moreover, the map $\varepsilon_0$ in \Cref{lem: dual epsilon} is bijective, and hence we can take the inverse $\varphi := \varepsilon^{-1}_0(\psi \circ p) \in C^*_0$. Note that since $\varepsilon \circ \varphi(x) = \psi(p(x))=1$, the element $\varphi \in C^*_0$ is a homological generator. Moreover, the equalities $$ \varepsilon \circ \varphi (C_{R^*}) = \psi \circ p(C_{R^*}) = 0 $$ hold, and hence $\varphi$ lies in $C^*_R$. This proves $R \in \widetilde{\mathcal{G}}_0(K^*)$. \end{proof} \subsection{Computing $\mathcal{G}_0(K^*)$ from $\mathcal{G}_0(K)$} \label{subsec: computing G_0(K^*) from G_0(K)} For any non-empty finite set $\mathcal{S}=\{R_i\}_{i=1}^N$ of semi-simple regions, let \[ S^{\perp} := \{ R \in \mathcal{CR}(\mathbb{Z}^2) \mid \forall R_i \in \mathcal{S},\ R \cap (-R_i) \neq \varnothing\}. \] Then we have the following duality theorem between $\mathcal{S}$ and $\min \mathcal{S}^{\perp}$. \begin{theorem} \label{char orthonormal} The equality $$ \min(\mathcal{S}^{\perp}) = \min \left\{cl(\{-p_i\}_{i=1}^N) \ \middle| \ (p_1, \ldots, p_N) \in \prod_{i=1}^N c(R_i) \right\} $$ holds. \end{theorem} \begin{remark} It can happen that $p_i \leq p_j$ for different $i$ and $j$, and hence the order of $\{p_i\}_{i=1}^N$ can be less than $N$. \end{remark} Recall that $\mathcal{G}_0(K) \subset \mathcal{CR}^{ss}(\mathbb{Z}^2)$ for any knot $K$, and hence, combining with \Cref{thm: duality}, we have the following formula. \begin{corollary} \label{char G_0 of mirror} Let $\mathcal{G}_0(K) = \{R_i\}_{i=1}^N$. Then, the equality $$ \mathcal{G}_0(K^*) = \min \left\{cl(\{-p_i\}_{i=1}^N) \ \middle| \ (p_1, \ldots, p_N) \in \prod_{i=1}^N c(R_i) \right\} $$ holds. In particular, $\mathcal{G}_0(K^*)$ is algorithmically determined from $\mathcal{G}_0(K)$. \end{corollary} Now we prove \Cref{char orthonormal}. \defProof{Proof of Theorem~\ref{char orthonormal}} \begin{proof} Set $\mathcal{S}' := \{cl(\{-p_i\}_{i=1}^N) \mid (p_1, \ldots, p_N) \in \prod_{i=1}^N c(R_i) \}$. We first note that for any $(p_1, \ldots, p_N) \in \prod_{i=1}^N c(R_i)$, we have \[ -p_i \in cl(\{-p_i\}_{i=1}^N) \cap (-R_i) \neq \varnothing. \] This implies $\mathcal{S}' \subset \mathcal{S}^{\perp}$. Now we prove $\min(\mathcal{S}^{\perp}) \subset \min \mathcal{S}'$. Let $R \in \min(\mathcal{S}^{\perp})$. Then $R \cap (-R_i) \neq \varnothing$ for any $1 \leq i \leq N$, which is equivalent to $R_i \not\subset R^*$. Here, \Cref{lem: char ss} gives $R_i = cl(c(R_i))$, and hence there exists an element $p_i \in c(R_i)$ such that $p_i \not\in R^*$. In particular, we have $-p_i \in R$, and hence $R \supset cl(\{-p_i\}_{i=1}^N)$. Since $cl(\{-p_i\}_{i=1}^N) \in \mathcal{S}' \subset \mathcal{S}^{\perp}$, it follows from the minimality of $R$ in $\mathcal{S}^{\perp}$ that $$ R= cl(\{-p_i\}_{i=1}^N) \in \mathcal{S}'.$$ The minimality of $R$ in $\mathcal{S}'$ directly follows from the minimality in $\mathcal{S}^{\perp}$. Therefore, we have $R \in \min S'$. Next, we prove $\min(\mathcal{S}^{\perp}) \supset \min \mathcal{S}'$. Let $cl(\{-p_i\}_{i=1}^N) \in \min \mathcal{S}'$. Then we have $cl(\{-p_i\}_{i=1}^N) \in \mathcal{S}^{\perp}$. Suppose that $R \in \mathcal{S}^{\perp}$ satisfies $R \subset cl(\{-p_i\}_{i=1}^N)$. Then, in a similar way to the converse case, we can find an element $cl(\{-q_i\}_{i=1}^N) \in \mathcal{S}'$ such that $cl(\{-q_i\}_{i=1}^N) \subset R$. In particular, we have $$ cl(\{-q_i\}_{i=1}^N) \subset R \subset cl(\{-p_i\}_{i=1}^N), $$ and hence the minimality of $cl(\{-p_i\}_{i=1}^N)$ in $\mathcal{S}'$ gives $$ cl(\{-q_i\}_{i=1}^N) = R = cl(\{-p_i\}_{i=1}^N). $$ This implies $cl(\{-p_i\}_{i=1}^N) \in \min (\mathcal{S}^{\perp})$. \end{proof} \defProof{Proof} \section{On sparse linear systems} \label{subsec:sparse-linear-system} Here we briefly discuss computational methods for handling large sparse linear system over any computational field, such as $\mathbb{Q}$ or $\mathbb{F}_p$. There are essentially two families of algorithms for handling linear systems: direct methods and iterative methods. Given a matrix $A$, direct methods (such as Gaussian elimination and LU factorization) perform elementary operations on $A$ so that information can be read directly from the modified matrix $A'$, whereas iterative methods (such as the Wiedemann algorithm) multiply $A$ repeatedly on a randomly generated vector seeking for a solution. The implementation of our program is based on the former. \textit{LU factorization} is one of the standard methods for handling (dense) linear systems. Suppose $A$ is an arbitrary $n$-by-$m$ matrix of rank $r$. It is known that, after some permutations of rows and columns, $A$ can be factored into a product of two matrices $L$ and $U$, where $L$ is a lower triangular $n$-by-$r$ matrix and $U$ is an upper triangular $r$-by-$m$ matrix. To be precise, there exists some permutation matrices $P$ and $Q$ such that \[ P A Q = L U \] holds (in some literature it is expressed in the form $A = PLUQ$ and called the $PLUQ$ factorization of $A$). This can be seen as factoring a linear map represented by $A$ into a composition of a surjection (represented by $U$) and an injection (represented by $L$). Having computed $PAQ = LU$, solving a linear system $Ax = b$ breaks up into solving $Ly = Pb$ and $U(Q^{-1}x) = y$. The former equation may or may not have a solution, and in case it does the solution is unique. The latter equation always has a solution with degree of freedom $m - r$. Both equations can be solved easily from the structure of $L$ and $U$. In the following, for simplicity, we omit the permutations $P, Q$ and write $A \sim LU$. LU factorization is obtained from Gaussian elimination. Suppose $A = (a_{ij})_{1 \leq i \leq n,\ 1 \leq j \leq m}$ is non-zero. Then there exists some non-zero component of $A$, which we may assume that it is $a_{11}$. We may write \[ A \sim \begin{pmatrix} 1 \\ a_{21} / a_{11} \\ \vdots \\ a_{n1} / a_{11} \\ \end{pmatrix} \begin{pmatrix} a_{11} & a_{12} & \cdots & a_{1m} \end{pmatrix} + \begin{pmatrix} 0 & 0 & \cdots & 0 \\ 0 & \\ \vdots & & \mbox{\large $A'$}\\ 0 & \end{pmatrix}. \] If $A' = (a'_{ij})_{2 \leq i \leq n,\ 2 \leq j \leq m}$ is non-zero, then we repeat the same process and obtain \[ A \sim \begin{pmatrix} 1 & 0 \\ a_{21} / a_{11} & 1\\ \vdots & \vdots \\ a_{n1} / a_{11} & a'_{n2} / a'_{22}\\ \end{pmatrix} \begin{pmatrix} a_{11} & a_{12} & \cdots & a_{1m} \\ 0 & a'_{22} & \cdots & a'_{2m} \end{pmatrix} + \begin{pmatrix} 0 & 0 & \cdots & 0 \\ 0 & 0 & \cdots & 0 \\ \vdots & \vdots & \mbox{\large $A''$}\\ 0 & 0 \end{pmatrix} \] Repeating this process until the remaining block becomes zero yields the desired factorization. Note that at each step, permutations of the complementary submatrix preserves the structure of the partially computed $L$ and $U$. This process can also be seen as iteratively computing the \textit{Schur complement} of the upper left non-zero component. In general, for a block matrix \[ X = \begin{pmatrix} A & B \\ C & D \end{pmatrix} \] with $A$ non-singular, the \textit{Schur complement} of $A$ is defined by \[ S := D - C A^{-1} B. \] Now we assume that the matrix $A$ is large and sparse, say $n, m > 500,000$ and its density is below $0.01\%$. In such case it is impractical to calculate its LU factorization directly. For this problem, Bouillaguet, Delaplace and Voge \cite{bouillaguet2017parallel} propose a method of finding a large amount of \textit{structural pivots}, and the algorithm is implemented in their program \texttt{SpaSM} \cite{spasm}. Structural pivots are non-zero entries of $A$ such that after some permutations $A$ is transformed into a matrix of the form \[ \begin{pmatrix} U & B \\ C & D \end{pmatrix} \] where $U$ is a regular upper triangular matrix with pivots lying in its diagonal. Given a set of structural pivots, we immediately obtain a partial LU factorization of $A$ as \[ A \sim \begin{pmatrix} I \\ C' \end{pmatrix} \begin{pmatrix} U & B \end{pmatrix} + \begin{pmatrix} O & O \\ O & S \end{pmatrix} \] where $C' = CU^{-1}$, and $S = D - C U^{-1} B$ is the Schur complement of $U$. To complete the LU factorization of $A$, we either iterate the same process on $S$, or switch to dense LU factorization, depending on the size and density of $S$. \begin{figure}[t] \centering \input{tikz/tikz_incremental_LU.tex} \caption{Incremental LU factorization} \label{fig:incremental-LU} \end{figure}{} The heavy part of the above described method is in the computation of the Schur complement. For our purpose, recall from \Cref{subsec:realizability-by-linear-system} that we only need to tell whether the system has a solution or not. Assuming that we have obtained pivots whose number is close to the rank of $A$, only a few more independent columns might suffice to find one solution. Likewise, only a few more independent rows might suffice to detect that the system has no solution. Thus the approach we take is the following: after we have obtained the pivots and permuted the matrix correspondingly, we divide the remaining columns (resp. rows) into small chunks and incrementally expand the factorization by performing the factorization on the submatrices (see \Cref{fig:incremental-LU}). At each step we check whether the subsystem has a solution or not. This enables us to exit the computation in an earlier stage whenever possible. \section{Glossary of notation} \label{sec: glossary} We give a quick glossary of notation by section number: \noindent \begin{multicols}{2} \begin{list}{}{} \item[$C_n$] the degree $n$ complex of $C$, \ref{poset filtered chain complexes}. \item[$C_R$] the subcomplex of $C$ associated to $R$, \ref{poset filtered chain complexes}. \item[$C^r$] the reflection of $C$, \ref{poset filtered chain complexes}. \item[$\mathcal{C}$] the knot concordance group, ***. \item[$\mathcal{C}_{\nu^+}$] the set of $\nu^+$-equivalence classes of knots, \ref{np equivalence}. \item[$CFK^{\infty}$] Heegaard Floer knot complex over $\mathbb{F}[U^{\pm 1}]$, \ref{subsubsec: CFK}. \item[$CFK^-$] Heegaard Floer knot complex over $\mathbb{F}[U]$ \item[$cl(S)$] the closure of $S$, \ref{subsec: computing G_0(K^*) from G_0(K)}. \item[$\mathcal{CR}(P)$] the set of closed regions of $P$, \ref{poset filtered chain complexes}. \item[$\mathcal{CR}^{ss}(\mathbb{Z}^2)$] the set of semi-simple regions of $\mathbb{Z}^2$, \ref{subsec: computing G_0(K^*) from G_0(K)}. \item[$d$] Ozsv\'ath-Szab\'o's $d$-invariant \item[$\partial$] boundary map, \ref{poset filtered chain complexes}. \item[$\mathbb{F}$] the field of order $2$, \ref{poset filtered chain complexes}. \item[$\mathcal{F}_i$] $\mathbb{Z}$-filtration, \ref{poset filtered chain complexes}. \item[$\mathcal{G}_0$] the invariant $\mathcal{G}_0$, \ref{the invariant G_0}. \item[$\widetilde{\mathcal{G}}_0$] the invariant $\widetilde{\mathcal{G}}_0$, \ref{the invariant G_0}. \item[$\mathcal{G}_0^{(n)}$] the invariant $\mathcal{G}_0^{(n)}$, \ref{generalization of G_0}. \item[$\widetilde{\mathcal{G}}^{(n)}_0$] the invariant $\widetilde{\mathcal{G}}_0^{(n)}$, \ref{generalization of G_0}. \item[$\mathrm{gen}_0(C;R)$] the set of realizers of $R$ in $C$, \ref{tG_0 and G_0}. \item[$\gr$] the homological grading, ***. \item[$H_*(C)$] the homology group of $C$, \ref{poset filtered chain complexes}. \item[$H_n(C)$] the degree $n$ homology group of $C$, \ref{poset filtered chain complexes}. \item[$K$] a knot \item[$-K$] the inverse of $K$ \item[$K^*$] the mirror image of $K$ \item[$\Lambda$] the ring $\mathbb{F}[U^{\pm 1}]$, \ref{subsubsec: CFK}. \item[$\nu^+$] Hom-Wu's $\nu^+$-invariant \item[$\overset{\np}{\sim}$] $\nu^+$-equivalence, \ref{np equivalence}. \item[$P$] a poset, \ref{poset filtered chain complexes}. \item[$\mathcal{P}(\mathcal{CR}(\mathbb{Z}^2))$] the power set of $\mathcal{CR}(\mathbb{Z}^2)$, \ref{the invariant G_0}. \item[$\heq{P}{\mathscr{R}}$] $P$-filtered homotopy equivalence over $\mathscr{R}$, \ref{poset filtered chain complexes}. \item[$R$] a closed region, \ref{poset filtered chain complexes}. \item[$S$] a subset of $\mathbb{Z}^2$, \ref{subsec: computing G_0(K^*) from G_0(K)}. \item[$\mathcal{S}$] a subset of $\mathcal{CR}(\mathbb{Z}^2)$, \ref{the invariant G_0}. \item[$S^3_{p/q}(K)$] the $p/q$-surgery along $K$ \item[$\tau$] Ozsv\'ath-Szab\'o's $\tau$-invariant \item[$\Upsilon$] Ozsv\'ath-Stipsicz-Szab\'o's $\Upsilon$-invariant \item[$V_k$] Ni-Wu's $V_k$ sequence \end{list} \end{multicols}

https://arxiv.org/abs/2209.02814

A Subexponential Quantum Algorithm for the Semidirect Discrete Logarithm Problem

Group-based cryptography is a relatively unexplored family in post-quantum cryptography, and the so-called Semidirect Discrete Logarithm Problem (SDLP) is one of its most central problems. However, the complexity of SDLP and its relationship to more well-known hardness problems, particularly with respect to its security against quantum adversaries, has not been well understood and was a significant open problem for researchers in this area. In this paper we give the first dedicated security analysis of SDLP. In particular, we provide a connection between SDLP and group actions, a context in which quantum subexponential algorithms are known to apply. We are therefore able to construct a subexponential quantum algorithm for solving SDLP, thereby classifying the complexity of SDLP and its relation to known computational problems.

\section{Introduction} The goal of Post-Quantum Cryptography ({\tt PQC}{}) is to design cryptographic cryptosystems which are secure against classical and quantum adversaries. A topic of fundamental research for decades, the status of {\tt PQC}{} drastically changed with the {\tt NIST}{} standardisation process \cite{NISTPQ}.\smallskip In July $2022$, after five years and three rounds of selection, {\tt NIST}{} selected a first set of {\tt PQC}{} standards for Key-Encapsulation Mechanism ({\tt KEM}) and Digital Signature Scheme {\tt DSS}{} based on lattices and hash functions. The standardization process is still ongoing with a fourth round for {\tt KEM}{} and a new {\tt NIST}{} call for post-quantum {\tt DSS}{} in $2023$. Recent attacks \cite{DBLP:journals/iacr/Beullens22,DBLP:conf/crypto/TaoPD21,DBLP:journals/iacr/BaenaBCPSV21} against round-$3$ multivariate signature schemes, {\tt Rainbow}{} \cite{DBLP:journals/iacr/Beullens22} and {\tt G{\it e}MSS}{} \cite{casanova2017gemss}, as well as the cryptanalysis of round-$4$ isogeny based {\tt KEM}{} {\tt SIKE}{} \cite{cryptoeprint:2022/975, cryptoeprint:2022/1026}, emphasise the need to continue the cryptanalysis effort in {\tt PQC}{} as well as the diversity in the potential post-quantum hard problems.\smallskip A relatively young family of such problems come from \textit{group-based} cryptography, see \cite{kahrobaeinotices}. The protocol of this type we are interested in is called Semidirect Product Key Exchange ({\tt SPDKE}{}), and was proposed by Habeeb et. al in 2013 \cite{habeeb2013public}. It is a Diffie-Hellman-like, non-interactive key exchange protocol, which uses the group-theoretic notion of the semidirect product. Roughly speaking\footnote{For more detail see Section~\ref{prelimsSDPKE}}, we are interested in products of the form $\phi^{x-1}(g)\cdot\ldots\cdot\phi(g)\cdot g$, where $g$ is an element of a (semi)group, $\phi$ is an endomorphism and $x\in\mathbb{N}$ is a positive integer.\smallskip Examples of concrete proposals for {\tt SPDKE}{} can be found in \cite{habeeb2013public, grigoriev2019tropical, kahrobaei2016using, rahman2022make, rahmanmobs}; respective cryptanalysis can be found in \cite{myasnikov2015linear, isaac2021closer, brown2021cryptanalysis, monico2021remarks, Battarbee_Kahrobaei_Shahandashti_2022, battarbee2021efficiency}. The authors have a survey paper giving a summary of the back-and-forth on this topic \cite{battarbee2022semidirect}. Analogously to the Discrete Logarithm Problem, the security of {\tt SPDKE}{} is heavily related to the following task: given a (semi)group element $g$, an endomorphism $\phi$, and for some $x\in\mathbb{N}$ a product of the form $\phi^{x-1}(g)\cdot\ldots\cdot\phi(g)\cdot g$, recover the integer $x$. We refer to this task as the \textit{Semidirect Discrete Logarithm Problem} ({\tt SDLP}{}); in particular, successfully carrying it out allows recovery of the private exponents of participants in the key exchange.\smallskip The extant body of cryptanalytic work in this area does not address {\tt SDLP}{}, instead using linear algebraic techniques to show that public information leads to key leakage. In fact, the complexity of {\tt SDLP}{} and its relationship to more well-known hardness problems has not been well understood and was a significant open problem for researchers in this area. Our work in this paper takes the first steps towards addressing this problem. In particular, we show that there is a reduction of {\tt SDLP}{} to the so-called \textit{Abelian Hidden Shift Problem}, for which quantum subexponential algorithms are available. \subsection{Organisation of the Paper and Main Results} In line with \cite{childs2014constructing}, the bulk of the work will be the construction of a free, transitive group action, from which the main results quickly follow. To aid with this construction it will be useful to change our perspective slightly - towards this goal we introduce some notation, and recall the appropriate background mathematics.\smallskip In Section~\ref{prelims} we start by giving the definitions that allow the paper to be self-contained. We give a more detailed review of the mechanics of {\tt SPDKE}{}, and introduce some new notation, and finish with a brief discussion of the Abelian Hidden Shift Problem.\smallskip The construction of the group action is contained entirely in Section~\ref{group-action-section}. Several technical lemmas are proved and assembled as Theorem~\ref{free-transitive-action}, in which the required group action is constructed.\smallskip In Section~\ref{headline-section}, we provide a preliminary reduction to the Abelian Hidden Shift Problem in Corollary~\ref{hidden-shift-reduction}. Before providing the full algorithm, we address some context-specific technicality as Procedure~\ref{tail-procedure}. The algorithm claimed in the title is finally given as Theorem~\ref{solving-sdlp}.\smallskip We conclude the paper with a Conclusions section, which will reflect on the manner in which the main results were proved - in particular noting a surprising connection to isogeny-based cryptography. \begin{Rmk*} Shortly before posting this manuscript the authors were made aware of the work in \cite{cryptoeprint:2022/1135}. The landmark result therein suggests that our {\tt SDLP}{} is equivalent to the natural contextual analogue of the Computational Diffie-Hellman problem, which significantly strengthens the claim of security of {\tt SPDKE}{}. We leave the full analysis to subsequent work. \end{Rmk*} \section{Preliminaries}\label{prelims} \subsection{Background Mathematics} We recall a number of group-theoretic notions used throughout this paper. A \textit{(semi)group} is a set $G$ together with a binary operation $G\times G\to G$. Unless otherwise specified, we will write this operation multiplicatively, and require the following: \begin{enumerate} \item $(x\cdot y)\cdot z = x\cdot(y\cdot z)$ for all $x,y,z\in G$ \item There is an identity element, written as $1$, such that $g\cdot 1=1\cdot g = g$ for all $g\in G$ \end{enumerate} Stopping here, we have a semigroup\footnote{More precisely, a monoid.}. Suppose for some $g\in G$ there exists an element $h\in G$ such that $g\cdot h=h\cdot g=1$; such an element is called the \textit{inverse} of $G$, and is necessarily unique. If every $g\in G$ has an inverse, the semigroup is instead a group. The constructions that follow are defined for both groups and semigroups - to reflect this scenario we will use the catch-all notation (semi)groups. For our purposes the (semi)groups will be non-abelian; that is, one cannot expect that $g\cdot h=h\cdot g$. One exception is the group we will write additively; in this case, the operation $g+h$ commutes, and we require an inverse for every element. Note also that the identity is written as $0$ in this case. It will be useful for us to build a new (semi)group from an existing (semi)group. One way of doing this is via a structure called the \textit{holomorph}\footnote{The holomorph is itself a special case of the notion of a semidirect product group, hence the terminology.}; let $G$ a (semi)group and $End(G)$ its endomorphism group. The holomorph is the set $G\times End(G)$ equipped with multiplication \[(g,\phi)\cdot (g',\phi') = (\phi'(g)\cdot g',\phi\circ\phi')\] where $\circ$ refers to function composition. Finally, we recall the notion of a \textit{group action}. Let $G$ a group and $X$ a set: a group action is a function $\psi:G\times X\to X$. By convention we write $\psi(g,x)$ as $g\ast x$, and require the following: $1\times x=x$ for all $x\in X$, and $g\ast(h\ast x)=(g\cdot h)\ast x$. The action is \textit{free} if $g\ast x = x$ forces $g=1$, and \textit{transitive} if for any pair $x,y\in X$ there is a group element $g\in G$ such that $g\ast x=y$. \subsection{Semidirect Product Key Exchange}\label{prelimsSDPKE} We here define in full {\tt SPDKE}{}. One verifies by induction that holomorph exponentiation takes the form \[(g,\phi)^x=(\phi^{x-1}(g)\cdot\ldots\phi(g)\cdot g, \phi^x)\] The central idea of {\tt SPDKE}{} is to use products of these form as a generalisation of Diffie-Helman Key-Exchange. It works as follows: \begin{enumerate} \item Suppose Alice and Bob agree on a public (semi)group $G$, as well as a group element $g$ and endomorphism of $G$, say $\phi$. \item Alice picks a random secret integer $x$, and calculates the holomorph exponent $(g,\phi)^x=(A,\phi^x)$. She sends \textbf{only} $A$ to Bob. \item Bob similarly calculates $(B,\phi^y)$ corresponding to random, private integer $y$, and sends only $B$ to Alice. \item With her private automorphism $\phi^x$ Alice can now calculate her key as the group element $K_A=\phi^x(B)\cdot A$; Bob similarly calculates his key $K_B=\phi^y(A)\cdot B$. \end{enumerate} We have \begin{align*} \phi^x(B)\cdot A &= \phi^x(\phi^{y-1}(g)\cdot\ldots\cdot g)\cdot (\phi^{x-1}(g)\cdot\ldots\cdot g) \\ &= (\phi^{x+y-1}\cdot\ldots\cdot \phi^x(g)) \cdot (\phi^{x-1}(g)\cdot\ldots\cdot g) \\ &= (\phi^{x+y-1}\cdot\ldots\cdot\phi^y(g))\cdot(\phi^{y-1}(g)\cdot\ldots\cdot g) \\ &= \phi^y(A)\cdot B \end{align*} so $K:=K_A=K_B$. Note that $A\cdot B\neq K$ as a consequence of our insistence that the group operation is non-commutative. It is immediate that the security of this scheme is related to {\tt SDLP}{}, since $A=\phi^{x-1}(g)\cdot\ldots\cdot g$, and $A,g,\phi$ are all available to an eavesdropper - recovering Alice's private exponent is therefore exactly {\tt SDLP}{}. Writing these products in full will quickly become rather cumbersome. We therefore introduce some non-standard notation, which is useful both for convenience of exposition and the required shift in perspective we will introduce in this paper. \begin{defn}\label{semidirect-product-function} Let $G$ be a finite, non-commutative (semi)group, $g\in G$, and $\phi\in End(G)$. We define the following function: \begin{align*} s:G\times End(G)\times \mathbb{N} &\to G \\ (g,\phi,x) &\mapsto \phi^{x-1}(g)\cdot\ldots\cdot \phi(g)\cdot g \end{align*} \end{defn} Notice also that when $g,\phi$ are fixed - as in the case of the key exchange - the function $s$ is really only taking integer arguments, continuing our analogy with the standard notion of group exponentiation. Indeed, in this language, {\tt SDLP}{} has the following form: \begin{defn}[Semidirect Discrete Logarithm Problem]\label{sdlp-defn} Let $G$ be a finite, non-commutative (semi)group, $g\in G$, and $\phi\in End(G)$. The \textit{Semidirect Discrete Logarithm Problem} is the following task: given $g,\phi$ and $s(g,\phi,x)$ for some $x\in\mathbb{N}$, recover the integer $x$. \end{defn} \subsection{Abelian Hidden Shift Problem}\label{prelims-HShP} Once we have constructed the requisite group action there is a canonical reduction to a known hardness problem, which will give us our quantum subexponential algorithms. The known hardness problem is as follows: \begin{defn}[Abelian Hidden Shift Problem]\label{hShp-defn} Let $A$ be a finite abelian group and $X$ a set. Suppose there are two injective functions $f,g:A\to S$ that differ by a shift: that is, there is a group element $s\in A$ such that $g(a)=f(a+s)$ for each $a\in A$. In this scenario we say that $f,g$ \textit{hide} $s$. Provided with $f,g$, the \textit{Abelian Hidden Shift Problem} is to recover the shift value $s$. \end{defn} We will make use of the following fact, due to Kuperberg \cite[Proposition~6.1]{kuperberg2005subexponential}: \begin{thm}\label{hShp-complexity} Let $|A|=N$ in the above setup. There exists a quantum algorithm of time complexity $2^{\mathcal{O}(\sqrt{\log N})}$ for solving the Abelian Hidden Shift Problem. \end{thm} \section{Structure of the Exponents}\label{group-action-section} In order to define our action we must first examine the structure of the set $X=\{s(g,\phi,i):i\in\mathbb{N}\}$. Certainly this is neither a group nor a semigroup - numerous counterexamples can be found whereby multiplication of elements in this set are not contained in the set - but we can make some progress by borrowing from the standard theory of monogenic semigroups; presented, for example, in \cite{howie1995fundamentals}. Since $X\subset G$, $X$ is finite - the set $\{x\in\mathbb{N}:\exists y\quad s(g,\phi,x)=s(g,\phi,y)\}$ must therefore be non-empty, else it is in bijection with the natural numbers. We may therefore choose its smallest element, say $n$. By definition of $n$ the set $\{x\in\mathbb{N}:s(g,\phi,n)=s(g,\phi,n+x)\}$ must also be non-empty, so we may again pick its smallest element and call it $r$. The structure of $X$ is further restricted by the following result: \begin{lem}\label{splitting-property} Let $x,y\in\mathbb{N}$, then \[\phi^x\left(s(g,\phi,y)\right)\cdot s(g,\phi,x)=s(g,\phi,x+y)\] \end{lem} \begin{proof} Note that $s(g,\phi, x+y)=\phi^{x+y-1}(g)\cdot\ldots\cdot g$. Since $\phi$ preserves multiplication, applying $\phi^x$ to $s(g,\phi,y)$ adds $x$ to the exponent of each term. Multiplication on the right by $s(g,\phi,x)$ then completes the remaining terms of $s(g,\phi,x+y)$. \end{proof} \begin{Rmk*} One can entirely symmetrically swap the roles of $x$ and $y$ in the above argument, which gives two ways of calculating $s(g,\phi,x+y)$. In essence, therefore, this result gives us a slightly more elegant proof of the correctness of {\tt SPDKE}{}. \end{Rmk*} As a consequence of Lemma~\ref{splitting-property} and the definitions of $n,r$ we have \begin{align*} s(g,\phi,n+2r) &= \phi^r(s(g,\phi,n+r))\cdot s(g,\phi,r) \\ &= \phi^r(s(g,\phi,n))\cdot s(g,\phi,r) \\ &= s(g,\phi,n+r) = s(g,\phi,n) \end{align*} We conclude, by extending this argument in the obvious way, that $s(g,\phi,n+qr)=s(g,\phi,n)$ for each $q\in\mathbb{N}$. In fact, we have the following: \begin{lem}\label{r-identity} Fix $g,\phi$ and define $n,r$ as above. One has that \[s(g,\phi,n+x+qr)=s(g,\phi,n+x)\] for all $x,q\in\mathbb{N}$. \end{lem} We will frequently invoke Lemma~\ref{r-identity}. Indeed, we immediately get that the set $X$ cannot contain values other than $\{g,...,s(g,\phi,n),...,s(g,\phi,n+r-1)\}$. If any of the values in $\{g,...,s(g,\phi,n-1)$ are equal we contradict the minimaltiy of $n$, and if any of the values in $\{s(g,\phi,n),...,s(g,\phi,n+r-1)\}$ are equal we contradict the minimality of $r$. We have shown the following: \begin{thm}\label{exponent-structure} Fix $g\in G$ and $\phi\in End(G)$. The set $X=\{s(g,\phi,i):i\in\mathbb{N}\}$ has size $n+r-1$ for integers $n,r$ dependent on $g,\phi$. In particular \[X=\{g,...,s(g,\phi,n),...,s(g,\phi,n+r-1)\}.\] \end{thm} We refer to the set $\{g,...,s(g,\phi,n-1)\}$ as the \textit{tail} of $X$, and the set $\mathcal{C}=\{s(g,\phi,n),...,s(g,\phi,n+r-1)\}$ as the \textit{cycle} of $X$. The values $n$ and $r$ are called the \textit{index} and \textit{period} of $X$. One can see that unique natural numbers correspond to each element in the tail, but infinitely many correspond to each element in the cycle. In fact, each element of the cycle corresponds to a unique residue class modulo $r$, shifted by the index $n$. This is a rather intuitive fact, but owing to its usefulness we will record it formally. In the following we assume the function $\mod$ returns the canonical positive residue. \begin{thm}\label{exponent-congruence} Let $x,y\in\mathbb{N}$. We have \[s(g,\phi,n+x)=s(g,\phi,n+y)\] if and only if $x\mod r=y\mod r$. \end{thm} \begin{proof} In the reverse direction, setting $x'=x\mod r$ and $y'=y\mod r$, we have by Lemma~\ref{r-identity} that $s(g,\phi,n+x)=s(g,\phi,n+x')$ and $s(g,\phi,n+y)=s(g,\phi,n+y')$. By assumption $x'=y'$, and $0\leq x',y'<r$. The claim follows since we know values in the range $\{s(g,\phi,n),...,s(g,\phi,n+r-1)\}$ are distinct by Theorem~\ref{exponent-structure}. On the other hand, suppose $s(g,\phi,n+y)=s(g,\phi,n+x)$ but $x\not\equiv y\mod r$. Without loss of generality we can write $y=x'+u+qr$ for some $q\in\mathbb{N}, 0<u<r$ and $x'=x\mod r$. By remarks made in the discussion of theorem 2.2, we must have \[s(g,\phi,n+x')=s(g,\phi,n+x'+u).\] There are now three cases to consider; we claim each of them gives a contradiction. First, suppose $x'+u=r$, then $s(g,\phi,n+x')=s(g,\phi,n)$. Since $x'<r$ we contradict minimality of $r$. The case $x'+u<r$ gives a similar contradiction. Finally, if $x'+u>r$, without loss of generality we can write $x'+u=r+v$ for some positive integer $v$, so we have $s(M,\phi,n+x')=s(M,\phi,n+v)$. Since $x'\neq v$ (else we contradict $u<r$), and both values are strictly less than $r$, we have a contradiction, since distinct integers of this form give distinct evaluations of $s$. \end{proof} We are almost ready to define our group action; first, however, we must specify the group acting on the cycle. The previous result implies that we are in some sense interested in the action of residue classes, but in order to use the usual notion of integers $\mod r$, denoted here by $\mathbb{Z}_r$, we would need to be comfortable letting the function $s$ take negative integer inputs. In fact, a well-behaved notion of the output of $s$ on negative integers can be constructed provided one is willing to let $g,\phi$ be invertible. Fortunately, we need not restrict ourselves to this case, and instead consider the following object: \begin{defn} We write $\mathbb{N}_r=\{\modd{i}:0\leq i<r\}$, where $\modd{i}=\{k\in\mathbb{N}:k\equiv i\mod r\}$. We define the operation, written additively, by $\modd{i}+\modd{j}=\modd{i+j}$. \end{defn} It is easy to see that $\mathbb{N}_r$ is a finite abelian group\footnote{For the reader uncomfortable with this unconventional definition, one can think of it as the group $<x:x^r=1>$; in particular, the element $x^{-1}$ is defined only in terms of positive powers of $x$.}. We conclude the section by defining its action on the cycle $\{s(g,\phi,n),...,s(g,\phi,n+r-1)\}$. \begin{thm}\label{free-transitive-action} Let $\mathcal{C}$ be the cycle as described above with size $r$. The additive group $\mathbb{N}_r$ acts freely and transitively on $\mathcal{C}$. \end{thm} \begin{proof} Define the action $\psi:\mathbb{N}_r\times \mathcal{C}\to\mathcal{C}$ by \[\psi(\modd{j},s(g,\phi,n+i))=\phi^{\modd{j}}\left(s(g,\phi,n+i)\right)\cdot s(g,\phi,\modd{j})\] This object is, of course, a set: indeed, it is equal to \[\{\phi^{j+kr}(s(g,\phi,n+i))\cdot s(g,\phi,j+kr):k\in\mathbb{N}_0\}\] which, by Proposition~\ref{splitting-property}, is exactly the set $\{s(g,\phi,n+i+j+kr):k\in\mathbb{N}\}$. By Theorem~\ref{exponent-congruence}, every element of this set is equal to $s(g,\phi,n+i+j)$, so we will harmlessly abuse notation by writing \[\psi(\modd{j},s(g,\phi,n+i))=s(g,\phi,n+i+j)\] Moreover, the output\footnote{Or, more accurately, the content of the singleton set that $\psi$ outputs is an element of $\mathcal{C}$.} of $\psi$ is indeed in $\mathcal{C}$, since $s(g,\phi,n+i+j)=s(g,\phi,n+(i+j)\mod r)$, and $0\leq(i+j)\mod r< r$. In general, when $y>r$ we are free to dispense with the slightly more cumbersome modular notation and write $s(g,\phi,n+y)$ instead of $s(g,\phi,n+(y\mod r))$. Let us check an action is indeed defined. Certainly $\modd{0}$ fixes every element of $\mathcal{C}$. Let $\modd{j},\modd{k}\in\mathbb{N}_r$; then, writing $\psi(\modd{j}, s(g,\phi,n+i))$ as $\modd{j}\ast s(g,\phi,n+i)$ in the conventional fashion, we have \begin{align*} \modd{k}\ast(\modd{j}\ast s(g,\phi,n+i)) &= \modd{k}\ast s(g,\phi,n+i+j) \\ &= s(g,\phi,n+i+j+k) \\ &= \modd{j+k}\ast s(g,\phi, n+i) \\ &= (\modd{k} + \modd{j}) \ast s(g,\phi, n+i) \end{align*} where any protests that the sum exceeds $r$ are countered by the well-definedness of modular addition. Now let us see that the action is free and transitive. If $s(g,\phi,n+i)$ is fixed by $\modd{j}$ then Theorem~\ref{exponent-congruence} gives that $i+j\equiv i\mod r$, so $\modd{j}=\modd{0}$. Thus the action is free. Fix $s(g,\phi,n+i)$ and $s(g,\phi,n+j)$; then $\modd{r-i+j}\in\mathbb{N}_r$ is such that \[\modd{r-i+j}\ast s(g,\phi,n+i)= s(g,\phi,n+i+r-i+j)=s(g,\phi,n+j)\] as required for transitivity. \end{proof} \section{Hidden Shift Problem}\label{headline-section} The connection between the Abelian Hidden Shift Problem problem and group actions has been noticed before, and is given in the context of isogeny-based cryptography in \cite{childs2014constructing}. In what follows, we use our Theorem \ref{free-transitive-action} to update their reduction. \begin{cor}\label{hidden-shift-reduction} Given $s(g,\phi,n+i),s(g,\phi,n+j)$, one can recover the value $\modd{k}$ such that $\modd{k}\ast s(g,\phi,n+i)=s(g,\phi,n+j)$ provided one can solve the Abelian Hidden Shift Problem. \end{cor} \begin{proof} Set $f_A,f_B:\mathbb{N}_r\to\mathcal{C}$ as $f_A(\modd{x})=\modd{x}\ast s(g,\phi,n+i)$ and $f_B(\modd{x})=\modd{x}\ast s(g,\phi,n+j)$. Then \begin{align*} f_B(\modd{x})&=\modd{x}\ast s(g,\phi,n+j) \\ &= \modd{x} \ast (\modd{k}\ast s(g,\phi,n+i)) \\ &= (\modd{x}+\modd{k})\ast s(g,\phi,n+i) \\ &= f_A(\modd{x}+\modd{k}) \end{align*} In other words, $f_A,f_B$ hide $\modd{k}$. To complete the setup of a hidden shift problem we require the functions to be injective, which follows from the action being free and transitive. \end{proof} Consider again the problem of recovering $x$ given $g,\phi,s(g,\phi,x)$. It follows from the above that if one can somehow find an $N$ with $N>n$, all that remains is to solve an instance of the Abelian Hidden Shift Problem. The point is that this latter part of the task is of quantum subexponential complexity, as discussed in Section~\ref{prelims-HShP}. Before we can prove our main theorem, however, we will have to deal with some context-specific technicality. \subsection{Non-empty Tails}\label{non-empty-tails} Notice in the case that the cycle constitutes the whole of the set of exponents (equivalently, when $n=1$) we have described a method of quantum subexponential complexity for recovering $x$ from $g,\phi,s(g,\phi,x)$ for any $x\in\mathbb{N}$, since in this case $g=s(g,\phi,1)$ is in the cycle. In particular, we have described a method of carrying out this task provided one has access to a commutative hidden shift problem oracle, and the best known such oracle is quantum subexponential. In order to deal with the case that $n>1$ and $g$ is in the tail, not the cycle, it is therefore reasonable to assume access to a commutative hidden shift oracle. We detail below how one gets around this technicality below; the procedure is adapted from \cite[Reduction~2.2, pp.~2--3]{banin2016reduction}. \begin{proc}\label{tail-procedure} \normalfont Pick some positive integer $N$. This integer $N$ must be, in a sense detailed below, sufficiently large, but at the outset we have no way of verifying if $N$ has such a property. We therefore need to perform a `sanity check' at the end of the procedure; if this is failed, we know our initial choice of $N$ was not sufficiently large, so we return to the outset replacing $N$ with $2N$. Choose some random $k\in\{\lceil N/2 \rceil,...,N\}$ and calculate $s(g,\phi,k)$. We assume $N$ is sufficiently large to guarantee $s(g,\phi,k)$ is in the cycle; that is, we assume $\lceil N/2 \rceil \geq n$. Choose another random $k'\in\{1,...,N\}$, set $h=s(g,\phi,k)$, and calculate $h'=\phi^{k'}(h)\cdot s(g,\phi,k')$. We now have two cycle elements; we know by our work above that with access to a hidden shift oracle we can recover the positive residue class $\modd{l}$ that acts on $h$ to give $h'$. In particular this gives us $r$, and therefore our sanity check: we should have that $\phi^r(h)\cdot s(g,\phi,r)=h$. If not, $N$ was not sufficiently large, and we return to the outset as described. With access to $r$ it is a simple matter to recover $n$, since by definition of $n,r$, if $s(g,\phi,m+r)\neq s(g,\phi,m)$ then $m<n$. On the other hand, if $s(g,\phi,m+r)=s(g,\phi,m)$ we must have $m\geq n$. We can therefore carry out binary search to find the smallest $m$ such that $s(g,\phi,m)=s(g,\phi,m+r)$; by definition, this smallest $m$ is exactly $n$. \end{proc} Since the requirement of $N$ to be sufficiently large is that $N\geq n$, we expect to have to make $\mathcal{O}(\log n)$ oracle calls. By Theorem~\ref{hShp-complexity}, each oracle call has quantum time complexity $2^{\mathcal{O}(\sqrt{log r})}$. We conclude after appropriate manipulation that the time complexity of finding $r$ is $2^{\mathcal{O}(\log\log n+\sqrt{\log r})}$. The final binary search procedure is again $\mathcal{O}(\log n)$-time, so the overall complexity is $2^{\mathcal{O}(\log\log n+\sqrt{\log r})}$. \subsection{Solving {\tt SDLP}{}} It remains to assemble the procedure given in Section~\ref{non-empty-tails} and Corollary~\ref{hidden-shift-reduction}. Indeed, doing so will gives the claim of the title, which we restate for completeness: \begin{thm}\label{solving-sdlp} There exists an algorithm which solves {\tt SDLP}{} in quantum subexponential time. \end{thm} \begin{proof} First, as described in Procedure~\ref{tail-procedure}, one uses the oracle to recover the index $n$, with complexity $2^{\mathcal{O}(\log\log n+\sqrt{\log r})}$. With knowledge of the index we can use the oracle again to find the positive residue class $\modd{k}$ such that $\modd{k}\ast s(g,\phi,n)=s(g,\phi,x)$. Without loss of generality let $k$ the smallest integer representative of this residue class; it follows that $x=n+k$. Again by Theorem~\ref{hShp-complexity}, the quantum time complexity of recovering this residue class is $2^{\mathcal{O}(\sqrt{log r})}$, from which the claim follows. \end{proof} \section{Conclusion} We have provided the first dedicated analysis of {\tt SDLP}{}, showing a reduction to a well-studied problem. Perhaps the most surprising aspect of the work is the progress made by a simple rephrasing; we made quite significant progress through rather elementary methods, and we suspect much more can be made within this framework. The reader may notice that we have shown that {\tt SPDKE}{} is an example of a commutative action-based key exchange, and that breaking all such protocols can be reduced to the Abelian Hidden Shift Problem. Indeed, this work shows the algebraic machinery of {\tt SPDKE}{} is a candidate for what Couveignes calls a \textit{hard homogenous space}\footnote{Another major example of which arises from the theory of isogenies between elliptic curves - see, for example, \cite{castryck2018csidh}} \cite{couveignes2006hard}, which was not known until now. In line with the naming conventions in this area we propose a renaming of {\tt SPDKE}{} to \spdh{}, which stands for `Semidirect Product Diffie Hellman', and should be pronounced \textit{spud}. We would also like to reiterate the sentiment expressed in the abstract. The purpose of this paper is not to claim a general purpose break of {\tt SPDKE}{} (or, indeed, \spdh{}) - the algorithm presented is subexponential in complexity, which has been treated as tolerable in classical contexts. Instead, the point is to show a connection between {\tt SDLP}{} and a known hardness problem, thereby providing insight on a problem about which little was known. \subsection{Acknowledgements} We thank Chloe Martindale of the University of Bristol and Samuel Jaques of the University of Oxford for fruitful discussion on this topic at the recent ICMS conference \textit{Foundations and Applications of Lattice-based Cryptography}. We furthermore thank the organisers of this conference for facilitating this conversation. \printbibliography \end{document}

https://arxiv.org/abs/2109.14945

Regular dessins with moduli fields of the form $\mathbb{Q}(ζ_p,\sqrt[p]{q})$

Gareth Jones asked during the 2014 SIGMAP conference for examples of regular dessins with nonabelian fields of moduli. In this paper, we first construct dessins whose moduli fields are nonabelian Galois extensions of the form $\mathbb{Q}(\zeta_p,\sqrt[p]{q})$, where $p$ is an odd prime and $\zeta_p$ is a $p$th root of unity and $q\in\mathbb{Q}$ is not a $p$th power, and we then show that their regular closures have the same moduli fields. Finally, in the special case $p=q=3$ we give another example of a regular dessin with moduli field $\mathbb{Q}(\zeta_3,\sqrt[3]{3})$ of degree $2^{19}\cdot3^4$ and genus $14155777$.

\section{Introduction} Grothendieck first coined the term {\em Dessin d'enfant} in {\em Esquisse d'un Programme} \cite{esquisse} to denote a connected bicolored graph embedded on a compact connected oriented topological surface. The study was motivated by the one to one correspondance between dessins d'enfant, the combinatorial data of the associated cartographical group, and the geometric concept of coverings of $\PP$ by compact Riemann surfaces ramified at most over three points. Moreover, by Belyi's theorem any such covering is given the structure of an algebraic curve defined over a number field, therefore we obtain a natural action of the absolute Galois group $\AGG$ on the set of isomorphism classes of dessins. A lot of the interest for dessins stems from the fact that this action is faithful, providing a way to study the absolute Galois group through its action on the set of dessins. A particularly interesting family of dessins is that of regular dessins, characterized by the fact that their automorphism groups act transitively on their sets of edges, and the Galois action was proved to remain faithful when restricted to the subset of isomorphism classes of regular dessins \cite{Gonzalez-Diez-Jaikin-Zapirain}. To any dessin we associate a number field called its moduli field, which is defined as the subfield of $\QQb$ fixed by the subgroup of $\AGG$ that fixes the dessin up to isomorphism. Conder, Jones, Streit and Wolfart noted in \cite{CJSW} that the moduli fields of all the examples of regular dessins known at the time were abelian Galois extensions of $\QQ$. Herradón constructed in \cite{Herradon} an explicit equation for a regular dessin whose moduli field $\QQ(\sqrt[3]{2})$ is not a Galois extension of $\QQ$, and Hidalgo later generalized his construction in \cite{Hidalgo} to produce regular dessins whose moduli fields are of the form $\QQ(\sqrt[p]{2})$ where $p$ is an odd prime number. However there is as of yet no known example of regular dessin whose moduli field is a nonabelian Galois extension of $\QQ$. This is the starting point of this paper, in which we will exhibit examples of regular dessins with moduli fields that are nonabelian Galois extensions of $\QQ$. In the present paper, we begin by recalling the main definitions and results on dessins d'enfant. We will then expose constructions of regular dessins whose moduli fields are nonabelian Galois extensions of $\QQ$. We first exhibit dessins whose moduli fields are of the form $\QQ(\zeta_3,\sqrt[3]{q})$, where $\zeta_3$ is a primitive third root of unity and $q\in\QQ$ is not a third power, and show that the regular closures of these dessins possess the same moduli fields. We then generalize this construction to show that there exist regular dessins with moduli fields $\QQ(\zeta_p,\sqrt[p]{q})$, where $\zeta_p$ is a primitive $p$th root of unity and $q\in\QQ_{>0}$ is not a $p$th power. Finally, we give an example of a regular dessin with moduli field $\QQ(\zeta_3,\sqrt[3]{3})$ of degree 42467328 and genus 14155777. \paragraph*{Aknowledgements} The authors are grateful to Professor Jürgen Wolfart for valuable comments. \bigskip \paragraph*{Notations} \begin{itemize} \item $\Sym_{E}$: the group of self-bijections of the set $E$, similarly $\Sym_n$ is the group of permutations of a set of $n$ elements (we favor a right action, hence we write the product $\sigma\tau\coloneqq\tau\circ\sigma$) \item $\Gal(E/F)$: the Galois group of $F$-automorphisms of $E$ \item $\zeta_k$: the $k$th primitive root of unity $\exp(\frac{2i\pi}{k})$ \item $F_2$: the free group of rank $2$ with generators $(\genx,\geny)$ \item $\Crit$: the set of critical values of a function \end{itemize} \section{Preliminaries on dessins d'enfant} We refer the reader to existing expositions of the theory such as \cite{Guillot}, \cite{Lando-Zvonkin}, \cite{Jones-Wolfart} and \cite{Girondo-Gonzalez-Diez} for proofs of the presented facts and further details. \bigskip A {\em dessin d'enfant} is a connected bipartite graph embedded on a compact connected orientable topological surface, such that the complement of the graph is a disjoint union of 2-cells. Two such dessins are equivalent if there exists an orientation preserving homeomorphism between the underlying surfaces that induces an isomorphism between the embedded bipartite graphs. A dessin is determined up to isomorphism by a pair $(C,\beta)$ where $C$ is a smooth algebraic curve and $\beta\colon C\to\PP$ is a meromorphic mapping ramified at most over $\{0,1,\infty\}$, and by Belyi's theorem we can further ask for $C$ and $\beta$ to both be defined over a number field. We call $(C,\beta)$ a {\em Belyi pair} and $\beta$ a {\em Belyi function}. The corresponding graph embedding on the underlying surface is recovered by pulling back the segment $[0,1]$ along $\beta$, we define black and white vertices as the preimages of $0$ and $1$ respectively, and the edges as the preimages of $]0,1[$. By covering theory a dessin is also determined up to isomorphism by the {\em monodromy action} of the fundamental group of the complex projective line $\pi_1(\PP)$ on the fiber over the point $\frac{1}{2}$ which is identified to the set of edges of the dessin. The fundamental group $\pi_1(\PP)$ is isomorphic to the free group of rank two $F_2=\langle\genx,\geny\rangle$ with generators $\genx$ and $\geny$ which are two loops with base point $\frac{1}{2}$ and circling counter-clockwise around $0$ and $1$ respectively. The monodromy action of the generators $\genx$ and $\geny$ then corresponds to the product of the counter-clockwise cyclic permutation of the edges around black and white vertices respectively. We call {\em monodromy map} $M\colon F_2\to\Sym_E$ the map that associates to each element of $F_2$ the corresponding permutation of the set of edges, and we call {\em cartographic group} the image of the monodromy map, which is a transitive subgroup of the group of permutations of the set of edges. \medskip When the {\em automorphism group} of a dessin $\cD$ acts transitively on the set of edges, we say that $\cD$ is a {\em regular dessin}. When that is the case the cartographic group $G$ acts transitively and freely on the set of edges, the monodromy action is thus given by the canonical action of $G$ on itself. There is a natural bijection between regular dessins and finite groups generated by two distinguished elements $\genx$ and $\geny$ up to isomorphism. Two regular dessins determined by $G_1=\langle\genx_1,\geny_1\rangle$ and $G_2=\langle\genx_2,\geny_2\rangle$ respectively are isomorphic if and only if there exists an isomorphism between $G_1$ and $G_2$ that preserves the distinguished generators. Given a dessin $\cD$, there exists a unique regular dessin $\widetilde{\cD}$ with a morphism $\phi\colon\widetilde{\cD}\to\cD$ such that any morphism from a regular dessin to $\cD$ factors through $\phi$. We call $\widetilde{\cD}$ the {\em regular closure} of $\cD$. Moreover, there exists an isomorphism $\Cart(\widetilde{\cD})\cong\Cart(\cD)$ that preserves the distinguished generators. There exists a natural action of the absolute Galois group $\AGG$ on the set of isomorphism classes of dessins, we denote by $\cD^\sigma$ the action of an automorphism $\sigma$ on a dessin $\cD$, and this Galois action commutes with regular closure, i.e. we have $(\widetilde{\cD})^\sigma\cong\widetilde{(\cD^\sigma)}$. \medskip Given a dessin $\cD$, we say that a number field $k$ is a {\em field of definition} of $\cD$ if $\cD$ is isomorphic to a dessin defined over $k$. However there does not necessarily exist a smallest field of definition. We thus define the {\em moduli field} of a dessin $\cD$ as the subfield of $\QQb$ fixed by the subgroup of $\AGG$ constituted of the elements fixing $\cD$ up to isomorphism. The moduli field of a dessin is contained in all fields of definition but is not necessarily itself a field of definition, however it is the case in particular for regular dessins. \section{Constructions of regular dessins with nonabelian moduli fields} We are now ready to give examples of regular dessins whose moduli fields are nonabelian Galois extensions of $\QQ$. To do so, we will first exhibit dessins with such moduli fields, and then prove that their regular closures admit the same moduli fields. \bigskip Before proceeding with the examples, let us first present a classic family of Belyi polynomials that we will use in the following constructions. For positive integers $m,n\in\NN$ we define the polynomial $$B_{m,n}\coloneqq\frac{(m+n)^{m+n}}{m^mn^n}X^m(1-X)^n\in\QQ[X].$$ By computing the derivative $B_{m,n}'=\frac{(m+n)^{m+n}}{m^mn^n}X^{m-1}(1-X)^{n-1}(m-(m+n)X)$ we verify that $B_{m,n}:\PP\to\PP$ is a Belyi function that ramifies only at $0$, $1$, $\infty$ and $\frac{m}{m+n}$ with ramification indices $m$, $n$, $m+n$ and $2$ respectively, and $B_{m,n}(0)=0$, $B_{m,n}(1)=0$, $B_{m,n}(\infty)=\infty$ and $B_{m,n}(\frac{m}{m+n})=1$ (see Figure \ref{dessin-Bmn}). \begin{figure}[H] \centering \includegraphics[height=2cm]{belyi-polynomial-mn} \caption{Dessin corresponding to the Belyi pair $(\PP,B_{m,n})$.} \label{dessin-Bmn} \end{figure} \subsection{Regular dessins with moduli fields of the form $\QQ(\zeta_3,\sqrt[3]{q})$}\label{ex1} Let $q\in\QQ_{>0}$ be a positive rational number that is not a third power. \medskip Let $m,n\in\NN$ be coprime positive integers such that $\frac{27}{27+q^2}=\frac{m}{m+n}$, and let \begin{align*} C&\colon y^2=x(x-(1-\zeta_3))(x-\sqrt[3]{q}),\\ \beta&\colon C\to\PP,\;(x,y)\mapsto\frac{1}{27^mq^{2n}}(x^6+27)^m(q^2-x^6)^n. \end{align*} The function $\beta$ is given by the composition $\beta=\beta_1\circ\beta_0\circ\pi$ of the following maps. \begin{enumerate} \item $\pi\colon C\to\PP$ is the projection on the coordinate $x$, which is ramified over $\{0,1-\zeta_3,\sqrt[3]{q},\infty\}$. \item $\beta_0\coloneqq X^6\in\QQ[X]$, $\Crit(\beta_0)=\{0\}$ so $\beta_1\circ\pi$ ramifies over $\{0,(1-\zeta_3)^6=-27,q^2,\infty\}$. \item $\beta_1\coloneqq B_{m,n}(\frac{X+27}{q^2+27})$, so $\beta=\beta_1\circ\beta_0\circ\pi$ ramifies over $\{0,1,\infty\}$. \end{enumerate} The pair $(C,\beta)$ is thus a Belyi pair, and we call $\cD$ the corresponding dessin. The dessin $\cD$ is defined over $\QQ(\zeta_3,\sqrt[3]{q})$, so its moduli field is a subfield of $\QQ(\zeta_3,\sqrt[3]{q})$. By taking the regular closure we then obtain the inclusion of moduli fields $\cM(\widetilde{D})\subseteq\cM(D)\subseteq\QQ(\zeta_3,\sqrt[3]{q})$, and moreover $\widetilde{D}$ is regular so it is defined over $\cM(\widetilde{D})$. We shall prove that $\cM(\widetilde{\cD})$ is in fact exactly $\QQ(\zeta_3,\sqrt[3]{q})$, which is a nonabelian Galois extension of $\QQ$ with Galois group $$\Gal(\QQ(\zeta_3,\sqrt[3]{q})/\QQ)\cong\Sym_3.$$ To that end we must show that an automorphism $\sigma\in\AGG$ fixes $\widetilde{\cD}$ if and only if it fixes $\zeta_3$ and $\sqrt[3]{q}$, or equivalently that $\Gal(\QQ(\zeta_3,\sqrt[3]{q})/\QQ)$ acts freely on the orbit of $\widetilde{\cD}$. \medskip Let $\sigma\in\AGG$, the Galois conjugate $\cD^\sigma$ is given by the Belyi pair $(C^\sigma,\beta^\sigma)$, where $$C^\sigma\colon y^2=x(x-(1-\sigma(\zeta_3)))(x-\sigma(\sqrt[3]{q})),$$ and $\beta^\sigma$ has the same expression as $\beta$ because all of its coefficients are rational. The orbit of the pair $(\zeta_3,\sqrt[3]{q})$ by $\AGG$ is $\{\zeta_3^i,\zeta_3^j\sqrt[3]{q}\}_{1\leq i\leq2,0\leq j\leq2}$. Elliptic curves given by equations of the form $y^2=(x-a)(x-b)(x-c)$ are isomorphic if and only if the cross-ratios of the tuples $(a,b,c,\infty)$ coincide. We verify that the cross-ratios are all distinct, so the orbit of $\cD$ is given by the six dessins $\cD^\sigma$ for $\sigma\in\Gal(\QQ(\zeta_3,\sqrt[3]{q})/\QQ)$. As a consequence $\cM(\cD)=\QQ(\zeta_3,\sqrt[3]{q})$. To prove that the regular closures $\widetilde{\cD^\sigma}$ constituting the orbit of $\widetilde{\cD}$ are also non isomorphic, we must first draw the dessins $\cD^\sigma$ to compute their cartographic groups. \medskip Let us first draw the dessin $\cD_0$ corresponding to the Belyi pair $(\PP,\beta_1\circ\beta_0)$ (see Figure \ref{beta-ex1}). The dessin $\cD_0$ is defined over $\QQ$, so the dessins $\cD^\sigma$ in the orbit are then obtained by lifting $\cD_0$ to the curves $C^\sigma$. To simplify the graphical representations of the dessins, we will use the notation in Figure \ref{simply} for consectutive edges incident to a vertex. \begin{figure}[H] \centering \includegraphics[height=2.5cm]{simply} \caption{Notation for consecutive edges.} \label{simply} \end{figure} \begin{figure}[H] \centering \includegraphics[width=6cm]{beta-ex1} \caption{Construction of $\cD_0$.} \label{beta-ex1} \end{figure} The dessins $\cD_1,\dots,\cD_6$ conjugate to $\cD$ are embedded on a torus, so in the representations in Figure \ref{ex1-dessins} we will identify the outermost edges on opposite sides. \begin{figure}[H] \begin{subfigure}{0.5\hsize} \begin{center} \includegraphics[height=0.9\hsize]{ex1-1} \end{center} \caption{$\cD_1\coloneqq\cD$} \label{ex1-1} \end{subfigure} \begin{subfigure}{0.5\hsize} \begin{center} \includegraphics[height=0.9\hsize]{ex1-2} \end{center} \caption{$\cD_2\coloneqq\cD^\sigma,\sigma\colon(\zeta_3,\sqrt[3]{q})\mapsto(\zeta_3^2,\zeta_3\sqrt[3]{q})$} \end{subfigure} \end{figure} \begin{figure}[H]\ContinuedFloat \begin{subfigure}{0.5\hsize} \begin{center} \includegraphics[height=0.9\hsize]{ex1-3} \end{center} \caption{$\cD_3\coloneqq\cD^\sigma,\sigma\colon(\zeta_3,\sqrt[3]{q})\mapsto(\zeta_3,\zeta_3\sqrt[3]{q})$} \end{subfigure} \begin{subfigure}{0.5\hsize} \begin{center} \includegraphics[height=0.9\hsize]{ex1-4} \end{center} \caption{$\cD_4\coloneqq\cD^\sigma,\sigma\colon(\zeta_3,\sqrt[3]{q})\mapsto(\zeta_3^2,\zeta_3^2\sqrt[3]{q})$} \end{subfigure} \end{figure} \begin{figure}[H]\ContinuedFloat \begin{subfigure}{0.5\hsize} \begin{center} \includegraphics[height=0.9\hsize]{ex1-5} \end{center} \caption{$\cD_5\coloneqq\cD^\sigma,\sigma\colon(\zeta_3,\sqrt[3]{q})\mapsto(\zeta_3,\zeta_3^2\sqrt[3]{q})$} \end{subfigure} \begin{subfigure}{0.5\hsize} \begin{center} \includegraphics[height=0.9\hsize]{ex1-6} \end{center} \caption{$\cD_6\coloneqq\cD^\sigma,\sigma\colon(\zeta_3,\sqrt[3]{q})\mapsto(\zeta_3^2,\sqrt[3]{q})$} \label{ex1-6} \end{subfigure} \caption{Dessins $\cD_1,\dots,\cD_6$ in the Galois orbit of $\cD$.} \label{ex1-dessins} \end{figure} We will now establish that $\widetilde{\cD_1}$ is not isomorphic to $\widetilde{\cD_2},\dots,\widetilde{\cD_6}$. To that end it suffices to show that there is no isomorphism between the cartographic groups fixing the canonical generators. We shall therefore exhibit an element $\omega\in F_2=\langle \genx,\geny\rangle$ such that $M_k(\omega)$ commutes with $M_k(\geny^2)$ only when $k=1$, where $M_k$ is the monodromy map of $\cD_k$. \medskip We have defined $m$ and $n$ to be positive coprime integers such that $\frac{27}{27+q^2}=\frac{m}{m+n}$, so we cannot have $m=n=1$. We will treat the case where $m\neq1$ does not divide $n$, the other case being treated similarly. Let $$\omega\coloneqq \genx^n\geny^{-1}\genx^{m-n}\geny\genx^n.$$ We shall show that $M_k(\omega)$ commutes with $M_k(\geny^2)$ only when $k=1$. Let $E_k\coloneqq\{1,2,\dots,24\}$ be the set of edges of $\cD_k$ incident to $0$. The action of $\geny$ fixes the set $E_k$ on which it induces the cyclic permutation $(1,2,\dots,24)$, and every white vertex except $0$ has degree one so the action of $\geny$ is trivial on the complement of $E_k$. We can write $E_k=E_k^{\odd}\sqcup E_k^{\even}$ as the disjoint union of the sets of respectively odd and even numbered edges incident to $0$, such that $\geny$ sends one to the other. The black vertices of $E_k^{\odd}$ are of degree $m$ except for the two black vertices of the edges $1$ and $13$ that are of degree $2m$. Therefore if $m$ does not divide some integer $l$ then $\genx^l$ sends every edge of $E_k^{\odd}$ to the complement of $E_k$, and otherwise the action of $\genx^m$ on $E_k^{\odd}$ corresponds to the sole transposition $(1,13)$. Similarly if $n$ does not divide $l$ then $\genx^l$ sends every edge of $E_k^{\even}$ to the complement of $E_k$, and the action of $\genx^n$ on $E_k^{\even}$ is the transposition $(2k,2k+12)$. In particular, by hypothesis $n$ is not a multiple of $m$, so $m-n$ is not a multiple of $m$ either, hence both $\genx^n$ and $\genx^{m-n}$ send the edges of $E_k^{\odd}$ to the complement of $E_k$. However $\geny$ acts trivially on the latter, so $\genx^n\geny^{-1}\genx^{m-n}$ and $\genx^{m-n}\geny\genx^n$ both fix the set $E_k^{\odd}$ on which they induce the same action as $\genx^m$, i.e. the transposition $(1,13)$. Therefore the action of $\omega=\genx^n\geny^{-1}\genx^{m-n}\geny\genx^n$ is the same as that of $\genx^m\geny\genx^n$ on $E_k^{\odd}$ and the same as that of $\genx^n\geny^{-1}\genx^m$ on $E_k^{\even}$. See Figure \ref{ex1-trans}. \begin{figure}[H] \centering \includegraphics[height=3cm]{ex1-trans} \caption{Action of $\omega$ on $E_k^{\odd}\setminus\{1,2k-1\}$ and on $E_k^{\even}\setminus\{2,2k\}$.} \label{ex1-trans} \end{figure} The action of $\omega$ fixes the set $E_k$ on which it induces the permutation $$M_k(\omega)|_{E_k}=(1,13)(2k,2k+12)\cdot(1,2)(3,4)\cdots(23,24)\cdot(1,13)(2k,2k+12)$$ Therefore for $k=1$, \begin{align*} M_1(\omega)|_{E_1}&=(1,13)(2,14)\cdot(1,2)(3,4)\cdots(23,24)\cdot(1,13)(2,14)\\ &=(1,2)(3,4)\cdots(23,24) \end{align*} so $\omega$ and $\geny^2$ commute on $E_1$. Moreover $\geny$ acts trivially on the complement of $E_1$ so $M_1(\omega)|_{\cD_1\setminus E_1}$ and $M_1(\geny^2)|_{\cD_1\setminus E_1}$ automatically commute. Finally, we obtain that $M_1(\omega)$ and $M_1(\geny^2)$ commute. \medskip For $k=2$, we observe that $4^{\omega\geny^2}=15^{\geny^2}=17$ but $4^{\geny^2\omega}=6^\omega=5$. Similarly, for $3\leq k\leq6$, we observe that $1^{\omega\geny^2}=14^{\geny^2}=16$ but $1^{\geny^2\omega}=3^\omega=4$. We have thus shown that $M_k(\omega)$ and $M_k(\geny^2)$ commute only for $k=1$. \medskip This concludes the proof that $\widetilde{\cD}$ is a regular dessin with moduli field $\QQ(\zeta_3,\sqrt[3]{q})$. \subsection{Regular dessins with moduli fields of the form $\QQ(\zeta_p,\sqrt[p]{q})$} Let $p$ be an odd prime, and $q\in\QQ_{>0}$ a positive rational number that is not a $p$th power. In this example we will need an additional parameter $\gamma\in\QQ\setminus\{0\}$. Let \begin{align*} C\colon y^2=x(x-(1-\zeta_p))(x-\gamma\sqrt[p]{q}). \end{align*} We construct the Belyi function $\beta\colon C\to\PP$ as the composition $\beta=\beta_2\circ\beta_1\circ\beta_0\circ\pi$ of the following maps. \begin{enumerate} \item $\pi\colon C\to\PP$ is the projection on the coordinate $x$, which ramifies over $\{0,1-\zeta_p,\gamma\sqrt[p]{q},\infty\}$. \item $\beta_0\coloneqq X^{2p}\in\QQ[X]$, and $\Crit(\beta_0)=\{0,\infty\}$ so $\beta_0\circ\pi$ ramifies over $\{0,(1-\zeta_p)^{2p},\gamma^{2p}q^2,\infty\}$. \item $\beta_1\in\QQ[X]$ is chosen independently of $\gamma$ such that $\Crit(\beta_1)\cup\{\beta_1((1-\zeta_p)^{2p})\}=\{0,1,\infty\}$, $\beta_1((1-\zeta_p)^{2p})=0<\beta_1(0)<1$ and $\beta_1'(0)>0$. The existence of $\beta_1$ verifying those conditions is assured by Proposition \ref{unramified-belyi-function} below. Under those assumptions $\beta_1\circ\beta_0\circ\pi$ ramifies over $\{0,1,\beta_1(0),\beta_1(\gamma^{2p}q^2),\infty\}$. \item $\gamma\in\QQ_{>0}$ is then chosen small enough so that $\beta_1'>0$ on $[0,\gamma^{2p}q^2]$. This guarantees us that we have $0<\beta_1(0)<\beta_1(\gamma^{2p}q^2)<1$. \item $\beta_2\coloneqq B_{r,s}\circ B_{m,n}$, where $(m,n)$ and $(r,s)$ are pairs of coprime positive integers such that $\beta_1(\gamma^{2p}q^2)=\frac{m}{m+n}$ and $B_{m,n}(\beta_1(0))=\frac{r}{r+s}$. Finally, $\beta=\beta_2\circ\beta_1\circ\beta_0\circ\pi$ ramifies over $\{0,1,\infty\}$. \end{enumerate} The pair $(C,\beta)$ is thus a Belyi pair, and we call $\cD$ the corresponding dessin. With the same arguments as before, the moduli field of $\cD$ is $\QQ(\zeta_p,\sqrt[p]{q})$, which is a nonabelian Galois extension of $\QQ$ with Galois group $$\Gal(\QQ(\zeta_p,\sqrt[p]{q})/\QQ)\cong\ZZn{p}\rtimes(\ZZn{p})^\times$$ generated by $\sigma\colon\zeta_p^i\sqrt[p]{q}\mapsto\zeta_p^{i+1}\sqrt[p]{q}$ and $\tau\colon\zeta_p^i\sqrt[p]{q}\mapsto\zeta_p^{gi}\sqrt[p]{q}$ where $g$ generates $(\ZZn{p})^\times$. We shall show that there exists $\gamma\in\QQ\setminus\{0\}$ such that the regular closure of the dessin $\cD$ thus obtained also has moduli field $\QQ(\zeta_p,\sqrt[p]{q})$. \begin{remark} In the previous subsection we treated the case $p=3$. In that specific case we gave a simpler expression for $\beta$, mainly due to the fact that $\beta_0\circ\pi$ already had all of its critical values in $\QQ\cup\{\infty\}$. However in the general case we must use the intermediate map $\beta_1$ as well as the parameter $\gamma$ to conclude the proof. \end{remark} \bigskip Let us first prove the existence of $\beta_1$. \begin{proposition}\label{unramified-belyi-function} Let $E\subset\QQb\cap\RR\setminus\{0\}$ be a finite set. Then there exists $P\in\QQ[X]$ such that $P(E)\subseteq\{0\}$, $\Crit(P)\subseteq\{0,1\}$, $0<P(0)<1$ and $P'(0)>0$. \end{proposition} \begin{remark} In the context of this proposition we only deal with polynomials so for $P\in\QQ[X]$ we define $\Crit(P)\coloneqq\{P(z)|\;z\in\CC,P'(z)=0\}$, which does not include the point at infinity to simplify notations. \end{remark} \begin{proof} To show this we will proceed similarly as in the proof of the {\em only if} part of Belyi's theorem, by applying additional transformations to ensure that $0<P(0)<1$. Let us first prove that we can reduce to the case where $E$ is a subset of rational numbers. \begin{lemma} Let $E\subset\QQb\cap\RR\setminus\{0\}$ be a finite set fixed by $\AGG$. Then there exists $P\in\QQ[X]$ such that $P(0)=0$ and $\Crit(P)\cup P(E)\subset\QQ\setminus\{0\}$. \end{lemma} \begin{proof} Let $\{a_1,\dots,a_m\}=E\cap\QQ$ and $\{b_1,\dots,b_n\}=E\setminus\QQ$. We construct $P$ by induction on the number $n$ of non rational elements of $E$. For $\alpha\in\QQ$, define $F_\alpha,G_\alpha\in\QQ[X]$ by $$F_\alpha\coloneqq\prod_{j=1}^n(X-(b_j-\alpha)^2)\quad\text{and}\quad G_\alpha\coloneqq F_\alpha((X-\alpha)^2)=\prod_{j=1}^n(X-b_j)(X+b_j-2\alpha).$$ Let us first assume that there exists $\alpha\in\QQ$ such that $G_\alpha(0)\not\in\Crit(G_\alpha)\cup G_\alpha(E)$. Define $P_1(X)\coloneqq G_\alpha(X)-G_\alpha(0)\in\QQ[X]$, then $P_1(0)=0\not\in E'\coloneqq\Crit(P_1)\cup P_1(E)\subset\QQb\cap\RR\setminus\{0\}$. Note that $E'$ is stable under the action of $\AGG$, and $|E'\setminus\QQ|=|\Crit(F_\alpha)\cup F_\alpha(0)\setminus\QQ|=|\Crit(F_\alpha)\setminus\QQ|<\deg F_\alpha=n$. By induction, there exists $P_2\in\QQ[X]$ such that $P_2(0)=0$ and $\Crit(P_2)\cup P_2(E')\subset\QQ\setminus\{0\}$. Now $P\coloneqq P_2\circ P_1$ has the desired properties, since $P(0)=0$ and $\Crit(P)\cup P(E)=\Crit(P_2)\cup P_2(\Crit(P_1))\cup P_2(P_1(E))=\Crit(P_2)\cup P_2(E')\subset\QQ\setminus\{0\}$. \medskip Let us now prove that there exists $\alpha\in\QQ$ such that $G_\alpha(0)\not\in\Crit(G_\alpha)\cup G_\alpha(E)$. Let us first treat the case where $0<b_1<b_2,\dots,b_n$. When $\alpha$ approaches $\frac{1}{2}$ , $G_\alpha(0)=\prod_{j=1}^n-b_j(b_j-2\alpha)$ approaches $0$ but the critical values of $G_\alpha$ do not. Indeed, $\Crit(G_\alpha)=\Crit F_\alpha\cup F_\alpha(\Crit((X-\alpha)^2))=\Crit(F_\alpha)\cup\{F_\alpha(0)\}$; $F_\alpha(0)$ approaches $F_{\frac{b_1}{2}}(0)\neq0$, and since $F_{\frac{b_1}{2}}$ does not have multiple roots, the critical values of $F_\alpha$ approach the critical values of $F_{\frac{b_1}{2}}$ which are all non zero. Therefore for $\alpha\neq\frac{b_1}{2}$ in the neighborhood of $\frac{b_1}{2}$ we have $G_\alpha(0)\not\in\Crit(G_\alpha)$. Moreover $G_\alpha(0),G_\alpha(a_1),\dots,G_\alpha(a_m)$ are all distinct polynomials in the indeterminate $\alpha$, so they coincide at only finitely many points. In particular for $\alpha\neq\frac{b_1}{2}$ in the neighborhood of $\frac{b_1}{2}$ we have $G_\alpha(0)\not\in\{G_\alpha(a_1),\dots,G_\alpha(a_m)\}$. Since $\alpha\in\QQ$ we also have $G_\alpha(0)\neq0=G_\alpha(b_1)=\dots=G_\alpha(b_n)$ hence $G_\alpha(0)\not\in G_\alpha(E)$, proving the existence of $\alpha$ as desired. \medskip Let us now treat the general case where $b_1,\dots,b_n$ are not assumed to be positive by reducing it to the previous case. For ${\alpha'}\in\QQ$, define $H_{\alpha'}\in\QQ[X]$ by $$H_{\alpha'}\coloneqq (X-{\alpha'})^2-{\alpha'}^2\in\QQ[X].$$ Note that $\Crit(H_{\alpha'})=\{-{\alpha'}\}$. For ${\alpha'}>0$ sufficiently small we have $-{\alpha'}^2<H_{\alpha'}(0)=0<H_{\alpha'}(a_1),\dots,H_{\alpha'}(a_m),H_{\alpha'}(b_1),\dots,H_{\alpha'}(b_n)$. Let $E''\coloneqq\Crit(H_{\alpha'})\cup H_{\alpha'}(E)$. The set $E''$ is a finite subset of $\QQb\cap\RR\setminus\{0\}$ fixed by $\AGG$, and $E''$ has at most $n$ non rational elements, which are all positive. By the above, there exists $P_3\in\QQ[X]$ such that $\Crit(P_3)\cup P_3(E'')\subset\QQ\setminus\{0\}$ and $P_3(0)=0$. Then $P\coloneqq P_3\circ H_{\alpha'}$ has the desired properties, since $P(0)=0$ and $\Crit(P)\cup P(E)=\Crit(P_3)\cup P_3(\Crit(H_{\alpha'}))\cup P_3(H_{\alpha'}(E))=\Crit(P_3)\cup P_3(E'')\subset\QQ\setminus\{0\}$. \end{proof} Let us denote by $P_1$ the polynomial obtained using this lemma, which verifies $P_1(0)=0$ and $E'\coloneqq\Crit(P_1)\cup P_1(E)\subset\QQ\setminus\{0\}$. We can further assume that $P_1'(0)>0$ by taking $(-P_1)$ if necessary. We now send the points $E'$ to $\{0,1\}$. \begin{lemma} Let $E\subset\QQ\setminus\{0\}$ a finite set. Then there exists $P\in\QQ[X]$ such that $P(E)\subseteq\{0\}$, $\Crit(P)\subseteq\{0,1\}$, $0<P(0)<1$ and $P'(0)>0$. \end{lemma} \begin{proof} For $\alpha\in\QQ$, let $F_\alpha\coloneqq(X-\alpha)^2\in\QQ[X]$, and note that $\Crit(F_\alpha)=\{0\}$. There exists $\alpha<0$ sufficiently small such that $0<F_\alpha(0)<F_\alpha(a)$ for all $a\in E$. We take $$F\coloneqq\frac{F_\alpha}{\max_{a\in E}F_\alpha(a)}.$$ Let $\{a_1,\dots,a_l\}=F(E)$ such that $0<F(0)<a_1<\dots<a_l=1$. We also add a rational point $a_0\in\QQ$ such that $F(0)<a_0<a_1$. \medskip Let $m$ and $n$ be the coprime positive integers such that $a_{l-1}=\frac{m}{m+n}$. We recall that $B_{m,n}$ verifies $\Crit(B_{m,n})=\{0,1\}$, $B_{m,n}(0)=B_{m,n}(1)=0$, $B_{m,n}(\frac{m}{m+n})=1$, and $B_{m,n}$ is strictly increasing between $0$ and $\frac{m}{m+n}$. Let $P_1\coloneqq B_{m,n}$, then $\Crit(P_1)=\{0,1\}$ and $0<P_1\circ F(0)<P_1(a_0)<\dots<P_1(a_{l-1})=1$. There is one point fewer than before, so we can iteratively construct $P_2,\dots,P_l$ in the same way, so that $P\coloneqq P_l\circ\cdots\circ P_1$ verifies $\Crit(P)\subseteq\{0,1\}$, $P(a_1)=\cdots=P(a_l)=0<P(F(0))<1=P(a_0)$ and $P'(F(0))>0$. Therefore $P\circ F$ has the desired properties. \end{proof} Let us denote by $P_2$ the polynomial obtained using this lemma with the finite set $E'$ obtained previously. Then the polynomial $P\coloneqq P_2\circ P_1$ verifies $P(E)\subseteq\{0\}$, $\Crit(P)\subseteq\{0,1\}$, $0<P(0)<1$ and $P'(0)>0$, thus concluding the proof of Proposition \ref{unramified-belyi-function}. \end{proof} We can now use Proposition \ref{unramified-belyi-function} with the finite set $$E\coloneqq\{(1-\zeta_p^k)^{2p}\}_{1\leq k\leq\frac{p-1}{2}}$$ to obtain the map $\beta_1$ as desired. For $1\leq k\leq\frac{p-1}{2}$ we have $(1-\zeta_p^k)^{2p}=(|1-\zeta_p^k|\zeta_{2p}^{2k-1})^{2p}=|1-\zeta_p^k|^{2p}\in\RR$ so $E\subset\QQb\cap\RR$, and the set $E$ is the Galois orbit of $(1-\zeta_p)^{2p}$ so it is fixed by $\AGG$, hence $E$ verifies the conditions of Proposition \ref{unramified-belyi-function}. Let us denote by $\cD(\beta_1)$ the dessin corresponding to the Belyi pair $(\PP,\beta_1)$. The Belyi pair $(\PP,\beta_1)$ is fixed by the action of the complex conjugation, so the embedding of $\cD(\beta_1)$ on $\PP$ admits a symmetry along the real line. Moreover the Belyi function $\beta_1$ is a polynomial, so $\cD(\beta_1)\cap\RR$ is a (graph theoretic) path. Let $v_l<\dots<v_1$ be the negative vertices on the path, and let $e_k$ denote the edge $(v_{k-1},v_k)$. By hypothesis $\beta_1'(0)>0$ so $v_1$ is a black vertex, and for $k<l$, the vertex $v_k$ is of even degree $2d_k$. We then have $e_k^{\genx^{d_k}}=e_{k+1}$ and $e_{k+1}^{\genx^{d_k}}=e_k$ if $k$ is odd, or $e_k^{\geny^{d_k}}=e_{k+1}$ and $e_{k+1}^{\geny^{d_k}}=e_k$ if $k$ is even. See Figure \ref{construction-beta-ex3}. \medskip As remarked earlier, the Galois orbit of $(1-\zeta_p)^{2p}$ is $\{(1-\zeta_p^k)^{2p}\}_{1\leq k\leq\frac{p-1}{2}}\subset\RR_-$, and $(1-\zeta_p^{\frac{p-1}{2}})^{2p}<\cdots<(1-\zeta_p)^{2p}<0$. By construction $\beta_1((1-\zeta_p)^{2p})=0$, so $(1-\zeta_p)^{2p}$ and all its Galois conjugates are black vertices of $\cD(\beta_1)$ lying on the path $(v_1,\cdots,v_l)$. Let $t>0$ be the index such that $v_t=(1-\zeta_p)^{2p}$, and $v_t$ is a black vertex so $t$ is odd. Then $$\mu_0\coloneqq\genx^{d_1}\geny^{d_2}\cdots\geny^{d_{t-1}}\genx^{2d_{t}}\geny^{d_{t-1}}\cdots\geny^{d_2}\genx^{d_1}$$ fixes the edge $e_1$ (Figure \ref{construction-beta-ex3}). \begin{figure}[H] \centering \includegraphics[height=2.2cm]{ex2-beta1} \caption{Dessin $\cD(\beta_1)$ corresponding to $(\PP,\beta_1)$.} \label{construction-beta-ex3} \end{figure} Let $\gamma>0$ small enough so that $\beta_1'>0$ on $[0,\gamma^{2p}q^2]$. Let us next draw the dessin $\cD(\beta_2)$ corresponding to the Belyi pair $(\PP,\beta_2=B_{r,s}\circ B_{m,n})$. See Figure \ref{ex2-beta2}. \begin{figure}[H] \centering \includegraphics[height=4.2cm]{ex2-beta2} \caption{Dessin $\cD(\beta_2)$ corresponding to $(\PP,\beta_2)$.} \label{ex2-beta2} \end{figure} By lifting the dessin $\cD(\beta_2)$ along $\beta_1$ we obtain the dessin $\cD(\beta_2\circ\beta_1)$ corresponding to the Belyi pair $(\PP,\beta_2\circ\beta_1)$. This amounts to replacing each edge of $\cD(\beta_1)$ by a copy of $\cD(\beta_2)$. Note that the degrees of the black and white vertices are thus multiplied by $mr$ and $nr$, respectively. Analogously to $\mu_0$ we define \begin{gather*} \begin{split} \mu\coloneqq\MoveEqLeft(\genx^{mrd_1}\geny\genx^s\geny)(\genx^{nrd_2}\geny\genx^s\geny)\cdots(\genx^{mrd_{t-2}}\geny\genx^s\geny)(\genx^{nrd_{t-1}}\geny\genx^s\geny)\\ &\cdot(\genx^{2mrd_t}\geny\genx^s\geny)(\genx^{nrd_{t-1}}\geny\genx^s\geny)(\genx^{mrd_{t-2}}\geny\genx^s\geny)\cdots(\genx^{nrd_2}\geny\genx^s\geny)\genx^{mrd_1} \end{split} \end{gather*} and we verify again that $\mu$ fixes the edge $(0,v_1)$. Note also that $\genx^{2s}$ fixes the edge $(0,\gamma\sqrt[p]{q})$. See Figure \ref{ex2-beta21}. \begin{figure}[H] \centering \includegraphics[height=3cm]{ex2-beta21} \caption{Dessin $\cD(\beta_2\circ\beta_1)$ corresponding to $(\PP,\beta_2\circ\beta_1)$.} \label{ex2-beta21} \end{figure} Let $\cD_0$ be the dessin corresponding to the Belyi pair $(\PP,\beta_2\circ\beta_1\circ\beta_0)$. To simplify the representations of the dessins we only show the vertices $0$, $\zeta_{2p}^k(1-\zeta_p)$, $1-\zeta_p^k$, and $\zeta_{2p}^k\gamma\sqrt[p]{q}$. We decorate the vertices $\zeta_{2p}^k(1-\zeta_p)$ (which map to $(1-\zeta_p)^{2p}\in\RR_-$ by $\beta_0$) and $\zeta_{2p}^k\gamma\sqrt[p]{q}$ (which map to $\gamma^{2p}q^2\in\RR_+$ by $\beta_0$) respectively with the symbols $\treem$ and $\treep$ to distinguish them. See Figure \ref{ex2-D0}. \ignore{The dessin $\cD$ corresponding to $(C,\beta)$ will then be obtained by lifting $\cD_0$ to $C$.} \begin{figure}[H] \centering \includegraphics[height=8.5cm]{ex2-D0} \caption{Dessin $\cD_0$ corresponding to $(\PP,\beta_2\circ\beta_1\circ\beta_0)$.} \label{ex2-D0} \end{figure} We may now draw the Galois conjugates $\cD^\sigma$ for $\sigma\in\AGG$ by lifting the dessin $\cD_0$ along the projection $\pi$, by treating separately the cases $\sigma(\zeta_p)\in\{\zeta_p,\bar{\zeta_p}\}$ and $\sigma(\zeta_p)\in\{\zeta_p^2,\dots,\zeta_p^{p-2}\}$. We call the dessins respectively $\cD_k$ and $\cD_k^j$, see Figure \ref{ex2-D}. We identify the outermost edges on opposite sides in the representations. \begin{figure}[H] \begin{subfigure}{0.5\hsize} \centering \includegraphics[width=\hsize]{ex2-D-1} \caption{$\cD_k\coloneqq\cD^\sigma$ with $\sigma(\zeta_p)=\zeta_p^{\pm1}$.} \label{ex2-Dk} \end{subfigure} \begin{subfigure}{0.5\hsize} \centering \includegraphics[width=\hsize]{ex2-D-2} \caption{$\cD_k^j\coloneqq\cD^\sigma$ with $\sigma(\zeta_p)=\zeta_p^j$ where $j\in\{2,\dots,p-2\}$.} \label{ex2-Dkj} \end{subfigure} \caption{Dessins (a) $\cD_k$ and (b) $\cD_k^j$ in the Galois orbit of $\cD$.} \label{ex2-D} \end{figure} For all $k$, we have in fact $\cD_{2k-1}=\cD^\sigma$ where $\sigma\colon(\zeta_p,\sqrt[p]{q})\mapsto(\zeta_p,\zeta_p^{k-1}\sqrt[p]{q})$, and $\cD_{2k}=\cD^\sigma$ where $\sigma\colon(\zeta_p,\sqrt[p]{q})\mapsto(\bar{\zeta_p},\zeta_p^{k}\sqrt[p]{q})$. We have similar expressions for the dessins $\cD_k^j$. \medskip \ignore{Let $k$ be fixed, and let us consider one of $\cD_k$ or $\cD_k^j$. Let $E$ denote the set of edges incident to $0$. The action of $\geny$ induces a cyclic permutation of the edges of $E$.} Let $k$ be fixed, and let us consider the dessin $\cD_k$. Let $A$ denote one of the two edges incident to $0$ and on the path to the ramification point $1-\zeta_p^{\pm1}$. We also call $B\coloneqq A^{\geny^{2k-1}},C\coloneqq A^{4p},D\coloneqq C^{\geny^{2k-1}}$ (See Figure \ref{ex2-Dk}). Let $E$ denote the set of edges incident to $0$. The action of $\geny$ induces the cyclic permutation of the edges of $E=\{A^{\geny^i}\}_{0\leq i<8p}$. Furthermore by construction every white vertex aside from $0$ has degree $1$ or $2$, so $\geny^2$ fixes every edge in the complement of $E$. We can write $E=E^{\treem}\sqcup E^{\treep}$ as the disjoint union of $E^{\treem}\coloneqq\{A^{2i}\}_{0\leq i<4p}$ and $E^{\treep}\coloneqq\{B^{2i}\}_{0\leq i<4p}$, such that $\geny$ sends one to the other. The action of $\mu$ on $E^\treem$ is the transposition $(A,C)$, and similarly the action of $\genx^{2s}$ on $E^\treep$ is the transposition $(B,D)$. We do the same for the dessins of the form $\cD_k^j$, with the only difference that this time the action of $\mu$ on $E^\treem$ is trivial, including on the edges $A$ and $C$ (see Figure \ref{ex2-Dkj}). \bigskip We are almost in the same configuration as in the first example. We define analogously $$\omega\coloneqq \mu\geny \mu^{-1}\genx^{2s}\geny^{-1}\mu,$$ and we shall prove that for some choices of $\gamma$, the actions of $\omega$ and of $\geny^2$ commute only for $\cD_1$. To reproduce the proof in the first example we need only show that for some choice of $\gamma$ the actions of $\mu\geny\mu^{-1}\genx^{2s}$ and $\mu^{-1}\genx^{2s}\geny^{-1}\mu$ on the set $E^\treep$ is the same as that of $\genx^{2s}$. Note that for any edge $e\in E^\treep$, the edge $e^{\genx^i}$ is fixed by $\geny$ if $i$ is not a multiple of $s$. To that end we shall show that for some choice of $\gamma$ the action of $\mu$ on $E^\treep$ is the same as that of $\genx^\delta$, where $\delta$ is the number of occurences of $\genx$ in the word $\mu$, and then that $\delta$ is not a multiple of $s$. \medskip We define the words $\rho_1,\rho'_1,\rho_2,\rho'_2,\dots,\rho_{2t-2},\rho'_{2t-2}\in F_2$ to be the increasing subsequence of the prefixes ending in $\eta$ of the word $\mu$ defined above, such that $\rho_1\coloneqq\genx^{mrd_1}\geny$, $\rho_1'\coloneqq\rho_1\genx^s\geny$, $\rho_2\coloneqq\rho_1'\genx^{nrd_2}\geny$, $\rho_2'\coloneqq\rho_2\genx^s\geny$, etc., and $\mu=\rho_{2t-2}'\genx^{mrd_1}$. We shall show by induction that for some choice of $\gamma$ the action of $\rho_i$ (resp. $\rho_i'$) is the same as the action of $\genx^{\delta_i}$ (resp. $\genx^{\delta_i'}$), where $\delta_i$ (resp. $\delta_i'$) is the number of occurences of $\genx$ in the word $\rho_i$ (resp. $\rho_i'$). By induction it suffices to show that $\delta_i$, $\delta_i'$ are not multiples of $s$. Modulo $s$ we have $\delta_i\equiv\delta_i'$ equal to the non empty partial sum of $$mrd_1+nrd_2+\cdots+mrd_{t-2}+nrd_{t-1}+2mrd_t+nrd_{t-1}+mrd_{t-2}+\cdots+nrd_2+mrd_1$$ consisting of the first $i$ terms. \bigskip To proceed we shall use the following result, but let us first introduce some notations. Let $P=\sum_{i=0}^dc_iX^i\in\ZZ[X]$ and $c\in\ZZ_{>0}$ such that $\beta_1=\frac{P}{c}$. Note that $P$ and $c$ do not depend on the choice of $\gamma$, and $0<\beta_1(0)=\frac{P(0)}{c}<1$ so $0<c_0,c-c_0$. We define $$\alpha\coloneqq v_2(\gamma^{2p}q^2),\quad\nu\coloneqq v_2(c_0)+v_2(c-c_0),$$ where $v_2$ denotes the $2$-valuation. \begin{lemma} If $\alpha>\nu$, then there exists $e\in\ZZ$ such that $em\equiv c_0\mod2^\alpha$ and $en\equiv c-c_0\mod2^\alpha$, and $v_2(s)\geq\alpha-\nu$. \end{lemma} \begin{proof} Let $a,b\in\ZZ$ coprime such that $\gamma^{2p}q^2=\frac{a}{b}2^\alpha$. Firstly, $$\frac{m}{m+n}=\beta_1(\frac{a}{b}2^\alpha)=\frac{P(\frac{a}{b}2^\alpha)}{c}=\frac{\sum_{i=0}^dc_ia^i2^{\alpha i}b^{d-i}}{b^dc},$$ so there exists $f\in\ZZ$ such that $fm=\sum_{i=0}^dc_ia^i2^{\alpha i}b^{d-i}$ and $f(m+n)=b^dc$, so $$em\equiv c_0\mod2^\alpha,\quad en\equiv c-c_0\mod2^\alpha$$ for $e\in\ZZ$ such that $eb^d\equiv f\mod2^\alpha$. Secondly, $$\frac{r}{r+s}=B_{m,n}(\beta_1(0))=\frac{\beta_1(0)^m(1-\beta_1(0))^n}{\beta_1(\frac{a}{b}2^\alpha)^m(1-\beta_1(\frac{a}{b}2^\alpha))^n}=\frac{b^{d(m+n)}c_0^m(c-c_0)^n}{(b^dP(\frac{a}{b}2^\alpha))^m(b^dc-b^dP(\frac{a}{b}2^\alpha))^n},$$ so there exists $g\in\ZZ$ such that $gr=b^{d(m+n)}c_0^m(c-c_0)^n$ and $g(r+s)=(b^dP(\frac{a}{b}2^\alpha))^m(b^dc-b^dP(\frac{a}{b}2^\alpha))^n$. In the expansion of $(b^dP(\frac{a}{b}2^\alpha))^m$, aside from the constant term $b^{dm}c_0^m$, every other term is a multiple of an integer of the form $c_0^i2^{\alpha j}$ with $i\leq m-1$ and $j\geq m-i$. By hypothesis $\alpha>\nu\geq v_2(c_0)$, so those other terms are all multiples of $2^{\alpha+(m-1)v_2(c_0)}$, hence there exists $A\in\ZZ$ such that $(b^dP(\frac{a}{b}2^\alpha))^m=b^{dm}c_0^m+A2^{\alpha+(m-1)v_2(c_0)}$. Similarly there exists $B\in\ZZ$ such that $(b^dc-b^dP(\frac{a}{b}2^\alpha))^n=b^{dn}(c-c_0)^n+B2^{\alpha+(n-1)v_2(c-c_0)}$. Then $g(r+s)=b^{d(m+n)}c_0^m(c-c_0)^n+C2^{\alpha+(m-1)v_2(c_0)+(n-1)v_2(c-c_0)}$ for some $C\in\ZZ$, so $gr=b^{d(m+n)}c_0^m(c-c_0)^n$ and $gs=C2^{\alpha+(m-1)v_2(c_0)+(n-1)v_2(c-c_0)}$. The integers $r$ and $s$ are coprime, so after dividing $gr$ and $gs$ by their greatest common dividor we obtain that $$v_2(s)\geq \alpha-\nu>0.$$ \end{proof} Using this lemma, we know that if $\alpha>\nu$, then there exists $e\in\ZZ$ such that $em\equiv c_0\mod2^\alpha$ and $en\equiv c-c_0\mod2^\alpha$, $v_2(s)\geq\alpha-\nu$ where $\nu$ does not depend on $\gamma$, and $r$ is coprime to $s$ so is not a multiple of $2$. Therefore there exists $e'\in\ZZ$ such that $e'mr\equiv c_0\mod2^\alpha$ and $e'nr\equiv c-c_0\mod2^\alpha$. Moreover $2^{\alpha-\nu}$ is a common divisor of $2^\alpha$ and $s$, so by the above modulo $2^{\alpha-\nu}$ we have $e'\delta_i\equiv e'\delta_i'$ equal to the non empty partial sum $\widetilde{\delta_i}$ consisting of the first $i$ terms of the sum $$c_0d_1+(c-c_0)d_2+\cdots+c_0d_{t-2}+(c-c_0)d_{t-1}+2c_0d_t+(c-c_0)d_{t-1}+c_0d_{t-2}+\cdots+(c-c_0)d_2+c_0d_1.$$ Similarly $e'\delta$ is equal modulo $2^{\alpha-\nu}$ to the whole sum $$\widetilde{\delta}\coloneqq2(c_0d_1+(c-c_0)d_2+\cdots+c_0d_{t-2}+(c-c_0)d_{t-1}+c_0d_t).$$ By construction $c_0,c-c_0,d_i$ are positive and do not depend on the choice of $\gamma$, so $0<c_0d_1\leq\widetilde{\delta}_i\leq\widetilde{\delta}$, thus for any choice of $\gamma$ such that $\alpha>\nu$ and $\widetilde{\delta}<2^{\alpha-\nu}$ (for instance $\gamma=\frac{2^u}{2^v+1}$ with $1\ll u\ll v$), we obtain $\widetilde{\delta}_i,\widetilde{\delta}\not\equiv 0\mod 2^{\alpha-\nu}$, and in consequence $\delta_i,\delta_i'$ and $\delta$ are not multiples of $s$. Therefore we can now conclude by induction that the actions of $\rho_i$ and $\rho_i'$ are the same as that of $\genx^{\delta_i}$ and $\genx^{\delta_i'}$, respectively. Indeed, $\delta_1$ is not a multiple of $s$ so $\rho_1=\genx^{\delta_1}\geny$ and $\genx^{\delta_1}$ have the same action on $E^\treep$. If $\rho_i$ has the same action as $\genx^{\delta_i}$ on $E^\treep$, then $\rho_i'=\rho_i\genx^s\geny$ has the same action as $\genx^{\delta_i}\genx^s\geny=\genx^{\delta_i'}\geny$ on $E^\treep$, and also the same action as $\genx^{\delta_i'}$ because $\delta_i'$ is not a multiple of $s$. Similarly, if $\rho_i'$ has the same action as $\genx^{\delta_i'}$ on $E^\treep$, then $\rho_{i+1}$ has the same action as $\genx^{\delta_{i+1}}\geny$ on $E^\treep$, and also the same action as $\genx^{\delta_{i+1}}$ because $\delta_{i+1}$ is not a multiple of $s$. \medskip We have thus proved that $\mu$ has the same action as $\genx^{\delta}$ on $E^\treep$, and by symmetry $\mu^{-1}$ has the same action as $\genx^{-\delta}$ on $E^\treep$. And $\delta$ and $2s-\delta$ are not multiples of $s$, so $\mu\geny\mu^{-1}\genx^{2s}$ and $\mu^{-1}\genx^{2s}\geny^{-1}\mu$ have the same action as $\genx^{2s}$ on $E^\treep$, as announced. We shall now observe the action of $\omega=\mu\geny\mu^{-1}\genx^{2s}\geny^{-1}\mu$ on $E$. Let $M_k$ and $M_k^j$ denote the monodromy maps of the dessins $\cD_k$ and $\cD_k^j$. \medskip For the dessins $\cD_k$ for $1\leq k\leq 2p$, the action of $\mu$ on $E^\treem$ is the transposition $(A,C)$, and the action of $\genx^{2s}$ on $E^\treep$ is the transposition $(B,D)$, therefore the action of $\omega$ fixes the set $E$ on which it induces the permutation $$M_k(\omega)|_E=(A,C)(B,D)\cdot\prod_{i=0}^{4p-1}(A^{\geny^{2i}},A^{\geny^{2i+1}})\cdot(A,C)(B,D).$$ Hence for $k=1$, \begin{align*} M_1(\omega)|_E&=(A,A^{\geny^{4p}})(A^\geny,A^{\geny^{4p+1}})\cdot\prod_{i=0}^{4p-1}(A^{\geny^{2i}},A^{\geny^{2i+1}})\cdot(A,A^{\geny^{4p}})(A^\geny,A^{\geny^{4p+1}})\\ &=\prod_{i=0}^{4p-1}(A^{\geny^{2i}},A^{\geny^{2i+1}}) \end{align*} so $\omega$ and $\geny^2$ commute on $E$. Moreover $\geny^2$ acts trivially on the complement of $E$, so finally $M_1(\omega)$ and $M_1(\geny^2)$ commute. \medskip For $k=2$, we observe that $B^{\omega\geny^2}=D^{\geny^{-1}\geny^2}=D^{\geny}$ but $B^{\geny^2\omega}=B^{\geny^2\geny^{-1}}=B^{\geny}$. Similarly, for $3\leq k\leq 2p$, we observe that $A^{\omega\geny^2}=C^{\geny\geny^2}=C^{\geny^3}$ but $A^{\geny^2\omega}=A^{\geny^2\geny}=A^{\geny^3}$. Therefore $M_k(\omega)$ and $M_k(\geny^2)$ do not commute for $2\leq k\leq 2p$. \medskip For the dessins $\cD_k^j$ for $1\leq k\leq 2p$ and $2\leq j\leq \frac{p-1}{2}$, $\genx^{2s}$ on $E^\treep$ is the transposition $(B,D)$, and $\mu$ acts trivially on $E^\treem$, therefore the action of $\omega$ fixes the set $E$ on which it induces the permutation $$M_k^j(\omega)|_E=(B,D)\cdot\prod_{i=0}^{4p-1}(A^{\geny^{2i}},A^{\geny^{2i+1}})\cdot(B,D).$$ Hence we observe that $B^{\omega\geny^2}=D^{\geny^{-1}\geny^2}=D^{\geny}$ but $B^{\geny^2\omega}=B^{\geny^2\geny^{-1}}=B^{\geny}$, so $M_k^j(\omega)$ and $M_k^j(\geny^2)$ do not commute. \medskip We have thus shown that the actions of $\omega$ and $\geny^2$ commute only for $\cD_1$, this concludes the proof that $\widetilde{\cD}$ is a regular dessin with moduli field $\QQ(\zeta_p,\sqrt[p]{q})$. \subsection{Regular dessin with moduli field $\QQ(\zeta_3,\sqrt[3]{3})$}\label{ex3} Finally, let us exhibit a regular dessin with moduli field $\QQ(\zeta_3,\sqrt[3]{3})$ of smaller degree by choosing a Belyi map that is a rational function instead of a polynomial as was done in the previous subsections. Let \begin{align*} C&\colon y^2=x(x-(1-\zeta_3))(x-\sqrt[3]{3}),\\ \beta&\colon C\to\PP,(x,y)\mapsto\frac{(x+3^3)^3}{3^5(x-3^2)^2}. \end{align*} The function $\beta$ is given by the composition of the following maps $\beta=\beta_1\circ\beta_0\circ\pi$. \begin{enumerate} \item $\pi\colon C\to\PP$ is the projection on the coordinate $x$, which ramifies over $\{0,1-\zeta_3,\sqrt[3]{3},\infty\}$. \item $\beta_0\coloneqq X^6\in\QQ[X]$, $\Crit(\beta_0)=\{0\}$ so $\beta_0\circ\pi$ ramifies over $\{0,(1-\zeta_3)^6=-3^3,3^2,\infty\}$. \item $\beta_1\coloneqq\frac{(X+3^3)^3}{3^5(X-3^2)^2}$, $\Crit(\beta_1)=\{0,1\}$ so $\beta=\beta_1\circ\beta_0\circ\pi$ ramifies over $\{0,1,\infty\}$. \end{enumerate} The pair $(C,\beta)$ is thus a Belyi pair, and we call $\cD$ the dessin corresponding to $(C,\beta)$. Similarly as in \ref{ex1}, $\cD$ has moduli field $\QQ(\zeta_3,\sqrt[3]{3})$. We will proceed analogously to show that the regular closure $\widetilde{\cD}$ has the same field of moduli. Let us first draw the dessin $\cD_0$ corresponding to the Belyi pair $(\PP,\beta_1\circ\beta_0)$ (see Figure \ref{D0-ex3}), and lift it to the conjugate curves $C^\sigma$ to obtain the conjugate dessins $\cD^\sigma$ for $\sigma\in\Gal(\QQ(\zeta_3,\sqrt[3]{3})/\QQ)$ (see Figure \ref{ex3-dessins}). \begin{figure}[H] \centering \includegraphics[width=6cm]{beta-ex3} \caption{Construction of $\cD_0$.} \label{D0-ex3} \end{figure} As usual we identify the outermost edges on opposite sides. \begin{figure}[H] \begin{subfigure}{0.5\hsize} \begin{center} \includegraphics[height=0.9\hsize]{ex3-1} \end{center} \caption{$\cD_1\coloneqq\cD$} \label{ex3-1} \end{subfigure} \begin{subfigure}{0.5\hsize} \begin{center} \includegraphics[height=0.9\hsize]{ex3-2} \end{center} \caption{$\cD_2\coloneqq\cD^\sigma,\sigma\colon(\zeta_3,\sqrt[3]{q})\mapsto(\zeta_3^2,\zeta_3\sqrt[3]{q})$} \end{subfigure} \end{figure} \begin{figure}[H]\ContinuedFloat \begin{subfigure}{0.5\hsize} \begin{center} \includegraphics[height=0.9\hsize]{ex3-3} \end{center} \caption{$\cD_3\coloneqq\cD^\sigma,\sigma\colon(\zeta_3,\sqrt[3]{q})\mapsto(\zeta_3,\zeta_3\sqrt[3]{q})$} \end{subfigure} \begin{subfigure}{0.5\hsize} \begin{center} \includegraphics[height=0.9\hsize]{ex3-4} \end{center} \caption{$\cD_4\coloneqq\cD^\sigma,\sigma\colon(\zeta_3,\sqrt[3]{q})\mapsto(\zeta_3^2,\zeta_3^2\sqrt[3]{q})$} \end{subfigure} \end{figure} \begin{figure}[H]\ContinuedFloat \begin{subfigure}{0.5\hsize} \begin{center} \includegraphics[height=0.9\hsize]{ex3-5} \end{center} \caption{$\cD_5\coloneqq\cD^\sigma,\sigma\colon(\zeta_3,\sqrt[3]{q})\mapsto(\zeta_3,\zeta_3^2\sqrt[3]{q})$} \end{subfigure} \begin{subfigure}{0.5\hsize} \begin{center} \includegraphics[height=0.9\hsize]{ex3-6} \end{center} \caption{$\cD_6\coloneqq\cD^\sigma,\sigma\colon(\zeta_3,\sqrt[3]{q})\mapsto(\zeta_3^2,\sqrt[3]{q})$} \label{ex3-6} \end{subfigure} \caption{Dessins $\cD_1,\dots,\cD_6$ in the Galois orbit of $\cD$.} \label{ex3-dessins} \end{figure} We can now compute the cartographic groups of the dessins. Let $M_k$ denote the monodromy map of $\cD_k$. Then $$M_k(\genx)=\begin{tabular}{c} (1,13,14,7,25,26)(2,15,16)(3,17,18)(4,19,20)(5,21,22)\\ (6,23,24)(8,27,28)(9,29,30)(10,31,32)(11,33,34)(12,35,36) \end{tabular}$$ for all $1\leq k\leq6$, and \begin{itemize} \item $M_1(\geny)=\begin{tabular}{c} (1,2,3,4,5,6,7,8,9,10,11,12)(13,36)(14,15)(16,17)(18,19)\\ (20,21)(22,23)(24,25)(26,27)(28,29)(30,31)(32,33)(34,35) \end{tabular}$, \item $M_2(\geny)=\begin{tabular}{c} (1,2,3,4,5,6,7,8,9,10,11,12)(13,36)(14,27)(15,26)(16,29)\\ (17,28)(18,19)(20,21)(22,23)(24,25)(30,31)(32,33)(34,35) \end{tabular}$, \item $M_3(\geny)=\begin{tabular}{c} (1,2,3,4,5,6,7,8,9,10,11,12)(13,36)(14,27)(15,26)(16,17)\\ (18,31)(19,30)(20,21)(22,23)(24,25)(28,29)(32,33)(34,35) \end{tabular}$, \item $M_4(\geny)=\begin{tabular}{c} (1,2,3,4,5,6,7,8,9,10,11,12)(13,36)(14,27)(15,26)(16,17)\\ (18,19)(20,33)(21,32)(22,23)(24,25)(28,29)(30,31)(34,35) \end{tabular}$, \item $M_5(\geny)=\begin{tabular}{c} (1,2,3,4,5,6,7,8,9,10,11,12)(13,36)(14,27)(15,26)(16,17)\\ (18,19)(20,21)(22,35)(23,34)(24,25)(28,29)(30,31)(32,33) \end{tabular}$, \item $M_6(\geny)=\begin{tabular}{c} (1,2,3,4,5,6,7,8,9,10,11,12)(13,24)(14,27)(15,26)(16,17)\\ (18,19)(20,21)(22,23)(25,36)(28,29)(30,31)(32,33)(34,35) \end{tabular}$. \end{itemize} Using the computer algebra system SageMath \cite{sagemath}, we determined that $$|\langle M_1(\genx),M_1(\geny)\rangle|=42467328=2^{19}\cdot3^4.$$ Moreover, $M_1(\genx)$, $M_1(\geny)$ and $M_1(\genx\geny)$ respectively have orders $6$, $12$ and $12$, so the Euler characteristic of the underlying surface of $\widetilde{\cD_1}$ is $\chi=|\langle M_1(\genx),M_1(\geny)\rangle|\cdot(\frac{1}{\ord M_1(\genx)}+\frac{1}{\ord M_1(\geny)}+\frac{1}{\ord M_1(\genx\geny)}-1)=-28311552=-2^{20}\cdot3^3$, and its genus is $g=1-\frac{\chi}{2}=14155777$. \medskip We will now show that $\widetilde{\cD_1}$ is not isomorphic to $\widetilde{\cD_2},\dots,\widetilde{\cD_6}$. We claim that $\omega\coloneqq[\genx^{-1}\geny^2\genx,\genx\geny]\in\ker M_1\setminus\bigcup_{2\leq k\leq6}\ker M_k$, thus concluding the proof. Indeed, we obtain: \begin{itemize} \item $M_1(w)=\id$; \item $M_2(w)=(13,25)(15,27)(21,33)(23,35)$; \item $M_3(w)=(17,29)(21,33)$; \item $M_4(w)=(13,25)(15,27)(19,31)(21,33)$; \item $M_5(w)=(13,25)(17,29)$; \item $M_6(w)=(13,25)(19,31)(21,33)(23,35)$. \end{itemize} We have thus constructed a regular dessin $\widetilde{\cD}$ of degree $2^{19}\cdot3^4$ and genus $14155777$ with moduli field $\QQ(\zeta_3,\sqrt[3]{3})$. \bibliographystyle{plain}

https://arxiv.org/abs/1503.05303

Asymptotic and chaotic solutions of a singularly perturbed Nagumo-type equation

We deal with the singularly perturbed Nagumo-type equation $$ \epsilon^2 u'' + u(1-u)(u-a(s)) = 0, $$ where $\epsilon > 0$ is a real parameter and $a: \mathbb{R} \to \mathbb{R}$ is a piecewise constant function satisfying $0 < a(s) < 1$ for all $s$. We prove the existence of chaotic, homoclinic and heteroclinic solutions, when $\epsilon$ is small enough. We use a dynamical systems approach, based on the Stretching Along Paths method and on the Conley-Wazewski's method.

\section{Introduction} \def5.\arabic{equation}{1.\arabic{equation}}\makeatother \setcounter{equation}{0} In this paper, we deal with the singularly perturbed Nagumo-type equation \begin{equation}\label{eqintro} \epsilon^2 u'' + u(1-u)(u-a(s)) = 0, \end{equation} where $\epsilon > 0$ is a small parameter and $a: \mathbb{R} \to \mathbb{R}$ is a locally integrable function satisfying $$ 0 < a(s) < 1, \quad \mbox{ for every } s \in \mathbb{R}. $$ Notice that equation \eqref{eqintro} has the two constant solutions $u \equiv 0$ and $u \equiv 1$ which actually behave like saddle points); we will be interested in the existence of solutions $u$ satisfying $0 < u(s) < 1$ for all $s$. \medbreak Our investigation is motivated by a classical paper by Angenent, Mallet-Paret and Peletier \cite{AngMalPel87}, dealing with the Neumann boundary value problem associated with \eqref{eqintro} in the framework of steady-states solutions of the corresponding parabolic problem (which arises in population genetics). On the lines of previous works \cite{ClePel85,Kur83}, in \cite{AngMalPel87} it is proved that, for $\epsilon$ small enough, the Neumann boundary value problem associated with \eqref{eqintro} has multiple solutions, whose shape and limit profile (for $\epsilon \to 0^+$) can be precisely described in terms of the zeros of the function $a - 1/2$. This analysis suggests that the dynamics of \eqref{eqintro} could be quite rich provided the function $a$ crosses the value $1/2$. \medbreak The aim of the present paper is indeed to prove the existence of solutions to \eqref{eqintro} defined on the whole real line and exhibiting complex behavior, when $a$ switches infinitely many times between two values $a_-,a_+$ with $$ 0 < a_- < \frac{1}{2} < a_+ < 1. $$ (see, however, Remark \ref{stabile} for possible generalizations of the above assumption). More precisely, we will provide solutions of essentially two types. On one hand, we find the existence of globally defined solutions rotating (in the phase-plane) a certain number of times around $(a_{\pm},0)$ in suitable intervals $I^{\pm}_j$ ($j \in \mathbb{Z}$) (solutions of this kind will be called \emph{chaotic}); the number of revolutions can be arbitrarily prescribed and $\cup_{j \in \mathbb{Z}}I^{\pm}_j$ is unbounded from below and from above, i.e., solutions oscillate infinitely many times on the whole real line. In particular, if $a$ is $T$-periodic for some $T > 0$, we find the existence of subharmonic solutions to \eqref{eqintro} with complex behavior; indeed, one could show that the Poincar\'e map associated with \eqref{eqintro} on a period is topologically semiconjugated to the Bernoulli shift on a suitable number of symbols (as in \cite{CapDamPap02}, see also Remark \ref{coniugio}). On the other hand, we are able to produce \emph{homoclinic} and \emph{heteroclinic} solutions, having the same nodal behavior as before on bounded intervals of arbitrarily large length and converging monotonically to one of the equilibria $(0,0)$ and $(1,0)$ for $t \to \pm \infty$. A similar nonlinearity is considered in the recent papers \cite{EllZan13, EllZan->, ZaZa-12, ZaZa-14}. However, those papers deal with situations which can be considered in a certain sense dual to ours, since they have a fixed central equilibrium (a center, in fact) while the two nearby saddle points move according to a suitable piecewise constant weight function. Moreover, we decided to focus on finding solutions that are obtained by exploiting quite different configurations than those investigated in the mentioned papers, and to omit the details in cases that are similar to those already treated (see Remark \ref{linkinginterno}). \medbreak For the proof of our results, we use the change of variable $x(t) = u(\epsilon t)$, so that the equivalent planar system associated with \eqref{eqintro} is transformed into \begin{equation}\label{sysintro2} x' = y, \qquad y' = x(x-1)(x-a(\epsilon t)); \end{equation} hence, the smallness of $\epsilon > 0$ reflects into the fact that the function $a_{\epsilon}(t) = a(\epsilon t)$ is constant on intervals of large amplitude. As a consequence, \eqref{sysintro2} can be regarded as a slowly varying perturbation of an autonomous system with two hyperbolic equilibrium points; in this setting, the arising of rich dynamics seems to be a quite common phenomenum (see, among others, \cite{FelMarTan06,GedKokMis02} as well as the bibliography in \cite{ZaZa-14}). \medbreak In order to detect such a complex behavior, we use a dynamical systems approach, based on a careful analysis of the trajectories of the (piecewise autonomous) system \eqref{sysintro2}. More in detail, we first rely on a topological technique of path stretching (the so-called SAP method) developed in \cite{PapZan04b,PapZan04, PirZan05,PirZan07} to detect the presence of symbolic dynamics (as well as of periodic points, when $a$ is periodic) for suitable Poincar\'e maps associated with \eqref{sysintro2}. As a by-product, this approach also provides us planar paths which can be used to connect the stable and unstable manifolds of the equilibria so as to obtain heteroclinic and homoclinic solutions with a complex nodal behavior (see, among others, \cite{EllZan13,Gav11,HolStu92,MarRebZan09} and the references therein for related results in this direction). It has to be noticed that, since the equation is non-autonomous, the existence of the stable/unstable manifolds is not straightforward; we use indeed the classical Conley-Wa\.zewski's method \cite{Con75,Waz47} (see also \cite{EllZan->,PapZan00}) to prove that these sets actually exist and can be suitably localized. \medbreak The plan of the paper is the following. In Section \ref{sec2}, we briefly discuss the autonomous case. In Section \ref{sec3}, we prove some stretching properties, as well as the existence of stable and unstable manifolds to the equilibria. In Section \ref{sec4}, we state and prove our main results. Finally, some basic facts about SAP method are collected in a final Appendix. \section{The autonomous case}\label{sec2} \def5.\arabic{equation}{2.\arabic{equation}}\makeatother \setcounter{equation}{0} In this section we collect some basic results for the autonomous equation \begin{equation}\label{eqaut} x'' + x(1-x)(x - a) = 0, \end{equation} where $a$ is a real constant such that $$ 0 < a < 1. $$ Precisely, we are going to perform a phase-plane analysis for the equivalent planar system \begin{equation}\label{sysaut} \left\{ \begin{array}{l} \vspace{0.1cm} x' = y \\ y'= x(x-1)(x-a). \end{array} \right. \end{equation} We will always confine our attention to the dynamics in the vertical strip $[0,1] \times \mathbb{R}$. \smallbreak It is immediately seen that the points $(0,0)$, $(a,0)$ and $(1,0)$ are the only equilibrium points of \eqref{sysaut}. To proceed further, we define the function $$ \mathcal{E}_a(x,y) = \frac{1}{2}y^2 + F_a(x),\quad \forall \ (x,y)\in {\mathbb R}^2, $$ where $$ F_a(x) = -\frac{1}{4}x^4 + \frac{1+a}{3} x^3 - \frac{a}{2}x^2,\quad \forall \ x\in {\mathbb R}. $$ As well-known, system \eqref{sysaut} is conservative and the function $t \mapsto \mathcal{E}_a(x(t),y(t))$ is constant along solutions $(x(t),y(t))$ to \eqref{sysaut}. To describe the global dynamics of \eqref{sysaut}, we can thus study the geometry of the level sets $\mathcal{E}_a^{-1}(c)$ for different values of $c \in \mathbb{R}$. \smallbreak It turns out that the value of the constant $a$ plays a significant role. We start by analyzing the case \begin{equation}\label{a=12} a = \frac{1}{2}, \end{equation} which indeed gives rise to the simplest picture. Precisely, the level set $\mathcal{E}_a^{-1}(c)$ can here be described as follows: \begin{itemize} \item[-] for $c = \tfrac{a-2}{12}a^2$, it is the point $(a,0)$; \item[-] for $\tfrac{a-2}{12}a^2 < c < 0$, it is a closed cycle around $(a,0)$; \item[-] for $c = 0$, it is the union of the points $(0,0)$, $(1,0)$ and of the heteroclinic orbits joining them; \item[-] for $c > 0$, it is the union of two curves, one in the half-plane $\{(x,y)\in {\mathbb R}^2:\ y>0\}$, one in the half-plane $\{(x,y)\in {\mathbb R}^2:\ y<0\}$, connecting two points of the form $(0,y_1)$ and $(1,y_1)$ for some $y_1 \in {\mathbb R}$. \end{itemize} The phase-portrait is shown in Figure \ref{fig1}. \begin{figure}[!h] \centering \includegraphics[height=8cm,width=10cm]{figura1} \caption{\small{The phase-portrait of the autonomous system \eqref{sysaut} for $a = 1/2$. The heteroclinic orbits connecting the equilibria $(0,0)$ and $(1,0)$ are painted with a darker color. For graphical reasons, a slightly different $x$ and $y$ scaling has been used.}} \label{fig1} \end{figure} We now turn our attention to the case \begin{equation}\label{a<12} 0 < a < \frac{1}{2}; \end{equation} now, for the level set $\mathcal{E}_a^{-1}(c)$ we have the following: \begin{itemize} \item[-] for $c = \tfrac{a-2}{12}a^2$, it is the point $(a,0)$; \item[-] for $\tfrac{a-2}{12}a^2 < c < 0$, it is a closed cycle around $(a,0)$; \item[-] for $c = 0$, it is the union of the point $(0,0)$ and its homoclinic orbit $H(a)$; for further convenience, we denote by $(z_a,0)$ the point of intersection between $H(a)$ and the positive $x$-semiaxis; \item[-] for $0 < c < \tfrac{1-2a}{12}$, it is made by a curve lying between the homoclinic to $(0,0)$ and the stable/unstable manifold $H^{\pm} (a)$ of $(1,0)$, connecting a point of the form $(0,y_1)$ with a point of the form $(0,-y_1)$, for some $y_1 > 0$; \item[-] for $c = \tfrac{1-2a}{12}$, it is the union of the point $(1,0)$ and its stable/unstable manifolds; \item[-] for $c > \tfrac{1-2a}{12}$, it is the union of two curves, one in the half-plane $\{(x,y)\in {\mathbb R}^2:\ y>0\}$, one in the half-plane $\{(x,y)\in {\mathbb R}^2:\ y<0\}$, connecting a point of the form $(0,y_1)$ with a point of the form $(1,y_2)$, for some $y_1, y_2\in {\mathbb R}$. \end{itemize} Finally, for \begin{equation}\label{a>12} \frac{1}{2} < a < 1 \end{equation} the phase-portrait can be obtained from the previous one with a symmetry with respect to the line $x = \tfrac{1}{2}$. The homoclinic orbit and the stable/unstable manifolds are defined in an analogous way, by swapping $(0,0)$ and $(1,0)$, and will be again denoted by $H(a)$ and $H^\pm (a)$; moreover, $(z_a,0)$ will be the point of intersection between $H(a)$ and the positive $x$-semiaxis. Both the phase-portraits are shown in Figure \ref{fig2}. \begin{figure}[!h] \includegraphics[height=5cm,width=7cm]{figura2} \hfill \includegraphics[height=5cm,width=7cm]{figura3} \caption{\small{On the left, the phase-portrait of the autonomous system \eqref{sysaut} for $0 < a < 1/2$; on the right, the phase-portrait of the autonomous system \eqref{sysaut} for $1/2 < a < 1$. The homoclinic orbits, as well as the stable/unstable manifolds, are painted with a darker color. For graphical reasons, a slightly different $x$ and $y$ scaling has been used.}} \label{fig2} \end{figure} \begin{remark}\label{bilanciato} \textnormal{ From the above discussion, it appears that the phase-portrait of system \eqref{sysaut} is completely different in the case $a = 1/2$ and in the case $a \neq 1/2$. Indeed, the heteroclinc orbit connecting $(0,0)$ and $(1,0)$ for $a = 1/2$ disappear as soon as $a \neq 1/2$, splitting into orbits of different type. In terms of the potential $F_a(x)$, we have indeed $F_a(0) = F_a(1)$ if and only if $a = 1/2$; in this case, the potential is said to be balanced. In this context, \eqref{eqmain} can be framed in the setting of equations with unbalanced potentials (compare with the introduction in \cite{NakTan03}).} \end{remark} \section{Topological lemmas}\label{sec3} \def5.\arabic{equation}{3.\arabic{equation}}\makeatother \setcounter{equation}{0} In this section we collect the preliminary technical lemmas which will be used in the proof of our main results. \subsection{Stretching properties}\label{stre} In this section, we fix two reals constants $a_-,a_+$ satisfying $$ 0 < a_- < \frac{1}{2} < a_+ < 1. $$ Our goal is to prove some results for the dynamics of the autonomous system \eqref{sysaut} for $a = a_-$ and $a= a_+$ on suitable sets which will be constructed below. Throughout this section, we always refer to the definitions given in the Appendix. Also, to simplify the notation, from now on we denote by $S(a_-)$ (resp., $S(a_+)$) the planar system \eqref{sysaut} for $a = a_-$ (resp., $a = a_+$); moreover, let $\Theta(a,z)$ be the orbit of \eqref{sysaut} passing through the point $z \in \mathbb{R}^2$. For $T>0$ we finally define the maps $\Psi_-^T$ and $\Psi_+^T$ as the (restriction to the vertical strip $[0,1] \times \mathbb{R}$) of Poincar\'e maps associated with systems $S(a_-)$ and $S(a_+)$, respectively, on the interval $[0,T]$, i.e. $$ \Psi_\pm^T(x_0,y_0) = (x(T;x_0,y_0),y(T;x_0,y_0)),\quad \forall \, (x_0,y_0) \in [0,1] \times \mathbb{R}, $$ where $(x(\cdot;x_0,y_0),y(\cdot;x_0,y_0))$ is the unique solution to $S(a_\pm)$ satisfying the initial condition $(x(0),y(0)) = (x_0,y_0)$. Since we will be interested in the dynamics on the strip $[0,1] \times \mathbb{R}$ only, we can assume that such maps are globally defined just by suitably modifying the nonlinearity $f_a(x) = x(1-x)(x-a)$ for $x \notin [0,1]$ (for instance, by setting $f_a(x) = 0$ for $x \notin [0,1]$). \smallbreak Let us fix $p_-$ and $p_+$ such that \begin{equation}\label{ordine0} \frac{1}{2} < \max\left\{z_{a_-},a_+ \right\} < p_- < 1 \qquad \mbox{ and } \qquad 0 < p_+ < \min\left\{a_-,z_{a_+}\right\} < \frac{1}{2}. \end{equation} Let $\mathcal{R}_1$ and $\mathcal{R}_3$ be the two connected components of the intersection of the following two strips: \begin{itemize} \item[-] the strip $S_-$ between the stable manifold $H^+(a_-)$ and the orbit $\Theta (a_-,(p_-,0))$ of $S(a_-)$, \item[-] the strip $S_+$ between the unstable manifold $H^+(a_+)$ and the orbit $\Theta (a_-,(p_+,0))$ of $S(a_+)$. \end{itemize} Notice that, since $p_+ < z_{a_+}$ and $z_{a_-} < p_-$, the above defined orbits passing through $(p_{\pm},0)$ lie between the homoclinics and the stable/unstable manifolds of the corresponding systems $S(a_\pm)$. Therefore, the defined regions $\mathcal{R}_1$ and $\mathcal{R}_3$ are topological rectangles, and we name $\mathcal{R}_1$ (resp., $\mathcal{R}_3$) the one contained in the upper (resp., lower) half-plane. \smallbreak In what follows we also need to provide an orientation for the above constructed rectangles; to this aim, we denote by $\mathcal{R}_1^\pm$ the components of the boundary of $\mathcal{R}_1$ lying on orbits of system $S(a_\pm)$ and by $\mathcal{R}_3^\pm$ the components of the boundary of $\mathcal{R}_1$ lying on orbits of system $S(a_\mp)$. \medbreak Now, let $\mathcal{R}_2$ be the intersection of the following regions: \begin{itemize} \item[-] the strip $S_-$ defined above, \item[-] the annular region between the homoclinic $H(a_+)$ and the closed orbit $\Theta(a_+,(q_+,0))$ of $S(a_+)$, for some $q_+$ such that \begin{equation}\label{ordine} p_- < q_+ < 1, \end{equation} \item[-] the upper half-plane. \end{itemize} Straightforward computations show that the homoclinic $H(a_+)$ is contained in the region bounded by the stable and unstable manifolds $H^\pm (a_-)$. Moreover, since $a_+ < p_- < q_+$ (recall both \eqref{ordine0} and \eqref{ordine}) the closed orbit $\Theta(a_+,(q_+,0))$ winds around the point $(p_-,0)$, as well. These facts together guarantee that $\mathcal{R}_2$ is a topological rectangle. \smallbreak In a similar way, we can define the rectangle $\mathcal{R}_4$ as the intersection of the following regions: \begin{itemize} \item[-] the strip $S_+$ defined above, \item[-] the annular region between the homoclinic $H(a_-)$ and the closed orbit $\Theta(a_-,(q_-,0))$ of $S(a_-)$, for some $q_-$ such that \begin{equation}\label{ordine1} 0 < q_- < p_+, \end{equation} \item[-] the lower half-plane. \end{itemize} Notice that the rectangles $\mathcal{R}_2$ and $\mathcal{R}_4$ depend also on the choice of the numbers $q_{\pm}$ satisfying \eqref{ordine} and \eqref{ordine1} which is not arbitrary and will be specified in Proposition \ref{stretching}. \smallbreak We can again orientate the above constructed rectangles, denoting by $\mathcal{R}_2^-$ the components of the boundary of $\mathcal{R}_2$ lying on orbits of the systems $S(a_+)$ and by $\mathcal{R}_2^+$ the remaining components; finally, $\mathcal{R}_4^-$ are the components of the boundary of $\mathcal{R}_4$ lying on orbits of the systems $S(a_-)$ and $\mathcal{R}_4^+$ are the remaining components. \medbreak We illustrate the whole construction of the rectangles $\mathcal{R}_i$, for $i=1,\ldots,4$ in Figure \ref{figrett}. \begin{figure}[!h] \centering \includegraphics[height=10.5cm,width=12.5cm]{R2.png} \caption{\small{The construction of the topological rectangles $\mathcal{R}_i$, for $i=1,\ldots,4$; in particular, we comment in detail the construction of $\mathcal{R}_1$ and $\mathcal{R}_2$. Consider the stable manifold (to $(1,0)$) $H^+(a_-)$ (in red), as well as the homoclinic (to $(0,0)$) $H(a_-)$ (in pink) and recall that such an orbit intersects the positive $x$-semiaxis at the point $(z_{a_-},0)$. The condition $p_- > z_{a_-}$ (see \eqref{ordine0}) then guarantees that the piece of orbit of $\Theta (a_-,(p_-,0))$ (also in red) in the upper half-plane lies between $H(a_-)$ and $H^+(a_-)$. With an analogous construction for the system $S(a_+)$, and taking into account that $p_+ < 1/2 < p_-$ (see \eqref{ordine0} again) we can thus determine the rectangle $\mathcal{R}_1$ (in orange) in the upper-half plane (as well as the rectangle $\mathcal{R}_3$ (in purple) in the lower half-plane. Now, consider the homoclinic (to $(1,0)$) $H(a_+)$ (in black), intersecting the positive $x$-semiaxis at the point $(z_{a_+},0)$. It is easy to check that this orbit is contained in the region bounded the stable and unstable manifold $H^\pm (a_-)$; moreover, the condition $z_{a_+} < a_+ < p_-$ (see \eqref{ordine}) implies that it intersects the orbit $\Theta (a_-,(p_-,0))$. Finally, focus on the orbit $\Theta(a_+,(q_+,0))$ (also in black). Since $a_+ < p_- < q_+$ (see both \eqref{ordine0} and \eqref{ordine}), such an orbit intersects $\Theta (a_-,(p_-,0))$ as well (that is, the closed orbit $\Theta(a_+,(q_+,0))$ winds around the point $(p_-,0)$). This determines the rectangle $\mathcal{R}_2$ (in green). An analogous construction gives $\mathcal{R}_4$ (in grey). Recall that, for the validity of the stretching properties in Proposition \ref{stretching}, $q_{\pm}$ cannot be arbitrary numbers satisfying \eqref{ordine} and \eqref{ordine1}, but they have to fulfill further conditions (see the proof).}} \label{figrett} \end{figure} We are now in position to prove our crucial result on the stretching: \begin{proposition}\label{stretching} The following stretching properties hold true. \begin{itemize} \item[1.] There exists $T_1^* > 0$ such that, for every $T_1 > T_1^*$, we have $$ \Psi_-^{T_1}: (\mathcal{R}_1,\mathcal{R}_1^-) \,\Bumpeq{\!\!\!\!\!\!\!\!{\longrightarrow}}\, (\mathcal{R}_2,\mathcal{R}_2^-), \qquad \Psi_-^{T_1}: (\mathcal{R}_2,\mathcal{R}_2^+) \,\Bumpeq{\!\!\!\!\!\!\!\!{\longrightarrow}}\, (\mathcal{R}_3,\mathcal{R}_3^-), $$ $$ \Psi_+^{T_1}: (\mathcal{R}_3,\mathcal{R}_3^-) \,\Bumpeq{\!\!\!\!\!\!\!\!{\longrightarrow}}\, (\mathcal{R}_4,\mathcal{R}_4^-), \qquad \Psi_+^{T_1}: (\mathcal{R}_4,\mathcal{R}_4^+) \,\Bumpeq{\!\!\!\!\!\!\!\!{\longrightarrow}}\, (\mathcal{R}_1,\mathcal{R}_1^-), $$ for a suitable choice of $q^+$ and $q^-$ (close enough to $1$ and to $0$, respectively). \item[2.] For any $N \in \mathbb{N}$, there exists $T_2^*(N) > 0$ such that, for every $T_2 > T_2^*(N)$, we have $$ \Psi_+^{T_2}: (\mathcal{R}_2,\mathcal{R}_2^-) \,\Bumpeq{\!\!\!\!\!\!\!\!{\longrightarrow}}^N\, (\mathcal{R}_2,\mathcal{R}_2^+), \qquad \Psi_-^{T_2}: (\mathcal{R}_4,\mathcal{R}_4^-) \,\Bumpeq{\!\!\!\!\!\!\!\!{\longrightarrow}}^N\, (\mathcal{R}_4,\mathcal{R}_4^+). $$ \end{itemize} \end{proposition} \begin{proof} 1. We give the details for the stretching property $$ \Psi_-^{T_1}: (\mathcal{R}_1,\mathcal{R}_1^-) \,\Bumpeq{\!\!\!\!\!\!\!\!{\longrightarrow}} \, (\mathcal{R}_2,\mathcal{R}_2^-), $$ the other three statements of the first group being analogous. We first observe that the time needed to cover the piece of the orbit $\Theta(a_-,(p_-,0))$ contained in the right half-plane is given by $$ \sqrt{2} \int_0^{p_-} \frac{dx}{\sqrt{F_{a_-}(p_-)-F_{a_-}(x)}}; $$ similarly, the time needed to cover the piece of the orbit $\Theta(a_+,(p_+,0))$ contained in the right half-plane is given by $$ \sqrt{2} \int_{p_+}^1 \frac{dx}{\sqrt{F_{a_+}(p_+)-F_{a_+}(x)}}. $$ Now, let $$ T_1^* = \sqrt{2} \max\left\{\int_0^{p_-} \frac{dx}{\sqrt{F_{a_-}(p_-)-F_{a_-}(x)}}, \int_{p_+}^1 \frac{dx}{\sqrt{F_{a_+}(p_+)-F_{a_+}(x)}}\right\} $$ and fix $T_1 > T_1^*$. Moreover, we define $$ q_+ = \sup \left\{ x \in [p_-,1] : \, (\Psi_-^{T_1})^{-1}(x,0) \in \mathcal{R}_1 \right\}. $$ Let us observe that the above set is non-empty since, in view of the choice of $T_1$, $x \mapsto (\Psi_-^{T_1})^{-1}(x,0)$ ($x \in [p_-,1]$) parameterizes a curve joining a point in the second quadrant with the point $(1,0)$, whose intersection in the first quadrant is contained in $S^-$. By a continuity argument, moreover, we can easily prove that $p_- < q_+ < 1$. In this manner, \eqref{ordine} is satisfied and this completely determines the rectangle $\mathcal{R}_2$. This choice of $ q_{+} $ helps in controlling the behavior of the solutions which move according to system $ S( a_{-} ) $ on a time interval of length $ T_{1} $ and either start from $ \mathcal{R}_{1} $ or arrive on $ \mathcal{R}_{3} $. In fact, such solutions evolve inside the strip $ S^{-} $ and, by the choice of $ q_{+} $, they can cross the positive $ x $ axis only at a point lying between $ ( p_{-}, 0 ) $ and $ ( q_{+}, 0 ) $. Similarly, we define \[ q_- = \inf \left\{ x \in [ 0, p_{+} ] : \, (\Psi_{+}^{T_1})^{-1}(x,0) \in \mathcal{R}_3 \right\}. \] in order to complete the definition of $ \mathcal{R}_{4} $. \vs{6} \noindent We now pass to the verification of the stretching property, according to Definition \ref{stretchdef}. Let $\gamma: [0,1] \to \mathcal{R}_1$ be a path such that $\gamma(0)$ and $\gamma(1)$ lies on different components of $\mathcal{R}_1^-$ and, just to fix the ideas, suppose that $\gamma(1)$ lies on the stable manifold $H^+(a_-)$. \noindent We observe that $\Psi_-^{T_1}(\gamma(1))$ remains on $H^+(a_-)$ in the first quadrant while, by the choice of $T_1$, $\Psi_-^{T_1}(\gamma(0))$ lies in the third quadrant; hence, by the positive invariance of the strip $S_-$ for the flow of $S(a_-)$, we have $$ \Psi_-^{T_1}(\gamma([0,1])) \cap \mathcal{R}_2 \neq \emptyset. $$ On the other hand, by the choices of $q_+$ and $T_1$, $\Psi_-^{T_1}(\gamma([0,1]))$ cannot intersect the boundary of $\mathcal{R}_2$ on the $x$-axis. As a consequence, it intersects the two components of $\mathcal{R}_2^-$, say $\mathcal{R}_{2,u}^-$ and $\mathcal{R}_{2,d}^-$ ($d$ means ``down'' and $u$ means ``up'', with obvious meaning). Now, let $$ t_\gamma^- = \sup \left\{ t \in [0,1] : \, \Psi_-^{T_1}(\gamma(t)) \in \mathcal{R}_{2,d}^- \right\} $$ and $$ t_\gamma^+ = \inf \left\{ t > t_\gamma^- : \, \Psi_-^{T_1}(\gamma(t)) \in \mathcal{R}_{2,u}^- \right\}. $$ By construction, the sub-path $\gamma_1 = \Psi_-^{T_1} \circ \gamma|_{[t_\gamma^-,t_\gamma^+]}$ has image contained in $\mathcal{R}_2$ and crosses it from $\mathcal{R}_{2,d}^-$ to $\mathcal{R}_{2,u}^-$, as desired. \vs{12} \noindent 2. We now turn our attention to the second group of statements, proving that $$ \Psi_+^{T_2}: (\mathcal{R}_2,\mathcal{R}_2^-) \,\Bumpeq{\!\!\!\!\!\!\!\!{\longrightarrow}}^N\, (\mathcal{R}_2,\mathcal{R}_2^+) $$ (the other one being analogous); we observe that here $\mathcal{R}_2$ appears with different orientations when considered as the domain or the target space of $\Psi_+^{T_2}$. We introduce the polar coordinate system, centered at $(a_+,0)$, $$ x = a_+ + r \cos \theta, \qquad y = -r \sin \theta; $$ then, for any $z \in \mathcal{R}_2$, we denote by $\theta(\cdot;z)$ the (unique) continuous angular function associated to the solution of $S(a_+)$ starting from $z$ at the time $t=0$, and such that $\theta(0;z) \in [-\pi,0]$. Moreover, we recall that the time needed by any solution to system $S(a_+)$ starting from a point on $\mathcal{R}_{2,d}^-$ to perform one turn around $(a_+,0)$ can be computed as $$ \sqrt{2} \int_{\eta^+}^{q^+} \frac{dx}{\sqrt{F_{a^+}(q^+)-F_{a^+}(x)}}, $$ where $\eta^+ \in (0,q^+)$ is the only number such that $F_{a^+}(\eta^+) = F_{a^+}(q^+)$. \vs{6} \noindent Given the natural number $N \geq 1$, we define $$ T_2^*(N) = (N+1)\sqrt{2} \int_{\eta^+}^{q^+} \frac{dx}{\sqrt{F_{a^+}(q^+)-F_{a^+}(x)}} $$ and let $T_2 > T_2^*$. For any $j=1,\ldots,N$ we set $$ \mathcal{H}_j = \{ z \in \mathcal{R}_2 : \, \theta(T_2,z) \in [(2j-1)\pi,2j\pi] \}; $$ let us observe that the sets $\mathcal{H}_1,\ldots,\mathcal{H}_N$ are compact and disjoint. \vs{6} \noindent Now, let $\gamma: [0,1] \to \mathcal{R}_2$ a path such that $\gamma(0)$ and $\gamma(1)$ lie on different components of $\mathcal{R}_2^-$ and, to fix the ideas, assume that $\gamma(1)$ lies on $\mathcal{R}_{2,u}^-$ (that is, on the stable manifold to $(1,0)$). Hence, $\theta(T_2,\gamma(1)) < 0$ while, in view of the choice of $T_2$, $\theta(T_2,\gamma(0)) > 2N\pi$. As a consequence, we can find $N$-disjoint subintervals $I_1,\ldots,I_N \subset [0,1]$ such that $\gamma(s) \in \mathcal{H}_j$ for any $s \in I_j$ and $\Psi_+^{T_2} \circ \gamma|_{I_j}$ crosses $\mathcal{R}_2$ from one component of $\mathcal{R}_{2}^+$ to the other component of $\mathcal{R}_{2}^+$. \end{proof} \begin{remark} \label{nodali} \textnormal{We observe that the stretching relationships proved in point 2. of Proposition \ref{stretching} naturally provide information about some nodal properties of the solutions of the systems $S(a_{\pm})$. In particular, the (clockwise) winding number around the point $(a_+,0)$ of solutions of $S(a_+)$ starting in $\mathcal{H}_j$ (for some $j=1, \ldots, N$), in a time interval of length $T_2$, lies in the interval $(j - 1/2,j + 1/2)$ and, more precisely, the first derivative of those solutions vanishes exactly $ 2 j $ times. Similar considerations holds of course for solutions to $S(a_-)$ around the point $(a_-,0)$. With a slight abuse of terminology, in our main results we will say that, in this case, solutions make $j$ turns around $(a_{\pm},0)$.} \end{remark} \begin{figure}[!h] \centering \includegraphics[height=6cm,width=13cm]{LTM.png} \caption{\small{On the left, the construction of the rectangles when $a_- = 0.4$ and $a_+ = 0.6$, so that the homoclinics intersect (this is indeed the same picture as in Figure \ref{figrett})}. On the right, the construction of the rectangles in a situation in which the homoclinics do not intersect; here $a_- = 0.3$ and $a_+ = 0.7$. One can immediately realize that nothing changes. However, on the left again, it is shown how the linking between the homoclinics provides two further topological rectangles, painted in yellow. It is well-known that such rectangles verify stretching properties, within a Linked Twist Maps framework (see \cite{MarRebZan10}).} \label{figltm} \end{figure} \begin{remark}\label{LTM} \textnormal{We observe that the construction of the rectangles $\mathcal{R}_i$ ($i=1,\ldots,4$) given in this section is independent from the fact that the homoclinics $H(a_-)$ or $H(a_+)$ intersect, that is, we are not assuming that $z_{a_+} \leq z_{a_-}$ (see Figure \ref{figltm}). However, it is worth noticing that when the strict inequality \begin{equation}\label{link} z_{a_+} < z_{a_-} \end{equation} holds true, one can easily construct two further rectangles satisfying stretching conditions. This is indeed a well-known geometrical configuration, related to the concept Linked Twist Map, which has been extensively discussed in \cite{MarRebZan10}. We refer again to Figure \ref{figltm}. With elementary computations, it is possible to show that condition \eqref{link} is satisfied when $a_{\pm}$ are not too far from the value $1/2$.} \end{remark} \subsection{Stable and unstable manifolds} In this section we prove the existence of stable and unstable manifolds to the equilibria of a nonautonomous planar system. Our result will be given in a slightly more general setting than needed; more precisely, we deal with the planar system \begin{equation}\label{sysq} \left\{ \begin{array}{l} \vspace{0.1cm} x' = y \\ y'= x(x-1)(x-q(t)), \end{array} \right. \end{equation} with $q: \mathbb{R} \to \mathbb{R}$ a locally integrable function. We use the notation $z(t;p,t_0)$ for the (unique) solution to \eqref{sysq} satisfying the condition $z(t_0) = p$. \begin{proposition}\label{waze} The following statements hold true. \begin{enumerate} \item Assume that $$ 0 < a_- \leq q(t) \leq a_+ < 1, \quad \mbox{ for a.e. } t \in (-\infty,t_0], $$ for some $t_0 \in \mathbb{R}$. Then there exist two continua (i.e., connected compact sets) $\Gamma_{-\infty}^0, \Gamma_{-\infty}^1 \subset \mathbb{R}^2$, lying between the unstable manifolds to $(0,0)$ and $(1,0)$ of the systems $S(a_-)$ and $S(a_+)$ respectively, and such that, for any $p \in \Gamma_{-\infty}^i$, it holds that $$ \lim_{t \to -\infty} z(t;p,t_0) \to (i,0), \quad \mbox{ for } i =0,1. $$ Moreover, there exist two constants $a_-^0 \in (0,a_-)$ and $a_+^1 \in (a_+,1)$ such that $$ \Gamma_{-\infty}^0 \cap \left([0,a_-^0] \times \mathbb{R} \right) = \{(x,y_{-\infty}^0(x)) : \, x \in [0,a_-^0] \} $$ and $$ \Gamma_{-\infty}^1 \cap \left([a_+^1,1] \times \mathbb{R} \right) = \{(x,y_{-\infty}^1(x)) : \, x \in [a_+^1,1] \}, $$ for suitable continuous functions $y_{-\infty}^0:[0,a_-^0] \to \mathbb{R}^+$ and $y_{-\infty}^1:[a_+^1,1] \to \mathbb{R}^-$. \item Assume that $$ 0 < a_- \leq q(t) \leq a_+ < 1, \quad \mbox{ for a.e. } t \in [t_0,+\infty), $$ for some $t_0 \in \mathbb{R}$. Then there exist two continua (i.e., connected compact sets) $\Gamma_{+\infty}^0, \Gamma_{+\infty}^1 \subset \mathbb{R}^2$, lying between the stable manifolds to $(0,0)$ and $(1,0)$ of the systems $S(a_-)$ and $S(a_+)$ respectively, and such that, for any $p \in \Gamma_{+\infty}^i$, it holds that $$ \lim_{t \to +\infty} z(t;p,t_0) \to (i,0), \quad \mbox{ for } i =0,1. $$ Moreover, there exist two constants $a_-^0 \in (0,a_-)$ and $a_+^1 \in (a_+,1)$ (the same ones as in the previous statement) such that $$ \Gamma_{+\infty}^0 \cap \left([0,a_-^0] \times \mathbb{R} \right) = \{(x,y_{+\infty}^0(x)) : \, x \in [0,a_-^0] \} $$ and $$ \Gamma_{+\infty}^1 \cap \left( [a_+^1,1] \times \mathbb{R} \right) = \{(x,y_{+\infty}^1(x)) : \, x \in [a_+^1,1] \}, $$ for suitable continuous functions $y_{+\infty}^0:[0,a_-^0] \to \mathbb{R}^-$ and $y_{+\infty}^1:[a_+^1,1] \to \mathbb{R}^+$. \end{enumerate} \end{proposition} \begin{proof} We observe that Statement 2 is a consequence of Statement 1 and the symmetry enjoyed by the vector field in \eqref{sysq} with respect to the axis $ y= 0 $. Hence we have that $ \Gamma^{i}_{+\infty} = \{ (x,y) : (x,-y) \in \Gamma^{i}_{-\infty}\} $ and $ y^{i}_{+\infty} \equiv -y^{i}_{-\infty} $ for $ i = 0, 1 $. Concerning the proof of Statement 1, we give only the details about the continuum $\Gamma_{-\infty}^0$, since the existence and the properties of $ \Gamma_{-\infty}^{1} $ can be deduced in an analogous way. Let us fix $\epsilon > 0$ such that $$ 0 < a_- -\epsilon < a_+ +\epsilon < 1 $$ and define $\Sigma_\epsilon \subset \mathbb{R}^2$ as the compact triangular region of the first quadrant of the phase-plane bounded by the unstable manifolds to $(0,0)$ for the systems $S(a_- -\epsilon)$ and $S(a_+ + \epsilon)$ and the vertical line $x = a_- - \epsilon$. \vs{4} \noindent We first prove that any solution $z(t) = (x(t),y(t))$ which remains in $\Sigma_\epsilon$ for every $t \leq t_0$ has to satisfy $$ \lim_{t \to -\infty} z(t) = (0,0). $$ Indeed, for such a solution we have that, for $t \leq t_0$, $$ x(t) \in \left( 0, a_- - \epsilon \right], \qquad x'(t) > 0, \qquad x''(t) > 0, $$ and, hence, there exist $$ \lim_{t \to -\infty} x(t) = L < a_- \quad \mbox{ and } \quad \lim_{t \to -\infty} x'(t) = 0. $$ Finally, $L$ must be zero; otherwise: \[ \liminf_{t \to -\infty}x''(t) = \liminf_{ t \to -\infty } \{ x(t) [ 1 - x(t) ][ a(t) - x(t) ] \} \ge L ( 1 - a_{-} ) \epsilon > 0 \] and, thus, we would have $\lim_{t \to -\infty} x'(t) \neq 0$. \vs{6} \noindent Now we write system \eqref{sysq} as an autonomous system in $\mathbb{R}^3$: \begin{equation}\label{sysr3} \left\{ \begin{array}{l} \vspace{0.1cm} x' = y \\ \vspace{0.1cm} y' = x (x-1)(x-q(t)) \\ t' = 1 \end{array} \right. \end{equation} and let $\pi(\cdot;s_0,P)$ be the unique solution to \eqref{sysr3} starting from $ P \in \mathbb{R}^{3}$ at $s = s_0$. We set $$ W = \Sigma_\epsilon \times (-\infty,t_0], \quad U = \left(\gamma_- \setminus \{(0,0)\} \right) \times (-\infty,t_0], \quad V = \left(\gamma_+ \setminus \{(0,0)\} \right) \times (-\infty,t_0], $$ where $\gamma_-$ and $\gamma_+$ are the portions of the boundary of $\Sigma_\epsilon$ lying on the unstable manifolds to $(0,0)$ for the systems $S(a_- -\epsilon)$ and $S(a_+ + \epsilon)$ respectively. \vs{4} \noindent Let us study the behavior of the vector field associated with system \eqref{sysr3} at any point $P = (x_1,y_1,t_1) \in \partial W$. First, if $(x_1,y_1) = (0,0)$, it is clear that $\pi(s;t_1,P) = (0,0,s)$ for all $s$. Next, if $P \in U \cup V$, the vector field points strictly inwards $W$; therefore, $\pi(s;t_1,P) \notin W$ for all $s$ in a left neighborhood of $t_1$. Finally, if either $$ (x_1,y_1) \in \partial\Sigma_\epsilon \setminus \left( \gamma_- \cup \gamma_+ \cup \{(0,0)\}\right) $$ or \[ (x_1,y_1) \in \Sigma_\epsilon \setminus \left( \gamma_- \cup \gamma_+ \cup \{(0,0)\}\right) \quad \text{and} \quad t_{1}=0 \] then $\pi(s;t_1,P) \in W$ for all $s$ in a left neighborhood of $t_1$, since in those points the vector field of \eqref{sysr3} points strictly outwards $W$. \vs{2} \noindent Now, we consider the set $D \subset W$ given by $$ D = \{ P = (x_1,y_1,t_1) \in W : \, \exists s < t_1 \,\mbox{s.t.}\, \pi(s;t_1,P) \notin W \} $$ and the map $$ \Phi: D \to \partial W $$ such that $\Phi(P)$ is the first backward exit point from $W$ for the solution of \eqref{sysr3} starting from the point $P$, namely $ \Phi(P) = \pi( s^{*}; t_{1}, P ) $ where \[ s^{*} = \sup \{ s < t_{1} : \pi( s; t_{1}, P ) \not \in D \}. \] It is proved in \cite{Con75} that $\Phi$ is continuous on $D$; moreover, by the previous discussion, $\Phi(D) = U \cup V$ is not connected and $U$ and $V$ are its connected components. \vs{4} \noindent Let $\gamma: [0,1] \to \Sigma_\epsilon$ be a continuous path such that $\gamma(0) \in \gamma_- \setminus \{ (0,0) \}$ and $\gamma(1) \in \gamma_+ \setminus \{ (0,0) \}$. Then we have that $\Phi(\gamma(0)) \in U$, $\Phi(\gamma(1)) \in V$. Since $[0,1]$ is connected, there must be $\tau \in (0,1)$ such that $\gamma(\tau) \notin D$. By the topological lemma \cite[Corollary 6]{RebZan00} there exists a continuum $\Gamma_{-\infty}^0 \subset \Sigma_\epsilon \setminus D$ such that $$ (0,0) \in \Gamma_{-\infty}^0 \quad \mbox{ and } \quad \Gamma_{-\infty}^0 \cap \left( \{a_- - \epsilon \} \times \mathbb{R} \right) \neq \emptyset. $$ Letting $\epsilon \to 0^+$, it is actually possible to show that $\Gamma_{-\infty}^0 \subset \Sigma_0$ (that is, it lies between the unstable manifolds to $(0,0)$ and $(1,0)$ of the systems $S(a_-)$ and $S(a_+)$ respectively) and reaches the vertical line $x = a_-$ (see \cite[Theorem 6, \Sigma 47, II, p.171]{Kur68}). \vs{4} \noindent The fact that, in a suitable vertical strip, $\Gamma_{-\infty}^0$ can be written as the graph of a continuous function can be proved as in \cite[Lemma 2.4]{Ure10}, taking into account the fact that, for $$ f(t,x) = x(1-x)(x-q(t)) \quad \mbox{ and } \quad a_-^0 = \frac{1+a_- - \sqrt{a_-^2-a_-+1}}{3}, $$ it holds $\partial_x f(t,x) \geq 0$ for a.e. $t \leq t_0$ and $x \in [0,a_-^0].$ \end{proof} \begin{remark} \label{stablemanif} \textnormal{We observe that the mere existence of the stable and unstable manifolds follows from the Stable Manifold Theorem (see\cite{Hal80}), since $(0,0)$ and $(1,0)$ are hyperbolic equilibrium points of system \eqref{sysq}. Here, we have used Wa\.zewski's method in order to provide, for such sets, a precise localization, which is indeed needed in our arguments. Notice that in this way we cannot directly claim that stable/unstable manifolds are curves (indeed, Wa\.zewski's method applies in more general cases in which this is not true); however, we can recover such an information in an elementary way, using the sign of the nonlinearity as in \cite{Ure10}.} \end{remark} \section{Main results}\label{sec4} \def5.\arabic{equation}{4.\arabic{equation}}\makeatother \setcounter{equation}{0} Let us consider the equation \begin{equation}\label{eqmain} \epsilon^2 u'' + f(s,u) = 0 \end{equation} where \begin{equation}\label{fmodificata} f(s,u) = \left\{\begin{array}{ll} \vspace{0.1cm} u(1-u)(u-a(s)) & \mbox{if } u \in [0,1], \\ 0 & \mbox{if } u \notin [0,1], \end{array} \right. \end{equation} with $a: \mathbb{R} \to \mathbb{R}$ a locally integrable function satisfying $$ 0 < a(s) < 1, \quad \mbox{ for every } s \in \mathbb{R}. $$ \begin{remark}\label{fmodif} \textnormal{We observe that, in view of the boundedness of $f$, any solution to \eqref{eqmain} is globally defined. Moreover, if a solution $u$ satisfies $u(s^*) \notin (0,1)$ for some $s^* \in \mathbb{R}$, then $u(s) \notin (0,1)$ either for all $s \leq s^*$ or for all $s \geq s^*$.} \end{remark} From now on, we assume \begin{equation} \label{definizionea} a(s) = \left\{\begin{array}{ll} \vspace{0.1cm} a_- & {\mbox{if $s_{2k} \leq s < s_{2k+1}$}} \\ a_+ & {\mbox{if $s_{2k+1} \leq s < s_{2k+2}$ }}, \end{array} \right. \end{equation} where $a_-,a_+$ are real constants with \begin{equation} \label{condizionea} 0 < a_- < \frac{1}{2} < a_+ < 1 \end{equation} and $(s_k)_{k \in \mathbb{Z}} \subset \mathbb{R}$ is a sequence such that: \begin{equation}\label{lowbound} 0 < \delta \leq \inf_{ k \in \mathbb{Z} } ( s_{k+1} - s_k ) \end{equation} and, thus, $s_k < s_{k+1}$ for every $k\in \mathbb{Z}$ and $\lim_{k \to \pm \infty}s_k = \pm \infty$. For every $j\in \mathbb{Z}$ let us set $$ I_j = [s_{6(j-1)},s_{6j}], \quad I_j^+ = [s_{6j-5},s_{6j-4}], \quad I_j^- = [s_{6j-2},s_{6j-1}]. $$ In this setting, we will prove the following results. \smallbreak The first one, Theorem \ref{chaos}, deals with chaotic solutions. \begin{Theorem}\label{chaos} For any integer $M \geq 1$, there exists $\epsilon^* = \epsilon^*(M) > 0$, such that, for any $\epsilon \in (0,\epsilon^*)$ and for any double infinite sequence ${\bf n} = \{(n_j^{+},n_j^-)\}_{j\in \mathbb{Z}}$, with $ n_j^{\pm} \in \{ 1, \dots, M \} $ for all $ j \in \mathbb{Z} $, there exists a globally defined solution $ u $ of \eqref{eqmain}, with $0 < u(s) < 1$ for all $s \in \mathbb{R}$, such that its trajectory $ ( u, u' ) $ makes $n_j^{\pm} $ turns around $ ( a_{\pm}, 0 ) $ in the interval $ I_j^{\pm} $, for all $ j \in \mathbb{Z} $. Moreover, if the sequence $ ( s_k - s_{k-1} ) $ is $ 6 $-periodic and the sequence $ {\bf n}$ is $ \ell $-periodic for some $\ell \in \mathbb{N}$, then this solution $ u $ is $\ell(s_6-s_0)$-periodic. \end{Theorem} Notice that the sentence ``$n_j^{\pm}$ turns around $ ( a_{\pm}, 0 ) $'' in the above statement has to be meant according to Remark \ref{nodali}; the same terminology will be used in Theorems \ref{eter} and \ref{omo} below, dealing with heteroclinic and homoclinic solutions, respectively. \begin{Theorem}\label{eter} For any integer $M \geq 1,$ there exists $\epsilon^* = \epsilon^*(M) > 0$, such that, for any $\epsilon \in (0,\epsilon^*)$, for any integer $K \geq 0$ and for any ${\bf n} = \{(n_j^{+},n_j^-)\}_{j=1,\ldots,K}$ (this $K$-uple is considered to be empty if $K = 0$) with $ n_j^{\pm} \in \{ 1, \dots, M \} $ for all $ j =1,\ldots,k$, there exists a globally defined solution $u$ of \eqref{eqmain}, with $0 < u(s) < 1$ for all $s \in \mathbb{R}$, such that: \begin{enumerate} \item $(u(-\infty),u'(-\infty)) = (0,0)$, \item $(u(+\infty),u'(+\infty)) = (1,0)$, \item the trajectory $ ( u, u' ) $ makes $n_j^{\pm} $ turns around $ ( a_{\pm}, 0 ) $ in the interval $ I_j^{\pm} $, for all $ j =1,\ldots,K $, \item $u$ is monotone in $(-\infty,s_0]$ and in $[s_{6K},+\infty)$. \end{enumerate} \end{Theorem} \begin{Theorem}\label{omo} For any integer $M \geq 1,$ there exists $\epsilon^* = \epsilon^*(M) > 0$, such that, for any $\epsilon \in (0,\epsilon^*)$, for any integer $K \geq 0$ and for any ${\bf n} = \{(n_j^{+},n_j^-)\}_{j=1,\ldots,K}$ (this $K$-uple is considered to be empty if $K = 0$) with $ n_j^{\pm} \in \{ 1, \dots, M \} $ for all $ j =1,\ldots,k$, there exists a globally defined solution $u$ of \eqref{eqmain}, with $0 < u(s) < 1$ for all $s \in \mathbb{R}$, such that: \begin{enumerate} \item $(u(-\infty),u'(-\infty)) = (0,0)$, \item $(u(+\infty),u'(+\infty)) = (0,0)$, \item the trajectory $ ( u, u' ) $ makes $n_j^{\pm} $ turns around $ ( a_{\pm}, 0 ) $ in the interval $ I_j^{\pm} $, for all $ j =1,\ldots,K $, \item $u$ is monotone in $(-\infty,s_0]$ and in $[s_{6K+1},+\infty)$ and $u'$ vanishes exactly once in $(s_{6K},s_{6K+1})$. \end{enumerate} \end{Theorem} \subsection{Proof of the main results} In order to prove the above theorems, we perform the change of variable $$ x(t) = u(\epsilon t), \quad t \in \mathbb{R}, $$ converting the equation \eqref{eqmain} into \[ x'' + f(\epsilon t , x ) = 0, \] with $$ f(\epsilon t, x ) = \begin{cases} x (1-x)(x-a_{-} ) & \text{if } x \in [ 0, 1 ] \text{ and } t \in \left[ t_{2k}, t_{2k+1} \right), k \in \mathbb{Z} \\ x (1-x)(x-a_{+} ) & \text{if } x \in [ 0, 1 ] \text{ and } t \in \left[ t_{2k+1}, t_{2k+2} \right), k \in \mathbb{Z} \\ 0 & \text{if } x \not\in [ 0, 1 ] \end{cases} $$ where we have set $$ t_k = \frac{s_k}{\epsilon}, \quad \mbox{ for every } k \in \mathbb{Z}. $$ Notice that, in view of \eqref{lowbound}, we now have $$ 0 < \frac{\delta}{\epsilon} \leq \inf_{k\in\mathbb{Z}} ( t_{k+1} - t_k ). $$ Let us define, for any $k \in \mathbb{Z}$ $$ \mathcal{D}_k = \mathcal{P}_k = \mathcal{R}_1 \quad \mbox{ and } \quad \widetilde{\mathcal{P}_k} = (\mathcal{R}_1,\mathcal{R}_1^-) $$ and \begin{align*} \phi_k = & \, \Psi_+^{t_{6k+6}-t_{6k+5}} \circ \Psi_-^{t_{6k+5}-t_{6k+4}} \circ \Psi_+^{t_{6k+4}-t_{6k+3}} \circ \Psi_-^{t_{6k+3}-t_{6k+2}} \circ \\ & \circ \Psi_+^{t_{6k+2}-t_{6k+1}} \circ \Psi_-^{t_{6k+1}-t_{6k}}. \end{align*} In other words, $ \Phi_{k}(p) = ( x( t_{6k+6} ), x'( t_{6k+6} ) ) $ where $ x $ is the solution of $ x'' + f( \epsilon t, x ) = 0 $ with $ ( x( t_{6k} ), x'( t_{6k} ) ) = p $. \begin{proof}[Proof of Theorem \ref{chaos}] Given $M \geq 1$, let us define $\epsilon^*(M) > 0$ as \begin{equation}\label{epsm} \epsilon^*(M) = \frac{\delta}{\max\{T_1^*,T_2^*(M)\}}, \end{equation} where $T_1^*,T_2^*(M)$ are given by Proposition \ref{stretching}. Now we take $\epsilon \in (0,\epsilon^*)$; notice that we have $$ t_{k+1} - t_k > \max\{T_1^*,T_2^*(M)\}, \quad \mbox{ for every } k\in \mathbb{Z}. $$ In view of Propositions \eqref{stretching} and \eqref{composizione}, we have $$ \phi_k: \widetilde{\mathcal{P}_k} \,\Bumpeq{\!\!\!\!\!\!\!\!{\longrightarrow}}^{N^2}\, \widetilde{\mathcal{P}_k}. $$ The conclusion thus follows from Theorem \ref{thchaos} in the Appendix. We finally notice that $x(t_{6k}) \in (0,1)$ for all $k \in \mathbb{Z}$, so that, in view of Remark \ref{fmodif}, $x(t) \in (0,1)$ for all $t \in \mathbb{R}$. \end{proof} \begin{proof}[Proof of Theorem \ref{eter}] We begin by applying Proposition \ref{waze} on the intervals $ \left( -\infty, t_{-1} \right] $ and $ \left[ t_{6k+1}, +\infty \right) $ in order to find the continuous functions $ y_{-\infty}^{0} : [ 0, a_{-}^{0} ] \to \mathbb{R}^{+} $ and $ y_{+\infty}^{1} : [ a_{+}^{1}, 1 ] \to \mathbb{R}^{+} $ such that $ ( x, y_{-\infty}^{0}(x) ) \in \Gamma_{-\infty}^{0} $ and $ ( x, y_{+\infty}^{1}(x) ) \in \Gamma_{+\infty}^{1} $ for all suitable $ x $. We denote by $x^*$ the abscissa of the intersection point in the first quadrant between the homoclinic to $(0,0)$ for the system $S(a_-)$ and the orbit $\Theta(a_+,(p_+,0))$ of system $S(a^+)$. Let $$ x_1 = \min\{x^*,a_-^0\}. $$ The time needed by a solution of system $S(a_+)$ to run along $\Theta(a_+,(p_+,0))$ from $x=x_1$ to $x=1$ is given by $$ \tau = \frac{1}{\sqrt{2}}\int_{x_1}^1 \frac{dx}{\sqrt{F_{a_+}(p_+)-F_{a_+}(x)}}. $$ In a similar way, we take $x^{**}$ as the abscissa of the intersection point in the first quadrant between the homoclinic to $(1,0)$ for the system $S(a_+)$ and the orbit $\Theta(a_-,(p^-,0))$ of system $S(a^-)$. Let $$ x_2 = \max\{x^{**},a_+^1\}. $$ The time needed by a solution of system $S(a_-)$ to go on $\Theta(a_-,(p^-,0))$ from $x=0$ to $x=x_2$ is given by $$ \tau' = \frac{1}{\sqrt{2}}\int_{0}^{x_2} \frac{dx}{\sqrt{F_{a_-}(p^-)-F_{a_-}(x)}}. $$ For any integer $M \geq 1$, we can thus define \begin{equation}\label{epsm2} \epsilon^*(M) = \frac{\delta}{\max\{T_1^*,T_2^*(M),\tau,\tau'\}}, \end{equation} where again $ T_{1}^{*} $ and $ T_{2}^{*}(M) $ are given by Proposition \ref{stretching}. With these positions, we have that $$ \Psi_+^{t_0-t_{-1}}(\Gamma_{-\infty}^0 \cap ([0,x_1] \times \mathbb{R})) $$ is a path which crosses $\mathcal{R}_1$ connecting the two components of $\mathcal{R}_1^-$. \vs{2} \noindent Indeed, this is a consequence of the fact that $(0,0)$ is a fixed point for $\Psi_+^{t_0-t_{-1}}$ and that the point $(x_1,y_1) \in \Gamma_{-\infty}^0$ is mapped by $\Psi_+^{t_0-t_{-1}}$ in the half-plane $\{ x \geq 1\}$ which follows from the choice of $\tau$ and the monotonicity of the map \[ c \mapsto \int_{x_1}^1 \frac{dx}{\sqrt{c-F_{a_+}(x)}}. \] Analogously (by the choice of $\tau'$) $$ (\Psi_-^{t_{6k+1}-t_{6k}})^{-1}(\Gamma_{+\infty}^1 \cap ([0,x_1] \times \mathbb{R})) $$ is a path which crosses $\mathcal{R}_1$ connecting the two components of $\mathcal{R}_1^+$. \vs{4} \noindent Now, corresponding to the choice of ${\bf n} = \{(n_j^{+},n_j^-)\}_{j=1,\ldots,K}$ with $ n_j^{\pm} \in \{ 1, \dots, M \} $, there exists a sub-path $\gamma$ of $\Psi_+^{t_0-t_{-1}}(\Gamma_{-\infty}^0)$ which is stretched across $(\mathcal{R}_1,\mathcal{R}_1^-)$ by the map $\phi_k \circ \phi_{k-1} \circ \ldots \circ \phi_1$. \vs{4} \noindent By the topological lemma \cite[Lemma 3]{MulWil74} the intersection $$ (\phi_k \circ \phi_{k-1} \circ \ldots \circ \phi_1)(\gamma) \cap (\Psi_-^{t_{6k+1}-t_{6k}})^{-1}(\Gamma_{+\infty}^1) $$ is not empty and any point in this intersection gives rise to the required solution. \end{proof} \begin{proof}[Proof of Theorem \ref{omo}] One can argue in a similar way as in the proof of Theorem \ref{eter}, using here the curves $\Gamma_{-\infty}^0$ and $\Gamma_{+\infty}^0$ given in Lemma \ref{waze}. \end{proof} \begin{remark}\label{stabile} \textnormal{We notice that the statements of our results can be modified (in some cases) to cover also more general equations $\epsilon^2 u'' + u(1-u)(u-\tilde{a}(t)) = 0$ with $\tilde a$ close to a stepwise function. For instance, in the setting of Theorems \ref{eter} and \ref{omo} and given integers $M, K$ and $\epsilon \in (0,\epsilon^*(M))$, the existence result still holds for all functions $\tilde{a}$ with $L^1$-norm on $[0,s_{6K}]$ smaller than a constant $\delta = \delta(M,K,\epsilon)$ (compare with \cite[Remark 4.1]{MarRebZan10} for more details on the stability of the stretching technique, and recall that Lemma \ref{waze} about the existence of stable/unstable manifolds is indeed proved for a more general, non-stepwise, function). We also observe that the value $\epsilon^*(M)$ is an explicit constant (see \eqref{epsm} and \eqref{epsm2}).} \end{remark} \begin{remark}\label{linkinginterno} \textnormal{We collect here some hints for possible variants and generalizations of our results. \begin{itemize} \item The role of the equilibria $(0,0)$ and $(1,0)$ may be switched in Theorems \ref{eter} and \ref{omo}. More precisely, we can provide also solutions homoclinic to $(1,0)$ as well as heteroclinic solutions converging to $(1,0)$ for $t \to -\infty$ and to $(0,0)$ for $t \to +\infty$. Moreover, the arguments used in the proof of Theorem \ref{eter} can be easily modified in order to find multiple solutions for Sturm-Liouville like boundary values problems (e.g., Dirichlet and Neumann ones). \item According to the discussion in Remark \ref{LTM}, it is possible to produce a further family of chaotic solutions when condition \eqref{link} is satisfied. \item In our main results we have considered functions of the form \eqref{definizionea} with $a_-$ and $a_+$ satisfying condition \eqref{condizionea}, suggested by the investigation in \cite{AngMalPel87}. When either $a_+<1/2$ or $a_->1/2$, the superposition of the phase-portraits of the systems $S(a_+)$ and $S(a_-)$ gives rise to a different configuration (see Figure \ref{zaza}) analogous to the one considered in the paper \cite{ZaZa-14}, where a Schr\"{o}dinger equation is studied. As a consequence, by combining the arguments therein together with Lemma \ref{waze}, it is possible to obtain the existence of chaotic dynamics and homoclinic orbits. \end{itemize}} \end{remark} \begin{figure}[!h] \centering \includegraphics[height=8cm,width=11cm]{zaza.png} \caption{\small{The superposition of the phase-portraits of the systems $S(a_-)$ and $S(a_+)$, with $a_- = 0.3$ and $a_+ = 0.4$. Orbits of the system $S(a_-)$ are painted in pink, orbits of the system $S(a_+)$ are painted in blue, topological rectangles verifying stretching properties are painted in orange. Compare with Figure 2 in \cite{ZaZa-14}.}} \label{zaza} \end{figure} \section{Appendix: SAP method} \def5.\arabic{equation}{5.\arabic{equation}} \makeatother\setcounter{equation}{0} In this appendix we collect the definitions and results on the Stretching Along Paths method which are needed in our paper. We refer to \cite{Bur14,PasPirZan08} for a comprehensive presentation of the theory and further references. \medbreak By a path $\gamma$ in $\mathbb{R}^2$ we mean a continuous mapping $\gamma:[0,1] \to \mathbb{R}^2$, while by a sub-path $\sigma$ of $\gamma$ we just mean the restriction of $\gamma$ to a compact subinterval of $[0,1]$. By an \emph{oriented rectangle} we mean a pair $$ \mathcal{\widetilde R} = (\mathcal{R},\mathcal{R}^-), $$ being $\mathcal{R} \subset \mathbb{R}^2$ homeomorphic to $[0,1]^2$ (namely, a \textit{topological rectangle}) and $$ \mathcal{R}^- = \mathcal{R}^-_1 \cup \mathcal{R}^-_2 $$ the disjoint union of two compact arcs (by definition, a compact arc is a homeomorphic image of $[0,1]$) $\mathcal{R}^-_1, \mathcal{R}^-_2 \subset \partial \mathcal{R}$. \smallbreak With these preliminaries, we can give the following definition. \begin{Definition}\label{stretchdef} Let $\mathcal{\widetilde A} = (\mathcal{A},\mathcal{A}^-)$, $\mathcal{\widetilde B} = (\mathcal{B},\mathcal{B}^-)$ be oriented rectangles and let $\Psi: \mathcal{D}_{\Psi} \subset \mathbb{R}^2 \to \mathbb{R}^2$ be a continuous map. \begin{itemize} \item[-] We say that $(\mathcal{H},\Psi)$ stretches $\mathcal{\widetilde A}$ to $\mathcal{\widetilde B}$ along the paths and write $$ (\mathcal{H},\Psi): \mathcal{\widetilde A} \, \Bumpeq{\!\!\!\!\!\!\!\!{\longrightarrow}} \, \mathcal{\widetilde B} $$ if $\mathcal{H} \subset \mathcal{A} \cap \mathcal{D}_{\Psi}$ is a compact subset and for every path $\gamma: [0,1] \to \mathcal{A}$ such that $\gamma(0) \in \mathcal{A}^-_1$ and $\gamma(1) \in \mathcal{A}^-_2$ (or $\gamma(0) \in \mathcal{A}^-_2$ and $\gamma(1) \in \mathcal{A}^-_1$), there exists a subinterval $[t',t''] \subset [0,1]$ such that for every $t \in [t',t'']$ $$ \gamma(t) \in \mathcal{H}, \qquad \Psi(\gamma(t)) \in \mathcal{B}, $$ and, moreover, $\Psi(\gamma(t'))$ and $\Psi(\gamma(t''))$ belong to different components of $\mathcal{B}^-$. \item[-] We say that $\Psi$ stretches $\mathcal{\widetilde A}$ to $\mathcal{\widetilde B}$ along the paths with crossing number $M \geq 1$ and write $$ \Psi: \mathcal{\widetilde A} \,\Bumpeq{\!\!\!\!\!\!\!\!{\longrightarrow}}^M \, \mathcal{\widetilde B} $$ if there exist $M$ pairwise disjoint compact sets $\mathcal{H}_1,\ldots,\mathcal{H}_M \subset \mathcal{A} \cap \mathcal{D}_{\Psi}$ such that $(\mathcal{H}_i,\Psi): \mathcal{\widetilde A} \Bumpeq{\!\!\!\!\!\!\!\!{\longrightarrow}} \mathcal{\widetilde B}$ for $i=1,\ldots,M$. \end{itemize} \end{Definition} As an easy consequence of the definition, we have that the stretching property has a good behavior with respect to compositions of maps. \begin{proposition}\label{composizione} Assume that $\Psi$ stretches $\mathcal{\widetilde A}$ to $\mathcal{\widetilde B}$ with crossing number $M$ and $\Phi$ stretches $\mathcal{\widetilde B}$ to $\mathcal{\widetilde C}$ with crossing number $N$. Then, the composition $\Phi \circ \Psi$ stretches $\mathcal{\widetilde A}$ to $\mathcal{\widetilde C}$ with crossing number $M \times N$. \end{proposition} We finally state the result which is employed in our paper. \begin{Theorem}\label{thchaos} Assume that there are double sequences of oriented rectangles $\left(\mathcal{\widetilde P}_k \right)_{k \in \mathbb{Z}}$ and of maps $\left(\phi_k \right)_{k \in \mathbb{Z}}$ such that $\phi_k$ stretches $\mathcal{\widetilde P}_k$ to $\mathcal{\widetilde P}_{k+1}$ with crossing number $M_k \geq 1$, for all $k \in \mathbb{Z}$. Let $\mathcal{H}_{k,j} \subset \mathcal{P}_{k}$, with $j=1,\ldots,M_k$, be the compact sets according to the definition of multiple stretching. Then, the following conclusion hold: \begin{itemize} \item for every sequence $(s_k)_{k \in \mathbb{Z}}$, with $s_k \in \{1,\ldots,M_k\}$, there exists a sequence $(w_k)_{k \in \mathbb{Z}}$ with $w_k \in \mathcal{H}_{k,s_k}$ and $\phi_k(w_k) = w_{k+1}$ for all $k \in \mathbb{Z}$; \item if there exists $h,k \in \mathbb{Z}$ with $h <k$ such that $\mathcal{\widetilde P}_h = \mathcal{\widetilde P}_k$, then there is a finite sequence $(w_i)_{h \leq i \leq k}$ with $w_i \in \mathcal{H}_{i,s_i}$, $\phi_i(w_i) = w_{i+1}$ for $i=h,\ldots,k-1$ and $w_h = w_k$, that is, $w_h$ is a fixed point of $\phi_{k-1} \circ \ldots \circ \phi_h$. \end{itemize} \end{Theorem} \begin{proof} By assumption, we have that $$ \left(\mathcal{H}_{k,s_k},\phi_k \right) : \mathcal{\widetilde P}_{k}\,\Bumpeq{\!\!\!\!\!\!\!\!{\longrightarrow}} \, \mathcal{\widetilde P}_{k+1}, \quad \mbox{ for every } k \in \mathbb{Z}. $$ The conclusion follows then from \cite[Theorem 2.2]{PapZan04}. \end{proof} \begin{remark}\label{coniugio} \textnormal{In the setting of Theorem \ref{thchaos}, when $\mathcal{\widetilde P}_k = \widetilde{\mathcal{P}}$, $\phi_k = \phi$ and $M_k = M$ for all $k \in \mathbb{Z}$, it is possible to prove that there is a compact invariant set $\Lambda \subset \mathcal{P}$ such that the map $\phi|_{\Lambda}$ is topologically semiconjugate on the Bernoulli shift on $M$ symbols (see \cite[Lemma 2.3]{PasPirZan08}).} \end{remark}

https://arxiv.org/abs/1706.09454

Drowning by numbers: topology and physics in fluid dynamics

Since its very beginnings, topology has forged strong links with physics and the last Nobel prize in physics, awarded in 2016 to Thouless, Haldane and Kosterlitz " for theoretical discoveries of topological phase transitions and topological phases of matter", confirmed that these connections have been maintained up to contemporary physics. To give some (very) selected illustrations of what is, and still will be, a cross fertilization between topology and physics, hydrodynamics provides a natural domain through the common theme offered by the notion of vortex, relevant both in classical (§2) and in quantum fluids (§3). Before getting into the details, I will sketch in §1 a general perspective from which this intertwining between topology and physics can be appreciated: the old dichotomy between discreteness and continuity, first dealing with antithetic thesis, eventually appears to be made of two complementary sides of a single coin.

\section{The arena of the discrete/continuous dialectic} One century after Thales of Miletus had proposed that water was the natural principle of all things, the first atomists Leucippus and Democritus advocated for a discrete conception of matter. The existence of an ultimate lower limit of divisibility, materialised by the atoms, may have been a logical answer to the Zeno's paradoxes~(\citealt[chap.~VIII]{Stokes71a}; \citealt[chap.~I]{Bell06a}). In some westernmost banks of the Mediterranean sea, the Pythagorean school was concerned by a line of thought following quite an opposite direction: the discovery of the irrational numbers counterbalanced the conception of a universe exclusively driven by the integer and rational---in the original acception of the word---numbers. For twenty-five centuries, the dialectic between continuity and discreteness has never stopped nurturing natural philosophy. At our daily life scales, the ones for which the brains have been shaped by Darwinian evolution\footnote{In modern times physics and chemistry were not, by far, the only scientific disciplines to be shaken by violent debates between discrete and continuous schools; in the \textsc{xix}\th century Lyell's uniformitarianism in geology, by contrast with catastrophism, had an important influence on the young Darwin. By the way, one can notice that the binary opposition between discreteness and continuity provides by itself a meta self-referring epistemological dichotomy, so to speak.}, discreteness appears to be an inevitable way for intelligence to model the world\footnote{However, neurology shows that numerical cognition is more analogical than numerical: beyond few units, the numbers are encoded and treated by the brain as fuzzy entities~\cite[specially part~I and chap.~9]{Dehaene97b}. }. Furthermore, operationally speaking, any measurement is reduced, in the last resort, to a reproducible counting~\cite[\S\;1.1]{Thouless98a}. Etymologically, ``discrete'', ``critical'', ``criterion'', and ``discernment'' share the same greek root~$\kappa\rho\acute\iota\nu\omega$ (\textit{kr}$\acute{\text{\it\={\i}}}$\textit{n\=o}, to judge)\footnote{The etymology lines of these words can be easily traced back with \texttt{www.wiktionary.org}.}. However, the boundaries of macroscopic objects, considered both in space and time, remain inevitably blurred. For instance, consider one cherry; through absorption and desorption, a perpetual exchange of matter takes place at small scales on the skin of the cherry, and no one can really identify with a precision of one second the time when this cherry has appeared from a blossom or destroyed by natural deterioration\footnote{In a contribution to the previous volume of this series~\cite[\S\;5]{Mouchet15a}\nocite{Emmer+15a} I have tried to show how symmetries play a crucial role in the process of abstraction and conceptualisation of a macroscopic object like a cherry.}. This ambiguity was known from antiquity and supply the sorites paradox (what is the minimum number of grains in a heap of sand?)---and the paradox of the ship of Theseus (Plutarch asks if, after decades of restauration, once her last plank has been replaced, the ship remains the same Theseus's ship~\cite[The life of Theseus \S\;XXIII.1]{Plutarch14a}). In the second part of the \textsc{xix}\th century, experiments allowed to move the debate beyond speculations into the microscopic world. In the same movement, mathematics saw the emergence of a new discipline, topology, where were identified some \emph{discrete} classifications---first in geometry, then in analysis and algebra---up to \emph{continuous} invertible transformations (homeomorphisms). The integer numbers upon which the classes of, say, graphs, knots, surfaces, fixed points of a flow, critical points of a real map, are discriminated provide, by essence, a robust quantization; they are topological invariant. To put it in a nutshell, there cannot be ``half a hole''. The dimension of a space\footnote{In fractal geometry, the Hausdorff dimension of a set, which can be irrationnal, is not preserved by a homeomorphism. }, its connectedness ($\pi_0$), its homotopy groups ($\pi_1$, $\pi_2$ and more generally~$\pi_n$), the signature of the Hessian of a function at a critical point, are examples of such discrete quantities. In the beginning of the \textsc{xx}\th century, quantum physics refuted so masterfully the Leibniz continuity principle (\textit{Nature does not make jumps}) that it bears this claim in its very name. The general rule---known by Pythagoreans for music---according to which a stable wave in a bounded domain has its frequencies quantized (that is, function of integer numbers) now applied at a fundamental level to the Schr\"odinger waves, which described the states of elementary particles, when bounded. The discrete classification of chemical elements successfully proposed in 1869 by Mendeleev and the discrete spectral lines corresponding to the Balmer series, the Paschen series, the Lyman series etc. observed in radiation, could be explained within a unifying scheme offered by quantum theory. Eventhough it appears that each atomic energy level has actually a continuous bandwidth, due to the coupling to the electromagnetic field whose scattering states belong to a continuum (the photon has no mass), it is nevertheless quantum theory that confered to ``being an integer'' a genuine physical property. So far, neither the quantification of the spin nor the quantification of the electric charge, say, can be seen as an approximation of a continuous model and the analogous of the Mendeleiev table in the Standard Model contains a finite number of species of elementary particles---about twenty, non counting as distinct a particle from its associated antiparticle---characterised by a handful of quantum numbers\footnote{The discrete character of some observable properties is all the more strengthened that there exists some superselection rules that make irrelevant any continuous superposition of states differing by some discrete values of this observable. }. Many attempts have been made for finding a topological origin of these quantum numbers, one of the motivation being that topological invariance is much harder to break than symmetry invariance. In condensed matter, topology offers a protection against the effects of impurities or out-of-control perturbations and therefore participates to the reproductibility and the fiability of measurements~\cite[\S\;1.3]{Thouless98a}. The seminal attempt in this direction is Dirac's model of magnetic monopole~\citep{Dirac31a} whose existence would imply the quantization of the electric charge; however, so far, all the quantizations that have been explained find their root in \emph{algebraic} properties of the symmetry groups used to build a basis of quantum states\footnote{Topological properties of these Lie groups, obviously their dimensions but also their compactness, their connectedness and their simple connectivity, do play a role but the algebraic commutation relations of their generators remain the main characteristics, which are local ones, that allow to build the irreducible representations defining the one-particle states. } (in the absence of evidence of elementary magnetic monopoles, the fact that the electric charges appear to be always an integer multiple of one unit remains mysterious). Despite these (temporary?) failures of finding topological rather than algebraic roots for the discrete characteristics of what appears to be elementary particles, the quantum theory of fields offers the possibility of describing some collective effects of those particles whose stability is guaranteed by topological considerations. There exists some configurations of a macroscopic number of degrees of freedom that cannot be created or destroyed by a smooth transformation without passing through an intermediate state having a macroscopic, and therefore redhibitory, energy. Depending on the dimension of the space and of the field describing the model, several such \emph{topological defects} can be considered (point, lines or surfaces) and have been observed in various condensed states~\citep[chap.~9]{Chaikin/Lubensky95a} including, of course, the quantum fluids where the defects are characterised by quantized numbers that can be interpreted as topological invariants. Vortices, which will be the object of the next two sections, provide typical examples of such topological defects along a line in a 3-dimensional space or localised at one point in a 2-dimensional space for a complex scalar field (or a real bidimensional vector field). Under certain circumstances, these collective effects share many properties with the so-called ordinary particles. Since, theoretically, the distinction between the quasi-particles and particles appears, after all, to be just a matter of convention on the choice of the vacuum and of the particles that are considered to be elementary, one may have the secret hope that at a more fundamental level, having the Standard Model as an effective theory, topology shall have the next, but presumably not the last, word. \section{Classical vortices} \begin{flushright} \textit{\dots when I first opened my eyes upon the wonders of the whirlpool\dots}\\ Edgar Allan Poe. \emph{A Descent into the Maelstr\"om} (1841). \end{flushright} \subsection{How vortices participate to the dynamics of the world according to Leonardo and Descartes} By strong contrast with the still, rather mineral, backgrounds of his paintings, Leonardo da Vinci's interest for the dynamics of water is manifest in his drawings and writtings all along his life. Vortices in water, in air, and even in blood \cite[\S\;3.3]{Pasipoularides10a}, were a recurrent source of fascination for him\footnote{\citet{Gombrich69a}\nocite{OMalley69a} saw in the exuberance of the terms used by Leonardo and in the profusion of his drawings an attempt to classify the vortices, a line of investigations he kept in mind throughout his life.}. Not only as esthetical motifs (fig.~\ref{fig:leonardtourbillons}), not only because of their crucial role for understanding hydraulics and fly, not only because they inspired him fear as a disordered manifestation of flooding or deluge, but also because they provided a central key for his global conception of the dynamics of the world: \textit{l'acqua, vitale omore della terreste macchina, mediante il suo natural calore si move.} (water, vital humour of the terrestrial machine, moves by means of its natural heat)\footnote{Folio H95r, whose facsimile and transcription can be found on \texttt{www.leonardodigitale.com}.} \cite[Chap.~\textit{Une science en mouvement}]{Arasse97a}. \begin{figure}[!ht] \begin{center} \includegraphics[width=\textwidth]{leonardtourbillons.eps} \caption{\label{fig:leonardtourbillons} Left) folio~w12380, A D\'eluge, $\sim$1517-18. Center) folio~w12663r, Studies of flowing water, $\sim$1510-13. Right) folio~w12518, Head of Leda, $\sim$1504-1506. \texttt{Wikimedia Foundation}. On the folio~w12579r Leonardo has drawn four studies of vortex alleys formed in water behind a parallelepipedic obstacle and writes \textit{Nota il moto del livello dell'acqua, il quale fa a uso de' capelli, che hanno due moti, de' quali l'uno attende al peso del vello, l'altro al liniamento delle volte: cos\`{\i} l'acqua ha le sue volte revertiginose, delle quali una parte attende al impeto del corso principale, l'altra attende al moto incidente e refresso.} (Observe the movement of the surface of water, like hair which has two movements, one due to its weight, the other following the lines of the curls: thus water has whirling eddies, in part following the impetus of the main stream, in part following the incidental and reversed motion, folio w12579r, trad.~\textsc{am}). } \end{center} \end{figure} More than a century later, most probably without any influence from Leonardo, Descartes put the vortices in the very core of his cosmological model. Rejecting the atomist concept of a vacuum separating matter \cite[part~II, 16th principle]{Descartes1644a}, he writes \begin{quotation} \textit{[\dots] putandum est, non tantum Solis \& Fixarum, sed totius etiam coeli materiam fluidam esse.}\\ ([\dots] we think that not only the matter of the Sun and of the Fixed Stars is fluid but also is the matter of all the sky, trad. \textsc{am}) \flushright \cite[\S~III.24 p. 79]{Descartes1644a} \end{quotation} Being aware of the proper rotation of the Sun (it takes 26 days for the sunspots to complete one turn \cite[\S~III.32 p.~83]{Descartes1644a}) and of the different orbital period of the planets, he pursues further the hydrodynamical analogy \begin{quotation} \textit{[\dots] putemus totam materiam coeli in qua Planetae versantur, in modum cuiusdam vorticis, in cuius centro est Sol, assidue gyrare, ac eius partes Soli viciniores celerius moveri quam remotiores [\dots]}\\ ([\dots] we think that all the matter of the sky, in which the Planets turn, rotates like a vortex with the Sun at its center; that the parts near the Sun move faster than the remote ones [\dots], trad. \textsc{am}) \flushright \cite[\S~III.30 pp. 81-82]{Descartes1644a} \end{quotation} \begin{figure}[!ht] \begin{center} \includegraphics[width=7cm]{descartestourbillons.eps} \caption{\label{fig:descartestourbillons} Descartes'vortex-based cosmology. Each star denotes by F, D, etc. is at the center of a vortex. The Sun is denoted by S \protect\cite[\S~III.23 p.~78]{Descartes1644a}. } \end{center} \end{figure} Descartes'model was overuled by Newton's theory planetary motion but, somehow, in contemporary astrophysics, vortices are still present---in a complete different way, of course, from Descartes'--- and triggered by gravitational field acting through the interstellar vacuum: one may think of protoplanetary accretion disks (turbulence plays a crucial role, in particular in the initial molecular cloud for explaining the scattered births of stars) and, at much larger scales, of galaxies, cosmic whirlpools spinning around a giant black hole. \subsection{Accompanying the birth of topology in the \textsc{xix}\th century} His study of the physical properties of organ pipes led Helmholtz to scrutinize the motion of the air near sharp obstacles and the influence of viscosity. The memoir he published in German in 1858 on the subject had a decisive influence on the physicists of the Scottish school including Maxwell, Rankine, Tait and Thomson (who was ennobled in 1892 as Lord Kelvin), all the more that Tait translated it into English in 1867 under the title \textit{On the integrals of the hydrodynamical equations, which express vortex-motion}~\citep{Helmholtz1867a}. Inspired by the parallel between mechanics of continuous media and electromagnetism \citep[chap.~4]{Darrigol05a}, Helmholtz showed that, given a field of velocities~$\vec{v}$, its curl, the vorticity field, \begin{equation}\label{eq:omega} \vec{\omega}=\overrightarrow{\mathrm{curl}}\,\vec{v} \end{equation} is a vector field proportional to the local rotation vector of the fluid. Helmholtz introduced the notion of vortex line (a curve tangent to~$\vec{\omega}$ at each of its points) and vortex filament/tube (a bunch of vortex lines) and proved that during its evolution each vortex line follows the motion of the fluid. The dynamical equation of~$\vec{\omega}$ allowed him to study precisely the dynamics of straight (fig.~\ref{fig:vortexstructure}) and circular vortex tubes (fig.~\ref{fig:vortexanneau}). A thin vortex ring whose radius~$R$ is much larger than the radius of the cross section of the tube that defines it moves perpendicularly to its plane with the velocity of its center increasing with~$R$\footnote{\label{fn:ringtango}In particular, when two rings moving along the same direction get close, the flow created around the leading ring tends to shrink the following one which, conversely, generates a flow that tends to expend the ring ahead. Therefore the leading ring slows down while the second one is sped up until it overtakes the former by passing through it, and the role of the rings are exchanged. This tango, predicted and observed by Helmholtz, is described in the end of his 1858 memoir.}. Based on the similar mathematical problem arose in electrostatics and magnetostatics, Helmholtz understood that the topology of the irrotational part of the flow was essential to determine \emph{globally} the velocity potential~$\alpha$: in the set of the points~$P$ where~$\vec{\omega}(P)=0$ one can always \emph{locally} define a scalar field~$\alpha$ such that \begin{equation}\label{eq:gradphi} \vec{v}=\overrightarrow{\mathrm{grad}}\,\alpha \end{equation} but \begin{quotation} If we consider [a vortex-filament] as always reentrant either within or without the fluid, the space for which [equation~\eqref{eq:gradphi}] holds is complexly connected, since it remains single if we conceive surfaces of separation through it, each of which is completely bounded by a vortex-filament. In such complexly connected spaces a function~[$\alpha$] which satisfies the above equation can have more than one value ; and it must be so if it represents currents reentering, since the velocity of the fluid outside the vortex-filaments are proportional to the differential coefficients of~[$\alpha$], and therefore the motion of the fluid must correspond to ever increasing values of~[$\alpha$]. If the current returns to itself, we come again to a point where it formely was, and find there a second greater value of~[$\alpha$]. Since this may occur indefinitely, there must be for every point of such a complexly-connected space an infinite number of distinct values of~[$\alpha$] differing by equal quantities like those of~$\tan^{-1}\frac{x}{y}$, which is such a many-valued function [\dots]. \hfill \cite[\S\;3, translation by Tait]{Helmholtz1867a}. \end{quotation} \begin{figure}[!ht] \begin{center} \parbox[c]{10cm}{\includegraphics[width=10cm]{vortexstructure.eps}} \parbox[c]{4cm}{\includegraphics[width=4cm]{tornade.eps} \includegraphics[width=4cm]{katrina.eps} \includegraphics[width=4cm]{Jupiter_tacherouge.eps} \includegraphics[width=4cm]{vortex_eau.eps}} \caption{\label{fig:vortexstructure} The same year Helmholtz published his seminal memoir, the simplest model of vortex was explicitely proposed by Rankine in \protect\cite[\S\S\;629-633]{Rankine1858a} who refers to some previous theoretical analysis made by the engineer and physicist James Thomson, inventor of the vortex wheel and brother of William. The vorticity~\eqref{eq:omega} is constant and uniform inside a cylinder---in green, where the fluid rotates as a solid core and the particles rotate around themselves (the axis of the gondola rotates)---and zero outside---in blue, where the fluid particles do not rotate around themselves (the axis of the gondola keeps the same direction). When coming closer to the axis of the vortex, the velocity increases with the inverse of the distance outside the cylindrical core (and then producing a spiral-like shape) and then linearly gets to zero inside the core. In a more or less realistic way, Rankine's vortex models hurricanes, tornados or simply water going down a plughole (image credit: wikipedia, \textsc{noaa}). } \end{center} \end{figure} \begin{figure}[!ht] \begin{center} \parbox[c]{12cm}{\includegraphics[width=5cm]{vortexanneau.eps} \includegraphics[width=7cm]{smokecanon_Taitp292_structureofmatter.eps}} \includegraphics[width=4cm]{etna_ronddefumee.eps}\ \ \ \ \includegraphics[width=4.15cm]{dauphin_bullesannulaires.eps} \caption{\label{fig:vortexanneau} When some vortex lines are bended into a circular tube (in green), each portion of the ring is dragged in the same direction by the fluid whose motion is induced by the other parts of the ring. As a result, a global translation perpendicular to the ring occurs. Helmholtz'study of the dynamics of the rings and the tango played by two interacting rings moving in the same direction, see footnote~\ref{fn:ringtango} p.~\pageref{fn:ringtango}, can be visualised with Tait's smoke box (upper right taken from \protect\cite[p.~292]{Tait1884a}). In exceptional circumstances vapour rings can be naturally produced by vulcanos (the lower left photograph is taken at Etna by the vulcanologist Boris Behncke, \textsc{invg}-Osservatorio Etneo). Dolfins and whales are able to produce vortex rings in water (lower right from youtube). } \end{center} \end{figure} The topological properties of vortices can also be understood from what is now known as Kelvin's circulation theorem \cite[\S\;59d]{Thomson1869a} which unified Helmholtz results: in an inviscid (no viscosity), barotropic (its density is a function of pressure only) fluid, the flux of the vorticity \begin{equation}\label{eq:circulationv} \Gamma=\int_{\mathscr{S}}\vec{\omega}\cdot\mathrm{d}\vec{S}=\int_{\partial\mathscr{S}}\vec{v}\cdot\mathrm{d}\vec{l} \end{equation} through a surface~$\mathscr{S}$ following the motion of the fluid---or equivalently, according to Stokes' theorem, the circulation of the velocity through the boundary~$\partial\mathscr{S}$ of~$\mathscr{S}$---is constant. As a consequence, we recover Helmholtz statement that the non simple connectedness of the space filled by the irrotational part of the flow, i.e. the complementary of the vortex tubes, prevents the existence of a continuous globally-defined~$\alpha$ and the circulation~$\Gamma$ depends on the homotopy class of the loop~$\mathscr{C}=\partial\mathscr{S}$. In such an ideal fluid, the vortex lines were therefore topologically stable and Thomson's saw in this stability a key for the description of atomic properties without referring to the corpuscular image inheritated from the atomists of antiquity, which was a too suspicious philosophy for Victorian times~\cite[\S\S\;2 and~9]{Kragh02a}\footnote{Some smoothness into the atom had already been introduced by Rankine in 1851 with his hypothesis of \emph{molecular vortices} according to which ``each atom of matter consists of a nucleus or central point enveloped by an elastic atmosphere, which is retained in its position by attractive forces, and that the elasticity due to heat arises from the centrifugal force of those atmospheres, revolving or oscillating about their nuclei or central points'' \cite[\S\;2]{Rankine1851a}. It is worth noting that Rankine acknowledges the pertinence of William Thomson's comments on the first version of this 1851's proposal. }. Since vortex tubes cannot cross transversaly\footnote{But, it seems that neither Helmholtz nor Thomson have considered the possibility of a longitudinal merging of vortex tubes, forming a trousers-like shape \cite[in particular fig.~6]{VelascoFuentes07a}.} otherwise it is easy to find a~$\mathscr{C}$ that does not satisfy Kelvin's theorem, the knot formed by a closed vortex tube and the intertwinning between several such closed loop remain topologically invariant. \begin{quotation} The absolute permanence of the rotation, and the unchangeable relation you have proved between it and the portion of the fluid once acquiring such motion in a perfect fluid, shows that if there is a perfect fluid all through space, constituting the substance of all matter, a vortex-ring would be as permanent as the solid hard atoms assumed by Lucretius and his followers (and predecessors) to account for the permanent properties of bodies (as gold, lead, etc.) and the differences of their characters. Thus, if two vortex-rings were once created in a perfect fluid, passing through one another like links of a chain, they never could come into collision, or break one another, they would form an indestructible atom; every variety of combinations might exist. \flushright Thomson to Helmholtz, January 22, 1867, quoted by \cite[p.~38]{Kragh02a}. \end{quotation} The theory of the vortex atoms offered to Thomson the possibility of making concrete his long-standing intuition of a continuous conception of the world, as he had confessed it to Stokes \begin{quotation} Now I think hydrodynamics is to be the root of all physical science, and is at present second to none in the beauty of mathematics. \flushright Thomson to Stokes, December, 20, 1857,\\ quoted in \cite[p.~35]{Kragh02a \end{quotation} Despite the physical failure of Thomson's ambitious aim \citep{Silliman63a,Epple98a,Kragh02a}\footnote{As far as classical hydrodynamics is concerned, some progress have been made in the \textsc{xx}\th century with, for instance, the identification of new integrals of motion constructed from topological invariants like the Calugareanu helicity \citep{Moffatt08a}\nocite{Borisov+08a} ; experimentally some not trivial knotted vortices could be produced only recently \citep{Kleckner/Irvine13a}.}, the identification of topological invariants on knots, upon which the classification of atoms and molecules would have been based, and the classification of the knots by Tait (see Fig.~\ref{fig:Taitknots} for instance) remains a groundbreaking mathematical work, with direct repercussions in contemporary topology. One of the Thomson's greatest hopes, while spectroscopy was gathering more and more precise data, was to explain the origin of the discrete spectral lines with `` [\dots] one or more fundamental periods of vibration, as has a stringed instrument of one or more strings [\dots]''~\cite[p.~96]{Thomson1867a}. One cannot prevent to find an echo of this motivation in modern string theory where ``each particle is identified as a particular vibrational mode of an elementary microscopic string''~\cite[\S\;1.2]{Zwiebach04a}---see also \cite[in particular \S\;19]{Cappelli+12a}. Not without malice, \citet{Kragh02a} was perfectly right to qualify Thomson's dream as a ``Victorian theory of everything''. \begin{figure}[!ht] \begin{center} \includegraphics[width=10cm]{Taitstableofknotsto7crossings.eps} \caption{\label{fig:Taitknots} List of knots up to the seventh order established by \citet[Plate XLIV between p.~338 \& 339]{Tait1884a}.} \end{center} \end{figure} \section{Quantum vortices} \subsection{Topological origin of quantized flux in quantum fluids} Unlike what occurs in classical fluids where viscosity eventually make the vortices smoothly vanish, quantum fluids provide a state of matter, much more similar to ideal fluids, where vortices are strongly protected from dissipative processes. Indeed, at low temperature, particles can condensate into a collective quantum state where transport can be dissipationless: this is one of the main characteristics of superconductivity (discovered in solid mercury below~4K by Onnes in 1911), superfluidity (discovered in liquid Helium-4 below~2K by Kapitsa and Allen \& Misener in 1938), and Bose-Einstein condensate of atoms (discovered for rubidium below 170~nK by Cornell \& Wieman and Ketterle in 1995)\footnote{One can find many textbooks at different levels and more or less specialised to one type of quantum fluids. To get an introductory bird's-eye view on quantum fluids and other matters in relation to statistical physics, my personal taste go to \citep{Chaikin/Lubensky95a}, \citep{Huang87a} and the particularly sound, concise, and pedagogical \citep{Sator/Pavloff16a} (in French). }. There is a second reason, of topological origin, that reinforces the stability of the vortices in quantum fluids: the scalar field~$\alpha$ whose gradient is proportional to the current is not a simple mathematical intermediate as in the classical case (see~\eqref{eq:gradphi}) but acquires the more physical status of being a phase (an angle) that may be measured in interference experiments like in the Aharonov-Bohm effect. As a consequence, on any closed loop~$\mathscr{C}$, the circulation~$\Gamma$ given by~\eqref{eq:circulationv} has to be an integer multiple of~$2\pi$: \begin{equation}\label{eq:winding} w[\mathscr{C}]\overset{\mathrm{def}}{=}\frac{1}{2\pi}\int_{\mathscr{C}}\overrightarrow{\mathrm{grad}}\,\alpha\cdot\mathrm{d}\vec{l}\in\mathbb{Z}\;. \end{equation} Since smooth transformations cannot provoque discrete jumps, $w$~is therefore topologically protected. In other words, the flux of~$\overrightarrow{\mathrm{curl}}\,\vec{v}$---which keeps its physical interpretation of being a vorticity in superfluids as well as in Bose-Einstein condensates of atoms, whereas it represents a magnetic field in superconductors\footnote{Compare~\eqref{eq:omega} with the relation~$\vec{B}=\overrightarrow{\mathrm{curl}}\,\vec{A}$ between the (gauge) vector potential~$\vec{A}$ and the magnetic field~$\vec{B}$.}---is quantized and naturally leads to elementary vortices carrying a unit flux quantum. As a matter of fact, the quantum fluid state is described by a complex field~$\psi=|\psi|{\large \mathrm{e}}^{\mathrm{i}\alpha}$ (the order parameter) and~$w[\mathscr{C}]\neq0$ denotes a singularity of the order parameter on any surface~$\mathscr{S}$ whose boundary is~$\mathscr{C}$. Vortices constitute a particular case of what is generally called a \emph{topological defect} whose dimension depends on the dimension of the order parameter and on the dimension of the space. \begin{figure}[!ht] \begin{center} \parbox[c]{5cm}{\includegraphics[width=5cm]{Hess+89a_fig2_wolegend.eps}} \parbox[c]{7cm}{\includegraphics[width=7cm]{Yarmchuk+79a_fig2.eps}} \parbox[c]{12cm}{\includegraphics[width=12cm]{ChevyDalibard06a_fig4.eps}} \caption{\label{fig:abrikosov} Up left)~Abrikosov lattice of vortices in a superconductor \protect\cite[fig.~2]{Hess+89a}. Up right) Vortices in superfluid helium \cite[fig.~2]{Yarmchuk+79a}. Below) Vortices in a rotating Bose-Einstein condensate obtained by a)~Dalibard's group (\protect\citealt[fig.~1]{Madison+00a} \& \protect\citealt[fig.~4]{Chevy/Dalibard06a}); b)~Ketterle's group \protect\cite[fig.~4c]{Raman+01a}. \copyright European Physical Society and American Physical Society. } \end{center} \end{figure} At microscopic scales, very much like in the Rankine model, the vortex is made of a core outside which~$\overrightarrow{\mathrm{curl}}\,\vec{v}=0$; the vorticity/magnetic lines are trapped inside the core where the density of the superfluid~$|\psi|^2$ tends to zero at its center. Not only, these vortices have been observed in all the three types of superfluids mentioned above but also the triangular lattice they form to minimize the (free) energy due to an effective repulsion between them first predicted by \citet{Abrikosov57a}, see fig~\ref{fig:abrikosov}). When the fluctuations of~$|\psi|$ in space and time are negligible, notably at sufficiently low temperatures, the quantum fluid is essentially described by the phase~${\mathrm{e}}^{\mathrm{i}\alpha}$ or equivalently by a bidimensional vector of unit norm oriented at angle~$\alpha$ with respect to a given direction (fig.~\ref{fig:modeleXY_def}). \begin{figure}[!bh] \begin{center} \includegraphics[width=8cm]{modeleXY_def.ps} \caption{\label{fig:modeleXY_def} The \textsc{xy}-model describes an interacting bidimensional vector field of constant and uniform norm. On a continuous space or on a lattice, the direction of the field at point~$\vec{r}$ is given by one angle $\alpha(\vec{r})$. } \end{center} \end{figure} \subsection{The \textsc{xy}-model} The latter picture is known as the \textsc{xy}-model, which is also relevant for some classical liquid crystals or for systems of classical spins \citep[chap.~6]{Chaikin/Lubensky95a}. At macroscopic scales, some collective effects of such model are not very sensitive to the details of the interaction nor to the geometry of the elementary cell in the case of a lattice but depend crucially on the dimension~$d$ of the space of positions (the number of components of~$\vec{r}$). Typically, the energy of the system increases when some differences in the orientation~$\alpha$ appears; more precisely the energy density contain a term proportional to~$(\overrightarrow{\mathrm{grad}}\,\alpha)^2$. It is not affected by a homogenous rotation of all the spins, \begin{equation}\label{eq:uniformrotation} \alpha(\vec{r}) \mapsto \alpha(\vec{r})+\alpha_0\;, \end{equation} where the angle~$\alpha_0$ does not depend on~$\vec{r}$. The absolute minimum of the total energy is obtained when all the vectors are aligned, which is the configuration at temperature~$T=0\;\mathrm{K}$. When~$T>0$, the equilibrium corresponds to more disordered configurations but, for~$d=3$\footnote{Surprisingly, as far as the computations are concerned, the integer nature of~$d$ becomes secondary and one can formally consider~$d$ as continuous. The condition for an order/disorder phase transition at~$T_{\mathrm{critical}}>0$ to exist is~$d>2$. }, some non-zero average value of~$\alpha$ can be maintained up to a critical temperature~$T_{\mathrm{critical}}$ beyond which the average value of~$\alpha$ is zero (fig.~\ref{fig:transitionordredesordre}). \begin{figure}[!ht] \begin{center} \parbox{17cm}{\includegraphics[width=5cm]{xy_T0.eps}\ \includegraphics[width=5cm]{xy_ordonne.eps}\ \includegraphics[width=5cm]{xy_aleatoire.eps}} \caption{\label{fig:transitionordredesordre} In three dimensions the \textsc{xy}-model presents order/disorder phase transition, very similar to the familiar solid/liquid phase transition. Below a critical temperature~$T_{\mathrm{critical}}>0$ some order is maintained throughout the system at macroscopic lengths (middle picture) with the perfect order obtained at~$T=0\mathrm{K}$ (left picture). Above~$T_{\mathrm{critical}}$, the average orientation is zero and no more order at large scales can be identified (right picture).} \end{center} \end{figure} At~$d=2$, on the contrary, the correlations between fluctuations never decrease sufficiently rapidly at large distances and the average value of~$\alpha$ is zero as soon as~$T>0$. However one can still identify, at some finite temperature~$T_{\mathrm{critical}}>0$, a qualitative change of behaviour in the correlation lengths, from a power-law decay at large distances to an exponential decay and this phase transition has observable repercussions, notably in superfluids helium films~\citep{Bishop/Reppy78a}. The theoretical description of what appeared to be a new kind of phase transition, now known as topological phase transitions, was proposed by \citet{Kosterlitz/Thouless72a} who showed that vortices were a cornerstone of the scheme. As soon as their first papers, Kosterlitz and Thouless, talked about ``topological order'' because they were perfectly aware that this type of phase transition, unlike all the phase transitions known at the time of their publication, relies on topology rather than on symmetry (breaking). As we have seen above on eq.~\eqref{eq:winding}, each vortex (now a topological defect of one dimension) is characterised by an integer, called the topological index of the vortex which can be reinterpreted using the concepts introduced by Poincar\'e in a series of papers that can be considered as the foundations of topology as a fully autonomous research discipline \cite[\S\;4]{Epple98a}. Any direction far away a topological defect of dimension~$f$ in a space of dimension~$d$ is represented by an element of the rotation group in~$n=d-f-1$ dimensions, in other words such a defect can completely enclosed by a~$n$-dimensional sphere~$S_n$. In~$d=3$ dimensions a wall (a surface of dimension~$f=2$) cannot be enclosed ($n=0$), a vortex-line ($f=1$) can be enclosed by a circle ($n=1$), a point ($f=0$) can be enclosed by a $n=2$-sphere. In~$d=2$ dimensions a wall (a line of dimension~$f=1$) cannot be enclosed ($n=0$) and a point can be enclosed by a circle ($n=1$). To each direction one can associate the value of the order parameter and therefore to each defect one gets a map from~$S_n$ to~$\mathscr{P}$ where~$\mathscr{P}$ denotes the space to which the order parameter belongs. In the examples above~$\mathscr{P}$ is just the set~$S_1$ of the angles~$\alpha$ but much more different situations may be encountered. For~$n=1$, any loop~$\mathscr{C}$ around a given point maps on a closed path~$\mathscr{C}'$ in~$\mathscr{P}=S_1$ and the topological index~$w$ of the point is just the winding number of~$\mathscr{C}'$ (figs.~\ref{fig:modeleXY_index} and \ref{fig:xy_index_01m1}) \newpage \begin{figure}[!hb] \begin{center} \includegraphics[width=11cm]{modeleXY_index.ps} \caption{\label{fig:modeleXY_index} For the \textsc{xy}-model, in~$d=2$, to each point on a loop~$\mathscr{C}$ enclosing any given point~$0$ (in red, on the left) is associated the direction of the order parameter on the circle~$\mathscr{P}$ (in black on the right). } \end{center} \end{figure} \begin{figure}[!hb] \begin{center} \includegraphics[width=7cm]{modeleXY_index_T0.ps} \includegraphics[width=7cm]{modeleXY_index_ordonne.ps} \includegraphics[width=7cm]{modeleXY_index_plus1_0.eps} \includegraphics[width=7cm]{modeleXY_index_plus1_100.eps} \includegraphics[width=7cm]{modeleXY_index_moins1.eps} \caption{\label{fig:xy_index_01m1} In the \textsc{xy}-model, the topological index~\eqref{eq:winding} of a point~$O$ is the winding number of the curve~$\mathscr{C}'$ (thick black line) defined to be the image of a closed loop~$\mathscr{C}$ (in red) in the circle~$\mathscr{P}$ (thin black circle) that indicate the direction~$\alpha$ of the order parameter. A smooth deformation deforms~$\mathscr{C}'$ but do not change~$w$ (we stay in the same homotopy class). The upper row provides two examples having~$w=0$ (with, on the left, a uniform order parameter, $\mathscr{C}'$ is just a point). The central row provides two elementary vortices ($w=1$) whose configurations differ from left to right by a rotation~\eqref{eq:uniformrotation} with~$\alpha_0=-\pi/2$. The lower row provides an example of configuration having an elementary antivortex ($w=-1$). } \end{center} \end{figure} \newpage \begin{figure}[!ht] \begin{center} \parbox{17cm}{\includegraphics[width=5cm]{xy_vortexpair_1.eps} \includegraphics[width=5cm]{xy_vortexpair_0.eps} \includegraphics[width=5cm]{xy_vortexpair_m1.eps}} \caption{\label{fig:xy_index_pair} A smooth transformation that does not require a macroscopic amount of energy can make a vortex/antivortex pair to spontaneously appear as a local fluctuation at non-zero temperature. The genericity and the structural stability of this scenario can be understood when considering the appearance of a fold (fig~\ref{fig:pli}). } \end{center} \end{figure} \begin{figure}[!ht] \begin{center} \includegraphics[width=\textwidth]{fold.eps} \caption{\label{fig:pli} The fold catastrophe is the simplest of the bifurcation scenario. It involves a one real parameter family of functions where two generic critical points (on the right), having opposite second derivatives, merge into a degenerate critical point (central graph) and disappear (on the left). It also represents how generically a non transversal crossing between two tangent curves (in the center) is unfolded from one (on the left) to three (on the right) transversal crossings. } \end{center} \end{figure} \newpage \ More generally, the topological invariants are given by the group~$\pi_n$ of~$\mathscr{P}$ (for~$n=0$ it provides the connectedness, for~$n=1$ it provides the first homotopy group that is the simple connectedness, etc.). A continuous transformation of the configuration cannot modify~$w$ at any point and physically it would require a macroscopic amount of energy to change~$w$. On the other hand, one configuration having one defect can be deformed continuously at low cost of energy into any other configuration having a defect with the same~$w$. In particular, the transformation~\eqref{eq:uniformrotation} does not cost any energy at all. One cannot therefore expect to \emph{isolated} elementary vortex ($w=1$) or \emph{isolated} elementary antivortex ($w=-1$) to be spontaneously created from a perfect ordered state. Nevertheless, a pair of vortex-antivortex is affordable when~$T>0$ (fig.~\ref{fig:xy_index_pair}). The continuous creation (or annihilation) of such a pair can be understood by considering the appearance of a fold on a drapery (back to Leonardo again?). One may intuitively see that this is a generic process, stable with respect to smooth transformations, that describes the creation or the annihilation of a pair of maximal-minimal points on a smooth function (fig.~\ref{fig:pli}) or, equivalently, the creation or annihilation of intersection points when two curves that cross transversaly are smoothly locally deformed\footnote{Topology is fully at work here and the study of the stability of the critical points of smooth mappings is the object of catastrophe theory whose greatest achievement is to have classified the generic possible scenarios; the simplest one being precisely the fold catastrophe, depicted in figure~\ref{fig:pli} \citep[for a general survey]{Poston/Stewart78a,Arnold84a}.}. The topological phase transition describes precisely how the creation of an increasing number of vortex-antivortex pairs as the temperature increases eventually lead from a topological order to a state where complete disorder reigns. \section{Concluding remark} To come back to issues mentioned in the last paragraph of~\S\;1, in quantum theory, the fundamental elementary particles stem from algebraic symmetry considerations. However, we have some clues (topological defects, solitons, instantons, monopoles, etc.) that topology may offer a complementary ground. The parallel between creation/annihilation of particle-antiparticle pairs and creation/annihilation of vortex-antivortex pairs may be more than a simple analogy. Thomson/Kelvin's intuition may take an unexpected but relevant form, after all. \bigskip \textbf{Acknowledgement}: I am particularly grateful to Pascal Brioist (Centre d'\'Etude Sup\'erieures de la Renaissance de l'Universit\'e de Tours) for his expert advices on Leonardo studies, to Boris Behncke (\textsc{invg}-Osservatorio Etneo)) for letting me use his photo of the vapour ring created by the Etna (see fig~\ref{fig:vortexanneau}) and to Michele Emmer who triggered the subject of this essay for the conference \textit{Matematica e Cultura 2017}, Imagine Math~6, at Venice. \nocite{James99a}

https://arxiv.org/abs/1711.00498

A conformal-type energy inequality on hyperboloids and its application to quasi-linear wave equation in $\mathbb{R}^{3+1}$

In the present work, we will develop a conformal inequality in the hyperbolic foliation context which is analogous to the conformal inequality in the classical time-constant foliation context. Then as an application, we will apply this a priori estimate to the problem of global existence of quasi-linear wave equations in three spatial dimensions under null condition. With the aid of this inequality, we can establish more precise decay estimates on the global solution.

\section{Introduction} In this article we will develop a conformal inequality in the hyperbolic foliation context analogue to the classical time-constant foliation context. Then we will apply this estimate to the problem of global existence of regular solution associated to small regular initial data (also called the {\sl global stability } for short in the following text) for quasi-linear wave equation in three spatial dimensions. More precisely, we are going to regard the following quasi-linear wave equation: \begin{equation}\label{eq main} \Box u + Q^{\alpha\beta\gamma}\del_{\gamma}u\del_{\alpha}\del_{\beta}u = 0 \end{equation} where $Q$ is supposed to be a {\sl null} cubic form. One can also consider a system with semi-linear terms such as: \begin{equation}\label{eq main'} \Box u + Q^{\alpha\beta\gamma}\del_{\gamma}u\del_{\alpha}\del_{\beta}u = N^{\alpha\beta}\del_{\alpha}u\del_{\beta}u \end{equation} with $N$ a {\sl null} quadratic form. But through an algebraic observation one can show that by a change of known $$ v = u + \frac{\sigma}{2}u^2, \quad \sigma = N^{00} $$ the semi-linear term can be eliminated with some high-order correction terms, which are negligible in dimension three. \subsection{The conformal inequality on hyperboloids} In the classical flat foliation context, the conformal energy inequality is a well-known $L^2$ type estimate which controls more quantities than the classical energy inequality. We recall that for $$ \Box u = f, $$ the conformal energy $\Ec(t,u)$ is defined as $$ \Ec(t,u) : =\int_{\RR^3}\big(|Su|^2 + \sum_a|L_au|^2 + \sum_{a\neq b}|\Omega_{ab}u|^2 + u^2\big)(x,t)\ dx $$ where $$ Su: = t\del_tu + x^a\del_au, \quad L_a u: = x^a\del_tu + t\del_au,\quad \Omega_{ab}u := x^a\del_bu - x^b\del_au. $$ And by using the multiplier $\left((r^2+t^2)\del_t + 2x^at\del_a + 2t \right)u$, we have the following estimate (see for example \cite{A1} section 6.7): There exists a positive constant $C$ such that, for all $u\in C^2\left(\RR_x^3\times[0,T)\right)$, sufficiently decaying when $|x|\rightarrow +\infty$, and all $t<T$, \begin{equation}\label{eq 1 30-05-2017} \Ec(t,u)^{1/2}\leq C\Ec(0,u)^{1/2} + C\int_0^t\big\|\sqrt{r^2+t^2}f(x,t)\big\|_{L^2_x(\RR^n)}\ dx. \end{equation} From the above estimates we can see the following facts. For the homogeneous wave equation $\Box u = 0$, if we differentiate the equation with respect to $\del^IL^J$, $$ \Box \del^IL^Ju = 0 $$ and by \eqref{eq 1 30-05-2017}, we see that $\Ec(t,\del^IL^Ju)^{1/2}$ is bounded by the initial conformal energy. This leads to the fact that \begin{equation}\label{eq 1 01-11-2017} \|S\del^IL^J u(t,\cdot )\|_{L^2(\RR^3)},\quad \|L_a\del^IL^J u(t,\cdot )\|_{L^2(\RR^3)},\quad \|\Omega_{ab}\del^IL^J u(t,\cdot )\|_{L^2(\RR^3)},\quad \|\del^IL^J u(t,\cdot)\|_{L^2(\RR^3)} \end{equation} are bounded. Then by some estimates on commutators and the Klainerman-Sobolev type inequality, we will see that $$ Su,\quad L_au,\quad \Omega_{ab}u $$ are bounded by $(|t-r|+1)^{-1/2}(t+1)^{-1}$, and especially: \begin{equation}\label{eq 1 29-08-2017} u\sim (t+1)^{-1}(|t-r|+1)^{-1/2}, \quad \del_{\alpha} u\sim (t+1)^{-1}(|t-r|+1)^{-3/2}. \end{equation} Compared with the classical energy-vector field argument, this conformal energy-vector field method supplies better decay bounds on the solution $u$: the classical energy only gives the following decay bounds: \begin{equation}\label{eq 2 01-11-2017} \del_{\alpha}u\sim (t+1)^{-1}(|t-r|+1)^{-1/2}. \end{equation} In the present article, we will establish a parallel estimate in the hyperbolic foliation context. More precisely, we will use the following multiplier: $$ (s/t)\left((t^2+r^2)\del_t + 2x^at\del_a + 2t\right)u $$ where $s = \sqrt{t^2-r^2}$ is the hyperbolic time. In the flat (Minkowski metric) case, this energy is written as $$ \Ec(s,u) = :\int_{\Hcal_s}\big|s(s/t)^2\del_tu\big|^2 + (s/t)^2\sum_a\big|L_au\big|^2 + (s/t)u^2 dx. $$ Here $\Hcal_s = \{x\in\RR^{1+3}|t = \sqrt{s^2+r^2}\}$, and for a $f$ defined in $\RR^{1+3}$, we define $$ \int_{\Hcal_s}fdx := \int_{\RR^3}f(\sqrt{s^2+|x|^2},x)dx. $$ For the linear equation $$ \Box u = f $$ the following estimate holds: \begin{equation} \Ec(s,u)^{1/2}\leq \Ec(s_0,u) + C\int_{s_0}^s \tau \|f\|_{L^2(\Hcal_{\tau})}d\tau \end{equation} For the homogeneous linear wave equation $\Box u = 0$, the above quantity is conserved. For the quasi-linear wave equation \eqref{eq main}, we will prove that this energy is also bounded up to the highest order. Combined with the global Sobolev's inequality, the above energy bounds leads to the following terms bounded by $C(t+1)^{-3/2}$: $$ (s/t)L_a u ,\quad s(s/t)^2\del_t \quad (s/t)u, $$ and these bounds gives the following decay rate: $$ u\sim (t+1)^{-1}(|t-r|+1)^{-1/2}, \quad \del_{\alpha} u\sim (t+1)^{-1}(|t-r|+1)^{-3/2} $$ which coincides with the classical conformal energy bounds. We will prove that in the quasi-linear case, these decay rates still hold. \subsection{The global existence result for quasi-linear wave equation} The problem on the stability of quasi-linear wave equations or systems has attracted lots of attentions of the mathematical community. Our method to be presented belongs to the ``vector-field method'' which was introduced by S. Klainerman for wave equation (\cite{Kl1}) and for Klein-Gordon equation (\cite{Kl2}). This method is then extended to many different cases. The global existence of \eqref{eq main} has been established by S. Klainerman in \cite{Kl1}. His method is based on the time-constant foliation and standard energy inequalities. The hyperbolic foliation and hyperbolic variables are introduced by S. Klainerman firstly for the analysis on Klein-Gordon equation in \cite{Kl2}, see also \cite{H1} for the ``alternative energy method''. This method is then revisited and applied by P. LeFloch et al. in \cite{PLF-MY-book} on system composed by wave equations and Klein-Gordon equations. The application of the hyperbolic foliation has the advantage that it does not require the scaling invariance of the system (for example, the wave-Klein-Gordon system). However, the scaling invariance of the wave equation does supply more conserved quantities. The conformal inequality \eqref{eq 1 30-05-2017} is essentially due to the scaling invariance of the wave equation (i.e. it does not hold for Klein-Gordon equation, which does not enjoy the scaling invariance), and it leads to better decay estimates (\eqref{eq 1 29-08-2017} compared with \eqref{eq 2 01-11-2017} supplied by the classical energy). In this article, we show that this conformal energy inequality on hyperboloids also leads to the global stability of \eqref{eq main} and it gives more precise decay rate for the global solution. \section{The conformal inequality on hyperboloids: flat case} \subsection{The hyperbolic variables and hyperbolic frame} In this subsection we briefly recall the notion of the hyperbolic frame. We denote by $\Kcal = \{(t,x)\in\RR^{1+3}|t>|x|+1\}$ the interior of the light cone. In this region we introduce the following parametrization by the hyperbolic variables: $$ \xb^0 := s := \sqrt{t^2-|x|^2},\quad \xb^a: = x^a. $$ The canonical frame associate to this parametrization is called the hyperbolic frame, denoted by: $$ \delb_0: = \del_s = \frac{s}{t}\del_t,\quad \delb_a : = \frac{x^a}{t}\del_t + \del_a. $$ The transition matrix between the hyperbolic frame and the natural frame is $$ \delb_\alpha = \Phib_{\alpha}^{\beta}\del_{\beta} $$ with $$ \left(\Phib_{\alpha}^{\beta}\right)_{\alpha\beta} := \left( \begin{array}{cccc} s/t &0 &0 &0 \\ x^1/t &1 &0 &0 \\ x^2/t &0 &1 &0 \\ x^3/t &0 &0 &1 \end{array} \right) $$ and its inverse is $$ \left(\Psib_{\alpha}^{\beta}\right)_{\alpha\beta} = \left( \begin{array}{cccc} t/s &0 &0 &0 \\ -x^1/s &1 &0 &0 \\ -x^2/s &0 &1 &0 \\ -x^3/s &0 &0 &1 \end{array} \right) $$ thus $$ \del_{\alpha} = \Psib_{\alpha}^{\beta}\delb_{\beta}. $$ Now we introduce the following notation. Let $T$ be a two-tensor, we recall that it can be written in the natural frame $\{\del_{\alpha}\}$ or in the hyperbolic frame $\{\delb_{\alpha}\}$. That is $$ T = T^{\alpha\beta}\del_{\alpha}\otimes\del_{\beta} = \Tb^{\alpha\beta}\delb_{\alpha}\otimes\delb_{\beta} $$ with $$ \Tb^{\alpha\beta} = T^{\alpha'\beta'}\Psib_{\alpha'}^{\alpha}\Psib_{\beta'}^{\beta}. $$ For a cubic form $Q$, we also have similar notation: $$ Q = Q^{\alpha\beta\gamma}\del_{\alpha}\otimes\del_{\beta}\otimes\del_{\gamma} = \Qb^{\alpha\beta\gamma}\delb_{\alpha}\otimes\delb_{\beta}\otimes\delb_{\gamma}, $$ and $$ \Qb^{\alpha\beta\gamma} = Q^{\alpha'\beta'\gamma'}\Psib_{\alpha'}^{\alpha}\Psib_{\beta'}^{\beta}\Psib_{\gamma'}^{\gamma}. $$ We also recall the dural co-frame to the hyperbolic frame: $$ \omegab^0 = (t/s)dt - \sum_a(x^a/s)dx^a,\quad \omegab^a = dx^a. $$ We recall the Minkowski metric $m^{\alpha\beta}$ written in the hyperbolic frame: $$ \mb^{\alpha\beta} = \left( \begin{array}{cccc} 1 &x^1/s &x^2/s &x^3/s \\ x^1/s &-1 &0 &0 \\ x^2/s &0 &-1 &0 \\ x^3/s &0 &0 &-1 \end{array} \right). $$ And we remark that for a second order differential operator $g^{\alpha\beta}\del_{\alpha}\del_{\beta}$, we see that $$ g^{\alpha\beta}\del_{\alpha}\del_{\beta} = g^{\alpha\beta}\Psib_{\alpha}^{\alpha'}\Psib_{\beta}^{\beta'}\delb_{\alpha'}\delb_{\beta'} + g^{\alpha\beta}\del_{\alpha}\left(\Psib_{\beta}^{\beta'}\right)\delb_{\beta'} = \gb^{\alpha\beta}\delb_{\alpha}\delb_{\beta} + g^{\alpha\beta}\del_{\alpha}\left(\Psib_{\beta}^{\beta'}\right)\delb_{\beta'}. $$ Here we list out $\del_{\alpha}\left(\Psib_{\beta}^{\beta'}\right)$: \begin{subequations} \begin{equation}\label{eq 2 14-08-2017} \del_t\left(\Psib_0^0\right) = \frac{-r^2}{s^3} = -s^{-1}\Psib_0^0\Psib_0^0 + s^{-1},\quad \del_t\left(\Psib_a^0\right) = \frac{x^a t}{s^3} = -s^{-1}\Psib_0^0\Psib_a^0 \end{equation} and \begin{equation}\label{eq 3 14-08-2017} \del_a\left(\Psib_0^0\right) = \frac{tx^a}{s^3} = -s^{-1}\Psib_0^0\Psib_a^0,\quad \del_a\left(\Psib_b^0\right) = -\frac{\delta_{ab}}{s} - s^{-1}\Psib_a^0\Psib_b^0. \end{equation} \end{subequations} The rest components of $\del_{\alpha}\left(\Psib_{\beta}^{\beta'}\right)$ are zero. Now we recall that $$ \Box = m^{\alpha\beta}\del_{\alpha}\del_{\beta} $$ thus in hyperbolic frame: \begin{equation}\label{eq 1 12-05-2017} \aligned \Box =& \mb^{\alpha\beta}\delb_{\alpha}\delb_{\beta} + m^{\alpha\beta}\del_{\alpha}\left(\Psib_{\beta}^{\beta'}\right)\delb_{\beta'} =\delb_s\delb_s + 2(x^a/s)\delb_s\delb_a- \sum_a\delb_a\delb_a + \frac{3}{s}\delb_s. \endaligned \end{equation} \subsection{Hyperbolic decomposition of $\Box$} We write \eqref{eq 1 12-05-2017} into the following form: \begin{equation}\label{eq1-11-02-2017} \aligned \Box u =& \delb_s\left(\delb_s + (2x^a/s)\delb_a\right)u - \sum_a\delb_a\delb_au + \frac{2x^a}{s^2}\delb_au + \frac{3}{s}\delb_su \\ =&\delb_s\left(\delb_s + (2x^a/s)\delb_a\right)u - \sum_a\delb_a\delb_au + s^{-1}\left(\delb_s + (2x^a/s)\delb_a\right)u + \frac{2}{s}\delb_su \\ =&s^{-1}\del_s\left(s\delb_s u+ 2x^a\delb_au\right) - \sum_a\delb_a\delb_au + \frac{2}{s}\delb_su \endaligned \end{equation} We denote by $$ Ku := \left(s\delb_s + 2x^a\delb_a\right)u = \frac{t^2+r^2}{t}\del_tu + 2r\del_ru $$ \subsection{The energy identity} We make the following calculation: \begin{equation}\label{eq1-10-05-2017} \aligned sKu \cdot \Box u =& \left(s\left(\delb_s + (2x^b/s)\delb_b\right)u\right)\delb_s\left(s\left(\delb_s + (2x^a/s)\delb_a\right)u\right) \\ &-\sum_a\delb_a\left(s^2\left(\delb_su + (2x^b/s)\delb_bu\right)\cdot\delb_au\right) + s^2\sum_a\delb_a\left(\left(\delb_su + (2x^b/s)\delb_bu\right)\right)\delb_au \\ &+2s\delb_su\cdot \left(\delb_su + (2x^b/s)\delb_bu\right) \\ =& \frac{1}{2}\delb_s\left(|Ku|^2\right) +\sum_a\left(s^2\delb_a\delb_su\cdot \delb_au + 2s|\delb_au|^2 + s^2(2x^b/s)\delb_a\delb_bu\cdot \delb_a u\right) \\ &+2s\delb_su\cdot \left(\delb_su + (2x^b/s)\delb_bu\right) - \sum_a\delb_a\left(s^2\left(\delb_su + (2x^b/s)\delb_bu\right)\cdot\delb_au\right) \\ =& \frac{1}{2}\delb_s\left(|Ku|^2\right) + \frac{1}{2}\sum_a\delb_s\left(|s\delb_au|^2\right) \\ & + 2s\left(\delb_su\cdot \left(\delb_su + (2x^b/s)\delb_bu\right) - \sum_a|\delb_au|^2\right) \\ &+ \sum_a\delb_b\left(sx^b|\delb_au|^2\right) - \sum_a\delb_a\left(s^2\left(\delb_su + (2x^b/s)\delb_bu\right)\cdot\delb_au\right). \endaligned \end{equation} We also remark that $$ \aligned u\cdot\Box u =&\del_s(u\delb_s u) - (\delb_su)^2 + s^{-1}\delb_s\left(2x^au\delb_a u\right)- (2x^a/s)\delb_au\delb_su \\ &- \sum_a\delb_a(u\delb_au) + \sum_a(\delb_au)^2 + \frac{n}{2}s^{-1}\delb_s (u^2) \\ =&s^{-1}\delb_s(su\delb_su) + s^{-1}\delb_s\left(u^2\right) + s^{-1}\delb_s\left(2x^au\delb_au\right) - \sum_a\delb_a\left(u\delb_au\right) \\ &-\left(|\delb_su|^2 + (2x^a/s)\delb_su\delb_au - \sum_a|\delb_au|^2 \right) \endaligned $$ thus $$ \aligned su\cdot\Box u =& \delb_s(su\delb_su) + \delb_s(u^2) + \delb_s\left(2x^a u\delb_au\right) - \sum_a\delb_a(su\delb_au) \\ &-s\left((\delb_su)^2 + 2(x^a/s)\delb_su\delb_au - \sum_a(\delb_au)^2\right). \endaligned $$ So we see that \begin{equation}\label{eq 2 10-05-2017} \aligned &2su\cdot \Box u + 2s\left((\delb_su)^2 + 2(x^a/s)\delb_su\delb_au - \sum_a(\delb_au)^2\right) \\ =& \delb_s\left(2su\delb_su + 4x^au\delb_au+ 2u^2\right) - \sum_a\delb_a\left(2su\delb_au\right). \endaligned \end{equation} Then we combine \eqref{eq1-10-05-2017} and \eqref{eq 2 10-05-2017}, \begin{equation}\label{eq 3 10-05-2017} 2s\left(Ku + 2u\right)\Box u = \delb_s\left(|Ku|^2+\sum_a|s\delb_au|^2 + 4uKu + 4u^2\right) + \delb_a(v^a) \end{equation} with $$ v^a := -4su\delb_au + 2sx^a\sum_{b}|\delb_bu|^2 - 2s\delb_au\cdot Ku $$ We define the {\bf flat energy} \begin{equation}\label{eq 4 10-05-2017} \aligned \Ec(s,u):=&\int_{\Hcal_s}\left((Ku)^2+\sum_a(s\delb_au)^2 + 4uKu + 4u^2\right)dx \\ =&\int_{\Hcal_s}\left(\sum_a(s\delb_au)^2 +\left(Ku + 2u\right)^2\right)dx \endaligned \end{equation} Then \eqref{eq 3 10-05-2017} leads to $$ 2\int_{\Hcal_s}s\left(K+2\right)u\Box u dx = \frac{d}{ds}\Ec(s,u) $$ Remark that $\|(Ku+2u)\|_{L^2(\Hcal_s)}\leq \Ec(s,u)^{1/2}$, $$ \frac{d}{ds}\Ec(s,u)^{1/2}\leq 2s\|\Box u\|_{L^2(\Hcal_s)} $$ which leads to \begin{equation}\label{eq 5 10-05-2017} \Ec(s,u)^{1/2}\leq \Ec(s_0,u)^{1/2} + 2\int_{s_0}^s\tau\|\Box u\|_{L^2(\Hcal_\tau)} d\tau. \end{equation} We conclude the above calculation by the following {\bf conformal energy estimate in flat case} \begin{proposition} Let $u$ be a function defined in $\Kcal_{[s_0,s_1]}$, sufficiently regular and vanishes near the conical boundary $\del\Kcal := \{(t,x)|t = r+1\}$. Then the estimate \eqref{eq 5 10-05-2017} holds. \end{proposition} \subsection{Analysis on the flat energy} It is clear that the flat energy can control the following quantities: $$ \|Ku + 2u\|_{L^2(\Hcal_s)},\quad s\|\delb_au\|_{L^2(\Hcal_s)} $$ and in this subsection we will prove the following proporty: \begin{proposition} Let $n=3$ and $u$ be a sufficiently regular function defined in the region $\Kcal_{[s_0,s_1]}$ and vanishes near the conical boundary. Then the following inequality holds for $s_0 \leq s\leq s_1$: \begin{equation}\label{eq 6 23-05-2017} \|(s/r)u\|_{L^2(\Hcal_s)}\leq 2\Ec(s,u)^{1/2}, \end{equation} \begin{equation}\label{eq 1 25-05-2017} \|s^2t^{-1}\delb_su\|_{L^2(\Hcal_s)}\leq C\Ec(s,u)^{1/2} \end{equation} and \begin{equation}\label{eq 7 22-07-2017} \|s(s/t)^2\del_a u\|_{L^2(\Hcal_s)}\leq C\Ec(s,u)^{1/2} \end{equation} with $C$ a universal constant. \end{proposition} \begin{proof} For \eqref{eq 6 23-05-2017}, we recall the Hardy's inequality for $\RR^3$. Let $w$ be a sufficiently regular function defined in $\RR^3$ and decreases sufficiently fast at infinity, then \begin{equation}\label{eq 2 25-05-2017} \|w/r\|_{L^2(\RR^3)}\leq 2\sum_a^3\|\del_a w\|_{L^2(\RR^3)}. \end{equation} We define (for a fixed $s$) $\tilde{u}_s:\RR^3\rightarrow \RR$: $$ \tilde{u}_s(x) := su(\sqrt{s^2+r^2},x). $$ We see that $$ \del_a \tilde{u}_s = s\delb_au $$ and we apply \eqref{eq 2 25-05-2017} with $w = \tilde{u}_s$ (remark that $\tilde{u}_s$ is compactly supported), this leads to $$ \|su/r\|_{L^2(\Hcal_s)}\leq 2\sum_a\|s\delb_au\|_{L^2(\Hcal_s)}\leq 2\Ec(s,u)^{1/2} $$ which is \eqref{eq 6 23-05-2017}. To establish \eqref{eq 1 25-05-2017}, we remark that $$ Ku + 2u = s\delb_su + 2x^a\delb_au + 2u $$ which leads to $$ \aligned (s^2/t)|\delb_su|\leq (s/t)|Ku+2u| + 2\sum_a|(x^a/t)s\delb_au| + 2|(s/t)u| \endaligned $$ thus $$ \aligned \|(s^2/t)\delb_su\|_{L^2(\Hcal_s)}\leq& \|(s/t)(Ku+2u)\|_{L^2(\Hcal_s)} + 2\sum_a\left\|s\delb_a u\right\|_{L^2(\Hcal_s)} + 2\|(s/r)u\|_{L^2(\Hcal_s)} \\ \leq& 7\Ec(s,u)^{1/2}. \endaligned $$ For the third one, we remark that $$ s(s/t)^2\del_a u = s(s/t)^2\delb_au - s(s/t)^2(x^a/t)\del_tu $$ which leads to $$ \|s(s/t)^2\del_au\|_{L^2(\Hcal_s)}\leq \|s\delb_au\|_{L^2(\Hcal_s)} + \|s(s/t)\delb_su\|_{L^2(\Hcal_s)}\leq C\Ec(s,u)^{1/2}. $$ \end{proof} \begin{remark} We see that when $n=3$ and when the flat energy is satisfies the following increasing condition: $$ \Ec(s,u)^{1/2}\leq Cs^{\delta} $$ then $$ \|sr^{-1}u\|_{L^2(\Hcal_s)}\leq Cs^{\delta}. $$ \end{remark} \section{The conformal inequality on hyperboloids: curved case} \subsection{Differential identity}\label{subsec energy-indentity} We suppose that $g^{\alpha\beta}$ is a (symmetric) Lorantzian metric defined in $\Kcal$, sufficiently regular, and coincides with the Minkowski metric near the light cone $\del\Kcal$. Then we remark the following calculation: $$ \aligned g^{\alpha\beta}\del_{\alpha}\del_{\beta}u =& \gb^{\alpha\beta}\delb_{\alpha}\delb_{\beta}u + g^{\alpha\beta}\del_{\alpha}\left(\Psib_{\beta}^{\beta'}\right)\delb_{\beta'}u \\ =& \gb^{\alpha\beta}\delb_{\alpha}\delb_{\beta}u + g^{\alpha\beta}\del_{\alpha}\left(\Psib_\beta^0\right)\delb_su. \endaligned $$ and we thus have $$ \aligned g^{\alpha\beta}\del_{\alpha}\del_{\beta}u =& \delb_s\left(\gb^{00}\delb_su + 2\gb^{0a}\delb_au\right) - \delb_s\gb^{00}\delb_su - 2\delb_s\gb^{a0}\delb_au \\ &+\gb^{ab}\delb_a\delb_bu + g^{\alpha\beta}\del_{\alpha}\left(\Psib_\beta^0\right)\delb_su \\ =&\delb_s\left(\gb^{00}\delb_su + 2\gb^{0a}\delb_au\right) + s^{-1}\left(\gb^{00}\delb_su + 2\gb^{a0}\delb_au\right) + \gb^{ab}\delb_a\delb_bu \\ &-\del_s\gb^{00}\delb_su - 2s^{-1}\left(\gb^{a0} + s\delb_s\gb^{a0}\right)\delb_au + \left( g^{\alpha\beta}\del_{\alpha}\left(\Psib_\beta^0\right) - s^{-1}\gb^{00}\right)\delb_su \\ =&s^{-1}\delb_s\left(s\left(\gb^{00}\delb_su + 2\gb^{0a}\delb_au\right)\right) + \gb^{ab}\delb_a\delb_bu \\ &-\delb_s\gb^{00}\delb_su - 2s^{-1}\left(\gb^{a0} + s\delb_s\gb^{a0}\right)\delb_au + \left( g^{\alpha\beta}\del_{\alpha}\left(\Psib_\beta^0\right) - s^{-1}\gb^{00}\right)\delb_su. \endaligned $$ We denote by $$ \mathscr{K}_g = s\left(\gb^{00}\delb_s + 2\gb^{a0}\delb_a\right). $$ and we calculate the following relations: $$ \gb^{0a} + s\del_s\gb^{0a} = s\Psib_{\alpha}^0\del_sg^{a\alpha} + (s/t)g^{0a} $$ and $$ \del_s\gb^{00} = \Psib_{\alpha}^0\Psib_{\beta}^0\del_sg^{\alpha\beta} - 2s^{-3}\left(r^2\gb^{00} + x^ax^bg^{ab}\right) - 2g^{a0}\Psib_a^0\frac{t^2+r^2}{ts^2}. $$ and \begin{equation}\label{eq 3 25-05-2017} \aligned g^{\alpha\beta}\del_{\alpha}\Psib_{\beta}^0 =& -s^{-1}\left(g^{00}(r/s)^2 - 2g^{a0}\frac{t}{s}\cdot\frac{x^a}{s} + g^{ab}\left(\frac{x^a}{s}\cdot\frac{x^b}{s} + \delta_{ab}\right)\right) \\ =& -s^{-1}\left(g^{00}(t/s)^2 - 2g^{a0}\frac{t}{s}\cdot\frac{x^a}{s} + g^{ab}\frac{x^a}{s}\cdot\frac{x^b}{s}\right) + s^{-1}\left(g^{00} - \sum_ag^{aa}\right) \\ =& -s^{-1}\gb^{00} + s^{-1}\left(g^{00} - \sum_ag^{aa}\right). \endaligned \end{equation} Then as the flat case, we use the multiplier $s\mathscr{K}_gu$ and see that $$ \aligned s\mathscr{K}_gu\cdot s^{-1}\del_s\left(\mathscr{K}_gu\right) = \frac{1}{2}\delb_s\left(|\mathscr{K}_gu|^2\right) \endaligned $$ $$ \aligned s\mathscr{K}_gu\cdot \gb^{ab}\delb_a\delb_bu =& s\delb_a\left(\mathscr{K}_gu\cdot \gb^{ab}\delb_bu\right) - s\delb_a(\mathscr{K}_gu)\gb^{ab}\delb_bu - s\delb_a\gb^{ab}\cdot \mathscr{K}_gu\delb_bu \endaligned $$ where $$ \aligned s\delb_a(\mathscr{K}_gu)\gb^{ab}\delb_bu =& s^2\delb_a\left(\gb^{00}\delb_su + 2\gb^{c0}\delb_cu\right)\cdot \gb^{ab}\delb_bu \\ =&s^2\gb^{00}\delb_a\delb_su\cdot\gb^{ab}\delb_bu + 2s^2\gb^{c0}\delb_a\delb_cu\cdot \gb^{ab}\delb_bu \\ &+s^2\delb_a\gb^{00}\cdot \gb^{ab}\delb_su\delb_bu + 2s^2\delb_a\gb^{c0}\cdot \gb^{ab}\delb_cu\delb_bu \\ =&\frac{s^2}{2}\delb_s\left(\gb^{00}\gb^{ab}\delb_au\delb_bu\right) + s^2\delb_c\left(\gb^{c0}\gb^{ab}\delb_au\delb_bu\right) \\ &-\frac{s^2}{2}\delb_s\left(\gb^{00}\gb^{ab}\right)\delb_au\delb_bu - s^2\delb_c\left(\gb^{c0}\gb^{ab}\right)\delb_au\delb_bu \\ &+s^2\delb_a\gb^{00}\cdot \gb^{ab}\delb_su\delb_bu + 2s^2\delb_a\gb^{c0}\cdot \gb^{ab}\delb_cu\delb_bu \\ =&\frac{1}{2}\delb_s\left(s^2\gb^{00}\gb^{ab}\delb_au\delb_bu\right) + \delb_c\left(s^2\gb^{c0}\gb^{ab}\delb_au\delb_bu\right) \\ &-s\gb^{00}\gb^{ab}\delb_au\delb_bu - \frac{s^2}{2}\delb_s\left(\gb^{00}\gb^{ab}\right)\delb_au\delb_bu -s^2\delb_c\left(\gb^{c0}\gb^{ab}\right)\delb_au\delb_bu \\ &+s^2\delb_a\gb^{00}\cdot \gb^{ab}\delb_su\delb_bu + 2s^2\delb_a\gb^{c0}\cdot \gb^{ab}\delb_cu\delb_bu, \endaligned $$ $$ s\mathscr{K}_gu\cdot \left(g^{\alpha\beta}\del_{\alpha}\left(\Psib_{\beta}^0\right)-s^{-1}\gb^{00}\right)\delb_su = s\left(sg^{\alpha\beta}\del_{\alpha}\left(\Psib_{\beta}^0\right) - \gb^{00}\right)\left(\gb^{00}\delb_su\delb_su + 2\gb^{a0}\delb_au\delb_su\right). $$ Thus we see that \begin{equation}\label{eq 1 21-05-2017} \aligned s\mathscr{K}_gu\cdot g^{\alpha\beta}\del_{\alpha}\del_{\beta}u =& \frac{1}{2}\delb_s\left(|\mathscr{K}_gu|^2 - s^2\gb^{00}\gb^{ab}\delb_au\delb_bu \right) + \delb_a(v_g^a) \\ &+ s\left(\gb^{00}\gb^{ab} + s\delb_c\left(\gb^{0c}\gb^{ab}\right) - 2s\delb_c\gb^{0a}\cdot\gb^{cb}\right)\delb_au\delb_bu \\ &+\left(sg^{\alpha\beta}\del_{\alpha}\left(\Psib_{\beta}^0\right) - \gb^{00}\right)\mathscr{K}_gu\cdot \delb_su \\ &\quad + \frac{s^2}{2}\delb_s\left(\gb^{00}\gb^{ab}\right)\delb_au\delb_bu - s^2\delb_a\gb^{00}\cdot\gb^{ab}\delb_su\delb_bu \\ &\quad - s\mathscr{K}_gu\cdot \delb_s\gb^{00}\delb_su - 2\mathscr{K}_gu\cdot \delb_s\left(s\gb^{a0}\right)\cdot \delb_au - s\delb_a\gb^{ab}\cdot \delb_b u \mathscr{K}_gu \endaligned \end{equation} with $$ v_g^a = s\mathscr{K}_g\cdot \gb^{ab}\delb_b u - s^2\gb^{a0}\gb^{cb}\delb_cu\delb_bu, $$ $$ \aligned su\cdot g^{\alpha\beta}\del_{\alpha}\del_{\beta}u =& su\cdot \gb^{\alpha\beta}\delb_{\alpha}\delb_{\beta}u + su\cdot g^{\alpha\beta}\del_{\alpha}\left(\Psib_{\beta}^0\right)\delb_su \\ =&\delb_s\left(u\cdot \mathscr{K}_g u + \frac{1}{2}\left(sg^{\alpha\beta}\del_{\alpha}\left(\Psib_\beta^0\right)-\delb_s(s\gb^{00})\right)u^2\right) +\delb_a\left(s u\gb^{ab}\delb_bu\right) \\ &-s\gb^{\alpha\beta}\delb_{\alpha}u\delb_{\beta}u \\ &+\frac{1}{2}\delb_s\left(\delb_s(s\gb^{00})-sg^{\alpha\beta}\del_{\alpha}\left(\Psib_\beta^0\right)\right)u^2 -2\delb_s\left(s\gb^{a0}\right)u\delb_au - \delb_a\left(s\gb^{ab}\right)u\delb_bu. \endaligned $$ We denote by $$ N_g := sg^{\alpha\beta}\del_{\alpha}\left(\Psib_\beta^0\right) - \delb_s(s\gb^{00}) $$ and by \eqref{eq 3 25-05-2017} we see that $$ N_g = g^{00} - \sum_ag^{aa} - 2\gb^{00} - s\delb_s\gb^{00}. $$ When $g^{\alpha\beta} = m^{\alpha\beta}$, we see that $N_m = n-1$. Then we see that \begin{equation}\label{eq 2 21-05-2017} \aligned N_gsu\cdot g^{\alpha\beta}\del_{\alpha}\del_{\beta}u =& -N_gs\gb^{\alpha\beta}\delb_{\alpha}u\delb_{\beta}u +\delb_s\left(N_gu\cdot \mathscr{K}_gu + \frac{1}{2}N_g^2u^2\right) +\delb_a\left(N_g\cdot s u\gb^{ab}\delb_bu\right) \\ &-\delb_sN_g\cdot \left(u\cdot \mathscr{K}_gu + \frac{1}{2}N_gu^2\right) - \delb_aN_g\cdot su\gb^{ab}\delb_bu \\ &-\frac{1}{2}N_g\delb_sN_g\cdot u^2 -N_g\left(2\delb_s\left(s\gb^{a0}\right)u\delb_au + \delb_a\left(s\gb^{ab}\right)u\delb_bu\right). \endaligned \end{equation} We combine \eqref{eq 1 21-05-2017} and \eqref{eq 2 21-05-2017}, and see that \begin{equation}\label{eq 3 21-05-2017} \aligned s\left(\mathscr{K}_gu + N_gu\right)\cdot g^{\alpha\beta}\del_{\alpha}\del_{\beta}u =& \frac{1}{2}\delb_s\left(|\mathscr{K}_gu + N_gu|^2 - s^2\gb^{00}\gb^{ab}\delb_au\delb_bu\right) + \delb_a(w_g^a) \\ & + R_g(\nabla u,\nabla u) + S_g[\nabla u]\cdot (\mathscr{K}_g+N_g)u + T_g[u] \endaligned \end{equation} where $R_g(\nabla u,\nabla u)$ is a quadratic form acting on the gradient of $u$: $$ \aligned R_g(\nabla u,\nabla u) :=& s\left(\gb^{00}\gb^{ab} + s\delb_c\left(\gb^{0c}\gb^{ab}\right) - 2s\delb_c\gb^{0a}\cdot\gb^{cb}\right)\delb_au\delb_bu \\ &+(sg^{\alpha\beta}\del_{\alpha}\Psib_{\beta}^0-\delb_s(s\gb^{00}))\mathscr{K}_gu\cdot \delb_su - N_gs\gb^{\alpha\beta}\delb_{\alpha}u\delb_{\beta}u, \\ & + \frac{s^2}{2}\delb_s\left(\gb^{00}\gb^{ab}\right)\delb_au\delb_bu - s^2\delb_a\gb^{00}\gb^{ab}\delb_su\delb_bu \\ =&sL_g^{ab}\delb_a u \delb_bu + N_g \mathscr{K}_gu\cdot \delb_su - N_gs\gb^{\alpha\beta}\delb_{\alpha}u\delb_{\beta}u \\ & + \frac{s^2}{2}\delb_s\left(\gb^{00}\gb^{ab}\right)\delb_au\delb_bu - s^2\delb_a\gb^{00}\gb^{ab}\delb_su\delb_bu \endaligned $$ with $$ L_g^{ab}: = \gb^{00}\gb^{ab} + s\delb_c\left(\gb^{0c}\gb^{ab}\right) - 2s\delb_c\gb^{0a}\cdot\gb^{cb}, $$ $$ \aligned \mathscr{K}_g u\cdot S_g[\nabla u] :=& - (\mathscr{K}_g + N_g)u\cdot\left(2\delb_s(s\gb^{a0})\delb_au + s\delb_a\gb^{ab}\delb_b u\right) \endaligned $$ and $$ \aligned T_g[u] := -\delb_sN_g\cdot u\left(\mathscr{K}_g+N_g\right)u - su\cdot \gb^{ab}\delb_aN_g\delb_b u. \endaligned $$ In \eqref{eq 3 21-05-2017}, $$ \aligned w_g^a =& v_g^a + N_gsu\cdot \gb^{ab}\delb_bu \\ =& s\mathscr{K}_g\cdot \gb^{ab}\delb_b u - s^2\gb^{a0}\gb^{cb}\delb_cu\delb_bu + N_gsu\cdot \gb^{ab}\delb_bu. \endaligned $$ We analyse the structure of $R_g(\nabla u,\nabla u)$: remark the coefficient $L_g^{ab}$ satisfies the following property: when $g^{\alpha\beta} = m^{\alpha\beta}$, $$ L_m^{ab} = 2\mb^{ab}. $$ We also see that $$ \aligned L_g^{ab} - L_m^{ab} =& \gb^{00}\gb^{ab} - \mb^{00}\mb^{ab} + s\delb_c\left(\gb^{0c}\gb^{ab} - \mb^{0c}\mb^{ab}\right) - 2s\delb_c\gb^{0a}\gb^{cb} + 2s\delb_c\mb^{0a}\mb^{cb}\\ \\ =&\gb^{00}\hb^{ab} + \hb^{00}\mb^{ab} + s\delb_c\left(\hb^{0c}\gb^{ab}\right) + s\delb_c\left(\hb^{ab}\mb^{0c}\right) - 2s\left(\delb_c\gb^{0a}\hb^{cb} + \delb_c\hb^{0a}\mb^{cb}\right) \endaligned $$ On the other hand, we see that $$ N_m = 2 $$ and $$ N_g - N_m = h^{00} - \sum_ah^{aa} - 2\hb^{00} - s\delb_s\hb^{00} $$ and $$ \aligned sN_g\gb^{\alpha\beta}\delb_{\alpha}u\delb_{\beta}u =& sN_g\left(\gb^{00}\delb_su + 2\gb^{a0}\delb_a u\right)\delb_su + sN_g\gb^{ab}\delb_au\delb_bu =N_g\mathscr{K}_gu \delb_su + sN_g\gb^{ab}\delb_au\delb_bu. \endaligned $$ Thus we see that \begin{equation}\label{eq R(du,du)} \aligned R_g(\nabla u,\nabla u) =& s\left(L_g^{ab}-L_m^{ab}\right)\delb_au\delb_bu - s\left(N_g\gb^{ab}-2\mb^{ab}\right)\delb_au\delb_bu \\ & + \frac{s^2}{2}\delb_s\left(\gb^{00}\gb^{ab}\right)\delb_au\delb_bu - s^2\delb_a\gb^{00}\gb^{ab}\delb_su\delb_bu \\ =& s\left(L_g^{ab}-L_m^{ab}\right)\delb_au\delb_bu - s(N_g-N_m)\gb^{ab}\delb_au\delb_bu - 2s\hb^{ab}\delb_au\delb_bu \\ &+ \frac{s^2}{2}\delb_s\left(\gb^{00}\gb^{ab}\right)\delb_au\delb_bu - s^2\delb_a\gb^{00}\gb^{ab}\delb_su\delb_bu. \endaligned \end{equation} \subsection{Energy estimate in curved case} We integrate the identity \eqref{eq 2 21-05-2017} in the region $\Kcal_{[s_0,s]}$, remark that we suppose that $u$ is sufficiently regular and vanishes near the conical boundary. By Stokes formula: \begin{equation}\label{eq 1 23-05-2017} \aligned \int_{\Kcal_{[s_0,s]}} s\left(\mathscr{K}_gu + N_gu\right)\cdot g^{\alpha\beta}\del_{\alpha}\del_{\beta}u \ dxds =& \frac{1}{2}\left(E_{\text{con},g}(s,u) - E_{\text{con},g}(s_0,u)\right) \\ &+ \int_{\Kcal_{[s_0,s]}}\left(R_g(\nabla u,\nabla u) + \mathscr{K}_g u\cdot S_g[\nabla u] + T_g[u]\right)\,dxds \endaligned \end{equation} with $$ E_{\text{con},g} := \int_{\Hcal_s}\left(|\mathscr{K}_gu + N_g u|^2 - s^2\gb^{ab}\delb_au\delb_bu\right)\,dxds, $$ called the {\bf curved energy}. Then we derive \eqref{eq 1 23-05-2017} with respect to $s$ and see that $$ \aligned \int_{\Hcal_s}s\left(\mathscr{K}_gu + N_gu\right)\cdot g^{\alpha\beta}\del_{\alpha}\del_{\beta}u \ dx =& \frac{d}{2ds}E_{\text{con},g}(s,u) \\ &+ \int_{\Hcal_s}\left(R_g(\nabla u,\nabla u) + \mathscr{K}_gu\cdot S_g[u] + T_g[u]\right)\ dx \endaligned $$ thus \begin{equation}\label{eq 2 23-05-2017} \aligned E_{\text{con},g}(s,u)^{1/2}\frac{d}{ds}E_{\text{con},g}(s,u)^{1/2} \leq& \|s\left(\mathscr{K}_gu + N_gu\right)\cdot g^{\alpha\beta}\del_{\alpha}\del_{\beta}u\|_{L^1(\Hcal_s)} \\ & + \|R_g(\nabla u,\nabla u) + \mathscr{K}_g u\cdot S_g[u] + T_g[u]\|_{L^1(\Hcal_s)} \endaligned \end{equation} We suppose that there exists a $\kappa\geq 1$ such that \begin{equation}\label{eq 3 23-05-2017} \kappa^{-1} E_{\text{con},g}(s,u)^{1/2}\leq \Ec(s,u)^{1/2} \leq \kappa E_{\text{con},g}(s,u)^{1/2} , \end{equation} \begin{equation}\label{eq 4 23-05-2017} \|R_g(\nabla u,\nabla u) + \mathscr{K}_gu\cdot S_g[u] + T_g[u]\|_{L^1(\Hcal_s)}\leq \Ec(s,u)^{1/2}M_g(s,u). \end{equation} Then we see that from \eqref{eq 2 23-05-2017} $$ \frac{d}{ds}E_{\text{con},g}(s,u)^{1/2}\leq \|sg^{\alpha\beta}\del_{\alpha}\del_{\beta}u\|_{L^2(\Hcal_s)} + M_g(s,u) $$ Thus \begin{equation}\label{eq 5 23-05-2017} E_{\text{con},g}(s,u)^{1/2}\leq E_{\text{con},g}(s_0,u)^{1/2} + \int_{s_0}^s\left(\tau\|g^{\alpha\beta}\del_{\alpha}\del_{\beta}u\|_{L^2(\Hcal_\tau)} + M_g(\tau,u)\right)\, d\tau \end{equation} which (combined with \eqref{eq 3 23-05-2017}) leads to \begin{equation}\label{eq energy} \Ec(s,u)^{1/2}\leq \kappa^2\Ec(s_0,u)^{1/2} + \kappa\int_{s_0}^s\left(\tau\|g^{\alpha\beta}\del_{\alpha}\del_{\beta}u\|_{L^2(\Hcal_\tau)} + M_g(\tau,u)\right)\, d\tau. \end{equation} which is the conformal energy estimate in curved case. \subsection{Analysis on curved energy} In this subsection we will analyse the structure of the curved energy and more precisely, we will give a sufficient condition for \eqref{eq 3 23-05-2017} . \begin{proposition}\label{prop 1 24-05-2017} There exists a constant $\vep_s$ (with the $s$ stands for ``structure'') which depends only on $n$ such that when \begin{equation}\label{eq 1 24-05-2017} \left\{ \aligned &|\hb^{00}|\leq (s/t)\vep_s,\quad |s\delb_s\hb^{00}|\leq \vep_s(s/t), \\ &|h^{\alpha\beta}|\leq (s/t)\vep_s. \endaligned \right. \end{equation} then \eqref{eq 3 23-05-2017} holds. \end{proposition} \begin{proof} Recall the structure of the curved energy, we consider first the term $\gb^{ab}\delb_au\delb_bu$. We see that $$ \gb^{ab} = g^{\alpha\beta}\Psib_{\alpha}^a\Psib_{\beta}^b = g^{ab}, $$ in the same way $\hb^{ab} = h^{ab}$. Thus there exists a positive constant $\vep_s$ such that if $|h^{ab}|\leq \vep_s$, \begin{equation}\label{eq 1 16-08-2017} \big\|s^2\gb^{ab}\delb_au\delb_bu - \sum_a|s\delb_au|^2\big\|_{L^2(\Hcal_s)}\leq C\vep_s \left\|\sum_a|s\delb_au|^2\right\|_{L^2(\Hcal_s)} \leq C\vep_s \Ec(s,u)^{1/2}. \end{equation} Then we regard the first term in $E_{\text{con},g}$. We first remark that $$ \aligned N_g - 2 =& g^{00} - \sum_ag^{aa} - 2\gb^{00} - s\delb_s\gb^{00} - \left(m^{00} - \sum_am^{aa} - 2\mb^{00} - s\delb_s\mb^{00}\right) \\ =& h^{00} - \sum_ah^{aa} - 2\hb^{00} - s\delb_s\hb^{00}. \endaligned $$ Then we see that (taking \eqref{eq 1 24-05-2017} with $\vep_s$ sufficiently small) \begin{equation}\label{eq 1 18-08-2017} |N_g|\leq 2 + \left|h^{00} - \sum_ah^{aa} - 2\hb^{00} - s\delb_s\hb^{00}\right|\leq 3. \end{equation} We see that \begin{equation}\label{eq 2 24-05-2017} \aligned \mathscr{K}_g u + N_g u - \gb^{00}\left(Ku + 2u\right) =& 2s\left(\gb^{a0} -\gb^{00}\mb^{a0} \right)\delb_au + \left(N_g - 2\gb^{00}\right)u \\ =& 2s\left(\gb^{a0} - \mb^{a0} + \mb^{00}\mb^{a0} - \gb^{00}\mb^{a0}\right)\delb_au + \left(N_g - 2\gb^{00}\right)u \\ =&2s\left(\hb^{a0}+\hb^{00}\mb^{a0}\right)\delb_au + \left(N_g - 2\gb^{00}\right)u \endaligned \end{equation} where we remark that $$ \hb^{\alpha\beta} = \gb^{\alpha\beta} - \mb^{\alpha\beta},\quad \text{and especially}\quad \hb^{00} = \gb^{00} - 1. $$ Thus when $|\hb^{a0}|\leq \vep_s$ and $|\hb^{00}|\leq \vep_s (s/t)$, we see that for the first term in right-hand-side of \eqref{eq 2 24-05-2017}: $$ \aligned 2s\left|\hb^{a0}+\hb^{00}\mb^{a0}\right|\delb_au \leq & C\vep_s\sum_a|s\delb_a u|. \endaligned $$ We remark that $$ \hb^{a0} = h^{\alpha\beta}\Psib_{\alpha}^a\Psib_{\beta}^0 = h^{a\beta}\Psib_{\beta}^0 = h^{a0}(t/s) - h^{ab}(x^b/s) $$ thus by \eqref{eq 1 24-05-2017}, $|\hb^{a0}|\leq C\vep_s$. On the other hand, we see that by \eqref{eq 3 25-05-2017}, $$ \aligned \left(N_g - 2\gb^{00}\right)u =& \left(-2\hb^{00} + h^{00} - \sum_ah^{aa} - 2\hb^{00} - s\delb_s\hb^{00}\right)u \\ =& \left(-4\hb^{00} + h^{00} - \sum_a\hb^{aa} - s\delb_s\hb^{00} \right)u. \endaligned $$ We see that under the assumption \eqref{eq 1 24-05-2017}, $$ \left|\left(N_g - 2\gb^{00}\right)u\right|\leq C\vep_s (s/t)|u| $$ thus $$ \left|\mathscr{K}_g u + N_g u - \gb^{00}\left(Ku + 2u\right)\right|\leq C\vep_s\sum_a|s\delb_au| + C\vep_s (s/t)|u|. $$ Then we see that \begin{equation}\label{eq 3 16-08-2017} \aligned \left|(\mathscr{K}_gu + N_gu) - (Ku+2u)\right|\leq& \left|(\mathscr{K}_gu + N_gu) - \gb^{00}(Ku+2u)\right| + |\hb^{00}||Ku+2u| \\ \leq & C\vep_s\sum_a|s\delu_au| + C\vep_s(s/t)|u| + C\vep_s|Ku+2u| \endaligned \end{equation} This leads to (recall \eqref{eq 6 23-05-2017}): \begin{equation}\label{eq 2 16-08-2017} \left\|(\mathscr{K}_gu + N_gu) - \left(Ku + 2u\right)\right\|_{L^2(\Hcal_s)}\leq C\vep_s\Ec(s,u)^{1/2} \end{equation} Then we see that (by \eqref{eq 1 16-08-2017} and \eqref{eq 2 16-08-2017}) \begin{equation}\label{eq 4 16-08-2017} (1-C\vep_s)\Ec(s,u)^{1/2}\leq E_{\text{con},g}(s,u)^{1/2} \leq (1+C\vep_s)\Ec(s,u)^{1/2} \end{equation} Then we take $\vep_s$ sufficiently small and the desired result is established. \end{proof} \section{Commutators and decay estimate} \subsection{Global Sobolev inequality on hyperboloids} In this subsection we recall the Klainerman-Sobolev inequality on hyperboloids due to H\"ormander \cite{H1}: \begin{proposition} Let $u$ be a sufficiently regular function defined in $\Kcal$ and vanishes near the conical boundary $\del_K\Kcal$. Then the following estimate holds: \begin{equation}\label{ineq 1 sobolev} \sup_{\Hcal_s}\big(t^{3/2}|u|\big)\leq C\sum_{|I|+|J|\leq 2}\|\del^IL^Ju\|_{L^2(\Hcal_s)}. \end{equation} Here $C$ is a positive constant independent of $u$. \end{proposition} This result is essentially due to H\"ormander (see in detail \cite{H1}(1997), Lemma 7.6.1). For this slightly improved version, see \cite{PLF-MY-book} section 5.1. This inequality helps us to get decay estimate via $L^2$ norm. So in the following we need to control on hyperboloids the $L^2$ norm of the following quantities for $|I'|+|J'|\leq 2$ $$ \del^{I'}L^{J'}\left((s^2/t)\delb_s\del^IL^J u\right),\quad \del^{I'}L^{J'}\left(s\delb_a\del^IL^J u\right),\quad \del^{I'}L^{J'}\left((s/t)\del^IL^Ju\right), $$ $$ \del^{I'}L^{J'}\left(s^2\delb_s\delb_a\del^IL^Ju\right),\quad \del^{I'}L^{J'}\left(st\delb_a\delb_b\del^IL^Ju\right), $$ $$ \del^{I'}L^{J'}\left((s^2/t)\del^IL^J\delb_s u\right),\quad \del^{I'}L^{J'} \left((s/t)\del^IL^Ju\right),\quad \del^{I'}L^{J'}\left(s\del^IL^J\delb_au\right). $$ and $$ \del^{I'}L^{J'}\left(s^2\del^IL^J\delb_s\delb_au\right),\quad\del^{I'}L^{J'}\left(st\del^IL^J\delb_a\delb_b u\right). $$ These are to be bounded by the conformal energies $\Ec(s,\del^{I''}L^{J''}u)$ with $|I''|+|J''|\leq |I|+|J|+2$. To do so, we need some estimates on commutators, which are studied in the following sections. \subsection{Commutators I} In this section we calculate the following quantities: $$ [L_a,\del_{\alpha}], \quad [L_a,\del^I],\quad [L^J,\del^I], $$ $$ [\del_{\alpha},\delb_\alpha],\quad [\del^I,\delb_{\alpha}] $$ and $$ [L_a,\delb_{\alpha}],\quad [L^J,\delb_{\alpha}]. $$ We begin with the first group. It is easy to see that $$ [L_a,\del_t] = -\del_a,\quad [L_a,\del_b] = -\delta_{ab}\del_b $$ and we denote by $$ [L_a,\del_{\alpha}] = \theta_{a\alpha}^\beta \del_{\beta} $$ where $\theta_{a\alpha}^{\beta}$ are constants. {\bf Important convention:} in the following we often make summation over multi-indices of order less than an integer. For the convenience of expression, we only give the upper bound of the order of a multi-index and omit the lower bound. For example when we write $$ \sum_{|I|\leq N} $$ we always mean $$ \sum_{0\leq|I|\leq N} $$ which is a sum over all multi-index of order from zero to $N$. In certain case, we take the sum from a positive order. In this case we will write $$ \sum_{n\leq |I|\leq N}. $$ We make the convention that when $n>N$, this sum is taken as zero. Then we have the following decompositions: \begin{lemma}\label{lem 1 19-08-2017} Let $u$ be a function defined in $\Kcal$, sufficiently regular. Then the following identities hold: \begin{equation}\label{eq 1 19-08-2017} [L_a,\del^I]u = \sum_{|I'|=|I|}\theta_{aI'}^{I}\del^{I'}u, \end{equation} \begin{equation}\label{eq 2 19-08-2017} [L^J,\del^I]u = \sum_{|I'|=|I|\atop |J'|<|J| }\theta_{I'J'}^{JI}\del^{I'}L^{J'}u, \end{equation} \begin{equation}\label{eq 6 22-08-2017} [L^J,\del_\alpha]u = \sum_{|J'|<|J|}\theta_{\alpha J'}^{J\beta}\del_{\beta}L^{J'}. \end{equation} where $\theta_{aI'}^{I}$ ,$\theta_{I'J'}^{JI}$ and $\theta_{\alpha J'}^{J\beta}$ are constants. \end{lemma} \begin{proof} These are by induction. For the first it is an induction on $|I|$. When $|I|=1$ it is already proved. We suppose that \eqref{eq 1 19-08-2017} holds for $|I|\leq m$. Then we consider $$ \aligned \,[L_a,\del_{\alpha}\del^I]u =& [L_a,\del_{\alpha}]\del^Iu + \del_{\alpha}[L_a,\del^I]u \\ =&\theta_{a\alpha}^{\beta}\del_{\beta}\del^Iu + \sum_{|I'|=|I|}\del_{\alpha}\left(\theta_{aI'}^I\del^{I'}u\right) =\theta_{a\alpha}^{\beta}\del_{\beta}\del^Iu + \sum_{|I'|=|I|}\theta_{aI'}^I\del_{\alpha}\del^{I'}u. \endaligned $$ which concludes \eqref{eq 1 19-08-2017} for $|I| = m+1$. For \eqref{eq 2 19-08-2017}, it is an induction on $|J|$. For $|J|=1$ it is already proved. Suppose that \eqref{eq 2 19-08-2017} holds for $|J|\leq m$. Then we consider $$ \aligned \,[L^J L_a,\del^I]u =& L^J[L_a,\del^I]u + [L^J,\del^I]L_au =L^J\left(\sum_{|I'|=|I|}\theta_{aI'}^{I}\del^{I'}u\right) + \sum_{|I'|=|I|\atop |J'|<|J|}\theta^{JI}_{I'J'}\del^{I'}L^{J'}L_au \\ =&\sum_{|I'|=|I|}\theta_{aI'}^IL^J\del^{I'}u + \sum_{|I'|=|I|\atop |J'|<|J|}\theta^{JI}_{I'J'}\del^{I'}L^{J'}L_au \\ =&\sum_{|I'|=|I|}\theta_{aI'}^I\del^{I'}L^Ju + \sum_{|I'|=|I|}\theta_{aI'}^I[L^J,\del^{I'}]u + \sum_{|I'|=|I|\atop |J'|<|J|}\theta^{JI}_{I'J'}\del^{I'}L^{J'}L_au \\ =&\sum_{|I'|=|I|}\theta_{aI'}^I\del^{I'}L^J u + \sum_{|I'|=|I|}\theta_{aI'}^I\sum_{|I''|=|I'|\atop |J''|<|J|}\theta_{I''J''}^{JI'}\del^{I''}L^{J''} u + \sum_{|I'|=|I|\atop |J'|<|J|}\theta^{JI}_{I'J'}\del^{I'}L^{J'}L_au \\ =&\sum_{|I'|=|I|}\theta_{aI'}^I\del^{I'}L^Ju + \sum_{{|I'|=|I|,|I''|=|I'|}\atop |J''|<|J|}\theta_{aI'}^I\theta_{I''J''}^{JI'}\del^{I''}L^{J''}u + \sum_{|I'|=|I|\atop |J'|<|J|}\theta^{JI}_{I'J'}\del^{I'}L^{J'}L_au. \endaligned $$ This concludes \eqref{eq 2 19-08-2017} for $|J| = m+1$. For \eqref{eq 6 22-08-2017} is a direct result of \eqref{eq 6 22-08-2017}. \end{proof} Here we also state a simple but frequently used result:\begin{lemma}\label{lem 2 29-08-2017} The following identity holds: \begin{equation}\label{eq 6 29-08-2017} \del^{I_1}L^{J_1}\del^{I_2}L^{J_2}u = \sum_{|I|= |I_1|+|I_2|\atop |J|\leq|J_1|+|J_2|}\zeta^{I_1I_2J_1J_2}_{IJ}\del^IL^Ju \end{equation} where $\zeta^{I_1I_2J_1J_2}_{IJ}$ are constants \end{lemma} \begin{proof} This is by \eqref{eq 2 19-08-2017}. We see that $$ \aligned \del^{I_1}L^{J_1}\del^{I_2}L^{J_2}u =& \del^{I_1}\del^{I_2}L^{J_1}L^{J_2}u + \del^{I_1}\left([L^{J_1},\del^{I_2}]L^{J_2}u\right) \\ =& \del^{I_1}\del^{I_2}L^{J_1}L^{J_2}u + \sum_{|I_2'|=|I_2|\atop |J_1'|<|J_1|}\theta_{I_2'J_1'}^{J_1I_2}\del^{I_1}\del^{I_2'}L^{J_1'}L^{J_2}u \endaligned $$ where $\theta_{I_2'J_1'}^{J_1I_2}$ are constants. This proves the desired result. \end{proof} Now we consider the commutator of $[\del_{\alpha},\delb_a]$. For the convenience of expression, we introduce the following notion of homogeneous function. Let $u$ be a $C^{\infty}$ function defined in $\{r<t\}$ and satisfies the following condition: \begin{equation}\label{eq 2 26-06-2017} u(\lambda,\lambda x) = \lambda^{n} u(1,x),\quad n\in \mathbb{Z}. \end{equation} for all $\lambda>0$ and $\del^I u(1,x)$ are bounded in $\{|x|\leq 1\}$. Such function $u$ is called a {\bf homogeneous function of degree $n$}. For example $x^a/t$ is a homogeneous function of degree zero. We state the following property of a homogeneous function of degree $n$: \begin{lemma}\label{lem 1 26-06-2017} Let $u$ be a homogeneous function of degree $n$. Then $\del^IL^Ju$ is homogeneous of degree $n-|I|$, and the following estimates holds in $\Kcal$: \begin{equation}\label{eq 1 lem 1 26-06-2017} \left|\del^IL^J u\right|\leq Ct^{n-|I|} \end{equation} with $C$ a constant determined by $u, I$ and $J$. \end{lemma} \begin{proof} We remak that \begin{equation}\label{eq 1 proof lem 1 26-06-2017} \left|\del^Iu\right|\leq Ct^{n-|I|} \end{equation} which is checked directly: $$ \aligned u(t,x) =& t^nu(1,x/t) \Rightarrow \del_t u(t,x) = nt^{n-1}u(1,x/t) + t^n(-x^a/t^2)\del_au(1,x/t) \\ =& t^{n-1}\left(nu(1,x/t) - (x^a/t)\del_au(1,x/t)\right) \endaligned $$ and $$ \del_au = t^n(-x^a/t^2)\del_au(1,x/t) = -t^{n-1}(x^a/t)\del_au(1,x/t) $$ where we remark that $$ nu(1,x/t) - (x^a/t)\del_au(1,x/t),\quad (x^a/t)\del_au(1,x/t) $$ are homogeneous of degree zero. Thus $\del_\alpha u$ is homogeneous of degree $n-1$. Then by recurrence we see that $\del^I u$ is homogeneous of degree $n-|I|$. This leads to \eqref{eq 1 proof lem 1 26-06-2017}. Then we prove that \begin{equation}\label{eq 2 proof lem 1 26-06-2017} L^Ju \quad \text{is homogeneous of degree n} \end{equation} This is also checked directly by $$ L_a u(t,x) = \left(t\del_a+x^a\del_t\right) u(t,x) = -t^{n}(x^a/t)\del_au(1,x/t) + t^n(x^a/t)\left(nu(1,x/t) - (x^a/t)\del_au(1,x/t)\right) $$ which is homogeneous of degree $n$. Thus $L^Ju$ is also homogeneous of degree zero, which leads to \eqref{eq 2 proof lem 1 26-06-2017}. The desired result is a combination of \eqref{eq 1 proof lem 1 26-06-2017} and \eqref{eq 2 proof lem 1 26-06-2017}. \end{proof} Now we see that $$ [\del_t,\delb_a] = -\frac{x^a}{t^2}\del_t,\quad [\del_a,\delb_b] = \frac{\delta_{ab}}{t}\del_t, $$ Then we denote by $$ [\del_{\alpha},\delb_b] = \sigma_{\alpha b}\del_t $$ with $\sigma_{\alpha\beta}$ homogeneous functions of degree $-1$. Then we establish the following result: \begin{lemma}\label{lem 2 19-08-2017} For $|I|\geq 1$, \begin{equation}\label{eq 3 19-08-2017} [\del^I,\delb_a] = \sum_{1\leq |J|\leq |I|}\sigma_{aJ}^I\del^J \end{equation} with $\sigma_{aJ}^I$ a homogeneous function of degree $|J|-|I|-1$. \end{lemma} \begin{proof} This is by induction on $|I|$. For $|I|=1$ we see that holds. We suppose that \eqref{eq 3 19-08-2017} holds for $|I|\leq m$, we consider $$ \aligned \,[\del_{\alpha}\del^I,\delb_a] =& \del_{\alpha}\left(\sum_{1\leq |J|\leq |I|}\sigma_{aJ}^I\del^J\right) + [\del_{\alpha},\delb_a]\del^I \\ =& \sum_{1\leq |J|\leq |I|}\sigma_{aJ}^I\del_{\alpha}\del^J + \sum_{1\leq|J|\leq|I|}\del_{\alpha}\sigma_{aJ}^I\del^J + \sigma_{\alpha a}\del_t\del^I. \endaligned $$ We see that $\del_{\alpha}\del^J$ is of order $|J|+1$, $\del^t\del^I$ is of order $|I|+1$; we recall that $\del_{\alpha}\sigma_{aJ}^I$ is homogeneous of order $|J|-|I|-2 = |J| - (|I|+1)-1$ (by the assumption of induction combined with lemma \ref{lem 1 26-06-2017}), and $\sigma_{\alpha a}$ is homogeneous of degree $-1$. Thus we see that the \eqref{eq 3 19-08-2017} is proved in the case $|I| = m+1$. \end{proof} Now we calculate: $$ [L_a,\delb_b] = -\frac{x^b}{t}\delb_a = \Psib_b^0\delb_a. $$ We denote by $$ [L_a,\delb_b] = \eta_{b}\delb_a $$ where $\eta_b$ is a homogeneous function of degree zero. Now we establish the following result: \begin{lemma}\label{lem 1 21-08-2017} Let $|I| = n$, then \begin{equation}\label{eq 2 21-08-2017} [L^I,\delb_a] = \sum_{|J|<|I|}\eta_{aJ}^{Ib}\delb_bL^J \end{equation} where $\eta_{aJ}^{Ib}$ are homogeneous functions of degree zero. \end{lemma} \begin{proof} This is by induction on $|I|$. We see that for $|I|=1$ \eqref{eq 2 21-08-2017} holds. Suppose that \eqref{eq 2 21-08-2017} holds for $|I|\leq m$, now we consider $$ \aligned \,[L_aL^I,\delb_b] =& L_a([L^I,\delb_b]) + [L_a,\delb_b]L^I = L_a\left(\sum_{|I'|<|I|}\eta_{bI'}^{Ic}\delb_cL^{I'}\right) + \eta_b\delb_aL^{I} \\ =& \sum_{|I'|<|I|}L_a\eta_{bI'}^{Ic}\delb_cL^{I'} + \sum_{|I'|<|I|}\eta_{bI'}^{Ic}L_a\delb_cL^{I'} + \eta_b\delb_aL^{I} \\ =& \sum_{|I'|<|I|}L_a\eta_{bI'}^{Ic}\delb_cL^{I'} + \sum_{|I'|<|I|}\eta_{bI'}^{Ic}\delb_c L_a L^{I'} + \sum_{|I'|<|I|}\eta_{bI'}^{Ic}[L_a,\delb_c]L^{I'} + \eta_b\delb_aL^{I} \\ =& \sum_{|I'|<|I|}L_a\eta_{bI'}^{Ic}\delb_cL^{I'} + \sum_{|I'|<|I|}\eta_{bI'}^{Ic}\delb_c L_a L^{I'} + \sum_{|I'|<|I|}\eta_{bI'}^{Ic}[L_a,\delb_c]L^{I'} + \eta_b\delb_aL^{I} \\ =& \sum_{|I'|<|I|}L_a\eta_{bI'}^{Ic}\delb_cL^{I'} + \sum_{|I'|<|I|}\eta_{bI'}^{Ic}\delb_c L_a L^{I'} + \sum_{|I'|<|I|}\eta_{bI'}^{Ic}\eta_a\delb_c L^{I'} + \eta_b\delb_aL^{I}. \endaligned $$ We recall that $\eta_{bI'}^{Ic}$ homogeneous of degree zero so $L^J\eta_{bI'}^{Ic}$ is again homogeneous of degree zero. This concludes \eqref{eq 2 21-08-2017} for the case $|I| = m+1$. \end{proof} Now we are ready to establish the following result: \begin{lemma}\label{lem 2 21-08-2017} \begin{equation}\label{eq 1 23-08-2017} [\del^IL^J,\delb_a] = \sum_{|I'|\leq|I|\atop |J'|<|J|}\rhob_{a I'J'}^{IJc}\delb_c\del^{I'}L^{J'} + \sum_{1\leq|I'|\leq |I|}\rho_{a I'}^{IJ}\del^{I'}L^{J} \end{equation} \end{lemma} where $\rhob_{a I'J'}^{IJc}$ are homogeneous functions of degree $(|I'|-|I|)$ and $\rho_{a I'}^{IJ}$ are homogeneous functions of degree $(|I'|-|I|-1)$. Furthermore, in $\Kcal$ for a function $u$ sufficiently regular, we have \begin{equation}\label{eq 2 23-08-2017} \left|[\del^IL^J,\delb_a]u\right|\leq C\sum_{{c,|I'|\leq|I|}\atop |J'|<|J|}|\delb_c\del^{I'}L^{J'}u| + Ct^{-1}\sum_{1\leq|I'|\leq |I|}\left|\del^{I'}L^J u\right|. \end{equation} \begin{proof} We remark that $$ \aligned \,[\del^IL^J,\delb_a] =& \del^I\left([L^J,\delb_a]\right) + [\del^I,\delb_a]L^J \\ =&\del^I\left(\sum_{|J'|<|J|}\eta_{aJ'}^{Jc}\delb_cL^{J'}\right) + \sum_{1\leq |I'|\leq |I|}\sigma_{aI'}^I\del^{I'}L^J \\ =&\sum_{I_1+I_2=I\atop |J'|<|J|}\del^{I_1}\eta_{aJ'}^{Jc}\cdot \del^{I_2}\delb_cL^{J'} + \sum_{1\leq |I'|\leq |I|}\sigma_{aI'}^I\del^{I'}L^J \\ =&\sum_{I_1+I_2=I\atop |J'|<|J|}\del^{I_1}\eta_{aJ'}^{Jc}\cdot \delb_c\del^{I_2}L^{J'} + \sum_{I_1+I_2=I\atop |J'|<|J|}\del^{I_1}\eta_{aJ'}^{Jc}[\del^{I_2},\delb_c]L^{J'} + \sum_{1\leq |I'|\leq |I|}\sigma_{aI'}^I\del^{I'}L^J \\ =&\sum_{I_1+I_2=I\atop |J'|<|J|}\del^{I_1}\eta_{aJ'}^{Jc}\cdot \delb_c\del^{I_2}L^{J'} + \sum_{I_1+I_2=I\atop |J'|<|J|}\del^{I_1}\eta_{aJ'}^{Jc}\left(\sum_{1\leq |I_2'|\leq |I_2|}\sigma_{cI_2'}^{I_2}\del^{I_2'}L^{J'}\right) \\ &+ \sum_{1\leq |I'|\leq |I|}\sigma_{aI'}^I\del^{I'}L^J \\ =&\sum_{I_1+I_2=I\atop |J'|<|J|}\del^{I_1}\eta_{aJ'}^{Jc}\cdot \delb_c\del^{I_2}L^{J'} + \sum_{{I_1+I_2=I,|J'|<|J|}\atop 1\leq |I_2'|\leq|I_2| }\del^{I_1}\eta_{aJ'}^{Jc}\cdot \sigma_{cI_2'}^{I_2}\del^{I_2'}L^{J'} \\ &+ \sum_{1\leq |I'|\leq|I|}\sigma_{aI'}^I\del^{I'}L^J. \endaligned $$ Now we recall that $\del^{I_1}\eta_{aJ'}^{Jc}$ is homogeneous of degree $-|I_1| = |I_2|-|I|$, $\del^{I_1}\eta_{aJ'}^{Jc}\cdot \sigma_{cI_2'}^{I_2}$ is homogeneous of degree $-|I_1| + |I_2'|-|I_2|-1 = |I_2'|-|I|-1$ and $\sigma_{aI'}^I$ is homogeneous of degree $|I'|-|I|-1$. Thus the desired result is established. \eqref{eq 2 23-08-2017} is direct by \eqref{eq 1 23-08-2017}. \end{proof} \subsection{Commutators II} In this subsection we consider the following quantities: $$ \del^IL^J(s/t), \quad \del^IL^J s. $$ And then based on these calculation, we analyse $[\del^I,\delb_s]$. We first remark the following result: \begin{equation}\label{eq 5 19-08-2017} \del_t (s/t) = \frac{r^2}{t^2} s^{-1}, \quad \del_a (s/t) = -\frac{x^a}{t} s^{-1}. \end{equation} We denote by $$ \del_{\alpha}(s/t) = \pi_{\alpha}s^{-1} $$ with $\pi_{\alpha}$ a homogeneous function of degree zero. \begin{equation}\label{eq 7 19-08-2017} \del_t s = \frac{t}{s}, \quad \del_a s = -\frac{x^a}{s} = -\frac{x^a}{t}(t/s). \end{equation} We denote by $$ \del_{\alpha}s = \rho_{\alpha}(s/t)^{-1} $$ with $\rho_{\alpha}$ a homogeneous function of degree zero. We also recall $$ L_a(s/t) = -\frac{x^a}{t}(s/t),\quad L_a s = 0. $$ We first establish the following relation: \begin{lemma}\label{lem 3 19-08-2017} \begin{equation}\label{eq 4 19-08-2017} L^J(s/t) = \lambda^J(s/t) \end{equation} with $\lambda^J$ a homogeneous function of degree zero. \end{lemma} \begin{proof} This is by induction. It is clear that \eqref{eq 4 19-08-2017} holds for $|J|=1$. Then we consider $$ L_a L^J(s/t) = L_a\left(\lambda^J(s/t)\right) = L_a\lambda^J \cdot (s/t) + \lambda^JL_a(s/t) = \left(L_a\lambda^J - (x^a/t)\right)(s/t). $$ We see that by \eqref{lem 1 26-06-2017}, $L_a\lambda^J$ is homogeneous of degree zero. Furthermore $x^a/t$ is also homogeneous of degree zero. Thus $\left(L_a\lambda^J - (x^a/t)\right)$ is homogeneous of degree zero. \end{proof} Then we establish the following result: \begin{lemma} For $|I|\geq 1$, \begin{equation}\label{eq 6 19-08-2017} \del^I (s/t) = \sum_{1\leq k\leq |I|}\pi^I_k(s/t)^{-k+1}s^{-k} \end{equation} with $\pi$ a sum of finite many homogeneous functions of degree $(k-|I|)$. Furthermore, in $\mathcal{K}$, \begin{equation}\label{eq 8 19-08-2017} \left|\del^I(s/t)\right|\leq Cs^{-1}. \end{equation} \end{lemma} \begin{proof} This is also by induction on $|I|$. For $|I|=1$, we see that it is established by \eqref{eq 5 19-08-2017}. We suppose that \eqref{eq 6 19-08-2017} holds for $|I|\leq m$, and we consider (where we use \eqref{eq 5 19-08-2017} and \eqref{eq 7 19-08-2017}) $$ \aligned \del_{\alpha}\del^I(s/t) =& \del_{\alpha}\left(\sum_{1\leq k\leq |I|}\pi^I_k(s/t)^{-k+1}s^{-k}\right) \\ =& \sum_{1\leq k\leq |I|}\del_{\alpha}(\pi^I_k)\cdot (s/t)^{-k+1}s^{-k} + \sum_{1\leq k\leq|I|}\pi^I_k\del_{\alpha}\left((s/t)^{-k+1}\right)\cdot s^{-k} \\ &+ \sum_{1\leq k\leq|I|}\pi^I_k(s/t)^{-k+1}\del_{\alpha}(s^{-k}) \\ =& \sum_{1\leq k\leq |I|}\del_{\alpha}(\pi^I_k)\cdot (s/t)^{-k+1}s^{-k} + \sum_{1\leq k\leq|I|}\pi^I_k(-k+1)(s/t)^{-k}\del_{\alpha}(s/t)\cdot s^{-k} \\ &+ \sum_{1\leq k\leq|I|}\pi^I_k(s/t)^{-k+1}(-k)s^{-k-1}\del_{\alpha}s \\ =&\sum_{1\leq k\leq |I|}\del_{\alpha}(\pi^I_k)\cdot (s/t)^{-k+1}s^{-k} + \sum_{1\leq k\leq|I|}(-k+1)\pi^I_k(s/t)^{-(k+1)-1}\pi_{\alpha}\cdot s^{-(k+1)} \\ &-\sum_{1\leq k\leq|I|}k\pi^I_k(s/t)^{-(k+1)+1}s^{-(k+1)}\rho_{\alpha} \\ =& \del_{\alpha}(\pi_1^I)s^{-1} + \sum_{2\leq k \leq |I|}\left(\del_{\alpha}(\pi_K^I)+(1-k)\pi_{k-1}^I\pi_{\alpha} + (1-k)\pi_{k-1}^I\rho_{\alpha}\right)(s/t)^{-k+1}s^{-k} \\ & + \left((1-|I|)\pi_{\alpha} - |I|\rho_{\alpha}\right)\pi_{|I|}^I(s/t)^{-(|I|+1)+1}s^{-(|I|+1)}. \endaligned $$ WE check that for each term the coefficients are homogeneous of degree $(k-(|I|+1))$, and this concludes the case where $|I| = m+1$. For \eqref{eq 8 19-08-2017}, we see that in \eqref{eq 6 19-08-2017}, $$ \left|\pi_k^I(s/t)^{-k+1}s^{-k}\right|\leq Ct^{k-1}s^{-2k+1} = Ct^{k-1}s^{-2k+2}s^{-1} = C(t/s^2)^{k-1} s^{-1}. $$ We remark that in $\mathcal{K}$, $t/s^2$ is bounded. Then \eqref{eq 8 19-08-2017} is established. \end{proof} Now we observe the quantity $\del^Is$ (for $|I|\geq 1$). We see that $$ \del^I(s) = \del^I(t\cdot (s/t)) = \sum_{I_1+I_2=I}\del^{I_1}t\cdot \del^{I_2}(s/t). $$ We see that for $|I_1|\geq 2$, $\del^{I_1}t = 0$. Thus we see $$ \sum_{I_1+I_2=I}\del^{I_1}t\cdot \del^{I_2}(s/t) = t\del^I(s/t) + \sum_{|I_1|=1\atop I_2+I_1=I}\del^{I_1}t\cdot \del^{I_2}(s/t) $$ where the second term does not exist when $|I|\leq 0$. Then combined with \eqref{eq 8 19-08-2017}, we see that \begin{equation}\label{eq 1 20-08-2017} |\del^I s|\leq\left\{ \aligned &C(t/s),\quad |I|\geq 1 \\ &s,\quad |I|=0. \endaligned \right. \end{equation} \begin{remark} In the following application, we see that in $\Kcal$, because $s^2\geq t$, we have $|\del^Is|\leq Cs$. Furthermore, we see that when $|J|\geq 1$, $$ \del^IL^J s = 0. $$ \end{remark} Combine \eqref{eq 4 19-08-2017} and \eqref{eq 6 19-08-2017}, we see that \begin{lemma} In $\Kcal$, \begin{equation}\label{eq 2 20-08-2017} |\del^IL^J(s/t)|\leq \left\{ \aligned &Cs^{-1},\quad |I|\geq 1, \\ &Cs/t,\quad |I|=0. \endaligned \right. \end{equation} Let $n\in\mathbb{N}^*$. Then \begin{equation}\label{eq 4 23-08-2017} |\del^IL^J \left((s/t)^n\right)|\leq \left\{ \aligned &Cs^{-1}(s/t)^{n-1},\quad |I|\geq 1, \\ &C(s/t)^n,\quad |I|=0. \endaligned \right. \end{equation} \end{lemma} \begin{proof} If $|I|=0$, by \eqref{eq 4 19-08-2017}, \eqref{eq 2 20-08-2017} is established. When $|I|\geq 1$, we denote by $\del^I = \del_{\alpha}\del^{I'}$ with $|I'|\geq 0$. Then $$ \aligned \del^IL^J(s/t) =& \del_{\alpha}\del^{I'}L^J(s/t) = \del_{\alpha}\del^{I'}\left(\lambda^J(s/t)\right) = \del_{\alpha}\left(\sum_{I_1'+I_2'=I'}\del^{I_1'}\lambda^J\cdot \del^{I_2'}(s/t)\right) \\ =& \sum_{I_1'+I_2'=I'}\del_{\alpha}\del^{I_1'}\lambda^J\cdot \del^{I_2'}(s/t) + \sum_{I_1'+I_2'=I'}\del^{I_1'}\lambda^J\cdot \del_{\alpha}\del^{I_2'}(s/t). \endaligned $$ Now we recall that $$ |\del^I(s/t)| \leq \left\{ \aligned &Cs^{-1},\quad |I|\geq 1, \\ &s/t,\quad |I|=0. \endaligned \right. $$ and the fact that $\big|\del_{\alpha}\del^{I_1'}\lambda^J\big|\leq Ct^{-1}\leq Cs^{-1}$. Thus \eqref{eq 2 20-08-2017} is established. For \eqref{eq 4 22-08-2017}, we see that $$ \del^IL^J\left((s/t)^n\right) = \sum_{I_1+\cdots I_n=I\atop J_1+\cdots J_n=J}\del^{I_1}L^{J_1}(s/t)\cdots \del^{I_n}L^{J_n}(s/t). $$ We apply \eqref{eq 2 20-08-2017} on each factor, \eqref{eq 4 23-08-2017} is established. \end{proof} Then we make the following estimate: $$ \del^IL^J(s^2/t) = \del^IL^J(s\cdot (s/t)) = \sum_{I_1+I_2=I\atop J_1+J_2=J}\del^{I_1}L^{J_1}s \cdot \del^{I_2}L^{J_2}(s/t). $$ Then we see that by when $|J_1|\geq 1$, we see that $\del^{I_1}L^{J_1}s=0$. So we see that $$ \del^IL^J(s^2/t) = \sum_{I_1+I_2=I\atop J_1+J_2=J}\del^{I_1}s \cdot \del^{I_2}L^{J}(s/t). $$ Now we see that when $|I_1|\geq 1$, we see that $$ |\del^{I_1}s \cdot \del^{I_2}L^{J}(s/t)|\leq C. $$ When $|I_1|=0$, suppose that $|I|\geq 1$, thus $$ |s\cdot\del^{I}L^{J}(s/t)| \leq C. $$ When $|I|=0$, we see that $$ |sL^J(s/t)\leq Cs^2/t. $$ Thus we conclude that \begin{equation}\label{eq 4 22-08-2017} |\del^IL^J(s^2/t)|\leq \left\{ \aligned &C,\quad |I|\geq 1, \\ &Cs^2/t,\quad |I|=0. \endaligned \right. \end{equation} Then we also establish the following result: \begin{lemma}\label{lem 4 31-08-2017} In $\Kcal$ the following bound holds: \begin{equation}\label{eq 3 31-08-2017} \left|\del^I (s^{-n})\right|\leq \left\{ \aligned &Cts^{-n-2},\quad |I|\geq 1, \\ &s^{-n},\quad |I| =0. \endaligned \right. \end{equation} and \begin{equation}\label{eq 3 02-09-2017} \left|\del^IL^J(t/s)^n\right|\leq \left\{ \aligned &C(t/s)^{n+1}s^{-1},\quad |I|\geq 1, \\ & (t/s)^n,\quad |I|=0. \endaligned \right. \end{equation} \end{lemma} \begin{proof} This is by the identity of Fa\`a di Bruno. We denote by $$ \aligned f:\RR^+ &\rightarrow \RR^+ \\ x &\rightarrow x^{-n}. \endaligned $$ Then we see that for $|I|\geq 1$, $$ \del^I f(s) = \sum_{1\leq k\leq |I|}\sum_{I_1+I_2+\cdots I_k = I}f^{(k)}(s)\del^{I_1}s\del^{I_2}s\cdots \del^{I_k}s. $$ We see that in the above expression, $$ \big|f^{(k)}(s)\big|\leq Cs^{-n-k}, \quad |\del^{I_j}s|\leq C(t/s) $$ This we see that $$ \big|f^{(k)}(s)\del^{I_1}s\del^{I_2}s\cdots \del^{I_k}s\big|\leq C s^{-k-n}(t/s)^k\leq Cs^{-n}(t/s^2)^k. $$ Recall that $k\geq 1$ and in $\Kcal$, $s^2\geq t$, \eqref{eq 3 31-08-2017} is established. For \eqref{eq 3 02-09-2017}, we see that $$ \aligned \del^IL^J\left(t^ns^{-n}\right) =& \sum_{I_1+I_2=I\atop J_1+J_2=J}\del^{I_1}L^{J_1}t^{n}\cdot\del^{I_2}L^{J_2}(s^{-n}). \endaligned $$ Then we see that $t^n$ is homogeneous of degree $n$ thus $\left|\del^{I_1}L^{J_1}t^{n}\right|\leq Ct^{n-|I_1|}$. Thus by \eqref{eq 3 31-08-2017}, \eqref{eq 3 02-09-2017} is proved. \end{proof} For simplicity of expression, we introduce the following notation: $$ \Lambda^{IJ}: = \del^IL^J(s/t). $$ Then by \eqref{eq 6 29-08-2017}, the following estimate is direct: \begin{equation}\label{eq 5 05-05-2017} \del^IL^J\Lambda^{I'J'}\leq \left\{ \aligned &Cs^{-1}\quad |I|+|I'|\geq 1, \\ &C(s/t)\quad |I|+|I'|=0. \endaligned \right. \end{equation} Now we are ready to calculate the commutator $[\del^IL^J,\delb_s]$. We have the following result: \begin{lemma}\label{lem 2 22-08-2017} Let $u$ be a function defined in $\Kcal$, sufficiently regular and vanishes near the conical boundary. Then the following estimate holds: \begin{equation}\label{eq 8 22-08-2017} \left|[\del^IL^J,\delb_s]u\right| \leq Cs^{-1}\sum_{|I'|<|I|}|\del_t\del^{I'}L^Ju| + C(s/t)\sum_{{\alpha,|I'|\leq|I|}\atop|J'|<J}|\del_\alpha\del^{I'}L^{J'}u|. \end{equation} where $C$ is a constant determined by $I,J$. \end{lemma} \begin{proof} We remark that $$ \aligned \,[\del^IL^J,\delb_s] =& [\del^IL^J,(s/t)\del_t] = \sum_{I_1+I_2=I,J_1+J_2=J\atop|I_2|+|J_2|<|I|+|J|}\Lambda^{I_1J_1}\del^{I_2}L^{J_2}\del_t + (s/t)[\del^IL^J,\del_t] \\ =&\sum_{I_1+I_2=I,J_1+J_2=J\atop|I_2|+|J_2|<|I|+|J|}\Lambda^{I_1J_1}\del_t\del^{I_2}L^{J_2} + \sum_{I_1+I_2=I,J_1+J_2=J\atop|I_2|+|J_2|<|I|+|J|}\Lambda^{I_1J_1}\del^{I_2}[L^{J_2},\del_t] + (s/t)\del^I[L^J,\del_t] \\ =&\sum_{I_1+I_2=I,J_1+J_2=J\atop|I_2|+|J_2|<|I|+|J|}\Lambda^{I_1J_1}\del_t\del^{I_2}L^{J_2} + \sum_{I_1+I_2=I\atop J_1+J_2=J}\Lambda^{I_1J_1}\del^{I_2}[L^{J_2},\del_t] \\ =&\sum_{I_1+I_2=I,J_1+J_2=J\atop|I_2|+|J_2|<|I|+|J|}\Lambda^{I_1J_1}\del_t\del^{I_2}L^{J_2} +\sum_{I_1+I_2=I, J_1+J_2=J\atop|J_2'|<|J_2|}\Lambda^{I_1J_1}\theta_{0J_2'}^{J_2\beta}\del_{\beta}\del^{I_2}L^{J_2'}. \endaligned $$ We decompose the first term in right-hand-side and see that $$ \aligned \,[\del^IL^J,\delb_s] =&\sum_{I_1+I_2=I\atop|I_2|<|I|}\Lambda^{I_1O}\del_t\del^{I_2}L^{J} \\ &+\sum_{I_1+I_2=I,J_1+J_2=J\atop|J_2|<|J|}\Lambda^{I_1J_1}\del_t\del^{I_2}L^{J_2} +\sum_{I_1+I_2=I, J_1+J_2=J\atop|J_2'|<|J_2|}\Lambda^{I_1J_1}\theta_{0J_2'}^{J_2\beta}\del_{\beta}\del^{I_2}L^{J_2'}. \endaligned $$ Were $\Lambda^{IO}$ means in $\del^I(s/t)$ where $|J|=0$. Recall the bound on $\Lambda^{IJ}$, the desired result is direct. \end{proof} Finally we establish the following estimates: \begin{lemma}\label{lem 6 31-08-2017} Let $u$ be a function defined in $\Kcal$, sufficiently regular, then \begin{equation}\label{eq 8 31-08-2017} \big|[\del^IL^J,\delb_s\delb_s]u\big|\leq C\sum_{\alpha,\beta,|I'|\leq|I|,|J'|\leq|J|\atop |I'|+|J'|<|I|+|J|}|(s/t)^2\del_{\alpha}\del_{\beta}\del^{I'}L^{J'}u| + C\sum_{\alpha,|I'|\leq|I|,|J'|\leq|J|\atop |I'|+|J'|<|I|+|J|}|t^{-1}\del_{\alpha}\del^{I'}L^{J'}u|. \end{equation} \end{lemma} \begin{proof} We remark that $$ \delb_s\delb_s = (s/t)^2\del_t\del_t + t^{-1}(r/t)^2\del_t. $$ Then we see that $$ [\del^IL^J,\delb_s\delb_s]u = [\del^IL^J,(s/t)^2\del_t\del_t]u + [\del^IL^J ,t^{-1}(r/t)^2\del_t]u =: T_1+T_2 $$ We see that $$ [\del^IL^J,(s/t)^2\del_t\del_t]u = \sum_{{I_1+I_2=I, J_1+J_2=J}\atop |I_2|+|J_2|\leq |I|+|J|-1}\del^{I_1}L^{J_1}(s/t)^2\del^{I_2}L^{J_2}\del_t\del_tu + (s/t)^2[\del^IL^J,\del_t\del_t]u. $$ We see that by \eqref{eq 6 22-08-2017}: \begin{equation}\label{eq 2 01-09-2017} \aligned \,[\del^IL^J,\del_t\del_t]u =& \del_t\left([\del^{I}L^{J},\del_t]u\right) + [\del^{I}L^{J},\del_t]\del_tu \\ =&\sum_{|J'|<|J|}\theta_{0J'}^{J\beta}\del_t\del_{\beta}\del^{I}L^{J'}u + \sum_{|J'|<|J|}\theta_{0J'}^{J\beta}\del_{\beta}\del^{I}L^{J'}\del_tu \\ =&2\sum_{|J'|<|J|}\theta_{0J'}^{J\beta}\del_t\del_{\beta}\del^{I}L^{J'}u +\sum_{|J'|<|J|}\theta_{0J'}^{J\beta}\del_{\beta}\del^{I}\left([L^{J'},\del_t] u\right) \\ =& 2\sum_{|J'|<|J|}\theta_{0J'}^{J\beta}\del_t\del_{\beta}\del^{I}L^{J'}u + \sum_{|J'|<|J|\atop|J''|<|J'|}\theta_{0J'}^{J_2\beta}\theta_{0J''}^{J'\gamma}\del_{\beta}\del_{\gamma}\del^{I}L^{J''}u. \endaligned \end{equation} This leads to \begin{equation}\label{eq 1 01-09-2017} [\del^{I}L^{J},\del_t\del_t u]\leq C\sum_{\alpha,\beta\atop|J'|<|J|}|\del_{\alpha}\del_{\beta}\del^{I}L^{J'}u| \end{equation} By the above inequality we also see that: $$ \aligned |\del^{I_2}L^{J_2}\del_t\del_tu| \leq& |\del_t\del_t\del^{I_2}L^{J_2}u| + |[\del^{I_2}L^{J_2},\del_t\del_t]u| \\ \leq& C\sum_{{\alpha,\beta}\atop|J_2'|\leq|J_2|}|\del_{\alpha}\del_{\beta}\del^{I_2}L^{J_2'}u| \endaligned $$ Thus we see that (recall $|\del^IL^J(s/t)^2|\leq C(s/t)^2$) $$ |T_1|\leq C\sum_{\alpha,\beta,|I'|\leq |I|,|J'|\leq |J|\atop|I'|+|J'|\leq|I|+|J|-1}|(s/t)^2\del_{\alpha}\del_{\beta}\del^{I'}L^{J'}u| $$ For the term $T_2$, by \eqref{eq 6 22-08-2017} and the fact that $r^2/t^3$ is homogeneous of degree $-1$: $$ \aligned |T_2|\leq& \sum_{I_1+I_2=I,J_1+J_2=J\atop|I_2|+|J_2|\leq |I|+|J|-1}\big|\del^{I_1}L^{J_1}(r^2/t^3)\cdot\del^{I_2}L^{J_2}\del_tu\big| + \big|(r^2/t^3)[\del^IL^J,\del_t]u\big| \\ \leq& \sum_{\alpha,|I'|\leq|I|,|J'|\leq |J|\atop |I'|+|J'|\leq|I|+|J|-1}|t^{-1}\del_\alpha\del^{I'}L^{J'} u|. \endaligned $$ The bounds on $T_1$ and $T_2$ concludes the desired result. \end{proof} \begin{lemma}\label{lem 3 03-09-2017} Let $u$ be a function defined in $\Kcal$ and sufficiently regular. Then the following estimate holds: \begin{equation} \aligned \big|[\del^IL^J,\delb_s\delb_a]u\big| \leq& C\sum_{|I''|\leq|I|\atop|J''|\leq|J|}\left(t^{-1}|\delb_s\del^{I''}L^{J''}u| + (s/t^2)|\del^{I''}L^{J''}u|\right) \\ &+ C(s/t^2)\sum_{a,|I''|\leq|I|\atop|J''|\leq|J|}|\delb_a\del^{I''}L^{J''}u|. \endaligned \end{equation} \end{lemma} \begin{proof} $$ [\del^IL^J,\delb_s\delb_a]u = [\del^IL^J,\delb_s]\delb_au + \delb_s\left([\del^IL^J,\delb_a]u\right) $$ For the first term in right-hand-side of the above equation, we see that by \eqref{eq 8 22-08-2017}, $$ \aligned \big|[\del^IL^J,\delb_s]\delb_au\big|\leq& Cs^{-1}\sum_{|I'|<|I|}|\del_t\del^{I'}L^J\delb_au| + C(s/t)\sum_{{\alpha,|I'|\leq|I|}\atop|J'|<J}|\del_\alpha\del^{I'}L^{J'}\delb_au| \\ = &Cs^{-1}\sum_{|I'|<|I|}|\del_t\del^{I'}L^J\left(t^{-1}L_au\right)| + C(s/t)\sum_{{\alpha,|I'|\leq|I|}\atop|J'|<J}|\del_\alpha\del^{I'}L^{J'}\left(t^{-1}L_a\right)u| \\ \leq& Cs^{-1}t^{-1}\sum_{|I'|<|I|\atop |J'|\leq|J|}|\del_t\del^{I'}L^JL_au| + C(s/t^2)\sum_{{\alpha,|I'|\leq|I|}\atop|J'|<J}|\del_\alpha\del^{I'}L^{J'}L_au| \\ \leq& C(s/t^2)\sum_{{\alpha\atop|I'|+|J'|\leq|I|+|J|}}|\del_\alpha\del^{I'}L^{J'}u| \endaligned $$ For the term $\delb_s\left([\del^IL^J,\delb_a]u\right)$, we see that by applying \eqref{eq 1 23-08-2017}, \begin{equation}\label{eq 1 05-09-2017} \aligned \delb_s\left([\del^IL^J,\delb_a]u\right) =&\sum_{|I'|\leq|I|\atop |J'|<|J|}\delb_s\left(\rhob_{a I'J'}^{IJc}\delb_c\del^{I'}L^{J'}u\right) + \sum_{1\leq|I'|\leq |I|}\delb_s\left(\rho_{a I'}^{IJ}\del^{I'}L^{J}u\right) \endaligned \end{equation} Now we for the first term in the right-hand-side, we see that $$ \delb_s\left(\rhob_{a I'J'}^{IJc}\delb_c\del^{I'}L^{J'}u\right) = \delb_s\rhob_{a I'J'}^{IJc}\cdot \delb_c\del^{I'}L^{J'}u + \rhob_{a I'J'}^{IJc}\delb_s\delb_c\del^{I'}L^{J'}u $$ We see that $$ \left|\delb_s\rhob_{a I'J'}^{IJc}\right|\leq C(s/t^2) $$ thus \begin{equation}\label{eq 2 05-09-2017} \left|\delb_s\rhob_{a I'J'}^{IJc}\cdot \delb_c\del^{I'}L^{J'}u\right|\leq C(s/t^2)\big|\delb_c\del^{I'}L^{J'}u \big|, \end{equation} Now we consider $\rhob_{a I'J'}^{IJc}\delb_s\delb_c\del^{I'}L^{J'}u$. This is by the following calculation (by \eqref{eq 1 19-08-2017}): $$ \aligned \delb_s\delb_c\del^{I'}L^{J'}u =& \delb_s\left(t^{-1}L_c\del^{I'}L^{J'}u\right) = \delb_s\left(t^{-1}\del^{I'}L_cL^{J'}u\right) + \sum_{|I''|=I'}\delb_s\left(t^{-1}\theta_{cI''}^{I'}\del^{I''}L^{J'}u\right) \endaligned $$ Thus we see that $$ \big|\rhob_{a I'J'}^{IJc}\delb_s\delb_c\del^{I'}L^{J'}u\big| \leq C\sum_{|I''|\leq|I|\atop |J''|\leq|J|}\left((s/t^2)|\del^{I''}L^{J''}u| + t^{-1}|\delb_s\del^{I''}L^{J''}u|\right). $$ For the second term in \eqref{eq 1 05-09-2017}, we see that $$ \delb_s\left(\rho_{a I'}^{IJ}\del^{I'}L^{J}u\right) = \delb_s\rho_{a I'}^{IJ}\cdot \del^{I'}L^{J}u + \rho_{a I'}^{IJ}\cdot\delb_s\del^{I'}L^Ju $$ and recall that $\rho_{a I'}^{IJ}$ is homogeneous of degree $\leq -1$. Thus we see that $|\delb_s\rho_{a I'}^{IJ}|\leq Cs/t^2$. So we see that $$ \big|\delb_s\left(\rho_{a I'}^{IJ}\del^{I'}L^{J}u\right)\big|\leq C\sum_{|I''|\leq|I|\atop|J'|\leq|J|}\left(\big|t^{-1}\delb_s\del^{I''}L^{J''}u\big| + (s/t^2)\big|\del^{I''}L^{J''}u\big|\right). $$ Thus we see that the desired estimate is established. \end{proof} \begin{lemma}\label{lem 1 05-09-2017} Let $u$ be a function defined in $\Kcal$ sufficiently regular. Then the following estimate holds: \begin{equation}\label{eq 3 05-09-2017} \big|[\del^IL^J,\delb_a\delb_b]u\big|\leq C\sum_{c,|I'|\leq|I|\atop|J'|\leq|J|}\left(t^{-1}|\delb_c\del^{I'}L^{J'}u| + t^{-2}|\del^{I'}L^{J'}u|\right). \end{equation} \end{lemma} \begin{proof} We remark that \begin{equation}\label{eq 4 05-09-2017} \aligned \,[\del^IL^J,\delb_a\delb_b]u =& [\del^IL^J,\delb_a] \delb_bu + \delb_a\left([\del^IL^J,\delb_b]u\right) \endaligned \end{equation} For the first term in right-hand-side of the above equation, we see that by \eqref{eq 1 23-08-2017}, $$ \aligned \,[\del^IL^J,\delb_a] \delb_bu =& \sum_{|I'|\leq|I|\atop |J'|<|J|}\rhob_{a I'J'}^{IJc}\delb_c\del^{I'}L^{J'}\delb_bu + \sum_{1\leq|I'|\leq |I|}\rho_{a I'}^{IJ}\del^{I'}L^{J}\delb_bu \endaligned $$ We see that (by homogeneity of $\rhob_{a I'J'}^{IJc}$ and $t^{-1}$) $$ \aligned \big|\rhob_{a I'J'}^{IJc}\delb_c\del^{I'}L^{J'}\delb_bu\big| =&\big|\rhob_{a I'J'}^{IJc}\delb_c\del^{I'}L^{J'}\left(t^{-1}L_bu\right)\big| \\ \leq &C\sum_{c,|I''|\leq|I|\atop|J''|<|J|}\left(t^{-1}|\delb_c\del^{I''}L^{J''}u| + t^{-2}|\del^{I''}L^{J''}u|\right). \endaligned $$ By \eqref{eq 2 23-08-2017} $$ \aligned \big|\rho_{aI'}^{IJ}\del^{I'}L^{J}\delb_bu\big|\leq& Ct^{-1}\sum_{{c,|I''|\leq|I'|}\atop |J''|<|J|}|\delb_c\del^{I''}L^{J''}u| + Ct^{-2}\sum_{1\leq|I''|\leq |I'|}\left|\del^{I''}L^Ju\right| \endaligned $$ For the second term in right-hand-side of \eqref{eq 4 05-09-2017}, we see that by \eqref{eq 1 23-08-2017} $$ \aligned \delb_a\left([\del^IL^J,\delb_b]u\right) =&\sum_{|I'|\leq|I|\atop |J'|<|J|}\delb_a\left(\rhob_{b I'J'}^{IJc}\delb_c\del^{I'}L^{J'}u\right) + \sum_{1\leq|I'|\leq |I|}\delb_a\left(\rho_{b I'}^{IJ}\del^{I'}L^{J}u\right). \endaligned $$ By homogeneity, we see that $$ \aligned \left|\delb_a\left([\del^IL^J,\delb_b]u\right)\right| \leq& C\sum_{c,|I'|\leq|I|\atop|J'|<|J|}\left(|\delb_a\delb_c\del^{I'}L^{J'}u| + t^{-1}|\delb_c\del^{I'}L^{J'}u|\right) \\ &+C\sum _{1\leq|I'|\leq|I|}\left(t^{-1}|\delb_a\del^{I'}L^{J}u| + t^{-2}|\del^{I'}L^{J}u|\right). \endaligned $$ Now we see that $$ \aligned |\delb_a\delb_c\del^{I'}L^{J'}u| =& \big|\delb_a\left(t^{-1}L_c\del^{I'}L^{J'}u\right)\big| \leq C t^{-2}|L_c\del^{I'}L^{J'}u| + Ct^{-1}|\delb_aL_c\del^{I'}L^Ju| \\ \leq& C\sum_{|I''|=|I'|\atop |J''|\leq|J'|}\left(t^{-2}|\del^{I''}L^{J''}u| + t^{-1}|\delb_a\del^{I''}L^{J''}u|\right) \endaligned $$ The above bounds conclude the desired result. \end{proof} \subsection{Estimates based on commutators I} In this subsection, we will control the following terms \begin{equation}\label{eq 3 24-08-2017} \|(s^2/t)\del^IL^J\delb_su\|_{L^2(\Hcal_s)},\quad \|s\del^IL^J\delb_au\|_{L^2(\Hcal_s)},\quad \|(s/t)\del^IL^J u\|_{L^2(\Hcal_s)} \end{equation} where $|I|+|J|\leq N$. \begin{equation}\label{eq 6 05-09-2017} \|s^2\del^IL^J\delb_s\delb_au\|_{L^2(\Hcal_s)},\quad \|st\del^IL^J\delb_a\delb_b u\|_{L^2(\Hcal_s)} \end{equation} where $|I|+|J|\leq N-1$. We have the following result \begin{lemma}\label{lem 2 24-08-2017} Let $u$ be a function defined in $\Kcal$, sufficiently regular. Then the terms in \eqref{eq 3 24-08-2017} and \eqref{eq 6 05-09-2017} are bounded by $$ \sum_{|I|+|J|\leq N}\Ec(s,\del^IL^Ju)^{1/2}. $$ \end{lemma} \begin{proof} These are by apply the decomposition of commutators. For the first term in \eqref{eq 3 24-08-2017}, $$ \aligned \big|(s^2/t)\del^IL^J\delb_su\big|\leq& \big|(s^2/t)\delb_s\del^IL^Ju\big| + \big|(s^2/t)[\del^IL^J,\delb_s]u\big| \\ \leq &\big|(s^2/t)\delb_s\del^IL^Ju\big| + C(s/t)\sum_{|I'|<|I|}|\del_t\del^{I'}L^Ju| + Cs(s/t)^2\sum_{{\alpha,|I'|\leq|I|}\atop|J'|<J}|\del_\alpha\del^{I'}L^{J'}u| \\ \leq& \big|(s^2/t)\delb_s\del^IL^Ju\big| + C\sum_{|I'|<|I|}|(1+s^2/t)\delb_s\del^{I'}L^Ju| \\ &+ C\sum_{{a,|I'|\leq|I|}\atop|J'|<J}|s(s/t)^2\del_a\del^{I'}L^{J'}u| \endaligned $$ thus (remark that for the second term, $s^2/t\geq 1\geq s/t$ in $\Kcal$) $$ \big\|(s^2/t)\del^IL^J\delb_su\big\|_{L^2(\Hcal_s)}\leq C\sum_{|I|+|J|\leq N}\Ec(s,\del^IL^Ju)^{1/2}. $$ For the second term we apply \eqref{eq 2 23-08-2017}, we omit the detail. The third is guaranteed by \eqref{eq 6 23-05-2017}. For the first term in \eqref{eq 6 05-09-2017}, we see that $$ \|s^2\del^IL^J\delb_s\delb_au\|_{L^2(\Hcal_s)}\leq \|s^2\delb_s\delb_a\del^IL^Ju\|_{L^2(\Hcal_s)} + \|s^2[\del^IL^J,\delb_s\delb_a]u\|_{L^2(\Hcal_s)} $$ By lemma \ref{lem 3 03-09-2017}, we see that $$ \|s^2[\del^IL^J,\delb_s\delb_a]u\|_{L^2(\Hcal_s)}\leq \sum_{|I'|+|J'|\leq N}\Ec(s,\del^{I'}L^{J'}u)^{1/2}. $$ $$ \aligned \|s^2\delb_s\delb_a\del^IL^Ju\|_{L^2(\Hcal_s)} =& \|s^2\delb_s\left(t^{-1}L_a\del^IL^Ju\right)\|_{L^2(\Hcal_s)} \\ \leq &\|(s^2/t)\delb_sL_a\del^IL^Ju\|_{L^2(\Hcal_s)} + \|(s/t)^3L_a\del^IL^Ju\|_{L^2(\Hcal_s)} \\ \leq &\|(s^2/t)\delb_s\del^IL_aL^Ju\|_{L^2(\Hcal_s)} + \|(s/t)^3\del^IL_aL^Ju\|_{L^2(\Hcal_s)} \\ &+\|(s^2/t)\delb_s\left([L_a,\del^I]L^Ju\right)\|_{L^2(\Hcal_s)} + \|(s/t)^3[L_a,\del^I]L^Ju\|_{L^2(\Hcal_s)} \endaligned $$ Then by \eqref{eq 1 19-08-2017}, we see that $$ \aligned \|s^2\delb_s\delb_a\del^IL^Ju\|_{L^2(\Hcal_s)}\leq& \sum_{|I'|\leq|I|\atop|J'|\leq|J|+1}\|(s^2/t)\delb_s\del^{I'}L^{J'}u\|_{L^2(\Hcal_s)} + \|(s/t)^3\del^{I'}L^{J'}u\|_{L^2(\Hcal_s)} \\ \leq&\sum_{|I'|+|J'|\leq N}\Ec(s,\del^{I'}L^{J'}u)^{1/2} \endaligned $$ and this proved the bound on the first term of \eqref{eq 6 05-09-2017}. Now we regard the term $\del^IL^J\delb_a\delb_b u$. \begin{equation}\label{eq 1 07-09-2017} \aligned \|st\del^IL^J\delb_a\delb_bu\|_{L^2(\Hcal_s)} =& \|st\delb_a\delb_b\del^IL^Ju\|_{L^2(\Hcal_s)} + \|st[\del^IL^J,\delb_a\delb_b]u\|_{L^2(\Hcal_s)}. \endaligned \end{equation} For the first term in the right-hand-side of the above equation, we see that $$ \aligned \|st\delb_a\delb_b\del^IL^Ju\|_{L^2(\Hcal_s)} =& \|st\delb_a\left(t^{-1}L_b\del^IL^Ju\right)\|_{L^2(\Hcal_s)} \\ \leq& \|s\delb_aL_b\del^IL^Ju\|_{L^2(\Hcal_s)} + \|(s/t)L_b\del^IL^Ju\|_{L^2(\Hcal_s)} \\ \leq& \|s\delb_a\del^IL_b L^Ju\|_{L^2(\Hcal_s)} + \|(s/t)\del^I L_b L^Ju\|_{L^2(\Hcal_s)} \\ &+\|s\delb_a\left([L_b,\del^I]L^Ju\right)\|_{L^2(\Hcal_s)} + \|(s/t)[L_b,\del^I]L^Ju\|_{L^2(\Hcal_s)}. \endaligned $$ Then, also by \eqref{eq 1 19-08-2017}, $$ \|st\delb_a\delb_b\del^IL^Ju\|_{L^2(\Hcal_s)} \leq C\sum_{|I'|=|I|\atop |J'|\leq|J|+1}\Ec(s,\del^{I'}L^{J'}u)^{1/2}. $$ For the second term in right-hand-side of \eqref{eq 1 07-09-2017}, by applying lemma \ref{lem 1 05-09-2017}, we see that it is also bounded by $$ C\sum_{|I|+|J|\leq N}\Ec(s,\del^IL^Ju) $$ Thus the desired bound is established. \end{proof} We also establish a rough bound on $\del^IL^J\delb_s\delb_s$: \begin{lemma}\label{lem 2 07-09-2017} Let $u$ be a function defined in $\Kcal$, sufficiently regular and vanishes near the conical boundary. Then the following bound holds for $|I|+|J|\leq N-1$: \begin{equation}\label{eq 5 07-09-2017} \|s\delb_s\delb_s\del^IL^Ju\|_{L^2(\Hcal_s)} +\|s\del^IL^J\delb_s\delb_su\|_{L^2(\Hcal_s)}\leq C\sum_{|I'|+|J'|\leq N}\Ec(s,\del^{I'}L^{J'}u)^{1/2}. \end{equation} \end{lemma} \begin{proof} For the first term, we recall that \begin{equation}\label{eq 9-07-09-2017} \aligned s\delb_s\delb_s\del^IL^Ju =& s\delb_s\left((s/t)\del_t\del^IL^Ju\right) =& s\delb_s(s/t)\cdot\del_t\del^IL^Ju + s(s/t)\delb_s\del_t\del^IL^Ju. \endaligned \end{equation} We see that for the first term in right-hand-side, by \eqref{eq 2 20-08-2017}, $$ \|s\delb_s\left((s/t)\del_t\del^IL^Ju\right)\|_{L^2(\Hcal_s)}\leq C\|(s/t)\del_t\del^IL^Ju\|_{L^2(\Hcal_s)} \leq C\sum_{|I'|+|J'|\leq N}\Ec(s,\del^{I'}L^{J'}u)^{1/2}. $$ For the second term in right-hand-side of \eqref{eq 9-07-09-2017}, We remark the following calculation (where we apply \eqref{eq 6 22-08-2017}): \begin{equation}\label{eq 6 07-09-2017} s\del^IL^J\delb_s\delb_su = \sum_{I_1+I_2=I\atop J_1+J_2=J}s\del^{I_1}L^{J_1}(s/t)\cdot \del_t\del^{I_2}L^{J_2}\delb_su + \sum_{I_1+I_2=I,J_1+J_2=J\atop |J_2'|<|J_2|}\!\!\!\!\!\!\!\!\!\!s\del^{I_1}L^{J_1}(s/t)\theta_{0J_2'}^{J_2\beta}\del_{\beta}\del^{I_1}L^{J_2'}\delb_su \end{equation} \end{proof} Then by applying \eqref{eq 2 20-08-2017} and the bounds on terms in \eqref{eq 3 24-08-2017}, we see that the desired result is proved. \subsection{Estimates based on commutators II} In view of the global Sobolev inequality \eqref{ineq 1 sobolev}, to turn the $L^2$ bounds (supplied by the energy) into $L^{\infty}$ bounds, we need to bound some terms. To do so , we need some preparations. Through out this subsection, we denote by $u$ a sufficiently regular function defined in $\Kcal$, and the following estimates are valid in $\Kcal$. \begin{lemma}\label{lem 1 29-08-2017} Let $u$ be a sufficiently regular function defined in $\Kcal$. Then the following estimates hold: \begin{equation}\label{eq 5 29-08-2017} \left|\del^{I'}L^{J'}\delb_s\del^IL^Ju\right|\leq C\sum_{|I''|\leq|I|+|I'|\atop |J''|\leq|J|+|J'|}\big|\delb_s\del^{I''}L^{J''}u\big| + C\sum_{{a,|I'|\leq|I'|+|I|}\atop{|J''|<|J'|+|J|}} |(s/t)\del_a\del^{I''}L^{J''}u|, \end{equation} \begin{equation}\label{eq 7 29-08-2017} \left|\del^{I'}L^{J'}\delb_a\del^IL^Ju\right|\leq C\sum_{{c,|I''|\leq|I|+|I'|}\atop{|J''|\leq|J|+|J'|}}\big|\delb_c\del^{I''}L^{J''}u\big| + Cs^{-1}\sum_{{1\leq|I''|\leq|I'|+|I|}\atop|J''|\leq|J'|+|J|}\big|(s/t)\del^{I''}L^{J''}u\big|. \end{equation} \end{lemma} \begin{proof} For the first term, we see that $$ \aligned \del^{I'}L^{J'}\delb_s\del^IL^Ju =& \delb_s\del^{I'}L^{J'}\del^IL^Ju + [\del^{I'}L^{J'},\delb_s]\del^IL^Ju =: T_1 +T_2. \endaligned $$ Then we see that by \eqref{eq 6 29-08-2017} $$ T_1 = \sum_{|I''|\leq|I'|+|I|\atop |J''|\leq|J'|+|J|}\zeta^{I'J'IJ}_{I''J''}\delb_s\del^{I''}L^{J''}u $$ which leads to $$ |T_1|\leq C\sum_{|I''|\leq|I|+|I'|\atop |J''|\leq|J|+|J'|}\big|\delb_s\del^{I''}L^{J''}u\big|. $$ On the other hand, $$ \aligned |T_2| \leq& Cs^{-1}\sum_{|I''|<|I'|}\big|\del_t\del^{I''}L^{J'}\del^IL^Ju\big| + C(s/t)\sum_{{\alpha,|I''|\leq|I'|}\atop |J''|<|J'|}\big|\del_\alpha\del^{I''}L^{J''}\del^IL^Ju\big| \\ =:& T_3+T_4 . \endaligned $$ We see that for each term of $T_3$, by \eqref{eq 6 29-08-2017} and the fact that in $\Kcal$ $s^2\geq t$, $$ \aligned s^{-1}\big|\del_t\del^{I''}L^{J'}\del^IL^Ju\big| \leq& C(t/s^2)\sum_{|I'''|\leq|I|\atop |J'''|\leq|J'|}\big|(s/t)\del_t\del^{I''}\del^{I'''}L^{J'''}L^Ju\big| \\ \leq& C\sum_{|I'''|\leq|I|\atop |J'''|\leq|J'|}\big|\delb_s\del^{I''}\del^{I'''}L^{J'''}L^Ju\big|. \endaligned $$ Also, for $T_4$, when $\alpha = 0$, $\del_{\alpha} = \del_t$, then $$ \aligned (s/t)\big|\del_t\del^{I''}L^{J''}\del^IL^Ju\big| \leq& C\sum_{|I'''|\leq |I|\atop|J'''|\leq|J''|}\big|(s/t)\del_t\del^{I''}\del^{I'''}L^{J'''}L^Ju\big| \\ \leq& C\sum_{|I'''|\leq |I|\atop|J'''|\leq|J''|}\big|\delb_s\del^{I''}\del^{I'''}L^{J'''}L^Ju\big|. \endaligned $$ For $\alpha>0$, we denote by $\alpha = a$, then $$ (s/t)\big|\del_a\del^{I''}L^{J''}\del^IL^Ju\big|\leq C\sum_{{a,|I'''|\leq |I|}\atop|J'''|\leq|J''|}\big|(s/t)\del_a\del^{I''}\del^{I'''}L^{J'''}L^Ju\big| $$ This leads to $$ |T_2|\leq C\sum_{|I''|\leq|I|+|I'|\atop |J''|\leq|J|+|J'|}\big|\delb_s\del^{I''}L^{J''}u\big| + C\sum_{{a,|I'''|\leq |I|}\atop|J'''|\leq|J''|}\big|(s/t)\del_a\del^{I''}\del^{I'''}L^{J'''}L^Ju\big|. $$ The bounds of $T_1$ and $T_2$ leads to \eqref{eq 5 29-08-2017}. For \eqref{eq 7 29-08-2017}, we see that $$ \aligned \del^{I'}L^{J'}\delb_a\del^IL^Ju =& \delb_a\del^{I'}L^{J'}\del^{I}L^Ju + [\del^{I'}L^{J'},\delb_a]\del^IL^Ju =: T_3 + T_4. \endaligned $$ We see that by \eqref{eq 6 29-08-2017} $$ |T_3|\leq C\sum_{|I''|\leq|I|+|I'|\atop|J''|\leq |J|+|J'|}\big|\delb_a\del^{I''}L^{J''}u\big|. $$ For $T_4$, we apply \eqref{eq 2 23-08-2017}: $$ \aligned |T_4|\leq & C\sum_{{c,|I''|\leq|I'|}\atop |J''|\leq|J'|}\big|\delb_c\del^{I''}L^{J''}\del^IL^Ju\big| + Ct^{-1}\sum_{1\leq|I''|\leq|I'|}\big|\del^{I''}L^{J'}\del^IL^Ju\big|. \endaligned $$ Also by \eqref{eq 6 29-08-2017}: $$ \aligned |T_4|\leq& C\sum_{{c,|I''|\leq|I|+|I'|}\atop{|J''|\leq|J|+|J'|}}\big|\delb_c\del^{I''}L^{J''}u\big| + Ct^{-1}\sum_{1\leq |I''|\leq|I'|+|I|}\big|\del^{I''}L^{J''}u\big| \\ =& C\sum_{{c,|I''|\leq|I|+|I'|}\atop{|J''|\leq|J|+|J'|}}\big|\delb_c\del^{I''}L^{J''}u\big| + Cs^{-1}\sum_{{1\leq|I''|\leq|I'|+|I|}\atop|J''|\leq|J'|+|J|}\big|(s/t)\del^{I''}L^{J''}u\big| \endaligned $$ Now the bounds on $|T_3|$ and $|T_4|$ leads to \eqref{eq 7 29-08-2017}. \end{proof} Then the following bounds are direct: \begin{lemma}\label{lem 2 31-08-2017} The following terms: \begin{equation}\label{eq 2 31-08-2017} \|(s^2/t)\del^{I'}L^{J'}\delb_s\del^IL^Ju\|_{L^2(\Hcal_s)},\quad \|s\del^{I'}L^{J'}\delb_a\del^IL^Ju\|_{L^2(\Hcal_s)} \end{equation} are bounded by $$ C\sum_{|I''|\leq|I|+|I'|\atop |J''|\leq|J|+|J'|}\Ec(s,\del^{I''}L^{J''}u)^{1/2}. $$ \end{lemma} Now we regard the following terms: \begin{equation}\label{eq 1 24-08-2017} \|\del^{I'}L^{J'}\left((s^2/t)\delb_s\del^IL^J u\right)\|_{L^2(\Hcal_s)},\quad \|\del^{I'}L^{J'}\left(s\delb_a\del^IL^Ju\right)\|_{L^2(\Hcal_s)} \end{equation} where $|I'|+|J'|\leq 2$ and $|I|+|J|\leq N-2$. \begin{equation}\label{eq 2 24-08-2017} \|\del^{I'}L^{J'}\left(s^2\delb_s\delb_a \del^IL^J u\right)\|_{L^2(\Hcal_s)},\quad \|\del^{I'}L^{J'}\left(st\delb_a\delb_b\del^IL^Ju\right)\|_{L^2(\Hcal_s)} \end{equation} where $|I'|+|J'|\leq 2$ and $|I|+|J|\leq N-3$. We have the following results: \begin{lemma}\label{lem 1 24-08-2017} The terms in \eqref{eq 1 24-08-2017} and \eqref{eq 2 24-08-2017} are bounded by $$ C\sum_{|I'|+|J'|\leq N}\Ec(s,\del^{I'}L^{J'}u)^{1/2}. $$ \end{lemma} \begin{proof} For the first term in \eqref{eq 1 24-08-2017}, we see that by \eqref{eq 4 22-08-2017}, \eqref{eq 5 29-08-2017} and lemma \ref{lem 2 31-08-2017}: $$ \aligned &\left\|\del^{I'}L^{J'}\left((s^2/t)\delb_s\del^IL^J u\right)\right\|_{L^2(\Hcal_s)} \leq \sum_{I_1'+I_2'=I'\atop J_1'+J_2'=J'}\left\|\del^{I_1'}L^{J_1'}(s^2/t)\cdot \del^{I_2'}L^{J_2'}\delb_s\del^IL^Ju\right\|_{L^2(\Hcal_s)} \\ \leq& C\sum_{I_1'+I_2'=I'\atop J_1'+J_2'=J'}\left\|(s^2/t)\del^{I_2'}L^{J_2'}\delb_s\del^IL^Ju\right\|_{L^2(\Hcal_s)} \leq C\sum_{|I''|\leq|I'|+|I|\atop |J''|\leq|J'|+|J|}\Ec(s,\del^{I''}L^{J''}u)^{1/2} \\ =& C\sum_{|I''|+|J''|\leq N}\Ec(s,\del^{I''}L^{J''}u)^{1/2}. \endaligned $$ The second term of \eqref{eq 1 24-08-2017} is controlled by \eqref{eq 1 20-08-2017} (combined with the fact that in $\Kcal$, $s\geq t/s$), \eqref{eq 7 29-08-2017} and lemma \ref{lem 2 31-08-2017}: $$ \aligned &\|\del^{I'}L^{J'}\left(s\delb_a\del^IL^Ju\right)\|_{L^2(\Hcal_s)} \\ \leq& C\sum_{I_1'+I_2'=I'\atop J_1'+J_2'=J'}\|\del^{I_1'}L^{J_1'}s\cdot\del^{I_2'}L^{J_2'}\delb_a\del^IL^Ju\|_{L^2(\Hcal_s)} \leq C\sum_{I_1'+I_2'=I'\atop J_1'+J_2'=J'}\|s\del^{I_2'}L^{J_2'}\delb_a\del^IL^Ju\|_{L^2(\Hcal_s)} \\ \leq& C\sum_{|I''|+|J''|\leq N}\Ec(s,\del^{I''}L^{J''}u)^{1/2}. \endaligned $$ The first term in \eqref{eq 2 24-08-2017} is bounded as following. First we remark that $$ \aligned \del^{I'}L^{J'}\left(s^2\delb_s\delb_a\del^IL^Ju\right) =& \del^{I'}L^{J'}\left(s^2\delb_s\left(t^{-1}L_a\del^IL^Ju\right)\right) \\ =&\del^{I'}L^{J'}\left((s^2/t)\delb_sL_a\del^IL^Ju\right) - \del^{I'}L^{J'}\left((s/t)^3L_a\del^IL^Ju\right) \\ =& T_1 + T_2. \endaligned $$ For $T_1$, we see that by \eqref{eq 4 22-08-2017} and lemma \ref{lem 2 31-08-2017}, we see that $$ \aligned \|T_1\|\leq& C\sum_{I_1'+I_2'=I'\atop J_1'+J_2'=J'}\del^{I_1'}L^{J_1'}(s^2/t)\cdot \del^{I_2'}L^{J_2'}\delb_sL_a\del^IL^Ju \\ \leq& C\sum_{|I_2'|\leq|I'|\atop|J_2'|\leq|J'|}\|(s^2/t)\del^{I_2'}L^{J_2'}\delb_sL_a\del^IL^Ju\|_{L^2(\Hcal_s)} \\ \leq& C\sum_{|I|+|J|\leq N}\Ec(s,\del^{I''}L^{J''}u)^{1/2}. \endaligned $$ Here by For the term $T_2$, we see that by \eqref{eq 6 29-08-2017} and lemma \ref{lem 2 31-08-2017}: $$ \aligned \|T_2\| \leq& C\sum_{I_1'+I_2'+I_3'+I_4'=I'\atop J_1'+J_2'+J_3'+J_4'=J'} \big\|\del^{I_1'}L^{J_1'}(s/t)\cdot\del^{I_2'}L^{J_2'}(s/t)\cdot\del^{I_3'}L^{J_3'}(s/t)\cdot\del^{I_4'}L^{J_4'}L_a\del^IL^Ju\big\|_{L^2(\Hcal_s)} \\ \leq& C\sum_{|I_4'|\leq|I'|\atop |J_4'|\leq|J'|}\|(s/t)\del^{I_4'}L^{J_4'}L_a\del^IL^Ju\|_{L^2(\Hcal_s)} \leq C\sum_{|I''|\leq|I_4'|+|I|\atop |J''|\leq|J_4''|+|J|+1}\|(s/t)\del^{I_4'}L^{J_4'}L_a\del^IL^Ju\|_{L^2(\Hcal_s)} \\ \leq& C\sum_{|I''|+|J''|\leq N}\Ec(s,\del^{I''}L^{J''}u)^{1/2}. \endaligned $$ The bounds on $T_1$ and $T_2$ give the bounds of the first term of \eqref{eq 2 24-08-2017}. Now we regard the second term in \eqref{eq 2 24-08-2017}: $$ \aligned &\del^{I'}L^{J'}\left(st\delb_a\delb_b\del^IL^Ju\right) = \del^{I'}L^{J'}\left(st\delb_a\left(t^{-1}L_b\del^IL^Ju\right)\right) \\ =&\del^{I'}L^{J'}\left(s\delb_aL_b\del^IL^Ju\right) - \del^{I'}L^{J'}\left((sx^a/t^2)L_a\del^IL^Ju\right) \\ =&\sum_{I_1'+I_2'=I'\atop J_1'+J_2'=J'}\del^{I_1'}L^{J_1'}s\cdot\del^{I_2'}L^{J_2'}\delb_aL_b\del^IL^Ju -\sum_{I_1'+I_2'+I_3=I'\atop J_1'+J_2'+J_3=J'}\del^{I_1'}L^{J_1'}(s/t)\del^{I_2'}L^{J_2'}(x^a/t)\del^{I_3'}L^{J_3'}L_a\del^IL^Ju \\ =:& T_3 + T_4. \endaligned $$ For $T_3$, we see that by \eqref{eq 1 20-08-2017} together with the fact that in $\Kcal$, $t/s\leq s$ and \ref{lem 2 31-08-2017}, $$ \aligned \|T_3\|\leq& C\sum_{|I_2'|\leq|I'|\atop|J_2'|\leq|J'|}\|s\del^{I_2'}L^{J_2'}\delb_aL_b\del^IL^Ju\|_{L^2(\Hcal_s)} \leq C\sum_{|I''|+|J''|\leq N}\Ec(s,\del^{I''}L^{J''}u)^{1/2}. \endaligned $$ For $T_4$, we see that by \eqref{eq 2 20-08-2017} and lemma \ref{lem 1 26-06-2017} ($x^a/t$ is homogeneous of degree zero) $$ \aligned \|\del^{I_1'}L^{J_1'}(s/t)\del^{I_2'}L^{J_2'}(x^a/t)\del^{I_3'}L^{J_3'}L_a\del^IL^Ju\|_{L^2(\Hcal_s)} \leq& C\|(s/t)\del^{I_3'}L^{J_3'}L_a\del^IL^Ju\|_{L^2(\Hcal_s)} \\ \leq& C\sum_{|I''|+|J''|\leq N}\|(s/t)\del^{I''}L^{J''}u\|_{L^2(\Hcal_s)} \\ \leq& C\sum_{|I''|+|J''|\leq N}\Ec(s,\del^{I''}L^{J''}u)^{1/2} \endaligned $$ The bounds on $T_3$ and $T_4$ concludes the second term in \eqref{eq 2 24-08-2017}. \end{proof} Now we list out a second group of terms: \begin{equation}\label{eq 1 22-08-2017} \|\del^{I'}L^{J'}\left((s^2/t)\del^IL^J\delb_s u\right)\|_{L^2(\Hcal_s)},\quad \|\del^{I'}L^{J'}\left(s\del^IL^J\delb_au\right)\|_{L^2(\Hcal_s)},\quad \|\del^{I'}L^{J'}\left((s/t)\del^IL^Ju\right)\|_{L^2(\Hcal_s)} \end{equation} where $|I'|+|J'|\leq 2$, $|I|+|J|\leq N-2$ and \begin{equation}\label{eq 2 22-08-2017} \|\del^{I'}L^{J'}\left(s^2\del^IL^J\delb_s\delb_a u\right)\|_{L^2(\Hcal_s)},\quad \|\del^{I'}L^{J'}\left(st\del^IL^J \delb_a\delb_b u\right)\|_{L^2(\Hcal_s)} \end{equation} where $|I'|+|J'|\leq 2$, $|I|+|J|\leq N-3$ with $N$ is an integer. In general we have the following estimates: \begin{lemma}\label{lem 2 05-09-2017} In $\Kcal$, the terms in \eqref{eq 1 22-08-2017} and \eqref{eq 2 22-08-2017} are bounded by $$ C\sum_{|I'|+|J'|\leq N}\Ec(s,\del^{I'}L^{J'}u)^{1/2}. $$ \end{lemma} \begin{proof} These are by lemma \ref{lem 2 24-08-2017} and the corresponding estimates of commutators. For the first term in \eqref{eq 1 22-08-2017}, by \eqref{eq 6 29-08-2017} $$ \aligned \del^{I'}L^{J'}\left((s^2/t)\del^IL^J\delb_s u\right) =& \sum_{I_1'+I_2'=I'\atop J_1'+J_2'=J'}\del^{I_1'}L^{J_1'}(s^2/t)\cdot\del^{I_2'}L^{J_2'}\del^{I_2}L^{J_2}\delb_s u \\ =& \sum_{I_1'+I_2'=I'\atop J_1'+J_2'=J'}\sum_{|I''|=|I_2'|+|I|\atop|J''|\leq|J_2'|+|J|} \del^{I_1'}L^{J_1'}(s^2/t)\cdot\theta_{I''J''}^{I_2'J_2'IJ}\del^{I''}L^{J''}\delb_s u. \endaligned $$ Then by \eqref{eq 4 22-08-2017} and lemma \ref{lem 2 24-08-2017}, the desired $L^2$ bound is direct. The second term in \eqref{eq 1 22-08-2017} is bounded similarly by \eqref{eq 6 29-08-2017}, \eqref{eq 1 20-08-2017} and lemma \ref{lem 2 24-08-2017}, we omit the detail. The third term in \eqref{eq 1 22-08-2017} is by \eqref{eq 6 29-08-2017}, \eqref{eq 2 20-08-2017} and lemma \ref{lem 2 24-08-2017}, we also omit the detail. The first term in \eqref{eq 2 22-08-2017} is by \eqref{eq 1 20-08-2017} combined with lemma \ref{lem 2 24-08-2017}: $$ \aligned \|\del^{I'}L^{J'}\left(s^2\del^IL^J\delb_s\delb_au\right)\|_{L^2(\Hcal_s)} \leq& \sum_{I_1'+I_2'+I_3'=I'\atop J_1'+J_2'+J_3'=J'} \|\del^{I_1'}L^{J_1'}s\cdot\del^{I_2'}L^{J_2'}s\cdot\del^{I_3'}L^{J_3'}\delb_s\delb_a\del^IL^Ju\|_{L^2(\Hcal_s)} \\ \leq& C\sum_{|I''|\leq|I'|\atop|J''|\leq|J'|}\|s^2\del^{I_3'}L^{J_3'}\delb_s\delb_a\del^IL^Ju\|_{L^2(\Hcal_s)} \endaligned $$ which is bounded by $C\sum_{|I'|+|J'|\leq N}\Ec(s,\del^{I'}L^{J'}u)^{1/2}$. The second term in \eqref{eq 2 22-08-2017} is by \eqref{eq 1 20-08-2017} combined with lemma \ref{lem 2 24-08-2017}, we omit the detail. \end{proof} \subsection{Decay bounds from global Sobolev inequality and commutators} In this section by applying lemma \ref{lem 1 24-08-2017} and lemma \ref{lem 2 05-09-2017} combined with \eqref{ineq 1 sobolev}, we will establish a series decay estimates. In general we have the following results: \begin{proposition}[Decay estimates by energy bounds]\label{prop 1 29-08-2017} Let $u$ be a function defined in $\Kcal$, sufficiently regular. Then for $|I|+|J|\leq N-2$, the following terms: \begin{subequations} \begin{equation}\label{eq 3 29-08-2017} \sup_{\Hcal_s}\{st^{1/2}|\del^IL^Ju|\},\quad \sup_{\Hcal_s}\{st^{3/2}\delb_a\del^IL^Ju\},\quad \sup_{\Hcal_s}\{s^2t^{1/2}\delb_s\del^IL^Ju\}, \end{equation} and \begin{equation}\label{eq 4 29-08-2017} \sup_{\Hcal_s}\{st^{3/2}\del^IL^J\delb_au\},\quad \sup_{\Hcal_s}\{s^2t^{1/2}\del^IL^J\delb_su\}, \end{equation} \end{subequations} are bounded by $$ C\sum_{|I|+|J|\leq N}\Ec(s,\del^IL^J u)^{1/2}. $$ For $|I|+|J|\leq N-3$, the following terms: \begin{subequations} \begin{equation} \sup_{\Hcal_s}\left\{s^2t^{3/2}\delb_s\delb_a\del^IL^J u\right\},\quad \sup_{\Hcal_s}\left\{st^{5/2}\delb_a\delb_b\del^IL^Ju\right\} \end{equation} and \begin{equation} \sup_{\Hcal_s}\left\{s^2t^{3/2}\del^IL^J\delb_s\delb_a u\right\},\quad \sup_{\Hcal_s}\left\{st^{5/2}\del^IL^J\delb_a\delb_bu\right\} \end{equation} \end{subequations} are bounded by $$ C\sum_{|I|+|J|\leq N}\Ec(s,\del^IL^J u)^{1/2}. $$ \end{proposition} Furthermore, we have the following rough decay: For $|I|+|J|\leq N-3$ \begin{equation}\label{eq 10 07-09-2017} \sup_{\Hcal_s}\{st^{3/2}\del^IL^J\delb_s\delb_su\}\leq C\sum_{|I'|+|J'|\leq N}\Ec(s,\del^{I'}L^{J'}u)^{1/2}. \end{equation} This is by \eqref{eq 6 07-09-2017} combined with \eqref{eq 2 20-08-2017} and \eqref{eq 4 29-08-2017}. \section{Estimates on Hessian form} \subsection{Objective and algebraic preparation} The purpose of this section is to give better $L^2$ and decay bounds on the following terms: $$ \del_{\alpha}\del_{\beta}\del^IL^J u,\quad \del^IL^J \del_{\alpha}\del_{\beta}u $$ We first make the following identities: $$ \aligned \del_t\del_t =& (t/s)^2\delb_s\delb_s - s^{-1}(r/s)^2\delb_s, \endaligned $$ $$ \aligned \del_t\del_a =& \del_a\del_t = \left(\delb_a - (x^a/s)\delb_s\right)\del_t = -\frac{tx^a}{s^2}\delb_s\delb_s + s^{-1}L_a\delb_s + \frac{tx^a}{s^3}\delb_s \endaligned $$ and $$ \aligned \del_a\del_b =& \left(\delb_a - (x^a/s)\delb_s\right)\left(\delb_b - (x^b/s)\delb_s\right) \\ =& \frac{x^ax^b}{s^2}\delb_s\delb_s + t^{-1}L_a\delb_b - \frac{x^a}{st}L_a\delb_s - \frac{x^a}{st}L_b\delb_s - s^{-1}\delb_s - \frac{x^ax^b}{s^3}\delb_s \endaligned $$ These identities show the fact that in the Hessian form, the component $\delb_s\delb_s$ has an essential contribution, because the rest terms as at least $s^{-1}$ as a supplementary decay factor. So we will concentrate on the bounds of $\delb_s\delb_s$. More precisely the above identities lead to the following result: \begin{lemma}\label{lem 1 01-09-2017} For $u$ a function defined in $\Kcal$, sufficiently regular, the following estimate holds: \begin{equation} \|s^2(s/t)^3\del_{\alpha}\del_{\beta}u\|_{L^2(\Hcal_s)}\leq C\|s^2(s/t)\delb_s\delb_s u\|_{L^2(\Hcal_s)} + C\sum_{|J|\leq 1}\Ec(s,L^J u)^{1/2}. \end{equation} \end{lemma} \begin{proof} This is by direct calculation. We see that $$ \aligned \|s^2(s/t)^3\del_t\del_tu\|_{L^2(\Hcal_s)}\leq& \|s^2(s/t)\delb_s\delb_su\|_{L^2(\Hcal_s)} + \|s(s/t)\delb_s u\|_{L^2(\Hcal_s)} \\ \leq& \|s^2(s/t)\delb_s\delb_su\|_{L^2(\Hcal_s)} + C\Ec(s,u)^{1/2}. \endaligned $$ For $\del_t\del_a u$, we need to apply lemma \ref{lem 2 24-08-2017}: $$ \aligned \|s^2(s/t)^3\del_t\del_au\|_{L^2(\Hcal_s)}\leq& \|s^2(s/t)\delb_s\delb_su\|_{L^2(\Hcal_s)} + \|s(s/t)^2L_a\delb_su\|_{L^2(\Hcal_s)} + \|s(s/t)\delb_su\|_{L^2(\Hcal_s)} \\ \leq& \|s^2(s/t)\delb_s\delb_su\|_{L^2(\Hcal_s)} + C\sum_{|J|\leq 1}\Ec(s,L^Ju)^{1/2}. \endaligned $$ The term $\del_a\del_bu$ is also by \eqref{lem 2 24-08-2017}: $$ \aligned \|s^2(s/t)^3\del_a\del_bu\|_{L^2(\Hcal_s)}\leq& \|s^2(s/t)\delb_s\delb_su\|_{L^2(\Hcal_s)} \\ &+ \|s(s/t)^4L_a\delb_bu\|_{L^2(\Hcal_s)} + \|s(s/t)^2L_b\delb_s\|_{L^2(\Hcal_s)} \\ &+ \|s(s/t)^3\delb_su\|_{L^2(\Hcal_s)} + \|s(s/t)\delb_su\|_{L^2(\Hcal_s)} \\ \leq& \|s^2(s/t)\delb_s\delb_su\|_{L^2(\Hcal_s)} + C\sum_{|J|\leq 1}\Ec(s,L^J u)^{1/2}. \endaligned $$ \end{proof} Then we remark the following identity (see \eqref{eq 1 12-05-2017}): $$ \aligned \Box =& \delb_s\delb_s + 2(x^a/s)\delb_s\delb_a - \sum_a\delb_a\delb_a + \frac{3}{s}\delb_s \\ =& \delb_s\delb_s + (2x^a/s)t^{-1}L_a\delb_s - t^{-1}L_a\delb_a + 3s^{-1}\delb_s \endaligned $$ This leads to \begin{equation}\label{eq 2 02-07-2017} \delb_s\delb_s u = \Box u + H_1 [u]. \end{equation} where $$ H_1[u]:= - (2x^a/s)t^{-1}L_a\delb_s + t^{-1}L_a\delb_a - 3s^{-1}\delb_s $$ \subsection{$L^2$ bounds} We remark that from \eqref{eq 2 02-07-2017}, $$ (s^3/t)\delb_s\delb_s \del^IL^J u = (s^3/t)\Box \del^IL^J u + (s^3/t)H_1[\del^IL^J u] $$ We remark the following property: \begin{lemma}\label{lem 1 31-08-2017} Let $u$ be a function defined in $\Kcal$, sufficiently regular. Then the following estimate holds: \begin{equation}\label{eq 1 31-08-2017} \|(s^3/t)H_1[\del^IL^J u]\|_{L^2(\Hcal_s)}\leq C\sum_{|I'|+|J'|\leq|I|+|J|+1}\Ec(s,\del^{I'}L^{J'}u)^{1/2}. \end{equation} \end{lemma} \begin{proof} We see that $$ \aligned (s^3/t)\big|H_1[\del^IL^Ju]\big| \leq& |(s^2/t)L_a\delb_s\del^IL^Ju| + |sL_a\delb_a\del^IL^Ju| + 3|(s^2/t)\delb_s\del^IL^Ju|. \endaligned $$ Then by lemma \ref{lem 2 31-08-2017}, the result is proved. \end{proof} On the other hand, $$ (s^3/t)\del^IL^J\delb_s\delb_s u = (s^2/t)\del^IL^J\Box u + \del^IL^J(H_1[u]) $$ So we establish the following result: \begin{lemma}\label{lem 3 31-08-2017} Let $u$ be a function defined in $\Kcal$, sufficiently regular. Then the following estimate holds: \begin{equation}\label{eq 1 31-08-2017} \|(s^3/t)\del^IL^J H_1[u]\|_{L^2(\Hcal_s)}\leq C\sum_{|I'|+|J'|\leq|I|+|J|+1}\Ec(s,\del^{I'}L^{J'}u)^{1/2}. \end{equation} \end{lemma} \begin{proof}[Proof of lemma \ref{lem 3 31-08-2017}] For the first term in $H_1$, we see that $$ \aligned (s^3/t)\del^IL^J\left((x^a/t)s^{-1}L_a\delb_su\right) =& \sum_{I_1+I_2+I_3\atop J_1+J_2+J_3=J}(s^3/t)\del^{I_1}L^{J_1}(x^a/t)\del^{I_2}L^{J_2}(s^{-1})\del^{I_3}L^{J_3}L_a\delb_su. \endaligned $$ Recall that \eqref{eq 3 31-08-2017} and the fact that $x^a/t$ is homogeneous of degree zero, we see that (by lemma \ref{lem 2 24-08-2017}) $$ \aligned \|(s^3/t)\del^IL^J\left((x^a/t)s^{-1}L_a\delb_su\right)\|_{\Hcal_s}\leq& C\sum_{|I_3|\leq||I\atop |J_3|\leq|J|}\|(s^2/t)\del^{I_3}L^{J_3}L_a\delb_su\|_{L^2(\Hcal_s)} \\ \leq& C\sum_{|I'|\leq|I|\atop|J'|\leq|J|+1}\Ec(s,\del^{I'}L^{J'}u)^{1/2}. \endaligned $$ The rest terms are bounded similarly and we omit the detail. \end{proof} Finally we conclude by the following estimate: \begin{proposition}\label{prop 1 31-08-2017} Let $u$ be a function defined in $\Kcal$, sufficiently regular. Then the following bounds hold: \begin{equation}\label{eq 4 31-08-2017} \aligned \|(s^3/t)\del^IL^J\delb_s\delb_su\|_{L^2(\Hcal_s)} + \|(s^3/t)\delb_s\delb_s\del^IL^Ju\|_{L^2(\Hcal_s)}\leq& C\|(s^3/t)\Box \del^IL^J u\|_{L^2(\Hcal_s)} \\ &+ C\sum_{|I'|\leq|I|\atop|J'|\leq|J|+1}\Ec(s,\del^{I'}L^{J'}u)^{1/2}. \endaligned \end{equation} \end{proposition} \subsection{$L^{\infty}$ bounds} The $L^{\infty}$ bounds are based on the above $L^2$ bounds and the global Sobolev inequality \eqref{ineq 1 sobolev}. We first establish the following results: \begin{lemma}\label{lem 5 31-08-2017} Let $u$ be a function defined in $\Kcal$, sufficiently regular. Then the following bounds hold: \begin{equation}\label{eq 5 31-08-2017} \aligned \|\del^{I'}L^{J'}\left((s^3/t)\del^IL^J\delb_s\delb_su\right)\|_{L^2(\Hcal_s)} \leq& C\sum_{|I''|\leq|I'|+|I|\atop|J''|\leq|J'|+|J|}\|(s^3/t)\Box \del^{I''}L^{J''} u\|_{L^2(\Hcal_s)} \\ &+ C\sum_{|I''|\leq|I|+|I'|\atop|J''|\leq|J|+|J'|+1}\Ec(s,\del^{I''}L^{J''}u)^{1/2}. \endaligned \end{equation} \begin{equation}\label{eq 7 31-08-2017} \aligned \|\del^{I'}L^{J'}\left((s^3/t)\delb_s\delb_s\del^IL^Ju\right)\|_{L^2(\Hcal_s)}\leq& C\sum_{|I''|\leq|I'|+|I|\atop|J''|\leq|J'|+|J|}\|(s^3/t)\Box \del^{I''}L^{J''} u\|_{L^2(\Hcal_s)} \\ &+ C\sum_{|I''|\leq|I|+|I'|\atop|J''|\leq|J|+|J'|+1}\Ec(s,\del^{I''}L^{J''}u)^{1/2}. \endaligned \end{equation} \end{lemma} To prove this we first remark the following bound: \begin{equation}\label{eq 6 31-08-2017} |\del^IL^J (s^3/t)|\leq C(s^3/t). \end{equation} This is checked by applying \eqref{eq 1 20-08-2017} and \eqref{eq 4 22-08-2017}: $$ |\del^IL^J (s^3/t)| = \del^IL^J(s\cdot s^2/t) \leq C\sum_{I_1+I_2=I\atop J_1+J_2=J}|\del^{I_1}L^{J_1}s\del^{I_2}L^{J_2}(s^2/t)| \leq Cs(s^2/t)\leq Cs^3/t. $$ \begin{proof}[Proof of lemma \ref{lem 5 31-08-2017}] For \eqref{eq 5 31-08-2017}, we see that by \eqref{eq 6 31-08-2017}: $$ \aligned \|\del^{I'}L^{J'}\left((s^3/t)\del^IL^J\delb_s\delb_su\right)\|_{L^2(\Hcal_s)} \leq& C\sum_{I_1'+I_2'=I'\atop J_1'+J_2'=J'}\|\del^{I_1'}L^{J_1'}(s^3/t)\cdot \del^{I_2'}L^{J_2'}\del^IL^J\delb_s\delb_su\|_{L^2(\Hcal_s)} \\ \leq& C\sum_{|I_2'|\leq|I'|\atop |J_2'|\leq|J'|}\|(s^3/t)\del^{I_2'}L^{J_2'}\del^IL^J\delb_s\delb_su\|_{L^2(\Hcal_s)} \\ \leq& C\sum_{|I''|\leq|I|+|I'|\atop |J''|\leq|J|+|J'|}\|(s^3/t)\del^{I''}L^{J''}\delb_s\delb_s u\|_{L^2(\Hcal_s)} \endaligned $$ where in the last inequality we have applied lemma \ref{lem 2 29-08-2017}. Then by \eqref{eq 4 31-08-2017}, we see that \eqref{eq 5 31-08-2017} is established. \eqref{eq 7 31-08-2017} is more complicated. We see that by \eqref{eq 6 31-08-2017}, $$ \aligned \|\del^{I'}L^{J'}\left((s^3/t)\delb_s\delb_s\del^IL^Ju\right)\|_{L^2(\Hcal_s)} \leq& C\sum_{|I_2'|\leq |I'|\atop|J_2'|\leq |J'|}\|(s^3/t)\del^{I_2'}L^{J_2'}\delb_s\delb_s\del^IL^Ju\|_{L^2(\Hcal_s)}. \endaligned $$ To bound this term we observe that by \eqref{eq 8 31-08-2017} $$ \aligned \|(s^3/t)\del^{I_2'}L^{J_2'}\delb_s\delb_s\del^IL^Ju\|_{L^2(\Hcal_s)} \leq& C\sum_{\alpha,\beta,|I''|\leq|I_2'|,|J''|\leq|J_2'|\atop|I''|+|J''|<|I_2'|+|J_2'|} \|s^2(s/t)^3\del_{\alpha}\del_{\beta}\del^{I''}L^{J''}\del^IL^Ju\|_{L^2(\Hcal_s)} \\ &+ C\sum_{\alpha,|I''|\leq|I_2'|,|J''|\leq|J_2'|\atop|I''|+|J''|<|I_2'|+|J_2'|}\|s(s/t)^2\del_{\alpha}\del^{I''}L^{J''}\del^IL^Ju\|_{\Hcal_s}. \endaligned $$ Then we apply \eqref{eq 6 29-08-2017} and see that by lemma \ref{lem 1 01-09-2017} $$ \aligned \|(s^3/t)\del^{I_2'}L^{J_2'}\delb_s\delb_s\del^IL^Ju\|_{L^2(\Hcal_s)}\leq& C\sum_{\alpha,\beta,|I''|\leq|I|+|I'|\atop|J''|\leq|J'|+|J|} \|s^2(s/t)^3\del_{\alpha}\del_{\beta}\del^{I''}L^{J''}u\|_{L^2(\Hcal_s)} \\ &+ C\sum_{\alpha,|I''|\leq|I'|+|I|\atop |J''|\leq|J'|+|J|}\|s(s/t)^2\del_{\alpha}\del^{I''}L^{J''}u\|_{\Hcal_s} \\ \leq& \sum_{|I''|\leq|I'|+|I|\atop |J''|\leq|J'|+|J|}\|(s^3/t)\delb_s\delb_s\del^{I''}L^{J''}u\|_{\Hcal_s} \\ &+\sum_{|I''|\leq|I'|+|I|\atop |J''|\leq|J'|+|J|+1} \Ec(s,\del^{I''}L^{J''}u)^{1/2} \endaligned $$ Then combined with \eqref{eq 4 31-08-2017}, the desired result is established. \end{proof} Now, combined with the global Sobolev's inequality \eqref{ineq 1 sobolev}, we have the following decay estimates: \begin{proposition} The $L^{\infty}$ bounds are based on the above $L^2$ bounds and the global Sobolev inequality \eqref{ineq 1 sobolev}. We first establish the following results: \begin{equation}\label{eq 3 01-09-2017} \aligned \sup_{\Hcal_s}\{s^3t^{1/2}|\del^IL^J\delb_s\delb_su|\} + \sup_{\Hcal_s}\{s^3t^{1/2}|\delb_s\delb_s\del^IL^Ju|\} \leq& C\sum_{|I'|+|J'|\leq|I|+|J|+2}\|(s^3/t)\Box \del^{I'}L^{J'} u\|_{L^2(\Hcal_s)} \\ &+ C\sum_{|I'|+|J'|\leq |I|+|J|+3}\Ec(s,\del^{I'}L^{J'}u)^{1/2}. \endaligned \end{equation} \end{proposition} \section{Null condition in hyperbolic frame} \subsection{Objective and basic calculations} The objective of this section is to give a first analysis on the following terms: $$ \del^IL^J\left(Q^{\alpha\beta\gamma}\del_{\gamma}u\del_{\alpha}\del_{\beta}u\right),\quad [\del^IL^J,Q^{\alpha\beta\gamma}\del_{\gamma}u\del_{\alpha}\del_{\beta}]u. $$ which will play essential role in the following sections. In this section we always suppose that $u$ is a function defined in $\Kcal$, sufficiently regular and vanishes near the conical boundary $\del\Kcal = \{t = |x|+1\}$. The first two subsections are preparations for the last one. Here we analyse the property of the quantity $r/t$ in the region $\{t/2<|x|<t\}$. We establish the following bounds: \begin{lemma}\label{lem 3 02-09-2017} In the region $\Kcal\cap \{t/2<|x|<t\}$, the following bounds hold with a constant $C$ determined by $I,J$: \begin{equation}\label{eq 5 02-09-2017} \left|\del^IL^J (r/t)\right|\leq Ct^{-|I|},\quad \left|\del^IL^J\left(t^{1/2}(t+r)^{-1/2}\right)\right|\leq Ct^{-|I|} \end{equation} \begin{equation}\label{eq 5' 02-09-2017} \del^IL^J\left((t-r)^{1/2}(t+r)^{-1/2}\right)\leq \left\{ \aligned &Cs^{-1},\quad |I|\geq 1, \\ &Cs/t,\quad |I|=0. \endaligned \right. \end{equation} \end{lemma} \begin{proof} We recall that $r^2/t^2$ is homogeneous of degree zero. For the convenience of expression, we de note by $$ \aligned f: (1/4,+\infty) &\rightarrow \RR \\ x &\rightarrow x^{1/2} \endaligned $$ and $v := (r/t)^2$. Thus we see that $r/t = f(v)$. Then (let $|I|+|J|=N\geq 1$) $$ \aligned \del^IL^J (r/t) = \del^IL^J\left(f(v)\right) = \sum_{1\leq k\leq N}\sum_{I_1+I_2+\cdots+I_k = I\atop J_1+J_2+\cdots+J_k=J}f^{(n)}(v)\cdot \del^{I_1}L^{J_1}v\cdots\del^{I_k}L^{J_k}v \endaligned $$ Thus we see that because $v$ is homogeneous of degree zero, and the fact that $f^{(n)}$ is bounded on $(1/4,+\infty)$ by a constant $C$ (determined by $n\geq 1$). Then we see that for $1/2<r/t<1$, $\del^IL^J (r/t)$ is bounded by $Ct^{-|I|}$. For $N=0$, we see that $r/t<1$. Thus the first term is correctly bounded. For the second term in \eqref{eq 5 02-09-2017}, we see that $$ \left(\frac{t}{t+r}\right)^{1/2} = (1+r/t)^{-1/2} = f(1+r/t). $$ Then we see that $$ \del^IL^J\left(f(1+r/t)\right) = \sum_{1\leq k\leq N}\sum_{I_1+I_2+\cdots+I_k = I\atop J_1+J_2+\cdots+J_k=J}f^{(n)}(1+(r/t))\cdot\del^{I_1}L^{J_1}(1+(r/t))\cdots\del^{I_k}L^{J_k}(1+(r/t)). $$ We see that $3/2<1+r/t<2$, and by the bounds on $r/t$, we see that the second bound is established. For the third term, we observe that $$ \del^IL^J\left((t-r)^{1/2}(t+r)^{-1/2}\right) = \del^IL^J\left((s/t)\frac{t}{t+r}\right) = \sum_{I_1+I_2=I\atop J_1+J_2=J}\del^{I_1}L^{J_1}(s/t)\cdot \del^{I_2}L^{J_2}\frac{t}{t+r}. $$ We recall \eqref{eq 2 20-08-2017} and for the second factor, we see that $$ \del^IL^J\left(\frac{t}{t+r}\right) = \del^IL^J\left((1+(r/t))^{-1}\right) = \sum_{I_1+I_2=I\atop J_1+J_2=J}\del^{I_1}L^{J_1}\left((1+(r/t))^{-1/2}\right)\del^{I_2}L^{J_2}\left(1+(r/t)^{-1/2}\right). $$ Then by the bound of the second term in \eqref{eq 5 02-09-2017}, the desired bound is established. \end{proof} \begin{corollary} By lemma \ref{lem 3 02-09-2017}, for a integer $k$, in the region $1/2\leq r/t\leq 1$, \begin{equation}\label{eq 6 02-09-2017} \bigg|\del^IL^J\left(\left(\frac{t-r}{t+r}\right)^k\right)\bigg|\leq \left\{ \aligned &C(s/t)^{k-1}s^{-1},\quad |I|\geq 1. \\ &C(s/t)^k ,\quad |I|=0 \endaligned \right. \end{equation} \begin{equation}\label{eq 6' 02-09-2017} \bigg|\del^IL^J\left(\left(\frac{t}{t+r}\right)^k\right)\bigg|\leq Ct^{-|I|}. \end{equation} \end{corollary} \subsection{Estimates on null forms} Recall the transition relation between $\Tb^{\alpha\beta}, \Qb^{\alpha\beta\gamma}$ and $T^{\alpha\beta},Q^{\alpha\beta\gamma}$, we see that the following terms are homogeneous of degree zero: \begin{equation}\label{eq 11 01-09-2017} \Tb^{ab}, \quad (s/t)\Tb^{a0},\quad (s/t)\Tb^{0a},\quad (s/t)^2\Tb^{00} \end{equation} and \begin{equation}\label{eq 2 02-09-2017} \aligned &\Qb^{abc},\quad (s/t)\Qb^{0bc},\quad (s/t)\Qb^{a0c},\quad (s/t)\Qb^{ab0}, \\ &(s/t)^2\Qb^{a00},\quad (s/t)^2\Qb^{0b0},\quad (s/t)^2\Qb^{00c},\quad (s/t)^3\Qb^{000}. \endaligned \end{equation} Based on the above observation, we have \begin{lemma}\label{lem 2 02-09-2017} In $\Kcal$, the following quantities are bounded by a constant $C$ which is determined by $I,J$: $$ \aligned &\del^IL^J\Qb^{abc},\quad \del^IL^J\Tb^{ab} \\ &(s/t)\del^IL^J\Qb^{0bc},\quad (s/t)\del^IL^J\Qb^{a0c},\quad (s/t)\del^IL^J\Qb^{ab0},\quad (s/t)\del^IL^J\Tb^{0b},\quad (s/t)\del^IL^J\Tb^{a0}, \\ &(s/t)^2\del^IL^J\Qb^{00c},\quad (s/t)^2\del^IL^J\Qb^{0b0},\quad (s/t)^2\del^IL^J\Qb^{a00},\quad (s/t)^2\del^IL^J\Tb^{00}, \\ &(s/t)^3\del^IL^J\Qb^{000}, \endaligned $$ and $$ \aligned &t\del_{\alpha}\del^IL^J\Qb^{abc},\quad t\del_{\alpha}\del^IL^J\Tb^{ab} \\ &s(s/t)^2\del_{\alpha}\del^IL^J\Qb^{0bc},\quad s(s/t)^2\del_{\alpha}\del^IL^J\Qb^{a0c},\quad s(s/t)^2\del_{\alpha}\del^IL^J\Qb^{ab0}, \\ &s(s/t)^2\del_{\alpha}\del^IL^J\Tb^{0b},\quad s(s/t)^2\del_{\alpha}\del^IL^J\Tb^{a0}, \\ &s(s/t)^3\del_{\alpha}\del^IL^J\Qb^{00c},\quad s(s/t)^3\del_{\alpha}\del^IL^J\Qb^{0b0},\quad s(s/t)^3\del_{\alpha}\del^IL^J\Qb^{a00},\quad s(s/t)^3\del_{\alpha}\del^IL^J\Tb^{00}, \\ &s(s/t)^4\del_{\alpha}\del^IL^J\Qb^{000}, \endaligned $$ \end{lemma} \begin{proof} This is by applying \eqref{eq 3 02-09-2017} and the fact that the terms in \eqref{eq 11 01-09-2017} and \eqref{eq 2 02-09-2017} are homogeneous of degree zero. We remark the following calculation: let $f$ be a homogeneous function of degree zero. We see that $$ \del^IL^J\left((t/s)^nf\right) = \sum_{I_1+I_2=I\atop J_1+J_2=J}\del^{I_1}L^{J_1}\left((t/s)^n\right)\cdot \del^{I_2}L^{J_2}f $$ and by applying \eqref{eq 3 02-09-2017} on the fist factor, we see that $$ \left|\del^IL^J\left((t/s)^nf\right)\right|\leq \left\{ \aligned &C(t/s)^{n+1}s^{-1},\quad |I|\geq 1, \\ &C(t/s)^n,\quad |I| = 0. \endaligned \right. $$ Then we observe that for $\Tb^{\alpha\beta}$ or $\Qb^{\alpha\beta\gamma}$, the expression $$ (t/s)^m \Tb^{\alpha\beta},\quad (t/s)^n\Qb^{\alpha\beta\gamma} $$ are homogeneous of degree zero where $m,n$ are the number of zero in $\alpha,\beta$ or $\alpha,\beta,\gamma$ respectively. This concludes the desired result. \end{proof} Remark the relation $$ \delb_a = t^{-1}L_a,\quad \delb_s = (s/t)\del_t $$ the following bounds are direct: \begin{lemma}\label{lem 1 16-09-2017} In $\Kcal$, the following terms are bounded by $C$: \begin{equation}\label{eq 4 16-09-2017} \aligned &t(t/s)\delb_s\Qb^{abc},\quad t(t/s)\delb_s\Tb^{ab}, \\ &s(s/t)\delb_s\Qb^{0bc},\quad s(s/t)\delb_s\Qb^{a0c},\quad s(s/t)\delb_s\Qb^{ab0},\quad s(s/t)\delb_s\Tb^{0b},\quad s(s/t)\delb_s\Tb^{a0}, \\ &s(s/t)^2\delb_s\Qb^{00c},\quad s(s/t)^2\delb_s\Qb^{0b0},\quad s(s/t)^2\delb_s\Qb^{a00},\quad s(s/t)^2\delb_s\Tb^{00}, \\ &s(s/t)^3\delb_s\Qb^{000}. \endaligned \end{equation} and \begin{equation}\label{eq 5 17-09-2017} \aligned &t\delb_a\Qb^{abc},\quad t\delb_a\Tb^{ab}, \\ &s\delb_a\Qb^{0bc},\quad s\delb_a\Qb^{a0c},\quad s\delb_a\Qb^{ab0},\quad s\delb_a\Tb^{0b},\quad s\delb_a\Tb^{a0}, \\ &s(s/t)\delb_a\Qb^{00c},\quad s(s/t)\delb_a\Qb^{0b0},\quad s(s/t)\delb_a\Qb^{a00},\quad s(s/t)\delb_a\Tb^{00}, \\ &s(s/t)^2\delb_a\Qb^{000}. \endaligned \end{equation} \end{lemma} Now we introduce the following notion of the null form. Let $T$ be a quadratic form defined in $\Kcal$ with constant coefficient (with respect to the canonical frame). We call $T$ a {\bf null quadratic form}, if for any $\xi\in\RR^4$ satisfying \begin{equation}\label{eq 8 01-09-2017} \xi_0^2 - \sum_{a=1}^3\xi_a^2 = 0 \end{equation} the following equation holds: \begin{equation}\label{eq 7 01-09-2017} T^{\alpha\beta}\xi_{\alpha}\xi_{\beta} = 0. \end{equation} We can also define the null condition for a cubic form: let $Q$ be a constant cubic form defined in $\Kcal$ and for any $\xi$ satisfying \eqref{eq 8 01-09-2017}, the following condition holds: \begin{equation}\label{eq 9 01-09-2017} Q^{\alpha\beta\gamma}\xi_{\alpha}\xi_{\beta}\xi_{\gamma} = 0. \end{equation} Then we establish the following important result: \begin{proposition}[Null condition in hyperbolic frame]\label{prop 1 02-09-2017} Let $T$ and $Q$ be bull quadratic and cubic form respectively. Then the following bounds hold: \begin{equation}\label{eq 10 01-09-2017} \big|\del^IL^J\Tb^{00}\big|\leq C,\quad \big|\del^IL^J\Qb^{000}\big|\leq C(t/s) \end{equation} and \begin{equation}\label{eq 10' 01-09-2017} \big|\del_{\alpha}\del^IL^J\Tb^{00}\big|\leq Cs^{-1},\quad \big|\del_{\alpha}\del^IL^J\Qb^{000}\big|\leq C(t/s)^2s^{-1} \end{equation} Furthermore, the following estimates hold: \begin{equation}\label{eq 6 17-09-2017} |\delb_\alpha \Tb^{00}|\leq Ct^{-1},\quad |\delb_s\Qb^{000}|\leq C(t/s)s^{-1},\quad |\delb_a\Qb^{000}|\leq Cs^{-1}. \end{equation} \end{proposition} \begin{proof}[Proof of proposition \ref{prop 1 02-09-2017}] We observe that in the region $\{(t,x)\in\RR^4||x|\leq t/2\}$, this is a direct result of the fact that $(s/t)^3\Qb^{000}$ and $(s/t)^2\Tb^{00}$ are homogeneous of degree zero. To see this, we denote by $$ f := (s/t)^3\Qb^{000},\quad g := (s/t)^2\Tb^{00}. $$ Then we see that $$ \del^IL^J \Qb^{000} = \del^IL^J \left((t/s)^3f\right) = \sum_{I_1+I_2=I\atop J_1+J_2=J}\del^{I_1}L^{J_1}(t/s)^3\cdot\del^{I_2}L^{J_2}f $$ then recalling that $|\del^IL^Jf|\leq C$, and by \eqref{eq 3 02-09-2017}, we see that $|\del^IL^J \Qb^{000}|\leq C(t/s)^3$. Remark that in the region $\{(t,x)\in\RR^4||x\leq t/2|\}$, $t/s\leq 4/3$. Thus the desired result is established. For $\Tb^{00}$ the argument is similar and we omit the detail. Then we discuss the region $\Kcal\cap\{t/2<|x|<t\}$. Let $$ \zeta_{\alpha} := \Psib_\alpha^0, \quad \xi = (r/t,x^1/t, x^2/t, x^3/t) . $$ We see that $\xi$ satisfies \eqref{eq 8 01-09-2017}. Furthermore, $$ \nu:= \zeta - (t/s)\xi = ((t-r)/s,0,0,0). $$ Now we see that $$ \aligned \Tb^{00} =& T^{\alpha\beta}\Psib_{\alpha}^0\Psib_{\beta}^0 = T^{\alpha\beta}(\nu_{\alpha} + (t/s)\xi_{\alpha})(\nu_{\beta} + (t/s)\xi_{\beta}) \\ =& T^{\alpha\beta}\nu_{\alpha}\nu_{\beta} + (t/s)T^{\alpha\beta}\nu_{\alpha}\xi_{\beta} + (t/s)T^{\alpha\beta}\xi_{\alpha}\nu_{\beta} + (t/s)^2T^{\alpha\beta}\xi_{\alpha}\xi_{\beta} \\ =& T^{\alpha\beta}\nu_{\alpha}\nu_{\beta} + (t/s)T^{\alpha\beta}\nu_{\alpha}\xi_{\beta} + (t/s)T^{\alpha\beta}\xi_{\alpha}\nu_{\beta} \\ =& \frac{t-r}{t+r}T^{00} + \frac{t}{t+r}T^{0\beta}\xi_{\beta} + \frac{t}{t+r}T^{\alpha 0}\xi_{\alpha}. \endaligned $$ where we have applied the null condition $T^{\alpha\beta}\xi_{\alpha}\nu_{\beta}=0$. Recall that $\xi_{\alpha}$ are homogeneous of degree zero, combined with \eqref{eq 6 02-09-2017}, $$ \big|\del^IL^J\Tb^{00}\big|\leq \left\{ \aligned &Cs^{-1},\quad |I|\geq 1, \\ &C,\quad |I| = 0. \endaligned \right. $$ Then we regard the cubic form. We see that similar to the quadratic case: $$ \aligned \Qb^{000} =& Q^{\alpha\beta\gamma}\nu_{\alpha}\nu_{\beta}\nu_{\gamma}+ (t/s)Q^{\alpha\beta\gamma}\left(\nu_{\alpha}\nu_{\beta}\xi_{\gamma} + \nu_{\alpha}\xi_{\beta}\nu_{\gamma} + \xi_{\alpha}\nu_{\beta}\nu_{\gamma}\right) \\ &+ (t/s)^2Q^{\alpha\beta\gamma}\left(\nu_{\alpha}\xi_{\beta}\xi_{\gamma} + \xi_{\alpha}\nu_{\beta}\xi_{\gamma} + \xi_{\alpha}\xi_{\beta}\nu_{\gamma}\right) +(t/s)^3Q^{\alpha\beta\gamma}\xi_{\alpha}\xi_{\beta}\xi_{\gamma} \\ =& Q^{\alpha\beta\gamma}\nu_{\alpha}\nu_{\beta}\nu_{\gamma}+ (t/s)Q^{\alpha\beta\gamma}\left(\nu_{\alpha}\nu_{\beta}\xi_{\gamma} + \nu_{\alpha}\xi_{\beta}\nu_{\gamma} + \xi_{\alpha}\nu_{\beta}\nu_{\gamma}\right) \\ &+ (t/s)^2Q^{\alpha\beta\gamma}\left(\nu_{\alpha}\xi_{\beta}\xi_{\gamma} + \xi_{\alpha}\nu_{\beta}\xi_{\gamma} + \xi_{\alpha}\xi_{\beta}\nu_{\gamma}\right) \\ =&(t/s)Q^{\alpha\beta\gamma}\left(\nu_{\alpha}\nu_{\beta}\xi_{\gamma} + \nu_{\alpha}\xi_{\beta}\nu_{\gamma} + \xi_{\alpha}\nu_{\beta}\nu_{\gamma}\right) \\ &+ (t/s)^2Q^{\alpha\beta\gamma}\left(\nu_{\alpha}\xi_{\beta}\xi_{\gamma} + \xi_{\alpha}\nu_{\beta}\xi_{\gamma} + \xi_{\alpha}\xi_{\beta}\nu_{\gamma}\right) \endaligned $$ where we have applied the null condition. We see that $$ (t/s)Q^{\alpha\beta\gamma}\left(\nu_{\alpha}\nu_{\beta}\xi_{\gamma} + \nu_{\alpha}\xi_{\beta}\nu_{\gamma} + \xi_{\alpha}\nu_{\beta}\nu_{\gamma}\right) = \frac{t(t-r)}{s^2}f = \frac{t}{t+r}f $$ where $f$ is a homogeneous function of degree zero. Also, $$ (t/s)^2Q^{\alpha\beta\gamma}\left(\nu_{\alpha}\xi_{\beta}\xi_{\gamma} + \xi_{\alpha}\nu_{\beta}\xi_{\gamma} + \xi_{\alpha}\xi_{\beta}\nu_{\gamma}\right) =\frac{t^2}{(t+r)^2}\left(\frac{t+r}{t-r}\right)^{1/2}f. $$ Then also by \eqref{eq 6 02-09-2017}, the bound \eqref{eq 10 01-09-2017} is established. For \eqref{eq 6 17-09-2017}, it is by direct calculation and the following relation in $\Kcal\cap \{r\geq t/2\}$: \begin{equation}\label{eq 7 17-09-2017} \left|\delb_s\left(\frac{t-r}{t+r}\right)\right|\leq C(s/t)t^{-1},\quad \left|\delb_a\left(\frac{t-r}{t+t}\right)\right|\leq C(s/t)^2t^{-1}. \end{equation} and \begin{equation}\label{eq 7 17-09-2017} \left|\del_\alpha\left(\frac{r}{t+r}\right)\right|\leq Ct^{-1}. \end{equation} \end{proof} \subsection{Analysis on null quadratic form} We first remark the following null decomposition for $u$ defined in $\Kcal$, sufficiently regular: \begin{equation}\label{eq 5 03-09-2017} \aligned Q^{\alpha\beta\gamma}\del_{\gamma}u\del_{\alpha}\del_{\beta}u =& \Qb^{\alpha\beta\gamma}\delb_{\gamma}u\delb_{\alpha}\delb_{\beta}u + Q^{\alpha\beta\gamma}\del_{\alpha}\Psib_{\beta}^{\beta'}\del_{\gamma}u\delb_{\beta'}u \\ =& \Qb^{000}\delb_su\delb_s\delb_su \\ &+ \Qb^{00c}\delb_cu\delb_s\delb_su + \Qb^{0b0}\delb_su\delb_s\delb_bu + \Qb^{a00}\delb_su\delb_a\delb_su \\ &+ \Qb^{0bc}\delb_cu\delb_s\delb_bu + \Qb^{a0c}\delb_cu\delb_a\delb_su + \Qb^{ab0}\delb_su\delb_a\delb_bu + \Qb^{abc}\delb_cu\delb_a\delb_bu \\ &+ Q^{\alpha\beta\gamma}\del_{\alpha}\Psib_{\beta}^{\beta'}\del_{\gamma}u\delb_{\beta'}u \\ =:& \Qb^{000}\delb_su\delb_s\delb_su + \Qb_1^{\alpha\beta}(\delb u)\delb_{\alpha}\delb_{\beta}u + Q^{\alpha\beta\gamma}\del_{\alpha}\Psib_{\beta}^{\beta'}\del_{\gamma}u\delb_{\beta'}u. \endaligned \end{equation} For the last term in the right-hand-side of the above equation, we recall \eqref{eq 2 14-08-2017} and \eqref{eq 3 14-08-2017}: \begin{equation}\label{eq 3 08-09-2017} \aligned &-s\del_t\Psib_0^0 = \Psib_0^0\Psib_0^0-1,\quad -s\del_t\Psib_a^0 = \Psib_0^0\Psib_a^0, \\ &-s\del_a\Psib_0^0 = \Psib_0^0\Psib_a^0,\quad -s\del_a\Psib_b^0 = \delta_{ab} + \Psib_a^0\Psib_b^0 \endaligned \end{equation} The rest component of $\del_{\alpha}\Psib_{\beta}^{\gamma}$ are zero. Thus we see that $$ \aligned Q^{\alpha\beta\gamma}\del_{\alpha}\Psib_{\beta}^{\beta'}\del_{\gamma}u\delb_{\beta'}u =&Q^{\alpha\beta\gamma}\del_{\alpha}\Psib_{\beta}^{0}\cdot\Psib_{\gamma}^{\gamma'} \delb_{\gamma'}u\delb_su \\ =&-s^{-1}Q^{00\gamma}(\Psib_0^0\Psib_0^0-1)\Psib_{\gamma}^{\gamma'}\delb_{\gamma'}u\delb_su - s^{-1}Q^{a0\gamma}\Psib_a^0\Psib_0^0\Psib_{\gamma}^{\gamma'}\delb_{\gamma'}u\delb_su \\ &-s^{-1}Q^{0a\gamma}\Psib_0^0\Psib_a^0\Psib_{\gamma}^{\gamma'}\delb_{\gamma'}u\delb_su -s^{-1}Q^{ab\gamma}\left(\delta_{ab} + \Psib_a^0\Psib_b^0\right)\Psib_{\gamma}^{\gamma'}\delb_{\gamma'}u\delb_su \\ =&-s^{-1}\Qb^{00\gamma}\delb_{\gamma}u\delb_su + s^{-1}Q^{00\gamma}\del_{\gamma}u\delb_su - s^{-1}\sum_aQ^{aa\gamma}\del_{\gamma}u\delb_su \endaligned $$ that is \begin{equation}\label{eq 4 08-09-2017} Q^{\alpha\beta\gamma}\del_{\alpha}\Psib_{\beta}^{\beta'}\del_{\gamma}u\delb_{\beta'}u =-s^{-1}\Qb^{00\gamma}\delb_{\gamma}u\delb_su + s^{-1}Q^{00\gamma}\del_{\gamma}u\delb_su - s^{-1}\sum_aQ^{aa\gamma}\del_{\gamma}u\delb_su \end{equation} We first concentrate on $\del^IL^J\left(Q^{\alpha\beta\gamma}\del_{\gamma}u\del_{\alpha}\del_{\beta}u\right)$. Applying lemma \ref{lem 2 02-09-2017} and proposition \ref{prop 1 02-09-2017}, we see that by \eqref{eq 5 03-09-2017} and \eqref{eq 4 08-09-2017} ,it is bounded by the sum of the following terms (modulo a constant determined by $I,J$): \begin{equation}\label{eq 1 03-09-2017} \aligned &(s/t)^{-1}\big|\del^{I_1}L^{J_1}\delb_su\del^{I_2}L^{J_2}\delb_s\delb_su\big|,\quad (s/t)^{-2}\big|\del^{I_1}L^{J_1}\delb_su\del^{I_2}L^{J_2}\delb_s\delb_au\big|, \\ &(s/t)^{-2}\big|\del^{I_1}L^{J_1}\delb_cu\del^{I_2}L^{J_2}\delb_s\delb_su\big|, \\ &(s/t)^{-1}\big|\del^{I_1}L^{J_1}\delb_cu\del^{I_2}L^{J_2}\delb_a\delb_su\big|,\quad (s/t)^{-1}\big|\del^{I_1}L^{J_1}\delb_su\del^{I_2}L^{J_2}\delb_a\delb_bu\big|, \\ &\big|\del^{I_1}L^{J_1}\delb_cu\del^{I_2}L^{J_2}\delb_a\delb_bu\big|, \\ & (s^2/t)^{-1}\big|\del^{I_1}L^{J_1}\delb_su\del^{I_2}L^{J_2}\delb_su\big|,\quad s^{-1}(s/t)^{-2}\big|\del^{I_1}L^{J_1}\delb_au\del^{I_2}L^{J_2}\delb_su\big|, \\ & s^{-1}|\del^{I_1}L^{J_1}\del_{\gamma}u\del^{I_2}L^{J_2}\delb_su| \endaligned \end{equation} where $|I_1|+|I_2|\leq|I|$ and $|J_1|+|J_2|\leq |J|$. For the last term we have applied the fact that $\del_{\alpha}\Psib_{\beta}^{\beta'}$ is homogeneous of degree $-1$. Then we regard $[\del^IL^J,Q^{\alpha\beta\gamma}\del_{\gamma}u\del_{\alpha}\del_{\beta}]u$. We see that by \eqref{eq 5 03-09-2017}, it is the sum of the following terms: $$ [\del^IL^J,\Qb^{000}\delb_su\delb_s\delb_s]u,\quad [\del^IL^J, \Qb_1^{\alpha\beta}(\delb u)\delb_{\alpha}\delb_{\beta}]u, \quad [\del^IL^J,Q^{\alpha\beta\gamma}\del_{\alpha}\Psib_{\beta}^{\beta'}\del_{\gamma}u\delb_{\beta'}]u $$ We see that in general the following calculation holds: \begin{equation}\label{eq 2 03-09-2017} \aligned \,[\del^IL^J,\Qb^{\alpha\beta\gamma}\delb_{\gamma}u\delb_{\alpha}\delb_{\beta}]u =& \sum_{{I_1+I_2+I_3=I\atop J_1+J_2+J_3=J}\atop |I_3|+|J_3|<|I|+|J|} \del^{I_1}L^{J_1}\Qb^{\alpha\beta\gamma}\del^{I_2}L^{J_2}\delb_{\gamma}u\del^{I_3}L^{J_3}\delb_{\alpha}\delb_{\beta}u \\ &+\Qb^{\alpha\beta\gamma}\delb_{\gamma}u[\del^IL^J,\delb_{\alpha}\delb_{\beta}]u \endaligned \end{equation} and \begin{equation}\label{eq 3 03-09-2017} \aligned \,[\del^IL^J,Q^{\alpha\beta\gamma}\del_{\alpha}\Psib_{\beta}^{\beta'}\del_{\gamma}u\delb_{\beta'}]u =& -[\del^IL^J,s^{-1}\Qb^{00\gamma}\delb_{\gamma}u\delb_s]u + [\del^IL^J,s^{-1}Q^{00\gamma}\del_{\gamma}u\delb_s]u \\ &- \sum_a[\del^IL^J,s^{-1}Q^{aa\gamma}\del_{\gamma}u\delb_s]u. \endaligned \end{equation} In the section \ref{sec 2 global 2} we will make $L^2$ estimates on these terms based on the bootstrap bounds. As an preparation, we establish the following bounds: \begin{lemma}\label{lem 1 07-09-2017} Let $u$ be a function defined in $\Kcal$, sufficiently regular. Then the following estimates hold: \begin{equation}\label{eq 2 07-09-2017} \aligned \big\|s^2(s/t)[\del^IL^J,\delb_s\delb_s]u\big\|_{L^2(\Hcal_s)}\leq& C\sum_{|I'|\leq|I|,|J'|\leq|J|\atop |I'|+|J'|<|I|+|J|}\|(s^3/t)\Box\del^{I'}L^{J'}u\|_{L^2(\Hcal_s)} \\ &+C\sum_{|I'|+|J'|\leq|I|+|J|}\!\!\!\!\!\!\!\!\Ec(s,\del^{I'}L^{J'}u)^{1/2} . \endaligned \end{equation} \begin{equation}\label{eq 3 07-09-2017} \|s^2[\del^IL^J,\delb_a\delb_s]u\|_{L^2(\Hcal_s)}\leq C\sum_{|I'|+|J'|\leq|I|+|J|}\!\!\!\!\!\!\!\!\Ec(s,\del^{I'}L^{J'}u)^{1/2}, \end{equation} and \begin{equation}\label{eq 4 07-09-2017} \|st[\del^IL^J,\delb_a\delb_b]u\|_{L^2(\Hcal_s)}\leq C\sum_{|I'|+|J'|\leq|I|+|J|}\!\!\!\!\!\!\!\!\Ec(s,\del^{I'}L^{J'}u)^{1/2}. \end{equation} \end{lemma} \begin{proof} We see first that by \eqref{eq 8 31-08-2017} and lemma \ref{lem 1 01-09-2017}: $$ \aligned &\big\|s^2(s/t)[\del^IL^J,\delb_s\delb_s]u\big\|_{L^2(\Hcal_s)} \\ \leq& C\sum_{\alpha,\beta,|I'|\leq|I|,|J'|\leq|J|\atop |I'|+|J'|<|I|+|J|}\!\!\!\!\!\!\!\! \big\|s^2(s/t)^3\del_{\alpha}\del_{\beta}\del^{I'}L^{J'}u\big\|_{L^2(\Hcal_s)} + C\sum_{\alpha,|I'|\leq|I|,|J'|\leq|J|\atop |I'|+|J'|<|I|+|J|}\!\!\!\!\!\!\!\! \big\|s(s/t)^2\del_{\alpha}\del^{I'}L^{J'}u\big\|_{L^2(\Hcal_s)} \\ \leq& C\sum_{|I'|\leq|I|,|J'|\leq|J|\atop |I'|+|J'|<|I|+|J|} \!\!\!\!\!\!\!\!\|s^2(s/t)\delb_s\delb_s\del^{I'}L^{J'}u\|_{L^2(\Hcal_s)} +C\sum_{|I'|+|J'|\leq|I|+|J|}\!\!\!\!\!\!\!\!\Ec(s,\del^{I'}L^{J'}u)^{1/2} \endaligned $$ Then by proposition \ref{prop 1 31-08-2017}, we see that $$ \aligned \big\|s^2(s/t)[\del^IL^J,\delb_s\delb_s]u\big\|_{L^2(\Hcal_s)} \leq& C\sum_{|I'|\leq|I|,|J'|\leq|J|\atop |I'|+|J'|<|I|+|J|}\|(s^3/t)\Box\del^{I'}L^{J'}u\|_{L^2(\Hcal_s)} \\ &+C\sum_{|I'|+|J'|\leq|I|+|J|}\!\!\!\!\!\!\!\!\Ec(s,\del^{I'}L^{J'}u)^{1/2} . \endaligned $$ The rest to estimates are direct by lemma \ref{lem 3 03-09-2017} and lemma \ref{lem 1 05-09-2017}, we omit the detail. \end{proof} \section{Global existence: bootstrap argument}\label{sec 1 global 1} \subsection{The bootstrap bounds} We consider the main equation of interest together with initial data : \begin{equation}\label{eq 1 30-06-2017} \left\{ \aligned &\Box u + Q^{\alpha\beta\gamma}\del_{\gamma}u\del_{\alpha}\del_{\beta}u = 0 \\ &u|_{\Hcal_2}= u_0, \quad \del_tu|_{\Hcal_2}=u_1. \endaligned \right. \end{equation} where $u_i$ are sufficiently regular functions defined on the hyperboloid $\Hcal_2$ and supported in $\Hcal_2\cap \Kcal$. \begin{remark} The fact that the initial data are posed on hyperboloid is not a standard but we can see it in the following way: we pose the initial data set on the hyperplane $\{t=2\}$ and supported in the unit disc. Then by standard local existence result, the associated local solution extends to region $\{(t,x)| 2\leq t\leq \sqrt{r^2+4}\}\cap \Kcal$. Then we can restrict the solution on $\Hcal_2$. Thus we for global result we can pose our initial data on $\Hcal_2$. For more detail, see for example \cite{H1} or \cite{PLF-MY-book}. \end{remark} We will apply the so-called bootstrap argument, explained in detail here: Let $u$ be the local-in-time solution associated to \eqref{eq 1 30-06-2017}. Assume that the largest hyperbolic time interval of existence is $\Kcal_{[2,s^*)}$. We define the bootstrap bounds for $s\in [2,s^*]$:, \begin{equation}\label{eq 5 24-08-2017} \sum_{|I|+|J|\leq N}E_{con}(s, \del^IL^Ju)^{1/2}\leq C_1\epsilon \end{equation} with $(C_1,\vep)$ a pair of positive constant to be determined. Then we define $s_1$ to be the largest (hyperbolic) time where $u$ satisfies this condition, that is, $$ s_1 :=\sup_s \left\{ s^*> s \geq 2| \text{\eqref{eq 5 24-08-2017} holds on} [s_0,s]\right\} $$ We suppose that \begin{equation}\label{eq 6 24-08-2017} \sum_{|I|+|J|\leq N}E_{con}(2, \del^IL^Ju)^{1/2}\leq C_0\epsilon. \end{equation} which can be guaranteed by the smallness of the initial data. We see that when taking $C_1>C_0$, by continuity, $s_1>2$. To argue by contradiction, we suppose that $s_1<s^*$. If we could deduce, for a suitable pair $(C_1,\vep_0)$, a {\sl improved} bound for all $0< \vep\leq \vep_0$: \begin{equation}\label{eq 2 29-08-2017} \sum_{|I|+|J|\leq N}E_{con}(s, \del^IL^Ju)^{1/2}\leq \frac{1}{2}C_1\epsilon \ \ \text{for}\ s\in [2,s_1], \end{equation} On the other hand, we see that by continuity, $$ \sum_{|I|+|J|\leq N}E_{con}(s_1, \del^IL^Ju)^{1/2}= C_1\epsilon . $$ This contradiction leads to the fact that $s_1=s^*$, then $$ \sum_{|I|+|J|\leq N}E_{con}(s^*, \del^IL^Ju)^{1/2}\leq C_1\epsilon. $$ Then by standard local-in-time theory (with $N$ sufficiently large), we see that $s^*$ could not be finite. This leads to the global existence result. Now we state the main result of this article: \begin{theorem}\label{thm main} There exists a constant $\vep_0>0$, determined only by the system \eqref{eq 1 30-06-2017}, such for all $0\leq \vep\leq \vep_0$, if \begin{equation}\label{eq 1 thm cond} \|u_0\|_{H^{N+1}(\Hcal_s)} + \|u_1\|_{H^{N}(\Hcal_s)}\leq \vep \end{equation} holds for $N$ sufficiently large ($N\geq 9$ is enough), then the associated local-in-time solution extends to time infinity. \end{theorem} Based on the above discussion on bootstrap argument, we see that the above result is deduced from the following proposition: \begin{proposition}\label{prop 1 01-07-2017} There exists a pair of positive constant $(C_1,\vep_0)$, determined by only by the system \eqref{eq 1 30-06-2017} such that if the initial data set satisfies \eqref{eq 6 24-08-2017} with $0< \vep\leq \vep_0$, then \eqref{eq 5 24-08-2017} leads to \eqref{eq 2 29-08-2017}. \end{proposition} The following sections from \ref{sec 1 global 1} to \ref{sec 4 global 4} are devoted to the proof of this proposition. \subsection{Basic bounds} The following bounds hold in the region $\Kcal_{[2,s_1]}$. The following terms are bounded by $CC_1\vep $ for $|I|+|J|\leq N$: \begin{equation}\label{eq 3 01-07-2017} \|s\delb_a\del^IL^Ju\|_{L^2(\Hcal_s)},\quad \|(s^2/t)\delb_s\del^IL^Ju\|_{L^2(\Hcal_s)},\quad \|(s/t)\del^IL^Ju\|_{L^2(\Hcal_s)}. \end{equation} Then by \eqref{eq 3 24-08-2017}, \eqref{eq 6 05-09-2017} and \eqref{eq 5 07-09-2017}, the following bounds are also bounded by $CC_1\vep $: \begin{equation}\label{eq 4 01-07-2017} \aligned &\|s\del^IL^J\delb_au\|_{L^2(\Hcal_s)},\quad \|(s^2/t)\del^IL^J\delb_su\|_{L^2(\Hcal_s)} \\ &\|s^2\del^IL^J\delb_s\delb_au\|_{L^2(\Hcal_s)},\quad \|st\del^IL^J\delb_a\delb_bu\|_{L^2(\Hcal_s)},\quad \|s\del^IL^J\delb_s\delb_su\|_{L^2(\Hcal_s)}. \endaligned \end{equation} By proposition \ref{prop 1 29-08-2017} and the global Sobolev inequality, for $|I|+|J|\leq N-2$, the following terms are bounded by $CC_1\vep $: \begin{equation}\label{eq 5 01-07-2017} \sup_{\Hcal_s}\left\{t^{3/2}s\delb_a\del^IL^J u\right\},\quad \sup_{\Hcal_s}\left\{t^{1/2}s^2 \delb_s\del^IL^J u\right\},\quad \sup_{\Hcal_s}\left\{t^{1/2}s\del^IL^J u\right\}, \end{equation} \begin{equation}\label{eq 4 01-09-2017} \sup_{\Hcal_s}\left\{t^{3/2}s\del^IL^J\delb_au\right\},\quad \sup_{\Hcal_s}\left\{t^{1/2}s^2\del^IL^J \delb_su\right\}, \end{equation} and for $|I|+|J|\leq N-3$, the following terms are bounded by $CC_1\vep $: \begin{subequations} \begin{equation}\label{eq 7 07-09-2017} \sup_{\Hcal_s}\left\{s^2t^{3/2}\delb_s\delb_a\del^IL^J u\right\},\quad \sup_{\Hcal_s}\left\{st^{5/2}\delb_a\delb_b\del^IL^Ju\right\} \end{equation} \begin{equation}\label{eq 8 07-09-2017} \sup_{\Hcal_s}\left\{s^2t^{3/2}\del^IL^J\delb_s\delb_a u\right\},\quad \sup_{\Hcal_s}\left\{st^{5/2}\del^IL^J\delb_a\delb_bu\right\} \end{equation} and \begin{equation}\label{eq 14 07-09-2017} \sup_{\Hcal_s}\left\{st^{3/2}\del^IL^J\delb_s\delb_su\right\}. \end{equation} where for the last term in the above list we applied \eqref{eq 10 07-09-2017} \end{subequations} \section{Global existence: refined bounds}\label{sec 2 global 2} \subsection{Estimates on Hessian form} We combine \eqref{eq 5 24-08-2017} together with proposition \eqref{prop 1 31-08-2017}: \begin{equation}\label{eq 5 01-09-2017} \|(s^3/t)\del^IL^J\delb_s\delb_su\|_{L^2(\Hcal_s)} + \|(s^3/t)\delb_s\delb_s\del^IL^Ju\|_{L^2(\Hcal_s)}\leq C\|(s^3/t)\Box \del^IL^J u\|_{L^2(\Hcal_s)} + CC_1\vep . \end{equation} Similar bounds hold for the combination of \eqref{eq 5 24-08-2017}with \eqref{eq 3 01-09-2017}. Thus we need to control $\|(s^3/t)\Box \del^IL^J u\|_{L^2(\Hcal_s)}$. This is by the following lemma: \begin{lemma}\label{lem 2 01-09-2017} Under the bootstrap bound \eqref{eq 5 24-08-2017}, the following estimate holds for $|I|+|J|\leq N-1$: \begin{equation}\label{eq 6 01-09-2017} \big\|(s^3/t)\del^IL^J\left(Q^{\alpha\beta\gamma}\del_{\gamma}u\del_{\alpha}\del_{\beta}u\right)\big\|_{L^2(\Hcal_s)}\leq C(C_1\vep)^2 . \end{equation} \end{lemma} \begin{proof} This is based on the $L^2$ bounds and $L^{\infty}$ bounds established in the last section. We need to bound each term in the list \eqref{eq 1 03-09-2017}. For each term in \eqref{eq 1 03-09-2017}, for $|I_1|+|J_1|\leq N-2$, we apply the decay bounds \eqref{eq 4 01-09-2017} on the first factor and the apply the $L^2$ bounds \eqref{eq 4 01-07-2017} on the second factor. We can check that for each term, the $L^2$ norm is bounded as $$ \|(s^3/t)X\|_{L^2(\Hcal_s)}\leq C(C_1\vep)^2 $$ where $X$ represents a term in \eqref{eq 4 01-09-2017}. When $|I_1|+|J_1|\geq N-1$, we see that $|I_2|+|J_2|\leq 1\leq N-3$. Thus in the similar way, we apply the decay estimates of \eqref{eq 8 07-09-2017}, \eqref{eq 14 07-09-2017} on the second factor and \eqref{eq 4 01-07-2017} (the first two terms) on the first factor. \end{proof} Now we are ready to establish the refined bound on $\del^IL^J\delb_s\delb_su$. \begin{lemma}\label{lem 3 07-09-2017} Under the bootstrap assumption, we see that \begin{equation}\label{eq 11 07-09-2017} \|(s^3/t)\del^IL^J\delb_s\delb_su\|_{L^2(\Hcal_s)} + \|(s^3/t)\delb_s\delb_s\del^IL^Ju\|_{L^2(\Hcal_s)}\leq CC_1\vep \end{equation} and \begin{equation}\label{eq 12 07-09-2017} \sup_{\Hcal_s}\{t^{1/2}s^3\del^IL^J\delb_s\delb_su\} + \sup_{\Hcal_s}\{s^3t^{1/2}\delb_s\delb_s\del^IL^Ju\}\leq CC_1\vep \end{equation} for $|I|+|J|\leq N-1$. \end{lemma} \begin{proof} This is by using the equation. We see that $$ \Box \del^IL^Ju = \del^IL^J\left(Q^{\alpha\beta\gamma}\del_{\gamma}u\del_{\alpha}\del_{\beta}u\right). $$ Thus by \eqref{eq 6 01-09-2017} \begin{equation}\label{eq 12' 07-09-2017} \|(s^3/t)\Box \del^IL^J u\|_{L^2(\Hcal_s)}\leq C(C_1\vep)^2 . \end{equation} Now we apply \eqref{eq 5 31-08-2017} and \eqref{eq 3 01-09-2017} together with the above bounds and the bootstrap bound on energy, and we see that the desired bounds is established. \end{proof} \subsection{Estimates on null form} In this section we concentrate on the $L^2$ bounds on $[\del^IL^J,Q^{\alpha\beta\gamma}\del_{\gamma}u\del_{\alpha}\del_{\beta}]u$. To get started we combine the bootstrap bounds with lemma \ref{lem 1 07-09-2017}, and we see that the following terms are bounded by $CC_1\vep $: \begin{equation}\label{eq 2' 07-09-2017} \aligned \big\|s^2(s/t)[\del^IL^J,\delb_s\delb_s]u\big\|_{L^2(\Hcal_s)}, \quad\|s^2[\del^IL^J,\delb_a\delb_s]u\|_{L^2(\Hcal_s)},\quad \|st[\del^IL^J,\delb_a\delb_b]u\|_{L^2(\Hcal_s)} \endaligned \end{equation} where $|I|+|J|\leq N$. Based on these bounds, we establish the following estimates: \begin{lemma}\label{lem 4 07-09-2017} Under the bootstrap assumption and assume that $Q^{\alpha\beta\gamma}$ be a null cubic form, then the following bounds hold: \begin{equation}\label{eq 13 07-09-2017} \|s\Qb^{\alpha\beta\gamma}\delb_{\gamma}u\cdot[\del^IL^J,\delb_{\alpha}\delb_{\beta}u]\|_{L^2(\Hcal_s)}\leq C(C_1\vep)^2 s^{-2 }. \end{equation} \end{lemma} \begin{proof} For $(\alpha,\beta,\gamma) = (0,0,0)$, we see that by \eqref{eq 10 01-09-2017}, \eqref{eq 4 01-09-2017} and \eqref{eq 2' 07-09-2017} (recall that in $\Kcal, s^2\geq t$): $$ \aligned \|s\Qb^{000}\delb_su\cdot[\del^IL^J,\delb_s\delb_su]\|_{L^2(\Hcal_s)} \leq& CC_1\vep\|s(t/s)t^{-1/2}s^{-2+\delta} \cdot (t/s)s^{-2}\cdot s^2(s/t)[\del^IL^J,\delb_s\delb_s]u\|_{L^2(\Hcal_s)} \\ \leq& C(C_1\vep)^2s^{-2 }. \endaligned $$ For the rest components, we apply lemma \ref{lem 2 02-09-2017}, \eqref{eq 4 01-09-2017} and \eqref{eq 2' 07-09-2017}. We omit the detail. \end{proof} \begin{lemma}\label{lem 5 07-09-2017} Under the bootstrap assumption, the $L^2$ norm of the following term $$ s\del^{I_1}L^{J_1}\Qb^{\alpha\beta\gamma}\del^{I_2}L^{J_2}\delb_{\gamma}u\del^{I_3}L^{J_3}\delb_{\alpha}\delb_{\beta}u $$ is controlled by $C(C_1\vep)^2s^{-2 }$, where $$ I_1+I_2+I_3=I,\quad J_1+J_2+J_3=J,\quad |I_3|+|J_3|<|I|+|J|\leq N. $$ \end{lemma} \begin{proof} This is also by applying \eqref{eq 10 01-09-2017} (for $(\alpha,\beta,\gamma) = (0,0,0)$) or lemma \ref{lem 2 02-09-2017} (for $(\alpha,\beta,\gamma)\neq (0,0,0)$). We see that when $|I_1|+|J_1|\leq N-2$, we apply \eqref{eq 4 01-09-2017} on the factor $\del^{I_2}L^{J_2}\delb_{\gamma}u$, \eqref{eq 4 01-07-2017} (the third and forth term) on the factor $\del^{I_3}L^{J_3}\delb_a\delb_{\beta}u$ and \eqref{eq 11 07-09-2017} on the factor $\del^{I_3}L^{J_3}\delb_s\delb_su$. For example for the component $(0,0,0)$, we see that $$ \aligned &\|s\del^{I_1}L^{J_1}\Qb^{000}\del^{I_2}L^{J_2}\delb_su\del^{I_3}L^{J_3}\delb_s\delb_su\|_{L^2(\Hcal_s)} \\ \leq& CC_1\vep\|s(t/s)\cdot t^{-1/2}s^{-2+\delta}\cdot(ts^{-3})\cdot (s^3/t)\del^{I_3}L^{J_3}\delb_s\delb_su\|_{L^2(\Hcal_s)}\leq C(C_1\vep)^2s^{-2 }. \endaligned $$ The rest components are verified similarly and we omit de detail. When $|I_1|+|I_2|\geq N-1$, we see that $|I_3|+|J_3|\leq 1\leq N-3$. Thus we apply \eqref{eq 4 01-07-2017} (the first two terms) on the factor $\del^{I_2}L^{J_2}\delb_\gamma u$, \eqref{eq 8 07-09-2017} on the factor $\del^{I_3}L^{J_3}\delb_a\delb_{\beta}u$ and \eqref{eq 12 07-09-2017} for $\del^{I_3}L^{J_3}\delb_a\delb_su$. For example for the component $(0,0,0)$ $$ \aligned &\|s\del^{I_1}L^{J_1}\Qb^{000}\del^{I_2}L^{J_2}\delb_su\del^{I_3}L^{J_3}\delb_s\delb_su\|_{L^2(\Hcal_s)} \\ \leq& CC_1\vep\|s(t/s)\cdot t^{-1/2}s^{-3+2\delta}\cdot (t/s^2)\cdot (s^2/t)\del^{I_2}L^{J_2}\delb_su \|_{L^2(\Hcal_s)}\leq C(C_1\vep)^2s^{-2 }. \endaligned $$ The rest components are verified similarly and we omit the detail. \end{proof} \begin{lemma}\label{lem 6 07-09-2017} Under the bootstrap assumption, the following estimate holds: \begin{equation}\label{eq 1 08-09-2017} \|s[\del^IL^J,Q^{\alpha\beta\gamma}\del_{\alpha}\Psib_{\beta}^{\beta'}\del_{\gamma}u\delb_{\beta'}]u\|_{L^2(\Hcal_s)}\leq C(C_1\vep)^2s^{-3/2 }. \end{equation} \end{lemma} \begin{proof} We recall \eqref{eq 3 03-09-2017}. For the first term in right-hand-side, we see that $$ \aligned \,[\del^IL^J,s^{-1}\Qb^{00\gamma}\delb_{\gamma}u\delb_s]u =& \sum_{I_1+I_2=I,J_1+J_2=J\atop |I_2|+|J_2|<|I|+|J|} \del^{I_1}L^{J_1}\left(s^{-1}\Qb^{00\gamma}\delb_{\gamma}u\right)\del^{I_2}L^{J_2}\delb_su \\ &+s^{-1}\Qb^{00\gamma}\delb_{\gamma}u[\del^IL^J,\delb_s]u \\ =:& T_1 + T_2. \endaligned $$ For $T_1$, we see that $$ \aligned \del^{I_1}L^{J_1}\left(s^{-1}\Qb^{00\gamma}\delb_{\gamma}u\right) =& \del^{I_{11}}L^{J_{11}}\left(s^{-1}\right)\del^{I_{12}}L^{J_{12}}\Qb^{00\gamma}\cdot\del^{I_{13}}L^{J_{13}}\delb_{\gamma}u \endaligned $$ Then applying \eqref{eq 3 31-08-2017} together with lemma \ref{lem 2 02-09-2017} (for $\gamma>0$) or \eqref{eq 10 01-09-2017} (for $\gamma=0$): $$ |T_1|\leq \left\{ \aligned &C(t/s^2)\sum_{|I_1|+|I_2|\leq|I|\atop|J_1|+|J_2|\leq|J|}|\del^{I_1}L^{J_1}\delb_su\del^{I_2}L^{J_2}\delb_su|, \gamma = 0, \\ &Cs^{-1}(t/s)^2\sum_{a,|I_1|+|I_2|\leq|I|\atop|J_1|+|J_2|\leq|J|}|\del^{I_1}L^{J_1}\delb_au\del^{I_2}L^{J_2}\delb_su|,\gamma = a >0. \endaligned \right. $$ Now, for $|I_1|+|J_1|\leq N-2$, we apply decay estimate \eqref{eq 4 01-09-2017} on the first factor and $L^2$ bounds \eqref{eq 4 01-07-2017} (the first two terms) on the second factor. When $|I_1|+|J_1|\geq N-1$, we see that $|I_2|+|J_2|\leq 1\leq N-2$. In this case we apply \eqref{eq 4 01-09-2017} on the second factor and \eqref{eq 4 01-07-2017} on the first factor. This leads to: $$ \|sT_1\|_{L^2(\Hcal_s)}\leq C(C_1\vep)^2s^{-2 }. $$ For the term $T_2$, we see that by \eqref{eq 4 01-09-2017} applied on $\del_{\gamma}u$, lemma \ref{lem 2 02-09-2017} (for $\gamma >0$) or \eqref{eq 10 01-09-2017} (for $\gamma = 0$) $$ |T_2|\leq\left\{ \aligned &Cs^{-4}t^{1/2}|[\del^IL^J,\delb_s]u|,\quad \gamma = 0, \\ &Cs^{-4}t^{1/2}|[\del^IL^J,\delb_a]u|,\quad \gamma>0. \endaligned \right. $$ Then we combine \eqref{eq 2 23-08-2017}, \eqref{eq 8 22-08-2017} together with \eqref{eq 3 01-07-2017}, the following bound is established $$ \|sT_2\|_{L^2(\Hcal_s)}\leq C(C_1\vep)^2s^{-2 }. $$ The rest terms in right-hand-side of \eqref{eq 3 03-09-2017} are bounded similarly, we omit the detail. \end{proof} Now we are ready to conclude the following result: \begin{proposition}\label{prop 1 08-9-2017} Under the bootstrap assumption, for $|I|+|J|\leq N$, the following estimate holds: \begin{equation}\label{eq 2 08-09-2017} \|s[\del^IL^J,Q^{\alpha\beta\gamma}\del_{\gamma}u\del_{\alpha}\del_{\beta}]u\|_{L^2(\Hcal_s)}\leq C(C_1\vep)^{2}s^{-2 }. \end{equation} \end{proposition} \section{Global existence: conclusion of bootstrap argument}\label{sec 4 global 4} Now we are ready to prove proposition \ref{prop 1 01-07-2017}. To do so we need to guarantee \eqref{eq 3 23-05-2017} and the bounds on $M_g(s)$ (with the notion in \eqref{eq 4 23-05-2017}). We first remark that by the notation in subsection \ref{subsec energy-indentity} $$ \hb^{\alpha\beta} = \Qb^{\alpha\beta\gamma}\delb_{\gamma}u. $$ For the convenience of discussion, we list out the following bounds on $\hb^{\alpha\beta}$. These are by \eqref{eq 4 01-09-2017} combined with lemma \ref{lem 2 02-09-2017}, \eqref{eq 10 01-09-2017}: \begin{equation}\label{eq 5 08-09-2017} \aligned &|\hb^{00}| \leq CC_1\vep t^{1/2}s^{-3 },\quad |\hb^{a0}| + |\hb^{0a}|\leq CC_1\vep t^{3/2}s^{-4}, \\ &|\hb^{ab}|\leq CC_1\vep t^{1/2}s^{-3 },\quad |h^{\alpha\beta}|\leq CC_1\vep t^{1/2}s^{-3 }. \endaligned \end{equation} Furthermore, we see that \begin{equation}\label{eq 1 25-10-2017} \delb_{\mu}\hb^{\alpha\beta} = \delb_{\mu}\Qb^{\alpha\beta\gamma}\cdot \delb_\gamma u + \Qb^{\alpha\beta\gamma}\delb_{\mu}\delb_{\gamma}u. \end{equation} We apply, lemma \ref{lem 2 02-09-2017} and \eqref{eq 10 01-09-2017} combined with \eqref{eq 4 01-09-2017}, \eqref{eq 8 07-09-2017} and \eqref{eq 12 07-09-2017}. \begin{equation}\label{eq 6 08-09-2017} \aligned &|\delb_s\hb^{00}|\leq C(C_1\vep)t^{1/2}s^{-4 },\quad |\delb_s\hb^{a0}| + |\delb_s\hb^{0a}|\leq C(C_1\vep)t^{3/2}s^{-5},\quad |\delb_s\hb^{ab}|\leq C(C_1\vep)t^{1/2}s^{-4 }, \\ &|\delb_a\hb^{00}|\leq C(C_1\vep)t^{-1/2}s^{-3 },\quad |\delb_c\hb^{a0}| + |\delb_c\hb^{0a}|\leq C(C_1\vep)t^{1/2}s^{-4 },\quad |\delb_c\hb^{ab}|\leq C(C_1\vep)t^{-1/2}s^{-3 }. \endaligned \end{equation} We also remark that $$ \delb_{\delta}h^{\alpha\beta} = Q^{\alpha\beta\gamma}\delb_{\delta}\del_{\gamma}u $$ that is $$ \aligned \delb_sh^{\alpha\beta} =& Q^{\alpha\beta0}\delb_s\del_tu + Q^{\alpha\beta c}\delb_s\del_cu \\ =&Q^{\alpha\beta0}\delb_s\left((t/s)\delb_su\right) + Q^{\alpha\beta c}\delb_s\left(\delb_cu - (x^c/s)\delb_s u\right) \\ =&\left((t/s)Q^{\alpha\beta0} - (x^c/s)Q^{\alpha\beta c}\right)\delb_s\delb_su + Q^{\alpha\beta c}\delb_s\delb_cu + \left((x^c/s^2)Q^{\alpha\beta c}- \frac{r^2}{ts^2}Q^{\alpha\beta 0}\right)\delb_su, \endaligned $$ which leads to \begin{equation}\label{eq 1 26-10-2017} |\delb_sh^{\alpha\beta}|\leq C(t/s)|\delb_s\delb_su| + C\sum_a|\delb_s\delb_a u| + C(t/s^2)|\delb_su|. \end{equation} Similar calculation shows that \begin{equation}\label{eq 2 26-10-2017} |\delb_ah^{\alpha\beta}|\leq C(t/s)|\delb_a\delb_su| + C\sum_c|\delb_a\delb_c u| + s^{-1}|\delb_su|. \end{equation} Then combined with \eqref{eq 12 07-09-2017}, \eqref{eq 7 07-09-2017} and \eqref{eq 5 01-07-2017}, we see that \begin{equation}\label{eq 3 26-10-2017} |\delb_sh^{\alpha\beta}|\leq Ct^{1/2}s^{-4 },\quad |\delb_ah^{\alpha\beta}|\leq Ct^{-1/2}s^{-3 }. \end{equation} Then we establish the following bounds: \begin{lemma}\label{lem 1 26-10-2017} Under the bootstrap assumption \eqref{eq 5 24-08-2017}, $$ |\delb_\alpha N_g|\leq CC_1\vep t^{1/2}s^{-4 }. $$ \end{lemma} \begin{proof} We recall that $$ N_g - 2 = h^{00} - \sum_ah^{aa} - 2\hb^{00} - s\delb_s\hb^{00} $$ thus $$ |\delb_{\alpha}N_g|\leq 2|\delb_{\alpha}\hb^{00}| + \sum_a|\delb_{\alpha}h^{aa}| + |\delb_{\alpha}h^{00}| + |\delb_{\alpha}(s\delb_s\hb^{00})| $$ The first three terms are bounded by \eqref{eq 6 08-09-2017} and \eqref{eq 3 26-10-2017}. For the last term, we remark the following relation: \begin{equation}\label{eq 4 26-10-2017} \aligned \big|\delb_s\left(s\delb_s\hb^{00}\right)\big|\leq& C\big|\delb_s\hb^{00}\big| + C\big|\delb_s\delb_s\Qb^{000}\delb_su\big| + C\big|\delb_s\Qb^{000}\delb_s\delb_su\big| + C|\Qb^{000}\delb_s\delb_s\delb_s u| \\ &+C\big|\delb_s\delb_s\Qb^{00c}\delb_cu\big| + C\big|\delb_s\Qb^{00c}\delb_cu\big| + C\big|\Qb^{00c}\delb_s\delb_s\delb_c u\big| \endaligned \end{equation} and \begin{equation}\label{eq 5 26-10-2017} \aligned \big|\delb_a\left(s\delb_s\hb^{00}\right)\big| =& s|\delb_a\delb_s\hb^{00}|\leq C\big|\delb_a\delb_s\Qb^{000}\delb_su\big| + C\big|\delb_a\Qb^{000}\delb_s\delb_su\big| + C|\Qb^{000}\delb_a\delb_s\delb_s u| \\ &+C\big|\delb_a\delb_s\Qb^{00c}\delb_cu\big| + C\big|\delb_a\Qb^{00c}\delb_cu\big| + C\big|\Qb^{00c}\delb_a\delb_s\delb_c u\big| \endaligned. \end{equation} And we see that by lemma \ref{lem 2 02-09-2017} and proposition \ref{prop 1 02-09-2017}, we see that \begin{equation}\label{eq 6 26-10-2017} \aligned &|\delb_s\delb_s\Qb^{000}|\leq (s/t)|\delb_s\del_t\Qb^{000}| + s^{-1}|\delb_s\Qb^{000}|\leq CC_1\vep s^{-1}, \\ &|\delb_s\delb_s\Qb^{00c}|\leq (s/t)|\delb_s\del_t\Qb^{00c}| + s^{-1}|\delb_s\Qb^{00c}|\leq CC_1(t/s)s^{-1}. \endaligned \end{equation} Similar calculation shows that \begin{equation}\label{eq 7 26 10-2017} |\delb_a\delb_s\Qb^{000}|\leq CC_1\vep s^{-2},\quad |\delb_a\delb_s\Qb^{00c}|\leq CC_1\vep (t/s)s^{-2}. \end{equation} We also see that by \eqref{eq 12 07-09-2017}: \begin{equation}\label{eq 8 26-10-2017} |\delb_s\delb_s\delb_su|\leq (s/t)|\del_t\delb_s\delb_su|\leq CC_1\vep t^{-3/2}s^{-2 },\quad |\delb_a\delb_s\delb_su|\leq t^{-1}|L_a\delb_s\delb_su|\leq CC_1\vep t^{-3/2}s^{-3 }. \end{equation} and by \eqref{eq 7 07-09-2017}, \begin{equation}\label{eq 9 26-10-2017} |\delb_a\delb_s\delb_c u|\leq t^{-1}|L_a\delb_c\delb_su|\leq CC_1\vep t^{-5/2}s^{-2 }. \end{equation} Now we substitute the bounds \eqref{eq 6 08-09-2017} together with \eqref{eq 6 26-10-2017}, \eqref{eq 7 26 10-2017}, lemma \ref{lem 2 02-09-2017}, proposition \ref{prop 1 02-09-2017}, \eqref{eq 8 26-10-2017}, \eqref{eq 9 26-10-2017}, \eqref{eq 12 07-09-2017}, \eqref{eq 8 07-09-2017} and \eqref{eq 5 01-07-2017}, we see that \begin{equation}\label{eq 10 26-10-2017} \big|\delb_s\left(s\delb_s\hb^{00}\right)\big|\leq CC_1\vep t^{1/2}s^{-4 },\quad \big|\delb_a\left(s\delb_s\hb^{00}\right)\big|\leq CC_1\vep t^{-1/2}s^{-3 }. \end{equation} Now combine \eqref{eq 6 08-09-2017}, \eqref{eq 3 26-10-2017} and \eqref{eq 10 26-10-2017}, we see that the desired bound is established. \end{proof} Then we establish the following bound: \begin{lemma}\label{lem 1 08-09-2017} Under the bootstrap assumption with $\vep$ sufficiently small, \eqref{eq 3 23-05-2017} holds for a $\kappa>1$. \end{lemma} \begin{proof} This is by verifying proposition \ref{prop 1 24-05-2017}. We see that by \eqref{eq 5 08-09-2017} and \eqref{eq 6 08-09-2017}, \eqref{eq 1 24-05-2017} is verified with $\vep\leq \frac{\vep_s}{CC_1}$. \end{proof} \begin{lemma}\label{lem 2 08-09-2017} Under the bootstrap assumption, we have $$ M_g(s,\del^IL^Ju) = C(C_1\vep)^2 s^{-2} $$ where $M_g$ is defined as in \eqref{eq 4 23-05-2017}. \end{lemma} \begin{proof} Recall \eqref{eq R(du,du)}, we see that $$ \aligned \|R_g(\nabla \del^IL^Ju,\nabla \del^IL^Ju)\|_{L^1(\Hcal_s)} \leq & \|s^{-1}\left(L_g^{ab}-L_m^{ab}\right)\|_{L^{\infty}(\Hcal_s)}\|s^2\delb_a\del^IL^Ju\delb_b\del^IL^Ju\|_{L^1(\Hcal_s)} \\ &+ \|s^{-1}(N_g-N_m)\gb^{ab}\|_{L^{\infty}(\Hcal_s)}\|s^2\delb_a\del^IL^Ju\delb_b\del^IL^Ju\|_{L^1(\Hcal_s)} \\ &+ 2\|s^{-1}\hb^{ab}\|_{L^{\infty}(\Hcal_s)}\|s^2\delb_a\del^IL^Ju\delb_b\del^IL^Ju\|_{L^1(\Hcal_s)} \\ &+ \frac{1}{2}\|\delb_s\left(\gb^{00}\gb^{ab}\right)\|_{L^{\infty}(\Hcal_s)}\|s^2\delb_a\del^IL^Ju\delb_b\del^IL^Ju\|_{L^1(\Hcal_s)} \\ &+ \|(t/s)\delb_a\gb^{00}\gb^{ab}\|_{L^{\infty}(\Hcal_s)} \cdot \|s^2(s/t)\delb_s\del^IL^Ju\delb_b\del^IL^Ju\|_{L^1(\Hcal_s)} \\ \leq& CM_1\Ec(s,\del^IL^Ju) \endaligned $$ where $$ \aligned M_1: =& \|s^{-1}\left(L_g^{ab}-L_m^{ab}\right)\|_{L^{\infty}(\Hcal_s)} + \|s^{-1}(N_g-N_m)\gb^{ab}\|_{L^{\infty}(\Hcal_s)} + \|s^{-1}\hb^{ab}\|_{L^{\infty}(\Hcal_s)} \\ & + \|\delb_s\left(\gb^{00}\gb^{ab}\right)\|_{L^{\infty}(\Hcal_s)} + \|(t/s)\delb_a\gb^{00}\gb^{ab}\|_{L^{\infty}(\Hcal_s)}. \endaligned $$ We remark the following bounds (by \eqref{eq 6 08-09-2017}): $$ \aligned &\|L_g^{ab} - L_m^{ab}\|_{L^{\infty}(\Hcal_s)}\leq CC_1\vep s^{-2 },\quad \|N_g-N_m\|_{L^{\infty}(\Hcal_s)}\leq CC_1\vep s^{-2 }, \\ &\|\hb^{ab}\|_{L^{\infty}(\Hcal_s)}\leq CC_1\vep s^{-2 }, \\ &\|\delb_s(\gb^{00}\gb^{ab})\|_{L^{\infty}(\Hcal_s)}\leq CC_1\vep s^{-3 },\quad \|(t/s)\delb_a\gb^{00}\gb^{ab}\|_{L^{\infty}(\Hcal_s)} \leq CC_1\vep s^{-3 } \endaligned $$ then we see that $$ M_1\leq CC_1\vep s^{-3 }. $$ To analysis the term $(\mathscr{K}_g+N_g) \del^IL^Ju\cdot S_g[\nabla \del^IL^Ju]$, we see that $$ \aligned &\big\|(\mathscr{K}_g + N_g)\del^IL^Ju\cdot S_g[\nabla \del^IL^Ju]\big\|_{L^1(\Hcal_s)} \leq&C\|S_g[\nabla \del^IL^J u]\|_{L^2(\Hcal_s)}\cdot \Ec(s,\del^IL^Ju)^{1/2}. \endaligned $$ Then we see that by \eqref{eq 6 08-09-2017} and especially the bound on $\hb^{a0},\delb_s\hb^{a0}$ $$ \aligned &\|S_g[\nabla \del^IL^J u]\|_{L^2(\Hcal_s)} \\ =& \|2s^{-1}\delb_s(s\mb^{a0}) + 2s^{-1}\delb_s(s\hb^{a0}) + \delb_b\gb^{ab}\|_{L^{\infty}(\Hcal_s)}\|s\delb_a\del^IL^Ju\|_{L^2(\Hcal_s)} \\ =&\|2s^{-1}\delb_s(s\hb^{a0}) + \delb_b\gb^{ab}\|_{L^{\infty}(\Hcal_s)}\|s\delb_a\del^IL^Ju\|_{L^2(\Hcal_s)} \\ \leq& CC_1\vep s^{-2}\Ec(s,\del^IL^Ju)^{1/2} \endaligned $$ Then we see that $$ \big\|\left(\mathscr{K}_g+N_g\right) \del^IL^Ju\cdot S_g[\nabla \del^IL^Ju]\big\|_{L^1(\Hcal_s)}\leq CC_1\vep s^{-2}\Ec(s,\del^IL^Ju). $$ For the term $T_g[\del^IL^Ju]$, we see that $$ \aligned T_g[\del^IL^J u] =& - (t/s)\delb_sN_g\cdot(s/t)\del^IL^Ju\cdot(\mathscr{K}_g+N_g)\del^IL^Ju \\ &-(t/s)\gb^{ab}\delb_aN_g\cdot s\delb_b\del^IL^Ju\cdot(s/t)\del^IL^Ju \endaligned $$ thus we see that $$ \|T_g[\del^IL^J u]\|_{L^1(\Hcal_s)}\leq C\|(t/s)\delb_\alpha N_g\|_{L^{\infty}(\Hcal_s)}\cdot \Ec(s,\del^IL^Ju). $$ Then we apply lemma \ref{lem 1 26-10-2017} and \eqref{eq 6 08-09-2017}, we see that $$ \|(t/s)\delb_\alpha N_g\|_{L^{\infty}(\Hcal_s)}\leq CC_1\vep s^{-2}. $$ Thus we see that the bound on $M_g$ is bounded by $CC_1\vep s^{-2}\Ec(s,\del^IL^Ju)^{1/2}$. Then we apply the bootstrap bound \eqref{eq 5 24-08-2017} and the desired result is established. \end{proof} Now we are ready to establish the improved energy bound. \begin{proof}[Proof of proposition \ref{prop 1 01-07-2017}] This is by applying the energy estimate \eqref{eq energy} on the following equation: \begin{equation}\label{eq energy final} \Box \del^IL^J u + Q^{\alpha\beta\gamma}\del_\gamma u \del_{\alpha}\del_{\beta}\del^IL^J u = - [\del^IL^J,Q^{\alpha\beta\gamma}\del_{\gamma}u\del_{\alpha}\del_{\beta}]u. \end{equation} We see that for $s\in[2,s_1]$, \begin{equation}\label{eq 4 28-10-2017} \Ec(s,\del^IL^J u)^{1/2} \leq C\Ec(2,\del^IL^J u) + C\int_{2}^s \ \left(\tau\|[\del^IL^J,Q^{\alpha\beta\gamma}\del_{\gamma}u\del_{\alpha}\del_{\beta}]u\|_{L^2(\Hcal_\tau)} + M_g(\tau)\right)\ d\tau \end{equation} Recall that the initial energy is determined by the initial data and thus can be bounded by $CC_0\vep$. Then we substitute the bounds \eqref{eq 2 08-09-2017} and lemma \ref{lem 2 08-09-2017}, we see that $$ \aligned \Ec(s,\del^IL^J u)^{1/2} \leq& CC_0\vep + C(C_1\vep)^2\int_{2}^{s}\tau^{-2 }d\tau \leq CC_0\vep + C(C_1\vep)^2. \endaligned $$ Then we see that we chose $C_1>2CC_0$ and $\vep_0\leq \frac{C_1-2CC_0}{CC_1^2}$. With this choice, we obtain that $$ \Ec(s,\del^IL^J u)^{1/2}\leq \frac{1}{2}C_1\vep $$ and this concludes the desired result. \end{proof}

https://arxiv.org/abs/1701.05798

Factorizable Module Algebras

The aim of this paper is to introduce and study a large class of $\mathfrak{g}$-module algebras which we call factorizable by generalizing the Gauss factorization of (square or rectangular) matrices. This class includes coordinate algebras of corresponding reductive groups $G$, their parabolic subgroups, basic affine spaces and many others. It turns out that tensor products of factorizable algebras are also factorizable and it is easy to create a factorizable algebra out of virtually any $\mathfrak{g}$-module algebra. We also have quantum versions of all these constructions in the category of $U_q(\mathfrak{g})$-module algebras. Quite surprisingly, our quantum factorizable algebras are naturally acted on by the quantized enveloping algebra $U_q(\mathfrak{g}^*)$ of the dual Lie bialgebra $\mathfrak{g}^*$ of $\mathfrak{g}$.

\section{Introduction} \label{sect:Introduction} The aim of this paper is to introduce and study a class of $\mathfrak{g}$-module algebras which we call \emph{factorizable} by generalizing the Gauss factorization of square or rectangular matrices. More precisely, let $\mathfrak{g}$ be a complex semisimple Lie algebra. A commutative $\mathbb{C}$-algebra $A$ which is also a $\mathfrak{g}$-module is called a $\mathfrak{g}$-module algebra if $\mathfrak{g}$ acts on $A$ by derivations. Given such a commutative $\mathfrak{g}$-module algebra $A$, we say that $A$ is \emph{factorizable} over a $\mathfrak{g}$-equivariant subalgebra $A_0$ if the restriction of the multiplication map of $A$, $\mu:A^+\otimes A_0\to A$, is an isomorphism of vector spaces. Here $A^+$ stands for the subalgebra of highest weight vectors in $A$, i.e., the kernel of the action of the maximal nilpotent subalgebra $\mathfrak{n}_+\subset \mathfrak{g}$. See Section \ref{sect:classical} for the precise definitions. Factorizable algebras abound in ``nature" with $A_0=\mathbb{C}[U]$, where $U$ is the maximal unipotent subgroup of the corresponding Lie group $G$, which is a $\mathfrak{g}$-module algebra isomorphic to the graded dual of the Verma module $M_0$, as a $\mathfrak{g}$-module. Our first result shows that the natural objects in the representation theory of $\mathfrak{g}$ are factorizable. \begin{theorem} \label{thm:G/U} Up to a localization, the algebra $A=\mathbb{C}[G/U]^{\otimes n}$ is factorizable over $A_0=\mathbb{C}[U]$ for any $n\ge 1$. \end{theorem} \begin{remark} \label{rem:classical Gauss} If we replace $\mathbb{C}[G/U]$ with $A_1=\mathbb{C}[x_1,\ldots,x_m]$ and let $\mathfrak{g}=\mathfrak{sl}_m(\mathbb{C})$, then after localization by leading principal minors, $A_1^{\otimes n}=\mathbb{C}[Mat_{m,n}]$ ($n\ge m-1$) is factorizable over $\mathbb{C}[U_m^-]$ where $U_m^-$ is the group of all lower unitriangular matrices in $SL_m(\mathbb{C})$. This recovers the Gauss factorization of $m\times n$ matrices. See Example \ref{ex:mat} for details in a particular case. \end{remark} \begin{remark} Our factorizations are rather different from the well-known ones such as $U(\mathfrak{g})=U(\mathfrak{g}_-)\otimes U(\mathfrak{g}_+)$ prescribed by the Poincar\'e-Birkhoff-Witt theorem for any decomposition of $\mathfrak{g}$ into a direct sum of Lie subalgebras $\mathfrak{g}_-$ and $\mathfrak{g}_+$ or the Kostant harmonic decompositions of $S(\mathfrak{g})$ into the $\mathfrak{g}$-invariants and $\mathfrak{g}$-harmonic elements. \end{remark} It turns out that factorizability of module algebras is not difficult to establish and reproduce. \begin{theorem} \label{thm:U+} For any complex semisimple Lie algebra $\mathfrak{g}$, we have: (a) (Theorems \ref{thm:b-} \& \ref{thm:g}) Let $A$ be a $\mathfrak{g}$-module algebra containing a $\mathfrak{g}$-module subalgebra isomorphic to $\mathbb{C}[U]$ and let $A^+$ denote the subalgebra of all highest weight vectors in $A$. Then the algebra $A'=A^+\otimes \mathbb{C}[U]$ has a natural $\mathfrak{g}$-module algebra structure such that $A'$ is factorizable over $\mathbb{C}[U]$. (b) The assignments $A\mapsto A'$ define a functor $R$ from the category $\mathcal{A}_\mathfrak{g}$ of $\mathfrak{g}$-module algebras containing a $\mathfrak{g}$-module subalgebra isomorphic to $\mathbb{C}[U]$ to the category ${\mathcal C}_\mathfrak{g}$ of $\mathfrak{g}$-module algebras which are factorizable over $\mathbb{C}[U]$, a full subcategory of $\mathcal{A}_\mathfrak{g}$. \end{theorem} We can think of $R$ as a ``remembering" functor because it is right adjoint to the ``forgetful" functor $F:{\mathcal C}_\mathfrak{g}\to \mathcal{A}_\mathfrak{g}$. Clearly, the composition $R\circ F$ is isomorphic to the identity functor on $\mathcal{C}_\mathfrak{g}$. Now $\mathcal{A}_\mathfrak{g}$ has a natural tensor multiplication such that for $A,B\in \mathcal{A}_\mathfrak{g}$, the tensor product $A\otimes B$ is a $\mathfrak{g}$-module algebra, plus the embedding $\mathbb{C}[U]=1\otimes \mathbb{C}[U]\subset A\otimes B$. \begin{proposition} (Proposition \ref{prop:tensor}) $\mathcal{C}_\mathfrak{g}$ is closed under the natural tensor multiplication on $\mathcal{A}_\mathfrak{g}$. \end{proposition} However, neither category is monoidal because each lacks a unit object. It is possible to fix the issue in both categories by tensoring over $\mathbb{C}[U]$ rather than over $\mathbb{C}$. See the discussion in Remark \ref{rmk:unitissue}, where we make this more concrete. It turns out that we can build factorizable algebras over $\mathbb{C}[U]$ out of $\mathfrak{b}_-$-module algebras, where $\mathfrak{b}_-$ is the lower Borel subalgebra of $\mathfrak{g}$. In fact, all factorizable algebras can be obtained this way. \begin{maintheorem}(Theorems \ref{thm:b-} \& \ref{thm:equiv}) \label{thm:construction} For any semisimple Lie algebra $\mathfrak{g}$, the assignments $A\mapsto A^+$ defines a functor $P$ from $\mathcal{A}_\mathfrak{g}$ to the category $\mathfrak{b}_- -{\bf ModAlg}$ of $\mathfrak{b}_-$-module algebras. Moreover, the composition $P\circ F$ is an equivalence of categories ${\mathcal C}_\mathfrak{g}\widetilde \to \mathfrak{b}_- -{\bf ModAlg}$. \end{maintheorem} \begin{remark} \label{rem:forgetmenot} Informally speaking, the theorem asserts that the ``forgetful" highest weight vector functor $A\mapsto A^+$ from $\mathcal{A}_\mathfrak{g}$ to $Alg_\mathbb{C}$, in fact, ``remembers almost everything." \end{remark} The functor $P$ from Theorem \ref{thm:construction} is highly nontrivial: it involves a quite mysterious $\mathfrak{b}_-$ action on $A$ such that $A^+$ is a $\mathfrak{b}_-$-invariant subalgebra (see Section \ref{sect:classical} for details). Namely, the action of the Cartan subalgebra of $\mathfrak{b}_-$ is inherited from that of $\mathfrak{g}$, but the action of the Chevalley generators $f_i$ of $\mathfrak{n}_-\subset \mathfrak{b}_-$ is given by the formula \begin{equation} \label{eq:fihat} f_i\triangleright a = f_i(a)-h_i(a)x_i \end{equation} for $a\in A$, where $h_i$ is the $i$-th Cartan generator and $x_i$ is the $i$-th ``near-diagonal element" in $\mathbb{C}[U]$ ($\subset A$). It is not difficult to show that $f_i\triangleright A^+\subset A^+$, but it is much harder to prove that the operators $f_i\triangleright$ satisfy the Serre relations (Theorem \ref{thm:b-}). It turns out that all of the above results, including the mysterious Serre relations, can be quantized as well. Namely, we replace $\mathfrak{g}$ with its quantized enveloping algebra $U_q(\mathfrak{g})$ and proceed as follows. Given a $U_q(\mathfrak{g})$-module algebra $A_0$, we say that a $U_q(\mathfrak{g})$-module algebra $A$ is \emph{factorizable} over a $U_q(\mathfrak{g})$-equivariant subalgbra $A_0$ if the restriction of the multiplication map of $A$, $\mu:A^+\otimes A_0\to A$, is an isomorphism of vector spaces. By quantizing our default choice of $A_0$ in the classical case, we will now focus on $A_0=\mathbb{C}_q[U]$, the quantized coordinate algebra of $U$, which is isomorphic to $U_q(\mathfrak{n}_+)$. See Section \ref{sect:quantum} for the precise definitions. \begin{theorem} \label{thm:qG/U} Up to a localization, the braided $n$-fold tensor power of $\mathbb{C}_q[G/U]$, $A=\mathbb{C}_q[G/U]^{\underline{\otimes} n}$, is factorizable over $A_0=\mathbb{C}_q[U]$ for any $n\ge 1$. \end{theorem} \begin{remark} Similarly to the classical case (Remark \ref{rem:classical Gauss}), if we replace $\mathbb{C}_q[G/U]$ with $A_1=\mathbb{C}_q[x_1,\ldots,x_m]$, the algebra of $q$-polynomials and let $\mathfrak{g}=\mathfrak{sl}_m(\mathbb{C})$, then the braided $n$-fold tensor power $A_1^{\underline{\otimes} n}=\mathbb{C}_q[Mat_{m,n}]$ ($n\ge m-1$) is factorizable over $\mathbb{C}_q[U_m]$, after localization by leading principal quantum minors. This recovers the Gauss factorization of quantum $m\times n$ matrices (see, e.g., \cite{dks}). See Example \ref{ex:qmat} for details in a particular case. \end{remark} As in the classical case, factorizability of module algebras in the quantum case is also easy to establish and reproduce. \begin{theorem} \label{thm:Uq+} For any complex semisimple Lie algebra $\mathfrak{g}$, we have: (a) (Theorems \ref{thm:qgd} \& \ref{thm:qg}) Let $A$ be a $U_q(\mathfrak{g})$-module algebra containing a $U_q(\mathfrak{g})$-module subalgebra isomorphic to $\mathbb{C}_q[U]$ and let $A^+$ denote the subalgebra of all highest weight vectors in $A$. Then the vector space $A'=A^+\otimes \mathbb{C}_q[U]$ has the structure of a $U_q(\mathfrak{g})$-module algebra and is factorizable over $\mathbb{C}_q[U]$. (b) The assignments $A\mapsto A'$ define a functor $R_q$ from the category ${\mathcal{A}}_{\mathfrak{g}}^q$ of $U_q(\mathfrak{g})$-module algebras containing a $U_q(\mathfrak{g})$-module subalgebra isomorphic to $\mathbb{C}_q[U]$ to the category ${\mathcal{C}}_{\mathfrak{g}}^q$ of $U_q(\mathfrak{g})$-module algebras which are factorizable over $\mathbb{C}_q[U]$, a full subcategory of ${\mathcal{A}}_{\mathfrak{g}}^q$. \end{theorem} We can think of $R_q$ as a ``remembering" functor because it is right adjoint to the ``forgetful" functor $F_q:{\mathcal{C}}_{\mathfrak{g}}^q\to {\mathcal{A}}_{\mathfrak{g}}^q$. Clearly, the composition $R_q\circ F_q$ is the identity functor on $\mathcal{C}_\mathfrak{g}^q$. In order to tensor multiply objects of these categories, we need to ``trim" it a bit. Namely, we consider the full subcategory $\underline{\mathcal{A}}_\mathfrak{g}^q$ consisting of \emph{weight} module algebras in $\mathcal{A}_\mathfrak{g}^q$ satisfying some additional mild conditions (see Section \ref{sect:quantum}). It turns out that $\underline{\mathcal{A}}_\mathfrak{g}^q$ has a natural \emph{braided} tensor product which we denote by $\underline{\otimes}$ (see, e.g. \cite{majid} and Section \ref{sect:quantum} below). Similar to the classical case, for $A, B\in \underline{\mathcal{A}}_{\mathfrak{g}}^{q}$, $A\underline{\otimes}B$ is naturally in $\underline{\mathcal{A}}_\mathfrak{g}^q$ with an embedding $\mathbb{C}_q[U]=1\underline{\otimes} \mathbb{C}_q[U]\subset A\underline{\otimes}B$. As in the classical case, this natural multiplication of course lacks a unit object. Since $\mathcal{C}_\mathfrak{g}^q$ is a full subcategory of $\mathcal{A}_\mathfrak{g}^q$, we can define $\underline{\mathcal{C}}_\mathfrak{g}^q$ as the intersection of $\mathcal{C}_\mathfrak{g}^q$ and $\underline{\mathcal{A}}_\mathfrak{g}^q$. \begin{proposition} (Proposition \ref{prop:qtensor}) $\underline{\mathcal{C}}_\mathfrak{g}^q$ is closed under the braided tensor multiplication in $\underline{\mathcal{A}}_{\mathfrak{g}}^{q}$. \end{proposition} That is, the category $\underline{\mathcal{C}}_\mathfrak{g}^q$ of factorizable $U_q(\mathfrak{g})$-weight module algebras is ``almost" monoidal but it lacks a unit object as in the classical case. Similar to the classical case (Theorem \ref{thm:construction}), we can build factorizable module algebras over $\mathbb{C}_q[U]$ out of some module algebras. However, unlike the expected $U_q(\mathfrak{b}_-)$-module algebras, we will deal with $U_q(\mathfrak{g}^*)$-module algebras, where $\mathfrak{g}^*$ is the dual Lie bialgebra of $\mathfrak{g}$ and all factorizable algebras are obtained this way. \begin{maintheorem} \label{thm:constructionq} (Theorems \ref{thm:qgd} \& \ref{thm:qequiv}) For any semisimple Lie algebra $\mathfrak{g}$, the assignments $A\mapsto A^+$ defines a functor $P_q$ from $\mathcal{A}_{\mathfrak{g}}^q$ to the category $U_q(\mathfrak{g}^*)-{\bf ModAlg}$ of $U_q(\mathfrak{g}^*)$-module algebras. Moreover, the composition $P_q\circ F_q$ is an equivalence of categories $\mathcal{C}_{\mathfrak{g}}^q\widetilde{\to}U_q(\mathfrak{g}^*)-{\bf ModAlg}$. \end{maintheorem} \begin{remark} \label{rem:forgetmenotq} In the spirit of Remark \ref{rem:forgetmenot}, the theorem asserts that the assignment $A\mapsto A^+$ is the forgetful functor which ``remembers almost everything." \end{remark} \begin{remark} We firmly believe that the emergence of the Lie bialgebra $\mathfrak{g}^*$ here is not a mere coincidence, but rather a manifestation of the ``semiclassical story" behind the quantum one. We plan to investigate it in a separate publication, when all relevant objects are Poisson algebras with a compatible action of the Poisson-Lie group $G$, where its Poisson-Lie dual $G^*$ emerges naturally. \end{remark} The functor $P_q$ from Theorem \ref{thm:constructionq} is highly nontrivial: it involves a quite mysterious $U_q(\mathfrak{g}^*)$ action on $A$ such that $A^+$ is a $U_q(\mathfrak{g}^*)$-invariant subalgebra (see Section \ref{sect:quantum} for details). Namely, the Cartan subalgebra action of $U_q(\mathfrak{g}^*)$ is inherited from that of $U_q(\mathfrak{g})$, but the action of the generators $F_{i,1}$ and $F_{i,2}$ of $U_q(\mathfrak{g}^*)$ is given by the formulas \begin{equation} \label{eq:fihatq} F_{i,1}\triangleright a = F_i(a)-\frac{x_ia-K_i^{-1}(a)x_i}{q_i-q_i^{-1}}, \quad F_{i,2}\triangleright a = \frac{x_iK_i^{-1}(a)-ax_i}{q_i-q_i^{-1}} \end{equation} for $a\in A^+$, where $K_i$ is the $i$-th Cartan generator of $U_q(\mathfrak{g})$, $q_i=q^{d_i}$, and $x_i$ is the $i$-th generator of $\mathbb{C}_q[U]\subset A$. As in the classical case, the most non-trivial part of the proof of the main theorem is to show that operators $F_{i,1}\rhd$ and $F_{i,2}\rhd$ given by \eqref{eq:fihatq} satisfy the quantum Serre relations, which we establish in Theorem \ref{thm:qgd}. Also, while not difficult, it is still surprising that in \eqref{eq:fihatq}, all operators $F_{i,1}\rhd$ commute with all $F_{i,2}\rhd$. \section{Definitions, Notation, and Results} \label{sect:Definitions, Notation, and Results} In this section, we will recall and introduce the relevant definitions and notation necessary to present our main results, which will also be included. We begin by defining the main object of study in this paper: module algebras. \begin{definition} Let $\Bbbk$ be an arbitrary field and $H$ a $\Bbbk$-bialgebra. A $\Bbbk$-algebra $A$ is called an \emph{$H$-module algebra} if it is an $H$-module, multiplication $(-)\cdot(-):A\otimes_\Bbbk A\to A$ is a homomorphism of $H$-modules, and $h(1)=\epsilon(h)$ for $h\in H$, where $\epsilon$ is the counit of $H$. That is to say, if we denote in sumless Sweedler notation $\Delta(h)=h_{(1)}\otimes h_{(2)}$ for $h\in H$, then for $a,b\in A$, $h(a\cdot b)=h_{(1)}(a)\cdot h_{(2)}(b)$ and $h(1)=\epsilon(h)$. A homomorphism of $H$-module algebras is a homomorphism of $H$-modules and algebras. We will denote the category of $H$-module algebras by $H$-{\bf ModAlg}. If $\mathfrak{g}$ is a Lie algebra over $\Bbbk$, then we will shorten ``$U(\mathfrak{g})$-module algebra" and ``$U(\mathfrak{g})-{\bf ModAlg}$" to ``$\mathfrak{g}$-module algebra" and ``$\mathfrak{g}$-{\bf ModAlg}", respectively. \end{definition} Let $\mathfrak{g}$ be a semisimple complex Lie algebra, with $I\times I$ symmetrizable Cartan matrix $C=(c_{i,j})$ and fixed choice of symmetrizers, $(d_i)_{i\in I}$, for $C$, i.e. $d_ic_{i,j}=d_jc_{j,i}$ for $i,j\in I$. Then $\mathfrak{g}$ has a triangular decomposition $\mathfrak{g}=\mathfrak{n}_-\oplus\mathfrak{h}\oplus\mathfrak{n}_+$. Here $\mathfrak{h}$ is a Cartan subalgebra with $\dim \mathfrak{h}=|I|$ and its dual $\mathfrak{h}^*$ has basis $\{\alpha_i\ |\ i\in I\}$, the simple roots of the associated root system. Let $(\cdot,\cdot)$ be the symmetric bilinear form on $\mathfrak{h}^*$ satisfying $(\alpha_i,\alpha_j)=d_ic_{i,j}$. As usual, we set $\alpha_i^\vee:=\frac{2\alpha_i}{(\alpha_i,\alpha_i)}$. By $\Lambda=\{\lambda\in \mathfrak{h}^*\ |\ (\alpha_i^\vee,\lambda)\in \mathbb{Z}\ \forall i\in I\}$, we denote the set of integral weights. Denote by $\omega_i\in \Lambda$ is the $i$-th fundamental weight, which satisfies $(\alpha_i^\vee,\omega_j)=\delta_{i,j}$ for $i,j\in I$. Let $W$ be the Weyl group of $\mathfrak{g}$, generated by the simple reflections $\{s_i\ |\ i\in I\}$, and with longest element $w_\mathrm{o}$. Given $w\in W$, $R(w)$ is the set of all ${\bf i}=(i_1,i_2,\ldots,i_m)\in I^m$ such that $s_{i_1}s_{i_2}\cdots s_{i_m}$ is a reduced expression for $w$. \subsection{Quantum Factorization} \label{sect:quantum} Throughout this section, all tensor products will be taken over $\mathbb{C}(q)$ unless otherwise specified and written $-\otimes-$ rather than $-\otimes_{\mathbb{C}(q)}-$. Recall that $U_q(\mathfrak{g})$ is the quantum enveloping algebra of $\mathfrak{g}$, generated by elements $\{K_i^{\pm 1},E_i,F_i\ |\ i\in I\}$ subject to the relations \begin{center} $\begin{array}{c c} K_iK_j=K_jK_i; &K_iK_i^{-1}=K_i^{-1}K_i=1;\\\\ K_iE_jK_i^{-1}=q_i^{c_{i,j}}E_j;& K_iF_jK_i^{-1}=q_i^{-c_{i,j}}F_j; \end{array}$ \[E_iF_j-F_jE_i=\delta_{i,j}\frac{K_i-K_i^{-1}}{q_i-q_i^{-1}};\] \[\sum_{k=0}^{1-c_{i,j}}(-1)^k E_i^{(k)}E_jE_i^{(1-c_{i,j}-k)}=0\text{ if }i\ne j;\] \[\sum_{k=0}^{1-c_{i,j}}(-1)^k F_i^{(k)}F_jF_i^{(1-c_{i,j}-k)}=0\text{ if }i\ne j;\] \end{center} where $y_i^{(n)}=\frac{1}{(n)_{q_i}!}y_i^n$, $(n)_{q_i}!=(1)_{q_i}(2)_{q_i}\cdots (n)_{q_i}$, $(m)_{q_i}=\dfrac{q_i^m-q_i^{-m}}{q_i-q_i^{-1}}$, and $q_i=q^{d_i}$. $U_q(\mathfrak{g})$ is a Hopf algebra with comultiplication $\Delta$, counit $\epsilon$, and antipode $S$ given on generators by \begin{center} $\begin{array}{c c c} \Delta(K_i^{\pm1})=K_i^{\pm1}\otimes K_i^{\pm1}; & \epsilon(K_i^{\pm1})=1 & S(K_i^{\pm1})=K_i^{\mp1};\\\\ \Delta(E_i)=E_i\otimes K_i+1\otimes E_i; & \epsilon(E_i)=0; & S(E_i)=-E_iK_i^{-1};\\\\ \Delta(F_i)=F_i\otimes 1+K_i^{-1}\otimes F_i; & \epsilon(F_i)=0; & S(F_i)=-K_iF_i.\\\\ \end{array}$ \end{center} Now, $\mathcal{K}$ is the Hopf subalgebra of $U_q(\mathfrak{g})$ generated by all $K_i$ as a $\mathbb{C}(q)$-algebra, while $U_q(\mathfrak{b}_+)$ (respectively $U_q(\mathfrak{b}_-)$) is the Hopf subalgebra of $U_q(\mathfrak{g})$ generated by all $K_i^{\pm1}$ and $E_i$ (respectively $K_i^{\pm 1}$ and $F_i$). We will assume henceforth that for any $i\in I$, the action of $E_i$ on any $U_q(\mathfrak{b}_+)$-modules is \emph{locally nilpotent}. In other words, if $M$ is a $U_q(\mathfrak{b}_+)$-module, we will assume that for each $x\in M$ and $i\in I$, there exists some $n\ge 0$ such that $E_i^n(x)=0$. Note that every $U_q(\mathfrak{g})$-module is a $U_q(\mathfrak{b}_+)$-module, so we are assuming these are ``bounded above". We do not assume the same for the action of $F_i$. For a $U_q(\mathfrak{b}_+)$-module $M$, we designate $M^+:=\{m\in M\ |\ E_i(m)=0\ \forall i\in I\}$, the set of \emph{highest weight vectors}. If $A$ is a $U_q(\mathfrak{b}_+)$-module algebra, then $A^+$ is a $U_q(\mathfrak{b}_+)$-module subalgebra. Suppose $M$ is a $U_q(\mathfrak{b}_+)$-module. For each $i\in I$ and $x\in M\setminus\{0\}$, set $\ell_i(x)=\max\{\ell\in\mathbb{Z}_{\ge0}\ |\ E_i^\ell(x)\ne 0\}$ and $E_i^{(top)}(x)=E_i^{(\ell_i(x))}(x)$. Given ${\bf i}\in I^m$ for some $m\ge 0$ and $x\in M\setminus\{0\}$, we also use the shorthand \[E_{\bf i}^{(top)}(x)=E_{i_m}^{(top)}E_{i_{m-1}}^{(top)}\cdots E_{i_1}^{(top)}(x)\] and define $\nu_{\bf i}:M\setminus \{0\}\to \mathbb{Z}_{\ge 0}^m$, $x\mapsto(a_1,a_2,\ldots,a_m)$ by the following: \[a_k=\ell_{i_k}(E_{i_{k-1}}^{(top)}E_{i_{k-2}}^{(top)}\cdots E_{i_1}^{(top)}(x)).\] Lastly, for ${\bf j}=(j_1,\ldots,j_m)\in \mathbb{Z}_{\ge 0}^m$, we set $E_{\bf i}^{(\bf j)}:=E_{i_m}^{(j_m)}E_{i_{m-1}}^{(j_{m-1})}\cdots E_{i_1}^{(j_1)}.$ \begin{definition} Let $A$ be a $U_q(\mathfrak{b}_+)$-module algebra, $w\in W$, and ${\bf i}\in R(w)$. If $E_{\bf i}^{ (top)}(x)\in A^+$ for all $x\in A\setminus\{0\}$, then we say $A$ is \emph{{\bf i}-adapted}. We say a basis $\mathcal{B}$ for $A$ is an \emph{{\bf i}-adapted basis} if \begin{enumerate} \item $E_{\bf i}^{(top)}(b)=1$ for all $b\in \mathcal{B}$. \item The restriction of $\nu_{\bf i}$ to $\mathcal{B}$ is an injective map $\mathcal{B}\hookrightarrow \mathbb{Z}_{\ge 0}^m$, where $m$ is the length of $w$. \end{enumerate} If there exists any $w\in W$ and ${\bf i}\in R(w)$ so that $A$ is {\bf i}-adapted, then we say more generally that $A$ is \emph{adapted}. \end{definition} \begin{remark} Our notion of an {\bf i}-adapted $U_q(\mathfrak{b}_+)$-module algebra is different than P. Caldero's notion of adapted algebra in \cite{caldero}, though they do have some examples in common. On the other hand, our notion of {\bf i}-adapted basis is stronger than the similar notion of an adapted basis for $(A,\nu_{\bf i})$ as in \cite{kavehmanon}. \end{remark} It turns out that if $A_0$ possesses an ${\bf i}$-adapted basis for some ${\bf i}\in R(w)$ and is a ``large enough" $U_q(\mathfrak{b}_+)$-module subalgebra of $A$, then $A$ is factorizable over $A_0$. The following theorem makes this precise. \begin{theorem} \label{thm:qfactoring} Let $A$ be a $U_q(\mathfrak{b}_+)$-module algebra. Suppose $A_0$ is a $U_q(\mathfrak{b}_+)$-module subalgebra of $A$ possessing an {\bf i}-adapted basis $\mathcal{B}$ for some reduced $\mathbf{i}$. Then: \begin{enumerate} \item The restriction $\mu:A^+\otimes A_0\to A$ of the multiplication in $A$ is an injective homomorphism of $U_q(\mathfrak{b}_+)$-modules. \item The map $\mu$ is an isomorphism if and only if $A$ is {\bf i}-adapted and $\nu_{\bf i}(A\setminus\{0\})=\nu_{\bf i}(A_0\setminus\{0\})$. \end{enumerate} \end{theorem} We will prove Theorem \ref{thm:qfactoring} in Section \ref{pf:qfactoring}. Theorem \ref{thm:qfactoring} demonstrates a close relationship between being {\bf i}-adapted and being factorizable over a $U_q(\mathfrak{b}_+)$-module subalgebra possessing an {\bf i}-adapted basis. The following theorem explores this relationship from a different angle. \begin{theorem} \label{thm:qclassify} Let $A$ be an {\bf i}-adapted $U_q(\mathfrak{b}_+)$-module algebra for some reduced $\mathbf{i}$ and suppose $A_0$ is a $U_q(\mathfrak{b}_+)$-module subalgebra of $A$. Then $\mu:A^+\otimes A_0\to A$ as in Theorem \ref{thm:qfactoring} is an isomorphism of $U_q(\mathfrak{b}_+)$-modules if and only if $A_0$ possesses an {\bf i}-adapted basis and $\nu_{\bf i}(A_0\setminus\{0\})=\nu_{\bf i}(A\setminus\{0\})$. \end{theorem} Theorem \ref{thm:qclassify} is proved in Section \ref{pf:qclassify}. We now restrict our focus to a specific $U_q(\mathfrak{g})$-module algebra, namely $\mathbb{C}_q[U]$. As a $\mathbb{C}(q)$-algebra, $\mathbb{C}_q[U]$ is generated by the set $\{x_i\ |\ i\in I\}$, subject to the quantum Serre relations: \[\sum_{k=0}^{1-c_{i,j}}(-1)^k x_i^{(k)}x_jx_i^{(1-c_{i,j}-k)}=0\text{ if }i\ne j.\] The $U_q(\mathfrak{g})$-module structure on $\mathbb{C}_q[U]$ is summarized in the following equations: \begin{align*} K_i^{\pm1}(x_j)&=q_i^{\mp c_{i,j}}x_j \quad \text{for all } i,j\in I\\ E_i(x_j)&=\delta_{i,j}\quad \text{for all } i,j\in I\\ F_i(x)&=\frac{x_ix-K_i^{-1}(x)x_i}{q_i-q_i^{-1}}\quad \text{for all } i\in I \text{ and } x\in \mathbb{C}_q[U]. \end{align*} Of course, the actions of $K_i^{\pm1}$ and $E_i$ must be extended to all of $\mathbb{C}_q[U]$ by the rules \begin{align*} K_i^{\pm1}(xx')&=K_i^{\pm1}(x)K_i^{\pm1}(x')\quad \text{for all } i\in I \text{ and } x,x'\in \mathbb{C}_q[U]\\ E_i(xx')&=E_i(x)K_i(x')+xE_i(x')\quad \text{for all } i\in I,\text{ and } x,x'\in \mathbb{C}_q[U]. \end{align*} The first author and A. Zelevinsky observed in (\cite{berensteinzelevinsky}, Proposition 3.5) that $\mathbb{C}_q[U]$ possesses a basis $\mathcal{B}^{dual}$ such that, for ${\bf i}\in R(w_\mathrm{o})$, the restriction of $\nu_{\bf i}$ to $\mathcal{B}^{dual}$ is injective. Note that they use the notation $\mathcal{A}$ in place of $\mathbb{C}_q[U]$ and view it only as a $U_q(\mathfrak{n}_+)$-module. As hinted by the notation, $\mathcal{B}^{dual}$ is the so-called \emph{dual canonical basis}. In Section \ref{pf:adapted}, we prove the following proposition. \begin{proposition} \label{prop:adapted} Given any ${\bf i}\in R(w_\mathrm{o})$, $\mathcal{B}^{dual}$ is an ${\bf i}$-adapted basis for $\mathbb{C}_q[U]$. \end{proposition} \begin{remark} Based on the recent paper \cite{kimuraoya}, we expect that the dual canonical basis $\mathcal{B}^{dual}\cap U_q(w)$ in each quantum Schubert cell $U_q(w)$ is $\mathbf{i}$-adapted for any reduced word $\mathbf{i}$ for $w$. \end{remark} Combining Proposition \ref{prop:adapted} with Theorem \ref{thm:qfactoring}, we are led to the following corollary, though it does still require some proof. \begin{corollary} \label{cor:qhighest} Let $A$ be a $U_q(\mathfrak{g})$-module algebra containing $\mathbb{C}_q[U]$ as a $U_q(\mathfrak{g})$-module subalgebra. If there exists a $U_q(\mathfrak{g})$-module algebra $A'$ containing $A$ as a $U_q(\mathfrak{g})$-module subalgebra, such that $A'$ is generated by $(A')^+$ as a $U_q(\mathfrak{g})$-module algebra, then $\mu:A^+\otimes \mathbb{C}_q[U]\to A$ as in Theorem \ref{thm:qfactoring} is an isomorphism of $U_q(\mathfrak{b}_+)$-modules. \end{corollary} Corollary \ref{cor:qhighest} will be proved in Section \ref{pf:qhighest} and provides us with the means to prove Theorem \ref{thm:qG/U}, which we do in Section \ref{pf:qG/U}. In the meantime, there are two families of quantities that arose in the proof of Corollary \ref{cor:qhighest}: \[x_ia-K_i(a)x_i\quad\text{and}\quad F_i(a)+x_i\frac{K_i^{-2}(a)-a}{q_i-q_i^{-1}}\] where $i\in I$ and $a\in A'$. These quantities are equally valid to consider for $a\in A$, without the assumed presence of $A'$. If $a\in A^+$, then both of these quantities are also in $A^+$. It is therefore natural to ask what relations the families of operators $L_i-R_iK_i$ and $F_i+L_i\frac{K_i^{-2}-1}{q_i-q_i^{-1}}$ satisfy, where $L_i$ (respectively $R_i$) represents left (respectively right) multiplication by $x_i$. Or to put it another way, do these operators indicate the action of a known algebra which is somehow related to $U_q(\mathfrak{g})$? We can answer in the affirmative. It can be proved that both of the families of operators observed in fact satisfy the quantum Serre relations and the two families ``almost" commute with each other. This resembles an action of the Hopf algebra $U_q(\mathfrak{g}^*)$, which we now define for the reader's convenience. \begin{definition} As an algebra, $U_q(\mathfrak{g}^*)$ is generated by $\{K_i^{\pm1}, F_{i,1}, F_{i,2}\ |\ i\in I\}$ subject to the following relations for $i,j\in I$ and $k\in \{1,2\}$: \begin{center} $\begin{array}{c c} K_i^{\pm1}K_j^{\pm1}=K_j^{\pm1}K_i^{\pm1}; & K_iK_i^{-1}=K_i^{-1}K_i=1;\\[.2cm] F_{i,1}F_{j,2}=F_{j,2}F_{i,1}; & K_iF_{j,k}K_i^{-1}=q_i^{-c_{i,j}}F_{j,k}; \end{array}$ \[\sum_{\ell=0}^{1-c_{i,j}}(-1)^kF_{i,k}^{(\ell)}F_{j,k}F_{i,k}^{(1-c_{i,j}-\ell)}=0 \text{ if }i\ne j.\] \end{center} The comultiplication $\Delta$, counit $\epsilon$, and antipode $S$ are given on generators by \begin{center} $\begin{array}{c c c} \Delta(K_i^{\pm1})=K_i^{\pm1}\otimes K_i^{\pm1}; & \epsilon(K_i^{\pm1})=1; & S(K_i^{\pm1})=K_i^{\mp1};\\\\ \Delta(F_{i,1})=F_{i,1}\otimes 1+K_i^{-1}\otimes F_{i,1}; & \epsilon(F_{i,1})=0; & S(F_{i,1})=-K_iF_{i,1};\\\\ \Delta(F_{i,2})=F_{i,2}\otimes K_i^{-1}+1\otimes F_{i,2}; & \epsilon(F_{i,2})=0; & S(F_{i,2})=-F_{i,2}K_i. \end{array}$ \end{center} \end{definition} After some tweaking and combining of our operators with the inherited Cartan action, we see that our operators really do indicate the presence of a $U_q(\mathfrak{g}^*)$-module algebra structure. The following theorem summarizes this and is proved in Section \ref{pf:qgd,qg}. \begin{theorem} \label{thm:qgd} Let $A$ be a $U_q(\mathfrak{g})$-module algebra containing $\mathbb{C}_q[U]$ as a $U_q(\mathfrak{g})$-module subalgebra. Then $A$ is a $U_q(\mathfrak{g}^*)$-module algebra with action given by \[K_i^{\pm 1}\triangleright a = K_i^{\pm1}(a),\quad F_{i,1}\triangleright a = F_i(a)-\frac{x_ia-K_i^{-1}(a)x_i}{q_i-q_i^{-1}},\quad F_{i,2}\triangleright a = \frac{x_iK_i^{-1}(a)-ax_i}{q_i-q_i^{-1}}.\] In particular, the subalgebra $A^+$ is invariant under this action of $U_q(\mathfrak{g}^*)$ and is therefore a $U_q(\mathfrak{g}^*)$-module subalgebra. \end{theorem} Theorem \ref{thm:qgd} is in some sense a statement about the existence of a functor. To make this precise, we introduce a category whose objects bear properties similar to those found in Theorem \ref{thm:qfactoring}. \begin{definition} Let $\mathcal{C}_\mathfrak{g}^q$ be the category whose objects consist of pairs $(A,\varphi_A)$, where \begin{itemize} \item $A$ is an adapted $U_q(\mathfrak{g})$-module algebra such that $\nu_{\bf i}(A\setminus \{0\})=\nu_{\bf i}(\mathbb{C}_q[U]\setminus\{0\})$ for all ${\bf i}\in R(w_\mathrm{o})$. \item $\varphi_A:\mathbb{C}_q[U]\hookrightarrow A$ is an embedding of $U_q(\mathfrak{g})$-module algebras. \end{itemize} A morphism $(A,\varphi_A)\to (B,\varphi_B)$ in $\mathcal{C}_\mathfrak{g}^q$ is a homomorphism of $U_q(\mathfrak{g})$-module algebras $\psi:A\to B$ such that $\psi\circ\varphi_A=\varphi_B$. \end{definition} Given a homomorphism of $U_q(\mathfrak{g})$-module algebras $\psi:A\to B$, it follows that $\psi(A^+)\subseteq B^+$, so $\psi|_{A^+}$ may be thought of as a map of $\mathbb{C}(q)$-algebras $A^+\to B^+$. If $\psi$ is a morphism in $\mathcal{C}_\mathfrak{g}^q$, $(A,\varphi_A){\to} (B,\varphi_B)$, then actually $\psi|_{A^+}$ is a homomorphism of $U_q(\mathfrak{g}^*)$-module algebras. As a consequence, we have the following corollary. \begin{corollary} There is a functor $(-)^+:\mathcal{C}_\mathfrak{g}^q\to U_q(\mathfrak{g}^*)-\textbf{ModAlg}$ (denoted $P_q\circ F_q$ in Section \ref{sect:Introduction}) which assigns to an object $(A,\varphi_A)$ of $\mathcal{C}_\mathfrak{g}^q$ its subalgebra of highest weight vectors $A^+$, equipped with the $U_q(\mathfrak{g}^*)$-module algebra structure of Theorem \ref{thm:qgd}. The functor $(-)^+$ is given on morphisms by restriction. \end{corollary} Theorem \ref{thm:qfactoring} strongly suggests that $(-)^+$ might actually be an equivalence of categories. In fact, this is the case, but in order to describe a quasi-inverse, we need the following theorem which describes a $U_q(\mathfrak{g})$-module algebra structure on $A\otimes \mathbb{C}_q[U]$ if $A$ is a $U_q(\mathfrak{g}^*)$-module algebra. \begin{theorem} \label{thm:qg} If $A$ is a $U_q(\mathfrak{g}^*)$-module algebra, then $A\otimes \mathbb{C}_q[U]$ has the structure of a $U_q(\mathfrak{g})$-module algebra determined by: \begin{align*} (1\otimes x_i)(a\otimes 1)&=K_i(a)\otimes x_i+(q_i-q_i^{-1})F_{i,2}K_i(a)\otimes 1\\ K_i^{\pm1}\triangleright(a\otimes x)&=K_i^{\pm1}(a)\otimes K_i^{\pm1}(x)\\ E_i\triangleright(a\otimes x)&=a\otimes E_i(x)\\ F_i\triangleright(a\otimes x)&=(F_{i,1}(a)+F_{i,2}K_i(a))\otimes x+\frac{K_i(a)-K_i^{-1}(a)}{q_i-q_i^{-1}}\otimes x_ix+K_i^{-1}(a)\otimes F_i(x). \end{align*} \end{theorem} Theorem \ref{thm:qg} will be proved in Section \ref{pf:qgd,qg}. Since the action of each $E_i$ is completely described on the $\mathbb{C}_q[U]$ factor and $\mathbb{C}_q[U]$ is adapted, we have the following corollary. \begin{corollary} \label{cor:qgcor} There is a functor $(-)\otimes \mathbb{C}_q[U]:U_q(\mathfrak{g}^*)-\textbf{ModAlg}\to \mathcal{C}_\mathfrak{g}^q$ which assigns to a $U_q(\mathfrak{g}^*)$-module algebra $A$, the pair $(A\otimes \mathbb{C}_q[U], 1\otimes \text{id})$, where $A\otimes \mathbb{C}_q[U]$ is given the $U_q(\mathfrak{g})$-module structure of Theorem \ref{thm:qg}. The functor $(-)\otimes \mathbb{C}_q[U]$ is given on morphisms by $\psi\mapsto \psi\otimes \text{id}$. \end{corollary} The following theorem says that $(-)\otimes \mathbb{C}_q[U]$ is the promised quasi-inverse for $(-)^+$. \begin{theorem} \label{thm:qequiv} The functors $(-)^+:\mathcal{C}_\mathfrak{g}^q\to U_q(\mathfrak{g}^*)-{\bf ModAlg}$ and $(-)\otimes \mathbb{C}_q[U]:U_q(\mathfrak{g}^*)-{\bf ModAlg}\to \mathcal{C}_\mathfrak{g}^q$ are quasi-inverses of each other and thus provide equivalences of categories. \end{theorem} Theorem \ref{thm:qequiv} is proved in section \ref{pf:qequiv}. Now, it is well-known that if $A$ and $B$ are $U_q(\mathfrak{g})$-weight module algebras, then so is the braided tensor product $A\underline{\otimes}B$. Here $A\underline{\otimes}B$ has multiplication given by $(a\otimes b)(a'\otimes b')=a\mathcal{R}_{(2)}(a')\otimes \mathcal{R}_{(1)}(b)b'$, where $\mathcal{R}$ is the universal $R$-matrix for $U_q(\mathfrak{g})$ and we use sumless Sweedler notation $\mathcal{R}=\mathcal{R}_{(1)}\otimes \mathcal{R}_{(2)}$. The $R$-matrix is of the form $\mathcal{R}=q^{\sum\limits_{i,j}d_i(C^{-1})_{i,j}H_i\otimes H_j}\left(\prod\limits_{\alpha>0}^{\leftarrow} e_{q_\alpha^{-2}}^{(q_{\alpha}-q_{\alpha}^{-1})E_\alpha\otimes F_\alpha}\right)$ (see Section 3.3 in \cite{majid} for details). At this point, we may need to use the field $\mathbb{C}(q^{1/d})$ where $d$ is the determinant of $C$ or else assume that $d_i(C^{-1})_{i,j}\in \mathbb{Z}$ for all $i,j\in I$, but this is a small matter which doesn't affect our overall approach. We define a subcategory of $\mathcal{C}_\mathfrak{g}^q$ on which the tensor product defined above will make sense. \begin{definition} Let $\underline{\mathcal{C}}_\mathfrak{g}^q$ be the full subcategory of $\mathcal{C}_\mathfrak{g}^q$ whose objects consist of pairs $(A,\varphi_A)$, where $A$ is additionally assumed to be a $U_q(\mathfrak{g})$-weight module algebra. \end{definition} The following proposition is then clear. \begin{proposition} \label{prop:qequiv} The functors $(-)^+$ and $(-)\otimes \mathbb{C}_q[U]$ restrict to equivalences between $\underline{\mathcal{C}}_\mathfrak{g}^q$ and $U_q(\mathfrak{g}^*)-{\bf WModAlg}$, the category of $U_q(\mathfrak{g}^*)$-weight module algebras. \end{proposition} If $(A,\varphi_A)$ and $(B,\varphi_B)$ are objects of $\underline{\mathcal{C}}_\mathfrak{g}^q$, then we already saw that $A\underline{\otimes}B$ is also a $U_q(\mathfrak{g})$-weight module algebra. Furthermore, it is obvious that $1\otimes \varphi_B$ and $\varphi_A\otimes 1$ are injections $\mathbb{C}_q[U]\hookrightarrow A\underline{\otimes} B$. However, it is not immediately obvious that $A\underline{\otimes} B$ is adapted with $\nu_{\bf i}(A\underline{\otimes} B\setminus\{0\})=\nu_{\bf i}(\mathbb{C}_q[U]\setminus\{0\})$ for all ${\bf i}\in R(w_\mathrm{o})$. Nevertheless, this is the case, which the following proposition asserts. \begin{proposition} \label{prop:qtensor} If $(A,\varphi_A)$ and $(B,\varphi_B)$ are objects of $\underline{\mathcal{C}}_\mathfrak{g}^q$, then $(A\underline{\otimes}B, 1\otimes \varphi_B)$ and $(A\underline{\otimes}B, \varphi_A\otimes 1)$ are objects of $\underline{\mathcal{C}}_\mathfrak{g}^q$ as well. \end{proposition} Proposition \ref{prop:qtensor} is proved in Section \ref{pf:qtensor}. Proposition \ref{prop:qequiv} allows us to turn Proposition \ref{prop:qtensor} into a statement about $U_q(\mathfrak{g}^*)$-module algebras. We define two ``fusion" products on the category of $U_q(\mathfrak{g}^*)$-weight module algebras, namely the following: \begin{align*} A*B&:=((A\otimes \mathbb{C}_q[U])\underline{\otimes} (B\otimes \mathbb{C}_q[U]),1\otimes 1\otimes 1\otimes \text{id})^+\\ A\star B&:=((A\otimes \mathbb{C}_q[U])\underline{\otimes} (B\otimes \mathbb{C}_q[U]),1\otimes \text{id}\otimes 1\otimes 1)^+ \end{align*} These fusion products are associative, but not monoidal due to the easy observation that there is no unit object. The reader may be bothered that the objects $(A\underline{\otimes}B,1\otimes \varphi_B)$ and $(A\underline{\otimes}B,\varphi_A\otimes 1)$ are not (necessarily at least) isomorphic despite having equal underlying $U_q(\mathfrak{g})$-module algebras. An attempt to force a common quotient leads to the discovery of an interesting $U_q(\mathfrak{g}^*)$-module algebra structure on $A\otimes B$ if $A$ and $B$ are $U_q(\mathfrak{g}^*)$-weight module algebras. \begin{proposition} \label{prop:qfusion} Let $A$ and $B$ be $U_q(\mathfrak{g}^*)$-weight module algebras. Then the $\mathbb{C}(q)$-vector space $A\otimes B$ has the structure of a $U_q(\mathfrak{g}^*)$-weight module algebra satisfying the following equations \begin{align*} (a\otimes b)(a'\otimes b')&=q^{(|a'|,|b|)}aa'\otimes bb'\\ K_i^{\pm1}\rhd(a\otimes b)&=K_i^{\pm1}(a)\otimes K_i^{\pm1}(b)\\ F_{i,1}\rhd(a\otimes b)&=K_i^{-1}(a)\otimes F_{i,1}(b)\\ F_{i,2}\rhd(a\otimes b)&=F_{i,2}(a)\otimes K_i^{-1}(b). \end{align*} for weight vectors $a,a'\in A$, $b,b'\in B$, and $i\in I$, where $|\cdot|$ indicates weight. \end{proposition} Proposition \ref{prop:qfusion} is proved in Section \ref{pf:qfusion} and induces a fusion product on $\underline{\mathcal{C}}_\mathfrak{g}^q$: \[(A,\varphi_A)\diamond (B,\varphi_B):=(((A,\varphi_A)^+\otimes (B,\varphi_B)^+)\otimes \mathbb{C}_q[U],1\otimes 1\otimes \text{id}).\] Just like for $*$ and $\star$, there is no unit object for $\diamond$, so it is not a monoidal tensor product. \subsection{Classical Factorization} \label{sect:classical} Throughout this section, all tensor products will be taken over $\mathbb{C}$ unless otherwise specified and written $-\otimes-$ rather than $-\otimes_\mathbb{C} -$. We also assume henceforth that every algebra is commutative unless otherwise stated, with the exception of previously referenced algebras such as $\mathbb{C}_q[U]$. The semisimple complex Lie algebra $\mathfrak{g}$ with Cartan matrix $C=(c_{i,j})$ is generated by elements $\{e_i,f_i,h_i\ |\ i\in I\}$ subject to the following relations: $ [h_i,h_j]=0,~[e_i,f_j]=\delta_{i,j}h_i,~\null[h_i,e_j]=c_{i,j}e_j,~[h_i,f_j]=c_{i,j}f_j,~ (\text{ad } e_i)^{1-c_{i,j}}(e_j)=(\text{ad } f_i)^{1-c_{i,j}}(f_j)=0,$$ where as usual $(\text{ad }x)(y)=[x,y]$ for $x,y\in \mathfrak{g}$. The universal enveloping algebra $U(\mathfrak{g})$ of $\mathfrak{g}$ is a noncommutative Hopf algebra on the same generators and relations, where $[x,y]=xy-yx$ for $x,y\in U(\mathfrak{g})$. The comultiplication of $U(\mathfrak{g})$ is given on generators by \[\Delta(x)=x\otimes 1+1\otimes x\] for $x\in \{e_i,f_i,h_i\ |\ i\in I\}$. We denote by $\mathfrak{n}_+$ (respectively $\mathfrak{b}_-$) the Lie subalgebra of $\mathfrak{g}$ generated by all $e_i$ (respectively $h_i$ and $f_i$). We will assume henceforth that for any $i\in I$, the action of $e_i$ on any $\mathfrak{n}_+$-module is \emph{locally nilpotent}. In other words, if $M$ is an $\mathfrak{n}_+$-module, we will assume that for each $x\in M$ and $i\in I$, there exists some $n\ge 0$ such that $e_i^n(x)=0$. Note that every $\mathfrak{g}$-module is also a $\mathfrak{n}_+$-module, so we are assuming these are ``bounded above" as well. We do not assume the same for the action of $f_i$. For an $\mathfrak{n}_+$-module $M$, we designate $M^+:=\{m\in M\ |\ e_i(m)=0\ \forall i\in I\}$, the set of \emph{highest weight vectors}. If $A$ is an $\mathfrak{n}_+$-module algebra, then $A^+$ is an $\mathfrak{n}_+$-module subalgebra. For $n\in \mathbb{Z}_{\ge 0}$ and $i\in I$, we will use the notation $e_i^{(n)}=\frac{1}{n!}e_i^n$. Suppose $M$ is an $\mathfrak{n}_+$-module algebra. For each $i\in I$ and $x\in M\setminus\{0\}$, set $\ell_i(x)=\max\{\ell\in\mathbb{Z}_{\ge0}\ |\ e_i^\ell(x)\ne 0\}$ and $e_i^{(top)}(x)=e_i^{(\ell_i(x))}(x)$. Given ${\bf i}\in I^m$ for some $m\ge 0$, and $x\in M\setminus\{0\}$, we also use the shorthand \[e_{\bf i}^{(top)}(x)=e_{i_m}^{(top)}e_{i_{m-1}}^{(top)}\cdots e_{i_1}^{(top)}(x)\] and define $\nu_{\bf i}:M\setminus \{0\}\to \mathbb{Z}_{\ge 0}^m$, $x\mapsto(a_1,a_2,\ldots,a_m)$ by the following: \[a_k=\ell_{i_k}(e_{i_{k-1}}^{(top)}e_{i_{k-2}}^{(top)}\cdots e_{i_1}^{(top)}(x)).\] Lastly, for ${\bf j}=(j_1,\ldots,j_m)\in \mathbb{Z}_{\ge 0}^m$, we set $e_{\bf i}^{({\bf j})}:=e_{i_m}^{(j_m)}e_{i_{m-1}}^{(j_{m-1})}\cdots e_{i_1}^{(j_1)}$. \begin{definition} Let $A$ be an $\mathfrak{n}_+$-module algebra, $w\in W$, and ${\bf i}\in R(w)$. If $e_{\bf i}^{(top)}(x)\in A^+$ for all $x\in A\setminus\{0\}$, then we say $A$ is \emph{{\bf i}-adapted}. We say a basis $\mathcal{B}$ for $A$ is an \emph{{\bf i}-adapted basis} if \begin{enumerate} \item $e_{\bf i}^{(top)}(b)=1$ for all $b\in \mathcal{B}$. \item The restriction of $\nu_{\bf i}$ to $\mathcal{B}$ is an injective map $\mathcal{B}\hookrightarrow \mathbb{Z}_{\ge0}^m$, where $m$ is the length of $w$. \end{enumerate} If there exists any $w\in W$ and ${\bf i}\in R(w)$ so that $A$ is {\bf i}-adapted, then we say more generally that $A$ is \emph{adapted}. \end{definition} As in the quantum case, if $A_0$ possesses an ${\bf i}$-adapted basis for some ${\bf i}\in R(w)$ and is a ``large enough" $\mathfrak{n}_+$-module subalgebra of $A$, then $A$ is factorizable over $A_0$. The following theorem makes this precise. \begin{theorem} \label{thm:factoring} Let $A$ be an $\mathfrak{n}_+$-module algebra. Suppose $A_0$ be an $\mathfrak{n}_+$-module subalgebra of $A$ possessing an {\bf i}-adapted basis $\mathcal{B}$ for some reduced $\mathbf{i}$. Then \begin{enumerate} \item The restriction $\mu:A^+\otimes A_0\to A$ of the multiplication in $A$ is an injective homomorphism of $\mathfrak{n}_+$-modules. \item The map $\mu$ is an isomorphism if and only if $A$ is {\bf i}-adapted and $\nu_{\bf i}(A\setminus\{0\})=\nu_{\bf i}(A_0\setminus\{0\})$. \end{enumerate} \end{theorem} Theorem \ref{thm:factoring} demonstrates a close relationship between being {\bf i}-adapted and being factorizable over an $\mathfrak{n}_+$-module subalgebra possessing an {\bf i}-adapted basis. The following theorem explores this relationship from a different angle. \begin{theorem} \label{thm:classify} Let $A$ be an ${\bf i}$-adapted $\mathfrak{n}_+$-module algebra for some reduced $\mathbf{i}$ and suppose $A_0$ is an $\mathfrak{n}_+$-module subalgebra of $A$. Then $A$ is factorizable over $A_0$ if and only if $A_0$ possesses an ${\bf i}$-adapted basis and $\nu_{\bf i}(A_0\setminus\{0\})=\nu_{\bf i}(A\setminus\{0\})$. \end{theorem} The proofs of Theorems \ref{thm:factoring} and \ref{thm:classify} are nearly identical to those of Theorems \ref{thm:qfactoring} and \ref{thm:qclassify}, respectively, so we do not replicate them here. We now restrict our focus to a specific $\mathfrak{g}$-module algebra, $\mathbb{C}[U]$. Actually, $\mathbb{C}[U]$ is a specialization of $\mathbb{C}_q[U]$ to $q=1$. This is accomplished as follows. It is well-known (see, e.g. \cite[Section 4]{berensteingreenstein}) that $\mathbb{C}_q[U]$ admits a form $\underline{\mathbb{C}_q[U]}$ over $\mathbb{A}=\mathbb{Z}[q,q^{-1}]$ which has both a PBW-basis and dual canonical basis. That is, the structure constants of the aforementioned bases belong to $\mathbb{A}$. It is also well-known (see, e.g., \cite[Section 3.3]{berensteingreenstein}) that $E_i^{(n)}(\underline{\mathbb{C}_q[U]})\subset \underline{\mathbb{C}_q[U]}$ for all $i\in I$ and $n\in \mathbb{Z}_{\ge 0}$. In particular, the quotient of $\underline{\mathbb{C}_q[U]}$ by the ideal $(q-1)$ generated by $q-1$ is a commutative algebra canonically isomorphic to $\mathbb{Z}[U]$. Tensoring by $\mathbb{C}$, we obtain $\mathbb{C}[U]$ as the classical limit of $\mathbb{C}_q[U]$. The action of $E_i$ specializes to the derivations which generate the action of $\mathfrak{n}_+$ on $\mathbb{C}[U]$. This in particular implies the well-known fact that $\mathbb{C}[U]$ is a Poisson algebra with the Poisson bracket given by \[\{f,g\}=\dfrac{\tilde{f}\tilde{g}-\tilde{g}\tilde{f}}{q-1} \mod (q-1)\] for all $f,g\in \mathbb{C}[U]$, where $\tilde{f}$ and $\tilde{g}$ denote any representatives of $f$ and $g$, respectively, modulo $(q-1)$. Since $\mathbb{C}_q[U]$ is generated by $\{x_i\ |\ i\in I\}$, $\mathbb{C}[U]$ has Poisson generators which we denote by slight abuse of notation $\{x_i\ |\ i\in I\}$. The quantum Serre relations for the quantum $x_i$ imply the following relations for the ``classical" versions \[\varepsilon(i,j,1-c_{i,j})=0 \text{ if }i\ne j,\] where $\varepsilon(i,j,n)$ is defined inductively by $\varepsilon(i,j,0)=x_j$, $\varepsilon(i,j,n+1)=\{x_i,\varepsilon(i,j,n)\}-d_i(c_{i,j}+2n)x_i\varepsilon(i,j,n)$. The $\mathfrak{g}$-module structure on $\mathbb{C}[U]$ is summarized in the following equations: \begin{align*} h_i(x_j)&=-c_{i,j}x_j\quad \forall i,j\in I\\ h_i(\{x,y\})&=\{h_i(x),y\}+\{x,h_i(y)\}\quad \forall i\in I,\ x,y\in \mathbb{C}[U]\\ e_i(x_j)&=\delta_{i,j} \quad \forall i,j\in I\\ e_i(\{x,y\})&=\{e_i(x),y\}+\{x,e_i(y)\}+d_ie_i(x)h_i(y)-d_ih_i(x)e_i(y)\quad \forall i\in I,\ x,y\in \mathbb{C}[U]\\ f_i(x)&=\frac{1}{2d_i}\{x_i,x\}+\frac{1}{2}x_ih_i(x)\quad \forall i\in I,\ x\in \mathbb{C}[U]. \end{align*} \begin{remark} \label{rem:nilpotentf} Comparing the defining relations of $\mathbb{C}[U]$ with the action of $f_i$ thereon, one sees that $f_i(x_i)=-x_i^{2}$ and $\varepsilon(i,j,n)=(2d_i)^nf_i^n(x_j)$, so that $f_i^{1-c_{i,j}}(x_j)=0$ if $i\ne j$. Observe also that $\{x_i,x\}=2d_if_i(x)-d_ix_ih_i(x)$ for $x\in \mathbb{C}[U]$ and $i\in I$. It follows that $\mathbb{C}[U]$ is generated as a $\mathfrak{g}$-module algebra by $\{x_i\ |\ i\in I\}$. \end{remark} \begin{example} \label{ex:mat} Consider a $3\times 2$ matrix with complex coefficients: $A=\begin{pmatrix*} a_{1,1} & a_{1,2}\\ a_{2,1} & a_{2,2}\\ a_{3,1} & a_{3,2} \end{pmatrix*}$. If $a_{1,1}\ne0$ and $a_{1,1}a_{2,2}-a_{1,2}a_{2,1}\ne0$, then $A$ has Gauss factorization \[A=\begin{pmatrix*} 1 & 0 & 0\\ \frac{a_{2,1}}{a_{1,1}} & 1 & 0\\ \frac{a_{3,1}}{a_{1,1}} & \frac{a_{1,1}a_{3,2}-a_{1,2}a_{3,1}}{a_{1,1}a_{2,2}-a_{1,2}a_{2,1}} & 1 \end{pmatrix*} \begin{pmatrix*} a_{1,1} & a_{1,2} \\ 0 & \frac{a_{1,1}a_{2,2}-a_{1,2}a_{2,1}}{a_{1,1}}\\ 0 & 0 \end{pmatrix*}.\] Denote by $x_{i,j}$, the $(i,j)$-th coordinate function in $\mathbb{C}[Mat_{3,2}]$, i.e. $x_{i,j}(A)=a_{i,j}$. The Gauss factorization above implies that upon localization of $\mathbb{C}[Mat_{3,2}]$ by the principal minors $x_{1,1}$ and $\Delta_2=x_{1,1}x_{2,2}-x_{1,2}x_{2,1}$, we obtain an isomorphism of algebras \[\mathbb{C}[Mat_{3,2}][x_{1,1}^{-1}, \Delta_2^{-1}]\cong \mathbb{C}\left[\frac{x_{2,1}}{x_{1,1}},\frac{x_{3,1}}{x_{1,1}},\frac{x_{1,1}x_{3,2}-x_{1,2}x_{3,1}}{\Delta_2}\right]\otimes \mathbb{C}\left[x_{1,1}^{\pm1},x_{1,2},\Delta_2^{\pm1}\right]\] The natural action of $\mathfrak{sl}_3(\mathbb{C})$ extends to the localized algebra and a short examination verifies that \[\left(\mathbb{C}[Mat_{3,2}][x_{1,1}^{-1}, \Delta_2^{-1}]\right)^+=\mathbb{C}\left[x_{1,1}^{\pm1},x_{1,2},\Delta_2^{\pm1}\right]~\text{and}~\mathbb{C}\left[\frac{x_{2,1}}{x_{1,1}},\frac{x_{3,1}}{x_{1,1}},\frac{x_{1,1}x_{3,2}-x_{1,2}x_{3,1}}{\Delta_2}\right]\cong \mathbb{C}[U],\] where the isomorphism is an isomorphism of $\mathfrak{sl}_3(\mathbb{C})$-module algebras and the generators $x_1$ and $x_2$ of $\mathbb{C}[U]$ are mapped to by $\dfrac{x_{2,1}}{x_{1,1}}$ and $\dfrac{x_{1,1}x_{3,2}-x_{1,2}x_{3,1}}{\Delta_2}$, respectively. \end{example} \begin{example} \label{ex:qmat} Recall that $\mathbb{C}_q[Mat_{3,2}]$ is generated by $\{x_{i,j}\ |\ 1\le i\le 3,\ 1\le j\le 2\}$, subject to relations \begin{align*} x_{k,j}x_{i,j}&=qx_{i,j}x_{k,j} \text{ if }i<k,\\ x_{i,k}x_{i,j}&=qx_{i,j}x_{i,k} \text{ if }j<k,\\ x_{k,\ell}x_{i,j}&=x_{i,j}x_{k,\ell}\text{ if }i<k\text{ and }j>\ell,\\ x_{k,\ell}x_{i,j}&=x_{i,j}x_{k,\ell}+(q-q^{-1})x_{i,\ell}x_{k,j}\text{ if }i<k\text{ and }j<\ell. \end{align*} Then $\mathbb{C}_q[Mat_{3,2}][x_{1,1}^{-1}, \Delta_2^{-1}]\cong \mathbb{C}_q\left[x_{1,1}^{\pm1},x_{1,2},\Delta_2^{\pm1}\right] \otimes \mathbb{C}_q\left[x_{1,1}^{-1}x_{2,1},x_{1,1}^{-1}x_{3,1},\Delta_2^{-1}(x_{1,1}x_{3,2}-q^{-1}x_{1,2}x_{3,1})\right]$, where $\Delta_2=x_{1,1}x_{2,2}-q^{-1}x_{1,2}x_{2,1}$ and $\mathbb{C}_q[-]$ denotes the subalgebra of $\mathbb{C}_q[Mat_{3,2}][x_{1,1}^{-1}, \Delta_2^{-1}]$ generated by those elements appearing inside the brackets. The natural action of $U_q(\mathfrak{sl}_m(\mathbb{C}))$ extends to the localized algebra and a short examination verifies that \[\left(\mathbb{C}_q[Mat_{m,n}][x_{1,1}^{-1}, (x_{1,1}x_{2,2}-q^{-1}x_{1,2}x_{2,1})^{-1}]\right)^+=\mathbb{C}_q\left[x_{1,1}^{\pm1},x_{1,2},(x_{1,1}x_{2,2}-q^{-1}x_{1,2}x_{2,1})^{\pm1}\right]\quad \text{and}\] \[\mathbb{C}_q\left[x_{1,1}^{-1}x_{2,1},x_{1,1}^{-1}x_{3,1},(x_{1,1}x_{2,2}-q^{-1}x_{1,2}x_{2,1})^{-1}(x_{1,1}x_{3,2}-q^{-1}x_{1,2}x_{3,1})\right]\cong \mathbb{C}[U],\] where the isomorphism is an isomorphism of $U_q(\mathfrak{sl}_m(\mathbb{C}))$-module algebras and the generators $x_1$ and $x_2$ of $\mathbb{C}_q[U]$ are mapped to by $x_{1,1}^{-1}x_{2,1}$ and $(x_{1,1}x_{2,2}-q^{-1}x_{1,2}x_{2,1})^{-1}(x_{1,1}x_{3,2}-q^{-1}x_{1,2}x_{3,1})$, respectively. \end{example} Now, since $\mathbb{C}_q[U]$ possessed an $\mathbf{i}$-adapted basis for every $\mathbf{i}\in R(w_\mathrm{o})$ and the action of $e_i$ on $\mathbb{C}[U]$ is induced by that of $E_i$ on $\mathbb{C}_q[U]$, it follows that $\mathbb{C}[U]$ possesses an $\mathbf{i}$-adapted basis for each $\mathbf{i}\in R(w_\mathrm{o})$. Combining this fact with Theorem \ref{thm:factoring} leads us to the following corollary. \begin{corollary} \label{cor:highest} Let $A$ be a $\mathfrak{g}$-module algebra containing $\mathbb{C}[U]$ as a $\mathfrak{g}$-module subalgebra. The map $\mu:A^+\otimes \mathbb{C}[U]\to A$ as in Theorem \ref{thm:factoring} is an isomorphism of $\mathfrak{n}_+$-modules if and only if there exists a $\mathfrak{g}$-module algebra $A'$ which contains $A$ as a $\mathfrak{g}$-module subalgebra and is generated by $(A')^+$ as a $\mathfrak{g}$-module algebra. \end{corollary} The proof of the ``if" part of Corollary \ref{cor:highest} is very similar to the proof of Corollary \ref{cor:qhighest} so we do not reproduce it here. The ``only if" part will be a very easy consequence of the discussion at the end of Section \ref{sect:classical}, so we will address it there. Also, just as in the quantum setting, Corollary \ref{cor:highest} leads to a proof of Theorem \ref{thm:G/U}, which is nearly identical to that of Theorem \ref{thm:qG/U}, so we will not include it here. We do, however, note that instead of two families of quantities as arose in the proof of Corollary \ref{cor:qhighest}, only one arises in the proof of the ``if" part of Corollary \ref{cor:highest}: \[f_i(a)-h_i(a)x_i\] where $i\in I$ and $a\in A'$. As in the quantum case, if $a\in A^+$, then $f_i(a)-h_i(a)x_i\in A^+$. Once again it is natural to ask what relations the family of operators $f_i-m_ih_i$ satisfies, where $m_i$ denotes multiplication by $x_i$. Or to put it another way, do these operators indicate the action of a known algebra or Lie algebra which is somehow related to $\mathfrak{g}$? Again, we can answer in the affirmative, which the following theorem summarizes. \begin{theorem} \label{thm:b-} Let $A$ be a $\mathfrak{g}$-module algebra containing $\mathbb{C}[U]$ as a $\mathfrak{g}$-module subalgebra. Then $A$ has another structure of a $\mathfrak{b}_-$-module algebra with action given by the formulas \[h_i\rhd a=h_i(a),\quad f_i\rhd a=f_i(a)-h_i(a)x_i.\] In particular, the subalgebra $A^+$ is invariant under this $\mathfrak{b}_-$ action and is therefore a $\mathfrak{b}_-$-module subalgebra. \end{theorem} Theorem \ref{thm:b-} is proved in Section \ref{pf:b-,g}. As in the quantum case, Theorem \ref{thm:b-} is in some sense a statement about the existence of a functor. To make this precise, we introduce a category whose objects bear properties similar to those found in Theorem \ref{thm:factoring}. \begin{definition} Let $\mathcal{C}_\mathfrak{g}$ be the category whose objects consist of pairs $(A,\varphi_A)$, where \begin{itemize} \item $A$ is an adapted $\mathfrak{g}$-module algebra such that $\nu_{\bf i}(A\setminus \{0\})=\nu_{\bf i}(\mathbb{C}[U]\setminus \{0\})$ for all ${\bf i}\in R(w_\mathrm{o})$; \item $\varphi_A:\mathbb{C}[U]\hookrightarrow A$ is an embedding of $\mathfrak{g}$-module algebras. \end{itemize} A morphism $(A,\varphi_A)\to (B,\varphi_B)$ in $\mathcal{C}_\mathfrak{g}$ is a homomorphism of $\mathfrak{g}$-module algebras $\psi:A\to B$ such that $\psi\circ \varphi_A=\varphi_B$. \end{definition} Given a homomorphism of $\mathfrak{g}$-module algebras $\psi:A\to B$, it follows that $\psi(A^+)\subseteq B^+$, so $\psi|_{A^+}$ may be thought of as a map of $\mathbb{C}$-algebras $A^+\to B^+$. If $\psi$ is a morphism in $\mathcal{C}_\mathfrak{g}$, $(A,\varphi_A)\to (B,\varphi_B)$, then actually $\psi|_{A^+}$ is a homomorphism of $\mathfrak{b}_-$-module algebras, where the $\mathfrak{b}_-$-module structure is the one given in Theorem \ref{thm:b-}. As a consequence, we have the following corollary. \begin{corollary} There is a functor $(-)^+:\mathcal{C}_\mathfrak{g}\to \mathfrak{b}_--{\bf ModAlg}$ (denoted $P\circ F$ in Section \ref{sect:Introduction}) which assigns to an object $(A,\varphi_A)$ of $\mathcal{C}_\mathfrak{g}$ its subalgebra of highest weight vectors $A^+$, equipped with the $\mathfrak{b}_-$-module algebra structure of Theorem \ref{thm:b-}. The functor $(-)^+$ is given on morphisms by restriction. \end{corollary} Again, we hope for a quasi-inverse functor for $(-)^+$, but we must first state the following theorem. \begin{theorem} \label{thm:g} If $A$ is a $\mathfrak{b}_-$-module algebra, then the usual tensor product of commutative algebras $A\otimes \mathbb{C}[U]$ has the structure of a $\mathfrak{g}$-module algebra determined by: \begin{align*} h_i\rhd(a\otimes x)&=h_i(a)\otimes x+a\otimes h_i(x)\\ e_i\rhd(a\otimes x)&=a\otimes e_i(x)\\ f_i\rhd(a\otimes x)&=f_i(a)\otimes x+h_i(a)\otimes x_ix+a\otimes f_i(x). \end{align*} \end{theorem} Theorem \ref{thm:g} will be proved in Section \ref{pf:b-,g}. Since the action of each $e_i$ is completely described on the $\mathbb{C}[U]$ factor and $\mathbb{C}[U]$ is adapted, the following corollary is almost immediate. \begin{corollary} There is a functor $(-)\otimes \mathbb{C}[U]:\mathfrak{b}_--{\bf ModAlg}\to \mathcal{C}_\mathfrak{g}$ which assigns to a $\mathfrak{b}_-$-module algebra $A$, the pair $(A\otimes \mathbb{C}[U],1\otimes \text{id})$, where $A\otimes \mathbb{C}[U]$ is given the $\mathfrak{g}$-module algebra structure of Theorem \ref{thm:g}. The functor $(-)\otimes \mathbb{C}[U]$ is given on morphisms by $\psi\mapsto \psi\otimes \text{id}$. \end{corollary} As promised, $(-)\otimes \mathbb{C}[U]$ is the desired quasi-inverse for $(-)^+$. \begin{theorem} \label{thm:equiv} The functors $(-)^+:\mathcal{C}_\mathfrak{g}\to \mathfrak{b}_--{\bf ModAlg}$ and $(-)\otimes \mathbb{C}[U]:\mathfrak{b}_--{\bf ModAlg}\to \mathcal{C}_\mathfrak{g}$ are quasi-inverses of each other and thus provide equivalences of categories. \end{theorem} The proof of Theorem \ref{thm:equiv} is very similar to that of \ref{thm:qequiv}, so we do not include it here. Now, it is well-known that if $A$ and $B$ are $\mathfrak{g}$-module algebras, then so is $A\otimes B$. Here $A\otimes B$ has the na\"ive multiplication $(a\otimes b)(a'\otimes b')=aa'\otimes bb'$. So if $(A,\varphi_A)$ and $(B,\varphi_B)$ are objects of $\mathcal{C}_\mathfrak{g}$, then $A\otimes B$ is a $\mathfrak{g}$-module algebra. Furthermore, it is obvious that $1\otimes \varphi_B$ and $\varphi_A\otimes 1$ are injections $\mathbb{C}[U]\hookrightarrow A{\otimes} B$. However, it is not immediately obvious that $A{\otimes} B$ is adapted with $\nu_{\bf i}(A{\otimes} B\setminus\{0\})=\nu_{\bf i}(\mathbb{C}[U]\setminus\{0\})\ \forall {\bf i}\in R(w_\mathrm{o})$. Nevertheless, this is the case and the following proposition asserts as much and is proved in Section \ref{pf:tensor}. \begin{proposition} \label{prop:tensor} If $(A,\varphi_A)$ and $(B,\varphi_B)$ are objects of $\mathcal{C}_\mathfrak{g}$, then $(A\otimes B, 1\otimes \varphi_B)$ and $(A\otimes B,\varphi_A\otimes 1)$ are objects of $\mathcal{C}_\mathfrak{g}$ as well. \end{proposition} Theorem \ref{thm:equiv} and Proposition \ref{prop:tensor} allow us to define two ``fusion" products on the category of $\mathfrak{b}_-$-module algebras, namely the following: \begin{align*} A*B&:=((A\otimes \mathbb{C}[U]){\otimes} (B\otimes \mathbb{C}[U]),1\otimes 1\otimes 1\otimes \text{id})^+\\ A\star B&:=((A\otimes \mathbb{C}[U]){\otimes} (B\otimes \mathbb{C}[U]),1\otimes \text{id}\otimes 1\otimes 1)^ \end{align*} Unfortunately, as in the quantum case, these fusion products are not monoidal as there is no unit object. Furthermore, the objects $(A\otimes B,1\otimes \varphi_B)$ and $(A\otimes B,\varphi_A\otimes 1)$ are not necessarily isomorphic, despite having equal underlying $\mathfrak{g}$-module algebras. However, given $\mathfrak{b}_-$-module algebras $A$ and $B$, we of course have the natural $\mathfrak{b}_-$-module algebra structure on $A\otimes B$ satisfying $$ (a\otimes b)(a'\otimes b')=aa'\otimes bb',~h_i(a\otimes b)=h_i(a)\otimes b+a\otimes h_i(b),~ f_i(a\otimes b)=f_i(a)\otimes b+a\otimes f_i(b)$$ for $a,a'\in A$, $b,b'\in B$, and $i\in I$. This induces a more symmetric fusion product on $\mathcal{C}_\mathfrak{g}$: \begin{equation} \label{eq:fusion} (A,\varphi_A)\diamond (B,\varphi_B):=(((A,\varphi_A)^+\otimes (B,\varphi_B)^+)\otimes \mathbb{C}[U],1\otimes 1\otimes \text{id}). \end{equation} \begin{remark} \label{rmk:unitissue} Since $\mathfrak{b}_-$-{\bf ModAlg} is a monoidal category with unit object $\mathbb{C}$, the product $\diamond$ makes $\mathcal{C}_\mathfrak{g}$ into a monoidal category with unit object $\mathbb{C}[U]=\mathbb{C}\otimes\mathbb{C}[U]$. Given objects $(A,\varphi_A)$ and $(B,\varphi_B)$ of $\mathcal{C}_\mathfrak{g}$, $(A,\varphi_A)\diamond(B,\varphi_B)$ is easily observed to be isomorphic to the quotient object \[(A\!\!\!\underset{\mathbb{C}[U]}{\otimes}\! \!\! B,\pi\circ(1\otimes \varphi_B))\cong(A\!\!\!\underset{\mathbb{C}[U]}{\otimes}\! \!\! B,\pi\circ(\varphi_A\otimes 1))\] where $A\!\!\!\underset{\mathbb{C}[U]}{\otimes}\! \!\! B$ is the quotient of the $\mathfrak{g}$-module algebra $A\otimes B$ by the ideal generated by all elements of the form $\varphi_A(x)\otimes 1-1\otimes \varphi_B(x)$ for $x\in \mathbb{C}[U]$ and $\pi:A\otimes B\to A\!\!\!\underset{\mathbb{C}[U]}{\otimes}\! \!\! B$ is the quotient map. \end{remark} This natural structure also results in a very easy proof of the ``only if" part of Corollary \ref{cor:highest}. Denote by $\mathbb{C}[T]$ the algebra with basis $\{v_\lambda\ |\ \lambda\in \Lambda\}$ (where $v_0=1$) and multiplication $v_\lambda v_\mu=v_{\lambda+\mu}$. We make it into a $\mathfrak{b}_-$-module algebra with $\mathfrak{b}_-$-module structure given by $h_i(v_\lambda)=(\alpha_i^\vee,\lambda)v_\lambda$ and $f_i(v_\lambda)=0$. Suppose $\mu:A^+\otimes \mathbb{C}[U]\to A$ as in Theorem \ref{thm:factoring} is an isomorphism. We give $A^+$ the structure of a $\mathfrak{b}_-$-module algebra as in Theorem \ref{thm:b-}. Then Theorem \ref{thm:g} allows us to make $(A^+\otimes \mathbb{C}[T])\otimes \mathbb{C}[U]$ into a $\mathfrak{g}$-module algebra. Now $A$ is clearly a $\mathfrak{g}$-module subalgebra and \begin{align*} ((1\otimes v_{-\omega_i})\otimes 1)\left[f_i\rhd ((1\otimes v_{\omega_i})\otimes 1)\right]&=((1\otimes v_{-\omega_i})\otimes 1)\left((1\otimes h_i(v_{\omega_i}))\otimes x_i\right)\\ &=(\alpha_i^\vee, \omega_i)(1\otimes v_{-\omega_i}v_{\omega_i}\otimes x_i)\\ &=1\otimes 1\otimes x_i \end{align*} showing that $(A^+\otimes \mathbb{C}[T])\otimes \mathbb{C}[U]$ is generated by $(A^+\otimes \mathbb{C}[T])\otimes \mathbb{C}=((A^+\otimes \mathbb{C}[T])\otimes \mathbb{C}[U])^+$ as a $\mathfrak{g}$-module algebra and proving the ``only if" part of Corollary \ref{cor:highest}. \section{Proofs} \label{sect:Proofs} In many proofs, we will use the fact that $\mathbb{Z}_{\ge 0}^m$ is well-ordered by the lexicographic order. For given $w\in W$, ${\bf i}\in R(w)$, and $U_q(\mathfrak{b}_+)$-module $M$, we have that $\nu_{\bf i}(M)$ is well-ordered, allowing us to induct on $\nu_{\bf i}(x)$ for $x\in M$. \subsection{Proof of Theorem \ref{thm:qfactoring}} \label{pf:qfactoring} For ${\bf j}\in \nu_{\bf i}(\mathcal{B})$, let $b_{\bf j}$ be the unique element of $\mathcal{B}$ such that $\nu_{\bf i}(b_{\bf j})={\bf j}$. \begin{enumerate} \item We first observe that since $A$ is a $U_q(\mathfrak{g})$-module algebra and $A^+$ and $A_0$ are $U_q(\mathfrak{b}_+)$-submodules, $\mu$ is a homomorphism of $U_q(\mathfrak{b}_+)$-modules. Hence we simply show that $\mu$ is injective. Now each nonzero element $a\in A^+\otimes A_0$ can be written \[a=\sum_{k=1}^n a_k\otimes b_{{\bf j}_k}\] for some $n>0$, $a_k\in A^+\setminus \{0\}$, and ${\bf j}_k\in \nu_{\bf i}(A_0\setminus\{0\})$. We may assume ${\bf j}_k<{\bf j}_{l}$ if $1\le k<l\le n$ so that \[E_{\bf i}^{{\bf j}_l}(b_{{\bf j}_k})=\begin{cases} 0 & \text{if } k<l\\ 1 & \text{if } k=\ell\end{cases}.\] Suppose for the sake of contradiction that $\mu(a)=0$. Then $$ 0=E_{\bf i}^{({\bf j}_n)}(\mu(a))=\mu\left(E_{\bf i}^{({\bf j}_n)}(a)\right)=\mu(a_n\otimes 1)=a_n$$ which is a contradiction. Hence $\mu(a)=0$ if and only if $a=0$, showing that $\mu$ is injective. \item ($\Rightarrow$) Suppose $\mu$ is an isomorphism. Given nonzero $a\in A$, write \[a=\mu\left(\sum_{k=1}^n a_k\otimes b_{{\bf j}_k}\right)\] as in (1). Then, since $\mu$ is injective, it is clear that $\nu_{\bf i}(a)={\bf j}_n=\nu_{\bf i}(b_{{\bf j}_n})$, showing that $\nu_{\bf i}(A\setminus\{0\})\subseteq\nu_{\bf i}(A_0\setminus\{0\})$. But since $A_0\subseteq A$, it follows that $\nu_{\bf i}(A\setminus\{0\})=\nu_{\bf i}(A_0\setminus\{0\})$. Also, as seen above \[E_{\bf i}^{(top)}(a)=E_{\bf i}^{({\bf j}_n)}(a)=\mu(a_n\otimes 1)=a_n\in A^+,\] so we see that $A$ is ${\bf i}$-adapted.\\ ($\Leftarrow$) Suppose that $A$ is {\bf i}-adapted and $\nu_{\bf i}(A\setminus\{0\})=\nu_{\bf i}(A_0\setminus\{0\})$. By (1), we already know that $\mu$ is an injective $U_q(\mathfrak{b}_+)$-module homomorphism. Hence we simply use induction to show that $\mu$ is surjective. We first note that since $A$ is {\bf i}-adapted, if $\nu_{\bf i}(a)=(0,0,\ldots,0)$, then $a=E_{\bf i}^{(top)}(a)\in A^+$. In other words, \[\{a\in A\setminus\{0\}\ |\ \nu_{\bf i}(a)=(0,0,\ldots,0)\}=A^+\setminus \{0\}\subset \mu(A^+\otimes A_0).\] Let $a\in A\setminus\{0\}$ and suppose $a'\in \mu(A^+\otimes A_0)$ for all $a'\in A\setminus\{0\}$ such that $\nu_{\bf i}(a')<\nu_{\bf i}(a)$. We have \[E_{\bf i}^{(\nu_{\bf i}(a))}(a-\mu(E_{\bf i}^{(top)}(a)\otimes b_{\nu_{\bf i}(a)}))=0.\] Hence either $a-\mu(E_{\bf i}^{(top)}(a)\otimes b_{\nu_{\bf i}(a)})=0$ or $\nu_{\bf i}(a-\mu(E_{\bf i}^{(top)}(a)\otimes b_{\nu_{\bf i}(a)}))<\nu_{\bf i}(a)$. In the former case, $a\in \mu(A^+\otimes A_0)$. In the latter case, $a-\mu(E_{\bf i}^{(top)}(a)\otimes b_{\nu_{\bf i}(a)})\in \mu(A^+\otimes A_0)$ and so \[a=(a-\mu(E_{\bf i}^{(top)}(a)\otimes b_{\nu_{\bf i}(a)}))+\mu(E_{\bf i}^{(top)}(a)\otimes b_{\nu_{\bf i}(a)})\in \mu(A^+\otimes A_0).\] So we have shown that $a\in \mu(A^+\otimes A_0)$. By induction, $\mu$ is surjective. Hence $\mu$ is an isomorphism. \end{enumerate}\qed \subsection{Proof of Theorem \ref{thm:qclassify}} \label{pf:qclassify} $(\Leftarrow)$ Suppose $A_0$ possesses an {\bf i}-adapted basis and $\nu_{\bf i}(A_0\setminus\{0\})=\nu_{\bf i}(A\setminus\{0\})$. Then by Theorem \ref{thm:qfactoring}, $\mu:A^+\otimes A_0\to A$ is an isomorphism of $U_q(\mathfrak{b}_+)$-modules. $(\Rightarrow)$ Suppose $\mu:A^+\otimes A_0\to A$ is an isomorphism of $U_q(\mathfrak{b}_+)$-modules. Hence we must have $(A_0)^+=\mathbb{C}(q)$ or else $\mu$ would fail to be injective. Also, since $A$ is {\bf i}-adapted, $A_0$ is as well. Now for each ${\bf j}\in \nu_{\bf i}(A_0\setminus\{0\})$ choose $b_{\bf j}\in A_0\setminus\{0\}$ such that $E_{\bf i}^{(top)}(b_{\bf j})=1$ and $\nu_{\bf i}(b_{\bf j})={\bf j}$ (note that $b_{(0,\cdots,0)}=1$). We claim that $\mathcal{B}=\{b_{\bf j}\ |\ {\bf j}\in \nu_{\bf i}(A_0\setminus \{0\})\}$ is an ${\bf i}$-adapted basis for $A_0$. To prove that $\mathcal{B}$ is linearly independent and spans $A_0$, we mimic the proofs that $\mu$ is injective and surjective (respectively) in Theorem \ref{thm:qfactoring}. Suppose \[\sum_{k=1}^n r_kb_{{\bf j}_k}=0\] for some $r_k\in \mathbb{C}(q)$ and ${\bf j}_k\in\nu_{\bf i}(A_0\setminus\{0\})$ such that ${\bf j}_k<{\bf j}_l$ if $k<l$. Then \[0=E_{\bf i}^{({\bf j}_n)}\left(\sum_{k=1}^n r_kb_{{\bf j}_k}\right)=r_n.\] By induction, each $r_k=0$. It follows that $\mathcal{B}$ is linearly independent. Note that \[\{a\in A_0\setminus \{0\}\ |\ \nu_{\bf i}(a)=(0,0,\ldots,0)\}=(A_0)^+\setminus\{0\}=\mathbb{C}(q)^{\times}\subseteq \text{span}_{\mathbb{C}(q)}(\mathcal{B}).\] Let $a\in A_0\setminus\{0\}$ and suppose $a'\in \text{span}_{\mathbb{C}(q)}(\mathcal{B})$ for all $a'\in A_0\setminus\{0\}$ such that $\nu_{\bf i}(a')<\nu_{\bf i}(a)$. We have \[E_{\bf i}^{(\nu_{\bf i}(a))}(a-E_{\bf i}^{(top)}(a)b_{\nu_{\bf i}(a)})=0.\] Hence either $a-E_{\bf i}^{(top)}(a)b_{\nu_{\bf i}(a)}=0$ or $\nu_{\bf i}(a-E_{\bf i}^{(top)}(a)b_{\nu_{\bf i}(a)})<\nu_{\bf i}(a)$. In the former case $a\in \text{span}_{\mathbb{C}(q)}(\mathcal{B})$. In the latter case $a-E_{\bf i}^{(top)}(a)b_{\nu_{\bf i}(a)}\in \text{span}_{\mathbb{C}(q)}(\mathcal{B})$ and so \[a=(a-E_{\bf i}^{(top)}(a)b_{\nu_{\bf i}(a)})+E_{\bf i}^{(top)}(a)b_{\nu_{\bf i}(a)}\in \text{span}_{\mathbb{C}(q)}(\mathcal{B}).\] So we have shown that $a\in \text{span}_{\mathbb{C}(q)}(\mathcal{B})$. By induction, $\mathcal{B}$ spans $A_0$. Hence we have shown that $\mathcal{B}$ is a basis for $A_0$. By construction, it is in fact an {\bf i}-adapted basis for $A_0$. In light of $\mathcal{B}$'s existence, a typical element of $A$ is of the form $\mu\left(\sum\limits_{k=1}^n a_k\otimes b_{{\bf j}_k}\right)$ for some $a_k\in A^+$ and ${\bf j}_k\in \nu_{\bf i}(A_0\setminus\{0\})$. It is now clear that $\nu_{\bf i}\left(\mu\left(\sum\limits_{k=1}^n a_k\otimes b_{{\bf j}_k}\right)\right)=\max\{{\bf j}_k\ |\ k=1,\ldots,n\}$ so that $\nu_{\bf i}(A_0\setminus\{0\})\supseteq \nu_{\bf i}(A\setminus\{0\})$. Since $A_0\subseteq A$, we have $\nu_{\bf i}(A_0\setminus\{0\})\subseteq\nu_{\bf i}(A\setminus\{0\})$ and so $\nu_{\bf i}(A_0\setminus\{0\})=\nu_{\bf i}(A\setminus\{0\})$. \qed \subsection{Proof of Proposition \ref{prop:adapted}} \label{pf:adapted} We have already observed that for any ${\bf i}\in R(w_\mathrm{o})$, the restriction of $\nu_{\bf i}$ to $\mathcal{B}^{dual}$ is an injective map $\mathcal{B}^{dual}\hookrightarrow \mathbb{Z}_{\ge 0}^m$, where $m$ is the length of $w_\mathrm{o}$. Hence it suffices to show that $E_{\bf i}^{(top)}(b)=1$ for all ${\bf i}\in R(w_\mathrm{o})$ and $b\in \mathcal{B}^{dual}$. To do this we need the following lemma. \begin{lemma} \label{lem:ewtop} Given $w\in W$ and ${\bf i, i'}\in R(w)$, $E_{\bf i}^{(top)}(b)=E_{\bf i'}^{(top)}(b)$ for all $b\in \mathcal{B}^{dual}$. \end{lemma} \begin{proof} Now $\mathbb{C}_q[U]$ factors as the product of two subalgebras: $\mathbb{C}_q[U]=(\mathbb{C}_q[U]_{>w})(\mathbb{C}_q[U]_{\le w})$ (see \cite{kimura}, for example, where they are respectively denoted ${\bf U}_q^-(>w,-1)$ and ${\bf U}_q^-(\le w,-1)$). In fact, these subalgebras can be described explicitly as follows. For any reduced word $\mathbf{i}\in R(w_\mathrm{o})$ such that $s_{i_1}\cdots s_{i_k}=w$, consider elements $X_1,\cdots,X_m$ as in \cite[Section 4]{berensteingreenstein}, where $m$ is the length of $w_\mathrm{o}$. This choice guarantees that monomials $X^a=X_1^{a_1}\cdots X_m^{a_m}$ for $a\in \mathbb{Z}_{\ge 0}^m$ form a basis for $\mathbb{C}_q[U]$. It follows that those $X^a$ with $a_\ell=0$ for $\ell>k$ form a basis for $\mathbb{C}_q[U]_{\le w}$ and those $X^a$ with $a_\ell=0$ for $\ell\le k$ form a basis for $\mathbb{C}_q[U]_{>w}$. Since $X_1=x_{i_1}$ and these two subalgebras are orthogonal with respect to Lusztig's pairing (under which multiplication by $x_i$ and action by $E_i$ are adjoint), we obtain the following well-known fact: \[E_{i}(\mathbb{C}_q[U]_{>w})=0\] for any $i\in I$ such that $\ell(s_iw)<\ell(w)$. In particular, this implies that \[E_{i}(\mathbb{C}_q[U]_{>w_{i,j}})=E_{j}(\mathbb{C}_q[U]_{>w_{i,j}})=0,\] where $w_{i,j}$ is the longest element in the subgroup generated by $s_i$ and $s_j$. It is well-known that any two reduced words for a fixed $w\in W$ are related by a series of rank two relations. Hence it suffices to show the lemma when ${\bf i}$ and ${\bf i'}$ differ by a single rank two relation. But it is also well-known that $E_j^{(top)}(b)\in \mathcal{B}^{dual}$ for all $j\in I$ and $b\in \mathcal{B}^{dual}$. For any ${\bf j}\in R(w)$, the operator $E_{\bf j}^{(top)}$ is by definition just the composition of operators $E_{j_\ell}^{(top)}\cdots E_{j_2}^{(top)}E_{j_1}^{(top)}$, where $\ell$ is the length of $w$. This reduces the problem to the case when $w$ is the longest element of a rank two parabolic subgroup of $W$. We will therefore assume for the rest of the proof that ${\bf i}=(i,j,\ldots)$ and ${\bf i'}=(j,i,\ldots)$ the only two distinct reduced words for $w_{i,j}$. An explicit (and apparently well-known) computation verifies that \[E_{\bf i}^{(top)}(\mathbb{C}_q[U]_{\le w_{i,j}}\cap \mathcal{B}^{dual})=E_{\bf i'}^{(top)}(\mathbb{C}_q[U]_{\le w_{i,j}}\cap \mathcal{B}^{dual})=1.\] According to (\cite{kimura}, Theorem 3.14), for each $b\in \mathcal{B}^{dual}$, there exist $b'\in \mathbb{C}_q[U]_{>w_{i,j}}\cap \mathcal{B}^{dual}$, $b''\in \mathbb{C}_q[U]_{\le w_{i,j}}\cap \mathcal{B}^{dual}$, and $\xi\in \mathbb{C}_q[U]$ such that $\nu_{\bf i}(\xi)<\nu_{\bf i}(b)$, $\nu_{\bf i'}(\xi)<\nu_{\bf i'}(b)$, and \[b'b''=b+\xi.\] Hence $$ E_{\bf i}^{(top)}(b)=E_{\bf i}^{(top)}(b+\xi)=E_{\bf i}^{(top)}(b'b'')=b'E_{\bf i}^{(top)}(b'')=b'.$$ Likewise, $E_{\bf i'}^{(top)}(b)=b'$, so the lemma is proved. \end{proof} In light of Lemma \ref{lem:ewtop}, given $w\in W$ and $b\in \mathcal{B}^{dual}$, we may unambiguously define $E_w^{(top)}(b):=E_{\bf i}^{(top)}(b)$ for any ${\bf i}\in R(w)$. Now given $j\in I$, there exists some ${\bf i}\in R(w_\mathrm{o})$ such that if ${\bf i}=(i_1,\ldots,i_m)$, then $i_m=j$. It follows that $E_j(E_{w_\mathrm{o}}^{(top)}(b))=0$ for all $j\in I$, i.e. $E_{w_\mathrm{o}}^{(top)}(b)\in (\mathbb{C}_q[U])^+=\mathbb{C}(q)$. Since $E_j^{(top)}(b)\in \mathcal{B}^{dual}$ for all $j\in I$ and $b\in \mathcal{B}^{dual}$, it follows that $E_{w_\mathrm{o}}^{(top)}(b)=1$ for $b\in \mathcal{B}^{dual}$. \qed \subsection{Proof of Corollary \ref{cor:qhighest}} \label{pf:qhighest} Let ${\bf i}\in R(w_\mathrm{o})$. As previously remarked, $\mathbb{C}_q[U]$ possesses an {\bf i}-adapted basis. Then Theorem \ref{thm:qfactoring} (1) says that $\mu':(A')^+\otimes \mathbb{C}_q[U]\to A'$ is an injective homomorphism of $U_q(\mathfrak{b}_+)$-modules. We now show that $\mu'((A')^+\otimes \mathbb{C}_q[U])$ is a $U_q(\mathfrak{g})$-module subalgebra of $A'$ and hence is equal to $A'$ by the assumption that $(A')^+$ generates $A'$. To see that $\mu'((A')^+\otimes \mathbb{C}_q[U])$ is a subalgebra of $A'$, it suffices to show that $x_ia\in \mu'((A')^+\otimes \mathbb{C}_q[U])$ for $i\in I$ and $a\in (A')^+$. For this, we observe that \[E_i(x_ja-K_j(a)x_j)=\delta_{ij}K_i(a)-\delta_{ij}K_j(a)=0\] and hence $x_ja-K_j(a)x_j\in (A')^+$. Then $$ x_ia=K_i(a)x_i+(x_ia-K_i(a)x_i)\\ =\mu'(K_i(a)\otimes x_i+(x_ia-K_i(a)x_i)\otimes 1)\\ \in \mu'((A')^+\otimes \mathbb{C}_q[U]).$$ So $\mu'((A')^+\otimes \mathbb{C}_q[U])$ is a subalgebra of $A'$. Now we need to show that $\mu'((A')^+\otimes \mathbb{C}_q[U])$ is closed under the action of $U_q(\mathfrak{g})$. By Theorem \ref{thm:qfactoring} (1), $\mu'$ is $U_q(\mathfrak{b}_+)$-invariant and hence it suffices to show that $\mu'((A')^+\otimes\mathbb{C}_q[U])$ is closed under the action of $F_i$ for $i\in I$. Observe that for $a\in (A')^+$ and $x\in \mathbb{C}_q[U]$, $\mu'(a\otimes x)=ax$, so we will simply compute the action of $F_i$ on such an element. However, before doing so, we note that for $a\in (A')^+$ and $i,j\in I$, we have \[E_i\left(F_j(a)+x_j\frac{K_j^{-2}(a)-a}{q_j-q_j^{-1}}\right)=\delta_{ij}\frac{K_i(a)-K_i^{-1}(a)}{q_i-q_i^{-1}}+\delta_{ij}\frac{K_i^{-1}(a)-K_i(a)}{q_i-q_i^{-1}}=0,\] showing that $F_j(a)+x_j\dfrac{K_j^{-2}(a)-a}{q_i-q_i^{-1}}\in (A')^+$. Now we compute: \begin{align*} F_i(ax)&=F_i(a)x+K_i^{-1}(a)F_i(x)\\ &=\left(F_i(a)+x_i\frac{K_i^{-2}(a)-a}{q_i-q_i^{-1}}\right)x-x_i\frac{K_i^{-2}(a)-a}{q_i-q_i^{-1}}x+K_i^{-1}(a)F_i(x)\\ &=\mu'\left(\left(F_i(a)+x_i\frac{K_i^{-2}(a)-a}{q_i-q_i^{-1}}\right)\otimes x+K_i^{-1}(a)\otimes F_i(x)\right)-\mu'(1\otimes x_i)\mu'\left(\frac{K_i^{-2}(a)-a}{q_i-q_i^{-1}}\otimes x\right)\\ &\in \mu'((A')^+\otimes \mathbb{C}_q[U]). \end{align*} So $\mu'((A')^+\otimes \mathbb{C}_q[U])$ is closed under the action of $U_q(\mathfrak{g})$ and we may conclude that $\mu'((A')^+\otimes \mathbb{C}_q[U])=A'$. Hence $\mu'$ is an isomorphism. By Theorem \ref{thm:qfactoring} (2), this implies that $A'$ is {\bf i}-adapted and $\nu_{\bf i}(A'\setminus\{0\})=\nu_{\bf i}(\mathbb{C}_q[U]\setminus\{0\})$. We deduce that $A$ is {\bf i}-adapted and $\nu_{\bf i}(\mathbb{C}_q[U]\setminus\{0\})\subseteq\nu_{\bf i}(A\setminus\{0\})\subseteq\nu_{\bf i}(A'\setminus\{0\})=\nu_{\bf i}(\mathbb{C}_q[U]\setminus\{0\})$, i.e. $\nu_{\bf i}(\mathbb{C}_q[U]\setminus\{0\})=\nu_{\bf i}(A\setminus\{0\})$. Hence applying Theorem \ref{thm:qfactoring} (2) again, $\mu$ is an isomorphism. \qed \subsection{Proof of Theorem \ref{thm:qG/U}} \label{pf:qG/U} It is well-known (see, e.g., \cite[Section 6.1]{berensteinrupel}) that $\mathbb{C}_q[B]:=U_q(\mathfrak{b}_+)^*$ is naturally a $U_q(\mathfrak{g})$-module algebra and its $U_q(\mathfrak{g})$-module subalgebra generated by all highest weight vectors $v_\lambda$ for $\lambda\in \Lambda^+$ is isomorphic as a $U_q(\mathfrak{g})$-module to the direct sum of all simple $V_\lambda$ for $\lambda\in \Lambda^+$ and thus identifies with the $q$-deformation $\mathbb{C}_q[G/U]$ of $\mathbb{C}[G/U]$. By the construction, the highest weight vectors satisfy $v_\lambda v_\mu=v_{\lambda+\mu}$ in $\mathbb{C}_q[B]$. Before showing the factorizability of $\mathbb{C}_q[G/U]$ after localization, we recall the definition of right Ore sets (which allow for Ore localizations and are sometimes also called right denominator sets) for the reader's convenience. \begin{definition} Let $R$ be any unital ring. A submonoid $\mathcal{S}\subset R\setminus \{0\}$ is called a \emph{right Ore set} if the following conditions are satisfied for $r\in R$ and $s\in \mathcal{S}$: \begin{enumerate} \item $r\mathcal{S}\cap sR\ne \varnothing$. \item If $sr=0$, then $\exists s'\in \mathcal{S}$ such that $rs'=0$. \end{enumerate} \end{definition} Recall (see, e.g., \cite{jordan}) that an element $p$ of a ring $R$ is \emph{normal} if $pR=Rp$. It is immediate (and well-known) that for any ring $R$, any submonoid $\mathcal{S}\subset R\setminus \{0\}$ consisting of normal elements that aren't zero-divisors is automatically both right and left Ore. In what follows, we will refer to these as \emph{normal} Ore sets. In particular, $\mathcal{S}=\{v_{\lambda}\ |\ \lambda\in \Lambda^+\}\subset \mathbb{C}_q[G/U]$ is a normal Ore set and the Ore localization $(\mathbb{C}_q[G/U])[\mathcal{S}^{-1}]$ is isomorphic to $\mathbb{C}_q[B]$ as $U_q(\mathfrak{g})$-module algebras. The following lemmas allow us to create normal Ore sets in the $n$-fold braided tensor product $\mathbb{C}_q[G/U]^{\underline{\otimes}n}$. \begin{lemma} \label{lem:ore} Let $\Bbbk$ be any field and suppose $A$ and $B$ are $\Bbbk$-algebras such that the $\Bbbk$-vector space $A\otimes_\Bbbk B$ has the structure of a $\Bbbk$-algebra satisfying $$ (a\otimes 1)(a'\otimes b')=aa'\otimes b',~(a\otimes b)(1\otimes b')=a\otimes bb' $$ for $a,a'\in A$ and $b,b'\in B$. If $\mathcal{S}$ is a normal Ore set in $B$ such that \[(1\otimes s)((A\setminus \{0\})\otimes 1)=((A\setminus\{0\})\otimes 1)(1\otimes s)\] for $s\in \mathcal{S}$, then $1\otimes \mathcal{S}:=\{1\otimes s\ |\ s\in \mathcal{S}\}$ is a normal Ore set in $A\otimes_\Bbbk B$. \end{lemma} \begin{proof} It is clear that $1\otimes \mathcal{S}$ is a multiplicative set containing $1\otimes 1$ and that $(1\otimes s)(A\otimes_\Bbbk B)=(A\otimes_\Bbbk B)(1\otimes s)$ for $s\in \mathcal{S}$, so we simply show that $1\otimes \mathcal{S}$ does not contain any zero-divisors. Fix $s\in \mathcal{S}$. Now an arbitrary nonzero element $x\in A\otimes_\Bbbk B$ can be written in the form $x=\sum\limits_{k=1}^n a_k\otimes b_k$ for some $a_k\in A\setminus \{0\}$ and $b_k\in B\setminus\{0\}$. We may assume that $\{b_k\}_{k=1}^n$ is a linearly independent set. Since $s$ is not a zero-divisor in $B$, it follows that $\{sb_k\}_{k=1}^n$ is a linearly independent set, as is $\{b_ks\}_{k=1}^n$. Also, by assumption, for each $k=1,\ldots, n$, there exists $a_k'\in A\setminus\{0\}$ such that $(1\otimes s)(a_k\otimes 1)=a_k'\otimes s$. Then $$(1\otimes s)x=(1\otimes s)\left(\sum_{k=1}^n a_k\otimes b_k\right)=\sum_{k=1}^n a_k'\otimes sb_k\ne 0,$$ $$x(1\otimes s)=\left(\sum_{k=1}^n a_k\otimes b_k\right)(1\otimes s)=\sum_{k=1}^na_k\otimes b_ks\ne 0.$$ Since $x$ was an arbitrary element of $A\otimes_\Bbbk B$, we have shown that $1\otimes s$ is not a zero-divisor in $A\otimes_\Bbbk B$. \end{proof} \begin{lemma} \label{lem:braidore} Let $A$ and $B$ be $U_q(\mathfrak{g})$-weight module algebras and let $\mathcal{S}$ be a normal Ore set in $B$ consisting of highest weight vectors. Then $1\otimes \mathcal{S}$ is a normal Ore set in the braided tensor product $A\underline{\otimes} B$. \end{lemma} \begin{proof} In light of Lemma \ref{lem:ore}, it suffices to show that $(1\otimes s)((A\setminus \{0\})\otimes 1)=((A\setminus\{0\})\otimes 1)(1\otimes s)$ for $s\in \mathcal{S}$. Since $\mathcal{S}$ consists of highest weight vectors in $B$, we have the commutation relation \[(1\otimes s)(a\otimes 1)=q^{(|a|,|s|)} a\otimes s\] for weight vectors $a\in A$ and $s\in \mathcal{S}$ of weight $|a|$ and $|s|$, respectively. Let us denote $q_{s,a}:=q^{(|a|,|s|)}$. Now an arbitrary nonzero element $a\in A\setminus\{0\}$ is of the form $\sum_{k=1}^n a_k$, where each $a_k\in A\setminus\{0\}$ is a weight vector. We may assume $|a_k|\ne |a_l|$ if $k\ne l$. Then for $s\in \mathcal{S}$, \[\sum_{k=1}^n q_{s,a_k}a_k\ne 0\quad \text{ and } \quad \sum_{k=1}^n q_{s,a_k}^{-1}a_k\ne 0.\] Therefore since \begin{align*} (1\otimes s)\left(\left(\sum_{k=1}^n a_k\right)\otimes 1\right)&=\left(\left(\sum_{k=1}^n q_{s,a_k}a_k\right)\otimes 1\right)(1\otimes s)\quad \text{ and }\\ \left(\left(\sum_{k=1}^n a_k\right)\otimes 1\right)(1\otimes s)&=(1\otimes s)\left(\left(\sum_{k=1}^n q_{s,a_k}^{-1}a_k\right)\otimes 1\right), \end{align*} it follows that $(1\otimes s)((A\setminus \{0\})\otimes 1)=((A\setminus\{0\})\otimes 1)(1\otimes s)$ for $s\in \mathcal{S}$ and so the lemma is proven. \end{proof} By Lemma \ref{lem:braidore} and induction, $\mathcal{S}':=1\otimes \cdots\otimes 1\otimes \mathcal{S}$ is a normal Ore set in $\mathbb{C}_q[G/U]^{\underline{\otimes}n}$. Furthermore, it is clear that \[\mathbb{C}_q[G/U]^{\underline{\otimes}n}[\mathcal{S}'^{-1}]\cong \mathbb{C}_q[G/U]^{\underline{\otimes}(n-1)}\underline{\otimes} \mathbb{C}_q[B]\] as $U_q(\mathfrak{g})$-module algebras and $\mathbb{C}_q[G/U]^{\underline{\otimes}(n-1)}\underline{\otimes}\mathbb{C}_q[B]$ is generated by $(\mathbb{C}_q[G/U]^{\underline{\otimes}(n-1)}\underline{\otimes}\mathbb{C}_q[B])^+$ as a $U_q(\mathfrak{g})$-module algebra. We now have an embedding of $U_q(\mathfrak{g})$-module algebras \[\mathbb{C}_q[U]\hookrightarrow \mathbb{C}_q[B]\subset\mathbb{C}_q[G/U]^{\underline{\otimes}(n-1)}\underline{\otimes}\mathbb{C}_q[B].\] Then by Corollary \ref{cor:qhighest}, $\mathbb{C}_q[G/U]^{\underline{\otimes}(n-1)}\underline{\otimes}\mathbb{C}_q[B]\cong (\mathbb{C}_q[G/U]^{\underline{\otimes} n})[\mathcal{S}'^{-1}]$ is factorizable over $\mathbb{C}_q[U]$.\qed \subsection{Proofs of Theorems \ref{thm:qgd} and \ref{thm:qg}} \label{pf:qgd,qg} Let $H$ be a Hopf algebra with invertible antipode (e.g. $H=\mathcal{K}$). We will refer to Yetter-Drinfeld modules of various kinds: ${}_H^H\mathcal{YD}$, ${}_H\mathcal{YD}^H$, and $\mathcal{YD}_H^H$ (see, e.g., \cite[Section 2]{cwy}). The side of the subscript denotes the side on which $H$ will act, while the side of the superscript denotes the side on which $H$ will coact. We use sumless Sweedler notation to write left coactions $x\mapsto x^{(-1)}\otimes x^{(0)}$ and right coactions $x\mapsto x^{(0)}\otimes x^{(1)}$. To distinguish the structure maps of a Nichols algebra (a Hopf algebra in the appropriate Yetter-Drinfeld category, see for example \cite{bazlov}) from those of $H$, we underline them. For instance, we write the braided comultiplication $\underline{\Delta}(b)=\underline{b}_{(1)}\otimes\underline{b}_{(2)}$. We start with some results that will play key roles in the proofs of the Theorems \ref{thm:qgd} and \ref{thm:qg}. \begin{theorem} \label{thm:adjoint} Let $A$ be a left $H$-module algebra and suppose $V\in {}_H^H\mathcal{YD}$ is such that the Nichols algebra $\mathcal{B}(V)$ is a left $H$-module subalgebra of $A$, where ${}_H^H\mathcal{YD}$ is the category of left-left Yetter-Drinfeld modules over $H$. Then $A$ can be given a left $\mathcal{B}(V)$-module structure via \[v\rhd a=va-(v^{(-1)}(a))v^{(0)}.\] \end{theorem} \begin{proof} Consider the Hopf algebra $\widetilde{H}:=\mathcal{B}(V)\rtimes H$, where $\Delta(u)=\underline{u}_{(1)}(\underline{u}_{(2)})^{(-1)}\otimes (\underline{u}_{(2)})^{(0)}$ and $S(u)=S(u^{(-1)})\underline{S}(u^{(0)})$ for $u\in \mathcal{B}(V)$. Then $\widetilde{H}$ can naturally be considered as a subalgebra of $\widetilde{A}:=A\rtimes H$. Hence $\widetilde{A}$ is an $\widetilde{H}$-module algebra under the adjoint action: \[\widetilde{h}\rhd \widetilde{a}=\widetilde{h}_{(1)}\widetilde{a} S(\widetilde{h}_{(2)}).\] We observe that $A$ is preserved under the restriction of this action to $\mathcal{B}(V)$ (note that for convenience we will write, e.g., $a$ instead of $a\otimes 1$): \begin{align*} u\rhd a&=u_{(1)}aS(u_{(2)})\\ &=\underline{u}_{(1)}(\underline{u}_{(2)})^{(-1)}aS(((\underline{u}_{(2)})^{(0)})^{(-1)})\underline{S}(((\underline{u}_{(2)})^{(0)})^{(0)})\\ &=\underline{u}_{(1)}(\underline{u}_{(2)})^{(-2)}aS((\underline{u}_{(2)})^{(-1)})\underline{S}((\underline{u}_{(2)})^{(0)})\\ &=\underline{u}_{(1)}((\underline{u}_{(2)})^{(-3)}(a))(\underline{u}_{(2)})^{(-2)}S((\underline{u}_{(2)})^{(-1)})\underline{S}((\underline{u}_{(2)})^{(0)})\\ &=\underline{u}_{(1)}((\underline{u}_{(2)})^{(-2)}(a))\varepsilon((\underline{u}_{(2)})^{(-1)})\underline{S}((\underline{u}_{(2)})^{(0)})\\ &=\underline{u}_{(1)}((\underline{u}_{(2)})^{(-1)}(a))\underline{S}((\underline{u}_{(2)})^{(0)})\in A \end{align*} for $u\in \mathcal{B}(V)$ and $a\in A$. In fact, it is clear that $A$ has become a left $\widetilde{H}$-module algebra. Now computing the given action for $v\in V$ and $a\in A$, we find $$ v\rhd a=\underline{v}_{(1)}((\underline{v}_{(2)})^{(-1)}(a))\underline{S}((\underline{v}_{(2)})^{(0)}) =va+v^{(-1)}(a)\underline{S}(v^{(0)}) =va-v^{(-1)}(a)v^{(0)},$$ as required. The second and third equalities follow from the fact that every element of $V$ is a primitive element of the braided Hopf algebra $\mathcal{B}(V)$. \end{proof} Of course, Theorem \ref{thm:adjoint} has a natural counterpart with ``left" replaced by ``right". \begin{theorem} \label{thm:op} Let $A$ be a right $H$-module algebra and suppose $V\in \mathcal{YD}_H^H$ is such that the Nichols algebra $\mathcal{B}(V)$ is a right $H$-module subalgebra of $A$. Then $A$ can be given a right $\mathcal{B}(V)$-module structure via \[a\blacktriangleleft v= av-v^{(0)}((a) v^{(1)}).\] \end{theorem} Given any ring $R$, a right $R$-module is naturally a left $R^{op}$-module, giving us the following obvious corollary. \begin{corollary} \label{cor:op} In the assumptions of Theorem \ref{thm:op}, if $H$ is commutative, then $A$ can be given a left $\mathcal{B}(V)^{op}$-module structure via \[v\blacktriangleright a=av-v^{(0)}(v^{(1)}(a)).\] \end{corollary} \begin{remark} If $A$ is a $(\mathcal{B}(V)\rtimes H)$-module algebra (e.g. Theorem \ref{thm:adjoint}), then we can form the \emph{braided cross product} $A\underline{\rtimes}\mathcal{B}(V)$ which, as a vector space, is just $A\otimes \mathcal{B}(V)\subset A\rtimes (\mathcal{B}(V)\rtimes H)$ and it is a subalgebra. Furthermore, it is an $H$-module algebra. We note that if $A$ is additionally a $\mathcal{B}(V)$-module algebra in ${}_H^H\mathcal{YD}$, then our definition of $A\underline{\rtimes}\mathcal{B}(V)$ matches that of $A\rtimes \mathcal{B}(V)$. However, we don't require that $A$ is even an $H$-comodule, which is why we use a different notation. Similarly, we can form the braided tensor product $A\underline{\otimes}\mathcal{B}(V)$ (which is an $H$-module algebra) even if $A$ is an $H$-module algebra and is not in ${}_H^H\mathcal{YD}$, simply satisfying $(1\otimes v)(a\otimes 1)=(v^{(-1)}\rhd a)\otimes v^{(0)}$. This corresponds to the braided cross product $A\underline{\rtimes}\mathcal{B}(V)$, where $\mathcal{B}(V)\rtimes H$ acts on $A$ by the ``trivial" action: $(u\otimes h)\rhd a=\underline{\varepsilon}(u)h\rhd a\quad \text{for }u\in \mathcal{B}(V),\ h\in H,\ a\in A$. \end{remark} \begin{theorem} \label{thm:braidedtriv} Let $V\in {}_H^H\mathcal{YD}$ and suppose $A$ is an $H$-module algebra containing $\mathcal{B}(V)$ as an $H$-module subalgebra. Then the linear map $\tau:=(\mu_A\otimes id)\circ(id\otimes \iota\otimes id)\circ (id\otimes \underline{\Delta}):A\underline{\rtimes}\mathcal{B}(V)\to A\underline{\otimes}\mathcal{B}(V)$ is an $H$-module algebra isomorphism with inverse $\tau^{-1}=(\mu_A\otimes id)\circ(id\otimes \iota\otimes id)\circ (id\otimes \underline{S}\otimes id)\circ (id\otimes \underline{\Delta})$, where $\iota: \mathcal{B}(V)\to A$ is the inclusion and the implied $\mathcal{B}(V)\rtimes H$ action on $A$ is that of Theorem \ref{thm:adjoint}. \end{theorem} \begin{proof} We first verify that $\tau$ and $\tau^{-1}$ are truly mutually inverse (and hence that we are justified in using the name $\tau^{-1}$). For $a\in A$ and $b\in \mathcal{B}(V)$, we directly compute $$(\tau\circ\tau^{-1})(a\otimes b)=\tau(a\underline{S}(\underline{b}_{(1)})\otimes \underline{b}_{(2)})=aS(\underline{b}_{(1)})\underline{b}_{(2)}\otimes \underline{b}_{(3)} =a\underline{\varepsilon}(\underline{b}_{(1)})\otimes \underline{b}_{(2)}=a\otimes\underline{\varepsilon}(\underline{b}_{(1)})\underline{b}_{(2)}=a\otimes b,$$ $$(\tau^{-1}\circ\tau)(a\otimes b)=\tau^{-1}(a\underline{b}_{(1)}\otimes \underline{b}_{(2)}) =a\underline{b}_{(1)}\underline{S}(\underline{b}_{(2)})\otimes \underline{b}_{(3)} =a\underline{\varepsilon}(\underline{b}_{(1)})\otimes \underline{b}_{(2)} =a\otimes\underline{\varepsilon}(\underline{b}_{(1)})\underline{b}_{(2)}=a\otimes b. $$ Since $\tau\circ \tau^{-1}$ and $\tau^{-1}\circ \tau$ act as the identity on pure tensors, they are both the identity homomorphism. Hence $\tau$ and $\tau^{-1}$ are mutually inverse. We conclude by verifying that $\tau$ is actually a homomorphism of algebras (and hence that $\tau^{-1}$ is as well). For $v\in V$, $a,a'\in A$, and $b\in \mathcal{B}(V)$, we have \begin{align*} \tau(a\otimes v)\tau(a'\otimes b)&=(av\otimes 1+a\otimes v)(a'\underline{b}_{(1)}\otimes \underline{b}_{(2)})\\ &=ava'\underline{b}_{(1)}\otimes \underline{b}_{(2)}+a(v^{(-1)} (a'\underline{b}_{(1)}))\otimes v^{(0)}\underline{b}_{(2)}\\ &=ava'\underline{b}_{(1)}\otimes \underline{b}_{(2)}+a(v^{(-2)} (a'))(v^{(-1)}( \underline{b}_{(1)}))\otimes v^{(0)}\underline{b}_{(2)}\\ &=ava'\underline{b}_{(1)}\otimes \underline{b}_{(2)}-a(v^{(-1)} (a'))v^{(0)}\underline{b}_{(1)}\otimes \underline{b}_{(2)}+a(v^{(-1)}(a'))v^{(0)}\underline{b}_{(1)}\otimes \underline{b}_{(2)}\\ &\hspace{0.5cm}+a(v^{(-2)}(a'))(v^{(-1)}( \underline{b}_{(1)}))\otimes v^{(0)}\underline{b}_{(2)}\\ &=\tau(ava'\otimes b-a(v^{(-1)}( a'))v^{(0)}\otimes b+a(v^{(-1)}( a'))\otimes v^{(0)}b)\\ &=\tau(a(va'-(v^{(-1)}(a'))v^{(0)})\otimes b+a(v^{(-1)}(a'))\otimes v^{(0)}b)\\ &=\tau(a(v\rhd a')\otimes b+a(v^{(-1)}(a'))\otimes v^{(0)}b)\\ &=\tau((a\otimes v)(a'\otimes b)). \end{align*} It is clear that $\{b\in \mathcal{B}(V)\ |\ \tau(a\otimes b)\tau(a'\otimes b')=\tau((a\otimes b)(a'\otimes b'))\ \forall a,a'\in A, b'\in \mathcal{B}(V)\}$ is a subalgebra of $A\underline{\rtimes}\mathcal{B}(V)$. We have shown it contains $V$, so it must be equal to $\mathcal{B}(V)$. Now, since pure tensors span $A\underline{\rtimes} \mathcal{B}(V)$ and $\tau$ is a linear map, it follows that $\tau$ respects multiplication. The theorem is proved. \end{proof} \begin{corollary} \label{cor:embed} Let $V\in {}_H^H\mathcal{YD}$. Then there are injective $H$-module algebra homomorphisms $\mathcal{B}(V)\to \mathcal{B}(V)\underline{\rtimes}\mathcal{B}(V)$ and $\mathcal{B}(V)\to \mathcal{B}(V)\underline{\otimes}\mathcal{B}(V)$ given by $v\mapsto 1\otimes v-v\otimes 1$ and $v\mapsto 1\otimes v+v\otimes 1$, respectively, for $v\in V$. \end{corollary} \begin{proof} Let $\tau$ be as in Theorem \ref{thm:braidedtriv}, where $A=\mathcal{B}(V)$. Restrict $\tau^{-1}$ to $\mathcal{B}(V)\cong 1\underline{\otimes} \mathcal{B}(V)\subset \mathcal{B}(V)\underline{\otimes}\mathcal{B}(V)$ and observe that $\tau^{-1}(1\otimes v)=1\otimes v-v\otimes 1$ for $v\in V$. Similarly, restrict $\tau$ to $\mathcal{B}(V)\cong 1\underline{\rtimes}\mathcal{B}(V)\subset \mathcal{B}(V)\underline{\rtimes}\mathcal{B}(V)$ and note that $\tau(1\otimes v)=1\otimes v+v\otimes 1$ for $v\in V$. \end{proof} \begin{theorem} \label{thm:diff} Let $V\in {}_H^H\mathcal{YD}$ and set $\widetilde{H}=\mathcal{B}(V)\rtimes H$. Let $A$ be a left $\widetilde{H}$-module algebra and suppose $A$ contains an $\widetilde{H}$-module subalgebra isomorphic to $\mathcal{B}(V)$ with the ``adjoint" action: \[(u\otimes h)\rhd u'=\underline{u}_{(1)}([(\underline{u}_{(2)})^{(-1)}h](u'))\underline{S}((\underline{u}_{(2)})^{(0)})\quad \text{for } u,u'\in \mathcal{B}(V),\ h\in H.\] Then there is a left $\mathcal{B}(V)$ action $\blacktriangleright$ on $A$ given by \[v\blacktriangleright a=(v\rhd a)-[va-(v^{(-1)}( a))v^{(0)}] \quad \text{for }v\in V,\ a\in A.\] \end{theorem} \begin{proof} We first observe that $\mathcal{B}(V)\underline{\rtimes}\mathcal{B}(V)$ is an $H$-module subalgebra of $A\underline{\rtimes}\mathcal{B}(V)$. By Corollary \ref{cor:embed}, the elements $1\otimes v-v\otimes 1\in A\underline{\rtimes}\mathcal{B}(V)$ ($v\in V$) generate an $H$-module algebra isomorphic to $\mathcal{B}(V)$. Then by Theorem \ref{thm:adjoint}, we can define an action of $\mathcal{B}(V)$ on $A\underline{\rtimes}\mathcal{B}(V)$ by \[v\cdot (a\otimes u)=(1\otimes v-v\otimes 1)(a\otimes u)-[v^{(-1)}(a\otimes u)](1\otimes v^{(0)}-v^{(0)}\otimes 1).\] Now we need only observe that this action preserves $A=A\underline{\rtimes} 1\subset A\underline{\rtimes}\mathcal{B}(V)$ and acts in the prescribed manner: \begin{align*} v\cdot (a\otimes 1)&=(1\otimes v-v\otimes 1)(a\otimes 1)-[v^{(-1)}(a\otimes 1)](1\otimes v^{(0)}-v^{(0)}\otimes 1)\\ &=(v\rhd a)\otimes 1+(v^{(-1)}( a))\otimes v^{(0)}-va\otimes 1-(v^{(-1)}( a))\otimes v^{(0)}+(v^{(-1)}( a))v^{(0)}\otimes 1\\ &=[(v\rhd a)-[va-(v^{(-1)}( a))v^{(0)}]]\otimes 1. \end{align*} \end{proof} \begin{theorem} \label{thm:sum} Let $V\in {}_H^H\mathcal{YD}$ and set $\widetilde{H}=\mathcal{B}(V)\rtimes H$. Let $A$ be a left $\widetilde{H}$-module algebra and suppose $A$ contains an $\widetilde{H}$-module subalgebra isomorphic to $\mathcal{B}(V)$ with the trivial action: \[(u\otimes h)\rhd u'=[\underline{\varepsilon}(u)h]( u')\quad \text{for } u,u'\in \mathcal{B}(V),\ h\in H.\] Then there is a left $\mathcal{B}(V)$ action $\blacktriangleright$ on $A$ given by \[v\blacktriangleright a=(v\rhd a)+[va-(v^{(-1)}( a))v^{(0)}] \quad \text{for }v\in V,\ a\in A.\] \end{theorem} \begin{proof} By Corollary \ref{cor:embed}, the elements $1\otimes v+v\otimes 1\in A\underline{\rtimes}\mathcal{B}(V)$ ($v\in V$) generate an $H$-module algebra isomorphic to $\mathcal{B}(V)$. Then by Theorem \ref{thm:adjoint}, we can define an action of $\mathcal{B}(V)$ on $A\underline{\rtimes}\mathcal{B}(V)$ by \[v\cdot (a\otimes u)=(1\otimes v+v\otimes 1)(a\otimes u)-[v^{(-1)} (a\otimes u)](1\otimes v^{(0)}+v^{(0)}\otimes 1).\] Now we need only observe that this action preserves $A=A\underline{\rtimes}1\subset A\underline{\rtimes}\mathcal{B}(V)$ and acts in the prescribed manner: \begin{align*} v\cdot (a\otimes 1)&=(1\otimes v+v\otimes 1)(a\otimes 1)-[v^{(-1)} (a\otimes 1)](1\otimes v^{(0)}+v^{(0)}\otimes 1)\\ &=(v\rhd a)\otimes 1+(v^{(-1)}( a))\otimes v^{(0)}+va\otimes 1-(v^{(-1)}( a))\otimes v^{(0)}-(v^{(-1)}( a))v^{(0)}\otimes 1\\ &=[(v\rhd a)+[va-(v^{(-1)}( a))v^{(0)}]]\otimes 1 \end{align*} \end{proof} \begin{theorem} \label{thm:commute} Let $\hat{A}$ be a left $H$-module algebra and suppose that $V_1\in {}_H^H\mathcal{YD}$ and $V_2\in {}_H\mathcal{YD}^H$ are $H$-submodules of $\hat{A}$. For $v_i\in V_i$, define the following actions on $\hat{A}$: \[v_1\rhd a=v_1a-v_1^{(-1)}(a)v_1^{(0)}\quad v_2\blacktriangleright a=v_2^{(0)}v_2^{(1)}(a)-av_2.\] If (1) $v_1\rhd v_2=v_2\blacktriangleright v_1=0$ and (2) $v_1^{(-2)}(v_2^{(0)})v_1^{(-1)}(v_2^{(1)}(a))v_1^{(0)}=v_2^{(0)}v_2^{(1)}(v_1^{(-1)}(a))v_2^{(2)}(v_1^{(0)})$ for all $v_i\in V_i$ and $a\in \hat{A}$, then $v_1\rhd (v_2\blacktriangleright a)=v_2\blacktriangleright(v_1\rhd a)$ for $v_i\in V_i$ and $a\in \hat{A}$. \end{theorem} \begin{proof} We first note that if $v_1\rhd v_2=v_2\blacktriangleright v_1=0$, then $v_1v_2=v_1^{(-1)}(v_2)v_1^{(0)}=v_2^{(0)}v_2^{(1)}(v_1)$ and for $a,b\in A$, we have $v_1\rhd(ab)=(v_1\rhd a)b+v_1^{(-1)}(a)(v_1^{(0)}\rhd b); \quad v_2\blacktriangleright(ab)=(v_2^{(0)}\blacktriangleright a)v_2^{(1)}(b)+a(v_2\triangleright b)$. Now we simply compute: \begin{align*} v_1\rhd (v_2\blacktriangleright a)&=v_1\rhd(v_2^{(0)}v_2^{(1)}(a)-av_2)\\ &=(v_1\rhd v_2^{(0)})v_2^{(1)}(a)+v_1^{(-1)}(v_2^{(0)})(v_1^{(0)}\rhd v_2^{(1)}(a))-(v_1\rhd a)v_2-v_1^{(-1)}(a)(v_1^{(0)}\rhd v_2)\\ &=v_1^{(-1)}(v_2^{(0)})(v_1^{(0)}\rhd v_2^{(1)}(a))-(v_1\rhd a)v_2\\ &=v_1^{(-1)}(v_2^{(0)})(v_1^{(0)}v_2^{(1)}(a)-(v_1^{(0)})^{(-1)}(v_2^{(1)}(a))(v_1^{(0)})^{(0)})-(v_1a-v_1^{(-1)}(a)v_1^{(0)})v_2\\ &=v_1^{(-1)}(v_2^{(0)})v_1^{(0)}v_2^{(1)}(a)-v_1^{(-2)}(v_2^{(0)})v_1^{(-1)}(v_2^{(1)}(a))v_1^{(0)}-v_1av_2+v_1^{(-1)}(a)v_1^{(0)}v_2\\\\ v_2\blacktriangleright(v_1\rhd a)&=v_2\blacktriangleright(v_1a-v_1^{(-1)}(a)v_1^{(0)})\\ &=(v_2^{(0)}\blacktriangleright v_1)v_2^{(1)}(a)+v_1(v_2\blacktriangleright a)-(v_2^{(0)}\blacktriangleright v_1^{(-1)}(a))v_2^{(1)}(v_1^{(0)}+v_1^{(-1)}(a)(v_2\triangleright v_1^{(0)}\\ &=v_1(v_2\blacktriangleright a)-(v_2^{(0)}\blacktriangleright v_1^{(-1)}(a))v_2^{(1)}(v_1^{(0)})\\ &=v_1(v_2^{(0)}v_2^{(1)}(a)-av_2)-((v_2^{(0)})^{(0)}(v_2^{(0)})^{(1)}(v_1^{(-1)}(a))-v_1^{(-1)}(a)v_2^{(0)})v_2^{(1)}(v_1^{(0)})\\ &=v_1v_2^{(0)}v_2^{(1)}(a)-v_1av_2-v_2^{(0)}v_2^{(1)}(v_1^{(-1)}(a))v_2^{(2)}(v_1^{(0)})+v_1^{(-1)}(a)v_2^{(0)}v_2^{(1)}(v_1^{(0)}) \end{align*} Comparing terms, we see that the two quantities are indeed equal. \end{proof} We are now ready to prove Theorem \ref{thm:qgd}. \noindent {\bf Proof of Theorem \ref{thm:qgd}}. Let $V=\text{span}_{\mathbb{C}(q)}\{F_i\ |\ i\in I\}\subset U_q(\mathfrak{b}_-)$. Then $V\in {}_\mathcal{K}^\mathcal{K}\mathcal{YD}$ with structure given by \[K_i^{\pm1}\rhd F_j=q_i^{\mp c_{i,j}}F_j;\quad \delta_L(F_i)=K_i^{-1}\otimes F_i.\] Let $V'=V$ as a vector space, but $V'\in {}_\mathcal{K}\mathcal{YD}^\mathcal{K}$ with structure given by \[K_i^{\pm1}\rhd F_j=q_i^{\mp c_{i,j}}F_j;\quad \delta_R(F_i)=F_i\otimes K_i^{-1}.\] Note that we can also consider $V'$ as an object of $\mathcal{YD}_\mathcal{K}^\mathcal{K}$ since $\mathcal{K}$ is commutative. It is well-known (see, e.g., \cite{ahs} or \cite{lusztig}, though Lusztig never used the term ``Nichols algebra") that the corresponding Nichols algebras are isomorphic to $U_q(\mathfrak{n}_-)$ as $\mathcal{K}$-module algebras in the obvious way, i.e. $F_i\mapsto F_i$. By assumption, $\mathbb{C}_q[U]\subset A$, so there is a natural embedding of $U_q(\mathfrak{b}_-)$-module algebras $U_q(\mathfrak{n}_-)\hookrightarrow A$ given by $F_i\mapsto \frac{x_i}{q_i-q_i^{-1}}$. Theorems \ref{thm:adjoint} and \ref{thm:diff} then imply that there is a $U_q(\mathfrak{n}_-)$ action on $A$ given by \[F_i\rhd a=F_i(a)-\frac{x_ia-K_i^{-1}(a)x_i}{q_i-q_i^{-1}},\] matching the proposed action of $F_{i,1}$. Now utilizing a slightly different embedding $U_q(\mathfrak{n}_-)\hookrightarrow A,\ F_i\mapsto -\frac{x_i}{q_i-q_i^{-1}}$, Corollary \ref{cor:op} gives another action of $U_q(\mathfrak{n}_-)\cong U_q(\mathfrak{n}_-)^{op}$ on $A$: \[F_i\rhd a=\frac{x_iK_i^{-1}(a)-ax_i}{q_i-q_i^{-1}},\] matching the proposed action of $F_{i,2}$. It is easily observed that we have made $A$ into both a $\mathcal{B}(V)\rtimes \mathcal{K}$-module algebra and a $\mathcal{B}(V')\rtimes \mathcal{K}$-module algebra. We now wish to show that the operators $F_{i,1}$ and $F_{j,2}$ commute. To do so, we construct the braided cross product $\hat{A}:=A\underline{\rtimes} U_q(\mathfrak{n}_-)$, where the $F_i$ act as $F_{i,1}$. As above, we define ``clever" embeddings of $V$ and $V'$ into $\hat{A}$, namely $F_i\mapsto \frac{x_i}{q_i-q_i^{-1}}\otimes1$ and $F_i\mapsto 1\otimes F_i$, respectively. It is easily checked that the hypotheses of Theorem \ref{thm:commute} are satisfied. Furthermore, the actions defined in Theorem \ref{thm:commute} preserve $A$ and match the actions of $F_{i,1}$ and $F_{j,2}$ on $A$, showing that the prescribed actions of $F_{i,1}$ and $F_{j,2}$ do, in fact, commute.\qed In light of Theorem \ref{thm:qgd}, $\mathbb{C}_q[U]$ is a $U_q(\mathfrak{g}^*)$-module algebra with action given by \[K_i^{\pm1}\rhd x_j=q_i^{\mp c_{i,j}}x_j; \quad F_{i,1}\rhd x_j=0;\quad F_{i,2}\rhd x_j=\frac{q_i^{c_{i,j}}x_ix_j-x_jx_i}{q_i-q_i^{-1}}.\] We make $\mathbb{C}_q[U]$ into a $U_q(\mathfrak{g}^*)$-comodule algebra via the algebra homomorphism $\delta:\mathbb{C}_q[U]\to U_q(\mathfrak{g}^*)\otimes \mathbb{C}_q[U]$ given on generators by \[\delta(x_i)=K_i\otimes x_i+(q_i-q_i^{-1})F_{i,2}K_i\otimes 1.\] The fact that this gives a well-defined algebra homomorphism follows immediately from the following lemma, which can be deduced from the fact that in \cite[1.2.6]{lusztig}, $r:{\bf f}\to {\bf f}\otimes {\bf f}$, $\theta_i\mapsto \theta_i\otimes 1+1\otimes \theta_i$ is well-defined. \begin{lemma} \label{lem:qsr} Let $R$ be any $\mathbb{C}(q)$-algebra and suppose $\{y_i\}_{i\in I},\{z_i\}_{i\in I}\subseteq R$ are two families of elements satisfying the quantum Serre relations. If \[z_jy_i=q_i^{c_{i,j}}y_iz_j\quad \text{for }i,j\in I\] then $\{y_i+z_i\}_{i\in I}$ also satisfies the quantum Serre relations. \end{lemma} It is easily checked that $(\text{id}\otimes \delta)\circ\delta=(\Delta\otimes \text{id})\circ \delta$ and $(\epsilon\otimes \text{id})\circ \delta=\text{id}$. \begin{proposition} \label{prop:ydCqU} The above action and coaction make $\mathbb{C}_q[U]$ into an algebra in the category ${}_{U_q(\mathfrak{g}^*)}^{U_q(\mathfrak{g}^*)}\mathcal{YD}$ of left-left Yetter-Drinfeld modules over $U_q(\mathfrak{g}^*)$. \end{proposition} \begin{proof} We need only verify that the compatibility condition is satisfied, i.e. that \begin{equation} \label{eq:ydcomp} h_{(1)}x^{(-1)}\otimes (h_{(2)}\rhd x^{(0)})=(h_{(1)}\rhd x)^{(-1)}h_{(2)}\otimes (h_{(1)}\rhd x)^{(0)} \end{equation} for $h\in U_q(\mathfrak{g}^*)$ and $x\in \mathbb{C}_q[U]$. It is easily checked that \eqref{eq:ydcomp} is satisfied for $h\in \{K_i^{\pm1},F_{i,1},F_{i,2}\ |\ i\in I\}$ and $x\in \{x_i\ |\ i\in I\}$. Suppose \eqref{eq:ydcomp} is satisfied for some $x,x'\in \mathbb{C}_q[U]$ and all $h\in \{K_i^{\pm1},F_{i,1},F_{i,2}\ |\ i\in I\}$. Then since $\Delta(\{K_i^{\pm1},F_{i,1},F_{i,2}\ |\ i\in I\})\subset \{K_i^{\pm1},F_{i,1},F_{i,2}\ |\ i\in I\}\otimes \{K_i^{\pm1},F_{i,1},F_{i,2}\ |\ i\in I\}$, we observe: \begin{align*} h_{(1)}(xx')^{(-1)}\otimes (h_{(2)}\rhd (xx')^{(0)})&=h_{(1)}x^{(-1)}(x')^{(-1)}\otimes (h_{(2)}\rhd (x^{(0)}(x')^{(0)}))\\ &=h_{(1)}x^{(-1)}(x')^{(-1)}\otimes (h_{(2)}\rhd x^{(0)})(h_{(3)}\rhd (x')^{(0)})\\ &=(h_{(1)}\rhd x)^{(-1)}h_{(2)}(x')^{(-1)}\otimes (h_{(1)}\rhd x)^{(0)}(h_{(3)}\rhd (x')^{(0)})\\ &=(h_{(1)}\rhd x)^{(-1)}(h_{(2)}\rhd x')^{(-1)}h_{(3)}\otimes (h_{(1)}\rhd x)^{(0)}(h_{(2)}\rhd x')^{(0)}\\ &=((h_{(1)}\rhd x)(h_{(2)}\rhd x'))^{(-1)}h_{(3)}\otimes ((h_{(1)}\rhd x)(h_{(2)}\rhd x'))^{(0)}\\ &=(h_{(1)}\rhd (xx'))^{(-1)}h_{(2)}\otimes (h_{(1)}\rhd (xx'))^{(0)} \end{align*} for $h\in \{K_i^{\pm1},F_{i,1},F_{i,2}\ |\ i\in I\}$. Hence we see that the set of all $x\in \mathbb{C}_q[U]$ such that \eqref{eq:ydcomp} holds for all $h\in \{K_i^{\pm1},F_{i,1},F_{i,2}\ |\ i\in I\}$ is a subalgebra of $\mathbb{C}_q[U]$ containing $\{x_i\ |\ i\in I\}$. Namely, \eqref{eq:ydcomp} holds for all $x\in \mathbb{C}_q[U]$ and $h\in \{K_i^{\pm1},F_{i,1},F_{i,2}\ |\ i\in I\}$. Now suppose \eqref{eq:ydcomp} holds for some $h,h'\in U_q(\mathfrak{g}^*)$ and all $x\in \mathbb{C}_q[U]$. Then we observe: \begin{align*} (hh')_{(1)}x^{(-1)}\otimes ((hh')_{(2)}\rhd x^{(0)})&=h_{(1)}h_{(1)}'x^{(-1)}\otimes ((h_{(2)}h_{(2)}')\rhd x^{(0)})\\ &=h_{(1)}h_{(1)}'x^{(-1)}\otimes (h_{(2)}\rhd(h_{(2)}'\rhd x^{(0)}))\\ &=h_{(1)}(h_{(1)}'\rhd x)^{(-1)}h_{(2)}'\otimes (h_{(2)}\rhd (h_{(1)}'\rhd x)^{(0)})\\ &=(h_{(1)}\rhd (h_{(1)}'\rhd x))^{(-1)}h_{(2)}h_{(2)}'\otimes (h_{(1)}\rhd (h_{(1)}'\rhd x))^{(0)}\\ &=((h_{(1)}h_{(1)}')\rhd x)^{(-1)}h_{(2)}h_{(2)}'\otimes ((h_{(1)}h_{(1)}')\rhd x)^{(0)}\\ &=((hh')_{(1)}\rhd x)^{(-1)}(hh')_{(2)}\otimes ((hh')_{(1)}\rhd x)^{(0)} \end{align*} for $x\in \mathbb{C}_q[U]$. Hence we see that the set of all $h\in U_q(\mathfrak{g}^*)$ such that \eqref{eq:ydcomp} holds for all $x\in \mathbb{C}_q[U]$ is a subalgebra of $U_q(\mathfrak{g}^*)$ containing $\{K_i^{\pm1},F_{i,1},F_{i,2}\ |\ i\in I\}$. Namely, \eqref{eq:ydcomp} holds for all $x\in \mathbb{C}_q[U]$ and $h\in U_q(\mathfrak{g}^*)$. \end{proof} The following proposition is probably well-known, but a source was not quickly found, so we provide a proof here. \begin{proposition} \label{prop:ydtensor} Let $H$ be a $\Bbbk$-bialgebra, $A$ an $H$-module algebra, and $B$ an algebra in ${}_H^H\mathcal{YD}$. Then the $H$-module $A\otimes_\Bbbk B$ is an $H$-module algebra with multiplication given by \[(a\otimes b)(a'\otimes b')=a(b^{(-1)}\rhd a')\otimes b^{(0)}b',\quad \text{for all }a,a'\in A,\ b,b'\in B,\] where $\rhd$ is the action of $H$ and $\delta(b)=b^{(-1)}\otimes b^{(0)}$ is the coaction of $H$ in sumless Sweedler notation. \end{proposition} \begin{proof} We first show that $A\otimes_\Bbbk B$ is indeed an associative algebra under the prescribed multiplication. For $a,a',a''\in A$ and $b,b',b''\in B$, we have \begin{align*} ((a\otimes b)(a'\otimes b'))(a''\otimes b'')&=(a(b^{(-1)}\rhd a')\otimes b^{(0)}b)(a''\otimes b'')\\ &=a(b^{(-1)}\rhd a')((b^{(0)}b')^{(-1)}\rhd a'')\otimes (b^{(0)}b)^{(0)}b''\\ &=a(b^{(-1)}\rhd a')((b^{(0)})^{(-1)}(b')^{(-1)}\rhd a'')\otimes (b^{(0)})^{(0)}(b')^{(0)}b''\\ &=a(b^{(-2)}\rhd a')(b^{(-1)}(b')^{(-1)}\rhd a'')\otimes b^{(0)}(b')^{(0)}b''\\ &=a(b^{(-1)}\rhd (a'((b')^{(-1)}\rhd a'')))\otimes b^{(0)}(b')^{(0)}b''\\ &=(a\otimes b)(a'((b')^{(-1)}\rhd a'')\otimes (b')^{(0)}b'')\\ &=(a\otimes b)((a'\otimes b')(a''\otimes b'')). \end{align*} Hence the prescribed multiplication is associative. We now verify that $A\otimes_\Bbbk B$ is indeed an $H$-module algebra. For $h\in H$, $a,a'\in A$, and $b,b'\in B$, we have \begin{align*} h\rhd ((a\otimes b)(a'\otimes b'))&=h\rhd(a(b^{(-1)}\rhd a')\otimes b^{(0)}b')\\ &=(h_{(1)}\rhd a)(h_{(2)}b^{(-1)}\rhd a')\otimes (h_{(3)}\rhd b^{(0)})(h_{(4)}\rhd b')\\ &=(h_{(1)}\rhd a)((h_{(2)}\rhd b)^{(-1)}h_{(3)}\rhd a')\otimes (h_{(2)}\rhd b)^{(0)}(h_{(4)}\rhd b')\\ &=(h_{(1)}\rhd a)((h_{(2)}\rhd b)^{(-1)}\rhd (h_{(3)}\rhd a'))\otimes (h_{(2)}\rhd b)^{(0)}(h_{(4)}\rhd b')\\ &=((h_{(1)}\rhd a)\otimes (h_{(2)}\rhd b))((h_{(3)}\rhd a')\otimes(h_{(4)}\rhd b'))\\ &=(h_{(1)}\rhd (a\otimes b))(h_{(2)}\rhd (a'\otimes b')). \end{align*} $$ h\rhd (1\otimes 1)=(h_{(1)}\otimes 1)\otimes (h_{(2)}\rhd 1) =\varepsilon(h_{(1)})\otimes \varepsilon(h_{(2)}) =\varepsilon(h_{(1)}\varepsilon(h_{(2)}))\otimes 1 =\varepsilon(h)\otimes 1.$$ The proposition is proven. \end{proof} \noindent {\bf Proof of Theorem \ref{thm:qg}}. By Propositions \ref{prop:ydCqU} and \ref{prop:ydtensor} we may give $A\otimes \mathbb{C}_q[U]$ a $U_q(\mathfrak{g}^*)$-module algebra structure satisfying \begin{align*} (1\otimes x_i)(a\otimes 1)&=K_i(a)\otimes x_i+(q_i-q_i^{-1})F_{i,2}K_i(a)\otimes 1\\ K_{i,1}^{\pm1}\rhd (a\otimes x)&=K_i^{\pm1}(a)\otimes K_i^{\pm1}(x)\\ F_{i,1}\rhd (a\otimes x)&=F_{i,1}(a)\otimes x+K_i^{-1}(a)\otimes F_{i,1}(x)\\ F_{i,2}\rhd (a\otimes x)&=F_{i,2}(a)\otimes K_i^{-1}(x)+a\otimes F_{i,2}(x) \end{align*} for $i\in I$, $a\in A$, and $x\in \mathbb{C}_q[U]$. Now by Theorem \ref{thm:sum}, there is a left action of $U_q(\mathfrak{n}_-)$ on $A\otimes \mathbb{C}_q[U]$ given by \begin{align*} F_i\rhd(a\otimes x)&=F_{i,1}\rhd(a\otimes x)+\frac{(1\otimes x_i)(a\otimes x)-(K_i^{-1}\rhd (a\otimes x))(1\otimes x_i)}{q_i-q_i^{-1}}\\ &=F_{i,1}(a)\otimes x+\frac{K_i(a)\otimes x_ix+(q_i-q_i^{-1})F_{i,2}K_i(a)\otimes x-K_i^{-1}(a)\otimes K_i^{-1}(x)x_i}{q_i-q_i^{-1}}\\ &=(F_{i,1}(a)+F_{i,2}K_i(a))\otimes x+\frac{K_i(a)\otimes x_ix-K_i^{-1}(a)\otimes K_i^{-1}(x)x_i}{q_i-q_i^{-1}}\\ &=(F_{i,1}(a)+F_{i,2}K_i(a))\otimes x+\frac{K_i(a)\otimes x_ix-K_i^{-1}(a)\otimes x_ix+K_i^{-1}(a)\otimes x_ix-K_i^{-1}(a)\otimes K_i^{-1}(x)x_i}{q_i-q_i^{-1}}\\ &=(F_{i,1}(a)+F_{i,2}K_i(a))\otimes x+\frac{K_i(a)-K_i^{-1}(a)}{q_i-q_i^{-1}}\otimes x_ix+K_i^{-1}(a)\otimes F_i(x), \end{align*} matching the proposed action of $F_i$. Furthermore, it is obvious that $E_i\rhd (a\otimes x)=a\otimes E_i(x)$ yields a well-defined action of $U_q(\mathfrak{n}_+)$ on $A\otimes \mathbb{C}_q[U]$. It is now straight-forward to check that \begin{align*} K_i^{\pm1}\triangleright (K_j^{\pm1}\triangleright (a\otimes x))&=K_j^{\pm1}\triangleright (K_i^{\pm 1}\triangleright (a\otimes x))\\ K_i\triangleright(E_{j}\triangleright(K_i^{-1}((a\otimes x))))&=q_i^{c_{i,j}}E_{j}\triangleright (a\otimes x)\\ K_i\triangleright(F_{j}\triangleright(K_i^{-1}((a\otimes x))))&=q_i^{-c_{i,j}}F_{j}\triangleright (a\otimes x)\\ E_i\triangleright (F_j\triangleright (a\otimes x))-F_j\triangleright (E_i\triangleright (a\otimes x))&=\delta_{i,j}\frac{K_i\triangleright (a\otimes x)-K_i^{-1}\triangleright (a\otimes x)}{q_i-q_i^{-1}} \end{align*} \begin{align*} K_i^{\pm1}(1\otimes 1)&=1\otimes 1\\ E_i(1\otimes 1)&=0\\ F_i(1\otimes 1)&=0. \end{align*} Hence we have given $A\otimes \mathbb{C}_q[U]$ the structure of a $U_q(\mathfrak{g})$-module. To see that it is in fact a module algebra, we need to check the following. \begin{align} \label{eq:Kcomp} K_i^{\pm1}\triangleright ((a\otimes x)(a'\otimes x'))&=(K_i^{\pm1}\triangleright (a\otimes x))(K_i^{\pm 1}\triangleright (a'\otimes x'))\\ \label{eq:Ecomp} E_{i}\triangleright ((a\otimes x)(a'\otimes x'))&=(E_{i}\triangleright (a\otimes x))(K_i\triangleright (a'\otimes x'))+(a\otimes x)(E_{i}\triangleright (a'\otimes x'))\\ \label{eq:Fcomp} F_{i}\triangleright ((a\otimes x)(a'\otimes x'))&=(F_{i}\triangleright (a\otimes x))(a'\otimes x')+(K_i^{-1}\triangleright(a\otimes x))(F_{i}\triangleright (a'\otimes x')). \end{align} Rather than direct verification, we begin by observing that $$ h\triangleright ((a\otimes 1)z)=(h_{(1)}\triangleright (a\otimes 1))(h_{(2)}\triangleright z)~\text{ and}~ h\triangleright ((a\otimes x_j)z)=(h_{(1)}\triangleright (a\otimes x_j))(h_{(2)}\triangleright z)$$ for $h\in \{K_i^{\pm 1}, E_i, F_i\ |\ i\in I\},\ j\in I,\ a\in A$, and $z\in A\otimes \mathbb{C}_q[U]$. Let $Y=\{x\in \mathbb{C}_q[U]\ |\ h\triangleright ((a\otimes x)z)=(h_{(1)}\triangleright (a\otimes x))(h_{(2)}\triangleright z)$ for all $a\in A,\ z\in A\otimes \mathbb{C}_q[U]$ and $h\in \{K_i^{\pm1},E_i,F_i\ |\ i\in I\}\}$. Then $Y$ is clearly a $\mathbb{C}(q)$-vector space (containing 1 and $x_j$). We show that $Y$ is closed under multiplication. Suppose $x,x'\in Y,$ $a\in A,$ and $z\in A\otimes \mathbb{C}_q[U]$. Then for $h\in \{K_i^{\pm 1},E_i,F_i\ |\ i\in I\}$, \begin{align*} h\triangleright ((a\otimes xx')z)&=h\triangleright ((a\otimes x)(1\otimes x')z)\\ &=(h_{(1)}\triangleright (a\otimes x))(h_{(2)}\triangleright ((1\otimes x')z))\\ &=(h_{(1)}\triangleright (a\otimes x))(h_{(2)}\triangleright (1\otimes x'))(h_{(3)}(z))\\ &=(h_{(1)}\triangleright (a\otimes x)(1\otimes x'))(h_{(2)}(z))\\ &=(h_{(1)}\triangleright (a\otimes xx'))(h_{(2)}(z)) \end{align*} Hence $xx'\in Y$ and we have shown that $Y$ is closed under multiplication. It follows that $Y$ is a $\mathbb{C}(q)$-subalgebra of $\mathbb{C}_q[U]$ containing $x_j$ and hence is actually $\mathbb{C}_q[U]$ itself. Hence we have verified equations \eqref{eq:Kcomp}, \eqref{eq:Ecomp}, and \eqref{eq:Fcomp}. It follows that the given structure makes $A\otimes \mathbb{C}_q[U]$ into a $U_q(\mathfrak{g})$-module algebra. \qed \subsection{Proof of Theorem \ref{thm:qequiv}} \label{pf:qequiv} We begin by constructing a natural isomorphism $\psi: (-)^+\otimes\mathbb{C}_q[U]\Rightarrow \text{id}_{\mathcal{C}_\mathfrak{g}^q}$. For every object $(A,\varphi_A)$ of $\mathcal{C}_\mathfrak{g}^q$, set $\psi_{(A,\varphi_A)}:=m_A\circ (\iota_A\otimes \varphi_A)$, where $\iota_A$ is the inclusion $A^+\hookrightarrow A$ and $m_A:A\otimes A\to A$ is multiplication. As an abuse of notation, we will write $\psi_A$ when context is clear. Since $\psi_A$ is clearly a linear map, we check that it respects multiplication and is $U_q(\mathfrak{g})$-equivariant. One easily computes \[\psi_A((a\otimes 1)(a'\otimes x'))=\psi_A(a\otimes 1)\psi_A(a'\otimes x')\quad \text{and} \quad \psi_A((a\otimes x_i)(a'\otimes x'))=\psi_A(a\otimes x_i)\psi_A(a'\otimes x').\] Let $Y=\{x\in \mathbb{C}_q[U]\ |\ \psi_A((a\otimes x)z)=\psi_A(a\otimes x)\psi_A(z)\ \forall a\in A, z\in A\otimes \mathbb{C}_q[U]\}$. We have seen that $1,x_i\in Y$ for $i\in I$, so the computations \begin{align*} \psi_A((a\otimes (x+y))z)&=\psi_A((a\otimes x)z+(a\otimes y)z)\\ &=\psi_A((a\otimes x)z)+\psi_A((a\otimes y)z)\\ &=\psi_A(a\otimes x)\psi_A(z)+\psi_A(a\otimes y)\psi_A(z)\\ &=(\psi_A(a\otimes x)+\psi_A(a\otimes y))\psi_A(z)\\ &=\psi_A(a\otimes (x+y))\psi_A(z) \end{align*} \begin{align*} \psi_A((a\otimes xy)z)&=\psi_A((a\otimes x)(1\otimes y)z)\\ &=\psi_A(a\otimes x)\psi_A((1\otimes y)z)\\ &=\psi_A(a\otimes x)\psi_A(1\otimes y)\psi_A(z)\\ &=\psi_A((a\otimes x)(1\otimes y))\psi_A(z)\\ &=\psi_A((a\otimes xy))\psi_A(z) \end{align*} show that $Y$ is a subalgebra of $\mathbb{C}_q[U]$ containing a generating set. Hence $Y=\mathbb{C}_q[U]$, i.e. $\psi_A$ is a homomorphism of algebras. Now we verify that $\psi_A$ is $U_q(\mathfrak{g})$-invariant. For $i\in I$, we have \begin{align*} K_i^{\pm1}(\psi_A(a\otimes x))&=K_i^{\pm1}(a\varphi_A(x))\\ &=K_i^{\pm1}(a)K_i^{\pm1}(\varphi_A(x))\\ &=K_i^{\pm1}(a)\varphi_A(K_i^{\pm1}(x))\\ &=\psi_A(K_i^{\pm1}(a)\otimes K_i^{\pm1}(x))\\ &=\psi_A(K_i^{\pm1}\rhd (a\otimes x)) \end{align*} \begin{align*} E_i(\psi_A(a\otimes x))&=E_i(a\varphi_A(x))\\ &=E_i(a)K_i(\varphi_A(x))+aE_i(\varphi_A(x))\\ &=a\varphi_A(E_i(x))\\ &=\psi_A(a\otimes E_i(x))\\ &=\psi_A(E_i\rhd(a\otimes x)) \end{align*} \begin{align*} F_i(\psi_A(a\otimes x))&=F_i(a\varphi_A(x))\\ &=F_i(a)\varphi_A(x)+K_i^{-1}(a)F_i(\varphi_A(x))\\ &=\left(F_i(a)-\frac{\varphi_A(x_i)a-K_i^{-1}(a)\varphi_A(x_i)}{q_i-q_i^{-1}}\right)\varphi_A(x)+\frac{\varphi_A(x_i)a-K_i(a)\varphi_A(x_i)}{q_i-q_i^{-1}}\varphi_A(x)\\ &\hspace{0.5cm}+\frac{K_i(a)-K_i^{-1}(a)}{q_i-q_i^{-1}}\varphi_A(x_i)\varphi_A(x)+K_i^{-1}(a)\varphi_A(F_i(x))\\ &=\left(F_{i,1}(a)+F_{i,2}K_i(a)\right)\varphi_A(x)+\frac{K_i(a)-K_i^{-1}(a)}{q_i-q_i^{-1}}\varphi(x_ix)+K_i^{-1}(a)\varphi_A(F_i(x))\\ &=\psi_A\left(\left(F_{i,1}(a)+F_{i,2}K_i(a)\right)\otimes x+\frac{K_i(a)-K_i^{-1}(a)}{q_i-q_i^{-1}}\otimes x_ix+K_i^{-1}(a)\otimes F_i(x)\right)\\ &=\psi_A(F_i\rhd (a\otimes x)) \end{align*} So $\psi_A$ is a homomorphism of $U_q(\mathfrak{g})$-modules and thus a homomorphism of $U_q(\mathfrak{g})$-module algebras. By Theorem \ref{thm:qfactoring}, $\psi_A$ is an isomorphism of $U_q(\mathfrak{g})$-module algebras. Now $$ \psi_A\circ (1\otimes \text{id})=m_A\circ (\iota_A\otimes \varphi_A)\circ(1\otimes \text{id})=m_A\circ(1\otimes \varphi_A) =\varphi_A.$$ Hence $\psi_A$ is a morphism of $\mathcal{C}_\mathfrak{g}^q$. To show that $\psi_A$ is an isomorphism in $\mathcal{C}_\mathfrak{g}^q$, we make the following easy observation. \begin{lemma} \label{lem:qA-iso} A morphism between objects of $\mathcal{C}_\mathfrak{g}^q$ is an isomorphism if and only if the underlying homomorphism of $U_q(\mathfrak{g})$-module algebras is an isomorphism. \end{lemma} \begin{proof} It is clear that the homomorphism of $U_q(\mathfrak{g})$-module algebras which underlies an isomorphism between objects of $\mathcal{C}_\mathfrak{g}^q$ is actually an isomorphism, so we simply show the converse. Let $(A,\varphi_A)$ and $(B,\varphi_B)$ be objects of $\mathcal{C}_\mathfrak{g}^q$ and $\xi:A\to B$ a morphism between them such that $\xi$ is an isomorphism of $U_q(\mathfrak{g})$-module algebras. Then $\xi\circ \varphi_A=\varphi_B$. Hence we have $\xi^{-1}\circ \varphi_B=\xi^{-1}\circ\xi\circ\varphi_A=\varphi_A$ and so $\xi^{-1}$ is a morphism $(B,\varphi_B)\to(A,\varphi_A)$. Thus $\xi$ is an isomorphism in $\mathcal{C}_\mathfrak{g}^q$. \end{proof} Hence $\psi_A$ is actually an isomorphism in $\mathcal{C}_\mathfrak{g}^q$. If we can show that $\psi:=(\psi_A)_{(A,\varphi_A)\in \mathcal{C}_\mathfrak{g}^q}$ is a natural transformation between $(-)^+\otimes \mathbb{C}_q[U]$ and $\text{id}_{\mathcal{C}_\mathfrak{g}^q}$, then we will have shown that it is a natural isomorphism. Let $(A,\varphi_A)$ and $(B,\varphi_B)$ be objects of $\mathcal{C}_\mathfrak{g}^q$ and $\xi:A\to B$ a morphism. Then \begin{align*} \psi_B\circ(\xi|_{A^+}\otimes \text{id})&=m_B\circ (\iota_B\otimes \varphi_B)\circ(\xi|_{A^+}\otimes \text{id})\\ &=m_B\circ(\xi|_{A^+}\otimes \varphi_B)\\ &=m_B\circ (\xi|_{A^+}\otimes (\xi\circ\varphi_A))\\ &=m_B\circ (\xi\otimes\xi)\circ(\iota_A\otimes \varphi_A)\\ &=\xi\circ m_A\circ (\iota_A\otimes \varphi_A)\\ &=\xi\circ \psi_A \end{align*} Hence $\psi:(-)^+\otimes\mathbb{C}_q[U]\Rightarrow \text{id}_{\mathcal{C}_\mathfrak{g}^q}$ is a natural transformation and therefore a natural isomorphism. Now for every $U_q(\mathfrak{g}^*)$-module $A$, let $\eta_A=\text{id}\otimes1:A\to (A\otimes \mathbb{C}_q[U])^+$. Then $\eta_A$ is obviously an injective homomorphism of algebras. We need to show that $\eta_A$ is a homomorphism of $U_q(\mathfrak{g}^*)$-module algebras, namely that $\eta_A$ respects the action of $U_q(\mathfrak{g}^*)$. So we make the following computations. $$ K_i^{\pm1}(\eta_A(a))=K_i^{\pm1}(a\otimes 1)=K_i^{\pm1}(a)\otimes 1 =\eta_A(K_i^{\pm1}(a)),$$ \begin{align*} F_{i,1}(\eta_A(a))&=F_{i,1}(a\otimes 1)\\ &=F_i(a\otimes 1)-\frac{(1\otimes x_i)(a\otimes 1)-K_i^{-1}(a\otimes 1)(1\otimes x_i)}{q_i-q_i^{-1}}\\ &=\left(F_{i,1}(a)+F_{i,2}K_i(a)\right)\otimes1+\frac{K_i(a)-K_i^{-1}(a)}{q_i-q_i^{-1}}\otimes x_i-\frac{K_i(a)\otimes x_i+F_{i,2}K_i(a)\otimes 1-K_i^{-1}(a)\otimes x_i}{q_i-q_i^{-1}}\\ &=F_{i,1}(a)\otimes 1\\ &=\eta_A(F_{i,1}(a)) \end{align*} \begin{align*} F_{i,2}(\eta_A(a))&=F_{i,2}(a\otimes 1)\\ &=\frac{(1\otimes x_i)K_i^{-1}(a\otimes 1)-(a\otimes 1)(1\otimes x_i)}{q_i-q_i^{-1}}\\ &=\frac{(1\otimes x_i)(K_i^{-1}(a)\otimes 1)-a\otimes x_i}{q_i-q_i^{-1}}\\ &=\frac{K_iK_i^{-1}(a)\otimes x_i+(q_i-q_i^{-1})F_{i,2}K_iK_i^{-1}(a)\otimes 1-a\otimes x_i}{q_i-q_i^{-1}}\\ &=F_{i,2}(a)\otimes 1\\ &=\eta_A(F_{i,2}(a)). \end{align*} Hence $\eta_A$ respects the action of $U_q(\mathfrak{g})$. Our last step is to show that $\eta_A$ is surjective. Given an arbitrary element $\displaystyle \sum_{k=1}^n a_k\otimes x_{{\bf j}_k}\in (A\otimes \mathbb{C}_q[U])^+$ with ${\bf j}_k<{\bf j}_l$ if $k<l$, we have $$ \sum_{k=1}^n a_k\otimes x_{{\bf j}_k}=E_{\bf i}^{(top)}\left(\sum_{k=1}^n a_k\otimes x_{{\bf j}_k}\right)=a_n\otimes 1.$$ Hence $(A\otimes \mathbb{C}_q[U])^+=A\otimes \mathbb{C}(q)$, so $\eta_A$ is surjective and therefore an isomorphism. One easily checks that $\eta:=(\eta_A)_{A\in U_q(\mathfrak{g}^*)-{\bf ModAlg}}$ is a natural transformation. Since each $\eta_A$ is an isomorphism, $\eta$ is a natural isomorphism $\eta:\text{id}_{U_q(\mathfrak{g}^*)-{\bf ModAlg}}\Rightarrow(-\otimes \mathbb{C}_q[U])^+$. We have now shown that $(-)^+\otimes \mathbb{C}_q[U]\cong \text{id}_{\mathcal{C}_\mathfrak{g}^q}$ and $(-\otimes \mathbb{C}_q[U])^+\cong \text{id}_{U_q(\mathfrak{g}^*)-{\bf ModAlg}}$, so $\mathbb{C}_q[U]\otimes -$ and $(-)^+$ are quasi-inverse equivalences of categories.\qed \subsection{Proof of Proposition \ref{prop:qtensor}} \label{pf:qtensor} We know by Theorem \ref{thm:qfactoring} that $A\cong A^+\otimes \mathbb{C}_q[U]$ and $B\cong B^+\otimes \mathbb{C}_q[U]$ and Theorem \ref{thm:qequiv} says this is an isomorphism of $U_q(\mathfrak{g})$-module algebras. We now consider the map \[\mu_L:(A\underline{\otimes} B)^+\otimes [\varphi_A(\mathbb{C}_q[U])\otimes \mathbb{C}(q)]\to A\underline{\otimes} B\cong (A^+\otimes \mathbb{C}_q[U])\underline{\otimes}(B^+\otimes \mathbb{C}_q[U])\] as in Theorem \ref{thm:qfactoring}. As in the proof of Corollary \ref{cor:qhighest} (Section \ref{pf:qhighest}), $\mu_L((A\underline{\otimes} B)^+\otimes [\varphi_A(\mathbb{C}_q[U])\otimes \mathbb{C}(q)])$ is a subalgebra of $A\underline{\otimes} B$. Since $A^+\otimes \mathbb{C}(q)$, $\mathbb{C}(q)\otimes B^+$, and $\{\varphi_A(x_i)\otimes 1-1\otimes \varphi_B(x_i)\ |\ i\in I\}$ are all contained in $(A\underline{\otimes} B)^+$ and $\{\varphi_A(x_i)\otimes 1\ |\ i\in I\}\subseteq \varphi_A(\mathbb{C}_q[U])\otimes \mathbb{C}(q)$, it follows that $\mu_L((A\underline{\otimes} B)^+\otimes [\varphi_A(\mathbb{C}_q[U])\otimes \mathbb{C}(q)])$ contains all of these sets. Hence $\mu_L((A\underline{\otimes} B)^+\otimes [\varphi_A(\mathbb{C}_q[U])\otimes \mathbb{C}(q)])$ contains a generating set of $A\underline{\otimes} B$. Being a subalgebra, it follows that $\mu_L((A\underline{\otimes} B)^+\otimes [\varphi_A(\mathbb{C}_q[U])\otimes \mathbb{C}(q)])=A\underline{\otimes} B$, i.e. $\mu_L$ is surjective. Hence $\mu_L$ is an isomorphism and by Theorem \ref{thm:qfactoring}, $A\underline{\otimes} B$ is adapted and $\nu_{\bf i}(A\underline{\otimes} B\setminus\{0\})=\nu_{\bf i}(\mathbb{C}_q[U]\setminus\{0\})\ \forall {\bf i}\in R(w_\mathrm{o})$. Since $A\underline{\otimes}B$ is a $U_q(\mathfrak{g})$-weight module algebra and $1\otimes \varphi_B$ and $\varphi_A\otimes 1$ are injections, the proposition follows.\qed \subsection{Proof of Proposition \ref{prop:qfusion}} \label{pf:qfusion} The vector space $A\otimes B$ is naturally viewed as a subspace of $A*B$ (or $A\star B$ if you prefer) via $a\otimes b\mapsto (a\otimes 1)\otimes (b\otimes 1)$. In fact, this subspace is actually a subalgebra since \[((a\otimes 1)\otimes (b\otimes 1))((a'\otimes 1)\otimes (b'\otimes 1))=q^{(|a'|,|b|)}(aa'\otimes 1)\otimes (bb'\otimes 1)\] for weight vectors $a,a'\in A\text{ and }b,b'\in B$ of weight $|a|,|a'|,|b|,$ and $|b'|$, respectively. Hence we may equip $A\otimes B$ with this multiplication. By design, the prescribed actions of $K_i$ and $F_{i,1}$ on $A\otimes B$ match those on $(A\otimes \mathbb{C}(q))\otimes (B\otimes \mathbb{C}(q))\subset A*B$, while the prescribed actions of $K_i^{\pm1}$ and $F_{i,2}$ match those on $(A\otimes \mathbb{C}(q))\otimes (B\otimes \mathbb{C}(q))\subset A\star B$. A straightforward check verifies that $F_{i,1}\rhd(F_{j,2}\rhd(a\otimes b))=F_{j,2}\rhd(F_{i,1}\rhd(a\otimes b))$ for $a\in A,\ b\in B$, and $i,j\in I$, so it follows that the prescribed action of $U_q(\mathfrak{g}^*)$ on $A\otimes B$ is well-defined and compatible with multiplication. \qed \subsection{Proofs of Theorems \ref{thm:b-} and \ref{thm:g}} \label{pf:b-,g} We begin with a theorem analogous to Theorem \ref{thm:braidedtriv}. Actually, it follows from \cite[7.3.3]{montgomery}, but for convenience, we give a self-contained proof. \begin{theorem} \label{thm:triv} Let $H$ be a Hopf algebra over a field $\Bbbk$ and suppose $A$ is an $H$-module algebra containing $H$ as a subalgebra. Then giving $A$ the structure of an $H$-module algebra via the adjoint action: \[h\rhd a=h_{(1)}aS(h_{(2)})\quad \text{for } h\in H,\ a\in A,\] the linear map $\tau:=(m\otimes id)\circ (id\otimes \iota\otimes id)\circ (id\otimes \Delta):A\rtimes H\to A\otimes_\Bbbk H$ is an algebra isomorphism with inverse $\tau^{-1}=(m\otimes id)\circ (id\otimes\iota\otimes id)\circ(id\otimes S\otimes id)\circ (id\otimes \Delta)$, where $\iota:H\hookrightarrow A$ is the inclusion. \end{theorem} \begin{proof} We first verify that $\tau$ and $\tau^{-1}$ are truly mutually inverse (and hence that we are justified in using the name $\tau^{-1}$). For $a\in A$ and $h\in H$, we directly compute $$ (\tau\circ\tau^{-1})(a\otimes h)=\tau(aS(h_{(1)})\otimes h_{(2)}) =aS(h_{(1)})h_{(2)}\otimes h_{(3)} =a\varepsilon(h_{(1)})\otimes h_{(2)} =a\otimes\varepsilon(h_{(1)})h_{(2)}=a\otimes h,$$ $$ (\tau^{-1}\circ\tau)(a\otimes h)=\tau^{-1}(ah_{(1)}\otimes h_{(2)}) =ah_{(1)}S(h_{(2)})\otimes h_{(3)} =a\varepsilon(h_{(1)})\otimes h_{(2)} =a\otimes\varepsilon(h_{(1)})h_{(2)} =a\otimes h.$$ Since $\tau\circ \tau^{-1}$ and $\tau^{-1}\circ \tau$ act as the identity on pure tensors, they are both the identity homomorphism. Hence $\tau$ and $\tau^{-1}$ are mutually inverse. We conclude by verifying that $\tau$ is actually a homomorphism of algebras (and then $\tau^{-1}$ automatically is as well). \begin{align*} \tau(a\otimes h)\tau(a'\otimes h')&=(ah_{(1)}\otimes h_{(2)})(a'h'_{(1)}\otimes h'_{(2)})\\ &=ah_{(1)}a'h'_{(1)}\otimes h_{(2)}h'_{(2)}\\ &=ah_{(1)}a'h'_{(1)}\otimes \varepsilon(h_{(2)})h_{(3)}h'_{(2)}\\ &=ah_{(1)}a'\varepsilon(h_{(2)})h'_{(1)}\otimes h_{(3)}h'_{(2)}\\ &=ah_{(1)}a'S(h_{(2)})h_{(3)}h'_{(1)}\otimes h_{(4)}h'_{(2)}\\ &=a(h_{(1)}\rhd a')h_{(2)}h'_{(1)}\otimes h_{(3)}h'_{(2)}\\ &=\tau(a(h_{(1)}\rhd a')\otimes h_{(2)}h')\\ &=\tau((a\otimes h)(a'\otimes h')) \end{align*} Again, since pure tensors span $A\otimes_\Bbbk H$ and $\tau$ is a linear map, it follows that $\tau$ respects multiplication. The theorem is proved. \end{proof} We now apply Theorem \ref{thm:triv} to the situation when $\Bbbk=\mathbb{C}(q)$ and $H=A=U_q(\mathfrak{b}_-)$, yielding a trivializing isomorphism $\tau:U_q(\mathfrak{b}_-)\rtimes U_q(\mathfrak{b}_-)\to U_q(\mathfrak{b}_-)\otimes U_q(\mathfrak{b}_-)$. Now we also have an embedding of $U_q(\mathfrak{b}_-)$-module algebras $\mathbb{C}_q[U]\hookrightarrow U_q(\mathfrak{b}_-)$, $x_i\mapsto (q_i-q_i^{-1})F_i$. This induces an embedding of algebras $\iota:\mathbb{C}_q[U]\rtimes U_q(\mathfrak{b}_-)\hookrightarrow U_q(\mathfrak{b}_-)\rtimes U_q(\mathfrak{b}_-)$. Then applying $\tau$, we have an embedding of algebras $\tau\circ\iota:\mathbb{C}_q[U]\rtimes U_q(\mathfrak{b}_-)\hookrightarrow U_q(\mathfrak{b}_-)\otimes U_q(\mathfrak{b}_-)$. Then we have \begin{align*} (\tau\circ \iota)\left(1\otimes F_i-x_i\otimes \frac{1-K_i^{-2}}{q_i-q_i^{-1}}\right)=&\tau(1\otimes F_i-F_i\otimes (1-K_i^{-2}))\\ =&F_i\otimes 1+K_i^{-1}\otimes F_i-F_i\otimes 1+F_iK_i^{-2}\otimes K_i^{-2}\\ =&K_i^{-1}\otimes F_i+F_iK_i^{-2}\otimes K_i^{-2}. \end{align*} Now the families $\{K_i^{-1}\otimes F_i\}_{i\in I}$ and $\{F_iK_i^{-2}\otimes K_i^{-2}\}_{i\in I}$ clearly satisfy the quantum Serre relations and \begin{align*} (F_jK_j^{-2}\otimes K_j^{-2})(K_i^{-1}\otimes F_i)&=F_jK_j^{-2}K_i^{-1}\otimes K_j^{-2}F_i\\ &=q_i^{c_{i,j}}K_i^{-1}F_jK_j^{-2}\otimes F_iK_j^{-2}\\ &=q_i^{c_{i,j}}(K_i^{-1}\otimes F_i)(F_jK_j^{-2}\otimes K_j^{-2}) \end{align*} for $i,j\in I$. Then by Lemma \ref{lem:qsr}, both families $\{K_i^{-1}\otimes F_i+F_iK_i^{-2}\otimes K_i^{-2}\}_{i\in I}$ and $\left\{1\otimes F_i-x_i\otimes \frac{1-K_i^{-2}}{q_i-q_i^{-1}}\right\}_{i\in I}$ must also satisfy the quantum Serre relations. Since \[(1\otimes K_i)\left(1\otimes F_i-x_i\otimes \frac{1-K_i^{-2}}{q_i-q_i^{-1}}\right)(1\otimes K_i^{-1})=q_i^{-c_{i,j}}\left(1\otimes F_i-x_i\otimes \frac{1-K_i^{-2}}{q_i-q_i^{-1}}\right),\] there is a well-defined homomorphism of algebras $U_q(\mathfrak{b}_-)\to \mathbb{C}_q[U]\rtimes U_q(\mathfrak{b}_-)$ such that ${K_i^{\pm1}}\mapsto1\otimes K_i^{\pm1}$, ${F_i}\mapsto1\otimes F_i-x_i\otimes \frac{1-K_i^{-2}}{q_i-q_i^{-1}}$. We now wish to dequantize the above map. Set $\mathscr{A}=\{g\in \mathbb{C}(q)\ |\ g\text{ is regular at }q=1\}$, a local subring of $\mathbb{C}(q)$ with maximal ideal $(q-1)\mathscr{A}$. Denote by $U_q(\mathfrak{b}_-)_{\mathscr{A}}$ the $\mathscr{A}$-subalgebra of $U_q(\mathfrak{b}_-)$ generated by $\left\{F_i,\frac{K_i^{\pm1}-1}{q_i-1}\ |\ \forall i\in I\right\}$. In fact, it is a Hopf subalgebra over $\mathscr{A}$. Then $(\mathscr{A}\mathcal{B}^{dual})\rtimes U_q(\mathfrak{b}_-)_\mathscr{A}$ is an $\mathscr{A}$-subalgebra of $\mathbb{C}_q[U]\rtimes U_q(\mathfrak{b}_-)$. Hence the above map restricts to a homomorphism of $\mathscr{A}$-algebras $U_q(\mathfrak{b}_-)_\mathscr{A}\to (\mathscr{A}\mathcal{B}^{dual})\rtimes U_q(\mathfrak{b}_-)_\mathscr{A}$, inducing a homomorphism of $\mathbb{C}$-algebras \[U(\mathfrak{b}_-)=U_q(\mathfrak{b}_-)_\mathscr{A}/(q-1)U_q(\mathfrak{b}_-)_\mathscr{A}\to [(\mathscr{A}\mathcal{B}^{dual})\rtimes U_q(\mathfrak{b}_-)_\mathscr{A}]/(q-1)[(\mathscr{A}\mathcal{B}^{dual})\rtimes U_q(\mathfrak{b}_-)_\mathscr{A}]=\mathbb{C}[U]\rtimes U(\mathfrak{b}_-).\] Hence the following proposition is proven. \begin{proposition} \label{prop:hat} The assignments $h_i\mapsto 1\otimes h_i$ and $f_i\mapsto 1\otimes f_i-x_i\otimes h_i$ define a homomorphism of algebras $\hat{\null}:U(\mathfrak{b}_-)\to \mathbb{C}[U]\rtimes U(\mathfrak{b}_-)$. \end{proposition} Another, less interesting proof of Proposition \ref{prop:hat} involves showing by induction that the elements $\hat{f}_i=1\otimes f_i-x_i\otimes h_i\in \mathbb{C}[U]\rtimes U(\mathfrak{b}_-)$ satisfy \begin{align*} (\text{ad }\hat{f}_i)^{(n)}(\hat{f}_j)=\sum_{k=0}^n\left(\prod_{\ell=n-k}^{n-1}(c_{i,j}+\ell)\right)&\left[x_i^{(k)}\otimes (\text{ad }f_i)^{(n-k)}(f_j)-\frac{d_i}{d_j}\,x_i^{(k-1)}f_i^{(n-k)}(x_j)\otimes f_i\right.\\ \nonumber &\hspace{0.25cm}\left.-\frac{1}{d_j}\,x_i^{(k)}f_i^{(n-k)}(x_j)\otimes ((n-k)d_ih_i+d_jh_j)\right] \end{align*} \noindent where we use the conventions $(\text{ad }x)(y)=xy-yx$ and $x_i^{(-1)}=0$. Setting $n=1-c_{i,j}$ and using Remark \ref{rem:nilpotentf}, we verify that $(\text{ad }\hat{f}_i)^{(1-c_{i,j})}(\hat{f}_j)=0$. In any case, we are now ready to prove Theorem \ref{thm:b-}. \noindent{\bf Proof of Theorem \ref{thm:b-}}. Since $A$ contains $\mathbb{C}[U]$ as a $\mathfrak{g}$-module subalgebra, there is an action of $\mathbb{C}[U]\rtimes U(\mathfrak{b}_-)$ on $A$ given by $(x\otimes g)\blacktriangleright a=xg(a)$. Then $\hat{\null}:U(\mathfrak{b}_-)\to \mathbb{C}[U]\rtimes U(\mathfrak{b}_-)$ gives rise to an action of $U(\mathfrak{b}_-)$ on $A$ via $g\rhd a=\hat{g}\blacktriangleright a$. Under this action, we have \[h_i\rhd a=(1\otimes h_i)\blacktriangleright a=h_i(a)\quad\text{and}\quad f_i\rhd a=(1\otimes f_i-x_i\otimes h_i)\blacktriangleright a=f_i(a)-x_ih_i(a)\] as prescribed. That $A^+$ is invariant under this action is an easy computation that we will not produce here. \qed Theorem \ref{thm:b-} of course gives rise to a ``new" $\mathfrak{b}_-$-module algebra structure on $\mathbb{C}[U]$, satisfying $$ h_i(x_j)=-c_{i,j}x_j,~h_i(\{x,y\})=\{h_i(x),y\}+\{x,h_i(y)\},~f_i(x)=\frac{1}{2d_i}\{x_i,x\}-\frac{1}{2}x_i h_i(x)$$ for $i\in I$ and $x,y\in \mathbb{C}[U]$. To distinguish $\mathbb{C}[U]$ equipped with this action from that with the usual action, we will use $\mathbb{C}[U]^{op}$ to denote the algebra $\mathbb{C}[U]$ equipped with the ``new" action. As with $\mathbb{C}[U]$, we can form the cross product $\mathbb{C}[U]^{op}\rtimes U(\mathfrak{b}_-)$. We now consider the well-known homomorphism of algebras $s:\mathbb{C}[U]\to \mathbb{C}[U]$ defined by the assignments $s(x_i)=-x_i$ for $i\in I$ and $s(\{x,y\})=-\{s(x),s(y)\}$ for $x,y\in \mathbb{C}[U]$. We will examine the linear map $s\otimes id: \mathbb{C}[U]\rtimes U(\mathfrak{b}_-)\to \mathbb{C}[U]^{op}\rtimes U(\mathfrak{b}_-)$, but first we need the following lemma. \begin{lemma} \label{lem:maps} Given a field $\Bbbk$, let $H$ be a $\Bbbk$-bialgebra generated as a $\Bbbk$-algebra by the subset $X$, $A$ an $H$-module algebra, and $B$ any $\Bbbk$-algebra with multiplication $\mu_B:B\otimes_\Bbbk B\to B$. Given homomorphisms of $\Bbbk$-algebras $\varphi_1:A\to B$ and $\varphi_2:H\to B$, the $\Bbbk$-linear map $\mu_B\circ (\varphi_1\otimes \varphi_2):A\rtimes H\to B$ is a homomorphism of algebras if and only if \begin{equation} \label{eq:commute} \varphi_2(h)\varphi_1(a)=\varphi_1(h_{(1)}(a))\varphi_2(h_{(2)}) \end{equation} for all $h\in X$ and $a\in A$, where we use sumless Sweedler notation: $\Delta(h)=h_{(1)}\otimes h_{(2)}$. \end{lemma} \begin{proof} To simplify notation, we write $\phi:=\mu_B\circ(\varphi_1\otimes \varphi_2)$. $(\Rightarrow)$ If $\phi$ is a homomorphism of algebras, then for $h\in X$ and $a\in A$, we have $$ \varphi_2(h)\varphi_1(a)=\phi(1\otimes h)\phi(a\otimes 1) =\phi((1\otimes h)(a\otimes 1)) =\phi(h_{(1)}(a)\otimes h_{(2)}) =\varphi_1(h_{(1)}(a))\varphi_2(h_{(2)}).$$ $(\Leftarrow)$ Let $Y$ be the subset of $H$ consisting of elements $h$ such that \eqref{eq:commute} holds for all $a\in A$. By assumption, $X\subset Y$, so showing that $Y$ is a subalgebra of $H$ is equivalent to showing that $Y=H$. We compute for $h,h'\in Y$ and $a\in A$: \begin{align*} \varphi_2(h+h')\varphi_1(a)&=(\varphi_2(h)+\varphi_2(h'))\varphi_1(a)\\ &=\varphi_2(h)\varphi_1(a)+\varphi_2(h')\varphi_1(a)\\ &=\varphi_1(h_{(1)}(a))\varphi_2(h_{(2)})+\varphi_1(h'_{(1)}(a))\varphi_2(h'_{(2)})\\ &=\varphi_1((h+h')_{(1)}(a))\varphi_2((h+h')_{(2)}) \end{align*} \begin{align*} \varphi_2(hh')\varphi_1(a)&=\varphi_2(h)\varphi_2(h')\varphi_1(a)\\ &=\varphi_2(h)\varphi_1(h'_{(1)}(a))\varphi_2(h'_{(2)})\\ &=\varphi_1(h_{(1)}h'_{(1)}(a))\varphi_2(h_{(2)})\varphi_2(h'_{(2)})\\ &=\varphi_1(h_{(1)}h'_{(1)}(a))\varphi_2(h_{(2)}h'_{(2)})\\ &=\varphi_1((hh')_{(1)}(a))\varphi_2((hh')_{(2)}). \end{align*} So we see that $Y$ is closed under addition and multiplication. Since it obviously contains $1$, $Y$ is a subalgebra of $H$ and hence $Y=H$. Thus \eqref{eq:commute} holds for all $h\in H$ and $a\in A$. Now we compute for $h,h'\in H$ and $a,a'\in A$: \begin{align*} \phi(a\otimes h)\phi(a'\otimes h')&=\varphi_1(a)\varphi_2(h)\varphi_1(a')\varphi_2(h')\\ &=\varphi_1(a)\varphi_1(h_{(1)}(a'))\varphi_2(h_{(2)})\varphi_2(h')\\ &=\varphi_1(ah_{(1)}(a'))\varphi_2(h_{(2)}h')\\ &=\phi(ah_{(1)}(a')\otimes h_{(2)}h')\\ &=\phi((a\otimes h)(a'\otimes h')). \end{align*} Since we already knew that $\phi$ was a $\Bbbk$-linear map, it follows that $\phi$ is a homomorphism of $\Bbbk$-algebras. \end{proof} \begin{proposition} The linear map $s\otimes id:\mathbb{C}[U]\rtimes U(\mathfrak{b}_-)\to \mathbb{C}[U]^{op}\rtimes U(\mathfrak{b}_-)$ is a homomorphism of algebras. \end{proposition} \begin{proof} By Lemma \ref{lem:maps}, it suffices to show that $(1\otimes h_i)(s(x)\otimes 1)=s(h_i(x))\otimes 1+s(x)\otimes h_i$ and $(1\otimes f_i)(s(x)\otimes 1)=s(f_i(x))\otimes 1+s(x)\otimes f_i$ for $i\in I$ and $x\in \mathbb{C}[U]$. It is clear that $s$ respects the action of $h_i$, namely $h_i(s(x))=s(h_i(x))$ for $i\in I$ and $x\in \mathbb{C}[U]$, so $$ (1\otimes h_i)(s(x)\otimes 1)=h_i(s(x))\otimes 1+s(x)\otimes h_i=s(h_i(x))\otimes 1+s(x)\otimes h_i,$$ \begin{align*} (1\otimes f_i)(s(x)\otimes 1)&=f_i(s(x))\otimes 1+s(x)\otimes f_i\\ &=\left(\frac{1}{2d_i}\{x_i,s(x)\}-\frac{1}{2}x_ih_i(s(x))\right)\otimes 1+s(x)\otimes f_i\\ &=\left(-\frac{1}{2d_i}\{s(x_i),s(x)\}+\frac{1}{2}s(x_i)s(h_i(x))\right)\otimes 1+s(x)\otimes f_i\\ &=\left(\frac{1}{2d_i}s(\{x_i,x\})+\frac{1}{2}s(x_ih_i(x))\right)\otimes 1 + s(x)\otimes f_i\\ &=s\left(\frac{1}{2d_i}\{x_i,x\}+\frac{1}{2}x_ih_i(x)\right)\otimes 1+s(x)\otimes f_i\\ &=s(f_i(x))\otimes 1+s(x)\otimes f_i. \end{align*} Hence $s\otimes \text{id}$ is a homomorphism of algebras. \end{proof} We are now ready to prove Theorem \ref{thm:g}. \noindent{\bf Proof of Theorem \ref{thm:g}}. We observe that $A\otimes \mathbb{C}[U]^{op}$ is naturally a $\mathfrak{b}_-$-module algebra and therefore is also a $\mathbb{C}[U]^{op}\rtimes U(\mathfrak{b}_-)$-module. \begin{lemma} \label{lem:b-embed} There is a homomorphism of algebras $U(\mathfrak{b}_-)\to \mathbb{C}[U]^{op}\rtimes U(\mathfrak{b}_-)$ given on generators by \[h_i\mapsto 1\otimes h_i\quad \text{and}\quad f_i\mapsto 1\otimes f_i+x_i\otimes h_i.\] \end{lemma} \begin{proof} We already saw that there is a homomorphism of algebras $\hat{\null}:U(\mathfrak{b}_-)\to \mathbb{C}[U]\rtimes U(\mathfrak{b}_-)$ given on generators by $\hat{h}_i= 1\otimes h_i$ and $\hat{f}_i=1\otimes f_i-x_i\otimes h_i$. Then the composition $(s\otimes \text{id})\circ \hat{\null}:U(\mathfrak{b}_-)\to \mathbb{C}[U]^{op}\rtimes U(\mathfrak{b}_-)$ has $ (s\otimes \text{id})(\hat{h}_i)=(s\otimes \text{id})(1\otimes h_i)=1\otimes h_i,~(s\otimes \text{id})(\hat{f}_i)=(s\otimes \text{id})(1\otimes f_i-x_i\otimes h_i)=1\otimes f_i+x_i\otimes h_i$ as desired. \end{proof} The homomorphism in Lemma \ref{lem:b-embed} induces a different action of $\mathfrak{b}_-$ on $A\otimes \mathbb{C}[U]^{op}$, namely \begin{align*} h_i\rhd(a\otimes x)&=h_i(a)\otimes x+a\otimes h_i(x)\\ f_i\rhd(a\otimes x)&=f_i(a)\otimes x+a\otimes f_i(x)+h_i(a)\otimes x_ix+a\otimes x_ih_i(x)\\ &=f_i(a)\otimes x+h_i(a)\otimes x_ix+a\otimes \left(\frac{1}{2d_i}\{x_i,x\}-\frac{1}{2}x_ih_i(x)+x_ih_i(x)\right)\\ &=f_i(a)\otimes x+h_i(a)\otimes x_ix+a\otimes \left(\frac{1}{2d_i}\{x_i,x\}+\frac{1}{2}x_ih_i(x)\right) \end{align*} Hence there is a well-defined action of $\mathfrak{b}_-$ on $A\otimes \mathbb{C}[U]$ (note the lack of $op$) given by \[h_i\rhd(a\otimes x)=h_i(a)\otimes x+a\otimes h_i(x),\quad f_i\rhd(a\otimes x)=f_i(a)\otimes x+h_i(a)\otimes x_ix+a\otimes f_i(x)\] as prescribed in the theorem. It is clear by definition that the family of operators $\{1\otimes e_i\}_{i\in I}$ which define the action of $e_i$ satisfy the Serre relations. It is also easily checked that \begin{align*} h_i\rhd (e_j\rhd (a\otimes x))-e_j\rhd (h_i\rhd (a\otimes x))&=c_{i,j}e_j\rhd(a\otimes x)\\ h_i\rhd (f_j\rhd (a\otimes x))-f_j\rhd (h_i\rhd (a\otimes x))&=-c_{i,j}f_j\rhd(a\otimes x)\\ e_i\rhd (f_j\rhd (a\otimes x))-f_j\rhd(e_i\rhd (a\otimes x))&=\delta_{i,j}h_i\rhd(a\otimes x). \end{align*} Hence we have a well-defined action of $\mathfrak{g}$ on $A\otimes \mathbb{C}[U]$. It is clear that each $e_i$ and $f_i$ acts by derivations and $e_i(1\otimes 1)=f_i(1\otimes 1)=0$, so the theorem is proved.\qed \subsection{Proof of Proposition \ref{prop:tensor}} \label{pf:tensor} We simply show that if $(A,\varphi_A)$ and $(B,\varphi_B)$ are objects of $\mathcal{C}_\mathfrak{g}$, then $A\otimes B$ is adapted with $\nu_{\bf i}(A\otimes B\setminus\{0\})=\nu_{\bf i}(\mathbb{C}[U]\setminus\{0\})$ for all ${\bf i}\in R(w_\mathrm{o})$. By Theorem \ref{thm:factoring}, it suffices to show that the map $\mu:(A\otimes B)^+\otimes (\varphi_A(\mathbb{C}[U])\otimes \mathbb{C}(q))\to A\otimes B$ is surjective. Again by Theorem \ref{thm:factoring}, we know that $A\cong A^+\otimes \mathbb{C}[U]$ and $B\cong B^+\otimes \mathbb{C}[U]$. Theorem \ref{thm:equiv} says that these are isomorphisms of $\mathfrak{g}$-module algebras. We observe that the following elements are contained in the image of $\mu$: $a\otimes 1$ for $a\in A^+$, $1\otimes b$ for $b\in B^+$, $\varphi_A(x_i)\otimes 1$ for $i\in I$, and $\varphi_A(x_i)\otimes 1-1\otimes \varphi_B(x_i)$ for $i\in I$. In fact, the image of $\mu$ is a $\mathfrak{g}$-module subalgebra of $A\otimes B$. \begin{lemma} Let $C$ be a $\mathfrak{g}$-module algebra containing $\mathbb{C}[U]$ as a $\mathfrak{g}$-module subalgebra and let $\mu:C^+\otimes \mathbb{C}[U]\to C$ be restriction of multiplication to these subalgebras. Then the image of $\mu$ is a $\mathfrak{g}$-module subalgebra of $C$. \end{lemma} \begin{proof} Since $\mu$ is $\mathbb{C}$-linear, it suffices to check that the subspace spanned by the images of pure tensors $c\otimes x$ is closed under multiplication and the $\mathfrak{g}$-action. Now since $\mu(c\otimes x)=cx$, we check only on elements of this form: $$ (cx)(c'x')=(cc')(xx')=\mu(cc'\otimes xx')\in \mu(C^+\otimes \mathbb{C}[U]), ~ e_i(cx)=ce_i(x)=\mu(c\otimes e_i(x))\in \mu(C^+\otimes \mathbb{C}[U])$$ $$ f_i(cx)=(f_i(c)-h_i(c)x_i)x+h_i(c)x_ix+cf_i(x)= \mu((f_i(c)-h_i(c)x_i)\otimes x+h_i(c)\otimes x_ix+c\otimes f_i(x))\in \mu(C^+\otimes \mathbb{C}[U]).$$ The lemma is proved. \end{proof} Since the image of $\mu$ contains a generating set for $A\otimes B$ as a $\mathfrak{g}$-module algebra (see Remark \ref{rem:nilpotentf}), we conclude that $\mu$ is surjective. Hence $\mu$ is an isomorphism and so, by Theorem \ref{thm:factoring}, $A\otimes B$ is adapted with $\nu_{\bf i}(A\otimes B\setminus\{0\})=\nu_{\bf i}(\mathbb{C}[U]\setminus\{0\})$ for all ${\bf i}\in R(w_\mathrm{o})$.\qed

https://arxiv.org/abs/1101.2302

Inverse spectral problems for Dirac operators on a finite interval

We consider the direct and inverse spectral problems for Dirac operators that are generated by the differential expressions $$ \mathfrak t_q:=\frac{1}{i}[I&0 0&-I]\frac{d}{dx}+[0&q q^*&0] $$ and some separated boundary conditions. Here $q$ is an $r\times r$ matrix-valued function with entries belonging to $L_2((0,1),\mathbb C)$ and $I$ is the identity $r\times r$ matrix. We give a complete description of the spectral data (eigenvalues and suitably introduced norming matrices) for the operators under consideration and suggest an algorithm of reconstructing the potential $q$ from the corresponding spectral data.

\section{Introduction} \label{} Direct and inverse spectral problems for Dirac and Sturm--Liouville operators are the objects of interest in plenty of papers. In 1966, M. Gasymov and B. Levitan solved the inverse spectral problem for Dirac operators on a half-line by using the spectral function \cite{GasLev1} and the scattering phase \cite{GasLev2}. Their investigations were continued and further developed in many directions. By now, the direct and inverse spectral problems for Dirac operators with potentials from different classes have been solved. For instance, the Dirac operators on a finite interval with continuous potentials were considered in \cite{GasDzab}, \cite{Malamud2} (reconstructing from two spectra), the ones on a half-line were treated in \cite{Malamud} (complete description of the spectral measures and the reconstruction procedure). The case of potentials belonging to $L_p(0,1)$, $p\ge1$, was considered in \cite{MykDirac} (reconstructing from two spectra and from one spectrum and the norming constants based on the Krein equation). The Weyl--Titchmarsh $m$-functions were used in \cite{ClarkGesz}, \cite{KisMakGesz} to recover the Dirac operators acting in $L_2(\mathbb R_+,\mathbb C^{2r})$. More general canonical systems on $\mathbb R$ were considered in \cite{Sakh1}, \cite{Sakh2}. The matrix-valued Weyl--Titchmarsh functions were recently used in \cite{Korot} for the characterization of vector-valued Sturm--Liouville operators on the unit interval. There are many other interesting papers concerning the direct and inverse spectral problems for Dirac and Sturm--Liouville operators besides those mentioned here. We refer the reader to the extensive bibliography cited in \cite{Malamud2}--\cite{sturm} for further results on that subject. The aim of the present paper is to extend the results of the recent paper \cite{sturm} by Ya. Mykytyuk and N. Trush concerning the inverse spectral problems for Sturm--Liouville operators with matrix-valued potentials to the case of Dirac operators on a finite interval with square-summable potentials. \subsection{Setting of the problem} \label{} Let $M_r$ denote the Banach algebra of $r\times r$ matrices with complex entries, which we identify with the Banach algebra of linear operators $\mathbb C^r\to\mathbb C^r$ endowed with the standard norm. We write $I=I_r$ for the unit element of $M_r$ and $M_r^+$ for the set of all matrices $A\in M_r$ such that $A=A^*\ge0$. Also we use the notations $$ \mathbb H:=L_2((0,1),\mathbb C^r)\times L_2((0,1),\mathbb C^r),\quad \mathfrak Q:=L_2((0,1),M_r). $$ Let $q\in\mathfrak Q$. Denote \begin{equation}\label{tqdef} \vartheta:=\frac{1}{i}\begin{pmatrix}I&0\\0&-I\end{pmatrix},\quad \bq:=\begin{pmatrix}0&q\\q^*&0\end{pmatrix} \end{equation} and consider the differential expression \begin{equation}\label{DiffExpr} \mathfrak t_q:=\vartheta\frac{d}{dx}+\bq \end{equation} on the domain $ D(\mathfrak t_q)=\{y=(y_1, \ y_2)^\top \ | \ y_1,y_2\in W_2^1((0,1),\mathbb C^r) \}$, where $W_2^1$ is the Sobolev space. The object of our investigation is a self-adjoint Dirac operator $T_q$ that is generated by the differential expression (\ref{DiffExpr}) and the separated boundary conditions $y_1(0)=y_2(0)$, $y_1(1)=y_2(1)$: $$ T_qy=\mathfrak t_q(y), \quad D(T_q)=\{y\in D(\mathfrak t_q) \ | \ y_1(0)=y_2(0), \ y_1(1)=y_2(1)\}. $$ The function $q\in\mathfrak Q$ will be conventionally called the \emph{potential} of $T_q$. The spectrum $\sigma(T_q)$ of the operator $T_q$ consists of countably many isolated real eigenvalues of finite multiplicity, accumulating only at $+\infty$ and $-\infty$. We denote by $\lambda_j(q)$, $j\in\mathbb Z$, the pairwise distinct eigenvalues of the operator $T_q$ labeled in increasing order so that $\lambda_0(q)\le0<\lambda_1(q)$: $$ \sigma(T_q)=\{\lambda_j(q)\}_{j\in\mathbb Z}. $$ Denote by $m_q$ the Weyl--Titchmarsh function of the operator $T_q$ that is defined as in \cite{ClarkGesz}. The function $m_q$ is a matrix-valued meromorphic Herglotz function (i.e. $\Im m_q(\lambda)\ge0$ whenever $\Im\lambda>0$), and $\{\lambda_j(q)\}_{j\in\mathbb Z}$ is the set of its poles. We put $$ \alpha_j(q):=-\underset{\lambda=\lambda_j(q)}\res m_q(\lambda),\quad j\in\mathbb Z, $$ and call $\alpha_j(q)$ the \emph{norming matrix} of the operator $T_q$ corresponding to the eigenvalue $\lambda_j(q)$. Note that the multiplicity of the eigenvalue $\lambda_j(q)$ of $T_q$ equals $\rank\alpha_j(q)$ and that $\alpha_j(q)\ge0$ for all $j\in\mathbb Z$. We call the collection $\mathfrak a_q:=((\lambda_j(q),\alpha_j(q)))_{j\in\mathbb Z}$ the \emph{spectral data} of the operator $T_q$, and the matrix-valued measure $$ \mu_q:=\sum\limits_{j=-\infty}^{\infty}\alpha_j(q)\delta_{\lambda_j(q)} $$ is called its \emph{spectral measure}. Here $\delta_\lambda$ is the Dirac delta-measure centered at the point $\lambda$. In particular, if $q\equiv0$ then $$ \mu_0=\sum\limits_{n=-\infty}^{\infty}I\delta_{\pi n}. $$ The aim is to give a complete description of the class $\mathfrak A:=\{\mathfrak a_q \ | \ q\in\mathfrak Q\}$ of spectral data for Dirac operators under consideration, which is equivalent to the description of the class $\mathfrak M:=\{\mu_q \ | \ q\in\mathfrak Q\}$ of spectral measures, and to suggest an efficient method of reconstructing the potential $q$ from the corresponding spectral data $\mathfrak a_q$. \subsection{{Main results}\label{subsec.12}} We start from the description of spectral data for operators under consideration. In what follows $\mathfrak a$ will stand for an arbitrary sequence $((\lambda_j,\alpha_j))_{j\in\mathbb Z}$, in which $(\lambda_j)_{j\in\mathbb Z}$ is a strictly increasing sequence of real numbers such that $\lambda_0\le0<\lambda_1$ and $\alpha_j$ are non-zero matrices in $M_r^+$. By $\mu^{\mathfrak a}$ we denote the matrix-valued measure given by \begin{equation}\label{MuAdef} \mu^{\mathfrak a}:=\sum\limits_{j=-\infty}^\infty \alpha_j\delta_{\lambda_j}. \end{equation} We partition the real axis into pairwise disjoint intervals $\Delta_n$, $n\in\mathbb Z$: $$ \Delta_n:=\left(\pi n-\tfrac{\pi}{2},\pi n+\tfrac{\pi}{2}\right],\ n \in \mathbb Z. $$ A complete description of the class $\mathfrak A$ is given by the following theorem. \begin{theorem}\label{Th1} In order that a sequence $\mathfrak a=((\lambda_j,\alpha_j))_{j\in\mathbb Z}$ should belong to $\mathfrak A$ it is necessary and sufficient that the following conditions are satisfied: \begin{itemize} \item[$(A_1)$]$\sup\limits_{n\in \mathbb Z}\sum\limits_{\lambda_j\in\Delta_n}1<\infty$, \ $\sum\limits_{n\in \mathbb Z}\sum\limits_{\lambda_j\in\Delta_n}|\lambda_j-\pi n|^2<\infty$, \ $\sum\limits_{n\in\mathbb Z}\|I-\sum\limits_{\lambda_k\in\Delta_n}\alpha_k\|^2<\infty$; \item[$(A_2)$]$\exists N_0\in\mathbb N \ \ \forall N\in \mathbb N:\ (N\ge N_0)\Rightarrow \sum\limits_{n=-N}^{N}\sum\limits_{\lambda_j\in\Delta_n}\rank \alpha_j=(2N+1)r$; \item[$(A_3)$]the system of functions $\{ de^{i\lambda_jt}\ | \ j\in\mathbb{Z},\ d\in\mathrm{Ran}\ \alpha_j \}$ is complete in $L_2((-1,1),\mathbb C^r)$. \end{itemize} \end{theorem} By definition, every $\mathfrak a\in\mathfrak A$ forms the spectral data for Dirac operator $T_q$ with some $q\in\mathfrak Q$. It turns out that this spectral data determine the potential $q$ uniquely: \begin{theorem}\label{Th2} The mapping $\mathfrak Q\owns q\mapsto\mathfrak a=\mathfrak a_q\in\mathfrak A$ is bijective. \end{theorem} We base our algorithm of reconstructing the potential $q$ from the corresponding spectral data $\mathfrak a_q$ on Krein's accelerant method. \begin{definition} We say that a function $H\in L_2((-1,1),M_r)$ is an accelerant if for every $a\in[0,1]$ the integral equation $$ f(x)+\int\limits_0^aH(x-t)f(t)dt=0 $$ has only trivial solution in $L_2((0,1),\mathbb C^r)$. We denote the set of accelerants by $\mathfrak H_2$ and endow it with the metric of the space $L_2((-1,1),M_r)$. \end{definition} We set $\mathfrak H_2^s:=\{H\in\mathfrak H_2 \ | \ H(x)^*=H(-x)\ \textrm{a.e. for}\ x\in(-1,1)\}$. The spectral data of the operator $T_q$ generate Krein's accelerant as explained in the following theorem. \begin{theorem}\label{Th3} Take a sequence $\mathfrak a=((\lambda_j,\alpha_j))_{j\in\mathbb Z}$ satisfying the asymptotics $(A_1)$, and set $\mu:=\mu^{\mathfrak a}$. Then the limit \begin{equation}\label{HMudef} H_\mu(x)=\lim\limits_{n\to\infty}\int\limits_{-\pi\left(n-\frac{1}{2}\right)}^ {\pi\left(n+\frac{1}{2}\right)}e^{2i\lambda x}d(\mu-\mu_0)(\lambda) \end{equation} exists in the topology of the space $L_2((-1,1),M_r)$. If, in addition, $(A_3)$ holds, then the function $H_\mu$ is an accelerant and belongs to $\mathfrak H_2^s$. \end{theorem} By virtue of Theorem \ref{Th1}, any $\mathfrak a\in\mathfrak A$ satisfies the conditions $(A_1)-(A_3)$. In addition, if $q\in\mathfrak Q$ and $\mathfrak a=\mathfrak a_q$, then $\mu^{\mathfrak a}=\mu_q$. Therefore according to Theorem \ref{Th3} we can define the mapping $q\mapsto\Upsilon(q):=H_{\mu_q}$ acting from $\mathfrak Q$ to $\mathfrak H_2^s$, and in order to solve the inverse spectral problem for the operator $T_q$ we have to find the inverse mapping $\Upsilon^{-1}$. As in \cite{sturm}, it can be done using the Krein equation. It is known that for all $H\in\mathfrak H_2$ the Krein equation \begin{equation}\label{KreinEq} R(x,t)+H(x-t)+\int\limits_0^xR(x,s)H(s-t)ds=0,\quad (x,t)\in\Omega^+, \end{equation} where $\Omega^+:=\{(x,t) \ | \ 0\le t\le x\le 1\}$, has a unique solution $R_H$ in the class $L_2(\Omega^+,M_r)$. Moreover, if we extend $R_H$ by zero to the triangle $\Omega^-:=\{(x,t) \ | \ 0\le x<t\le 1\}$, we obtain that $R_H\in G_2(M_r)$ (see \ref{add.1}). Thus we can define the mapping $\Theta:\mathfrak H_2^s\to \mathfrak Q$ given by \begin{equation}\label{ThetaDef} \Theta(H):=iR_H(\cdot,0). \end{equation} The following theorem explains how to solve the inverse spectral problem for the operator $T_q$. \begin{theorem}\label{Th4} $\Upsilon^{-1}=\Theta$. In particular, if $q\in\mathfrak Q$, $\mathfrak a=\mathfrak a_q$, $\mu=\mu^{\mathfrak a}$, then \begin{equation}\label{QafterTheta} q=\Theta(H_\mu). \end{equation} \end{theorem} According to this theorem the reconstruction algorithm can proceed as follows. Given $\mathfrak a\in\mathfrak A$ we construct the matrix-valued measure $\mu:=\mu^{\mathfrak a}$ via (\ref{MuAdef}), which generates the accelerant $H:=H_\mu$ via (\ref{HMudef}). Solving the Krein equation (\ref{KreinEq}) we find the function $R_H$, which gives us $q$ via the formulas (\ref{QafterTheta}) and (\ref{ThetaDef}). That $q$ is the function looked for follows from the fact that the Dirac operator $T_q$ has the spectral data $\mathfrak a$ we have started with. We visualize the reconstruction algorithm by means of the following diagram: $$ \mathfrak a \xrightarrow[s_1]{(\ref{MuAdef})} \mu^{\mathfrak a}=:\mu \xrightarrow[s_2]{(\ref{HMudef})}H_\mu=:H \xrightarrow[s_3]{(\ref{KreinEq})}R_H \xrightarrow[s_4]{(\ref{ThetaDef})}\Theta(H)=q. $$ Here $s_j$ denotes the step number $j$. Steps $s_1$, $s_2$, $s_4$ are trivial. The basic and non-trivial step is $s_3$. \begin{remark} One can also consider the case of more general separated self-adjoint boundary conditions. Denote by $T_{q,a,b}$ the operator generated by the differential expression (\ref{DiffExpr}) and the boundary conditions $$ ay(0)=0,\quad by(1)=0, $$ where $a$ and $b$ are $r\times2r$ matrices with complex entries such that (see \cite{ClarkGesz}) $$ aa^*=bb^*=I,\quad a\vartheta a^*=b\vartheta b^*=0. $$ For the operator $T_{q,a,b}$, the analogues of Theorems \ref{Th1}--\ref{Th4} can be proved, but their formulations are more complicated since the spectrum of the non-perturbed operator $T_{0,a,b}$ has a more involved structure. Namely, it consists of $2r$ eigenvalue sequences of the form $(\lambda^0_j+2\pi k)_{k\in\mathbb{Z}}$, $j=1,\dots,2r$, counting multiplicities. The authors plan to consider the case of general (not necessarily separated) boundary conditions in a forthcoming paper. \end{remark} \section{Direct spectral analysis} In this section we study the properties of the spectral data for operators under consideration. \subsection{Basic properties of the operator $T_q$} Here we prove self-adjointness of $T_q$, construct its resolvent and the resolution of identity. Let $\lambda\in\mathbb C$. For an arbitrary $q\in\mathfrak Q$ denote by $u_q=u_q(\cdot,\lambda)\in W_2^1((0,1),M_{2r})$ a solution of the Cauchy problem \begin{equation}\label{probl1} \vartheta \tfrac{d}{dx}u+\bq u=\lambda u,\quad u(0,\lambda)=I_{2r}, \end{equation} where $\vartheta$ and $\bq$ are defined via (\ref{tqdef}). Note that if $q\equiv0$ then \begin{equation}\label{Phi0} u_0(x,\lambda)=\begin{pmatrix}e^{i\lambda x}I&0\\0&e^{-i\lambda x}I\end{pmatrix}. \end{equation} Denote \begin{equation}\label{PhiPsiDef} \varphi_q(\cdot,\lambda):=u_q(\cdot,\lambda)\vartheta a^*,\quad \psi_q(\cdot,\lambda):=u_q(\cdot,\lambda)a^*, \end{equation} where $$ a:=\tfrac{1}{\sqrt{2}}\begin{pmatrix}I,&-I\end{pmatrix}, $$ and set $s(\lambda,q):=a\varphi_q(1,\lambda)$, $c(\lambda,q):=a\psi_q(1,\lambda)$, $m_q(\lambda):=-s(\lambda,q)^{-1}c(\lambda,q)$. We call $m_q$ \emph{the Weyl--Titchmarsh} function of the operator $T_q$. Some basic properties of the objects just introduced are described in the following lemma. \begin{lemma}\label{Th5} \begin{itemize} \item[$(i)$]For every $q\in\mathfrak Q$ there exists a unique matrix-valued function $K_q\in G_2^+(M_{2r})$ such that for any $\lambda\in\mathbb C$ and $x\in[0,1]$, \begin{equation}\label{varphiRepr} \varphi_q(x,\lambda)=\varphi_0(x,\lambda)+\int\limits_0^x K_q(x,s)\varphi_0(s,\lambda)ds; \end{equation} \item[$(ii)$]the mapping $\mathfrak Q\owns q\mapsto K_q\in G_2^+(M_{2r})$ is continuous; \item[$(iii)$]the matrix-valued functions $\lambda\mapsto s(\lambda,q)$ and $\lambda\mapsto c(\lambda,q)$ are entire and allow the representations \begin{equation}\label{slambdarepr} s(\lambda,q)=(\sin\lambda) I+\int\limits_{-1}^{1}e^{i\lambda t}g_1(t)dt,\quad c(\lambda,q)=(\cos\lambda) I+\int\limits_{-1}^{1}e^{i\lambda t}g_2(t)dt, \end{equation} where $g_1$ and $g_2$ are some (depending on $q$) functions from the space $L_2((-1,1),M_r)$; \item[$(iv)$]for every $q\in\mathfrak Q$ the following relation holds: \begin{equation}\label{PhiPsiRel} -\psi_q(x,\lambda)\varphi_q(x,\overline\lambda)^*+\varphi_q(x,\lambda)\psi_q(x,\overline\lambda)^*\equiv\vartheta. \end{equation} \end{itemize} \end{lemma} \begin{proof} Let us fix $q\in\mathfrak Q$ and set $q_1:=-\Im q=-\frac{1}{2}(q-q^*)$, $q_2:=\Re q=\frac{1}{2}(q+q^*)$. Consider the Cauchy problem $$ B\tfrac{d}{dx}v+Qv=\lambda v, \quad v(0,\lambda)=I_{2r}, $$ where $$ B:=\begin{pmatrix}0&I\\-I&0\end{pmatrix},\quad Q:=\begin{pmatrix}q_1&q_2\\q_2&-q_1\end{pmatrix}. $$ It follows from \cite{cauchy} that this problem has a unique solution $v_q=v_q(\cdot,\lambda)$ in $W_2^1((0,1),M_{2r})$ and that $v_q(\cdot,\lambda)$ can be represented in the form \begin{equation}\label{Vform} v_q(x,\lambda)=e^{-\lambda xB} + \int\limits_0^x P^+(x,s)e^{-\lambda (x-2s) B}ds + \int\limits_0^x P^-(x,s)e^{\lambda (x-2s) B}ds, \end{equation} where $ e^{xB}=(\cos x)I_{2r}+(\sin x)B$. Note that $\vartheta=W^{-1}BW$ and $\bq=W^{-1}Q W$, where $W$ is the unitary matrix $$ W=\frac{1}{\sqrt{2}}\begin{pmatrix}I&-iI\\-iI&I\end{pmatrix}. $$ Therefore the function $u_q(\cdot,\lambda)=W^{-1}v_q(\cdot,\lambda)W$ solves the Cauchy problem (\ref{probl1}). Note that $$ e^{xB}J=Je^{-xB},\quad J=\begin{pmatrix}0&I\\I&0\end{pmatrix}. $$ Using now (\ref{Vform}) and performing some calculations we easily obtain that $$ \varphi_q(x,\lambda)=\varphi_0(x,\lambda)+\int\limits_0^x K_q(x,s)\varphi_0(s,\lambda)ds, $$ where $K_q(x,t)=W^{-1}P_Q(x,t)W$ and $$ P_Q(x,t)=\tfrac{1}{2}\left\{P^+\left(x,\tfrac{x-t}{2}\right)+P^+\left(x,\tfrac{x+t}{2}\right)J +P^-\left(x,\tfrac{x-t}{2}\right)J+P^-\left(x,\tfrac{x+t}{2}\right)\right\}. $$ It follows from \cite{cauchy} that the function $P_Q$ belongs to $G_2^+(M_{2r})$ and that the mapping $ L_2((0,1),M_{2r})\owns Q\mapsto P_Q\in G_2^+(M_{2r}) $ is continuous. Therefore the first two statements of the present lemma will be proved if we prove the uniqueness of the representation (\ref{varphiRepr}), but this can be easily done repeating the proof given in \cite{cauchy}. Now let us prove $(iii)$. By virtue of the definition of $s(\lambda,q)$ and the representation (\ref{varphiRepr}) we obtain that $$ s(\lambda,q)=(\sin\lambda)I+\int\limits_0^1aK_q(1,s)\varphi_0(s,\lambda)ds $$ and simple calculations yield the formula for $s(\lambda)$ in (\ref{slambdarepr}) with some $g_1\in L_2((-1,1),M_r)$. Having noted that $\psi_q(x,\lambda)=u_q(x,\lambda)a^*=W^{-1}v_q(x,\lambda)Wa^*$ and taking into consideration (\ref{Vform}) we can analogously obtain the formula for $c(\lambda)$. It remains to prove $(iv)$. A direct verification shows that $$ \tfrac{d}{dx}\{u_q(x,\overline\lambda)^*\vartheta u_q(x,\lambda)\}\equiv0, $$ and therefore we obtain the relation $u_q(x,\overline\lambda)^*\vartheta u_q(x,\lambda)\equiv\vartheta$. From this equality we obtain that $\vartheta u_q(x,\lambda)\vartheta u_q(x,\overline\lambda)^*\equiv -I_{2r}$, and thus $u_q(x,\lambda)\vartheta u_q(x,\overline\lambda)^*\equiv\vartheta$. Having noted that $\vartheta=a^*a\vartheta+\vartheta a^*a$ we conclude that $$ u_q(x,\lambda)a^*a\vartheta u_q(x,\overline\lambda)^*+u_q(x,\lambda)\vartheta a^*a u_q(x,\overline\lambda)^*\equiv\vartheta, $$ which proves the relation (\ref{PhiPsiRel}). \end{proof} For $\lambda\in\mathbb C$ denote by $\Phi_q(\lambda)$ the operator acting from $\mathbb C^r$ to $\mathbb H$ by the formula \begin{equation}\label{PhiOperDef} [\Phi_q(\lambda)c](x):=\varphi_q(x,\lambda)c. \end{equation} Taking into consideration (\ref{varphiRepr}) we obtain that \begin{equation}\label{PhiKrel} \Phi_q(\lambda)=(\mathscr I+\mathscr K_q)\Phi_0(\lambda),\quad \lambda\in\mathbb C, \end{equation} where $\mathscr K_q$ is an integral operator with kernel $K_q$ and $\mathscr I$ is the identity operator in $\mathscr B(\mathbb H)$, which is the algebra of bounded linear operators acting in $\mathbb H$. Note that since $K_q$ belongs to $G_2^+(M_{2r})$, the operator $\mathscr K_q$ belongs to $\mathscr G_2^+(M_{2r})$ (see \ref{add.1}), and hence it is a Volterra operator (see \cite{kreinvolterra}). Some properties of the operators $\Phi_q(\lambda)$ and the Weyl--Titchmarsh function $m_q(\lambda)$ are formulated in the following lemma. \begin{lemma}\label{Lemma1} Let $q\in\mathfrak Q$. Then the following statements hold: \begin{itemize} \item[$(i)$]the operator function $\lambda\mapsto\Phi_q(\lambda)$ is analytic in $\mathbb C$; moreover, for $\lambda\in\mathbb C$ \begin{equation}\label{KerRanPhi} \ker\Phi_q(\lambda)=\{0\},\quad \Ran\Phi_q(\lambda)^*=\mathbb C^r, \end{equation} \begin{equation}\label{KerPhiRel} \ker(T_q-\lambda\mathscr I)=\Phi_q(\lambda)\ker s(\lambda,q). \end{equation} \item[$(ii)$]the operator functions $\lambda\mapsto s(\lambda,q)^{-1}$ and $$ \lambda\mapsto m_q(\lambda)=-s(\lambda,q)^{-1}c(\lambda,q) $$ are meromorphic in $\mathbb C$; moreover, $m_0(\lambda)=-\cot\lambda I$ and \begin{equation}\label{mAsymp} \|m_q(\lambda)+\cot\lambda I\|=o(1) \end{equation} as $\lambda\to\infty$ within the domain $\mathcal O=\{z\in\mathbb C \ | \ \forall n\in\mathbb Z \ |z-\pi n|>1\}$. \end{itemize} \end{lemma} \begin{proof} The proof of this lemma is analogous to the proof of Lemma 2.3 in \cite{sturm}. \end{proof} Finally, basic properties of the operator $T_q$ are described in the following theorem. \begin{theorem}\label{Th6} Let $q\in\mathfrak Q$. Then the following statements hold: \begin{itemize} \item[$(i)$]the operator $T_q$ is self-adjoint; \item[$(ii)$]the spectrum $\sigma(T_q)$ of $T_q$ consists of isolated real eigenvalues and $$ \sigma(T_q)=\{\lambda \ | \ \ker s(\lambda,q)\neq\{0\}\}; $$ \item[$(iii)$]let $\lambda_j=\lambda_j(q)$ and let $P_{j,q}$ be the orthogonal projector on $\ker(T_q-\lambda\mathscr I)$, then $$ \sum\limits_{j=-\infty}^{\infty}P_{j,q}=\mathscr I; $$ \item[$(iv)$]the norming matrices $\alpha_j=\alpha_j(q)$ satisfy the relations $\alpha_j\ge0$, $j\in\mathbb Z$; moreover, for all $j\in\mathbb Z$ we have $$ P_{j,q}=\Phi_q(\lambda_j)\alpha_j\Phi_q^*(\lambda_j), $$ where $\Phi_q^*(\lambda):=[\Phi_q(\lambda)]^*$. \end{itemize} \end{theorem} \begin{proof} A direct verification shows that the operator $T_q$ is symmetric. Take an arbitrary $\lambda$ such that the matrix $s(\lambda,q)$ is non-singular, and let $f\in\mathbb H$. Then the function $$ g(x)=[\mathscr T(\lambda)f](x):=\psi_q(x,\lambda)\int\limits_0^x \varphi_q(t,\overline\lambda)^*f(t)dt+ \varphi_q(x,\lambda)\int\limits_x^1 \psi_q(t,\overline\lambda)^*f(t)dt $$ belongs to the domain of differential expression $\mathfrak t_q$ and solves the Cauchy problem $$ \mathfrak t_q(g)=\lambda g+f,\quad ag(0)=0, $$ as can be directly verified using (\ref{PhiPsiRel}). A generic solution of this problem takes the form $h=\varphi_q(\cdot,\lambda)c+\mathscr T(\lambda) f$, $c\in\mathbb C^r$. The choice $$ c=m_q(\lambda)\int\limits_0^1 \varphi_q(t,\overline\lambda)^*f(t)dt $$ gives that $ah(1)=0$, i.e. the boundary conditions $h_1(0)=h_2(0)$, $h_1(1)=h_2(1)$ are satisfied. This implies that $\lambda$ is a resolvent point of the operator $T_q$, and the resolvent of $T_q$ is given by $$ (T_q-\lambda\mathscr I)^{-1}=\Phi_q(\lambda)m_q(\lambda)\Phi_q^*(\overline\lambda)+\mathscr T(\lambda). $$ Since $\mathscr T(\lambda)$ is a Hilbert--Schmidt operator, the operator $T_q$ has a compact resolvent, and therefore the statements $(i)-(iii)$ are proved. Recall that $-\alpha_j(q)$ is a residue of the Weyl--Titchmarsh function at the point $\lambda_j=\lambda_j(q)$, $j\in\mathbb Z$. Taking $\varepsilon>0$ small enough we obtain that $$ P_{j,q}=-\frac{1}{2\pi i}\oint\limits_{|\lambda-\lambda_j| =\varepsilon} (T_q-\lambda\mathscr I)^{-1}d\zeta =\Phi_q(\lambda_j) \alpha_j(q)\Phi_q^*(\lambda_j) $$ for every $j\in\mathbb Z$. By virtue of (\ref{KerRanPhi}) we obtain that $\alpha_j(q)\ge0$ for all $j\in\mathbb Z$, and the statement $(iv)$ is also proved. \end{proof} \subsection{Description of the spectral data: the necessity part} Here we show that if $q\in\mathfrak Q$, then the spectral data $\mathfrak a_q$ satisfy the conditions $(A_1)-(A_3)$, which is the necessity part of Theorem \ref{Th1}. \subsubsection{The condition $(A_1)$} In the sequel we shall use the following notations. If $(\lambda_{j})_{j\in\mathbb Z}$ is a strictly increasing sequence of non-negative real numbers and $(\alpha_j)_{j\in\mathbb Z}$ is a sequence in $M^+_r$, then \begin{equation}\label{eq.215} \beta_n:= I-\sum\limits_{\lambda_k\in\Delta_n}\alpha_k, \qquad \widetilde\lambda_j:=\lambda_j-\pi n, \quad \lambda_j\in\Delta_n,\quad n\in\mathbb Z, \end{equation} with $\Delta_n$ being defined in Subsection \ref{subsec.12}. We start from the condition $(A_1)$, which describes the asymptotics of spectral data. \begin{theorem}\label{a1th1} Let $q\in\mathfrak Q.$ Then for the sequence $\mathfrak a=\mathfrak a_q$ the condition $(A_1)$ holds. \end{theorem} \begin{proofsketch} The proof of this theorem is analogous to the proof in \cite{sturm}, and therefore we give here only its sketch. Let $q\in\mathfrak Q$ and $\lambda_j=\lambda_j(q)$, $\alpha_j=\alpha_j(q)$ for $j\in\mathbb Z$. The eigenvalues $\lambda_j$ are zeros of the sine-type function $\lambda\mapsto s(\lambda)$ (see\eqref{slambdarepr}) that belongs to the following class of functions $ \mathbb C\to M_r$: $$ \mathcal{F}_f(\lambda):=\sin{\lambda}I + \int_{-1}^1 f(t)e^{i\lambda t}\,d t, \qquad \lambda\in\mathbb C, $$ where $f\in L_2((-1,1),M_r)$. It is shown in \cite{zeros} that the set of zeros of a function $\det \mathcal{F}_f$, with ${\mathcal F}_f$ as above, can be indexed (counting multiplicities) by the set $\mathbb Z$ so that the corresponding sequence $(\omega_n)_{n\in\mathbb Z}$ of its zeros has the asymptotics $$ \omega_{kr+j}=\pi k +\widehat\omega_{j,k},\qquad k\in\mathbb{Z},\quad j=0,\dots,r-1, $$ where the sequences $(\widehat\omega_{j,k})_{k\in\mathbb{Z}}$ belong to $\ell_2(\mathbb Z)$. Therefore, \begin{equation}\label{eq.216} \sup\limits_{n\in\mathbb Z}\sum\limits_{\lambda_j\in\Delta_n} 1<\infty, \qquad \sum_{n\in\mathbb Z}\sum\limits_{\lambda_j\in\Delta_n}| \widetilde\lambda_j|^2<\infty, \end{equation} and thus it is left to prove only that (see (\ref{eq.215})) $$ \sum\limits_{n=-\infty}^\infty\|\beta_n\|^2<\infty. $$ It can be done in exactly the same way as in \cite{sturm}. \end{proofsketch} \subsubsection{The condition $(A_2)$} We start from proving the following lemma, which is an analogue of Lemma 2.12 in \cite{sturm}. \begin{lemma}\label{Lemma3} Assume that $q\in\mathfrak Q$, and let $\mathfrak a$ be a collection satisfying the asymptotics $(A_1)$. For $j\in\mathbb Z$ set $\hat P_j:=\Phi_q(\lambda_j)\alpha_j\Phi_q^*(\lambda_j)$. Then the series $\sum_{j\in\mathbb Z}\hat P_j$ converges in the strong operator topology and \begin{equation}\label{PSeries} \sum\limits_{n=-\infty}^\infty\|P_{n,0}-\sum\limits_{\lambda_j\in\Delta_n}\hat P_j\|^2<\infty. \end{equation} \end{lemma} \begin{proofsketch} Let the assumptions of the present lemma hold, and let $\mathfrak a=((\lambda_j,\alpha_j))_{j\in\mathbb Z}$. Using (\ref{PhiKrel}) and the fact that $\mathscr K_q$ is a Hilbert--Schmidt operator, it can be observed that $$ \sum\limits_{n\in\mathbb Z}\sum\limits_{\lambda_j\in \Delta_n}\|\Phi_q(\pi n)-\Phi_0(\pi n)\|^2<\infty. $$ Since $\|\Phi_q(\lambda_j)-\Phi_q(\pi n)\|\le C|\widetilde\lambda_j|$ ($\lambda_j\in \Delta_n$, $n\in\mathbb Z$) for some $C>0$, \begin{equation}\label{PhiSeries} \sum\limits_{n\in\mathbb Z}\sum\limits_{\lambda_j\in \Delta_n}\|\Phi_q(\lambda_j)-\Phi_0(\pi n)\|^2<\infty. \end{equation} From ~\eqref{PhiSeries} we easily obtain that \begin{equation}\label{PhiFSeries} \sum\limits_{j\in\mathbb Z}\|\Phi_q^*(\lambda_j)f\|^2<\infty \end{equation} for all $f\in\mathbb H$. Indeed, it is enough to note that $ \sum_{n\in\mathbb Z}\|\Phi_0(\pi n)^*f\|^2= \|f\|^2$, $f\in\mathbb H$, and that $\sup\limits_{n\in \mathbb Z}\sum_{\lambda_j\in\Delta_n}1<\infty$. Taking into account that the sequence $(\alpha_j)_{j\in\mathbb Z}$ is bounded we conclude that $ \sum_{j\in\mathbb Z}\|\alpha_j\Phi^*(\lambda_j)f\|^2<\infty$, $f\in\mathbb H$. Moreover, it can also be shown that for every sequence $c\in l_2(\mathbb Z,\mathbb C^r)$ the series $\sum_{j\in\mathbb Z}\Phi_q(\lambda_j)c_j$ is convergent, which justifies the convergence of $\sum_{j\in\mathbb Z}\hat P_j$. Now let us prove (\ref{PSeries}). Recall that $P_{n,0}=\Phi_0(\pi n)\Phi_0^*(\pi n)$. By virtue of the definition of $\beta_n$ we obtain that $$ P_{n,0}=\Phi_0(\pi n)\beta_n\Phi_0^*(\pi n)+\sum\limits_{\lambda_j\in\Delta_n}\Phi_0(\pi n)\alpha_j\Phi_0^*(\pi n), $$ and thus we can write $$ P_{n,0}-\sum\limits_{\lambda_j\in\Delta_n} \hat P_j=\Phi_0(\pi n)\beta_n\Phi_0^*(\pi n)+\sum\limits_{\lambda_j\in\Delta_n}[\Phi_0(\pi n)\alpha_j\Phi_0^*(\pi n)-\Phi_q(\lambda_j)\alpha_j\Phi_q^*(\lambda_j)]. $$ Thus, since the sequences $(\Phi_q(\lambda_j))$ and $(\alpha_j)$ are bounded, we obtain that $$ \|P_{n,0}-\sum\limits_{\lambda_j\in\Delta_n}\hat P_j\|^2\le C_1\|\beta_n\|^2+C_2\sum\limits_{\lambda_j\in\Delta_n}\|\Phi_q(\lambda_j)-\Phi_0(\pi n)\|^2, $$ where $C_1$ and $C_2$ are non-negative constants independent of $n$. Taking now into consideration (\ref{PhiSeries}) and $(A_1)$, we obtain (\ref{PSeries}). \end{proofsketch} \noindent The following lemma is proved in \cite{sturm} (Lemma B.1). \begin{lemma}\label{Lemma4} Suppose that H is a Hilbert space. Let $(P_n)_{n=1}^\infty$ and $(G_n)_{n=1}^\infty$ be sequences of pairwise orthogonal projectors of finite rank in $H$ such that $\sum_{n=1}^\infty P_n=\sum_{n=1}^\infty G_n=\mathscr I_H$, where $\mathscr I_H$ is the identity operator in $H$, and let $\sum_{n=1}^\infty\|P_n-G_n\|^2<\infty$. Then there exists $N_0\in\mathbb N$ such that for all $N\ge N_0$, $$ \sum\limits_{n=1}^N \rank P_n=\sum\limits_{n=1}^N \rank G_n. $$ \end{lemma} We use Lemmas \ref{Lemma3} and \ref{Lemma4} to prove $(A_2)$. If $\mathfrak a=\mathfrak a_q$, then the operators $\hat P_j$, $j\in\mathbb Z$, from Lemma \ref{Lemma3} coincide with the orthogonal projectors $P_{j,q}$ corresponding to the eigenvalues $\lambda_j$ (see Theorem \ref{Th6}). Since $\{P_{j,q}\}$, $j\in\mathbb Z$, forms a complete system of orthogonal projectors, by virtue of Lemmas \ref{Lemma3} and \ref{Lemma4} we justify that $$ \sum\limits_{n=-N}^N \rank P_{n,0}=\sum\limits_{n=-N}^N \sum\limits_{\lambda_j\in\Delta_n} \rank P_{j,q} $$ for $N\ge N_0$. Taking into consideration (\ref{KerRanPhi}), we obtain that $\rank P_{j,q}=\rank \alpha_j$ and $\rank P_{n,0}=r$ for all $j,n\in\mathbb Z$, and thus we justify that the condition $\mathrm{(A_2)}$ is satisfied. \subsubsection{The operators $\mathscr U_{\mathfrak a,q}$} Before proving the condition $(A_3)$ we have to introduce some operators that play an important role below. Let $q\in\mathfrak Q$, and let $\mathfrak a=((\lambda_j,\alpha_j))_{j\in\mathbb Z}$ be any collection satisfying the asymptotics $(A_1)$. Construct the operator $\mathscr U_{\mathfrak a,q}:\mathbb H\to\mathbb H$ by the formula \begin{equation}\label{UaqDef} \mathscr U_{\mathfrak a,q}:=\sum\limits_{j\in\mathbb Z}\Phi_q(\lambda_j)\alpha_j\Phi_q^*(\lambda_j). \end{equation} By virtue of Lemma \ref{Lemma3} the operator $\mathscr U_{\mathfrak a,q}$ is continuous, and, since $\alpha_j\ge0$ for all $j\in\mathbb Z$ (see Theorem \ref{Th6}), it is also non-negative. In particular, \begin{equation}\label{UaqI} \mathscr U_{\mathfrak a_q,q}=\mathscr I, \end{equation} as follows from Theorem \ref{Th6}. Now we are going to show that the operator $\mathscr U_{\mathfrak a,q}$ is the sum of the identity one and a compact one. We start from proving the following lemma. \begin{lemma}\label{LemmaHconv} Let $\mathfrak a$ be any collection satisfying the condition $(A_1)$. Then the limit (\ref{HMudef}) exists in the topology of the space $L_2((-1,1),M_r)$, and the following relation holds: \begin{equation}\label{HadjReal} H_\mu(x)^*=H_\mu(-x). \end{equation} \end{lemma} \begin{proof} Taking into consideration the definitions of measures $\mu$ and $\mu_0$ it is easy to observe that the function $H:=H_\mu$ can be rewritten as \begin{equation}\label{HSeries} H(x)=\sum\limits_{n\in\mathbb Z}\left\{\left(\sum\limits_{\lambda_j\in\Delta_n}e^{2i\lambda_jx}\alpha_j\right)-e^{2i\pi nx}I\right\}, \end{equation} and thus we have to show that the series (\ref{HSeries}) is convergent in $L_2((-1,1),M_r)$. Note that \begin{equation}\label{eq.226} \left(\sum\limits_{\lambda_j\in\Delta_n}e^{2i\lambda_jx}\alpha_j\right)-e^{2i\pi nx}I= e^{2i\pi nx}\gamma_n(x) +x e^{2i\pi nx}\eta_n-e^{2i\pi nx}\beta_n, \end{equation} where $$ \gamma_n(x):=\sum\limits_{\lambda_j\in\Delta_n} (e^{2i\widetilde\lambda_jx}-1-2i\widetilde\lambda_jx)\alpha_j, \qquad \eta_n:=\sum\limits_{\lambda_j\in\Delta_n} 2i\widetilde\lambda_j\alpha_j, $$ $\beta_n$ and $\widetilde\lambda_j$ are given by (\ref{eq.215}). Since the sequence $(\alpha_j)_{j\in\mathbb Z}$ is bounded and $|e^z-1-z|\le|z|^2e^{|z|}$, $z\in\mathbb C$, in view of the condition $\mathrm{(A_1)}$ we obtain that $$ \sum\limits_{n\in\mathbb Z} \sup_{x\in[0,1]} \|\gamma_n(x)\|< \infty, \qquad \sum\limits_{n\in\mathbb Z} \|\eta_n\|^2 <\infty, \qquad\sum\limits_{n\in\mathbb Z} \|\beta_n\|^2<\infty. $$ Therefore, taking into consideration (\ref{eq.226}) it is easy to observe that the series (\ref{HSeries}) is convergent in the topology of the space $L_2((-1,1),M_r).$ The relation (\ref{HadjReal}) follows directly from the formula (\ref{HSeries}). \end{proof} For $H\in L_2((-1,1),M_r)$ denote \begin{equation}\label{FHdef} F_H(x,t):=\frac{1}{2}\begin{pmatrix}H\left(\frac{x-t}{2}\right)&H\left(\frac{x+t}{2}\right)\\ H^\sharp\left(\frac{x+t}{2}\right)&H^\sharp\left(\frac{x-t}{2}\right)\end{pmatrix}, \end{equation} where $H^\sharp(x):=H(-x)$. Note that $F_H\in G_2(M_{2r})$. \begin{proposition} Let $\mathfrak a$ be any collection satisfying the asymptotics $(A_1)$, and set $\mu:=\mu^{\mathfrak a}$, $H:=H_\mu$. Then \begin{equation}\label{UF} \mathscr U_{\mathfrak a,0}=\mathscr I+\mathscr F_H, \end{equation} where $\mathscr F_H$ is a Hilbert--Schmidt operator with kernel $F_H$, i.e. $$ (\mathscr F_Hf)(x)=\int\limits_0^1F_H(x,s)f(s)ds,\quad f\in L_2((0,1),\mathbb C^{2r}). $$ \end{proposition} \begin{proof} The proof can be obtained by direct verification. \end{proof} \subsubsection{The condition $(A_3)$} Now let us prove that for all $q\in\mathfrak Q$ the spectral data $\mathfrak a_q$ satisfy the condition $(A_3)$. In view of \eqref{UaqI}, this fact directly follows from the following lemma. \begin{lemma}\label{A3Lemma} Let $q\in\mathfrak Q$, and let $\mathfrak a=((\lambda_j,\alpha_j))_{j\in\mathbb Z}$ be any collection satisfying the asymptotics $(A_1)$. Then \begin{equation}\label{A3Equiv} (A_3)\Longleftrightarrow \mathscr U_{\mathfrak a,q}>0. \end{equation} \end{lemma} \begin{proof} Taking into consideration the relation (\ref{PhiKrel}), we obtain that \begin{equation}\label{UaqUa0} \mathscr U_{\mathfrak a,q}=(\mathscr I+\mathscr K_q)\mathscr U_{\mathfrak a,0}(\mathscr I+\mathscr K_q^*). \end{equation} Since the operator $\mathscr I+\mathscr K_q$ is a homeomorphism of the space $\mathbb H$, it is enough to prove the equivalence (\ref{A3Equiv}) only for the case $q=0$. Since $\mathscr U_{\mathfrak a,0}\ge0$ and the operator $\mathscr F_H$ in (\ref{UF}) is compact, we obtain that $\mathscr U_{\mathfrak a,0}>0$ if and only if $\ker\mathscr U_{\mathfrak a,0}=\{0\}$. Thus it is enough to prove the equivalence \begin{equation}\label{A3aux} (A_3)\Longleftrightarrow \ker\mathscr U_{\mathfrak a,0}=\{0\}. \end{equation} Set $\mathcal X:=\{ e^{i\lambda_jt}d\ | \ j\in\mathbb{Z},\ d\in\mathrm{Ran}\ \alpha_j \}\subset L_2((-1,1),\mathbb C^r)$ and note that the condition $(A_3)$ is equivalent to the equality $\mathcal X^\perp=\{0\}$. Consider the unitary transformation $ U: L_2((-1,1),\mathbb C^r) \to \mathbb H$ given by $$ (Uf)(x):= (f(-x),f(x))\in\mathbb C^{2r}, \qquad x\in(0,1). $$ It follows from the definitions of $\mathscr U_{\mathfrak a,0} $ and $\Phi_0(\lambda)$ that $$ \ker\mathscr U_{\mathfrak a,0}=\bigcap\limits_{j\in\mathbb Z}\ker\alpha_j\Phi_0^*(\lambda_j) =(U\mathcal X)^\perp=U\mathcal X^\perp, $$ and therefore (\ref{A3aux}) is proved. \end{proof} \section{Inverse spectral problem} In this section we solve the inverse spectral problem for the operator $T_q$. We show that if a collection $\mathfrak a$ satisfies the conditions $(A_1)-(A_3)$, then $\mathfrak a=\mathfrak a_q$ for some $q\in\mathfrak Q$ and suggest a method of constructing such $q$. \subsection{The Krein accelerant: proof of Theorem \ref{Th3}} Here we prove Theorem \ref{Th3}, i.e. we show that any collection $\mathfrak a$ satisfying the conditions $(A_1)$ and $(A_3)$ generates the Krein accelerant belonging to $\mathfrak H_2^s$. Since the convergence of (\ref{HMudef}) was already proved, it is left to prove only the following lemma. \begin{lemma} Let $\mathfrak a$ satisfy the conditions $(A_1)$ and $(A_3)$, $\mu:=\mu^{\mathfrak a}$, $H:=H_\mu$. Then the function $H$ is an accelerant and belongs to $\mathfrak H_2^s$. \end{lemma} \begin{proof} Let us prove that $H$ belongs to $\mathfrak H_2$. It is enough to prove that the operator $\mathscr I+\mathscr H$ is positive in $L_2((0,1),\mathbb C^r)$, where $\mathscr H$ is given by \begin{equation}\label{WienerHopfDef} (\mathscr Hf)(x)=\int\limits_0^1H(x-t)f(t)dt. \end{equation} By virtue of (\ref{UF}) and Lemma \ref{A3Lemma}, the condition $(A_3)$ implies the \linebreak positivity of the operator $\mathscr I+\mathscr F_H$ in the space $L_2((0,1),\mathbb C^{2r})$. Consider the unitary transformation $V:L_2((0,1),\mathbb C^{2r})\to L_2((0,1),\mathbb C^r)$, $$ (Vf)(t)=\begin{cases}\sqrt{2}f_2(1-2t),&t\in(0,1/2],\\\sqrt{2}f_1(2t-1),&t\in(1/2,1).\end{cases} $$ A direct verification shows that $\mathscr I+\mathscr H=V(\mathscr I+\mathscr F_H)V^{-1}$, and thus the operators $\mathscr I+\mathscr H$ and $\mathscr I+\mathscr F_H$ are unitary equivalent. Therefore $I+\mathscr H>0$ in $L_2((0,1),\mathbb C^r)$. It is left to notice that by virtue of the relation (\ref{HadjReal}) the function $H$ belongs to $\mathfrak H_2^s$. \end{proof} \subsection{Factorization of $\mathscr U_{\mathfrak a,0}$} Given a collection $\mathfrak a$ satisfying the asymptotics $(A_1)$, put $\mu:=\mu^{\mathfrak a}$, $H:=H_\mu$, and construct the operator $\mathscr U_{\mathfrak a,0}$ via (\ref{UaqDef}). In this subsection we show that $\mathscr U_{\mathfrak a,0}$ admits a factorization in $\mathscr G_2(M_r)$. Some statements concerning the theory of factorization can be found in \ref{add.2}. \subsubsection{Basic properties of $R_H$} Recall that for $H\in\mathfrak H_2$ we denote by $R_H$ the solution of the Krein equation (\ref{KreinEq}). Here we prove some basic properties of $R_H$. \begin{lemma}\label{Lemma2} \begin{itemize} \item[$(i)$]If $H\in\mathfrak H_2$, then $R_H\in G_2^+(M_r)$ and the mapping $$ \mathfrak H_2\owns H\mapsto R_H\in G_2^+(M_r) $$ is continuous; \item[$(ii)$]if $H\in\mathfrak H_2^s$, then $H^\sharp\in\mathfrak H_2^s$ and \begin{equation}\label{LemmaRel} R_{H^\sharp}(\cdot,0)=[R_H(\cdot,0)]^*; \end{equation} \item[$(iii)$]the mapping $\Theta:\mathfrak H_2^s\to \mathfrak Q$ given by $\Theta(H):=iR_H(\cdot,0)$ is continuous; \item[$(iv)$]if $H\in\mathfrak H_2\cap C^1([-1,1],M_r)$, then $R_H\in C^1(\Omega^+,M_r)$. \end{itemize} \end{lemma} \begin{proof} We start from proving $(i)$. Suppose that $H\in\mathfrak H_2$. Denote by $\mathscr H$ the operator given by (\ref{WienerHopfDef}), and set $\mathscr H^a:=\chi_a\mathscr H\chi_a$ (see \ref{add.2}). Since $H\in\mathfrak H_2$, $\ker(\mathscr I+\mathscr H^a)=\{0\}$ for all $a\in[0,1]$, and the operator $\mathscr I+\mathscr H^a$ is invertible in the algebra $\mathscr B(L_2)$ of bounded linear operators acting in $L_2((0,1),\mathbb C^r)$. Since $\mathscr H^a$ depends continuously on $a\in[0,1]$, the mapping $[0,1]\owns a\mapsto(\mathscr I+\mathscr H^a)^{-1}\in\mathscr B(L_2)$ is continuous. Denote by $\Gamma_{a,H}$ the kernel of the integral operator $-\mathscr H^a(\mathscr I+\mathscr H^a)^{-1}$. Since $\mathscr H$ is a Hilbert--Schmidt operator, the mapping $$ [0,1]\times\mathfrak H_2\owns(a,H)\mapsto\Gamma_{a,H}\in L_2((0,1)^2,M_r) $$ is also continuous. For $(x,t)\in\Omega^+$ put \begin{equation}\label{bRHrepr} \hat R_H(x,t):=\int\limits_0^xH(x-y)H(y-t)dy+ \int\limits_0^x\int\limits_0^xH(x-u)\Gamma_{x,H}(u,v)H(v-t)dvdu. \end{equation} It is easily seen that the mapping $\mathfrak H_2\owns H\mapsto \hat R_H\in C(\Omega^+,M_r)$ is continuous. A direct verification shows that the function \begin{equation}\label{RHrepr} R_H(x,t):=\begin{cases}\hat R_H(x,t)-H(x-t),&(x,t)\in\Omega^+,\\0,&(x,t)\in\Omega^-\end{cases} \end{equation} solves the Krein equation (\ref{KreinEq}). Therefore $R_H$ belongs to $G_2^+(M_r)$, and the mapping $\mathfrak H_2\owns H\mapsto R_H\in G_2^+(M_r)$ is continuous. Let us prove $(ii)$. Assume that $H\in\mathfrak H_2^s$. First let us show that $H^\sharp\in\mathfrak H_2^s$. Construct the integral operator $\mathscr H^\sharp$ via the formula (\ref{WienerHopfDef}) with $H^\sharp$ instead of $H$. The operators $\mathscr I+\mathscr H^\sharp$ and $\mathscr I+\mathscr H$ are unitary equivalent under the unitary transformation $f(t)\mapsto f(1-t)$. Therefore $I+\mathscr H>0$ if and only if $I+\mathscr H^\sharp>0$, and thus $H^\sharp\in\mathfrak H_2^s$. The equality (\ref{LemmaRel}) can be easily verified having noted that $\Gamma_{a,H}(a-x,a-t)=\Gamma_{a,H^\sharp}(x,t)$ and $\Gamma_{a,H}(x,t)=[\Gamma_{a,H}(t,x)]^*$ for all $x,t\in[0,a]$ and for all $H\in\mathfrak H_2^s$. The continuity of $\Theta$ easily follows from its definition and continuity of the mapping $H\mapsto R_H$, and thus the statement $(iii)$ is proved. It is left to prove $(iv)$. It follows from \cite[Chapter IV]{kreinvolterra} that if $H\in\mathfrak H_2\cap C^1([-1,1],M_r)$, then the function $a\mapsto\Gamma_{a,H}(u,v)$ is continuously differentiable for $a\ge\max\{u,v\}$. Therefore taking into consideration (\ref{RHrepr}) and (\ref{bRHrepr}) we conclude that $R_H\in C^1(\Omega^+,M_r)$. \end{proof} \subsubsection{The GLM equation} Here we establish structure of the solution of Gelfand--Levitan--Marchenko (GLM) equation. \begin{lemma}\label{Th7} Let $H\in L_2((-1,1),M_r)$. If $H\in\mathfrak H_2^s$, then the GLM equation \begin{equation}\label{GLM} L(x,t)+F_H(x,t)+\int\limits_0^xL(x,s)F_H(s,t)ds=0,\quad (x,t)\in\Omega^+ \end{equation} has a unique solution in the class $L_2(\Omega^+,M_{2r})$; moreover, this solution \linebreak belongs to $G_2^+(M_{2r})$ and takes the form \begin{equation}\label{KHform} L_H(x,t)=\frac{1}{2}\begin{pmatrix}R_H\left(x,\frac{x+t}{2}\right)&R_H\left(x,\frac{x-t}{2}\right)\\ R_{H^\sharp}\left(x,\frac{x-t}{2}\right)&R_{H^\sharp}\left(x,\frac{x+t}{2}\right)\end{pmatrix}. \end{equation} \end{lemma} \begin{proof} A direct verification shows that the function $L_H$ given by (\ref{KHform}) solves the GLM equation (\ref{GLM}). Since $F_H\in G_2(M_{2r})$ and $L_H\in L_2(\Omega^+,M_{2r})$, the results of \ref{add.2} yield that $L_H\in G_2^+(M_{2r})$. \end{proof} \begin{remark}\label{Lrem} Since the mapping $H\mapsto R_H$ is continuous, it is easily seen that the mapping $H\mapsto L_H$ given by (\ref{KHform}) is also continuous. \end{remark} \subsubsection{Theorem on factorization of $\mathscr U_{\mathfrak a,0}$} Main result of the present subsection is the following theorem. \begin{theorem}\label{Th8} Let $\mathfrak a=((\lambda_j,\alpha_j))_{j\in\mathbb Z}$ be a collection satisfying the conditions $(A_1)$ and $(A_3)$, $\mu:=\mu^{\mathfrak a}$, $H:=H_\mu$. Set $q:=\Theta(H)$. Then \begin{equation}\label{Ua0Fact} \mathscr U_{\mathfrak a,0}=(\mathscr I+\mathscr K_q)^{-1}(\mathscr I+\mathscr K_q^*)^{-1}, \end{equation} where $\mathscr K_q$ is an integral operator with kernel $K_q$ (see Lemma \ref{Th5}). \end{theorem} \begin{proof} By virtue of Lemma \ref{Th7}, the function $L_H$ given by (\ref{KHform}) solves the GLM equation (\ref{GLM}). Thus, as follows from \ref{add.2}, the equality $$ \mathscr U_{\mathfrak a,0}=(\mathscr I+\mathscr L_H)^{-1}(\mathscr I+\mathscr L_H^*)^{-1} $$ takes place, where $\mathscr L_H$ is an integral operator with kernel $L_H$. Therefore it is left to show that $\mathscr L_H=\mathscr K_q$, i.e. it suffices to show that \begin{equation}\label{LK} L_H=K_q,\quad q=\Theta(H). \end{equation} Notice that it is enough to prove (\ref{LK}) only for the case $H\in\mathfrak H_2^s\cap C^1([-1,1],M_r)$. Indeed, the set $\mathfrak H_2^s\cap C^1([-1,1],M_r)$ is dense everywhere in $\mathfrak H_2^s$, and the mappings $q\mapsto K_q$, $\Theta$ and $H\mapsto L_H$ are continuous (see Lemma \ref{Th5}, Lemma \ref{Lemma2} and Remark \ref{Lrem} respectively). Let $H\in\mathfrak H_2^s\cap C^1([-1,1],M_r)$. Taking into consideration Lemma \ref{Th5}, it is easily seen that the equality (\ref{LK}) is equivalent to the fact that the function \begin{equation}\label{PhiDef} \varphi(x,\lambda):=\varphi_0(x,\lambda)+\int\limits_0^x L_H(x,s)\varphi_0(s,\lambda)ds \end{equation} solves the Cauchy problem \begin{equation}\label{PhiEq} \vartheta\tfrac{d}{dx}\varphi+\bq\varphi=\lambda\varphi,\quad \varphi(0,\lambda)=\vartheta a^*. \end{equation} Thus it is left to prove (\ref{PhiEq}). Let us introduce the auxiliary functions $$ \tH:=\begin{pmatrix}H&0\\0&H^\sharp\end{pmatrix},\quad \tR_H:=\begin{pmatrix}R_H&0\\0&R_{H^\sharp}\end{pmatrix}. $$ The definitions of $R_H$ and $\tR_H$ yield that the following relation holds: \begin{equation}\label{KreinEq1} \tR_H(x,t)+\tH(x-t)+\int\limits_0^x\tR_H(x,s)\tH(s-t)ds=0,\quad (x,t)\in\Omega^+. \end{equation} Moreover, by virtue of Lemma \ref{Lemma2} we obtain that $\tR_H\in C^1(\Omega^+,M_{2r})$. Noting that $$ L_H(x,t)=\tfrac{1}{2}\left\{\tR_H\left(x,\tfrac{x+t}{2}\right)+\tR_H\left(x,\tfrac{x-t}{2}\right)J\right\} $$ and $$ J\varphi_0(x,\lambda)=\varphi_0(-x,\lambda),\quad J=\begin{pmatrix}0&I\\I&0\end{pmatrix}, $$ we can rewrite (\ref{PhiDef}) as $$ \varphi(x,\lambda)=\varphi_0(x,\lambda)+\int\limits_0^x \tR_H(x,x-s)\varphi_0(x-2s,\lambda)ds. $$ From this equality, taking into consideration that $\vartheta\tfrac{d}{dx}\varphi_0(x,\lambda)-\lambda\varphi_0(x,\lambda)=0$, we easily obtain that $$ \vartheta\tfrac{d}{dx}\varphi(x,\lambda)+\bq(x)\varphi(x,\lambda)-\lambda\varphi(x,\lambda) =\{\vartheta \tR_H(x,0)J\varphi_0(x,\lambda)+\bq(x)\varphi_0(x,\lambda)\} $$ \begin{equation}\label{SecondInt} +\int\limits_0^x\left\{\vartheta \tfrac{\partial}{\partial x}[ \tR_H(x,x-s)]+\bq(x)\tR_H(x,x-s)\right\}\varphi_0(x-2s,\lambda)ds. \end{equation} Taking into consideration (\ref{LemmaRel}) we conclude that $\bq(x)=-\vartheta\tR_H(x,0)J$ and thus the relation (\ref{SecondInt}) can be rewritten as \begin{eqnarray*} &\vartheta\frac{d}{dx}\varphi(x,\lambda)+\bq(x)\varphi(x,\lambda)-\lambda\varphi(x,\lambda) \\ &=\vartheta\int\limits_0^x\left\{\frac{\partial}{\partial x}[ \tR_H(x,x-s)]-\tR_H(x,0)J \tR_H(x,s)J\right\}\varphi_0(x-2s,\lambda)ds. \end{eqnarray*} If we show that \begin{equation}\label{ZeroRel} \tfrac{\partial}{\partial x}[ \tR_H(x,x-s)]-\tR_H(x,0)J \tR_H(x,s)J=0 \end{equation} for $(x,t)\in\Omega^+$, then (\ref{PhiEq}) will be proved. Let us show (\ref{ZeroRel}). From (\ref{KreinEq1}) we obtain that $$ \tR_H(x,x-t)+\tH(t)+\int\limits_0^x\tR_H(x,x-s)\tH(t-s)ds=0,\quad (x,t)\in\Omega^+, $$ and differentiating this expression with respect to $x$ we can write \begin{equation}\label{DifEq} \tfrac{\partial}{\partial x}\tR_H(x,x-t)+ \tR_H(x,0)\tH(t-x)+\int\limits_0^x\tfrac{\partial}{\partial x}[ \tR_H(x,x-s)]\tH(t-s)ds=0. \end{equation} Now we multiply the relation (\ref{KreinEq1}) by $\tR_H(x,0)J$ from the left and by $J$ from the right, and write \begin{eqnarray}\nonumber \tR_H(x,0)J\tR_H(x,t)J+\tR_H(x,0)J\tH(x-t)J \\\label{MultEq} +\int\limits_0^x\tR_H(x,0)J\tR_H(x,s)\tH(s-t)Jds=0 \end{eqnarray} for $(x,t)\in\Omega^+$. Subtracting now (\ref{DifEq}) from (\ref{MultEq}) and taking into consideration that $\tH(x)J=J\tH(-x)$, we obtain that the function $$ X(x,t)=\tfrac{\partial}{\partial x}[ \tR_H(x,x-s)]-\tR_H(x,0)J \tR_H(x,s)J $$ solves the equation $$ X(x,t)+\int\limits_0^xX(x,s)\tH(s-t)ds=0,\quad (x,t)\in\Omega^+. $$ Since $\tR_H\in C^1(\Omega^+,M_{2r})$, $X\in C(\Omega^+,M_{2r})$ and thus by virtue of Lemma \ref{LemmaIntEq} and the relation (\ref{KreinEq1}) we conclude that $X(x,t)\equiv0$. Therefore the relation (\ref{ZeroRel}) follows, and the proof is complete. \end{proof} \begin{remark} Let $\mathfrak a=((\lambda_j,\alpha_j))_{j\in\mathbb Z}$ be a collection satisfying the conditions $(A_1)$ and $(A_3)$, $\mu:=\mu^{\mathfrak a}$, $H:=H_\mu$, $q:=\Theta(H)$. Then from the equalities (\ref{Ua0Fact}) and (\ref{UaqUa0}) we obtain that \begin{equation}\label{eq1} \mathscr U_{\mathfrak a,q}=\mathscr I. \end{equation} \end{remark} \subsection{Description of the spectral data: the sufficiency part} In this subsection we show that if a collection $\mathfrak a$ satisfies the conditions $(A_1)-(A_3)$, then it belongs to $\mathfrak A$, i.e. that $\mathfrak a=\mathfrak a_q$ for some $q\in\mathfrak Q$. This is the sufficiency part of Theorem \ref{Th1}. We start from proving the following lemma. \begin{lemma}\label{Lemma5} Let $\mathfrak a=((\lambda_j,\alpha_j))_{j\in\mathbb Z}$ be a collection satisfying the conditions $(A_1)-(A_3)$, $\mu:=\mu^{\mathfrak a}$, $H:=H_\mu$, $q:=\Theta(H)$. Set $$ \hat P_j:=\Phi_q(\lambda_j)\alpha_j\Phi_q^*(\lambda_j). $$ Then the collection $\{\hat P_j\}_{j\in\mathbb Z}$ forms a complete system of pairwise orthogonal projectors. \end{lemma} \noindent In order to prove this lemma we need the following additional statements, that are proved in \cite{sturm} (Lemmas B.2 and B.3 respectively). \begin{lemma}\label{LemmaPr1} Let $H$ be a Hilbert space. Assume that $(A_j)_{j=1}^\infty$ is a sequence in $\mathscr B(H)$ and that $(G_j)_{j=1}^\infty$ is a sequence of pairwise orthogonal projectors such that the following statements hold: \begin{itemize} \item[(i)] the series $\sum_{j=1}^\infty A_j$ converges in the strong operator topology to an operator $A$; \item[(ii)] the orthogonal projector $G:=\mathscr I_H-\sum_{j=1}^\infty G_j$ is of finite rank; \item[(iii)] $\sum_{j=1}^\infty\|A_j-G_j\|^2<1$ and $\rank A_j\le\rank G_j<\infty$ for every $j\in\mathbb N$. \end{itemize} Then $\mathrm{codim}\ \mathrm{Ran}\ A\ge\rank G$. \end{lemma} \begin{lemma}\label{LemmaPr2} Let $H$ be a Hilbert space, and let $\{A_j\}_{j=0}^n$ be a set of self-adjoint operators from the algebra $\mathscr B(H)$ that are of finite rank for $j\neq0$. If $$ \sum\limits_{j=0}^nA_j=\mathscr I_H,\quad \sum\limits_{j=1}^n\rank A_j\le\mathrm{codim}\ \mathrm{Ran}\ A_0, $$ then $\{A_j\}_{j=0}^n$ is the set of pairwise orthogonal projectors. \end{lemma} \noindent We use these statements to prove Lemma \ref{Lemma5}. \begin{prooflemma34} It follows from Lemma \ref{Lemma3} that the series $\sum_{j\in\mathbb Z}\hat P_j$ converges in the strong operator topology, and in view of (\ref{UaqDef}) and (\ref{eq1}) we obtain that $$ \sum\limits_{j=-\infty}^\infty\hat P_j=\mathscr I. $$ Thus, it is enough to show that the operators $\hat P_j$, $j\in\mathbb Z$ are pairwise orthogonal projectors. Denote $$ A_n:=\sum\limits_{\lambda_j\in\Delta_n}\hat P_j. $$ By virtue of Lemma \ref{Lemma3} we obtain that $\sum_{n=-\infty}^\infty\|P_{n,0}-A_n\|^2<\infty$, and therefore there exists an $N_0\in\mathbb N$ such that $\sum_{|n|>N_0}\|P_{n,0}-A_n\|^2<1$. Moreover, due to the conditions $\mathrm{(A_1)}$ and $\mathrm{(A_2)}$ we conclude that $N_0$ can be taken so large that \begin{equation}\label{eq3} \sum\limits_{n=-N}^N\sum\limits_{\lambda_j\in\Delta_n}\rank \alpha_j=(2N+1)r,\qquad N\ge N_0, \end{equation} \begin{equation}\label{eq2} \|\sum\limits_{\lambda_j\in\Delta_n}\alpha_j-I\|<1,\qquad |n|\ge N_0. \end{equation} First let us show that \begin{equation}\label{eq4} \sum\limits_{\lambda_j\in\Delta_n}\rank \alpha_j=r,\qquad |n|\ge N_0. \end{equation} Indeed, it follows from (\ref{eq2}) that $\sum_{\lambda_j\in\Delta_n}\rank \alpha_j\ge r$ and from (\ref{eq3}) that $ \sum_{\lambda_j\in\Delta_n}\rank \alpha_j+\sum_{\lambda_j\in\Delta_{-n}}\rank \alpha_j=2r $ for $|n|\ge N_0$, and thus we obtain (\ref{eq4}). Fix $N>N_0$ and set $$ P:=\mathscr I-\sum\limits_{|n|>N}P_{n,0},\quad A:=\sum\limits_{|n|>N}A_n. $$ Since $\rank \hat P_j=\rank \alpha_j$ for all $j\in\mathbb Z$ (which follows directly from the definition of $\hat P_j$ and (\ref{KerRanPhi})), taking into consideration (\ref{eq4}) we conclude that $$ \rank A_n\le\sum\limits_{\lambda_j\in\Delta_n}\rank \hat P_j=\sum\limits_{\lambda_j\in\Delta_n}\rank \alpha_j=r=\rank P_{n,0},\qquad |n|>N_0. $$ Recalling also that $\{P_{n,0}\}_{j\in\mathbb Z}$ forms a complete system of pairwise orthogonal projectors, by virtue of Lemma \ref{LemmaPr1} we obtain that $$ \mathrm{codim}\ \Ran A\ge\rank P=(2N+1)r. $$ Moreover, $ A+\sum_{n=-N}^NA_n=\sum_{j\in\mathbb Z}\hat P_j=\mathscr I$ and \begin{eqnarray*} &\sum\limits_{n=-N}^N\rank A_n\le\sum\limits_{n=-N}^N\sum\limits_{\lambda_j\in\Delta_n}\rank \hat P_j \\ &=\sum\limits_{n=-N}^N\sum\limits_{\lambda_j\in\Delta_n}\rank \alpha_j=(2N+1)r\le\mathrm{codim}\ \mathrm{Ran}\ A. \end{eqnarray*} Therefore, since the operators $A$ and $A_n$, $|n|\le N$, are self-adjoint, by virtue of Lemma \ref{LemmaPr2} we obtain that the set $$ \{\hat P_j\ | \ \lambda_j\in \bigcup\limits_{n=-N}^N \Delta_n\} $$ is a set of pairwise orthogonal projectors. Since $N$ is arbitrary, we conclude that projectors $\{\hat P_j\}_{j\in\mathbb Z}$ are orthogonal ones. \end{prooflemma34} In order to prove the sufficiency part of Theorem \ref{Th1} it obviously suffices to find $q\in\mathfrak Q$ such that $\mathfrak a=\mathfrak a_q$. \begin{theorem}\label{Th9} Let $\mathfrak a=((\lambda_j,\alpha_j))_{j\in\mathbb Z}$ be a collection satisfying the conditions $(A_1)-(A_3)$, $\mu:=\mu^{\mathfrak a}$, $H:=H_\mu$, $q:=\Theta(H)$. Then $\mathfrak a=\mathfrak a_q$. \end{theorem} \begin{proof} It is enough to prove the relation \begin{equation}\label{PrRel} \mathrm{Ran}\ \hat P_j\subset\ker(T_q-\lambda_j\mathscr I),\quad j\in\mathbb Z. \end{equation} Indeed, taking into account the completeness of system $\{\hat P_j\}_{j\in\mathbb Z}$, from (\ref{PrRel}) we immediately conclude that $\lambda_j(q)=\lambda_j$ for all $j\in\mathbb Z$, where $\lambda_j(q)$ are the eigenvalues of $T_q$. From this equality and (\ref{PrRel}) we obtain the relation $P_{j,q}-\hat P_j\ge0$, $j\in\mathbb Z$, where $P_{j,q}$ are corresponding orthogonal projectors of $T_q$. However, by virtue of completeness of the systems $\{\hat P_j\}_{j\in\mathbb Z}$ and $\{P_{j,q}\}_{j\in\mathbb Z}$ we conclude that $ \sum_{j\in\mathbb Z}(P_{j,q}-\hat P_j)=0$, and therefore $P_{j,q}-\hat P_j=0$ for all $j\in\mathbb Z$. Therefore, taking into account Lemma \ref{Lemma5} and the definition of $\hat P_j$, we conclude that $$ \Phi_q(\lambda_j)\{\alpha_j(q)-\alpha_j\}\Phi_q^*(\lambda_j)=0, \quad j\in\mathbb Z, $$ and by virtue of (\ref{KerRanPhi}) we justify that $\alpha_j(q)=\alpha_j$, which, together with $\lambda_j(q)=\lambda_j$, means that $\mathfrak a=\mathfrak a_{q}$. Thus it only remains to prove (\ref{PrRel}). Due to the definition of $\Phi_q(\lambda)$ and (\ref{KerRanPhi}) we obtain that $ \mathrm{Ran}\ \hat P_j=\{\varphi_{q}(\cdot,\lambda_j)\alpha_jc\ | \ c\in\mathbb C^r\}$. From the other side, by virtue of Lemma \ref{Lemma1} we obtain that $$ \ker (T_q-\lambda_j \mathscr I)=\{ \varphi_q(\cdot,\lambda_j)c \ | \ a\varphi_q(1,\lambda_j)c=0 \}. $$ Therefore we conclude that it is enough to show that \begin{equation}\label{FinalRel} a\varphi_q(1,\lambda_j)\alpha_j=0. \end{equation} Let $j,k\in\mathbb Z$ and $c,d\in\mathbb C^r$. Then, taking into account that $\bq=\bq^*$, $\vartheta^*=-\vartheta$ and integrating by parts, we obtain that $$ \lambda_j(\Phi_q(\lambda_j)c\ | \ \Phi_q(\lambda_k)d)=(\vartheta\varphi_q(1,\lambda_j)c\ | \ \varphi_q(1,\lambda_k)d)+\lambda_k(\Phi_q(\lambda_j)c\ | \ \Phi_q(\lambda_k)d), $$ \begin{equation}\label{Rel1} \lambda_j\Phi_q(\lambda_k)^*\Phi_q(\lambda_j)- \lambda_k\Phi_q(\lambda_k)^*\Phi_q(\lambda_j)=\varphi_q(1,\lambda_k)^*\vartheta\varphi_q(1,\lambda_j). \end{equation} Since $\hat P_k\hat P_j=0$ if $k\neq j$, we obtain that $ \Phi_q(\lambda_k)\alpha_k\Phi_q^*(\lambda_k)\Phi_q(\lambda_j)\alpha_j\Phi_q^*(\lambda_j)=0$, and by virtue of (\ref{KerRanPhi}) we conclude that $ \alpha_k\Phi_q^*(\lambda_k)\Phi_q(\lambda_j)\alpha_j=0$. Multiplying now (\ref{Rel1}) by $\alpha_k$ from the left and by $\alpha_j$ from the right we obtain that $$ \alpha_k\varphi_q(1,\lambda_k)^*\vartheta\varphi_q(1,\lambda_j)\alpha_j=0, $$ and therefore we can write \begin{equation}\label{Rel2} \left\{\sum\limits_{\lambda_k\in\Delta_n}(-1)^n\alpha_k\varphi_q(1,\lambda_k)^*\right\}\vartheta \varphi_q(1,\lambda_j)\alpha_j=0, \qquad \lambda_j\notin\Delta_n. \end{equation} Taking into account (\ref{varphiRepr}), it follows from the Riemann--Lebesgue lemma and the asymptotic behavior of the sequences $(\lambda_k)$ and $(\alpha_k)$ that $$ \lim\limits_{n\to\infty}\left\{\sum\limits_{\lambda_k\in\Delta_n}(-1)^n\varphi_q(1,\lambda_k)\alpha_k\right\}=\vartheta a^*, $$ and passing to the limit in (\ref{Rel2}) we obtain the relation (\ref{FinalRel}). \end{proof} \subsection{Potential reconstruction: proof of Theorems \ref{Th2} and \ref{Th4}} Finally, we prove Theorems \ref{Th2} and \ref{Th4}. \begin{proofth12} Suppose that $q_1,q_2\in\mathfrak Q$, and let $\mathfrak a_{q_1}=\mathfrak a_{q_2}$. Let us show that $q_1=q_2$. Write $\mathfrak a_{q_1}=\mathfrak a_{q_2}=:\mathfrak a$ for short, and set $\mu:=\mu^{\mathfrak a}$, $H:=H_\mu$. Then by virtue of Theorem \ref{Th8} the operator $\mathscr U_{\mathfrak a,0}=\mathscr I+\mathscr F_H$ admits a factorization, and we can write $$ \mathscr U_{\mathfrak a,0}=(\mathscr I+\mathscr K_{q_1})^{-1}(\mathscr I+\mathscr K_{q_1}^*)^{-1}=(\mathscr I+\mathscr K_{q_2})^{-1}(\mathscr I+\mathscr K_{q_2}^*)^{-1}. $$ Since any operator may admit at most one factorization of the above form (see \ref{add.2}), we conclude that $$ \mathscr K_{q_1}=\mathscr K_{q_2}. $$ It is left to notice that $\mathscr K_{q_1}=\mathscr K_{q_2}\ \Rightarrow\ q_1=q_2$. Taking into account (\ref{varphiRepr}) we conclude that $\mathscr K_{q_1}=\mathscr K_{q_2}\ \Rightarrow\ \varphi_{q_1}(\cdot,0)=\varphi_{q_2}(\cdot,0)=:\varphi$, and therefore we obtain $$ \vartheta\varphi'+\bq_1\varphi=\vartheta\varphi'+\bq_2\varphi=0, $$ and thus $\{\bq_1-\bq_2\}\varphi=0$. Thus it is left to show that for all $x$ the matrix $\varphi(x)$ is invertible. Assume the contrary. Then there exist $x_0\in(0,1]$ and $c\in\mathbb C^r\setminus\{0\}$ such that $\varphi(x_0)c=0$, and therefore the function $f=\varphi_{q_1}(\cdot,0)c$ is a non-zero solution of the Cauchy problem $\vartheta f'+\bq_1f=0$, $f(x_0)=0$. But this is in contradiction with the uniqueness theorem; thus $\varphi(x)$ is non-singular, and $\bq_1=\bq_2$. Besides this, by definition of the spectral data we obviously have $q_1=q_2\ \Rightarrow\ \mathfrak a_{q_1}=\mathfrak a_{q_2}$, and therefore we conclude that the mapping $\mathfrak Q\owns q\mapsto\mathfrak a_q\in\mathfrak A$ is bijective. \end{proofth12} \begin{proofth14} Theorem \ref{Th4} now directly follows from Theorems \ref{Th2} and \ref{Th9}. \end{proofth14} \section*{Acknowledgements} The authors are grateful to Rostyslav Hryniv for helpful discussions and valuable suggestions in preparing this manuscript.

https://arxiv.org/abs/2111.09837

Markov chains on hyperbolic-like groups and quasi-isometries

We propose the study of Markov chains on groups as a "quasi-isometry invariant" theory that encompasses random walks. In particular, we focus on certain classes of groups acting on hyperbolic spaces including (non-elementary) hyperbolic and relatively hyperbolic groups, acylindrically hyperbolic 3-manifold groups, as well as fundamental groups of certain graphs of groups with edge groups of subexponential growth. For those, we prove a linear progress result and various applications, and these lead to a Central Limit Theorem for random walks on groups quasi-isometric to the ones we consider.

\section{Introduction} Random walks on groups have been studied extensively since, at least, work of Kesten \cite{Kesten} in the fifties. In particular, there are various ways in which random walks have been examined in the context of geometric group theory, and most relevantly for us a substantial amount of work has been devoted to understanding random walks on hyperbolic groups and various generalisations (see below for some relevant references). Given the strength of the connections between random walks and geometric group theory, it is natural to wonder how random walks interact with a crucial notion in geometric group, namely that of quasi-isometry. Unfortunately, random walks and quasi-isometries are not compatible, as given a random walk on a group $G$ and another group $H$ quasi-isometric to $G$, there is no "corresponding" random walk on $H$ in any meaningful sense. The main goal of this paper is to suggest that this conflict can be resolved if one studies more general Markov chains rather than random walks. Indeed, more general Markov chains can be ``pushed forward" via quasi-isometries in a natural way. (We note that we cannot restrict to Markov chains that are finite-state in any meaningful way if we want this push-forward property; we explain this in Remark \ref{rmk:no_finite_state}.) To illustrate the usefulness of this approach, we now state a theorem about random walks (not Markov chains) that can be deduced from our initial results on Markov chains on groups together with known results about random walks. The groups that we can consider are those satisfying either of two assumptions (Assumptions \ref{assump:superdiv} and \ref{assump:graph} below) which we do not state in this introduction, but rather we list some of the groups to which they apply (we note that we only consider finitely generated groups in this paper): \begin{itemize} \item non-elementary hyperbolic and relatively hyperbolic groups, \item acylindrically hyperbolic 3-manifold groups, \item right-angled Artin groups whose defining graph is a tree of diameter at least 3, \item groups acting acylindrically and non-elementarily on a tree with nilpotent and undistorted edge stabilisers. \end{itemize} The theorem is a central limit theorem for the distance from the identity. More general random walks can be included, but we state it only for simple random walks only for now. \begin{theorem} \label{thm:clt_intro} Let $G$ be a group quasi-isometric to some group satisfying either of Assumption \ref{assump:superdiv} or Assumption \ref{assump:graph} below (e.g., a group from the list above), fix a word metric $d_G$ on $G$, and let $(Z_n)$ be a simple-random walk on $G$. Then there exist constants $L, \sigma >0$ such that we have the following convergence in distribution: $$\frac{d_{G}(1,Z_n)-Ln}{\sigma^2\sqrt{n}} \to \mathcal{N}(0,1),$$ where $\mathcal N(0,1)$ denotes the normal distribution with mean $0$ and variance $1$. \end{theorem} The interesting feature of the theorem is that it is a theorem about random walks on a group for which the only assumption is that it is quasi-isometric to a group of a certain kind. As we explain below, the proof uses more general Markov chains than random walks, and this should give an idea of the potential of the Markov chain approach. We should note that the groups satisfying Assumptions \ref{assump:superdiv} or Assumption \ref{assump:graph} are acylindrically hyperbolic, and there is a central limit theorem for random walks on acylindrically hyperbolic groups \cite{MathieuSisto}. However, it is not known whether being acylindrically hyperbolic is a quasi-isometry invariant, and in fact it is not even known whether it is a commensurability invariant, see \cite[Question 2]{MinasyanOsinErratum} for a discussion of this. We also note that the theorem does not hold for Markov chains, as those might not even have well-defined drift; an example of this is described in work in progress of the first author. This should also be taken as a warning that Markov chains can have more exotic behaviour than one might expect based on random walks. \subsection*{Literature on Markov chains and quasi-isometries} Before discussing our results, we should mention that Markov chains on quasi-isometric graphs have been considered in the literature before, see e.g. \cite{benjaminiLiouville,soardiyamasakiParabolicindex,hermonperes,DingPeres,BerryMixingtimeRobust} (the term ``rough isometry" is used in this context, rather than ``quasi-isometry"). It is shown in \cite{kanai, SaloffCosteCoulhon} that transience (resp. recurrence) of a graph is a quasi-isometric invariant. We are however not aware of any paper that aims to study quasi-isometric \emph{groups} from this perspective. While, for example, \cite{soardiyamasakiParabolicindex} studies a property of Markov chains which turns out to be quasi-isometry invariant, most of the papers mentioned above actually show that certain properties of Markov chains are not quasi-isometry invariants in a suitable sense. For example \cite{hermonperes, DingPeres} shows that the mixing time can vary drastically for (finite) quasi-isometric graphs (with bounded quasi-isometry constants). In a different direction, we should also mention that there are results in the literature on random walks on quasi-isometric groups. For example, the fact that (non-)amenability is a quasi-isometry invariant implies that having spectral radius for random walks less than 1 or equal to 1 are quasi-isometry invariants; a stronger result for amenable groups is proved in \cite{SaloffCostePittet}. \subsection*{Linear progress} The groups we will be dealing with all come with an action on a hyperbolic space. In the case of random walks, it turns out that there is one result in particular that is a very useful starting point for other applications, namely linear progress. This is a result of Maher-Tiozzo \cite{maher2015random} who proved that a (non-elementary) random walk on a group acting on a hyperbolic space $X$ makes linear progress in $X$. This result feeds into the proof of several others, for example on the translation length \cite{maher2015random}, random subgroups \cite{MS-rndsubgp,Abbott2021RandomWA}, various kinds of projections \cite{SistoTaylor}, and deviation from quasi-geodesics \cite{MathieuSisto}, as well as the already mentioned central limit theorem for acylindrically hyperbolic groups, and results on counting measures, see e.g. \cite{GTT:counting_lox}. Therefore, since we want to initiate the study of more general Markov chains on groups acting on hyperbolic spaces, it is natural to start by generalising the linear progress result. Unfortunately, one cannot just use the proof from \cite{maher2015random}, and not even the proof from \cite{MathieuSisto} for the case where the action on the hyperbolic space is acylindrical, or the related arguments from, e.g. \cite{sunderlandlinearprogress, sisto2013contracting, gouezel2021exponential, boulanger2020large}. The reasons for this are, for the arguments that use boundary theory, the fact that it is not even clear what should play the role of stationary measures, and for the arguments that do not, the fact that, when decomposing a Markov path in 2 or more parts, subsequent parts are not independent from the previous ones. We note moreover that since the Markov chains we consider are not finite-state in any meaningful way, they are not the same kind of Markov chains considered in e.g. \cite{CalegariFujiwara,GTT:counting_lox}, and the arguments from those papers do not apply here either. Still, the theorem below gives linear progress in $X$ in certain cases. The notion of tameness is defined and discussed in Subsection \ref{tameMarkov}, but it is not so important for this general discussion so we omit it here. \begin{theorem} \label{linear progress} Let $G$ be a group acting on a corresponding hyperbolic space $X$ as in either of Assumption \ref{assump:superdiv} or Assumption \ref{assump:graph}. Let $(w^o_n)_{n\geq 0}$ be a tame Markov chain in $G$, starting at an element $o\in G$ and consider any basepoint $x_0 \in X$. Then there exist constants $L, C>0$ such that $$\mathbb{P}\Big[ d_{X}(x_0,w^{o}_nx_0) \geq L n\Big] \geq 1 - C e^{-n/C}.$$ \end{theorem} Our proof, while inspired by that in \cite{MathieuSisto}, is different from either of the aforementioned proofs, and in fact we even need additional assumptions of a geometric nature. It is possible that our techniques can be pushed to deal with more cases, and we elaborate on this in the outline of the paper below. In particular, it would be very interesting to deal with general acylindrically hyperbolic groups, or at least with some other subfamilies. We record this in the following question. \begin{problem} Does the conclusion of Theorem \ref{linear progress} hold when $G$ is only assumed to act acylindrically and non-elementarily on $X$? What if in addition $G$ is a hierarchically hyperbolic group? \end{problem} Just as in the case of random walks, we can deduce various other results from Theorem \ref{linear progress}. While, as mentioned above, we had to use different techniques than for random walks to prove Theorem \ref{linear progress}, for the following applications we can use known arguments. The first item says in particular that (under the assumptions of the theorem) loxodromic elements are generic with respect to Markov chains. \begin{theorem} \label{thm:qgeod_intro} Let $G$ be a group acting on a hyperbolic space $X$, with a fixed word metric $d_G$. Let $(w^o_n)_{n\geq 0}$ be a tame Markov chain satisfying the conclusion of Theorem \ref{linear progress}. Then: \begin{itemize} \item (Translation length) There exist constants $L_1$, $C_1>0$ such that $$\mathbb{P}\big[ \tau(w^{o}_n) >L_1 n \big] \geq 1-C_1e^{-n / C_1} $$ where $\tau(h)$ is the translation length in $X$ of the element $h \in G$. \item (Deviation from quasi-geodesics) For every $D>0$, there exists $C_2>0$ such that for all $l>0$ and $0 \leq k \leq n$ we have : $$ \mathbb{P}\left[ \sup_{\alpha \in QG_{D}(o,w^{o}_n)} d_{G}(w^{o}_k, \alpha) \geq l \right] \leq C_2 e^{-l/C_1} $$ where $QG_{D}(a,b)$ is the set of all $(D,D)$- quasi-geodesics from $a$ to $b$ (with respect to $d_{G}$). \end{itemize} \end{theorem} Note that the second item makes no reference to $X$, and actually this will be the crucial statement about Markov chains that we need to prove the central limit theorem, Theorem \ref{thm:clt_intro}. The sketch of proof is as follows. Consider a group $G$ and a random walk on it as in the statement. Then the random walk can be pushed-forward to a Markov chain (probably not a random walk) on another group $H$ quasi-isometric to $G$. In $H$ we know that, in the sense of the second item above, Markov paths stay close to quasi-geodesics. We deduce that sample paths of the random walk in $G$ stay close (in the same sense) to geodesics, and at this point we can simply apply results from \cite{MathieuSisto} to conclude. \subsection{Outline of the paper} In Section \ref{preliminaries}, we look at some basic definitions as well as discussing the setup for Markov chains and random walks. This allows us to consider a Markov chain under some assumptions, which we shall call a tame Markov chain. We show that a random walk (satisfying standard conditions) 'pushed-forward' via a quasi-isometry gives rise to a tame Markov chain, see Lemma \ref{pushforwardistame}. This, as explained above, is the key fact about Markov chains that we are interested in. Section \ref{sec:WPD} has two purposes. First, we introduce the notion of super-divergent element, which is the key notion for one of the assumptions under which we can prove linear progress. Second, we review and set up notation regarding loxodromic WPD element, as well as showing that super-divergent elements are WPD. In Section \ref{sec:assump} we state the two sets of assumptions under which we can prove linear progress, and discuss various examples. In Section \ref{geometricarguments} we state Proposition \ref{axiom}, which we then prove under either of the assumptions for linear progress. This proposition is a key intermediate step in the proof of linear progress, and we highlight it because of the following two reasons. First, its conclusion can be stated whenever one has a group acting on a hyperbolic space with a loxodromic WPD element. So, potentially the proposition could be true for all acylindrically hyperbolic groups. Second, in Section \ref{probabilisticarguments} we show that whenever the conclusion of the proposition holds, then linear progress also holds. Hopefully Proposition \ref{axiom} will be used in the future to cover more groups and group actions. We note that the proof of Proposition \ref{axiom} in both cases uses a key idea of the form "if you have a path that, in the Cayley graph, moves away from certain specified parts, then a certain projection moves slowly". See for example Lemma \ref{20francs} and Lemma \ref{o(log)}. Finally, in Section \ref{applications} we prove all the applications of linear progress mentioned above. \subsection*{Acknowledgements} Part of the content of this paper originated from discussions with Dominik Gruber, without whom this paper would not exist, and the authors are very thankful for his contributions. The author would also like to thank Adrien Boulanger for pointing out useful references and Cagri Sert for useful remarks that led to clarifications in the introduction. The work of the first author was supported by the EPSRC-UKRI studentship EP/V520044/1. \section{Tame Markov chains} \label{preliminaries} \subsection{Basic geometric group definitions} We recall some of the main notions from geometric group theory that will be used throughout the paper. A geodesic metric space $(X,d_X)$ is \textit{$\delta$-hyperbolic} for some $\delta>0$ (or simply \textit{hyperbolic}) if for any geodesic triangle in $X$ the following holds: any side of the triangle is contained in the closed $\delta$-neighborhood of the union of the other two sides. (We recall that the $M$-neighborhood of a subset $A$ of the metric space $(X,d)$ is $\mathcal N_M(A) = \{ x \in X : d(x,A) \leq M \}$, where $ d(x,A)=\inf \{ d(x,a) : a \in A \}$.) A finitely generated group is called \textit{hyperbolic} if its Cayley graph with respect to some (equivalently, any) finite generating set is a hyperbolic metric space, when endowed with the graph metric. For two metric spaces $(X, d_X)$ and $(Y, d_Y)$, we say that a function $f:X \to Y$ is a \textit{quasi-isometry} if there exist constants $A\geq 1, B \geq 0$ and $C \geq 0$ such that: \begin{itemize} \item For all $x,y \in X$ : $d_{X}(x,y)/A -B \leq d_Y(f(x),f(y)) \leq A d_{X}(x,y)+B$, \item $\forall z \in Y$ $ \exists x \in X : d_{Y}(z,f(x))\leq C$. \end{itemize} Relatedly, a function $f$ as above is $L$-coarsely Lipschitz if $d_Y(f(x),f(y))\leq Ld_X(x,y)$ for all $x,y\in X$. The following is a well-known exercise in hyperbolic-geometry; we omit the proof. First, we recall the relevant notions. A subset $A$ of a geodesic metric space $X$ is \emph{quasi-convex} if there exists $C\geq 0$ such that all geodesics with endpoints in $A$ are contained in the $C$-neighborhood of $A$. Given a subset $A$ of a metric space $X$, we call a map $\pi:X\to A$ a closest-point projection if $d_X(x,\pi(x))=d_X(x,A)$ for all $x\in X$. \begin{lemma} \label{lem:exo_hyperbolicspace} Let $X$ be a $\delta$-hyperbolic space. Let $Q$ be a quasi-convex set and $\pi_Q:X \rightarrow Q$ a closest-point projection. Then there exists a constant $R>0$ only depending on $\delta$ and the quasi-convexity constant such that the following hold. \begin{enumerate} \item $\pi_Q$ is $R$-coarsely Lipschitz. \item For all $x,y \in X$, if $d_{X}(\pi_Q(x), \pi_Q(y)) \geq R$ then there are points $m_1, m_2 \in [x,y]$ such that $d_X(m_1,\pi_Q(x)) \leq R$ and $d_X(m_2,\pi_Q(y)) \leq R$ where $[x,y]$ is a geodesic between $x$ and $y$. Further, the subgeodesic of $[x,y]$ between $m_1$ and $m_2$ lies in the $R$-neighborhood of $Q$. \end{enumerate} \end{lemma} \subsection{General Markov chains} We explain the formal setup of general Markov chains and then the specific case of random walks. This setup is based on Chapter 1 from \cite{woess_2000}. \\ Given a finitely generated group $G$, we define a \textit{Markov chain on } $G$. This Markov chain has \textit{state space} $G$ and \textit{transition operator} $P=(p_G(g,h))_{g,h \in G}$, with the requirement that each $p_G(g,h)$ is non-negative and for each $g\in G$ we have $\sum_{h\in G} p_G(g,h)=1$. Each $p_G(g,h)$ represents the probability of moving from $g$ to $h$ in one step. When the group being considered is clear, we will drop the subscript. \\ When a basepoint $o\in G$ is fixed, a Markov chain gives a sequence of random variables $(w^{o}_n)_{n \geq 0}$ which describe the position of the Markov chain starting at $o$ and after $n$ steps. To model this we can choose as probability space the \textit{path space} $\Omega = G^{\mathbb{N}}$ equipped with the product $\sigma$-algebra arising from the discrete $\sigma$-algebra on $G$. We equip $\Omega$ with the probability measure $\mathbb P_{o}$ defined on cylinder sets (and extended via the Kolmogorov extension theorem) by$$ \mathbb P_o \Big[\ \{(h_i)\in G^{\mathbb{N}} : h_j = g_j\ \forall j=0,\dots n\}\ \Big] =\delta_{o}(g_0)p(g_0,g_1)\cdots p(g_{n-1}, g_n)$$ for all given $g_0,\dots, g_n\in G$. Then $w^o_n$ is just the $n$-th projection $\Omega \rightarrow G$. Note that a reformulation of the above is that for all given $o,g_i \in G$ we have $$\mathbb P \Big[ w^{o}_0=g_0,w^{o}_1=g_1 \cdots ,w^{o}_n=g_n \Big]=\delta_{o}(g_0)p(g_0,g_1)\cdots p(g_{n-1}, g_n).$$ We will often consider $\mathbb{P}[w^g_n=h]$ which is the probability that a Markov chain starting at the element $g$ ends up on the element $h$ after $n$ steps. Note that this can be written as a certain sum over all sequences of points in $G$ of length $n$ that start at $o$ and end in $g$, but we will never use this sum explicitly. Instead, similarly to the above, we note that the following holds for all $o,g,g_i\in G$ and $k\leq n$ (using a similar argument summing over suitable sequences): \begin{equation} \label{eqn:change_of_basepoint} \mathbb P_o \Big[ w^{o}_k=g_0,w^{o}_{k+1}=g_1 \cdots ,w^{o}_n=g_n | w^o_k=h\Big]= \mathbb P_o \Big[ w^{h}_0=g_0,w^{h}_{1}=g_1 \cdots ,w^{h}_{n-k}=g_{n-k} \Big]. \end{equation} \subsubsection{A note on a Markov property} \label{note_Markov_property} Usually, Markov chains are defined as stochastic processes satisfying the \textit{Markov property}: For all $h,g_i \in G$ and $n$: $$\mathbb{P}\big[ w^h_n=g_n \quad \vert w^h_{n-1}=g_{n-1}, \dots, w^h_0=g_0 \big]=\mathbb{P}\big[ w^h_n=g_n \quad \vert w^h_{n-1}=g_{n-1}\big].$$ This is well-known to be equivalent to the above set up of Markov chains, see e.g. \cite[Theorem 1.1.1]{markovnorris}. We will also need a similar property, which we record in the following lemma. \begin{lemma} \label{lem:Markov_property_set} For all $k \leq n$, all $A \subseteq G^{n-k+1}$, and $o,h \in G$ we have $$\mathbb P \Big[ (w^{o}_k,w^{o}_{k+1},\cdots, w^{o}_n) \in A \quad \Big\vert \quad w^{o}_{k}=h\Big]= \mathbb P \Big[ (w^{h}_0, w^{h}_2, \cdots, w^{h}_{n-k}) \in A \Big].$$ Moreover, for a subset $A \subseteq G$ and for all $o,h \in G$ and $k \leq n$: $$ \mathbb P \Big[ w^{o}_n \in A \Big\vert w^{o}_{n-k} =h\Big] =\mathbb P \Big[ w^{h}_k \in A\Big] .$$ \end{lemma} \begin{proof} This just follows from summing Equation \eqref{eqn:change_of_basepoint} over all elements of $A$. For the ``moreover" part, we can deduce it from the first statement applied to the set of all sequences of appropriate length where the last entry is in $A$. \end{proof} With a slight abuse of notation, we will often write ``Let $(w^o_n)$ be a Markov chain", and similar, to mean that we fix transition probabilities and denote $w^o_n$ the corresponding random variables.\\ \subsubsection{Random walks} A random walk is an equivariant Markov chain, meaning that for all $x,g,y \in G$ we have $$p_G(x,y)=p_G(gx,gy).$$ Equivalently, for all $x,y\in G$ we have $$ p_G(x,y) = \mu(x^{-1}y) $$ for some fixed probability measure $\mu$ on $G$, which is called the \textit{driving measure of the random walk}. In the case of a random walk, the $n$-step transition probabilities are obtained by $$ p^{(n)}(x,y)=\mu^{(n)}(x^{-1}y) $$ where $\mu^{(n)}$ is the $n$-fold convolution of $\mu$ with itself. Due to equivariance, the starting point of a random walk is often not as important, while for Markov chains we will have to ensure that all our statements are ``uniform" over all choices of starting point. In proofs, we will sometimes write $w_n$ instead of $w^{o}_n$ Markov path starting at a fixed basepoint $o$, but only when it is safe to do so. \\ \subsubsection{Push-forwards} Random walks can be pushed-forward via group homomorphisms, while we are interested in the fact that Markov chains can be pushed forward via bijections (or more general maps, but we will not need it in this paper). We now make this precise in the following definition. \begin{definition} \label{defn:push_forward} Consider now two finitely generated groups $G,H$ and $ f: H \rightarrow G$ a bijective map (which later on will be a quasi-isometry). Given a Markov chain on $H$, we define a Markov chain on $G$ given by $$p_{G}(g,h) = p_{H}(f^{-1}(g), f^{-1}(h)) $$ for all $g,h \in G$. We will call this Markov chain, the $\textit{push-forward Markov chain}$. We note that in general this is not a random walk even when the Markov chain on $H$ is a random walk. \end{definition} \subsection{Tame Markov chains} \label{tameMarkov} \begin{definition}[Tame Markov chain] \label{defn:tame} We shall say that a Markov chain on $G$ is \textit{tame} if it satisfies the following conditions:\\ \begin{enumerate} \item\label{item:bounded_jumps} {\bf (Bounded jumps)} There exists a finite set $S\subseteq G$ such that $p(g,h)=\mathbb P[w^g_1=h]=0$ if $h\notin gS$. \item\label{item:non-amen} {\bf (Non-amenability)} There exist $A>0$ and $\rho<1$ such that for all $x,y\in G$ and $n\geq 0$ we have $$\mathbb P[w^x_n=y]\leq A\rho^n.$$ \item\label{item:irred} {\bf (Irreducibility)} For all $s \in G$ there exist constants $\epsilon_s, K_s>0$ such that for all $g \in G$ we have $$\mathbb P[w^g_k=gs] \geq \epsilon_s$$ for some $k \leq K_s$. \end{enumerate} \end{definition} \begin{remark} \label{rmk:bounded_jumps} Once we have fixed a word metric $d_G$ on $G$ then the assumption of Bounded jumps \ref{defn:tame}-\eqref{item:bounded_jumps} is equivalent to the following: There exists a constant $K>0$ such that for all $n \in \mathbb N$ and starting point $o \in G$ we have $d_G(w^o_n, w^o_{n+1}) \leq K$. \end{remark} \begin{remark}\label{rmk:no_finite_state} (We cannot reduce to the finite-state case.) One can show that there are uncountably many tame Markov chains on a free group with all transition probabilities of the form $i/10$. This boils down to constructing uncountably many labellings on the edges of a standard Cayley graph with the property that for each vertex the sum of the labels emanating from it is 1, and all edges have weight at least 1/10. Note that, instead, there are only countably many random walks satisfying the same requirement on transition probabilities. What is more, the same is true for finite-state Markov chains, again with the same requirement, and this is one way to see that the study of tame Markov chains cannot be reduced to that of finite-state Markov chains. In fact, one can even construct uncountably many tame Markov chains on the free group that are all push-forwards of the same random walk via \emph{isometries} of the standard Cayley graph, using a similar strategy. Therefore, for our purposes of establishing a quasi-isometry invariant theory, we cannot reduce to finite-state Markov chains, even if we further restricted the notion of tameness in some way. \end{remark} In the following lemma we consider a bijective quasi-isometry. Due to a result of Whyte \cite[Theorem 2]{whyte} every quasi-isometry between two non-amenable groups lies at finite distance from a bijection, so bijectivity will not be a real constraint for us. Note also that the lemma applies with $G=H$ and $f$ the identity map, and in this case it says that random walks with suitable driving measures on non-amenable groups are tame Markov chains. \begin{lemma} \label{pushforwardistame} Let $G,H$ be finitely generated non-amenable groups and let $f:H \to G$ be a bijective quasi-isometry. Let $\mu$ be the driving measure for a random walk on $H$ such that $\mu$ has finite support and generates $H$ as a semigroup. Then the push-forward Markov chain on $G$ (see Definition \ref{defn:push_forward}) is a tame Markov chain. \end{lemma} \begin{proof} For convenience, we recall that the transition probabilities defining the Markov chains under consideration are $$p_G(g,h)=p_H(f^{-1}(g),f^{-1}(h))= \mu \Big( (f^{-1}(g))^{-1} f^{-1}(h) \Big).$$ \eqref{item:bounded_jumps} We first show that the Markov chain on $G$ has bounded jumps. Let $S'$ be the support of $\mu$, and note that there exists a finite subset $S$ of $G$ such that for all $h\in H$ we have $f(hS')\subseteq f(h)S$. Indeed, since $f$ is a quasi-isometry and $S'$ is finite, $f(hS')$ lies in a ball of uniformly bounded radius around $f(h)$ in any fixed word metric on $G$. Thus, for any $g,h\in G$ we have $p_G(g,h)=0$ unless $h\in gS$, as required. \par\medskip \eqref{item:non-amen} The analogous statement for the random walks on $H$, which is a well-known consequence of non-amenability \cite{Kesten,Day}, readily implies this condition \ref{item:non-amen} for the Markov chain on $G$. \par\medskip \eqref{item:irred} We now verify the third point in the definition of tameness. First note that for any $s\in G$ there exists a finite set $S\subseteq H$ such that for any $g\in G$ we have $f^{-1}(gs)\in f^{-1}(g) S$; again this is a consequence of $f$ being a quasi-isometry. Fix now $s\in G$. Given $g\in G$ there exists $s'\in S$ such that for all $k$ we have $$p^{(k)}(g,gs)= p^{(k)}(f^{-1}(g), f^{-1}(g)s')=\mu^{(k)}(s').$$ Since the support of $\mu$ generates $G$ as a semigroup, for any $s'\in S$ there exists $K_{s'}$ such that $\mu^{(K_{s'})}(s')>\epsilon_{s'}$ for some $\epsilon_{s'}>0$. We can then take $\epsilon_s=\min_{s'\in S}(\epsilon_{s'})$ and $K_s=\max_{s'\in S}K_s$. Hence, the Markov chain defined on $G$ is tame. \end{proof} Throughout this paper, we will be working with tame Markov chains. \subsection{Logarithmic subpaths } In this subsection we prove two lemmas regarding what kinds of paths Markov chains could or are likely to follow. The first lemma roughly speaking gives a lower bound on the probability that the Markov chain ends up at a specified element. \begin{lemma} \label{epsilonlength} Let $(w^o_n)_n$ be a tame Markov chain on a finitely generated group $G=\langle S \rangle$. Then there exist constants $\epsilon_0,U>0$ such that the following holds. For all $y \in G$ there exists $k \leq \ell_S(y) U$ such that for all $h \in G$ we have $$\mathbb{P}\big[ w_k^h=hy\big] \geq \epsilon_0^{\ell_S(y)}$$ where $\ell_S(y)$ is the word length of the element $y$ with respect to the finite generating set $S$. \end{lemma} \begin{proof} We can take $\epsilon_0=\min \{\epsilon_s : s\in S\}$ and $U=\max \{K_s : s\in S\}$, where $\epsilon_s,K_s$ are as in Definition \ref{defn:tame}-\eqref{item:irred}. We will work by induction on the length $l=\ell_S(y)$ of the element $y$. For $l=1$, this is clear as we can choose $k =k_s \leq U$ where $y=s\in S$ and $k_s$ is from Definition \ref{defn:tame}-\eqref{item:irred}. We assume the lemma holds for all elements of length $l$, and we shall show it for elements of length $l+1$. Let $y=s_1\cdots s_ls_{l+1}$. By the inductive hypothesis, there exists some $k \leq l U$ such that for all $h \in G$ we have $\mathbb{P}\big[ w_k^h=hs_1\cdots s_l\big] \geq \epsilon_0^{l}$. \\ We let $k_{s_{l+1}} \leq U$, from the Definition \ref{defn:tame}-\eqref{item:irred}, be such that $ \mathbb{P}\big[ w_{k_{s_{l+1}}}^{hs_1\cdots s_l}=hs_1\cdots s_ls_{l+1}\big] \geq \epsilon_0 $. Let $k'=k+k_{s_{l+1}}$. Hence, for all $h \in G$: \begin{align*} \begin{split} \mathbb P \big[ w_{k'}^h=hy\big]& \geq \mathbb P \big[ (w_{k'}^h=hy) \cap (w^h_k=hs_1\cdots s_l)\big] \\ & = \mathbb P \big[ w_{k'}^h=hy \quad \vert \quad w^h_k=hs_1\cdots s_l\big] \mathbb P \big[w^h_k=hs_1\cdots s_l\big] \\ & = \mathbb P \big[ w^{hs_1\cdots s_l}_{k_s}=hy\big] \mathbb P \big[w^h_k=hs_1\cdots s_l\big] \\ & \geq \epsilon_0 \epsilon^l_0 \end{split} \end{align*} where we used Lemma \ref{lem:Markov_property_set} for the ``change of basepoint'' from the second to the third line. Further $k' \leq (l+1)U$ as desired. This proves the lemma. \end{proof} The following lemma will be essential when proving Lemma \ref{etalogproj} later, and roughly speaking it says that it is very likely that a Markov path of length $n$ contains ``a copy'' of any given path of length about $\log(n)$ that it can possibly contain. \begin{lemma} \label{markovword} Let $G$ be a finitely generated group with a fixed generating set $S$. Let $w^{o}_{n}$ be a tame Markov chain. Then there exist constants $\eta, U>0$ such that the following holds. For all $n\geq 1$ and all elements $y \in G$ with $\ell_S(y) \leq \eta \log(n)$ we have: $$\mathbb{P}\Big[\exists i,j \leq n \hspace{2mm} : \hspace{2mm} (w^{o}_{i})^{-1}w^{o}_{j}=y\Big] \geq 1-e^{-\sqrt{n}/U}. $$ \end{lemma} \begin{proof} We fix $\epsilon_0, U,k>0$ as in Lemma \ref{epsilonlength} (note that we will increase $U$ in the last step of the proof). We let $u^o_i=(w^o_{(i-1)k})^{-1} w^o_{ik}$. Our goal, roughly, is to show that the probability that no $u^o_i$ equals $y$ is small (for any basepoint $o$). More precisely, we claim that for all $o\in G$ we have $$\mathbb{P}(\forall 1\leq i \leq r \hspace{2mm} \hspace{2mm} u^o_{i} \neq y)\geq (1-\epsilon_0^l)^r,$$ where $l=\ell_S(y)$. For $r=1$, we have $$\mathbb{P}(u^o_{1} \neq y)=\mathbb{P}(w^o_{k} \neq y)\geq 1-\epsilon_0^l$$ due to our choice of $k$ and $\epsilon_0$. To show the inductive step we compute \begin{equation*} \begin{split} \mathbb{P}\big[\forall 1\leq i \leq r \hspace{2mm} \hspace{2mm} u^o_{i} \neq y\big] &= \mathbb{P}\big[u^o_{1} \neq y \hspace{2mm} \text{and} \hspace{2mm} \forall 1<i \leq r \ u^o_{i} \neq y\big] \\ &=\sum_{g \neq y} \mathbb{P}\big[\forall 1<i \leq r \hspace{2mm} u^o_{i} \neq y | u^o_{1} = g \big]\mathbb{P}\big[u^o_{1}=g\big].\\ &=\sum_{g \neq y} \mathbb{P}\big[ \forall 1 \leq i\leq r-1 \hspace{2mm} u^g_{i} \neq y \hspace{2mm} \big]\mathbb{P}\big[u^o_{1}=g\big].\\ &\leq (1-\epsilon_0^l)^{r-1}\sum_{g \neq y} \mathbb{P}\big[u^o_{1}=g\big]\leq (1-\epsilon_0^l)^{r}, \end{split} \end{equation*} where in the third line we used Lemma \ref{lem:Markov_property_set}, and in the last line we used the induction hypothesis and then again the choice of $\epsilon_0$ and $k$. Using the claim we get (for any $y$, and the associated $k$ and $l=\ell_S(y)$) \begin{equation*} \begin{split} \mathbb{P}\Big[\forall i,j \leq n \hspace{2mm} \vert \hspace{2mm} (w_{i})^{-1}w_{j}\neq y\Big] & \leq \mathbb{P}\Big[\forall 1\leq i \leq \lfloor n/k \rfloor \quad u_i \neq y \Big] \\ &\leq (1-\epsilon_0^l)^{\lfloor n/k \rfloor}\\ &\leq (1-\epsilon_0^l)^{n/(lU)-1}, \end{split} \end{equation*} where for the last line recall that $k\leq U l$. We now assume that $y$ satisfies $l=\ell_S(y)\leq \eta\log(n)$ for $\eta = \frac{-1}{3\log(\epsilon_0)}$ and we show that for all but finitely many $n$ we have $(1-\epsilon_0^l)^{n/U-1}\leq e^{-\sqrt{n}}$. This suffices to prove the lemma for a large enough $U$, and specifically we will now consider $n\geq 1$ large enough that $n/(lU)-1\geq n^{5/6}$. In this setup we have $\epsilon_0^l \geq \epsilon_0^{\eta\log(n)}= n^{-1/3}$ and $$(1-\epsilon_0^l)^{n/(lU)-1}\leq ((1-n^{-1/3})^{n^{1/3}})^{\sqrt{n}}\leq e^{-\sqrt{n}},$$ where we used that $(1-a^{-1})^a\leq e^{-1}$ for all $a\geq 1$. This concludes the proof. \end{proof} \section{WPD and super-divergent elements} \label{sec:WPD} \subsection{Super-divergent elements} We define super-divergent elements, study some basic properties of these elements, and give some examples. For a subset $A \subseteq Z$ of a metric space $Z$ and a fixed map $\pi:Z \rightarrow A$ we write $d_{A}(x,y):= d(\pi(x), \pi(y)).$ \begin{definition}[Super-divergence] Let $A$ be a subset of a metric space $Z$ and let $\pi:Z \rightarrow A$ be a projection, that is, a map such that $\pi|_{A}$ is the identity. We say that the projection $\pi$ is \textit{super-divergent} if there exists a constant $\theta > 0$ and a super-linear function $f:\mathbb{R}^{+} \to \mathbb{R}^{+}$ such that the following holds. For all $d >0$ and paths $p$ remaining outside of the $d$-neighbourhood of $A$, if $d_{A}(\pi(p_{-}),\pi(p_{+})) > \theta$ then $\ell(p) > f(d)$, where $p_{-}$ and $p_{+}$ are the endpoints of $p$ and $\ell(p)$ is its length. \end{definition} \begin{figure*}[h] \centering \begin{tikzpicture}[scale=0.8] \draw[black] (0,0) to[out=20,in=160] node [pos=0.75] (mid1) {} node [pos=0.55] (mid2) {} node [pos=0.85] (mid3) {} (10,0); \filldraw[black] (2,0.63) circle (1pt) node[anchor=north]{$\pi_A(p_{-})$}; \filldraw[black] (8,0.63) circle (1pt) node[anchor=north]{$\pi_A(p_{+})$}; \draw[densely dashed][red] (0,1.5) to[out=20,in=160] node [pos=0.75] (mid1) {} node [pos=0.55] (mid2) {} node [pos=0.85] (mid3) {} (10,1.5); \filldraw[black] (1,3) circle (1pt) node[anchor=east]{$p_{-}$}; \filldraw[black] (9,3) circle (1pt) node[anchor=west]{$p_{+}$}; \draw[-stealth] (1,3) --(1.97,0.68); \draw[-stealth] (9,3) --(8,0.68); \draw [->,decorate,decoration=snake] (1,3) -- (9,3); \path [draw=blue,snake it] (1,3) -- (9,3); \draw[blue] (6,3.4) node{$\ell(p) >f(d)$}; \draw[>=triangle 45, <->,purple] (2.3,0.6)-- (7.7,0.6) ; \draw[purple] (5,0.3) node{$\geq \theta$}; \draw[red] (10,1.2) node{$\mathcal N_{d}(A)$}; \draw[black] (10,-0.3) node{$A$}; \end{tikzpicture} \caption{The projection $\pi_A$ is super-divergent.} \end{figure*} \begin{definition} \label{defn:X_proj} Let $G$ act on a hyperbolic space $X$. If $A\subseteq G$, an $X$-\textit{projection} $\pi: G \rightarrow A$ is a map such that for all $h\in G$ we have that $\pi(h)x_0$ is a closest point in $Ax_0$ to $hx_0$, for a fixed basepoint $x_{0} \in X$. \end{definition} Recall that an element $g \in G$ \textit{acts loxodromically} on $X$ if the map $\mathbb Z \rightarrow X$ given by $n \rightarrow g^nx_0$ is a quasi-isometric embedding for some (equivalently, any) $x_0 \in X$. We note that if $g$ is loxodromic, then $\langle g \rangle x_0$ is quasi-convex in $X$. Hence we can apply Lemma \ref{lem:exo_hyperbolicspace} to $\langle g \rangle$, which we will do in Lemma \ref{lem:examples_of_superdivgelements} below. \begin{definition} Let $G$ act on a hyperbolic space $X$ and let $g \in G$. We say that $g$ is a \textit{super-divergent element} if \begin{itemize} \setlength\itemsep{1em} \item $g$ acts loxodromically on $X$, and \item an $X$-projection $\pi_{g}=\pi : G \rightarrow \langle g \rangle$ is super divergent, where $G$ is endowed with a fixed word metric $d_G$. We shall call $\theta$ the \textit{super-divergent constant} for $g$. \end{itemize} \end{definition} In the following lemmas we give examples of groups that have a super-divergent elements. \begin{lemma} \label{lem:examples_of_superdivgelements} Infinite hyperbolic groups have a super-divergent element where the hyperbolic space acted on can be taken to be either \begin{enumerate} \item the Cayley graph with respect to a finite generating set, or (more generally) \item the coned-off Cayley graph with respect to a family of uniformly quasi-convex infinite index subgroups. \end{enumerate} \end{lemma} \begin{proof} While technically the second case is more general, we prove the first one separately for clarity and since we will reduce the second case to the first one. (1) Here, the $\delta$-hyperbolic space $X$ is the Cayley graph of $G$ with respect to a finite generating set $S$, for some $\delta \geq 0$. We know that as $G$ is hyperbolic, every infinite order element is loxodromic for the action of $G$ on its Cayley graph $X$. Let $g$ be such a loxodromic element. Let $R>0$ be the constant from Lemma \ref{lem:exo_hyperbolicspace} for the quasi-convex subgroup $\langle g\rangle$. Let $d>R$ and $\alpha$ be a path staying outside of the $d$-neighbourhood of the axis $\langle g \rangle $ and such that the $X$-projection of $\alpha $ satisfies $d(\pi_{g}(\alpha_{-}), \pi_{g}(\alpha_{+})) \geq R $ where $\alpha_{-}$ and $\alpha_{+}$ are the endpoints of this path. Then by Lemma \ref{lem:exo_hyperbolicspace}, there is a point $m$ on the geodesic from $\alpha_{-}$ to $\alpha_{+}$ such that $d(m, \langle g \rangle) \leq R$. Hence $d(m, \alpha) \geq d-R$ and by \cite[Proposition III.H.1.6]{bridsonhaefliger} (taking exponentials on both sides), this leads to $\ell(\alpha) \geq 2^{(d-{R}-1)/ \delta}$ which is super linear in $d$ and hence $g$ is a super-divergent element. \\ (2) We will show that $G$ contains a super divergent element for the action on the coned-off Cayley graph with respect to one quasi-convex infinite-index subgroup, the general case follows from the fact that a family of uniformly quasi-convex subgroups will be finite. Let $X$ be the Cayley graph of $G$ with respect to a generating set $S$ and $\hat{X}$ be the coned-off Cayley graph over the quasi-convex subgroup $H$, which is hyperbolic by \cite[Proposition 5.6]{kapovichrafi}. Consider a loxodromic element $g$ for the action on $\hat{X}$; such an element exists in view of \cite[Theorem 2.4]{mahermasaischleimer}, applied to the element given by \cite[Theorem 1]{minasyan}. We note that for a point $x \in G$ we will interchangeably say that $x \in X$ or $x \in \hat{X}$. When considering geodesics in either of these spaces we will be more precise as to where they are. We note that by definition of the coned-off Cayley graph we have $d_{\hat{X}} \leq d_X=d_G$.\\ \textbf{Claim}: There exists a constant $M>0$ such that the following holds. For $x \in G $, let $\hat{\pi}_g(x)$ be coned-off projection ($\hat{X}$-projection) of $x$ and let $\pi_g(x)$ be the closest-point projection of $x$ on $\langle g \rangle$. Then for all $x\in G$, we have $d_X(\hat{\pi}_g(x),\pi_g(x)) \leq M$. \begin{proof}[Proof of claim] We will show that there is a constant $M'>0$ such that $d_{\hat{X}}(\hat{\pi}_g(x),\pi_g(x)) \leq M'$. By the fact that $\langle g \rangle \xhookrightarrow{q.i.} \hat{X}$ is a quasi-isometric embedding, this is enough to prove the claim. We first note that by \cite[Proposition 5.6]{kapovichrafi}, there exists a constant $N>0$ such that the following holds. Any geodesic $ [p_1,p_2]_{X}$ in $X$ between points $p_1$ and $p_2$ gets mapped $N$-Hausdorff close to a geodesic $ [p_1, p_2]_{\hat{X}}$ in $\hat{X}$ between $p_1$ and $p_2$. \\ Let $M'=2(N+R)$ where $R$ is from Lemma \ref{lem:exo_hyperbolicspace}, for the hyperbolic space $\hat{X}$ and quasi-convex subgroup $\langle g \rangle$. For a contradiction, assume that $d_{\hat{X}}(\hat{\pi}_g(x),\pi_g(x))> M'$. Hence $d_X(\hat{\pi}_g(x),\pi_g(x))> M'$ and by Lemma \ref{lem:exo_hyperbolicspace} there is a point $m \in [x, \hat{\pi}_g(x)]_X$ (a geodesic in $X$) such that $d_X(m,\pi_g(x)) \leq R$. Let $q \in [x, \hat{\pi}_g(x)]_{\hat{X}}$ (this is a geodesic in $\hat{X}$) be such that $d_{\hat{X}}(q,m) \leq N $, which exists by the fact that $[x, \hat{\pi}_g(x)]_X$ gets mapped $N$-Hausdorff close to $[x, \hat{\pi}_g(x)]_{\hat{X}}$, with respect to $d_{\hat{X}}$. \\ Hence $d_{\hat{X}}(q,\pi_g(x)) < R+N$ and so $d_{\hat{X}}(q,\hat{\pi}_g(x)) < R+N$ (otherwise $\hat{\pi}_g(x)$ would not be the $\hat{X}$-projection of $x$). This leads to $$ 2(N+R)= M'< d_{\hat{X}}(\hat{\pi}_g(x),\pi_g(x)) \leq d_{\hat{X}}(q,\pi_g(x))+d_{\hat{X}}(q,\hat{\pi}_g(x)) \leq 2(N+R) $$ a contradiction, proving the claim. \end{proof} Let $\alpha $ be a path in $G$ staying outside of the $d$-neighbourhood of $\langle g \rangle$ and such that $d_{X}(\hat{\pi}_{g}(\alpha_{-}),\hat{\pi}_g( \alpha_{+})) > \theta_1 +2M$ where $\theta_1$ is the super-divergent constant from the bullet point (1) above and $M$ is from the claim. The claim then leads to \begin{align*} \begin{split} d_X(\pi_g(\alpha_{-}),\pi_g(\alpha_{+}))) &\geq d_{X}(\hat{\pi}_{g}(\alpha_{-}),\hat{\pi}_g( \alpha_{+}))-d_X(\hat{\pi}_g(\alpha_{-}),\pi_g(\alpha_{-}))-d_X(\hat{\pi}_g(\alpha_{+}),\pi_g(\alpha_{+})) \\ &\geq \theta_1+2M-2M = \theta_1 \end{split} \end{align*} We can therefore apply bullet point (1) above to conclude. \end{proof} \begin{lemma} \label{lem:rh_superdiv} Let $(G,\mathcal P)$ be a relatively hyperbolic group, where all $P\in\mathcal P$ have infinite index in $G$. Then $G$ contains a super-divergent element for the action on the coned-off Cayley graph $\Cay(G,S\cup \mathcal P)$, where $S$ is a finite generating set for $G$. \end{lemma} \begin{proof} We denote by $\hat{G}$ the coned-off Cayley graph of $G$, with respect to $\mathcal P$. Further, denote by $d_G$ the word metric on $G$ with respect to the generating set $S$ and by $d_{\hat{G}}$ the metric in $\Cay(G,S\cup \mathcal P)$ with respect to $S \cup \mathcal P$. Clearly $d_{\hat{G}} \leq d_G$. Similarly to what we did above, for a $x \in G$, we will interchangeably say that $x \in G$ or $x \in \hat{G}$. \\ By \cite[Theorem 3.26]{osinrelativelyhyperbolic}, there exists a constant $\nu>0$ such that the following holds. Let $\Delta = \xi_1\xi_2\xi_3$ be a geodesic triangle whose sides $\xi_i$ are geodesics in $\hat{G}$. Then for any vertex $a \in \xi_1$ there is a vertex $b \in \xi_2\cup \xi_3$ such that $d_G(a,b) \leq \nu$.\\ Let $g$ be an element acting loxodromically on $\hat{G}$ with axis $\langle g \rangle$, which exists by \cite[Corollary 4.5]{osinelementary}. As $\langle g \rangle$ is quasi-isometrically embedded in $\hat{G}$ (say with quasi-isometric embedding constants $(a,b)$), the subgroup $\langle g \rangle$ is $\iota$-quasi-convex in $\hat{G}$, for some $\iota>0$. Let $\alpha $ be a path in $G$, between points $\alpha_{-},\alpha_{+} \in G$, staying outside of the $d$-neighbourhood (with respect to $d_{G}$) of $\langle g \rangle$ and such that $d_{G}(\pi_g(\alpha_{-}),\pi_g(\alpha_{+}) )\geq a(10 \nu + 10\iota +b)=: \theta$. Let $\hat{\beta_1}$ be a geodesic, in $\hat{G}$, between $\pi_g(\alpha_{-})$ and $\pi_g(\alpha_{+})$. By the choice of $\theta$, we can find a point $m \in \hat{\beta_1}$ at distance at least $4\nu +3 \iota $ from both $\pi_g(\alpha_{-})$ and $\pi_g(\alpha_{+})$. Let $\hat{\beta_2}=[\alpha_{-},\alpha_{+}]_{\hat{G}}$ be a geodesic in $\hat{G}$, similarly $\hat{\beta_3}=[\alpha_{-},\pi_g(\alpha_{-})]_{\hat{G}}$ and $\hat{\beta_4}=[\alpha_{+},\pi_g(\alpha_{+})]_{\hat{G}}$. Then, by \cite[Theorem 3.26]{osinrelativelyhyperbolic} mentioned above and by a usual quadrangle argument, this leads to $d_{G}(m,\hat{\beta_2} \cup \hat{\beta_3} \cup \hat{\beta_4}) \leq 2\nu$. Now, by our choice of $m$ and the fact that $\pi_{g}(\alpha_{-}), \pi_g(\alpha_{+}) $ are closest point projections, we have $d_G(m, \hat{\beta_3} \cup \hat{\beta_4}) > 2\nu$. Hence there is a point $y \in \hat{\beta_2}$ such that $d_G(m,y) \leq 2 \nu $.\\ By \cite[Lemma 8.8]{hruskarelativelyhyperbolic} and by the fact that $\langle g \rangle$ is also a quasi-geodesic in $G$, there is a constant $L_1>0$ (only depending on the quasi-geodesic constants of $\langle g \rangle$) and a point $m'\in \langle g \rangle$ such that $d_G(m,m') \leq L_1$. \\ Let $\beta_2$ be a geodesic, in $G$, from $\alpha_{-}$ to $\alpha_{+}$. Then by \cite[Proposition 8.13]{hruskarelativelyhyperbolic}, there exists a constant $L_2$ such that for some transient point $z \in \beta_2$ we have that $d_G(y,z) \leq L_2$. We do not give the definition of a transient point and refer to \cite{hruskarelativelyhyperbolic} and \cite{sisto2013tracking} for definitions and properties. \\ Combining all of the above yields, $$ d_G(\langle g \rangle, z) \leq d_G(m',m) + d_G(m,y)+d_G(y,z) \leq L_1 + 2\nu +L_2 $$ Hence $d_G(z,\alpha) \geq d-(L_1 + 2\nu +L_2 )$. Using \cite[Lemma 4.3]{sisto2013tracking} and the fact that $z$ is a transient point, we get that there exists a constant $R$ such that $$R\log_2(\ell(\alpha)+1) +R \geq d-(L_1 + 2\nu +L_2 )$$ This leads to \[ \ell(\alpha) \geq 2^{(d-(L_1 + 2\nu +L_2 )-R)/R} \] which is super-linear in $d$, giving the required function $f$. \end{proof} \subsection{A digression on divergence} We now note a result, relating the existence of super-divergent elements and the \textit{divergence} of the ambient group; this will not be needed in the rest of the paper. We invite the reader to read \cite{DrutuMozesSapir} for equivalent definitions of divergence in a group. We will only use the observation from \cite[Section 3]{divergencedrutu} that, given a bi-infinite quasi-geodesic $q$ in a metric space $X$, seen as a function $q : \mathbb{R} \to X$, the \textit{divergence function of the quasi-geodesic} $Div^{q}_{\eta}$ is a lower bound for the divergence of $X$. For fixed constants $0 <\delta <1$ and $\eta \geq 0$, $Div^{q}_{\eta}(m)$ is defined as the infimum of the length of the paths from $q(m)$ to $q(-m)$ avoiding the ball $B(q(0), \delta m-\eta)$. (We note that the choice of constants is irrelevant provided that they are in the range given in [Corollary 3.12, \cite{DrutuMozesSapir}]). \begin{lemma} \label{divergence} If $G$ has a super divergent element $g$ for an action on a hyperbolic space $X$, then $G$ has at least super-quadratic divergence. \end{lemma} \begin{proof} We fix a word metric $d_G$ on $G$. Consider a super-divergent element $g$ and the $(a,b)$-quasi-geodesic $\langle g \rangle x_0 $. Fix constants $0 <\delta <1$ and $\eta \geq 0$. We will show that the function $Div^{q}_{\eta}(m)$ is super-quadratic in $m$. It is enough to consider this function for $m \geq m_0$ where $m_0 > \frac{2a^2(\eta-b)}{a\delta}=:M'$. \\ \textbf{Claim}: There is a constant $D>0$ such that the following holds. If $h\not\in B^G(1,\delta m-\eta)$ and $d_G(1,\pi_g(h)) \leq \delta m/2a -\eta $ then $h \not\in \mathcal{N}_{\delta m/D -\eta}(\langle g \rangle)$, all considered with respect to the metric $d_G$. \begin{proof}[ Proof of claim] Let $D > \frac{2a^2\delta m_0}{M' }$ where $M'$ is defined above. For a contradiction, assume that there is a point $h' \in \langle g \rangle$ such that $d_G(h,h') \leq \delta m/D-\eta$. Hence $d_X(hx_0,h'x_0) \leq \delta m/D-\eta$. This leads to $d_G(1,h') \geq d_{G}(1,h)-d_G(h,h') \geq \delta m(D -1) /D $. \\ Hence $d_X(x_0, h'x_0) \geq \delta m(D -1) /aD-b$ as $\langle gx_0 \rangle$ is a $(a,b)$-quasi-geodesic in $X$. Therefore: \begin{align*} \begin{split} 2(\delta m/D-\eta) \geq 2d_X(hx_0, h'x_0) &\geq d_X(\pi_g(h)x_0, h'x_0) \geq d_X(x_0, h'x_0) -d_X(x_0,\pi_g(h)x_0) \\ &\geq \delta m(D -1) /aD-b-(\delta m/2a -\eta ) \end{split} \end{align*} which by the choice of $D$ is a contradiction. Hence there is no such $h' \in \langle g \rangle$ and $h \not\in \mathcal{N}_{\delta m/D -\eta}(\langle g \rangle)$ as claimed. \end{proof} Let $p$ be a path in (the Cayley graph of) $G$ from $g^{-m}$ to $g^{m}$ staying outside of the ball $B(1,\delta m-\eta)$; for convenience we will actually consider a discrete path (that is, a sequence of vertices with consecutive ones connected by an edge). We will think of the path $p$ as being 'oriented' from $g^{-m}$ to $g^{m}$, this allows us to talk about 'first points' and 'last points'. \\ Let $y_1$ be the first point on $p$ such that $d_G(1, \pi_g(y_1)) \leq \delta m/2a -\eta$. We then define the point $y_2$ to be the first point after $y_1$ such that $d_{\langle g \rangle}(y_1,y_2) \geq \theta$ where $\theta$ is the super-divergent constant for $g$. By minimality of $y_2$, and the fact that closest-point projections to quasi-convex sets are coarsely Lipschitz (Lemma \ref{lem:exo_hyperbolicspace}) this also means that $d_{\langle g \rangle}(y_1,y_2) \leq \theta +L$, for some constant $L$.\\ We recursively define the point $y_i$ as the first point, after $y_{i-1}$, such that $d_{\langle g \rangle}(y_{i-1},y_i) \geq \theta$. We again note that $d_{\langle g \rangle}(y_{i-1}, y_i) \leq \theta + L$. We stop this process as soon as $y_{s+1}$ projects on $\langle g \rangle$ outside of $B(1,\delta m/2a-\eta)$.\\ Say that there are $s$ of these points $(y_i)_{i=1}^s$. By the fact that $\langle g \rangle$ is a $(a,b)$-quasi geodesic, this leads to $s(\theta +L) > \frac{1}{a}(\delta m/2a-\eta) -b-2L$. By the claim above, each subpath $p_i$ of $p$ from $y_i$ to $y_{i+1}$ is outside of the $(\delta/D m-\eta)$-neighbourhood of $\langle g \rangle$.\\ By super-divergence of $g$, we get that $$ \ell(p) > \sum_{i=1}^{s} \ell(p_i) \geq s f(\delta m/3 -\eta) > \frac{1}{\theta+L}(\delta m/ 2a^2 -\eta /a -b-2L)f(\delta m/D -\eta) $$ which is super-quadratic in $m$ as $f$ is super linear in $m$. Hence $Div^{q}_{\eta}(m)$ is super-quadratic in $m$. Therefore the divergence of $G$ is at least super-quadratic. \end{proof} We note that since mapping class groups have quadratic divergence \cite{behrstock,Duchin_2009} (with finitely many exceptions), they cannot contain super-divergent elements. \subsection{WPD elements} \label{subsec:WPD} We recall the definition of a WPD element from \cite{bestvinafujiwara}. \begin{definition}[WPD element] Let $G$ be a group acting on a hyperbolic space $X$ and $g$ an element of $G$. We say that $g$ satisfies the \textit{weak proper discontinuity condition} (or that $g$ is a $WPD$ element) if for all $\kappa >0$ and $x_0 \in X$ there exists a $N \in \mathbb{N}$ such that \[ \# \{ h \in G \vert \quad d_X(x_0, hx_0)< \kappa, \quad d_X(g^{N}x_0, hg^{N}x_0) < \kappa \} < \infty \] \end{definition} We fix a group $G$ acting on a hyperbolic space $X$ throughout this subsection. We now show that super-divergent elements are WPD, and we list some known consequences of being WPD. \begin{lemma} \label{lem:super_div_WPD} If $g \in G$ is a super-divergent element for the action on a hyperbolic space $X$ then it is a WPD element. \end{lemma} \begin{proof} As $g$ acts loxodromically on $X$, the axis of $g$ is a $(a, b)$-quasi geodesic both in $X$ and in $G$, where at this point we fixed a word metric $d_G$ on $G$. For a contradiction, assume that $g$ is not a WPD element. Hence there exists $\kappa>0$ such that for all $m \in \mathbb{N}$ there are infinitely many elements $h \in G$ satisfying: \begin{equation} \label{not_wpd_defn} d_X(x_{0},hx_{0}) <\kappa \hspace{2mm} \text{and} \hspace{2mm} \hspace{1mm}d_X(g^{m}x_{0},hg^{m}x_{0}) < \kappa \end{equation} We let $m$ be such that $m > a(\theta+b+2\kappa)$ where $\theta$ is the super-divergent constant for $g$. Let $d$ be such that $f(d) >a m +b$ where $f$ is the super-linear function from the definition of $g$ being super-divergent. As we have infinitely many $h$ satisfying (\ref{not_wpd_defn}), we can find $h$ such that $d_G(1,h) > 10d+2(am+b)$. By definition, in $X$ we have that $\pi_g(h)x_0$ and $\pi_g(hg^m)x_0$ are the closest points on $\langle g x_0 \rangle$ to $h$ and $hg^m$ respectively. Hence $d_X(hg^mx_0, \pi_g(hg^m)x_0) \leq d_X(hg^mx_0, g^mx_0) < \kappa$ and similarly $d_X(hx_0, \pi_g(h)x_0) < \kappa$. Therefore \[ d_X\left(\pi_{g}(h)x_0,\pi_{g}(hg^m)x_0\right) > d_X(hx_0, hg^{m}x_0) -2\kappa >m/a-b-2\kappa > \theta \] by the choice of $m$. Hence $d_G(\pi_{g}(h),\pi_{g}(hg^m)) > \theta$. Further, by the choice of $h$, the geodesic from $h$ to $hg^{m}$ stays outside of the $d$-neighbourhood of $\langle g \rangle $. By super divergence of $g$, this leads to \[am+b \geq d_G(h,hg^{m}) > f(d) > am+b\] a contradiction. Hence $g$ is a WPD element. \end{proof} Recall that any loxodromic WPD element $g$ is contained in a unique maximal elementary subgroup of $G$, denoted $E(g)$ and called the \textit{elementary closure} of $g$, see \cite[Theorem 1.4]{Osin_acylindrical}. We will refer to the following lemma as the strong Behrstock inequality, and we will use it very often. The lemma follows from \cite[Theorem 4.1]{bromberg2017acylindrical}. \begin{lemma} \label{lem:Behrstock} Let $g$ be a loxodromic WPD element. Then, for $\gamma=E(g)$, there is a $g$-equivariant map $\pi_{\gamma}:G\to \mathcal P(\gamma)$, where $\mathcal P(\gamma)$ is the set of all subsets of $\gamma$, and a constant $B$ with the following property. For all $x\in G$ and $h\gamma \neq h'\gamma$ we have \[ d_X(\pi_{h\gamma}(x),\pi_{h\gamma}(h'\gamma)) >B \implies \pi_{h'\gamma}(x)=\pi_{h'\gamma}(h\gamma), \] where $\pi_{k\gamma}(z)=k\pi_{\gamma}(k^{-1}z)$. Moreover, for all $x\in G$ the Hausdorff distance between $\pi_\gamma(x)$ and an $X$-projection of $x$ to $\langle g\rangle$ is bounded by $B$. \end{lemma} \begin{figure}[h] \centering \begin{tikzpicture}[scale=0.9] \draw (0,0) arc(0:180:3cm and 1cm); \draw (7,0) arc(0:180:3cm and 1cm); \draw[ultra thick, blue] (0,0) arc(0:40:3cm and 1cm); \draw[blue] (-1,0) node{$\pi_{h\gamma}(h'\gamma)$}; \draw[ultra thick, red] (1,0) arc(180:140:3cm and 1cm); \draw[black] (-3,1.4) node{$h\gamma$}; \filldraw[black] (-6,2.5) circle (1pt) node[anchor=east]{$x$}; \draw[-stealth, dashed] (-6,2.5) --(-5.5,0.6); \filldraw[black] (-5.5,0.6) circle (1pt) node[anchor=east]{$\pi_{h\gamma}(x)$}; \draw[red] (1.8,-0.3) node{$\pi_{h'\gamma}(x)=\pi_{h'\gamma}(h\gamma)$}; \draw[black] (4,1.4) node{$h'\gamma$}; \draw[>=triangle 45, <->,black] (-5.2,0.55)-- (-0.9,0.55); \draw[black] (-3,0.2) node{$>B$}; \end{tikzpicture} \caption{The strong Behrstock inequality.} \end{figure} \begin{notation} \label{not:d_hgamma} For a loxodromic WPD element $g$, consider $\gamma=E(g)$ and the maps $\pi_{h\gamma}$ as in Lemma \ref{lem:Behrstock}. For $h,x,y\in G$ we denote $$d_{h\gamma}(x,y)=\diam (\pi_{h\gamma}(x)\cup\pi_{h\gamma}(y)).$$ \end{notation} As in the statement of Lemma \ref{lem:Behrstock}, we will often denote $\gamma=E(g)$ when a loxodromic WPD element $g$ has been fixed. We will often look at the set of cosets where two given elements have far away projections, as captured by the following definition. \begin{definition} \label{defn:H_T} Given a loxodromic WPD element $g$, for $x,y\in G$ and $T\geq 0$ we define the set \[ \mathcal{H}_T(x,y) :=\{ h\gamma \hspace{1mm} : \hspace{1mm} d_{h\gamma}(x,y) \geq T\} \] Throughout, we will consider the following (``distance formula") sum, for $x,y,z_1,z_2 \in G$, once a loxodromic WPD element $g$ has been fixed: \[ \sum_{\mathcal{H}_T(x,y)}[z_1,z_2] := \sum_{\substack{h\gamma \in \mathcal{H}_T(x,y) \\ \pi_{h\gamma}(z_1)\neq \pi_{h\gamma}(z_2)}} d_{h\gamma}(z_1,z_2). \] \end{definition} \begin{remark} \label{rem:triangle_inequality} (Triangle inequality) Note that for all $x,y,z,o,p\in G$ we have $$\sum_{\mathcal{H}_T(o,p)} [x,z] \leq \sum_{\mathcal{H}_T(o,p)} [x,y] + \sum_{\mathcal{H}_T(o,p)} [y,z],$$ since projection distances satisfy the triangle inequality as well. \end{remark} Let $o,p\in G$. In view of the fact that the projections on the $h\gamma$ satisfy the projection axioms of \cite{bestvina2014constructing}, by \cite[Theorem 3.3 (G)]{bestvina2014constructing}, we have a linear order on $\mathcal{H}(o,p)$: \begin{lemma}(Consequence of \cite[Theorem 3.3 (G)]{bestvina2014constructing}) \label{linearorder} Fix a loxodromic WPD element $g$. For any sufficiently large $T$, the set $\mathcal{H}_T(o,p) \cup \{o,p\}$ is totally ordered with least element $o$ and greatest element $p$. The order is given by $h\gamma \prec h'\gamma$ if one, and hence all, of the following equivalent conditions hold for $B$ as in Lemma \ref{lem:Behrstock}:\\ \begin{itemize} \item $d_{h\gamma}(o,h'\gamma) > B$. \item $\pi_{h'\gamma}(o)= \pi_{h'\gamma}(h\gamma)$. \item $d_{h'\gamma}(p,h\gamma) > B$. \item $\pi_{h\gamma}(p)= \pi_{h\gamma}(h'\gamma)$. \end{itemize} \end{lemma} \begin{figure}[h] \begin{tikzpicture}[scale=0.9] \draw (0,0) arc(0:180:1cm and 0.5cm); \draw (3,0) arc(0:180:1cm and 0.5cm); \draw (6,0) arc(0:180:1cm and 0.5cm); \draw (9,0) arc(0:180:1cm and 0.5cm); \draw (12,0) arc(0:180:1cm and 0.5cm); \draw[black] (0.5,0) node {$\cdots$}; \draw[black] (9.5,0) node {$\cdots$}; \draw[ultra thick, red] (0,0) arc(0:40:1cm and 0.5cm); \draw[ultra thick, red] (3,0) arc(0:40:1cm and 0.5cm); \draw[ultra thick, red] (6,0) arc(0:40:1cm and 0.5cm); \draw[ultra thick, red] (9,0) arc(0:40:1cm and 0.5cm); \draw[ultra thick, red] (12,0) arc(0:40:1cm and 0.5cm); \draw[ultra thick, blue] (-2,0) arc(180:140:1cm and 0.5cm); \draw[ultra thick, blue] (1,0) arc(180:140:1cm and 0.5cm); \draw[ultra thick, blue] (4,0) arc(180:140:1cm and 0.5cm); \draw[ultra thick, blue] (7,0) arc(180:140:1cm and 0.5cm); \draw[ultra thick, blue] (10,0) arc(180:140:1cm and 0.5cm); \draw[black] (5,-0.5) node {$h\gamma$}; \draw[blue] (4,-0.2) node {\tiny$\pi_{h\gamma}(h'\gamma)=\pi_{h\gamma}(o)$}; \filldraw[black] (-2,2) circle (1pt) node[anchor=east]{$o$}; \filldraw[black] (12,2) circle (1pt) node[anchor=east]{$p$}; \draw[dashed] (-2,1.9) --(-2,0); \draw[dashed] (12,2) --(12,0); \draw[red] (6,0.8) node {\tiny $\pi_{h\gamma}(h''\gamma)=\pi_{h\gamma}(p)$}; \draw [decorate,decoration={brace,amplitude=5pt,mirror,raise=4ex}] (-2,0) -- (3,0) node[midway,yshift=-3em]{$h'\gamma \prec h\gamma$}; \draw [decorate,decoration={brace,amplitude=5pt,mirror,raise=4ex}] (7,0) -- (12,0) node[midway,yshift=-3em]{$h\gamma \prec h''\gamma$}; \end{tikzpicture} \caption{Linear order on $\mathcal H_T(o,p)$: all the cosets on the left of $h\gamma$ have the same projection on $h\gamma$, and similarly for all the cosets on the right of $h\gamma$.} \end{figure} \begin{proof} We refer to \cite[Theorem 3.3(G)]{bestvina2014constructing}] for the linear order; the equivalence of the various conditions follows from the strong Behrstock inequality (Lemma \ref{lem:Behrstock}). \end{proof} The following lemma says that distance formula sums can be used to get lower bounds on distances in $X$. \begin{lemma} \label{distance} Fix a loxodromic WPD element $g$, and a basepoint $x_0\in X$. Then for all sufficiently large $T$ the following holds. For all $a,b \in G$, we have \[ d_{X}(ax_0,bx_0) \geq \frac{1}{2} \sum_{\mathcal{H}_T(a,b)} [a,b]. \] \end{lemma} \begin{proof} The following claim is shown in the proof of \cite[Theorem 6.8]{dahmani2014hyperbolically} (showing that $E(g)$ is geometrically separated) via \cite[Theorem 4.42 (c)]{dahmani2014hyperbolically}. \\ \textbf{Claim 1}\cite[Theorem 6.8, Theorem 4.42 (c) ]{dahmani2014hyperbolically}: For all constants $R>0$, there exists a constant $B_1>0$ such that the following holds. For cosets $h\gamma\neq h'\gamma$ we have \[\diam\Big( \mathcal N_R(h\gamma x_0) \cap \mathcal N_R(h'\gamma x_0) \Big) < B_1. \] \par\medskip Let $\alpha$ be a geodesic in $X$ between $ax_0,bx_0$. By Lemma \ref{lem:exo_hyperbolicspace}, there is a constant $R$ such that for each $h_i \gamma \in \mathcal H_T(a,b)$ there exist points $p_i, q_i \in \alpha$ such that $d_{X}(p_i, \pi_{h_i\gamma x_0}(ax_0)) \leq R$ and $d_{X}(q_i, \pi_{h_i\gamma x_0}(bx_0)) \leq R$. Further, the geodesic between $p_i$ and $q_i$ lies within the $R$-neighbourhood of $(h_i\gamma) x_0$. For each $i$, let $\beta_i$ be the sub-geodesic of $[p_i,q_i]$ with endpoints at distance $B_1$ from $p_i$ and $q_i$ respectively, where $B_1$ is from the claim above. If we choose $T$ large enough, then we can ensure that this subgeodesic $\beta_i$ exists for all $h_i \gamma \in \mathcal H_T(a,b)$. We note that $\beta_i \subset [p_i,q_i] \subseteq \mathcal N_R(h_i \gamma x_0)$. \\ \textbf{Claim 2}: For $i \neq j$, we have $\beta_i \cap \beta_j = \emptyset$ \begin{proof}[Proof of claim 2] For a contradiction, assume there exists $x \in \beta_i \cap \beta_j$. Then $[p_i,q_i]$ and $[p_j,q_j]$ share a subgeodesic (containing $x$) of length $B_1$. Such subgeodesic is contained in both $\mathcal N_R(h_i \gamma x_0)$ and $\mathcal N_R(h_i \gamma x_0)$, a contradiction with Claim 1. \end{proof} Now, if $T$ is large enough, for all $h_i \gamma \in \mathcal H_T(a,b)$ we have \[\ell(\beta_i) \geq d_X(p_i,q_i)-2B_1 \geq d_{h_i\gamma}(ax_0,bx_0) -2B_1-2R \geq d_{h_i\gamma}(ax_0,bx_0)/2.\] Hence \[d_X(ax_0,bx_0) \geq \sum_{h_i\gamma\in\mathcal{H}_T(a,b)} \ell(\beta_i) \geq \frac{1}{2}\sum_{h_i\gamma\in\mathcal{H}_T(a,b)} [a,b] \] as the $(\beta_i)_i$ are pairwise disjoint. \end{proof} The following lemma will be used below to prove that the ``distance formula" is coarsely Lipschitz and it will also be used in the proof of lemma \ref{o(log)}. \begin{lemma} \label{lem:2cosetsmax} Fix a loxodromic WPD element $g$. For any $T\geq 10 B$ which satisfies Lemma \ref{linearorder}, with $B$ as in Lemma \ref{lem:Behrstock}, the following holds. Let $o,p,a\in G$. Then there at most $2$ cosets $h\gamma \in \mathcal{H}_T(o,p)$ such that $\pi_{h\gamma}(a)$ is distinct from both $\pi_{h\gamma}(o)$ and $\pi_{h\gamma}(p)$. \end{lemma} \begin{proof} Assume this is not the case and let $h_1\gamma \prec h_2\gamma \prec h_3\gamma$ in $\mathcal{H}_T(o,p)$ be such that $\pi_{h_i\gamma}(a)\neq \pi_{h_i\gamma}(o)$ and $\pi_{h_i\gamma}(a)\neq\pi_{h_i\gamma}(p)$ for $i=1,2,3$. Since $d_{h_2\gamma}(o,p)\geq T\geq 10B$, we have either $d_{h_2\gamma}(a,p)>B$ or $d_{h_2\gamma}(a,o)>B$. The two cases are symmetrical, hence without loss of generality assume $d_{h_2\gamma}(a,p)>B$. Since $\pi_{h_2\gamma}(p)=\pi_{h_2\gamma}(h_3\gamma)$ by Lemma \ref{linearorder} this leads to $d_{h_2\gamma}(a,h_3\gamma)>B$ and so by the strong Behrstock inequality (Lemma \ref{lem:Behrstock}), we get $\pi_{h_3\gamma}(a)=\pi_{h_3\gamma}(h_2\gamma)$. Again by by Lemma \ref{linearorder}, we have $\pi_{h_3\gamma}(h_2\gamma)=\pi_{h_3\gamma}(o)$, a contradiction. \end{proof} We now show that the distance formula sum is coarsely Lipschitz. \begin{lemma} \label{lem:coarsely_lip_sum} Fix a loxodromic WPD element $g$, and a basepoint $x_0\in X$. For any $T\geq 10 B$ which satisfies Lemma \ref{linearorder} and Lemma \ref{distance}, with $B$ as in Lemma \ref{lem:Behrstock}, there exists a constant $L$ such that the following holds. Let $o,p,a,b\in G$. Then $$\sum_{\mathcal H_T(o,p)}[a,b]\leq L d_X(ax_0,bx_0)+L.$$ \end{lemma} \begin{proof} By Lemma \ref{lem:2cosetsmax}, there is a set $\mathcal A$ consisting of at most 4 cosets $h\gamma\in \mathcal H_T(o,p)$ such that one of $\pi_{h\gamma}(a)$ and $\pi_{h\gamma}(b)$ does not coincide with either $\pi_{h\gamma}(o)$ or $\pi_{h\gamma}(p)$. Note also that, for a given $h\gamma\in \mathcal H_T(o,p)$, if both $\pi_{h\gamma}(a)$ and $\pi_{h\gamma}(b)$ coincide with either $\pi_{h\gamma}(o)$ or $\pi_{h\gamma}(p)$, then either $\pi_{h\gamma}(a)=\pi_{h\gamma}(b)$ or $h\gamma\in \mathcal H_T(a,b)$. In view of this, we have $$\sum_{\mathcal H_T(o,p)}[a,b]\leq \sum_{h\gamma\in \mathcal A} d_{h\gamma}(a,b)+\sum_{\mathcal H_T(a,b)}[a,b].$$ Since closest-point projections on quasi-convex sets in hyperbolic spaces are coarsely Lipschitz, and the first sum has at most 4 terms, we see that the first sum is bounded linearly in $d_X(ax_0,bx_0)$. The second one is also bounded linearly in $d_X(ax_0,bx_0)$ by Lemma \ref{distance}, and we are done. \end{proof} \subsection{Logarithmic random projections} We conclude this section with a lemma about projections of Markov chains (where in the statement we use the notation associated to the loxodromic WPD element $g$ established above). The lemma says, roughly, that it is very likely that the Markov chain creates logarithmically sized projections, this follows from Lemma \ref{markovword}. \begin{lemma} \label{etalogproj} Let $(w_n^o)$ be a tame Markov chain on a group $G$ containing a loxodromic WPD element for the action on some hyperbolic space. Then there exist constants $ \eta, U >0$ such that for each basepoint $p\in G$ and each $n\geq 1$ we have \[ \mathbb{P}\left[\exists m\leq n,\ h\in G: d_{h\gamma}(p,w^p_m)\geq \eta\log(n)\right]\geq 1- e^{-\sqrt{n}/U}. \] \end{lemma} \begin{proof} Let $\eta,U$ be as in Lemma \ref{markovword} (but note that the $\eta$ that satisfies the corollary will be smaller). By Lemma \ref{markovword}, for any $n$ with probability at least $1- e^{-\sqrt{n}/U}$ there exist $i,j\leq n$ such that $(w^o_{i})^{-1}w^o_{j}=g^{\lfloor\eta \log(n)/l_G(g)\rfloor}$. Then $d_{w_i\gamma}(w^o_i,w^o_j)\geq d_{\gamma}(1,g^{\lfloor\eta \log(n)/l_G(g)\rfloor})\geq \eta'\log(n)$, for some sufficiently small $\eta'$, so that $d_{w_i\gamma}(o,w^o_m)\geq \eta'\log(n)/2$ for either $m=i$ or $m=j$, as required. \end{proof} \section{Our assumptions} \label{sec:assump} In this short section we spell out the two sets of assumptions under which we work for the rest of the paper, and point out examples of groups and spaces that satisfy them. \begin{assumptions} \label{assump:superdiv} $G$ is a finitely generated group acting on a hyperbolic space $X$ and containing a super-divergent element. \end{assumptions} In particular, note that Assumption \ref{assump:superdiv} is satisfied by hyperbolic and relatively hyperbolic groups and corresponding spaces as in Lemma \ref{lem:examples_of_superdivgelements} and Lemma \ref{lem:rh_superdiv}. Before stating the next assumption, we need a definition. \begin{definition} \label{defn:subexp} Let $G$ be a group with a fixed word metric $d_G$, and let $H<G$. We say that $H$ has \emph{subexponential growth in the ambient word metric} if the function $\psi:\mathbb{N} \to \mathbb{N}$ defined as $\psi(n)= \vert \{h \in H : d_G(1, h) \leq n \} \vert$ is subexponential in $n$. \end{definition} \begin{assumptions} \label{assump:graph} $G$ is the finitely generated fundamental group of a graph of groups where the edge groups (more precisely, their images in $G$) have subexponential growth in the ambient word metric, and there exists a loxodromic WPD element for the action on the Bass-Serre tree $X$. \end{assumptions} The assumption applies, for example, to many right-angled Artin group, as we now show. \begin{lemma} Any right-angled Artin group $A_\Gamma$ whose defining graph $\Gamma$ has a separating simplex and diameter at least 3 satisfies Assumption \ref{assump:graph}. \end{lemma} \begin{proof} The separating simplex $\Delta$ gives a splitting over an abelian subgroup $A_\Gamma=A_{\Gamma_1} *_\Delta A_{\Gamma_2}$ where the $\Gamma_i$ are subgraphs of $\Gamma$ intersecting at $\Delta$. Since $A_\Gamma$ is CAT(0), said subgroup is undistorted and hence has subexponential growth in the ambient word metric (as well as its own word metric). Let $a, b$ be vertices of $\Gamma$ in two distinct connected components of the complement of the disconnecting simplex. Then the element $ab$ of $\Gamma$ is not contained in a vertex group of the splitting (for example this can be checked by passing to the abelianisation of $A_\Gamma$) and hence it is loxodromic for the action on the Bass-Serre tree. Towards proving that $ab$ is WPD, we note that the vertices $A_{\Gamma_1}$ and $ab A_{\Gamma_1}$ of the Bass-Serre tree have trivial common stabiliser (recall that the vertices of the Bass-Serre tree are cosets of the vertex groups). This can be seen for example using the normal form from \cite[Proposition 3.2]{HermillerMeier}: The stabilisers of $A_{\Gamma_1}$ and $ab A_{\Gamma_1}$ are $A_{\Gamma_1}$ and $ab A_{\Gamma_1}b^{-1}a^{-1}$, so that the normal form for elements in the former will not contain the letter $b$, while those for the latter will. This fact suffices to show that $ab$ is WPD since it implies that two vertices on the axis of $ab$ (the closest-point projections of the aforementioned vertices) have trivial common stabiliser, which means that \cite[Corollary 4.3]{minasyan2017acylindrical} applies and gives that $ab$ is WPD. \end{proof} \subsection{3-manifold groups} In this subsection we show that each acylindrically hyperbolic 3-manifold group satisfies one of the assumptions above, for a suitable hyperbolic space $X$. We recall the following well-known result on some obstructions to acylindrical hyperbolicity, see for example \cite{minasyan2017acylindrical}. \begin{lemma}[ \cite{minasyan2017acylindrical}] \label{obstructionacylindricalhyperbolicity} Let $G$ be a group such that one of the following holds: \begin{itemize} \item $G$ contains an infinite cyclic normal subgroup $Z$, or \item $G$ is virtually solvable. \end{itemize} Then $G$ is not acylindrically hyperbolic. \end{lemma} By a result of Minasyan-Osin \cite{minasyan2017acylindrical} we know that most $3$-manifold groups are acylindrically hyperbolic; we will need a different version of their characterisation of when a $3$-manifold group is acylindrically hyperbolic. Using various results from the literature, including the geometrisation theorem, we can get the following result. \begin{proposition} \label{acylindricallyhyperbolicmanifold} Let $M$ be a closed, connected, oriented $3$-manifold. Let $G:= \pi_1(M)$. Then $G$ is acylindrically hyperbolic if and only if $M$ satisfies one of the following: \begin{itemize} \item $M$ is not prime and not $\mathbb{RP}^3 \# \mathbb{RP}^3$, \item $M$ is a geometric manifold with Thurston geometry $\mathbb{H}^3$, \item $M$ is a non-geometric prime manifold. \end{itemize} Moreover, when $G$ is acylindrically hyperbolic then it either contains a super-divergent element for some action on a hyperbolic space, or it admits a graph of groups decomposition as in Assumption \ref{assump:graph}. \end{proposition} \begin{proof} We use the geometrisation theorem, see \cite{perelman2003ricci}, \cite{perelman2}, \cite{LottKleiner}, \cite{morganTian}, \cite{CaoZhu}. If $M= M_1 \# M_2\#\cdots \# M_n$ is the prime decomposition of $M$, then $G:=\pi_1(M)$ is hyperbolic relative to $\pi_1(M_i)$. In this case, $G$ is acylindrically hyperbolic unless it is virtually cyclic. This only happens when $M$ is the connected sum of two real projective $3$-spaces. Indeed, the only free product of non-trivial groups which is virtually cyclic is the free product of two copies of $\mathbb Z/2\mathbb Z$, and the only closed, connected, oriented $3$-manifold with this fundamental group is $\mathbb{RP}^3$. We can therefore assume that $M$ is a prime manifold. First, we consider the case where $M$ is geometric. By using Lemma \ref{obstructionacylindricalhyperbolicity}, we will show that the only acylindrically hyperbolic geometric $3$-manifolds are the ones with geometry $\mathbb{H}^3$. We list the other possible geometries for a prime manifold as well as the reason why they are not acylindrically hyperbolic. We only note the possible fundamental groups, non acylindrical hyperbolicity will follow from Lemma \ref{obstructionacylindricalhyperbolicity}. We refer to \cite{scottmanifolds} for the details on the possible fundamental groups for manifolds with the following Thurston geometries: \\ \begin{itemize} \item $S^3$: finite fundamental group. \item $\mathbb{R}^3$: virtually abelian fundamental group. \item $\mathbb{H}^3$: non-elementary hyperbolic hence acylindrically hyperbolic. \item $S^2 \times \mathbb{R}$: virtually cyclic fundamental group. \item $\mathbb{H}^2 \times \mathbb{R}$: fundamental group contains an infinite normal cyclic subgroup. \item $\widetilde{SL_2(\mathbb{R})}$: fundamental group contains an infinite normal cyclic subgroup. \item Nil: virtually nilpotent fundamental group. \item Sol: virtually solvable fundamental group. \end{itemize} We therefore only need to consider the case where $M$ is a non-geometric prime manifold. We have $2$ cases to consider: \begin{itemize} \item $M$ is a non-geometric graph manifold. In this case, $\pi_1(M)$ acts acylindrically on the Bass-Serre tree, see \cite[Lemma 2.4]{Wilton2008ProfinitePO}. Hence $G$ is acylindrically hyperbolic. \item $M$ contains a hyperbolic component. In view of Dahmani's combination theorem \cite{Dahmani}, $\pi_1(M)$ is hyperbolic relative to infinite-index abelian and graph manifold groups and hence acylindrically hyperbolic. \end{itemize} Note that by the arguments above, if $M$ is not prime (and not $\mathbb{RP}^3 \# \mathbb{RP}^3$) or if $M$ contains a hyperbolic component then it is relatively hyperbolic with infinite-index peripherals, and so $G$ contains a super-divergent element by Lemma \ref{lem:rh_superdiv}. If $M$ is prime with geometry $\mathbb{H}^3$, then $G$ is non-elementary hyperbolic and hence contains a super-divergent element by Lemma \ref{lem:examples_of_superdivgelements}. \\ Therefore, the only acylindrically hyperbolic $3$ manifold groups left to consider are non-geometric graph manifolds. It might be worth noting that by a result of Gersten \cite{Gersten}, graph manifolds are exactly the closed $3$-manifolds with quadratic divergence, and it particular non-geometric graph manifolds do not contain super-divergent elements by Lemma \ref{divergence}. The fundamental group of a non-geometric graph manifold admits a graph of groups decomposition (the one coming from the geometric decomposition) where the edge groups are isomorphic to $\mathbb{Z}^2$. Moreover, the edge groups are undistorted. This can be seen, for example, from the fact that the universal cover of the graph manifold is quasi-isometric to that of a \emph{flip} graph manifold in a way that preserves the geometric decomposition \cite{KapovichLeeb}. Since flip graph manifolds are CAT(0), their abelian subgroups are undistorted, we see that edge groups are undistorted in fundamental groups of flip graph manifolds, and hence the same holds for any graph manifold. \end{proof} \section{Geometric arguments} \label{geometricarguments} Our proof of Theorem \ref{linear progress} (linear progress for Markov chains) is in two parts. The goal of the first part is to establish Proposition \ref{axiom} below. We do so separately for the two assumptions from Section \ref{sec:assump} and in both cases we make use of the geometric features of the groups under consideration. For the second part of the proof, which takes place in the next section and is probabilistic in nature, Proposition \ref{axiom} can be used as a black box. In particular, we note that whenever one establishes Proposition \ref{axiom} for some group $G$, then $G$ will also satisfy the conclusion of Theorem \ref{linear progress}. The content of the proposition is roughly the following. Suppose that we start a Markov path at $o$ and at some stage we get to $p$. At that point we have created certain large projections on cosets of some $E(g)$, for $g$ a fixed loxodromic WPD, and the proposition says that, continuing the Markov path past $p$, it is unlikely that we undo much of these large projections. In Subsection \ref{subsec:WPD} we introduced various objects and notations related to a loxodromic WPD $g$ and the unique maximal elementary subgroup $E(g)=\gamma$ containing $g$. In particular we introduced certain projections $\pi_\gamma$ and projection distances $d_{h\gamma}$ (see Lemma \ref{lem:Behrstock} and Notation \ref{not:d_hgamma}), certain sets of cosets $\mathcal H_T(o,p)$ where projections are large, and sums of projection distances over such sets $\sum_{\mathcal H(x,y)}[z_1,z_2]$ (Definition \ref{defn:H_T}). Once a loxodromic WPD has been fixed, we will freely refer to these objects constructed from the given loxodromic WPD. \begin{proposition} \label{axiom} Let $G$ be a group acting on a hyperbolic space $X$ satisfying one of Assumption \ref{assump:superdiv} or Assumption \ref{assump:graph}, and fix a basepoint $x_0\in X$. Then there exists a loxodromic WPD $g$ and a constant $T_0$ such that the following holds. Consider a tame Markov chain $(w^o_n)$ on $G$. For each $T\geq T_0$ there exists a constant $C$ such that for all $o,p\in G$, $n \in \mathbb{N}$, and $t>0$ we have \[ \mathbb{P}\left[\exists r\leq n : \sum_{\mathcal{H}_T(o,p)} [p,w^p_r] \geq t \right] \leq Ce^{-t/C}. \] \end{proposition} Each subsection will be about proving this Proposition for the different assumptions on $G$. \subsection{Proof of Proposition \ref{axiom} for groups containing a super-divergent element} We now fix some notation and constants. In this subsection, let $G$ be a group containing a super-divergent element $g$ for the action on a hyperbolic space $X$, and we fix the rest of the notation set above Proposition \ref{axiom}, for $g$ the given super-divergent element (which is loxodromic WPD by Lemma \ref{lem:super_div_WPD}), and $T$ chosen as described below. We also fix a word metric $d_G$ on $G$, with finite generating set $S$. Since the projections $\pi_{h\gamma}$ from Lemma \ref{lem:Behrstock} are bounded distance from $X$-projections onto $h\langle g \rangle$, the super-divergent property holds with $\pi=\pi_{h\gamma}$, with respect to the projection distances as in Notation \ref{not:d_hgamma}, meaning the following. There exist a constant $\theta > 0$ and a super-linear function $f:\mathbb{R}^{+} \to \mathbb{R}^{+}$ (possibly larger than the ones for $\langle g\rangle$) such that the following holds for all $h\in G$: for all $d >0$ and paths $p$ remaining outside of the $d$-neighbourhood of $h\gamma$, if $d_{h\gamma}(p_{-},p_{+}) > \theta$ then $\ell(p) > f(d)$. Furthermore, we fix the constant $B$ of Lemma \ref{lem:Behrstock} and we also fix any $T\geq 10(B+\theta)$ that satisfies Lemma \ref{linearorder} and such that Lemma \ref{distance} holds with $T$ replaced by $T-4B$. Finally, we set $\mathcal H(x,y)=\mathcal H_T(x,y)$. The following lemma says that with high probability, a tame Markov chain moves away at linear speed from the fixed cosets we are projecting to. \begin{lemma} \label{20francs} There exist constants $D,C_1\geq 1$ such that for all $o,p,q\in G$ and $n \in \mathbb{N}$ we have \[\mathbb{P}\left[ w_n^q \in \mathcal{N}_{n/D}\left(\bigcup_{h\gamma\in \mathcal H(o,p)}h\gamma \right) \right] \leq C_1 e^{-n/C_1}.\] \end{lemma} \begin{proof} Let $K$ be a bound on the size of the jumps of $w_n^o$, that is, let $K$ be such that $d_G(w^p_n,w^p_{n+1})\leq K$ for all $p\in G$ and $n$ (see Remark \ref{rmk:bounded_jumps}). Up to increasing $K$ we can also assume that for all $a,b\in G$ we have $d_X(ax_0,bx_0)\leq K d_G(a,b)$. We begin with the following claim, stating that for all points $q\in G$ the number of cosets from $\mathcal{H}(o,p)$ that intersect a ball of radius $Kn$ and centre $q$ is linear in $n$. \\ \textbf{Claim}: For all $n$, we have $$\# \{ h\gamma \in \mathcal H(o,p) : h\gamma \cap B(q,Kn) \neq\emptyset \} \leq 4K^2n/(T-4B)+2 $$ where $B$ is from Lemma $\ref{lem:Behrstock}$ and $T$ is the threshold for $\mathcal{H}(o,p)$, see the discussion above the statement of the lemma for details. \begin{proof}[Proof of claim] Fix $n$ and let $h_1\gamma \prec h_2\gamma \prec \dots h_r\gamma $ be the cosets such that $h_i\gamma \in \mathcal{H}(o,p) $ and $h_i\gamma \cap B(q,Kn) \neq \emptyset$ (the linear order is from Lemma \ref{linearorder}). We therefore want to show that $r \leq 4K^2n/(T-4B)+2$.\\ If $r<2$ we are done, so let us assume $r\geq 2$ and pick $a\in h_1\gamma\cap B(q,Kn)$ and $b\in h_r\gamma\cap B(q,Kn)$. For each $1<i<r$ we have $\pi_{h_i\gamma}(a)\subseteq \pi_{h_i\gamma}(h_1\gamma)=\pi_{h_i\gamma}(o)$ (where we used the equivalent characterisation of the linear order), as well as $\pi_{h_i\gamma}(b)\subseteq \pi_{h_i\gamma}(p)$. In particular, we have $d_{h_i\gamma}(a,b)\geq T-4B$ (the "$-4B$" is due to the fact that projections have diameter at most $2B$ as a consequence of Lemma \ref{lem:Behrstock}). Hence, for all $1<i<r$ we have $d_{h_i\gamma}(o,p)=d_{h_i\gamma}(x, y)$. This leads to \[ 2Kn\geq d_G(a,b) \geq d_X(a x_0, bx_0)/K\geq \frac{1}{2K} \sum_{\mathcal{H}_{T-4B}(x,y)}[x,y] \geq (r-2)(T-4B)/(2K) \] where we used Lemma \ref{distance} (and our choice of $T$). This proves the claim. \end{proof} Now, the inclusion of $E(g)$ into $G$ is a quasi-isometric embedding. Hence for each coset $h\gamma \in \mathcal H(o,p)$ there are at most $K'n$ elements of $h\gamma$ in the ball $B(q,Kn)$ for some constant $K'$. Hence, by the fact that our Markov chain is tame (and we use Definition \ref{defn:tame}-\ref{item:non-amen}), for $S$ the fixed finite generating set for $G$, we get that \begin{align*} \begin{split} \mathbb{P}\left[ w_n^q \in \mathcal{N}_{n/D}\left(\bigcup_{h\gamma\in \mathcal H(o,p)}h\gamma \right) \right] &\leq \# \{ h\gamma \in \mathcal H(o,p) : h\gamma \cap B(q,Kn) \neq\emptyset \} K'n\vert S\vert^{n/D} A \rho^n \\ &\leq (4K^2n/(T-4B)+2 ) K'n \vert S\vert^{n/D} A \rho^n \\ &\leq Mn^2(\vert S\vert^{1/D} \rho)^n \end{split} \end{align*} for some $M>0$, where $A, \rho $ are from the `non-amenability' condition from the definition of tameness (Defintion \ref{defn:tame}-\eqref{item:non-amen}). This decays exponentially if we choose $D$ large enough ($D$ such that $1/D < -log_{\vert S\vert }(\rho)$ works). \end{proof} In view of the previous lemma, it is of interest to study how projections of a path change assuming that the path moves away at linear speed from the cosets we are projecting to. We will apply the following lemma to a sample path of our Markov chain, but it holds for all discrete paths with $K$-bounded jumps for some $K$, that is, sequences of points with consecutive ones being distance at most $K$ apart. \begin{lemma} \label{o(log)} For each $D,K>0$, there exists a function $\phi : \mathbb{R}^{+} \to \mathbb{R}^{+} $ with $\phi(n)=o(\log(n))$ such that the following holds. Let $o,p\in G$ and consider a discrete path $\alpha=(\alpha_n)_{n\geq 0}$ with $K$-bounded jumps and such that for all $n \in \mathbb{N}$ we have $$d\left(\bigcup_{h\gamma \in \mathcal{H}(o,p)} h\gamma, \alpha_n\right) \geq n/D.$$ Then \[ \sum_{\mathcal{H}(o,p)} [\alpha_0,\alpha_n] \leq \phi(n).\] \end{lemma} \begin{proof} Recall that $f$ is, roughly, the function controlling super-divergence of the fixed super-divergent element. We define the sequence $(\Psi_i)_{i\geq 1}$ recursively: \\ \begin{itemize} \item Let $\Psi_1$ be such that for all $\Psi\geq \Psi_1$ we have $\frac{1}{K}f\left(\frac{\Psi}{D}-K\right)\geq 2\Psi$ (which exists since $f$ is super-linear). \item $\Psi_{j+1}=\frac{1}{K}f\left(\frac{\sum_{i\leq j}\Psi_i}{D}-K\right)$. \end{itemize} Let \[u_{j}:= \sum_{i\leq j} \Psi_{i}.\] Note that by our choice of $\Psi_1$, we have $u_j\geq\Psi_j\geq 2^{j-1} \Psi_1$ for all $j$. \par\medskip \textbf{Claim 1:} For all $j\geq 1$ and $u_j< n\leq u_{j+1}$ we have \[ \sum_{ \mathcal{H}(o,p)} [\alpha_{u_j},\alpha_{n}] \leq 4\theta. \] \begin{proof}[Proof of Claim 1] \vspace{2mm} Assume that the claim does not hold. Define $\alpha_{u,v}$ to be a concatenation of geodesics from $\alpha_i$ to $\alpha_{i+1}$ for $u\leq i<v$. We show below that there exists $h\gamma$ such that $d_{h\gamma}(\alpha_{u_j},\alpha_{n}) \geq \theta$. This is enough to prove the claim. Indeed, since $\alpha_{u_j,n} $ stays outside of the $(u_j/D-K)$-neighborhood of $h\gamma$, by super-divergence of $g$, this leads to \[K( u_{j+1}-u_j) \geq \ell(\alpha_{u_j,n}) > f\left(\frac{u_j}{D}-K\right)= K\Psi_{j+1}. \] a contradiction as $u_{j+1}-u_j= \Psi_{j+1}$. \\ Therefore to prove the claim it remains to show that there exists an $h\gamma$ such that $d_{h\gamma}(\alpha_{u_j},\alpha_{n}) \geq \theta$. Note that if for some $h\gamma\in\mathcal H(o,p)$ we have $\pi_{h\gamma}(\alpha_{u_j})=\pi_{h\gamma}(o)$ and $\pi_{h\gamma}(\alpha_{n})=\pi_{h\gamma}(p)$ (or vice versa), then we are done since $T\geq \theta$. Hence, assume that this is not the case. By Lemma \ref{lem:2cosetsmax}, there are at most 4 cosets $h\gamma\in \mathcal H(o,p)$ where one of $\pi_{h\gamma}(\alpha_{u_j})$ and $\pi_{h\gamma}(\alpha_{n})$ does not coincide with either $\pi_{h\gamma}(o)$ or $\pi_{h\gamma}(p)$. We have that $\sum_{ \mathcal{H}(o,p)} [\alpha_{u_j},\alpha_{n}]$ is in fact a sum over the aforementioned cosets, and since there are at most 4 of them, one of the cosets satisfies the required property. \end{proof} In view of Lemma \ref{lem:coarsely_lip_sum}, there exists $L$ (independent of $\alpha$, $o$, $p$) such that for all $i$ we have $\sum_{\mathcal H(o,p)}[\alpha_i,\alpha_{i+1}]\leq L$. We note that this $L$ is actually bigger than the one from Lemma \ref{lem:coarsely_lip_sum} and depends on $K$. In view of the claim we have, for $\phi(n):=\max \{j: u_j \leq n \}$, \[ \sum_{\mathcal{H}(o,p)} [\alpha_0,\alpha_n] \leq\sum_{\mathcal{H}(o,p)} [\alpha_0,\alpha_{u_1}]+\dots \sum_{\mathcal{H}(o,p)} [\alpha_{u_{\phi(n)-1}},\alpha_{n}] \leq L\Psi_1+ 4\theta\phi(n). \] Therefore, we are done once we prove the following claim. \textbf{Claim 2:} \[\phi(n)=o(\log(n))\] \begin{proof}[Proof of Claim 2] It suffices to show that for all $E>1$ we have $u_j\geq E^j$ for all sufficiently large $j$, which implies that $\phi(n)\leq \log_E(n)$ for all sufficiently large $n$. We have $u_{j+1}=u_j+g(u_j)$ for $g(x):= \frac{1}{K}f(\frac{x}{D}-K)$, and recall that $u_j\geq 2^{j-1}\psi_1$. Since $g$ is super-linear, for all sufficiently large $j$ we have $g(u_j)\geq 2E u_j$, and hence for those same $j$ we have $u_{j+1}\geq 2E u_j$, and the claim follows. \end{proof} This proves the Lemma. \end{proof} The consequence of the lemma that we will need is the following Corollary, which says, roughly, that it is very unlikely that projections of a sample path of our Markov chain change faster than logarithmically in the number of steps. \begin{corollary} \label{undoproba} For all $m, m' >0$, there exists a constant $C_2 >0$ such that the following holds. Let $o,p \in G$ be basepoints and $w^{o}_r$ a tame Markov chain. Then for all $t>0$ and $k \leq e^{mt/ \eta}$, where $\eta$ is as in Lemma \ref{etalogproj}, we have \[ \mathbb{P}\left[ \exists r \leq k : \sum_{\mathcal{H}(o,p)} [p,w^p_r] \geq t/m' \right] \leq C_2 e^{-t/C_2}. \] \end{corollary} \begin{proof} For all $t>0$, we let $i_t = \lfloor \frac{t}{(m'+1)LK} \rfloor $ where $L$ is from Lemma \ref{lem:coarsely_lip_sum} and $K$ is a bound on the size of the jumps of the Markov chain in $X$, meaning that $d_X(w_n^qx_0,w_{n+1}^qx_0)\leq K$ for all $q\in G$ and $n$ (see Remark \ref{rmk:bounded_jumps}). Let $A_{r,t}$ be the event ``for all $i_t \leq n \leq r : d\left(w^p_{n}, \bigcup_{h\gamma \in \mathcal{H}(o,p)} h\gamma\right) \geq n/D$", where $D$ is from Lemma \ref{20francs}. \\ \textbf{Claim:} There exists $t_1$ such that for $t \geq t_1$ and $r \leq e^{mt/\eta}$, the event $A_{r,t}$ implies that $$\sum_{\mathcal{H}(o,p)} [p,w^p_r] < t/m'.$$ \begin{proof}[Proof of Claim] We note that the $K$-discrete path $(w^p_r)_{r\geq i_t}$ satisfies Lemma \ref{o(log)}. Hence there exists a function (independent of $o,p$) $\phi(n)=o(\log(n))$ such that if $A_{r,t}$ holds then $\sum_{\mathcal{H}(o,p)} [w^p_{i_t},w^p_r] \leq \phi(r-i_t)$. Using the triangular inequality (Remark \ref{rem:triangle_inequality}), the coarsely Lipschitz property (Lemma \ref{lem:coarsely_lip_sum}) and the fact that $(w^p_n)_{n\geq 0}$ has $K$-bounded jumps, we get the following: \begin{align*} \begin{split} \sum_{\mathcal{H}(o,p)} [p,w^p_r] &\leq \sum_{\mathcal{H}(o,p)} [p,w^p_{i_t}]+\sum_{\mathcal{H}(o,p)} [w^p_{i_t},w^p_r] \\ & \leq L d_X(px_0,w^p_{i_t}x_0)+L+ \phi(r-i_t) \\ & \leq LKi_t+L+\phi(r-i_t) \\ \end{split} \end{align*} Now by the fact that $\phi(r-i_t) =o(\log(r-i_t)) $ we can find $t_1$ such that for all $t \geq t_1$ : $$ t/m'-LKi_t-L > \phi(r-i_t) $$ Hence, for $t\geq t_1$ we get that $\sum_{\mathcal{H}(o,p)} [p,w^p_r] < t/m'$ as required. \end{proof} By the claim, for $t\geq t_1$ we can have $\sum_{\mathcal{H}(o,p)} [p,w^p_r] \geq t/m'$ only in $A_{r,t}^c$. Also, we have $$\mathbb P\left[\bigcup_r A_{r,t}^c\right] \leq \sum_{i\geq i_t} C_1 e^{-i}/C_1\leq C_2 e^{-i_t/C_2},$$ for $C_1$ as in Lemma \ref{20francs} and some $C_2$ depending on $C_1$. Therefore, $$\mathbb{P}\left[ \exists r \leq k : \sum_{\mathcal{H}(o,p)} [p,w^p_r] \geq t/m' \right] \leq \mathbb P\left[\bigcup_r A_{r,t}^c\right]\leq C_2 e^{-i_t/C_2}\leq C_2e^{- t /(m'+1)KLC_2}e^{1/C_2},$$ as required for $t\geq t_1$. To cover the case $t\leq t_1$ we can just increase the constant resulting from the previous computation. \end{proof} Finally, the following lemma will be used to say that if $h\gamma, h'\gamma \in \mathcal{H}(o,p)$ satisfy $h\gamma \prec h'\gamma$, then before the tame Markov chain can ``undo" projections on $h\gamma$ it first has to "undo" the projections on $h'\gamma$. (This is just a version of Lemma \ref{linearorder}, with the same proof.) \begin{lemma} \label{close} Let $o,p$ be basepoints and $h\gamma, h'\gamma \in \mathcal{H}(o,p)$ with $h\gamma \prec h'\gamma$. Then for all $x \in G$, if $d_{h\gamma}(p,x) > B$ then $\pi_{h'\gamma}(o)=\pi_{h'\gamma}(x)$. \end{lemma} \begin{proof} By the equivalent characterisation of the linear order from Lemma \ref{linearorder} on $\mathcal H(o,p)$, we have that $\pi_{h\gamma}(p)=\pi_{h\gamma}(h'\gamma)$. Hence $d_{h\gamma}(x,h'\gamma)=d_{h\gamma}(x,p)>B$ and by the strong Behrstock inequality (Lemma \ref{lem:Behrstock}), we get that \[ \pi_{h'\gamma}(x)=\pi_{h'\gamma}(h\gamma)=\pi_{h'\gamma}(o), \] where we again use the linear order on $\mathcal H(o,p)$. \end{proof} \begin{proof}[ Proof of Proposition \ref{axiom} for groups containing a super-divergent element] Recall that we have to prove that there exists a constant $C$ such that for all $o,p\in G$, $n \in \mathbb{N}$, and $t>0$ we have $$ \mathbb{P}\left[\exists r\leq n : \sum_{\mathcal{H}_T(o,p)} [p,w^p_r] \geq t \right] \leq Ce^{-t/C}. $$ The idea of the proof is the following. As we run the Markov chain starting at $p$, it might start creating projections on the cosets in $\mathcal H(o,p)$, but it is unlikely that this happens fast by Corollary \ref{undoproba}. At the same time, it is likely that the Markov chain will create a new projection, by Lemma \ref{etalogproj}. Before creating more projections on the cosets in $\mathcal H(o,p)$, the Markov chain would have to undo said new projection, and this can be turned into a sort of recursive formula which will yield the desired bound. We now proceed to the actual proof.\\ Let $t\geq 10T$ and define $$p_{n,t}=\sup_{o,p \in G} \mathbb P\Big[\quad \exists r\leq n \quad : \sum_{\mathcal{H}(o,p)} [p,w^p_r] \geq t\Big] $$\\ We will show that there exists a constant $C>0$ such that $p_{n,t} \leq Ce^{-t/C}$, which is exactly the required inequality (except for the requirement "$t\geq 10 T$" but we can increase the constant to cover the other case).\\ The main claim is the following.\\ {\bf Claim.} There exists $C'$ such that we have $p_{n,t}\leq C' e^{-t/C'}+p_{n,2t}$. \begin{proof}[Proof of Claim] Let $\epsilon>0$ be arbitrary. Choose $o,p$ such that $$p_{n,t}\leq (1+\epsilon) \mathbb P\Big[\quad \exists r\leq n \quad : \sum_{\mathcal{H}(o,p)} [p,w^p_r] \geq t\Big].$$ For $o,p \in G, t>0$, let $\mathcal A_{o,p,t} \subseteq G$ be the set of all $q\in G$ such that: \begin{enumerate} \item There exists $h\gamma$ such that $d_{h\gamma}(p,q)\geq 5t$. \item $$\sum_{ \mathcal H(o,p)} [p,q]\leq t/2.$$ \end{enumerate} \begin{figure}[h] \centering \begin{tikzpicture}[scale=0.6] \draw[black] (0,0) to[out=20,in=160] (2.5,0); \draw[black] (3,0) to[out=20,in=160] (5.5,0); \draw[black] (-3,0) to[out=20,in=160] (-0.5,0); \draw[black] (6,0) to[out=20,in=160] (8.5,0); \draw[black] (9,0) to[out=20,in=160] (11.5,0); \draw[red] (7,2) to[out=100,in=230] (7.5,7); \draw[red] (8,7) node{$h_q\gamma$}; \filldraw[black] (5,7.5) circle (1pt) node[anchor=east]{$q$}; \draw[dashed] (5,7.5) --(6.9,6); \draw[blue] (8.9,4.4) node {$\geq 5t$}; \filldraw[black] (6.9,6) circle (1pt) node[anchor=west]{$\pi_{h_q\gamma}(q)$}; \filldraw[black] (6.9,2.5) circle (1pt) node[anchor=west]{$\pi_{h_q\gamma}(p)$}; \filldraw[black] (-3,2.5) circle (1pt) node[anchor=east]{$o$}; \filldraw[black] (12,2.5) circle (1pt) node[anchor=east]{$p$}; \filldraw[black] (11.6,0) circle (1pt) node[anchor=west]{$\pi_{h'\gamma}(p)$}; \draw[dashed] (-3,2.5) --(-2.5,0.2); \draw[dashed] (12,2.5) --(11.6,0); \draw[dashed] (5,7.5) --(7,0.2); \filldraw[black] (7,0.2) circle (1pt) node[anchor=north]{$\pi_{h\gamma}(q)$}; \draw [decorate,decoration={brace,amplitude=5pt,mirror,raise=4ex}] (7,0.2) -- (11.6,0.2) node[midway,yshift=-3em]{$\sum_{\mathcal H(o,p)}[p,q] \leq t/2$}; \draw [decorate,decoration={brace,amplitude=5pt,raise=4ex}, blue] (6.9,5.8) -- (6.9,2.5) node[midway,yshift=-3em]{}; \end{tikzpicture} \caption{Illustration of $q\in A_{o,p,t}$} \end{figure} We consider the following events, where $\eta$ is from Lemma \ref{etalogproj}. \begin{enumerate} \item $\mathcal B^{t}_1$ : `` $\forall k \leq e^{5t/\eta}, \quad \forall h \in G : \quad d_{h\gamma}(p,w^p_{k}) <5t $". This is the event that the Markov chain is not creating a new projection quickly. \\ \item $\mathcal B^{t}_2$ : ``$\exists k <e^{5t/\eta} : \quad \sum_{\mathcal H(o,p)} [p,w^p_k] > t/2 $". This is the event that the Markov chain is undoing $t/2$-worth of projections on the original cosets $\mathcal H (o,p)$ quickly.\\ \item Let $\mathcal B^{t}_3(k,q)$ be the event ``$w^p_k=q$ and for all $i<k : w^p_i \not\in \mathcal A_{o,p,t}$ ". \end{enumerate} We call $B_3^t=\bigsqcup_{\substack{k \leq e^{5t/\eta} \\ q\in \mathcal A_{o,p,t}}} \mathcal B^{t}_3(k,q)$. \par\medskip \textbf{Subclaim: } Let $q\in \mathcal A_{o,p,t}$ and let $w \in G $ be such that $\sum_{\mathcal H(o,p)}[p,w]\geq t $. Then $\sum_{\mathcal H(o,q)}[q,w] \geq 2t$. \begin{proof}[Proof of Subclaim] We let $h_q \gamma$ be the coset satisfying the first item in the definition of $q$ belonging to the set $\mathcal A_{o,p,t}$. We first note that $h_q \gamma \not\in \mathcal H(o,p)$. Indeed, if it was then $\sum_{\mathcal H(o,p)}[p,q] \geq 5t$, a contradiction with the second item in the definition of $\mathcal A_{o,p,t}$. Hence $d_{h_q\gamma}(o,p) <T $. To prove the claim, it is therefore enough to show that $d_{h_q\gamma}(q,w) \geq 2t$, as $h_q \gamma \in \mathcal H(o,q)$. Let $k\gamma \in \mathcal H(o,p)$ be any coset. Then either $d_{k\gamma}(o,h_q\gamma)>B$ or $d_{k\gamma}(p,h_q\gamma)>B$ and so by the strong Behrstock inequality (Lemma \ref{lem:Behrstock}) we get that either $\pi_{h_q\gamma}(o)=\pi_{h_q\gamma}(k\gamma)$ or $\pi_{h_q\gamma}(p)=\pi_{h_q\gamma}(k\gamma)$. \\ For a contradiction, assume that $d_{h_q\gamma}(q,w) < 2t$. Then $$d_{h_q\gamma}(p,w) \geq d_{h_q\gamma}(p,q)-d_{h_q\gamma}(w,q) \geq 3t,$$ so that $d_{h_q\gamma}(o,w)\geq 3t-T$. The cases $\pi_{h_q\gamma}(p)=\pi_{h_q\gamma}(k\gamma)$ or $\pi_{h_q\gamma}(o)=\pi_{h_q\gamma}(k\gamma)$ are similar so assume that $\pi_{h_q\gamma}(o)=\pi_{h_q\gamma}(k\gamma)$. Then $d_{h_q\gamma}(w,k\gamma)=d_{h_q\gamma}(w,o)\geq 3t-T>B$ and so by the strong Behrstock inequality (Lemma \ref{lem:Behrstock}), this leads to $$\pi_{k\gamma}(w)=\pi_{k\gamma}(h_q\gamma).$$ We also have that $d_{h_q\gamma}(q,k\gamma) = d_{h_q\gamma}(q,o)>5t-T >B$. Again by the strong Behrstock inequality, this leads to $$\pi_{k\gamma}(h_q\gamma)=\pi_{k\gamma}(q)$$ for all $k\gamma \in \mathcal H(o,p)$ and so $ \pi_{k\gamma}(q)=\pi_{k\gamma}(w)$ for all $k\gamma \in \mathcal H(o,p)$. Hence, $$ t\leq \sum_{\mathcal H(o,p)} [p,w] \leq \sum_{\mathcal H(o,p)} [p,q] \leq t/2, $$ a contradiction. Hence $d_{h_q\gamma}(q,w) \geq 2t$ and in particular $\sum_{\mathcal H(o,q)}[q,w] \geq 2t$, proving the claim. \end{proof} We note that the event ``$\exists r\leq n : \sum_{\mathcal{H}(o,p)} [p,w^p_r] \geq t$" is contained in the union of the events $\mathcal B^{t}_1$, $\mathcal B^{t}_2$, and the intersection of $B^t_3$ with the event ``$\exists e^{5t/\eta} \leq r\leq n : \sum_{\mathcal{H}(o,p)} [p,w^p_r] \geq t$". This leads to the following chain of inequalities, which we explain further below and where the constant $C_2$ is from Corollary \ref{undoproba} with $m=5$ and $m'=2$. \begin{align*} \begin{split} \mathbb P\Big[\exists r\leq n : &\sum_{\mathcal{H}(o,p)} [p,w^p_r] \geq t\Big] \\ & \leq \mathbb{P}[\mathcal B^{t}_1] + \mathbb{P}[\mathcal B^{t}_2]+ \sum_{\substack{k \leq e^{5t/\eta} \\ q\in \mathcal A_{o,p,t} }} \mathbb P\left[\exists e^{5t/\eta} \leq r\leq n : \sum_{\mathcal{H}(o,p)} [p,w^p_r] \geq t \quad \Big\vert \quad \mathcal B^{t}_3(q,k) \right] \mathbb P \Big[ B_3^t(k,q)\Big] \\ & \leq \mathbb{P}[\mathcal B^{t}_1] + \mathbb{P}[\mathcal B^{t}_2]+ \sum_{\substack{k \leq e^{5t/\eta} \\ q\in \mathcal A_{o,p,t} }} \mathbb P\left[\exists e^{5t/\eta} \leq r\leq n : \sum_{\mathcal{H}(o,q)} [q,w^p_r] \geq 2t \quad \Big\vert \quad \mathcal B^{t}_3(q,k) \right] \mathbb P \Big[ B_3^t(k,q)\Big] \\ &\leq \mathbb{P}[\mathcal B^{t}_1] + \mathbb{P}[\mathcal B^{t}_2]+ \sum_{\substack{k \leq e^{5t/\eta} \\ q\in \mathcal A_{o,p,t} }} \mathbb P\left[\exists r'\leq n-k : \sum_{\mathcal H(o,q)}[q,w_{r'}^q] \geq 2t \right] \mathbb P \Big[w^p_k=q\Big] \\ &\leq e^{-\sqrt{e^{5t/\eta}}/U} + C_2e^{-t/C_2}+p_{n,2t} \end{split} \end{align*} The first inequality is a union bound plus conditioning. We then use the subclaim for the second inequality. The third inequality follows from the usual Markov property (which says that we can just condition on "$w^p_k=q$" instead) and Lemma \ref{lem:Markov_property_set}. We then use Lemma \ref{etalogproj} to get the term ``$e^{-\sqrt{e^{5t/\eta}}/U}$" and Corollary \ref{undoproba} (with $m=5$ and $m'=2$) to get the term ``$C_2e^{-t/C_2}$". The last probability is from the definition of $p_{n,t}$ which is defined as the supremum over all basepoints $o,p$. Hence, $$ p_{n,t} \leq (1+\epsilon)(e^{-\sqrt{e^{5t/\eta}}/U} + C_2e^{-t/C_2}+p_{n,2t}) \leq (1+\epsilon)C'e^{-t/C'}+(1+\epsilon)p_{n,2t} $$ for some $C'>0$ depending on $C_2,\eta,U$. Since $\epsilon$ was arbitrary, this proves the claim.\\ \end{proof} We note that for a fixed $n$, if $u>L(Kn+1)$ then $p_{n,u}=0$ where $K$ is from the bounded jumps condition (see Remark \ref{rmk:bounded_jumps}) and $L$ is the coarsely Lipschitz constant from Lemma \ref{lem:coarsely_lip_sum}. Let $s= \lceil \log_2(L(Kn+1)/t) \rceil$ so that $p_{n,2^st}=0$. Inductively, we get \begin{align*} \begin{split} p_{n,t} &\leq C'\sum_{i=1}^{s} e^{-2^{i}t/C'} \\ & \leq C'\sum_{i=1}^{\infty} e^{-it/C'} \\ & \leq Ce^{-t/C}, \end{split} \end{align*} for some $C$ depending on $C'$, as we wanted. This proves Proposition \ref{axiom} for groups containing a super-divergent element. \end{proof} We proved Proposition \ref{axiom} for groups containing a super divergent element. We will now show that this result also holds for some particular graphs of groups. \subsection{Proof of Proposition \ref{axiom} for graph of groups with subexponential growth edge groups} In this subsection let $G = \pi_1( \mathcal{G})$ be the fundamental group of a graph of groups $\mathcal{G}$ where each edge group has subexponential growth, in a fixed ambient word metric $d_G$, and fix any loxodromic WPD $g$ for the action on the Bass-Serre tree $X$, for which we also fix a basepoint $x_0$. We also fix the rest of the notation set above Proposition \ref{axiom} for the given $g$, for some fixed $T\geq 10B$, for $B$ as in Lemma \ref{lem:Behrstock}, that satisfies both Lemma \ref{linearorder} and Lemma \ref{distance}. We also assume that $T\geq 10M$ for $M$ as in Lemma \ref{lem:X_e_proj} below (we will not use $T$ between now and then). We set $\mathcal H(x,y)=\mathcal H_T(x,y)$. \subsubsection{Bass-Serre theory preliminaries} \label{subsec:BS_prelims} Since $\gamma=E(g)$ is virtually cyclic, with $g\in G$ acting loxodromically, it stabilises a line in the Bass-Serre tree. Similarly: \begin{definition} For every coset $h \gamma$, there is a bi-infinite line which is stabilised by $h\gamma h^{-1}$. We call this line the \textit{axis} of $h\gamma$ and denote it by $Axis(h\gamma)$. Note that $Axis(h\gamma)=h Axis(\gamma)$ for al $h$. \end{definition} \begin{lemma} \label{lem:geodesic_in_axis} There exists a constant $D$, depending only on the loxodromic WPD $g$ such that the following holds. If $a,b \in G$ are such that $d_{h\gamma}(a,b) >D$ then any geodesic in $X$ from $ax_0$ to $bx_0$ contains an edge in the axis of $h\gamma$, and moreover this edge can be chosen at distance at least $d_{h\gamma}(a,b)/10$ from $ax_0$. \end{lemma} \begin{proof} Note that there exists $C$ such that for all $h\in G$ we have $d_{\text{Haus}}(h \gamma x_0, Axis(h\gamma)) \leq C$. In fact, $\gamma x_0$ and $Axis(\gamma)$ both lie within finite Hausdorff distance of an orbit of $\langle g\rangle$ (and proving the required bound for $h=1$ suffices). Hence, if $D$ is sufficiently large and $d_{h\gamma}(a,b)\geq D$ then $ax_0$ and $bx_0$ have closest-point projections onto $Axis(h\gamma)$ that are at least $d_{h\gamma}(a,b)/2$ apart; this is because the projections used to define $d_{h\gamma}$ are within bounded distance of closest-point projections to $h\gamma x_0$, and in turn those lie within bounded distance of closest-point projections to $Axis(h\gamma)$. Since $Axis(h\gamma)$ is a bi-infinite geodesic line in a tree, we have that the geodesic $[ax_0, bx_0]$ contains the subgeodesic of $Axis(h\gamma)$ connecting the closest-point projections of $ax_0$ and $bx_0$. Said subgeodesic contains an edge at distance $\geq d_{h\gamma}(a,b)/10$ from the closest-point projection $a'$ of $ax_0$ to $Axis(h\gamma)$, and this edge will also have distance $\geq d_{h\gamma}(a,b)/10$ from $ax_0$, as required. \end{proof} We now establish some notation related to Bass-Serre theory. In short, to each vertex and edge of the Bass-Serre tree we associate a suitable subspace of $G$. We fix orbit representatives $v_1,\dots,v_a$ for the action of $G$ on the vertices of $X$, and similarly let $e_1,\dots,e_b$ be orbit representatives for the $G$-action on the edges. We let $G_{v_i}=\Stab(v_i)$ and $G_{e_i}=\Stab(e_i)$. Corresponding to each vertex $v$ of $X$ we have a coset of one of the $G_{v_i}$, which we denote by $X_v$ (namely, if $v=gv_i$ then $X_v=gG_{v_i}$). Similarly, we can associate to each edge $e$ a coset $X_e$ of some $G_{e_i}$. We define $\Psi: \mathbb{R} \to \mathbb{R}$ to be the maximal of the all the growth functions $\psi_{e_i}$ of the $G_{e_i}$ with respect to the ambient word metric $d_G$, i.e. $\Psi(n):= \max \{ \psi_{e_i}(n) \} $ (see Definition \ref{defn:subexp}). This function is well-defined as $E(\Gamma)$ is finite, and is subexponential as each $\psi_e$ is subexponential.\\ Now, for all $K$ there exists a constant $K'$ such that the following holds. Any $K$-path in the Cayley graph, from $a$ to $b$ intersects $\mathcal N_{K'}(X_e)$ for all $X_e \in \mathcal I_{a,b}$, where $$\mathcal I_{a,b} = \{X_e \quad \vert \quad e \in [v,v'] \}$$ where $v,v'$ are vertices of the Bass-Serre tree $X$ such that $a \in X_v$ and $b \in X_{v'}$, and $[v,v']$ is the (unique) geodesic between $v $ and $v'$ in $X$.\\ Finally, we note the following technical lemma. \begin{lemma} \label{lem:X_e_proj} There exists $M>0$ such that the following holds. Let $e$ be an edge of $Axis(h\gamma)$ for some $h\in G$. Then for all $h'\gamma\neq h\gamma$ and all $a\in X_e$ we have $d_{h'\gamma}(x,h\gamma)\leq M$. \end{lemma} \begin{proof} Note that there is a uniform bound on the distance between $ax_0$ and $e$, and in turn $e$ lies uniformly close to $h\gamma x_0$ as noted at the beginning of the proof of Lemma \ref{lem:geodesic_in_axis}. Hence, the claim follows from the fact that closest-point projections to quasi-convex sets in hyperbolic spaces are coarsely Lipschitz. \end{proof} \subsubsection{Back to the proof} Recall that we have to prove that there exists a constant $C$ such that for all $o,p\in G$, $n \in \mathbb{N}$, and $t>0$ we have $$\mathbb{P}\left[\exists r\leq n : \sum_{\mathcal{H}_T(o,p)} [p,w^p_r] \geq t \right] \leq Ce^{-t/C}. $$ We note the following claim which will allow us to reduce the probability we are trying to bound to the problem of bounding the probability of two events. Let $o,p \in G$ be any two basepoints. \\ \textbf{Claim:} For any $w \in G$, if $\sum_{\mathcal H_T (o,p)}[p,w] \geq t$ then either \begin{itemize} \item there exists a coset $h\gamma \in \mathcal H_T(o,p)$ such that $d_{h\gamma}(p,w) \geq t/10$, or \item $$\sum_{\substack{h\gamma \in \mathcal H_T(o,p) \\ \pi_{h\gamma}(o) =\pi_{h\gamma}(w)}}[p,w] \geq t/20.$$ \end{itemize} \begin{proof}[Proof of Claim] We note that by Lemma \ref{lem:2cosetsmax} there are most 2 cosets such that $\pi_{h\gamma}(w) \neq \pi_{h\gamma}(o), \pi_{h\gamma}(p) $, say $h_1\gamma$ and $h_2\gamma$. Hence, if the first case does not hold we have \begin{align*} \begin{split} \sum_{\mathcal H_T (o,p)}[p,w] &= \sum_{\substack{h\gamma \in \mathcal H_T(o,p) \\ \pi_{h\gamma}(o) =\pi_{h\gamma}(w)}}[p,w] + \sum_{\substack{h\gamma \in \mathcal H_T(o,p) \\ \pi_{h\gamma}(o) \neq \pi_{h\gamma}(w)}}[p,w] \\ & =\sum_{\substack{h\gamma \in \mathcal H_T(o,p) \\ \pi_{h\gamma}(o) =\pi_{h\gamma}(w)}}[p,w]+d_{h_1\gamma}(p,w)+d_{h_2\gamma}(p,w) \\ & \leq \sum_{\substack{h\gamma \in \mathcal H_T(o,p) \\ \pi_{h\gamma}(o) =\pi_{h\gamma}(w)}}[p,w]+t/5 \end{split} \end{align*} Hence if $\sum_{\mathcal H_T (o,p)}[p,w] \geq t$, the second case holds, proving the claim. \end{proof} We let $A_1 \subseteq G$ be the set of elements $w$ that satisfy the first bullet point of the claim and $A_2 \subseteq G$ the set of elements satisfying the second bullet point. Hence, by the claim if $\sum_{\mathcal H_T (o,p)}[p,w] \geq t$ then $w \in A_1 \cup A_2$. Therefore, to prove the proposition it suffices to find a constant $C$ (independent of $p$ and $n$) such that we have $$\mathbb P[\exists r\leq n : w^p_r\in A_1\cup A_2]\leq C e^{-n/C}.$$ The goal is now to show, roughly, that if the event from the equation above holds then the Markov path must have passed close to some edge space corresponding to an edge contained in some $Axis(h\gamma)$ for $h\gamma\in \mathcal H(o,p)$. From now on we always assume that $t$ satisfies $t \geq 10 D$ where $D$ is from Lemma \ref{lem:geodesic_in_axis} (as usual, it suffices to consider $t$ large enough). Also, let $K$ be a bound on the size of the jumps of the Markov chain (see Remark \ref{rmk:bounded_jumps}), and let $K'$ be the corresponding constant from the end of Subsection \ref{subsec:BS_prelims}. For all $h\gamma \in \mathcal H (o,p)$ let $$ \mathcal I_{h\gamma}=\{X_e \quad \vert \quad e \in Axis(h\gamma)\}. $$ Further, we let $$ \mathcal {NI} =\bigcup_{h\gamma \in \mathcal H (o,p)}\bigcup_{X_e \in \mathcal I_{h\gamma}} \mathcal N_{K'}(X_e)\subseteq G, $$ where we emphasise that the neighborhood is taken in $G$. \par\medskip {\bf Claim:} There exists $\epsilon>0$ (independent of $t$) such that if the event "$\exists r\leq n : w^p_r\in A_1\cup A_2$" holds, then the event "$\exists \epsilon t \leq k\leq n: w^p_k \in \mathcal {NI}$" also holds. \begin{proof} If $w^p_r\in A_1$ for some $r$, then by Lemma \ref{lem:geodesic_in_axis} we have that a geodesic $[p,w_p^r]$ contains an edge $e$ of $Axis(h\gamma)$ for some $h\gamma\in \mathcal H(o,p)$. Moreover, said edge lies at distance bounded from below in terms of $t$ from $px_0$. By the defining property of $K'$ we have that $w^p_k \in X_e$ for some $k$, and further $k$ needs to be bounded from below in terms of $t$ because the Markov chain makes bounded jumps and the distance from $p$ to $X_e$ can be bounded linearly from below in terms of the corresponding distance in $X$. The proof in the case where $w^p_r\in A_2$ for some $r$ is similar, but in this case the choice of $e$ is trickier. We consider a minimal element $h\gamma$ (with respect to the order from Lemma \ref{linearorder}) of $\{h'\gamma\in \mathcal H(o,p): \pi_{h'\gamma}(o)=\pi_{h'\gamma}(w^p_r)\}$. Again by Lemma \ref{lem:geodesic_in_axis} we have that a geodesic $[p,w_r^p]$ contains an edge $e$ of $Axis(h\gamma)$. As above, we have to prove that the distance from $p$ to $X_e$ is large. Actually, if $d_{h\gamma}(p,w^p_r)\geq t/40$ we can apply the same argument as above, so we will further assume that this is not the case. Hence $$\sum_{\substack{h'\gamma \neq h\gamma \\ \pi_{h\gamma}(o) =\pi_{h\gamma}(w^p_r)}}[p,w^p_r] \geq t/40,$$ and in view of Lemma \ref{lem:X_e_proj} (and the fact that $T$ is much larger than the constant from that lemma) we also have a similar bound on the corresponding sum for any given element of $X_e$. Hence, we can conclude using Lemma \ref{distance}. \end{proof} Now, an argument similar to the Claim in the proof of Lemma \ref{20francs} yields that there exists $M$ such that for each $n\geq 1$ we have $$\mathcal J_n=\bigcup_{h\gamma \in \mathcal H (o,p)} \{e\in I_{h\gamma}: X_e\cap B(p,Kn+K')\neq \emptyset\}\leq Mn.$$ Setting $\mathcal NI_n=\bigcup_{e\in \mathcal J_n} \mathcal N_{K'}(X_e\cap B(p,Kn+K'))$ we have that the events "$\exists \epsilon t \leq k\leq n: w^p_k \in \mathcal {NI}$" and "$\exists \epsilon t \leq k\leq n: w^p_k \in \mathcal {NI}_k$" coincide, because of the bounded jumps of the tame Markov chain. Noticing that $|X_e\cap B(p,Kn+K')|\leq \psi(M'n)$ for some $M'$ (where recall that $\psi$ is an upper bound for the growth functions of the edge groups), we can now estimate \begin{align*} \begin{split} \mathbb P \left[ \exists r \leq n \quad : \sum_{\mathcal H (o,p)}[p,w^p_r] \geq t \right] &\leq \mathbb P \Big[ \exists \epsilon t \leq k\leq n: w^p_k \in \mathcal {NI}_n\Big]\\ &\leq \sum_{i\geq \epsilon t} A\rho^{i} Mi \psi(M'i). \end{split} \end{align*} where $A,\rho$ are from the non-amenability condition on the Markov chain, Definition \ref{defn:tame}-\ref{item:non-amen}. Since $\psi$ is subexponential, the last term decays exponentially fast in $t$, as required. This concludes the proof of Proposition \ref{axiom} under Assumption \ref{assump:graph}. \section{Probabilistic arguments} \label{probabilisticarguments} The main result of this section is that if Proposition \ref{axiom} holds, then we get that a tame Markov chain makes linear progress in the hyperbolic space being acted on. \subsection{Proof of Theorem \ref{linear progress}} We fix the setup of Theorem \ref{linear progress}, and in particular a group $G$ acting on a hyperbolic space $X$ and a tame Markov chain on $G$ starting at $o\in G$ (note that the constants below will be independent of $o$, as stated in the theorem). We fix a $T \geq T_0$, for $T_0$ as in Proposition \ref{axiom}. We also require that $T$ satisfies both Lemma $\ref{linearorder}$ and Lemma $\ref{distance}$. As $T$ is fixed, for ease of notation, we drop the subscript throughout this proof, i.e. $\mathcal H (o,p) := \mathcal{H}_T (o,p)$. \\ We will show that there exist constant $L>0,C_0 \geq 1$ and $m \in \mathbb N_{>0}$ such that for all integers $j$ we have $$ \mathbb{P}\left[ \sum_{\mathcal{H}(o,w_{jm})}[o,w_{jm}] < jm/C_0\right]\leq C_0e^{-jm/C_0}, $$ This is enough to prove Theorem \ref{linear progress} as we now argue. By Lemma \ref{distance}, we get \begin{equation} \label{star} \mathbb{P}\Big[ d_{X}(x_0, w_{jm} x_0) < jm/C'_0 \Big] \leq \mathbb{P}\left[ \sum_{\mathcal{H}(o,w_{jm})}[o,w_{jm}] < jm/C_0\right] \leq C_0e^{-jm/C_0} \end{equation} where $C'_0=C_0/2$, and we dropped the basepoint from the notation for simplicity. Now, let $n$ be any positive integer, which we can assume to be large enough that $n/(C_0) \geq Km$, where $K$ is a bound on the size of the jumps of the Markov chain (Remark \ref{rmk:bounded_jumps}). Write $n=jm+r$ where $0 \leq r \leq m-1$, we note that $d_{X}(x_0, w_nx_0) \geq d_{X}(x_0, w_{jm}x_0)-d_{X}(w_{jm}x_0, w_nx_0)$. Hence \begin{equation*} \begin{split} \mathbb{P}\big[ d_X(x_0,w_nx_0) < n/(2C_0) \big]& \leq \mathbb{P}\big[ d_{X}(x_0, w_{jm}x_0)< n/C_0 \big]\\ & \leq C_0e^{-jm/C_0} \leq Ce^{-n/C}, \end{split} \end{equation*} for a suitable $C$, as required. We now prove a similar result to the Claim in the proof of \cite[Theorem 9.1]{MathieuSisto}. \begin{proposition} \label{prop:exp_lambda} \label{expectation} Let $G$ be a group satisfying the conclusion of Proposition \ref{axiom} with $T_0$ defined as in Proposition \ref{axiom}. Consider a tame Markov chain $(w^o_n)$ on $G$. Then for all $T \geq T_0$, there exist constants $ \lambda, \epsilon_1 >0$ and $m \in \mathbb N$ such that the following holds. For all $o,p\in G$: $$ \mathbb{E}\left[\exp\left(\lambda \left(\sum_{\mathcal{H}_T(o,p)}[o,p]-\sum_{\mathcal{H}_T(o,w^p_m)}[o,w^p_m]\right)\right)\right] \leq 1-\epsilon_1 $$ \end{proposition} We will denote by $\mathcal{H}(o,p)^C$ to be the complement of $\mathcal{H}(o,p)$ in $\mathcal{H}(o,w^p_m)$, i.e. $\mathcal{H}(o,w^p_m)=\mathcal{H}(o,p) \cup \mathcal{H}(o,p)^C$. \\ We begin with a lemma. \begin{lemma} \label{lem:small_complement} There exist $\eta >0$ and a function $f:\mathbb{N}^{+} \to \mathbb{R}$, with $f(m) \to 0 $ as $m \to + \infty$, such that $$ \mathbb{P} \Big[ \sum_{\mathcal{H}(o,p)^C} [o,w^p_m] \leq \eta \log(m) \Big] \leq f(m) $$ \end{lemma} \begin{proof} Let $\eta, U> 0$ be as in Lemma \ref{etalogproj} and fix $m \geq e^{6T/\eta}$; the "$\eta$" satisfying the conclusion of the lemma will actually be $\eta/2$. Let $\mathcal A$ be the set of all $q\in G$ such that there exists $h_q\gamma\notin \mathcal H(o,p)$ with $d_{h_q\gamma}(o,q)\geq \eta \log(m)$; for each such $q$ we fix a corresponding $h_q$. For $k \in \mathbb{N}, q \in G$, we consider the events $A_{k,q}= ``q \in \mathcal A, w^p_k=q, w^p_i\notin \mathcal A \quad\forall i<k $". Note that if both events $A_{k,q}$ and ``$d_{h_q\gamma}(w^p_k,w^p_m)< \eta \log(m)/2$" hold then $\sum_{\mathcal{H}(o,p)^C} [o,w^p_m] > \eta \log(m)/2$. Indeed, the sum includes the term $d_{h_q\gamma}(o,w^p_m)\geq \eta \log(m)/2$. \par\medskip \textbf{Claim 1}: There exists a function $g$ with with $g(m) \to 0 $ as $m \to + \infty$ such that for all $m$ we have $$\mathbb P[\exists k\leq m : w^p_k\in \mathcal A]\geq 1- g(m).$$ \begin{proof} If we did not have the requirement $h_q\gamma\notin \mathcal H(o,p)$ then this would follow directly from Lemma \ref{etalogproj}. However, it suffices to combine said lemma with the fact that Proposition \ref{axiom} (together with the fact that projections are coarsely Lipschitz) implies that $$\mathbb P\Big[\exists k\leq m , h\gamma\in \mathcal H(o,p): d_{h\gamma}( w^p_k, o)\geq \eta \log(m)/2\Big]$$ is bounded by some function of $m$ that goes to $0$ as $m$ tends to infinity. \end{proof} Note that for each $k\leq m$ and $q\in \mathcal A$ we have $$\mathbb P[d_{h_q\gamma}(w^p_k,w^p_m)< \eta \log(m)/2| A_{k,q}]\ =\ \mathbb P[d_{h_q\gamma}(q,w^q_{m-k})< \eta \log(m)/2]\ \geq \ 1- Ce^{-\eta\log(m)/(2C)},$$ where the equality follows from the Markov property (Lemma \ref{lem:Markov_property_set}), while the inequality (and corresponding constant) comes from Proposition \ref{axiom}. The last term can be rewritten as $1-h(m)$ for some function going to $0$ as $m\to +\infty$. Since the events $A_{k,q}$ are disjoint we can now estimate \begin{align*} \mathbb{P} \Big[ \sum_{\mathcal{H}(o,p)^C} [o,w^p_m] > &\eta \log(m)/2 \Big] \\ &\geq \sum_{k\leq m, q\in \mathcal A} \mathbb P[d_{h_q\gamma}(w^p_k,w^p_m) < \eta \log(m)/2\ |\ A_{k,q}]\ \mathbb P[A_{k,q}]\\ & \geq (1-h(m)) \sum_{k\leq m, q\in \mathcal A}\mathbb P[A_{k,q}]\\ & \geq (1-h(m))(1-g(m)), \end{align*} where we used both the Claim and the estimate on conditional probabilities explained above. This proves the lemma. \end{proof} \begin{proof}[Proof of Proposition \ref{prop:exp_lambda}] Using the Cauchy-Schwarz inequality and Remark \ref{rem:triangle_inequality} (the triangular inequality), we get \begin{align*} \begin{split} \mathbb{E}\left[\exp\Big(\lambda (\sum_{\mathcal{H}(o,p)}[o,p]-\sum_{\mathcal{H}(o,w^p_m)}[o,w^p_m])\Big)\right] &=\mathbb{E}\left[\exp\left(\lambda \left(\sum_{\mathcal{H}(o,p)}[o,w]-\sum_{\mathcal{H}(o,p)}[o,w^p_m]\right)\right)\exp\left(-\lambda \sum_{\mathcal{H}(o,p)^C} [o,w^p_m]\right)\right] \\ &\leq \mathbb{E}\left[\exp\left(2\lambda \left(\sum_{\mathcal{H}(o,p)}[o,p]-\sum_{\mathcal{H}(o,p)}[o,w^p_m]\right)\right)\right]^{\frac{1}{2}}\mathbb{E}\left[\exp\left(-2\lambda \sum_{\mathcal{H}(o,p)^C}[o,w^p_m]\right)\right]^{\frac{1}{2}} \\ & \leq \mathbb E \left[\exp \left(2\lambda \sum_{\mathcal{H}(o,p)}[p,w^p_m] \right) \right]^{\frac{1}{2}}\mathbb{E}\left[\exp\left(-2\lambda \sum_{\mathcal{H}(o,p)^C}[o,w^p_m]\right)\right]^{\frac{1}{2}}. \end{split} \end{align*} Now, by Proposition \ref{axiom} there exists a constant $C>0$ such that for all $t>0$: $$ \mathbb{P}\left[ \sum_{\mathcal{H}(o,p)}[p,w^p_m] \geq t\right]\leq Ce^{-t/C}. $$ which leads to $$ \mathbb{P}\left[\exp \left(2\lambda \sum_{\mathcal{H}(o,p)}[p,w^p_m] \right)\geq s\right]=\mathbb{P}\left[ \sum_{\mathcal{H}(o,p)}[p,w^p_m] \geq \frac{\log(s)}{2\lambda}\right] \leq Cs^{-1/(2\lambda C)} $$ We recall that for a non-negative random variable $Y$, the expected value is given by: $$\mathbb{E}(Y) = \int_{0}^{+\infty} \mathbb{P}(Y \geq s) ds$$ If we choose $\lambda < \frac{1}{2C}$ this leads to \begin{equation*} \begin{split} \mathbb E \left[\exp \left(2\lambda \sum_{\mathcal{H}(o,p)}[p,w^p_m] \right) \right] &= \int_{0}^{+\infty} \mathbb{P}\left[\exp (2\lambda \sum_{\mathcal{H}(o,p)}[p,w^p_m] )\geq s\right]ds \\ &\leq \int_{0}^{1} \mathbb{P}\left[\exp (2\lambda \sum_{\mathcal{H}(o,p)}[p,w^p_m] )\geq s\right] ds + C\int_{1}^{+\infty} s^{-1/(2\lambda C)} ds \\ &\leq 1+\frac{2C^2\lambda}{1-2\lambda C} \\ &=\frac{1-2\lambda C (1-C)}{1-2\lambda C} \\ \end{split} \end{equation*} Now, letting $B$ be the event $``\sum_{\mathcal{H}(o,p)^{C}}[o,w^p_m] \leq \eta \log(m)" $ where $\eta $ is from Lemma \ref{lem:small_complement}, we get: \begin{align*} \begin{split} \mathbb{E}\left[\exp(-2\lambda (\sum_{\mathcal{H}(o,p)^C}[o,w^p_m])\right] &= \mathbb{E}\left[\exp(-2\lambda (\sum_{\mathcal{H}(o,p)^C}[o,w^p_m]))\mathbb{1}_{B}\right] + \mathbb{E}\left[\exp(-2\lambda (\sum_{\mathcal{H}(o,p)^C}[o,w^p_m]))\mathbb{1}_{B^{C}}\right] \\ & \leq \mathbb{P}[B] + m^{-2\lambda \eta} \\ & \leq f(m)+m^{-2\lambda \eta} \\ \end{split} \end{align*} where we go from the second to the third by Lemma \ref{lem:small_complement}, with $f(m) \to 0$ as $m \to +\infty$. This all leads to $$ \mathbb{E}\left[\exp\left(\lambda (\sum_{\mathcal{H}(o,p)}[o,w]-\sum_{\mathcal{H}(o,w^p_m)}[o,w^p_m])\right)\right] \leq \Big(\frac{1-2\lambda(1-C)}{1-2\lambda C}\Big)^{1/2}(f(m)+m^{-2\lambda \eta})^{1/2} $$ Recall that we have fixed some $0<\lambda <\frac{1}{2 C}$. We can therefore increase $m$ to get $$ \mathbb{E}\left[\exp\left(\lambda (\sum_{\mathcal{H}(o,p)}[o,w]-\sum_{\mathcal{H}(o,w^p_m)}[o,w^p_m])\right)\right] \leq 1-\epsilon_1$$ for some $\epsilon_1>0$ as required, proving the Proposition. \end{proof} We now prove Equation \eqref{star}. Fix $\epsilon_1, \lambda$ and $m$ from Proposition \ref{expectation}. For any positive integer $j$, we have: $$ \sum_{\mathcal{H}(o,w_{m+jm})}[o,w_{m+jm}]=\sum_{\mathcal{H}(o,w_{m+jm})}[o,w_{jm+m}]+\sum_{\mathcal{H}(o,w_{jm})}[o,w_{jm}]-\sum_{\mathcal{H}(o,w_{jm})}[o,w_{jm}] $$ and, by Proposition \ref{prop:exp_lambda}, for all $g\in G$ we have $$ \mathbb{E}\left[\exp(-\lambda\big( \sum_{\mathcal{H}(o,w_{m+jm})}[o,w_{jm+m}] -\sum_{\mathcal{H}(o,w_{jm})}[o,w_{jm}] ) \hspace{2mm} \Big\vert \hspace{2mm} w_{jm}=g\right] \leq 1-\epsilon_1. $$ Summing over all $g$, this leads to \begin{equation*} \begin{split} \mathbb{E}&\Big[\exp(-\lambda \sum_{\mathcal{H}(o,w_{m+jm})}[o,w_{jm+m}])\Big] \\ &\leq \sum_{g\in G}\mathbb{E}\Big[e^{-\lambda( \sum_{\mathcal{H}(o,w_{m+jm})}[o,w_{jm+m}] -\sum_{\mathcal{H}(o,w_{jm})}[o,w_{jm}])} \ e^{-\lambda \sum_{\mathcal{H}(o,g)}[o,g]}\hspace{2mm} \Big\vert \hspace{2mm} w_{jm}=g \Big]\mathbb{P}[w_{jm}=g] \\ &\leq (1-\epsilon_1)\sum_{g\in G} e^{-\lambda \sum_{\mathcal{H}(o,g)}[o,g]}\mathbb{P}[w_{jm}=g] \\ &\leq (1-\epsilon_1)\mathbb{E}\Big[\exp(-\lambda(\sum_{\mathcal{H}(o,w_{jm})}[o,w_{jm}])\Big]. \end{split} \end{equation*} Inductively, we get for all $j$ :$$ \mathbb{E}\left[\exp(-\lambda(\sum_{\mathcal{H}(o,w_{jm})}[o,w_{jm}]))\right] \leq (1-\epsilon_1)^{j} $$ Hence, by Markov's inequality, for any $L>0$ we have \begin{equation*} \begin{split} \mathbb{P}\left[ \sum_{\mathcal{H}(o,w_{jm})}[o,w_{jm}] < L jm\right] & = \mathbb{P}\Big[\exp(-\lambda(\sum_{\mathcal{H}(o,w_{jm})}[o,w_{jm}])) > \exp(-\lambda L jm)\Big] \\ & \leq \frac{\mathbb{E}\left[\exp(-\lambda(\sum_{\mathcal{H}(o,w_{jm})}[o,w_{jm}]))\right]}{e^{-\lambda L jm}}\\ &\leq (1-\epsilon)^{j}e^{\lambda L jm} \\ \end{split} \end{equation*} if we choose $L$ small enough, we can find a $C_0 \geq 1$ such that $$ \mathbb{P}\Big[ \sum_{\mathcal{H}(o,w_{jm})}[o,w_{jm}] < jm/C_0\Big] \leq C_0e^{-jm/C_0} $$ as required. By the discussion at the start of this section, this is enough to prove Theorem \ref{linear progress}. \qed \section{Applications of linear progress} \label{applications} In Theorem \ref{linear progress} we have established that, under suitable conditions and with probability tending to 1 exponentially fast, tame Markov chains are at a `linear' distance from the starting point in a hyperbolic space being acted on. We now encapsulate this notion of `linear progress' in a definition and then state some of the applications one can derive from a tame Markov chain making linear progress. These will be similar to the applications given in \cite{maher2015random} and \cite{MathieuSisto}. Specifically, we will study deviations from quasi-geodesics in Subsection \ref{subsec:deviation_qgeod}, a central limit theorem in Subsection \ref{subsec:clt} and the translation lengths of Markov elements in Subsection \ref{subsec:transl_length}. In all cases, and in contrast with the proof of linear progress, we can follow known arguments for random walks with small modifications. \begin{definition} Consider a group $G$ acting on a metric space $X$, with a basepoint $x_0$. We say that a Markov chain $(w_n^{o})_n$ on $G$ makes \textit{linear progress in $X$ with exponential decay} if there exists a constant $ C>0$ such that for all basepoints $o \in G$ we have $$ \mathbb{P}\Big[ d_{X}(ox_0,w^{o}_nx_0) \geq n/C \Big] \geq 1-Ce^{-n/C} $$ \end{definition} \subsection{Deviations from quasi-geodesics for tame Markov chains} \label{subsec:deviation_qgeod} We start by showing that, in a suitable sense, tame Markov chains tend not to deviate too far from quasi-geodesics. The arguments are almost identical to \cite[Section 10]{MathieuSisto}, in particular we show a similar result to Theorem 10.7 in the case of tame Markov chains. \\ We will state the deviation inequality in an \textit{acylindrically intermediate space} for $(G,X)$. We omit this definition here but refer the reader to \cite[Definition 10.1]{MathieuSisto}. Instead, we just note that the following spaces $Y$ are acylindrically intermediate for the specified group $G$ and metric space $X$ (see \cite[Proposition 10.3]{MathieuSisto} for these and more examples). \begin{example} \begin{itemize} \item If $G$ is a finitely generated group acting acylindrically on the geodesic hyperbolic space $X$, then $Y=X$ and $Y= \Cay(G,S)$, for a finite generating set $S$, are both acylindrically intermediate spaces for $(G,X)$. \\ \item If $G$ is relatively hyperbolic, $X$ is its coned-off Cayley graph and $Y$ is its Bowditch space, then $Y$ is acylindrically intermediate for $(G,X)$. \end{itemize} \end{example} \begin{theorem} \label{thmdeviationquasigeo} Let $G$ be a finitely generated group acting non-elementarily on a hyperbolic space $X$ and let $Y$ be acylindrically intermediate for $(G,X)$ with basepoint $y_0$. Consider a tame Markov chain $(w^{o}_n)_n$ on $G$ making linear progress with exponential decay in $X$. Then for every $D$, there exists a constant $D'$ such that for all $l$, $k <n$, and $p\in G$ we have $$ \mathbb{P}\Big[ \sup_{\alpha \in QG_{D}(p,w^p_n)} d_{Y}(w^p_k y_0, \alpha) \geq l \Big] \leq D'e^{-l/D'} $$ where $QG_{D}(a,b)$ is the set of all $(D,D)$- quasi-geodesics from $ay_0$ to $by_0$ (with respect to $d_{Y}$). \end{theorem} \begin{proof} We drop the basepoints from the notation for Markov chains as well as from expressions such as $d_X(ax_0,bx_0)$, as they will not be very important in this proof. We fix a bound $K$ on the size of the jumps of the Markov chain in $Y$ (see Remark \ref{rmk:bounded_jumps}), that is $d_{Y}(w_{n},w_{n+1})\leq K$ for all $n$ (and all basepoints). Similarly to \cite[Section 10]{MathieuSisto}, we say that a discrete path $(w_i)_{i\leq k}$ in $G$ is \textit{tight around} $w_k$ \textit{at scale $l$} if for any $k_1 \leq k \leq k_2$ with $k_2-k \geq l$ we have: \begin{enumerate} \item $d_{X}(w_{k_1}, w_{k_2}) \geq \frac{k_2-k_1}{C} $, \item $d_{Y}(w_{k'},w_{k'+1})\leq K$ for every $k'$. \end{enumerate} where the constant $C>0$ is from the definition of linear progress with exponential decay. We note that in \cite[Section 10]{MathieuSisto}, the definition of tightness has three conditions, the first one being the same as above, while the other two are implied by our second condition. That is, our definition is more restrictive, but it is simpler and sufficient for our purposes. \begin{lemma}(cfr. \cite[Lemma 10.11]{MathieuSisto}) \label{probatightscale} There exists a constant $F>0$ so that for all $n$ and $k \leq n $ and all $l \geq 1$ : $$ \mathbb{P}\Big[ \hspace{2mm} (w_i)_{i \leq n} \text{ is tight around } w_k \text{\ at scale l} \hspace{2mm} \Big] \geq 1-Fe^{-l/ F}. $$ \end{lemma} \begin{proof} Note that the fact that our Markov chain has bounded jumps yields point 2 of the definition, with probability 1. We now estimate the probability that point 1 above does not hold. We claim that for all $k_1\leq k_2$ we have $\mathbb{P}\Big[d_{X}(w_{k_1},w_{k_2}) < (k_2-k_1) / C \Big]\leq Ce^{-(k_2-k_1)/C}$, where $C$ is the constant of linear progress with exponential decay. To see this, we need Lemma \ref{lem:Markov_property_set}, which allows us to perform a "change of basepoint" computation: $$\mathbb{P}\Big[d_{X}(w_{k_1},w_{k_2}) < (k_2-k_1) / C \Big]$$ $$\sum_{g\in G}\mathbb{P}\Big[d_{X}(w_{k_1},w_{k_2}) < (k_2-k_1) / C \ | \ w_{k_1}=g\Big]\ \mathbb P[w_{k_1}=g]=$$ $$\sum_{g\in G}\mathbb{P}\Big[d_{X}(g,w^g_{k_2-k_1}) < (k_2-k_1) / C\Big] \mathbb P[w_{k_1}=g],$$ after which we can apply linear progress. For fixed $k, k_1$ we have $$\mathbb{P}\Big[ \exists k_2 \geq k : k_2-k \geq l, d_{X}(w_{k_1},w_{k_2}) < (k_2-k_1) / C \Big] \leq \sum_{k_2-k_1=j \geq \max\{l,k-k_1\}} e^{-j/C}\leq F_1 e^{-l/C_1}. $$ for some constant $F_1>0$; this is where we use linear progress with exponential decay, in the form explained above. We also note that we are implicitly using the Markov property from lemma \ref{lem:Markov_property_set} and the fact that $k_1$ has been fixed. We now sum over all possible $k_1$: $$\mathbb{P}\big[ \exists k_1\leq k\leq k_2 : k_2-k \geq l, d_{X}(w_{k_1},w_{k_2}) < (k_2-k_1) / C \big] $$ $$\leq \sum_{k_1\leq k, k-k_1\leq l}F_1 e^{-l/C_1} + \sum_{ k-k_1=j> l} F_1 e^{-l/C}$$ $$\leq F_1le^{-l/C_1}+F_2e^{-l/F_2},$$ for some constant $F_2>0$, and we are done since this decays exponentially fast in $l$. \end{proof} In view of Lemma \ref{probatightscale}, the theorem follows from \cite[Lemma 10.12]{MathieuSisto}, which says that if $(w_i)_{0 \leq i \leq n}$ is tight around $w_k$ at scale $l$ then there exists a constant $F_3>0$ such that $d_{Y}(w_k, \alpha(w_0, w_n)) \leq F_3 l$ (note that the lemma holds for the notion of tightness as in \cite{MathieuSisto}, hence, a fortiori, for the notion of tightness we use here). \end{proof} Now that we know that if a tame Markov chain on $G$ makes linear progress in the hyperbolic space $X$ on which it acts, we get ``deviation from quasi-geodesics" results in an acylindrically intermediate space, we can get the following corollary, in the case when $G$ acts acylindrically on the hyperbolic space $X$. We note that if $G$ acts acylindrically on $X$ and any tame Markov chain makes linear progress with exponential decay in $X$ then the action of $G$ on $X$ is non-elementary in view of the classification of acylindrical actions on hyperbolic spaces \cite[Theorem 1.1]{Osin_acylindrical}. From here on, given two elements $a,b$ in a group with a fixed word metric, we denote by $[a,b]$ any choice of geodesic in the corresponding Cayley graph connecting them. \begin{corollary} \label{devgeod} Let $G$ be a finitely generated group acting acylindrically (and non-elementarily) on a hyperbolic space $X$, such that any tame Markov chain in $G$ makes linear progress with exponential decay in $X$. Let $H$ be a group quasi-isometric to $G$, with a fixed word metric $d_H$. Consider a random walk $(Z_n)_n$ on $H$ with driving measure whose support is finite and generates $H$ as a semigroup. Then there exists a constant $R$ such that for all $k \leq n$ and $l>0$: $$ \mathbb{P}\big[ d_{H}(Z_k, [1,Z_n]) \geq l \big] \leq R e^{-l / R}. $$ \end{corollary} \begin{proof} Since $H$ and $G$ are quasi-isometric, and $G$ is non-amenable then by \cite[Theorem 2]{whyte}] there exists a bijective $R'$-quasi-isometry $f:H\to G$. For $g,h\in G$ we denote by $\alpha(g,h)$ the quasi-geodesic $f([f^{-1}(g),f^{-1}(h)])$ in $G$ between $g$ and $h$. Let $(w^o_n)$ be the push-forward of $(Z_n)$ to $G$ (see Definition \ref{defn:push_forward}), which is a tame Markov chain by Lemma \ref{pushforwardistame}. We let the basepoint of the Markov chain be $o=f(1)$. Note that $$\mathbb{P}\big[ d_{H}(Z_k, [1,Z_n]) \geq l \big] \leq \mathbb{P}\big[ d_{G}(w^o_k, \alpha(o,w^o_k)) \geq l/R'-R' \big].$$ Theorem \ref{thmdeviationquasigeo} then gives the required bound. \end{proof} \subsection{Central limit theorem} \label{subsec:clt} We recall that, given points $x,y,z$ in a metric space $Z$, the Gromov product $(x,y)_z$ is defined as $\left(d(x,z)+d(z,y)-d(x,y)\right)/2$. If $z$ lies within $d$ of some geodesic between $x$ and $y$, then $(x,y)_z\leq d$. All Gromov products \begin{lemma}\label{lem:second_moment_dev} Let $H$ be a group with a given word metric $d_H$ and consider a random walk $(Z_n)$ on $H$ with driving measure whose support is finite and generates $H$ as a semigroup. Then there exists a constant $R$ such that the following holds. For all $k \leq n$ and $l>0$:$$ \mathbb{P}\Big[ d_{H}(Z_k, [1,Z_n]) \geq l \Big] \leq R e^{-l / R}. $$ Then there exists a constant $R_2>0$, depending only on $R$, such that $$ \mathbb{E}\big[ (1,Z_n)_{Z_k}^2 \big] \leq R_2, $$ where the Gromov product is measured in the metric $d_H$. \end{lemma} \begin{proof} We have \begin{align*} \begin{split} \mathbb{E}\big[ ((1,Z_n)_{Z_k})^2 \big] & = \int_{0}^{+\infty} \mathbb{P}\big[ ((1,Z_n)_{Z_k})^2 \geq t \big] \quad dt \\ & \leq \int_{0}^{+\infty} \mathbb{P}\big[ d_H(Z_k, [1,Z_n]) \geq \sqrt{t} \big] \quad dt\\ &\leq R \int_{0}^{+\infty} e^{-\sqrt{t}/R} dt \\ & \leq R_2 \end{split} \end{align*} as the integral converges and is equal to a constant only depending on $R$. \end{proof} The lemma above allows us to apply the machinery of \cite{MathieuSisto} to obtain the following central limit theorem. \begin{theorem} Let $G$ be a finitely generated group acting acylindrically (and non elementarily) on a hyperbolic space $X$, where any tame Markov chain makes linear progress in $X$ with exponential decay. Let $H$ be a group quasi-isometric to $G$ and let $(Z_n)_n$ be a random walk on $H$ with driving measure whose support is finite and generates $H$ as a semigroup. Then $(Z_n)_n$ satisfies the Central Limit Theorem. i.e. there exist constants $l>0$ and $\sigma>0$ such that $$ \frac{d_H(1,Z_n)-ln}{\sigma^2\sqrt{n}} \to \mathcal{N}(0,1), $$ where the convergence is a convergence in law. \end{theorem} \begin{proof} By Corollary \ref{devgeod} and Lemma \ref{lem:second_moment_dev}, there exists a constant $R_2>0$ such that $$ \mathbb{E}\big[ (1,Z_n)_{Z_k}^2 \big] \leq R_2$$ (which, in the language of \cite{MathieuSisto}, says that the random walk under consideration satisfies the second moment deviation inequality). \\ Hence, by \cite[Theorem 4.2]{MathieuSisto}, this yields that the random walk $(Z_n)_n$ satisfies the Central Limit Theorem. \end{proof} \subsection{Translation lengths of Markov elements} \label{subsec:transl_length} We recall that, given a group $G$ acting on a metric space $X$, the \textit{translation length} of an element $g\in G$ is defined as$$\tau(g) = \liminf_{n \to + \infty} \frac{d_X(x_0,g^nx_0)}{n} $$ which is well defined and independent of the point $x_0 \in X$. We recall that an element $g$ is \emph{loxodromic} if and only if $\tau(g) >0$. \\ We will establish the following result for a Markov chain, similar to a result of Maher-Tiozzo for random walks \cite[Theorem 1.4]{maher2015random}. In particular, the result says that loxodromic elements are generic with respect to tame Markov chains. \begin{theorem} \label{translationlength} Let $G$ be a group acting acylindrically on a hyperbolic space $X$. Consider a tame Markov chain $(w^{o}_n)$ on $G$, making linear progress in $X$ with exponential decay. Then there exist constants $L_1, C_2> 0$ such that $$ \mathbb{P}\Big[ \tau(w_n^{o}) > L_1 n \Big] \geq 1- C_{2}e^{-n/C_{2}}$$ \end{theorem} We will need the following lemma. \begin{lemma} \label{small_gromov_2} There exist constants $M,C>0$ such that the following holds. For all $h \in G$ and $k\geq 1$: $$ \mathbb P\big[ (x_0,(w^h_k)^2x_0)_{w^h_kx_0} \geq Mk\big] \leq Ce^{-k/C} $$ \end{lemma} In turn, in order to prove Lemma \ref{small_gromov_2} we will use the following (deterministic) lemma from \cite{MathieuSisto}. The notation $B^G$ and $\diam^X$ indicate balls in $G$ and diameter in the metric of $X$, respectively, for emphasis. We note that the lemma requires that the action of $G$ on the hyperbolic space $X$ is acylindrical. \begin{lemma}\cite[Lemma 9.5]{MathieuSisto} \label{diam} There exist constants $C', D>0$ with the following property. For each $u,v \in G$ and $k \geq 0$, the set $B(u,v,k)$ of elements $s \in B^{G}(id,k)$ with $\diam^{X} \Big( [x_0,ux_0] \cap \mathcal N_{2 \delta}([usx_0,usvx_0])\Big) \geq D$ has cardinality at most $C'k^2$. \end{lemma} \begin{figure}[h] \centering \begin{tikzpicture}[scale=0.8] \filldraw[black] (0,0) circle (1pt) node[anchor=east]{$x_0$}; \filldraw[black] (1.5,7) circle (1pt) node[anchor=east]{$hx_0$}; \filldraw[black] (3,7) circle (1pt) node[anchor=south]{$w^h_kx_0$}; \filldraw[black] (8,3) circle (1pt) node[anchor=south]{$w^h_khx_0$}; \filldraw[black] (9,1.5) circle (1pt) node[anchor=east]{$(w^h_k)^2x_0$}; \draw[black] (0,0) to[out=60,in=300] (1.5,7); \draw[black] (0,0) to[out=40,in=280] node[pos=0.70] (mid1) {} node[pos=0.90] (mid2) {} (3,7); \draw[black] (3,7) to[out=290,in=170] (8,3); \path [draw=blue,snake it] (1.5,7) -- (3,7); \path [draw=blue,snake it] (8,3) -- (9,1.5); \draw[blue] (9.3,2.5) node {$\leq kK $}; \draw[blue] (2,8) node {$\leq kK $}; \end{tikzpicture} \caption{Points and geodesics relevant to the proof of Lemma \ref{small_gromov_2}.} \end{figure} \begin{proof}[Proof of Lemma \ref{small_gromov_2}] We let $K$ be a bound on the size of the jumps of the Markov chain in $X$ (see Remark \ref{rmk:bounded_jumps}), that is, $d_X(w^p_n,w^p_{n+1})\leq K$ for all $p$ and $n$. First we note that if $(x_0,(w^h_k)^2x_0)_{w^h_kx_0} \geq Mk$ then $(x_0,w^h_k hx_0)_{w^h_kx_0}\geq Mk-Kk$, since $d_X(w^h_k hx_0,(w^h_k)^2x_0)\leq Kk$. In this case, two geodesics $[w^h_kx_0, x_0]$ and $[w^h_kx_0,w^h_k hx_0 ]$ will have initial subgeodesics of length $Mk-Kk-E$ that stay within $\delta$ of each other, where $\delta$ is a hyperbolicity constant for $X$ and $E$ only depends on $\delta$. Since $d_X(hx_0, w^h_kx_0)\leq Kk$, we further have that geodesics $[hx_0,x_0]$ and $[w^h_kx_0,w^h_k hx_0 ]$ have subgeodesics of length $Mk-2Kk-2E$ that stay within $2\delta$ of each other. Now, if $M$ was chosen large enough so that $Mk-2Kk-2E\geq D$, then by Lemma \ref{diam} we have that $h^{-1}w^h_k\in B(h,h, Kk)$ and $\# B(h,h,Kk) \leq C'(Kk)^2$. Hence, the non-amenability assumption from tameness of the Markov chain (\ref{defn:tame}-\eqref{item:non-amen}) gives the following, where $\rho <1$: \begin{align*} \begin{split} \mathbb P\Big[ (x_0,(w^h_k)^2x_0)_{w^h_kx_0} \geq Mk\Big] &\leq \# B(h,h,Kk) A \rho^k\\ & \leq C'(Kk)^2A\rho^k \\ & \leq Ce^{-k/C} \end{split} \end{align*} for some constant $C >0$, proving the lemma. \end{proof} \begin{lemma} \label{smallgromov} For all $\epsilon >0$, there exists a constant $C_{\epsilon}$ such that the following holds. For all $o \in G$ and $n$: $$ \mathbb{P}\big[ (x_0,(w^{o}_n)^2 x_0)_{w^{o}_nx_0} \geq \epsilon n \big] \leq C_{\epsilon}e^{-n/ C_{\epsilon}} $$ \end{lemma} \begin{proof} Fix $\epsilon >0$ and $n$. Let $k = \lfloor \frac{\epsilon n}{M}\rfloor$ where $M $ is from Lemma \ref{small_gromov_2}. We have \begin{align*} \begin{split} \mathbb{P}\Big[ (x_0,(w^{o}_n)^2 x_0)_{w^{o}_nx_0} \geq \epsilon n \Big] &= \sum_{h \in G}\mathbb P \Big[ (x_0, (w^o_n)^2x_0)_{w^o_n x_0} \geq \epsilon n \quad \Big\vert \quad w^{o}_{n-k}=h\Big]\mathbb P \Big[ w^{o}_{n-k}=h\Big] \\ & \leq \sum_{h \in G}\mathbb P \Big[ (x_0, (w^h_k)^2x_0)_{w^h_k x_0} \geq \epsilon n \Big] \mathbb P \Big[ w^{o}_{n-k}=h\Big] \\ & \leq \sum_{h \in G} \mathbb P \Big[ (x_0, (w^h_k)^2x_0)_{w^h_k x_0} \geq Mk \Big] \mathbb P \Big[ w^{o}_{n-k}=h\Big]\\ & \leq Ce^{-k/C} \\ & \leq C''e^{-n/C''} \end{split} \end{align*} where we use the Markov property from lemma \ref{lem:Markov_property_set} to go from the first to the second line. We then use Lemma \ref{small_gromov_2} and the fact that $ \frac{\epsilon n}{M} \geq k \geq \frac{\epsilon n}{M}-1$. \end{proof} We need another result, from \cite{maher2015random}. \begin{lemma}\cite[Proposition 5.8]{maher2015random} \label{lowerbound_translationlength} There exists a constant $C_{1} > 0$, depending only on $\delta $ such that the following holds. For any isometry $g$ of the $\delta$-hyperbolic space $X$ the translation length of $g$ satisfies $$ \tau(g) \geq d_X(x_0,gx_0)-2(gx_0,g^{-1}x_0)_{x_0} - S \delta$$ for some constant $S$. \end{lemma} We now have everything in order to prove that with high probability, the translation length of a Markov element is at least linear in $n$, and in particular that loxodromic elements are generic. \begin{proof}[Proof of Theorem \ref{translationlength}] Let $L_1 < C$, where $C$ is the constant of linear progress with exponential decay. Fix the constant $S$ from Lemma \ref{lowerbound_translationlength}. Also, let $C_\epsilon$ be the constant from Lemma \ref{smallgromov} for $\epsilon = (C-L_1)/3>0$. The for all $n$ large enough (it suffices to consider this case) we have \begin{align*} \begin{split} \mathbb{P}\Big[ \tau(w_n) \leq L_1 n \Big] &\leq \mathbb{P}\big[ d_X(x_0,w_nx_0)-2(w_nx_0,w_n^{-1}x_0)_{x_0} -S\delta \leq L_1 n \big] \\ & \leq \mathbb{P}\big[ d_X(x_0,w_nx_0) <Ln \big]+ \mathbb{P}\big[ ((w^{o}_n)^2x_0,x_0)_{w_nx_0} \geq (L - L_1) n/2-S\delta /2\big] \\ &\leq Ce^{-n/C}+C_{\epsilon} e^{-n/C_{\epsilon}}, \\ \end{split} \end{align*} where we uses linear progress and Lemma \ref{smallgromov} in the last step. The latter term decays exponentially in $n$ so we are done. \end{proof}

https://arxiv.org/abs/1907.11905

Coloring rings

A ring is a graph $R$ whose vertex set can be partitioned into $k \geq 4$ nonempty sets, $X_1, \dots, X_k$, such that for all $i \in \{1,\dots,k\}$, the set $X_i$ can be ordered as $X_i = \{u_i^1, \dots, u_i^{|X_i|}\}$ so that $X_i \subseteq N_R[u_i^{|X_i|}] \subseteq \dots \subseteq N_R[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$. A hyperhole is a ring $R$ such that for all $i \in \{1,\dots,k\}$, $X_i$ is complete to $X_{i-1}\cup X_{i+1}$. In this paper, we prove that the chromatic number of a ring $R$ is equal to the maximum chromatic number of a hyperhole in $R$. Using this result, we give a polynomial-time coloring algorithm for rings.Rings formed one of the basic classes in a decomposition theorem for a class of graphs studied by Boncompagni, Penev, and Vušković in [Journal of Graph Theory 91 (2019), 192--246]. Using our coloring algorithm for rings, we show that graphs in this larger class can also be colored in polynomial time. Furthermore, we find the optimal $\chi$-bounding function for this larger class of graphs, and we also verify Hadwiger's conjecture for it.

\section{Introduction} All graphs in this paper are finite, simple, and nonnull. As usual, the vertex set and edge set of a graph $G$ are denoted by $V(G)$ and $E(G)$, respectively; for a vertex $v$ of $G$, $N_G(v)$ is the set of neighbors of $v$ in $G$, and $N_G[v] = N_G(v) \cup \{v\}$. A {\em ring} is a graph $R$ whose vertex set can be partitioned into $k \geq 4$ nonempty sets $X_1, \dots, X_k$ (whenever convenient, we consider indices of the $X_i$'s to be modulo $k$), such that for all $i \in \{1,\dots,k\}$ the set $X_i$ can be ordered as $X_i = \{u_i^1, \dots, u_i^{|X_i|}\}$ so that \begin{displaymath} X_i \subseteq N_R[u_i^{|X_i|}] \subseteq \dots \subseteq N_R[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}. \end{displaymath} (Note that this implies that $X_1,\dots,X_k$ are all cliques\footnote{A {\em clique} is a set of pairwise adjacent vertices.} of $R$, and that $u_1^1,u_2^1,\dots,u_k^1,u_1^1$ is a hole\footnote{A {\em hole} is an induced cycle of length at least four.} of length $k$ in $R$.) Under such circumstances, we also say that the ring $R$ is of {\em length} $k$, or that $R$ is a {\em $k$-ring}; furthermore, $(X_1, \dots, X_k)$ is called a {\em ring partition} of $R$. A ring is {\em even} or {\em odd} depending on the parity of its length. Rings played an important role in~\cite{VIK}: they formed a ``basic class'' in the decomposition theorems for a couple of graph classes defined by excluding certain ``Truemper configurations'' as induced subgraphs (more on this in subsection~\ref{subsec:Truemper}). In that paper, the complexity of the optimal vertex coloring problem for rings was left as an open problem.\footnote{In fact, only odd rings are difficult in this regard; even rings are readily colorable in polynomial time (see Lemma~\ref{lemma-even-ring-col-alg}).} In the present paper, we give a polynomial-time coloring algorithm for rings (see Theorems~\ref{thm-ring-col-alg} and~\ref{thm-ring-chi-alg}). It can easily be shown that every ring is a circular-arc graph. Furthermore, rings have unbounded clique-width. To see this, let $k \geq 3$ be an integer, and let $R$ be a $(k+1)$-ring with ring partition $(X_1,\dots,X_k,X_{k+1})$ such that the cliques $X_i$ are all of size $k+1$, with vertices labeled $0,1,\dots,k$, and furthermore, assume that vertices labeled $p$ and $q$ from consecutive cliques of the ring partition are adjacent if and only if $p+q \leq k$. Now, the graph obtained from $R$ by first deleting $X_{k+1}$, and then deleting all the vertices labeled 0, is precisely the permutation graph $H_k$ defined in~\cite{gr}, and the clique-width of $H_k$ is at least $k$ (see Lemma~5.4 from~\cite{gr}). Given graphs $H$ and $G$, we say that $G$ {\em contains} $H$ if $G$ contains an induced subgraph isomorphic to $H$; if $G$ does not contain $H$, then $G$ is {\em $H$-free}. For a family $\mathcal{H}$ of graphs, we say that a graph $G$ is {\em $\mathcal{H}$-free} if $G$ is $H$-free for all $H \in \mathcal{H}$. Given a graph $G$, a {\em clique} of $G$ is a (possibly empty) set of pairwise adjacent vertices of $G$, and a {\em stable set} of $G$ is a (possibly empty) set of pairwise nonadjacent vertices of $G$. The {\em clique number} of $G$, denoted by $\omega(G)$, is the maximum size of a clique of $G$, and the {\em stability number} of $G$, denoted by $\alpha(G)$, is the maximum size of a stable set of $G$. A {\em proper coloring} of $G$ is an assignment of colors to the vertices of $G$ in such a way that no two adjacent vertices receive the same color. For a positive integer $r$, $G$ is said to be {\em $r$-colorable} if there is a proper coloring of $G$ that uses at most $r$ colors. The {\em chromatic number} of $G$, denoted by $\chi(G)$, is the minimum number of colors needed to properly color $G$. An {\em optimal coloring} of $G$ is a proper coloring of $G$ that uses only $\chi(G)$ colors. Given a graph $G$, a vertex $v \in V(G)$, and a set $S \subseteq V(G) \setminus \{v\}$, we say that $v$ is {\em complete} (resp.\ {\em anticomplete}) to $S$ in $G$ provided that $v$ is adjacent (resp.\ nonadjacent) to every vertex of $S$; given disjoint sets $X,Y \subseteq V(G)$, we say that $X$ is {\em complete} (resp.\ {\em anticomplete}) to $Y$ in $G$ provided that every vertex in $X$ is complete (resp.\ anticomplete) to $Y$ in $G$. A {\em hole} is a chordless cycle on at least four vertices; the {\em length} of a hole is the number of its vertices, and a hole is \emph{even} or \emph{odd} according to the parity of its length. When we say ``$H$ is a hole in $G$,'' we mean that $H$ is a hole that is an induced subgraph of $G$. A \emph{hyperhole} is any graph $H$ whose vertex set can be partitioned into $k \geq 4$ nonempty cliques $X_1, \dots, X_k$ (whenever convenient, we consider indices of the $X_i$'s to be modulo $k$) such that for all $i \in \{1,\dots,k\}$, $X_i$ is complete to $X_{i-1}\cup X_{i+1}$ and anticomplete to $V(H) \setminus (X_{i-1} \cup X_i \cup X_{i+1})$; under such circumstances, we also say that $H$ is a hyperhole of {\em length} $k$, or that $H$ is a {\em $k$-hyperhole}. A hyperhole is {\em even} or {\em odd} according to the parity of its length. Note that every hole is a hyperhole, and every hyperhole is a ring. When we say ``$H$ is a hyperhole in $G$,'' we mean that $H$ is a hyperhole that is an induced subgraph of $G$. Hyperholes can be colored in linear time~\cite{hyperhole}. Furthermore, the following lemma gives a formula for the chromatic number of a hyperhole. \begin{lemma} \cite{hyperhole} \label{lemma-hyperhole-chi-formula} Let $H$ be a hyperhole. Then $\chi(H) = \max\Big\{\omega(H),\Big\lceil \frac{|V(H)|}{\alpha(H)} \Big\rceil\Big\}$. \end{lemma} The main result of the present paper is the following theorem. \begin{theorem} \label{thm-ring-hyperhole} Let $k \geq 4$ be an integer, and let $R$ be a $k$-ring. Then $\chi(R) = \max\{\chi(H) \mid \text{$H$ is a $k$-hyperhole in $R$}\}$. \end{theorem} It was shown in~\cite{VIK} that all holes of a $k$-ring ($k \geq 4$) are of length $k$;\footnote{In the present paper, this result is stated as Lemma~\ref{lemma-ring-chordal}(b).} consequently, all hyperholes in a $k$-ring are of length $k$. Thus, Theorem~\ref{thm-ring-hyperhole} in fact establishes that the chromatic number of a ring is equal to the maximum chromatic number of a hyperhole in the ring. It is easy to see that the stability number of any $k$-hyperhole ($k \geq 4$) is $\lfloor k/2 \rfloor$. Thus, the following is an immediate corollary of Lemma~\ref{lemma-hyperhole-chi-formula} and Theorem~\ref{thm-ring-hyperhole}. \begin{corollary} \label{cor-ring-chi-formula} Let $k \geq 4$ be an integer, and let $R$ be a $k$-ring. Then $\chi(R) = \max\Big(\{\omega(R)\} \cup \Big\{\Big\lceil \frac{|V(H)|}{\lfloor k/2 \rfloor} \Big\rceil \mid \text{$H$ is a $k$-hyperhole in $R$}\Big\}\Big)$. \end{corollary} Using Theorem~\ref{thm-ring-hyperhole},\footnote{More precisely, we use Lemma~\ref{lemma-reduce-chi}, which is a corollary of Theorem~\ref{thm-ring-hyperhole} and Lemma~\ref{lemma-main-technical}. Lemma~\ref{lemma-main-technical}, in turn, is the main part of the proof of Theorem~\ref{thm-ring-hyperhole}.} we construct an $O(n^6)$ algorithm that computes an optimal coloring of an input ring (see Theorem~\ref{thm-ring-col-alg}). Furthermore, using Corollary~\ref{cor-ring-chi-formula}, we also give an $O(n^3)$ time algorithm that computes the chromatic number of a ring without actually finding an optimal coloring of that ring (see Theorem~\ref{thm-ring-chi-alg}). \subsection{Terminology, notation, and paper outline} \label{subsec:Truemper} For a function $f:A \rightarrow B$ and a set $A' \subseteq A$, we denote by $f \upharpoonright A'$ the restriction of $f$ to $A'$. The complement of a graph $G$ is denoted by $\overline{G}$. For a nonempty set $X \subseteq V(G)$, we denote by $G[X]$ the subgraph of $G$ induced by $X$; for vertices $x_1,\dots,x_t \in V(G)$, we often write $G[x_1,\dots,x_t]$ instead of $G[\{x_1,\dots,x_t\}]$. For a set $S \subsetneqq V(G)$, we denote by $G \setminus S$ the subgraph of $G$ obtained by deleting $S$, i.e.\ $G \setminus S = G[V(G) \setminus S]$; if $G$ has at least two vertices and $v \in V(G)$, then we often write $G \setminus v$ instead of $G \setminus \{v\}$.\footnote{Since our graphs are nonnull, if $G$ has just one vertex, say $v$, then $G \setminus v$ is undefined.} A class of graphs is {\em hereditary} if it is closed under isomorphism and induced subgraphs. More precisely, a class $\mathcal{G}$ of graphs is {\em hereditary} if for every graph $G \in \mathcal{G}$, the class $\mathcal{G}$ contains all isomorphic copies of induced subgraphs of $G$. A {\em theta} is any subdivision of the complete bipartite graph $K_{2,3}$; in particular, $K_{2,3}$ is a theta. A {\em pyramid} is any subdivision of the complete graph $K_4$ in which one triangle remains unsubdivided, and of the remaining three edges, at least two edges are subdivided at least once. A {\em prism} is any subdivision of $\overline{C_6}$ in which the two triangles remain unsubdivided; in particular, $\overline{C_6}$ is a prism. A {\em three-path-configuration} (or {\em 3PC} for short) is any theta, pyramid, or prism. A {\em wheel} is a graph that consists of a hole and an additional vertex that has at least three neighbors in the hole. If this additional vertex is adjacent to all vertices of the hole, then the wheel is said to be a {\em universal wheel}; if the additional vertex is adjacent to three consecutive vertices of the hole, and to no other vertex of the hole, then the wheel is said to be a {\em twin wheel}. A {\em proper wheel} is a wheel that is neither a universal wheel nor a twin wheel. A {\em Truemper configuration} is any 3PC or wheel (for a survey, see~\cite{Truemper-survey}). Note that every Truemper configuration contains a hole. Note, furthermore, that every prism or theta contains an even hole, and every pyramid contains an odd hole. Thus, even-hole-free graphs contain no prisms and no thetas, and odd-hole-free graphs contain no pyramids. $\mathcal{G}_{\text{T}}$ is the class of all (3PC, proper wheel, universal wheel)-free graphs; thus, the only Truemper configurations that a graph in $\mathcal{G}_{\text{T}}$ may contain are the twin wheels. Clearly, the class $\mathcal{G}_{\text{T}}$ is hereditary. A decomposition theorem for $\mathcal{G}_{\text{T}}$ (where rings form one of the ``basic classes'') was obtained in~\cite{VIK},\footnote{In the present paper, this decomposition theorem is stated as Theorem~\ref{thm-GT-decomp}.} as were polynomial-time algorithms that solve the recognition, maximum weight clique, and maximum weight stable set problems for the class $\mathcal{G}_{\text{T}}$. The complexity of the optimal coloring problem for $\mathcal{G}_{\text{T}}$ was left open in~\cite{VIK}, and the main obstacle in this context were rings. In the present paper, we show that graphs in $\mathcal{G}_{\text{T}}$ can be colored in polynomial time (see Theorems~\ref{thm-GT-col-alg} and~\ref{thm-GT-chi-alg}). A {\em simplicial vertex} is a vertex whose neighborhood is a (possibly empty) clique. For an integer $k \geq 4$, let $\mathcal{R}_k$ be the class of all graphs $G$ that have the property that every induced subgraph of $G$ either is a $k$-ring or has a simplicial vertex; clearly, $\mathcal{R}_k$ is hereditary, and furthermore (by Lemma~\ref{lemma-RRk-hered}) it contains all $k$-rings. We remark that graphs in $\mathcal{R}_k$ are precisely the chordal graphs,\footnote{A graph is {\em chordal} if it contains no holes.} and the graphs that can be obtained from a $k$-ring by (possibly) repeatedly adding simplicial vertices (see Lemma~\ref{lemma-RRk-description}). Further, for all integers $k \geq 4$, we set $\mathcal{R}_{\geq k} = \bigcup\limits_{i=k}^{\infty} \mathcal{R}_i$; clearly, $\mathcal{R}_{\geq k}$ is hereditary, and furthermore (by Lemma~\ref{lemma-RRk-hered}) it contains all rings of length at least $k$. In particular, the class $\mathcal{R}_{\geq 4}$ is hereditary and contains all rings. We show that graphs in $\mathcal{R}_{\geq 4}$ can be colored in polynomial time (see Theorems~\ref{thm-ring-col-alg} and~\ref{thm-ring-chi-alg}). A {\em clique-cutset} of a graph $G$ is a (possibly empty) clique $C \subsetneqq V(G)$ of $G$ such that $G \setminus C$ is disconnected. A {\em clique-cut-partition} of a graph $G$ is a partition $(A,B,C)$ of $V(G)$ such that $A$ and $B$ are nonempty and anticomplete to each other, and $C$ is a (possibly empty) clique. Clearly, a graph admits a clique-cutset if and only if it admits a clique-cut-partition. A graph is {\em perfect} if all its induced subgraphs $H$ satisfy $\chi(H) = \omega(H)$. The Strong Perfect Graph Theorem~\cite{SPGT} states that a graph $G$ is perfect if and only if neither $G$ nor $\overline{G}$ contains an odd hole. $\mathbb{N}$ is the set of all positive integers. A hereditary class $\mathcal{G}$ is {\em $\chi$-bounded} if there exists a function $f:\mathbb{N} \rightarrow \mathbb{N}$ (called a {\em $\chi$-bounding function} for $\mathcal{G}$) such that all graphs $G \in \mathcal{G}$ satisfy $\chi(G) \leq f(\omega(G))$. For a hereditary $\chi$-bounded class $\mathcal{G}$ that contains all complete graphs (equivalently: that contains graphs of arbitrarily large clique number), we say that a $\chi$-bounding function $f:\mathbb{N} \rightarrow \mathbb{N}$ for $\mathcal{G}$ is {\em optimal} if for all $n \in \mathbb{N}$, there exists a graph $G \in \mathcal{G}$ such that $\omega(G) = n$ and $\chi(G) = f(n)$. It was shown in~\cite{VIK} that $\mathcal{G}_{\text{T}}$ is $\chi$-bounded by a linear function; more precisely, it was shown that every graph $G \in \mathcal{G}_{\text{T}}$ satisfies $\chi(G) \leq \Big\lfloor \frac{3}{2}\omega(G) \Big\rfloor$.\footnote{See Theorem~7.6 from~\cite{VIK}.} In the present paper, we improve this $\chi$-bounding function, and in fact, we find the optimal $\chi$-bounding function for the class $\mathcal{G}_{\text{T}}$ (see Theorem~\ref{thm-GT-chi}). Finally, we consider Hadwiger's conjecture. Let $H$ be an $n$-vertex graph with vertex set $V(H) = \{v_1,\dots,v_n\}$. We say that a graph $G$ {\em contains $H$ as a minor} if there exist pairwise disjoint, nonempty subsets $S_1,\dots,S_n \subseteq V(G)$ (called {\em branch sets}) such that $G[S_1],\dots,G[S_n]$ are all connected, and such that for all distinct $i,j \in \{1,\dots,n\}$ with $v_iv_j \in E(H)$, there is at least one edge between $S_i$ and $S_j$ in $G$. As usual, the complete graph on $k$ vertices is denoted by $K_k$. Hadwiger's conjecture states that every graph $G$ contains $K_{\chi(G)}$ as a minor. Using Theorem~\ref{thm-ring-hyperhole}, we prove that rings satisfy Hadwiger's conjecture (see Lemma~\ref{lemma-Hadwiger-ring}), and as a corollary, we obtain that graphs in $\mathcal{G}_{\text{T}}$ also satisfy Hadwiger's conjecture (see Theorem~\ref{thm-Hadwiger-GT}). A {\em hyperantihole} is a graph $A$ whose vertex set can be partitioned into nonempty cliques $X_1,\dots,X_k$ ($k \geq 4$)\footnote{Whenever convenient, we consider indices of the $X_i$'s to be modulo $k$.} such that for all $i \in \{1,\dots,k\}$, $X_i$ is complete to $V(A) \setminus (X_{i-1} \cup X_i \cup X_{i+1})$ and anticomplete to $X_{i-1} \cup X_{i+1}$.\footnote{Note that the complement of a hyperantihole need not be a hyperhole.} Under these circumstances, we also say that the hyperantihole $A$ is of {\em length} $k$, and that $A$ is a {\em $k$-hyperantihole}. A hyperantihole is {\em odd} or {\em even} depending on the parity of its length. The remainder of this paper is organized as follows. In section~\ref{sec:prelim}, we state a few results from the literature that we need in the remainder of the paper; in section~\ref{sec:prelim}, we also prove a few easy lemmas about rings and their induced subgraphs, and about classes $\mathcal{R}_k$ and $\mathcal{R}_{\geq k}$ ($k \geq 4$). In section~\ref{sec:main}, we prove Theorem~\ref{thm-ring-hyperhole}, and we also give a polynomial-time coloring algorithm for even rings (see Lemma~\ref{lemma-even-ring-col-alg}). In section~\ref{sec:col}, we give an $O(n^6)$ time coloring algorithm for rings (see Theorem~\ref{thm-ring-col-alg}).\footnote{In fact, this is a coloring algorithm for graphs in $\mathcal{R}_{\geq 4}$. By Lemma~\ref{lemma-RRk-hered}, $\mathcal{R}_{\geq 4}$ contains all rings.} Even rings are easy to color (see Lemma~\ref{lemma-even-ring-col-alg}); our coloring algorithm for odd rings relies on ideas from the proof of Theorem~\ref{thm-ring-hyperhole}. Using our coloring algorithm for rings, as well as various results from the literature, we also construct an $O(n^7)$ time coloring algorithm for graphs in $\mathcal{G}_{\text{T}}$ (see Theorem~\ref{thm-GT-col-alg}). In section~\ref{sec:chi}, we construct an $O(n^3)$ time algorithm that computes the chromatic number of a ring (see Theorem~\ref{thm-ring-chi-alg}),\footnote{In fact, our algorithm computes the chromatic number of graphs in $\mathcal{R}_{\geq 4}$.} and more generally, we construct an $O(n^5)$ time algorithm that computes the chromatic number of graphs in $\mathcal{G}_{\text{T}}$ (see Theorem~\ref{thm-GT-chi-alg}).\footnote{The difference between the algorithms from Theorems~\ref{thm-ring-col-alg} and~\ref{thm-GT-col-alg} on the one hand, and the algorithms from Theorems~\ref{thm-ring-chi-alg} and~\ref{thm-GT-chi-alg} on the other, is that the former compute an optimal coloring of the input graph from the relevant class, whereas the latter only compute the chromatic number (but are significantly faster than the former).} In section~\ref{sec:chibound}, we obtain the optimal $\chi$-bounding function for the class $\mathcal{G}_{\text{T}}$ (see Theorem~\ref{thm-GT-chi}). Furthermore, in section~\ref{sec:chibound}, for each odd integer $k \geq 5$, we obtain the optimal bound for the chromatic number in terms of the clique number for $k$-hyperholes and $k$-hyperantiholes.\footnote{We only defined $\chi$-boundedness for hereditary classes, and so, technically, these are not ``$\chi$-bounding functions'' for the classes of $k$-hyperholes and $k$-hyperantiholes. They are, however, the optimal $\chi$-bounding functions for the closures of these classes under induced subgraphs. See section~\ref{sec:chibound} for the details.} Finally, in section~\ref{sec:Hadwiger}, we prove Hadwiger's conjecture for the class $\mathcal{G}_{\text{T}}$ (see Theorem~\ref{thm-Hadwiger-GT}). \section{A few preliminary lemmas} \label{sec:prelim} In this section, we state a few results from the literature, which we use later in the paper. We also prove a few easy results about rings and their induced subgraphs, and about classes $\mathcal{R}_k$ and $\mathcal{R}_{\geq k}$ ($k \geq 4$). Given a graph $G$ and distinct vertices $u,v \in V(G)$, we say that $u$ {\em dominates} $v$ in $G$, and that $v$ is {\em dominated} by $u$ in $G$, whenever $N_G[v] \subseteq N_G[u]$. The following lemma was stated without proof in~\cite{VIK} (see Lemma~1.4 from~\cite{VIK}); it readily follows from the definition of a ring, as the reader can check. \begin{lemma} \cite{VIK} \label{lemma-ring-char} Let $G$ be a graph, and let $(X_1,\dots,X_k)$, with $k \geq 4$, be a partition of $V(G)$. Then $G$ is a $k$-ring with ring partition $(X_1,\dots,X_k)$ if and only if all the following hold:\footnote{As usual, indices of the $X_i$'s are understood to be modulo $k$.} \begin{itemize} \item[(a)] $X_1,\dots,X_k$ are cliques; \item[(b)] for all $i \in \{1,\dots,k\}$, $X_i$ is anticomplete to $V(G) \setminus (X_{i-1} \cup X_i \cup X_{i+1})$; \item[(c)] for all $i \in \{1,\dots,k\}$, some vertex of $X_i$ is complete to $X_{i-1} \cup X_{i+1}$; \item[(d)] for all $i \in \{1,\dots,k\}$, and all distinct $y_i,y_i' \in X_i$, one of $y_i,y_i'$ dominates the other. \end{itemize} \end{lemma} Recall that a graph is {\em chordal} if it contains no holes. The following is Lemma~2.4(a)-(d) from~\cite{VIK}. \begin{lemma} \cite{VIK} \label{lemma-ring-chordal} Let $R$ be a $k$-ring ($k \geq 4$) with ring partition $(X_1,\dots,X_k)$. Then all the following hold: \begin{itemize} \item[(a)] every hole in $R$ intersects each of $X_1,\dots,X_k$ in exactly one vertex; \item[(b)] every hole in $R$ is of length $k$; \item[(c)] for all $i \in \{1,\dots,k\}$, $R \setminus X_i$ is chordal; \item[(d)] $R \in \mathcal{G}_{\text{T}}$. \end{itemize} \end{lemma} Note that Lemma~\ref{lemma-ring-chordal}(b) states that, for an integer $k \geq 4$, every hyperhole in a $k$-ring is of length $k$. On the other hand, Lemma~\ref{lemma-ring-chordal}(d) implies that $\mathcal{R}_{\geq 4} \subseteq \mathcal{G}_{\text{T}}$,\footnote{Indeed, suppose that $G \in \mathcal{R}_{\geq 4}$, and assume inductively that all graphs in $\mathcal{R}_{\geq 4}$ on fewer than $|V(G)|$ vertices belong to $\mathcal{G}_{\text{T}}$. If $G$ is a ring, then Lemma~\ref{lemma-ring-chordal}(d) guarantees that $G \in \mathcal{G}_{\text{T}}$. So suppose that $G$ is not a ring. Then by the definition of $\mathcal{R}_{\geq 4}$, $G$ has a simplicial vertex, call it $v$. Obviously, $K_1 \in \mathcal{G}_{\text{T}}$, and so we may assume that $|V(G)| \geq 2$. Note that no Truemper configuration contains a simplicial vertex, and so $v$ does not belong to any induced Truemper configuration in $G$. Since $\mathcal{G}_{\text{T}}$ was defined by forbidding certain Truemper configurations as induced subgraphs, we deduce that $G$ belongs to $\mathcal{G}_{\text{T}}$ if and only if $G \setminus v$ does. Now, since $\mathcal{R}_{\geq 4}$ is hereditary and contains $G$, we see that $G \setminus v$ belongs to $\mathcal{R}_{\geq 4}$. It then follows from the induction hypothesis that $G \setminus v$ belongs to $\mathcal{G}_{\text{T}}$, and we deduce that $G \in \mathcal{G}_{\text{T}}$.} but we will not need this in the remainder of the paper. Rings can be recognized in polynomial time. More precisely, the following is Lemma~8.14 from~\cite{VIK}. (In all our algorithms, $n$ denotes the number of vertices and $m$ the number of edges of the input graph.) \begin{lemma} \cite{VIK} \label{lemma-detect-ring} There exists an algorithm with the following specifications: \begin{itemize} \item Input: A graph $G$; \item Output: Either the true statement that $G$ is a ring, together with the length and ring partition of the ring, or the true statement that $G$ is not a ring; \item Running time: $O(n^2)$. \end{itemize} \end{lemma} As an easy corollary of Lemma~\ref{lemma-detect-ring}, we can obtain Lemma~\ref{lemma-detect-ring-ord} (below). We remark that the proof (but not the statement) of Lemma~8.14 from~\cite{VIK} in fact gives precisely Lemma~\ref{lemma-detect-ring-ord}. For the sake of completeness, we give a full proof. \begin{lemma} \label{lemma-detect-ring-ord} There exists an algorithm with the following specifications: \begin{itemize} \item Input: A graph $G$; \item Output: Exactly one of the following: \begin{itemize} \item the true statement that $G$ is a ring, together with the length $k$ and a ring partition $(X_1,\dots,X_k)$ of the ring $G$, and for each $i \in \{1,\dots,k\}$, an ordering $X_i = \{u_i^1,\dots,u_i^{|X_i|}\}$ of $X_i$ such that $X_i \subseteq N_G[u_i^{|X_i|}] \subseteq \dots \subseteq N_G[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$, \item the true statement that $G$ is not a ring; \end{itemize} \item Running time: $O(n^2)$. \end{itemize} \end{lemma} \begin{proof} We first run the algorithm from Lemma~\ref{lemma-detect-ring} with input $G$; this takes $O(n^2)$ time. If the algorithm returns the statement that $G$ is not a ring, then we return this statement as well, and we stop. So assume that the algorithm returned the statement that $G$ is a ring, together with the length $k$ and ring partition $(X_1,\dots,X_k)$ of the ring. We then find the degrees of all vertices of $G$, and for each $i \in \{1,\dots,k\}$, we order $X_i$ as $X_i = \{u_i^1,\dots,u_i^{|X_i|}\}$ so that ${\rm deg}_G(u_i^1) \geq \dots \geq {\rm deg}_G(u_i^{|X_i|})$; this takes $O(n^2)$ time. Since we already know that $(X_1,\dots,X_k)$ is a ring partition of $G$, it is easy to see that for all $i \in \{1,\dots,k\}$, we have that $X_i \subseteq N_G[u_i^{|X_i|}] \subseteq \dots \subseteq N_G[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$. We now return the statement that $G$ is a ring of length $k$, the ring partition $(X_1,\dots,X_k)$ of $G$, and for each $i \in \{1,\dots,k\}$, the ordering $X_i = \{u_i^1,\dots,u_i^{|X_i|}\}$ of $X_i$, and we stop. Clearly, the algorithm is correct, and its running time is $O(n^2)$. \end{proof} We remind the reader that a {\em simplicial vertex} is a vertex whose neighborhood is a (possibly empty) clique. A {\em simplicial elimination ordering} of a graph $G$ is an ordering $v_1,\dots,v_n$ of the vertices of $G$ such that for all $i \in \{1,\dots,n\}$, $v_i$ is simplicial in the graph $G[v_i,v_{i+1},\dots,v_n]$. It is well known (and easy to show) that a graph is chordal if and only if it has a simplicial elimination ordering (see~\cite{FulkersonGross}); in particular, every chordal graph contains a simplicial vertex. We also note that there is an $O(n+m)$ time algorithm that either produces a simplicial elimination ordering of the input graph, or determines that the graph is not chordal~\cite{RTL76}. Recall that a graph is {\em perfect} if all its induced subgraphs $H$ satisfy $\chi(H) = \omega(H)$; it is well known (and easy to show) that chordal graphs are perfect~\cite{Berge-German, D61}. The following algorithm is a minor modification of the algorithm described in the introduction of~\cite{HHMMSimplicial}.\footnote{The algorithm from~\cite{HHMMSimplicial} produces a maximal sequence $v_1,\dots,v_t$ ($t \geq 0$) of pairwise distinct vertices of the input graph $G$ such that for all $i \in \{1,\dots,t\}$, $v_i$ is simplicial in either $G \setminus \{v_1,\dots,v_{i-1}\}$ or $\overline{G} \setminus \{v_1,\dots,v_{i-1}\}$. Thus, the algorithm from Lemma~\ref{lemma-simplicial-list-alg} is in fact obtained from the algorithm from~\cite{HHMMSimplicial} by omitting some steps. The running time of the two algorithms is the same. For the sake of completeness, we give all the details for the algorithm that we need (i.e.\ for the algorithm from Lemma~\ref{lemma-simplicial-list-alg}).} \begin{lemma} \label{lemma-simplicial-list-alg} There exists an algorithm with the following specifications: \begin{itemize} \item Input: A graph $G$; \item Output: A maximal sequence $v_1,\dots,v_t$ ($t \geq 0$) of pairwise distinct vertices of $G$ such that for all $i \in \{1,\dots,t\}$, $v_i$ is simplicial in the graph $G \setminus \{v_1,\dots,v_{i-1}\}$;\footnote{If $t = 0$, then the sequence $v_1,\dots,v_t$ is empty and $G$ has no simplicial vertices.} \item Running time: $O(n^3)$. \end{itemize} \end{lemma} \begin{proof} \textbf{Step~0.} First, for all distinct $x,y \in V(G)$, we set \begin{displaymath} \begin{array}{ccc} \text{diff}(x,y) & = & \left\{\begin{array}{lll} |N_G[x] \setminus N_G[y]| & \text{if} & xy \in E(G) \\ \\ 0 & \text{if} & xy \notin E(G) \end{array}\right. \end{array} \end{displaymath} Clearly, computing $\text{diff}(x,y)$ for all possible choices of distinct $x,y \in V(G)$ can be done in $O(n^3)$ time. We will update $\text{diff}(x,y)$ as the algorithm proceeds. Note that a vertex $x \in V(G)$ is simplicial in $G$ if and only if for all $y \in V(G) \setminus \{x\}$, we have that $\text{diff}(x,y) = 0$. Let $L$ be the empty list. We now go to Step~1. \textbf{Step~1.} We first check if there is a vertex $x \in V(G)$ such that for all $y \in V(G) \setminus \{x\}$, we have that $\text{diff}(x,y) = 0$; this can be done in $O(n^2)$ time. If we found no such vertex, then $G$ has no simplicial vertices; in this case, we return the list $L$ and stop. Suppose now that we found such a vertex $x$. First, we set $L := L,x$ (i.e.\ we update $L$ by adding $x$ to the end of $L$). If $x$ is the only vertex of $G$, then we return $L$ and stop. Suppose now that $G$ has at least two vertices. Then, for all distinct $x',y \in V(G) \setminus \{x\}$, we update $\text{diff}(x',y)$ as follows: if $x \in N_G[x'] \setminus N_G[y]$, then we set $\text{diff}(x',y) := \text{diff}(x',y)-1$, and otherwise, we do not change $\text{diff}(x',y)$; this update takes $O(n^2)$ time. Finally, we update $G$ by setting $G := G \setminus x$, and we go to Step~1 with input $G$, $L$, and $\text{diff}(x',y)$ for all distinct $x',y \in V(G)$. Clearly, the algorithm terminates and is correct. Step~0 takes $O(n^3)$ time. We make $O(n)$ calls to Step~1, and otherwise, the slowest step of Step~1 takes $O(n^2)$ time. Thus, the total running time of the algorithm is $O(n^3)$. \end{proof} Recall that chordal graphs are precisely those graphs that admit a simplicial elimination ordering~\cite{FulkersonGross}. So, the algorithm from Lemma~\ref{lemma-simplicial-list-alg} can be used to recognize chordal graphs in $O(n^3)$ time.\footnote{Indeed, suppose that, given an $n$-vertex input graph $G$, the algorithm from Lemma~\ref{lemma-simplicial-list-alg} returned the sequence $v_1,\dots,v_t$. If $t = n$ (i.e.\ $V(G) = \{v_1,\dots,v_t\}$), then $v_1,\dots,v_t$ is a simplicial elimination ordering of $G$, and therefore (by~\cite{FulkersonGross}) $G$ is chordal. Suppose now that $t < n$. Then the maximality of $v_1,\dots,v_t$ guarantees that $G \setminus \{v_1,\dots,v_t\}$ has no simplicial vertices. Then by~\cite{FulkersonGross}, $G \setminus \{v_1,\dots,v_t\}$ is not chordal, and consequently, $G$ is not chordal either.} Lemma~\ref{lemma-omega-ring} (below) follows immediately from Theorem~8.25 from~\cite{VIK}. \begin{lemma} \cite{VIK} \label{lemma-omega-ring} There exists an algorithm with the following specifications: \begin{itemize} \item Input: A graph $G$; \item Output: Either $\omega(G)$, or the true statement that $G \notin \mathcal{G}_{\text{T}}$; \item Running time: $O(n^3)$. \end{itemize} \end{lemma} By Lemma~\ref{lemma-ring-chordal}(d), rings belong to $\mathcal{G}_{\text{T}}$, and so Lemma~\ref{lemma-omega-ring} guarantees that the clique number of a ring can be computed in $O(n^3)$ time. \begin{lemma} \label{lemma-ring-simplicial} Let $k \geq 4$ be an integer. Then every induced subgraph of a $k$-ring either contains a simplicial vertex or is a $k$-ring. More precisely, let $R$ be a $k$-ring with ring partition $(X_1,\dots,X_k)$, and let $Y \subseteq V(R)$ be a nonempty set. Then either $R[Y]$ contains a simplicial vertex, or $R[Y]$ is a $k$-ring with ring partition $(X_1 \cap Y,\dots,X_k \cap Y)$. \end{lemma} \begin{proof} For all $i \in \{1,\dots,k\}$, we set $X_i = \{u_i^1, \dots, u_i^{|X_i|}\}$ so that $X_i \subseteq N_R[u_i^{|X_i|}] \subseteq \dots \subseteq N_R[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$, as in the definition of a ring. For all $i \in \{1,\dots,k\}$, we set $Y_i = X_i \cap Y$. If at least one of $Y_1,\dots,Y_k$ is empty, then Lemma~\ref{lemma-ring-chordal}(c) implies that $R[Y]$ is chordal, and consequently (by~\cite{FulkersonGross}), $R[Y]$ contains a simplicial vertex. So from now on, we assume that $Y_1,\dots,Y_k$ are all nonempty. For all $i \in \{1,\dots,k\}$, let $j_i \in \{1,\dots,|X_i|\}$ be maximal with the property that $u_i^{j_i} \in Y_i$; then $u_i^{j_i}$ is dominated in $R[Y]$ by all other vertices in $Y_i$. If for some $i \in \{1,\dots,k\}$, $u_i^{j_i}$ is anticomplete to $Y_{i-1}$ or $Y_{i+1}$, then it is easy to see that $u_i^{j_i}$ is a simplicial vertex of $R[Y]$, and we are done; otherwise, Lemma~\ref{lemma-ring-char} implies that $R[Y]$ is a ring with ring partition $(Y_1,\dots,Y_k)$. \end{proof} \begin{lemma} \label{lemma-RRk-hered} For all integers $k \geq 4$, both the following hold: \begin{itemize} \item the class $\mathcal{R}_k$ is hereditary and contains all $k$-rings; \item the class $\mathcal{R}_{\geq k}$ is hereditary and contains all rings of length at least $k$. \end{itemize} In particular, the class $\mathcal{R}_{\geq 4}$ is hereditary and contains all rings. \end{lemma} \begin{proof} This follows immediately from Lemma~\ref{lemma-ring-simplicial} and from the relevant definitions. \end{proof} The following lemma (Lemma~\ref{lemma-RRk-description}) will not be used in the remainder of the paper, but the reader may find it informative. We remark that, for each integer $k \geq 4$, Lemmas~\ref{lemma-detect-ring},~\ref{lemma-simplicial-list-alg}, and~\ref{lemma-RRk-description} readily yield $O(n^3)$ time recognition algorithms for the classes $\mathcal{R}_k$ and $\mathcal{R}_{\geq k}$. However, we will not need these algorithms in the remainder of the paper, and so we leave the details to the reader. \begin{lemma} \label{lemma-RRk-description} Let $k \geq 4$ be an integer, and let $G$ be a graph. Then the following are equivalent: \begin{itemize} \item[(a)] $G \in \mathcal{R}_k$; \item[(b)] either $G$ is chordal, or $G$ is a $k$-ring, or $G$ can be obtained from a $k$-ring by repeatedly adding simplicial vertices. \end{itemize} \end{lemma} \begin{proof} Suppose first that (a) holds, i.e.\ that $G \in \mathcal{R}_k$. Let $v_1,\dots,v_t$ ($t \geq 0$) be a maximal sequence of pairwise distinct vertices of $G$ such that for all $i \in \{1,\dots,t\}$, $v_i$ is simplicial in $G \setminus \{v_1,\dots,v_{i-1}\}$. If $V(G) = \{v_1,\dots,v_t\}$, then $v_1,\dots,v_t$ is a simplicial elimination ordering of $G$, and so (by~\cite{FulkersonGross}) $G$ is chordal. Suppose now that $\{v_1,\dots,v_t\} \subsetneqq V(G)$. Set $R = G \setminus \{v_1,\dots,v_t\}$. Since $G \in \mathcal{R}_k$, and since $\mathcal{R}_k$ is hereditary, we see that $R \in \mathcal{R}_k$. On the other hand, by the maximality of $v_1,\dots,v_t$, we know that $R$ has no simplicial vertices. So, by the definition of $\mathcal{R}_k$, $R$ is a $k$-ring. If $t = 0$, then $G = R$, and we have that $G$ is a $k$-ring. On the other hand, if $t \geq 1$, then $G$ can be obtained from the $k$-ring $R$ by adding simplicial vertices $v_t,\dots,v_1$ (in that order). So, (b) holds. Suppose now that (b) holds. Clearly, every induced subgraph of a chordal graph is chordal. Furthermore, by~\cite{FulkersonGross}, every chordal graph has a simplicial vertex. So, if $G$ is chordal, then all its induced subgraphs contain a simplicial vertex, and it follows that $G \in \mathcal{R}_k$. Suppose now that $G$ can be obtained from a $k$-ring by (possibly) repeatedly adding simplicial vertices. But then Lemma~\ref{lemma-ring-simplicial} implies that every induced subgraph of $G$ either is a $k$-ring or has a simplicial vertex, and so $G \in \mathcal{R}_k$. Thus, (a) holds. \end{proof} \begin{lemma} \label{lemma-simplicial-chi} Let $G$ be a graph on at least two vertices, and let $v$ be a simplicial vertex of $G$. Then $\omega(G) = \max\{|N_G[v]|,\omega(G \setminus v)\}$ and $\chi(G) = \max\{\omega(G),\chi(G \setminus v)\}$. \end{lemma} \begin{proof} We first show that $\omega(G) = \max\{|N_G[v]|,\omega(G \setminus v)\}$. Since $v$ is simplicial, $N_G[v]$ is a clique, and we deduce that $\max\{|N_G[v]|,\omega(G \setminus v)\} \leq \omega(G)$. To prove the reverse inequality, let $K$ be a clique of size $\omega(G)$ in $G$. If $v \notin K$, then $K$ is a clique of $G \setminus v$, and so $\omega(G) = |K| \leq \omega(G \setminus v) \leq \max\{|N_G[v]|,\omega(G \setminus v)\}$. So suppose that $v \in K$. Since $K$ is a clique, it follows that $K \subseteq N_G[v]$, and so $\omega(G) = |K| \leq |N_G[v]| \leq \max\{|N_G[v]|,\omega(G \setminus v)\}$. This proves that $\omega(G) = \max\{|N_G[v]|,\omega(G \setminus v)\}$. It remains to show that $\chi(G) = \max\{\omega(G),\chi(G \setminus v)\}$. It is clear that $\max\{\omega(G),\chi(G \setminus v)\} \leq \chi(G)$. For the reverse inequality, we set $\ell = \max\{\omega(G),\chi(G \setminus v)\}$, and we construct a proper coloring of $G$ that uses at most $\ell$ colors. First, we properly color $G \setminus v$ with colors $1,\dots,\ell$. Next, since $N_G[v]$ is a clique, we see that $|N_G(v)| = |N_G[v]|-1 \leq \omega(G)-1 \leq \ell-1$; thus, at least one of our $\ell$ colors was not used on $N_G(v)$, and we can assign this ``unused'' color to $v$. This produces a proper coloring of $G$ that uses at most $\ell$ colors, and we are done. \end{proof} We complete this section by stating the decomposition theorem for the class $\mathcal{G}_{\text{T}}$ proven in~\cite{VIK} (this is Theorem~1.8 from~\cite{VIK}). \begin{theorem} \label{thm-GT-decomp} \cite{VIK} Let $G \in \mathcal{G}_{\text{T}}$. Then one of the following holds: \begin{itemize} \item $G$ is a complete graph, a ring, or a 7-hyperantihole; \item $G$ admits a clique-cutset. \end{itemize} \end{theorem} Finally, we remark that graphs in $\mathcal{G}_{\text{T}}$ can be recognized in $O(n^3)$ time (see Theorem~8.23 from~\cite{VIK}), but we do not need this result in the remainder of the paper. \section{Proof of Theorem~\ref{thm-ring-hyperhole}} \label{sec:main} In this section, we prove Theorem~\ref{thm-ring-hyperhole}. We begin with an easy lemma. \begin{lemma} \label{lemma-ring-large-omega} Let $R$ be a $k$-ring (with $k \geq 4$) such that $\chi(R) = \omega(R)$. Then $R$ contains a $k$-hyperhole $H$ such that $\chi(H) = \chi(R)$. \end{lemma} \begin{proof} Let $(X_1,\dots,X_k)$ be a ring partition of $R$, and for all $i \in \{1,\dots,k\}$, let $X_i = \{u_i^1,\dots,u_i^{|X_i|}\}$ be an ordering of $X_i$ such that $X_i \subseteq N_R[u_i^{|X_i|}] \subseteq \dots \subseteq N_R[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$, as in the definition of a ring. Let $Q$ be a clique of size $\omega(R)$ in $R$. By the definition of a ring, and by symmetry, we may assume that $Q \subseteq X_1 \cup X_2$. Since $u_1^1$ is complete to $X_2$, and since $u_2^1$ is complete to $X_1$, the maximality of $Q$ guarantees that $u_1^1,u_2^1 \in Q$, and in particular, $Q$ intersects both $X_1$ and $X_2$. Set $H = R[Q \cup \{u_3^1,u_4^1,\dots,u_k^1\}]$. Clearly, $H$ is a $k$-hyperhole. Furthermore, we have that $\omega(R) = |Q| \leq \omega(H) \leq \chi(H) \leq \chi(R)$; since $\chi(R) = \omega(R)$, it follows that $\chi(H) = \chi(R)$. \end{proof} In view of Lemma~\ref{lemma-ring-large-omega}, our next lemma (Lemma~\ref{lemma-even-ring-col-alg}) shows that Theorem~\ref{thm-ring-hyperhole} holds for even rings. We will also rely on Lemma~\ref{lemma-even-ring-col-alg} in our coloring algorithm for rings in section~\ref{sec:col}. \begin{lemma} \label{lemma-even-ring-col-alg} Even rings are perfect.\footnote{We remind the reader that a graph is {\em perfect} if all its induced subgraphs $H$ satisfy $\chi(H) = \omega(H)$. In particular, every perfect graph $G$ satisfies $\chi(G) = \omega(G)$. The fact that even rings are perfect easily follows from the Strong Perfect Graph Theorem~\cite{SPGT}. However, here we give an elementary proof of this fact.} Furthermore, there exists an algorithm with the following specifications: \begin{itemize} \item Input: A graph $G$; \item Output: Either an optimal coloring of $G$, or the true statement that $G$ is not an even ring; \item Running time: $O(n^3)$. \end{itemize} \end{lemma} \begin{proof} We begin by constructing the algorithm. We first call the algorithm from Lemma~\ref{lemma-detect-ring-ord} with input $G$; this takes $O(n^2)$ time. If the algorithm returns the answer that $G$ is not a ring, then we return the answer that $G$ is not an even ring, and we stop. So from now on, we assume that the algorithm returned all the following: \begin{itemize} \item the true statement that $G$ is a ring; \item the length $k$ and a ring partition $(X_1,\dots,X_k)$ of the ring $G$; \item for each $i \in \{1,\dots,k\}$, an ordering $X_i = \{u_i^1,\dots,u_i^{|X_i|}\}$ of $X_i$ such that $X_i \subseteq N_G[u_i^{|X_i|}] \subseteq \dots \subseteq N_G[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$. \end{itemize} If $k$ is odd, then we return the answer that $G$ is not an even ring, and we stop. So assume that $k$ is even. Since $G$ is a ring, Lemma~\ref{lemma-ring-chordal}(d) guarantees that $G \in \mathcal{G}_{\text{T}}$, and so we can compute $\omega(G)$ by running the algorithm from Lemma~\ref{lemma-omega-ring} with input $G$; this takes $O(n^3)$ time. We now color $G$ as follows. For all odd $i \in \{1,\dots,k\}$ and all $j \in \{1,\dots,|X_i|\}$, we assign color $j$ to the vertex $u_i^j$; and for all even $i \in \{1,\dots,k\}$ and all $j \in \{1,\dots,|X_i|\}$, we assign color $\omega(G)-j+1$ to the vertex $u_i^j$. Since $|X_i| \leq \omega(G)$ for all $i \in \{1,\dots,k\}$, we see that our coloring uses only colors $1,\dots,\omega(G)$. Let us show that the coloring is proper. Suppose otherwise. By Lemma~\ref{lemma-ring-char}(b), there exist some $i \in \{1,\dots,k\}$, $j \in \{1,\dots,|X_i|\}$, and $\ell \in \{1,\dots,|X_{i+1}|\}$ such that $u_i^j$ and $u_{i+1}^{\ell}$ are adjacent in $G$ and were assigned the same color. Since $u_i^j$ and $u_{i+1}^{\ell}$ are adjacent, we see that $\{u_i^1,\dots,u_i^j\}$ and $\{u_{i+1}^1,\dots,u_{i+1}^{\ell}\}$ are cliques, complete to each other;\footnote{This follows from the properties of our orderings of $X_i$ and $X_{i+1}$.} thus, $\{u_i^1,\dots,u_i^j\} \cup \{u_{i+1}^1,\dots,u_{i+1}^{\ell}\}$ is a clique, and consequently, $j+\ell \leq \omega(G)$. On the other hand, by construction, we have that: \begin{itemize} \item if $i$ is odd, then $u_i^j$ received color $j$, and $u_{i+1}^{\ell}$ received color $\omega(G)-\ell+1$; \item if $i$ is even, then $u_i^j$ received color $\omega(G)-j+1$, and $u_{i+1}^{\ell}$ received color $\ell$. \end{itemize} Since vertices $u_i^j$ and $u_{i+1}^{\ell}$ received the same color, it follows that either $j = \omega(G)-\ell+1$ or $\omega(G)-j+1 = \ell$; in either case, we get that $j+\ell = \omega(G)+1$, contrary to the fact that $j+\ell \leq \omega(G)$. This proves that our coloring of $G$ is indeed proper. Furthermore, as pointed out above, this coloring uses at most $\omega(G)$ colors. Since $\omega(G) \leq \chi(G)$, we deduce that our coloring is optimal, and that $\chi(G) = \omega(G)$. We now return this coloring of $G$, and we stop. Clearly, the algorithm is correct, and its running time is $O(n^3)$. Note, furthermore, that we have established that all even rings $R$ satisfy $\chi(R) = \omega(R)$. The fact that even rings are perfect now follows from Lemmas~\ref{lemma-ring-simplicial} and~\ref{lemma-simplicial-chi} by an easy induction. \end{proof} As we pointed out above, Lemmas~\ref{lemma-ring-large-omega} and~\ref{lemma-even-ring-col-alg} together imply that even rings satisfy Theorem~\ref{thm-ring-hyperhole}. We devote the remainder of the section to proving Theorem~\ref{thm-ring-hyperhole} for odd rings. Given a graph $G$, a coloring $c$ of $G$, and distinct colors $a,b$ used by $c$, we set $T^{a,b}_{G,c} = G[\{x \in V(G) \mid \text{$c(x) = a$ or $c(x) = b$\}}]$;\footnote{Thus, $T^{a,b}_{G,c}$ is the subgraph of $G$ induced by the vertices colored $a$ or $b$.} note that if $c$ is a proper coloring of $G$, then $T^{a,b}_{G,c}$ is a bipartite graph, and if, in addition, $G$ contains no even holes, then $T^{a,b}_{G,c}$ is a forest. After introducing a few more definitions, we describe the structure of the components $Q$ of $T^{a,b}_{G,c}$ when $G$ is an induced subgraph of an odd ring (see Lemma~\ref{lemma-Rab}). We now need a few more definitions. Let $k \geq 5$ be an odd integer, let $R$ be a $k$-ring with ring partition $(X_1,\dots,X_k)$, and for each $i \in \{1,\dots,k\}$, let $X_i = \{u_i^1,\dots,u_i^{|X_i|}\}$ be an ordering of $X_i$ such that $X_i \subseteq N_R[u_i^{|X_i|}] \subseteq \dots \subseteq N_R[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$, as in the definition of a ring. For all $i \in \{1,\dots,k\}$ and $j,\ell \in \{1,\dots,|X_i|\}$ such that $j \leq \ell$ (resp. $j < \ell$), we say that $u_i^j$ is {\em lower} (resp.\ {\em strictly lower}) than $u_i^{\ell}$, and that $u_i^{\ell}$ is {\em higher} (resp.\ {\em strictly higher}) than $u_i^j$; under these circumstances, we also write $u_i^j \leq u_i^{\ell}$ (resp.\ $u_i^j < u_i^{\ell}$) and $u_i^{\ell} \geq u_i^j$ (resp.\ $u_i^{\ell} > u_i^j$). For each $i \in \{1,\dots,k\}$ let $s_i = u_i^1$ and $t_i = u_i^{|X_i|}$.\footnote{Thus, $s_i$ is the lowest and $t_i$ the highest vertex in $X_i$. Note that this means that $s_i$ is a highest-degree and $t_i$ a lowest-degree vertex in $X_i$.} Further, suppose that $c$ is a proper coloring of $R \setminus t_2$. For all $X \subseteq V(R) \setminus \{t_2\}$, set $c(X) = \{c(x) \mid x \in X\}$. Given distinct colors $a,b \in c(V(R) \setminus \{t_2\})$ and an index $i \in \{1,\dots,k\}$, we say that $a$ is {\em lower} than $b$ in $X_i$ with respect to $c$, and that $b$ is {\em higher} than $a$ in $X_i$ with respect to $c$, provided that either \begin{itemize} \item $a \in c(X_i \setminus \{t_2\})$ and $b \notin c(X_i \setminus \{t_2\})$,\footnote{Obviously, if $i \neq 2$, then $X_i \setminus \{t_2\} = X_i$.} or \item there exist indices $j,\ell \in \{1,\dots,|X_i|\}$ such that $j < \ell$, $c(u_i^j) = a$, and $c(u_i^{\ell}) = b$. \end{itemize} Let $c_1 = c(s_1)$.\footnote{Note that this means that $c_1 \notin c(X_2 \setminus \{t_2\})$. This is because $c(s_1) = c_1$, $s_1$ is complete to $X_2$ in $R$, and $c$ is a proper coloring of $R \setminus t_2$.} We say that $c$ is {\em unimprovable} if for all colors $a \in c(V(R) \setminus \{t_2\})$ such that $a \neq c_1$, and all components $Q$ of $T^{c_1,a}_{R \setminus t_2,c}$ that do not contain $s_1$, both the following are satisfied: \begin{itemize} \item for all odd $i \in \{3,\dots,k\}$ such that $Q$ intersects $X_i$, $c_1$ is lower than $a$ in $X_i$ with respect to $c$; \item for all even $i \in \{3,\dots,k\}$ such that $Q$ intersects $X_i$, $c_1$ is higher than $a$ in $X_i$ with respect to $c$. \end{itemize} We remark that if $c$ is an unimprovable coloring of $R \setminus t_2$, then by definition, $c$ is a proper coloring of $R \setminus t_2$, but it need not be an optimal coloring of $R \setminus t_2$, i.e.\ it may possibly use more than $\chi(R \setminus t_2)$ colors. \begin{lemma} \label{lemma-Rab} Let $k \geq 5$ be an odd integer, let $R$ be a $k$-ring with ring partition $(X_1,\dots,X_k)$, and for each $i \in \{1,\dots,k\}$, let $X_i = \{u_i^1,\dots,u_i^{|X_i|}\}$ be an ordering of $X_i$ such that $X_i \subseteq N_R[u_i^{|X_i|}] \subseteq \dots \subseteq N_R[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$. Let $G$ be an induced subgraph of $R$, let $c$ be a proper coloring of $G$, let $a,b$ be distinct colors used by $c$, and let $Q$ be any component of $T^{a,b}_{G,c}$. Then there are integers $i,j \in \{1,\dots,k\}$ such that $V(Q)\subseteq X_i\cup X_{i+1}\cup\cdots\cup X_{j-1}\cup X_j$,\footnote{As usual, indices are understood to be modulo $k$.} and such that $Q$ consists of an induced path $p_i,\dots,p_j$, where $p_{\ell}\in X_{\ell}$ for all $\ell \in \{i, \dots, j\}$, plus, optionally for each $\ell \in \{i,\dots,j\}$, a vertex $p'_{\ell}\in X_{\ell}$, strictly higher than $p_{\ell}$ in $X_{\ell}$,\footnote{So, if $p_{\ell}'$ exists, then it is dominated by $p_{\ell}$ in $R$.} with $N_Q(p'_{\ell})=\{p_{\ell}\}$. \end{lemma} \begin{proof} By Lemma~\ref{lemma-ring-chordal}, all holes in $R$ are of length $k$, and in particular, $R$ contains no even holes. The result now readily follows from the relevant definitions. \end{proof} Our next lemma shows that any proper coloring of $R \setminus t_2$ (where $R$ and $t_2$ are as above) can be turned into an unimprovable coloring that uses no more colors than the original coloring of $R \setminus t_2$.\footnote{In particular, this implies that if $R \setminus t_2$ is $r$-colorable, then there exists an unimprovable coloring of $R \setminus t_2$ that uses at most $r$ colors.} \begin{lemma} \label{lemma-unimprov-alg} There exists an algorithm with the following specifications: \begin{itemize} \item Input: An odd ring $R$ with ring partition $(X_1,\dots,X_k)$, for each $i \in \{1,\dots,k\}$, an ordering $X_i = \{u_i^1,\dots,u_i^{|X_i|}\}$ of $X_i$ such that $X_i \subseteq N_R[u_i^{|X_i|}] \subseteq \dots \subseteq N_R[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$, and a proper coloring $c$ of $R \setminus u_2^{|X_2|}$; \item Output: An unimprovable coloring of $R \setminus u_2^{|X_2|}$ that uses no more colors than $c$ does; \item Running time: $O(n^4)$. \end{itemize} \end{lemma} \begin{proof} To simplify notation, for all $i \in \{1,\dots,k\}$, we set $s_i = u_i^1$ and $t_i = u_i^{|X_i|}$. (Thus, $c$ is a proper coloring of $R \setminus t_2$.) Let $r$ be the number of colors used by $c$; by symmetry, we may assume that $c:V(R) \setminus \{t_2\} \rightarrow \{1,\dots,r\}$. Set $c_1 = c(s_1)$. Now, for every proper coloring $\widetilde{c}:V(R) \setminus \{t_2\} \rightarrow \{1,\dots,r\}$ of $R \setminus t_2$ such that $\widetilde{c}(s_1) = c_1$,\footnote{Note that this implies that $c_1 \notin \widetilde{c}(X_2 \setminus \{t_2\})$. This is because $\widetilde{c}(s_1) = c_1$, $s_1$ is complete to $X_2$ in $R$, and $\widetilde{c}$ is a proper coloring of $R \setminus t_2$.} we define the {\em rank} of $\widetilde{c}$, denoted by $\text{rank}(\widetilde{c})$, as follows. \begin{itemize} \item For all odd $i \in \{3,\dots,k\}$, if there exists an index $j \in \{1,\dots,|X_i|\}$ such that $\widetilde{c}(u_i^j) = c_1$,\footnote{Note that if $j$ exists, then it is unique. This is because $X_i$ is a clique of $R \setminus t_2$, and $\widetilde{c}$ is a proper coloring of $R \setminus t_2$.} then we set $r_i(\widetilde{c}) = j$, and otherwise, we set $r_i(\widetilde{c}) = |X_i|+1$. \item For all even $i \in \{3,\dots,k\}$, if there exists an index $j \in \{1,\dots,|X_i|\}$ such that $\widetilde{c}(u_i^j) = c_1$,\footnote{As before, if $j$ exists, then it is unique.} then we set $r_i(\widetilde{c}) = |X_i|-j+2$, and otherwise, we set $r_i(\widetilde{c}) = 1$. \item We set $\text{rank}(\widetilde{c}) = \sum\limits_{i=3}^k r_i(\widetilde{c})$.\footnote{Note that $k-2 \leq \text{rank}(\widetilde{c}) \leq k-2+\sum\limits_{i=3}^k |X_i|$. So, rank can take at most $1+\sum\limits_{i=3}^k |X_i| < n$ different values.} \end{itemize} The algorithm proceeds as follows. We check whether the input coloring $c$ is unimprovable by examining all colors $a \in \{1,\dots,r\} \setminus \{c_1\}$, and all components $Q$ of $T^{c_1,a}_{R \setminus t_2,c}$ that do not contain $s_1$; this can be done in $O(n^3)$ time. If $c$ is unimprovable, then we return $c$, and we stop. Otherwise, the algorithm found some color $a \in \{1,\dots,r\} \setminus \{c_1\}$, some component $Q$ of $T^{c_1,a}_{R \setminus t_2,c}$ that does not contain $s_1$, and some index $i^* \in \{3,\dots,k\}$ such that $Q$ intersects $X_{i^*}$ and either \begin{itemize} \item $i^*$ is odd, and $a$ is lower than $c_1$ in $X_{i^*}$ with respect to $c$; or \item $i^*$ is even, and $a$ is higher than $c_1$ in $X_{i^*}$ with respect to $c$. \end{itemize} Lemma~\ref{lemma-Rab} then implies that both the following hold: \begin{itemize} \item for all odd $i \in \{3,\dots,k\}$ such that $Q$ intersects $X_i$, $a$ is lower than $c_1$ in $X_i$ with respect to $c$; \item for all even $i \in \{3,\dots,k\}$ such that $Q$ intersects $X_i$, $a$ is higher than $c_1$ in $X_i$ with respect to $c$. \end{itemize} Let $c'$ be the coloring of $R \setminus t_2$ obtained from $c$ by swapping colors $c_1$ and $a$ on $Q$.\footnote{Since $Q$ does not contain $s_1$, we have that $c'(s_1) = c(s_1) = c_1$.} Note that $\text{rank}(c') < \text{rank}(c)$. We now update the coloring $c$ by setting $c := c'$, and we obtain an unimprovable coloring of $R \setminus t_2$ by making a recursive call to the algorithm. The algorithm terminates because the rank of the coloring $c$ decreases before each recursive call. We make $O(n)$ recursive calls,\footnote{This is because rank can take at most $n$ different values, and the rank of our coloring decreases before each recursive call.} and otherwise, the slowest step of the algorithm takes $O(n^3)$ time. So, the total running time of the algorithm is $O(n^4)$. \end{proof} We now prove a technical lemma (Lemma~\ref{lemma-main-technical}) that is at the heart of our proof of Theorem~\ref{thm-ring-hyperhole} for odd rings. We also rely on Lemma~\ref{lemma-main-technical} in our coloring algorithm for rings.\footnote{More precisely, our coloring algorithm for rings relies on Lemma~\ref{lemma-reduce-chi}, which is an easy corollary of Lemma~\ref{lemma-main-technical} and Theorem~\ref{thm-ring-hyperhole}.} We remark that in our proof of Lemma~\ref{lemma-main-technical}, we repeatedly rely on Lemma~\ref{lemma-Rab} without explicitly stating this.\footnote{Essentially, every time we consider a component $Q$ as in Lemma~\ref{lemma-Rab}, we keep in mind the structure of $Q$, as described in Lemma~\ref{lemma-Rab}.} \begin{lemma} \label{lemma-main-technical} Let $k \geq 5$ be an odd integer, let $R$ be a $k$-ring with ring partition $(X_1,\dots,X_k)$, and for each $i \in \{1,\dots,k\}$, let $X_i = \{u_i^1,\dots,u_i^{|X_i|}\}$ be an ordering of $X_i$ such that $X_i \subseteq N_R[u_i^{|X_i|}] \subseteq \dots \subseteq N_R[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$. For all $i \in \{1,\dots,k\}$, set $s_i = u_i^1$ and $t_i = u_i^{|X_i|}$. Let $c$ be an unimprovable coloring of $R \setminus t_2$, and let $r$ be the number of colors used by $c$.\footnote{In particular, $c$ is a proper coloring of $R \setminus t_2$. Furthermore, we have that $\chi(R \setminus t_2) \leq r$, and this inequality may possibly be strict.} Let $c_1 = c(s_1)$, and let $S = \{x \in V(R) \mid x \neq t_2, c(x) = c_1\}$.\footnote{Note that $S$ is a stable set in $R \setminus t_2$. Furthermore, $s_1 \in S$, and in particular, $S \neq \emptyset$.} Then both the following hold: \begin{itemize} \item[(a)] either $\omega(R \setminus S) \leq r-1$, or $R$ contains a $k$-hyperhole of chromatic number $r+1$; \item[(b)] if every $k$-ring $R'$ such that $|V(R')| < |V(R)|$ contains a $k$-hyperhole of chromatic number $\chi(R')$, then either $\chi(R \setminus S) \leq r-1$, or $R$ contains a $k$-hyperhole of chromatic number $r+1$. \end{itemize} \end{lemma} \begin{proof} By hypotheses, we have that $\chi(R \setminus t_2) \leq r$; it follows that $\omega(R) \leq \chi(R) \leq r+1$. If $\omega(R) = r+1$, then both (a) and (b) follow from Lemma~\ref{lemma-ring-large-omega}; thus, we may assume that $\omega(R) \leq r$. Set $Y_1 = N_R(t_2) \cap X_1$, $X_2' = X_2 \setminus \{t_2\}$, and $Y_3 = N_R(t_2) \cap X_3$. Note that $N_R(t_2) = Y_1 \cup X_2' \cup Y_3$, with $Y_1,X_2',Y_3$ pairwise disjoint. Furthermore, we have that $s_1 \in Y_1$ and $s_3 \in Y_3$, and in particular, $Y_1$ and $Y_3$ are nonempty (the set $X_2'$ may possibly be empty). Finally, we remark that $Y_1 \cup X_2$ and $X_2 \cup Y_3$ are maximal cliques of $R$. Let $C$ be the set of colors used by $c$; then $|C| = r$. To simplify notation, for all distinct colors $a,b \in C$, we write $T^{a,b}$ instead of $T^{a,b}_{R \setminus t_2,c}$. Further, for all $i \in \{1,\dots,k\} \setminus \{2\}$ and $a \in c(X_i)$, we denote by $x_i^a$ the (unique) vertex of $X_i$ to which $c$ assigned color $a$; similarly, for all $a \in c(X_2')$, we denote by $x_2^a$ the (unique) vertex of $X_2'$ to which $c$ assigned color $a$. Finally, when we say that some color is higher or lower than some other color in some $X_i$, we always mean this with respect to our coloring $c$. \begin{quote} \begin{claim} \label{claim-a} Either $\omega(R \setminus S) \leq r-1$, or $R$ contains a $k$-hyperhole of chromatic number $r+1$. In other words, (a) holds. \end{claim} \end{quote} {\em Proof of Claim~\ref{claim-a}.} Since $\omega(R) \leq r$, we have that $\omega(R \setminus S) \leq r$. Thus, we may assume that $\omega(R \setminus S) = r$, for otherwise we are done; since $\omega(R) \leq r$, this implies that $\omega(R) = r$. Since $c$ is a proper coloring of $R \setminus t_2$ that uses only $r$ colors, and since $S$ is a color class of the coloring $c$, we see that $S$ intersects all cliques of size $r$ in $R$ that do not contain $t_2$. Furthermore, there are exactly two maximal cliques in $R$ that contain $t_2$, namely $Y_1 \cup X_2$ and $X_2 \cup Y_3$. Since $S$ intersects $Y_1 \cup X_2$ (because $s_1 \in Y_1 \cap S$), we deduce that $X_2 \cup Y_3$ is the unique clique of $R \setminus S$ of size $r$. (Note that this implies that $X_2' \cup Y_3$ is a clique of size $r-1$.) In particular, $c_1 \notin c(X_2' \cup Y_3)$. Consider any color $a \in c(Y_3)$, and let $Q$ be the component of $T^{c_1,a}$ that contains the vertex of $Y_3$ colored $a$. Since $c_1 \notin c(Y_3)$, we see that $a \neq c_1$, and furthermore, $a$ is lower than $c_1$ in $X_3$. So, since $c$ is unimprovable, we have that $s_1 \in V(Q)$. Further, since $c_1 \notin c(X_2')$, we see that $V(Q) \cap X_2' = \emptyset$. We now deduce that the following hold: \begin{itemize} \item for every odd $i \neq 1$, we have that $c(Y_3) \subseteq c(X_i)$; \item for every even $i \neq 2$, some vertex of $X_i$ is colored $c_1$,\footnote{Recall that this vertex is called $x_i^{c_1}$.} and furthermore, this vertex is adjacent to all vertices of $X_{i-1} \cup X_{i+1}$ that received a color used on $Y_3$. \end{itemize} For odd $i \geq 5$, let $h_i$ be the highest vertex of $X_i$ that is adjacent both to $x_{i-1}^{c_1}$ and to $x_{i+1}^{c_1}$.\footnote{Let us check that such an $h_i$ exists. Since $i \geq 5$ is odd, we see that either $5 \leq i \leq k-2$ or $i = k$. If $5 \leq i \leq k-2$, then $i-1,i+1 \geq 4$ are both even, and so by what we just showed, $x_{i-1}^{c_1}$ and $x_{i+1}^{c_1}$ are both defined. If $i = k$, then once again, $i-1 \geq 4$ is even, and so $x_{i-1}^{c_1}$ is defined, and furthermore, since our subscripts are understood to be modulo $k$, we have that $x_{i+1}^{c_1} = x_1^{c_1} = s_1$. So, in either case, $x_{i-1}^{c_1}$ and $x_{i+1}^{c_1}$ are both defined. Moreover, at least one vertex of $X_i$ (namely, the vertex $s_i$) is adjacent both to $x_{i-1}^{c_1}$ and to $x_{i-1}^{c_1}$. So, $h_i$ exists.} We now define sets $Z_1,\dots,Z_k$ as follows: \begin{itemize} \item let $Z_1 = \{s_1\}$, $Z_2 = X_2$, and $Z_3 = Y_3$; \item for all even $i \geq 4$, let $Z_i = \{x \in X_i \mid x \leq x_i^{c_1}\}$; \item for all odd $i \geq 5$, let $Z_i = \{x \in X_i \mid x \leq h_i\}$. \end{itemize} Finally, let $H = R[Z_1 \cup Z_2 \cup \dots \cup Z_k]$. By construction, $H$ is a $k$-hyperhole of $R$; thus, $\chi(H) \leq \chi(R) \leq r+1$. If $\chi(H) = r+1$, then we are done. So assume that $\chi(H) \leq r$. Then $\Big\lceil \frac{2|V(H)|}{k-1} \Big\rceil = \Big\lceil \frac{|V(H)|}{\alpha(H)} \Big\rceil \leq \chi(H) \leq r$. It follows that $|V(H)| \leq \frac{k-1}{2}r$, and consequently, $|V(H) \setminus \{t_2\}| < \frac{k-1}{2}r$. Now, $X_2' \cup Y_3$ is a clique of size $r-1$ in $R \setminus t_2$, and so $|c(X_2' \cup Y_3)| = r-1$. Furthermore, we know that $c_1 \notin c(X_2' \cup Y_3)$, and so $|\{c_1\} \cup c(X_2' \cup Y_3)| = r$. Since $|V(H) \setminus \{t_2\}| < \frac{k-1}{2}r$, we see that some color from $\{c_1\} \cup c(X_2' \cup Y_3)$ appears on fewer than $\frac{k-1}{2}$ vertices of $H \setminus t_2$. Now, by construction, every color from $\{c_1\} \cup c(Y_3)$ appears $\frac{k-1}{2}$ times on $H \setminus t_2$. It follows that some color $d \in c(X_2')$ appears fewer than $\frac{k-1}{2}$ times on $H \setminus t_2$. Thus, there exists some even $i \geq 4$ such that $d \notin c(Z_i)$;\footnote{By the construction of $Z_i$, this implies that $d \neq c_1$, and that $d$ is higher than $c_1$ in $X_i$.} let $i$ be the smallest such index. Thus, $d$ appears on each $Z_j$, for even $j < i$, and there are $\frac{i}{2}-1$ such $j$'s. On the other hand, let $Q$ be the component of $T^{c_1,d}$ that contains $x_i^{c_1}$. Now, we have that $i \geq 4$ is even, and that $d$ is higher than $c_1$ in $X_i$; since $c$ is unimprovable, we deduce that $s_1 \in V(Q)$. It follows that each $Z_j$, for odd $j > i$, contains a vertex colored $d$; there are $\lceil \frac{k-i}{2} \rceil = \frac{k-i+1}{2}$ such $j$'s. So, in total, at least $(\frac{i}{2}-1)+\frac{k-i+1}{2} = \frac{k-1}{2}$ vertices of $H \setminus t_2$ are colored $d$, contrary to our choice of $d$.~$\blacksquare$ \medskip It remains to prove (b). For this, we assume that both the following hold: \begin{itemize} \item every $k$-ring $R'$ such that $|V(R')| < |V(R)|$ contains a $k$-hyperhole of chromatic number $\chi(R')$; \item $\chi(R \setminus S) \geq r$; \end{itemize} and we prove that $R$ contains a $k$-hyperhole of chromatic number $r+1$. Since $S$ is a color class of a proper coloring of $R \setminus t_2$ that uses at most $r$ colors, we see that $\chi\Big(R \setminus (S \cup \{t_2\})\Big) \leq r-1$; consequently, $\chi(R \setminus S) \leq r$. Since $\chi(R \setminus S) \geq r$, it follows that $\chi(R \setminus S) = r$. Further, in view of (a), we may assume that $\omega(R \setminus S) \leq r-1$. \begin{quote} \begin{claim} \label{claim-R-S-hyperhole} $R \setminus S$ contains a $k$-hyperhole $H$ such that $\chi(H) = \Big\lceil \frac{2|V(H)|}{k-1} \Big\rceil = r$. \end{claim} \end{quote} {\em Proof of Claim~\ref{claim-R-S-hyperhole}.} Let $v_1,\dots,v_t$ (with $t \geq 0$) be a maximal sequence of pairwise distinct vertices in $R \setminus S$ such that for all $i \in \{1,\dots,t\}$, $v_i$ is simplicial in $R \setminus (S \cup \{v_1,\dots,v_{i-1}\})$. Set $A = \{v_1,\dots,v_t\}$. Suppose first that $R \setminus S = A$. Then $v_1,\dots,v_t$ is a simplicial elimination ordering of $R \setminus S$, and so by coloring $R \setminus S$ greedily using the ordering $v_t,\dots,v_1$, we obtain a proper coloring of $R \setminus S$ that uses only $\omega(R \setminus S)$ colors, contrary to the fact that $\chi(R \setminus S) = r > \omega(R \setminus S)$. So, $R \setminus S \neq A$. Lemma~\ref{lemma-ring-simplicial} and the maximality of $A$ now imply that $R \setminus (S \cup A)$ is a $k$-ring. Since $S \neq \emptyset$, the $k$-ring $R \setminus (S \cup A)$ has fewer vertices than $R$, and so $R \setminus (S \cup A)$ contains a $k$-hyperhole $H$ such that $\chi(H) = \chi\Big(R \setminus (S \cup A)\Big)$. Now, Lemma~\ref{lemma-simplicial-chi} and an easy induction guarantee that \begin{displaymath} \chi(R \setminus S) = \max\Big\{\omega(R \setminus S),\chi\Big(R \setminus (S \cup A)\Big)\Big\}. \end{displaymath} Since $\chi(R \setminus S) = r$, $\omega(R \setminus S) \leq r-1$, and $\chi(H) = \chi\Big(R \setminus (S \cup A)\Big)$, we deduce that $\chi(H) = r$. Since $\omega(H) \leq \omega(R \setminus S) \leq r-1$, we see that $\omega(H) < \chi(H)$, and so Lemma~\ref{lemma-hyperhole-chi-formula} implies that $\chi(H) = \Big\lceil \frac{|V(H)|}{\alpha(H)} \Big\rceil = \Big\lceil \frac{2|V(H)|}{k-1} \Big\rceil$. Thus, $\chi(H) = \Big\lceil \frac{2|V(H)|}{k-1} \Big\rceil = r$.~$\blacksquare$ \medskip From now on, let $H$ be as in Claim~\ref{claim-R-S-hyperhole}. Our goal is to find a hyperhole in $R$ of size at least $|V(H)|+\frac{k+1}{2}$; this will imply\footnote{The details are given at the end of the proof of the lemma.} that the chromatic number of that hyperhole is at least $r+1$,\footnote{Since $\chi(R) \leq r+1$, we see that any hyperhole in $R$ of chromatic number at least $r+1$ in fact has chromatic number exactly $r+1$.} which is what we need. For each $i \in \{1,\dots,k\}$, let $h_i$ be the highest vertex of $X_i \cap V(H)$. Then $h_1,\dots,h_k,h_1$ is a $k$-hole in $R$, and we may assume that for all $i \in \{1,\dots,k\}$, we have that $V(H) \cap X_i = \{x \in X_i \mid x \leq h_i\} \setminus S$.\footnote{Indeed, set $H' = R[\bigcup\limits_{i=1}^k (\{x \in X_i \mid x \leq h_i\} \setminus S)]$. It is clear that $H'$ is a $k$-hyperhole in $R \setminus S$, and that $H'$ contains $H$ as an induced subgraph. So, $r = \chi(H) \leq \chi(H') \leq \chi(R \setminus S) = r$, and it follows that $\chi(H') = r$. On the other hand, $\omega(H') \leq \omega(R \setminus S) \leq r-1 < \chi(H')$, and so Lemma~\ref{lemma-hyperhole-chi-formula} implies that $\chi(H') = \Big\lceil \frac{|V(H')|}{\alpha(H')} \Big\rceil = \Big\lceil \frac{2|V(H')|}{k-1} \Big\rceil$. Thus, $\chi(H') = \Big\lceil \frac{2|V(H')|}{k-1} \Big\rceil = r$. So, if $H' \neq H$, then from now on, instead of $H$, we simply consider $H'$.} In particular, we have that $s_1,\dots,s_k \in V(H) \cup S$. Recall that $c_1 = c(s_1)$. Let $j$ be the largest odd index such that $c(s_i) = c_1$ for all odd $i \in \{1,\dots,j\}$. Then $j \leq k-2$.\footnote{Indeed, $s_k$ and $s_1$ are adjacent, and $c(s_1) = c_1$; so, $c(s_k) \neq c_1$, and it follows that $j \neq k$. Since $j$ and $k$ are both odd, we deduce that $j \leq k-2$.} Furthermore, since $s_j$ is complete to $X_{j+1}$, we have that $c_1 \notin c(X_{j+1})$. \begin{quote} \begin{claim} \label{claim-c1-in-even} $c_1 \in c(X_i)$ for every even index $i \geq j+3$.\footnote{So, $x_i^{c_1}$ is defined for every even index $i \geq j+3$.} \end{claim} \end{quote} {\em Proof of Claim~\ref{claim-c1-in-even}.} Suppose otherwise, and fix the smallest even index $i \geq j+3$ such that $c_1 \notin c(X_i)$. If $c(s_{i-1}) = c_1$, then: \begin{itemize} \item if $i-1 = j+2$, then the choice of $j$ is contradicted; \item if $i-1 \geq j+4$, then the choice of $i$ is contradicted.\footnote{We are using the fact that $s_{i-1}$ is complete to $X_{i-2}$, and so $c(s_{i-1}) \notin c(X_{i-2})$.} \end{itemize} It follows that $c(s_{i-1}) \neq c_1$. Set $c_{i-1} = c(s_{i-1})$; since $s_{i-1}$ is complete to $X_i$, we have that $c_{i-1} \notin c(X_i)$. Let $Q$ be the component of $T^{c_1,c_{i-1}}$ that contains $s_{i-1}$. We know that $c_1,c_{i-1} \notin c(X_i)$, and so $V(Q) \cap X_i = \emptyset$. On the other hand, by the parity of $i$ and $j$, and by the fact that $c_1 \notin c(X_{j+1})$, we have that $V(Q) \cap X_{j+1} = \emptyset$. Thus, $V(Q) \subseteq X_{j+2} \cup \dots \cup X_{i-1}$. We now have that $s_1 \notin V(Q)$, that $i-1 \geq 3$ is odd, that $Q$ intersects $X_{i-1}$, and that $c_{i-1}$ is lower than $c_1$ in $X_{i-1}$. But this contradicts the fact that $c$ is unimprovable.~$\blacksquare$ \medskip Recall that $h_i$ be the highest vertex of $X_i \cap V(H)$. Let $\ell$ be the largest odd index such that for every odd $i \in \{1,\dots,\ell\}$, the coloring $c$ assigns color $c_1$ to some vertex of $X_i$ lower than $h_i$.\footnote{So, for all odd $i \in \{1,\dots,\ell\}$, we have that $x_i^{c_1}$ is defined and satisfies $x_i^{c_1} \leq h_i$.} Clearly, $j \leq \ell \leq k-2$.\footnote{The fact that $j \leq \ell$ is immediate from the the choice of $j$ and $\ell$. The fact that $\ell \neq k$ follows from the fact that $c(s_1) = c_1$, and that $s_1$ is complete to $X_k$, so that $c_1 \notin c(X_k)$. Since $\ell$ and $k$ are both odd, it follows that $\ell \leq k-2$.} Now, we define vertices $w_1,\dots,w_k$ as follows: \begin{itemize} \item for $i \leq \ell+2$, let $w_i = h_i$; \item for even $i \geq \ell+3$, let $w_i = \max\{h_i,x_i^{c_1}\}$;\footnote{Claim~\ref{claim-c1-in-even} guarantees that $c_1 \in c(X_i)$, and so $x_i^{c_1}$ is defined.} \item for odd $i \geq \ell+4$, let $w_i$ be the highest vertex of $X_i \cap V(H)$ that is adjacent to $x_{i-1}^{c_1}$.\footnote{Let us check that such a $w_i$ exists. First, Claim~\ref{claim-c1-in-even} guarantees that $x_{i-1}^{c_1}$ is defined. It now suffices to show that some vertex of $X_i \cap V(H)$ is adjacent to $x_{i-1}^{c_1}$. Clearly, $s_i$ is adjacent to $x_{i-1}^{c_1}$. Since $c(x_{i-1}^{c_1}) = c_1$, and since $c$ is a proper coloring of $R \setminus t_2$, we have that $c(s_i) \neq c_1$; consequently, $s_i \notin S$. Since $s_1,\dots,s_k \in V(H) \cup S$, we deduce that $s_i \in V(H)$. So, $X_i \cap V(H)$ contains a vertex (namely $s_i$) that is adjacent to $x_{i-1}^{c_1}$.} \end{itemize} For all $i \in \{1,\dots,k\}$, let $W_i = \{x \in X_i \mid x \leq w_i\}$. Further, let $W = R[W_1 \cup \dots \cup W_k]$. Finally, let $S_W = \{x \in V(W) \mid x \neq t_2, c(x) = c_1\}$. We note that, by construction, $|S_W| \geq \frac{k-1}{2}$. \begin{quote} \begin{claim} \label{claim-W-hyperhole} $W$ is a $k$-hyperhole. \end{claim} \end{quote} {\em Proof of Claim~\ref{claim-W-hyperhole}.} Suppose otherwise. Then there exists some even $i \geq \ell+3$ such that $x_i^{c_1}$ is nonadjacent to $w_{i-1}$.\footnote{Let us justify this. By supposition, $W$ is not a $k$-hyperhole, and so there exists some $i \in \{1,\dots,k\}$ such that $w_i$ is nonadjacent to $w_{i-1}$ in $R$. By the construction of $W$, we have that $w_1,\dots,w_{\ell+2} \in V(H)$, as well as that $w_k \in V(H)$; since $H$ is a hyperhole, we deduce that $i \geq \ell+3$. Suppose that $i$ is odd (thus, $i \geq \ell+4$). By Claim~\ref{claim-c1-in-even}, $x_{i-1}^{c_1}$ is defined, and since $i \geq \ell+4$ is odd, we know that $w_i$ is adjacent to $x_{i-1}^{c_1}$. On the other hand, since $i \geq \ell+4$ is odd, we have that $w_i \in V(H)$, and so $w_i$ is adjacent to $h_{i-1}$. But since $i-1 \geq \ell+3$ is even, we have by construction that $w_{i-1} = \max\{h_{i-1},x_{i-1}^{c_1}\}$; so, $w_i$ is adjacent to $w_{i-1}$, a contradiction. This proves that $i$ is even. Since $i \geq \ell+3$ is even, we have that $w_i = \max\{h_i,x_i^{c_1}\}$. Furthermore, $i-1$ is odd, and so $w_{i-1} \in V(H)$; since $H$ is a hyperhole, it follows that $h_i$ is adjacent to $w_{i-1}$. Since $w_i$ is nonadjacent to $w_{i-1}$, we deduce that $w_i = x_i^{c_1}$, and that $x_i^{c_1}$ is nonadjacent to $w_{i-1}$.} Let $a = c(w_{i-1})$. Suppose that $i = \ell+3$. Then by the choice of $\ell$, no vertex in $W_{i-1} = W_{\ell+2}$ is colored $c_1$. So, $a \neq c_1$, and $c_1$ is higher than $a$ in $X_{i-1} = X_{\ell+2}$. Let $Q$ be the component of $T^{c_1,a}$ that contains $w_{i-1}$. By construction, $V(Q) \cap X_i = \emptyset$, i.e.\ $V(Q) \cap X_{\ell+3} = \emptyset$; on the other hand, by the parity of $i$ and $j$, and by the fact that $c_1 \notin c(X_{j+1})$, we see that $V(Q) \cap X_{j+1} = \emptyset$. Thus, $V(Q) \subseteq X_{j+2} \cup \dots \cup X_{\ell+2}$. We now have that $s_1 \notin V(Q)$, that $\ell+2 \geq 3$ is odd, that $Q$ intersects $X_{\ell+2}$, and that $a$ is lower than $c_1$ in $X_{\ell+2}$. But this contradicts the fact that $c$ is unimprovable. Thus, $i \geq \ell+5$. By construction, $x_{i-2}^{c_1}$ is adjacent to $w_{i-1}$, and so if $c_1 \in c(X_{i-1})$, then $w_{i-1} < x_{i-1}^{c_1}$. Thus, $a \neq c_1$, and $a$ is lower than $c_1$ in $X_{i-1}$. Let $Q$ be the component of $T^{c_1,a}$ that contains $w_{i-1}$. Then $V(Q) \cap X_{j+1} = V(Q) \cap X_i = \emptyset$,\footnote{As before, the fact that $V(Q) \cap X_{j+1} = \emptyset$ follows from the parity of $i$ and $j$, and from the fact that $c_1 \notin c(X_{j+1})$. The fact that $V(Q) \cap X_i = \emptyset$ follows from the fact that $w_{i-1}$ is nonadjacent to $x_i^{c_1}$.} and we deduce that $V(Q) \subseteq X_{j+2} \cup \dots \cup X_{i-1}$. But now $s_1 \notin V(Q)$, $i-1 \geq 3$ is odd, $Q$ intersects $X_{i-1}$, and $a$ is lower than $c_1$ in $X_{i-1}$; this contradicts the fact that $c$ is unimprovable.~$\blacksquare$ \begin{quote} \begin{claim} \label{claim-W-big} $|V(W)| \geq |V(H)|+\frac{k-1}{2}$. \end{claim} \end{quote} {\em Proof of Claim~\ref{claim-W-big}.} To simplify notation, for all $i \in \{1,\dots,k\}$, we set $H_i = V(H) \cap X_i$. Recall that $S_W = \{x \in V(W) \mid x \neq t_2, c(x) = c_1\}$. We then have that $S_W \subseteq S$, that $S_W$ is a stable set in $R \setminus t_2$, and that $V(H) \cap S_W = \emptyset$. Further, recall that $|S_W| \geq \frac{k-1}{2}$. Thus, it suffices to show that $|V(H)| \leq |V(W) \setminus S_W|$. By the construction of $W$, for all indices $i \in \{1,\dots,k\}$ such that either $i \leq \ell+2$ or $i$ is even, we have that $H_i \subseteq W_i \setminus S_W$. We may now assume that for some even index $i \geq \ell+3$, we have that $|W_i \setminus (H_i \cup S_W)| < |H_{i+1} \setminus W_{i+1}|$, for otherwise we are done. Since $W_i \setminus (H_i \cup S_W)$ and $H_{i+1} \setminus W_{i+1}$ are both cliques of $R \setminus t_2$, and since $c$ is a proper coloring of $R \setminus t_2$, we have that $|c(W_i \setminus (H_i \cup S_W))| < |c(H_{i+1} \setminus W_{i+1})|$; fix $a \in c(H_{i+1} \setminus W_{i+1}) \setminus c(W_i \setminus (H_i \cup S_W))$. Then $a \neq c_1$.\footnote{This is because $a \in c(H_{i+1})$, and $c$ does not assign color $c_1$ to any vertex in $V(H) \setminus \{t_2\}$.} Furthermore, we have that $a \notin c(W_i)$,\footnote{By construction, $a \notin c(W_i \setminus (H_i \cup S_W))$, and since $a \neq c_1$, we also have that $a \notin c(S_W)$. Further, $a \in c(H_{i+1})$, and so since $H_i$ is complete to $H_{i+1}$, we have that $a \notin c(H_i)$. Thus, $a \notin c(W_i)$.} whereas by the construction of $W$, and by the fact that $i \geq \ell+3$ is even, we have that $c_1 \in c(W_i)$. It then follows from the construction of $W$ that $a$ is higher than $c_1$ in $X_i$ (possibly $a \notin c(X_i)$). Since $a \in c(H_{i+1} \setminus W_{i+1})$, we have that $x_{i+1}^a \in H_{i+1} \setminus W_{i+1}$. Since $i+1$ is odd with $i+1 \geq \ell+4$, we see from the construction of $W$ that $x_{i+1}^a$ is nonadjacent to $x_i^{c_1}$. Let $Q$ be the component of $T^{c_1,a}$ that contains $x_i^{c_1}$. Then $V(Q) \cap X_{j+1} = V(Q) \cap X_{i+1} = \emptyset$,\footnote{The fact that $V(Q) \cap X_{i+1} = \emptyset$ follows from the fact that $a$ is higher than $c_1$ in $X_i$, and $x_i^{c_1}$ is nonadjacent to $x_{i+1}^a$. The fact that $V(Q) \cap X_{j+1} = \emptyset$ follows from the parity of $i$ and $j$, and from the fact that $c_1 \notin c(X_{j+1})$.} and it follows that $V(Q) \subseteq X_{j+2} \cup \dots \cup X_i$. We now have that $s_1 \notin V(Q)$, that $i \geq 4$ is even, that $Q$ intersects $X_i$, and that $a$ is higher than $c_1$ in $X_i$. But this contradicts the fact that $c$ is unimprovable.~$\blacksquare$ \medskip By Claim~\ref{claim-W-hyperhole}, $W$ is a $k$-hyperhole; since $k$ is odd, we see that $\alpha(W) = \frac{k-1}{2}$. Using Claims~\ref{claim-R-S-hyperhole} and~\ref{claim-W-big}, we now get that \begin{displaymath} \begin{array}{ccc ccc ccc} \chi(W) & \geq & \Big\lceil \frac{|V(W)|}{\alpha(W)} \Big\rceil & = & \Big\lceil \frac{2|V(W)|}{k-1} \Big\rceil & \geq & \Big\lceil \frac{2|V(H)|}{k-1} \Big\rceil+1 & = & r+1. \end{array} \end{displaymath} On the other hand, we have that $\chi(W) \leq \chi(R) \leq r+1$, and we deduce that $\chi(W) = r+1$. This proves (b), and we are done. \end{proof} We are now ready to prove Theorem~\ref{thm-ring-hyperhole}, restated below for the reader's convenience. \begin{thm-ring-hyperhole} Let $k \geq 4$ be an integer, and let $R$ be a $k$-ring. Then $\chi(R) = \max\{\chi(H) \mid \text{$H$ is a $k$-hyperhole in $R$}\}$. \end{thm-ring-hyperhole} \begin{proof} If $k$ is even, then the result follows from Lemmas~\ref{lemma-ring-large-omega} and~\ref{lemma-even-ring-col-alg}. So from now on, we assume that $k$ is odd. Clearly, it suffices to show that $R$ contains a $k$-hyperhole of chromatic number $\chi(R)$. We assume inductively that this holds for smaller $k$-rings, i.e.\ we assume that every $k$-ring $R'$ such that $|V(R')| < |V(R)|$ contains a $k$-hyperhole of chromatic number $\chi(R')$. Let $(X_1,\dots,X_k)$ be a ring partition of $R$. For each $i\in\{1, \dots , k\}$, let $X_i = \{u_i^1, \dots, u_i^{|X_i|}\}$ be an ordering of $X_i$ such that $X_i \subseteq N_R[u_i^{|X_i|}] \subseteq \dots \subseteq N_R[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$, as in the definition of a ring. For all $i \in \{1,\dots,k\}$, set $s_i = u_i^1$ and $t_i = u_i^{|X_i|}$. Set $r = \chi(R \setminus t_2)$, and note that this implies that $r \leq \chi(R) \leq r+1$. Thus, we may assume that $R$ contains no hyperhole of chromatic number $r+1$, for otherwise we are done. Let $c$ be an unimprovable coloring of $R \setminus t_2$ that uses exactly $r$ colors (the existence of such a coloring follows from Lemma~\ref{lemma-unimprov-alg}). Let $C$ be the set of colors used by $c$ (thus, $|C| = r$), and set $c_1 = c(s_1)$ and $S = \{x \in V(R) \mid x \neq t_2, c(x) = c_1\}$. Lemma~\ref{lemma-main-technical} now implies that $\omega(R \setminus S) \leq r-1$ and $\chi(R \setminus S) \leq r-1$. Since $S$ is a stable set in $R$, we see that $\chi(R) \leq \chi(R \setminus S)+1 \leq r$; we already saw that $r \leq \chi(R) \leq r+1$, and so we deduce that $\chi(R) = r$. Further, since $\omega(R \setminus S) \leq r-1$, and since $S$ is a stable set, we see that $\omega(R) \leq r$. If $\omega(R) = r$, then $\chi(R) = \omega(R)$, and the result follows from Lemma~\ref{lemma-ring-large-omega}. Thus, we may assume that $\omega(R) \leq r-1$. Clearly, this implies that $\omega(R \setminus t_2) \leq r-1$. Since $\chi(R \setminus t_2) = r$, we have that $\omega(R \setminus t_2) < \chi(R \setminus t_2)$. Suppose that $|X_2| = 1$, i.e.\ that $X_2 = \{t_2\}$. Then by Lemma~\ref{lemma-ring-chordal}(c), $R \setminus t_2$ is chordal, and therefore (by~\cite{Berge-German, D61}) perfect. So, $\chi(R \setminus t_2) = \omega(R \setminus t_2)$, contrary to the fact that $\omega(R \setminus t_2) < \chi(R \setminus t_2)$. So, $|X_2| \geq 2$. Since every vertex in $X_2 \setminus \{t_2\}$ dominates $t_2$ in $R$, Lemma~\ref{lemma-ring-char} readily implies that $R \setminus t_2$ is a $k$-ring with ring partition $(X_1,X_2 \setminus \{t_2\},X_3,\dots,X_k)$. So, by the induction hypothesis, $R \setminus t_2$ contains a $k$-hyperhole $H$ of chromatic number $\chi(R \setminus t_2)$. But recall that $\chi(R \setminus t_2) = r = \chi(R)$. So, $H$ is a $k$-hyperhole in $R$ of chromatic number $\chi(R)$, which is what we needed. \end{proof} \section{Coloring rings} \label{sec:col} We remind the reader that $\mathcal{R}_{\geq 4}$ is the class of all graphs $G$ that have the property that every induced subgraph of $G$ either is a ring or has a simplicial vertex. By Lemma~\ref{lemma-RRk-hered}, $\mathcal{R}_{\geq 4}$ is hereditary and contains all rings. Our goal in this section is to construct a polynomial-time coloring algorithm for graphs in $\mathcal{R}_{\geq 4}$ (see Theorem~\ref{thm-ring-col-alg}), and more generally, for graphs in $\mathcal{G}_{\text{T}}$ (see Theorem~\ref{thm-GT-col-alg}). We already know how to color even rings (see Lemma~\ref{lemma-even-ring-col-alg}). In the remainder of the section, we focus primarily on odd rings. The following lemma is an easy corollary of Theorem~\ref{thm-ring-hyperhole} and Lemma~\ref{lemma-main-technical}, and it is at the heart of our coloring algorithm for odd rings. \begin{lemma} \label{lemma-reduce-chi} Let $k \geq 5$ be an odd integer, let $R$ be a $k$-ring with ring partition $(X_1,\dots,X_k)$, and for each $i \in \{1,\dots,k\}$, let $X_i = \{u_i^1,\dots,u_i^{|X_i|}\}$ be an ordering of $X_i$ such that $X_i \subseteq N_R[u_i^{|X_i|}] \subseteq \dots \subseteq N_R[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$. For all $i \in \{1,\dots,k\}$, set $s_i = u_i^1$ and $t_i = u_i^{|X_i|}$. Let $c$ be an unimprovable coloring of $R \setminus t_2$, and let $r$ be the number of colors used by $c$.\footnote{In particular, $c$ is a proper coloring of $R \setminus t_2$. Furthermore, we have that $\chi(R \setminus t_2) \leq r$, and this inequality may possibly be strict.} Let $c_1 = c(s_1)$, and let $S = \{x \in V(R) \mid x \neq t_2, c(x) = c_1\}$.\footnote{Note that $S$ is a stable set in $R \setminus t_2$.} Then either $\chi(R \setminus S) \leq r-1$ or $\chi(R) = r+1$. \end{lemma} \begin{proof} By Theorem~\ref{thm-ring-hyperhole}, the hypotheses of Lemma~\ref{lemma-main-technical}(b) are satisfied, and we deduce that either $\chi(R \setminus S) \leq r-1$, or $R$ contains a $k$-hyperhole of chromatic number $r+1$. In the former case, we are done. So assume that $R$ contains a $k$-hyperhole $H$ such that $\chi(H) = r+1$. But then \begin{displaymath} \begin{array}{ccc cc cc cc} r+1 & = & \chi(H) & \leq & \chi(R) & \leq & \chi(R \setminus t_2)+1 & \leq & r+1, \end{array} \end{displaymath} and we deduce that $\chi(R) = r+1$. \end{proof} \begin{lemma} \label{lemma-extend-optimal-coloring} There exists an algorithm with the following specifications: \begin{itemize} \item Input: All the following: \begin{itemize} \item an odd ring $R$, \item a ring partition $(X_1,\dots,X_k)$ of $R$, \item for all $i \in \{1,\dots,k\}$, an ordering $X_i = \{u_i^1,\dots,u_i^{|X_i|}\}$ of $X_i$ such that $X_i \subseteq N_R[u_i^{|X_i|}] \subseteq \dots \subseteq N_R[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$, \item a proper coloring $c$ of $R \setminus u_2^{|X_2|}$; \end{itemize} \item Output: A proper coloring of $R$ that uses at most $\max\{\chi(R),r\}$ colors, where $r$ is the number of colors used by the input coloring $c$ of $R \setminus u_2^{|X_2|}$;\footnote{Thus, the algorithm outputs a proper coloring of $R$ that is either optimal or uses no more colors than the input coloring $c$ of $R \setminus u_2^{|X_2|}$ does. Clearly, $\chi(R) \leq \chi(R \setminus t_2)+1 \leq r+1$. So, the output coloring of $R$ uses at most $\max\{\chi(R),r\} \leq r+1$ colors. Furthermore, if it uses exactly $r+1$ colors, then $\chi(R) = r+1$, and our output coloring of $R$ is optimal. However, if our output coloring of $R$ uses at most $r$ colors, then we do not know whether or not the coloring is optimal.} \item Running time: $O(n^5)$. \end{itemize} \end{lemma} \begin{proof} To simplify notation, for all $i \in \{1,\dots,k\}$, we set $s_i = u_i^1$ and $t_i = u_i^{|X_i|}$. So, $c$ is a proper coloring of $R \setminus t_2$. We may assume that $c$ uses the color set $\{1,\dots,r\}$, i.e.\ that $c:V(R) \setminus \{t_2\} \rightarrow \{1,\dots,r\}$. First, we update $c$ by running the algorithm from Lemma~\ref{lemma-unimprov-alg} and transforming it into an unimprovable coloring of $R \setminus t_2$ that uses only colors from the set $\{1,\dots,r\}$; this takes $O(n^4)$ time. We may assume that $c(s_1) = r$. Let $S = \{x \in V(R) \mid x \neq t_2, c(x) = r\}$.\footnote{Clearly, $S$ is a stable set.} Our first goal is to compute a proper coloring $\widetilde{c}$ of $R \setminus S$ that uses at most $\max\{\chi(R \setminus S),r-1\}$ colors. Then, depending on how many colors $\widetilde{c}$ uses, we will construct the needed coloring of $R$ by either extending the coloring $c$ of $R \setminus t_2$ or by extending the coloring $\widetilde{c}$ of $R \setminus S$. Let $v_1,\dots,v_t$ ($t \geq 0$) be a maximal sequence of pairwise distinct vertices of $R \setminus S$ such that for all $i \in \{1,\dots,t\}$, $v_i$ is simplicial in $R \setminus (S \cup \{v_1,\dots,v_{i-1}\})$; this sequence can be found in $O(n^3)$ time by running the algorithm from Lemma~\ref{lemma-simplicial-list-alg} with input $R \setminus S$. Suppose first that $V(R) \setminus S = \{v_1,\dots,v_t\}$. Then $v_1,\dots,v_t$ is a simplicial elimination ordering of $R \setminus S$, and we construct the coloring $\widetilde{c}$ by coloring $R \setminus S$ greedily using the ordering $v_t,\dots,v_1$. Clearly, $\widetilde{c}$ uses only $\omega(R \setminus S)$ colors, and we have that $\omega(R \setminus S) \leq \chi(R \setminus S) \leq \max\{\chi(R \setminus S),r-1\}$. Suppose now that $V(R) \setminus S \neq \{v_1,\dots,v_t\}$. Set $R' := R \setminus (S \cup \{v_1,\dots,v_t\})$. The maximality of $v_1,\dots,v_t$ guarantees that $R'$ has no simplicial vertices, and so it follows from Lemma~\ref{lemma-ring-simplicial} that $R'$ is a $k$-ring with ring partition $\Big(X_1 \cap V(R'),\dots,X_k \cap V(R')\Big)$. Let $c' = c \upharpoonright (V(R') \setminus \{t_2\})$, and note that $c'$ uses only colors from the set $\{1,\dots,r-1\}$. If $t_2 \notin V(R')$, then we set $c'' := c'$. On the other hand, if $t_2 \in V(R')$, then we make a recursive call to the algorithm with input $R'$ and $c'$,\footnote{We also input the ring partition $\Big(X_1 \cap V(R'),\dots,X_k \cap V(R')\Big)$ of $R'$, and well as the orderings of the sets $X_i \cap V(R')$ inherited from our input orderings of the sets $X_i$.} and we obtain a proper coloring $c''$ of $R'$ that uses at most $\max\{\chi(R'),r-1\}$ colors. So, in either case (i.e.\ independently of whether $t_2$ does or does not belong to $V(R')$), we have now obtained a proper coloring $c''$ of $R'$ that uses at most $\max\{\chi(R'),r-1\}$ colors. We now extend $c''$ to a proper coloring $\widetilde{c}$ of $R \setminus S$ by assigning colors greedily to the vertices $v_t,\dots,v_1$ (in that order). Note that the coloring $\widetilde{c}$ uses at most $\max\{\omega(R \setminus S),\chi(R'),r-1\} \leq \max\{\chi(R \setminus S),r-1\}$ colors. In either case,\footnote{That is: both in the case when $V(R) \setminus S = \{v_1,\dots,v_t\}$ and in the case when $V(R) \setminus S \neq \{v_1,\dots,v_t\}$.} we have constructed a proper coloring $\widetilde{c}$ of $R \setminus S$ that uses at most $\max\{\chi(R \setminus S),r-1\}$ colors. If $\widetilde{c}$ uses at most $r-1$ colors, then we extend $\widetilde{c}$ to a proper coloring of $R$ that uses at most $r$ colors by assigning the same new color to all the vertices of the stable set $S$; we then return this coloring of $R$, and we stop. Suppose now that the coloring $\widetilde{c}$ uses at least $r$ colors. Then $\chi(R \setminus S) \geq r$, and so Lemma~\ref{lemma-reduce-chi} implies that $\chi(R) = r+1$. We now extend the coloring $c$ of $R \setminus t_2$ to a proper coloring of $R$ by assigning color $r+1$ to the vertex $t_2$. Our coloring of $R$ uses at most $r+1 = \chi(R)$ colors,\footnote{So, in fact, our coloring of $R$ uses exactly $\chi(R)$ colors, and it is therefore optimal.} we return this coloring, and we stop. Clearly, the algorithm is correct. We make $O(n)$ recursive calls, and otherwise, the slowest step of the algorithm takes $O(n^4)$ time. Thus, the total running time of the algorithm is $O(n^5)$. \end{proof} \begin{theorem} \label{thm-ring-col-alg} There exists an algorithm with the following specifications: \begin{itemize} \item Input: A graph $G$; \item Output: Either an optimal coloring of $G$, or the true statement that $G \notin \mathcal{R}_{\geq 4}$; \item Running time: $O(n^6)$. \end{itemize} \end{theorem} \begin{proof} First, we form a maximal sequence $v_1,\dots,v_t$ ($t \geq 0$) of pairwise distinct vertices of $G$ such that, for all $i \in \{1,\dots,t\}$, $v_i$ is simplicial in $G \setminus \{v_1,\dots,v_{i-1}\}$; this can be done in $O(n^3)$ time by running the algorithm from Lemma~\ref{lemma-simplicial-list-alg} with input $G$. Suppose first that $t \geq 1$. If $V(G) = \{v_1,\dots,v_t\}$, so that $v_1,\dots,v_t$ is a simplicial elimination ordering of $G$, then we color $G$ greedily in $O(n^2)$ time using the ordering $v_t,\dots,v_1$; clearly, the resulting coloring of $G$ is optimal, we return this coloring, and we stop. So assume that $V(G) \setminus \{v_1,\dots,v_t\} \neq \emptyset$. We then make a recursive call to the algorithm with input $G \setminus \{v_1,\dots,v_t\}$. If we obtain an optimal coloring of $G \setminus \{v_1,\dots,v_t\}$, then we greedily extend this coloring to an optimal coloring of $G$ using the ordering $v_t,\dots,v_1$, we return this coloring of $G$, and we stop. On the other hand, if the algorithm returns the statement that $G \setminus \{v_1,\dots,v_t\} \notin \mathcal{R}_{\geq 4}$, then we return the answer that $G \notin \mathcal{R}_{\geq 4}$ (this is correct because $R_{\geq 4}$ is hereditary), and we stop. From now on, we assume that $t = 0$. Thus, $G$ contains no simplicial vertices, and so by the definition of $\mathcal{R}_{\geq 4}$, either $G$ is a ring, or $G \notin \mathcal{R}_{\geq 4}$. We now run the algorithm from Lemma~\ref{lemma-detect-ring-ord} input $G$; this takes $O(n^2)$ time. If the algorithm returns the answer that $G$ is not a ring, then we return the answer that $G \notin \mathcal{R}_{\geq 4}$. So assume the algorithm returned the statement that $G$ is a ring, along with the length $k$ and ring partition $(X_1,\dots,X_k)$ of $G$, and for each $i \in \{1,\dots,k\}$ an ordering $X_i = \{u_i^1,\dots,u_i^{|X_i|}\}$ of $X_i$ such that $X_i \subseteq N_G[u_i^{|X_i|}] \subseteq \dots \subseteq N_G[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$. If $k$ is even, then we obtain an optimal coloring of $G$ in $O(n^3)$ time by running the algorithm from Lemma~\ref{lemma-even-ring-col-alg}, we return this coloring, and we stop. So from now on, we assume that $k$ is odd, so that $G$ is an odd ring. For each $i \in \{1,\dots,k\}$, we set $t_i = u_i^{|X_i|}$. Since $G$ is a ring, Lemma~\ref{lemma-RRk-hered} guarantees that $G \in \mathcal{R}_{\geq 4}$, and since $\mathcal{R}_{\geq 4}$ is hereditary, we see that $G \setminus t_2$ belongs to $\mathcal{R}_{\geq 4}$. We now obtain an optimal coloring $c$ of $G \setminus t_2$ by making a recursive call to the algorithm. We then call the algorithm from Lemma~\ref{lemma-extend-optimal-coloring} with input $G$ and $c$,\footnote{We also input the ring partition $(X_1,\dots,X_k)$ of the ring $G$, as well as our orderings of the sets $X_1,\dots,X_k$.} and we obtain a proper coloring of $G$ that uses at most $\max\{\chi(G),\chi(G \setminus t_2)\} = \chi(G)$ colors;\footnote{Clearly, this coloring of $G$ is optimal.} this takes $O(n^5)$ time. We now return this coloring of $G$, and we stop. Clearly, the algorithm is correct. We make $O(n)$ recursive calls to the algorithm, and otherwise, the slowest step of the algorithm takes $O(n^5)$ time. Thus, the total running time of the algorithm is $O(n^6)$. \end{proof} We complete this section by giving a polynomial-time coloring algorithm for graphs in $\mathcal{G}_{\text{T}}$. \begin{theorem} \label{thm-GT-col-alg} There exists an algorithm with the following specifications: \begin{itemize} \item Input: A graph $G$; \item Output: Either an optimal coloring of $G$, or the true statement that $G \notin \mathcal{G}_{\text{T}}$; \item Running time: $O(n^7)$. \end{itemize} \end{theorem} \begin{proof} We first check whether $G$ has a clique-cutset, and if so, we obtain a clique-cut-partition $(A,B,C)$ of $G$ such that $G[A \cup C]$ does not admit a clique-cutset; this can be done in $O(n^3)$ time by running the algorithm from~\cite{Tarjan} with input $G$. If we obtained the answer that $G$ does not admit a clique-cutset, then we set $A = V(G)$, $B = \emptyset$, and $C = \emptyset$. On the other hand, if we obtained $(A,B,C)$, then we make a recursive call to the algorithm with input $G[B \cup C]$; if we obtained the answer that $G[B \cup C] \notin \mathcal{G}_{\text{T}}$, then we return the answer that $G \notin \mathcal{G}_{\text{T}}$ (this is correct because $\mathcal{G}_{\text{T}}$ is hereditary), and we stop. So from now on, we assume that one of the following holds: \begin{itemize} \item $B = C = \emptyset$; \item $(A,B,C)$ is a clique-cut-partition of $G$, and we recursively obtained an optimal coloring $c_B$ of $G[B \cup C]$. \end{itemize} In either case, we also have that $G[A \cup C]$ does not admit a clique-cutset. We now run the algorithm from Theorem~\ref{thm-ring-col-alg} with input $G[A \cup C]$; this takes $O(n^6)$ time. The algorithm either returns an optimal coloring $c_A$ of $G[A \cup C]$, or it returns the answer that $G[A \cup C] \notin \mathcal{R}_{\geq 4}$. If the algorithm returned the answer that $G[A \cup C] \notin \mathcal{R}_{\geq 4}$, then our goal is to either produce an optimal coloring $c_A$ of $G[A \cup C]$ in another way, or to determine that $G \notin \mathcal{G}_{\text{T}}$. In this case (i.e.\ if the algorithm returned the answer that $G[A \cup C] \notin \mathcal{R}_{\geq 4}$), we proceed as follows. Since $\mathcal{R}_{\geq 4}$ contains all rings (by Lemma~\ref{lemma-RRk-hered}), we have that $G[A \cup C]$ is not a ring. Recall that $G[A \cup C]$ does not admit a clique-cutset. Thus, Theorem~\ref{thm-GT-decomp} implies that either $G[A \cup C]$ is a complete graph, or $G[A \cup C]$ is a 7-hyperantihole, or $G[A \cup C] \notin \mathcal{G}_{\text{T}}$ (in which case, $G \notin \mathcal{G}_{\text{T}}$, since $\mathcal{G}_{\text{T}}$ is hereditary). Clearly, complete graphs have stability number one, and hyperantiholes have stability number two. Thus, either $\alpha(G[A \cup C]) \leq 2$ or $G \notin \mathcal{G}_{\text{T}}$. We determine whether $\alpha(G[A \cup C]) \leq 2$ by examining all triples of vertices in $G[A \cup C]$; this takes $O(n^3)$ time. If $\alpha(G[A \cup C]) \geq 3$, then we return the answer that $G \notin \mathcal{G}_{\text{T}}$, and we stop. So suppose that $\alpha(G[A \cup C]) \leq 2$. This means that each color class of a proper coloring of $G[A \cup C]$ is of size at most two, and that, taken together, color classes of size exactly two correspond to a matching of $\overline{G}[A \cup C]$ (the complement of $G[A \cup C]$). So, we form the graph $\overline{G}[A \cup C]$ in $O(n^2)$ time, and we find a maximum matching $M$ in $\overline{G}[A \cup C]$ in $O(n^4)$ time by running the algorithm from~\cite{matching}. We now color $G[A \cup C]$ as follows: each member of $M$ is a two-vertex color class,\footnote{By construction, members of $M$ are edges of $\overline{G}[A \cup C]$; consequently, members of $M$ are stable sets of size two in $G[A \cup C]$.} and each vertex in $A \cup C$ that is not an endpoint of any member of $M$ forms a one-vertex color class.\footnote{So, in total, we used $|M|+(|A \cup C|-2|M|) = |A \cup C|-|M|$ colors.} This produces an optimal coloring $c_A$ of $G[A \cup C]$. So from now on, we may assume that we have obtained an optimal coloring $c_A$ of $G[A \cup C]$. If $B = C = \emptyset$, then $c_A$ is in fact an optimal coloring of $G$; in this case, we return $c_A$, and we stop. So assume that $B \cup C \neq \emptyset$. Then we have already obtained an optimal coloring $c_B$ of $G[B \cup C]$. After possibly renaming colors, we may assume that the color set used by one of $c_A,c_B$ is included in the color set used by the other one. Now, $C$ is a clique in $G$, and so $c_A$ assigns a different color to each vertex of $C$, and the same is true for $c_B$. So, after possibly permuting colors, we may assume that $c_A$ and $c_B$ agree on $C$, i.e.\ that $c_A \upharpoonright C = c_B \upharpoonright C$. Now $c := c_A \cup c_B$ is an optimal coloring of $G$. We return $c$, and we stop. Clearly, the algorithm is correct. We make $O(n)$ recursive calls to the algorithm, and otherwise, the slowest step takes $O(n^6)$ time. Thus, the total running time of the algorithm is $O(n^7)$. \end{proof} \section{Computing the chromatic number of a ring} \label{sec:chi} In section~\ref{sec:col}, we gave an $O(n^6)$ time coloring algorithm for graphs in $\mathcal{R}_{\geq 4}$ (see Theorem~\ref{thm-ring-col-alg}),\footnote{Recall that, by Lemma~\ref{lemma-RRk-hered}, the class $\mathcal{R}_{\geq 4}$ contains all rings.} and we also gave an $O(n^7)$ time coloring algorithm for graphs in $\mathcal{G}_{\text{T}}$ (see Theorem~\ref{thm-GT-col-alg}). These algorithms produce optimal colorings of input graphs from the specified classes; however, for some applications, it is enough to compute the chromatic number, without actually finding an optimal coloring of the input graph. In this section, we use Corollary~\ref{cor-ring-chi-formula} and Lemma~\ref{lemma-omega-ring} to construct an $O(n^3)$ time algorithm that computes the chromatic number of a graph in $\mathcal{R}_{\geq 4}$ (see Theorem~\ref{thm-ring-chi-alg}), and using this result, we construct an $O(n^5)$ time algorithm that computes the chromatic number of a graph in $\mathcal{G}_{\text{T}}$ (see Theorem~\ref{thm-GT-chi-alg}). First, we give an $O(n^3)$ time algorithm that computes a maximum hyperhole in a ring (see Lemma~\ref{lemma-normal-hyperhole-alg}).\footnote{We remind the reader that, by Lemma~\ref{lemma-ring-chordal}(b), every hyperhole in a ring is of the same length as that ring. As usual, a {\em maximum hyperhole} in a ring is a hyperhole of maximum size in that ring.} We begin with some terminology and notation. Let $k \geq 4$ be an integer, let $R$ be a $k$-ring with ring partition $(X_1,\dots,X_k)$, and for all $i \in \{1,\dots,k\}$, let $X_i = \{u_i^1,\dots,u_i^{|X_i|}\}$ be an ordering of $X_i$ such that $X_i \subseteq N_G[u_i^{|X_i|}] \subseteq \dots \subseteq N_G[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$, as in the definition of a ring. Let $H$ be a hyperhole in $R$. By Lemma~\ref{lemma-ring-chordal}, the hyperhole $H$ is of length $k$, and it intersects each of the sets $X_1,\dots,X_k$. For all $i \in \{1,\dots,k\}$, let $\ell_i = \max\{\ell \in \{1,\dots,|X_i|\} \mid u_i^{\ell} \in V(H)\}$ and $Y_i = \{u_i^1,\dots,u_i^{\ell_i}\}$. Finally, let $\widetilde{H} = R[Y_1 \cup \dots \cup Y_k]$ and $C_H = \{u_1^{\ell_1},\dots,u_k^{\ell_k}\}$. Clearly, $\widetilde{H}$ is a hyperhole, with $V(H)\subseteq V(\widetilde{H})$. Furthermore, $C_H$ induces a hole in $R$, and it uniquely determines $\widetilde{H}$. We say that $H$ is \emph{normal} in $R$ if $H = \widetilde{H}$. Clearly, any maximal hyperhole (and therefore, any hyperhole of maximum size) in $R$ is normal. Thus, to find a maximum hyperhole in an input ring, we need only consider normal hyperholes in that ring. \begin{lemma} \label{lemma-normal-hyperhole-alg} There exists an algorithm with the following specifications: \begin{itemize} \item Input: A graph $R$; \item Output: Either a maximum hyperhole $H$ in $R$, or the true statement that $R$ is not a ring; \item Running time: $O(n^3)$. \end{itemize} \end{lemma} \begin{proof} We first run the algorithm from Lemma~\ref{lemma-detect-ring-ord} with input $R$; this takes $O(n^2)$ time. If the algorithm returns the answer that $R$ is not a ring, then we return that answer as well and stop. So assume the algorithm returned the statement that $R$ is a ring, along with the length $k$ and ring partition $(X_1,\dots,X_k)$ of $R$, and for each $i \in \{1,\dots,k\}$ an ordering $X_i = \{u_i^1,\dots,u_i^{|X_i|}\}$ of $X_i$ such that $X_i \subseteq N_R[u_i^{|X_i|}] \subseteq \dots \subseteq N_R[u_i^1] = X_{i-1} \cup X_i \cup X_{i+1}$. For each $j \in \{1,\dots,|X_1|\}$, we will find a normal hyperhole $H_j$ of $R$ such that $V(H_j) \cap X_1 = \{u_1^1,\dots,u_1^j\}$, and subject to that, $|V(H_j)|$ is maximum. We will then compare the sizes of all the $H_j$'s (with $1 \leq j \leq |X_1|$), and we will return an $H_j$ of maximum size. We begin by constructing an auxiliary weighted digraph $(D,w)$,\footnote{The weight function $w$ will assign nonnegative interger weights to the arcs of $D$.} as follows. First, we construct a set of $|X_1|$ new vertices, $X_{k+1} = \{u_{k+1}^1,\dots,u_{k+1}^{|X_1|}\}$, with $X_{k+1} \cap V(R) = \emptyset$.\footnote{So, $|X_{k+1}| = |X_1|$, and we think of $X_{k+1}$ as a copy of $X_1$.} Let ${D}$ be the digraph with vertex set $V({D}) = V(R) \cup X_{k+1}$ and arc set: \begin{displaymath} \begin{array}{rcl} A({D}) & = & \bigcup\limits_{i=1}^{k-1} \Big(\Big\{\overrightarrow{x y} \mid x\in X_i, \ y\in X_{i+1}, \ xy \in E(R)\Big\} \\ & & \phantom{iiiii} \cup \Big\{\overrightarrow{x u_{k+1}^{\ell}} \mid x \in X_k, \ x u_1^{\ell} \in E(R), \ 1 \leq \ell \leq |X_1|\Big\}\Big). \end{array} \end{displaymath} Finally, for every arc $\overrightarrow{u_i^pu_{i+1}^q}$ in $A({D})$, with $i \in \{1, \dots, k\}$, $p \in \{1, \dots, |X_i|\}$, and $q \in \{1, \dots, |X_{i+1}|\}$, we set $w(\overrightarrow{u_i^pu_{i+1}^q}) = (|X_i|-p)+(|X_{i+1}|-q)$. Now, for a fixed $j \in \{1,\dots,|X_1|\}$, we find the hyperhole $H_j$ as follows. Let ${P_j}$ be a minimum weight directed path between $u_1^j$ and $u_{k+1}^j$ in the weighted digraph $({D}, w)$. Such a path can be found in $O(n^2)$ time using Dijkstra's algorithm~\cite{D59, Sch}. For each $i \in \{1, \dots, k\}$, let $\ell_{i, j} \in \{1, \dots, |X_i|\}$ be the (unique) index such that $u_i^{\ell_{i, j}} \in V({P_j})$, and let $Y_{i, j} = \{u_i^{\ell} \mid 1 \leq \ell \leq \ell_{i, j}\}$. Finally, let $H_j = R[Y_{1, j}\cup\dots\cup Y_{k, j}]$. Clearly, $H_j$ is a normal hyperhole of $R$, and $V(H_j) \cap X_1 = \{u_1^1,\dots,u_1^j\}$. Moreover, we have that \begin{displaymath} \begin{array}{rcl} |V(H_j)| & = & \sum\limits_{i=1}^k |Y_{i, j}| \\ \\ & = & \sum\limits_{i=1}^k \ell_{i, j} \\ \\ & = & |V(R)|-\sum\limits_{i=1}^k (|X_i|-\ell_{i, j}) \\ \\ & = & |V(R)|-\frac{1}{2}w({P_j}), \end{array} \end{displaymath} and so the fact that $P_j$ has minimum weight implies that $H_j$ has maximum size among all hyperholes $H$ in $R$ that satisfy $V(H) \cap X_1 = \{u_1^1,\dots,u_1^j\}$. So, $H_j$ is the desired hyperhole for a given $j$. We now compare the sizes of the hyperholes $H_1,\dots,H_{|X_1|}$ (this takes $O(n^2)$ time), and we return one of maximum size. Clearly, the algorithm is correct. The total running time is $O(n^3)$, since computing $H_j$ (for fixed $j$) takes $O(n^2)$ time, and we do this for $O(n)$ values of $j$. \end{proof} We now give a polynomial-time algorithm that computes the chromatic number of graphs in $\mathcal{R}_{\geq 4}$. \begin{theorem} \label{thm-ring-chi-alg} There exists an algorithm with the following specifications: \begin{itemize} \item Input: A graph $G$; \item Output: Either $\chi(G)$, or the true statement that $G \notin \mathcal{R}_{\geq 4}$; \item Running time: $O(n^3)$. \end{itemize} \end{theorem} \begin{proof} First, we form a maximal sequence $v_1,\dots,v_t$ ($t \geq 0$) of pairwise distinct vertices of $G$ such that, for all $i \in \{1,\dots,t\}$, $v_i$ is simplicial in $G \setminus \{v_1,\dots,v_{i-1}\}$; this can be done in $O(n^3)$ time by calling the algorithm from Lemma~\ref{lemma-simplicial-list-alg} with input $G$. Suppose first that $V(G) = \{v_1,\dots,v_t\}$, so that $v_1,\dots,v_t$ is a simplicial elimination ordering of $G$. In this case, we greedily color $G$ using the ordering $v_t,\dots,v_1$,\footnote{Clearly, this produces an optimal coloring of $G$.} we return the number of colors that we used, and we stop; this takes $O(n^2)$ time. From now on, we assume that $V(G) \neq \{v_1,\dots,v_t\}$, and we form the graph $R := G \setminus \{v_1,\dots,v_t\}$ in $O(n^2)$ time. The maximality of $v_1,\dots,v_t$ guarantees that $R$ contains no simplicial vertices, and so by the definition of $\mathcal{R}_{\geq 4}$, we have that either $R$ is a ring, or $G \notin \mathcal{R}_{\geq 4}$.\footnote{Indeed, suppose that $G \in \mathcal{R}_{\geq 4}$. Since $\mathcal{R}_{\geq 4}$ is hereditary, it follows that $R \in \mathcal{R}_{\geq 4}$. By the maximality of $v_1,\dots,v_t$, the graph $R$ has no simplicial vertices. So, by the definition of $\mathcal{R}_{\geq 4}$, we have that $R$ is a ring.} We now run the algorithm from Lemma~\ref{lemma-detect-ring} with input $R$; this takes $O(n^2)$ time. If the algorithm returns the answer that $R$ is not a ring, then we return the answer that $G \notin \mathcal{R}_{\geq 4}$, and we stop. So assume the algorithm returned the statement that $R$ is a ring, along with the length $k$ and ring partition $(X_1,\dots,X_k)$ of $R$. Next, we call the algorithm from Lemma~\ref{lemma-omega-ring}; this takes $O(n^3)$ time. Since $R$ is a ring, Lemma~\ref{lemma-ring-chordal}(d) guarantees that $R \in \mathcal{G}_{\text{T}}$, and so the algorithm returns $\omega(R)$. Next, we run the algorithm from Lemma~\ref{lemma-normal-hyperhole-alg} with input $R$; this takes $O(n^3)$ time. Since $R$ is a ring, we know that the algorithm returns a hyperhole $H$ of $R$ of maximum size; since $R$ is a $k$-ring, Lemma~\ref{lemma-ring-chordal}(b) guarantees that $H$ is a $k$-hyperhole. Set $r := \max\Big\{\omega(R),\Big\lceil \frac{|V(H)|}{\lfloor k/2 \rfloor} \Big\rceil\Big\}$; by Corollary~\ref{cor-ring-chi-formula}, we have that $\chi(R) = r$. If $t = 0$ (so that $G = R$), then we return $r$, and we stop. So assume that $t \geq 1$. For each $i \in \{1,\dots,t\}$, set $r_i = |N_G[v_i] \setminus \{v_1,\dots,v_{i-1}\}|$; computing the constants $r_1,\dots,r_t$ takes $O(n^2)$ time. An easy induction using Lemma~\ref{lemma-simplicial-chi} now establishes that $\chi(G) = \max\{r_1,\dots,r_t,r\}$. So, we return $\max\{r_1,\dots,r_t,r\}$, and we stop. Clearly, the algorithm is correct, and its running time is $O(n^3)$. \end{proof} We complete this section by showing how to compute the chromatic number of graphs in $\mathcal{G}_{\text{T}}$ in polynomial time. We remark that this algorithm is very similar to the one from Theorem~\ref{thm-GT-col-alg}, except that we use Theorem~\ref{thm-ring-chi-alg} instead of Theorem~\ref{thm-ring-col-alg}. Nevertheless, for the sake of completeness, we give all the details. \begin{theorem} \label{thm-GT-chi-alg} There exists an algorithm with the following specifications: \begin{itemize} \item Input: A graph $G$; \item Output: Either $\chi(G)$, or the true statement that $G \notin \mathcal{G}_{\text{T}}$; \item Running time: $O(n^5)$. \end{itemize} \end{theorem} \begin{proof} We first check whether $G$ has a clique-cutset, and if so, we obtain a clique-cut-partition $(A,B,C)$ of $G$ such that $G[A \cup C]$ does not admit a clique-cutset; this can be done by running the algorithm from~\cite{Tarjan} with input $G$, and it takes $O(n^3)$ time. If we obtained the answer that $G$ does not admit a clique-cutset, then we set $A = V(G)$, $B = \emptyset$, and $C = \emptyset$, and we set $r = 0$. On the other hand, if we obtained $(A,B,C)$, then we make a recursive call to the algorithm with input $G[B \cup C]$; if we obtained the answer that $G[B \cup C] \notin \mathcal{G}_{\text{T}}$, then we return the answer that $G \notin \mathcal{G}_{\text{T}}$ and stop,\footnote{This is correct because $\mathcal{G}_{\text{T}}$ is hereditary.} and otherwise (i.e.\ if we obtained the chromatic number of $G[B \cup C]$) we set $r = \chi(G[B \cup C])$. We may now assume that we have obtained the number $r$ (for otherwise, we terminated the algorithm). Clearly, $\chi(G) = \max\{\chi(G[A \cup C]),r\}$. Next, we run the algorithm from Theorem~\ref{thm-ring-chi-alg} with input $G[A \cup C]$; this takes $O(n^3)$ time. If the algorithm returned $\chi(G[A \cup C])$, then we return the number $\max\{\chi(G[A \cup C]),r\}$, and we stop. So assume the algorithm returned the answer that $G[A \cup C]$ is not a ring. So far, we know that $G[A \cup C]$ does not admit a clique-cutset and is not a ring. Theorem~\ref{thm-GT-decomp} now guarantees that either $G[A \cup C]$ is a complete graph or a 7-hyperantihole, or $G[A \cup C] \notin \mathcal{G}_{\text{T}}$ (in which case, $G \notin \mathcal{G}_{\text{T}}$, since $\mathcal{G}_{\text{T}}$ is hereditary). Clearly, complete graphs have stability number one, and hyperantiholes have stability number two. Thus, either $\alpha(G[A \cup C]) \leq 2$ or $G \notin \mathcal{G}_{\text{T}}$. Now, we determine whether $\alpha(G[A \cup C]) \leq 2$ by examining all triples of vertices in $G[A \cup C]$; this takes $O(n^3)$ time. If $\alpha(G[A \cup C]) \geq 3$, then we return the answer that $G \notin \mathcal{G}_{\text{T}}$ and stop. Assume now that $\alpha(G[A \cup C]) \leq 2$. Then we form the graph $\overline{G}[A \cup C]$ (the complement of $G[A \cup C]$) in $O(n^2)$ time, and we find a maximum matching $M$ in $\overline{G}[A \cup C]$ by running the algorithm from~\cite{matching}; this takes $O(n^4)$ time. Since $\alpha(G[A \cup C]) \leq 2$, we see that $\chi(G[A \cup C]) = |A \cup C|-|M|$; we now return the number $\max\{|A \cup C|-|M|,r\}$, and we stop. Clearly, the algorithm is correct. The slowest step takes $O(n^4)$ time, and we make $O(n)$ recursive calls. Thus, the total running time of the algorithm is $O(n^5)$. \end{proof} \section{Optimal $\chi$-bounding functions} \label{sec:chibound} For all integers $k \geq 4$, we let $\mathcal{H}_k$ be the class of all induced subgraphs of $k$-hyperholes, and we let $\mathcal{A}_k$ be the class of all induced subgraphs of $k$-hyperantiholes; clearly, classes $\mathcal{H}_k$ and $\mathcal{A}_k$ are both hereditary, and they contain all complete graphs.\footnote{The reason we emphasize that these classes contain all complete graphs is that we defined optimal $\chi$-bounding functions only for hereditary, $\chi$-bounded classes that contain all complete graphs.} Recall that for all integers $k \geq 4$, $\mathcal{R}_k$ is the class of all graphs $G$ that have the property that every induced subgraph of $G$ either is a $k$-ring or has a simplicial vertex; clearly, $\mathcal{R}_k$ is hereditary and contains all complete graphs, and by Lemma~\ref{lemma-RRk-hered}, all $k$-rings belong to $\mathcal{R}_k$ (in particular, $\mathcal{H}_k \subseteq \mathcal{R}_k$). In this section, for all integers $k \geq 4$, we find the optimal $\chi$-bounding functions for the classes $\mathcal{H}_k$ (see Theorem~\ref{thm-HHk-chi-optimal}), $\mathcal{A}_k$ (see Theorem~\ref{thm-AAk-chi-optimal}), and $\mathcal{R}_k$ (see Theorem~\ref{thm-RRk-chi-optimal}). Further, for all integers $k \geq 4$, we set $\mathcal{H}_{\geq k} = \bigcup\limits_{i=k}^{\infty} \mathcal{H}_i$ and $\mathcal{A}_{\geq k} = \bigcup\limits_{i=k}^{\infty} \mathcal{A}_i$, and we remind the reader that $\mathcal{R}_{\geq k} = \bigcup\limits_{i=k}^{\infty} \mathcal{R}_i$.\footnote{Clearly, for all integers $k \geq 4$ we have that: $\mathcal{H}_{\geq k}$ is the class of all induced subgraphs of hyperholes of length at least $k$; $\mathcal{A}_{\geq k}$ is the class of all induced subgraphs of hyperantiholes of length at least $k$; and $\mathcal{R}_{\geq k}$ contains all induced subgraphs of rings of length at least $k$. In particular, $\mathcal{H}_{\geq k} \subseteq \mathcal{R}_{\geq k}$. It is clear that $\mathcal{H}_{\geq k}$, $\mathcal{A}_{\geq k}$, and $\mathcal{R}_{\geq k}$ are hereditary and contain all complete graphs.} For all integers $k \geq 4$, we find the optimal $\chi$-bounding functions for the classes $\mathcal{H}_{\geq k}$ (see Corollary~\ref{cor-HHk-chi-optimal}), $\mathcal{A}_{\geq k}$ (see Corollary~\ref{cor-AAk-chi-optimal}), and $\mathcal{R}_{\geq k}$ (see Corollary~\ref{cor-RRk-chi-optimal}); see also Theorem~\ref{thm-HH-AA-RR-chi-optimal}. Finally, we find the optimal $\chi$-bounding function for the class $\mathcal{G}_{\text{T}}$ (see Theorem~\ref{thm-GT-chi}). Recall that $\mathbb{N}$ is the set of all positive integers, and let $i_\mathbb{N}$ be the identity function on $\mathbb{N}$, i.e.\ let $i_{\mathbb{N}}:\mathbb{N} \rightarrow \mathbb{N}$ be given by $i_{\mathbb{N}}(n) = n$ for all $n \in \mathbb{N}$. We define the function $f_{\text{T}}:\mathbb{N} \rightarrow \mathbb{N}$ by setting \begin{displaymath} \begin{array}{lll} f_{\text{T}}(n) & = & \left\{\begin{array}{lll} \lfloor 5n/4 \rfloor & \text{if} & \text{$n \equiv 0,1$ (mod $4$)} \\ \\ \lceil 5n/4 \rceil & \text{if} & \text{$n \equiv 2,3$ (mod $4$)} \end{array}\right. \end{array} \end{displaymath} for all $n \in \mathbb{N}$. For all odd integers $k \geq 5$, we define the function $f_k:\mathbb{N} \rightarrow \mathbb{N}$ by setting \begin{displaymath} \begin{array}{lll} f_k(n) & = & \left\{\begin{array}{lll} \Big\lfloor \frac{kn}{k-1} \Big\rfloor & \text{if} & \text{$n \equiv 0,1$ (mod $k-1$)} \\ \\ \Big\lceil \frac{kn}{k-1} \Big\rceil & \text{if} & \text{$n \equiv 2,\dots,k-2$ (mod $k-1$)} \end{array}\right. \end{array} \end{displaymath} for all $n \in \mathbb{N}$. For all odd integers $k \geq 5$, we define the function $g_k:\mathbb{N} \rightarrow \mathbb{N}$ by setting \begin{displaymath} \begin{array}{lll} g_k(n) & = & \left\{\begin{array}{lll} \Big\lfloor \frac{kn}{k-1} \Big\rfloor & \text{if} & \text{$n \equiv 0,\dots,\frac{k-3}{2}$ (mod $k-1$)} \\ \\ \Big\lceil \frac{kn}{k-1} \Big\rceil & \text{if} & \text{$n \equiv \frac{k-1}{2},\dots,k-2$ (mod $k-1$)} \end{array}\right. \end{array} \end{displaymath} for all $n \in \mathbb{N}$. Note that $f_{\text{T}} = f_5 = g_5$. Before turning to the classes mentioned at the beginning of this section, we prove a few technical lemmas concerning functions $f_{\text{T}}$, $f_k$, and $g_k$. \begin{lemma} \label{lemma-fk-calculation} Let $k \geq 5$ be an odd integer, and let $n \in \mathbb{N}$. Then $f_k(n) = n+\Big\lceil \frac{2 \lfloor n/2 \rfloor}{k-1} \Big\rceil$. \end{lemma} \begin{proof} Set $m = \lfloor \frac{n}{k-1} \rfloor$ and $\ell = n-(k-1)m$. Clearly, $m$ is a nonnegative integer, $\ell \in \{0,\dots,k-2\}$, $n = (k-1)m+\ell$, and $n \equiv \ell$ (mod $k-1$). Since $k$ is odd, we have that $k-1$ is even, and so \begin{displaymath} \begin{array}{ccc cc} \Big\lceil \frac{2 \lfloor n/2 \rfloor}{k-1} \Big\rceil & = & \Bigg\lceil \frac{2 \Big\lfloor \frac{(k-1)m+\ell}{2} \Big\rfloor}{k-1} \Bigg\rceil & = & m+\Big\lceil \frac{2\lfloor \ell/2 \rfloor}{k-1} \Big\rceil. \end{array} \end{displaymath} If $0 \leq \ell \leq 1$, then $f_k(n) = \lfloor \frac{kn}{k-1} \rfloor$, and we have that \begin{displaymath} \begin{array}{rcl} n+\Big\lceil \frac{2 \lfloor n/2 \rfloor}{k-1} \Big\rceil & = & n+m+\Big\lceil \frac{2\lfloor \ell/2 \rfloor}{k-1} \Big\rceil \\ \\ & = & n+m \\ \\ & = & \lfloor \frac{kn}{k-1} \rfloor \\ \\ & = & f_k(n), \end{array} \end{displaymath} and we are done. Suppose now that $2 \leq \ell \leq k-2$; then $f_k(n) = \lceil \frac{kn}{k-1} \rceil$. First, we have that \begin{displaymath} \begin{array}{ccc cccc} n+\Big\lceil \frac{2 \lfloor n/2 \rfloor}{k-1} \Big\rceil & = & n+m+\Big\lceil \frac{2\lfloor \ell/2 \rfloor}{k-1} \Big\rceil & = & n+m+1 & = & \lfloor \frac{kn}{k-1} \rfloor+1. \end{array} \end{displaymath} Since $\ell \neq 0$, we see that $\frac{kn}{k-1}$ is not an integer, and so $\lfloor \frac{kn}{k-1} \rfloor+1 = \lceil \frac{kn}{k-1} \rceil$. It now follows that \begin{displaymath} \begin{array}{ccc cccc} n+\Big\lceil \frac{2 \lfloor n/2 \rfloor}{k-1} \Big\rceil & = & \lfloor \frac{kn}{k-1} \rfloor+1 & = & \lceil \frac{kn}{k-1} \rceil & = & f_k(n), \end{array} \end{displaymath} which is what we needed. This completes the argument. \end{proof} \begin{lemma} \label{lemma-gk-calculation} Let $k \geq 5$ be an odd integer, and let $n \in \mathbb{N}$. Then $g_k(n) = n+\Big\lceil \lfloor \frac{2n}{k-1} \rfloor/2 \Big\rceil$. \end{lemma} \begin{proof} Set $m = \lfloor \frac{n}{k-1} \rfloor$ and $\ell = n-(k-1)m$. Clearly, $m$ is a nonnegative integer, $\ell \in \{0,\dots,k-2\}$, $n = (k-1)m+\ell$, and $n \equiv \ell$ (mod $k-1$). First, we have that \begin{displaymath} \begin{array}{ccc cc} \Big\lceil \lfloor \frac{2n}{k-1} \rfloor/2 \Big\rceil & = & \Big\lceil \Big\lfloor \frac{2((k-1)m+\ell)}{k-1} \Big\rfloor/2 \Big\rceil & = & m+\Big\lceil \lfloor \frac{2\ell}{k-1} \rfloor/2 \Big\rceil. \end{array} \end{displaymath} Suppose first that $0 \leq \ell \leq \frac{k-3}{2}$; then $g_k(n) = \lfloor \frac{kn}{k-1} \rfloor$. We now have that \begin{displaymath} \begin{array}{rcl} n+\Big\lceil \lfloor \frac{2n}{k-1} \rfloor/2 \Big\rceil & = & n+m+\Big\lceil \lfloor \frac{2\ell}{k-1} \rfloor/2 \Big\rceil \\ \\ & = & n+m \\ \\ & = & \lfloor \frac{kn}{k-1} \rfloor \\ \\ & = & g_k(n), \end{array} \end{displaymath} which is what we needed. Suppose now that $\frac{k-1}{2} \leq \ell \leq k-2$; then $g_k(n) = \lceil \frac{kn}{k-1} \rceil$. Now, note that \begin{displaymath} \begin{array}{rcl} n+\Big\lceil \lfloor \frac{2n}{k-1} \rfloor/2 \Big\rceil & = & n+m+\Big\lceil \lfloor \frac{2\ell}{k-1} \rfloor/2 \Big\rceil \\ \\ & = & n+m+1 \\ \\ & = & \lfloor \frac{kn}{k-1} \rfloor+1. \end{array} \end{displaymath} Since $\ell \neq 0$, we see that $\frac{kn}{k-1}$ is not an integer, and so $\lfloor \frac{kn}{k-1} \rfloor+1 = \lceil \frac{kn}{k-1} \rceil$. We now have that \begin{displaymath} \begin{array}{ccc cccc} n+\Big\lceil \lfloor \frac{2n}{k-1} \rfloor/2 \Big\rceil & = & \lfloor \frac{kn}{k-1} \rfloor+1 & = & \lceil \frac{kn}{k-1} \rceil & = & g_k(n), \end{array} \end{displaymath} which is what we needed. This completes the argument. \end{proof} Given functions $f,g:\mathbb{N} \rightarrow \mathbb{N}$, we write $f \leq g$ and $g \geq f$, if for all $n \in \mathbb{N}$, we have that $f(n) \leq g(n)$. As usual, a function $f:\mathbb{N} \rightarrow \mathbb{N}$ is said to be {\em nondecreasing} if for all $n_1,n_2 \in \mathbb{N}$ such that $n_1 \leq n_2$, we have that $f(n_1) \leq f(n_2)$. \begin{lemma} \label{lemma-fTfkgk} Function $f_{\text{T}}$ is nondecreasing, and $f_{\text{T}} = f_5 = g_5$. Furthermore, for all odd integers $k \geq 5$, all the following hold: \begin{itemize} \item[(a)] $f_{\text{T}} \geq f_k \geq g_k$; \item[(b)] functions $f_k$ and $g_k$ are nondecreasing; \item[(c)] $f_k \geq f_{k+2}$ and $g_k \geq g_{k+2}$. \end{itemize} \end{lemma} \begin{proof} The fact that $f_{\text{T}}$ is nondecreasing, and that $f_{\text{T}} = f_5 = g_5$, follows from the definitions of $f_{\text{T}}$, $f_5$, and $g_5$. Further, it follows from construction that for all odd integers $k \geq 5$, we have that $f_k \geq g_k$. The rest readily follows from Lemmas~\ref{lemma-fk-calculation} and~\ref{lemma-gk-calculation}. \end{proof} \begin{lemma} \label{lemma-hyp-chi} Let $k \geq 5$ be an odd integer. Then all $k$-hyperholes $H$ satisfy $\chi(H) \leq f_k(\omega(H))$. Furthermore, there exists a sequence $\{H_n^k\}_{n=2}^{\infty}$ of $k$-hyperholes such that for all integers $n \geq 2$, we have that $\omega(H_n^k) = n$ and $\chi(H_n^k) = f_k(n)$. \end{lemma} \begin{proof} We begin by proving the first statement of the lemma. Let $H$ be a $k$-hyperhole, and let $(X_1,\dots,X_k)$ be a partition of $V(H)$ into nonempty cliques such that for all $i \in \{1,\dots,k\}$, $X_i$ is complete to $X_{i-1} \cup X_{i+1}$ and anticomplete to $V(H) \setminus (X_{i-1} \cup X_i \cup X_{i+1})$, as in the definition of a $k$-hyperhole. Since $H$ is a $k$-hyperhole, and since $k$ is odd, we have that $\alpha(H) = \lfloor k/2 \rfloor = \frac{k-1}{2}$. Then by Lemma~\ref{lemma-hyperhole-chi-formula}, we have that \begin{displaymath} \begin{array}{ccccc} \chi(H) & = & \max\Big\{\omega(H),\Big\lceil \frac{|V(H)|}{\alpha(H)} \Big\rceil\Big\} & = & \max\Big\{\omega(H),\Big\lceil \frac{2|V(H)|}{k-1} \Big\rceil\Big\}. \end{array} \end{displaymath} It is clear that $\omega(H) \leq f_k(\omega(H))$, and so it suffices to show that $\Big\lceil \frac{2|V(H)|}{k-1} \Big\rceil \leq f_k(\omega(H))$. Clearly, for all $i \in \{1,\dots,k\}$, $X_i \cup X_{i+1}$ is a clique, and so $|X_i|+|X_{i+1}| \leq \omega(H)$. In particular, $|X_k|+|X_1| \leq \omega(H)$, and so either $|X_k| \leq \lfloor \omega(H)/2 \rfloor$ or $|X_1| \leq \lfloor \omega(H)/2 \rfloor$; by symmetry, we may assume that $|X_k| \leq \lfloor \omega(H)/2 \rfloor$. Now, using the fact that $k$ is odd, we get that \begin{displaymath} \begin{array}{rcl} |V(H)| & = & \sum\limits_{i=1}^k |X_i| \\ \\ & = & \Big(\sum\limits_{i=1}^{(k-1)/2} (|X_{2i-1}|+|X_{2i}|)\Big)+|X_k| \\ \\ & \leq & \frac{k-1}{2}\omega(H)+\lfloor \omega(H)/2 \rfloor. \end{array} \end{displaymath} But now by Lemma~\ref{lemma-fk-calculation}, we have that \begin{displaymath} \begin{array}{rcl} \Big\lceil \frac{2|V(H)|}{k-1} \Big\rceil & \leq & \Bigg\lceil \frac{2\Big(\frac{k-1}{2}\omega(H)+\lfloor \omega(H)/2 \rfloor\Big)}{k-1} \Bigg\rceil \\ \\ & = & \omega(H)+\Big\lceil \frac{2\lfloor \omega(H)/2 \rfloor}{k-1} \Big\rceil \\ \\ & = & f_k(\omega(H)), \end{array} \end{displaymath} which is what we needed. This proves the first statement of the lemma. It remains to prove the second statement of the lemma. We fix an integer $n \geq 2$, and we construct $H_n^k$ as follows. Let $X_1,\dots,X_k$ be pairwise disjoint sets such that for all $i \in \{1,\dots,k\}$, \begin{itemize} \item if $i$ is odd, then $|X_i| = \lfloor n/2 \rfloor$, and \item if $i$ is even, then $|X_i| = \lceil n/2 \rceil$. \end{itemize} Since $n \geq 2$, sets $X_1,\dots,X_k$ are all nonempty. Now, let $H_n^k$ be the graph whose vertex set is $V(H_n^k) = X_1 \cup \dots \cup X_k$, and with adjacency as follows: \begin{itemize} \item $X_1,\dots,X_k$ are all cliques; \item for all $i \in \{1,\dots,k\}$, $X_i$ is complete to $X_{i-1} \cup X_{i+1}$ and anticomplete to $V(H_n^k) \setminus (X_{i-1} \cup X_i \cup X_{i+1})$. \end{itemize} Clearly, $H_n^k$ is a $k$-hyperhole, and $\omega(H_n^k) = \lfloor n/2 \rfloor+\lceil n/2 \rceil = n$. It remains to show that $\chi(H_n^k) = f_k(n)$. But by the first statement of the lemma, we have that $\chi(H_n^k) \leq f_k(n)$, and so in fact, it suffices to show that $\chi(H_n^k) \geq f_k(n)$. It is clear that $\chi(H_n^k) \geq \Big\lceil \frac{|V(H_n^k)|}{\alpha(H_n^k)} \Big\rceil$. Further, by construction, and by the fact that $k$ is odd, we have that \begin{itemize} \item $\alpha(H_n^k) = \lfloor k/2 \rfloor = \frac{k-1}{2}$, and \item $|V(H_n^k)| = \lceil k/2 \rceil \lfloor n/2 \rfloor+\lfloor k/2 \rfloor \lceil n/2 \rceil = \frac{k-1}{2}n+\lfloor n/2 \rfloor$. \end{itemize} Thus, \begin{displaymath} \begin{array}{ccc cccc} \chi(H_n^k) & \geq & \Big\lceil \frac{|V(H_n^k)|}{\alpha(H_n^k)} \Big\rceil & = & \Bigg\lceil \frac{2\Big(\frac{k-1}{2}n+\lfloor n/2 \rfloor\Big)}{k-1} \Bigg\rceil & = & n+\Big\lceil \frac{2 \lfloor n/2 \rfloor}{k-1} \Big\rceil. \end{array} \end{displaymath} Lemma~\ref{lemma-fk-calculation} now implies that \begin{displaymath} \begin{array}{ccc cc} \chi(H_n^k) & \geq & n+\Big\lceil \frac{2 \lfloor n/2 \rfloor}{k-1} \Big\rceil & = & f_k(n), \end{array} \end{displaymath} which is what we needed. This proves the second statement of the lemma. \end{proof} \begin{theorem} \label{thm-HHk-chi-optimal} Let $k \geq 4$ be an integer. Then $\mathcal{H}_k$ is $\chi$-bounded. Furthermore, if $k$ is even, then the identity function $i_{\mathbb{N}}$ is the optimal $\chi$-bounding function for $\mathcal{H}_k$, and if $k$ is odd, then $f_k$ is the optimal $\chi$-bounding function for $\mathcal{H}_k$. \end{theorem} \begin{proof} Note that every induced subgraph of a $k$-hyperhole is either a $k$-hyperhole or a chordal graph.\footnote{This is easy to see by inspection, but it also follows from Lemma~\ref{lemma-ring-chordal}(c).} Since chordal graphs are perfect (by~\cite{Berge-German, D61}), it follows that all graphs in $\mathcal{H}_k$ are either $k$-hyperholes or perfect graphs. Furthermore, by construction, $\mathcal{H}_k$ contains all $k$-hyperholes. Thus, if $k$ is odd, then Lemma~\ref{lemma-hyp-chi} implies that $f_k$ is the optimal $\chi$-bounding function for $\mathcal{H}_k$.\footnote{We are also using the fact that $K_1 \in \mathcal{H}_k$, $\omega(K_1) = 1$, and $\chi(K_1) = 1 = f_k(1)$.} Suppose now that $k$ is even. By Lemma~\ref{lemma-even-ring-col-alg}, all even hyperholes are perfect, and we deduce that all graphs in $\mathcal{H}_k$ are perfect. Furthermore, $\mathcal{H}_k$ contains all complete graphs. So, $i_{\mathbb{N}}$ is the optimal $\chi$-bounding function for $\mathcal{H}_k$. \end{proof} \begin{corollary} \label{cor-HHk-chi-optimal} Let $k \geq 4$ be an integer. Then $\mathcal{H}_{\geq k}$ is $\chi$-bounded. Furthermore, if $k$ is even, then $f_{k+1}$ is the optimal $\chi$-bounding function for $\mathcal{H}_{\geq k}$, and if $k$ is odd, then $f_k$ is the optimal $\chi$-bounding function for $\mathcal{H}_{\geq k}$. \end{corollary} \begin{proof} This follows immediately from Lemma~\ref{lemma-fTfkgk}(c) and Theorem~\ref{thm-HHk-chi-optimal}. \end{proof} \begin{lemma} \label{lemma-ring-chi-optimal} Let $k \geq 5$ be an odd integer. Then all $k$-rings $R$ satisfy $\chi(R) \leq f_k(\omega(R))$. Furthermore, there exists a sequence $\{R_n^k\}_{n=2}^{\infty}$ of $k$-rings such that for all integers $n \geq 2$, we have that $\omega(R_n^k) = n$ and $\chi(R_n^k) = f_k(n)$. \end{lemma} \begin{proof} Since every $k$-hyperhole is a $k$-ring, the second statement of the lemma follows immediately from the second statement of Lemma~\ref{lemma-hyp-chi}. It remains to prove the first statement. Let $R$ be a $k$-ring. Then by Theorem~\ref{thm-ring-hyperhole}, there exists a $k$-hyperhole $H$ in $R$ such that $\chi(R) = \chi(H)$. By Lemma~\ref{lemma-hyp-chi}, we have that $\chi(H) \leq f_k(\omega(H))$. Clearly, $\omega(H) \leq \omega(R)$, and by Lemma~\ref{lemma-fTfkgk}(b), $f_k$ is a nondecreasing function. We now have that \begin{displaymath} \begin{array}{ccc cccc} \chi(R) & = & \chi(H) & \leq & f_k(\omega(H)) & \leq & f_k(\omega(R)), \end{array} \end{displaymath} which is what we needed. This completes the argument. \end{proof} \begin{theorem} \label{thm-RRk-chi-optimal} Let $k \geq 4$ be an integer. Then $\mathcal{R}_k$ is $\chi$-bounded. Furthermore, if $k$ is even, then the identity function $i_{\mathbb{N}}$ is the optimal $\chi$-bounding function for $\mathcal{R}_k$, and if $k$ is odd, then $f_k$ is the optimal $\chi$-bounding function for $\mathcal{R}_k$. \end{theorem} \begin{proof} Suppose first that $k$ is even. By Lemma~\ref{lemma-even-ring-col-alg}, every $k$-ring $R$ satisfies $\chi(R) = \omega(R)$. Lemma~\ref{lemma-simplicial-chi} and an easy induction now imply that $\mathcal{R}_k$ is $\chi$-bounded by $i_{\mathbb{N}}$, and it is obvious that this $\chi$-bounding function is optimal. Suppose now that $k$ is odd. By Lemma~\ref{lemma-RRk-hered}, all $k$-rings belong to $\mathcal{R}_k$. Thus, it suffices to show that $\mathcal{R}_k$ is $\chi$-bounded by $f_k$, for optimality will then follow immediately from Lemma~\ref{lemma-ring-chi-optimal}.\footnote{We are also using the fact that $K_1 \in \mathcal{R}_k$, $\omega(K_1) = 1$, and $\chi(K_1) = 1 = f_k(1)$.} So, fix $G \in \mathcal{R}_k$, and assume inductively that all graphs $G' \in \mathcal{R}_k$ with $|V(G')|<|V(G)|$ satisfy $\chi(G') \leq f_k(\omega(G'))$. We must show that $\chi(G) \leq f_k(\omega(G))$. If $G$ is a complete graph, then $\chi(G) = \omega(G) \leq f_k(\omega(G))$, and we are done. So assume that $G$ is not complete, and in particular, $G$ has at least two vertices. Suppose that $G$ has a simplicial vertex $v$. Then by Lemma~\ref{lemma-simplicial-chi}, $\chi(G) = \max\{\omega(G),\chi(G \setminus v)\}$. Clearly, $\omega(G) \leq f_k(\omega(G))$. On the other hand, using the induction hypothesis and the fact that $f_k$ is nondecreasing (by Lemma~\ref{lemma-fTfkgk}(b)), we get that $\chi(G \setminus v) \leq f_k(\omega(G \setminus v)) \leq f_k(\omega(G))$. It now follows that $\chi(G) = \max\{\omega(G),\chi(G \setminus v)\} \leq f_k(\omega(G))$, which is what we needed. Suppose now that $G$ does not contain a simplicial vertex. Then by the definition of $\mathcal{R}_k$, $G$ is a $k$-ring, and so Lemma~\ref{lemma-ring-chi-optimal} implies that $\chi(G) \leq f_k(\omega(G))$. This completes the argument. \end{proof} \begin{corollary} \label{cor-RRk-chi-optimal} Let $k \geq 4$ be an integer. Then $\mathcal{R}_{\geq k}$ is $\chi$-bounded. Furthermore, if $k$ is even, then $f_{k+1}$ is the optimal $\chi$-bounding function for $\mathcal{R}_{\geq k}$, and if $k$ is odd, then $f_k$ is the optimal $\chi$-bounding function for $\mathcal{R}_{\geq k}$. \end{corollary} \begin{proof} This follows immediately from Lemma~\ref{lemma-fTfkgk}(c) and Theorem~\ref{thm-RRk-chi-optimal}. \end{proof} A {\em cobipartite graph} is a graph whose complement is bipartite. Equivalently, a graph is {\em cobipartite} if its vertex set can be partitioned into two (possibly empty) cliques. \begin{lemma} \label{lemma-hypanti-perfect} Let $k \geq 4$ be an integer, let $A$ be a $k$-hyperantihole, and let $(X_1,\dots,X_k)$ be a partition of $V(A)$ into nonempty cliques such that for all $i \in \{1,\dots,k\}$, $X_i$ is complete to $V(A) \setminus (X_{i-1} \cup X_i \cup X_{i+1})$ and anticomplete to $X_{i-1} \cup X_{i+1}$. Then for all $i \in \{1,\dots,k\}$, $A \setminus X_i$ is perfect. Furthermore, if $k$ is even, then $A$ is perfect. \end{lemma} \begin{proof} The Perfect Graph Theorem~\cite{Lovasz} states that a graph is perfect if and only if its complement is perfect; bipartite graphs are obviously perfect, and it follows that cobipartite graphs are also perfect. Clearly, for all $i \in \{1,\dots,k\}$, $A \setminus X_i$ is cobipartite and consequently perfect. Furthermore, if $k$ is even, then $A$ is cobipartite and consequently perfect. \end{proof} \begin{lemma} \label{lemma-hypanti-chi} Let $k \geq 5$ be an odd integer. Then all $k$-hyperantiholes $A$ satisfy $\omega(A) \geq \frac{k-1}{2}$ and $\chi(A) \leq g_k(\omega(A))$. Furthermore, there exists a sequence $\{A_n^k\}_{n=\frac{k-1}{2}}^{\infty}$ of $k$-hyperantiholes such that for all integers $n \geq \frac{k-1}{2}$, we have that $\omega(A_n^k) = n$ and $\chi(A_n^k) = g_k(n)$.\end{lemma} \begin{proof} We begin by proving the first statement of the lemma. Let $A$ be a $k$-hyperantihole, and let $(X_1,\dots,X_k)$ be a partition of $V(A)$ into nonempty cliques such that for all $i \in \{1,\dots,k\}$, $X_i$ is complete to $V(A) \setminus (X_{i-1} \cup X_i \cup X_{i+1})$ and anticomplete to $X_{i-1} \cup X_{i+1}$, as in the definition of a $k$-hyperantihole. Since $A$ is a $k$-hyperantihole, and since $k$ is odd, we see that $\omega(A) \geq \lfloor \frac{k}{2} \rfloor = \frac{k-1}{2}$. By symmetry, we may assume that $|X_2| = \min\{|X_1|,\dots,|X_k|\}$. Since $\bigcup\limits_{i=1}^{(k-1)/2} X_{2i}$ is a clique, we see that $\sum\limits_{i=1}^{(k-1)/2} |X_{2i}| \leq \omega(A)$, and so by the minimality of $|X_2|$, we have that $|X_2| \leq \Big\lfloor \frac{2\omega(A)}{k-1} \Big\rfloor$. By construction, $X_2$ is anticomplete to $X_1 \cup X_3$ in $A$, and $|X_2| \leq |X_1|,|X_3|$. Fix any $X_1^2 \subseteq X_1$ and $X_3^2 \subseteq X_3$ such that either $|X_1^2| = \Big\lfloor |X_2|/2 \Big\rfloor$ and $|X_3^2| = \Big\lceil |X_2|/2 \Big\rceil$, or $|X_1^2| = \Big\lceil |X_2|/2 \Big\rceil$ and $|X_3^2| = \Big\lfloor |X_2|/2 \Big\rfloor$.\footnote{This way, we maintain full symmetry between $X_1$ and $X_3$.} Let $X_2^* = X_1^2 \cup X_2 \cup X_3^2$. Note that $X_2$ and $X_2^* \setminus X_2 = X_1^2 \cup X_3^2$ are cliques in $A$, they are anticomplete to each other in $A$, and they are both of size $|X_2|$. Thus, $\chi(A[X_2^*]) = |X_2|$. By Lemma~\ref{lemma-hypanti-perfect}, $A \setminus X_2$ is perfect. Since $A \setminus X_2^*$ is an induced subgraph of $A \setminus X_2$, it follows that $\chi(A \setminus X_2^*) = \omega(A \setminus X_2^*)$. Let $K$ be a maximum clique of $A \setminus X_2^*$. (In particular, $K \cap X_2 = \emptyset$.) Then \begin{displaymath} \begin{array}{rcl} \chi(A) & \leq & \chi(A \setminus X_2^*)+\chi(A[X_2^*]) \\ \\ & = & \omega(A \setminus X_2^*)+|X_2| \\ \\ & = & |K|+|X_2| \\ \\ & = & |K \cup X_2|. \end{array} \end{displaymath} Suppose first that $K$ intersects neither $X_1 \setminus X_1^2$ nor $X_3 \setminus X_3^2$. Since $K \subseteq V(A) \setminus X_2^*$, it follows that $K \cap (X_1 \cup X_3) = \emptyset$. Then $X_2$ is complete to $K$. Thus, $K \cup X_2$ is a clique of $A$, and it follows that $|K \cup X_2| \leq \omega(A)$; consequently, \begin{displaymath} \begin{array}{ccc cccc} \chi(A) & \leq & |K \cup X_2| & \leq & \omega(A) & \leq & g_k(\omega(A)), \end{array} \end{displaymath} and we are done. Suppose now that $K$ intersects at least one of $X_1 \setminus X_1^2$ and $X_3 \setminus X_3^2$; by symmetry, we may assume that $K \cap (X_1 \setminus X_1^2) \neq \emptyset$. Then $K \cup X_1^2$ is a clique of $A$,\footnote{Since $K \subseteq V(A) \setminus X_2^*$ and $X_1^2 \subseteq X_2^*$, we have that $K$ and $X_1^2$ are disjoint.} and it follows that $|K \cup X_1^2| \leq \omega(A)$; consequently, \begin{displaymath} \begin{array}{ccc cc} |K| & \leq & \omega(A)-|X_1^2| & \leq & \omega(A)-\Big\lfloor |X_2|/2 \Big\rfloor, \end{array} \end{displaymath} and so \begin{displaymath} \begin{array}{rcl} \chi(A) & \leq & |K|+|X_2| \\ \\ & \leq & (\omega(A)-\Big\lfloor |X_2|/2 \Big\rfloor)+|X_2| \\ \\ & = & \omega(A)+\Big\lceil |X_2|/2 \Big\rceil \\ \\ & \leq & \omega(A)+\Big\lceil \Big\lfloor \frac{2\omega(A)}{k-1} \Big\rfloor/2 \Big\rceil. \end{array} \end{displaymath} By Lemma~\ref{lemma-gk-calculation}, we now have that \begin{displaymath} \begin{array}{ccc cc} \chi(A) & \leq & \omega(A)+\Big\lceil \Big\lfloor \frac{2\omega(A)}{k-1} \Big\rfloor/2 \Big\rceil & = & g_k(\omega(A)), \end{array} \end{displaymath} and again we are done. This proves the first statement of the lemma. It remains to prove the second statement of the lemma. We fix an integer $n \geq \frac{k-1}{2}$, and we construct $A_n^k$ as follows. Set $m = \lfloor \frac{n}{k-1} \rfloor$ and $\ell = n-(k-1)m$. Clearly, $m$ is a nonnegative integer, $\ell \in \{0,\dots,k-2\}$, $n = (k-1)m+\ell$, and $n \equiv \ell$ (mod $k-1$). Now, let $X_1,\dots,X_k$ be pairwise disjoint sets such that for all $i \in \{1,\dots,k\}$, \begin{itemize} \item if $0 \leq \ell \leq \frac{k-3}{2}$, then $|X_1| = \dots = |X_{2\ell}| = 2m+1$ and $|X_{2\ell+1}| = \dots = |X_k| = 2m$; \item if $\frac{k-1}{2} \leq \ell \leq k-2$, then $|X_1| = \dots = |X_{2\ell-k+1}| = 2m+2$ and $|X_{2\ell-k+2}| = \dots = |X_k| = 2m+1$. \end{itemize} Since $n \geq \frac{k-1}{2}$, sets $X_1,\dots,X_k$ are all nonempty. Let $A_n^k$ be the graph with vertex set $V(A_n^k) = X_1 \cup \dots \cup X_k$, and with adjacency as follows: \begin{itemize} \item $X_1,\dots,X_k$ are all cliques; \item for all $i \in \{1,\dots,k\}$, $X_i$ is complete to $V(A_n^k) \setminus (X_{i-1} \cup X_i \cup X_{i+1})$ and anticomplete to $X_{i-1} \cup X_{i+1}$. \end{itemize} Clearly, $A_n^k$ is a $k$-hyperantihole. We must show that $\omega(A_n^k) = n$ and $\chi(A_n^k) = g_k(n)$. We first show that $\omega(A_n^k) = n$. Suppose first that $0 \leq \ell \leq \frac{k-3}{2}$. Now $2\ell$ consecutive $X_i$'s are of size $2m+1$ (since they are consecutive, at most $\ell$ of them can be included in a clique of $A_n^k$), and all the other $X_i$'s are of size $2m$. So, a maximum clique of $A_n^k$ is the union of $\ell$ sets $X_i$ of size $2m+1$, and of $\frac{k-1}{2}-\ell$ sets $X_i$ of size $2m$. It follows that \begin{displaymath} \begin{array}{ccc ccc ccc} \omega(A_n^k) & = & \ell(2m+1)+\Big(\frac{k-1}{2}-\ell\Big)2m & = & (k-1)m+\ell & = & n, \end{array} \end{displaymath} which is what we needed. Suppose now that $\frac{k-1}{2} \leq \ell \leq k-2$. Then $2\ell-k+1$ consecutive $X_i$'s are of size $2m+2$ (since they are consecutive, at most $\lceil \frac{2\ell-k+1}{2} \rceil = \ell-\frac{k-1}{2}$ of them can be included in a clique of $A_n^k$), and all the other $X_i$'s are of size $2m+1$. So, a maximum clique of $A_n^k$ is the union of $\ell-\frac{k-1}{2}$ sets $X_i$ of size $2m+2$, and of $\frac{k-1}{2}-(\ell-\frac{k-1}{2}) = k-\ell-1$ sets $X_i$ of size $2m+1$. It follows that \begin{displaymath} \begin{array}{rcl} \omega(A_n^k) & = & \Big(\ell-\frac{k-1}{2}\Big)(2m+2)+(k-\ell-1)(2m+1) \\ \\ & = & (k-1)m+\ell \\ \\ & = & n, \end{array} \end{displaymath} which is what we needed. We have now shown that $\omega(A_n^k) = n$. It remains to show that $\chi(A_n^k) = g_k(n)$. But by the first statement of the lemma, we have that $\chi(A_n^k) \leq g_k(n)$, and so in fact, it suffices to show that $\chi(A_n^k) \geq g_k(n)$. Clearly, $\chi(A_n^k) \geq \Big\lceil \frac{|V(A_n^k)|}{\alpha(A_n^k)} \Big\rceil$, and since $A_n^k$ is a hyperantihole, we see that $\alpha(A_n^k) = 2$. Thus, $\chi(A_n^k) \geq \Big\lceil \frac{1}{2}|V(A_n^k)| \Big\rceil$. Suppose first that $0 \leq \ell \leq \frac{k-3}{2}$. Then $g_k(n) = \lfloor \frac{kn}{k-1} \rfloor$, and we have that \begin{displaymath} \begin{array}{rcl} \chi(A_k^n) & \geq & \Big\lceil \frac{1}{2}|V(A_n^k)| \Big\rceil \\ \\ & = & \Big\lceil \frac{1}{2}\Big(2\ell(2m+1)+(k-2\ell)2m\Big) \Big\rceil \\ \\ & = & km+\ell \\ \\ & = & n+m \\ \\ & = & \lfloor \frac{kn}{k-1} \rfloor \\ \\ & = & g_k(n), \end{array} \end{displaymath} which is what we needed. Suppose now that $\frac{k-1}{2} \leq \ell \leq k-2$. Since $\ell \neq 0$, we see that $\frac{kn}{k-1}$ is not an integer, and so $\lfloor \frac{kn}{k-1} \rfloor+1 = \lceil \frac{kn}{k-1} \rceil$. Further, since $\frac{k-1}{2} \leq \ell \leq k-2$, we have that $g_k(n) = \lceil \frac{kn}{k-1} \rceil$. We then see that \begin{displaymath} \begin{array}{rcl} \chi(A_k^n) & \geq & \Big\lceil \frac{1}{2}|V(A_n^k)| \Big\rceil \\ \\ & = & \Big\lceil \frac{1}{2}\Big((2\ell-k+1)(2m+2)+(2k-2\ell-1)(2m+1)\Big) \Big\rceil \\ \\ & = & \Big\lceil km+\ell+\frac{1}{2} \Big\rceil \\ \\ & = & km+\ell+1 \\ \\ & = & n+m+1 \\ \\ & = & \lfloor \frac{kn}{k-1} \rfloor+1 \\ \\ & = & \lceil \frac{kn}{k-1} \rceil \\ \\ & = & g_k(n), \end{array} \end{displaymath} which is what we needed. This proves the second statement of the lemma. \end{proof} \begin{theorem} \label{thm-AAk-chi-optimal} Let $k \geq 4$ be an integer. Then $\mathcal{A}_k$ is $\chi$-bounded. Furthermore, if $k$ is even, then the identity function $i_{\mathbb{N}}$ is the optimal $\chi$-bounding function for $\mathcal{A}_k$, and if $k$ is odd, then $g_k$ is the optimal $\chi$-bounding function for $\mathcal{A}_k$. \end{theorem} \begin{proof} If $k$ is even, then by Lemma~\ref{lemma-hypanti-perfect}, all graphs in $\mathcal{A}_k$ are perfect, and it follows that $i_{\mathbb{N}}$ is the optimal $\chi$-bounding function for $\mathcal{A}_k$. Suppose now that $k$ is odd. Clearly, all $k$-hyperantiholes belong to $\mathcal{A}_k$; on the other hand, it follows from Lemma~\ref{lemma-hypanti-perfect} that all graphs in $\mathcal{A}_k$ are either $k$-hyperantiholes or perfect graphs. So, by Lemma~\ref{lemma-hypanti-chi}, $\mathcal{A}_k$ is $\chi$-bounded by $g_k$. It remains to establish the optimality of $g_k$. Fix $n \in \mathbb{N}$. If $n \leq \frac{k-3}{2}$, then $g_k(n) = n$, and we observe that $K_n \in \mathcal{A}_k$, $\omega(K_n) = n$, and $\chi(K_n) = n = g_k(n)$. On the other hand, if $n \geq \frac{k-1}{2}$, then we let $A_n^k$ be as in Lemma~\ref{lemma-hypanti-chi}, and we observe that $A_n^k \in \mathcal{A}_k$, $\omega(A_n^k) = n$, and $\chi(A_n^k) = g_k(n)$. This proves that the $\chi$-bounding function $g_k$ for $\mathcal{A}_k$ is indeed optimal. \end{proof} \begin{corollary} \label{cor-AAk-chi-optimal} Let $k \geq 4$ be an integer. Then $\mathcal{A}_{\geq k}$ is $\chi$-bounded. Furthermore, if $k$ is even, then $g_{k+1}$ is the optimal $\chi$-bounding function for $\mathcal{A}_{\geq k}$, and if $k$ is odd, then $g_k$ is the optimal $\chi$-bounding function for $\mathcal{A}_{\geq k}$. \end{corollary} \begin{proof} This follows immediately from Lemma~\ref{lemma-fTfkgk}(c) and Theorem~\ref{thm-AAk-chi-optimal}. \end{proof} We remind the reader that the function $f_{\text{T}}:\mathbb{N} \rightarrow \mathbb{N}$ is given by \begin{displaymath} \begin{array}{lll} f_{\text{T}}(n) & = & \left\{\begin{array}{lll} \lfloor 5n/4 \rfloor & \text{if} & \text{$n \equiv 0,1$ (mod $4$)} \\ \\ \lceil 5n/4 \rceil & \text{if} & \text{$n \equiv 2,3$ (mod $4$)} \end{array}\right. \end{array} \end{displaymath} for all $n \in \mathbb{N}$. Note that $\mathcal{H}_{\geq 4}$ is the class of all induced subgraphs of hyperholes, and that $\mathcal{A}_{\geq 4}$ is the class of all induced subgraphs of hyperantiholes. Furthermore, by Lemma~\ref{lemma-RRk-hered}, $\mathcal{R}_{\geq 4}$ contains all induced subgraphs of rings. In particular, $\mathcal{H}_{\geq 4} \subseteq \mathcal{R}_{\geq 4}$. \begin{theorem} \label{thm-HH-AA-RR-chi-optimal} Classes $\mathcal{H}_{\geq 4}$, $\mathcal{A}_{\geq 4}$, and $\mathcal{R}_{\geq 4}$ are $\chi$-bounded. Furthermore, $f_{\text{T}}$ is the optimal $\chi$-bounding function for all three classes. \end{theorem} \begin{proof} By Lemma~\ref{lemma-fTfkgk}, we have that $f_{\text{T}} = f_5 = g_5$. The result now follows immediately from Corollaries~\ref{cor-HHk-chi-optimal},~\ref{cor-RRk-chi-optimal}, and~\ref{cor-AAk-chi-optimal}. \end{proof} \begin{theorem} \label{thm-GT-chi} $\mathcal{G}_{\text{T}}$ is $\chi$-bounded. Furthermore, $f_{\text{T}}$ is the optimal $\chi$-bounding function for $\mathcal{G}_{\text{T}}$. \end{theorem} \begin{proof} We begin by showing $f_{\text{T}}$ is a $\chi$-bounding function for $\mathcal{G}_{\text{T}}$. First, by Lemma~\ref{lemma-fTfkgk}, we have that $f_{\text{T}}$ is nondecreasing, and that $f_{\text{T}} = f_5 = g_5$. Now, fix $G \in \mathcal{G}_{\text{T}}$, and assume inductively that for all $G' \in \mathcal{G}_{\text{T}}$ such that $|V(G')| < |V(G)|$, we have that $\chi(G') \leq f_{\text{T}}(\omega(G'))$. By Theorem~\ref{thm-GT-decomp}, we know that either $G$ is a complete graph, a ring, or a 7-hyperantihole, or $G$ admits a clique-cutset. If $G$ is a complete graph, a ring, or a 7-hyperantihole, then $G \in \mathcal{R}_{\geq 4} \cup \mathcal{A}_{\geq 4}$, and Theorem~\ref{thm-HH-AA-RR-chi-optimal} guarantees that $\chi(G) \leq f_{\text{T}}(\omega(G))$. It remains to consider the case when $G$ admits a clique-cutset. Let $(A,B,C)$ be a clique-cut-partition of $G$, and set $G_A = G[A \cup C]$ and $G_B = G[B \cup C]$. Clearly, $\chi(G) = \max\{\chi(G_A),\chi(G_B)\}$. Using the induction hypothesis and the fact that $f_{\text{T}}$ is nondecreasing, we now get that \begin{displaymath} \begin{array}{rcl} \chi(G) & = & \max\{\chi(G_A),\chi(G_B)\} \\ \\ & \leq & \max\{f_{\text{T}}(\omega(G_A)),f_{\text{T}}(\omega(G_B))\} \\ \\ & \leq & f_{\text{T}}(\omega(G)), \end{array} \end{displaymath} which is what we needed. This proves that $f_{\text{T}}$ is indeed a $\chi$-bounding function for $\mathcal{G}_{\text{T}}$. It remains to establish the optimality of $f_{\text{T}}$. Let $n \in \mathbb{N}$; we must exhibit a graph $G \in \mathcal{G}_{\text{T}}$ such that $\omega(G) = n$ and $\chi(G) = f_{\text{T}}(n)$. If $n = 1$, then we observe that $K_1 \in \mathcal{G}_{\text{T}}$, $\omega(K_1) = 1$, and $\chi(K_1) = 1 = f_{\text{T}}(1)$. So assume that $n \geq 2$. Let $H_n^5$ be as in the statement of Lemma~\ref{lemma-hyp-chi}. Then $H_n^5$ is a 5-hyperhole, and it is easy to see that all hyperholes belong to $\mathcal{G}_{\text{T}}$;\footnote{Alternatively, we observe that every hyperhole is a ring, and by Lemma~\ref{lemma-ring-chordal}(d), all rings belong to $\mathcal{G}_{\text{T}}$.} thus, $H_n^5 \in \mathcal{G}_{\text{T}}$. Further, since $f_{\text{T}} = f_5$, Lemma~\ref{lemma-hyp-chi} guarantees that $\omega(H_n^5) = n$ and $\chi(H_n^5) = f_5(n) = f_{\text{T}}(n)$. Thus, $f_{\text{T}}$ is indeed the optimal $\chi$-bounding function for $\mathcal{G}_{\text{T}}$. \end{proof} \section{Class $\mathcal{G}_{\text{T}}$ and Hadwiger's conjecture} \label{sec:Hadwiger} In this section, we prove Hadwiger's conjecture for the class $\mathcal{G}_{\text{T}}$ (see Theorem~\ref{thm-Hadwiger-GT}). Recall that a graph is {\em perfect} if all its induced subgraphs $H$ satisfy $\chi(H) = \omega(H)$. Obviously, Hadwiger's conjecture is true for perfect graphs: every perfect graph $G$ contains $K_{\chi(G)}$ an induced subgraph, and therefore as a minor as well. \begin{lemma} \label{lemma-Hadwiger-hyperhole} Every hyperhole $H$ contains $K_{\chi(H)}$ as a minor. \end{lemma} \begin{proof} Let $H$ be a hyperhole, and let $k$ be its length. Let $(X_1,\dots,X_k)$ be a partition of $V(H)$ into nonempty cliques such that for all $i \in \{1,\dots,k\}$, $X_i$ is complete to $X_{i-1} \cup X_{i+1}$ and anticomplete to $V(H) \setminus (X_{i-1} \cup X_i \cup X_{i+1})$. By symmetry, we may assume that $|X_1| = \min\{|X_1|,|X_2|,\dots,|X_k|\}$. Clearly, $\chi(H \setminus X_1) = \omega(H \setminus X_1)$,\footnote{By Lemma~\ref{lemma-ring-chordal}(c), $H \setminus X_1$ is chordal, and by~\cite{Berge-German, D61}, chordal graphs are perfect. So, $H \setminus X_1$ is perfect and therefore satisfies $\chi(H \setminus X_1) = \omega(H \setminus X_1)$.} and furthermore, there exists some index $j \in \{2,\dots,k-1\}$ such that $\omega(H \setminus X_1) = |X_j \cup X_{j+1}|$. By the choice of $X_1$, we see that there are $|X_1|$ vertex-disjoint induced paths between $X_{j-1}$ and $X_{j+2}$, none of them passing through $X_j \cup X_{j+1}$. We then take our $|X_1|$ paths and the vertices of $X_j \cup X_{j+1}$ as branch sets, and we obtain a $K_{|X_1|+\omega(H \setminus X_1)}$ minor in $G$. Since $\chi(H) \leq |X_1|+\chi(H \setminus X_1) = |X_1|+\omega(H \setminus X_1)$, we conclude that $H$ contains $K_{\chi(H)}$ as a minor. \end{proof} \begin{lemma} \label{lemma-Hadwiger-ring} Every ring $R$ contains $K_{\chi(R)}$ as a minor. \end{lemma} \begin{proof} This follows immediately from Theorem~\ref{thm-ring-hyperhole} and Lemma~\ref{lemma-Hadwiger-hyperhole}. \end{proof} \begin{lemma} \label{lemma-Hadwiger-hyperantihole} Every hyperantihole $A$ contains $K_{\chi(A)}$ as a minor. \end{lemma} \begin{proof} Let $A$ be a hyperantihole, and let $(X_1,\dots,X_k)$, with $k \geq 4$, be a partition of $V(A)$ into nonempty cliques, such that for all $i \in \{1,\dots,k\}$, $X_i$ is complete to $A \setminus (X_{i-1} \cup X_i \cup X_{i+1})$ and anticomplete to $X_{i-1} \cup X_{i+1}$, as in the definition of a hyperantihole. If $k = 4$, then $V(K)$ can be partitioned into two cliques (namely $X_1 \cup X_3$ and $X_2 \cup X_4$), anticomplete to each other, and the result is immediate. From now on, we assume that $k \geq 5$. By symmetry, we may assume that $|X_1| = \min\{|X_1|,|X_2|,\dots,|X_k|\}$. Clearly, $\chi(A) \leq \chi(A \setminus X_1)+|X_1|$. On the other hand, by Lemma~\ref{lemma-hypanti-perfect}, $A \setminus X_1$ is perfect, and in particular, $\chi(A \setminus X_1) = \omega(A \setminus X_1)$. Let $K$ be a clique of size $\omega(A \setminus X_1)$ in $A \setminus X_1$. Then, $\chi(A) \leq |K|+|X_1|$, and so it suffices to show that $A$ contains $K_{|K|+|X_1|}$ as a minor. If $K \cap (X_k \cup X_2) = \emptyset$, then $X_1$ is complete to $K$ in $A$, $K \cup X_1$ is a clique of size $|K|+|X_1|$ in $A$, and we are done. From now on, we assume that $K$ intersects at least one of $X_2$ and $X_k$. By symmetry, we may assume that $K \cap X_2 \neq \emptyset$. Since $X_2$ is anticomplete to $X_3$, and since $K$ is a clique, we see that $K \cap X_3 = \emptyset$. Since $X_{k-1}$ and $X_k$ are anticomplete to each other, and since $K$ is a clique, we see that $K$ intersects at most one of $X_{k-1},X_k$, and we deduce that $|(X_{k-1} \cup X_k) \setminus K| \geq \min\{|X_{k-1}|,|X_k|\} \geq |X_1|$. So, there exist $|X_1|$ pairwise disjoint three-vertex subsets of $V(A) \setminus K$, each of them containing exactly one vertex from each of the sets $X_1$, $X_3$, and $X_{k-1} \cup X_k$. Clearly, each of these three-vertex sets induces a connected subgraph of $A$. We now take our $|X_1|$ three-vertex sets and all the vertices of $K$ as branch sets, and we obtain a $K_{|K|+|X_1|}$ minor in $A$. This completes the argument. \end{proof} \begin{theorem} \label{thm-Hadwiger-GT} Every graph $G \in \mathcal{G}_{\text{T}}$ contains $K_{\chi(G)}$ as a minor. \end{theorem} \begin{proof} Fix $G \in \mathcal{G}_{\text{T}}$, and assume inductively that every graph $G' \in \mathcal{G}_{\text{T}}$ with $|V(G')| < |V(G)|$ contains $K_{\chi(G')}$ as a minor. We must show that $G$ contains $K_{\chi(G)}$ as a minor. We apply Theorem~\ref{thm-GT-decomp}. Suppose first that $G$ admits a clique-cutset, and let $(A,B,C)$ be a clique-cut-partition of $G$. Clearly, $\chi(G) = \max\{\chi(G[A \cup C]),\chi(G[B \cup C])\}$, and the result follows from the induction hypothesis. So assume that $G$ does not admit a clique-cutset. Then Theorem~\ref{thm-GT-decomp} implies that $G$ is a complete graph, a ring, or a 7-hyperantihole; in the first case, the result is immediate, in the second, it follows from Lemma~\ref{lemma-Hadwiger-ring}, and in the third, it follows from Lemma~\ref{lemma-Hadwiger-hyperantihole}. \end{proof} \small{

https://arxiv.org/abs/2002.01589

Mixed Hodge Structures on Alexander Modules

Motivated by the limit mixed Hodge structure on the Milnor fiber of a hypersurface singularity germ, we construct a natural mixed Hodge structure on the torsion part of the Alexander modules of a smooth connected complex algebraic variety. More precisely, let $U$ be a smooth connected complex algebraic variety and let $f\colon U\to \mathbb{C}^*$ be an algebraic map inducing an epimorphism in fundamental groups. The pullback of the universal cover of $\mathbb{C}^*$ by $f$ gives rise to an infinite cyclic cover $U^f$ of $U$. The action of the deck group $\mathbb{Z}$ on $U^f$ induces a $\mathbb{Q}[t^{\pm 1}]$-module structure on $H_*(U^f;\mathbb{Q})$. We show that the torsion parts $A_*(U^f;\mathbb{Q})$ of the Alexander modules $H_*(U^f;\mathbb{Q})$ carry canonical $\mathbb{Q}$-mixed Hodge structures. We also prove that the covering map $U^f \to U$ induces a mixed Hodge structure morphism on the torsion parts of the Alexander modules. As applications, we investigate the semisimplicity of $A_*(U^f;\mathbb{Q})$, as well as possible weights of the constructed mixed Hodge structures. Finally, in the case when $f\colon U\to \mathbb{C}^*$ is proper, we prove the semisimplicity and purity of $A_*(U^f;\mathbb{Q})$, and we compare our mixed Hodge structure on $A_*(U^f;\mathbb{Q})$ with the limit mixed Hodge structure on the generic fiber of $f$.

\section{Introduction} Let $U$ be a connected topological space of finite homotopy type, and let \begin{align*} \xi\colon\pi_1(U) \twoheadrightarrow \mathbb{Z}\end{align*} be an epimorphism. Denote by $U^\xi$ the infinite cyclic cover of $U$ corresponding to $\ker \xi$. Let $k$ be a subfield of $\mathbb{R}$, and denote by $R=k[t^{\pm 1}]$ the ring of Laurent polynomials in variable $t$ with $k$-coefficients. The group of covering transformations of $U^\xi$ is isomorphic to $\mathbb{Z}$, and it induces an $R$-module structure on each group $H_i(U^\xi;k)$. By analogy with knot theory, the $R$-module $H_i(U^\xi;k)$ is called the $i$-th (homology) {\it $k$-Alexander module} of the pair $(U,\xi)$. Since $U$ is homotopy equivalent to a finite CW-complex, $H_i(U^\xi;k)$ is a finitely generated $R$-module, for each integer $i$. Note that $\xi\colon \pi_1(U) \to \mathbb{Z}$ can be regarded as an element in $H^1(U;\mathbb{Z})$ via the canonical identification: \[ \Hom\big(\pi_1(U),\mathbb{Z}\big) \cong \Hom\big(H_1(U;\mathbb{Z}),\mathbb{Z}\big) \cong H^1(U;\mathbb{Z}). \] Moreover, any such class in $H^1(U;\mathbb{Z})$ is represented by a homotopy class of continuous maps $U \to S^1$. Whenever such a representative $f\colon U \to S^1$ for $\xi$ is fixed (that is, $\xi=f_{*}$), we will also use the notation $U^f$ for the corresponding infinite cyclic cover of $U$. For example, if $f\colon U \to S^1$ is a fiber bundle with connected fiber $F$ a finite CW-complex, then $\xi=f_*\colon \pi_1(U) \to \pi_1(S^1)=\mathbb{Z}$ is surjective, and the corresponding infinite cyclic cover $U^f$ is homeomorphic to $F \times \mathbb{R}$ and hence homotopy equivalent to $F$. The deck group action on $H_i(U^f;k)$ is isomorphic (up to a choice of orientation on $S^1$, as described in Lemma \ref{lem:fiberMonodromy}) to the monodromy action on $H_i(F;k)$, which gives the latter vector spaces $R$-module structures. Therefore $H_i(U^f;k)\cong H_i(F;k)$ is a torsion $R$-module for all $i \geq 0$. This applies in particular to the following geometric situations: \begin{enumerate}[(a)] \item a smooth proper surjective submersion (with connected fibers) $f\colon U \to \Delta^*$ from an open set of a smooth complex algebraic variety to a punctured disc. \item the Milnor fibration $f\colon U \to S^1$ associated to a reduced complex hypersurface singularity germ, with $F$ the corresponding Milnor fiber. \item \label{item:milnor} the global (affine) Milnor fiber $f\colon U=\mathbb{C}^n \setminus \{f=0\} \to \mathbb{C}^*$, where $f$ is a square-free homogeneous polynomial in $n$ complex variables and $F=f^{-1}(1)$. \end{enumerate} In a different vein, it was shown in \cite{LMW} that if $U$ is a smooth quasi-projective variety of complex dimension $n$, admitting a proper semi-small map (e.g., a finite map or a closed embedding) to some complex semiabelian variety, then for any {\it generic} epimorphism $\xi\colon \pi_1(U) \to \mathbb{Z}$ the corresponding Alexander modules $H_i(U^\xi; k)$ are torsion $R$-modules for all $i \neq n$. However, for an arbitrary topological space $U$ of finite homotopy type, the Alexander modules $H_i(U^\xi; k)$ are not torsion $R$-modules in general. One then considers the torsion part \[ A_i(U^\xi;k)\coloneqq \Tors_R H_i(U^\xi;k) \] of the $R$-module $H_i(U^\xi;k)$. This is a $k$-vector space of finite dimension on which a generating covering transformation (i.e., $t$-multiplication) acts as a linear automorphism. Moreover, if $U$ is a smooth complex algebraic variety, then all eigenvalues of the $t$-action on $A_i(U^\xi;k)$ are roots of unity, for any integer $i$ (see Proposition \ref{eig1}). \medskip In this paper, {we assume that $U$ is a smooth complex algebraic variety and} we investigate the existence of mixed Hodge structures on $A_i(U^\xi;k)$ for $k=\mathbb{Q}$ or $\mathbb{R}$. Note that $U^{\xi}$ is not in general a complex algebraic variety, so the classical Deligne theory does not apply. Specifically, we address the following question, communicated to the authors by \c{S}tefan Papadima: \begin{quest}\label{conj} Let $U$ be a smooth connected complex algebraic variety. Let $\xi\colon \pi_1(U) \to \mathbb{Z}$ be an epimorphism with corresponding infinite cyclic cover $U^{\xi}$. Is there a natural $\Q$-mixed Hodge structure on the torsion part $A_i(U^\xi;k)$ of the $\Q[t^{\pm 1}]$-module $H_i(U^{\xi};\Q)$, for all $i \geq 0$? \end{quest} The purpose of this paper is to give a positive answer to Question \ref{conj} in the case when the epimorphism $\xi\colon \pi_1(U) \to \mathbb{Z}$ is realized by an {\it algebraic} map $f\colon U \to \mathbb{C}^*$. We remark here that a homomorphism $\xi\colon \pi_1(U)\to \Z$ is induced by an algebraic map $f\colon U\to \C^*$ if and only if, when considered as an element in $H^1(U; \Z)$, $\xi$ is of type $(1, 1)$, that is $\xi\in F^1H^1(U; \C)\cap \overline{F^1H^1(U; \C)}.$ This is a consequence of Deligne's theory of $1$-motives (cf. \cite[(10.I.3)]{De3}). In this algebraic context, Question \ref{conj} has already been answered positively in the following special situations: \begin{enumerate} \item\label{hain} When $H_i(U^{\xi};\Q)$ is $\Q[t^{\pm 1}]$-torsion for all $i \geq 0$ and the $t$-action is unipotent, see \cite{hain1987rham}; \item \label{dlt}When $f\colon U=\mathbb{C}^n \setminus \{f=0\}\to \mathbb{C}^*$ is induced by a reduced complex polynomial $f\colon \mathbb{C}^n \to \mathbb{C}$ which is {\it transversal at infinity} (i.e., the hyperplane at infinity in $\mathbb{C}P^n$ is transversal in the stratified sense to the projectivization of $\{f=0\}$), for $\xi=f_*$ and $i<n$; see \cite{DL, Liu}. In fact, in this case it was shown in \cite{Max06} that the corresponding Alexander modules $H_i(U^\xi;\mathbb{Q})$ are torsion $\mathbb{Q}[t^{\pm 1}]$-modules for $i<n$, while $H_n(U^\xi;\mathbb{Q})$ is free and $H_i(U^\xi;\mathbb{Q})=0$ for $i>n$. Furthermore, the $t$-action on $H_i(U^\xi;\mathbb{C})$ is diagonalizable (semisimple) for $i<n$, and the corresponding eigenvalues are roots of unity of order $d=\deg(f)$. \item\label{lkk} When $f\colon U=\mathbb{C}^n \setminus \{f=0\}\to \mathbb{C}^*$ is induced by a complex polynomial $f\colon \mathbb{C}^n \to \mathbb{C}$ which has at most isolated singularities, including at infinity, in the sense that both the projectivization of $\{f=0\}$ and its intersection with the hyperplane at infinity have at most isolated singularities. In this case, and with $\xi=f_*$, there is only one interesting Alexander module, $H_{n-1}(U^\xi;\mathbb{Q})$, which is torsion (see \cite[Theorem 4.3, Remark 4.4]{Lib94}), and a mixed Hodge structure on it was constructed in \cite{Lib96}; see also \cite{KK} for the case of plane curves under some extra conditions. \end{enumerate} In this paper we prove the following general statement (see Corollary \ref{halexandermhs}): \begin{thm}\label{mhsexistence} Let $U$ be a smooth connected complex algebraic variety, with an algebraic map $f\colon U \to \mathbb{C}^*$. Assume that $\xi=f_*\colon\pi_1(U) \to \mathbb{Z}$ is an epimorphism, and denote by $U^f = \{(x,z)\in U\times \C \mid f(x) = e^z \}$ the corresponding infinite cyclic cover. Then the torsion part $A_i(U^f;\Q)$ of the $\mathbb{Q}[t^{\pm 1}]$-module $H_i(U^f;\mathbb{Q})$ carries a canonical $\Q$-mixed Hodge structure for any $i\geq 0$. Suppose $N$ is a positive integer, chosen such that $t^N$ acts unipotently on $A_i(U^f;\Q)$. Let $\log(t^N)$ denote the Taylor series centered at $t^N=1$. Then the action of $\log(t^N)$ is a mixed Hodge structure morphism $A_i(U^f;\Q)\to A_i(U^f;\Q)(-1)$, where $(-1)$ denotes the Tate twist of mixed Hodge structures. \end{thm} In Section~\ref{ss:DLandLiu} we prove that, if $U$ and $f$ are as in case (\ref{dlt}) above, the mixed Hodge structure of Theorem \ref{mhsexistence} recovers the mixed Hodge structure obtained by different means in both \cite{DL, Liu}. We expect, but have not proven, the same to be true for cases (\ref{hain}) and (\ref{lkk}). Moreover, in Section~\ref{ss:genFiber}, we also prove that, in the global affine Milnor fibration case \ref{item:milnor} mentioned above, the mixed Hodge structure of Theorem~\ref{mhsexistence} recovers Deligne's mixed Hodge structure on the fiber (Corollary \ref{cor:quasihom}). The proof of Theorem~\ref{mhsexistence} makes use of a sequence of reductions (e.g., after pulling back to a finite cover, one may assume that the $t$-action on $A_*(U^\xi;\Q)$ is unipotent, it also suffices to work with cohomological Alexander modules, etc.), and it relies on the construction of a suitable {\it thickening} of the Hodge-de Rham complex (see Section \ref{mhsal} for details). There are choices made in the construction, but we prove that the resultant mixed Hodge structure is {independent} of them (see Theorems \ref{indcompactification} and \ref{indN}) and further that it behaves functorially with respect to algebraic maps over $\C^*$ (see Theorem~\ref{functorial}). The only choice that affects the mixed Hodge structure is the choice of infinite cyclic cover $U^f$. This is because $\Z$ acts on this space as deck transformations and the mixed Hodge structure is {\bf not} preserved in general (see Proposition~\ref{prop:depf}). However, this is consistent with the behavior of the limit mixed Hodge structure (see Theorem~\ref{comp}), which is also not preserved by deck transformations. We proceed to relate our mixed Hodge structures on the torsion parts of the Alexander modules to known mixed Hodge structures, then derive consequences of our construction and these relations. Centrally, in Section~\ref{sec:main} we prove that the infinite cyclic covering map induces a morphism of mixed Hodge structures: \begin{thm}\label{geoIntro} In the setting of Theorem~\ref{mhsexistence}, the vector space map $A_i(U^f; \Q) \to H_i(U;\Q)$ induced by the covering $U^f \rightarrow U$ is a morphism of mixed Hodge structures for all $i \geq 0$, where $H_i(U;\Q)$ is equipped with (the dual of) Deligne's mixed Hodge structure. \end{thm} We use this theorem and our construction to obtain several results, including a bound on the weight filtrations of the mixed Hodge structures on the torsion parts of the Alexander modules (Theorem \ref{thm:boundedWeights}). This bound coincides with the known bound for the homology of smooth algebraic varieties of the same dimension as the generic fiber of $f$ (cf. \cite[Corollaire 3.2.15]{De2}). \begin{thm}\label{boundsIntro} Assume the setting of Theorem~\ref{mhsexistence}. Let $i \geq 0$. If $k\notin [i,2i]\cap [i,2\dim_\C(U)-2]$, then \[ \Gr^W_{-k} A_i(U^f;\Q) = 0 \] where $\Gr^W_{-k}$ denotes the $-k$th graded piece of the weight filtration. \end{thm} Other consequences of our construction and Theorem~\ref{geoIntro} are related to the $t$-action on the torsion parts of the Alexander modules. For example, we apply it to determine bounds on the size of the Jordan blocks of this $t$-action (see Corollary \ref{cor:jordan}), which nearly cut in half existing bounds, as in \cite[Proposition 1.10]{BudurLiuWang}. We also address conditions under which this $t$-action is a mixed Hodge structure morphism. We prove that this is the case if and only if the $t$-action is semisimple (see Corollary \ref{cor:t} and Proposition \ref{prop:semisimpleIsEasyConverse}). \begin{thm}\label{tsemisimpleIntro} Assume the setting of Theorem \ref{mhsexistence}. Let $i \geq 0$. The $t$-action on $A_i(U^f; \Q)$ is a mixed Hodge structure morphism if and only if it is semisimple. \end{thm} Semisimplicity has other pleasing consequences. We prove that, when the $t$-action is semisimple, the mixed Hodge structure on the torsion parts of the Alexander modules can be constructed directly using a finite cyclic cover, which, unlike an infinite cyclic cover, is always a complex algebraic variety. This bypasses our rather abstract general construction of the mixed Hodge structure. We present two different viewpoints. In the first, we utilize cap product with the pullback of a generator of $H^1(\C^*;\Q)$. In the second, we utilize a generic fiber of the algebraic map, which is always a complex algebraic variety. For the following result see Corollary \ref{cor:cup} and Corollary \ref{cor:fiber}. \begin{thm}\label{finiteIntro} Assume the setting of Theorem \ref{mhsexistence}. Let $i \geq 0$ and assume that the $t$-action on $A_i(U; \Q)$ is semisimple. Let $N$ be such that the action of $t^N$ on $A_i(U^f;\Q)$ is unipotent, and let $U_N = \{(x,z) \in U \times \C^*\mid f(x) = z^N\}$ denote the corresponding $N$-fold cyclic cover. Equip the rational homology of $U_N$ with the (dual of) Deligne's mixed Hodge structure. \begin{enumerate}[(A)] \item Let $f_N\colon U_N \rightarrow \C^*$ denote the algebraic map induced by projection onto the second component, and let $\gen \in H^1(\C^*;\Q)$ be a generator. Then $A_i(U^f;\Q)$ is isomorphic as a mixed Hodge structure to the image of the mixed Hodge structure morphism induced by cap product with $f^*_N(\gen)$ \[(-) \frown f^*_N(\gen) \colon H_{i+1}(U_N;\Q)(-1) \rightarrow H_i(U_N;\Q) ,\] where $(-1)$ denotes the $-1$th Tate twist of a mixed Hodge structure. \item Let $F \hookrightarrow U$ be the inclusion of any generic fiber of $f$ and let $F \hookrightarrow U_N$ be any lift of this inclusion. Then $A_i(U^f;\Q)$ is isomorphic as a mixed Hodge structure to the image of the mixed Hodge structure morphism \[H_i(F; \Q) \rightarrow H_i(U_N; \Q)\] induced by the inclusion, where $H_i(F;\Q)$ is equipped with (the dual of) Deligne's mixed Hodge structure. \end{enumerate} \end{thm} Theorem \ref{finiteIntro}, when it applies, brings the mixed Hodge structures on the torsion parts of the Alexander modules down to earth, and reinforces the significance of semisimplicity. The first viewpoint granted by semisimplicity, in terms of cap products (Theorem \ref{finiteIntro}A), is suggested by the thickened complexes that play the central role in our construction (see Section \ref{ss:cupcap}). Regarding the second viewpoint, note that the homologies of different choices of generic fibers in the same degree may have different mixed Hodge structures, but \emph{any} choice is allowed in Theorem \ref{finiteIntro}B. This shows that the mixed Hodge structures on the torsion parts of the Alexander modules are common quotients of the homologies of all generic fibers, when semisimplicity holds. Our results in this paper show that semisimplicity is not a rare occurrence. In fact, we have proven that semisimplicity holds in many situations: it always holds on the torsion part of the first Alexander module (see Corollary \ref{cor:semisimple}) and, when $f$ is proper, it necessarily holds on the torsion parts of all Alexander modules (see Theorem \ref{thmsimple} and Corollary \ref{corss}). In fact, we do not know of any example where semisimplicity does not hold--see the open questions at the end of Section \ref{sec:examples}. This lack of examples is mainly due to the fact that higher Alexander modules are harder to compute than the first. An interesting point is that we have used the mixed Hodge structure constructed in this paper to prove the semisimplicity of the first Alexander module for arbitrary $U$ and $f$. This was previously known in the case of certain Alexander modules associated to affine curve complements, as is explained in \cite[Corollary 1.7]{DL}, which combines results from \cite{KK,DN,dimca-jag} (see Remark~\ref{remk:A1Semisimple} for details). We use the above-mentioned semisimplicity in the case when $f$ is proper to show the following (see Corollary \ref{cor:pure}). \begin{thm}\label{thm:pur} If $f\colon U \to \C^*$ is a proper algebraic map, then $A_i(U^f;\Q)$ carries a pure Hodge structure of weight $-i$. \end{thm} \medskip There is a well-known and motivating (for our construction) mixed Hodge structure that we would be remiss not to explicitly address. In the situations (a)--(c) mentioned above, the mixed Hodge structure on $A_i(U^\xi;\Q)$ coincides with the {\it limit} mixed Hodge structure on $H_i(F;\Q)$. In the situation considered in this paper, the epimorphism $\xi$ is realized by an algebraic map $f\colon U \to \mathbb{C}^*$. Let $D^*$ be a sufficiently small punctured disk centered at $0$ in $\C$, such that $f\colon f^{-1}(D^*)\rightarrow D^*$ is a fibration, and let $T^*=f^{-1}(D^*)$. The infinite cyclic cover $(T^*)^f$ is homotopy equivalent to $F$, where $F$ denotes any fiber of the form $f^{-1}(c)$, for $c\in D^*$. With this in mind, $(T^*)^f$ can be regarded as the canonical fiber of $f\colon f^{-1}(D^*)\rightarrow D^*$. If $f$ is proper, $H_i((T^*)^f;\Q)$ is also endowed with a limit mixed Hodge structure, which we compare with the one we construct on $A_i(U^f;\Q)$. In Section \ref{rellim} we show the following: \begin{thm}\label{comp} In the setup of Theorem \ref{mhsexistence}, assume $f$ is proper, and let $D^*$ be a small enough punctured disk centered at $0$ in $\C$ such that $f\colon f^{-1}(D^*)\rightarrow D^*$ is a fibration. Let $T^*=f^{-1}(D^*)$, and let $(T^*)^f$ be its corresponding infinite cyclic cover induced by $f$. Then, for all $i\geq 0$, the inclusion $(T^*)^f\subset U^f$ induces an epimorphism of $\Q$-mixed Hodge structures \begin{align*} H_i((T^*)^f;\Q) \twoheadrightarrow A_i(U^f;\Q), \end{align*} where $H_i((T^*)^f;\Q)$ is endowed with its limit mixed Hodge structure. If, moreover, $f$ is a fibration, then the two mixed Hodge structures are isomorphic. \end{thm} \medskip The paper is structured as follows. In Section \ref{sec2}, we recall the relevant background and set notations for the rest of the paper. We introduce here the Alexander modules (homological and cohomological versions, and their descriptions in terms of local systems), differential graded algebras (dga), mixed Hodge complexes, and recall how the latter are used to get mixed Hodge structures on smooth varieties. We also summarize here the construction of the limit mixed Hodge structure associated to a family of projective manifolds. Section \ref{sec3} is devoted to the theory of thickened complexes, which already made an appearance in \cite{budur2018monodromy}. The purpose of Section \ref{sec4} is to show that, under certain technical assumptions, the thickened complex of a multiplicative mixed Hodge complex is again a mixed Hodge complex (Theorem \ref{mhsthickened}). In Section \ref{mhsal}, we perform the construction of a suitable thickening of the Hodge-de Rham complex (see Theorem \ref{logthickenedmhs}) and prove Theorem \ref{mhsexistence} (see Corollary \ref{halexandermhs}). We also show that, fixing $f$, the mixed Hodge structure constructed here is independent of all choices (finite cover used to make the monodromy unipotent, and good compactification) and is functorial. The content of the remaining sections can largely be classified either as (1) comparing our mixed Hodge structures on the torsion parts of the Alexander modules with known mixed Hodge structures or (2) investigating applications. In Section~\ref{sec:main}, we prove that the infinite cyclic covering map induces morphisms of mixed Hodge structures from the torsion parts of the Alexander modules into the homology vector spaces of the base space (see Theorem \ref{geoIntro}). The consequences of this fact are discussed in Section~\ref{sec:consequences}, in particular we obtain sharp bounds on the weight filtrations (Theorem~\ref{boundsIntro}) and explore how the $t$-action interacts with the mixed Hodge structures. It is here that we hit on the significance of semisimplicity, and prove theorems to which it relates (see Theorems \ref{tsemisimpleIntro} and \ref{finiteIntro}). We also show that our construction coincides with the mixed Hodge structure of case (\ref{dlt}) above, when the latter is defined. In Section~\ref{sect:ss}, we prove the semisimplicity of $A_i(U^f; \Q)$ in the case when $f\colon U \to \C^*$ is a proper map (see Corollary \ref{corss}), as well as the purity statement of Theorem~\ref{thm:pur}. In Section \ref{rellim}, we compare the mixed Hodge structure we constructed on the torsion part of the Alexander modules with the limit mixed Hodge structure on the generic fiber of the algebraic map realizing the infinite cyclic cover. In particular, Theorem \ref{comp} is proved in this section. We apply our results to specific examples in Section \ref{sec:examples}, more exactly to the first Alexander modules of hyperplane arrangements. We conclude by listing open questions regarding our constructions, as well as some of our expectations for what the answers will be. \medskip \textbf{Acknowledgements.} We thank Lingquan Ma and Mircea Mus\-ta\-\c{t}\u{a} for helpful discussions. E. Elduque is partially supported by an AMS-Simons Travel Grant. L. Maxim is partially supported by the Simons Foundation Collaboration Grant \#567077. B. Wang is partially supported by the NSF grant DMS-1701305 and by a Sloan Fellowship. \section{Preliminaries}\label{sec2} In this section, we recall some standard results about Alexander modules, mixed Hodge complexes, differential graded algebras and limit mixed Hodge structures. We closely follow the presentation of \cite{peters2008mixed}. However, we choose different conventions in several places. \subsection{Denotations and Assumptions}\label{ss:notations} In this section, we fix the notations for the rest of the paper. Let $k$ denote a field of characteristic zero. Unless otherwise stated, $\otimes$ denotes $\otimes_k$. If $V$ is a graded vector space, we denote by $V^i$ its $i$-th graded component. Let $R$ denote the ring $k[t^{\pm 1}]$ of Laurent polynomials in the variable $t$ over the field $k$. Let $R_\infty$ denote the ring $k[[s]]$ of formal power series in the variable $s$ over the field $k$. We identify $R$ with a distinguished subring of $R_\infty$ by setting $t = 1+s$. For $m \geq 1$ let $R_m$ denote the quotient ring $R/((t-1)^m) = R_\infty/(s^m)$. Throughout, $M^\vee$ will denote the (sometimes derived) dual of $M$ as an $R$-module, and we will use $M^{\vee_k}$ for the dual as a $k$-vector space. If $U$ is a smooth manifold, let $\calE^\bullet_U$ denote the real de Rham complex of sheaves on $U$. If $U$ is moreover a complex manifold, let $\Omega^\bullet_U$ denote the holomorphic de Rham complex of sheaves on $U$. If $X$ is a complex manifold, and $D \subset X$ is a simple normal crossing divisor, let $\logdr{X}{D}$ denote the log de Rham complex of sheaves on $X$. When working with sheaves, we follow the notations of \cite{dimca2004sheaves}. In particular, the pullback is denoted $f^{-1}$. For a module $M$ (over $R$ or $k$) and a space $X$, we will use $\ul M_X$ to denote the constant sheaf on $X$ with stalk $M$. If the differential of a cochain complex is not specified, it is assumed to be denoted by $d$. Given any double complex $\calA^{\bullet,\bullet}$, we denote the total complex by $\Tot \calA^{\bullet, \bullet}$. We use the sign conventions that agree with \cite[III \S 3.2]{gelfandmanin}. Most importantly, given a map of complexes $f\colon A^\bullet\to B^\bullet$, its cone is $\Cone(f)^i = B^i\oplus A^{i+1}$, with differential $\begin{pmatrix} d_B & f\\ 0 & -d_A \end{pmatrix}$. The maps $B^\bullet\to \Cone(f)\to A^\bullet[1]$ are induced by $\Id_B$ and $\Id_A$. \subsection{Alexander Modules} \label{ssAlex} Let $U$ be a smooth connected complex algebraic variety, and let $f\colon U\rightarrow \C^*$ be an algebraic map inducing an epimorphism $f_*\colon\pi_1(U)\twoheadrightarrow \Z$ on fundamental groups. Let $\exp\colon\C\to \C^*$ be the infinite cyclic cover, and let $U^f$ be the following fiber product: \[ \begin{tikzcd} U^f\subset U\times \C\arrow[r,"f_\infty"] \arrow[d,"\pi"]\arrow[dr,phantom,very near start, "\lrcorner"]& \C \arrow[d,"\exp"] \\ U\arrow[r,"f"] & \C^*. \end{tikzcd} \] Under this presentation $U^f$ is embedded in $U\times \C$: \[ U^f = \{ (x,z)\in U\times \C\mid f(x) = e^z\}. \] Let $f_\infty$ be the restriction to $U^f$ of the projection $U\times \C\to \C$. Since $\exp$ is an infinite cyclic cover, $\pi\colon U^f\to U$ is the infinite cyclic cover induced by $\ker f_*$. The group of covering transformations of $U^f$ is isomorphic to $\mathbb{Z}$, and it induces an $R$-module structure on each group $H_i(U^f;k)$, with $1\in \mathbb{Z}$ corresponding to $t \in R = k[t^{\pm 1}]$. We will also say $t$ acts on $U^f$ as the deck transformation $(x,z)\mapsto (x,z+2\pi i)$. \begin{dfn} The \bi{$i$-th homological Alexander module} of $U$ associated to the algebraic map $f\colon U\rightarrow \C^*$ is the $R$-module $$ H_i(U^f;k). $$ \end{dfn} \begin{remk}\label{rem:ACHTUNG} The deck transformation group, generated by $t$, induces an automorphism of $H_i(U^f;k)$, but in principle it does not preserve the mixed Hodge structure that we will define on its torsion. We will show in Corollary~\ref{cor:t} and Proposition~\ref{prop:semisimpleIsEasyConverse} that this is the case if and only if the $t$-action is semisimple on $\Tors_R H_i(U^f;k)$. Therefore, it is important that $f_\infty$ is fixed, as choosing a map of the form $f_\infty\circ t^m = f_\infty + 2\pi i m$ for any $m\in \Z$ would still identify $U^f$ with the fiber product $U\times_{\C^*}\C$, but it could give rise to an isomorphic, but not equal, mixed Hodge structure on the same vector space. This is to be expected: the limit mixed Hodge structure exhibits the same behavior, see \cite[11.2]{peters2008mixed}. $\sslash$ \end{remk} For our purposes, it is more convenient to realize the Alexander modules as homology groups of a certain local system on $U$, which we now define. Let $\calL = \pi_! \ul k_{U^f}$. The action of $t$ on $U^f$ as deck transformations induces an automorphism of $\calL$, making ${\mathcal L}$ into a local system of rank $1$ free $R$-modules. For any point $x\in U$, the stalks are given by $ {\mathcal L}_x = \bigoplus_{x'\in \pi^{-1}(x)} k$. The monodromy action of a loop $\gamma$ on ${\mathcal L}_x$ interchanges the summands according to the monodromy action of $\gamma$ on $\pi^{-1}(x)$. Therefore, the monodromy of ${\mathcal L}$ is the representation: $$ \begin{array}{ccc} \pi_1(U) & \longrightarrow & \Aut_R(R)\\ \gamma & \mapsto & \left(1\mapsto t^{f_*(\gamma)}\right). \end{array} $$ \begin{remk}\label{remk:isoLocal-Uf} By \cite[Theorem 2.1]{KL}, there are natural isomorphisms of $R$-modules for all $i$: $ H_i(U;\calL)\cong H_i(U^f;k)$. These come from an isomorphism at the level of chain complexes.$\sslash$ \end{remk} \begin{remk}\label{homologyAM} The definition of the homological Alexander modules can be applied to every connected space $Y$ of finite homotopy type and any epimorphism $\pi_1(Y)\twoheadrightarrow \Z$. See for example \cite{DN, budur2018monodromy, Max06, MaxTommy}. Note that every complex algebraic variety has the homotopy type of a finite CW-complex \cite[p.27]{dimca1992hypersurfaces}.$\sslash$ \end{remk} \begin{remk}\label{rem:conjugate} If $M$ is a (sheaf of) $R$-modules, we will use $\ov{M}$ to denote $M$ with the conjugate $R$-module structure, where $t$ acts by $t^{-1}$. This is the same as saying $\ov{M} = M\otimes_R R$, where the tensor is via the map $R\to R$ sending $t$ to $t^{-1}$. Note that every (sheaf of) $R$-modules has a canonical $k$-linear isomorphism $M\cong \ov{M}$, namely $m\mapsto \ov m\coloneqq m\otimes 1$.$\sslash$ \end{remk} \begin{dfn}\label{def:2.2.6} The \bi{$i$-th cohomology Alexander module} of $U$ associated to the algebraic map $f$ is the $R$-module $$ \overline{H^i(U;\calL)}. $$ Note that by flatness there is an isomorphism \[ \ov{H^i(U;{\mathcal L})} = H^i(U;{\mathcal L})\otimes_R R\cong H^i(U;{\mathcal L}\otimes_R R) = H^i(U;\ov{\mathcal L}). \] \end{dfn} \begin{remk}\label{UCT} The Universal Coefficients Theorem (UCT) relates the two notions of Alexander modules as follows (for a proof see Lemma~\ref{lem:UCT}). There is a natural short exact sequence of $R$-modules \[ 0\to \Ext^1_R(H_{i-1}(U;\calL),R)\to H^i(U;\Homm_R(\calL,\ul R)) \to \Hom_R(H_i(U;\calL),R)\to 0. \] Moreover, this short exact sequence splits, but the splitting is not natural. Taking into account Remark~\ref{rem:oppositeDual} below, the middle term can be identified with $\ov{H^i(U;{\mathcal L})}$.$\sslash$ \end{remk} \begin{remk}\label{rem:oppositeDual} Both rank 1 local systems $\Homm_R({\mathcal L},\ul R)$ and $\ov{\mathcal L}$ have monodromy $\gamma\mapsto t^{-f_*(\gamma)}$. Let us make a canonical choice of isomorphism between them. Since ${\mathcal L}=\pi_!\ul k$, for any $x\in U$, the stalk ${\mathcal L}_x$ has a $k$-basis parametrized by $\pi^{-1}(x)$. Let us call this basis $\{\delta_{x'} \}_{x'\in \pi^{-1}(x)}$, and note that $t\delta_{x'} = \delta_{tx'}$. Then we have bases: \[ \ov{\mathcal L}_x = k\langle \ov\delta_{x'}\rangle_{x'\in \pi^{-1}(x)} ;\quad \Homm_R({\mathcal L},\ul R)_x \cong k\langle \delta_{x'}^\wedge\rangle_{x'\in \pi^{-1}(x)}. \] Here $\delta_{x'}^\wedge$ is the element in $\Hom_R({\mathcal L}_x,R)$ mapping $\delta_{x'}\mapsto 1$. Let $\gamma$ be a loop acting on the stalks by the monodromy. Then: \[ \gamma\ov\delta_{x'}= t^{-f_*\gamma}\ov\delta_{x'}; \quad t\ov\delta_{x'} = \ov{\delta}_{t^{-1}x'} ; \quad \gamma\delta_{x'}^\wedge = t^{-f_*\gamma}\delta_{x'}^\wedge;\quad t\delta_{x'}^\wedge = \delta_{t^{-1}x'}^\wedge . \] Therefore, mapping $\ov\delta_{x'}\mapsto \delta_{x'}^\wedge$ on every stalk gives an isomorphism $\ov{\mathcal L}\cong \Homm_R({\mathcal L},\ul R)$. We will freely identify these local systems from now on.$\sslash$ \end{remk} \begin{remk}\label{rem:cohom_not_iso} In general, the cohomological Alexander modules are not isomorphic as $R$-modules to the cohomology of the corresponding infinite cyclic cover. Indeed, $\overline{H^i(U;\calL)}$ is a finitely generated $R$-module for all $i$. However, if $H_i(U^f;k)$ is not a finite dimensional $k$-vector space, then $H^i(U^f;k)$ is not a finitely generated $R$-module. Alternatively, $H^*(U^f;k)$ and $H^*(U;{\mathcal L})$ are the hypercohomology of $U$ with coefficients in $R\pi_*\pi^{-1}\ul k_U=\pi_*\ul k_{U^f}$ and ${\mathcal L}=\pi_!\ul k_{U^f}$ respectively, so they need not be isomorphic. However, the torsion parts of the cohomological Alexander modules and the cohomology of $U^f$ are related in a way which will be made more precise in Proposition~\ref{propcanon}.$\sslash$ \end{remk} \subsection{Universal Coefficient Theorem} In this paper, we work with local systems of $R$-modules. For an $R$-module $M$ (or a complex thereof), we denote $M^\vee = R\Hom_R^\bullet(M,R)$. Similarly, for a sheaf of $R$-modules $\mathcal{M}$ on a topological space $X$ (or a complex thereof), we denote $\mathcal{M}^\vee = R\Homm_R^\bullet(\mathcal{M},\underline{R}_X)$. \begin{dfn}\label{dfn:complex} Let $U$ be a connected locally contractible space, and let ${\mathcal F}^\bullet$ be a bounded complex of local systems of finitely generated \textit{free} $R$-modules on $U$. Let $x\in U$, and denote $\pi_1\coloneqq\pi_1(U,x)$. Let $\wt U$ be any covering space induced by a normal subgroup $H$ of $\pi_1$ on which the inverse image of all the local systems ${\mathcal F}^i$ becomes trivial. Following \cite[VI.3]{whitehead}, we define \begin{itemize} \item $C_\bullet({\mathcal F}^\bullet) \coloneqq C_\bullet(\wt U;R)\otimes_{R[\pi_1/H]} {\mathcal F}^\bullet_x$ \item $C^\bullet({\mathcal F}^\bullet) \coloneqq \Hom_{R[\pi_1/H]}^\bullet(C_\bullet(\wt U;R),{\mathcal F}^\bullet_x)$ \end{itemize} Here $C_\bullet(\wt U;R)$ are the singular chains with coefficients in $R$ with a right action of $\pi_1/H$ by deck transformations. \end{dfn} \begin{remk}\label{remk:dfncomplex} Let ${\mathcal M}$ be a local system of $R$-modules on $U$, and let ${\mathcal F}^\bullet\rightarrow {\mathcal M}$ be a resolution of ${\mathcal M}$, where $U$ and ${\mathcal F}^\bullet$ are as in Definition \ref{dfn:complex}. $C_\bullet({\mathcal F}^\bullet)$ is a chain complex computing the homology of ${\mathcal M}$. Indeed, since $C_\bullet(\wt U;R)$ is a complex of free $R[\pi_1/H]$-modules, the tensor product appearing in the formula defining $C_\bullet({\mathcal F}^\bullet)$ coincides with the left derived tensor product, so $ C_\bullet({\mathcal F}^\bullet)$ is quasiisomorphic to $C_\bullet(\wt U;R)\otimes_{R[\pi_1/H]} {\mathcal M}_x= C_{\bullet}({\mathcal M})$ (the quasiisomorphism given by the resolution ${\mathcal F}^\bullet\rightarrow {\mathcal M}$), which is the complex computing the homology of ${\mathcal M}$ by \cite[VI.3]{whitehead}. Similarly, $C^\bullet({\mathcal F}^\bullet)$ is quasiisomorphic to $\Hom^\bullet_{R[\pi_1/H]}(C_\bullet(\wt U;R), {\mathcal M}_x)= C^{\bullet}({\mathcal M})$ (the quasiisomorphism given by the resolution ${\mathcal F}^\bullet\rightarrow {\mathcal M}$). $C^\bullet({\mathcal M}$) is a chain complex computing the cohomology of ${\mathcal M}$ (by loc. cit.). Again, note that the derived $R\Hom_{R[\pi_1/H]}^\bullet$ coincides with the usual $\Hom_{R[\pi_1/H]}^\bullet$ in the definition of $C^\bullet({\mathcal F}^\bullet)$, by the freeness of $C_\bullet(\wt U;R)$. $\sslash$\end{remk} The Universal Coefficients Theorem will be important in Section \ref{sec:main}. Below is the version that we will need to use, which contains the definitions of the maps $\UCT_R$ and $\UCT_{k}$. We include the proof, since we will later make use of the spectral sequence that appears in it, and we have not been able to find this version of the Universal Coefficients Theorem in the literature. When using homological notation, we let $H_j=H^{-j}$. \begin{lem}[Definition of the maps $\UCT_R$ and $\UCT_k$]\label{lem:UCT} Let $U$ be a locally contractible space, and let ${\mathcal F}^\bullet$ be a bounded complex of local systems of finitely generated \textit{free} $R$-modules. Let $x\in U$, and let $\pi_1\coloneqq\pi_1(U,x)$. Since $R$ is a PID, we have the following natural Universal Coefficients Theorem short exact sequence \[ 0\to \Ext^1_R(H_{i-1}(C_\bullet(({\mathcal F}^\bullet)^\vee));R) \to H^i(C^\bullet({\mathcal F}^\bullet)) \to \Hom_R(H_i(C_\bullet(({\mathcal F}^\bullet)^\vee)),R)\to 0, \] which gives us a natural isomorphism $$\UCT_R \colon \Ext^1_R(H_{i-1}(C_\bullet(({\mathcal F}^\bullet)^\vee));R) \xrightarrow{\sim} \Tors_R H^{i}(C^\bullet({\mathcal F}^\bullet)).$$ Now, suppose that ${\mathcal F}^\bullet$ is a single local system ${\mathcal M}$ located at degree 0. If the stalk of ${\mathcal M}$ at $x$ is a finitely generated \textit{torsion} $R$-module, we can consider ${\mathcal M}$ as a local system of $k$-vector spaces. Analogously to the short exact sequence above, we can replace $R$ by $k$ and ${\mathcal F}^\bullet$ by ${\mathcal M}$, to obtain the natural isomorphism $\UCT_k\colon H^i(U;{\mathcal M}) \xrightarrow{\sim} \Hom_{k} ( H_i( U;\Homm_{k}({\mathcal M},k)), k)$. \end{lem} \begin{proof} By tensor-hom adjunction, we have the following isomorphism. \[ C^\bullet({\mathcal F}^\bullet) = \Hom_{R[\pi_1/H]}^\bullet\left(C_\bullet(\wt U;R),\Hom^\bullet_R(({\mathcal F}^\bullet)^\vee_x,R)\right) \cong\] \[ \cong \Hom_{R}^\bullet\left(C_\bullet(\wt U;R)\otimes_{R[\pi_1/H]}({\mathcal F}^\bullet)^\vee_x ,R\right) = \Hom_{R}^\bullet\left(C_\bullet(({\mathcal F}^\bullet)^\vee) ,R\right). \] Note that we are using $({\mathcal F}^\bullet)^\vee = R\Homm_R^\bullet({\mathcal F}^\bullet,\underline R)$, but since ${\mathcal F}^\bullet$ is a complex of local systems of free $R$-modules, then we have $({\mathcal F}^\bullet)^\vee = \Homm_R^\bullet({\mathcal F}^\bullet,\underline R)$, which is again a bounded complex of local systems of free $R$-modules. We are also using that for bounded complexes of local systems, $\Homm_R^\bullet$ commutes with taking stalks. Now, since $R$ is a PID, we can proceed as in the proof of the Universal Coefficient Theorem in \cite[Chapter 3.1]{hatcher}, to arrive at the natural Universal Coefficients short exact sequence from the statement of the lemma. \begin{equation}\label{eq:UCTses} 0\to \Ext^1_R(H_{i-1}(C_\bullet(({\mathcal F}^\bullet)^\vee));R) \to H^i(C^\bullet({\mathcal F}^\bullet)) \to \Hom_R(H_i(C_\bullet(({\mathcal F}^\bullet)^\vee)),R)\to 0. \end{equation} Alternatively, we can obtain this short exact sequence using a spectral sequence, namely the one obtained from the Grothendieck spectral sequence for the composition of two derived functors $RF\circ RG$, taking $G = \Id$ and $F = \Hom_R(\bullet, R)$. Later on we will need to worry about what the maps are, so let us recall how it is obtained. Let $K=k(t)$ be the field of fractions of $R$, viewed as an $R$-module. Let $I^\bullet$ be the following injective resolution of $R$, with the obvious maps: \[ R\to \underset{(=I^0)}{K}\to \underset{(=I^1)}{K/R.} \] Consider the map of complexes induced by the map $I^0\to I^1$: \[ \Hom_R^\bullet(C_\bullet(({\mathcal F}^\bullet)^\vee),I^0) \to \Hom_R^\bullet(C_\bullet(({\mathcal F}^\bullet)^\vee),I^1). \] We can turn this map into a double complex where each of the complexes above is a row, and the map between them (with some sign changes) becomes the vertical differential. We consider the spectral sequences associated to this double complex. Consider the spectral sequence that starts by taking the vertical cohomology. Since $R\to I^\bullet$ is an injective resolution, the cohomology of the columns is $\Ext^i(C_\bullet(({\mathcal F}^\bullet)^\vee),R)$, which vanishes if $i\neq 0$ since $C_\bullet(({\mathcal F}^\bullet)^\vee)$ is a complex of free $R$-modules. We are left with only one nonzero row, namely the complex $\Hom^{\bullet}_R(C_\bullet(({\mathcal F}^\bullet)^\vee),R)$. Therefore, the spectral sequence converges in the second page, to the cohomology of $\Hom^{\bullet}_R(C_\bullet(({\mathcal F}^\bullet)^\vee),R)$. In other words, the cohomology of the total complex is the cohomology of $C^\bullet({\mathcal F}^\bullet)$. Now we consider the other spectral sequence, starting with the horizontal cohomo\-logy. Since $I^i$ is injective, there are natural isomorphisms: \[ H^j(\Hom_R^\bullet(C_\bullet(({\mathcal F}^\bullet)^\vee),I^i)) \cong \Hom_R(H_j(C_\bullet(({\mathcal F}^\bullet)^\vee)),I^i). \] Since $I^\bullet$ is an injective resolution of $R$, we have isomorphisms as follows: \[ H^i(\Hom_R(H_j(C_\bullet(({\mathcal F}^\bullet)^\vee)),I^\bullet)) \cong \Ext^i_R(H_j(C_\bullet(({\mathcal F}^\bullet)^\vee)),R ) . \] Further, the map $R\to I^\bullet$ induces maps $R\to K$ and $\beta\colon K/R[-1]\to R$, which means that the isomorphisms above are induced by the inclusion $R\to K$ if $i=0$ and by $\beta$ if $i=1$ (we will need to use this fact in Proposition~\ref{prop:mapsAreEqual}). At this point, the spectral sequence has converged (just by looking at the degrees), yielding a filtration of $H^{i}(C^\bullet({\mathcal F}^\bullet))$ by the cohomology groups above. In other words, we obtain the desired short exact sequence~(\ref{eq:UCTses}). The isomorphism $\UCT_k$ is obtained by replacing $R$ by $k$, and ${\mathcal F}^\bullet$ by ${\mathcal M}$ in the proof above. \end{proof} \begin{remk}\label{remk:UCT} If we take ${\mathcal F}^\bullet=\ov{\mathcal L}$ seen as a complex in degree $0$, we get that $({\mathcal F}^\bullet)^\vee\cong {\mathcal L}$. Taking into account Remark \ref{remk:dfncomplex}, the UCT short exact sequence in Lemma \ref{lem:UCT} (in $R$) becomes $$ 0\to \Ext^1_R(H_{i-1}(U;{\mathcal L});R) \to H^i(U;\ov{\mathcal L}) \to \Hom_R(H_i(U;{\mathcal L}),R)\to 0, $$ that is, the UCT short exact sequence in Remark~\ref{UCT}.$\sslash$ \end{remk} \subsection{Cohomology Alexander modules and duality}\label{ss:duality} The relation between the cohomology Alexander modules and the corresponding infinite cyclic cover can be made more precise as follows. \begin{prop} \label{propcanon} There is a natural $R$-module isomorphism between $\Tors_R H^i(U;\ov\calL)$ and $(\Tors_R H_{i-1}(U^f;k))^{\vee_k}$, where $^{\vee_k}$ denotes the dual as a $k$-vector space. Moreover, this isomorphism is functorial in the following sense. Let $U_1$ and $U_2$ be smooth connected complex algebraic varieties, with algebraic maps $f_j\colon U_j\rightarrow \C^*$ such that $f_j$ induces an epimorphism in fundamental groups for $j=1,2$, and assume that there exists an algebraic map $g\colon U_1\rightarrow U_2$ that makes the following diagram commutative. \begin{center} \begin{tikzcd}[row sep = 1em] U_1\arrow[rd,"f_1"']\arrow[rr, "g"] & \ & U_2\arrow[ld,"f_2"] \\ \ & \C^* & \ \end{tikzcd} \end{center} Let $\pi_j\colon U_j^{f_j}\to U_j$ be the corresponding infinite cyclic covers, and let $\calL_j = \pi_! \ul k_{U_j^{f_j}}$ be the local systems of $R$-modules induced by $f_j$ for $j=1,2$. Then, $g$ induces a commutative diagram of $R$-modules compatible with the above isomorphism as follows \begin{center} \begin{tikzcd}[row sep = 1.4em] \Tors_RH^i(U_2;\ov{\calL_2})\arrow[r,"\cong"]\arrow[d, "g^*"]& (\Tors_R H_{i-1}(U_2^{f_2},k))^{\vee_k}\arrow[d,"(g_*)^{\vee_k}"] \\ \Tors_R H^i(U_1;\ov{\calL_1})\arrow[r,"\cong"]& (\Tors_R H_{i-1}(U_1^{f_1},k))^{\vee_k}. \end{tikzcd} \end{center} \end{prop} \begin{proof} By the UCT (Remark~\ref{remk:UCT}), there is a short exact sequence of $R$-modules \[ 0\to \Ext^1_R\big(H_{i-1}(U;\calL),R\big)\to H^i(U;\ov\calL) \to \Hom_R\big(H_i(U;\calL),R\big)\to 0. \] Since $R$ is a PID, the $R$-module $\Hom_R(H_i(U;\calL),R)$ is free and $\Ext^1_R(H_{i-1}(U;\calL),R)$ is a torsion $R$-module. Moreover, there is a natural isomorphism of $R$-modules \[\Ext^1_R(H_{i-1}(U;\calL),R)\cong\Ext^1_R(\Tors_R H_{i-1}(U^f;k),R).\] Therefore, we have a natural isomorphism of $R$-modules \begin{equation} \Tors_R H^i(U;\ov \calL)\cong \Ext^1_R(\Tors_R H_{i-1}(U^f;k),R). \label{eqn1} \end{equation} Using equation (\ref{eqn1}) and applying Lemma \ref{lemcanon} (see below) to the finitely generated torsion $R$-module $\Tors_R H_{i-1}(U^f;k)$ yields a natural isomorphism of $R$-modules $$ \Tors_R H^i(U;\ov\calL)\xrightarrow{\cong}(\Tors_R H_{i-1}(U^f;k))^{\vee_k}, $$ concluding our proof of the first part of this proposition. The functoriality follows from the functoriality of the UCT and Lemma \ref{lemcanon}. \end{proof} \begin{lem} Let $A$ be a finitely generated torsion $R$-module. Then, there exists a natural isomorphism of $R$-modules $$ \Res\colon\Ext^1_R(A,R)\xrightarrow{\cong} A^{\vee_k}, $$ where $^{\vee_k}$ denotes the dual as a $k$-vector space. \label{lemcanon} \end{lem} This lemma is essentially a special case of the relative local duality theorem of Smith, see \cite[Theorem 1.5]{localduality}. Since we are only using a very special version of the local duality theorem, we sketch a more elementary proof. We first make the following definition, which we will use in the proof. \begin{dfn}\label{dfn:Res} Let $K = k(t)$, and let $\beta\colon K/R[-1]\to R$ be the map in the derived category of $R$-modules given by the floor diagram: \[ \begin{tikzcd}[row sep = 1.4em] 0 \arrow[d]& K\arrow[r,"1",leftarrow]\arrow[l,leftarrow]\arrow[d,"1"]\arrow[dr,phantom,"\sim"] & R\arrow[d]\\ K/R & K/R \arrow[l,"1",leftarrow]\arrow[r,leftarrow]& 0. \end{tikzcd} \] For an $R$-module $M$, this induces a map \[ R^1\Hom_R^\bullet(M,\beta)\colon \Hom_R(M,K/R)\to \Ext^1_R(M,R). \] Let $\res_*$ be the composition $\Hom_R(\cdot,K/R)\to \Hom_k(\cdot,K/R)\to \Hom_k(\cdot,k)$, where the first arrow is just the forgetful arrow and the second arrow is the residue of a rational function (see below for details). For every torsion $R$-module $A$, the isomorphism of $R$-modules $\Res$ is defined as $\Res=\res_*\circ (R^1\Hom_R^\bullet(A,\beta))^{-1}\colon\Ext^1_R(A,R)\rightarrow \Hom_k(A,k)$. That is, it is the composition of the inverse of the map $\Hom_R(A,K/R)\rightarrow \Ext^1_R(A,R)$ in the long exact sequence corresponding to $R\Hom_R(A,\cdot)$ applied to the short exact sequence of $R$-modules $0\rightarrow R\rightarrow K\rightarrow K/R\rightarrow 0$ with $\res_*$. \end{dfn} \begin{proof}[Sketch of proof of Lemma~\ref{lemcanon}] First, we prove that $\Res$ is an isomorphism assuming that $k$ is algebraically closed. Since $R$ is a PID, the following is an injective resolution of $R$: \[ 0\rightarrow R\rightarrow K \rightarrow K/R\rightarrow 0. \] By definition, there is a natural isomorphism $ \Ext^1_R(A, R)\cong \Hom_R(A,K/R)/\Hom_R(A,K)$. Since $A$ is a torsion $R$-module, $\Hom_R(A,K)=0$ and hence we have a natural isomorphism \begin{equation}\label{eqiso} \Ext^1_R(A, R)\cong \Hom_R(A,K/R). \end{equation} We next define a $k$-linear map $\mathrm{res}\colon K/R\to k$ as follows. Using the division algorithm and partial fraction decomposition, every element $b\in K$ can be written uniquely as a finite sum \[ b=v(t)+\sum_{i=1}^r\sum_{j=1}^{q_i}\frac{\alpha_{i,j}}{(t-\beta_i)^j}, \] where $v(t)\in R$, and $\alpha_{i,j}\in k, \beta_i\in k\setminus \{0\}$ are constants. We define $\mathrm{res}(b)=\sum_{i=1}^{r}\alpha_{i, 1}$. Composing with the $k$-linear map $\mathrm{res}$, we get a $k$-linear map \[ \mathrm{res}_*\colon \Hom_R(A,K/R)\to \Hom_k(A, k)=A^{\vee_k}. \] One can check that this is a homomorphism of $R$-modules. In fact, it is easy to check this after localizing at each $\beta_i\in k\setminus \{0\}=\mathrm{mSpec}(R)$. Since $A$ is a finitely generated torsion $R$-module, both $\Hom_R(A,K/R)$ and $\Hom_k(A, k)$ decompose into direct sums of the localizations. Next, we prove that $\mathrm{res}_*$ is an isomorphism of $R$-modules. We claim that $\ker(\mathrm{res})\subset K/R$ does not contain any non-zero $R$-submodule $M$. In fact, without loss of generality, we can assume that $M$ is supported at one point $\beta_1$. Then a nonzero element of $M$ is of the form $\big[\frac{w(t)}{(t-\beta_1)^q}\big]$, where $q\geq 1$ and $w(\beta_1)\neq 0$. Now, \[ (t-\beta_1)^{q-1}\cdot \Big[\frac{w(t)}{(t-\beta_1)^q}\Big]=\Big[\frac{w(t)}{t-\beta_1}\Big]\in M \] and $\mathrm{res}(\frac{w(t)}{t-\beta_1})=w(\beta_1)\neq 0$. Thus, the above claim follows, which then implies that $\mathrm{res}_*$ is injective. Using the primary decomposition of finitely generated $R$-modules, it is easy to see that as $k$-vector spaces, $\Hom_R(A,K/R)$ and $\Hom_k(A, k)$ have the same dimensions. Therefore $\mathrm{res}_*$ is an isomorphism. Hence, by equation \eqref{eqiso}, we have a functorial $R$-module isomorphism \[ \Ext^1_R(A,R)\xrightarrow{\cong} A^{\vee_k}. \] When $k$ is not algebraically closed, we let $\bar{k}$ be its algebraic closure. The above arguments show that there exists a natural isomorphism \begin{equation}\label{eqkbar} \Ext^1_{R\otimes_k \bar{k}}(A\otimes_k \bar{k},R\otimes_k \bar{k})\xrightarrow{\cong} (A\otimes_k \bar{k})^{\vee_k}. \end{equation} Notice that the division algorithm and the partial fraction decomposition are invariant under the Galois action of $\mathrm{Gal}(\bar{k}/k)$. Therefore, the isomorphism \eqref{eqkbar} descends to an isomorphism over $k$. \end{proof} \begin{remk}\label{rem:annoyingSign} Let $\beta$ be as in Definition~\ref{dfn:Res}. In the derived category, there is an exact triangle: \[ R\xrightarrow{1} K\xrightarrow{1} K/R \xrightarrow{\beta[1]} R[1]. \] Here $\beta[1]$ is the following floor diagram. Note the sign: \[ \begin{tikzcd}[row sep = 1.4em] 0 \arrow[d]& K\arrow[r,"1",leftarrow]\arrow[l,leftarrow]\arrow[d,"-1"] & R\arrow[d]\\ K/R & K/R \arrow[l,"1",leftarrow]\arrow[r,leftarrow]& 0. \end{tikzcd} \] This sign means that the map $\Hom_R(\cdot ,K/R)\to \Ext^1_R(\cdot,R)$ appearing in the obvious Ext long exact sequence is opposite to the map in the derived category induced by $\beta[1]$. We will use the former map, but this sign won't affect any of our computations in this paper.$\sslash$ \end{remk} \begin{cor} \label{isocohom} In the above notations, assume moreover that $H_{i}(U^f;k)$ is a torsion $R$-module for some $i\geq 0$. Then, there exists a canonical isomorphism $$ \Tors_R H^{i+1}(U;\ov\calL)\cong H^{i}(U^f;k). $$ Moreover, if $H_{i+1}(U^f;k)$ is also a torsion $R$-module, then, so is $H^{i+1}(U;\calL)$. Hence, in that case, $H^{i+1}(U;\ov\calL)$ and $H^{i}(U^f;k)$ are naturally isomorphic. \end{cor} \begin{proof} \[ \Tors_R{H^{i+1}(U;\ov{\mathcal L})} \overset{(\ref{propcanon})}\cong \Hom_k(\Tors_R H_{i}(U^f;k),k) = \Hom_k( H_{i}(U^f;k),k) \overset{\UCT_k}\cong H^i(U^f;k). \] The rest of the proof follows from the UCT (written as in Remark~\ref{remk:UCT}). \end{proof} The result of the above corollary applies, in particular, to the special case when $U=\C^n \setminus \{f=0\}$ and $i \neq n$, with $f\colon U\to \C^*$ induced by a complex polynomial in general position at infinity; see \cite{DL,Max06}. \subsection{Relationship with the generic fiber}\label{genfiber} Let $U,f,\pi$ be as in Section~\ref{ssAlex}. By Verdier's generic fibration theorem \cite[Corollary 5.1]{verdier}, there exists a finite set of points $\calB\subset\C^*$ such that $ f\colon f^{-1}(\C^*\setminus\calB)\longrightarrow \C^*\setminus\calB$ is a locally trivial fibration. Being the pullback of a locally trivial fibration, $f_\infty$ is a locally trivial fibration away from $\exp^{-1}(\calB)$ as well. \begin{dfn}\label{def:genericFiber} We will say a fiber $f_\infty^{-1}(c)$ (resp. $f^{-1}(e^c)$) is generic if $c\notin\exp^{-1}(\calB)$ (resp. if $e^c\notin\calB$). \end{dfn} Let $F$ be a generic fiber of $f$. For any $c\in \C$, $\pi$ is an isomorphism between a neighborhood of $f_\infty^{-1}(c)$ and a neighborhood of $f^{-1}(e^c)$ (recall that $\pi$ is the pullback of $\exp\colon\C\to \C^*$). In particular, if $F = f^{-1}(e^c)$, there is a unique lift $i_{\infty}\colon F \hookrightarrow U^f$ of the inclusion $i\colon F\hookrightarrow U$ inducing an isomorphism $F\cong f_\infty^{-1}(c)$. Let us recall the description of the monodromy action of $F$, along with its compatibility with the deck action on $U^f$. \begin{lem}\label{lem:fiberMonodromy} For any $j\geq 0$, the map $i_{\infty}$ induces a map: \[ (i_{\infty})_j\colon H_j(F;k)\longrightarrow H_j(U^f;k). \] This map is an $R$-module homomorphism if we let $t\in R$ act on the right hand side as the deck transformation of $U^f$ and on the left-hand side as the \emph{clockwise} monodromy of $f$, i.e. the action of $-1\in \Z\cong \pi_1(\C^*)$. Further, $H_j(F;k)$ is finite dimensional, so it is necessarily a torsion $R$-module. Therefore, the image of $(i_{\infty})_j$ is contained in $\Tors_R H_j(U^f;k) = A_j(U^f;k)$. \end{lem} \begin{proof} Suppose we have any two generic fibers $F_1,F_2$, with $F_i = f_{\infty}^{-1}(c_i)$, and a path $\gamma$ connecting $c_1$ with $c_2$ in $\C\setminus \exp^{-1}(\calB)$. Then, $f_\infty$ restricted to $f^{-1}_\infty(\gamma)$ is a locally trivial fibration over a contractible base, and therefore $f^{-1}_\infty(\gamma)$ deformation retracts to any fiber. In particular, we obtain a homotopy equivalence $F_1\simeq F_2$. Therefore, we are free to choose a fixed fiber $F\subseteq U^f$ such that $\exp(f_\infty(F))$ is contained in a small neighborhood of $0$. Recall that we can use $\pi$ to identify $F\cong \pi(F)$. Let $c_0 = f_\infty(F)$. Let $\gamma$ be a counterclockwise circle centered at $0$ passing through $e^{c_0}$. By definition, the monodromy action of $\gamma$ on $\pi(F)\cong F$ is given as follows. Lifting $\gamma$ to a line segment $\wt \gamma = [c_0,c_0+2\pi i]\subset \C$, we use the same deformation retraction argument from above to obtain the monodromy map as the following composition from left to right (defined up to homotopy): \[ \pi(F)\xleftarrow[\cong]{\pi} f_\infty^{-1}(c_0) \xhookrightarrow[\simeq]{} f_\infty^{-1}(\wt\gamma) \xhookleftarrow[\simeq]{} f_\infty^{-1}(c_0+2\pi i) \xrightarrow[\cong]{\pi} \pi(F). \] Let us see how this relates to the deck action. The preimage $\exp^{-1}(\gamma)$ is a line, and therefore it is contractible. Therefore, $\wt F\coloneqq f_\infty^{-1}(\exp^{-1}(\gamma))$ deformation retracts to all its fibers. Fix the homotopy equivalence $\alpha\colon\pi(F)\cong f_\infty^{-1}(c_0)\xhookrightarrow{\simeq} \wt F$. The deck transformation $t$ acts on $\wt F$ by restricting the action on $U^f$. Consider the following diagram. \[ \begin{tikzcd}[row sep = 1.6em] \pi(F) \arrow[dd,"\alpha"',bend right = 50]\arrow[r,equals]& \pi(F)\arrow[from=r,"\gamma"']& \pi(F) \arrow[dd,"\alpha",bend left = 50]\\ f_\infty^{-1}(c_0)\arrow[u,"\pi","\cong"']\arrow[d,hookrightarrow,"\simeq"] \arrow[r,"t"]& f_\infty^{-1}(c_0+2\pi i)\arrow[d,hookrightarrow,"\simeq"] \arrow[u,"\pi","\cong"']& f_\infty^{-1}(c_0)\arrow[u,"\pi","\cong"']\arrow[d,hookrightarrow,"\simeq"] \\ \wt F\arrow[r,"t"]& \wt F\arrow[r,equals] & \wt F. \end{tikzcd} \] Here $\gamma$ denotes the homotopy action of $\gamma$, as defined above. The diagram is commutative, except for the right hand rectangle, which only commutes up to homotopy (by definition of the monodromy). From this diagram we see that on $\pi(F)$, $\alpha^{-1}\circ t \circ \alpha = \gamma^{-1}$, as desired. Choosing a different lift $F$ of $\pi(F)$, we could repeat this proof to obtain the same result. \end{proof} \begin{prop}\label{genf} Under the notations of Lemma~\ref{lem:fiberMonodromy}, the homomorphism $(i_{\infty})_j\colon H_j(F;k)\longrightarrow \Tors_R H_j(U^f;k)$ is an epimorphism for all $j\geq 0$. In other words, the $R$-module $H_j(F;k)$ maps onto to the torsion part of the $j$-th homology Alexander module. Moreover, the kernel of $(i_{\infty})_j$ can be described as follows: for any $b\in \C^*$, let $T_b=f^{-1}(D_b)$, where $D_b$ is a small disk around $b$. Using a path to identify $F$ with a general fiber in $T_b$, let \[ K_b = \ker \left( H_j(F;k)\to H_j(T_b;k) \right). \] The kernel of $(i_{\infty})_j$ is the $R$-module generated by the subgroups $K_b$ for $b\in \calB$, where $\calB\subset \C^*$ consists of the points over which $f$ is not a locally trivial fibration. \end{prop} \begin{proof} We use an argument similar to \cite[2.3]{DN2}. As in the proof of Lemma \ref{lem:fiberMonodromy}, we may choose $F$ to be a generic fiber of $f_\infty$ such that $|\exp(f_\infty(F))|\ll 1$. Recall that $\pi$ induces an isomorphism between $F$ and a generic fiber of $f$. For all $c\in \C$, let $D_c$ be a small disk centered at $c$, $D_c^*=D_c\setminus\{c\}$, $T_c\coloneqq f^{-1}(D_c)$ and $T_c^*\coloneqq f^{-1}(D_c^*)$. There is a deformation retraction of $\C^*$ to a subspace we will call $X$, which is the union of $D_0^*$ and $D_b$ for $b\in \calB$, together with a path connecting $D_0^*$ with $D_b$ for each $b$. We choose these paths to be contractible and pairwise disjoint. Then, $\C$ has a deformation retraction to $\exp^{-1}(X)$, which is the union of $D_{\wt b}$ for $\wt b\in \exp^{-1}(\calB)$, the preimages of all the paths in $X$ and a left half-plane, which is the preimage of $D_0^*$. We can further deformation retract $\C$ to $\wt X$, defined as follows: for every $\wt b\in \exp^{-1}\calB$, fix a contractible path from $f_\infty(F)$ to the boundary of $D_{\wt b}$, so that all the paths are disjoint away from $f_\infty(F)$. Let $\wt X$ be the union of all of the $D_{\wt b}$'s and all those paths joining each of the disks to $f_\infty(F)$. It will later be convenient for us to use closed disks to construct $\wt X$. Recall that $f_\infty$ is a locally trivial fibration away from $\exp^{-1}(\calB)$. Let $Y = f_\infty^{-1}(\wt X)$. Using the fibration, we have that $U^f$ deformation retracts to $Y$. Therefore, the inclusion $Y\to U^f$ induces isomorphisms for all $i$: \[ H_j(Y, F;k)\cong H_j(U^f, F;k). \] Now we consider two contractible neighborhoods $V_1\subsetneq V_2$ of $f_{\infty}(F)$ in $\wt X$: let $V_1$ be the complement of the closed disks around each $\wt b\in \calB$ (i.e. a wedge sum of segments), and let $V_2$ be a contractible open neighborhood of $\ov V_1$ that doesn't intersect $\exp^{-1}\calB$. For example, $V_2$ can be taken to be the union of $V_1$ with a small open disk around $\ov V_1\cap D_{\wt b}$ for every $\wt b\in \exp^{-1}(\calB)$. Since $V_i$ are both contractible and $f_\infty$ is a locally trivial fibration over them, we have that the inclusions are homotopy equivalences $F\simeq f_\infty^{-1}(V_1) \simeq f_\infty^{-1}(V_2)$. Therefore, by these equivalences and by excision, we have isomorphisms induced by inclusions: \[ H_j(Y,F;k) \cong H_j(Y,f_\infty^{-1}(V_2);k) \cong H_j(Y\setminus f_\infty^{-1}(V_1),f_\infty^{-1}(V_2\setminus V_1);k). \] Since $\wt X\setminus V_1$ is a disjoint union of $D_{\wt b}$, \[ H_j(Y,F;k) \cong \bigoplus_{\wt b\in \exp^{-1}(\calB)} H_j(f^{-1}_{\infty}(D_{\wt b}),f_\infty^{-1}(V_2\cap D_{\wt b}) ;k).\ Over $V_2\cap D_{\wt b}$, $f_\infty$ is a trivial fibration (since the base is contractible). Therefore, any path connecting $f_\infty (F)$ with some point $c\in V_2\cap D_{\wt b}$ induces a homotopy equivalence $F\simeq f_\infty^{-1}(c)$, which in turn is homotopy equivalent to $f_\infty^{-1}(V_2\cap D_{\wt b})$. If we just denote $F=f_\infty^{-1}(c)$, we have \[ \bigoplus_{\wt b\in \exp^{-1}(\calB)} H_j(f^{-1}_{\infty}(D_{\wt b}),f_\infty^{-1}(V_2\cap D_{\wt b}) ;k) \cong \bigoplus_{\wt b\in \exp^{-1}(\calB)} H_j(f^{-1}_{\infty}(D_{\wt b}),F ;k) . \] Now we consider the $R$-module structure on the cohomology of the pair $(U^f,F)$. If $F = f_\infty^{-1}(c)$ for $\Re c \ll 0$ ($\Re$ denotes the real part), $F$ is homotopy equivalent to $\wt F \coloneqq f_{\infty}^{-1}( \{z \mid \Re z = \Re c \})$. Then the deck transformations ($z\mapsto z+2\pi i$) act on the pair $(U^f,\wt F)$, making the maps in the long exact sequence $R$-linear. Recall from Lemma~\ref{lem:fiberMonodromy} that the deck action on $H_j(\wt F;k)$ coincides with the monodromy action on $H_j(F;k)$. Note that the action of $t\in R$ on $U^f$ restricts to $\bigsqcup_{\wt b\in \exp^{-1}(\calB)}(f^{-1}_{\infty}(D_{\wt b}),F ;k)$: we can choose the disks to be $\Z$-invariant, and we can choose the paths identifying the fibers so that the monodromy action agrees with the action on $F$. Therefore, the homology group above is an $R$-module. Further, it is a free $R$-module, since $f^{-1}_{\infty}(D_{\wt b}) \cong f^{-1}(D_b)\times \Z$: \[ \bigoplus_{\wt b\in \exp^{-1}(\calB)} H_j(f^{-1}_{\infty}(D_{\wt b}),F ;k) \cong \bigoplus_{b\in \calB} R\otimes_k H_j(f^{-1}(D_{ b}),F ;k). \] We have the long exact sequence of the pair $(U^f,F)$: \[ H_{j+1}(U^f,F;k)\to H_j(F;k)\to H_j(U^f;k)\to H_j(U^f,F;k) . \] Since the last group is free, the restriction $\Tors_R H_j(U^f;k)\to H_j(U^f,F;k)$ must vanish. We have seen above that the inclusion induces an isomorphism $H_{j+1}(U^f,F;k) \cong \bigoplus_{b\in \calB} H_{j+1}(\Z\times T_b,F;k) $, so the sequence above restricts to an exact sequence of $R$-modules: \[ \bigoplus_{b\in \calB} R\otimes_k H_{j+1}(T_b,F;k)\to H_j(F;k)\to \Tors_R H_j(U^f;k)\to 0. \] By the exactness, the kernel of $(i_{\infty})_j$ is the image of the first map. This image is the $R$-module generated by $H_{j+1}(T_b,F;k)$ for all $b\in \C^*$ (note that for $b\notin \calB$, this homology group vanishes). Therefore, it is enough to describe the image of: \[ H_{j+1}(T_b,F;k) \to H_j(F;k). \] By the long exact sequence of the pair $(T_b,F)$, the image of the above map is $K_b$ \end{proof} \begin{remk}\label{remk:tube} The map in Proposition~\ref{genf} factors through $H_j((T^*)^f;k)$, where $T^*=f^{-1}(D^*)$ for $D^*$ a punctured disk centered at $0$ in $\C$. We choose $D^*$ to be small enough that the restriction of $f$ to $T^*$ is a fibration, so $(T^*)^f$ is homotopy equivalent to $F$, and the epimorphism $H_j(F; k)\twoheadrightarrow \Tors_{R}H_j(U^f;k)$ factors as \begin{equation}\label{eqn:tubeTorsion} H_j(F; k)\xrightarrow{\cong}H_j((T^*)^f;k)\twoheadrightarrow \Tors_{R}H_j(U^f;k), \end{equation} where the second morphism doesn't depend on the choice of base points. Now, if we take the $k$-dual of $(i_\infty)_j$, we get a monomorphism of $R$-modules $$ \big(\Tors_{R}H_j(U^f;k)\big)^\vee\hookrightarrow \big(H_j(F; k)\big)^\vee. $$ By Proposition \ref{propcanon} and the isomorphism $(H_j(F; k))^\vee\cong H^j(F;k)$, we get a monomorphism of $R$-modules, only depending on a choice of lift of the base point: \begin{equation}\label{eqn:monoCohom} {\Tors_R H^{j+1}(U;\overline\calL)}\hookrightarrow H^j(F;k). \end{equation} By the isomorphism in (\ref{eqn:tubeTorsion}), $H_j((T^*)^f;k)$ is finite dimensional, so it is a torsion $R$-module. Hence, by Corollary~\ref{isocohom}, $H^j((T^*)^f;k)\cong H^{j+1}(T^*;\overline\calL)$ (note that all the results of Section~\ref{ss:duality} are purely topological, so they can be applied to $T^*$). This implies that for $D^*$ small enough, the monomorphism in (\ref{eqn:monoCohom}) factors in this way: $$ {\Tors_R H^{j+1}(U,\overline\calL)}\hookrightarrow {H^{j+1}(T^*;\overline\calL)}\xrightarrow{\cong}H^j(F;k), $$ where the first map is given by restriction and does not depend on the choice of base points.$\sslash$ \end{remk} \begin{remk} If $f$ is proper, then Proposition~\ref{genf} can be expressed using the nearby cycles of $f$ at $b\in \C^*$, as follows. By \cite[Part II, Section 6.13]{stratifiedmorsetheory}, $T_b$ has a deformation retraction to $F_b=f^{-1}(b)$, and we can compose with the inclusion of $F$ into $T_b$ to obtain the specialization map in cohomology $H^j(F_b;k)\to H^j(F;k)$. Letting $\psi_{f-b} \ul k$ be the nearby cycles of $f$ supported on $f=b$, there is an isomorphism $H^j(F;k) \cong H^j(F_b;\psi_{f-b} \ul k)$ by \cite[Expos\'e XIII p.103]{SGA72}. Under this isomorphism, the specialization map agrees with the map $H^j(F_b;k)\to H^j(F_b;\psi_{f-b} \ul k)$ induced by the canonical morphism $i^{-1}_b\ul k\to \psi_{f-b} \ul k$, where $i_b\colon F_b\hookrightarrow U$ is the inclusion. Let $K_b^{\vee_k}\subseteq H^j(F;k)$ be the image of the specialization map. Using Proposition~\ref{propcanon}, we have the dual map $\left((i_{\infty})_j\right)^{\vee_k}\colon\Tors_R H^{j+1}(U;\ov{\mathcal L})\to H^j(F;k)$. The image of $\left((i_{\infty})_j\right)^{\vee_k}$ is the largest $R$-module contained in $K_b^{\vee_k}$ for all $b\in \calB$.$\sslash$ \end{remk} \subsection{Monodromy action on Alexander Modules and finite cyclic covers}\label{sscover} Let $k=\C$, let $U$ be a smooth connected complex algebraic variety, and let $f\colon U\rightarrow \C^*$ be an algebraic map inducing a local system $\calL$ of $R$-modules as in Section~\ref{ssAlex}. Since $R$ is a PID, we have the primary decomposition $$ A_i(U^f;\C)=\Tors_R H_i(U;\calL)\cong \bigoplus_{i=1}^r R/\big((t-\lambda_i)^{p_i}\big) $$ with $p_i\geq 1$ for all $i=1,\ldots, r$. The set $\{\lambda_i \in \C \mid i=1, \ldots, r\}$ is uniquely determined by $A_i(U^f;\C)$. \begin{prop}\label{eig1} Every $\lambda_i$ defined above is a root of unity. \label{monodromy} \end{prop} \begin{proof} When $U$ is quasi-projective, this fact is proved in \cite[Proposition 1.4]{budur2018monodromy} by reducing to a structure theorem for the cohomology jump loci of a smooth quasi-projective variety \cite[Theorem 1.1]{budur2015jump}. The structure theorem of cohomology jump loci is generalized to arbitrary smooth complex algebraic varieties in \cite[Theorem 1.4.1]{budur2017absolute}. So the statement of \cite[Proposition 1.4]{budur2018monodromy} applies to any smooth complex algebraic variety. \end{proof} In particular, applying Proposition \ref{propcanon}, one has the following. \begin{cor} Let $k=\C$. The eigenvalues of the action of $t$ on $\Tors_R H^*(U;\ov\calL)$ are all roots of unity. \label{roots} \end{cor} The main goal of this paper is to construct a natural mixed Hodge structure on $\Tors_R H^*(U;\ov\calL)$ (for $k=\Q,\R$). For reasons that will become apparent later on, we would like to be able to reduce this problem to the case where the eigenvalues of the action of $t$ on $\Tors_R H^*(U;\ov\calL)$ are not just roots of unity, but equal to $1$. The rest of this section is devoted to justifying such a reduction. Let $N\in\N$ be such that the $N$-th power of all the eigenvalues of the action of $t$ on $\Tors_R H_*(U;\calL)$ (equiv. on $\Tors_R H^*(U;\ov\calL)$) is $1$ for all $*$. Consider the following pull-back diagram: $$ \begin{tikzcd} U_N = \{(x,z)\in U\times \C^*\mid f(x) = z^N \}\arrow[d,"p"] \arrow[dr,phantom,very near start,"\lrcorner"] \arrow[r,"f_N"] & \C^*\arrow[d,"z\mapsto z^N"] \\ U \arrow[r,"f"] & \C^*. \end{tikzcd} $$ Here $p$ is an $N$-sheeted cyclic cover. Notice that all of the maps involved in this diagram are algebraic, and that $U_N$ is also a smooth algebraic variety. We can define $(U_N)^{f_N}$, $(f_N)_\infty$, $\pi_N$ and ${\mathcal L}_N$ as in Section~\ref{ssAlex} for the map $f_N\colon U_N\to \C^*$. We define: \[ \f{\theta_N}{U^f}{U^{f_N}}{U\times \C\ni (x,z)}{(x,e^{z/N}, z/N)\in U_N\times \C\subset U\times \C^* \times \C.} \] It fits into the following commutative diagram: \begin{equation}\label{eq:UN} \begin{tikzcd}[column sep = 7em] U^f \arrow[d,"f_\infty"]\arrow[r,"\sim"',"\theta_N", dashed] \arrow[rrr,"\pi",rounded corners,to path= { --([yshift = 0.7em]\tikztostart.north) -- ([yshift = 0.7em]\tikztotarget.north)\tikztonodes -- (\tikztotarget)}] & U^{f_N}_N \arrow[d,"(f_N)_\infty"]\arrow[r,"\pi_N"]\arrow[dr,phantom,very near start,"\lrcorner"] & U_N \arrow[d,"f_N"] \arrow[r,"p"]\arrow[dr,phantom,very near start,"\lrcorner"] & U \arrow[d,"f"] \\ \C \arrow[r,"z\mapsto\frac{z}{N}"]\arrow[rrr,"\exp", rounded corners,to path= { --([yshift = -0.7em]\tikztostart.south) -- ([yshift = -0.7em]\tikztotarget.south)\tikztonodes -- (\tikztotarget)}] & \C \arrow[r,"\exp"]& \C^* \arrow[r,"z\mapsto z^N"]& \C^*. \end{tikzcd} \end{equation} The map $\theta_N$ allows us to identify $U^f$ with $U_N^{f_N}$ in a canonical way, which we will do from now on. In particular, we can also identify the constant sheaves $\ul k_{U_N^{f_N}}$ and $\ul k_{U^f}$ canonically. On fundamental groups, $p$ induces an isomorphism: \[ \pi_1(U_N) \cong \{ \gamma\in \pi_1(U) \mid f_*(\gamma)\in N\Z\}. \] Let $R(N)\coloneqq k[t^{N},t^{-N}]$. Since $p$ is an $N$-sheeted covering space, $p_*\calL_N =p_!\calL_N$ is a local system of $R(N)$-modules of rank $N$ on $U$. Since $R$ is a rank $N$ free $R(N)$-module, we can also consider $\calL$ as a local system of rank $N$ free $R(N)$-modules on $U$. In fact, $\theta_N$ induces the following isomorphism of local systems of $R(N)$-modules (we are using equal signs for canonical isomorphisms): \[ p_*{\mathcal L}_N= p_!{\mathcal L}_N =p_! \left((\pi_N)_! \ul k_{U_N^{f_N}} \right)\overset{\theta_N}{\cong} p_!\left((\pi_N \circ \theta_N)_!\ul k_{U^f}\right) = \pi_! \ul k_{U^f} = {\mathcal L}. \] Since $p$ is a finite covering, $p_*=p_!$ is the left and right adjoint to the sheaf inverse image $p^{-1}$. In particular $p_*$ is exact. As in Remark~\ref{rem:oppositeDual}, there is a basis of the stalk of ${\mathcal L}$ (resp. ${\mathcal L}_N$) parametrized by the fiber of $\pi$ (resp. $\pi_N$). If we denote $\delta_{x'}$ the elements in this basis, then $\theta_N$ maps an element of the form $\delta_{(x,e^{z/N},z)}$ (in the stalk of ${\mathcal L}_N$ at $(x,e^{z/N})\in U_N$) to $\delta_{(x,z)}$. This discussion and some immediate consequences are summarized in the following lemma. \begin{lem}\label{233} \label{lemLocal} In the above notations, $\theta_N$ induces a canonical isomorphism of local systems of $R(N)$-modules $ \theta_{{\mathcal L}_N}\colon p_*\calL_N \cong \calL$. This induces canonical isomorphisms in the homology of these local systems, which is the same as the map induced by $\theta_N$, i.e. the following diagram commutes: \[ \begin{tikzcd}[row sep = 1em] H_i(U_N;{\mathcal L}_N) \arrow[r,"\theta_{{\mathcal L}_N}","\sim"']\arrow[d,"\sim"] & H_i(U;{\mathcal L}) \arrow[d,"\sim"] \\ H_i(U_N^{f_N};k) \arrow[r,"\theta_N","\sim"'] & H_i(U^f;k). \end{tikzcd} \] Identifying each local system with its conjugate, we obtain another canonical isomorphism $\ov\theta_{{\mathcal L}_N}\colon\ov{p_*{\mathcal L}_N}\cong p_*\ov\calL_N \cong \ov\calL$. This induces in cohomology the same map as $\theta_N$, i.e. the following diagram commutes: \[ \begin{tikzcd}[row sep = 1em] \Tors_R H^{i+1}(U;\ov{\mathcal L}) \arrow[from = r,"\ov\theta_{{\mathcal L}_N}"',"\sim"]\arrow[d,"\sim"] & \Tors_{R(N)}H^{i+1}(U_N;\ov{\mathcal L}_N) \arrow[d,"\sim"] \\ (\Tors_R H_i(U^f;k))^{\vee_k} \arrow[from=r,"\theta_N^{\vee_k}"',"\sim"] & (\Tors_{R(N)} H_i(U_N^{f_N};k))^{\vee_k}. \end{tikzcd} \] The vertical isomorphisms come from Proposition~\ref{propcanon}. \end{lem} \begin{proof} The statement about homology is a straightforward application of the discussion in \cite[VI.3]{whitehead}. For the cohomology, we will outline the proof and leave the details to the reader. First, it is enough to show that this diagram commutes, since the isomorphism in Proposition~\ref{propcanon} is natural: \begin{equation}\label{eq:thetaN} {\footnotesize \begin{tikzcd}[column sep = 2em] H^{i+1}(U;\ov{\mathcal L}) \arrow[d,leftrightarrow] \arrow[r,leftrightarrow,"\ov\theta_{{\mathcal L}_N}"] & H^{i+1}(U;\ov{p_!{\mathcal L}_N})\arrow[d,leftrightarrow] \arrow[r,leftrightarrow,"(1)" ]& H^{i+1}(U_N;\ov{{\mathcal L}_N})\arrow[d,leftrightarrow] \\ H^{i+1}(U;\Homm_R({\mathcal L},\ul R))\arrow[r,leftrightarrow,"\theta_{{\mathcal L}_N}^\vee"] & H^{i+1}(U;\Homm_{R}(p_!{\mathcal L}_N,\ul R)) \arrow[r,leftrightarrow,"(2)"] & H^{i+1}(U_N;\Homm_{R(N)}({\mathcal L}_N,\ul {R(N)})) \\ \Ext^1_R(H_{i}(U;{\mathcal L}),R) \arrow[u,"\UCT_R"]\arrow[r,leftrightarrow,"\theta_{{\mathcal L}_N}"]& \Ext^1_R(H_{i}(U;p_!{\mathcal L}_N),R) \arrow[u,"\UCT_R"]\arrow[r,leftrightarrow,"(3)"]& \Ext^1_{R(N)}(H_{i}(U_N;{\mathcal L}_N),R(N))\arrow[u,"\UCT_{R(N)}"] \end{tikzcd} } \end{equation} The top row of vertical arrows is the identification of the dual of a rank 1 free (local system of) modules with its conjugate, as in Remark~\ref{rem:oppositeDual}. The top left square commutes because this identification is natural, as one can easily verify. By $\theta_{{\mathcal L}_N}^\vee$ (resp. $\ov\theta_{{\mathcal L}_N}$) we mean the result of applying $\Homm_R(\cdot, \ul R)$ (resp. the conjugate structure) to $\theta_{{\mathcal L}_N}$. The bottom row of vertical arrows are all the maps of Lemma~\ref{lem:UCT}, which are isomorphisms onto the torsion of the codomain. The bottom left square commutes by the naturality of these maps. The arrow labeled (1) comes from the relation between cohomology and pushforward. According to the statement we want to prove, the arrow labeled (3) must come from the isomorphism $\Tors_RH_i(U;p_!{\mathcal L}_N)\cong \Tors_{R(N)} H_i(U;p_!{\mathcal L})$, after taking $k$-duals and using Lemma~\ref{lemcanon} together with the fact that $\Tors_R = \Tors_{R(N)}$. The arrow labeled (2) is a composition of two maps. We start with the adjunction between restriction of scalars and $\Hom$: $\Hom_R(A,\Hom_{R(N)}(R,B))\cong \Hom_{R(N)}(A_{R(N)},B) $. We have an $R$-linear isomorphism $R\cong\Hom_{R(N)}(R,R(N))$ where $g(t)$ goes to the map $f(t)\mapsto \frac{1}{N}\sum_{\xi^N = 1} g(\xi t)f(\xi t) $. This gives us an isomorphism \[ \begin{array}{rcl} \Homm_R(p_!{\mathcal L}_N,\ul R)&\cong &\Homm_{R(N)}(p_!{\mathcal L}_N,\ul{R(N)})\\ \left(\delta_{(x,e^{z/N},z)}^\wedge\colon\delta_{(x,e^{z/N},z)}\mapsto 1\right) &\mapsto & \left(\delta_{(x,e^{z/N},z)}^\wedge\colon\delta_{(x,e^{z/N},z)}\mapsto 1\right). \end{array} \] Note that on the left-hand side, $\delta_{(x,e^{z/N},z)}^\wedge$ is $R$-linear, so it sends $\delta_{(x,e^{(z+2\pi i k)/N},z+2\pi i k)}\mapsto t^k$. On the right-hand side, its image sends $ \delta_{(x,e^{(z+2\pi i k)/N},z+2\pi i k)}$ to $0$ whenever $k$ is not a multiple of $N$. Verdier duality (\cite{verdierDuality}, see also \cite[V, 7.17]{borel}) then gives the following isomorphism, taking into account that, since $p$ is a finite covering map, $p^!= p^{-1}$: \[ \begin{array}{rcl} \Homm_{R(N)}(p_!{\mathcal L}_N,\ul{R(N)})&\cong & p_*\Homm_{R(N)}({\mathcal L}_N,\ul{R(N)}) = p_*{\mathcal L}_N^\vee\\ \delta_{(x,e^{z/N},z)}^\wedge &\mapsto & \left(\delta_{(x,e^{z/N},z)}^\wedge\colon\delta_{(x,e^{z/N},z)}\mapsto 1\right). \end{array} \] The stalk at $x$ of $p_*{\mathcal L}_N^\vee$ is the sum of the stalks of ${\mathcal L}_N^\vee$ at $p^{-1}(x)$, and $\delta_{(x,e^{z/N},z)}^\wedge$ is $0$ on all these stalks except at $(x,e^{z/N})$. We can now see that the top right square above commutes, as it comes from applying cohomology to the following commutative square, where the elements are mapped as in the right-hand side (the notation $\ov\cdot$ is as in Remark~\ref{rem:conjugate}): \[ \begin{tikzcd} \ov{p_!{\mathcal L}_N} \arrow[r,leftrightarrow]\arrow[d] & p_!\ov{{\mathcal L}_N} \arrow[d]& \ov{\delta_{(x,e^{z/N},z/N)}} \arrow[r,mapsto]\arrow[d,mapsto] & \ov{\delta_{(x,e^{z/N},z/N)}} \arrow[d,mapsto]\\ \Homm_R(p_!{\mathcal L}_N,\ul R) \arrow[r,leftrightarrow] & p_*\Homm_{R(N)}({{\mathcal L}_N},\ul{R(N)}) & \delta_{(x,e^{z/N},z/N)}^\wedge \arrow[r,mapsto] & \delta_{(x,e^{z/N},z/N)}^\wedge . \end{tikzcd} \] It remains to show that the bottom right square of (\ref{eq:thetaN}) commutes. To this end, consider the arrow labeled $(3)$. Up to composing with the isomorphism $ H_i(U;p_!{\mathcal L}_N)\cong H_i(U_N;{\mathcal L}_N)$, it is the bottom path in the following diagram. \[ {\footnotesize \begin{tikzcd}[row sep = 1.4em, column sep = 0.1em] \Ext^1_R(H_i(U;p_!{\mathcal L}_N),R) \arrow[d,"\Res","\sim"'] \arrow[r]& \Ext^1_R(H_i(U;p_!{\mathcal L}_N),\Hom_R(R,R(N)))\arrow[r] & \Ext^1_{R(N)}(H_i(U;p_!{\mathcal L}_N),R(N))\arrow[d,"\Res","\sim"'] \\ \Hom_k(\Tors_R H_i(U;p_!{\mathcal L}_N),k) \arrow[rr,equals]& & \Hom_k(\Tors_{R(N)} H_i(U;p_!{\mathcal L}_N),k) \end{tikzcd} } \] On the top row we have the same $R$-linear isomorphism $R\cong \Hom_R(R,R(N))$ that we used above, composed with (Ext applied to) the adjunction map $\Hom_R(R,R(N))\mapsto R(N)$, sending $\varphi\mapsto \varphi(1)$. We leave to the reader the verification that this diagram commutes, noting that one can use the classification of modules over a PID to reduce to the case where $H_i(U;p_!{\mathcal L}_N)$ is a cyclic module. Therefore, to show that the bottom right square in (\ref{eq:thetaN}) commutes, we may use the description of the map above. We show that it commutes by taking the complexes whose cohomologies give rise to this square. Take a base point $x_N\in U_N$, let $x=p(x_N)$ and $\pi_1 \coloneqq \pi_1(U,x)\supset \pi_1(U_N,x_N) \eqqcolon \pi_1'$. Denote the stalk $M \coloneqq ({\mathcal L}_N)_{x_N}$, seen as a $R(N)[\pi_1']$-module. Note that $(p_*{\mathcal L}_N)_x\cong k[\pi_1]\otimes_{k[\pi_1']} M$, which becomes an $R$-module by letting $t\in R$ act as $\gamma\in \pi_1$ such that $f_*(\gamma)$ is the counterclockwise generator of $\pi_1(\C^*)$. Let $C_\bullet$ be the singular chain complex of the universal cover of $U$ with $k$ coefficients, seen as a right $k[\pi_1]$-module by the inverse of deck transformations. Then, the square we are interested in arises from the cohomology of the commutative diagram below. \[ \begin{tikzcd} \Hom_{k[\pi_1]}^\bullet(C_{\bullet},\Hom_{R}(k[\pi_1]\otimes_{k[\pi_1']} M, R)) \arrow[r] & \Hom_{k[\pi_1']}^\bullet(C_{\bullet},\Hom_{R(N)}(M,R(N))) \\ \Hom^\bullet_R(C_\bullet \otimes_{k[\pi_1]} k[\pi_1]\otimes_{k[\pi_1']} M,R) \arrow[u,"\text{t-h}"]\arrow[r]& \Hom^\bullet_{R(N)}(C_\bullet \otimes_{k[\pi_1']} M,R(N)).\arrow[u,"\text{t-h}"] \end{tikzcd} \] To obtain the bottom right square in (\ref{eq:thetaN}), we must take cohomology and then restrict to the torsion of the bottom row, applying the universal coefficient theorem. The vertical arrows are tensor-hom adjunction. We can directly verify that the diagram commutes. Note that $C_\bullet$ is a free $k[\pi_1]$-module, with a basis that we will denote $\{c_j\}_{j\in J}$. $M$ is a rank $1$ free module over $R(N)$. Let us call a basis (for both module structures) $\{m\}$, and note that $m$ also generates $M$ over $k[\pi_1']$. Take $\gamma\in \pi_1$ to be any lift of the generator of $\pi_1/\pi_1'\cong \Z/N\Z$. Then $\{ \gamma^k\}_{0\le k<N}$ is a $k[\pi_1']$-basis of $k[\pi_1]$. Let $P\colon R\to R(N)$ be the projection $f(t) \mapsto \frac{1}{N}\sum_{\xi^N=1} f(\xi t)$. The elements of the groups above have the following form. They are parametrized by picking $a_{j}\in R$ for $j\in J$. \[ \begin{tikzcd} \left\{ c_{j} \mapsto (\gamma^{k}\otimes m \mapsto t^ka_{j}) \right\}_{(a_j)_{j\in J}} \arrow[r] & \left\{ \gamma^{-k}c_{j} \mapsto ( m \mapsto P(t^k a_{j})) \right\}_{(a_j)_{j\in J}} \\ \left\{ c_{j}\otimes \gamma^{k} \otimes m \mapsto t^ka_{j} \right\}_{(a_j)_{j\in J}} \arrow[u] \arrow[r]& \left\{ \gamma^{-k}c_{j}\otimes m \mapsto P(t^ka_{j}) \right\}_{(a_j)_{j\in J}}.\arrow[u] \end{tikzcd} \] We will leave to the reader the verification that the maps are indeed defined as the diagram suggests, so that the diagram commutes as desired. \end{proof} \begin{remk} Notice that the only eigenvalue of the action of $t^N$ in $\Tors_{R(N)} H^*(U_N;\ov\calL_N)$ is $1$. So Lemma \ref{233} allows us to reduce the problem of constructing a mixed Hodge structure on $\Tors_{R} H^*(U;\ov\calL)$ to the case when the only eigenvalue is $1$.$\sslash$ \label{remEigenvalue} \end{remk} \subsection{Differential Graded Algebras} By a \bi{commutative differential graded $k$-algebra (cdga)} we mean a triple $(A,d,\wedge)$ such that: \begin{itemize} \item $(A, \wedge)$ is a positively graded unital $k$-algebra. \item $a \wedge b = (-1)^{\deg(a)\deg(b)} b \wedge a$ for homogeneous $a,b \in A$. \item $(A,d)$ is a cochain complex. \item $d(a \wedge b) = da \wedge b + (-1)^{\deg(a)} a \wedge db$ for homogeneous $a,b \in A$. \end{itemize} Notice that when we write cdga, the field $k$ is implicit. We often write $A$ instead of $(A,d,\wedge)$ when the differential and multiplication are understood. When we discuss Hodge complexes, we will often work with filtered cdgas whose filtrations are compatible with the differential and the multiplication. Suppose $(A,d,\wedge)$ is a cdga. By an \bi{increasing cdga-filtration} on $(A, d, \wedge)$ we mean an increasing filtration $W_{\lc}$ on $A$ such that \[ W_iA \wedge W_jA \subset W_{i+j}A \quad\text{and}\quad d(W_iA)\subset W_iA \] for all integers $i$ and $j$. By a \bi{decreasing cdga-filtration} on $(A, d, \wedge)$ we mean a decreasing filtration $F^{\lc}$ on $A$ such that \[ F^iA \wedge F^jA \subset F^{i+j}A \quad\text{and}\quad d(F^iA)\subset F^iA \] for all integers $i$ and $j$. One defines \bi{cdga-filtrations on a sheaf of cdgas} analogously, by looking at the cdgas of sections over arbitrary open subsets. \subsection{Mixed Hodge Structures and Complexes}\label{ss:MHSsAndComplexes} \textit{Let $k$ denote a subfield of $\R$ and $X$ denote a topological space in this subsection.} The purpose of this subsection is to compile relevant definitions and to set notations related to mixed Hodge structures and complexes. \begin{dfn}[{\cite[Definition 2.1]{peters2008mixed}}]\label{dfn:pureHS} Let $m\in\Z$, $V$ be a finite dimensional $k$-vector space, and let $V_{\C}=V\otimes_k \C$ be its complexification. A \bi{$k$-Hodge structure of weight $m$ }on $V$ is a direct sum decomposition (\bi{Hodge decomposition}): $$ V_{\C}=\bigoplus_{\substack{p,q\in\Z\\p+q=m}}V^{p,q}\quad\text{ such that }V^{p,q}=\overline{V^{q,p}}. $$ The data of the Hodge decomposition is equivalent to a decreasing filtration $F^p$ on $V_{\C}$ (the \bi{Hodge filtration}) with the property that $V_{\C} = F^p\oplus\ov{F^q}$ for all $p,q\in\Z$ such that $p+q=m+1$. More precisely, one can be obtained from the other as follows: $$ F^p=\bigoplus_{r\geq p} V^{r,m-r};\quad V^{p,q}=F^p\cap\ov{F^q}. $$ \end{dfn} \begin{dfn}[{\cite[Definition 3.1]{peters2008mixed}}]\label{dfn:MHS} Let $V$ be a finite dimensional $k$-vector space, and let $V_{\C}=V\otimes_k \C$ be its complexification. A \bi{$k$-mixed Hodge structure (MHS)} on $V$ is the data of an increasing filtration $W_{\lc}$ on $V$ (\bi{weight filtration}) and a decreasing filtration $F^{\lc}$ on $V_{\C}$ (\bi{Hodge filtration}), with the following property: for every $m\in \Z$, on $\Gr_m^WV\coloneqq \frac{W_{m}}{W_{m+1}}$, the filtration induced $F^p$ on the complexification gives $\Gr_m^WV$ a pure Hodge structure of weight $m$. \end{dfn} \begin{dfn}[{\cite[Definition 2.32]{peters2008mixed}}] If $m \in \Z$ then a \bi{$k$-Hodge complex of sheaves on $X$ of weight $m$} is a triple $\calK^\bullet = (\calK^\bullet_k, (\calK^\bullet_{\C},F), \alpha)$ where: \begin{itemize} \item $\calK^\bullet_k$ is a bounded below cochain complex of sheaves of $k$-vector spaces on $X$ such that the $\mathbb{H}^*(X,\calK_k^\bullet)$ are finite-dimensional. \item $\calK^\bullet_\C$ is a bounded below cochain complex of sheaves of $\C$-vector spaces on $X$ and $F$ is a decreasing filtration on $\calK^\bullet_\C$. \item $\alpha\colon \calK^\bullet_k \dashrightarrow \calK^\bullet_\C$ is a \bi{pseudo-morphism} of complexes of sheaves of $k$-vector spaces on $X$ (i.e.\ a morphism in the derived category) that induces a \bi{pseudo-isomorphism} $\alpha \otimes 1\colon \calK^\bullet_k \otimes \C \dashrightarrow \calK^\bullet_\C$ (i.e.\ an isomorphism in the derived category). \item the filtration induced by $F$ and $\alpha$ on $\mathbb{H}^*(X,\calK^\bullet_k) \otimes \C \cong \mathbb{H}^*(X,\calK^\bullet_\C)$ endows the $k$-vector space $\mathbb{H}^*(X, \calK^\bullet_k)$ with a $k$-Hodge structure of weight $*+m$. \item The spectral sequence $\H^{p+q}(X,\Gr_F^p \calK^\bullet_{\C})\Rightarrow \H^{p+q}(X,\calK^\bullet_{\C})$ associated to $(\calK^\bullet_\C, F)$ degenerates at the $E_1$-page (see \cite[Definition 2.32]{peters2008mixed} for more details). \end{itemize} \end{dfn} \begin{dfn}[{\cite[Definition 3.13]{peters2008mixed}}] A \bi{$k$-mixed Hodge complex of sheaves on $X$} is a triple $\calK^\bullet = ((\calK^\bullet_k,W_{\lc}), (\calK^\bullet_\C, W_{\lc},F^{\lc}), \alpha)$ where: \begin{itemize} \item $\calK^\bullet_k$ is a bounded below cochain complex of sheaves of $k$-vector spaces on $X$ such that $\mathbb{H}^*(X,\calK_k^\bullet)$ are finite-dimensional, and $W_{\lc}$ is an increasing (weight) filtration on $\calK^\bullet_k$. \item $\calK^\bullet_\C$ is a bounded below cochain complex of sheaves of $\C$-vector spaces on $X$; $W_{\lc}$ is an increasing (weight) filtration and $F^{\lc}$ a decreasing (Hodge) filtration on $\calK^\bullet_\C$. \item $\alpha\colon (\calK^\bullet_k,W_{\lc}) \dashrightarrow (\calK^\bullet_\C,W_{\lc})$ is a \bi{pseudo-morphism of filtered complexes} of sheaves of $k$-vector spaces on $X$ (i.e.\ a chain of morphisms of bounded-below complexes of sheaves as in \cite[Definition 2.31]{peters2008mixed} except that each complex in the chain is filtered, as are all the morphisms) that induces a filtered pseudo-isomorphism \[ \alpha \otimes 1\colon (\calK^\bullet_k \otimes \C, W_{\lc} \otimes \C) \dashrightarrow (\calK^\bullet_\C,W_{\lc}) \] that is, a pseudo-isomorphism on each graded component. \item for $m \in \Z$, the $m$-th $W$-graded component \[ \Gr_m^W\calK^\bullet = \Big(\Gr_m^W\calK^\bullet_k, \big(\Gr_m^W\calK^\bullet_\C, F^{\lc}\big), \Gr_m^W\alpha\Big) \] is a $k$-Hodge complex of sheaves \cite[Definition 2.32]{peters2008mixed} on $X$ of weight $m$, where $F^{\lc}$ denotes the induced filtration. \end{itemize} We will sometimes introduce a $k$-mixed Hodge complex of sheaves on $X$ simply as $\calK^\bullet$ and implicitly assume the components of the triple to be notationally the same as in the above definition. \end{dfn} \begin{dfn} A \bi{multiplicative $k$-mixed Hodge complex of sheaves on $X$} is a $k$-mixed Hodge complex of sheaves $\calK^\bullet$ on $X$ such that the pseudo-morphism $\alpha$ has a distinguished representative given by a chain of morphisms of sheaves of cdgas on $X$ (with all but $\calK^\bullet_k$ being a sheaf of $\C$-cdgas), and such that \textit{all} filtrations (including those in the chain) are cdga-filtrations (over $\C$ except for the weight filtration on $\calK^\bullet_k$). \end{dfn} From a given Hodge complex, one can construct others as follows. Suppose $m \in \Z$ and $\calK^\bullet = (\calK^\bullet_k, (\calK^\bullet_\C, F^{\lc}), \alpha)$ is a (pure) $k$-Hodge complex of sheaves \cite[Definition 2.32]{peters2008mixed} on $X$ of weight $m$. If $j \in \Z$, the \bi{$j$-th Tate twist of $\calK^\bullet$} is the triple \[\calK^\bullet(j) = \Big( \calK^\bullet_k, \big(\calK^\bullet_\C, F[j]\big), \alpha \Big)\] where $F[j]^i = F^{j+i}$ is the shifted filtration. $\calK^\bullet(j)$ is a $k$-Hodge complex of sheaves on $X$ of weight $m-2j$. For details see \cite[Definition 2.35]{peters2008mixed} and notice the we have changed the convention by selecting not to multiply by $(2 \pi i )^j$. Tate twists can also be defined for mixed Hodge complexes of sheaves. Suppose $\calK^\bullet$ is a $k$-mixed Hodge complex of sheaves on $X$. If $j \in \Z$, the \bi{$j$-th Tate twist of $\calK^\bullet$} is the triple \[ \calK^\bullet(j) = \Big(\big(\calK^\bullet_k,W[2j]_{\lc}\big), \big(\calK^\bullet_\C, W[2j]_{\lc},F[j]^{\lc}\big), \alpha\Big) \] where $W[2j]_i = W_{2j+i}$ and $F[j]^i = F^{j+i}$ are shifted filtrations. $\calK^\bullet(j)$ is again a $k$-mixed Hodge complex of sheaves on $X$. For details see \cite[Definition 3.14]{peters2008mixed} and notice again that we have changed convention by selecting not to multiply by $(2 \pi i )^j$. We will also refer to Tate twists of $k$-mixed Hodge structures. These are defined by shifting the weight and Hodge filtrations with the same formula we used for mixed Hodge complexes above. See \cite[Example 3.2(3)]{peters2008mixed} for an explicit definition, except that, as expected, we opt not to multiply by a power of $2\pi i$. We will be interested in shifting mixed Hodge complexes. If $\calF^\bullet$ is a complex of sheaves on $X$, then its \bi{translation} is the complex $\calF^{\bullet}[1] = \calF^{\bullet+1}$ with differential $d^\bullet[1] = -d^{\bullet+1}$. Suppose $\calK^\bullet$ is a $k$-mixed Hodge complex of sheaves on $X$. The \bi{translation of $\calK^\bullet$} is the triple \[ \calK^\bullet[1] = ((\calK^\bullet_k[1], W[1]_{\lc}), (\calK^\bullet_\C[1], W[1]_{\lc}, F[1]^{\lc}), \alpha[1]) \] where the filtrations are described by: \begin{align*} &(W[1])_i\left(\calK^\bullet_k[1]\right) = \left(W_{i+1}\calK^\bullet_k\right)[1],\ (W[1])_i\left(\calK^\bullet_\C[1]\right) = \left(W_{i+1}\calK^\bullet_\C\right)[1],\ i \in \Z\\ &\left(F[1]\right)^p\left(\calK^\bullet_\C[1]\right) = \left(F^{p+1}\calK^\bullet_\C\right)[1],\ p \in \Z. \end{align*} This is again a $k$-mixed Hodge complex of sheaves on $X$. Note that this agrees with the translation of a pure Hodge complex as defined in \cite[2.35]{peters2008mixed} for pure Hodge complexes, but not with the translation of mixed Hodge complexes implicit in loc. cit. 3.22. \begin{remk}\label{transvstate} Suppose $\calK^\bullet$ is a $k$-mixed Hodge complex of sheaves on $X$. By \cite[Theorem 3.18II]{peters2008mixed} the hypercohomology vector spaces $\mathbb{H}^*(X, \calK^\bullet_k)$ inherit $k$-mixed Hodge structures. There is a relation between Tate twists and the translation of a $k$-mixed Hodge complex $\calK^\bullet$ of sheaves on $X$. Namely it can be shown that: \[\mathbb{H}^{*}(X, \calK^\bullet_k[1]) \cong \mathbb{H}^{*+1}(X, \calK^\bullet_k)(1)\] where a Tate twist has been taken on the right-hand side, and the $k$-mixed Hodge structure on the left-hand side has been induced by the translated $k$-mixed Hodge complex $\calK^\bullet[1]$.$\sslash$ \end{remk} \begin{remk}[Derived direct image of a mixed Hodge complex of sheaves.] \label{directim} Suppose $\calK^\bullet$ is a $k$-mixed Hodge complex of sheaves on $X$ where the filtrations $W_{\lc}$ and $F^{\lc}$ are biregular (i.e. for all $m$, the filtrations induced on $\calK^m$ are finite). Let $g\colon X\rightarrow Y$ be a continuous map between two topological spaces. The derived direct image of $\calK^\bullet$ via $g$ is again a mixed Hodge complex of sheaves, and it is defined as follows (\cite[B.2.5]{peters2008mixed}). Let $\Tot[\calC_{\Gdm}^\bullet\calF^\bullet]$ be the Godement resolution of a complex of sheaves $\calF^\bullet$ as defined in \cite[B.2.1]{peters2008mixed}, which is a flabby resolution. Here, $\Tot[\calC_{\Gdm}^\bullet\calF^\bullet]$ denotes the simple complex associated to the double complex $\calC_{\Gdm}^\bullet\calF^\bullet$. We define $Rg_*\calK^\bullet$ to be the following triple \[ \Big(\big(g_* \Tot[\calC_{\Gdm}^\bullet \calK^\bullet_k],g_* \Tot[\calC_{\Gdm}^\bullet W_{\lc}]\big), \big(g_* \Tot[\calC_{\Gdm}^\bullet \calK^\bullet_\C], g_* \Tot[\calC_{\Gdm}^\bullet W_{\lc}],g_* \Tot[\calC_{\Gdm}^\bullet F^{\lc}]\big), g_*\alpha\Big) \] where $g_*\alpha$ is the pseudo-morphism of filtered complexes of sheaves of $k$-vector spaces induced by $\alpha$ and the functoriality of both $g_*$ and the Godement resolution.$\sslash$ \end{remk} \subsection{Real Mixed Hodge Complexes on Smooth Varieties}\label{realmixedhodgecomplexonsmoothvarieties} \textit{Throughout this subsection let $k = \R$, $U$ denote a smooth complex algebraic variety, and $X$ denote a good compactification \cite[Definition 4.1]{peters2008mixed} of $U$ with simple normal crossing divisor $D = X\setminus U$. Let $j\colon U \rightarrow X$ denote the inclusion.} In this section, we remind the reader of the $\R$-mixed Hodge complex of sheaves associated to the log de Rham complex $\logdr{X}{D}$. For a definition of the log de Rham complex and its properties, see \cite[Section 4.1]{peters2008mixed}. Before describing the particular filtrations on the log de Rham complex, we define a pair of filtrations that can be applied to any cochain complex. Suppose $\calF^\bullet$ is a cochain complex of sheaves. Then we have the increasing \bi{canonical filtration} $\tau_{\lc}$, which is described by: \begin{align*} \tau_m(\calF^\bullet) = \left\{ \cdots \rightarrow \calF^{m-2} \rightarrow \calF^{m-1} \rightarrow \ker d^m \rightarrow 0 \rightarrow 0 \rightarrow \cdots\right\} \end{align*} and we have the decreasing \bi{trivial filtration}, whose $p$-th filtered subcomplex is $\calF^{\geq p}$. We now look at the log de Rham complex in particular. On it we have the weight filtration $W_{\lc}$, which is described by: \begin{align*} W_m\Omega^p_X(\log D) = \left\{\begin{array}{ll} 0 & \mbox{if}\ m < 0\\ \Omega_X^p(\log D) & \mbox{if}\ m \geq p\\ \Omega_X^{p-m} \wedge \Omega_X^m(\log D) & \mbox{if}\ 0 \leq m \leq p \end{array}\right. \end{align*} and we have Hodge filtration $F^{\lc}$, defined to be the trivial filtration. \begin{thm}\label{logmhs} The triple \[ \calH dg^\bullet(X\,\log D) \coloneqq \Big(\big(j_*\calE^\bullet_U, \tau_{\lc}\big), \big(\logdr{X}{D}, W_{\lc}, F^{\lc}\big), \alpha \Big) \] where $\calE^\bullet_U$ is the real de Rham complex and $\alpha$ is represented by: \begin{equation}\label{eqmorphisms} \begin{tikzcd} (j_*\calE^\bullet_U, \tau_{\lc}) \ar[r, "j_*(\id \otimes 1)"] & (j_*(\calE^\bullet_U \otimes \C), \tau_{\lc}) & \ar{l}{\simeq}[swap]{\textnormal{incl}} (\logdr{X}{D}, \tau_{\lc}) \arrow{r}{\id}[swap]{\simeq} & (\logdr{X}{D}, W_{\lc}) \end{tikzcd} \end{equation} determines an $\R$-mixed Hodge complex of sheaves on $X$. \end{thm} \begin{proof} This is the content of \cite[Proposition-Definition 4.11]{peters2008mixed} except that here we work over $\R$ instead of $\Q$. Notice that we do not need derived pushforwards, because the objects being pushed forward are complexes of soft sheaves. \end{proof} \begin{remk}\label{multiplicative} All the complexes of sheaves appearing in \eqref{eqmorphisms} are cdgas over $\C$ except the first one, and the morphisms are morphisms of sheaves of cdgas. Furthermore, all filtrations are cdga-filtrations. In other words, $\calH dg^\bullet(X\, \log D)$ is a multiplicative $\R$-mixed Hodge complex of sheaves.$\sslash$ \end{remk} \subsection{Rational Mixed Hodge Complexes on Smooth Varieties}\label{rationalmixedhodgecomplexonsmoothvarieties} \textit{Throughout this subsection let $k = \Q$, let $U$ denote a smooth complex algebraic variety, and $X$ denote good compactification of $U$ with simple normal crossing divisor $D = X\setminus U$. Let $j\colon U \rightarrow X$ denote the inclusion.} In \cite[Section 4.4]{peters2008mixed} the authors define a multiplicative $\Q$-mixed Hodge complex of sheaves associated to the log de Rham complex whose pseudo-morphism is actually a morphism. The rational component of this mixed Hodge complex is perhaps less familiar than the de Rham complex, which is the reason we have separately considered the real case. Let us outline the $\Q$-mixed Hodge complex. Let $\calO_X$ denote the sheaf of holomorphic functions on $X$. Let $\calO_U^*$ denote the sheaf of invertible holomorphic functions on $U$. Let $\calM_{X,D}^{\mathrm{gp}}$ be the sheaf of abelian groups associated to $\calM_{X,D} = \calO_X \cap j_*\calO_U^*$, defined by the following universal property: there is a universal map $c\colon\calM_{X,D}\rightarrow \calM_{X,D}^{\mathrm{gp}}$ such that every homomorphism of monoid sheaves from $\calM_{X,D}$ to a sheaf of groups on $X$ factorizes uniquely over $c$. In other words, $\calM_{X,D}^{\mathrm{gp}}$ is the sheaf of meromorphic functions on $X$ which are holomorphic on $U$ and whose inverse is holomorphic on $U$ as well. For $i \in \Z$ let $\mathrm{Sym}^i_\Q(\calO_X)$ denote the $i$-th symmetric tensor sheaf on $\calO_X$, where $\calO_X$ has been interpreted as a sheaf of $\Q$-vector spaces. As in \cite[Section 4.4]{peters2008mixed} for integers $m$ and $p$ we define: \begin{align*} \calK^p_m = \left\{\begin{array}{ll} \mathrm{Sym}^{m-p}_{\Q}(\calO_X) \otimes_\Q \bigwedge_{\Q}^p (\calM_{X,D}^{\mathrm{gp}} \otimes_\Z \Q) & \mbox{if}\ m \geq p \geq 0\\ 0 & \mbox{otherwise}, \end{array}\right. \end{align*} and the differential $d\colon \calK^p_m \rightarrow \calK^{p+1}_m$ is given by: \begin{align*} d(f_1 \cdots f_{m-p} \otimes y) = \sum_{k=1}^{m-p} f_1 \cdots f_{k-1} \cdot f_{k+1} \cdots f_{m-p} \otimes \exp(2\pi i f_k) \wedge y \end{align*} for $m \geq p \geq 0$, where $f_1, \dots, f_{m-p}$ are sections of $\calO_X$ and $y$ is a section of $\bigwedge_{\Q}^p (\calM_{X,D}^{\mathrm{gp}} \otimes_\Z \Q)$. For fixed $m$, the differential $d$ endows $\calK^\bullet_m$ with the structure of a cochain complex of sheaves on $X$. It can be related to the log de Rham complex via the map $\varphi_m\colon \calK^\bullet_m \rightarrow W_m\logdr{X}{D}$ described by: \[ \varphi_m(f_1 \cdots f_{m-p} \otimes y_1 \wedge \cdots \wedge y_p) = \frac{1}{(2\pi i)^p} \left(f_1 \cdots f_{m-p}\right) \frac{dy_1}{y_1} \wedge \cdots \wedge \frac{dy_p}{y_p} \] for $0 \leq p \leq m$. The map $\varphi_m \otimes 1\colon \calK^\bullet_m \otimes \C \rightarrow W_m \logdr{X}{D}$ is a quasi-isomorphism by \cite[Theorem~4.15]{peters2008mixed}. For any $m$, we have an inclusion of complexes $\calK^\bullet_m \rightarrow \calK^\bullet_{m+1}$ given by: \[ f_1 \cdots f_{m-p} \otimes y \mapsto 1 \cdot f_1 \cdots f_{m-p} \otimes y \] for $0 \leq p \leq m$. Therefore we may consider the direct limit \[ \calK^\bullet_\infty \coloneqq \varinjlim_m \calK^\bullet_m. \] Define a weight filtration $W_{\lc}$ on $\calK^\bullet_\infty$ by declaring $W_m\calK^\bullet_\infty$ to be the image of $\calK^\bullet_m$ in the direct limit. $\calK^\bullet_\infty$ can be equipped with the structure of a sheaf of cdgas by defining: \begin{align*} (f_1 \cdots f_r \otimes y) \wedge (g_1 \cdots g_s \otimes z) = (f_1 \cdots f_r \cdot g_1 \cdots g_s) \otimes (y \wedge z). \end{align*} The weight filtration $W_{\lc}$ is then a cdga-filtration. The direct limit $\varphi_\infty\colon \calK^\bullet_\infty \rightarrow \logdr{X}{D}$ is a morphism of cdgas and moreover by \cite[Corollary 4.16]{peters2008mixed} is a $W_{\lc}$-filtered quasi-isomorphism after tensoring with $\C$. Moreover, by \cite[Corollary 4.17]{peters2008mixed}, the triple \[ \Big(\big(\calK^\bullet_\infty, W_{\lc}\big), \big(\logdr{X}{D}, W_{\lc}, F^{\lc}\big), \varphi_\infty\Big) \] is a multiplicative $\Q$-mixed Hodge complex of sheaves on $X$. Note that the filtration $W_{\lc}$ on $\calK^\bullet_\infty$ is not bounded above, though bounded above filtrations will be important later. This can be easily corrected by replacing $W_{\lc}$ on $\calK^\bullet_{\infty}$ with the cdga-filtration $\tilde{W}_{\lc}$ described by: \begin{align*} \tilde{W}_m \calK^\bullet_\infty = \begin{cases} W_m \calK^\bullet_\infty & \textnormal{if}\, m \leq \dim X\\ \calK^\bullet_\infty & \textnormal{otherwise.} \end{cases} \end{align*} The morphism $\varphi_\infty$ maps $\tilde{W}_{\lc}$ into $W_{\lc}$ because $W_{\dim X} \logdr{X}{D} = \logdr{X}{D}$. For the same reason, the morphism \[ \varphi_\infty \otimes 1\colon \big(\calK_\infty^\bullet \otimes \C, \tilde{W}_{\lc} \otimes \C\big) \rightarrow \big(\logdr{X}{D}, W_{\lc}\big) \] continues to be a filtered quasi-isomorphism. And because $\C$ is faithfully flat over $\Q$, the inclusion $(\calK^\bullet_\infty, W_{\lc}) \rightarrow (\calK^\bullet_\infty, \tilde{W}_{\lc})$ is also a filtered quasi-isomorphism. Therefore we have: \begin{thm} The triple $((\calK^\bullet_\infty, \tilde{W}_{\lc}), (\logdr{X}{D}, W_{\lc}, F^{\lc}), \varphi_\infty)$ is a multiplicative $\Q$-mixed Hodge complex of sheaves on $X$, and $\calK^\bullet_\infty$ is pseudo-isomorphic to $Rj_*\underline{\Q}$, where $\underline{\Q}$ is the constant sheaf on $U$. \end{thm} \begin{proof} This is the content of \cite[Corollary 4.17]{peters2008mixed} except that $W_{\lc}$ has been replaced by $\tilde{W}_{\lc}$. By the discussion preceding the theorem, this replacement is inconsequential. \end{proof} \subsection{Limit Mixed Hodge Structure}\label{LimitMixedHodgeStructure} After we have identified the mixed Hodge structure of Theorem \ref{mhsexistence}, we will compare it with the well-studied limit mixed Hodge structure on the generic fiber of $f$. We review the construction of this structure over $\Q$, considered in our setting, and under an assumption of properness. \textit{Throughout this subsection, let $U$ denote a smooth, connected $n$-dimensional complex algebraic variety and $f\colon U \rightarrow \C^*$ a proper algebraic map inducing an epimorphism on fundamental groups.} Select a good compactification $X$ of $U$ with simple normal crossing divisor $D = X\setminus U$ such that $f\colon U \rightarrow \C^*$ extends to an algebraic map $\bar{f}\colon X \rightarrow \C P^1$. By replacing $f\colon U \rightarrow \C^*$ with a finite cyclic cover $f_N\colon U_N \rightarrow \C^*$ if necessary, we may assume that $\bar{f}^{-1}(0)$ is reduced, by \cite[Semi-stable Reduction Theorem]{kempf2006toroidal}. \textit{Fix $\bar{f}\colon X \rightarrow \C P^1$ as above for the rest of this subsection. Let $j\colon U \rightarrow X$ and $i\colon E \hookrightarrow X$ denote inclusions, where $E$ denotes the divisor $\bar{f}^{-1}(0)$, which is reduced.} Select an open disk $\Delta \subset \C$ centered at the origin such that $f$ is submersive over $\Delta^*$, the punctured disk. The \bi{nearby cycles functor} $\psi_{\bar{f}}\colon D^+\left(\bar{f}^{-1}\Delta\right) \rightarrow D^+(E)$ is a functor between derived categories of bounded below complexes of sheaves of vector spaces, defined for example in \cite[Section 11.2.3]{peters2008mixed}. Importantly, $\mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q}) \cong H^*(F; \Q)$ where $F$ is any fiber of $f$ over $\Delta^*$. A clockwise loop in $\Delta^*$ determines a monodromy homeomorphism from $F$ to itself and so equips $\mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q})$ with the structure of a torsion module over $\Q[t^{\pm 1}]$. The limit mixed Hodge complex is assigned to $\psi_{\bar{f}}\underline{\Q}$. To define complex weight and Hodge filtrations, Peters and Steenbrink in \cite[Section 11.2.5]{peters2008mixed} work with a double complex of sheaves that is quasi-isomorphic to $\psi_{\ov f} \underline{\Q}$. In this spirit, but with minor shifts of convention, we define a double complex $\calA^{\bullet,\bullet}$ of sheaves of $\C$-vector spaces on $X$ as follows. Let \[ \calA^{r,s} = \begin{cases} \frac{\Omega_X^{r+s+1}(\log D)}{W_r\Omega_X^{r+s+1}(\log D)}, &\text{ if }r\geq 0,\\ 0, &\text{ if }r< 0, \end{cases} \] with differentials $d' = \frac{1}{2\pi i} \frac{df}{f} \wedge - \colon \calA^{r,s} \rightarrow \calA^{r+1,s}$, and $d'' = d\colon \calA^{r,s} \rightarrow \calA^{r,s+1}$ being the differential from $\logdr{X}{D}$. According to \cite[Section 11.2.5]{peters2008mixed}, the associated total complex $\Tot\calA^{\bullet,\bullet}$ is, after applying the functor $i^{-1}$, pseudo-isomorphic to $\psi_{\ov f} \underline{\C}$. The \bi{monodromy weight filtration} on $\Tot\calA^{\bullet, \bullet}$ is denoted by $W(M)_{\lc}$ and described for integers $r \geq 0$, $s$, and $m$ by: \begin{align*} W(M)_m\, \calA^{r,s} = \mbox{image of}\ W_{m+2r+1}\Omega_X^{r+s+1}(\log D)\ \mbox{in}\ \calA^{r,s}. \end{align*} The Hodge filtration on $\Tot\calA^{\bullet, \bullet}$ is denoted by $F^{\lc}$ and defined for an integer $p$ by: \begin{align*} F^p(\Tot\calA^{\bullet,\bullet}) = \bigoplus_{r} \bigoplus_{s \geq p} \calA^{r,s}. \end{align*} This concludes the description of the $\C$-component of the limit mixed Hodge complex. Following the same blueprint, we construct a double complex pseudo-isomorphic to $\psi_{\ov f} \underline{\Q}$. We make use of the pair $(\calK^\bullet_\infty, \tilde{W}_{\lc})$ from Section \ref{rationalmixedhodgecomplexonsmoothvarieties}. Define $\tilde{\calC}^{\bullet,\bullet}$ as follows: \[ \tilde{\calC}^{r,s} = \begin{cases} \frac{\calK_\infty^{r+s+1}}{\tilde{W}_r \calK_\infty^{r+s+1}}, &\text{ if }r\geq 0,\\ 0, &\text{ if }r< 0, \end{cases} \] with differentials $d'=(1 \otimes f) \wedge -\colon \tilde{\calC}^{r,s} \rightarrow \tilde{\calC}^{r+1,s}$, and $d''=d\colon \tilde{\calC}^{r,s} \rightarrow \tilde{\calC}^{r,s+1}$ being the differential from $\calK^\bullet_\infty$. Equip the associated total complex $\Tot\calC^{\bullet,\bullet}$ with the \bi{monodromy weight filtration} denoted by $\tilde{W}(M)_{\lc}$ and described for integers $r \geq 0$, $s$, and $m$ by: \begin{align*} \tilde{W}(M)_m\, \tilde{\calC}^{r,s} = \mbox{image of}\ \tilde{W}_{m+2r+1}\calK_\infty^\bullet\ \mbox{in}\ \tilde{\calC}^{r,s}. \end{align*} The map $\varphi_\infty\colon \calK^\bullet_\infty \rightarrow \logdr{X}{D}$ induces a monodromy-filtered morphism $\varphi_{\infty}\colon \Tot\tilde{\calC}^{\bullet,\bullet} \rightarrow \Tot\calA^{\bullet,\bullet}$. \begin{thm}\label{rationallimitmhs} The restricted triple: \begin{align*} \psi_{ f}^{\textnormal{Hdg}} \coloneqq i^{-1}\left(\big[\Tot\tilde{\calC}^{\bullet,\bullet}, \tilde{W}(M)_{\lc}\big], \big[\vphantom{\tilde{A}} \Tot\calA^{\bullet,\bullet}, W(M)_{\lc}, F^{\lc}\big], \varphi_{\infty}\right) \end{align*} is a $\Q$-mixed Hodge complex of sheaves on $E$, and $i^{-1}\Tot\tilde{\calC}^{\bullet,\bullet}$ is pseudo-isomorphic to $\psi_{\bar{f}}\underline{\Q}$. \end{thm} \begin{proof} This is established (up to minor modifications owing to moderate changes in the definitions of both double complexes) by \cite[Theorem 11.22]{peters2008mixed} except that in place of $\tilde{\calC}^{\bullet, \bullet}$ and its filtration $\tilde{W}(M)_{\lc}$ the authors opt for $\calC^{\bullet,\bullet}$ and $W(M)_{\lc}$, which are built using the unbounded filtration $W_{\lc}$ on $\calK^\bullet_\infty$ instead of $\tilde{W}_{\lc}$. Because the inclusion $(\calK^\bullet_\infty, W_{\lc}) \rightarrow (\calK^\bullet_\infty, \tilde{W}_{\lc})$ is a filtered quasi-isomorphism of cgdas, the induced map $(\Tot\tilde{\calC}^{\bullet,\bullet}, \tilde{W}(M)_{\lc}) \rightarrow \left(\Tot\calC^{\bullet,\bullet}, W(M)_{\lc}\right)$ can also be shown to be a filtered quasi-isomorphism. \end{proof} The $\Q[t^{\pm 1}]$-module structure on $\mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q}) \cong H^*(F; \Q)$ induced by the monodromy action can actually be realized on this mixed Hodge complex of sheaves. Specifically, for integers $r \geq 0$ and $s$ define: \begin{align*} &\Theta\colon \tilde{\calC}^{r,s} \rightarrow \tilde{\calC}^{r+1,s-1},\ \Theta(-) = (-)\, \textnormal{mod}\, \tilde{W}_{r+1} \calK_\infty^{r+s+1}\\ &\Theta\colon \calA^{r,s} \rightarrow \calA^{r+1,s-1},\ \Theta(-) = (-)\, \textnormal{mod}\, W_{r+1} \Omega^{r+s+1}_X(\log D). \end{align*} \begin{thm}\label{limitmhsmonodromy} Under the above notations, the map $\Theta$ gives rise to a morphism of mixed Hodge complexes of sheaves: \begin{align*} i^{-1}\Theta\colon \psi_f^{\textnormal{Hdg}} \rightarrow \psi_f^{\textnormal{Hdg}}(-1) \end{align*} which induces the map $\log t \colon \mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q}) \rightarrow \mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q})(-1)$. Here $\log t$ is interpreted as its power series representation at $t = 1$. \end{thm} \begin{proof} The assertion follows from \cite[Section 11.3.1]{peters2008mixed} with minor modifications, because we have selected moderately different definitions for Tate twists and the pair of double complexes appearing in $\psi_f^{\textnormal{Hdg}}$. We remark that $\log t$ is in fact well-defined, because $E$ being reduced implies that the monodromy action on $\mathbb{H}^*(E; \psi_{\ov f}\underline{\Q})$ is unipotent (for a proof see \cite[Corollary 11.19]{peters2008mixed}). \end{proof} \section{Thickened Complexes}\label{sec3} \subsection{Thickened Complex of a Differential Graded Algebra} Suppose $(A,d,\wedge)$ is a cdga {over $k$}, $\eta \in A^1 \cap \ker d$, and $m \geq 1$. {Recall that $R_m$ is the quotient ring $R_\infty/(s^m)=k[[s]]/(s^m)$. We define the \bi{$m$-thickening of $A$ in the direction $\eta$} to be the cochain complex of $R_m$-modules denoted by \begin{align*} A(\eta,m) = (A {\otimes_k} R_m, d_\eta) \end{align*} and described by: \begin{itemize} \item for $p \in \Z$, the $p$-th graded component of $A(\eta,m)$ is the free $R$-module {$A^p {\otimes_k} R_m$.} \item for $\omega \in A$ and $\phi \in R_m$, we set $d_\eta(\omega \otimes \phi) = d\omega \otimes \phi + (\eta \wedge \omega) \otimes {s\phi} $. \end{itemize} A straightforward calculation verifies that $d_{\eta}^2$ vanishes. If $i\geq 1,j \geq 0$, then the quotient map $R_{i+j} \twoheadrightarrow R_i$ induces a surjective morphism $\phi_{ji}\colon A(\eta,i+j) \twoheadrightarrow A(\eta,i)$. These maps endow $\{A(\eta,i)\}$ with the structure of an inverse system of $\{R_i\}$-modules. Using that $R_\infty = \varprojlim R_i$, we define the \bi{$\infty$-thickening of $A$ in the direction $\eta$} to be the cochain complex of $R_\infty$-modules \begin{align*} A(\eta, \infty) = \varprojlim A(\eta,i). \end{align*} \begin{remk}\label{thickenedmultiplication} The inverse system $\{A \otimes R_i\}$ admits a natural multiplication operation, defined on simple tensors by $(\omega \otimes \phi) \wedge (\omega' \otimes \phi') = (\omega \wedge \omega') \otimes \phi \phi'$. However, this multiplication does not in general endow $\{A(\eta,i)\}$ with the structure of an inverse system of \textit{cdgas}, because it is not compatible with the differential $d_{\eta}$. Nevertheless, we will find an opportunity to utilize it shortly.$\sslash$ \end{remk} \begin{remk}\label{filtration} The complex $A(\eta,m)$ can be equipped with a decreasing filtration $G$, described by \begin{align*} G^pA(\eta,m) = A(\eta,m)s^p \end{align*} for $p \geq 0$. Its graded components are, for $p \geq 0$: \begin{align*} \Gr^p_G A(\eta,m) = (A \otimes k\langle s^p \rangle, d \otimes 1). \end{align*} In particular, $H^*( \Gr^p_G A(\eta,m)) \cong H^*(A)$ $\otimes k\langle s^p \rangle$.$\sslash$ \end{remk} We next state two lemmas concerning thickened complexes, which are essentially listed in \cite[Section~3]{budur2018monodromy}. For future reference, we also include their proofs. \begin{lem}\label{cohomologous} Suppose $(A,d,\wedge)$ is a cdga and $m \geq 1$. Suppose that $\eta_1$ and $\eta_2$ are cohomologous elements in $A^1 \cap \ker d$. Then $A(\eta_1, m)$ and $A(\eta_2, m)$ are isomorphic. More precisely: if $a \in A^0$ is such that $\eta_1-\eta_2 = da$, then the morphism \[\exp(a \otimes s) \wedge -\colon A(\eta_1,m) \rightarrow A(\eta_2,m)\] is defined in the proof and is an isomorphism of cochain complexes. \end{lem} \begin{proof} Let $a \in A^0$ be such that $da = \eta_1 - \eta_2$. The element \[ \exp(a \otimes s) = \sum_{n = 0}^\infty \frac{1}{n!}\, a^n \otimes s^n \] belongs to $\varprojlim A^0 \otimes R_i$ and determines, via the multiplication described in Remark \ref{thickenedmultiplication}, a morphism \[ \exp(a \otimes s) \wedge -\colon A \otimes R_m \rightarrow A \otimes R_m \] of $R_m$-modules. We will show that it in fact determines a morphism \[ \exp(a \otimes s) \wedge -\colon A(\eta_1, m) \rightarrow A(\eta_2,m) \] by verifying commutativity with the differentials. It suffices to investigate its application to a simple tensor of the form $\omega \otimes 1$: \begin{align*} &d_{\eta_2}\left(\exp(a \otimes s)\wedge (\omega \otimes 1)\right)=\\ &= \sum_{n = 0}^{m-1} \frac{1}{n!}~ d_{\eta_2}\left(a^n \omega \otimes s^n\right)=\\ &= \sum_{n = 0}^{m-1} \frac{1}{n!}~ d\left(a^n\omega\right) \otimes s^n + \sum_{n = 1}^{m-1} \frac{1}{(n-1)!}a^{n-1}\eta_2 \wedge \omega \otimes s^n=\\ &=\sum_{n = 1}^{m-1}\frac{1}{(n-1)!}~ a^{n-1} da \wedge \omega \otimes s^n + \sum_{n=0}^{m-1}\frac{1}{n!}~a^n \wedge d\omega \otimes s^n+ \sum_{n = 1}^{m-1} \frac{1}{(n-1)!}a^{n-1}\eta_2 \wedge \omega \otimes s^n \end{align*} and since $da = \eta_1-\eta_2$, the above simplifies to: \begin{align*} \sum_{n = 1}^{m-1}\frac{1}{(n-1)!}~ &a^{n-1} \eta_1 \wedge \omega \otimes s^n + \sum_{n=0}^{m-1}\frac{1}{n!}~a^n \wedge d\omega \otimes s^n=\\ &= \exp(a \otimes s) \wedge (\eta_1 \wedge \omega \otimes s) + \exp(a \otimes s) \wedge (d\omega \otimes s)=\\ &= \exp(a \otimes s) \wedge d_{\eta_1}(\omega \otimes 1). \end{align*} Therefore $\exp(a \otimes s) \wedge -$ is a morphism of cochain complexes. It is in fact an isomorphism, because $\exp(-a \otimes s) \wedge -$ is its inverse. \end{proof} \begin{lem}\label{inducedmap} Suppose $(A,d,\wedge)$ and $(B, d, \wedge)$ are cdgas, $\eta \in A^1 \cap \ker d$, and $m \geq 1$. If $F\colon A \rightarrow B$ is a morphism of dgas, then there is a natural induced morphism \[ F_{\#}\colon A(\eta, m) \rightarrow B(F(\eta), m). \] If $F$ is a quasi-isomorphism, then so is $F_{\#}$. \end{lem} \begin{proof} The morphism $F_\#$ is given by $F \otimes \id\colon A \otimes R_m \rightarrow B \otimes R_m$. $F_\#$ preserves the decreasing filtration $G$ of Remark \ref{filtration}. In particular, for $p \geq 0$ it induces commutative diagrams of short exact sequences: $$ \begin{tikzcd} 0 \ar[r] & G^{p+1} A(\eta,m) \ar[d] \ar[r] & G^p A(\eta,m) \ar[d] \ar[r] & \Gr^p_G A(\eta,m) \ar[d] \ar[r] & 0\\ 0 \ar[r] & G^{p+1} B(F(\eta),m) \ar[r] & G^p B(F(\eta),m) \ar[r] & \Gr^p_G B(F(\eta),m) \ar[r] & 0. \end{tikzcd} $$ From the associated commutative diagram of long exact sequences, the central map is a quasi-isomorphism if the outer two are. For all $p \geq 0$, the induced map on cohomology of graded components is the isomorphism $F^* \otimes \id\colon H^*(A) \otimes k\langle s^p \rangle \rightarrow H^*(B) \otimes k\langle s^p \rangle$. Therefore the induced map on graded components is a quasi-isomorphism. Since $m$ is finite, both $G^mA(\eta,m)$ and $G^mB(F(\eta), m)$ are zero. This provides the starting point to an induction concluding that $G^pA(\eta,m) \rightarrow G^pB(F(\eta),m)$ is a quasi-isomorphism for all $p \geq 0$. Take $p = 0$ to conclude. \end{proof} \begin{remk}\label{remk:psi-ij} Besides the maps associated to the inverse system, there is another natural collection of maps between the thickened complexes. If $i \geq 1,j\geq 0$ then there is an inclusion $R_i \hookrightarrow R_{i+j}$ mapping $R_i$ isomorphically onto $s^jR_{i+j}$. We denote the induced inclusion of thickened complexes by $\psi_{ij}\colon A(\eta,i) \hookrightarrow A(\eta,i+j)$.$\sslash$ \end{remk} \begin{remk}\label{pid} Note that $R_\infty$ is a PID. Therefore all finitely generated $R_\infty$-modules split into torsion and free parts. We let $\mathrm{Tors}_{R_{\infty}}$ denote torsion as an $R_\infty$-module.$\sslash$ \end{remk} \begin{remk}\label{rem:tensorRm} For all $m \geq 1$ we have $A(\eta,m) \cong A(\eta,\infty) \otimes_{R_\infty} R_m$ as $R_\infty$-modules. Therefore, the ring map $R_\infty\to R_m$ induces a map \[ A(\eta,\infty) =A(\eta,\infty) \otimes_{R_{\infty}} R_{\infty} \to A(\eta,\infty) \otimes_{R_{\infty}} R_{m} \cong A(\eta,m).\sslash \] \end{remk} The following lemma allows us to interpret $\mathrm{Tors}_{R_{\infty}}H^*A(\eta,\infty)$ as the kernel of a map between finitely thickened complexes. \begin{lem}\label{lem:torsion2} Suppose $C^\bullet$ is a complex of torsion-free $R_\infty$-modules, and suppose that $m$ is such that $s^m$ annihilates $\Tors_{R_\infty}H^*(C^\bullet)$. Let $\psi_{ij} \colon C^\bullet \otimes_{R_\infty} R_i \to C^\bullet \otimes_{R_\infty} R_{i+j}$ be the map induced by the map $R_i \hookrightarrow R_{i+j}$ as in Remark~\ref{remk:psi-ij}, and let $\psi_{ij}^*$ the map induced in cohomology. Let $\phi_{ji}\colon C^\bullet \otimes_{R_\infty} R_{i+j} \to C^\bullet \otimes_{R_\infty} R_{i}$ be the map induced by the projection $R_{i+j}\twoheadrightarrow R_i$. The natural map $C^\bullet\to C^\bullet\otimes_R R_i$ induces isomorphisms \begin{align*} \Tors_{R_\infty} H^*(C^\bullet) &\cong \ker \psi_{ij}^* \subseteq H^* (C^\bullet\otimes_R R_i)\quad \text{ if $i,j\ge m$};\\ H^*(C^\bullet)\otimes_{R_\infty} R_i &\cong \im \phi_{ji}^* \subseteq H^* (C^\bullet\otimes_R R_i)\quad \text{ if $j\ge m$}. \end{align*} \end{lem} \begin{proof} Since $C^\bullet$ is torsion-free, we have the following short exact sequences with maps between them: \[ \begin{tikzcd}[row sep = 1.5em, column sep = 1.8em] 0 \to C^\bullet \arrow[r,"s^i"]\arrow[d,"=", shift left = 1.7ex] & C^\bullet \arrow[r,twoheadrightarrow]\arrow[d,"s^j"] & C^\bullet \otimes_{R_\infty} R_{i} \arrow[d,hookrightarrow,"\psi_{ij}", shift right = 3ex] \to 0 & 0 \to C^\bullet \arrow[r,"s^{i+j}"]\arrow[d,"s^j", shift left = 1.7ex] & C^\bullet \arrow[r,twoheadrightarrow]\arrow[d,"="] & C^\bullet \otimes_{R_\infty} R_{i+j} \arrow[d,twoheadrightarrow,"\phi_{ji}", shift right = 3ex] \to 0\\ 0 \to C^\bullet \arrow[r,"s^{i+j}"] & C^\bullet \arrow[r,twoheadrightarrow] & C^\bullet \otimes_{R_\infty} R_{i+j} \to 0; & 0 \to C^\bullet \arrow[r,"s^{i}"] & C^\bullet \arrow[r,twoheadrightarrow] & C^\bullet \otimes_{R_\infty} R_{i} \to 0. \end{tikzcd} \] We take the cohomology long exact sequences corresponding to the sequences above, and we consider the map between them. We obtain the following diagrams, where we denote $\Tors_{s^i} M = \{ a\in M \mid s^ia=0\}$: \[ \begin{tikzcd}[row sep = 1.3em] 0 \arrow[r] & H^*(C^\bullet)\otimes_{R_\infty} R_{i\phantom{+j}} \arrow[r]\arrow[d,"s^j\cdot ","(1)"'] & H^*\left(C^\bullet \otimes_{R_\infty} R_{i\phantom{+j}}\right) \arrow[r]\arrow[d,"\psi_{ij}^*"] & \Tors_{s^{i\phantom{+j}}}H^{*+1}(C^\bullet) \arrow[r]\arrow[d,hookrightarrow,"\Id_{H^{*+1}(C^\bullet)}"] & 0\\ 0 \arrow[r] & H^*(C^\bullet)\otimes_{R_\infty} R_{i+j} \arrow[r] & H^*\left(C^\bullet \otimes_{R_\infty} R_{i+j}\right) \arrow[r] & \Tors_{s^{i+j}}H^{*+1}(C^\bullet) \arrow[r] & 0;\end{tikzcd}\] \[ \begin{tikzcd}[row sep = 1.3em] 0 \arrow[r] & H^*(C^\bullet)\otimes_{R_\infty} R_{i+j} \arrow[r]\arrow[d,"\Id_{H^{*}(C^\bullet)}",twoheadrightarrow,"(2)"'] & H^*\left(C^\bullet \otimes_{R_\infty} R_{i+j}\right) \arrow[r]\arrow[d,"\phi_{ji}^*"] & \Tors_{s^{i+j}}H^{*+1}(C^\bullet) \arrow[r]\arrow[d,"s^j\cdot "] & 0\\ 0 \arrow[r] & H^*(C^\bullet)\otimes_{R_\infty} R_{i\phantom{+j}} \arrow[r] & H^*\left(C^\bullet \otimes_{R_\infty} R_{i\phantom{+j}}\right) \arrow[r] & \Tors_{s^{i\phantom{+j}}}H^{*+1}(C^\bullet) \arrow[r] & 0. \end{tikzcd} \] Let $i\ge m$. Since $s^m$ annihilates $\Tors_{R_\infty} H^*(C^\bullet)$, the following composition of obvious maps is an injection: \[ \Tors_{R_\infty} H^*(C^\bullet)\hookrightarrow H^*(C^\bullet) \to \frac{H^*(C^\bullet)}{s^iH^*(C^\bullet)}. \] Seeing $\Tors_{R_\infty} H^*(C^\bullet)$ as a submodule of $\frac{H^*(C^\bullet)}{s^iH^*(C^\bullet)}$ in this way, we have that as long as $j\ge m$, $\Tors_{R_\infty} H^*(C^\bullet)$ is the kernel of the multiplication by the $s^j$ map labeled (1). Applying the snake lemma to the diagram above yields the first statement. For the second statement, note that if $j\ge m$, then the right hand map $s^j$ vanishes. Therefore, the image of $\phi_{ji}^*$ equals the image of the map labeled (2). \end{proof} \begin{cor}\label{torsion} Suppose $(A,d,\wedge)$ is a cdga and $\eta \in A^1 \cap \ker d$. Assume $H^*A(\eta,\infty)$ is a finitely generated $R_\infty$-module. Consider the map $H^*A(\eta,\infty)\to H^*A(\eta,i)$ from Remark~\ref{rem:tensorRm}. This map induces the following isomorphisms: \begin{align*} \mathrm{Tors}_{R_\infty}\,H^*A(\eta,\infty)&\cong \ker \psi_{ij}^* &\text{if $i,j\gg 0$}\\ H^*A(\eta,\infty)\otimes_{R_\infty} R_i&\cong \im \phi_{ji}^* &\text{if $j\gg 0$.} \end{align*} \end{cor} \begin{remk}\label{torsionrelations} Suppose $i, j$ are sufficiently large as in Corollary \ref{torsion}. The following statements follow from the proof of Lemma~\ref{lem:torsion2}. If $j' \geq j$ then $\ker\psi_{ij'}^* = \ker\psi_{ij}^*$ and if $i' \geq i$ then $\ker \psi_{i'j}^* \cong \ker \psi_{ij}^*$ via $\phi_{i'-i,i}^*$. Similarly, $\im \phi_{j'i}^* = \im \phi_{ji}^*$, and $\phi_{i'-i,i}^*$ induces the projection map $\im \phi_{ji'}^* \cong H^*A(\eta,\infty)\otimes_{R_\infty} R_{i'}\twoheadrightarrow H^*A(\eta,\infty)\otimes_{R_\infty} R_{i}\cong \im\phi_{ji}^*$. Finally, $\psi_{i,i'-i}$ induces the map coming from the inclusion $R_i\hookrightarrow{s^{i'-i}} R_{i'}$, tensored with $H^*A(\eta,\infty)$.$\sslash$ \end{remk} \begin{remk}\label{twistedmodule} Suppose $1 \leq m \leq \infty$. Later on we will be interested not in the given $R_\infty$-module structure on $A(\eta,m)$ but rather a twisted one. Namely, if $\tilde{s} \in R_\infty$ satisfies $\tilde{s} = s u$ where $u \in R_\infty$ is some unit, then define $A(\eta,m)_{\tilde{s}}$ to be the vector space cochain complex $A(\eta,m)$ but with $R_\infty$-module multiplication $*_{\tilde{s}}$ described by: \begin{center} if $g(s) \in R_\infty$, then $- *_{\tilde{s}} g(s)$ = $- \cdot g(\tilde{s})$. \end{center} The induced $R_\infty$-module structure on $H^*A(\eta,m)_{\tilde{s}}$ is also given by $- *_{\tilde{s}} g(s)$ = $- \cdot g(\tilde{s})$. Clearly the maps $\phi_{ji}\colon A(\eta,i+j) \twoheadrightarrow A(\eta,i)$ continue to induce $R_\infty$-module morphisms $\phi_{ji}\colon A(\eta,i+j)_{\tilde{s}} \twoheadrightarrow A(\eta,i)_{\tilde{s}}$. Similarly, the maps $\psi_{ij}\colon A(\eta,i) \hookrightarrow A(\eta,i+j)$ continue to induce $R_\infty$-module morphisms $\psi_{ij}\colon A(\eta,i)_{\tilde{s}} \hookrightarrow A(\eta,i+j)_{\tilde{s}}$. Even better, since $\tilde{s}$ and $s$ differ by multiplication with a unit, under the hypotheses of Corollary \ref{torsion}, we have that $H^*A(\eta,\infty)_{\tilde{s}}$ is a finitely generated $R_\infty$-module and for all sufficiently large $i,j$, $\ker\big(\psi_{ij}^*\colon H^*A(\eta,i)_{\tilde{s}} \rightarrow H^*A(\eta,i+j)_{\tilde{s}}\big)$ is naturally isomorphic to $\mathrm{Tors}_{R_\infty}\, H^*A(\eta,\infty)_{\tilde{s}}$ as $R_\infty$-modules.$\sslash$ \end{remk} \subsection{Thickened Complexes and Filtrations}\label{thickenedcomplexesandfiltrations} \textit{Fix $m \geq 1$ for this subsection.} When we discuss mixed Hodge complexes, we will need to find weight and Hodge filtrations on thickened complexes that are induced by filtrations on the original cdga. In this subsection, we describe a general procedure for inducing appropriate filtrations on the thickened complex. Suppose $(A,d,\wedge)$ is a cdga with an increasing cdga-filtration $W_{\lc}$ and $\eta \in W_1A^1 \cap \ker d$. Note the requirement that $\eta \in W_1A$. We start by defining a weight filtration $w_{\lc}$ on $R_m$. For $i \geq 0$, define $$w_iR_m = R_m \quad \text{and}\quad w_{-2i}R_m = w_{-2i+1}R_m = s^iR_m.$$ We equip $A(\eta, m)$ with the tensor weight filtration, which we also denote by $W_{\lc}$ since it does not depend on $\eta$. In other words, for $i \in \Z$, we set: \begin{align*} W_iA(\eta,m) &= \sum_j W_{i+j}A \otimes w_{-j}R_m \end{align*} where $\sum$ denotes the internal sum, not a direct sum. Observe that we can rewrite the expression with a direct sum using the definition of $w_{\lc}$: \begin{align*} W_iA(\eta,m) &= \bigoplus_{j=0}^{m-1} W_{i+2j}A \otimes k\langle s^j \rangle. \end{align*} Since $\eta \wedge W_i A \subset W_{i+1} A$, we have $s\eta\wedge W_iA(\eta,m)\subset W_iA(\eta,m)$. Therefore, $W_iA(\eta,m)$ is closed under $d_\eta$. One technical reason for skipping by twos in the direct sum description is explained by the following lemma. \begin{lem}\label{gradedcomponents} Suppose $(A,d,\wedge)$ is a cdga with an increasing cdga-filtration $W_{\lc}$ and suppose $\eta \in W_1A^1 \cap \ker d$. Then, \begin{align*} \Gr_i^W A(\eta,m) &= \bigoplus_{j=0}^{m-1} \Gr_{i+2j}^W A \otimes k\langle s^j \rangle \end{align*} with differential given by $d \otimes \id$. \end{lem} \begin{proof} This follows from the definition of the weight filtration. In particular, skipping by twos ensures that the $(\eta \wedge - )\otimes s$ component of the differential vanishes on the graded pieces. \end{proof} We can state a filtered analogue of Lemma \ref{cohomologous}. \begin{lem}\label{cohomologousfiltered} Suppose $(A,d,\wedge)$ is a cdga with an increasing cdga-filtration $W_{\lc}$. Suppose that $\eta_1$ and $\eta_2$ are elements of $W_1A^1 \cap \ker d$ that are cohomologous in $W_1A$. If $a \in W_1A^0$ is such that $\eta_1-\eta_2 = da$, then the morphism $\exp(a \otimes s) \wedge -$ of Lemma \ref{cohomologous} gives a filtered isomorphism \[\exp(a \otimes s) \wedge -\colon \left(A(\eta_1, m), W_{\lc}\right) \rightarrow \left(A(\eta_2,m), W_{\lc}\right)\] that induces the identity map on each graded component associated to $W_{\lc}$. \end{lem} \begin{proof} Let $a \in W_1A^0$ be such that $da = \eta_1-\eta_2$. By Lemma \ref{cohomologous}, we know that \[ \exp(a \otimes s) \wedge - \colon A(\eta_1,m) \rightarrow A(\eta_2,m) \] is an isomorphism. Because $a \in W_1A$, this isomorphism preserves filtrations. Furthermore, wedging by $\exp(a \otimes s) - 1 \otimes 1$ takes $W_i$ into $W_{i-1}$ for all integers $i$. Consequently, $\exp(a \otimes s) \wedge -$ coincides with $(1 \otimes 1) \wedge -$ on each graded component. \end{proof} We can also expand on Lemma \ref{inducedmap}. \begin{lem}\label{inducedquasi} Suppose $(A, d, \wedge)$ and $(B,d,\wedge)$ are cdgas with increasing cdga-filtrations both denoted by $W_{\lc}$, and $\eta \in W_1A^1 \cap \ker d$. If $F\colon (A,W_{\lc}) \rightarrow (B,W_{\lc})$ is a filtered morphism of cdgas, then $F_{\#}\colon A(\eta, m) \rightarrow B(F(\eta), m)$ is a filtered morphism with respect to $W_{\lc}$. Moreover for $i \in \Z$, $F_{\#}$ is given on the $i$-th graded component by: \begin{align*} \Gr^W_i(F_{\#}) = \bigoplus_{j=0}^{m-1} \Gr_{i+2j}^W F \otimes \id\colon \bigoplus_{j=0}^{m-1} \Gr_{i+2j}^W A \otimes k\langle s^j \rangle \rightarrow \bigoplus_{j=0}^{m-1} \Gr_{i+2j}^W B \otimes k\langle s^j \rangle. \end{align*} \end{lem} \begin{proof} Because $F_{\#} = F \otimes \id$, the lemma is immediate. \end{proof} Suppose now that $(A,d,\wedge)$ is a cdga with two cdga-filtrations, an increasing filtration $W_{\lc}$ and a decreasing filtration $F^{\lc}$. Suppose also that $\eta \in W_1A^1 \cap F^1A^1 \cap \ker\, d$. We next equip $A(\eta, m)$ with a decreasing filtration, which we also denote by $F^{\lc}$ because it does not depend on $\eta$. For $p \in \Z$, set: \begin{align*} F^pA(\eta,m) = \bigoplus_{j = 0}^{m-1} F^{p+j}A \otimes k\langle s^j \rangle. \end{align*} This is closed under $d_{\eta}$, because $\eta \wedge F^pA \subset F^{p+1}A$ for all $p \in \Z$. Unlike the filtration $W_{\lc}$ on $A(\eta,m)$, the decreasing filtration $F^{\lc}$ is \textit{not} a filtration by cochain complexes of $R_m$-modules, only by cochain complexes of $k$-vector spaces. The following lemma describes the interplay of the two filtrations we have defined. \begin{lem}\label{filtrationinterplay} Suppose $(A,d,\wedge)$ is a cdga with an increasing cdga-filtration $W_{\lc}$ and a decreasing cdga-filtration $F^{\lc}$. Suppose also that $\eta \in W_1A^1 \cap F^1A^1 \cap \ker d$. Then for $i \in \Z$ the filtration on $\Gr_i^W A(\eta,m)$ induced by the filtration $F^{\lc}$ of $A(\eta,m)$ is described by: \begin{align*} F^p \Gr_i^WA(\eta,m) = \bigoplus_{j=0}^{m-1}\left(F^{p+j}\Gr^W_{i+2j}A\right) \otimes k \langle s^j \rangle. \end{align*} where on the right hand side we use the filtration on $\Gr_{i+2j}^W A$ induced by the filtration $F^{\lc}$ of $A$. \end{lem} \begin{proof} Using the definition of the two filtrations, for all integers $i$ and $p$ we find: \begin{align*} F^pA(\eta,m)\cap W_iA(\eta,m) = \bigoplus_{j=0}^{m-1}\left(F^{p+j}A \cap W_{i+2j}A\right) \otimes k \langle s^j \rangle \end{align*} which, under the quotient map to $\Gr_i^WA(\eta,m)$, maps to: \[ F^p\Gr_i^WA(\eta,m) = \bigoplus_{j=0}^{m-1}\left(F^{p+j}\Gr^W_{i+2j}A\right) \otimes k \langle s^j \rangle. \qedhere \] \end{proof} \subsection{Thickened Complexes of Sheaves of Differential Graded Algebras} Suppose $X$ is a topological space and $(\calA, \wedge, d)$ is a sheaf of cdgas on $X$. If $m \geq 1$ and $\eta \in \Gamma(X, \calA^1) \cap \ker d$, then we define the \bi{$m$-thickening of $\calA$ in direction $\eta$}, denoted by $\calA(\eta, m)$, to be the cochain complex of sheaves of $R_\infty$-modules on $X$ defined by: \begin{align*} U \mapsto \Gamma(U, \calA)(\eta|_U, m) \end{align*} for an open subset $U \subset X$. Observe that this does in fact define a cochain complex of sheaves, not just presheaves. By definition, for all $p \geq 0$ we have $\calA^p(\eta,m) = \calA^p \otimes_k R_m$. Recalling Remark \ref{twistedmodule}, if $\tilde{s} \in R_\infty$ differs from $s$ by multiplication with a unit, then we define $\calA(\eta,m)_{\tilde{s}}$ to be the cochain complex of sheaves of $R_\infty$-modules described for open $U \subset X$ by: \begin{align*} U \mapsto \Gamma(U, \calA)(\eta|_U, m)_{\tilde{s}}. \end{align*} As before, if $i \geq 1,j \geq 0$ then we have a surjective morphism $\phi_{ji}\colon \calA(\eta,i+j) \twoheadrightarrow \calA(\eta,i)$. This endows $\{\calA(\eta,i)\}$ with the structure of an inverse system of sheaves of $\{R_i\}$-modules. We avoid discussing the inverse limit, as it may not coincide with the derived inverse limit. Again as before, if $i \geq 1,j \geq 0$ then we also have an injective morphism $\psi_{ij}\colon \calA(\eta,i) \hookrightarrow \calA(\eta,i+j)$. \begin{remk}\label{sheafgeneralities} The analogues of Lemmas \ref{cohomologous} and \ref{inducedmap} hold for sheaves of cdgas and the dga-sheaf-morphisms between them. If a sheaf of cdgas $(\calA, \wedge, d)$ is equipped with an increasing cdga-filtration $W_{\lc}$, then as in Subsection \ref{thickenedcomplexesandfiltrations}, the thickened complex of sheaves $\calA(\eta, m)$ inherits this filtration, provided that $\eta$ is a global section of $W_1\calA$. The same holds if we replace $W_{\lc}$ by a decreasing filtration. The analogues of Lemmas \ref{cohomologousfiltered}, \ref{inducedquasi}, and \ref{filtrationinterplay} are then also valid.$\sslash$ \end{remk} \begin{remk}\label{remk:torsionForSheaves} The analogue of Lemma~\ref{lem:torsion2} holds for a complex of sheaves of free $R_\infty$-modules. This can be seen by applying Lemma~\ref{lem:torsion2} to a free resolution of a complex representing the hypercohomology. Therefore, the analogue of Corollary~\ref{torsion} holds for the hypercohomology sheaves of a thickening of a sheaf of cdgas $(\calA, \wedge, d)$.$\sslash$ \end{remk} \section{Thickened Complexes and Mixed Hodge Complexes}\label{sec4} The main goal of this section is to prove that the thickened complex of a multiplicative mixed Hodge complex is again a mixed Hodge complex, provided that the $1$-forms used to conduct thickenings belong to $W_1$ and $F^1$ when applicable. \subsection{Denotations and Assumptions}\label{denotationsandassumptionsthickenedcomplexesandmixedhodgecomplexes} Let $k$ denote a subfield of $\R$ and $X$ denote a topological space. Let $\calK^\bullet = ((\calK_k^\bullet, W_{\lc}), (\calK^\bullet_\C, W_{\lc},F^{\lc}), \alpha)$ denote a multiplicative $k$-mixed Hodge complex on $X$ with distinguished representative for $\alpha$ given by: $$ \begin{tikzcd}[column sep = small] (\calK^\bullet_k, W_{\lc}) = (\calL^{\bullet}_0, W_{\lc}) \arrow{r}{\alpha_1} & (\calL^\bullet_1, W_{\lc})& \arrow{l}[swap]{\alpha_2} (\calL^\bullet_2, W_{\lc}) \arrow{r}{\alpha_3} & \cdots \arrow{r}{\alpha_{2r+1}} & (\calL^\bullet_{2r+1}, W_{\lc}) = (\calK^\bullet_\C,W_{\lc}), \end{tikzcd} $$ where $\alpha_{\textrm{odd}}$ are right arrows and $\alpha_{\textrm{even}}$ are left arrows. To thicken $\calK^\bullet$ as a mixed Hodge complex, we need to thicken each $\calL^{\bullet}_i$ with an element from $\Gamma(X, \calL^1_{i})\cap \ker d$. It is in fact enough to choose such global $1$-cycles from the even-numbered sheaves $\calL^{1}_{\textrm{even}}$, provided their images in the odd-numbered $\calL^{1}_{\textrm{odd}}$ are cohomologous. Specifically, we need the following assumption on chosen sections $\eta_\ell\in \Gamma(X, \calL^1_{2\ell})\cap \ker d$ to guarantee that the differentials in the lifting are compatible with the filtrations and that the maps $\alpha_i$ extend to psuedo-isomorphisms. \begin{assumption}\label{as} The elements $\eta_\ell\in \Gamma(X, \calL^1_{2\ell})\cap \ker d$ where $0 \leq \ell \leq r$ satisfy: \begin{enumerate} \item $\eta_\ell\in W_1\Gamma(X, \calL^1_{2\ell})\cap \ker d$; \item $\alpha_{2r+1}\eta_r\in F^1\Gamma(X, \calK^\bullet_\C)$; \item the images of $\eta_\ell$ in $\calL^1_{\textrm{odd}}$ are cohomologous, i.e., \[ \alpha_1 \eta_0 \simeq_{W_1} \alpha_2 \eta_1,\quad \alpha_3 \eta_1 \simeq_{W_1} \alpha_4 \eta_2,\quad \dots,\quad \alpha_{2r-1} \eta_{r-1} \simeq_{W_1} \alpha_{2r} \eta_r, \] where $\simeq_{W_1}$ denotes the relation of being cohomologous in $W_1$. For $\ell \geq 1$ fix choices of $a_\ell \in W_1\Gamma(X,\calL_{2\ell-1}^0)$ such that $\alpha_{2\ell-1}\eta_{\ell-1} - \alpha_{2\ell} \eta_\ell = da_\ell$. \end{enumerate} \end{assumption} \subsection{Thickened Complexes of Mixed Hodge Complexes}\label{thickenedcomplexesofmixedhodgecomplexes} \textit{Fix $m \geq 1$ for this subsection.} Before thickening our complexes, we point out that, if $\ell \geq 1$, then $\calL^\bullet_\ell$ is a sheaf of $\C$-cdgas with filtration(s) defined over $\C$. This is part of the definition of a multiplicative mixed Hodge complex. Using Remark \ref{sheafgeneralities} and Assumption \ref{as}, we obtain a chain of morphisms of thickened complexes: $$ \begin{tikzcd}[column sep = small] \calK^\bullet_k(\eta_0,m) \arrow{r}{\alpha_{1\#}} & \calL^\bullet_1(\alpha_1\eta_0, m) \ar[rr, "\exp(a_1 \otimes s)", "\cong"'] & & \calL^\bullet_1(\alpha_2\eta_1,m) & \arrow{l}[swap]{\alpha_{2\#}} \calL^\bullet_2(\eta_1,m) \arrow{r}{\alpha_{3\#}} & \cdots \arrow{r}{\alpha_{2r+1\#}} & \calK^\bullet_\C(\alpha_{2r+1}\eta_r,m) \end{tikzcd} $$ where all but the leftmost complex have been thickened over $\C$, not $k$. Note that $\alpha_{1\#}$ is technically a composition: first the induced morphism of thickened $k$-complexes, then the map from the thickened $k$-complex to the thickened $\C$-complex induced by $ R_m = R_m \otimes k \xrightarrow{\id \otimes \mathrm{incl}} R_m \otimes \C$. We define $\alpha_{\#}\colon \calK^\bullet_k(\eta_0,m) \dashrightarrow \calK_\C^\bullet(\alpha_{2r+1}\eta_r,m)$ to be the pseudo-morphism associated to the above chain. Because the thickenings are conducted with sections of $W_{1}\calL^1$, it follows by Remark \ref{sheafgeneralities} that the thickened complexes of sheaves inherit their own increasing filtrations $W_{\lc}$. Moreover, $\calK^\bullet_\C(\alpha_{2r+1}\eta_r, m)$ inherits a decreasing filtration $F^{\lc}$, because $\alpha_{2r+1}\eta_r$ belongs to $F^1\Gamma(X, \calK^\bullet_C)$. \begin{thm}\label{mhsthickened} Under the above notations and assumptions, the triple \[ \calK^\bullet(\eta,m) \coloneqq \Big(\big(\calK_k^\bullet(\eta_0,m),W_{\lc}\big), \big(\calK^\bullet_\C(\alpha_{2r+1}\eta_r,m),W_{\lc},F^{\lc}\big), \alpha_{\#}\Big) \] is a $k$-mixed Hodge complex of sheaves on $X$. \end{thm} \begin{proof} We need to verify the following assertions: \begin{enumerate}[(i)] \item $\alpha_\#$ preserves the $W_{\lc}$-filtrations; \item $ \alpha_\# \otimes 1\colon \calK_k^\bullet(\eta_0,m) \otimes \C \dasharrow \calK^\bullet_\C(\alpha_{2r+1}\eta_r,m) $ is a $W_{\lc}$-filtered pseudo-isomorphism; \item the $i$-th $W$-graded component of $\calK^\bullet(\eta,m)$ is a pure Hodge complex of weight $i$. \end{enumerate} To show statements (i) and (ii), we apply the sheaf analogues of Lemma \ref{cohomologousfiltered} and Lemma \ref{inducedquasi}. The chain of morphisms representing $\alpha_\#$ is $W_{\lc}$-filtered, and it induces on the $i$-th $W$-graded component the pseudo-morphism: \begin{align*} \Gr^W_i(\alpha_{\#}) = \bigoplus_{j=0}^{m-1} \Gr_{i+2j}^W \alpha \otimes \mathrm{incl}\colon \bigoplus_{j=0}^{m-1} \Gr_{i+2j}^W \calK^\bullet_k \otimes k\langle s^j \rangle \dasharrow \bigoplus_{j=0}^{m-1} \Gr_{i+2j}^W \calK^\bullet_\C \otimes_\C \C\langle s^j \rangle. \end{align*} Therefore $\Gr^W_i(\alpha_{\#} \otimes 1) = \bigoplus_{j=0}^{m-1} \Gr_{i+2j}^W (\alpha \otimes 1) \otimes_\C \id$. Each $\Gr_{i+2j}^W (\alpha \otimes 1)$ is a quasi-isomorphism, because $\calK^\bullet$ is a mixed Hodge complex. Therefore, as their direct sum, $\Gr^W_i(\alpha_{\#}\otimes 1)$ is also a quasi-isomorphism. To prove (iii), let us fix $i \in \Z$. We have just verified that the $i$-th graded component of $\calK^\bullet(\eta,m)$ is: \begin{align*} \bigoplus_{j = 0}^{m-1}\Big(\Gr^W_{i+2j}\calK^\bullet_k \otimes k\langle s^j \rangle, \left(\Gr^W_{i+2j}\calK^\bullet_\C \otimes \C\langle s^j \rangle, F^{\lc}\right), \Gr_{i+2j}^W\alpha \otimes \mathrm{incl}\Big) \end{align*} and we need to show that this is a Hodge complex of weight $i$. It is enough to check that, for all $0\leq j\leq m-1$, the $j$-th direct summand above is a Hodge complex of weight $i$. By Lemma \ref{filtrationinterplay}, we have: \begin{align*} F^p \left(\Gr_{i+2j}^W\calK^\bullet_\C \otimes \C \langle s^j \rangle\right) = \left(F^{p+j}\Gr_{i+2j}^W\calK^\bullet_\C\right) \otimes \C\langle s^j \rangle \end{align*} for $p\in \Z$. Therefore, the $j$-th direct summand is isomorphic to the $j$-th Tate twist of the following weight $i+2j$ Hodge complex of sheaves: \begin{align*} \Big(\Gr^W_{i+2j}\calK^\bullet_k, \left(\Gr^W_{i+2j}\calK^\bullet_\C, F^{\lc}\right), \Gr_{i+2j}^W\alpha\Big). \end{align*} The $j$-th Tate twist of a weight $i+2j$ Hodge complex is a weight $i$ Hodge complex. \end{proof} \begin{cor}\label{mhshypercohomology} The $k$-mixed Hodge complex of sheaves $\calK^\bullet(\eta,m)$ induces a $k$-mixed Hodge structure on the hypercohomology $\mathbb{H}^*(X, \calK^\bullet_k(\eta_0,m))$. \end{cor} \begin{proof} See Remark \ref{transvstate}. \end{proof} Recall that if $ i \geq 1,j \geq 0$, then the natural map $R_{i+j} \twoheadrightarrow R_i$ induces a morphism $\phi_{ji}$ of thickened complexes. Furthermore, the natural map $R_i \hookrightarrow R_{i+j}$ induces a morphism $\psi_{ij}$ of thickened complexes. \begin{lem}\label{inducedmhsmaps} If $i \geq 1,j \geq 0$, then the following are morphisms of $k$-mixed Hodge complexes of sheaves. The numbers $(-j)$ and $(-1)$ denote Tate twists. \begin{enumerate} \item $\phi_{ji}\colon \calK^\bullet(\eta,j+i) \twoheadrightarrow \calK^\bullet(\eta,i),$ \item $\psi_{ij}\colon \calK^\bullet(\eta,i) \hookrightarrow \calK^\bullet(\eta,i+j)(-j),$ \item Multiplication by $s$, $S_i:\calK^\bullet(\eta,i) \rightarrow \calK^\bullet(\eta,i)(-1)$. \end{enumerate} In particular, the following are morphisms of $k$-mixed Hodge structures: \begin{enumerate} \item $\phi_{ji}^*\colon \mathbb{H}^*(X,\calK^\bullet_k(\eta_0,j+i)) \rightarrow \mathbb{H}^*(X,\calK^\bullet_k(\eta_0,i)),$ \item $\psi_{ij}^*\colon \mathbb{H}^*(X,\calK^\bullet_k(\eta_0,i)) \rightarrow \mathbb{H}^*(X,\calK^\bullet_k(\eta_0,i+j))(-j),$ \item Multiplication by $s$, $ \mathbb{H}^*(X,\calK^\bullet_k(\eta_0,i)) \rightarrow \mathbb{H}^*(X,\calK^\bullet_k(\eta_0,i))(-1)$. \end{enumerate} \end{lem} \begin{proof} Recall the definition of the projection $\phi_{ji}$ (at the beginning of Section~\ref{sec3}) and the inclusion $\psi_{ij}$ (Remark~\ref{remk:psi-ij}). Observe that $\phi_{ji}$ and $\psi_{ij}$ do indeed induce morphisms between the pseudo-morphism representatives of the two mixed Hodge complexes. It remains to check their compatibility with the weight and Hodge filtrations. The morphism $\phi_{ji}$ clearly preserves both filtrations; furthermore, a morphism of $k$-mixed Hodge complexes of sheaves always induces morphisms of $k$-mixed Hodge structures on hypercohomology \cite[Theorem 3.18III]{peters2008mixed}. It is also straightforward to check that $\psi_{ij}$ maps $W_{\lc}$ into $W_{{\lc}-2j}$ and $F^{\lc}$ into $F^{{\lc}-j}$. In other words, $\psi_{ij}\colon \calK^\bullet(\eta,i) \rightarrow \calK^\bullet(\eta,i+j)(-j)$ is a morphism of $k$-mixed Hodge complexes of sheaves. Pass to hypercohomology and use that Tate twists commute with taking hypercohomology (see \cite[Theorem 3.18IIi]{peters2008mixed}) to obtain the second statement. To show part (3), we observe that $S_i$ equals the composition $\psi_{i-1,1}\circ \phi_{1,i-1}$. \end{proof} \begin{remk}\label{translatedlemmas} Later in the paper, we will be interested not in $\calK^\bullet(\eta,m)$ itself, but rather its translation $\calK^\bullet(\eta,m)[1]$. The valid analogue of Lemma \ref{mhshypercohomology} is that $\calK^\bullet(\eta,m)[1]$ induces a $k$-mixed Hodge structure on $\mathbb{H}^*(X, \calK^\bullet_k(\eta_0,m)[1])$. Note that, by Remark \ref{transvstate}, $\mathbb{H}^*(X, \calK^\bullet_k(\eta_0,m)[1]) \cong \mathbb{H}^{*+1}(X, \calK^\bullet_k(\eta_0,m))(1)$ where a Tate twist has been taken on the right hand side. The valid analogue of Lemma \ref{inducedmhsmaps} is that for $ i \geq 1,j\ge 0$, the following are morphisms of mixed Hodge structures: \begin{align*} \phi_{ji}[1]^*&\colon \mathbb{H}^*(X,\calK^\bullet_k(\eta_0,i+j)[1]) \rightarrow \mathbb{H}^*(X,\calK^\bullet_k(\eta_0,i)[1]);\\ \psi_{ij}[1]^*&\colon \mathbb{H}^*(X,\calK^\bullet_k(\eta_0,i)[1]) \rightarrow \mathbb{H}^*(X,\calK^\bullet_k(\eta_0,i+j)[1])(-j);\\ (s\cdot)^*&\colon \mathbb{H}^*(X,\calK^\bullet_k(\eta_0,i)[1]) \rightarrow \mathbb{H}^*(X,\calK^\bullet_k(\eta_0,i)[1])(-1). \sslash \end{align*} \end{remk} \section{Mixed Hodge Structures on Alexander Modules}\label{mhsal} \subsection{Denotations and Assumptions} Let $k = \R$. Let $U$ be a smooth connected complex algebraic variety of dimension $n$. Let $f\colon U \rightarrow \C^*$ be an algebraic map such that $f_*\colon \pi_1(U)\to \pi_1(\C^*)$ is surjective. Let $U^f \rightarrow U$ denote the infinite cyclic cover as in Section~\ref{ssAlex}. Choose a good compactification $X$ of $U$ with simple normal crossing divisor $D = X\setminus U$, such that $f\colon U \rightarrow \C^*$ extends to an algebraic map $\bar{f}\colon X \rightarrow \C P^1$. Let \[ \calH dg^\bullet(X\, \log D) = \Big(\big(j_*\calE^\bullet_U, \tau_{\lc}\big), \big(\logdr{X}{D}, W_{\lc}, F^{\lc}\big), \alpha \Big) \] denote the $\R$-mixed Hodge complex of sheaves on $X$ of Theorem \ref{logmhs}. We will consider a choice of $X$ fixed, until Theorem~\ref{indcompactification}, when we prove that our constructions are independent of this choice. \subsection{Local Systems and Thickened Real de Rham Complexes}\label{slocal} The main result of this subsection is Proposition~\ref{propLocal}. Throughout this section, we will let $k=\R$. As in Section~\ref{ssAlex}, Let $\calL=\pi_!\underline{\mathbb R}_{U^f}$, and let $\ov{\mathcal L}$ have the conjugate $R$ action. $\ov{\mathcal L}$ has stalks isomorphic to $R = \R[t^{\pm 1}]$ and action $\pi_1(U) \rightarrow \Aut(R)$ given by $\gamma \mapsto \left(t^{-f_*(\gamma)}\right)\, \cdot\, $, multiplication by $t^{-f_*(\gamma)}$. This is a rank one local system of $R$-modules. Consider $\ov\calL \otimes_R R_m$, which has stalks isomorphic to $R_m = R/(t-1)^mR$ and action $\pi_1(U) \rightarrow \Aut(R_m)$ given by $\gamma \mapsto \left((1+s)^{-f_*(\gamma)}\right)\, \cdot\, $, multiplication by $(1+s)^{-f_*(\gamma)}$ (recall that $1+s=t$). This is a rank one local system of $R_m$-modules. Next, consider the thickened complex $\calE^{\bullet}_{U}\left(\Im\, df/f,m\right)$, given by \begin{align*} \calE^0_{U} \otimes R_m \xrightarrow{d^0_m = d +\Im\, df/f \wedge \otimes s} \calE^1_{U} \otimes R_m \xrightarrow{d^1_m = d +\Im\, df/f \wedge \otimes s} \calE^2_{U} \otimes R_m \rightarrow \cdots \end{align*} with differential given by $d_m\colon \omega \otimes 1 \mapsto d\omega \otimes 1 + \left(\Im\, df/f \wedge \omega\right) \otimes s$. Here, $\Im$ denotes the imaginary part. \begin{prop} Let \[ \tilde{s}\coloneqq \exp(2\pi s)-1=\sum_{n=1}^\infty \frac{(2\pi s)^{n}}{n!}\in R_\infty. \] Then \[ \calE^\bullet_U\left(\Im\, df/f,\infty\right)_{\tilde s}\coloneqq \displaystyle\varprojlim_m \left(\calE^\bullet_U\left(\Im\, df/f,m\right)_{\tilde s}\right) \] is a resolution of $\ov\calL \otimes_R R_\infty$, via a canonical map $\ov\calL \otimes_R R_\infty\to \calE^0_U\left(\Im\, df/f,\infty\right)_{\tilde s}$ (defined in Remark~\ref{rem:nu}). \label{propLocal} \end{prop} Note that $\tilde{s}=su$, where $u$ is a unit in $R_\infty$, so we are in the situation described by Remark \ref{twistedmodule}. We will prove Proposition \ref{propLocal} at the end of this subsection. In simply connected open sets, one can check locally that $\ker d^k_m= \im\, d^{k-1}_m$ for all $k\geq 1$. Thus, $\calE^\bullet_U\left(\Im\, df/f,m\right)$ is a resolution of $\ker\, d^0_m$, which in turn is a rank one local system of $R_m$-modules. We give a description of this local system. \begin{lem}\label{lemKer} If $V \subset U$ is a small enough connected open subset (in the analytic topology) such that $f(V)$ is simply-connected, then $\ker\, d^0_m(V)$ consists of all functions of the form $\exp\left(-\arg f \otimes s\right)g(s) \in \calE^0_U(V) \otimes R_m$, where $g(s)\in R_m$ and $\arg$ can be taken to be any branch of the argument on $f(V)$. \end{lem} \begin{proof} On $V$, the differential is given by $d^0_m(V)(\omega \otimes 1) = d\omega \otimes 1+\left(d\arg(f) \wedge \omega\right) \otimes s$. Let $\sum_{k=0}^{m-1} g_k\otimes s^k$ be an element of $\ker d^0_m(V)$. Then, \[ 0=d^0_m\left(\sum_{k=0}^{m-1} g_k\otimes s^k\right)= d g_0\otimes 1+ \sum_{k=1}^{m-1} \big(d g_k+g_{k-1} d\arg(f)\big)\otimes s^k . \] Hence, $g_0=c_0\in\R$, $dg_1=-c_0 d\arg f$, so $g_1=-c_0 \arg(f)+c_1$ for some $c_1\in \R$ (choosing a different branch of the argument will result in a different $c_1$). Inductively, we get that $$ g_k=\sum_{j=0}^k \frac{(-1)^j}{j!} c_{k-j} \arg^j(f) $$ for some $c_j\in\R$, $j=0, 1, \ldots, k$. Hence, \begin{align*} &\sum_{k=0}^{m-1} g_k\otimes s^k =\sum_{k=0}^{\infty}\left(\sum_{j=0}^k \frac{(-1)^j}{j!} c_{k-j} \arg^j f\right)\otimes s^k=\\ & =\sum_{j=0}^{\infty}\sum_{k=j}^{\infty} \frac{(-1)^j}{j!} c_{k-j} \arg^j f \otimes s^k= \sum_{j=0}^{\infty}\sum_{l=0}^{\infty} \left(\frac{(-1)^j}{j!} \arg^j f \otimes s^j\right)c_ls^l=\\ &=\left(\sum_{l=0}^{\infty} c_l s^l\right)\exp\big(-\arg(f)\otimes s\big)= \left(\sum_{l=0}^{m-1} c_l s^l\right)\exp\big(-\arg(f)\otimes s\big), \end{align*} which finishes our proof. \end{proof} \begin{remk}\label{rem:IVP} By the proof of the previous lemma, the elements of $\ker d^0_m$ form a local system of rank $m$ over $\R$. In particular, the following map is an $R$-linear isomorphism for any simply connected open set $V$ and any $x\in V$: \[ \begin{array}{rcl} \Gamma(V,\ker d^0_m) & \longrightarrow & R_m \\ \sum_{k=0}^{m-1} g_k\otimes s^k &\longmapsto & \sum_{k=0}^{m-1} g_k(x)s^k.\sslash \end{array} \] \end{remk} \begin{lem} The $\pi_1(U)$-action on the local system $\ker\, d^0_m$ is given by $\gamma \mapsto \exp(-2\pi f_*(\gamma) s)\, \cdot\, $, multiplication by $\exp(-2\pi f_*(\gamma) s)$. \label{lemPi1} \end{lem} \begin{proof} The stalk of $\ker\, d^0_m$ at any point in $U$ is generated by $\exp(-\arg(f) \otimes s)$. A loop $\gamma\in\pi_1(U)$ takes $\arg(f)$ to $\arg(f)+2\pi f_*(\gamma) $. Thus, the action of $\gamma \in \pi_1(U)$ on $\ker\, d^0_m$ is: \begin{align*} & \gamma \cdot \exp(-\arg(f) \otimes s) = \exp\left(-\left(\arg(f) + 2\pi f_*(\gamma) \right) \otimes s\right) \\ & =\exp(-2\pi f_*(\gamma) \otimes s) \cdot \exp(-\arg(f) \otimes s)= \exp(-2\pi f_*(\gamma) s) \cdot \exp(-\arg(f) \otimes s). \qedhere \end{align*} \end{proof} Consequently, we have two rank one local systems of $R_m$-modules on $U$: \begin{itemize} \item $\ov\calL \otimes_R R_m$ with action $\gamma \mapsto (1+s)^{-f_*(\gamma)} \, \cdot\, $ (multiplication by $(1+s)^{-f_*(\gamma)}$), for all $\gamma\in\pi_1(U)$. \item $\ker\, d^0_m$ with action $\gamma \mapsto \exp(-2\pi f_*(\gamma) s)\, \cdot \,$ (multiplication by $\exp(-2\pi f_*(\gamma) s)$), for all $\gamma\in\pi_1(U)$. \end{itemize} Because these actions are distinct, $\ov\calL \otimes_R R_m$ and $\ker\, d^0_m$ are not isomorphic as local systems of $R_m$-modules, but they are if one considers the twisted module structure $(\ker\, d^0_m)_{\tilde s}$. The following lemma is a consequence of the discussion above. \begin{lem} The kernel of the $0$-th differential of $\calE_{U}^\bullet\left(\Im\, df/f,m\right)_{\tilde s}$ is $(\ker\, d^0_m)_{\tilde s}$, which is isomorphic to $\ov\calL\otimes_R R_m$ as local systems of $R_m$-modules. \label{lemRm} \end{lem} \begin{remk}\label{rem:nu} We can fix a canonical isomorphism $\nu\colon \ov\calL\otimes_R R_m\cong (\ker\, d^0_m)_{\tilde s}$ as follows. Recall that at any $x\in U$, $\ov\calL_x$ has an $\R$-basis $\{\ov\delta_{x'}\}$ parametrized by $x'\in \pi^{-1}(x)$ (see Remark~\ref{rem:oppositeDual}). For any such $x'$, let $\nu_x( \ov\delta_{x'}) $ be the unique germ $g$ in the stalk of $(\ker\, d^0_m)$ at $x$ such that $ g(x) = \exp(-\Im f_\infty(x')\otimes s)$. Recall that according to our definitions of $U^f$ and ${\mathcal L}$ in Section~\ref{ssAlex}, the map $f_\infty$ is fixed as the projection $U^f\subset U\times \C\to \C$. Since $\exp\circ f_\infty = f\circ \pi$, we can describe $\nu$ as follows. On any simply connected neighborhood $V$ of $x$, let $\iota\colon V\to U^f$ be the section of $\pi$ such that $\iota(x)=x'$. Then $\Im f_\infty \circ \iota$ is a branch of $\arg f$, so \[ \nu_x(\ov\delta_{x'}) =\exp(-\Im f_\infty\circ \iota\otimes s). \] Let us check that $\nu_x$ is an $R_m$-linear map on stalks. \begin{align*} \nu_x(t\ov\delta_{x'})& = \nu_x(\ov\delta_{t^{-1}x'}) = \exp(-\Im f_\infty\circ t^{-1}\circ \iota\otimes s)= \exp(-\Im f_\infty\circ \iota\otimes s + 2\pi \otimes s) = \\ &=\nu_x(\ov\delta_{x'})e^{2\pi s} =\nu_x(\ov\delta_{x'})*_{\widetilde s}t. \end{align*} Note that, in the formula above, $t$ is seen both as an element of $R_m$ and as a deck transformation. We can see that $\nu$, which we have defined on stalks, gives us a morphism of local systems over $R_m$. Indeed, from the definition of $\nu$, it agrees with the monodromy action of any path $\gamma$ from $x$ to any given $y\in U$, i.e. $\gamma\cdot \nu_x(\ov\delta_{x'}) = \nu_y(\ov\delta_{\gamma \cdot x'}) =\nu_y(t^{-1}\ov\delta_{ x'}) = \nu_y(\gamma\cdot \ov\delta_{x'})$.$\sslash$ \end{remk} \begin{remk}\label{remk:thickenedComplex} The proof of the claims in this remark are straightforward following the definitions of the objects involved, so we omit them. The thickened complex ${\mathcal E}^\bullet_U\left( \Im\frac{df}{f} ,m\right)$ of Lemma~\ref{lemRm} is isomorphic (as sheaves of $\R$-vector spaces) to the complex ${\mathcal E}^\bullet_U(m)_{\log}$, which is defined as ${\mathcal E}^i_U(m)_{\log} = {\mathcal E}^i_U \otimes_\R R_m$ and whose differential is the following, where $\log(1+s)$ represents the power series at $s=0$: \[ d_{\log}\left(\alpha\otimes s^j\right) = d\alpha\otimes s^j+\Im\frac{df}{f} \wedge \alpha\otimes \frac{ \log(1+s)}{2\pi}s^j. \] The ($\R$-linear) isomorphism is determined by: \[ \ff{{\mathcal E}_U^i\left( \Im\frac{df}{f} ,m\right)}{{\mathcal E}^i_U(m)_{\log}}{\alpha\otimes s^j}{\alpha\otimes \left(\frac{\log(1+s)}{2\pi}\right)^j.} \] Note that the above isomorphism becomes $R_m$-linear if we give ${\mathcal E}_U^{\bullet}\left( \Im\frac{df}{f} ,m\right)$ the twisted $R_m$-module structure ${\mathcal E}_U^{\bullet}\left( \Im\frac{df}{f} ,m\right)_{\wt s}$ of Proposition~\ref{propLocal}. \textit{We will use this $\R$-linear identification between ${\mathcal E}_U^i\left( \Im\frac{df}{f} ,m\right)$ and ${\mathcal E}^i_U(m)_{\log}$ from now on whenever it is convenient}. It will simplify our notation later, especially in Section~\ref{sec:main}. Using Lemma~\ref{lemRm}, we see that the composition $\overline{{\mathcal L}}\otimes_R R_m \xrightarrow{\nu}{\mathcal E}_U^{\bullet}\left( \Im\frac{df}{f} ,m\right)_{\wt s}\to {\mathcal E}^{\bullet}_U(m)_{\log}$ is an \bi{$\bm{ R_m}$-linear} quasiisomorphism, i.e. ${\mathcal E}^{\bullet}_U(m)_{\log}$ is a resolution of $\overline{{\mathcal L}}\otimes_R R_m$ as sheaves of $R_m$-modules. The resolution $\overline{{\mathcal L}}/s^m\xrightarrow{\eta_m} {\mathcal E}^{\bullet}_U(m)_{\log}$ is determined by the map $\eta_m$, analogous to the map $\nu$ in Remark~\ref{rem:nu}: using the same notation, $\eta_m$ is given by: $$ \f{(\eta_m)_x}{(\overline{{\mathcal L}}/s^m)_x}{({\mathcal E}^0_U(m)_{\log})_x}{\ov\delta_{x'}}{\exp(-\Im f_\infty \circ \iota \otimes \frac{\log(1+s)}{2\pi}).} $$ Manipulating power series we can see the expression above as \[ \eta_m(\ov\delta_{x'}) = (1+s)^{-\frac{1}{2\pi}\Im f_\infty \circ \iota} = \sum_{i=0}^{m-1} \binom{-\frac{1}{2\pi}\Im f_\infty \circ \iota}{i} \otimes s^i.\sslash \] \end{remk} We now prove a result that will help us understand $\displaystyle\varprojlim_m \left(\calE^\bullet_U\left(\Im\, df/f,m\right)_{\tilde s}\right)$. \begin{lem} Let $\left(\{\calF_m^{\bullet}\}_{m\in \N},\left\{\alpha^\bullet_{m,j}\right\}_{m\geq j\in \N}\right)$ be an inverse system of complexes of sheaves of $R_{\infty}$-modules on a topological space that has a basis of simply connected open sets, such that the maps $\alpha_{m,j}^k(V)\colon \calF_m^k(V)\rightarrow\calF_j^k(V)$ are surjective for all $m\geq j$ and for every simply connected open set $V$. Suppose that $\calF_m^{\bullet}(V)$ are exact complexes for all $m\in\N$ and for every simply connected open set $V$. Then, $\varprojlim_m \calF_m^\bullet$ is also an exact complex of sheaves of $R_\infty$-modules. \label{clExact} \end{lem} \begin{proof} We need to show that the stalk of $\varprojlim_m \calF_m^\bullet$ at every point is exact. Since $U$ has a basis of simply connected open sets, and since \[ \Big(\displaystyle\varprojlim_m \calF_m^k\Big)(V)\simeq \displaystyle\varprojlim_m \calF_m^k(V) \] for every $k$, it suffices to show that $\displaystyle\varprojlim_m \calF_m^\bullet(V)$ is exact for any simply connected open subset $V$ of $U$. For a fixed $m$, the exact complex $\calF_m^{\bullet}(V)$ gives rise to short exact sequences \[ 0\to \ker(d^k_m)\to \calF_m^{k}(V) \to \ker (d^{k+1}_m)\to 0, \] where $d^k_m\colon \calF_m^{k}(V)\to \calF_m^{k+1}(V)$ are the differentials. By assumption, the homomorphisms $\alpha_{m, j}^k$ in the inverse system $\displaystyle\varprojlim_m \calF_m^k(V)$ are surjective, so the inverse system satisfies the Mittag-Leffler (ML) condition. Using the fact that the spaces $\ker (d^{k+1}_m)$ are quotients of $\calF_m^{k}(V)$ for all $k$, we can see that the restrictions $\alpha_{m, j}^k\colon \ker (d^{k+1}_{l})\to \ker (d^{k+1}_{m})$ are surjective, so the inverse system $\displaystyle\varprojlim_m \ker(d^k_m)$ also satisfies (ML). By \cite[Chapter II, Proposition 9.1 (b)]{Hartshorne}, we have exact sequences, \[ 0\to \displaystyle\varprojlim_m\ker(d^k_m)\to \displaystyle\varprojlim_m\calF_m^{k}(V) \to\displaystyle\varprojlim_m \ker (d^{k+1}_m)\to 0, \] for every $k$. The above short exact sequences for all $k$ induce the following long exact sequence \[ \cdots\to \displaystyle\varprojlim_m\calF_m^{k-1}(V) \to \displaystyle\varprojlim_m\calF_m^{k}(V) \to \displaystyle\varprojlim_m\calF_m^{k+1}(V) \to \cdots, \] thus completing the proof of the lemma. \end{proof} Now, we are ready to prove Proposition \ref{propLocal}, the main result of this section. \begin{proof}[Proof of Proposition~\ref{propLocal}] The result about the resolution follows from applying Lemma \ref{clExact} to the exact complex of sheaves of $R_m$-modules \[ \ov\calL\otimes_R R_m\rightarrow \calE_U^\bullet\left(\Im\, df/f,m\right)_{\tilde s}. \] Notice that $\varprojlim \ov\calL \otimes_R R_m$ is naturally isomorphic to $\ov\calL \otimes_R R_{\infty}$. \end{proof} \subsection{Local Systems and Thickened Complexes: Rational Version} We now state the analogue of the results in Section~\ref{slocal} in the $k=\Q$ case, for the thickened complex $j^{-1}\calK_\infty^\bullet(1\otimes f, m)$ constructed from the complex $\calK_\infty^\bullet$ of Section \ref{rationalmixedhodgecomplexonsmoothvarieties}. In the remainder of this section, $k=\Q$, $\ov\calL$ is a local system of $\Q[t^{\pm 1}]$-modules, $d^i_m$ is the $i$-th differential of $j^{-1}\calK_\infty^\bullet(1\otimes f, m)$, and $\tilde{s}_2\coloneqq \exp(s)-1\in R_\infty$. Abusing notation, we will consider $\tilde{s}_2$ as an element of $R_m$ or $\C[t^{\pm 1}]_m$ for all $m\geq 1$ whenever it makes sense. As stated in Section \ref{rationalmixedhodgecomplexonsmoothvarieties}, $\varphi_{\infty}\otimes 1\colon (\calK_\infty^\bullet,W_{\lc})\otimes \C\rightarrow (\Omega_X^\bullet(\log D),W_{\lc})$ is a filtered quasi-isomorphism, which takes $1\otimes f$ to $\frac{1}{2\pi i}\frac{df}{f}$. So it induces a quasi-isomorphism $j^{-1}\calK_\infty^\bullet\otimes \C\rightarrow \Omega_U^\bullet$. By Remark \ref{sheafgeneralities}, this induces a quasi-isomorphism \begin{equation} F\colon j^{-1}\calK_\infty^\bullet(1\otimes f, m)\otimes \C\rightarrow \Omega_U^\bullet\left(\frac{1}{2\pi i}\frac{df}{f}, m\right) \label{eqnF} \end{equation} for all $m\geq 1$. One can check locally in simply connected open sets that $\Omega_U^\bullet\left(\frac{1}{2\pi i}\frac{df}{f}, m\right)$ is exact at place $k$ for all $k\geq 1$. Therefore, since $\C$ is faithfully flat over $\Q$, $\ker d^{k}_m=\im d_m^{k-1}$ for all $k\geq 1$. Hence, $j^{-1}\calK_\infty^\bullet(1\otimes f, m)$ is a resolution of $\ker\, d^0_m$, which in turn is a rank one local system of $R_m$-modules. Now, we give a description of this local system. \begin{lem}[Analogue of Lemma \ref{lemKer}] If $V \subset U$ is a small enough open subset (in the analytic topology) such that $f(V)$ is simply-connected, then $\ker\, d^0_m(V)$ is generated by $\exp\left(-\frac{\log(f)}{2\pi i}\otimes 1 \otimes s\right) \in \calK^0_\infty(V) \otimes R_m$ as an $R_m$-module, where $\log$ is taken to be a branch of the logarithm function which is defined on $f(V)$. \end{lem} \begin{proof} A direct computation shows that $\exp\left(-\frac{\log(f)}{2\pi i}\otimes 1 \otimes s\right)\in \ker\, d^0_m(V)$. We have that $$ F\left(\exp\Big(-\frac{\log(f)}{2\pi i}\otimes 1 \otimes s\Big)\otimes 1\right)=\exp\left(-\frac{\log(f)}{2\pi i} \otimes s\right)\in \Omega_U^0\otimes \C[t^{\pm 1}]_m, $$ where $F$ is the quasi-isomorphism in equation (\ref{eqnF}). Following a similar argument as in Lemma \ref{lemKer} for the thickened {holomorphic} de Rham complex $\Omega_U^\bullet\left(\frac{1}{2\pi i}\frac{df}{f}, m\right)$ instead of $\calE_U^\bullet\left(\Im\big(\frac{df}{f}\big),m \right)$, we get that $\exp\Big(-\frac{\log(f)}{2\pi i} \otimes s\Big)$ generates the kernel of the $0$-th differential of $\Omega_U^\bullet\left(\frac{1}{2\pi i}\frac{df}{f}, m\right)$. Let $A$ be the $R_m$-module generated by $\exp\left(-\frac{\log(f)}{2\pi i}\otimes 1 \otimes s\right)$. Using the map $F$, we see that the inclusion $A\hookrightarrow \ker\, d^0_m(V)$ becomes an isomorphism after tensoring by $\C$. Since $\C$ is faithfully flat over $\Q$, it follows that $A\cong \ker\, d^0_m(V)$. \end{proof} The proof of the rest of the results in this sections follow the same steps as the proof of their analogue results in the real case, and we will omit them. \begin{lem}[Analogue of Lemma \ref{lemPi1}] The $\pi_1(U)$-action on the local system $j^{-1}\ker\, d^0_m$ is given by $\gamma \mapsto \exp(-f_*(\gamma) s)\, \cdot\, $, multiplication by $\exp(- f_*(\gamma) s)$. \end{lem} \begin{lem}[Analogue of Lemma \ref{lemRm}]\label{lemRmQ} The kernel of the $0$-th differential of the complex $j^{-1}\calK_{\infty}^\bullet\left(1\otimes f,m\right)_{\tilde{s}_2}$ is $(j^{-1}\ker\, d^0_m)_{\tilde{s}_2}$, which is isomorphic to $\ov\calL\otimes_R R_m$ as local systems of $R_m$-modules. \end{lem} \begin{remk}[Analogue of Remark \ref{rem:nu}]\label{rem:nuQ} We can fix a canonical isomorphism $\nu_{\Q}\colon \ov\calL\otimes_R R_m\cong (\ker\, d^0_m)_{\tilde s_2}$ as follows. Recall that at any $x\in U$, $\ov\calL_x$ has an $\Q$-basis $\{\ov\delta_{x'}\}$ parametrized by $x'\in \pi^{-1}(x)$. For any such $x'$, let $(\nu_{\Q})_x(\ov \delta_{x'}) $ be the unique germ $g$ in the stalk of $(j^{-1}\ker\, d^0_m)$ at $x$ such that $ g(x) = \exp\left(-\frac{f_\infty(x')}{2\pi i}\otimes 1\otimes s\right)$. On any simply connected neighborhood $V$ of $x$, let $\iota\colon V\to U^f$ be the section of $\pi$ such that $\iota(x)=x'$. Then, $\nu_{\Q}$ can be described by \[ (\nu_\Q)_x(\ov\delta_{x'}) =\exp\left(-\frac{f_\infty\circ\iota}{2\pi i}\otimes 1\otimes s\right). \] Similarly as in Remark \ref{rem:nu}, we can check that $\nu_{\Q}$ is an $R_m$-linear map on stalks (after twisting the $R_m$ module structure by $\widetilde{s_2}$ in the target) and that it defines a morphism of local systems of $R_m$-modules on $U$.$\sslash$ \end{remk} \begin{prop}[Analogue of Proposition \ref{propLocal}] \begin{align*} j^{-1}\calK^\bullet_\infty\left(1\otimes f,\infty\right)_{\tilde{s}_2}\coloneqq \displaystyle\varprojlim_m \left(j^{-1}\calK^\bullet_\infty\left(1\otimes f,m\right)_{\tilde{s}_2}\right) \end{align*} is a resolution of $\ov\calL \otimes_R R_\infty$, via a canonical map $\ov\calL \otimes_R R_\infty\to j^{-1}\calK^\bullet_\infty\left(1\otimes f,\infty\right)_{\tilde{s}_2}$ described in Remark~\ref{rem:nuQ}. \label{propLocalQ} \end{prop} \begin{remk} As stated in \cite[Corollary 4.17]{peters2008mixed}, the adjunction map $\calK_\infty^\bullet\rightarrow Rj_*j^{-1}\calK_\infty^\bullet$ is a quasi-isomorphism. Hence, using Remark \ref{sheafgeneralities} and Proposition \ref{propLocalQ}, for all $1 \leq m \leq \infty$ we get natural isomorphisms $$ \H^*\big(X,\calK_\infty^\bullet(1\otimes f,m)_{\tilde{s}_2}\big)\cong \H^*\big(U,j^{-1}\calK_\infty^\bullet(1\otimes f,m)_{\tilde{s}_2}\big)\cong H^*(U,\ov\calL\otimes_R R_m). $$ Furthermore for large integers $m$ and $j$, the submodule $\Tors_{R_\infty} H^*(U,\ov\calL\otimes_R R_\infty)$ is shown to be contained in $H^*(U, \ov\calL \otimes_R R_m)$ as the kernel of $\psi_{mj}^*$ applying Lemma \ref{lem:torsion2} to the cochain complexes of these local systems.$\sslash$ \label{remQ} \end{remk} \subsection{Thickened Complex of the Log {de Rham} Complex} \label{sthick} It is easy to see that the form $df/f$ is an element of the intersection $\Gamma(U, \Omega^1_U) \cap \ker d $. In fact, we have a stronger result: \begin{lem}\label{logarithmicform} We have \[ df/f\in W_1\Gamma\big(X, \Omega^1_X(\log D)\big) \cap F^1\Gamma\big(X, \Omega^1_X(\log D)\big) \cap \ker d. \] \end{lem} \begin{proof} We will check the stalk condition described in \cite[Section 4.1]{peters2008mixed}. Let $x \in X$ be given. Select an open $V$ admitting local holomorphic coordinates $(z_1, \dots, z_n)$ at $x$ in which $D$ is given by $z_1\dots z_\ell = 0$. By shrinking $V$ if necessary and taking the appropriate affine chart of $\C P^1$, we may assume that $\bar{f}\colon X \rightarrow \C P^1$ determines by restriction a holomorphic map $\bar{f}|_V\colon V \rightarrow \C$, where $\left(\bar{f}|_V\right)^{-1}(0)\subset D\cap V$. In particular, we can write $\bar{f}|_V = z_1^{a_1} \cdots z_\ell^{a_\ell} \cdot g$ for some $a_1, \dots, a_\ell \geq 0$ and invertible $g\colon V \rightarrow \C^*$. Application of the Leibniz rule to the following: \begin{align*} \frac{df}{f} = \frac{d(z_1^{a_1} \cdots z_\ell^{a_\ell} \cdot g)}{z_1^{a_1} \cdots z_\ell^{a_\ell} \cdot g}, \end{align*} where we have implicitly further restricted to $U \cap V$, shows that $df/f$ is of the desired form on $V$. \end{proof} Our goal is to conduct a thickening of $\Hdg{X}{D}$ so that the hypotheses of Section \ref{denotationsandassumptionsthickenedcomplexesandmixedhodgecomplexes} are satisfied. Recall that our pseudo-morphism $\alpha$ is given by the chain: $$ \begin{tikzcd} (j_*\calE^\bullet_U, \tau_{\lc}) \ar[r, "j_*(\id \otimes 1)"] & (j_*(\calE^\bullet_U \otimes \C), \tau_{\lc}) & \ar[l, "\simeq"] (\logdr{X}{D}, \tau_{\lc}) \arrow{r}{\id}[swap]{\simeq} & (\logdr{X}{D}, W_{\lc}) \end{tikzcd} $$ where the central map is the inclusion. As required by Section \ref{denotationsandassumptionsthickenedcomplexesandmixedhodgecomplexes}, we begin by selecting closed global $1$-forms: $\Im (df/f)$ of $\tau_1j_*\calE^\bullet_U$ (where $\Im$ denotes the imaginary part) and $(1/i)df/f$ of $W_1\logdr{X}{D} \cap F^1\logdr{X}{D}$, allowable by Lemma \ref{logarithmicform}. We must verify that these two global 1-forms are cohomologous in $\tau_1$: \begin{lem}\label{imaginarycohomologous} The forms $\Im(df/f)$ and $(1/i)df/f$ are cohomologous in $\tau_1\Gamma(U, \calE^\bullet_U \otimes \C)$ via: \[\Im \frac{df}{f}-\frac{1}{i}\frac{df}{f} = d\left(\frac{\Log(|f|)}{-i}\right)\] where $\Log\colon \R_{>0} \rightarrow \R$ is the real logarithm. \end{lem} \begin{proof} Since $\tau_1\Gamma(U, \calE^0_U \otimes \C) = \Gamma(U, \calE^0_U \otimes \C)$ it follows that $\frac{\Log(|f|)}{-i} \in \tau_1\Gamma(U, \calE^\bullet_U \otimes \C)$. It only remains to verify the presented equality. Note that: \[\Im \frac{df}{f}-\frac{1}{i}\frac{df}{f} = -\frac{1}{i}\Re \frac{df}{f}\] where $\Re$ denotes the real part. On the other hand, for any branch $\log$ of the complex logarithm, we have: \[d\left(\frac{\Log(|f|)}{-i}\right) =-\frac{1}{i}d(\Re \log (f)) = -\frac{1}{i}\Re \frac{df}{f}, \] as desired. \end{proof} Accounting for Remark \ref{multiplicative}, the assumptions of Section \ref{denotationsandassumptionsthickenedcomplexesandmixedhodgecomplexes} are satisfied (as required by assumption (3) we fix the choice $\Log(|f|)/(-i)$ as witness to $\Im(df/f)$ and $(1/i) df/f$ being cohomologous), and we conclude: \begin{thm}\label{logthickenedmhs} Consider the $\R$-mixed Hodge complex of sheaves: \begin{align*} \calH dg^\bullet(X\,\log D) = \left((j_*\calE^\bullet_U, \tau_{\lc}), (\logdr{X}{D}, W_{\lc}, F^{\lc}), \alpha \right) \end{align*} described in Theorem \ref{logmhs}. Suppose $m \geq 1$. Then we have a thickened triple as in Theorem \ref{mhsthickened}: \begin{align*} \calH dg^\bullet&(X\,\log D)\left(\frac{1}{i}\frac{df}{f},m\right)\\ &\coloneqq \left(\left[j_*\calE^\bullet_U\left(\Im\, \frac{df}{f},m\right), \tau_{\lc}\right], \left[\logdr{X}{D}\left(\frac{1}{i}\frac{df}{f},m\right), W_{\lc}, F^{\lc} \right], \alpha_{\#}\right) \end{align*} which is an $\R$-mixed Hodge complex of sheaves on $X$. \end{thm} In this paper, we focus on the torsion part of the Alexander modules, but the next Corollary shows that we have a MHS that involves (a quotient of) the free part as well. \begin{cor}\label{cor:alex/s^mMHS} Let $m,j\in \mathbb N$. Recall that we define $R_m= R/(s^m)$, where $s=t-1$. The $\R$-vector spaces $M_m \coloneqq H^*(U; \ov\calL) \otimes_R R_m$ admit natural $\R$-mixed Hodge structures for which the following maps are morphisms of mixed Hodge structures: \begin{enumerate} \item The projection $M_{m+j} \twoheadrightarrow M_m$. \item The map induced by multiplication by $(\log(t))^{j}\colon R_m\hookrightarrow R_{m+j}$ after tensoring, that is $(\log(t))^{j}\colon M_{m} \to M_{m+j}(-j)$. \item Multiplication by $\log(t)$, $M_m\to M_m(-1)$. \end{enumerate} Here $\log(t)$ represents the Taylor series centered at $0$, and $(-j)$, $(-1)$ denote Tate twists. \end{cor} \begin{proof} By the flatness of $R_\infty$ over $R$, we have \[ M_m =H^*(U;\ov\calL)\otimes_R R_m \cong H^*(U;\ov\calL)\otimes_{R} R_\infty \otimes_{R_\infty} R_m \cong H^*(U;\ov\calL\otimes_{R} R_\infty) \otimes_{R_\infty} R_m . \] For $m \geq 1$, it turns out that the translated mixed Hodge complex of sheaves \[\calH dg^\bullet(X\,\log D)\left(\frac{1}{i}\frac{df}{f},m\right)[1]\] is better suited for comparison to known mixed Hodge structures, so we use it to endow $M_m$ with a mixed Hodge structure. This will enable us to state Theorems~\ref{geoIntro} and \ref{comp} without involving Tate twists. This translated $\R$-mixed Hodge complex of sheaves endows, by Remark \ref{translatedlemmas}, $\R$-mixed Hodge structures on the hypercohomology: \[ \mathbb{H}^*(U, \calE^\bullet_U(\Im df/f, m)[1]) \] which we fix for the rest of the proof. We will also use Lemma~\ref{inducedmhsmaps} for the maps between these spaces. Let $\tilde s= \exp(2\pi s)-1\in R_\infty$. Then $\calE_U^\bullet\left(\Im\, df/f,m\right)_{\tilde s}$ is a complex of soft sheaves for all $m$, since it is a complex of $\calE_U^0$-modules (\cite[Proposition 2.1.8]{dimca2004sheaves}). Hence, $\calE_U^\bullet\left(\Im\, df/f,m\right)_{\tilde s}$ is $\Gamma$-acyclic and therefore, Proposition \ref{propLocal} provides isomorphisms of $R_\infty$-modules $H^*\left(U; \ov\calL\otimes_R R_\infty \right) \cong H^*\Gamma\left(U, \calE^\bullet_U(\Im\, df/f, \infty)\right)_{\tilde s} \cong \mathbb{H}^{*-1}(U, \calE^\bullet_U(\Im df/f, \infty)[1])_{\tilde{s}}$. Since every complex algebraic variety has the homotopy type of a finite CW complex \cite[p. 27]{dimca1992hypersurfaces}, the above isomorphism yields that $H^*\Gamma\left(U, \calE^\bullet_U(\Im\, df/f, \infty)\right)$ is a finitely generated $R_\infty$-module. In particular, we can apply Corollary~\ref{torsion}: $M_m$ is isomorphic to the image of $\phi_{m'm}[1]^*$ for $m'\gg 0$. We will use that all the maps in Lemma~\ref{inducedmhsmaps} are MHS morphisms (rather, their translations as in Remark~\ref{translatedlemmas}). Being the image of a MHS morphism, $M_m$ is a sub-MHS of $\mathbb{H}^{*-1}(U, \calE^\bullet_U(\Im df/f, m)[1])_{\tilde{s}}$. The fact that this MHS is independent of $m'$ and $m$ is a consequence of Remark~\ref{torsionrelations}. Finally, we show that the maps in the statement are MHS morphisms. Note that multiplication by $\frac{\log(t)}{2\pi}$ in $\mathbb{H}^{*-1}\big(U, \calE^\bullet_U(\Im df/f, m)[1]\big)_{\tilde s}$ corresponds to multiplication by $s$ in $\mathbb{H}^{*-1}\big(U, \calE^\bullet_U(\Im df/f, m)[1]\big)$. Using Remark~\ref{torsionrelations}: the map (1) is induced by $\phi_{jm}[1]^*$, the map (2) is induced by $(2\pi)^j \psi_{mj}[1]^*$ and (3) is multiplication by $2\pi s$ (taking into account the different module structures). Since all three maps induce MHS morphisms in cohomology by Lemma~\ref{inducedmhsmaps} and Remark~\ref{translatedlemmas}, we are done. \end{proof} \begin{cor}\label{torsionmhs} Suppose that the action of $t$ on $\Tors_{R}H^*(U; \ov\calL)$ is unipotent. The $\R$-vector spaces $\Tors_{R}H^*(U; \ov\calL)$ admit natural $\R$-mixed Hodge structures for which multiplication by $\log(t)$ determines a morphism of mixed Hodge structures into the $-1$st Tate twist. \end{cor} \begin{proof} Since every complex algebraic variety has the homotopy type of a finite CW complex \cite[p. 27]{dimca1992hypersurfaces}, the modules $H^*(U; \ov\calL)$ are finitely generated over $R$. In particular, there is an $m\ge 0$ such that $s^m$ annihilates $\Tors_{R}H^*(U; \ov\calL)$. For such an $m$, $\Tors_{R}H^*(U; \ov\calL)$ is canonically isomorphic to the image of the following map, since $s$ is nilpotent by hypothesis: \begin{equation} \label{eqn:imageTorsion} \Tors_{s^m} \left(H^*(U; \ov\calL) \otimes_{R} R_{2m}\right) \hookrightarrow H^*(U; \ov\calL) \otimes_{R} R_{2m} \twoheadrightarrow H^*(U; \ov\calL) \otimes_{R} R_{m}. \end{equation} Since $s$ and $\log(t)$ differ by a unit, $\Tors_{s^m} H^*(U; \ov\calL)$ is the kernel of the multiplication by $(\log(t))^m$, which is a MHS morphism by Corollary~\ref{cor:alex/s^mMHS}, part (3). Hence, the inclusion in equation (\ref{eqn:imageTorsion}) is a MHS morphism. The second map in equation (\ref{eqn:imageTorsion}) is also a MHS morphism by Corollary~\ref{cor:alex/s^mMHS}, part (1). Therefore, the canonical isomorphism between $\Tors_{R}H^*(U; \ov\calL)$ and the image of the map in equation (\ref{eqn:imageTorsion}) endows $\Tors_{R}H^*(U; \ov\calL)$ with a MHS such that multiplication by $\log(t)$ is a morphism of MHS, as it is the restriction of a MHS morphism. By Corollary~\ref{cor:alex/s^mMHS}, part (1), this MHS is independent of $m$, provided that $s^m$ annihilates $\Tors_{R}H^*(U; \ov\calL)$. \end{proof} \begin{cor}\label{alexandermhs} The $\R$-vector spaces $\mathrm{Tors}_R\, H^*(U;\ov\calL)$ admit canonical $\R$-mixed Hodge structures for which multiplication by $\log (t^N)$ is a morphism of mixed Hodge structures into the $-1$st Tate twist. Here $\log$ is the Taylor series centered at $1$, and $N$ is chosen so that $t^N$ acts unipotently on $\mathrm{Tors}_R\, H^*(U;\ov\calL)$. \end{cor} \begin{proof} If $1$ is not the only eigenvalue of the action of $t$ on $H^*(U;\ov\calL\otimes \C)$, pick $N$ such that $\lambda^N=1$ for all $\lambda$ eigenvalue of the action of $t$ on $H^*(U;\ov\calL\otimes\C)$. We can reduce this case to the one where $1$ is the only eigenvalue by Lemma \ref{lemLocal} and Remark \ref{remEigenvalue}, obtaining a $k$-linear isomorphism $\mathrm{Tors}_R\, H^*(U;\ov\calL) \cong \Tors_{R(N)} H^*(U_N;\ov\calL_N)$. We give the left hand side the MHS of the right hand side, which we constructed in Corollary \ref{torsionmhs}. As we will see in Theorem \ref{indN} below, this construction is independent of the choice of suitable $N$ \end{proof} \begin{thm}[Independence of the compactification] Suppose that the action of $t$ on $\Tors_{R}H^*(U;\ov\calL)$ is unipotent. The mixed Hodge structure on $\Tors_R H^*(U;\ov\calL)$ obtained in Corollary \ref{torsionmhs} is independent of the compactification $X$ of $U$ such that $D=X\setminus U$ is a simple normal crossing divisor, which we used to construct the mixed Hodge structure. \label{indcompactification} \end{thm} \begin{proof} Suppose that $X$ and $Y$ are two such compactifications, called \emph{good} compactifications. Let $Z$ be a resolution of the closure of the diagonal $\Delta$ of $U\times U$ inside of $X\times Y$ such that $Z$ is a good compactification of $U$ as well. The two projections $p_X\colon Z\rightarrow X$ and $p_Y\colon Z\rightarrow Y$ induce the identity on $U$. It suffices to show that the mixed Hodge structure on $\Tors_R H^*(U;\ov\calL)$ obtained using $Z$ is the same as the one obtained using $X$. Let $E=Z\setminus U$. By \cite[Lemma 4.12]{peters2008mixed} (recalling Remark \ref{directim}), we have a canonical morphism $$ \calH dg^\bullet (X\log D)\rightarrow R(p_X)_*\calH dg^\bullet (Z\log E). $$ By construction, this morphism factors through $(p_X)_*\calH dg^\bullet (Z\log E)$ as maps of triples. Since $$(p_X)_*\left(\calH dg^\bullet (Z\log E)\left(\frac{1}{i}\frac{df}{f},m\right)\right)=\left((p_X)_*\calH dg^\bullet (Z\log E)\right)\left(\frac{1}{i}\frac{df}{f},m\right),$$Lemma \ref{inducedmap} tells us that this map of triples induces a canonical morphism between their corresponding thickenings $$ \calH dg^\bullet (X\log D)\left(\frac{1}{i}\frac{df}{f},m\right)\rightarrow (p_X)_*\left(\calH dg^\bullet (Z\log E)\left(\frac{1}{i}\frac{df}{f},m\right)\right). $$ Composing with $(p_X)_*$ of the canonical map into the Godement resolution, we get a morphism of mixed Hodge complexes of sheave $$ \calH dg^\bullet (X\log D)\left(\frac{1}{i}\frac{df}{f},m\right)\rightarrow R(p_X)_*\left(\calH dg^\bullet (Z\log E)\left(\frac{1}{i}\frac{df}{f},m\right)\right). $$ Let $j_X:U\hookrightarrow X$ and $j_Z:U\hookrightarrow Z$ be the inclusions, and note that $p_X\circ j_Z=j_X$. The $[1]$ translation of the map of mixed Hodge complexes above induces isomorphisms between the $\R$-MHS on $\H^*(U,\calE_U^\bullet(\Im df/f,m)[1])$ obtained by the compactification $X$ and the $\R$-MHS on $\H^*(U,\calE_U^\bullet(\Im df/f,m)[1])$ obtained by the compactification $Z$. Hence, the MHS obtained on $\Tors_{R}H^*(U;\ov\calL)$ using these MHS on $\H^{*-1}(U,\calE_U^\bullet(\Im df/f,m)[1])$ for large enough $m$ is independent of the compactification. \end{proof} \begin{thm}[Independence of $N$] \label{indN} The mixed Hodge structure on $\Tors_R H^*(U,\ov\calL)$ obtained in the proof of Corollary \ref{alexandermhs} is independent of the choice of cover $U_N$ used to construct it. \end{thm} \begin{proof} It suffices to show that if $1$ is the only eigenvalue of the action of $t$ on $H^*(U;\ov\calL\otimes \C)$, then the mixed Hodge structure defined on $\Tors_R H^*(U;\ov\calL)$ is the same as the mixed Hodge structure obtained on $\Tors_{R(N)} H^*(U_N;\ov\calL_N)$, where $N\geq 1$ and $p:U_N\rightarrow U$ is a degree $N$ cover as in Section \ref{sscover}. We use the same notation from the discussion preceding Lemma \ref{lemLocal}. In particular, we defined an isomorphism ${\mathcal L}\cong p_*{\mathcal L}_N$ of sheaves of $R(N)$-modules, which we will refer to as ``the canonical isomorphism'' throughout the proof. By Lemma~\ref{233}, this is the obvious isomorphism appearing in this situation, since it induces in cohomology the map coming from the isomorphism of infinite cyclic covers. As explained in \cite[Section 4.5.1]{peters2008mixed}, it is possible to find smooth compactifications $X$ of $U$ and $Y$ of $U_N$ such that $D\coloneqq X\backslash U$ and $E\coloneqq Y\backslash U_N$ are simple normal crossing divisors, and such that $p:U_N\rightarrow U$ extends to $p: Y\rightarrow X$, and $p^{-1}(D)=E$ by construction. By Remark~\ref{directim}, applying $Rp_*$ to $\calH dg^\bullet (Y\log E)\left(\frac{1}{i}\frac{df_N}{f_N},m\right)$ yields a mixed Hodge complex of sheaves on $X$. We need to define a map of $\R$-mixed Hodge complexes of sheaves as below. In Corollary~\ref{cor:alex/s^mMHS} we used a translation of these complexes to define the mixed Hodge structures, which according to Remark~\ref{transvstate} will result in a Tate twist in their cohomology groups. We will omit the translation in this proof for brevity. \[ \wt p:\Hdg{X}{D}\left(\frac{1}{i}\frac{df}{f},m\right) \to Rp_*\left( \calH dg^\bullet (Y\log E)\left(\frac{1}{i}\frac{df_N}{f_N},m\right) \right).\] Note that the twisted de Rham complexes on the left use the ring $R$, and the twisted de Rham complexes on the right use the ring $R(N)=k[t^{\pm N}]$. We recall the notation $R(N)_m=R(N)/((s_N)^m)\subset R$, with $s_N=t^N-1$. We define $\wt p$ as a composition: \[ \wt p:\Hdg{X}{D}\left(\frac{1}{i}\frac{df}{f},m\right) \xrightarrow{\wh p} p_*\left( \calH dg^\bullet (Y\log E)\left(\frac{1}{i}\frac{df_N}{f_N},m\right) \right) \to\]\[\to Rp_*\left( \calH dg^\bullet (Y\log E)\left(\frac{1}{i}\frac{df_N}{f_N},m\right) \right).\] The second arrow is the natural transformation $p_*\to Rp_*$ (it comes from the map including a complex of sheaves into its Godement resolution and then applying $p_*$). For our purposes, we do not need the object in the middle to be a mixed Hodge complex, only the first and last. For $\omega\in {\mathcal E}^*_U $ defined on any open subset of $U$, we let: \[ \wh p_{\R} ( \omega\otimes s^k) = p^* \omega \otimes \frac{(s_N)^k}{N^k}. \] where $p^* \omega \otimes \frac{(s_N)^k}{N^k}\in{\mathcal E}^*_U\left(\Im\frac{df_N}{f_N},m\right)$. This commutes with the differential: \begin{align*} d_{m}(\wh p_{\R}(\omega\otimes s^k) ) &= d_{m}\left( p^* \omega \otimes \frac{(s_N)^k}{N^k}\right) = dp^* \omega \otimes \frac{(s_N)^k}{N^k} + p^*\omega \wedge \Im\frac{df}{f} \otimes \frac{(s_N)^{k+1}}{N^k};\\ \wh p_{\R} \circ d_{m}(\omega\otimes s^k) &= \wh p_{\R} \left( d\omega\otimes s^k + \omega \wedge \Im\frac{df}{f}\otimes s^{k+1} \right) \overset{p^*\frac{df}{f}=N\frac{df_N}{f_N}}{=}\\ &= dp^*\omega\otimes \frac{(s_N)^k}{N^k} + p^*\omega \wedge \Im\frac{df_N}{f_N}N\otimes \frac{(s_N)^{k+1}}{N^{k+1}}. \end{align*} And similarly we can define $\wh p_{\C}$ for the complex part. Using Remark~\ref{directim}, the filtrations on $\calH dg^\bullet (Y\log E)\left(\frac{1}{i}\frac{df_N}{f_N},m\right)$ induce filtrations on both its underived and derived pushforwards, and the map between them preserves these filtrations. Then $\wt p$ preserves the filtrations, given that $\wh p$ does, which is straightforward to verify. The morphism of pseudomorphisms is the natural one, constructed using the pullback of forms similarly to the definition of $\wt p_{\R}$. Hence, $\wt p$ is a morphism of mixed Hodge complexes of sheaves. Taking cohomology and using Remark~\ref{rem:nu}, we obtain a morphism of mixed Hodge structures $H^i(U;\ov{\mathcal L}\otimes_R R_m)\to H^i(U_N;\ov{\mathcal L}_N\otimes_{R(N)} R(N)_m)$. We want to show that this morphism is related to the canonical isomorphism $\ov{\mathcal L}\cong p_*\ov{\mathcal L}_N$ from Lemma~\ref{233}. Since ${\mathcal E}_{U_N}^\bullet(m,\Im df_N/f_N)$ is a complex of soft sheaves, the map $p_*{\mathcal E}_{U_N}^\bullet(m,\Im df_N/f_N)\rightarrow Rp_*{\mathcal E}_{U_N}^\bullet(m,\Im df_N/f_N)$ is a quasi-isomorphism. Since $\wt p$ is a morphism of mixed Hodge complexes, this means that $\widehat p_\R$ induces the same morphism of MHS $H^i(U;\ov{\mathcal L}\otimes_R R_m)\to H^i(U_N;\ov{\mathcal L}_N\otimes_{R(N)} R(N)_m)$ as $\wt p_\R$. We work with $\widehat p_\R$ from now on. The map $\widehat p_\R$ is split injective. A left inverse is $\widehat p_\R':p_*{\mathcal E}_{U_N}^\bullet(m,\Im df_N/f_N)\to {\mathcal E}_U^\bullet(m,\Im df/f)$ defined on open sets as follows: on an open set $V$, \[ \Gamma(V;p_*{\mathcal E}_{U_N}^\bullet(m,\Im df_N/f_N)) = \Gamma(p^{-1}(V);{\mathcal E}_{U_N}^\bullet(m,\Im df_N/f_N)). \] Considering $\alpha\otimes (s_N)^{a}\in {\mathcal E}_{U_N}^\bullet(m,\Im df_N/f_N)$ defined over $p^{-1}(V)$, we let $t$ act as the generator of the deck group of $p:U_N\to U$, and define \[ \ov \alpha = \frac{1}{N}\sum_{k=0}^{N-1} (t^k)^*\alpha. \] By construction, $\ov \alpha$ is $t$-invariant. Therefore, we can define $\wh p'_{\R}\alpha$ as the unique form such that $ p^*\wh p'_{\R}(\alpha) = \ov \alpha$. We extend $\wt p'_\R$ to the thickening by letting $\wt p'_\R (\alpha\otimes (s_N)^a) = (\wt p'_\R \alpha)\otimes N^as^a$. Direct computation shows that $\wh p'_\R \circ \wh p_\R=\Id$. We consider the following commutative diagram, in which the horizontal arrows compose to the identity. \begin{equation}\label{eq:indepNAttempt2} \begin{tikzcd}[column sep = 6em, row sep = 1.5em] {\mathcal E}_U^0\otimes_\R R_m\arrow[r,"\wh p_{\R}"] & p_*{\mathcal E}_{U_N}^0\otimes_{\R} R(N)_m \arrow[r,"\wh p_{\R}'"] & {\mathcal E}_U^0\otimes_\R R_m \\ \varprojlim_m{\mathcal E}_U^0\otimes_\R R_m\arrow[r,"\wh p_{\R}"]\arrow[u] & \varprojlim_m p_*{\mathcal E}_{U_N}^0\otimes_{\R} R(N)_m \arrow[r,"\wh p_{\R}'"]\arrow[u] & \varprojlim_m {\mathcal E}_U^0\otimes_\R R_m \arrow[u] \\ \ov{\mathcal L}_U \otimes_R R_\infty \arrow[r,"\sigma" ]\arrow[u,"\nu"] & p_*\ov{\mathcal L}_N \otimes_{R(N)} R(N)_\infty\arrow[u,"p_*\nu_N"]\arrow[r,"\sigma'"]& \ov{\mathcal L}_U \otimes_R R_\infty \arrow[u,"\nu"]. \end{tikzcd} \end{equation} The top row of vertical arrows is induced by the limit. $\nu$ and $\nu_N$ are defined as (the inverse limit of the maps) in Remark~\ref{rem:nu}. Since $\nu$ and $\nu_N$ are isomorphisms onto the kernel of the differential (by Proposition~\ref{propLocal}), $\sigma$ and $\sigma'$ are uniquely determined as the restrictions of $\wh p_\R$ and $\wh p_\R'$ in order to make the diagram commute. One can check that, switching the $R$ and $R(N)$-module structure respectively, the map $$ \wh p_{\R}:{\mathcal E}_U^\bullet(\Im df/f,m)_{\wt s}\rightarrow p_*{\mathcal E}_{U_N}^\bullet(\Im df_N/f_N,m)_{\wt{s_N}} $$ is $R(N)_\infty$-linear, where $\wt s=\exp(2\pi s)-1$ and $\wt{s_N}=\exp(2\pi s_N)-1$. Since $\nu$ and $\nu_N$ are $R_\infty$ and $R(N)_\infty$-linear respectively after these changes in the $R$ and $R(N)$-module structures on the de Rham complexes, we get that $\sigma$ is $R(N)_{\infty}$-linear as well. We will need a formula for $\sigma'$. First, recall our notation for the sections of these sheaves. Let $x\in U$. The stalk at $x$ of $\ov{\mathcal L}$ is generated over $k$ by elements of the form $\ov\delta_{(x,z)}$ for some $(x,z)\in U^f$. The stalk of $p_*\ov{\mathcal L}_N \otimes_{R(N)} R(N)_m$ at $x$ is the direct sum of the stalks of $p^{-1}(x)$, where the stalk at $(x,e^{z/N})$ is generated by sections of the form $\ov\delta_{(x,e^{z/N},z/N + 2\pi i k)}$ with $k\in \Z$. Each element $\ov\delta_{(x,e^{y},y)}$ must be interpreted as a section of $\ov{\mathcal L}_N$ around the point $(x,e^{y})\in U_N\subset U\times \C^*$, as in diagram (\ref{eq:UN}). We claim that $\sigma'$ is the map defined on stalks as the $\R$-linear map satisfying \[ \sigma'\ov\delta_{(x,e^{z/N},z/N)} = \frac{1}{N} \ov\delta_{(x,z)}. \] A straightforward computation shows that this map on stalks gives a well-defined map of local systems of $R(N)$-modules. Another straightforward computation shows that this is the correct formula for $\sigma'$, that is, we gave the formula that makes the bottom right hand square commute. Now we take hypercohomology of diagram (\ref{eq:indepNAttempt2}), and we restrict to the torsion part of the left half. We obtain the following diagram. Recall that $\sigma$ is $R(N)_{\infty}$-linear, and that $\Tors_{R_\infty} = \Tors_{R(N)_\infty}$. \[ \begin{tikzcd} H^j\Gamma(U;{\mathcal E}_U^\bullet\otimes_\R R_m) \arrow[r,"\wh p_{\R}"] & H^j\Gamma(U_N;{\mathcal E}_{U_N}^\bullet\otimes_{\R} R(N)_m ) \\ \Tors_{R_\infty} H^j(U;\ov{\mathcal L} \otimes_R R_\infty) \arrow[r,"\sigma" ,hookrightarrow ]\arrow[u,"\nu",hookrightarrow] & \Tors_{R(N)_\infty}H^j(U_N;\ov{\mathcal L}_{N} \otimes_{R(N)} R(N)_\infty)\arrow[u,"p_*\nu_N",hookrightarrow]. \end{tikzcd} \] Suppose $m$ is large enough. The top arrow is a MHS morphism because it comes from a mixed Hodge complex morphism. The vertical arrows are MHS morphisms by definition of the MHS morphism on the domains (in Corollary~\ref{alexandermhs}). Also, they are injective. Therefore, $\sigma$ is a MHS morphism and it is injective, since it comes from a split injective morphism of sheaves. Also, note that by flatness and the isomorphism $p_*\ov{\mathcal L}_N\cong \ov{\mathcal L}$: \[ \Tors_{R_\infty} H^j(U;\ov{\mathcal L} \otimes_R R_\infty) \cong \Tors_{R} H^j(U;\ov{\mathcal L} ) \cong \Tors_{R(N)} H^j(U;\ov{\mathcal L} ) \cong\]\[\cong \Tors_{R(N)} H^j(U_N;\ov{\mathcal L}_N ) \cong \Tors_{R(N)_\infty} H^j(U_N;\ov{\mathcal L}_N \otimes_{R(N)} R(N)_\infty) . \] In particular, these two spaces have the same dimension. So $\sigma$, which is an injection, must be an isomorphism of MHS. Since $\sigma'$ is its right inverse, it is also an isomorphism of MHS. Now we just need to observe that the following diagram clearly commutes (up to multiplication by $N$), where the top horizontal arrow is the canonical map as we've defined it in Lemma~\ref{233}: \[ \begin{tikzcd} p_*\ov{\mathcal L}_N \arrow[r,"\sim"]\arrow[d] & \ov{\mathcal L}\arrow[d] \\ p_*\ov{\mathcal L}_N\otimes_{R(N)} R(N)_\infty \arrow[r,"\sigma'"] & \ov{\mathcal L}\otimes_R R_\infty. \end{tikzcd} \] This shows that the canonical isomorphism $p_*\ov{\mathcal L}_N\cong \ov{\mathcal L}$ coincides with $\sigma'$ up to multiplication by an integer constant, so the canonical isomorphism induces a MHS isomorphism: $$ \Tors_R H^*(U;\ov\calL)\cong \Tors_{R(N)} H^*(U_N;\ov\calL_N). $$ This concludes the proof. \end{proof} \begin{thm}[Functoriality of the mixed Hodge structure.]\label{functorial} Let $U_1$ and $U_2$ be smooth connected complex algebraic varieties, with algebraic maps $f_i:U_i\rightarrow \C^*$ such that $f_i$ induces an epimorphism in fundamental groups for $i=1,2$, and assume that there exists an algebraic map $g:U_1\rightarrow U_2$ that makes the following diagram commutative. \begin{center} \begin{tikzcd}[row sep = 1.4em] U_1\arrow[rd,"f_1"']\arrow[rr, "g"] & \ & U_2.\arrow[ld,"f_2"] \\ \ & \C^* & \ \end{tikzcd} \end{center} Let $\ov\calL_i$ be the local system of $R$-modules induced by $f_i$ for $i=1,2$, where $k=\R$. Then, the natural map $\Tors_R H^*(U_2,\ov\calL_2)\rightarrow \Tors_R H^*(U_1,\ov\calL_1)$ induced by $g$ is a morphism of mixed Hodge structures. \end{thm} \begin{proof} Note that $\ov\calL_1=g^{*}\ov\calL_2$. Let $N\in\N$ such that the action of $t^N$ on $\Tors_R H^*(U_i;\ov\calL_i)$ is unipotent, for $i=1,2$. The map $g$ lifts to a map $g_N:(U_1)_N\rightarrow (U_2)_N$ such that $(\ov\calL_1)_N=g_N^{*}(\ov\calL_2)_N$. As explained in \cite[Section 4.5.1]{peters2008mixed}, it is possible to find smooth compactifications $X_i$ of $(U_i)_N$ such that $D_i\coloneqq X_i\backslash (U_i)_N$ is a simple normal crossing divisor for $i=1,2$ and such that $g_N$ extends to $g_N: X_1\rightarrow X_2$, and $(g_N)^{-1}(D_2)=D_1$ by construction. By \cite[Lemma 4.12]{peters2008mixed}, there is a canonical morphism of mixed Hodge complexes of sheaves $$ (g_N)^*:\Hdg{X_2}{D_2}\rightarrow R{g_N}_*\Hdg{X_1}{D_1}. $$ By construction, this morphism factors trough $(g_N)_*\Hdg{X_1}{D_1}$ as maps of triples. Since \small $$(g_N)_*\left(\Hdg{X_1}{D_1}\left(\frac{1}{i}\frac{d(f_1)_N}{(f_1)_N},m\right)\right)=\left((g_N)_*\Hdg{X_1}{D_1}\right)\left(\frac{1}{i}\frac{d(f_1)_N}{(f_1)_N},m\right),$$\normalsize and using that $(g_N)^* (f_2)_N=(f_1)_N$, Lemma \ref{inducedmap} tells us that this map of triples induces a canonical morphism between their corresponding thickenings $$ \Hdg{X_2}{D_2}\left(\frac{1}{i}\frac{d(f_2)_N}{(f_2)_N},m\right)_{\tilde s_N}\rightarrow {g_N}_*\left(\Hdg{X_1}{D_1}\left(\frac{1}{i}\frac{d(f_1)_N}{(f_1)_N},m\right)\right)_{\tilde s_N}. $$ Composing with $(g_N)_*$ of the canonical map into the Godement resolution (and twisting the $R(N)_m$-module structure by $\tilde s_N$), we get a morphism of mixed Hodge complexes of sheaves $$ \Hdg{X_2}{D_2}\left(\frac{1}{i}\frac{d(f_2)_N}{(f_2)_N},m\right)_{\tilde s_N}\rightarrow R{g_N}_*\left(\Hdg{X_1}{D_1}\left(\frac{1}{i}\frac{d(f_1)_N}{(f_1)_N},m\right)\right)_{\tilde s_N}. $$ Therefore, after $[1]$ translating the above morphism, we see that the canonical morphism induced by $g$: $$ \Tors_R H^*(U_2,\ov\calL_2)\rightarrow \Tors_R H^*(U_1,\ov\calL_1) $$ is a morphism of $\R$-mixed Hodge structures. \end{proof} \begin{thm}[$\Q$-MHS]\label{Qalexandermhs} The mixed Hodge structure on $\Tors_R H^i(U;\ov{\mathcal L})$ defined for $k=\R$ in Corollary~\ref{alexandermhs} comes from a (necessarily unique) mixed Hodge structure defined for $k=\Q$. \end{thm} \begin{proof} In this proof, let $k=\Q$ and therefore $R=\Q[t^{\pm 1}]$. We start by proving the theorem in the case where the action of $t$ on $H^*(U;\ov{\mathcal L})$ is unipotent. Let $\tilde{s}_2\coloneqq \exp(s)-1=\Sigma_{n=1}^{\infty}\frac{s^n}{n!}$. Similarly one obtains a $\Q$-mixed Hodge complex: recalling the notation of Section \ref{rationalmixedhodgecomplexonsmoothvarieties} for $m \geq 1$, we select the thickened triple: \begin{align*} \Hdg{X}{D}&\left(\frac{1}{2\pi i}\frac{df}{f},m\right)_{\tilde{s}_2}\\ &\coloneqq \left(\left[\calK^\bullet_\infty\left(1 \otimes f,m\right), \tilde{W}_{\lc}\right], \left[\logdr{X}{D}\left(\frac{1}{2\pi i}\frac{df}{f},m\right), W_{\lc}, F^{\lc}\right], \varphi_{\infty\#}\right)_{\tilde{s}_2} \end{align*} which also satisfies the hypotheses of Section \ref{denotationsandassumptionsthickenedcomplexesandmixedhodgecomplexes}, therefore is a $\Q$-mixed Hodge complex of sheaves on $X$. Using Remark \ref{remQ}, we see that the translated mixed Hodge complex \[\Hdg{X}{D}\left(\frac{1}{2\pi i}\frac{df}{f},m\right)_{\tilde{s}_2}[1]\] induces $\Q$-mixed Hodge structures on $H^*(U;\ov\calL\otimes_R R_m) \cong \mathbb{H}^{*-1}(X, \calK^\bullet_\infty(1 \otimes f,m)[1])_{\tilde{s}_2}$ for which multiplication by $\log t$ is a morphism of mixed Hodge structures into the $-1$st Tate twist, because it is induced by $S_m[1]$ (multiplication by $s$) on the complex level (recall Lemma \ref{inducedmhsmaps}). For large enough $m$, this $\Q$-MHS on $H^*(U;\ov\calL\otimes_R R_m)$ induces a $\Q$-MHS on $\mathrm{Tors}_{R}H^*(U;\ov\calL)$ via the map of sheaves of $R$-modules $\ov\calL\rightarrow \ov\calL\otimes_R R_m$, just like we had in the case of $\R$ coefficients (as in the proof of Corollary~\ref{torsionmhs}). Let $\ov{\mathcal L}_\R \coloneqq \ov{\mathcal L} \otimes_{\Q} \R$, seen as a local system of $\R[t^{\pm 1}]$-modules. Our goal is to see that the $\R$-MHS on $H^*(U;\ov\calL\otimes_R R_m)\otimes_{\Q} \R\cong H^*(U;\ov\calL_\R\otimes_{\R[t^{\pm 1}]}\R[t^{\pm 1}]/(s^m))$ induced by the $\Q$-MHS on $H^*(U;\ov\calL\otimes_R R_m)$ coincides with the $\R$-MHS on $H^*(U;\ov\calL_\R\otimes_{\R[t^{\pm 1}]}\R[t^{\pm 1}]/(s^m))$ obtained using the real mixed Hodge complex of sheaves $\Hdg{X}{D}\left(\frac{1}{i}\frac{df}{f},m\right)_{\tilde{s}}[1]$. Indeed, this will imply our claim about the $\Q$ and $\R$-MHS on $\Tors_R H^i(U;\ov{\mathcal L})$. Using the definitions of the filtrations on the thickened complex (Section \ref{thickenedcomplexesandfiltrations}), it is straightforward to check that the following map is an isomorphism of bi-filtered complexes \begin{equation} \f{G_m}{\left(\logdr{X}{D}\left(\frac{1}{2\pi i}\frac{df}{f},m\right), W_{\lc}, F^{\lc}\right)}{ \left(\logdr{X}{D}\left(\frac{1}{ i}\frac{df}{f},m\right), W_{\lc}, F^{\lc}\right)}{\omega\otimes s^j}{(2\pi)^j\omega\otimes s^j} \label{eqnFilter} \end{equation} for all $m\geq 1$. Recalling the definitions of the mixed Hodge complexes of sheaves involved, our claim follows if we show that the following two maps are the same in the derived category. The first map is the composition \begin{align*} (Rj_*\ov\calL\otimes_R R_m)\otimes_{\Q}\C&\xrightarrow{Rj_*\nu_{\Q}\otimes_{\Q}\C} Rj_*j^{-1}\calK^{\bullet}_{\infty}(1\otimes f,m)\otimes_{\Q} \C\xleftarrow{\cong} \calK^{\bullet}_{\infty}(1\otimes f,m)\otimes_{\Q} \C\\ &\xrightarrow{\varphi_{\infty}\otimes 1} \Omega_X^\bullet(\log D)\left(\frac{1}{2\pi i}\frac{df}{f},m\right)\xrightarrow{G_m} \logdr{X}{D}\left(\frac{1}{ i}\frac{df}{f},m\right), \end{align*} which (after a $[1]$ translation) endows the cohomology of $\ov\calL\otimes_R R_m$ with a $\Q$-MHS. $\varphi_\infty$ is defined in Section~\ref{rationalmixedhodgecomplexonsmoothvarieties} and it induces a map on thickenings via Lemma~\ref{inducedmap}; and $\nu_\Q$ is defined in Remark~\ref{rem:nuQ}. The second morphism is the composition \begin{align*} (j_*\ov\calL_\R\otimes_{\R[t^{\pm 1}]} \R[t^{\pm 1}]/(s^m))\otimes_{\R}\C& \xrightarrow{j_*\nu \otimes_{\R}\C} j_*\calE^{\bullet}_{U}\left(\Im\frac{df}{f},m\right)\otimes_{\R} \C\\ &\xrightarrow{\exp\left(\frac{\Log(|f|)}{-i}\right)\wedge} j_*\calE^{\bullet}_{U}\left(\frac{1}{ i}\frac{df}{f},m\right)\otimes_{\R} \C \xleftarrow{\cong} \logdr{X}{D}\left(\frac{1}{ i}\frac{df}{f},m\right) \end{align*} which (after a $[1]$ translation) endows the cohomology of $\ov\calL\otimes_{\R[t^{\pm 1}]} \R[t^{\pm 1}]/(s^m)$ with an $\R$-MHS. Note that the two domains are canonically identified. Instead of proving that those two maps are the same, we will prove that the (post) composition of them with the quasi-isomorphism given by inclusion $\logdr{X}{D}\left(\frac{1}{ i}\frac{df}{f},m\right)\xrightarrow{\cong}j_*\calE^{\bullet}_{U}\left(\frac{1}{ i}\frac{df}{f},m\right)\otimes_{\R} \C$ give us the same map in the derived category. Note that we are now dealing with two maps from $Rj_*(\ov\calL\otimes_R R_m)\otimes_{\Q}\C$ to $Rj_*\calE^{\bullet}_{U}\left(\frac{1}{ i}\frac{df}{f},m\right)$ (the sheaves involved are $j_*$-acyclic). Since $Rj_*$ is fully faithful, it suffices to check that these two maps of sheaves of $\C$-vector spaces are the same on stalks at points in $U$. \begin{align*} \ov\calL\otimes_R R_m\otimes_{\Q}\C&\xrightarrow{\nu_{\Q}\otimes_{\Q}\C} j^{-1}\calK^{\bullet}_{\infty}(1\otimes f,m)\otimes_{\Q} \C \xrightarrow{\varphi_{\infty}\otimes 1} j^{-1}\Omega_X^\bullet(\log D)\left(\frac{1}{2\pi i}\frac{df}{f},m\right)\\ &\xrightarrow{G_m} j^{-1}\logdr{X}{D}\left(\frac{1}{ i}\frac{df}{f},m\right)\xrightarrow{\cong} \calE^{\bullet}_{U}\left(\frac{1}{ i}\frac{df}{f},m\right)\otimes_{\R} \C , \end{align*} $$ \ov\calL\otimes_{\R[t^{\pm 1}]} \R[t^{\pm 1}]/(s^m))\otimes_{\R}\C \xrightarrow{\nu \otimes_{\R}\C} \calE^{\bullet}_{U}\left(\Im\frac{df}{f},m\right)\otimes_{\R} \C \xrightarrow{\exp\left(\frac{\Log(|f|)}{-i}\otimes s\right)\wedge} \calE^{\bullet}_{U}\left(\frac{1}{ i}\frac{df}{f},m\right)\otimes_{\R} \C $$ Now, it is a straightforward computation to check that the image of $\ov\delta_{x'}$ in $\calE^{\bullet}_{U}\left(\frac{1}{ i}\frac{df}{f},m\right)\otimes_{\R}\C$ is $\exp(-\frac{ f_{\infty}\circ \iota}{i}\otimes s)$ via both those two maps, where $x\in U$, $x'\in\pi^{-1}(x)$ and $\iota$ is a section of $\pi$ defined on a simply connected neighborhood of $x$ that takes $x$ to $x'$. Hence, we have proved the theorem in the case when $t$ acts unipotently on $H^*(U;\ov{\mathcal L})$. In the case where the action is not unipotent, let $U_N$ be as in Remark~\ref{remEigenvalue}. Then we have the isomorphism $H^i(U;\ov{\mathcal L})\otimes_{\Q} \R\cong H^i(U_N;\ov{\mathcal L}_N)\otimes_{\Q} \R$ which is defined over $\Q$. The right hand side has a MHS defined over $\Q$, and therefore the left hand side's MHS is also defined over $\Q$. \end{proof} \begin{cor} The MHS on $\Tors_R H^i(U;\ov{\mathcal L})$ defined for $k=\Q$ is independent of the compactification, independent of $N$, and functorial (as in Theorems \ref{indcompactification}, \ref{indN}, and \ref{functorial}). \end{cor} \begin{proof} This is a consequence of the faithful flatness of $\R$ over $\Q$: the $\R$-MHSs determine the MHSs over $\Q$, and all the maps involved in the statements and proofs are defined over $\Q$ as maps of vector spaces. \end{proof} Note that, in view of Proposition \ref{propcanon}, results of this section also yield the following result, which motivated our paper (Theorem~\ref{mhsexistence}): \begin{cor}\label{halexandermhs} The $\Q$-vector spaces $A_*(U^f;\Q)\coloneqq \mathrm{Tors}_R\, H_*(U^f;\Q)$ admit natural $\Q$-mixed Hodge structures, so that multiplication by $\log t^N$ is a morphism of mixed Hodge structures into the $-1$st Tate twist. Here $\log$ is the Taylor series centered at $1$, and $N$ is chosen so that $t^N$ acts unipotently on $A_*(U^f;\Q)$. \end{cor} \begin{proof} By Remark~\ref{remk:isoLocal-Uf}, we have a canonical isomorphism $H_i(U^f;\Q)\cong H_{i}(U;{\mathcal L})$. We use Proposition~\ref{propcanon} to identify $\Tors_R H_{i}(U;{\mathcal L})$ with the dual MHS of $\Tors_R {H^{i+1}(U;\ov{\mathcal L})}$. The dual of multiplication by $t$ (resp. $\log(t^N)$) is multiplication by $t$ (resp. $\log(t^N)$), so we obtain the dual MHS morphism: \[ \log(t^N)\colon A_*(U^f;\Q)(1) = \left( \Tors_R {H^{i+1}(U;\ov{\mathcal L})}(-1) \right)^{\vee_k} \longrightarrow \left( \Tors_R {H^{i+1}(U;\ov{\mathcal L})} \right)^{\vee_k} = A_*(U^f;\Q). \] It suffices to apply the Tate twist $(-1)$ to this map. \end{proof} \begin{remk}\label{remk:mhsSummary} Let us give an overview of where the MHS on $\Tors_R H^{i+1}(U;\ov{\mathcal L})$ comes from. First, we pass to a finite cover $U_N\to U$ such that the action of $t^N$ on $\Tors_R H^{i+1}(U;\ov{\mathcal L})$ is unipotent. Lemma~\ref{233} provides an isomorphism between the Alexander modules of $U_N$ and $U$, so it suffices to give the former a MHS. Let us denote $U_N=U$. Let $k=\R$. We consider the shift of the thickening of the Hodge-de Rham complex \linebreak$\calH dg^\bullet(X\,\log D)\left(\frac{1}{i}\frac{df}{f},m\right)[1]$, which is a mixed Hodge complex of sheaves by Theorem~\ref{logthickenedmhs}. Hence, its $i$th cohomology, $H^{i}\Gamma(U; {\mathcal E}_U^\bullet(\Im df/f,m)[1])$, carries an $\R$-mixed Hodge structure. For every $m$, we have a map $\nu[1]\colon \ov{\mathcal L}\otimes_R R_m[1]\to {\mathcal E}_U^0(\Im df/f,m)[1]$ as in Remark~\ref{rem:nu}. This map makes the right hand side into a soft resolution of $\ov{\mathcal L} \otimes_R R_m[1]$, providing an isomorphism $H^{*}(U;\ov{\mathcal L}\otimes_R R_m) \cong H^{*-1}\Gamma(U; {\mathcal E}_U^\bullet(\Im df/f,m)[1])$, which we use to give $H^{*}(U;\ov{\mathcal L}\otimes_R R_m)$ a MHS Following the proof of Corollary~\ref{cor:alex/s^mMHS}, the MHS on $H^*(U; \ov\calL)\otimes R_m$ is the unique one that makes the following injective (for $m\gg 0$) map into a MHS morphism: \[ H^*(U; \ov\calL)\otimes R_m \hookrightarrow H^*(U; \ov\calL\otimes_R R_m). \] Following the proof of Corollary~\ref{torsionmhs}, the MHS on $\Tors_R H^*(U; \ov\calL)$ is the unique one that makes the following injective (for $m\gg 0$) map into a MHS morphism: \[ \Tors_R H^*(U; \ov\calL)\hookrightarrow H^*(U; \ov\calL)\otimes R_m. \] Combining the two, we get that the MHS on $\Tors_R H^*(U; \ov\calL)$ is the unique one that makes the following injective (for $m\gg 0$) map into a MHS morphism, for $k=\R$ and also $k=\Q$ by Theorem~\ref{Qalexandermhs}: \[ \Tors_R H^*(U; \ov\calL) \hookrightarrow H^*(U; \ov\calL\otimes_R R_m). \] Note that the image of $\Tors_{R}H^*(U; \ov\calL)$ in $H^*(U; \ov\calL\otimes_R R_m)$ is the kernel of $\psi_{mm}^*$, by Corollary~\ref{torsion}.$\sslash$ \end{remk} \subsection{Dependence on the function} We have seen that the construction of the mixed Hodge structure is well-defined, but we have made a somewhat arbitrary choice of the map $f_\infty\colon U^f\to \C$. We will now see that changing this choice (and even changing $f$) gives an isomorphic MHS. Note that these are MHS's on the same vector space, so they can be isomorphic but unequal. \begin{prop}[Dependence on the function]\label{prop:depf} Let $c\in \C$. Consider the function $\wt f= e^{2\pi c}f$, and consider the following diagram: \[ \begin{tikzcd}[column sep = 6em, row sep = 1.3em] U^{\wt f}\arrow[rrr,"\wt f_\infty"]\arrow[ddd,"\wt \pi"'] & & & \C \arrow[ddd,"\exp"]\\ &U^{f} \arrow[r,"f_{\infty}"]\arrow[d,"\pi"]\arrow[ul,"{(x,z)\mapsto (x,z+2\pi c) }"',"\chi"] & \C \arrow[ur,"+{2\pi c}"']\arrow[d,"\exp"] & \\ &U \arrow[r,"f"] & \C^* \arrow[dr,"e^{2\pi c} "] & \\ U \arrow[rrr,"\wt f"] \arrow[ur,equals]& & & \C^*. \end{tikzcd} \] Let ${\mathcal L} = \pi_! \ul k_{U^f}, {\mathcal L}' = \wt\pi_! \ul k_{U^{\wt f}}$. $\chi$ induces an isomorphism $\chi_{{\mathcal L}}\colon {\mathcal L} = \pi_!\ul k_{U^f} \cong \wt\pi_! \chi_!\ul k_{U^f} = \wt \pi_! \ul k_{U^{\wt f}} = {\mathcal L}'$. Let $\Tors_R H^i(U;\ov{\mathcal L})$ (resp. $\Tors_R H^i(U;\ov{\mathcal L}')$) be the mixed Hodge structure obtained from Corollary~\ref{alexandermhs} using the function $f$ (resp. $\wt f$). Then $\ov\chi_{{\mathcal L}}\circ t^{\Im c}$ induces an isomorphism of MHS $\Tors_R H^i(U;\ov {\mathcal L}) \to \Tors_R H^i(U;\ov{\mathcal L}')$, where \[ t^{\Im c} = \sum_{j=0}^{m-1} \binom{\Im c/N}{j} (t^N-1)^j. \] Here $N, m$ are any natural numbers for which the action of $(t^N-1)^m$ is $0$ (in which case the above expression doesn't depend on $N,m$). \end{prop} \begin{proof} On sections, $\ov\chi_{{\mathcal L}}$ is given by $\ov\delta_{(x,z)} \mapsto \ov\delta_{(x,z+2\pi c)}$. Let us start with the case where the only eigenvalue of the action of $t$ on $\Tors_R H^i(U;\ov{\mathcal L})$ is $1$. Let $m$ be large enough so that $(t-1)^m$ annihilates $\Tors_R H^i(U;\ov{\mathcal L})$. First, ${\mathcal E}^\bullet_U(\Im\frac{df}{f} ,m) = {\mathcal E}^\bullet_U(\Im \frac{d e^{2\pi c}f}{e^{2\pi c}f} ,m)$ has a canonical structure of a mixed Hodge complex independent of $c$. Let \[ \nu_f:\ov{\mathcal L}\otimes_R R_m \to {\mathcal E}^0_U\left( \Im\frac{df}{f}, m\right);\quad \nu_{\wt f}:\ov{\mathcal L}'\otimes_R R_m \to {\mathcal E}^0_U\left( \Im\frac{df}{f}, m\right) \] be constructed as in Remark~\ref{rem:nu}. Explicitly: for $(x,z)\in U^f$ and $\iota$ a local section of $\pi$ such that $\iota(x)=(x,z)$, we have that $\chi\circ \iota$ is a section of $\wt\pi$. Note that $\wt f_{\infty} \circ \chi= f_\infty + 2\pi c$, so: \begin{align*} \nu_{\wt f} \circ \ov\chi_{{\mathcal L}} (\ov\delta_{(x,z)}) &= \nu_{\wt f} (\ov\delta_{(x,z+2\pi c)}) = \exp(-\Im \wt f_{\infty}\circ \chi\circ \iota\otimes s) = e^{-2\pi \Im c\otimes s}\exp(-\Im f_{\infty}\circ \iota\otimes s) \\ &= \nu_f(t^{-\Im c}\ov\delta_{ (x,z)}). \end{align*} In other words, $\nu_{\wt f}\circ \ov\chi_{{\mathcal L}} = \nu_f\circ t^{-\Im c}$. Recall that $\nu_f$ is not $R_m$-linear unless we change the $R_m$-module structure. In the notation from Remark~\ref{twistedmodule}, we see that $\ul{\phantom{x}}\cdot e^{2\pi \Im c\otimes s}=\ul{\phantom{x}}*_{\widetilde s} t^{\Im c}$, so, since $\widetilde s=\exp(2\pi s)-1$, multiplication by $e^{-2\pi \Im c\otimes s}$ on ${\mathcal E}^0_U(\Im\frac{df}{f},m)$ corresponds to multiplication by $t^{-\Im c}$ on $\ov{\mathcal L}\otimes_R R_m$. The maps $\nu_f$ and $\nu_{\wt f}$ are used to give mixed Hodge structures to $H^i(U;\ov{\mathcal L}\otimes_R R_m)$ and $H^i(U;\ov{\mathcal L}'\otimes_R R_m)$, respectively. This shows that the following is a MHS isomorphism: \[ \ov\chi_{{\mathcal L}}\circ t^{\Im c}\colon H^i(U;\ov{\mathcal L}\otimes_R R_m) \to H^i(U;\ov{\mathcal L}'\otimes_R R_m). \] By Remark~\ref{remk:mhsSummary}, the MHS on $\Tors_R H^i(U;\ov{\mathcal L})$ is constructed as a sub-MHS of $H^i(U;\ov{\mathcal L}\otimes_R R_m)$ by the map induced by $R\to R_m$. This means that $\ov\chi_{{\mathcal L}}\circ t^{\Im c}$ is a MHS morphism $\Tors_R H^i(U;\ov{\mathcal L}) \to \Tors_R H^i(U;\ov{\mathcal L}')$. We now turn to the case where $t$ is not unipotent. As in Remark~\ref{remEigenvalue}, let $N$ be such that $t^N$ is unipotent, which will correspond to deck transformations for the degree $N$ cyclic cover. Let $\wt{U_{ N}} = \{(x,z)\in U\times \C^*\mid \wt f(x) = z^N \}$. We can draw the maps in Diagram (\ref{eq:UN}) for both $(U,f)$ and $(U,\wt f)$, and they will be connected by the following isomorphisms: \[ \begin{tikzcd}[column sep = 6em] U^{f} \arrow[r,"\theta_N"] \arrow[d,"\chi"',"{z\mapsto z+2\pi c}"]& U_N^{f_N} \arrow[r,"\pi_N"]\arrow[d,"\chi'"',"{z\mapsto z+2\pi c/N}"] & U_N\arrow[r,"p"] \arrow[d,"\chi_N"',"{z\mapsto e^{2\pi c/N}z}"]& U \\ U^{\wt f} \arrow[r,"\wt \theta_N"] & \wt{U_{ N}}^{\wt f_N} \arrow[r,"\wt\pi_N"] & \wt{U_{ N}}\arrow[r,"\wt p"] & U \arrow[u, equals]. \end{tikzcd} \] Let ${\mathcal L}_N = (\pi_N)_! \ul k_{U_N^{f_N}}, {\mathcal L}'_N = (\wt\pi_N)_! \ul k_{\wt{U_{ N}}^{\wt f_N}}$. Let $t_N = t^N$ be the deck transformation on the cohomology of $\ov{\mathcal L}_N$. Applying the already proven result in the unipotent case to $U_N$ and $c' = c/N$, we have that $\chi_{\ov{\mathcal L}_N} \circ t_N^{-\Im c/N}$ is an isomorphism of MHS between $\Tors_{R(N)} H^i(U_N;\ov{\mathcal L}_N)$ and $\Tors_{R(N)} H^i(\wt{U_{ N}};\ov{\mathcal L}'_N)$. Now, $t_N =t^N$ and the horizontal maps $\theta_N$ give isomorphisms of MHS (Lemma~\ref{233} and Theorem~\ref{indN}), so we have the desired result. \end{proof} \begin{cor}\label{cor:semisimpleIsEasy} If the action of $t$ on $\Tors_R H^i(U;\ov{\mathcal L})$ is semisimple, then the MHS on the Alexander module is independent of $f_\infty$. \end{cor} \begin{proof} If the action of $t$ is semisimple, then in the Taylor series of $t^{\Im c}$ we may take $m=1$, so $t^{\Im c} = \Id$. Proposition~\ref{prop:depf} implies that $\ov\chi_{{\mathcal L}}$ is a MHS isomorphism for any $c\in \C$. \end{proof} \begin{remk} In particular, if the action of $t$ on $\Tors_R H^i(U;\ov{\mathcal L})$ is semisimple, then taking $c= i$ in Corollary~\ref{cor:semisimpleIsEasy}, we conclude that the MHS is preserved by deck transformations.$\sslash$ \end{remk} \begin{remk} The converse of Corollary~\ref{cor:semisimpleIsEasy} holds as well, see Proposition~\ref{prop:semisimpleIsEasyConverse}.$\sslash$ \end{remk} \section{The geometric map is a morphism of MHS}\label{sec:main} The goal of this section is to prove the following result (Theorem~\ref{geoIntro} in the introduction). We use all the notations of Section~\ref{ssAlex}. \begin{thm}\label{thm:geomIsMHS} The covering space map $\pi\colon U^f\to U$ induces a map in homology $$H_i(\pi)\colon \Tors_R H_i(U^f;k)\to H_i(U;k)$$ and the $k$-dual of $H_i(\pi)$ is a map in cohomology $$H^i(\pi)\colon H^i(U;k)\to \Tors_R H^{i+1}(U;\ov{\mathcal L}).$$ Both are morphisms of mixed Hodge structures. \end{thm} The MHS's on $\Tors_R H^{i+1}(U;\ov{\mathcal L})$ and $\Tors_R H_i(U;{\mathcal L})$ are the ones in Corollary~\ref{alexandermhs} and Corollary~\ref{halexandermhs}, respectively. $H^i(U;k)$ is endowed with Deligne's MHS, and $H_i(U;k)$ is endowed with its dual MHS. Before proving Theorem \ref{thm:geomIsMHS} in Section~\ref{ssec:proofOfGeoThm} we have to do some preliminary work. \subsection{Maps between local systems}\label{sec:maps} Let $R$ be as in Section~\ref{ssAlex}, and let $R_m = \frac{R}{s^mR}$. We work with local systems of $R$-modules on $U$. For such a local system ${\mathcal F}$, we will abbreviate ${\mathcal F}\otimes_R R_m$ to ${\mathcal F}/s^m$. We make $\ul{k}$ into a sheaf of $R$-modules by letting $s$ act as $0$. For any $R$-module $M$ (or sheaf of $R$-modules), we will denote by $\ov M$ the conjugate structure as in Remark~\ref{rem:conjugate}. Throughout the whole section, the derived functor $R\Homm_R(\cdot,\ul R)$ (or $R\Hom_R(\cdot,R)$) will be abbreviated $\cdot^\vee$. The covering map $\pi$ induces a map in homology $H_i(U^f;k)\rightarrow H_i(U;k)$. From Remark~\ref{remk:isoLocal-Uf}, there is a canonical isomorphism of $R$-modules $H_i(U;{\mathcal L})\cong H_i(U^f;k)$. As discussed in Remark~\ref{rem:cohom_not_iso}, the same is not true for cohomology in general, see Corollary~\ref{isocohom} for the hypotheses needed to have such isomorphism in cohomology. Restricting the map induced by $\pi$ in homology to the torsion, we get that $\pi$ induces a map \[ H_i(\pi)\colon \Tors_R H_i(U;\calL)\to H_i(U;k). \] Taking the dual as $k$-vector spaces, and using the isomorphism $\Res$ from Definition~\ref{dfn:Res}, we obtain a map $ H^i(U;k)\to \Ext_R^1(\Tors_R H_i(U;\calL),R)$. Finally, the Universal Coefficients Theorem (as in Remark~\ref{remk:UCT}) gives us an isomorphism $$\Ext_R^1(\Tors_R H_i(U;\calL),R)\rightarrow \Tors_R H^{i+1}(U;\ov{\mathcal L}).$$ Precomposing with the inverse of this isomorphism, we get that the covering space map $\pi$ induces a map \[ H^i(\pi)\colon H^i(U;k)\to \Tors_R H^{i+1}(U;\ov{\mathcal L}). \] We will use the notation $H_i(\pi)$ and $H^i(\pi)$ from now on to refer to these maps, as well as $H_i(\pi_N)$ and $H^i(\pi_N)$ for the analogous definition using $\pi_N$ (as in Section~\ref{sscover}) instead of $\pi$. The goal of this section is to realize both as the maps in (co)homology arising from maps of local systems (Propositions~\ref{prop:geomMapPoincare} and \ref{prop:mapsAreEqual}). Corollary~\ref{isocohom} justifies the notation $H^i(\pi)$. Indeed, in its proof, we see that under the identification we are using, and under the condition that $H_i(U^f;k)$ is a torsion $R$-module, $H^i(\pi)$ (as defined by us) is indeed the map $H^i(U;k)\rightarrow H^i(U^f;k)$ induced in cohomology by $\pi$ (the dual of the map induced in homology). \begin{prop}\label{prop:geomMapPoincare} Let $\pi_{{\mathcal L}}\colon {\mathcal L}\to \frac{{\mathcal L}}{s} \xrightarrow{\sim} \ul{k}$ be determined by $\pi_{{\mathcal L}}(\delta_{(x,z)}) = 1$ for all $(x,z)\in U^f\subset U\times \C$. Up to the identification $H_i(U^f;k)\cong H_i(U;{\mathcal L})$ (Remark~\ref{remk:isoLocal-Uf}), the two maps $H_i(U;{\mathcal L})\to H_i(U;k)$ induced in homology by $\pi$ and $\pi_{{\mathcal L}}$ coincide. \end{prop} \begin{proof} This is a straightforward computation given the isomorphism $C_\bullet(U;{\mathcal L})\cong C_\bullet(U^f;k)$ of Remark~\ref{remk:isoLocal-Uf}. From the definitions, all we need is to show that the following diagram is commutative: \[ \begin{tikzcd}[row sep = 1.3em] C_\bullet(U^f;k)\arrow[r,"\sim"]\arrow[d,"\pi"] & C_\bullet(U;{\mathcal L}) = C_\bullet(U^f;k)\otimes_R R \arrow[d,"\pi_{\mathcal L}:{\mathcal L}_x=R\to k"]\\ C_\bullet(U;k) \arrow[r,"\sim"]& C_\bullet(U;\ul{k}) = C_\bullet(U^f;k)\otimes_R k. \end{tikzcd} \] We can check directly that a chain $c\in C_i(U^f;k)$ is mapped to $\pi_*(c)\in C_\bullet(U;k)$ via both paths. \end{proof} Our next goal is to find a map (in the derived category) of local systems $\ul k[-1]\to \ov{\mathcal L}$ that will induce $H^i(\pi)$ in cohomology. Let us fix the free resolution ${\mathcal L}\xrightarrow{s} {\mathcal L}$ of $\ul{k}$ (see the maps below), so that $\pi_{{\mathcal L}}$ becomes a composition as the left hand diagram below (the bottom row sits in degree $0$): \[ \begin{tikzcd}[row sep = 1.4em] 0 \arrow[r]\arrow[d]& {\mathcal L}\arrow[r]\arrow[d,"s"] & 0\arrow[d] & & \ov{\mathcal L}\arrow[r,"1"]\arrow[d,"s"]\arrow[dr,phantom,"\pi_{{\mathcal L}}^\vee"] &\ov{\mathcal L} \arrow[d] \\ {\mathcal L} \arrow[r,"1"]\arrow[rr,"\pi_{{\mathcal L}}"',dashrightarrow, bend right =10]& {\mathcal L} \arrow[r,"\pi_{{\mathcal L}}"]& \ul{k}; & & \ov{\mathcal L} \arrow[r]& 0. \end{tikzcd} \] Let $\pi_{{\mathcal L}}^\vee$ be defined by applying $R\Homm_R(\cdot,\ul R)$ to $\pi_{{\mathcal L}}$. If we identify $\ul k$ with ${\mathcal L}\xrightarrow{s}{\mathcal L}$ as above, then it corresponds to the right hand map of complexes above (recall that we have fixed a way of identifying $\Homm_R({\mathcal L},\ul R)\cong \ov{\mathcal L}$, as in Remark~\ref{rem:oppositeDual}). \begin{remk}\label{rem:Rvee} Let $\pi_{\ov{\mathcal L}}\colon \ov{\mathcal L}\to \ul k$ be the result of conjugating $\pi_{\mathcal L}$. This induces a quasi-isomorphism $R\Homm_R(\ul k,\ul R)\cong \ul k[-1]$ by using the resolution ${\mathcal L}\xrightarrow{s}{\mathcal L}$ above: $$ R\Homm_R(\ul k,\ul R)\cong (\ov{\mathcal L}\xrightarrow{s}\ov{\mathcal L} )\cong \ul k[-1], $$ where $\ov{\mathcal L}\xrightarrow{s}\ov{\mathcal L}$ sits in degrees $0$ and $1$ and the last $\cong$ is given by the map $\pi_{\ov {\mathcal L}}$. Definition~\ref{dfn:Res} provides another such isomorphism: \[ R\Homm_R(\ul k{,}\ul R) \cong \Extt^1_R(\ul k, \ul R)[-1] \xlongrightarrow[\sim]{\Res} \Homm_k(\ul k,\ul k)[-1] \cong \ul k[-1]. \] Since $\Aut(\ul k)\cong k^*$ (both in the abelian and the derived categories), these two isomorphisms differ by mutiplication by a nonzero constant (as maps in the derived category of sheaves of $R$-modules on $U$). This constant turns out to be $1$, keeping in mind Remark~\ref{rem:annoyingSign}. For all our purposes, it will not matter which constant this is, so we will omit the computation.$\sslash$ \end{remk} We have two exact triangles, where we are using $\cdot^\vee$ to denote the functor $R\Homm_R(\cdot,\ul R)$. \begin{equation}\label{eq:picLTriangles} \begin{tikzcd}[column sep = 3.5em, row sep = 1.3em] 0 \arrow[r]\arrow[d] & 0 \arrow[r]\arrow[d]\arrow[dr,phantom,"\pi_{{\mathcal L}}"] & {\mathcal L} \arrow[r,"1"]\arrow[d,"s"]\arrow[dr,phantom,"-\pi_{\ov{\mathcal L}}^\vee{[1]}"] & {\mathcal L} \arrow[d]& 0 \arrow[r]\arrow[d] & 0 \arrow[r]\arrow[d]\arrow[dr,phantom,"\pi_{\ov{\mathcal L}}"] & \ov{\mathcal L} \arrow[r,"1"]\arrow[d,"s"]\arrow[dr,phantom,"-\pi_{{\mathcal L}}^\vee{[1]}"] & \ov{\mathcal L} \arrow[d]\\ {\mathcal L} \arrow[r,"s"] & {\mathcal L} \arrow[r,"1"] & {\mathcal L} \arrow[r] & 0; & \ov{\mathcal L} \arrow[r,"s"] & \ov{\mathcal L} \arrow[r,"1"] & \ov{\mathcal L} \arrow[r] & 0. \end{tikzcd} \end{equation} Note that the labeling on the maps is consistent with $\cdot^\vee = R\Homm(\cdot,R)$. Below, the first diagram represents $\pi_{{\mathcal L}}^\vee$, the second diagram is $\pi_{{\mathcal L}}^\vee[1]$ and the third diagram is the same map, after the isomorphism $\left(\ov{\mathcal L}\xrightarrow{s}\ov{\mathcal L}\right)\cong \left(\ov{\mathcal L}\xrightarrow{-s}\ov{\mathcal L}\right)$ that is $\Id_{\ov{\mathcal L}}$ in degree $0$ and $-\Id_{\ov{\mathcal L}}$ in degree $-1$ (which is the identity morphism in the derived category). \[ \hspace{-1.5cm} \begin{tikzcd}[column sep = 1.75em, row sep = 1.3em] \deg = 0 \arrow[r,dashrightarrow]& \ov{\mathcal L} \arrow[r,"1"]\arrow[d,"s"] & \ov{\mathcal L}\arrow[d] & & \ov{\mathcal L} \arrow[r,"1"]\arrow[d,"-s"] & \ov{\mathcal L} \arrow[d]& & \ov{\mathcal L} \arrow[r,"-1"]\arrow[d,"s"] & \ov{\mathcal L} \arrow[d]\\ & \ov{\mathcal L} \arrow[r] & 0 & \deg = 0 \arrow[r,dashrightarrow]& \ov{\mathcal L} \arrow[r] & 0 & \deg = 0 \arrow[r,dashrightarrow] & \ov{\mathcal L} \arrow[r] & 0 \end{tikzcd} \] We will identify $\ul k$ with its free resolutions, hence our naming of the maps above, with a slight abuse of notation. \begin{prop}\label{prop:mapsAreEqual} The map $H^i(U;k)\to\Tors_R H^{i+1}(U;\ov{\mathcal L})$ induced by $\pi^\vee_{{\mathcal L}}$ in cohomology coincides with $H^i(\pi)$. In other words, the following diagram commutes. The dotted arrow is the desired map computed in two different ways. \[ {\footnotesize \begin{tikzcd}[column sep = 3em, row sep = 1.3em] H^{i+1}(U;\ul{k}^\vee) \arrow[d,"H^i(\pi_{{\mathcal L}}^\vee)"] \arrow[r,"\Res","\sim"']& H^{i}(U;k)\arrow[r,"\UCT_k",leftarrow,"\sim"'] \arrow[dl,dashrightarrow] & \Hom_k( H_i(U;k),k)\arrow[d,"\Hom(H_i(\pi){,}k)"]\\ H^{i+1}(U;\ov {\mathcal L})\supseteq \Tors_R H^{i+1}(U;\ov {\mathcal L}) & \Ext^1_R(H_i(U;{\mathcal L}){,}R)\arrow[l,"\UCT_R","\sim"'] & \Hom_k(\Tors_R H_i(U;{\mathcal L}),k).\arrow[l,"\Res","\sim"',leftarrow] \end{tikzcd} } \] Here ``UCT'' stands for the maps defined in \ref{lem:UCT}, and $\Res$ (abusing notation) is the canonical isomorphism induced by the isomorphism of $R$-modules $\Res$ of Definition \ref{dfn:Res}. Here we only assume that $U$ is a topological space and $f:U\to \C^*$ is a continuous map. \end{prop} \begin{proof} Throughout this proof, all homology and cohomology groups will be on the space $U$, so we will ommit it for brevity. Also, $k$ is always viewed as the $R$-module $R/sR$, via the map sending $1$ to $1$. The naturality of tensor-hom adjunction in the definition of $\UCT_R$ in Lemma \ref{lem:UCT} tells us that the following diagram commutes: \[ \begin{tikzcd}[column sep = 5em, row sep =1.3em] \Ext^1_R(H_{i}(\ul k),R) \arrow[d,"\Ext^1_R(H_i(\pi_{{\mathcal L}}){,}R)"]\arrow[r,"\UCT_R","\sim"'] & H^{i+1}(\ul{k}^\vee)=H^{i}(\Extt^1_R(\ul{k},\underline R))=H^{i+1}(C^\bullet(\ul{k}^\vee)) \arrow[d,"H^i(\pi_{{\mathcal L}}^\vee)"]\\ \Ext^1_R(H_{i}( {\mathcal L}),R)\arrow[r,"\UCT_R","\sim"'] & \Tors_R H^{i+1}(\ov {\mathcal L}). \end{tikzcd} \] Recall Remark \ref{rem:Rvee}. For the purposes of the map $H^i(\pi_{{\mathcal L}}^\vee)$, $\ul{k}^\vee$ is seen as the bounded complex $\ov{\mathcal L}\xrightarrow{s}\ov{\mathcal L}$, sitting in degrees $0$ and $1$. Recall that, by truncation, we also identified $\ul{k}^\vee$ with $\Extt^1_R(\ul{k},\underline R)[-1]$. In view of the diagram above, this lemma is equivalent to the commutativity of the following diagram: \[ \begin{tikzcd}[column sep = 5em] H^{i+1}(\ul{k}^\vee) \arrow[r,"\Res","\sim"'] & H^{i}(\ul k) \arrow[r,"\UCT_k","\sim"'] & \Hom_k( H_i(\ul k),k)\arrow[d,"\Hom(H_i(\pi){,}k)"] \\ \Ext^1_R( H_{i}(\ul{k}),R) \arrow[u,"\UCT_R","\sim"'] \arrow[r,"\Ext^1_R(H_i(\pi_{{\mathcal L}}){,}R)"] & \Ext^1_R(H_i({\mathcal L}){,}R) & \Hom_k(\Tors_R H_i({\mathcal L}),k).\arrow[l,"\Res",leftarrow,"\sim"']\end{tikzcd} \] We claim that the following diagram of isomorphisms commutes: \begin{equation}\label{eq:resUCTresUCT} \begin{tikzcd}[row sep = 1.3em] H^{i+1}(\ul{k}^\vee) \arrow[d,"\Res"]\arrow[r,"\UCT_R",leftarrow] & \Ext^1_R( H_{i}(\ul k),R)\arrow[d,"\Res"] \\ H^{i}(\ul k) \arrow[r,"\UCT_k"]& \Hom_{k}( H_i(\ul k),k). \end{tikzcd} \end{equation} Let us write the complexes representing these (co)homology groups, like in Definition \ref{dfn:complex} and Remark \ref{remk:dfncomplex}. We will draw a commutative diagram with the corresponding maps between complexes. Let $C_\bullet(\ul{k})$ be as in Remark \ref{remk:dfncomplex}. Note that the coefficients are constant, so we can take $\wt U=U$ in the definition. Then, $C_\bullet(\ul{k}) =C_\bullet(U,R)\otimes_R k$. The proof amounts to looking at the diagram (\ref{eq:betatensorhom}). \begin{equation}\label{eq:betatensorhom} {\footnotesize \hspace{-1cm} \begin{tikzcd}[column sep = 1em, row sep = 1.4em] H^{i+1}(\ul{k}^\vee) \arrow[r,no head, dashed,"a"]\arrow[ddd,"\Res",=rightarrow] \arrow[rrr,"\UCT_R ", bend left = 10,leftarrow]& \Hom_R^\bullet(C_\bullet(U,R),({\mathcal F}^\bullet)^\vee_x)[1]\arrow[r,"\text{t-h}"] & \Hom_R^\bullet(C_\bullet({\mathcal F}^\bullet),R)[1]\arrow[d,"\beta",leftarrow] & \Ext^1_R(H_{i}(\ul k),R) \arrow[l,no head, dashed,"c"] \\ & \Hom_R^\bullet(C_\bullet(U,R),\Hom_R^\bullet({\mathcal F}^\bullet_x,K/R))\arrow[u,"\beta"] \arrow[r,"\text{t-h}"] & \Hom_{R}^\bullet(C_\bullet({\mathcal F}^\bullet),K/R) & \\ & \Hom_R^\bullet(C_\bullet(U,R),\Hom_R(k,K/R))\arrow[r,"\text{t-h}"]\arrow[u,"\rho"] \arrow[d,"\Hom_R^\bullet(C_\bullet(U{,}R){,}\res_*)"] & \Hom_{R}^\bullet(C_\bullet(\ul{k}),K/R),\arrow[u,"\rho"]\arrow[d,"\res_*"]& \Hom_R(H_i(\ul k),K/R)\arrow[l,no head, dashed,"b"]\arrow[uu,"\beta"]\arrow[d,"\res_*"] \\ H^i(\ul k) \arrow[r,no head, dashed,"a"]\arrow[rrr,"\UCT_{k}", bend right = 10]& \Hom_R^\bullet(C_\bullet(U,R),\Hom_k(k,k))\arrow[r,"\text{t-h}"] & \Hom_{k}^\bullet(C_\bullet(\ul{k}),k)& \Hom_{k}(H_i(\ul k),k) \arrow[l,no head, dashed,"b"]. \end{tikzcd} } \end{equation} The four corners and the arrows between them coincide with those of diagram (\ref{eq:resUCTresUCT}). ${\mathcal F}^\bullet := \ul R \xrightarrow{s} \ul R$ (in degrees $[-1,0]$) is a free resolution of $\ul{k}$, and $\rho\colon {\mathcal F}^\bullet\to \ul{k}$ is the corresponding quasi-isomorphism (we've abused notation by labeling $\rho$ the maps that are induced by $\rho$ after applying some functors). Note that $({\mathcal F}^{\bullet})^\vee[1]$ is a resolution of $\Extt_R^1(\ul{k},\underline R)=\underline{\Ext_R^1(k,R)}$ (recall that we are using $\ul M$ to denote the constant sheaf with stalk $M$). Similarly, $\beta \colon K/R\to R[1]$ is defined as in Definition~\ref{dfn:Res}, and we've labeled $\beta$ the arrows that are induced by $\beta$. The arrows ``t-h'' represent tensor-hom adjunction. First, note that the ``inside'' rectangle made of solid straight arrows is commutative. The top two squares commute by the naturality of tensor-hom adjunction, and the bottom square can be seen to commute by a direct calculation. \textbf{The dashed lines are not maps}, rather they represent the cohomology of the corresponding complex. We would like to show that the maps drawn ``outside'' in cohomology are induced by the maps of complexes depicted in the diagram. The lines labeled ``$a$" are obtained by simply taking the cohomology, by definition (together with identifying $\Hom_{k}(k,k) = k$). The lines labeled ``$b$" and ``$c$" arise in the universal coefficient theorem. The $b$'s come from the obvious map $H^i(\Hom_A^\bullet(B^\bullet,C))\to \Hom_A(H^i(B^\bullet),C)$ for any ring $A$, any $A$-module $C$ and any complex $B^\bullet$ of $A$-modules. The line ``$c$" is obtained from the universal coefficient theorem. We claim that the following diagram commutes: \[ \begin{tikzcd}[row sep = 0.8em] H^{i+1}(\Hom_R^\bullet(C_\bullet({\mathcal F}^\bullet),R) & \Ext^1_R(H_i(C_\bullet({\mathcal F}^\bullet)),R) \arrow[l,"c"]\\ H^{i}(\Hom_R^\bullet(C_\bullet({\mathcal F}^\bullet),K/R) \arrow[u,"\beta"]\arrow[r,"b"]& \Hom_R(H_i(C_\bullet({\mathcal F}^\bullet)),K/R) \arrow[u,"\beta"] \end{tikzcd} \] Note that $\rho$ induces the identity in cohomology, so we are ignoring its role above. The fact that this diagram commutes follows from following the proof of Lemma~\ref{lem:UCT}. This, together with the definitions, shows that the maps ``inside'' the diagram (\ref{eq:betatensorhom}) induce in cohomology the maps ``outside'', which correspond to diagram (\ref{eq:resUCTresUCT}). Now, the commutativity of diagram (\ref{eq:resUCTresUCT}) follows from the commutativity of the ``inside'' diagram (\ref{eq:betatensorhom}). After having proved that the diagram (\ref{eq:resUCTresUCT}) commutes, the desired statement is equivalent to proving that the following diagram commutes: \[ \begin{tikzcd}[column sep = 5em, row sep = 1.3em] \Ext^1_R(H_{i}(\ul k),R) \arrow[r,"\Ext^1_R(H_i(\pi_{{\mathcal L}}){,}R)"]\arrow[d,"\Res","\sim"'] & \Ext^1_R(H_{i}( {\mathcal L}),R)\arrow[d,"\Res","\sim"']\\ \Hom_{k}(H_{i}(\ul k),k) \arrow[r,"\Hom_{k}(H_i(\pi){,}k)"] & \Hom_{k}(\Tors_R H_{i}({\mathcal L}),k). \end{tikzcd} \] But now using the naturality of the ``Res'' isomorphism, this is equivalent to $ H_i(\pi) = H_i(\pi_{{\mathcal L}})$, which is the content of Proposition~\ref{prop:geomMapPoincare}. \end{proof} \subsection{Proof of Theorem~\ref{thm:geomIsMHS}}\label{ssec:proofOfGeoThm} We use all the notations of Section~\ref{sec:maps}: we have the map $\pi_{{\mathcal L}}$ from Proposition~\ref{prop:geomMapPoincare}, and its dual $\pi_{{\mathcal L}}^\vee$, where we are still writing $\cdot^\vee = R\Homm_R(\cdot,\ul R)$. We will identify $\ul k^\vee\cong \ul k[-1]$ as in Remark~\ref{rem:Rvee}. For a (sheaf of) $R$-modules $M$, we may abbreviate $\frac{M}{s^m}\coloneqq M\otimes_R R_m$. \begin{remk} Since $\R$ is faithfully flat over $\Q$, it suffices to prove the claim when $k=\R$. \textit{For the rest of this section, we will only deal with real coefficients.}$\sslash$ \end{remk} \begin{remk} Let $N\in\N$ be a natural number such that the action of $t^N$ on $\Tors_R H^i(U;\calL)$ is unipotent, and consider the $N$-fold cover $U_N$ of $U$ as in Lemma~\ref{233} and Remark~\ref{remEigenvalue}. It suffices to show that the map in cohomology $H^i(\pi_N)\colon H^i(U_N;\R)\to\Tors_R H^{i+1}(U;{\mathcal L})$ is a MHS morphism, from which the result follows, since $U_N\to U$ is an algebraic map, and therefore it induces a MHS morphism in cohomology. \textit{Throughout the whole proof we will denote $U:=U_N$, $f:=f_N$, $\pi=\pi_N$ and assume that the $t$-action on $\Tors_R H_i(U;{\mathcal L})$ is unipotent for all $i$}.$\sslash$ \end{remk} \begin{proof}[Proof of Theorem~\ref{thm:geomIsMHS}] In Corollary~\ref{halexandermhs}, the MHS on $\Tors_R H_i(U^f;\R)$ is defined as the dual of the MHS defined on $\Tors_R H^{i+1}(U;\ov{\mathcal L})$, via a series of isomorphisms: \[ (\Tors_R H^{i+1}(U;\ov{\mathcal L}) )^\vee \overset{\Res}\cong \Ext^1_R(H^{i+1}(U;\ov{\mathcal L}),R) \overset{\UCT_R}\cong \Tors_R H_i(U;{\mathcal L}) \overset{\text{}}\cong \Tors_R H_i(U^f;\R). \] We consider the projection $\phi_m=(\phi_U)_m\colon \ov {\mathcal L}\to \frac{\ov {\mathcal L}}{s^m} $, which arises from $R\to R_m$ after tensoring with $\ov{\mathcal L}$. We are interested in $\phi_m\circ \pi_{{\mathcal L}}^\vee\colon \underline{\mathbb R}_U^\vee\to \frac{\ov{\mathcal L}}{s^m}$, where $\pi_{{\mathcal L}}$ is defined as in Proposition~\ref{prop:geomMapPoincare}. As in Remark~\ref{rem:Rvee}, we identify $\underline{\mathbb R}_U^\vee$ with $\underline{\mathbb R}_U[-1]$. By Proposition~\ref{prop:mapsAreEqual}, it is enough to prove that $H^i(\pi_{{\mathcal L}}^\vee)\colon H^i(U;\R)\to \Tors_R H^{i+1}(U;\ov{\mathcal L})$ is a MHS morphism. We will prove the theorem in a series of lemmas. We will state them all, then we will prove the theorem using them, and after that we will include the proofs of all the lemmas in order. \begin{lem}\label{lem:theMapA} Let $X$ be a good compactification of $U$, such that $D:=X\backslash U$ is a normal crossing divisor, as in Section~\ref{ss:MHSsAndComplexes} and \cite[Definition 4.1]{peters2008mixed}. Let $j\colon U\hookrightarrow X$ be the inclusion. There is a morphism of mixed Hodge complexes $A_{\calH dg}\colon \Hdg{X}{D}\rightarrow \linebreak\Hdg{X}{D}\left(\frac{1}{i}\frac{df}{f}, m \right)[1]$ given, in the real part, by $j_*A_\R$, where: $$ \f{A_\R}{\calE_{U}^\bullet}{\calE_{U}^{\bullet+1}\otimes R_m}{\alpha}{\left(\Im\frac{df}{f}\wedge \alpha\right)\otimes 1,} $$ in the complex part, by $$ \f{A_\C}{\Omega_{X}^\bullet(\log D)}{\Omega_{X}^{\bullet+1}(\log D)\otimes R_m}{\omega}{\left(\frac{1}{i}\frac{df}{f}\wedge \omega\right)\otimes 1.} $$ Being a morphism of mixed Hodge complexes, it induces a MHS morphism on the cohomology. \[ H^i(A)\colon H^i(U;\R)\to H^{i+1}\left(U;\frac{\ov{\mathcal L}}{s^m}\right). \] Here $A$ is the composition: \[ \underline{\mathbb R}_U\to {\mathcal E}^\bullet_U \onarrow{A_{\R}} {\mathcal E}^\bullet_U\left(\Im\frac{df}{f}, m\right)[1] \cong {\mathcal E}^\bullet_U(m)_{\log}[1] \onarrow{\eta_m^{-1}} \frac{\ov{\mathcal L}}{s^m}[1], \] where $\eta_m$ is as in Remark \ref{remk:thickenedComplex}. Note that $A$ is in principle just a morphism in the derived category of sheaves of $\R$-vector spaces. \end{lem} \begin{lem}\label{lem:proofInCircle} Let $\phi_{m,\C^*}, \pi_{{\mathcal L},\C^*}^\vee, A_{\C^*},{\mathcal L}_{\C^*}, \ov{\mathcal L}_{\C^*}$ be defined analogously to $\phi_m,\pi_{{\mathcal L}}^\vee,A,{\mathcal L},\ov{\mathcal L}$ in the particular case where $U=\C^*$ and $f=\Id$. There is a constant $c\in \R\setminus \{0\}$ such that \[ (\phi_{m,\C^*}\circ \pi_{{\mathcal L},\C^*}^\vee)[1] = c\cdot A_{\C^*}\colon \underline{\mathbb R}_{\C^*}\longrightarrow \frac{\ov{\mathcal L}_{\C^*}}{s^m}[1]. \] These are equal as maps in the derived category of sheaves of $\R$-vector spaces. In particular, $A_{\C^*}$ comes from a map in the derived category of sheaves of $R$-modules, after restriction of scalars. \end{lem} \begin{lem}\label{lem:pullbacks} Here $A_U=A$, ${\mathcal L}_U={\mathcal L}$, etc. We have two commutative diagrams, whose horizontal arrows are isomorphisms: \[ \begin{tikzcd} f^{-1}\underline{\mathbb R}_{\C^*} \arrow[d,"f^{-1}A_{\C^*}"]\arrow[r] & \underline{\mathbb R}_{U} \arrow[d,"A_U"] & & & f^{-1}\underline{\mathbb R}_{\C^*} \arrow[d,"f^{-1}(\phi_{m,\C^*}\circ \pi_{{\mathcal L},\C^*}^\vee){[1]}"']\arrow[r] & \underline{\mathbb R}_{U} \arrow[d,"(\phi_{m,U}\circ \pi_{{\mathcal L},U}^\vee){[1]}"] \\ f^{-1}\frac{\ov{\mathcal L}_{\C^*}}{s^m}[1]\arrow[r] & \frac{\ov{\mathcal L}_{U}}{s^m}[1]; & & & f^{-1}\frac{\ov{\mathcal L}_{\C^*}}{s^m}[1]\arrow[r] & \frac{\ov{\mathcal L}_{U}}{s^m}[1]. \end{tikzcd} \] The bottom arrow comes from the (dual of) the base change isomorphism $f^{-1}{\mathcal L}_{\C^*} = f^{-1}\exp_!\underline{\mathbb R}_{\C} \cong \pi_!f_\infty^{-1}\underline{\mathbb R}_{\C} \cong \pi_! \underline{\mathbb R}_{U^f} = {\mathcal L}_{U}$. The isomorphism $f^{-1}\underline{\mathbb R}_{\C^*}\to \underline{\mathbb R}_U$ sends $1$ to $1$. In particular, there is a constant $c\in \R\setminus \{0\}$ for which \[ c A_U = \phi_{m,U} \circ \pi_{{\mathcal L},U}^\vee. \] These are equal as maps in the derived category of sheaves of $\R$-vector spaces. In particular, $A_{U}$ comes from a map in the derived category of sheaves of $R$-modules, after restriction of scalars. \end{lem} The proof of all the lemmas in order can be found below. Before, let us show how to prove the theorem. By Lemma~\ref{lem:theMapA}, $A$ induces a MHS morphism $H^i(U;\R)\to H^{i+1}(U;\frac{\ov{\mathcal L}}{s^m})$, and by Lemma~\ref{lem:pullbacks}, $\phi_{m} \circ \pi_{{\mathcal L}}^\vee$ is its multiple. Therefore, $\phi_{m} \circ \pi_{{\mathcal L}}^\vee$ induces a MHS morphism in cohomology as well. Now, since the image of a torsion module must lie in the torsion, there is a unique map as follows, which we will also denote $H^{i}(\pi_{{\mathcal L}}^\vee)$: \[ \begin{tikzcd} H^{i-1}(U;\R) \arrow[r,"H^{i}(\pi_{{\mathcal L}}^\vee)"]\arrow[dr,dashrightarrow,"\exists!"] & H^i(U;\ov{\mathcal L}) \arrow[r,"H^i(\phi_m)"] & H^i(U;\frac{\ov{\mathcal L}}{s^m}) \\ & \Tors_R H^i(U;\ov{\mathcal L})\arrow[ur,hookrightarrow,"\iota"]\arrow[u,hookrightarrow] & \end{tikzcd} \] By Remark~\ref{remk:mhsSummary}, $\iota$ is a MHS morphism. Since $H^i(\pi_{{\mathcal L}}^\vee)\colon H^{i-1}(U;\R)\to \Tors_R H^i(U;\ov{\mathcal L})$ amounts to restricting the codomain of $H^i(\phi_m\circ \pi_{{\mathcal L}}^\vee)\colon H^{i-1}(U;\R)\to H^i(U;\frac{\ov{\mathcal L}}{s^m})$ to a sub-MHS, it is itself a MHS morphism. \begin{proof}[Proof of Lemma~\ref{lem:theMapA}] Let us recall the filtrations: \begin{align*} \tau_k{\mathcal E}_U^\bullet\left(\Im\frac{df}{f}, m\right)[1] &= \bigoplus_j \tau_{k+2j+2}{\mathcal E}_U^\bullet\otimes \R\langle s^j\rangle; \\ W_k\Omega_X^\bullet(\log D)\left(\frac{1}{i}\frac{df}{f}, m\right)[1] &= \bigoplus_j W_{k+2j+2}\Omega_X^\bullet(\log D)\otimes \C\langle s^j\rangle; \\ F^p\Omega_X^\bullet(\log D)\left(\frac{1}{i}\frac{df}{f}, m\right)[1] &= \bigoplus_j F^{p+j+1}\Omega_X^\bullet(\log D)\otimes \C\langle s^j\rangle . \end{align*} One can easily check that both components of $A_{\calH dg}$ respect the filtrations above. We must define a morphism of pseudomorphisms that makes the definition of $A_{\calH dg}$ on the real and complex part agree. Let us start by recalling the definitions. The pseudomorphism in the definition of $\Hdg XD$ is given in (\ref{eqmorphisms}). According to the construction in Theorem~\ref{mhsthickened}, the pseudomorphism used in Theorem~\ref{logthickenedmhs} is: \begin{align*} j_*\calE^{\bullet}_{U}\left(\Im\frac{df}{f},m\right)& \hookrightarrow j_*\calE^{\bullet}_{U}\left(\Im\frac{df}{f},m\right)\otimes_{\R} \C\\ &\xrightarrow{\exp\left(\frac{\Log(|f|)}{-i} \otimes s\right)} j_*\calE^{\bullet}_{U}\left(\frac{1}{ i}\frac{df}{f},m\right)\otimes_{\R} \C \xleftarrow{\cong} \logdr{X}{D}\left(\frac{1}{ i}\frac{df}{f},m\right). \end{align*} This is followed by the identity map of $\logdr{X}{D}\left(\frac{1}{ i}\frac{df}{f},m\right)$, where the domain has the filtration induced by $\tau_{\lc}$ and the target has the filtration induced by $W_{\lc}$. If we extend the pseudomorphism in the obvious way it will not define a morphism of pseudomorphisms in the sense of \cite[3.16]{peters2008mixed}, namely the following diagram does not commute: \begin{equation}\label{eq:nastyDiagram} \begin{tikzcd}[column sep = 5em] \calE^{\bullet}_{U}\otimes_\R \C \arrow[r,equals]\arrow[d,"\Im\frac{df}{f}\wedge \cdot "] & \calE^{\bullet}_{U}\otimes_\R \C \arrow[d,"\frac{1}{i}\frac{df}{f}\wedge \cdot "] \\ \calE^{\bullet}_{U}\left(\Im \frac{df}{f},m\right)\otimes_\R \C[1] \arrow[r,"\exp\left(\frac{\Log(|f|)}{-i} \otimes s\right)"] & \calE^{\bullet}_{U}\left(\frac{1}{i}\frac{df}{f},m\right)\otimes_\R \C[1] . \end{tikzcd} \end{equation} Instead, it commutes up to a homotopy $h:\calE^{\bullet}_{U}\otimes_\R \C \to \calE^{\bullet}_{U}\left(\frac{1}{i}\frac{df}{f},m\right)\otimes_\R \C$, given by the following equation \[ h(\alpha) = \frac{\exp\left(\frac{\Log(|f|)}{-i} \otimes s\right) \alpha - \alpha}{s} = \sum_{j\ge 0} \frac{s^j \left( \frac{\Log(|f|)}{-i}\right)^{j+1}}{(j+1)!} \alpha. \] A direct computation shows that indeed we have the desired homotopy (recall that the shifted differential on the target is $d[1] = -d$): \[ (-dh+hd)(\alpha) = \frac{1}{i} \frac{df}{f}\wedge \alpha - \exp\left(\frac{\Log(|f|)}{-i} \otimes s\right) \Im \frac{df}{f}\wedge \alpha . \] Hence, the diagram (\ref{eq:nastyDiagram}) commutes in the (filtered) homotopy category, so in particular it commutes in the filtered derived category. Therefore, we have defined a morphism of mixed Hodge complexes in the sense of \cite[8.1.5]{De3}, but not necessarily in the stricter sense of \cite[3.16]{peters2008mixed}. Nonetheless, since the diagram commutes up to homotopy, the diagram in hypercohomology will commute, which is enough to show that $A_\R$ induces a MHS morphism in hypercohomology following \cite[3.18]{peters2008mixed}. Note that the sheaves involved in the $\R$ part are soft. We now take global sections in the $\R$ part, to get an induced map in cohomology for all $i$: $$ H^i(A)\colon H^i(U;\R)\rightarrow H^{i+1}\Gamma\left(U;\calE_{U}^\bullet\left(\Im\frac{df}{f}, m\right)\right)\cong H^{i+1}\Gamma\left(U;\calE_{U}^\bullet(m)_{\log}\right). $$ Since it comes from a map of mixed Hodge complexes, it is a morphism of MHS. Equivalently, we can define the composition $A\colon \underline{\mathbb R}_U\to {\mathcal E}^\bullet_U\to {\mathcal E}_U^\bullet(m)_{\log} \onarrow{\eta_m^{-1}} \frac{\ov{\mathcal L}}{s^m}$, which gives us a morphism of MHS: \[ H^i(U;\R)\to H^i\Gamma(U;{\mathcal E}^\bullet_U)\to H^{i+1}\Gamma(U;{\mathcal E}_U^\bullet(m)_{\log}) \onarrow{\eta_m^{-1}} H^{i+1}\left(U;\frac{\ov{\mathcal L}}{s^m}\right). \] Indeed, the first and third arrows are MHS morphisms since they are used to define the MHS on $H^i(U;\R)$ and $H^{i+1}(U;\ov{\mathcal L}/s^m)$, respectively. Note that $\eta_m^{-1}$ is the inverse of a quasi-isomorphism, so it exists in the derived category and it induces maps in cohomology. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:proofInCircle}] First, recall that $R\Hom_{D^b_{\R}(\C^*)}^\bullet(\underline{\mathbb R}_{\C^*},\cdot)$ (morphisms in the derived category of sheaves of $\R$-vector spaces on $\C^*$) and $H^\bullet(\C^*,\cdot)$ are both the derived functor of global sections, so $\Hom_{D^b_{\R}(\C^*)}\left(\underline{\mathbb R}_{\C^*},\frac{\ov {\mathcal L}_{\C^*}}{s^m}[1]\right) \cong H^1\left({\C^*};\frac{\ov{\mathcal L}_{\C^*}}{s^m} \right)$. Let us show that this space is one-dimensional Using \cite[Example 2.5.7]{dimca2004sheaves}, $H^1(\C^*;\ov{\mathcal L}_{\C^*}/s^m)$ is isomorphic to the cokernel of $T-\Id$ acting on the stalk of $\ov{\mathcal L}_{\C^*}/s^m$, where $T$ is the monodromy action of the generator of $\pi_1(\C^*)$. This generator acts as $t^{-1} = (1+s)^{-1}\in R_m$. We can see directly that $t^{-1}-1$ equals $s$ up to multiplication by a unit in $R_m$ (namely $-t$), so its cokernel is one-dimensional. This shows that $(\phi_{m,\C^*}\circ \pi_{{\mathcal L},\C^*}^\vee)[1]$ and $A_{\C^*}$ are scalar multiples of each other as long as they are both nonzero (as classes of maps in the derived category). Let us show that $\phi_{m,\C^*}\circ \pi_{{\mathcal L},\C^*}^\vee\neq 0$. First, recall that $\pi_{{\mathcal L},\C^*}^\vee\colon \underline{\mathbb R}_{\C^*}[-1]\to \ov{\mathcal L}_{\C^*}$ is the roof diagram: \[ \begin{tikzcd} 0 \arrow[d]& \ov{\mathcal L}_{\C^*}\arrow[d,"s"]\arrow[r,"="]\arrow[l] & \ov{\mathcal L}_{\C^*}\arrow[d] \\ \underline{\mathbb R}_{\C^*} & \ov{\mathcal L}_{\C^*}\arrow[l,twoheadrightarrow] \arrow[r]& 0. \end{tikzcd} \] Seeing $\pi_{{\mathcal L},\C^*}^\vee\in \Hom_{D^b_{\R}(\C^*)}\left(\underline{\mathbb R}_{\C^*},\ov {\mathcal L}_{\C^*}[1]\right) \cong \Ext^1_{\R}(\underline{\mathbb R}_{\C^*},\ov {\mathcal L}_{\C^*})$ as the class of an extension, the roof diagram shows that it corresponds to the short exact sequence $0\to \ov{\mathcal L}_{\C^*}\onarrow{s} \ov{\mathcal L}_{\C^*} \to \underline{\mathbb R}_{\C^*}\to 0$. Note that for our purposes it doesn't matter which surjection $\ov{\mathcal L}_{\C^*}\to \underline{\mathbb R}_{\C^*}$ we use. Using the Yoneda product (see \cite[III.5, III.6]{mac}), $\phi_{m,\C^*}\circ \pi_{{\mathcal L},\C^*}^\vee$ is the extension class in $\Ext^1_R(\underline{\mathbb R},\frac{\ov{\mathcal L}}{s^m})$ given by the pushout of this short exact sequence by $\phi_{m,\C^*}$, i.e. the class of the following short exact sequence: \[ \begin{tikzcd} 0 \arrow[r]& \ov{\mathcal L} \arrow[r,"s"]\arrow[d,"\phi_{m,\C^*}"]& \ov {\mathcal L} \arrow[r,twoheadrightarrow]\arrow[d,"\phi_{m+1,\C^*}"] & \underline{\mathbb R} \arrow[r]\arrow[d,equals] & 0 \\ 0 \arrow[r]& \frac{\ov{\mathcal L}}{s^m} \arrow[r,"s"]& \frac{\ov {\mathcal L}}{s^{m+1}} \arrow[r,twoheadrightarrow]\arrow[ul, phantom, "\scalebox{1.25}{$\ulcorner$}" , very near start] & \underline{\mathbb R} \arrow[r] & 0. \end{tikzcd} \] Since $\frac{\ov{\mathcal L}}{s^{m+1}}$ is indecomposable, the sequence is not split. Therefore $\phi_{m,\C^*}\circ \pi_{{\mathcal L},\C^*}^\vee\neq 0$. To see that $0\neq A_{\C^*}$, we can see that the image of $1\in H^0(\C^*;\R)$ is $\Im\frac{dz}{z} \neq 0 \in H^1\Gamma\left(\C^*,{\mathcal E}^\bullet_{\C^*}\left(\Im\frac{dz}{z},m\right)\right) \cong H^1(\C^*;\frac{\ov{\mathcal L}}{s^m})$, so $A_{\C^*}\neq 0$. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:pullbacks}] We want to show that the following diagram is commutative: \begin{equation} \begin{tikzcd}[column sep = 4em] f^{-1}\underline{\mathbb R}_{\C^*} \arrow[r,"f^{-1}\pi_{{\mathcal L},\C^*}^\vee{[1]}"]\arrow[d]& f^{-1} \ov{\mathcal L}_{\C^*}[1] \arrow[r,"f^{-1}\phi_{m,\C^*}{[1]}"] \arrow[d]& f^{-1} \frac{\ov{\mathcal L}_{\C^*}}{s^m}[1]\arrow[d]\\ \underline{\mathbb R}_{U} \arrow[r,"\pi_{{\mathcal L},U}^\vee{[1]}"] & \ov{\mathcal L}_{U}[1] \arrow[r,"\phi_{m,U}{[1]}"] & \frac{\ov{\mathcal L}_{U}}{s^m}[1]. \end{tikzcd} \end{equation} For the right hand square, we see that it commutes because it amounts to tensoring the isomorphism $f^{-1}\ov{\mathcal L}_{\C^*}\to \ov{\mathcal L}_U$ with the map $R\to R_m$. Let us apply $\cdot^\vee= R\Homm_R(\cdot,\ul R)$ to the left hand square. Note that, for locally constant sheaves of $R$-modules, pullback is restriction of scalars on the stalks from $R[\pi_1(\C^*)]$ to $R[\pi_1(U)]$, so there is a natural isomorphism of $\pi_1(U,x)$-modules: \[ (f^{-1}\Homm_R(\calF,\calG))_x\cong (\Hom_{R}(\calF_{f(x)},\calG_{f(x)}))_{R[\pi_1(U)]}\cong \Hom_{R}((\calF_{f(x)})_{R[\pi_1(U)]},(\calG_{f(x)})_{R[\pi_1(U)]})\cong\]\[\cong\Hom_{R}((f^{-1}\calF)_x,(f^{-1}\calG)_x)\cong \Homm_{R}(f^{-1}\calF,f^{-1}\calG)_x. \] Applying this to resolutions by locally constant sheaves of free $R$-modules, we have that there is also a natural isomorphism $f^{-1}R\Homm_R^\bullet(\calF^\bullet,\calG^\bullet)\cong R\Homm_R^\bullet(f^{-1}\calF^\bullet,f^{-1}\calG^\bullet)$ for bounded complexes of locally constant sheaves. Therefore, showing that the left hand square above commutes is equivalent to showing that the following square commutes, obtained by applying $\cdot^\vee[1]$: \[ \begin{tikzcd} f^{-1}\underline{\mathbb R}_{\C^*} & f^{-1}{\mathcal L}_{\C^*} = f^{-1}\exp_!\underline{\mathbb R}_{\C} \arrow[l,"f^{-1}\pi_{{\mathcal L},\C^*}"] \\ \underline{\mathbb R}_{U} \arrow[u]& {\mathcal L}_{U} = \pi_!\underline{\mathbb R}_{U^f} \arrow[l,"\pi_{{\mathcal L},U}"] \arrow[u," \pi_!f_{\infty}^{-1}\cong f^{-1}\exp_!"']. \end{tikzcd} \] We can see that on a stalk $({\mathcal L}_{U})_x$ for some $x\in U$, the generators $\delta_{(x,z)}$ (as in Remark~\ref{rem:oppositeDual}) are all mapped to $f^{-1}1\in f^{-1}\underline{\mathbb R}_{\C^*}$ via both paths. Therefore, the square commutes and so does its dual. We consider now the following diagram, where $\eta_m$ is the map in Remark~\ref{remk:thickenedComplex}: \[ \begin{tikzcd}[column sep = 5em] f^{-1}\underline{\mathbb R}_{\C^*} \arrow[r,"f^{-1}(\eta_{m,\C^*}\circ A_{\C^*})"]\arrow[d] & f^{-1} {\mathcal E}^\bullet_{\C^*}(m)_{\log}[1] \arrow[d,"f^*"]& f^{-1} (\ov{\mathcal L}_{\C^*}/s^m)[1]\arrow[l,"\eta_{m,\C^*}"] \arrow[d,"\pi_!f^{-1}_{\infty}\cong f^{-1}\exp_!"] \\ \underline{\mathbb R}_{U} \arrow[r,"\eta_{m,U}\circ A_U"] & {\mathcal E}_{U}^\bullet(m)_{\log} [1]& \ov{\mathcal L}_{U}/s^m[1]\arrow[l,"\eta_{m,U}"] \end{tikzcd} \] The arrow $f^*$ is induced by sending $\alpha\otimes s^j\in {\mathcal E}_{\C^*}^i(m)_{\log}$ defined on an open set $V$ to $f^*\alpha\otimes s^j\in {\mathcal E}_{U}^i(m)_{\log}$ defined on $f^{-1}(V)$. Just from the definitions we can see that the left hand square commutes, taking into account that $f^*z= f$. Let us see that the right hand square commutes. Note that applying section~\ref{ssAlex} to the special case $f= \Id_{\C^*}$, the stalk of $\ov{\mathcal L}_{\C^*}$ at some $y\in \C^*$ is generated by sections of the form $\ov\delta_{\wt y}$ for $\wt y\in \C$ such that $e^{\wt y} = y$. Let us look at the stalk at $x\in U$, and let $z\in \C$ be such that $(x,z)\in U^f$. Then $\ov\delta_{z}\in (\ov{\mathcal L}_{\C^*}/s^m)_{f(x)}$. Since $f_{\infty}$ maps $(x,z)\in U^f$ to $z\in \C$, the image of $f^{-1}\ov\delta_{z}$ in the stalk of $\ov{\mathcal L}_{U}/s^m$ is $\ov\delta_{(x,z)}$. Now, let $\wt\log$ be a local branch of the logarithm such that $\wt\log(f(x)) = z$. Then, $(\Id_U,\wt\log\circ f)\colon U\to U^f\subset U\times \C$ is a local section of $\pi$ mapping $x$ to $(x,z)$ (it is only defined on a neighborhood $V$ of $x$ small enough that $f(V)$ is contained in the domain of $\wt\log$). Then, from Remark~\ref{remk:thickenedComplex}, $\eta_{m,\C^*}(\ov\delta_{z}) = \exp\left(-\frac{\Im\wt\log }{2\pi}\otimes \log(1+s)\right)$, so taking the image by $f^*$ we obtain $$ f^*\circ \eta_{m,\C^*}(\ov\delta_z)=\exp\left(-\frac{\Im\wt\log \circ f}{2\pi}\otimes \log(1+s)\right)=\eta_{m,U}(\ov\delta_{(x,z)}), $$ as desired. We have shown that both diagrams in the lemma commute. Now, we claim that there is some $c\in \R\setminus \{0\}$ for which the vertical arrows in this commutative diagram coincide: \[ \begin{tikzcd}[column sep = 6em] f^{-1}\underline{\mathbb R}_{\C^*}[-1] \arrow[d,"c\cdot f^{-1}A_{\C^*}", shift left = 0.5ex] \arrow[d,"f^{-1}(\phi_{m,\C^*}\circ \pi_{{\mathcal L},\C^*}^\vee)"', shift right = 0.5ex]\arrow[r,"\sim"] & \underline{\mathbb R}_{U}[-1] \arrow[d,"c\cdot A", shift left = 0.5ex] \arrow[d,"\phi_{m}\circ \pi_{{\mathcal L}}^\vee"', shift right = 0.5ex] \\ f^{-1}\frac{\ov{\mathcal L}_{\C^*}}{s^m}\arrow[r,"\sim"] & \frac{\ov{\mathcal L}_{U}}{s^m}. \end{tikzcd} \] By Lemma~\ref{lem:proofInCircle}, there exists some $c\in \R\setminus \{0\}$ for which the left hand vertical arrows coincide. By the discussion above, this implies that the right hand vertical arrows coincide as well. \end{proof} This concludes the proof of Theorem \ref{thm:geomIsMHS}. \end{proof} \begin{remk}\label{rem:holyGrail} Suppose that the action of $t$ on all cohomology groups $\Tors_R H^{i+1}(U;\ov{\mathcal L})$ is unipotent and that $m$ is large enough that $(t-1)^m$ annihilates $\Tors_R H^{i+1}(U;\ov{\mathcal L})$. The map $A\colon \underline{\mathbb R}\to \ov{\mathcal L}/s^m [1]$ used in the proof of Theorem~\ref{thm:geomIsMHS} induces a map in cohomology $H^i(A)\colon H^i(U;\R)\to H^{i+1}(U;\ov{\mathcal L}/s^m)$ which factors (up to some $c\in \R\setminus \{0\}$) into the following $R$-linear morphisms of MHS: \[ \begin{tikzcd} H^i(U;\R) \arrow[rr,"cH^i(A)"] \arrow[dr,"H^i(\pi)"']& & H^{i+1}\left(U;\frac{\ov{\mathcal L}}{s^m}\right).\\ & \Tors_R H^{i+1}(U;\ov{\mathcal L}) \arrow[ur,"H^{i+1}(\phi_m)"',hookrightarrow] \end{tikzcd} \] Indeed: by Lemma~\ref{lem:pullbacks}, there is a $c$ for which $cA = \phi_m\circ \pi_{{\mathcal L}}^\vee$. By Proposition~\ref{prop:mapsAreEqual}, $\pi_{{\mathcal L}}^\vee$ induces the map $H^i(\pi)$ in cohomology (which is a MHS morphism by Theorem~\ref{thm:geomIsMHS}); and by Remark~\ref{remk:mhsSummary}, $H^{i+1}(\phi_m)$ is the injection used to define the MHS on $\Tors_R H^{i+1}(U;\ov{\mathcal L})$, so in particular it is also a morphism of MHS. From the relation between $A$ and $A_\R$ described in Lemma~\ref{lem:theMapA}, we have that $H^i(A_\R)=H^i(A)$, up to identifying the cohomology of the domain and target local systems with that of the global sections of the corresponding sheaf of cdgas. Recall that these identifications are morphisms of MHS by definition.$\sslash$ \end{remk} \section{Consequences of Theorem \ref{thm:geomIsMHS}}\label{sec:consequences} We start by identifying some properties of the maps induced by covering spaces in homology and cohomology. Let $N\in\Z_{>0}$, and let $p\colon U_N\rightarrow U$ be the $N$-fold cover described in Section \ref{sscover}. The covering space $\pi\colon U^f\rightarrow U$ factors through $p\colon U_N\rightarrow U$ by $\pi_N\colon U^f\rightarrow U_N$. Recall that the deck transformation of $\pi_N$ is $t^N$. \begin{prop}\label{prop:kerim} The kernel of the map $H_j(\pi_N)\colon \Tors_R H_{j}(U;{\mathcal L})\to H_j(U_N;k)$ is $(t^N-1)\cdot \Tors_R H_{j}(U;{\mathcal L})$. The image of $H^j(\pi_N)\colon H^j(U_N;k)\rightarrow \Tors_R H^{j+1}(U;\ov{\mathcal L})$ is the $(t^N-1)$-torsion of the target, namely $$ \Tors_{(t^N-1)} H^{j+1}(U;\ov{\mathcal L})\coloneqq\{a\in \Tors_R H^{j+1}(U;\ov{\mathcal L})\ \vert\ (t^N-1)a=0\}. $$ \end{prop} \begin{proof} We use a similar argument to \cite[Assertion 5]{milnor}. We consider the short exact sequence of chain complexes of $R$-modules $$ 0\to C_\bullet(U^f;k)\xrightarrow{\cdot (t^N-1)}C_\bullet(U^f;k)\xrightarrow{(\pi_N)_*}C_\bullet(U_N;k)\to 0, $$ where $t$ acts on both $U^f$ and $U_N$ by deck transformations. It gives us a long exact sequence in homology (the Milnor sequence), from which we get the exact sequence $$ H_j(U;{\mathcal L})\xrightarrow{\cdot(t^N-1)}H_j(U;{\mathcal L})\xrightarrow{H_j(\pi_N)} H_j(U_N;k). $$ This in turn tells us that the following sequence is exact $$ \Tors_R H_j(U;{\mathcal L})\xrightarrow{\cdot(t^N-1)}\Tors_R H_j(U;{\mathcal L})\xrightarrow{H_j(\pi_N)} H_j(U_N;k), $$ which implies the assertion about the kernel of $H_j(\pi_N)$. We take $\Hom_k(\cdot,k)$ of the above exact sequence and get the exact sequence \[ H^j(U_N;k)\xrightarrow{H^j(\pi_N)} \Hom_k(\Tors_R H_j(U;{\mathcal L}),k)\xrightarrow{\cdot(t^N-1)} \Hom_k(\Tors_R H_j(U;{\mathcal L}),k). \] from which we get the claim about the image of $H^j(\pi_N)$. \end{proof} The following corollary is a direct consequence of Proposition \ref{prop:kerim}. It says that, if the monodromy action is semisimple, the MHS on $\Tors_R H^{j+1}(U;\ov{\mathcal L})$ (resp. $\Tors_R H_{j}(U;{\mathcal L})$) is determined by $H^j(\pi_N)$ and Deligne's MHS on $H^j(U_N;k)$ (resp. by $H_j(\pi_N)$ and Deligne's MHS on $H_j(U_N;k)$, which is the dual MHS of $H^j(U_N;k)$). \begin{cor}\label{cor:surj} Suppose that the $t$-action on $\Tors_R H^{j+1}(U;\ov {\mathcal L})$ (equiv. on $\Tors_R H_{j}(U;{\mathcal L})$) is semisimple for some $j$, and let $N$ be such that the action of $t^N$ on $\Tors_R H_j(U;{\mathcal L})$ is unipotent. Then $H^j(\pi_N)\colon H^j(U_N;k)\to \Tors_R H^{j+1}(U;\ov{\mathcal L})$ is a surjective morphism of MHS. Equivalently, $H_j(\pi_N)\colon \Tors_R H_{j}(U;{\mathcal L})\rightarrow H_j(U_N;k)$ is an injective morphism of MHS. \end{cor} \begin{remk}\label{rem:decomposition} Let $A$ be a torsion $R$-module which is annihilated by $t^N-1$. Then, there is a canonical isomorphism $$ A_1\oplus A_{\neq 1}\cong A $$ where $A_1=\ker(A\xrightarrow{\cdot(t-1)}A)$ and $A_{\neq 1}=\ker(A\xrightarrow{\cdot(t^{N-1}+\ldots+t+1)}A)$.$\sslash$ \end{remk} \begin{cor}[of Corollary \ref{cor:surj}]\label{cor:t} Suppose that the $t$-action on $\Tors_R H^{j+1}(U;\ov {\mathcal L})$ (equiv. on $\Tors_R H_{j}(U;{\mathcal L})$) is semisimple for some $j$. Then, multiplication by $t$ induces a MHS morphism from $\Tors_R H^{j+1}(U;\ov {\mathcal L})$ (equiv. on $\Tors_R H_{j}(U;{\mathcal L})$) to itself. In particular, we have MHS isomorphisms $$ \Tors_R H^{j+1}(U;\ov {\mathcal L})\cong\Tors_R H^{j+1}(U;\ov {\mathcal L})_1\oplus\Tors_R H^{j+1}(U;\ov {\mathcal L})_{\neq 1} $$ and $$ \Tors_R H_j(U;{\mathcal L})\cong\Tors_R H_j(U;{\mathcal L})_1\oplus\Tors_R H_j(U;{\mathcal L})_{\neq 1} $$ \end{cor} \begin{proof} We write the proof in the cohomology case. Let $N$ be such that the action of $t^N$ on $\Tors_R H_j(U;{\mathcal L})$ is unipotent. Since $t$ acts on $U_N$ by deck transformations, and $U_N\rightarrow U$ is an algebraic covering map, we get that the deck transformation corresponding to $t$ is actually an algebraic map, so multiplication by $t$ on $H^j(U_N)$ is a morphism of MHS. The result follows from the commutativity of the following diagram and the fact that all the arrows except for the bottom one are known to be morphism of MHS. $$ \begin{tikzcd} H^j(U_N;k)\arrow[r,"t"]\arrow[d, "H^j(\pi_N)",twoheadrightarrow]& H^j(U_N;k)\arrow[d, "H^j(\pi_N)",twoheadrightarrow]\\ \Tors_R H^{j+1}(U;\ov {\mathcal L})\arrow[r,"t"] & \Tors_R H^{j+1}(U;\ov {\mathcal L}). \end{tikzcd} $$ The last statement about the direct sum decomposition follows from Remark \ref{rem:decomposition}. \end{proof} The converse to Corollary~\ref{cor:t} also holds. \begin{prop}\label{prop:semisimpleIsEasyConverse} Suppose that the $t$-action on $\Tors_R H^{j+1}(U;\ov {\mathcal L})$ (equiv. on $\Tors_R H_{j}(U;{\mathcal L})$) is a MHS morphism for some $j$. Then, the action of $t$ is semisimple. \end{prop} \begin{proof} Let $H =\Tors_R H^{j+1}(U;\ov {\mathcal L})$. Let $N$ be such that $t^N$ is unipotent (and a morphism of MHS on $H$). By Corollary~\ref{alexandermhs}, $\log(t^N)$ is a morphism $H\to H(-1)$ (viewed as a power series around $t^N=1$). Therefore, we have a MHS $H$ with a nilpotent morphism of MHS $\theta \coloneqq \log(t^N)\colon H\to H(-1)$ such that $\exp(\theta)\colon H\to H$ is also a MHS morphism. We want to prove that $\theta=0$. For this it is enough to show that $\ker \theta = \ker \theta^2$, so we can assume that $\theta^2 (H) = 0$. Therefore, $\exp(\theta) = 1+\theta$. Since $1+\theta$ is a MHS morphism, then $\theta$ is as well. Therefore, we have to show that if $\theta:H\to H$ and $\theta:H\to H(-1)$ are MHS morphisms, then $\theta=0$. This is a direct consequence of \cite[Corollary 3.6]{peters2008mixed}, where it is shown that a MHS morphism preserves the weight filtration strictly. Consider the image MHS $\theta (H)\subseteq H$. Then, \[ W_k(\theta (H))\coloneqq W_kH\cap \theta (H) = \theta (W_kH) = \theta (W_{k+2}H) \quad\forall k\Rightarrow W_k(\theta (H)) = W_{k+1}(\theta (H))\quad\forall k. \] Since the filtration $W_k$ is finite, for $k\gg 0$, $0=W_{-k}(\theta(H)) = W_{k}(\theta (H)) = \theta (H)$. \end{proof} Corollary~\ref{cor:t} and Proposition~\ref{prop:semisimpleIsEasyConverse} together give us Theorem \ref{tsemisimpleIntro} in the introduction, which is stated in homological notation. \subsection{Relationship with the cup and cap products.}\label{ss:cupcap} Consider the real component\linebreak$j_*\calE^\bullet_U\left(\Im\, \frac{df}{f},m\right)$ of the $\R$-mixed Hodge complex of sheaves $\calH dg^\bullet(X\,\log D)\left(\frac{1}{i}\frac{df}{f},m\right)$ used to endow the torsion part of the Alexander modules with a MHS. The differential of $\calE^\bullet_U\left(\Im\, \frac{df}{f},m\right)$ involves a wedge with $\Im\, \frac{df}{f}$. Wedging real differential forms corresponds to cup products in cohomology, which loosely suggests a relation between Alexander modules and the cup (and cap) products arising from the thickened complexes. In this section, we explore this relation and make it explicit. \begin{prop}\label{prop:cup} Let $N$ be such that the action of $t^N$ on $\Tors_R H^{j+1}(U;\ov{\mathcal L})$ is unipotent. Let $\gen\in H^1(\C^*;\Z)\cong \Z$ be a generator. We consider $\gen$ as an element of $H^1(\C^*;k)$, and $f_N^*(\gen)\in H^1(U_N;k)$. There exists an arrow (the dashed one) that makes the following diagram a commutative diagram of morphisms of MHS (and of $R$-modules, if $t$ acts on $U_N$ by deck transformations). Up to multiplication by a nonzero constant, it is induced by $\pi_{\ov{\mathcal L}}\colon \ov{\mathcal L}\to\ul k$. $$ \begin{tikzcd} H^j(U_N;k)\arrow[rr, " f_N^*(\gen)\smile"]\arrow[dr, "H^j(\pi_N)"]&\ & H^{j+1}(U_N;k)(1).\\ \ &\Tors_R H^{j+1}(U;\ov{\mathcal L})\arrow[ru, dashed]&\ \end{tikzcd} $$ Here $(1)$ denotes the Tate twist, and $f_N^*(\gen)\smile $ denotes the cup product by $f_N^*(\gen)$. Moreover, if the $t$-action on $\Tors_R H^{j+1}(U;\ov {\mathcal L})$ is semisimple, the dashed arrow is injective. \end{prop} \begin{proof} To make notation simpler, we may replace $U_N$ by $U$, $f_N$ by $f$, $\pi_N$ by $\pi$ and assume that the action of $t$ on $\Tors_R H^*(U;{\mathcal L})$ is unipotent. Let $m\in \N$ be the minimum natural number such that $(t-1)^m$ annhilates $\Tors_R H^{i+1}(U;\calL)$. Note that the $t$-action on $\Tors_R H^{i+1}(U;\calL)$ is semisimple if and only if $m=1$. We continue the rest of the proof in the case when $k=\R$, from which the case $k=\Q$ will follow. Let $A_{\calH dg}\colon \Hdg{X}{D}\rightarrow \Hdg{X}{D}\left(\frac{1}{i}\frac{df}{f}, m \right)[1]$ be as in Lemma \ref{lem:theMapA}, which induces a map of MHS $H^j(A_\R)\colon H^j(U)\rightarrow H^{j+1}\Gamma(U;{\mathcal E}_U^\bullet(\Im df/f,m))$. Note that $\Im df/f=f^*(\Im dz/z)$, so there exists a non-zero constant $b\in\R^*$ such that $f^*(\gen)=b\Im df/f$. Let $\phi_{m-1,1}^*$ be the map induced in cohomology, corresponding to the map ${\mathcal E}_U^\bullet(\Im df/f, m)\rightarrow {\mathcal E}_U^\bullet(\Im df/f, 1)\cong {\mathcal E}_U^\bullet$ coming from the projection $R_m\rightarrow R_1$. Note that $\phi_{0,1}^*$ is the identity map. By Lemma~\ref{inducedmhsmaps}, $\phi_{m-1,1}^*$ is a MHS morphism. Let $c\in\R^*$ be as in Remark \ref{rem:holyGrail}. By Remark \ref{rem:holyGrail}, we get the following commutative diagram of morphisms of MHS, which finishes the proof. $$ \begin{tikzcd}[column sep = 7em] H^j(U)\cong H^j\Gamma(U;{\mathcal E}_U^\bullet)\arrow[r," f^*(\gen)\smile =b\cdot( \Im df/f \smile )"]\arrow[dr, "H^j(A_\R)"]\arrow[d, "H^j(\pi)"] & H^j\Gamma(U;{\mathcal E}_U^\bullet(\Im df/f,1)[1])\cong H^{j+1}(U)(1)\\ \Tors_R H^{j+1}(U;\ov{\mathcal L})\arrow[r,"\frac{1}{c}\cdot H^{j+1}(\phi_m)"',hookrightarrow] & H^j\Gamma(U;{\mathcal E}_U^\bullet(\Im df/f,m)[1])\cong H^{j+1}(U;\ov{\mathcal L}/s^m)\arrow[u, "b\cdot\phi_{m-1,1}^*{[}1{]}"]. \end{tikzcd} $$ \end{proof} \begin{remk}\label{rem:cup} We have not used it in the proof above, but the cup product $H^m(Y;k)\otimes H^1(Y;k)\rightarrow H^{m+1}(Y;k)$ is a morphism of MHS for every algebraic variety $Y$ (\cite[Corollary 5.45]{peters2008mixed}). The Tate twist in the target of the horizontal map in the commutative diagram above agrees with the fact that $\gen$ has weight $2$ in $H^1(\C^*;\R)$ (the MHS on $H^1(\C^*;\R)$ is pure of type $(1,1)$).$\sslash$ \end{remk} The following result follows from Corollary \ref{cor:surj} and Proposition \ref{prop:cup}. The result in homology, which already appeared in the introduction as Theorem~\ref{finiteIntro}A, follows from taking $\Hom_k(\cdot,k)$ in the commutative diagram of Proposition \ref{prop:cup}. \begin{cor}\label{cor:cup} Let $N$ be such that the action of $t^N$ acts unipotently on $\Tors_R H^{j+1}(U;\ov{\mathcal L})$. Let $\gen\in H^1(\C^*;\Z)\cong \Z$ be a generator. We consider $\gen$ as an element of $H^1(\C^*;k)$, and $f_N^*(\gen)\in H^1(U_N;k)$. Suppose that the $t$-action on $\Tors_R H_j(U;{\mathcal L})$ (equivalently, on $\Tors_R H^{j+1}(U,\ov{\mathcal L})$) is semisimple. Then, \begin{itemize} \item $\Tors_R H^{j+1}(U,\ov{\mathcal L})$ is isomorphic, both as MHS and as $R$-modules, to the image of the cup product map $$H^j(U_N;k)\xrightarrow{ f_N^*(\gen)\smile } H^{j+1}(U_N;k)(1).$$ \item $\Tors_R H_j(U;{\mathcal L})$ is isomorphic, both as MHS and as $R$-modules, to the image of the cap product map $$ H_{j+1}(U_N;k)(-1)\xrightarrow{\frown f_N^*(\gen) } H_j(U_N;k) $$ \end{itemize} Here $H^*(U_N;k)$ are endowed with Deligne's MHS, and $H_*(U_N;k)$ are endowed with the dual of Deligne's MHS in cohomology. \end{cor} \subsection{Relationship with Deligne's MHS on the generic fiber.}\label{ss:genFiber} Let $F$ be a generic fiber of $f\colon U\rightarrow \C^*$ (as in Definition~\ref{def:genericFiber}), and let $N$ be chosen such that the action of $t^N$ on $\Tors_R H^{j+1}(U;\ov{\mathcal L})$ is unipotent. Let $i\colon F\hookrightarrow U$ be the inclusion. As in Section~\ref{genfiber}, $F$ lifts to $U_N$ and $U^f$ via maps $i_N$ and $i_\infty$, making the following diagram commutative, where the vertical arrows are covering space maps. $$ \begin{tikzcd} \ & U^f\arrow[d, "\pi_N"]\arrow[dd, "\pi", bend left=90]\\ \ & U_N\arrow[d,"p"]\\ F\arrow[r,"i",hook]\arrow[ru, "i_N" pos=1, hook]\arrow[ruu, "i_{\infty}",hook] & U. \end{tikzcd} $$ Note that, in homology, the composition $i_N=\pi_N\circ i_{\infty}$ factors through $\Tors_R H_j(U^f;k)\cong \Tors_R H_j(U;{\mathcal L})$. Hence, we get \begin{equation}\label{eq:fiberhom} \begin{tikzcd} H_j(F;k)\arrow[r,"H_j(i_\infty)",two heads]\arrow[rr, "H_j(i_N)", bend right = 20] & \Tors_R H_j(U;{\mathcal L})\arrow[r, "H_j(\pi_N)"] & H_j(U_N;k), \end{tikzcd} \end{equation} and taking $\Hom_k(\cdot,k)$ and the appropriate identifications, we get \begin{equation}\label{eq:fibercohom} \begin{tikzcd} H^j(U_N;k)\arrow[r,"H^j(\pi_N)"]\arrow[rr, "H^j(i_N)", bend right = 20] & \Tors_R H^{j+1}(U;\ov{\mathcal L})\arrow[r, "H^j(i_\infty)",hook] & H^j(F;k). \end{tikzcd} \end{equation} Here $H_j(i_\infty)$ and $H^j(i_\infty)$ are surjective and injective respectively by Proposition~\ref{genf} and Remark~\ref{remk:tube}. Note that $H_j(i_N)$ and $H^j(i_N)$ are MHS morphisms, since $i_N$ is an algebraic map. Recall that $H_j(\pi_N)$ and $H^j(\pi_N)$ are MHS morphisms by Theorem \ref{thm:geomIsMHS}. If the monodromy is semisimple, Corollary \ref{cor:surj} tells us that $H_j(\pi_N)$ and $H^j(\pi_N)$ are injective and surjective respectively. The information in this paragraph and the commutativity of the two diagrams above tell us that $H_j(i_\infty)$ and $H^j(i_\infty)$ are also MHS morphisms. \begin{cor}\label{cor:fiber} Let $N$ be such that the action of $t^N$ on $\Tors_R H^{j+1}(U;\ov{\mathcal L})$ is unipotent. Suppose that the $t$-action on $\Tors_R H_j(U;{\mathcal L})$ (equivalently, on $\Tors_R H^{j+1}(U,\ov{\mathcal L})$) is semisimple. Then, we have the following commutative diagrams, where all the arrows are morphisms of MHS. $$ \begin{tikzcd} H_j(F;k)\arrow[r,"H_j(i_\infty)",two heads]\arrow[rr, "H_j(i_N)", bend right = 20] & \Tors_R H_j(U;{\mathcal L})\arrow[r, "H_j(\pi_N)", hook] & H_j(U_N;k), \end{tikzcd} $$ and $$ \begin{tikzcd} H^j(U_N;k)\arrow[r,"H^j(\pi_N)", two heads]\arrow[rr, "H^j(i_N)", bend right = 20] & \Tors_R H^{j+1}(U;\ov{\mathcal L})\arrow[r, "H^j(i_\infty)",hook] & H^j(F;k). \end{tikzcd} $$ Therefore, $\Tors_R H_j(U;{\mathcal L})$ (resp. $\Tors_R H^{j+1}(U;\ov{\mathcal L})$) is isomorphic as MHS to the image of the MHS morphism $H_j(i_N)$ (resp. $H^j(i_N)$). \end{cor} Note that Theorem~\ref{finiteIntro}B is an immediate consequence of the statement in homology of Corollary ~\ref{cor:fiber}. \begin{remk}\label{rem:indepF} In general, the MHS on $F$ depends on the specific choice of the fiber, even if $F$ is generic. Corollary \ref{cor:fiber} above tells us that the map induced by a lift of the inclusion of the generic fiber $F$ of $f$ into the infinite cyclic cover $U^f$ induces a MHS morphism in both homology and cohomology for \textbf{any} choice of generic fiber $F$.$\sslash$ \end{remk} \begin{cor}\label{cor:fiberConverse} The $t$-action on $\Tors_R H_j(U;{\mathcal L})$ (equivalently, on $\Tors_R H^{j+1}(U;\ov{\mathcal L})$) is semisimple if and only if for any generic fiber $F\subset U^f$, the induced map in homology \linebreak$H_j(i_\infty)\colon H_j(F;k)\to \Tors_R H_j(U^f;k)$ is a MHS morphism (equivalently, the dual map \linebreak$H^j(i_\infty)\colon \Tors_R H^{j+1}(U;\ov{\mathcal L})\to H^j(F;k)$ is a MHS morphism). \end{cor} \begin{proof} The forward direction is Corollary~\ref{cor:fiber}. For the converse, consider an inclusion or a fiber $i_\infty\colon F\to U^f$ and the deck transformation $t\colon U^f\to U^f$. Consider the induced maps in homology: \[ \begin{tikzcd} H_i(F;k) \arrow[r,"i_\infty",twoheadrightarrow]\arrow[d,equals]& \Tors_R H_i(U^f;k)\subseteq H_i(U^f;k)\arrow[d,"t",shift right = 2ex] \\ H_i(F;k) \arrow[r,"t\circ i_\infty",twoheadrightarrow]& \Tors_R H_i(U^f;k)\subseteq H_i(U^f;k). \end{tikzcd} \] The horizontal arrows are surjections by Proposition~\ref{genf}, and by hypothesis they are MHS morphisms. Therefore, the map ``$t$'' must be a MHS morphism as well. By Proposition~\ref{prop:semisimpleIsEasyConverse}, this implies that $t$ is semisimple. \end{proof} Let $f\in\C[x_1,\ldots,x_n]$ be a weighted homogeneous polynomial, and let $U=\linebreak\C^n\setminus\{f=0\}$. In this case, we have a global Milnor fibration $f\colon U\to\C^*$. Moreover, assume that $f\colon U\to\C^*$ induces an epimorphism on fundamental groups, which happens if and only if the greatest common divisor of the exponents of the distinct irreducible factors of $f$ is $1$. Let $F$ be a fiber of $f\colon U\to\C^*$, and let $i_\infty\colon F\hookrightarrow U^f$ be a lift of the inclusion $i\colon F\hookrightarrow U$. Note that, since $f\colon U\rightarrow \C^*$ is a fibration, $i_\infty$ is a homotopy equivalence, so it induces isomorphisms $H_j(F;k)\rightarrow H_j(U^f;k)$ for all $j$, which are compatible with the $t$-action (see Lemma~\ref{lem:fiberMonodromy}). Moreover, the $t$-action on $F$ comes from an algebraic map $F\rightarrow F$ of finite order, so the $t$-action on $H_j(F)$ and $H^j(F)$ is semisimple. Applying Corollary \ref{cor:fiber}, we have arrived at the following result. \begin{cor}[The Alexander modules recover the MHS on the global Milnor fiber]\label{cor:quasihom} Let $f\in\C[x_1,\ldots,x_n]$ be a weighted homogeneous polynomial, and let $U=\C^n\setminus\{f=0\}$. Assume that $f\colon U\rightarrow \C^*$ induces an epimorphism in fundamental groups. Let $F$ be a fiber of $f\colon U\rightarrow \C^*$, and let $i_\infty\colon F\hookrightarrow U^f$ be a lift of the inclusion $i\colon F\hookrightarrow U$. The map $$ \Tors_R H^{j+1}(U;\ov{\mathcal L})\rightarrow H^j(F;k) $$ induced by $i_\infty$ is a MHS isomorphism, where $H^j(F;k)$ is endowed with Deligne's MHS. \end{cor} \subsection{Dimca-Libgober and Liu's MHS}\label{ss:DLandLiu} In this section, we recall other MHS on Alexander modules in the literature, and we show that the MHS that we consider in this paper generalizes them. Let $F$ be a generic fiber of $f\colon U\rightarrow \C^*$, as in Definition~\ref{def:genericFiber}. Let $\pi\colon U^f\rightarrow U$ be the covering space induced by $f_*\colon \pi_1(U)\twoheadrightarrow \Z$, Let $i\colon F\hookrightarrow U$ be the inclusion, and let $i_{\infty}\colon F\hookrightarrow U^f$ be a lift of $i$, as in Section~\ref{genfiber}. We make the following assumption in this section. \begin{assumption} Fix $j\geq 0$, and suppose that $\pi^*\colon H^j(U;\Q)\rightarrow H^j(U^f;\Q)$ is an epimorphism, and $(i_{\infty})^*\colon H^j(U^f;\Q)\rightarrow H^j(F;\Q)$ is a monomorphism. \label{a1} \end{assumption} Endow $H^j(U;\Q)$ and $H^j(F;\Q)$ with Deligne's MHS. Under Assumption \ref{a1}, we have the following commutative diagram: \begin{center} \begin{tikzcd} H^j(U;\Q) \arrow[rd, "i^*"] \arrow[r, two heads, "\pi^*"] & H^j(U^f;\Q)\arrow[d, hook,"(i_{\infty})^*"] \\ \ & H^j(F;\Q) \end{tikzcd} \end{center} \begin{remk} \label{nota} Since $i^*$ is a map of MHS, we can endow $H^j(U^f;\Q)$ with a unique MHS such that both $\pi^*$ and $(i_{\infty})^*$ are maps of MHS. Indeed, the image of $i^*$ is a sub-MHS of $H^j(F;\Q)$ that is identified through $(i_{\infty})^*$ with $H^j(U^f;\Q)$.$\sslash$ \end{remk} \begin{set}\label{set:DL} In their paper \cite{DL}, Dimca and Libgober consider the following setting: Let $W'=W_0'\cup\ldots\cup W_m'$ be a hypersurface arrangement in ${\mathbb P}^N$ for $N > 1$, where $W_j'$ is a hypersurface of degree $d_j$ defined by the equation $g_j=0$, where $g_j$ is a reduced homogeneous polynomial. Let $Z\subset {\mathbb P}^N$ be a smooth complete intersection of dimension $n>1$ which is not contained in $W'$, and let $W_j =W_j'\cap Z$ for $j=0,...,m$ be the corresponding hypersurface in $Z$. Let $W = W_0\cup\ldots\cup W_m$ denote the corresponding hypersurface arrangement in $Z$. Moreover, assume that the following three conditions hold: \begin{itemize} \item All the hypersurfaces $W_j$ are distinct, reduced and irreducible; moreover $W_0$ is smooth. \item The hypersurface $W_0$ is transverse in the stratified sense to $T = W_1\cup\ldots\cup W_m$, i.e., if $\mathcal{S}$ is a Whitney regular stratification of $T$ , then $W_0$ is transverse to any stratum $S\in\mathcal{S}$. \item $d_0$ divides the sum $\sum_{j=1}^m d_j$ , say $dd_0 = \sum_{j=1}^m d_j$. \end{itemize} The complement $M= Z \backslash W_0$ is a smooth affine variety. Let $Y$ be the hypersurface $Y=M\cap T$ in $M$. The authors consider the variety $U =M\backslash Y$, with the map $$ \f{f}{U}{\C^*}{x}{\frac{g_1(x)\cdot\ldots\cdot g_m(x)}{g_0(x)^d}}. $$ By \cite[Theorem 1.2]{DL}, the generic fiber of this map is connected. \end{set} \begin{remk} Note that, in this setting, $U$ is an affine variety of dimension $n$, so it has the homotopy type of a finite $n$-dimensional CW-complex, which implies that $H^j(U^f,\Q)$ is a free $R$-module for $j=n$, and $0$ if $j>n$. Moreover, by \cite[Corollary 1.6]{DL}, $H^j(U^f,\Q)$ is annihilated by $t^d-1$ (and therefore is $R$-torsion and semisimple) for $j<n$.$\sslash$ \label{affine} \end{remk} Let $p\colon U_d\rightarrow U$ be the $d$-fold cover of $U$ obtained as the pullback by $f$ of the map $g\colon \C^*\rightarrow \C^*$ given by $z\mapsto z^d$, as exemplified by the following commutative diagram: \begin{center} \begin{tikzcd} U_d \arrow[d, "p"] \arrow[r, "f_d"] & \C^*\arrow[d, "g"] \\ U \arrow[r, "f"] & \C^*. \end{tikzcd} \end{center} Note that $F$ is the generic fiber of $f_d$ as well. Let $\pi_d\colon U^f\rightarrow U_d$ be the covering space such that $\pi=p\circ \pi_d$, and let $i_d$ be a lift of $i$ to $U_d$. Note that $i_{\infty}$ can be considered to be a lift of $i_d$. \begin{remk}[Dimca and Libgober's MHS] By \cite[Theorem 1.5]{DL}, $(\pi_d)^*\colon H^j(U_d;\Q)\rightarrow H^j(U^f;\Q)$ is an epimorphism for all $j<n$. By \cite[Corollary 1.3]{DL}, $(i_{\infty})^*\colon H^j(U^f;\Q)\rightarrow H^j(F;\Q)$ is an isomorphism if $j<n-1$ and a monomorphism if $j=n-1$. This means that if we switch $U$ for $U_d$ and $\pi$ for $\pi_d$, we are under the conditions of Assumption \ref{a1}. Hence, there is a unique MHS on $H^j(U^f;\Q)$ that makes $\pi_d^*$ and $(i_{\infty})^*$ into morphisms of MHS for $j<n$, and this MHS is the one that Dimca and Libgober construct in \cite{DL} (recall Remark \ref{nota}).$\sslash$ \label{remDL} \end{remk} In \cite{Liu}, Liu considers a particular case of the situation described above, which was initially studied in \cite{Max06}. In his setting, $N=n$, $W_0'$ is the hyperplane at infinity and $Z={\mathbb P}^{n}$. Then, $U$ is the affine complement of a hypersurface transversal at infinity, and $f$ induces the linking number homomorphism in fundamental groups. That is, the image by $f_*$ of a positively oriented meridian around each of the irreducible components $W_j$ of $X$ gets mapped to $1\in\Z$, for $j=1,\ldots,m$. He recovers Dimca and Libgober's MHS with a completely different approach, using nearby cycles. \begin{cor}[of Corollary \ref{cor:fiber}]\label{cor:recoverMHS} Suppose $U$ and $f$ are as in Setting \ref{set:DL}. Under the isomorphism given by Corollary~\ref{isocohom}, the MHS on $\Tors_R H^{j+1}(U;\ov\calL)$ constructed in this paper coincides with the MHS on $H^j(U^f;\Q)$ constructed in \cite{DL} , for $0\leq j\leq n-1$. \end{cor} \subsection{Bounding the weights} In this section we prove the following result, which was already stated in the introduction as Theorem~\ref{boundsIntro} (using homological instead of cohomological notation). \begin{thm}\label{thm:boundedWeights} If $k\notin [i,2i]\cap [i,2n-2]$, then \[ \Gr^W_k \Tors_RH^{i+1}(U;\ov{\mathcal L}) = 0, \] where $n$ is the dimension of $U$. \end{thm} Before the proof, we will discuss some consequences. In the context of Alexander modules realized by an algebraic map (as considered in this paper), the following corollary improves on \cite[Proposition 1.10]{BudurLiuWang}, as the bound below is the ceiling of half the bound in loc. cit. \begin{cor}\label{cor:jordan} Every Jordan block of the action of $t$ on $\Tors_R H_i(U^f;k)$ (respectively, on $\Tors_R H^{i+1}(U;\ov{\mathcal L})$) has size at most $\min\{ \lceil (i+1)/2\rceil, n- \lfloor (i+1)/2\rfloor\}$. (Here, Jordan blocks are considered over the algebraic closure of the field $k$.) \end{cor} \begin{proof} Let $K=\min\{ \lceil (i+1)/2\rceil, n-\lfloor(i+1)/2\rfloor\}$. We will start by showing the corollary in the case where the action of $t$ is unipotent. By Corollary~\ref{alexandermhs}, $\log(t) = \sum_{j=0}^\infty (-1)^{j-1}\frac{(t-1)^j}{j}$ is a morphism of MHS into the $(-1)$st Tate twist, i.e. $$\log(t)W_k\Tors_R H^{i+1}(U;\ov{\mathcal L})\subseteq W_{k-2}\Tors_R H^{i+1}(U;\ov{\mathcal L}).$$ Theorem~\ref{thm:boundedWeights} implies: \[ \log(t)^{K}\Tors_R H^{i+1}(U;\ov{\mathcal L}) = \log(t)^{K}W_{\min\{2i,2n-2\}}\Tors_R H^{i+1}(U;\ov{\mathcal L})\subseteq\]\[ \subseteq W_{\min\{2i-2K,2n-2-2K\}}\Tors_R H^{i+1}(U;\ov{\mathcal L})\subseteq W_{i-1}\Tors_R H^{i+1}(U;\ov{\mathcal L})= 0. \] This shows that $\log(t)^{K}$ annihilates the whole module $\Tors_R H^{i+1}(U;\ov{\mathcal L})$. Since $\log(t)/(t-1)$ is a unit in $k[[t-1]]$, this implies that $(t-1)^{K}=0$ as well. Therefore, all its Jordan blocks have size at most $K$, as desired. The general case can be reduced to the unipotent case as described in Section \ref{sscover} through the following isomorphism of $R(N)$-modules, where $R(N) = k[t^{\pm N}]$ \[ \Tors_{R}H^{i+1}(U;{\mathcal L})=\Tors_{R(N)}H^{i+1}(U;{\mathcal L}) \cong\Tors_{R(N)}H^{i+1}(U_N;{\mathcal L}_N). \] The action of $t^N$ is then unipotent on $\Tors_{R(N)}H^{i+1}(U_N;{\mathcal L}_N) $, so it has Jordan blocks of size at most $K$, i.e. $(t^N-1)^{K}=0$. Therefore, the minimal polynomial of $t$ acting on $\Tors_{R}H^{i+1}(U;{\mathcal L})$ divides $(t^N-1)^K$, so the multiplicity of any root is at most $K$. \end{proof} Using the canonical isomorphisms of $R$-modules $$\Tors_R H^{i+1}(U;\ov{\mathcal L})\cong \Hom_k(\Tors_R H_i(U;{\mathcal L}),k)$$ (obtained from the composition $\Res^{-1} \circ \UCT_R$) and $\Tors_R H_i(U;{\mathcal L})\cong \Tors_R H_i(U^f;k)$, we obtain the analogous statement to Corollary \ref{cor:jordan} for the $t$ action by deck transformations on $\Tors_R H_i(U^f;k)$. In particular, if $i=1$, we obtain the following. \begin{cor}\label{cor:semisimple} If $U$ is an algebraic variety and $f\colon U\to \C^*$ is an algebraic map inducing epimorphisms on fundamental groups and inducing an infinite cyclic cover $U^f$, then the $t$-action on $H_1(U^f;k)$ is semisimple. \end{cor} \begin{remk}\label{remk:A1Semisimple} Let $f\colon \C^2\rightarrow \C$ be a polynomial function such that $f^{-1}(0)$ is reduced and connected. Let $U=\C^2\setminus f^{-1}(0)$, and we use the same letter $f$ to denote the induced map $f\colon U\rightarrow \C^*$. By \cite[Corollary 1.7]{DL}, the action of $t$ on $H_1(U^f;k)$ is semisimple. The corollary above generalizes this result, not only to algebraic varieties $U$ that are not affine connected curve complements, but also to connected curve complements in which the corresponding $f$ is given by a non-reduced polynomial, or even a rational function.$\sslash$ \end{remk} \begin{remk} The bound in Theorem~\ref{thm:boundedWeights} is sharp in the sense that for any $k\in [i,2i]\cap [i,2n-2]$, $\Gr^W_k \Tors_RH^{i+1}(U;\ov{\mathcal L})$ can be nonzero. If $U=F\times \C^*$, where $F$ is a smooth algebraic variety and $f$ is the projection, then $U^f\cong F\times \C$, which is homotopy equivalent to $F$. Therefore, it is straightforward to check that $t-1$ annihilates the Alexander module, and in this case the map $F\to U^f$ induces a MHS isomorphism in cohomology by Corollary~\ref{cor:fiber}. For example, we can let $F = E^{n_1}\times (\C^*)^{n_2}$, where $E$ is an elliptic curve. In this case, using the K\"unneth formula (which is a MHS morphism by \cite[Theorem 5.44]{peters2008mixed}), we will have: \[ \Gr^W_{k} H^{i}(F;k)\neq 0 \Leftrightarrow \left\{ \begin{array}{rl} \Gr^W_{2i-k} H^{2i-k}(E^{n_1};k) & \neq 0;\\ \Gr^W_{2k-2i} H^{k-i}((\C^*)^{n_2};k) &\neq 0 \end{array} \right\} \Leftrightarrow k \in [i,2i] \cap [2i-2n_1,n_2+i]. \] Varying $n_1,n_2\in \Z_{\ge 0}$ with $n_1+n_2 = \dim F = n-1$, one can have nonzero graded pieces for any weight $k\in [i,2i]\cap [i,2n-2]$.$\sslash$ \end{remk} \begin{proof}[Proof of Theorem~\ref{thm:boundedWeights}] First, we claim that it suffices to prove the theorem assuming that $t$ acts unipotently on $\Tors_R H^{i+1}(U;\ov{\mathcal L})$. For a general $U$, we let $N$ be such that $t^N$ acts unipotently on $\Tors_R H^{i+1}(U;\ov{\mathcal L})$ and we proceed as in Lemma~\ref{233} to obtain $U_N,f_N,{\mathcal L}_N$. By Remark~\ref{remk:mhsSummary}, there is a MHS isomorphism $\Tors_R H^{i+1}(U;\ov{\mathcal L}) \cong \Tors_R H^{i+1}(U_N;\ov{\mathcal L}_N)$. Therefore, if the statement of Theorem~\ref{thm:boundedWeights} holds for $(U_N,f_N)$, it will hold for $(U,f)$. From now on, we will assume that $U=U_N$, and that $t$ acts unipotently on $\Tors_R H^{i+1}(U;\ov{\mathcal L})$. Let $s=t-1\in R$, and let $R_m = \frac{R}{(s^m)}$. We work with local systems of $R$-modules on $U$. For such a local system ${\mathcal F}$, we will abbreviate ${\mathcal F}\otimes_R R_m$ to ${\mathcal F}/s^m$. We make $\ul k$ into a sheaf of $R$-modules by letting $s$ act as $0$. Recall the notation $\ov M$ from Remark~\ref{rem:conjugate}. Consider the following map of short exact sequences, with $\pi_{\ov{\mathcal L}}$ defined in Remark~\ref{rem:Rvee}. Later we will assume that $m$ is large enough that $s^m$ annihilates $\Tors_R H^{i}(U;\ov{\mathcal L})$ for all $i$. \begin{equation}\label{eq:twoSessBoundingWeights} { \begin{tikzcd}[row sep = 1.3em] 0\arrow[r] & \ov{\mathcal L}\arrow[r,"s"]\arrow[d,"\phi_m"] & \ov{\mathcal L}\arrow[r,"\pi_{\ov{\mathcal L}}"]\arrow[d,"\phi_{m+1}"] & \ul k \arrow[r]\arrow[d,"="]& 0 \\ 0\arrow[r] & \frac{\ov{\mathcal L}}{s^m}\arrow[r,"\psi_{m1}"] & \frac{\ov{\mathcal L}}{s^{m+1}}\arrow[r,"\phi_{m1}"]& \ul k \arrow[r] & 0. \end{tikzcd} } \end{equation} Here $\phi_j$ is induced by the projection $R\to R_j$ after tensoring with $\ov{\mathcal L}$. Since the vertical arrows are surjections, the arrows in the bottom row are uniquely determined as the arrows closing the diagram. By the discussion in Section~\ref{sec:maps}, concretely from (\ref{eq:picLTriangles}), the top row can be completed to an exact triangle with connecting map $-\pi_{{\mathcal L}}^\vee[1]\colon \ul k\to \ov{\mathcal L}[1]$. Let us see that the bottom row can be completed to an exact triangle with the connecting map $-(\phi_m\circ \pi_{{\mathcal L}}^\vee)[1]$. Recall that $\phi_m$ is the result of tensoring the quotient map $R\to R_m$ by $\ov{\mathcal L}$. Using the map $\phi_j$, $\overset{(-1)}{\ov{\mathcal L}}\xrightarrow{s^j}\overset{(0)}{\ov{\mathcal L}}$ becomes a free resolution of $\ov{\mathcal L}/s^j$, as in the following diagram: \[ { \begin{tikzcd}[row sep = 1.3em] \ov{\mathcal L} \arrow[d,"s^j"] \arrow[r]& 0 \arrow[d]\\ \ov{\mathcal L}\arrow[r,"\phi_j"] & \frac{\ov{\mathcal L}}{s^j}. \end{tikzcd} } \] Recall from the definition of $\pi_{{\mathcal L}}^\vee$ that $\phi_m\circ\pi_{{\mathcal L}}^\vee$ is given by the diagram below, on the left. Therefore, $-\phi_m\circ\pi_{{\mathcal L}}^\vee[1]$ is given by the middle diagram. We apply an isomorphism of resolutions which induces the identity on $\ul k$, we have $-\phi_m\circ\pi_{{\mathcal L}}^\vee[1]$ represented in a different way by the rightmost diagram: \[ \hspace{-0.25cm}\begin{tikzcd}[row sep = 1.3em] & 0 \arrow[r]\arrow[d]& \ov{\mathcal L}\arrow[d,"s^m"] & & 0 \arrow[r]\arrow[d]& \ov{\mathcal L}\arrow[d,"-s^m"]& & 0 \arrow[r]\arrow[d]& \ov{\mathcal L}\arrow[d,"-s^m"] \\ \deg = 0 \arrow[r,dashrightarrow] & \ov{\mathcal L} \arrow[r,"1"]\arrow[d,"s"]& \ov{\mathcal L}\arrow[d] & & \ov{\mathcal L} \arrow[r,"-1"]\arrow[d,"-s"]\arrow[rrr,bend right = 20,"-1"',dashrightarrow]& \ov{\mathcal L}\arrow[d] & & \ov{\mathcal L} \arrow[r,"1"]\arrow[d,"s"]& \ov{\mathcal L}\arrow[d] \\ & \ov{\mathcal L}\arrow[r] & 0 & \deg = 0\arrow[r,dashrightarrow] & \ov{\mathcal L}\arrow[r]\arrow[rrr,bend right = 20,"1"',dashrightarrow] & 0& & \ov{\mathcal L}\arrow[r] & 0 \end{tikzcd} \] We can check that the short exact sequence together with the connecting map can be lifted (uniquely up to homotopy) to the following diagram: \[ \begin{tikzcd}[column sep = 5em] \frac{\ov{\mathcal L}}{s^m} \arrow[r,"\psi_{m1}"]\arrow[d,phantom,"\rotatebox{-90}{$\cong$}"] & \frac{\ov{\mathcal L}}{s^{m+1}} \arrow[r,"\phi_{m1}"]\arrow[d,phantom,"\rotatebox{-90}{$\cong$}"] & \ul k \arrow[r,"-(\phi_m\circ \pi_{{\mathcal L}}^\vee){[1]}" ]\arrow[d,phantom,"\rotatebox{-90}{$\cong$}"] & \frac{\ov{\mathcal L}}{s^m} [1] \arrow[d,phantom,"\rotatebox{-90}{$\cong$}"] \\ 0\arrow[r]\arrow[d] & 0\arrow[r]\arrow[d] & 0\arrow[r]\arrow[d] & \ov{\mathcal L} \arrow[d,"-s^m"] \\ \ov{\mathcal L} \arrow[d,"s^m"]\arrow[r,"1"] & \ov{\mathcal L} \arrow[d,"s^{m+1}"] \arrow[r,"s^m"] & \ov {\mathcal L} \arrow[d,"s"] \arrow[r,"1"] & \ov{\mathcal L}\arrow[d] \\ \ov{\mathcal L} \arrow[r,"s"] & \ov{\mathcal L} \arrow[r,"1"] & \ov {\mathcal L} \arrow[r] & 0. \end{tikzcd} \] To see that this is an exact triangle we need to find a quasiisomorphism between the free resolution of $\ov{\mathcal L}/s^{m+1}$ and the cone $C:=\Cone(-\phi_m\circ \pi_{{\mathcal L}}^\vee\colon \ul k[-1]\to \ov{\mathcal L}/s^m)$, compatible with the maps coming from $\ov{\mathcal L}/s^m$ and going to $\ul k$ (i.e. the maps using the resolutions). Note that we are resolving $\ul k[-1]$ by $\ov{\mathcal L}\xrightarrow{-s} \ov{\mathcal L}$, the shift of the resolution above. Here is the quasiisomorphism: the map $\star$ is given by the matrix $\left(\begin{matrix} s^m & -1\\ 0 & s \end{matrix}\right)$. \[ \begin{tikzcd}[column sep = 5em] \ov{\mathcal L}/s^m \arrow[d,phantom,"\rotatebox{-90}{$\cong$}"] & C \arrow[d,phantom,"\rotatebox{-90}{$\cong$}"] & \ov{\mathcal L}/s^{m+1} \arrow[d,phantom,"\rotatebox{-90}{$\cong$}"] & \ul k \arrow[d,phantom,"\rotatebox{-90}{$\cong$}"] \\ \ov{\mathcal L} \arrow[d,"s^m"]\arrow[r,"(1{,}0)^t" & \ov{\mathcal L}\oplus \ov{\mathcal L} \arrow[d,"\star"] \arrow[r,"(1{,}0)"] \arrow[dr,phantom,"\cong"] & \ov{\mathcal L} \arrow[r,"s^m"]\arrow[d,"s^{m+1}"] & \ov{\mathcal L} \arrow[d,"s"]\\ \ov{\mathcal L} \arrow[r,"(1{,}0)^t"] & \ov{\mathcal L}\oplus \ov{\mathcal L} \arrow[r,"(s{,}1)" & \ov{\mathcal L} \arrow[r,"1"] & \ov{\mathcal L} \end{tikzcd} \] From this diagram, we automatically see that the quasiisomorphism is compatible with the map from $\ov{\mathcal L}_m$. For the map to $\ul k$, the map $C\to \ul k$ above should coincide with the projection on the first coordinate, which is realized by the following homotopy: \[ \begin{tikzcd}[column sep = 8em, row sep =3.5em] \ov{\mathcal L}\oplus \ov{\mathcal L} \arrow[r,"(s^m{,}0)",shift left = 1.5ex]\arrow[r,"(0{,}1)"', shift right = -0.5 ex]\arrow[d,"\star"] & \ov{\mathcal L} \arrow[d,"s"] \\ \ov{\mathcal L}\oplus \ov{\mathcal L} \arrow[r,"(s{,}1)",shift left = 0.5ex]\arrow[r,"(0{,}1)"',shift right = 0.5ex]\arrow[ur,"(-1{,}0)"] & \ov{\mathcal L} \end{tikzcd} \] So the short exact sequences in (\ref{eq:twoSessBoundingWeights}) can be filled in to exact triangles in the derived category, using the maps $-\pi_{{\mathcal L}}^\vee[1]\colon \ul k\to \ov{\mathcal L}$ and $-(\phi_m\circ \pi_{{\mathcal L}}^\vee)[1]\colon \ul k\to \ov{\mathcal L}/s^m$. Note that clearly, these will create a map between triangles. Therefore, we have the following map of cohomology long exact sequences, given in the two bottom rows below \begin{equation}\label{eq:bigExactSeqs} \begin{tikzcd} H^i(U;k)\arrow[r,"\pi_{{\mathcal L}}^\vee"]\arrow[d,"="] & \Tors_R H^{i+1}(U;\ov{\mathcal L}) \arrow[r,"s"]\arrow[d,hookrightarrow] & \Tors_R H^{i+1}(U;\ov{\mathcal L}) \arrow[r,"\pi_{\ov{\mathcal L}}"]\arrow[d,hookrightarrow] & H^{i+1}(U;k) \arrow[d,"="]\\ H^i(U;k)\arrow[r,"\pi_{{\mathcal L}}^\vee"]\arrow[d,"="] & H^{i+1}(U;\ov{\mathcal L}) \arrow[r,"s"]\arrow[d,"\phi_m"] & H^{i+1}(U;\ov{\mathcal L}) \arrow[r,"\pi_{\ov{\mathcal L}}"]\arrow[d,"\phi_{m+1}"] & H^{i+1}(U;k)\arrow[d,"="] \\ H^i(U;k)\arrow[r,"\phi_m\circ \pi_{{\mathcal L}}^\vee"] & H^{i+1}(U;\frac{\ov{\mathcal L}}{s^m}) \arrow[r,"s"] & H^{i+1}(U;\frac{\ov{\mathcal L}}{s^{m+1}}) \arrow[r,"\phi_{m1}"] & H^{i+1}(U;k) . \end{tikzcd} \end{equation} We are abusing notation by using the same letters to denote the maps of (the derived category of) sheaves and their induced maps in cohomology. From now on, assume that $m$ is large enough that $s^m$ annihilates $\Tors_R H^{i+1}(U;\ov{\mathcal L})$. By Remark~\ref{remk:torsionForSheaves}, the maps $\Tors_R H^*(U;{\mathcal L})\rightarrow H^*(U;{\mathcal L}/s^m)$ induced by $\phi_m$ (the composition of the two central vertical arrows above) are injective. Also, note that on $\Tors_R H^{i+1}(U;\ov{\mathcal L})$, $s$ and $\log(t) = \log(1+s)$ differ by a unit, namely the Taylor polynomial of degree $m-1$ of $\frac{s}{\log(1+s)}$, and therefore their kernel and cokernel coincide. This means that the bottom row in the diagram above will remain exact if we replace $s$ by $\log(t)$. We have the following commutative diagram, formed by the first and last rows of equation (\ref{eq:bigExactSeqs}). \begin{equation}\label{eq:kerOfLogS} \begin{tikzcd} H^i(U;k)\arrow[r,"\pi_{{\mathcal L}}^\vee"]\arrow[d,"="] & \Tors_R H^{i+1}(U;\ov{\mathcal L}) \arrow[r,"\log t"]\arrow[d,"\phi_m",hookrightarrow] & \Tors_R H^{i+1}(U;\ov{\mathcal L})(-1) \arrow[r,"\pi_{\ov{\mathcal L}}"]\arrow[d,"\phi_{m+1}",hookrightarrow] & H^{i+1}(U;k) \arrow[d,"="]\\ H^i(U;k)\arrow[r,"\pi_{{\mathcal L}}^\vee"] & H^{i+1}(U;\frac{\ov{\mathcal L}}{s^m}) \arrow[r,"\log t"] & H^{i+1}(U;\frac{\ov{\mathcal L}}{s^{m+1}})(-1)\arrow[r,"\phi_{m1}"] & H^{i+1}(U;k) . \end{tikzcd} \end{equation} We make the following claims: \begin{enumerate} \item $\pi_{{\mathcal L}}^\vee$ is a MHS morphism (Theorem~\ref{thm:geomIsMHS} together with Proposition~\ref{prop:mapsAreEqual}). \item $\log t$ is a MHS morphism (note the Tate twist), by Corollary~\ref{alexandermhs}. \item $\phi_m$ and $\phi_{m+1}$ are MHS morphisms (Remark~\ref{remk:mhsSummary}). \item $\phi_{m1}$ is a MHS morphism. This follows from Lemma~\ref{inducedmhsmaps}. The Tate twist is due to the fact that the MHS on the cohomology $\ov{\mathcal L}/s$ comes from the shifted Hodge-de Rham complex, as in Remark~\ref{remk:mhsSummary}. Remark~\ref{transvstate} tells us that shifts of mixed Hodge complexes result in Tate twists in their cohomologies. \item It follows that $\pi_{\ov{\mathcal L}} = \phi_{m1}\circ \phi_{m+1}$ is a MHS morphism. \item The top row is exact. This is a straightforward verification. \end{enumerate} We have now arrived at the following exact sequence of mixed Hodge structures: \[ H^i(U;k) \xrightarrow{\pi_{{\mathcal L}}^\vee} \Tors_R H^{i+1}(U;\ov{\mathcal L}) \xrightarrow{\log(t)} \Tors_R H^{i+1}(U;\ov{\mathcal L})(-1) \xrightarrow{\pi_{\ov{\mathcal L}}} H^{i+1}(U;k). \] Taking the associated graded sequence for the weight filtration, we obtain the following exact sequence of pure Hodge structures for any $k$: \[ \Gr^W_k H^i(U;k) \xrightarrow{\pi_{{\mathcal L}}^\vee} \Gr^W_k \Tors_R H^{i+1}(U;\ov{\mathcal L}) \xrightarrow{\log(t)} \Gr^W_{k-2} \Tors_R H^{i+1}(U;\ov{\mathcal L}) \xrightarrow{\pi_{\ov{\mathcal L}}} \Gr^W_k H^{i+1}(U;k). \] By \cite[Corollaire 3.2.15]{De2}, the nonzero graded pieces of the weight filtration of $H^i(U;k)$ are in the interval $[i,\min\{2i,2n\}]$. Therefore, if either $k\ge \min\{2i+3,2n+1\}$ or $k\le i-1$, we have an isomorphism \[ \Gr^W_k \Tors_R H^{i+1}(U;\ov{\mathcal L}) \overset{\log(t)}{\cong} \left(\Gr^W_{k-2} \Tors_R H^{i+1}(U;\ov{\mathcal L})\right)(-1). \] Since the MHS are finite dimensional, these groups must be $0$ for $k\ll 0$ and $k\gg 0$, which implies that \[ \Gr^W_k \Tors_R H^{i+1}(U;\ov{\mathcal L}) = 0 \text{ unless } k\in [i,\min\{2i,2n-2\}]. \] \end{proof} \section{Semisimplicity for proper maps}\label{sect:ss} Let $k=\Q$. In this section, assuming that $f\colon U \to \C^*$ is proper, we will prove that the torsion part $A_i(U^f; \Q)$ of the Alexander module $H_i(U^f; \Q)$ is a semisimple $R$-module, that is, the $t$-action on $A_i(U^f; \Q)$ is semisimple, for all $i \geq 0$. \begin{thm}\label{thmsimple} Let $U$ be a smooth complex algebraic variety, and let $f\colon U\to \C^*$ be a proper algebraic map. Let $\calL$ be the rank one $R=\Q[t^{\pm 1}]$-local system on $U$ defined as in Section \ref{ssAlex}. Then $\Tors_R {H^i(U;\calL)}$ is a semisimple $R$-module. \end{thm} Since the operations of taking the conjugate $R$-module structure and taking the $\Q$-vector space dual preserve semisimplicity, the following is an immediate consequence of Proposition~\ref{propcanon} and the above theorem. \begin{cor}\label{corss} Under the assumptions of Theorem \ref{thmsimple}, the torsion part $A_i(U^f; \Q)$ of the homology Alexander module $H_i(U^f; \Q)$ is a semisimple $R$-module, for all $i \geq 0$. \end{cor} In the rest of this section, we prove Theorem \ref{thmsimple}. Let $\calL_{\C^*}$ be the tautological rank one $R$-local system on $\C^*$. Under the natural isomorphism $R\cong \Q[\pi_1(\C^*)]$, the monodromy action is the natural $\pi_1(\C^*)$-action on $\Q[\pi_1(\C^*)]$. Then $\calL\cong f^{-1}(\calL_{\C^*})$ as $R$-local systems. By the projection formula, we have \[ H^i(U;\calL)\cong H^i(U; f^{-1}(\calL_{\C^*}))\cong \mathbb{H}^i(\C^*, Rf_*(\Q_U)\otimes_\Q \calL_{\C^*}). \] Since $U$ is smooth and $f$ is a proper map, the decomposition theorem of \cite{BBD} yields that the push-forward $Rf_*(\Q_U)$ decomposes as \[ Rf_*(\Q_U)\cong \bigoplus_{\lambda\in \Lambda}\mathcal{P}_\lambda[d_\lambda], \] where $\Lambda$ is a finite index set, $d_\lambda\in \Z$, and each $\mathcal{P}_\lambda$ is a simple $\Q$-perverse sheaf on $\C^*$. Therefore, to prove Theorem \ref{thmsimple}, it suffices to show the following proposition, which is essentially a special case of the theory of the Mellin transformation developed by Gabber-Loeser \cite{GL}. \begin{prop}\label{propMellin} If $\mathcal{P}$ is a simple $\Q$-perverse sheaf on $\C^*$, we have: \begin{enumerate} \item $\mathbb{H}^i(\C^*; \mathcal{P}\otimes \calL_{\C^*})=0$ for $i\neq 0$; \item $\mathbb{H}^0(\C^*; \mathcal{P}\otimes \calL_{\C^*})$ is a simple $R$-module when $\mathcal{P}$ is smooth (i.e., the shift of a local system on $\C^*$), and a free $R$-module when $\mathcal{P}$ is not smooth. \end{enumerate} \end{prop} Here, we recall that an $R$-module is {\em simple} if it does not contain any proper nonzero $R$-submodule. Since $R=\Q[t^{\pm 1}]$, an $R$-module is simple if and only if it is a finite dimensional $\Q$-vector space and it is a simple $\Q$-representation of $\Z$. Before proving the above proposition, we need a few lemmas. \begin{lem}\label{lemsmooth} Let $\mathcal{P}$ be a simple $\Q$-perverse sheaf on $\C^*$, and let $\Q\subset k$ be a field extension. Then $\mathcal{P}\otimes_\Q k$ is a semisimple $k$-perverse sheaf. Moreover, $\mathcal{P}$ is smooth if and only if one of the simple summands of $\mathcal{P}\otimes_\Q k$ is smooth. \end{lem} \begin{proof} Since $\mathcal{P}$ is a simple perverse sheaf, it is either a skyscraper sheaf or it is isomorphic (after a shift of degree one) to the intermediate extension of a simple local system on a Zariski open subset of $\C^*$. After a field extension, a simple local system becomes at worst semisimple. Since taking intermediate extension commutes with taking field extension, the first half of the lemma follows. The ``only if" part of the second half is obvious. To show the ``if" part, we first assume that the field extension $\Q\subset k$ is Galois. Then its Galois group acts transitively on the simple summands of $\mathcal{P}\otimes_\Q k$. Thus, in this case, the second half of the lemma follows. In general, we can take a further field extension $\Q\subset k\subset k'$, such that $\Q\subset k'$ is Galois. Then the second part holds for $\mathcal{P}\otimes_\Q k'$, which implies that it also holds for $\mathcal{P}\otimes_\Q k$. \end{proof} \begin{lem}\label{lemH1} Let $k$ be a field, and let $\mathcal{P}_k$ be a simple $k$-perverse sheaf on $\C^*$. If $\mathbb{H}^{-1}(\C^*; \mathcal{P}_k)\neq 0$, then $\mathcal{P}_k\cong k_{\C^*}[1]$. \end{lem} \begin{proof} By the natural isomorphism \[ \mathbb{H}^{-1}(\C^*; \mathcal{P}_k)\cong \Hom(k_{\C^*}[1], \mathcal{P}_k), \] if $\mathbb{H}^{-1}(\C^*; \mathcal{P}_k)\neq 0$, then there exists a nontrivial map between $k_{\C^*}[1]$ and $\mathcal{P}_k$. Since both $k_{\C^*}[1]$ and $\mathcal{P}_k$ are simple perverse sheaves, it follows that they are isomorphic. \end{proof} \begin{proof}[Proof of Proposition \ref{propMellin}] \textbf{Case 1.} Assume $\mathcal{P}$ is smooth. In this case, $\mathcal{P}\cong L[1]$ for a simple $\Q$-local system $L$ on $\C^*$. As in \cite[Example 2.5.7]{dimca2004sheaves}, we have natural isomorphisms of $R$-modules \[ \mathbb{H}^i(\C^*; \mathcal{P}\otimes \calL_{\C^*})\cong H^{i+1}(\C^*; L\otimes_\Q \calL_{\C^*})\cong \begin{cases} 0 \quad &\text{if}\quad i\neq 0 \\ \overline{V_L}\quad &\text{if} \quad i=0, \end{cases} \] where $V_L$ is the $R$-module associated to the monodromy representation of $L$, and $\overline{\cdot}$ denotes the conjugate $R$-module structure, with $t$ acting by $t^{-1}$ (see Remark \ref{rem:conjugate}). By the equivalence of categories between $\Q$-local systems on $\C^*$ and $\Q[\Z]$-modules, the local system $L$ is simple if and only if $V_L$ is simple. Evidently, $V_L$ is simple if and only if $\overline{V_L}$ is simple. Thus, in this case, the proposition follows. \noindent\textbf{Case 2.} Assume $\mathcal{P}$ is not smooth. Let $p: \C^*\to \mathrm{pt}$ be the projection to a point. We claim that there is a natural isomorphism \begin{equation}\label{eqpf} Rp_*(\mathcal{P}\otimes_{\Q} \calL_{\C^*})\stackrel{L}{\otimes}_R R/\mathfrak{m}\cong Rp_*(\mathcal{P}\otimes_{\Q} \calL_{\C^*}\otimes_R R/\mathfrak{m}), \end{equation} where $\mathfrak{m}$ is a maximal ideal of $R$ and $\stackrel{L}{\otimes}$ denotes the derived tensor product. In fact, take a complex of injective sheaves $\mathcal{I}^\bullet$ on $\C^*$ representing $\mathcal{P}\otimes_{\Q}\calL_{\C^*}$ and take a free resolution $F^\bullet$ of the $R$-module $R/\mathfrak{m}$. Then the total complex of $\Gamma(\C^*, \mathcal{I}^\bullet)\otimes_R F^\bullet$ is a complex representing both sides of \eqref{eqpf}. Let $k\coloneqq R/\mathfrak{m}$ and let $L_k\coloneqq \calL_{\C^*}\otimes_R R/\mathfrak{m}$. Then $L_k$ is a rank one $k$-local system on $\C^*$. By Lemma \ref{lemsmooth}, \[\mathcal{P}\otimes_{\Q} L_k\cong (\mathcal{P}\otimes_{\Q} k)\otimes_k L_k \] is a semisimple $k$-perverse sheaf on $\C^*$. By the support condition for perverse sheaves we have that $\mathcal{H}^i(\mathcal{P}\otimes_{\Q} L_k)=0$ for $i<-1$. Thus, the hypercohomology spectral sequence yields that $\mathbb{H}^i(\C^*; \mathcal{P}\otimes_{\Q} L_k)=0$ for $i<-1$. Additionally, by Lemma \ref{lemH1}, $\mathbb{H}^i(\mathcal{P}\otimes_{\Q} L_k)=0$ for $i\leq -1$. On the other hand, we get by Artin's vanishing theorem that $\mathbb{H}^i(\C^*; \mathcal{P}\otimes_{\Q} L_k)=0$ for $i\geq 1$. Therefore, \[ \dim_k \mathbb{H}^0(\C^*; \mathcal{P}\otimes_{\Q} L_k)=\chi(\C^*; \mathcal{P}) \] is independent of the choice of the maximal ideal $\mathfrak{m}$ of $R$. By the isomorphism \eqref{eqpf} and the universal coefficient theorem for cochain complexes, we have the following short exact sequence \[ 0\to \mathbb{H}^i(\C^*; \mathcal{P}\otimes_{\Q} \calL_{\C^*})\otimes_R k\to \mathbb{H}^i(\C^*; \mathcal{P}\otimes_{\Q} L_k)\to \mathrm{Tor}^R_1(\mathbb{H}^{i+1}(\C^*; \mathcal{P}\otimes_{\Q} \calL_{\C^*}), k) \to 0. \] The results in the previous paragraph then yield that, for any maximal ideal $\mathfrak{m}$ of $R$, we have \[ \mathbb{H}^i(\C^*; \mathcal{P}\otimes_{\Q} \calL_{\C^*})\otimes_R R/\mathfrak{m}=0 \quad \text{for all } i\neq 0, \] and \[ \mathrm{Tor}^R_1(\mathbb{H}^{i}(\C^*; \mathcal{P}\otimes_{\Q} \calL_{\C^*}), R/\mathfrak{m})=0 \quad \text{for all } i\neq 1. \] Thus, $\mathbb{H}^i(\C^*; \mathcal{P}\otimes_{\Q} \calL_{\C^*})=0$ for all $i\neq 0$ and $\mathbb{H}^0(\C^*; \mathcal{P}\otimes_{\Q} \calL_{\C^*})$ is free. \end{proof} As a consequence of Corollary \ref{corss}, we get the following purity result (Theorem~\ref{thm:pur} in the introduction). \begin{cor}\label{cor:pure} If $f\colon U\to \C^*$ is a proper algebraic map, then $A_i(U^f; \Q)$ carries a pure Hodge structure of weight $-i$. \end{cor} \begin{proof} The generic fiber $F$ of the proper morphism $f\colon U\to \C^*$ between smooth complex algebraic varieties is a complete smooth algebraic variety, whose rational homology and cohomology groups have pure Hodge structures. By Corollary \ref{corss} and Corollory \ref{cor:fiber}, $A_i(U^f; \Q)$ is a quotient of the weight $-i$ pure Hodge structure on $H_i(F; \Q)$, which proves our claim. \end{proof} \begin{remk}\label{rem:smf} If $f\colon U \to \C^*$ is a projective submersion of smooth complex algebraic varieties, then $f$ is a fibration, and let $F$ denote its fiber. In this case, the semisimplicity of $A_i(U^f;\Q)=H_i(U^f;\Q)\cong H_i(F;\Q)$ is a direct consequence of Deligne's decomposition theorem \cite{De, De2} (compare with \cite[Remark 1.12]{BudurLiuWang}). Indeed, Deligne's theorem implies that the local systems $R^if_*\Q_U$ are semisimple on $\C^*$, or equivalently, the monodromy representation on $H^i(F;\Q)$ is semisimple. The semisimplicity of $A_i(U^f;\Q)$ follows then by using Lemma \ref{lem:fiberMonodromy}.$\sslash$ \end{remk} \section{Relation to the Limit MHS}\label{rellim} In this section, we compare the mixed Hodge structure on $\Tors_R H^*(U;\ov\calL)$ with the limit mixed Hodge structure on the generic fiber of $f$ (as recalled in Section \ref{LimitMixedHodgeStructure}). \textit{Throughout this subsection assume the ground field $k = \Q$, that the algebraic map $f\colon U \rightarrow \C^*$ is in addition a proper map with generic fiber $F$, and that the good compactification $\bar{f}\colon X \rightarrow \C P^1$ of $f$ has the property that $\bar{f}^{-1}(0)$ is reduced.} Let $i\colon E \rightarrow X$ denote the inclusion of the reduced divisor $\bar{f}^{-1}(0)$. \begin{remk}\label{unipotent} Under the above assumptions, the monodromy of $\Tors_R H^*(U,\ov\calL)$ is unipotent, see \cite[Corollary 11.19]{peters2008mixed}.$\sslash$ \end{remk} \begin{remk}\label{proper} Let $f\colon U\rightarrow \C^*$ be a proper map. Recall the notations of the beginning of Section \ref{sscover}. By \cite[Semi-stable Reduction Theorem]{kempf2006toroidal}, there is a finite cover $U_N\rightarrow U$ for some $N$ such that $f_N\colon U_N\rightarrow \C^*$ is also a proper map, and there is a compactification $\bar{f_N}\colon X \rightarrow \C P^1$ satisfying the conditions that we assume throughout this subsection. Further, by loc. cit. $X\setminus U_N = (\ov f_N)^{-1}(\{0,\infty\})$. Note that the eigenvalues of the action of $t^N$ in $\Tors_{R(N)}H^*(U_N,\ov\calL_N)$ will be all $1$ by Remark \ref{unipotent}. Recall also that the way we induced $\Tors_R H^*(U;\ov\calL)$ with a MHS was using the natural isomorphism of $R(N)$-modules $\Tors_R H^*(U;\ov\calL)\cong \Tors_{R(N)}H^*(U_N,\ov\calL_N)$ of Lemma \ref{lemLocal}. Hence, the only meaningful conditions we are assuming in this subsection is that the map $f\colon U\rightarrow\C^*$ be \textit{proper}.$\sslash$ \end{remk} We wish to relate the mixed Hodge structure on $\mathrm{Tors}_R\, H^*(U; \ov\calL)$ with the limit mixed Hodge structure on $\H^*(E; \psi_{\bar{f}} \underline{\Q})$ discussed in Section \ref{LimitMixedHodgeStructure}. In fact, we will identify morphisms of mixed Hodge complexes of sheaves that realize this relationship. We compare the subsequent pair of mixed Hodge complexes. On the one hand, Theorem \ref{Qalexandermhs} provides us for $m \geq 1$: \begin{align*} \Hdg{X}{D}&\left(\frac{1}{2\pi i}\frac{df}{f},m\right)\\ &= \left(\left[\calK^\bullet_\infty\left(1 \otimes f,m\right), \tilde{W}_{\lc}\right], \left[\logdr{X}{D}\left(\frac{1}{2\pi i}\frac{df}{f},m\right), W_{\lc}, F^{\lc}\right], \varphi_{\infty\#}\right) \end{align*} On the other hand, Theorem \ref{rationallimitmhs} offers us, independent of $m$: \begin{align*} \psi_f^{\textnormal{Hdg}} = i^{-1}\left(\left[\Tot\tilde{\calC}^{\bullet,\bullet}, \tilde{W}(M)_{\lc}\right], \left[\vphantom{\tilde{A}} \Tot\calA^{\bullet,\bullet}, W(M)_{\lc}, F^{\lc}\right], \varphi_{\infty}\right). \end{align*} Our first goal is to show that, pre-application of $i^{-1}$, the second triple is a quotient of a translation of the first, as long as $m$ is large enough. We obtain our map of triples from the quotient maps (up to a sign): \begin{align*} &(-1)^\ell\Phi^\ell_\Q(m)\colon \bigoplus_{j=0}^{m-1} \calK^{\ell+1}_\infty \otimes \Q \langle s^j \rangle \rightarrow \bigoplus_{j=0}^{m-1} \calK^{\ell+1}_\infty/\tilde{W}_j \calK^{\ell+1}_\infty\\ &(-1)^\ell\Phi^\ell_\C(m)\colon \bigoplus_{j=0}^{m-1} \Omega_X^{\ell+1}(\log D) \otimes_\C \C \langle s^j \rangle \rightarrow \bigoplus_{j=0}^{m-1} \Omega_X^{\ell+1}(\log D)/ W_j \Omega_X^{\ell+1}(\log D) \end{align*} where $\ell$ and $m \geq 1$ denote integers. We define $\Phi^\ell_\Q(\infty),\Phi^\ell_\C(\infty)$ using the same formulas. If $m \geq \dim X$ then the codomains of the above maps are $(\Tot \tilde{\calC}^{\bullet,\bullet})^\ell$ and $(\Tot\calA^{\bullet, \bullet})^\ell$ respectively, because the filtrations $\tilde{W}_{\lc}$ and $W_{\lc}$ fill out at index $\dim X$ (this is where boundedness of filtrations is important). This enables us to, assuming $m \geq \dim X$ (or even $m=\infty$), define natural surjections \[\Phi_\Q(m)\colon \calK^\bullet_\infty(1 \otimes f, m)[1] \rightarrow \Tot\tilde{\calC}^{\bullet, \bullet} \] and \[ \Phi_\C(m)\colon \logdr{X}{D}\left(\frac{1}{2 \pi i} \frac{df}{f}, m\right)[1] \rightarrow \Tot\calA^{\bullet, \bullet}. \] The compatibility with differentials $\Phi(m) \circ d[1] = (d'+d'') \circ \Phi(m)$ follows by definition, where $d'$ and $d''$ are the differentials on the double complexes. \begin{lem}\label{surj} Suppose $m \geq \dim X$. Then: \begin{align*} \Phi(m)\colon \Hdg{X}{D}&\left(\frac{1}{2\pi i}\frac{df}{f},m\right)[1] \rightarrow \left(\left[\Tot\tilde{\calC}^{\bullet,\bullet}, \tilde{W}(M)_{\lc}\right], \left[\vphantom{\tilde{A}} \Tot\calA^{\bullet,\bullet}, W(M)_{\lc}, F^{\lc}\right], \varphi_{\infty}\right) \end{align*} is a (surjective) map of triples. \end{lem} \begin{proof} The commutativity $\Phi_\C(m) \circ \varphi_{\infty\#} = \varphi_\infty \circ \Phi_\Q(m)$ follows from the definitions. We next verify that filtrations are preserved for the $\C$-component of the triple. The same arguments apply to the $\Q$-component as well. The $W[1]_{\lc}$-filtered subcomplex indexed by integer $i$ is: \begin{align*} \left[W_{i+1}\logdr{X}{D}\left(\frac{1}{2\pi i}\frac{df}{f},m\right)\right][1] = \bigoplus_{j = 0}^{m-1} \left[\vphantom{\tilde{A}}W_{i+2j+1} \Omega^{\bullet}_X(\log D)\right][1] \otimes_\C \C\langle s^j \rangle \end{align*} which is mapped under $\Phi_\C(m)$ into $W(M)_i \left(\Tot\calA^{\bullet,\bullet}\right)$. The $F[1]^{\lc}$-filtered subcomplex indexed by integer $p$ is: \begin{align*} \left[F^{p+1}\logdr{X}{D}\left(\frac{1}{2\pi i}\frac{df}{f},m\right) \right][1] &= \bigoplus_{j=0}^{m-1}\left[F^{p+j+1}\logdr{X}{D}\right][1] \otimes_\C \C \langle s^j \rangle\\ &= \bigoplus_{j=0}^{m-1}\left[\Omega^{\geq p + j + 1}_X (\log D)\right][1] \otimes_\C \C \langle s^j \rangle \end{align*} which is mapped under $\Phi_\C(m)$ into $F^p \left(\Tot\calA^{\bullet,\bullet}\right)$. \end{proof} The adjunction $\Id\rightarrow i_*i^{-1}$ applied to $\left(\left[\Tot\tilde{\calC}^{\bullet,\bullet}, \tilde{W}(M)_{\lc}\right], \left[\vphantom{\tilde{A}} \Tot\calA^{\bullet,\bullet}, W(M)_{\lc}, F^{\lc}\right], \varphi_{\infty}\right)$ gives us a surjective map of triples $$\Ad\colon \left(\left[\Tot\tilde{\calC}^{\bullet,\bullet}, \tilde{W}(M)_{\lc}\right], \left[\vphantom{\tilde{A}} \Tot\calA^{\bullet,\bullet}, W(M)_{\lc}, F^{\lc}\right], \varphi_{\infty}\right)\rightarrow i_*\psi_{\bar{f}}^{\textnormal{Hdg}}$$ Note that, since $i$ is proper, $i_*=i_!$ is an exact functor, so we may identify $i_*$ with $Ri_*$. Composing $\Phi(m)$ with the adjunction $\Ad$, we immediately get the following. \begin{cor}\label{relnlimitmhc} Suppose $m \geq \dim X$. Then: \begin{align*} \Ad\circ\Phi(m)\colon \Hdg{X}{D}&\left(\frac{1}{2\pi i}\frac{df}{f},m\right)[1] \rightarrow i_*\psi_{\bar{f}}^{\textnormal{Hdg}} \end{align*} is a surjective morphism of mixed Hodge complexes of sheaves. It satisfies the equality: \begin{align*} \Phi(m) \circ S_m[1] = \Theta \circ \Phi(m) \end{align*} where $S_m$ is multiplication by $s$ (as in Lemma \ref{inducedmhsmaps}), and $\Theta$ is defined in Section \ref{LimitMixedHodgeStructure}. \end{cor} \begin{proof} The equality is obtainable by expanding definitions. \end{proof} Recall that we defined the MHS on $H^{*+1}(U; \ov\calL \otimes R_m)$ using the $[1]$ shift of the thickened Hodge complex $\Hdg{X}{D}\left(\frac{1}{2\pi i}\frac{df}{f},m\right)_{\widetilde{s}_2}$. Twisting the $R$-module structure and taking global hypercohomology in Corollary \ref{relnlimitmhc}, we get: \begin{cor}\label{relnlimitmhs} For sufficiently large $m$, $$\Ad\circ\Phi(m)\colon \Hdg{X}{D}\left(\frac{1}{2\pi i}\frac{df}{f},m\right)_{\widetilde{s}_2}[1] \rightarrow i_*\psi_{\bar{f}}^{\textnormal{Hdg}}$$ induces a morphism of $\Q$-mixed Hodge structures: \begin{align*} \Phi^*\colon \mathrm{Tors}_R H^{*+1}(U;\ov\calL) \rightarrow \mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q}) \end{align*} which is also an $R$-module homomorphism. \end{cor} \begin{proof} Let $k=\Q$, and $m\geq \dim X$. Since $\Theta$ corresponds to multiplication by $\log(t)$, and multiplication by $s$ on $\Hdg{X}{D}\left(\frac{1}{2\pi i}\frac{df}{f},m\right)$ corresponds to multiplication by $\log(t)$ on $\Hdg{X}{D}\left(\frac{1}{2\pi i}\frac{df}{f},m\right)_{\widetilde{s}_2}$, Corollary \ref{relnlimitmhc} tells us that \[\Ad\circ\Phi(m)\colon \Hdg{X}{D}\left(\frac{1}{2\pi i}\frac{df}{f},m\right)_{\widetilde{s}_2}[1] \rightarrow i_*\psi_{\bar{f}}^{\textnormal{Hdg}}\] induces the following morphism of mixed Hodge structures and $R$-modules in hypercohomology: \begin{align*} H^{*+1}(U;\ov\calL \otimes_R R_m) \rightarrow \mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q}). \end{align*} which we denote by $\Phi(m)^*$. The torsion submodule $\Tors_{R_\infty}H^{*+1}(U; \ov\calL \otimes_R R_\infty)$ is contained in $H^{*+1}(U;\ov\calL \otimes_R R_m)$ as the kernel of $\psi_{mj}[1]^*$ for sufficiently large $m$ and $j$ (Remark \ref{remQ}). And because the monodromy action is unipotent (Remark \ref{unipotent}), the torsion submodule is isomorphic to $\Tors_R H^{*+1}(U; \ov\calL)$ as $R$-modules. Stitching these observations together, for sufficiently large $m$ the morphism $\Ad\circ\Phi(m)$ induces an $R$-module morphism: \begin{align*} \Phi^*\colon \mathrm{Tors}_R\,H^{*+1}(U;\ov\calL) \rightarrow \mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q}) \end{align*} Recall Remark \ref{torsionrelations} and the proof of Corollary \ref{torsionmhs}, where we see that the MHS constructed on $\mathrm{Tors}_R\,H^{*+1}(U;\ov\calL)$ is independent of $m$, for $m$ sufficiently large. Independence of the map $\Phi^*$ of large enough $m$ is determinable through examination of the maps between the various thickened complexes associated to their inverse limit structure. \end{proof} Let $D$ and $D^*$ be an open disk and a punctured open disc centered at $0$ in $\C$, res\-pec\-tive\-ly. Let $T\coloneqq (\overline f)^{-1}(D)\subset X$ and $T^*\coloneqq f^{-1}(D^*)\subset U$. Assume further that $D$ is sufficiently small such that $$f_{|T^*}\colon T^*\rightarrow D^*$$ is a fibration. Note that by Remark~\ref{proper}, we can assume that $X\backslash U$ contains only vertical divisors, so $T\backslash E=T^*$. Recall that by definition, $\psi_{\ov f}=i^{-1} \circ R(j\circ \pi)_*\circ (j\circ \pi)^{-1} $, where, abusing notation, $\pi$ is seen as a map $\pi\colon (T^*)^f\rightarrow T^*$, $j$ is seen as $j:T^*\hookrightarrow T$, and $i$ is seen as $i:E\hookrightarrow T$. We will use this notation in the rest of this section. In the following definition, we describe a map which has the same domain and target as the MHS morphism $\Phi^*$ of Corollary \ref{relnlimitmhs}, but is defined explicitely in cohomology and in a more geometric and down-to-earth way. \begin{dfn}\label{def:relLimit} The map $r^*:\Tors_R H^{*+1}(U;\ov{\mathcal L})\hookrightarrow \mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q})$ is defined as the following composition $$ \Tors_R H^{*+1}(U;\ov{\mathcal L})\hookrightarrow H^{*+1}(T^*;\ov{\mathcal L})\cong H^{*}((T^*)^f;\Q)\xrightarrow{\cong} \mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q}). $$ Here, the first map is given by the restriction from $U$ to $T^*$ (hence the name $r^*$), and was shown to be injective in Remark \ref{remk:tube}. The isomorphism $H^{*+1}(T^*;\ov{\mathcal L})\cong H^{*}((T^*)^f;\Q)$ is given by Corollary \ref{isocohom} and was already described in Remark \ref{remk:tube}. Finally, the map $H^{*}((T^*)^f;\Q)\to \mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q})$ is given by the adjunction $\Id\to i_*i^{-1}$ applied to $R(j\circ \pi)_*(j\circ\pi)^{-1}\underline \Q_T=R(j\circ \pi)_*\underline \Q_{(T^*)^f}$. The map $H^{*}((T^*)^f;\Q)\to \mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q})$ is an isomorphism because by \cite[Remark 2.6.9]{KS}, \[ \mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q}) \cong \varinjlim_{V\supseteq E} \mathbb{H}^*(V; R(j\circ \pi)_*\underline{\Q}_{(T^*)^f}), \] where the limit is taken over open sets $V$. Since $f$ is proper, every such $V$ contains an open set of the form $T=(\ov f)^{-1}(D)$, for $D$ a small enough disk around $0$. The isomorphism follows from the fact that all sufficiently small tubes $T$ are fibrations over the disk, so they are homotopy equivalent to each other. \end{dfn} Note that, up to the natural identification $H^{*}((T^*)^f;\Q)\cong \mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q})$ given by adjunction and the natural identification $(\Tors_R H_*(U^f;\Q))^{\vee_{\Q}}\cong \Tors_R H^{*+1}(U;\ov{\mathcal L})$ given in Proposition \ref{propcanon}, the map $r^*$ is just the dual as $\Q$-vector spaces of the map $H_*((T^*)^f;\Q)\twoheadrightarrow \Tors_R H_*(U^f;\Q)$ given by the inclusion of infinite cyclic covers $(T^*)^f\subset U^f$. Moreover, $r^*$ allows us to see the MHS on the $\Tors_R H^{*+1}(U;\ov\calL)$ as a sub-MHS of the limit MHS on $\H^*(E;\psi_{\bar{f}}\underline{\Q})$ for all $*$, as exemplified in the following result. \begin{thm}\label{thm:limitMap} The maps $r^*$ and $\Phi^*$ are equal up to multiplication by a rational constant. In particular, \[ r^*:\Tors_R H^{*+1}(U;\ov{\mathcal L})\hookrightarrow \mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q}) \] is a morphism of MHS and of $R$-modules. \end{thm} \begin{proof} In order to prove this theorem, we must first recall how the hypercohomology of $\Tot \calA^{\bullet,\bullet}$ is identified with the nearby cycle cohomology in \cite{peters2008mixed}, sections 11.2.4 and 11.2.5. First, note that since $\pi$ is a covering map, we can see that $\pi_*$ is exact by checking over an open cover of $U$. Therefore, there are natural isomorphisms $ H^*((T^*)^f;\C) \cong \H^*(T^*;R\pi_*\C) \cong \H^*(T^*;\pi_*\C) $. We define the complex $\Omega_T^\bullet(\log E)[f_\infty]\coloneqq \Omega_T^\bullet(\log E)\otimes_{\C} \C[f_\infty] $, where the differential is given by letting $df_\infty = \frac{df}{f}$ and using the Leibniz rule. By \cite[Theorem 11.16]{peters2008mixed} and the discussion preceding it, we have quasi-isomorphisms: \[ i^{-1}Rj_*\pi_*\ul \C_{(T^*)^f} \xhookrightarrow{\sim} i^{-1}Rj_*\pi_*\Omega_{(T^*)^f}^\bullet \xhookleftarrow{\sim} i^{-1}\Omega_T^\bullet(\log E)[f_\infty] . \] The right hand arrow is given by pulling back forms and seeing $f_\infty$ as $f_\infty\colon(T^*)^f\subseteq U^f\to \C$. Next, combining Theorem 11.16 and the discussion in 11.2.5 in loc. cit., we conclude that there is a quasiisormophism as follows: \begin{equation}\label{eqn:qisoSteenbrink} \begin{split} \displaystyle {i_*i^{-1}\Omega_T^\ell(\log E)[f_\infty]} & \overset{\sim}{\longrightarrow} \displaystyle {(\Tot \calA^{\bullet,\bullet})^\ell = \bigoplus_{p=0}^{\ell+1} \frac{\Omega^{\ell + 1}_T(\log E)}{W_p\Omega^{\ell + 1}_T(\log E)}}\\ \displaystyle {\sum_{j} (f_\infty)^j\omega_j } & \longmapsto \displaystyle {(-1)^\ell\frac{1}{2\pi i}\frac{df}{f}\wedge \omega_0\in \frac{\Omega^{\ell + 1}_T(\log E)}{W_0\Omega^{\ell + 1}_T(\log E)}.} \end{split} \end{equation} We should note that the $\frac{1}{2\pi i}$ in the formula above does not appear in \cite{peters2008mixed} but it needs to appear here due to the difference in our conventions regarding Tate twists. As mentioned in Section \ref{ss:MHSsAndComplexes}, Tate twists in loc. cit. are defined using powers of $2\pi i$. The rational part of $\psi_{\bar{f}}^{\textnormal{Hdg}}$ is defined in loc. cit. using Tate twists, and the extra $\frac{1}{2\pi i}$ in (\ref{eqn:qisoSteenbrink}) is necessary for that quasiisomorphism to also be defined with rational coefficients, following our conventions. Needless to say, both conventions give rise to the same MHS. Let us fill the following commutative diagram. The diagonal arrow is defined above. \begin{equation}\label{eqn:commPhi} \begin{tikzcd} \Omega^\bullet_T(\log E)[f_\infty]\arrow[r,"i^{-1}\vdash i_*"]\arrow[d,dashrightarrow] & i_*i^{-1}\Omega^\bullet_T(\log E)[f_\infty]\arrow[dr, bend left = 20]\arrow[d,dashrightarrow] & \\ \Omega_T^{\bullet+1}(\log E)\left(\frac{1}{2\pi i}\frac{df}{ f},\infty\right)[1] \arrow[r,"i^{-1}\vdash i_*"] & i_*i^{-1}\Omega_T^{\bullet+1}(\log E)\left(\frac{1}{2\pi i}\frac{df}{ f},\infty\right)[1] \arrow[r,"\Phi^\ell_{\C}(\infty)"] & \Tot \calA^{\bullet,\bullet} . \end{tikzcd} \end{equation} One can check directly that the map $\sum_{j} (f_\infty)^j\omega_j \mapsto \frac{1}{2\pi i}\frac{df}{f}\wedge \omega_0 \otimes 1\in \Omega_T^\bullet(\log E)\otimes R_\infty$ makes the diagram commute and it induces a map of complexes (taking into account that $d[1]=-d$ on the target). Consider now the following diagram, which includes the maps explained above. The thickened complex is a resolution of $\ov{\mathcal L}\otimes_R R_\infty$ using Proposition~\ref{propLocal}. The arrow $\pi_{{\mathcal L}}^\vee$ is the one in Proposition~\ref{prop:mapsAreEqual}. \begin{equation}\label{eqn:commLimit} \begin{tikzcd}[column sep = 1em] Rj_*\ul \C_{T^*} \arrow[dr,"\pi_{{\mathcal L}}^\vee",bend right = 10] \arrow[r,"\pi^{-1}\vdash \pi_*"] & Rj_*\pi_*\ul\C_{(T^*)^f} \to i_*i^{-1}Rj_*\pi_*\Omega_{(T^*)^f}^\bullet & i_*i^{-1}\Omega^\bullet_T(\log E)[f_\infty]\arrow[d,dashrightarrow]\arrow[l,"\sim"']\\ & Rj_*\ov{\mathcal L}[1]\otimes_R R_\infty \xrightarrow{\sim} \Omega_T^{\bullet}(\log E)\left(\frac{1}{2\pi i}\frac{df}{ f},\infty\right)[1] \arrow[r] & i_*i^{-1}\Omega_T^{\bullet}(\log E)\left(\frac{1}{2\pi i}\frac{df}{ f},\infty\right)[1]. \end{tikzcd} \end{equation} We are going to show that it commutes up to multiplication by a nonzero real constant. Before we show this, let us explain why the commutativity (up to constant) of this diagram finishes the proof. Identifying $\H^j(E;\psi_{\ov f}\C)$, $\H^j(E;\Omega^\bullet_T(\log E)[f_\infty])$ and $\H^j(T;\Tot \calA^{\bullet,\bullet})$ through the quasiisomorphisms between the corresponding complexes that we have described above, taking hypercohomology in the commutative diagrams (\ref{eqn:commPhi}) and (\ref{eqn:commLimit}) yields the following commutative (up to a non-zero constant) diagram. Note that we can identify $H^{j+1}(T^*;\ov{\mathcal L}\otimes_R R_\infty)$ with $H^{j+1}(T^*;\ov{\mathcal L})$ naturally, since the latter is annihilated by $s^m$ for some $m$, using Remark~\ref{proper}. $$ \begin{tikzcd}[column sep = 1.2em] H^j(T^*;\C) \arrow[dr,"H^j(\pi_{{\mathcal L}}^\vee)",bend right = 10] \arrow[r,"\pi^*"] & H^j((T^*)^f;\C) \arrow[r, "i^{-1}\vdash i_*"]& \H^j(E,\psi_{\ov f}\C) \arrow[r,equals]\arrow[dr,dashrightarrow] & \H^j(T;\Tot \calA^{\bullet,\bullet}) \\ & H^{j+1}(T^*;\ov{\mathcal L}) \arrow[rr, "i^{-1}\vdash i_*"] & & \H^{j+1}\left(E;\Omega_T^{\bullet}(\log E)\left(\frac{1}{2\pi i}\frac{df}{ f},\infty\right)\right)\arrow[u,"H^j(\Phi^\ell_{\C}(\infty))",swap]. \end{tikzcd} $$ If the above commutes (up to a scalar), we incorporate to this diagram arrows given by the inclusion of $T^*$ in $U$ in the left side. \begin{equation}\label{eqn:commCohom} \begin{tikzcd}[column sep = 0.8em] H^j(U;\C)\arrow[r]\arrow[dr,"H^j(\pi_{{\mathcal L}}^\vee)",bend right = 10, two heads] &H^j(T^*;\C) \arrow[dr,"H^j(\pi_{{\mathcal L}}^\vee)",bend right = 10] \arrow[r,"\pi^*"] & H^j((T^*)^f;\C) \arrow[r, "i^{-1}\vdash i_*"]& \H^j(E,\psi_{\ov f}\C) \\ & \Tors_R H^{j+1}(U;\ov{\mathcal L})\arrow[r] & H^{j+1}(T^*;\ov{\mathcal L}) \arrow[r & \H^{j+1}\left(E;\Omega_T^{\bullet}(\log E)\left(\frac{1}{2\pi i}\frac{df}{ f},\infty\right)\right)\arrow[u,"H^j(\Phi^\ell_{\C}(\infty))",swap]. \end{tikzcd} \end{equation} Let us recall why $H^j(\pi_{{\mathcal L}}^\vee)\colon H^j(U;\C)\to \Tors_R H^{j+1}(U;\ov{\mathcal L})$ is surjective. By Theorem \ref{thmsimple}, $\Tors_R H^{j+1}(U;\ov{\mathcal L})$ is a semisimple $R$-module. Recall that, in this section, we have made the assumption that the monodromy action on $\Tors_R H^{j+1}(U;\ov{\mathcal L})$ is unipotent (Remark \ref{unipotent}). The surjectivity now follows from Proposition~\ref{prop:mapsAreEqual}, which shows that $H^j(\pi_{{\mathcal L}}^\vee)$ is the dual of the map induced by $U^f\to U$, and Corolary \ref{cor:surj}, which shows that if $t$ acts as the identity on $\Tors_R H^{j+1}(U;\ov{\mathcal L})$, this map is surjective. We also note that, by Proposition \ref{prop:mapsAreEqual}, we have the following commutative diagram, where the vertical arrow comes from Proposition~\ref{propcanon}. The vertical isomorphism appears in Definition~\ref{def:relLimit}. $$ {\footnotesize \begin{tikzcd} H^j(T^*;\C) \arrow[dr,"H^j(\pi_{{\mathcal L}}^\vee)"',bend right = 10] \arrow[r,"\pi^*"] & H^j((T^*)^f;\C)\\ & H^{j+1}(T^*;\ov{\mathcal L}).\arrow[u,"\cong"] \end{tikzcd} } $$ Using the diagram above, the map $H^j(U;\C)\rightarrow \H^j(E;\psi_{\ov f}\C)$ in (\ref{eqn:commCohom}) given by the composition of all the arrows in the top row is none other than $(r^*\otimes\C)\circ H^j(\pi_{{\mathcal L}}^\vee)$. The map $\Tors_R H^{j+1}(U;\ov{\mathcal L})\rightarrow \H^j(E;\psi_{\ov f}\C)$ in (\ref{eqn:commCohom}) given by the composition of all the arrows in the bottom row and the vertical arrow on the right is none other that $\Phi^*\otimes\C$: recall that when we defined $\Phi^*$ in the proof of Corollary~\ref{relnlimitmhs}, we showed that by definition it is induced by $\Phi_{\Q}(\infty)$ taking cohomology and using our identifications. Further, by Lemma~\ref{surj} $\Phi_{\C}(\infty)$ and $\Phi_{\Q}(\infty)\otimes \C$ induced the same morphism in cohomology $\Tors_R H^{j+1}(U;\ov{\mathcal L}\otimes \C)\to H^j(E;\psi_{\ov f}\C) $. The commutativity (up to non-zero constant) of (\ref{eqn:commCohom}) implies that for some $c\in \R\setminus\{0\}$, $c \cdot (r^*\otimes\C)\circ H^j(\pi_{{\mathcal L}}^\vee)=(\Phi^*\otimes\C)\circ H^j(\pi_{{\mathcal L}}^\vee)$. The surjectivity of $H^j(\pi_{{\mathcal L}}^\vee)$ implies that $c \cdot (r^*\otimes_{\Q} \C)=(\Phi^*\otimes_{\Q} \C)$. Finally, the fact that $\C$ is fully faithful over $\Q$ implies that $c\cdot r^*=\Phi^*$ and $c$ must actually be a rational number, as desired. To conclude the proof, we want to show that diagram (\ref{eqn:commLimit}) commutes up to a real constant. We use the de Rham resolution to identify $\ul \C_{T^*}$ with $\Omega_{T^*}^\bullet$, and similarly on $(T^*)^f$. Then, the commutativity of (\ref{eqn:commLimit}) is reduced to the commutativity of (\ref{eqn:commLimit-deRham}) below. We are denoting by $\star$ the map of de Rham complexes induced by $\pi_{{\mathcal L}}^\vee$, and we explain below why the dashed arrow $\omega\mapsto \frac{1}{i}\frac{df}{ f} \wedge \omega$ is indeed induced by $\pi_{\mathcal L}^\vee$. \begin{equation}\label{eqn:commLimit-deRham} \begin{tikzcd}[column sep = 1.5em, row sep = 2.6 em] i_*i^{-1}\Omega_T^\bullet(\log E) \arrow[dr,dashrightarrow,"\Id_{\Omega_T^\bullet(\log E)}", bend left = 15,end anchor={[xshift=9ex, yshift = -0.5ex]}] \\ \Omega_T^\bullet(\log E)\xrightarrow{\pi^{-1}\vdash \pi_*} \arrow[u,"i^{-1}\vdash i_*",start anchor={[xshift=-5ex, yshift = -1ex]}] \arrow[d,"(\Id \to i_*i^{-1})\circ \star"',start anchor={[xshift=-5ex, yshift = 1ex]}, bend right = 15] \arrow[dr,dashrightarrow," \omega\mapsto \frac{1}{i}\frac{df}{ f} \wedge \omega" ,start anchor={[xshift=-12ex, yshift = 1ex]}] Rj_*\pi_*\Omega_{(T^*)^f}^\bullet \arrow[r] & i_*i^{-1}Rj_*\pi_*\Omega_{(T^*)^f}^\bullet \xleftarrow{\sim} i_*i^{-1}\Omega^\bullet_T(\log E)[f_\infty] \arrow[d," \sum (f_\infty)^j \omega_j\mapsto \frac{1}{2\pi i}\frac{df }{f}\wedge \omega_0",start anchor={[xshift=5ex, yshift = 1ex]}, bend left = 10] \\ i_*i^{-1}\Omega_T^{\bullet}(\log E)\left(\frac{1}{ i}\frac{df}{ f},\infty\right)[1] \arrow[r,"G_\infty^{-1}"] & i_*i^{-1}\Omega_T^{\bullet}(\log E)\left(\frac{1}{2\pi i}\frac{df}{ f},\infty\right)[1]. \end{tikzcd} \end{equation} First, a direct computation shows that the inclusion of $\Omega_T^\bullet(\log E)$ in $\Omega^\bullet_T(\log E)[f_\infty]$ makes the top portion of the above diagram commute. Therefore, the composition (in the derived category) of the middle row is the inclusion $\Omega_T^\bullet(\log E)\hookrightarrow\Omega^\bullet_T(\log E)[f_\infty]$ composed with the adjunction map $\Id\to i_*i^{-1}$. Next, consider the map $\star$ induced by $\pi_{{\mathcal L}}^\vee$, which is the complexification of the map $\pi_{{\mathcal L}}^\vee$ appearing in Lemma~\ref{lem:pullbacks}. By this Lemma (up to a real constant and taking the inverse limit as $m\to \infty$), $\pi_{{\mathcal L}}^\vee$ corresponds to the map of real de Rham complexes $A:{\mathcal E}_T^\bullet\to {\mathcal E}_T^{\bullet} \left( \Im\frac{df}{f},\infty \right)[1]$ given by $\alpha\mapsto \Im \frac{df}{f}\wedge \alpha$. By Lemma~\ref{lem:theMapA} and its proof, the complexification of this map of real de Rham complexes induces in the logarithmic de Rham complexes the map $\omega\mapsto \frac{1}{i} \frac{df}{f}\wedge \omega:\Omega_{T^*}^\bullet(\log E)\to \Omega_{T^*}^{\bullet}(\log E)\left(\frac{1}{ i}\frac{df}{ f},\infty\right)[1]$ (these are maps in the derived category, in particular, two homotopic maps are equal, as in the proof of Lemma~\ref{lem:theMapA}). The map $G_\infty$ is the inverse limit of the maps $G_m$ defined in the proof of Theorem~\ref{Qalexandermhs}. Namely, its inverse is given by \[ G_\infty^{-1}(\omega\otimes s^j) = (2\pi)^{-j}\omega\otimes s^j. \] As in the proof of Theorem~\ref{Qalexandermhs}, a straightforward computation shows that the following diagram commutes. Recall that $\nu_{\Q}$ (resp. $\nu$) is defined in Remark~\ref{rem:nuQ} (resp. Remark~\ref{rem:nu}). \[ \begin{tikzcd}[row sep = 1.2em] {\mathcal L} \arrow[r] \arrow[dr,"\nu_\Q"']& {\mathcal L}\otimes_{\Q} \C \otimes_R R_\infty \arrow[r,"\nu\otimes \C"]& \Omega_T^{\bullet}(\log E)\left(\frac{1}{ i}\frac{df}{ f},\infty\right) \arrow[d,"G_\infty^{-1}"]\\ & \calK^\bullet_\infty(1\otimes f,m)\arrow[r,"\varphi_\infty"] & \Omega_T^{\bullet}(\log E)\left(\frac{1}{ 2 \pi i}\frac{df}{ f},\infty\right) \end{tikzcd} \] Using this diagram, we see that the lower left triangle in (\ref{eqn:commLimit-deRham}) commutes, since up to the resolutions above it is the following diagram: \[ \begin{tikzcd}[row sep = 0.8 em] \ul {\C}_{T^*} \arrow[d,"\pi_{{\mathcal L}}^\vee"'] \arrow[dr,"\pi_{{\mathcal L}}^\vee"]\\ \ov{\mathcal L} \arrow[r,equals] & \ov{\mathcal L}. \end{tikzcd} \] Finally, it follows that (\ref{eqn:commLimit-deRham}) commutes up to real constant, by a computation on the remaining triangle. \end{proof} \begin{cor}\label{limit} Let $f\colon U\rightarrow \C^*$ be a fibration. Then, the map $$ r^*\colon \mathrm{Tors}_R\,H^{*+1}(U;\ov\calL) \rightarrow \mathbb{H}^*(E; \psi_{\bar{f}}\underline{\Q}) $$ of Definition~\ref{def:relLimit} is an isomorphism of MHS for all $*$. In other words, the MHS described in Corollary \ref{alexandermhs} coincides with the limit MHS. \end{cor} \begin{proof} If $f$ is a fibration, the inclusion $T^*\hookrightarrow U$ is a homotopy equivalence, so $H^{*}(U;\ov\calL)$ is $R$-torsion, and the monomorphism $\Tors_R H^{*+1}(U;\ov{\mathcal L})\hookrightarrow H^{*+1}(T^*;\ov{\mathcal L})$ in Definition~\ref{def:relLimit} is actually an isomorphism. \end{proof} The content of Theorem~\ref{comp} is the result of combining Theorem~\ref{thm:limitMap} and Corollary~\ref{limit} in homological notation (dualizing via Corollary~\ref{isocohom}). \section{Examples and open questions}\label{sec:examples} \subsection{Hyperplane arrangements}\label{sec:hyp} Let $n\geq 2$. Let $f_1,\ldots,f_d$ be degree $1$ polynomials in $\C[x_1,\ldots,x_n]$ defining $d$ distinct hyperplanes and let $f=f_1\cdot\ldots\cdot f_d$. The zeros of $f$ define a hyperplane arrangement ${\mathcal A}$ of $d$ hyperplanes in $\C^n$. Let $U\subset \C^n$ be the corresponding arrangement complement. \begin{dfn}\label{def:rankessential} The rank of a hyperplane arrangement ${\mathcal A}$ in $\C^n$ is the maximal codimension of a non-empty intersection of some subfamily of ${\mathcal A}$. We say that ${\mathcal A}$ is an essential hyperplane arrangement if its rank is equal to $n$. \end{dfn} \begin{remk}[Reducing to the case where ${\mathcal A}$ is essential]\label{rem:essential} Suppose that ${\mathcal A}$ has rank $l<n$. By \cite[Proposition 6.1]{kohnopajitnov}, there exists an affine subspace $L$ of dimension $l$ such that $U_L:=U\cap L$ (seen in $\C^l$) is an essential hyperplane arrangement complement, and the inclusion $U_L\hookrightarrow U$ is a homotopy equivalence. Let $f_L$ be the restriction of $f\colon U\rightarrow \C^*$ to $U_L$. The functoriality of the MHS (Theorem \ref{functorial}), as well as the homotopy invariance of cohomology with local systems, ensures that the map $$ \Tors_R H^j(U;\overline {\mathcal L})\rightarrow \Tors_R H^j(U_L;\overline{{\mathcal L}_L}) $$ induced by inclusion is a MHS isomorphism between cohomology Alexander modules for all $j$, where ${\mathcal L}_L$ is the local system induced by $f_L$. Therefore, the study of the MHS on $\Tors_R H^j(U;\overline {\mathcal L})$ for ${\mathcal A}$ a hyperplane arrangement can be reduced to the case where ${\mathcal A}$ is an essential hyperplane arrangement.$\sslash$ \end{remk} \begin{remk}[The cohomology groups of the infinite cyclic cover.]\label{rem:cohomologyhyperplanes} Let ${\mathcal A}$ be an essential hyperplane arrangement in $\C^n$ defined by the zeroes of a reduced polynomal $f$. Let $U$ be the corresponding arrangement complement, and ${\mathcal L}$ the local system on $U$ induced by $f\colon U\rightarrow \C^*$. By \cite[Theorem 4]{thesiseva}, we have that $ H_j(U;{\mathcal L})$ is a torsion $R$-module for all $j<n$, a free $R$-module for $j=n$, and $0$ for $j>n$. Hence, by Proposition \ref{propcanon} and Corollary \ref{isocohom}, we have canonical isomorphisms $$\Tors_R H^{j+1}(U;\overline{{\mathcal L}})\cong H^{j}(U^{f};k)$$ for $0\leq j<n$, and $$\Tors_R H^{j+1}(U;\overline{{\mathcal L}})\cong 0$$ for $j\geq n$. We use this canonical isomorphism to endow $H^j(U^{f};k)$ with a MHS, for $0\leq j\leq n-1$. In this section, we will talk about the MHS on $H^j(U^{f};k)$ instead of the isomorphic MHS on $\Tors_R H^{j+1}(U;\overline{{\mathcal L}})$, both to simplify the notation and to highlight the geometric nature of the situation.$\sslash$ \end{remk} \begin{remk}[Connectivity of the fiber]\label{rem:connectedfiber} Let ${\mathcal A}$ be an essential arrangement of $d$ hyperplanes in $\C^n$ defined by the zeros of a reduced polynomial $f$ of degree $d$, with $n>1$. Then, by \cite[Theorem 2.1]{DimcaTame} (and the discussion following it), the generic fiber of $f$ is connected.$\sslash$ \end{remk} By Corollary \ref{cor:semisimple}, the $t$-action on $H^1(U^f;k)$ is semisimple, so by Corollary \ref{cor:t}, we have an isomorphism of MHS $H^1(U^f;k)\cong H^1(U^f;k)_1\oplus H^1(U^f;k)_{\neq 1}$. The goal of this section is to arrive at the following result. \begin{thm}\label{thm:hyperplanes} Let ${\mathcal A}$ be an arrangement of $d$ hyperplanes in $\C^n$ defined by the zeros of a reduced polynomial $f$ of degree $d$, for $n\geq 2$. Assume that not all the hyperplanes of ${\mathcal A}$ are parallel, or equivalently, that the rank of ${\mathcal A}$ is greater than or equal to $2$. Then, \begin{enumerate} \item $H^1(U^f;k)_1$ is a pure Hodge structure of type $(1,1)$, and has dimension $d-1$. \item $H^1(U^f;k)_{\neq 1}$ is a pure Hodge structure of weight $1$. \end{enumerate} \end{thm} \begin{remk}[The Alexander polynomial of an essential line arrangement]\label{rem:alexpoly} The first Alexander polynomial $\Delta_1(t)$ of an essential line arrangement is defined as the order of the torsion $R$-module $H_1(U;{\mathcal L})\cong H_1(U^f;k)$, or equivalently, as a generator of the $0$-th Fitting Ideal of the $R$-module $H_1(U;{\mathcal L})$. Hence, it is well defined up to multiplication by a unit of $R$. Since $H_1(U;{\mathcal L})$ is semisimple, $\Delta_1(t)$ determines the $R$-module structure of both $H_1(U;{\mathcal L})$ and $H^1(U^f;k)$, its dual as a vector space. Note that Theorem \ref{thm:hyperplanes} tells us that the first Alexander polynomial of an essential line arrangement complement determines the Hodge numbers of $H^1(U^f;k)$.$\sslash$ \end{remk} If ${\mathcal A}$ is a \textit{central} hyperplane arrangement ($f$ is a homogeneous polynomial), $f$ determines a global Milnor fibration with fiber $F$, so $H^{j}(U^f;k)\cong H^j(F;k)$ for all $j$. In particular, $H^{j}(U^f;k)$ is a finite dimensional vector space for all $j$, so by Remarks \ref{rem:essential} and \ref{rem:cohomologyhyperplanes}, $H^j(U^f;k)=0$ for all $j\geq \text{ rank }{\mathcal A}$. Hence, Corollary \ref{cor:quasihom} and Remark \ref{rem:cohomologyhyperplanes} tell us that the isomorphism $H^{j}(U^f;k)\cong H^j(F;k)$ is a MHS isomorphism for all $j$. Thus, Theorem \ref{thm:hyperplanes} is a direct generalization of parts (1) (in the case $j=1$) and (3) of the following result regarding \textit{central} hyperplane arrangements, which can be found in \cite[Theorem 7.7]{dimca2017hyperplanes}. Note that the last assertion in the result below follows from the second to last one. \begin{thm}[\cite{dimca2017hyperplanes}, Theorem 7.7]\label{thm:central} Let ${\mathcal A}$ be a central hyperplane arrangement in $\C^{n}$ defined by a homogeneous reduced polynomial $f$. Let $F$ denote its (global) Milnor fiber, given by the equation $f=1$. \begin{enumerate} \item $H^j(F;k)_1$ is a pure Hodge structure of type $(j,j)$ for any $j\leq n-1$. \item $\Gr_{2j}^W H^j(F,k)_{\neq 1}=0$ for any $j\leq n-1$. \item $H^1(F;k)_{\neq 1}$ is a pure Hodge structure of weight $1$. \end{enumerate} \end{thm} Before we prove Theorem \ref{thm:hyperplanes}, we need the following lemma regarding the MHS on the generic fiber of $f$. \begin{lem}\label{lem:genFiberLines} Let ${\mathcal A}$ be an essential line arrangement in $\C^2$, given by the zeros of a reduced polynomial $f$ of degree $d$. Let $c\in \C$ be generic, and let $F = f^{-1}(c)\subset \C^2$. Then \[ \dim \Gr^W_2 H^1(F;k) = d-1. \] \end{lem} \begin{proof} By generic smoothness and Remark \ref{rem:connectedfiber}, $F$ is a smooth connected curve, whose genus we will denote $g$. Let $\ov F$ be its closure in $\C P^2$, and let $\wt F$ be the normalization of $\ov F$. By \cite[Corollaire 3.2.15 and Corollaire 3.2.17]{De2}, the mixed Hodge structure on $H^1(F;k)$ has $W_0H^1(F;k)=0$, $W_2H^1(F;k) = H^1(F;k)$ and $W_1H^1(F;k)$ is the image of $H^1(\wt F;k)$. The latter has dimension $2g$, and the map $H^1(\wt F;k)\to H^1(F;k)$ is injective, since $F$ is a punctured genus $g$ orientable surface. Let $\#p$ be the number of punctures. Then: \[ \dim \Gr^W_2 H^1(F;k) = \dim \frac{ H^1(F;k)}{W_1 H^1(F;k)} = (2g + \#p -1) - 2g = \# p-1. \] So all we need to show is that $\#p = d$. Take the set of points at infinity $\{p_i\}\coloneqq \ov{f^{-1}(0)}\setminus \C^2$. Let $r_i$ be the number of lines of the arrangement passing through $p_i$. Locally around $p_i$, the closure of $\{f=0\}$ has an ordinary singularity of multiplicity $r_i$. The fibers $\ov F$ are the curves in the pencil generated by $\{f=0\}$ and $d\cdot L_\infty$, where $L_\infty$ denotes the line at infinity. Since the multiplicity of $d\cdot L_\infty$ at $p_i$ is $d>r_i$ (because the arrangement is essential), all the fibers $\ov F$ have ordinary singularities of multiplicity $r_i$ at $p_i$. This means that $\wt F$ has $r_i$ many branches over $p_i$, and this is all we need: $\#p = \sum_i r_i = d$. \end{proof} Now, we can finally prove Theorem \ref{thm:hyperplanes}. \begin{proof}[Proof of Theorem \ref{thm:hyperplanes}] This result deals with the MHS on $H^1(U^f;k)\cong (H_1(U;{\mathcal L}))^{\vee_k}$. In light of Remark \ref{rem:essential}, we see that to study $H^1(U^f;k)$, it suffices to consider the case in which ${\mathcal A}$ is an essential hyperplane arrangement. After intersecting with enough generic hyperplanes, we can and will assume in this proof that ${\mathcal A}$ is an essential line arrangement in $\C^2$, by a Lefschetz type argument. Let us start by proving that $H^1(U^f;k)_1$ is a pure Hodge structure of type $(1,1)$. By Proposition \ref{prop:kerim}, the map $H^1(U;k)\rightarrow H^{2}(U;\ov{\mathcal L})\cong H^1(U^f;k)$ induced by the covering space map $\pi\colon U^f\rightarrow U$ is surjective onto the $(t-1)$-torsion of $H^1(U^f;k)$. Since the $t$-action on $H^1(U^f;k)$ is semisimple, we get that $H^1(U;k)\rightarrow H^1(U^f;k)_1$ is a surjective MHS morphism, and the purity result follows from the fact that $H^j(U;k)$ is a pure Hodge structure of type $(j,j)$ for all $U$ affine hyperplane arrangement complement, by \cite{shapiro}. Now, we prove that $\dim_k H^1(U^f;k)_1=\dim_k H_1(U^f;k)_1= d-1$. We start with the Milnor long exact sequence (already discussed in Proposition \ref{prop:kerim}) $$\cdots \to H_1(U^f;k) \overset{t-1}{\to} H_1(U^f;k) \to H_1(U;k) \overset{\partial}{\to} H_0(U^f;k) \overset{t-1}\to H_0(U^f;k)\to H_0(U;k) \to 0.$$ Since $U^f$ and $U$ are connected, $\dim_k H_0(U^f;k)=1=\dim_k H_0(U;k)$, so $H_0(U^f;k) \overset{t-1}\to H_0(U^f;k)$ is the zero map. Now, since $H_1(U^f;k)$ is a semisimple $R$-module, then $H_1(U^f;k)/(t-1) H_1(U^f;k)\cong H_1(U^f;k)_1$. Hence, the Milnor long exact sequence gives us the short exact sequence $$ 0\to H_1(U^f;k)_1 \to H_1(U;k) \to H_0(U^f;k)\to 0. $$ Since $\dim_k H_1(U;k)=d$, this finishes our proof of the equality \begin{center}$\dim_k H^1(U^f;k)_1=\dim_k H_1(U^f;k)_1= d-1$.\end{center} Recall that, by Theorem \ref{thm:boundedWeights}, $\Gr_j^W H^1(U^f;k)=0$ for all $j\neq 1,2$. Let $F\subset U$ be a generic fiber of $f$. By Corollary \ref{cor:fiber}, the map $$ H^1(U^f;k)\hookrightarrow H^1(F;k) $$ induced by inclusion is a morphism of MHS. We know $H^1(U^f;k)_1$ is pure Hodge structure of weight $2$ and dimension $d-1$. By Lemma \ref{lem:genFiberLines} and the inclusion above, we get that $\Gr_2^W H^1(U^f;k)_{\neq 1}=0$, concluding our proof. \end{proof} In light of Theorem \ref{thm:hyperplanes} and Remark \ref{rem:alexpoly}, one might wonder in which cases the MHS of $H^1(U^f;k)$ is pure. If $d>1$, this amounts to $H^1(U^f;k)_{\neq 1}=0$, or equivalently, $\Delta_1=(t-1)^{d-1}$, where $\Delta_1$ is the first Alexander polynomial of the line arrangement complement. One can find sufficient conditions for $\Delta_1=(t-1)^{d-1}$ in \cite[Theorem 6]{thesiseva}, for example, which translated to the notation of this paper reads as follows. \begin{prop}[\cite{thesiseva}, Theorem 6]\label{prop:purity} Let ${\mathcal A}=\{L_1,\ldots,L_d\}$ be an essential line arrangement of $d$ lines in $\C^2$, and, after reordering, let ${\mathcal B}=\{L_1,\ldots,L_l\}$ be the set of lines in ${\mathcal A}$ such that for each line in ${\mathcal B}$ no other line in ${\mathcal A}$ is parallel to it, where $0\leq l\leq d$. Suppose that ${\mathcal B}\neq\emptyset$. If for every $m>2$, there exists a line in ${\mathcal B}$ with no points of multiplicity divisible by $m$, then $\Delta_1$ is a power of $t-1$, or equivalently, $$ H^1(U^f;k)=H^1(U^f;k)_1\cong k^{d-1} $$ is pure of type $(1,1)$. In particular, if there exists a line in ${\mathcal B}$ with only double points, the hypotheses of this proposition are satisfied. \end{prop} By Proposition \ref{prop:purity}, we know that there are many examples of essential line arrangements such that $H^1(U^f;k)_{\neq 1}=0$ and $H^1(U^f;k)$ is pure of type $(1,1)$. Here are some examples illustrating Theorem \ref{thm:hyperplanes} and Remark \ref{rem:alexpoly}, in which the MHS is not pure. \begin{ex} Consider a central line arrangement of $d$ lines, i.e., defined by the equation $x^d=y^d$. A simple application of the Thom-Sebastiani theorem yields that the Alexander polynomial of the complement is $\prod_{\alpha,\beta}(t-\alpha \beta)=(t^d-1)^{d-2}(t-1)$, where the product runs over $n$th roots of unity $\alpha, \beta$, with $\alpha\neq 1, \beta \neq 1$. Hence, the non-zero Hodge numbers of $H^1(U^f;k)$ are $h^{1,1}=d-1$, $h^{0,1}=h^{1,0}=\frac{(d-1)(d-2)}{2}$. By Corollary \ref{cor:quasihom}, we have an isomorphism of MHS $H^1(U^f;k)\cong H^1(F;k)$, where $F$ is the global Milnor fiber of the homogeneous polynomial $f$. If $d=3$, the MHS (not just the Hodge numbers) on $H^1(U^f;k)$ is determined as follows. The closure $\ov F$ of $F$ in $\C P^2$ is the elliptic curve whose $j$-invariant is $0$, and $F$ is $\ov F$ with three points removed. Following the proof of Lemma \ref{lem:genFiberLines}, we have the MHS isomorphism $H^1(U^f;k)_{\neq 1}\cong H^1(\ov F;k)$. \end{ex} \begin{ex}[A non-central line arrangement with non-trivial $H^1(U^f;k)_{\neq 1}$] Let ${\mathcal A}$ be the line arrangement defined by the zeros of $f(x,y)=x(x-1)y(y-1)(x+y-1)$. $$ \begin{tikzpicture} \draw (-1,1.5) -- (-1,-1.5); \draw (1,1.5) -- (1,-1.5); \draw (-1.5,1.5) -- (1.5,-1.5); \draw (-1.5,1) -- (1.5,1); \draw (-1.5,-1) -- (1.5,-1); \end{tikzpicture}$$ The Alexander polynomial is $(t-1)^4(t^2+t+1)$. Hence, the non-zero Hodge numbers of $H^1(U^f;k)$ are $h^{1,1}=4$, $h^{0,1}=h^{1,0}=1$. \end{ex} We end this section with several open questions regarding the MHS on Alexander modules for hyperplane arrangement complements. \begin{open} Is $H^j(U^f;k)$ a semisimple $R$-module for $j>1$? \end{open} If the question above had a positive answer, then Corollary~\ref{cor:t} would give us a MHS isomorphism $H^j(U^f;k)\cong H^j(U^f;k)_1\oplus H^j(U^f;k)_{\neq 1}$. In that case, $H^j(U^f;k)_1$ would be pure of type $(j,j)$, like part (1) of Theorem \ref{thm:central}, and the proof would be the same as the $j=1$ case in Theorem \ref{thm:hyperplanes}. Moreover, it would make sense to ask if the generalization of part (2) of Theorem \ref{thm:central} holds, namely, \begin{open} Assume $H^j(U^f;k)$ is a semisimple $R$-module. Is $\Gr_{2j}^W H^j(U^f;k)_{\neq 1}=0$? \end{open} \subsection{Future directions. Open questions.}\label{sec:open} In addition to the questions already mentioned in Section \ref{sec:hyp}, we list here several open problems we hope to address in the future. Most of these are motivated by corresponding results for the (co)homology of the Milnor fiber $F_x$ associated to a complex hypersurface singularity germ $f\colon (\C^n,x) \to (\C,0)$. We aim to globalize such statements by replacing the (co)homology of the Milnor fiber $F_x$ with the torsion part $$A_*(U^f;\Q):=\Tors_R H_*(U^f;\Q) $$ of the homology Alexander modules associated to an algebraic map $f\colon U\to \C^*$. \subsubsection{Semisimplicity} We begin with the following. \begin{open} Let $t_s$ denote the semisimple part of the $t$-action on $A_*(U^f;\Q)$. Is $t_s$ a morphism of MHS? \end{open} This question is a generalization of our result from Corollary \ref{cor:t}, which provides an affirmative answer in the case $t=t_s$. It is motivated by the corresponding results for the semisimple part of the monodromy operator acting on the Milnor fiber cohomology, and, respectively, on the cohomology of the generic fiber of a proper family $f\colon U \to \Delta^*$ over a punctured disc (see \cite[Th\'eor\`eme 15.13]{NA}). It would also be interesting to find examples of pairs $(U,f)$, with $U$ a smooth connected complex algebraic manifold and $f\colon U \to \C^*$ an algebraic map, for which $A_i(U^f;\Q)$ is not a semisimple $R$-module for some $i$. In many of the algebraic situations considered in this paper, we have in fact that $A_i(U^f;\Q)$ is semisimple for all $i$. This applies, in particular, to the following cases: \begin{itemize} \item When $f\colon U=\C^n \setminus \{f=0\} \to \C^*$ is induced by a complex polynomial $f\colon \C^n \to \C$ which is transversal at infinity (see \cite{DL, Max06}), e.g., $f$ could be a homogeneous polynomial. More generally, $A_i(U^f;\Q)$ is semisimple in the case of Setting \ref{set:DL} (see \cite{DL}). \item When $f\colon U \to \C^*$ is a projective submersion of smooth complex algebraic varieties, the semisimplicity of $A_i(U^f;\Q)$ is a consequence of Deligne's decomposition theorem (see Remark \ref{rem:smf}). \item When $f\colon U \to \C^*$ is a proper algebraic map, the semisimplicity of $A_i(U^f;\Q)$ is a consequence of the decomposition theorem of Beilinson-Bernstein-Deligne \cite{BBD} (see Corollary \ref{corss}). \end{itemize} Let us also point out here that the semisimplicity property does not hold in general in the local situation, that is, for the monodromy operator acting on the (co)homology of the Milnor fiber associated to a complex hypersurface singularity germ (see, e.g., the discussion in \cite[Section I.9]{Ku}). \subsubsection{Finite type invariants} Despite the fact that (unlike the Milnor fiber of a hypersurface singularity germ) the infinite cyclic cover $U^f$ is not in general a CW complex of finite type, one can associate finite type invariants to $U^f$ (or better said, to the pair $(U,f)$) in terms of the $R$-torsion part $A_*(U^f;\Q)$ of the Alexander modules. For instance, one can define: \begin{itemize} \item Betti numbers: $b_i(U,f):=\dim_\Q A_i(U^f;\Q)$. \item mixed Hodge numbers: $h^{p,q,i}(U,f):=\dim_\C Gr^p_FGr_{p+q}^WA_i(U^f;\C)$. \item spectral pairs: if the semisimple part $t_s$ of the $t$-action is a MHS morphism (e.g., if $t=t_s$), let $h_\alpha^{p,q,i}(U,f)$ denote the dimension of the $\lambda$-eigenspace for the $t_s$-action on $Gr^p_FGr_{p+q}^WA_i(U^f;\C)$, where $\lambda=\exp(2\pi i \alpha)$ and $\alpha \in [0,1)$. The collection $\{h_\alpha^{p,q,i}(U,f)\}$ forms the {\it spectral pairs} of the $t_s$-action on the MHS $A_i(U^f;\Q)$. \end{itemize} In future work, we aim to investigate such finite-type invariants of the pair $(U,f)$; compare with \cite{LiMa} for a special case. In the case when $U$ is the complement of an essential hyperplane arrangement, it is also natural to ask about the combinatorial nature of such finite type invariants on $A_i(U^f;\Q)$. This question is motivated by similar open problems in the case of central arrangements, where, for instance, it is still unknown if the Betti numbers of the associated Milnor fiber are determined by the intersection lattice of the arrangement. See, e.g., \cite{LiMa} for results on the case of complements of line arrangements which are transversal at infinity, and also \cite{budursaito}, \cite{DiLe}, \cite{Yo} for the combinatorial invariance of the {\it Hodge spectrum} of a central arrangement and variants of this result. \subsubsection{Motivic realization} Motivated by connections between the Igusa zeta functions, Bernstein--Sato polynomials and the topology of hypersurface singularities, Denef and Loeser introduced in \cite{DeLo} the {\it motivic Milnor fiber} of a hypersurface singularity germ. This is a virtual variety endowed with an action of the group scheme of roots of unity, from which one can retrieve several invariants of the (topological) Milnor fiber, such as the Euler characteristic, Hodge spectrum, etc. More generally, to any (finite type) infinite cyclic cover associated to a punctured neighborhood of a divisor on a smooth quasiprojective variety, one attached in \cite{GLM1, GLM2} a {\it motivic infinite cyclic cover}. This is an element in the Grothendieck ring $K_0(\text{\rm Var}^{\hat \mu}_\C) $ of complex algebraic varieties endowed with a good action of the group scheme $\hat \mu$ of roots of unity, whose Betti realization recovers (upon taking degrees) the Euler characteristic of the (topological) infinite cyclic cover of the punctured neighborhood (see \cite[Proposition 4.2]{GLM1}). In the terminology of \cite[Section 4]{GLM1}, we can therefore ask the following. \begin{open} With $U$ and $f\colon U \to \C^*$ as above, does there exist an element in (a certain localization of) $K_0(\text{\rm Var}^{\hat \mu}_\C) $, whose Betti realization yields $\sum_i (-1)^i [A_i(U^f;\Q)] \in K_0(V_\Q^{\text{aut}})$? Under the semisimplicity assumption for the $t$-action, a similar question can be asked about the Hodge realization of such a motive. \end{open} Here, $K_0(V_\Q^{\text{aut}})$ denotes the Grothendieck ring of the category of finite dimensional $\Q$-vector spaces endowed with a finite order automorphism (which in our case is given by the semisimple part of the $t$-action). \subsubsection{Mixed Hodge module realization} To each complex algebraic variety $X$, M. Saito \cite{Sai} associated an abelian category $\text{MHM}(X)$ of algebraic mixed Hodge modules on $X$, in such a way that Deligne's category of mixed Hodge structures is recovered as mixed Hodge modules over a point space. Mixed Hodge modules are extensions in the singular context of (admissible) variations of mixed Hodge structures, and can be regarded, informally, as sheaves of mixed Hodge structures. Hypercohomology groups of a variety, with coefficients in a complex of mixed Hodge modules, are naturally endowed with mixed Hodge structures. In recent decades, Saito's theory has been very successful at constructing mixed Hodge structures on new entities (e.g., on intersection cohomology groups of complex algebraic varieties), as well as recovering previously known such structures (see, e.g., \cite[Chapter 11]{Max-book} for an overview). It is therefore natural to ask the following. \begin{open} Given the pair $(U,f)$ as before, can one recover the mixed Hodge structures on $A_*(U^f;\Q)$ via Saito's mixed Hodge module theory? \end{open} For instance, motivated by Corollary~\ref{torsion}, one can try to define a (complex of) mixed Hodge module(s) whose underlying rational complex is $\ov{\mathcal L}\otimes_R R_m$, and such that the map $\psi_{m,m}\colon \ov{\mathcal L}\otimes_R R_m \hookrightarrow \ov{\mathcal L}\otimes_R R_{2m}$ comes from a map in $D^b\text{MHM}(U)$. Then for $m\gg 0$, the torsion part of the cohomology Alexander modules would inherit the MHS on the kernel of this map in cohomology induced from $\psi_{m,m}$. Let us just note here that the limit mixed Hodge structure of Section \ref{rellim} has such a mixed Hodge module realization. This is due to the fact that the nearby cycle functor of constructible sheaves lifts to the derived category of bounded complexes of mixed Hodge modules. Therefore, one can ask the following. \begin{open} Is the comparison map of Theorem \ref{thm:limitMap} induced by a map of complexes of mixed Hodge modules? \end{open} \subsubsection{Comparison to the limit mixed Hodge structure in the nonproper case} Under certain assumptions on $f$, a limit mixed Hodge structure can be defined even if the map $f\colon U \to \C^*$ is not proper, see \cite[Section 5]{SZ} and also \cite{EZ1,EZ2}. It would therefore be interesting to see if Theorem \ref{comp} holds without the properness assumption; we hope to address this general situation in future work. \subsubsection{Generalizations to other algebraic maps} Let $f\colon X\to Y$ be an algebraic map of connected complex algebraic varieties. Any such map is homotopy equivalent to a fibration over $Y$. The fiber of this fibration, denoted by $E_f$, is called the homotopy fiber of $f$. For instance, if $Y$ is an aspherical space and $f$ induces an epimorphism on $\pi_1$, then $E_f$ is the covering space of $X$ defined by the kernel of $\pi_1(X) \to \pi_1(Y)$. In \cite{hain1987rham}, Hain proved that if the cohomology groups of $E_f$ are finite dimensional and $\pi_1(Y)$ acts unipotently on them, then $H^*(E_f;\Q)$ have natural mixed Hodge structures. This suggests generalizations of our results to allow singularities as well as to more general algebraic maps. \begin{open} Can Theorem \ref{mhsexistence} be generalized to arbitrary complex algebraic varieties $U$? \end{open} \begin{open} Does Theorem \ref{mhsexistence} generalize to other algebraic maps $f\colon X\to Y$? \end{open} Let $f\colon X\to Y$ and $E_f$ be defined as above. Each cohomology group $H^i(E_f; \Q)$ has natural $\Q[\pi_1(Y)]$-module structure. If $H^i(E_f; \Q)$ is Noetherian, then it contains a unique maximal $\Q[\pi_1(Y)]$-submodule that is a finite dimensional $\Q$-vector space and which is a natural generalization to the torsion part of the Alexander module considered in this paper. We can therefore ask whether this submodule admits a natural mixed Hodge structure. \subsubsection{Coverings which are not realized by algebraic maps} The fact that the epimorphism $\xi\colon \pi_1(U) \to \Z$ is realized by an algebraic map $f\colon U \to \C^*$ plays an essential role in proving the results of this paper. It is however natural to investigate (variants of) our original Question \ref{conj} in more topological contexts when an algebraic realization of $\xi$ is not readily available (e.g., if $U$ is a smooth complex projective variety). \bibliographystyle{plain}

https://arxiv.org/abs/0803.3164

Symmetric jump processes: localization, heat kernels, and convergence

We consider symmetric processes of pure jump type. We prove local estimates on the probability of exiting balls, the Hölder continuity of harmonic functions and of heat kernels, and convergence of a sequence of such processes.

\section{Introduction}\label{S:I} Suppose $J:{\mathbb R}^d\times {\mathbb R}^d\to [0,\infty)$ is a symmetric function satisfying $$ \frac{c_1}{|y-x|^{\beta_1}}\leq J(x,y)\leq \frac{c_2}{|y-x|^{\beta_2}} $$ if $|y-x|\leq 1$ and 0 otherwise. Define the Dirichlet form \begin{equation}\label{DFdedef} \sE(f,f)=\int\int (f(y)-f(x))^2 J(x,y)\, dy\, dx, \end{equation} and we take as the domain of $\sE$ the closure with respect to the norm $(\norm{f}_{L^2({\mathbb R}^d)}$ $+\sE(f,f))^{1/2}$ of the Lipschitz functions with compact support. When $\beta_1=\beta_2$, the Dirichlet form and associated infinitesimal generator are said to be of fixed order, namely, $\beta_1$, while if $\beta_1<\beta_2$, the generator is of variable order. The variable order case allows for considerable variability in the jump intensities and directions. In \cite{BBCK} a number of results were proved for the Hunt process $X$ associated with $\sE$, including exit probabilities, heat kernel estimates, a parabolic Harnack inequality, and the lack of continuity of harmonic functions. The last is perhaps the most interesting: it was shown that there exist harmonic functions that are not continuous. This paper could be considered a sequel to \cite{BBCK}, although the set of authors for the present paper neither contains nor is contained in the set of authors of \cite{BBCK}. We prove three main results, which we discuss in turn. First we discuss estimates on exit probabilities. In \cite{BBCK} some estimates were obtained on ${\mathbb P}^x(\tau_{B(x,r)}<t)$. These estimates held for all $x$, but were very crude, and were not sensitive to the behavior of $J(x,y)$ when $y$ is close to $x$. We show in the current paper that to a large extent the behavior of these exit probabilities depend on the size of $J(x,y)$ for $y$ near $x$. We also allow large jumps, which translates to allowing $J(x,y)\ne 0$ for $|y-x|>1$. In Example \ref{Ex1} we show how under some smoothness in $J$, we can get fairly precise estimates. Our motivation for obtaining better bounds on exit probabilities is to consider the question of when harmonic functions and the heat kernel are continuous. The example in \cite{BBCK} shows this continuity need not always hold. However, when $J$ possesses a minimal amount of smoothness, we establish that indeed harmonic functions are H\"older continuous, and the heat kernel is also H\"older continuous. The technique for showing the H\"older continuity of harmonic functions is based on ideas from \cite{BKa05}, where the non-symmetric case was considered. More interesting is the part of the proof where we show that H\"older continuity of harmonic functions plus global bounds on the heat kernel imply H\"older continuity of the heat kernel. This argument is of independent interest, and should be applicable in many other situations. Finally, we suppose we have a sequence of functions $J_n$ with corresponding Dirichlet forms and Hunt processes. We show that if the $J_n$ converge weakly to $J$, and some uniform integrability holds, then the corresponding processes converge. Note only weak convergence of the $J_n$ is needed. This is in contrast to the diffusion case, where it is known that weak convergence is not sufficient, and a much stronger type of convergence of the Dirichlet forms is required; see \cite{SZ}. Our assumptions and results are stated and proved in the next three sections, the exit probabilities in Section \ref{S:EP}, the regularity in Section \ref{S:R}, and the weak convergence in Section \ref{S:WC}. \section{Exit probabilities}\label{S:EP} Suppose $J:{\mathbb R}^d\times {\mathbb R}^d\to [0,\infty)$ is jointly measurable. We suppose throughout this paper that there exist constants $\kappa_1,\kappa_2, \kappa_3>0$ and $\beta_1, \beta_2\in (0,2)$ such that \begin{equation}\label{E01} \frac{\kappa_1}{|x-y|^{d+\beta_1}}\leq J(x,y)\leq \frac{\kappa_2}{|x-y|^{d+\beta_2}}, \qq |x-y|\leq 1, \end{equation} and \begin{equation}\label{E02} \int_{|x-y|>1} J(x,y)\, dy\leq \kappa_3, \qq x\in {\mathbb R}^d. \end{equation} The constants $\beta_1, \beta_2, c_1, c_2, c_3$ play only a limited role in what follows and \eqref{E01} and \eqref{E02} are used to guarantee a certain amount of regularity. Much more important is the ${\alpha}$ that is introduced in \eqref{E1}. Define a Dirichlet form $\sE=\sE_J$ by \begin{equation}\label{E03} \sE(f,f)=\int\int (f(y)-f(x))^2 J(x,y)\, dy\, dx, \end{equation} where we take the domain to be the closure of the Lipschitz functions with compact support with respect to the norm $(\norm{f}_2+(\sE(f,f))^{1/2}$. Let $X$ be the Hunt process associated with the Dirichlet form $\sE$. Let $B(x,r)$ denote the open ball of radius $r$ centered at $x$. The letter $c$ with or without subscripts will denotes constants whose exact values are unimportant and which may change from line to line. We remark that if we define $J_1(x,y)=J(x,y)1_{(|x-y|\leq 1)}$ and define the corresponding Dirichlet form in terms of $J_1$, then the Hunt process $X^{(1)}$ corresponding to this Dirichlet form is conservative by \cite[Theorem 1.1]{BBCK}. Using a construction due to Meyer (see \cite[Remark 3.4]{BBCK} and \cite[Section 3.1]{BGK}) we can use $X^{(1)}$ to obtain $X$. This is a probabilistic procedure that involves adding jumps. Only finitely many jumps are added in any finite time interval, and we deduce from this construction that $X$ is also conservative. We now fix $z_0\in {\mathbb R}^d$ and assume that there exist constants $\kappa_4$ and ${\alpha}\in (0,2)$ such that \begin{equation}\label{E1} J(x,y) \geq \kappa_4 |x-y|^{-d-{\alpha}}, \qq x,y\in B(z_0,3r). \end{equation} Here ${\alpha}$ may depend on $z_0$. Define \begin{equation}\label{E2} L_1(x,s)=\int_{|x-w|\geq s} J(x,w)\, dw, \end{equation} \begin{equation}\label{E3} L_2(x,s) =\int_{|x-w|\leq s} |x-w|^2 J(x,w)\, dw, \end{equation} and let \begin{equation}\label{E4} L(z_0,r) =\sup_{x\in B(z_0,3r)} L_1(x,r)+\sup_{x\in B(z_0,3r)} \sup_{s\leq r}s^{d} [s^{-2}L_2(x,s)]^{\frac{d+{\alpha}}{{\alpha}}}. \end{equation} From \eqref{E1} we see that \begin{equation}\label{E4.5} L(z_0,r)\geq cr^{-{\alpha}}. \end{equation} \begin{theorem}\label{T1} Suppose \eqref{E01}, \eqref{E02}, and \eqref{E1} hold. There exists $c_1$ (depending only on $d$, $\kappa_4$, and ${\alpha}$) such that if $r\in (0,1)$, then for $x\in B(z_0,r)$, $${\mathbb P}^x(\tau_{B(x,r)}<t)\leq c_1tL(z_0,r).$$ \end{theorem} {\medskip\noindent {\bf Proof. }} Let $x_0,y_0$ be fixed, let $R=|y_0-x_0|$, and suppose $R\geq 18(d+{\alpha})r/{\alpha}$. By \eqref{E4.5}, the result is immediate if $t>r^{\alpha}$, so let us suppose $t\leq r^{\alpha}$. Define \begin{equation}\label{E5} \widetilde J(x,y)=\begin{cases} J(x,y) & \mbox{ if }x,y \mbox{ are both in }B(z_0,3r) \mbox{ and } |x-y|< R,\\ \kappa_4|x-y|^{-d-{\alpha}}& \mbox{ if at least one of $x$ and $y$ is not in }\\ &\qq B(z_0,3r) \mbox{ and } |x-y|< R,\\ 0 &\mbox{ otherwise.} \end{cases} \end{equation} Define $$\widetilde J_\delta(x,y)=\widetilde J(x,y) 1_{(|x-y|\leq \delta)},$$ where we will choose $\delta \in [6r, R)$ in a moment. Let $\widetilde X$ be the Hunt process corresponding to $\widetilde J$ and $\widetilde X^{(\delta)}$ the Hunt process associated with $\widetilde J_\delta$. We have the Nash inequality (see, e.g., (3.9) of \cite{BBCK}): \begin{equation}\label{E6} \norm{u}_2^{2+\frac{2{\alpha}}{d}} \leq c\Big(\int\int_{|x-y|<\delta} \frac{(u(x)-u(y))^2}{|x-y|^{d+{\alpha}}}\, dy\, dx +\delta^{-{\alpha}}\|u\|_2^2\Big) \, \norm{u}_1^{2{\alpha}/d}. \end{equation} Using \eqref{E1} we obtain from this that \begin{equation}\label{E7} \norm{u}_2^{2+\frac{2{\alpha}}{d}} \leq c\Big(\int\int (u(x)-u(y))^2 \widetilde J_\delta (x,y)\, dy\, dx +\delta^{-{\alpha}} \norm{u}_2^2\Big) \, \norm{u}_1^{2{\alpha}/d}. \end{equation} Let \begin{align} \delta&=\frac{R{\alpha}}{3(d+{\alpha})},\label{E71}\\ N(\delta)&=\delta^{-{\alpha}}+\sup_{x\in B(z_0,3r)}\delta^{-2}L_2(x,\delta), \label{E72}\\ {\lambda}&=\frac{1}{3\delta} \log(1/(N(\delta)t)). \label{E73} \end{align} Let $\psi(x)={\lambda}(R-|x-x_0|)^+$. Set $$\Gamma(f,f)(x)=\int (f(y)-f(x))^2 \widetilde J_\delta(x,y)\, dy.$$ Since $|e^t-1|^2\leq t^2 e^{2t}$, $|\psi(x)-\psi(y)|\leq {\lambda} |x-y|$, and $\widetilde J_\delta(x,y)=0$ unless $|x-y|<\delta$, then \begin{align} e^{-2\psi(x)}\Gamma(e^\psi, e^\psi)(x)&=\int_{|x-y|\leq \delta} \Big( e^{\psi(x)-\psi(y)}-1\Big)^2 \widetilde J_\delta(x,y)\, dy{\nonumber}\\ &\leq e^{2{\lambda} \delta} {\lambda}^2 \int_{|x-y|\leq \delta} |x-y|^2 \widetilde J_\delta(x,y)\, dy.\label{E74} \end{align} Since $\delta \geq 6r$, then by our definition of $\widetilde J$ we have that the integral on the last line of \eqref{E74} is bounded by $\sup_{x\in B(z_0,3r)}L_2(x,\delta)+\delta^{2-{\alpha}}.$ We therefore have \begin{align*} e^{-2\psi(x)}\Gamma(e^\psi, e^\psi)(x) &\leq e^{2{\lambda} \delta} {\lambda}^2 \delta^2 N(\delta)\\ &\leq e^{3{\lambda} \delta} N(\delta). \end{align*} We obtain in the same way the same upper bound for $e^{2\psi(x)} \Gamma (e^{-\psi}, e^{-\psi})(x)$. So by \cite[Theorem 3.25]{CKS} we have \begin{equation}\label{E9} p_\delta(t,x_0,y_0)\leq ct^{-d/{\alpha}} e^{ct\delta^{-{\alpha}}} e^{-{\lambda} R+c e^{3{\lambda} \delta}N(\delta)t}, \end{equation} where $p_\delta$ is the transition density for $\widetilde X^{(\delta)}$. (Note that by \cite[Theorem 3.1]{BBCK}, the transition density $p_\delta(t,x,y)$ exists for $x,y\in {\mathbb R}^d\setminus {\cal N}$, where $\cal N$ is a set of capacity zero, called a properly exceptional set. We will take $x_0,y_0\in {\mathbb R}^d\setminus {\cal N}$.) Since $t\leq r^{\alpha}\leq c\delta^{{\alpha}}$, we then get $$p_\delta(t,x_0,y_0)\leq ct^{-d/{\alpha}} e^{-{\lambda} R} =ct^{-d/{\alpha}} (N(\delta)t)^{R/3\delta}.$$ Our bound now becomes \begin{align*} p_\delta(t,x_0,y_0)&\leq ct^{-d/{\alpha}} t^{1+\frac{d}{{\alpha}}} N(\delta)^{(d+{\alpha})/{\alpha}}\\ &= ct N(\delta)^{(d+{\alpha})/{\alpha}}. \end{align*} Since $\delta\geq 6r$, then $$\norm{\widetilde J-\widetilde J_\delta}_\infty \leq c\delta^{-(d+{\alpha})},$$ so by \cite[Lemma 3.1]{BGK} and by (\ref{E71}) \begin{align*} p(t,x_0,y_0)&\leq p_\delta(t,x_0,y_0)+ct\delta^{-(d+{\alpha})}\\ &\leq ct[\sup_{x\in B(z_0,3r)}\delta^{-2} L_2(x,\delta)+\delta^{-{\alpha}}]^{\frac{d+{\alpha}}{{\alpha}}}+ct R^{-(d+{\alpha})}. \end{align*} Since \begin{align*} \sup_{x\in B(z_0,3r)}\delta^{-2}L_2(x,\delta)+\delta^{-{\alpha}} &\leq c\delta^{-2}[ \sup_{x\in B(z_0,3r)}L_2(x,r)+\delta^{2-{\alpha}}]\\ &\leq cR^{-2}\sup_{x\in B(z_0,3r)}L_2(x,r)+cR^{-{\alpha}}, \end{align*} then, because $\widetilde X$ is conservative, integrating over $R\geq r/2$ gives us $${\mathbb P}^{x_0}(|\widetilde X_t-x_0|\geq r/2)\leq ctr^{d} \Big[r^{-2} \sup_{x\in B(z_0,3r)}L_2(x,r)\Big]^{\frac{d+{\alpha}}{{\alpha}}}+ctr^{-{\alpha}}\le ctL(z_0,r).$$ By \cite[Lemma 3.8]{BBCK} we then have $${\mathbb P}^{x_0}(\sup_{s\leq t} |\widetilde X_s-x_0|>r)\leq ctL(z_0,r).$$ We now use Meyer's construction to compare $\widetilde X$ to $X$. Using this construction we obtain, for $x\in B(z_0,r)$, \begin{align*} {\mathbb P}^x( X_s\ne \widetilde X_s\mbox{ for some }s\leq t)&\leq t\sup_{x'\in B(z_0,2r)}\int_{B(z_0,3r)^c}|J(x',y)-\widetilde J(x',y)|dy\\ &\le ctL(z_0,r). \end{align*} (The first inequality can be obtained by observing the processes $X$ and $\widetilde X$ killed on exiting $B(z_0, 2r)$.) Therefore, for $x\in B(z_0,r)$, \begin{align*} {\mathbb P}^x(\sup_{s\leq t}|X_s-x|>r) &\leq {\mathbb P}^x(\sup_{s\leq t} |\widetilde X_s-x|>r)+{\mathbb P}^x(X_s\ne \widetilde X_s \mbox{ for some }s\leq t)\\ &\leq ctL(z_0,r). \end{align*} {\hfill $\square$ \bigskip} \begin{corollary}\label{C1} Suppose \eqref{E01} and \eqref{E02} hold. Suppose instead of \eqref{E1} we have \begin{equation}\label{E101} J(x,y)\geq \kappa_4 |x-y|^{-d-{\alpha}}-K(x,y), \qq x,y\in B(z_0,3r), \end{equation} where \begin{equation}\label{E102} \int_{|x-y|\leq \delta} K(x,y)\, dy\leq \kappa_5 \delta^{-{\alpha}} \end{equation} for all $x\in B(z_0,3r)$ and all $\delta\leq r$. Then the conclusion of Theorem \ref{T1} still holds. \end{corollary} {\medskip\noindent {\bf Proof. }} The only place the lower bound on $J(x,y)$ plays a role is in deriving \eqref{E7} from \eqref{E6}. If we have \eqref{E101} instead of \eqref{E1}, then in place of \eqref{E7} we now have \begin{align} \norm{u}_2^{2+\frac{2{\alpha}}{d}} &\leq c\Big(\int\int (u(x)-u(y))^2 \widetilde J_\delta (x,y)\, dy\, dx\label{E103}\\ &\qq +\int\int_{|x-y|\leq \delta} (u(x)-u(y))^2 K(x,y)\, dy\, dx +\delta^{-{\alpha}} \norm{u}_2^2\Big) \, \norm{u}_1^{2{\alpha}/d}.{\nonumber} \end{align} But by our assumption on $K(x,y)$, the double integral with $K$ in the integrand is bounded by $$ \int \Big(\int_{|x-y|\leq \delta} K(x,y)\, dy\Big)\ u(x)^2\, dx \leq c\delta^{-{\alpha}}\norm{u}_2^2.$$ {\hfill $\square$ \bigskip} \begin{example}\label{Ex1}{\rm Suppose $\varepsilon>0$ and there exists a function $s:{\mathbb R}^d\to (\varepsilon, 2-\varepsilon)$ such that \begin{equation}\label{E1031} |s(x)-s(y)|\leq c \log(2/|x-y|), \qq |x-y|<1. \end{equation} Suppose there exist constants $c_1, c_2$ such that \begin{equation}\label{E1032}\frac{c_1}{|x-y|^{d+s(x)\land s(y)}}\leq J(x,y) \leq \frac{c_2}{|x-y|^{d+s(x)\lor s(y)}}. \end{equation} Suppose further that (\ref{E02}) holds. We show that $L(z_0,r)$ is comparable to $r^{-s(z_0)}$ if $r<1$. To see this, note that \begin{equation}\label{E104} |x-y|^{s(x)-s(y)}\leq |x-y|^{-c/\log(2/|x-y|)}\leq e^c, \end{equation} if $|x-y|\leq 1$ and similarly we have \begin{equation}\label{E105}|x-y|^{s(x)-s(y)}\geq |x-y|^{c/\log(2/|x-y|)}\geq e^{-c}. \end{equation} If we fix $x$ and let $$M(v)=\sup_{|x-w|=v} J(x,w),$$ then for $v\leq 1$ \begin{align*} M(v)&\leq \sup_{|x-w|=v} \frac{c}{v^{d+s(x)}} v^{-|s(x)-s(w)|}\\ &\leq \frac{c}{v^{d+s(x)}}. \end{align*} We then estimate for $r\leq 1$ \begin{align*} L_2(x,r)&\leq c\int_0^r v^2 M(v) v^{d-1}\, dv\\ &\leq c\int_0^r v^{1-s(x)}\, dv=cr^{2-s(x)}, \end{align*} We can similarly obtain a bound for $L_1(x,r)$: \begin{align*} L_1(x,r)&\leq c\int_r^1 M(v)v^{d-1}\, dv+\int_{|x-w|>1}J(x,w)\, dw\\ &\leq c\int_r^1 v^{-1-s(x)} \, dv+c\\ &\leq cr^{-s(x)}+c\leq cr^{-s(x)} \end{align*} if $r\leq 1$. Next, for $x\in B(z_0,3r)$, we have $r^{-s(x)}$ is comparable to $r^{-s(z_0)}$ for $r\leq 1$. To see this, $$c\leq r^{s(x)-s(z_0)}\leq r^{-|s(x)-s(z_0)|}\leq c'$$ as in \eqref{E104} and \eqref{E105}. If we take ${\alpha}$ in \eqref{E1} to be $\inf_{x\in B(z_0,3r)} s(x)$, then we conclude $$L(x,r)\leq cr^{-s(z_0)}+cr^{-{\alpha}}\leq cr^{-s(z_0)},$$ so $${\mathbb P}^x(\tau_r\leq t)\leq ct r^{-s(z_0)}, \qq x\in B(z_0,r).$$ } \end{example} \section{Regularity}\label{S:R} We suppose throughout this section that \eqref{E01} and \eqref{E02} hold. We suppose in addition first that there exists $c$ such that \begin{equation}\label{C01} \int_A J(z,y)\, dy\geq c L(x,r) \end{equation} whenever $r\in (0,1)$, $A\subset B(x,3r)$, $|A|\geq \frac13 |B(x,r)|$, $x\in {\mathbb R}^d$, and $z\in B(x,r/2)$ and second there exist $\sigma$ and $c$ such that \begin{equation}\label{C02} \frac{L_1(x,{\lambda} r)}{L_1(x,r)}\leq c{\lambda}^{-\sigma}, \qq x\in {\mathbb R}^d, r\in (0,1), {\lambda}\in (1,1/r). \end{equation} It is easy to check that \eqref{C01} and \eqref{C02} hold for Example \ref{Ex1}. We say a function $h$ is harmonic in a ball $B(x_0,r)$ if $h(X_{t\land \tau_{B(x_0,r(1-\varepsilon))}})$ is a ${\mathbb P}^x$ martingale for q.e. $x$ and every $\varepsilon\in (0,1)$. \begin{theorem}\label{T.R1} Suppose \eqref{E01}, \eqref{E02}, \eqref{C01}, and \eqref{C02} hold. There exist $c_1$ and $\gamma$ such that if $h$ is bounded in ${\mathbb R}^d$ and harmonic in a ball $B(x_0,r)$, then \begin{equation}\label{C03} |h(x)-h(y)|\leq c_1\Big(\frac{|x-y|}{r}\Big)^\gamma\norm{h}_\infty, \qq x,y\in B(x_0,r/2). \end{equation} \end{theorem} {\medskip\noindent {\bf Proof. }} As in \cite{CK03, CK07} we have the L\'evy system formula: \begin{equation}\label{C04} {{\mathbb E}\,}^x\Big[\sum_{s\leq T} f(X_{s-},X_s)\Big]= {{\mathbb E}\,}^x\Big[\int_0^T \Big(\int f(X_s,y) J(X_s,y)\, dy\Big)\, ds\Big] \end{equation} for any nonnegative $f$ that is 0 on the diagonal, for every bounded stopping time $T$, and q.e. starting point $x$. Given this, the proof is nearly identical to that in \cite[Theorem 2.2]{BKa05}. {\hfill $\square$ \bigskip} We obtain a crude estimate on the expectation of the exit times. \begin{lemma}\label{L.R2} Assume the lower bound of \eqref{E01}. Then there exists $c_1$ such that $${{\mathbb E}\,}^x \tau_r\leq c_1 r^{\beta_1}, \qq x\in {\mathbb R}^d, r\in (0,1/2).$$ \end{lemma} {\medskip\noindent {\bf Proof. }} The expression $\sum_{s\leq t\land \tau_r} 1_{(|X_s-X_{s-}|>2r)}$ is 1 if there is a jump of size at least $2r$ before time $t\land \tau_r$, in which case the process exits $B(x,r)$ before or at time $t$, or 0 if there is no such jump. So \begin{align*} {\mathbb P}^x(\tau_r\leq t)&\geq {{\mathbb E}\,}^x \sum_{s\leq t\land \tau_r} 1_{(|X_s-X_{s-}|>2r)}\\ &={{\mathbb E}\,}^x \int_0^{t\land \tau_r} \int_{B(x,2r)^c} J(X_s,y)\, dy\, ds\\ &\geq cr^{-\beta_1} {{\mathbb E}\,}^x[t\land \tau_r]\\ &\geq cr^{-\beta_1}t{\mathbb P}^x(\tau_r>t), \end{align*} using the lower bound of \eqref{E01}. Thus $${\mathbb P}^x(\tau_r>t)\leq 1-cr^{-\beta_1}t {\mathbb P}^x(\tau_r>t),$$ or ${\mathbb P}^x(\tau_r>t)\leq 1/2$ if we take $t=c^{-1}r^{\beta_1}$. This holds for every $x\in {\mathbb R}^d$. Using the Markov property at time $mt$, $${\mathbb P}^x(\tau_r>(m+1)t)\leq {{\mathbb E}\,}^x[{\mathbb P}^{X_{mt}}(\tau_r>t); \tau_r>mt]\leq \tfrac12 {\mathbb P}^x(\tau_r>mt).$$ By induction ${\mathbb P}^x(\tau_r>mt)\leq 2^{-m}.$ With this choice of $t$, our lemma follows. {\hfill $\square$ \bigskip} We next show ${\lambda}$-potentials are H\"older continuous. Let $$U^{\lambda} f(x)={{\mathbb E}\,}^x\int_0^\infty e^{-{\lambda} t} f(X_t)\, dt.$$ \begin{proposition}\label{PR3} Under the same assumption as in Theorem \ref{T.R1}, there exist $c_1=c_1(\lambda)$ and $\gamma'$ such that if $f$ is bounded, then $$|U^{\lambda} f(x)-U^{\lambda} f(y)|\leq c_1 |x-y|^{\gamma'} \norm{f}_\infty.$$ \end{proposition} {\medskip\noindent {\bf Proof. }} Fix $x_0$, let $r\in (0,1/2)$, and suppose $x,y\in B(x_0,r/2)$. By the strong Markov property, \begin{align*} U^{\lambda} f(x)&={{\mathbb E}\,}^x\int_0^{\tau_r} e^{-{\lambda} t} f(X_t) \, dt +{{\mathbb E}\,}^x (e^{-{\lambda} \tau_r}-1)U^{\lambda} f(X_{\tau_r})\\ &\qq + {{\mathbb E}\,}^x U^{\lambda} f(X_{\tau_r})\\ &=I_1+I_2+I_3, \end{align*} and similarly when $x$ is replaced by $y$. We have by Lemma \ref{L.R2} $$|I_1|\leq \norm{f}_\infty {{\mathbb E}\,}^x \tau_r\leq cr^{\beta_1}\norm{f}_\infty$$ and by the mean value theorem and Lemma \ref{L.R2} $$|I_2|\leq {\lambda} {{\mathbb E}\,}^x \tau_r \norm{U^{\lambda} f}_\infty\leq cr^{\beta_1}\norm{f}_\infty,$$ and similarly when $x$ is replaced by $y$. So \begin{equation}\label{C05} |U^{\lambda} f(x)-U^{\lambda} f(y)|\leq cr^{\beta_1}\norm{f}_\infty+|{{\mathbb E}\,}^x U^{\lambda} f(X_{\tau_r}) -{{\mathbb E}\,}^y U^{\lambda} f(X_{\tau_r})|. \end{equation} But $z\to {{\mathbb E}\,}^z U^{\lambda} f(X_{\tau_r})$ is bounded in ${\mathbb R}^d$ and harmonic in $B(x_0,r)$, so by Theorem \ref{T.R1} the second term in \eqref{C05} is bounded by $$c\Big(\frac{|x-y|}{r}\Big)^\gamma \norm{U^{\lambda} f}_\infty.$$ If we use $\norm{U^{\lambda} f}_\infty\leq \frac1{{\lambda}}\norm{f}_\infty$ and set $r=|x-y|^{1/2}$, then \begin{equation}\label{C06}|U^{\lambda} f(x)-U^{\lambda} f(y)|\leq (c|x-y|^{\beta_1/2}+c|x-y|^{\gamma/2})\norm{f}_\infty, \end{equation} and our result follows. {\hfill $\square$ \bigskip} Using the spectral theorem, there exists projection operators $E_\mu$ on the space $L^2({\mathbb R}^d, dx)$ such that \begin{align} f&=\int_0^\infty \, dE_\mu(f),{\nonumber}\\ P_tf&=\int_0^\infty e^{-\mu t} \, dE_\mu(f),{\nonumber}\\ U^{\lambda} f&=\int_0^\infty \frac{1}{{\lambda}+\mu} \, dE_\mu(f).\label{C07} \end{align} \begin{proposition}\label{PR4} Under the same assumptions as in Theorem \ref{T.R1}, if $f$ is in $L^2$, then $P_tf$ is equal a.e. to a function that is H\"older continuous. \end{proposition} {\medskip\noindent {\bf Proof. }} Write $\brack{f,g}$ for the inner product in $L^2$. Note that in what follows $t$ is fixed. Each of our constants may depend on $t$. If $X^{(1)}$ is the Hunt process associated with the Dirichlet form defined in terms of the kernel $J_1(x,y)=J(x,y)1_{(|x-y|<1)}$, we know from \cite[Theorem 2.1]{BBCK} that $X^{(1)}$ has a transition density $p(t,x,y)$ bounded by $c$. Using \cite[Lemma 3.1]{BGK} and Meyer's construction (cf.\ \cite[Section 3]{BBCK}), we then can conclude that $X$ also has a transition density bounded by $c$. Define $$h=\int_0^\infty ({\lambda}+\mu)e^{-\mu t}\, dE_\mu(f).$$ Since $\sup_\mu ({\lambda}+\mu)^2 e^{-2\mu t}\leq c$, then $$\int_0^\infty ({\lambda}+\mu)^2 e^{-2\mu t} \, d\brack{E_\mu(f), E_\mu(f)} \leq c\int_0^\infty \, d\brack{E_\mu(f), E_\mu(f)}=c\norm{f}_2^2,$$ we see that $h$ is a well defined function in $L^2$. Suppose $g\in L^1$. Then $\norm{P_t g}_1\leq \norm{g}_1$ by Jensen's inequality, and $$|P_tg(x)|=\Big|\int p(t,x,y)g(y)\, dy\Big|\leq c\norm{g}_1$$ by the fact that $p(t,x,y)$ is bounded. So $\norm{P_tg}_\infty\leq c\norm{g}_1$, and it follows that $\norm{P_tg}_2\leq c\norm{g}_1$. Using Cauchy-Schwarz and the fact that $$\sup_\mu ({\lambda}+\mu) e^{-\mu t/2}\leq c<\infty,$$ we have \begin{align*} \brack{h,g}&=\int_0^\infty ({\lambda}+\mu)e^{-\mu t}\, d\brack{E_\mu(f), E_\mu(g)}\\ &\leq \Big(\int_0^\infty ({\lambda}+\mu)e^{-\mu t} \, d\brack{E_\mu(f), E_\mu(f)}\Big)^{1/2}\\ &\qq\qq\times \Big(\int_0^\infty ({\lambda}+\mu)e^{-\mu t} \, d\brack{E_\mu(g), E_\mu(g)}\Big )^{1/2}\\ &\leq c\Big(\int_0^\infty \, d\brack{E_\mu(f), E_\mu(f)}\Big )^{1/2} \Big(\int_0^\infty e^{-\mu t/2} \, d\brack{E_\mu(g), E_\mu(g)}\Big )^{1/2}\\ &=c \norm{f}_2\norm{P_{t/2}g}_2\\ &\leq c \norm{f}_2 \norm{g}_1. \end{align*} Taking the supremum over $g\in L^1$ with $L^1$ norm less than 1, $\norm{h}_\infty\leq c\norm{f}_2$. But by \eqref{C07} $$U^{\lambda} h=\int_0^\infty e^{-\mu t} \, dE_\mu(f)=P_tf,\qq a.e.,$$ and the H\"older continuity of $P_tf$ follows by Proposition \ref{PR3}. {\hfill $\square$ \bigskip} Finally we have \begin{theorem}\label{PT6} Under the same assumption as in Theorem \ref{T.R1}, we can choose $p(t,x,y)$ to be jointly continuous. \end{theorem} {\medskip\noindent {\bf Proof. }} Fix $y$ and let $f(z)=p(t/2,z,y)$. $f$ is bounded by $c$ (depending on $t$) and has $L^1$ norm equal to 1, hence $f\in L^2$ with norm bounded by $c$. Note $$P_{t/2}f(x)=\int p(t/2,x,z) f(z)\, dz=\int p(t/2,x,z)p(t/2,z,y) =p(t,x,y).$$ Using Proposition \ref{PR4} shows that $p(t,x,y)$ is H\"older continuous with constants independent of $x$ and $y$. This and symmetry gives the result. {\hfill $\square$ \bigskip} \begin{remark}\label{R3.6}{\rm The argument we gave deriving the H\"older continuity of the transition densities from the boundedness of the transition densities plus the H\"older continuity of harmonic functions holds in much more general contexts than just jump processes in ${\mathbb R}^d$.} \end{remark} \section{Convergence}\label{S:WC} Suppose now that we have a sequence of jump kernels $J^n(x,y)$ satisfying \eqref{E01}, \eqref{E02}, \eqref{C01}, and \eqref{C02} with constants independent of $n$. Suppose in addition that \begin{equation}\label{WCA1} \limsup_{\eta\to 0} \sup_{n,x} \int_{|y-x|\geq \eta^{-1}} J_n(x,y)\, dy\, dx=0, \end{equation} \begin{equation}\label{WCA2} \limsup_{\eta\to 0} \sup_{n,x} \int_{|y-x|\leq \eta} |y-x|^2 J_n(x,y)\, dy\, dx=0, \end{equation} and for almost every $\eta$ \begin{equation}\label{WCA3} J_n(x,y)1_{(\eta,\eta^{-1})}(|y-x|)\, dx\, dy\to J(x,y)1_{(\eta,\eta^{-1})}(|y-x|)\, dx\, dy \end{equation} weakly as $n\to \infty$. Let $\sE^n$ be the Dirichlet forms defined in terms of the $J^n$ with $P_t^n$, $U^{\lambda}_n$, and ${\mathbb P}^x_n$ the associated semigroup, resolvent, and probabilities. Let $P_t$, $U^{\lambda}$, and ${\mathbb P}^x$ be the semigroup, resolvent, and probabilities corresponding to the Dirichlet form $\sE_J$ defined in terms of the kernel $J$. Under the above set-up we have \begin{theorem}\label{WT1} If $f$ is bounded and continuous, then $P_t^n f$ converges uniformly on compacts to $P_tf$. For each $t$, for q.e.\ $x$, ${\mathbb P}^x_n$ converges weakly to ${\mathbb P}^x$ with respect to the space $D([0,t])$. \end{theorem} {\medskip\noindent {\bf Proof. }} The first step is to show that any subsequence $\{n_j\}$ has a further subsequence $\{n_{j_k}\}$ such that $U^{\lambda}_{n_{j_k}}f$ converges uniformly on compacts whenever $f$ is bounded and continuous. The proof of this is very similar to that of \cite[Proposition 6.2]{BKu08}, and we refer the reader to that paper. Now suppose we have a subsequence $\{n'\}$ such that the $U^{\lambda}_{n'} f$ are equicontinuous and converge uniformly on compacts whenever $f$ is bounded and continuous with compact support. Fix such an $f$ and let $H$ be the limit of $U^{\lambda}_{n'}f$. We will show \begin{equation}\label{WCE1} \sE_J(H,g)=\angel{f,g}-{\lambda}\angel{H,g} \end{equation} whenever $g$ is a Lipschitz function with compact support, where $\sE_J$ is the Dirichlet form corresponding to the kernel $J$. This will prove that $H$ is the ${\lambda}$-resolvent of $f$ with respect to $\sE_J$, that is, $H=U^{\lambda} f$. We can then conclude that the full sequence $U^{\lambda}_n f$ converges to $U^{\lambda} f$ whenever $f$ is bounded and continuous with compact support. The assertions about the convergence of $P_t^n$ and ${\mathbb P}^x_n$ then follow as in \cite[Proposition 6.2]{BKu08}. So we need to prove $H$ satisfies \eqref{WCE1}. We drop the primes for legibility. We know \begin{equation}\label{WCE3} \sE^n(U^{\lambda}_n f, U^{\lambda}_n f)=\angel{f,U^{\lambda}_n f}-{\lambda} \angel{U^{\lambda}_n f, U^{\lambda}_n f}. \end{equation} Since $\norm{U^{\lambda}_n f}_2\leq (1/{\lambda}) \norm{f}_2$ (by Jensen's inequality), we have by Cauchy-Schwarz that $$\sup_n \sE^n (U^{\lambda}_n f, U^{\lambda}_n f)\leq c<\infty.$$ Since the $U^{\lambda}_n f$ are equicontinuous and converge uniformly to $H$ on $B(0, \eta^{-1})-\overline{B(0,\eta)}$ for almost every $\eta$, then \begin{align*} \int\int_{\eta<|y-x|<\eta^{-1}} &(H(y)-H(x))^2 J(x,y)\, dy\, dx\\ &\leq \limsup_{n\to \infty} \int\int_{\eta<|y-x|<\eta^{-1}} (U^{\lambda}_n f(y)-U^{\lambda}_n f(x))^2 J_n(x,y)\, dy\, dx\\ &\leq \limsup_n \sE^n(U^{\lambda}_n f, U^{\lambda}_n f)\leq c<\infty. \end{align*} Letting $\eta\to 0$ (while avoiding the null set), we have \begin{equation}\label{WCE2} \sE_J(H,H)<\infty. \end{equation} Fix a Lipschitz function $g$ with compact support and choose $M$ large enough so that the support of $g$ is contained in $B(0,M)$. Then \begin{align*} \Big| \int\int_{|y-x|\geq \eta^{-1}} & (U^{\lambda}_n f(y)-U^{\lambda}_n f(x))(g(y)-g(x)) J_n(x,y) \, dy\, dx\Big|\\ &\leq \Big( \int\int (U^{\lambda}_n f(y)-U^{\lambda}_n f(x))^2 J_n(x,y)\, dy\, dx\Big)^{1/2}\\ &\qq\times \Big(\int\int_{|y-x|\geq \eta^{-1}} (g(y)-g(x))^2 J_n(x,y)\, dy\, dx\Big)^{1/2}. \end{align*} The first factor is $(\sE^n(U^{\lambda}_n f, U^{\lambda}_nf))^{1/2}$, while the second factor is bounded by $$\norm{g}_\infty \Big(\int_{B(0,M)}\int_{|y-x|\geq \eta^{-1}} J_n(x,y)\, dx\, dy\Big)^{1/2},$$ which, in view of \eqref{WCA1}, will be small if $\eta$ is small. Similarly, \begin{align*} \Big| \int\int_{|y-x|\leq \eta} & (U^{\lambda}_n f(y)-U^{\lambda}_n f(x))(g(y)-g(x)) J_n(x,y) \, dy\, dx\Big|\\ &\leq \Big( \int\int (U^{\lambda}_n f(y)-U^{\lambda}_n f(x))^2 J_n(x,y)\, dy\, dx\Big)^{1/2}\\ &\qq\times \Big(\int\int_{|y-x|\leq \eta} (g(y)-g(x))^2 J_n(x,y)\, dy\, dx\Big)^{1/2}. \end{align*} The first factor is as before, while the second is bounded by $$\norm{{\nabla} g}_\infty \Big(\int_{B(0,M)} \int_{|y-x|\leq \eta}|y-x|^2 J_n(x,y)\, dx\, dy\Big)^{1/2}.$$ In view of \eqref{WCA2}, the second factor will be small if $\eta$ is small. Similarly, using \eqref{WCE2}, we have $$\Big|\int\int_{|y-x|\notin (\eta,\eta^{-1})} (H(y)-H(x))(g(y)-g(x)) J(x,y)\, dy\, dx\Big|$$ will be small if $\eta$ is taken small enough. By \eqref{WCA3}, \eqref{E01}, \eqref{E02}, and the fact that the $U^{\lambda}_nf$ are equicontinuous and converge to $H$ uniformly on compacts, for almost every $\eta$ \begin{align*} \int\int_{|y-x|\in (\eta, \eta^{-1})}& (U^{\lambda}_nf(y)-U^{\lambda}_nf(x))(g(y)-g(x))J_n(x,y)\, dy\, dx\\ &\to \int\int_{|y-x|\in (\eta, \eta^{-1})} (H(y)-H(x))(g(y)-g(x))J(x,y)\, dy\, dx. \end{align*} It follows that \begin{equation}\label{WCE4} \sE^n(U^{\lambda}_nf,g)\to \sE_J(H,g). \end{equation} But $$\sE^n(U^{\lambda}_nf,g)=\angel{f,g}-{\lambda}\angel{U^{\lambda}_nf,g}\to \angel{f,g}-{\lambda}\angel{H,g}.$$ Combining with \eqref{WCE4} proves \eqref{WCE1}. {\hfill $\square$ \bigskip} \begin{remark}\label{WR1} {\rm One can modify the above proof to obtain a central limit theorem for symmetric Markov chains. Suppose for each $n$ we have a symmetric Markov chain on $n^{-1}{\mathbb Z}^d$ with unbounded range with conductances $C^n_{xy}$. If $\nu_n$ is the measure that gives mass $n^{-d}$ to each point in $n^{-1}{\mathbb Z}^d$, then we can define the Dirichlet form $$\sE_n(f,f)=\sum_{x,y\in n^{-1}{\mathbb Z}^d} (f(x)-f(y))^2 C^n_{xy}$$ with respect to the measure $\nu_n$. Under appropriate assumptions analogous to those in Sections \ref{S:EP} and \ref{S:R}, one can show that the semigroups corresponding to $\sE_n$ converge to those of $\sE$ and in addition there is weak convergence of the probability laws. Since the details are rather lengthy, we leave this to the interested reader. } \end{remark} \begin{remark}\label{WR121} {\rm We can also prove the following approximation of a jump process by Markov chains, which is a generalization of \cite[Theorem 2.3]{HK07}. Suppose $J:{\mathbb R}^d\times {\mathbb R}^d\to [0,\infty)$ is a symmetric measurable function satisfying \eqref{E01}, \eqref{E02}, \eqref{C01}, and \eqref{C02}. Define the conductivity functions $C^n\colon n^{-1}{\mathbb Z}^d \times n^{-1}{\mathbb Z}^d \to[0,\infty)$ by \[ C^n(x,y) = n^{2d}\int_{|x-\xi|_\infty< \frac 1{2n}}\int_{|y-\zeta|_\infty< \frac 1{2n}} J(\xi,\zeta) d\xi\,d\zeta ~\text{ for }~ x\ne y\in n^{-1}{\mathbb Z}^d, \] and $C^n(x,x) = 0$, where $|x-y|_\infty=\max_{1\le i\le d}|x_i-y_i|$. Let $X$ be the Hunt process corresponding to the Dirichlet form given by (\ref{DFdedef}). Then the sequence of processes corresponding to $C^n$ converges weakly to $X$. Given Remark \ref{WR1}, the proof is standard. } \end{remark} \begin{remark}\label{WR221} {\rm As we mentioned at the beginning of Section 2, the assumptions \eqref{E01} and \eqref{E02} are used to guarantee a certain amount of regularity, namely, conservativeness and the existence of the heat kernel. However, one can relax these assumptions. All of the results in this paper hold if instead of \eqref{E01} and \eqref{E02} we assume \eqref{E02}, \eqref{E1} for all $z_0\in {\mathbb R}^d$ and the following: \[ \int_{|x-y|\le 1} |x-y|^2J(x,y)\, dy\leq \kappa_5, \qq x\in {\mathbb R}^d. \] } \end{remark}

https://arxiv.org/abs/1010.1941

Ramanujan-Type Series Related to Clausen's Product

Infinite series are evaluated through the manipulation of a series for $\cos(2t \sin^{-1}x)$ resulting from Clausen's Product. Hypergeometric series equal to an expression involving $\frac{1} {\pi}$ are determined. Techniques to evaluate generalized hypergeometric series are discussed through perspectives of experimental mathematics.

\section{Introduction} Clausen's product is discussed in the fascinating book by Borwein, Bailey, and Girgensohn, \emph{Experimentation in Mathematics: Computational Paths to Discovery} \cite{bbg}. New Ramanujan-type series derived from Clausen's product such as the following are given in this paper: $\sum_{n=2}^{\infty} ( \frac{1}{4})^{3n-1} \frac{(2n-3)!(4n+3)!}{((2n)!(n+1)(2n+1))^{2}(n-2)!n!}$; the following section ({\bf Ramanujan-Type Series}) explains the derivation of this series and a related series. We discuss the evaluation of hypergeometric series through integral substitution. For example, we shall soon consider an infinite array of hypergeometric series of the form $_{3} F _{2} (\frac{1}{2}, \frac{2k}{n},\frac{2n-2k}{n};\frac{3}{2},2;1)$. Furthermore, we discuss methods to incorporate the Riemann-zeta function into hypergeometric-like series, and discuss related new hypergeometric generalizations, evaluating hypergeometric series such as $_{6} F _{5} (1,1,\frac{9}{4},\frac{5}{2},\frac{11}{4},3;2,\frac{37}{16},\frac{41}{16},\frac{45}{16},\frac{49}{16};1)$. We are interested in the evaluation of infinite series. There has been much work discussing hypergeometric identities of the form $_{p} F _{q} (a_{1},a_{2},...,a_{p};b_{1},b_{2},...,b_{q};x)$ for constant $x$ such that $a_{1}$,$a_{2}$,...,$a_{p}$ and/or $b_{1}$,$b_{2}$,...,$b_{q}$ contain a variable/variables. There does not seem to be as much work on hypergeometric series of the form $_{p} F _{q} (a_{1},a_{2},...,a_{p};b_{1},b_{2},...,b_{q};x)$ for variable $x$, which can be manipulated through integration with respect to $x$ (and by replacing $x$ by $f(x)$ and integrating). For example, only series of the form $_{p} F _{q} (a_{1},a_{2},...,a_{p};b_{1},b_{2},...,b_{q};x)$ for constant $x$ are discussed in \cite{ira} (which is cited in \cite{ab}). ``Common", modern computational software is usually unable to directly evaluate hypergeometric series of the form $_{p} F _{q} (a_{1},a_{2},...,a_{p};b_{1},b_{2},...,b_{q};x)$ for variable $x$; we thus consider the evaluation of such series to be important. Consider infinite series of the form $\sum g(n) (1-x^a)^{bn} x^{dn} $, where $g(n)$ represents a combination of a finite number of elementary functions (and $a$, $b$, and $d$ are positive integers). Integration of the expression equal to such series is often very difficult, yielding complicated results, some of which we shall here present. The evaluation of a series such as $\sum_{k=0}^{\infty} \frac{4^k} {\binom{2k}{k}(2k+1)(k+1)^{2}}$ is an example of a tangible approach to exploring series representations of important constants through integral substitution. Although the theorems below have somewhat more cumbersome derivations, briefly describing our evaluation of the aforementioned series will hopefully serve as a straightforward introduction to this paper. The following well-known and important series (for $x^{2} \leq 1$) is given in \cite{gradsh}: $ (\sin^{-1} x )^{2} = \sum_{k=0}^{\infty} \frac{4^{k} (k!)^{2} x^{2k+2}} {(2k+1)!(k+1)}$. Divide both sides of this equation by $x$ and integrate: \begin{equation*} \int \frac{(\sin^{-1} x)^2}{x} dx = \frac{\operatorname{Li}_{3}(e^{2i \sin^{-1}x})}{2} + i \sin^{-1}x\operatorname{Li}_{2}(e^{2i \sin^{-1}x}) + \frac{1}{3} i (\sin^{-1} x)^{3} + (\sin^{-1} x)^{2} \ln (1-e^{2i \sin^{-1}x}) \end{equation*} Give the above integral an upper limit of one and a lower limit of zero, resulting in an integral evaluated in \cite{jchoi} (described as ``an interesting definite integral"). It is now clear that $\sum_{k=0}^{\infty} \frac{4^k} {\binom{2k}{k}(2k+1)(k+1)^{2}} = \frac{\pi^2}{2} \ln (2) - \frac{7 \zeta (3)}{4}$. The following related series is elegantly proven in \cite{borweinchamber}: $\sum_{n=1}^{\infty} \frac{4^n}{n^{3}\binom{2n}{n}} = \pi^{2} \ln(2) - \frac{7 \zeta (3)}{2}$. Let us continue by providing the infinite array of $_{3} F _{2}$ series we have in mind, which is based upon Clausen's product. Consider: \begin{equation*} (_{2} F _{1} (\frac{k}{n},\frac{n-k}{n};\frac{3}{2};\sin ^{2} x))^2 = ( \frac{ \csc x \sin ( (-1 + 2 (1 - \frac{k}{n})) \sin ^{-1 } (\sin x) ) }{-1 + 2 (1 - \frac{k}{n})} ) ^2 \end{equation*} \begin{theorem \begin{eqnarray*} \lefteqn{\int_{0}^{\frac{\pi}{2}} ( \frac{ \csc x \sin ( (-1 + 2 (1 - \frac{k}{n})) \sin ^{-1 } (\sin x) ) }{-1 + 2 (1 - \frac{k}{n})} ) ^2 dx = } \\ & & - \frac{1}{4 (-2 k + n)} i e^{ \frac{2 i k \pi}{n} } n ( \frac{n}{2 k - n} + \frac{e^{\frac{4 i k \pi}{n}} n}{2 k - n} - 2 e^{ \frac{2 i k \pi}{n} } \pi \cot ( \frac{2 k \pi}{n}) - \psi_{0} (\frac{1}{2} - \frac{k}{n}) + \psi_{0}( 1 - \frac{k}{n}) \\ & & + e^{\frac{4 i k \pi}{n}} \psi_{0} (-\frac{1}{2} + \frac{k}{n}) - e^{\frac{4 i k \pi}{n}} \psi_{0}(\frac{k}{n})) \end{eqnarray*} $=$ \begin{equation*} _{3} F _{2} (\frac{1}{2}, \frac{2k}{n},\frac{2n-2k}{n};\frac{3}{2},2;1) \frac{\pi}{2} \end{equation*} \end{theorem \begin{proof} Given Clausen's product, the above theorem holds. \end{proof} Perhaps it is probable that the above theorem has been considered before. Integrating $ ( _{2} F _{1} ( 2 - \frac{k}{n} , \frac{k}{n} ; \frac{5}{2} ; \sin ^{2} x ))^2$ proves to be much more cumbersome, involving multifarious hypergeometric series. Now, let us briefly discuss some more simple results. Clausen's product can be used to obtain the following relationships \cite{bbg}: \begin{equation} \label{eq:cos2t} \sum_{n=0}^{\infty} \frac{(t)_{n}(-t)_{n}} {(2n)!} (2x)^{2n} = \cos(2t \sin^{-1}x) \end{equation} \begin{equation*} -\frac{1}{2} \sum_{n=1}^{\infty} \frac{(t)_{n}(-t)_{n}} {(2n)!} ( 4\sin^{2} x)^{n} = \sin^{2} (tx) \end{equation*} A generalization of (\ref{eq:cos2t}) which does not involve the Pochhammer symbol is given in \cite{jolley}. Consider integrating (\ref{eq:cos2t}) for $\cos(\sin^{-1}(\sqrt{x}))$: \begin{equation} \label{eq:2/3} \sum_{n=0}^{\infty} \frac{(\frac{1}{2})_{n}(-\frac{1}{2})_{n}} {(2n)!(n+1)} 4^{n} x^{n+1} = -\frac{2}{3} (1-x)^{\frac{3}{2}}+\frac{2}{3} \end{equation} Consider integrating (\ref{eq:2/3}): \begin{equation} \label{eq:2/5} \sum_{n=0}^{\infty} \frac{(\frac{1}{2})_{n}(-\frac{1}{2})_{n}} {(2n)!(n+1)(n+2)} 4^{n} x^{n+2} =-\frac{2}{3} ( - \frac{2}{5} (1-x)^{\frac{5}{2}} -x ) - \frac{4}{15} \end{equation} (\ref{eq:2/3}) can be transformed as follows: \begin{equation} \label{eq:5/8} \sum_{n=0}^{\infty} \frac{(\frac{1}{2})_{n}(-\frac{1}{2})_{n}} {(2n)!(n+1)(2n+3)} 4^{n} x^{2n+3} =-\frac{2}{3} ( \sqrt{1-x^2} (\frac{5x}{8}-\frac{x^3}{4}) -x+\frac{3}{8} \sin^{-1} x ) \end{equation} Using such techniques, it is not difficult to evaluate series such as $\sum_{n=0}^{\infty} \frac{(\frac{1}{6})_{n}(-\frac{1}{6})_{n}} {(2n)!(2n+3)}$ or $\sum_{n=0}^{\infty} \frac{(\frac{1}{8})_{n}(-\frac{1}{8})_{n}} {(2n)!(n+2)} $ . Simple variations of series such as (\ref{eq:2/3}) often lead to outlandish results. Consider multiplying both sides of (\ref{eq:5/8}) by $x$ and integrating: \begin{equation} \label{eq:144} \sum_{n=0}^{\infty} \frac{(\frac{1}{2})_{n}(-\frac{1}{2})_{n}} {(2n)!(n+1)(2n+3)(2n+5)} 4^{n} x^{2n+5} =\frac{1}{144} ( 32 x^{3} -18 x^{2} \sin^{-1} (x) +\sqrt{1-x^2} (4x^{2} -16x^{2} -3)x + 3 \sin^{-1} x ) \end{equation} The Ramanujan-type series which follow are proven using the following very important definite integral: \begin{equation} \label{eq:extremelybeautiful} \int_{0}^{\frac{\pi}{2}} \sin^{2n} t dt = \frac{\pi}{2} \frac{(\frac{1}{2})_n}{n!} \end{equation} Consider substituting (\ref{eq:extremelybeautiful}) into (\ref{eq:cos2t}) and variations of (\ref{eq:cos2t}), yielding fairly simple series, the partial sums of many of which are known. For example consider the partial sums of $\sum_{n=0}^{\infty} 4^{n} \frac{(\frac{1}{32})_{n}(-\frac{1}{32})_{n}(\frac{1}{2})_{n}} {n!(2n)!}$ or $ \sum_{n=2}^{\infty} ( \frac{1}{4})^{2n-1} \frac{\binom{2n-3}{n-1}\binom{2n+1}{n}} {n(n+1)(n+2)}$. Relationships such as (\ref{eq:144}) can be manipulated using (\ref{eq:extremelybeautiful}), resulting in series which are perhaps new; even if such series are not \emph{outright} new, some computational software is unable to evaluate such series, and they do not appear in authoritative mathematical literature on the topic. The following integral is worthy of mention: \begin{equation} \label{eq:alsobeautiful} \int_{0}^{\pi} x \sin^{n} x dx = \frac{\pi^{\frac{3}{2}} \Gamma (\frac{n+1}{2}) } {2 \Gamma (1+ \frac{n}{2})} \end{equation} Now, consider using relationships such as the following to manipulate series such as (\ref{eq:2/3}): \begin{equation} \label{eq:23vari} \int_{0}^{\pi} x ( \frac{2}{3} - \frac{2}{3} (1 - \sin x)^{\frac{3}{2}}) dx = \frac{\pi}{9} (20-16 \sqrt{2} + 3 \pi) \end{equation} Let us begin by acknowledging that it seems to us that the following relationship is highly important, at least in terms of the evaluation of hypergeommetric series: \begin{equation*} \sum_{n=1}^{\infty} (((1 - x^a)^{(b n) + c}) (n + g)^{d} f^n = f (1 - x^a)^{b + c} \Phi(f (1 - x^a)^{b}, -d, 1 + g) \end{equation*} So, the above series can be transformed into a hypergeometric series. If $\operatorname{Re} (a) > 0 $ and $\operatorname{Re}(c + b n) > -1$, then: \begin{equation} \label{eq:notgeneral} \int_{0}^{1} (((1 - x^a)^{(b n) + c}) (n + g)^{d} f^n dx = f^{n} (g + n)^{d} \frac{\Gamma(1 + \frac{1}{a}) \Gamma(1 + c + b n)}{\Gamma(1 + \frac{1}{a} + c + b n)} \end{equation The above relationships lead to an extremely important question, which is author is presently unable to answer: {\bf Is it possible to determine a meaningful, general evaluation of the following integral?} \begin{equation*} \int_{0}^{1} f (1 - x^a)^{b + c} \Phi(f (1 - x^a)^{b}, -d, 1 + g) dx \end{equation* The above integral leads to very beautiful evaluations of hypergeometric series. Questions related to the above question such as the following arise: ``Is it possible to meaningfully evaluate $\int_{0}^{\frac{\pi}{2}} \sin^{-1} \sin^{2} x dx$?" Given that $\sum_{n=2}^{\infty} \frac{(1-x^a)^{bn}}{c^n} = - \frac{(1 - x^a)^{2 b} }{c (-c + (1 - x^a)^b)}$ , is it possible to determine a meaningful, general evaluation of $\int_{0}^{1} - \frac{(1 - x^a)^{2 b} }{c (-c + (1 - x^a)^b)} dx $? ``General" evaluations of this integral for only a \emph{single} variable are very complicated: consider {\bf Theorem 17}. Given that $\sum_{n=2}^{\infty} \frac{(1-x^a)^{bn}}{n^c} = -(1 - x^a)^{b} + \operatorname{Li}_{c}((1 - x^a)^b)$, is it possible to determine a meaningful, general evaluation of $\int_{0}^{1} -(1 - x^a)^{b} + \operatorname{Li}_{c}((1 - x^a)^b) dx $? These questions seem to us to be highly important. In \cite{weiss} it is stated that ``No general algorithm is known for integration of polylogarithms of functions." Such general series can be used to evaluate generalized hypergeometric series $_{p} F _{q}$ for relatively arbitrary $p$ and $q$. More generally, compared to (\ref{eq:notgeneral}), we are interested in integrals of the form $\int^p_q (f(x))^{n} g(x) dx$ which are equal to a finite number of combinations of elementary functions, where $f(x)$ and $g(x)$ represent a finite number of combinations of elementary functions such that ${n} \in \mathbb{Z}^+$; such integrals can be used to establish new hypergeometric series. Let $h(x) = (f(x))^{n} g(x)$. Consider the following straightforward two steps 1) If possible, evaluate $\sum_{n=1}^{\infty} h(x)j(n)$, where $j(n)$ represents a finite combination of elementary functions (e.g. $j(n)=\frac{n^{45}+1}{n+2}$ 2) Integrate the above sum (i.e. using $\int^p_{q} h(x) dx$) and its equivalent expressio Let the above simple two-step procedure be called the HJ-algorithm. In this paper, we present some general tendencies of the HJ-algorithm, through perspectives of experimental mathematics. As stated in the excellent \emph{Mathematics by Experiment: Plausible Reasoning in the 21st Century} \cite{bo}, \begin{quote} The computer provides the mathematician with a ``laboratory" in which he or she can perform experiments: analyzing examples, testing out new ideas, or searching for patterns.\end{quote} Most of the theorems in this paper are not given rigorous analyses; indeed, most of the series are presented for their own sake. Although the ideas (i.e. theorems) being ``tested" in this paper are new in the sense that they may not have been previously published or widely recognized, the HJ-algorithm is so simple that it can hardly be described as a ``new idea". Finally, although we present some seemingly unpredictable tendencies of the HJ-algorithm, we presently do not claim to have considerable knowledge of general patterns in the hypergeometric series. Indeed, some modern computational software is unable to evaluate extremely important integrals such as $\int_0^{1} \frac{(1 - x^a)^b}{-c + (1 - x^a)^b} dx$. Many of the evaluations of the hypergeometric series of the forms we analyze in this paper involve special functions such as the polylogarithm function; for the time being, we mostly discuss hypergeometric series which have a ``nice" (although often long and intricate) form consisting of a finite number of combinations of elementary functions; evaluating an infinite series (i.e. a hypergeometric series) using another infinite series (i.e. the polylogarithm function) seems somwhat redundant. Having said that, we present a way to incorporate the Riemann-zeta function into certain forms of hypergeometric series. We will now present a variety of infinite series. {\bf Theorem 2} is particularly worthy of note, because it is most likely to be original, and does not seem to be equal to a single hypergeometric function. \section{Ramanujan-Type Series} \begin{theorem \begin{equation} \sum_{n=2}^{\infty} ( \frac{1}{4})^{3n-1} \frac{(2n-3)!(4n+3)!}{((2n)!(n+1)(2n+1))^{2}(n-2)!n!} = -\frac{553}{96} - \frac{42 \ln(\frac{2-\sqrt{2}}{2+\sqrt{2}}) -34\sqrt{8}} {9\pi} \end{equation} \end{theorem \begin{proof} {\bf Theorem 2} spawns from (\ref{eq:2/3}). Determine $\int (1-\sin^{4} x)^{\frac{3}{2}} dx$ and substitute (\ref{eq:extremelybeautiful}) into (\ref{eq:2/3}). \end{proof} \begin{theorem \begin{equation} _{3} F _{2} ( - \frac{5}{2}, \frac{1}{4}, \frac{3}{4} ; \frac{1}{2} , 1 ; 1 ) = \frac{1141}{960 \sqrt{2} \pi} - \frac{103 \ln ( \frac{2 - \sqrt{2}}{2 + \sqrt{2}} )}{ 256 \pi} \end{equation} \end{theorem \begin{proof} Use the same technique as in {\bf Theorem 2}, using (\ref{eq:2/5}). \end{proof} Use similar techniques to prove the following very similar equation: \begin{equation*} _{3} F _{2} (-\frac{5}{2},\frac{1}{4},\frac{3}{4};\frac{3}{2},2;1) = \frac{1067}{6720 \sqrt{2} \pi} + \frac{383 \tanh^{-1}(\frac{1}{\sqrt{2}}) }{128 \pi} \end{equation*} Consider how zeta-type sums can be determined using (\ref{eq:extremelybeautiful}). In the great book \emph{Computation Techniques for the Summation of Series} \cite{sofo}, Anthony Sofo provides the following series, which results from integrating $\frac{1}{2} \sum_{n=1}^{\infty} \frac{(2x)^{2n}}{n^{2} \binom{2n}{n} }$ repeatedly: \begin{equation*} \sum_{n=1}^{\infty} \frac{(2x)^{2n+3}} {n^{2}\binom{2n}{n} 4 (2n+3)(2n+2)(2n+1) } = \frac{x(2x^{2}+3)(\sin^{-1}x)^2}{3} + \frac{2\sqrt{1-x^2}(11x^{2}+4)\sin^{-1}x}{9}-\frac{85x^3}{27}-\frac{8x}{9} \end{equation*} We leave it as an exercise to demonstrate how (\ref{eq:extremelybeautiful}) and the above series can be used to prove the equation $\sum_{n=1}^{\infty} \frac{1}{n^{2}(n+1)^{2}(n+2)}=\frac{2\pi^{2}-19}{8}$, the partial sums of the series of which are known. The above series may be simplified as follows: \begin{equation*} _{3} F _ {2} (1,1,1;2,\frac{7}{2};x^2) \frac{x^5}{15} = \frac{x(2x^{2}+3)(\sin^{-1}x)^2}{3} + \frac{2\sqrt{1-x^2}(11x^{2}+4)\sin^{-1}x}{9}-\frac{85x^3}{27}-\frac{8x}{9} \end{equation*} It is worthy of note that the following infinite series, which is reminiscent of the series under discussion, is indicated in \cite{jolley}: \begin{equation*} \sum_{n=1}^{\infty} \frac{(-1)^{n} \theta^{2n-1} (-m)_{2n-1} }{(2n-1)!} = \sin (m \tan^{-1} \theta) (1+\theta^2)^{\frac{m}{2}} \end{equation*} Consider incorporating alternating harmonic numbers into similar such series: $\sum_{j=0}^{\infty} \frac{\binom{2j}{j}}{4^{j} (2j+1)}H^{'}_{2n+1} = -2C + \frac{3\pi \ln 2}{2} $. Consider using the integrals $\int_{0}^{1} \frac{x^{2n+1}}{x+1} dx = H^{'}_{2n+1} - \ln 2$ and $\int_{0}^{1} \frac {\sin^{-1}x}{x+1} dx = -2C+\pi \ln 2$ to establish the above equation. It can be established that: \begin{equation*} \int_{0}^{1} x^{2n} (1-x^2)^{n} dx = \frac{4^n}{(2n+1)\binom{4n+1}{2n}} \end{equation*} Integrals of such forms are very important identities. We are presently interested in series involving the expression $x^{2n} (1-x^2)^{n}$. In the excellent \emph{ Some New Formulas for} $\pi$ \cite{almkv}, integrals of the form $\int x^{pn} (1-x)^{(m-p)n} dx$ are analyzed. In \cite{ramanujan1}, the equation $ \int_{0}^{1} x^{m} (1-x)^{n} dx = \frac{\Gamma (m+1) \Gamma (n+1)} {\Gamma (m+n+2 } $ is given, and is described as ``...an extremely well-known formula for the beta-function... ." Let us consider the integral $\int_{0}^{1} x^{2n} (1-x^2)^{n} dx$ and similar integrals. We intend to establish new infinite series, and classificationss of integrals (which are similar to the beta function). \begin{theorem \begin{equation*} _{4} F _ {3} (1,1,1,\frac{3}{2};\frac{7}{4},2,\frac{9}{4};\frac{1}{4}) = \frac{15} {8} (-64+\pi(8+\pi)+(\cos^{-1}2)^{2}+8\sqrt{3}\cosh^{-1}(7)+\cosh^{-1}(26)\ln(2-\sqrt{3})) \end{equation*} \vspace{.11in} \end{theorem \begin{proof} \begin{equation*} \sum_{n=1}^{\infty} \frac{x^{2n} (1-x^2)^n}{n^2} = \operatorname{Li}_{2} (x^{2}-x^{4}) \end{equation*} Integrate both sides of the above equation. \end{proof} \begin{example \begin{equation*} _{3} F _{2} (\frac{3}{2},2,2;\frac{7}{4},\frac{9}{4};\frac{1}{4}) = \frac{15}{36} (6- \sqrt{3} \ln(2+\sqrt{3}) ) \end{equation*} \end{example \begin{example \begin{equation*} \sum_{n=1}^{\infty} \frac{4^{n}(n+1)^2} {(2n+1)\binom{4n+1}{2n}} = \frac{-36+45\pi+10\sqrt{3}\cosh^{-1}2} {144} \end{equation*} \end{example It can be established that: \begin{equation*} \int_{0}^{1}x^{4n+1} (1-x^4)^{2n+\frac{1}{2}} dx = \pi \frac{(2n-1)!!(4n+1)!!} {2^{3n+3}(3n+1)!} \end{equation*} The above equation can be proven using the gamma function of half-integer expressions. It is not difficult to consider an infinitude of integrals of the form $\int_{0}^{1} x^{a} (1-x^b)^{c}$ equal to an expression involving the gamma function of half-integer expressions. Consider: \begin{equation*} \sum_{n=2}^{\infty} (\frac{1}{32})^{n} \frac{\binom{2n}{n} \binom{2n-3}{n-2} (2n-1)}{n} = \frac{(\Gamma (\frac{1}{4})) ^2}{8\pi^{\frac{3}{2}}} - \frac{9}{32} \end{equation*} This series is determined through the complete elliptic integral of the first kind, being based upon $_{2} F _{1} (\frac{1}{2},\frac{1}{2};2;-1)$. In \emph{Theory and Problems of Complex Variables with an Introduction to Conformal Mapping and its Applications} \cite{spiegel} it is indicated that: \begin{equation*} \int_{0}^{\frac{\pi}{2}} \frac{1}{ \sqrt{1-\frac{\sin^{2}x}{2}}} dx = \frac{(\Gamma(\frac{1}{4}))^2}{4 \sqrt{\pi}} \end{equation*} Use the binomial theorem as follows: \begin{equation*} \frac{1}{\sqrt{1-x}} = -\sum_{n=0}^{\infty} x^{n} \frac{(-\frac{1}{2})_{n} (2n-1)}{n!} \end{equation*} Substitute (\ref{eq:extremelybeautiful}): \begin{equation*} \frac{1}{\sqrt{1-\frac{\sin^{2}x}{2}}} = \sum_{n=0}^{\infty} \sin^{2n} x \frac{(-\frac{1}{2})_{n}(2n-1)}{2^{n}n!} \end{equation*} \section{Computational Perspectives} Let us discuss some results, some of which are possibly original (given that computational software is unable to evaluate them), from perspectives of experimental mathematics. \begin{example \begin{equation*} _{3} F_ {2} (\frac{1}{2},\frac{2}{3},\frac{4}{3};\frac{3}{2},\frac{3}{2};1) = \frac{3\sqrt{3} \coth^{-1} \sqrt{3} }{2} \end{equation*} \end{example \begin{proof} Consider: \begin{equation*} _{2} F_{1} (\frac{1}{3},\frac{2}{3};\frac{3}{2};z^2) ^{2} = (\frac{3\sin(\frac{\sin^{-1} z}{3} )}{z} )^2 \end{equation*} Integrate both sides of the above equation \end{proof} \begin{example \begin{equation*} _{3} F _{2} (\frac{1}{3},\frac{1}{2},\frac{5}{3};\frac{3}{2},\frac{3}{2};1) = \frac{9}{4} - \frac{3}{4} \sqrt{3} \coth^{-1}\sqrt{3} \end{equation*} \end{example \begin{example \begin{equation*} _{3} F _{2} (\frac{1}{4},\frac{1}{2},\frac{7}{4};\frac{3}{2},\frac{3}{2};1) = \frac{1}{9} (12 \sqrt{2} - 3 \pi i + 12 \tanh^{-1}(-1)^{\frac{3}{4}}) \end{equation*} \end{example \begin{example \begin{eqnarray*} \lefteqn{ _{3} F _{2} (\frac{1}{8},\frac{1}{2},\frac{15}{8};\frac{3}{2},\frac{3}{2};1) = } \\ & & \frac{ 16 \sqrt{2 + \sqrt{2}}}{21} + \frac{2 i \pi}{7} - \frac{8}{7} i \tan^{-1}((-1)^{\frac{1}{8}}) + \frac{4}{7} i \sqrt{2} \tan^{-1}(1 - \sqrt{2}) \\ & & - \frac{4}{7} i \sqrt{2} \tan^{-1}(1 + \sqrt{2}) - \frac{4}{7} i \sqrt{2} \tan^{-1}(1 - (-1)^{\frac{1}{8}} \sqrt{2}) \\ & & + \frac{4}{7} i \sqrt{2} \tan^{-1}(1 + (-1)^{\frac{1}{8}} \sqrt{2}) \end{eqnarray*} \end{example \begin{example \begin{equation*} _{3} F _{2} (\frac{1}{6},\frac{1}{2},\frac{11}{6};\frac{3}{2},\frac{3}{2};1) = \frac{9 \sqrt{3}}{10} + \frac{\sqrt{3} \pi i}{10} - \frac{3 \sqrt{3} \tanh^{-1}( \frac{3 + 2 i \sqrt{3}}{7} )}{5} \end{equation*} \end{example We consider integrals of the form $\int_{0}^{1} (\frac{\sin(a \sin^{-1}x)}{x} )^{2} dx$ to be important. Consider series for $\frac{1}{\pi}$ using the above latter technique presented in hypergeometric form: \begin{theorem \begin{equation*} _{3} F_{2} (\frac{1}{2},\frac{2}{3},\frac{4}{3}; \frac{3}{2}, 2;1) = \frac{3\sqrt{3}(3-\ln 4)}{2 \pi} \end{equation*} \end{theorem \begin{proof} Consider: \begin{equation*} (_{2} F_{1} ( \frac{1}{3} , \frac{2}{3} ; \frac{3}{2} ; \sin^{2} z ))^2 = (3 \csc (z) \sin (\frac{1}{3} \sin^{-1} (\sin z)))^2 \end{equation*} Integrate both sides of the above equation. Recall (\ref{eq:extremelybeautiful}). \end{proof} \begin{theorem \begin{equation*} _{3} F_{2} (\frac{1}{2}, \frac{1}{3}, \frac{5}{3} ; \frac{3}{2}, 2 ; 1 ) = \frac{3 \sqrt{3} (3+ \ln 16 )} {8 \pi} \end{equation*} \end{theorem \begin{proof} Consider: \begin{equation*} (_{2} F_{1} ( \frac{1}{6} , \frac{5}{6} ; \frac{3}{2} ; \sin^{2} z ))^2 = (\frac{3}{2} \csc (z) \sin (\frac{2}{3} \sin^{-1} (\sin z)))^2 \end{equation*} Given the HJ algorithm, {\bf Theorem 6} holds. \end{proof} \begin{theorem \begin{equation*} _{3} F_{2} (\frac{1}{2}, \frac{1}{4}, \frac{7}{4}; \frac{3}{2}, 2 ; 1) = \frac{8 (\sqrt{2}+6\tanh^{-1} \sqrt{\frac{2 - \sqrt{2}}{2 + \sqrt{2}} } )}{9 \pi} \end{equation*} \end{theorem \begin{proof} \begin{equation*} (_{2} F_{1} ( \frac{1}{8} , \frac{7}{8} ; \frac{3}{2} ; \sin^{2} z ))^2 = (\frac{4}{3} \csc (z) \sin (\frac{3}{4} \sin^{-1} (\sin z)))^2 \end{equation*} Given the HJ algorithm, {\bf Theorem 7} holds. \end{proof} \begin{theorem \begin{equation*} _{3} F_{2} (\frac{1}{2}, \frac{1}{6}, \frac{11}{6}; \frac{3}{2}, 2 ; 1) = \frac{\frac{18}{25}+\frac{12}{5}\sqrt{3}\coth^{-1}\sqrt{3}} {\pi} \end{equation*} \end{theorem \begin{proof} \begin{equation*} (_{2} F_{1} ( \frac{1}{12} , \frac{11}{12} ; \frac{3}{2} ; \sin^{2} z ))^2 = (\frac{6}{5} \csc (z) \sin (\frac{5}{6} \sin^{-1} (\sin z)))^2 \end{equation*} Given the HJ algorithm, {\bf Theorem 8} holds. \end{proof} The following very useful relationship is given in ``Some New Formulas for $\pi$" \cite{almkv}: \begin{equation*} \frac{1}{\binom{3n}{n}(3n+1)} = \int_{0}^{1} x^{2n} (1-x)^{2n} dx \end{equation*} Series of the form $\pi = \sum_{n=0}^{\infty} \frac{S(n)}{\binom{mn}{pn}a^n}$ are discussed in \cite{almkv}. Consider another example of the above generalization: \begin{theorem \begin{equation*} _{3} F_{2} ( \frac{1}{2}, \frac{8}{5}, \frac{2}{5}; \frac{3}{2}, 2; 1 ) = \frac{5 (5 (5 + 3 \sqrt{5}) + 3 (5 + 3 \sqrt{5}) \ln 5 - 12 (5 + 2 \sqrt{5}) \ln(5 - \sqrt{5}) + 6 (5 + \sqrt{5}) \ln (5 + \sqrt{5}))}{9 \pi (1 + \sqrt{5}) \sqrt{2 (5 + \sqrt{5})}} \end{equation*} \end{theorem \begin{proof} Consider: $(_{2} F_{1} ( \frac{1}{5} , \frac{4}{5} ; \frac{3}{2} ; \sin^{2} z ))^2 = (\frac{5}{3} \csc (z) \sin (\frac{3}{5} \sin^{-1} (\sin z)))^2$. \end{proof} The following integral is given in \cite{gradsh}: $\int_0^{\pi} x \sin^{p} x dx = \frac{\pi^2}{2^{p+1}} \frac{\Gamma(p+1)}{(\Gamma(\frac{p}{2}+1))^2}$. We presently evaluate new hypergeometric series using such integrals through manipulations of the left-hand side of the following equation discussed in \cite{bo}: \begin{eqnarray*} \lefteqn{_{3}F _{2} (a,2b-a-1,a-2b+2;b,a-b+\frac{3}{2};\frac{x}{4}) = } \\ & & _{3}F _{2} (\frac{a}{3}, \frac{a+1}{3}, \frac{a+2}{3};b,a-b+\frac{3}{2}; \frac{-27x}{4(1-x)^3}) (1-x)^{-a} \end{eqnarray*} Replace $x$ with $4\sin^{2}(x)$ and integrate. Let $\heartsuit (n,x)$ = $_{2}F_{1} (-n-1,n+2; \frac{3}{2}; \sin^{2}x)$. \section{Further Results} \begin{theorem \begin{eqnarray*} \lefteqn{_{3} F_{2} (-\frac{6}{5},\frac{1}{2},\frac{11}{5};1,\frac{3}{2};1) = } \\ & & ( 5 (\sqrt{5 (5 + \sqrt{5})} (25 - 6 \ln(-1 + \sqrt{5})) + \sqrt{5 + \sqrt{5}} (-25 + 6 \ln(-1 + \sqrt{5})) \\ & & + \sqrt{5 - \sqrt{ 5}} (1 + \sqrt{5}) (\ln(64) - 6 \ln(3 + \sqrt{5}))) )/ ( 204 \pi \sqrt{2} ) \end{eqnarray*} \end{theorem \begin{proof} Consider: $\heartsuit (\frac{1}{5},x) = \frac{5}{17} \csc(x) \sin(\frac{17}{5} \sin^{-1}(\sin x))$. Use (\ref{eq:extremelybeautiful}). \end{proof} \begin{theorem \begin{eqnarray*} \lefteqn{_{3} F_{2} (-\frac{9}{8}, \frac{1}{2}, \frac{17}{8}; 1, \frac{3}{2}; 1) = } \\ & & (( \frac{2}{ 117 \pi} (128 \sqrt{2 - \sqrt{2}} + 9 (\ln(161 - 112 \sqrt{2} - 4 \sqrt{2 (1594 - 1127 \sqrt{2})}) \\ & & - 2 \sqrt{2} \ln(\sqrt{2} + \sqrt{2 - \sqrt{2}}) + \sqrt{2} \ln(4 - \sqrt{2} - 2 \sqrt{2 (2 - \sqrt{2})}))))) \end{eqnarray*} \end{theorem \begin{proof} Integrate $\heartsuit (\frac{1}{8},x)$. \end{proof} \begin{theorem \begin{equation*} _{3} F _{2} (1,1,\frac{3}{2};\frac{5}{3},\frac{13}{6};1) = \frac{28}{3} + \frac{ 2^{\frac{1}{3}} 7 \pi}{9 \sqrt{3}} - \frac{ 2^{\frac{1}{3}} 14 \tan^{-1}( \frac{1 + 2^{\frac{2}{3}}}{\sqrt{3}} )} {3 \sqrt{3}} + \frac{14}{9} 2^{\frac{1}{3}} \ln(2 - 2^{\frac{2}{3}}) - \frac{7}{9} 2^{\frac{1}{3}} \ln(2 (2 + 2^{\frac{1}{3}} + 2^{\frac{2}{3}})) \end{equation*} \end{theorem \begin{proof} Use the HJ-algorithm. \begin{equation*} \sum_{n=1}^{\infty} \frac{(1 - x^3)^{2n}}{n} = -\ln(-x^{3} (-2 + x^{3})) \end{equation* Integrate both sides of the above equation. \end{proof} \begin{example \begin{eqnarray*} _{3} F _{2} (1,1,\frac{3}{2};\frac{19}{12}, \frac{25}{12};1) & = & \frac{91}{72} ( 12 - 2^{\frac{1}{6}} \sqrt{3} \tan^{-1} (\frac{-1 + 2^{\frac{5}{6}}}{\sqrt{3}}) - 2^{\frac{1}{6}} \sqrt{3} \tan^{-1}(\frac{1 + 2^{\frac{5}{6}}} {\sqrt{3}}) + 2^{\frac{1}{6}} \ln(2 - 2^{\frac{5}{6}}) \\ & & \mbox{} + \frac{\ln(2 + 2^{\frac{2}{3}} - 2^{\frac{5}{6}})} {2^{\frac{5}{6}}} - 2^{\frac{1}{6}} \ln(2 + 2^{\frac{5}{6}}) - \frac{\ln(2 + 2^{\frac{2}{3}} + 2^{\frac{5}{6}})}{2^{\frac{5}{6}}} ) \end{eqnarray*} \end{example \begin{proof} Use the HJ-algorithm. \begin{equation*} \sum _{n=1}^{\infty} \frac{(1-x^6)^{2n}}{n} = -\ln(-x^{6}(-2+x^6)) \end{equation* Integrate both sides of the above equation. \end{proof} \begin{example \begin{eqnarray*} _{4} F _{3} (1,\frac{5}{4},\frac{7}{4},2;\frac{3}{2},\frac{11}{6},\frac{13}{6};\frac{2}{27}) = \frac{7}{2}(-10+\pi+(-2\sqrt{-7 - i}+\sqrt{2-14i})\tan^{-1}(\sqrt{\frac{-1-i}{2}}) \\ -2\sqrt{-7+i} \tan^{-1}(\sqrt{\frac{-1+i}{2}})+\sqrt{2+14i}\tan^{-1}(\sqrt{\frac{-1+i}{2}})) \end{eqnarray*} \end{example \begin{proof} Use the HJ-algorithm. \begin{equation*} \sum_{n=1}^{\infty} \frac{(1-x^2)^{n} x^{4n}}{2^n} = - \frac{x^{4} (-1 + x^2)}{2 - x^{4} + x^6} \end{equation* Consider: \begin{equation*} \int_0 ^1 (1-x^2)^{n} x^{4n} dx = \frac{\Gamma(1 + n) \Gamma(\frac{1}{2} + 2 n)}{2 \Gamma(\frac{3}{2} + 3 n)} = \frac {2^{n} n! (4n-1)!!} { (6n+1)!!} \end{equation* \end{proof} \begin{example \begin{eqnarray*} \lefteqn{_{6} F _{5} (1,1,1,\frac{5}{4},\frac{3}{2},\frac{7}{4}; \frac{11}{8},\frac{13}{8},\frac{15}{8},2,\frac{17}{8};1) = } \\ & & \frac{315}{64} (-32 + 3 \pi^{2} - 4 (-1)^{\frac{1}{8}} 2^{\frac{1}{4}} \ln(-1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) + 2 i \cot^{-1} ( \frac{2^{\frac{3}{4}} - \sqrt{2 + \sqrt{2}}}{\sqrt{ 2 - \sqrt{2}} } ) \\ & & \ln(-1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) + 4 (-1)^{\frac{1}{8}} 2^{\frac{1}{4}} \ln(1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) - 2 i \cot^{-1} ( \frac{2^{\frac{3}{4}} - \sqrt{2 + \sqrt{2}}}{\sqrt{2 - \sqrt{2}}} ) \\ & & \ln( i + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) + \ln ^{2} ( \frac{1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}}{1 - (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}} ) - 4 (-1)^{\frac{7}{8}} 2^{\frac{1}{4}} \ln(-1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) \\ & & + \ln ^{2} ( \frac{1 - (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}}{1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}} ) + 2 i \cot^{-1} ( \frac{2^{\frac{3}{4}} + \sqrt{2 + \sqrt{2}}}{\sqrt{ 2 - \sqrt{2}}} ) (\ln(-1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) - \ln(1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) \\ & & + \ln(-1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) - \ln(1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}})) + 4 (-1)^{\frac{7}{8}} 2^{\frac{1}{4}} \ln(1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) \\ & & - 2 i \cot^{-1} ( \frac{2^{\frac{3}{4}} - \sqrt{2 + \sqrt{2}}}{\sqrt{2 - \sqrt{2}}} ) \ln(1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) + 2 \ln(-1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) \ln(1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) \\ & & - 2 i \pi ( \ln ( \frac{(-1)^{\frac{1}{8}} }{(-1)^{\frac{1}{8}} - 2^{\frac{1}{4}}} ) - \ln(-1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) + \ln(1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) + 2 \ln(1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}})) \\ & & - 4 \sqrt{2} \ln(-1 + \sqrt{2}) - 2 \ln ^{2}(-1 + \sqrt{2}) + 4 \sqrt{2} \ln(1 + \sqrt{2}) + 4 \ln(-1 + \sqrt{2}) \ln(1 + \sqrt{2}) \\ & & - 2 \ln(1 + \sqrt{2})^2 + \ln(3 + 2 \sqrt{2})^2 - \ln(1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) \ln( \operatorname{Im}(\sqrt{1 - i})^2 \\ & & + (-1 + \operatorname{Re}(\sqrt{1 - i}))^2) - \ln(-1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) \ln( \operatorname{Im}(\sqrt{1 - i})^2 \\ & & + (1 + \operatorname{Re}(\sqrt{1 - i}))^2) + \ln(1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) \ln( \operatorname{Im}(\sqrt{1 - i})^2 \\ & & + (1 + \operatorname{Re}(\sqrt{1 - i}))^2) - \ln(-1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) \ln( \operatorname{Im}(\sqrt{1 + i})^2 \\ & & + (-1 + \operatorname{Re}(\sqrt{1 + i}))^2) + \ln(1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) \ln( \operatorname{Im}(\sqrt{1 + i})^2 \\ & & + (-1 + \operatorname{Re}(\sqrt{1 + i}))^2) + \ln(-1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) \ln( \operatorname{Im}(\sqrt{1 + i})^2 \\ & & + (1 + \operatorname{Re}(\sqrt{1 + i}))^2) - \ln(1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) \ln( \operatorname{Im}(\sqrt{1 + i})^2 + (1 + \operatorname{Re}(\sqrt{1 + i}))^2)) \end{eqnarray*} \end{example \begin{proof} Use the HJ-algorithm. \begin{equation*} \sum_{n=1}^{\infty} \frac{(1 - x^2)^{4 n}}{n^2} = \operatorname{Li}_{2}((-1+x^2)^4) \end{equation*} \end{proof} \begin{example \begin{eqnarray*} \lefteqn{_{5} F _{4} (\frac{1}{4},\frac{3}{4},1,1,2;\frac{1}{2},\frac{5}{6},\frac{7}{6},3;\frac{2}{27}) = } \\ & & \frac{18}{5} - \frac{491 \pi}{35} + \frac{81 \pi^2}{4} + 9 \ln ^{2} ( \frac{ (-1)^{\frac{1}{8}} + 2^{\frac{1}{4}}}{(-1)^{\frac{1}{8}} - 2^{\frac{1}{4}}} ) - (\frac{522}{35} - \frac{124 i}{5}) (-1)^{\frac{1}{8}} 2^{\frac{1}{4}} \ln(-1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) \\ & & + 18 \pi i \ln(-1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) - 18 i \cot^{-1} ( \frac{-2 + 2^{\frac{1}{4}} \sqrt{2 + \sqrt{2}}}{2^{\frac{1}{4}} \sqrt{2 - \sqrt{2}}} ) \ln(-1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) \\ & & + 18 i \cot^{-1}( \frac{2 + 2^{\frac{1}{4}} \sqrt{2 + \sqrt{2}}}{2^{\frac{1}{4}} \sqrt{2 - \sqrt{2}}} ) \ln(-1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) + (\frac{522}{35} - \frac{124 i}{5}) (-1)^{\frac{1}{8}} 2^{\frac{1}{4}} \ln(1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) \\ & & - 18 i \pi \ln(1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) + 18 i \cot^{-1}( \frac{-2 + 2^{\frac{1}{4}} \sqrt{2 + \sqrt{2}}}{2^{\frac{1}{4}} \sqrt{2 - \sqrt{2}}} ) \ln(1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) \\ & & - 18 i \cot^{-1} ( \frac{2 + 2^{\frac{1}{4}} \sqrt{2 + \sqrt{2}}}{ 2^{\frac{1}{4}} \sqrt{2 - \sqrt{2}}} ) \ln(1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) + 9 \ln ^{2} ( \frac{1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}}{ 1 - (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}} ) \\ & & - (\frac{522}{35} + \frac{124 i}{5}) (-1)^{\frac{7}{8}} 2^{\frac{1}{4}} \ln(-1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) + 18 i \cot^{-1} ( \frac{2 + 2^{\frac{1}{4}} \sqrt{2 + \sqrt{2}}}{ 2^{\frac{1}{4}} \sqrt{2 - \sqrt{2}}} ) \ln(-1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) \\ & & + (\frac{522}{35} + \frac{124i}{5} ) (-1)^{\frac{7}{8}} 2^{\frac{1}{4}} \ln(1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) - 18 \pi i \ln(1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) + 18 i \cot^{-1} ( \frac{-2 + 2^{\frac{1}{4}} \sqrt{2 + \sqrt{2}}}{2^{\frac{1}{4}} \sqrt{2 - \sqrt{2}}} ) \\ & & \ln(1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) - 18 i \cot^{-1} ( \frac{2 + 2^{\frac{1}{4}} \sqrt{2 + \sqrt{2}}}{ 2^{\frac{1}{4}} \sqrt{2 - \sqrt{2}}} ) \ln[1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}] \\ & & + 18 \ln(-1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) \ln(1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) \\ & & - 9 \ln(1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) \ln^{2 }( \operatorname{Im}(\sqrt{1 - i})+ (-1 + \operatorname{Re}(\sqrt{1 - i}))^2) \\ & & - 9 \ln(-1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) \ln^{2}( \operatorname{Im}(\sqrt{1 - i}) + (1 + \operatorname{Re}(\sqrt{1 - i}))^2) \\ & & + 9 \ln(1 + (-1)^{\frac{7}{8}} 2^{\frac{1}{4}}) \ln( \operatorname{Im}(\sqrt{1 - i})^2 + (1 + \operatorname{Re}(\sqrt{1 - i}))^{2}) \\ & & - 9 \ln(-1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) \ln( \operatorname{Im}(\sqrt{1 + i})^2 + (-1 + \operatorname{Re}(\sqrt{1 + i}))^2) \\ & & + 9 \ln(1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) \ln( \operatorname{Im}(\sqrt{1 + i})^2 + (-1 + \operatorname{Re}(\sqrt{1 + i}))^2) \\ & & + 9 \ln(-1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) \ln( \operatorname{Im}(\sqrt{1 + i})^2 + (1 + \operatorname{Re}(\sqrt{1 + i}))^2) \\ & & - 9 \ln(1 + (-1)^{\frac{1}{8}} 2^{\frac{1}{4}}) \ln( \operatorname{Im}(\sqrt{1 + i})^2 + (1 + \operatorname{Re}(\sqrt{1 + i}))^2) \end{eqnarray*} \end{example \begin{proof} Use the HJ-algorithm. \begin{equation*} \sum_{n=1}^{\infty} \frac{(1 - x^2)^{n} x^{4 n}}{2^{n} (n + 2) } = \frac{-4 x^{4} + 4 x^{6} - x^{8} + 2 x^{10} - x^{12} - 8 \ln(\frac{1}{2} (2 - x^{4} + x^{6}))}{2 x^{8} (-1 + x^2)^2} \end{equation* \end{proof} \begin{theorem \begin{equation*} _{3} F _{2} (\frac{3}{2},2,2;\frac{7}{4},\frac{9}{4};\frac{1}{3}) = \frac{45}{128} (6 - \sqrt{3 (1 + \sqrt{3})} \coth^{-1}(\sqrt{1 + \sqrt{3}} ) + \sqrt{3 (-1 + \sqrt{3})} \csc^{-1}(3^{\frac{1}{4}})) \end{equation*} \end{theorem \begin{proof} Use the HJ-algorithm. \begin{equation*} \sum_{n=1}^{\infty} \frac{(1 - x^2)^{2n} n} {3^n} = \frac{3 (-1 + x^2)^2}{(-2 - 2 x^{2} + x^4)^2} \end{equation*} \end{proof} \begin{theorem \begin{equation*} _{5} F _{4} (\frac{3}{2},2,2,2,2;1,\frac{7}{4},\frac{9}{4},3;\frac{1}{4}) = \frac{5}{24} (9 (-6 + \pi) \pi + 4 (6 + 14 \sqrt{3} \cosh^{-1}(2) + 9 \cosh^{-1}(2)^{2} - 72 \coth^{-1}(\sqrt{3})^2)) \end{equation*} \end{theorem \begin{proof} Use the HJ-algorithm. \begin{eqnarray*} \lefteqn{\sum_{n=1}^\infty \frac{(1 - x^2)^{2 n} n^2}{4^{n} (n + 1)} = } \\ & & -\frac{1}{(3 - x^{2} - 3 x^{4} + x^{6})^2} 4 (2 - 8 x^{4} + 8 x^{6} - 2 x^{8} + 9 \ln(\frac{1}{4} (3 + 2 x^{2} - x^{4})) + 12 x^{2} \ln(\frac{1}{4} (3 + 2 x^{2} - x^{4})) \\ & & - 2 x^{4} \ln(\frac{1}{4} (3 + 2 x^{2} - x^{4})) - 4 x^{6} \ln(\frac{1}{4} (3 + 2 x^{2} - x^{4})) + x^{8} \ln(\frac{1}{4} (3 + 2 x^{2} - x^{4}))) \end{eqnarray*} \end{proof} \begin{theorem \begin{eqnarray*} \lefteqn{_{3} F _{2} (\frac{3}{2},2,2;\frac{5}{3},\frac{13}{6};\frac{1}{4} ) = } \\ & & \frac{14}{2187} ( 216 + \sqrt{3} (18+53^{\frac{1}{3}}) \pi - 3 ^ {\frac{5}{6}} 30 \tan^{-1} (\frac{2+3^{\frac{1}{3}}}{3^\frac{5}{6}}) + 54 \ln (2) \\ & & + 3 ^ \frac{1}{3} 30 \ln(3-3^\frac{2}{3}) - 3 ^ \frac{1}{3} 15 \ln (3(3+3^{\frac{1}{3}}+3^{\frac{2}{3}})) ) \end{eqnarray*} \end{theorem \begin{proof} Use the HJ-algorithm. \begin{equation*} \sum_{n=1}^{\infty} \frac{(1 - x^3)^{2 n} n}{4^n} = \frac{4 (-1 + x^3)^2}{(-3 - 2 x^{3} + x^6)^2} \end{equation* \end{proof} \begin{theorem \begin{equation*} _{4} F _{3} (\frac{3}{2}, 2, 2, 2; 1, \frac{13}{8}, \frac{17}{8}; \frac{1}{4}) = \frac{5}{1024}(88 + 63 \sqrt{2} \pi + 3^{\frac{1}{4}} 4 \cot^{-1}(3^{\frac{1}{4}}) + 126 \sqrt{2} \coth^{-1}(\sqrt{2}) + 3^{\frac{1}{4}}4 \coth^{-1}(3^{\frac{1}{4}})) \end{equation*} \end{theorem \begin{proof} Use the HJ-algorithm. \begin{equation*} \sum_{n=1}^{\infty} \frac{(1 - x^4)^{2 n} n^2}{4^n} = -\frac{4 (-1 + x^4)^2 (5 - 2 x^{4} + x^8)}{(-3 - 2 x^4 + x^8)^3} \end{equation*} \end{proof} \begin{example \begin{equation*} _{3} F _{2} (-\frac{3}{2},\frac{1}{4},\frac{3}{4};\frac{3}{2},2;1) = - \frac{9}{40 \sqrt{2} \pi} + \frac{55 \tanh^{-1}(\frac{1}{\sqrt{2}})}{16 \pi} \end{equation*} \end{example \begin{example \begin{eqnarray*} \lefteqn{ _{4} F _{3} (1,1,1,\frac{3}{2};\frac{7}{4},2,\frac{9}{4};1) = } \\ & & (-8 + \frac{7 \pi^{2}}{6} + \frac{1}{2} \ln^{2}(-i - 2^{\frac{1}{4}}) + \ln^{2}(i - 2^{\frac{1}{4}}) + 2 \sqrt{2} \ln(1 - i 2^{\frac{1}{4}}) - \ln(2) \ln(1 - i 2^{\frac{1}{4}}) \\ & & + 2 \sqrt{2} \ln(1 + i 2^{\frac{1}{4}}) - \ln(2) \ln(1 + i 2^{\frac{1}{4}}) - 2 \sqrt{2} \ln(-1 + 2^{\frac{1}{4}}) + \ln(2) \ln(-1 + 2^{\frac{1}{4}}) \\ & & - \ln(-i - 2^{\frac{1}{4}}) \ln(-1 + 2^{\frac{1}{4}}) - \ln(i - 2^{\frac{1}{4}}) \ln(-1 + 2^{\frac{1}{4}}) + \ln(1 - i 2^{\frac{1}{4}}) \ln(-1 + 2^{\frac{1}{4}}) \\ & & + \ln(1 + i 2^{\frac{1}{4}}) \ln(-1 + 2^{\frac{1}{4}}) - \frac{1}{2} \ln^{2}(-1 + 2^{\frac{1}{4}}) + \ln(1 + i 2^{\frac{1}{4}}) \ln(- \frac{i}{-i + 2^{\frac{1}{4}}} ) \\ & & + \ln(1 - i 2^{\frac{1}{4}}) \ln ( - \frac{2 i}{-i + 2^{\frac{1}{4}}} ) - \ln(-1 + 2^{\frac{1}{4}}) \ln ( \frac{1 - i}{-i + 2^{\frac{1}{4}}} ) + \frac{1}{2} \ln ^{2} ( - \frac{1}{i + 2^{\frac{1}{4}}} ) \\ & & + \ln(1 - i 2^{\frac{1}{4}} ) \ln(\frac{i}{i + 2^{\frac{1}{4}}} ) + \ln(1 + i 2^{\frac{1}{4}}) \ln ( \frac{2 i}{i + 2^{\frac{1}{4}}} ) + i \pi (\ln ( 1 - i 2^{\frac{1}{4}} ) \\ & & - \ln(1 + i 2^{\frac{1}{4}} ) + \ln ( - \frac{2 i}{-i + 2^{\frac{1}{4}}} ) - \ln ( \frac{1 - i}{-i + 2^{\frac{1}{4}}} ) + \ln ( \frac{i}{i + 2^{\frac{1}{4}}} ) - \ln ( \frac{1 + i}{i + 2^{\frac{1}{4}}} ) ) \\ & & - \ln ( -1 + 2^{\frac{1}{4}} ) \ln ( \frac{1 + i}{i + 2^{\frac{1}{4}}} ) - 2 \sqrt{2} \ln ( 1 + 2^{\frac{1}{4}} ) + \ln(2) \ln ( 1 + 2^{\frac{1}{4}}) - \ln ( i - 2^{\frac{1}{4}} ) \ln (1 + 2^{\frac{1}{4}}) \\ & & + \ln (1 - i 2^{\frac{1}{4}}) \ln(1 + 2^{\frac{1}{4}}) + \ln(1 + i 2^{\frac{1}{4}}) \ln(1 + 2^{\frac{1}{4}}) - \ln(-1 + 2^{\frac{1}{4}}) \ln(1 + 2^{\frac{1}{4}}) \\ & & - \ln ( - \frac{1 + i}{-i + 2^{\frac{1}{4}}} ) \ln(1 + 2^{\frac{1}{4}}) + \ln ( - \frac{1}{i + 2^{\frac{1}{4}}} ) \ln(1 + 2^{\frac{1}{4}}) \\ & & - \ln( - \frac{1 - i}{i + 2^{\frac{1}{4}}} ) \ln(1 + 2^{\frac{1}{4}}) - \frac{1}{2} \ln^{2}(1 + 2^{\frac{1}{4}}) + \frac{1}{3} (\pi^{2} - 6 \pi \cot^{-1} (2^{\frac{1}{4}}) + 6 \cot^{-1}(2^{\frac{1}{4}})^2)) \frac{15}{4} \end{eqnarray*} \end{example Given that $\sum_{n=1}^{\infty} \frac{((1 - x^4)^{2 n}) x}{n^2} = x \operatorname{Li}_{2} ((-1 + x^4)^2)$, integrate both sides of this equation. What are alternative series representations of series of the form $\sum_{n=2}^{\infty} \frac{(\zeta (n) - 1) (p)_{an}} {(q)_{bn}} $ ? \begin{theorem \begin{eqnarray*} \lefteqn{\sum_{n=2}^{\infty} \frac{(\zeta (n) - 1)(2 n)!} {(\frac{5}{4})_{2n}} = } \\ & & \sum_{n=2}^{\infty} -(\frac{1}{8} (8 - \frac{ \sqrt{n} \ln(1 - (1 - \sqrt{n})^{\frac{1}{4}}) }{ (1 - \sqrt{n})^{\frac{3}{4}} } - \frac{i \sqrt{n} \ln( 1 - i (1 - \sqrt{n})^{\frac{1}{4}})}{ (1 - \sqrt{n})^{\frac{3}{4}}} + \frac{i \sqrt{n} \ln(1 + i (1 - \sqrt{n})^{\frac{1}{4}})}{(1 - \sqrt{n})^{\frac{3}{4}}} \\ & & + \frac{\sqrt{n} \ln( 1 + (1 - \sqrt{n})^{\frac{1}{4}})}{(1 - \sqrt{n})^{\frac{3}{4}} } + \frac{\sqrt{n} \ln(1 - (1 + \sqrt{n})^{\frac{1}{4}})}{(1 + \sqrt{n})^{\frac{3}{4}}} + \frac{i \sqrt{n} \ln(1 - i (1 + \sqrt{n})^{\frac{1}{4}})}{(1 + \sqrt{n})^{\frac{3}{4}}} \\ & & - \frac{i \sqrt{n} \ln( 1 + i (1 + \sqrt{n})^{\frac{1}{4}})}{(1 + \sqrt{n})^{\frac{3}{4}}} - \frac{\sqrt{ n} \ln(1 + (1 + \sqrt{n})^{\frac{1}{4}})}{(1 + \sqrt{n})^{\frac{3}{4}}} \\ & & - \frac{\sqrt{ n}}{32 (1 - n)^{\frac{3}{4}}} (-4 (1 + \sqrt{n})^{\frac{3}{4}} \ln(-(1 - \sqrt{n})^{\frac{1}{4}}) 4 i (1 + \sqrt{n})^{\frac{3}{4}} \ln(-i (1 - \sqrt{n})^{\frac{1}{4}}) \\ & & + 4 i (1 + \sqrt{n})^{\frac{3}{4}} \ln( i (1 - \sqrt{n})^{\frac{1}{4}}) + (1 + \sqrt{n})^{\frac{3}{4}} \ln( 1 - \sqrt{n}) + 4 (1 - \sqrt{n})^{\frac{3}{4}} \ln(-(1 + \sqrt{n})^{\frac{1}{4}}) \\ & & + 4 i (1 - \sqrt{n})^{\frac{3}{4}} \ln(-i (1 + \sqrt{n})^{\frac{1}{4}}) - 4 i (1 - \sqrt{n})^{\frac{3}{4}} \ln( i (1 + \sqrt{n})^{\frac{1}{4}}) \\ & & - (1 - \sqrt{n})^{\frac{3}{4}} \ln( 1 + \sqrt{n})) - \frac{32}{45 n}) \end{eqnarray*} \end{theorem \begin{proof} Consider: \begin{equation*} \label{eq:expan1} \sum_{n=2}^{\infty} \frac{(1-x^4)^{2n}}{c^n} = \frac{(-1 + x^4)^4}{c (-1 + c + 2 x^{4} - x^8)} \end{equation*} Integrate both sides of the above equation; integrating the right-hand side proves to be rather challenging. \end{proof} A question which immediately comes to mind is: ``Can the right-hand side of {\bf Theorem 17} be simplified at all?" We presently leave open to discussion how such $\zeta (x)$ expansions can be manipulated and simplified. We should note that such $\zeta (x) $ expansions may be expressed in integral form by evaluating series of the form $\sum_{n=2}^{\infty} \zeta (x) f(x)^{n}$. Such definite integrals are often rather striking. The following relationship is related to {\bf Theorem 17}. \begin{theorem \begin{eqnarray*} \lefteqn{ _{3} F _{2} ( 1,3,\frac{5}{2} ; \frac{21}{8} , \frac{25}{8} ; \frac{1}{n}) \frac{8192}{n^{2}13260} = } \\ & & -(\frac{1}{8} (8 - \frac{ \sqrt{n} \ln(1 - (1 - \sqrt{n})^{\frac{1}{4}}) }{ (1 - \sqrt{n})^{\frac{3}{4}} } - \frac{i \sqrt{n} \ln( 1 - i (1 - \sqrt{n})^{\frac{1}{4}})}{ (1 - \sqrt{n})^{\frac{3}{4}}} + \frac{i \sqrt{n} \ln(1 + i (1 - \sqrt{n})^{\frac{1}{4}})}{(1 - \sqrt{n})^{\frac{3}{4}}} \\ & & + \frac{\sqrt{n} \ln( 1 + (1 - \sqrt{n})^{\frac{1}{4}})}{(1 - \sqrt{n})^{\frac{3}{4}} } + \frac{\sqrt{n} \ln(1 - (1 + \sqrt{n})^{\frac{1}{4}})}{(1 + \sqrt{n})^{\frac{3}{4}}} + \frac{i \sqrt{n} \ln(1 - i (1 + \sqrt{n})^{\frac{1}{4}})}{(1 + \sqrt{n})^{\frac{3}{4}}} \\ & & - \frac{i \sqrt{n} \ln( 1 + i (1 + \sqrt{n})^{\frac{1}{4}})}{(1 + \sqrt{n})^{\frac{3}{4}}} - \frac{\sqrt{ n} \ln(1 + (1 + \sqrt{n})^{\frac{1}{4}})}{(1 + \sqrt{n})^{\frac{3}{4}}} \\ & & - \frac{\sqrt{ n}}{32 (1 - n)^{\frac{3}{4}}} (-4 (1 + \sqrt{n})^{\frac{3}{4}} \ln(-(1 - \sqrt{n})^{\frac{1}{4}}) 4 i (1 + \sqrt{n})^{\frac{3}{4}} \ln(-i (1 - \sqrt{n})^{\frac{1}{4}}) \\ & & + 4 i (1 + \sqrt{n})^{\frac{3}{4}} \ln( i (1 - \sqrt{n})^{\frac{1}{4}}) + (1 + \sqrt{n})^{\frac{3}{4}} \ln( 1 - \sqrt{n}) + 4 (1 - \sqrt{n})^{\frac{3}{4}} \ln(-(1 + \sqrt{n})^{\frac{1}{4}}) \\ & & + 4 i (1 - \sqrt{n})^{\frac{3}{4}} \ln(-i (1 + \sqrt{n})^{\frac{1}{4}}) - 4 i (1 - \sqrt{n})^{\frac{3}{4}} \ln( i (1 + \sqrt{n})^{\frac{1}{4}}) \\ & & - (1 - \sqrt{n})^{\frac{3}{4}} \ln( 1 + \sqrt{n})) - \frac{32}{45 n}) \end{eqnarray*} \end{theorem \begin{proof} Integrate (\ref{eq:expan1}). \end{proof} \begin{theorem \begin{equation*} _{4} F _{3} (1,1,\frac{5}{2},3;2,\frac{21}{8},\frac{25}{8};1) = \frac{221}{276480} (12584 - 2^{\frac{1}{4}} 1305 \pi -2^{\frac{1}{4}} 2610 \coth^{-1}(2^{\frac{1}{4}}) + 2^{\frac{1}{4}} 2610 \tan^{-1}(2^{\frac{1}{4}})) \end{equation*} \end{theorem \begin{proof} Integrate both sides of {\bf Theorem 18}. \end{proof} \begin{theorem \begin{equation} _{3} F _{2} ( 1, \frac{3}{2} , 3 ; \frac{21}{8} , \frac{25}{8} ; 1 ) = \frac{221 }{4096} ( 8 + 2^{\frac{1}{4}} 27 \pi + 2^{\frac{1}{4}} 54 \coth^{-1}(2^{\frac{1}{4}}) - 2^{\frac{1}{4}} 54 \tan^{-1}(2^{\frac{1}{4}}) ) \end{equation} \end{theorem \begin{proof} Replace $x$ by $x^2$ of and integrate. \end{proof} Clearly, integrating ``generalized" power series may lead to outlandish results. Let us continue in a similar vein. \begin{theorem \begin{eqnarray*} \lefteqn{ _{5} F _{4} (1,\frac{9}{4},\frac{5}{2},\frac{11}{4},3;\frac{37}{16},\frac{41}{16},\frac{45}{16},\frac{49}{16} ;\frac{1}{c}) \frac{8388608 }{15862275 c^2} = } \\ & & \frac{1}{c} ( - \frac{2048}{3315} - c + \frac{ c^{\frac{5}{4}} \ln ( 1 - (1 - c^{\frac{1}{4}} )^{\frac{1}{4}} ) }{16 ( 1 - c^{\frac{1}{4}})^{\frac{3}{4}}} + \frac{ i c^{\frac{5}{4}} \ln ( 1 - i (1 - c^{\frac{1}{4}})^{\frac{1}{4}} ) }{ 16 (1 - c^{\frac{1}{4}})^{\frac{3}{4}} } - \frac{ i c^{\frac{5}{4}} \ln(1 + i (1 - c^{\frac{1}{4}})^{\frac{1}{4}}) }{ 16 (1 - c^{\frac{1}{4}})^{\frac{3}{4}}} \\ & & - \frac{ c^{\frac{5}{4}} \ln(1 + (1 - c^{\frac{1}{4}})^{\frac{1}{4}}) }{ 16 (1 - c^{\frac{1}{4}})^{\frac{3}{4}}} + \frac{ i c^{\frac{5}{4}} \ln(1 - (1 - i c^{\frac{1}{4}})^{\frac{1}{4}}) }{ 16 (1 - i c^{\frac{1}{4}})^{\frac{3}{4}}} - \frac{ c^{\frac{5}{4}} \ln(1 - i (1 - i c^{\frac{1}{4}})^{\frac{1}{4}}) }{ 16 (1 - i c^{\frac{1}{4}})^{\frac{3}{4}}} + \frac{ c^{\frac{5}{4}} \ln(1 + i (1 - i c^{\frac{1}{4}})^{\frac{1}{4}}) }{ 16 (1 - i c^{\frac{1}{4}})^{\frac{3}{4}}} \\ & & - \frac{ i c^{\frac{5}{4}} \ln(1 + (1 - i c^{\frac{1}{4}})^{\frac{1}{4}}) }{ 16 (1 - i c^{\frac{1}{4}})^{\frac{3}{4}}} - \frac{ i c^{\frac{5}{4}} \ln(1 - (1 + i c^{\frac{1}{4}})^{\frac{1}{4}}) }{ 16 (1 + i c^{\frac{1}{4}})^{\frac{3}{4}}} + \frac{ c^{\frac{5}{4}} \ln(1 - i (1 + i c^{\frac{1}{4}})^{\frac{1}{4}}) }{ 16 (1 + i c^{\frac{1}{4}})^{\frac{3}{4}}} - \frac{ c^{\frac{5}{4}} \ln(1 + i (1 + i c^{\frac{1}{4}})^{\frac{1}{4}}) }{ 16 (1 + i c^{\frac{1}{4}})^{\frac{3}{4}}} \\ & & + \frac{ i c^{\frac{5}{4}} \ln(1 + (1 + i c^{\frac{1}{4}})^{\frac{1}{4}})}{ 16 (1 + i c^{\frac{1}{4}})^{\frac{3}{4}}} - \frac{ c^{\frac{5}{4}} \ln(1 - (1 + c^{\frac{1}{4}})^{\frac{1}{4}}) }{16 (1 + c^{\frac{1}{4}})^{\frac{3}{4}}} - \frac{ i c^{\frac{5}{4}} \ln(1 - i (1 + c^{\frac{1}{4}})^{\frac{1}{4}})}{ 16 (1 + c^{\frac{1}{4}})^{\frac{3}{4}}} + \frac{ i c^{\frac{5}{4}} \ln(1 + i (1 + c^{\frac{1}{4}})^{\frac{1}{4}})}{ 16 (1 + c^{\frac{1}{4}})^{\frac{3}{4}}} \\ & & + \frac{c^{\frac{5}{4}} \ln(1 + (1 + c^{\frac{1}{4}})^{\frac{1}{4}}) }{ 16 (1 + c^{\frac{1}{4}})^{\frac{3}{4}}} - \frac{1}{64} c^{\frac{5}{4}} \frac{(4 \ln(-(1 - c^{\frac{1}{4}})^{\frac{1}{4}})}{(1 - c^{\frac{1}{4}})^{\frac{3}{4}} } + \frac{ 4 i \ln(-i (1 - c^{\frac{1}{4}})^{\frac{1}{4}})}{(1 - c^{\frac{1}{4}})^{\frac{3}{4}}} - \frac{ 4 i \ln(i (1 - c^{\frac{1}{4}})^{\frac{1}{4}})}{(1 - c^{\frac{1}{4}})^{\frac{3}{4}} } \\ & & - \frac{\ln(1 - c^{\frac{1}{4}}) }{ (1 - c^{\frac{1}{4}})^{\frac{3}{4}} } + \frac{ 4 i \ln(-(1 - i c^{\frac{1}{4}})^{\frac{1}{4}}) }{ (1 - i c^{\frac{1}{4}})^{\frac{3}{4}} } - \frac{ 4 \ln(-i (1 - i c^{\frac{1}{4}})^{\frac{1}{4}})}{ (1 - i c^{\frac{1}{4}})^{\frac{3}{4}}} + \frac{ 4 \ln(i (1 - i c^{\frac{1}{4}})^{\frac{1}{4}}) }{ (1 - i c^{\frac{1}{4}})^{\frac{3}{4}}} - \frac{ i \ln(1 - i c^{\frac{1}{4}}) }{ (1 - i c^{\frac{1}{4}})^{\frac{3}{4}} } \\ & & - \frac{ 4 i \ln(-(1 + i c^{\frac{1}{4}})^{\frac{1}{4}})}{(1 + i c^{\frac{1}{4}})^{\frac{3}{4}}} + \frac{ 4 \ln(-i (1 + i c^{\frac{1}{4}})^{\frac{1}{4}})}{ (1 + i c^{\frac{1}{4}})^{\frac{3}{4}}} - \frac{ 4 \ln(i (1 + i c^{\frac{1}{4}})^{\frac{1}{4}})}{ (1 + i c^{\frac{1}{4}})^{\frac{3}{4}}} + \frac{ i \ln(1 + i c^{\frac{1}{4}})}{ (1 + i c^{\frac{1}{4}})^{\frac{3}{4}} } - \frac{ 4 \ln(-(1 + c^{\frac{1}{4}})^{\frac{1}{4}}) }{ (1 + c^{\frac{1}{4}})^{\frac{3}{4}}} \\ & & - \frac{ 4 i \ln(-i (1 + c^{\frac{1}{4}})^{\frac{1}{4}}) }{ (1 + c^{\frac{1}{4}})^{\frac{3}{4}} } + \frac{ 4 i \ln(i (1 + c^{\frac{1}{4}})^{\frac{1}{4}}) }{ (1 + c^{\frac{1}{4}})^{\frac{3}{4}}} + \frac{ \ln(1 + c^{\frac{1}{4}}) }{ (1 + c^{\frac{1}{4}})^{\frac{3}{4}})} ) \end{eqnarray*} \end{theorem \begin{proof} Consider the appealing sum $\sum_{n=2}^{\infty} \frac{(1-x^4)^{4n}}{c^n} = \frac{(-1 + x^4)^8}{c (-1 + c + 4 x^{4} - 6 x^{8} + 4 x^{12} - x^16)} $. Integrate both sides of this equation. \end{proof} Consider integrating both sides of the above theorem. \begin{eqnarray*} \lefteqn{ _{6} F _{5} (1,1,\frac{9}{4},\frac{5}{2},\frac{11}{4},3;2,\frac{37}{16},\frac{41}{16},\frac{45}{16},\frac{49}{16};1) = } \\ & & ( \frac{332799152 }{ 32967675 } + \frac{ ( (272 - 1923 i) (1 - i)^{\frac{1}{4}} + (272 + 1923 i) (1 + i)^{\frac{1}{4}} - 2^{\frac{1}{4}} 28976 ) \pi }{53040 } \\ & & + \frac{( (1923 + 272 i) (1 - i)^{\frac{1}{4}} + (1923 - 272 i) (1 + i)^{\frac{1}{4}} ) \ln 2 }{ 26520 } - (\frac{641}{1105} + \frac{16 i}{195} ) (1 - i)^{\frac{1}{4}} \\ & & \ln(-(1 - i)^{\frac{1}{4}}) + (\frac{16}{195} - \frac{641 i}{1105}) (1 - i)^{\frac{1}{4}} \ln(-i (1 - i)^{\frac{1}{4}}) - (\frac{16}{195} - \frac{641 i}{1105}) (1 - i)^{\frac{1}{4}} \\ & & \ln(i (1 - i)^{\frac{1}{4}}) - (\frac{641}{1105} - \frac{16 i}{195} ) (1 + i)^{\frac{1}{4}} \ln(-(1 + i)^{\frac{1}{4}}) - (\frac{16}{195} + \frac{641 i}{1105}) (1 + i)^{\frac{1}{4}} \\ & & \ln(-i (1 + i)^{\frac{1}{4}}) + (\frac{16}{195} + \frac{641 i}{1105}) (1 + i)^{\frac{1}{4}} \ln(i (1 + i)^{\frac{1}{4}}) + (\frac{641}{1105} + \frac{16 i}{195} ) (1 - i)^{\frac{1}{4}} \\ & & \ln(1 - (1 - i)^{\frac{1}{4}}) - (\frac{16}{195} - \frac{641 i}{1105}) (1 - i)^{\frac{1}{4}} \ln(1 - i (1 - i)^{\frac{1}{4}}) + (\frac{16}{195} - \frac{641 i}{1105}) (1 - i)^{\frac{1}{4}} \\ & & \ln(1 + i (1 - i)^{\frac{1}{4}}) - (\frac{641}{1105} + \frac{16 i}{195} ) (1 - i)^{\frac{1}{4}} \ln(1 + (1 - i)^{\frac{1}{4}}) + (\frac{641}{1105} - \frac{16 i}{195} ) (1 + i)^{\frac{1}{4}} \\ & & \ln(1 - (1 + i)^{\frac{1}{4}}) + (\frac{16}{195} + \frac{641 i}{1105}) (1 + i)^{\frac{1}{4}} \ln(1 - i (1 + i)^{\frac{1}{4}}) - (\frac{16}{195} + \frac{641 i}{1105}) (1 + i)^{\frac{1}{4}} \\ & & \ln(1 + i (1 + i)^{\frac{1}{4}}) - (\frac{641}{1105} - \frac{16 i}{195} ) (1 + i)^{\frac{1}{4}} \ln(1 + (1 + i)^{\frac{1}{4}}) + \frac{ 1811 i 2^{\frac{1}{4}} \ln(1 - i 2^{\frac{1}{4}})}{ 3315 } \\ & & - \frac{ 1811 i 2^{\frac{1}{4}} \ln(1 + i 2^{\frac{1}{4}}) }{ 3315} + \frac{ 1811 2^{\frac{1}{4}} \ln(-1 + 2^{\frac{1}{4}}) }{ 3315 } - \frac{ 2^{\frac{1}{4}} 1811\ln(1 + 2^{\frac{1}{4}})}{3315} ) \frac{15862275}{8388608} \end{eqnarray*} Have hypergeometric series such as the above one been evaluated, published, or widely recognized? \begin{theorem \begin{eqnarray*} \lefteqn{ _{3} F _{2} (\frac{3}{2},2,2;\frac{13}{8},\frac{17}{8};\frac{1}{c}) = } \\ & & 1-\frac{45}{32} ( - ( \frac{1}{ 2880 (1 - c)^{\frac{7}{4}} } ) (-2048 (1 - c)^{\frac{3}{4}} + 2048 (1 - c)^{\frac{3}{4}} c - 360 (1 - c)^{\frac{3}{4}} c^{2} + 45 (1 + \sqrt{c})^{\frac{3}{4}} c^{\frac{3}{2}} \\ & & (-4 - 3 \sqrt{c} + c) \ln(1 - (1 - \sqrt{c})^{\frac{1}{4}}) + 45 i (1 + \sqrt{c})^{\frac{3}{4}} c^{\frac{3}{2}} (-4 - 3 \sqrt{c} + c) \ln(1 - i (1 - \sqrt{c})^{\frac{1}{4}}) \\ & & + 180 i (1 + \sqrt{c})^{\frac{3}{4}} c^{\frac{3}{2}} \ln(1 + i (1 - \sqrt{c})^{\frac{1}{4}}) + 135 i (1 + \sqrt{c})^{\frac{3}{4}} c^2 \ln(1 + i (1 - \sqrt{c})^{\frac{1}{4}}) - 45 i (1 + \sqrt{c})^{\frac{3}{4}} c^{\frac{5}{2}} \\ & & \ln(1 + i (1 - \sqrt{c})^{\frac{1}{4}}) + 180 (1 + \sqrt{c})^{\frac{3}{4}} c^{\frac{3}{2}} \ln(1 + (1 - \sqrt{c})^{\frac{1}{4}}) + 135 (1 + \sqrt{c})^{\frac{3}{4}} c^{2} \ln(1 + (1 - \sqrt{c})^{\frac{1}{4}}) \\ & & - 45 (1 + \sqrt{c})^{\frac{3}{4}} c^{\frac{5}{2}} \ln(1 + (1 - \sqrt{c})^{\frac{1}{4}}) + 180 (1 - \sqrt{c})^{\frac{3}{4}} c^{\frac{3}{2}} \ln(1 - (1 + \sqrt{c})^{\frac{1}{4}}) - 135 (1 - \sqrt{c})^{\frac{3}{4}} c^{2} \\ & & \ln(1 - (1 + \sqrt{c})^{\frac{1}{4}}) - 45 (1 - \sqrt{c})^{\frac{3}{4}} c^{\frac{5}{2}} \ln(1 - (1 + \sqrt{c})^{\frac{1}{4}}) + 180 i (1 - \sqrt{c})^{\frac{3}{4}} c^{\frac{3}{2}} \ln(1 - i (1 + \sqrt{c})^{\frac{1}{4}}) - \\ & & 135 i (1 - \sqrt{c})^{\frac{3}{4}} c^{2} \ln(1 - i (1 + \sqrt{c})^{\frac{1}{4}}) - 45 i (1 - \sqrt{c})^{\frac{3}{4}} c^{\frac{5}{2}} \ln(1 - i (1 + \sqrt{c})^{\frac{1}{4}}) - 180 i (1 - \sqrt{c})^{\frac{3}{4}} c^{\frac{3}{2}} \\ & & \ln(1 + i (1 + \sqrt{c})^{\frac{1}{4}}) + 135 i (1 - \sqrt{c})^{\frac{3}{4}} c^2 \ln(1 + i (1 + \sqrt{c})^{\frac{1}{4}}) + 45 i (1 - \sqrt{c})^{\frac{3}{4}} c^{\frac{5}{2}} \ln(1 + i (1 + \sqrt{c})^{\frac{1}{4}}) \\ & & - 180 (1 - \sqrt{c})^{\frac{3}{4}} c^{\frac{3}{2}} \ln(1 + (1 + \sqrt{c})^{\frac{1}{4}}) + 135 (1 - \sqrt{c})^{\frac{3}{4}} c^2 \ln(1 + (1 + \sqrt{c})^{\frac{1}{4}}) + 45 (1 - \sqrt{c})^{\frac{3}{4}} c^{\frac{5}{2}} \\ & & \ln(1 + (1 + \sqrt{c})^{\frac{1}{4}})) + ( \frac{1}{ 256 (1 - c)^{\frac{7}{4}} } ) c^{\frac{3}{2}} (4 (1 + \sqrt{c})^{\frac{3}{4}} (-4 - 3 \sqrt{c} + c) \ln(-(1 - \sqrt{c})^{\frac{1}{4}}) + 4 i (1 + \sqrt{c})^{\frac{3}{4}} \\ & & (-4 - 3 \sqrt{c} + c) \ln(-i (1 - \sqrt{c})^{\frac{1}{4}}) + 16 i (1 + \sqrt{c})^{\frac{3}{4}} \ln(i (1 - \sqrt{c})^{\frac{1}{4}}) + 12 i (1 + \sqrt{c})^{\frac{3}{4}} \sqrt{c} \ln(i (1 - \sqrt{c})^{\frac{1}{4}}) \\ & & - 4 i (1 + \sqrt{c})^{\frac{3}{4}} c \ln(i (1 - \sqrt{c})^{\frac{1}{4}}) + 4 (1 + \sqrt{c})^{\frac{3}{4}} \ln(1 - \sqrt{c}) + 3 (1 + \sqrt{c})^{\frac{3}{4}} \sqrt{c} \ln(1 - \sqrt{c}) - (1 + \sqrt{c})^{\frac{3}{4}} c \\ & & \ln(1 - \sqrt{c}) + 16 (1 - \sqrt{c})^{\frac{3}{4}} \ln(-(1 + \sqrt{c})^{\frac{1}{4}}) - 12 (1 - \sqrt{c})^{\frac{3}{4}} \sqrt{c} \ln(-(1 + \sqrt[c])^{\frac{1}{4}}) - 4 (1 - \sqrt{c})^{\frac{3}{4}} c \\ & & \ln(-(1 + \sqrt{c})^{\frac{1}{4}}) + 16 i (1 - \sqrt{c})^{\frac{3}{4}} \ln(-i (1 + \sqrt{c})^{\frac{1}{4}}) - 12 i (1 - \sqrt{c})^{\frac{3}{4}} \sqrt{c} \ln(-i (1 + \sqrt[c])^{\frac{1}{4}}) - 4 i (1 - \sqrt{c})^{\frac{3}{4}} \\ & & c \ln(-i (1 + \sqrt{c})^{\frac{1}{4}}) - 16 i (1 - \sqrt{c})^{\frac{3}{4}} \ln(i (1 + \sqrt{c})^{\frac{1}{4}}) + 12 i (1 - \sqrt{c})^{\frac{3}{4}} \sqrt{c} \ln(i (1 + \sqrt{c})^{\frac{1}{4}}) + 4 i (1 - \sqrt{c})^{\frac{3}{4}} c \\ & & \ln(i (1 + \sqrt{c})^{\frac{1}{4}}) - 4 (1 - \sqrt{c})^{\frac{3}{4}} \ln(1 + \sqrt{c}) + 3 (1 - \sqrt{c})^{\frac{3}{4}} \sqrt{c} \ln(1 + \sqrt{c}) + (1 - \sqrt[c])^{\frac{3}{4}} c \ln(1 + \sqrt{c}))) \end{eqnarray*} \end{theorem \begin{proof} Consider: $\sum_{n=2}^{\infty} \frac{((1 - x^4)^{2 n}) n}{c^n} = \frac{(-1 + x^4)^{4} (-1 + 2 c + 2 x^{4} - x^{8})}{c (-1 + c + 2 x^{4} - x^{8})^2} $. \end{proof} Such elaborate generalized hypergeometric series will perhaps serve as a decent conclusion to this paper. In summary, the author has presented somewhat straightforward techniques to evaluate multifarious hypergeometric series, and has evaluated a variety of hypergeometric series (some of which could possibly be original), and series which cannot be expressed by a single hypergeometric series (some of which could possibly be original).

https://arxiv.org/abs/1606.02320

On additive bases of sets with small product set

We prove that finite sets of real numbers satisfying $|AA| \leq |A|^{1+\epsilon}$ with sufficiently small $\epsilon > 0$ cannot have small additive bases nor can they be written as a set of sums $B+C$ with $|B|, |C| \geq 2$. The result can be seen as a real analog of the conjecture of Sárközy that multiplicative subgroups of finite fields of prime order are additively irreducible.

\section{Introduction} \label{sec:introduction} The duality between the additive and multiplicative structure of an arithmetic set, now called `sum-product phenomena' has been extensively studied since the seminal paper of Erd\H{o}s and Szemer\'edi \cite{ErdosSzemeredi}. In its classical formulation the duality is expressed in the fact that for a finite set of integers or reals $A$ either the set of pairwise sums $A+A$ or pairwise products $AA$ is significantly larger then the original set, unless $A$ is close to a subring of the ambient ring, see \cite{TaoVu} for details. In the current paper we consider an intrinsic version of the sum-product phenomenon in the following setting. Assume that a set $A \subset \mathbb{R}$ has small multiplicative doubling, so that $|AA| \leq |A|^{1+\epsilon}$ for some sufficiently small $\epsilon > 0$ (we assume that $\epsilon$ is fixed and $|A| > C(\epsilon)$ is large). Recall that a set $B$ is called a \emph{basis}\footnote{All bases in the present paper are assumed to be of order two.} for $A$ if each element in $A$ can be represented as a sum of two elements in $B$ or simply $A \subseteq B+B$. First, we show that if $A$ is multiplicatively structured in the above sense, then it does not admit small additive bases of order $|A|^{1/2+c}$ for an explicit constant $c > 0$. \begin{theorem} \label{thm:smalldoubling} There is an $\epsilon > 0$ such that the following holds: if $B$ is an arbitrary additive basis for a real set $A$ with $|AA| \leq |A|^{1+\epsilon}$ then $$ |B| \gg |A|^{1/2 + 1/442 - o(1)}. $$ \end{theorem} Theorem \ref{thm:smalldoubling} implies a `power-saving' estimate for the multiplicative energy $E_\times$ of a sumset or a difference set, significantly improving on the sub-exponential bound of \cite{RocheNewtonZhelezov}. \begin{corr} \label{corr:energy} There is an effective constant $\epsilon_0 > 0$ such that for any set $B$ of real numbers holds $$ E_{\times} (B \pm B) \leq |B|^{6 - \epsilon_0}. $$ \end{corr} In fact, we prove an even stronger statement than that of Theorem \ref{thm:smalldoubling}, replacing the condition $A \subseteq B + B$ by a weaker assumption that the number of pairs $(b_1, b_2) \in B \times B : b_1 + b_2 \in A$ is large, see the upcoming Lemma \ref{lm:convolutions} for details. While slightly more technical, such a reformulation may be useful if one wishes to take into account the number of ways an element in $A$ can be represented as a sum $b_1 + b_2$, see Remark \ref{r:sigma}. Second, we show that $A$ is \emph{additively irreducible}, i.e. it cannot be written as a set of sums $B+C$ unless one of the sets consists of a single element. \begin{theorem} \label{thm:irreducibility} There is an $\epsilon > 0$ such that for all sufficiently large $A \subset \mathbb{R}$ with $|AA| \leq |A|^{1+\epsilon}$ there is no decomposition $A = B + C$ with $|B|, |C| \geq 2$. \end{theorem} Both theorems substantially extend the results of Roche-Newton and the second author \cite{RocheNewtonZhelezov}, where a much more restrictive condition $|AA| \leq K|A|$ with $K$ fixed, was assumed. In particular, in the case at hand deep structural results such as the Freiman-type theorem of Sanders (see e.g. \cite{SandersExp}) and the Subspace theorem of Evertse, Schlickewei and Schmidt \cite{evertse2002linear} are no longer available. Instead, we rely solely on techniques of additive combinatorics, which we see as the main innovation of the present paper. On the other hand, Theorem \ref{thm:irreducibility} extends the result of the first author \cite{ShkredovSmallProductIrreducibility}, which is the specialisation of Theorem \ref{thm:irreducibility} to the case $C = -B$. Clearly, Theorem \ref{thm:irreducibility} has the following sum-product result as a corollary. \begin{corr} \label{corr:sum-product} There is an absolute constant $c > 0$ such that for any real sets $B, C$ with $|B|, |C| \geq 2$ holds $$ \left| (B+C)(B+C) \right| \gg |B+C|^{1+c}. $$ \end{corr} Both questions considered in the paper arise as a natural extension of classical problems concerning thin additive bases and additive reducibility of integer sequences to the finite setting. One of the basic questions, posed by Erd\H{o}s and Newman \cite{ErdosNewman} in 1976, is to estimate the minimal basis size for a given set $A$ and, in particular, to decide if there is a very thin basis for $A$ of size of order $|A|^{1/2}$. They noted that while for a randomly picked set $A$ one should expect that no such basis exists, it is in general hard to prove such a claim for a specific set $A$. The question of additive irreducibility of integer sequences was posed by Ostmann \cite{Ostmann} back in 1956. Perhaps the most notable conjecture of Ostmann, still wide open, is that the set of prime numbers cannot be written as a non-trivial set of sums $B+C$ even if one allows to discard a finite number of elements. We refer the reader to \cite{elsholtz2006additive} for the history of the problem and \cite{ElsholtzHarper} for the state of the art partial results (see also \cite{GreenHarper}). S\'ark\"ozy \cite{SarkozyDecompositions} (see also \cite{gyarmati2013reducible} and references therein) extended such problems to the finite field setting, perhaps led by the intuition that in general `multiplicatively structured' sets should be additively irreducible (though this intuition may sometimes be misleading, as showed by Elsholtz in \cite{Elsholtz2008}). As a special case of this program, S\'ark\"ozy conjectured that multiplicative subgroups of finite fields of prime order are additively irreducible, in particular the set of quadratic residues modulo a prime. Despite some progress (see e.g. \cite{ShkredovQuadraticResidues} and references therein), the conjecture of S\'ark\"ozy is an important open problem with rich connections (to e.g. multiplicative character sums). From this point of view, Theorem \ref{thm:irreducibility} can be seen as a real analog of S\'ark\"ozy's conjecture for small subgroups, and in fact the proof can be transferred to the finite field setting except a certain sum-product-type estimate not currently available in finite fields. We discuss this matter in more detail at the end of the paper. The notation is briefly explained in the next section and is relatively standard in additive combinatorics. We recommend the reader to consult \cite{TaoVu} for further details when needed. \section{Notation} \label{sec:notation} The following notation is used throughout the paper. The expressions $X \gg Y$, $Y \ll X$, $Y = O(X)$, $X = \Omega(Y)$ all have the same meaning that there is an absolute constant $c > 0$ such that $|Y| \leq c|X|$. The expressions $X \gtrsim Y$ and $Y \lesssim X$ both mean that $|Y| = O(|X| \log^C |X|)$ for some absolute constant $C > 0$. For a graph $G$, $E(G)$ denotes the set of edges and $V(G)$ denotes the set of vertices. If $X$ is a set then $|X|$ denotes its cardinality. For sets of numbers $A$ and $B$ the sumset $A + A$ is the set of all pairwise sums $\{ a + a' : a, a' \in A \}$, and similarly $AA$, $A-A$ denotes the set of products and differences, respectively. If $G$ is some graph with the vertex set identified with $A$ then $A\stackrel{G}{+}A$ denotes the set of sums $\{ a + a' : (a, a') \in E(G) \}$ (with the obvious generalisation to all arithmetic operations). For a bipartite graph $G$ with parts $(V_1, V_2)$ we will occasionally use the term \emph{bipartite density} for the quantity $\frac{|E(G)|}{|V_1||V_2|}$. For a general graph, the edge \emph{density} is defined as the quotient $\frac{|E(G)|}{|V(G)|^2}$, where the standard notation $E(G), V(G)$ is used for the set of edges and vertices, respectively. For a vertex $v \in V(G)$, $N(v)$ denotes the set of neighbours of $v$. The additive energy $E_+(A)$ (see \cite{TaoVu}, Definition 2.8) denotes the number of additive quadruples $(a_1, a_2, a_3, a_4)$ such that $a_1 + a_2 = a_3 + a_4$. The multiplicative energy $E_\times$ is defined similarly as the number of multiplicative quadruples. \section{Proof of Theorem \ref{thm:smalldoubling}} \label{sec:proof_smd} In what follows we will always assume that $0 \notin A$ but $1 \in A$ which one can do without loss of generality. We will use the following result of the first author (Theorem 5.4 in \cite{shkredov2013}). \begin{theorem} \label{thm:ShkredovEnergyBound} Let $A \subset \mathbb{R}$ be such that $|AA| = M|A|$. Then $$ E_{+}(A) \ll M^{14/13}|A|^{32/13} \log^{71/65} |A|. $$ \end{theorem} Combining with the Pl\"unnecke-Ruzsa inequality (\cite{TaoVu}, Corollary 6.29), we have the following corollary. \begin{corr} \label{corr:energy_bound_upper} For any $c < 1/26$ there is $\eps > 0$ such that for any $A \subset \mathbb{R}$ with $|AA| \leq |A|^{1+\eps}$ holds $$ E_+(AA/A) \ll |A|^{5/2 - c}. $$ \end{corr} Recall the following graph--theoretic lemma of Gowers (see e.g. \cite{TaoVu}, Lemma 6.19). \begin{lemma} \label{lm:pairs} Let $G$ be a bipartite graph on $(B_1, B_2)$ where $|B_1| = |B_2| = n$ and $0 < \alpha = E(B_1, B_2)/n^2$. Let $0 < \epsilon < 1$ be fixed. Then there is a subset $B'_1 \subseteq B_1$ with $|B'_1| \geq \alpha n/2$ such that for at least $(1-\epsilon)|B'_1|^2$ of the ordered pairs of vertices $(v_1, v_2) \in B'_1 \times B'_1$ the following holds: $$ |N(v_1) \cap N(v_2)| \geq \frac{\epsilon \alpha^2 n}{2}. $$ \end{lemma} In what follows, we use a weaker definition of a basis, merely assuming that the number of pairs $(b_1, b_2) \in B \times B: b_1 + b_2 \in A$ is large. First, it allows one to take into account `multiplicities', i.e. elements in $A$ which have many representations as a sum. Second, it gives additional flexibility by relaxing the condition that $B+B$ contains \emph{all} elements in $A$. For this matter, let us call a set $B$ an $(L, K)$\emph{-basis} for $A$, with $K,L \ge 1$, if $|B| = K|A|^{1/2}$ and $$ \left| \{(b_1, b_2) : b_1 + b_2 \in A \} \right| = L^{-1}|A|. $$ We will also make a use of the \emph{containment graph} $G$, which is a bipartite graph $G$ on $(B, B)$ such that $(b_1, b_2) \in E(G)$ if and only if $b_1 + b_2 \in A$. In particular, if $B$ is an $(L, K)$-basis, one has $|E(G)| = L^{-1}|A| = |B|^2/LK^2$, so the bipartite edge density of $G$ is $L^{-1}K^{-2}$. \begin{lemma}[The set of popular differences contains an almost closed difference set] \label{lm:convolutions} Let $B$ be an $(L, K)$-basis for $A$. Then there is $B' \subset B$ with $|B'| \gg L^{-1}K^{-2}|B|$ such that for at least $0.99 |B'|^2$ of the ordered pairs $(b_1, b_2) \in B' \times B'$ the equation \beq \label{eq:representations} b_1 - b_2 = a - a' \eeq has at least $\Omega(L^{-2}K^{-3}|A|^{1/2})$ solutions $(a, a') \in A \times A$. \end{lemma} \begin{proof} Applying Lemma \ref{lm:pairs} to the containment graph $G$ we obtain a set $B'$ of size $\Omega(L^{-1}K^{-2}|B|)$ such that for $0.99 |B'|^2$ of the ordered pairs $(b_1, b_2) \in B' \times B'$ holds $$ |N(b_1) \cap N(b_2)| \geq \Omega(L^{-2}K^{-4}|B|). $$ On the other side, if $b \in N(b_1) \cap N(b_2)$ then by construction $b + b_1 = a$ and $b + b_2 = a'$ for some $a, a' \in A$, which after rearranging gives $$ b_1 - b_2 = a - a'. $$ Since, $b + b_1$ are all distinct as $b$ varies, we obtain at least $|N(b_1) \cap N(b_2)|$ distinct solutions of (\ref{eq:representations}) for a fixed pair $(b_1, b_2)$ (and $(b_2, b_1)$ as well). The result follows. \end{proof} Another ingredient is a beautiful incidence result, originally proved by Jones \cite{Jones} with a shorter proof discovered by Roche-Newton \cite{roche2015short}. \begin{lemma} \label{thm:RN} Let $A \subset \mathbb{R}$. Then the number of solutions to \beq \label{eq:triangles_collisions} (a-b)(a'-c') = (a-c)(a'-b') \eeq such that $a, a', b, b', c, c' \in A$ is $O(|A|^4 \log |A|)$. \end{lemma} The following lemma is crucial for the proof. \begin{lemma}[Generating a large set of popular differences] \label{lm:popular_diff} Let $B$ be an $(L, K)$-basis for $A$. Then there is a set $R$ such that the following holds. \begin{enumerate} \item[(i)] $R \subseteq A/A$. \item[(ii)] $|R| \gtrsim L^{-8}K^{-14}|A|$. \item[(iii)] For any $x \in R$, the equation $$ 1-x = \alpha_1 - \alpha_2 $$ has at least $\Omega(L^{-2}K^{-3}|A|^{1/2})$ distinct solutions $(\alpha_1, \alpha_2)$ with $\alpha_1, \alpha_2 \in A/A$. \end{enumerate} \end{lemma} \begin{proof} Let $B'$ be the set given by Lemma \ref{lm:convolutions}. Let us call a pair $(b_1, b_2) \in B' \times B'$ \emph{rich} if $|N(b_1) \cap N(b_2)| \gg L^{-2}K^{-3}|A|^{1/2}$ in the containment graph $G$, so that at least 99\% of pairs are rich. Take $$ R = \left \{ \frac{b_2 + b}{b_1 + b} : (b_1, b_2) \text{ is rich}, b \in N(b_1) \cap N(b_2) \right\}. $$ The first claim follows from the construction of the containment graph, so it remains to prove (ii) and (iii). For $x \in R$, define $$ n(x) := \left| \left\{(b_2, b_1, b) : x = \frac{b_2 + b}{b_1 + b}, (b_1, b_2) \text{ is rich}, b \in N(b_1) \cap N(b_2) \right\} \right|. $$ and $$ Q := \left| \left\{ (b_2, b_1, b, b'_2, b'_1, b')\in B^6 : \frac{b_2 + b}{b_1 + b} = \frac{b'_2 + b'}{b'_1 + b'} \right\} \right|. $$ We get by the Cauchy-Schwartz inequality $$ L^{-2}K^{-3}|B'|^2|A|^{1/2} \ll \sum_{x \in R} n(x) \leq |R|^{1/2} \left(\sum_{x \in R} n^2(x) \right)^{1/2} \leq |R|^{1/2}Q^{1/2}. $$ But by Lemma \ref{thm:RN} we have $Q \ll |B|^4 \log |B|$ and combining with the bound $|B'| \gg L^{-1}K^{-1}|A|^{1/2}$ we obtain $$ L^{-8}K^{-14}|A| \lesssim |R|. $$ For the third bullet, for an $x \in R$ fix $(b_1, b_2, b)$ such that $x = (b_2 + b)/(b_1 + b)$ and observe that \begin{equation}\label{f:1-x} 1 - \frac{b_2 + b}{b_1 + b} = \frac {b_1 - b_2}{b_1 + b} = \frac {b_1+ b'}{b_1 + b} - \frac{b_2 + b'}{b_1 + b}. \end{equation} If we vary over $b' \in N(b_1) \cap N(b_2)$, the pairs $(b_1 + b')/(b_1 + b), (b_2 + b')/(b_1 + b))$ are all distinct, and the claim follows since $b_1 + b', b_2 + b', b_1 + b$ are all in $A$ by construction. \end{proof} The next lemma is simple but powerful. \begin{lemma}[Generating a larger set additive quadruples] \label{lm:more_popular_diff} Assume $1 \in X$. Let $R \subset X$ such that for any $x \in R$ the equation $$ 1 - x = \alpha_1 - \alpha_2 $$ has at least $N$ solutions $(\alpha_1, \alpha_2) \in X \times X$. Then for any set $Y$ one has the estimate $$ E_+(YX) \geq N|Y||R|. $$ \end{lemma} \begin{proof} By definition, $E_+(YX)$ is the number of additive quadruples $(y_1, y_2, y_3, y_4)$ such that $y_i \in YX$ and $y_1 - y_2 = y_3 - y_4$. If $1 - x = \alpha_1 - \alpha_2$ with $x, \alpha_1, \alpha_2 \in X$, then clearly $(y, yx, y\alpha_1, y\alpha_2)$ is an additive quadruple with elements in $YX$ (remember that $1 \in X$ so $y \in YX$). It remains to check that such quadruples are all distinct. But if $$ (y, yx, y\alpha_1, y\alpha_2) = (y', y'x', y'\alpha'_1, y'\alpha'_2) $$ then $y = y'$ and $(x, \alpha_1, \alpha_2) = (x', \alpha'_1, \alpha'_2)$, therefore the number of additive quadruples is at least $|Y|$ times the number of distinct triples $(x, \alpha_1, \alpha_2)$, which is at least $N|R|$. \end{proof} It remains to put everything together. \begin{proof}{(of Theorem \ref{thm:smalldoubling})} Let $B$ be an $(L, K)$-basis. Applying Lemma \ref{lm:popular_diff} and then Lemma \ref{lm:more_popular_diff} with $Y = A$ and $X = R$ from Lemma \ref{lm:popular_diff} we obtain \begin{equation} \label{eq:LKenergybound} E_+(AA/A) \gg L^{-2}K^{-3}|A|^{3/2}|R| \gtrsim L^{-10}K^{-17}|A|^{5/2}. \end{equation} On the other hand, by Corollary \ref{corr:energy_bound_upper} $$ E_+(AA/A) \lesssim |A|^{5/2 - c} $$ provided $|AA|/|A|$ is small enough. We then have \begin{equation} \label{eq:LKbound} L^{10}K^{17} \gtrsim |A|^{c}. \end{equation} In particular, if $B$ is a basis for $A$ then $$ |B| \gtrsim |A|^{1/2 + c/17}. $$ \end{proof} We record an immediate Corollary of Theorem \ref{thm:smalldoubling}. \begin{corr} There is an $\epsilon > 0$ and an absolute constant $c > 0$, such that for any $A \subset \mathbb{R}$ with $|AA| \leq |A|^{1+\epsilon}$ the following holds. If $A \subset B+C$ for some real sets $B, C$ then $$ \max(|B|, |C|) \gg |A|^{1/2 + c}. $$ If $B + C \subset A$ for some real sets $B, C$ then $$ \min(|B|, |C|) \ll |A|^{1/2 - c}. $$ \end{corr} \begin{proof} The first claim follows immediately from Theorem \ref{thm:smalldoubling} and the fact that $B \cup C$ is a basis for $A$ of size $O(\max(|B|, |C|))$. In the second case, assume $|B| \leq |C|$ and let $B'$ be an arbitrary subset of $C$ of size $|B|$. Then $B \cup B'$ is a basis for $A' := B+B'$ of size at most twice $\min(|B|, |C|)$. Applying Lemma \ref{lm:popular_diff} and then Lemma \ref{lm:more_popular_diff} we obtain similarly to (\ref{eq:LKenergybound}) $$ E_+(A'A'/A') \gtrsim |B'|^2 |B'| |B'|^2 = |B'|^{5}. $$ But clearly $A' \subset A$, so $E_+(A'A'/A') \leq E_+(AA/A)$. On the other hand, by Corollary \ref{corr:energy_bound_upper} $$ E_+(AA/A) \lesssim |A|^{5/2 - c} $$ and the claim follows. \end{proof} The proof of Corollary \ref{corr:energy} is a simple application of the Balog-Szemer\'edi-Gowers theorem (BSG) (see \cite{TaoVu}) so we present it in a sketchy manner. Let $X = B \pm B$. Since trivially $E_{\times}(X) \ll |X|^3$ we may assume that $|X| \gg |B|^{2 - \epsilon}$ with $\epsilon$ to be defined in due course. If now one assumes that $$ E_\times(X) \gg |X|^{3-\epsilon}, $$ then by BSG there is $X' \subset X$ such that $|X'X'| \leq |X'|^{1+\epsilon'}$ and $|X'| \gg |X|^{1-\epsilon'}$ with $\epsilon'$ depending polynomially on $\epsilon$. But $X' \subset X = B \pm B$ and whence $B \cup (-B)$ is a basis for $X'$ of size $O(|X'|^{1/2 + \epsilon''})$ with polynomial dependence of $\epsilon''$ on $\epsilon$. Taking $\epsilon$ small enough, we contradict Theorem \ref{thm:smalldoubling} and the claim follows. \begin{remark} \label{r:sigma} The following question was posed in \cite{ShkredovSmallProductIrreducibility}. Assume that $A\subseteq B-B$, $|AA| \le |A|^{1+\varepsilon}$, where $\varepsilon > 0$ is small. Is it true that there is $\delta= \delta(\varepsilon)$ such that $$ \sigma_A (B) := \sum_{x\in A} |\{ b_1-b_2 = x ~:~ b_1,b_2 \in B \}| \ll |B|^{2-\delta} \quad \mbox{?} $$ It is easy to see that our methods allow one to resolve the problem in the affirmative. Indeed, put $\sigma = \sigma_A (B)$ and consider the graph with $B$ as the vertex set and two vertices $x,y$ adjacent iff $x-y \in A$. Then the edge density of the graph is $\alpha := \sigma / |B|^2$. Next, repeat the arguments of Section \ref{sec:proof_smd} and obtain that for a sufficiently small $\varepsilon$ holds $$ |A|^{5/2-c} \gtrsim \alpha^2 |B| |A| |R| \gtrsim \alpha^{10} |B|^3 |A| $$ and hence since $A\subseteq B-B$ \begin{equation}\label{tmp:06.06.2016} \sigma \lesssim |B|^2 \cdot |A|^{-c/10} \cdot \left( \frac{|A|^{3/2}}{|B|^3} \right)^{1/10} \leq |B|^2 \cdot |A|^{-c/10} \,. \end{equation} The asymmetric case with two different sets $B,C$ is considered in the next section. In this case one can obtain an analog of the upper bound (\ref{tmp:06.06.2016}) of the form $$ \sum_{x\in A} |\{ b+c = x ~:~ b\in B,\, c\in C \}| \lesssim |B||C| \cdot |A|^{-c/10} \,. $$ \end{remark} \begin{remark} It is highly probable that the exponent $1/2+1/442$ in Theorem \ref{thm:smalldoubling} can be improved, perhaps even to $1-o(1)$. We pose a much more modest question, which however does not seem to follow from the results of the present paper. Is it true that there exists $c, \epsilon > 0$ such that for all sufficiently large real sets $B$ such that $A \subset B+B $ with $|AA| \leq |A|^{1+\epsilon}$ holds $$ |B+B| \gg |A|^{1+c} \,\, ? $$ \end{remark} \section{Additive irreducibility of multiplicative sets} \label{sec:decomposition} In this section we prove Theorem \ref{thm:irreducibility}. From now on we assume for the sake of contradiction that $A = B+C$ with $|B| \geq |C| \geq 2$ and $|AA| \leq |A|^{1+\epsilon}$ with $\epsilon$ small enough. First, note that it follows from Lemma 29 of \cite{ShkredovSmallProductIrreducibility} that for $\alpha \neq 0$ and $A$ such that $|AA| \leq M|A|$ holds \begin{equation} \label{eq:convolution_bound} |A \cap (A + \alpha)| \leq M^{4/3}|A|^{2/3}. \end{equation} For any $c_1 \neq c_2 \in C$ one has $(B + c_1) \subseteq A \cap (A + (c_1 - c_2))$, and thus by (\ref{eq:convolution_bound}) and the trivial bound $|B||C| \geq |A|$ one concludes that $|B|, |C| \gg |A|^{1/3 - \epsilon}$, say. Hence, we can safely assume that both $|B|, |C|$ are large. An inspection of the proof of Theorem \ref{thm:smalldoubling} reveals that if we put $$ X = \left\{ \frac{b_1+c}{b_2+c} : b_1, b_2 \in B, c \in C \right\} $$ and $$ Y = \left\{ \frac{c_1+b}{c_2+b} : b \in B, c_1, c_2 \in C \right\} $$ then one has that $$ E_+(AA/A) \gg \min \large( |A||X||C|, |A||Y||B| \large). $$ Thus, by Corollary \ref{corr:energy_bound_upper} one has $$ |X||C| \ll |A|^{3/2 - c} $$ and \beq \label{eq:energy_C} |Y||B| \ll |A|^{3/2 - c} \eeq for some explicit $c > 0$. However, sufficiently good lower bounds for $|X|$ and $|Y|$ are not readily available. Let $T(X, Y, Z)$ be the number of collinear triples of distinct points $(x, y, z)$ with $x \in X \times X$, $y \in Y \times Y$, $z \in Z \times Z$. Following the lines of Lemma \ref{lm:popular_diff} we have by the Cauchy--Schwarz inequality \beq \label{eq:bound_X} |X| \gg \frac{|B|^4|C|^2}{T(B, B, -C)} \,, \eeq and \beq \label{eq:bound_Y} |Y| \gg \frac {|C|^4|B|^2}{T(C, C, -B)} \,, \eeq Indeed, without loss of generality it suffices to check that if, say, $$ y := \frac{c_1+b}{c_2+b} = \frac{c'_1+b'}{c'_2+b'} $$ then either the points $(c_1, c'_1), (c_2, c'_2), (-b, -b')$ are all distinct and collinear or $y \in \{0, 1, \infty \}$. But we can simply exclude such degenerate values from $X$ and $Y$ and thus justify (\ref{eq:bound_X}) and (\ref{eq:bound_Y}). We prove the following estimate with $|C| \leq |B|$ which improves on Lemma \ref{thm:RN} when $|B|$ is significantly larger than $|C|$. \begin{lemma}[Bounding collinear triples for different sets] \label{lm:T_CCB} For sets $C \times C$ and $B \times B$ with $|B| \geq |C|$ holds $$T(C, C, B) \ll |B|^{4/3}|C|^{8/3} \log^2 |B| \,.$$ \end{lemma} \begin{proof} Write $L_{i, j}$, $i,j\ge 0$ for the set of lines $\ell$ such that $2^{i} \leq |\ell \cap (C \times C)| < 2^{i+1}$ and $2^{j} \leq |\ell \cap (B \times B)| < 2^{j+1}$. Then $$ T(C, C, B) \ll \sum^{\log |C|}_{i = 0}\, \sum^{\log |B|}_{j = 0} |L_{i, j}| 2^{2i}2^{j}. $$ Since the number of summands is at most $\log^2 |B|$ it is enough to bound each term by $|B|^{4/3}|C|^{8/3}$. For the sake of notation, denote $k = 2^i$ and $l = 2^j$, $L = L_{i,j}$ so that out task is to estimate $|L|k^2l$ where $L$ is the set of lines intersecting $C \times C$ in $k$ (up to a factor of two) points and $B \times B$ in $l$ points (again, up to a factor of two). By the Szemer\'edi-Trotter theorem \cite{TaoVu}, for $k \geq 2, l \geq 2$ holds $$ |L| \ll \min \left( \frac{|C|^4}{k^3} + \frac{|C|^2}{k}, \frac{|B|^4}{l^3} + \frac{|B|^2}{l} \right), $$ so $$ T := k^2l|L| \ll \min \left( \frac{ l |C|^4}{k} + kl|C|^2, \frac{k^2|B|^4}{l^2} + k^2|B|^2 \right) := \min (M_1, M_2). $$ Before we proceed, let us rule out the cases not covered by the Szemer\'edi-Trotter theorem as stated above. Since each line contains at least two distinct points in $C \times C$, only the case $l = 1$ should be considered separately. In this case we have $$ T \ll \frac{|C|^4}{k} + k|C|^2 \ll |C|^4 \leq |B|^{4/3}|C|^{8/3}, $$ which is the desired bound. Next, since always $k \leq |C|$, $l \leq |B|$, we obtain that $\frac{ l |C|^4}{k} \gg M_1$ and $\frac{k^2 |B|^4}{l^2} \gg M_2$, so $$ T^3 \ll M_1^2 M_2 \ll \left( \frac{ l |C|^4}{k} \right)^2 \cdot \frac{k^2 |B|^4}{l^2} = |C|^8 |B|^4 $$ or $$ T \ll |B|^{4/3} |C|^{8/3} \,. $$ This finishes the proof. \end{proof} Combining Lemma \ref{lm:T_CCB}, estimate (\ref{eq:bound_Y}) and the trivial bounds $|B||C| \geq |A|, |B| \geq |A|^{1/2}$, we have $T(C, C, B) \lesssim |C|^{8/3}|B|^{4/3} $ and $$ |Y||B| \gtrsim |C|^{4/3}|B|^{2/3}|B| = (|C||B|)^{4/3} |B|^{1/3} \geq |A|^{4/3 + 1/6} = |A|^{3/2}, $$ which contradicts (\ref{eq:energy_C}) if $|A|$ is large enough. Thus, $A$ is additively irreducible. \begin{remark} Instead of using the graph approach as it was done in Sections \ref{sec:proof_smd} and \ref{sec:decomposition} one can apply in the proof of Theorem \ref{thm:irreducibility} the following analog of the formula (\ref{f:1-x}) $$ 1-\frac{b_1+c}{b_2+c} = \frac{b_2+c'}{b_2+c} \left( 1 - \frac{b_1+c'}{b_2+c'} \right), $$ which holds for any $b_1,b_2\in B$ and all $c,c'\in C$. It means in particular, in the notation of Section \ref{sec:decomposition}, that taking an arbitrary $x\in X$, we have at least $|C|$ solutions to the equation $$ 1-x = y (1-x_*) \,, $$ where $y\in Y$ and $x_* \in X$ are fixed. Of course, one can replace $X$ and $Y$ in the last formula. It immediately gives that $E_{+} (AA/A) \ge |X||C||A|$ and $E_{+} (AA/A) \ge |Y||B||A|$. \end{remark} \section{Discussion} \label{sec:discussion} First, let us note that the only feature of the reals we have used, which is not available in an arbitrary field, is the Szemer\'{e}di-Trotter theorem (which is also used in the proof of Theorem \ref{thm:ShkredovEnergyBound}). Thus, our results can be readily extended to sets of complex numbers thanks to the extensions of the Szemer\'{e}di-Trotter theorem to the complex plane due to Zahl \cite{MR3392965} and Toth \cite{toth2015szemeredi}. However, when we replace $A$ with a sufficiently small multiplicative subgroup of $\mathbb{F}_p$, the situation becomes more subtle. For multiplicative subgroups Theorem \ref{thm:ShkredovEnergyBound} can be substituted with a similar energy bound, beating the exponent $5/2$, as was shown by the first author \cite{Sh_ineq}. Next, it follows from the result of Shkredov and Vyugin \cite{MR2984656} (see also the paper \cite{ShparlinskiGroupDecomposition} by Shparlinski) that if a multiplicative subgroup $G$ is written as a non-trivial sumset $B+C$, then necessarily $|G|^{1/2+o(1)} \ll |B|, |C| \ll |G|^{1/2+o(1)}$. Further, a slightly weaker form of the Szemer\'{e}di-Trotter bound for the number of incidences for Cartesian products in $\mathbb{F}^2_p$ is now available, see \cite{AMRS}. Thus, the only remaining obstacle for translating Theorem \ref{thm:irreducibility} to subgroups is to find a substitute for Lemma \ref{thm:RN}, which we were unable to do. However, as was shown in \cite{AMRS}, see also \cite{RudnevRocheShkredov} the bound $O(|A|^4 \log |A|)$ of Lemma 4 can indeed be replaced with $O(|A|^{9/2})$, which only barely falls short of the desired bound. Finally, let us mention that similarly to Section \ref{sec:proof_smd} one can slightly weaken the hypothesis of Theorem \ref{thm:irreducibility} and assume only that the number of pairs $(b, c) \in B \times C : b + c \in A$ is large. We leave the details for the interested reader. \section*{Acknowledgements} The authors would like to thank Tomasz Schoen, Misha Rudnev and Oliver Roche-Newton for useful discussions and the anonymous referees for valuable suggestions. The second author would like to acknowledge the hospitality of Adam Mickiewicz University in Poznan at which the work on the present paper was partially conducted. The work of I. D. Shkredov was supported by Russian Scientific Foundation grant 14-11-00433. \bibliographystyle{plain}

https://arxiv.org/abs/2103.05418

Poisson Approximations and Convergence Rates for Hyperbolic Dynamical Systems

We prove the asymptotic functional Poisson laws in the total variation norm and obtain estimates of the corresponding convergence rates for a large class of hyperbolic dynamical systems. These results generalize the ones obtained before in this area. Applications to intermittent solenoids, Axiom A systems, Hénon attractors and to billiards, are also considered.

\section{Introduction}\ \par The studies of Poisson approximations of the process of recurrences to small subsets in the phase spaces of chaotic dynamical systems, started in \cite{Pitskel}, are developed now into a large active area of the dynamical systems theory. Another view at this type of problems is a subject of the theory of open dynamical systems \cite{yorke}, where some positive measure subset $A$ of the phase space is named a hole, and hitting and escape the hole processes are studied. The third view at this type of problems concerns statistics of extreme events (``record values") in the theory of random processes \cite{Freitas}. In this paper we present new advances in this area. In a general set up, one picks a small measure subset $A$ in the phase space $\mathcal{M}$ of hyperbolic (chaotic) ergodic dynamical system and attempts to prove that in the limit, when the measure of $A$ approaches zero, the corresponding process of recurrences to $A$ converges to the Poisson process. This area received an essential boost in L-S.Young papers \cite{Y,Y1}, where a new general framework for analysis of statistical properties of hyperbolic dynamical systems was introduced. This approach employs representation of the phase space of a dynamical system as a tower (later called Young tower, Gibbs-Markov-Young tower, etc), which allow to study dynamics by analysing recurrences to the base of this tower. Several developments of this method were proposed later, essentially all focused on the dynamical systems with weak hyperbolicity (slow decay of correlations). For such systems the method of inducing was employed, when the base of the tower is chosen as such subset of the phase space where the induced dynamics, generated by the recurrences to the base, is strongly hyperbolic \cite{chernovhighdim, markarian,chernov-dolgopyat}. Our approach to the Poisson approximations is slightly different. It employs pulling back a hole $A$ to a nice (strongly hyperbolic) reference set in the phase space, e.g., the base of the Young tower. This pull back method allows to improve various results previously obtained in this area. The main results (Theorems \ref{thm} and \ref{thm2}) of the paper are dealing with convergence of a random process, generated by the measure preserving dynamics, to the functional Poisson law in the total variation (TV) norm. We also obtain estimates of the corresponding convergence rates in the form \begin{equation}\label{equaintro} d_{TV}(N^{r,z,T}, P) \precsim_{T,z} r^{a} \text{ for almost every } z \in \mathcal{M} , \end{equation} where $P$ is a Poisson point process and $N^{r,z,T}$ is a dynamical point process which counts a number of entrances by an orbit to a metric ball $B_{r}(z)$ with radius $r$ and the center $z$ in the phase space of a dynamical system during the time interval $[0,T]$. The notation $\precsim_{T,z}$ means that a constant in (\ref{equaintro}) depends only on $z$ and $T$ (see the Definition \ref{dynamicptprocess} for more details). These results on convergence to the Poisson distribution are stronger than the ones obtained previously \cite{ ptrf, nicol,pene,collet,haydndcds,haydncmp}. Namely \begin{enumerate} \item In \cite{ptrf,haydncmp,nicol} the following forms of convergence were obtained \[\lim_{r \to 0}\mathbb{P}\{N^{r,z,T}([0,T])=k\}=\mathbb{P}\{P([0,T]=k)\}\] and/or \[(N^{r,z,T}(I_1), \cdots , N^{r,z,T}(I_m)) \xrightarrow[r \to 0]{distribution} (P(I_1), \cdots ,P(I_m)),\] where $m, k \in \mathbb{N}$ and intervals $I_1, \cdots I_m \subseteq [0,T]$. Clearly, (\ref{equaintro}) implies these two forms. \item In \cite{ptrf,pene,nicol,haydncmp} only convergence to the Poisson law was considered, while the estimations of the convergence rates were not studied because the approaches used there did not allow for such estimates. \item In \cite{collet, haydndcds} the convergence rates were obtained in a weaker form. Namely, $ \forall r \in (0,1)$ there exist positive constants $a$, $b$ and a set $M_r \subseteq \mathcal{M}$ with $\mathbb{P}(M_r)\le r^b$ such that $\forall z \notin M_r$ \begin{equation}\label{weak} \sum_{k}|\mathbb{P}\{N^{r,z,T}([0,T])=k\}-\mathbb{P}\{P([0,T])=k\}|\precsim r^a. \end{equation} \end{enumerate} Besides, the following generalizations of the previous results are obtained \begin{enumerate} \item In \cite{collet, haydndcds} a relatively high regularity (at least bounded derivatives) of the dynamics was required, while we just need it to be a local $C^1$-diffeomorphism. Particularly, the derivatives can be unbounded. \item Unlike \cite{collet}, we do not assume that the unstable manifolds are one-dimensional. \item In \cite{ haydndcds, collet, pene,haydncmp} sufficiently fast decay rates of return times on hyperbolic towers were required. Our proofs of the existence of the Poisson limit laws use only contraction rates $\alpha$ of the (un)stable manifolds. Particularly, a simple, easy to verify, following criterion for the existence of the Poisson limit law is obtained \[\alpha> \frac{2}{\dim \gamma^u}-\frac{1}{\dim_H \mu},\] where $\dim \gamma^u$ is the dimension of unstable manifolds and $\dim_H \mu$ is the Hausdorff dimension of the SRB measure on the Gibbs-Markov-Young tower (see the details in the Theorems \ref{thm} and \ref{thm2}). This criterion allows to skip verification of the so called corona conditions (see the Definition \ref{defshortcoro} or \cite{pene}), which is usually rather cumbersome even for uniformly hyperbolic dynamical systems. Such verification becomes even more involved in case of non-uniformly hyperbolic systems. \end{enumerate} Also, some applications of our approach are considered to obtain the asymptotic functional Poisson laws and the corresponding convergence rates for Axiom A attractors, H\'enon attractors, intermittent solenoids and billiards, which improve various previously known results for these systems. The structure of the paper is the following one. In section \ref{def} we introduce notations, which are used throughout the paper, give the necessary definitions and formulate the main results. The section \ref{fpl} presents a proof of the functional Poisson law (with the error term) for systems admitting Young towers of general type. The section \ref{provethm1} contains a proof of the Theorem \ref{thm}. A proof of the Theorem \ref{thm2} is in the section \ref{provethm2}. In the section \ref{app} applications to Axiom A attractors, intermittent solenoids, billiards and Henon attractors are considered. \section{Definitions and Main Results }\label{def} We start by introducing some notations \begin{enumerate} \item $C_z$ denotes a constant depending on $z$. \item The notation $a\precsim_z b$ means that there is a constant $C_z>0$, depending on $z$, such that (s.t.) $a \le C_z \cdot b$. \item The relation $a=C_z^{\pm 1} \cdot b$ means that there is a constant $C_z\ge 1$ s.t. $C_z^{-1} \cdot b\le a \le C_z \cdot b$. \item The notation $\mathbb{P}$ refers to a probability distribution on the probability space, where a random variable lives, and $\mathbb{E}$ denotes the expectation of a random variable. \end{enumerate} \begin{definition}[Dynamical point processes]\label{dynamicptprocess}\ \par Let $(\mathcal{M},d)$ be a Riemannian manifold (with or without boundaries, connected or non-connected, compact or non-compact), $d$ is the Riemannian metric on $\mathcal{M}$ and $B_{r}(z)$ is a geodesic ball in $\mathcal{M}$ with radius $r$ and center $z \in \mathcal{M}$. We assume that dynamics $f: (\mathcal{M},\mu) \to (\mathcal{M}, \mu)$ is ergodic with respect to (w.r.t.) some invariant probability measure $\mu$. Let $T>0$. Consider a dynamical point process on $[0,T]$, so that for any $t \in [0,T]$ \[N^{r,T, z}_{t}:=\sum_{i=0}^{ \frac{t}{\mu(B_r(z))}} 1_{B_r(z)} \circ f^i.\] Thus the dynamical point process $N^{r,T, z}$ is a random counting measure on $[0,T]$. \end{definition} \begin{definition}[Poisson point processes]\label{poisson}\ \par For any $T>0$, we say that $P$ is a Poisson point process on $[0,T]$ if \begin{enumerate} \item $P$ is a random counting measure on $[0,T]$. \item $P(A)$ is a Poisson-distributed random variable for any Borel set $A \subseteq [0,T]$. \item If $A_1, A_2, \cdots, A_n \subseteq [0,T]$ are pairwise disjoint, then $P(A_1), \cdots, P(A_n)$ are independent. \item $\mathbb{E}P(A)=\Leb(A)$ for any Borel set $A\subseteq [0,T]$. \end{enumerate} \end{definition} \begin{definition}[Total variation norms of point processes]\label{totalnorm}\ \par For any $T>0$ consider the $\sigma$-algebra $\mathcal{C}$ on the space of point processes on $[0,T]$, defined as \begin{equation}\label{sigmaalg} \sigma \{\pi^{-1}_AB: A \subseteq [0,T], \end{equation} where $B \subseteq \mathbb{N}$ is a Borel set and $\pi_A$ is an evaluation map defined on the space of counting measures, so that for any counting measure $N$ \[\pi_AN:=N(A).\] Now we can define the total variation norm for the Poisson approximation of a dynamical point process as \[d_{TV}(N^{r,T,z},P):=\sup_{C\in \mathcal{C}}|\mu(N^{r,T,z} \in C)-\mathbb{P}(P \in C)|\] \end{definition} \begin{definition}[Convergence rates of the Poisson approximations]\label{convergencerate}\ \par Suppose that for any $T>0$ there exists a constant $a>0$ s.t. for almost every $z\in \mathcal{M}$ \[d_{TV}(N^{r,T,z},P) \precsim_{T,z} r^a \to 0.\] Then $a$ is called a rate of a Poisson approximation. \end{definition} We now turn to the definition of the Gibbs-Markov-Young structures \cite{Alves, Y, Y1}: \begin{definition}[Gibbs-Markov-Young structures]\label{gibbs}\ \par Introduce at first several notions concerning hyperbolic dynamics $f$ on the Riemannian manifolds $(\mathcal{M},d)$. \begin{enumerate} \item An embedded disk $\gamma^u$ is called an unstable manifold if for every $x,y \in \gamma^u$ \[\lim_{n \to \infty}d(f^{-n}(x), f^{-n}(y))=0\] \item An embedded disk $\gamma^s$ is called a stable manifold if for every $x,y \in \gamma^s$ \[\lim_{n \to \infty}d(f^{n}(x), f^{n}(y))=0\] \item $\Gamma^u:=\{\gamma^u\}$ is called a continuous family of $C^1$-unstable manifolds if there is a compact set $K^s$, a unit disk $D^u$ in some $\mathbb{R}^n$ and a map $\phi^u: K^s \times D^u \to \mathcal{M}$ such that \begin{enumerate} \item $\gamma^u=\phi^u(\{x\}\times D^u)$ is an unstable manifold, \item $\phi^u$ maps $K^s \times D^u$ homeomorphically onto its image, \item $x \to \phi^u|_{\{x\} \times D^u}$ defines a continuous map from $K^s$ to $Emb^1(D^u, \mathcal{M})$, where $Emb^1(D^u, \mathcal{M})$ is the space of $C^1$-embeddings of $D^u$ into $\mathcal{M}$. \end{enumerate} \end{enumerate} A continuous family of $C^1$-stable manifolds $\Gamma^s:=\{\gamma^s\}$ is defined similarly. \item We say that a compact set $ \Lambda \subseteq \mathcal{M}$ has a hyperbolic product structure if there exist continuous families of stable manifolds $\Gamma^s:=\{\gamma^s\}$ and of unstable manifolds $\Gamma^u:=\{\gamma^u\}$ such that \begin{enumerate} \item $\Lambda=(\cup \gamma^s) \cap (\cup \gamma^u)$, \item $\dim \gamma^s+\dim \gamma^u=\dim \mathcal{M}$, \item each $\gamma^s$ intersects each $\gamma^u$ at exactly one point, \item stable and unstable manifolds are transversal, and the angles between them are uniformly bounded away from 0. \end{enumerate} A subset $\Lambda_1 \subseteq \Lambda$ is called a s-subset if $\Lambda_1$ has a hyperbolic product structure and, moreover, the corresponding families of stable and unstable manifolds $\Gamma^s_1$ and $\Gamma^u_1$ can be chosen so that $\Gamma^s_1 \subseteq \Gamma^s$ and $\Gamma^u_1 = \Gamma^u$. Analogously, a subset $\Lambda_2 \subseteq \Lambda$ is called an $u$-subset if $\Lambda_2$ has a hyperbolic product structure and the families $\Gamma^s_2$ and $\Gamma^u_2$ can be chosen so that $\Gamma^u_2 \subseteq \Gamma^u$ and $\Gamma^s_2 = \Gamma^s$. \item For $x \in \Lambda$, denote by $\gamma^u(x)$ (resp. $\gamma^s(x)$) the element of $\Gamma^u$ (resp. $\Gamma^s$) which contains $x$. Also, for each $n \ge 1$, denote by $(f^n)^u$ the restriction of the map $f^n$ to $\gamma^u$-disks, and by $\det D(f^n)^u$ denote the Jacobian of $(f^n)^u$. We say that the set $\Lambda$ with hyperbolic product structure has also a \textbf{Gibbs-Markov-Young structure} if the following properties are satisfied \begin{enumerate} \item Lebesgue detectability: there exists $\gamma \in \Gamma^u$ such that $\Leb_{\gamma}(\Lambda \bigcap \gamma) > 0$. \item Markovian property: there exist pairwise disjoint $s$-subsets $\Lambda_1,\Lambda_2, \cdots \subseteq \Lambda$ such that \begin{enumerate} \item $\Leb_{\gamma}(\Lambda \setminus (\bigcup_{i \ge 1}\Lambda_i))=0$ on each $\gamma \in \Gamma^u$, \item for each $i \in \mathbb{N}$ there exists such $R_i \in \mathbb{N}$ that $f^{R_i} (\Lambda_i)$ is an $u$-subset, and for all $x \in \Lambda_i$ \[f^{R_i}(\gamma^s(x)) \subseteq \gamma^s(f^{R_i}(x))\] and \[f^{R_i}(\gamma^u(x)) \supseteq \gamma^u(f^{R_i}(x)).\] \end{enumerate} Define now a return time function $R : \Lambda \to \mathbb{N}$ and a return function $f^R: \Lambda \to \Lambda$, so that for each $i \in \mathbb{N}$ \[R|_{\Lambda_i}=R_i \text{ and } f^R|_{\Lambda_i}=f^{R_i}|_{\Lambda_i}\] The separation time $s(x, y)$ for $x, y \in \Lambda$ is defined as \[s(x,y):=\min \{n \ge 0: (f^R)^n(x) \text{ and } (f^R)^n(y) \text{belong to the different sets } \Lambda_i\}.\] We also assume that there are constants $C > 1, \alpha > 0$ and $0 < \beta < 1$, which depend only on $f$ and $\Lambda$, such that the following conditions hold \item Polynomial contraction on stable leaves: \[\forall \gamma^s \in \Gamma^s, \forall x,y \in \gamma^s, \forall n \in \mathbb{N}, d(f^n(x), f^n(y)) \le \frac{C}{n^{\alpha}}.\] \item Backward polynomial contraction on unstable leaves: \[\forall \gamma^u \in \Gamma^u, \forall x,y \in \gamma^u, \forall n \in \mathbb{N}, d(f^{-n}(x), f^{-n}(y)) \le \frac{C}{n^{\alpha}}.\] \item Bounded distortion: $\forall \gamma \in \Gamma^u \text{ and } x, y \in \gamma \bigcap \Lambda_i$ for some $\Lambda_i$, \[\log \frac{\det D(f^R)^u(x)}{\det D(f^R)^u(y)} \le C \cdot \beta^{s(f^R(x), f^R(y))}.\] \item Regularity of the stable foliations: for each $\gamma, \gamma'\in \Gamma^u$ denote \[\Theta_{\gamma, \gamma'}: \gamma' \bigcap \Lambda \to \gamma \bigcap \Lambda: x \to \gamma^s(x)\bigcap \gamma.\] Then the following properties hold \begin{enumerate} \item $\Theta_{\gamma, \gamma'}$ is absolutely continuous and $\forall x \in \gamma \bigcap \Lambda$ \[\frac{d(\Theta_{\gamma, \gamma'})_{*}\Leb_{\gamma'}}{d\Leb_{\gamma}}(x)=\prod_{n \ge 0} \frac{\det Df^u(f^n(x))}{\det Df^u(f^n(\Theta_{\gamma, \gamma'}^{-1}x))},\] \[ \frac{d(\Theta_{\gamma, \gamma'})_{*}\Leb_{\gamma'}}{d\Leb_{\gamma}}(x)= C^{\pm 1},\] \item for any $x,y \in \gamma \bigcap \Lambda$ \[\log \frac{\frac{d(\Theta_{\gamma, \gamma'})_{*}\Leb_{\gamma'}}{d\Leb_{\gamma}}(x)}{\frac{d(\Theta_{\gamma, \gamma'})_{*}\Leb_{\gamma'}}{d\Leb_{\gamma}}(y)} \le C \cdot \beta^{s(x,y)}.\] \end{enumerate} \item Aperiodicity: $\gcd(R_i, i \ge 1)= 1$. \item A decay rate of the return times $R$: there exist $\xi>1$ and $\gamma\in \Gamma^u$ such that \[\Leb_{\gamma}(R>n) \le \frac{C}{n^{\xi}}.\] \end{enumerate} \textbf{SRB measures}: Let the dynamics $f: (\mathcal{M}, \mu) \to (\mathcal{M}, \mu)$ has a Gibbs-Markov-Young structure. It was proved in \cite{Alves, Y, Y1} that there exists an ergodic probability measure $\mu$ such that for any unstable manifold $\gamma^u$ (including $\Gamma^u$) $\mu_{\gamma^u} \ll \Leb_{\gamma^u}$, where $\mu_{\gamma^u}$ is the conditional measure of $\mu$ on the unstable manifold $\gamma^u$. Such $\mu$ is called a Sinai-Ruelle-Bowen measure (SRB measure). \end{definition} \begin{assumption}[Geometric regularities]\label{geoassumption}\ \par Assume that $f: \mathcal{M}\to \mathcal{M}$ has the Gibbs-Markov-Young structure, as described in the Definition \ref{gibbs}, and \begin{enumerate} \item $f$ is bijective and a local $C^1$-diffeomorphism on $\bigcup_{i \ge 1} \bigcup_{0 \le j < R_i} f^j(\Lambda_i)$. \item the following limit exists \[ \dim_H\mu:=\lim_{r \to 0} \frac{\log \mu(B_r(z))}{\log r}\] for almost every $z\in \mathcal{M}$. Then $dim_H\mu$ is called a Hausdorff dimension of the measure $\mu$. \item $\alpha \cdot \dim_H\mu>1$, where $\alpha$ is the contraction rate of the (un)stable manifolds in the Definition \ref{gibbs}. \end{enumerate} \end{assumption} \begin{assumption}[The first returns \& interior assumptions on $\Lambda$]\label{assumption}\ \par Let $f: \mathcal{M}\to \mathcal{M}$ has the Gibbs-Markov-Young structure, and there are constants $C>1$ and $\beta\in (0,1)$ (the same as in the Definition \ref{gibbs}) such that \begin{enumerate} \item $R: \Lambda \to \mathbb{N}$ is the first return time and $f^R: \Lambda \to \Lambda$ is the first return map for $\Lambda$. This implies that $f^R$ is actually bijective (see the Lemma \ref{inducemapbi} below). \item $\forall \gamma \in \Gamma^s, \gamma_1 \in \Gamma^u, \forall x, y \in \gamma \bigcap \Lambda, \forall x_1, y_1 \in \gamma_1 \bigcap \Lambda $, \[d((f^R)^n(x), (f^R)^n(y)) \le C \cdot \beta^n,\] and \[d((f^R)^{-n}(x_1), (f^R)^{-n}(y_1)) \le C \cdot \beta^n \cdot d(x_1,y_1).\] \item $\mu\{\interior{(\Lambda)}\}>0$ and $\mu(\partial \Lambda)=0$, where \[\interior{\Lambda}:=\{x \in \Lambda: \exists r_x>0 \text{ s.t. } \mu(B_{r_x}(x)\setminus \Lambda)=0 \},\] \[\partial \Lambda:=\Lambda \setminus \interior{\Lambda}.\] In other words, $x \in \interior{\Lambda}$ if and only if $x \in \Lambda$ and there is a small ball $B_{r_x}(x)$ s.t. $B_{r_x}(x) \subseteq \Lambda$ $\mu$-almost surely. \end{enumerate} \end{assumption} Now we can formulate the first main result of the paper. \begin{theorem}[Convergence rates for functional Poisson laws \Romannum{1}]\label{thm}\ \par Assume that the dynamics $f: (\mathcal{M}, \mu) \to (\mathcal{M}, \mu)$ has a Gibbs-Markov-Young structure (see the Definition \ref{gibbs}) and satisfies the Assumptions \ref{geoassumption} and \ref{assumption}. Then for any $T>0$ the following results hold \begin{enumerate} \item $\dim_H\mu \ge \dim {\gamma}^u$. \item If $\alpha> \frac{2}{\dim \gamma^u}-\frac{1}{\dim_H \mu}$, then for almost every $z \in \mathcal{M}$ \[d_{TV}(N^{r,z,T}, P) \precsim_{T,\xi,z} r^{a},\] where the constant $a>0$ depends on $\xi>1$, $\dim_H \mu, \dim \gamma^u $ and $\alpha$, but it does not depend upon $z\in \mathcal{M}$. The expression for $a$ can be found in the Lemma \ref{lastproofs}. \item If $\mu \ll \Leb_{\mathcal{M}}$ and $\frac{d\mu}{d\Leb_{\mathcal{M}}} \in L^{\infty}_{loc}(\mathcal{M})$, then \[d_{TV}(N^{r,z,T}, P) \precsim_{T,\xi,z} r^{a},\] where $a>0$ depends on $\xi>1$, and $\dim_H \mu, \dim \gamma^u, \alpha$, but does not depend upon $z\in \mathcal{M}$. The expression for $a$ can be found in the Lemma \ref{lastproofs}. \end{enumerate} \end{theorem} \begin{definition}[Induced measurable partitions]\label{mpartition}\ \par We say a probability measure $\mu$ for the dynamics $f: \mathcal{M} \to \mathcal{M}$ has an induced measurable partition if there are constants $\beta \in (0,1), C>1$ (the same as in the Definition \ref{gibbs}) and $b>0$ such that \begin{enumerate} \item There exists an area $U \subseteq \mathcal{M}$ with $\mu \{\interior{(U)}\}>0$, $\mu(\partial U)=0$. \item The area $U$ has a measurable partition $\Theta:=\{\gamma^u(x)\}_{x \in U}$ (which could be different from $\Gamma^u$), such that the elements of $\Theta$ are disjoint connected unstable manifolds, so that $\mu$-almost surely $U=\bigsqcup_{x \in U} \gamma^u(x)$ and \[\mu_{U}(\cdot)=\int_U \mu_{\gamma^u(x)}(\cdot) d\mu_U(x),\] where $\mu_U:=\frac{\mu|_{U}}{\mu(U)}$ and $\mu_{\gamma^u(x)}$ is the conditional probability induced by $\mu$ on $\gamma^u(x)\in \Theta$. \item Each $\gamma^u\in \Theta$ is (at least $C^1$) smooth. \item All $\gamma^u\in \Theta$ have uniformly bounded sectional curvatures and the same dimensions. \item For any $\epsilon \in (0,1)$\[\mu_U \{x \in U: |\gamma^u(x)|< \epsilon\}\le C \cdot \epsilon^b,\] where $|\gamma^u(x)|$ is the radius of the largest inscribed geodesic ball in $\gamma^u(x) \in \Theta$, and a geodesic ball is defined with respect to the distance $d_{\gamma^u(x)}$ on $\gamma^u(x)$, induced by the Riemannian metric. This property implies that almost every $\gamma^u(x)\in \Theta$ is non-degenerated, i.e. $|\gamma^u(x)|>0$ for almost every $x\in U$. \item For almost every point $x\in U$ we have $\mu_{\gamma^u(x)} \ll \Leb_{\gamma^u(x)}$, $\mu_{\gamma^u(x)}(\gamma^u(x))>0$, and for any $y, z \in\gamma^u(x)$ \[\frac{d\mu_{\gamma^u(x)}}{d\Leb_{\gamma^u(x)}}(y)=C^{\pm 1} \cdot \frac{d\mu_{\gamma^u(x)}}{d\Leb_{\gamma^u(x)}}(z).\] \item Denote by $\overline{R}$ the first return time to $U$ for $f$. Then the first return map $f^{\overline{R}}: U \to U$ has an exponential u-contraction, i.e. $\forall \gamma^u \in \Theta, \forall x,y \in \gamma^u, n \ge 1$ \[d((f^{\overline{R}})^{-n}(x),(f^{\overline{R}})^{-n}(y)) \le C \cdot \beta ^n \cdot d(x,y),\] and an exponential decay of correlation, i.e. for any $h \in \mbox{\rm Lip}(U)$ \[\int h \circ (f^{\overline{R}})^n \cdot h d\mu_{U}-(\int h d\mu_U)^2= O(\beta^n) \cdot ||h||^2_{\mbox{\rm Lip}}. \] \end{enumerate} \end{definition} Now we are able to formulate the second main result of the paper. \begin{theorem}[Convergence rates for the functional Poisson laws \Romannum{2}]\label{thm2}\ \par Assume that the dynamics $f: (\mathcal{M}, \mu) \to (\mathcal{M}, \mu)$ has the Gibbs-Markov-Young structure (see the Definition \ref{gibbs}), satisfies the Assumption \ref{geoassumption} and $\mu$ has an induced measurable partition (see the Definition \ref{mpartition}). Then for any $T>0$, the following results hold. \begin{enumerate} \item $\dim_H \mu \ge \frac{ b}{b+\dim \gamma^u} \cdot \dim \gamma^u$. \item If $\alpha> \frac{2}{\dim \gamma^u}\cdot \frac{b+\dim \gamma^u}{b}-\frac{1}{\dim_H \mu} $, then for almost every (a.e.) $z \in \mathcal{M}$, \[d_{TV}(N^{r,z,T}, P) \precsim_{T, \xi, z} r^{a},\] where a constant $a>0$ depends on $\xi>1$, $\dim_H \mu, \dim \gamma^u, b$ and $\alpha$, but it does not depend on $z\in \mathcal{M}$. The expression for $a$ can be found in the Lemma \ref{lastlastproof}. \item If $\mu \ll \Leb_{\mathcal{M}}$ with $\frac{d\mu}{d\Leb_{\mathcal{M}}} \in L^{\infty}_{loc}(\mathcal{M})$, then \[d_{TV}(N^{r,z,T}, P) \precsim_{T, \xi,z} r^{a},\] where $a>0$ depends on $\xi>1, \dim_H \mu, \dim \gamma^u, \alpha$ and $b$, but it does not depend on $z\in \mathcal{M}$. The expression for $a$ can be found in the Lemma \ref{lastlastproof}. \end{enumerate} \end{theorem} \begin{remark}\label{remark}\ \par \begin{enumerate} \item The Assumption \ref{assumption} that $R$ is the first return time and $f^R$ is the first return map of $\Lambda$ is natural if the system has a Markov partition. Otherwise, we assume that the system has an area $U$ with an induced measurable partition (see the Definition \ref{mpartition}). It will be shown in what follows that the Theorems \ref{thm} and \ref{thm2} work efficiently for various systems in applications (see the Section \ref{app}). Clearly, a key issue is a choice of the reference sets $\Lambda$ and $U$. \item Under similar conditions to the Definition \ref{mpartition}, \cite{vaientizhang} proved that $\mu_{U}(\overline{R}>n)$ characterizes the optimal bound for the decay rates of correlations for sufficiently good observables supported on $U$ (see the Theorem 1.3 in \cite{vaientizhang}); \cite{melboune} uses operator renewal theory as a method to prove also sharp results on polynomial decay of correlations (see Theorem 3.1 in \cite{melboune}). For many purposes the aperiodicity in the Definition \ref{gibbs} is irrelevant provided the dynamic $f: (\mathcal{M}, \mu) \to (\mathcal{M}, \mu) $ is mixing (see the Remark 2.2 in \cite{melboune}). Indeed all dynamical systems, which we consider in applications (section \ref{app}), do have a Markov partition. Also an ergodic completely hyperbolic (all Lyapunov exponents do not vanish) dynamical system is mixing. Therefore Young towers are mixing. So, to simplify the argument of our proof, we only assume the aperiodicity in the Gibbs-Markov-Young structures. \item When dealing with applications, (see the section \ref{app}), it is always assumed that $\mu$ is a hyperbolic measure (i.e. the Lyapunov exponents do not vanish almost everywhere, see \cite{pesin}). Also, in applications most often there is an explicit natural invariant measure (sometimes called a physical measure). Therefore the Assumption \ref{geoassumption} that $\dim_H\mu:=\lim_{r \to 0} \frac{\log \mu(B_r(z))}{\log r}$ is relevant to such approach. (However, another dimension conditions, like e.g., \cite{haydncmp}, could be used as well). \item If an SRB measure $\mu$ is explicitly known, then the Poisson approximations are usually well understood \cite{demers, haydncmp, nomixing, dolgopyat, nicol}. However, if it is not the case, then often essential difficulties arise, e.g. for intermittent solenoid attractors, Axiom A attractors, etc (see \cite{pene}). The Theorem \ref{thm} provides an useful, easy to verify, criterion. Indeed, if $\alpha>\frac{2}{\dim \gamma^u}$, then there is no need to know $\dim_H \mu$. Moreover, estimations of the corresponding convergence rates can be obtained as well. \item According to the Theorems \ref{thm} and \ref{thm2}, it is only required that $\xi>1$. In fact, it is a minimal requirement for the existence of the SRB measures (see \cite{Alves}). Observe that for our approach only the contraction rate $O(\frac{1}{n^{\alpha}})$ along (un)stable manifolds matters, which is different from the ones employed in \cite{pene, haydncmp, collet}. \item If $f$ has a sufficiently good regularity, then $\dim_H \mu \ge \dim \gamma^u$ \cite{pesin,ledra1,ledra2}. Our only assumption is that $f$ is a local $C^1$-diffeomorphism. Observe that we do not even assume that $\mathcal{M}$ is a compact manifold (see the Definition \ref{gibbs} and the Assumption \ref{geoassumption}). Therefore the Theorem \ref{thm2} does not provide a good lower bound for $\dim_H \mu$. It is worthwhile to mention also that for all applications considered below (see the section \ref{app}) the relation $\dim_H\mu \ge \dim \gamma^u$ always holds. \end{enumerate} \end{remark} \begin{corollary}[The first hitting and survival probabilities]\label{cor}\ \par Under the same conditions as in the Theorem \ref{thm} or \ref{thm2} consider the first hitting moment of time $\tau_{B_r(z)}(x):=\inf\{n \ge 0: f^n(x) \in B_r(z)\}$. Then for almost every $z \in \mathcal{M}$, any $T>0$ and any $t\le T$ the following relation holds for the first hitting probability \begin{equation}\label{escape1} \mu(\tau_{B_r(z)}>\frac{t}{\mu(B_r(z))})-e^{-t}=O_{T,\xi, z}(r^a). \end{equation} Particularly, the survival probability at time $T$ can be approximated as \[\mu(\tau_{B_r(z)}>T)=e^{-T\cdot \mu(B_r(z))}+\min\{O_{T,\xi,z}(r^a),2\}.\] Moreover, the following limiting relations hold \begin{equation}\label{escape} \lim_{T \to \infty}\lim_{r \to 0} \frac{\log \mu(\tau_{B_r(z)}>T )}{-T\cdot \mu(B_r(z))}=1, \end{equation} and for any $ T>0$ \begin{equation}\label{escape2} \lim_{r\to 0} \frac{\log \mu(\tau_{B_r(z)}>\frac{T}{\mu(B_r(z))})}{-T}=1. \end{equation} \end{corollary} \begin{proof} Clearly $\mu(\tau_{B_r(z)}>\frac{t}{\mu(B_r(z))})=\mu(N^{r,T,z}[0,\frac{t}{\mu(B_r(z))}]=0)$. Apply now a relevant one of the Theorems \ref{thm} and \ref{thm2}. Then $O_{T,\xi,z}(r^a)$ is the error term with the convergence rate $a$. For the survival probability at time $T$ take $t=T \cdot \mu(B_r(z))$. The relation (\ref{escape1}) implies (\ref{escape2}). According to the Assumption \ref{geoassumption}, $f$ is a local diffeomorphism almost everywhere. Besides, the set of all periodic points has measure zero. Hence $\mu(\tau_{B_r(z)}>T)=1-\mu(\bigcup_{i\le T}f^{-i}B_r(z) )=1-(T+1) \cdot \mu(B_r(z))$, if $r$ is small enough. Therefore (\ref{escape}) holds. \end{proof} \section{Functional Poisson Limit Laws}\label{fpl} This section deals with the functional Poisson limit laws and with convergence rates of $d_{TV}(N^{r,z,T}, P)$ for the dynamics $f$ described in the Definition \ref{gibbs} and satisfying the Assumption \ref{geoassumption} only. For any $n \ge 0, I \subseteq[0,n]$, let \[X_i:=1_{B_r(z)} \circ f^i, \text{ } X_I:=\sum_{i \in I} 1_{B_r(z)} \circ f^i.\] Denote by $\{\hat{X}_i\}_{i\ge 0}$ the i.i.d. random variables, such that for each $i \ge 0$ \[X_i\overset{d}{=}\hat{X}_i.\] Let \[\hat{X_I}:=\sum_{i \in I}\hat{X}_i.\] Observe that generally $\hat{X_I}$ and ${X_I}$ are not identically distributed. \begin{lemma}\label{poissonappro} For any disjoint sets $I_1, I_2, \cdots, I_m \subseteq [0,n]$ and any integer $p\in (0,n)$, \[d_{TV}((X_{I_1}, \cdots, X_{I_m}),(\hat{X}_{I_1}, \cdots, \hat{X}_{I_m})) \] \[\le 2\cdot \sum_{0 \le l \le n-p}\sup_{h\in[0,1]} | \mathbb{E}[1_{X_0=1} \cdot h(X_p,\cdots, X_{n-l})]-\mathbb{E}1_{X_0=1} \cdot \mathbb{E}[h(X_p,\cdots, X_{n-l})]|\] \[+4(n-p)\cdot \mathbb{E} 1_{X_0=1} \cdot 1_{\sum_{1\le j \le p-1}X_j\ge 1}+4p \cdot (n-p) \cdot \mu(B_{r}(z))^2+4 p \cdot \mu(B_r(z)),\] where $h$ is a measurable function with values in $[0,1]$ . Observe that we obtain a slightly better error bound here, compared to the Theorem 2.1 in \cite{collet}. \end{lemma} \begin{proof} By the definition of the total variation norm \[d_{TV}((X_{I_1}, \cdots, X_{I_m}),(\hat{X}_{I_1}, \cdots, \hat{X}_{I_m})) =\sup_{h\in[0,1]} |\mathbb{E}h(X_{I_1}, \cdots, X_{I_m})-h(\hat{X}_{I_1}, \cdots, \hat{X}_{I_m})|\] \[\le \sup_{h\in[0,1]} |\mathbb{E}h(X_{1},\cdots, X_n)- h(\hat{X}_{1}, \cdots, \hat{X}_{n})|= d_{TV}((X_{1}, \cdots, X_{n}),(\hat{X}_{1}, \cdots, \hat{X}_{n})).\] Hence, it suffices to estimate \[d_{TV}((X_{1}, \cdots, X_{n}),(\hat{X}_{1}, \cdots, \hat{X}_{n}))=\sup_{h\in[0,1]} |\mathbb{E}h(X_{1},\cdots, X_{n})- h(\hat{X}_{1}, \cdots, \hat{X}_{n})| \] \[= \sup_{h\in[0,1]}|\sum_{0 \le l \le n} \mathbb{E} h(\hat{X}_1, \cdots, \hat{X}_{l-1},X_l, \cdots, X_n)- \mathbb{E} h(\hat{X}_1, \cdots, \hat{X}_{l-1},\hat{X}_l, \cdots, X_n)|\] \[\le \sup_{h\in[0,1]}|\sum_{0 \le l \le n} \mathbb{E} h_l(X_l,,X_{l+1}, \cdots,X_n)- \mathbb{E}h_l(\hat{X}_l, X_{l+1}, \cdots, X_n)|,\] here $h_l(\sbullet[.75]):=h(\hat{X}_1, \cdots,\hat{X}_{l-1}, \sbullet[.75])$. Since $\hat{X}_1, \cdots, \hat{X}_{l-1}$ are independent of other random variables, without loss of generality, $h_l$ can be regarded as a function which does not depend on $\hat{X}_1, \cdots, \hat{X}_{l-1}$. Note that $X_l \overset{d}{=} \hat{X}_l$ are $\{0,1\}$-valued random variables. Thus \[\mathbb{E} h_l(X_l,,X_{l+1}, \cdots,X_n)- \mathbb{E}h_l(\hat{X}_l, X_{l+1}, \cdots, X_n)|\] \[=|\mathbb{E}[ 1_{X_l=0} \cdot h_l(0,X_{l+1}, \cdots,X_n)+\mathbb{E}[ 1_{X_l=1} \cdot h_l(1, X_{l+1}, \cdots,X_n)]\] \[-\mathbb{E}1_{\hat{X}_l=0} \cdot \mathbb{E}h_l(0,X_{l+1},\cdots, X_n)- \mathbb{E}1_{\hat{X}_l=1} \cdot \mathbb{E}h_l(1,X_{l+1},\cdots, X_n)|\] \[=|\mathbb{E}[ 1_{X_l=1} \cdot h_l(0,X_{l+1}, \cdots,X_n)+\mathbb{E}[ 1_{X_l=1} \cdot h_l(1, X_{l+1}, \cdots,X_n)]\] \[-\mathbb{E}1_{\hat{X}_l=1} \cdot \mathbb{E}h_l(0,X_{l+1},\cdots, X_n)- \mathbb{E}1_{\hat{X}_l=1} \cdot \mathbb{E}h_l(1,X_{l+1},\cdots, X_n)|\] \[\le 2 \cdot \sup_{h\in[0,1]} |\mathbb{E}[ 1_{X_l=1} \cdot h(X_{l+1}, \cdots, X_n)]-\mathbb{E}1_{X_l=1} \cdot \mathbb{E}h(X_{l+1}, \cdots, X_n)|.\] Therefore, \[d_{TV}((X_{1}, \cdots, X_{n}),(\hat{X}_{1}, \cdots, \hat{X}_{n}))\] \begin{equation}\label{sum} \le 2 \sum_{0 \le l \le n} \sup_{h\in[0,1]} |\mathbb{E}[ 1_{X_l=1} \cdot h(X_{l+1}, \cdots, X_n)]-\mathbb{E}1_{X_l=1} \cdot \mathbb{E}h(X_{l+1}, \cdots, X_n)|. \end{equation} We will first estimate the terms with $l\le n-p$ in the sum (\ref{sum}). \[\mathbb{E}[ 1_{X_l=1} \cdot h(X_{l+1},\cdots, X_n)]-\mathbb{E}1_{X_l=1} \cdot \mathbb{E}h(X_{l+1}, \cdots, X_n)\] \[\le \mathbb{E}[ 1_{X_l=1} \cdot h(X_{l+1},\cdots, X_n)]-\mathbb{E}[ 1_{X_l=1} \cdot h(0,\cdots, 0,X_{l+p},\cdots,X_n)]\] \[+\mathbb{E}[ 1_{X_l=1} \cdot h(0,\cdots, 0,X_{l+p},\cdots,X_n)]-\mathbb{E}1_{X_l=1} \cdot \mathbb{E}h(X_{l+1}, \cdots, X_n)\] \[+\mathbb{E}1_{X_l=1} \cdot \mathbb{E}h(0,\cdots, 0,X_{l+p},\cdots,X_n)-\mathbb{E}1_{X_l=1} \cdot \mathbb{E}h(0,\cdots, 0,X_{l+p},\cdots,X_n)\] \[\le \mathbb{E}\big\{ 1_{X_l=1} \cdot [h(X_{l+1},\cdots, X_n)-h(0,\cdots, 0,X_{l+p},\cdots,X_n)]\big\}\] \[+\mathbb{E}1_{X_l=1} \cdot \mathbb{E}[h(0,\cdots, 0,X_{l+p},\cdots,X_n)-h(X_{l+1}, \cdots, X_n)]\] \[+\mathbb{E}[ 1_{X_l=1} \cdot h(0,\cdots, 0,X_{l+p},\cdots,X_n)]-\mathbb{E}1_{X_l=1} \cdot \mathbb{E}h(0,\cdots, 0,X_{l+p},\cdots,X_n)\] Observe that \[ |h(X_{l+1}, \cdots, X_n)- h(0, \cdots, 0,X_{l+p}, \cdots, X_n)| \le 2 \cdot 1_{\sum_{l+1\le j \le l+p-1}X_j\ge 1}.\] Now, because of the stationarity of $(X_i)_{i\ge 0}$, we can continue the estimates as \[\le | \mathbb{E}[1_{X_l=1} \cdot h(0, \cdots, 0,X_{l+p}, \cdots, X_n)]-\mathbb{E}1_{X_l=1} \cdot \mathbb{E}h(0, \cdots, 0,X_{l+p}, \cdots, X_n)|\] \[+2|\mathbb{E} 1_{X_l=1} \cdot 1_{\sum_{l+1\le j \le l+p-1}X_j\ge 1}|+2\mathbb{E}1_{X_l=1} \cdot \mathbb{E}1_{\sum_{l+1\le j \le l+p-1}X_j\ge 1}\] \[\le | \mathbb{E}[1_{X_l=1} \cdot h(0, \cdots, 0,X_{l+p}, \cdots, X_n)]-\mathbb{E}1_{X_l=1} \cdot \mathbb{E}h(0, \cdots, 0,X_{l+p}, \cdots, X_n)|\] \[+2|\mathbb{E} 1_{X_0=1} \cdot 1_{\sum_{1\le j \le p-1}X_j\ge 1}|+2\mathbb{E}1_{X_0=1} \cdot \mathbb{E}1_{\sum_{1\le j \le p-1}X_j\ge 1}.\] Note that $1_{\sum_{1\le j \le p-1}X_j\ge 1}=1_{\cup_{1 \le j \le p-1} f^{-j} B_{r}(z)}$. Hence, we can continue the sequence of the inequalities above as \[\le | \mathbb{E}[1_{X_l=1} \cdot h(0, \cdots, 0,X_{l+p}, \cdots, X_n)]-\mathbb{E}1_{X_l=1} \cdot \mathbb{E}h(0, \cdots, 0,X_{l+p}, \cdots, X_n)|\] \[+2|\mathbb{E} 1_{X_0=1} \cdot 1_{\sum_{1\le j \le p-1}X_j\ge 1}|+2 (p-1)\mu(B_{r}(z))^2.\] Therefore for the terms with $l\le n-p$ in the sum (\ref{sum}) we have \[\mathbb{E}[ 1_{X_l=1} \cdot h(X_{l+1}, \cdots, X_n)]-\mathbb{E}1_{X_l=1} \cdot \mathbb{E}h(X_{l+1}, \cdots, X_n)\] \[\le | \mathbb{E}[1_{X_l=1} \cdot h(0, \cdots, 0,X_{l+p}, \cdots, X_n)]-\mathbb{E}1_{X_l=1} \cdot \mathbb{E}h(0, \cdots, 0,X_{l+p}, \cdots, X_n)|\] \[+2\mathbb{E} 1_{X_0=1} \cdot 1_{\sum_{1\le j \le p-1}X_j\ge 1}+2p \cdot \mu(B_{r}(z))^2.\] Consider now the terms with $l> n-p$ in the sum (\ref{sum}). Since $||h||_{\infty} \le 1$, then \[\mathbb{E}[ 1_{X_l=1} \cdot h(X_{l+1},\cdots, X_n)]-\mathbb{E}1_{X_l=1} \cdot \mathbb{E}h(X_{l+1},\cdots, X_n) \le 2 \mu(B_r(z)).\] Therefore \[(\ref{sum})=2 \cdot \sum_{0 \le l \le n} \sup_{h\in[0,1]}|\mathbb{E}[ 1_{X_l=1} \cdot h(X_{l+1},\cdots, X_n)]-\mathbb{E}1_{X_l=1} \cdot \mathbb{E}h(X_{l+1},\cdots,X_n)|\] \[\le 2\cdot \sum_{0 \le l \le n-p} \sup_{h\in[0,1]} |\mathbb{E}[1_{X_l=1} \cdot h(X_{l+p},\cdots, X_n)]-\mathbb{E}1_{X_l=1} \cdot \mathbb{E}[h(X_{l+p},\cdots, X_n)]|\] \[+4(n-p)\cdot \mathbb{E} 1_{X_0=1} \cdot 1_{\sum_{1\le j \le p-1}X_j\ge 1}+4p \cdot (n-p) \cdot \mu(B_{r}(z))^2+4 p \cdot \mu(B_r(z)).\] Again, by making use of the stationarity of $(X_i)_{i \ge 0}$, the last expression above can be estimated as \[\le 2\cdot \sum_{0 \le l \le n-p}\sup_{h\in[0,1]} | \mathbb{E}[1_{X_0=1} \cdot h(X_p,\cdots, X_{n-l})]-\mathbb{E}1_{X_0=1} \cdot \mathbb{E}[h(X_p,\cdots, X_{n-l})]|\] \[+4(n-p)\cdot \mathbb{E} 1_{X_0=1} \cdot 1_{\sum_{1\le j \le p-1}X_j\ge 1}+4p \cdot (n-p) \cdot \mu(B_{r}(z))^2+4 p \cdot \mu(B_r(z)).\] \end{proof} For further estimates we will need the following lemma. \begin{lemma}[Hyperbolic towers, see \cite{pene}, page 2609]\label{towerresult}\ \par Define a tower $\Delta$ and a map $F : \Delta \to \Delta$ as \[\Delta:=\{(x,l) \in \Lambda \times \mathbb{N}: 0 \le l < R(x) \},\] \[F(x,l):=\begin{cases} (x,l+1), &l < R(x)-1\\ (f^R(x),0), & l=R(x)-1\\ \end{cases}.\] The equivalence relation $\sim$ on $\Lambda$ is then \[x \sim y \text{ if and only if } x,y \in \gamma^s \text{ for some } \gamma^s \in \Gamma^s.\] Now we can define a quotient tower $\widetilde{\Delta}:=\Delta/\sim$, a quotient Gibbs-Markov-Young product structure $\widetilde{\Lambda}:=\Lambda/\sim$, quotient maps $\widetilde{F} : \widetilde{\Delta} \to \widetilde{\Delta}, \widetilde{f^R}: \widetilde{\Lambda} \to \widetilde{\Lambda}$, and canonical projections $\widetilde{\pi}_{\Delta}:\Delta \to \widetilde{\Delta}$ and $\widetilde{\pi}_{\Lambda}:\Lambda \to \widetilde{\Lambda}$. At first, we introduce a family of partitions $(\mathcal{Q}_k)_{k \ge 0}$ of $\Delta$ as \[\mathcal{Q}_0:=\{\Lambda_i \times \{l\}, i \ge 1, l < R_i\}, \mathcal{Q}_k:= \bigvee_{0 \le i \le k} F^{-i} \mathcal{Q}_0.\] Next, a projection $\pi: \Delta \to M $ is defined as \[\pi(x,l):=f^l(x).\] Then there exists a constant $C>1$ (the same as in the Definition \ref{gibbs}) such that for any $Q \in \mathcal{Q}_{2k}$ \begin{equation}\label{coronadiam} \diam (\pi \circ F^k(Q)) \le \frac{C}{k^{\alpha}}. \end{equation} There exist also probability measures $\mu_{\Delta}, \mu_{\Lambda}$ on $\Delta$ and $\Lambda$, respectively, such that \begin{equation}\label{allmeasure} \pi_{*}\mu_{\Delta}=\mu, \text{ } F_{*}\mu_{\Delta}=\mu_{\Delta}, \text{ } f_{*}\mu=\mu, \text{ } (f^R)_{*}\mu_{\Lambda}=\mu_{\Lambda}. \end{equation} Further, there exist probability measures $\mu_{\widetilde{\Delta}}, \mu_{\widetilde{\Lambda}}$ on $\widetilde{\Delta}$ and $\widetilde{\Lambda}$ respectively, such that \begin{equation}\label{allquotientmeasure} (\widetilde{\pi}_{\Delta})_{*}\mu_{\Delta}=\mu_{\widetilde{\Delta}}, \text{ } (\widetilde{\pi}_{\Lambda})_{*}\mu_{\Lambda}=\mu_{\widetilde{\Lambda}}, \text{ } \widetilde{F}_{*}\mu_{\widetilde{\Delta}}=\mu_{\widetilde{\Delta}}, \text{ } (\widetilde{f^R})_{*}\mu_{\widetilde{\Lambda}}=\mu_{\widetilde{\Lambda}}. \end{equation} Thus $\mu$ is supported on $\bigcup_{i \ge 1} \bigcup_{j < R_i} f^j(\Lambda_i)$, i.e. \[\mu(\bigcup_{i \ge 1} \bigcup_{j < R_i} f^j(\Lambda_i))=1.\] Moreover \begin{equation}\label{horsesrbleb} (\mu_{\Lambda})_{\gamma^u} \ll \Leb_{\gamma^u}, \text{ } \frac{d(\mu_{\Lambda})_{\gamma^u}}{d\Leb_{\gamma^u}}= C^{\pm 1}, \end{equation} where $(\mu_{\Lambda})_{\gamma^u}$ is the conditional measure of $\mu_{\Lambda}$ on $\gamma^u \in \Gamma^u$. Since $R$ is the first return time, (see the Assumption \ref{assumption}), then \[\mu_{\Lambda}=\frac{\mu|_{\Lambda}}{\mu (\Lambda)}.\] Finally, for any $k\ge 1$ and any $(Q_i)_{i \ge 1} \subseteq \mathcal{Q}_k$, any $h: \Delta \to \mathbb{R}$ satisfying $||h||_{\infty}\le 1$ and $h(x,l)=h(y,l)$ for any $x,y \in \gamma^s \in \Gamma^s$ , and any allowable $l \in \mathbb{N}$, (i.e. $h$ is $\sigma(\bigcup_{k\ge 0}\mathcal{Q}_k)$-measurable), we have the following estimate for decay of correlations \begin{equation}\label{decorrelation} |\int 1_{\bigcup_{i \ge 1} Q_i} \cdot h \circ F^{2k} d\mu_{\Delta}-\mu_{\Delta}(\bigcup_{i \ge 1} Q_i) \cdot \int h d\mu_{\Delta}| \le \frac{C}{k^{\xi-1}} \cdot \mu_{\Delta}(\bigcup_{i \ge 1} Q_i). \end{equation} \end{lemma} \begin{lemma}\label{decaylemma1} For any $l\ge 0, \frac{1}{p^{\alpha}}\ll r$ and any measurable function $h$ with values in $[0, 1]$ \[| \mathbb{E}[1_{X_0=1} \cdot h(X_p,\cdots, X_{p+l})]-\mathbb{E}1_{X_0=1} \cdot \mathbb{E}h(X_p,\cdots, X_{p+l})|\] \[\le 4\cdot \frac{C}{p^{\xi-1}} \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z))+[2 +4l \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z))] \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z)), \] where a constant $C$ is the same as in the Definition \ref{gibbs}. \end{lemma} \begin{proof} Let $m:=\lfloor \frac{p}{4} \rfloor$. By (\ref{allmeasure}) and invariance of $F$ (i.e. $F_{*}\mu_{\Delta}=\mu_{\Delta}$) we have \[ \mathbb{E}[1_{X_0=1} \cdot h(X_p,\cdots, X_{p+l})]=\int 1_{B_r(z)} \cdot h(1_{B_r(z)} \circ f^p,\cdots, 1_{B_r(z)} \circ f^{p+l} ) d\mu\] \[=\int 1_{B_r(z)} \circ \pi \circ F^{m}\cdot h(1_{B_r(z)} \circ \pi \circ F^{p+m-p},\cdots, 1_{B_r(z)} \circ \pi \circ F^{p+l+m-p} )\circ F^p d\mu_{\Delta}.\] Denote $A_1:=F^{-m} \pi^{-1} B_r(z), A_0:=\bigsqcup_{Q \in \mathcal{Q}_{2m}:Q \bigcap A_1\neq \emptyset}Q$ and $A_2:=\bigsqcup_{Q \in \mathcal{Q}_{2m}:Q \bigcap (A_0\setminus A_1)\neq \emptyset}Q$. Then $A_1 \bigcup A_2=A_0$. The sets $A_0,A_2$ are $\sigma(\bigcup_{k\ge 0}\mathcal{Q}_k)$-measurable. Therefore we can continue the equality above as \[=\int 1_{A_1} \cdot h(1_{A_1},\cdots, 1_{A_1}\circ F^{l}) \circ F^p d\mu_{\Delta}=\int 1_{A_1} \cdot h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l}) \circ F^p d\mu_{\Delta}\] \[+\int 1_{A_1} \cdot h(1_{A_1},\cdots, 1_{A_1}\circ F^{l}) \circ F^p d\mu_{\Delta}-\int 1_{A_1} \cdot h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l}) \circ F^p d\mu_{\Delta}.\] {\bf Claim}: $| h(1_{A_1},\cdots, 1_{A_1}\circ F^{l})- h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l})|\le 2\cdot 1_{\cup_{j \le l}F^{-j}A_2}$. Indeed, if $F^j(x,l) \notin A_2 $ for all $j\le l$, then $1_{A_1}\circ F^{j-p}(x,l)=1_{A_1\cup A_2}\circ F^{j-p}(x,l)$. On the other hand, $||h_j||_{\infty} \le 1$. Hence, the claim holds. Therefore, \[| \mathbb{E}[1_{X_0=1} \cdot h(X_p,\cdots, X_{p+l})]-\mathbb{E}1_{X_0=1} \cdot \mathbb{E}h(X_p,\cdots, X_{p+l})|\] \[=|\int 1_{A_1} \cdot h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l}) \circ F^p d\mu_{\Delta}- \int 1_{A_1} d\mu_{\Delta} \cdot \int h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l}) \circ F^p d\mu_{\Delta}\] \[+\int 1_{A_1} \cdot h(1_{A_1},\cdots, 1_{A_1}\circ F^{l}) \circ F^p d\mu_{\Delta}-\int 1_{A_1} \cdot h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l}) \circ F^p d\mu_{\Delta}\] \[-\int 1_{A_1} d\mu_{\Delta} \cdot \int h(1_{A_1},\cdots, 1_{A_1}\circ F^{l}) \circ F^p d\mu_{\Delta}+\int 1_{A_1} d\mu_{\Delta} \cdot \int h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l}) \circ F^p d\mu_{\Delta}|\] \[\le |\int 1_{A_1} \cdot h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l}) \circ F^p d\mu_{\Delta}- \int 1_{A_1} d\mu_{\Delta} \cdot \int h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l}) \circ F^p d\mu_{\Delta}|\] \[+2\int 1_{A_1} \cdot 1_{\bigcup_{j \le l}F^{-j}A_2} \circ F^p d\mu_{\Delta}+ 2\int 1_{A_1} d\mu_{\Delta} \cdot \int 1_{\bigcup_{j \le l}F^{-j}A_2}\circ F^p d\mu_{\Delta}\] \[\le |\int 1_{A_0} \cdot h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l}) \circ F^p d\mu_{\Delta}- \int 1_{A_0} d\mu_{\Delta} \cdot \int h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l}) \circ F^p d\mu_{\Delta}|\] \[+ |\int 1_{A_0\setminus A_1} \cdot h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l})\circ F^p d\mu_{\Delta}- \int 1_{A_0\setminus A_1} d\mu_{\Delta} \cdot \int h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l})\circ F^p d\mu_{\Delta}|\] \[+2\int 1_{A_1} \cdot 1_{\bigcup_{j \le l}F^{-j}A_2}\circ F^p d\mu_{\Delta}+ 2\int 1_{A_1} d\mu_{\Delta} \cdot \int 1_{\bigcup_{j \le l}F^{-j}A_2} \circ F^p d\mu_{\Delta}.\] Note that $A_0\setminus A_1 \subseteq A_2, A_1 \subseteq A_0$, which means that we can continue the estimate above as \[\le |\int 1_{A_0} \cdot h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l}) \circ F^p d\mu_{\Delta}- \int 1_{A_0} d\mu_{\Delta} \cdot \int h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l})\circ F^p d\mu_{\Delta}|\] \[+ \int 1_{A_2} \cdot h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l}) \circ F^p d\mu_{\Delta}+ \int 1_{A_2} d\mu_{\Delta} \cdot \int h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l})\circ F^p d\mu_{\Delta}\] \[+2\int 1_{A_0} \cdot 1_{\bigcup_{j \le l}F^{-j}A_2} \circ F^p d\mu_{\Delta}+ 2\int 1_{A_1} d\mu_{\Delta} \cdot \int 1_{\bigcup_{j \le l}F^{-j}A_2} d\mu_{\Delta}\] \[\le |\int 1_{A_0} \cdot h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l}) \circ F^p d\mu_{\Delta}- \int 1_{A_0} d\mu_{\Delta} \cdot \int h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l})\circ F^p d\mu_{\Delta}|\] \[+ |\int 1_{A_2} \cdot h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l}) \circ F^p d\mu_{\Delta}- \int 1_{A_2} d\mu_{\Delta} \cdot \int h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l}) \circ F^p d\mu_{\Delta}|\] \[+2\int 1_{A_2} d\mu_{\Delta} \cdot \int h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l})\circ F^p d\mu_{\Delta}\] \[+2|\int 1_{A_0} \cdot 1_{\bigcup_{j \le l}F^{-j}A_2} \circ F^p d\mu_{\Delta}-\int 1_{A_0}d\mu_{\Delta} \cdot \int 1_{\bigcup_{j \le l}F^{-j}A_2} \circ F^p d\mu_{\Delta}|\] \begin{equation}\label{1} +2\int 1_{A_0}d\mu_{\Delta} \cdot \int 1_{\bigcup_{j \le l}F^{-j}A_2} \circ F^p d\mu_{\Delta}+ 2\int 1_{A_1} d\mu_{\Delta} \cdot \int 1_{\cup_{j \le l}F^{-j}A_2} d\mu_{\Delta}. \end{equation} {\bf Claim}: $h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l})$ is $\sigma(\bigcup_{k\ge 0}\mathcal{Q}_k)$-measurable. Observe, that for any $(x,l), (y,l) \in \Delta$, $x,y\in \gamma^s \in \Gamma^s$ we have $F^{j-p}(x,l)=(x',l'), F^{j-p}(y,l)=(y',l')$ for some $l'\in \mathbb{N}$ and some $x',y'\in (\gamma^s)'\in \Gamma^s$. Since $1_{A_1 \bigcup A_2}$ is $\sigma(\bigcup_{k\ge 0}\mathcal{Q}_k)$-measurable, $1_{A_1\cup A_2}\circ F^{j-p}(x,l)=1_{A_1\cup A_2}\circ F^{j-p}(y,l)$. Therefore $h(1_{A_1\bigcup A_2},\cdots,1_{A_1\bigcup A_2}\circ F^{l})$ is $\sigma(\cup_{k\ge 0}\mathcal{Q}_k)$-measurable. {\bf Claim:} $1_{\bigcup_{j \le l}F^{-j}A_2} $ is also $\sigma(\bigcup_{k\ge 0}\mathcal{Q}_k)$-measurable. Indeed, each set $F^{-j}A_2$ is $\sigma(\bigcup_{k\ge 0}\mathcal{Q}_k)$-measurable. So their union is also $\sigma(\bigcup_{k\ge 0}\mathcal{Q}_k)$-measurable. {\bf Claim:} $\mu_{\Delta}(A_2)\le \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z))$. Observe that \[\mu_{\Delta}(A_2)\le \mu_{\Delta}(F^{-m}\pi^{-1}\pi( F^m A_2))=\mu(\pi( F^m A_2)).\] By the definition of $A_2:=\bigsqcup_{Q \in \mathcal{Q}_{2m}:Q \cap (A_0\setminus A_1)\neq \emptyset}Q$, for each $Q$, contained in $A_2$, there exist $x_1, x_2 \in Q $, such that $\pi(F^mx_1)\in B_r(z), \pi(F^mx_2)\notin B_r(z)$. Now, by making use of (\ref{coronadiam}) and taking $m=\lfloor \frac{p}{4} \rfloor$, we obtain $\pi(F^m A_2) \subseteq B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z)$. Hence the claim holds. Having these claims and (\ref{decorrelation}), we can continue the estimate of (\ref{1}) as \[\le \frac{C}{p^{\xi-1}} \cdot \mu_{\Delta}(A_0) +\frac{C}{p^{\xi-1}} \cdot \mu_{\Delta}(A_2)+2\mu_{\Delta}(A_2) +2 \cdot \frac{C}{p^{\xi-1}} \cdot \mu_{\Delta}(A_0)\] \[+2\mu_{\Delta}(A_0) \cdot \mu_{\Delta}(\bigcup_{j \le l}F^{-j}A_2)+ 2 \mu_{\Delta}(A_1) \cdot \mu_{\Delta}(\bigcup_{j \le l}F^{-j}A_2)\] \[\le \frac{C}{p^{\xi-1}} \cdot \mu_{\Delta}(A_1)+\frac{C}{p^{\xi-1}} \cdot \mu_{\Delta}(A_2) +\frac{C}{p^{\xi-1}} \cdot \mu_{\Delta}(A_2)+2\mu_{\Delta}(A_2) +2 \cdot \frac{C}{p^{\xi-1}} \cdot \mu_{\Delta}(A_1)\] \begin{equation}\label{10} +2 \cdot \frac{C}{p^{\xi-1}} \cdot \mu_{\Delta}(A_2)+2[\mu_{\Delta}(A_1)+\mu_{\Delta}(A_2)] \cdot \mu_{\Delta}(\bigcup_{j \le l}F^{-j}A_2)+ 2 \mu_{\Delta}(A_1) \cdot \mu_{\Delta}(\bigcup_{j \le l}F^{-j}A_2). \end{equation} Recall that $\mu(B_r(z))=\mu_{\Delta}(A_1)$ and $\mu_{\Delta}(A_2)\le \mu(B_{r+\frac{C}{p^{\alpha}}}(z)\setminus B_{r-\frac{C}{p^{\alpha}}}(z))$. Therefore the estimate of (\ref{10}) can be continued as \[\le \frac{C}{p^{\xi-1}} \cdot \mu(B_r(z))+2\cdot \frac{C}{p^{\xi-1}} \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z)) +2\mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z)) \] \[+2 \cdot \frac{C}{p^{\xi-1}} \cdot \mu(B_r(z))+2 \cdot \frac{C}{p^{\xi-1}} \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z))\] \[+2[\mu(B_r(z))+\mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z))] \cdot l \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z))\] \[+ 2 \mu(B_r(z)) \cdot l \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z))\] \[\le 4\cdot \frac{C}{p^{\xi-1}} \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z))+2\mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z)) +4l \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)) \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z)).\] \end{proof} \begin{proposition}[Functional Poisson laws]\ \label{rate} Let $p:=\lfloor {\lfloor \frac{T}{\mu(B_r(z))}\rfloor}^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu}}\rfloor$, where $\epsilon$ is so small that \[\alpha \cdot (\dim_H\mu-\epsilon)>\frac{\dim_H\mu}{\dim_H\mu-\epsilon}>1 \text{ (see the Assumption } \ref{geoassumption}).\] Then for almost any $z \in \mathcal{M}$ there is $r_z>0$, such that $\forall r< r_z$ \[d_{TV}(N^{r,T,z},P) \precsim_{T, \xi, \epsilon}\] \[r^{\dim_H\mu-\epsilon}+r^{\frac{(\dim_H\mu-\epsilon)^2}{\dim_H\mu} \cdot (\xi-1)}+ r^{ \frac{\epsilon \cdot(\dim_H\mu-\epsilon)}{\dim_H\mu}}+\frac{\mu(B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z))}{\mu(B_r(z))}\] \[+[\frac{\mu(B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}}}(z))}{\mu(B_r(z))}]^2+\frac{1}{\mu(B_r(z))}\cdot \int_{B_r(z)} 1_{\bigcup_{1\le j \le p}f^{-j}B_r(z)}d\mu,\] where $C'$ depends on $\alpha, \epsilon $ and on the constant $C$ in the Definition \ref{gibbs}. \end{proposition} \begin{proof} Let $n:=\lfloor \frac{T}{\mu(B_r(z))}\rfloor, p:=\lfloor n^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu}}\rfloor$, where $\epsilon$ is so small that \[\alpha \cdot (\dim_H\mu-\epsilon)>\frac{\dim_H\mu}{\dim_H\mu-\epsilon}>1 \text{ (in view of the Assumption } \ref{geoassumption}).\] Therefore for a.e. $z \in \mathcal{M}$ there exists $r_z>0$ such that $\forall r < r_z$, in view of the Assumption \ref{geoassumption}, \[\frac{T}{r^{\dim_H\mu-\epsilon}} \precsim n \precsim \frac{T}{r^{\dim_H\mu+\epsilon}}\] \[\frac{1}{p^{\alpha}} \precsim_{T, \alpha} \frac{1}{n^{\frac{\alpha \cdot (\dim_H\mu-\epsilon)}{\dim_H\mu}}}\precsim_{T, \alpha} r^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu} \cdot \alpha \cdot (\dim_H\mu-\epsilon)} \ll r. \] Hence by the Lemmas \ref{poissonappro} and \ref{decaylemma1}, for any disjoint sets $I_1, \cdots, I_m \subseteq [0,n]$ we have \[d_{TV}((X_{I_1}, \cdots, X_{I_m}),(\hat{X}_{I_1}, \cdots, \hat{X}_{I_m}))\] \[\le 2\cdot \sum_{0 \le l \le n-p}\sup_{h\in[0,1]} | \mathbb{E}[1_{X_0=1} \cdot h(X_p,\cdots, X_{n-l})]-\mathbb{E}1_{X_0=1} \cdot \mathbb{E}[h(X_p,\cdots, X_{n-l})]|\] \[+4(n-p)\cdot \mathbb{E} 1_{X_0=0} \cdot 1_{\sum_{1\le j \le p-1}X_j\ge 1}+4p \cdot (n-p) \cdot \mu(B_{r}(z))^2+4 p \cdot \mu(B_r(z))\] \[\le \sum_{0 \le l \le n-p} 8\cdot \frac{C}{p^{\xi-1}} \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z))+[4 +8(n-l-p) \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z))] \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z)) \] \[+4(n-p)\cdot \mathbb{E} 1_{X_0=0} \cdot 1_{\sum_{1\le j \le p-1}X_j\ge 1}+4p \cdot (n-p) \cdot \mu(B_{r}(z))^2+4 p \cdot \mu(B_r(z))\] \[\le 8\cdot n\cdot \frac{C}{p^{\xi-1}} \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z))+[4 +8n \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z))] \cdot n \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z)) \] \[+4n\cdot \mathbb{E} 1_{X_0=0} \cdot 1_{\sum_{1\le j \le p-1}X_j\ge 1}+4p \cdot n \cdot \mu(B_{r}(z))^2+4 p \cdot \mu(B_r(z))\] \[\le 8\cdot n\cdot \frac{C}{p^{\xi-1}} \cdot \mu(B_{r}(z))+ 8\cdot n\cdot \frac{C}{p^{\xi-1}} \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z))\] \[+[4 +8n \cdot \mu(B_{r}(z))+8n \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z))] \cdot n \cdot \mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z)) \] \[+4n\cdot \mathbb{E} 1_{X_0=0} \cdot 1_{\sum_{1\le j \le p}X_j\ge 1}+4p \cdot n \cdot \mu(B_{r}(z))^2+4 p \cdot \mu(B_r(z)).\] Note that \[p \approx_{T,\epsilon} [\frac{1}{\mu(B_r(z))}]^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu}}, \text{ } n \cdot \mu(B_r(z))\le T.\] Thus we can continue the inequality above as \[\precsim_{T, \xi, \epsilon} \mu(B_r(z))^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu} \cdot (\xi-1)}+ \frac{\mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z))}{\mu(B_r(z))}+[\frac{\mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z))}{\mu(B_r(z))}]^2\] \[+\frac{1}{\mu(B_r(z))}\cdot \int_{B_r(z)} 1_{\bigcup_{1\le j \le p}f^{-j}B_r(z)}d\mu+\mu(B_r(z))^{\frac{\epsilon}{\dim_H\mu}}.\] By applying the Theorem 2 of \cite{chenmethod} to $(\hat{X}_{I_1}, \cdots, \hat{X}_{I_m})$ one gets \[d_{TV}((\hat{X}_{I_1}, \cdots, \hat{X}_{I_m}), (P(I_1), \cdots, P(I_m)))\le 4 \cdot n \cdot \mu(B_r(z))^2 \precsim_T \mu(B_r(z)). \] Approximate now $(X_{I_1}, \cdots, X_{I_m})$ by the Poisson point process $P$. Then \[d_{TV}((X_{I_1}, \cdots, X_{I_m}),(P(I_1), \cdots, P(I_m)))\] \[\le d_{TV}((\hat{X}_{I_1}, \cdots, \hat{X}_{I_m}), (P(I_1), \cdots, P(I_m)))+d_{TV}((X_{I_1}, \cdots, X_{I_m}),(\hat{X}_{I_1}, \cdots, \hat{X}_{I_m}))\] \[\precsim_{T, \xi, \epsilon}\mu(B_r(z))+ \mu(B_r(z))^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu} \cdot (\xi-1)}+ \frac{\mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z))}{\mu(B_r(z))}\] \[+[\frac{\mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z))}{\mu(B_r(z))}]^2+\frac{1}{\mu(B_r(z))}\cdot \int_{B_r(z)} 1_{\bigcup_{1\le j \le p}f^{-j}B_r(z)}d\mu+\mu(B_r(z))^{\frac{\epsilon}{\dim_H\mu}}.\] Since $\mathcal{C}':=\{\pi_{I'_1}^{-1}A_1 \cap \cdots \cap \pi_{I'_m}^{-1}A_m: \forall m \ge 1, A_i \subseteq \mathbb{N}, \forall \text{ disjoint sets } I'_1, \cdots, I'_m \subseteq [0,T]\}$ generates $\sigma$-algebra $\mathcal{C}=\sigma \{\pi^{-1}_AB: \text{ any Borel sets } A \subseteq [0,T], B \subseteq \mathbb{N}\}$, we obtain $\forall r< r_z$ the following functional Poisson approximation \[d_{TV}(N^{r,T,z},P)=\sup_{\text{disjoint }I_i \subseteq [0,n]}d_{TV}((X_{I_1}, \cdots, X_{I_m}), (P(I_1), \cdots, P(I_m)))\] \[\precsim_{T, \xi, \epsilon}\mu(B_r(z))+ \mu(B_r(z))^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu} \cdot (\xi-1)}+\mu(B_r(z))^{\frac{\epsilon}{\dim_H\mu}} +\frac{\mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z))}{\mu(B_r(z))}\] \[+[\frac{\mu(B_{r+\frac{4^{\alpha}C}{p^{\alpha}}}(z)\setminus B_{r-\frac{4^{\alpha}C}{p^{\alpha}}}(z))}{\mu(B_r(z))}]^2+\frac{1}{\mu(B_r(z))}\cdot \int_{B_r(z)} 1_{\bigcup_{1\le j \le p}f^{-j}B_r(z)}d\mu\] \[\precsim_{T, \xi, \epsilon}r^{\dim_H\mu-\epsilon}+r^{\frac{(\dim_H\mu-\epsilon)^2}{\dim_H\mu} \cdot (\xi-1)}+ +r^{ \frac{\epsilon \cdot(\dim_H\mu-\epsilon)}{\dim_H\mu}}+\frac{\mu(B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z))}{\mu(B_r(z))}\] \[+[\frac{\mu(B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}}}(z))}{\mu(B_r(z))}]^2+\frac{1}{\mu(B_r(z))}\cdot \int_{B_r(z)} 1_{\bigcup_{1\le j \le p}f^{-j}B_r(z)}d\mu,\] where $C'$ depends upon $\alpha, \epsilon $ and the constant $C$ in the Definition \ref{gibbs}. \end{proof} \begin{definition}[Short returns and Coronas]\label{defshortcoro}\ \par Let $p$ be the one in the Proposition \ref{rate}. Define \begin{enumerate} \item {\bf Short returns}: \[\int_{B_r(z)} 1_{\bigcup_{1\le j \le p}f^{-j}B_r(z)}d\mu.\] \item {\bf Coronas}: \[\mu(B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)).\] \end{enumerate} It will be shown below that these quantities tend to $0$ for almost all $z \in \mathcal{M}$ with certain convergence rates. \end{definition} \section{Proof of the Theorem \ref{thm}}\label{provethm1} \subsection{Properties of the First Return to $\Lambda$} Before studying convergence rates for short returns and coronas we will prove several lemmas for the first return time $R$ and for the first return map $f^R: \Lambda \to \Lambda$ under the Assumptions \ref{geoassumption} and \ref{assumption}. \begin{lemma}\label{inducemapbi} The map $f^R: \Lambda \to \Lambda$ is bijective. \end{lemma} \begin{proof} We show first that $f^R$ is {\bf one-to-one} for any $x,y \in \Lambda$, and $f^R(x)=f^R(y)$. If $x, y \in \Lambda_i $ for some $i$, then $f^{R_i}(x)=f^{R_i}(y)$. On the other hand, it follows from the Assumption \ref{geoassumption} that $f$ is bijective on $\cup_{i \ge 1} \cup_{j < R_i} f^j(\Lambda_i)$. Thus we have inductively the following reduction \[f(f^{R_i-1}x)=f(f^{R_i-1}y) \Rightarrow f(f^{R_i-2}x)=f(f^{R_i-2}y)\Rightarrow \cdots \Rightarrow f(x)=f(y) \Rightarrow x=y. \] Let $x \in \Lambda_i, y \in \Lambda_j$ for some $i \neq j$ and $f^R(x)=f^R(y)$. Without any loss of generality, we may assume that $R_i<R_j$. Then $f^{R_i}(x)=f^{R_j}(y)$. Again, by the Assumptions \ref{geoassumption} \[f(f^{R_i-1}x)=f(f^{R_j-1}y) \Rightarrow f(f^{R_i-2}x)=f(f^{R_j-2}y)\Rightarrow \cdots \Rightarrow x=f^{R_j-R_i}y \in \Lambda. \] But the first return time of $y$ to $\Lambda$ is $R_j$, i.e., $f^{R_j-R_i}y \notin \Lambda$. So we came to a contradiction, and therefore this case can not occur. Show now that $f^R$ is {\bf onto} map. Let $y \in \Lambda$ and $y\in \Lambda_i$ for some $i$. Then $y \in \cup_{i \ge 1} \cup_{j < R_i} f^j(\Lambda_i)$. By the Assumption \ref{geoassumption} $f$ is bijective on $\cup_{i \ge 1} \cup_{j < R_i} f^j(\Lambda_i)$. Therefore there exists $x' \in \cup_{i \ge 1} \cup_{j < R_i} f^j(\Lambda_i)$, i.e. there is $x\in \Lambda_k$ such that $f^j(x)=x'$, where $j < R_k$ and $f(x')=f^{j+1}(x)=y$. Since $R_k$ is the first return time for $x$, then $j+1=R_k$ and $f^R(x)=y$. \end{proof} \begin{lemma}\label{pibi} The following maps satisfy the properties that \[\pi: \Delta \to \cup_{i \ge 1} \cup_{ j < R_i} f^j(\Lambda_i) \text{ is bijective,}\] \[\pi: \Delta_0 \to \Lambda \text{ is identity},\] \[\pi:\Delta_{\ge 1} \to \bigcup_{i \ge 1}\bigcup_{ 1 \le j < R_i} f^j(\Lambda_i) \text{ is bijective},\] \[\text{where } \Delta_{\ge 1}:=\{(x,l) \in \Lambda \times \mathbb{N}: 1 \le l < R(x)\},\] \[\Delta_0:=\{(x,0): x \in \Lambda \}\subseteq \Delta.\] \end{lemma} \begin{proof} Clearly, it is enough to prove just the first statement. By the definition of $\Delta$ the first map $\pi$ is onto. Let us show now that it is actually one-to-one map. For all $ (x,l), (x',l')\in \Delta $ with $\pi(x,l)=\pi(x',l')$ it holds that $f^l(x)=f^{l'}(x')$. Without loss of generality, let $l \le l'$. By the Assumption \ref{geoassumption} $f$ is bijective on $\cup_{i \ge 1} \cup_{j < R_i} f^j(\Lambda_i)$. Then \[f(f^{l-1}x)=f(f^{l'-1}y) \Rightarrow f(f^{l-2}x)=f(f^{l'-2}y)\Rightarrow \cdots \Rightarrow x=f^{l'-l}y. \] Since $x,y \in \Lambda$ and $l'-l$ is less than the first return time of $y$, one gets that $l'=l$ and $x=y$. \end{proof} \par By the Birkhoff's Ergodic Theorem, for almost every $z \in \bigcup_{i \ge 1} \bigcup_{0 \le j < R_i} f^j(\Lambda_i)$ we have $z=f^{j_z}(z')$ for some $z' \in \interior{(\Lambda)}$ and $j_z \in \mathbb{N}$. Recall that $\mu\{\interior{(\Lambda)}\}>0$. \begin{lemma}[Pulling metric balls back to $\Lambda$]\ \label{ballinhorseshoe}\ \par There exists a small enough $r$ such that \[\mu(f^{-j_z}B_{r}(z) \bigcap \Lambda^c)=0,\] \[\mu_{\Delta}(\{\pi^{-1}f^{-j_z}B_r(z)\} \bigcap \Delta_{\ge 1})=0,\] and \[\mu_{\Delta}(\{\pi^{-1}f^{-j_z}B_r(z)\} \bigcap \Delta_{0})=1.\] \end{lemma} \begin{proof} By the Assumption \ref{assumption} there is a small neighborhood $U_{z'}\subseteq \mathcal{M}$ of $z'\in \interior{(\Lambda)}$ such that \[\mu(U_{z'}\bigcap \Lambda^c)=\mu(U_{z'} \bigcap \{\bigcup_{i \ge 1} \bigcup_{1\le j < R_i} f^j(\Lambda_i)\} )=0.\] Because $f^{j_z}$ is a local $C^1$-diffeomorphism, there exists a small ball $B_r(z)$ such that $f^{-j_z} B_r(z) \subseteq U_{z'}$. So $\mu(f^{-j_z}B_{r}(z) \bigcap \Lambda^c)=\mu(f^{-j_z}B_{r}(z) \bigcap \{\bigcup_{i \ge 1} \cup_{1 \le j < R_i} f^j(\Lambda_i)\})=0$. Hence by Lemma \ref{pibi}, \[\mu_{\Delta}(\{\pi^{-1}f^{-j_z}B_r(z)\}\bigcap \Delta_{\ge 1})=0 \text{ and } \mu_{\Delta}(\{\pi^{-1}f^{-j_z}B_r(z)\} \bigcap \Delta_{0})=1.\] \end{proof} \begin{definition}[Topological balls]\ \par We say that a set $TB_r(z') \subseteq \mathcal{M}$ is a topological ball if there is a ball $B_r(z) \subseteq \mathcal{M}$ and a map $T$ of $B_r(z)$, such that \[T:B_r(z) \to TB_r(z') \text{ is a } C^1 \text{-diffeomorphism and } T(z)=z'.\] Denote by $r, z'$ the radius and the center of $TB_r(z')$. \end{definition} For almost every $z \in \bigcup_{i \ge 1} \bigcup_{0 \le j < R_i} f^{j_z}(\Lambda_i)$ we have $z=f^j(z')$, where $z' \in \interior{(\Lambda)}$ and $j_z \in \mathbb{N}$. Since $f^{j_z}$ is a local diffeomorphism (by the Assumption \ref{geoassumption}), the set $TB_r(f^{-j_z}z):=f^{-j_z}B_r(z)$ is a topological ball for sufficiently small $r>0$. \begin{lemma}[Comparisons of topological and metric balls]\ \label{comparetopballmetricball}\ \par There exist constants $C_z\ge 1$ and $ r_z>0$, such that $\forall r < r_z$ \[B_{C^{-1}_z\cdot r} (f^{-j_z}z) \subseteq TB_r(f^{-j_z}z) \subseteq B_{C_z\cdot r} (f^{-j_z}z).\] \end{lemma} \begin{proof} Since $f^{-j_z}$ is a local diffeomorphism near $z$, then $f^{-j_z}(\partial B_r(z))=\partial TB_r(f^{-j_z}z)$. We will estimate $\sup_{x \in \partial TB_r(f^{-j_z}z)}d(x, f^{-j_z}z)$ and $\inf_{x \in \partial TB_r(f^{-j_z}z)}d(x, f^{-j_z}z)$. For any $x\in \partial TB_r(f^{-j_z}z)$ one has $f^{j_z}(x) \in \partial B_r(z)$. Let $(\gamma_t)_{0 \le t \le 1}$ be the geodesic connecting $x$ and $f^{-j_z}z$, and a curve $\hat{\gamma}:=f^{j_z}\gamma$ is connecting $f^{j_z}x$ and $z$. Then \begin{equation}\label{6} d(x, f^{-j_z}z)=\int_0^1 \sqrt{\langle \gamma_t',\gamma_t'\rangle_{\gamma_t}}dt=\int_0^1 \sqrt{\langle Df^{-j_z}\hat{\gamma}_t',Df^{-j_z}\hat{\gamma}_t'\rangle_{\gamma_t}}dt. \end{equation} If $r$ is sufficiently small (i.e. $r< r_z$ for some $r_z>0$), then $Df^{-j_z}$ and the Riemannian metric $\langle \cdot, \cdot \rangle_{\gamma_t}$ are close to $Df^{-j_z}(z)$ and $\langle \cdot, \cdot \rangle_{z}$, respectively. Then there exists $C_z\ge 1$ such that \[d(x, f^{-j_z}z) \ge C^{-1}_z \cdot \int_0^1 \sqrt{\langle\hat{\gamma}_t',\hat{\gamma}_t'\rangle_{\hat{\gamma}_t}}dt \ge C^{-1}_z \cdot d(f^{j_z}x,z)=C^{-1}_z \cdot r.\] Similarly, let $(\gamma_t)_{0 \le t \le 1}$ be a curve connecting $x$ and $f^{-j_z}z$, such that $\hat{\gamma}:=f^{j_z}\gamma$ is a geodesic connecting $f^{j_z}x$ and $z$. Then \[d(x, f^{-j_z}z)\le \int_0^1 \sqrt{\langle\gamma_t',\gamma_t'\rangle_{\gamma_t}}dt=\int_0^1 \sqrt{\langle Df^{-j_z}\hat{\gamma}_t',Df^{-j_z}\hat{\gamma}_t'\rangle_{\gamma_t}}dt \] \[\le C_z \cdot \int_0^1 \sqrt{\langle\hat{\gamma}_t',\hat{\gamma}_t'\rangle_{\hat{\gamma}_t}}dt = C_z \cdot d(f^{j_z}x,z)=C_z \cdot r,\] which proves the lemma. \end{proof} \begin{definition}[Two-sided cylinders in $\Lambda$]\ \par Since $f^R: \Lambda \to \Lambda $ is bijective, we can define two-sided cylinders as \[\xi_{i_{-n}\cdots i_{0} \cdots i_{n}}:=(f^R)^n\Lambda_{i_{-n}} \bigcap (f^R)^{n-1}\Lambda_{i_{-(n-1)}} \bigcap \cdots \bigcap \Lambda_{i_0} \bigcap \cdots \bigcap (f^R)^{-(n-1)}\Lambda_{i_{n-1}} \bigcap (f^R)^{-n}\Lambda_{i_{n}}. \] Let us introduce now a new partition of $\Lambda$ as \[\mathcal{M}_0:=\{\Lambda_i, i \ge 1\}, \mathcal{M}_k:= \bigvee_{0 \le i \le k} (f^R)^{-i} \mathcal{M}_0,\] and a quotient partition of $\widetilde{\Lambda}$ \[\widetilde{\mathcal{M}}_0:=\{\widetilde{\Lambda}_i, i \ge 1\}, \widetilde{\mathcal{M}}_k:= \bigvee_{0 \le i \le k} (\widetilde{f^R})^{-i} \widetilde{\mathcal{M}}_0,\] \[\widetilde{\mathcal{M}}_{\infty}:= \bigvee^{\infty}_{i=0} (\widetilde{f^R})^{-i} \widetilde{\mathcal{M}}_0 \text{ is the } \sigma \text{-algebra of }\widetilde{\Lambda}.\] \end{definition} \begin{lemma}[Diameters of two-sided cylinders]\ \label{diamcylinder}\ \par \[\diam \xi_{i_{-n}\cdots i_{0} \cdots i_{n}} \le 2C \cdot \beta^n,\] where $C\ge 1$ and $\beta\in (0,1)$ are the same as in the Assumption \ref{assumption}. \end{lemma} \begin{proof} Observe that $\xi_{i_{-n}\cdots i_{0} \cdots i_{n}}=\xi_{i_{-n}\cdots i_{0}} \bigcap \xi_{ i_{0} \cdots i_{n}}$ is an intersection of two one-sided cylinders $\xi_{i_{-n}\cdots i_{0}}$ and $\xi_{ i_{0} \cdots i_{n}}$, where $\xi_{i_{-n}\cdots i_{0}}$ is long and $\xi_{ i_{0} \cdots i_{n}}$ is slim. We will now estimate their widths. By the Assumption \ref{assumption} $(f^R)^{-1}$ contracts exponentially along any $\gamma^u \in \Gamma^u$. Hence $\diam \gamma^u \bigcap \xi_{ i_{0} \cdots i_{n}} \le C \cdot \beta^n$. Similarly, since \[\xi_{ i_{-n} \cdots i_{0}}=(f^R)^n\Lambda_{i_{-n}} \bigcap (f^R)^{n-1}\Lambda_{i_{-(n-1)}} \bigcap \cdots \bigcap \Lambda_{i_0}\] \[=(f^R)^n[\Lambda_{i_{-n}} \bigcap (f^R)^{-1}\Lambda_{i_{-(n-1)}} \bigcap \cdots \bigcap (f^R)^{-n}\Lambda_{i_0}]\] and $f^R$ contracts exponentially along any $\gamma^s \in \Gamma^s$, we have $\diam \gamma^s \bigcap \xi_{ i_{-n} \cdots i_{0}} \le C \cdot \beta^n$. The properties of the Gibbs-Markov-Young structures ensure that $\forall x,y \in \xi_{i_{-n}\cdots i_{0} \cdots i_{n}}$ there is $z \in \gamma^u(x)$, $z\in \gamma^s(y)$, such that \[d(x,z) \le C \cdot \beta^n, \text{ } d(y,z) \le C \cdot \beta^n.\] Hence $d(x,y) \le 2C\cdot \beta^n$, i.e. $\diam \xi_{i_{-n}\cdots i_{0} \cdots i_{n}} \le 2C \cdot \beta^n$. \end{proof} \begin{lemma}[Decay of correlations for $f^R: \Lambda \to \Lambda$]\label{decorrelationcylinder}\ \par There exist constants $C''>1, \beta_1 \in (0,1)$, such that for any one-sided cylinder $\xi_{ i_{0} \cdots i_{n}} \in \mathcal{M}_n $ and any $A \in \sigma(\cup_{k\ge 0}\mathcal{M}_k)$ \begin{equation} |\int 1_{\xi_{ i_{0} \cdots i_{n}}} \cdot 1_A \circ (f^R)^{2n}d\mu_{\Lambda}-\int 1_{\xi_{ i_{0} \cdots i_{n}}} d\mu_{\Lambda} \cdot \int 1_A d\mu_{\Lambda}| \le C'' \cdot \beta_1^n \cdot \mu_{\Lambda}(\xi_{ i_{0} \cdots i_{n}}). \end{equation} \end{lemma} \begin{proof} At first, we prove that \[|\int 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}} \cdot 1_{\widetilde{A}} \circ (\widetilde{f^R})^{2n}d\mu_{\widetilde{\Lambda}}-\int 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}} d\mu_{\widetilde{\Lambda}} \cdot \int 1_{\widetilde{A}} d\mu_{\widetilde{\Lambda}}| \le C'' \cdot \beta_1^n \cdot \mu_{\widetilde{\Lambda}}(\widetilde{\xi_{ i_{0} \cdots i_{n}}}),\] where $\widetilde{\xi_{ i_{0} \cdots i_{n}}} \in \widetilde{\mathcal{M}_n}, \widetilde{A} \in \sigma(\cup_{k\ge 0}\widetilde{\mathcal{M}}_k)$. Denote the transfer operator of $\widetilde{f^R}$ by $\widetilde{P}$. {\bf Claim:} \[||\widetilde{P}^{n+1} 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}||_{\infty} \precsim \mu_{\widetilde{\Lambda}}(\widetilde{\xi_{ i_{0} \cdots i_{n}}}),\] where a constant in $\precsim$ depends on the constant $C$ in the Definition \ref{gibbs}. Indeed, $\forall \widetilde{x} \in \widetilde{\Lambda}$ we have $(\widetilde{f^R})^{n+1}(\xi_{ i_{0} \cdots i_{n}})=\widetilde{\Lambda}$, and \[\widetilde{P}^{n+1} 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}(\widetilde{x})=\sum_{(\widetilde{f^R})^{n+1}(\widetilde{y})=\widetilde{x}} \frac{1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}(\widetilde{y})}{|\det D(\widetilde{f^R})^{n+1}(\widetilde{y})|}=\frac{1}{|\det D(\widetilde{f^R})^{n+1}[(\widetilde{f^R})^{n+1}|_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}^{-1}\widetilde{x}]|}.\] By making use of the distortion estimate of $\widetilde{f^R}$ (see the Lemma 3.2 in \cite{Alves} or the Lemma 1 in \cite{Y}) and (\ref{horsesrbleb}) we get \[\widetilde{P}^{n+1} 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}(\widetilde{x})=\frac{1}{|\det D(\widetilde{f^R})^{n+1}[(\widetilde{f^R})^{n+1}|_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}^{-1}\widetilde{x}]|} \precsim \mu_{\widetilde{\Lambda}}(\widetilde{\xi_{ i_{0} \cdots i_{n}}}).\] Therefore $||\widetilde{P}^{n+1} 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}||_{\infty} \precsim \mu_{\widetilde{\Lambda}}(\widetilde{\xi_{ i_{0} \cdots i_{n}}})$. {\bf Claim:} \[|\widetilde{P}^{n+1} 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}(\widetilde{x}_1)-\widetilde{P}^{n+1} 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}(\widetilde{x}_2)| \precsim \cdot \beta^{\widetilde{s}(\widetilde{x}_1,\widetilde{x}_2)} \cdot \mu_{\widetilde{\Lambda}}(\widetilde{\xi_{ i_{0} \cdots i_{n}}}),\] where the constant in $\precsim$ depends on the constant $C$ in the Definition \ref{gibbs}, $\beta \in (0,1)$ is the same as in the Definition \ref{gibbs}, $\widetilde{x}_1, \widetilde{x_2} \in \widetilde{\Lambda}_i$ for some $\widetilde{\Lambda}_i \in \widetilde{\mathcal{M}}_0$ and \[\widetilde{s}(\widetilde{x}_1,\widetilde{x}_2):=\inf \{n: (\widetilde{f^R})^n(\widetilde{x}_1), (\widetilde{f^R})^n(\widetilde{x}_2) \text{ lie in different elements of } \widetilde{\mathcal{M}}_0\}.\] Again, by using the distortion of $\widetilde{f^R}$, we get \[|\widetilde{P}^{n+1} 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}(\widetilde{x}_1)-\widetilde{P}^{n+1} 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}(\widetilde{x}_2)|\] \[=|\frac{1}{|\det D(\widetilde{f^R})^{n+1}[(\widetilde{f^R})^{n+1}|_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}^{-1}\widetilde{x}_1]|}-\frac{1}{|\det D(\widetilde{f^R})^{n+1}[(\widetilde{f^R})^{n+1}|_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}^{-1}\widetilde{x}_2]|}|\] \[=\frac{1}{|\det D(\widetilde{f^R})^{n+1}[(\widetilde{f^R})^{n+1}|_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}^{-1}\widetilde{x}_1]|} \cdot |1-\frac{|\det D(\widetilde{f^R})^{n+1}[(\widetilde{f^R})^{n+1})^{n+1}|_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}^{-1}\widetilde{x}_1]|}{|\det D(\widetilde{f^R})^{n+1}[(\widetilde{f^R})^{n+1}|_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}^{-1}\widetilde{x}_2]|}|\] \[\precsim \beta^{\widetilde{s}(\widetilde{x}_1,\widetilde{x}_2)} \cdot \mu_{\widetilde{\Lambda}}(\widetilde{\xi_{ i_{0} \cdots i_{n}}}).\] So the claim holds. \textbf{Claim}: $\widetilde{f^R}: (\widetilde{\Lambda}, \mu_{\widetilde{\Lambda}}) \to (\widetilde{\Lambda}, \mu_{\widetilde{\Lambda}})$ is exact, so mixing. Take a set $A\in \bigcap_{i \ge 0}(\widetilde{f^R})^{-i}\widetilde{\mathcal{M}}_{\infty}$. So $\mu_{\widetilde{\Lambda}}((\widetilde{f^R})^{i}A)=\mu_{\widetilde{\Lambda}}(A), \forall i \ge 0$. If $\mu_{\widetilde{\Lambda}}(A)>0$, then there is $B_k \in \widetilde{\mathcal{M}}_k$ for a large $k \gg 1$ s.t. $(\widetilde{f^R})^k: B_k \to \widetilde{\Lambda}$ is bijective and \[\frac{\mu_{\widetilde{\Lambda}}(B_k \bigcap A) }{\mu_{\widetilde{\Lambda}}(B_k)}\approx 1.\] By the distortion of $\widetilde{f^R}$, we have \[\mu_{\widetilde{\Lambda}}(A)=\frac{\mu_{\widetilde{\Lambda}}\{(\widetilde{f^R})^k(A)\}}{\mu_{\widetilde{\Lambda}}(\widetilde{\Lambda})}\ge \frac{\mu_{\widetilde{\Lambda}}\{(\widetilde{f^R})^k(B_k \bigcap A)\} }{\mu_{\widetilde{\Lambda}}\{\widetilde{f^R}(B_k)\}}\approx\frac{\mu_{\widetilde{\Lambda}}(B_k \bigcap A) }{\mu_{\widetilde{\Lambda}}(B_k)}\approx 1.\] So this claim holds. From the these three claims it follows that $\widetilde{P}^{n+1} 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}$ is a locally Lipschitz function on $\widetilde{\Lambda}$ with a Lipschitz constant $\mu_{\widetilde{\Lambda}}(\widetilde{\xi_{ i_{0} \cdots i_{n}}})$ and $\widetilde{f^R}: (\widetilde{\Lambda}, \mu_{\widetilde{\Lambda}}) \to (\widetilde{\Lambda}, \mu_{\widetilde{\Lambda}})$ is mixing. By applying the Corollary 2.3 (b) of \cite{asipmelbourne} to $\widetilde{P}^{n+1} 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}}$ we obtain that there exist constants $C''>1, \beta_1 \in (0,1)$, such that \[|\int (\widetilde{P}^{n+1} 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}} )\cdot 1_{\widetilde{A}} \circ (\widetilde{f^R})^{n}d\mu_{\widetilde{\Lambda}}-\int \widetilde{P}^{n+1} 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}} d\mu_{\widetilde{\Lambda}} \cdot \int 1_{\widetilde{A}} d\mu_{\widetilde{\Lambda}}| \le C'' \cdot \beta_1^n \cdot \mu_{\widetilde{\Lambda}}(\widetilde{\xi_{ i_{0} \cdots i_{n}}}).\] Therefore \[|\int 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}} \cdot 1_{\widetilde{A}} \circ (\widetilde{f^R})^{2n}d\mu_{\widetilde{\Lambda}}-\int 1_{\widetilde{\xi_{ i_{0} \cdots i_{n}}}} d\mu_{\widetilde{\Lambda}} \cdot \int 1_{\widetilde{A}} d\mu_{\widetilde{\Lambda}}| \le C'' \cdot \beta_1^n \cdot \mu_{\widetilde{\Lambda}}(\widetilde{\xi_{ i_{0} \cdots i_{n}}}).\] Now, by (\ref{allquotientmeasure}) $(\widetilde{\pi}_{\Lambda})_{*}\mu_{\Lambda}=\mu_{\widetilde{\Lambda}}$, we have \[|\int 1_{\xi_{ i_{0} \cdots i_{n}}} \cdot 1_A \circ (f^R)^{2n}d\mu_{\Lambda}-\int 1_{\xi_{ i_{0} \cdots i_{n}}} d\mu_{\Lambda} \cdot \int 1_A d\mu_{\Lambda}| \le C'' \cdot \beta_1^n \cdot \mu_{\Lambda}(\xi_{ i_{0} \cdots i_{n}}).\] \end{proof} \subsection{Short Returns} Recall that $n:=\lfloor \frac{T}{\mu(B_r(z))}\rfloor, p:=\lfloor n^{\frac{\text{dim}_H\mu-\epsilon}{\text{dim}_H\mu}}\rfloor$ (see the Proposition \ref{rate}). We now consider short returns to $\Lambda$. A reason to do this is that the short returns problem on $\mathcal{M}$ can be turned into the short returns problem on $\Lambda$ (see the Lemma \ref{pullbackhorseshoe}). For any $z' \in \interior{(\Lambda)}$, a fixed positive integer $M>0$, sufficiently small constants $\epsilon'>0, r>0$, such that $ n^{\epsilon'}\ll p$, $B_{M\cdot r} (z') \subseteq \interior{(\Lambda)}$ almost surely and \begin{equation}\label{sumhorshose} \int_{B_{M\cdot r} (z')} 1_{\bigcup_{1\le k \le p}(f^R)^{-k}B_{M\cdot r} (z')}d\mu_{\Lambda}=\int_{B_{M\cdot r} (z')} 1_{\bigcup_{1\le k \le N}(f^R)^{-k}B_{ M\cdot r} (z')}d\mu_{\Lambda} \end{equation} \[+\int_{B_{M\cdot r} (z')} 1_{\bigcup_{N\le k \le n^{\epsilon'}}(f^R)^{-k}B_{ M\cdot r} (z')}d\mu_{\Lambda}+\int_{B_{ M\cdot r} (z')} 1_{\bigcup_{n^{\epsilon'}\le k \le p}(f^R)^{-k}B_{ M\cdot r} (z')}d\mu_{\Lambda},\] where $N= \lfloor\frac{-\log 2C }{\log \beta}\rfloor+1$, and the constants $C$ and $\beta$ are defined in the Assumption \ref{assumption}. We begin with the short fixed-length returns described by the integral \[\int_{B_{M\cdot r} (z')} 1_{\bigcup_{1\le k \le N}(f^R)^{-k}B_{ M\cdot r} (z')}d\mu_{\Lambda}.\] \begin{lemma}[Short fixed-length returns]\label{fixreturn}\ \par For almost every $z' \in \interior{(\Lambda)}$ and sufficiently small $r_{N,M, z'}>0$ we have for any $ r < r_{N,M,z'}$ \[\int_{B_{M\cdot r} (z')} 1_{\bigcup_{1\le k \le N}(f^R)^{-k}B_{ M\cdot r} (z')}d\mu_{\Lambda}=0.\] (Actually, a stronger result will be proved, i.e. $\forall k\in \mathbb{N}$ a map $(f^{{R}})^k$ is a local diffeomorphism at almost every point $z' \in \interior{(\Lambda)}$). \end{lemma} \begin{proof} From the Assumption \ref{assumption} $\mu(\partial \Lambda)=0$, and then $\mu(\bigcup_{i \in \mathbb{Z}} f^{-i}\partial \Lambda)=0$. By making use of the decay of correlations for the quotient map $\widetilde{f^R}: \widetilde{\Lambda} \to \widetilde{\Lambda}$, we have that $(\widetilde{f^R})^i$ are ergodic for all $i \ge 1$. Therefore the set of periodic points $A_{per}$ of $\widetilde{f^R}$ has $\mu_{\widetilde{\Lambda}}(A_{per})=0$, and $\mu_{\Lambda}(\widetilde{\pi}_{\Lambda}^{-1}(A_{per}))=0$ due to (\ref{allquotientmeasure}). Choose now $z' \in \widetilde{\pi}_{\Lambda}^{-1}(A_{per}^c)\bigcap \interior{(\Lambda)} \bigcap [\bigcup_{i \in \mathbb{Z}} f^{-i}\partial \Lambda]^c$. Then there is $r_M>0$ s.t. for any $r<r_M$ almost surely $B_{M \cdot r} \subseteq \interior{(\Lambda)}$. We will make now several claims. \textbf{Claim:} For any $k\in \mathbb{N}$ let $R^k$ be the k-th return time. Then ${R}^k|_{B_{M \cdot r}(z')}={R}^k(z'), \forall k\in[1, N]$ if $r< r'_{N,M, z'}$ for a small enough $ r'_{N,M, z'}>0$. From the choice of $z'$ we have \[f^{m}(z') \notin \Lambda, \forall m \in [{R}^k(z')+1, {R}^{k+1}(z')-1], \forall k\in [0,N-1]\] and \[f^{{R}^k(z')}(z')\in \interior{(\Lambda)}, \forall k\in [0,N].\] Due to the Assumption \ref{geoassumption}, there is $r_{N,M,z'}>0$ such that, if $r<r'_{N,M,z'}$, $B_{M \cdot r}(z') \subseteq \interior{(\Lambda)}$ almost surely, then \[f^{m}(B_{M \cdot r}(z')) \subseteq \Lambda^c, \forall m \in [{R}^k(z')+1, {R}^{k+1}(z')-1], \forall k\in [0,N-1]\] and \[f^{{R}^k(z')}(B_{M \cdot r}(z'))\subseteq \Lambda \text{ almost surely}, \forall k\in [0,N].\] Since ${R}, {R}^2, \cdots, {R}^N$ are consecutive return times to $\Lambda$, we have ${R}^k|_{B_{M \cdot r}(z')}={R}^k(z'), \forall k\in[1, N]$. Thus this claim holds. \textbf{Claim:} $f^{{R}^k}$ for all $k\in[1, N]$ is a local diffeomorphism at $z'$. This claim holds because $f$ is a local diffeomorphism on $\bigcup_{i \ge 1} \bigcup_{0 \le j < R_i} f^j(\Lambda_i)$ and ${R}^k|_{B_{M \cdot r}(z')}={R}^k(z')$ for sufficiently small $r>0$. These two claims, together with the fact that $f^{{R}^k}(z')\in \interior{(\Lambda)}$ are distinct $\forall k\in[0, N],$ imply that there exists a small enough $r_{z',M,N}>0$, such that $\forall r<r_{z',M,N}$ the sets $f^{{R}^k}(B_{M \cdot r}(z'))$ are disjoint for all $k\in [0,N]$. So the lemma holds. \end{proof} Before estimating the super-short returns $\int_{B_{M\cdot r} (z')} 1_{\bigcup_{N\le k \le n^{\epsilon'}}(f^R)^{-k}B_{ M\cdot r} (z')}d\mu_{\Lambda}$, we need one more lemma. \begin{lemma}[Recurrences]\label{recurrence} There exists $r_M>0$, such that $ \forall r < r_M$ and $\forall \gamma^u \in \Gamma^u, i\ge N$ the following inequality holds \[\Leb_{\gamma^u}\{z'\in \Lambda \bigcap \gamma^u: d((f^{{R}})^{-i}z',z') \le M \cdot r\}\precsim_{\dim \gamma^u} (M \cdot r)^{\dim \gamma^u},\] where a constant in $\precsim_{\dim \gamma^u}$ depends upon $\dim \gamma^u$, but it does not depend upon $i\ge {N}$ and $\gamma^u \in \Gamma^u$. \end{lemma} \begin{proof} We start with making \textbf{Claim}: there are finitely many balls $\{B_{r'_i}(z'_i): 1 \le i \le N'\}$, where $N'\in \mathbb{N}$ depends only on $\Lambda$, such that all unstable fibers $\gamma^u \in\Gamma^u$ are almost flat in each $B_{r_i'}(z'_i)$. In view of the Definition \ref{gibbs} of $\Gamma^s$ (respectively, of $\Gamma^u$), for any $z'' \in \Lambda$ there exists a small open ball $B_{r''}(z'')$, such that all $\gamma^s \in \Gamma^s$ (respectively, $\gamma^u \in \Gamma^u)$ intersecting $B_{r''}(z'')$ are almost flat and parallel. Since $\Lambda$ is compact, one can find finitely many open balls $\{B_{{{r_1''}}}(z''_1), \cdots, B_{{r''_{N'}}}(z''_{N'})\}$, which cover $\Lambda$. Hence, the claim holds. Take now any of these balls, say $B_{r''_1}(z''_1)$, and any $\gamma^u \in \Gamma^u, z_1',z_2' \in \{z'\in \Lambda \bigcap \gamma^u\bigcap B_{r_1''}(z_1''): d((f^{{R}})^{-i}z',z')\le M \cdot r\} $. Then for any $r<\tau_M:=\min\{\frac{r_1''}{8M}, \cdots, \frac{r''_{N'}}{8M}\}$, \[d((f^{{R}})^{-i}z'_1,z_1') \le M \cdot r,\text{ } d((f^{{R}})^{-i}z'_2,z_2') \le M \cdot r.\] By making use of the Assumption \ref{assumption} together with $i \ge N$, we get that $C \cdot \beta^i \le C \cdot \beta^{N}< \frac{1}{2}$ and \[d(z'_1,z'_2)\le d(z'_1, (f^{{R}})^{-i}z'_1)+d((f^{{R}})^{-i}z'_1, (f^{{R}})^{-i}z'_2)+d((f^{{R}})^{-i}z'_2, z'_2)\] \[\le 2M \cdot r+ C \cdot \beta^i \cdot d(z'_1,z'_2) \le 2M \cdot r +\frac{1}{2} d(z'_1,z'_2).\] Thus $d(z'_1,z'_2) \le 4M \cdot r < \frac{r_1''}{2} \implies \diam \{z'\in \Lambda \bigcap B_{r_1''}(z_1'') \bigcap \gamma^u: d((f^{{R}})^{-i}z',z') \le M \cdot r\} \le 4M \cdot r$. Since $\gamma^u$ in the ball $B_{r_1''}(z_1'')$ is almost flat, then its Lebesgue measure can be estimated by diameters, i.e. \[\Leb_{\gamma^u}\{z'\in \Lambda \bigcap B_{r''_1}(z''_1): d((f^{{R}})^{-i}z',z') \le M \cdot r\} \precsim_{\dim \gamma^u} (M \cdot r)^{\dim \gamma^u}.\] This estimate also holds for the balls $B_{r_2''}(z_2''), \cdots, B_{r_{N'}''}(z_{N'}'')$. By summing over all the balls and noting that $\Lambda \subseteq \bigcup_{1\le i\le {N'}} B_{r_i''}(z_i'')$, we get \[\Leb_{\gamma^u}\{z'\in \Lambda: d((f^{{R}})^{-i}z',z') \le M \cdot r\} \precsim_{\dim \gamma^u} (M \cdot r)^{\dim \gamma^u},\] where a constant in $\precsim_{\dim \gamma^u}$ depends upon $\dim \gamma^u$, but it does not depend upon $i\ge {N}$ and $\gamma^u \in \Gamma^u$. \end{proof} \begin{lemma}[Super-short returns]\label{veryshortreturn}\ \par Choose $n= \lfloor \frac{T}{\mu(B_r(z))}\rfloor$. Then for almost every $z\in \mathcal{M}, z'\in \Lambda$ there exists $r_{z,z',M}>0$, such that $\forall r< r_{z,z',M}$ \[\int_{B_{M\cdot r} (z')} 1_{\bigcup_{N\le k \le n^{\epsilon'}}(f^{{R}})^{-k}B_{ M\cdot r} (z')}d\mu_{\Lambda}\precsim_{T,\epsilon} (M \cdot r)^{\dim_H\mu -\epsilon} \cdot M^{\frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{6}}\cdot r^{\frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{12}},\] where $\epsilon>0$ is the same as in the Proposition \ref{rate} and $\epsilon'<\frac{\min\{\dim_H \mu, { \dim \gamma^u}\}}{12 \dim_H \mu+12 \epsilon}$. \end{lemma} \begin{proof} It follows from (\ref{horsesrbleb}), $(f^R)_{*}\mu_{\Lambda}=\mu_{\Lambda}$ and Lemma \ref{recurrence} that \[\mu_{\Lambda}\{z'\in \Lambda: d((f^{{R}})^{k}z',z')\le M\cdot r\}=\mu_{\Lambda}\{z'\in \Lambda: d((f^{{R}})^{-k}z',z')\le M\cdot r\}\] \[= \int \mu_{\gamma^u}\{z'\in \Lambda: d((f^{{R}})^{-k}z',z')\le M\cdot r\}d\mu_{\Lambda}\] \[\precsim \int \Leb_{\gamma^u}\{z'\in \Lambda: d((f^{{R}})^{-k}z',z')\le M\cdot r\}d\mu_{\Lambda} \precsim_{\dim \gamma^u} (M \cdot r)^{\dim \gamma^u}.\] By the Assumption \ref{geoassumption}, for $\delta=\frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{6}$ and almost every $z'\in \Lambda$ there exists $r_{z',\delta}>0$, such that $r^{\dim_H\mu +\delta} \le \mu (B_r(z'))\le r^{\dim_H\mu-\delta}, \forall r< r_{z',\delta}$. Let $A_m:=\{z' \in \Lambda: r_{z',\delta}>\frac{1}{m}\}$. Then $\bigcup_{m} A_m=\Lambda$ and $\forall z'\in A_m$, $\forall r< \frac{1}{m}$ \begin{equation}\label{11'} r^{\dim_H\mu +\delta} \le \mu (B_r(z'))\le r^{\dim_H\mu-\delta} \end{equation} and \[\int_{B_{M\cdot r} (z')} 1_{\bigcup_{N\le k \le n^{\epsilon'}}(f^{{R}})^{-k}B_{ M\cdot r} (z')}d\mu_{\Lambda} \le \int_{B_{M\cdot r} (z')} 1_{\bigcup_{N\le k\le n^{\epsilon'}}d((f^{{R}})^{k}y,y)\le 2M\cdot r}d\mu_{\Lambda}(y).\] Let a kernel on $\Lambda\times A_m$ be $K(y,z'):=1_{B_{M \cdot r}(z')}(y)$. If $r<\frac{1}{3m \cdot M}$ then by (\ref{11'}) \[\int K(y,z') d\mu_{\Lambda}(y)=\mu_{\Lambda}(B_{M \cdot r}(z'))\precsim (M \cdot r)^{\dim_H \mu-\delta}.\] In order to estimate $\int K(y,z')\cdot 1_{A_m}(z')d\mu_{\Lambda}(z')=\mu_{\Lambda}(A_m \bigcap B_{M \cdot r}(y))$, observe that if $z'' \in A_m \bigcap B_{M \cdot r}(y)\neq \emptyset$, then $B_{M \cdot r}(y)\subseteq B_{3M \cdot r}(z'')$. By (\ref{11'}) again, $\int K(y,z')\cdot 1_{A_m}(z')d\mu_{\Lambda}(z')=\mu_{\Lambda}(A_m \bigcap B_{M \cdot r}(y)) \le \mu_{\Lambda}(B_{3M \cdot r}(z'')) \precsim (3M \cdot r)^{\dim_H \mu-\delta}$. Having the estimates of $K(y,z')$ above and Lemma \ref{recurrence}, we can use the Schur's test (see the Theorem 5.6 in \cite{tao}), i.e. for all $r<\min\{\frac{1}{3m \cdot M}, r_M\}$ \[\int 1_{A_m}(z')d\mu_{\Lambda}(z')\int_{B_{M\cdot r} (z')} 1_{d((f^{{R}})^{k}y,y)\le 2M\cdot r}d\mu_{\Lambda}(y) \precsim_{\dim \gamma^u} (3M\cdot r)^{\dim_H \mu-\delta} \cdot (M \cdot r)^{{ \dim \gamma^u}},\] where a constant in $\precsim_{\dim \gamma^u}$ does not depend upon $M,r,m,k$. Since $r^{\dim_H\mu +\delta} \le \mu (B_r(z'))$, one gets \[\int_{A_m}\frac{\int_{B_{M\cdot r} (z')} 1_{\bigcup_{N \le k \le [2^{\frac{4(\dim_H\mu+\epsilon)}{\min\{\dim_H\mu, \dim \gamma^u\}}}\cdot \frac{T}{r^{\dim_H\mu +\epsilon}}]^{\epsilon'}}d((f^{{R}})^{k}y,y)\le 2M\cdot r}d\mu_{\Lambda}(y) }{(\frac{T}{r^{\dim_H \mu+\epsilon}})^{\epsilon'}\cdot \mu (B_{M\cdot r}(z')) \cdot (M \cdot r)^{\delta}}d\mu_{\Lambda}(z')\] \[\precsim_{\dim_H \mu, \dim \gamma^u, \epsilon, \epsilon'} \frac{(M\cdot r)^{\dim_H \mu-\delta+\dim \gamma^u}}{(M\cdot r)^{\dim_H \mu+2\delta}} \le(M \cdot r)^{\frac{\min\{\dim_H\mu,{\dim \gamma^u}\}}{2}}.\] Choose now $r_i=i^{-\frac{4}{\min\{\dim_H\mu,\dim \gamma^u\}}}<\min\{\frac{1}{3m\cdot M},r_M\}$. Then, by the Borel-Cantelli Lemma, for almost every $z'\in A_m$ there exists $N_{M, m, z'}>\min\{\frac{1}{3m\cdot M}, r_M\}^{-\frac{\min\{\dim_H\mu,\dim \gamma^u\}}{4}}$, such that $\forall i> N_{M,m,z'}$ \begin{equation}\label{21} \int_{B_{M\cdot r_i} (z')} 1_{\bigcup_{N \le k \le [2^{\frac{4(\dim_H\mu+\epsilon)}{\min\{\dim_H\mu, \dim \gamma^u\}}}\cdot \frac{T}{r_i^{\dim_H\mu +\epsilon}}]^{\epsilon'}}d((f^{{R}})^{k}y,y)\le 2M\cdot r_i}d\mu_{\Lambda}(y)\le (\frac{T}{r_i^{\dim_H \mu+\epsilon}})^{\epsilon'}\cdot \mu (B_{M\cdot r_i}(z')) \cdot (M \cdot r_i)^{\delta}. \end{equation} Hence for almost every $z'\in \Lambda$ (in particular, for $z'\in \interior{(\Lambda)}$) there is $m_{z'}>0$, such that $z' \in A_{m_{z'}}$. Let $r_{z',M}:=\min\{r_M, N_{M,m_{z'},z'}^{-\frac{4}{\min\{\dim_H\mu, \dim \gamma^u\}}}\}$. By the Assumption \ref{geoassumption}, for $\epsilon$ from the Proposition \ref{rate} there exists such $r_{z,z',M}\in (0,r_{z',M})$ that $\forall r<r_{z,z',M}$ \[n\le \frac{T}{\mu(B_r(z))} \le \frac{T}{r^{\dim_H \mu +\epsilon}} \text{ and } \mu(B_{M \cdot r}(z'))\le (M \cdot r)^{\dim_H\mu -\epsilon}.\] Then $\forall r< r_{z,z',M}$, $\exists i >0$, such that $r_{i+1}\le r \le r_i$, and the following estimates hold \[\int_{B_{M\cdot r} (z')} 1_{\bigcup_{N\le k \le n^{\epsilon'}}(f^{{R}})^{-k}B_{ M\cdot r} (z')}d\mu_{\Lambda} \le \int_{B_{M\cdot r_i} (z')} 1_{\bigcup_{N\le k\le (\frac{T}{r^{\dim_H \mu +\epsilon}})^{\epsilon'}}d((f^{{R}})^{k}y,y)\le 2M\cdot r_i}d\mu_{\Lambda}(y)\] \[\le \int_{B_{M\cdot r_i} (z')} 1_{\bigcup_{N\le k\le (\frac{T}{r_{i+1}^{\dim_H \mu +\epsilon}})^{\epsilon'}}d((f^{{R}})^{k}y,y)\le 2M \cdot r_i}d\mu_{\Lambda}(y)\] \[\le \int_{B_{M\cdot r_i} (z')} 1_{\bigcup_{N\le k\le [(\frac{r_i}{r_{i+1}})^{\dim_H \mu +\epsilon}\cdot (\frac{T}{r_{i}^{\dim_H \mu +\epsilon}})]^{\epsilon'}}d((f^{{R}})^{k}y,y)\le 2M \cdot r_i}d\mu_{\Lambda}(y)\] \begin{equation}\label{12'} \le \int_{B_{M\cdot r_i} (z')} 1_{\bigcup_{N \le k \le (2^{\frac{4(\dim_H\mu+\epsilon)}{\min\{\dim_H\mu, \dim \gamma^u\}}}\cdot \frac{T}{r_i^{\dim_H\mu +\epsilon}})^{\epsilon'}}d((f^{{R}})^{k}y,y)\le 2M\cdot r_i}d\mu_{\Lambda}(y). \end{equation} Hence, if $\epsilon'< \frac{\min\{\dim \gamma^u, \dim_H \mu\}}{12\dim_H \mu+12\epsilon} $, then we can use (\ref{21}) to continue the estimate (\ref{12'}). Namely, \[\int_{B_{M\cdot r} (z')} 1_{\bigcup_{N\le k \le n^{\epsilon'}}(f^{{R}})^{-k}B_{ M\cdot r} (z')}d\mu_{\Lambda}\le (M \cdot r_i)^{\dim_H\mu -\epsilon} \cdot (M \cdot r_i)^{\frac{\min\{\dim \gamma^u, \dim_H \mu\}}{6}}\cdot (\frac{T}{r_i^{\dim_H \mu+\epsilon}})^{\epsilon'}\] \[\le (M \cdot r \cdot \frac{r_i}{r_{i+1}})^{\dim_H\mu -\epsilon} \cdot (M \cdot r \cdot \frac{r_i}{r_{i+1}})^{\frac{\min\{\dim \gamma^u, \dim_H \mu\}}{6}}\cdot (\frac{T}{r^{\dim_H \mu+\epsilon}})^{\epsilon'}\] \[\precsim_{T,\epsilon} (M \cdot r)^{\dim_H\mu -\epsilon} \cdot M^{\frac{\min\{\dim \gamma^u, \dim_H \mu\}}{6}}\cdot r^{\frac{\min\{\dim \gamma^u, \dim_H \mu\}}{12}}.\] The last inequality holds because $\frac{r_i}{r_{i+1}} \precsim 1$. \end{proof} \begin{lemma}[Not super-short, but short returns]\label{notveryshort}\ \par Let $n= \lfloor \frac{T}{\mu(B_r(z))}\rfloor,p=\lfloor {\lfloor \frac{T}{\mu(B_r(z))}\rfloor}^{\frac{\text{dim}_H\mu-\epsilon}{\text{dim}_H\mu}}\rfloor$ and $n^{\epsilon'}\ll p$. Then, for almost all $z\in \mathcal{M}$, $z'\in \interior{(\Lambda)}$, there exists $r_{z,z',M}>0$, such that $\forall r< r_{z,z',M}$ \[\int_{B_{ M\cdot r} (z')} 1_{\bigcup_{n^{\epsilon'}\le k \le p}(f^R)^{-k}B_{ M\cdot r} (z')}d\mu_{\Lambda}\] \[\precsim_{T,\epsilon} (M\cdot r)^{\dim_H\mu-\epsilon^3} \cdot {\beta_2}^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}+\mu_{\Lambda}[ B_{M \cdot r + 2C \cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}}(z') \setminus B_{M \cdot r}(z')]\cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}\] \[+r^{-(\dim_H\mu+\epsilon) \cdot \frac{\text{dim}_H\mu-\epsilon}{\text{dim}_H\mu}} \cdot \{(M\cdot r)^{\dim_H\mu-\epsilon^3} +\mu_{\Lambda}[B_{M \cdot r + 2C \cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}}(z') \setminus B_{M \cdot r}(z')]\}^2,\] where $C>1$ is the same as in the Definition \ref{gibbs}, $\beta_2\in (0,1) $ does not depend on $z,z',r,M$. \end{lemma} \begin{proof} Observe first that $n^{\epsilon'}\ll p$ implies $\epsilon'<\frac{\dim_H \mu-\epsilon}{\dim_H \mu}$. Cover $B_{M \cdot r}(z')$ by two-sided cylinders $\xi_{i_{-m}\cdots i_0 \cdots i_m}$, where $m:=\frac{n^{\epsilon'}}{4}$ and $\xi_{i_{-m}\cdots i_0 \cdots i_m} \bigcap B_{M \cdot r}(z')\neq \emptyset$. By Lemma \ref{diamcylinder}, we have $\diam \xi_{i_{-m}\cdots i_0 \cdots i_m} \le 2C \cdot \beta^m$. So $(\bigcup_{}\xi_{i_{-m}\cdots i_0 \cdots i_m}) \setminus B_{M\cdot r}(z') \subseteq B_{M \cdot r +2 C \cdot \beta^m}(z') \setminus B_{M \cdot r}(z')$. Denote $i_0:=p-n^{\epsilon'}$. Then \[\int 1_{B_{M \cdot r}(z')} \cdot 1_{\ge 1} \circ [1_{B_{M \cdot r}(z')}+ \cdots +1_{B_{M \cdot r}(z')}\circ (f^R)^{i_0}] \circ (f^R)^{n^{\epsilon'}} d\mu_{\Lambda}\] \begin{equation}\label{2} =\int 1_{(f^R)^{-m}B_{M \cdot r}(z')} \cdot 1_{\ge 1} \circ [1_{(f^R)^{-m}B_{M \cdot r}(z')}+ \cdots +1_{(f^R)^{-m}B_{M \cdot r}(z')}\circ (f^R)^{i_0}] \circ (f^R)^{n^{\epsilon'}} d\mu_{\Lambda}. \end{equation} Let $A_1:=(f^R)^{-m}B_{M \cdot r}(z'), A_2:=(f^R)^{-m} \bigcup_{\xi_{i_{-m}\cdots i_0 \cdots i_m} \bigcap \partial B_{M\cdot r}(z')\neq \emptyset} \xi_{i_{-m}\cdots i_0 \cdots i_m}$, $A_0:=(f^R)^{-m} \bigcup \xi_{i_{-m}\cdots i_0 \cdots i_m}$. Observe that \[A_0,A_2 \in \mathcal{M}_{2m}, A_0=A_1 \bigcup A_2, \text{ and } (f^R)^mA_2 \subseteq B_{M \cdot r + 2C \cdot \beta^m}(z') \setminus B_{M \cdot r-2C \cdot \beta^m}(z'),\] and, moreover, the function $1_{\ge 1} \circ [1_{A_0}+ \cdots +1_{A_0}\circ (f^R)^{i_0}]$ is constant along each $\gamma^s\in \Gamma^s$. Then we can continue the equality (\ref{2}) as \[=[\int 1_{A_1} \cdot 1_{\ge 1} \circ [1_{A_1}+ \cdots +1_{A_1}\circ (f^R)^{i_0}] \circ (f^R)^{n^{\epsilon'}} d\mu_{\Lambda}\] \[-\int 1_{A_1} \cdot 1_{\ge 1} \circ [1_{A_0}+ \cdots +1_{A_0}\circ (f^R)^{i_0}] \circ (f^R)^{n^{\epsilon'}} d\mu_{\Lambda}]\] \[-\int 1_{A_0\setminus{A_1}} \cdot 1_{\ge 1} \circ [1_{A_0}+ \cdots +1_{A_0}\circ (f^R)^{i_0}] \circ (f^R)^{n^{\epsilon'}} d\mu_{\Lambda}\] \[+[\int 1_{A_0} \cdot 1_{\ge 1} \circ [1_{A_0}+ \cdots +1_{A_0}\circ (f^R)^{i_0}] \circ (f^R)^{n^{\epsilon'}} d\mu_{\Lambda}\] \[-\int 1_{A_0} d\mu_{\Lambda} \cdot \int 1_{\ge 1} \circ [1_{A_0}+ \cdots +1_{A_0}\circ (f^R)^{i_0}] \circ (f^R)^{n^{\epsilon'}} d\mu_{\Lambda}]\] \begin{equation}\label{7} +\int 1_{A_0} d\mu_{\Lambda} \cdot \int 1_{\ge 1} \circ [1_{A_0}+ \cdots +1_{A_0}\circ (f^R)^{i_0}] \circ (f^R)^{n^{\epsilon'}} d\mu_{\Lambda}. \end{equation} Apply now Lemma \ref{decorrelationcylinder}, the relation $A_0\setminus A_1 \subseteq A_2$ and the inequality \[|1_{\ge 1} \circ [1_{A_1}+ \cdots +1_{A_1}\circ (f^R)^{i_0}]- 1_{\ge 1} \circ [1_{A_0}+ \cdots +1_{A_0}\circ (f^R)^{i_0}]|\le 1_{\ge 1} \circ [1_{A_2}+ \cdots +1_{A_2}\circ (f^R)^{i_0}] \] to the right hand side of (\ref{7}). Then it can be estimated as \[\le \int 1_{A_1} \cdot 1_{\ge 1} \circ [1_{A_2}+ \cdots +1_{A_2}\circ (f^R)^{i_0}] \circ (f^R)^{n^{\epsilon'}} \] \[+\int 1_{A_2} \cdot 1_{\ge 1} \circ [1_{A_0}+ \cdots +1_{A_0}\circ (f^R)^{i_0}] \circ (f^R)^{n^{\epsilon'}} d\mu_{\Lambda}+C'' \cdot \mu_{\Lambda}(A_0) \cdot \beta_1^{n^{\epsilon'}}+\mu_{\Lambda}(A_0)^2 \cdot i_0\] \[\le 2\int 1_{A_0} \cdot 1_{\ge 1} \circ [1_{A_0}+ \cdots +1_{A_0}\circ (f^R)^{i_0}] \circ (f^R)^{n^{\epsilon'}} d\mu_{\Lambda}+C'' \cdot \mu_{\Lambda}(A_0) \cdot \beta_1^{n^{\epsilon'}}+\mu_{\Lambda}(A_0)^2 \cdot i_0\] \[-2\int 1_{A_0} d\mu_{\Lambda}\cdot \int 1_{\ge 1} \circ [1_{A_0}+ \cdots +1_{A_0}\circ (f^R)^{i_0}] \circ (f^R)^{n^{\epsilon'}} d\mu_{\Lambda}\] \[+2\int 1_{A_0} d\mu_{\Lambda}\cdot \int 1_{\ge 1} \circ [1_{A_0}+ \cdots +1_{A_0}\circ (f^R)^{i_0}] \circ (f^R)^{n^{\epsilon'}} d\mu_{\Lambda}.\] By applying Lemma \ref{decorrelationcylinder} again, with $i_0 \le p, m=\frac{n^{\epsilon'}}{4}$, we have \[\le 3C'' \cdot \mu_{\Lambda}(A_0) \cdot \beta_1^{n^{\epsilon'}}+3 i_0 \cdot \mu_{\Lambda}(A_0)^2\le 3C'' \cdot \mu_{\Lambda}(A_0) \cdot \beta_1^{\frac{n^{\epsilon'}}{2}}+3 p \cdot \mu_{\Lambda}(A_0)^2\] \[=3C'' \cdot \mu_{\Lambda}((f^R)^mA_0) \cdot \beta_1^{\frac{n^{\epsilon'}}{2}}+3 p \cdot \mu_{\Lambda}((f^R)^mA_0)^2\precsim \mu_{\Lambda}(B_{M\cdot r}(z')) \cdot \beta_1^{\frac{n^{\epsilon'}}{2}}\] \begin{equation}\label{3} +\mu_{\Lambda}[ B_{M \cdot r + 2C \cdot \beta^{\frac{n^{\epsilon'}}{4}}}(z') \setminus B_{M \cdot r}(z')]\cdot \beta_1^{\frac{n^{\epsilon'}}{2}}+p \cdot \{\mu_{\Lambda}(B_{M\cdot r}(z')) +\mu_{\Lambda}[B_{M \cdot r + 2C \cdot \beta^{\frac{n^{\epsilon'}}{4}}}(z') \setminus B_{M \cdot r}(z')]\}^2. \end{equation} By making use of the Assumption \ref{geoassumption}, we choose now the same $\epsilon>0$ as in the Proposition \ref{rate}. Then for almost all $z\in \mathcal{M}$, $ z'\in \interior{(\Lambda)}$ there is $r_{z,z',M}>0$ s.t. $\forall r< r_{z,z',M}$ \[B_{M \cdot r}(z') \subseteq \Lambda \text{ almost surely, and } \frac{T}{r^{\dim_H\mu-\epsilon}}\precsim n \precsim \frac{T}{r^{\dim_H\mu+\epsilon}},\] \[r^{\dim_H\mu+\epsilon}\le \mu(B_r(z)) \le r^{\dim_H\mu-\epsilon}, \frac{(M \cdot r)^{\dim_H\mu+\epsilon^3}}{\mu(\Lambda)}\le \mu_{\Lambda}(B_{M \cdot r}(z')) \le \frac{(M\cdot r)^{\dim_H\mu-\epsilon^3}}{\mu(\Lambda) },\] \[ (\frac{T}{r^{\dim_H\mu-\epsilon}})^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \precsim p\precsim (\frac{T}{r^{\dim_H\mu+\epsilon}})^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu}},(\frac{T}{r^{\dim_H\mu-\epsilon}})^{\epsilon'}\precsim n^{\epsilon'} \precsim (\frac{T}{r^{\dim_H\mu+\epsilon}})^{\epsilon'}.\] Then we can continue the estimate in the inequality (\ref{3}) as \[\precsim_{T,\epsilon} (M\cdot r)^{\dim_H\mu-\epsilon^3} \cdot {\beta_2}^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}+\mu_{\Lambda}[ B_{M \cdot r + 2C \cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}}(z') \setminus B_{M \cdot r}(z')]\cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}\] \[+r^{-(\dim_H\mu+\epsilon) \cdot \frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \cdot \{(M\cdot r)^{\dim_H\mu-\epsilon^3} +\mu_{\Lambda}[B_{M \cdot r + 2C \cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}}(z') \setminus B_{M \cdot r}(z')]\}^2,\] where $\beta_2 \in (0,1)$ depends on $T, \epsilon', \beta, \beta_1$. \end{proof} By combining the Lemmas \ref{fixreturn}, \ref{veryshortreturn} and \ref{notveryshort}, we obtain the following summary of the results. \begin{proposition}[Rates of short returns ]\label{shortreturnrate}\ \par Let $\epsilon, \epsilon'>0$ satisfy the relations $\alpha \cdot (\dim_H\mu-\epsilon)>\frac{\dim_H\mu}{\dim_H\mu-\epsilon}>1$, $p=\lfloor {\lfloor \frac{T}{\mu(B_r(z))}\rfloor}^{\frac{\text{dim}_H\mu-\epsilon}{\text{dim}_H\mu}}\rfloor$, $\epsilon'<\min\{\frac{\min\{\dim_H \mu, { \dim \gamma^u}\}}{12 \dim_H \mu+12 \epsilon}, \frac{\dim_H\mu-\epsilon}{\dim_H\mu}\}$. Then for almost all $z\in \mathcal{M}$, $z'\in \interior{(\Lambda)}$ and for each integer $M>0$ there exists a small enough $r_{z,z',M}>0$ such that $\forall r< r_{z,z',M}$ \[\int_{B_{ M\cdot r} (z')} 1_{\bigcup_{1\le k \le p}(f^R)^{-k}B_{ M\cdot r} (z')}d\mu_{\Lambda}\precsim_{T, \epsilon} (M \cdot r)^{\dim_H\mu -\epsilon} \cdot M^{\frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{6}}\cdot r^{\frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{12}}\] \[+(M\cdot r)^{\dim_H\mu-\epsilon^3} \cdot {\beta_2}^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}+\mu_{\Lambda}[ B_{M \cdot r + 2C \cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}}(z') \setminus B_{M \cdot r}(z')]\cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}\] \[+r^{-(\dim_H\mu+\epsilon) \cdot \frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \cdot \{(M\cdot r)^{\dim_H\mu-\epsilon^3} +\mu_{\Lambda}[B_{M \cdot r + 2C \cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}}(z') \setminus B_{M \cdot r}(z')]\}^2,\] where a constant $C>1$ is the same as in the Definition \ref{gibbs}, a constant $\beta_2\in (0,1) $ is independent on $z,z',r,M$. \end{proposition} \begin{proof} Recall that in the Lemmas \ref{fixreturn} \ref{veryshortreturn}, \ref{notveryshort}, we fixed an integer $M>0$. Then for almost all $z \in \mathcal{M}$ and $z' \in \interior{(\Lambda)}$ the desired estimates hold. The set of such points in $ \mathcal{M} \times \interior{(\Lambda)}$ has the full measure with respect to $\mu(\mathcal{M}) \times \mu(\interior{(\Lambda)})$, and it does depend on $M$. However, since $M$ is an integer, we can find a smaller set of points $(z,z') \in \mathcal{M} \times \interior{(\Lambda)}$ such that it does not depend on $M>0$ and has full measure $\mu(\mathcal{M}) \times \mu(\interior{(\Lambda)})$. Therefore the required estimate holds. \end{proof} \subsection{Coronas} Here we study two coronas, which appeared in the previous sections. One of them is in $\mathcal{M}$ and has a measure $\mu[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)] $ (see the Proposition \ref{rate}). The other one is in $\Lambda$ and its measure is $\mu_{\Lambda}[ B_{M \cdot r + 2C \cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}}(z') \setminus B_{M \cdot r}(z')]$ (see the Proposition \ref{shortreturnrate}). \begin{proposition}[Coronas in $\mathcal{M}$ and $\Lambda$]\label{coronaestimate}\ \par For almost every $z \in \mathcal{M}, z' \in \interior{(\Lambda)}$ there are $r_z, r_{z',M}>0$ such that $\forall r< \min \{r_z, r_{z',M}\} $ \[\mu\{B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)\} \precsim_{z, \dim \gamma^u} r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u}{2}},\] \[\mu_{\Lambda}[ B_{M \cdot r + 2C \cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}}(z') \setminus B_{M \cdot r}(z')]\precsim_{z, \dim \gamma^u} (M\cdot r)^{\frac{\dim \gamma^u }{2}} \cdot \beta_2^{\frac{r^{-(\dim_H\mu-\epsilon)\epsilon'}\cdot \dim \gamma^u}{2}}.\] \end{proposition} \begin{proof} For corona in $\mathcal{M}$ we have from Lemma \ref{ballinhorseshoe} for almost all $z \in \bigcup_{i \ge 1} \bigcup_{0 \le j < R_i} f^j(\Lambda_i)$, $z=f^{j_z}(z')$, $z' \in \interior{(\Lambda)}$ that there exists such $r'_z>0$, that $\forall r< r'_z$ \[\mu(f^{-j_z}[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)] \bigcap \{\bigcup_{i \ge 1} \bigcup_{1 \le j < R_i} f^j(\Lambda_i)\})=0.\] Hence by the invariance of $\mu$ (i.e. $f_* \mu=\mu$) we have \[\mu([B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)] \bigcap f^{j_z}\{\bigcup_{i \ge 1} \bigcup_{1 \le j < R_i} f^j(\Lambda_i)\})=0.\] Therefore \[[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)] \subseteq f^{j_z}\Lambda \text{ almost surely}.\] Because $f^{j_z}$ is a local diffeomorphism, then for a sufficiently small $r$ all manifolds $f^{j_z}\gamma^u$ ($\gamma^u\in \Gamma^u$) in any non-empty set $(f^{j_z}\gamma^u) \bigcap [B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)] $ are almost flat. From the Gauss lemma for the exponential map $\exp_{z}$ we have, that in a neughborhood of $z \in \mathcal{M}$ \[(\exp_z)^{-1}[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)] \] \[\subseteq \Cor:=\{v \in \mathbb{R}^{\dim \mathcal{M}}: \langle v,v \rangle_{z}\in [r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}}, r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}}]\}, \] where $\langle \cdot, \cdot \rangle_z$ is the Riemannian metric at $z\in \mathcal{M}$. This is a corona in an ellipse. If $r$ is sufficiently small, say $r<r_z<r_z'$ for some $r_z>0$, then all manifolds $(\exp_z)^{-1} f^{j_z}\gamma^u$ in $\Cor$ are almost flat. So their diameters (in Euclidean norm) satisfy the inequality \[\diam \{[(\exp_z)^{-1} f^{j_z}\gamma^u]\bigcap \Cor\} \precsim_z \cdot \sqrt{(r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}})^2-(r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}})^2}\] $\precsim \sqrt{r \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}}}=r^{\frac{1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}}{2}}$. Therefore the Lebesgue measure of $\Cor$ on $(\exp_z)^{-1} f^{j_z}\gamma^u$ can be controlled by diameters, i.e. \[\Leb_{(\exp_z)^{-1} f^{j_z}\gamma^u}\{[(\exp_z)^{-1} f^{j_z}\gamma^u]\bigcap \Cor\} \precsim_{z, \dim \gamma^u} r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u}{2}}.\] Since $\exp_z$ is a local diffeomorphism, then \[\Leb_{f^{j_z}\gamma^u} \{B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)\} \precsim_{z, \dim \gamma^u} r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u}{2}}.\] Now, since $f^{j_z}$ is also a local diffeomorphism, we have \[\Leb_{\gamma^u} \{f^{-j_z}[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)]\} \precsim_{z, \dim \gamma^u} r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u}{2}}.\] By making use of (\ref{horsesrbleb}) and integrating over all $\gamma^u \in \Gamma^u$, we get \[\mu\{f^{-j_z}[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)]\} \precsim_{z, \dim \gamma^u} r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u}{2}}.\] Hence, thanks to the invariance of the measure $\mu$ ($f_{*}\mu=\mu$) \[\mu\{B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)\} \precsim_{z, \dim \gamma^u} r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u}{2}}.\] For the corona in $\Lambda$, i.e. for $B_{M \cdot r +2 C \cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}}(z') \setminus B_{M \cdot r}(z')$, a trick is the same and even more straightforward. Indeed, because $\gamma^u\in \Gamma^u$ intersects with this corona directly, there is $r_{z', M}>0$, such that $\forall r<r_{z',M}$ \[\mu_{\Lambda}[ B_{M \cdot r + 2C \cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}}(z') \setminus B_{M \cdot r}(z')]\precsim_{z', \dim \gamma^u} (M \cdot r)^{\frac{\dim \gamma^u }{2}} \cdot \beta_2^{ \frac{r^{-(\dim_H\mu-\epsilon)\epsilon'}\cdot\dim \gamma^u}{2}}.\] \end{proof} \subsection{Conclusion of the Proof of the Theorem \ref{thm}} We will start with a rough estimate of $\dim_H \mu$. \subsection*{Proof that $\dim_H\mu \ge \dim {\gamma}^u$:} \begin{lemma}\label{roughdim} It follows from the Assumptions \ref{geoassumption} and \ref{assumption} that $\dim_H\mu \ge \dim {\gamma}^u$. \end{lemma} \begin{proof} Due to the ergodicity of $\mu$, it is enough to consider $z' \in \interior{(\Lambda)}$ and a ball $B_r(z')$ in $\Lambda$. If $r$ is sufficiently small then $\gamma^u \in \Gamma^u$ is almost flat in $B_r(z')\bigcap \gamma^u$. Therefore $\diam \{B_r(z')\bigcap \gamma^u\} \precsim_{z'} r$. By the same trick as in the Proposition \ref{coronaestimate} we get \[\Leb_{\gamma^u}(B_r(z')) \precsim_{z',\dim \gamma^u}r^{\dim \gamma^u} \text{ and } \mu(B_r(z')) \precsim_{z',\dim \gamma^u}r^{\dim \gamma^u}.\] On the other hand, by the Assumption \ref{geoassumption} for almost all $z' \in \mathcal{M}$ and any $m \in \mathbb{N}$ there is $r_{z',m}>0$, such that $\mu(B_r(z')) \ge r^{\dim_H \mu+ \frac{1}{m}}$ for any $r<r_{z',m}$. This implies that $\dim \gamma^u \le \dim_H \mu + \frac{1}{m}, \forall m \in \mathbb{N}$. By taking the limit $m\to \infty$ one gets $\dim_H\mu \ge \dim {\gamma}^u$. \end{proof} We will address now the convergence rates in the Theorem \ref{thm}. According to the Proposition \ref{rate} it is enough to find the convergence rates for short returns and coronas. A proof will consist of several steps. \subsection*{Pull back short returns $\int_{B_r(z)} 1_{\bigcup_{1\le j \le p}f^{-j}B_r(z)}d\mu$.} For almost any $z \in \bigcup_{i \ge 1} \bigcup_{0 \le j < R_i} f^j(\Lambda_i)$ we have $z=f^{j_z}(z')$ for some $z' \in \interior{\Lambda}$. By the Lemma \ref{comparetopballmetricball}, there is a topological ball $TB_r(z')=TB_r(f^{-j_z}(z))$ and a constant $C_z>1$, such that $B_{C^{-1}_z\cdot r} (f^{-j_z}z) \subseteq TB_r(f^{-j_z}z) \subseteq B_{C_z\cdot r} (f^{-j_z}z)$. \begin{lemma}[Pull short returns back to $\Lambda$]\label{pullbackhorseshoe}\ \par There exists a small enough $r_z>0$ such that $\forall r< r_z$ the following inequality holds \[ \int_{B_r(z)} 1_{\bigcup_{1\le k \le p}f^{-k}B_r(z)}d\mu \le \mu(\Lambda) \cdot \int_{B_{C_z\cdot r} (f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}(f^R)^{-k}B_{C_z\cdot r} (f^{-j_z}z)}d\mu_{\Lambda} .\] \end{lemma} \begin{proof} It follows from the invariance of $\mu$ (i.e. $f_*\mu=\mu$) that \[ \int_{B_r(z)} 1_{\bigcup_{1\le k \le p}f^{-k}B_r(z)}d\mu=\int_{TB_r(f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}f^{-k}TB_r(f^{-j_z}z)}d\mu.\] By Lemma \ref{ballinhorseshoe}, we have $TB_r(f^{-j_z}z) \subseteq \Lambda$ almost surely $\forall r< r_z$ for some small enough $r_z>0$. Then the equality above can be continued as \begin{equation}\label{4} = \int_{TB_r(f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}f^{-k}TB_r(f^{-j_z}z)}d\mu|_{\Lambda} = \mu(\Lambda) \cdot \int_{TB_r(f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}f^{-k}TB_r(f^{-j_z}z)}d\mu_{\Lambda}. \end{equation} Denote $R^i:=R^{i-1}+R \circ f^R, R^1:=R\ge 1$. Then $p\le R^p$. Also note that, since $R$ is the first return time, then $f^k \notin \Lambda$ almost surely (a.s.) if $R^i<k<R^{i+1}$. Therefore \[1_{TB_r(f^{-j_z}z)} \cdot 1_{\bigcup_{1\le k \le R^p}f^{-k}TB_r(f^{-j_z}z)}=1_{TB_r(f^{-j_z}z)} \cdot 1_{\bigcup_{1\le k \le p}(f^R)^{-k}TB_r(f^{-j_z}z)} \text{ a.s.}.\] Then we can estimate (\ref{4}) as \[\le \mu(\Lambda) \cdot \int_{TB_r(f^{-j_z}z)} 1_{\bigcup_{1\le k \le R^p}f^{-k}TB_r(f^{-j_z}z)}d\mu_{\Lambda}\] \begin{equation}\label{5} =\mu(\Lambda) \cdot \int_{TB_r(f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}(f^R)^{-k}TB_r(f^{-j_z}z)}d\mu_{\Lambda}. \end{equation} By Lemma \ref{comparetopballmetricball} there is a constant $C_z\ge 1$, such that $TB_r(f^{-j_z}z) \subseteq B_{C_z\cdot r} (f^{-j_z}z)$, and we can continue the equality (\ref{5}) as \[\le\mu(\Lambda) \cdot \int_{B_{C_z\cdot r} (f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}(f^R)^{-k}B_{C_z\cdot r} (f^{-j_z}z)}d\mu_{\Lambda}.\] \end{proof} Hence the short returns problem on $\mathcal{M}$ becomes a short returns problem on $\Lambda$. The next task is to \subsection*{Estimate $\frac{1}{\mu(B_r(z))}\cdot \int_{B_{C_z\cdot r} (f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}(f^R)^{-k}B_{C_z\cdot r} (f^{-j_z}z)}d\mu_{\Lambda}$.} \begin{lemma}\label{afterpullback} If $\epsilon<\min\{\frac{\min\{\dim \gamma^u, \dim_H \mu\}}{24},\frac{1}{3\dim_H\mu}\}$ satisfies $\alpha \cdot (\dim_H\mu-\epsilon)>\frac{\dim_H\mu}{\dim_H\mu-\epsilon}>1$, then for almost every $z\in \mathcal{M}$ there is $r_z>0$, such that $\forall r<r_z$ \[\frac{1}{\mu(B_r(z))}\cdot \int_{B_{C_z\cdot r} (f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}(f^R)^{-k}B_{C_z\cdot r} (f^{-j_z}z)}d\mu_{\Lambda}\precsim_{T,z,\epsilon} r^{\ \frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{12}-2\epsilon}+ r^{\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3}\to 0.\] \end{lemma} \begin{proof} Observe that $C_z < \infty $ for almost all $z\in \mathcal{M}$. Denote $F_1:=\{C_z<\infty\}$, $F_2 \times F_3:=\{(z,z')\in \mathcal{M}\times \interior{(\Lambda)}: \text{ the Propositions }\ref{shortreturnrate}, \ref{coronaestimate} \text{ hold}\}$. Define a new measure one set in $\mathcal{M}$ as $F:=F_1 \bigcap F_2 \bigcap \{\bigcup_{j\ge 0}f^j(\interior{(\Lambda)}\bigcap F^c_3 )\}^c$. Then for any $z \in F$ we have $ C_z < \infty$ and $(z,f^{-j_z}z)\in F_2 \times F_3$. Let $M:=\lfloor C_z \rfloor+1, z'=f^{-j_z}z $. Then by the Proposition \ref{shortreturnrate} there is $r_{z,z',M}=r_{z,f^{-j_z}z,\lfloor C_z \rfloor+1}>0$ such that $\forall r< r_{z,f^{-j_z}z,\lfloor C_z \rfloor+1}$ \[\int_{B_{C_z\cdot r} (f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}(f^R)^{-k}B_{C_z\cdot r} (f^{-j_z}z)}d\mu_{\Lambda}\le \int_{B_{M\cdot r} (z')} 1_{\bigcup_{1\le k \le p}(f^R)^{-k}B_{M\cdot r} (z')}d\mu_{\Lambda}\] \[\precsim_{T,\epsilon} (M \cdot r)^{\dim_H\mu -\epsilon} \cdot M^{\frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{6}}\cdot r^{\frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{12}}\] \[+(M\cdot r)^{\dim_H\mu-\epsilon^3} \cdot {\beta_2}^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}+\mu_{\Lambda}[ B_{M \cdot r +2 C \cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}}(z') \setminus B_{M \cdot r}(z')]\cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}\] \[+r^{-(\dim_H\mu+\epsilon) \cdot \frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \cdot \{(M\cdot r)^{\dim_H\mu-\epsilon^3} +\mu_{\Lambda}[B_{M \cdot r + 2C \cdot \beta_2^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}}(z') \setminus B_{M \cdot r}(z')]\}^2.\] By the Proposition \ref{coronaestimate}, the right hand side of the inequality above can be estimated as \[\precsim_{T,C_z,z, \epsilon} r^{\dim_H\mu -\epsilon+\frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{12}}+ r^{\dim_H \mu-\epsilon^3} \cdot {\beta_2}^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}+r^{\frac{\dim \gamma^u }{2}} \cdot \beta_2^{\frac{r^{-(\dim_H\mu-\epsilon)\epsilon'}\cdot \dim \gamma^u}{2}+r^{-(\dim_H\mu-\epsilon)\epsilon'}}\] \[+r^{-(\dim_H\mu+\epsilon) \cdot \frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \cdot \{ r^{\dim_H\mu-\epsilon^3} +r^{\frac{\dim \gamma^u }{2}} \cdot \beta_2^{\frac{r^{-(\dim_H\mu-\epsilon)\epsilon'}\cdot \dim \gamma^u}{2}}\}^2.\] By the Assumption \ref{geoassumption} for almost all $z\in \mathcal{M}$ there is $r_z>0$ such that $\forall r<r_{z}$ \[r^{\dim_H\mu+\epsilon^3}\le \mu(B_r(z)) \le r^{\dim_H\mu-\epsilon^3}.\] Then $\forall r<\min \{ r_{z,f^{-j_z}z,\lfloor C_z \rfloor+1}, r_z\}$, \[\frac{1}{\mu(B_r(z))}\cdot \int_{B_{C_z\cdot r} (f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}(f^R)^{-k}B_{C_z\cdot r} (f^{-j_z}z)}d\mu_{\Lambda}\] \[\precsim_{T,z, \epsilon} \frac{1}{r^{\dim_H\mu+\epsilon^3}} \cdot \{r^{\dim_H\mu -\epsilon+\frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{12}}+r^{\dim_H \mu-\epsilon} \cdot {\beta_2}^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}\] \[+r^{\frac{\dim \gamma^u }{2}} \cdot \beta_2^{\frac{r^{-(\dim_H\mu-\epsilon)\epsilon'}\cdot \dim \gamma^u}{2}+r^{-(\dim_H\mu-\epsilon)\epsilon'}}+r^{-(\dim_H\mu+\epsilon) \cdot \frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \cdot ( r^{\dim_H\mu-\epsilon^3} +r^{\frac{\dim \gamma^u }{2}} \cdot \beta_2^{\frac{r^{-(\dim_H\mu-\epsilon)\epsilon'}\cdot \dim \gamma^u}{2}})^2\}\] \[\le r^{\ -2\epsilon+\frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{12}}+r^{-\epsilon-\epsilon^3} \cdot {\beta_2}^{r^{-(\dim_H\mu-\epsilon)\epsilon'}}+r^{\frac{\dim \gamma^u }{2}-\epsilon^3-\dim_H\mu} \cdot \beta_2^{\frac{r^{-(\dim_H\mu-\epsilon)\epsilon'}\cdot \dim \gamma^u}{2}+r^{-(\dim_H\mu-\epsilon)\epsilon'}}\] \[+\frac{1}{r^{\dim_H\mu+\epsilon^3}} \cdot r^{-(\dim_H\mu+\epsilon) \cdot \frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \cdot \{ r^{\dim_H\mu-\epsilon^3} +r^{\frac{\dim \gamma^u }{2}} \cdot \beta_2^{\frac{r^{-(\dim_H\mu-\epsilon)\epsilon'}\cdot \dim \gamma^u}{2}}\}^2\] \[\precsim r^{\ \frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{12}-2\epsilon}+r^{\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3}\to 0,\] this last $\precsim$ is because $\beta_2^{r^{-c}}\ll r^{c'}$ for any $ c,c'>0$. \end{proof} By combining the Lemmas \ref{pullbackhorseshoe} and \ref{afterpullback} we get an \subsection*{Estimate of the short returns rates $\frac{1}{\mu(B_r(z))}\int_{B_r(z)} 1_{\bigcup_{1\le j \le p}f^{-j}B_r(z)}d\mu$.} \begin{lemma}\label{lastshortreturnrate} If $\epsilon<\min\{\frac{\min\{\dim \gamma^u, \dim_H \mu\}}{24},\frac{1}{3\dim_H\mu}\}$ satisfies $\alpha \cdot (\dim_H\mu-\epsilon)>\frac{\dim_H\mu}{\dim_H\mu-\epsilon}>1$, then for almost all $z\in \mathcal{M}$ \[\frac{1}{\mu(B_r(z))}\int_{B_r(z)} 1_{\bigcup_{1\le j \le p}f^{-j}B_r(z)}d\mu \precsim_{z,T,\epsilon} r^{\ \frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{12}-2\epsilon}+ r^{\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3}\to 0.\] \end{lemma} We finished now the estimates of short returns and move to \subsection*{Estimate of coronas rates $\frac{\mu[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)]}{\mu(B_r(z))}$. } \begin{lemma}\label{lastcoronarate} If $\alpha> \frac{2}{\dim \gamma^u}-\frac{1}{\dim_H\mu}$ and $\epsilon<\min\{\frac{\min\{\dim \gamma^u, \dim_H \mu\}}{24},\frac{1}{3\dim_H\mu}\}$ is small enough, so that $\alpha >\frac{\frac{2}{\dim \gamma^u}-\frac{1}{\dim_H \mu}+\frac{2\epsilon}{\dim \gamma^u \cdot \dim_H \mu}}{(1-\frac{\epsilon}{\dim_H\mu})^2}$, then for almost all $z\in \mathcal{M}$ \[\frac{\mu[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)]} {\mu(B_r(z))}\precsim_{z} r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u-2\dim_H \mu -2\epsilon}{2}}\to 0.\] \end{lemma} \begin{proof} By the Assumption \ref{geoassumption} for almost all $z\in \mathcal{M}$ there exists $r_z>0$, such that $\forall r<r_{z}$ \[r^{\dim_H\mu+\epsilon}\le \mu(B_r(z)) \le r^{\dim_H\mu-\epsilon}.\] Now, in view of the Proposition \ref{coronaestimate} and the choice of $\epsilon$, we have \[(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u-2\dim_H \mu -2\epsilon>0,\] \[\frac{\mu[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)] }{\mu(B_r(z))}\precsim_{z, \dim \gamma^u} r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u-2\dim_H \mu -2\epsilon}{2}}.\] \end{proof} By that we finished the estimation of the rates for coronas, and can now conclude a proof of the Theorem \ref{thm}. \subsection*{Convergence rates $a>0$ in $d_{TV}(N^{r,T,z},P)\precsim_{T,z}r^a$.} \begin{lemma}\label{lastproofs} If $\alpha> \frac{2}{\dim \gamma^u}-\frac{1}{\dim_H\mu}$ and $\epsilon<\min\{\frac{\min\{\dim \gamma^u, \dim_H \mu\}}{24},\frac{1}{3\dim_H\mu}\}$ are such that \[\alpha >\max \{\frac{\frac{1}{\dim_H\mu}}{(1-\frac{\epsilon}{\dim_H\mu})^2}, \frac{\frac{2}{\dim \gamma^u}-\frac{1}{\dim_H \mu}+\frac{2\epsilon}{\dim \gamma^u \cdot \dim_H \mu}}{(1-\frac{\epsilon}{\dim_H\mu})^2}\},\] then from the Proposition \ref{rate}, Lemma \ref{lastcoronarate} and Lemma \ref{lastshortreturnrate} we have \[a:=\min \{ \frac{(\dim_H\mu-\epsilon)^2(\xi-1)}{\dim_H\mu}, \frac{\epsilon\cdot (\dim_H\mu-\epsilon)}{\dim_H\mu}, \frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u-2\dim_H \mu -2\epsilon}{2},\] \[\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3, \frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{12}-2\epsilon\}\] \[=\min \{ \frac{(\dim_H\mu-\epsilon)^2(\xi-1)}{\dim_H\mu}, \frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u-2\dim_H \mu -2\epsilon}{2},\] \[\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3, \frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{12}-2\epsilon\},\] where the last equality comes from the relation $\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3 \le \frac{\epsilon\cdot (\dim_H\mu-\epsilon)}{\dim_H\mu}.$ If $\frac{d\mu}{d\Leb_{\mathcal{M}}}\in L^{\infty}_{loc}(\mathcal{M})$ and $\epsilon<\min\{\frac{\min\{\dim \gamma^u, \dim_H \mu\}}{24},\frac{1}{3\dim_H\mu}\}$ are such that $\alpha >\frac{\frac{1}{\dim_H\mu}}{(1-\frac{\epsilon}{\dim_H\mu})^2}$, then \[\frac{\mu(B_{r+C \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}}}(z))}{\mu(B_r(z))}\precsim_{z} r^{\frac{(\dim_H \mu-\epsilon)^2\cdot \alpha}{\dim_H\mu}-1},\] and from the Proposition \ref{rate} and Lemma \ref{lastshortreturnrate} we have \[a:=\min \{\frac{(\dim_H\mu-\epsilon)^2(\xi-1)}{\dim_H\mu}, \frac{\epsilon\cdot (\dim_H\mu-\epsilon)}{\dim_H\mu},\frac{(\dim_H \mu-\epsilon)^2\cdot \alpha}{\dim_H\mu}-1, \frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3,\] \[\frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{12}-2\epsilon \}=\min \{\frac{(\dim_H\mu-\epsilon)^2(\xi-1)}{\dim_H\mu}, \frac{(\dim_H \mu-\epsilon)^2\cdot \alpha}{\dim_H\mu}-1,\] \[\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3, \frac{\min\{{\dim \gamma^u}, \dim_H \mu\}}{12}-2\epsilon \},\] where again the last equality holds because $\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3 \le \frac{\epsilon\cdot (\dim_H\mu-\epsilon)}{\dim_H\mu}$. \end{lemma} This completes a proof of the Theorem \ref{thm}. \section{Proof of the Theorem \ref{thm2}}\label{provethm2} The scheme of a proof of the Theorem \ref{thm2} is analogous to the one of the Theorem \ref{thm}, i.e. it contains estimates of the short returns and coronas. But the ways to establish these estimates are the other ones, because of the differences of assumptions in the Theorem \ref{thm2} and in the Theorem \ref{thm}. \subsection{Properties of the First Return Map $f^{\overline{R}}$} \begin{lemma}[Properties of the first returns]\ \par The map $f^{\overline{R}}$ is ergodic with respect to the probability $\mu_U:=\frac{\mu|_{U}}{\mu(U)}$, and it is bijective on $U \bigcap \bigcup_{i \ge 1} \bigcup_{0 \le j < R_i} f^j(\Lambda_i)$. \end{lemma} \begin{proof} A proof that $f^{\overline{R}}$ is one-to-one is the same as in the Lemma \ref{inducemapbi}, which uses the first return $\overline{R}$ and replaces $R_i,R_j$ by $\overline{R}(x), \overline{R}(y)$. Clearly, the map $f^{\overline{R}}$ is ergodic due to the exponential decay of correlations, and it is also onto due to ergodicity. \end{proof} Now, since $\mu\{\interior{(U)}\}>0$, then by the Birkhoff's Ergodic Theorem for almost every $z \in \bigcup_{i \ge 1} \bigcup_{0 \le j < R_i} f^j(\Lambda_i)$ we have $z=f^{j_z}(z')$ for some $z' \in \interior{(U)}$ and $j_z \in \mathbb{N}$. Analogously to the proof of Lemma \ref{ballinhorseshoe}, we obtain \begin{lemma}[Pull metric balls back to $U$]\ \label{ballinhorseshoe1}\ \par For almost every $z \in \bigcup_{i \ge 1} \bigcup_{0 \le j < R_i} f^j(\Lambda_i)$ there exists $j_z \in \mathbb{N}$, such that $f^{-j_z}B_{r}(z) \subseteq U$ $\mu$-almost surely, i.e. \[\mu(f^{-j_z}B_{r}(z) \bigcap U^c)=0.\] \end{lemma} \subsection{Short returns} Let $z' \in \interior{(U)}$. Take now any fixed positive integer $M>0$, a sufficiently small constant $\epsilon'>0$ such that $ n^{\epsilon'}\ll p$ and $B_{M\cdot r} (z') \subseteq \interior{(U)}$ almost surely, where $n=\lfloor \frac{T}{\mu(B_r(z))}\rfloor, p=\lfloor {\lfloor \frac{T}{\mu(B_r(z))}\rfloor}^{\frac{\text{dim}_H\mu-\epsilon}{\text{dim}_H\mu}}\rfloor$, and the same $\epsilon$ as in the Proposition \ref{rate}. We now consider short returns for the induced map $f^{\overline{R}}:U\to U$, namely \begin{equation}\label{sumhorshose1} \int_{B_{M\cdot r} (z')} 1_{\bigcup_{1\le k \le p}(f^{\overline{R}})^{-k}B_{M\cdot r} (z')}d\mu_{U}=\int_{B_{M\cdot r} (z')} 1_{\bigcup_{1\le k \le N}(f^{\overline{R}})^{-k}B_{ M\cdot r} (z')}d\mu_{U} \end{equation} \[+\int_{B_{M\cdot r} (z')} 1_{\bigcup_{N\le k \le n^{\epsilon'}}(f^{\overline{R}})^{-k}B_{ M\cdot r} (z')}d\mu_{U}+\int_{B_{ M\cdot r} (z')} 1_{\bigcup_{n^{\epsilon'}\le k \le p}(f^{\overline{R}})^{-k}B_{ M\cdot r} (z')}d\mu_{U},\] where $N= \lfloor\frac{-\log 2C }{\log \beta}\rfloor+1$, and $C, \beta$ are the ones from the Definition \ref{mpartition}. \begin{lemma}[Short fixed-length returns]\label{fixreturn1}\ \par For almost every $z' \in \interior{(U)} $ and sufficiently small $r_{N,M, z'}>0$ we have $\forall r < r_{N,M,z'}$ \[\int_{B_{M\cdot r} (z')} 1_{\bigcup_{1\le k \le N}(f^{\overline{R}})^{-k}B_{ M\cdot r} (z')}d\mu_{U}=0.\] (Actually, a stronger result will be proved, i.e. $\forall k\in \mathbb{N}$ a map $f^{\overline{R}^k}$ is a local diffeomorphism at almost every point $z' \in \interior{U}$). \end{lemma} \begin{proof} According to the Definition \ref{mpartition}, we have $\mu(\partial U)=0$. Also, the exponential decay of correlations property holds for $f^{\overline{R}}$. Let $A_{per}$ be a set of all periodic points for $f^{\overline{R}}$. Then $\mu(A_{per})=0$ and $\mu(\bigcup_{n \in \mathbb{Z}}f^{-n}\partial U)=0$. Choose $z'\notin \bigcup_{n \in \mathbb{Z}}f^{-n}\partial U \bigcup A_{per}$. The rest of the proof is exactly the same as in the Lemma \ref{fixreturn}, replacing $R$ by $\overline{R}$ and $\Lambda$ by $U$. \end{proof} Before estimating the super-short returns $\int_{B_{M\cdot r} (z')} 1_{\bigcup_{N\le k \le n^{\epsilon'}}(f^{\overline{R}})^{-k}B_{ M\cdot r} (z')}d\mu_{U}$ we will need one more lemma. \begin{lemma}[Recurrences]\label{recurrence1}\ \par There exists $r_M>0$, such that $ \forall r < r_M$ and $i\ge N$ \[\mu_U \{z'\in U: d((f^{\overline{R}})^{i}z',z') \le M \cdot r\}\precsim_{\dim \gamma^u} (M \cdot r)^{\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}},\] where a constant in $\precsim_{\dim \gamma^u}$ depends upon $\dim \gamma^u$, but it does not depend upon $i\ge {N}$ and $\gamma^u \in \Theta$. \end{lemma} \begin{proof} For any $\gamma^u \in \Theta, z_1',z_2' \in \{z'\in U\bigcap \gamma^u: d((f^{\overline{R}})^{-i}z',z')\le M \cdot r\} $ we have \[d((f^{\overline{R}})^{-i}z'_1,z_1') \le M \cdot r,\text{ } d((f^{\overline{R}})^{-i}z'_2,z_2') \le M \cdot r.\] The u-contraction (see the Definition \ref{mpartition}), together with $i \ge N$, give that $C \cdot \beta^i \le C \cdot \beta^{N}< \frac{1}{2}$ and \[d(z'_1,z'_2)\le d(z'_1, (f^{\overline{R}})^{-i}z'_1)+d((f^{\overline{R}})^{-i}z'_1, (f^{\overline{R}})^{-i}z'_2)+d((f^{\overline{R}})^{-i}z'_2, z'_2)\] \[\le 2M \cdot r+ C \cdot \beta^i \cdot d(z'_1,z'_2) \le 2M \cdot r +\frac{1}{2} d(z'_1,z'_2).\] So $d(z'_1,z'_2) \le 4M \cdot r \implies \diam \{z'\in U \bigcap \gamma^u: d((f^{\overline{R}})^{-i}z',z') \le M \cdot r\} \le 4M \cdot r$. Since each $\gamma^u\in \Theta$ has uniformly bounded sectional curvature, then $\Leb_{\gamma^u}(B_{4M\cdot r}(z')) \precsim (4M \cdot r)^{\dim \gamma^u}$ for any $z' \in \gamma^u$ and \[\Leb_{\gamma^u}\{z'\in U: d((f^{\overline{R}})^{-i}z',z') \le M \cdot r\} \precsim_{\dim \gamma^u} (M \cdot r)^{\dim \gamma^u},\] where a constant in $\precsim_{\dim \gamma^u}$ depends only on $\dim \gamma^u$. From the Definition \ref{mpartition} we get $r_M:=\frac{1}{M}$ such that $\forall r<r_M$, $\mu_U \{x \in U: |\gamma^u(x)|<(M \cdot r)^{\frac{\dim \gamma^u}{b+\dim \gamma^u}}\}\le C \cdot (M \cdot r)^{\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}}$, and for any $y \in\gamma^u \in \Theta$, $\frac{d\mu_{\gamma^u}}{d\Leb_{\gamma^u}}(y)=C^{\pm} \cdot \frac{1}{\Leb_{\gamma^u} (\gamma^u)}$. Hence \[\mu_U \{z'\in U: d((f^{\overline{R}})^{-i}z',z') \le M \cdot r\}=\int \mu_{\gamma^u(x)}\{z'\in U: d((f^{\overline{R}})^{-i}z',z') \le M \cdot r\} d\mu_U(x)\] \[=\int_{|\gamma^u(x)|\le (M \cdot r)^{\frac{\dim \gamma^u}{b+\dim \gamma^u}}} \mu_{\gamma^u(x)}\{z'\in U: d((f^{\overline{R}})^{-i}z',z') \le M \cdot r\} d\mu_U(x)\] \[+\int_{|\gamma^u(x)|\ge (M \cdot r)^{\frac{\dim \gamma^u}{b+\dim \gamma^u}}} \mu_{\gamma^u(x)}\{z'\in U: d((f^{\overline{R}})^{-i}z',z') \le M \cdot r\} d\mu_U(x) \] \[\le C \cdot (M \cdot r)^{\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}} +\int_{|\gamma^u(x)|\ge (M \cdot r)^{\frac{\dim \gamma^u}{b+\dim \gamma^u}}} \mu_{\gamma^u(x)}\{z'\in U: d((f^{\overline{R}})^{-i}z',z') \le M \cdot r\} d\mu_U(x) \] \[\precsim_{\dim \gamma^u} (M \cdot r)^{\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}}+\int_{|\gamma^u(x)|\ge (M \cdot r)^{\frac{\dim \gamma^u}{b+\dim \gamma^u}}} \frac{\Leb_{\gamma^u(x)}\{z'\in U: d((f^{\overline{R}})^{-i}z',z') \le M \cdot r\}}{\Leb_{\gamma^u(x)} (\gamma^u(x))} d\mu_U(x)\] \[\precsim_{\dim \gamma^u} (M \cdot r)^{\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}}+ (M\cdot r)^{\dim \gamma^u-\dim \gamma^u\cdot {\frac{\dim \gamma^u}{b+\dim \gamma^u}}} \precsim (M \cdot r)^{\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}},\] where in the last two inequalities we used that the sectional curvature of $\gamma^u(x) \in \Theta$ is uniformly bounded, and $\dim \gamma^u-\dim \gamma^u\cdot {\frac{\dim \gamma^u}{b+\dim \gamma^u}}=\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}$. Finally, $(f^{\overline{R}})_{*} \mu_U=\mu_U$ implies that \[\mu_U \{z'\in U: d((f^{\overline{R}})^{i}z',z') \le M \cdot r\}\precsim_{\dim \gamma^u} (M \cdot r)^{\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}}.\] \end{proof} \begin{lemma}[Super-short returns]\label{veryshortreturn1}\ \par For almost every $z' \in \interior{(U)}, z \in \mathcal{M}$ there exist $r_{M,z'}>0, r_z>0$, such that $\forall r< \min \{r_{z',M}, r_z\}$ \[\int_{B_{M\cdot r} (z')} 1_{\bigcup_{N\le k \le n^{\epsilon'}}(f^{\overline{R}})^{-k}B_{ M\cdot r} (z')}d\mu_{U}\precsim_{T,b} (M \cdot r)^{\dim_H\mu -\epsilon} \cdot M^{\frac{\min\{{\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}}, \dim_H \mu\}}{6}}\cdot r^{\frac{\min\{{\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}}, \dim_H \mu\}}{12}},\] where $\epsilon>0$ is the same as in the Proposition \ref{rate} and $\epsilon'<\frac{\min\{\dim_H \mu, {\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}}\}}{12 \dim_H \mu+12 \epsilon}$. \end{lemma} \begin{proof} In the Lemma \ref{recurrence1} we already proved that \[\mu_U \{z'\in U: d((f^{\overline{R}})^{i}z',z') \le M \cdot r\}\precsim_{\dim \gamma^u} (M \cdot r)^{\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}}.\] The rest of the proof is exactly the same as in the Proposition \ref{veryshortreturn}, where one should replace $\dim \gamma^u, \Lambda, R$ by $\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}, U, \overline{R}$. \end{proof} \begin{lemma}[Not super-short, but short returns]\label{notveryshort1}\ \par Let $n= \lfloor \frac{T}{\mu(B_r(z))}\rfloor,p=\lfloor {\lfloor \frac{T}{\mu(B_r(z))}\rfloor}^{\frac{\text{dim}_H\mu-\epsilon}{\text{dim}_H\mu}}\rfloor$ and $n^{\epsilon'}\ll p$. Then for almost all $z'\in \interior{(U)}$ and $z\in \mathcal{M}$ there is $r_{z,z',M}>0$, such that $\forall r< r_{z,z',M}$ \[\int_{B_{ M\cdot r} (z')} 1_{\bigcup_{n^{\epsilon'}\le k \le p}(f^R)^{-k}B_{ M\cdot r} (z')}d\mu_{U}\] \[\precsim_{T,\epsilon} (\frac{1}{r^{\dim_H\mu+\epsilon}})^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \cdot [(M\cdot r)^{2\dim_H\mu-2\epsilon^3}+\mu_U(B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z')\setminus B_{M \cdot r}(z'))],\] where $\beta\in (0,1) $ is the same as in the Definition \ref{mpartition}. \end{lemma} \begin{proof} The approach to proving the required estimate is quite standard. It uses Lipschitz functions to approximate $B_{M \cdot r}(z')$. However, we will write it down for completeness. Let $L\in \mbox{\rm Lip}(U)$, such that $L=1$ on ${B_{M \cdot r}(z')}$, $L=0 $ on $ {B^c_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z')} $, and $L$ is linear on $ B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z') \setminus B_{M \cdot r}(z')$, where $\epsilon$ and $\alpha$ are the same as in the Proposition \ref{rate}. Hence $L \in \mbox{\rm Lip}(U)$ with a Lipschitz constant $(\sqrt[4]{\beta})^{-T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}$. Therefore for any $k \in [n^{\epsilon'}, p]$ \[\int_{B_{ M\cdot r} (z')} 1_{B_{ M\cdot r} (z')} \circ (f^R)^{k}d\mu_{U} \le \int L \cdot L \circ (f^R)^{k}d\mu_{U}+2 \mu_U( B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z') \setminus B_{M \cdot r}(z')) \] \[\le |\int L \cdot L \circ (f^R)^{k}d\mu_{U}-\int L d\mu_U \cdot \int L \circ (f^R)^{k}d\mu_{U}|+\int L d\mu_U \cdot \int L \circ (f^R)^{k}d\mu_{U}\] \[+2 \mu_U(B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z')\setminus B_{M \cdot r}(z')).\] Now, by making use of the exponential decay of correlation in the Definition \ref{mpartition}, we can continue the estimate as \[\precsim \mbox{\rm Lip}(L)^2 \cdot \beta^{n^{\epsilon'}}+\mu_U(B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z'))^2\] \[+2 \mu_U(B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z')\setminus B_{M \cdot r}(z')) \le (\sqrt[4]{\beta})^{-T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}\cdot \beta^{n^{\epsilon'}}\] \[+\mu_U(B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z'))^2+2 \mu_U(B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z')\setminus B_{M \cdot r}(z')).\] Hence \[\int_{B_{ M \cdot r} (z')} 1_{\bigcup_{n^{\epsilon'}\le k \le p}(f^R)^{-k}B_{ M\cdot r} (z')}d\mu_{U} \precsim p \cdot (\sqrt{\beta})^{-T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}\cdot \beta^{n^{\epsilon'}}\] \begin{equation}\label{9} +p \cdot \mu_U(B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z'))^2+2 p \cdot \mu_U(B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z')\setminus B_{M \cdot r}(z')). \end{equation} Choose now from the Assumption \ref{geoassumption} the same $\epsilon>0$ as in the Proposition \ref{rate}. Then, for almost all $z\in \mathcal{M}$ and $ z'\in \interior{(U)}$, there exists such $r_{z,z',M}>0$, that $\forall r< r_{z,z',M}$ \[B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z') \subseteq U, \frac{T}{r^{\dim_H\mu-\epsilon}}\precsim n \precsim \frac{T}{r^{\dim_H\mu+\epsilon}},\] \[r^{\dim_H\mu+\epsilon}\le \mu(B_r(z)) \le r^{\dim_H\mu-\epsilon}, \frac{(M \cdot r)^{\dim_H\mu+\epsilon^3}}{\mu(U)}\le \mu_{U}(B_{M \cdot r}(z')) \le \frac{(M\cdot r)^{\dim_H\mu-\epsilon^3}}{\mu(U) },\] \[ (\frac{T}{r^{\dim_H\mu-\epsilon}})^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \precsim p\precsim (\frac{T}{r^{\dim_H\mu+\epsilon}})^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu}},(\frac{T}{r^{\dim_H\mu-\epsilon}})^{\epsilon'}\precsim n^{\epsilon'} \precsim (\frac{T}{r^{\dim_H\mu+\epsilon}})^{\epsilon'}.\] Therefore, the estimates from $(\ref{9})$ can be continued as \[\precsim (\frac{T}{r^{\dim_H\mu+\epsilon}})^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \cdot (\sqrt{\beta})^{-T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}} \cdot \beta^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}+(\frac{T}{r^{\dim_H\mu+\epsilon}})^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu}}\] \[\times [\mu_U(B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z'))^2+\mu_U(B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z')\setminus B_{M \cdot r}(z'))]\] \[\precsim_{T,\epsilon} (\frac{1}{r^{\dim_H\mu+\epsilon}})^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \cdot [(M\cdot r)^{2\dim_H\mu-2\epsilon^3}+\mu_U(B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z')\setminus B_{M \cdot r}(z'))].\] The last inequality holds because $\beta^{r^{-c}} \precsim r^{c'}$ for any $c',c>0$. \end{proof} Combining now the Lemmas \ref{fixreturn1}, \ref{veryshortreturn1}, \ref{notveryshort1}, and arguing as in the Proposition \ref{shortreturnrate}, we can formulate the following summary of the obtained results. \begin{proposition}[Short returns rates]\label{shortreturnrate1}\ \par Let $\epsilon, \epsilon'>0$ satisfy the relations $\alpha \cdot (\dim_H\mu-\epsilon)>\frac{\dim_H\mu}{\dim_H\mu-\epsilon}>1$, $p=\lfloor {\lfloor \frac{T}{\mu(B_r(z))}\rfloor}^{\frac{\text{dim}_H\mu-\epsilon}{\text{dim}_H\mu}}\rfloor$, $\epsilon'<\min\{\frac{\min\{{\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}}, \dim_H \mu\}}{12\dim_H \mu+12\epsilon} , \frac{\dim_H\mu-\epsilon}{\dim_H\mu}\}$. Then for almost all $z\in \mathcal{M}$, $z'\in \interior{(U)}$ and for each integer $M>0$ there exists a small enough $r_{z,z',M}>0$, such that $\forall r< r_{z,z',M}$ \[\int_{B_{M\cdot r} (z')} 1_{\bigcup_{1\le k \le p}(f^{\overline{R}})^{-k}B_{M\cdot r} (z')}d\mu_{U}\precsim_{\epsilon, T,b} (M \cdot r)^{\dim_H\mu -\epsilon}\cdot M^{\frac{\min\{{\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}}, \dim_H \mu\}}{6}}\cdot r^{\frac{\min\{{\frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}}, \dim_H \mu\}}{12}}\] \[+^{-(\dim_H\mu+\epsilon) \cdot \frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \cdot [(M\cdot r)^{2\dim_H\mu-2\epsilon^3}+\mu_U(B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z')\setminus B_{M \cdot r}(z'))],\] where $\beta \in (0,1)$ is the same as in the Definition \ref{mpartition}. \end{proposition} \subsection{Coronas} \begin{proposition}[Coronas in $\mathcal{M}$ and $U$]\label{coronaestimate1}\ \par For almost every $z \in \mathcal{M}$ and $z' \in \interior{(U)}$ there exist $r_z, r_{z',M}>0$, such that $\forall r< \min \{r_z, r_{z',M}\} $ the following estimate holds for the corona in $\mathcal{M}$ \[\mu\{B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)\} \precsim_{z, \dim \gamma^u} r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u \cdot b}{2(b+\dim \gamma^u)}},\] and the estimate for the corona in $U$ is \[\mu_U[B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z')\setminus B_{M \cdot r}(z')]\precsim_{z, \dim \gamma^u} (M\cdot r)^{\frac{\dim \gamma^u \cdot b }{2(b+\dim \gamma^u)}} \cdot \beta^{ \frac{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}\cdot\dim \gamma^u \cdot b}{8(b+\dim \gamma^u)}}.\] \end{proposition} \begin{proof} For corona in $\mathcal{M}$ we have from the Lemma \ref{ballinhorseshoe1} for almost all $z \in \bigcup_{i \ge 1} \bigcup_{0 \le j < R_i} f^j(\Lambda_i)$ that $z=f^{j_z}(z')$ for some $z' \in \interior{(U)}$. Moreover, there exists $r'_z>0$, such that $\forall r< r'_z$ \[\mu(f^{-j_z}[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)] \bigcap U^c)=0.\] Since $f_* \mu=\mu$ and $f$ is bijective on $\bigcup_{i \ge 1} \bigcup_{0 \le j < R_i} f^j(\Lambda_i)$, then \[\mu([B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)] \bigcap f^{j_z} U^c)=0.\] Therefore \[[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)] \subseteq f^{j_z}U \text{ almost surely}.\] The rest of the proof is exactly the same as in the Proposition \ref{coronaestimate}, where one should replace $\gamma^u\in \Gamma^u$ by $\gamma^u \in \Theta$, and then use that the sectional curvatures of all $\gamma^u \in \Theta$ are uniformly bounded. Then there exists $r_z>0$ such that $\forall r< r_z$ \[\Leb_{\gamma^u} \{f^{-j_z}[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)]\} \precsim_{z, \dim \gamma^u} r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u}{2}}.\] Similarly to the argument in a proof in the Lemma \ref{recurrence1}, we have \[\mu\{f^{-j_z}[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)]\}\] \[=\int \mu_{\gamma^u(x)}\{f^{-j_z}[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)]\}d\mu(x)\] \[=\int_{|\gamma^u(x)|\ge r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u}{2(b+\dim \gamma^u)}}} \mu_{\gamma^u(x)}\{f^{-j_z}[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)]\} d\mu(x) \] \[+\int_{|\gamma^u(x)|\le r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u}{2(b+\dim \gamma^u)}}} \mu_{\gamma^u(x)}\{f^{-j_z}[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)]\} d\mu(x) \] \[\precsim_{z, \dim \gamma^u} r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u \cdot b}{2(b+\dim \gamma^u)}}+r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u}{2}} \cdot r^{-\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u \cdot \dim \gamma^u}{2(b+\dim \gamma^u)}}\] \[\precsim_{z, \dim \gamma^u} r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u \cdot b}{2(b+\dim \gamma^u)}}.\] Then, using that $f_*\mu=\mu$, we have \[\mu\{B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)\} \precsim_{z, \dim \gamma^u} r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \dim \gamma^u \cdot b}{2(b+\dim \gamma^u)}}.\] For corona in $U$ there is such $r_{z', M}\in (0, \frac{1}{M})$, that $\forall r<r_{z',M}$ \[B_{M \cdot r+(\sqrt{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z') \subseteq U,\] and, using the same argument as in the Lemma \ref{recurrence1}, we have \[\mu_U[B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z')\setminus B_{M \cdot r}(z')]\] \[=\int_{|\gamma^u(x)|\ge (M\cdot r)^{\frac{\dim \gamma^u }{2(b+\dim \gamma^u)}} \cdot \beta^{ \frac{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}\cdot\dim \gamma^u}{8(b+\dim \gamma^u)}}} \mu_{\gamma^u(x)}\{B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z')\setminus B_{M \cdot r}(z')\} d\mu(x) \] \[+\int_{|\gamma^u(x)|\le (M\cdot r)^{\frac{\dim \gamma^u }{2(b+\dim \gamma^u)}} \cdot \beta^{ \frac{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}\cdot\dim \gamma^u}{8(b+\dim \gamma^u)}}} \mu_{\gamma^u(x)}\{B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z')\setminus B_{M \cdot r}(z')\} d\mu(x) \] \[\precsim_{z, \dim \gamma^u} (M\cdot r)^{\frac{\dim \gamma^u \cdot b }{2(b+\dim \gamma^u)}} \cdot \beta^{ \frac{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}\cdot\dim \gamma^u \cdot b}{8(b+\dim \gamma^u)}}. \] \end{proof} \subsection{Conclusion of the Proof of the Theorem \ref{thm2}} We will start with a rough estimate of $\dim_H \mu$. \subsection*{Proof that $\dim_H\mu \ge \frac{b}{b+\dim \gamma^u} \cdot \dim {\gamma}^u$:} \begin{lemma}\label{roughdim1} It follows from the Assumptions \ref{geoassumption} and the Definition \ref{mpartition} that $\dim_H\mu \ge \frac{b}{b+\dim \gamma^u} \cdot \dim {\gamma}^u$. \end{lemma} \begin{proof} Due to the ergodicity of $\mu$, it is enough to consider $z' \in \interior{(U)}$ and a ball $B_r(z')$ in $U$. By the same trick as in the Proposition \ref{coronaestimate1} we get for any $\gamma^u \in \Theta$ \[\Leb_{\gamma^u}(B_r(z')) \precsim_{z',\dim \gamma^u}r^{\dim \gamma^u} \text{ and } \mu(B_r(z')) \precsim_{z',\dim \gamma^u}r^{\frac{\dim \gamma^u \cdot b}{b + \dim \gamma^u}}.\] On the other hand, by the Assumption \ref{geoassumption} for almost all $z' \in \mathcal{M}$ and any $m \in \mathbb{N}$ there is $r_{z',m}>0$ such that $\mu(B_r(z')) \ge r^{\dim_H \mu+ \frac{1}{m}}$ for any $r<r_{z',m}$. This implies that $\dim \gamma^u \cdot \frac{b}{b+\dim \gamma^u} \le \dim_H \mu + \frac{1}{m}, \forall m \in \mathbb{N}$. By taking the limit $m\to \infty$ one gets $\dim_H\mu \ge \dim \gamma^u \cdot \frac{b}{b+\dim \gamma^u}$. \end{proof} We will obtain now the estimates of convergence rates in the Theorem \ref{thm2}. According to the Proposition \ref{rate} it is enough to find convergence rates for short returns and coronas. A proof will consist of several steps. \subsection*{Pull back short returns $\int_{B_r(z)} 1_{\bigcup_{1\le j \le p}f^{-j}B_r(z)}d\mu$} For almost any $z \in \bigcup_{i \ge 1} \bigcup_{0 \le j < R_i} f^j(\Lambda_i)$ we have $z=f^{j_z}(z')$ for some $z' \in \interior{(U)}$. By Lemma \ref{comparetopballmetricball}, there exist a topological ball $TB_r(z')=TB_r(f^{-j_z}(z))$ and a constant $C_z>1$ such that $B_{C^{-1}_z\cdot r} (f^{-j_z}z) \subseteq TB_r(f^{-j_z}z) \subseteq B_{C_z\cdot r} (f^{-j_z}z)$. \begin{lemma}[Pull short returns back to $U$]\label{pullbackhorseshoe1}\ \par There exists a small enough $r_z>0$, such that $\forall r< r_z$ the following inequality holds \[ \int_{B_r(z)} 1_{\bigcup_{1\le k \le p}f^{-k}B_r(z)}d\mu \le \mu(U) \cdot \int_{B_{C_z\cdot r} (f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}(f^{\overline{R}})^{-k}B_{C_z\cdot r} (f^{-j_z}z)}d\mu_{U} .\] \end{lemma} \begin{proof} It follows from the invariance of $\mu$ (i.e. $f_*\mu=\mu$) that \[ \int_{B_r(z)} 1_{\bigcup_{1\le k \le p}f^{-k}B_r(z)}d\mu=\int_{TB_r(f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}f^{-k}TB_r(f^{-j_z}z)}d\mu.\] By Lemma \ref{ballinhorseshoe1}, we have $TB_r(f^{-j_z}z) \subseteq U$ for any $ r< r_z$ if $r_z>0$ is small enough. The rest of the proof is the same as in Lemma \ref{pullbackhorseshoe}, where one should use the first return map $f^{\overline{R}}$. \end{proof} \subsection*{Estimate of $\frac{1}{\mu(B_r(z))}\cdot \int_{B_{C_z\cdot r} (f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}(f^{\overline{R}})^{-k}B_{C_z\cdot r} (f^{-j_z}z)}d\mu_{U}$.} \begin{lemma}\label{afterpullback1} For $\epsilon<\min\{\frac{\min\{\frac{b\cdot \dim \gamma^u}{b+\dim \gamma^u}, \dim_H \mu\}}{24}, \frac{1}{3\dim_H\mu}\}$, satisfying $\alpha \cdot (\dim_H\mu-\epsilon)>\frac{\dim_H\mu}{\dim_H\mu-\epsilon}>1$, and for almost every $z\in \mathcal{M}$, there exists $r_z>0$, such that $\forall r<r_z$ \[\frac{1}{\mu(B_r(z))}\cdot \int_{B_{C_z\cdot r} (f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}(f^{\overline{R}})^{-k}B_{C_z\cdot r} (f^{-j_z}z)}d\mu_{U}\precsim_{T,z,\epsilon} r^{\frac{\min\{\frac{b\cdot \dim \gamma^u}{b+\dim \gamma^u}, \dim_H \mu\}}{12}-2\epsilon}+r^{\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3}.\] \end{lemma} \begin{proof} Note that $C_z < \infty $ for almost all $z\in \mathcal{M}$. Denote $F_1:=\{C_z<\infty\}$, $F_2 \times F_3:=\{(z,z')\in \mathcal{M}\times \interior{(U)}: \text{ the Propositions }\ref{shortreturnrate1}, \ref{coronaestimate1} \text{ hold}\}$. Define a new measure one subset in $\mathcal{M}$ as $F:=F_1 \bigcap F_2 \bigcap \{\bigcup_{j\ge 0}f^j(\interior{(U)}\bigcap F^c_3 )\}^c$. Then for any $z \in F$ we have $ C_z < \infty$ and $(z,f^{-j_z}z)\in F_2 \times F_3$. Let now $M:=\lfloor C_z \rfloor+1, z'=f^{-j_z}z $. By the Proposition \ref{shortreturnrate1}, there exists $r_{z,z',M}=r_{z,f^{-j_z}z,\lfloor C_z \rfloor+1}>0$, such that $\forall r< r_{z,f^{-j_z}z,\lfloor C_z \rfloor+1}$ \[\int_{B_{C_z\cdot r} (f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}(f^{\overline{R}})^{-k}B_{C_z\cdot r} (f^{-j_z}z)}d\mu_{U}\le \int_{B_{M\cdot r} (z')} 1_{\bigcup_{1\le k \le p}(f^{\overline{R}})^{-k}B_{M\cdot r} (z')}d\mu_{U}\] \[\precsim_{T,b,\epsilon} (M \cdot r)^{\dim_H\mu -\epsilon} \cdot M^{\frac{\min\{\frac{b\cdot \dim \gamma^u}{b+\dim \gamma^u}, \dim_H \mu\}}{6}}\cdot r^{\frac{\min\{\frac{b\cdot \dim \gamma^u}{b+\dim \gamma^u}, \dim_H \mu\}}{12}}\] \[+(\frac{1}{r^{\dim_H\mu+\epsilon}})^{\frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \cdot [(M\cdot r)^{2\dim_H\mu-2\epsilon^3}+\mu_U(B_{M \cdot r+(\sqrt[4]{\beta})^{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}}}(z')\setminus B_{M \cdot r}(z'))].\] Next, by the Proposition \ref{coronaestimate1}, the right hand side of the inequality above can be estimated as \[\precsim_{T,C_z,z, \epsilon} (M \cdot r)^{\dim_H\mu -\epsilon} \cdot r^{\frac{\min\{\frac{b\cdot \dim \gamma^u}{b+\dim \gamma^u}, \dim_H \mu\}}{12}}+r^{-(\dim_H\mu+\epsilon) \cdot \frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \cdot r^{2\dim_H\mu-2\epsilon^3}\] \[+r^{-(\dim_H\mu+\epsilon) \cdot \frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \cdot (M\cdot r)^{\frac{\dim \gamma^u \cdot b }{2(b+\dim \gamma^u)}} \cdot \beta^{ \frac{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}\cdot\dim \gamma^u \cdot b}{8(b+\dim \gamma^u)}}.\] Now, by the Assumption \ref{geoassumption} for almost all $z\in \mathcal{M}$ there exists $r_z>0$, such that $\forall r<r_{z}$ \[r^{\dim_H\mu+\epsilon^3}\le \mu(B_r(z)) \le r^{\dim_H\mu-\epsilon^3}.\] Then $\forall r<\min \{ r_{z,f^{-j_z}z,\lfloor C_z \rfloor+1}, r_z\}$ \[\frac{1}{\mu(B_r(z))}\cdot \int_{B_{C_z\cdot r} (f^{-j_z}z)} 1_{\bigcup_{1\le k \le p}(f^{\overline{R}})^{-k}B_{C_z\cdot r} (f^{-j_z}z)}d\mu_{\Lambda}\] \[\precsim_{T,z, \epsilon} \frac{1}{r^{\dim_H\mu+\epsilon^3}}\{(M \cdot r)^{\dim_H\mu -\epsilon} \cdot r^{\frac{\min\{\frac{\dim \gamma^u \cdot b }{b+\dim \gamma^u}, \dim_H \mu\}}{12}}+r^{-(\dim_H\mu+\epsilon) \cdot \frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \cdot r^{2\dim_H\mu-2\epsilon^3}\] \[+r^{-(\dim_H\mu+\epsilon) \cdot \frac{\dim_H\mu-\epsilon}{\dim_H\mu}} \cdot (M\cdot r)^{\frac{\dim \gamma^u \cdot b}{2(b+\dim \gamma^u)}} \cdot \beta^{ \frac{T^{\epsilon'}\cdot{r^{-(\dim_H\mu-\epsilon) \epsilon'}}\cdot\dim \gamma^u \cdot b}{8(b+\dim \gamma^u)}}\}\] \[\precsim_{T,z,\epsilon,C_z} r^{\frac{\min\{\frac{\dim \gamma^u \cdot b }{b+\dim \gamma^u}, \dim_H \mu\}}{12}-2\epsilon}+r^{\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3}\to 0,\] The last inequality holds because $\beta_2^{r^{-c}}\ll r^{c'}, \forall c,c'>0.$ \end{proof} By combining the Lemmas \ref{pullbackhorseshoe1} and \ref{afterpullback1}, we get an \subsection*{Estimate of short returns rates $\frac{1}{\mu(B_r(z))}\int_{B_r(z)} 1_{\bigcup_{1\le j \le p}f^{-j}B_r(z)}d\mu$.} \begin{lemma}\label{lastshortreturnrate1}. For $\epsilon<\min \{\frac{\min\{\frac{\dim \gamma^u \cdot b }{b+\dim \gamma^u}, \dim_H \mu\}}{24}, \frac{1}{3\dim_H \mu}\}$, satisfying $\alpha \cdot (\dim_H\mu-\epsilon)>\frac{\dim_H\mu}{\dim_H\mu-\epsilon}>1$, and for almost every $z\in \mathcal{M}$ \[\frac{1}{\mu(B_r(z))}\int_{B_r(z)} 1_{\bigcup_{1\le j \le p}f^{-j}B_r(z)}d\mu \precsim_{T,z,\epsilon} r^{\frac{\min\{\frac{\dim \gamma^u \cdot b }{b+\dim \gamma^u}, \dim_H \mu\}}{12}-2\epsilon}+r^{\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3}\to 0.\] \end{lemma} \subsection*{Estimate of coronas rates $\frac{\mu[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)]}{\mu(B_r(z))}$ }. \begin{lemma}\label{lastcoronarate1}. Let $\alpha>\frac{2}{\dim \gamma^u} \cdot \frac{b+\dim \gamma^u}{b}-\frac{1}{\dim_H\mu}$ and $\epsilon<\min \{\frac{\min\{\frac{\dim \gamma^u \cdot b }{b+\dim \gamma^u}, \dim_H \mu\}}{24},\frac{1}{3\dim_H\mu}\}$ is small enough, so that $\alpha >\frac{(\frac{2}{\dim \gamma^u}+\frac{2\epsilon}{\dim \gamma^u \cdot \dim_H \mu})\cdot \frac{b+\dim \gamma^u}{b}-\frac{1}{\dim_H \mu}}{(1-\frac{\epsilon}{\dim_H\mu})^2}$. Then for almost all $z\in \mathcal{M}$ \[\frac{\mu[B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}} }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} }}(z)]} {\mu(B_r(z))}\precsim_{z} r^{\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \frac{\dim \gamma^u \cdot b }{b+\dim \gamma^u}-2\dim_H \mu -2\epsilon}{2}}\to 0,\] The calculations here are exactly the same as in the Lemma \ref{lastcoronarate}, using the Proposition \ref{coronaestimate1}. Therefore we will not repeat them. \end{lemma} \subsection*{Convergence rates $a>0$ in $d_{TV}(N^{r,T,z},P)\precsim_{T,z}r^a$.} \begin{lemma}\label{lastlastproof}. Let $\alpha> \frac{2}{\dim \gamma^u}-\frac{1}{\dim_H\mu}$ and $\epsilon<\min \{\frac{\min\{\frac{\dim \gamma^u \cdot b }{b+\dim \gamma^u}, \dim_H \mu\}}{24},\frac{1}{3\dim_H\mu}\}$, so that \[\alpha >\max \{\frac{\frac{1}{\dim_H\mu}}{(1-\frac{\epsilon}{\dim_H\mu})^2}, \frac{(\frac{2}{\dim \gamma^u}+\frac{2\epsilon}{\dim \gamma^u \cdot \dim_H \mu})\cdot \frac{b+\dim \gamma^u}{b}-\frac{1}{\dim_H \mu}}{(1-\frac{\epsilon}{\dim_H\mu})^2}\}.\] Now from the Proposition \ref{rate}, Lemma \ref{lastcoronarate1} and Lemma \ref{lastshortreturnrate1} we have \[a:=\min \{ \frac{(\dim_H\mu-\epsilon)^2(\xi-1)}{\dim_H\mu}, \frac{\epsilon\cdot (\dim_H\mu-\epsilon)}{\dim_H\mu}, \frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \frac{b \cdot \dim \gamma^u}{b+\dim \gamma^u}-2\dim_H \mu -2\epsilon}{2},\] \[\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3, \frac{\min\{\frac{\dim \gamma^u \cdot b }{b+\dim \gamma^u}, \dim_H \mu\}}{12}-2\epsilon \}=\min \{ \frac{(\dim_H\mu-\epsilon)^2(\xi-1)}{\dim_H\mu}, \frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3,\] \[\frac{(1+\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}) \cdot \frac{\dim \gamma^u \cdot b}{b+\dim \gamma^u}-2\dim_H \mu -2\epsilon}{2}, \frac{\min\{\frac{\dim \gamma^u \cdot b }{b+\dim \gamma^u}, \dim_H \mu\}}{12}-2\epsilon\},\] where the last equality holds because $\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3 \le \frac{\epsilon\cdot (\dim_H\mu-\epsilon)}{\dim_H\mu}$. If $\frac{d\mu}{d\Leb_{\mathcal{M}}}\in L^{\infty}_{loc}(\mathcal{M})$, then choose $\epsilon<\min \{\frac{\min\{\frac{\dim \gamma^u \cdot b }{b+\dim \gamma^u}, \dim_H \mu\}}{24},\frac{1}{3\dim_H\mu}\}$, so that $\alpha >\frac{\frac{1}{\dim_H\mu}}{(1-\frac{\epsilon}{\dim_H\mu})^2}$. Then \[\frac{\mu(B_{r+C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu} } }(z)\setminus B_{r-C' \cdot r^{\frac{(\dim_H\mu-\epsilon)^2\cdot \alpha }{\dim_H\mu}}}(z))}{\mu(B_r(z))}\precsim_{z} r^{\frac{(\dim_H \mu-\epsilon)^2\cdot \alpha}{\dim_H\mu}-1}.\] Then from the Proposition \ref{rate} and Lemma \ref{lastshortreturnrate1} we have $a:=$ \[\min \{\frac{(\dim_H\mu-\epsilon)^2(\xi-1)}{\dim_H\mu}, \frac{\epsilon\cdot (\dim_H\mu-\epsilon)}{\dim_H\mu},\frac{(\dim_H \mu-\epsilon)^2\cdot \alpha}{\dim_H\mu}-1,\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3, \] \[\frac{\min\{\frac{\dim \gamma^u \cdot b }{b+\dim \gamma^u}, \dim_H \mu\}}{12}-2\epsilon\}=\min \{\frac{(\dim_H\mu-\epsilon)^2(\xi-1)}{\dim_H\mu},\frac{(\dim_H \mu-\epsilon)^2\cdot \alpha}{\dim_H\mu}-1,\] \[\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3, \frac{\min\{\frac{\dim \gamma^u \cdot b }{b+\dim \gamma^u}, \dim_H \mu\}}{12}-2\epsilon\},\] where the last equality holds because $\frac{\epsilon^2}{\dim_H \mu}-3\epsilon^3 \le \frac{\epsilon\cdot (\dim_H\mu-\epsilon)}{\dim_H\mu}.$ \end{lemma} This completes a proof of the Theorem \ref{thm2}. \section{Applications}\label{app} For all classes of dynamical systems, which will be considered in this section, there exist Gibbs-Markov-Young structures (see the Definition \ref{gibbs}) and SRB measures $\mu$. The Assumption \ref{geoassumption}, sometimes except $\alpha \cdot \dim_H\mu>1$, also always holds. Therefore, only the following conditions must be verified: \begin{itemize} \item for the Theorem \ref{thm} \begin{enumerate} \item $R$ is the first return time for $\Lambda$ and $f$, \item there exist constants $\alpha>0$ and $C>0$ such that \[ \sup_{x,y \in \gamma^s \in \Gamma^s, x',y' \in \gamma^u \in \Gamma^u} \{d(f^nx, f^ny), d(f^{-n}x',f^{-n}y')\} \le \frac{C}{n^{\alpha}},\] \item $\mu\{\interior{(\Lambda)}\}>0$ and $\mu(\partial \Lambda)=0$, \item verify, whether or not $\alpha>\frac{2}{\dim \gamma^u}-\frac{1}{\dim_H \mu}> \frac{1}{\dim_H\mu}$, and whether or not $\frac{d\mu}{d\Leb_{\mathcal{M}}}\in L_{loc}^{\infty}(\mathcal{M})$. \end{enumerate} \item for the Theorem \ref{thm2} \begin{enumerate} \item find such a reference area $U \subseteq \mathcal{M}$ that its first return map $f^{\overline{R}}$ has an exponential decay of correlations, \item find a measurable partition $\Theta:=\{\gamma^u(x)\}_{x \in U}$ with required properties, \item check that the estimate $\mu_U \{x \in U: |\gamma^u(x)|< \epsilon\}\le C \cdot \epsilon^b$ holds, \item get an estimate of distortion, i.e. $\frac{d\mu_{\gamma^u(x)}}{d\Leb_{\gamma^u(x)}}(y)=C^{\pm 1} \cdot \frac{d\mu_{\gamma^u(x)}}{d\Leb_{\gamma^u(x)}}(z) \text{ for any }y, z \in \gamma^u(x) \in \Theta$, \item show that there exist constants $\alpha>0$ and $C>0$, such that \[\sup_{x,y \in \gamma^s \in \Gamma^s, x',y' \in \gamma^u \in \Gamma^u} \{d(f^nx, f^ny), d(f^{-n}x',f^{-n}y')\} \le \frac{C}{n^{\alpha}},\] \item verify, whether or not $\alpha>\frac{2}{\dim \gamma^u}\cdot \frac{b+\dim \gamma^u}{b}-\frac{1}{\dim_H \mu}> \frac{1}{\dim_H \mu}$, and whether or not $\frac{d\mu}{d\Leb_{\mathcal{M}}}\in L_{loc}^{\infty}(\mathcal{M})$. \end{enumerate} \end{itemize} All other required conditions hold for the studied below classes of dynamical systems. \subsection{Intermittent solenoids}\label{appsolenoid} Following \cite{pene, Alves} let $\mathcal{M}=S^1 \times \mathbb{D}$, $f_{\gamma}(x,z)=(g_{\gamma}(x), \theta \cdot z+\frac{1}{2}e^{2\pi i x})$, where $g_{\gamma}: S^1 \to S^1$ is a continuous map of degree $d\ge 2$ and $\gamma \in (0, +\infty)$ such that \begin{enumerate} \item $g_{\gamma}$ is $C^2$ on $S^1 \setminus \{0\}$ and $Dg_{\gamma}>1$ on $S^1 \setminus \{0\}$, \item $g_{\gamma}(0)=0, Dg_{\gamma}(0+)=1$ and $xD^2g_{\gamma}(x) \sim x^{\gamma}$ for sufficiently small positive $x$, \item $Dg_{\gamma}(0-)>1$, \item $\theta>0$ is so small that $\theta \cdot ||Dg_{\gamma}||_{\infty}< 1-\theta$. \end{enumerate} It was proved in \cite{Alves} that the SRB probability measure $\mu$ exists iff $\gamma\in (0,1)$, the attractor is $A:=\bigcap_{i\ge 0}f_{\gamma}^i(\mathcal{M}), \xi=\frac{1}{\gamma}>1, \alpha=1+\frac{1}{\gamma}$, $\Lambda=(I \times \mathbb{D}) \bigcap A$, where $I$ is one of the intervals of hyperbolicity, and $f_{\gamma}: I \to S^1$ is a $C^2$-diffeomorphism. Then $\partial \Lambda=(\partial I \times \mathbb{D}) \bigcap A$ and $\mu(\partial \Lambda)\precsim \Leb_{S^1}(\partial I)=0$ due to (\ref{allquotientmeasure}) and (\ref{horsesrbleb}). Let $R$ be the first return time constructed in \cite{Alves}, and $f_{\gamma}^R:\Lambda \to \Lambda$ is the corresponding first return map. Clearly, $\dim \gamma^u=1$. With all of these, we have that $\alpha=1+\frac{1}{\gamma}>2>\frac{2}{\dim \gamma^u}-\frac{1}{\dim_H \mu}$. The condition $\mu\{\interior{(\Lambda)}\}>0$ holds also. Hence, by the Theorem \ref{thm} we have the following \begin{corollary}\label{fplsolenoid} The functional Poisson limit laws hold for $f_{\gamma}$ for any $\gamma \in (0,1)$ with convergence rates specified in the Lemma \ref{lastproofs}. \end{corollary} \begin{remark}\ \par \begin{enumerate} \item We could also use here the Theorem \ref{thm2}. Indeed, let $U=I \times \mathbb{D}$, $\overline{R}$ is the first return time to $U$, $\Theta$ is the set of all unstable manifolds in $U$ (observe that their union is, actually, $\Lambda$). The lengths of all $\gamma^u \subseteq \Lambda$ are uniformly bounded from below. Therefore, if $\epsilon$ is small enough, then $\mu_U \{x \in U: |\gamma^u(x)|< \epsilon\}=0 \le \epsilon^b$. Note, that here $b$ is arbitrary large. For each $\gamma^u \in \Theta$, $\mu_{\gamma^u} \approx \Leb_{\gamma^u}$ and $\alpha=1+\frac{1}{\gamma}$. It is well known that correlations for $f^{\overline{R}}: U \to U$ decay exponentially. Therefore the Corollary \ref{fplsolenoid} holds. \item In \cite{pene} a ``maximum" metric was chosen, instead of the Riemannian metric. It was proved there that the Poisson limit laws hold for $\gamma \in (0,\frac{\sqrt{2}}{2})$. After checking the details therein, we found that our approach allows to improve the results proved there to $\gamma\in (0,1)$, i.e. by using their metric. Note however, that we consider only a Riemannian metric everywhere in the present paper. Therefore we omit here these calculations. \end{enumerate} \end{remark} \subsection{Axiom A attractors}\label{appA} It was proved in \cite{collet} that Axiom A attractors $\Sigma \subset \mathcal{M}$ with $\dim \gamma^u=1$ satisfy the Poisson limit laws. Later \cite{pene} it was established that the Poisson limit laws hold for the ergodic dynamics $f: \Sigma \to \Sigma$ if $\dim_H\mu>\dim{\mathcal{M}}-1$. We will show that the conditions on $\dim_H\mu$ and $\dim \gamma^u$ can be dropped. \begin{definition}[Axiom A attractors, see \cite{Bowen, sinai, Y}]\ \par Let $f : \mathcal{M} \to \mathcal{M}$ be a $C^2$-diffeomorphism. A compact set $\Sigma \subseteq \mathcal{M}$ is called an Axiom A attractor if \begin{enumerate} \item There is a neighborhood $U$ of $\Sigma$, called its basin, such that $f^n(x) \to \Sigma$ for every $x \in U$. \item The tangent bundle over $\Sigma$ is split into $E^u \oplus E^s$, where $E^u$ and $E^s$ are $df$-invariant subspaces. \item $df|_{E^u}$ is uniformly expanding and $df|_{E^s}$ is uniformly contracting. \item $f: \Sigma \to \Sigma$ is topologically mixing. \end{enumerate} \end{definition} Before turning to the proofs, we need one lemma from \cite{Bowen}: \begin{lemma}[Markov partitions, see the chapter 3 of \cite{Bowen}]\ \par The set $\Sigma$ has a Markov partition $\{\Sigma_1, \Sigma_2, \cdots, \Sigma_m\}$ into the elements with arbitrarily small diameters. Here the sets $\Sigma_i$ are proper rectangles (i.e., $\Sigma_i=\overline{\interior{(\Sigma_i)}}$ and $\interior{(\Sigma_i)}\bigcap \interior{(\Sigma_j)}=\emptyset$ for $i \neq j$, where the interior and the closure are taken with respect to the topology of $\Sigma$, rather than to the topology of $\mathcal{M}$). \end{lemma} We will verify now conditions imposed in our main theorems. \begin{enumerate} \item Let the horseshoe $\Lambda$ coincides with $\Sigma_1$. Then a return time for a hyperbolic tower $\Delta$ is actually the first return due to the existence of a Markov partition. \item The contraction rates of (un)stable manifolds are exponential, i.e. faster than required $O(\frac{1}{n^{\alpha}})$. \item a constant $\alpha$ can be in this case arbitrary large, namely ($\alpha>\max\{2,\frac{1}{\dim_H \mu}\}$). Then $\alpha> 2> \frac{2}{\dim \gamma^u}-\frac{1}{\dim_H \mu}>\frac{1}{\dim_H \mu}$. \item Finally, from the existence of a finite Markov partition follows that $\mu(\interior{(\Lambda)})>0$. And $\mu(\partial \Lambda)=0$ due to the structure of $\partial \Sigma_1$, according to the Lemma 3.11 of \cite{Bowen}. \end{enumerate} Therefore the Theorem \ref{thm} holds, and we obtain the following \begin{corollary}\label{axiomfpl} The functional Poisson limit laws hold for Axiom A attractors with a convergence rate specified in the Lemma \ref{lastproofs}. \end{corollary} \subsection{Dispersing billiards with and without finite horizon} The existence of the Gibbs-Markov-Young structure for dispersing billiards was established in \cite{chernovhighdim}. Denote by $D$ a billiard table, i.e. a closed region on the Euclidean plane with a piecewise $C^3$-smooth boundary $\partial D$. The phase space of a billiard $\mathcal {M}$ is $\partial D \times [-\frac{\pi}{2}, \frac{\pi}{2}]$. In dispersing billiards the boundary is convex inwards. These billiards are hyperbolic dynamical systems with singularities, which appear because of the orbits tangent to the boundary and orbits hitting the singularities of the boundary. For some technical reasons (see \cite{bunimovich2}) it is convenient to introduce some extra (artificial) singularities, and represent the phase space $\mathcal{M}$ as $\bigsqcup_{k \ge k_0} \partial D \times [-\frac{\pi}{2}+\frac{1}{(k+1)^2}, -\frac{\pi}{2}+\frac{1}{k^2}] \bigsqcup \bigsqcup_{-k \ge k_0} \partial D \times [\frac{\pi}{2}-\frac{1}{k^2}, \frac{\pi}{2}-\frac{1}{(k+1)^2}] \bigsqcup \partial D \times [-\frac{\pi}{2}+\frac{1}{k_0^2}, \frac{\pi}{2}-\frac{1}{k_0^2}]$, where $k_0 \gg 1$. Then $\mathcal{M}$ becomes formally closed, non-compact and disconnected. Moreover, the billiard map $f: \partial{\mathcal{M}} \to \partial{\mathcal{M}}$ becomes multi-valued because the phase space $\mathcal{M}$ gets partitioned into infinitely many pieces, and the boundary $\partial \mathcal{M}$ acquires infinitely many new components. As a result, the billiard map acquires additional singularities. However, this trick allows to get proper estimates of the distortions, probability densities and the Jacobian of the holonomy map due to partition of unstable manifolds into homogeneous ones (see the details in \cite{bunimovich2} or in the chapter 5 of \cite{CMbook}). Denote by $\mathbb{S}$ the union of all singular manifolds. We will verify now for dispersing billiards conditions of our main theorems. \begin{enumerate} \item $\overline{R}=1, U=\mathcal{M}, \mu \ll \Leb_{\mathcal{M}}$. Also, it was proved in \cite{Y, chernovhighdim} that correlations decay exponentially. \item $\alpha>0$ is arbitrarily large, $\dim_H \mu=2, \dim \gamma^u=1$. \item Let $\mathcal{Q}^{\mathbb{H}}_n(x)$ be a connected component of $\mathcal{M} \setminus \bigcup^n_{m=0}f^m(\mathbb{S})$ containing a point $x$. The partition $\Theta:=(\bigcap_{n\ge 1}\overline{\mathcal{Q}^{\mathbb{H}}_n(x)})_{x\in \mathcal{M}}$, which consists of maximal homogeneous unstable manifolds, is measurable. \item A required distortion's estimate holds for each $\gamma^u(x) \in \Theta$ by the Corollary 5.30 in \cite{CMbook}. \item The Theorem 5.17 of \cite{CMbook} gives the estimate $\mu_U \{x \in U: |\gamma^u(x)|< \epsilon\}\le C \cdot \epsilon$. (Observe that here $b=1$). \end{enumerate} Hence the Theorem \ref{thm2} can be applied, and we have \begin{corollary}\label{fpldispersing} The functional Poisson limit laws hold for two-dimensional dispersing billiards with or without a finite horizon, and the corresponding convergence rates satisfy the estimates from Lemma \ref{lastlastproof}. \end{corollary} \subsection{Billiards with focusing components of the boundary} In this section we consider two-dimensional hyperbolic billiards, which have convex outwards of the billiard table circular boundary components together with dispersing and neutral (zero curvature) components of the boundary. The main assumption is that the entire circle, which contains any focusing component, belongs to a billiard table $D$. This class of billiards was introduced and studied in \cite{buni74,bunimovich3}. Standard coordinates for the billiard map $f$ are $(r,\phi)$, where $r$ fixes a point on the boundary of a billiard table and $\phi$ is an angle of reflection off the boundary at this point. To simplify the exposition, we will consider now only the most studied and popular example in this class, called a stadium billiard. (Actually, all the reasoning for a general case is the same \cite{Bun90}). The boundary of a stadium consists of two semicircles of the same radius connected by two tangent to them neutral components. The existence of the Gibbs-Markov-Young structure for the stadium was proved in \cite{hongkun,markarian}. The phase space in this case is $\mathcal{M}:=\partial D \times [-\frac{\pi}{2}, \frac{\pi}{2}]$, where $\partial D$ is the boundary of a stadium. \begin{enumerate} \item Let $U \subseteq \mathcal{M}$ consists of all points, where the first or last collisions of the billiard map orbits with the semicircles occur. By the first (resp., last) collisions we mean here the first (resp., last) collisions with a circular component of the boundary, which occur after (resp., before) the last (resp., first) collision of the orbit in a series of consecutive collisions with the neutral part of the boundary or with another focusing component. Clearly this set is a disjoint union of two similar hexagons. Hence, it is enough to consider one of them, say the hexagon attached to $\{(r,\phi) \in \mathcal{M}: r=0\}$, (see the Figure 8.10 of \cite{CMbook}). We have $\mu \ll \Leb_{\mathcal{M}}$ and $\mu(\partial U)=0$. Let $\overline{R}$ be the first return time to $U$. Using the Theorems 4 and 5 in \cite{hongkun} we can prove that the first return map $f^{\overline{R}}: U \to U$ has an exponential decay of correlations. Consider the set of singular points $\mathbb{S}$ which correspond to hitting the four singular points of the boundary, where focusing and neutral components meet and generate jumps of the curvature. Denote $\mathbb{S}_1:=(f^{\overline{R}})^{-1}(\mathbb{S})$. \item Let $\mathcal{Q}_n(x)$ be the connected component of $\mathcal{M} \setminus \bigcup^n_{m=0}(f^{\overline{R}})^m(\mathbb{S}\bigcup \mathbb{S}_1)$ containing a point $x$. The partition $\Theta:=(\bigcap_{n\ge 1}\overline{\mathcal{Q}_n(x)})_{x\in \mathcal{M}}$ is measurable. \item The required estimate of distortion holds for each $\gamma^u(x) \in \Theta$ by the Corollary 8.53 in \cite{CMbook}. \item We will prove now that $\mu_U \{x \in U: |\gamma^u(x)|< r\}\precsim r^{\frac{1}{2}}$. Let $\mathcal{Q}'_n(x)$ be the connected component of $\mathcal{M} \setminus \bigcup^n_{m=0}(f^{\overline{R}})^m(\mathbb{S})$ which contains a point $x$. Then some smooth unstable manifolds $\gamma^{u'}(x) \in \Theta':=(\bigcap_{n\ge 1}\overline{\mathcal{Q}'_n(x)})_{x\in \mathcal{M}}$ are cut by the set $\mathbb{S}_1$ into smaller pieces, which belong to $\Theta$. (Observe that some of them could be disjoint with $\mathbb{S}_1$). It follows from the Theorem 8.42 of \cite{CMbook} that $\mu_U \{x \in U: |\gamma^{u'}(x)|< r\}\le C \cdot r$. The connected components of the set $\mathbb{S}_1$ are of two types: \begin{enumerate} \item $L_k$ is a straight (increasing in the $(r, \phi)$- coordinates) segment in $U$ with the slope $\frac{1}{k}$, representing $k$ successive reflections at one and the same semicircle (see e.g. the Figure 8.11 in \cite{CMbook}). \item $F_m$ is an increasing curve (in $(r,\phi)$-coordinates) in $U$ with slope $\approx 1$ (i.e. it is bounded away from $0$ and $+\infty$), which corresponds to $m$ successive bounces on the flat sides of the boundary (see e.g. the Figure 8.12 of \cite{CMbook}). \end{enumerate} Moreover, $L_k$ is located at the distance $\approx \frac{1}{k}$ from the set $\{(r,\phi) \in \mathcal{M}: \phi=\pm \frac{\pi}{2}\}$, and $F_m$ is at the distance $\approx \frac{1}{m}$ from $\{(r,\phi) \in \mathcal{M}: \phi=0\}$. Let \[V_1:=\{(r,\phi) \in \mathcal{M}: \phi \in [\frac{\pi}{2}-\sqrt{r}, \frac{\pi}{2}] \bigcup [-\sqrt{r}, \sqrt{r}] \bigcup [-\frac{\pi}{2}, -\frac{\pi}{2}+\sqrt{r}]\},\] \[V_2:=\bigcup_{k \le \frac{1}{\sqrt{r}}}B_{r}(L_k) \bigcup \bigcup_{m \le \frac{1}{\sqrt{r}}}B_{r}(F_m),\] where $B_r(L_k), B_r(F_m)$ are the $r$-neighborhoods of $L_k, F_m$. Then $\mu_U(V_1 \bigcup V_2) \precsim \sqrt{r}+\frac{1}{\sqrt{r}} \cdot r \precsim \sqrt{r}$. A curve $\gamma^u(x)\in \Theta$ is decreasing (in $(r,\phi)$ coordinates) for almost every $x \in U \bigcap (V_1\bigcup V_2)^c$. Also, if its length $|\gamma^u(x)|<r $, then $\gamma^u(x)$ is disjoint with $\mathbb{S}_1$, and thus $\gamma^u(x) \in \Theta'$. Therefore, \[\mu_U \{x \in U: |\gamma^u(x)|< r\} \precsim \mu_U(V_1\bigcup V_2)+\mu_U\{x\in (V_1 \bigcup V_2)^c: |\gamma^u(x)|< r\}\] \[\precsim r^{\frac{1}{2}}+\mu_U\{x\in (V_1\bigcup V_2)^c: |\gamma^{u'}(x)|< r\} \precsim \sqrt{r}.\] \item It was proved in \cite{pene} that $\alpha=1$. Also, $\dim_H \mu=2$ and $\dim \gamma^u=1$. \end{enumerate} Therefore all conditions of the Theorem \ref{thm2} are satisfied, and we have \begin{corollary}\label{fplstaduim} The functional Poisson limit laws hold for the stadium-type billiards, and the corresponding convergence rates are provided by Lemma \ref{lastlastproof}. \end{corollary} \begin{remark}[A general remark on billiards]\ \par All the considerations in our paper were traditionally dealing with hitting of small sets (e.g. small balls) in the phase spaces of hyperbolic (chaotic) dynamical systems. However, in case of billiards, the most interesting and natural questions are about hitting (or escape through) some small sets (particularly ``holes") on the boundary of billiard tables, rather than in the interior of a billiard table. These sets are small in the space (e.g. $r$) coordinate, but they are large (have a ``full" size) along the angle ($\phi$) coordinate. It is worthwhile to mention though, that there are some real life situations, when actually escape (radiation, emission) from various physical devices (cavities, lasers, etc) occurs only in some small range of angles (see e.g. \cite{nature,science}). Our results could be directly applied to such cases. However, when a target set is a strip (or a cylinder) with a finite fixed height in the angle $\phi$-coordinate, the results of the present paper can also be used/adapted by cutting a cylinder into small sets. Then the obtained estimates are valid for these pieces of a cylinder. Clearly, this approach does not generally work for recurrences, but it could be applied for the first hitting probabilities because an orbit cannot escape through one hole and then again escape through another hole. (By holes we mean here disjoint pieces of a cylinder). Therefore, one can take in such cases a relevant maximum or minimum of the obtained estimates for ``small" sets, i.e. for the pieces of a cylinder in the phase space of a billiard. It is worthwhile to mention though, that the functional Poisson limit laws for billiards with holes in the boundary (``cylindric" holes in the phase space) is a work in progress. It requires some new arguments and lengthy computations. \end{remark} \subsection{H\'enon Attractors}\label{henon} The Poisson limit laws for the H\'enon attractors (obeying hyperbolic Young towers) have been proved in \cite{collet}. However, the convergence rate for this class of dynamical systems was obtained in a weaker form (\ref{weak}). Here we will derive a stronger rate of convergence (\ref{equaintro}) by making use of some tricks from the proofs of the main Theorems \ref{thm} and \ref{thm2}. \begin{corollary}\label{fplhenon} The functional Poisson limit laws with convergence rates (\ref{equaintro}) hold for H\'enon attractors (modelled by Young towers). \end{corollary} \begin{proof} For H\'enon attractors $\alpha$ is an arbitrary large number and the decay of correlation is exponential. Hence, according to the Proposition \ref{rate}, we just need to estimate convergence rates of \[ \frac{\int_{B_r(z)} 1_{\bigcup_{1\le j \le [\frac{T}{\mu(B_r(z))}]^{\delta}}f^{-j}B_r(z)}d\mu}{\mu(B_r(z))} \text{ and } \frac{\mu(B_{r+C' \cdot r^{\alpha'} }(z)\setminus B_{r-C' \cdot r^{\alpha'}}(z))}{\mu(B_r(z))},\] where $\alpha'>0$ is an arbitrary large number and $\delta>0$ is an arbitrary small number. It follows from the Proposition 4.1 in \cite{collet} that there are constants $c,d, C>0$ and a set $\mathcal{U}_r $ with $\mu(\mathcal{U}_r) \le C \cdot r^c$, such that $\frac{\int_{B_r(z)} 1_{\bigcup_{1\le j \le q}f^{-j}B_r(z)}d\mu}{\mu(B_r(z))} \precsim_{T} C \cdot (r^d + q \cdot r^c)$ for any $z \notin \mathcal{U}_r, T>0$ and $q \ge 2$. By the Borel–Cantelli lemma $\mu(\bigcap_{N >0} \bigcup_{n \ge N} \mathcal{U}_{n^{-\frac{2}{c}}})=0$. So for almost every $z \in \mathcal{M}$, $\exists N_z, r_z>0$, such that $\forall n >N_z, z \notin \mathcal{U}_{n^{-\frac{2}{c}}}$ and $\mu(B_r(z)) \in [r^{\dim_H \mu+\delta}, r^{\dim_H \mu-\delta}]$ for any $r< r_z$ (see the Assumption \ref{geoassumption}). Then for $r< \min \{ N_z^{-\frac{2}{c}}, r_z\}$, $\exists n_r\ge N_z$, such that $(n_r+1)^{-\frac{2}{c}} \le r \le n_r^{-\frac{2}{c}}$. Hence, we have \[\frac{(1+n_r)^{-\frac{2}{c} \cdot (\dim_H \mu +\delta)}}{n_r^{-\frac{2}{c} \cdot (\dim_H \mu-\delta)}} \le \frac{\mu(B_{n_r^{-\frac{2}{c}}}(z))}{\mu(B_r(z))} \le \frac{n_r^{-\frac{2}{c} \cdot (\dim_H \mu-\delta)}}{(1+n_r)^{-\frac{2}{c} \cdot (\dim_H \mu +\delta)}} \] \[\frac{\int_{B_r(z)} 1_{\bigcup_{1\le j \le [\frac{T}{\mu(B_r(z))}]^{\delta}}f^{-j}B_r(z)}d\mu}{\mu(B_r(z))} \le \frac{\int_{B_{n_r^{-\frac{2}{c}}}(z)} 1_{\bigcup_{1\le j \le [\frac{T}{\mu(B_r(z))}]^{\delta}}f^{-j}B_{n_r^{-\frac{2}{c}}}(z)}d\mu}{\mu(B_{n_r^{-\frac{2}{c}}}(z))} \cdot \frac{\mu(B_{n_r^{-\frac{2}{c}}}(z))}{\mu(B_r(z))}\] \[\precsim_{T} [n_r^{-\frac{2d}{c}}+\frac{n_r^{-2}}{\mu(B_r(z))^{\delta}}] \cdot \frac{n_r^{-\frac{2}{c} \cdot (\dim_H \mu-\delta)}}{(1+n_r)^{-\frac{2}{c} \cdot (\dim_H \mu +\delta)}} \precsim [r^d+r^c \cdot r^{-\delta \cdot (\dim_H \mu+\delta)}] \cdot r^{-2\delta} \to 0, \] provided that $\delta=\frac{\min\{\frac{d}{3}, \dim_H\mu, \frac{c}{2\dim_H \mu +3}\}}{2}$. In other words, \[\frac{\int_{B_r(z)} 1_{\bigcup_{1\le j \le [\frac{T}{\mu(B_r(z))}]^{\delta}}f^{-j}B_r(z)}d\mu}{\mu(B_r(z))} \precsim_{z, T} r^{\frac{\min\{\frac{d}{3}, \dim_H\mu, \frac{c}{2\dim_H \mu +3}\}}{2}}\] for almost every $z\in \mathcal{M}$. It follows from the last paragraph in the proof of the Proposition 4.2 in \cite{collet}, that there are such constants $e, C>0$ that for any $f>0, z \in \mathcal{M}$ \[\mu(B_{r+C' \cdot r^{\alpha'} }(z)\setminus B_{r-C' \cdot r^{\alpha'}}(z))\le C \cdot (r^{\frac{\alpha'}{2}} \cdot r^{-e}+r^f).\] Hence, for almost every $z \in \mathcal{M}$ and $\forall r< r_z$ \[\frac{\mu(B_{r+C' \cdot r^{\alpha'} }(z)\setminus B_{r-C' \cdot r^{\alpha'}}(z))}{\mu(B_r(z))}\le C \cdot \frac{(r^{\frac{\alpha'}{2}} \cdot r^{-e}+r^f)}{r^{\dim_H \mu+\delta}} \to 0,\] if $\alpha' > 2(e+\delta + \dim_H \mu)$ and $f> \dim_H \mu+ \delta$. Thus \[\frac{\mu(B_{r+C' \cdot r^{\alpha'} }(z)\setminus B_{r-C' \cdot r^{\alpha'}}(z))}{\mu(B_r(z))}\precsim_{z} \frac{(r^{\frac{\alpha'}{2}} \cdot r^{-e}+r^{2(\dim_H \mu +\delta)})}{r^{\dim_H \mu+\delta}} \to 0\] for almost every $z \in \mathcal{M}$. Now the required estimates of the convergence rates for short returns and coronas are obtained, and the Corollary \ref{fplhenon} holds, according to the Proposition \ref{rate}. \end{proof} \section*{Acknowledgements} We are indebted to I. Melbourne for useful comments and for pointing to the paper \cite{melboune}. \bibliographystyle{amsalpha}%

https://arxiv.org/abs/2112.11148

Schurian-finiteness of blocks of type $A$ Hecke algebras

For any algebra $A$ over an algebraically closed field $\mathbb{F}$, we say that an $A$-module $M$ is Schurian if $\mathrm{End}_A(M) \cong \mathbb{F}$. We say that $A$ is Schurian-finite if there are only finitely many isomorphism classes of Schurian $A$-modules, and Schurian-infinite otherwise. By work of Demonet, Iyama and Jasso it is known that Schurian-finiteness is equivalent to $\tau$-tilting-finiteness, so that we may draw on a wealth of known results in the subject. We prove that for the type $A$ Hecke algebras with quantum characteristic $e\geq 3$, all blocks of weight at least $2$ are Schurian-infinite in any characteristic. Weight $0$ and $1$ blocks are known by results of Erdmann and Nakano to be representation finite, and are therefore Schurian-finite. This means that blocks of type $A$ Hecke algebras (when $e\geq 3$) are Schurian-infinite if and only if they have wild representation type if and only if the module category has finitely many wide subcategories. Along the way, we also prove a graded version of the Scopes equivalence, which is likely to be of independent interest.

\section{Introduction} Let $A$ be a finite-dimensional algebra over an algebraically closed field $\bbf$. All the $A$-modules we consider in this paper are assumed to be finite-dimensional left modules. Then, an $A$-module $M$ is called Schurian if $\End_A(M) = \bbf$. Schurian modules appear in various places. Among them, the following two examples are well-known. \begin{itemize} \item[(1)] Let $\theta:K_0(A\textup{-mod})\to \bbz$ be a linear map. An $A$-module $M$ is called stable if $\theta(M)=0$ and any nonzero proper submodule $N$ satisfies $\theta(N)<0$. Stable modules are Schurian. \item[(2)] Let $A$ be an $\bbf$-algebra. If an $A$-module $M$ is an indecomposable module which belongs to the preprojective or the preinjective component of the Auslander--Reiten quiver of $A$, then $M$ is Schurian and $\operatorname{Ext}^1_A(M,M)=0$ -- for example, see~\cite[Chap.VIII, Lemma~2.7]{ASS1}. \end{itemize} An algebra $A$ is called Schurian-finite if the number of isomorphism classes of Schurian modules is finite. \begin{rem} In \cite{CKW}, they call an algebra $A$ Schur-representation-finite if there are finitely many isomorphism classes of Schurian $A$-modules of a fixed dimensional vector $d$, for each $d$. \end{rem} There had been few results on Schurian modules until recently, but the $\tau$-tilting theory initiated by Adachi, Iyama and Reiten \cite{AIR} has changed the perspective. An $A$-module $M$ is $\tau$-rigid if $\Hom_A(M,\tau(M))=0$, where $\tau$ is the Auslander--Reiten translate. An algebra $A$ is then called $\tau$-tilting finite if there are only finitely many isomorphism classes of indecomposable $\tau$-rigid $A$-modules. Demonet, Iyama and Jasso \cite[Theorem~4.2]{DIJ} proved that the map that sends an $A$-module $M$ to $M/\operatorname{Rad}_B(M)$, where $B=\End_A(M)$, induces a bijection between isomorphism classes of indecomposable $\tau$-rigid $A$-modules and isomorphism classes of Schurian $A$-modules. Thus, Schurian-finiteness has reappeared under the name $\tau$-tilting finiteness. Let $M$ be a $\tau$-rigid $A$-module, and $P$ be a projective $A$-module. We call the pair $(M,P)$ support $\tau$-tilting if $\Hom_A(P,M)=0$ and the numbers of pairwise non-isomorphic indecomposable direct summands of $M$ and $P$ sum up to the number of the isomorphism classes of simple $A$-modules. Then, support $\tau$-tilting pairs are in bijection with functorially-finite torsion classes in $A\textup{-mod}$, and with two-term silting complexes in $K^b(A\textup{-proj})$~\cite{AIR}. In other words, the study of Schurian modules has applications to those representation theoretic classifications. Another application is that if $A$ is wild and Schurian-finite, then it gives an example of a wild algebra which is not strictly wild. Recently, the relationship with polytope theory has also been pursued. Because of its importance, researchers who study $\tau$-tilting theory look at various examples to check whether they are $\tau$-tilting finite or not. For example, we know that preprojective algebras of Dynkin type are Schurian-finite by~\cite{Mizuno14}. Note that if $A$ is representation-finite, then $A$ is Schurian-finite. An immediate consequence of this is that blocks of Hecke algebras of weight $0$ or $1$ are Schurian-finite, by \cite{enreptype,APA1}. The converse may not hold. Indeed, preprojective algebras of type other than $A_n$ for $1\leq n \leq 4$ are representation-infinite. Another example is a multiplicity-free Brauer cyclic graph algebra with an odd number of vertices~\cite{Ad2}. However, it is observed that the converse \emph{does} hold for many classes of algebras. The first obvious examples are those which admit preprojective component, such as the class of path algebras of acyclic quivers, quasi-tilted algebras~\cite{ch97}, and algebras that satisfy the separation condition~\cite[Chapter IX, Theorem 4.5]{ASS1}. Other examples include cycle-finite algebras~\cite{MS16}, gentle algebras~\cite{Plam19}, tilted and cluster-tilted algebras~\cite{zito20}, locally hereditary algebras~\cite{AHMW}, and simply-connected algebras~\cite{qiwangsimplyconnected}. As algebras satisfying the separation condition are simply-connected~\cite{skow93}, it also implies that representation finiteness coincides with Schurian-finiteness for them. In this article, we initiate the study of Schurian-finiteness for block algebras of the Hecke algebra $\hhh$ of the symmetric group and add a new family of algebras for which Schurian-finiteness coincides with representation finiteness. The difficulties to overcome were twofold. Recall that block algebras of the Hecke algebra of the symmetric group $\sss$ are labeled by $e$-cores $\kappa$ such that $n-|\kappa|\in e\bbz_{\geq0}$, where $e$ is the \emph{quantum characteristic}. The nonnegative integer $(n-|\kappa|)/e$ is called the weight of the block algebra. As all of the known methods to determine the Schurian-infiniteness of a finite-dimensional algebra are based on the bound quiver presentation of the algebra, we have to obtain information of the Gabriel quiver of each block algebra of $\hhh$ which suffices to determine the Schurian-infiniteness. Then, the first difficulty was the lack of a method of reduction to small $n$ (except for the Scopes equivalence), that would allow us to determine the Schurian-infiniteness by explicit computation. For example, the induction functor does not behave well with the Schurian property. On the other hand, we may use the induction functor to determine the representation type of block algebras of $\hhh$. Thus, we must control the Gabriel quiver of the block algebra, for large $n$, itself. Then, the second difficulty was the fact that it is not easy to compute the Gabriel quiver of the block even if one tries to find a small piece of information on the structure of indecomposable projective modules. Indeed, when we study algebras outside of bound quiver algebras, this is the most difficult part. Curiously, this reality is often overlooked. The key idea to overcome those difficulties is \cref{prop:matrixtrick}. It asserts that if we find a certain set of partitions for which the graded decomposition numbers labelled by them satisfy certain conditions, then we may find enough information to determine Schurian-infiniteness. Because of the proposition, we are able to focus on graded decomposition numbers exclusively to prove our main theorems. This assertion may sound a bit surprising because graded decomposition numbers of a graded cellular algebra are determined by the Grothendieck group of its graded modules and they do not contain information on the Gabriel quiver in general. When we consider the blocks of $\hhh$ in positive characteristic that appear in \cref{prop:matrixtrick}, the submatrix of graded decomposition numbers that satisfy the necessary extra conditions still does not give us information on extensions between the corresponding simple modules for $\hhh$. Rather, we deduce extensions between simples for some idempotent truncation, from which we can deduce Schurian-infiniteness of the truncated algebra, and lift the property back to $\hhh$. To apply \cref{prop:matrixtrick}, we must find partitions that satisfy the assumptions of the proposition. We note that this requires insight, and another new idea here is in the novel use of various runner-removal and row-removal theorems. In small characteristics, we often see different behaviour than the general patterns. Thus, we need a different method to determine the Schurian-finiteness. In particular, we assume that the quantum characteristic is $e\ne 2$ for all three of our main theorems. \begin{rem} Morita classes of tame blocks are classified by the first author~\cite{ariki21} under the assumption $p\ne2$. Then Wang showed in~\cite[Theorem~5.5]{qiwang2point} that all those tame blocks of Hecke algebras are $\tau$-tilting finite (and therefore Schurian-finite). In~\cite[Theorem~3.4]{qiwang22}, Wang also showed that blocks of tame Schur algebras $S(n, r)$, which occur if $p=2$, are $\tau$-tilting finite, from which it follows that the (indecomposable) Hecke algebra $\hhh[4]$ is also Schurian-finite if $e=p=2$. This is the only example of a block of $\hhh$ of weight at least $2$, for $e=p=2$, for which we can determine whether it is Schurian-finite. \end{rem} The following theorems are the main results of this article. The proofs utilise recent advances on graded decomposition numbers. For an explanation of the notation, we refer the reader to \cref{subsec:abacus}. \begin{thm}\label{thm:wt2} Suppose $e\geq 3$, and that $B$ is any block of weight $2$. Then $B$ is Schurian-infinite in any characteristic. \end{thm} \begin{rem} We need to study the case $e=3$, $p=2$, $_10_2$, and $_00_1$ separately. This case consists of the (Scopes class of) Rouquier blocks. These blocks are Scopes equivalent to the block of $\hhh[11]$ with core $(3,1^2)$. \end{rem} \begin{thm}\label{thm:wt3} Suppose $e\geq 3$, and that $B$ is any block of weight $3$. Then $B$ is Schurian-infinite in any characteristic. \end{thm} \begin{rem} Once again, we need to study the case $e=3$, $p=2$, $_10_2$, and $_00_1$ separately. This case now consists of four Scopes classes of blocks, one of which is the class of Rouquier blocks, whose minimal representative is the weight 3 Rouquier block of $\hhh[25]$. \end{rem} \begin{thm}\label{thm:principalblock} Suppose $e\geq 3$, and that $B$ is the principal block of the Hecke algebra $\hhh$ of weight greater than or equal to $2$. Then $B$ is Schurian-infinite in any characteristic. In particular, if $e\geq 3$, then $B$ is Schurian-finite if and only if it is representation-finite. \end{thm} \begin{rem} In an earlier version of this paper, we needed Fock space theory, cf.~\cite{arikibook}, and a result by the first author and Mathas~\cite[Corollary 3.7]{am04}, to settle several cases in weight four. In the current version we have developed a more combinatorial method to replace the canonical basis computation. \end{rem} The paper is organised as follows. In \cref{sec:background}, we recall various runner-, row-, and column-removal theorems in \cref{subsec:grdec} after we fix notations for Hecke algebras, partitions and abacus displays. Our method to show Schurian-infiniteness by using graded decomposition numbers is explained in \cref{subsec:reduction}. We also prove a graded analogue of the Scopes equivalence in \cref{sec:gradedScopes}, which will be of independent interest. After \cref{sec:gradedScopes}, we exclusively work with graded decomposition numbers -- except for in a few cases -- to find partitions that satisfy the assumptions of \cref{prop:matrixtrick}. As much is known for blocks of weight $2$ and $3$, we give a classification of Schurian-finiteness/infiniteness for those blocks in \cref{sec:wt2,sec:wt3} whenever $e\geq 3$, proving \cref{thm:wt2,thm:wt3}. Our approach there largely hinges on the runner-removal result of James and Mathas. In \cref{sec:highwt}, we assume that $e\geq3$ and focus on blocks of weight greater than or equal to $4$. There, we deduce that any block whose core satisfies any of several conditions is Schurian-infinite, by using a row- and column-removal theorem of Chuang, Tan and Miyachi~\cite{cmt02} along with a special case of a result of Bowman and the second author \cite{bs15}, that generalises a result of Kleshchev~\cite{klesh1box} and Chuang, Tan and Miyachi~\cite{cmt08} -- see \cref{prop:2runners,prop:4runners,prop:3runners,prop:3runnersalt,cor:conjugateblocks}. As a by-product, we will easily prove \cref{thm:principalblock} whenever $e\geq 4$. The case $e=3$ requires some special treatment. By using the aforementioned results, we may easily prove that principal blocks are Schurian-infinite in the following cases. \begin{enumerate}[label=(\roman*)] \item The core is the empty partition and the weight is $w\geq 4$. \item The core is $(1)$ and the weight is $w\geq 5$. \item The core is $(2)$ and the weight is $w\geq 6$. \end{enumerate} The remaining cases are then just the principal blocks of $\hhh[13]$, $\hhh[14]$, and $\hhh[17]$ with $e=3$, which we resolve using explicit submatrices of the graded decomposition matrices. These are all resolved in \cref{thm:principalblockmain}. Note that the principal blocks of weights $2$ and $3$ are covered by \cref{thm:wt2,thm:wt3}. \begin{ack} The first author is partially supported by JSPS Kakenhi grants number 18K03212 and 21K03163. The second author is partially supported by JSPS Kakenhi grant number 20K22316. The second author thanks Osaka University for their hospitality during his visit, in which this research was initiated. The first author is grateful for a pleasant stay at OIST in November 2021. During the visit, we completed most of the work on this paper. We thank Matt Fayers for making his LaTeX style file for abacus displays publicly available, for pointing out a useful reference within \cite{richardswt2} that we use to characterise \emph{adjacent} partitions in \cref{sec:wt2}, and for helpful conversations when we applied his results to a concrete example. We also thank Andrew Mathas for providing us with the graded decomposition matrices for various blocks of low rank when $e=3$ and $p>0$ -- these were a great help in proving our result for those blocks. The second author thanks Eoghan McDowell for help with TikZ. The first author considered the Schurian-finiteness of block algebras of Hecke algebras two years ago with Ryoichi Kase, Kengo Miyamoto, Euiyong Park, and Qi Wang, trying to find an explicit subquotient algebra and prove its Schurian-infiniteness. Although unsuccessful in this endeavour, the first author thanks those four colleagues for their time spent together on the project in 2019. \end{ack} \section{Background}\label{sec:background} Throughout, we let $\bbf$ denote an algebraically closed field of characteristic $p\geq 0$. All our modules are left modules. \subsection{Partitions}\label{subsec:partitions} A partition of $n$ is a weakly decreasing sequence of nonnegative integers $\la = (\la_1, \la_2, \dots)$ that sum to $n$. We write $\la \vdash n$ to mean `$\la$ is a partition of $n$'. It will be useful to consider each partition to have infinite length, though it will only have finitely many nonzero terms. Thus we will omit the trailing zeroes when writing $\la$, and always consider $\la_r = 0$ for $r$ large enough. We will also group together equal terms, so that, for example, we will write $(3,1^2)$ instead of $(3,1,1)$. We denote the unique partition of $0$ by $\varnothing$. For $e \in \bbz_{\geq 2}$, we say that a partition $\la$ is \emph{$e$-regular} if it does not have any $e$ nonzero parts being equal, and \emph{$e$-singular} if it does. For example, $(5,1^2)$ is $3$-regular, but $(4,1^3)$ and $(3,1^4)$ are both $3$-singular. The \emph{conjugate partition} $\la'$ is defined by \[ \la'_i = |\{ j\geq 1 \mid \la_j \geq i\}|. \] For a partition $\la$, its \emph{Young diagram} is the set $[\la] = \{(r,c) \in \bbn \times \bbn \mid c \leq \la_r\}$, which we may depict as boxes in the plane, following the English convention, as in the example below. To each node $A = (r,c) \in [\la]$, we assign the \emph{$e$-residue} $i \in \{0, 1, \dots, e-1\}$ with $i \equiv c-r \pmod e$. If a node $A \in [\la]$ (respectively $B \notin [\la]$) is such that $[\la]\setminus A$ (respectively $[\la] \cup B$) is a Young diagram for a partition, then we say that $A$ is a \emph{removable node} (respectively $B$ is an \emph{addable node}). We refer to an addable or removable node $(r,c)$ as an \emph{addable $i$-node} or a \emph{removable $i$-node}, respectively, if $i \equiv c-r \pmod e$. \begin{eg} Let $e=4$, and take $\la = (6,4,3,1^2)$. Then $[\la]$ is drawn below, with the residues written in each node, as well as the residues of addable nodes. \[ \Yaddables1 \yngres(4,6,4,3,1^2) \] \end{eg} If $\la$ and $\mu$ are partitions of $n$, we say that $\la$ dominates $\mu$, and write $\la \dom \mu$, if $\la_1 + \la_2 + \dots + \la_r \geq \mu_1 + \mu_2 + \dots + \mu_r$ for all $r$, and write $\la \doms \mu$ if $\la \dom \mu$ and $\la \neq \mu$. \subsection{Hecke algebras}\label{subsec:hecke} Let $q\in \bbf^\times$. The Hecke algebra of the symmetric group, denoted by $\hhh$, is the unital associative $\bbf$-algebra generated by $T_1, T_2, \dots, T_{n-1}$ subject to the relations \begin{alignat*}3 (T_i - q)(T_i + 1) &= 0 \qquad &&\text{ for } i = 1,\dots, n-1;\\ T_i T_j &= T_j T_i \qquad &&\text{ for } 1\leq i,j \leq n-1 \text{ with } |i-j|>1;\\ T_i T_{i+1} T_i &= T_{i+1} T_i T_{i+1} \qquad &&\text{ for } i = 1,\dots, n-2. \end{alignat*} An excellent introduction to the representation theory of $\hhh$, which largely parallels that of the symmetric group $\sss$, can be found in \cite{mathas}. Here, we will briefly recall some key aspects that we will require. The \emph{quantum characteristic} of $\hhh$ is the smallest positive integer $e$ such that $1 + q + q^2 + \dots + q^{e-1} = 0$, if such an $e$ exists, and we set $e = \infty$ otherwise. The Hecke algebras are cellular algebras. For each partition $\la$ of $n$, one may construct the Specht module $\spe\la$, which is also a cell module, with cellular basis given by the \emph{dual Murphy basis} -- see, for example, \cite{durui01,hm10}. If $e > n$, $\hhh$ is semisimple, and the set $\{\spe\la \mid \la \vdash n\}$ is a complete set of pairwise non-isomorphic irreducible $\hhh$-modules. If $e \leq n$, then $\spe\la$ has a simple head, denoted $\D\la$, whenever $\la$ is $e$-regular. The set $\{\D\la \mid \la \vdash n, \text{ $\la$ is $e$-regular}\}$ is a complete set of pairwise non-isomorphic irreducible $\hhh$-modules when $e < \infty$. The Specht modules $\spe\la$ can be constructed very explicitly, and have a basis indexed by \emph{standard $\la$-tableaux}. However, the simple modules $\D\la$ are much harder to explicitly construct, except in special cases. In general, even the dimensions of $\D\la$ are unknown. It is well-known that the Specht modules are indecomposable if $e\neq 2$ -- for example, by \cite[Corollary 8.6]{dj91}. We will abuse notation a little and say that two partitions $\la$ and $\mu$ lie in the same block of $\hhh$ whenever $\spe\la$ and $\spe\mu$ do. If we remove all length $e$ rim hooks from a Young diagram $[\la]$, then we obtain a Young diagram for its \emph{$e$-core}, i.e.~a partition whose Young diagram has no removable length $e$ rim hooks. The following result is well-known, and is often referred to as Nakayama's conjecture. \begin{thmc}{mathas}{Corollary 5.38} Two partitions of $\hhh$ are in the same block if and only if they have the same $e$-core. \end{thmc} More recently, Khovanov and Lauda~\cite{kl09}, and, independently, Rouquier~\cite{rouq}, have introduced \emph{quiver Hecke algebras}, or \emph{KLR algebras}, in order to categorify the negative halves of quantum groups. Khovanov and Lauda also introduced cyclotomic quotients in their paper, that we refer to as \emph{cyclotomic KLR algebras}, that categorify highest weight irreducible modules over quantum groups~\cite{kk12}. Importantly for us, Brundan and Kleshchev showed in \cite{bkisom} that if $e = \infty$, $\hhh$ is isomorphic to a level 1 cyclotomic KLR algebra of type $A_\infty$, while if $e<\infty$, $\hhh$ is isomorphic to a level 1 cyclotomic KLR algebra of type $A^{(1)}_{e-1}$. We will not recall the (long!)~presentation of the cyclotomic KLR algebras, as we will not be directly working with the definition. For our purposes, it will suffice to note that this framework allows us to study the graded representation theory of $\hhh$, which is further developed in \cite{bk09,bkw11}. In \cref{subsec:grdec} we will discuss this further. \subsection{Abacus combinatorics}\label{subsec:abacus} For a partition $\la$ of $n$, we define its \emph{beta-numbers} to be $\beta_i = \la_i - i$, for all $i\geq 1$. The $e$-runner abacus is drawn with $e$ infinite vertical runners, with marked positions increasing from left-to-right along successive `rows'. Our convention is that the 0 position is on the leftmost runner. For example, if $e=4$, our abacus is marked as follows. {\scriptsize \[ \begin{tikzpicture}[scale=.6] \foreach\x in{0,...,3}{\draw(\x,-2.5)--++(0,5);\draw[dashed](\x,-2.5)--++(0,-1);\draw[dashed](\x,2.5)--++(0,1);} \foreach\x in{0,...,3}{\draw(\x,0)node[fill=white]{$\x$};}; \foreach\x in{-8,...,-5}\draw(\x+8,2)node[fill=white]{$\x$}; \foreach\x in{-4,...,-1}\draw(\x+4,1)node[fill=white]{$\x$}; \foreach\x in{4,...,7}\draw(\x-4,-1)node[fill=white]{$\x$}; \foreach\x in{8,...,11}\draw(\x-8,-2)node[fill=white]{$\x$}; \end{tikzpicture} \] } The \emph{abacus display} for $\la$ is then obtained by placing a bead in position $\beta_i$, for each $i \geq 1$. The $e$-core of $\la$ is obtained by stripping off all possible $e$-rim hooks from the Young diagram of $\la$, which translates to pushing all beads up as high as possible on the abacus. We will adopt the notation and conventions of \cite{fayerswt2data}, which is in turn a sleeker presentation of the results of \cite{richardswt2}. If $B$ is a block with core $\kappa$, we take the abacus display for $\kappa$ and define the integers $p_0 < p_1 < \dots < p_{e-1}$ so that each is the position of the lowest bead on one of the runners. If $B$ is a weight 2 block, then the abacus display for any partition in $B$ is obtained from that for its $e$-core by sliding two beads down one place, or by sliding one bead down two places. Then, for $0 \leq i \leq j < e$, define \[ _iB_j = \begin{cases} 1 &\text{if } p_j - p_i < e,\\ 0 &\text{if } p_j - p_i > e. \end{cases} \] Collectively, the array $_iB_j$ is the \emph{pyramid} of the weight 2 block corresponding to the $e$-core we started with. We extend this to all pairs of integers by setting $_iB_j = 0$ if $i<0$ or $j\geq e$, and $_iB_j = 1$ if $i>j$. Finally, we adopt the shorthand notation $_i0_j$ and $_i1_j$ to mean that $_iB_j = 0$ and $1$, respectively. In particular, we note that $_i0_e$ for any $i$. \begin{eg} Let $e=4$, and let $\rho$ be the 4-core $(2^2)$. Then the corresponding beta-numbers are $\beta = (1,0,-3,-4,\dots)$, and the corresponding abacus display is below. \[ \abacus(vvvv,bbbb,bbbb,bbnn,bbnn,nnnn,vvvv) \] Note that we have $p_0 = -6$, $p_1 = -5$, $p_2 = 0$, $p_3 = 1$, and thus that $_01_1$, $_00_2$, $_00_3$, $_10_2$, $_10_3$, $_21_3$. We always have that $_i1_i$, since $p_i - p_i = 0 < e$. \end{eg} Next, we introduce notation for the partitions in a weight 2 block. Number the runners from $0$ to $e-1$ so that runner $i$ contains the marked position $p_i$, recalling that $p_0 < p_1 < \dots < p_{e-1}$. If the abacus display of $\la$ is obtained from that of its $e$-core by sliding the lowest beads on runners $i$ and $j$ each down one spot, with $i<j$, we denote the partition by $\langle i, j \rangle$. If it is obtained by sliding the lowest bead on runner $i$ down two spaces, we denote it by $\langle i \rangle$. Finally, if it is obtained by sliding each of the bottom two beads on runner $i$ down one space, we denote it $\langle i^2 \rangle$. Thus we may denote any partition in $B$ by one of the above, for some $i$ (and possibly $j$). These partitions thus index all Specht modules in the weight two block $B$. \begin{eg} Continuing our previous example, where $e=4$, and we consider a block $B$ with 4-core $\rho = (2^2)$, and weight $2$. Then the abacus displays for $\langle 3 \rangle$, $\langle 2 \rangle$, $\langle 1 \rangle$, and $\langle 0 \rangle$ are below, yielding partitions $(10,2)$, $(9,3)$, $(4,3^2,1^2)$, $(3^3,1^3)$, respectively. \[ \abacus(vvvv,bbbb,bbbb,bbnn,bnnn,nnnn,nbnn,nnnn,vvvv) \qquad \qquad \abacus(vvvv,bbbb,bbbb,bbnn,nbnn,nnnn,bnnn,nnnn,vvvv) \qquad \qquad \abacus(vvvv,bbbb,bbbn,bbnn,bbnb,nnnn,nnnn,nnnn,vvvv) \qquad \qquad \abacus(vvvv,bbbb,bbnb,bbnn,bbbn,nnnn,nnnn,nnnn,vvvv) \] The abacus displays for $\langle 2, 3 \rangle$, $\langle 1,3 \rangle$, $\langle 3^2 \rangle$, and $\langle 2^2 \rangle$ are below, yielding partitions $(6^2)$, $(6,2^2,1^2)$, $(6,3^2)$, $(5,3^2,1)$, respectively. \[ \abacus(vvvv,bbbb,bbbb,bbnn,nnnn,bbnn,nnnn,nnnn,vvvv) \qquad \qquad \abacus(vvvv,bbbb,bbbn,bbnb,bnnn,nbnn,nnnn,nnnn,vvvv) \qquad \qquad \abacus(vvvv,bbbb,bbbb,bnnn,bbnn,nbnn,nnnn,nnnn,vvvv) \qquad \qquad \abacus(vvvv,bbbb,bbbb,nbnn,bbnn,bnnn,nnnn,nnnn,vvvv) \] \end{eg} \subsection{Graded decomposition matrices}\label{subsec:grdec} Let $\la, \mu \vdash n$, with $\mu$ $e$-regular. The corresponding \emph{decomposition number} is the multiplicity $d_{\la\mu}^{e,p}(1) = [\spe\la : \D\mu]$ of $\D\mu$ in $\spe\la$. For a graded $\hhh$-module $D$, let $D\langle d \rangle$ denote the graded shift (by $d$) of the module $D$ -- in other words $D\langle d \rangle_r = D_{r-d}$. Then the corresponding \emph{graded} decomposition number is the Laurent polynomial \[ d_{\la\mu}^{e,p}(v) = [\spe\la : \D\mu]_v = \sum_{d\in\bbz} [\spe\la : \D\mu \langle d \rangle] v^d \in \bbn[v,v^{-1}]. \] It is known that $d_{\la\la}^{e,p}(v) = 1$ and $d_{\la\mu}^{e,p}(v) \neq 0$ only if $\la \domby \mu$ -- this follows, for instance, from the graded cellular basis of Hu--Mathas~\cite[Lemma~2.13, Theorem~6.11 and Section~6.4]{hm10}. It is also well-known that the \emph{ungraded} decomposition matrices $D_p = (d_{\la\mu}^{e,p}(1))_{\la,\mu}$ and $D_0 = (d_{\la\mu}^{e,0}(1))_{\la,\mu}$ are related by multiplication by the so-called \emph{adjustment matrix}. For us, we will need the more recent development of the \emph{graded adjustment matrix}, introduced in~\cite[Section~5.6]{bk09}. By~\cite[Theorem~5.17]{bk09}, $D_p$, the characteristic $p$ graded decomposition matrix, is obtained from the characteristic 0 one $D_0$ by post-multiplication by the lower-unitriangular adjustment matrix, whose entries $a_{\la\mu}(v)$ -- indexed by $\la$ and $\mu$ both $e$-regular partitions of $n$ -- are Laurent polynomials with nonnegative integral coefficients, symmetric in $v, v^{-1}$. In other words, \[ d_{\la\mu}^{e,p}(v) = d_{\la\mu}^{e,0}(v) + \sum_{\nu \domsby \mu} d_{\la\nu}^{e,0}(v) a_{\nu\mu}(v). \] \begin{lem}\label{lem:charfree} Suppose that $\la$ and $\mu$ are partitions, with $\mu$ $e$-regular, and that $d_{\la\mu}^{e,p}(1) = d_{\la\mu}^{e,0}(1)$ for a prime $p$. Then $d_{\la\mu}^{e,p}(v) = d_{\la\mu}^{e,0}(v)$. \end{lem} \begin{proof} Recall that \[ d_{\la\mu}^{e,p}(v) = d_{\la\mu}^{e,0}(v) + \sum_{\nu \domsby \mu} d_{\la\nu}^{e,0}(v) a_{\nu\mu}(v). \] Since $d_{\la\mu}^{e,p}(1) = d_{\la\mu}^{e,0}(1)$, we have that $\sum_{\nu \domsby \mu} d_{\la\nu}^{e,0}(v) a_{\nu\mu}(v) = 0$, and the result follows. \end{proof} Next, we discuss the main results of \cite{bs15}, one of which will be crucial to our proof of \cref{thm:principalblock}. First, we must adopt a mirrored-Russian convention for drawing Young diagrams, so that for example the partition $(4,2,1)$ is drawn as follows. \[ \YRussian \yng(3,2,1,1) \] Now, let $\gamma$ be a ($\ell$-multi)partition that has no addable $i$-nodes -- called `$i$-admissible' in \cite{bs15}. In \cite[Section~4.1]{bs15}, diagonals of residue $i$ in the Young diagram for $\gamma$, along with those surrounding diagonals of residue $i\pm 1$ are collectively referred to as $i$-diagonals. These may be built up vertically from the lower edge of the Young diagram by stacking certain bricks, therein denoted $\mathbf{B}_1$, $\mathbf{B}_2$, $\mathbf{B}_3$, $\mathbf{B}_4$, $\mathbf{B}_5$, and $\mathbf{B}_6$. The brick $\mathbf{B}_6$ is the `empty brick' that marks an $i$-diagonal if it is the leading diagonal in the partition -- in this paper, this will only correspond to $i=0$, but \cite{bs15} works more generally with $\ell$-multipartitions. \cite[Figures~9 and 10]{bs15} are reproduced below, which help to visualise how the $i$-diagonals are built from these bricks. See also \cite[Figures~11 and 12]{bs15} for some more helpful depictions of building these $i$-diagonals. From these it is clear that \begin{itemize} \item the diagonals are mostly built by stacking $\mathbf{B}_1$ bricks. \item the leading diagonal must start with the empty brick $\mathbf{B}_6$ at the bottom, any bricks to the left of the leading diagonal start with the brick $\mathbf{B}_4$ at the bottom, while any to the right of it start with the brick $\mathbf{B}_5$ at the bottom. \item the top of a diagonal consists of a $\mathbf{B}_2$ or $\mathbf{B}_3$ brick if and only if there is no addable $i$-node at the top of this diagonal. \end{itemize} \begin{figure*}[h]\captionsetup{labelformat=empty} \begin{center}\scalefont{0.55} \begin{tikzpicture}[scale=.7] \path (0,0) coordinate (origin); \draw[thick] (origin) --++(130:2*0.8) --++(40:1*0.8) --++(-50:1*0.8) --++(40:1*0.8) --++(-50:1*0.8) --++(220:2*0.8) ; \clip (origin) (origin) --++(130:2*0.8) --++(40:1*0.8) --++(-50:1*0.8) --++(40:1*0.8) --++(-50:1*0.8) --++(220:2*0.8); \path (40:1cm) coordinate (A1); \path (40:2cm) coordinate (A2); \path (40:3cm) coordinate (A3); \path (40:4cm) coordinate (A4); \path (130:1cm) coordinate (B1); \path (130:2cm) coordinate (B2); \path (130:3cm) coordinate (B3); \path (130:4cm) coordinate (B4); \path (A1) ++(130:3cm) coordinate (C1); \path (A2) ++(130:2cm) coordinate (C2); \path (A3) ++(130:1cm) coordinate (C3); \foreach \i in {1,...,19} { \path (origin)++(40:0.8*\i cm) coordinate (a\i); \path (origin)++(130:0.8*\i cm) coordinate (b\i); \path (a\i)++(130:4cm) coordinate (ca\i); \path (b\i)++(40:4cm) coordinate (cb\i); \draw[thin,gray] (a\i) -- (ca\i) (b\i) -- (cb\i); \draw (origin) ++(40:0.4)++(130:0.4) node {${i}$} ; \draw (origin) ++(40:0.4)++(130:1.2) node {${i+1}$} ; \draw (origin) ++(40:1.2)++(130:0.4) node {${i-1}$} ; } \end{tikzpicture} \quad \begin{tikzpicture}[scale=.7] \path (0,0) coordinate (origin); \draw[thick] (origin) --++(130:1*0.8) --++(40:2*0.8) --++(-50:1*0.8) --++(220:2*0.8) ; \clip (origin) --++(130:1*0.8) --++(40:2*0.8) --++(-50:1*0.8) --++(220:2*0.8) ; \path (40:1cm) coordinate (A1); \path (40:2cm) coordinate (A2); \path (40:3cm) coordinate (A3); \path (40:4cm) coordinate (A4); \path (130:1cm) coordinate (B1); \path (130:2cm) coordinate (B2); \path (130:3cm) coordinate (B3); \path (130:4cm) coordinate (B4); \path (A1) ++(130:3cm) coordinate (C1); \path (A2) ++(130:2cm) coordinate (C2); \path (A3) ++(130:1cm) coordinate (C3); \foreach \i in {1,...,19} { \path (origin)++(40:0.8*\i cm) coordinate (a\i); \path (origin)++(130:0.8*\i cm) coordinate (b\i); \path (a\i)++(130:4cm) coordinate (ca\i); \path (b\i)++(40:4cm) coordinate (cb\i); \draw[thin,gray] (a\i) -- (ca\i) (b\i) -- (cb\i); \draw (origin) ++(40:0.4)++(130:0.4) node {${i}$} ; \draw (origin) ++(40:0.4)++(130:1.2) node {${i+1}$} ; \draw (origin) ++(40:1.2)++(130:0.4) node {${i-1}$} ; } \end{tikzpicture} \quad \begin{tikzpicture}[scale=.7] \path (0,0) coordinate (origin); \draw[thick] (origin) --++(130:2*0.8) --++(40:1*0.8) --++(-50:2*0.8) --++(220:1*0.8) ; \clip (origin) --++(130:2*0.8) --++(40:1*0.8) --++(-50:2*0.8) --++(220:1*0.8) ; \path (40:1cm) coordinate (A1); \path (40:2cm) coordinate (A2); \path (40:3cm) coordinate (A3); \path (40:4cm) coordinate (A4); \path (130:1cm) coordinate (B1); \path (130:2cm) coordinate (B2); \path (130:3cm) coordinate (B3); \path (130:4cm) coordinate (B4); \path (A1) ++(130:3cm) coordinate (C1); \path (A2) ++(130:2cm) coordinate (C2); \path (A3) ++(130:1cm) coordinate (C3); \foreach \i in {1,...,19} { \path (origin)++(40:0.8*\i cm) coordinate (a\i); \path (origin)++(130:0.8*\i cm) coordinate (b\i); \path (a\i)++(130:4cm) coordinate (ca\i); \path (b\i)++(40:4cm) coordinate (cb\i); \draw[thin,gray] (a\i) -- (ca\i) (b\i) -- (cb\i); \draw (origin) ++(40:0.4)++(130:0.4) node {${i}$} ; \draw (origin) ++(40:0.4)++(130:1.2) node {${i+1}$} ; \draw (origin) ++(40:1.2)++(130:0.4) node {${i-1}$} ; } \end{tikzpicture} \quad \begin{tikzpicture}[scale=.7] \path (0,0) coordinate (origin); \draw[thick] (origin) --++(130:1*0.8) --++(40:1*0.8) --++(-50:1*0.8) --++(220:1*0.8) ; \clip (origin) (origin) --++(130:1*0.8) --++(40:1*0.8) --++(-50:1*0.8) --++(220:1*0.8) ; \path (40:1cm) coordinate (A1); \path (40:2cm) coordinate (A2); \path (40:3cm) coordinate (A3); \path (40:4cm) coordinate (A4); \path (130:1cm) coordinate (B1); \path (130:2cm) coordinate (B2); \path (130:3cm) coordinate (B3); \path (130:4cm) coordinate (B4); \path (A1) ++(130:3cm) coordinate (C1); \path (A2) ++(130:2cm) coordinate (C2); \path (A3) ++(130:1cm) coordinate (C3); \foreach \i in {1,...,19} { \path (origin)++(40:0.8*\i cm) coordinate (a\i); \path (origin)++(130:0.8*\i cm) coordinate (b\i); \path (a\i)++(130:4cm) coordinate (ca\i); \path (b\i)++(40:4cm) coordinate (cb\i); \draw[thin,gray] (a\i) -- (ca\i) (b\i) -- (cb\i); \draw (origin) ++(40:0.4)++(130:0.4) node {${i-1}$} ; \draw (origin) ++(40:0.4)++(130:1.2) node {${i}$} ; \draw (origin) ++(40:1.2)++(130:1.2) node {${i-1}$} ; } \end{tikzpicture} \quad \begin{tikzpicture}[scale=.7] \path (0,0) coordinate (origin); \draw[thick] (origin) --++(130:1*0.8) --++(40:1*0.8) --++(-50:1*0.8) --++(220:1*0.8) ; \clip (origin) (origin) --++(130:1*0.8) --++(40:1*0.8) --++(-50:1*0.8) --++(220:1*0.8) ; \path (40:1cm) coordinate (A1); \path (40:2cm) coordinate (A2); \path (40:3cm) coordinate (A3); \path (40:4cm) coordinate (A4); \path (130:1cm) coordinate (B1); \path (130:2cm) coordinate (B2); \path (130:3cm) coordinate (B3); \path (130:4cm) coordinate (B4); \path (A1) ++(130:3cm) coordinate (C1); \path (A2) ++(130:2cm) coordinate (C2); \path (A3) ++(130:1cm) coordinate (C3); \foreach \i in {1,...,19} { \path (origin)++(40:0.8*\i cm) coordinate (a\i); \path (origin)++(130:0.8*\i cm) coordinate (b\i); \path (a\i)++(130:4cm) coordinate (ca\i); \path (b\i)++(40:4cm) coordinate (cb\i); \draw[thin,gray] (a\i) -- (ca\i) (b\i) -- (cb\i); \draw (origin) ++(40:0.4)++(130:0.4) node {${i+1}$} ; } \end{tikzpicture} \end{center} \caption{\cite[Figure~10]{bs15}, depicting the bricks $\mathbf{B}_1$, $\mathbf{B}_2$, $\mathbf{B}_3$, $\mathbf{B}_4$, $\mathbf{B}_5$ respectively. The $\mathbf{B}_6$ brick is the `empty brick'.} \label{bricksleft} \end{figure*} \tikzset{wei/.style={red,double=red,double distance=0.5pt}} \tikzset{wei2/.style={red,double=red,double distance=0.5pt}} \begin{figure}[h] \captionsetup{labelformat=empty} \begin{center}\scalefont{0.6} \begin{tikzpicture}[scale=1] \path (0,0) coordinate (origin); \draw[wei2] (0,0) circle (2pt); \draw[thick] (origin) --++(130:10*0.4) --++(40:1*0.4) --++(-50:1*0.4) --++(40:2*0.4) --++(-50:3*0.4) --++(40:1*0.4) --++(-50:1*0.4) --++(-50:1*0.4)--++(40:1*0.4) --++(40:1*0.4) --++(-50:1*0.4) --++(40:1*0.4) --++(-50:1*0.4) --++(40:1*0.4) --++(-50:1*0.4) --++(40:2*0.4) --++(-50:1*0.4) --++(220:10*0.4); \clip (origin) --++(130:10*0.4) --++(40:1*0.4) --++(-50:1*0.4) --++(40:2*0.4) --++(-50:3*0.4) --++(40:1*0.4) --++(-50:1*0.4) --++(-50:1*0.4)--++(40:1*0.4) --++(40:1*0.4) --++(-50:1*0.4) --++(40:1*0.4) --++(-50:1*0.4) --++(40:1*0.4) --++(-50:1*0.4) --++(40:2*0.4) --++(-50:1*0.4) --++(220:10*0.4); \path (40:1cm) coordinate (A1); \path (40:2cm) coordinate (A2); \path (40:3cm) coordinate (A3); \path (40:4cm) coordinate (A4); \path (130:1cm) coordinate (B1); \path (130:2cm) coordinate (B2); \path (130:3cm) coordinate (B3); \path (130:4cm) coordinate (B4); \path (A1) ++(130:3cm) coordinate (C1); \path (A2) ++(130:2cm) coordinate (C2); \path (A3) ++(130:1cm) coordinate (C3); \foreach \i in {1,...,19} { \path (origin)++(40:0.4*\i cm) coordinate (a\i); \path (origin)++(130:0.4*\i cm) coordinate (b\i); \path (a\i)++(130:4cm) coordinate (ca\i); \path (b\i)++(40:4cm) coordinate (cb\i); \draw[thin,gray] (a\i) -- (ca\i) (b\i) -- (cb\i); \draw (origin)++(40:0.2)++(130:0.2) node {${0}$} ; \draw (origin) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${0}$} ; \draw (origin) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${0}$} ; \draw (origin) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${0}$} ; \draw (origin) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${0}$} ; \draw (origin) ++(40:0.4) ++(40:0.2)++(130:0.2) node {${4}$} ; \draw (origin) ++(40:0.4)++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${4}$} ; \draw (origin) ++(40:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${4}$} ; \draw (origin) ++(40:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${4}$} ; \draw (origin) ++(40:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${4}$} ; \draw (origin) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \draw (origin) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \draw (origin) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \draw (origin) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \draw (origin) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \path (origin) ++(130:2cm) coordinate (XXXXXX); \draw (XXXXXX)++(40:0.2)++(130:0.2) node {${0}$} ; \draw (XXXXXX) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${0}$} ; \draw (XXXXXX) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${0}$} ; \draw (XXXXXX) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${0}$} ; \draw (XXXXXX) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${0}$} ; \draw (XXXXXX) ++(40:0.4) ++(40:0.2)++(130:0.2) node {${4}$} ; \draw (XXXXXX) ++(40:0.4)++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${4}$} ; \draw (XXXXXX) ++(40:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${4}$} ; \draw (XXXXXX) ++(40:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${4}$} ; \draw (XXXXXX) ++(40:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${4}$} ; \draw (XXXXXX) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \draw (XXXXXX) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \draw (XXXXXX) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \draw (XXXXXX) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \draw (XXXXXX) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \draw (XXXXXX)++(40:0.2)++(130:0.2)++(-50:0.4) node {${4}$} ; \draw (XXXXXX)++(40:0.2)++(130:0.2)++(130:0.4*4) node {${4}$} ; \path (origin) ++(40:2cm) coordinate (XXXXXX); \draw (XXXXXX)++(40:0.2)++(130:0.2)++(220:0.4) node {${1}$} ; \draw (XXXXXX)++(40:0.2)++(130:0.2) node {${0}$} ; \draw (XXXXXX) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${0}$} ; \draw (XXXXXX) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${0}$} ; \draw (XXXXXX) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${0}$} ; \draw (XXXXXX) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${0}$} ; \draw (XXXXXX) ++(40:0.4) ++(40:0.2)++(130:0.2) node {${4}$} ; \draw (XXXXXX) ++(40:0.4)++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${4}$} ; \draw (XXXXXX) ++(40:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${4}$} ; \draw (XXXXXX) ++(40:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${4}$} ; \draw (XXXXXX) ++(40:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${4}$} ; \draw (XXXXXX) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \draw (XXXXXX) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \draw (XXXXXX) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \draw (XXXXXX) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \draw (XXXXXX) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.4) ++(130:0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \draw (origin) ++(40:9*0.4) ++(40:0.2)++(130:0.2) node {${1}$} ; \draw[wei2] (0,0) ++(130:9*0.4) --++ (130:0.4) --++ (40:0.4) --++ (-50:0.4) --++ (220:0.4); \draw[wei2] (0,0) ++(130:4*0.4) --++ (130:0.4) --++ (40:0.4) --++ (-50:0.4) --++ (220:0.4); \draw[wei2] (0,0) ++(130:5*0.4) --++ (130:2*0.4) --++ (40:0.4) --++ (-50:0.4) --++ (40:0.4) --++ (-50:0.4) --++ (220:0.4); \draw[wei2] (0,0) ++ (130:6*0.4) ++ (40:0.4) --++ (130:2*0.4) --++ (40:0.4) --++ (-50:0.4) --++ (40:0.4) --++ (-50:0.4) --++ (220:0.4); \draw[wei2] (0,0) ++ (130:7*0.4) ++ (40:2*0.4) --++ (130:2*0.4) --++ (40:0.4) --++ (-50:2*0.4) \draw[wei2] (0,0) --++ (130:2*0.4) --++ (40:0.4) --++ (-50:0.4) --++ (40:0.4) --++ (-50:0.4) --++ (220:2*0.4); \draw[wei2] (0,0) ++ (130:0.4) ++ (40:0.4) --++ (130:2*0.4) --++ (40:0.4) --++ (-50:0.4) --++ (40:0.4) --++ (-50:0.4) --++ (220:0.4); \draw[wei2] (0,0) ++ (130:2*0.4) ++ (40:2*0.4) --++ (130:2*0.4) --++ (40:0.4) --++ (-50:0.4) --++ (40:0.4) --++ (-50:0.4) --++ (220:0.4); \draw[wei2] (0,0) ++ (130:3*0.4) ++ (40:3*0.4) --++ (130:2*0.4) --++ (40:0.4) --++ (-50:0.4) --++ (40:0.4) --++ (-50:0.4) --++ (220:0.4); \draw[wei2] (0,0) ++ (40:4*0.4) --++ (130:0.4) --++ (40:0.4) --++ (-50:0.4) --++ (220:0.4); \draw[wei2] (0,0) ++ (40:5*0.4) --++ (130:2*0.4) --++ (40:0.4) --++ (-50:0.4) --++ (40:0.4) --++ (-50:0.4) --++ (220:2*0.4); \draw[wei2] (0,0) ++ (40:6*0.4) ++(130:0.4) --++ (130:2*0.4) --++ (40:0.4) --++ (-50:0.4) --++ (40:0.4) --++ (-50:0.4) --++ (220:0.4); \draw[wei2] (0,0) ++ (40:9*0.4) --++ (130:0.4) --++ (40:0.4) --++ (-50:0.4) --++ (220:0.4); } \end{tikzpicture} \end{center} \caption{\cite[Figure~9]{bs15}, with individual bricks highlighted. The partition $\gamma=(10,9^2,6,4^2,3,2,1^2)$ with $e=5$, with the diagonals of interest for $i=0$. The diagram features two $i$-diagonals to the left of the leading diagonal, the leading diagonal itself (with the `empty brick' denoted by the red circle), and two $i$-diagonals to the right of leading diagonal.} \label{level1} \end{figure} Since each diagonal is characterised by which bricks sit at their base and top, along with how many $\mathbf{B}_1$ bricks they contain, which we denote by $b_1$, we may present this information compactly, giving the information for each $i$-diagonal in the Young diagram, read from left-to-right in the mirrored-Russian convention above. In fact, all that is needed from the $\mathbf{B}_1$ bricks turns out to be a sign. To each diagonal we associate the symbol $(-1)^{b_1}\mathbf{d}^{j}_{k}$, where the subscript $k=4,5,$ or $6$ if the bottom brick is $\mathbf{B}_4,\mathbf{B}_5$ or $\mathbf{B}_6$, respectively and where the superscript $j=2$ or $3$ if the top brick is $\mathbf{B}_2$ or $\mathbf{B}_3$, respectively, and $j=0$ if there is an addable $i$-node there. Finally, we encode these symbols as a string, denoted $\chi(\gamma)$, by reading the symbol for each $i$-diagonal in $\gamma$, from left to right. For example, looking at the diagram above for $e=5$, $i=0$, and $\gamma = (10,9^2,6,4^2,3,2,1^2)$, we may read off that \[ \chi(\gamma) = (\mathbf{d}^{0}_{4},\mathbf{d}^{3}_{4},\mathbf{d}^{0}_{6},\mathbf{d}^{0}_{5},\mathbf{d}^{0}_{5}). \] In \cite[Definition~4.26]{bs15}, some local relations are introduced, that may be applied to the symbols of $\chi(\gamma)$. We do not recall the exact relations, just that the most important one for us in our proof below allows us to turn a $\pm \mathbf{d}^{j}_{4}$ symbol into a $\mp \mathbf{d}^{j}_{5}$ symbol, and vice versa. We write $\chi(\gamma) \sim \chi(\bar\gamma)$ if $\chi(\gamma)$ may be obtained from $\chi(\bar\gamma)$ by applying a sequence of these relations. Finally, we may introduce the important results. They concern the diagrammatic Cherednik algebras, $A(n,\theta,\kappa)$, first introduced by Webster. If one of the parameters -- the \emph{weighting} -- is chosen to be \emph{well-separated}, then the resulting (graded) algebras are Morita equivalent to cyclotomic $q$-Schur algebras. These algebras have subquotients $A_\Gamma$ (defined in \cite[Definition~3.5]{bs15}) whose cell modules are indexed by those ($\ell$-multi)partitions obtained from $\gamma$ by added some fixed number of $i$-nodes, and it is shown that the graded decomposition numbers for $A_\Gamma$ match those labelled by the same (multi)partitions for the diagrammatic Cherednik algebra, and therefore for the (cyclotomic) Hecke algebra. \begin{thmc}{bs15}{Theorem~4.28}\label{thm:bs4.28} Suppose $\gamma$ and $\overline\gamma$ are two (multi)partitions that are $i$- and $\overline i$-admissible, respectively, for quantum characteristics $e$ and $\bar{e}$, respectively, with corresponding subquotients $A_{\Gamma}$ and $A_{\overline{\Gamma}}$, respectively. If $\chi(\gamma) \sim \chi(\overline\gamma)$, then $A_{\Gamma} \cong A_{\overline{\Gamma}}$ as graded $\bbf$-algebras; the isomorphism is given explicitly in \cite[Proposition~4.11]{bs15}. \end{thmc} The surprising nature of this isomorphism is that it allows us to compare subquotients of algebras, and corresponding graded decomposition numbers, that may have different quantum characteristics, ranks, and even levels. The original proof of the following theorem contains a typo and a minor oversight, making it a little unclear, so we provide a clarification below. For the notation in the explicit formula, the reader is invited to check \cite[Section~2]{bs15}. \begin{thmc}{bs15}{Theorem~4.30}\label{thm:bs4.30} Let $\gamma$ be an $i$-admissible multipartition and suppose that the $e$-multicharge $\kappa\in(\bbz/e\bbz)^\ell$ contains $i\in\bbz/e\bbz$ as a constituent with multiplicity 0 or 1. The graded decomposition numbers of $A(n,\theta,\kappa)$ over $\bbf$ can be given in terms of nested sign sequences as follows: \[ d^{e,p}_{\la\mu}(v) = \sum_{\mathclap{\omega \in \Omega(\la',\mu')}} \; v^{\|\omega\|} \] for $\la,\mu\in\Gamma$ such that $\la \domby \mu$. \end{thmc} \begin{proof} Under the condition on the multicharge, $\chi(\gamma)$ is a sequence which has a maximum of one entry $\pm \mathbf{d}_6^{j}$ for $j=0,2,3$. One can apply the local relations to $\chi(\gamma)$ to obtain some $\chi(\gamma) \sim \overline{\chi}$ that consists of three parts, in order: $(i)$ a sequence of $\pm \mathbf{d}_4^j$ symbols for $j=0,2,3$; $(ii)$ either one or zero $\pm \mathbf{d}_6^j$ symbols for $j=0,2,3$; and $(iii)$ a sequence of $\pm \mathbf{d}_5^j$ symbols for $j=0,2,3$ in order. Then there clearly exists a (level 1!) partition $\overline{\gamma}$ such that for $\bar{e} = \nchar(\bbf)$, $\chi(\gamma) \sim \overline{\chi} \sim \chi(\overline\gamma)$. Applying \cref{thm:bs4.28} yields the result for $d^{e,0}_{\la\mu}(v)$ by applying \cite[Theorem~4.4]{tanteo13}, and for $d^{e,p}_{\la\mu}(1)$ by applying \cite[Theorem~4.6]{tanteo13}, where this formula gives the graded decomposition numbers for the classical Schur algebras in characteristic $p$. The result now follows by applying \cref{lem:charfree}. \end{proof} We will actually only use a special case of the above, so we give the graded decomposition number explicitly in this case below. In characteristic $0$, it is a special case of~\cite[Theorem~4.2]{cmt08}, which in turn is a Hecke algebra analogue of a result of Kleshchev for the symmetric group~\cite{klesh1box}, though Kleshchev's (ungraded) result applies in positive characteristic. If the quantum characteristic $e$ and the field characteristic $p$ coincide (i.e.~our Hecke algebra is just the symmetric group), then the result is a special case of \cite[Theorem~4.6]{tanteo13}. The positive characteristic graded version we need, where $e$ and $p$ need not be equal, is a very special case of \cref{thm:bs4.30}, but nonetheless one that does not follow readily from other results in the literature. \begin{thm}\label{thm:kleshdecomp} Suppose that $\la \domsby \mu$ are two partitions, with $\mu$ being $e$-regular, that differ by moving a single node of some residue $i$ from row $r$ of $\mu$ to row $r+s$ of $\la$, and that there are no addable or removable $i$-nodes in rows $r+1, r+2,\dots, r+s-1$. Then $d_{\la\mu}^{e,p}(v) = v$. \end{thm} Our next result will ensure that our calculations for weight 2 and 3 blocks may be characteristic-free, so long as the characteristic is larger than the weight of the block. For $w=2$, $3$, or $4$, the following is proved in \cite{richardswt2}, \cite{fay08wt3} and \cite{fay07wt4}, respectively. \begin{thm}\label{thm:jamesconj} \label{thm:jamesconjsmallwt} If $B$ is a block of weight $w \leq 4$, then the adjustment matrix is the identity matrix whenever $p>w$. \end{thm} Note that Low has also extended the above results to $q$-Schur algebras~\cite{lowwt34}. In weight 2, the characteristic $2$ situation has been solved by Fayers; we will use the following two results in \cref{sec:wt2}, in which we employ the pyramid notation introduced in \cref{subsec:abacus}. \begin{thmc}{faywt2}{Corollary 2.4}\label{thm:wt2adjust} Let $\nu$ and $\mu$ be $e$-regular partitions in a weight $2$ block of $\hhh$, and let $p=2$. Then \[ a_{\nu\mu}(v) = \begin{cases} 1 &\text{ if } \nu = \langle i^2 \rangle, \ \mu =\langle i\rangle, \ _{i-1}0_i \text{ and } _i0_{i+1} \text{ for } 1\leq i \leq e-1;\\ 1 &\text{ if } \nu = \langle i^2 \rangle, \ \mu =\langle i, i+1\rangle, \ _{i-1}0_i \text{ and } _i1_{i+1} \text{ for } 1\leq i < e-1;\\ \delta_{\la\mu} &\text{ otherwise.}\\ \end{cases} \] \end{thmc} Recall that the Mullineux map $m$ is a bijection on the set of $e$-regular partitions satisfying $\D\la \otimes \sgn \cong \D{m(\la)}$ if $e=p$ (so that $\hhh \cong \bbf\sss$), and an analogous statement for the Hecke algebra, corresponding to twisting simple modules by a certain sign automorphism of $\hhh$. There are several combinatorial algorithms to compute $m$, due to several different authors, but we will not need these here. The following result will be useful for us later. \begin{propc}{fay09}{Lemma 3.6 and Proposition 3.7}\label{prop:degrees} Let $\la$ and $\mu$ be partitions of $n$, and suppose that $\mu$ is $e$-regular. Then $d_{\la\mu}^{e,0}(v) = 0$ unless $\mu \dom \la \dom m(\mu)'$, where $m$ is the Mullineux map. Moreover, if $\la = m(\mu)'$, then $d_{\la\mu}^{e,0}(v) = v^w$, where $w$ is the weight of the block containing $\la$ and $\mu$, and $d_{\la\mu}^{e,0}(v)$ has degree at most $w-1$ otherwise. \end{propc} In \cite{faywt2}, a partition $\la$ is said to be \emph{adjacent} to an $e$-regular partition $\mu$ if $\mu \dom \la \dom m(\mu)'$ and $|\partial\la - \partial\mu| = 1$, where $\partial$ is defined in \cite[Section 1.4]{faywt2}. We will not require this definition, and since we only consider situations where both $\la$ and $\mu$ are $e$-regular we will say that $\la$ and $\mu$ are adjacent, without worrying about which is more dominant. Just above \cite[Theorem~4.4]{richardswt2}, Richards defines a map $\diamond$ on $e$-regular weight two partitions that satisfies $\partial \mu^{\diamond'} = \partial \mu$, and on p399 of loc.~cit., Richards shows that this map agrees with the Mullineux map. Then by \cite[Theorem~4.4]{richardswt2}, combined with \cref{prop:degrees}, $\la$ and $\mu$ are adjacent if and only if $d_{\la\mu}^{e,0}(v) = v$, where $\mu \doms \la \doms m(\mu)'$. \begin{thmc}{faywt2}{Theorem 3.2}\label{thm:wt2ext} Suppose $p=2$, $B$ is a weight two block and $\la$, $\mu$ are $e$-regular partitions in $B$. Then: \begin{itemize} \item If neither of $\la$ and $\mu$ is of the form $\langle i^2 \rangle$ for some $i$ such that $_{i-1}0_i$, then $\operatorname{Ext}^1(\D\la, \D\mu) \cong \bbf$ if $\la$ and $\mu$ are adjacent, and is trivial otherwise. \item Suppose $\la = \langle i^2 \rangle$ for some $i$ such that $_{i-1}0_i$. Then $\operatorname{Ext}^1(\D\la, \D\mu) \cong \bbf$ if $\mu = \langle i \rangle$ with $_i0_{i+1}$ or if $\mu = \langle i, i+1 \rangle$ with $_i1_{i+1}$, and is trivial otherwise. \end{itemize} \end{thmc} \subsection{Runner removal}\label{subsec:runnerem} Our next result is a runner removal theorem of James and Mathas, though we use a useful reformulation of the statement due to Fayers~\cite[Theorem~2.15]{fay07wt4}. \begin{thmc}{jm02}{Theorem 3.2}\label{thm:runnerrem} Suppose $e\geq 3$ and that $\la$ and $\mu$ are partitions of $n$, $\mu$ is $e$-regular, and that we take abacus displays for $\la$ and $\mu$. Suppose that for some $i$, the last bead on runner $i$ occurs before every unoccupied space on both abacus displays, and define two abacus displays with $e-1$ runners by deleting runner $i$ from the abacus displays of $\la$ and $\mu$. Let $\la^-$ and $\mu^-$ be the partitions defined by these displays. If $\mu^-$ is $(e-1)$-regular, then \[ d^{e,0}_{\la\mu}(v) = d^{e-1,0}_{\la^- \mu^-}(v). \] \end{thmc} \subsection{Row-removal}\label{subsec:rowrem} The following so-called row- and column-removal theorems will be used in \cref{sec:highwt} for determining that the principal blocks of weight at least 4 are Schurian-infinite. \begin{thmc}{cmt02}{Theorem 1}\label{thm:rowremFock} Let $\la = (\la_1, \la_2, \dots)$ and $\mu = (\mu_1, \mu_2, \dots)$. \begin{enumerate}[label=(\roman*)] \item If $\la_1 + \la_2 + \dots + \la_r = \mu_1 + \mu_2 + \dots + \mu_r$ for some $r$, and we let \begin{alignat*}3 \la^{(0)} &= (\la_1, \la_2, \dots, \la_r ), \qquad &&\mu^{(0)} = (\mu_1, \mu_2, \dots, \mu_r),\\ \la^{(1)} &= (\la_{r+1}, \la_{r+2}, \dots ), \qquad &&\mu^{(1)} = (\mu_{r+1}, \mu_{r+2}, \dots), \end{alignat*} then $d_{\la\mu}^{e,0}(v) = d_{\la^{(0)}\mu^{(0)}}^{e,0}(v) d_{\la^{(1)}\mu^{(1)}}^{e,0}(v)$. \item If $\la_1' + \la_2' + \dots + \la_r' = \mu_1' + \mu_2' + \dots + \mu_r'$ for some $r$, and we let \begin{alignat*}3 \la^{(0)} &= (\min(\la_1,r), \min(\la_2,r), \dots), \qquad &&\mu^{(0)} = (\min(\mu_1,r), \min(\mu_2,r), \dots),\\ \la^{(1)} &= (\max(\la_1-r,0), \max(\la_2-r,0), \dots ), \qquad &&\mu^{(1)} = (\max(\mu_1-r,0), \max(\mu_2-r,0), \dots), \end{alignat*} then $d_{\la\mu}^{e,0}(v) = d_{\la^{(0)}\mu^{(0)}}^{e,0}(v) d_{\la^{(1)}\mu^{(1)}}^{e,0}(v)$. \end{enumerate} \end{thmc} \begin{thmc}{Donkin}{4.2(9) and 4.2(15)}\label{thm:rowremDecomp} If $\la$, $\mu$, $\la^{(0)}$, $\mu^{(0)}$, $\la^{(1)}$, and $\mu^{(1)}$ are as in either case above, then $d_{\la\mu}^{e,p}(1) = d_{\la^{(0)}\mu^{(0)}}^{e,p}(1) d_{\la^{(1)}\mu^{(1)}}^{e,p}(1)$. \end{thmc} \begin{rem} In the special case of $r=1$, this was proved earlier, in \cite[Theorems~1 and 2]{j81}. \end{rem} \subsection{Scopes equivalences}\label{subsec:scopes} In several places, we will make use of certain Morita equivalences between blocks of Hecke algebras, known as Scopes equivalences. Scopes introduced these for the symmetric groups in~\cite{scopes}, and this theory was easily generalised to the Hecke algebras by Jost~\cite{jost}. First, note that the definition of abacus that we are using is convenient for our purpose, but usually a truncated version of this abacus is used, where we do not allow the runners to extend infinitely upwards. In truncating the abacus, we are essentially forcing an abacus display to have finitely many beads. In this setting, it is common to then use beta-numbers $\beta_i = \la_i - i + r$, for $r\geq \la'_1$, to yield an abacus display with $r$ beads. We may adjust $r$ in order to give a clean description of the Scopes equivalence. So a given partition can have many different abacus displays. When applying the Scopes equivalence, we number the runners of a chosen abacus display $0, 1, \dots, e-1$, so that position $i$ is on runner $i$ for each $i=0,1,\dots,e-1$. Suppose that $B$ is a block of $\hhh$ of weight $w$, and core $\kappa$, and that for some $i$ an abacus display for $\kappa$ (or equivalently, for any partition in $B$) has $k$ more beads on runner $i$ than on runner $i-1$, for some $k\geq w$. Let $A$ be the block of $\hhh[n-k]$ of weight $w$ and core $\Phi(\kappa)$, whose abacus display is obtained from that of $\kappa$ by swapping runners $i$ and $i-1$. We may define this map $\Phi$ in the same way for any partition $\la \in B$. That is, we swap runners $i$ and $i-1$ of the corresponding abacus display for $\la$, yielding a partition $\Phi(\la) \in A$. If $i=0$, we actually want to swap runners $0$ and $e-1$, and in doing so we need $k+1$ more beads on runner $0$ than runner $e-1$. We will favour changing $r$ to avoid the need for this exceptional treatment. For example, taking (core) partitions $(2^2,1^2)$ and $(2,1^2)$, with $e=3$, we have the following two abacus displays and their truncations. Choosing $r=9$, we have \[ \abacus(vvv,bbb,bbn,bbn,bbn,nnn,vvv) \quad \leftrightarrow \quad \abacus(lmr,bbb,bbn,bbn,bbn,nnn,vvv) \] while $r=10$ gives us the below abacus display, for which we may apply the map $\Phi$ as depicted. \[ \abacus(vvv,bbb,bbb,nbb,nbb,nnn,vvv) \quad \leftrightarrow \quad \abacus(lmr,bbb,bbb,nbb,nbb,nnn,vvv) \qquad \xrightarrow[]{\phantom{aa} \Phi \phantom{aa}} \qquad \abacus(lmr,bbb,bbb,bnb,bnb,nnn,vvv) \quad \leftrightarrow \quad \abacus(vvv,bbb,bbb,bnb,bnb,nnn,vvv) \] Scopes showed that this $\Phi$ is a bijection from $B$ to $A$, that maps $e$-regular partitions to $e$-regular partitions. We say that such a pair of blocks \emph{forms a $[w:k]$ pair}. Note that many authors do not assume that $k\geq w$ when using this terminology, unlike in the present paper. \begin{thmc}{jost}{Theorem 7.3}\label{thm:scopes} If $B$ and $A$ are blocks of $\hhh$ and $\hhh[n-k]$ as above, forming a $[w:k]$ pair for $k\geq w$, then they are Morita equivalent, and the equivalence is realised via $\D\la \leftrightarrow \D{\Phi(\la)}$. In particular, if $\la$ and $\mu$ are partitions of $n$, with $\mu$ $e$-regular, then for any $p\geq0$, $d_{\la\mu}^{e,p}(1) = d_{\Phi(\la)\Phi(\mu)}^{e,p}(1)$. \end{thmc} \subsection{Reduction theorem and Jantzen filtration}\label{subsec:reduction} \begin{prop}\label{reduction} Let $A$ be a Schurian-finite algebra. \begin{enumerate}[label=(\roman*)] \item If $B$ is a factor algebra of $A$, then $B$ is Schurian-finite. \item If $B=eAe$, for an idempotent $e\in A$, then $B$ is Schurian-finite. \end{enumerate} \end{prop} \begin{proof} (i) is obvious. (ii) follows from \cite[Proposition~2.4]{Ad2}. \end{proof} \begin{cor}\label{cor:Schurianinfinitequiver} If the Gabriel quiver of a finite-dimensional algebra $A$ over $\bbf$ contains the quiver of an affine Dynkin diagram with zigzag orientation (i.e.~such that every vertex is a sink or a source) as a subquiver, then $A$ is Schurian-infinite. \end{cor} \begin{proof} We may assume that $A$ is a basic algebra without loss of generality. Then, $A\cong \bbf Q/I$, for a finite quiver $Q$ and an admissible ideal $I\subseteq \bbf Q$. Let $e\in A$ be the sum of the idempotents associated with the vertices of the subquiver. Then, since the path algebra of the subquiver we consider here is radical square zero, we have a surjective algebra homomorphism from $eAe$ to the path algebra of the subquiver. The latter has preprojective and preinjective components with infinitely many vertices, by \cite[Chap.~VIII, Corollary~2.3]{ASS1}, each of which are Schurian, by \cite[Chap.~VIII, Lemma~2.7]{ASS1}. Finally, \cref{reduction} implies that $A$ is Schurian-infinite. \end{proof} The following is Shan's theorem \cite[Theorem~0.1]{ShanJantzenfilt}. \begin{thmc}{ShanJantzenfilt}{Theorem~0.1}\label{thm:Shan} Suppose that $q=\exp(-2\pi i/e)\in \bbc$ with $e\geq 3$. Let $\la, \mu$ be partitions of $n$ and the modules $W(\la')$ and $L(\mu')$ over the $q$-Schur algebra $S_q(n,n)$ are the Weyl module with highest weight $\la'$ and the irreducible module with highest weight $\mu'$, respectively. Then, the graded decomposition numbers are given by the Jantzen filtration of $W(\la')$ as follows. \[ d^{e,0}_{\la\mu}(v) = \sum_{i\geq 0} [J^iW(\la')/J^{i+1}W(\la'):L(\mu')] v^i \] \end{thmc} Furthermore, as is pointed out in \cite[Remark~6.6]{ShanJantzenfilt}, the radical filtration of $W(\la')$ coincides with the Jantzen filtration by the dual statement of \cite[Lemma~5.2.2]{BB93} and \cite[Proposition~5.5]{ShanJantzenfilt} which proves that the geometric Jantzen filtration gives rise to the Jantzen filtration of $W(\la')$. The assumption of \cite[Lemma~5.2.2]{BB93} for finite-dimensional Schubert varieties is known to hold. In particular, if $[J^1W(\la')/J^{2}W(\la'):L(\mu')] \neq 0$, then $W(\la')$ has a quotient that is a uniserial module of length two whose head is $L(\la')$ and whose socle is $L(\mu')$. \begin{lem}\label{lem:uniserial in char=0} Suppose that $q=\exp(-2\pi i/e)\in \bbc$ with $e\geq 3$ as above. If $\la, \mu$ are $e$-regular partitions of $n$ and the coefficient of $v$ in $d^{e,0}_{\la\mu}(v)$ is nonzero, then $\operatorname{Ext}^1(\D\la, \D\mu) = \operatorname{Ext}^1(\D\mu,\D\la)\neq 0$. \end{lem} \begin{proof} By the remark after \cref{thm:Shan}, $L(\mu')$ appears in $\operatorname{Rad}(W(\la'))/\operatorname{Rad}^2(W(\la'))$. Since $\mu'$ is $e$-restricted, the Schur functor sends $W(\lambda')$ to the dual Specht module $\rrspe{\la'} \cong \spe\la\otimes\sgn$ and $L(\mu')$ to $\D{\mu}\otimes\sgn$. That $\D\la$ is the unique head of $\spe\la$ implies that $\D\mu$ appears in the second layer of the radical series of $\spe\la$. \end{proof} \begin{prop}\label{prop:matrixtrick} Suppose that $e\geq3$ and $\bbf$ has characteristic $p\geq 0$. If a submatrix of the graded decomposition matrix in characteristic $0$ is one of the following matrices, and $d_{\la\mu}^{e,p}(1) = d_{\la\mu}^{e,0}(1)\in \{0,1\}$ holds, for all $e$-regular partitions $\la, \mu$ that label rows of the submatrix, then the block in which those partitions belong is Schurian-infinite. \begin{multicols}{2} \begin{equation*}\label{targetmatrix}\tag{\(\dag\)} \begin{pmatrix} 1\\ v & 1\\ 0 & v & 1\\ v & v^2 & v & 1 \end{pmatrix} \end{equation*} \begin{equation*}\label{targetmatrix2}\tag{\(\dag'\)} \begin{pmatrix} 1\\ v & 1\\ v^2 & v & 1\\ v & 0 & v & 1 \end{pmatrix} \end{equation*} \begin{equation*}\label{targetmatrix3}\tag{\(\dag''\)} \begin{pmatrix} 1\\ v & 1\\ v^2 & v & 1\\ v & v^2 & v & 1 \end{pmatrix} \end{equation*} \begin{equation*}\label{targetmatrixalt}\tag{\(\ddag\)} \begin{pmatrix} 1\\ v & 1\\ v & 0 & 1\\ v^2 & v & v & 1 \end{pmatrix} \end{equation*} \begin{equation*}\label{targetmatrixaltsquare}\tag{\(\clubsuit\)} \begin{pmatrix} 1\\ 0 & 1\\ v & v & 1\\ v & v & 0 & 1 \end{pmatrix} \end{equation*} \begin{equation*}\label{targetmatrixstar}\tag{\(\spadesuit\)} \begin{pmatrix} 1\\ 0 & 1\\ v & v & 1\\ 0 & v^2 & v & 1\\ v^2 & 0 & v & 0 &1 \end{pmatrix} \end{equation*} \end{multicols} \end{prop} \begin{proof} Let $B$ and $B_\bbf$ be the block in characteristic $0$ or $p$ in which the four or the five partitions that label the rows and columns of the submatrix belong, and denote the $e$-regular partitions by $\la^{(i)}$, for $1\leq i\leq 4$ or $1\leq i\leq 5$. Then, \cref{lem:uniserial in char=0} implies that we have an $A^{(1)}_3$ quiver (square) or a $D^{(1)}_4$ quiver (4-pointed star) with zigzag orientation as a subquiver of the Gabriel quiver of $B$. By \cref{lem:charfree}, $d_{\la\mu}^{e,p}(v) = d_{\la\mu}^{e,0}(v)$. Let $\D\mu$ and $\D\mu_\bbf$ be the simple $B$-module and the simple $B_\bbf$-module labeled by $\mu$, respectively. We denote by $\spe\la_\bbf$ the Specht $B_\bbf$-module labeled by $\la$. Now we consider the modular reduction of $\D\la/\operatorname{Rad}^2(\D\la)$ as a factor module of $\spe\la_\bbf$. If $d_{\la\mu}^{e,0}(v) = v^2$, then $[\spe\la/\operatorname{Rad}^2(\spe\la):\D\mu] = 0$. Then, $d_{\la\mu}^{e,p}(1) = d_{\la\mu}^{e,0}(1)\in \{0,1\}$ implies that $\D\mu_\bbf$ does not appear in the modular reduction of $\D\la/\operatorname{Rad}^2(\D\la)$ as a composition factor. The same argument shows that $\D\la_\bbf$ appears with multiplicity $1$ as the unique head of the modular reduction of $\D\la/\operatorname{Rad}^2(\D\la)$, and $\D\mu_\bbf$ appears with multiplicity $1$ as one of the composition factors of the modular reduction if $d^{e,0}_{\la\mu}(v) = v$, and all the other composition factors of the modular reduction are $\D\nu_\bbf$ where $\nu$ is not among the four or five partitions $\la^{(i)}$. Therefore, we may obtain an indecomposable $B_\bbf$-module that has the unique head $\D\la_\bbf$ and submodules $\D\mu_\bbf$ for $\mu$ with $d^{e,0}_{\la\mu}(v)=v$, and all the other composition factors of the module are $\D\nu_\bbf$ for partitions $\nu$ that are not among those we have labelled by $\la^{(i)}$. Let $P_\bbf^\mu$ be the projective cover of $\D\mu_\bbf$ and let $t$ be the sum of the idempotents in the basic algebras of $B_\bbf$ that are projectors to $P_\bbf^{\la^{(i)}}$, summing over all $i$. Then $t$ kills every simple module $\D\nu_\bbf$ that is not labelled by some $\la^{(i)}$. Since $\operatorname{Rad}(t{B_\bbf}t)=t\operatorname{Rad}(B_\bbf)t$, it follows that we have $\operatorname{Ext}^1(t\D\la_\bbf, t\D\mu_\bbf)\ne0$, for simple $t{B_\bbf}t$-modules $t\D\la_\bbf$ and $t\D\mu_\bbf$ with $d^{e,0}_{\la\mu}(v) = v$. This implies that the Gabriel quiver of $t{B_\bbf}t$ contains either $A^{(1)}_3$ or $D^{(1)}_4$ quiver with zigzag orientation, so that the Gabriel quiver of $B_\bbf$ contains one of them as a subquiver. We apply \cref{cor:Schurianinfinitequiver} to conclude that $B_\bbf$ is Schurian-infinite. \end{proof} \section{Graded Scopes Equivalence}\label{sec:gradedScopes} We consider the Scopes equivalence (c.f.~\cref{subsec:scopes}) in the graded setting. Throughout this section we assume that either $p \nmid e$ or $p = e$, so that there is a block of some Hecke algebra that this is isomorphic to. Suppose that $B$ and $A$ form a $[w : k]$ pair, and consider its graded version, which we denote by the cyclotomic KLR algebra $R^{\La_0}(\beta)$ and $R^{\La_0}(\beta-k\alpha_i)$, respectively. We may define the Scopes restriction and induction functors in the graded setting to be the (graded) cyclotomic divided powers $e_i^{(k)}$ and $f_i^{(k)}$ (c.f.~\cite[Section 4.6]{bk09}), but we take a different approach. Recall that if a pair of blocks $B$ and $A$ forms a $[w:k]$ pair, then the ungraded Scopes restriction functor is the composition of four functors \cite{jost}, and that the ungraded Scopes induction functor is its adjoint functor. We may consider the graded version of the Scopes restriction functor. Define $\mathcal{N}_k$ to be the graded algebra with generators $u_j$ with $\deg u_j=-2$, $1\leq j\leq k-1$, obeying $u_j^2=0$ and the braid relations. Let $\bbf[x_1,\dots,x_k]$ be the polynomial ring with degree given by $\deg x_a=2$. Let $e_j$, for $1\leq j\leq k$, be the elementary symmetric polynomials in $x_1,\dots,x_k$, and $I = (e_1,\dots,e_k)$ the ideal generated by them. Then, $\mathcal{N}_k\otimes \bbf[x_1,\dots,x_k]$, with the relations \[ u_jx_a = x_au_j\;(a\neq j,j+1), \quad u_jx_{j+1}-x_ju_j=1=x_{j+1}u_j-u_jx_j \] is a graded algebra that is isomorphic to $R(k\alpha_i)$. We denote by $Y$ the $R(k\alpha_i)$-module realised on $\bbf[x_1,\dots,x_k]/I$, where $x_a$ acts by multiplication and $u_j$ acts as the divided difference $\partial_i = (x_{i+1} - x_i)^{-1}(1-s_i)$. The graded $R(k\alpha_i)$-module $Y$ is the unique irreducible graded $R(k\alpha_i)$-module up to shift. Let $\hhh[k]$ be the subalgebra of $R(k\alpha_i)$ generated by $T_i = u_i x_i - q x_i u_i$, for $1\leq i\leq k-1$. We consider $\hhh[k]$ as a graded algebra concentrated in degree $0$, and may view any $R(k\alpha_i)$-module as a graded $\hhh[k]$-module by restriction. Then, $Y$ is a graded $\hhh[k]$-module whose degree zero component is isomorphic to the sign module $\spe{(1^k)}$. We denote by $P$ the projective cover of the $\hhh[k]$-module $\spe{(1^k)}$ concentrated in degree zero. Let $e_{\beta-k\alpha_i, k\alpha_i}\in R^{\La_0}(\beta)$ be the sum of the idempotents $e(\res(T)*i^k)$, where $T$ runs through standard tableaux of partitions of $n-k$ which belong to the block $A$. Then, for a graded $R^{\La_0}(\beta)$-module $M$, $e_{\beta-k\alpha_i,k\alpha_i} M$ is a graded $R^{\La_0}(\beta-k\alpha_i)\boxtimes \hhh[k]$-module. \begin{defn}\label{def:grscopes} The graded Scopes functor is the functor from the category of finite-dimensional graded $R^{\La_0}(\beta)$-modules to the category of finite-dimensional graded $R^{\La_0}(\beta-k\alpha_i)$-modules given by \[ M \longmapsto \Hom_{\hhh[k]}(P, e_{\beta-k\alpha_i,k\alpha_i} M), \] where homomorphism are degree preserving homomorphisms. This is an exact functor. \end{defn} Since $w\leq k$, every partition $\la$ that belongs to $R^{\La_0}(\beta)$ satisfies $\varepsilon_i(\la)=k$, as was proved in the proof of \cite[Lemma~2.1]{scopes}, so that the graded Scopes functor is a subfunctor of $e_i^{\rm max}=e_i^k$. The correspondence $\la\mapsto \Phi(\la)$ is nothing but $\la\mapsto \tilde{e}_i^{\rm max}\la=\tilde{e}_i^{k}\la$. \begin{prop}\label{prop:grScopes} Suppose $B$ and $A$ form a $[w:k]$ pair. \begin{enumerate}[label=(\roman*)] \item The graded Scopes functor sends graded Specht module $\spe\la$ to $\spe{\Phi(\la)}$, for partitions $\la$. \item The graded Scopes functor sends graded simple module $\D\la$ to $\D{\Phi(\la)}$, for $e$-regular partitions $\la$. \item If we forget the grading, the graded Scopes functor coincides with the ungraded Scopes functor defined in \cite{jost}. \item $d^{e,p}_{\la\mu}(v) = d^{e,p}_{\Phi(\la)\Phi(\mu)}(v)$, for any $p\geq0$. \item The category of finite-dimensional graded $B$-modules is category equivalent to the category of finite-dimensional graded $A$-modules. \end{enumerate} \end{prop} \begin{proof} \begin{enumerate}[label=(\roman*)] \item We see that $e_{\beta-k\alpha_i,k\alpha_i} \spe\la\cong \spe{\Phi(\la)}\otimes Y$ by~\cite[Corollary~5.8]{Mathas17}. Then, $\Hom_{\hhh[k]}(P, Y) = \Hom_{\hhh[k]}(P, \spe{(1^k)}) = \bbf$ and \[ \Hom_{\hhh[k]}(P, \spe{\Phi(\la)}\otimes Y)\cong \spe{\Phi(\la)}\otimes \Hom_{\hhh[k]}(P, Y)\cong \spe{\Phi(\la)} \] follows. \item Note that the composition factors of $\spe\la$ are $\D{\mu}$ with $\mu \doms \la$, up to shift. Since $\la\mapsto \Phi(\la)$ respects the lexicographic order~\cite[Lemma~2.2]{scopes}, and $d^{e,p}_{\la\mu}(1) = d^{e,p}_{\Phi(\la)\Phi(\mu)}(1)$, for any $p\geq0$ \cite[Lemma~5.3]{jost}, we have the result for the graded Scopes functor by using (i) and induction on the reverse lexicographic order. \item Recall the definition of Jost's Scopes functor. After restricting an $\hhh$-module to $\hhh[n-k]\boxtimes \hhh[k]$, we tensor it with the ungraded right $\hhh[k]$-module ${\spe{(k)}}^*$. He showed that the restriction of $\D\la$ is isomorphic to $\D{\Phi(\la)}\boxtimes \hhh[k]$ \cite[Corollary~6.3]{jost}. Hence, the restricted module viewed as an $\hhh[k]$-module is free of finite rank. \begin{itemize} \item[(a)] If we restrict an ungraded $R^{\La_0}(\beta)$-module to $\hhh[n-k]\boxtimes \hhh[k]$, and view it as an $R^{\La_0}(\beta-k\alpha_i)\boxtimes \hhh[k]$-module through the algebra homomorphism \[ R^{\La_0}(\beta-k\alpha_i)\boxtimes \hhh[k] \hookrightarrow \hhh[n-k]\boxtimes \hhh[k] \hookrightarrow \hhh, \] each composition factor $\D\la$ changes to $\D{\Phi(\la)}\otimes \hhh[k]$, and tensoring it with ${\spe{(k)}}^*$ over $\hhh[k]$ has the effect that we replace $\hhh[k]$ with $\bbf$. \item[(b)] If we restrict a graded $R^{\La_0}(\beta)$-module to $R^{\La_0}(\beta-k\alpha)\boxtimes \hhh[k]$, each composition factor $\D{\la}\langle d\rangle$ changes to $\D{\Phi(\la)}\langle d\rangle\otimes Y$, and taking the space of degree preserving $\hhh[k]$-module homomorphisms from $P$ has the effect that we replace $Y$ with $\bbf$. \end{itemize} Comparing (a) and (b), we know that our graded Scopes functor coincides with Jost's, if we forget the grading. More precisely, we prove the next result. \begin{quote} Let $\Phi_1$ be our functor defined in \cref{def:grscopes}, and let $\Phi_2$ be Jost's functor. Then, ${\rm For}\circ\Phi_1\cong \Phi_2\circ {\rm For}$, where ${\rm For}$ is the forgetful functor. \end{quote} \medskip For any $R^{\La_0}(\beta)$-module $M$, we have the inclusion of graded $R^{\La_0}(\beta-k\alpha_i)$-modules \[ \Phi_1(M) = (e_{\beta-k\alpha_i,k\alpha_i}M)_0 \longrightarrow M, \] where $(e_{\beta-k\alpha_i,k\alpha_i}M)_0$ is the degree $0$ part of $e_{\beta-k\alpha_i,k\alpha_i}M$ with respect to the graded $R(k\alpha_i)$-module structure. We choose a composition series \[ 0=M_0\subset M_1\subset \cdots \subset M_l=M. \] Then, $0 \to M_{l-1} \to M \to \D{\la}\langle d\rangle \to 0$, for some $\la$ and $d$. We show by induction on the length of modules that the inclusion above induces ${\rm For}\circ\Phi_1(M)\cong\Phi_2\circ {\rm For}(M)$. Assume that ${\rm For}\circ\Phi_1(M_k)\cong \Phi_2\circ {\rm For}(M_k)$ holds for $1\leq k\leq l-1$, and we restrict the module $M$ to $R^{\La_0}(n-k)\boxtimes R(k)$. Then, we have the commutative diagram of $\hhh[n-k]$-modules \[ \xymatrix@C=1em{ 0 \ar[rr] && {\rm For}(e_{\beta-k\alpha_i,k\alpha_i}M_{l-1})_0 \ar[rr] \ar[d] && {\rm For}(e_{\beta-k\alpha_i,k\alpha_i}M)_0 \ar[rr] \ar[d] && {\rm For}(e_{\beta-k\alpha_i,k\alpha_i}\D\la\langle d\rangle)_0 \ar[rr] \ar[d] && 0 \\ 0 \ar[rr] && {\rm For}(M_{l-1}) \ar[rr] \ar[d] && {\rm For}(M) \ar[rr] \ar[d] && {\rm For}(\D{\la}) \ar[rr] \ar[d] && 0 \\ && {\spe{(k)}}^*\otimes_{\hhh[k]} {\rm For}(M_{l-1}) \ar[rr] \ar[d] && {\spe{(k)}}^*\otimes_{\hhh[k]} {\rm For}(M) \ar[rr] \ar[d] && {S^{(k)}}^*\otimes_{\hhh[k]} ({\rm For}(\D{\Phi(\la)})\boxtimes \hhh[k]) \ar[rr] \ar[d] && 0 \\ 0 \ar[rr] && \Phi_2( {\rm For}(M_{l-1})) \ar[rr] && \Phi_2({\rm For}(M)) \ar[rr] && \Phi_2( {\rm For}(\D\la\langle d\rangle)) \ar[rr] && 0 } \] where the first vertical arrow is the inclusion, the third vertical arrow is the block truncation. Note that ${\rm For}(\D\la)$ restricts to ${\rm For}(\D{\Phi(\la)})\boxtimes \hhh[k]$. Hence, we have the commutative diagram \[ \xymatrix@C=1em{ 0 \ar[rr] && {\rm For}(\Phi_1(M_{l-1})) \ar[rr] \ar[d] && {\rm For}(\Phi_1(M)) \ar[rr] \ar[d] && {\rm For}(\Phi_1(\D\la\langle d\rangle)) \ar[rr] \ar[d] && 0 \\ 0 \ar[rr] && \Phi_2( {\rm For}(M_{l-1})) \ar[rr] && \Phi_2({\rm For}(M)) \ar[rr] && \Phi_2( {\rm For}(\D\la\langle d\rangle)) \ar[rr] && 0 } \] such that the left vertical arrow and the right vertical arrow are isomorphisms by the induction hypothesis. By the five lemma, we deduce that ${\rm For}\circ\Phi_1(M)\cong \Phi_2\circ{\rm For}(M)$. \item Since $d^{e,p}_{\la\mu}(1) = d^{e,p}_{\Phi(\la)\Phi(\mu)}(1)$, for any $p\geq0$, this follows from (i) and (ii). \item We have to show that the graded Scopes functor is fully faithful and dense. Since the ungraded Scopes functor induces category equivalence, (iii) implies the fully faithfulness. To see that it is dense, we argue by induction on the length of graded modules. Suppose that \[ 0 \rightarrow \D{\Phi(\la)} \rightarrow M \rightarrow N \rightarrow 0 \] and that $L$ maps to $N$ under the graded Scopes functor. Choose the element in \linebreak $\operatorname{EXT}^1(N,\D{\Phi(\la)})$ which represents this extension. Since $\operatorname{Ext}^1(N,D^{\Phi(\la)})\cong \operatorname{Ext}^1(L,D^\la)$ by the ungraded Scopes equivalence~\cite[Theorem~7.3]{jost}, we may choose the corresponding element in $\operatorname{EXT}^1(L,D^\la)$ which gives a short exact sequence of graded modules \[ 0 \rightarrow D^\la \rightarrow X \rightarrow L \rightarrow 0. \] Then, it induces $0 \rightarrow D^{\Phi(\la)} \rightarrow X' \rightarrow N \rightarrow 0$, where $X'$ is the image of $X$ under the graded Scopes functor, and we have an isomorphism $X'\cong M$ if we forget the grading. This isomorphism is a sum of homomorphisms of various degrees. However, since the homomorphisms in the short exact sequences are all degree $0$ homomorphisms, this isomorphism is pure of degree $0$. Hence the graded Scopes functor is a dense functor. \qedhere \end{enumerate} \end{proof} \begin{rem} Since the indecomposable direct summands of the regular module of $R(k\alpha_i)$ are $\bbf[x_1,\dots,x_k]\langle -2j \rangle$, where $0\leq j\leq k(k-1)/2$, and \[ f_i^kN = R^{\La_0}(\beta)e_{\beta-k\alpha_i,k\alpha_i}\otimes_{R^{\La_0}(\beta-k\alpha_i)} N \cong R^{\La_0}(\beta)e_{\beta-k\alpha_i,k\alpha_i}\otimes_{R^{\La_0}(\beta-k\alpha_i)\boxtimes R(k\alpha_i)} N\otimes R(k\alpha_i), \] we may define the graded Scopes induction functor by \[ N \longmapsto R^{\La_0}(\beta)e_{\beta-k\alpha_i,k\alpha_i}\otimes_{R^{\La_0}(\beta-k\alpha_i)\boxtimes R(k\alpha_i)} N\otimes \bbf[x_1,\dots,x_k]\langle -k(k-1)\rangle. \] This is an exact functor. Moreover, \cite[Corollary~4.7]{hm12} implies that the graded Scopes functor sends $\spe{\Phi(\la)}$ to $\spe\la$. Unlike the Scopes restriction functor, we may generalise Scopes and Jost's induction functor in a straightforward manner. Namely, we may define \[ N \longmapsto R^{\La_0}(\beta)e_{\beta-k\alpha_i,k\alpha_i}\otimes_{R^{\La_0}(\beta-k\alpha_i)\boxtimes R(k)} N\otimes \spe{(k)}. \] We have not checked whether it coincides with our definition (up to shift) or not. \end{rem} \section{Weight 2 blocks}\label{sec:wt2} In this section, we will prove \cref{thm:wt2}. We will follow the notations and conventions of the note of Fayers~\cite{fayerswt2data}. This will allow us to work in a great deal of generality. As we noted in the remark before the statement of \cref{thm:wt2}, if $e=2$ and $p\neq 2$, weight 2 blocks of the Hecke algebra are known to be Schurian-finite, and we are unable to handle the case $e=p=2$. So we will assume throughout this paper that $e\geq 3$. Originally, we were able to prove \cref{thm:wt2} for $p\neq 2$ by applying the results of~\cite{fayerswt2data}, but the proof we present here extends more readily into the weight $3$ case. By~\cref{thm:jamesconj}, when $w=2$, $d_{\la\mu}^{e,0}(v) = d_{\la\mu}^{e,p}(v)$ for all $p\neq 2$, so we may rely on characteristic 0 results throughout this section, unless $p=2$. Thus, we also assume that $p\neq 2$ in this section, except where explicitly stated, but otherwise we allow $e$ and $p$ to be arbitrary. We work across five separate cases, depending on the abacus for the core indexing a given block, finding the submatrices, (\ref{targetmatrix}), (\ref{targetmatrixalt}), or (\ref{targetmatrixstar}) from \cref{prop:matrixtrick} in each case. The runner-removal result in \cref{thm:runnerrem} will be key to reducing our necessary computations to known small rank cases. \subsection{$p_{e-1} - p_{e-3} < e$}\label{subsec:wt2firstcase} First, suppose that we are in a block such that $p_{e-1} - p_{e-3} < e$, and $p\neq2$. We will obtain the $4\times4$ submatrix of the graded decomposition matrix for the block, with rows and columns indexed by the partitions $\langle e-1\rangle$, $\langle e-2\rangle$, $\langle e-3\rangle$, and $\langle e-3, e-1\rangle$, and see that it is (\ref{targetmatrix}). It is clear that for these four partitions, we may apply the runner removal result \cref{thm:runnerrem} to remove all but runners $e-3$, $e-2$, and $e-1$, leaving the abacus configuration of the core as one of those depicted below. \[ \abacus(vvv,bbb,bbb,nnn,nnn,vvv) \qquad\qquad \abacus(vvv,bbb,bbb,bnn,nnn,vvv) \qquad\qquad \abacus(vvv,bbb,bbb,bbn,nnn,vvv) \] We place $p_{e-3}$ on the leftmost runner. Then, we may, without loss of generality, work with the first display above. The corresponding four partitions then become $(6)$, $(5,1)$, $(4,1^2)$, and $(3,2,1)$, from which we may deduce that the corresponding submatrix (for $e=3$) is known to be (\ref{targetmatrix}) -- for instance, see \cite[Appendix~B]{mathas}. The result now follows.\\ Now, if $p=2$, we may directly apply \cref{thm:wt2ext}; note that we do so for our four partitions on the $e$-runner abacus, as we cannot appeal to a runner removal result here. None of our partitions are of the form $\langle i^2 \rangle$ for any $i$, and $\la$ and $\mu$ are adjacent whenever $d_{\la\mu}^{e,0}(v) = v$, so that we get the same subquiver of the Gabriel quiver of $B$ in characteristic $2$ as characteristic $0$ -- i.e.~we have an $A^{(1)}_3$ quiver (square) with zigzag orientation as a subquiver. Alternatively, we may apply \cref{thm:wt2adjust}. Since \[ d_{\la\mu}^{e,2}(v) = d_{\la\mu}^{e,0}(v) + \sum_{\nu \domsby \mu} d_{\la\nu}^{e,0}(v) a_{\nu\mu}(v), \] and $d_{\la\nu}^{e,0}(v) = 0$ unless $\la \domby \nu$, it suffices to note that no $\langle e-3, e-1\rangle \domby \nu \domsby \langle e-1 \rangle$ can be of the form $\langle i^2 \rangle$, and thus that every summand on the right has either $d_{\la\nu}^{e,0}(v) = 0$ or $a_{\nu\mu}(v) = 0$. In other words, the submatrix we found in other characteristics is identical in characteristic $2$, and we may apply \cref{prop:matrixtrick}. \subsection{$p_{e-1} - p_{e-2} < e$ and $p_{e-2} - p_{e-3} < e$, but $p_{e-1} - p_{e-3} > e$}\label{subsec:wt2secondcase} Next, suppose that $p_{e-1} - p_{e-2} < e$ and $p_{e-2} - p_{e-3} < e$, but $p_{e-1} - p_{e-3} > e$. We will obtain the $4\times4$ submatrix of the graded decomposition matrix for the block, with rows and columns indexed by the partitions $\langle e-1\rangle$, $\langle e-2\rangle$, $\langle e-2, e-1\rangle$, and $\langle (e-1)^2\rangle$, and see that it is (\ref{targetmatrix}). As in \cref{subsec:wt2firstcase}, it is clear that for these four partitions, we may apply the runner removal result \cref{thm:runnerrem} to remove all but runners $e-3$, $e-2$, and $e-1$, leaving the abacus configuration of the core as depicted below, where we place $p_{e-3}$ on the leftmost runner. Then it is easy to see that $p_{e-2}$ cannot be on the middle runner. \[ \abacus(vvv,bbb,bbb,nbn,nnn,vvv) \] The corresponding four partitions then become $(7)$, $(5,2)$, $(4,3)$, and $(4,2,1)$, from which we may deduce that the corresponding submatrix (for $e=3$) is known to be (\ref{targetmatrix}) -- for instance, see \cite[Appendix~B]{mathas}. The result now follows.\\ Now, if $p=2$, we may directly apply \cref{thm:wt2ext} as in \cref{subsec:wt2firstcase} -- only one of our partitions is of the form $\langle i^2 \rangle$ for any $i$ (that is, when $i=e-1$), but we do not have $_{e-2}0_{e-1}$. Again, $\la$ and $\mu$ are adjacent whenever $d_{\la\mu}^{e,0}(v) = v$, so that we get the same subquiver of the Gabriel quiver in characteristic $2$ as characteristic $0$ -- i.e.~an $A^{(1)}_3$ quiver (square) with zigzag orientation. Alternatively, we may apply \cref{thm:wt2adjust} as in \cref{subsec:wt2firstcase}. Now, note that the only $\langle (e-1)^2 \rangle \domby \nu \domsby \langle e-1 \rangle$ of the form $\langle i^2 \rangle$ is $\langle (e-1)^2 \rangle$, and thus that every summand on the right has either $d_{\la\nu}^{e,0}(v) = 0$ or $a_{\nu\mu}(v) = 0$, except the summand with $\nu = \langle (e-1)^2 \rangle$, which only occurs if $\la = \langle (e-1)^2 \rangle$ to. In this case, if $\mu$ is any one of our four possible partitions, the above becomes \begin{align*} d_{\langle (e-1)^2 \rangle\mu}^{e,2}(v) &= d_{\langle (e-1)^2 \rangle\mu}^{e,0}(v) + d_{\langle (e-1)^2 \rangle \langle (e-1)^2 \rangle}^{e,0}(v) a_{\langle (e-1)^2 \rangle \mu}(v)\\ &= d_{\langle (e-1)^2 \rangle\mu}^{e,0}(v) + a_{\langle (e-1)^2 \rangle \mu}(v)\\ &= d_{\langle (e-1)^2 \rangle\mu}^{e,0}(v) \text{ by \cref{thm:wt2adjust}, as $p_{e-1}-p_{e-2}<e$, or $_{e-2}1_{e-1}$}. \end{align*} In other words, the submatrix we found in other characteristics is again identical in characteristic $2$. \subsection{$p_{e-1} - p_{e-2} > e$ and $p_{e-2} - p_{e-3} < e$}\label{subsec:wt2thirdcase} Next, suppose that $p_{e-1} - p_{e-2} > e$ and $p_{e-2} - p_{e-3} < e$. We will obtain the $4\times4$ submatrix of the graded decomposition matrix for the block, with rows and columns indexed by the partitions $\langle e-1\rangle$, $\langle e-2, e-1\rangle$, $\langle e-2\rangle$, and $\langle e-3\rangle$, and see that it is (\ref{targetmatrix}). As in \cref{subsec:wt2firstcase,subsec:wt2secondcase}, it is clear that for these four partitions, we may apply the runner removal result \cref{thm:runnerrem} to remove all but runners $e-3$, $e-2$, and $e-1$, leaving abacus displays with $e=3$. We must have either $p_{e-2} = p_{e-3} + 1$ and $p_{e-1} = p_{e-2} + (k-1)e + 1$ or $p_{e-2} = p_{e-3} + 2$ and $p_{e-1} = p_{e-2} + (k-1)e + 2$, for some $k\geq 2$. The corresponding $k+3$ bead abacus displays are depicted below. \[ \abacus(vvv,bbb,bnn,bnn,vvv,bnn,bnn,nnn,vvv) \qquad\qquad \abacus(vvv,bbb,nbn,nbn,vvv,nbn,nbn,nnn,vvv) \] Here, the diagrams have $k$ more beads on the first and second runners, respectively, than on the other two. Because of this, swapping the $0$- and $1$-runners (i.e.~the leftmost and middle runners) or the $1$- and $2$-runners (i.e.~the middle and rightmost runners) always involves swapping runners on which the number of beads differ by at least $w=2$. For $k\geq 2$, we let $\rho^{(2k-3)} := (2k-2, 2k-4,\dots, 2)$ denote the partition with abacus display on the left above (satisfying $p_{e-2} = p_{e-3} + 1$ and $p_{e-1} = p_{e-2} + (k-1)e + 1$) and $\rho^{(2k-2)} := (2k-1,2k-3,\dots,1)$ denote the partition with abacus display on the right above (satisfying $p_{e-2} = p_{e-3} + 2$ and $p_{e-1} = p_{e-2} + (k-1)e + 2$) Then our first few partitions are $\rho^{(1)}=(2)$, $\rho^{(2)}=(3,1)$, $\rho^{(3)}=(4,2)$, $\rho^{(4)}=(5,3,1)$, and $\rho^{(5)}=(6,4,2)$. It is clear that swapping the $0$- and $1$-runners swaps $\rho^{(2k-3)}$ and $\rho^{(2k-2)}$, while swapping the $1$- and $2$-runners of the above abacus display for $\rho^{(2k-2)}$ yields a $k+3$ bead abacus display for $\rho^{(2k-1)}$. But $\rho^{(2k-1)}$ has a $(k+1)+3$ bead abacus display as above, so that we can again swap the $0$- and $1$-runners to obtain $\rho^{(2k)}$, and so on. Then we see that all possible cores in this case are among our partitions $\rho^{(i)}$, and they may be recursively obtained from $\rho^{(1)} = (2)$ by swapping adjacent runners, where one has $k\geq 2 = w$ more beads than the others. So we may apply \cref{prop:grScopes}, and reduce our computation to the case of the block with core $(2)$. Notice that when we swap the $0$- and $1$-runners to interchange $\rho^{(2k-3)}$ and $\rho^{(2k-2)}$, this corresponds to performing $(1-k)$-induction or $(1-k)$-restriction on the core (where we take the residue of $1-k$ modulo $3$). Similarly, when swapping the $1$- and $2$-runners to interchange $\rho^{(2k-2)}$ and $\rho^{(2k-1)}$, this corresponds to performing $(2-k)$-induction or $(2-k)$-restriction on the core. Thus, we may alternatively characterise our runner swaps above as applying (maximal powers of) Kashiwara operators recursively to the cores, and the partitions in those blocks, as follows. For $i \geq 0$, \begin{align*}\label{eq:inducecase3cores} \begin{split} \rho^{(6i+2)} = \tilde{f_2}^{3i+2} \rho^{(6i+1)}, \quad \rho^{(6i+3)} &= \tilde{f_0}^{3i+2} \rho^{(6i+2)}, \quad \rho^{(6i+4)} = \tilde{f_1}^{3i+3} \rho^{(6i+3)},\\ \rho^{(6i+5)} = \tilde{f_2}^{3i+3} \rho^{(6i+4)}, \quad \rho^{(6i+6)} &= \tilde{f_0}^{3i+4} \rho^{(6i+5)}, \quad \rho^{(6i+7)} = \tilde{f_1}^{3i+4} \rho^{(6i+6)}. \end{split} \end{align*} Since Scopes equivalences preserve our labelling of Specht modules and simple modules when we reduce to the block with core $(2)$, we have for this block that $\langle e-1\rangle = (8)$, $\langle e-2, e-1\rangle = (5,2,1)$, $\langle e-2\rangle = (4,3,1)$, and $\langle e-3\rangle = (3^2,1^2)$, from which we may deduce that the corresponding submatrix is known to be (\ref{targetmatrix}) -- for instance, see \cite[Appendix~B]{mathas}. The result now follows.\\ Now, if $p=2$, we may argue as in \cref{subsec:wt2firstcase}, by directly applying \cref{thm:wt2ext} -- none of our partitions are of the form $\langle i^2 \rangle$ for any $i$, and $\la$ and $\mu$ are adjacent whenever $d_{\la\mu}^{e,0}(v) = v$, so that we get the same subquiver of the Gabriel quiver in characteristic $2$ as characteristic $0$ -- i.e.~an $A^{(1)}_3$ quiver (square) with zigzag orientation. \subsection{$p_{e-1} - p_{e-2} < e$ and $p_{e-2} - p_{e-3} > e$}\label{subsec:wt2fourthcase} Next, suppose that $p_{e-1} - p_{e-2} < e$ and $p_{e-2} - p_{e-3} > e$. We will obtain the $4\times4$ submatrix of the graded decomposition matrix for the block, with rows and columns indexed by the partitions $\langle e-1\rangle$, $\langle e-2\rangle$, $\langle e-2, e-1\rangle$, and $\langle (e-1)^2\rangle$, and see that it is (\ref{targetmatrix}). This case proceeds analogously to the one handled in \cref{subsec:wt2thirdcase}. As before, it is clear that for these four partitions, we may apply the runner removal result \cref{thm:runnerrem} to remove all but runners $e-3$, $e-2$, and $e-1$, leaving abacus displays with $e=3$. We must have either $p_{e-2} = p_{e-3} + (k-1)e + 1$ and $p_{e-1} = p_{e-2} + 1$ or $p_{e-2} = p_{e-3} + (k-1)e + 2$ and $p_{e-1} = p_{e-2} + 2$, for some $k\geq 2$. The corresponding $(2k+3)$ bead abacus displays are depicted below. \[ \abacus(vvv,bbb,bbn,bbn,vvv,bbn,bbn,nnn,vvv) \qquad\qquad \abacus(vvv,bbb,bnb,bnb,vvv,bnb,bnb,nnn,vvv) \] Here, the diagrams have $k$ more beads on the leftmost and middle runners than the rightmost, or $k$ more beads on the leftmost and rightmost runners than the middle one, respectively. Because of this, swapping the $1$- and $2$-runners (i.e.~the middle and rightmost runners) or the $0$- and $1$-runners (i.e.~the leftmost and middle runners) always involves swapping runners on which the number of beads differ by at least $w=2$. For $k\geq 2$, we let $\sigma^{(2k-3)} = (\rho^{(2k-3)})' = ((k-1)^2, (k-2)^2,\dots, 1^2)$ denote the partition with abacus display on the left above (satisfying $p_{e-2} = p_{e-3} + (k-1)e + 1$ and $p_{e-1} = p_{e-2} + 1$) and $\sigma^{(2k-2)} = (\rho^{(2k-2)})' = (k,(k-1)^2, (k-2)^2,\dots, 1^2)$ denote the partition with abacus display on the right above (satisfying $p_{e-2} = p_{e-3} + (k-1)e + 2$ and $p_{e-1} = p_{e-2} + 2$). Then our first few partitions are $\sigma^{(1)}=(1^2)$, $\sigma^{(2)}=(2,1^2)$, $\sigma^{(3)}=(2^1,1^2)$, $\sigma^{(4)}=(3,2^2,1^2)$, and $\sigma^{(5)}=(3^2,2^2,1^2)$. It is clear that swapping the $1$- and $2$-runners swaps $\sigma^{(2k-3)}$ and $\sigma^{(2k-2)}$, while swapping the $0$- and $1$-runners of the above abacus display for $\sigma^{(2k-2)}$ yields a $2k+3$ bead abacus display for $\sigma^{(2k-1)}$. But $\sigma^{(2k-1)}$ has a $2(k+1)+3$ bead abacus display as above, so that we can again swap the $0$- and $1$-runners to obtain $\sigma^{(2k)}$, and so on. Then we see that all possible cores in this case are among our partitions $\sigma^{(i)}$, and they may be recursively obtained from $\sigma^{(1)} = (1^2)$ by swapping adjacent runners, where one has $k\geq 2 = w$ more beads than the others. So we may apply \cref{prop:grScopes}, and reduce our computation to the case of the block with core $(1^2)$. Notice that when we swap the $0$- and $1$-runners to interchange $\sigma^{(2k-3)}$ and $\sigma^{(2k-2)}$, this corresponds to performing $k$-induction on the core (where we take the residue of $k$ modulo $3$). Similarly, when swapping the $1$- and $2$-runners to interchange $\sigma^{(2k-2)}$ and $\sigma^{(2k-1)}$, this corresponds to performing $(k+1)$-induction on the core. Thus, we may alternatively characterise our runner swaps above as applying (maximal powers of) Kashiwara operators recursively to the cores, and the partitions in those blocks, as follows. \begin{align*}\label{eq:inducecase4cores} \begin{split} \sigma^{(6i+2)} = \tilde{f_1}^{3i+2} \sigma^{(6i+1)}, \quad \sigma^{(6i+3)} &= \tilde{f_0}^{3i+2} \sigma^{(6i+2)}, \quad \sigma^{(6i+4)} = \tilde{f_2}^{3i+3} \sigma^{(6i+3)},\\ \sigma^{(6i+5)} = \tilde{f_1}^{3i+3} \sigma^{(6i+4)}, \quad \sigma^{(6i+6)} &= \tilde{f_0}^{3i+4} \sigma^{(6i+5)}, \quad \sigma^{(6i+7)} = \tilde{f_2}^{3i+4} \sigma^{(6i+6)}. \end{split} \end{align*} Noting, as before, that Scopes equivalences preserve our labelling of Specht modules and simple modules when we reduce to the block with core $(1^2)$, we have for this block that $\langle e-1\rangle = (7,1)$, $\langle e-2\rangle = (6,2)$, $\langle e-2, e-1\rangle = (4^2)$, and $\langle (e-1)^2\rangle = (4,2^2)$, from which we may deduce that the corresponding submatrix is known to be (\ref{targetmatrix}) -- for instance, see \cite[Appendix~B]{mathas}. The result now follows.\\ Now, if $p=2$, the exact same argument we used in \cref{subsec:wt2secondcase} shows that we get the same subquiver of the Gabriel quiver in characteristic $2$ as characteristic $0$ -- i.e.~an $A^{(1)}_3$ quiver (square) with zigzag orientation. \subsection{$p_{e-1} - p_{e-2} > e$ and $p_{e-2} - p_{e-3} > e$}\label{subsec:wt2fifthcase} Finally, suppose that $p_{e-1} - p_{e-2} > e$ and $p_{e-2} - p_{e-3} > e$. We will obtain the $5\times5$ submatrix of the graded decomposition matrix for the block, with rows and columns indexed by the partitions $\langle e-1\rangle$, $\langle (e-1)^2\rangle$, $\langle e-2, e-1\rangle$, $\langle e-2\rangle$, and $\langle (e-2)^2\rangle$, and see that it is (\ref{targetmatrixstar}). This case proceeds analogously to the one handled in \cref{subsec:wt2thirdcase}. As before, it is clear that for these five partitions, we may apply the runner removal result \cref{thm:runnerrem} to remove all but runners $e-3$, $e-2$, and $e-1$, leaving abacus displays with $e=3$. Then we have either $p_{e-1} = p_{e-2} + ke + 1$ and $p_{e-2} = p_{e-3} + je + 1$ or $p_{e-1} = p_{e-2} + ke + 2$ and $p_{e-2} = p_{e-3} + je + 2$, for some $j, k \geq 1$. In this way, we index the cores by the pair of integers $(ke+1,je+1)$ or $(ke+2, je+2)$, so we will denote such core by $\kappa_{(ke+1,je+1)}$ or $\kappa_{(ke+2,je+2)}$, as appropriate. We will draw the $(k+2j+3)$ bead abacus display for $\kappa_{(ke+1,je+1)}$ as below. \[ \abacus(vvv,bbb,nbb,nbb,vvv,nbb,nnb,nnb,vvv,nnb,vvv) \] Note that the number of beads on the $0$- and $1$-runners differ by $j$, while those on the $1$- and $2$-runners differ by $k$. Similarly, we draw the $(k+2j+5)$ bead abacus display for $\kappa_{(ke+2,je+2)}$ as below. \[ \abacus(vvv,bbb,bnb,bnb,vvv,bnb,nnb,nnb,vvv,nnb,vvv) \] Note that the number of beads on the $0$- and $1$-runners differ by $j+1$, while those on the $1$- and $2$-runners differ by $k+j+1$. It is easy to see that swapping the $1$- and $2$-runners on the above display yields a $(k+2j+5)$ bead abacus display for $\kappa_{(ke+1,je+1)}$. On the other hand, if we start with our $(k+2j+3)$ bead abacus display for $\kappa_{(ke+1,je+1)}$, swapping the $0$- and $1$-runners (which differ by $j$ beads) yields a $(k+2j+3)$ bead abacus display for $\kappa_{(ke+2,(j-1)e+2)}$, while swapping the $1$- and $2$-runners (which differ by $k$ beads) yields a $(k+2j+3)$ bead abacus display for $\kappa_{((k-1)e+2,je+2)}$. Thus we may use the graded Scopes equivalences of \cref{prop:grScopes} to pass between the blocks with these cores, as in \cref{subsec:wt2thirdcase,subsec:wt2fourthcase}. Noting, as before, that Scopes equivalences preserve our labelling of Specht modules and simple modules, it is clear that we can use Scopes equivalences to reduce all the way down to the block with core $\kappa_{(e+1,e+1)}$, for which the corresponding five partitions then become $(9,1^2)$, $(6,4,1)$, $(6,3,2)$, $(5,4,2)$, and $(3^2,2^2,1)$, from which we may deduce that the corresponding submatrix is known to be (\ref{targetmatrixstar}) -- for instance, see \cite[Appendix~B]{mathas}. Our result now follows.\\ Now, if $p=2$, we must use different partitions. If $e=3$, then the block is labelled by the $3$-core $(3,1^2)$. By applying \cref{thm:wt2ext} our Gabriel quiver is \[ \begin{tikzcd} {\langle (e-1)^2 \rangle} & {\langle e-1 \rangle} & {\langle e-2,e-1 \rangle} & {\langle e-2 \rangle} & {\langle (e-2)^2 \rangle} \arrow[from=1-1, to=1-2, shift left=.9ex, "\alpha_1" above] \arrow[from=1-2, to=1-1, shift left=.9ex, "\beta_1" below] \arrow[from=1-2, to=1-3, shift left=.9ex, "\alpha_2" above] \arrow[from=1-3, to=1-2, shift left=.9ex, "\beta_2" below] \arrow[from=1-3, to=1-4, shift left=.9ex, "\alpha_3" above] \arrow[from=1-4, to=1-3, shift left=.9ex, "\beta_3" below] \arrow[from=1-4, to=1-5, shift left=.9ex, "\alpha_4" above] \arrow[from=1-5, to=1-4, shift left=.9ex, "\beta_4" below] \end{tikzcd} \] so that our methods are unable to determine the Schurian-infiniteness of the block. We use a different method to prove that it is Schurian-infinite, so we will come back to this case at the end of the section. Until then, we will assume that $e\geq4$. If $_{e-4}0_{e-3}$, we take our partitions to be $\langle e-2, e-1 \rangle$, $\langle e-3, e-1 \rangle$, $\langle e-2 \rangle$, $\langle e-3, e-2 \rangle$. Then \cref{thm:wt2ext} gives us that \begin{multline*} \operatorname{Ext}^1(\D{\langle e-2, e-1 \rangle}, \D{\langle e-3, e-1 \rangle}) \cong \operatorname{Ext}^1(\D{\langle e-2, e-1 \rangle}, \D{\langle e-2 \rangle})\\ \cong \operatorname{Ext}^1(\D{\langle e-3, e-2 \rangle}, \D{\langle e-3, e-1 \rangle}) \cong \operatorname{Ext}^1(\D{\langle e-3, e-2 \rangle}, \D{\langle e-2 \rangle}) \cong \bbf, \end{multline*} provided we can show that the corresponding pairs of partitions are adjacent. In order to do this, we should again show that these partitions give us (\ref{targetmatrixalt}) as the corresponding submatrix of the characteristic $0$ graded decomposition matrix -- in fact it suffices to pick out the four $v$ entries. We may use \cref{thm:runnerrem} to reduce to the $e=4$ situation. Then we may argue as for $p\neq 2$, but this time we must keep track of $p_3-p_2$, $p_2 - p_1$, and $p_1 - p_0$ -- we will again denote the corresponding core by $\kappa_{p_3-p_2, p_2 - p_1, p_1 - p_0}$. Then since we can't ever have $p_i - p_j \equiv 0 \pmod 4$, it is not so difficult to check that all cores must fit into one of the following six families: \[ \kappa_{ie+3, je+3, ke+3}, \; \kappa_{ie+2, je+3, ke+2}, \; \kappa_{ie+1, je+2, ke+3}, \; \kappa_{ie+3, je+2, ke+1}, \; \kappa_{ie+2, je+1, ke+2}, \text{ or } \kappa_{ie+1, je+1, ke+1}, \] for $i,j,k \in \bbz_{>0}$. We may check that, as before, we may perform runner swaps to get between these blocks, and that these will always involve runners whose number of beads differ by at least $2$, so that we may again apply \cref{prop:grScopes} to reduce down to the block with core $\kappa_{e+1, e+1, e+1} = (6,3^2,1^3)$. In this block, we use the partitions $\langle e-2, e-1 \rangle = (10,6,4,1^3)$, $\langle e-3, e-1 \rangle = (10,3^3,2^2)$, $\langle e-2 \rangle = (9,7,4,1^3)$, and $\langle e-3, e-2 \rangle = (6^2,4,3,2^2)$, and the result follows. To see the runner swaps explicitly, we use the Kashiwara operators $\tilde{e}_i^{\rm max}$ on the cores, so that it is presented more cleanly. Then, taking subscripts modulo $e$, we have \begin{itemize} \item $\tilde{e}_{3i+2j+k+2}^{i+j+k+2} \kappa_{ie+3,je+3,ke+3} = \kappa_{ie+2, je+3, ke+2}$; \item $\tilde{e}_{3i+2j+k+1}^{i+j+1} \kappa_{ie+2,je+3,ke+2} = \kappa_{ie+1, je+2, ke+3}$ and $\tilde{e}_{3i+2j+k-1}^{j+k+1} \kappa_{ie+2,je+3,ke+2} = \kappa_{ie+3, je+2, ke+1}$; \item $\tilde{e}_{3i+2j+k-1}^{j+k+1} \kappa_{ie+1, je+2, ke+3} = \kappa_{ie+2, je+1, ke+2} = \tilde{e}_{3i+2j+k+1}^{i+j+1} \kappa_{ie+3, je+2, ke+1}$; \item $\tilde{e}_{3i+2j+k}^{i} \kappa_{ie+1, je+2, ke+3} = \kappa_{(i-1)e+3, je+3, ke+3}$, $\tilde{e}_{3i+2j+k-2}^{j} \kappa_{ie+2, je+1, ke+2} = \kappa_{ie+3, (j-1)e+3, ke+3}$, and\linebreak $\tilde{e}_{3i+2j+k}^{k} \kappa_{ie+3, je+2, ke+1} = \kappa_{ie+3, je+3, (k-1)e+3}$; \item $\tilde{e}_{3i+2j+k}^{i+j+k+1} \kappa_{ie+2, je+1, ke+2} = \kappa_{ie+1, je+1, ke+1}$; \item $\tilde{e}_{3i+2j+k-1}^{i} \kappa_{ie+1, je+1, ke+1} = \kappa_{(i-1)e+3, je+2, ke+1}$, $\tilde{e}_{3i+2j+k-2}^{j} \kappa_{ie+1, je+1, ke+1} = \kappa_{ie+2, (j-1)e+3, ke+2}$, and\linebreak $\tilde{e}_{3i+2j+k-3}^{k} \kappa_{ie+1, je+1, ke+1} = \kappa_{ie+1, je+2, (k-1)e+3}$. \end{itemize} Each of these may be easily deduced by analysing the beta-numbers for these cores. On the other hand, if $_{e-4}1_{e-3}$, we take our partitions to be $\langle e-2 \rangle$, $\langle e-3, e-2 \rangle$, $\langle e-3 \rangle$, $\langle e-4 \rangle$. Then \cref{thm:wt2ext} gives us that \[ \operatorname{Ext}^1(\D{\langle e-2 \rangle}, \D{\langle e-3, e-2 \rangle}) \cong \operatorname{Ext}^1(\D{\langle e-2 \rangle}, \D{\langle e-4 \rangle}) \cong \operatorname{Ext}^1(\D{\langle e-3 \rangle}, \D{\langle e-3, e-2 \rangle}) \cong \operatorname{Ext}^1(\D{\langle e-3 \rangle}, \D{\langle e-4 \rangle}) \cong \bbf, \] provided we can show that the corresponding pairs of partitions are adjacent. This latter point is proved in an analogous manner to the previous case, essentially setting $k=0$ in the previous case. Then Scopes equivalences reduce this to the case of the block with core $\kappa_{e+1,e+1,1} = (5,2^2)$, so that our four partitions are $\langle e-2 \rangle = (8,6,3)$, $\langle e-3, e-2 \rangle = (5^2,3,2,1^2)$, $\langle e-3 \rangle = (5,4,3^2,1^2)$, and $\langle e-4 \rangle = (5,3^3,1^3)$, giving the matrix (\ref{targetmatrix}). In both cases, we obtain an $A^{(1)}_3$ quiver (square) with zigzag orientation as a subquiver of the Gabriel quiver.\\ Finally, we return to the case $e=3$, recalling that up to Scopes equivalence, we are looking at the block with core $(3,1^2)$. As preparation, we first show that Specht modules in this block are all uniserial. The simples in this block are labelled by the five partitions $\la_1 := \langle 2 \rangle = (9,1^2)$, $\la_2 := \langle 2^2 \rangle = (6,4,1)$, $\la_3 := \langle 1,2 \rangle = (6,3,2)$, $\la_4 := \langle 1 \rangle = (5,4,2)$ and $\la_5 := \langle 1^2 \rangle = (3^2,2^2,1)$. The block also contains the $3$-singular partitions $\la_6 := (6,1^5)$, $\la_7 := (3^2,2,1^3)$, $\la_8 := (3,2^3,1^2)$, and $\la_9 := (3,1^8)$. Using the LLT algorithm and \cref{thm:wt2adjust}, we may deduce that the (ungraded) decomposition matrix is as below. \[ \begin{array}{c|ccccc} & \la_1 & \la_2 & \la_3 & \la_4 & \la_5 \\\hline \la_1 & 1 & \cdot & \cdot & \cdot & \cdot \\ \la_2 & 1 & 1 & \cdot & \cdot & \cdot \\ \la_3 & 2 & 1 & 1 & \cdot & \cdot \\ \la_4 & 1 & 1 & 1 & 1 & \cdot \\ \la_5 & 1 & 0 & 1 & 1 & 1 \\ \la_6 & 0 & 0 & 1 & 0 & 0 \\ \la_7 & 0 & 0 & 1 & 2 & 1 \\ \la_8 & 0 & 0 & 0 & 1 & 1 \\ \la_9 & 0 & 0 & 0 & 1 & 0 \end{array} \] The Mullineux map swaps $\la_1$ and $\la_4$, $\la_2$ and $\la_5$, and fixes $\la_3$. In the following, we use the terminology of string modules and band modules, for a general finite-dimensional algebra~\cite{stv21}. However, as we use left modules, some care must be taken. To define the string module $M(w)$, for a walk $w$ on the double quiver of the Gabriel quiver, we read the walk $w$ from right to left. For example, $M(\beta_2\beta_1)$ is the uniserial module such that the head is $\D{\la_2}$, the heart is $\D{\la_1}$ and the socle is $\D{\la_3}$. By the decomposition matrix, $\spe{\la_9}\cong \D{\la_4}$. As $\spe{\la_8}$ is a submodule of $P^{\la_5}$ and $[\spe{\la_8}] = [\D{\la_4}] + [\D{\la_5}]$, $\spe{\la_8}$ is the uniserial module of length two whose head is $\D{\la_4}$ and whose socle is $\D{\la_5}$. This is the string module $M(\beta_4)$. Next we consider $\spe{\la_7}$. As it is a submodule of $P^{\la_3}$, and $[S^{\la_7}/\operatorname{Soc}(\spe{\la_7})] = 2[\D{\la_4}] + [\D{\la_5}]$, $\operatorname{Ext}^1(\D{\la_4},\D{\la_4}) = 0$ and $\operatorname{Ext}^1(\D{\la_4},\D{\la_5})=\operatorname{Ext}^1(\D{\la_5},\D{\la_4}) = \bbf$, we must have that $\spe{\la_7}/\operatorname{Soc}(\spe{\la_7})\cong M(\alpha_4\beta_4)$, so that $\spe{\la_7}\cong M(\alpha_3\alpha_4\beta_4)$. Hence \[ \spe{\la_9}\cong \D{\la_4}, \quad \spe{\la_8}\cong M(\beta_4), \quad \spe{\la_7}\cong M(\alpha_3\alpha_4\beta_4). \] By similar arguments, we obtain \begin{gather*} \spe{\la_6}\cong \D{\la_3}, \quad \spe{\la_5}\cong M(\alpha_2\alpha_3\alpha_4), \quad \spe{\la_4}\cong M(\alpha_1\alpha_2\alpha_3), \\ \spe{\la_3}\cong M(\beta_1\alpha_1\alpha_2),\quad \spe{\la_2}\cong M(\beta_1), \quad \spe{\la_1}\cong \D{\la_1}. \end{gather*} We have Specht filtration $P^{\la_2} = V_0\supset V_1\supset V_2$ such that \[ V_0/V_1\cong \spe{\la_2}, \quad V_1/V_2\cong \spe{\la_3}, \quad V_2\cong \spe{\la_4}. \] By the shape of the Gabriel quiver, the head and the socle of $\operatorname{Rad}(P^{\la_2})/\operatorname{Soc}(P^{\la_2})$ is $\D{\la_1}$ and its composition factors are $4[\D{\la_1}]+[\D{\la_2}]+2[\D{\la_3}]+[\D{\la_4}]$. Since $\operatorname{Rad}(P^{\la_2})/\operatorname{Soc}(P^{\la_2})$ is self-dual, $\D{\la_i}$ appears in the head of $\operatorname{Rad}(P^{\la_2})/\operatorname{Soc}(P^{\la_2})$ only when $\D{\la_i}$ appears at least twice as the composition factor, so that $i=1$ or $i=3$, and we may rule out $i=3$ since $\operatorname{Ext}^1(\D{\la_2},\D{\la_3}) = 0$. Hence the head and the socle of $\operatorname{Rad}(P^{\la_2})/\operatorname{Soc}(P^{\la_2})$ is $D^{\la_1}$ and $\operatorname{Rad}^2(P^{\la_2})/\operatorname{Soc}^2(P^{\la_2})$ is self-dual with composition factors $2[\D{\la_1}] + [\D{\la_2}] + 2[\D{\la_3}] + [\D{\la_4}]$. Then, by the same argument as above, $\D{\la_i}$ appears in the head of $\operatorname{Rad}^2(P^{\la_2})/\operatorname{Soc}^2(P^{\la_2})$ only when $\D{\la_i}$ appears at least twice as the composition factor, so that $i=1$ or $i=3$, and we may rule out $i=1$ since $\operatorname{Ext}^1(\D{\la_1},\D{\la_1}) = 0$. Hence the head and the socle of $\operatorname{Rad}^2(P^{\la_2})/\operatorname{Soc}^2(P^{\la_2})$ is $\D{\la_3}$ and $\operatorname{Rad}^3(P^{\la_2})/\operatorname{Soc}^3(P^{\la_2})$ is self-dual. It follows that \[ \operatorname{Rad}^3(P^{\la_2})/\operatorname{Soc}^3(P^{\la_2})\cong M(\beta_1\alpha_1)\oplus \D{\la_4}. \] Therefore, relabelling the vertices of the Gabriel quiver as $1,2,\dots,5$ from left to right, we may identify the corresponding indecomposable projective module of the basic algebra of the block as \begin{gather*} \bbf e_1 \\ \bbf \beta_1 \\ \bbf \beta_2\beta_1 \\ H \\ \bbf \alpha_3\beta_3\beta_2\beta_1 \\ \bbf \alpha_2\alpha_3\beta_3\beta_2\beta_1 \\ \bbf \alpha_1\alpha_2\alpha_3\beta_3\beta_2\beta_1 \end{gather*} where $H$ is the direct sum of $\bbf \beta_3\beta_2\beta_1$ and the uniserial module \begin{gather*} \bbf \alpha_2\beta_2\beta_1 \\ \bbf \alpha_1\alpha_2\beta_2\beta_1 \\ \bbf \beta_1\alpha_1\alpha_2\beta_2\beta_1. \end{gather*} Further, we may obtain several relations, which importantly include $\alpha_1\beta_1 = 0$. By swapping $\la_1$ and $\la_4$, $\la_2$ and $\la_5$, we obtain the analogous result for $P^{\la_5}$. In the following, we denote by $P^{\la_i}$ the indecomposable projective modules over the basic algebra by abuse of notation. Then right multiplication by $\alpha_1$ gives the unique homomorphism $P^{\la_2}\to P^{\la_1}$ up to nonzero scalar multiple. Let $S$ and $T$ be the submodule of $P^{\la_1}$ generated by $\beta_2\beta_1\alpha_1$ and $\beta_2$, respectively. Then, $\operatorname{Rad}(P^{\la_1})$ is the sum of $T$ and the module \begin{gather*} \bbf \alpha_1 \\ \bbf \beta_1 \alpha_1 \\ S. \end{gather*} Further, $\operatorname{Rad}(S)$ is generated by $\alpha_2\beta_2\beta_1\alpha_1$ and $\beta_3\beta_2\beta_1\alpha_1$, $\operatorname{Rad}(T)$ is generated by $\alpha_2\beta_2$ and $\beta_3\beta_2$. We have a similar result for $P^{\la_4}$. On the other hand, $\operatorname{Rad}(P^{\la_3})$ is generated by $\alpha_2$ and $\beta_3$. Let $A$ be the quotient algebra of the basic algebra of the block by setting $\alpha_2 = 0$ and $\beta_3 = 0$. Further, let $t = e_2+e_3+e_4$ and consider $tAt$. It is easy to check that $tAt\cong \bbf Q/(\varepsilon^2, \varphi^2)$, where $Q$ is the following quiver. \[ \begin{tikzcd}[scale=2] {2} & {3} & {4} \arrow[from=1-1, to=1-1, loop left, "\varepsilon" above] \arrow[from=1-2, to=1-1, "\beta" above] \arrow[from=1-2, to=1-3, "\alpha" above] \arrow[from=1-3, to=1-3, loop right, "\phi" above] \end{tikzcd} \] Thus, $tAt$ is a gentle algebra. Then $tAt$ is Schurian-infinite, by \cite[Proposition~3.3]{Plam19}. We may also see this by applying \cite[Theorem~1.1]{Plam19}, observing that there is a band $\beta^-\alpha\varphi\alpha^-\beta\varepsilon$, which implies that $tAt$ is representation-infinite. The result now follows by \cref{reduction}. \section{Weight 3 blocks}\label{sec:wt3} In this section, we will prove \cref{thm:wt3}. As in \cref{sec:wt2}, we will heavily rely on \cref{prop:matrixtrick}, and thus, by \cref{thm:jamesconjsmallwt} we may assume that $p=0$, and deduce the result for all $p>3$. We once again split our blocks into five separate cases, as in \cref{sec:wt2}. In order to index partitions nicely, we tweak our notation from \cref{subsec:abacus}. Our notation is analogous to that used for example in \cite{fay08wt3}, but slightly different, to keep it compatible with our weight 2 notation choices. Recalling that runner $i$ of the abacus display is by definition the runner containing the position $p_i$, in this subsection, $\langle i \rangle$ will denote the partition obtained from the core by sliding the a bead on runner $i$ down three places, while $\langle i, j \rangle$ will denote the partition obtained from the core by sliding a bead on runner $i$ down two places, and the lowest bead on runner $j$ down one place. Note that we allow $i=j$, so that $\langle i, i \rangle$ is obtained by sliding the lowest bead on runner $i$ down two places, and the second lowest bead down one place. We let $\langle i^2, j \rangle$ denote the partition whose abacus display is obtained from that of the core by sliding down the lowest two beads on runner $i$ and the lowest bead on runner $j$ one place each. We let $\langle i^3 \rangle$ denote the partition whose abacus display is obtained from that of the core by sliding the lowest three beads on runner $i$ down one place each. Finally, we let $\langle i, j, k\rangle$, for $i<j<k$, denote the partition whose abacus display is obtained from that of the core by sliding down the lowest bead on runners $i$, $j$, and $k$ down one place each. In order to also compute the graded decomposition numbers in characteristics 2 and 3, we will use results of Fayers and Tan to determine the adjustment matrices. \begin{defn} If $B$ is a weight 3 block of $\hhh$ given by a core $\rho$ whose abacus display satisfies $p_i - p_{i-1} > 2e$ for all $i = 1,\dots, e-1$, then we say that $B$ is a \emph{Rouquier block}. The collection of all such blocks of weight 3 forms a Scopes equivalence class. A Rouquier block can be thought of as the maximal Scopes class for each weight. \end{defn} \begin{thmc}{faytan06}{Proposition 3.1} Suppose that $B$ is a weight $3$ Rouquier block of $\hhh$. \begin{enumerate}[label=(\roman*)] \item If $\nchar \bbf = 2$, then $a_{\nu\mu}(v) = 1$ if $(\nu,\mu) = (\langle i^3 \rangle, \langle i \rangle)$ or $(\langle i^2,k \rangle, \langle i,k \rangle)$, for $1\leq i,k \leq e-1$ with $i\neq k$. \item If $\nchar \bbf = 3$, then $a_{\nu\mu}(v) = 1$ if $(\nu,\mu) = (\langle i^3 \rangle, \langle i,i \rangle)$ or $(\langle i,i \rangle, \langle i \rangle)$, for $1\leq i \leq e-1$. \end{enumerate} For all other $\nu$ and $\mu$, $a_{\nu\mu}(v) = \delta_{\nu\mu}$. \end{thmc} For the following result, the reader is invited to see \cite{faytan06} for the exact definition of inducing semi-simply and almost semi-simply -- we will only require the explicit determination of when the former occurs. \begin{thmc}{faytan06}{Theorem 3.3 and Proposition 3.4}\label{thm:wt3adj} Suppose that $B$ is a weight $3$ block of $\hhh$. \begin{enumerate}[label=(\roman*)] \item If $\nchar \bbf = 2$, then $a_{\nu\mu}(v) = 1$ if there is a Rouquier block $C$ and some $1\leq i\neq k \leq e-1$ such that \begin{itemize} \item $\nu$ induces semi-simply or almost semi-simply to $\langle i^3 \rangle$ in $C$, while $\mu$ induces semi-simply to $\langle i \rangle$ in $C$; or such that \item $\nu$ induces semi-simply to $\langle i^2, k \rangle$ in $C$, while $\mu$ induces semi-simply to $\langle i, k \rangle$ in $C$. \end{itemize} \item If $\nchar \bbf = 3$, then $a_{\nu\mu}(v) = 1$ if there is a Rouquier block $C$ and some $1\leq i \leq e-1$ such that \begin{itemize} \item $\nu$ induces semi-simply to $\langle i^3 \rangle$ in $C$, while $\mu$ induces semi-simply to $\langle i, i \rangle$ in $C$; or such that \item $\nu$ induces semi-simply to $\langle i, i \rangle$ in $C$, while $\mu$ induces semi-simply to $\langle i \rangle$ in $C$. \end{itemize} \end{enumerate} For all other $\nu$ and $\mu$, $a_{\nu\mu}(v) = \delta_{\nu\mu}$. In particular, a partition $\la$ in $B$ induces up semi-simply to a partition $\omega$ in $C$ of one of the forms above if and only if $\omega$, $\la$ and $B$ satisfy one of the following sets of conditions, where $1 \leq i < k$. (Note that as a matter of convention, we consider $p_e$ to be an arbitrary integer satisfying $p_e > p_{e-1} + 2e$.) \begin{table} \centering \begin{tabular}{ccc}\toprule $\omega$ & $\la$ & Conditions on $B$\\\midrule & $\langle i \rangle$ & $p_{i+1} - p_i > 2e$\\\cmidrule{2-3} $\langle i \rangle$ & $\langle i, i+1 \rangle$ & $p_{i+1} - p_i < 2e$, $p_{i+2} - p_i > e$\\\cmidrule{2-3} & $\langle i, i+1, i+2 \rangle$ & $p_{i+2} - p_i < e$\\\midrule \multirow{2}{*}{$\langle i, i \rangle$} & $\langle i, i \rangle$ & $p_{i+1} - p_i > e$, $p_i - p_{i-1} > e$\\\cmidrule{2-3} & $\langle i^2, i+1 \rangle$ & $p_{i+1} - p_i < e$, $p_i - p_{i-1} > e$\\\midrule $\langle i^3 \rangle$ & $\langle i^3 \rangle$ & $p_i - p_{i-1} > 2e$\\\midrule & $\langle i, k \rangle$ & $p_k - p_i > 2e$, $p_{i+1} - p_i > e$\\\cmidrule{2-3} $\langle i, k \rangle$ & $\langle i, i+1, k \rangle$ & $p_k - p_{i+1} > e$, $p_{i+1} - p_i < e$\\\cmidrule{2-3} & $\langle k^2, i \rangle$ & $p_k - p_{i+1} < e$, $p_k - p_i < 2e$, $p_k - p_{i-1} > e$ \\\midrule & $\langle k, i \rangle$ & $p_{k+1} - p_k > e$, $p_k - p_i > e$\\\cmidrule{2-3} \multirow{2}{*}{$\langle k, i \rangle$} & $\langle k, k \rangle$ & $p_{k+1} - p_k > e$, $p_k - p_i < e$, $p_k - p_{i-1} > e$\\\cmidrule{2-3} & $\langle i, k, k+1 \rangle$ & $p_{k+1} - p_k < e$, $p_k - p_i > e$\\\cmidrule{2-3} & $\langle k^2, k+1 \rangle$ & $p_{k+1} - p_k < e$, $p_k - p_i < e$, $p_k - p_{i-1} > e$\\\midrule \multirow{2}{*}{$\langle i^2, k \rangle$} & $\langle i^2, k \rangle$ & $p_k - p_i > e$, $p_i - p_{i-1} > e$\\\cmidrule{2-3} & $\langle k^3 \rangle$ & $p_k - p_i < e$, $p_k - p_{i-1} > 2e$\\\midrule \multirow{2}{*}{$\langle k^2, i \rangle$} & $\langle k^2, i \rangle$ & $p_k - p_i > 2e$, $p_k - p_{k-1} > e$\\\cmidrule{2-3} & $\langle k^3 \rangle$ & $2e > p_k - p_i > e$, $p_k - p_{i-1} > 2e$, $2e > p_k - p_{k-1} > e$\\ \bottomrule \end{tabular} \end{table} \FloatBarrier \end{thmc} We will use the above result in the following way. Whenever we can choose four partitions such that $a_{\nu\mu}(v) = \delta_{\nu\mu}$ for each $\mu$ among those four partitions, then $d_{\la\mu}^{e,p}(v) = d_{\la\mu}^{e,0}(v)$ for $\la$ and $\mu$ among our four partitions, and thus we may apply \cref{prop:matrixtrick}. Where possible, we choose our partitions so that $\mu$ does not induce up semi-simply to $\langle i \rangle$ or $\langle i,k \rangle$ if $p=2$, or likewise does not induce up semi-simply to $\langle i, i \rangle$ or $\langle i \rangle$ if $p=3$. When this is not possible, it can be checked that no partition $\nu$ induces up semi-simply to $\langle i^3 \rangle$ or $\langle i^2, k \rangle$, or to $\langle i^3 \rangle$ or $\langle i, i \rangle$, respectively. \subsection{$p_{e-1} - p_{e-3} < e$}\label{subsec:wt3firstcase} We take the four partitions $\langle e-3 \rangle$, $\langle e-1, e-3 \rangle$, $\langle e-2, e-3 \rangle$, and $\langle e-3, e-2 \rangle$. To compute the corresponding graded decomposition numbers, note that all but runners $e{-}3$, $e{-}2$ and $e{-}1$ may be removed, using \cref{thm:runnerrem}, leaving an abacus display with just three runners. Since $p_{e-1} - p_{e-3} < e$, the abacus display for the core must be one of the following. \[ \abacus(vvv,bbb,bbb,nnn,vvv) \qquad\qquad \abacus(vvv,bbb,bbn,nnn,vvv) \qquad\qquad \abacus(vvv,bbb,bnn,nnn,vvv) \] They are equivalent, and by convention we work with the first, as in \cref{subsec:wt2firstcase}. Then our four partitions mentioned above become $(7,1^2)$, $(6,2,1)$, $(5,2^2)$ and $(4,3,2)$, respectively, from which we may deduce that the corresponding submatrix (for $e=3$) is known to be (\ref{targetmatrix}) -- for instance, see \cite[Appendix~B]{mathas}.\\ If $p=2$ or $3$, we may apply \cref{thm:wt3adj} and note that we have chosen our four partitions so that $a_{\nu\mu}(v) = \delta_{\nu\mu}$ for any $\mu$ among them, in both characteristics (since none of these partitions induce up semi-simply to any $\langle i \rangle$, $\langle i, i \rangle$, or $\langle i, k \rangle$). Thus we obtain the same submatrix of the graded decomposition matrix in any characteristic. \subsection{$p_{e-1} - p_{e-2} < e$ and $p_{e-2} - p_{e-3} < e$, but $p_{e-1} - p_{e-3} > e$}\label{subsec:wt3secondcase} The four partitions we must examine are $\langle e-2 \rangle$, $\langle e-1, e-2 \rangle$, $\langle e-3 \rangle$, and $\langle e-2, e-3 \rangle$. We may again remove all but runners $e{-}3$, $e{-}2$ and $e{-}1$, using \cref{thm:runnerrem}, leaving an abacus display with just three runners. After doing so, the remaining configuration for the $e=3$ abacus display is as below. \[ \abacus(vvv,bbb,bbn,bnn,nnn,vv) \] Thus, our runner removal applied to the abacus displays for $\langle e-2 \rangle$, $\langle e-1, e-2 \rangle$, $\langle e-3 \rangle$, and $\langle e-2, e-3 \rangle$ yields the four partitions $(8,2)$, $(7,3)$, $(6,2,1^2)$, and $(5,2^2,1)$, respectively. Now, the corresponding submatrix is known to be (\ref{targetmatrix}) -- e.g.~by \cite[Appendix~B]{mathas} -- which completes this case.\\ If $p=2$ or $3$, we may apply \cref{thm:wt3adj} and note, as in \cref{subsec:wt3firstcase}, that we have chosen our four partitions so that $a_{\nu\mu}(v) = \delta_{\nu\mu}$ for any $\mu$ among them, in both characteristics (since none of these partitions induce up semi-simply to any $\langle i \rangle$, $\langle i, i \rangle$, or $\langle i, k \rangle$). Thus we obtain the same submatrix of the graded decomposition matrix in any characteristic. \subsection{$p_{e-1} - p_{e-2} > e$ and $p_{e-2} - p_{e-3} < e$}\label{subsec:wt3thirdcase} Our choice of partitions depends on the values of $p_{e-1}, p_{e-2}$ and $p_{e-3}$, or in other words on which Scopes class our block lies in. The four partitions we will use are either \begin{enumerate}[label=(\roman*)] \item $\langle e-1,e-1 \rangle$, $\langle e-1, e-2 \rangle$, $\langle (e-1)^2, e-2 \rangle$, and $\langle (e-1)^2,e-3 \rangle$; \item $\langle (e-1)^2, e-2 \rangle$, $\langle e-3, e-1 \rangle$, $\langle e-3 \rangle$, and $\langle e-2, e-3 \rangle$; or \item $\langle e-2, e-1 \rangle$, $\langle e-3, e-1 \rangle$, $\langle e-2 \rangle$, and $\langle e-3 \rangle$. \end{enumerate} For each case, to compute the corresponding graded decomposition numbers, we may remove all but the $(e{-}3)$-, $(e{-}2)$-, and $(e{-}1)$-runners from the abacus displays, leaving an abacus display for $e=3$. The remaining core may be any of the cores $\rho^{(i)}$ from \cref{subsec:wt2thirdcase}, and the reader may refer to that section for abacus displays for the first few. Recall that for $j\geq 1$, we may swap adjacent runners to get between $\rho^{(2j)}$ and either $\rho^{(2j-1)}$ or $\rho^{(2j+1)}$, each involve swapping runners on which the number of beads differs by $j+1$. Then if $j\geq 2$, this gives us a Scopes equivalence between blocks. In other words, if we can prove our result for the blocks with cores $\rho^{(1)}=(2)$, $\rho^{(2)}=(3,1)$, and $\rho^{(3)}=(4,2)$, the result will follow. For the block with core $(2)$ (after reduction by runner removal), we take the four partitions in (i) above, which are $(8,3)$, $(8,2,1)$, $(5,3^2)$, and $(5,3,2,1)$, respectively. One can check that the corresponding submatrix is (\ref{targetmatrix}), e.g.~by the LLT algorithm. For the block with core $(3,1)$, we take the four partitions in (ii) above, which are $(6,4,3)$, $(6,3,2,1^2)$, $(5,4,2,1^2)$, and $(4^2,2^2,1)$, respectively. One can check that the corresponding submatrix is (\ref{targetmatrix3}). For the block with core $(4,2)$, we take the four partitions in (iii) above, which are $(7,4,3,1)$, $(7,3^2,1^2)$, $(6,5,3,1)$, $(5^2,3,1^2)$, respectively. Then one can check that the corresponding submatrix is (\ref{targetmatrixalt}), e.g.~by the LLT algorithm. The result now follows when $p>3$.\\ If $p=2$ or $3$, we may apply \cref{thm:wt3adj} and note, as in \cref{subsec:wt3firstcase}, that we have chosen our four partitions so that $a_{\nu\mu}(v) = \delta_{\nu\mu}$ for any $\mu$ among them, in both characteristics for cores $(2)$ and $(3,1)$, and in characteristic $3$ for the core $(4,2)$. To see this, note the following. \begin{enumerate}[label=(\roman*)] \item In the block with core $(2)$, we have that \begin{itemize} \item $\langle e-1, e-1 \rangle$ induces up semi-simply to $\langle e-1, e-1 \rangle$ in a Rouquier block; \item $\langle e-1, e-2 \rangle$ induces up semi-simply to $\langle e-1, e-2 \rangle$ in a Rouquier block; \item $\langle (e-1)^2, e-2 \rangle$ induces up semi-simply to $\langle e-2, e-1 \rangle$ in a Rouquier block; \item $\langle (e-1)^2, e-3 \rangle$ doesn't induce up semi-simply to any $\langle i \rangle$, $\langle i, i \rangle$, or $\langle i, k \rangle$ in a Rouquier block. \end{itemize} Since no $\nu$ induces up semi-simply to $\langle (e-1)^3 \rangle$, $\langle (e-1)^2, e-2 \rangle$, or $\langle (e-2)^2, e-1 \rangle$ in a Rouquier block, we still get that $a_{\nu\mu}(v) = \delta_{\nu\mu}$ for any $\mu$ among our four partitions when $p=2$ or $3$. \item In the block with core $(3,1)$, we have that none of our partitions induce up semi-simply to any $\langle i \rangle$, $\langle i, i \rangle$, or $\langle i, k \rangle$ except for $\langle (e-1)^2, e-2 \rangle$, which induces up semi-simply to $\langle e-2, e-1 \rangle$ in a Rouquier block. Since no $\nu$ induces up semi-simply to $\langle (e-2)^2, e-1 \rangle$ in a Rouquier block, we still get that $a_{\nu\mu}(v) = \delta_{\nu\mu}$ for any $\mu$ among our four partitions when $p=2$ or $3$. \item In the block with core $(4,2)$, none of our partitions induce up semi-simply to any $\langle i \rangle$ or $\langle i, i \rangle$, except for $\langle e-2 \rangle$, which induces up semi-simply to $\langle e-2 \rangle$ in a Rouquier block. Since no $\nu$ induces up semi-simply to $\langle e-2, e-2 \rangle$ in a Rouquier block, we still get that $a_{\nu\mu}(v) = \delta_{\nu\mu}$ for any $\mu$ among our four partitions when $p = 3$. \end{enumerate} Thus we obtain the same submatrix of the graded decomposition matrix in any characteristic, except for the block with core $(4,2)$ in characteristic $2$. For this, we must choose different partitions -- we take $\langle e-1, e-1 \rangle$, $\langle (e-1)^2, e-2 \rangle$, $\langle e-2, e-1 \rangle$, and $\langle e-3, e-1 \rangle$. We may again use runner-removal to reduce the computation of the characteristic zero graded decomposition numbers, yielding partitions $(10,5)$, $(7,5,2,1)$, $(7,4,3,1)$, and $(7,3^2,1^2)$, whence we obtain the submatrix (\ref{targetmatrix3}). Now, since none of our four partitions induce semi-simply to any $\langle i \rangle$ or $\langle i, k \rangle$, by \cref{thm:wt3adj}, we have $a_{\nu\mu}(v) = \delta_{\nu\mu}$ for any $\mu$ among them, and thus that we obtain the same submatrix of the graded decomposition matrix in characteristics $0$ and $2$. Thus, in all cases above, \cref{prop:matrixtrick} yields our result. \subsection{$p_{e-1} - p_{e-2} < e$ and $p_{e-2} - p_{e-3} > e$}\label{subsec:wt3fourthcase} Our choice of partitions depends on the values of $p_{e-1}, p_{e-2}$ and $p_{e-3}$, or in other words on which Scopes class our block lies in. The four partitions we will use are either \begin{enumerate}[label=(\roman*)] \item $\langle e-1,e-1 \rangle$, $\langle e-2, e-2 \rangle$, $\langle (e-1)^2, e-2 \rangle$, and $\langle (e-2)^2,e-1 \rangle$; \item $\langle e-2 \rangle$, $\langle e-1, e-2 \rangle$, $\langle e-1, e-1 \rangle$, and $\langle e-2, e-2 \rangle$; or \item $\langle e-1 \rangle$, $\langle e-2 \rangle$, $\langle e-1, e-2 \rangle$, and $\langle e-1, e-1 \rangle$. \end{enumerate} For each case, to compute the corresponding graded decomposition numbers, we may remove all but the $(e{-}3)$-, $(e{-}2)$-, and $(e{-}1)$-runners from the abacus displays, leaving an abacus display for $e=3$. The remaining core may be any of the cores $\sigma^{(i)}$ from \cref{subsec:wt2fourthcase}, and the reader may refer to that section for abacus displays for the first few. As in \cref{subsec:wt3thirdcase}, our Scopes equivalences reduce the problem down to solving for the blocks with cores $\sigma^{(1)} = (1^2)$, $\sigma^{(2)} = (2,1^2)$, and $\sigma^{(3)} = (2^2,1^2)$. For the block with core $(1^2)$ (after reduction by runner removal), we take the four partitions in (i) above, which are $(7,2^2)$, $(6,2^2,1)$, $(4^2,3)$, and $(4^2,2,1)$, respectively. One can check that the corresponding submatrix is (\ref{targetmatrixalt}), e.g.~by the LLT algorithm. For the block with core $(2,1^2)$, we take the four partitions in (ii) above, which are $(9,3,1)$, $(8,4,1)$, $(8,3,2)$, and $(6,3,2^2)$, respectively. One can check that the corresponding submatrix is (\ref{targetmatrix3}). For the block with core $(2^2,1^2)$, we take the four partitions in (iii) above, which are $(11,2,1^2)$, $(10,3,1^2)$, $(8,5,1^2)$, and $(8,3^2,1)$, respectively. One can check that the corresponding submatrix is (\ref{targetmatrix}). The result now follows when $p>3$.\\ If $p=2$ or $3$, we may apply \cref{thm:wt3adj} and note, as in \cref{subsec:wt3firstcase}, that we have chosen our four partitions so that $a_{\nu\mu}(v) = \delta_{\nu\mu}$ for any $\mu$ among them, in both characteristics for cores $(1^2)$ and $(2,1^2)$, and in characteristic $3$ for the core $(2^2,1^2)$. To see this, note the following. \begin{enumerate}[label=(\roman*)] \item In the block with core $(1^2)$, we have that \begin{itemize} \item $\langle e-1, e-1 \rangle$ induces up semi-simply to $\langle e-1, e-2 \rangle$ in a Rouquier block $C$; \item $\langle e-2, e-2 \rangle$ doesn't induce up semi-simply to any $\langle i \rangle$, $\langle i, i \rangle$, or $\langle i, k \rangle$ in a Rouquier block $C$; \item $\langle (e-1)^2, e-2 \rangle$ induces up semi-simply to $\langle e-2, e-1 \rangle$ in a Rouquier block $C$; \item $\langle (e-2)^2, e-1 \rangle$ induces up semi-simply to $\langle e-2, e-2 \rangle$ in a Rouquier block $C$. \end{itemize} Since no $\nu$ induces up semi-simply to $\langle (e-1)^2, e-2 \rangle$, $\langle (e-2)^2, e-1 \rangle$, or $\langle (e-2)^3 \rangle$ in $C$, we still get that $a_{\nu\mu}(v) = \delta_{\nu\mu}$ for any $\mu$ among our four partitions when $p=2$ or $3$. \item In the block with core $(2,1^2)$, we have that none of our partitions induce up semi-simply to any $\langle i \rangle$, $\langle i, i \rangle$, or $\langle i, k \rangle$ except for $\langle e-1, e-1 \rangle$, which induces up semi-simply to $\langle e-1, e-2 \rangle$ in a Rouquier block $C$. Since no $\nu$ induces up semi-simply to $\langle (e-1)^2, e-2 \rangle$ in a Rouquier block $C$, we still get that $a_{\nu\mu}(v) = \delta_{\nu\mu}$ for any $\mu$ among our four partitions when $p=2$ or $3$. \item In the block with core $(2^2,1^2)$, none of our partitions induce up semi-simply to any $\langle i \rangle$ or $\langle i, i \rangle$, except for $\langle e-1 \rangle$, which induces up semi-simply to $\langle e-1 \rangle$ in a Rouquier block. Since no $\nu$ induces up semi-simply to $\langle e-1, e-1 \rangle$ in a Rouquier block, we still get that $a_{\nu\mu}(v) = \delta_{\nu\mu}$ for any $\mu$ among our four partitions when $p = 3$. \end{enumerate} Thus we obtain the same submatrix of the graded decomposition matrix in any characteristic, except for the block with core $(2^2,1^2)$ in characteristic $2$. For this, we must choose different partitions -- we take $\langle e-2 \rangle$, $\langle e-1, e-2 \rangle$, $\langle e-1, e-1 \rangle$, and $\langle e-2, e-2 \rangle$. We may again use runner-removal to reduce the computation of the characteristic zero graded decomposition numbers, yielding partitions $(10,3,1^2)$, $(8,5,1^2)$, $(8,3^2,1)$, and $(7,3^2,2)$, whence we obtain the submatrix (\ref{targetmatrix3}). Now, by \cref{thm:wt3adj}, none of our partitions induce semi-simply to any $\langle i \rangle$ or $\langle i, k \rangle$, except for $\langle e-1, e-1 \rangle$, which induces semi-simply to $\langle e-1, e-2 \rangle$ in a Rouquier block. But since no $\nu$ induces semi-simply to $\langle (e-1)^2, e-2 \rangle$, we have $a_{\nu\mu}(v) = \delta_{\nu\mu}$ for any $\mu$ among our four partitions, and thus we obtain the same submatrix of the graded decomposition matrix in characteristics $0$ and $2$. Thus, in all cases above, \cref{prop:matrixtrick} yields our result. \subsection{$p_{e-1} - p_{e-2} > e$ and $p_{e-2} - p_{e-3} > e$}\label{subsec:wt3fifthcase} Across various cases, we will use partitions $\langle e-1,e-1 \rangle, \langle e-1,e-2 \rangle, \langle (e-1)^2,e-2 \rangle, \langle e-2,e-1 \rangle, \langle (e-1)^3 \rangle, \langle (e-2)^2,e-1 \rangle, \langle e-2, e-2 \rangle$. Notice that in all cases, we are able to apply \cref{thm:runnerrem}, reducing our computations down to the $e=3$ situation. Recall that in \cref{subsec:wt2fifthcase}, we were able to index the cores that arise in this case by the pair of integers $(ke+1,je+1)$ or $(ke+2, je+2)$, where $k, j \in \bbz_{>0}$, and denote these cores by $\kappa_{(ke+1,je+1)}$ or $\kappa_{(ke+2,je+2)}$. As stated in \cref{subsec:wt2fifthcase}, we may pass between these cores by runner swaps, and these swaps involve swapping runners on which the number of beads differs by either $k+j+1$ (for passing between $\kappa_{(ke+1,je+1)}$ and $\kappa_{(ke+2,je+2)}$), $k$ (for passing between $\kappa_{((k-1)e+2,je+2)}$ and $\kappa_{(ke+1,je+1)}$), or $j$ (for passing between $\kappa_{(ke+2,(j-1)e+2)}$ and $\kappa_{(ke+1,je+1)}$). Since we are now in weight $3$, these are Scopes equivalences except for small $k$ and $j$, and we can see that there are four Scopes classes in total to check, with (minimal) representatives $\kappa_{(e+1,e+1)} = (3,1^2)$, $\kappa_{(2e+1,e+1)} = (5,3,1^2)$, $\kappa_{(e+1,2e+1)} = (4,2^2,1^2)$, and $\kappa_{(2e+1,2e+1)} = (6,4,2^2,1^2)$. For the block with core $(3,1^2)$, we may take the four partitions $\langle e-1, e-1 \rangle = (9,4,1)$, $\langle e-1,e-2 \rangle = (9,3,2)$, $\langle (e-1)^2,e-2 \rangle = (6,4^2)$, and $\langle (e-1)^3 \rangle = (6,4,2^2)$, respectively. One can check that the corresponding submatrix is (\ref{targetmatrix}), e.g.~by the LLT algorithm. For the block with core $(5,3,1^2)$, we may take the four partitions $\langle (e-1)^2,e-2 \rangle = (8,6,3,2)$, $\langle e-2,e-1 \rangle = (8,5,4,2)$, $\langle (e-2)^2,e-1 \rangle = (8,3^2,2^2,1)$, and $\langle e-2, e-2 \rangle = (5^2,4,2^2,1)$, respectively. One can check that the corresponding submatrix is (\ref{targetmatrixalt}), e.g.~by the LLT algorithm. For the block with core $(4,2^2,1^2)$, we may take the four partitions $\langle e-1,e-2 \rangle = (10,4,3,1^2)$, $\langle (e-1)^2,e-2 \rangle = (7,5^2,1^2)$, $\langle (e-1)^3 \rangle = (7,5,3^2,1)$, and $\langle (e-2)^2,e-1 \rangle = (7,4,3^2,2)$, respectively. One can check that the corresponding submatrix is (\ref{targetmatrix2}). For the block with core $(6,4,2^2,1^2)$, we may take the four partitions $\langle e-1,e-2 \rangle = (12,4^2,3,1^2)$, $\langle (e-1)^2,e-2 \rangle = (9,7,4,3,1^2)$, $\langle e-2,e-1 \rangle = (9,6,5,3,1^2)$, and $\langle (e-2)^2,e-1 \rangle = (9,4^2,3^2,2)$, respectively. One can check that the corresponding submatrix is (\ref{targetmatrixaltsquare}). The result now follows if $p>3$. If $p=3$, we may apply \cref{thm:wt3adj} and note, as in \cref{subsec:wt3firstcase}, that we have chosen our four partitions so that $a_{\nu\mu}(v) = \delta_{\nu\mu}$ for any $\mu$ among them. Checking this is similar to previous cases, and we leave the details to the reader. If $p=2$, then these four blocks are not susceptible to our methods when $e=3$. Thus we will return to this $p=2$ case later. For now we will assume that $e\geq 4$. Under this assumption, we will take either the partition $\langle (e-1)^3 \rangle$ or the partition $\langle (e-1)^2, e-2 \rangle$ as our most dominant partition, along with the three partitions $\langle (e-1)^2, e-3 \rangle$, $\langle (e-2)^2, e-1 \rangle$, and $\langle e-3, e-2, e-1 \rangle$. Then, for any pair of these partitions, we may apply \cref{thm:runnerrem} to remove all but four runners when computing the characteristic zero graded decomposition numbers. One may easily check, using the system of runner swaps at the end of \cref{subsec:wt2fifthcase}, that up to Scopes equivalence, these decomposition numbers boil down to those of just 10 blocks, with the following cores (in the notation of \cref{subsec:wt2fifthcase}). \begin{multicols}{4} \begin{enumerate}[label=(\roman*)] \item $\kappa_{e+1, e+1, 1}$, \item $\kappa_{e+1, e+2, 3}$, \item $\kappa_{e+1, e+1, e+1}$, \item $\kappa_{e+1, e+1, 2e+1}$, \item $\kappa_{e+1, 2e+1, e+1}$, \item $\kappa_{e+1, 2e+1, 2e+1}$, \item $\kappa_{2e+1, e+1, e+1}$, \item $\kappa_{2e+1, e+1, 2e+1}$, \item $\kappa_{2e+1, 2e+1, e+1}$, \item $\kappa_{2e+1, 2e+1, 2e+1}$. \end{enumerate} \end{multicols} One may check that if we choose the partitions $\langle (e-1)^3 \rangle$, $\langle (e-1)^2, e-3 \rangle$, $\langle (e-2)^2, e-1 \rangle$, and $\langle e-3, e-2, e-1 \rangle$ in cases (i)--(vi) and choose the partitions $\langle (e-1)^2, e-2 \rangle$, $\langle (e-1)^2, e-3 \rangle$, $\langle (e-2)^2, e-1 \rangle$, and $\langle e-3, e-2, e-1 \rangle$ in cases (vii)--(x), respectively, every case yields (\ref{targetmatrixalt}) as the corresponding submatrix of the graded decomposition matrix. Concretely, in each case the corresponding partitions are as follows. \begin{enumerate}[label=(\roman*)] \item $(9,6,3^2)$, $(9,6,2^2,1^2)$, $(9,5,3^2,1)$, $(9,5,3,2,1^2)$; \item $(10,7,4^2,1)$, $(10,7,3,2^3)$, $(10,6,4^2,2)$, $(10,6,4,2^3)$; \item $(10,7,4^2,1^2)$, $(10,7,3^2,2^2)$, $(10,6,4^2,2,1)$, $(10,6,4,3,2^2)$; \item $(11,5^2,2^2,1^3)$, $(11,8,4^2,3^2,1^3)$, $(11,7,5^2,3,2,1^3)$, $(11,7,5,4,3^2,1^3)$; \item $(12,9,6^2,3,1^3)$, $(12,9,5,3^3,2^2)$, $(12,8,6^2,4,1^3)$, $(12,8,6,3^3,2^2)$; \item $(13,10,7^2,4,2^3,1^3)$, $(13,10,6,4^3,3^2,1^3)$, $(13,9,7^2,5,2^3,1^3)$, $(13,9,7,4^3,3^2,1^3)$; \item $(13,10,6,4,1^3)$, $(13,10,3^3,2^2)$, $(13,6^2,4^2,2,1)$, $(13,6^2,4,3,2^2)$; \item $(14,11,7,5,2^3,1^3)$, $(14,11,4^3,3^2,1^3)$, $(14,7^2,5^2,3,2,1^3)$, $(14,7^2,5,4,3^2,1^3)$; \item $(15,12,8,6,3^2,1^3)$, $(15,12,5^2,3^3,2^2)$, $(15,8^2,6^2,4,1^3)$, $(15,8^2,6,3^3,2^2)$; \item $(16,13,9,7,4^2,2^3,1^3)$, $(16,13,6^2,4^3,3^2,1^3)$, $(16,9^2,7^2,5,2^3,1^3)$, $(16,9^2,7,4^3,3^2,1^3)$. \end{enumerate} Now, for any $e$, one may check that none of $\langle (e-1)^2, e-3 \rangle$, $\langle (e-2)^2, e-1 \rangle$, or $\langle e-3, e-2, e-1 \rangle$ induce up semi-simply to any $\langle i \rangle$ or $\langle i, k \rangle$ in any of these cases (i.e.~whenever $p_{e-1} - p_{e-2} > e$ and $p_{e-2} - p_{e-3} > e$), while $\langle (e-1)^3 \rangle$ and $\langle (e-1)^2, e-2 \rangle$ don't induce up semi-simply to any $\langle i \rangle$ or $\langle i, k \rangle$ in cases (i)--(vi) and (vii)--(x), respectively (i.e.~whenever $2e > p_{e-1} - p_{e-2} > e$ and $p_{e-2} - p_{e-3} > e$, or $p_{e-1} - p_{e-2} > 2e$ and $p_{e-2} - p_{e-3} > e$, respectively). Thus $a_{\nu \mu} = 0$ whenever $\mu$ is any one of our chosen partitions, and the result follows as in the previous sections. It is worth noting that even though our chosen partitions involve sliding beads on only three runners, if we use runner-removal to reduce to abacus displays with three runners, our partitions are not all $3$-regular. We now return to the $p=2$ case, and will directly show that the four blocks are Schurian-infinite. Note that we may use the LLT algorithm to compute the characteristic $0$ decomposition matrices, and then apply \cref{thm:wt3adj} (along with \cite[Proposition~3.5]{faytan06} in order to determine when a partition induces almost semi-simply to another) to determine those in characteristic $2$. In the block with core $(6,4,2^2,1^2)$, the five most dominant $3$-regular partitions are $\la_1 := (15,4,2^2,1^2)$, $\la_2 := (12,7,2^2,1^2)$, $\la_3 := (12,4^2,3,1^2)$, $\la_4 := (9,7,5,2,1^2)$, and $\la_5 := (9,7,4,3,1^2)$. The five most dominant $3$-regular partitions in the block with core $(5,3,1^2)$ may be obtained from those by removing the first column (of length 6) from each. In both blocks, we obtain characteristic $2$ decomposition matrices that start with the following identical 5 rows. \[ \begin{array}{c|ccccc} & \la_1 & \la_2 & \la_3 & \la_4 & \la_5 \\\hline \la_1 & 1 & \cdot & \cdot & \cdot & \cdot \\ \la_2 & 0 & 1 & \cdot & \cdot & \cdot \\ \la_3 & 1 & 1 & 1 & \cdot & \cdot \\ \la_4 & 1 & 0 & 0 & 1 & \cdot \\ \la_5 & 1 & 1 & 1 & 1 & 1 \\ \end{array} \] It follows that in each case, we have \[ \spe{\la_1} = \D{\la_1}, \quad \spe{\la_2} = \D{\la_2}, \quad \operatorname{Rad}(\spe{\la_4}) \cong \D{\la_1}. \] We may compute that there are homomorphisms from $\spe{\la_1}$ and $\spe{\la_2}$ to $\spe{\la_3}$, so that $\operatorname{Rad}(\spe{\la_3}) \cong \D{\la_1} \oplus \D{\la_2}$. For example, working with graded (column) Specht modules for KLR algebras \cite[Section~7]{kmr}, one may check that these homomorphisms are given by \[ z^{\la_1} \longmapsto (\psi_{16} \psi_{15} \psi_{14} \psi_{13} \psi_{17} \psi_{16} \psi_{18} \psi_{17} \dots \psi_{14} + \psi_{22} \psi_{21} \dots \psi_{13} \psi_{23} \psi_{22} \dots \psi_{16} \psi_{24} \psi_{23} \dots \psi_{14}) z^{\la_3} \] and \[ z^{\la_2} \longmapsto \psi_{15} \psi_{14} \psi_{13} \psi_{17} \psi_{16} \psi_{19} \psi_{18} \psi_{17} \psi_{16} \psi_{15} \psi_{14} z^{\la_3}. \] It is easy to check that these define nonzero elements in $\spe{\la_3}$ -- replacing each generator $\psi_i$ with the corresponding basic transposition $s_i$ and acting naturally on the (column) initial $\la_3$-tableau yields a standard $\la_3$-tableau, and indeed each expression is equal to the element of the standard homogenous basis (c.f.~\cite[Corollary~7.20]{kmr}) indexed by that tableau, up to some lower order terms. Moreover, it is easy to check that we may apply column-removal (c.f.~\cite[Theorems~3.6 and 4.1]{fs16}) to yield corresponding homomorphisms in the smaller block (by shifting all indices down by $6$). We may also compute that there are homomorphisms from $\spe{\la_1}$ and $\spe{\la_2}$ to $\spe{\la_5}$, so that $\D{\la_1} \oplus \D{\la_2}$ is a submodule of $\spe{\la_5}$, and that there are homomorphisms from $\spe{\la_3}$ and $\spe{\la_4}$ to $\spe{\la_5}$. These are given by \[ z^{\la_1} \longmapsto \psi_{19} \psi_{18} \psi_{17} \psi_{16} \psi_{20} \psi_{19} \psi_{18} \psi_{21} \psi_{20} \psi_{22} \psi_{21} \dots \psi_{13} \psi_{23} \psi_{22} \dots \psi_{16} \psi_{24} \psi_{23} \dots \psi_{14} z^{\la_5}, \] \[ z^{\la_2} \longmapsto \psi_{15} \psi_{14} \psi_{13} \psi_{17} \psi_{16} \psi_{19} \psi_{18} \dots \psi_{14} \psi_{22} \psi_{21} \psi_{20} \psi_{19} \psi_{23} \psi_{22} \psi_{21} \psi_{24} \psi_{23} z^{\la_5}, \] \[ z^{\la_3} \longmapsto \psi_{22} \psi_{21} \psi_{20} \psi_{19} \psi_{23} \psi_{22} \psi_{21} \psi_{24} \psi_{23} z^{\la_5}, \] and \[ z^{\la_4} \longmapsto \psi_{18} \psi_{17} \psi_{16} \psi_{15} \psi_{14} z^{\la_5}. \] Once again, it is easy to check that the elements in $\spe{\la_5}$ are nonzero, and we may apply column-removal to obtain the homomorphisms for the smaller block by shifting all indices down by $6$. One may check that (in both blocks) composing homomorphisms from $\spe{\la_1}$ and $\spe{\la_2}$ to $\spe{\la_3}$ and from $\spe{\la_1}$ to $\spe{\la_4}$ with these latter homomorphisms gives nonzero homomorphisms, so that $\operatorname{Soc}(\spe{\la_5}) \cong \D{\la_1} \oplus \D{\la_2}$ and $\spe{\la_5}$ has heart $\D{\la_3} \oplus \D{\la_4}$. It follows that we have extensions yielding $A^{(1)}_3$ with zigzag orientation as a subquiver of the Gabriel quiver, on vertices $\la_1$, $\la_3$, $\la_4$, and $\la_5$. The result now follows by \cref{cor:Schurianinfinitequiver}. For the block with core $(4,2^2,1^2)$, we note that the $\#$-automorphism on $\hhh$ (e.g.~see \cite[Exercise~3.14]{mathas}) induces an isomorphism between this block and the one with core $(5,3,1^2)' = (4,2^2,1^2)$. Finally, let $B$ be the weight $3$ block with core $(3,1^2)$. Using the LLT algorithm and applying \cref{thm:wt3adj} and \cite[Proposition~3.5]{faytan06}, we may compute that the characteristic $2$ decomposition matrix has the following first 5 rows. \[ \begin{array}{c|ccccc} & (12,1^2) & (9,4,1) & (9,3,2) & (8,4,2) & (6^2,2) \\\hline (12,1^2) & 1 & \cdot & \cdot & \cdot & \cdot \\ (9,4,1) & 0 & 1 & \cdot & \cdot & \cdot \\ (9,3,2) & 1 & 1 & 1 & \cdot & \cdot \\ (8,4,2) & 1 & 1 & 1 & 1 & \cdot \\ (6^2,2) & 1 & 0 & 0 & 1 & 1 \\ \end{array} \] We see immediately that $\spe{(12,1^2)} = \D{(12,1^2)}$ and $\spe{(9,4,1)} = \D{(9,4,1)}$. Moreover, we may check that there are homomorphisms from $\spe{(12,1^2)}$ and $\spe{(9,4,1)}$ to $\spe{(9,3,2)}$ given by \[ z^{(12,1^2)} \longmapsto (\psi_5 \psi_7 \psi_6 + \psi_{11} \psi_{10} \dots \psi_5 \psi_{12} \psi_{11} \dots \psi_7 \psi_{13} \psi_{12} \dots \psi_6) z^{(9,3,2)} \] and \[ z^{(9,4,1)} \longmapsto \psi_8 \psi_7 \psi_6 z^{(9,3,2)}, \] yielding that $\operatorname{Rad}(\spe{(9,3,2)}) \cong \D{(12,1^2)} \oplus \D{(9,4,1)}$. Now, one may compute that there is no homomorphism $\spe{(12,1^2)} \rightarrow \spe{(8,4,2)}$, but there are homomorphisms $\spe{(9,4,1)} \rightarrow \spe{(8,4,2)}$ and $\spe{(9,3,2)} \rightarrow \spe{(8,4,2)}$, given by \[ z^{(9,4,1)} \longmapsto \psi_8 \psi_7 \psi_6 \psi_{13} \psi_{12} \psi_{11} \psi_{10} z^{(8,4,2)} \] and \[ z^{(9,3,2)} \longmapsto \psi_{13} \psi_{12} \psi_{11} \psi_{10} z^{(8,4,2)}, \] and moreover that the composition of homomorphisms $\spe{(9,4,1)}\rightarrow \spe{(9,3,2)} \rightarrow \spe{(8,4,2)}$ is nonzero. This implies that $\spe{(8,4,2)}$ has a submodule consisting of a simple head isomorphic to $\D{(9,3,2)}$ and a simple socle isomorphic to $\D{(9,4,1)}$. Now consider the module $M := \spe{(8,4,2)}/\operatorname{Soc}(\spe{(8,4,2)})$. Since $M$ has simple head $\D{(8,4,2)}$, $\operatorname{Soc}(M)$ contains $\D{(9,3,2)}$, and $\D{(12,1^2)}$ does not appear in $\operatorname{Soc}(\spe{(8,4,2)})$, it follows that we must have $\D{(12,1^2)}$ in the head of $\operatorname{Rad}(M)$, and therefore in the head of $\operatorname{Rad}(\spe{(8,4,2)})$. Then we have \[ \operatorname{Ext}^1(\D{(8,4,2)},\D{(12,1^2)}) \neq 0. \] Finally, there is a homomorphism $\spe{(12,1^2)} \rightarrow \spe{(6^2,2)}$ given by \[ z^{(12,1^2)} \longmapsto \psi_8 \psi_7 \psi_6 \psi_5 \psi_9 \psi_8 \psi_7 \psi_{10} \psi_9 \psi_{11} \psi_{13} \psi_{12} \psi_6 z^{(9,3,2)}, \] and therefore $\D{(12,1^2)}$ is in the socle of $\spe{(6^2,2)}$, and thus $\D{(8,4,2)}$ must be in $\operatorname{Rad}{(\spe{(6^2,2)})}/\operatorname{Rad}^2{(\spe{(6^2,2)})}$. Thus we have \[ \operatorname{Ext}^1(\D{(6^2,2)},\D{(8,4,2)}) \neq 0. \] Now, since the Mullineux map switches $(12,1^2)$ with $(6^2,2)$, and switches $(8,4,2)$ with $(4^2,2^2,1^2)$, we must also have that \[ \operatorname{Ext}^1(\D{(4^2,2^2,1^2)},\D{(6^2,2)}) \neq 0 \quad \text{ and } \quad \operatorname{Ext}^1(\D{(12,1^2)},\D{(4^2,2^2,1^2)}) \neq 0. \] Then the four partitions $(12,1^2)$, $(8,4,2)$, $(6^2,2)$ and $(4^2,2^2,1^2)$ form a square in the Gabriel quiver, and we may apply \cref{cor:Schurianinfinitequiver} to complete the result. \section{Blocks of weight at least 4}\label{sec:highwt} The goal of this section is to prove \cref{thm:principalblock}, i.e.~to show that if the quantum characteristic of $q\in \bbf^\times$ is $e\geq3$, then the principal block of the Hecke algebra $\hhh$ is Schurian-infinite whenever it has weight at least $2$. This means that the principal block of $\hhh$ is Schurian-finite if and only if it is representation-finite, that is, its weight is $w\leq1$. In fact, we will prove that a much larger family of blocks are Schurian-infinite, and deduce the results for most principal blocks as a corollary. We begin with the following simple observation, which is an abacus realisation of \cref{thm:kleshdecomp}. \begin{lem}\label{lem:abacusklesh} Suppose $\mu$ has an abacus display in which adjacent runners $i$ and $i+1$ are such that the lowest beads on runners $i$ and $i+1$ are in positions $r$ and $s$, respectively, for some $r<s$, and that positions $s-1$ and positions $s-e, s-2e, \dots, r+1$ are all unoccupied. Let $\la$ denote the partition whose abacus display is obtained from this one by pushing the bead in position $r$ to position $r+1$, and the bead in position $s$ to position $s-1$. Then, $d_{\la\mu}^{e,p}(v) = v$. \end{lem} \begin{rem} The conditions in the above lemma amount to looking for the following configuration of beads at the bottom of two adjacent runners in an abacus display. \[ \mu: \abacus(vv,bn,nn,vv,nn,nb,nn,vv) \qquad \qquad \longmapsto \qquad \qquad \la: \abacus(vv,nb,nn,vv,nn,bn,nn,vv) \] Note that we allow the case where there are no empty rows between the beads on the above displays, i.e.~the case $s=r+e+1$. \end{rem} By combining \cref{lem:abacusklesh} with \cref{thm:rowremFock,thm:rowremDecomp}, we may independently move $i$-nodes and $j$-nodes to yield decomposition numbers of $v$ and $v^2$ that are independent of the characteristic of the field. This is actually sufficient to find the matrix (\ref{targetmatrixalt}) as a submatrix of the decomposition matrix of many blocks, from which the result follows by \cref{prop:matrixtrick}. If $i=j$ or $j\pm 1$, we will actually need to cut the Young diagram into a top and bottom part, independently moving $i$-nodes in the top part and $j$-nodes in the bottom part to yield the same result. First, we will handle a case in which we assume $i=j$. \begin{prop}\label{prop:2runners} Suppose $B$ is a block of $\hhh$ whose core $\rho$ has an abacus display in which two adjacent runners are as follows, \[ \makeatletter \begin{tikzpicture}[scale=\abas*1.2,,baseline={([yshift=-.8ex]current bounding box.center)}] \@bacus vv,bb,bb,bb,bn,bn,vv,bn,nn,vv . \coordinate[at={(0cm,-\abav*3.5)}] (first2); \coordinate[at={($(first2)+(0,-\abav*3)$)}] (last2); \tikzbracel{first2}{last2}{k $ beads$} \node at (5cm,-3.8cm){or}; \abax=9cm\abay=0cm\abaxorig=9cm \@bacus vv,bb,bb,bb,nb,nb,vv,nb,nn,vv . \coordinate[at={(11.8cm,-\abav*3.5)}] (first1); \coordinate[at={($(first1)+(0,-\abav*3)$)}] (last1); \tikzbracer{first1}{last1}{k $ beads$} \end{tikzpicture} \makeatother \] and which has weight $w\geq 2k+4$, for some $k\geq 0$ in the first case above, or $k>0$ in the second. Let $r$ be the position of the bottom left bead on these two runners in the first case above, and $r+1$ the position of the bottom right bead on them in the second (so that position $r$ always lies on the left-hand runner). Suppose further that not all of the positions $r-e+2, r-e+3, \dots, r-1$ are occupied by beads. Then $B$ is Schurian-infinite. \end{prop} \begin{proof} First, suppose that our abacus display for $\rho$ has the former pair of adjacent runners in it, and that $w=2k+4$ for $k\geq0$. Looking at just these two runners, we may obtain a partition $\la^{(1)}$ in $B$ by sliding the bottom left bead down 1 space, the bottom right bead down $k+2$ spaces, and the bead above it down $k+1$ spaces. In other words, $\la^{(1)}$ has an abacus display that agrees with $\rho$ on all but these two runners, that now look as follows. \[ \abacus(vv,bb,bn,bn,vv,bn,bn,nb,bn,nb,nn,vv) \] Now we have two pairs of beads satisfying the conditions of \cref{lem:abacusklesh}. If $w>2k+4$, slide the lowest bead down $w-2k-4$ more spaces (for a total of $w-k-2$). Now define $\la^{(2)}$ to be the partition whose abacus display is obtained from that of $\la^{(1)}$ by pushing the lowest bead one space to the left, and the second-lowest bead one space to the right. Similarly, construct $\la^{(3)}$ by performing the same moves to the third and fourth lowest beads, and $\la^{(4)}$ by performing both of the above `flips'. The relevant two runners of these abacus displays are pictured below. \[ \la^{(2)}: \abacus(vv,bb,bn,bn,vv,bn,bn,nb,nb,bn,nn,vv) \qquad \qquad \la^{(3)}: \abacus(vv,bb,bn,bn,vv,bn,nb,bn,bn,nb,nn,vv) \qquad \qquad \la^{(4)}: \abacus(vv,bb,bn,bn,vv,bn,nb,bn,nb,bn,nn,vv) \] The condition that not all of $r-e+2, r-e+3, \dots, r-1$ are occupied by beads ensures that all four of these partitions are $e$-regular. Thus, we may apply \cref{lem:abacusklesh,thm:rowremFock,thm:rowremDecomp} to obtain \[ d_{\la^{(2)} \la^{(1)}}^{e,p}(v) = d_{\la^{(3)} \la^{(1)}}^{e,p}(v) = d_{\la^{(4)} \la^{(2)}}^{e,p}(v) = d_{\la^{(4)} \la^{(3)}}^{e,p}(v) = v, \;\; d_{\la^{(4)} \la^{(1)}}^{e,p}(v) = v^2. \] Finally, since $\la^{(2)}$ and $\la^{(3)}$ are incomparable in the dominance order, we have $d_{\la^{(3)} \la^{(2)}}^{e,p}(v) = 0$. Hence, the assumption of \cref{prop:matrixtrick} are satisfied -- in particular, finding the submatrix (\ref{targetmatrixalt}) -- and the result follows. Next, suppose that our abacus display for $\rho$ has the latter pair of adjacent runners in it, and that $w=2k+4$ for $k>0$. Looking at just these two runners, we may obtain a partition $\la^{(1)}$ in $B$ by sliding the bead at the bottom right down $2$ spaces, the bead above it down $1$ space, the bottom left bead down $k+1$ spaces, and the bead above it down $k$ spaces. In other words, $\la^{(1)}$ has an abacus display that agrees with $\rho$ on all but these two runners, that now look as follows. \[ \abacus(vv,bb,bn,nb,nb,vv,nb,nb,bn,nb,bn,nb,nn,vv) \] Now we have two pairs of beads satisfying the conditions of \cref{lem:abacusklesh}. If $w>2k+4$, slide the lowest bead down $w-2k-4$ more spaces (for a total of $w-2k-2$). Now define $\la^{(2)}$, $\la^{(3)}$, and $\la^{(4)}$ exactly as in the previous case. The rest of the proof follows identically to the previous case, noting once again that our condition on unoccupied positions ensures that all four partitions are $e$-regular. \end{proof} \begin{egs} \begin{enumerate}[label=(\roman*)] \item Let $e=3$, $\rho = (2^2,1^2)$, and $w=4$. Then $\rho$ has the following abacus display, which satisfies the hypotheses of \cref{prop:2runners} on the first two runners, landing in the $k=0$ case therein. \[ \abacus(vvv,bbb,bbn,bbn,bbn,nnn,nnn,vvv) \] Then, using the notation in the proof of \cref{prop:2runners}, we construct the four partitions: \begin{align*} \la^{(1)}=(8,5,4,1)&: \abacus(vvv,bbb,bbn,bnn,nbn,bnn,nbn,nnn,vvv) \qquad \qquad \la^{(2)}=(7,6,4,1): \abacus(vvv,bbb,bbn,bnn,nbn,nbn,bnn,nnn,vvv)\\ \\ \la^{(3)}=(8,5,3,2)&: \abacus(vvv,bbb,bbn,nbn,bnn,bnn,nbn,nnn,vvv) \qquad \qquad \la^{(4)}=(7,6,3,2): \abacus(vvv,bbb,bbn,nbn,bnn,nbn,bnn,nnn,vvv) \end{align*} By cutting the partitions after the second row of each, we may apply \cref{lem:abacusklesh} (or \cref{thm:kleshdecomp}) and \cref{thm:rowremFock,thm:rowremDecomp} to obtain $d_{\la^{(4)} \la^{(1)}}^{e,p}(v) = v^2$. Note that if we had a block of weight $w>4$ and the same core, we would simply slide the lowest bead down further in each of the above abacus displays, or equivalently add $(w-4)e$ to the first part of each of the partitions. \item Let $e=3$, $\rho = (6,4,2^2,1^2)$, and $w=8$. Then $\rho$ has the following abacus display, which satisfies the hypotheses of \cref{prop:2runners} on the second and third runners, landing in the $k=2$ case therein. \[ \abacus(vvv,bbb,nbb,nbb,nnb,nnb,nnn,vvv) \] The process in the proof of \cref{prop:2runners} constructs $\la^{(1)} = (12,9,8,5,4,2)$, with abacus display \[ \abacus(vvv,bbb,nnb,nnb,nbn,nnb,nbn,nnb,nnn,vvv) \] \end{enumerate} \end{egs} Analogously to the two runner situation covered by \cref{prop:2runners}, similar methods can be applied to abacus displays consisting of two such adjacent pairs of runners. This corresponds to adding and removing both an $i$-node and a $j$-node in the Young diagram, for $i\notin \{j,j\pm1\}$. This case will actually suffice for proving \cref{thm:principalblock} when $e\geq 4$. \begin{prop}\label{prop:4runners} Suppose $B$ is a block of $\hhh$ of weight $w\geq j + k + 2$, for some $j,k\geq 0$, whose core $\rho$ has an abacus display in which two pairs of adjacent runners are as follows. \[ \makeatletter \begin{tikzpicture}[scale=\abas*1.2,,baseline={([yshift=-.8ex]current bounding box.center)}] \node at (-6cm,-3.8cm){One pair is either}; \@bacus vv,bb,bb,bb,bn,bn,vv,bn,nn,vv . \coordinate[at={(0cm,-\abav*3.5)}] (first2); \coordinate[at={($(first2)+(0,-\abav*3)$)}] (last2); \tikzbracel{first2}{last2}{j} \node at (3cm,-3.8cm){or}; \abax=5cm\abay=0cm\abaxorig=5cm \@bacus vv,bb,bb,bb,nb,nb,vv,nb,nn,vv . \coordinate[at={(7.8cm,-\abav*3.5)}] (first1); \coordinate[at={($(first1)+(0,-\abav*3)$)}] (last1); \tikzbracer{first1}{last1}{j} \node at (-7.5cm,-13.3cm){and the other pair is either}; \abax=0cm\abay=-9.5cm\abaxorig=0cm \@bacus vv,bb,bb,bb,bn,bn,vv,bn,nn,vv . \coordinate[at={(0cm,-9.5cm-\abav*3.5)}] (first3); \coordinate[at={($(first3)+(0,-\abav*3)$)}] (last3); \tikzbracel{first3}{last3}{k} \node at (3cm,-13.3cm){or}; \abax=5cm\abay=-9.5cm\abaxorig=5cm \@bacus vv,bb,bb,bb,nb,nb,vv,nb,nn,vv . \coordinate[at={(7.8cm,-9.5cm-\abav*3.5)}] (first4); \coordinate[at={($(first4)+(0,-\abav*3)$)}] (last4); \tikzbracer{first4}{last4}{k} \end{tikzpicture} \makeatother \] Let $r_1, r_2$ denote the position of lowest bead on each of these two pairs of runners. If $j=0$ or $k=0$, we let the corresponding $r_i$ denote the position of the bottom-leftmost bead. If they are not in the same row, let $r'$ be the higher of the two (i.e.~$r'$ is equal to either $r_1$ or $r_2$), and let $r = r'$ if the pair of runners is in the first configuration above and $r=r'-1$ if the pair is in the second configuration above, and suppose further that not all of the positions $r+2, r+3, \dots, r+e-1$ are occupied by beads. Then $B$ is Schurian-infinite. \end{prop} \begin{rem} If the lowest bead on each pair of runners is in the same row, then no further conditions are required to ensure we construct four $e$-regular partitions. If they are not in the same row, the definition above sets $r$ to be a position on the left-hand runner of its pair, for either configuration. \end{rem} \begin{proof} For each pair of runners, if $j\geq 0$ (respectively $k\geq 0$) we may slide the lowest bead on the right-hand runner down $j+1$ (respectively $k+1$) places in the first configuration, or if $j> 0$ (respectively $k> 0$) slide the lowest bead on the left-hand runner down $j$ (respectively $k$) places and the lowest bead on the right-hand runner down $1$ place in the second configuration, for a total of $j+k+2$ places, as necessary for the a block of weight $j+k+2$. If $w>j+k+2$, we may slide the lowest bead on these four runners down its runner further. In any case, this constructs a partition $\la^{(1)}$, from which we can construct another three partitions as in the proof of \cref{prop:2runners}, and proceed in an identical fashion. As in that proof, the conditions on unoccupied abacus positions ensures that all four partitions are $e$-regular. In order to obtain the decomposition number $d_{\la^{(4)} \la^{(1)}}^{e,p}(v) = v^2$, it suffices to note that the residues of the two pairs of bead moves (i.e.~of the two nodes in the Young diagram that are being moved) are not equal or adjacent, so that the decomposition number is $d_{\la^{(4)} \la^{(1)}}^{e,p}(v) = d_{\la^{(2)} \la^{(1)}}^{e,p}(v)d_{\la^{(4)} \la^{(2)}}^{e,p}(v) = v^2$, for example by \cite[Theorem~5.4]{bs15}. \end{proof} Now we may deduce \cref{thm:principalblock}, mostly using \cref{prop:4runners}. \begin{thm}\label{thm:principalblockmain} If $e\geq3$, then the principal block of the Hecke algebra $\hhh$ is Schurian-infinite whenever it has weight $w\geq 2$. \end{thm} \begin{proof} First, note that if the block has weight $2$ or $3$, it is Schurian-infinite in any characteristic, by \cref{thm:wt2,thm:wt3}. For now, assume that $e\geq 4$. If the block has empty core, then there is an abacus display for it in which we may find four adjacent runners of the following form. \[ \abacus(vvvv,bbbb,bbbb,nnnn,vvvv) \] We may now apply \cref{prop:4runners} with $j=k=0$ to deduce that the block is Schurian-infinite whenever it has weight $w\geq 2$. Similarly, if the core is $(1)$ or $(k)$, for $2\leq k \leq e-2$, we may find four adjacent runners of the following forms. \[ \abacus(vvvv,bbbb,bbnb,nnnn,vvvv) \qquad \text{or} \qquad \abacus(vvvv,bbbb,bbbb,nnbn,vvvv) \qquad \text{respectively.} \] Then we apply \cref{prop:4runners} with $j=0$, $k=1$ to deduce that the block is Schurian-infinite whenever it has weight $w\geq 3$. Finally, if the core is $(e-1)$, then we may find four adjacent runners of the following form. \[ \abacus(vvvv,bbbb,bbbb,nnnb,vvvv) \] We may again apply \cref{prop:4runners} with $j=0$, $k=1$ to deduce that the block is Schurian-infinite whenever it has weight $w\geq 3$. Now, we let $e=3$. This case is harder to visualise on the abacus, as some `wrapping around' is necessary, so we instead directly appeal to \cref{lem:abacusklesh,thm:rowremFock,thm:rowremDecomp}. If the block has core $\varnothing$ and weight $w\geq 4$, then we consider the four $3$-regular partitions \begin{alignat*}3 \la^{(1)} &= (3w-6,3^2) \qquad &&\la^{(2)} = (3w-6,3,2,1)\\ \la^{(3)} &= (3w-7,4,3) \qquad &&\la^{(4)} = (3w-7,4,2,1) \end{alignat*} of $3w$. We divide $\la^{(i)}$ into the first two rows and the remaining rows. Then, the first two rows of $\la^{(i)}$ are either $((w-2)e,e)$ or $((w-2)e-1,e+1)$, and the remaining rows of $\la^{(i)}$ are either $(e)$ or $(e-1,1)$. Applying \cref{thm:kleshdecomp} to the pairs $\{((w-2)e,e), ((w-2)e-1,e+1)\}$ and $\{(e), (e-1,1)\}$, respectively, and using row-removal \cref{thm:rowremFock,thm:rowremDecomp}, we obtain \[ d_{\la^{(2)} \la^{(1)}}^{e,p}(v) = d_{\la^{(3)} \la^{(1)}}^{e,p}(v) = d_{\la^{(4)} \la^{(2)}}^{e,p}(v) = d_{\la^{(4)} \la^{(3)}}^{e,p}(v) = v, \;\; d_{\la^{(4)} \la^{(1)}}^{e,p}(v) = v^2 \] for any $p\geq 0$. Since $\la^{(2)}$ and $\la^{(3)}$ are incomparable in the dominance, we have $d_{\la^{(3)} \la^{(2)}}^{e,0}(v) = 0$ and $d_{\la^{(3)} \la^{(2)}}^{e,p}(1) = 0$. Hence, the assumption of \cref{prop:matrixtrick} are satisfied -- in particular, finding the submatrix (\ref{targetmatrixalt}) -- and the result follows. Similarly, if the core is $(1)$ and the weight is $w\geq 5$, the result follows in the same way by considering the four $3$-regular partitions \begin{alignat*}3 \la^{(1)} &= (3w-8,4^2,1) \qquad &&\la^{(2)} = (3w-8,4,3,2)\\ \la^{(3)} &= (3w-9,5,4,1) \qquad &&\la^{(4)} = (3w-9,5,3,2) \end{alignat*} of $3w+1$. If the core is $(2)$ and the weight is $w\geq 6$, the result follows in the same way by considering the four $3$-regular partitions \begin{alignat*}3 \la^{(1)} &= (3w-9,6,4,1) \qquad &&\la^{(2)} = (3w-9,6,3,2)\\ \la^{(3)} &= (3w-10,7,4,1) \qquad &&\la^{(4)} = (3w-10,7,3,2) \end{alignat*} of $3w+2$. To handle the remaining cases, we must work slightly harder. First, suppose that $B$ is the principal block of $\hhh[13]$ (so that the core is $(1)$ and the weight is $4$). We consider the four $3$-regular partitions \begin{alignat*}3 \la^{(1)} &= (7,5,1) \qquad &&\la^{(2)} = (7,4,1^2)\\ \la^{(3)} &= (7,3^2) \qquad &&\la^{(4)} = (7,3,2,1) \end{alignat*} of $13$, and will show that the corresponding graded decomposition matrix is (\ref{targetmatrixalt}) in any characteristic. By \cref{thm:kleshdecomp}, $d^{e,p}_{\la^{(4)} \la^{(2)}}(v) = d^{e,p}_{\la^{(4)} \la^{(3)}}(v) = d^{e,p}_{\la^{(2)} \la^{(1)}}(v) = v$, while $d^{e,p}_{\la^{(3)} \la^{(2)}}(v) = 0$ since $\la^{(2)}$ and $\la^{(3)}$ are incomparable in the dominance order. To compute $d^{e,p}_{\la^{(4)} \la^{(1)}}(v)$, we may use row-removal \cref{thm:rowremFock,thm:rowremDecomp} along with the decomposition numbers computed in \cref{subsec:wt2firstcase}, yielding \[ d_{\la^{(4)} \la^{(1)}}^{e,0}(v) = d_{(3,2,1) (5,1)}^{e,0}(v) = v^2, \;\; d_{\la^{(4)} \la^{(1)}}^{e,p}(1) = d_{(3,2,1) (5,1)}^{e,p}(1) = 1. \] Similarly, to compute $d^{e,p}_{\la^{(3)} \la^{(1)}}(v)$, we may use row-removal \cref{thm:rowremFock} to obtain $d_{\la^{(3)} \la^{(1)}}^{e,0}(v) = d_{(3^2) (5,1)}^{e,0}(v) = v$, and \cref{thm:rowremDecomp} to see that $d_{\la^{(3)} \la^{(1)}}^{e,p}(1) = d_{(3^2) (5,1)}^{e,p}(1)$. To see that $d_{(3^2) (5,1)}^{e,p}(1) = 1$ for any $p$, we note that the adjustment matrices are always trivial for $p>2$, and that we already showed in \cref{subsec:wt2firstcase} that for $p=2$, $a_{\nu (5,1)}(v) = 0$ for any $\nu$, and thus $d_{(3,2,1) (5,1)}^{e,2}(1) = d_{(3,2,1) (5,1)}^{e,0}(1) = 1$. The result now follows. Finally, we let $B$ be the principal block of $\hhh[14]$ or $\hhh[17]$, respectively -- i.e.~of $\hhh$ for $n = 3w+2$ for $w=4$ or $5$. By considering the four $e$-regular partitions \begin{alignat*}3 \la^{(1)} &= ((w-2)e+2,5,1) \qquad &&\la^{(2)} = ((w-2)e+2,4,1^2)\\ \la^{(3)} &= ((w-2)e+2,3^2) \qquad &&\la^{(4)} = ((w-2)e+2,3,2,1) \end{alignat*} of $3w+2$, the argument may be completed identically to the previous case, showing that the corresponding graded decomposition matrix is (\ref{targetmatrixalt}) in any characteristic, and applying \cref{prop:matrixtrick} to yield the result. \end{proof} \begin{cor}\label{cor:signblock} Suppose $e\geq 3$, $n\geq 2e$, and that $B$ is the block of $\hhh$ that contains the sign representation $\spe{(1^n)}$. Then $B$ is Schurian-infinite in any characteristic. \end{cor} \begin{proof} The result follows immediately from \cref{thm:principalblockmain} by applying the $\#$-automorphism on $\hhh$ (e.g.~see \cite[Exercise~3.14]{mathas}), which induces an isomorphism between $B$ and the principal block. \end{proof} Finally, we return to determining that blocks are Schurian-infinite based on abacus displays for their cores. It remains to handle moving nodes of consecutive residues in a Young diagram. We may do so as long as we can use row-removal to factorise the decomposition numbers as a product of those corresponding to just moving a single node. In other words, an $j$-node in the top part of the Young diagram moves independently to an $i=j\pm 1$-node in the bottom part, and we may factorise by our row removal theorems. \begin{prop}\label{prop:3runners} Suppose $B$ is a block of $\hhh$ of weight $w\geq j+k+1$, for some $j,k\geq 1$, whose core $\rho$ has an abacus display in which three adjacent runners are as follows. \[ \makeatletter \begin{tikzpicture}[scale=\abas*1.2,,baseline={([yshift=-.8ex]current bounding box.center)}] \@bacus vvv,bbb,nbb,nbb,vvv,nbb,nnb,nnb,vvv,nnb,nnn,vvv . \coordinate[at={(3.8cm,-\abav*1.5)}] (first1); \coordinate[at={($(first1)+(0,-\abav*3)$)}] (last1); \tikzbracer{first1}{last1}{j $ beads$} \coordinate[at={(3.8cm,-\abav*5.5)}] (first2); \coordinate[at={($(first2)+(0,-\abav*3)$)}] (last2); \tikzbracer{first2}{last2}{k $ beads$} \node at (7.8cm,-3.8cm){or}; \abax=13.5cm\abay=0cm\abaxorig=13.5cm \@bacus vvv,bbb,bbn,bbn,vvv,bbn,bnn,bnn,vvv,bnn,nnn,vvv . \coordinate[at={(13.5cm,-\abav*1.5)}] (first3); \coordinate[at={($(first3)+(0,-\abav*3)$)}] (last3); \tikzbracel{first3}{last3}{j $ beads$} \coordinate[at={(13.5cm,-\abav*5.5)}] (first4); \coordinate[at={($(first4)+(0,-\abav*3)$)}] (last4); \tikzbracel{first4}{last4}{k $ beads$} \end{tikzpicture} \makeatother \] Let $r$ be the position of the lowest bead on the middle runner of these three, and suppose further that not all of the positions $r-e+2, r-e+3, \dots, r-2$ are occupied by beads. Then $B$ is Schurian-infinite. \end{prop} \begin{rem} If the above configuration is such that $j=0$ or $k=0$, working with three runners cannot get as strong a result as we can already obtain by applying \cref{prop:2runners}. \end{rem} \begin{proof} The argument is almost identical to the proof of \cref{prop:2runners,prop:4runners}, so we only explain how to obtain the partition $\la^{(1)}$ therein. To obtain the abacus display for $\la^{(1)}$ from that of $\rho$ in the first case, we slide down the bottom bead on the left-hand runner $j-1$ places down, the bottom bead on the right-hand runner down $1$ place, the bottom bead on the middle runner down $k$ places, and the bead above it down $1$ place. In the second case, we instead slide the bottom bead on the middle runner down $k+1$ spaces, and the bottom bead on the right-hand runner $j$ spaces (and do not slide any beads on the left-hand runner). Once again, we may slide the lowest bead down further if the weight is $w>j+k+1$, and once again the condition on unoccupied positions ensures that all four constructed partitions will be $e$-regular. \end{proof} Similar methods also prove the following proposition, the details of which we leave to the reader. We note only that -- unlike \cref{prop:2runners,prop:4runners,prop:3runners} -- it's easy to construct four $e$-regular partitions with no extra conditions required. \begin{prop}\label{prop:3runnersalt} Suppose $B$ is a block of $\hhh$ of weight $w\geq 2k+3$, for some $k\geq 1$, whose core $\rho$ has an abacus display in which three adjacent runners are as follows. \[ \makeatletter \begin{tikzpicture}[scale=\abas*1.2,,baseline={([yshift=-.8ex]current bounding box.center)}] \@bacus vvv,bbb,nbn,nbn,vvv,nbn,nnn,vvv . \coordinate[at={(3.8cm,-\abav*1.5)}] (first1); \coordinate[at={($(first1)+(0,-\abav*3)$)}] (last1); \tikzbracer{first1}{last1}{k $ beads$} \end{tikzpicture} \makeatother \] Then $B$ is Schurian-infinite. \end{prop} \begin{rem} Having extra beads on the leftmost or rightmost runner and trying to construct partitions with these three runners ends up producing a weaker result than just focussing on two of the three runners, so that the above proposition involves first and third runners with the same number of beads. \end{rem} In fact, an identical argument to the proof of \cref{cor:signblock} can be used to conjugate the core of any block that we have already shown to be Schurian-infinite. \begin{cor}\label{cor:conjugateblocks} Suppose $B$ is a block of $\hhh$ with core $\rho$, such that its conjugate $\rho'$ has an abacus display satisfying the conditions of Propositions \ref{prop:2runners}, \ref{prop:4runners}, \ref{prop:3runners}, or \ref{prop:3runnersalt}, and $B$ has weight as in the corresponding proposition. Then $B$ is Schurian-infinite in any characteristic. \end{cor} We now give an example of a block that our methods are unable to show is Schurian-infinite. This example is the smallest rank weight 4 block when $e=4$ that we cannot verify. \begin{eg} Let $e=4$ , and let $B$ be the weight $4$ block with core $(3,2,1)$, which has an abacus configuration \[ \abacus(vvvv,bbbb,bbbb,nbnb,nbnn,nnnn,vvvv) \] This is in fact the best choice of abacus display for us, because we may apply \cref{prop:4runners} to determine that any block of weight at least $5$ with core $(3,2,1)$ is Schurian-infinite. The weight $4$ block, however, is not resolved by our results in this paper. \end{eg} We end with a conjecture. \begin{conj} Suppose $e\geq3$. Then all blocks of $\hhh$ of weight at least 2 are Schurian-infinite in any characteristic. \end{conj} \bibliographystyle{amsalpha} \addcontentsline{toc}{section}{\refname}

https://arxiv.org/abs/2004.04298

Modular forms and ellipsoidal T-designs

In recent work, Miezaki introduced the notion of a $spherical$ $T$-d$esign$ in $\mathbb{R}^2$, where $T$ is a potentially infinite set. As an example, he offered the $\mathbb{Z}^2$-lattice points with fixed integer norm (a.k.a. shells). These shells are $maximal$ spherical $T$-designs, where $T=\mathbb{Z}^+\setminus 4\mathbb{Z}^+$. We generalize the notion of a spherical $T$-design to special ellipses, and extend Miezaki's work to the norm form shells for rings of integers of imaginary quadratic fields with class number 1.

\section{Introduction and statement of results}\label{Intro} $Spherical$ $t$-$designs$ were introduced in 1977 by Delsarte, Goethals and Seidel \cite{Delsarte}, and they have played an important role in algebra, combinatorics, number theory and quantum mechanics (for background see \cite{Bannai}, \cite{BOT}, \cite{Chen}, \cite{quantum}, \cite{Go}, \cite{Miezaki}). A spherical $t$-design is a nonempty finite set of points on the unit sphere with the property that the average value of any real polynomial of degree $\leq t$ over this set equals the average value over the sphere. Namely, if $S^{n-1}$ denotes the unit sphere in $\mathbb{R}^n$ centered at the origin, then a finite nonempty subset $X\subset S^{n-1}$ is a spherical $t$-design if \begin{equation}\label{E1} \dfrac{1}{|X|}\sum_{x\in X}P(x)= \dfrac{1}{\operatorname{Vol}(S^{n-1})}\int_{S^{n-1}} P(x)d\sigma(x) \end{equation} for all polynomials $P(x)$ of degree $\leq t$. The right-hand side of ($\ref{E1}$) is the usual surface integral over $S^{n-1}$. In general, a finite nonempty subset $X$ of $S_{n-1}(r)$, the sphere of radius $r$ centered at the origin, is a spherical $t$-design if $\frac{1}{r}X$ satisfies (\ref{E1}). Since a spherical $t$-design is also a spherical $t^{\prime}$-design for all $ t^{\prime}\leq t$, we say that $X$ has $strength$ $t$ if it is the maximum of all such numbers. Delsarte, Goethals and Seidel developed a very simple criterion for determining spherical $t$-designs. This criterion involves {\it homogeneous harmonic} polynomials of bounded degree. A polynomial in $n$ variables is $harmonic$ if it is annihilated by the Laplacian operator $\Delta:=\sum^{n}_{i=1}\partial^2/\partial x^2_i$, and they showed \cite{Delsarte} that $X\subset S^{n-1}$ is a spherical $t$-design if \begin{equation}\label{P} \sum_{x\in X}P(x) = 0 \end{equation} for all homogeneous harmonic polynomials $P(x)$ of nonzero degree $\leq t$. This criterion is a consequence of two results from harmonic analysis. The first result is the mean value property for harmonic functions \cite[p.~5]{HFT}, which implies that the integral of a harmonic polynomial over a sphere centered at the origin vanishes, combined with the fact that homogeneous polynomials of fixed degree are spanned by certain harmonic polynomials \cite[Th. ~5.7]{HFT}. In view of this framework, it is natural to ask whether there are generalizations of spherical $t$-designs to other curves, surfaces and varieties. Here we consider certain $ellipsoids\footnote{We do not use the term $ellipse$ to avoid possible confusion that might arise with the term $elliptical$.}$ in dimension two. To be precise, for square-free $D\geq 1$ we define the norm $r$ ellipses \begin{equation} C_D(r) := \begin{cases} \{(x,y)\in \mathbb{R}^2 : x^2+Dy^2=r\} & \text{ if } D\equiv 1,2\pmod{4},\\ \{(x,y)\in \mathbb{R}^2 : x^2+xy+\frac{1+D}{4}y^2=r\} & \text{ if } D\equiv 3\pmod{4}. \end{cases} \end{equation} \begin{rmk} These ellipses arise from certain imaginary quadratic orders. \end{rmk} For $D\equiv 1,2\pmod 4$, we say that a finite nonempty subset $X\subset C_D(r)$ is an $ellipsoidal$ $t$-$design$ if \begin{equation}\label{HE1} \dfrac{1}{|X|}\sum_{(x,y)\in X}P(x,y)= \dfrac{1}{2\pi\sqrt{D}}\mathlarger\int_{C_D(r)}\dfrac{P(x,y)}{\sqrt{x^2/D^2+y^2}}d\sigma(x,y) \end{equation} for all polynomials $P(x,y)$ of degree $\leq t$ over $\mathbb{R}$. For $ D\equiv 3\pmod 4$, instead we require \begin{equation}\label{HE2} \dfrac{1}{|X|}\sum_{(x,y)\in X}P(x,y)= \dfrac{\sqrt{D}}{\pi}\mathlarger\int_{C_D(r)}\dfrac{P(x,y)}{\sqrt{20x^2+(D^2+2D+5)y^2+(20+4D)xy}}d\sigma(x,y). \end{equation} Here the right-hand sides are line integrals. As in the case of spherical $t$-designs, every ellipsoidal $t$-design is also an ellipsoidal $t^{\prime}$-design for all $ t^{\prime}\leq t$, and the maximum of all such $t$'s is called the $strength$ of $X$.These definitions coincide with the notion of a spherical $t$-design when $D=1$. In analogy to Delsarte, Goethals and Seidel, we have a natural criterion for confirming ellipsoidal $t$-designs. To this end, we consider the 2-dimensional real vector space \begin{equation}\label{D-Harmonic} H^{\mathbb{R}}_{D,j}[x,y] := \begin{cases} \langle \text{Re}(x+\sqrt{-D}y)^j,\text{Im}(x+\sqrt{-D}y)^j\rangle & \text{ if } D\equiv 1,2 \pmod{4},\\ \langle \text{Re}(x+\frac{1+\sqrt{-D}}{2}y)^j,\text{Im}(x+\frac{1+\sqrt{-D}}{2}y)^j\rangle & \text{ if } D\equiv 3 \pmod{4}. \end{cases} \end{equation} In terms of these vector spaces of polynomials, we have the following ellipsoidal $t$-design criterion. \begin{theorem}\label{CD} A finite nonempty set $X\subset C_D(r)$ is an ellipsoidal $t$-design if \begin{equation*} \sum_{x\in X}P(x,y)=0 \end{equation*} for all $P(x,y)\in H^{\mathbb{R}}_{D,j}[x,y]$ for all $0<j\leq t$. \end{theorem} \begin{rmk} 1)Observe that if $X\subset S^{1}$ is a spherical $t$-design, then $Y=\{(x,y/\sqrt{D})|(x,y)\in X\}\subset C_D$ (resp. $Y=\{(x+y/\sqrt{D},2y/\sqrt{D}|(x,y)\in X\}\subset C_D$) is an ellipsoidal $t$-design for $D\equiv 1,2 \pmod{4}$ (resp. $D\equiv 3 \pmod{4}$). Therefore, the existence of a spherical $t$-design implies the existence of a corresponding ellipsoidal $t$-design. In fact, there is a one-to-one correspondence between spherical $t$-designs and ellipsoidal $t$-designs. However, the proof of Theorem \ref{CD} is not a direct consequence because care is required for justifying the role of the vector spaces $H^{\mathbb{R}}_{D,j}[x,y]$.\\ 2)Since there is one-to-one correspondence between spherical and ellipsoidal $t$-designs, we get a lower bound \cite[pg 2]{Delsarte} on the size of ellipsoidal $t$-design $X$, $$|X|\geq t+1. $$ \end{rmk} Recently, Miezaki in \cite{Miezaki} introduced a generalization of the notion of spherical $t$-designs. Instead of restricting to polynomials of degree $\leq t$, he considered harmonic polynomials of degree $j\in T\subset \mathbb{N}$, where $T$ is a potentially infinite set. The main theorem from \cite{Miezaki} gives infinitely many spherical $T$-designs for $T :=\mathbb{Z}^{+}\setminus 4\mathbb{Z}^+$ in dimension two. Namely, he considered norm $r$ shells, integer points on $x^2+y^2=r$ for fixed $r\in \mathbb{Z}^+$. He showed that these $r$-shells are spherical $T$-designs. Moreover, these sets have strength $T$, meaning that (\ref{P}) fails if any multiple of 4 is added to $T$. His proof makes use of theta functions arising from complex multiplication by $\mathbb{Z}[i]$. We generalize Miezaki's work to ellipsoidal $T$-designs. We call $X\subset C_D$ an $ellipsoidal$ $T$-$design$ if the condition in Theorem \ref{CD} is satisfied for all polynomials in $H^{\mathbb{R}}_{D,j}[x,y]$ with $j\in T$. We say $X$ has strength $T$ if it is maximal among such sets. For each square-free positive integer $D$, let $\mathcal{O}_D$ be the ring of integers of $\mathbb{Q}(\sqrt{-D})$. In particular, this means that \begin{equation}\label{OD} \mathcal{O}_D=\begin{cases} \mathbb{Z}[\sqrt{-D}] \ \ \ \ \ & \text{ if } D\equiv 1,2 \pmod{4}, \\ \mathbb{Z}[\frac{1+\sqrt{-D}}{2}] \ \ \ \ \ & \text{ if } D\equiv 3 \pmod{4}. \end{cases} \end{equation} We consider $D\in\{1,2,3,7,11,19,43,67,163\}$, the square-free positive integers for which $\mathcal{O}_D$ has class number 1. To make this precise, we define the {\it norm $r$ shells} in $C_D(r)$ by \begin{equation}\label{NormShells} \Lambda_D^r:= \mathcal{O}_D \cap C_D(r). \end{equation} Generalizing Miezaki’s work for $D=1$, we obtain the following theorem. \begin{theorem}\label{Main} If $D\in \{1, 2, 3, 7, 11, 19, 43, 67, 163\},$ then every non-empty shell $\Lambda_D^r$ is an ellipsoidal $T_D$ design with strength $T_D$, where $$ T_D:=\begin{cases} \mathbb{Z}^+\setminus 4\mathbb{Z}^+ \ \ \ \ \ &{\text {\rm if}}\ D=1,\\ \mathbb{Z}^+\setminus 6\mathbb{Z}^{+} \ \ \ \ &{\text {\rm if}}\ D=3,\\ \mathbb{Z}^+\setminus 2\mathbb{Z}^+ \ \ \ \ &{\text {\rm otherwise.}} \end{cases} $$ \end{theorem} \begin{rmk} The method used here seems to be well-poised only for the dimension 2 cases. It would be interesting to obtain higher dimensional analogues. \end{rmk} \begin{example} We consider $D=3$, and $r=691$. Then we have \begin{align*} \Lambda^{691}_3=&\{(11,19),(-11,-19),(19,11),(-19,-11),(11,-30),(-11,30),(30,-19),(-30,19),\\ &(30,-11),(-30,11),(19,-30),(-19,30)\}. \end{align*} We consider the polynomial $P(x,y)=2x^2+3462xy+1729y^2\in H^{\mathbb{R}}_{3,2}[x,y]$, and we find that $\sum_{(x,y)\in \Lambda^{691}_3}P(x,y)=0$ which shows that $\Lambda^{691}_3$ is an elliptical $2$-design and $2\in T_3$. On the other hand, Theorem \ref{Main} implies that $\Lambda^{691}_3$ is not an ellipsoidal $6$-design.To see this we choose $Q(x,y)=2x^2+6x^5y-15x^4y^2-40x^3y^3-15x^2y^4+6xy^5+2y^6\in H^{\mathbb{R}}_{3,6}(x,y),$ and we find that $\sum_{(x,y)\in \Lambda^{691}_3}Q(x,y)=-4818834696\not=0$. \end{example} In Section $2$ we prove Theorem \ref{CD}, criterion for confirming that a set is an ellipsoidal $t$-design, and in Section $3$ we recall the theory of theta functions arising from complex multiplication, and we prove Theorem \ref{Main}. \section*{Acknowledgement} I would like to thank Prof Ken Ono for suggesting me this problem and guiding through. I also thank Will Craig and Wei-Lun Tsai for reviewing my paper and giving useful comments. I thank Matthew McCarthy for helping me with Sage Math. Lastly, I would like to thank the reviewers for their useful comments. \section{Criterion for ellipsoidal $t$-Design} In this section we prove Theorem \ref{CD}, criterion for confirming ellipsoidal $t$-designs. Throughout this section we assume that $D\geq 1$ is square-free and $j\geq 1$. To prove that Theorem \ref{CD} is indeed a criterion for confirming ellipsoidal $t$-designs, we first need to show that the spaces $H^{\mathbb{R}}_{D,k}[x,y]$, for $0<k\leq j$, generate all the polynomials of degree $\leq j$ when restricted to $C_D(r)$. It suffices to show this for $P^{\mathbb{R}}_j[x,y]$, the set of homogeneous polynomials of degree $j$. \begin{lemma}\label{H-Span} If $D\geq 1$ is square-free and $j\geq 1$, then the following are true:\\ 1) If $D\equiv 3\mod{4}$, then we have \begin{equation*} P^{\mathbb{R}}_j[x,y] = \bigoplus^{\floor{j/2}}_{k=0} (x^2+Dy^2)^{k}H^{\mathbb{R}}_{D,j-2k}[x,y]. \end{equation*} 2) If $D\equiv 3\mod{4}$, then we have \begin{equation*} P^{\mathbb{R}}_j[x,y] = \bigoplus^{\floor{j/2}}_{k=0} \Big(x^2+xy+\frac{1+D}{4}y^2\Big)^kH^{\mathbb{R}}_{D,j-2k}[x,y]. \end{equation*} \end{lemma} \begin{proof} The lemma is well known for homogeneous harmonic polynomials (for example, see \cite[Thm~5.7]{HFT}). Namely, if $H^{\mathbb{R}}_{k}[x,y]$ is the set of homogeneous harmonic polynomials of degree $k$ then \begin{equation*} P^{\mathbb{R}}_j(x,y) = \bigoplus^{\floor{j/2}}_{k=0} (x^2+y^2)^{k}H^{\mathbb{R}}_{j-2k}[x,y]. \end{equation*} We extend it to general $D$. It is well known that $H^{\mathbb{R}}_j[x,y]=\langle\text{Re}(x+iy)^j, \text{Im}(x+iy)^j\rangle$, and so if we do the change of variable for $D\equiv 1,2\mod{4}$ (resp. $D\equiv 3\mod{4}$), $x'=x$,$y'=\sqrt{D}y$ (resp. $x'=x+y/2$,$y'=2y/\sqrt{D}$), then $H^{\mathbb{R}}_{j-2}(x',y')=\langle \text{Re}(x'+ iy')^j,\text{Im}(x'+iy')^j\rangle$ gives \begin{equation*} P^{\mathbb{R}}_j[x',y'] = \bigoplus^{\floor{j/2}}_{k=0} (x'^2+y'^2)^{k}H^{\mathbb{R}}_{j-2k}[x',y']. \end{equation*} Therefore, if $D\equiv 1,2\mod{4}$, then we have \begin{equation*} P^{\mathbb{R}}_j(x,y) = \bigoplus^{\floor{j/2}}_{k=0} (x^2+Dy^2)^{k}H^{\mathbb{R}}_{D,j-2k}[x,y]. \end{equation*} If $D\equiv 3\mod{4}$, then we have \begin{equation*} P^{\mathbb{R}}_j(x,y) = \bigoplus^{\floor{j/2}}_{k=0} \Big(x^2+xy+\frac{1+D}{4}y^2\Big)^kH^{\mathbb{R}}_{D,j-2k}[x,y]. \end{equation*} \end{proof} We now prove Theorem \ref{CD}. \begin{proof}[Proof of Theorem \ref{CD}]\label{IN} Lemma \ref{H-Span} shows that the set of polynomials when restricted to $C_D$ are generated by the spaces $H^{\mathbb{R}}_{D,j}[x,y]$ since $x^2+Dy^2=r$ (resp., $x^2+xy+\frac{1+D}{4}y^2=r$) on $C_D(r)$. Therefore, it suffices to show that if $P(x,y)\in H^{\mathbb{R}}_{D,j}[x,y]$, then the following are true:\\ 1) If $D\equiv 1,2\mod{4}$, then we have $$\int_{C_D(r)} \dfrac{P(x,y)}{\sqrt{x^2/D^2+y^2}}d\sigma(x,y)=0.$$ 2) If $D\equiv 3\mod{4}$, then we have $$\int_{C_D(r)}\dfrac{P(x,y)}{\sqrt{20x^2+(D^2+2D+5)y^2+(20+4D)xy}}d\sigma(x,y)=0.$$ As $H^{\mathbb{R}}_{D,j}[x,y]$ is a vector space, it is enough to show these claims for basis vectors. Since $X\subset C_D(r)$ is an ellipsoidal $t$-design if and only if $\frac{1}{r}\subset C_D(1)$ is an ellipsoidal $t$-design, it's enough to consider $r=1$. For $D\equiv 1,2\pmod{4}$, $H^{\mathbb{R}}_{D,j}[x,y]=\langle\text{Re}(x+\sqrt{-D}y)^j,\text{Im}(x+\sqrt{-D}y)^j\rangle$. By the parametrization of $C_D(1):x^2+Dy^2=1$ as $\gamma:=\{(\cos{\theta},\sin{\theta}/\sqrt{D})|0\leq \theta\leq 2\pi\}$, we have \begin{align*} \int_{C_D(1)} \dfrac{\text{Re}(x+\sqrt{-D}y)^j}{\sqrt{x^2/D^2+y^2}}d\sigma(x,y) &= \int^{2\pi}_{0}\dfrac{\text{Re}(\cos{\theta}+\sqrt{-D}(\sin{\theta}/\sqrt{D}))^j}{\sqrt{\cos{\theta}^2/D^2+\sin{\theta}^2/D}}\sqrt{\sin{\theta}^2+\cos{\theta}^2/D}d\theta \\ &=\sqrt{D}\int^{2\pi}_{0}\text{Re}(\cos{\theta}+i\sin\theta)^jd\theta =\sqrt{D}\int_{S^1}\text{Re}(x+i y)^jdz =0. \end{align*} Since $\text{Re}(x+i y)^j$ is harmonic, the last integral over $S^1$ is $0$. A similar argument shows that $$\int_{C_D(1)} \dfrac{\text{Im}(x+\sqrt{-D}y)^j}{\sqrt{x^2/D^2+y^2}}d\sigma(x,y)=0.$$ If $D\equiv 3\pmod{4}$, $H^{\mathbb{R}}_{D,j}[x,y]=\langle\text{Re}(x+\frac{1+\sqrt{-D}}{2}y)^j,\text{Im}(x+\frac{1+\sqrt{-D}}{2}y)^j\rangle$. By the parametrization of $C_D(1):x^2+xy+\frac{1+D}{4}y^2=1$ as $\gamma:=\{(\cos{\theta}-\sin{\theta}/\sqrt{D},2\sin{\theta}/\sqrt{D}):0\leq \theta\leq 2\pi\}$, we have \begin{align*} & \int_{C_D(1)} \dfrac{\text{Re}(x+(1+\sqrt{-D})y/2)^j}{\sqrt{20x^2+(D^2+2D+5)y^2+(20+4D)xy}}d\sigma(x,y)\\ = &\int^{2\pi}_{0}\frac{\text{Re}(\cos{\theta}-\sin{\theta}/\sqrt{D}+(1+\sqrt{-D}\sin{\theta}/\sqrt{D})^j}{\sqrt{4D\sin{\theta}^2+20\cos{\theta}^2+8\sqrt{D}\sin{\theta}\cos{\theta}}}\sqrt{\sin{\theta}^2+5\cos{\theta}^2/D+2\sin{\theta}\cos{\theta}/\sqrt{D}}d\theta \\ = &\dfrac{1}{2\sqrt{D}}\int^{2\pi}_{0}\text{Re}(\cos{\theta}+i\sin\theta)^jd\theta = \dfrac{1}{2\sqrt{D}}\int_{S^1}\text{Re}(x+i y)^jdz=0. \end{align*} A similar argument shows that $$\int_{C_D(1)}\dfrac{P(x)}{\sqrt{20x^2+(D^2+2D+5)y^2+(20+4D)xy}}d\sigma(x,y)=0.$$ \end{proof} \section{ellipsoidal T-Designs} Here we prove Theorem \ref{Main}, the construction of ellipsoidal $T$-designs arising from the ring of integers of imaginary quadratic fields with class number 1. We use the theory of theta functions with complex multiplication. Throughout, we shall assume that $D\in\{1,2,3,7,11,19,43,67,163\}.$ \subsection{Theta functions} Given an $n$-dimensional lattice $\Lambda$ and a polynomial $P(x)$ of degree $j$ in $n$ variables, the theta function of $P(x)$ over the lattice $\Lambda$ is defined by the Fourier series (note $q:=e^{2\pi iz}$) \begin{equation} \Theta(\Lambda,P;z):=\sum_{x\in\Lambda} P(x)q^{N(x)} = \Theta(\Lambda,P;z)=\sum^\infty_{n=0} a(\Lambda,P,n)q^n, \end{equation} where $N(x)$ is the standard norm in $\mathbb{R}^n$. The theta functions for $\Lambda_D= \mathcal{O}_{D}$ play an important role in the study of ellipsoidal $T$-designs. Namely, if $\Theta(\Lambda_D,P;z)=\sum^\infty_{r=0} a(\Lambda_D,P,r)q^r,$ then \begin{equation}\label{Coeff} a(\Lambda_D,P,r)= \mathlarger\sum_{(x,y)\in \Lambda^r_D}P(x,y). \end{equation} The theta function $\Theta(\Lambda_D,P;z)\in \mathcal{M}_k(\Gamma_0(4D),\chi)$, the space of holomorphic modular forms with weight $k=j+1$ and nebentypus $\chi(A)=(\frac{-D}{d}),$ where $A=\Big(\begin{array}{cc} a & b \\ c & d \end{array}\Big)$ \cite[Thm~10.8]{AF}. Moreover, $\Theta(\Lambda_D,P;z)$ is a cusp form when $j>0$. To ease the study of these theta function, it is convenient to introduce the following the polynomials for each $j\geq 1$: \begin{equation} R_{D,j}(x,y):=\begin{cases} \text{Re}(x+\sqrt{-D}y)^j \ \ \ \ \ & \text{ if } D\equiv 1,2\pmod{4}, \\ \text{Re}(x+\frac{1+\sqrt{-D}}{2}y)^j \ \ \ \ \ & \text{ if } D\equiv 3\pmod{4}, \\ \end{cases} \end{equation} and \begin{equation} I_{D,j}(x,y):=\begin{cases} \text{Im}(x+\sqrt{-D}y)^j \ \ \ \ \ & \text{ if } D\equiv 1,2\pmod{4}, \\ \text{Im}(x+\frac{1+\sqrt{-D}}{2}y)^j \ \ \ \ \ & \text{ if } D\equiv 3\pmod{4}. \\ \end{cases} \end{equation} By definition, we have that $H^{\mathbb{R}}_{D,j}[x,y]=\langle R_{D,j}(x,y),I_{D,j}(x,y)\rangle$. In particular, $\Theta(\Lambda_D,R_{D,j};z)$ and $\Theta(\Lambda_D,I_{D,j};z)$ are cusp forms. Theorem \ref{CD} together with the discussion above gives the following lemma which transforms the problem of determining ellipsoidal $T$-designs into the vanishing of certain coefficients of special theta functions. \begin{lemma}\label{Equiv} The norm $r$ shell $\Lambda^r_D=\Lambda_D\cap C_D(r)$ is an ellipsoidal $T$-design if and only if $a(\Lambda_D,R_{D,j},r)=0$ and $a(\Lambda_D,I_{D,j},r)=0$ for all $j\in T$. \end{lemma} We require some standard facts from the theory of newforms. Since $\mathcal{O}_D$ has class number 1, each {\it Hecke character} mod $\mathcal{O}_D$ is defined by its values on principal ideals. Let $(\alpha)\subset \mathcal{O}_D$ be a principal ideal. Let $u_D$ be the number of units in $\mathcal{O}_D$, namely \begin{equation} u_D:=\begin{cases} 4 \ \ \ \ \ & \text{ if } D=1, \\ 6 \ \ \ \ \ & \text{ if } D=3, \\ 2 \ \ \ \ \ & \text{ otherwise. } \end{cases} \end{equation} For each positive $j_D\equiv 0\pmod{u_D}$, define Hecke characters mod $\mathcal{O}_D$ by: $$\zeta_{j_D}((\alpha))=\Big(\dfrac{\alpha}{|\alpha|}\Big)^{j_D}$$ Then by \cite[Thm~4.8.2]{Miyake}, we have the following well known lemma about the modular form \begin{equation*} f_{j_D}(\zeta_{j_D};z):=\begin{cases} \Theta(\Lambda_D,(x+\sqrt{-D}y)^j;z) & \text{ if } D\equiv 1,2\pmod{4}, \\ \Theta\Big(\Lambda_D,\Big(x+\frac{1+\sqrt{-D}}{2}y\Big)^j;z\Big) & \text{ if } D\equiv 3\pmod{4} \end{cases} \end{equation*} \begin{lemma}\label{newform} Assuming the notations above, we have \begin{equation*} f_{j_D}(\zeta_{j_D};z)=\sum_{(\alpha)\subset \mathcal{O}_D} \zeta_{j_D}((\alpha))N(\alpha)^{j/2} q^{N(\alpha)} \in \mathcal{S}_{k_D}(\Gamma_0(N),\chi), \end{equation*} the space of cusp forms of weight $k_D=j_D+1$ with nebentypus $\chi\pmod{N}$. Here $N:=|\Delta_{\mathcal{O}_D}|$, the absolute value of the discriminant of $\mathcal{O}_D$. Moreover, $f_{j_D}(\zeta_{j_D};z)$ is a {\it newform.} \end{lemma} \subsection{Other Propositions and Lemmas} Recall that $\Lambda^r_D=C_D(r)\cap \mathcal{O}_D$. Using well known facts about the positive definite binary quadratic forms corresponding to class number 1 norm forms, we have the following lemma. \begin{lemma}\label{prime-decom} Suppose $r$ is a positive integer. Then $\Lambda^r_D$ is nonempty if and only if $ord_p(r)$ is even for every prime $p\nmid r$ for which $\Lambda^p_D$ is nonempty. \end{lemma} Rewriting (\ref{Coeff}), we have \begin{equation} a(\Lambda_D,P,r)= \mathlarger\sum_{(x,y)\in\Lambda^r_D}P(x,y). \end{equation} Lemma \ref{Equiv} implies that $\Lambda^r_D$ is an ellipsoidal $T$-design if and only if $a(\Lambda_D,R_{D,j},r)$ and $a(\Lambda_D,I_{D,j},r)$ vanish for all $j\in T$. Since $\Lambda^r_D$ is antipodal ($i.e.$ $-\Lambda^r_D=\Lambda^r_D$ for all $r$), $a(\Lambda_D,R_{D,j},r)$ and $a(\Lambda_D,I_{D,j},r)$ are 0 for all $j\in \mathbb{Z}^+\setminus2\mathbb{Z}^+$. Therefore, we have that following proposition. \begin{proposition}\label{odd} Suppose $r\in \mathbb{Z}^+$ such that $\Lambda^r_D$ is nonempty. Then $\Lambda^r_D$ is an ellipsoidal $\mathbb{Z}^+\setminus 2\mathbb{Z}^+$-design. \end{proposition} Our objective is to find maximal set $T_D$ for which $\Lambda^r_D$ is ellipsoidal $T$-design. By proposition above we have that $\mathbb{Z}^+\setminus 2\mathbb{Z}^+\subset T_D$. So we only look for all even $j$ which can be in $T_D$. \begin{proposition}\label{I} Suppose $j\equiv 0\pmod{2}$, and $r\in \mathbb{Z}^+$. Then the following are true:\\ 1) We have that $a(\Lambda_D,I_{D,j},r)=0.$\\ 2) We have that $a(\Lambda_D,R_{D,j},r)= \begin{cases} \mathlarger\sum_{(x_0,y_0)\in \Lambda^r_D}(x+\sqrt{-D}y)^j \ \ \ \ & \text{\rm if } D\equiv 1,2\pmod{4}, \\ \mathlarger\sum_{(x_0,y_0)\in \Lambda^r_D}\Big(x+\frac{1+\sqrt{-D}}{2}y\Big)^j \ \ \ \ & \text{\rm if } D\equiv 3\pmod{4} \end{cases}$ \end{proposition} \begin{proof} Part{\it (2)} is an obvious consequence of part{\it (1)}. So it is enough to prove part{\it (1)}. The idea is to show that points in $\Lambda^r_D$ occur in pairs on which value of $I_{D,j}$ cancel. If $D\equiv 1,2\pmod{4}$, then $I_{D,j}={\rm Im}(x+\sqrt{-D}y)^j$. In this case $(a,b), (a,-b)\in \Lambda^r_D$ such that $I_{D,j}(a,b)+I_{D,j}(a,-b)=0$. This is true because each term of $I_{D,j}(x,y)$ has odd power in both the variables $x,y$. If $D\equiv 3\pmod{4}$, then $I_{D,j}={\rm Im}((x+\frac{1}{2}y)+\frac{\sqrt{-D}}{2}y)^j$. In this case $(a,b),(a+b,-b)\in \Lambda^j_D$ such that $I_{D,j}(a,b)+I_{D,j}(a+b,-b)=0$. This is because each term of $I_{D,j}(x,y)$ has odd power in $x+y/2,y.$ \end{proof} We notice that if $(x_0,y_0)\in \mathcal{O}_D,$ then we have \begin{equation}\label{unit} \sum_{\alpha_D \in \mathcal{O}_D:|\alpha_D|=1}R_{D,j}(\alpha_D(x_0,y_0))=R_{D,j}(x_0,y_0)\sum_{\alpha_D \in \mathcal{O}_D:|\alpha_D|=1}\alpha^j_D. \end{equation} \begin{proposition}\label{R} If $r\geq 1$, $1\leq j\not\equiv 0\pmod{u_D}$, and $\Lambda^r_D$ nonempty, then $a(\Lambda^r_D,R_{D,j},r)=0$ \end{proposition} \begin{proof} The idea is that if $(x_0,y_0)\in \Lambda^r_D$ then $\alpha_D(x_0,y_0) \in \Lambda^r_D$ where $\alpha_D$ is a unit in $\mathcal{O}_D$. Therefore enough to show that the sum in RHS of (\ref{unit}) is 0. For $D=1$, number of units in $\mathcal{O}_D,$ $u_D=4$ which are $\{1,-1,i,-i\}$. We have $1^j+(-1)^j+i^j+(-i)^j=0.$ For $D=3$, number of units in $\mathcal{O}_D,$ $u_D=6$ which are $\{ \pm 1 , \frac{\pm 1\pm\sqrt{-3}}{2}\}$. A brute force calculation shows the result. For other $D$, the number of units in $\mathcal{O}_D,$ $u_D=2$ which are $\{1,-1\}$. For all $j$ odd, $(1)^j+(-1)^j=0$ \end{proof} From here on we will only consider the theta function $\Theta\Big(\Lambda_D,\frac{1}{u_D}R_{D,j};z\Big)$ so let's give its coefficients a shorthand. \begin{equation}\label{Theta-short} \Theta\Big(\Lambda_D,\frac{1}{u_D}R_{D,j};z\Big)=\sum^\infty_{r=0} a(D,j,r)q^r \end{equation} Proposition \ref{I} together with Lemma \ref{newform} give us that if $j\equiv0\pmod{u_D}$, then the theta function $\Theta\Big(\Lambda_D,\frac{1}{u_D}R_{D,j};z\Big)\in \mathcal{S}_{j+1}(\Gamma_0(N),\chi)$ is a Hecke eigenform. So we have the following lemma. \begin{lemma}\label{Hecke prop} Suppose $j\in u_D\mathbb{Z}^+$. Then the following is true:\\ 1) If $\gcd{(r_1,r_2)}=1$ then $$ a(D,j,r_1r_2)=a(D,j,r_1)a(D,j,r_2) $$ 2) For $p$ prime and $\alpha>0$, we have $$ a(D,j,p^{\alpha})= a(D,j,p)a(D,j,p^{\alpha-1})-\chi(p)p^ja(D,j,p^{\alpha-2}) $$ 3) For $p$ prime and $\alpha>0$, we have $$ a(D,j,p^{\alpha})= a(D,j,p)^{\alpha}\pmod{p} $$ \end{lemma} Suppose $p$ be a prime such that $\Lambda^p_D$ be nonempty. Let $(x_p,y_p)\in \Lambda^p_D$ and $j\equiv 0\pmod{u_D}$. When $p=D$ then it ramifies in $\mathcal{O}_D$ and there are exactly $u_D$ points in $\Lambda^p_D$. From (\ref{unit}) we have $a(D,j,p)=R_{D,j}(x_p,y_p).$ If $p\not=D$ then it's unramified and we get exactly $2u_D$ solutions. In this case $a(D,j,p)=2R_{D,j}(x_p,y_p).$ \begin{lemma}\label{non-zero} Suppose $j\in u_D\mathbb{Z}^+$ and $p$ be an odd prime such that $\Lambda^p_D$ is nonempty. Let $(x_p,y_p)\in \Lambda^p_D$ then $R_{D,j}(x_p,y_p)\not\equiv 0\pmod{p}$. In particular, $a(D,j,p)$ is non-zero. \end{lemma} \begin{proof} We will consider two cases, $D\equiv 1,2\pmod{4}$ and $D\equiv 3\pmod{4}$. Proof is essentially same in both the cases.\\ If $D\equiv 1,2 \pmod{4}$ then $p=x_p^2+Dy_p^2$, in particular $x_p\not\equiv 0\pmod{p}$. we consider the binomial expansion \begin{align*} R_{D,j}(x_p,y_p) &= \text{Re}(x_p+\sqrt{-D}y_p)^j \\ &= \frac{1}{2}\sum^{j/2}_{n=0}\binom{j}{2n}x^{j-2n}_p(-1)^n(Dy^2_p)^n = \frac{1}{2}\sum^{j/2}_{n=0}\binom{j}{2n}x^{j-2n}_p(-1)^n(p-x^2_p)^n \\ &\equiv \frac{1}{2}x^j_p\sum^{j/2}_{n=0}\binom{j}{2n} \equiv 2^{j-2}x^j_p \not\equiv 0 \pmod{p} \end{align*} If $D\equiv 1,2 \pmod{4}$ then $p=(x_p+y_p/2)^2+Dy_p^2/4$, in particular $x_p+y_p/2\not\equiv 0\pmod{p}$. we consider the binomial expansion \begin{align*} R_{D,j}(x_p,y_p) &= \text{Re}\Big(x_p+y_p/2+\sqrt{-D}y_p/2\Big)^j = \frac{1}{2} \sum^{j/2}_{n=0}\binom{j}{2n}\Big(x_p+\frac{y_p}{2}\Big)^{j-2n}(-1)^n\Big(\frac{Dy^2_p}{4}\Big)^n \\ &= \frac{1}{2} \sum^{j/2}_{n=0}\binom{j}{2n}\Big(x_p+\frac{y_p}{2}\Big)^{j-2n}(-1)^n\Big(p-\Big(x_p+\frac{y_p}{2}\Big)^2\Big)^n \\ &\equiv \frac{1}{2} \Big(x_p+\frac{y_p}{2}\Big)^j\sum^{j/2}_{n=0}\binom{j}{2n} \equiv 2^{j-2}\Big(x_p+\frac{y_p}{2}\Big)^j \not\equiv 0 \pmod{p} \end{align*} \end{proof} \begin{proposition}\label{2} For prime 2, $\Lambda^2_D$ is nonempty only for $D=1,2,7$. In this case $a(D,j,2)$ does not vanish for all $j\in 2\mathbb{Z}^+$. Moreover, we have that $a(7,j,2)\equiv1\pmod{2}$ \end{proposition} \begin{proof} For $D=1,2$, $2|\Delta_{\mathcal{O}_D}(=-4D)$ so the ideal $(2)$ is ramified in $\mathcal{O}_D$, in particular there are elements of norm 2. For $D\in\{3,7,11,19,43,67,163\}$, $2\nmid\Delta_{\mathcal{O}_D}(=-D)$. So the ideal $(2)$ is unramified in $\mathcal{O}_D$. Here we need to check whether 2 splits or not. We have the condition that 2 splits if and only if $-D\equiv 1\pmod{8}$. Only $D=7$ satisfies the condition. A brute force calculation shows that $a(1,j,2)=(1+i)^j\not=0$, $a(2,j,2)=i^j2^{j+1}\not=0,$ and $a(7,j,2)= 4{\rm Re}\Big(\frac{1+\sqrt{-7}}{2}\Big)^j\not=0.$ We prove that $a(7,j,2)\equiv1\pmod{2}$ using induction on even $j$. First, note that $a(7,2,2)=-3\equiv1\pmod{2}$. Now we assume that $a(7,j,2)\equiv1\pmod{2}$, which implies that ${\rm Re}\Big(\frac{1+\sqrt{-7}}{2}\Big)^j=(2k+1)/2$ for some $k$. The norm of $\Big(\frac{1+\sqrt{-7}}{2}\Big)^j$ is even, so we get that ${\rm Im}\Big(\frac{1+\sqrt{-7}}{2}\Big)^j=\sqrt{7}(2k'+1)/2$ for some $k'$. An easy calculation shows that $a(7,j+2,2)=-3{\rm Re}\Big(\frac{1+\sqrt{-7}}{2}\Big)^j-\sqrt{7}{\rm Im}\Big(\frac{1+\sqrt{-7}}{2}\Big)^j\equiv1\pmod{2}.$ \end{proof} \subsection{Proof of Theorem \ref{Main}} Proposition \ref{odd}, \ref{I} and \ref{R} together imply that $a(\Lambda_D,R_{D,j},r)$ and $a(\Lambda_D,I_{D,j},r)$ vanish for all $j\not\equiv 0\pmod{u_D}$, which implies that every nonempty shell $\Lambda^r_D$ is an ellipsoidal $T_D$-design (remember that $T_D=\mathbb{Z}^+\setminus u_D\mathbb{Z}^+$). Now we prove the maximality of $T_D$. We show that $a(D,j,r)\not=0$ (note that $a(D,j,r)= \frac{1}{u_D}a(\Lambda_D,R_{D,j},r)$) for all $j\not\in T_D$ and $\Lambda^r_D$ nonempty. By Lemma \ref{Hecke prop}, enough to take $r$ to be a prime power. Suppose $p$ be a prime and $\alpha\geq 1$ be such that $\Lambda^{p^{a}}_D\not=\phi$. There are two cases possible, either $\Lambda^p_D$ is empty or it is not. First suppose $\Lambda^p_D$ is nonempty. If $p$ is 2 then $a(D,j,2)\not=0$ by Proposition \ref{2}. By part{\it (2)} of Lemma \ref{Hecke prop}, we have that $a(D,j,2^{\alpha})=a(D,j,2)^{\alpha}\not=0$ for $D=1,2$ since $\chi(2)=0$. When $D=7$ then part{\it (3)} of Lemma \ref{Hecke prop}, we have $a(7,j,2^{\alpha})\not=0$. If $p$ is an odd prime, then Lemma \ref{non-zero} implies that $a(D,j,p)\not=0$. Now using part{\it (3)} of Lemma \ref{Hecke prop} again, we have $a(D,j,p^{\alpha})\not=0$. Suppose $\Lambda^p_D$ is empty then $a(D,j,p)=0$ and Lemma \ref{prime-decom} implies $\alpha$ is even. Now by part{\it (2)} of Lemma \ref{non-zero}, we get $a(D,j,p^{\alpha})=p^{j\alpha/2}\not=0$ (note that this case includes 2 too). So we get that $a(D,j,p^{\alpha})\not=0$ whenever $\Lambda^{p^{\alpha}}_D$ is nonempty.

https://arxiv.org/abs/1909.00801

Optimal energy decay in a one-dimensional wave-heat-wave system

Harnessing the abstract power of the celebrated result due to Borichev and Tomilov (Math.\ Ann.\ 347:455--478, 2010, no.\ 2), we study the energy decay in a one-dimensional coupled wave-heat-wave system. We obtain a sharp estimate for the rate of energy decay of classical solutions by first proving a growth bound for the resolvent of the semigroup generator and then applying the asymptotic theory of $C_0$-semigroups. The present article can be naturally thought of as an extension of a recent paper by Batty, Paunonen, and Seifert (J.\ Evol.\ Equ.\ 16:649--664, 2016) which studied a similar wave-heat system via the same theoretical framework.

\chapter[#1]{#1\\[1ex]\large#2}} \theoremstyle{theorem} \newtheorem{theorem}{Theorem}[section] \newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma} \newtheorem*{lemma*}{Lemma} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \newtheorem{example}[theorem]{Example} \newtheorem{opq}[theorem]{Open Question} \newtheorem{xca}[theorem]{Exercise} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \numberwithin{equation}{section} \newcommand{\mathcal{S}}{\mathcal{S}} \newcommand{\mathcal{B}}{\mathcal{B}} \newcommand{\mathcal{G}}{\mathcal{G}} \newcommand{\mathbb{Z}}{\mathbb{Z}} \newcommand{\mathbb{R}}{\mathbb{R}} \newcommand{\mathbb{N}}{\mathbb{N}} \newcommand{\mathbb{C}}{\mathbb{C}} \newcommand{\mathbb{Q}}{\mathbb{Q}} \newcommand{\mathcal{F}}{\mathcal{F}} \newcommand{\mathcal{H}}{\mathcal{H}} \newcommand{\int_\Omega^\oplus H_s \,d\mu(s)}{\int_\Omega^\oplus H_s \,d\mu(s)} \newcommand{\tilde{u}}{\tilde{u}} \newcommand{\tilde{v}}{\tilde{v}} \newcommand{\tilde{f}}{\tilde{f}} \newcommand{\tilde{g}}{\tilde{g}} \newcommand{\tilde{a}}{\tilde{a}} \newcommand{\tilde{b}}{\tilde{b}} \newcommand{(u,v,w,\ut,\vt)}{(u,v,w,\tilde{u},\tilde{v})} \newcommand{(f,g,h,\ft,\gt)}{(f,g,h,\tilde{f},\tilde{g})} \newcommand{e^{\sqrt{\lambda}}}{e^{\sqrt{\lambda}}} \newcommand{e^{-\sqrt{\lambda}}}{e^{-\sqrt{\lambda}}} \newcommand{\lambda}{\lambda} \newcommand{\mathbf{M}}{\mathbf{M}} \newcommand{\sqrt{\lambda}}{\sqrt{\lambda}} \newcommand{\tilde{h}}{\tilde{h}} \newcommand\longmapsfrom{\mathrel{\reflectbox{\ensuremath{\longmapsto}}}} \newcommand{\abs}[1]{\lvert#1\rvert} \DeclareMathOperator{\spn}{span} \DeclareMathOperator{\esssup}{ess-sup} \DeclareMathOperator{\ob}{ob} \DeclareMathOperator{\real}{Re} \DeclareMathOperator{\imag}{Im} \DeclareMathOperator{\diag}{diag} \DeclareMathOperator{\Ker}{Ker} \DeclareMathOperator{\Ran}{Ran} \DeclareMathOperator{\ran}{Ran} \DeclareMathOperator{\dist}{dist} \DeclareMathOperator{\spann}{span} \title[Energy decay in a 1-D wave-heat-wave system]{Optimal energy decay in a one-dimensional wave-heat-wave system} \author[A.C.S.\ Ng]{Abraham C.S.\ Ng} \address[A.C.S.\ Ng]{St Edmund Hall, Queen's Lane, Oxford OX1 4AR, UK} \email{abraham.ng@maths.ox.ac.uk} \begin{document} \begin{abstract} Harnessing the abstract power of the celebrated result due to Borichev and Tomilov (Math.\ Ann.\ 347:455--478, 2010, no.\ 2), we study the energy decay in a one-dimensional coupled wave-heat-wave system. We obtain a sharp estimate for the rate of energy decay of classical solutions by first proving a growth bound for the resolvent of the semigroup generator and then applying the asymptotic theory of $C_0$-semigroups. The present article can be naturally thought of as an extension of a recent paper by Batty, Paunonen, and Seifert (J.\ Evol.\ Equ.\ 16:649--664, 2016) which studied a similar wave-heat system via the same theoretical framework. \end{abstract} \subjclass[2010]{35M33, 35B40, 47D06 (34K30).} \keywords{Wave equation, heat equation, coupled, energy, rates of decay, $C_0$-semigroups, resolvent estimates.} \maketitle \section{Introduction} In this article, we apply the theorem of Borichev-Tomilov \cite[Theorem 4.1]{BoTo} to a one-dimensional system with coupled wave and heat parts. This application is modelled upon the 2016 paper of Batty, Paunonen, and Seifert \cite{BPS1} where the `optimal energy decay in a one-dimensional coupled wave-heat system' with finite Neumann wave and Dirichlet heat parts was studied by analysing the following system: \begin{equation}\label{WHeq1} \begin{cases}\begin{aligned} & u_{tt}(\xi,t) = u_{\xi\xi}(\xi,t), & \xi \in (-1,0), \ & t>0,\\ & w_t(\xi,t) = w_{\xi\xi}(\xi,t), & \xi \in (0,1), \ & t>0, \\ & u_t(0,t) = w(0,t), \ \ \ u_\xi(0,t) = w_\xi(0,t), & & t>0, \\ & u_\xi(-1,t) = 0, \ \ \ w(1,t) = 0, & & t>0, \\ & u(\xi,0)=u(\xi), \ \ \ u_t(\xi,0) = v(\xi) \ & \xi \in (-1,0), \\ & w(\xi,0) = w(\xi), \ & \xi \in (0,1), \end{aligned}\end{cases} \end{equation} where the initial data $u,v,$ and $w$ lived in $H^1(-1,0), L^2(-1,0)$ and $L^2(0,1)$ respectively. The energy was then defined, given a vector of initial data $x=(u,v,w)$, as $$E_x(t) = \frac{1}{2}\int_{-1}^1 |u_\xi(\xi,t)|^2 + |u_t(\xi,t)|^2 + |w(\xi,t)|^2 \ d\xi, \ \ \ t\geq 0,$$ with all the functions being understood to have been extended by zero in $\xi$ to the interval $(-1,1)$. If the solution is sufficiently regular, a routine calculation via integration by parts shows that $$E'_x(t) = -\int_0^1|w_\xi(\xi,t)|^2\ d\xi, \ \ \ t \geq 0,$$ and, in particular, that the energy of any such solution is non-increasing with respect to time. The main goal of analysing such a model is to quantitatively estimate the rate of energy decay of a given solution. The system (\ref{WHeq1}) was first studied (with Dirichlet boundary at $\xi=-1$ and a slightly different coupling condition) in \cite{ZZ1}, yielding the sharp decay rate $E_x(t) = O(t^{-4}), t\to\infty$ (see below for the meaning of `big O' notation). The approach in \cite{ZZ1} relied on a rather complicated spectral analysis used in conjunction with the theory of Riesz spectral operators. In contrast to \cite{ZZ1}, however, the approach in \cite{BPS1} was based on the semigroup methods of non-uniform stability pioneered by Batty and Duyckaerts in \cite{BaDu}, widely popularised by Borichev and Tomilov in \cite{BoTo}, and largely completed by Rozendaal, Seifert, and Stahn in \cite{RSS}, greatly simplifying the analysis necessary to obtain the rate of decay. The motivation of studying models like this and, in particular, the one in this article presented below, stems mainly from the study of fluid-structure models where, often in higher-dimensional settings, the Navier-Stokes equations (the fluid half) are coupled with the nonlinear elasticity equation (the structure half). We refer to \cite[Section 1]{BPS1} and \cite{AvTr} for surveys of similar problems (see also \cite{BPS2} where the same approach with suitable adjustments is applied to study a wave-heat system on a rectangular domain). In this article, we add an extra wave component to the system (\ref{WHeq1}) and take Dirichlet boundary conditions on both ends, analysing the following wave-heat-wave system: \begin{equation}\label{WHWeq1} \begin{cases}\begin{aligned} & u_{tt}(\xi,t) = u_{\xi\xi}(\xi,t), & \xi \in (0,1), \ & t>0,\\ & w_t(\xi,t) = w_{\xi\xi}(\xi,t), & \xi \in (1,2), \ & t>0, \\ & \tilde{u}_{tt}(\xi,t) = \tilde{u}_{\xi\xi}(\xi,t), & \xi \in (2,3), \ & t>0, \\ & u(0,t) = \tilde{u}(3,t) = 0, & & t>0, \\ & u_t(1,t) = w(1,t),\ \ \ u_\xi(1,t) = w_\xi(1,t), & & t>0, \\ & \tilde{u}_t(2,t) = w(2,t),\ \ \ \tilde{u}_\xi(2,t) = w_\xi(2,t), & & t>0, \\ & u(\xi,0) = u(\xi), \ \ \ u_t(\xi,0) = v(\xi), & \xi \in (0,1), \ & \\ & w(\xi,0) = w(\xi), & \xi \in (1,2), \ & \\ & \tilde{u}(\xi,0) = \tilde{u}(\xi), \ \ \ \tilde{u}_t(\xi,0) = \tilde{v}(\xi), & \xi \in (2,3). \ & \\ \end{aligned}\end{cases} \end{equation} The initial data is required to satisfy $u = u(\xi,0) \in H^1(0,1), v = u_t(\xi,0) \in L^2(0,1), w = w(\xi,0) \in L^2(1,2), \tilde{u} = \tilde{u}(\xi,0) \in H^1(2,3),$ and $\tilde{v} = \tilde{u}_t(\xi,0) \in L^2(2,3)$. As in \cite{BPS1}, the aim here is to find a quantitative estimate for the rate of energy decay of a given solution. Given a vector of initial data $x=(u,v,w,\tilde{u},\tilde{v})$ satisfying the conditions above, we similarly define the energy of the corresponding solution as $$E_x(t) = \frac{1}{2}\int_0^3 |u_\xi(\xi,t)|^2 + |u_t(\xi,t)|^2 + |w(\xi,t)|^2 + |\tilde{u}_\xi(\xi,t)|^2 + |\tilde{u}_t(\xi,t)|^2 \ d\xi, \ \ \ t\geq 0.$$ Again, all functions have been extended by zero in $\xi$ to the interval $(0,3)$. Provided we have sufficient regularity of the solution, a simple calculation via integration by parts shows that $$E'_x(t) = \real{\left\{\tilde{u}_\xi(3,t)\overline{\tilde{u}_t(3,t)} - u_\xi(0,t)\overline{u_t(0,t)}\right\}} - \int_1^2|w_\xi(\xi,t)|^2\ d\xi, \ \ \ t\geq0.$$ Since $u_t(0,t) = \frac{\partial}{\partial t}u(0,t) = \tilde{u}_t(3,t) = \frac{\partial}{\partial t}\tilde{u}(3,t) = 0$ for $t >0$, the energy of any such solution is non-increasing with respect to time. The remaining sections are devoted to obtaining a sharp quantitative estimate for the rate of this decay for classical solutions of (\ref{WHWeq1}), but first, we detail below, the mostly standard notation used in this article. Closely following the notation of \cite{BPS1}, the domain, kernel, range, spectrum, and range of a closed operator $A$ acting on a Hilbert space (always complex by assumption) will be denoted by $D(A), \Ker{A}, \Ran{A}, \sigma(A)$ and $\rho(A)$ respectively. For $\lambda \in\rho(A)$, we write $R(\lambda,A)$ to signify the resolvent operator $(\lambda-A)^{-1}$. For $\lambda \in \mathbb{C}$, we define the square root $\sqrt{\lambda}$ by taking the branch cut along the negative real axis, that is, for $\lambda =re^{i\theta}$ where $r\geq0$ and $\theta \in (-\pi,\pi]$, we let $\sqrt{\lambda} = r^{1/2}e^{i\theta/2}$. We also denote the closed complex left half-plane by $\mathbb{C}_- := \{z \in \mathbb{C} : \real{z} <0\}$. Finally, given two functions $f,g:(0,\infty) \to [0,\infty]$ and $a \in [0,\infty]$ fixed, we write $f(t) = O(g(t)),\ t\to \infty$, to indicate that there exists some constant $C>0$ such that $f(t)\leq Cg(t)$ for all $t$ sufficiently large, the so-called `big O notation'. If $g$ is strictly positive for all sufficiently large $t>0$, we write $f(t)=o(g(t)),\ t \to \infty$, to mean that $f(t)/g(t) \to 0$ as $t\to \infty$, the so-called `little o notation'. If $p$ and $q$ are non-negative real-valued quantities, the notation $p \lesssim q$ denotes that $p\leq Cq$ for some constant $C>0$ that is independent of any varying parameters in a given context. \subsection*{Acknowledgements} The author thanks David Seifert and Charles Batty for helpful discussions on the topic of this article and is especially indebted to David for his careful reading and feedback of several drafts of this article. The author is also grateful to the University of Sydney for funding this work through the Barker Graduate Scholarship. \section{Well-posedness -- the Semigroup and its Generator} In this section, we first prove that (\ref{WHWeq1}) is well posed and has solution given by the orbits of a $C_0$-semigroup of contractions $(T(t))_{t\geq0}$, before turning to analyse the spectrum of the generator $A$ of $(T(t))_{t\geq0}$. \subsection{Existence of the Semigroup} We start by recasting (\ref{WHWeq1}) into an abstract Cauchy problem in order to later apply the methods of non-uniform stability. Consider the Hilbert space $$X_0 = H^1(0,1)\times L^2(0,1)\times L^2(1,2) \times H^1(2,3) \times L^2(2,3)$$ and define $$X= \{(u,v,w,\tilde{u},\tilde{v}) \in X_0 : u(0) = \tilde{u}(3) = 0\}$$ endowed with the norm (and corresponding inner product) given by $$\|(u,v,w,\tilde{u},\tilde{v})\|_X^2 = \|u'\|_{L^2}^2 + \|v\|_{L^2}^2 + \|w\|_{L^2}^2+\|\tilde{u}'\|_{L^2}^2+\|\tilde{v}\|_{L^2}^2$$ which is non-degenerate because the fundamental theorem of calculus applied in conjunction with the boundary conditions $u(0) = \tilde{u}(3) = 0$ implies that $\|u\|_{L^2}\lesssim \|u\|_{L^2}$ and $\|\tilde{u}\|_{L^2}\lesssim \|\tilde{u}'\|_{L^2}$. Here and in the rest of the article, the intervals for function spaces appearing as subscripts will often be omitted if they are clear from the context. Let $$X_1=X\cap [H^2(0,1)\times H^1(0,1) \times H^2(1,2) \times H^2(2,3)\times H^1(2,3)]$$ and define the operator $A$ on $X$ by $Ax = (v,u'',w'',\tilde{v},\tilde{u}'')$ for $x=(u,v,w,\tilde{u},\tilde{v})$ in the domain \begin{equation*}\begin{split}D(A) = \{(u,v,w,\tilde{u},\tilde{v}) \in X_1 : & \ v(0) = \tilde{v}(3) = 0, u'(1) = w'(1), \\ & \ \ \ v(1) = w(1), \tilde{u}'(2) = w'(2), \tilde{v}(2) = w(2)\}. \end{split}\end{equation*} \begin{lemma}\label{biglemma1} The following hold: \begin{enumerate}[(i)] \item $A$ is closed; \item $A$ is densely defined; \item $A$ is dissipative; \item $1-A$ is surjective. \end{enumerate} \end{lemma} \begin{proof} (i) Let $x_n = (u_n,v_n,w_n,\tilde{u}_n,\tilde{v}_n) \in D(A)$ be such that $$x_n\to x = (u,v,w,\tilde{u},\tilde{v}), \ Ax_n = (v_n,u_n'',w_n'',\tilde{v}_n,\tilde{u}_n'') \to y = (f,g,h,\tilde{f},\tilde{g})$$ in $X$. Then $u_n$ converges to $u$ in $H^1(0,1)$ and $u_n''$ converges to $g$ in $L^2(0,1)$. Hence \begin{equation}\label{wkd}\int u\varphi'' = \lim_{n\to \infty}\int u_n\varphi'' = \lim_{n\to\infty}\int u_n'' \varphi = \int g\varphi, \ \ \ \varphi \in C_c^\infty(0,1),\end{equation} where the integral is taken over $((0,1),d\xi)$ so that $u \in H^2(0,1)$ and $u'' = g$. As $v_n$ converges to both $v$ and $f$ in $L^2(0,1)$, $v=f$. In particular, $v \in H^1(0,1)$. The same argument shows that $\tilde{u} \in H^2(2,3)$ with $\tilde{u}'' = \tilde{g}$ and $\tilde{v} = \tilde{f} \in H^1(2,3)$. Next, $w_n$ converges to $w$ and $w_n''$ to $h$ in $L^2(1,2)$. Standard Sobolev theory (see for example \cite[Page ~217]{Brezis}) ensures the existence of a constant $C$ such that $$\|\psi'\|_{L^2(1,2)} \leq \|\psi''\|_{L^2(1,2)} + C\|\psi\|_{L^2(1,2)} \ \ \ \psi \in H^2(1,2).$$ Hence, the sequence $w_n'$ is Cauchy and converges to some $H$ in $L^2(1,2)$. Using similar reasoning to that in (\ref{wkd}), we see that $w\in H^2(1,2)$ with $w' = H$ and $w'' = h$. To check that the coupling conditions for $x$ to be in the domain $D(A)$ are satisfied, it is enough to pass to a subsequence $x_{n_k}$ that converges pointwise a.e.\ and note the continuity of $u', v, w', w, \tilde{u}',\tilde{v}$. It follows that $Ax = y$. (ii) Consider the subspace $X_1$ equipped with the $X$ norm, which is dense in $X$. The linear functional $\phi_1 : x = (u,v,w,\tilde{u},\tilde{v}) \mapsto v(0)$ is unbounded on $X_1$, and hence $$X_2 = \Ker \phi_1 = \{(u,v,w,\tilde{u},\tilde{v}) \in X_1 : v(0)=0 \}$$ is dense in $X_1$. Similarly, $$X_3 = \Ker \phi_2 = \{(u,v,w,\tilde{u},\tilde{v}) \in X_2 : v(1)=w(1)\}$$ is dense in $X_2$ where $\phi_2$ is the unbounded linear functional on $X_2$ defined by $x \mapsto v(1)-w(1)$. Again, by considering the unbounded linear functional $\phi_3 : x \mapsto u'(1) - w'(1)$ on $X_3$, we see that $$X_4 = \Ker \phi_3 = \{(u,v,w,\tilde{u},\tilde{v}) \in X_3 : u'(1)=w'(1)\}$$ is dense in $X_3$. The same argument can be repeated for the coupling and boundary conditions for $w,\tilde{u},$ and $\tilde{v}$ to produce a decreasing finite chain of subspaces $$X \supset X_1 \supset X_2 \supset ... \supset D(A),$$ where each subspace is dense in the preceding one under the $X$ norm. Hence $A$ is densely defined. (iii) Let $x \in D(A)$. Assuming the appropriate intervals over which to take the $L^2$ inner products, we have, through integration by parts and the coupling and boundary conditions, \begin{align*}\langle Ax,x\rangle & = \langle v',u'\rangle_{L^2} + \langle u'',v\rangle_{L^2} + \langle w'',w\rangle_{L^2} + \langle \tilde{v}',\tilde{u}'\rangle_{L^2} + \langle \tilde{u}'',\tilde{v}\rangle_{L^2} \\ & = - \overline{\langle u'',v\rangle_{L^2}} + \langle u'',v\rangle_{L^2} - \langle w',w'\rangle_{L^2} - \overline{\langle \tilde{u}'',\tilde{v}\rangle_{L^2}} + \langle \tilde{u}'',\tilde{v}\rangle_{L^2}. \end{align*} Hence $$ \real{\langle Ax,x\rangle} = -\|w'\|_{L^2}^2 \leq 0,$$ showing that $A$ is dissipative. (iv) Though in the setting of this lemma, we only need to work with $1-A$, we perform a procedure here with $\lambda-A$ for general $\lambda \neq 0$ in order to avoid repetition that otherwise would be inevitable in later sections. Note that we are closely following the proof of \cite[Theorem 3.1]{BPS1}. Let $x = (u,v,w,\ut,\vt)$ and $y = (f,g,h,\tilde{f},\tilde{g})$ be in $X$. Then the equation $(\lambda-A)x = y$ can be rewritten as the following system of boundary value problems: \begin{subequations}\begin{align} u'' & = \lambda^2 u-\lambda f -g, & \xi \in (0,1), \label{eqa}\\ v & = \lambda u-f, & \xi \in (0,1), \\ w'' & = \lambda w-h, & \xi \in (1,2), \label{eqc}\\ \tilde{u}'' & = \lambda^2\tilde{u} - \lambda\tilde{f} - \tilde{g}, & \xi \in (2,3), \label{eqd}\\ \tilde{v} & = \lambda \tilde{u} - \tilde{f}, & \xi \in (2,3), \\ u(0) = v(0) = 0, \ \ \ v(1) & = w(1), \ \ \ u'(1) = w'(1), \\ \tilde{u}(3) = \tilde{v}(3)=0, \ \ \ \tilde{v}(2) & = w(2), \ \ \ \tilde{u}'(2) = w'(2). \end{align}\end{subequations} Let $$U_\lambda(\xi) = \frac{1}{\lambda}\int_{0}^\xi \sinh(\lambda(\xi-r))(\lambda f(r)+g(r))\ dr, \ \ \ \xi \in [0,1],$$ which has derivative $$U_\lambda'(\xi) =\int_{0}^\xi \cosh(\lambda(\xi-r))(\lambda f(r)+g(r))\ dr, \ \ \ \xi \in [0,1].$$ The differential equation (\ref{eqa}) with the boundary condition $u(0) = 0$ has the general solution \begin{equation}\label{equ} u(\xi) = a(\lambda)\sinh(\lambda\xi) - U_\lambda(\xi), \ \ \ \xi \in [0,1], \end{equation} where $a(\lambda) \in \mathbb{C}$ is a parameter free to be varied. In particular, \begin{equation}\label{equ'} u'(\xi) = \lambda a(\lambda)\cosh(\lambda\xi) - U_\lambda'(\xi), \ \ \ \xi \in [0,1]. \end{equation}. Clearly $u\in H^2(0,1)$ and hence $v\in H^1(0,1)$ with $v(0)=\lambda u(0)-f(0) =0$. Similarly, the general solution of (\ref{eqd}) with boundary condition $\tilde{u}(3)=0$ can be written as \begin{equation}\label{equt} \tilde{u}(\xi) = \tilde{a}(\lambda)\sinh(\lambda(3-\xi)) + \tilde{U}_\lambda(\xi), \ \ \ \xi \in [2,3], \end{equation} where $\tilde{a}(\lambda) \in \mathbb{C}$ can be varied freely and $$\tilde{U}_\lambda(\xi) = \frac{1}{\lambda}\int_{\xi}^3 \sinh(\lambda(r-\xi))(\lambda\tilde{f}(r)+\tilde{g}(r))\ dr, \ \ \ \xi \in [2,3].$$ Thus \begin{equation}\label{equt'} \tilde{u}'(\xi) = -\lambda\tilde{a}(\lambda)\cosh(\lambda(3-\xi)) + \tilde{U}_\lambda'(\xi), \ \ \ \xi \in [2,3], \end{equation} where $$\tilde{U}_\lambda'(\xi) = -\int_{\xi}^3 \cosh(\lambda(r-\xi))(\lambda\tilde{f}(r)+\tilde{g}(r))\ dr, \ \ \ \xi \in [2,3].$$ Again, it follows that $\tilde{u} \in H^2(2,3)$ and $\tilde{v} \in H^1(2,3)$ with $\tilde{v}(3)=0$. In the same spirit, let $$W_\lambda(\xi) = \frac{1}{\sqrt{\lambda}}\int_{1}^\xi \sinh(\sqrt{\lambda}(\xi-r))h(r)\ dr, \ \ \ \xi \in [1,2],$$ which has derivative $$W_\lambda'(\xi) = \int_{1}^\xi \cosh(\sqrt{\lambda}(\xi-r))h(r)\ dr, \ \ \ \xi \in [1,2].$$ The general solution of (\ref{eqc}) can then be written as \begin{equation}\label{eqw} w(\xi) = b(\lambda)\cosh(\sqrt{\lambda}(\xi-1))+c(\lambda)\sinh(\sqrt{\lambda}(\xi-1)) - W_\lambda(\xi), \ \ \ \xi \in [1,2], \end{equation} where $b(\lambda), c(\lambda) \in \mathbb{C}$ are free parameters and in particular, \begin{equation}\label{eqw'} w'(\xi) = \sqrt{\lambda}b(\lambda)\sinh(\sqrt{\lambda}(\xi-1))+\sqrt{\lambda}c(\lambda)\cosh(\sqrt{\lambda}(\xi-1)) - W_\lambda'(\xi), \ \ \ \xi \in [1,2]. \end{equation} It remains to choose specific constants $a(\lambda),b(\lambda),c(\lambda)$ and $\tilde{a}(\lambda)$ in order to satisfy the coupling conditions. Using (\ref{equ}) and (\ref{eqw}), the requirement $\lambda u(1)-f(1) = v(1) = w(1)$ holds if and only if $$\lambda a(\lambda)\sinh(\lambda) - b(\lambda) = \lambda U_\lambda(1) + f(1).$$ Likewise, the conditions $u'(1) = w'(1)$, $\lambda \tilde{u}(2)-\tilde{f}(2) = w(2)$, and $\tilde{u}'(2) = w'(2)$ are equivalent to $$\lambda a(\lambda)\cosh(\lambda) -\sqrt{\lambda}c(\lambda) = U_\lambda'(1),$$ $$\lambda \tilde{a}(\lambda) \sinh(\lambda) - b(\lambda)\cosh(\sqrt{\lambda})-c(\lambda)\sinh(\sqrt{\lambda}) = -\lambda\tilde{U}_\lambda(2) + \tilde{f}(2) - W_\lambda(2),$$ and $$ -\lambda \tilde{a}(\lambda) \cosh(\lambda) -\sqrt{\lambda}b(\lambda)\sinh(\sqrt{\lambda}) - \sqrt{\lambda}c(\lambda)\cosh(\sqrt{\lambda}) = -\tilde{U}_\lambda'(2) - W_\lambda'(2)$$ respectively. These four equations can be written in matrix form as \begin{equation}\label{eqM} M_\lambda \cdot \begin{pmatrix} a(\lambda) \\ b(\lambda) \\ c(\lambda) \\ \tilde{a}(\lambda) \end{pmatrix} = \mathbf{b}, \end{equation} where \begin{equation} M_\lambda = \begin{pmatrix} \lambda\sinh(\lambda) & -1 & 0 & 0 \\ \lambda \cosh(\lambda) & 0 & -\sqrt{\lambda} & 0 \\ 0 & -\cosh(\sqrt{\lambda}) & -\sinh(\sqrt{\lambda}) & \lambda\sinh(\lambda) \\ 0 & \sqrt{\lambda}\sinh(\sqrt{\lambda}) & \sqrt{\lambda}\cosh(\sqrt{\lambda}) & \lambda\cosh(\lambda) \end{pmatrix} \end{equation} and \begin{equation}\mathbf{b} = \begin{pmatrix}\lambda U_\lambda(1) + f(1) \\ U_\lambda'(1) \\ -\lambda\tilde{U}_\lambda(2) +\tilde{f}(2) - W_\lambda(2) \\ \tilde{U}_\lambda'(2) +W_\lambda'(2) \end{pmatrix}. \end{equation} Thus, (\ref{eqM}) has a solution for any given $y = (f,g,h,\tilde{f},\tilde{g})$ in $X$ if and only if $$\det M_\lambda = -\lambda^2[2\sqrt{\lambda}\cosh(\sqrt{\lambda})\cosh(\lambda)\sinh(\lambda)+\sinh(\sqrt{\lambda})(\lambda\sinh^2(\lambda)+\cosh^2(\lambda))]$$ is non-zero. For $\lambda = 1$, $$\det M_1 = -\sinh(1)[4\cosh^2(1) - 1]\neq 0,$$ proving (4) \end{proof} All the dirty work has now been done (ahead of time). The following theorem follows immediately from Lemma \ref{biglemma1} and the Lumer-Phillips theorem. \begin{theorem}\label{wpthm} $A$ generates a contractive $C_0$-semigroup $T(t)$ on $X$. \end{theorem} \subsection{Spectrum of the Generator} From Theorem \ref{wpthm} and the Hille-Yosida theorem, we know that $\sigma(A)$ is contained in the closed left half-plane. However, we can say more about the spectrum. \begin{theorem}\label{WHWgenspec} The spectrum of $A$ consists of isolated eigenvalues and is given by $$\sigma(A) = \{\lambda \in \mathbb{C}_- : \det M_\lambda = 0\}.$$ In particular, $\sigma(A) \cap i\mathbb{R} = \emptyset$. \end{theorem} We will need the following lemma in order to prove the theorem above. \begin{lemma}\label{compres} If $\lambda \in \rho(A)$, then $R(\lambda,A)$ is a compact operator. \end{lemma} \begin{proof} Let $\lambda \in \rho(A)$. Then $\lambda -A$ is a bijective bounded (and in particular, closed) linear map from $D(A)$ endowed with the graph norm onto $X$. Hence the inverse map $R(\lambda,A)$ maps $X$ isomorphically onto $(D(A),\|\cdot\|_{D(A)})$. Since \begin{align*}\|(u,v,w,\tilde{u},\tilde{v})\|_{D(A)} & = \|(u,v,w,\tilde{u},\tilde{v})\|_X+ \|(v,u'',w'',\tilde{v},\tilde{u}'')\|_X \\ & \lesssim \|u'\|_{L^2} + \|v\|_{L^2} + \|w\|_{L^2} + \|\tilde{u}'\|_{L^2} + \|\tilde{v}\|_{L^2} \\ & \qquad + \|v'\|_{L^2} + \|u''\|_{L^2} + \|w''\|_{L^2} + \|\tilde{v}'\|_{L^2} + \|\tilde{u}''\|_{L^2}, \end{align*} it follows that $(D(A),\|\cdot\|_{D(A)})$ embeds continuously into $$H^2(0,1)\times H^1(0,1) \times H^2(1,2) \times H^2(2,3)\times H^1(2,3)$$ endowed with its natural norm (see \cite[Page 217]{Brezis}). This space in turn embeds compactly into $X$ by the Rellich-Kondrachov theorem of Sobolev theory. Stringing together these embeddings, $R(\lambda,A)$ is a compact operator on $X$. \end{proof} \begin{proof}[Proof of Theorem \ref{WHWgenspec}] We first show that not only is $\lambda-A$ surjective as shown in Lemma \ref{biglemma1} whenever $\det M_\lambda \neq 0$, it is also injective. Indeed, suppose $(\lambda-A)x = 0$. Then, $x$ is obtained in the same way as in the proof of Lemma \ref{biglemma1}(4) with $\mathbf{b} =0$ in (\ref{eqM}). As $\det M_\lambda \neq 0$, we get that $x = 0$. Hence $\lambda-A$ is closed and bijective, so has bounded inverse by the closed graph theorem. In particular, $1 \in \rho(A)$ and so the resolvent is non-empty. The spectral theorem for compact operators used in conjunction with Lemma \ref{compres} implies that the spectrum of $R(1,A)$ consists only of eigenvalues of finite multiplicity with the only possible accumulation point being the origin. By the spectral mapping theorem for the resolvent, $$\sigma(A) = \{1-\nu^{-1} : \nu \in \sigma(R(1,A))\setminus\{0\}\}$$ and furthermore, a simple calculation shows that if $\nu$ is an eigenvalue of $R(1,A)$, then $1- \nu^{-1}$ is an eigenvalue of $A$. Hence $\sigma(A)$ consists only of eigenvalues of finite multiplicity with the only possible accumulation point being at infinity. Thus, $\lambda \in \sigma(A)$ if and only if $\det M_\lambda = 0$. To show the final statement, suppose that $s\in \mathbb{R}$ with $s\neq 0$ and that $x = (u,v,w,\tilde{u},\tilde{v}) \in \Ker (is-A)$. From the proof of Lemma \ref{biglemma1}(3), we have \begin{equation}\label{w=0}0 = \real{\langle (is-A)x,x\rangle} = -\real{\langle Ax,x\rangle} =\|w'\|_{L^2}.\end{equation} Thus $w = (is)^{-1}(w')' =0$. As in the proof for Lemma \ref{biglemma1}(4), we have $$u(\xi) = a(is)\sinh(is\xi), \ \ \ v(\xi) = is\ u(\xi), \ \ \ \xi \in [0,1].$$ The coupling conditions imply that $u'(1) = v(1) =0$. Thus, $$is\ a(is)\cosh(is) = is\ a(is)\sinh(is) =0,$$ implying that $a(is) =0$. Similarly, $\tilde{a}(is) =0$ so that $x=0$. Consider now the case $s=0$. Rewriting $Ax =0$ into component differential equations, we get that $u''=0$ and $v=0$ as well as $w=0'$ as in (\ref{w=0}). As $u'(1)=w'(1)$ and $u'$ is constant, $u' = 0$ and hence $u(0)=0$ implies that $u=0$. Similarly, $v(1)=0$ implies that $w=0$. The same is true for $\tilde{u}$ and $\tilde{v}.$ It follows that $\sigma(A) \cap i\mathbb{R} = \emptyset$. \end{proof} \section{Resolvent Estimates}\label{resolventestimates} We turn now to obtaining an upper bound on the growth of $\|R(is,A)\|$ as $|s| \to \infty$ which will allow us to deduce a quantitative estimate on the rate of energy decay in the next section. \begin{theorem}\label{mainthm} We have $\|R(is,A)\| = O(|s|^{1/2})$ as $|s| \to \infty$. \end{theorem} To prove this theorem, we will need explicit forms for the $a(\lambda),b(\lambda),c(\lambda),\tilde{a}(\lambda)$ found in the proof of Lemma \ref{biglemma1}(4) for the case where $\lambda = is$ and to this end, we invert $M_\lambda$ to get that \begin{equation} (\det M_\lambda)^{-1} C^{T} \mathbf{b} = \begin{pmatrix}a(\lambda) \\ b(\lambda) \\ c(\lambda) \\ \tilde{a}(\lambda)\end{pmatrix}, \end{equation} where $C$ is the cofactor matrix of $M_\lambda$. First, we rewrite $\det M_\lambda$ and define two terms which are ubiquitous in this section: \begin{align*} \det M_\lambda & = -\lambda^2[2\sqrt{\lambda}\cosh(\sqrt{\lambda})\cosh(\lambda)\sinh(\lambda)+\sinh(\sqrt{\lambda})(\lambda\sinh^2(\lambda)+\cosh^2(\lambda)] \\ & = \frac{\lambda^2}{2}[-e^{\sqrt{\lambda}}(\lambda \sinh^2(\lambda)+2\sqrt{\lambda}\cosh(\lambda)\sinh(\lambda)+\cosh^2(\lambda)) \\ & \ \ \ \ \ \ \ \ \ \ \ + e^{-\sqrt{\lambda}}(\lambda \sinh^2(\lambda)-2\sqrt{\lambda}\cosh(\lambda)\sinh(\lambda)+\cosh^2(\lambda))] \\ & = \lambda^2[-e^{\sqrt{\lambda}}T_+^2(\lambda) + e^{-\sqrt{\lambda}}T_-^2(\lambda)], \end{align*} where $$T_+(\lambda) =\frac{1}{2}[\cosh(\lambda)+\sqrt{\lambda}\sinh(\lambda)], \ \ \ T_-(\lambda) = \frac{1}{2}[\cosh(\lambda)- \sqrt{\lambda}\sinh(\lambda)].$$ The functions $T_+$ and $T_-$ are useful because they obey convenient lower bounds on the one hand, and appear many times in the entries of $C = \{c_{ij}\}_{i,j}$ on the other hand. As an example of this, $c_{11}$ is explicitly computed and stated here: \begin{align*} c_{11} & = -\lambda^{3/2}[\cosh(\sqrt{\lambda})\cosh(\lambda) +\sqrt{\lambda}\sinh(\sqrt{\lambda})\sinh(\lambda)] \\ & = -\lambda^{3/2}[e^{\sqrt{\lambda}}T_+(\lambda) + e^{-\sqrt{\lambda}}T_-(\lambda)]. \end{align*} The expressions for the other entries can be found in the appendix. We will also need the following two lemmas, the first of which is proved in \cite[Lemma 3.3]{BPS1} (over the interval $[-1,0]$ rather than $[0,1]$ or $[2,3]$ as we have here). \begin{lemma}\label{1U} There exists a constant $C\geq 0$ such that, for all $f \in H^1(0,1), g\in L^2(0,1), \tilde{f} \in H^1(2,3), \tilde{g} \in L^2(2,3),$ and $\lambda\in i\mathbb{R}$, \begin{align*} \left|\int_{0}^\xi \sinh(\lambda(\xi-r))(\lambda f(r)+g(r))dr\right| & \leq C\|f\|_{H^1}+\|g\|_{L^2}, \ \ \ \xi \in [0,1], \\ \left|\int_{0}^\xi \cosh(\lambda(\xi-r))(\lambda f(r)+g(r))dr\right| & \leq C\|f\|_{H^1}+\|g\|_{L^2}, \ \ \ \xi \in [0,1], \\ \left|\int_{\xi}^3 \sinh(\lambda(r-\xi))(\lambda\tilde{f}(r)+\tilde{g}(r))dr\right| & \leq C\|\tilde{f}\|_{H^1}+\|\tilde{g}\|_{L^2}, \ \ \ \xi \in [2,3], \\ \left|\int_{\xi}^3 \cosh(\lambda(r-\xi))(\lambda\tilde{f}(r)+\tilde{g}(r))dr\right| & \leq C\|\tilde{f}\|_{H^1}+\|\tilde{g}\|_{L^2}, \ \ \ \xi \in [2,3]. \end{align*} \end{lemma} \begin{lemma}\label{pm} For $\lambda\in i\mathbb{R}$ with $|\lambda| \geq \left(\frac{1}{\sqrt{2}}+1\right)^2$, we have $$|T_+(\lambda)|,|\ T_-(\lambda)| \geq 1/4.$ \end{lemma} \begin{proof} We prove this for $T_+(\lambda)$ where $\lambda = is$ with $s\in\mathbb{R}$ and note that $2T_+(is) = \cos(s)+i\sqrt{is}\sin(s)$. Explicit calculation yields \begin{align*}4|T_+(\lambda)|^2 & = \left|\frac{1}{\sqrt{is}}(\sqrt{is}\cos(s)-s\sin(s))\right|^2 \\ & = \frac{1}{|s|}\left(\frac{|s|}{2}\cos^2(s) + \left(\sqrt{\frac{|s|}{2}}\cos(s)-s\sin(s)\right)^2\right),\end{align*} as $\real{\sqrt{\lambda}} \geq 0$ for all $\lambda\in\mathbb{C}$ since we have taken the branch cut of the square root along the negative real axis. In the case where $\cos^2(s) \geq 1/2$, it follows that $4|T_+(\lambda)|^2 \geq 1/4$. However, if $\cos^2(s) < 1/2$, then $|\sin(s)|^2\geq1/2$, so that $$2|T_+(\lambda)| \geq |i\sqrt{is}\sin(s)| - |\cos(s)| \geq \sqrt{\frac{|s|}{2}} - \frac{1}{\sqrt{2}} \geq \frac{1}{2}$$ whenever $|\lambda| = |s| \geq \left(\frac{1}{\sqrt{2}}+1\right)^2$. The case for $T_-(\lambda)$ is similar. \end{proof} \begin{proof}[Proof of Theorem \ref{mainthm}] Let $\lambda =is$ for $s\in \mathbb{R}$ and let $y= (f,g,h,\tilde{f},\tilde{g}) \in Z$, further defining $x = (u,v,w,\tilde{u},\tilde{v}) \in D(A)$ by $x = R(\lambda,A)y.$ As $v = \lambda u -f$ and $\tilde{v} = \lambda\tilde{u}$, we have that $$ \|x\| \lesssim \|\lambda u\|_{L^2} + \|f\|_{L^2}+ \|u'\|_{L^2} +\|w\|_{L^2} +\|\lambda \tilde{u}\|_{L^2} +\|\tilde{f}\|_{L^2}+ \|\tilde{u}'\|_{L^2}.$$ Thus the result will follow once we have established that each of the summands in the above equation are bounded by $C\sqrt{|\lambda|}\|y\|$ for $|s| \geq N$, where $C,N >0$ are constants independent of $y$. Consider $u$ given by (\ref{equ}). By Lemma \ref{1U}, it is enough to consider $|\lambda a(\lambda)|$ in order to estimate $\|\lambda u\|_{L^2}$ and $\|u'\|_{L^2}$. Now \begin{equation}\label{c11}\lambda a(\lambda) = \frac{\lambda}{\det M_\lambda}(c_{11}b_1 + c_{21}b_2+c_{31}b_3 + c_{41}b_4)\end{equation} where $b_i$ are the components of the vector $\mathbf{b}$ in (\ref{eqM}). We consider each of these terms. Note that by lemma \ref{1U}, the only terms in the components of $\mathbf{b}$ that are not automatically bounded by some constant multiple of $\|y\|$ are $W_\lambda(2)$ and $W_\lambda'(2)$. Looking at the first term in (\ref{c11}), $$\left|\frac{\lambda}{\det M_\lambda}c_{11}\right| = \sqrt{|\lambda|}\left|\frac{e^{\sqrt{\lambda}} T_+(\lambda) + e^{-\sqrt{\lambda}} T_-(\lambda)}{-e^{\sqrt{\lambda}} T_+(\lambda)^2 + e^{-\sqrt{\lambda}} T_-(\lambda)^2}\right| \lesssim \sqrt{|\lambda|}|T_{+}(\lambda)|^{-1} \lesssim \sqrt{|\lambda|},$$ since $\real{\sqrt{\lambda}}>0$ for $\lambda = \in i\mathbb{R}\setminus\{0\}$ as before, so that $e^{\sqrt{\lambda}}$ dominates $e^{-\sqrt{\lambda}}$. Thus, $$\left|\frac{\lambda}{\det M_\lambda}c_{11}b_1\right|\lesssim \sqrt{|\lambda|}\|y\|,$$ where the implicit constant is independent of $\lambda$ and $y$. Likewise, $$\left|\frac{\lambda}{\det M_\lambda}c_{21}\right|= \left|\frac{e^{\sqrt{\lambda}} T_+(\lambda) - e^{-\sqrt{\lambda}} T_-(\lambda)}{-e^{\sqrt{\lambda}} T_+(\lambda)^2 + e^{-\sqrt{\lambda}} T_-(\lambda)^2}\right| \lesssim 1,$$ so that $$\left|\frac{\lambda}{\det M_\lambda}c_{21}b_2\right|\lesssim \|y\|.$$ Noting that $|\cosh(\lambda)|,|\sinh(\lambda)| \leq 1$, a similar argument shows that the remaining terms in (\ref{c11}) that do not include $W_\lambda(2)$ and $W_\lambda(2)'$ are bounded by a constant times $\sqrt{|\lambda|}\|y\|$. Consider now \begin{align*}\left|\frac{\lambda}{\det M_\lambda}c_{31}W_\lambda(2)\right| & \leq \int_1^2 \left|\frac{\sinh(\sqrt{\lambda}(2-r))h(r)}{-e^{\sqrt{\lambda}} T_+(\lambda)^2+e^{-\sqrt{\lambda}} T_-(\lambda)^2}\right|dr \\ & \leq \frac{1}{2}\int_1^2 \left|\frac{e^{\sqrt{\lambda}(2-r)}-e^{-\sqrt{\lambda}(2-r)}}{-e^{\sqrt{\lambda}} T_+(\lambda)^2 +e^{-\sqrt{\lambda}} T_-(\lambda)^2}\right||h(r)| dr \\ & \lesssim |T_+(\lambda)|^{-2}\|h\|_{L^2} \lesssim \|h\|_{L^2} \end{align*} where the inequality in the final line is justified as before noting that $2-r \in[0,1]$. Similarly, $$\left|\frac{\lambda}{\det M_\lambda}c_{41}W_\lambda'(2)\right| \lesssim \sqrt{|\lambda|}\|h\|_{L^2}.$$ These inequalities combined with (\ref{c11}) imply that $$|\lambda a(\lambda)| \lesssim \sqrt{|\lambda|}\|y\| \ \ \ (|\lambda|\geq N)$$ for some constant $N>0$ independent of $y$ and in particular, $$\|\lambda u\|_{L^2}, \ \|u'\|_{L^2} \lesssim \|y\| \ \ \ (|\lambda| \geq N).$$ The same arguments show that this also holds for $\|\lambda \tilde{u}\|_{L^2}$ and $\|\tilde{u}'\|_{L^2}$. We must now estimate $w$ given by (\ref{eqw}), noting that $$b(\lambda) = \frac{1}{\det M_\lambda}(c_{12}b_1+c_{22}b_2+c_{32}b_2+c_{42}b_4)$$ and $$c(\lambda) = \frac{1}{\det M_\lambda}(c_{13}b_1+c_{23}b_2+c_{33}b_2+c_{43}b_4).$$ The trick to estimating $w$ is to group the terms together in a specific way. As before, the only terms in the components of $\mathbf{b}$ which are not bounded by $\|y\|$ are $W_\lambda(2)$ and $W_\lambda'(2)$. Hence it is enough to estimate the moduli of \begin{align*} w_1(\xi) & = \frac{1}{\det M_\lambda}[c_{12}\cosh(\sqrt{\lambda}(\xi-1))+c_{13}\sinh(\sqrt{\lambda}(\xi-1))],\\ w_2(\xi) & = \frac{1}{\det M_\lambda}[c_{22}\cosh(\sqrt{\lambda}(\xi-1))+c_{23}\sinh(\sqrt{\lambda}(\xi-1))],\\ w_3(\xi) & = \frac{1}{\det M_\lambda}[c_{32}\cosh(\sqrt{\lambda}(\xi-1))+c_{33}\sinh(\sqrt{\lambda}(\xi-1))],\\ \omega_4(\xi) & = \frac{1}{\det M_\lambda}[c_{42}\cosh(\sqrt{\lambda}(\xi-1))+c_{43}\sinh(\sqrt{\lambda}(\xi-1))],\end{align*} and \begin{equation*}\begin{split} w_5(\xi) & = \frac{1}{\det M_\lambda}\Big[\big(-c_{32}W_\lambda(2)+c_{42}W_\lambda'(2)\big)\cosh(\sqrt{\lambda}(\xi-1))\\ & \qquad \ \ + \big(-c_{33}W_\lambda(2)+c_{43}W_\lambda'(2)\big)\sinh(\sqrt{\lambda}(\xi-1)) - \det M_\lambda W_\lambda(\xi)\Big],\end{split} \end{equation*} where $\xi \in[1,2]$, since the sum of the $w_i$ is equal to the $w$ after removing the terms of $\mathbf{b}$ that do not include $W_\lambda(2)$ and $W_\lambda'(2)$. These removed terms can be shown to obey the desired estimates using the previous method. Plugging in the appropriate values gives \begin{align*}w_1 & = \frac{\lambda^2\cosh(\lambda)}{\det M_\lambda}\Big[e^{\sqrt{\lambda}} T_+(\lambda)\big(\cosh(\sqrt{\lambda}(\xi-1)) - \sinh(\sqrt{\lambda}(\xi-1))\big)\\ & \quad \quad \quad \quad \quad - e^{-\sqrt{\lambda}} T_-(\lambda) \big(\cosh(\sqrt{\lambda}(\xi-1)) +\sinh(\sqrt{\lambda}(\xi-1))\big)\Big] \\ & = \frac{\lambda^2\cosh(\lambda)}{\det M_\lambda} \left(e^{\sqrt{\lambda}} T_+(\lambda)e^{-\sqrt{\lambda}(\xi-1)} - e^{-\sqrt{\lambda}} T_-(\lambda)e^{\sqrt{\lambda}(\xi-1)}\right)\\ & = \cosh(\lambda)\frac{e^{\sqrt{\lambda}(2-\xi)}T_+(\lambda) - e^{-\sqrt{\lambda}(2-\xi)}T_-(\lambda)}{-e^{\sqrt{\lambda}} T_+(\lambda)^2+e^{-\sqrt{\lambda}} T_-(\lambda)^2}.\end{align*} Since $2-\xi \in [0,1]$, as in the case for $u$, $$|w_1(\xi)| \lesssim|T_\pm(\lambda)|^{-1} \lesssim 1,$$ where the sign of $\pm$ is determined by that of $s$ and the growth bound is independent of $\xi$. Likewise, $|w_2(\xi)| \lesssim 1$ with the bound independent of $\xi$. Next we have that \begin{align*} w_3 & = \cosh\lambda \frac{\sqrt\lambda \sinh(\lambda)\cosh(\sqrt\lambda(\xi-1))+\cosh(\lambda)\sinh(\sqrt\lambda(\xi-1))}{-e^{\sqrt{\lambda}} T_+(\lambda)^2 + e^{-\sqrt{\lambda}} T_-(\lambda)^2} \\ & = \cosh(\lambda)\frac{e^{\sqrt\lambda(\xi-1)}T_+(\lambda) - e^{-\sqrt\lambda(\xi-1)}T_-(\lambda)}{-e^{\sqrt{\lambda}} T_+(\lambda)^2 + e^{-\sqrt{\lambda}} T_-(\lambda)^2}. \end{align*} Since $\xi -1 \in[0,1]$, the previous argument again shows that $|w_3(\xi)| \lesssim 1$ with the bound independent of $\xi$. The same holds for $w_4$. Thus, it remains to estimate $w_5$ and after some simple manipulation, we can rewrite this as \begin{equation}\label{wt}w_5 = \frac{\lambda^2}{\det M_\lambda}\left[\frac{\cosh^2(\lambda)}{\sqrt{\lambda}}\Omega_1(\xi) +\sqrt{\lambda}\sinh^2(\lambda) \Omega_2(\xi) + \cosh(\lambda)\sinh(\lambda)\Omega_3(\xi)\right],\end{equation} where \begin{align*}\Omega_1(\xi) & = -\int_1^2 \sinh(\sqrt{\lambda}(2-r))\sinh(\sqrt{\lambda}(\xi-1))h(r)\ dr \\ & \ \ \ \ \ \ \ \ \ \ \ \ + \int_1^\xi \sinh(\sqrt{\lambda}(\xi-1))\sinh(\sqrt{\lambda})h(r)\ dr \\ & = \frac{1}{2}\Big[-\int_\xi^2 \cosh(\sqrt{\lambda}(1+\xi-r))h(r)\ dr \\ & \ \ \ \ \ \ \ \ \ \ \ \ + \int_1^2 \cosh(\sqrt{\lambda}(3-\xi-r))h(r)\ dr \\ & \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ - \int_1^\xi \cosh(\sqrt{\lambda}(\xi-r-1))h(r)\ dr \Big],\end{align*} with the second equality following from the use of identities such as \begin{equation*}\begin{split} 2\sinh(\sqrt{\lambda}(2-r))& \sinh(\sqrt{\lambda}(\xi-1))\\ & = \cosh(\sqrt{\lambda}(2-r)+\sqrt{\lambda}(\xi-1)) \\ & \quad\quad -\cosh(\sqrt{\lambda}(2-r) -\sqrt{\lambda}(\xi-1)),\end{split}\end{equation*} \begin{align*}\Omega_2(\xi) & = -\int_1^2 \cosh(\sqrt{\lambda}(2-r))\cosh(\sqrt{\lambda}(\xi-1))h(r)\ dr \\ & \ \ \ \ \ \ \ \ \ \ \ \ + \int_1^\xi \sinh(\sqrt{\lambda}(\xi-r))\sinh(\sqrt{\lambda})h(r)\ dr \\ & = -\frac{1}{2}\Big[\int_\xi^2 \cosh(\sqrt{\lambda}(1+\xi-r))h(r)\ dr \\ & \ \ \ \ \ \ \ \ \ \ \ \ + \int_1^2 \cosh(\sqrt{\lambda}(3-\xi-r))h(r)\ dr \\ & \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ + \int_1^\xi \cosh(\sqrt{\lambda}(\xi-r-1))h(r)\ dr \Big],\end{align*} and \begin{align*}\Omega_3(\xi) & = 2\int_1^\xi \sinh(\sqrt{\lambda}(\xi-r))\cosh(\sqrt{\lambda})h(r)\ dr \\ & \ \ \ \ \ \ -\int_1^2 \sinh(\sqrt{\lambda}(2-r))\cosh(\sqrt{\lambda}(\xi-1))h(r)\ dr \\ & \ \ \ \ \ \ \ \ \ \ \ \ -\int_1^2 \cosh(\sqrt{\lambda}(2-r))\sinh(\sqrt{\lambda}(\xi-1))h(r)\ dr \\ & = -\int_1^\xi \sinh(\sqrt{\lambda}(1-\xi+r))h(r)\ dr -\int_\xi^2 \sinh(\sqrt{\lambda}(1+\xi-r))h(r)\ dr.\end{align*} Note that for $\xi \in [1,2]$, $|1+\xi-r|\leq 1$ whenever $r\in[\xi,2]$, and $|3-\xi-r|\leq 1$ whenever $r\in[1,2]$, and $|\xi-r-1|\leq 1$ whenever $r\in[1,\xi]$. Thus by pulling the factor of $\big(-e^{\sqrt{\lambda}} T_+(\lambda)^2 + e^{-\sqrt{\lambda}} T_-(\lambda)^2\big)^{-1}$ into the integrand of $\Omega_1$, we see that $$\left|\frac{\lambda^2}{\det M_\lambda} \frac{\cosh^2(\lambda)}{\sqrt{\lambda}}\Omega_1(\xi)\right| \lesssim \frac{1}{\sqrt{\lambda}}\|h\|_{L^2}.$$ Arguing similarly for $\Omega_2$ and $\Omega_3$, we get from (\ref{wt}) that $$|w_5| \lesssim \left(\frac{1}{\sqrt{|\lambda|}} + \sqrt{|\lambda|} + 1\right)\|h\|_{L^2} \lesssim \sqrt{|\lambda|}\|h\|_{L^2},$$ with the implicit constant independent of $\xi$. It follows that $$\|w\|_{L^2} \lesssim \|y\|, \ \ \ |\lambda|\geq N,$$ for some constant $N>0$ independent of $y$ and, in particular, $$\|x\| \lesssim |\lambda|^{1/2}\|y\|,$$ with the implicit constant independent of the specific $y$ and $x$. \end{proof} \section{Optimal Energy Decay for Classical Solutions} For the reader's convenience, we state below the version of the Borichev-Tomilov theorem used in \cite{BPS1}. \begin{theorem}[{\cite[Theorem 4.1]{BPS1}}]\label{boto} Let $(T(t))_{t\geq0}$ be a bounded $C_0$-semigroup on a Hilbert space $X$ with generator $A$ such that $\sigma(A)\cap i\mathbb{R} = \emptyset$. Then for any $\alpha >0$, the following are equivalent: \begin{enumerate}[(i)] \item $\|R(is,A)\| = O(|s|^\alpha)$ as $|s|\to\infty$; \item $\|T(t)A^{-1}\| = O(t^{-1/\alpha})$ as $t\to \infty$; \item $\|T(t)x\| = o(t^{-1/\alpha})$ as $t\to \infty$ for all $x\in D(A)$. \end{enumerate} \end{theorem} Using the abstract but powerful tool above, we can convert the resolvent estimate in Theorem \ref{mainthm} into a rate of energy decay of classical solutions of (\ref{WHWeq1}), deriving the main result of the article. The rate itself will follow easily from Theorem \ref{boto} as we shall soon see, but optimality will require a little more work. \begin{theorem}\label{opt} If $x \in D(A)$, then $E_x(t) = o(t^{-4})$ as $t\to \infty$. Moreover, this rate is optimal in the sense that, given any positive function $r$ satisfying $r(t) = o(t^{-4})$ as $t \to \infty$, there exists $x \in D(A)$ such that $E_x(t) \neq o(r(t))$ as $t \to \infty$. \end{theorem} Before we begin the proof, we state the following summary proposition needed to show optimality. What is stated below is more or less a collection of results from \cite{BPS1}. \begin{proposition}\label{halfprop} Let $B$ be the generator of the $C_0$-semigroup $S(t)$ on the Hilbert space $$Z_* = \{(u,v,w) \in H^1(0,1)\times L^2(0,1) \times L^2(1,3/2) : u(0)=0\}$$ that solves the following well-posed problem: \begin{equation}\label{WHWeq2} \begin{cases}\begin{aligned} & u_{tt}(\xi,t) = u_{\xi\xi}(\xi,t), & \xi \in (0,1), \ & t>0,\\ & w_t(\xi,t) = w_{\xi\xi}(\xi,t), & \xi \in (1,3/2), \ & t>0, \\ & u(0,t) = w(3/2,t) = 0, & & t>0, \\ & u_t(1,t) = w(1,t),\ \ \ u_\xi(1,t) = w_\xi(1,t), & & t>0, \\ & u(\xi,0) = u(\xi), \ \ \ u_t(\xi,0) = v(\xi), & \xi \in (0,1), \ & \\ & w(\xi,0) = w(\xi), & \xi \in (1,3/2). \ & \end{aligned}\end{cases}\end{equation} Then \begin{equation}\label{limsup}\limsup_{|s|\to\infty}|s|^{-1/2}\|R(is,B)\| >0.\end{equation} In particular, for any positive function $r$ satisfying $r=o(t^{-4})$ as $t\to \infty$, there exists \begin{align*}x_* \in D(B) = \{(u,v,w) \in H^2(0,1)&\times H^1(0,1)\times H^2(1,3/2) \\ & : u(0)=v(0)=w(3/2) = 0, \\ & \ \ \ \ \ \ v(1)=w(1), \ u'(1)=w'(1)\}\end{align*} such that $$E_{x_*}(t) = \int_0^1 |u'(\xi,t)|^2 + |v(\xi,t)|^2 \ d\xi + \int_1^{3/2} |w(\xi,t)|^2 \ d\xi \neq o(r(t))$$ as $t\to \infty$. \end{proposition} \begin{proof} After a rescaling of the heat component by a factor of $2$, this is the same problem as what is studied in \cite[Section 5]{BPS1}, namely the coupled wave-heat equation that leads to the optimal resolvent bound in \cite[Theorem 3.1]{BPS1}, but with Dirichlet wave condition. The problem is well-posed, therefore, and the same resolvent estimates hold up to a constant and they remain optimal in the sense of (\ref{limsup}). This is again proved in the same exact way as \cite[Theorem 3.4]{BPS1} by using the argument found there based on Rouch\'e's theorem. Note that in this case, however, $\sigma(B)\cap i\mathbb{R} = \emptyset$. The final part of the proposition follows from (\ref{limsup}) and is proved along the lines of \cite[Remark 4(a)]{BPS1}. We flesh that remark out here. Assume for a contradiction that there exists a positive function $r$ satisfying $r(t)=o(t^{-2})$ as $t\to \infty$ such that for all $x \in D(B)$ $\|S(t)x\| = o(r(t))$. Without loss of generality, $r$ is non-increasing since we can replace $r$ with $r_1(t) = \sup_{t\leq \tau}r(\tau)$ which also satisfies $r_1(t) = o(t^{-2})$ and $\|S(t)x\|=o(r_1(t))$ for all $x\in D(B)$. Then for all $y \in X$, there exists a constant $C_y$ such that $$r(t)^{-1}\|S(t)R(1,B)y\| \leq C_y, \ \ \ t\geq 0.$$ Hence by the uniform boundedness principle, there exists $C>0$ independent of $y$ such that $$\|S(t)R(1,B)\| \leq Cr(t).$$ In particular, $m(t) \leq Cr(t)$ where $m(t)=\sup_{t\leq \tau} \|S(\tau)R(1,B)\|.$ By \cite[Proposition 1.3]{BaDu}, $$\|R(is,B)\| \lesssim 1+ m_{*}^{-1}\left(\frac{1}{2(|s|+1)}\right), \ \ \ s \in \mathbb{R},$$ where $m_*^{-1}$ is a right inverse of the function $m$, mapping $(0,m(0)]$ onto $[0,\infty)$. This contradicts (\ref{limsup}) if $|s|^{-1/2}m_{*}^{-1}\left(\frac{1}{2(|s|+1)}\right)\to 0$ as $|s| \to \infty$, which we now show. Notice first that because $t^2Cr(t) \to 0$ as $t\to \infty$ and $(Cr)_*^{-1}(|s|)\to \infty$ as $|s|\to 0$, we have that $(Cr)_*^{-1}(|s|)^2|s| \to 0$ as $|s| \to 0$, where $(Cr)_*^{-1}$ is a right inverse of the function $Cr$, mapping $(0,Cr(0)]$ onto $[0,\infty)$. Hence $$(Cr)_*^{-1}\left(\frac{1}{2(|s|+1)}\right)^2\frac{1}{2(|s|+1)}\to 0$$ as $|s| \to \infty$. But since $m\leq Cr$ and both functions are non-increasing, it follows that $m_*^{-1} \leq (Cr)_*^{-1}$ on the interval $(0,m(0)]$ and we are done. \end{proof} \begin{remark} In the above proof, we can alternatively prove the simpler optimality statement that $$\|S(t)R(1,B)\| \geq ct^{-2}, \ \ \ t\ge1,$$ for some constant $c>0$ by combining \cite[Theorem 4.4.14]{ABHN} with the fact that the specific $\lambda_n^\pm$ in \cite[Theorem 3.4]{BPS1} are evenly spaced. \end{remark} We finally prove the decay rate in Theorem \ref{opt} using Theorem \ref{boto} as promised and its optimality by showing that the system (\ref{WHWeq2}) is effectively contained within (\ref{WHWeq1}). \begin{proof}[Proof of Theorem \ref{opt}] By Theorem \ref{boto}, we have that $$E_x(t) = \frac{1}{2}\|T(t)x\|^2 = o(t^{-4})$$ as $t \to \infty$ for any $x \in D(A)$ since Theorem \ref{mainthm} gives us the rate $\|R(is,A)\| = O(|s|^{1/2})$ as $s \to \infty$. To show optimality, assume that there exists a positive function $r$ satisfying $r=o(t^{-4})$ as $t\to \infty$. Proposition \ref{halfprop} produces an $x_* \in D(B) \subset Z_*$ for which $E_{x_*}(t) \neq o(r(t))$ as $t \to \infty$. Define $\tilde{x}:[0,3]\to\mathbb{C}$ by $$\tilde{x}(\xi) = \begin{cases} x_*(\xi), & \xi \in [0,3/2], \\ -x_*(3-\xi), & \xi \in (3/2,3], \end{cases}$$ and note that $\tilde{x}$ satisfies all the conditions necessary to be in $D(A)$, including the $H^2(1,2)$ condition since on a symmetric interval around the only potentially problematic point $\xi=3/2$, the function $\tilde{x}$ is the negative reflection of an $H^2$ function around a point at which it is $0$. Morever, the classical solution to (\ref{WHWeq1}) of initial data $\tilde{x}$ is given by $$\tilde{x}(\xi,t) = \begin{cases} x_*(\xi,t), & \xi \in [0,3/2], \ t>0, \\ -x_*(3-\xi,t), & \xi \in (3/2,3], \ t>0, \end{cases}$$ where $x_*(\xi,t)$ is the classical solution to (\ref{WHWeq2}) for intial data $x_*$. It follows that $$E_{\tilde{x}}(t) = 2E_{x_*}(t) \neq o(r(t))$$ as $t \to \infty$. \end{proof} \section{Possible Future Directions} In this final section, we pose and comment on a few questions about possible future directions arising out of systems similar to that described by (\ref{WHWeq1}). The last of these questions could potentially be very interesting and not easily tractable. Note, however, that the question likely to be asked first -- what happens when the Dirichlet conditions are replaced by Neumann conditions -- is easily answered. In this case, the semigroup is actually unbounded. The function $x(\xi,t) = (at,a,at)$ for any constant $a\neq 0$ solves the variant of (\ref{WHWeq1}) where the fourth line is changed to $u_t(0,t)=\tilde{u}_t(3,t) = 0$ for the initial condition $(0,a,0)$, which yields an unbounded orbit of the semigroup in this case. That said, an alternative formulation involving a different state space can be chosen for the Neumann problem, one that is more physically intuitive and for which the same method as for the Dirichlet case can be applied to obtain the same rate of decay. With that out of the way, we ask the following natural two part question. \begin{opq} \begin{enumerate}[(i)] \item Does the rate of energy decay remain the same up to a constant when extra wave and heat parts are added, for example, in a wave-heat-wave-heat or a wave-heat-wave-heat-wave system? \item If the energy remains optimally bounded by $Ct^{-4}$ for $C$ dependent on the particular system, can we find an explicit $N$-formula for the multiplicative constant $C_N$ bounding the energy of the system composed of $N$ wave-heat pairs all coupled together? \end{enumerate} \end{opq} Our first reaction at the thought of answering this question is one of horror, as the methods used in this article involved inverting a $4\times4$ matrix, and a system composed of $N$ wave-heat pairs would require the inversion of a $(4N-2)\times(4N-2)$ matrix. Though we believe that the answer to the first part of the above question is affirmative, the second part would require some clever matrix tricks to avoid total carnage. It is notable, however, that the matrices would have $0$ entries everywhere, except off of a diagonal of width at most four. So perhaps it is doable. The idea of, perhaps inductively, obtaining a formula for $C_N$ as above leads to the question of homogenisation, that is, the computation of a limit equation. For $N \in \mathbb{N}$ and a given smooth function $f$, consider the following system of mixed hyperbolic and elliptic type that was studied in \cite{Wa1}: \begin{equation}\label{HEHomo} \begin{cases}\begin{aligned} & \partial^2_{t}u_N(\xi,t) - \partial^2_{\xi}u_N(\xi,t) = \partial_t f(\xi,t), & \xi \in \bigcup_{j \in \{1,\dots,N\}}\left(\frac{j-1}{N},\frac{2j-1}{2N}\right), \ & t\in \mathbb{R},\\ & u_N(\xi,t) - \partial^2_{\xi}u_N(\xi,t) = f(\xi,t), & \xi \in \bigcup_{j \in \{1,\dots,N\}}\left(\frac{2j-1}{2N},\frac{j}{N}\right), \ & t\in \mathbb{R},\\ & \partial_\xi u_N(0,t) = \partial_\xi u_N(1,t), & \ & t \in \mathbb{R}, \end{aligned}\end{cases}\end{equation} subject to zero initial conditions and the requirement that the $u_N$ and their derivatives are continuous. We use the notation $\partial$ for derivatives as in \cite{Wa1} to avoid a mess involving the subscript $N$. Note that requiring conditions of continuity at the junction points results in the coupling considered in \cite{ZZ1} rather than that of \cite{BPS1} and (\ref{WHWeq1}). Waurick showed in \cite{Wa1} that as $N\to \infty$, the sequence of solutions $(u_N)_{N\in \mathbb{N}}$ converges weakly in $L_{loc}^2([0,1]\times \mathbb{R})$ to $u$, the solution to the limit equation $$\frac{1}{2}\partial_t^2 u(\xi,t)+\partial_t u(\xi,t)+\frac{1}{2}u(\xi,t) - 2\partial_\xi^2 u(\xi,t) = f(\xi,t) + \partial_t f(\xi,t), \ \ \ t \in \mathbb{R},$$ subject to zero initial conditions and Neumann boundary on both ends. Furthermore, he showed that this limit admitted exponentially stable solutions in the sense of \cite{Tros1}. However, when the elliptic part, $u_N(\xi,t) - \partial_\xi^2 u_N(\xi,t) = f(\xi,t)$, is replaced with the corresponding parabolic part, $\partial_t u_N(\xi,t) - \partial_\xi^2 u_N(\xi,t) = f(\xi,t)$, the limit equation becomes $$\frac{1}{2}\partial_t^2 u(\xi,t)+\partial_t u(\xi,t)- 2\partial_\xi^2 u(\xi,t) = f(\xi,t) + \partial_t f(\xi,t), \ \ \ t \in \mathbb{R},$$ subject again to zero initial conditions and Neumann boundary. Crucially, this limit equation is not exponentially stable, raising the following question. \begin{opq} For the homogenised limit equation with mixed hyperbolic and parabolic parts as above, can the limit solution be posed and solved by a non-uniformly stable semigroup? \end{opq} In \cite{FrWa}, resolvent estimates of some kind are calculated in a way that depends on $N$ via the Gelfand transform, before a numerical analysis is conducted. How this might somehow be converted to a resolvent estimate for the limit problem itself remains an interesting unanswered question. \section*{Appendix} \subsection*{Entries for the Cofactor Matrix $C$ in Section \ref{resolventestimates}} \addcontentsline{toc}{chapter}{Appendix} \begin{align*} c_{11} & = -\lambda^{3/2}[e^{\sqrt{\lambda}}T_+(\lambda) + e^{-\sqrt{\lambda}}T_-(\lambda)], \\ c_{12} & = \lambda^2\cosh(\lambda)(e^{\sqrt{\lambda}}T_+ - e^{-\sqrt{\lambda}}T_-), \\ c_{13} & = -\lambda^2\cosh(\lambda)[e^{\sqrt{\lambda}} T_+(\lambda) +e^{-\sqrt{\lambda}} T_-(\lambda)], \\ c_{14} & = \lambda^{3/2} \cosh(\lambda), \\ c_{21} & = -\lambda[e^{\sqrt{\lambda}} T_+(\lambda) - e^{-\sqrt{\lambda}} T_-(\lambda)],\\ c_{22} & = -\lambda^2\sinh(\lambda)[e^{\sqrt{\lambda}} T_+(\lambda) -e^{-\sqrt{\lambda}} T_-(\lambda)],\\ c_{23} & = \lambda^2\sinh(\lambda)[e^{\sqrt{\lambda}} T_+(\lambda) + e^{-\sqrt{\lambda}} T_-(\lambda)],\\ c_{24} & = -\lambda^{3/2}\sinh(\lambda), \\ c_{31} & = \lambda^{3/2}\cosh(\lambda), \\ c_{32} & = \lambda^{5/2}\sinh(\lambda)\cosh(\lambda) \\ c_{33} & = \lambda^2\cosh^2(\lambda), \\ c_{34} & = -\lambda^{3/2}[e^{\sqrt{\lambda}} T_+(\lambda) e^{-\sqrt{\lambda}} T_-(\lambda)],\\ c_{41} & = -\lambda^{3/2}\sinh(\lambda), \\ c_{42} & = -\lambda^{5/2} \sinh^2(\lambda), \\ c_{43} & = -\lambda^2\cosh(\lambda)\sinh(\lambda), \\ c_{44} & = -\lambda[e^{\sqrt{\lambda}} T_+(\lambda) - e^{-\sqrt{\lambda}} T_-(\lambda)]. \end{align*} \providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace} \providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR } \providecommand{\MRhref}[2]{% \href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2} } \providecommand{\href}[2]{#2}

https://arxiv.org/abs/1702.07365

The primes are not metric Poissonian

It has been known since Vinogradov that, for irrational $\alpha$, the sequence of fractional parts $\{\alpha p\}$ is equidistributed in $\mathbb{R}/\mathbb{Z}$ as $p$ ranges over primes. There is a natural second-order equidistribution property, a pair correlation of such fractional parts, which has recently received renewed interest, in particular regarding its relation to additive combinatorics. In this paper we show that the primes do not enjoy this stronger equidistribution property.

\section{Introduction} Let $\mathcal{A}\subset \mathbb{N}$ be an infinite sequence of natural numbers, and let $A_N$ denote the first $N$ elements of $\mathcal{A}$. For $\alpha \in [0,1]$, we consider the sequence $\alpha \mathcal{A}$, taken modulo 1. Recall that the sequence $\alpha \mathcal{A}$ is \emph{equidistributed} in $\mathbb{R}/\mathbb{Z}$ if for every interval $I\subset \mathbb{R}/\mathbb{Z}$ one has \begin{equation} \label{equidistribution definition} \lim\limits_{N\rightarrow \infty}\frac{1}{N}\sum\limits_{x\in A_N}\mathbbm{1}_{I}(\alpha x) = \vert I\vert. \end{equation} \noindent For many arithmetic sequences $\mathcal{A}$ of interest, the sequence $\alpha \mathcal{A}$ is equidistributed in $\mathbb{R}/\mathbb{Z}$ for all irrational $\alpha$. This is true for $\mathcal{A}=\mathbb{N}$ itself, or more generally the set of $k^{th}$ powers for any $k\in \mathbb{N}$, and, most pertinently for us, the set of primes. In this paper we will consider a strictly stronger notion of equidistribution. With notation as above, we define the pair correlation function \begin{equation} \label{key definition of pair correlation function} F(\mathcal{A},\alpha,s,N) :=\frac{1}{N} \sum\limits_{\substack{x_i,x_j\in A_N\\ x_i\neq x_j}} \mathbbm{1}_{[-s/N,s/N]}( \alpha(x_i - x_j)), \end{equation} \noindent where both the interval $[-s/N,s/N]$ and the sequence $\alpha \mathcal{A}$ are considered modulo $1$.\\ Informally, $F(\mathcal{A},\alpha,s,N)$ counts the number of pairs $(\alpha x_i,\alpha x_j)$ such that the distance $\alpha x_i - \alpha x_j \text{ mod } 1$ is approximately $s$ times the average gap length of the sequence $\alpha A_N\text{ mod } 1$. Analysing the behaviour of $F(\mathcal{A},\alpha,s,N)$ for a specific $\alpha$ can require delicate Diophantine information about $\alpha$ (see \cite{H-B10}, \cite{RSZ01}), but one may instead settle for results which hold for almost all $\alpha$. In the setting of (\ref{equidistribution definition}), \emph{any} $\mathcal{A}$ satisfies the equidistribution property for almost all $\alpha$ (the sharpest results in this direction are due to Baker \cite{Ba81}). However, in the setting of pair correlations, the situation is more subtle. We say that the sequence $\mathcal{A}$ is `metric Poissonian' if for almost all $\alpha\in [0,1]$, and for all fixed $s>0$, we have \begin{equation} \label{def of met poiss} F(\mathcal{A},\alpha,s,N) = 2s(1+o_{\mathcal{A},\alpha,s}(1)) \end{equation} \noindent as $N\rightarrow \infty$. Notice that if we had picked $N$ i.i.d. random variables $(X_n)_{n\in [N]}$ uniformly distributed on $\mathbb{R}/\mathbb{Z}$, instead of the sequence $\alpha A_N \text{ mod }1$, then as $N$ tends to infinity the equivalent pair correlation function would tend to $2s$ with high probability. Therefore (\ref{def of met poiss}) may be viewed as some strong indication that $\alpha \mathcal{A}\text{ mod }1$ exhibits the behaviour of a random sequence. The connection to the equidistribution property (\ref{equidistribution definition}) was recently made rigorous: indeed, three simultaneous papers \cite{ALP16,GL16, St17} recently showed that if (\ref{def of met poiss}) holds, for some fixed $\alpha$ and for all $s$, then for the same $\alpha$ one has that $\alpha \mathcal{A}$ is equidistributed in $\mathbb{R}/\mathbb{Z}$. \\ One might expect that, for the classical sequences $\mathcal{A}$ where $\alpha \mathcal{A}$ is equidistributed for irrational $\alpha$, one could prove that these sequences $\mathcal{A}$ are metric Poissonian. Indeed, for $k\geqslant 2$ and $\mathcal{A}$ the set of $k^{th}$ powers this was shown by Rudnick and Sarnak \cite{RS98}. However, the sequence $\mathcal{A}=\mathbb{N}$ is \emph{not} metric Poissonian. This follows from consideration of the continued fraction expansion of $\alpha$, but is in fact a special case of a more general phenomenon, connected to the large \emph{additive energy} of this particular set $\mathcal{A}$. For a finite set $B\subset \mathbb{N}$ we define the additive energy $E(B)$ to be the number of quadruples $(b_1,b_2,b_3,b_4)\in B^4$ such that $b_1+b_2 = b_3+b_4$. If $\vert B\vert =N$, then we have the trivial bounds $N^2\ll E(B)\leqslant N^3$. For $x\in\mathbb{R}$, let us write $\Vert x\Vert$ for $\operatorname{min}_{y\in \mathbb{Z}}\vert x-y\vert$. Then, for an additive quadruple $(b_1,b_2,b_3,b_4)$ satisfying $b_1+b_2 = b_3+b_4$, obviously if $\Vert \alpha(b_1-b_3)\Vert \leqslant \frac{s}{N}$ then $\Vert \alpha(b_2-b_4)\Vert\leqslant \frac{s}{N}$. This is extremely different behaviour than that which would be seen if $\alpha b_1$, $\alpha b_2$, $\alpha b_3$, $\alpha b_4$ were genuinely i.i.d. uniform random variables on $\mathbb{R}/\mathbb{Z}$, and indeed we have the following result of Bourgain, which shows that all sets of nearly maximal energy fail to have the metric Poissonian property. \begin{Theorem}{\cite[Appendix]{ALLB16}} \label{large energy theorem} Suppose $$\operatorname{limsup}\limits_{N\rightarrow \infty} \frac{E(A_N)}{N^3}>0.$$ Then $\mathcal{A}$ is not metric Poissonian. \end{Theorem} \noindent It is clear that the sequence $\mathcal{A}=\mathbb{N}$ satisfies the hypotheses of this theorem, and therefore this sequence is not metric Poissonian. Remarkably, a near-converse to this theorem has also been proved to be true. \begin{Theorem} \label{small energy theorem} Let $\delta>0$ be fixed, and suppose that $E(A_N) \ll_\delta N^{3-\delta}$ for this fixed $\delta$ and for every $N$. Then $\mathcal{A}$ is metric Poissonian. \end{Theorem} \noindent This theorem first appears as stated\footnote{The same result may be deduced from Theorem 3.2 of Harman's earlier book \cite{Ha98}, combined with the relevant modification of the variance estimate from page 69 of the same volume.} in the recent work of Aistleitner, Larcher and Lewko \cite{ALLB16}. It immediately implies the theorem of Rudnick-Sarnak on $k^{th}$ powers, and also earlier work on lacunary sequences \cite{RZ99}. \\ It is natural to wonder whether there is a tight energy threshold for this problem. Although the truth seems unlikely to be so clean, it is certainly interesting to consider the behaviour of specific sets $\mathcal{A}$ which satisfy $N^{3-\varepsilon} \ll_\varepsilon E(A_N)\ll o(N^3)$ for all $\varepsilon>0$. In this paper, we prove the following result (answering a question posed by Nair\footnote{at the ELAZ 2016 conference in Strobl.}). \begin{Theorem}[Main Theorem] \label{Main Theorem} The primes are not metric Poissonian. \end{Theorem} When $\mathcal{A}$ is the set of primes one has $E(A_N)\asymp N^3(\log N)^{-1}$, so certainly the primes are not included in the range of applicability of either Theorem \ref{large energy theorem} or Theorem \ref{small energy theorem}. In \cite{ALLB16}, Bourgain constructs a sequence $\mathcal{A}$ which is not metric Poissonian but nonetheless has $E(A_N) = o(N^3)$, thereby showing that the converse to Theorem \ref{large energy theorem} is false. A quantitative analysis of his argument shows that $E(A_N)\ll_{\varepsilon} N^3 (\log\log N)^{-\frac{1}{4} + \varepsilon}$ is achievable, for any $\varepsilon>0$. So, as an immediate corollary to Theorem \ref{Main Theorem}, we have an improved bound for the smallest energy $E(A_N)$ of the initial segments of a set $\mathcal{A}$ which is not metric Poissonian. \\\\\\ \noindent\textbf{Acknowledgements} \noindent The author would like to thank Prof. R. Nair for making him aware of the central question of this paper, Prof. C. Aistleitner for comments on an earlier version, and Prof. B. J. Green for his doctoral supervision. A helpful conversation was also had with Sam Chow. The work was completed while the author was a Program Associate at the Mathematical Sciences Research Institute in Berkeley, which provided excellent working conditions. The author is supported by EPSRC grant no. EP/M50659X/1. \section{Proof of Theorem \ref{Main Theorem}} The plan of the proof is as follows. For each fixed $\alpha$, we will try to find infinitely many $n$ such that $\Vert \alpha n\Vert $ is extremely small. Using such an $n$ we will be able to construct a scale $N$ and a small constant $s$ such that $\Vert \alpha mn\Vert\leqslant s/N$ for some initial segment of integers $m$. By a variant of a well-known result concerning the exceptional set for the Goldbach problem, we may show that many such $mn$ are represented many times as $p_i-p_j$ for two primes $p_i,p_j\leqslant p_N$, the $N^{th}$ prime. Combining all these observations will enable us to conclude, provided $s$ is small enough, that $F(\mathcal{P},\alpha,s,N)\geqslant c$ for some constant $c>2s$. Since this holds for infinitely many $N$, we cannot have $ F(\mathcal{P},\alpha,s,N) = 2s(1+o_{\alpha,s}(1))$ for all almost all $\alpha$ and for all $s>0$. In fact, we will show that, for almost all $\alpha$, this asymptotic \emph{fails} to hold. \\ We now begin to consider the details of this argument. We will use the following result of Harman on Diophantine approximation (\cite[Theorem 4.2]{Ha98}, \cite{Ha88}). \begin{Theorem} \label{Harman's theorem} Let $\psi(n)$ be a non-increasing function with $0<\psi(n)\leqslant \frac{1}{2}$. Suppose that $$\sum\limits_n\psi(n)=\infty.$$ Let $\mathcal{B}$ be an infinite set of integers, and let $S(\mathcal{B},\alpha,N)$ denote the number of $n\leqslant N$, $n\in \mathcal{B}$, such that $\Vert n\alpha\Vert < \psi(n)$. Then for almost all $\alpha$ we have \begin{equation} S(\mathcal{B},\alpha,N) = 2\Psi(N,\mathcal{B}) + O_\varepsilon(\Psi(N)^{\frac{1}{2}}(\log \Psi(N) )^{2+\varepsilon}) \end{equation} \noindent for all $\varepsilon>0$, with implied constant uniform in $\alpha$, where $$\Psi(N) = \sum\limits_{n\leqslant N} \psi(n)$$ and $$\Psi(N,\mathcal{B}) = \sum\limits_{\substack{n\leqslant N\\n\in \mathcal{B}}} \psi(n).$$ \end{Theorem} \noindent This theorem may be thought of as a flexible version of Khintchines's theorem on Diophantine approximation, in which one can further pass to approximations coming from a set $\mathcal{B}$, provided $\mathcal{B}$ is relatively dense. The quality of the error term in this theorem is much better than we need in our application, although it is important that there is no dependence on $N$ except through $\Psi(N)$. Earlier results of this type include a $\log N$ factor in the error, which would not have been adequate. \\ The other technical tool will be the standard bound on the size of the exceptional set in a Goldbach-like problem. \begin{Theorem} \label{binary Goldbach} For a large quantity $X$, and natural number $n\leqslant X$, define $$r(n): = \sum\limits_{\substack{p_i,p_j\leqslant X\\ p_i-p_j = n}} \log p_i \log p_j.$$ Then for any $B>0$, and for all but $O_B(\frac{X}{\log ^B X})$ exceptional values of $n\leqslant X$, we have the approximation \begin{equation} \label{asymptotic formula for binary Goldbach} r(n)= \mathfrak{S}(n)J(n) + O_B(\frac{X}{\log ^B X}), \end{equation} \noindent where \begin{equation*} \mathfrak{S}(n) := \begin{cases} 2\prod\limits_{p\geqslant 3}\left(1-\frac{1}{(p-1)^2}\right)\prod\limits_{\substack{p\vert n\\p\geqslant 3}}\frac{p-1}{p-2} & n\text{ even, }\\ 0 & n\text{ odd} \end{cases} \end{equation*} \noindent is the singular series, and $$J(n) = \int\limits_{-\infty}^{\infty}\left\vert\int\limits_{0}^{X} e(\beta u) \, du \right\vert^2 e(-\beta n) \, d\beta$$ is the singular integral. \end{Theorem} \begin{proof} This result follows by trivial modifications of the usual argument for the binary Goldbach problem, originally due (independently) to van der Corput, Estermann, and \u{C}udakov. The clearest reference is \cite[Chapter 3.2]{Va97}, or, for a more modern approach, one may consider the proof of Theorem 19.1 in \cite[Chapter 19]{IwKo04}. \end{proof} We combine these two key ingredients in the following proposition. \begin{Proposition} \label{key proposition} There exists a small absolute $c>0$, such that for almost all $\alpha\in [0,1]$, and for all fixed $s>0$, there exist infinitely many $n$ satisfying: \begin{enumerate}[(i)] \item $\Vert \alpha n\Vert < \frac{s}{ n\log n}$\\ \item At least $c\log n$ of the numbers $n$, $2n$, $\cdots$, $\lfloor \frac{1}{10}\log n\rfloor n$ are expressible in at least $c\frac{n}{\log n}$ ways as the difference $p_1 - p_2$ of two primes $p_1,p_2 \leqslant \frac{1}{2} n\log n$. \end{enumerate} \end{Proposition} \begin{proof}[Proof of Proposition \ref{key proposition}] Let $c>0$ be a quantity to be specified later. With this $c$, let $\mathcal{B}$ be the set of natural numbers $n$ which satisfy (ii), and let $\psi(n) = \operatorname{min}(\frac{1}{2},\frac{s}{ n \log n})$. It is to this $\mathcal{B}$ and this $\psi$ that we will apply Theorem \ref{Harman's theorem}.\\ We claim that $\mathcal{B}$ is relatively dense. Indeed, let $K$ be a large integer, and let $n$ and $m$ be natural numbers restricted to the ranges $K\leqslant n<2K$ and $1\leqslant m\leqslant \lfloor \frac{1}{10}\log 2K\rfloor$. For notational convenience we let $X$ denote the quantity $\frac{1}{2} K\log K$, and we consider Theorem \ref{binary Goldbach} with this $X$. We say that the pair $(n,m)$ is \emph{exceptional} if $nm$ lies in the exceptional set from Theorem \ref{binary Goldbach} for which the asymptotic formula (\ref{asymptotic formula for binary Goldbach}) fails to hold. [Note that $nm\leqslant X$, so Theorem \ref{binary Goldbach} applies in this setting.] The map $(n,m)\mapsto nm$ is at most $\frac{1}{10}\log 2K$-to-$1$, due to the restricted range of $m$. Since the exceptional set from Theorem \ref{binary Goldbach} has size at most $O_B(\frac{K}{\log ^B K})$, for all $B>0$, we conclude that there are at most $O_B(\frac{K}{\log ^B K})$ exceptional pairs $(n,m)$, for all $B>0$. So certainly there are at least $(1-O_B(\log ^{-B} K))K$ values of $n\in [K,2K)$ such that the asymptotic formula for $r(nm)$ holds for all $m\leqslant \frac{1}{10} \log 2K$. Let $D_K$ denote this set of $n$. \\ $D_K$ is certainly very dense in $[K,2K)$, and we claim further that $D_K\subset \mathcal{B}$, provided we choose $c$ small enough. Combining the different scales $K$ will allow us to show that $\mathcal{B}$ is suitably dense. Indeed, let us analyse the asymptotic formula for $r(nm)$. When $nm$ is even, the singular series $\mathfrak{S}(nm)$ is always $\Omega(1)$. By Fourier inversion, the singular integral is exactly $(\mathbbm{1}_{[0,X]}\ast \mathbbm{1}_{[-X,0]})(nm)$, which is $\Omega(X)$ provided that $\vert nm\vert \leqslant \frac{1}{5} X$. But, by the choice of ranges for $n$ and $m$, this inequality is always satisfied. So, for all $n\in D_K$ and for all $m\leqslant \frac{1}{10}\log n$, such that $nm$ is even, we have $r(nm)\gg X$. Removing the log weights on the primes, and recalling the definition of $X$, in particular we notice that there is some small absolute constant $c_1$ such the following holds: if $n\in D_K$ and if $nm$ is even, there are at least $c_1K/\log K$ pairs of primes $(p_i,p_j)$ with $p_i,p_j\leqslant \frac{1}{2}K\log K$ such that $p_i - p_j$ = $nm$. By the definition of $\mathcal{B}$, provided $c$ is chosen smaller than $\operatorname{min}(c_1,\frac{1}{100})$, we have that $D_K\subset \mathcal{B}$. \\ We may now prove that $\Psi(\mathcal{B},N)\gg_s \Psi(N)$ for large $N$. Indeed, \begin{align*} \Psi(\mathcal{B},N)& = \sum\limits_{\substack{n\leqslant N\\ n\in \mathcal{B}}} \operatorname{min}(\frac{1}{2},\frac{s}{n\log n})\\ &\gg -O_s(1) + \sum\limits_{k=k_0}^{\lfloor\log_2 N\rfloor - 1 }\sum\limits_{\substack{2^k\leqslant n< 2^{k+1}\\n\in \mathcal{B}}} \frac{s}{2^k\log (2^k)}\\ &\gg -O_s(1) + \sum\limits_{k=k_0}^{\lfloor\log_2 N\rfloor -1}\sum\limits_{2^k\leqslant n< 2^{k+1}}(1-O_B(k^{-B})) \frac{s}{k 2^k}\\ &\gg -O_s(1) + \sum\limits_{k=1}^{\lfloor \log N \rfloor}\frac{s}{k}\\ &\gg_s \log\log N\\ &\gg_s \Psi(N). \end{align*} Therefore, applying Theorem \ref{Harman's theorem} to this set $\mathcal{B}$ and this function $\psi$, the main term from the conclusion of Theorem \ref{Harman's theorem} dominates the error term, and we conclude that for almost all $\alpha$ there are infinitely many $n\in \mathcal{B}$ satisfying $\Vert \alpha n\Vert < \frac{s}{n\log n}$. The proposition is proved. \end{proof} With this moderately technical proposition proved, the deduction of Theorem \ref{Main Theorem} is extremely short. \begin{proof}[Proof of Theorem \ref{Main Theorem}] Let $\Omega\subset [0,1]$ be the full-measure set of $\alpha$ for which Proposition \ref{key proposition} holds. Let $c$ be the constant from Proposition \ref{key proposition}, and fix some $s$ satisfying $0<2s<c^2$. Let $\alpha\in \Omega$, and fix a large $N$ to be one of the infinitely many natural numbers which satisfy the conclusions of Proposition \ref{key proposition}. By construction, we know that $$\Vert \alpha N\Vert < \frac{s}{ N\log N}.$$ Therefore, for all $d\leqslant \frac{1}{10}\log N$, we have $$\Vert \alpha dN\Vert < \frac{s}{ N}.$$ But by the second conclusion of Proposition \ref{key proposition}, this implies that there are at least $c^2N$ pairs of distinct primes $p_i,p_j\leqslant \frac{1}{2}N\log N$ such that $$\Vert \alpha (p_i-p_j)\Vert < \frac{s}{ N}.$$ Since $P_N\sim N\log N$, and $N$ is large, this certainly implies that $$F(\mathcal{P},\alpha,s,N)\geqslant c^2 >2s.$$ This holds for infinitely many $N$, and therefore for all $\alpha \in \Omega$ we have $$F(\mathcal{P},\alpha,2s,N)\neq 2s(1+o(1))$$ as $N\rightarrow \infty$. Since $\Omega$ has measure 1, Theorem \ref{Main Theorem} is proved. \end{proof} \bibliographystyle{plain}

https://arxiv.org/abs/1611.08921

Estimating volume and surface area of a convex body via its projections or sections

The main goal of this paper is to present a series of inequalities connecting the surface area measure of a convex body and surface area measure of its projections and sections. We present a solution of a question from S. Campi, P. Gritzmann and P. Gronchi regarding the asymptotic behavior of the best constant in a recently proposed reverse Loomis-Whitney inequality. Next we give a new sufficient condition for the slicing problem to have an affirmative answer, in terms of the least "outer volume ratio distance" from the class of intersection bodies of projections of at least proportional dimension of convex bodies. Finally, we show that certain geometric quantities such as the volume ratio and minimal surface area (after a suitable normalization) are not necessarily close to each other.

\section{Introduction} \hspace*{1.5em}In the past decades a lot of effort has been put in the study of problems of estimating volumetric quantities of a convex body (i.e. a convex compact set with non-empty interior) in terms of the corresponding functionals of its sections or projections. We refer to the books \cite{Ga, Ko1, KoY, RZ, S} for more information, examples and the history of those problems. Our aim is to continue this investigation by considering a number of problems of this type. In Section 3, we study inequalities involving the size of projections of $n$-dimensional convex body and study the following problem proposed in \cite{CGG}: \begin{question}\label{qu:lw} Find the smallest constant $\Lambda_n$, such that the following inequality holds for all convex bodies in $\mathbb{R}^n$: \begin{equation}\label{eq:lum} \min_{\{e_1,\dots,e_n\}\in{\mathcal F}_n}\prod_{i=1}^n\big|K|e_i^{\perp}\big|\leq \Lambda_n|K|^{n-1}, \end{equation} where ${\mathcal F}_n$ denotes the set of all orthonormal basis' in $\mathbb{R}^n$. \end{question} Inequality (\ref{eq:lum}) can be viewed as a reverse to the classical Loomis-Whitney \cite{LW} inequality for compact subset $A \subset \mathbb{R}^n$: \begin{equation}\label{eq:lw} \prod_{i=1}^n\big|A|e_i^{\perp}\big|\ge |A|^{n-1}. \end{equation} The authors in \cite{CGG} asked for the correct asymptotical behavior of $\Lambda_n^{1/n}$. In section 3 (Theorem \ref{thm-reverse-lw}) we show that $\Lambda_n^{1/n}\leq c\sqrt{n}$, for some absolute constant $c>0$. Moreover, we prove that this estimate is the best possible up to an absolute constant $c$. We note that other variants of reverse Loomis-Whitney type inequalities where considered in \cite{CGaG}. In Section 4, we study a number of inequalities which arise in the study of comparison problems. For example, what information on convex bodies we can get from comparison inequalities for its curvature functions. We also study the relationship of volume ratio, curvature measure and surface area of convex bodies. For instance we prove that a convex body (even highly symmetric) can have large volume ratio (close to the volume ratio of the cube of the same volume) but small minimal surface area (close to the surface area of the ball of the same volume). Notice that it is well known that the opposite cannot happen (see (\ref{eq-minimal-surface-vr}) below). Bourgain's slicing problem \cite{Bou1} asks whether any convex body in $\mathbb{R}^n$ of volume $1$ has a hyperplane section of volume greater than $c>0$, where $c$ is an absolute constant. It follows from the work of the first named author \cite[Theorem 1]{Ko2} (Theorem \ref{thm-Ko} in Section 5) that if a centrally symmetric convex body $K$ has bounded outer volume ratio with respect to the class of intersection bodies (see Section 5 for definition), then the slicing problem has an affirmative answer for $K$ (actually this result extends to general measures in place of volume). In Section 5, we extend this result (however, not bodywise) as follows (Theorem \ref{thm-slicing}): If every convex body in $\mathbb{R}^n $ has a projection of dimension at least proportional to $n$ which is close to an intersection body (in the sense of outer volume ratio), then the slicing problem will have an affirmative answer. \smallbreak \noindent {\bf Acknowledgment}. We are indebted to Matthieu Fradelizi for many valuable discussions and anonymous referee for many suggestions which led to a better presentation of the above results. \section{Background and notation} \hspace*{1.5em}We use the notation $a\sim b$ to declare that the ratio $a/b$ is bounded from above and from below by absolute constants. We denote by $D_n$ the standard $n$-dimensional Euclidean ball of volume $1,$ and by $B_2^n$ the Euclidean ball of radius 1. We denote the volume of $B_2^n$ by $\omega_n$. A set $L$ is called a star body if it has non-empty interior, it is star-shaped at the origin and its radial function $\rho_L$ is continuous. We remind that the radial function $\rho_L$ of $L$ is defined as: $$\rho_L(u)=\max\{\lambda>0:\lambda u\in L\}\ ,\qquad u\in S^{n-1}\ ,$$ where $S^{n-1}=\{x\in\mathbb{R}^n:|x|=1\}$ is the unit sphere in $\mathbb{R}^n$. In this section, $K$ will always denote a convex body in $\mathbb{R}^n$. The support function of $K$ is defined as $$h_K(x)=\max_{y\in K}\langle x,y\rangle\ ,\qquad x\in\mathbb{R}^n \ .$$ The surface area measure (or curvature measure) $S_K$ of $K$ is a measure on $S^{n-1}$, defined by $$S_K(\Omega)=\mathcal{H}^{n-1}\Big(\big\{x\in bdK:\exists u\in \Omega, \textnormal{ such that }\langle x,u\rangle=h_K(u)\big\}\Big)\ ,\qquad \Omega\textnormal{ is a Borel subset of }S^{n-1}\ ,$$ where $\mathcal{H}^{n-1}$ is the $(n-1)$-dimensional Haussdorff measure. If $S_K$ is absolutely continuous with respect to the Lebesgue measure, its density is denoted by $f_K$ and it is called the curvature function of $K$. A symmetric convex body $\Pi K$ is ``the projection body of $K$'' if its support function is given by \begin{equation}\label{eq:zonoid} h_{\Pi K}(u)=|K|u^{\perp}|= \frac{1}{2}\int\limits_{S^{n-1}} |x \cdot u| d S_K(x) \mbox{ for all } u\in S^{n-1}, \end{equation} where the last equality follows from the Cauchy formula for the volume of the orthogonal projection. Next, if $K_1,\dots,K_n$ are compact convex sets in $\mathbb{R}^n$, we denote their mixed volume by $V(K_1,\dots,K_n)$. We refer to \cite{S} or \cite{Ga} for an extensive discussion on the theory of mixed volumes. Note that if $S(K)$ is the surface area of $K$, then \begin{equation*}\label{eq-S(K)=V(...)} S(K)=nV(K[n-1],B_2^n)=\int_{S^{n-1}}dS_K. \end{equation*} In general, $$V(K[n-1],L)=\frac{1}{n}\int_{S^{n-1}}h_LdS_K$$ and $$|K|=\frac{1}{n}\int_{S^{n-1}}h_KdS_K.$$ The Minkowski inequality for mixed volumes states that \begin{equation}\label{eq-Minkowski} V(K[n-1],L)\geq |K|^{(n-1)/n}|L|^{1/n}, \end{equation} with equality if and only if $K$ and $L$ are homothetic. As F. John \cite{J} (see also \cite{AGM}, page 50) proved, there exists a unique ellipsoid $JK$ of maximal volume contained in $K$, the so-called ``John ellipsoid'' of $K$. Since, for $T\in GL(n)$ we have that $J(TK)=T(JK)$, there always exists $T\in SL(n)$, such that $J(TK)$ is a ball. The quantity $vr(K):=\big(|K|/|JK|\big)^{1/n}$ is called the volume ratio of $K$. Consider a convex body $K$, with $|K|=1$ and such that $JK$ is an Euclidean ball. Then, $$1=|K|=\frac{1}{n}\int_{S^{n-1}}h_KdS_K\geq \frac{1}{n}\int_{S^{n-1}}h_{JK}dS_K=\frac{|JK|^{1/n}}{n\omega_n^{1/n}}S(K).$$ We know by Stirling's formula that $\omega_n^{1/n}\sim 1/\sqrt{n}$, hence \begin{equation}\label{eq-surface-vr} S(K)\leq c\sqrt{n}\frac{1}{|JK|^{1/n}}=c\sqrt{n}vr(K)\ . \end{equation} Let us also define the quantity $$\partial(K):=\min_{T\in GL(n)}\frac{S(TK)}{|T K|^{(n-1)/n}}\ .$$ We say that $K$ is in minimal surface area position if $S(K)=\partial (K)$ and $|K|=1$ (see \cite{AGM}, Section 2.3). With this definition, (\ref{eq-surface-vr}) gives: \begin{equation}\label{eq-minimal-surface-vr} \partial(K)\leq c\sqrt{n}vr(K)\ . \end{equation} The parameters $\partial (K)$ and $vr(K)$ are affine invariants. A useful characterization of the minimal surface area position due to Petty \cite{Pe} states that a convex body $K$ of volume $1$ is in minimal surface area position if and only if its surface area measure $S_K$ is isotropic, i.e. \begin{equation}\label{eq-S_K-isotropic} \int_{S^{n-1}}|\langle x,u\rangle|^2dS_K(x)=\frac{1}{n}S(K)\ ,\qquad \textnormal{ for all }u\in S^{n-1}\ . \end{equation} We say that the convex body $K$ of volume $1$ is in minimal mean width position (see \cite[Section 2.2]{AGM}) if $$ \frac{1}{|S^{n-1}|}\int_{S^{n-1}} h_K(u) du \le \frac{1}{|S^{n-1}|}\int_{S^{n-1}} h_{TK}(u) du $$ for every $T\in SL(n)$. A very useful result which follows from Figiel, Tomczak-Jaegermann, Lewis and Pisier estimates (see \cite[Corollary 6.5.3]{AGM}) gives that for a symmetric convex body $K\subset \R^n$ in minimal mean width position and of volume $1$ we get \begin{equation}\label{eq:min} \frac{1}{|S^{n-1}|}\int_{S^{n-1}} h_K(u) du \le C \sqrt{n} \log n, \end{equation} for some absolute constant $C>0$. If $K$ contains the origin in its interior, the polar body $K^{\circ}$ of $K$ is defined to be the convex body $$K^{\circ}=\{x\in\mathbb{R}^n:\langle x,y\rangle\leq 1,\ \forall y\in K\}\ .$$ If $T\in GL(n)$ and $H\in G_{n,k}$, the following formulas hold: $$(TK)^{\circ}=T^{-\ast}K^{\circ}\qquad \textnormal{and}\qquad (K|H)^{\circ}=K^{\circ}\cap H\ .$$ Here, $G_{n,k}$ denotes the Grassmannian manifold of $k$-dimensional subspaces of $\mathbb{R}^n$ and the notation $A|H$ denotes the orthogonal projection of $A$ onto the subspace $H$. Let us assume that the origin is the centroid of $K$. The Blaschke-Santal\'o inequality (see \cite{San, RZ}) together with its reverse (see \cite{BM, RZ}) give: \begin{equation}\label{eq-Santalo} (|K|\cdot|K^{\circ}|)^{1/n}\sim 1/n\ . \end{equation} Set $N(K,D_n)$ to be the covering number of $K$ by translates of $D_n$, i.e. $$ N(K,D_n)=\min\{N\in\mathbb{N}: \exists \,\, x_1,\dots,x_N\in\mathbb{R}^n,\textnormal{ such that }K\subseteq \cup_{i=1}^N(x_i+D_n)\}\ . $$ Milman (\cite{M}, see also \cite[Chapter 8]{AGM}) proved that there exists an absolute constant $C>0$ and a linear image $K'$ of $K$ of volume 1, such that $$\max\Big\{N(K',D_n), N(|K'^{\circ}|^{-1/n}K'^{\circ},D_n), |K'\cap D_n|^{-1},|(|K'^{\circ}|^{-1/n}K'^{\circ})\cap D_n|^{-1}\Big\}\leq C^{n}\ .$$ If the previous inequality holds for $K'$, we say that $K'$ is in $M$-position. We will also need to use the notion of isotropic position of a convex body (see \cite{MP, BGVV}): there exists $T\in SL(n)$, such that $$\int_{TK}\langle x,y\rangle^2dy=L_K^2|K|^{(n+2)/n}|x|^2\ ,\qquad x\in\mathbb{R}^n \,$$ we will call $TK$ an isotropic convex body or a body in isotropic position. The parameter $L_K$ is called the isotropic constant of $K$. $L_K$ depends only on $K$ and it is invariant under invertible linear maps. It turns out that \begin{equation}\label{eq-L_K-variational} L_K^2=\frac{1}{n}\min_{T\in SL(n)}\frac{1}{|K|^{(n+2)/n}}\int_{TK}|x|^2dx=\frac{1}{n}\min_{T\in GL(n)}\frac{1}{|TK|^{(n+2)/n}}\int_{TK}|x|^2dx\ . \end{equation} It is a major problem to show that $L_K$ is uniformly bounded from above by an absolute constant (the fact that $L_K>c$ can be proved by comparison with an Euclidean ball; see \cite{MP, BGVV}). The best estimate currently is due to Klartag \cite{K}: $L_K\leq C'n^{1/4}$ who removed the logarithmic term in the previous estimate of Bourgain \cite{Bou2}. It should be noted that (see \cite{MP, BGVV}): \begin{equation}\label{eq-hensley} |K\cap u^{\perp}|\sim \bigg(\int_K\langle x,u\rangle^2dx\bigg)^{-1/2}\ ,\qquad u\in S^{n-1}\ . \end{equation} Thus, the problem of bounding $L_K$ is equivalent to Bourgain's slicing problem. \section{A remark on the reverse Loomis-Whitney inequality} The main goal of this section is to provide a sharp asymptotic estimate for the quantity $\Lambda_n^{1/n}$ in the reverse Loomis-Whitney inequality. \begin{theorem}\label{thm-reverse-lw} There exists an absolute constant $c>0$, such that \begin{equation}\label{eq-thm-lw-1} \Lambda_n\leq (c\sqrt{n})^n. \end{equation} Moreover, there exists a symmetric convex body $L$ in $\mathbb{R}^n$, such that \begin{equation}\label{eq-thm-lw-2}\min_{\{e_1,\dots,e_n\}\in{\mathcal F}_n}\prod_{i=1}^n\big|L|e_i^{\perp}\big|\geq (c'\sqrt{n})^n|L|^{n-1}\ , \end{equation} where $c'>0$ is another absolute constant. \end{theorem} The proof of Theorem \ref{thm-reverse-lw} follows from two theorems due to K. Ball. The first is K. Ball's reverse isoperimetric inequality (for symmetric convex bodies): \begin{oldtheorem}\label{oldthm-ball-1}\cite{Ba2} Let $K$ be a convex body and $\Delta$ be a simplex in $\mathbb{R}^n$. Then, $$\partial (K)\leq \partial (\Delta)\leq cn\ .$$ \end{oldtheorem} The second is a remarkable example of a convex body whose projections all have large volumes (comparing to its volume): \begin{oldtheorem}\label{oldthm-ball-2}\cite{Ba1} There exists a symmetric convex body $L$ in $\mathbb{R}^n$ of volume 1, such that $\big|L|u^{\perp}\big|\sim\sqrt{n}$, for all $u\in S^{n-1}$. \end{oldtheorem} \textit{}\\ \noindent{\it Proof of Theorem \ref{thm-reverse-lw}:} First, note that the convex body $L$ from Theorem \ref{oldthm-ball-2} serves as an example of a convex body that satisfies (\ref{eq-thm-lw-2}). To prove (\ref{eq-thm-lw-1}), we may assume that $K$ is of volume 1. Let us first consider a convex body $M$ in minimal surface area position (thus $|M|=1$). It was observed in \cite{GP}, that, in this case, the measure $S_M$ is isotropic. By equation (\ref{eq-S_K-isotropic}), Theorem \ref{oldthm-ball-1} and the Cauchy-Schwartz inequality \begin{eqnarray}\label{eq-proj-up-bound} \big|M|u^{\perp}\big|&=&\frac{1}{2}\int_{S^{n-1}}|\langle x,u\rangle|dS_M(x)\leq \frac{1}{2}\bigg(\int_{S^{n-1}}|\langle x,u\rangle|^2dS_M(x)\bigg)^{1/2} \bigg(\int_{S^{n-1}}dS_M(x)\bigg)^{1/2}\nonumber\\ &=&\frac{1}{2}\bigg(\frac{1}{n}\int_{S^{n-1}}|x|^2dS_M(x)\bigg)^{1/2} \bigg(\int_{S^{n-1}}dS_M(x)\bigg)^{1/2}\nonumber\\ &=&\frac{1}{2\sqrt{n}}\partial(M)|M|^{(n-1)/n}\leq c\sqrt{n}\ , \end{eqnarray} for all $u\in S^{n-1}$. Note that for every convex body $K$, $|K|=1$ there exists $T\in SL(n)$, such that $TK$ is in minimal surface area position. Since both (\ref{eq-S_K-isotropic}) and (\ref{eq-thm-lw-1}) are invariant under orthogonal transformations, we may assume that $T$ is a diagonal matrix with strictly positive entries. Then, $T^{-1}=\textnormal{diag}(\lambda_1,\dots,\lambda_n)$ for some $\lambda_1,\dots,\lambda_n>0$ with $\prod\lambda_i=1$. Consequently, if $\{e_1,\dots,e_n\}$ is the standard orthonormal basis in $\mathbb{R}^n$, it follows by (\ref{eq-proj-up-bound}) and the fact that $T^{-1}$ is diagonal that: \begin{eqnarray*} \prod_{i=1}^n\big|K|e_i^{\perp}\big|&=&\prod_{i=1}^n \big|T^{-1}T(K|e_i^{\perp})\big|=\prod_{i=1}^n\big|TK|e_i^{\perp}\big|\prod_{j\in\{1,\dots,n\}\setminus\{i\}}\lambda_j\\ &=&\prod_{i=1}^n\big|TK|e_i^{\perp}\big|\frac{1}{\lambda_i}\leq (c\sqrt{n})^n\prod_{i=1}^n\frac{1}{\lambda_i}=(c\sqrt{n})^n\ . \end{eqnarray*} \endproof \section{Curvature measures and comparison of volumes} Let us continue in another direction with the following observation: \begin{proposition}\label{planar} Let $K$, $L$ be centrally symmetric convex bodies in $\mathbb{R}^2$. If $S_K\leq S_L$, then $K\subseteq L$. \end{proposition} \noindent{\it Proof:} Using the definition of projection body of a convex body (see equation (\ref{eq:zonoid})) we get that $\Pi K\subseteq \Pi L$. However one can notice that for any symmetric convex body $M \subset \mathbb{R}^2$ we have $\Pi M=2O M$, where $O$ is the rotation by $\pi/2$. The result follows. \endproof \begin{remark} After a suitable translation, Proposition \ref{planar} remains true in the non-symmetric planar case as well. This is due to the additivity of the curvature measures of planar convex bodies. Indeed, set $\mu:=S_L-S_K$. Minkowski's Existence and Uniqueness Theorem (see \cite[Section 8.2]{S}) states that there exists a unique (up to translation) compact convex set $M$, such that $S_M=\mu$. Assuming that $K$, $L$ and $M$ contain the origin, we get: $L=K+M\supseteq K$, where we used the fact (see \cite[Section 8.3]{S}) that in the plane $S_{K+M}=S_K+S_M$ and again Minkowski's Existence and Uniqueness Theorem. \end{remark} One cannot naturally expect Proposition \ref{planar} to hold true in any dimension. Indeed, let $K=[-1,1]^{n-3}\times[-1/10,1/10]^2\times[-2,2]$ and $L=[-1,1]^n$, $n\ge 3$, then $S_K \le S_L$, but $K \not\subset L$. From another point of view, one can always guarantee the comparison of volumes: \begin{proposition}\label{notplanar} Let $K$, $L$ be centrally symmetric convex bodies in $\mathbb{R}^n$. If $S_K\leq S_L$, then $|K| \le |L|$. \end{proposition} \noindent{\it Proof:} Indeed, from $S_K\leq S_L$ we get $\int_{S^{n-1}} h_L S_K \le \int_{S^{n-1}} h_L S_L$, and can use Minkowski inequality (\ref{eq-Minkowski}) to finish the proof.\endproof Thus it is interesting to ask for comparison with a Euclidean Ball: If $K$ is a symmetric convex body of volume $1$, such that $JK$ is an Euclidean ball and $S_K\geq S_{cD_n}$, for some absolute constant $c>0$, is it true that $K$ has bounded volume ratio? Actually, one might ask a weaker (as (\ref{eq-minimal-surface-vr}) shows) version of the previous question: \begin{question}\label{question 1} Let $K$ be a symmetric convex body, such that $JK$ is a ball. If $S_K\geq S_{cD_n}$, is it true that the quantity $\partial(K)/\sqrt{n}$ is bounded from above by an absolute constant? \end{question} An extra motivation for Question \ref{question 1} is Proposition \ref{prop-conditional} (see below). As we show the answer to this question is negative. Recall that a convex body is called 1-symmetric if its symmetry group contains the symmetry group of the standard $n$-cube. \begin{theorem}\label{thm-curvature-vr} There exists an absolute contant $c>0$, such that for each positive integer $n$, there exists an 1-symmetric convex body $L$ of volume 1 in $\mathbb{R}^n$, such that $S_L>cS_{D_n}$, but $\partial(L)/\sqrt{n}\geq c\sqrt{n}$. \end{theorem} \begin{remark}\textit{} \begin{enumerate} \item Note that if $L$ as in Theorem \ref{thm-curvature-vr}, i.e. $1$-symmetric, then $JL$ is an Euclidean ball and $\partial(L)=S(L)$. \item K. Ball's reverse isoperimetric inequality shows that the factor $1/\sqrt{n}$ gives the worst possible order in Theorem \ref{thm-curvature-vr}. \item We also note that $cS_{D_n}=S_{c^{1/(n-1)}D_n}$, so the assumption in Theorem \ref{thm-curvature-vr} is much stronger than the assumption in Question \ref{question 1}. \end{enumerate} \end{remark} \noindent{\it Proof of Theorem \ref{thm-curvature-vr}:} Let $C_n$ be the standard $n$-cube of volume 1. Consider the convex body $K$ defined as the Blaschke average of $C_n$ and $D_n$: $$S_K=\frac{1}{2}S_{C_n}+\frac{1}{2}S_{D_n}\ .$$ Note that $C_n$ is in $M$-position (this, follows, for example from covering $C_n$ by copies of $D_n$, see Lemma 8.1.3 in \cite{AGM}), thus:$$|C_n\cap D_n|\geq C^{-n} \ .$$ Using the Minkowski inequality (\ref{eq-Minkowski}), we obtain: \begin{eqnarray*} C^{-1}|K|^{(n-1)/n}&\leq& |C_n\cap D_n|^{1/n}|K|^{(n-1)/n}\leq\frac{1}{n}\int_{S^{n-1}}h_{C_n\cap D_n}dS_K\\ &=&\frac{1}{2n}\int_{S^{n-1}}h_{C_n\cap D_n}dS_{C_n}+\frac{1}{2n}\int_{S^{n-1}}h_{C_n\cap D_n}dS_{D_n}\\ &\leq&\frac{1}{2n}\int_{S^{n-1}}h_{C_n}dS_{C_n}+\frac{1}{2n}\int_{S^{n-1}}h_{ D_n}dS_{D_n}\\ &=&\frac{1}{2}|C_n|+\frac{1}{2}|D_n|=1\ . \end{eqnarray*} This shows that \begin{equation}\label{eq-|K|^{(n-1)/n}} |K| ^{(n-1)/n}\leq C \ . \end{equation} Moreover, using the fact that $S(D_n)\sim \sqrt{n}$ and $S(C_n)=2n$, we get: $$\frac{1}{2}\frac{2n+c'\sqrt{n}}{C}\leq \frac{S(C_n)+S(D_n)}{2C}\leq \frac{S(K)}{|K|^{(n-1)/n}}\ ,$$ therefore $$\frac{S(K)}{|K|^{(n-1)/n}}\geq c''n\ .$$ Set $L:=(1/|K|^{1/n})K$. Then, $|L|=1$ and $$S_L\geq \frac{1}{2}\frac{1}{|K|^{(n-1)/n}}S_{D_n}\geq [1/(2C)]S_{D_n}\ .$$However, $$\partial(L)=\frac{S(K)}{|K|^{(n-1)/n}}\geq c''n\ ,$$as claimed. \endproof Let $K_0$ be a centrally symmetric star body in $\mathbb{R}^n$. Let us recall the definition of the curvature image $C(K_0)$ body of $K_0$ which is defined via the curvature function of $C(K_0)$ (see \cite{L1} for more information): $$f_{C(K_0)}(x)=\frac{\rho_{K_0}^{n+1}(x)}{n+1}\ ,\qquad x\in S^{n-1}\ .$$ A convex body $L$ is called ``body of elliptic type'' if there exists a convex body $K_0$, such that $L=C(K_0)$. \begin{question}\label{question 1'} Is Question \ref{question 1} true for symmetric bodies of elliptic type? \end{question} \begin{remark}\label{prop-conditional} We note that following ideas of the proof from \cite[Proposition 5.3]{Sa} one can show that if Question \ref{question 1'} had an affirmative answer, then the slicing problem would have an affirmative answer as well. \end{remark} As shown in \cite{Sa}, for any symmetric convex body $K$, $L_K\sim \partial(C(K))$ (so the reverse implication to the statement from Remark \ref{prop-conditional} obviously holds). Hence, it follows by (\ref{eq-minimal-surface-vr}) that if all centrally symmetric bodies of elliptic type have uniformly bounded volume ratios, then the isotropic constant would be uniformly bounded. It would be, therefore, natural to ask if the reverse inequality of (\ref{eq-minimal-surface-vr}) is true. \begin{question}\label{question 2} Is there an absolute constant $c'>0$. such that $$c'\sqrt{n}vr(K)\leq \partial(K)\ ,$$ for all centrally symmetric bodies of elliptic type $K$? \end{question} If Question \ref{question 2} had an affirmative answer, then for any symmetric convex body $K$, the following would hold true: $L_K$ is bounded if and only if $vr(C(K))$ is bounded. This provides a good motivation for our next result: a reverse inequality to (\ref{eq-minimal-surface-vr}) cannot hold true for general symmetric convex bodies. \begin{theorem}\label{counterexample} There exists a convex body $K$ in $\mathbb{R}^n$, such that $$\partial(K)\leq c'vr(K)\sqrt{\log(n+1)}\ .$$ \end{theorem} \begin{remark}\text{} \begin{enumerate} \item Actually $K$, from Theorem \ref{counterexample} can be taken to be 1-symmetric. \item It follows by K. Ball's volume ratio inequality \cite{Ba2} and the classical isoperimetric inequality that Theorem \ref{counterexample} gives the worst possible case up to a logarithmic factor. \end{enumerate} \end{remark} \noindent{\it Proof of Theorem \ref{counterexample}:} We first notice that the mean width of $B_{\infty}^n$ (the unit ball of $\ell_\infty^n$) is of the order $\sqrt{(\log (n+1))/n}$ (this can be calculated by passing to the integral over the gaussian measure and using standard estimates for the sequence of standard normal variables, see \cite{AGM} and \cite[Chapter 3, page 79]{LT} for more details). Therefore, integrating in polar coordinates, we get: \begin{eqnarray*} \int_{D_n}\|x\|_{\infty}dx&=&\frac{1}{n+1}\int_{S^{n-1}}\|x\|_{\infty}\rho^{n+1}_{D_n}dx\\ &=&\frac{\omega_n^{-\frac{n+1}{n}}}{n+1}|S^{n-1}|\int_{S^{n-1}}\|x\|_{\infty}\frac{dx}{|S^{n-1}|}\\ &\sim&\omega_n^{-1/n}\int_{S^{n-1}}\|x\|_{\infty}d\sigma(x)\\ &\sim&\sqrt{n}\sqrt{(\log (n+1))/n}=\sqrt{\log(n+1)}\ , \end{eqnarray*} where $\sigma$ is the Haar probability measure on $S^{n-1}$. We note that \begin{eqnarray*} \int_{D_n}\|x\|_{\infty}dx=\int_0^{\infty}|D_n\cap \{\|x\|_{\infty}> s\}|ds=\int_0^{\infty}|D_n\setminus sB^n_{\infty}|ds \ . \end{eqnarray*} Let $s_0>0$ be such that \begin{equation}\label{eq-s_0} |D_n\cap s_0B_{\infty}^n|=|D_n\setminus s_0B^n_{\infty}|=1/2\ . \end{equation} \begin{eqnarray*} \int_{D_n}\|x\|_{\infty}dx&=&\int_0^{s_0}|D_n\setminus sB^n_{\infty}|ds+\int_{s_0}^{\infty}|D_n\setminus sB^n_{\infty}|ds\\ &>&\int_0^{s_0}|D_n\setminus s_0B^n_{\infty}|ds+\int_{s_0}^{\infty}|D_n\setminus sB^n_{\infty}|ds>s_0/2\ . \end{eqnarray*} It follows that $s_0\leq C_1\sqrt{\log(n+1)}$. Take, now, $$K:=D_n\cap s_0B_{\infty}^n\ .$$ Note that $K$ is 1-symmetric, with $|K|=1/2$ and thus $$\partial(K)=S(K)\leq S(D_n)\leq C_2\sqrt{n}\ .$$ Moreover, since $K$ is 1-symmetric, $JK$ is an Euclidean ball and the largest Euclidean ball contained in $K$. Using the definition of $K$ we get that $JK$ is not larger then the largest Euclidean ball contained in $s_0B_{\infty}^n$, so $$JK\subseteq s_0B_2^n\leq (1/c_1)\sqrt{\log(n+1)}B_2^n\ ,$$for some absolute constant $c_1>0$. Thus, $$vr(K)\geq c_1\bigg(\frac{|K|}{|\sqrt{\log(n+1)}B_2^n|}\bigg)^{1/n}\geq c_1\frac{1/2}{\sqrt{\log(n+1)}\omega_n^{1/n}}\geq c_2\frac{\sqrt{n}}{\sqrt{\log(n+1)}}\geq \frac{c_2}{C_2}\frac{\partial(K)}{\sqrt{\log(n+1)}}\ .$$ \endproof \begin{remark} By a well known lemma of C. Borell \cite{Bor}, we have $$|D_n\setminus sB^n_{\infty}|\leq C_0e^{-c_0s}\ ,$$ for $s\geq 2s_0$, where $c_0, C_0>0$ are absolute constants and $s_0$ is defined by (\ref{eq-s_0}). Therefore, $$\int_{2s_0}^{\infty}|D_n\setminus sB_{\infty}^n|ds\leq \int_0^{\infty}C_0e^{-c_0s}ds\leq C_0'$$ and $$\int_0^{2s_0}|D_n\setminus sB^n_{\infty}ds|\leq 2s_0|D_n|=2s_0\ .$$ Thus, $s_0\geq c_3\sqrt{\log(n+1)}$, which shows that the logarithmic factor in the example of Theorem \ref{counterexample} cannot be removed. \end{remark} \begin{remark} Note that if $K$ is the example from Theorem \ref{counterexample} (recall that $|K|=1/2$), by (\ref{eq-S_K-isotropic}) and (\ref{eq-proj-up-bound}), we have $ |K|u^{\perp}|\leq c_4\ ,$ where we used the fact that since $K$ is 1-symmetric, $S_K$ is isotropic. Also, $|K|u^{\perp}|\geq |K\cap u^{\perp}|\geq c_{5}|K|^{(n-1)/n}\geq c_6$, since $K$ is in isotropic position and it has bounded isotropic constant, as 1-symmetric. Consequently (after a suitable dilation), we may take $K$ in Theorem \ref{counterexample} such that all its projections have volumes of constant order, when it is in minimal surface area position. \end{remark} The next two theorems continue the discussion started in Proposition \ref{notplanar}. Our goal is to provide bounds on the difference of volumes of convex bodies via its curvature functions. \begin{thm}\label{th:min} Let $K, L \subset \R^n$ be convex bodies with absolutely continuous surface area measure, such that $f_K(\theta) \le f_L(\theta)$, for all $\theta\in S^{n-1}$. Then $$ |L|^{\frac{n-1}{n}} - |K|^{\frac{n-1}{n}} \ge \omega_n^{\frac{n-1}{n}} \min\limits_{\theta \in S^{n-1}} (f_L(\theta)- f_K(\theta)). $$ \end{thm} \noindent{\it Proof:} Consider two convex bodies $K,L \subset \R^n$ such that \begin{equation}\label{eq:minf} f_K(\theta) \le f_L(\theta)- \epsilon, \end{equation} for some $\epsilon \ge 0$ and all $\theta \in S^{n-1}$. Then $$ \int\limits_{S^{n-1}} h_L(\theta) f_K(\theta) d \theta \le \int\limits_{S^{n-1}} h_L(\theta) f_L(\theta) d \theta - \epsilon \int\limits_{S^{n-1}} h_L(\theta) d \theta $$ or $$ V(K[n-1], L) \le |L| -\epsilon V(B_2^n[n-1], L). $$ Applying the Minkowski inequality (\ref{eq-Minkowski}) to $V(K[n-1], L)$ and $V(B_2^n[n-1], L)$ we get $$ |L|^{\frac{1}{n}}|K|^{\frac{n-1}{n}} \le |L| -\epsilon |L|^{\frac{1}{n}}|B_2^n|^{\frac{n-1}{n}}. $$ Thus $$ |L|^{\frac{n-1}{n}} - |K|^{\frac{n-1}{n}} \ge \epsilon |B_2^n|^{\frac{n-1}{n}}. $$ To finish the proof of the theorem we note that (\ref{eq:minf}) is always true with $\epsilon = \min\limits_{\theta \in S^{n-1}} (f_L(\theta)- f_K(\theta))$. \endproof We note that if we can consider $K$ to be $rB_2^n$ in the above theorem and send $r \to 0$ to get that $$ |L|^{\frac{n-1}{n}} \ge \omega_n^{\frac{n-1}{n}} \min\limits_{\theta \in S^{n-1}} (f_L(\theta)), $$ for any $L$ with absolutely continuous surface area measure. \begin{thm}\label{th:max} There exists an absolute constant $C>0$ such that for any $K, L \subset \R^n$, two convex bodies with absolutely continuous surface area measure, with $f_L(\theta) \ge f_K(\theta),$ for all $\theta \in S^{n-1}$ and such that $K$ is in the minimal width position, we have $$ |L|^{\frac{n-1}{n}} - |K|^{\frac{n-1}{n}} \le C \log n \,\, \omega_n^{\frac{n-1}{n}} \max \limits_{\theta \in S^{n-1}} (f_L(\theta)- f_K(\theta)). $$ \end{thm} \noindent{\it Proof:} Consider two convex bodies $K,L \subset \R^n$ such that \begin{equation}\label{eq:max} f_L(\theta) \le f_K(\theta)+ \epsilon, \end{equation} for some $\epsilon \ge 0$ and all $\theta \in S^{n-1}$. Then $$ \int\limits_{S^{n-1}} h_K(\theta) f_L(\theta) d \theta \le \int\limits_{S^{n-1}} h_K(\theta) f_K(\theta) d \theta + \epsilon \int\limits_{S^{n-1}} h_K(\theta) d \theta $$ or $$ V(L[n-1], K) \le |K| +\epsilon \frac{1}{n} \int\limits_{S^{n-1}} h_K(\theta) d \theta. $$ Now we can apply the Minkowski inequality (\ref{eq-Minkowski}) to $V(L[n-1], K)$. We can also use the fact that $K$ is in minimal width position and apply inequality (\ref{eq:min}) to claim that $$ \int\limits_{S^{n-1}} h_K(\theta) d \theta \le C |S^{n-1}| |K|^{\frac{1}{n}} \sqrt{n} \log n = C \omega_n |K|^{\frac{1}{n}} n \sqrt{n} \log n. $$ Thus $$ |K|^{\frac{1}{n}}|L|^{\frac{n-1}{n}} \le |K| + \epsilon \frac{1}{n} C \omega_n |K|^{\frac{1}{n}} n \sqrt{n} \log n. $$ or $$ |L|^{\frac{n-1}{n}} - |K|^{\frac{n-1}{n}} \le \epsilon C |B_2^n|^{\frac{n-1}{n}} \log n. $$ To finish the proof of the theorem we note that (\ref{eq:max}) is always true for $\epsilon = \max\limits_{\theta \in S^{n-1}} (f_L(\theta)- f_K(\theta))$. \endproof \section{A note on the slicing problem} Recall that the notion of an {\it intersection body of a star body} was introduced by E. Lutwak \cite{L2}: $IL$ is called the intersection body of $L$ if the radial function of $IL$ in every direction is equal to the $(n-1)$-dimensional volume of the central hyperplane section of $L$ perpendicular to this direction: $\forall \xi\in S^{n-1},$ $$ \label{eq:int}\rho_{IL}(u)= |L\cap u^{\perp}|, $$ As shown in \cite{L2} (using a result of Busemann's theorem \cite{Bu}), if $L$ happens to be a symmetric convex body, then $IL$ is also a symmetric convex body. A more general class of {\it intersection bodies} was defined by R. Gardner \cite{Ga1} and G. Zhang \cite{Zh} as the closure of intersection bodies of star bodies in the radial metric $d(K,L)=\sup_{\xi\in S^{n-1}} |\rho_K(\xi)- \rho_L(\xi)|$. We refer reader to books \cite{Ga, Ko1, KoY, RZ} for more information on definition and properties of Intersection body and their applications in Convex Geometry and Geometric Tomography. Define the quantity $L_n:=\max\{L_K:K\textnormal{ is a convex body in }\mathbb{R}^n\}.$ For $n,k\in\mathbb{N}$, $k\leq n$ and $c>0$, define the class $\mathcal{C}_{n,k,c}$ of convex bodies $K$ in $\mathbb{R}^n$, that have a $l$-dimensional projection $P$, with $l\geq k$, for which there exists an $l$-dimensional intersection body, such that $P\subseteq L$ and $|L|/|P|\leq c^n$. Finally, define $I_{n,k}$ to be the smallest constant $t$, for which $\mathcal{C}_{n,k,t}$ contains all centrally symmetric convex bodies in $\mathbb{R}^n$. The following was proved in \cite[Corollary 1]{Ko2}: \begin{oldtheorem}\label{thm-Ko} There exists an absolute constant $\overline{C}_0>0$, such that if $L$ is a symmetric convex body in $\mathbb{R}^n$, such that $L\in \mathcal{C}_{n,n,a}$, for some $a>0$ and $\mu $ is any even measure with continuous density in $\mathbb{R}^n$, then \begin{equation}\label{eq-general-slicing} \mu(L)\leq (\overline{C}_0/a)\max_{u\in S^{n-1}}\mu(L\cap u^{\perp})|L|^{1/n} \ . \end{equation} \end{oldtheorem} It follows by Theorem \ref{thm-Ko}, applied to $\mu$ being standard Lebesgue measure, together with (\ref{eq-hensley}) that $L_n\leq \overline{C}_0'I_{n,n}$, for some absolute constant $\overline{C}_0'>0$. Our goal is to replace $I_{n,n}$ with $I_{n,k}$, where $k$ is at least proportional to $n$. \begin{theorem}\label{thm-slicing} There exists an absolute constant $\widetilde{C}>0$, such that for any $n\in\mathbb{N}$ and for any $\lambda\in(0,1]$, it holds $$L_n\leq \widetilde{C}^{1/\lambda}I_{n,\lfloor\lambda n\rfloor}\ ,$$where $\lfloor\cdot\rfloor$ denotes the floor function. \end{theorem} We also note that if $K$ has bounded outer volume ratio (in which case $K$ also has bounded isotropic constant), then $K^\circ$ has a bounded volume ratio and thus (see for example \cite[Theorem 6.1]{Pi}) $K^\circ$ has an almost Euclidean section of proportional dimension, which gives a projection of $K$ of proportional dimension which is almost Euclidean. The following question is therefore natural: \begin{question} Is it true that any convex body $K \subset \mathbb{R}^n$ has projections of at least proportional dimension that are close to some intersection body in the sense of the Banach-Mazur distance? \end{question} Before proving Theorem \ref{thm-slicing}, we will need some geometric statements. \begin{lemma}\label{lemma-small-projection} Let $K$ be a symmetric convex body of volume 1, in $M$ position. There exists an absolute constant $C_0>0$, such that for any subspace $H$ of $\mathbb{R}^n$, it holds $$1/C_0\leq|K\cap H|^{1/n}\leq |K|H|^{1/n}\leq C_0\ .$$ \end{lemma} \noindent{\it Proof:} By assumption, there exist points $x_1,\dots,x_N\in\mathbb{R}^n$, with $N<C^n$, such that $$K\subseteq \bigcup_{i=1}^N(x_i+D_n)\ .$$ Let $H\in G_{n,k}$. Then, since $D_n|H=\omega_n^{-1/n}B_2^k$, we obtain \begin{eqnarray*} |K|H|&\leq& \Big|\bigcup_{i=1}^N(x_i+D_n) | H \Big|\leq N|D_n|H| = N\omega_n^{-k/n}|B_2^k|\\ &=&N\omega_n^{-k/n}\omega_k\leq NC'^k(\sqrt{n})^{k}(\sqrt{k})^{-k} \leq (CC')^n(\sqrt{n/k})^k \ , \end{eqnarray*} where we used the fact that $\omega_n^{1/n}\sim 1/\sqrt{n}$. Thus, $$|K|H|^{1/n}\leq \overline{C} \ .$$ On the other hand, if we replace $K$ by $(1/|K^{\circ}|)^{1/n}K^{\circ}$ and since $\big|(1/|K^{\circ}|)^{1/n}K^{\circ}\big|=1$, we get: $$\Big|\big(\frac{1}{|K^{\circ}|^{1/n}}K^{\circ}\big)|H\Big|^{1/n}\leq \overline{C} \ ,$$ which gives \begin{equation}\label{eq-n^2} |K^{\circ}|H|^{1/n}\leq \overline{C}|K^{\circ}|^{k/n^2}\ . \end{equation} Using (\ref{eq-Santalo}) and the fact that $(K^{\circ}|H)^{\circ}=K\cap H$, we obtain $$(c_1/k)^{k/n}|K\cap H|^{-1/n}\leq |K^{\circ}|H|^{1/n}.$$ This together with (\ref{eq-n^2}) give $$(c_1/k)^{k/n}|K\cap H|^{-1/n}\leq \overline{C}|K^{\circ}|^{k/n^2}\leq \overline{C}(c_2/n)^{k/n}$$ or $$|K\cap H|^{1/n}\geq (nc_1/(c_2k))^{k/n}/\overline{C}\geq \overline{c}\ ,$$as required. \endproof The next proposition is based on the existence of an $M$-position. \begin{proposition}\label{prop-small-proj} Let $K'$ be a symmetric convex body of volume 1, be such that $|K'\cap H|^{1/n}\leq C'$, for all subspaces $H$ of $\mathbb{R}^n$, for some absolute constant $C'>0$. Then, there exists an absolute constant $C'''$, such that $$|K'|H|^{1/n}\leq C'''\ ,$$for all subspaces $H$ of $\mathbb{R}^n$. \end{proposition} \noindent{\it Proof:} Consider a convex body $K$ of volume $1$, in $M$-position, such that $K'=TK$, for some $T\in SL(n)$. As in the proof of Theorem \ref{thm-reverse-lw}, we may assume that $T$ is diagonal; actually we are only going to need the fact that $T$ is symmetric, i.e. $T^{\ast}=T$. We will make use of a special case of a result from \cite{F} (see also \cite[(5.28)]{S}), stating that if $M$ is a convex body in $\mathbb{R}^n$, $H\in G_{n,k}$ and $P$ is a convex body in $H^{\perp}$, then \begin{equation}\label{eq-projections-general} |M|H|=\binom{n}{k}\frac{1}{|P|}V\big(M[k],P[n-k]\big) \ . \end{equation} Let $k\in\{1,\dots,n-1\}$ and $H\in G_{n,k}$. Now we apply Lemma \ref{lemma-small-projection} together with (\ref{eq-Santalo}) to get: \begin{equation}\label{eq-polar-sections} \big|\big((TK)\cap H\big)^{\circ}\big|^{1/n}\geq (c_1/k)^{k/n}|(TK)\cap H|^{-1/n}\geq (c_1/k)^{k/n}/c' \end{equation} and similarly, \begin{equation}\label{eq-polar-projections} |K^{\circ}|H|^{1/n}\leq (c_2/k)^{k/n}|K\cap H|^{-1/n}\leq (c_2/k)^{k/n}/c''\ . \end{equation} Moreover, for any convex body $P$ in $H^{\perp}$, using (\ref{eq-projections-general}) we obtain: \begin{eqnarray}\label{eq-after-polar-projections} \big|(TK)^{\circ}|H\big|&=&\binom{n}{k}\frac{1}{|P|}V\big((T^{-\ast}K^{\circ})[k],P[n-k]\big)\nonumber \\ &=&\binom{n}{k}\frac{1}{|P|}V\big(K^{\circ}[k],(T^{\ast}P)[n-k]\big)\nonumber\\ &=&\binom{n}{k}\frac{1}{|P|}V\big(K^{\circ}[k],(TP)[n-k]\big)\nonumber\\ &=&\frac{|TP|}{|P|}\binom{n}{k}\frac{1}{|TP|}V\big(K^{\circ}[k],(TP)[n-k]\big)\nonumber\\ &=&\frac{|TP|}{|P|}\big|K^{\circ}|(T^{-\ast}H)^{\perp}\big|\ , \end{eqnarray} where we used the fact that $TP$ is a convex body in $(T^{-\ast}H)^{\perp}$. Thus, (\ref{eq-polar-sections}), (\ref{eq-polar-projections}) and (\ref{eq-after-polar-projections}) imply $$\bigg(\frac{|TP|}{|P|}\bigg)^{1/n}\geq c''' \ .$$ Replacing $P$ by $T^{-1}P$, we get: \begin{equation}\label{eq-T^{-1}P} |P|^{1/n}\geq c'''|T^{-1}P|^{1/n}\ , \end{equation} for all convex bodies $P$ in $H$, for all subspaces $H$ of $\mathbb{R}^n$. On the other hand, as before one can write: \begin{eqnarray*} |(TK)|H|&=&\binom{n}{k}\frac{1}{|P|}V\big((TK)[k],P[n-k]\big)\\ &=&\binom{n}{k}\frac{1}{|P|}V\big(K[k],(T^{-1}P)[n-k]\big)\\ &=&\frac{|T^{-1}P|}{|P|}\binom{n}{k}\frac{1}{|T^{-1}P|}V\big(K[k],(T^{-1}P)[n-k]\big)\\ &=&\frac{|T^{-1}P|}{|P|}\big|K|(T^{\ast}H)\big|\ . \end{eqnarray*} Therefore, the previous identity, Lemma \ref{lemma-small-projection} and (\ref{eq-T^{-1}P}) give: $$|K'|H|^{1/n}=\big|(TK)|H\big|^{1/n}\leq \big(|T^{-1}P|/|P|\big)^{1/n}\big|K|(T^{\ast}H)\big|^{1/n}\leq C'''/c'''\ ,$$ for any subspace $H$ of $\mathbb{R}^n$. \endproof \begin{proposition}\label{prop-slicing-main} Let $\lambda\in(0,1]$ and $K$ be a symmetric convex body in $\mathbb{R}^n$ of volume $1$, such that $\big|K|H\big|^{1/n}\leq A$, for some $A>0$ and for all subspaces $H$ of $\mathbb{R}^n$. If, in addition, $K\in \mathcal{C}_{n,k,t}$, for some $t>0$, $k\geq \lambda n$, then there exists an $(n-1)$-dimensional subspace $F$, such that $$|K\cap F|\geq \frac{\widetilde{c}}{A^{1/\lambda}t}\ ,$$ for some absolute constant $\widetilde{c}>0$. \end{proposition} Let $H\in G_{n,k}$. By Fubini's Theorem, we have: \begin{equation}\label{eq-1st-fubini} 1=|K|=\int_{K|H}|K\cap (H^{\perp}+x)|dx\ . \end{equation} Let $G$ be a codimension $1$ subspace of $H$. Then, the subspace $F:=\textnormal{span}\{G\cup H^{\perp}\}$ has dimension $n-1$. The following claim is the key step in the proof of Proposition \ref{prop-slicing-main}:\\ \noindent\textbf{Claim.} $(K\cap F)|H=(K|H)\cap G$ and $(K\cap F)\cap (H^{\perp}+x)=K\cap (H^{\perp}+x)$, for all $x\in G$. \noindent{\it Proof of the claim:} To prove the first part, let $x\in(K|H)\cap G $. Since $x\in K|H$, there exists $y\in H^{\perp}$, such that $x+y\in K$. Since $x\in G$, $x+y\in F$, it follows that $x+y\in K\cap F$, so $x\in (K\cap F)|H$. Thus, $(K\cap F)|H\subseteq (K|H)\cap G$. Conversely, let $x\in (K\cap F)|H$. Then, there exists $y\in H^{\perp}$, such that $x+y\in K\cap F$. Thus, $x+y\in K$ and $x+y\in F$, so $x\in K|H$ and $x\in F|H=G$. Consequently, $x\in (K|H)\cap G$, which shows that $ (K|H)\cap G \subseteq(K\cap F)|H$, as required. To prove the second part, note that for any $x\in G$, $x+H^{\perp}\subseteq\textnormal{span}(G\cup H^{\perp})=F$, thus $K\cap(x+H^{\perp})\subseteq (K\cap F)\cap (x+H^{\perp})$. Since the opposite inclusion is trivial, our claim is proved. \endproof \noindent{\it Proof of Proposition \ref{prop-slicing-main}:} Using again Fubini's Theorem and the previous claim, we get: \begin{equation}\label{eq-2nd-fubini} |K\cap F|=\int_{(K\cap F)|H}|K\cap F\cap (H^{\perp}+x)|dx=\int_{(K|H)\cap G}|K\cap F\cap (H^{\perp}+x)|dx=\int_{(K|H)\cap G}|K\cap (H^{\perp}+x)|dx\ . \end{equation} Assume, now, that $K|H\subseteq L$ and $|L|/|K|H|<t^n$, where $L$ is an intersection body in some $k$-dimensional subspace $H$ of $\mathbb{R}^n$. Set $\mu$ to be the measure with density $H\in x\mapsto |K\cap (H^{\perp}+x)|$ in $H$ (i.e. $\mu$ is the marginal of the uniform measure on $K$ with respect to the subspace $H$; see e.g. \cite{Ba3}) and $L:=K|H\subseteq H$ . Using Theorem \ref{thm-Ko}, we conclude that there exists a codimension-1 subspace $G$ of $H$, such that: \begin{equation}\label{eq-slicing-last} \mu((K|H)\cap G)|K|H|^{1/k}\geq (\widetilde{c}/t)\mu(K|H)\ , \end{equation} for some absolute constant $\widetilde{c}>0$. Since $$|K|H|^{1/k}\leq A^{1/\lambda}\ ,$$ if we set $F=\textnormal{span}\{G\cup H^{\perp}\}$, applying (\ref{eq-1st-fubini}) and (\ref{eq-2nd-fubini}) to (\ref{eq-slicing-last}), we obtain: $$|K\cap F|=\mu((K|H)\cap G)\geq \frac{\widetilde{c}_0}{A^{1/\lambda}t}\ ,$$as claimed. \endproof \noindent{\it Proof of Theorem \ref{thm-slicing}.} As shown in \cite{K}, there exists a symmetric convex body $K$, such that $L_K\geq c_0L_n$. We may assume that $K$ is of volume 1 and in isotropic position. Then, as it was proved in \cite[Corollary 3.5]{BKM}, $|K\cap H|^{1/n}$ is bounded from above by an absolute constant. It follows immediately by Propositions \ref{prop-small-proj}, \ref{prop-slicing-main} and the definition of $I_{n,k}$ that there exists an $(n-1)$-dimensional subspace $F$, such that $$|K\cap F|\geq \frac{\widetilde{c}}{C^{1/\lambda}I_{n,[\lambda n]}}\ ,$$ for some absolute constant $C>0$. Our claim follows immediately from the previous inequality and (\ref{eq-hensley}). \endproof We remark that our method can be used to increase the class of symmetric convex bodies that are known to satisfy (\ref{eq-general-slicing}) (with an absolute constant). Indeed, one can replace the Lebesgue measure in Proposition \ref{prop-slicing-main} by any even measure (one has to repeat the steps of the proof) to obtain the following: \begin{proposition}\label{prop-main-general} Let $\mu$ be an even measure in $\mathbb{R}^n$, with continuous density, $\lambda\in(0,1]$ and $K$ be a 0-symmetric convex body in $\mathbb{R}^n$, such that $\big|K|H\big|^{1/k}\leq A|K|^{1/n}$, for some $A>0$ and for all $k\geq \lambda n$ and $H\in G_{n,k}$. If, in addition, $K\in \mathcal{C}_{n,k_0,t}$, for some $t>0$ and for some $k_0\geq \lambda n$, then there exists an $(n-1)$-dimensional subspace $F$, such that $$\mu(K\cap F)|K|^{1/n}\geq \frac{\widetilde{c}}{A^{1/\lambda}t}\mu(K)\ ,$$ for some absolute constant $\widetilde{c}>0$. \end{proposition} In particular, it follows from the previous statement and from Lemma \ref{lemma-small-projection} that if $K\in \mathcal{C}_{n,k,c}$, for some $k\geq \lambda n$ and $K$ is in $M$-position, then (\ref{eq-general-slicing}) holds for $K$ (with an absolute constant) for all even measures $\mu$ with continuous density. As an example, take $K=aB_2^n\times L$, where $L$ is any convex body in $M$-position, $\textnormal{dim}L=n$ and $|L|=|aB_2^n|$.

https://arxiv.org/abs/1910.03241

A Fast Exact Algorithm for Airplane Refueling Problem

We consider the airplane refueling problem, where we have a fleet of airplanes that can refuel each other. Each airplane is characterized by specific fuel tank volume and fuel consumption rate, and the goal is to find a drop out order of airplanes that last airplane in the air can reach as far as possible. This problem is equivalent to the scheduling problem $1||\sum w_j (- \frac{1}{C_j})$. Based on the dominance properties among jobs, we reveal some structural properties of the problem and propose a recursive algorithm to solve the problem exactly. The running time of our algorithm is directly related to the number of schedules that do not violate the dominance properties. An experimental study shows our algorithm outperforms state of the art exact algorithms and is efficient on larger instances.

\section{Introduction} \label{sec:introduction} The \emph{airplane refueling problem}, originally introduced by Gamow and Stern \cite{puzzle_math_1958}, is a special case of single machine scheduling problem. Consider a fleet of several airplanes with certain fuel tank volume and fuel consumption rate. Suppose all airplanes travel at the same speed and they can refuel each other in the air. An airplane will drop out of the fleet if it has given out all its fuel to other airplanes. The goal is to find a drop out order so that the last airplane can reach as far as possible. \subsubsection{Problem definition} \label{ssub:problem_definition} In the original description of airplane refueling problem, all airplanes are defaulted to be identical. Woeginger \cite{chrobaketal:DSP:2010:2536} generalized this problem that each airplane $j$ can have arbitrary tank volume $w_j$ and consumption rate $p_j$. Denote the set of all airplanes by $J$, a solution is a drop out order $\sigma: \{1, 2, \dots, |J|\} \mapsto J$, where $\sigma(i) = j$ if airplane $j$ is the $i^{\text{\tiny th}}$ airplane to leave the fleet. As a result, the objective value of the drop out order $\sigma$ is: \[ \sum_{j = 1}^{n} \Big( {w_{\sigma(j)}} \text{\bfseries \large{/}} \text{\tiny \ } {\sum_{k = j}^{n}p_{\sigma(k)}} \Big). \] As pointed out by Vásquez \cite{vasquez_for_2015}, we can rephrase the problem as a single machine scheduling problem, which is equivalent to finding a permutation $\pi$(the reverse of $\sigma$) that minimizes: \[ \sum_{j = 1}^{n} \Big( - {w_{\pi(j)}} \text{\bfseries \large{/}} \text{\tiny \ } {\sum_{k = j}^{n}p_{\pi(k)}} \Big) = \sum_{j=1}^{n} -w_j / C_j, \] where $C_j$ is the completion time of job $j$, and $p_j, w_j$ correspond to its processing time and weight, respectively. This scheduling problem is specified as $1||\sum w_j (- \frac{1}{C_j})$ using the classification scheme introduced by Graham et al. \cite{graham_optimization_1979}. In the rest of this paper, we study the equivalent scheduling problem instead. \subsubsection{Dominance properties} \label{ssub:dominance_relation} Since the computational complexity status of airplane scheduling problem is still open \cite{chrobaketal:DSP:2010:2536}, existing algorithms find the optimal solution with branch and bound method. While making branching decisions in a branch and bound search, it would be much more useful if we know the dominance relation among jobs. For example, if we know job $i$ always precedes job $j$ in an optimal solution, we can speed up the searching process by pruning all the branches with job $i$ processed after job $j$. Let the start time of job $j$ be $t_j$, we refer to \cite{durr_order_2014} for the definition of \emph{local dominance} and \emph{global dominance} : \begin{itemize} \item \textit{local dominance:} Suppose job $j$ starts at time $t$ and is followed directly by job $i$ in a schedule. If exchanging the positions of jobs $i,j$ strictly improves the cost, we say that \emph{job $i$ locally dominants job $j$ at time $t$} and denote this property by $i \prec_{l(t)} j$. If $i \prec_{l(t)} j$ holds for all $t \in [a, b]$, we denote it by $i \prec_{l[a, b]} j$. \item \textit{global dominance:} Suppose schedule $S$ satisfies $a \leq t_j \leq t_i - p_j \leq b$. If exchanging the positions of jobs $i,j$ strictly improves the cost, we say that \emph{job $i$ globally dominants job $i$ in time interval $[a, b]$} and denote this property by $i \prec_{g[a, b]} j$. If it holds that $i \prec_{g[0, \infty)} j$, we denote this property by $i \prec_{g} j$. \end{itemize} We call the schedule that do not violate the dominance properties as \emph{potential schedule}. The effect of dominance properties is to narrow the search space to the set of all potential schedules, whose cardinality is much smaller than $n!$, the number of all job permutations. \subsubsection{Related work} \label{ssub:related_work} Airplane refueling problem is a special case of a more general scheduling problem $1||\sum w_j C_j^\beta$. For most problems of this form, including airplane refueling problem, no polynomial algorithm has been found. Existing methods resort to approximation algorithms or branch and bound schemes. For approximations, several constant factor approximations and polynomial time approximation scheme(PTAS) have been devised for different cost functions \cite{bansal_geometry_2014,cheung_primal-dual_2017,hohn_how_2018,megow_dual_2013}. Recently, Gamzu and Segev\cite{gamzu_polynomial-time_2019} gave the first polynomial-time approximation scheme for airplane refueling problem. The focus of exact methods is to find stronger dominance properties. Following a long series of improvements \cite{bagga_node_1980,croce_minimizing_1993,bader_experimental_2012,mondal_improved_2000-1,sen_minimizing_1990}, Dürr and Vásquez \cite{durr_order_2014} conjectured that for all cost functions of the form $f_j(t) = w_jt^\beta, \beta >0$ and all jobs $i, j$, $i \prec_l j$ implies $i \prec_g j$. Latter, Bansal et al. \cite{bansal_localglobal_2017} confirmed this conjecture, and they also gave a counter example of the generalized conjecture that $i \prec_{l[a, b]} j$ implies $i \prec_{g[a, b]} j$. For airplane refueling problem, Vásquez \cite{vasquez_for_2015} proved that $i \prec_{l[a, b]} j$ implies $i \prec_{g[a, b]} j$. The establish dominance properties are commonly incorporated into a branch and bound scheme, such as Algorithm $A^*$ \cite{hart_formal_1968}, to speed up the searching process. \subsubsection{Our contribution} \label{ssub:our_contribution} Existing branch and bound algorithms search for potential schedules in a trail and error manner. Specifically, when making branching decisions, it is unknown whether current branch contains any potential schedule unless we exhaust the entire branch. So if we can prune all the branches that do not contain potential schedule, it will considerably improve the efficiency of the searching process. In this paper we give an exact algorithm with the merit above for airplane refueling problem. Specifically, every branch explored by our algorithm is guaranteed to contain at least one potential schedule, and the time to find each potential schedule is bounded by $\bigO(n^2)$. Numerical experiments exhibit empirical advantage of our algorithm over the $A^*$ algorithms proposed by former studies, and the advantage is more significant on instances with more jobs. The main difference between previous methods and our algorithm is that instead of branching on possible succeeding jobs, we branch on the possible start times of a certain job. To this end, we introduce a prefixing technique to determine the possible start times of a certain job in a potential schedule. In addition, the relative order of other jobs regarding that certain job is also decided. Thus, each branch divides the original problem into subproblems with fewer jobs and we can solve the problem recursively. \subsubsection{Organization} \label{ssub:organization} The rest of the paper is organized as follows. In section 2, we introduce a new auxiliary function and give a concise representation of dominance property. Section 3 establishes some useful lemmas. We present our algorithm in section 4 and experimental results in section 5. Finally, section 6 concludes the paper. \section{Preliminaries} \label{sec:preliminaries} A \emph{dominance rule} stipulates the local and global dominance properties among jobs. We can refer to a dominance rule by the set $R:= \cup_{i, j \in J} \{r_{ij}\}$, where $r_{ij}$ represents the dominance properties between jobs $i, j$ specified by the rule. Given a dominance rule $R$, we formally define the potential schedule as follows. \begin{definition}[Potential Schedule] We call $S$ a potential schedule with respect to dominance rule R if for all $r_{ij} \in R$, start times of jobs $i,j$ in $S$ do not violate relation $r_{ij}$. \end{definition} \subsubsection{Auxiliary function} \label{ssub:auxiliary_function} For each job $j$, we introduce an auxiliary function $\varphi_j(t)$, where $t \geq 0$ represents the possible start time: \[ \varphi_j(t) = \frac{w_j}{p_j(p_j + t)}. \] We remark that $\varphi_j(t)$ is well defined since the processing time $p_j$ is positive. With the help of this auxiliary function, we can obtain a concise characterization of local precedence between two jobs. \begin{lemma} \label{lemma: local_new} For any two jobs $i, j$ and time $t \geq 0$, $i \prec_{l(t)} j$ if and only if $\varphi_{i}(t) > \varphi_{j}(t)$. \end{lemma} \begin{proof} Suppose $S = AijB$ is an arbitrary schedule, where job $i$ starts at time $t$ and job $j$ is processed directly after job $i$. We use $F(S)$ to represent the cost of the schedule $S$. To explore the local dominance relation between job $i$ and job $j$, we need to check the impact of swapping the order of $i, j$, which leads to: \begin{align*} F(AijB) - F(AjiB) & = \frac{w_i}{p_i + t} + \frac{w_j}{p_j + p_i + t} - \frac{w_j}{p_j + t} - \frac{w_i}{p_i + p_j + t} \\ & = w_i \cdot \frac{p_j}{(p_i + t)(p_i + p_j + t)} - w_j \cdot \frac{p_i}{(p_j + t)(p_i + p_j + t)} \\ & = \frac{p_ip_j}{p_i + p_j + t}(\frac{w_i}{p_i(p_i + t)} - \frac{w_j}{p_j(p_j + t)}) \\ & = \frac{p_ip_j}{p_i + p_j + t}(\varphi_i(t) - \varphi_j(t)). \end{align*} Since $\frac{p_ip_j}{p_i + p_j + t} > 0$, the effect of exchanging jobs $i, j$ depends on the term $(\varphi_i(t) - \varphi_j(t))$, which completes the proof. \qed \end{proof} \subsubsection{Dominance rule} \label{ssub:modified_rule} Vásquez \cite{vasquez_for_2015} proved the following dominance properties of airplane refueling problem. \begin{theorem}[Vásquez] \label{theorem: vas} For all jobs $i, j$ and time points $a, b$, the dominance property $i \prec_{l[a, b]} j$ implies $i \prec_{g[a, b]} j$. \end{theorem} Based on Theorem \ref{theorem: vas} and Lemma \ref{lemma: local_new} we obtain the following dominance rule of airplane refueling problem. \begin{corollary} \label{rule: new_rule} For any two jobs $i, j$ with $w_i > w_j$: \begin{enumerate} \item If $\varphi_i(t) \geq \varphi_j(t)$ for $t \in [0, \infty)$, $i \prec_g j$; \item If $\varphi_j(t) \geq \varphi_i(t)$ for $t \in [0, \infty)$, $j \prec_g i$; \item else $\exists$ $t^*_{ij} = \frac{w_jp_i^2 - w_ip_j^2}{w_i p_j - w_j p_i} > 0$ with: \begin{itemize} \item $\varphi_i(t) > \varphi_j(t)$ for $t \in [0, t^*_{ij})$, $i \prec_{g[0, t^*_{ij})} j$; \item $\varphi_j(t) \geq \varphi_i(t)$ for $t \in [t^*_{ij}, \infty)$, $j \prec_{g[t^*_{ij}, \infty)}i$. \end{itemize} \end{enumerate} \end{corollary} Actually, this rule is equivalent to the rule given in \cite{vasquez_for_2015}, whereas we use $\varphi$ to indicate the dominance relation between any jobs $i, j$. Figure \ref{fig: auxilary_function} gives an illustrative example of the scenario where $t^*_{ij} = \frac{w_jp_i^2 - w_ip_j^2}{w_i p_j - w_j p_i} > 0$. In this paper, we care about potential schedules with respect to Corollary \ref{rule: new_rule}. We refer to these schedules as \emph{potential schedules} for short. \begin{figure}[H] \centering \begin{tikzpicture}[domain=-2:4,yscale=1,samples=200,>=latex,thick] \draw[thick,->] (-0.5,0.5) -- (4,0.5) node[right] {$t$} \draw[thick,->] (0,0) -- (0,4) node[below left] {$\varphi(t)$} \draw (0,0.5) node[below left] {O} \coordinate (O) at (0,0); \draw[name path=i, domain=0:4, color=black] plot (\x, {(12)/(2*(2+ \x)}) node[right] {$y=\varphi_i(t)$}; \draw[name path=j, domain=0:4, color=black] plot (\x, {(162)/(9*(9 + \x)}) node[right] {$y=\varphi_j(t)$}; \draw [name intersections={of=i and j, by=x}, dashed, thin] (x) -- (1.5, 0.5) node[below] {$t^*_{ij}$}; \end{tikzpicture} \caption{Illustration of dominance property between job $i, j$. For $t \in [0, t^*_{ij})$, $\varphi_i(t) > \varphi_j(t)$, $i \prec_{g[0, t^*_{ij})} j$; for $t \in [t^*_{ij}, \infty)$, $\varphi_i(t) \leq \varphi_j(t)$, $j \prec_{g[t^*_{ij}, \infty)}i$.} \label{fig: auxilary_function} \end{figure} \section{Technical lemmas} \label{sec:technical_lemmas} In this section, we establish several important properties concerning potential schedules. \subsection{Relative order between two jobs} \label{sub:relative_order_between_two_jobs} To begin with, we show that the dominance rule makes it impossible for a job to start within some time intervals in a potential schedule. Let $T:= \sum_{j \in J} p_j$ be the total processing time, we have: \begin{lemma} \label{lemma: banned_interval} For two jobs $i, j \in J$ with $\varphi_i(0) > \varphi_j(0)$ and $t^*_{ij} \in (0, T)$, if in a complete schedule $S$ job $i$ starts in time interval $[t^*_{ij}, t^*_{ij} + p_j)$, then $S$ is not a potential schedule. \end{lemma} \begin{proof} By Corollary \ref{rule: new_rule} it follows that $i \prec_{g[0, t^*_{ij})} j$ and $j \prec_{g[t^*_{ij},\infty)} i$. If job $j$ precedes job $i$ in $S$, then $t_j < t_j + p_j \leq t_i$. Since $t_i \in [t^*_{ij}, t^*_{ij} + p_j)$, we have: \[ 0 \leq t_j \leq t_i - p_j < t^*_{ij}, \] which leads to a violation of the relation $i \prec_{g[0, t^*_{ij})} j$ . Otherwise job $j$ is processed after job $i$ in $S$. In a similar manner we can derive: \[ t^*_{ij} \leq t_i \leq t_j - p_i < \infty. \] Again relation $j \prec_{g[t^*_{ij}, \infty)} i$ is not satisfied. \qed \end{proof} According to Lemma \ref{lemma: banned_interval}, job $i$ starts either in $[0, t^*_{ij})$ or in $[t^*_{ij} + p_j, T - p_i)$ in a potential schedule. Next lemma shows that if we fix the start time of job $i$ to one of these intervals, the relative order of job $i, j$ in a potential schedule is already decided. \begin{lemma} \label{lemma: left_or_right} For two jobs $i, j \in J$ with $\varphi_i(0) > \varphi_j(0)$ and $t^*_{ij} \in (0, T)$, suppose $S$ is a schedule with $t_j \not \in [t^*_{ij}, t^*_{ij} + p_j)$. \begin{enumerate} \item If $t_i < t^*_{ij}$, job $i$ should be processed before job $j$, otherwise $S$ is not a potential schedule; \item If $t_i \geq t^*_{ij} + p_j$, job $i$ should be processed after job $j$, otherwise $S$ is not a potential schedule. \end{enumerate} \end{lemma} \begin{proof} According to Corollary \ref{rule: new_rule} we have $i \prec_{g[0, t^*_{ij})} j$ and $j \prec_{g[t^*_{ij}, \infty)} i$. For the first scenario, processing job $j$ before job $i$ in $S$ would imply $t_j + p_j < t_i$, which leads to the following inequalities: \[ 0 \leq t_j \leq t_i - p_j < t^*_{ij}. \] Thus we have reached a negation to $i \prec_{g[0, t^*_{ij})} j$. Similarly, for the second scenario if job $i$ precedes job $j$ we will have: \[ t^*_{ij} \leq t_i \leq t_j - p_i < T. \] Relation $j \prec_{g[t^*_{ij}, \infty)} i$ does not hold, which concludes the proof. \qed \end{proof} Conditioning on the start time of job $i$, Lemma \ref{lemma: banned_interval} and Lemma \ref{lemma: left_or_right} provide a complete characterization of the relative order between jobs $i, j$ in a potential schedule. See Figure \ref{fig: relative_order} for an overview of these relations. \begin{figure} \centering \begin{tikzpicture}[domain=-2:8,yscale=1,samples=200,>=latex,thick] \draw[thick,-] (0,0) -- (8,0) \draw (0, 0.05) -- (0, -0.05) node[below] {$0$}; \draw (3, 0.05) -- (3, -0.05) node[below] {$t^*_{ij}$}; \draw (5, 0.05) -- (5, -0.05) node[below] {$t^*_{ij} + p_j$}; \draw (8, 0.05) -- (8, -0.05) node[below] {$T$}; \draw[thick, decoration={brace, amplitude=5}, decorate] (0,0.1) -- (2.95,0.1); \draw[thick, decoration={brace, amplitude=5}, decorate] (3.05,0.1) -- (4.95,0.1); \draw[thick, decoration={brace, amplitude=5}, decorate] (5.05,0.1) -- (8,0.1); \draw (1.5, 0.2) node[above] {$i$ before $j$}; \draw (4, 0.2) node[above] {banned interval}; \draw (6.5, 0.2) node[above] {$j$ before $i$}; \end{tikzpicture} \caption{The relative order of job $i$ and job $j$ in a potential schedule conditioned on the start time of job $i$ with $t^*_{ij} \in (0, T)$ and $\varphi_i(0) > \varphi_j(0)$. When $t_i \in [0, t^*_{ij})$, job $i$ should precede job $j$; when $t_i \in [t^*_{ij}, t^*_{ij} + p_j)$, the resulting schedule will violate dominance property; when $t_i \in [t^*_{ij} + p_j, T)$, job $j$ should precede job $i$.} \label{fig: relative_order} \end{figure} \subsection{Possible start times in a potential schedule} \label{sub:possible_start_times_in_a_potential_schedule} In this subsection we consider a general \emph{sub-problem} scenario. Let $J' \subseteq J$ be a set of jobs to be processed consecutively at start time $t_o$, where $t_o \in [0, T - \sum_{j \in J'} p_j]$. For convenience we denote the completion time $t_o + \sum_{j \in J'} p_j$ by $t_e$. Notice that when $J' = J$, we will have $t_o = 0$ by definition, the setting above describes the original problem. We further denote the partial schedule of $J'$ by $S'$. If $S'$ does not violate any dominance rule, we call it as a \emph{partial potential schedule}. Now consider job $\alpha \in J'$ such that: \[ \alpha = \operatorname*{arg\,max}_i \varphi_i(t_o). \] When there are more than one job with the maximum $\varphi(t_o)$, the tie breaking rule is to choose the job with the maximum $\varphi(t_e)$. Our goal is to study the possible start times of job $\alpha$ in a partial potential schedule $S'$. We start with analyzing the precedence relation of job $\alpha$ with other jobs in a partial potential schedule. To that end, we divide time interval $[t_o, t_e]$ into consecutive subintervals with the time set: \[ C_{J'}^{\alpha} := \{c_{\alpha j} | c_{\alpha j} = t^*_{\alpha j} + p_j, c_{\alpha j} \in (t_o, t_e), j \in J'\setminus \{\alpha\} \} \cup \{t_o, t_e\}. \] We re-index the set $C_{J'}^{\alpha}$ according to their value rank, that is, if a time point ranks the $q^{\text{\tiny th}}$ in the set, it will be denoted by $c_q$. Besides, we use mapping $M_{J'}^{\alpha}: J' \mapsto \mathbb{Z^+}$ to maintain job information of the index. The mapping is defined as: \[ M_{J'}^{\alpha}(j) = \begin{cases} 1, & \text{if $j = \alpha$}; \\ \text{the rank of $c_{\alpha j}$ in $C_{J'}^\alpha$}, & \text{if $t^*_{\alpha j} \in (t_o, t_e)$}; \\ |J'| + 1, & \text{if $t^*_{\alpha j} \notin (t_o, t_e)$}. \end{cases} \] We consider the start time of job $\alpha$ in each subinterval $[c_q, c_{q + 1})$. Next lemma shows that once we fix $t_\alpha$ to a subinterval, the positions of all remaining jobs in a potential schedule relative to $\alpha$ are determined. \begin{lemma} \label{lemma: J_l_J_r} Suppose $S'$ is partial potential schedule that job $\alpha$ starts within time interval $[c_q, c_{q + 1})$, then all $j$ with $M_{J'}^{\alpha}(j) \leq q$ should proceed $\alpha$, and the rest jobs with $M_{J'}^{\alpha}(j) > q$ should come after $\alpha$. \end{lemma} \begin{proof} Suppose $j$ is an arbitrary job in the set $J'$ other than $\alpha$, by the definition of job $\alpha$ and Corollary \ref{rule: new_rule} we know either it is dominated by job $\alpha$ in time interval $[t_o, t_e]$, or there exists $t^*_{ij} \in (t_o, t_e)$ that $\alpha \prec_{g[t_o, t^*_{ij})} j$ and $j \prec_{g[t^*_{ij}, t_e)} \alpha$. For the first case, we always have $M_{J'}^{\alpha}(j) = |J'| + 1 > q$, job $j$ comes after job $\alpha$. While for the second case, Lemma \ref{lemma: banned_interval} and Lemma \ref{lemma: left_or_right} apply, we have: \begin{itemize} \item If $M_{J'}^{\alpha}(j) \leq q$, it implies that $t_\alpha \geq t^*_{\alpha j} + p_j$. By Lemma \ref{lemma: left_or_right} job $j$ should precedes job $\alpha$ in a potential schedule. \item Otherwise $M_{J'}^{\alpha}(j) < q$, it follows that $t_\alpha < t^*_{\alpha j} + p_j$. Since $S'$ is a partial potential schedule, Lemma \ref{lemma: banned_interval} rules out the possibility that $t_\alpha \in [t^*_{\alpha j}, t^*_{\alpha j} + p_j)$, we have $t_\alpha < t^*_{\alpha j}$. Again by Lemma \ref{lemma: left_or_right} job $j$ should come after job $\alpha$. \end{itemize} \qed \end{proof} At last, for the possible start times of job $\alpha$ in a partial potential schedule, we have: \begin{lemma} \label{lemma: at_most_start_times} For every time interval $[c_q, c_{q + 1})$, there is at most one possible start time for job $\alpha$ in $[c_q, c_{q + 1})$. Thus, there are at most $|J'|$ possible start times of job $\alpha$ in a partial potential schedule $S'$. \end{lemma} \begin{proof} Suppose that $t_{\alpha} \in [c_q, c_{q + 1})$, according to Lemma \ref{lemma: left_or_right}, the positions of all remaining jobs in $J' \setminus \{\alpha\}$ relative to $\alpha$ are determined. We denote the set of jobs before and after job $\alpha$ as $J_l, J_r$, respectively. The start time of job $\alpha$ is determined by $J_l$. More precisely, we have: $t_\alpha = t_o + \sum_{j \in J_l} p_j.$ While if $t_o + \sum_{j \in J_l} p_j \notin [c_q, c_{q + 1})$, we have reached a conflict, which means there is no partial potential schedule that job $\alpha$ starts in current subinterval. As a result, there is at most one possible start time for job $\alpha$ in each subinterval. \qed \end{proof} \section{Algorithm} \label{sec:algorithm} In this section we devise an exact algorithm for airplane refueling problem and analyze its running time. \subsection{An exact algorithm for air plane refueling problem} \label{sub:an_exact_algorithm_for_air_plane_refueling_problem} We start with some notations. For each job pair $i, j$ of Lemma \ref{lemma: banned_interval}, we name the time interval $b_{ij} = [t^*_{ij}, t^*_{ij} + p_j)$ as the \emph{banned interval} of job $i$ imposed by job $j$, and denote $B_i := \cup_{j \in J \setminus {i}} b_{ij}$ as the union of all banned intervals of job $i$. \begin{figure}[H] \centering \begin{tikzpicture}[domain=-2:8,yscale=1,samples=200,>=latex,thick] \draw[thick,-] (-1,0) -- (8,0); \draw (-1, 0.05) -- (-1, -0.05) node[below] {$t^*_{\alpha i}$}; \draw (0, 0.075) -- (0, -0.075) node[below] {$t_o$}; \draw (0, -0.5) node [below] {$(c_1)$}; \draw (2, 0.05) -- (2, -0.05) node[below] {$c_{\alpha j}$}; \draw (2, -0.5) node [below] {$(c_2)$}; \draw (4, 0.05) -- (4, -0.05) node[below] {$c_{\alpha k}$}; \draw (4, -0.5) node [below] {$(c_3)$}; \draw (6, 0.05) -- (6, -0.05) node[below] {$c_{\alpha l}$}; \draw (6, -0.5) node [below] {$(c_4)$}; \draw (8, 0.075) -- (8, -0.075) node[below] {$t_e$}; \draw (8, -0.5) node [below] {$(c_5)$}; \draw (0, 0.05) node [above] {$t^1_{\alpha}$}; \draw (4.5, 0) -- (4.5, 0.05) node [above] {$t^2_{\alpha}$}; \draw (5, 0) -- (5, 0.05) node [above] {$t^3_{\alpha}$}; \draw (7, 0) -- (7, 0.05) node [above] {$t^4_{\alpha}$}; \draw (0, -1.5) node [left] {$B_\alpha :$}; \draw[thick, -] (1,-1.5) -- (4,-1.5); \draw[dashed, thin] (1, -1.5) -- (1, 0); \draw[dashed, thin] (4, -1.5) -- (4, -1); \draw[dashed, thin] (4, -0.7) -- (4, -0.35); \draw[thick, -] (4.75, -1.5) -- (6, -1.5); \draw[dashed, thin] (4.75, -1.5) -- (4.75, 0); \draw[dashed, thin] (6, -1.5) -- (6, -1); \draw[dashed, thin] (6, -0.7) -- (6, -0.35); \draw (2.5, -1.5) node [below] {$b_{\alpha j} \cup b_{\alpha k}$}; \draw (5.5, -1.5) node [below] {$b_{\alpha l}$}; \end{tikzpicture} \caption{Possible start times of job $\alpha$.} \label{fig: pos_start_time} \end{figure} Given an instance of airplane refueling problem, we find the job $\alpha$ and try to start $\alpha$ in each interval induced by $C_{J'}^{\alpha}$. For an interval $[c_q, c_{q+1})$ and corresponding $J_l, J_r$, the possible start time of $\alpha$ will be $t_\alpha = \sum_{j \in J_l} p_j$. If $t_\alpha \in [c_q, c_{q+1})$ and $t_\alpha \notin B_\alpha$, then we have found a potential start time of job $\alpha$. Figure \ref{fig: pos_start_time} provides an illustrative example of the above procedure, where $J' = \{i, j, k, l, \alpha \}$. As shown in the figure, the set $C^{\alpha}_{J'}$ divided time interval $[t_o, t_e)$ into 4 subintervals. Job $i$ should always behind job $\alpha$ since $t^*_{\alpha i} < t_o$. For other jobs in $J'$, condition on the start time of job $\alpha$ we have: \begin{enumerate} \item $t_\alpha \in [c_1, c_2)$: Job $\alpha$ serves as the first job and $t_\alpha^1 = t_o$. \item $t_\alpha \in [c_2, c_3)$: Job $j$ precedes job $\alpha$ in a potential schedule. As $t_\alpha^2 = t_o + p_j > c_3$, job $\alpha$ can not start in this interval. \item $t_\alpha \in [c_3, c_4)$: Job $j, k$ precedes job $\alpha$ in a potential schedule. As $t_\alpha^3 = t_o + p_j + p_k$ is in banned intervals of job $\alpha$, job $\alpha$ can not start in this interval. \item $t_\alpha \in [c_4, c_5)$: Job $j, k, l$ precedes job $\alpha$, $t_\alpha^4 = t_o + p_j + p_k + p_l$. \end{enumerate} Once we find a possible start time of job $\alpha$, we can solve $J_l$ and $J_r$ recursively with the procedure above until the subproblem has only one job. See Algorithm \ref{algo: fast_schedule} for the pseudo code of our algorithm, where the initial inputs are the complete job set $J$ and original start time $t=0$. \begin{algorithm}[H] \caption{Fast Schedule} \label{algo: fast_schedule} \begin{algorithmic}[1] \Require{Job set $J$ and start time $t$.} \Ensure{The optimal schedule and the optimal cost.} \Function{FastSchedule}{$J, t$} \If{$J$ is empty} \State \Return{$0$, $[]$} \ElsIf{$J$ contains only one job $j$} \State \Return{$\frac{w_j}{p_j + t}$, $[j]$} \EndIf \State Find job $\alpha$ \State $J_l, J_r, opt \gets [], J, 0$ \For{$c_{q} \in C_J^\alpha$} \State Find job $i$ with $M_{J}^{\alpha}(i) = q$ \Comment{$i = \alpha$ for $q=1$, i.e. job $\alpha$ is the first job} \State $J_l \gets J_l \cup \{i\}$ \State $J_r \gets J_r \setminus \{i\}$ \If{$c_{q} \leq t + \sum_{i \in J_l} p_i <{c}_{q+1} \And \ (t + \sum_{i \in J_l} p_i) \notin B_\alpha$} \label{line: new_branch} \Comment{new branch} \State $opt_l, seq_l \gets $ \Call{FastSchedule}{$J_l \setminus \{\alpha\}, t$} \State $opt_r, seq_r \gets $ \Call{FastSchedule}{$J_r, t + \sum_{i \in J_l} p_i + p_\alpha$} \If{$opt_l + opt_r + \frac{w_\alpha}{p_\alpha + \sum_{i \in J_l} p_i} > opt$} \State $opt, seq \gets (opt_l + opt_r + \frac{w_\alpha}{p_\alpha + \sum_{i \in J_l} p_i})$, $[seq_l, \alpha, seq_r]$ \EndIf \EndIf \EndFor \State \Return{opt, seq} \EndFunction \end{algorithmic} \end{algorithm} Next, we prove the correctness of Algorithm \ref{algo: fast_schedule}. \begin{theorem} Algorithm \ref{algo: fast_schedule} returns the optimal solution of airplane refueling problem. \end{theorem} \begin{proof} We only need to show that Algorithm \ref{algo: fast_schedule} does not eliminate any potential schedules. While locating possible start times of job $\alpha$, our algorithm drops out schedules with $t_\alpha$ in banned intervals or $t_\alpha$ exceeds current interval. According to Lemma \ref{lemma: banned_interval} and Lemma \ref{lemma: at_most_start_times}, all the excluded schedules violate Corollary \ref{rule: new_rule}. Each recursive call also drops out schedules that have jobs in $J_l$ processed after job $\alpha$ or jobs in $J_r$ processed before job $\alpha$. By Lemma \ref{lemma: left_or_right}, these schedules are not potential schedules either. \qed \end{proof} \subsection{Running time of Algorithm \ref{algo: fast_schedule}} \label{sub:running_time_of_algorithm_algo: fast_schedule} In this section we analyze the running time of Algorithm \ref{algo: fast_schedule} with respect to the number of potential schedules. Our main result can be stated as follows: \begin{theorem} \label{theorem: num_sol} Algorithm \ref{algo: fast_schedule} finds the optimal solution of airplane refueling problem in $\bigO(n^2(\log n + K))$ time, where $K$ is the number of potential schedules with respect to Corollary \ref{rule: new_rule}. \end{theorem} Before proving this theorem, we need to establish some properties of Algorithm \ref{algo: fast_schedule}. We start by showing that for any job set $J'$ that starts at time $t$, there is at least one partial potential schedule. \begin{lemma} \label{lemma: greedy} Suppose $J' \subseteq J$ is a set of job that starts at time $t$, then there exists a procedure that can find one partial potential schedule for $J'$. \end{lemma} \begin{proof} Consider the following procedure that construct the schedule successively. At each stage, choose the job with the maximum $\varphi(C)$ from the unscheduled jobs as the next job, where $C$ is the total processing time of the scheduled jobs plus $t$. We claim this procedure returns a partial potential schedule of $J'$. Let $S'$ be the schedule returned by the procedure above. Assume by contradiction that there are two jobs $i,j $ that violate dominance properties. Suppose $i$ precedes $j$ in $S'$, then we have $j \prec_{g[t_i, t_j - p_i)} i$, which implies $\varphi_j(t_i) > \varphi_i(t_i)$ by Lemma \ref{lemma: local_new}. In that case, we should have chosen job $j$ at time $t_i$ instead of $i$, absurdity is obtained. \qed \end{proof} While Lemma \ref{lemma: greedy} concerns about a single job set, the following lemma describes the relationship between two job sets $J_l$ and $J_r$. \begin{lemma} \label{lemma: independent} Given start time $t_{\alpha} \in [c_{q}, c_{q + 1})$ of job $\alpha$ with $t_\alpha \in [t_o, t_e-p_\alpha]$ and $t_\alpha \notin B_\alpha$, for any two jobs $j \in J_l$ and $k \in J_r$ with $M_{J'}^{\alpha}(j) \leq q < M_{J'}^{\alpha}(k)$, if job $j$ is scheduled before job $\alpha$ and job $k$ is scheduled after job $\alpha$, then no matter where job $j$ and job $k$ start, dominance properties among $\alpha,j$ and $k$ will not be violated. \end{lemma} \begin{proof} The dominance relations between jobs $\alpha,j$ and jobs $\alpha,k$ are satisfied, we only need to consider the relation between job $j$ and job $k$. From the definition of $c_q$, it follows that \[ t^*_{\alpha j} + p_j \leq t_{\alpha} < t^*_{\alpha k} + p_k. \] Since $t_\alpha$ is not in banned interval $b_{\alpha k} = [t^*_{\alpha k}, t^*_{\alpha k} + p_k)$, the equation above leads to: \[ t^*_{\alpha j} < t_\alpha < t^*_{\alpha k}. \] We distinguish the cases whether job $j$ and job $k$ have global precedence in time interval $[t_o, t_e]$. \begin{case}[Global dominance]\\ Job $j$ and job $k$ have global precedence in time interval $[t_o, t_e]$. \begin{enumerate} \item $j \prec_{g[t_o, t_e]} k$: Since job $j$ is scheduled before job $k$, precedence rule is satisfied. \item $k \prec_{g[t_o, t_e]} j$: We will show this scenario is impossible. At time $t_\alpha$, since $t^*_{\alpha j} < t_\alpha < t^*_{\alpha k}$ we have $j \prec_{l(t_\alpha)} \alpha$ and $\alpha \prec_{l(t_\alpha)} k$, which implies $j \prec_{l(t_\alpha)} k$. Thus we have constructed a contradiction to global dominance $j \prec_{g[t_o, t_e]} k$. \end{enumerate} \end{case} \begin{case}[Partial dominance] \\ Job $j$ and job $k$ have no global precedence in time interval $[t_o, t_e]$. In this case there exists $t^*_{jk} \in (t_o, t_e)$. \begin{enumerate} \item $t^*_{jk} < t^*_{\alpha j}$: On one hand by the definition of $\alpha$ we have $\varphi_k(t_o) > \varphi_j(t_o)$. On the other hand for the start time of job $k$, $t_k > t_\alpha > t^*_{\alpha j} + p_j > t^*_{jk} + p_j$. According to Lemma \ref{lemma: left_or_right} job $j$ precedes job $k$ does not violate the dominance properties between $i, j$. Figure \ref{fig: case_1} depicts the auxiliary functions of job $\alpha, j, k$ of this scenario. \item $t^*_{jk} = t^*_{\alpha j}$: We claim this scenario is impossible. If $t^*_{jk} = t^*_{\alpha j}$ we will have $t^*_{jk} = t^*_{\alpha j} = t^*_{\alpha k}$. By the definition of $i, j$ this relation will lead to $t^*_{\alpha k} < t^*_{\alpha j} + p_j < t_{\alpha} < t^*_{\alpha k} + p_k$, that is, $t_\alpha \in [t^*_{\alpha k}, t^*_{\alpha k} + p_k)$. Since $t_\alpha \notin B_\alpha$ and $[t^*_{\alpha k}, t^*_{\alpha k} + p_k) \subseteq B_\alpha$, we have reached a contradiction. See Figure \ref{fig: case_3} for the auxiliary functions of job $\alpha, j, k$ of this scenario \item $t^*_{jk} > t^*_{\alpha j}$: At first we show $t^*_{jk} > t^*_{\alpha k}$. Assume by contradiction that $t^*_{\alpha j} < t^*_{jk} < t^*_{\alpha k}$, which would imply $\alpha \prec_{g(t^*_{\alpha j}, t_e)} j$ and $\alpha \prec_{g(t_o, t^*_{\alpha k})} k$. Then for any time $t \in (t^*_{\alpha j}, t^*_{\alpha k})$ we have the relation $j \prec_{l(t^*_{\alpha j})} \alpha \prec_{l(t^*_{\alpha j})} k$. However, this is impossible since $t^*_{jk} \in (t^*_{\alpha j}, t^*_{\alpha k})$. Therefore, we have $t^*_{jk} > t^*_{\alpha k}$, and job $j$ starts before time $t^*_{kj}$. In that case, having job $j$ precedes job $k$ does not violate the dominance properties between $i, j$, our claim holds. See Figure \ref{fig: case_2} for this scenario. \end{enumerate} \end{case} \qed \end{proof} \begin{figure} \begin{minipage}[b]{.3\textwidth} \centering \begin{tikzpicture}[domain=-2:5,yscale=.7, xscale=.7, samples=200,>=latex,thick] \draw[thick,->] (-0.5,0.5) -- (3.7,0.5) node[right] {\scriptsize $t$} \draw[thick,->] (0,0) -- (0,4.7) node[left] {\scriptsize $\varphi(t)$} \draw (0,0.5) node[below left] {\scriptsize $t_o$} \coordinate (O) at (0,0); \draw[name path=job_alpha, domain=0:3.7, color=black] plot (\x, {(3.95)/(0.98*(0.98 + \x)}); \draw (-0.1, 4) node[left] {\scriptsize $\varphi_\alpha(t)$}; \draw[name path=job_j, domain=-0:3.7, color=black] plot (\x, {(15)/(2.5*(2.5+ \x)}); \draw (-0.1, 2.5) node[left] {\scriptsize $\varphi_k(t)$}; \draw[name path=job_k, domain=-0:3.7, color=black] plot (\x, {(160)/(9*(9 + \x)}); \draw (-0.1, 1.95) node[left] {\scriptsize $\varphi_j(t)$}; \draw [dashed, thin] (1.3714309278350516, 1.7141104155710498) -- (1.3714309278350516, 0.5) node[below] {\scriptsize $t^*_{\alpha j}$}; \draw [dashed, thin] (2.1308808290155445, 1.2956498388829214) -- (2.1308808290155445, 0.5) node[below] {\scriptsize $t^*_{\alpha k}$}; \draw [dashed, thin] (0.8113207547169812, 1.8119658119658117) -- (0.8113207547169812, 0.5) node[below] {\scriptsize $t^*_{j k}$}; \end{tikzpicture} \caption{$t^*_{jk} < t^*_{\alpha j}$.} \label{fig: case_1} \end{minipage} \begin{minipage}[b]{.3\textwidth} \centering \begin{tikzpicture}[domain=2:5,yscale=.7, xscale=.7, samples=200,>=latex,thick] \draw[thick,->] (-0.5,0.5) -- (3.7,0.5) node[right] {\scriptsize \scriptsize$t$} \draw[thick,->] (0,0) -- (0,4.7) node[left] {\scriptsize $\varphi(t)$} \draw (0 ,0.5) node[below left] {\scriptsize $t_o$} \coordinate (O) at (0,0); \draw[name path=job_alpha, domain=0:3.7, color=black] plot (\x, {(3.95)/(0.98*(0.98 + \x)}); \draw (-0.1, 4) node[left] {\scriptsize $\varphi_\alpha(t)$}; \draw[name path=job_j, domain=0:3.7, color=black] plot (\x, {(15)/(2.5*(2.5+ \x) - 1.05}); \draw (-0.1, 2.95) node[left] {\scriptsize $\varphi_k(t)$}; \draw[name path=job_k, domain=0:3.7, color=black] plot (\x, {(160)/(9*(9 + \x)}); \draw (-0.1, 2) node[left] {\scriptsize $\varphi_j(t)$}; \draw [dashed, thin] (1.3714309278350516, 1.7141104155710498) -- (1.3714309278350516, 0.5) node[below] {\scriptsize $t^*_{\alpha j}$}; \end{tikzpicture} \caption{$t^*_{jk} = t^*_{\alpha j}$.} \label{fig: case_3} \end{minipage} \begin{minipage}[b]{.3\textwidth} \centering \begin{tikzpicture}[domain=-2:5,yscale=.7, xscale=.7, samples=200,>=latex,thick] \draw[thick,->] (-0.5,0.5) -- (3.7,0.5) node[right] {\scriptsize$t$} \draw[thick,->] (0,0) -- (0,4.7) node[left] {\scriptsize$\varphi(t)$} \draw (0,0.5) node[below left] {\scriptsize$t_o$} \coordinate (O) at (0,0); \draw[name path=job_alpha, domain=0:3.7, color=black] plot (\x, {(5.5)/(1.2*(1.2 + \x)}); \draw (-0.1, 1.85) node[left] {\scriptsize$\varphi_k(t)$}; \draw[name path=job_j, domain=0:3.7, color=black] plot (\x, {(15)/(2.2*(2.2+ \x)}); \draw (-0.1, 3.82) node[left] {\scriptsize$\varphi_\alpha(t)$}; \draw[name path=job_k, domain=0:3.7, color=black] plot (\x, {(150)/(9*(9 + \x)}); \draw (-0.2, 3.1) node[left] {\scriptsize$\varphi_j(t)$}; \draw [dashed, thin] (0.85084745762712, 2.234848484848484) -- (0.85084745762712, 0.5) node[below] {\scriptsize$t^*_{\alpha j}$}; \draw [dashed, thin] (1.7586206896551724, 1.5491452991452992) -- (1.7586206896551724, 0.5) node[below] {\scriptsize$t^*_{\alpha k}$}; \draw [dashed, thin] (2.507692307692307, 1.448306595365419) -- (2.507692307692307, 0.5) node[below] {\scriptsize$t^*_{j k}$}; \end{tikzpicture} \caption{$t^*_{jk} > t^*_{\alpha j}$.} \label{fig: case_2} \end{minipage} \end{figure} Combining Lemma \ref{lemma: greedy} and Lemma \ref{lemma: independent}, we can identify a strong connection between Algorithm \ref{algo: fast_schedule} and potential schedules. \begin{lemma} \label{lemma: least_one} Whenever Algorithm \ref{algo: fast_schedule} adds a new branch for an instance $(J, t)$, there is at least one potential schedule on that branch. \end{lemma} \begin{proof} Suppose after we find a possible start time of job $\alpha$ the remaining jobs are divided into $J_l$ and $J_r$. By Lemma \ref{lemma: greedy} both job sets have at least one partial potential schedule. Denote the partial schedule of $J_l$ by $S_l$ and the partial schedule of $J_r$ by $S_r$, then according to Lemma \ref{lemma: independent} the jointed schedule $[S_l, \alpha, S_r]$ is also a potential schedule. \qed \end{proof} Now we are ready to prove Theorem \ref{theorem: num_sol}. \begin{proof}[Proof of Theorem \ref{theorem: num_sol}] We only need to generate all $t^*_{ij}$ and sorted them once, which takes $\bigO(n^2 \log n)$ time. While finding the possible start times of $\alpha$, with the already calculated $t^*_{ij}$, we can construct the set $C_J^\alpha$ and the mapping $M_{J}^{\alpha}$ in $\bigO(n)$ time. The iteration over each subinterval $[c_q, c_{q + 1})$ also needs at most $\bigO(n)$ time. Therefore, it takes at most $\bigO(n)$ time to find a potential start time of a job. By Lemma \ref{lemma: least_one} there exists at least one potential schedule with job $\alpha$ starting at that time. For any potential schedule there are $n$ start times to be decided, as a consequence, Algorithm \ref{algo: fast_schedule} finds each potential schedule in at most $\bigO(n^2)$ time. To find all the potential schedules it will take $\bigO(n^2 (\log n + K))$ time. \qed \end{proof} \section{Experimental Study} \label{sec:experimental_study} We code our algorithm with Python 3 and perform the experiment on a Linux machine with one Intel Core i7-9700K @ 3.6Ghz $\times$ 8 processor and 16Gb RAM. Notice that our implementation only invokes one core at a time. For experimental data set we adopt the method introduced by Höhn and Jacobs \cite{bader_experimental_2012} to generate random instances. For an instance with $n$ jobs, the processing time $p_i$ of job $i$ is an integer generated from uniform distribution ranging from 1 to 100, whereas the priority weight $w_i = 2^{N(0, \sigma^2)} \cdot p_i$ with $N$ being normal distribution. Therefore, a random instance is characterized by two parameters $(n, \sigma)$. According to previous results \cite{bader_experimental_2012,vasquez_for_2015}, instances generated with smaller $\sigma$ are more likely to be harder to solve, which means we can roughly tune the hardness of instances by adjusting the value $\sigma$. At first we compare our algorithm with the $A^*$ algorithm given by Vásquez \cite{vasquez_for_2015}. This algorithm modeles airplane refueling problem as a shortest path problem on a directed acyclic graph $G$. The vertexes of $G$ consist of all subset of $J$ and arcs are linked according to Corollary \ref{rule: new_rule}， and a path from vertex $\{\}$ to $J$ corresponds to a schedule. Since there will be $2^n$ vertexes in $G$ and it takes $\bigO(n)$ time to process each vertex in the worst case, the complexity of this algorithm is $\bigO(n \cdot 2^n)$. This part of experiment is conducted on data set $S_1$, which is generated with $\sigma=0.1$ and job size ranges from $\{10, 20, \dots, 140\}$, where for each configuration there are 50 instances. Secondly, we evaluate the empirical performance of Algorithm \ref{algo: fast_schedule} on data set $S_2$. In detail, for each job size $n$ in $\{100, 500, 1000, 2000, 3000 \}$ and for each $\sigma$ value $\{0.1, 0.101, 0.102, \dots, 1 \}$, there are 5 instances, that is, $S_2$ has 4505 instances in total. At last, we generate data set $S_3$ to examine the relations of instance hardness with the number of potential schedules and the value of $\sigma$. This data set has 5 instances of 500 jobs for each $\sigma$ from $\{0.100, 0.101, 0.102, \dots, 1\}$. \subsubsection{Comparison with $A^*$} \label{ssub:comparison_with_} We consider the ratio between the average running time of $A^*$ and Algorithm \ref{algo: fast_schedule} on data set $S_1$. As shown in Figure \ref{fig: compare}, our algorithm outperforms $A^*$ on all sizes and the speed up is more significant on instances with larger size. For instances with 140 jobs, our algorithm is over 100 times faster. \begin{figure}[H] \centering \includegraphics[width=.6\textwidth]{compare_nodes.eps} \caption{Speed up factor on dependence of instance size, on data set $S_1$.} \label{fig: compare} \end{figure} \subsubsection{Empirical performance} \label{ssub:empirical_performance} Table \ref{tab: running_time} presents the running time of Algorithm \ref{algo: fast_schedule} on data set $S_2$. We set a 10000 seconds timeout while performing the experiment. When our algorithm does not solve all the instances within this time bound, we present the percentage of the solved instances. For all instances with less than 2000 jobs as well as most instances with 2000 and 3000 jobs, our algorithm returns the optimal solution within the time bound. While for hard instances of 2000 and 3000 jobs generated with $\sigma$ less than $0.2$, our algorithm can solve $92\%$ and $79\%$ of the instances within the time bound respectively. \begin{table}[H] \centering \caption{Running time on data set $S_2$.} \label{tab: running_time} \scriptsize \begin{tabularx}{\textwidth}{@{}p{.1\textwidth}<{\centering}YYYYYYYYYY@{}} \toprule \multirow{2}{*}{$|J|$} & \multicolumn{2}{c}{100} & \multicolumn{2}{c}{500} & \multicolumn{2}{c}{1000} & \multicolumn{2}{c}{2000} & \multicolumn{2}{c}{3000} \\ \cmidrule(lr){2-3} \cmidrule(lr){4-5} \cmidrule(lr){6-7} \cmidrule(lr){8-9} \cmidrule(lr){10-11} & $\mathrm{avg.}$ & $\mathrm{std.}$ & $\mathrm{avg.}$ & $\mathrm{std.}$ & $\mathrm{avg.}$ & $\mathrm{std.}$ & $\mathrm{avg.}$ & $\mathrm{std.}$ & $\mathrm{avg.}$ & $\mathrm{std.}$ \\ \midrule $[0.1, 0.2)$ & 0.37 & 0.16 & 62.26 & 128 & 645.4 & 2223 & 92.40\% & - & 79.60\% & - \\ $[0.2, 0.3)$ & 0.30 & 0.05 & 13.38 & 8.5 & 65.08 & 43.23 & 328.6 & 210.2 & 955.4 & 871.5 \\ $[0.3, 0.4)$ & 0.27 & 0.03 & 7.11 & 9.27 & 41.91 & 12.46 & 222.7 & 128.0 & 538.6 & 339.4 \\ $[0.4, 0.5)$ & 0.26 & 0.02 & 8.16 & 0.91 & 35.74 & 5.00 & 172.8 & 38.70 & 421.5 & 87.15 \\ $[0.5, 0.6)$ & 0.25 & 0.01 & 7.64 & 0.59 & 33.27 & 3.24 & 159.9 & 28.54 & 392.0 & 53.67 \\ $[0.6, 0.7)$ & 0.25 & 0.01 & 7.37 & 0.38 & 31.84 & 2.74 & 151.0 & 14.37 & 367.3 & 41.37 \\ $[0.7, 0.8)$ & 0.25 & 0.01 & 7.18 & 0.35 & 30.52 & 1.52 & 142.1 & 12.97 & 349.4 & 22.09 \\ $[0.8, 0.9)$ & 0.24 & 0.01 & 7.07 & 0.32 & 30.10 & 1.62 & 144.8 & 9.07 & 341.4 & 22.79 \\ $[0.9, 1.0]$ & 0.24 & 0.01 & 6.94 & 0.23 & 29.28 & 1.00 & 141.5 & 7.41 & 327.3 & 16.93 \\ \bottomrule \end{tabularx} \end{table} \subsubsection{Instance hardness} \label{ssub:instance_hardness} Figure \ref{fig: hardness} depicts relations of different hardness indicators. The chart on the left shows the number of potential schedules and the running time of Algorithm \ref{algo: fast_schedule} on data set $S_3$. Since both variables cover a large range, we present the figure as a log-log plot. As indicated by Theorem \ref{theorem: num_sol}, there is a strong correlation between the number of potential schedules and the running time of Algorithm \ref{algo: fast_schedule}. Therefore, the number of potential schedules can serve as a rough measure of the instance hardness. For the relation between the number of potential schedules and the $\sigma$ value, the second chart exhibits that instances generated with small $\sigma$ are more likely to have a large number of potential schedules. The relation above is more obvious on smaller $\sigma$, while on larger $\sigma$ the difference between the number of potential schedules is less significant. \begin{figure}[H] \begin{minipage}[b]{.5\textwidth} \centering \includegraphics[width=\textwidth]{sol_time.eps} \end{minipage} \begin{minipage}[b]{.5\textwidth} \centering \includegraphics[width=\textwidth]{sol_sigma.eps} \end{minipage} \caption{Different indicators of instance hardness, on data set $S_3$.} \label{fig: hardness} \end{figure} \section{Conclusions} \label{sec:conclusions} We devise an efficient exact algorithm for airplane refueling problem. Based on the dominance properties of the problem, we propose a method that can prefix some jobs' start times and determine the relative orders among jobs in a potential schedule. This technique enables us to solve airplane refueling problem in a recursive manner. Our algorithm outperforms the state of the art exact algorithm on random generated data sets. For large instances with hard configurations that can not be tackled by previous algorithms, our algorithm can solve most of them in a reasonable time. The empirical efficiency of our algorithm can be attributed to two factors. First, our algorithm explores only the branches that contain potential schedules. Second, on the root node of each branch, the problem is further divided into smaller subproblems, which can also speedup the searching process. Theoretically, we prove that the running time of our algorithm is upper bounded by the number of potential schedules times a polynomial overhead in the worst case. Another contribution of this work is that we give some new structural properties of airplane refueling problem, which may be helpful in understanding the computational complexity of the problem.

https://arxiv.org/abs/1709.03867

An Efficient Approximation Algorithm for the Steiner Tree Problem

The Steiner tree problem is one of the classic and most fundamental $\mathcal{NP}$-hard problems: given an arbitrary weighted graph, seek a minimum-cost tree spanning a given subset of the vertices (terminals). Byrka \emph{et al}. proposed a $1.3863+\epsilon$-approximation algorithm in which the linear program is solved at every iteration after contracting a component. Goemans \emph{et al}. shown that it is possible to achieve the same approximation guarantee while only solving hypergraphic LP relaxation once. However, optimizing hypergraphic LP relaxation exactly is strongly NP-hard. This article presents an efficient two-phase heuristic in greedy strategy that achieves an approximation ratio of $1.4295$.

\section{Introduction}\label{sec:Introduction} The \textit{Steiner tree} problem is one of the classic and most fundamental $\mathcal{NP}$-hard problems. Given an arbitrary weighted graph with a distinguished vertex subset, the Steiner tree problem asks for a shortest tree spanning the distinguished vertices. This problem is widely used in many fields, such as VLSI routing~\cite{Kahng95}, wireless communications~\cite{Min2006,Liang:2002}, transportation~\cite{Hwang92}, wirelength estimation~\cite{Caldwell98}, and network routing~\cite{Korte90}. The Steiner tree problem is $\mathcal{NP}$-hard even in the very special cases of Euclidean or rectilinear metrics~\cite{garey02}. In fact, it is $\mathcal{NP}$-hard to approximate the Steiner tree problem within a factor $96/95$~\cite{Chlebik08}. Hence, an approximation algorithm with a small and provable guarantee is thirsted by researchers. Recall that an $\alpha$-approximation algorithm for a minimization problem is a polynomial-time algorithm that finds approximate solutions to $\mathcal{NP}$-hard optimization problems with cost at most $\alpha$ times the optimum value. Arora~\cite{Arora98} established that Euclidean and rectilinear minimum-cost Steiner trees can be approximated in polynomial time within a factor of $1 + \epsilon$ for any constant $\epsilon>0$. For arbitrary weighted graphs, a sequence of improved approximation algorithms appeared in the literatures~\cite{Takahashi80, Zel93, Berman94, Zel96, Pro97, Karpinski97, Hougardy99, Robins05, Byrka13} and the best approximation ratio achievable within polynomial time was improved from 2 to 1.39. Byrka \emph{et al}. proposed an LP-based approximation algorithm that achieves approximation ratio of $\ln4+\epsilon$ for general graphs~\cite{Byrka13}. However, the linear program is solved at every iteration after contracting a component. Goemans \emph{et al}.~\cite{Goemans2012} shown that it is possible to achieve the same approximation guarantee while only solving hypergraphic LP relaxation once. However, optimizing hypergraphic LP relaxation exactly is strongly NP-hard~\cite{Goemans2012}. Borchers and Du~\cite{Borchers97} show that $\rho_k \leq 1+\left\lfloor \log_{2}k\right\rfloor^{-1}$ where $\rho_k$ is the worst-case ratio of the cost of optimal $k$-restricted Steiner tree to the cost of optimal Steiner tree. We may therefore choose $k=2^{1/\epsilon}$ appropriately to obtain a $1+\epsilon$ approximation to hypergraphic LP relaxation, for any $\epsilon>0$. The number of variables and constraints will consequently be more than $n^{2^{1/\epsilon}}$ where $n$ is the number of terminals~\cite{Feldmann2016}. \section{Notation and Preliminaries}\label{sec:Preliminaries} Given a graph $G=(V,E)$ with nonnegative edge costs (or weights) $cost:E\rightarrow \mathbb{R}^{+}$ and a subset $R \subseteq V$ of terminals of the vertices of $G$, the Steiner tree problem asks for a minimum-cost Steiner tree spanning $R$. Any tree in $G$ spanning $R$ is called a Steiner tree, and any non-terminal vertices contained in a Steiner tree are referred to as Steiner points. The cost of a tree is the sum of its edge costs. The graph $G$ is assumed to be a complete graph and let $G_R$ be a complete graph that induced by $R$. For any graph $H$, we denote by $MST(H)$ a minimum spanning tree of a graph $H$ and by $cost(H)$ the sum of the costs of all edges in $H$. We thus abbreviate $mst(H)=cost(MST(H))$, i.e., the cost of a minimum spanning tree of $H$. A \textit{terminal-spanning tree} is a Steiner tree that does not contain any Steiner points. Let $mst$ be the cost of minimum terminal-spanning tree $MST(G_R)$. A minimum-cost Steiner tree spanning subset $R'\subset R$ in which all terminals are leaves is called a \textit{full component}. Any Steiner tree can be decomposed into full components by splitting all the non-leaf terminals~\cite{Robins05}. Our algorithm will starts with a minimum-cost terminal spanning tree, and iteratively adds full components to improve it. Any full component is assumed to have its own copy of each Steiner point so that full components chosen by our algorithm do not share Steiner points. Let $\Gamma(K)$ be the terminal set of a given full component $K$. Let $E_{0}(R')$ be the set of zero-cost edges in which all edges connect all pairs of terminals in $R'$. For brevity, let $E_{0}(H)=E_{0}(\Gamma(H))$. We call a Steiner tree $S$ is a \textit{well solution} if $\left|\Gamma(K_{i})\cap \Gamma(K_{j})\right|\leq 1$ for any two full components $K_{i}$ and $K_{j}$ in $S$. Let $Loss(K)$ be the minimum-cost sub-forest of $K$. A simple method of computing $Loss(K)$ is given by the following lemma. \begin{lem}\label{lem:lem-2} {\rm\cite{Robins05}}. For any full component $K$, $Loss(K)=MST(K\cup E_0(K))-E_0(K)$. \end{lem} We denote the cost of $Loss(K)$ by $loss(K)$. Let $\mathcal{C}[K]$ be a loss-contracted full component that can be obtained by collapsing each connected component of $Loss(K)$ into a single node. We denote by $Opt_{k}$ an optimal $k$-restricted Steiner tree. Let $opt_{k}$ and $loss_{k}$ be the cost and loss of $Opt_{k}$, respectively. Let $opt$ be the cost of the optimal Steiner tree. For brevity, this article uses $T/ E_{0}(R')$ to denote the minimum spanning tree of $T\cup E_{0}(R')$ for $R'\subset R$. The \textit{gain} of a full component $K$ with respect to $T$ is defined as \begin{eqnarray*} gain_{T}(K) = cost(T)-mst(T\cup E_{0}(K))-cost(K), \end{eqnarray*} and the \textit{load} of of a full component $K$ with respect to $T$ is defined as \begin{eqnarray*} load_{T}(K) = cost(K)+mst(T\cup E_{0}(K))- cost(T). \end{eqnarray*} Let $\Psi_{T_{1},T_{2}}(K)=cost(T_{1})-cost(T_{2})-mst(T_{1}\cup E_{0}(K))+mst(T_{2}\cup E_{0}(K))$. The following lemma shows that if no full component can improve a terminal-spanning tree $T$, then $cost(T)\leq opt_{k}$. \begin{lem}\label{lem08} {\rm\cite{Robins05}}. Let $T$ be a terminal-spanning tree; if $gain_{T}(K)\leq 0$ for any k-restricted full component $K$, then $cost(T)\leq opt_{k}$. \end{lem} \section{Two-phase Algorithm}\label{sec:Algorithm3} This section proposes a $k$-restricted two-phase heuristic ($k$-TPH) which is described in \textbf{Algorithm~\ref{Alg1}}. Let $T^t$ be the terminal-spanning tree at the end of iteration $t$ and let $K_t$ be the chosen full component at the end of iteration $t$. The first phase finds a terminal-spanning tree $T_{base}$ such that no full component can improve it. The terminal-spanning tree $T_{base}$ is a based criterion for the second phase. We denote by $S_{1}$ the solution in the first phase, and by $S_{2}$ the solution in the second phase. The first phase is a loss-contracting algorithm. The criterion function of $K$ with respect to $T^{t-1}$ is defined as \begin{eqnarray*} r=\frac{gain_{T^{t-1}}(K)}{loss(K)}. \end{eqnarray*} A chosen full component $K_i$ may be modified by other chosen full component. Assume that some edges $\left\{e_1,e_2,\ldots\right\}$ in $T^{t-1}$ that are corresponding to $\mathcal{C}[K_i]$ are deleted when adding $\mathcal{C}[K_t]$ to $T^{t-1}$. Some components are obtained by $K_i-\left\{e_1,e_2,\ldots\right\}$ and each component can be replaced by a full component with same terminals. The full component $K_i$ is replaced by these full components. That is because we want to ensure that $\frac{1}{2}\cdot cost(S_{1})\leq cost(T_{base})$. If no edge in $T^{t-1}$ is corresponding to $\mathcal{C}[K_i]$, we keep a \textit{basic component} from $K_i$ that is a Steiner point directly connect to two terminals in which an edge belongs to $Loss(K_i)$ and another edge belongs to $K_i-Loss(K_i)$ (see Figure~\ref{fig3}). It guarantees that the chosen full components never be chosen again. However, it may bring that some Steiner points are leaves in $S_{1}$. Fortunately, these Steiner points can be removed. Therefore, this paper assume that no Steiner point is leaf in $S_{1}$. \begin{figure}[t] \centering \includegraphics[width=3 in]{figure1.pdf} \caption{A full component $K$: squares denote terminals, circles denote Steiner and bold black edges indicate $K-Loss(K)$. A subgraph $B=\left(\left\{a,b,c\right\},\left\{\left\{a,c\right\},\left\{b,c\right\}\right\}\right)$ is a basic component in $K$ where an edge $\left\{a,c\right\}$ belongs to $Loss(K)$ and another edge $\left\{b,c\right\}$ belongs to $K-Loss(K)$.} \label{fig3} \end{figure} The second phase calls the $k$-restricted enhanced relative greedy heuristic ($k$-ERGH), which is described in \textbf{Algorithm~\ref{Alg2}}, to obtain a Steiner tree $S_{2}$. The $k$-ERGH iteratively finds a full component $K$ for modifying the terminal-spanning trees $T_{origin}^{0}=MST(G_{R})$ and $T_{base}^{0}$. When a full component $K_{t}$ has been chosen, the algorithm contracts the cost of the corresponding edges in $T_{origin}^{t-1}$ to zero, that is, $T_{origin}^{t}=MST(T_{origin}^{t-1}\cup E_{0}(K_{t}))$. Similarly, $T_{base}^{t}=MST(T_{base}^{t-1}\cup E_{0}(K_{t}))$. The criterion function of $K$ with respect to $T_{origin}^{t-1}$ and $T_{base}^{t-1}$ is defined as \begin{eqnarray*} f(K)=\frac{load_{T_{base}^{t-1}}(K)}{\Psi_{T_{origin}^{t-1},T_{base}^{t-1}}(K)}. \end{eqnarray*} The following steps analyze the complexity of $k$-TPH. Recall that, $n$ is the number of terminals. In the first phase, the number of iterations cannot exceed the number of full Steiner components $O(n^k)$. The gain of a full component $K$ can be found in time $O(k)$ after precomputing the longest edges between any pair of nodes in the current minimum spanning tree, which may be accomplished in time $O(n \log n)$~\cite{Robins05}. Thus, the runtime of all the iterations in the first phase can be bounded by $O(k n^{2k+1}\log n)$. We also can obtain the runtime of all the iterations in the second phase is bounded by $O(k n^{2k+1}\log n)$. Thus, the total runtime is $O(k n^{2k+1}\log n)$. \begin{algorithm} \caption{The $k$-restricted two-phase heuristic ($k$-TPH)} \begin{algorithmic}[1] \STATE \textbf{--------------------The first phase--------------------} \STATE $T^{0}=MST(G_{S})$ \FOR{$t=1,2,\ldots$} \STATE Find a $k$-restricted full component $K_{t}=K$ with maximizes \[ r=\frac{gain_{T^{t-1}}(K)}{loss(K)} \] \IF{$r\leq 0$} \STATE $T_{base} = T^{t-1}$ and exit for-loop \ENDIF \IF{there exist some edges $\left\{e_1,e_2,\ldots\right\}\subseteq T^{t-1}-MST(T^{t-1}\cup E_0(K_t))$ and $\left\{e_1,e_2,\ldots\right\}\subseteq \mathcal{C}(K_i)$ for $i\neq t$} \STATE Some components are obtained by $K_i-\left\{e_1,e_2,\ldots\right\}$ and each components can be replaced by a full component with same terminals. \STATE Replaced the full component $K_i$ by these full components. \STATE (for convenient to describe algorithm, we reuse the notain $K_i$ to represent these full components.) \ENDIF \STATE $T^{t}=MST(T^{0}\cup \mathcal{C}[K_1]\cup \cdots \mathcal{C}[K_t])$ \STATE $S_{1}=MST(T^{0}\cup K_1\cup \cdots \cup K_{t})$ \IF{no edge in $T^{t}$ is corresponding to $\mathcal{C}[K_i]$ for $i\neq t$} \STATE Keep a basic component from $K_i$. \STATE (we also reuse the notain $K_i$ to represent this basic component.) \ENDIF \ENDFOR \STATE \textbf{--------------------The second phase--------------------} \STATE $S_{2} = k\mbox{-ERGH}(T_{base})$ \RETURN the minimum-cost tree $S$ between $S_{1}$ and $S_{2}$. \end{algorithmic} \label{Alg1} \end{algorithm} \begin{algorithm} \caption{The $k$-restricted enhanced relative greedy heuristic ($k$-ERGH)} \begin{algorithmic}[1] \REQUIRE $T_{base}$. \STATE $T_{base}^{0}=T_{base}$ and $T_{origin}^{0}=MST(G_{S})$ \FOR{$t=1,2,\ldots$} \STATE Find a $k$-restricted full component $K_{t}=K$ which minimizes \[ f(K)=\frac{load_{T_{base}^{t-1}}(K)}{\Psi_{T_{origin}^{t-1},T_{base}^{t-1}}(K)} \] \STATE $T_{origin}^{t}=MST(T_{origin}^{t-1}\cup E_{0}(K_{t}))$ \STATE $T_{base}^{t}=MST(T_{base}^{t-1}\cup E_{0}(K_{t}))$ \IF{$c(T_{origin}^{t}) = c(T_{base}^{t})$} \RETURN $MST(T_{origin}^{0}\cup K_1 \cup K_2 \cdots \cup K_t)$ \ENDIF \ENDFOR \end{algorithmic} \label{Alg2} \end{algorithm} \section{Approximation ratio of the $k$-TPH} This section shows the approximation result of the $k$-TPH. When a full component $K$ has been chosen, the following lemma shows that the first phase never choose the full component $K$ even it has been replaced by some full components. \begin{lem}\label{lem09} The first phase never choose the chosen full components again. \end{lem} \begin{proof} Assume that the first phase choose a full component $K_t=K$. If $MST(T\bigcup_{i=1}^{t'}\mathcal{C}[K_i])$ contain all edges $e\in \mathcal{C}[K]$ in the iteration $t'>t$, $gain_{T}(K)\leq 0$ and the first phase never choose the full component $K$ again. If $MST(T\bigcup_{i=1}^{t'}\mathcal{C}[K_i])$ does not contain some edge $e\in \mathcal{C}[K]$ in the iteration $t'>t$, the edge $e$ has been improved by a chosen full component. The full component $K$ is divided into two components by removing the edge $e$. Let $A$ and $B$ be two connected components of $K-\left\{e\right\}$. The full component $K_t$ is replaced by two full components $K_A$ and $K_B$ with terminals sets $\Gamma(A)$ and $\Gamma(B)$, respectively. We have $T^{t'}=MST(T\bigcup_{i=1}^{t-1}\mathcal{C}[K_i]\cup K_A \cup K_B \bigcup_{i=t+1}^{t'}C[K^i])$, $gain_{T^{t'}}(K)\leq gain_{T^{t'}}(A\cup B)\leq gain_{T^{t'}}(A)+gain_{T^{t'}}(B)$ and $loss(K)=loss(A)+loss(B)$. Finally, \begin{eqnarray*} \frac{gain_{T^{t'}}(K)}{loss(K)} & \leq & \frac{gain_{T^{t'}}(A))+gain_{T^{t'}}(B)}{loss(B)+loss(B)} \\ & \leq & \max\left\{\frac{gain_{T^{t'}}(A)}{loss(A)},\frac{gain_{T^{t'}}(B)}{loss(B)}\right\}. \end{eqnarray*} We show that $K_A$ never be replaced by $A$. We knows that $cost(K_A)\leq cost(A)$ and $gain_{T^{t'}}(K_A)\leq 0$. The full component $K_A$ is superior to $A$. We also can obtain that $K_B$ never be replaced by $B$. The first phase never choose the full component $K$ again. If no edge in $T^{t'}$ is corresponding to $\mathcal{C}[K]$, we keep a basic component in $K$. Then, we can find a full component that superior to $K$. The chosen full components never be chosen again. \end{proof} \begin{lem}\label{lem01} $cost(T_{base}^{0} ) \geq \frac{1}{2}\cdot cost(S_{1})$. \end{lem} \begin{proof} The cost of the Steiner tree in the first phase is \begin{eqnarray*} cost(S_{1}) & = & cost(T_{base}^{0})+ \sum_{K_{j}\in S_{1}} loss(K_{j}). \end{eqnarray*} Since $loss(K) \leq \frac{1}{2}\cdot cost(K)$~\cite{Robins05} for any full component $K$, \begin{eqnarray*} cost(S_{1}) & \leq & cost(T_{base}^{0})+ \sum_{K_{j}\in S_{1}} \frac{1}{2}\cdot cost(K_{j}) \\ & \leq & cost(T_{base}^{0})+ \frac{1}{2}\cdot cost(S_{1}) \end{eqnarray*} which yields $cost(T_{base}^{0}) \geq \frac{1}{2}\cdot cost(S_{1})$. \end{proof} \begin{lem}\label{lem06} If no full component can improve the terminal-spanning tree $T$, \[ load_{T}\left(\bigcup_{i=1}^{n} K_{i}\right) \geq \sum_{i=1}^{n} load_{T}(K_{i}) \] for full components $K_1,K_2,\ldots,K_n$. \end{lem} \begin{proof} The proof can be obtained by the following chain of inequalities: \begin{eqnarray*} load_{T}\left(\bigcup_{i=1}^{n} K_{i}\right) & = & cost\left(\bigcup_{i=1}^{n} K_{i}\right)+mst\left(T\cup \bigcup_{i=1}^{n} E_{0}\left( K_{i}\right)\right)-cost(T)\\ & = & \sum_{i=1}^{n} cost(K_{i})+mst\left(T\cup \bigcup_{j=1}^{i} E_{0}\left( K_{j}\right)\right)-cost\left(T/\bigcup_{j=1}^{i-1} E_{0}\left( K_{j}\right)\right)\\ & \geq & \sum_{i=1}^{n} cost(K_{i})+mst\left(T\cup E_{0}\left(K_{i}\right)\right)-cost(T) \\ & = & \sum_{i=1}^{n} load_{T}(K_{i}). \end{eqnarray*} \end{proof} The following lemma guarantees that the solution of $k$-TPH at the second phase is a well solution. \begin{lem}\label{lem03} For any chosen full components $K_{i}$ and $K_{j}$, $\left|\Gamma(K_{i})\cap \Gamma(K_{j})\right|\leq 1$. \end{lem} \begin{proof} Without loss of generality, assume that $\left|\Gamma(K_{i})\cap \Gamma(K_{j})\right|= 2$ and $j< i$. Both $T_{origin}^{i-1}-MST(T_{origin}^{i-1}\cup E_{0}(K_i))$ and $T_{base}^{i-1}-MST(T_{base}^{i-1}\cup E_{0}(K_i))$ contain a zero-cost edge that is from $E_{0}(K_{j})$. Since any full component cannot improve $T_{base}^{0}$, $MST(T_{base}^{0}\cup K) = T_{base}^{0}\cup Loss(K)$ for any full component $K$. We can find a edge $e\in K_{i}-Loss(K_{i})$ such that $\Psi_{T_{origin}^{i-1},T_{base}^{i-1}}(K_i)=\Psi_{T_{origin}^{i-1},T_{base}^{i-1}}(A)+\Psi_{T_{origin}^{i-1},T_{base}^{i-1}}(B)$ and $load_{T_{base}^{i-1}}(K_i)\geq load_{T_{base}^{i-1}}(A \cup B) \geq load_{T_{base}^{i-1}}(A)+load_{T_{base}^{i-1}}(B)$ (from Lemma~\ref{lem06}) where $A$ and $B$ are two connected components of $K_{i}-\left\{e\right\}$. Finally, \begin{eqnarray*} \frac{load_{T_{base}^{i-1}}(K_i)}{\Psi_{T_{origin}^{i-1},T_{base}^{i-1}}(K_i)}\geq \frac{load_{T_{base}^{i-1}}(A)+load_{T_{base}^{i-1}}(B)}{\Psi_{T_{origin}^{i-1},T_{base}^{i-1}}(A)+\Psi_{T_{origin}^{i-1},T_{base}^{i-1}}(B)} \geq \min\left\{\frac{load_{T_{base}^{i-1}}(A)}{\Psi_{T_{origin}^{i-1},T_{base}^{i-1}}(A)},\frac{load_{T_{base}^{i-1}}(B)}{\Psi_{T_{origin}^{i-1},T_{base}^{i-1}}(B)}\right\} \end{eqnarray*} which contradicts the choice of $K_{i}$. \end{proof} \begin{lem}\label{lem04} For any Steiner tree $S$, $load_{T_{base}^{0}}(S) \geq load_{T_{base}^{i-1}}\left(S/\bigcup_{j=1}^{i-1}E_{0}\left(K_{i}\right)\right)$. \end{lem} \begin{proof} Since no full component can improve the terminal-spanning tree $T_{base}^{0}$, $cost(S)-cost(T_{base}^{0})-mst\left(S\cup \bigcup_{j=1}^{i-1} E_{0}\left(K_{i}\right)\right)+mst\left(T_{base}^{0}\cup \bigcup_{j=1}^{i-1} E_{0}\left(K_{i}\right)\right)\geq 0$. The proof can be obtained by the following chain of inequalities: \begin{eqnarray*} load_{T_{base}^{0}}(S) & = & cost(S)-cost(T_{base}^{0}) \\ & = & mst\left(S\cup \bigcup_{j=1}^{i-1} E_{0}\left(K_{i}\right)\right)-mst\left(T_{base}^{0}\cup \bigcup_{j=1}^{i-1} E_{0}\left(K_{i}\right)\right) \\ & & + cost(S)-cost(T_{base}^{0})-mst\left(S\cup \bigcup_{j=1}^{i-1} E_{0}\left(K_{i}\right)\right)+mst\left(T_{base}^{0}\cup \bigcup_{j=1}^{i-1} E_{0}\left(K_{i}\right)\right)\\ & \geq & mst\left(S\cup \bigcup_{j=1}^{i-1} E_{0}\left(K_{i}\right)\right)-mst\left(T_{base}^{0}\cup \bigcup_{j=1}^{i-1} E_{0}\left(K_{i}\right)\right) \\ & = & load_{T_{base}^{i-1}}\left(S/\bigcup_{j=1}^{i-1}E_{0}\left(K_{i}\right)\right). \end{eqnarray*} \end{proof} \begin{lem}\label{lem11} If $load_{T_{base}^{i-1}/E_{0}\left(C\right)}(K)\leq \Psi_{T_{origin}^{i-1},T_{base}^{i-1}/ E_{0}\left(C\right)}(K)$ for any full components $C$ and $K$, \[ \frac{load_{T_{base}^{i-1}/ E_{0}\left(C\right)}(K)}{\Psi_{T_{origin}^{i-1},T_{base}^{i-1}/ E_{0}\left(C\right)}(K)} \geq \frac{load_{T_{base}^{i-1}}(K)}{\Psi_{T_{origin}^{i-1},T_{base}^{i-1}}(K)}. \] \end{lem} \begin{proof} Since $load_{T_{base}^{i-1}/E_{0}\left(C\right)}(K)\leq \Psi_{T_{origin}^{i-1},T_{base}^{i-1}/ E_{0}\left(C\right)}(K)$ and $cost(T_{base}^{i-1}/ E_{0}\left(C\right))-mst(T_{base}^{i-1}\cup E_{0}\left(C\right)\cup E_{0}\left(K\right))\leq cost(T_{base}^{i-1})-mst(T_{base}^{i-1}\cup E_{0}\left(K\right))$, the proof can be obtained by the following chain of inequalities: \begin{eqnarray*} & & \frac{load_{T_{base}^{i-1}/ E_{0}\left(C\right)}(K)}{\Psi_{T_{origin}^{i-1},T_{base}^{i-1}/ E_{0}\left(C\right)}(K)} \\ & = & \frac{cost(K)+mst(T_{base}^{i-1}\cup E_{0}\left(C\right)\cup E_{0}\left(K\right))-cost(T_{base}^{i-1}/ E_{0}\left(C\right))}{cost(T_{origin}^{i-1})-cost(T_{base}^{i-1}/ E_{0}\left(C\right))-mst(T_{origin}^{i-1}\cup E_{0}(K))+mst(T_{base}^{i-1}\cup E_{0}\left(C\right)\cup E_{0}(K))} \\ & \geq & \frac{cost(K)+mst(T_{base}^{i-1}\cup E_{0}\left(K\right))-cost(T_{base}^{i-1})}{cost(T_{origin}^{i-1})-cost(T_{base}^{i-1})-mst(T_{origin}^{i-1}\cup E_{0}(K))+mst(T_{base}^{i-1}\cup E_{0}(K))} \\ & = & \frac{load_{T_{base}^{i-1}}(K)}{\Psi_{T_{origin}^{i-1},T_{base}^{i-1}}(K)}. \end{eqnarray*} \end{proof} Based on the analysis in~\cite{Zel96}, the bound on the cost of our solution is as follows. \begin{thm}\label{thm:0001} The $k$-TPH finds a Steiner tree $S$ such that \[ cost(S) \leq \left(\ln \frac{mst -cost(T_{base}^{0})}{opt_{k} -cost(T_{base}^{0})}+ 1\right)\cdot \left(opt_{k}-cost(T_{base}^{0})\right) + cost(T_{base}^{0}). \] \end{thm} \begin{proof} Let $M_i=cost(T_{origin}^i)-cost(T_{base}^i)$ and $m_i=M_{i-1}-M_{i}$. Therefore, $f(K_i)=\frac{load_{T_{base}^{i-1}}(K_i)}{m_i}$. Let $Opt_{k}^{i-1}=\left(Opt_{k}/\bigcup_{l=1}^{i-1} E_{0}\left(k_{l}\right)\right)-\bigcup_{l=1}^{i-1} E_{0}\left(K_{l}\right)$. For $i=1,\ldots, r+1$ and $load_{T_{base}^{0}}(Opt_{k})\leq M_{i-1}$, we have \begin{eqnarray*} \frac{load_{T_{base}^{0}}(Opt_{k})}{M_{i-1}} & = & \frac{load_{T_{base}^{0}}(Opt_{k})}{\Psi_{T_{origin}^{i-1},T_{base}^{i-1}}(Opt_{k})} \\ & \overset{\mbox{Lem~\ref{lem04}}}{\geq} & \frac{load_{T_{base}^{i-1}}(Opt_{k}^{i-1})}{\Psi_{T_{origin}^{i-1},T_{base}^{i-1}}(Opt_{k}^{i-1})} \\ & = & \frac{\sum_{X_{j}\in Opt_{k}^{i-1}} load_{T_{base}^{i-1}/ \bigcup_{l=1}^{j-1} E_{0}\left(X_{l}\right)}(X_{j})}{\sum_{X_{j}\in Opt_{k}^{i-1}} \Psi_{T_{origin}^{i-1}/ \bigcup_{l=1}^{j-1} E_{0}\left(X_{l}\right),T_{base}^{i-1}/ \bigcup_{l=1}^{j-1} E_{0}\left(X_{l}\right)}(X_{j})} \\ & \geq & \frac{\sum_{X_j\in Opt_{k}^{i-1}} load_{T_{base}^{i-1}/ \bigcup_{l=1}^{j-1} E_{0}\left(X_{l}\right)}(X_j)}{\sum_{X_j\in Opt_{k}^{i-1}} \Psi_{T_{origin}^{i-1},T_{base}^{i-1}/ \bigcup_{l=1}^{j-1} E_{0}\left(X_{l}\right)}(X_j)} \\ & \overset{\mbox{Lem~\ref{lem11}}}{\geq} & \frac{\sum_{X_j\in Opt_{k}^{i-1}} load_{T_{base}^{i-1}}(X_j)}{\sum_{X_j\in Opt_{k}^{i-1}} \Psi_{T_{origin}^{i-1},T_{base}^{i-1}}(X_j)} \\ & \geq & \min_{X_j\in Opt_{k}^{i-1}} \left\{\frac{load_{T_{base}^{i-1}}(X_j)}{\Psi_{T_{origin}^{i-1},T_{base}^{i-1}}(X_j)}\right\} \\ & \geq & \frac{load_{T_{base}^{i-1}}(K_{i})}{m_i}. \end{eqnarray*} Replacing $m_i=M_{i-1}-M_{i}$ into the above inequality yields \begin{eqnarray}\label{eq:1} M_{i} \leq M_{i-1}\left(1-\frac{load_{T_{base}^{i-1}}(K_{i})}{load_{T_{base}^{0}}(Opt_{k})}\right) \end{eqnarray} for $i = 1, 2, \ldots, t$. From the inequality~(\ref{eq:1}), \begin{eqnarray*} M_{r} & \leq & M_{0}\prod_{i=1}^{t}\left(1-\frac{load_{T_{base}^{i-1}}(K_{i})}{load_{T_{base}^{0}}(Opt_{k})}\right). \end{eqnarray*} Taking the natural logarithms of both sides and using the inequality $\ln(1+x)\leq x$, \begin{eqnarray}\label{eq:001} \ln \frac{M_{0}}{M_{r}} & \geq & -\sum_{i=1}^{t} \ln \left(1-\frac{load_{T_{base}^{i-1}}(K_{i})}{load_{T_{base}^{0}}(Opt_{k})}\right) \notag \\ & \geq & \frac{\sum_{i=1}^{t} load_{T_{base}^{i-1}}(K_{i})}{load_{T_{base}^{0}}(Opt_{k})}. \end{eqnarray} Since $k$-TPA interrupts at $M_{t}=c(T_{origin}^{t})-c(T_{base}^{t})=0$, there exists $M_{r}>load_{T_{base}^{0}}(Opt_{k}) \geq M_{r+1}$ for some $r< t$. The value $m_{r+1}$ can be split into two values $m^*$ and $m'$ such that \begin{eqnarray}\label{enq:0003} m^*=M_{r}-load_{T_{base}^{0}}(Opt_{k}), \end{eqnarray} \begin{eqnarray}\label{enq:0004} m'=load_{T_{base}^{0}}(Opt_{k})-M_{r+1}, \end{eqnarray} According to inequality (\ref{enq:0003}), we have \begin{eqnarray}\label{enq:0005} M_{r+1}^*=M_{r}-m^*=M_{r}-M_{r}+load_{T_{base}^{0}}(Opt_{k})=load_{T_{base}^{0}}(Opt_{k}). \end{eqnarray} The value $load_{T_{base}^{r}}(K_{r+1})$ also can be split into $w^*$ and $w'$ such that $\frac{load_{T_{base}^{r}}(K_{r+1})}{m_{r+1}} = \frac{w^*}{m^*} = \frac{w'}{m'}$. Since $\frac{load_{T_{base}^{r}}(K_{r+1})}{m_{r+1}} = \frac{w^*}{m^*}$, inequality (\ref{eq:001}) implies that \begin{eqnarray}\label{eq:002} \ln \frac{M_{0}}{M_{r+1}^*} \geq \frac{\sum_{i=1}^{r} load_{T_{base}^{i-1}}(K_{i})+w^*}{load_{T_{base}^{0}}(Opt_{k})}. \end{eqnarray} Since $\frac{load_{T_{base}^{r}}(K_{r+1})}{m_{r+1}}\leq \frac{load_{T_{base}^{0}}(Opt_{k})}{M_{r}}\leq 1$, we have \begin{eqnarray}\label{enq:0006} w'\leq m'. \end{eqnarray} The ratio related to the cost of approximate Steiner tree after $r + 1$ iterations is at most \begin{eqnarray*} \frac{cost(S_{2})-cost(T_{base}^{0})}{opt_{k}-cost(T_{base}^{0})} & = & \frac{mst(T_{origin}^{0}\cup\bigcup_{i=1}^{t} K_i)-cost(T_{base}^{0})}{load_{T_{base}^{0}}(Opt_{k})} \\ & \overset{\mbox{Lem~\ref{lem03}}}{\leq} & \frac{\sum_{i=1}^{r+1} load_{T_{base}^{i-1}}(K_{i}) +M_{r+1}}{load_{T_{base}^{0}}(Opt_{k})} \\ & = & \frac{\sum_{i=1}^{r} load_{T_{base}^{i-1}}(K_{i}) +w^*+w' +M_{r+1}}{load_{T_{base}^{0}}(Opt_{k})} \\ & \overset{\mbox{(\ref{eq:002})}}{\leq} & \ln \frac{M_{0}}{M_{r+1}^*}+ \frac{w' +M_{r+1}}{load_{T_{base}^{0}}(Opt_{k})} \\ & \overset{\mbox{(\ref{enq:0006})}}{\leq} & \ln \frac{M_{0}}{M_{r+1}^*}+ \frac{m' +M_{r+1}}{load_{T_{base}^{0}}(Opt_{k})} \\ & \overset{\mbox{(\ref{enq:0004})}}{=} & \ln \frac{M_{0}}{M_{r+1}^*}+ 1 \\ & \overset{\mbox{(\ref{enq:0005})}}{=} & \ln \frac{M_{0}}{load_{T_{base}^{0}}(Opt_{k})}+ 1 \\ & = & \ln \frac{cost(T_{origin}^{0}) -cost(T_{base}^{0})}{opt_{k} -cost(T_{base}^{0})}+ 1 \\ & = & \ln \frac{mst -cost(T_{base}^{0})}{opt_{k} -cost(T_{base}^{0})}+ 1 \end{eqnarray*} which yields \begin{eqnarray}\label{enq:0007} cost(S) \leq cost(S_{2}) \leq \left(\ln \frac{mst -cost(T_{base}^{0})}{opt_{k} -cost(T_{base}^{0})}+ 1\right)\cdot \left(opt_{k}-cost(T_{base}^{0})\right) + cost(T_{base}^{0}). \end{eqnarray} \end{proof} Since $cost(T_{base}^{0})\leq opt_k$ (from Lemma~\ref{lem08}) and $cost(T_{base}^{0})\geq \frac{1}{2}\cdot cost(S_1) \geq \frac{1}{2}\cdot opt_k$ (from Lemma~\ref{lem01}), we can assume that $cost(T_{base}^{0}) = \alpha \cdot opt_k$ for $\alpha \in \left(\frac{1}{2},1\right)$. The following result can be obtained. \begin{thm}\label{thm:0003} If $cost(T_{base}^{0}) = \alpha \cdot opt_k$ for $\alpha \in \left(\frac{1}{2},1\right)$, the $k$-TPH finds a Steiner tree $S$ such that \begin{eqnarray*} cost(S) \leq \left(\ln \frac{mst -\alpha \cdot opt_k}{opt_{k} -\alpha \cdot opt_k}+ 1\right)\cdot \left(opt_{k}-\alpha \cdot opt_k\right) + \alpha \cdot opt_k. \end{eqnarray*} and \begin{eqnarray*} cost(S) \leq 2\cdot \alpha \cdot opt_k. \end{eqnarray*} \end{thm} \begin{proof} From Theorem \ref{thm:0001}, we have \begin{eqnarray*} cost(S) \leq \left(\ln \frac{mst -\alpha \cdot opt_k}{opt_{k} -\alpha \cdot opt_k}+ 1\right)\cdot \left(opt_{k}-\alpha \cdot opt_k\right) + \alpha \cdot opt_k. \end{eqnarray*} According to Lemma~\ref{lem01}, $cost(S) \leq 2\cdot cost(T_{base}^{0}) =2\cdot \alpha \cdot opt_k$. \end{proof} \section{Performance of the $k$-TPH in general graphs} The following corollaries gives a bound on the cost of the Steiner tree generated by $k$-TPH. \begin{coro}\label{thm0006} The $k$-TPH has an approximation ratio of at most $1.4295$. \end{coro} \begin{proof} We have $mst \leq 2 \cdot opt$ (see~\cite{Takahashi80}). Theorem~\ref{thm:0003} yield \begin{eqnarray*} \frac{cost(S)}{opt} & \leq & \left(\ln \frac{2 \cdot opt -\alpha \cdot opt_k}{opt_{k} -\alpha \cdot opt_k}+ 1\right)\cdot \left(1-\alpha\right)\frac{opt_{k}}{opt} + \alpha \cdot \frac{opt_{k}}{opt} \\ & = & \left(\ln \frac{\frac{2}{\rho_k} -\alpha}{1 -\alpha}+ 1\right)\cdot \left(1-\alpha\right)\rho_k + \alpha \cdot \rho_k \end{eqnarray*} and \begin{eqnarray*} \frac{cost(S)}{opt} \leq 2\cdot \alpha \cdot \rho_k, \end{eqnarray*} where $\rho_k$ is the worst-case ratio of $\frac{opt_{k}}{opt}$. Borchers and Du~\cite{Borchers97} show that $\rho_k \leq 1+\left\lfloor \log_{2}k\right\rfloor^{-1}$ and $\lim_{k\rightarrow \infty}\rho_k =1$. When $k\rightarrow \infty$, the approximation ratio of the $k$-TPH converges to \begin{eqnarray*} A(\alpha) = \left(\ln \frac{2 -\alpha}{1 -\alpha}+ 1\right)\cdot \left(1-\alpha\right) + \alpha. \end{eqnarray*} and \begin{eqnarray*} B(\alpha) = 2\cdot \alpha. \end{eqnarray*} Since $A(\alpha)$ is decreasing in $\alpha$ and $B(\alpha)$ is increasing in $\alpha$, solving $A(\alpha)=B(\alpha)$ yeilds $\alpha^*\approx 0.7147$. The $k$-TPH has an approximation ratio of at most $A(\alpha^*)\approx 1.4295$. \end{proof}

https://arxiv.org/abs/2010.16391

Error bounds, facial residual functions and applications to the exponential cone

We construct a general framework for deriving error bounds for conic feasibility problems. In particular, our approach allows one to work with cones that fail to be amenable or even to have computable projections, two previously challenging barriers. For the purpose, we first show how error bounds may be constructed using objects called one-step facial residual functions. Then, we develop several tools to compute these facial residual functions even in the absence of closed form expressions for the projections onto the cones. We demonstrate the use and power of our results by computing tight error bounds for the exponential cone feasibility problem. Interestingly, we discover a natural example for which the tightest error bound is related to the Boltzmann-Shannon entropy. We were also able to produce an example of sets for which a Hölderian error bound holds but the supremum of the set of admissible exponents is not itself an admissible exponent.

\section{Introduction} Our main object of interest is the following conic feasibility problem: \begin{align} \text{find} & \quad \ensuremath{\mathbf x} \in ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \cap \ensuremath{\mathcal{K}} \label{eq:feas}\tag{Feas}, \end{align} where $ \ensuremath{\mathcal{L}}$ is a subspace contained in some finite-dimensional real Euclidean space $\ensuremath{\mathcal{E}}$, $\ensuremath{\mathbf a} \in \ensuremath{\mathcal{E}}$ and $ \ensuremath{\mathcal{K}} \subseteq \ensuremath{\mathcal{E}}$ is a closed convex cone. In particular, we are interested in how to obtain \emph{error bounds} for \eqref{eq:feas}. That is, assuming $( \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})\cap \ensuremath{\mathcal{K}}\neq \emptyset$, we want an inequality that, given some arbitrary $\ensuremath{\mathbf x} \in \ensuremath{\mathcal{E}}$, relates the \emph{individual distances} $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}), \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}})$ to the \emph{distance to the intersection} $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, ( \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})\cap \ensuremath{\mathcal{K}})$. Considering that $\ensuremath{\mathcal{E}}$ is equipped with some norm $\norm{\cdot}$ induced by some inner product $\inProd{\cdot}{\cdot}$, we recall that the \emph{distance function to a convex set $C$} is defined as follows: \begin{equation*} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, C) \coloneqq \inf _{\ensuremath{\mathbf y} \in C} \norm{\ensuremath{\mathbf x}-\ensuremath{\mathbf y}}. \end{equation*} When $ \ensuremath{\mathcal{K}}$ is a polyhedral cone, the classical Hoffman's error bound \cite{HF57} gives a relatively complete picture of the way that the individual distances relate to the distance to the intersection. If $ \ensuremath{\mathcal{K}}$ is not polyhedral, but $ \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ intersects $ \ensuremath{\mathcal{K}}$ in a sufficiently well-behaved fashion, we may still expect ``good'' error bounds to hold, e.g., \cite[Corollary~3]{BBL99}. Here, however, we focus on error bound results that \emph{do not} require any assumption on the way that the affine space $ \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ intersects $ \ensuremath{\mathcal{K}}$. So, for example, we want results that are valid even if, say, $ \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ fails to intersect the relative interior of $ \ensuremath{\mathcal{K}}$. Inspired by Sturm's pioneering work on error bounds for positive semidefinite systems \cite{St00}, the class of \emph{amenable cones} was proposed in \cite{L17} and it was shown that the following three ingredients can be used to obtain general error bounds for \eqref{eq:feas}: (i) amenable cones, (ii) facial reduction \cite{BW81,Pa13,WM13} and (iii) the so-called facial residual functions (FRFs). In this paper, we will show that, in fact, it is possible to obtain error bounds for \eqref{eq:feas} by using the facial residual functions (with respect to $ \ensuremath{\mathcal{K}}$) directly in combination with facial reduction. It is fair to say that computing the facial residual functions is the most critical step in obtaining error bounds for \eqref{eq:feas}. We will demonstrate several techniques that are readily adaptable for the purpose. In particular, we introduce a generalization of amenability called \emph{$\ensuremath{\mathfrak{g}}$-amenability} and discuss how it helps in the computation of the facial residual functions. All the techniques discussed here will be showcased with error bounds for the so-called \emph{exponential cone} which is defined as follows: \begin{align*} \ensuremath{K_{\exp}}:=&\left \{(x,y,z)\in \ensuremath{\mathbb R}^3\;|\;y>0,z\geq ye^{x/y}\right \} \cup \left \{(x,y,z)\;|\; x \leq 0, z\geq 0, y=0 \right \}. \end{align*} Put succinctly, the exponential cone is the closure of the epigraph of the perspective function of $z=e^x$. It is quite useful in entropy optimization, see \cite{CS17}. Furthermore, it is also implemented in the MOSEK package, see \cite[Chapter 5]{MC2020} and the many modelling examples in Section~5.4 therein. So convex optimization with exponential cones is widely available even if, as of this writing, it is not as widespread as, say, semidefinite programming. The exponential cone $\ensuremath{K_{\exp}}$ appears, at a glance, to be simple. However, it possesses a very intricate geometric structure that illustrates a wide range of challenges practitioners may face in computing error bounds. First of all, it is not amenable (Proposition~\ref{prop:exp_am}), so the theory developed in \cite{L17} does not directly apply to it. Another difficulty is that not many analytical tools have been developed to deal with the projection operator onto $\ensuremath{K_{\exp}}$ (compared with, for example, the projection operator onto PSD cones) which is only implicitly specified. Until now, these issues have made it challenging the establishment of error bounds for objects like $\ensuremath{K_{\exp}}$, many of which are of growing interest in mathematical programming. Our research is at the intersection of two topics: \emph{error bounds} and the \emph{facial structure of cones}. General information on the former can be found, for example, in \cite{Pang97,LP98}. Classically, there seems to be a focus on the so-called \emph{H\"olderian error bounds} (see also \cite{Li10,Li13,LMP15}) but we will see in this paper that non H\"olderian behavior can still appear even in relatively natural settings such as conic feasibility problems associated to the exponential cone. Facts on the facial structure of convex cones can be found, for example, in \cite{Ba73,Ba81,Pa00}. We recall that a cone is said to be \emph{facially exposed} if each face arises as the intersection of the whole cone with some supporting hyperplane. Stronger forms of facial exposedness have also been studied to some extent, here are some examples: projectional exposedness~\cite{BW81,ST90}, niceness~\cite{Pa13_2,Ve14}, tangential exposedness~\cite{RT19}, amenability \cite{L17}. These notions are useful in many topics, e.g.: regularization of convex programs and extended duals~\cite{BW81,Pa13,LiuPat18}, studying the closure of certain linear images~\cite{Pa07,LiuPat18}, lifts of convex sets~\cite{GPR13} and error bounds \cite{L17}. However, as can be seen in Figure~\ref{fig:cone}, the exponential cone is not even a facially exposed cone, so none of the aforementioned notions apply. This was one of the motivations for looking beyond facial exposedness and considering the class of {$\ensuremath{\mathfrak{g}}$-amenable cones}, which does include the exponential cone. \begin{figure} \begin{center} \begin{tikzpicture} [scale=0.76] \node[anchor=south west,inner sep=0] (image) at (0,0) {\includegraphics[width=.4\linewidth]{exponential_cone_fig1.png}}; \begin{scope}[x={(image.south east)},y={(image.north west)}] \node[purple] at (0.65,0.025) {$x$}; \node[purple] at (0.9,0.36) {$y$}; \node[purple] at (0.06,0.9) {$z$}; \end{scope} \end{tikzpicture} \end{center} \caption{The exponential cone}\label{fig:cone} \end{figure} \subsection{Outline and results} The goal of this paper is to build a robust framework that may be used to obtain error bounds for previously inaccessible cones, and to demonstrate the use of this framework by applying it to fully describe error bounds for \eqref{eq:feas} with $ \ensuremath{\mathcal{K}} = \ensuremath{K_{\exp}}$. In Section~\ref{sec:preliminaries}, we recall preliminaries. New contributions begin in Section~\ref{sec:frf}. We first recall some rules for chains of faces and the diamond composition. Then we show how error bounds may be constructed using objects known as facial residual functions (FRFs), whose definition we recall and motivate. In Sections~\ref{ss:alpha-amenable} and \ref{sec:frf_comp}, we introduce the notion of $\ensuremath{\mathfrak{g}}$-amenable cones, and build our general framework for constructing FRFs. Our key result, Theorem~\ref{thm:1dfacesmain}, obviates the need of computing explicitly the projection onto the cone. Instead, we make use of the parametrization of the boundary of the cone. We emphasize that \textit{all} of the results of Section~\ref{sec:frf} are applicable to a \textit{general closed convex cone}. In Section~\ref{sec:exp_cone}, we use our new framework to fully describe error bounds for \eqref{eq:feas} with $ \ensuremath{\mathcal{K}} = \ensuremath{K_{\exp}}$. This was previously a problem lacking a clear strategy, because all projections onto $\ensuremath{K_{\exp}}$ are implicitly specified. However, having obviated the projection barrier, we successfully obtain all the necessary FRFs. Surprisingly, we discover that different collections of faces and exposing hyperplanes admit very different FRFs. In Section~\ref{sec:freas_2d}, we show that for the unique 2-dimensional face, any exponent in $\left(0,1\right)$ may be used to build a valid FRF, while the supremum over all the admissible exponents \textit{cannot} be. As a surprising consequence, the exponential cone is, in a sense, arbitrarily close to being amenable, but it is \textit{not} an amenable cone (Proposition~\ref{prop:exp_am}). Furthermore, a better FRF for the 2D face can be obtained if we go beyond H\"olderian error bounds and consider a so-called \emph{entropic error bound} which uses a modified Boltzmann-Shannon entropy function, see Theorem~\ref{thm:entropic}. The curious discoveries continue; for infinitely many 1-dimensional faces, the FRF, and the final error bound, feature exponent $1/2$. For the final outstanding 1-dimensional face, the FRF, and the final error bound, are Lipschitzian for all exposing hyperplanes except exactly one, for which \textit{no exponent} will suffice. However, for this exceptional case, our framework \textit{still} successfully finds an FRF, which is logarithmic in character (Corollary~\ref{col:1dfaces_infty}). Consequentially, the system consisting of $\{(0,0,1)\}^\perp$ and $\ensuremath{K_{\exp}}$ possesses a kind of ``logarithmic error bound" (see Example~\ref{ex:non_hold}) instead of a H\"olderian error bound. In Theorems~\ref{thm:main_err} and \ref{theo:sane}, we give explicit error bounds by using our FRFs and the suite of tools we developed in Section~\ref{sec:frf}. We also show that the error bound given in Theorem~\ref{thm:main_err} is tight, see Remark~\ref{rem:opt}. These findings about the exponential cone are surprising, since we are not aware of other objects having this litany of odd behaviour hidden in its structure all at once.\footnote{To be fair, the exponential function is the classical example of a non-semialgebraic analytic function. Given that semialgebraicity is connected to the KL property (which is related to error bounds), one may argue that it is not \emph{that} surprising that the exponential cone has its share of quirks. Nevertheless, given how natural the exponential cone is, the amount of quirks is still somewhat surprising.} One possible reason for the absence of previous reports on these phenomena might have been the sheer absence of tools for obtaining error bounds for general cones. In this sense, we believe that the machinery developed in Section~\ref{sec:frf} might be a reasonable first step towards filling this gap. In Section~\ref{sec:odd}, we document additional odd consequences and connections to other concepts, with particular relevance to the \emph{Kurdyka-{\L}ojasiewicz (KL) property} \cite{BDL07,BDLS07,ABS13,ABRS10,BNPS17,LP18}. In particular, we have two sets satisfying a H\"olderian error bound for every $\gamma \in \left(0,1\right)$ but the supremum of allowable exponents is not allowable. Consequently, one obtains a function with the $KL$ property with exponent $\alpha$ for any $\alpha \in \left(1/2,1\right)$ at the origin, but not for $\alpha = 1/2$. We conclude in Section~\ref{sec:conclusion}. \section{Preliminaries}\label{sec:preliminaries} We recall that $\ensuremath{\mathcal{E}}$ denotes an arbitrary finite-dimensional real Euclidean space. We will adopt the following convention, vectors will be boldfaced while scalars will use normal typeface. For example, if $\ensuremath{\mathbf p} \in \ensuremath{\mathbb R}^3$, we write $\ensuremath{\mathbf p} = (p_x,p_y,p_z)$, where $p_x,p_y,p_z \in \ensuremath{\mathbb R}$. We denote by $B(\eta)$ the closed ball of radius $\eta$ centered at the origin, i.e., $B(\eta) = \{\ensuremath{\mathbf x}\in\ensuremath{\mathcal{E}} \mid \norm{\ensuremath{\mathbf x}} \leq \eta\}$. Let $C\subseteq \ensuremath{\mathcal{E}}$ be a convex set. We denote the relative interior and the linear span of $C$ by $\ensuremath{\mathrm{ri}\,} C$ and $\ensuremath{\mathrm{span}\,} C$, respectively. We also denote the boundary of $C$ by $\partial C$, and $\mathrm{cl}\, C$ is the closure of $C$. We denote the projection operator onto $C$ by $P_C$, so that $P_C(\ensuremath{\mathbf x}) = \ensuremath{\operatorname*{argmin}}_{\ensuremath{\mathbf y} \in C} \norm{\ensuremath{\mathbf x}-\ensuremath{\mathbf y}}$. Given closed convex sets $C_1,C_2 \subseteq \ensuremath{\mathcal{E}}$, we note the following properties of the projection operator \begin{align} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x},C_1) &\leq \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x},C_2) + \ensuremath{\operatorname{d}}(P_{C_2}(\ensuremath{\mathbf x}),C_1) \label{proj:p1}\\ \ensuremath{\operatorname{d}}(P_{C_2}(\ensuremath{\mathbf x}),C_1) &\leq \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x},C_2) + \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x},C_1). \label{proj:p2} \end{align} \subsection{Cones and their faces} Let $ \ensuremath{\mathcal{K}}$ be a closed convex cone. We say that $ \ensuremath{\mathcal{K}}$ is \emph{pointed} if $ \ensuremath{\mathcal{K}} \cap - \ensuremath{\mathcal{K}} = \{\mathbf{0}\}$. A \emph{face} of $ \ensuremath{\mathcal{K}}$ is a closed convex cone $ \ensuremath{\mathcal{F}}$ satisfying $ \ensuremath{\mathcal{F}} \subseteq \ensuremath{\mathcal{K}}$ and the following property \[ \ensuremath{\mathbf x},\ensuremath{\mathbf y} \in \ensuremath{\mathcal{K}}, \ensuremath{\mathbf x}+\ensuremath{\mathbf y} \in \ensuremath{\mathcal{F}} \Rightarrow \ensuremath{\mathbf x},\ensuremath{\mathbf y} \in \ensuremath{\mathcal{F}}. \] In this case, we write $ \ensuremath{\mathcal{F}} \mathrel{\unlhd} \ensuremath{\mathcal{K}}$. We say that $ \ensuremath{\mathcal{F}}$ is \emph{proper} if $ \ensuremath{\mathcal{F}} \neq \ensuremath{\mathcal{K}}$. A face is said to be \emph{nontrivial} if $ \ensuremath{\mathcal{F}} \neq \ensuremath{\mathcal{K}}$ and $ \ensuremath{\mathcal{F}} \neq \ensuremath{\mathcal{K}} \cap - \ensuremath{\mathcal{K}}$. In particular, if $ \ensuremath{\mathcal{K}}$ is pointed (as is the case of the exponential cone), a nontrivial face is neither $ \ensuremath{\mathcal{K}}$ nor $\{\mathbf{0}\}$. Next, let $ \ensuremath{\mathcal{K}}^*$ denote the dual cone of $ \ensuremath{\mathcal{K}}$, i.e., $ \ensuremath{\mathcal{K}}^* = \{\ensuremath{\mathbf z} \in \ensuremath{\mathcal{E}} \mid \inProd{\ensuremath{\mathbf x}}{\ensuremath{\mathbf z}} \geq 0, \forall \ensuremath{\mathbf x} \in \ensuremath{\mathcal{K}} \}$. We say that $ \ensuremath{\mathcal{F}}$ is an \emph{exposed face} if there exists $\ensuremath{\mathbf z} \in \ensuremath{\mathcal{K}}^*$ such that $ \ensuremath{\mathcal{F}} = \ensuremath{\mathcal{K}} \cap \{\ensuremath{\mathbf z}\}^\perp$. A \emph{chain of faces} of $ \ensuremath{\mathcal{K}}$ is a sequence of faces satisfying $ \ensuremath{\mathcal{F}}_\ell \subsetneq \cdots \subsetneq \ensuremath{\mathcal{F}}_{1}$ such that each $ \ensuremath{\mathcal{F}}_{i}$ is a face of $ \ensuremath{\mathcal{K}}$ and the inclusions $ \ensuremath{\mathcal{F}} _{i+1} \subsetneq \ensuremath{\mathcal{F}}_{i}$ are all proper. The length of the chain is defined to be $\ell$. With that, we define the \emph{distance to polyhedrality of $ \ensuremath{\mathcal{K}}$} as the length {\em minus one} of the longest chain of faces of $ \ensuremath{\mathcal{K}}$ such that $ \ensuremath{\mathcal{F}} _{\ell}$ is polyhedral and $ \ensuremath{\mathcal{F}}_{i}$ is not polyhedral for $i < \ell$. We denote the distance to polyhedrality by $\ell _{\text{poly}}( \ensuremath{\mathcal{K}})$. \subsection{Lipschitzian and H\"olderian error bounds} In this subsection, suppose that $C_1,\ldots, C_{\ell} \subseteq \ensuremath{\mathcal{E}}$ are convex sets with nonempty intersection. We recall the following definitions. \begin{definition}[H\"olderian and Lipschitzian error bounds]\label{def:hold} The sets $C_1,\ldots, C_\ell$ are said to satisfy a \emph{H\"olderian error bound} if for every bounded set $B \subseteq \ensuremath{\mathcal{E}}$ there exist some $\theta_B > 0$ and an exponent $\gamma _B\in(0, 1]$ such that \begin{equation*} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \cap _{i=1}^\ell C_i) \le \theta_B\max_{1\le i\le \ell}\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \, C_i)^{\gamma_B}, \qquad \forall\ \ensuremath{\mathbf x}\in B. \end{equation*} If we can take the same exponent $\gamma _B = \gamma \in (0,1]$ for all $B$, then we say that the bound is \emph{uniform}. Furthermore, if the bound is uniform with $\gamma = 1$, we call it a \emph{Lipschitzian error bound}. \end{definition} H\"olderian and Lipschitzian error bounds will appear frequently in our results, but we also encounter examples of non-H\"olderian bounds as in Theorem~\ref{thm:entropic} and Theorem~\ref{thm:nonzerogammasec72}. Next, we recall the following result which ensures a Lipschitzian error bound holds between families of convex sets when a constraint qualification is satisfied. \begin{proposition}[{\cite[Corollary~3]{BBL99}}]\label{prop:cq_er} Let $C_1,\ldots, C_{\ell} \subseteq \ensuremath{\mathcal{E}}$ be convex sets such that $C_{1},\ldots, C_{k}$ are polyhedral. If \begin{equation*} \left(\bigcap _{i=1}^k C_i\right) \bigcap \left(\bigcap _{j=k+1}^\ell \ensuremath{\mathrm{ri}\,} C_j\right) \neq \emptyset, \end{equation*} then for every bounded set $B$ there exists $\kappa _B>0$ such that \[ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \cap _{i=1}^\ell C_i) \leq \kappa _B(\max_{1 \leq i \leq \ell} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, C_i)), \qquad \forall \ensuremath{\mathbf x} \in B. \] \end{proposition} In view of \eqref{eq:feas}, we say that \emph{Slater's condition} is satisfied if $( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \cap \ensuremath{\mathrm{ri}\,} \ensuremath{\mathcal{K}} \neq \emptyset$. If $ \ensuremath{\mathcal{K}}$ can be written as $ \ensuremath{\mathcal{K}}^1 \times \ensuremath{\mathcal{K}}^2\subseteq \ensuremath{\mathcal{E}}^1 \times \ensuremath{\mathcal{E}}^2$, where $\ensuremath{\mathcal{E}}^1$ and $\ensuremath{\mathcal{E}}^2$ are real Euclidean spaces and $ \ensuremath{\mathcal{K}}^1\subseteq \ensuremath{\mathcal{E}}^1$ is polyhedral, we say that the \emph{partial polyhedral Slater's (PPS) condition} is satisfied if \begin{equation}\label{eq:ppsc} ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \cap ( \ensuremath{\mathcal{K}}^1 \times (\ensuremath{\mathrm{ri}\,} \ensuremath{\mathcal{K}}^2) ) \neq \emptyset. \end{equation} Adding a dummy coordinate, if necessary, we can see Slater's condition as a particular case of the PPS condition. By convention, we consider that the PPS condition is satisfied for \eqref{eq:feas} if one of the following is satisfied: 1) $ \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ intersects $\ensuremath{\mathrm{ri}\,} \ensuremath{\mathcal{K}}$; 2) $( \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})\cap \ensuremath{\mathcal{K}}\neq \emptyset$ and $ \ensuremath{\mathcal{K}}$ is polyhedral; or 3) $ \ensuremath{\mathcal{K}}$ can be written as a direct product $ \ensuremath{\mathcal{K}}^1 \times \ensuremath{\mathcal{K}}^2$ where $ \ensuremath{\mathcal{K}}^1$ is polyhedral and \eqref{eq:ppsc} is satisfied. Noting that $( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \cap ( \ensuremath{\mathcal{K}}^1 \times (\ensuremath{\mathrm{ri}\,} \ensuremath{\mathcal{K}}^2) ) = ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a})\cap ( \ensuremath{\mathcal{K}}^1 \times \ensuremath{\mathcal{E}}^2) \cap (\ensuremath{\mathcal{E}}^1 \times (\ensuremath{\mathrm{ri}\,} \ensuremath{\mathcal{K}}^2) )$, we deduce the following result from Proposition~\ref{prop:cq_er}. \begin{proposition}[Error bound under PPS condition]\label{prop:pps_er} Suppose that \eqref{eq:feas} satisfies the \emph{partial polyhedral Slater's condition}. Then, for every bounded set $B$ there exists $\kappa _B>0$ such that \[ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, ( \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})\cap \ensuremath{\mathcal{K}}) \leq \kappa _B \max\{\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}),\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})\}, \qquad \forall \ensuremath{\mathbf x} \in B. \] \end{proposition} \section{Facial residual functions and error bounds}\label{sec:frf} In this section, we discuss a strategy for obtaining error bounds for the conic linear system \eqref{eq:feas} based on the so-called \emph{facial residual functions} that were introduced in \cite{L17}. In contrast to \cite{L17}, we will not require that $ \ensuremath{\mathcal{K}}$ be amenable. The motivation for our approach is as follows. If it were the case that \eqref{eq:feas} satisfies some constraint qualification, we would have a Lipschitizian error bound per Proposition~\ref{prop:pps_er}. Unfortunately, this does not happen in general. However, as long as \eqref{eq:feas} is feasible, there is always a face of $ \ensuremath{\mathcal{K}}$ that contains the feasible region of \eqref{eq:feas} and for which a constraint qualification holds. The error bound computation essentially boils down to understanding how to compute the distance to this special face. The first result towards our goal is the following. \begin{proposition}[An error bound when a face satisfying a CQ is known]\label{prop:err_cq2} Suppose that \eqref{eq:feas} is feasible and let $ \ensuremath{\mathcal{F}} \mathrel{\unlhd} \ensuremath{\mathcal{K}}$ be a face such that \begin{enumerate}[{\rm (a)}] \item $ \ensuremath{\mathcal{F}}$ contains $ \ensuremath{\mathcal{K}} \cap ( \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})$ \item $ \ensuremath{\mathcal{F}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ satisfy the PPS condition.\footnote{As a reminder, the PPS condition is, by convention, a shorthand for three closely related conditions, see remarks after \eqref{eq:ppsc}.} \end{enumerate} Then, for every bounded set $B$, there exists $\kappa_B > 0$ such that \[ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}} \cap ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a})) \leq \kappa_B(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}) + \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})), \qquad \forall \ensuremath{\mathbf x} \in B. \] \end{proposition} \begin{proof} Since $ \ensuremath{\mathcal{F}}$ is a face of $ \ensuremath{\mathcal{K}}$, assumption (a) implies $ \ensuremath{\mathcal{K}} \cap ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) = \ensuremath{\mathcal{F}} \cap ( \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})$. Then, the result follows from assumption (b) and Proposition~\ref{prop:pps_er}. \end{proof} From Proposition~\ref{prop:err_cq2} we see that the key to obtaining an error bound for the system \eqref{eq:feas} is to find a face $ \ensuremath{\mathcal{F}} \mathrel{\unlhd} \ensuremath{\mathcal{K}}$ satisfying (a), (b) \emph{and} we must know how to estimate the quantity $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}})$ from the available information $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}})$ and $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})$. This is where we will make use of facial reduction and facial residual functions. The former will help us find $ \ensuremath{\mathcal{F}}$ and the latter will be instrumental in upper bounding $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}})$. First, we recall below a result that follows from the analysis of the FRA-poly facial reduction algorithm developed in \cite{LMT18}. \begin{proposition}[{\cite[Proposition~5]{L17}}\footnote{Although {\cite[Proposition~5]{L17}} was originally stated for pointed cones, it holds for general closed convex cones. Indeed, its proof only relies on {\cite[Proposition~8]{LMT18}}, which holds for general closed convex cones.}]\label{prop:fra_poly} Let $ \ensuremath{\mathcal{K}} = \ensuremath{\mathcal{K}}^1\times \cdots \times \ensuremath{\mathcal{K}}^s$, where each $ \ensuremath{\mathcal{K}}^i$ is a closed convex cone. Suppose \eqref{eq:feas} is feasible. Then there is a chain of faces \begin{equation}\label{eq:chain} \ensuremath{\mathcal{F}} _{\ell} \subsetneq \cdots \subsetneq \ensuremath{\mathcal{F}}_1 = \ensuremath{\mathcal{K}} \end{equation} of length $\ell$ and vectors $\{\ensuremath{\mathbf z}_1,\ldots, \ensuremath{\mathbf z}_{\ell-1}\}$ satisfying the following properties. \begin{enumerate}[{\rm (i)}] \item \label{prop:fra_poly:1} $\ell -1\leq \sum _{i=1}^{s} \ell _{\text{poly}}( \ensuremath{\mathcal{K}} ^i) \leq \dim{ \ensuremath{\mathcal{K}}}$ \item \label{prop:fra_poly:2} For all $i \in \{1,\ldots, \ell -1\}$, we have \begin{flalign*}% \ensuremath{\mathbf z}_i \in \ensuremath{\mathcal{F}} _i^* \cap \ensuremath{\mathcal{L}}^\perp \cap \{\ensuremath{\mathbf a}\}^\perp \ \ \ {and}\ \ \ \ensuremath{\mathcal{F}} _{i+1} = \ensuremath{\mathcal{F}} _{i} \cap \{\ensuremath{\mathbf z}_i\}^\perp. \end{flalign*} \item \label{prop:fra_poly:3} $ \ensuremath{\mathcal{F}} _{\ell} \cap ( \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}) = \ensuremath{\mathcal{K}} \cap ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a})$ and $ \ensuremath{\mathcal{F}} _{\ell}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ satisfy the PPS condition. \end{enumerate} \end{proposition} In view of Proposition~\ref{prop:fra_poly}, we define the \emph{distance to the PPS condition} $d_{\text{PPS}}( \ensuremath{\mathcal{K}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})$ as the length \emph{minus one} of the shortest chain of faces (as in \eqref{eq:chain}) satisfying item (iii) in Proposition~\ref{prop:fra_poly}. For example, if \eqref{eq:feas} satisfies the PPS condition, we have $d_{\text{PPS}}( \ensuremath{\mathcal{K}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}) = 0$. Next, we recall the definition of facial residual functions from \cite[Definition~16]{L17}. \begin{definition}[Facial residual functions\footnote{The only difference between Definition~\ref{def:ebtp} and the definition of facial residual functions in {\cite[Definition~16]{L17}} is that we added the ``with respect to $ \ensuremath{\mathcal{K}}$'' part, to emphasize the dependency on $ \ensuremath{\mathcal{K}}$.}]\label{def:ebtp} Let $ \ensuremath{\mathcal{K}}$ be a closed convex cone, $ \ensuremath{\mathcal{F}} \mathrel{\unlhd} \ensuremath{\mathcal{K}}$ be a face, and let $\ensuremath{\mathbf z} \in \ensuremath{\mathcal{F}}^*$. Suppose that $\psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}} : \ensuremath{\mathbb{R}_+} \times \ensuremath{\mathbb{R}_+} \to \ensuremath{\mathbb{R}_+}$ satisfies the following properties: \begin{enumerate}[label=({\roman*})] \item $\psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}$ is nonnegative, monotone nondecreasing in each argument and $\psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}(0,t) = 0$ for every $t \in \ensuremath{\mathbb{R}_+}$. \item The following implication holds for any $\ensuremath{\mathbf x} \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}$ and any $\epsilon \geq 0$: \[ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \leq \epsilon, \quad \inProd{\ensuremath{\mathbf x}}{\ensuremath{\mathbf z}} \leq \epsilon, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}} ) \leq \epsilon \quad \Rightarrow \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}} \cap \{\ensuremath{\mathbf z}\}^{\perp}) \leq \psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}} (\epsilon, \norm{\ensuremath{\mathbf x}}). \] \end{enumerate} Then, $\psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}$ is said to be \emph{a facial residual function for $ \ensuremath{\mathcal{F}}$ and $\ensuremath{\mathbf z}$ with respect to $ \ensuremath{\mathcal{K}}$}. \end{definition} \begin{remark}[Concerning the implication in {Definition~\ref{def:ebtp}}] In view of the monotonicity of $\psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}$, the implication in Definition~\ref{def:ebtp} can be equivalently and more succinctly written as \[ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}\cap\{\ensuremath{\mathbf z}\}^\perp) \le \psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}(\max\{d(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}), \inProd{\ensuremath{\mathbf x}}{\ensuremath{\mathbf z}},d(\ensuremath{\mathbf x},\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}})\}, \norm{\ensuremath{\mathbf x}}),\ \ \forall \ensuremath{\mathbf x} \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}. \] The {\em unfolded} form presented in Definition~\ref{def:ebtp} is more handy in our discussions and analysis below. \end{remark} Facial residual functions always exist (see \cite[Section~3.2]{L17}), but their computation is often nontrivial. Next, we review a few examples. \begin{example}[Examples of facial residual functions]\label{ex:frf} If $ \ensuremath{\mathcal{K}}$ is a symmetric cone (i.e., a self-dual homogeneous cone), then given $ \ensuremath{\mathcal{F}} \mathrel{\unlhd} \ensuremath{\mathcal{K}}$ and $\ensuremath{\mathbf z} \in \ensuremath{\mathcal{F}}^*$, there exists a $\kappa > 0$ such that $\psi _{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}(\epsilon,t) \coloneqq \kappa \epsilon + \kappa \sqrt{\epsilon t}$ is a facial residual function for $ \ensuremath{\mathcal{F}}$ and $\ensuremath{\mathbf z}$ with respect to $ \ensuremath{\mathcal{K}}$, see \cite[Theorem~35]{L17}. If $ \ensuremath{\mathcal{K}}$ is a polyhedral cone, the function $\psi _{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}(\epsilon,t) \coloneqq \kappa \epsilon$ can be taken instead, with no dependency on $t$, see \cite[Proposition~18]{L17}. \end{example} Moving on, we say that a function $\tilde \psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}$ is a \emph{positive rescaling of $\psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}$} if there are positive constants $M_1,M_2,M_3$ such that \begin{equation}\label{eq:pos_rescale} \tilde \psi _{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}(\epsilon,t) = M_3\psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}} (M_1\epsilon,M_2t). \end{equation} We also need to compose facial residual functions in a special manner. Let $f:\ensuremath{\mathbb{R}_+}\times \ensuremath{\mathbb{R}_+} \to \ensuremath{\mathbb{R}_+}$ and $g:\ensuremath{\mathbb{R}_+}\times \ensuremath{\mathbb{R}_+} \to \ensuremath{\mathbb{R}_+}$ be functions. We define the \emph{diamond composition} $f\diamondsuit g$ to be the function satisfying \begin{equation}\label{eq:comp} (f\diamondsuit g)(a,b) = f(a+g(a,b),b), \qquad \forall a,b \in \ensuremath{\mathbb{R}_+}. \end{equation} Note that the above composition is not associative in general. When we have functions $f_i:\ensuremath{\mathbb{R}_+}\times \ensuremath{\mathbb{R}_+} \to \ensuremath{\mathbb{R}_+}$, $i = 1,\ldots,m$ with $m\ge 3$, we define $f_m\diamondsuit \cdots \diamondsuit f_1$ inductively as the function $\varphi _m$ such that \begin{align*} \varphi_i &\coloneqq f_i \diamondsuit \varphi_{i-1},\qquad i \in \{2,\ldots,m\}\\ \varphi_1 &\coloneqq f_1. \end{align*} With that, we have $ f_m\diamondsuit f_{m-1} \diamondsuit \cdots \diamondsuit f_2 \diamondsuit f_1 \coloneqq f_m\diamondsuit (f_{m-1}\diamondsuit(\cdots \diamondsuit (f_2\diamondsuit f_1)))$. \begin{lemma}[Diamond composing facial residual functions, {\cite[Lemma~22]{L17}}\footnote{In \cite{L17}, although there is a blanket assumption that the cones are pointed, the proof of Lemma~22 does not require this assumption. For the sake of self-containment, we present here a more streamlined proof where $ \ensuremath{\mathcal{K}}$ is a general closed convex cone.}]\label{lem:chain} Suppose \eqref{eq:feas} is feasible and let \[ \ensuremath{\mathcal{F}} _{\ell} \subsetneq \cdots \subsetneq \ensuremath{\mathcal{F}}_1 = \ensuremath{\mathcal{K}} \] be a chain of faces of $ \ensuremath{\mathcal{K}}$ together with $\ensuremath{\mathbf z}_i \in \ensuremath{\mathcal{F}} _i^*\cap \ensuremath{\mathcal{L}}^\perp \cap \{\ensuremath{\mathbf a}\}^\perp$ such that $ \ensuremath{\mathcal{F}}_{i+1} = \ensuremath{\mathcal{F}} _i\cap \{\ensuremath{\mathbf z}_i\}^\perp$, for $i = 1,\ldots, \ell - 1$. For those $i$, let $\psi _{i}$ be a facial residual function for $ \ensuremath{\mathcal{F}}_{i}$ and $\ensuremath{\mathbf z}_{i}$ with respect to $ \ensuremath{\mathcal{K}}$. Then, there is a positive rescaling of $\psi_i$ (still denoted as $\psi_i$ by an abuse of notation) so that $$ \ensuremath{\mathbf x} \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \leq \epsilon, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \leq \epsilon \quad \Rightarrow\quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}} _{\ell}) \leq \varphi (\epsilon,\norm{\ensuremath{\mathbf x}}), $$ where $\varphi = \psi _{{\ell-1}}\diamondsuit \cdots \diamondsuit \psi_{{1}}$, if $\ell \geq 2$. If $\ell = 1$, we let $\varphi$ be the function satisfying $\varphi(\epsilon, t) = \epsilon$. \end{lemma} \begin{proof} For $\ell = 1$, we have $ \ensuremath{\mathcal{F}}_{\ell} = \ensuremath{\mathcal{K}}$, so the lemma follows immediately. Now, we consider the case $\ell \geq 2$. First we note that $ \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}$ is contained in all the $\{\ensuremath{\mathbf z}_{i}\}^\perp$ for $i = 1, \ldots, \ell-1$. Since the distance of $\ensuremath{\mathbf x} \in \ensuremath{\mathcal{E}}$ to $\{\ensuremath{\mathbf z}_{i}\}^\perp$ is given by $\frac{|\inProd{\ensuremath{\mathbf x}}{\ensuremath{\mathbf z}_i}|}{\norm{\ensuremath{\mathbf z}_i}}$, we have the following chain of implications \begin{equation}\label{eq:hyper} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \leq \epsilon\quad \Rightarrow\quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x},\{\ensuremath{\mathbf z}_i\}^\perp) \leq \epsilon \quad \Rightarrow\quad \inProd{\ensuremath{\mathbf x}}{\ensuremath{\mathbf z}_i} \leq \epsilon \norm{\ensuremath{\mathbf z}_i}. \end{equation} Next, we proceed by induction. If $\ell = 2$, we have that $\psi_1$ is a facial residual function for $ \ensuremath{\mathcal{K}}$ and $\ensuremath{\mathbf z}_1$ with respect to $ \ensuremath{\mathcal{K}}$. By Definition~\ref{def:ebtp}, we have \[ \ensuremath{\mathbf x} \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \leq \epsilon, \quad \inProd{\ensuremath{\mathbf x}}{\ensuremath{\mathbf z}_1} \leq \epsilon \quad \Rightarrow \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}_{2}) \leq \psi_{1} (\epsilon, \norm{\ensuremath{\mathbf x}}). \] In view of \eqref{eq:hyper} and the monotonicity of $\psi_1$, we see further that \[ \ensuremath{\mathbf x} \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \leq \epsilon, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \leq \epsilon \quad \Rightarrow \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}_{2}) \leq \psi_{1} (\epsilon(1+\norm{\ensuremath{\mathbf z}_{1}}), \norm{\ensuremath{\mathbf x}}), \] which proves the lemma for chains of length $\ell = 2$. Now, suppose that the lemma holds for chains of length $\hat \ell$ and consider a chain of length $\hat \ell + 1$. By the induction hypothesis, we have \begin{equation}\label{eq:inductive} \ensuremath{\mathbf x} \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \leq \epsilon, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \leq \epsilon \quad \Rightarrow \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}_{\hat \ell}) \leq \varphi (\epsilon, \norm{\ensuremath{\mathbf x}}), \end{equation} where $\varphi = \psi _{{\hat\ell-1}}\diamondsuit \cdots \diamondsuit \psi_1$ and the $\psi_i$ are facial residual functions (which might have been positively rescaled). By the definition of $\psi _{{\hat\ell}}$ and using \eqref{eq:hyper}, we may positively rescale $\psi _{{\hat\ell}}$ so that for $\ensuremath{\mathbf x} \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}$ and $\hat \epsilon \ge 0$, \begin{equation}\label{eq:eps_hat_ell} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \leq \hat\epsilon, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \leq \hat\epsilon, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}_{\hat \ell}) \leq \hat \epsilon \quad \Rightarrow \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}_{\hat \ell+1}) \leq \psi_{\hat\ell}(\hat \epsilon, \norm{\ensuremath{\mathbf x}}). \end{equation} In addition, if $\ensuremath{\mathbf x} \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}$ and $\epsilon \ge 0$ satisfy $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \leq \epsilon$ and $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \leq \epsilon$, then $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \leq \epsilon + \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x},\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}_{\hat \ell})$ and $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \leq \epsilon+\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x},\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}_{\hat \ell})$, which together with \eqref{eq:eps_hat_ell} gives \begin{align*} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}_{\hat \ell+1}) &\leq \psi_{\hat\ell}(\epsilon+\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x},\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}_{\hat \ell}) , \norm{\ensuremath{\mathbf x}})\\ &\leq\psi_{\hat\ell}(\epsilon+\varphi(\epsilon, \norm{\ensuremath{\mathbf x}}), \norm{\ensuremath{\mathbf x}})=(\psi_{\hat\ell} \diamondsuit \varphi)(\epsilon,\norm{\ensuremath{\mathbf x}}), \end{align*} where the second inequality follows from \eqref{eq:inductive}, the fact that $ \ensuremath{\mathcal{F}}_{\hat \ell} \subseteq \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}_{\hat \ell}$ and the monotonicity of $\psi_{\hat\ell}$. \end{proof} We now have all the pieces to state an error bound result for \eqref{eq:feas} that does not require any constraint qualifications. In comparison to \cite[Theorem~23]{L17}, it is not necessary that $ \ensuremath{\mathcal{K}}$ be amenable. \begin{theorem}[Error bound based on facial residual functions]\label{theo:err} Suppose \eqref{eq:feas} is feasible and let \[ \ensuremath{\mathcal{F}} _{\ell} \subsetneq \cdots \subsetneq \ensuremath{\mathcal{F}}_1 = \ensuremath{\mathcal{K}} \] be a chain of faces of $ \ensuremath{\mathcal{K}}$ together with $\ensuremath{\mathbf z}_i \in \ensuremath{\mathcal{F}} _i^*\cap \ensuremath{\mathcal{L}}^\perp \cap \{\ensuremath{\mathbf a}\}^\perp$ such that $ \ensuremath{\mathcal{F}} _{\ell}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ satisfy the PPS condition and $ \ensuremath{\mathcal{F}}_{i+1} = \ensuremath{\mathcal{F}} _i\cap \{\ensuremath{\mathbf z}_i\}^\perp$ for every $i$. For $i = 1,\ldots, \ell - 1$, let $\psi _{i}$ be a {facial residual function} for $ \ensuremath{\mathcal{F}}_{i}$ and $\ensuremath{\mathbf z}_i$ with respect to $ \ensuremath{\mathcal{K}}$. Then, there is a suitable positive rescaling of the $\psi _{i}$ (still denoted as $\psi_i$ by an abuse of notation) such that for any bounded set $B$ there is a positive constant $\kappa_B$ (depending on $B, \ensuremath{\mathcal{L}}, \ensuremath{\mathbf a}, \ensuremath{\mathcal{F}} _{\ell}$) such that \[ \ensuremath{\mathbf x} \in B\cap\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \leq \epsilon, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \leq \epsilon\quad \Rightarrow \quad \ensuremath{\operatorname{d}}\left(\ensuremath{\mathbf x}, ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \cap \ensuremath{\mathcal{K}}\right) \leq \kappa _B (\epsilon+\varphi(\epsilon,M)), \] where $M = \sup _{\ensuremath{\mathbf x} \in B} \norm{\ensuremath{\mathbf x}}$, $\varphi = \psi _{{\ell-1}}\diamondsuit \cdots \diamondsuit \psi_{{1}}$, if $\ell \geq 2$. If $\ell = 1$, we let $\varphi$ be the function satisfying $\varphi(\epsilon, M) = \epsilon$. \end{theorem} \begin{proof} The case $\ell = 1$ follows from Proposition \ref{prop:err_cq2}, by taking $ \ensuremath{\mathcal{F}} = \ensuremath{\mathcal{F}}_1$. Now, suppose $\ell \geq 2$. We apply Lemma \ref{lem:chain}, which tells us that, after positively rescaling the $\psi_i$, we have: \[ \ensuremath{\mathbf x}\in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \leq \epsilon, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \leq \epsilon \implies \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}} _{\ell}) \leq \varphi(\epsilon,\norm{\ensuremath{\mathbf x}}), \] where $\varphi = \psi _{{\ell-1}}\diamondsuit \cdots \diamondsuit \psi_{{1}} $. In particular, since $\norm{\ensuremath{\mathbf x}} \leq M$ for $\ensuremath{\mathbf x} \in B$ we have \begin{equation}\label{eq:aam_aux1} \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \leq \epsilon, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \leq \epsilon \implies \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}} _{\ell}) \leq \varphi(\epsilon,M), \qquad \forall \ensuremath{\mathbf x} \in B\cap (\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}). \end{equation} By assumption, $ \ensuremath{\mathcal{F}} _{\ell}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ satisfy the PPS condition. We invoke Proposition~\ref{prop:err_cq2} to find $ \kappa _B > 0$ such that \begin{equation}\label{eq:aam_aux2} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}} \cap ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a})) \leq \kappa_B(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}_{\ell}) + \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})), \qquad \forall \ensuremath{\mathbf x} \in B. \end{equation} Combining \eqref{eq:aam_aux1}, \eqref{eq:aam_aux2}, we conclude that if $\ensuremath{\mathbf x} \in B \cap (\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}})$ and $\epsilon\ge 0$ satisfy $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \leq \epsilon$ and $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \leq \epsilon$, then we have $\ensuremath{\operatorname{d}}\left(\ensuremath{\mathbf x}, ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \cap \ensuremath{\mathcal{K}}\right) \leq \kappa_B(\epsilon+\varphi(\epsilon,M))$. This completes the proof. \end{proof} Theorem~\ref{theo:err} is an abstract error bound result; whether some concrete inequality can be written down depends on obtaining a formula for the $\varphi$ function. To do so, it would require finding expressions for the facial residual functions. In the next subsections, we will address this challenge. \subsection{$\ensuremath{\mathfrak{g}}$-amenability}\label{ss:alpha-amenable} A result analogous to Theorem~\ref{theo:err} was obtained for the so-called \emph{amenable cones} in \cite[Theorem~23]{L17}. Although Theorem~\ref{theo:err} does not require amenability, we will see in this subsection that amenability may be helpful in the computation of the facial residual functions. However, in view of the fact that our main target cone (i.e., the exponential cone) is not \emph{amenable}, we need to consider the following generalization. \begin{definition}[$\ensuremath{\mathfrak{g}}$-amenability]\label{def:am} Let $ \ensuremath{\mathcal{K}}$ be a closed convex cone and let $\ensuremath{\mathfrak{g}}:\ensuremath{\mathbb{R}_+}\to\ensuremath{\mathbb{R}_+}$ be a monotone nondecreasing function such that $\ensuremath{\mathfrak{g}}(0) = 0$. We say that $ \ensuremath{\mathcal{F}}\mathrel{\unlhd} \ensuremath{\mathcal{K}} $ is a \emph{$\ensuremath{\mathfrak{g}}$-amenable} face if for every bounded set $B$, there exists $\kappa > 0$ such that \begin{equation}\label{eq:def_aam} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}) \leq \kappa \ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}})), \quad \forall \ensuremath{\mathbf x} \in (\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}) \cap B. \end{equation} If $ \ensuremath{\mathcal{F}} \mathrel{\unlhd} \ensuremath{\mathcal{K}}$ is $\ensuremath{\mathfrak{g}}$-amenable with $\ensuremath{\mathfrak{g}} = |\cdot|^\alpha$ for some $\alpha \in (0,1]$, then we say that $ \ensuremath{\mathcal{F}}$ is \emph{$\alpha$-amenable}. If all faces of $ \ensuremath{\mathcal{K}}$ are $\ensuremath{\mathfrak{g}}$-amenable (respectively, $\alpha$-amenable), then $ \ensuremath{\mathcal{K}}$ is said to be \emph{$\ensuremath{\mathfrak{g}}$-amenable} (respectively, $\alpha$-amenable). \end{definition} Recall that $\ensuremath{\operatorname{d}}(\beta \ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) = \beta\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}})$ if $\beta \geq 0$ and $ \ensuremath{\mathcal{K}}$ is a cone. Therefore, in the case of $1$-amenable faces, \eqref{eq:def_aam} holds globally. That is, a face $ \ensuremath{\mathcal{F}} \mathrel{\unlhd} \ensuremath{\mathcal{K}}$ is $1$-amenable if and only if there exists $\kappa > 0$ such that \begin{equation*} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}) \leq \kappa \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}), \quad \forall \ensuremath{\mathbf x} \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}. \end{equation*} Therefore, the notion of $1$-amenable cones coincides with the notion of amenable cones introduced in \cite{L17}. Next, we check how $\ensuremath{\mathfrak{g}}$-amenability behaves under direct-products. \begin{proposition}[$\ensuremath{\mathfrak{g}}$-amenability and products]\label{prop:direct_prod} Let $ \ensuremath{\mathcal{K}}^i \subseteq \ensuremath{\mathcal{E}}^i$ be closed convex cones for every $i \in \{1,\ldots,m\}$ and let $ \ensuremath{\mathcal{K}} = \ensuremath{\mathcal{K}}^1\times \cdots \times \ensuremath{\mathcal{K}}^m$. Let $ \ensuremath{\mathcal{F}} \mathrel{\unlhd} \ensuremath{\mathcal{K}}$ and suppose that $ \ensuremath{\mathcal{F}} = \ensuremath{\mathcal{F}}^1\times \cdots \times \ensuremath{\mathcal{F}}^m$ where $ \ensuremath{\mathcal{F}}^i \mathrel{\unlhd} \ensuremath{\mathcal{K}}^i$ is a $\ensuremath{\mathfrak{g}}$-amenable face of $ \ensuremath{\mathcal{K}}^i$ for every $i \in \{1,\ldots,m\}$. Then, $ \ensuremath{\mathcal{F}}$ is a $\ensuremath{\mathfrak{g}}$-amenable face of $ \ensuremath{\mathcal{K}}$. \end{proposition} \begin{proof} Let $B$ be a bounded set in $ \ensuremath{\mathcal{E}}^1\times \cdots \times \ensuremath{\mathcal{E}}^m$ and $B^i$ be the projection of $B$ onto $\ensuremath{\mathcal{E}}^i$. Since every $ \ensuremath{\mathcal{F}}^i$ is $\ensuremath{\mathfrak{g}}$-amenable, by taking the maximum of the constants appearing in \eqref{eq:def_aam}, we see that there exists $\kappa$ such that \begin{equation}\label{eq:d_prod_amb2} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}^i) \leq \kappa \ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}^i)), \quad \forall \ensuremath{\mathbf x} \in (\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}^i) \cap B^i, \,\, \forall i \in \{1,\ldots,m\}. \end{equation} To complete the proof, let $\ensuremath{\mathbf x} = (\ensuremath{\mathbf x}_1\ldots, \ensuremath{\mathbf x}_m) \in (\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}})\cap B$. Then, $\ensuremath{\mathbf x}_i \in (\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}^i)\cap B^i$ for every $i$. We have \begin{align*} &\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}) = \sqrt{\sum _{i=1}^m\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}_i, \ensuremath{\mathcal{F}}^{i})^2} \le {\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}_1, \ensuremath{\mathcal{F}}^{1}) + \cdots+ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}_m, \ensuremath{\mathcal{F}}^{m})}\\ &\overset{\rm (a)}\leq {\kappa \ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}_1, \ensuremath{\mathcal{K}}^1)) + \cdots + \kappa\ensuremath{\mathfrak{g}} (\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}_m, \ensuremath{\mathcal{K}}^m))}\overset{\rm (b)}\leq {\kappa \ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}})) + \cdots + \kappa\ensuremath{\mathfrak{g}} (\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}))} \leq m\kappa \ensuremath{\mathfrak{g}} (\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}})), \end{align*} where (a) holds because of \eqref{eq:d_prod_amb2}, (b) follows from the monotonicity of $\ensuremath{\mathfrak{g}}$. \end{proof} The notion of $\ensuremath{\mathfrak{g}}$-amenability has two helpful applications. The first is that if $ \ensuremath{\mathcal{K}}^1$ and $ \ensuremath{\mathcal{K}}^2$ are $\ensuremath{\mathfrak{g}}$-amenable, then we can obtain facial residual functions for $ \ensuremath{\mathcal{K}}^1 \times \ensuremath{\mathcal{K}}^2$ from the facial residual functions of $ \ensuremath{\mathcal{K}}^1$ and $ \ensuremath{\mathcal{K}}^2$. The second application will be discussed in Proposition~\ref{prop:frf_am}. \begin{proposition}[Facial residual functions for products in the presence of $\ensuremath{\mathfrak{g}}$-amenability]\label{prop:frf_prod} Let $ \ensuremath{\mathcal{K}}^i \subseteq \ensuremath{\mathcal{E}}^i$ be closed convex cones for every $i \in \{1,\ldots,m\}$ and let $ \ensuremath{\mathcal{K}} = \ensuremath{\mathcal{K}}^1 \times \cdots \times \ensuremath{\mathcal{K}}^m$. Let $ \ensuremath{\mathcal{F}} \mathrel{\unlhd} \ensuremath{\mathcal{K}}$, $\ensuremath{\mathbf z} \in \ensuremath{\mathcal{F}}^*$ and suppose that $ \ensuremath{\mathcal{F}} = \ensuremath{\mathcal{F}}^1\times \cdots \times \ensuremath{\mathcal{F}}^m$ where $ \ensuremath{\mathcal{F}}^i \mathrel{\unlhd} \ensuremath{\mathcal{K}}^i$ is a $\ensuremath{\mathfrak{g}}$-amenable face of $ \ensuremath{\mathcal{K}}^i$ for every $i \in \{1,\ldots,m\}$. Write $\ensuremath{\mathbf z} = (\ensuremath{\mathbf z}_1,\ldots,\ensuremath{\mathbf z}_m)$ with $\ensuremath{\mathbf z}_i \in ( \ensuremath{\mathcal{F}}^i)^*$. For every $i$, let $\psi_{ \ensuremath{\mathcal{F}}^i,\ensuremath{\mathbf z}_i}$ be a facial residual function for $ \ensuremath{\mathcal{F}}^i$ and $z_i$ with respect to $ \ensuremath{\mathcal{K}}^i$. Then, there exists a nonnegative monotone nondecreasing function $\sigma: \ensuremath{\mathbb R}_+ \to \ensuremath{\mathbb R}_+$ such that the function $\psi _{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}$ satisfying \[ \psi _{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}(\epsilon,t) = \sum _{i=1}^m \psi_{ \ensuremath{\mathcal{F}}^i,\ensuremath{\mathbf z}_i}(\sigma(t)\max\{\epsilon,\ensuremath{\mathfrak{g}}(2\epsilon) \},t) \] is a facial residual function for $ \ensuremath{\mathcal{F}}$ and $\ensuremath{\mathbf z}$ with respect to $ \ensuremath{\mathcal{K}}$. In addition, if all the $ \ensuremath{\mathcal{F}}^i$ are amenable faces (i.e., $1$-amenable), then $\sigma(\cdot)$ can be taken to be a constant function. \end{proposition} \begin{proof} Let $\rho > 0$ and suppose that $\ensuremath{\mathbf x} \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}$ and $\epsilon\ge 0$ satisfy $\norm{\ensuremath{\mathbf x}} \leq \rho$ and the inequalities \begin{equation}\label{eq:prod} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \leq \epsilon, \quad \inProd{\ensuremath{\mathbf x}}{\ensuremath{\mathbf z}} \leq \epsilon, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}} ) \leq \epsilon. \end{equation} We note that \[ \ensuremath{\mathcal{F}}\cap \{\ensuremath{\mathbf z}\}^{\perp} = ( \ensuremath{\mathcal{F}}^{1} \cap\{\ensuremath{\mathbf z}_1\}^\perp) \times \cdots \times ( \ensuremath{\mathcal{F}}^{m} \cap\{\ensuremath{\mathbf z}_m\}^\perp), \] and that for every $i \in \{1,\ldots,m\}$: \begin{align} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \leq \epsilon \quad & \Rightarrow\quad \ensuremath{\operatorname{d}}({\ensuremath{\mathbf x}}_i, \ensuremath{\mathcal{K}}^i) \leq \epsilon \label{eq:im1}\\ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x},\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}} ) \leq \epsilon\quad &\Rightarrow\quad \ensuremath{\operatorname{d}}({\ensuremath{\mathbf x}}_i,\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}^i) \leq \epsilon \label{eq:im2} \end{align} By the $\ensuremath{\mathfrak{g}}$-amenability of $ \ensuremath{\mathcal{F}}^i$, there exists a $\kappa_i > 0$ (depending on $\rho$), such that \begin{equation}\label{eq:f1_am} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf y}, \ensuremath{\mathcal{F}}^i) \leq \kappa_i \ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf y}, \ensuremath{\mathcal{K}}^i)), \quad \forall \ensuremath{\mathbf y} \in (\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}^i),\quad \norm{\ensuremath{\mathbf y}} \leq \rho. \end{equation} We have \begin{align} \ensuremath{\operatorname{d}}({\ensuremath{\mathbf x}}_i, \ensuremath{\mathcal{F}}^i) &\overset{\rm (a)}\leq \ensuremath{\operatorname{d}}({\ensuremath{\mathbf x}}_i,\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}^i) + \ensuremath{\operatorname{d}}(P_{\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}^i}({\ensuremath{\mathbf x}}_i), \ensuremath{\mathcal{F}}^i) \overset{\rm (b)}\leq \epsilon + \kappa_i\ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(P_{\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}^i}({\ensuremath{\mathbf x}}_i), \ensuremath{\mathcal{K}}^i)) \notag \\ & \overset{\rm (c)}\leq \epsilon + \kappa_i\ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}({\ensuremath{\mathbf x}}_i,\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}^i)+\ensuremath{\operatorname{d}}({\ensuremath{\mathbf x}}_i, \ensuremath{\mathcal{K}}^i)) \overset{\rm (d)}\leq \epsilon + \kappa_i \ensuremath{\mathfrak{g}}(2\epsilon), \label{eq:f1} \end{align} where (a) follows from \eqref{proj:p1}, (b) follows from \eqref{eq:im2}, \eqref{eq:f1_am} since $\norm{{\ensuremath{\mathbf x}}_i}\leq \norm{\ensuremath{\mathbf x}}\leq \rho$ and projections are nonexpansive operators, (c) follows from \eqref{proj:p2} and the monotonicity of $\ensuremath{\mathfrak{g}}$, and (d) is true because of \eqref{eq:im1} and \eqref{eq:im2}. Since $\ensuremath{\mathbf z}_i \in ( \ensuremath{\mathcal{F}}^i)^*$, we have from \eqref{eq:f1} that \begin{equation}\label{eq:p_f1} 0 \leq \inProd{\ensuremath{\mathbf z}_i}{P_{ \ensuremath{\mathcal{F}}^{i}}({\ensuremath{\mathbf x}}_i)} = \inProd{\ensuremath{\mathbf z}_i}{P_{ \ensuremath{\mathcal{F}}^{i}}({\ensuremath{\mathbf x}}_i) - {\ensuremath{\mathbf x}}_i + {\ensuremath{\mathbf x}}_i } \leq \norm{\ensuremath{\mathbf z}_i}(\epsilon + \kappa_i \ensuremath{\mathfrak{g}}(2\epsilon)) + \inProd{\ensuremath{\mathbf z}_i}{{\ensuremath{\mathbf x}}_i}. \end{equation} Using \eqref{eq:p_f1} for all $i$ and recalling that $\inProd{\ensuremath{\mathbf z}}{\ensuremath{\mathbf x}}\leq \epsilon$, we have \begin{align} \inProd{(\ensuremath{\mathbf z}_1,\ldots,\ensuremath{\mathbf z}_m)}{(P_{ \ensuremath{\mathcal{F}}^{1}}({\ensuremath{\mathbf x}}_1),\ldots,P_{ \ensuremath{\mathcal{F}}^{m}}({\ensuremath{\mathbf x}}_m)) } \leq \sum _{i=1}^m\left[\norm{\ensuremath{\mathbf z}_i}(\epsilon + \kappa_i \ensuremath{\mathfrak{g}}(2\epsilon)) + \inProd{\ensuremath{\mathbf z}_i}{\ensuremath{\mathbf x}_i}\right] \leq \hat\kappa \epsilon + \hat\kappa \ensuremath{\mathfrak{g}}(2\epsilon), \label{eq:p_f} \end{align} where $\hat\kappa = \max\{1+\sum_{i=1}^m\norm{\ensuremath{\mathbf z}_i},\sum_{i=1}^m\kappa_i\norm{\ensuremath{\mathbf z}_i}\}$. Since $\inProd{\ensuremath{\mathbf z}_i}{P_{ \ensuremath{\mathcal{F}}^{i}}(\ensuremath{\mathbf x}_i)} \geq 0$ for $i \in \{1,\ldots,m\}$, from \eqref{eq:p_f} we obtain \begin{equation}\label{eq:z_pf} \inProd{\ensuremath{\mathbf z}_i}{P_{ \ensuremath{\mathcal{F}}^{i}}(\ensuremath{\mathbf x}_i)} \leq \hat\kappa \epsilon + \hat\kappa \ensuremath{\mathfrak{g}}(2\epsilon), \qquad i \in \{1,\ldots,m\}. \end{equation} This implies that for $i \in\{1,\ldots,m\}$ we have \begin{equation*} \inProd{\ensuremath{\mathbf z}_i}{\ensuremath{\mathbf x}_i} = \inProd{\ensuremath{\mathbf z}_i}{\ensuremath{\mathbf x}_i -P_{ \ensuremath{\mathcal{F}}^{i}}(\ensuremath{\mathbf x}_i) + P_{ \ensuremath{\mathcal{F}}^{i}}(\ensuremath{\mathbf x}_i) } \leq \norm{\ensuremath{\mathbf z}}(\epsilon + \kappa_i \ensuremath{\mathfrak{g}}(2\epsilon)) + \hat\kappa \epsilon + \hat\kappa \ensuremath{\mathfrak{g}}(2\epsilon), \end{equation*} where the inequality follows from \eqref{eq:f1} and \eqref{eq:z_pf}. Then, letting $M_{\rho} = 1+2(\hat \kappa + \norm{\ensuremath{\mathbf z}} \max\{1, \kappa_1,\ldots, \kappa_m\})$, we have \begin{equation} \label{eq:zixi} \inProd{\ensuremath{\mathbf z}_i}{\ensuremath{\mathbf x}_i} \leq M_{\rho}\max\{\epsilon,\ensuremath{\mathfrak{g}}(2\epsilon)\}, \qquad i \in \{1,\ldots,m\}, \end{equation} where we write $M_{\rho}$ to emphasize the dependency on $\rho$. Now, recapitulating, the facial residual function $\psi_{ \ensuremath{\mathcal{F}}^i,\ensuremath{\mathbf z}_i}$ has the property that the relations \begin{align*} \ensuremath{\mathbf y}_i\in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}^{i},\quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf y}_i, \ensuremath{\mathcal{K}}^{i}) \leq \gamma_1,\quad \inProd{\ensuremath{\mathbf y}_i}{\ensuremath{\mathbf z}_i} \leq \gamma_2, \quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf y}_i, \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}^i ) \leq \gamma_3 \end{align*} imply $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf y}_i, \ensuremath{\mathcal{F}}^i\cap \{\ensuremath{\mathbf z}_i\}^\perp) \leq \psi_{ \ensuremath{\mathcal{F}}^i,\ensuremath{\mathbf z}_i}(\max_{1\leq j \leq 3}\{\gamma_j \},\norm{\ensuremath{\mathbf y}_i})$. Therefore, from \eqref{eq:im1}, \eqref{eq:im2}, \eqref{eq:zixi} and the monotonicity of $\psi_{ \ensuremath{\mathcal{F}}^i,\ensuremath{\mathbf z}_i}$, we conclude that \begin{equation}\label{eq:x_df} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}_i, \ensuremath{\mathcal{F}}^i\cap \{\ensuremath{\mathbf z}_i\}^\perp) \leq \psi_{ \ensuremath{\mathcal{F}}^i,\ensuremath{\mathbf z}_i}(M_{\rho}\max\{\epsilon,\ensuremath{\mathfrak{g}}(2\epsilon)\},\norm{\ensuremath{\mathbf x}_i}) \leq \psi_{ \ensuremath{\mathcal{F}}^i,\ensuremath{\mathbf z}_i}(M_{\rho}\max\{\epsilon,\ensuremath{\mathfrak{g}}(2\epsilon)\},\norm{\ensuremath{\mathbf x}}), \end{equation} where the first inequality follows from the fact that $M_{\rho} \geq 1$ so that $\max\{\epsilon, M_{\rho} \max\{\epsilon,\ensuremath{\mathfrak{g}}(2\epsilon)\} \} = M_{\rho}\max\{\epsilon,\ensuremath{\mathfrak{g}}(2\epsilon)\}$. From \eqref{eq:x_df}, we obtain that \begin{align} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}\cap \{\ensuremath{\mathbf z}\}^\perp) \le \sum _{i=1}^m{\ensuremath{\operatorname{d}}({\ensuremath{\mathbf x}}_i, \ensuremath{\mathcal{F}}^i \cap \{\ensuremath{\mathbf z}_i\}^\perp)} \leq \sum _{i=1}^m\psi_{ \ensuremath{\mathcal{F}}^i,\ensuremath{\mathbf z}_i}(M_{\rho}\max\{\epsilon,\ensuremath{\mathfrak{g}}(2\epsilon) \},\norm{\ensuremath{\mathbf x}}). \label{eq:res_sum} \end{align} We have shown that for every $\rho > 0$, there exists a constant $M_{\rho}$ such that if $\norm{\ensuremath{\mathbf x}} \leq \rho$, $\ensuremath{\mathbf x}\in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}$, and $\ensuremath{\mathbf x}$ and $\epsilon$ satisfy \eqref{eq:prod}, then \eqref{eq:res_sum} holds. From the monotonicity of the facial residual functions, if $M$ is greater than $M_{\rho}$, then \eqref{eq:res_sum} is still valid if we replace $M_{\rho}$ by $M$. Therefore, there exists a monotone nondecreasing function $\sigma:\ensuremath{\mathbb R}_+ \to \ensuremath{\mathbb R}_+$ with the property that \eqref{eq:res_sum} holds if $M_{\rho}$ is replaced\footnote{\label{foot:sigma}Let $\sigma _1 \coloneqq M_1$ and for every positive integer $i > 1$ let $\sigma _{i} \coloneqq \max\{\sigma_{i-1},M_i\}$. Then, we define $\sigma:\ensuremath{\mathbb R}_+ \to \ensuremath{\mathbb R}_+$ to be the function such that $\sigma(\rho) \coloneqq {\sigma _{\ceil{\rho}}} $, where $\ceil{\rho}$ is the smallest integer greater or equal than $\rho$. Then $\sigma$ has the required properties.} by $\sigma(\rho)$. We conclude that \begin{align*} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}\cap \{\ensuremath{\mathbf z}\}^\perp) \leq \psi _{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}(\epsilon,\norm{\ensuremath{\mathbf x}}) = \sum _{i=1}^m \psi_{ \ensuremath{\mathcal{F}}^i,\ensuremath{\mathbf z}_i}(\sigma(\norm{\ensuremath{\mathbf x}})\max\{\epsilon,\ensuremath{\mathfrak{g}}(2\epsilon) \},\norm{\ensuremath{\mathbf x}}) , \end{align*} where the function on right-hand-side has all the monotonicity properties required of facial residual functions. Finally, if all the $ \ensuremath{\mathcal{F}}^i$ are $1$-amenable, the fact that $\sigma(\cdot)$ can be taken to be a constant function follows from \cite[Proposition~17]{L17}. \end{proof} \subsection{How to compute facial residual functions?}\label{sec:frf_comp} In this section, we present some general tools for computing facial residual functions. \begin{lemma}[Facial residual functions from error bounds]\label{lem:facialresidualsbeta} Suppose that $ \ensuremath{\mathcal{K}}$ is a closed convex cone and let $\ensuremath{\mathbf z}\in \ensuremath{\mathcal{K}}^*$ be such that $ \ensuremath{\mathcal{F}} = \{\ensuremath{\mathbf z} \}^\perp\cap \ensuremath{\mathcal{K}}$ is a proper face of $ \ensuremath{\mathcal{K}}$. Let $\ensuremath{\mathfrak{g}}:\ensuremath{\mathbb R}_+\to\ensuremath{\mathbb R}_+$ be monotone nondecreasing with $\ensuremath{\mathfrak{g}}(0)=0$, and let $\kappa_{\ensuremath{\mathbf z},\ensuremath{\mathfrak{s}}}$ be a finite monotone nondecreasing nonnegative function in $\ensuremath{\mathfrak{s}}\in \ensuremath{\mathbb R}_+$ such that \begin{equation}\label{assumption:q} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{F}}) \leq \kappa_{\ensuremath{\mathbf z},\norm{\ensuremath{\mathbf q}}} \ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}}))\ \ \mbox{whenever}\ \ \ensuremath{\mathbf q} \in \{\ensuremath{\mathbf z}\}^\perp. \end{equation} Define the function $\psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}:\ensuremath{\mathbb{R}_+}\times \ensuremath{\mathbb{R}_+}\to \ensuremath{\mathbb{R}_+}$ by \begin{equation*} \psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}(s,t) := \max \left\{s,s/\|\ensuremath{\mathbf z}\| \right\} + \kappa_{\ensuremath{\mathbf z},t}\ensuremath{\mathfrak{g}} \left(s +\max \left\{s,s/\|\ensuremath{\mathbf z}\| \right\} \right). \end{equation*} Then we have \begin{equation}\label{haha} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf p}, \ensuremath{\mathcal{F}}) \leq \psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}(\epsilon,\norm{\ensuremath{\mathbf p}}) \mbox{\ \ \ \ whenever\ \ \ \ $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf p}, \ensuremath{\mathcal{K}}) \leq \epsilon$\ \ and\ \ $\inProd{\ensuremath{\mathbf p}}{\ensuremath{\mathbf z}} \leq \epsilon$.} \end{equation} Moreover, $\psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}$ is a facial residual function for $ \ensuremath{\mathcal{K}}$ and $\ensuremath{\mathbf z}$ with respect to $ \ensuremath{\mathcal{K}}$. \end{lemma} \begin{proof} Suppose that $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf p}, \ensuremath{\mathcal{K}}) \leq \epsilon$ and $\inProd{\ensuremath{\mathbf p}}{\ensuremath{\mathbf z}} \leq \epsilon$. We first claim that \begin{equation}\label{dpzperp} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf p},\{\ensuremath{\mathbf z} \}^\perp) \leq \max \left\{\epsilon,\epsilon/\|\ensuremath{\mathbf z}\| \right\}. \end{equation} This can be shown as follows. Since $\ensuremath{\mathbf z} \in \ensuremath{\mathcal{K}}^*$, we have $\inProd{\ensuremath{\mathbf p}+P_{ \ensuremath{\mathcal{K}}}(\ensuremath{\mathbf p})-\ensuremath{\mathbf p}}{\ensuremath{\mathbf z}} \geq 0$ and \[ \inProd{\ensuremath{\mathbf p}}{\ensuremath{\mathbf z}} \geq - \inProd{P_{ \ensuremath{\mathcal{K}}}(\ensuremath{\mathbf p})-\ensuremath{\mathbf p}}{\ensuremath{\mathbf z}} \geq -\epsilon \norm{\ensuremath{\mathbf z}}. \] We conclude that $|\langle \ensuremath{\mathbf p},\ensuremath{\mathbf z}\rangle| \leq \max\{\epsilon \norm{\ensuremath{\mathbf z}},\epsilon\}$. This, in combination with $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf p},\{\ensuremath{\mathbf z} \}^\perp) = |\langle \ensuremath{\mathbf p},\ensuremath{\mathbf z}\rangle|/\|\ensuremath{\mathbf z}\|$, leads to \eqref{dpzperp}. Next, let $\ensuremath{\mathbf q}:=P_{\{\ensuremath{\mathbf z} \}^\perp}\ensuremath{\mathbf p}$. Then we have that \begin{equation*} \begin{aligned} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf p}, \ensuremath{\mathcal{F}}) &\leq \|\ensuremath{\mathbf p} - \ensuremath{\mathbf q}\|+\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{F}})\overset{\rm (a)}\leq \max \left\{\epsilon,\epsilon/\|\ensuremath{\mathbf z}\| \right\} + \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{F}})\\ &\overset{\rm (b)}\leq \max \left\{\epsilon,\epsilon/\|\ensuremath{\mathbf z}\| \right\} + \kappa_{\ensuremath{\mathbf z},\norm{\ensuremath{\mathbf q}}} \ensuremath{\mathfrak{g}}\left(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}})\right)\\ &\overset{\rm (c)}\leq \max \left\{\epsilon,\epsilon/\|\ensuremath{\mathbf z}\| \right\} + \kappa_{\ensuremath{\mathbf z},\norm{\ensuremath{\mathbf p}}} \ensuremath{\mathfrak{g}}\left(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}})\right)\\ &\overset{\rm (d)}\leq \max \left\{\epsilon,\epsilon/\|\ensuremath{\mathbf z}\| \right\} + \kappa_{\ensuremath{\mathbf z},\norm{\ensuremath{\mathbf p}}} \ensuremath{\mathfrak{g}}\left(\epsilon +\max \left\{\epsilon,\epsilon/\|\ensuremath{\mathbf z}\| \right\} \right), \end{aligned} \end{equation*} where (a) follows from \eqref{dpzperp}, (b) is a consequence of \eqref{assumption:q}, (c) holds because $\|\ensuremath{\mathbf q}\| = \|P_{\{\ensuremath{\mathbf z} \}^\perp}\ensuremath{\mathbf p}\| \leq \|\ensuremath{\mathbf p}\|$ so that $\kappa_{\ensuremath{\mathbf z},\|\ensuremath{\mathbf q}\|}\leq \kappa_{\ensuremath{\mathbf z},\|\ensuremath{\mathbf p}\|}$, and (d) holds because $\ensuremath{\mathfrak{g}}$ is monotone nondecreasing and $$ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}}) \leq \ensuremath{\operatorname{d}}(\ensuremath{\mathbf p}, \ensuremath{\mathcal{K}}) + \|\ensuremath{\mathbf q} - \ensuremath{\mathbf p}\| \leq \epsilon + \max \left\{\epsilon,\epsilon/\|\ensuremath{\mathbf z}\| \right\}; $$ here, the second inequality follows from \eqref{dpzperp} and the assumption that $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf p}, \ensuremath{\mathcal{K}})\le \epsilon$. This proves \eqref{haha}. Finally, notice that $\psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}$ is nonnegative, monotone nondecreasing in each argument, and that $\psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}(0,\delta)=0$ for every $\delta \in \ensuremath{\mathbb R}_+$. Hence, $\psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}$ is a facial residual function for $ \ensuremath{\mathcal{K}}$ and $\ensuremath{\mathbf z}$ with respect to $ \ensuremath{\mathcal{K}}$. \end{proof} \begin{figure} \begin{center} \begin{tikzpicture}[scale=6] \draw[->,gray] (0,0) -- (1.1,0) node[right] {$x$-axis}; \draw[->,gray] (0,0) -- (0,1.0) node[above] {$y$-axis}; \draw[scale=1,domain=0:1,smooth,variable=\t,red] plot ({\t*exp(-\t)},{exp(-\t)}); \draw[scale=1,domain=0.001:1,smooth,variable=\u,blue] plot ({exp(-1/\u)/\u},{exp(-1/\u)}); \draw[scale=1,domain=0.1613:0.35,smooth,variable=\u,black] plot ({\u},{1.213*\u+-0.1956}); \draw[scale=1,domain=0.35:.45,smooth,variable=\u,black,dotted] plot ({\u},{1.213*\u+-0.1956}); \draw[scale=1,domain=0.45:.6,smooth,variable=\u,black] plot ({\u},{1.213*\u+-0.1956}); \draw[scale=1,domain=0.6:.8,smooth,variable=\u,black,dotted] plot ({\u},{1.213*\u+-0.1956}); \draw[scale=1,domain=0.8:1,smooth,variable=\u,black] plot ({\u},{1.213*\u+-0.1956}); \draw[scale=1,domain=0.0:0.5,smooth,variable=\u,purple,dashed] plot ({\u},{(1.61/2.94)*\u+0}); \shade[shading=ball, ball color=purple] (.294,.161) circle (0.015) node [above left,purple] {$\ensuremath{\mathbf u} = P_{\ensuremath{{\mathcal F}}}\ensuremath{\mathbf w}\;$}; \shade[shading=ball, ball color=black] (0.294,0.6239) circle (0.015) node [left,black] {$\ensuremath{\mathbf v}=P_{ \ensuremath{\mathcal{K}}}\ensuremath{\mathbf q}\;$}; \shade[shading=ball, ball color=black] (0.9699,0.9809) circle (0.015) node [right,black] {$\;\;\ensuremath{\mathbf q} \in \{\ensuremath{\mathbf z}\}^\perp\cap B(\eta)\backslash \ensuremath{\mathcal{F}}$}; \shade[shading=ball, ball color=black] (0.52137,0.43676) circle (0.015) node [below right,black] {$\ensuremath{\mathbf w}=P_{\{\mathbf{z} \}^\perp}\ensuremath{\mathbf v}$}; \draw[->,thick,blue] (0.915828,0.952340) -- (0.348072,0.652460); \draw[->,thick,blue] (0.339474,0.586472) -- (0.475896,0.474188); \draw[<-,thick,blue] (0.3271055 +0.025,0.2023640-0.025) -- (0.4872645+0.025,0.395390 -0.025); \end{tikzpicture} \end{center} \caption{The construction of $\ensuremath{\mathbf u},\ensuremath{\mathbf v},\ensuremath{\mathbf w}$ for one choice of $\ensuremath{\mathbf q}$. For describing a possibly higher dimensional problem in 2D, we represent $\{\ensuremath{\mathbf z}\}^\perp$ with a line, $\mathcal{K}$ with a horizontal slice, and $\ensuremath{{\mathcal F}}$ with a dot; of course, this is an oversimplification, since $\ensuremath{\mathbf q},\ensuremath{\mathbf w},\ensuremath{\mathbf u}$ are \textit{not} generically colinear, nor would any of the points necessarily lie in the same horizontal slice. The exact scenario shown is meant to suggest intuition, but it is not a plausible configuration of points.}\label{fig:uvw} \end{figure} In view of Lemma~\ref{lem:facialresidualsbeta}, one may construct facial residual functions after establishing the error bound \eqref{assumption:q}. In the next theorem, we present a characterization for the existence of such an error bound. Our result is based on the quantity \eqref{gammabetaeta} defined below being {\em nonzero}. Note that this quantity {\em does not} explicitly involve projections onto $ \ensuremath{\mathcal{K}}$; this enables us to work with the exponential cone later, whose projections do not seem to have simple expressions. \begin{theorem}[Characterization of the existence of error bounds]\label{thm:1dfacesmain} Suppose that $ \ensuremath{\mathcal{K}}$ is a closed convex cone and let $\ensuremath{\mathbf z}\in \ensuremath{\mathcal{K}}^*$ be such that $ \ensuremath{\mathcal{F}} = \{\ensuremath{\mathbf z} \}^\perp \cap \ensuremath{\mathcal{K}}$ is a nontrivial exposed face of $ \ensuremath{\mathcal{K}}$. Let $\eta \ge 0$, $\alpha \in (0,1]$ and let $\ensuremath{\mathfrak{g}}:\ensuremath{\mathbb R}_+\to \ensuremath{\mathbb R}_+$ be monotone nondecreasing with $\ensuremath{\mathfrak{g}}(0) = 0$ and $\ensuremath{\mathfrak{g}} \geq |\cdot|^\alpha$. Define \begin{equation}\label{gammabetaeta} \gamma_{\ensuremath{\mathbf z},\eta} := \inf_{\ensuremath{\mathbf v}} \left\{\frac{\ensuremath{\mathfrak{g}}(\|\ensuremath{\mathbf w}-\ensuremath{\mathbf v}\|)}{\|\ensuremath{\mathbf w}-\ensuremath{\mathbf u}\|}\;\bigg|\; \ensuremath{\mathbf v}\in \bd \ensuremath{\mathcal{K}}\cap B(\eta)\backslash \ensuremath{\mathcal{F}},\ \ensuremath{\mathbf w} = P_{\{\ensuremath{\mathbf z} \}^\perp}\ensuremath{\mathbf v},\ \ensuremath{\mathbf u} = P_{ \ensuremath{\mathcal{F}}}\ensuremath{\mathbf w},\ \ensuremath{\mathbf w}\neq \ensuremath{\mathbf u}\right\}. \end{equation} Then the following statements hold. \begin{enumerate}[{\rm (i)}] \item\label{thm:1dfacesmaini} If $\gamma_{\ensuremath{\mathbf z},\eta} \in (0,\infty]$, then it holds that \begin{equation}\label{haha2} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{F}}) \leq \kappa_{\ensuremath{\mathbf z},\eta} \ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}}))\ \ \mbox{whenever\ $\ensuremath{\mathbf q} \in \{\ensuremath{\mathbf z} \}^\perp \cap B(\eta)$}, \end{equation} where $\kappa_{\ensuremath{\mathbf z},\eta} := \max \left \{2\eta^{1-\alpha}, 2\gamma_{\ensuremath{\mathbf z},\eta}^{-1} \right \} < \infty$. \item\label{thm:1dfacesmainii} If there exists $\kappa_{_B} \in (0,\infty)$ so that \begin{equation}\label{hahaww2} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{F}}) \leq \kappa_{_B} \ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}}))\ \ \mbox{whenever\ $\ensuremath{\mathbf q} \in \{\ensuremath{\mathbf z} \}^\perp \cap B(\eta)$}, \end{equation} then $\gamma_{\ensuremath{\mathbf z},\eta} \in (0,\infty]$. \end{enumerate} \end{theorem} \begin{proof} We first consider item (i). If $\eta = 0$ or $\ensuremath{\mathbf q} \in \ensuremath{\mathcal{F}}$, the result is vacuously true, so let $\eta > 0$ and $\ensuremath{\mathbf q} \in \{\ensuremath{\mathbf z} \}^\perp \cap B(\eta) \backslash \ensuremath{\mathcal{F}}$. Then $\ensuremath{\mathbf q}\notin \ensuremath{\mathcal{K}}$ because $ \ensuremath{\mathcal{F}} = \{\ensuremath{\mathbf z}\}^\perp\cap \ensuremath{\mathcal{K}}$. Define\footnote{A configuration of these points is shown in Figure~\ref{fig:uvw}.} \[ \ensuremath{\mathbf v}=P_{ \ensuremath{\mathcal{K}}}\ensuremath{\mathbf q},\quad \ensuremath{\mathbf w} = P_{\{\ensuremath{\mathbf z} \}^\perp}\ensuremath{\mathbf v},\quad \text{and}\quad \ensuremath{\mathbf u} = P_{ \ensuremath{\mathcal{F}}}\ensuremath{\mathbf w}. \] Then $\ensuremath{\mathbf v}\in \bd \ensuremath{\mathcal{K}}\cap B(\eta)$ because $\ensuremath{\mathbf q}\notin \ensuremath{\mathcal{K}}$ and $\|\ensuremath{\mathbf q}\|\le \eta$. If $\ensuremath{\mathbf v}\in \ensuremath{\mathcal{F}}$, then we have $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{F}}) = \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}})$ and hence \[ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{F}}) = \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}})^{1-\alpha}\cdot \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}})^\alpha\le \eta^{1-\alpha}\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}})^\alpha\le \kappa_{\ensuremath{\mathbf z},\eta} \ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}})), \] where the first inequality holds because $\norm{\ensuremath{\mathbf q}}\le \eta$, and the last inequality follows from the definitions of $\ensuremath{\mathfrak{g}}$ and $\kappa_{\ensuremath{\mathbf z},\eta}$. Thus, from now on, we assume that $\ensuremath{\mathbf v}\in \bd \ensuremath{\mathcal{K}}\cap B(\eta)\backslash \ensuremath{\mathcal{F}}$. Next, since $\ensuremath{\mathbf w}=P_{\{\ensuremath{\mathbf z}\}^\perp}\ensuremath{\mathbf v}$, it holds that $\ensuremath{\mathbf v}-\ensuremath{\mathbf w}\in \{\ensuremath{\mathbf z}\}^{\perp\perp}$ and hence $\|\ensuremath{\mathbf q}-\ensuremath{\mathbf v}\|^2 = \|\ensuremath{\mathbf q}-\ensuremath{\mathbf w}\|^2+\|\ensuremath{\mathbf w}-\ensuremath{\mathbf v}\|^2$. In particular, we have \begin{equation}\label{cinequality_infty} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}}) = \|\ensuremath{\mathbf q}-\ensuremath{\mathbf v}\| \geq \max\{\|\ensuremath{\mathbf v}-\ensuremath{\mathbf w}\|,\|\ensuremath{\mathbf q}-\ensuremath{\mathbf w}\|\}, \end{equation} where the equality follows from the definition of $\ensuremath{\mathbf v}$. Now, to establish \eqref{haha2}, we consider two cases. \begin{enumerate}[(I)] \item\label{thMbetafacescase1_infty} $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{F}}) \leq 2\ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}, \ensuremath{\mathcal{F}})$; \item\label{thMbetafacescase2_infty} $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{F}}) > 2\ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}, \ensuremath{\mathcal{F}})$. \end{enumerate} \ref{thMbetafacescase1_infty}: In this case, we have from $\ensuremath{\mathbf u} = P_{ \ensuremath{\mathcal{F}}}\ensuremath{\mathbf w}$ and $\ensuremath{\mathbf q}\notin \ensuremath{\mathcal{F}}$ that \begin{equation}\label{haha3} 2\|\ensuremath{\mathbf w} - \ensuremath{\mathbf u}\|= 2\ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}, \ensuremath{\mathcal{F}}) \ge \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{F}}) > 0, \end{equation} where the first inequality follows from the assumption in this case \ref{thMbetafacescase1_infty}. Hence, \begin{equation}\label{biginequalitybetageneral_infty} \frac{1}{\kappa_{\ensuremath{\mathbf z},\eta}} \stackrel{\rm (a)}{\le} \frac{1}{2}\gamma_{\ensuremath{\mathbf z},\eta} \stackrel{\rm (b)}{\leq} \frac{\ensuremath{\mathfrak{g}}(\|\ensuremath{\mathbf w}-\ensuremath{\mathbf v}\|)}{2\|\ensuremath{\mathbf w}-\ensuremath{\mathbf u}\|} \stackrel{\rm (c)}{\leq} \frac{\ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}}))}{2\|\ensuremath{\mathbf w}-\ensuremath{\mathbf u}\|}\stackrel{\rm (d)}{\leq} \frac{\ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}}))}{\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{F}})}, \end{equation} where (a) is true by the definition of $\kappa_{\ensuremath{\mathbf z},\eta}$, (b) uses the condition that $\ensuremath{\mathbf v}\in \bd \ensuremath{\mathcal{K}}\cap B(\eta)\backslash \ensuremath{\mathcal{F}}$, \eqref{haha3} and the definition of $\gamma_{\ensuremath{\mathbf z},\eta}$, (c) is true by \eqref{cinequality_infty} and the monotonicity of $\ensuremath{\mathfrak{g}}$, and (d) follows from \eqref{haha3}. This concludes case \ref{thMbetafacescase1_infty}.\footnote{In particular, in view of \eqref{biginequalitybetageneral_infty}, we see that this case only happens when $\gamma_{\ensuremath{\mathbf z},\eta} < \infty$.} \noindent\ref{thMbetafacescase2_infty}: Using the triangle inequality, we have \begin{equation*} 2\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{F}}) \leq 2\|\ensuremath{\mathbf q}-\ensuremath{\mathbf w}\|+2\ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}, \ensuremath{\mathcal{F}})< 2\|\ensuremath{\mathbf q}-\ensuremath{\mathbf w}\|+\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{F}}), \end{equation*} where the strict inequality follows from the condition for this case \ref{thMbetafacescase2_infty}. Consequently, we have $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{F}}) \leq 2\|\ensuremath{\mathbf q}-\ensuremath{\mathbf w}\|$. Combining this with \eqref{cinequality_infty}, we deduce further that \[ \begin{aligned} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{F}}) &\le 2\|\ensuremath{\mathbf q}-\ensuremath{\mathbf w}\|\le 2 \max\{\|\ensuremath{\mathbf v}-\ensuremath{\mathbf w}\|,\|\ensuremath{\mathbf q}-\ensuremath{\mathbf w}\|\} \le 2\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}}) = 2\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}})^{1-\alpha}\cdot \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}})^\alpha\\ & \le 2\eta^{1-\alpha}\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}})^\alpha\le \kappa_{\ensuremath{\mathbf z},\eta} \ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{\mathcal{K}})), \end{aligned} \] where the fourth inequality holds because $\norm{\ensuremath{\mathbf q}}\le \eta$, and the last inequality follows from the definitions of $\ensuremath{\mathfrak{g}}$ and $\kappa_{\ensuremath{\mathbf z},\eta}$. This proves item (i). We next consider item (ii). Again, the result is vacuously true if $\eta = 0$, so let $\eta > 0$. Let $\ensuremath{\mathbf v}\in \bd \ensuremath{\mathcal{K}}\cap B(\eta)\backslash \ensuremath{\mathcal{F}}$, $\ensuremath{\mathbf w} = P_{\{\ensuremath{\mathbf z} \}^\perp}\ensuremath{\mathbf v}$ and $\ensuremath{\mathbf u} = P_{ \ensuremath{\mathcal{F}}}\ensuremath{\mathbf w}$ with $\ensuremath{\mathbf w}\neq \ensuremath{\mathbf u}$. Then $\ensuremath{\mathbf w}\in B(\eta)$, and we have in view of \eqref{hahaww2} that \[ \|\ensuremath{\mathbf w}-\ensuremath{\mathbf u}\| \overset{\rm (a)}= \ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}, \ensuremath{\mathcal{F}}) \overset{\rm (b)}\le \kappa_{_B}\ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}, \ensuremath{\mathcal{K}})) \overset{\rm (c)}\le \kappa_{_B}\ensuremath{\mathfrak{g}}(\|\ensuremath{\mathbf w} - \ensuremath{\mathbf v}\|), \] where (a) holds because $\ensuremath{\mathbf u} = P_{ \ensuremath{\mathcal{F}}}\ensuremath{\mathbf w}$, (b) holds because of \eqref{hahaww2}, $\ensuremath{\mathbf w}\in \{\ensuremath{\mathbf z}\}^\perp$ and $\|\ensuremath{\mathbf w}\|\le \eta$, and (c) is true because $\ensuremath{\mathfrak{g}}$ is monotone nondecreasing and $\ensuremath{\mathbf v}\in \ensuremath{\mathcal{K}}$. Thus, we have $\gamma_{\ensuremath{\mathbf z},\eta} \ge 1/\kappa_{_B} > 0$. This completes the proof. \end{proof} \begin{remark}[About $\kappa_{\ensuremath{\mathbf z},\eta}$ and $\gamma_{\ensuremath{\mathbf z},\eta}^{-1}$]\label{rem:kappa} As $\eta$ increases, the infimum in \eqref{gammabetaeta} is taken over a larger region, so $\gamma_{\ensuremath{\mathbf z},\eta}$ does not increase. Accordingly, $\gamma_{\ensuremath{\mathbf z},\eta}^{-1}$ does not decrease when $\eta$ increases. Therefore, the $\kappa_{\ensuremath{\mathbf z},\eta}$ and $\gamma_{\ensuremath{\mathbf z},\eta}^{-1}$ considered in Theorem~\ref{thm:1dfacesmain} are monotone nondecreasing as function of $\eta$ when $\ensuremath{\mathbf z}$ is fixed. We are also using the convention that $1/\infty = 0$ so that $\kappa_{\ensuremath{\mathbf z},\eta} = 2\eta^{1-\alpha}$ when $\gamma_{\ensuremath{\mathbf z},\eta} = \infty$. \end{remark} Thus, to establish an error bound as in \eqref{haha2}, it suffices to show that $\gamma_{\ensuremath{\mathbf z},\eta} \in (0,\infty]$ for the choice of $\ensuremath{\mathfrak{g}}$ and $\eta\ge 0$. Clearly, $\gamma_{\ensuremath{\mathbf z},0} = \infty$. The next lemma allows us to check whether $\gamma_{\ensuremath{\mathbf z},\eta} \in (0,\infty]$ for an $\eta > 0$ by considering {\em convergent} sequences. \begin{lemma}\label{lem:infratio} Suppose that $ \ensuremath{\mathcal{K}}$ is a closed convex cone and let $\ensuremath{\mathbf z}\in \ensuremath{\mathcal{K}}^*$ be such that $ \ensuremath{\mathcal{F}} = \{\ensuremath{\mathbf z} \}^\perp \cap \ensuremath{\mathcal{K}}$ is a nontrivial exposed face of $ \ensuremath{\mathcal{K}}$. Let $\eta > 0$, $\alpha \in (0,1]$ and let $\ensuremath{\mathfrak{g}}:\ensuremath{\mathbb R}_+\to \ensuremath{\mathbb R}_+$ be monotone nondecreasing with $\ensuremath{\mathfrak{g}}(0) = 0$ and $\ensuremath{\mathfrak{g}} \geq |\cdot|^\alpha$. Let $\gamma_{\ensuremath{\mathbf z},\eta}$ be defined as in \eqref{gammabetaeta}. If $\gamma_{\ensuremath{\mathbf z},\eta} = 0$, then there exist $\ensuremath{\bar{\vv}} \in \ensuremath{\mathcal{F}}$ and a sequence $\{\ensuremath{\mathbf v}^k\}\subset \bd \ensuremath{\mathcal{K}}\cap B(\eta) \backslash \ensuremath{\mathcal{F}}$ such that \begin{subequations}\label{infratiohk} \begin{align} \underset{k \rightarrow \infty}{\lim}\ensuremath{\mathbf v}^k = \underset{k \rightarrow \infty}{\lim}\ensuremath{\mathbf w}^k &= \ensuremath{\bar{\vv}} \label{infratiohka} \\ {\rm and}\ \ \ \lim_{k\rightarrow\infty} \frac{\ensuremath{\mathfrak{g}}(\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf v}^k\|)}{\|\ensuremath{\mathbf w}^k- \ensuremath{\mathbf u}^k\|} &= 0, \label{infratiohkb} \end{align} \end{subequations} where $\ensuremath{\mathbf w}^k = P_{\{\ensuremath{\mathbf z}\}^\perp}\ensuremath{\mathbf v}^k$, $\ensuremath{\mathbf u}^k = P_{ \ensuremath{\mathcal{F}}}\ensuremath{\mathbf w}^k$ and $\ensuremath{\mathbf w}^k\neq \ensuremath{\mathbf u}^k$. \end{lemma} \begin{proof} Suppose that $\gamma_{\ensuremath{\mathbf z},\eta} = 0$. Then, by the definition of infimum, there exists a sequence $\{\ensuremath{\mathbf v}^k\}\subset\bd \ensuremath{\mathcal{K}}\cap B(\eta) \backslash \ensuremath{\mathcal{F}}$ such that \begin{equation}\label{infratio1} \underset{k \rightarrow \infty}{\lim} \frac{\ensuremath{\mathfrak{g}}(\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf v}^k\|)}{\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf u}^k\|} = 0, \end{equation} where $\ensuremath{\mathbf w}^k = P_{\{\ensuremath{\mathbf z}\}^\perp}\ensuremath{\mathbf v}^k$, $\ensuremath{\mathbf u}^k = P_{ \ensuremath{\mathcal{F}}}\ensuremath{\mathbf w}^k$ and $\ensuremath{\mathbf w}^k\neq \ensuremath{\mathbf u}^k$. Since $\{\ensuremath{\mathbf v}^k\}\subset B(\eta)$, by passing to a convergent subsequence if necessary, we may assume without loss of generality that \begin{equation}\label{infratio4} \underset{k \rightarrow \infty}{\lim}\ensuremath{\mathbf v}^k = \ensuremath{\bar{\vv}} \end{equation} for some $\ensuremath{\bar{\vv}}\in \ensuremath{\mathcal{K}}\cap B(\eta)$. In addition, since $0\in \ensuremath{\mathcal{F}}\subseteq \{\ensuremath{\mathbf z}\}^\perp$, and projections onto closed convex sets are nonexpansive, we see that $\{\ensuremath{\mathbf w}^k\}\subset B(\eta)$ and $\{\ensuremath{\mathbf u}^k\}\subset B(\eta)$, and hence the sequence $\{\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf u}^k\|\}$ is bounded. Then we can conclude from \eqref{infratio1} and the assumption $\ensuremath{\mathfrak{g}} \geq |\cdot|^\alpha$ that \begin{equation}\label{infratio3} \underset{k \rightarrow \infty}{\lim}\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\| =0. \end{equation} Now \eqref{infratio3}, \eqref{infratio4}, and the triangle inequality give $\ensuremath{\mathbf w}^k\to \ensuremath{\bar{\vv}}$. Since $\{\ensuremath{\mathbf w}^k\}\subset \{\ensuremath{\mathbf z}\}^\perp$, it then follows that $\ensuremath{\bar{\vv}} \in \{\ensuremath{\mathbf z} \}^\perp$. Thus, $\ensuremath{\bar{\vv}}\in \{\ensuremath{\mathbf z} \}^\perp \cap \ensuremath{\mathcal{K}}= \ensuremath{\mathcal{F}}$. This completes the proof. \end{proof} Let $ \ensuremath{\mathcal{K}}$ be a closed convex cone. Lemma~\ref{lem:facialresidualsbeta}, Theorem~\ref{thm:1dfacesmain} and Lemma~\ref{lem:infratio} are tools to obtain facial residual functions for $ \ensuremath{\mathcal{K}}$ when the $ \ensuremath{\mathcal{F}}$ in Definition \ref{def:ebtp} is the whole $ \ensuremath{\mathcal{K}}$ itself and $\ensuremath{\mathbf z} \in \ensuremath{\mathcal{K}}^*$. However, when $ \ensuremath{\mathcal{F}}$ is not $ \ensuremath{\mathcal{K}}$, these results cannot be used directly. In this final part of Section~\ref{sec:frf}, we will see how $\ensuremath{\mathfrak{g}}$-amenability may help to bridge this gap. \begin{proposition}[From FRFs of a face to FRFs for the whole cone]\label{prop:frf_am} Let $ \ensuremath{\mathcal{K}}$ be a closed convex cone and let $ \ensuremath{\mathcal{F}} \mathrel{\unlhd} \ensuremath{\mathcal{K}}$ be a $\ensuremath{\mathfrak{g}}$-amenable face. Let $\ensuremath{\mathbf z} \in \ensuremath{\mathcal{F}}^*$ and let $\psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}$ be a facial residual function for $ \ensuremath{\mathcal{F}}$ and $\ensuremath{\mathbf z}$ with respect to $ \ensuremath{\mathcal{F}}$. Then, there exists a monotone nondecreasing function $\sigma: \ensuremath{\mathbb R}_+ \to \ensuremath{\mathbb R}_+$ such that the function $\hat \psi _{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}$ satisfying \[ \hat \psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}(\epsilon,t) = \epsilon + \psi _{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}(\sigma(t)\max\{\epsilon,\ensuremath{\mathfrak{g}}(2\epsilon)\},t) \] is a {facial residual function} for $ \ensuremath{\mathcal{F}}$ and $\ensuremath{\mathbf z}$ with respect to $ \ensuremath{\mathcal{K}}$. \end{proposition} \begin{proof} Since $\psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}$ is a facial residual function for $ \ensuremath{\mathcal{F}}$ and $\ensuremath{\mathbf z}$ with respect to $ \ensuremath{\mathcal{F}}$, we have \begin{equation}\label{eq:frf_am1} \ensuremath{\mathbf y} \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}, \,\, \ensuremath{\operatorname{d}}(\ensuremath{\mathbf y}, \ensuremath{\mathcal{F}}) \leq \epsilon, \,\, \inProd{\ensuremath{\mathbf y}}{\ensuremath{\mathbf z}} \leq \epsilon\quad \Longrightarrow\quad \ensuremath{\operatorname{d}}(\ensuremath{\mathbf y}, \ensuremath{\mathcal{F}}\cap \{\ensuremath{\mathbf z}\}^\perp) \leq\ \psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}} (\epsilon, \norm{\ensuremath{\mathbf y}}). \end{equation} Also, since $ \ensuremath{\mathcal{F}}$ is a $\ensuremath{\mathfrak{g}}$-amenable face of $ \ensuremath{\mathcal{K}}$, for every $\rho \geq 0$, there exists $\kappa _{\rho} > 0$ such that \begin{equation}\label{eq:frf_am} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf y}, \ensuremath{\mathcal{F}}) \leq \kappa _{\rho} \ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf y}, \ensuremath{\mathcal{K}})), \qquad \forall \ensuremath{\mathbf y} \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}},\ \ \ \norm{\ensuremath{\mathbf y}} \leq \rho. \end{equation} Now, suppose that $\ensuremath{\mathbf x} \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}$ and $\epsilon\ge 0$ satisfy $\norm{\ensuremath{\mathbf x}} \leq \rho$ and \begin{equation}\label{eq:frf_x} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \leq \epsilon, \,\, \inProd{\ensuremath{\mathbf x}}{\ensuremath{\mathbf z}} \leq \epsilon, \,\, \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}} ) \leq \epsilon. \ \end{equation} The goal is to find a bound on $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}\cap \{\ensuremath{\mathbf z}\}^\perp)$ using $\psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}$. First, we note that \eqref{eq:frf_x} implies \begin{equation}\label{eq:px_z} \inProd{P_{\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}}(\ensuremath{\mathbf x})}{\ensuremath{\mathbf z}} = \inProd{P_{\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}}(\ensuremath{\mathbf x}) - \ensuremath{\mathbf x} + \ensuremath{\mathbf x}}{\ensuremath{\mathbf z}} \leq (1+\norm{\ensuremath{\mathbf z}})\epsilon. \end{equation} Since $0 \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}$, we have $\|P_{\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}}(\ensuremath{\mathbf x})\| \leq \norm{\ensuremath{\mathbf x}} \leq \rho$. Therefore, \begin{align}\label{eq:px_f} \ensuremath{\operatorname{d}}(P_{\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}}(\ensuremath{\mathbf x}), \ensuremath{\mathcal{F}}) \overset{\rm (a)}\leq \kappa _{\rho}\ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(P_{\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}}(\ensuremath{\mathbf x}), \ensuremath{\mathcal{K}})) \overset{\rm (b)}\leq \kappa _{\rho}\ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x},\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}})+\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}})) \overset{\rm (c)}\leq \kappa _{\rho} \ensuremath{\mathfrak{g}}(2\epsilon), \end{align} where (a) follows from \eqref{eq:frf_am}, (b) follows from \eqref{proj:p2} and the monotonicity of $\ensuremath{\mathfrak{g}}$, and (c) follows from \eqref{eq:frf_x}. Let $\hat\kappa_{\rho} \coloneqq \max\{\kappa_\rho,1+\norm{\ensuremath{\mathbf z}} \}$. Then, combining \eqref{eq:px_z}, \eqref{eq:px_f} and \eqref{eq:frf_am1} we obtain \begin{equation}\label{eq:px_hatf} \ensuremath{\operatorname{d}}(P_{\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}}(\ensuremath{\mathbf x}), \ensuremath{\mathcal{F}}\cap \{\ensuremath{\mathbf z}\}^\perp) \leq \psi _{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}(\hat\kappa_{\rho}\max\{\epsilon,\ensuremath{\mathfrak{g}}{(2\epsilon)}\},\rho ). \end{equation} Finally, from \eqref{proj:p1}, \eqref{eq:frf_x} and \eqref{eq:px_hatf} we have \begin{align} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{F}}\cap \{\ensuremath{\mathbf z}\}^\perp) \leq \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x},\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}) + \ensuremath{\operatorname{d}}(P_{\ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}}(\ensuremath{\mathbf x}), \ensuremath{\mathcal{F}}\cap \{\ensuremath{\mathbf z}\}^\perp) \leq \epsilon + \psi _{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}(\hat\kappa_{\rho}\max\{\epsilon,\ensuremath{\mathfrak{g}}{(2\epsilon)}\},\rho ) \label{eq:x_hatf}. \end{align} Thus we have shown that for every $\rho > 0$, there exists a constant $\hat\kappa_{\rho}$ such that if $\norm{\ensuremath{\mathbf x}} \leq \rho$, $\ensuremath{\mathbf x}\in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{K}}$, and $\ensuremath{\mathbf x}$ and $\epsilon$ satisfy \eqref{eq:frf_x}, then \eqref{eq:x_hatf} holds. From the monotonicity of $\psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}$, if $M$ is greater than $\hat\kappa_{\rho}$ then \eqref{eq:x_hatf} is still valid if we replace $\hat\kappa_{\rho}$ by $M$. Therefore, there exists a monotone nondecreasing function $\sigma:\ensuremath{\mathbb R}_+ \to \ensuremath{\mathbb R}_+$ with the property that \eqref{eq:x_hatf} holds if $\hat\kappa_{\rho}$ is replaced by $\sigma(\rho)$, see footnote~\ref{foot:sigma}. With that, let $\hat \psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}$ be such that \[ \hat \psi _{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}(\epsilon,t) = \epsilon + \psi _{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}(\sigma(t)\max\{\epsilon,\ensuremath{\mathfrak{g}}{(2\epsilon)}\},t), \] then $\hat \psi_{ \ensuremath{\mathcal{F}},\ensuremath{\mathbf z}}$ is a facial residual function for $ \ensuremath{\mathcal{F}}$ and $\ensuremath{\mathbf z}$ with respect to $ \ensuremath{\mathcal{K}}$. \end{proof} \section{The exponential cone}\label{sec:exp_cone} In this section, we will use all the techniques developed so far to obtain error bounds for the 3D exponential cone $\ensuremath{K_{\exp}}$. We will start with a study of its facial structure in Section~\ref{sec:facial_structure}, then we will compute its facial residual functions and analyze $\ensuremath{K_{\exp}}$ from the point of view of $\ensuremath{\mathfrak{g}}$-amenability in Section~\ref{sec:exp_frf}. Finally, error bounds will be presented in Section~\ref{sec:exp_err}. In Section~\ref{sec:odd}, we summarize odd behaviour found in the facial structure of the exponential cone. \subsection{Facial structure}\label{sec:facial_structure} Recall that the exponential cone is defined as follows: \begin{align}\label{d:Kexp} K_{\exp}:=&\left \{(x,y,z)\;|\;y>0,z\geq ye^{x/y}\right \} \cup \left \{(x,y,z)\;|\; x \leq 0, z\geq 0, y=0 \right \}. \end{align} \begin{figure} \begin{center} \begin{tikzpicture} [scale=0.76] \node[anchor=south west,inner sep=0] (image) at (0,0) {\includegraphics[width=.75\linewidth]{cone_and_dual_paper}}; \begin{scope}[x={(image.south east)},y={(image.north west)}] \node [left] at (0.05,0.4) {$\ensuremath{{\mathcal F}}_{\beta>1}$}; \draw [->,thick] (0.05,0.4) -- (0.1,0.425); \node [above] at (0,0.72) {$\ensuremath{{\mathcal F}}_{\beta=1}$}; \draw [->,thick] (0,0.72) -- (0.01,0.68); \node [above] at (0.04,0.82) {$\ensuremath{{\mathcal F}}_{\beta \in \left(0,1 \right) }$}; \draw [->,thick] (0.04,0.82) -- (0.05,0.76); \node [above] at (0.12,.88) {$\ensuremath{{\mathcal F}}_{\beta =0 }$}; \draw [->,thick] (0.12,.88) -- (0.14,0.84); \node [above] at (0.25,.88) {$\ensuremath{{\mathcal F}}_{\beta <0 }$}; \draw [->,thick] (0.25,.88) -- (0.25,0.85); \node [below] at (0.55,.07) {$\ensuremath{{\mathcal F}}_{\infty}$}; \draw [->,thick] (0.55,.07) -- (0.55,0.15); \node [above] at (0.48,.6) {$\ensuremath{{\mathcal F}}_{-\infty}$}; \draw [->,thick] (0.48,.6) -- (0.48,0.5); \node [above] at (0.6,.62) {$\ensuremath{\mathbf z}_{\beta>1}$}; \draw [->,thick] (0.6,.62) -- (0.6,0.58); \node [above] at (0.75,.60) {$\ensuremath{\mathbf z}_{\beta=1}$}; \draw [->,thick] (0.75,.60) -- (0.75,0.55); \node [above] at (0.88,.56) {$\ensuremath{\mathbf z}_{\beta \in \left(0,1 \right)}$}; \draw [->,thick] (0.88,.56) -- (0.88,0.51); \node [right] at (0.79,.16) {$\ensuremath{\mathbf z}_{\beta <0}$}; \draw [->,thick] (0.79,.16) -- (0.74,0.19); \node [right] at (0.87,.25) {$\ensuremath{\mathbf z}_{\beta =0}$}; \draw [->,thick] (0.87,.25) -- (0.83,0.26); \node [below] at (0.15,.11) {$\ensuremath{\mathbf z}_{\beta =-\infty}$}; \draw [->,thick] (0.15,.11) -- (0.15,0.16); \node [right] at (0.45,.72) {$\ensuremath{\mathbf z}_{\beta =\infty}, \, \ensuremath{\operatorname{cone}}\{\ensuremath{\mathbf z}_{\infty}\} = \ensuremath{{\mathcal F}}_{ne}$}; \draw [->,thick] (0.45,.72) -- (0.42,0.72); \end{scope} \end{tikzpicture} \end{center} \caption{The exponential cone and its dual, with faces and exposing vectors labeled according to our index $\beta$.}\label{fig:exp} \end{figure} \noindent Its dual cone is given by \begin{align*} K_{\exp}^*:=&\left \{(x,y,z)\;|\;x<0, ez\geq -xe^{y/x}\right \} \cup \left \{(x,y,z)\;|\; x = 0, z\geq 0, y\geq0 \right \}. \end{align*} It may therefore be readily seen that $K_{\exp}^*$ is a scaled and rotated version of $K_{\exp}$. In this subsection, we will describe the nontrivial faces of $K_{\exp}$; see Figure~\ref{fig:exp}. We will show that we have the following types of nontrivial faces: \begin{enumerate}[{\rm (a)}] \item infinitely many exposed extreme rays (1D faces) parametrized by $\beta \in \ensuremath{\mathbb R}$ as follows: \begin{equation}\label{eq:exp_1d_beta} \ensuremath{{\mathcal F}}_\beta \coloneqq \left\{\left(-\beta y+y,y,e^{1-\beta}y \right)\;\bigg|\;y \in [0,\infty )\right\}. \end{equation} \item a single ``exceptional'' exposed extreme ray denoted by $\ensuremath{{\mathcal F}}_{\infty}$: \begin{equation}\label{eq:exp_1d_exc} \ensuremath{{\mathcal F}}_{\infty}\coloneqq \{(x,0,0)\;|\;x\le0 \}. \end{equation} \item a single non-exposed extreme ray denoted by $\ensuremath{{\mathcal F}} _{ne} $: \begin{equation}\label{eq:exp_1d_ne} \ensuremath{{\mathcal F}} _{ne}\coloneqq \{(0,0,z)\;|\;z\ge0 \}. \end{equation} \item a single 2D exposed face denoted by $\ensuremath{{\mathcal F}}_{-\infty}$: \begin{equation}\label{eq:exp_2d} \ensuremath{{\mathcal F}}_{-\infty}\coloneqq \{(x,y,z)\;|\;x\le0, z\ge0, y=0\}, \end{equation} where we note that $\ensuremath{{\mathcal F}}_{\infty}$ and $\ensuremath{{\mathcal F}} _{ne}$ are the extreme rays of $\ensuremath{{\mathcal F}}_{-\infty}$. \end{enumerate} Notice that except for the case (c), all faces are exposed and thus arise as an intersection $\{\ensuremath{\mathbf z}\}^\perp\cap K_{\exp}$ for some $\ensuremath{\mathbf z} \in K_{\exp}^*$. To establish the above characterization, we start by examining how the components of $\ensuremath{\mathbf z}$ determine the corresponding exposed face. \subsubsection{Exposed faces}\label{sec:exposed} Let $\ensuremath{\mathbf z}\in \ensuremath{K_{\exp}}^*$ be such that $\{\ensuremath{\mathbf z}\}^\perp\cap \ensuremath{K_{\exp}}$ is a nontrivial face of $\ensuremath{K_{\exp}}$. Then $\ensuremath{\mathbf z}\neq 0$ and $\ensuremath{\mathbf z}\in \partial K_{\exp}^*$. We consider the following cases. \noindent \underline{$z_x < 0$}: Since $\ensuremath{\mathbf z}\in \partial K_{\exp}^*$, we must have $z_z e = -z_x e^{\frac{z_y}{z_x}}$ and hence \begin{equation}\label{d:pz} \ensuremath{\mathbf z}=(z_x,z_y,-z_x e^{\frac{z_y}{z_x}-1}). \end{equation} Since $z_x \neq 0$, we see that $\ensuremath{\mathbf q} \in \{\ensuremath{\mathbf z}\}^\perp$ if and only if \begin{equation}\label{qdotp1} q_x+ q_y\left(\frac{z_y}{z_x}\right)-q_z e^{\frac{z_y}{z_x}-1} = 0. \end{equation} Solving \eqref{qdotp1} for $q_z$ and letting $\beta:=\frac{z_y}{z_x}$ to simplify the exposition, we have \begin{equation}\label{qdotp4} q_z = e^{1-\frac{z_y}{z_x}}\left(q_x+q_y\cdot \frac{z_y}{z_x} \right) = e^{1-\beta}\left(q_x+q_y\beta \right) \;\; \text{with}\;\;\beta:=\frac{z_y}{z_x}\in (-\infty,\infty). \end{equation} Thus, we obtain that $\{\ensuremath{\mathbf z}\}^\perp = \left \{ \left(x,y, e^{1-\beta}\left(x+y\beta \right) \right)\;\big|\; x,y \in \ensuremath{\mathbb R} \right \}$. Combining this with the definition of $K_{\exp}$ and the fact that $\{\ensuremath{\mathbf z}\}^\perp$ is a supporting hyperplane (so that $K_{\exp} \cap \{\ensuremath{\mathbf z}\}^\perp = \partial K_{\exp}\cap \{\ensuremath{\mathbf z}\}^\perp$) yields \begin{equation}\label{capperp} \begin{aligned} &K_{\exp} \cap \{\ensuremath{\mathbf z}\}^\perp = \partial K_{\exp}\cap \{\ensuremath{\mathbf z}\}^\perp \\ =& \left \{ \left(x,y, e^{1-\beta}\left(x+y\beta \right) \right) \;\;\big|\;\; e^{1-\beta}\left(x+y\beta \right) = ye^{\frac{x}{y}},y>0 \right \} \\ & \bigcup \left \{ \left(x,y, e^{1-\beta}\left(x+y\beta \right) \right) \;\;\big|\;\; x \leq 0, e^{1-\beta}\left(x+y\beta \right) \geq 0,y=0 \right \} \\ =&\left \{ \left(x,y, e^{1-\beta}\left(x+y\beta \right) \right) \;\;\big|\;\; e^{1-\beta}\left(x+y\beta \right) = ye^{\frac{x}{y}},y>0 \right \} \cup \{0\}. \end{aligned} \end{equation} We now refine the above characterization in the next proposition. \begin{proposition}[Characterization of $\ensuremath{{\mathcal F}}_\beta$, $\beta\in \ensuremath{\mathbb R}$]\label{d:FcapbdKexp} Let $\ensuremath{\mathbf z} \in K_{\exp}^*$ satisfy $\ensuremath{\mathbf z}=(z_x,z_y,z_z)$, where $z_z e = -z_x e^{\frac{z_y}{z_x}}$ and $z_x <0$. Define $\beta=\frac{z_y}{z_x}$ as in \eqref{qdotp4} and let $\ensuremath{{\mathcal F}}_\beta:= K_{\exp}\cap \{\ensuremath{\mathbf z}\}^\perp$. Then \begin{align*} \ensuremath{{\mathcal F}}_\beta = \left\{\left(-\beta y+y,y,e^{1-\beta}y \right)\;\bigg|\;y \in [0,\infty )\right\}. \end{align*} \end{proposition} \begin{proof} Let $\Omega := \left\{\left(-\beta y+y,y,e^{1-\beta}y \right)\;\big|\;y \in [0,\infty )\right\}$. In view of \eqref{capperp}, we can check immediately that $\Omega\subseteq \ensuremath{{\mathcal F}}_\beta$. To prove the converse inclusion, pick any $\ensuremath{\mathbf q}=\left(x,y,e^{1-\beta}(x+y\beta)\right) \in \ensuremath{{\mathcal F}}_\beta$. We need to show that $\ensuremath{\mathbf q}\in \Omega$. To this end, we note from \eqref{capperp} that if $y = 0$, then necessarily $\ensuremath{\mathbf q} = \mathbf{0}$ and consequently $\ensuremath{\mathbf q}\in \Omega$. On the other hand, if $y > 0$, then \eqref{capperp} gives $ye^{x/y}= (x+\beta y)e^{1-\beta}$. Then we have the following chain of equivalences: \begin{equation}\label{xneq0case} \begin{array}{crl} & ye^{x/y}&\!\!\!= (x+\beta y)e^{1-\beta} \\ \iff & -e^{-1} &\!\!\!=-(x/y+\beta)e^{-(x/y+\beta)} \\ \overset{\rm (a)}\iff & -x/y-\beta&\!\!\!=-1 \\ \iff & x&\!\!\!=y-y\beta, \end{array} \end{equation} where (a) follows from the fact that the function $t\mapsto te^t$ is strictly increasing on $[-1,\infty)$. Plugging the last expression back into $\ensuremath{\mathbf q}$, we may compute \begin{equation}\label{xneq0case6} q_z = e^{1-\beta}(x+y\beta) = e^{1-\beta}(y-y\beta+y\beta)=ye^{1-\beta}. \end{equation} Altogether, \eqref{xneq0case}, \eqref{xneq0case6} together with $y>0$ yield \begin{equation*} \ensuremath{\mathbf q}=\left(y-\beta y ,y,ye^{1-\beta} \right)\in \Omega. \end{equation*} This completes the proof \end{proof} Next, we move on to the two remaining cases. \noindent{\underline{$z_x = 0$, $z_z > 0$}}: Notice that $\ensuremath{\mathbf q}\in \ensuremath{K_{\exp}}$ means that $q_y \ge 0$ and $q_z\ge 0$. Since $z_z > 0$ and $z_y \ge 0$, in order to have $\ensuremath{\mathbf q}\in \{\ensuremath{\mathbf z}\}^\perp$, we must have $q_z = 0$. The the definition of $K_{\exp}$ also forces $q_y = 0$ and hence \begin{equation}\label{Finfinity} \{\ensuremath{\mathbf z}\}^\perp\cap K_{\exp}=\{(x,0,0)\;|\; x \leq 0\} =:\ensuremath{{\mathcal F}}_{\infty}. \end{equation} This one-dimensional face is exposed by any hyperplane with normal vectors coming from the set $\{(0,z_y,z_z):\; z_y\ge 0, z_z > 0)\}$. \noindent{\underline{$z_x = 0$, $z_z = 0$}}: In this case, we have $z_y > 0$. In order to have $\ensuremath{\mathbf q}\in \{\ensuremath{\mathbf z}\}^\perp$, we must have $q_y = 0$. Thus \begin{equation}\label{Fneginfinity} \{\ensuremath{\mathbf z}\}^\perp\cap K_{\exp} = \{(x,y,z)\;|\;x\le0, z\ge0, y=0\} =: \ensuremath{{\mathcal F}}_{-\infty}, \end{equation} which is the unique two-dimensional face of $K_{\exp}$. \subsubsection{The single non-exposed face and completeness of the classification} The face $\ensuremath{{\mathcal F}}_{ne}$ is non-exposed because, as shown in Proposition~\ref{d:FcapbdKexp}, \eqref{Finfinity} and \eqref{Fneginfinity}, it never arises as an intersection of the form $\{\ensuremath{\mathbf z}\}^\perp\cap K_{\exp}$, for $\ensuremath{\mathbf z} \in K_{\exp}^*$. We now show that all nontrivial faces of $K_{\exp}$ were accounted for in \eqref{eq:exp_1d_beta}, \eqref{eq:exp_1d_exc}, \eqref{eq:exp_1d_ne}, \eqref{eq:exp_2d}. First of all, by the discussion in Section~\ref{sec:exposed}, all exposed faces must be among the ones in \eqref{eq:exp_1d_beta}, \eqref{eq:exp_1d_exc} and \eqref{eq:exp_2d}. So, let $ \ensuremath{\mathcal{F}}$ be a non-exposed face of $K_{\exp}$. Then, it must be contained in a nontrivial exposed face of $K_{\exp}$.\footnote{This is a general fact. A proper face of a closed convex cone is contained in a proper exposed face, e.g., \cite[Proposition~3.6]{BW81}.} Therefore, $ \ensuremath{\mathcal{F}}$ must be a proper face of the unique 2D face \eqref{eq:exp_2d}. This implies that $ \ensuremath{\mathcal{F}}$ is one of the extreme rays of \eqref{eq:exp_2d}: $\ensuremath{{\mathcal F}}_{\infty}$ or $\ensuremath{{\mathcal F}}_{ne}$. By assumption, $ \ensuremath{\mathcal{F}}$ is non-exposed, so it must be $\ensuremath{{\mathcal F}}_{ne}$. \subsection{Facial residual functions}\label{sec:exp_frf} In this subsection, we will use the machinery developed in Section~\ref{sec:frf} to obtain the facial residual functions for $K_{\exp}$. Let us first discuss how the discoveries were originally made, and how that process motivated the development of the framework we built in Section~\ref{sec:frf}. The FRFs proven here were initially found by using the characterizations of Theorem~\ref{thm:1dfacesmain} and Lemma~\ref{lem:infratio} together with numerical experiments. Specifically, we used \emph{Maple}\footnote{Though one could use any suitable computer algebra package.} to numerically evaluate limits of relevant sequences \eqref{infratiohk}, as well as plotting lower dimensional slices of the function $\ensuremath{\mathbf v} \mapsto \ensuremath{\mathfrak{g}}(\|\ensuremath{\mathbf v}-\ensuremath{\mathbf w}\|)/\|\ensuremath{\mathbf w}-\ensuremath{\mathbf u}\|$, where $\ensuremath{\mathbf w}$ and $\ensuremath{\mathbf u}$ are defined similarly as in \eqref{gammabetaeta}. A natural question is whether it might be simpler to change coordinates and work with the nearly equivalent $\ensuremath{\mathbf w} \mapsto \ensuremath{\mathfrak{g}}(\|\ensuremath{\mathbf v}-\ensuremath{\mathbf w}\|)/\|\ensuremath{\mathbf w}-\ensuremath{\mathbf u}\|$, since $\ensuremath{\mathbf w} \in \{\ensuremath{\mathbf z} \}^\perp$. However, $P_{\{\ensuremath{\mathbf z}\}^\perp}^{-1}\{\ensuremath{\mathbf w}\}\cap \partial \ensuremath{\mathcal{K}}$ may contain multiple points, which creates many challenges. We encountered an example of this when working with the exponential cone, where the change of coordinates from $\ensuremath{\mathbf v}$ to $\ensuremath{\mathbf w}$ necessitates the introduction of the two real branches of the Lambert $\mathcal W$ function (see, for example, \cite{BL2016,bauschke2018proximal,burachik2019generalized} or \cite{Se15} for the closely related Wright Omega function). With terrible effort, one can use such a parametrization to prove the FRFs for $\ensuremath{{\mathcal F}}_{\beta}, \beta \in \left[-\infty,\infty \right]\setminus \{\hat{\beta}:=-\mathcal W_{{\rm principal}}(2e^{-2})/2 \}$. However, the change of branches inhibits proving the result for the exceptional number $\hat{\beta}$. The change of variables to $\ensuremath{\mathbf v}$ cures this problem by obviating the need for a branch function in the analysis; see \cite{WWJD} for additional details. This is why we present Theorem~\ref{thm:1dfacesmain} in terms of $\ensuremath{\mathbf v}$. Computational investigation also pointed to the path of proof, though the proof we present may be understood without the aid of a computer. \subsubsection{$\ensuremath{{\mathcal F}}_{-\infty}$: the unique 2D face} \label{sec:freas_2d} Define the piecewise modified Boltzmann--Shannon entropy $\ensuremath{\mathfrak{g}}_{-\infty}:\ensuremath{\mathbb{R}_+}\to \ensuremath{\mathbb{R}_+}$ as follows: \begin{equation}\label{d:entropy} \ensuremath{\mathfrak{g}}_{-\infty}(t) := \begin{cases} 0 & \text{if}\;\; t=0,\\ -t \ln(t) & \text{if}\;\; t\in \left(0,1/e^2\right],\\ t+ \frac{1}{e^2} & \text{if}\;\; t>1/e^2. \end{cases} \end{equation} For more on its usefulness in optimization, see, for example, \cite{bauschke2018proximal,burachik2019generalized}. We note that $\ensuremath{\mathfrak{g}}_{-\infty}$ is monotone increasing and there exists $L>0$ such that the following inequalities hold for every $t \in \ensuremath{\mathbb{R}_+}$ and $M > 0$: \begin{equation}\label{d:entropy_p} |t| \leq \ensuremath{\mathfrak{g}}_{-\infty}(t), \quad \ensuremath{\mathfrak{g}}_{-\infty}(2t) \leq L\ensuremath{\mathfrak{g}}_{-\infty}(t),\quad \ensuremath{\mathfrak{g}}_{-\infty}(Mt) \leq L^{1+\log_2(M)}\ensuremath{\mathfrak{g}}_{-\infty}(t). \end{equation} With that, we prove in the next theorem that $\gamma_{\ensuremath{\mathbf z},\eta}$ is positive for $\ensuremath{{\mathcal F}}_{-\infty}$, which implies that an \emph{entropic error bound} holds. \begin{theorem}[Entropic error bound concerning $\ensuremath{{\mathcal F}}_{-\infty}$]\label{thm:entropic} Let $\ensuremath{\mathbf z}\in K_{\exp}^*$ with $z_x=z_z=0$ and $z_y > 0$ so that $\{\ensuremath{\mathbf z} \}^\perp \cap K_{\exp}=\ensuremath{{\mathcal F}}_{-\infty}$ is the two-dimensional face of $K_{\exp}$. Let $\eta > 0$ and let $\gamma_{\ensuremath{\mathbf z},\eta}$ be defined as in \eqref{gammabetaeta} with $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_{-\infty}$ in \eqref{d:entropy}. Then $\gamma_{\ensuremath{\mathbf z},\eta} \in (0,\infty]$ and \begin{equation}\label{eq:entropic} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},\ensuremath{{\mathcal F}}_{-\infty})\le \max\{2,2\gamma_{\ensuremath{\mathbf z},\eta}^{-1}\}\cdot\ensuremath{\mathfrak{g}}_{-\infty}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},K_{\exp}))\ \ \ \mbox{whenever $\ensuremath{\mathbf q}\in \{\ensuremath{\mathbf z}\}^\perp\cap B(\eta)$.} \end{equation} \end{theorem} \begin{proof} In view of Lemma~\ref{lem:infratio}, take any $\ensuremath{\bar{\vv}} \in \ensuremath{{\mathcal F}}_{-\infty}$ and a sequence $\{\ensuremath{\mathbf v}^k\}\subset \bd K_{\exp}\cap B(\eta) \backslash \ensuremath{{\mathcal F}}_{-\infty}$ such that \begin{equation}\label{infratiohk_contradiction_neg_infty} \underset{k \rightarrow \infty}{\lim}\ensuremath{\mathbf v}^k = \underset{k \rightarrow \infty}{\lim}\ensuremath{\mathbf w}^k =\ensuremath{\bar{\vv}} \end{equation} where $\ensuremath{\mathbf w}^k = P_{\{\ensuremath{\mathbf z}\}^\perp}\ensuremath{\mathbf v}^k$, $\ensuremath{\mathbf u}^k = P_{\ensuremath{{\mathcal F}}_{-\infty}}\ensuremath{\mathbf w}^k$, and $\ensuremath{\mathbf w}^k\neq \ensuremath{\mathbf u}^k$. We will show that \eqref{infratiohkb} does not hold for $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_{-\infty}$. Since $\ensuremath{\mathbf v}^k \notin \ensuremath{{\mathcal F}}_{-\infty}$, in view of \eqref{d:Kexp} and \eqref{Fneginfinity}, we have $v^k_y>0$ and \begin{equation}\label{formulav_entropic} \ensuremath{\mathbf v}^k = (v^k_x,v^k_y,v^k_ye^{v^k_x/v^k_y}) = (v^k_y\ln(v^k_z/v^k_y),v^k_y,v^k_z), \end{equation} where the second representation is obtained by solving for $v^k_x$ from $v^k_z = v^k_ye^{v^k_x/v^k_y} > 0$. Using the second representation in \eqref{formulav_entropic}, we then have \begin{equation}\label{formulawu_entropic} \ensuremath{\mathbf w}^k = (v^k_y\ln(v^k_z/v^k_y),0,v^k_z)\ \ {\rm and}\ \ \ensuremath{\mathbf u}^k = (0,0,v^k_z); \end{equation} here, we made use of the fact that $\ensuremath{\mathbf w}^k\neq \ensuremath{\mathbf u}^k$, which implies that $v^k_y\ln(v^k_z/v^k_y) > 0$ and thus the resulting formula for $\ensuremath{\mathbf u}^k$. In addition, we also note from $v^k_y\ln(v^k_z/v^k_y) > 0$ (and $v^k_y>0$) that \begin{equation}\label{inequality_entropic} v^k_z > v^k_y > 0. \end{equation} Furthermore, since $\bar \ensuremath{\mathbf v} \in \ensuremath{{\mathcal F}}_{-\infty}$, we see from \eqref{Fneginfinity} and \eqref{infratiohk_contradiction_neg_infty} that \begin{equation}\label{limitvyzero} \lim_{k\to\infty} v_y^k = 0. \end{equation} Now, using \eqref{formulav_entropic}, \eqref{formulawu_entropic}, \eqref{limitvyzero} and the definition of $\ensuremath{\mathfrak{g}}_{-\infty}$, we see that for $k$ sufficiently large, \begin{equation}\label{e:entropic} \frac{\ensuremath{\mathfrak{g}}_{-\infty}(\|\ensuremath{\mathbf v}^k-\ensuremath{\mathbf w}^k\|)}{\|\ensuremath{\mathbf u}^k-\ensuremath{\mathbf w}^k\|} = \frac{-v_y^k \ln(v_y^k)}{v_y^k\ln(v_z^k/v_y^k)} = \frac{-\ln(v_y^k)}{\ln(v_z^k)-\ln(v_y^k)}. \end{equation} We will show that \eqref{infratiohkb} does not hold for $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_{-\infty}$ in each of the following cases. \begin{enumerate}[(I)] \item\label{entropiccase1} $\bar{v}_z >0$. \item\label{entropiccaseinfty} $\bar{v}_z =0$. \end{enumerate} \noindent \ref{entropiccase1}: In this case, we deduce from \eqref{limitvyzero} and \eqref{e:entropic} that \begin{equation*} \lim_{k\rightarrow\infty} \frac{\ensuremath{\mathfrak{g}}_{-\infty}(\|\ensuremath{\mathbf v}^k-\ensuremath{\mathbf w}^k\|)}{\|\ensuremath{\mathbf u}^k-\ensuremath{\mathbf w}^k\|} = 1. \end{equation*} Thus \eqref{infratiohkb} does not hold for $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_{-\infty}$. \noindent \ref{entropiccaseinfty}: By passing to a subsequence if necessary, we may assume that $v^k_z < 1$ for all $k$. This together with \eqref{inequality_entropic} gives $\frac{\ln(v_z^k)}{\ln(v_y^k)} \in (0,1)$ for all $k$. Thus, we conclude from \eqref{e:entropic} that for all $k$, \begin{align*} \frac{\ensuremath{\mathfrak{g}}_{-\infty}(\|\ensuremath{\mathbf v}^k-\ensuremath{\mathbf w}^k\|)}{\|\ensuremath{\mathbf u}^k-\ensuremath{\mathbf w}^k\|} & = \frac{-\ln(v_y^k)}{\ln(v_z^k)-\ln(v_y^k)} = \frac{1}{1 - \frac{\ln(v_z^k)}{\ln(v_y^k)}}> 1 \end{align*} Consequently, \eqref{infratiohkb} also fails for $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_{-\infty}$ in this case. Having shown that \eqref{infratiohkb} does not hold for $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_{-\infty}$ in any case, we conclude by Lemma~\ref{lem:infratio} that $\gamma_{\ensuremath{\mathbf z},\eta} \in \left(0,\infty \right]$. With that, \eqref{eq:entropic} follows from Theorem~\ref{thm:1dfacesmain} and \eqref{d:entropy_p}. \end{proof} Using Theorem~\ref{thm:entropic}, we can also show weaker H\"olderian error bounds. \begin{corollary}\label{col:2d_hold} Let $\ensuremath{\mathbf z}\in K_{\exp}^*$ with $z_x=z_z=0$ and $z_y > 0$ so that $\{\ensuremath{\mathbf z} \}^\perp \cap \ensuremath{K_{\exp}}=\ensuremath{{\mathcal F}}_{-\infty}$ is the two-dimensional face of $\ensuremath{K_{\exp}}$. Let $\eta>0$, $\alpha \in (0,1)$, and $\gamma_{\ensuremath{\mathbf z},\eta}$ be as in \eqref{gammabetaeta} with $\ensuremath{\mathfrak{g}} = |\cdot|^\alpha$. Then $\gamma_{\ensuremath{\mathbf z},\eta} \in (0,\infty]$ and \begin{equation*} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},\ensuremath{{\mathcal F}}_{-\infty})\le \max\{2\eta^{1-\alpha},2\gamma_{\ensuremath{\mathbf z},\eta}^{-1}\}\cdot\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},K_{\exp})^\alpha\ \ \ \mbox{whenever $\ensuremath{\mathbf q}\in \{\ensuremath{\mathbf z}\}^\perp\cap B(\eta)$.} \end{equation*} \end{corollary} \begin{proof} Suppose that $\gamma_{\ensuremath{\mathbf z},\eta} = 0$ and let sequences $\{\ensuremath{\mathbf v}^k\},\{\ensuremath{\mathbf w}^k\},\{\ensuremath{\mathbf u}^k\}$ be as in Lemma~\ref{lem:infratio}. Then $\ensuremath{\mathbf v}^k\neq \ensuremath{\mathbf w}^k$ for all $k$ because $\{\ensuremath{\mathbf v}^k\}\subset \ensuremath{K_{\exp}}\backslash \ensuremath{{\mathcal F}}_{-\infty}$, $\{\ensuremath{\mathbf w}^k\}\subset\{\ensuremath{\mathbf z}\}^\perp$, and $\ensuremath{{\mathcal F}}_{-\infty} = \ensuremath{K_{\exp}}\cap \{\ensuremath{\mathbf z}\}^\perp$. Since $\ensuremath{\mathfrak{g}}_{-\infty}(t)/|t|^{\alpha} \downarrow 0 $ as $t \downarrow 0$ we have \begin{equation*} \liminf_{k\rightarrow\infty} \frac{\ensuremath{\mathfrak{g}}_{-\infty}(\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf v}^k\|)}{\|\ensuremath{\mathbf w}^k- \ensuremath{\mathbf u}^k\|} = \liminf_{k\rightarrow\infty} \frac{\ensuremath{\mathfrak{g}}_{-\infty}(\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf v}^k\|)}{\|\ensuremath{\mathbf w}^k- \ensuremath{\mathbf v}^k\|^\alpha} \frac{\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf v}^k\|^{\alpha}}{\|\ensuremath{\mathbf w}^k- \ensuremath{\mathbf u}^k\|} = 0, \end{equation*} which contradicts Theorem~\ref{thm:entropic} because the quantity in \eqref{gammabetaeta} should be positive for $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}} _{-\infty}$. \end{proof} Recalling \eqref{d:entropy_p}, we obtain facial residual functions using Theorem~\ref{thm:entropic} and Corollary~\ref{col:2d_hold} in combination with Theorem~\ref{thm:1dfacesmain}, Remark~\ref{rem:kappa} and Lemma~\ref{lem:facialresidualsbeta}. \begin{corollary}[Facial residual functions concerning $\ensuremath{{\mathcal F}}_{-\infty}$]\label{col:frf_2dface_entropic} Let $\ensuremath{\mathbf z}\in K_{\exp}^*$ be such that $\{\ensuremath{\mathbf z} \}^\perp \cap K_{\exp}=\ensuremath{{\mathcal F}}_{-\infty}$ is the two-dimensional face of $K_{\exp}$. Let $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_{-\infty}$ in \eqref{d:entropy} or $\ensuremath{\mathfrak{g}} = |\cdot|^\alpha$ for $\alpha\in (0,1)$. Let $\kappa_{\ensuremath{\mathbf z},t}$ be defined as in \eqref{haha2}. Then the function $\psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}:\ensuremath{\mathbb R}_+\times\ensuremath{\mathbb R}_+\to \ensuremath{\mathbb{R}_+}$ given by \begin{equation*} \psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}(\epsilon,t):=\max \left\{\epsilon,\epsilon/\|\ensuremath{\mathbf z}\| \right\} + \kappa_{\ensuremath{\mathbf z},t}\ensuremath{\mathfrak{g}}(\epsilon +\max \left\{\epsilon,\epsilon/\|\ensuremath{\mathbf z}\| \right\} ) \end{equation*} is a facial residual function for $K_{\exp}$ and $\ensuremath{\mathbf z}$ with respect to $K_{\exp}$. In particular, there exist $\kappa > 0$ and a nonnegative mononotone nondecreasing function $\rho:\ensuremath{\mathbb R}_+ \to \ensuremath{\mathbb R}_+$ such that the function $\hat \psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}$ given by $\hat \psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}(\epsilon,t) \coloneqq \kappa \epsilon + \rho(t)\ensuremath{\mathfrak{g}}(\epsilon)$ is a facial residual function for $K_{\exp}$ and $\ensuremath{\mathbf z}$ with respect to $K_{\exp}$. \end{corollary} \subsubsection{$\ensuremath{{\mathcal F}}_\beta$: the family of one dimensional faces $\beta \in \ensuremath{\mathbb R}$ We will now show that for $\ensuremath{{\mathcal F}}_\beta$, $\beta\in \ensuremath{\mathbb R}$, the $\gamma_{\ensuremath{\mathbf z},\eta}$ defined in Theorem~\ref{thm:1dfacesmain} is positive when $\ensuremath{\mathfrak{g}} = |\cdot|^\frac12$. Our discussion will be centered around the following quantities, which were also defined and used in the proof of Theorem~\ref{thm:1dfacesmain}. Specifically, for $\ensuremath{\mathbf z} \in \ensuremath{K_{\exp}}^*$ such that $\ensuremath{{\mathcal F}}_{\beta} = \ensuremath{K_{\exp}}\cap \{\ensuremath{\mathbf z}\}^\perp$, we let $\ensuremath{\mathbf v}\in \bd K_{\exp}\cap B(\eta)\backslash \ensuremath{{\mathcal F}}_\beta$ and define \begin{align} \label{d:w} \ensuremath{\mathbf w}:=P_{\{\ensuremath{\mathbf z} \}^\perp }\ensuremath{\mathbf v}\ \ \ {\rm and}\ \ \ \ensuremath{\mathbf u}:=P_{\ensuremath{{\mathcal F}}_\beta}\ensuremath{\mathbf w} \end{align} We first note the following three important vectors: \begin{equation}\label{veczfp} \ensuremath{\widehat{\mathbf z}} := \begin{bmatrix} 1\\ \beta\\ -e^{\beta-1} \end{bmatrix},\ \ \ \ensuremath{\widehat{\mathbf f}} = \begin{bmatrix} 1-\beta \\ 1\\ e^{1-\beta} \end{bmatrix},\ \ \ \ensuremath{\widehat{\mathbf p}} = \begin{bmatrix} \beta e^{1-\beta} + e^{\beta-1}\\ -e^{1-\beta} - (1-\beta)e^{\beta-1}\\ \beta^2-\beta+1 \end{bmatrix}. \end{equation} Note that $\ensuremath{\widehat{\mathbf z}}$ is parallel to $\ensuremath{\mathbf z}$ in \eqref{d:pz} (recall that $z_x < 0$ for $\ensuremath{{\mathcal F}}_\beta$, where $\beta := \frac{z_y}{z_x}\in \ensuremath{\mathbb R}$), $\ensuremath{{\mathcal F}}_\beta$ is the conic hull of $\{\ensuremath{\widehat{\mathbf f}}\}$ according to Proposition~\ref{d:FcapbdKexp}, $\langle\ensuremath{\widehat{\mathbf z}},\ensuremath{\widehat{\mathbf f}}\rangle=0$ and $\ensuremath{\widehat{\mathbf p}} = \ensuremath{\widehat{\mathbf z}}\times\ensuremath{\widehat{\mathbf f}}\neq {\bf 0}$. These three {\em nonzero} vectors form an orthogonal set. The next lemma represents $\|\ensuremath{\mathbf u} - \ensuremath{\mathbf w}\|$ and $\|\ensuremath{\mathbf w} - \ensuremath{\mathbf v}\|$ in terms of inner products of $\ensuremath{\mathbf v}$ with these vectors, whenever possible. \begin{lemma}\label{lem:distance} Let $\beta\in \ensuremath{\mathbb R}$ and $\ensuremath{\mathbf z}\in K_{\exp}^*$ with $z_x<0$ so that $\ensuremath{{\mathcal F}}_\beta = \{\ensuremath{\mathbf z} \}^\perp \cap K_{\exp}$ is a one-dimensional face of $K_{\exp}$. Let $\eta >0$, $\ensuremath{\mathbf v}\in \bd K_{\exp}\cap B(\eta)\backslash \ensuremath{{\mathcal F}}_\beta$ and define $\ensuremath{\mathbf w}$ and $\ensuremath{\mathbf u}$ as in \eqref{d:w}. Let $\{\ensuremath{\widehat{\mathbf z}},\ensuremath{\widehat{\mathbf f}},\ensuremath{\widehat{\mathbf p}}\}$ be as in \eqref{veczfp}. Then \[ \|\ensuremath{\mathbf w} - \ensuremath{\mathbf v}\| = \frac{|\langle \ensuremath{\widehat{\mathbf z}},\ensuremath{\mathbf v}\rangle|}{\|\ensuremath{\widehat{\mathbf z}}\|}\ \ \ {\rm and}\ \ \ \|\ensuremath{\mathbf w} - \ensuremath{\mathbf u}\| =\begin{cases} \frac{|\langle \ensuremath{\widehat{\mathbf p}},\ensuremath{\mathbf v}\rangle|}{\|\ensuremath{\widehat{\mathbf p}}\|} & {\rm if}\ \langle\ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf v}\rangle\ge 0,\\ \|\ensuremath{\mathbf w}\| & {\rm otherwise}. \end{cases} \] Moreover, when $\langle\ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf v}\rangle\ge 0$, we have $\ensuremath{\mathbf u} = P_{{\rm span}\ensuremath{{\mathcal F}}_\beta}\ensuremath{\mathbf w}$. \end{lemma} \begin{proof} Since $\{\ensuremath{\widehat{\mathbf z}},\ensuremath{\widehat{\mathbf f}},\ensuremath{\widehat{\mathbf p}}\}$ is orthogonal, one can decompose $\ensuremath{\mathbf v}$ as \begin{equation}\label{eq:vdecompose} \ensuremath{\mathbf v} = \lambda_1 \ensuremath{\widehat{\mathbf z}} + \lambda_2 \ensuremath{\widehat{\mathbf f}} + \lambda_3 \ensuremath{\widehat{\mathbf p}}, \end{equation} with \begin{equation}\label{eq:lambda} \lambda_1 = \langle \ensuremath{\widehat{\mathbf z}},\ensuremath{\mathbf v}\rangle/\|\ensuremath{\widehat{\mathbf z}}\|^2, \ \ \lambda_2 = \langle \ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf v}\rangle/\|\ensuremath{\widehat{\mathbf f}}\|^2\ \ {\rm and}\ \ \lambda_3 = \langle \ensuremath{\widehat{\mathbf p}},\ensuremath{\mathbf v}\rangle/\|\ensuremath{\widehat{\mathbf p}}\|^2. \end{equation} Also, since $\ensuremath{\widehat{\mathbf z}}$ is parallel to $\ensuremath{\mathbf z}$, we must have $\ensuremath{\mathbf w} = \lambda_2 \ensuremath{\widehat{\mathbf f}} + \lambda_3 \ensuremath{\widehat{\mathbf p}}$. Thus, it holds that $\|\ensuremath{\mathbf w}-\ensuremath{\mathbf v}\| = |\lambda_1|\|\ensuremath{\widehat{\mathbf z}}\|$ and the first conclusion follows from this and \eqref{eq:lambda}. Next, we have $\ensuremath{\mathbf u} = \hat t \,\ensuremath{\widehat{\mathbf f}}$, where \[ \hat t= \ensuremath{\operatorname*{argmin}}_{t\ge 0}\|\ensuremath{\mathbf w} - t\ensuremath{\widehat{\mathbf f}}\| = \begin{cases} \frac{\langle \ensuremath{\mathbf w},\ensuremath{\widehat{\mathbf f}}\rangle}{\|\ensuremath{\widehat{\mathbf f}}\|^2} & {\rm if}\ \langle \ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf w}\rangle\ge 0,\\ 0 & {\rm otherwise}. \end{cases} \] Moreover, observe from \eqref{eq:vdecompose} that $\langle \ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf w}\rangle = \langle \ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf v} - \lambda_1\ensuremath{\widehat{\mathbf z}}\rangle = \langle \ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf v}\rangle$. These mean that when $\langle \ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf v}\rangle< 0$, we have $\ensuremath{\mathbf u} = 0$, while when $\langle \ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf v}\rangle\ge 0$, we have $\ensuremath{\mathbf u} = \frac{\langle \ensuremath{\mathbf w},\ensuremath{\widehat{\mathbf f}}\rangle}{\|\ensuremath{\widehat{\mathbf f}}\|^2}\ensuremath{\widehat{\mathbf f}}=P_{{\rm span}\ensuremath{{\mathcal F}}_\beta}\ensuremath{\mathbf w}$ and \[ \|\ensuremath{\mathbf w} - \ensuremath{\mathbf u}\| = \left\|\ensuremath{\mathbf w} - \frac{\langle \ensuremath{\mathbf w},\ensuremath{\widehat{\mathbf f}}\rangle}{\|\ensuremath{\widehat{\mathbf f}}\|^2}\ensuremath{\widehat{\mathbf f}}\right\| = |\lambda_3|\|\ensuremath{\widehat{\mathbf p}}\| = |\langle\ensuremath{\widehat{\mathbf p}},\ensuremath{\mathbf v}\rangle|/\|\ensuremath{\widehat{\mathbf p}}\|, \] where the second and the third equalities follow from \eqref{eq:vdecompose}, \eqref{eq:lambda}, and the fact that $\ensuremath{\mathbf w} = \lambda_2 \ensuremath{\widehat{\mathbf f}} + \lambda_3 \ensuremath{\widehat{\mathbf p}}$. This completes the proof. \end{proof} We now prove our main theorem in this section. \begin{theorem}[H\"{o}lderian error bound concerning $\ensuremath{{\mathcal F}}_\beta$, $\beta\in \ensuremath{\mathbb R}$]\label{thm:nonzerogamma} Let $\beta\in \ensuremath{\mathbb R}$ and $\ensuremath{\mathbf z}\in K_{\exp}^*$ with $z_x<0$ so that $\ensuremath{{\mathcal F}}_\beta = \{\ensuremath{\mathbf z} \}^\perp \cap K_{\exp}$ is a one-dimensional face of $K_{\exp}$. Let $\eta > 0$ and let $\gamma_{\ensuremath{\mathbf z},\eta}$ be defined as in \eqref{gammabetaeta} with $\ensuremath{\mathfrak{g}} = |\cdot|^\frac12$. Then $\gamma_{\ensuremath{\mathbf z},\eta} \in (0,\infty]$ and \[ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},\ensuremath{{\mathcal F}}_\beta)\le \max\{2\sqrt{\eta},2\gamma_{\ensuremath{\mathbf z},\eta}^{-1}\}\cdot\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},\ensuremath{K_{\exp}})^\frac12\ \ \ \mbox{whenever $\ensuremath{\mathbf q}\in \{\ensuremath{\mathbf z}\}^\perp\cap B(\eta)$.} \] \end{theorem} \begin{proof} In view of Lemma~\ref{lem:infratio}, take any $\ensuremath{\bar{\vv}} \in \ensuremath{{\mathcal F}}_\beta$ and a sequence $\{\ensuremath{\mathbf v}^k\}\subset \bd K_{\exp}\cap B(\eta) \backslash \ensuremath{{\mathcal F}}_\beta$ such that \begin{equation}\label{infratiohk_contradiction} \underset{k \rightarrow \infty}{\lim}\ensuremath{\mathbf v}^k = \underset{k \rightarrow \infty}{\lim}\ensuremath{\mathbf w}^k =\ensuremath{\bar{\vv}} \end{equation} where $\ensuremath{\mathbf w}^k = P_{\{\ensuremath{\mathbf z}\}^\perp}\ensuremath{\mathbf v}^k$, $\ensuremath{\mathbf u}^k = P_{\ensuremath{{\mathcal F}}_\beta}\ensuremath{\mathbf w}^k$, and $\ensuremath{\mathbf w}^k\neq \ensuremath{\mathbf u}^k$. We will show that \eqref{infratiohkb} does not hold for $\ensuremath{\mathfrak{g}} = |\cdot|^\frac12$. We first suppose that $\ensuremath{\mathbf v}^k\in \ensuremath{{\mathcal F}}_{-\infty}$ infinitely often. By extracting a subsequence if necessary, we may assume that $\ensuremath{\mathbf v}^k\in \ensuremath{{\mathcal F}}_{-\infty}$ for all $k$. From the definition of $\ensuremath{{\mathcal F}}_{-\infty}$ in \eqref{Fneginfinity}, we have $v_x^k\le 0$, $v_y^k=0$ and $v_z^k \ge 0$. Thus, recalling the definition of $\ensuremath{\widehat{\mathbf z}}$ in \eqref{veczfp}, it holds that \begin{equation}\label{easycase1} \begin{aligned} |\langle\ensuremath{\widehat{\mathbf z}},\ensuremath{\mathbf v}^k\rangle| &= |v^k_x - v_z^ke^{\beta-1}| = -v^k_x + v_z^ke^{\beta-1} = |v^k_x| + e^{\beta-1}|v_z^k| \\ &\ge \min\{1,e^{\beta-1}\}\sqrt{|v^k_x|^2 + |v_z^k|^2} = \min\{1,e^{\beta-1}\}\|\ensuremath{\mathbf v}^k\|, \end{aligned} \end{equation} where the last equality holds because $v_y^k=0$. Next, using properties of projections onto subspaces, we have $\|\ensuremath{\mathbf w}^k\| \le \|\ensuremath{\mathbf v}^k\|$. This together with Lemma~\ref{lem:distance} and \eqref{easycase1} shows that \[ \|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf u}^k\| = \ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k,\ensuremath{{\mathcal F}}_\beta)\le \|\ensuremath{\mathbf w}^k\| \le \|\ensuremath{\mathbf v}^k\|\le \frac{1}{\min\{1,e^{\beta-1}\}}|\langle\ensuremath{\widehat{\mathbf z}},\ensuremath{\mathbf v}^k\rangle| = \frac{\|\ensuremath{\widehat{\mathbf z}}\|}{\min\{1,e^{\beta-1}\}}\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf v}^k\|. \] Since $\ensuremath{\mathbf w}^k \to \ensuremath{\bar{\vv}}$ and $\ensuremath{\mathbf v}^k \to \ensuremath{\bar{\vv}}$, the above display shows that \eqref{infratiohkb} does not hold for $\ensuremath{\mathfrak{g}} = |\cdot|^\frac12$ in this case. Next, suppose that $\ensuremath{\mathbf v}^k\notin \ensuremath{{\mathcal F}}_{-\infty}$ for all large $k$ instead. By passing to a subsequence, we assume that $\ensuremath{\mathbf v}^k\notin \ensuremath{{\mathcal F}}_{-\infty}$ for all $k$. In view of \eqref{d:Kexp} and \eqref{Fneginfinity}, this means in particular that \begin{equation}\label{formulavstar} \ensuremath{\mathbf v}^k = (v^k_x,v^k_y,v^k_ye^{v^k_x/v^k_y})\ \ {\rm and}\ \ v^k_y>0\ \mbox{ for all}\ k. \end{equation} We consider two cases and show that \eqref{infratiohkb} does not hold for $\ensuremath{\mathfrak{g}} = |\cdot|^\frac12$ in either of them: \begin{enumerate}[(I)] \item\label{ebsubcasev0} $\langle \ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf v}^k\rangle\ge 0$ infinitely often; \item\label{ebsubcasevneq0} $\langle \ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf v}^k\rangle< 0$ for all large $k$. \end{enumerate} \ref{ebsubcasev0}: Since $\langle \ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf v}^k\rangle\ge 0$ infinitely often, by extracting a subsequence if necessary, we assume that $\langle \ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf v}^k\rangle\ge 0$ for all $k$. Now, consider the following functions: \[ \begin{aligned} h_1(\zeta)&:= \zeta + \beta - e^{\beta+\zeta-1},\\ h_2(\zeta)&:= (\beta e^{1-\beta} + e^{\beta-1})\zeta - e^{1-\beta}-(1-\beta)e^{\beta-1} + (\beta^2-\beta+1)e^\zeta. \end{aligned} \] Using these functions, Lemma~\ref{lem:distance}, \eqref{veczfp} and \eqref{formulavstar}, one can see immediately that \begin{equation}\label{hahahehe1} \|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf v}^k\| = \frac{|\langle\ensuremath{\widehat{\mathbf z}},\ensuremath{\mathbf v}^k\rangle|}{\|\ensuremath{\widehat{\mathbf z}}\|} = \frac{v^k_y |h_1(v^k_x/v^k_y)|}{\|\ensuremath{\widehat{\mathbf z}}\|}\ \ {\rm and}\ \ \|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf u}^k\| = \frac{|\langle\ensuremath{\widehat{\mathbf p}},\ensuremath{\mathbf v}^k\rangle|}{\|\ensuremath{\widehat{\mathbf p}}\|} = \frac{v^k_y |h_2(v^k_x/v^k_y)|}{\|\ensuremath{\widehat{\mathbf p}}\|}. \end{equation} Note that $h_1$ is zero {\bf if and only if} $\zeta = 1 - \beta$. Furthermore, we have $h_1'(1-\beta) = 0$ and $h_1''(1-\beta) = -1$. Then, considering the Taylor expansion of $h_1$ around $1-\beta$ we have \begin{equation*} h_1(\zeta) = 1 + (\zeta + \beta-1) - e^{\beta+\zeta-1} = -\frac{(\zeta+\beta-1)^2}{2} + O(|\zeta+\beta-1|^3)\ \ \ {\rm as}\ \ \zeta\to 1-\beta. \end{equation*} Also, one can check that $h_2(1-\beta)=0$ and that \[ h_2'(1-\beta) = \beta e^{1-\beta} + e^{\beta-1} + (\beta^2-\beta+1)e^{1-\beta} = e^{\beta-1} + (\beta^2+1)e^{1-\beta} > 0. \] Therefore, we have the following Taylor expansion of $h_2$ around $1-\beta$: \begin{equation}\label{newlyadded} h_2(\zeta) = (e^{\beta-1} + (\beta^2+1)e^{1-\beta})(\zeta + \beta-1) + O(|\zeta+\beta-1|^2)\ \ \ {\rm as}\ \ \zeta\to 1-\beta. \end{equation} Thus, using the Taylor expansions of $h_1$ and $h_2$ at $1-\beta$ we have \begin{equation}\label{taylor_limit} \lim\limits_{\zeta\to 1-\beta}\frac{|h_1(\zeta)|^\frac12}{|h_2(\zeta)|} = \frac1{\sqrt{2}(e^{\beta-1} + (\beta^2+1)e^{1-\beta})}> 0. \end{equation} Hence, there exist $C_h > 0$ and $\epsilon>0$ so that \begin{equation}\label{maybeusefulrel} |h_1(\zeta)|^\frac12 \ge C_h|h_2(\zeta)| \ \ {\rm whenever}\ |\zeta - (1-\beta)| \le \epsilon. \end{equation} Next, consider the following function\footnote{Notice that this function is well defined because $h_1$ is zero only at $1-\beta$ and thus we will not end up with $\frac00$.} \[ H(\zeta):=\begin{cases} \frac{|h_1(\zeta)|}{|h_2(\zeta)|} & {\rm if}\ |\zeta -(1-\beta)|\ge \epsilon, h_2(\zeta)\neq 0,\\ \infty & {\rm otherwise}. \end{cases} \] Then it is easy to check that $H$ is proper closed and is never zero. Moreover, by direct computation, we have $\lim\limits_{\zeta\to \infty}H(\zeta) = \frac{e^{\beta-1}}{\beta^2-\beta+1} > 0$ and \[ \lim\limits_{\zeta\to -\infty}H(\zeta) = \begin{cases} |\beta e^{1-\beta} + e^{\beta-1}|^{-1} & {\rm if}\ \beta e^{1-\beta} + e^{\beta-1}\neq 0,\\ \infty & {\rm otherwise}. \end{cases} \] Thus, we deduce that $\inf H > 0$. Now, if it happens that $|v^k_x/v^k_y - (1-\beta)|> \epsilon$ for all large $k$, upon letting $\zeta_k:= v^k_x/v^k_y$, we have from \eqref{hahahehe1} that for all large $k$, \begin{equation}\label{finalcase2} \begin{aligned} \frac{\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\|}{\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf u}^k\|} &=\frac{\|\ensuremath{\widehat{\mathbf p}}\|}{\|\ensuremath{\widehat{\mathbf z}}\|}\frac{|h_1(\zeta_k)|}{|h_2(\zeta_k)|} = \frac{\|\ensuremath{\widehat{\mathbf p}}\|}{\|\ensuremath{\widehat{\mathbf z}}\|}H(\zeta_k)\ge \frac{\|\ensuremath{\widehat{\mathbf p}}\|}{\|\ensuremath{\widehat{\mathbf z}}\|}\inf H >0, \end{aligned} \end{equation} where the second equality holds because of the definition of $H$ and the facts that $\ensuremath{\mathbf w}^k\neq \ensuremath{\mathbf u}^k$ (so that $h_2(\zeta_k)\neq 0$ by \eqref{hahahehe1}) and $|v^k_x/v^k_y - (1-\beta)|> \epsilon$ for all large $k$. On the other hand, if it holds that $|v^k_x/v^k_y - (1-\beta)|\le \epsilon$ infinitely often, then by extracting a further subsequence, we may assume that $|v^k_x/v^k_y - (1-\beta)|\le \epsilon$ for all $k$. Upon letting $\zeta_k:= v^k_x/v^k_y$, we have from \eqref{hahahehe1} that \begin{equation}\label{finalcase1} \begin{aligned} \frac{\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf u}^k\|}{\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\|^\frac12} =\frac{\sqrt{v^k_y\|\ensuremath{\widehat{\mathbf z}}\|}}{\|\ensuremath{\widehat{\mathbf p}}\|}\frac{|h_2(\zeta_k)|}{|h_1(\zeta_k)|^\frac12} \le \frac{\sqrt{v^k_y\|\ensuremath{\widehat{\mathbf z}}\|}}{C_h\|\ensuremath{\widehat{\mathbf p}}\|} \le \frac{\sqrt{\eta\|\ensuremath{\widehat{\mathbf z}}\|}}{C_h\|\ensuremath{\widehat{\mathbf p}}\|}, \end{aligned} \end{equation} where the first inequality holds thanks to $|v^k_x/v^k_y - (1-\beta)|\le \epsilon$ for all $k$, \eqref{maybeusefulrel} and the fact that $\ensuremath{\mathbf w}^k\neq {\ensuremath{\mathbf u}}^k$ (so that $h_2(\zeta_k)\neq 0$ and hence $h_1(\zeta_k)\neq 0$), and the second inequality holds because $\ensuremath{\mathbf v}^k\in B(\eta)$. Using \eqref{finalcase2} and \eqref{finalcase1} together with \eqref{infratiohk_contradiction}, we see that \eqref{infratiohkb} does not hold for $\ensuremath{\mathfrak{g}} = |\cdot|^\frac12$. This concludes case~\ref{ebsubcasev0}. \ref{ebsubcasevneq0}: By passing to a subsequence, we may assume that $\langle \ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf v}^k\rangle< 0$ for all $k$. Then we see from \eqref{veczfp} and \eqref{formulavstar} that \[ (1-\beta)\frac{v_x^k}{v_y^k} + 1 + e^{1-\beta}e^{v_x^k/v_y^k} = \frac1{v_y^k} \langle \ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf v}^k\rangle< 0. \] Using this together with the fact that $(1-\beta)^2+1+e^{2(1-\beta)} > 0$, we deduce that there exists $\epsilon> 0$ so that \begin{equation}\label{eq:keyrelation} \left|\frac{v_x^k}{v_y^k}-(1-\beta)\right|\ge \epsilon\ \ \mbox{for all}\ k. \end{equation} Now, consider the following function \[ G(\zeta):= \frac{|\zeta+\beta-e^{\beta+\zeta-1}|}{\sqrt{\zeta^2+1+e^{2\zeta}}}. \] Then $G$ is continuous and is zero {\bf if and only if} $\zeta = 1-\beta$. Moreover, by direct computation, we have $\lim\limits_{\zeta\to\infty} G(\zeta) = e^{\beta-1} > 0$ and $\lim\limits_{\zeta\to-\infty} G(\zeta) = 1 > 0$. Thus, it follows that \begin{equation}\label{infG} \underline{G}:= \inf_{|\zeta +\beta-1|\ge \epsilon}G(\zeta) > 0. \end{equation} Finally, since $\langle \ensuremath{\widehat{\mathbf f}},\ensuremath{\mathbf v}^k\rangle< 0$ for all $k$, we see that \[ \begin{aligned} \frac{\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\|}{\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf u}^k\|}&\overset{\rm (a)}\ge \frac{\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\|}{\|\ensuremath{\mathbf v}^k\|} \overset{\rm (b)}= \frac{|{\widehat{z}}_xv^k_x + {\widehat{z}}_yv^k_y + {\widehat{z}}_zv^k_ye^{v^k_x/v^k_y}|}{\|\ensuremath{\widehat{\mathbf z}}\|\sqrt{(v^k_x)^2+(v^k_y)^2+(v^k_y)^2e^{2v^k_x/v^k_y}}}\\ &\overset{\rm (c)}=\frac{|{\widehat{z}}_x\zeta_k + {\widehat{z}}_y + {\widehat{z}}_ze^{\zeta_k}|}{\|\ensuremath{\widehat{\mathbf z}}\|\sqrt{(\zeta_k)^2 + 1 + e^{2\zeta_k}}} \overset{\rm (d)}= \frac{1}{\|\ensuremath{\widehat{\mathbf z}}\|}G(\zeta_k)\overset{\rm (e)}\ge \frac{1}{\|\ensuremath{\widehat{\mathbf z}}\|}\underline{G} > 0, \end{aligned} \] where (a) follows from $\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf u}^k\| = \|\ensuremath{\mathbf w}^k\|$ (see Lemma~\ref{lem:distance}) and $\|\ensuremath{\mathbf w}^k\|\leq \norm{\ensuremath{\mathbf v}^k}$ (because $\ensuremath{\mathbf w}^k$ is the projection of ${\ensuremath{\mathbf v}}^k$ onto a subspace), (b) follows from Lemma~\ref{lem:distance} and \eqref{formulavstar}, (c) holds because $v^k_y > 0$ (see \eqref{formulavstar}) and we defined $\zeta_k:= v_x^k/v_y^k$, (d) follows from \eqref{veczfp} and the definition of $G$, and (e) follows from \eqref{eq:keyrelation} and \eqref{infG}. The above together with \eqref{infratiohk_contradiction} shows that \eqref{infratiohkb} does not hold for $\ensuremath{\mathfrak{g}} = |\cdot|^\frac12$, which is what we wanted to show in case~\ref{ebsubcasevneq0}. Summarizing the above cases, we conclude that there does not exist the sequence $\{\ensuremath{\mathbf v}^k\}$ and its associates so that \eqref{infratiohkb} holds for $\ensuremath{\mathfrak{g}} = |\cdot|^\frac12$. By Lemma~\ref{lem:infratio}, it must then hold that $\gamma_{\ensuremath{\mathbf z},\eta}\in (0,\infty]$ and we have the desired error bound in view of Theorem~\ref{thm:1dfacesmain}. This completes the proof. \end{proof} Combining Theorem~\ref{thm:nonzerogamma}, Theorem~\ref{thm:1dfacesmain} and Lemma~\ref{lem:facialresidualsbeta}, and using the observation that $\gamma_{\ensuremath{\mathbf z},0}=\infty$ (see \eqref{gammabetaeta}), we obtain a facial residual function in the following corollary. \begin{corollary}[Facial residual function concerning $\ensuremath{{\mathcal F}}_\beta$, $\beta\in \ensuremath{\mathbb R}$]\label{col:frf_fb} Let $\beta\in \ensuremath{\mathbb R}$ and $\ensuremath{\mathbf z}\in K_{\exp}^*$ with $z_x<0$ so that $\ensuremath{{\mathcal F}}_\beta = \{\ensuremath{\mathbf z} \}^\perp \cap K_{\exp}$ is a one-dimensional face of $K_{\exp}$. Let $\kappa_{\ensuremath{\mathbf z},t}$ be defined as in \eqref{haha2} with $\ensuremath{\mathfrak{g}} = |\cdot|^\frac12$. Then the function $\psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}:\ensuremath{\mathbb R}_+\times \ensuremath{\mathbb R}_+\to \ensuremath{\mathbb R}_+$ given by \begin{equation*} \psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}(\epsilon,t) := \max \left\{\epsilon,\epsilon/\|\ensuremath{\mathbf z}\| \right\} + \kappa_{\ensuremath{\mathbf z},t}(\epsilon +\max \left\{\epsilon,\epsilon/\|\ensuremath{\mathbf z}\| \right\} )^\frac12 \end{equation*} is a facial residual function for $K_{\exp}$ and $\ensuremath{\mathbf z}$ with respect to $K_{\exp}$. In particular, there exist $\kappa > 0$ and a nonnegative mononotone nondecreasing function $\rho:\ensuremath{\mathbb R}_+ \to \ensuremath{\mathbb R}_+$ such that the function $\hat \psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}$ given by $\hat \psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}(\epsilon,t) \coloneqq \kappa \epsilon + \rho(t)\sqrt{\epsilon}$ is a facial residual function for $K_{\exp}$ and $\ensuremath{\mathbf z}$ with respect to $K_{\exp}$. \end{corollary} \subsubsection{$\ensuremath{{\mathcal F}}_{\infty}$: the exceptional one dimensional face We first show that we have a Lipschitz error bound for any exposing normal vectors $\ensuremath{\mathbf z} = (0, z_y,z_z)$ with $z_y > 0$ and $z_z > 0$. \begin{theorem}[Lipschitz error bound concerning $\ensuremath{{\mathcal F}}_{\infty}$]\label{thm:nonzerogammasec71} Let $\ensuremath{\mathbf z}\in K_{\exp}^*$ with $z_x=0$, $z_y > 0$ and $z_z>0$ so that $\{\ensuremath{\mathbf z} \}^\perp \cap K_{\exp}=\ensuremath{{\mathcal F}}_\infty$. Let $\eta > 0$ and let $\gamma_{\ensuremath{\mathbf z},\eta}$ be defined as in \eqref{gammabetaeta} with $\ensuremath{\mathfrak{g}} = |\cdot|$. Then $\gamma_{\ensuremath{\mathbf z},\eta} \in (0,\infty]$ and \[ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},\ensuremath{{\mathcal F}}_{\infty})\le \max\{2,2\gamma_{\ensuremath{\mathbf z},\eta}^{-1}\}\cdot\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},K_{\exp})\ \ \ \mbox{whenever $\ensuremath{\mathbf q}\in \{\ensuremath{\mathbf z}\}^\perp\cap B(\eta)$.} \] \end{theorem} \begin{proof} Without loss of generality, upon scaling, we may assume that $\ensuremath{\mathbf z} = (0,a,1)$ for some $a > 0$. Similarly as in the proof of Theorem~\ref{thm:nonzerogamma}, we will consider the following vectors: \begin{equation*} \ensuremath{\widetilde{\mathbf z}} := \begin{bmatrix} 0\\a\\1 \end{bmatrix}, \ \ \ensuremath{\widetilde{\mathbf f}} := \begin{bmatrix} -1\\0\\0 \end{bmatrix},\ \ \ensuremath{\widetilde{\mathbf p}} := \begin{bmatrix} 0\\1\\-a \end{bmatrix}. \end{equation*} Here, $\ensuremath{{\mathcal F}}_\infty$ is the conical hull of $\ensuremath{\widetilde{\mathbf f}}$ (see \eqref{Finfinity}), and $\ensuremath{\widetilde{\mathbf p}}$ is constructed so that $\{\ensuremath{\widetilde{\mathbf z}},\ensuremath{\widetilde{\mathbf f}},\ensuremath{\widetilde{\mathbf p}}\}$ is orthogonal. Now, let $\ensuremath{\mathbf v}\in \bd K_{\exp}\cap B(\eta)\backslash \ensuremath{{\mathcal F}}_\infty$, $\ensuremath{\mathbf w} = P_{\{\ensuremath{\mathbf z}\}^\perp }\ensuremath{\mathbf v}$ and $\ensuremath{\mathbf u}=P_{\ensuremath{{\mathcal F}}_\infty}\ensuremath{\mathbf w}$ with $\ensuremath{\mathbf u}\neq \ensuremath{\mathbf w}$. Then, as in Lemma~\ref{lem:distance}, by decomposing $\ensuremath{\mathbf v}$ as a linear combination of $\{\ensuremath{\widetilde{\mathbf z}},\ensuremath{\widetilde{\mathbf f}},\ensuremath{\widetilde{\mathbf p}}\}$, we have \begin{equation}\label{haha5} \|\ensuremath{\mathbf w} - \ensuremath{\mathbf v}\| = \frac{|\langle\ensuremath{\widetilde{\mathbf z}},\ensuremath{\mathbf v}\rangle|}{\|\ensuremath{\widetilde{\mathbf z}}\|}\ \ {\rm and}\ \ \|\ensuremath{\mathbf w}-\ensuremath{\mathbf u}\| = \begin{cases} \frac{|\langle\ensuremath{\widetilde{\mathbf p}},\ensuremath{\mathbf v}\rangle|}{\|\ensuremath{\widetilde{\mathbf p}}\|} & {\rm if}\ \langle\ensuremath{\widetilde{\mathbf f}},\ensuremath{\mathbf v}\rangle\ge 0,\\ \|\ensuremath{\mathbf w}\| & {\rm otherwise}. \end{cases} \end{equation} We consider the following cases for estimating $\gamma_{\ensuremath{\mathbf z},\eta}$. \begin{enumerate}[(I)] \item\label{inftyebcase1} $\ensuremath{\mathbf v}\in \ensuremath{{\mathcal F}}_{-\infty}\backslash \ensuremath{{\mathcal F}}_\infty$; \item\label{inftyebcase2} $\ensuremath{\mathbf v}\notin \ensuremath{{\mathcal F}}_{-\infty}$ with $v_x \le 0$; \item\label{inftyebcase3} $\ensuremath{\mathbf v}\notin \ensuremath{{\mathcal F}}_{-\infty}$ with $v_x > 0$. \end{enumerate} \ref{inftyebcase1}: In this case, $\ensuremath{\mathbf v} = (v_x,0,v_z)$ with $v_x\le 0\le v_z$; see \eqref{Fneginfinity}. Then $\langle\ensuremath{\widetilde{\mathbf f}},\ensuremath{\mathbf v}\rangle = -v_x\ge 0$ and $|\langle\ensuremath{\widetilde{\mathbf z}},\ensuremath{\mathbf v}\rangle| = |v_z| = \frac1a|\langle\ensuremath{\widetilde{\mathbf p}},\ensuremath{\mathbf v}\rangle|$. Consequently, we have from \eqref{haha5} that \[ \|\ensuremath{\mathbf w} - \ensuremath{\mathbf v}\| = \frac{|\langle\ensuremath{\widetilde{\mathbf z}},\ensuremath{\mathbf v}\rangle|}{\|\ensuremath{\widetilde{\mathbf z}}\|} = \frac{|\langle\ensuremath{\widetilde{\mathbf p}},\ensuremath{\mathbf v}\rangle|}{a\|\ensuremath{\widetilde{\mathbf z}}\|} = \frac{\|\ensuremath{\widetilde{\mathbf p}}\|}{a\|\ensuremath{\widetilde{\mathbf z}}\|}\|\ensuremath{\mathbf w} - \ensuremath{\mathbf u}\|. \] \ref{inftyebcase2}: In this case, in view of \eqref{d:Kexp} and \eqref{Fneginfinity}, we have $\ensuremath{\mathbf v} = (v_x,v_y,v_ye^{v_x/v_y})$ with $v_x\le 0$ and $v_y > 0$. Then $\langle\ensuremath{\widetilde{\mathbf f}},\ensuremath{\mathbf v}\rangle = -v_x\ge 0$. Moreover, since $v_y>0$, we have \[ \begin{aligned} |\langle\ensuremath{\widetilde{\mathbf z}},\ensuremath{\mathbf v}\rangle| &= |av_y+v_ye^{v_x/v_y}| = av_y+v_ye^{v_x/v_y} \ge \min\{1,a\}(v_y+v_ye^{v_x/v_y})\\ & \ge \frac{\min\{1,a\}}{\max\{1,a\}}(v_y+av_ye^{v_x/v_y})\ge \frac{\min\{1,a\}}{\max\{1,a\}}|v_y-av_ye^{v_x/v_y}| = \frac{\min\{1,a\}}{\max\{1,a\}}|\langle\ensuremath{\widetilde{\mathbf p}},\ensuremath{\mathbf v}\rangle|. \end{aligned} \] Using \eqref{haha5}, we then obtain that $\|\ensuremath{\mathbf w} - \ensuremath{\mathbf v}\|\ge \frac{\min\{1,a\}\|\ensuremath{\widetilde{\mathbf p}}\|}{\max\{1,a\}\|\ensuremath{\widetilde{\mathbf z}}\|}\|\ensuremath{\mathbf w} - \ensuremath{\mathbf u}\|$. \ref{inftyebcase3}: In this case, in view of \eqref{d:Kexp} and \eqref{Fneginfinity}, $\ensuremath{\mathbf v} = (v_x,v_y,v_ye^{v_x/v_y})$ with $v_x> 0$ and $v_y > 0$. Then $\langle\ensuremath{\widetilde{\mathbf f}},\ensuremath{\mathbf v}\rangle = -v_x< 0$ and hence $\|\ensuremath{\mathbf w} - \ensuremath{\mathbf u}\|= \|\ensuremath{\mathbf w}\|\le \|\ensuremath{\mathbf v}\|$, where the equality follows from \eqref{haha5} and the inequality holds because $\ensuremath{\mathbf w}$ is the projection of $\ensuremath{\mathbf v}$ onto a subspace. Since $v_y > 0$, we have \[ \begin{aligned} |\langle\ensuremath{\widetilde{\mathbf z}},\ensuremath{\mathbf v}\rangle| &= |av_y+v_ye^{v_x/v_y}| = av_y+0.5v_ye^{v_x/v_y} + 0.5v_ye^{v_x/v_y}\\ & \overset{\rm (a)}\ge av_y+0.5v_y(1 + v_x/v_y) + 0.5v_ye^{v_x/v_y} \ge 0.5v_y + 0.5v_x + 0.5v_ye^{v_x/v_y}= 0.5\|\ensuremath{\mathbf v}\|_1, \end{aligned} \] where we used $v_y > 0$ and $e^t\ge 1+t$ for all $t$ in (a) and $\|\ensuremath{\mathbf v}\|_1$ denotes the $1$-norm of $\ensuremath{\mathbf v}$. Combining this with \eqref{haha5} and the fact that $\|\ensuremath{\mathbf w}\|\le \|\ensuremath{\mathbf v}\|$, we see that \[ \|\ensuremath{\mathbf w} - \ensuremath{\mathbf v}\| = \frac{|\langle\ensuremath{\widetilde{\mathbf z}},\ensuremath{\mathbf v}\rangle|}{\|\ensuremath{\widetilde{\mathbf z}}\|}\ge\frac{\|\ensuremath{\mathbf v}\|_1}{2\|\ensuremath{\widetilde{\mathbf z}}\|}\ge \frac{\|\ensuremath{\mathbf v}\|}{2\|\ensuremath{\widetilde{\mathbf z}}\|} \ge \frac{\|\ensuremath{\mathbf w}\|}{2\|\ensuremath{\widetilde{\mathbf z}}\|} = \frac{\|\ensuremath{\mathbf w}-\ensuremath{\mathbf u}\|}{2\|\ensuremath{\widetilde{\mathbf z}}\|}. \] Summarizing the three cases, we conclude that $\gamma_{\ensuremath{\mathbf z},\eta} \ge \min\left\{\frac{\|\ensuremath{\widetilde{\mathbf p}}\|}{a\|\ensuremath{\widetilde{\mathbf z}}\|},\frac{\min\{1,a\}\|\ensuremath{\widetilde{\mathbf p}}\|}{\max\{1,a\}\|\ensuremath{\widetilde{\mathbf z}}\|},\frac{1}{2\|\ensuremath{\widetilde{\mathbf z}}\|}\right\} > 0$. In view of Theorem~\ref{thm:1dfacesmain}, we have the desired error bound. This completes the proof. \end{proof} We next turn to the supporting hyperplane defined by $\ensuremath{\mathbf z} = (0,0,z_z)$ for some $z_z>0$ and so $\{\ensuremath{\mathbf z} \}^\perp$ is the $xy$-plane. The following lemma demonstrates that the H\"{o}lderian-type error bound in the form of \eqref{haha2} with $\ensuremath{\mathfrak{g}} = |\cdot|^\alpha$ for some $\alpha\in (0,1]$ no longer works in this case. \begin{lemma}[Nonexistence of H\"{o}lderian error bounds]\label{lem:non_hold} Let $\ensuremath{\mathbf z}\in K_{\exp}^*$ with $z_x= z_y = 0$ and $z_z>0$ so that $\{\ensuremath{\mathbf z} \}^\perp \cap K_{\exp}=\ensuremath{{\mathcal F}}_\infty$. Let $\alpha\in (0,1]$ and $\eta > 0$. Then \[ \inf_\ensuremath{\mathbf q}\left\{\frac{\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},K_{\exp})^\alpha}{\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},\ensuremath{{\mathcal F}}_\infty)}\;\bigg|\;\ensuremath{\mathbf q}\in \{\ensuremath{\mathbf z}\}^\perp\cap B(\eta)\backslash\ensuremath{{\mathcal F}}_\infty\right\} = 0. \] \end{lemma} \begin{proof} For each $k\in \mathbb{N}$, let $\ensuremath{\mathbf q}^k := (-\frac\eta2,\frac\eta{2k},0)$. Then $\ensuremath{\mathbf q}^k\in \{\ensuremath{\mathbf z}\}^\perp\cap B(\eta)\backslash\ensuremath{{\mathcal F}}_\infty$ and we have $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}^k,\ensuremath{{\mathcal F}}_\infty) = \frac\eta{2k}$. Moreover, since $(q^k_x,q^k_y,q^k_ye^{q^k_x/q^k_y})\in K_{\exp}$, we have $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}^k,K_{\exp})\le q^k_ye^{q^k_x/q^k_y} = \frac\eta{2k}e^{-k}$. Then it holds that \[ \frac{\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}^k,K_{\exp})^\alpha}{\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}^k,\ensuremath{{\mathcal F}}_\infty)} \le \frac{\eta^{\alpha-1}}{2^{\alpha-1}}k^{1-\alpha}e^{-\alpha k} \to 0 \ \ \ {\rm as}\ \ k\to \infty \] since $\alpha\in (0,1]$. This completes the proof. \end{proof} Since a zero-at-zero monotone nondecreasing function of the form $(\cdot)^\alpha$ no longer works, we opt for the following function $\ensuremath{\mathfrak{g}}_\infty:\ensuremath{\mathbb{R}_+}\to \ensuremath{\mathbb{R}_+}$ that grows faster around $t=0$: \begin{equation}\label{def:frakg} \ensuremath{\mathfrak{g}}_\infty(t) := \begin{cases} 0 &\text{if}\;t=0,\\ -\frac{1}{\ln(t)} & \text{if}\;0 < t\leq \frac{1}{e^2},\\ \frac{1}{4}+\frac{1}{4}e^2t & \text{if}\;t>\frac{1}{e^2}. \end{cases} \end{equation} Similar to $\ensuremath{\mathfrak{g}}_{-\infty}$ in \eqref{d:entropy}, $\ensuremath{\mathfrak{g}}_{\infty}$ is monotone increasing and there exists a constant $\widehat L > 0$ such that the following inequalities hold for every $t \in \ensuremath{\mathbb{R}_+}$ and $M > 0$: \begin{equation}\label{def:frakg_p} |t| \leq \ensuremath{\mathfrak{g}}_{\infty}(t), \quad \ensuremath{\mathfrak{g}}_{\infty}(2t) \leq \widehat L\ensuremath{\mathfrak{g}}_{\infty}(t),\quad \ensuremath{\mathfrak{g}}_{\infty}(Mt) \leq {\widehat L}^{1+\log_2(M)}\ensuremath{\mathfrak{g}}_{\infty}(t). \end{equation} We next show that error bounds in the form of \eqref{haha2} holds for $\ensuremath{\mathbf z} = (0,0,z_z)$, $z_z>0$, if we use $\ensuremath{\mathfrak{g}}_\infty$. \begin{theorem}[Log-type error bound concerning $\ensuremath{{\mathcal F}}_{\infty}$]\label{thm:nonzerogammasec72} Let $\ensuremath{\mathbf z}\in \ensuremath{K_{\exp}}^*$ with $z_x=z_y = 0$ and $z_z>0$ so that $\{\ensuremath{\mathbf z} \}^\perp \cap \ensuremath{K_{\exp}}=\ensuremath{{\mathcal F}}_\infty$. Let $\eta > 0$ and let $\gamma_{\ensuremath{\mathbf z},\eta}$ be defined as in \eqref{gammabetaeta} with $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_\infty$ in \eqref{def:frakg}. Then $\gamma_{\ensuremath{\mathbf z},\eta} \in (0,\infty]$ and \[ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},\ensuremath{{\mathcal F}}_{\infty})\le \max\{2,2\gamma_{\ensuremath{\mathbf z},\eta}^{-1}\}\cdot\ensuremath{\mathfrak{g}}_\infty(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},K_{\exp}))\ \ \ \mbox{whenever $\ensuremath{\mathbf q}\in \{\ensuremath{\mathbf z}\}^\perp\cap B(\eta)$.} \] \end{theorem} \begin{proof} Take $\ensuremath{\bar{\vv}} \in \ensuremath{{\mathcal F}}_\infty$ and a sequence $\{\ensuremath{\mathbf v}^k\}\subset \bd K_{\exp}\cap B(\eta) \backslash \ensuremath{{\mathcal F}}_\infty$ such that \begin{equation*} \underset{k \rightarrow \infty}{\lim}\ensuremath{\mathbf v}^k = \underset{k \rightarrow \infty}{\lim}\ensuremath{\mathbf w}^k =\ensuremath{\bar{\vv}}, \end{equation*} where $\ensuremath{\mathbf w}^k = P_{\{\ensuremath{\mathbf z}\}^\perp}\ensuremath{\mathbf v}^k$, $\ensuremath{\mathbf u}^k = P_{\ensuremath{{\mathcal F}}_\infty}\ensuremath{\mathbf w}^k$, and $\ensuremath{\mathbf w}^k\neq \ensuremath{\mathbf u}^k$. Since $\ensuremath{\mathbf w}^k\neq \ensuremath{\mathbf u}^k$, in view of \eqref{Finfinity} and \eqref{Fneginfinity}, we must have $\ensuremath{\mathbf v}^k\notin \ensuremath{{\mathcal F}}_{-\infty}$. Then, from \eqref{d:Kexp} and \eqref{Finfinity}, we have \begin{equation}\label{vwurepre} \ensuremath{\mathbf v}^k = (v^k_x,v^k_y,v^k_ye^{v^k_x/v^k_y}) \mbox{ with } v^k_y>0,\ \ \ensuremath{\mathbf w}^k = (v^k_x,v^k_y,0)\ \ {\rm and}\ \ \ensuremath{\mathbf u}^k = (\min\{v^k_x,0\},0,0). \end{equation} Since $\ensuremath{\mathbf w}^k\to \ensuremath{\bar{\vv}}$ and $\ensuremath{\mathbf v}^k\to \ensuremath{\bar{\vv}}$, without loss of generality, by passing to a subsequence if necessary, we assume in addition that $\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf v}^k\|\le e^{-2}$ for all $k$. From \eqref{vwurepre} we conclude that $\ensuremath{\mathbf v}^k \neq \ensuremath{\mathbf w}^k$, hence $\ensuremath{\mathfrak{g}}_\infty(\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf v}^k\|) = -(\ln\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf v}^k\|)^{-1}$. We consider the following two cases in order to show that \eqref{infratiohkb} does not hold for $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_\infty$: \begin{enumerate}[(I)] \item\label{finalebcasevneq0} $\ensuremath{\bar{\vv}}\neq \mathbf{0}$; \item\label{finalebcasev0} $\ensuremath{\bar{\vv}} = \mathbf{0}$. \end{enumerate} \ref{finalebcasevneq0}: In this case, we have $\ensuremath{\bar{\vv}} = ({\bar{v}}_x,0,0)$ for some ${\bar{v}}_x < 0$. This implies that $v^k_x<0$ for all large $k$. Hence, we have from \eqref{vwurepre} that for all large $k$, \begin{equation*} \frac{\ensuremath{\mathfrak{g}}_\infty(\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf v}^k\|) }{\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf u}^k \|} = \frac{-(\ln\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\|)^{-1}}{\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf u}^k\|} = -\frac{1}{v^k_y\left(v^k_x/v^k_y + \ln v^k_y \right)} = -\frac{1}{v^k_y\ln v^k_y + v^k_x} \to -\frac1{{\bar{v}}_x} > 0 \end{equation*} since $v^k_y\to 0$ and $v^k_x\to {\bar{v}}_x < 0$. This shows that \eqref{infratiohkb} does not hold for $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_\infty$. \ref{finalebcasev0}: If $v_x^k\le 0$ infinitely often, by extracting a subsequence, we assume that $v_x^k\le 0$ for all $k$. Since $\ensuremath{\mathbf w}^k\neq \ensuremath{\mathbf u}^k$ (and $\ensuremath{\mathbf w}^k\neq \ensuremath{\mathbf v}^k$), we note from \eqref{vwurepre} that \[ -\frac{1}{v^k_y\ln v^k_y + v^k_x} = \frac{-(\ln\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\|)^{-1}}{\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf u}^k\|} \in (0,\infty)\ \ \mbox{for all}\ k. \] Since $\{-(v^k_y\ln v^k_y + v^k_x)\}$ is a positive sequence and it converges to zero as $(v^k_x,v^k_y)\to 0$, it follows that $\lim\limits_{k\to\infty}\frac{-(\ln\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\|)^{-1}}{\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf u}^k\|}=\infty$. This shows that \eqref{infratiohkb} does not hold for $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_\infty$. Now, it remains to consider the case that $v_x^k>0$ for all large $k$. By passing to a subsequence if necessary, we assume that $v_x^k>0$ for all $k$. By solving for $v_x^k$ from $v^k_z=v^k_y e^{v^k_x/v^k_y} > 0$ and noting \eqref{vwurepre}, we obtain that \begin{equation}\label{vwurepre2} \ensuremath{\mathbf v}^k = (v^k_y\ln(v_z^k/v_y^k),v^k_y,v^k_z) \mbox{ with } v^k_y>0,\ \ \ensuremath{\mathbf w}^k = (v^k_y\ln(v_z^k/v_y^k),v^k_y,0)\ \ {\rm and}\ \ \ensuremath{\mathbf u}^k = (0,0,0). \end{equation} Also, we note from $v_x^k=v^k_y\ln(v_z^k/v_y^k)>0$, $v_y^k>0$ and the monotonicity of $\ln(\cdot)$ that for all $k$, \begin{equation}\label{bugfix} v^k_z>v^k_y>0. \end{equation} Next consider the function $h(t) := \frac{1}t\sqrt{1+(\ln t)^2}$ on $[1,\infty)$. Then $h$ is continuous and positive. Since $h(1)=1$ and $\lim_{t\to \infty}h(t) = 0$, there exists $M_h$ such that $h(t)\le M_h$ for all $t\ge 1$. Now, using \eqref{vwurepre2}, we have, upon defining $t_k:= v_z^k/v_y^k$ that \[ \begin{aligned} \frac{\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf u}^k\|}{-(\ln\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\|)^{-1}} &= \frac{v_y^k\sqrt{1+[\ln(v_z^k/v_y^k)]^2}}{-(\ln v_z^k)^{-1}} = -v_y^k\sqrt{1+[\ln(v_z^k/v_y^k)]^2}\ln v_z^k\\ & \overset{\rm (a)}= -\frac{v_y^k}{v_z^k}\sqrt{1+\left[\ln\left(\frac{v_z^k}{v_y^k}\right)\right]^2}v_z^k\ln v_z^k \overset{\rm (b)}= - h(t_k)v_z^k\ln v_z^k \overset{\rm (c)}\le -M_hv_z^k\ln v_z^k, \end{aligned} \] where the division by $v_z^k$ in (a) is legitimate because $v_z^k>0$, (b) follows from the definition of $h$ and the fact that $t_k > 1$ (see \eqref{bugfix}), and (c) holds because of the definition of $M_h$ and the fact that $-\ln v_z^k > 0$ (thanks to $v_z^k = \|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf v}^k\|\le e^{-2}$). Since $v^k_z\to 0$, it then follows that $\left\{\frac{\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf u}^k\|}{-(\ln\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\|)^{-1}}\right\}$ is a positive sequence that converges to zero. Thus, $\lim\limits_{k\to\infty}\frac{-(\ln\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\|)^{-1}}{\|\ensuremath{\mathbf w}^k - \ensuremath{\mathbf u}^k\|}=\infty$, which again shows that \eqref{infratiohkb} does not hold for $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_\infty$. Having shown that \eqref{infratiohkb} does not hold for $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_\infty$, in view of Lemma~\ref{lem:infratio}, we must have $\gamma_{\ensuremath{\mathbf z},\eta} \in (0,\infty]$. Then the desired error bound follows from Theorem~\ref{thm:1dfacesmain} and \eqref{def:frakg_p}. \end{proof} Combining Theorem~\ref{thm:nonzerogammasec71}, Theorem~\ref{thm:nonzerogammasec72} and Lemma~\ref{lem:facialresidualsbeta}, and noting that $\gamma_{\ensuremath{\mathbf z},0}=\infty$ (see \eqref{gammabetaeta}), we can now summarize the facial residual functions derived in this section in the following corollary. \begin{corollary}[Facial residual functions concerning $\ensuremath{{\mathcal F}}_{\infty}$]\label{col:1dfaces_infty} Let $\ensuremath{\mathbf z}\in K_{\exp}^*$ with $z_x=0$ and $\{\ensuremath{\mathbf z} \}^\perp \cap K_{\exp}=\ensuremath{{\mathcal F}}_\infty$. \begin{enumerate}[{\rm (i)}] \item \label{lem:facialresidualsbeta:1}In the case when $z_y > 0$, let $\kappa_{\ensuremath{\mathbf z},t}$ be defined as in \eqref{haha2} with $\ensuremath{\mathfrak{g}} = |\cdot|$. Then the function $\psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}:\ensuremath{\mathbb R}_+\times\ensuremath{\mathbb R}_+\to \ensuremath{\mathbb R}_+$ given by \begin{equation*} \psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}(\epsilon,t):=\max \left\{\epsilon,\epsilon/\|\ensuremath{\mathbf z}\| \right\} + \kappa_{\ensuremath{\mathbf z},t}(\epsilon +\max \left\{\epsilon,\epsilon/\|\ensuremath{\mathbf z}\| \right\} ) \end{equation*} is a facial residual function for $K_{\exp}$ and $\ensuremath{\mathbf z}$ with respect to $K_{\exp}$. \item \label{lem:facialresidualsbeta:2} In the case when $z_y = 0$, let $\kappa_{\ensuremath{\mathbf z},t}$ be defined as in \eqref{haha2} with $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_\infty$ in \eqref{def:frakg}. Then the function $\psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}:\ensuremath{\mathbb R}_+\times\ensuremath{\mathbb R}_+\to \ensuremath{\mathbb R}_+$ given by \begin{equation*} \psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}(\epsilon,t):=\max \left\{\epsilon,\epsilon/\|\ensuremath{\mathbf z}\| \right\} + \kappa_{\ensuremath{\mathbf z},t}\ensuremath{\mathfrak{g}}_\infty(\epsilon +\max \left\{\epsilon,\epsilon/\|\ensuremath{\mathbf z}\| \right\} ) \end{equation*} is a facial residual function for $K_{\exp}$ and $\ensuremath{\mathbf z}$ with respect to $K_{\exp}$. \end{enumerate} \end{corollary} \subsubsection{$\ensuremath{\mathfrak{g}}$-amenability and FRF for the non-exposed face $\ensuremath{{\mathcal F}}_{ne}$} In this subsection we take a look at how the exponential cone fits inside the family of $\ensuremath{\mathfrak{g}}$-amenable cones. First, we observe that, for an arbitrary pointed cone $ \ensuremath{\mathcal{K}}$, extreme rays (one-dimensional faces) are always $1$-amenable faces. The argument is essentially the same as the proof that strictly convex cones are amenable in \cite[Proposition~9]{L17}. \begin{proposition}[Extreme rays are always amenable]\label{prop:err_oned} Let $ \ensuremath{\mathcal{F}} \mathrel{\unlhd} \ensuremath{\mathcal{K}}$ be an extreme ray of a pointed closed convex cone $ \ensuremath{\mathcal{K}}$. Then, there exists $\kappa > 0$ such that \[ \ensuremath{\operatorname{d}} (\ensuremath{\mathbf p}, \ensuremath{\mathcal{F}}) \leq \kappa \ensuremath{\operatorname{d}}(\ensuremath{\mathbf p}, \ensuremath{\mathcal{K}})\ \ \ \mbox{whenever \ $\ensuremath{\mathbf p} \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}$.} \] \end{proposition} \begin{proof} Because $ \ensuremath{\mathcal{F}}$ is one-dimensional, there exists $\ensuremath{\mathbf v}$ such that $\norm{\ensuremath{\mathbf v}} = 1$ and \[ \ensuremath{\mathcal{F}} = \{t\ensuremath{\mathbf v} \mid t \geq 0 \}. \] Let $\ensuremath{\mathbf p} \in \ensuremath{\mathrm{span}\,} \ensuremath{\mathcal{F}}$. Then $\ensuremath{\mathbf p} = t\ensuremath{\mathbf v}$ for some $t \in \ensuremath{\mathbb R}.$ First, suppose that $t < 0$ and let $\ensuremath{\mathbf u} \in \ensuremath{\mathcal{K}}$ such that \[ \ensuremath{\operatorname{d}} (-\ensuremath{\mathbf v}, \ensuremath{\mathcal{K}}) = \norm{\ensuremath{\mathbf u} + \ensuremath{\mathbf v}}. \] With that, when $t < 0$, we have \[ \ensuremath{\operatorname{d}}(t \ensuremath{\mathbf v}, \ensuremath{\mathcal{F}}) = -t, \qquad \ensuremath{\operatorname{d}}(t \ensuremath{\mathbf v}, \ensuremath{\mathcal{K}}) = \ensuremath{\operatorname{d}}(-t(-\ensuremath{\mathbf v}), \ensuremath{\mathcal{K}})= -t \norm{\ensuremath{\mathbf u} + \ensuremath{\mathbf v}}. \] Because $ \ensuremath{\mathcal{K}}$ is pointed, $-\ensuremath{\mathbf v} \not \in \ensuremath{\mathcal{K}}$, and so $\ensuremath{\operatorname{d}} (-\ensuremath{\mathbf v}, \ensuremath{\mathcal{K}}) = \norm{\ensuremath{\mathbf u} + \ensuremath{\mathbf v}}$ is nonzero. Thus the two equalities above may be combined to obtain \begin{equation} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf p}, \ensuremath{\mathcal{F}}) = \frac{1}{\norm{\ensuremath{\mathbf u} + \ensuremath{\mathbf v}}} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf p}, \ensuremath{\mathcal{K}}). \label{eq:oned_bound} \end{equation} To conclude, we observe that if $t \geq 0$ then \eqref{eq:oned_bound} also holds because $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf p}, \ensuremath{\mathcal{F}}) = \ensuremath{\operatorname{d}}(\ensuremath{\mathbf p}, \ensuremath{\mathcal{K}}) = 0$. This shows that it is enough to take $\kappa$ to be $\frac{1}{\norm{\ensuremath{\mathbf u} + \ensuremath{\mathbf v}}}$. \end{proof} In view of Proposition~\ref{prop:err_oned}, we have the following result. \begin{proposition}[$\ensuremath{\mathfrak{g}}$-amenability and the exponential cone]\label{prop:exp_am} Let $ \ensuremath{\mathcal{F}} \mathrel{\unlhd} \ensuremath{K_{\exp}}$ be a nontrivial face. If $ \ensuremath{\mathcal{F}}$ is an extreme ray, then it is $1$-amenable. If $ \ensuremath{\mathcal{F}} = \ensuremath{{\mathcal F}}_{-\infty}$ (see \eqref{eq:exp_2d}) then $ \ensuremath{\mathcal{F}}$ is $\ensuremath{\mathfrak{g}}_{-\infty}$-amenable, where $\ensuremath{\mathfrak{g}}_{-\infty}$ is as in \eqref{d:entropy}. $\ensuremath{{\mathcal F}}_{-\infty}$ is also $\alpha$-amenable for $\alpha \in (0,1)$ but not for $\alpha = 1$. \end{proposition} \begin{proof} By Proposition~\ref{prop:err_oned}, if $ \ensuremath{\mathcal{F}}$ is an extreme ray, then it is $1$-amenable. Next, suppose that $ \ensuremath{\mathcal{F}}$ is the unique 2D face $ \ensuremath{{\mathcal F}}_{-\infty}$. Let $\ensuremath{\mathbf z} \in \ensuremath{K_{\exp}}^*$ be such that $z_x = z_z = 0$ and $z_y = 1$. By Theorem~\ref{thm:entropic}, given $\eta > 0$, there exists a constant $\kappa > 0$ such that \begin{equation}\label{eq:2d_am} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},\ensuremath{{\mathcal F}}_{-\infty})\le \kappa \cdot\ensuremath{\mathfrak{g}}_{-\infty}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},\ensuremath{K_{\exp}}))\ \ \ \mbox{whenever $\ensuremath{\mathbf q}\in \{\ensuremath{\mathbf z}\}^\perp\cap B(\eta)$.} \end{equation} By \eqref{Fneginfinity}, we have that $\ensuremath{\mathrm{span}\,} \ensuremath{{\mathcal F}}_{-\infty} = \{\ensuremath{\mathbf z}\}^\perp$. Therefore, \eqref{eq:2d_am} is valid for $\ensuremath{\mathbf q} \in (\ensuremath{\mathrm{span}\,} \ensuremath{{\mathcal F}}_{-\infty})\cap B(\eta)$, showing that $\ensuremath{{\mathcal F}}_{-\infty}$ is $\ensuremath{\mathfrak{g}}_{-\infty}$-amenable. Similarly, from Corollary~\ref{col:2d_hold}, we conclude that $\ensuremath{{\mathcal F}}_{-\infty}$ is $\alpha$-amenable for $\alpha \in (0,1)$. However, $\ensuremath{{\mathcal F}}_{-\infty}$ cannot be $1$-amenable because, this would imply that $\ensuremath{K_{\exp}}$ is an amenable cone, i.e., all the faces of $\ensuremath{K_{\exp}}$ would be $1$-amenable. This cannot happen because amenable cones must be facially exposed (see \cite[Proposition~17]{L17}) and $\ensuremath{K_{\exp}}$ has the non-exposed face $\ensuremath{{\mathcal F}}_{ne}$. \end{proof} Clarifying how the exponential cone stands with respect to $\ensuremath{\mathfrak{g}}$-amenability allows us to compute a facial residual function for the remaining face $\ensuremath{{\mathcal F}}_{ne}$ with little effort. \begin{corollary}[The facial residual function for $\ensuremath{{\mathcal F}}_{ne}$]\label{col:frf_ne} Let $\ensuremath{\mathbf z} \in \ \ensuremath{{\mathcal F}}_{-\infty}^*$ be such that $\ensuremath{{\mathcal F}}_{ne} = \ensuremath{{\mathcal F}}_{-\infty} \cap \{\ensuremath{\mathbf z}\}^\perp$. Let $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_{-\infty}$ in \eqref{d:entropy}, or $\ensuremath{\mathfrak{g}} = |\cdot|^\alpha$ for $\alpha \in (0,1)$. Then, there exists a nonnegative monotone nondecreasing function $\sigma: \ensuremath{\mathbb R}_+ \to \ensuremath{\mathbb R}_+$ such that \[ \psi _{\ensuremath{{\mathcal F}}_{-\infty},\ensuremath{\mathbf z} }(\epsilon,t) \coloneqq \sigma(t) \epsilon + \sigma(t)\ensuremath{\mathfrak{g}}(\epsilon) \] is a facial residual function for $\ensuremath{{\mathcal F}}_{-\infty}$ and $\ensuremath{\mathbf z}$ with respect to $\ensuremath{K_{\exp}}$. \end{corollary} \begin{proof} Since $\ensuremath{{\mathcal F}}_{ne}$ is a face of the polyhedral cone $\ensuremath{{\mathcal F}}_{-\infty}$, there exists a constant $\kappa > 0$ such that \[\hat \psi _{\ensuremath{{\mathcal F}}_{-\infty},\ensuremath{\mathbf z}}(\epsilon,t) \coloneqq \kappa \epsilon \] is a facial residual function for $\ensuremath{{\mathcal F}}_{-\infty}$ and $\ensuremath{\mathbf z}$ with respect to $\ensuremath{{\mathcal F}}_{-\infty}$, see Example~\ref{ex:frf}. By Proposition~\ref{prop:exp_am}, $\ensuremath{{\mathcal F}}_{-\infty}$ is a $\ensuremath{\mathfrak{g}}$-amenable face of $\ensuremath{K_{\exp}}$. Then, we invoke Proposition~\ref{prop:frf_am}, to obtain a monotone nondecreasing function $\hat \sigma: \ensuremath{\mathbb R}_+ \to \ensuremath{\mathbb R}_+$ such that \[ \hat \psi _{\ensuremath{{\mathcal F}}_{-\infty},\ensuremath{\mathbf z} }(\epsilon,t) \coloneqq \epsilon + \kappa \hat \sigma(t) \max\{\epsilon,\ensuremath{\mathfrak{g}}(2\epsilon)\}, \] is a facial residual function for $\ensuremath{{\mathcal F}}_{-\infty}$ and $\ensuremath{\mathbf z}$ with respect to $\ensuremath{K_{\exp}}$. Then, recalling \eqref{d:entropy_p}, upon letting $\sigma(t) \coloneqq 1+\kappa\hat\sigma(t)\max\{1,2^\alpha,L\}$, where $L$ is as in \eqref{d:entropy_p}, we have that $\psi _{\ensuremath{{\mathcal F}}_{-\infty},\ensuremath{\mathbf z} }(\epsilon,t) \coloneqq \sigma(t)\epsilon + \sigma(t)\ensuremath{\mathfrak{g}}{(\epsilon)}$ is a facial residual function for $\ensuremath{{\mathcal F}}_{-\infty}$ and $\ensuremath{\mathbf z}$ with respect to $\ensuremath{K_{\exp}}$. \end{proof} \subsection{Error bounds}\label{sec:exp_err} In this subsection, we return to the feasibility problem \eqref{eq:feas} and consider the case where $ \ensuremath{\mathcal{K}} = \ensuremath{K_{\exp}}$. We now have all the tools for obtaining error bounds. Recalling Definition~\ref{def:hold}, we can state the following result. \begin{theorem}[Error bounds for \eqref{eq:feas} with $ \ensuremath{\mathcal{K}} = \ensuremath{K_{\exp}}$]\label{thm:main_err} Let $ \ensuremath{\mathcal{L}} \subseteq \ensuremath{\mathbb R}^3$ be a subspace and $\ensuremath{\mathbf a} \in \ensuremath{\mathbb R}^3$ such that $( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \cap \ensuremath{K_{\exp}} \neq \emptyset$. Then the following items hold. \begin{enumerate} \item The distance to the PPS condition of $\ensuremath{K_{\exp}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ satisfies $d_{\text{PPS}}(\ensuremath{K_{\exp}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}) \leq 1$. \item\label{thm:mainii} If $d_{\text{PPS}}(\ensuremath{K_{\exp}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})=0 $, then $\ensuremath{K_{\exp}}$ and $ \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ satisfy a Lipschitzian error bound. \item\label{thm:mainiii} Suppose $d_{\text{PPS}}(\ensuremath{K_{\exp}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})=1$ and let $ \ensuremath{\mathcal{F}} \subsetneq \ensuremath{K_{\exp}}$ be a chain of faces of length $2$ satisfying items {\rm(ii)} and {\rm(iii)} of Proposition~\ref{prop:fra_poly}. We have the following possibilities. \begin{enumerate}[label={\rm (\alph*)}] \item\label{thm:mainiiia} If $ \ensuremath{\mathcal{F}} = \ensuremath{{\mathcal F}}_{-\infty}$ then $\ensuremath{K_{\exp}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ satisfy an entropic error bound as in \eqref{eq:entropic_err}. In addition, for all $\alpha \in (0,1)$, a uniform H\"olderian error bound with exponent $\alpha$ holds. \item\label{thm:mainiiib} If $ \ensuremath{\mathcal{F}} = \ensuremath{{\mathcal F}}_{\beta}$, with $\beta \in \ensuremath{\mathbb R}$, then $\ensuremath{K_{\exp}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ satisfy a uniform H\"olderian error bound with exponent $1/2$. \item\label{thm:mainiiic} Suppose that $ \ensuremath{\mathcal{F}} = \ensuremath{{\mathcal F}}_{\infty}$. If there exists $\ensuremath{\mathbf z}\in \ensuremath{K_{\exp}}^* \cap \ensuremath{\mathcal{L}}^\perp \cap \{a\}^\perp$ with $z_x=0$, $z_y > 0$ and $z_z>0$ then $\ensuremath{K_{\exp}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ satisfy a Lipschitzian error bound. Otherwise, $\ensuremath{K_{\exp}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ satisfy a log-type error bound as in \eqref{eq:log_err}. \item \label{thm:mainivc} If $ \ensuremath{\mathcal{F}} = \{\mathbf{0} \}$, then $\ensuremath{K_{\exp}}$ and $ \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ satisfy a Lipschitzian error bound. \end{enumerate} \end{enumerate} \end{theorem} \begin{proof} (i): All proper faces of $\ensuremath{K_{\exp}}$ are polyhedral, therefore $\ell_{\text{poly}}(\ensuremath{K_{\exp}}) = 1$. By item \ref{prop:fra_poly:1} of Proposition~\ref{prop:fra_poly}, there exists a chain of length $2$ satisfying item \ref{prop:fra_poly:3} of Proposition~\ref{prop:fra_poly}. Therefore, $d_{\text{PPS}}(\ensuremath{K_{\exp}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}) \leq 1$. (ii): If $d_{\text{PPS}}(\ensuremath{K_{\exp}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}) = 0$, it is because $\ensuremath{K_{\exp}}$ and $ \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ satisfy the PPS condition, which implies a Lipschitzian error bound by Proposition~\ref{prop:cq_er}. (iii): Next, suppose $d_{\text{PPS}}(\ensuremath{K_{\exp}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})=1$ and let $ \ensuremath{\mathcal{F}} \subsetneq \ensuremath{K_{\exp}}$ be a chain of faces of length $2$ satisfying items \ref{prop:fra_poly:2} and \ref{prop:fra_poly:3} of Proposition~\ref{prop:fra_poly}, together with $\ensuremath{\mathbf z}\in K_{\exp}^* \cap \ensuremath{\mathcal{L}}^\perp \cap \{\ensuremath{\mathbf a}\}^\perp$ such that \[ \ensuremath{\mathcal{F}} = \ensuremath{K_{\exp}} \cap \{\ensuremath{\mathbf z}\}^\perp. \] Since positively scaling $\ensuremath{\mathbf z}$ does not affect the chain of faces, we may assume that $\norm{\ensuremath{\mathbf z}} = 1$. Also, in what follows, for simplicity, we define \[ d(\ensuremath{\mathbf x}) \coloneqq \max \{\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}), \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{K_{\exp}}) \}. \] Then, we prove each item by applying Theorem~\ref{theo:err} with the corresponding facial residual function. \begin{enumerate}[{\rm (a)}] \item If $ \ensuremath{\mathcal{F}} = \ensuremath{{\mathcal F}}_{-\infty}$, the facial residual functions are given by Corollary~\ref{col:frf_2dface_entropic}. First we consider the case where $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_{-\infty}$ and we have \begin{equation*} \psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}(\epsilon,t):= \epsilon + \kappa_{\ensuremath{\mathbf z},t}\ensuremath{\mathfrak{g}}_{-\infty}(2\epsilon ), \end{equation*} where $\ensuremath{\mathfrak{g}}_{-\infty}$ is as in \eqref{d:entropy}. Then, if $\psi$ is a positively rescaling of $\psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}$, using the monotonicity of $\ensuremath{\mathfrak{g}}_{-\infty}$ and of $\kappa_{\ensuremath{\mathbf z},t}$ as a function of $t$, we conclude that there exists $\widehat M > 0$ such that \begin{equation*} \psi(\epsilon,t) \leq \widehat M \epsilon + \widehat M \kappa_{\ensuremath{\mathbf z}, \widehat M t}\ensuremath{\mathfrak{g}}_{-\infty}(\widehat M\epsilon ). \end{equation*} Invoking Theorem~\ref{theo:err}, using the monotonicity of all functions involved in the definition of $\psi$ and recalling \eqref{d:entropy_p}, we conclude that for every bounded set $B$, there exists $\kappa _B > 0$ \begin{equation}\label{eq:entropic_err} \ensuremath{\operatorname{d}}\left(\ensuremath{\mathbf x}, ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \cap \ensuremath{K_{\exp}}\right) \leq \kappa_{B}\ensuremath{\mathfrak{g}}_{-\infty}(d(\ensuremath{\mathbf x})), \qquad \forall \ensuremath{\mathbf x} \in B, \end{equation} which shows that an entropic error bound holds. Next, we consider the case $\ensuremath{\mathfrak{g}} = |\cdot|^{\alpha}$. Given $\alpha \in (0,1)$, we have the following facial residual function: \begin{equation*} \psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}(\epsilon,t):= \epsilon + \kappa_{\ensuremath{\mathbf z},t}2^\alpha \epsilon^\alpha, \end{equation*} where $\kappa_{\ensuremath{\mathbf z},t}$ is defined as in \eqref{haha2}. Invoking Theorem~\ref{theo:err}, we conclude that for every bounded set $B$, there exists $\kappa _B > 0$ such that \[ \ensuremath{\operatorname{d}}\left(\ensuremath{\mathbf x}, ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \cap \ensuremath{K_{\exp}}\right) \leq \kappa _B d(\ensuremath{\mathbf x}) + \kappa_{B} d(\ensuremath{\mathbf x})^\alpha, \qquad \forall \ensuremath{\mathbf x} \in B, \] In addition, for $\ensuremath{\mathbf x} \in B$, we have $d(\ensuremath{\mathbf x}) \leq d(\ensuremath{\mathbf x})^\alpha {M}$, where $M = \sup _{\ensuremath{\mathbf x} \in B} d(\ensuremath{\mathbf x})^{1-\alpha}$. In conclusion, for $\kappa = 2\kappa_{B}\max\{M,1\}$, we have \[ \ensuremath{\operatorname{d}}\left(\ensuremath{\mathbf x}, ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \cap \ensuremath{K_{\exp}}\right) \leq \kappa d(\ensuremath{\mathbf x})^\alpha, \qquad \forall \ensuremath{\mathbf x} \in B. \] That is, a uniform H\"olderian error bound holds with exponent $\alpha$. \item If $ \ensuremath{\mathcal{F}} = \ensuremath{{\mathcal F}}_{\beta}$, with $\beta \in \ensuremath{\mathbb R}$, then the facial residual function is given by Corollary~\ref{col:frf_fb}, that is, we have \begin{equation*} \psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}(\epsilon,t) := \epsilon + \kappa_{\ensuremath{\mathbf z},t}\sqrt{2} \epsilon^{1/2}, \end{equation*} Then, following the same argument as in the second half of item (a), we conclude that a uniform H\"olderian error bound holds with exponent $1/2$. \item If $ \ensuremath{\mathcal{F}} = \ensuremath{{\mathcal F}}_{\infty}$, the facial residual functions are given by Corollary~\ref{col:1dfaces_infty} and they depend on $\ensuremath{\mathbf z}$. Since $ \ensuremath{\mathcal{F}} = \ensuremath{{\mathcal F}}_{\infty}$, we must have $z_x = 0$ and $z_z > 0$, see Section~\ref{sec:exposed}. The deciding factor is whether $z_y$ is positive or zero. If $z_y > 0$, then we have the following facial residual function: \begin{equation*} \psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}(\epsilon,t):= (1+2 \kappa_{\ensuremath{\mathbf z},t})\epsilon, \end{equation*} where $\kappa_{\ensuremath{\mathbf z},t}$ is defined as in \eqref{haha2}. In this case, analogously to items (a) and (b) we have a Lipschitzian error bound. If $z_y = 0$, we have \begin{equation*} \psi_{ \ensuremath{\mathcal{K}},\ensuremath{\mathbf z}}(\epsilon,t):= \epsilon + \kappa_{\ensuremath{\mathbf z},t}\ensuremath{\mathfrak{g}}_\infty(2\epsilon ), \end{equation*} where $\ensuremath{\mathfrak{g}}_\infty$ is as in \eqref{def:frakg}. Analogous to the proof of item (a) but making use of \eqref{def:frakg_p} in place of \eqref{d:entropy_p}, we conclude that for every bounded set $B$, there exists $\kappa _B > 0$ such that \begin{equation}\label{eq:log_err} \ensuremath{\operatorname{d}}\left(\ensuremath{\mathbf x}, ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \cap \ensuremath{K_{\exp}}\right) \leq \kappa_{B}\ensuremath{\mathfrak{g}}_\infty(d(\ensuremath{\mathbf x}))), \qquad \forall \ensuremath{\mathbf x} \in B. \end{equation} \item See \cite[Proposition~27]{L17}. \end{enumerate} \end{proof} \begin{remark}[Tightness of Theorem~\ref{thm:main_err}]\label{rem:opt} We will argue that Theorem~\ref{thm:main_err} is tight by showing that for every situation described in item (iii), there is a specific choice of $ \ensuremath{\mathcal{L}}$ and a sequence $\{\ensuremath{\mathbf w}^k\}$ in $ \ensuremath{\mathcal{L}}$ with $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k,\ensuremath{K_{\exp}}) \to 0$ along which the corresponding error bound for $\ensuremath{K_{\exp}}, \ensuremath{\mathcal{L}}$ is off by at most a multiplicative constant. \begin{enumerate}[{\rm (a)}] \item\label{opt:a} Let $ \ensuremath{\mathcal{L}} = \ensuremath{\mathrm{span}\,} \ensuremath{{\mathcal F}}_{-\infty} = \{(x,y,z) \mid y = 0 \}$ (see \eqref{eq:exp_2d}) and consider the sequence $\{\ensuremath{\mathbf w}^k\}$ where $\ensuremath{\mathbf w}^k = ((1/k)\ln(k),0,1)$, for every $k \in \ensuremath{\mathbb N}$. Then, $ \ensuremath{\mathcal{L}} \cap \ensuremath{K_{\exp}} = \ensuremath{{\mathcal F}}_{-\infty}$ and we are under the conditions of item~\ref{thm:mainiii}\ref{thm:mainiiia} of Theorem~\ref{thm:main_err}. Since $\{\ensuremath{\mathbf w}^k\} =: B \subseteq \ensuremath{\mathcal{L}}$, there exists $\kappa_B$ such that \[ \ensuremath{\operatorname{d}}\left(\ensuremath{\mathbf w}^k, \ensuremath{\mathcal{L}} \cap \ensuremath{K_{\exp}}\right) \leq \kappa_{B}\ensuremath{\mathfrak{g}}_{-\infty}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k, \ensuremath{K_{\exp}})), \quad \forall k \in \ensuremath{\mathbb N}. \] Then, the projection of $\ensuremath{\mathbf w}^k$ onto $\ensuremath{{\mathcal F}}_{-\infty}$ is given by $(0,0,1)$. Therefore, \[ \frac{\ln(k)}{k} = \ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k, \ensuremath{\mathcal{L}}\cap \ensuremath{K_{\exp}}) \leq \kappa_{B}\ensuremath{\mathfrak{g}}_{-\infty}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k, \ensuremath{K_{\exp}})). \] Let $\ensuremath{\mathbf v}^k = ((1/k)\ln(k),(1/k),1)$ for every $k$. Then, we have $\ensuremath{\mathbf v}^k \in \ensuremath{K_{\exp}}$. Therefore, $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k, \ensuremath{K_{\exp}}) \leq 1/k$. In view of the definition of $\ensuremath{\mathfrak{g}}_{-\infty}$ (see \eqref{d:entropy}), we conclude that for some constant $L>0$ and for large enough $k$ we have \[ \frac{\ln(k)}{k} = \ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k, \ensuremath{\mathcal{L}}\cap \ensuremath{K_{\exp}}) \leq \kappa_{B}\ensuremath{\mathfrak{g}}_{-\infty}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k, \ensuremath{K_{\exp}})) \leq L\frac{\ln(k)}{k}, \] which shows that the error bound is tight up to a constant. \item Let $\beta \in \ensuremath{\mathbb R}$ and let $\ensuremath{\widehat{\mathbf z}}$, $\ensuremath{\widehat{\mathbf p}}$ and $\ensuremath{\widehat{\mathbf f}}$ be as in \eqref{veczfp}. Let $ \ensuremath{\mathcal{L}} = \{\ensuremath{\mathbf z}\}^\perp$ with $z_x < 0$ so that $\ensuremath{K_{\exp}} \cap \ensuremath{\mathcal{L}} = \ensuremath{{\mathcal F}}_{\beta}$. We are then under the conditions of item~\ref{thm:mainiii}\ref{thm:mainiiib} of Theorem~\ref{thm:main_err}. We consider the following sequences \begin{equation*} \ensuremath{\mathbf v}^k = \begin{bmatrix} 1-\beta +1/k\\ 1\\ e^{1-\beta + 1/k} \end{bmatrix},\quad \ensuremath{\mathbf w}^k = P_{\{\ensuremath{\mathbf z}\}^\perp}\ensuremath{\mathbf v}^k,\quad \ensuremath{\mathbf u}^k = P_{\ensuremath{{\mathcal F}}_\beta}\ensuremath{\mathbf w}^k. \end{equation*} For every $k$ we have $\ensuremath{\mathbf v}^k \in \partial \ensuremath{K_{\exp}} \setminus \ensuremath{{\mathcal F}}_{\beta}$, and $\ensuremath{\mathbf v}^k \neq \ensuremath{\mathbf w}^k$ (because otherwise, we would have $\ensuremath{\mathbf v}^k \in \ensuremath{K_{\exp}} \cap \{\ensuremath{\mathbf z}\}^\perp = \ensuremath{{\mathcal F}}_{\beta}$). In addition, we have $\ensuremath{\mathbf v}^k \to \ensuremath{\widehat{\mathbf f}}$ and, since $\ensuremath{\widehat{\mathbf f}} \in \ensuremath{{\mathcal F}}_{\beta}$, we have $\ensuremath{\mathbf w}^k \to \ensuremath{\widehat{\mathbf f}}$ as well. Next, notice that we have $\langle \ensuremath{\widehat{\mathbf f}}, \ensuremath{\mathbf v}^k \rangle \geq 0$ for $k$ sufficiently large and $|v_x^k/v_y^k - (1-\beta)| \rightarrow 0$. Then, following the computations outlined in case~\ref{ebsubcasev0} of the proof of Theorem~\ref{thm:nonzerogamma} and letting $\zeta_k\coloneqq v_x^k/v_y^k$, we have from \eqref{hahahehe1} and \eqref{newlyadded} that $h_2(\zeta_k)\neq 0$ for all large $k$ (hence, $\ensuremath{\mathbf w}^k\neq \ensuremath{\mathbf u}^k$ for all large $k$), and that \begin{equation}\label{eq:beta_limit} L_{\beta} \coloneqq \lim_{k \rightarrow \infty}\frac{\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\|^{\frac{1}{2}}}{\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf u}^k\|}=\lim_{k \rightarrow \infty}\frac{\|\ensuremath{\widehat{\mathbf p}}\|}{\|\ensuremath{\widehat{\mathbf z}}\|^{\frac{1}{2}}}\frac{|h_1(\zeta_k)|^{\frac{1}{2}}}{|h_2(\zeta_k)|} = \frac{\|\ensuremath{\widehat{\mathbf p}}\|}{\|\ensuremath{\widehat{\mathbf z}}\|^{\frac{1}{2}}}\frac1{\sqrt{2}(e^{\beta-1} + (\beta^2+1)e^{1-\beta})} \in(0,\infty), \end{equation} where the latter equality is from \eqref{taylor_limit}. On the other hand, from item~\ref{thm:mainiii}\ref{thm:mainiiib} of Theorem~\ref{thm:main_err}, for $B \coloneqq \{\ensuremath{\mathbf w}^k\}$, there exists $\kappa_B > 0$ such that for all $k\in \ensuremath{\mathbb N}$, \[ \|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf u}^k\| = \ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k, \ensuremath{\mathcal{L}}\cap\ensuremath{K_{\exp}}) \leq \kappa_B \ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k,\ensuremath{K_{\exp}})^{\frac{1}{2}} \leq \kappa_B\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\|^{\frac{1}{2}} \] However, from \eqref{eq:beta_limit}, for large enough $k$, we have $\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf u}^k\| \geq 1/(2L_{\beta})\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\|^{\frac{1}{2}}$. Therefore, for large enough $k$ we have \[ \frac{1}{2L_{\beta}}\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\|^{\frac{1}{2}} \leq \ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k, \ensuremath{\mathcal{L}}\cap\ensuremath{K_{\exp}})\leq \kappa_B \ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k,\ensuremath{K_{\exp}})^{\frac{1}{2}} \leq \kappa_B\|\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\|^{\frac{1}{2}}. \] Recalling that $\ensuremath{\mathbf w}^k\neq \ensuremath{\mathbf v}^k$ for all $k$ and $\ensuremath{\mathbf w}^k-\ensuremath{\mathbf v}^k\to 0$, we see that the error bound is tight up to a constant. \item Let $\ensuremath{\mathbf z}= (0,0,1)$ and $ \ensuremath{\mathcal{L}} = \{(x,y,0) \mid x,y \in \ensuremath{\mathbb R} \} = \{\ensuremath{\mathbf z}\}^\perp$. Then, from \eqref{Finfinity}, we have $ \ensuremath{\mathcal{L}} \cap \ensuremath{K_{\exp}} = \ensuremath{{\mathcal F}}_{\infty}$. We are then under case \ref{thm:mainiii}\ref{thm:mainiiic} of Theorem~\ref{thm:main_err}. Because there is no $\hat \ensuremath{\mathbf z} \in \ensuremath{\mathcal{L}}^\perp $ with $\hat z _y > 0$, we have a log-type error bound as in \eqref{eq:log_err}. We proceed as in item \ref{opt:a} using sequences such that $\ensuremath{\mathbf w}^k=(-1,1/k,0)$, $\ensuremath{\mathbf v}^k=(-1,1/k,(1/k)e^{-k})$, $\ensuremath{\mathbf u}^k=(-1,0,0)$, for every $k$. Note that $\ensuremath{\mathbf w}^k \in \ensuremath{\mathcal{L}}, \ensuremath{\mathbf v}^k \in \ensuremath{K_{\exp}}$ and $\proj{\ensuremath{{\mathcal F}}_\infty}(\ensuremath{\mathbf w}^k) = \ensuremath{\mathbf u}^k$, for every $k$. Therefore, there exists $\kappa_B > 0$ such that \begin{equation*} \frac{1}{k} = \ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k, \ensuremath{\mathcal{L}} \cap \ensuremath{K_{\exp}}) \leq \kappa_B \ensuremath{\mathfrak{g}}_{\infty}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k,\ensuremath{K_{\exp}}))\leq \kappa_{B}\ensuremath{\mathfrak{g}}_{\infty}\left(\frac{1}{ke^k}\right), \quad \forall k \in \ensuremath{\mathbb N}. \end{equation*} In view of the definition of $\ensuremath{\mathfrak{g}}_{\infty}$ (see \eqref{def:frakg}), there exists $L > 0$ such that for large enough $k$ we have \begin{equation*} \frac{1}{k} = \ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k, \ensuremath{\mathcal{L}} \cap \ensuremath{K_{\exp}}) \le \kappa_B \ensuremath{\mathfrak{g}}_{\infty}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf w}^k,\ensuremath{K_{\exp}})) \leq \frac{L}{k}, \end{equation*} which shows the tightness of the error bound. \end{enumerate} Note that a Lipschitz error bound is always tight up to a constant, because $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}\cap ( \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})) \geq \max\{\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}),\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})\}$. Therefore, the error bounds in items~\ref{thm:mainii}, \ref{thm:mainiii}\ref{thm:mainivc} and in the first half of \ref{thm:mainiii}\ref{thm:mainiiic} are tight. \end{remark} Sometimes we may need to consider direct products of multiple copies of $\ensuremath{K_{\exp}}$ in order to model certain problems, i.e., our problem of interest could have the following shape: \begin{equation*} \text{find} \quad \ensuremath{\mathbf x} \in ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \cap \ensuremath{\mathcal{K}}, \label{eq:multiple_exp} \end{equation*} where $ \ensuremath{\mathcal{K}} = \ensuremath{K_{\exp}} \times \cdots \times \ensuremath{K_{\exp}}$ is a direct product of $m$ exponential cones. Fortunately, we already have all the tools required to extend Theorem~\ref{thm:main_err} and compute error bounds for this case too. We recall that the faces of a direct product of cones are direct products of the faces of the individual cones.\footnote{Here is a sketch of the proof. If $ \ensuremath{\mathcal{F}}^1 \mathrel{\unlhd} \ensuremath{\mathcal{K}}^1, \ensuremath{\mathcal{F}}^{2} \mathrel{\unlhd} \ensuremath{\mathcal{K}}^2$, then the definition of face implies that $ \ensuremath{\mathcal{F}}^1 \times \ensuremath{\mathcal{F}}^{2} \mathrel{\unlhd} \ensuremath{\mathcal{K}}^1 \times \ensuremath{\mathcal{K}}^2$. For the converse, let $ \ensuremath{\mathcal{F}} \mathrel{\unlhd} \ensuremath{\mathcal{K}}^1 \times \ensuremath{\mathcal{K}}^2$ and let $ \ensuremath{\mathcal{F}} ^1, \ensuremath{\mathcal{F}}^2$ be the projection of $ \ensuremath{\mathcal{F}}$ onto the first variable and second variables, respectively. Suppose that $\ensuremath{\mathbf x},\ensuremath{\mathbf y} \in \ensuremath{\mathcal{K}}^1$ are such that $\ensuremath{\mathbf x}+\ensuremath{\mathbf y} \in \ensuremath{\mathcal{F}}^1$. Then, $(\ensuremath{\mathbf x}+\ensuremath{\mathbf y},\ensuremath{\mathbf z}) \in \ensuremath{\mathcal{F}}$ for some $\ensuremath{\mathbf z} \in \ensuremath{\mathcal{K}}^2$. Since $(\ensuremath{\mathbf x}+\ensuremath{\mathbf y},\ensuremath{\mathbf z}) = (\ensuremath{\mathbf x},\ensuremath{\mathbf z}/2) + (\ensuremath{\mathbf y},\ensuremath{\mathbf z}/2)$ and $ \ensuremath{\mathcal{F}}$ is a face, we conclude that $(\ensuremath{\mathbf x},\ensuremath{\mathbf z}/2), (\ensuremath{\mathbf y},\ensuremath{\mathbf z}/2) \in \ensuremath{\mathcal{F}}$ and $\ensuremath{\mathbf x}, \ensuremath{\mathbf y} \in \ensuremath{\mathcal{F}}^1$. Therefore $ \ensuremath{\mathcal{F}}^1 \mathrel{\unlhd} \ensuremath{\mathcal{K}}^1$ and, similarly, $ \ensuremath{\mathcal{F}}^2 \mathrel{\unlhd} \ensuremath{\mathcal{K}}^2$. Then, the equality $ \ensuremath{\mathcal{F}} = \ensuremath{\mathcal{F}}^1 \times \ensuremath{\mathcal{F}}^2$ is proven using the definition of face and the fact that $(\ensuremath{\mathbf x},\ensuremath{\mathbf z}) = (\ensuremath{\mathbf x},0) + (0,\ensuremath{\mathbf z})$. } Therefore, using Propositions~\ref{prop:direct_prod}, \ref{prop:frf_prod} and \ref{prop:exp_am}, we are able to compute all facial residual functions for all the faces of $ \ensuremath{\mathcal{K}}$. Once they are obtained we can invoke Theorem~\ref{theo:err}. Unfortunately, there is quite a number of different cases one must consider, so we cannot give a concise statement of an all-encompassing tight error bound result. We will, however, given an error bound result under the following \emph{simplifying assumption of non-exceptionality} or SANE. \begin{assumption}[SANE: simplifying assumption of non-exceptionality] Suppose \eqref{eq:feas} is feasible with $ \ensuremath{\mathcal{K}} = \ensuremath{K_{\exp}} \times \cdots \times \ensuremath{K_{\exp}}$ being a direct product of $m$ exponential cones. We say that $ \ensuremath{\mathcal{K}}$ and $ \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ satisfy the \emph{simplifying assumption of non-exceptionality} (SANE) if there exists a chain of faces $ \ensuremath{\mathcal{F}} _{\ell} \subsetneq \cdots \subsetneq \ensuremath{\mathcal{F}}_1 = \ensuremath{\mathcal{K}} $ as in Proposition~\ref{prop:fra_poly} with $\ell - 1 = {d_{\text{PPS}}( \ensuremath{\mathcal{K}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})}$ such that for all $i$, the exceptional face $\ensuremath{{\mathcal F}}_{\infty}$ of $\ensuremath{K_{\exp}}$ never appears as one of the blocks of $ \ensuremath{\mathcal{F}}_{i}$. \end{assumption} Under SANE, we can state the following result. \begin{theorem}[Error bounds for direct products of exponential cones]\label{theo:sane} Suppose \eqref{eq:feas} is feasible with $ \ensuremath{\mathcal{K}} = \ensuremath{K_{\exp}} \times \cdots \times \ensuremath{K_{\exp}}$ being a direct product of $m$ exponential cones. Then the following hold. \begin{enumerate} \item The distance to the PPS condition of $ \ensuremath{\mathcal{K}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ satisfies $d_{\text{PPS}}( \ensuremath{\mathcal{K}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}) \leq m$. \item If SANE is satisfied then $ \ensuremath{\mathcal{K}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}$ satisfy a uniform mixed entropic-H\"olderian error bound as in \eqref{eq:mixed_err}. \end{enumerate} \end{theorem} \begin{proof} (i): All proper faces of $\ensuremath{K_{\exp}}$ are polyhedral, therefore $\ell_{\text{poly}}(\ensuremath{K_{\exp}}) = 1$. By item \ref{prop:fra_poly:1} of Proposition~\ref{prop:fra_poly}, there exists a chain of length $\ell$ satisfying item \ref{prop:fra_poly:3} of Proposition~\ref{prop:fra_poly} such that $\ell-1 \leq m$. Therefore, $d_{\text{PPS}}( \ensuremath{\mathcal{K}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})\leq \ell-1 \leq m$. (ii): If SANE is satisfied, then there exists a chain $ \ensuremath{\mathcal{F}} _{\ell} \subsetneq \cdots \subsetneq \ensuremath{\mathcal{F}}_1 = \ensuremath{\mathcal{K}} $ of length $\ell \leq m +1$ as in Proposition~\ref{prop:fra_poly}, together with the corresponding $\ensuremath{\mathbf z}_{1},\ldots,\ensuremath{\mathbf z}_{\ell-1}$. Furthermore, the exceptional face $\ensuremath{{\mathcal F}}_{\infty}$ never appears as one of the blocks of the $ \ensuremath{\mathcal{F}} _i$. In what follows, for simplicity, we define \[ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}) \coloneqq \max \{\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a}), \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, \ensuremath{\mathcal{K}}) \}. \] Then, we invoke Theorem~\ref{theo:err}, which implies that given a bounded set $B$, there exists a constant $\kappa _B > 0$ such that \begin{equation}\label{eq:bound_sane} \ensuremath{\operatorname{d}}\left(\ensuremath{\mathbf x}, ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \cap \ensuremath{\mathcal{K}}\right) \leq \kappa _B (d(\ensuremath{\mathbf x})+\varphi(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}),M)), \end{equation} where $M = \sup _{\ensuremath{\mathbf x}\in B} \norm{\ensuremath{\mathbf x}}$ and there are two cases for $\varphi$. If $\ell = 1$, $\varphi$ is the function such that $\varphi(\epsilon,M) = \epsilon$. If $\ell \geq 2$, we have $\varphi = \psi _{{\ell-1}}\diamondsuit \cdots \diamondsuit \psi_{{1}}$, where $\psi _{i}$ is a (suitable positive rescaling of a) facial residual function for $ \ensuremath{\mathcal{F}}_{i}$ and $\ensuremath{\mathbf z}_i$ with respect to $ \ensuremath{\mathcal{K}}$. In the former case, the PPS condition is satisfied, we have a Lipschitzian error bound and we are done. We therefore assume that the latter case occurs with $\ell - 1 = {d_{\text{PPS}}( \ensuremath{\mathcal{K}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})}$. First, we compute facial residual functions for each $ \ensuremath{\mathcal{F}}_i$. In order to do that, we recall that each $ \ensuremath{\mathcal{F}}_{i}$ is a direct product $ \ensuremath{\mathcal{F}}_{i}^1\times \cdots \times \ensuremath{\mathcal{F}}_{i}^m$ where each $ \ensuremath{\mathcal{F}}_{i}^j$ is a face of $\ensuremath{K_{\exp}}$, excluding $\ensuremath{{\mathcal F}}_{\infty}$ by SANE. Therefore, a facial residual function for $ \ensuremath{\mathcal{F}}_{i}^j$ can be obtained from Corollary~\ref{col:frf_2dface_entropic}, \ref{col:frf_fb} or \ref{col:frf_ne}. In particular, taking the worst\footnote{$\sqrt{\cdot}$ is ``worse'' than $\ensuremath{\mathfrak{g}}_{-\infty}$ in that, near zero, $\sqrt{t} \geq \ensuremath{\mathfrak{g}}_{-\infty}(t)$. The function $\ensuremath{\mathfrak{g}}_{\infty}$ need not be considered because, by SANE, $ \ensuremath{\mathcal{F}}_{\infty}$ never appears.} case in consideration, and taking the maximum of the facial residual functions, there exists a nonnegative monotone nondecreasing function $\rho :\ensuremath{\mathbb R}_+ \to \ensuremath{\mathbb R}_+$ such that the function $\psi$ given by \[ \psi(\epsilon,t) \coloneqq \rho(t) \epsilon +\rho(t)\sqrt{\epsilon } \] is a facial residual function for each $ \ensuremath{\mathcal{F}}_{i}^j$. By Proposition~\ref{prop:exp_am}, each face of $\ensuremath{K_{\exp}}$ is $\ensuremath{\mathfrak{g}}_{-\infty}$-amenable so we can invoke Proposition~\ref{prop:frf_prod} to obtain a facial residual function for $ \ensuremath{\mathcal{F}}_i$ and $\ensuremath{\mathbf z}_i$ with respect to $ \ensuremath{\mathcal{K}}$. In what follows, in order to simplify the notation, we define $\ensuremath{\mathfrak{g}} = \ensuremath{\mathfrak{g}}_{-\infty}$, $\hat\amf \coloneqq \sqrt{\ensuremath{\mathfrak{g}}}$. Also, for every $j$, we use $\hat\amf_j$ to denote the composition of $\hat\amf$ with itself $j$-times, i.e., \begin{equation}\label{eq:hatg} \hat\amf_j = \underbrace{\hat\amf\circ \cdots \circ \hat\amf}_{j \text{ times}}. \end{equation} Then, using \eqref{d:entropy_p}, gathering the constants and adjusting the function $\sigma$ in Proposition~\ref{prop:frf_prod} if necessary, we conclude the existence of a nonnegative monotone nondecreasing function $\sigma: \ensuremath{\mathbb R}_+ \to \ensuremath{\mathbb R}_+$ such that the function $\psi _{i}$ given by \begin{align} \psi _{i}(\epsilon,t) \coloneqq \sigma(t)\ensuremath{\mathfrak{g}}(\epsilon)+ \sigma(t)\hat\amf{(\epsilon)} \notag \end{align} is a facial residual function for $ \ensuremath{\mathcal{F}}_i$ and $\ensuremath{\mathbf z}_i$ with respect to $ \ensuremath{\mathcal{K}}$. Therefore, for $\ensuremath{\mathbf x} \in B$, we have \begin{align} \psi _{i}(\epsilon,\norm{\ensuremath{\mathbf x}}) \leq \sigma(M)\ensuremath{\mathfrak{g}}(\epsilon)+ \sigma(M)\hat\amf{(\epsilon)} = \psi _{i}(\epsilon,M), \label{eq:frf_fi} \end{align} where $M = \sup _{\ensuremath{\mathbf x}\in B} \norm{\ensuremath{\mathbf x}}$. Next we are going make a series of arguments related to the following informal principle: over a bounded set only the terms $\hat\amf_j$ with largest $j$ matter. First, we will prove that there exists $\hat \kappa>0$ such that for $\ensuremath{\mathbf x} \in B$ and $j \in \{1, \cdots, \ell-2 \}$ we have \begin{equation}\label{eq:kappa} \ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})) \leq \hat \kappa \hat\amf(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})),\quad \max\{\ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})),\ensuremath{\mathfrak{g}}(\hat\amf_j(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})))\} \leq \hat \kappa \hat\amf_{j+1}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})). \end{equation} Letting $ \hat\kappa_{j} = \sup _{\ensuremath{\mathbf x} \in B} \{ \hat\amf_{j+1}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})) \}$ for $j\in\{0,\cdots,\ell-2\}$, we deduce that \begin{align} \ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})) & = \hat\amf(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}))\hat\amf(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})) \leq \hat \kappa_0 \hat\amf(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}))\notag \\ &\stackrel{\rm (a)}{\le} \hat \kappa_0 \ensuremath{\mathfrak{g}}{(\hat\amf(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})))} = \hat \kappa_0 \hat\amf_2(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}))\hat\amf_2(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})) \notag\\ & \leq \hat \kappa_0 \hat\kappa_1 \hat\amf_2(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})),\notag \end{align} where (a) follows from \eqref{d:entropy_p}. Therefore, letting $\tilde \kappa = \max \{\hat \kappa_0,\hat \kappa_1,\hat\kappa_0\hat\kappa_1 \}$, we have \[ \ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})) \leq \tilde \kappa \hat\amf(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})), \quad \max\{\ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})),\ensuremath{\mathfrak{g}}(\hat\amf(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}))) \} \leq \tilde \kappa\hat\amf_2(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})). \] Similarly, for every $j \geq 1$ we have \[ \hat\amf_j(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})) \leq \ensuremath{\mathfrak{g}}(\hat\amf_j(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}))) = (\hat\amf_{j+1}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})))^2 =\hat\amf_{j+1}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}))\hat\amf_{j+1}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})) \leq \hat\kappa _j\hat\amf_{j+1}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})). \] With that, by induction, we can show the existence of a constant $\hat \kappa$ such that \eqref{eq:kappa} holds for every $j \in \{1,\ldots, \ell-2\}$. Then, taking $\epsilon = \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})$ in \eqref{eq:frf_fi} and invoking \eqref{eq:kappa}, we conclude that there exists $\kappa > 0$ (not depending on $i$) such that \begin{equation}\label{eq:frf_fi_b} \psi _{i}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}),M) \leq \kappa \hat\amf(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})), \qquad \forall \ensuremath{\mathbf x} \in B. \end{equation} Let $\varphi_j \coloneqq \psi _{{j}}\diamondsuit \cdots \diamondsuit \psi_{{1}}$, where $\diamondsuit$ is the diamond composition defined in \eqref{eq:comp}. We will show by induction that for every $j \leq \ell-1$ there exists $\kappa _j$ such that \begin{equation}\label{eq:diamond_bound} \varphi_j(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}),M) \leq \kappa_j\hat\amf_{j}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})), \qquad \forall \ensuremath{\mathbf x} \in B. \end{equation} For $j = 1$, it follows from \eqref{eq:frf_fi_b}. Now, suppose that the claim is valid for some $j$ such that $j+1 \leq \ell-1$. By \eqref{d:entropy_p} and the inductive hypothesis, we have \begin{align} \varphi_{j+1}(d(\ensuremath{\mathbf x}),M) &= \psi _{j+1}(d(\ensuremath{\mathbf x})+ \varphi _{j}(d(\ensuremath{\mathbf x}),M),M) \notag \\ & \leq \psi _{j+1}(\ensuremath{\mathfrak{g}}(d(\ensuremath{\mathbf x}))+ \kappa_j\hat\amf_{j}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})),M) \notag\\ & \leq \psi _{j+1}(\tilde \kappa_j\hat\amf_{j}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})),M), \label{eq:varphi_j} \end{align} where $\tilde \kappa_j \coloneqq 2\max\{\hat \kappa,\kappa_j\}$ and the last inequality follows from \eqref{eq:kappa}. From \eqref{d:entropy_p} and the definition of $\hat\amf_j$, there exists a nonnegative function $\hat \rho$ such that $\ensuremath{\mathfrak{g}}(at) \leq \hat \rho(a) \ensuremath{\mathfrak{g}}(t)$ and $\hat\amf_j(at) \leq \hat \rho(a) \hat\amf_j(t)$ holds for every $t,a \in \ensuremath{\mathbb R}_+$. Then, we plug $\epsilon = \tilde \kappa_j\hat\amf_{j}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}))$ in \eqref{eq:frf_fi} to obtain \begin{align} \psi _{j+1}(\tilde \kappa_j\hat\amf_{j}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})),M) & = \sigma(M)\ensuremath{\mathfrak{g}}(\tilde \kappa_j\hat\amf_{j}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}))) + \sigma(M)\hat\amf(\tilde \kappa_j\hat\amf_{j}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}))) \notag\\ & \le \sigma(M)\hat \rho(\tilde \kappa_j)\ensuremath{\mathfrak{g}}(\hat\amf_{j}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}))) + \sigma(M)\hat \rho(\tilde \kappa_j)\hat\amf_{j+1}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})) \notag \\ & \le \sigma(M)\hat \rho(\tilde \kappa_j)\hat\kappa\hat\amf_{j+1}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})) + \sigma(M)\hat \rho(\tilde \kappa_j)\hat\amf_{j+1}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})) \label{eq:varphi_j2}, \end{align} where the last inequality follows from \eqref{eq:kappa}. Combining \eqref{eq:varphi_j} and \eqref{eq:varphi_j2}, we conclude that there exists $\kappa_{j+1}$ such that \begin{equation*} \varphi_{j+1}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}),M) \leq \kappa_{j+1}\hat\amf_{j+1}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})), \qquad \forall \ensuremath{\mathbf x} \in B. \end{equation*} This concludes the induction proof. In particular, \eqref{eq:diamond_bound} is valid for $j = \ell-1$. We also recall that, from \eqref{d:entropy_p}, we have $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}) \leq \ensuremath{\mathfrak{g}}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}))$. Then, taking into account some positive rescaling (see \eqref{eq:pos_rescale}) and adjusting constants, from \eqref{eq:bound_sane}, \eqref{eq:diamond_bound} and \eqref{eq:kappa} we conclude that there exists $\kappa > 0$ such that \begin{equation}\label{eq:mixed_err} \ensuremath{\operatorname{d}}\left(\ensuremath{\mathbf x}, ( \ensuremath{\mathcal{L}} + \ensuremath{\mathbf a}) \cap \ensuremath{\mathcal{K}}\right) \leq \kappa \hat\amf_{\ell-1}(\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x})), \qquad \forall \ensuremath{\mathbf x} \in B \end{equation} where $\hat\amf = \sqrt{\ensuremath{\mathfrak{g}}_{-\infty}}$, with $\ensuremath{\mathfrak{g}}_{-\infty}$ as in \eqref{d:entropy} and $\hat\amf_{\ell-1}$ as in \eqref{eq:hatg}. We also recall that ${d_{\text{PPS}}( \ensuremath{\mathcal{K}}, \ensuremath{\mathcal{L}}+\ensuremath{\mathbf a})} = \ell-1$. \end{proof} \begin{remark}[Variants of Theorem~\ref{theo:sane}] Theorem~\ref{theo:sane} is not very tight and admits a number of variants that are somewhat cumbersome to describe precisely. For example, the presence of the $\ensuremath{\mathfrak{g}}_{-\infty}$ function makes the proof of Theorem~\ref{theo:sane} notationally challenging, so one might wonder whether we could do without it. In fact, because all the faces of $\ensuremath{K_{\exp}}$ are $\alpha$-amenable for $\alpha \in (0,1)$ (see Proposition~\ref{prop:exp_am}), we could have taken $\ensuremath{\mathfrak{g}} = |\cdot|^\alpha$ when invoking Proposition~\ref{prop:direct_prod} so that $\hat\amf = |\cdot|^{\alpha/2}$ and the final error bound would be uniform H\"olderian with exponent ${(\alpha/2)}^{\ell-1}$. We note, however, that we may not take $\alpha=1$, therefore the resulting error bound would be worse than \eqref{eq:mixed_err}. That said, the 2D face $\ensuremath{{\mathcal F}}_{-\infty}$ is the only one that is not $1$-amenable. Therefore, if we have a face $ \ensuremath{\mathcal{F}}_{i}$ of $ \ensuremath{\mathcal{K}}$ which does not have $\ensuremath{{\mathcal F}}_{-\infty}$ as one of its blocks, we can indeed use $\ensuremath{\mathfrak{g}} = |\cdot|$ instead of $\ensuremath{\mathfrak{g}}_{-\infty}$ when invoking Proposition~\ref{prop:frf_prod}. In particular, if we assume besides SANE that we can take a chain of faces where the 2D face $\ensuremath{{\mathcal F}}_{-\infty}$ never appears as well, then we can take $\ensuremath{\mathfrak{g}} = |\cdot|$ in \eqref{eq:mixed_err}, so that the resulting error bound is uniform H\"olderian with exponent ${(1/2)}^{\ell-1}$. Going for greater generality, we can also drop the SANE assumption altogether and try to be as tight as our analysis permits when dealing with possibly inSANE instances. Although there are several possibilities one must consider, the overall strategy is the same as outlined in the proof of Theorem~\ref{theo:sane}: invoke Theorem~\ref{theo:err}, fix a bounded set $B$, pick a chain of faces as in Proposition~\ref{prop:fra_poly} and the first task is to find a bound for each facial residual function as in \eqref{eq:frf_fi_b}. Intuitively, whenever function compositions appear, only the ``higher'' compositions matter. However, the analysis must consider the possibility of $\ensuremath{\mathfrak{g}}_{-\infty}$ or $\ensuremath{\mathfrak{g}}_{\infty}$ appearing. Then, once suitable bounds for the facial residual functions are identified, one would need to give an upper bound to their diamond composition as in \eqref{eq:diamond_bound}. After this is done, it is just a matter to plug this upper bound into \eqref{eq:bound_sane}. \end{remark} \subsection{Miscellaneous odd behavior and connections to other notions}\label{sec:odd} In this final subsection, we collect several instances of pathological behaviour that can be found inside the facial structure of the exponential cone. \begin{example}[H\"olderian bounds and the non-attainment of admissible exponents]\label{ex:exponents} We recall Definition~\ref{def:hold} and we consider the special case of two closed convex sets $C_1,C_2$ with non-empty intersection. We say that $\gamma \in (0,1]$ is an \emph{admissible exponent} for $C_1, C_2$ if $C_1$ and $C_2$ satisfy a uniform H\"olderian error bound with exponent $\gamma$. It turns out that the supremum of the set of admissible exponents is not itself admissible. In particular if $C_1 = \ensuremath{K_{\exp}}$ and $C_2 = \ensuremath{\mathrm{span}\,} \ensuremath{{\mathcal F}}_{-\infty}$, we have $C_1 \cap C_2 = \ensuremath{{\mathcal F}}_{-\infty}$ and that $C_1$ and $C_2$ satisfy a uniform H\"olderian error bound for all $\gamma \in (0,1)$ but not for $\gamma = 1$ (see Proposition~\ref{prop:exp_am}). In fact, from Theorem~\ref{thm:main_err} and Remark~\ref{rem:opt}, $C_1$ and $C_2$ satisfy an entropic error bound which is tight and is, in a sense, better than any H\"olderian error bound with $\gamma \in (0,1)$ but worse than a Lipschitzian error bound. \end{example} \begin{example}[Non-H\"olderian error bound]\label{ex:non_hold} The facial structure of $\ensuremath{K_{\exp}}$ can be used to derive an example of two sets that provably do not have a H\"olderian error bound. Let $C_1 = \ensuremath{K_{\exp}}$ and $C_2 = \{\ensuremath{\mathbf z}\}^\perp$, where $z_x=z_y = 0$ and $z_z=1$ so that $C_1\cap C_2=\ensuremath{{\mathcal F}}_\infty$. Then, for every $\eta > 0$ and every $\alpha \in (0,1]$, there is no constant $\kappa > 0$ such that \[ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x},\ensuremath{{\mathcal F}}_\infty) \leq \kappa \max\{\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x},\ensuremath{K_{\exp}})^\alpha, \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x},\{\ensuremath{\mathbf z}\}^\perp)^\alpha \}, \qquad \forall \ \ensuremath{\mathbf x} \in B(\eta). \] This is because if there were such a positive $\kappa$, the infimum in Lemma~\ref{lem:non_hold} would be positive, which it is not. This shows that $C_1$ and $C_2$ do not have a H\"olderian error bound. However, as seen in Theorem~\ref{thm:nonzerogammasec72}, $C_1$ and $C_2$ have a log-type error bound. In particular if $\ensuremath{\mathbf q} \in B(\eta)$, using \eqref{proj:p1} and \eqref{proj:p2} together with Theorem~\ref{thm:nonzerogammasec72} we have \begin{align} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}, \ensuremath{{\mathcal F}}_\infty) & \leq \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},\{\ensuremath{\mathbf z}\}^\perp) + \ensuremath{\operatorname{d}}(P_{\{\ensuremath{\mathbf z}\}^\perp}(\ensuremath{\mathbf q}),\ensuremath{{\mathcal F}}_\infty) \notag\\ & \leq \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},\{\ensuremath{\mathbf z}\}^\perp) + \max\{2,2\gamma_{\ensuremath{\mathbf z},\eta}^{-1}\}\ensuremath{\mathfrak{g}}_\infty (\ensuremath{\operatorname{d}}(P_{\{\ensuremath{\mathbf z}\}^\perp}(\ensuremath{\mathbf q}),\ensuremath{K_{\exp}})) \notag\\ &\leq \ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}) + \max\{2,2\gamma_{\ensuremath{\mathbf z},\eta}^{-1}\}\ensuremath{\mathfrak{g}}_\infty (2\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q})) \label{eq:non_hold}, \end{align} where $\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q}) \coloneqq \max\{\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},\ensuremath{K_{\exp}}),\ensuremath{\operatorname{d}}(\ensuremath{\mathbf q},\{\ensuremath{\mathbf z}\}^\perp) \}$ and in the last inequality we used the monotonicity of $\ensuremath{\mathfrak{g}}_\infty$. \end{example} Let $C_1, \cdots, C_m$ be closed convex sets having nonempty intersection and let $C \coloneqq \cap _{i=1}^m C_i$. Following \cite{LL20}, we say that $\varphi : \ensuremath{\mathbb R}_+\times \ensuremath{\mathbb R}_+ \to \ensuremath{\mathbb R}_+ $ is a \emph{consistent error bound function (CEBF)} for $C_1, \ldots, C_m$ if the following inequality holds \begin{equation*} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf x},\, C) \le \varphi\left(\max_{1 \le i \le m}\ensuremath{\operatorname{d}}(\ensuremath{\mathbf x}, C_i), \, \|\ensuremath{\mathbf x}\|\right) \ \ \ \forall\ \ensuremath{\mathbf x}\in\mathcal{E}; \end{equation*} and the following technical conditions are satisfied for every $a,b\in \ensuremath{\mathbb R}_+$: $\varphi(\cdot,b)$ is monotone nondecreasing, right-continuous at $0$ and $\varphi(0,b) = 0$; $\varphi(a,\cdot)$ is mononotone nondecreasing. CEBFs are a framework for expressing error bounds and can be used in the convergence analysis of algorithms for convex feasibility problems, see \cite[Sections~3 and 4]{LL20}. For example, $C_1,\ldots, C_m$ satisfy a H\"olderian error bound (Definition~\ref{def:hold}) if and only if these sets admit a CEBF of the format $\varphi(a,b) \coloneqq \rho(b)\max\{a,a^{\gamma(b)}\}$, where $\rho:\ensuremath{\mathbb R}_+ \to \ensuremath{\mathbb R}_+$ and $\gamma:\ensuremath{\mathbb R}_+ \to (0,1]$ are monotone nondecreasing functions \cite[Theorem~3.4]{LL20}. We remark that in Example~\ref{ex:non_hold}, although the sets $C_1, C_2$ do not satisfy a H\"olderian error bound, the log-type error bound displayed therein is covered under the framework of consistent error bound functions. This is because $\ensuremath{\mathfrak{g}}_\infty$ is a continuous monotone nondecreasing function and $\gamma_{\ensuremath{\mathbf z},\eta}^{-1}$ is monotone nondecreasing as a function of $\eta$ (Remark~\ref{rem:kappa}). Therefore, in view of \eqref{eq:non_hold}, the function given by $\varphi(a,b) \coloneqq a + \max\{2,2\gamma_{\ensuremath{\mathbf z},b}^{-1}\}\ensuremath{\mathfrak{g}}_\infty (2a)$ is a CEBF for $C_1$ and $C_2$. By the way, it seems conceivable that many of our results in Section~\ref{sec:frf_comp} can be adapted to derive CEBFs for arbitrary convex sets. Specifically, Lemma~\ref{lem:facialresidualsbeta}, Theorem~\ref{thm:1dfacesmain}, and Lemma~\ref{lem:infratio} only rely on convexity rather than on the more specific structure of cones. Next, we will see that we can also adapt Examples~\ref{ex:exponents} and \ref{ex:non_hold} to find instances of odd behavior of the so-called \emph{Kurdyka-{\L}ojasiewicz (KL) property} \cite{BDL07,BDLS07,ABS13,ABRS10,BNPS17,LP18}. First, we recall some notations and definitions. Let $f: \ensuremath{\mathbb R}^n\to \ensuremath{\mathbb R} \cup \{+\infty\}$ be a proper closed convex extended-real-valued function. We denote by $\ensuremath{\operatorname{dom}} \partial f$ the set of points for which the subdifferential $\partial f(\ensuremath{\mathbf x})$ is non-empty and by $[a < f < b]$ the set of $\ensuremath{\mathbf x}$ such that $a < f(\ensuremath{\mathbf x}) < b$. As in \cite[Section~2.3]{BNPS17}, we define for $r_0\in (0,\infty)$ the set \begin{equation*} {\cal K}(0,r_0) := \{\phi\in C[0,r_0)\cap C^1(0,r_0)\;|\; \phi \mbox{ is concave}, \ \phi(0) = 0, \ \phi'(r) > 0\ \forall r\in (0,r_0)\}. \end{equation*} Let $B(\ensuremath{\mathbf x},\epsilon)$ denote the closed ball of radius $\epsilon > 0$ centered at $\ensuremath{\mathbf x}$. With that, we say that $f$ satisfies the KL property at $\ensuremath{\mathbf x} \in \ensuremath{\operatorname{dom}} \partial f$ if there exist $r_0 \in (0,\infty)$, $\epsilon > 0$ and $\phi \in \mathcal{K}(0,r_0)$ such that for all $\ensuremath{\mathbf y} \in B(\ensuremath{\mathbf x},\epsilon) \cap [f(\ensuremath{\mathbf x}) < f < f(\ensuremath{\mathbf x}) + r_0 ]$ we have \[ \phi'(f(\ensuremath{\mathbf y})-f(\ensuremath{\mathbf x}))\ensuremath{\operatorname{d}}(0,\partial f(\ensuremath{\mathbf y})) \geq 1. \] In particular, as in \cite{LP18}, we say that $f$ satisfies the \emph{KL property with exponent $\alpha\in [0,1)$ at $\ensuremath{\mathbf x} \in \ensuremath{\operatorname{dom}} \partial f$}, if $\phi$ can be taken to be $\phi(t) = ct^{1-\alpha}$ for some positive constant $c$. Next, we need the following result, which is an immediate consequence of \cite[Theorem~5]{BNPS17}. \begin{proposition}\label{prop:error_kl} Let $C_1, C_2 \subseteq \ensuremath{\mathbb R}^n$ be closed convex sets with $C_1 \cap C_2 \neq \emptyset$. Define $f: \ensuremath{\mathbb R}^n \to \ensuremath{\mathbb R}$ as \[ f(\ensuremath{\mathbf y}) = \ensuremath{\operatorname{d}}(\ensuremath{\mathbf y},C_1)^2 + \ensuremath{\operatorname{d}}(\ensuremath{\mathbf y},C_2)^2. \] Let $\ensuremath{\mathbf x} \in C_1\cap C_2$, $\gamma \in (0,1]$. Then, there exist $\kappa > 0$ and $\epsilon > 0 $ such that \begin{equation}\label{eq:error} \ensuremath{\operatorname{d}}(\ensuremath{\mathbf y}, C_1\cap C_2) \leq \kappa \max\{\ensuremath{\operatorname{d}}(\ensuremath{\mathbf y},C_1),\ensuremath{\operatorname{d}}(\ensuremath{\mathbf y},C_2)\}^\gamma,\qquad \forall \ensuremath{\mathbf y} \in B(\ensuremath{\mathbf x},\epsilon) \end{equation} if and only if $f$ satisfies the KL property with exponent $1-\gamma/2$ at $\ensuremath{\mathbf x}$. \end{proposition} \begin{proof} Note that $\inf f = 0$ and $\ensuremath{\operatorname*{argmin}} f = C_1\cap C_2$. Furthermore, \eqref{eq:error} is equivalent to the existence of $\kappa' > 0$ and $\epsilon > 0$ such that \[ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf y}, \ensuremath{\operatorname*{argmin}} f) \leq \varphi(f(\ensuremath{\mathbf y})),\qquad \forall \ensuremath{\mathbf y} \in B(\ensuremath{\mathbf x},\epsilon), \] where $\varphi$ is the function given by $\varphi(r) = \kappa' r^{\gamma/2}$. With that, the result follows from \cite[Theorem~5]{BNPS17}. \end{proof} \begin{example}[Examples in the KL world] In Example~\ref{ex:exponents}, we have two sets $C_1, C_2$ satisfying a uniform H\"olderian error bound for $\gamma \in (0,1)$ but not for $\gamma = 1$. Because $C_1$ and $C_2$ are cones and the corresponding distance functions are positively homogeneous, this implies that for $\mathbf{0} \in C_1 \cap C_2$, a Lipschitzian error bound never holds at any neighbourhood of $\mathbf{0}$. That is, given $\eta > 0$, there is no $\kappa > 0$ such that \[ \ensuremath{\operatorname{d}}(\ensuremath{\mathbf y}, C_1\cap C_2) \leq \kappa \max\{\ensuremath{\operatorname{d}}(\ensuremath{\mathbf y},C_1),\ensuremath{\operatorname{d}}(\ensuremath{\mathbf y},C_2)\},\qquad \forall \ensuremath{\mathbf y} \in B(\eta) \] holds. We conclude that the corresponding function $f$ in Proposition~\ref{prop:error_kl} satisfies the KL property with exponent $\alpha$ for any $\alpha \in (1/2,1)$ at the origin, but not for $\alpha = 1/2$. Similarly, from Example~\ref{ex:non_hold} we obtain $C_1,C_2$ for which \eqref{eq:error} does not hold for $\mathbf{0} \in C_1\cap C_2$ with any chosen $\kappa,\varepsilon>0,\;\gamma \in \left(0,1 \right]$. Thus from Proposition~\ref{prop:error_kl} we obtain a function $f$ that does not satisfy the KL property with exponent $\beta \in [1/2,1)$ at the origin. Since a function satisfying the KL property with exponent $\alpha\in [0,1)$ at an $\ensuremath{\mathbf x}\in \ensuremath{\operatorname{dom}} \partial f$ necessarily satisfies it with exponent $\beta$ for any $\beta \in [\alpha,1)$ at $\ensuremath{\mathbf x}$, we see further that this $f$ does not satisfy the KL property with any exponent at the origin. \end{example} \section{Concluding remarks}\label{sec:conclusion} In this work, we presented an extension of the results of \cite{L17} and showed how to obtain error bounds for conic linear systems using facial residual functions and facial reduction (Theorem~\ref{theo:err}) even when the underlying cone is not amenable. Nevertheless, if the cone is indeed amenable or, more generally, $\ensuremath{\mathfrak{g}}$-amenable, then one can use extra tools to obtain the facial residual functions, e.g., Propositions~\ref{prop:frf_prod} and ~\ref{prop:frf_am}. Related to facial residual functions, we also developed several techniques that aid in their computation; see Section~\ref{sec:frf_comp}. Finally, all techniques and results developed in Section~\ref{sec:frf} were used in some shape or form in order to obtain error bounds for the exponential cone in Section~\ref{sec:exp_cone}. Our new framework unlocks analysis for cones not reachable with the techniques developed in \cite{L17}; these include cones that are not facially exposed, as well as cones for which the projection operator has no simple closed form or is only implicitly specified. These were, until now, significant barriers against error bound analysis for many cones of interest. As future work, we are planning to use the techniques developed in this paper to analyze and obtain error bounds for some of these other cones that have been previously unapproachable. Potential examples include the cone of $n\times n$ completely positive matrices and its dual, the cone of $n\times n$ copositive matrices. The former is not facially exposed when $n\geq 5$ (see \cite{Zh18}) and the latter is not facially exposed when $n \geq 2$. It would be interesting to clarify where these cones stand from the point of view of $\ensuremath{\mathfrak{g}}$-amenability. Or, more ambitiously, we could try to obtain some of the facial residual functions and some error bound results. Of course, a significant challenge is that their facial structure is not completely understood, but we believe that even partial results for general $n$ or complete results for specific values of $n$ would be relevant and, possibly, quite non-trivial. \bibliographystyle{abbrvurl}

https://arxiv.org/abs/2101.05083

Functional calculi for sectorial operators and related function theory

We construct two bounded functional calculi for sectorial operators on Banach spaces, which enhance the functional calculus for analytic Besov functions, by extending the class of functions, generalizing and sharpening estimates, and adapting the calculus to the angle of sectoriality. The calculi are based on appropriate reproducing formulas, they are compatible with standard functional calculi and they admit appropriate convergence lemmas and spectral mapping theorems. To achieve this, we develop the theory of associated function spaces in ways which are interesting and significant. As consequences of our calculi, we derive several well-known operator norm-estimates and provide generalizations of some of them.

\section{Introduction} The theory of functional calculi forms a basis for the study of sectorial operators and semigroup generators. In particular, there were two functional calculi used extensively in the research on operator semigroups and sectorial operators through the past fifty years. One of them, the Hille-Phillips (HP) functional calculus for semigroup generators, probably stemmed from the foundational monograph \cite{HP}, and it became an indispensable part of semigroup theory. The systematic approach to the other one, the holomorphic functional calculus for sectoral operators, was initiated by McIntosh and his collaborators in the 1980s. While the two calculi appeared to be very useful in applications, the operator norm-estimates within them are often problematic. The estimates within the HP-calculus are direct but rather crude, and the task of getting bounds within the holomorphic functional calculus is a priori cumbersome since the calculus is not, in general, a bounded Banach algebra homomorphism. To circumvent those problems, a number of additional tools and methods appeared in the literature. In particular, the advanced notions and techniques related to bounded $H^\infty$-calculus, $R$-boundedness, Fourier multipliers and transference were developed in depth, and one may consult \cite{Hytonen}, \cite{KW} and \cite{HaaseB} for many of these function-theoretical developments. Moreover, various implications of positivity of functions and their derivatives (completely monotone, Bernstein, $\mathcal{NP}_+$-functions) were adjusted to the operator-theoretical set-up. For clarification of the role of positivity, see \cite{Schill}, \cite{GT15} and \cite{BGTad}, for example. Recently, in \cite{BGT}, a new functional calculus was constructed, the so-called $\mathcal B$-calculus. First of all, the $\mathcal B$-calculus offers a simple and efficient route to operator norm-estimates for functions of semigroup generators, thus unifying a number of estimates in the literature and leading to new ones. No supplementary arguments are required and the estimates underline the strength of the calculi. Moreover, the $\mathcal B$-calculus possesses all attributes of classical functional calculi, see \cite{BGT}. When combined properly they lead to new spectral mapping theorems and generalizations of fundamentals of semigroup theory, see \cite{BGT2}. To put our results into a proper context and to use some of the $\mathcal B$-calculus properties in the sequel, we briefly recall the set-up for the $\mathcal B$-calculus, see \cite{BGT} for more details. Let $\mathcal B$ be the algebra of holomorphic functions on the right half-plane $\mathbb{C}_+$ such that \begin{equation} \label{bdef0} \|f\|_{\Bes_0}:=\int_{0}^{\infty} \sup_{\beta \in \mathbb{R} }|f'(\alpha+i\beta)| \, d\alpha <\infty. \end{equation} These functions are sometimes called analytic Besov functions, and they have been considered in some detail in \cite{BGT}. . In particular, every $f \in \mathcal B$ belongs to $H^\infty(\mathbb C_+)\cap C(\overline{\mathbb C}_+)$, and $\mathcal B$ is a Banach algebra with the norm \begin{equation}\label{b-norm} \|f\|_{\mathcal{B}}:=\|f\|_\infty+ \|f\|_{\Bes_0}. \end{equation} In the setting of power bounded operators on Hilbert spaces, the unit disc counterpart of $\mathcal B$ was employed for the study of functional calculi in \cite{Pel}. Let $A$ be a densely defined closed operator on a Banach space $X$ such that $\sigma(A) \subseteq \overline{\mathbb{C}}_+$, and \begin{equation} \label{8.1} \sup_{\alpha >0} \alpha \int_{\mathbb R} |\langle (\alpha +i\beta + A)^{-2}x, x^* \rangle| \, d\beta <\infty \end{equation} for all $x \in X$ and $x^* \in X^*$. This class of operators includes two substantial subclasses, namely the (negative) generators of bounded $C_0$-semigroups on Hilbert spaces $X$ and the generators of (sectorially) bounded holomorphic $C_0$-semigroups on Banach spaces $X$. On the other hand, every operator in the class is the negative generator of a bounded $C_0$-semigroup. Some results involving functional calculus of operators in $\mathcal{B}$ were obtained in \cite{Haase}, \cite{Sch1}, \cite{V1} and \cite{White}, but each of them applied only to one of the two subclasses of operators, and only \cite{Sch1} and \cite{V1} considered all functions in $\mathcal{B}$ applied to generators of bounded holomorphic semigroups. In \cite{BGT} we introduced a bounded functional calculus for all operators satisfying \eqref{8.1}, and we extended the theory in \cite{BGT2}. For $f \in \mathcal{B}$, set \begin{align} \label{fcdef} \lefteqn{\langle f(A)x, x^* \rangle} \quad\\ &= f(\infty) \langle x, x^* \rangle - \frac{2}{\pi} \int_0^\infty \alpha \int_{\mathbb{R}} {f'(\alpha +i\beta)} \langle (\alpha -i\beta +A)^{-2}x, x^* \rangle \, d\beta\,d\alpha \nonumber \end{align} for all $x \in X$ and $x^* \in X^*$, where $f(\infty) = \lim_{\Re z\to\infty} f(z)$. Using \eqref{8.1} and the definition of $\mathcal B$ (and the Closed Graph Theorem), it is easy to show that $f(A)$ is a bounded linear mapping from $X$ to $X^{**}$, and that the linear mapping \[ \widetilde\Phi_A : \mathcal{B} \to \mathcal L(X,X^{**}), \qquad f \mapsto f(A), \] is bounded. It was discovered in \cite{BGT} that much more is true. If $A$ belongs to any of the classes of semigroup generators mentioned above, then the formula \eqref{fcdef} defines a bounded algebra homomorphism \[ \Phi_A: \mathcal{B} \to L (X),\qquad \Phi_A (f):=f(A). \] It is natural to call the homomorphism $\Phi_A$ the ($\mathcal{B}$-)calculus of $A$. It was proved in \cite{BGT} that $\Phi_A$ possesses a number of useful properties. In particular, it admits the spectral inclusion (spectral mapping, in the case of bounded holomorphic semigroups) theorem and a Convergence Lemma of appropriate form. The utility of the $\mathcal B$-calculus depends on the facts that it (strictly) extends the Hille-Phillips (HP-) calculus and it is compatible with the holomorphic functional calculi for sectorial and half-plane type operators. Moreover, the $\mathcal{B}$-calculus $\Phi_A$ is the only functional calculus that one can define for $A$ satisfying \eqref{8.1} and for functions in $\mathcal{B}$. Indeed, let $A$ be an operator on $X$ with dense domain, and assume that $\sigma(A) \subseteq \overline{\mathbb{C}}_+$. A \emph{(bounded) $\mathcal{B}$-calculus} for $A$ is, by definition, a bounded algebra homomorphism $\Phi : \mathcal{B} \to L(X)$ such that $\Phi((z+\cdot)^{-1}) = (z+A)^{-1}$ for all $z \in \mathbb{C}_+$. As shown in \cite{BGT2}, if $A$ admits a $\mathcal B$-calculus, then the resolvent assumption \eqref{8.1} holds, and the calculus is $\Phi_A$. While the $\mathcal B$-calculus is optimal and unique for generators of Hilbert space semigroups, the situation is far from being so for generators of bounded holomorphic semigroups on Banach spaces (as this paper will, in particular, show). Thus using the $\mathcal B$-calculus ideology as a guiding principle, it is natural to try to extend it beyond the Besov algebra $\mathcal B$ keeping all of its useful properties such as availability of good norm-estimates, spectral mapping theorems, Convergence Lemmas, compatibility with the other calculi, etc. Moreover, it is desirable to cover all sectorial operators regardless their sectoriality angle. In this paper we will construct some functional calculi encompassing wider classes of functions (including some with singularities on $i\mathbb R$) and providing finer estimates for all negative generators of (sectorially) bounded holomorphic semigroups, and eventually for all sectorial operators. Functional calculi for generators of some classes of bounded holomorphic semigroups were constructed in \cite{GaPy97}, \cite{GaMiPy02}, \cite{GaMi06}, \cite{GaMi08}, and \cite{KrieglerWeis}. However, most of the results in those papers concern sectorial operators of angle zero, and the approaches there are based on fine estimates for the corresponding semigroups rather than fine analytic properties of resolvents. Let $A$ be a densely defined sectorial operator of sectorial angle $\theta_A \in [0,\pi)$ on a Banach space $X$. It is well known that $-A$ is the generator of a (sectorially) bounded holomorphic $C_0$-semigroup on $X$ if and only if and only if $A$ is sectorial and $\theta_A < \pi/2$ (we may write $A \in \Sect(\pi/2)$ for this class). In this paper we address the question whether the $\mathcal{B}$-calculus for $A$ can be extended to more functions. Since the resolvent of $A$ satisfies the estimate \begin{equation*} \label{Aa0_intro} M_{\psi}(A) := \sup_{z\in\Sigma_{\pi-\psi}} \|z(z+A)^{-1}\| < \infty, \qquad z \in \Sigma_{\pi-\psi}. \end{equation*} for all $\psi\in(\theta_A,\pi)$, a direct way to define an appropriate function algebra would be to introduce a Banach space of functions $f$ which are holomorphic on sectors $\Sigma_{\psi}:=\{z\in \mathbb C: |\arg(z)| < \psi \}$ such that \begin{equation}\label{assump_f} \|f\|_{\psi}:=\int_{\partial \Sigma_{\psi}} \frac{|f(z)|}{|z|}|dz| <\infty. \end{equation} In order to apply this to all $A \in \Sect(\pi/2-)$, $f$ should be holomorphic on $\mathbb{C}_+$ and the assumption \eqref{assump_f} should hold for all $\psi \in (0,\pi/2)$, and in order to provide an estimate which is uniform in $\theta$ it is desirable to have $\sup_{\psi\in (0,\pi/2)} \|f\|_{\psi}<\infty$. To our knowledge, no spaces of this type have been studied systematically in the literature, although they appear naturally in \cite[Appendix H2 and Chapter 10.2]{Hytonen}, \cite[Section 6]{Haase_Sp} or \cite[Appendix C]{Haak_H}. This class of functions is strictly included in each of the spaces $\mathcal D_s, s>0$ (see Proposition \ref{weight1} and a discussion following it), which we now define. To define a functional calculus for all $A \in \Sect(\pi/2-)$, we let $\mathcal{D}_s, s >-1$, be the linear space of all holomorphic functions $f$ on $\mathbb{C}_{+}$ such that \begin{equation}\label{vintro} \|f\|_{\mathcal D_{s,0}}:= \int_0^\infty \alpha^s\int_{-\infty}^\infty\frac{|f'(\alpha +i\beta)|}{(\alpha^2+\beta^2)^{(s+1)/2}}\,d\beta\,d\alpha <\infty. \end{equation} If $f \in \mathcal D_s$, then there exists a finite limit $f(\infty):=\lim_{|z|\to\infty,\,z\in {\Sigma}_\psi}\,f(z)$ for all $\psi\in(0.\pi/2)$. For every $s>-1$ the space $\mathcal D_s$ equipped with the norm \[ \|f\|_{\mathcal D_s}:= |f(\infty)|+\|f\|_{\mathcal D_{s,0}}, \qquad f \in\mathcal D_s, \] is a Banach space, but not an algebra. However, the spaces $\mathcal{D}_s$ increase with $s$, and we prove in Lemma \ref{Alg1} that \[ \mathcal D_\infty:=\bigcup_{s>-1} \mathcal D_s \] is an algebra. Let $f\in \mathcal{D}_\infty$, so $f \in \mathcal{D}_s$ for some $s>-1$, and let $A$ be sectorial with $\theta_A < \pi/2$. Define \begin{equation}\label{formulaD_intro} f_{\mathcal{D}_s}(A):=f(\infty)- \frac{2^s}{\pi}\int_0^\infty \alpha^s\int_{-\infty}^\infty f'(\alpha+i\beta)(A+\alpha-i\beta)^{-(s+1)}\,d\beta\,d\alpha. \end{equation} Then $f_{\mathcal{D}_\sigma}(A) = f_{\mathcal{D}_s}(A)$ whenever $\sigma>s$. The following result sets out other properties of this functional calculus. The proof will be given in Section \ref{defD}. \begin{thm} \label{dcalculus_intro} Let A be a densely defined, closed operator on a Banach space $X$ such that $\sigma(A) \subset \overline{\mathbb{C}}_+$. The following are equivalent: \begin{enumerate}[\rm(i)] \item $A \in {\rm Sect}(\pi/2-)$, \item There is an algebra homomorphism $\Psi_A : \mathcal{D}_\infty \to L(X)$ such that \begin{equation*} \Psi_A((z+\cdot)^{-1}) = (z+A)^{-1}, \qquad z \in \mathbb{C}_+, \end{equation*} and $\Psi_A$ is bounded in the sense that there exist constants $C_s, \, s>-1$, such that, for every $f \in \mathcal D_s$, \begin{equation}\label{bounded_intro} \|\Psi_A(f)\|\le C_s \|f\|_{\mathcal D_s}. \end{equation} \end{enumerate} When these properties hold, $\Psi_A$ is unique, and it is defined by the formula \eqref{formulaD_intro}: \begin{eqnarray*} \Psi_A : \mathcal D_{\infty} \mapsto L(X), \qquad \Psi_A(f)=f_{\mathcal{D}_s}(A), \quad f \in \mathcal{D}_s. \end{eqnarray*} \end{thm} The homomorphism $\Psi_A$ will be called the {\it $\mathcal D$-calculus} for $A$. It will be shown in Section \ref{defD} that the $\mathcal{D}$-calculus is compatible with the Hille-Phillips calculus and the holomorphic calculus for sectorial operators, and a spectral mapping theorem is given in Theorem \ref{SMT}. Corollary \ref{Compat} provides a version of this functional calculus based on the Banach algebra $H^\infty(\mathbb C_+)\cap \mathcal D_s$ for a fixed value of $s$. The $\mathcal D$-calculus defined as above does not take into account the sectoriality angle of $A \in {\rm Sect}(\pi/2-)$. However, it can be used to construct a functional calculus which does not have this drawback. To achieve this aim we introduce the Hardy-Sobolev spaces $\mathcal{H}_\psi$, on sectors $\Sigma_\psi$. First, for any $\psi\in (0,\pi)$, we define the Hardy space $H^1(\Sigma_\psi)$ as the linear space of functions $f\in \operatorname{Hol}(\Sigma_\psi)$ such that \begin{equation}\label{CC2_intro} \|f\|_{H^1(\Sigma_{\psi})}:=\sup_{|\varphi|< \psi}\, \int_0^\infty \bigl(|f(te^{i\varphi})|+ |f(te^{-i\varphi})| \bigr) \,dt <\infty. \end{equation} Note that $H^1(\Sigma_{\pi/2})$ coincides with the classical Hardy space $H^1(\mathbb C_+)$ in the right half-plane $\mathbb C_+$. It is well-known that $(H^1(\Sigma_\psi), \|\cdot\|_{H^1(\Sigma_\psi)})$ is a Banach space, and every $f \in H^1(\Sigma_\psi)$ has a boundary function on $\partial \Sigma_\psi$. The boundary function exists as the limit of $f$ in both an $L^1$-sense and a pointwise (a.e.) sense. Moreover, the norm of $f$ in $H^1(\Sigma_\psi)$ is attained by the $L^1$-norm of its boundary function. See Section \ref{hardy_rays} for a succinct approach to the Hardy spaces on sectors. The space $H^1(\Sigma_\psi)$ induces the corresponding Hardy-Sobolev space $\mathcal{H}_\psi$ on $\Sigma_\psi$ as \[ \mathcal{H}_\psi:=\left\{f \in \operatorname{Hol}(\Sigma_\psi): f' \in H^1(\Sigma_\psi)\right\}. \] Any function $f \in \mathcal{H}_\psi$ has a finite limit $f(\infty):=\lim_{t \to \infty}f(t)$, and moreover $f \in H^\infty(\Sigma_\psi)$. Then $\mathcal{H}_\psi$ becomes a Banach algebra in the norm \[ \|f\|_{\mathcal \mathcal{H}_\psi}:=\|f\|_{H^\infty(\Sigma_\psi)} +\|f'\|_{H^1(\Sigma_\psi)}, \quad f \in \mathcal{H}_\psi. \] The relationship between these spaces and the spaces $\mathcal D_s$ for all $s>-1$ is set out in Corollary \ref{HHH} and Lemma \ref{Dh1}; in particular, for each $s>-1$, $\mathcal{H}_{\pi/2}$ is contained in $\mathcal{D}_s$, and $\mathcal{D}_s$ is embedded in $\mathcal{H}_\psi$ for $\psi<\pi/2$. To make use of the angle of sectoriality of $A$, we can adjust the $\mathcal D$-calculus to sectors as follows. If $f\in \mathcal{H}_\psi$ where $\psi \in (\theta_A,\pi)$, $\gamma = \pi/(2\psi)$, and $f_{1/\gamma}(z):=f(z^{1/\gamma})$, then $f'_{1/\gamma}\in H^1(\mathbb{C}_{+})$ and $f_{1/\gamma}(\infty)=f(\infty)$, and hence $f_{1/\gamma}\in \mathcal{D}_0$. This observation allows us to extend the $\mathcal D$-calculus to the class of all sectorial operators, and makes the next definition (based on the $\mathcal{D}$-calculus) natural and plausible. If $A$ is sectorial and $\psi \in (\theta_A,\pi)$, define \begin{equation}\label{sigma_def_intro} f_{\mathcal H}(A):= f(\infty)-\frac{1}{\pi}\int_0^\infty\int_{-\infty}^\infty f'_{1/\gamma}(\alpha+i\beta) (A^{\gamma}+\alpha-i\beta)^{-1}\, d\beta\,d\alpha. \end{equation} One can prove (see \eqref{A2Dop}) that \begin{equation}\label{A2Dop_intro} \|f_{\mathcal H}(A)\|\le |f(\infty)|+\frac{M_{\pi/2}{(A^\gamma)}}{\pi}\|f_{1/\gamma}\|_{\mathcal{D}_0} \le |f(\infty)|+ M_{\pi/2}({A^\gamma}) \|f\|_{\mathcal{H}_\psi}. \end{equation} Then \eqref{sigma_def_intro} and \eqref{A2Dop_intro} hold for any $\gamma \in (1,\pi/(2\theta))$, and the definition of $f_\mathcal{H}(A)$ does not depend on the choice of $\psi$. Now we are able to formalize our extension of the $\mathcal D$-calculus as follows. \begin{thm}\label{sigma_cal_intro} Let $A$ be a densely defined operator on a Banach space $X$ such that $\sigma(A) \subset \overline{\Sigma_\theta}$, where $\theta \in [0,\pi)$. The following are equivalent: \begin{enumerate}[\rm(i)] \item $A \in \Sect(\theta)$; \item For each $\psi \in (\theta,\pi)$, there is a bounded Banach algebra homomorphism $\Upsilon_A : \mathcal \mathcal{H}_\psi \mapsto L(X)$ such that $\Upsilon_A((z+\cdot)^{-1}) = (z+A)^{-1}, \quad z \in \Sigma_{\pi-\psi}$. \end{enumerate} When these properties hold, the homomorphism $\Upsilon_A$ is unique for each value of $\psi$, and it is defined by the formula \eqref{sigma_def_intro}: \begin{equation*} \Upsilon_A: \mathcal{H}_\psi \to L(X), \qquad \Upsilon_A(f) = f_\mathcal{H}(A), \quad f \in \mathcal{H}_\psi. \end{equation*} \end{thm} The homomorphism $\Upsilon_A$ will be called the {\it $\mathcal H$-calculus} for $A$. The $\mathcal{D}$-calculus can be given a more succinct form, by replacing \eqref{sigma_def_intro} with the somewhat more transparent formula \eqref{formula_a_Intro} below. \begin{thm}\label{Sovp_Intro} Let $A \in \Sect(\theta)$, $\theta < \psi < \pi$, and $\gamma=\pi/(2\psi)$. For $f\in \mathcal{H}_\psi$, let \[ f_\psi(s):=\frac{f(e^{i\psi}t)+f(e^{-i\psi}t)}{2}, \quad t \ge 0. \] Then \begin{equation}\label{formula_a_Intro} f_{\mathcal H}(A)= f(\infty)-\frac{2}{\pi}\int_0^\infty f_\psi'(t){\arccot}(A^\gamma/t^{\gamma})\,dt \end{equation} where the integral converges in the uniform operator topology, and \begin{equation \|f_{\mathcal H}(A)\|\le |f(\infty)|+ M_\psi(A)\|f_\psi'\|_{L^1(\mathbb{R}_+)} \le M_\psi(A) \|f\|_{\mathcal{H}_\psi}. \end{equation} Thus $\|\Upsilon_A\| \le M_\psi(A)$. \end{thm} See Section \ref{alternative} for details. The $\mathcal D$-calculus and the $\mathcal H$-calculus possess natural properties of functional calculi such as spectral mapping theorems and Convergence Lemmas. These properties are studied in Section \ref{SMT_sec}. The strength of the constructed calculi is illustrated by several examples showing that they lead to sharper estimates than those offered by other calculi (see Section \ref{VS} for one example). Moreover, the theory developed in this paper is successfully tested by deriving several significant estimates for functions of sectorial operators from the literature. In particular, in Section \ref{norm_estimates}, we provide a proof of permanence of the class of sectorial operators under subordination and we revisit a few basic results from semigroup theory. In developing the $\mathcal D$- and $\mathcal H$-calculi we prove a number of results of independent interest in function theory. Apart from the theory of the spaces $\mathcal D_s$ and $\mathcal{H}_\psi$, their reproducing formulas, and boundedness of the associated operators elaborated in this paper, we emphasize the property \eqref{n_equality} in Corollary \ref{HHH} yielding isometric coincidence of spaces of Hardy type, Theorem \ref{ACM} on Laplace representability of Hardy-Sobolev functions, and Theorem \ref{DM122} on the density of rational functions in Hardy-Sobolev spaces. \section{Preliminaries} \label{prelims} \subsection*{Notation} Throughout the paper, we shall use the following notation: \begin{enumerate}[\null\hskip1pt] \item $\mathbb{R}_+ :=[0,\infty)$, \item $\mathbb{C}_+ := \{z \in\mathbb{C}: \Re z>0\}$, $\overline{\mathbb{C}}_+ = \{z \in\mathbb{C}: \Re z\ge0\}$, \item $\Sigma_\theta := \{z\in\mathbb{C}: z \ne 0, |\arg z|<\theta\}$ for $\theta \in (0,\pi)$. \item \end{enumerate} For $f : \mathbb{C}_+ \to \mathbb{C}$, we say that $f$ has a \emph{sectorial limit at infinity} if \[ \lim_{|z|\to\infty, z \in \Sigma_\psi} f(z) \] exists for every $\psi \in (0,\pi/2)$. Similarly, $f$ has a \emph{sectorial limit at $0$} if \[ \lim_{|z|\to0, z \in \Sigma_\psi} f(z) \] exists for every $\psi \in (0,\pi/2)$. We say that $f$ has a \emph{half-plane limit at infinity} if \[ \lim_{\Re z \to \infty} f(z) \] exists in $\mathbb{C}$. We say that $f$ has a \emph{full limit at infinity} or \emph{at zero} if \[ \lim_{|z| \to \infty, z\in\mathbb{C}_+} f(z) \quad \text{or} \quad \lim_{|z| \to 0, z\in\mathbb{C}_+} f(z) \] exists in $\mathbb{C}$. The notation $f(\infty)$ and $f(0)$ may denote a sectorial limit, a half-plane limit, or a full limit, according to context. \noindent For $a \in \overline{\mathbb{C}}_+$, we define functions on $\mathbb{C}$ by \[ e_a(z) = e^{-az}; \qquad r_a(z) = (z+a)^{-1}, \, z\ne-a. \] \noindent We use the following notation for spaces of functions or measures, and transforms, on $\mathbb{R}$ or $\mathbb{R}_+$: \begin{enumerate}[\null\hskip1pt] \item $\operatorname{Hol}(\Omega)$ denotes the space of holomorphic functions on an open subset $\Omega$ of $\mathbb{C}$, $H^\infty(\Omega)$ is the space of bounded homorphic functions on $\Omega$, and $\|f\|_{H^\infty(\Omega)} = \sup_{\Omega} |f(z)| $. \item $H^p(\mathbb{C}_+), \, 1 \le p \le \infty$, are the standard Hardy spaces on the (right) half-plane. \item $M(\mathbb{R}_+)$ denotes the Banach algebra of all bounded Borel measures on $\mathbb{R}_+$ under convolution. We identify $L^1(\mathbb{R}_+)$ with a subalgebra of $M(\mathbb{R}_+)$ in the usual way. We write $\mathcal L\mu$ for the Laplace transform of $\mu \in M(\mathbb{R}_+)$. \item $\mathcal{LM}$ is the Hille-Phillips algebra, $\mathcal{LM} := \{\mathcal{L}\mu : \mu \in M(\mathbb{R}_+)\}$, with the norm $\|\mathcal{L}\mu\|_{\mathrm{HP}}:= |\mu|(\mathbb{R}_+)$, and $\mathcal{L}L^1 := \{\mathcal{L} f: f \in L^1(\mathbb{R}_+)\}$. \item $dS$ denotes area measure on $\mathbb{C}_+$. \end{enumerate} \noindent For a Banach space $X$, $X^*$ denotes the dual space of $X$ and $L(X)$ denotes the space of all bounded linear operators on $X$. The domain, spectrum and resolvent set of an (unbounded) operator $A$ on $X$ are denoted by $D(A)$, $\sigma(A)$ and $\rho(A)$, respectively. If $(\mathcal{X},\|\cdot\|_{\mathcal{X}})$ and $(\mathcal{Y},\|\cdot\|_{\mathcal{Y}})$ are normed spaces of holomorphic functions on domains $\Omega_{\mathcal{X}}$ and $\Omega_{\mathcal{Y}}$, we will use notation as follows: \begin{enumerate}[\null\hskip1pt] \item $\mathcal{Y} \overset{i}{\hookrightarrow} \mathcal{X}$ if $\Omega_{\mathcal{Y}} = \Omega_{\mathcal{X}}$, $\mathcal{Y}$ is a subset of $\mathcal{X}$ and the inclusion map is continuous; \item $\mathcal{Y} \subset \mathcal{X}$ if $\Omega_{\mathcal{Y}} = \Omega_{\mathcal{X}}$, $\mathcal{Y}$ is a subset of $\mathcal{X}$ and $\mathcal{Y}$ inherits the norm from $\mathcal{X}$; \item $\mathcal{Y} \overset{r}{\hookrightarrow} \mathcal{X}$ if $\Omega_{\mathcal{Y}} \subset \Omega_{\mathcal{X}}$, and the restriction map $f \mapsto f|_{\Omega_{\mathcal{Y}}}$ is a continuous map from $\mathcal{Y} \to \mathcal{X}$. \end{enumerate} \subsection*{Elementary inequalities} We will need the following lemma which gives elementary inequalities for $|z+\lambda|$, when $z$, $\lambda \in \mathbb{C}$. \begin{lemma} \label{trig} \begin{enumerate}[\rm(i)] \item Let $z=|z|e^{i\psi}$ and $\lambda=|\lambda|e^{i\varphi} \in \mathbb{C}$, where $|\psi-\varphi|\le\pi$. Then \begin{equation} \label{Fc} |z+\lambda|\ge \cos\left(\frac{\psi-\varphi}{2}\right) (|z|+|\lambda|). \end{equation} \item Let $z \in \overline{\Sigma}_{\psi}$ and $\lambda \in \overline{\Sigma}_\varphi$, where $\psi,\varphi>0$ and $\varphi+\psi<\pi$. Then \begin{equation} \label{Fd} |z+\lambda| \ge \cos\left(\frac{\psi+\theta}{2}\right) (|z|+|\lambda|). \end{equation} \item Let $z=|z|e^{i\psi}$ and $\lambda=|\lambda|e^{i\varphi} \in \mathbb{C}$, where $|\psi|<\pi/2$ and $|\varphi|\le \pi/2$. Then \begin{equation}\label{Fa} |z+\lambda|\ge \cos\psi \, |\lambda|, \end{equation} and \begin{equation}\label{Fb} |z+\lambda|\ge \cos\psi\, |z|. \end{equation} \end{enumerate} \end{lemma} \begin{proof} For \eqref{Fc}, we may assume that $\varphi\ge\psi$ and let $\theta := (\pi-\varphi+\psi)/2 \in[0,\pi/2]$. By applying a rotation of $\mathbb{C}$ we may further assume that $\varphi=\pi-\theta$ and $\psi = \theta$. Then \[ |z+\lambda| \ge \Im z + \Im\lambda = \sin\theta (|z|+|\lambda|) = \cos \left(\frac{\varphi-\psi}{2}\right) (|z|+|\lambda|). \] The inequality \eqref{Fd} follows from \eqref{Fc}, since $\psi+\varphi$ is the maximum value of $|\psi'-\varphi'|$ for $\psi'\in[-\psi,\psi]$ and $\varphi' \in [-\varphi,\varphi]$. The inequality \eqref{Fb} is obtained by considering $\Re(z+\lambda)$. For the inequality \eqref{Fa}, note that \begin{align*} |\lambda+z|^2 -|\lambda|^2\cos^2\psi &= \left(|z|+|\lambda|\cos(\varphi-\psi)\right)^2 + |\lambda|^2 \left(\sin^2\varphi - \cos^2(\varphi-\psi)\right). \end{align*} This is clearly non-negative if $\cos(\varphi-\psi)\ge0$. If $\cos(\varphi-\psi)<0$, and assuming without loss of generality that $\sin\varphi\ge 0$, we have \[ \sin\varphi - |\cos(\varphi-\psi)| = \sin\varphi \, (1+\sin\psi) + \cos \varphi\cos\psi \ge 0. \] This completes the proof. \end{proof} \subsection*{Beta function} The Beta function appears in many places in the paper. It is defined for $s,t>0$ by \[ B(s,t) = B(t,s) := \int_0^1 \tau^{s-1} (1-\tau)^{t-1} \, d\tau = 2 \int_0^{\pi/2} \cos^{2s-1} \psi \sin^{2t-1} \psi \, d\psi. \] In particular, for $s>-1$ we will use the relations \[ B\left(\frac{s+1}{2},\frac{1}{2}\right) = \int_{-\pi/2}^{\pi/2}\cos^s\psi\,d\psi = \int_{-\infty}^\infty \frac{dt}{(1+t^2)^{(s+1)/2}} =\frac{\sqrt{\pi}\Gamma((s+1)/2)}{\Gamma(s/2+1)}, \] see \cite[p.\ 386]{Prud}. We note also the following limit properties: \[ \lim_{s\to -1}\,(s+1)\,B\left(\frac{s+1}{2},\frac{1}{2}\right)=2,\quad \lim_{s\to\infty}\,\sqrt{s}B\left(\frac{s+1}{2},\frac{1}{2}\right)=\sqrt{2\pi}. \] \subsection* {Proof conventions} We will make extensive use of the dominated convergence theorem, often for vector-valued functions. With a few exceptions, we will not give details of the relevant dominating functions, as they are usually easily identified. We will also use the following elementary lemma on several occasions. See \cite[p.21, Lemma 1]{Duren1} for a proof. \begin{lemma} \label{duren21} Let $(\Omega,\mu)$ be a $\sigma$-finite measure space, and $(f_n)_{n\ge1} \subset L^p(\Omega,\mu)$, where $p\in[1,\infty)$. If $f_n \to f_0$ a.e., and $\|f_n\|_{L^p(\Omega,\mu)} \to \|f_0\|_{L^p(\Omega,\mu)}$, then $\|f_n-f_0\|_{L^p(\Omega,\mu)} \to 0$ as $n\to\infty$. \end{lemma} We will use Vitali's theorem several times, usually for holomorphic vector-valued functions. We refer to the version given in \cite[Theorem A.5]{ABHN}. Let $\mathcal{X}$ be a Banach space of holomorphic functions on a domain $\Omega_{\mathcal{X}}$ such that the point evaluations $\delta_z : f \mapsto f(z), \, z \in \Omega$, are continuous on $X$. Let $(\Omega,\mu)$ be either an interval in $\mathbb{R}$ with length measure or an open set in $\mathbb{C}$ with area measure, and $F : \Omega \to \mathcal{X}$ be a continuous function such that $\int_\Omega \|F(t)\|_{\mathcal{X}} \, d\mu(t) < \infty$. Then the integral \[ G := \int_I F(t) \, dt \] exists as a Bochner integral in $\mathcal{X}$ and it can be approximated by Riemann sums. It follows that $G$ belongs to the closed linear span of $\{F(t) : t\in I\}$ in $\mathcal{X}$. Now assume that $F : \Omega \to \mathcal{X}$ is locally bounded, where $\Omega$ is an open set in $\mathbb{C}$, and that $\lambda \mapsto F(\lambda)(z)$ is holomorphic on $\Omega$ for all $z \in \Omega_{\mathcal{X}}$. We shall use the fact that $F: \Omega \to \mathcal{X}$ is holomorphic in the vector-valued sense, without further comment. The result at this level of generality can be seen from \cite[Corollary A.7]{ABHN}, using the point evaluations as separating functionals. An alternative is to show that $F$ is continuous, and then apply Morera's theorem. If the definition of $F$ is by an integral formula, it may also be possible to apply a standard corollary of the dominated convergence theorem which leads to an integral formula for $F'$. \section{The Banach spaces $\mathcal{D}_s$ and their reproducing formulas} In this section we introduce some spaces of holomorphic functions to which we will extend the $\mathcal{B}$-calculus of operators in Section \ref{defD} onwards. \subsection{The spaces $\mathcal{V}_s$} Let $s>-1$, $z=\alpha +i\beta$, and let $\mathcal{V}_s$ be the Banach space of (equivalence classes of) measurable functions $g: \mathbb{C}_{+}\to \mathbb C$ such that the norm \begin{align}\label{norm} \|g\|_{\mathcal{V}_s}:&= \int_{\mathbb{C}_+} \frac{(\Re z)^s |g(z)|}{|z|^{s+1}} \, dS(z) \\ &= \int_0^\infty \alpha^s\int_{-\infty}^\infty\frac{|g(\alpha +i\beta)|}{(\alpha^2+\beta^2)^{(s+1)/2}}\,d\beta\,d\alpha \notag\\ &=\int_{-\pi/2}^{\pi/2}\cos^s\varphi\int_0^\infty |g(\rho e^{i\varphi})|\,d\rho\,d\varphi,\notag \end{align} is finite, where $S$ is the area measure on $\mathbb{C}_+$. Note that \begin{equation} \label{vst} \mathcal{V}_s\subset \mathcal{V}_{\sigma} \qquad \text{and} \qquad \|g\|_{\mathcal{V}_{\sigma}}\le \|g \|_{\mathcal{V}_s}, \qquad g\in \mathcal{V}_s,\; s < \sigma, \end{equation} and \begin{equation} \label{vest} \int_{\Sigma_\psi} \frac{|g(z)|}{|z|} \, dS(z) \le \max\left\{1,\frac{1}{\cos^s\psi}\right\} \|g\|_{\mathcal{V}_s}, \qquad g \in \mathcal{V}_s, \, \psi \in (0,\pi/2). \end{equation} The following property of functions from $\mathcal V_s$ is an essential element in the arguments which lead to the representations for functions in $\mathcal{V}_s$ in Proposition \ref{prim} and for $\mathcal{D}_s$ in Corollary \ref{Repr}, and eventually to the definition of a functional calculus for operators in \eqref{formulaD}. \begin{lemma}\label{Derin} Let $g\in \mathcal{V}_s$ be holomorphic, where $s>-1$. For every $k\ge 1$ and every $\psi \in (0,\pi/2)$, \begin{equation}\label{der1} \lim_{|z|\to\infty ,\;z\in \Sigma_{\psi}}\,z^k g^{(k-1)}(z)=0. \end{equation} \end{lemma} \begin{proof} Let $g \in\mathcal V_s$ be holomorphic, $\psi\in (0,\pi/2)$, $\psi' = (\pi/2 + \psi)/2$ and $b_\psi={\sin((\pi/2-\psi)/2)} = \cos\psi'$. If $z \in \Sigma_\psi$, then \[ \{\lambda\in\mathbb{C}: |\lambda-z|\le b_\psi|z|\}\subset \{\lambda\in\Sigma_{\psi'}: |\lambda|\ge(1-b_\psi)|z|\}. \] Let $r \in (0, b_\psi|z|)$. By Cauchy's integral formula for derivatives, \[ g^{(k-1)}(z)=\frac{(k-1)!}{2\pi i}\int_{\{\lambda:\,|\lambda-z|=r\}} \frac{g(\lambda)}{(\lambda-z)^{k}}\,d\lambda. \] Multiplying by $r^k$ and integrating with respect to $r$ over $(0, b_\psi|z|)$ gives \[ \frac{(b_\psi|z|)^{k+1}}{k+1} |g^{(k-1)}(z)| \le \frac{(k-1)!}{2\pi} \int_{\{\lambda: |\lambda-z|\le b_\psi|z|\}} |g(\lambda)| \, dS(\lambda), \] and then \[ |z|^k |g^{(k-1)}(z)| \le \frac{(k+1)(k-1)!(1+b_\psi)}{2\pi b_\psi^{k+1}} \int_{\{\lambda\in\Sigma_{\psi'}: |\lambda|\ge (1-b_\psi)|z|\}} \frac{|g(\lambda)|}{|\lambda|} \, dS(\lambda). \] It now follows from \eqref{vest} that $|z^k g^{(k-1)}(z)| \to 0$ as $|z|\to\infty, \, z \in \Sigma_\psi$. \end{proof} \subsection{The spaces $\mathcal{D}_s$ and the operators $Q_s$} \label{3.02} We now define a linear operator $Q_s$ on $\mathcal{V}_s, \, s>-1$. It will play a similar role to the operator $Q$ on $\mathcal{W}$ considered in \cite[Section 3]{BGT2}, where $\mathcal{W}$ is the Banach space of all (equivalence classes of) measurable functions $g : \mathbb{C}_+ \to \mathbb{C}$ such that \begin{equation} \label{3.2} \|g\|_{\mathcal{W}} := \int_0^\infty \sup_{\beta\in \mathbb{R}} |g(\alpha+i\beta)| \, d\alpha < \infty. \end{equation} Indeed the definition of $Q_1$ is formally the same as the definition of $Q$ in \cite{BGT2}, but the domain $\mathcal{V}_1$ of $Q_1$ is larger than $\mathcal{W}$. For $g \in \mathcal{V}_s$, let \begin{equation} \label{qdef} (Q_s g)(z):= - \frac{2^s}{\pi}\int_0^\infty \alpha^s\int_{-\infty}^\infty\frac{g(\alpha+i\beta)}{(z+\alpha-i\beta)^{s+1}} \, d\beta\,d\alpha, \quad z\in \mathbb{C}_{+} \cup \{0\}. \end{equation} By \eqref{Fa}, the integral is absolutely convergent, and \begin{equation}\label{angle} |(Q_s g)(z)|\le \frac{2^s\|g\|_{\mathcal{V}_s}}{\pi \cos^{s+1}\psi}, \qquad z \in \Sigma_\psi, \, \psi \in (0,\pi/2). \end{equation} The dominated convergence theorem implies that $Q_s g$ is continuous on $\mathbb{C}_+$, with sectorial limits at infinity and $0$: \begin{equation} \label{qinf} \lim_{|z|\to\infty,\,z\in {\Sigma}_\psi}\,(Q_s g)(z)=0, \qquad \psi \in (0,\pi/2), \end{equation} and \begin{equation} \label{qzero} \lim_{|z|\to 0,\,z\in {\Sigma}_\psi}\,(Q_s g)(z)=(Q_s g)(0), \qquad \psi \in (0,\pi/2). \end{equation} Thus $Q_s g$ is bounded and continuous on $\overline{\Sigma}_\psi$ for every $\psi\in (0,\pi/2)$. Moreover, $Q_s g$ is holomorphic on $\mathbb{C}_+$ and \begin{equation} \label{qder} (Q_s g)'(z) = (s+1)\frac{2^s}{\pi}\int_0^\infty \alpha^s\int_{-\infty}^\infty\frac{g(\alpha + i\beta)}{(z+\alpha-i\beta)^{s+2}}\,d\beta \, d\alpha, \qquad z\in \mathbb{C}_{+}. \end{equation} Using this, (\ref{Fa}) and (\ref{Fb}), we obtain that \begin{equation}\label{Dangle} |(Q_s g)'(z)|\le \frac{(s+1)2^s}{\pi\cos^{s+2}\psi \,|z|}\|g\|_{\mathcal{V}_s},\qquad z\in \Sigma_\psi, \end{equation} and \begin{equation*} |z(Q_s g)'(z)|\le \frac{(s+1)2^s}{\pi\cos^{s+2}\psi}\int_{\mathbb{C}_{+}} \frac{|z|(\Re \lambda)^s\,|g(\lambda)|}{|z+\overline{\lambda}||\lambda|^{s+1}}\,dS(\lambda), \qquad z\in \Sigma_\psi. \end{equation*} Using the dominated convergence theorem again, we obtain \begin{equation} \label{Lebeg11} \lim_{|z|\to0,z\in\Sigma_\psi} z(Q_sg)'(z) = 0. \end{equation} We now give another formula for $Q_s$. Let $s=n+\delta>-1$ where $n\in \mathbb{N}\cup\{-1,0\}$ and $\delta\in [0,1)$, and let \begin{equation}\label{Const46} C_s:=\int_0^\infty\frac{dt}{(t+1)^{n+2}t^\delta}= \int_0^1\frac{(1-\tau)^s}{\tau^\delta}\,d\tau=B(1-\delta,s+1). \end{equation} Then \[ \int_0^\infty \frac{dt}{(\lambda+t)^{n+2}t^\delta} = \frac{C_s}{\lambda^{s+1}},\quad \lambda\in \mathbb{C}_{+}. \] Indeed, both sides of this equation are holomorphic functions of $\lambda\in\mathbb{C}_+$, and they coincide for $\lambda \in (0,\infty)$, so they coincide for all $\lambda\in\mathbb{C}_+$, by the identity theorem for holomorphic functions. Putting $\lambda = z+\alpha-i\beta$, we obtain \begin{equation}\label{46RRR} (Q_s g)(z) =- \frac{2^s}{\pi C_s}\int_0^\infty \int_0^\infty\int_{-\infty}^\infty \frac{\alpha^s\,g(\alpha+i\beta)\,d\beta\,d\alpha}{(z+\alpha-i\beta+t)^{n+2}}\,\frac{dt}{t^\delta},\quad z\in \mathbb{C}_{+}. \end{equation} For $s>-1$ let $\mathcal{D}_s$ be the linear space of all holomorphic functions $f$ on $\mathbb{C}_{+}$ such that \[ f'\in \mathcal{V}_s, \] equipped with the semi-norm \[ \|f\|_{\mathcal D_{s,0}}:=\|f'\|_{\mathcal{V}_s}, \qquad f \in \mathcal D_s. \] If $\sigma>s>-1$, then it is immediate from \eqref{vst} that $\mathcal{D}_s \subset \mathcal{D}_\sigma$. We will exhibit some functions in $\mathcal{D}_s$ later in this section and in Section \ref{some_functions}. In the rest of this section we will also obtain a reproducing formula \eqref{qnrep} for functions from $\mathcal{D}_s$ and describe some basic properties which will be relevant for the sequel. To this aim, we first define and study the behaviour of operators $Q_s$ on the scale of $\mathcal D_s$-spaces. Recall that in \cite[Proposition 3.1]{BGT2} we showed that $Q$ maps $\mathcal{W}$ into $\mathcal{B}$. However $Q_s$ does not map the whole of $\mathcal{V}_s$ into $\mathcal{D}_s$. For $s > -1$, a function $g \in \mathcal{V}_s$ for which $Q_sg \notin \mathcal{D}_s$, can be defined as follows: \begin{multline*} g(\rho e^{i\varphi}) := \left(\cos^{s+1} \varphi (\rho-\sin\varphi) \log^2(\rho-\sin\varphi)\right)^{-1}, \\ 1< \rho < 2 - \sin\varphi, \, \pi/4 < \varphi < \pi/2, \end{multline*} and $g(z)= 0$ for all other points in $\mathbb{C}_+$. We do not give details in this paper. Instead we will show in Propositions \ref{Furt} and \ref{prim} that $Q_s$ maps $\mathcal{V}_s$ boundedly into $\mathcal{D}_{\sigma}$ for any $\sigma>s$, and it maps holomorphic functions in $\mathcal{V}_s$ into $\mathcal{D}_s$. We need the following auxiliary lemma which will be useful in a number of instances. \begin{lemma}\label{D0012} Let $h\in L^1[0,1] \cap L^\infty[1/2,1]$ be a positive function. Let $s>-1$, $\beta>1/2$, and \begin{equation} \label{defG} G_{h,\beta,s}(\varphi):= \int_{-\pi/2}^{\pi/2}\cos^s\psi \int_0^1\frac{h(t)\,dt}{(t^2+2t\cos(\varphi+\psi)+1)^\beta}\,d\psi,\quad \varphi\in (-\pi/2,\pi/2). \end{equation} \begin{enumerate}[\rm(a)] \item If $2\beta-s-2\le0$, then \begin{equation}\label{DerAn00} K_{h,\beta,s} := \sup_{|\varphi|<\pi/2}\,G_{h,\beta,s}(\varphi)<\infty. \end{equation} \item If $2\beta-s-2\ge0$, then \begin{equation}\label{DerAn01} \tilde{K}_{h,\beta,s} := \sup_{|\varphi|<\pi/2}\, \cos^{2\beta-s-2}\varphi \, G_{h,\beta,s}(\varphi)<\infty. \end{equation} \end{enumerate} \end{lemma} \begin{proof} Since $G_{h,\beta,s}(-\varphi)=G_{h,\beta,s}(\varphi)$, we may assume that $\varphi\in [0,\pi/2)$. Now \begin{align*} G_{h,\beta,s}(\varphi) &= \int_{-\pi/2}^{0}\cos^s\psi \int_0^1\frac{h(t)\,dt}{(t^2+2t\cos(\varphi+\psi)+1)^\beta}\,d\psi\\ &\null \hskip30pt + \int_0^{\pi/2}\cos^s\psi\int_0^1\frac{h(t)\,dt}{(t^2+2t\cos(\varphi+\psi)+1)^\beta}\,d\psi\\ &=: G_{h,\beta, s}^{-}(\varphi)+G_{h,\beta, s}^+(\varphi), \end{align*} and we estimate these two integrals separately. Since $\varphi \in [0,\pi/2)$, $\beta>1/2$, and $s>-1$, we have \begin{align*}\label{Eqa1} G_{h,\beta, s}^{-}(\varphi) &\le \int_{-\pi/2}^{0}\cos^s\psi \,d\psi\int_0^1 h(t)\,dt = \frac{\|h\|_{L^1[0,1]}}{2}B\left(\frac{s+1}{2},\frac{1}{2}\right). \end{align*} For the second integral, note that if $\phi = \varphi+\psi \in [0, \pi)$, \[ t^2+2t\cos\phi+1\ge \begin{cases} \frac{1}{4},\quad &t\in [0,1/2], \\ (1-t)^2 + 1 + \cos\phi, \quad &t\in [1/2,1]. \end{cases} \] Hence \begin{align*} \lefteqn{\hskip-30pt\int_0^1\frac{h(t)}{(t^2+2t\cos\phi+1)^\beta}\,dt } \\ &\le 4^\beta\int_0^{1/2} h(t)\,dt+ \int_{1/2}^1\frac{h(t)}{((t-1)^2+(1+\cos\phi))^\beta}\,dt\\ &\le 4^\beta \|h\|_{L^1[0,1/2]}+ \|h\|_{L^\infty[1/2,1]}\int_0^\infty\frac{d\tau}{(\tau^2+ 2\cos^2(\phi/2))^{\beta}} \\ &\le \frac{C_{h,\beta}}{\cos^{2\beta-1}(\phi/2)}, \end{align*} for some constant $C_{h,\beta}>0$. Replacing $\varphi$ by $\pi/2 -\varphi$ and $\psi$ by $\pi/2-\psi$, and using \[ \omega\ge \sin \omega\ge \frac{2}{\pi}\omega,\qquad \omega\in (0,\pi/2), \] we infer that if $\varphi\in [0,\pi/2)$, then \begin{align*} G^+_{h,\beta,s}(\pi/2-\varphi) &\le C_{h,\beta} \int_0^{\pi/2}\frac{\sin^s\psi}{\sin^{2\beta-1}((\varphi+\psi)/2)} \,d\psi\\ &\le C_{h,\beta}\pi^{2\beta-1} \int_0^{\pi/2}\frac{\psi^s}{(\varphi+\psi)^{2\beta-1}} \,d\psi. \end{align*} In case (a), when $2\beta-s-2 \le 0$, we have \[ \int_0^{\pi/2} \frac{\psi^s}{(\varphi+\psi)^{2\beta-1}} \,d\psi \le \int_0^{\pi/2} \psi^{s-2\beta+1} \,d\psi = \frac{\pi^{s-2\beta+2}}{(s-2\beta+2)2^{s-2\beta+2}}<\infty. \] In case (b), when $2\beta - s -2 \ge 0$, we obtain \[ \int_0^{\pi/2} \frac{\psi^s}{(\varphi+\psi)^{2\beta-1}} \,d\psi \le \varphi^{s+2-2\beta} \int_0^\infty \frac{t^s}{(t+1)^{2\beta-1}} \,dt, \] and hence \begin{multline*} \cos^{2\beta-s-2}(\pi/2-\varphi) \, G_{h,\beta,s}(\pi/2-\varphi) \le \varphi^{2\beta-s-2} G_{h,\beta,s}(\pi/2-\varphi)\\ \null \hskip-20pt \le C_{h,\beta} \pi^{2\beta-1} \int_0^\infty \frac{t^s}{(t+1)^{2\beta-1}} \,dt < \infty, \end{multline*} for some constant $C_{h,\beta}>0$. \end{proof} Let $\mathcal{D}_{s,0}$ be the space of all functions $f \in \mathcal{D}_s$ such that $f$ has a sectorial limit $0$ at infinity. Then $(\mathcal{D}_{s,0}, \|\cdot\|_{\mathcal{D}_{s,0}})$ is a normed space, and we shall see in Corollary \ref{Ban_D} that it is a Banach space. The following basic examples will play roles in several estimates later in the paper. We start with the the resolvent functions and their powers. \begin{exa} \label{resd} Let $\lambda=|\lambda|e^{i\varphi}\in \mathbb{C}_{+}$, and $r_\lambda(z)= (z+\lambda)^{-1}$. Let $\gamma>0$, and consider \[ r_{\lambda}^{\gamma}(z):=(z+\lambda)^{-\gamma},\quad z \in \mathbb{C}_+. \] Let $s>-1$. Then \begin{align*} \|r_{\lambda}^{\gamma}\|_{\mathcal{D}_{s,0}} &=\gamma \int_{-\pi/2}^{\pi/2} \cos^s\psi\int_0^\infty \frac{d\rho}{|\rho e^{i\psi}+r e^{i\varphi}|^{\gamma+1}}\,d\psi\\ &=\frac{\gamma}{|\lambda|^{\gamma}}\int_{-\pi/2}^{\pi/2} \cos^s\psi\int_0^\infty \frac{d\rho}{|\rho+e^{i(\varphi-\psi)}|^{\gamma+1}}\,d\psi\\ &=\frac{\gamma}{|\lambda|^{\gamma}} \int_{-\pi/2}^{\pi/2} \cos^s\psi\int_0^1 \frac{1+t^{\gamma-1}} {(t^2+2t\cos(\varphi+\psi)+1)^{(\gamma+1)/2}}\,dt\,d\psi, \end{align*} where we have put $t = \rho^{-1}$ for $\rho>1$. Now we apply Lemma \ref{D0012} with $h(t) = 1+t^{\gamma-1}$, $\beta = (\gamma+1)/2$, so $2\beta - s -2 = \gamma -s -1$. Thus we obtain \begin{equation} \label{rtl} \|r_{\lambda}^{\gamma}\|_{\mathcal{D}_{s,0}} \le \begin{cases} \dfrac{\gamma K_{h,(\gamma+1)/2,s}}{|\lambda|^\gamma}, &s>\gamma-1>-1, \\ \noalign{\vskip8pt} \dfrac{\gamma \tilde{K}_{h,(\gamma+1)/2,s}}{|\lambda|^{\gamma}\cos^{\gamma-s-1}\varphi}, &\gamma-1> s>-1. \end{cases} \end{equation} In particular, taking $\gamma=1$ and a fixed $s>0$, \begin{equation} \label{r_l} \|r_\lambda\|_{\mathcal{D}_{s,0}} = \int_{-\pi/2}^{\pi/2} \int_0^\infty \frac{\cos^s \psi}{|\lambda+\rho e^{i\psi}|^2} \,d\rho \,d\psi \le \frac{C_s}{|\lambda|}, \qquad\lambda \in \mathbb{C}_+. \end{equation} This estimate will play a crucial role in the proof of Theorem \ref{dcalculus_intro}. \end{exa} Next we consider some functions which appear frequently in the studies of holomorphic $C_0$-semigroups. \begin{exa}\label{dm} Let \[ f_\nu(z):=z^{\nu} e^{-z},\quad z\in \mathbb{C}_{+},\quad \nu\ge0. \] We shall show here that $f_\nu\in \mathcal{D}_s$ if and only if $s>\nu$. Moreover, if $s >\nu$, then \begin{equation}\label{frac} \|f_\nu\|_{\mathcal{D}_{s,0}}\le 2 B\left(\frac{s-\nu}{2},\frac{1}{2}\right)\Gamma(\nu+1). \end{equation} This estimate will be crucial for operator estimates in Section \ref{norm_estimates}. We have \[ f_\nu'(z)=e^{-z}\left(\nu{z^{\nu-1}}-z^\nu\right) \] and \begin{align*} \|f_\nu\|_{\mathcal{D}_{s,0}}&=\int_{-\pi/2}^{\pi/2} \cos^s\varphi\int_0^\infty {e^{-\rho\cos\varphi}}{\rho^{\nu-1}}\left|{\rho e^{i\varphi}}-\nu\right|\,d\rho\,d\varphi\\ &=\int_{-\pi/2}^{\pi/2} \cos^{s-\nu-1}\varphi\int_0^\infty e^{-r}r^{\nu-1}| r e^{-i\varphi}-\nu\cos\varphi|\,dr\,d\varphi. \end{align*} We use the estimates \[ r |\sin\varphi| \le |r e^{-i\varphi}-\nu\cos\varphi| \le r+\nu. \] If $s > \nu$, we obtain \begin{align*} \|f_\nu\|_{\mathcal{D}_{s,0}} &\le 2\int_0^{\pi/2} \cos^{s-\nu-1}\varphi\,d\varphi \int_0^\infty e^{-r} r^{\nu-1}(r+\nu) \,dr \\ &= 2 B\left(\frac{s-\nu}{2},\frac{1}{2}\right)\Gamma(\nu+1). \notag \end{align*} If $s \le \nu$, then \[ \|f_\nu\|_{\mathcal{D}_{s,0}} \ge 2\int_0^{\pi/2} \cos^{s-\nu-1}\varphi \, \sin\varphi\,d\varphi \int_0^\infty e^{-\tau} \tau^{\nu} \,d\tau = \infty. \] This establishes the claims above. \end{exa} Finally we consider a function which will play an important role in our constructions of functional calculi in Section \ref{hardy_sobolev}. \begin{exa} \label{arc} The function $\arccot$ is defined by \begin{equation} \label{arcdef} \arccot(z) = \frac{1}{2i} \log \left(\frac{z+i}{z-i}\right), \quad z \in \mathbb{C}_+. \end{equation} Then $|\Re \arccot(z)| \le \pi/2$, $\arccot$ has sectorial limit $0$ at infinity, and its derivative is $- (z^2+1)^{-1}$. It is easy to see that $\arccot \in \mathcal{D}_s$ for all $s>-1$. For $s = 0$, we have \begin{align} \label{arccot_D} &\\ \lefteqn{\|\arccot\|_{\mathcal{D}_{0,0}}} \notag\\ &=\int_{-\pi/2}^{\pi/2}\int_0^\infty \frac{d\rho}{|1+\rho^2e^{2i\varphi}|}\,d\varphi =\int_0^\infty \int_0^{\pi}\frac{d\psi}{(\rho^4+2\rho^2\cos\psi+1)^{1/2}}\,d\rho \notag\\ &\le \sqrt{\pi}\int_0^\infty \left(\int_0^{\pi} \frac{d\psi}{\rho^4+2\rho^2\cos\psi+1}\right)^{1/2}\,d\rho = \sqrt{\pi}\int_0^\infty \left(\frac{\pi}{|\rho^4-1|}\right)^{1/2}\,d\rho\notag\\ &=\frac{\pi}{2} B(1/4,1/2) < 3 \pi.\notag \end{align} See \cite[2.5.16, (38)]{Prud} for the evaluation of the integral with respect to $\psi$. \end{exa} \begin{prop}\label{Furt} Let $\sigma > s > -1$. The following hold. \begin{enumerate}[\rm(i)] \item $\mathcal{D}_s$ is continuously embedded in $\mathcal{D}_\sigma$. \item The restriction of $Q_\sigma$ to $\mathcal{V}_s$ is in $L(\mathcal{V}_s,\mathcal{D}_s)$. \item $Q_s \in L (\mathcal{V}_s, \mathcal{D}_{\sigma})$. \end{enumerate} \end{prop} \begin{proof} The first statement is immediate from the definitions of the spaces and \eqref{vst}. For the second statement, let $g \in \mathcal{V}_s$. From \eqref{norm} and \eqref{qder}, we have \begin{align*} \lefteqn{\frac{\pi}{2^\sigma(\sigma+1)} \|Q_{\sigma}g\|_{\mathcal{D}_{s,0}}\hskip10pt}\\ &\le \frac{1}{\sigma+1} \int_{-\pi/2}^{\pi/2}\cos^{s}\psi \int_0^\infty \int_{-\pi/2}^{\pi/2} \cos^\sigma \varphi \int_0^\infty \frac{r^{\sigma+1}|g(t e^{i\varphi})|}{|\rho e^{i\psi}+t e^{-i\varphi}|^{\sigma+2}}\,d\rho\,d\varphi \,dt\,d\psi\\ &= \int_{-\pi/2}^{\pi/2} \cos^\sigma \varphi \int_0^\infty t^{\sigma+1}\|r_{t e^{-i\varphi}}^{\sigma+1}\|_{\mathcal{D}_{s,0}} |g(t e^{i\varphi})|\,dt\,d\varphi \\ &\le K_{h,(s+2)/2,\sigma} \|g\|_{\mathcal{V}_s}, \end{align*} where $h(t) = 1+t^s$. Here we have used the first case of \eqref{rtl} with $s$ and $\sigma$ interchanged. This establishes the second statement. For the third statement, the same estimation, but with $s$ and $\sigma$ interchanged, setting $h(t) = 1+t^\sigma$, and using the second case of \eqref{rtl}, shows that \[ \frac{\pi}{2^s(s+1)} \|Q_{s}g\|_{\mathcal{D}_{\sigma,0}}\le \tilde{K}_{h,(s+2)/2,\sigma} \|g\|_{\mathcal{V}_s}. \] This establishes the third statement. \end{proof} For insight on why $\|r_{\lambda}^{\sigma+1}\|_{\mathcal{D}_{s,0}}$ appears in the proof above, we refer the reader to the proof of Theorem \ref{D00}. \begin{prop} \label{prim} Let $g \in \mathcal{V}_s$ be holomorphic, where $s>-1$. Then $Q_s g \in \mathcal{D}_s$ and $(Q_s g)'= g$. \end{prop} \begin{proof} First we consider the case when $g\in \mathcal{V}_n$, where $n \in \mathbb{N}\cup\{0\}$. Then $Q_n g$ is holomorphic and \eqref{qdef} holds for $s=n$. It suffices to show that $(Q_ng)' = g$. Let \begin{align*} I(\alpha) &= \int_{-\infty}^\infty \frac{|g(\alpha+i\beta)|}{(\alpha^2+\beta^2)^{(n+1)/2}} \,d\beta, \\ J(\beta) &= \int_0^\infty \frac{\alpha^n|g(\alpha+i\beta)|}{(\alpha^2+\beta^2)^{(n+1)/2}} \,d\alpha. \end{align*} From \eqref{norm} and Fubini's theorem, we see that $\int_0^\infty \alpha^n I(\alpha) \, d\alpha < \infty$. Hence $I(\alpha)<\infty$ for almost all $\alpha>0$ and it follows that there exists a sequence $(\alpha_j)_{j\ge 1}$ such that \[ \lim_{j \to \infty}\alpha_j =\infty, \qquad \lim_{j\to\infty} I(\alpha_j) = 0. \] Similarly $\int_{-\infty}^\infty J(\beta)\,d\beta < \infty$, and so there exist sequences $(\beta_k^{\pm})_{k \ge 1}$ such that \[ \lim_{k \to \infty}\beta_k^\pm = \pm\infty, \qquad \lim_{k\to\infty} J(\beta_k^\pm) = 0. \] Let $z \in \mathbb{C}_+$ be fixed. Take $\alpha>0$ with $I(\alpha)<\infty$, and let $j$ be sufficiently large that $\alpha_j > 2\alpha + \Re z$, and $k$ be sufficiently large that $\beta_k^- < \Im z < \beta_k^+$. We may apply the Cauchy integral formula around the rectangle with vertices $\alpha +i\beta_k^\pm$ and $\alpha_j + i \beta_k^{\pm}$, and we obtain \begin{align*} \lefteqn{\frac{2\pi}{n!} g^{(n)}(2\alpha+z)} \\ &= \int_{\beta_k^-}^{\beta_k^+} \frac{g(\alpha_j+i\beta)}{(\alpha_j+i\beta - z -2\alpha)^{n+1}} d\beta - \int_{\beta_k^-}^{\beta_k^+} \frac{g(\alpha+i\beta)}{(\alpha+i\beta - z -2\alpha)^{n+1}} d\beta \\ &\null\hskip10pt -i \int_{\alpha}^{\alpha_j}\frac{g(s+i\beta_k^{-})}{(s+i\beta_k^{-}- z-2\alpha)^{n+1}}\,ds + i \int_{\alpha}^{\alpha_j}\frac{g(s+i\beta_k^{+})}{(s+i\beta_k^{+} - z -2\alpha)^{n+1}}\,ds. \end{align*} Letting $k \to \infty$, we obtain \[ (-1)^n \frac{2\pi}{n!} g^{(n)}(2\alpha+z) = \int_{-\infty}^\infty \frac{g(\alpha+i\beta)}{(z+\alpha-i\beta)^{n+1}} \,d\beta - \int_{-\infty}^\infty \frac{g(\alpha_j+i\beta)}{(z+\alpha_j-i\beta)^{n+1}} \,d\beta. \] Letting $j \to \infty$, we obtain \begin{equation} \label{Qn} (-1)^n \frac{2\pi}{n!} g^{(n)}(2\alpha+z) = \int_{-\infty}^\infty \frac{g(\alpha+i\beta)}{(z+\alpha-i\beta)^{n+1}} \,d\beta. \end{equation} This holds for almost all $\alpha>0$. Substituting this into \eqref{qdef}, and then integrating by parts and using Lemma \ref{Derin}, we infer that \begin{align} \label{Qnn} (Q_n g)(z) &= (-1)^{n+1} \frac{2^{n+1}}{n!} \int_0^\infty \alpha^{n} g^{(n)}(2\alpha+z)\,d\alpha \\ &=\frac{(-1)^{n+1}}{n!}\int_0^\infty \alpha^n g^{(n)}(\alpha+z)\,d\alpha = - \int_0^\infty g(\alpha+z)\,d\alpha. \nonumber \end{align} Here the first two integrals are absolutely convergent, by Fubini's theorem. The final integral may be improper, but it follows that $(Q_n g)' = g$. Now consider the case when $s=n+\delta>-1$, where $n \in \mathbb{N}\cup\{-1,0\}$, $\delta \in (0,1)$, and $g \in \mathcal{V}_n$ is holomorphic. Then $g\in \mathcal{V}_{n+1}$, $n+1\in \mathbb{N}\cup\{0\}$, and \eqref{Qn} gives \[ \int_{-\infty}^\infty \frac{g(\alpha+i\beta)}{(z+\alpha-i\beta+t)^{n+2}} \,d\beta= (-1)^{n+1}\frac{2\pi}{(n+1)!} g^{(n+1)}(2\alpha+z+t), \] for $z \in \mathbb{C}_+$, $t>0$ and almost all $\alpha>0$. We obtain from (\ref{46RRR}) and \eqref{Const46} that \begin{align*} C_s (Q_s g)(z) &=- \frac{2^s}{\pi}\int_0^\infty \int_0^\infty\int_{-\infty}^\infty \frac{\alpha^s\,g(\alpha+i\beta)\,d\beta\,d\alpha}{(z+\alpha-i\beta+t)^{n+2}}\,\frac{dt}{t^\delta}\\ &=(-1)^n\frac{2^{s+1}}{(n+1)!}\int_0^\infty \int_0^\infty \alpha^s g^{(n+1)}(2\alpha+z+t)\, d\alpha\,\frac{dt}{t^\delta}\\ &=\frac{(-1)^n}{(n+1)!}\int_0^\infty \int_t^\infty (\tau-t)^s g^{(n+1)}(\tau+z) \, d\tau\,\frac{dt}{t^\delta}\\ &=C_s\frac{(-1)^n}{(n+1)!}\int_0^\infty \tau^{n+1} \,g^{(n+1)}(\tau+z)\,d\tau. \end{align*} As in \eqref{Qnn}, it follows that $(Q_sg)' = g$. \end{proof} \begin{cor} If $g \in \mathcal{V}_s$ is holomorphic, then $Q_\sigma g = Q_s g$ for all $\sigma\ge s$. \end{cor} \begin{proof} This is immediate from Proposition \ref{prim} and \eqref{qinf}. \end{proof} The following representation of functions in $\mathcal{D}_s$ has appeared in \cite[Corollary 4.2]{Al2014} (see also \cite[Lemma 3.13.2]{AlPel} for the case $s=1$). \begin{cor}\label{Repr} Let $f\in \mathcal{D}_s$, $s>-1 $. Then the sectorial limits \begin{align} f(\infty): &=\lim_{z\to\infty ,\;z\in \Sigma_{\psi}}\,f(z),\label{finf}\\ f(0):&= \lim_{z\to 0,\;z\in \Sigma_{\psi}}\,f(z)\label{fzero}, \end{align} exist in $\mathbb{C}$ for every $\psi \in (0,\pi/2)$. Moreover, for all $z \in \mathbb{C}_{+}\cup \{0\}$, \begin{align} \label{qnrep} f(z)&=f(\infty)+(Q_s f')(z)\\ &=f(\infty)- \frac{2^s}{\pi}\int_0^\infty \alpha^s\int_{-\infty}^\infty \frac{f'(\alpha+i\beta)\,d\beta}{(z+\alpha-i\beta)^{s+1}} \, d\alpha.\notag \end{align} With $f(0)$ defined as above, $f \in C(\overline{\Sigma}_\psi)$ for every $\psi\in (0,\pi/2)$. \end{cor} \begin{proof} It follows from Proposition \ref{prim} that $(Q_s f')' = f'$. The statements follow from \eqref{qinf} and \eqref{qzero}. \end{proof} \begin{cor}\label{Ban_D} For every $s>-1$ the space $\mathcal D_s$ equipped with the norm \[ \|f\|_{\mathcal D_s}:= |f(\infty)|+\|f\|_{\mathcal D_{s,0}}, \qquad f \in\mathcal D_s, \] is a Banach space. \end{cor} \begin{proof} Let $s>-1$ be fixed and let $(f_k)_{k=1}^\infty$ be a Cauchy sequence in $\mathcal D_s$. Then \eqref{angle} and Vitali's theorem imply that $(Q_s f_k')_{k=1}^\infty$ converges uniformly on each $\Sigma_\psi$ to a limit $g$ which is holomorphic on $\mathbb{C}_+$. Moreover $(f_k(\infty))_{k=1}^{\infty}$ converges to a limit $\zeta \in\mathbb{C}$. It follows from Proposition \ref{prim} that $(f_k)_{k=1}^\infty$ converges uniformly on $\Sigma_\psi$ to $f := \zeta + g$. Then $(f'_k)_{k=1}^\infty$ converges pointwise on $\mathbb{C}_+$ to $g' = f'$. Applying Fatou's lemma to the sequences $(\|f'_k-f'_n\|_{\mathcal{V}_s})_{n=k}^\infty$ for fixed $k$, one sees that $\|f_k' - f'\|_{\mathcal V_s} \to 0$. So $f' \in \mathcal{V}_s$ and $f \in \mathcal{D}_s$. By \eqref{qinf}, $f(\infty)=\zeta$ and so $\|f_k - f\|_{\mathcal{D}_s} \to 0$, $k\to\infty$. \end{proof} The argument used in the proof of Corollary \ref{Ban_D} also provides the following lemma of Fatou type (see also Lemma \ref{FatouH}). \begin{cor}\label{Fatou} Let $s>-1$ and $(f_k)_{k=1}^\infty \subset \mathcal D_s$ be such that $\sup_{k \ge 1} \|f_k\|_{\mathcal D_s}<\infty$ and $f(z)=\lim_{k \to \infty} f_k(z)$ exists for all $z \in \mathbb C_+$. Then $f \in \mathcal D_s$. \end{cor} Now employing \eqref{qnrep}, the estimates (\ref{angle}), (\ref{Dangle}) and (\ref{Lebeg11}), and Lemma \ref{Derin}, we obtain the following estimates. \begin{cor}\label{Cangle} Let $\psi \in (0,\pi/2)$. For all $f\in \mathcal{D}_s$, $s>-1$, \[ |f(z)|\le \max\left(1,\frac{2^s}{\pi\cos^{s+1}\psi}\right)\|f\|_{\mathcal{D}_s},\qquad z\in \Sigma_\psi, \] and \[ |f'(z)|\le \frac{(s+1)2^s}{\pi |z| \cos^{s+2}\psi}\|f'\|_{\mathcal{V}_s},\qquad z\in\Sigma_\psi. \] Moreover, $f$ is continuous on $\Sigma_\psi \cup \{0\}$ and \[ \lim_{|z|\to0, z \in \Sigma_\psi} z f'(z) = 0. \] \end{cor} \begin{rem} \label{remdr} Corollary \ref{Cangle} implies that the point evaluation functionals $\delta_z, \, z \in \mathbb{C}_+$, are continuous on $\mathcal{D}_s, \, s>-1$. Using \eqref{rtl} and the principle set out in Section \ref{prelims}, we see that the function $\lambda \mapsto r_{\lambda}^{\gamma}$ is holomorphic from $\mathbb{C}_+$ to $\mathcal{D}_s$, for any $s>-1$, $\gamma>0$. \end{rem} \subsection{More functions in $\mathcal{D}_s$ and their properties}\label{some_functions} In this section we give more examples of functions from $\mathcal{D}_s$ and note some additional elementary properties which will be relevant for the sequel. \begin{prop} \label{BDs} For $s > 0$, $\mathcal{B} \overset{i}{\hookrightarrow} \mathcal{D}_s$. \end{prop} \begin{proof} For $s>0$ and $f\in \mathcal{B}$ we have \begin{align*} \|f\|_{\mathcal{D}_{s,0}}&=|f(\infty)| +\int_0^\infty \alpha^s\int_{-\infty}^\infty\frac{|f'(\alpha+i\beta)|} {|\alpha+i\beta|^{s+1}} \,d\beta\,d\alpha \label{inclus_b}\\ &\le \|f\|_{\infty} +2\int_0^\infty \sup_{\beta\in \mathbb{R}}\,|f'(\alpha+i\beta)|\, \int_{0}^\infty\frac{dt} {(t^2+1)^{(s+1)/2}}\,d\alpha\notag\\ &\le \max\{1, B(1/2,s/2)\}\|f\|_\mathcal{B}.\notag \end{align*} Thus $\mathcal{B}$ is continuously included in $\mathcal{D}_s$. \end{proof} \begin{rem} \label{Brep} The representation \eqref{qnrep} in Corollary \ref{Repr} extends the reproducing formula for $\mathcal{B}$, i.e., \eqref{qnrep} for $s=1$, to a larger class of functions. \end{rem} Recall that $\mathcal{LM} \overset{i}{\hookrightarrow} \mathcal{B}$, see \cite[Section 2.4]{BGT}, Thus, in view of Proposition \ref{BDs}, we have \[ \mathcal{LM} \overset{i}{\hookrightarrow} \mathcal{B} \overset{i}{\hookrightarrow} \mathcal{D}_s, \qquad s>0. \] We will show in Corollary \ref{Bdens} that $\mathcal B$ is dense in $\mathcal D_s$ for every $s>0$, and hence $\mathcal D_s$ is dense in $\mathcal D_\sigma$ for all $\sigma > s > 0$. On the other hand, we will show in Corollary \ref{Bdens} that $\mathcal{B}$ is not dense in $\mathcal{D}^\infty_s$ for $s>0$. For $f \in \operatorname{Hol}(\mathbb{C}_+)$, let \[ \tilde{f}(z):=f(1/z), \qquad f_t(z): = f(tz), \qquad t >0, \quad z\in \mathbb{C}_{+}. \] \begin{lemma} \label{leminv} Let $s>-1$ and $t>0$. Then \begin{enumerate}[\rm(i)] \item $f\in \mathcal{D}_s$ if and only if $\tilde{f}\in \mathcal{D}_s$ and \[ \|f-f(\infty)\|_{\mathcal{D}_s}=\|\tilde{f}- f(0)\|_{\mathcal{D}_s}. \] \item If $f$ is bounded away from $0$ and $f \in \mathcal{D}_s$, then $1/f \in \mathcal{D}_s$. \item $f \in \mathcal{D}_s$ if and only if $f_t \in \mathcal{D}_s$, and $\|f\|_{\mathcal{D}_s} = \|f_t\|_{\mathcal{D}_s}$. \end{enumerate} \end{lemma} \begin{proof} Note that \begin{align*} \|\tilde{f}'\|_{\mathcal{V}_s}&=\int_{-\pi/2}^{\pi/2}\cos^s\varphi\int_0^\infty \frac{|f'(\rho^{-1}e^{-i\varphi})|}{\rho^2}\,d\rho\,d\phi\\ &=\int_{-\pi/2}^{\pi/2}\cos^s\varphi\int_0^\infty |f'(r e^{i\varphi})|\,dr\,d\varphi\\ &=\|f'\|_{\mathcal{V}_s}. \end{align*} Moreover, by Corollary \ref{Repr}, $\tilde{f}(\infty)=f(0)$. This proves (i). The other statements are very easy. \end{proof} \begin{rems} \label{e-z} 1. Neither of the spaces $\mathcal{B}$ and $\mathcal{D}_0$ is contained in the other. Indeed, the function $e^{-z} \in \mathcal{LM} \subset \mathcal{B}$ but $e^{-z} \not \in \mathcal D_0$ (see Example \ref{dm}). On the other hand, there are bounded functions in $\mathcal{D}_0$ which are not in $\mathcal{B}$, for example the function $\exp(\arccot z) \in \mathcal D_0$ and is bounded but it is not in $\mathcal{B}$ (see Example \ref{eac}). More generally, for $\nu\ge0$, let $f_\nu(z) = z^\nu e^{-z}, \, z\in \mathbb{C}_+$. Then $f_\nu \in \mathcal{D}_s$ if and only if $s>\nu$ (see Example \ref{dm}). Note that, if $\nu>0$, $f_\nu$ is not bounded on any right half-plane. One can show that if $f \in \mathcal{D}_{s, 0}$, then \[ |f(z)| \le \frac{2^s}{\pi} \|f\|_{\mathcal{D}_{s,0}}\left(1+\frac{4\beta^2}{\alpha^2}\right)^{(s+1)/2}, \quad z =\alpha+i\beta \in \mathbb{C}_+. \] The function $\log(1+z) e^{-z}$ is in $\mathcal{D}_s, s>0$, but is also unbounded on every right half-plane. \noindent 2. Since $e^{-z} \in \mathcal{D}_s$ for $s>0$, it follows from Lemma \ref{leminv} that the functions $e^{-t/z}$ are in $\mathcal D_s$ for all $t>0, \, s >0$. This shows that functions $f \in \mathcal D_s$ may not have full limits at infinity or at zero. However, the properties \eqref{finf} and \eqref{fzero} in Corollary \ref{Repr} establish values for $f$ at infinity and at zero as sectorial limits. \noindent 3. The spaces $\mathcal{D}_s, \, s>-1$, are invariant under shifts given by \[ (T(\tau)f)(z) = f(z+\tau), \qquad f \in \mathcal{D}_s, \, \tau \in \mathbb{C}_+, \, z \in \mathbb{C}_+. \] Indeed these operators form a bounded $C_0$-semigroup on $\mathcal{D}_s$. See Section \ref{shifts} for a proof. On the other hand, $\mathcal{D}_s$ are not invariant under the vertical shifts when $\tau \in i\mathbb{R}$, as we see in the following example. \end{rems} \begin{exa} \label{eac} As stated in Example \ref{arc}, the function $\arccot$ is in $\mathcal{D}_s$ for all $s>-1$. Let \[ g(z)=\exp(\arccot(z)), \qquad \quad z\in \mathbb{C}_{+}, \] Since $|\Re \arccot(z)| \le \pi/2$, $\|g\|_\infty = \exp(\pi/2)$. For $s>-1$, we have \begin{align*} \|g\|_{\mathcal{D}_s^\infty}&\le \|g\|_\infty+\|g \cdot (\arccot)'\|_{\mathcal{V}_s} \le \exp(\pi/2)(1+\|\arccot\|_{\mathcal{D}_s}). \end{align*} Thus $g \in \mathcal{D}_s^\infty$ for all $s>-1$. However the boundary function of $g$ is not continuous at $z=\pm i$. Indeed, for a fixed $\epsilon>0$, \[ \arccot (i+i\epsilon)=\frac{1}{2i}\log\left(1+\frac{2}{\epsilon}\right), \] hence \[ g(i+i\epsilon)=\exp{(-i\log(1+2/\epsilon)^{1/2})} \] does not have a limit as $\epsilon\to 0+$. Note that $g(\epsilon + i)$ does not converge as $\epsilon\to0+$. This means that if $f(z):=g(z-i)$, then $f$ does not have a sectorial limit at 0 and therefore does not belong to $\mathcal{D}_s$ for any $s>-1$. Thus $\mathcal{D}_s$ and $\mathcal{D}_s^\infty$ are not invariant under vertical shifts. \end{exa} \subsection{Bernstein functions and $\mathcal{D}_s$} Recall that a holomorphic function $g :\mathbb{C}_+ \to \mathbb{C}_+$ is a \emph{Bernstein function} if it is of the form \begin{equation} \label{bsfn} g(z) = a + bz + \int_{(0,\infty)} (1-e^{-zs}) \, d\mu(s), \end{equation} where $a\ge 0$, $b\ge0$ and $\mu$ is a positive Borel measure on $(0,\infty)$ such that $\int_{(0,\infty)} \frac{s}{1+s} \,d\mu(s) < \infty$. The following properties of Bernstein functions $g$ will be used (these properties differ slightly from those used in \cite{BGTad}): \begin{enumerate} \item[(B1)] $g$ maps $\Sigma_\psi$ into $\Sigma_\psi$ for each $\psi \in [0,\pi/2]$, see \cite[Proposition 3.6]{Schill} or \cite[Proposition 2.1(1)]{BGTad}. \item[(B2)] $g$ is increasing on $(0,\infty)$. \item[(B3)] For all $z \in \mathbb{C}_+$, \[ g(\Re z) \le \Re g(z) \le |g(z)|, \qquad |g'(z)| \le g'(\Re z). \] Here the first inequality follows from taking the real parts in \eqref{bsfn}, and the second inequality is shown in \cite[Section 3.5, (B3)]{BGT}. \end{enumerate} Further information on Bernstein functions can be found in \cite{Schill}. \begin{lemma} \label{bsds} Let $g$ be a Bernstein function, $\lambda\in\mathbb{C}_+$, and \[ f(z;\lambda): =(\lambda+g(z))^{-1},\qquad z\in \mathbb{C}_{+}. \] Then $f(\cdot;\lambda) \in \mathcal{D}_s$ for $s>2$ and \[ \|f(\cdot;\lambda)\|_{\mathcal{D}_{s,0}} \le \frac{2^s}{(s-2)|\lambda|} . \] \end{lemma} \begin{proof} We have \[ f'(z;\lambda)=-\frac{g'(z)}{(\lambda+g(z))^2}, \] and then, for $\psi\in (-\pi/2,\pi/2)$, using Lemma \ref{trig}, (B2) and (B3), \begin{align*} \int_0^\infty |f'(\rho e^{i\psi};\lambda)|\,d\rho &=\int_0^\infty \frac{|g'(\rho e^{i\psi})|} {|\lambda+g(\rho e^{i\psi})|^2} \,d\rho\\ &\le \frac{1}{\cos^2((|\psi|+\pi/2)/2)} \int_0^\infty \frac{g'(\rho\cos\psi)} {(|\lambda|+|g(\rho e^{i\psi})|)^2} \,d\rho\\ &\le\frac{1}{\cos^2((|\psi|+\pi/2)/2)} \int_0^\infty \frac{g'(\rho\cos\psi)} {(|\lambda|+g(\rho\cos\psi))^2} \,d\rho\\ &\le \frac{1}{\cos^2((|\psi|+\pi/2)/2)\cos\psi} \int_0^\infty \frac{dt} {(|\lambda|+t)^2}\\ &= \frac{1}{\cos^2((|\psi|+\pi/2)/2)\cos\psi}\cdot \frac{1}{|\lambda|}. \end{align*} If $s>2$, then \begin{align*} |\lambda|\|f(\cdot;\lambda)\|_{\mathcal{D}_{s,0}} &=|\lambda|\int_{-\pi/2}^{\pi/2}\cos^s\psi \int_0^\infty |f'(\rho e^{i\psi};\lambda)|\,d\rho\,d\psi\\ &\le 2\int_0^{\pi/2} \frac{\cos^{s-1}\psi}{\cos^2((\psi+\pi/2)/2)}\,d\psi = 4\int_0^{\pi/2} \frac{\cos^{s-1}\psi}{1-\sin\psi} \,d\psi\\ &=4\int_0^1 (1+t)^{s-2}(1-t)^{s-3}\,dt < \frac{2^s}{s-2}. \qedhere \end{align*} \end{proof} \subsection{Algebras associated with $\mathcal D_s$} The spaces $\mathcal{D}_s, \, s>-1$, are not algebras, but there are some related algebras. Consider the Banach spaces $\mathcal{D}_s^\infty:= \mathcal{D}_s\cap H^\infty (\mathbb{C}_{+})$ equipped with the norm \[ \|f\|_{\mathcal{D}_s^\infty}:=\|f\|_\infty+\|f'\|_{\mathcal{V}_s}. \] Thus $\mathcal{D}_s^\infty$ is the space of bounded holomorphic functions on $\mathbb{C}_{+}$ such that \[ f(\infty):=\lim_{|z|\to\infty, z\in \Sigma_\psi}\,f(z) \] exists for every $\psi\in (0,\pi/2)$, and \[ \|f\|_{\mathcal{D}_s^\infty}:=\|f\|_{\infty}+\int_0^\infty \alpha^s \int_{-\infty}^\infty \frac{|f'(\alpha+i\beta)|}{(\alpha^2+\beta^2)^{(s+1)/2}}\,d\alpha\,d\beta < \infty. \] Then $(\mathcal{D}_s^\infty, \|\cdot\|_{\mathcal{D}_s})$ is a Banach algebra, and in particular \begin{equation}\label{algebra} \|fg\|_{\mathcal{D}_s^\infty}\le \|f\|_{\mathcal{D}_s^\infty} \|g\|_{\mathcal{D}_s^\infty},\quad f,g\in \mathcal{D}_s^\infty. \end{equation} By Proposition \ref{BDs}, $\mathcal{B}\overset{i}{\hookrightarrow} \mathcal{D}_s^\infty$ for $s>0$, and the embeddings are continuous. It follows from Lemma \ref{leminv} that \[ f\in \mathcal{D}_s^\infty \quad \text{if and only if}\quad \tilde{f}(z):=f(1/z)\in \mathcal{D}_s^\infty, \] and \[ \|f\|_{\mathcal{D}_s^\infty}=\|\tilde{f}\|_{\mathcal{D}_s^\infty}. \] Moreover, the spectrum of $f$ in $\mathcal{D}_s^\infty$ is the closure of the range of $f$. In particular, the spectral radius of $f$ is $\|f\|_\infty$. Now we consider the linear space \[ \mathcal D_\infty:=\bigcup_{s>-1} \mathcal{D}_s. \] We will show that $\mathcal D_\infty$ is an algebra, which opens the way to an operator functional calculus on $\mathcal D_\infty$. \begin{lemma}\label{Alg1} For $s,\sigma>-1$, let $f\in \mathcal{D}_s$ and $g\in \mathcal{D}_\sigma$. Then \[ h:=fg \in \mathcal{D}_{s+\sigma+1}, \] and \begin{equation} \label{fgest} \|h\|_{\mathcal{D}_{s+\sigma+1}} \le \left(2+\frac{2^s+2^\sigma}{\pi}\right)\|f\|_{\mathcal{D}_s}\,\|g\|_{\mathcal{D}_\sigma}. \end{equation} Hence, $\mathcal{D}_\infty$ is an algebra. \end{lemma} \begin{proof} By Corollary \ref{Cangle}, for $\rho>0$ and $|\varphi|< \pi/2$, we have \begin{align*} |f(\rho e^{i\varphi})| &\le \left(1+\frac{2^s}{\pi\cos^{s+1}\varphi}\right)\|f\|_{\mathcal{D}_s},\\ |g(\rho e^{i\varphi})| &\le \left(1+\frac{2^\sigma}{\pi\cos^{\sigma+1}\varphi}\right)\|g\|_{\mathcal{D}_\sigma}. \end{align*} Hence \begin{align*} \|h\|_{\mathcal{D}_{s+\sigma+1,0}} &= \int_{-\pi/2}^{\pi/2} \cos^{s+\sigma+1}\varphi \int_0^\infty \left| f'(\rho e^{i\varphi}) g(\rho e^{i\varphi}) + f(\rho e^{i\varphi}) g'(\rho e^{i\varphi})\right| \,d\rho\,d\varphi \\ &\le \left(1+\frac{2^\sigma}{\pi}\right)\|f\|_{\mathcal{D}_{s,0}}\,\|g\|_{\mathcal{D}_\sigma} +\left(1+\frac{2^s}{\pi}\right)\|f\|_{\mathcal{D}_s}\,\|g\|_{\mathcal{D}_{\sigma,0}}. \end{align*} This shows that $h \in \mathcal{D}_q$, and \eqref{fgest} follows easily. \end{proof} \subsection{Derivatives of functions in $\mathcal{D}_s$} This section further clarifies the behaviour of the derivatives of functions from $\mathcal D_s$, and Lemma \ref{D001} is of independent interest. Corollary \ref{D00n} will be used in Section \ref{AnalF}. Here $zf'$ denotes the function $z \mapsto zf'(z)$. \begin{lemma}\label{D001} Let $f\in \mathcal{D}_s$, $s>-1$. Then $zf'\in \mathcal{D}_{s+1}$, and there exists $C_{s}$ (independent of $f$) such that \begin{equation}\label{DerAnD} \|zf'(z)\|_{\mathcal{D}_{s+1}}\le {C}_{s}\|f\|_{\mathcal{D}_s}. \end{equation} and \begin{equation} \label{DerAnE} \|f(tz)-f(\tau z)\|_{\mathcal{D}_{s+1}} \le \frac{C_s \|f\|_{\mathcal{D}_s}}{\min\{t,\tau\}} |t-\tau|, \qquad t,\tau>0. \end{equation} \end{lemma} \begin{proof} Note that \[ \|zf'\|_{\mathcal{D}_{s+1}}\le \|f'\|_{\mathcal{V}_{s+1}}+\|zf''\|_{\mathcal{V}_{s+1}} \le \|f\|_{\mathcal{D}_{s}}+\|zf''\|_{\mathcal{V}_{s+1}}. \] So, for (\ref{DerAnD}), it suffices to consider $\|zf^{''}\|_{\mathcal V_{s+1}}$. The argument is similar to Example \ref{resd} and the proof of Proposition $\ref{Furt}$. By Corollary \ref{Repr}, for fixed $\sigma>s$, \[ f''(z)=-c_\sigma \int_{-\pi/2}^{\pi/2} \cos^\sigma \varphi \int_0^\infty \frac{\rho^{\sigma+1} f'(\rho e^{i\varphi})}{(z+\rho e^{-i\varphi})^{\sigma+3}}\,d\rho\,d\varphi,\quad z\in \mathbb{C}_{+}, \] where $c_\sigma=(\sigma+1)(\sigma+2)\frac{2^\sigma}{\pi}$. Then we have \begin{align*} \lefteqn{c_\sigma^{-1}\|zf''(z)\|_{\mathcal{V}_{s+1}}}\\ &=\int_{-\pi/2}^{\pi/2}\cos^{s+1}\psi \int_0^\infty r \left| \int_{-\pi/2}^{\pi/2} \cos^\sigma\varphi \int_0^\infty \frac{\rho^{\sigma+1}f'(\rho e^{i\varphi})}{(r e^{i\psi}+\rho e^{-i\varphi})^{\sigma+3}}\,d\rho\,d\varphi\right| \,dr\,d\psi\\ &\le \int_{-\pi/2}^{\pi/2} \cos^\sigma \varphi \int_{-\pi/2}^{\pi/2} \cos^{s+1}\psi \int_{0}^\infty\int_0^\infty \frac{t |f'(\rho e^{i\varphi})|\,dt\,d\rho}{(t^2+2t\cos(\varphi+\psi)+1)^{(\sigma+3)/2}} \,d\psi\,d\varphi \\ &= \int_{-\pi/2}^{\pi/2} \cos^{\sigma-s}\varphi \, G_{h,\beta,s+1}(\varphi) \int_0^\infty |f'(\rho e^{i\varphi})| \,d\rho\,d\varphi, \end{align*} where $h(t)=t+t^\sigma, \, t\in (0,1)$, $\beta = (\sigma+3)/2$ and $G_{h,\beta,s+1}(\varphi)$ is defined in \eqref{defG}, noting that $2\beta - (s+1)-2 = \sigma-s$. Now the estimate (\ref{DerAnD}) follows from Lemma \ref{D0012}(b). For \eqref{DerAnE}, let $g(z) = f(\tau z)-f(tz)$. Without loss, assume that $0 < t < \tau$. Then \[ g'(z) = \tau f'(\tau z) - t f'(tz) = \int_t^\tau \left(f'(rz) + rz f''(rz)\right) \, dr = \int_t^\tau \frac{d}{dz} (zf'(rz)) \,dr. \] Hence, by Fubini's theorem, \begin{align*} \|g\|_{\mathcal{D}_{s+1}} &\le \int_{\mathbb{C}_+} \int_t^\tau \frac{(\Re z)^{s+1}}{|z|^{s+2}} \left|\frac{d}{dz} (zf'(rz))\right|\,dr \,dS(z)\\ &= \int_t^\tau \frac{\|zf'(z)\|_{\mathcal{D}_{s+1}}}{r} \,dr \le \frac{C_s\|f\|_{\mathcal{D}_s}}{t} (\tau-t), \end{align*} since the $\mathcal{D}_{s+1}$-norm is invariant under the change of variable $z\mapsto rz$ (Lemma \ref{leminv}(iii)). \end{proof} The following corollary is easily proved by induction. \begin{cor} \label{D00n} If $f \in \mathcal{D}_\infty$ and $n \in \mathbb{N}$, then $z^nf^{(n)} \in \mathcal{D}_\infty$. \end{cor} \begin{remark} Lemma \ref{D001} is sharp in the sense that for any $s>0$ and $\sigma \in (-1,s+1)$, there exist functions $f \in \mathcal{D}_s$ for which $zf' \notin \mathcal{D}_\sigma$. For example, the function $f_\nu(z) := z^\nu e^{-z}$ has these properties if $\max\{0,\sigma-1\}<\nu< s$. This follows directly from Example \ref{dm}. \end{remark} \section{Hardy-Sobolev algebras on sectors}\label{hardy-sobolev} \subsection{$H^p$-spaces on the right half-plane and their norms} \label{hardy_rays} In this section and in Section \ref{hardy_sec} we will study the Hardy spaces $H^1(\Sigma_\psi)$ defined on sectors $\Sigma_\psi, \, \psi \in (0,\pi)$. The properties of such spaces are similar to the properties of the classical Hardy space $H^1(\mathbb{C}_+)$, though their theory seems to be more involved. The Hardy spaces $H^p(\Sigma_\psi)$ have been studied, mostly for $p > 1$, but the results are scattered around various places in the literature, which is often obscure, and some proofs contain rather complicated, incomplete or vague arguments. We propose a streamlined (and probably new) approach avoiding the use of Carleson measures or log-convexity, and we obtain a new result (Corollary \ref{HHH}) on the way. The case $p = 1$ does not require any significant adjustments as we illustrate below. Standard references for the theory of Hardy spaces on the right half-plane are \cite{Duren1} and \cite{Garnett}. We set out the situation when $\psi=\pi/2$ in this section, and the case of general $\psi$ in Section \ref{hardy_sec}. Although we are mainly interested in $H^1$-spaces, we present statements which are valid for $H^p$-spaces with $p \in [1, \infty)$, since the arguments are the same for all such $p$. Let $1\le p<\infty$. The classical Hardy space $H^p(\mathbb{C}_{+})$ in the right half-plane $\mathbb{C}_+$ is defined as \[ H^p(\mathbb{C}_{+})=\left \{g\in \mbox{Hol}(\mathbb{C}_{+}): \|g\|_{p}:=\sup_{\alpha>0}\, \left(\int_{-\infty}^\infty |g(\alpha+i\beta)|^p\,d\beta\right)^{1/p}<\infty \right \}. \] It is well-known that $\|\cdot\|_p$ is a norm on $H^p(\mathbb C_+)$ and $(H^p(\mathbb{C}_+),\|\cdot\|_{p})$ is a Banach space. Moreover, for almost every $t \in \mathbb R$ there exists a sectorial limit $g(it):=\lim_{z \to it} g(z)$ in $\mathbb{C}$. For every $g \in H^p(\mathbb{C}_+)$ one has $\lim_{\alpha \to 0} g(\alpha+i\cdot)= g(i\cdot)$ in $L^p(\mathbb R)$, and $\|g\|_{H^p(\mathbb{C}_+)} := \|g\|_p = \|g(i\cdot)\|_{L^p(\mathbb{R})}$. One may also consider the normed space $(H^{p}(\Sigma_{\pi/2}), \|\cdot\|_{H^p(\Sigma_{\pi/2})})$ as the space of all $g\in \operatorname{Hol}(\mathbb{C}_{+})$ such that \[\|g\|_{H^p(\Sigma_{\pi/2})}:=\sup_{|\varphi|<\pi/2}\, \left(\int_0^\infty \left(|g(te^{i\varphi})|^p+ |g(te^{-i\varphi})|^p\right)\,dt\right)^{1/p}<\infty. \] and $(H^{p}_{*}(\Sigma_{\pi/2}), \|\cdot\|_{H^p_{*}(\Sigma_{\pi/2})})$ as \[ \left \{g\in \operatorname{Hol}(\mathbb{C}_{+}): \|g\|_{H^p_*(\Sigma_{\pi/2})}:=\sup_{|\varphi|<\pi/2}\left(\int_0^\infty |g(te^{i\varphi})|^p\,dt\right)^{1/p} < \infty \right\}. \] It is clear that $H^p(\Sigma_{\pi/2})$ and $H^p_*(\Sigma_{\pi/2})$ coincide as vector spaces, and \[ \|g\|_{H^p_*(\Sigma_{\pi/2})} \le \|g\|_{H^p(\Sigma_{\pi/2})}\le 2^{1/p}\|g\|_{H^p_*(\Sigma_{\pi/2})}. \] \begin{lemma}\label{11L1A} Let $p\in [1,\infty)$. Then $H^{p}(\Sigma_{\pi/2})\subset H^{p}(\mathbb{C}_{+})$ and \begin{equation}\label{RelAB} \|g\|_{p}\le \|g\|_{H^{p}(\Sigma_{\pi/2})}, \qquad g\in H^{p}(\Sigma_{\pi/2}). \end{equation} \end{lemma} \begin{proof} Fix $p \in [1,\infty)$. For fixed $\gamma\in (1/2,1)$ define \[ g_\gamma(z):=(\gamma z^{\gamma-1})^{1/p}g(z^\gamma)\in \operatorname{Hol}(\Sigma_{\pi/(2\gamma)}). \] Note that \begin{equation}\label{Al11} \int_0^\infty |g_\gamma(te^{i\varphi})|^p\,dt= \int_0^\infty |g(te^{i\gamma\varphi})|^p\,dt, \end{equation} and \[ \int_{-\pi/2}^{\pi/2} \int_0^\infty |g_\gamma(te^{i\varphi})|^p\,dt \, d\varphi\le \frac{\pi}{2} \|g\|^p_{{H^{p}(\Sigma_{\pi/2})}}. \] By Fubini's theorem and H\"older's inequality there exist sequences $(t_{1,n})_{n\ge 1}$ and $(t_{2,n})_{n\ge 1}$ such that $0<t_{1,n}<t_{2,n}$, $t_{1,n}\to0$, $t_{2,n}\to \infty$ as $n\to\infty$, and \begin{align}\label{seq112} \lim_{n\to\infty}\, t_{1,n} \int_{-\pi/2}^{\pi/2} |g_\gamma (t_{1,n}e^{i\varphi})|\,d\varphi&=0, \\ \lim_{n\to\infty}\,\int_{-\pi/2}^{\pi/2} |g_\gamma (t_{2,n}e^{i\varphi})|\,d\varphi&=0. \label{seq113} \end{align} Let $\Omega_n:=\{z\in \mathbb{C}_{+}:\, t_{1,n}<|z|<t_{2,n}\}$. By Cauchy's formula, \begin{equation}\label{LL112} g_\gamma(z)=\frac{\alpha}{\pi i}\int_{\partial\Omega_n}\frac{g_\gamma(\lambda) }{(\lambda-z)(\lambda+\overline{z})} \,d\lambda,\quad z= \alpha+i\beta \in \Omega_n, \end{equation} for large $n$. Passing to the limit in \eqref{LL112} as $n \to \infty$, and using (\ref{Al11}), (\ref{seq112}) and (\ref{seq113}), we obtain the Poisson formula \begin{equation}\label{Puas2} g_\gamma(z)=\frac{\alpha}{\pi}\int_{-\infty}^\infty \frac{g_\gamma(it)}{(t-\beta)^2+\alpha^2}\,dt. \end{equation} Hence, by Young's inequality and (\ref{Al11}), for every $\alpha>0$, \[ \|g_\gamma(\alpha+i\cdot)\|_{L^p(\mathbb{R})}\le \|g_\gamma(i\cdot)\|_{L^p(\mathbb R)}\le \|g\|_{H^{p}(\Sigma_{\pi/2})}. \] Letting $\gamma\to 1$, Fatou's Lemma implies (\ref{RelAB}). \end{proof} \begin{lemma}\label{11L2A} Let $p\in [1,\infty)$. Then $H^{p}(\mathbb{C}_{+})\subset H^{p}(\Sigma_{\pi/2})$ and \begin{equation}\label{RelABC} \|g\|_{H^{p}(\Sigma_{\pi/2})}\le \|g\|_{p}, \qquad g\in H^{p}(\mathbb{C}_{+}). \end{equation} \end{lemma} \begin{proof} First let $p=2$ so that $g\in H^2(\mathbb{C}_{+})$. Then, by \cite[Ch.VIII, p.508]{Djrbashian}, there exists $f\in L^1(\mathbb{R}, e^{\pi |t|}\,dt)$ such that $f \ge 0$ on $\mathbb R$, and for all $|\varphi|\le \pi/2$, \[ \int_0^\infty |g(te^{i\varphi})|^2\,dt= \int_{-\infty}^\infty e^{2\varphi t} f(t)\,dt. \] Hence \begin{align*} \int_0^\infty \bigl(|g(te^{i\varphi})|^2+|g(t e^{-i\varphi})|^2\bigr)\,dt&= 2\int_{-\infty}^\infty \cosh(2\varphi t)f(t)\,dt \\ \le 2\int_{-\infty}^\infty \cosh(\pi t)f(t)\,dt&=\int_0^\infty \bigl(|g(it)|^2+|g(-it)|^2 \bigr)\,dt,\notag \end{align*} and (\ref{RelAB}) holds for $p=2$. Let $p\in [1,\infty), \, p\not=2$, be fixed, and $g\in H^p(\mathbb{C}_{+})$. Then $g(z)=B(z)\tilde g(z)$, $z \in \mathbb C_+$, where $B$ is the Blaschke product associated with $g$ and $\tilde g$ has no zeros in $\mathbb{C}_+$. Then there is a well-defined holomorphic function $g_p(z)=[\tilde g (z)]^{p/2}$ on $\mathbb{C}_+$ and $g_p \in H^2(\mathbb{C}_{+})$. Using (\ref{RelAB}) for $p=2$, for all $|\varphi|<\pi/2$, we have \begin{gather} \int_0^\infty \bigl(|g(te^{i\varphi})|^p+|g(te^{-i\varphi})|^p\bigr)\,dt = \int_0^\infty \bigl(|g_p(te^{i\varphi})|^2+|g_p(te^{-i\varphi})|^2\bigr)\,dt\label{RRD} \\ \le \int_{-\infty}^\infty |g_p(it)|^2\,dt=\int_{-\infty}^\infty |g(it)|^p\,dt=\|g\|_p^p, \notag \end{gather} and (\ref{RelABC}) follows. \end{proof} Lemmas \ref{11L1A} and \ref{11L2A} imply the next statement. \begin{cor}\label{HHH} Let $p\in [1,\infty)$. Then $H^{p}(\Sigma_{\pi/2})=H^{p}(\mathbb{C}_+)$, and for every $g\in H^{p}(\mathbb{C}_{+})$, \begin{equation}\label{n_equality} \|g\|_{H^{p}(\Sigma_{\pi/2})}=\|g\|_{H^{p}(\mathbb{C}_+)}, \end{equation} and \begin{equation}\label{Lux11} \|g\|_{H^{p}_{*}(\Sigma_{\pi/2})}\le \|g\|_{H^p(\mathbb{C}_+)}\le 2^{1/p}\|g\|_{H^{p}_{*}(\Sigma_{\pi/2})}. \end{equation} \end{cor} Note that the two-sided estimate (\ref{Lux11}) was proved in \cite{Sedl} and \cite{Luxemburg} in a more complicated way (see also \cite{Akopjan}, \cite{dil}, \cite{Martirosian}, \cite{Kev1}, \cite{Kev2}). The coincidence of norms in \eqref{n_equality} seems not to have been noted before. It appears to be quite useful, as we will see in the proof of Corollary \ref{Zb} below. \begin{remark}\label{Lux} The two-sided estimate (\ref{Lux11}) is sharp (and cannot be improved). Indeed, let $p\in [1,\infty)$ and let \[ f_k(z):=\frac{1}{\pi^{1/p}(z+1+ik)^{2/p}},\qquad k\in\mathbb{N}. \] Then for all $k$, we have $\|f_k\|_{H^p(\mathbb{C}_{+})}=1$ and, by direct estimates, \begin{align*} &\left( \|f_k\|_{H^p_{*}(\Sigma_{\pi/2})} \right)^p=\frac{1}{\pi}\int_0^\infty |f_k(te^{-i\pi/2})|^p\,dt\\ &=\frac{1}{\pi}\int_0^\infty \frac{dt}{(t-k)^2+1}\,dt =\frac{1}{2}+\frac{\arctan k}{\pi}. \end{align*} [In fact, for all $f \in \mathcal{H}^p_{*}(\Sigma_{\pi/2})$, $\|f\|_{H^p_{*}(\Sigma_{\pi/2})}$ is attained at the boundary of $\Sigma_{\psi/2}$.] Thus, \[ \|f_0\|_{H^p(\mathbb{C}_{+})}=2^{1/p}\|f_0\|_{H^p_{*}(\Sigma_{\pi/2})} \qquad \text{and}\qquad \lim_{k\to\infty}\,\frac{\|f_k\|_{H^p_{*}(\Sigma_{\pi/2})}}{\|f_k\|_{H^p(\mathbb{C}_{+})}}=1. \] \end{remark} \begin{cor}\label{Zb} Let $g \in H^{p}(\Sigma_{\pi/2})$, $p\in [1,\infty)$. Then there exist $g(\pm it):=\lim_{\varphi \to \pm \pi/2} g(te^{\pm i\varphi})$ for a.e.\ $t \in \mathbb R_+$, $g(\pm i \cdot) \in L^p(\mathbb R_+)$, and \begin{equation}\label{DH11} \lim_{\varphi\to\pm \pi/2}\, \int_0^\infty |g(te^{i\varphi})-g(\pm it)|^p\,dt=0. \end{equation} As a consequence, for every $g \in H^p(\Sigma_{\pi/2})$, \begin{equation}\label{norm_h} \|g\|_{H^p(\Sigma_{\pi/2})}=\|g(i\cdot)\|_{L^p(\mathbb R)}. \end{equation} \end{cor} \begin{proof} Let $p \in [1,\infty)$ be fixed. By Corollary \ref{HHH}, it suffices to prove \eqref{DH11} for $g\in H^p(\mathbb{C}_{+})$. If $g\in H^p(\mathbb{C}_{+})$, then as recalled above, for almost all $t \in \mathbb R$ there exists a sectorial limit $g(it):=\lim_{z \to it} g(z)$, and $g(i\cdot) \in L^p(\mathbb{R})$. Therefore, we also have $\lim_{\varphi \to \pm \pi/2} g(te^{\pm i\varphi})=g(\pm it)$ for almost all $t$. Using this and Fatou's Lemma, we infer from (\ref{RRD}) that \begin{align*} \limsup_{\varphi\to \pi/2}\,&\int_0^\infty \bigl(|g(te^{i\varphi})|^p+|g(te^{-i\varphi})|^p \bigr)\,dt\le \int_{-\infty}^\infty |g(it)|^p\,dt\\ &\le \liminf_{\varphi\to\pi/2}\,\int_0^\infty \bigl(|g(te^{i\varphi})|^p+|g(te^{-i\varphi})|^p\bigr)\,dt, \end{align*} and hence \begin{equation}\label{Duren} \lim_{\varphi\to\pi/2}\,\int_0^\infty \bigl(|g(te^{i\varphi})|^p+|g(te^{-i\varphi})|^p \bigr)\,dt= \int_{-\infty}^\infty |g(it)|^p\,dt. \end{equation} Then, by \eqref{Duren} and Lemma \ref{duren21}, using once again the pointwise a.e.\ convergence of $g(te^{\pm i\varphi})$ to $g(\pm it)$ as $\varphi\to\pm \pi/2$, we obtain \eqref{DH11}. Since $\|g\|_{H^p(\mathbb{C}_+)}=\|g(i\cdot)\|_{L^p(\mathbb R)},$ we get \eqref{norm_h} as well. \end{proof} For (formally) more general versions of \eqref{n_equality} and \eqref{Lux11} see \eqref{equivalent} and \eqref{iso} below. \subsection{The spaces $H^1(\Sigma_\psi)$}\label{hardy_sec} Now using the results of Section \ref{hardy_rays} for $\psi=\pi/2$, we develop basic properties of $H^1(\Sigma_\psi)$ for any $\psi \in (0,\pi)$. Define the Hardy space $H^1(\Sigma_\psi)$ on the sector $\Sigma_\psi$ to be the space of all functions $f \in \operatorname{Hol}(\Sigma_\psi)$ such that \begin{equation}\label{CC2} \|f\|_{H^1(\Sigma_\psi)} := \sup_{|\varphi|<\psi} \int_0^\infty \bigl(|f(te^{i\varphi})|+ |f(te^{-i\varphi})| \bigr) \,dt <\infty. \end{equation} We shall also consider a non-symmetric version of $H^1(\Sigma_\psi)$, defined as \begin{equation*} H^1_{*}(\Sigma_{\psi}):=\left\{f \in \operatorname{Hol}\, (\Sigma_{\psi}): \|f\|_{H^1_{*}(\Sigma_{\psi})}:=\sup_{|\varphi|< \psi}\,\int_0^\infty |f(t e^{i\varphi})|\,dt<\infty\right\}. \end{equation*} \begin{thm}\label{hardy0} Let $\psi, \psi_1, \psi_2 \in (0,\pi)$. \begin{enumerate}[\rm(i)] \item $f \in H^1 (\Sigma_{\psi})$ if and only if $f \in H^1_*(\Sigma_\psi)$, and then \begin{equation}\label{equivalent} 2^{-1} \|f\|_{H^1(\Sigma_{\psi})} \le \|f\|_{H^1_*(\Sigma_{\psi})}\le \|f\|_{H^1(\Sigma_{\psi})}. \end{equation} \item For any $\psi_1,\psi_2 \in (0,\pi)$, the map \begin{eqnarray*} H^1(\Sigma_{\psi_1}) &\to& H^1(\Sigma_{\psi_2})\\ f(z) &\mapsto &\frac{\psi_1}{\psi_2} z^{(\psi_1/\psi_2) -1}f(z^{\psi_1/\psi_2}) \end{eqnarray*} is an isometric isomorphism of $H^1(\Sigma_{\psi_1})$ onto $H^1(\Sigma_{\psi_2})$, and of $H^1_*(\Sigma_{\psi_1})$ onto $H^1_*(\Sigma_{\psi_2})$. \item If $f \in H^1(\Sigma_\psi)$, then the limits $f(re^{\pm i\psi}):=\lim_{\varphi \to \pm \psi} f(re^{i\varphi})$ exist a.e., and in the $L^1$-sense with respect to $r$. Moreover \begin{equation}\label{Boch1} f(z)= \frac{1}{2\pi i} \int_{\partial\Sigma_{\psi}}\,\frac{f(\lambda)}{\lambda-z} \,d\lambda, \quad z\in \Sigma_\psi, \end{equation} where the integral is taken downwards. \item If $f \in H^1(\Sigma_\psi)$, then \begin{align \|f\|_{H^1(\Sigma_\psi)}&=\int_{0}^{\infty}\bigl(|f(te^{i\psi})|+ |f(te^{-i\psi})| \bigr) \,dt. \label{iso} \end{align} \item $H^1(\Sigma_{\psi})$ and $H^1_*(\Sigma_{\psi})$ are Banach spaces. \end{enumerate} \end{thm} \begin{proof} The proof of (i) is clear, and (ii) is a direct verification. For $\psi = \pi/2$, the statements (iii) and (iv), excluding \eqref{Boch1}, were proved in Corollaries \ref{HHH} and \ref{Zb}, and (v) is well-known. Then the general cases are reduced to the case when $\psi=\pi/2$, by means of (ii). The Cauchy formula \eqref{Boch1} is well-known for $\psi=\pi/2$ (see, for example, \cite[Theorem 11.8]{Duren1}). For general $\psi$, we may argue similarly to the proof of Lemma \ref{11L1A}, as follows. Since $f \in H^1(\Sigma_{\psi})$, \[ \int_0^\infty \int_{-\psi}^\psi |f(te^{i\varphi})| \,d\varphi\,dt < \infty. \] Hence there exist sequences $(t_{1,n})_{n\ge1}$ and $(t_{2,n})_{n\ge1}$ such that $0 < t_{1,n} < t_{2,n}$, $t_{1,n} \to 0, \, t_{2,n} \to \infty$ as $n\to\infty$, and \[ \lim_{n\to \infty} \int_{-\psi}^\psi t_{1,n} |f(t_{1,n} e^{i\varphi})| \,d\varphi = \lim_{n\to\infty} \int_{-\psi}^\psi |f(t_{2,n} e^{i\varphi})| \,d\varphi = 0. \] By applying Cauchy's theorem around the boundary of \[ \left\{z \in \partial \Sigma_{\psi - n^{-1}} : t_{1,n} < |z| < t_{2,n} \right\} \] for large $n$, and taking the limit, we obtain \eqref{Boch1}. \end{proof} \begin{rem} In addition to \eqref{iso}, it is also possible to prove that \[ \|f\|_{H^1_{*}(\Sigma_\psi)} = \max \left(\int_{0}^{\infty}|f(te^{-i\psi})|\, dt,\int_{0}^{\infty}|f(te^{i\psi})|\,dt\right). \] This requires additional techniques, and it is not used in this paper. \end{rem} \subsection{Functions with derivatives in $H^1(\Sigma_\psi)$} For $\psi\in (0,\pi)$, let us introduce the space \[\mathcal{H}_\psi:=\left\{ f \in \operatorname{Hol}(\Sigma_\psi): f' \in H^1 (\Sigma_{\psi})\right\}.\] In view of Corollary \ref{HHH}, \begin{equation} \label {HH11} \mathcal{H}_{\pi/2} = H^{1,1}(\mathbb{C}_+):=\{f \in \operatorname{Hol}(\mathbb{C}_+): f'\in H^1(\mathbb{C}_+)\}, \end{equation} and we may sometimes use the notation $H^{1,1}(\mathbb{C}_+)$ instead of $\mathcal{H}_{\pi/2}$. Such function spaces are often called Hardy-Sobolev spaces, and we shall also use this terminology sporadically. Spaces more general than $\mathcal{H}_\psi$ appear in \cite{Domelevo}. Namely, for $f \in \operatorname{Hol}(\Sigma_\psi)$ it was required in \cite{Domelevo} that the boundary values of $f$ exist and belong to a Besov space $B^s_{\infty, 1}, \, s>0$. One can develop a similar approach to those spaces, but we do not see much advantage in such generality within the present context. \begin{thm}\label{hardy2} Let $f \in \mathcal{H}_\psi,\, \psi\in (0,\pi)$. \begin{enumerate}[\rm(i)] \item The function $f$ extends to a continuous bounded function on $\overline{\Sigma}_{\psi}$. \item The limit \[ f(\infty):=\lim_{|z|\to \infty, z \in \Sigma_\psi}\,f(z) \] exists. \item One has \begin{equation}\label{bound_for_sup} \|f\|_{H^\infty(\Sigma_\psi)} \le |f(\infty)| + \|f'\|_{H^1(\Sigma_\psi)}. \end{equation} In particular, the evaluation functionals $\delta_z,\, z\in\mathbb{C}_+$, are continuous on $\mathcal{H}_\psi$. \end{enumerate} \end{thm} \begin{proof} Let $\psi = \pi/2$. Since $\mathcal H_{\pi/2}=H^{1,1}(\mathbb C_+) \subset \mathcal{B}$ by \eqref{HH11} and \cite[Proposition 2.4]{BGT}, the statement (i) follows from \cite[Proposition 2.2(iv)]{BGT}, and (ii) follows from \cite[Proposition 2.4]{BGT}. In the general case, the map $f(z) \mapsto f(z^{2\psi/\pi})$ is an isomorphism of $\mathcal{H}_\psi$ onto $H^{1,1}(\mathbb{C}_+)$, by Theorem \ref{hardy0}(ii), so (i) and (ii) hold for $\mathcal{H}_\psi$. The statement (iii) is easily seen. \end{proof} It follows from Theorem \ref{hardy2} that $\mathcal{H}_\psi$ is an algebra for every $\psi\in (0,\pi)$. We define a norm on $\mathcal{H}_\psi$ by \begin{equation} \|f \|_{\mathcal{H}_\psi}:=\|f\|_{H^{\infty}(\Sigma_\psi)}+\|f'\|_{H^1(\Sigma_\psi)}, \qquad f \in \mathcal{H}_\psi. \end{equation} This is easily seen to be an algebra norm. Theorem \ref{hardy2}(iii) shows that \begin{equation} \label{hnorms} \|f\|'_{\mathcal{H}_\psi} := |f(\infty)| + \|f'\|_{H^1(\Sigma_\psi)} \end{equation} is an equivalent norm on $\mathcal{H}^1(\Sigma_\psi)$. The following lemma is simple, but crucial for our theory. The completeness of the norm is a standard fact, the scale-invariance is trivial, and the final isomorphism follows from Theorem \ref{hardy0}(ii). \begin{lemma}\label{simple} For every $\psi\in (0,\pi)$, the space $(\mathcal{H}_\psi, \|\cdot\|_{\mathcal{H}_\psi})$ is a Banach algebra. For $t>0$, the map $f(z) \mapsto f(tz)$ is an isometric algebra isomorphism on $\mathcal{H}_\psi$. Moreover, for any $\psi_1,\psi_2 \in (0,\pi)$, the map \begin{eqnarray*} \mathcal{H}_{\psi_1}&\to& \mathcal{H}_{\psi_2}\\ f(z)&\mapsto& f(z^{\psi_1/\psi_2}) \end{eqnarray*} is an isometric algebra isomorphism. \end{lemma} We now give some examples of functions in $\mathcal{H}_\psi$ which will play important roles in subsequent sections of this paper. The first example is of similar type to Example \ref{resd}. \begin{exas} \label{hexs} Let $\psi \in (0,\pi/2)$, and $\lambda = |\lambda| e^{i\varphi} \in \Sigma_{\pi-\psi}$. \begin{enumerate}[\rm1.] \item Let $\gamma >0$, and consider the function $r_\lambda^\gamma(z) = (\lambda+z)^{-\gamma}, \,z \in \Sigma_{\psi}$. Then $r_{\lambda}^{\gamma}\in \mathcal{H}_\psi$ and \begin{equation} \label{ff} \|r_{\lambda}^{\gamma}\|_{\mathcal{H}_\psi} = \int_{\partial\Sigma_\psi} \frac{\gamma |dz|}{|z+\lambda|^{\gamma+1}} \le \frac{2}{\sin^{\gamma+1} ((\varphi-\psi)/2) \,|\lambda|^{\gamma}}, \end{equation} where we have used Lemma \ref{trig}. Thus $r_\lambda^\gamma \in \mathcal{H}_\psi$, and there exists $C_{\varphi,\psi,\nu}$ such that \begin{equation} \label{rtl2} \|r_\lambda^\gamma\|'_{\mathcal{H}_\psi} \le \frac{C_{\varphi,\psi,\gamma}}{|\lambda|^\nu}, \qquad \lambda \in \Sigma_\varphi, \, \varphi < \pi - \psi. \end{equation} In particular, if $\gamma=1$, then \begin{equation}\label{res_sigma_cal_intro} \|r_{\lambda}\|_{\mathcal{H}_\psi} \le \frac{2}{\sin^{2} ((\varphi-\psi)/2) \,|\lambda|}, \qquad \lambda \in \Sigma_{\pi-\varphi}. \end{equation} This property will be important for the proof of Theorem \ref{sigma_cal_intro} (see Theorem \ref{sigma_cal}). A more general estimate will be given in Corollary \ref{BF}. \item Let $\gamma \in (0,\pi/(2\psi)$ and $e_{\gamma,\lambda}(z) := e^{-\lambda z^\gamma}, \,z \in \Sigma_\psi$. Then $e_{\gamma,\lambda} \in \mathcal{H}_\psi$ and \[ \|e_{\gamma,\lambda}\|'_{\mathcal{H}_{\psi}} = \|e_{1,\lambda} \|'_{\mathcal{H}_{\gamma\psi}} \le \int_{\partial\Sigma_\psi} |\lambda| e^{-\Re\lambda z} \, |dz| \le \frac{1}{\cos(\varphi+\psi)} + \frac{1}{\cos (\varphi-\psi)}. \] \end{enumerate} \end{exas} More examples can be found in Sections \ref{handd} and \ref{bandh}. In particular, Lemma \ref{Dh1} shows that the restriction of any function in $\mathcal{D}_\infty$ to $\Sigma_\psi, \, \psi \in (0,\pi/2)$, belongs to $\mathcal{H}_\psi$. The following lemma is a result of Fatou type closely related to Corollary \ref{Fatou}. \begin{lemma}\label{FatouH} Let $\psi \in (0,\pi)$ and $(f_k)_{k=1}^\infty \subset \mathcal{H}_\psi$ be such that $\sup_{k \ge 1} \|f_k\|_{\mathcal{H}_\psi}<\infty$ and $f(z)=\lim_{k \to \infty} f_k(z)$ exists for all $z \in \mathbb C_+$. Then $f \in \mathcal{H}_\psi$. \end{lemma} \begin{proof} By Theorem \ref{hardy2}, the functions $\{f_k: k\ge 1\}$ are uniformly bounded on $\Sigma_\psi$. By Vitali's theorem, $f$ is holomorphic, and $f_k'(z) \to f'(z)$ as $k \to \infty$ for each $z \in \Sigma_\psi$. By Fatou's Lemma, for $|\varphi| < \psi$, \[ \int_0^\infty |f'(te^{i\varphi})| \,dt \le \liminf_{k\to\infty} \int_0^\infty |f_k'(te^{i\varphi})| \,dt \le \sup_{k\ge1} \|f_k\|_{\mathcal{H}_\psi}. \] Thus $f \in \mathcal{H}_\psi$. \end{proof} \subsection{The spaces $H^{1,1}(\mathbb{C}_+)$ and $\mathcal{L} L^1$} In \cite[Proposition 2.4]{BGT}, we showed that $\mathcal{H}_{\pi/2} \overset{i}{\hookrightarrow} \mathcal{B}$. We will now show a stronger result that $H^{1,1}(\mathbb{C}_+) \overset{i}{\hookrightarrow} \mathcal{L} L^1 + \mathbb{C} \subset \mathcal{LM}$, where $(\mathcal{LM},\|\cdot\|)_{\mathrm{HP}}$ is the Hille-Phillips algebra as in Section \ref{prelims}. In particular, it shows that the Laplace transforms of singular measures on $(0,\infty)$ are not in $H^{1,1}(\mathbb{C}_+)$, which may be of interest. \begin{thm}\label{ACM} If $f\in H^{1,1}(\mathbb{C}_{+})$ then there exists $g \in L^1(\mathbb{R}_+)$ such that $f = f(\infty) + \mathcal{L} g$. Moreover, there is an absolute constant $C$ such that \begin{equation} \label{AW1} \|f\|_{\mathrm{HP}} \le C\|f\|_{H^{1,1}(\mathbb{C}_+)}. \end{equation} \end{thm} \begin{proof} Let $f\in H^{1,1}(\mathbb{C}_{+})$, and, for $n\in\mathbb{N}$, let \begin{align*} f_n(z) &:=f(z)-f(z+n),\qquad z \in \mathbb{C}_+, \\ g_n(t) &:= -\frac{2}{\pi} t(1-e^{-nt})\int_0^\infty \alpha e^{-\alpha t} \int_{-\infty}^\infty f'(\alpha+i\beta) e^{-i\beta t}\,d\beta\,d\alpha, \quad t>0. \end{align*} Then $f_n \in H^{1,1}(\mathbb{C}_+)$, and $\|f_n\|_{H^{1,1}(\mathbb{C}_+)} \le 2\|f\|_{H^{1,1}(\mathbb{C}_+)}$. Moreover, \begin{align*} |g_n(t)| &\le \frac{2}{\pi}t(1-e^{-nt})\int_0^\infty \alpha e^{-\alpha t} \int_{-\infty}^\infty |f'(\alpha+i\beta)|\,d\beta\,d\alpha\\ &\le \frac{2}{\pi} \|f'\|_{H^1(\mathbb{C}_{+})}t(1-e^{-nt})\int_0^\infty \alpha e^{-\alpha t} \,d\alpha= \frac{2}{\pi}\|f'\|_{H^1(\mathbb{C}_{+})}\frac{(1-e^{-nt})}{t}\\ &\le \frac{2n}{\pi} \|f'\|_{H^1(\mathbb{C}_{+})}. \end{align*} By the reproducing formula for $\mathcal{B}$ (see Remark \ref{Brep} or \cite[Proposition 2.20]{BGT}) and Fubini's theorem, \begin{align*} f_n(z)=&-\frac{2}{\pi}\int_0^\infty \alpha\int_{-\infty}^\infty f'(\alpha+i\beta) \left(\frac{1}{(z+\alpha-i\beta)^{2}}- \frac{1}{(z+n+\alpha-i\beta)^{2}}\right) d\beta\,d\alpha\\ =&-\frac{2}{\pi} \int_0^\infty \alpha \int_{-\infty}^\infty f'(\alpha+i\beta) \left(\int_0^\infty e^{-(z+\alpha-i\beta)t} t(1-e^{-nt})\,dt\right)\,d\beta\,d\alpha\\ =&-\frac{2}{\pi}\int_0^\infty e^{-zt} g_n(t)\,dt,\quad z\in \mathbb{C}_{+}. \end{align*} It follows that $(\pi/2) f_n'$ is the Laplace transform of $t g_n(t)$, and then $tg_n(t)$ is the inverse Fourier transform of $(\pi/2)f_n'(i\cdot)\in L^1(\mathbb{R})$. By Hardy's inequality in the form of \cite[p.198]{Duren1}, \begin{equation}\label{H11} \int_0^\infty |g_n(t)|\,dt =\int_0^\infty \frac{|tg_n(t)|}{t}\,dt \le \frac{\pi}{4}\|f_n'\|_{H^1(\mathbb{C}_{+})} \le \frac{\pi}{2}\|f'\|_{H^1(\mathbb{C}_{+})}. \end{equation} Moreover, \[ f(z)=f(\infty)+\lim_{n\to\infty} f_n(z),\qquad z\in \overline{\mathbb{C}}_{+}, \] and then by \cite[Theorem 1.9.2]{Rud} we obtain that $f\in \mathcal{LM}$ and (\ref{AW1}) holds. Now let \[ u(z) := f'(z+1) -\int_0^\infty e^{-zt}e^{-t}t \, d\mu(t),\qquad z\in \mathbb{C}_{+}. \] Since $f\in H^\infty(\mathbb{C}_{+})$ and $f'\in H^1(\mathbb{C}_{+})$, we have that $u\in H^1(\mathbb{C}_+) \cap H^\infty(\mathbb{C}_+) \subset H^2(\mathbb{C}_{+})$. Hence $u = \mathcal{L} h$ for some $h \in L^2(\mathbb{R}_+)$, so \[ u(z) = \int_0^\infty e^{-zt} h(t) \, dt, \qquad z \in \mathbb{C}_+. \] From the uniqueness properties of Laplace transforms it follows that \[ -e^{-t}t\,\mu(dt)=h(t)\,dt. \] Thus $\mu$ is absolutely continuous on $(0,\infty)$, with Radon-Nikodym derivative $g(t)$ on $(0,\infty)$. Since $\mu$ is a bounded measure, $g \in L^1(\mathbb{R}_+)$, and \[ \mu(dt)=f(\infty)\delta_0 +g(t)\,dt. \] Hence $f = f(\infty) + \mathcal{L} g$. \end{proof} \subsection{The spaces $\mathcal{H}_\psi$ and $\mathcal{D}_s$} \label{handd} Since $H^{1,1}(\mathbb{C}_+) = \mathcal{H}_{\pi/2}$ (by Corollary \ref{HHH}), we have shown in Theorem \ref{ACM} that \[ \mathcal{H}_{\pi/2} \overset{i}{\hookrightarrow} \mathcal{L} L^1 + \mathbb{C} \subset \mathcal{LM} \overset{i}{\hookrightarrow} \mathcal{B} \overset{i}{\hookrightarrow} \mathcal{D}_s \] if $s >0$. In the next lemma, we show that, for all $s>-1$, $\mathcal{H}_{\pi/2} \overset{i}{\hookrightarrow} \mathcal{D}_s$ and $\mathcal D_s \overset{r}{\hookrightarrow} \mathcal{H}_\psi$ for every $\psi\in (0,\pi/2)$. Moreover it follows that $\mathcal{D}_s^\infty \overset{r}{\hookrightarrow} \mathcal{H}_\psi$. \begin{lemma}\label{Dh1} \begin{enumerate}[\rm(i)] \item If $f\in \mathcal{H}_{\pi/2}$, then $f\in \mathcal{D}_s^\infty$ for every $s>-1$, and \begin{align}\label{embedh} \|f\|_{\mathcal D_{s,0}} &\le B\left(\frac{s+1}{2},\frac{1}{2}\right)\|f'\|_{H^1(\mathbb{C}_+)}, \\ \|f\|_{\mathcal D_{s}^\infty } &\le \max \left\{1, B\left(\frac{s+1}{2},\frac{1}{2}\right)\right\}\|f\|_{\mathcal{H}_{\pi/2}}. \notag \end{align} \item If $f\in \mathcal{D}_s$, $s>-1$, then $f \in \mathcal{H}_\psi$ for every $\psi \in (0,\pi/2)$, and \begin{equation}\label{Haab} \|f'\|_{H^1(\Sigma_\psi)}\le \frac{2^{s}}{\pi\cos^{s+2}(\psi/2+\pi/4)}\|f\|_{\mathcal{D}_{s,0}}. \end{equation} \end{enumerate} \noindent Thus, for all $\psi \in (0,\pi/2)$ and $s>-1$, there are natural continuous embeddings \[ \mathcal{H}_{\pi/2} \overset{i}{\hookrightarrow} \mathcal D_s \overset{r}{\hookrightarrow} \mathcal{H}_\psi. \] \end{lemma} Note that the estimates \eqref{rtl2} and \eqref{embedh} for functions $r_\lambda^\gamma$ reproduce the estimate \eqref{rtl}, with different constants. \begin{proof} Let $s >-1$ be fixed, and let $f \in \mathcal{H}_{\pi/2}$. Using \eqref{equivalent}, we have \begin{align*} \|f\|_{\mathcal D_{s,0}}&= \int_{-\pi/2}^{\pi/2}\cos^s\varphi \int_0^\infty |f'(te^{i\varphi})|\,dt\,d\varphi \\ &\le \left(\int_{-\pi/2}^{\pi/2}\cos^s\varphi\,d\varphi\right) \sup_{|\varphi|\le \pi/2}\int_0^\infty |f'(te^{i\varphi})|\,dt \\ &= B\left(\frac{s+1}{2},\frac{1}{2}\right)\|f'\|_{H^{1}(\mathbb{C}_+)}, \end{align*} and (i) follows. To prove (ii), note that if $f \in \mathcal D_s$, then by Corollary \ref{Repr}, \begin{equation*} f'(z)=\frac{(s+1)2^s}{\pi}\int_0^\infty \alpha^s\int_{-\infty}^\infty \frac{f'(\alpha+i\beta)}{(z+\alpha-i\beta)^{s+2}}\,d\beta\,d\alpha, \quad z\in \mathbb{C}_{+}. \end{equation*} Hence using \eqref{Fc}, for every $\psi \in (0,\pi/2)$, we obtain \begin{align*} \lefteqn {\hskip-10pt 2^{-s}\int_0^\infty |f'(te^{\pm i\psi})|\,dt}\\ &\le \frac{(s+1)}{\pi}\int_0^\infty \int_0^\infty \alpha^s\int_{-\infty}^\infty \frac{|f'(\alpha+i\beta)|}{|te^{\pm i\psi}+\alpha-i\beta|^{s+2}}\,d\beta\,d\alpha\,dt\\ &\le \frac{s+1}{\pi\cos^{s+2}(\psi/2+\pi/4)} \int_0^\infty \alpha^s\int_{-\infty}^\infty \left( \int_0^\infty \frac{|f'(\alpha+i\beta)|}{(t+|\alpha+i\beta|)^{s+2}}\right) \,dt\,d\beta\,d\alpha\\ &= \frac{1}{\pi\cos^{s+2}(\psi/2+\pi/4)} \int_0^\infty \alpha^s\int_{-\infty}^\infty \frac{|f'(\alpha+i\beta)|}{|\alpha+i\beta|^{s+1}}\,d\beta\,d\alpha\\ &=\frac{1}{\pi\cos^{s+2}(\psi/2+\pi/4)}\|f'\|_{\mathcal{V}_s}, \end{align*} and \eqref{Haab} follows. \end{proof} For a function $f \in \operatorname{Hol}(\Sigma_\psi)$, $\gamma>0$ and $0<\varphi\le \min\{\pi,\psi/\gamma\}$ , define $f_\gamma \in \operatorname{Hol}(\Sigma_{\varphi})$ by \[ f_\gamma(z) = f(z^\gamma), \quad z \in \Sigma_{\varphi}. \] \begin{cor} \label{gggC} Let $f \in \mathcal{D}_s, \, s>-1$, and let $\gamma \in (0,1)$. Then $f_\gamma \in \mathcal{D}_\sigma^\infty \cap \mathcal{H}_{\pi/2}$ for all $\sigma>-1$. Moreover, for each $s>-1$ and $\sigma>-1$, there exist constants $C_{s,\sigma,\gamma}$ and $\tilde C_{s,\gamma}$ such that \[ \|f_\gamma\|_{\mathcal{D}_\sigma^\infty} \le C_{s,\sigma,\gamma} \|f\|_{\mathcal{D}_s} \quad \text{and} \quad \|f_\gamma\|_{\mathcal{H}_{\pi/2}} \le \tilde C_{s,\gamma} \|f\|_{\mathcal{D}_s}, \qquad f \in \mathcal{D}_s. \] \end{cor} \begin{proof} Using Lemma \ref{Dh1}, (i) and (ii), and Lemma \ref{simple}, we see firstly that $f \in \mathcal{H}_{\pi\gamma/2}$, and then $f_\gamma \in \mathcal{H}_{\pi/2} \subset \mathcal{D}_\sigma^\infty$. Moreover each of the embeddings from Lemma \ref{Dh1} is continuous, and the map $f \mapsto f_\gamma$ is isometric. \end{proof} Now we relate the spaces $\mathcal H_\psi$ and $\mathcal D_s$ to another class of spaces used in the literature on functional calculi. For $\psi \in (0,\pi)$, let \begin{equation} \label{edef} \mathcal E_\psi:=\left \{f \in {\operatorname{Hol}} (\Sigma_\psi): \|f\|_\psi:=\sup_{\varphi \in (0,\psi)}\int_{\partial \Sigma_{\varphi}} \frac{|f(z)|}{|z|}|dz| <\infty\right \}. \end{equation} It is easy to see that $(\mathcal E_\psi, \|\cdot \|_{\psi})$ is a Banach space, and that \[ \mathcal E_\psi=\{f \in {\operatorname{Hol}} (\Sigma_\psi): f(z)/z \in H^1(\Sigma_\psi)\}. \] \begin{prop}\label{weight1} Let $f \in \mathcal E_\psi$ and let $g(z):=f(z)/z$. Then, for every $\varphi\in (0,\psi)$, \begin{equation}\label{Q2} \|f'\|_{H^{1}(\Sigma_\varphi)} \le \frac{1}{2\pi} \left(\frac{\pi-\psi-\varphi}{\sin(\psi+\varphi)}+ \frac{\pi-\psi+\varphi}{\sin(\psi-\varphi)}\right)\|g\|_{H^1(\Sigma_\psi)}. \end{equation} Thus, $\mathcal E_\psi \overset{i}{\hookrightarrow} \mathcal H_{\varphi}, \, \varphi\in (0,\psi)$, and $\mathcal E_{\pi/2} \overset{i}{\hookrightarrow} \mathcal D_s,\, s>0$. \end{prop} \begin{proof} By Cauchy's theorem, for every $z\in\Sigma_\psi$, \[ f'(z) =\frac{1}{2\pi i}\int_{\partial\Sigma_\psi}\, \frac{\lambda g(\lambda)\,d\lambda}{(z-\lambda)^2}, \] Hence, for every $\varphi\in(0,\psi)$, by Fubini's theorem, \begin{align*} \int_0^\infty \left(|f'(\rho e^{i\varphi})| + |f'(\rho e^{-i\varphi})|\right)\,d\rho \le\frac{1}{2\pi} \int_{\partial\Sigma_\psi}|g(\lambda)| \, J(\lambda, \varphi) \, |d\lambda|, \end{align*} where \begin{align*} J(\lambda, \varphi)&:= \int_0^\infty\left(\frac{|\lambda|}{|\rho e^{i\varphi}-\lambda|^2}+ \frac{|\lambda|}{|\rho e^{-i\varphi}-\lambda|^2}\right)\ d\rho\\ \\ &=\int_0^\infty\frac{d\rho}{\rho^2-2\rho\cos(\psi+\varphi)+1} +\int_0^\infty\frac{d\rho}{\rho^2-2\rho\cos(\psi-\varphi)+1} \\ &=\frac{\pi-\psi-\varphi}{\sin(\psi+\varphi)}+ \frac{\pi-\psi+\varphi}{\sin(\psi-\varphi)}, \end{align*} in view of \cite[p.311]{Prud}. Hence $f \in \mathcal{H}_\varphi$ and \eqref{Q2} follows. Thus, $\mathcal E_\psi \overset{r}{\hookrightarrow} \mathcal H_{\varphi}$ for all $\varphi\in (0,\psi)$. Recalling Lemma \ref{Dh1} and that $\mathcal{H}_{\pi/2} = H^{1,1}(\mathbb C_+)$, we have proved the inclusion $\mathcal E_{\pi/2} \overset{i}{\hookrightarrow} \mathcal D_s, \, s>0$, as well. \end{proof} Note that the inclusions in Lemma \ref{weight1} are strict. Indeed, if $f(z)=z(z+1)^{-1}$ then one has $f \in \mathcal H_\psi$ for every $\psi \in (0,\pi)$, but $f \notin \mathcal E_{\psi}$ for any $\psi \in (0,\pi)$. Moreover, if \[ f(z)=\frac{z\,e^{-z}}{(z+1)\log^2(z+2)}, \] then $z^{-1}f \in H^1(\mathbb{C}_{+})$, and $f \in \mathcal{D}_0$, but $f \not\in \mathcal{D}_s$ for any $s\in (-1,0)$. The spaces $\mathcal E_\psi$ are studied in \cite[Chapter 10]{Hytonen}; see also \cite[Appendix H2]{Hytonen}, \cite[Section 6]{Haase_Sp} and \cite[Appendix C]{Haak_H}. To ensure the algebra property and to relate the spaces to the $H^\infty$-calculus, the authors considered the algebras $H^\infty(\Sigma_\psi)\cap \mathcal E_\psi$. Lemma \ref{weight1} shows that the spaces $\mathcal E_\psi$ are fully covered within the framework of the algebras $\mathcal D_\infty$ and $\mathcal H_\psi$. These algebras will be associated to the more powerful functional calculi constructed in Corollary \ref{Compat} and Theorem \ref{sigma_cal}. \subsection{Bernstein functions and $\mathcal{H}_\psi$} \label{bandh} To illustrate the relevance of the Hardy-Sobolev spaces, we show that the ``resolvent'' of a Bernstein function belongs to an appropriate Hardy-Sobolev space. This observation will be used in Section \ref{norm_estimates} to provide a new proof of the permanence of subordination for holomorphic semigroups, one of the main results of \cite{GT15}; see also \cite{BGTad} and \cite{BGT}. Let $g$ be a Bernstein function, $\psi\in (0,\pi/2)$, and $\lambda \in \Sigma_{\pi-\psi}$. Let \[ f(z, \lambda):=(\lambda+g(z))^{-1},\qquad z\in \Sigma_\psi. \] If $\lambda \in \mathbb{C}_+$, it follows from Lemmas \ref{bsds} and \ref{Dh1} that $f(\cdot, \lambda) \in \mathcal \mathcal{H}_\psi$, and $\|f(\cdot,\lambda)\|_{\mathcal{H}_\psi} \le C_\psi/|\lambda|$, where $C_\psi$ is independent of $g$ and $\lambda\in\mathbb{C}_+$. In order to obtain the correct angle, we will need to extend this to $\lambda \in \Sigma_{\pi-\varphi}$, where $\varphi \in (\psi,\pi/2)$. \begin{cor}\label{BF} Let $g$ be a Bernstein function, $\psi\in (0,\pi/2)$, $\varphi \in (\psi,\pi)$ and $\lambda \in \Sigma_{\pi-\varphi}$. Then \begin{equation}\label{bernst} \|f(\cdot, \lambda)\|_{\mathcal \mathcal{H}_\psi} \le 2\left(\frac{1}{\sin(\min(\varphi,\pi/2))}+\frac{2}{\cos\psi \sin^2((\varphi-\psi)/2)}\right)\frac{1}{|\lambda|}. \end{equation} \end{cor} \begin{proof} For fixed $\psi \in (0,\pi/2)$, $\varphi\in (\psi,\pi)$, and $\lambda\in \Sigma_{\pi-\varphi}$, observe that \[ \|f'(\cdot,\lambda)\|_{H^1(\Sigma_\psi)} \le \int_0^\infty \left( \frac{|g'(te^{i\psi})|}{|\lambda+g(te^{i\psi})|^2} + \frac{|g'(te^{-i\psi})|} {|\lambda+g(te^{-i\psi})|^2} \right) \,dt. \] Using the property (B1) for Bernstein functions and \eqref{Fc}, we have \[ |\lambda+g(te^{\pm i\psi})|\ge \sin((\varphi-\psi)/2)(|\lambda|+|g(te^{\pm i\psi})|), \qquad t \ge 0. \] Moreover, in view of (B3), for all $t \ge 0$, \begin{align*} |g(te^{\pm i\psi})|\ge \Re g(te^{\pm i\psi})\ge g(t\cos\psi)\quad \text{and} \quad |g'(e^{\pm i\psi}t)|&\le g'(t\cos\psi). \end{align*} Using (B2), we have \begin{align*} \|f'(\cdot, \lambda)\|_{H^1(\Sigma_\psi)} &\le \frac{2}{\sin^2((\varphi-\psi)/2)} \int_0^\infty\frac{g'(t\cos\psi)} {(|\lambda|+g(t\cos\psi ))^2} \,dt\\ &\le \frac{2}{\cos\psi \sin^2((\varphi-\psi)/2)} \int_0^\infty\frac{ds} {(|\lambda|+s)^2}\\ &= \frac{2}{\cos\psi \sin^2((\varphi-\psi)/2)}\frac{1}{|\lambda|}. \end{align*} Thus $f(\cdot,\lambda) \in \mathcal{H}_\psi$ and $f(\infty,\lambda) = (\lambda + g(\infty))^{-1}$. Since $|\arg\lambda|<\pi-\varphi$ and $g(\infty) \in [0,\infty]$, \begin{equation}\label{res_limit} |f(\infty,\lambda)| \le \frac{1}{\sin(\min(\varphi,\pi/2))|\lambda|}. \end{equation} Now \eqref{bernst} follows from Theorem \ref{hardy2}(ii). \end{proof} \subsection{Representations for functions in $\mathcal{H}_\psi$} \label{arccotrep} In this section we derive a reproducing formula for functions from $\mathcal{H}_\psi$ and obtain certain alternative representations for its kernel. \begin{prop}\label{FrepH} Let $f \in \mathcal{H}_\psi, \, \psi \in (0,\pi)$, and let \[ \gamma:=\frac{2\psi}{\pi}\in (0,2),\quad f_\gamma(z):=f(z^{\gamma}),\quad z\in \mathbb{C}_{+}. \] Then \begin{equation}\label{RepA1} f(z)=f(\infty)-\frac{1}{\pi}\int_0^\infty \int_{-\infty}^\infty \frac{f_\gamma'(\alpha+i\beta)}{z^{1/\gamma}+\alpha-i\beta}\,d\beta\,d\alpha,\qquad z\in \Sigma_\psi \cup \{0\}. \end{equation} \end{prop} \begin{proof} Since $f \in \mathcal{H}_\psi$, Lemma \ref{simple} implies that $f'_\gamma \in H^1(\mathbb{C}_{+})$. Hence, by Lemma \ref{Dh1} and Theorem \ref{hardy2}, we have $f_\gamma\in \mathcal{D}_0\cap C(\overline {\mathbb C}_+)$. Then, in view of Corollary \ref{Repr}, \[ f_\gamma(z)=f_\gamma(\infty)-\frac{1}{\pi}\int_0^\infty \int_{-\infty}^\infty \frac{f_\gamma'(\alpha+i\beta)}{z+\alpha-i\beta}\,d\beta\,d\alpha, \qquad z\in \mathbb{C}_{+}\cup \{0\}, \] and (\ref{RepA1}) follows. \end{proof} \begin{cor}\label{alpha} Let $f\in \mathcal{D}_\infty$. For $\gamma\in (0,1)$, set $f_\gamma(z):=f(z^\gamma), \, z\in\mathbb{C}_+$. Then \begin{equation}\label{RepA} f(z)=f(\infty)-\frac{1}{\pi}\int_0^\infty \int_{-\infty}^\infty \frac{f_\gamma'(\alpha+i\beta)}{z^{1/\gamma}+\alpha-i\beta}\,d\beta\,d\alpha,\quad z\in \Sigma_{\pi\gamma/2}\cup\{0\}. \end{equation} \end{cor} \begin{proof} By Lemma \ref{Dh1}, $f_\gamma \in \mathcal{H}_{\pi\gamma/2}$, so \eqref{RepA} follows from \eqref{RepA1}. \end{proof} The next reproducing formula for functions in $\mathcal{H}_\psi$ resembles \cite[Lemma 7.4]{Boy}, and it was, in fact, inspired by \cite[Lemma 7.4]{Boy}. In particular, this formula replaces the double (area) integral in \eqref{RepA1} with a line integral, it involves boundary values of $f'$ rather than scalings of $f'$ (such as in \eqref{RepA1}), and it offers a different kernel which might sometimes be easier to deal with. The $\arccot$ function has been defined in \eqref{arcdef}. \begin{prop}\label{L7.4} Let $f \in \mathcal \mathcal{H}_\psi$, $\psi\in(0,\pi)$. Let $\nu:=\pi/(2\psi)$, and \begin{equation}\label{fpsi} f_\psi(t):=\frac{f(e^{i\psi}t)+f(e^{-i\psi}t)}{2}, \quad t>0. \end{equation} Then \begin{equation}\label{arctan} f(z)=f(\infty)-\frac{2}{\pi}\int_0^\infty f_\psi'(t)\, {\arccot}(z^\nu/t^{\nu})\,dt,\qquad z\in \Sigma_\psi\cup \{0\}. \end{equation} \end{prop} \begin{proof} Since $\arccot \lambda \in \mathcal D_0$ (see Example \ref{arc}), \eqref{qnrep} shows that \begin{equation}\label{ArcN} \arccot(\lambda)=\frac{1}{\pi}\int_0^\infty\int_{-\infty}^\infty \frac{(\lambda+u-iv)^{-1}}{(u+iv)^2+1}\,dv\, du, \quad \lambda\in \mathbb{C}_{+}. \end{equation} Let $\gamma=1/\nu$, and $f_\gamma(\lambda):=f(\lambda^{\gamma})$, $\lambda\in \mathbb{C}_{+}$. By Proposition \ref{FrepH}, for $z \in \Sigma_\psi \cup \{0\}$, \begin{equation}\label{AlA} f(z)-f(\infty)= \frac{1}{\pi}\int_0^\infty \int_{-\infty}^\infty \frac{f_\gamma'(\alpha+i\beta)}{z^{\nu}+\alpha-i\beta}\,d\beta\,d\alpha. \end{equation} It follows from Lemma \ref{simple} that $f_\gamma' \in H^1(\Sigma_{\pi/2}) = H^1(\mathbb{C}_+)$, so by Cauchy's formula for functions in $H^1(\mathbb{C}_+)$ \cite[Theorem 11.8]{Duren1}, we have, for $\lambda \in \mathbb{C}_{+}$, \begin{align*} f_\gamma'(\lambda) &= \frac{1}{2\pi} \int_{-\infty}^\infty \frac{f_\gamma'(it)}{\lambda-it} \,dt + \frac{1}{2\pi}\int_{-\infty}^\infty \frac{f_\gamma'(it)}{\lambda+it} \,dt \\ &= - \frac{i}{\pi}\int_{-\infty}^\infty \frac{tf_\gamma'(it)}{\lambda^2+t^2} \,dt\\ &=\frac{i\gamma}{\pi }\int_0^\infty \frac{e^{i\pi(\gamma-1)/2}f'(e^{i\pi\gamma/2}t^\gamma)t^\gamma - e^{-i\pi(\gamma-1)/2}f'(e^{-i\pi\gamma/2}t^\gamma)t^\gamma}{\lambda^2+t^2} \,dt\\ &=\frac{\gamma}{\pi}\int_0^\infty \left(e^{i\psi}f'(e^{i\psi}t^\gamma) +e^{-i\psi}f'(e^{-i\psi}t^\gamma)\right)\frac{t^{\gamma}}{\lambda^2+t^2} \,dt\\ &=\frac{2}{\pi}\int_0^\infty \frac{f_\psi'(s) s^{\nu}}{\lambda^2+s^{2\nu}} \,ds. \end{align*} Therefore, by (\ref{AlA}) and (\ref{ArcN}), we obtain \begin{align*} &f(z)-f(\infty)\\ &= -\frac{2}{\pi^2}\int_0^\infty f'_\psi(s) \left(\int_0^\infty \int_{-\infty}^\infty \frac{s^\nu\,d\beta\, d\alpha}{(z^{\nu}+\alpha-i\beta)((\alpha+i\beta)^2+s^{2\nu})}\right)\,ds\\ &=-\frac{2}{\pi^2}\int_0^\infty f'_\psi(s) \left(\int_0^\infty \int_{-\infty}^\infty \frac{d\beta\, d\alpha}{((z/s)^{\nu}+\alpha-i\beta)((\alpha+i\beta)^2+1)}\right)\,ds\\ &=-\frac{2}{\pi}\int_0^\infty f'_\psi(s)\arccot(z^\nu/s^\nu)\,ds, \end{align*} so \eqref{arctan} holds for $z \in \Sigma_\psi \cup \{0\}$. \end{proof} Proposition \ref{L7.4} motivates a more careful study of the kernel $\arccot (z^\nu)$. The integral representation of this kernel will be crucial in deriving fine estimates for the $\mathcal{H}$-calculus for operators in Section \ref{alternative}. \begin{lemma}\label{log} Let $\psi\in (0,\pi)$ and $\nu={\pi}/{(2\psi)}$. Then \begin{align}\label{rew} \arccot(z^{\nu})&=\frac{1}{2\pi}\int_0^\infty V_\psi(z,t)\log\left|\frac{1+t^\nu}{1-t^\nu}\right| \,\frac{dt}{t}+\frac{1}{4 i}\int_{\Gamma_\psi}\, \frac{d\lambda}{\lambda-z},\quad z\in \Sigma_\psi, \end{align} where \begin{align}\label{vpsi} V_\psi(z,t)&:=-\frac{t}{2} \left(\frac{ e^{-i\psi}} {z-t e^{-i\psi}} +\frac{ e^{i\psi}}{z-t e^{i\psi}}\right), \\ \Gamma_\psi&:= \{\lambda: |\lambda|=1,\,\arg\lambda \in (\psi,2\pi-\psi)\}. \notag \end{align} \end{lemma} \begin{proof} We have \begin{align*} \arccot(z^{\nu}) =\frac{1}{2i}\log\left|\frac{z^\nu+i}{z^\nu-i}\right| +\frac{1}{2}\arg\left(\frac{z^\nu+i} {z^\nu-i}\right),\qquad z\in \Sigma_\psi. \end{align*} Since $\nu\psi=\pi/2$, for every $t>0, \, t\ne1$, \[ \lim_{z\in \Sigma_\psi,\,z\to te^{i\psi}} \arccot(z^{\nu}) =\frac{1}{2i}\log\left|\frac{1+t^\nu}{1-t^\nu}\right| +\frac{1}{2}\arg\left(\frac{t^\nu+1} {t^\nu-1}\right), \] and \[ \lim_{z\in \Sigma_\psi,\,z\to te^{-i\psi}}\arccot(z^{\nu}) =-\frac{1}{2i}\log\left|\frac{1+t^\nu}{1-t^\nu}\right| +\frac{1}{2}\arg\left(\frac{t^\nu-1} {t^\nu+1}\right). \] Here, \[ \arg\left(\frac{t^\nu+1} {t^\nu-1}\right)=\arg\left(\frac{t^\nu-1} {t^\nu+1}\right)=\pi\chi_{(0,1)}(t),\quad t>0, \, t\ne1, \] where $\chi_{[0,1]}$ is the characteristic function of $(0,1)$. So, \begin{equation}\label{zv} \lim_{z\in \Sigma_\psi,\,z\to te^{\pm i\psi}} \arccot(z^{\nu}) =\mp\frac{i}{2}\log\left|\frac{1+t^\nu}{1-t^\nu}\right| +\frac{\pi}{2}\chi_{[0,1]}(t), \end{equation} Now fix $z \in \Sigma_\psi$. Using $\limsup_{|\lambda|\to\infty,\,\lambda\in \mathbb{C}_{+}}\,|\lambda\arccot \lambda|<\infty$, it follows from Cauchy's theorem that \[ \arccot(z^{\nu})=-\frac{1}{2\pi i}\int_{\partial\Sigma_\psi}\, \frac{\arccot(\lambda^{\nu})}{\lambda-z}\,d\lambda, \quad \int_{\lambda\in\partial\Sigma_\psi, |\lambda|<1} \frac{d\lambda}{\lambda-z} = - \int_{\Gamma_\psi} \frac{d\lambda}{\lambda-z}. \] Thus by \eqref{zv}, \begin{align*} \arccot(z^{\nu}) &=-\frac{1}{2\pi }\int_0^\infty \frac{\Im\arccot(it^\nu)}{t-e^{-i\psi}z} \,dt +\frac{1}{2\pi}\int_0^\infty \frac{\Im\arccot(-it^\nu)}{t-e^{i\psi}z}\,dt \\ &\null \hskip30pt - \frac{1}{4 i}\int_{\lambda\in \partial\Sigma_\psi,\,|\lambda|\le 1}\, \frac{d\lambda}{\lambda-z}\\ &=\frac{1}{4\pi}\int_0^\infty \log\left|\frac{1+t^\nu}{1-t^\nu}\right| \left(\frac{1}{t-e^{i\psi}z} +\frac{1}{t-e^{-i\psi}z}\right)\,dt +\frac{1}{4 i}\int_{\Gamma_\psi}\, \frac{d\lambda}{\lambda-z}\\ &=\frac{1}{2\pi}\int_0^\infty V_\psi(z,t)\log\left|\frac{1+t^\nu}{1-t^\nu}\right| \,\frac{dt}{t}+\frac{1}{4 i}\int_{\Gamma_\psi}\, \frac{d\lambda}{\lambda-z}. \qedhere \end{align*} \end{proof} \begin{rem}\label{Value} Letting $z \to 0$ in (\ref{rew}) we obtain \begin{align*} \frac{\pi}{2} &=\frac{1}{2\pi}\int_0^\infty \log\left|\frac{1+t^\nu}{1-t^\nu}\right| \frac{dt}{t} +\frac{\pi-\psi}{2}, \end{align*} hence \begin{equation}\label{remA} \frac{1}{2\pi}\int_0^\infty \log\left|\frac{1+t^\nu}{1-t^\nu}\right| \frac{dt}{t}=\frac{\psi}{2}. \end{equation} \end{rem} \section{Dense sets in $\mathcal{D}_s$ and $\mathcal{H}_\psi$} \label{density} In this section we establish some results concerning density and approximations in our spaces. \subsection{Dense subsets of $\mathcal{D}_s$ and some applications} Let $\mathcal{R}(\mathbb{C}_+)$ be the linear span of $\{r_\lambda : \lambda \in \mathbb{C}_+\}$, and $\widetilde{\mathcal{R}}(\mathbb{C}_+)$ be the sum of $\mathcal{R}(\mathbb{C}_+)$ and the constant functions. Using Example \ref{hexs}(1) and Lemma \ref{Dh1}, we have \begin{equation} \label{rinh} \mathcal{\widetilde{R}}(\mathbb{C}_+) \subset \mathcal{H}_{\pi/2} \overset{i}{\hookrightarrow} \mathcal{D}_s, \quad s>-1. \end{equation} \begin{thm} \label{D00} The space \[ \widetilde{\mathcal{R}}(\mathbb{C}_+):=\left\{a_0 + \sum_{k=1}^n a_k(\lambda_k+z)^{-1}: n \in \mathbb N, \; a_k \in \mathbb{C}, \; \lambda_k\in \mathbb{C}_{+}\right\} \] is dense in $\mathcal{D}_s$ for each $s>-1$. \end{thm} \begin{proof} Let $\mathcal{R}_{\mathcal{D}_s}(\mathbb{C}_+)$ be the closure of $\mathcal{R}(\mathbb{C}_+)$ in $\mathcal{D}_{s}$, F and let $f \in \mathcal{D}_s$. It follows from Example \ref{resd} and Remark \ref{remdr} (or a direct estimate) that the function \[ (R_f)(\lambda) := - \frac{2^s}{\pi} f'(\lambda) r_{\overline{\lambda}} \] is continuous from $\mathbb{C}_+$ to $\mathcal{D}_s$, and it is Bochner integrable with respect to area measure $S$ on $\mathbb{C}_+$. Since point evaluations are continuous on $\mathcal{D}_s$ (Corollary \ref{Cangle}), it follows from Corollary \ref{Repr} that \[ Q_s f' = \int_{\mathbb{C}_+} R_f(\lambda) \,dS(\lambda) \] as the Bochner integral of a continuous function. Hence $Q_s f'$ belongs to the closure in $\mathcal{D}_s$ of the linear span of the range of the integrand, which is contained in $\mathcal{R}_{\mathcal{D}_s}(\mathbb{C}_+)$. Now $f = f(\infty) + Q_sf'$ which is in the closure of $\widetilde{\mathcal{R}}(\mathbb{C}_+)$ in $\mathcal{D}_s$. \end{proof} From Proposition \ref{Furt}, we have the continuous inclusion \[ \mathcal{D}_s \overset{i}{\hookrightarrow} \mathcal{D}_\sigma \quad \text{if $\sigma>s>-1$}, \] and from \eqref{rinh}, Theorem \ref{ACM} and Proposition \ref{BDs}, we have \[ \widetilde{\mathcal{R}}(\mathbb{C}_+) \subset H^{1,1}(\mathbb{C}_+) \overset{i}{\hookrightarrow} \mathcal{L} L^1 + \mathbb{C} \subset \mathcal{LM} \overset{i}{\hookrightarrow} \mathcal{B} \overset{i}{\hookrightarrow} \mathcal{D}_s^\infty \overset{i}{\hookrightarrow} \mathcal{D}_s \quad \text{if $s>0$}. \] Here $\mathcal{L} L^1+\mathbb{C}$ is the sum of $\mathcal{L} L^1$ and the constant functions, and it is a closed subspace of $\mathcal{LM}$. The following density results hold. \begin{cor} \label{Bdens} \begin{enumerate}[\rm1.] \item If $\sigma>s>-1$, then $\mathcal{D}_s$ is dense in $\mathcal{D}_\sigma$. \item For $s>-1$, the spaces $H^{1,1}(\mathbb{C}_+)$ and $\mathcal{D}_s^\infty$ are dense in $\mathcal{D}_s$. \item For $s>0$, the spaces $\mathcal{L}L^1+\mathbb{C}$, $\mathcal{LM}$ and $\mathcal{B}$ are dense in $\mathcal{D}_s$. \item For $s>0$, the spaces $\widetilde{\mathcal{R}}(\mathbb{C}_+)$, $H^{1,1}(\mathbb{C}_+)$, $\mathcal{L}L^1 + \mathbb{C}$, $\mathcal{LM}$ and $\mathcal{B}$ are not dense in $\mathcal{D}_s^\infty$. \end{enumerate} \end{cor} \begin{proof} The first three statements are immediate from Theorem \ref{D00}. Since any function in $\mathcal{B}$ extends continuously to $i\mathbb{R}$, the same holds for the closure of $\mathcal{B}$ in $\mathcal{D}_s^\infty$ when $s>0$. The function $f(z)=e^{-1/z}\in \mathcal{D}_s^\infty$ for $s>0$, (see Remarks \ref{e-z} and Example \ref{dm} with $\nu=0$), but $f$ is not continuous at $z=0$. This establishes the final statement. \end{proof} The function $g(z) = \exp(\arccot z)$ considered in Example \ref{eac} provides another example of a function from $\mathcal{D}_0^\infty$ which is discontinuous on $i\mathbb{R}$, and so does not belong to the closure of $\mathcal{B}$ in $\mathcal{D}^\infty_s$ for $s>0$. In order to obtain operator norm-estimates for functions $f^{(n)}$ applied to semigroup generators (see Theorem \ref{AnalF}), we will need a stronger version of Corollary \ref{D00n} on differentiability of $t \to f(t\cdot)$ in the $D_s$-norm. We first prove a lemma, and we present the stronger statement in Corollary \ref{AlgDer}. \begin{lemma}\label{raF} Let $\lambda\in \mathbb{C}_{+}$, $\tau>0$, $t \in (\tau/2,2\tau)$, and define \[ g_{t,\tau,\lambda}(z):=\frac{r_\lambda(tz)-r_\lambda(\tau z)}{t-\tau}-zr_\lambda'(\tau z),\qquad z\in \mathbb{C}_{+}. \] Then, for every $s>-1$, \begin{equation}\label{Nder} \lim_{t\to\tau}\,\|g_{t,\tau,\lambda}\|_{\mathcal{D}_s}=0. \end{equation} \end{lemma} \begin{proof} We have \begin{align*} g_{t,\tau,\lambda}'(z)&=\left(-\frac{t}{(tz+\lambda)^2}+ \frac{\tau}{(\tau z+\lambda)^2}\right)\frac{1}{t-\tau} +\frac{1}{(\tau z+\lambda)^2}-\frac{2\tau z}{(\tau z+\lambda)^3}\\ &= \frac{\tau t z^2 - \lambda^2}{(tz+\lambda)^2(\tau z+\lambda)^2} + \frac{\lambda-\tau z}{(\tau z+\lambda)^3} \to0, \qquad t \to \tau. \end{align*} Hence \[ |g_{t,\tau,\lambda}'(z)|\le \frac{C_{\tau,\lambda}}{1+|z|^2}, \qquad t \in (\tau/2,2\tau), \, z \in \mathbb{C}_+, \] for some $C_{\tau,\lambda}$. Since \[ \int_{-\pi/2}^{\pi/2} \cos^s\psi\int_0^\infty \frac{d\rho}{1+\rho^2}\,d\psi<\infty \] for any $s>-1$, the dominated convergence theorem implies \eqref{Nder}. \end{proof} \begin{cor}\label{DcorA} Let $f\in \mathcal{D}_s,$ $s>-1$. For every $\tau>0$, \[ \lim_{t\to \tau}\left\| \frac{f(t z)-f(\tau z)}{t-\tau}-zf'(\tau z)\right\|_{\mathcal{D}_{s+1}}=0. \] \end{cor} \begin{proof} Let $\tau>0$ be fixed, and \[ (R_{t,\tau} f)(z) := \frac{f(t z)-f(\tau z)}{t-\tau}-zf'(\tau z), \qquad f \in\mathcal{D}_s, \, t>\tau/2. \] By Lemma \ref{D001}, $\{R_{t,\tau} : t> \tau/2\}$ is a bounded subset of $L(\mathcal{D}_s,\mathcal{D}_{s+1})$. By Lemma \ref{raF}, \[ \lim_{t\to\tau} \|R_{t,\tau}r_\lambda\|_{\mathcal{D}_{s+1}} = 0, \qquad \lambda \in \mathbb{C}_+. \] Since the linear span of the functions $r_\lambda$ and the constants is dense in $\mathcal{D}_s$ (see Theorem \ref{D00}), the assertion follows. \end{proof} \begin{cor}\label{AlgDer} Let $f\in\mathcal{D}_s$, $s>-1$, and let \[ G(t)(z) : = f(tz), \qquad F_n(t)(z) := z^nf^{(n)}(tz), \quad n\in \mathbb{N}, \,t>0, \, z \in \mathbb{C}_+. \] Then $G$ and $F_n$ map $(0,\infty)$ into $\mathcal{D}_{s+n}$, $G$ is $n$-times differentiable as a function from $(0,\infty)$ to $\mathcal{D}_{s+n}$, and \[ F_n = G^{(n)}. \] \end{cor} \begin{proof} Firstly, $f(tz) \in \mathcal{D}_s \subset \mathcal{D}_{s+n}$, so $G$ maps $(0,\infty)$ into $\mathcal{D}_{s+n}$. The proof is by induction on $n$. The case $n=1$ is given by Corollary \ref{DcorA}. Assume that $G^{(n)} = F_n$ with values in $\mathcal{D}_{s+n}$, and let $f_n(z) = z^n f^{(n)}(z)$. Then \[ G^{(n)}(t)(z) = F_n(t)(z) = t^{-n}f_n(tz). \] By Corollary \ref{DcorA} applied to $f_n\in\mathcal{D}_{s+n}$, $G^{(n)}$ is differentiable with respect to $t$, when considered as a function with values in $\mathcal{D}_{s+n+1}$. Finally, \[ G^{(n+1)}(t)(z) = \frac{d}{dt} (z^n f^{(n)}(tz)) = z^{n+1} f^{(n+1)}(tz) = F_{n+1}(t)(z). \qedhere \] \end{proof} \subsection{Approximations via change of variables} Here we consider approximations of $f$ from $\mathcal{D}_s$ and $\mathcal{H}_\psi$ by the functions $f_\gamma(z) := f(z^\gamma)$ as $\gamma \to1-$. \begin{prop} \label{gggP} Let $\gamma \in (0,1)$. The following hold. \begin{enumerate}[\rm1.] \item Let $s>-1$, and $f \in \mathcal{D}_s$. Then \begin{equation} \label{Dc1} \|f_\gamma\|_{\mathcal{D}_s} \le \|f\|_{\mathcal{D}_s}, \end{equation} and \begin{equation} \label{Dc2} \lim_{\gamma\to1-} \|f_\gamma-f\|_{\mathcal{D}_s} = 0. \end{equation} \item Let $\psi\in (0,\pi)$and $g\in \mathcal{H}_{\psi}$. Then \begin{equation} \label{Sal1} \lim_{\gamma\to1-} \|g_\gamma -g\|_{\mathcal{H}_\psi}= 0. \end{equation} \end{enumerate} \end{prop} \begin{proof} 1. First let $g \in H^{1,1}(\mathbb{C}_{+})$. If $0<\psi<\varphi\le \pi/2$, then it follows from Theorem \ref{hardy0}(iv) that \begin{equation}\label{Afs} \int_0^\infty \left(|g'(te^{i\varphi})|+|g'(te^{-i\varphi})|\right)\,dt \ge \int_0^\infty \left(|g'(te^{i\psi})|+|g'(t e^{-i\psi})|\right)\,dt. \end{equation} Hence \begin{align*} \|g\|_{\mathcal{D}_s} &= |g(\infty)|+ \int_0^{\pi/2}\cos^s\varphi \int_0^\infty \left(|g'(te^{i\varphi})|+|g(te^{-i\varphi})|\right)\,dt\,d\varphi \\ &\ge |g(\infty)|+ \int_0^{\pi/2}\cos^s\varphi \int_0^\infty \left(|g'(te^{i\gamma\varphi})|+|g'(te^{-i\gamma\varphi})|\right)\,dt\,d\varphi \\ &= \|g_\gamma\|_{\mathcal{D}_s}. \end{align*} Since $\mathcal{H}_{\pi/2}$ is dense in $\mathcal{D}_s$ (see Corollary \ref{Bdens}), it follows that the map $g \mapsto g_\gamma$ extends to a contraction on $\mathcal{D}_s$, and this contraction maps $f$ to $f_\gamma$. Then \eqref{Dc1} holds. Now \eqref{Dc2} follows from \eqref{Sal1}, \eqref{Dc1}, and the fact that $\mathcal{H}_{\pi/2}$ is continuously and densely embedded in $\mathcal{D}_s$ (see Proposition \ref{Dh1} and Corollary \ref{Bdens}). 2. Since the norms $\|\cdot\|_{\mathcal{H}_\psi}$ and $\|\cdot\|'_{\mathcal{H}_\psi}$ are equivalent and $g_\gamma(\infty) = g(\infty)$, it suffices to show that \[ \|g'_\gamma - g'\|_{H^1(\Sigma_\psi)} = \int_{\partial\Sigma_\psi} |g_\gamma'(z)-g(z)| \,|dz| \to 0. \] By Lemma \ref{simple}, we have \[ \int_{\partial\Sigma_\psi}|g'_\gamma(z)|\,|dz| =\int_{\partial\Sigma_{\gamma\psi}}|g'(z)|\,|dz|. \] Applying Theorem \ref{hardy0}(iii) to $g'$, \[ \lim_{\gamma\to 1-} \int_{\partial\Sigma_{\gamma\psi}}|g'(z)|\,|dz| =\int_{\partial\Sigma_{\psi}}|g'(z)|\,|dz| \] and \[ \lim_{\gamma\to1-} g'_\gamma(z) = g'(z), \quad \text{for almost all $z \in\partial \Sigma_\psi$}. \] Now the statement \eqref{Sal1} follows from Lemma \ref{duren21}. \end{proof} \subsection{Density of rational functions in $\mathcal{H}_\psi$} In Theorem \ref{D00} and Corollary \ref{Bdens}, we established that $\widetilde{\mathcal{R}}(\mathbb{C}_+)$ and several larger spaces are dense in $\mathcal{D}_s$, for $s>-1$ or $s>0$. In particular, we noted that $H^{1,1}(\mathbb{C}_+)$ is dense in $\mathcal{D}_s$. Let $\psi\in (0,\pi)$ and \[ \mathcal{H}_{\psi,0}=\{f\in \mathcal{H}_\psi:\,f(\infty)=0\}, \] with the norm \[ \|f\|_{\mathcal{H}_{\psi,0}}=\|f'\|_{H^1(\Sigma_\psi)}. \] By \eqref{bound_for_sup}, this norm is equivalent to $\|\cdot\|_{\mathcal{H}_\psi}$ on $\mathcal{H}_{\psi,0}$. Let $\psi \in (0,\pi)$ and $\mathcal{R}(\Sigma_\psi)$ be the linear span of $\{r_\lambda :\lambda \in \Sigma_{\pi-\psi}\}$. Let $\mathcal{R}_\mathcal{H}(\Sigma_\psi)$ be the closure of $\mathcal{R}(\Sigma_\psi)$ in $\mathcal{H}_{\psi,0}$. We shall prove that $\mathcal{H}_{\psi,0} = \mathcal R_{\mathcal H}(\Sigma_\psi)$. Thus the rational functions which vanish at infinity and have simple poles outside $\overline\Sigma_\psi$ are dense in $\mathcal{H}_\psi$ modulo constants. This fact may be known, but we did not find it in the literature. Our proof involves several lemmas given below and it may be of interest as a piece of function theory. The following lemma, relating to the function spaces $\mathcal E_\varphi$ from Definition \eqref{edef}, is the key step in our proof. \begin{lemma}\label{DH111} Let $\psi\in (0,\pi)$, $\varphi\in (\psi,\pi)$ and let $f\in H^1(\Sigma_{\varphi})$. If \[ \int_{\partial\Sigma_{\varphi}}\frac{|f(\lambda)|}{|\lambda|}\,|d\lambda|<\infty, \] then \begin{equation}\label{GGG1} f\in \mathcal{R}_\mathcal{H}(\Sigma_\psi). \end{equation} \end{lemma} \begin{proof} From \eqref{ff} for $\gamma=1$, the function $F(\lambda) := r_\lambda$ maps $\Sigma_{\pi-\psi}$ into $\mathcal{H}_{\psi,0}$, and is locally bounded. Moreover $F$ is holomorphic (see Theorem \ref{hardy2}(iii) and Section \ref{prelims}) and its derivative is $-r_\lambda^2$. The Cauchy integral formula \eqref{Boch1} may be written as \begin{equation*} f(z)= \frac{1}{2\pi i} \int_{\partial\Sigma_{\varphi}} f(\lambda) F(-\lambda)(z) \,d\lambda, \quad z\in \Sigma_\psi. \end{equation*} From \eqref{ff}, we obtain \[ \int_{\partial\Sigma_\varphi} \|f(\lambda)F(-\lambda)\|_{\mathcal{H}_{\psi,0}} \,|d\lambda| \le \frac{1}{2\pi \cos^2 ((\varphi-\psi)/2) } \int_{\partial\Sigma_\varphi} \frac{|f(\lambda)|}{|\lambda|}\,d\lambda < \infty. \] Thus \[ f = \frac{1}{2\pi i} \int_{\partial\Sigma_\varphi} f(\lambda)F(-\lambda) \,d\lambda \] as a Bochner integral in $\mathcal{H}_{\psi,0}$, with continuous integrand, so it may be approximated in the $\mathcal{H}_{\psi,0}$-norm by Riemann sums of the integrand, which lie in $\mathcal{R}(\Sigma_\psi)$. Hence $f\in \mathcal{R}_\mathcal{H}(\Sigma_\psi)$. \end{proof} The next step in the proof is to construct a family of functions which serve as an approximate identity for $\mathcal H_{\varphi,0}$ when restricted to any sector smaller than $\Sigma_\varphi$. \begin{lemma}\label{Add} Let $\varphi \in (0,\pi)$ and $\epsilon\in (0,1)$, and let \begin{equation}\label{Deps} g_\epsilon(z):= \frac{2z^\epsilon}{1+z^\epsilon}(1+\epsilon z)^{-2},\qquad z\in \mathbb{C}\setminus (-\infty,0]. \end{equation} Then $g_\epsilon\in \mathcal{H}_{\varphi,0} \cap H^1(\Sigma_\varphi)$ and \begin{equation}\label{DR1} \sup_{\epsilon \in (0,1)} \|g_\epsilon'\|_{H^1(\Sigma_\varphi)} <\infty. \end{equation} Moreover, \begin{equation}\label{DR2} \lim_{\epsilon\to 0}\,g_\epsilon(z)=1,\quad z\in \mathbb{C}\setminus (-\infty,0], \end{equation} and, for $0<a<b<\infty$, there exists $C_{\varphi,a,b}$ such that \begin{equation}\label{DR3} |g'_\epsilon(z)|\le C_{\varphi,a,b}\cdot \epsilon, \quad z\in\partial\Sigma_\varphi,\, |z|\in (a,b). \end{equation} \end{lemma} \begin{proof} It is clear that $g_\epsilon \in H^1(\Sigma_\varphi)$ and $g_\epsilon(\infty)=0$. Moreover \begin{equation*} g'_\epsilon(z)=\frac{2\epsilon z^{\epsilon}}{z(1+z^\epsilon)^2(1+\epsilon z)^2}- \frac{4\epsilon z^{\epsilon}}{(1+z^\epsilon)(1+\epsilon z)^3}. \end{equation*} Applying Lemma \ref{trig}, there is a constant $C_\varphi$ such that, for $z \in \partial\Sigma_\varphi$ and $t=|z|$, \begin{equation} \label{DerD} |g'_\epsilon(z)|\le C_\varphi \epsilon t^\epsilon \left(\frac{2}{t(1+t^\epsilon)^2(1+\epsilon t)^2}+ \frac{4}{(1+t^\epsilon)(1+\epsilon t)^3}\right). \end{equation} Hence \begin{align*} \|g'_\epsilon\|_{H^1(\Sigma_\varphi)} &\le 2\epsilon C_\varphi\int_0^\infty \frac{t^\epsilon\,dt}{t(1+t^\epsilon)^2(1+\epsilon t)^2} + 4\epsilon C_\varphi\int_0^\infty \frac{t^{\epsilon}\,dt}{(1+t^\epsilon)(1+\epsilon t)^3} \\ &\le 2 \epsilon C_\varphi \int_1^\infty \frac{dt}{(1+\epsilon t)^2} + 2\epsilon C_\varphi \int_0^1 t^{\epsilon-1}\,dt + 4 \epsilon C_\varphi \int_0^\infty \frac{dt}{(1+\epsilon t)^3} \\ &= 6 C_\varphi. \end{align*} This yields \eqref{DR1}. The property \eqref{DR2} is straightforward, and \eqref{DR3} follows from \eqref{DerD}. \end{proof} Lemmas \ref{DH111} and \ref{Add} enable us to show that any function $f \in \mathcal{H}_{\varphi,0}$, when restricted to $\Sigma_\psi, \, \psi \in (0,\varphi)$, can be approximated by rational functions in $\mathcal{H}_{\psi,0}$. \begin{lemma}\label{DM1} Let $\psi \in (0,\pi)$, $\varphi\in (\psi,\pi)$, and let $ f\in \mathcal{H}_{\varphi,0}. $ Then \begin{equation}\label{dens} f\in \mathcal{R}_\mathcal{H}(\Sigma_\psi). \end{equation} \end{lemma} \begin{proof} Assume first that $f\in H_{\varphi,0}$ and $f(0)=0$, and let $g_\epsilon$ be defined by \eqref{Deps}. Then $f g_\epsilon \in \mathcal{H}_{\varphi,0} \cap H^1(\Sigma_{\varphi})$, and \[ \int_{\partial \Sigma_{\varphi}} \frac{|f(z)g_\epsilon(z)|}{|z|}\,|dz|<\infty. \] By Lemma \ref{DH111}, $fg_\epsilon \in \mathcal{R}_\mathcal{H}(\Sigma_\psi)$. Note that \[ \|f(1-g_\epsilon)\|_{\mathcal{H}_{\varphi,0}} \le \|f'(1-g_\epsilon)\|_{H^1(\Sigma_{\varphi})}+ \|f g'_\epsilon\|_{H^1(\Sigma_{\varphi})}. \] We will prove that \begin{equation}\label{Dlim} \lim_{\epsilon\to 0}\,\|f'(1-g_\epsilon)\|_{H^1(\Sigma_\varphi)}=0, \end{equation} and \begin{equation}\label{Dlim1} \lim_{\epsilon\to 0}\,\|fg'_\epsilon\|_{H^1(\Sigma_{\varphi})}=0. \end{equation} By \eqref{bound_for_sup} and \eqref{DR1}, \[ \sup_{\epsilon\in(0,1) } \|g_\epsilon\|_{H^\infty(\Sigma_\varphi)}\le \sup_{\epsilon \in (0,1)} \|g'_\epsilon\|_{H^1(\Sigma_\varphi)}<\infty, \] so, using \eqref{DR2} and the dominated convergence theorem, we have \[ \|f'(1-g_\epsilon)\|_{H^1(\Sigma_{\varphi})}= \int_{\partial\Sigma_\varphi} |f'(\lambda)(1-g_\epsilon(\lambda))|\,|d\lambda| \to 0,\quad \epsilon\to 0. \] For $0<a<b<\infty$, \begin{align*} \|f g'_\epsilon\|_{H^1(\Sigma_{\varphi})} &=\int_{\partial\Sigma_{\varphi}}|f(\lambda)||g_\epsilon'(\lambda)|\,|d\lambda|\\ &\le \sup_{|z|>b,\,z\in\partial\Sigma_{\varphi}}\,|f(z)|\, \int_{|\lambda|>b, \,\lambda\in \partial\Sigma_{\varphi}} |g'_\epsilon(\lambda)|\,|d\lambda|\\ &\null\hskip20pt +\sup_{|z|<a,\,z\in\partial\Sigma_{\varphi}}\,|f(z)|\, \int_{|\lambda|<a, \,\lambda\in \partial\Sigma_{\varphi}} |g'_\epsilon(\lambda)|\,|d\lambda|\\ &\null\hskip20pt +\|f\|_{H^{\infty}(\Sigma_{\varphi})} \int_{|\lambda|\in (a,b), \,\lambda\in \partial\Sigma_{\varphi}} |g'_\epsilon(\lambda)|\,|d\lambda|\\ &\le \left(\sup_{|z|>b,\,z\in\partial\Sigma_{\varphi}}\,|f(z)| +\sup_{|z|<a,\,z\in\partial\Sigma_{\varphi}}\,|f(z)|\right) \|g'_\epsilon\|_{H^1(\Sigma_{\varphi})}\\ &\null\hskip20pt +2(b-a) \|f\|_{H^{\infty}(\Sigma_{\varphi})} \sup_{|z|\in(a,b),\,z\in \partial\Sigma_{\varphi}}\,|g'_\epsilon(z)|. \end{align*} Letting first $\epsilon\to0$, using \eqref{DR1}-\eqref{DR3} along with Vitali's theorem, and then letting $a\to0$ and $b \to \infty$, using $f(0)=f(\infty)=0$, we obtain \[ \lim_{\epsilon\to 0}\|f g'_\epsilon\|_{H^1(\Sigma_{\varphi})}=0. \] We have now proved the assertions \eqref{Dlim} and \eqref{Dlim1}. Thus we obtain \eqref{dens} under the additional assumption that $f(0)=0$. Now let $f\in \mathcal{H}_{\varphi,0}$ be arbitrary. Then consider \[ f_0(z):=f(z)-2f(0)\left(\frac{1}{z+1}-\frac{1}{z+2}\right), \] and note that \[ f_0 \in \mathcal{H}_{\varphi,0} \cap H^1(\Sigma_{\varphi}),\qquad f_0(0)=0. \] Then, by the above, \[ f_0\in \mathcal{R}_\mathcal{H}(\Sigma_\varphi), \] and hence \eqref{dens} holds. \end{proof} We now approximate functions $f \in \mathcal{H}_{\psi,0}$ by functions from $\mathcal{H}_{\psi',0}, \, \psi'>\psi$, using the change of variables from Proposition \ref{gggP}. For $f\in \mathcal{H}_{\psi,0}$ and $f_\gamma(z) := f(z^\gamma)$ for $\gamma \in (0,1)$, we now have \[ f_\gamma\in \mathcal{H}_{\varphi_\gamma,0},\qquad \varphi_\gamma:=\min\{\gamma^{-1}\psi,\pi\}>\psi. \] By \eqref{Sal1}, \[ \lim_{\gamma\to 1-}\,\|f-f_\gamma\|_{\mathcal{H}_{\psi,0}}=0. \] By Lemma \ref{DM1}, $f_\gamma \in \mathcal{R}_\mathcal{H}(\Sigma_\psi)$. Thus we obtain the following result that the rational functions with simple poles are dense in $\mathcal{H}_{\psi,0}$. \begin{thm}\label{DM122} Let $\psi\in (0,\pi)$. Then \begin{equation}\label{GGG122} \mathcal{H}_{\psi,0}=\mathcal{R}_\mathcal{H}(\Sigma_\psi). \end{equation} \end{thm} \section{Convergence Lemmas} In this section we formulate Convergence Lemmas for functions in $\mathcal{D}_s$ and $\mathcal{H}_\psi$, composed with fractional powers. \begin{lemma}\label{48St2} Let $s>-1$ and $(f_k)_{k=1}^\infty\subset \mathcal{D}_s$ be such that \begin{equation*}\label{Bound50A} \sup_{k \ge 1}\|f_k\|_{\mathcal{D}_s}<\infty, \end{equation*} and for every $z\in \mathbb{C}_{+}$ there exists \[ f_0(z):=\lim_{k\to\infty}\,f_k(z). \] Let $g\in \mathcal{D}_s$ satisfy \[ g(0)=g(\infty)=0. \] For $\gamma\in (0,1)$, let \[ f_{k,\gamma}(z) = f_k(z^\gamma), \quad g_\gamma(z) = g(z^\gamma), \qquad z \in \mathbb{C}_+. \] Then \begin{equation}\label{Se50A} \lim_{k\to\infty}\,\|f_{k,\gamma}g_\gamma\|_{\mathcal{D}_s}=0. \end{equation} \end{lemma} \begin{proof} By Corollary \ref{Fatou}, $f_0 \in \mathcal D_s$. So without loss of generality, we can assume that $f_0=0$. By Corollary \ref{gggC}, $g_{\gamma}$ and $f_{k, \gamma}$ belong to the algebra $\mathcal{D}_s^\infty$, so $f_{k, \gamma}g_{\gamma}\in \mathcal{D}_s$. Moreover \[ B:=\sup_{k \ge 1}\left(\|f_{k, \gamma}\|_{\mathcal{D}_s}+\|f_{k, \gamma}\|_{\infty}\right)< \infty. \] Let $0 < r < R < \infty$ and $\Omega_{r,R} = \{z \in \mathbb{C}_+ : r \le |z| \le R\}$. By Vitali's theorem, \[ \lim_{k\to\infty}\,\left(|f_{k, \gamma}(z)|+|f_{k, \gamma}'(z)|\right)=0 \] uniformly on $\Omega_{r,R}$. Therefore the integrals \[ \int_{\Omega_{r,R}} \frac{(\Re z)^s}{|z|^{s+1}} |f'_{k,\gamma}(z) g_\gamma(z)| \,dS(z), \quad \int_{\Omega_{r,R}} \frac{(\Re z)^s}{|z|^{s+1}} |f_{k,\gamma}(z) g'_\gamma(z)| \,dS(z), \] tend to $0$ as $k\ \to \infty$. Moreover, \[ \int_{\mathbb{C}_+ \setminus \Omega_{r,R}} \frac{(\Re z)^s}{|z|^{s+1}} |f'_{k,\gamma}(z) g_\gamma(z)| \,dS(z) \le \sup_{z \in \mathbb{C}_+ \setminus \Omega_{r,R}} |g_\gamma(z)| \, \|f_{k,\gamma}\|_{\mathcal{D}_s}, \] and \[ \int_{\mathbb{C}_+ \setminus \Omega_{r,R}} \frac{(\Re z)^s}{|z|^{s+1}} |f_{k,\gamma}(z) g'_\gamma(z)| \,dS(z) \le \|f_{k,\gamma}\|_\infty \int_{\mathbb{C}_+ \setminus\Omega_{r,R}} \frac{(\Re z)^s}{|z|^{s+1}} |g'_\gamma(z)| \,dS(z). \] Hence \[ \limsup_{k\to\infty} \| f_{k,\gamma}g_\gamma\|_{\mathcal{D}_s} \le B \left( \sup_{\mathbb{C}_+ \setminus \Omega_{r,R}} |g_\gamma(z)| + \int_{\mathbb{C}_+ \setminus \Omega_{r,R}} \frac{(\Re z)^s}{|z|^{s+1}} |g'_\gamma(z)| \,dS(z) \right). \] Letting $r\to0$ and $R \to \infty$, we obtain the assertion \eqref{Se50A}. \end{proof} The following result is a convergence lemma for $\mathcal{H}_\psi$, analogous to Lemma \ref{48St2}. \begin{lemma}\label{H11cl} Let $\psi\in(0,\pi)$ and $(f_k)_{k=1}^\infty\subset \mathcal{H}_\psi$ be such that \begin{equation*}\label{Bound50B} \sup_{k \ge 1}\|f_k\|_{\mathcal{H}_\psi}<\infty, \end{equation*} and for every $z\in \mathbb{C}_{+}$ there exists \[ f_0(z):=\lim_{k\to\infty}\,f_k(z). \] Let $g\in \mathcal{H}_\psi$ satisfy \[ g(0)=g(\infty)=0. \] For $\gamma\in (0,1)$, let \[ f_{k,\gamma}(z) = f_k(z^\gamma), \quad g_\gamma(z) = g(z^\gamma), \qquad z \in \mathbb{C}_+. \] Then \begin{equation}\label{Se50B} \lim_{k\to\infty}\,\|f_{k,\gamma}g_\gamma\|_{\mathcal{H}_\psi}=0. \end{equation} \end{lemma} \begin{proof} The proof is similar to Lemma \ref{Se50A}. By Lemma \ref{FatouH}, $f_0 \in \mathcal H_\psi$. Thus, we will assume that $f_0=0$. Let $\gamma \in (\psi/\pi,1)$. By Lemma \ref{simple}, $g_{\gamma}$ and $f_{k, \gamma}$ belong to the algebra $\mathcal{H}_{\psi/\gamma} \subset \mathcal{H}_\psi$, ao $f_{k, \gamma}g_{\gamma}\in \mathcal{H}_\psi$. Moreover \[ B:=\sup_{k \ge 1}\left(\|f_{k, \gamma}\|_{\mathcal{H}_\psi}+\|f_{k, \gamma}\|_{H^\infty(\Sigma_\psi)}\right)< \infty. \] Let $0 < r < R < \infty$ and $I_{r,R} = \{z \in \partial\Sigma_{\psi}: r \le |z| \le R\}$. By Vitali's theorem, \[ \lim_{k\to\infty}\,\left(|f_{k, \gamma}(z)|+|f_{k, \gamma}'(z)|\right)=0 \] uniformly on $I_{r,R}$. Therefore the integrals \[ \int_{I_{r,R}} |f'_{k,\gamma}(z) g_\gamma(z)| \,|dz|, \quad \int_{I_{r,R}} |f_{k,\gamma}(z) g'_\gamma(z)| \,|dz|, \] tend to $0$ as $k\ \to \infty$. Moreover, \[ \int_{\partial\Sigma_\psi \setminus I_{r,R}} |f'_{k,\gamma}(z) g_\gamma(z)| \,|dz| \le \sup_{z \in \partial\Sigma_\psi \setminus I_{r,R}} |g_\gamma(z)| \, \|f_{k,\gamma}\|_{\mathcal{H}_\psi}, \] and \[ \int_{\partial\Sigma_\psi \setminus I_{r,R}} |f_{k,\gamma}(z) g'_\gamma(z)| \,|dz| \le \|f_{k,\gamma}\|_{H^\infty(\Sigma_\psi)} \int_{\partial\Sigma_\psi \setminus I_{r,R}} |g'_\gamma(z)| \,|dz|. \] Hence \[ \limsup_{k\to\infty} \| f_{k,\gamma}g_\gamma\|_{\mathcal{H}_\psi} \le B \left( \sup_{\partial\Sigma_\psi \setminus I_{r,R}} |g_\gamma(z)| + \int_{\partial\Sigma_\psi \setminus I_{r,R}} |g'_\gamma(z)| \,|dz| \right). \] Letting $r\to0$ and $R \to \infty$, we obtain \eqref{Se50B}. \end{proof} \section{The $\mathcal D$-calculus and its compatibility}\label{defD} Here we discuss functional calculus for sectorial operators $A$ of angle less than $\pi/2$ and functions $f \in \mathcal{D}_\infty$. Since $f$ is bounded on a closed sector containing the spectrum of $A$ (Corollary \ref{Repr}), $f(A)$ may be considered via the extended holomorphic (sectorial) calculus. If $A$ is injective then $f(A)$ can be defined that way as a closed operator, but we will show that $f(A)$ is a bounded operator when $f \in \mathcal{D}_\infty$. Our methods provide estimates for $\|f(A)\|$, and we will adapt the results in Section \ref{hardy_sobolev} to take account of the angle of sectoriality, by using fractional powers of operators (cf.\ Corollary \ref{gggC}). Recall that a densely defined operator $A$ on a Banach space $X$ is {\it sectorial of angle} $\theta \in [0,\pi/2)$ if $\sigma(A) \subset \overline \Sigma_\theta$ and, for each $\varphi \in (\theta,\pi]$, \begin{equation} \label{Aa0} M_{\varphi}(A) := \left\{\|z(z+A)^{-1}\| : z \in \Sigma_{\pi-\varphi}\right\} < \infty. \end{equation} The {\it sectorial angle} $\theta_A$ of $A$ is the minimal such $\theta$. Note that $M_\varphi(A)$ is a decreasing function of $\varphi$. Let $\Sect(\theta)$ stand for the class of all sectorial operators of angle $\theta$ for $\theta \in [0,\pi/2)$ on Banach spaces, and denote $\Sect(\pi/2-) := \bigcup_{\theta \in [0,\pi/2)} \Sect(\theta)$. Then $A \in \Sect(\pi/2-)$ if and only if $-A$ generates a (sectorially) bounded holomorphic $C_0$-semigroup on $X$ of angle $(\pi/2)-\theta_A$, in the sense that the semigroup has a holomorphic extension to $\Sigma_{(\pi/2)-\theta_A}$ which is bounded on each smaller subsector. Note that these semigroups are sometimes called {\it sectorially} bounded holomorphic semigroups in the literature. However, in this paper, we shall adopt the convention that bounded holomorphic semigroups are bounded on sectors. We shall denote the semigroup as $(e^{-tA})_{t\ge0}$, and $e^{-tA}$ then agrees with $e_t(A)$ defined in the Hille-Phillips calculus, where $e_t(z) = e^{-tz}$. One may consult \cite{HaaseB} for the general theory of sectorial operators, and \cite[Section 3.7]{ABHN} for the theory of holomorphic semigroups. Let $A$ be a closed, densely defined operator on a Banach space $X$ such that \begin{equation} \label{Aa} \sigma(A) \subset \overline\mathbb{C}_+ \qquad \text{and} \qquad M_A:= M_{\pi/2}(A) = \sup_{z\in \mathbb{C}_{+}}\,|z|\,\|(A+z)^{-1}\|<\infty. \end{equation} Then $\|A(z+A)^{-1}\| \le M_A+1, \,z\in \mathbb{C}_+$, and Neumann series (see \cite[Lemma 1.1]{V1}) imply that $\sigma(A) \subset \Sigma_\theta \cup \{0\}$ and \begin{equation}\label{Aaa1} \|z(z+A)^{-1}\|\le 2M_A,\qquad z\in \Sigma_{\pi-\theta}, \end{equation} where \[ \theta:=\arccos(1/(2M_A))<\pi/2. \] So $A \in \Sect(\theta) \subset \Sect(\pi/2-)$. Conversely, if $A \in \Sect(\theta)$ where $\theta \in [0,\pi/2)$, then \eqref{Aa} holds. Thus $-A$ generates a bounded holomorphic semigroup if and only if (\ref{Aa}) holds. The constant $M_A$ is a basic quantity associated with $A$, and we call it the {\it sectoriality constant} of $A$. Note that $M_{tA} = M_A$ for all $t>0$. A set $S$ of sectorial operators on the same Banach space $X$ is {\it uniformly sectorial} of angle $\theta$ if $S \subset \Sect(\theta)$ and, for each $\varphi \in (\theta,\pi)$, there exists $C_{\varphi}$ such that $M_\varphi(A) \le C_\varphi$ for all $A \in S$. Thus $S$ is uniformly sectorial of some angle $\theta<\pi/2$ if and only if each $A \in S$ satisfies \eqref{Aa} and $\sup_{A \in S} M_A < \infty$. In the presentation of the $\mathcal{D}$-calculus that follows, we assume that the reader is familiar with the holomorphic functional calculus for sectorial operators, as in \cite{HaaseB}, and in particular with the Hille-Phillips calculus for negative generators of bounded $C_0$-semigroups. We will make extensive use of fractional powers of sectorial operators in the form $(A+z)^{-\gamma}$ where $\gamma >0$. If $\gamma$ is not an integer, these operators are fractional powers which can be defined in many ways (see \cite{MS}), including using the holomorphic functional calculus (see \cite[Chapter 3]{HaaseB}). All these approaches are consistent with each other. Since $\mathcal{D}_\infty = \bigcup_{n=0}^\infty \mathcal{D}_n$, it is possible to define the $\mathcal{D}$-calculus without using fractional powers, and this would simplify some proofs (for example, Lemma \ref{fractional} becomes trivial, and the formulas \eqref{Sti} and \eqref{s1} would not be needed). Thus we could define the $\mathcal{D}$-calculus without using fractional powers, and in particular we could define the fractional powers $(z+A)^{-\gamma}$ for all $\gamma > 0$. This definition would be consistent with other definitions (see Theorem \ref{Compatible}). Then we could define the $\mathcal{D}_s$-calculus for all $s>-1$ in the way described below, using fractional powers in \eqref{formulaD}. The following simple lemma for fractional powers is a version of the moment inequality applied to the sectorial operator $(A+z)^{-1}$. \begin{lemma}\label{fractional} Let $A \in \operatorname{Sect}(\pi/2-)$, and $\gamma>0$. Let $\lceil\gamma\rceil$ be the ceiling function of $\gamma$, i.e.\ the smallest integer in $[\gamma,\infty)$. Then \[ \|(A+z)^{-\gamma}\|\le M_A^{\lceil\gamma\rceil}/|z|^\gamma,\quad z\in \mathbb{C}_{+}. \] \end{lemma} \begin{proof} When $\gamma \in \mathbb{N}$, the estimate is trivial. Let $\gamma \in (0,1)$. By the compatibility of our calculus with the holomorphic functional calculus for fractional powers, we may use the following standard Stiletjes formula (see \cite[(3.52)]{ABHN}, for example): \begin{equation}\label{Sti} (A+z)^{-\gamma}=\frac{\sin(\pi\gamma)}{\pi}\int_0^\infty t^{-\gamma}(A+z+t)^{-1}\,dt,\quad z\in \mathbb{C}_{+}. \end{equation} Next, let $z=\rho e^{i\varphi}\in \mathbb{C}_{+}$. Then, using Cauchy's theorem, \begin{align*} (A+z)^{-\gamma}&=\frac{\sin(\pi\gamma)}{\pi}\int_0^{e^{i\varphi \infty}} w^{-\gamma}(A+\rho e^{i\varphi}+w)^{-1}\,dw\\ &=e^{i(1-\gamma )\varphi}\frac{\sin(\pi\gamma)}{\pi}\int_0^\infty s^{-\gamma}(A+\rho e^{i\varphi}+s e^{i\varphi})^{-1}\,ds. \end{align*} So, \[ \|(A+z)^{-\gamma}\|\le M_A\frac{\sin(\pi\gamma)}{\pi} \int_0^\infty s^{-\gamma}(\rho+s)^{-1} \,ds = \frac{M_A}{\rho^\gamma}. \] In other cases, $\gamma = (\lceil\gamma\rceil - 1) + \delta$ where $\delta \in (0,1)$, and the estimate follows from the two previous cases. \end{proof} Now let $s >-1$ be fixed and let $f\in \mathcal{D}_s$. We define \begin{equation}\label{formulaD} f_{\mathcal{D}_s}(A):=f(\infty)- \frac{2^s}{\pi}\int_0^\infty \alpha^s\int_{-\infty}^\infty f'(\alpha+i\beta)(A+\alpha-i\beta)^{-(s+1)}\,d\beta\,d\alpha. \end{equation} Note that when $s = 1$ and $f \in \mathcal{B}$, \eqref{formulaD} coincides with the definition of $f(A)$ as given by the $\mathcal{B}$-calculus in \cite{BGT}, cf.\ \eqref{fcdef}. This definition is valid as the following simple proposition shows. \begin{prop}\label{bounded_s} Let $A \in \Sect(\pi/2-)$, and $s>-1$. \begin{enumerate}[\rm1.] \item The map $f \mapsto f_{\mathcal{D}_s}(A)$ is bounded from $\mathcal{D}_s$ to $L(X)$. \item For $f \in \mathcal{D}_s$, \begin{equation} \lim_{\epsilon\to0+}\,f_{\mathcal D_s}(A+\epsilon) =f_{\mathcal D_s}(A), \label{epsd11} \end{equation} in the operator norm topology. \end{enumerate} \end{prop} \begin{proof} Lemma \ref{fractional} and \eqref{formulaD} imply that \begin{equation}\label{estimateD} \|f_{\mathcal{D}_s}(A)\|\le |f(\infty)|+ \frac{2^s M_A^{\lceil{s+1}\rceil}}{\pi}\|f'\|_{\mathcal{V}_s}. \end{equation} Thus the boundedness of the map $f \mapsto f_{\mathcal{D}_s}(A)$ from $\mathcal{D}_s$ to $L(X)$ follows. By a standard variant of the Stieltjes representation \eqref{Sti} for fractional powers \cite[(3.56)]{ABHN}, \begin{equation} \label{s1} (A+z)^{-(s+1)} = \frac{1}{\Gamma(s+1)} \int_0^\infty t^s e^{-tz} e^{-tA} \,dt, \qquad z \in \mathbb{C}_+. \end{equation} By the dominated convergence theorem, $z \mapsto (A+z)^{-(s+1)}$ is continuous (even holomorphic) on $\mathbb{C}_+$ in the operator norm topology. The operators $(A+\epsilon)_{\epsilon\ge0}$ are uniformly sectorial of angle $\theta$, more precisely, $M_{A+\epsilon} \le M_A$ \cite[Proposition 2.1.1 f)]{HaaseB}. By Lemma \ref{fractional}, this implies that \begin{equation}\label{UnEE} \|(A+\epsilon+z)^{-(s+1)}\|\le M_A^{\lceil{s+1}\rceil} |z|^{-(s+1)}, \qquad \epsilon\ge 0, \quad z\in \mathbb C_{+}. \end{equation} Now \eqref{epsd11} follows from applying \eqref{formulaD} to $A+\epsilon$, letting $\epsilon\to0+$, and using the dominated convergence theorem. \end{proof} \begin{rem} \label{A+e} The property \eqref{epsd11} can be compared with Corollary \ref{SemDa} where a direct proof is given that the shifts form a bounded holomorphic $C_0$-semigroup on $\mathcal{D}_s$. To deduce \eqref{epsd11}, one also needs that if $f_\epsilon(z) = f(z+\epsilon)$, then $(f_\epsilon)_{\mathcal{D}_s}(A) = f_{\mathcal{D}_s}(A+\epsilon)$. By density of the rational functions in $\mathcal{D}_s$, it suffices that this holds for $f = r_\lambda, \, \lambda\in \mathbb{C}_+$, i.e., to show that $(r_\lambda)_{\mathcal{D}_s}(A) = (\lambda+A)^{-1}$. We show this in Theorem \ref{dcalculus}, but the argument uses \eqref{epsd11}. \end{rem} Let $f_{\text{HP}}(A)$ stand for a function $f$ of $A$ defined by the Hille-Phillips functional calculus when $f$ is in the Hille-Phillips algebra $\mathcal{LM}$, and let $f_{\operatorname{Hol}}(A)$ denote a function $f$ of $A$ given by the holomorphic functional calculus when $f$ is in the domain of that calculus. The following statement shows that both calculi are compatible with the $\mathcal D$-calculus, and moreover that the definitions of $f_{\mathcal{D}_s}(A)$ agree for the various values of $s$ for which $f \in \mathcal{D}_s$. \begin{thm}\label{Compatible} Let $A \in \Sect(\pi/2-)$, and let $f \in \mathcal{D}_s,\, s>-1$. \begin{enumerate}[\rm(i)] \item If $A$ is injective, then \begin{equation}\label{Ds11} f_{\mathcal{D}_s}(A)=f_{\mathrm{Hol}}(A). \end{equation} \item If $\sigma\ge s$ then \begin{equation}\label{Ds12} f_{\mathcal{D}_\sigma}(A)=f_{\mathcal{D}_s}(A). \end{equation} \item If $f \in \mathcal{LM} \cap \mathcal{D}_s$, then \begin{equation}\label{Ds13} f_{\mathcal{D}_s}(A)=f_{{\rm HP}}(A). \end{equation} In particular, \eqref{Ds13} holds if $f \in \mathcal{LM}$ and $s>0$. \end{enumerate} \end{thm} \begin{proof} We start by proving \eqref{Ds11}. Assume that $A$ is injective, and $A \in \Sect(\theta)$, where $\theta \in (0, \pi/2)$. Let $\psi \in (\theta, \pi/2)$. Let $f \in \mathcal{D}_s, \, s>-1$, and assume (without loss of generality) that $f(\infty)=0$. By the definition of the holomorphic functional calculus, \[ A(1+A)^{-2}f_{\operatorname{Hol}}(A)=\frac{1}{2\pi i}\int_{\partial \Sigma_\psi} \frac{\lambda f(\lambda)}{(\lambda+1)^2}(\lambda-A)^{-1}\,d\lambda. \] Since \[ f(\lambda)=-\frac{2^s}{\pi}\int_0^\infty \alpha^s\int_{-\infty}^\infty\frac{f'(\alpha+i\beta)} {(\lambda+\alpha-i\beta)^{s+1}}\,d\beta\,d\alpha,\quad \lambda\in \mathbb{C}_{+}, \] Fubini's theorem and Cauchy's theorem imply that \begin{align*} \lefteqn{A(1+A)^{-2}f_{\operatorname{Hol}}(A) \hskip10pt} \\ &=- \frac{2^s}{\pi} \int_0^\infty \alpha^s \int_{-\infty}^\infty f'(\alpha+i\beta) \left(\frac{1}{2\pi i} \int_{\partial \Sigma_\psi}\,\frac{\lambda(\lambda-A)^{-1}}{(\lambda+1)^2(\lambda+\alpha-i\beta)^{s+1}}\,d\lambda\right) d\beta \, d\alpha\\ &= -\frac{2^s}{\pi}A(1+A)^{-2} \int_0^\infty\alpha^s \int_{-\infty}^\infty f'(\alpha+i\beta)(A+\alpha-i\beta)^{-(s+1)} d\beta \, d\alpha\\ &=A(1+A)^{-2}f_{\mathcal{D}_s}(A). \end{align*} Hence (\ref{Ds11}) holds. Now we no longer assume that $A$ is injective. We infer by \eqref{Ds11} that \[ f_{\mathcal{D}_s}(A+\epsilon)=f_{\mathcal{D}_\sigma}(A+\epsilon) \] for all $\epsilon>0$ and $\sigma \ge s$. Letting $\epsilon\to 0$ and using Proposition \ref{bounded_s}, we obtain the assertion (\ref{Ds12}). Finally, if $f\in \mathcal{LM} \cap \mathcal D_s$ for some $s > -1$, then $f\in \mathcal{B}\cap \mathcal D_s$, and using \eqref{Ds12}, we have \[ f_{\text{HP}}(A)=\Phi_A(f)=f_{\mathcal{D}_1}(A)=f_{\mathcal{D}_s}(A). \qedhere \] \end{proof} \begin{rem}\label{coincid} If $f$ has zero polynomial limits at zero and at infinity in the sense of \cite[p.27]{HaaseB}, then the proof above does not require the regularisation factor $\lambda(\lambda+1)^{-2}$. Hence $f_{\mathcal D}(A)=f_{\mathrm{Hol}}(A)$ regardless of the injectivity of $A$. One can show that $f_{\mathcal D}(A)=f_{\mathrm{Hol}}(A)$ even when $f$ belongs to the extended Riesz-Dunford function class, (for example, $f(z)=e^{-tz},\, t >0$), but we omit a discussion of this here (cf.\ the proof of Lemma \ref{resap}). \end{rem} Recall that $\mathcal D_s \subset \mathcal D_\sigma$ if $-1< s \le \sigma$, and the space \[\mathcal D_\infty=\bigcup_{s>-1} \mathcal D_s\] is an algebra, by Lemma \ref{Alg1}. Thus it is a plausible and natural domain for a functional calculus, which we now define. \begin{thm}\label{dcalculus} Let $A \in \Sect(\pi/2-)$. The formula \eqref{formulaD} defines an algebra homomorphism: \begin{eqnarray*} \Psi_A : \mathcal D_{\infty} \mapsto L(X), \qquad \Psi_A(f)=f_{\mathcal{D}_s}(A), \qquad f \in \mathcal{D}_s, \, s>-1. \end{eqnarray*} Moreover, \begin{enumerate}[\rm(i)] \item $\Psi_A(r_\lambda) = (\lambda+A)^{-1}, \quad \lambda \in \mathbb{C}_+$. \item $\Psi_A$ is bounded in the sense that for every $s>-1$ there exists $C_s(A)$ such that \begin{equation}\label{bounded} \|\Psi_A(f)\|\le |f(\infty)| + C_s(A)\|f\|_{\mathcal D_s}, \qquad f \in \mathcal{D}_s. \end{equation} Specifically, \eqref{estimateD} holds. \end{enumerate} Moreover, $\Psi_A$ is the unique algebra homomorphism from $\mathcal{D}_\infty$ to $L(X)$ which satisfies {\rm(i)} and {\rm(ii)}. \end{thm} The homomorphism $\Psi_A$ will be called the {\it $\mathcal D$-calculus}. \begin{proof} It follows from \eqref{Ds12} that $\Psi_A$ is well-defined by \eqref{formulaD}, and from \eqref{Ds11} that $(r_\lambda)_{\mathcal{D}_s}(A+\epsilon) = (\lambda+\epsilon+A)^{-1}$ for $\epsilon>0$. Letting $\epsilon \to 0$ and using \eqref{epsd11} gives (i). Moreover \eqref{bounded} is a direct consequence of Proposition \ref{bounded_s}. We will now prove that $\Psi_A$ is a homomorphism. Let $f,g\in \mathcal D_\infty$. Then $f\in \mathcal D_r$ and $g\in \mathcal D_t$ for some strictly positive $s$ and $t$, hence $f g\in\mathcal D_{s+t+1}$ by Lemma \ref{Alg1}. Since $\mathcal {LM}$ is dense in $\mathcal D_s$ for every $s>0$ by Corollary \ref{Bdens}, there exist $(f_n)_{n=1}^\infty$ and $(g_n)_{n=1}^\infty$ from $\mathcal {LM}$ such that \[ f_n \to f \quad \text{in} \quad \mathcal D_s \qquad \text{and} \qquad g_n \to g \quad \text{in}\quad \mathcal D_t, \qquad n \to \infty, \] and then, in view of Lemma \ref{Alg1}, $f_n g_n \to f g$ in $\mathcal D_{s+t+1}$ as $n \to \infty$. By the product rule for the HP-calculus and \eqref{Ds13}, \[ \Psi_A(f_n g_n)=(f_n g_n)_{\mathrm{HP}}(A)=\Psi_A(f_n)\Psi_A(g_n), \qquad n \ge 1. \] Passing to the limit when $n\to\infty$ and using \eqref{bounded}, it follows that \[ \Psi_A(fg)=\Psi_A(f)\Psi_A(g). \] Let $\Psi : \mathcal{D}_\infty \to L(X)$ be an algebra homomorphism satisfying (i) and (ii). Then $\Psi$ and $\Psi_A$ coincide on $\{r_\lambda : \lambda \in \mathbb{C}_+\}$. Since $A$ is densely defined, $\Psi(1)=1$ (see \cite[Section 6, Observation (2)]{BGT2}, so $\Psi$ and $\Psi_A$ coincide on the span of these functions which is dense in $\mathcal{D}_s$ (Theorem \ref{D00}). Since $\Psi$ and $\Psi_A$ are both bounded on $\mathcal{D}_s$, it follows that they coincide on each $\mathcal{D}_s$ and hence on $\mathcal{D}_\infty$. \end{proof} \begin{remark} \label{Dsn} If $A$ is an operator for which a $\mathcal{D}$-calculus exists with the properties (i) and (ii) given in Theorem \ref{dcalculus}, then $A\in\Sect(\pi/2-)$. This follows from \eqref{r_l} and the properties (i) and (ii). By combining this with Theorem \ref{dcalculus}, we obtain Theorem \ref{dcalculus_intro}. Note also that, if (i) holds for some $\lambda\in\mathbb{C}_+$, then it holds for all $\lambda\in\mathbb{C}_+$, by the resolvent identity. \end{remark} The Banach algebras $\mathcal{D}_s^\infty, \, s>-1$, are subalgebras of $\mathcal{D}_\infty$, so we obtain the following corollary by restricting the $\mathcal{D}$-calculus. \begin{cor}\label{Compat} Let $A \in \Sect(\pi/2-)$ and $s>-1$. Then there exists a bounded homomorphism $\Psi_A^s: \mathcal D^\infty_s \mapsto L(X)$ such that \begin{enumerate}[\rm(i)] \item $\Psi_A^s(r_\lambda)=(\lambda+A)^{-1}, \quad \lambda \in \mathbb C_+$. \item $\Psi^s_A$ is bounded in the $\mathcal{D}_s$-norm, i.e., there exists $C_s(A)$ such that \begin{equation*} \|\Psi_A(f)\|\le C_s(A) \|f\|_{\mathcal D_s}, \qquad f \in \mathcal{D}_s. \end{equation*} \end{enumerate} Moreover $\Psi_A^s$ is the unique algebra homomorphism from $\mathcal{D}_s^\infty$ to $L(X)$ which satisfies {\rm(i)} and {\rm(ii)}. \end{cor} From now onwards, we will write $f_\mathcal{D}(A)$ instead of $\Psi_A(f)$, for $f \in \mathcal{D}_\infty$. When $f(z) = e^{-tz}$, we will continue to use the notation $e^{-tA}$ for $f(A)$, since the $\mathcal D$-calculus agrees with the HP-calculus by Theorem \ref{Compatible}. So $(e^{-tA})_{t\ge 0}$ form the $C_0$-semigroup generated by $-A$, and it extends to a bounded holomorphic semigroup. Let $g : \mathbb{C}_+ \to \mathbb{C}_+$ be a holomorphic function and assume that $r_\lambda \circ g \in \mathcal{D}_\infty$ for all $\lambda \in \mathbb{C}_+$. Since the functions $(r_\lambda)_{\lambda\in\mathbb{C}_+}$ satisfy the resolvent identity, the operators \[ ((r_\lambda \circ g)_\mathcal{D}(A))_{\lambda\in\mathbb{C}_+} \subset L(X) \] also satisfy the resolvent identity, i.e., they form a pseudo-resolvent. In particular their kernels and their ranges are independent of $\lambda$, and they form the resolvent of an operator $B$ if and only if the common kernel is $\{0\}$, and the domain of $B$ is the common range of the pseudo-resolvent (see \cite[Section VIII.4]{Yos}). \begin{cor} \label{dcomp} Let $A, B \in \Sect(\pi/2-)$, and let $g : \mathbb{C}_+ \to \mathbb{C}_+$ be holomorphic. Assume that, for each $s>-1$, there exists $\sigma>-1$ such that \begin{enumerate}[\rm(a)] \item For all $f \in \mathcal{D}_s$, $f\circ g \in \mathcal{D}_\sigma$, and \item For all $\lambda\in\mathbb{C}_+$, $(r_\lambda \circ g)_\mathcal{D}(A) = (\lambda+B)^{-1}$. \end{enumerate} Then $(f \circ g)_\mathcal{D}(A) = f_\mathcal{D}(B)$ for all $f \in \mathcal{D}_\infty$. \end{cor} \begin{proof} By assumption (a), Corollary \ref{Cangle} and the Closed Graph Theorem, $f \mapsto f \circ g$ is a bounded map from $\mathcal{D}_s$ to $\mathcal{D}_\sigma$. Moreover the $\mathcal{D}$-calculus for $A$ is a bounded map from $\mathcal{D}_\sigma$ to $L(X)$. Hence the composition is a bounded map from $\mathcal{D}_s$ to $L(X)$, and by assumption (b) it sends $r_\lambda$ to $(\lambda+B)^{-1}$ for all $\lambda \in \mathbb{C}_+$. Moreover the maps collectively form an algebra homomorphism from $\mathcal{D}_\infty$ to $L(X)$. By the uniqueness in Theorem \ref{dcalculus}, this map is the $\mathcal{D}$-calculus for $B$. \end{proof} In the context of Corollary \ref{dcomp}, it is natural to write $g(A)$ for the operator $B$. There is also a version of Corollary \ref{dcomp} for a single value of $s$ and $\sigma$, and the $\mathcal{D}_s^\infty$-calculus. \begin{exas} \label{gcomp} Examples of functions $g$ and operators $B$ which satisfy the conditions of Corollary \ref{dcomp} include the following. \begin{enumerate}[\rm1.] \item $g(z) = z^{-1}$, if $A$ is injective (with dense range); $\sigma=s$, $B=A^{-1}$. Then $f \circ g$ is the function $\tilde f \in \mathcal{D}_s$ as in Lemma \ref{leminv}. Note that $\tilde f_\mathcal{D}(A)$ is defined as a bounded operator on $X$, even if $A$ is not injective. If $A_0$ is the restriction of $A$ to $X_0$, the closure of the range of $A$, then $\tilde f_\mathcal{D}(A)$ acts as $A_0^{-1}$ on $X_0$ and as the sectorial limit $f(\infty)$ on the kernel of $A$. If $X$ is reflexive this determines $\tilde f_\mathcal{D}(A)$ on $X$. \item $g(z) = tz$, where $t>0$; $\sigma=s$, $B = tA$. See Lemma \ref{leminv}. \item $g(z) = z + \eta$, where $\eta \in \mathbb{C}_+$; $\sigma=s$, $B = A +\eta$. See Remark \ref{A+e}. \item $g(z) = z^\gamma$, where $\gamma \in (0,1)$; $s>\-1,\sigma>-1$; $B = A^\gamma$ (as defined in the holomorphic functional calculus). See Corollary \ref{gggC} for the assumption (a) in Corollary \ref{dcomp}, and Corollary \ref{fpD} below for the assumption (b). The result also holds for $\gamma \in (1, \pi/(2\theta))$. \item Examples (2), (3) for $\eta \in \mathbb{R}_+,$ and (4) above are Bernstein functions. By Lemma \ref{bsds}, $r_\lambda \circ g \in \mathcal{D}_s$ for all Bernstein functions $g$ and $s>2$. We will show in the proof of Theorem \ref{BF1} that $(r_\lambda\circ g)_D(A) = (\lambda+g(A))^{-1}$, where $g(A)$ is a sectorial operator. \end{enumerate} \end{exas} In Example \ref{gcomp}(4) above, we have introduced a fractional power $A^\gamma$, where $\gamma>0$. These operators are defined in various ways, including the extended holomorphic functional calculus. To justify the example, we need the following lemma about fractional powers, which is probably known at least in simpler form. For $\gamma \in (0,1)$ and $\nu\in\mathbb{N}$, it follows easily from a standard result \cite[Proposition 5.1.4]{MS}. We give a proof that uses the holomorphic functional calculus for fractional powers as in \cite[Section 3.1]{HaaseB}. \begin{lemma} \label{resap} Let $A \in \Sect(\pi/2-)$, $\gamma \in (0,1)$, and $\nu>0$. In the operator norm topology, \[ \lim_{\epsilon\to0+} ((A+\epsilon)^\gamma + z)^{-\nu} = (A^\gamma+z)^{-\gamma}, \quad z\in \mathbb{C}_+. \] \end{lemma} \begin{proof} Let $z \in \mathbb{C}_+$ be fixed, and let \[ f(\lambda) = \left(\lambda^\gamma + z\right)^{-\nu}, \qquad \lambda\in \overline{\mathbb{C}}_+. \] Then $f \in H^\infty(\mathbb{C}_+)$, and, by considering the derivative of $\mu \mapsto (\mu+z)^{-\nu}$, we see that there exists a constant $C$ (depending on $z$) such that \begin{equation*}\label{f2} |f(\lambda)-f(0)| \le C |\lambda|^{\gamma}, \quad |f(\lambda)| \le C |\lambda|^{-\gamma\nu}, \qquad \lambda\in\mathbb{C}_+. \end{equation*} Thus $f$ has polynomial limits at $0$ and $\infty$, and so $f$ belongs to the extended Riesz-Dunford class defined in \cite[Lemma 2.2.3]{HaaseB}. In other words, \[ g_0(\lambda) := f(\lambda) - z^{-\nu}(1+\lambda)^{-1} \] has polynomial decay at $0$ and $\infty$. Moreover, there exists a constant $C'$ (independent of $\epsilon$) such that \begin{equation}\label{f3} |f(\lambda+\epsilon)-f(\epsilon)| \le C' |\lambda|^{\gamma}, \quad |f(\lambda+\epsilon)| \le C' |\lambda|^{-\gamma\nu}, \qquad \lambda\in\mathbb{C}_+, \, \epsilon \in(0,1). \end{equation} Let \[ g_\epsilon(\lambda) = f(\lambda+\epsilon) - f(\epsilon)(1+\lambda)^{-1}. \] Using the definition of the primary functional calculus \cite[Section 2.3.1]{HaaseB}, we have \begin{align*} f(A+\epsilon)-f(A)&= g_\epsilon(A)-g_0(A) + (f(\epsilon) -f(0))(I+A)^{-1}, \\ g_\epsilon(A)-g_0(A)&=\frac{1}{2\pi i} \int_{\Sigma_\psi} \left(g_\epsilon(\lambda)-g_0(\lambda)\right) (\lambda-A)^{-1} \,d\lambda, \end{align*} where $\psi \in (\theta,\pi/2)$. By the dominated convergence theorem, \[ \lim_{\epsilon \to 0+} \|(g_\epsilon(A) - g_0(A)\|=0. \] The pointwise convergence of $g_\epsilon -g_0$ to zero is clear, and the existence of an integrable majorant follows easily from \eqref{f3}. \end{proof} \begin{cor} \label{fpD} Let $A\in\Sect(\pi/2-)$, and $\gamma \in (0,1)$. Let $f \in \mathcal{D}_\infty$ and $h(z) = f(z^\gamma)$. Then \begin{enumerate}[\rm1.] \item In operator norm, \[ \lim_{\epsilon\to0+} f_{\mathcal{D}}((A+\epsilon)^\gamma) = f_{\mathcal{D}}(A^\gamma), \] \item $h_{\mathcal{D}}(A) = f_{\mathcal{D}}(A^\gamma)$. \end{enumerate} \end{cor} \begin{proof} The proof of (1) follows from Lemma \ref{resap} in essentially the same way as the last paragraph of the proof of Proposition \ref{bounded_s}. By Corollary \ref{gggC}, $h \in \mathcal{D}_\infty$. By \eqref{Ds11} and the Composition Theorem for the holomorphic functional calculus \cite[Theorem 2.4.2]{HaaseB}, we have $h_{\mathcal{D}}(A+\epsilon) = f_{\mathcal{D}}((A+\epsilon)^\gamma)$. Letting $\epsilon\to0+$ and using \eqref{epsd11} and (1), we obtain (2). \end{proof} \section{The calculus on Hardy-Sobolev algebras}\label{hardy_sobolev} Given the negative generator $A$ of a bounded holomorphic $C_0$-semigroup on a Banach space $X$, the $\mathcal D$-calculus allows us to extend the $\mathcal B$-calculus to a much larger class of functions. A drawback of the $\mathcal D$-calculus is that it does not respect the sectoriality angle of $A$, so the results within the $\mathcal D$-calculus are independent of the sectoriality angle and confined to holomorphic functions on $\mathbb{C}_+$. To remedy that problem, we introduce in this section a version of the $\mathcal D$-calculus adjusted to an appropriate Hardy-Sobolev algebra on a sector in the right half-plane. While the Hardy-Sobolev algebra has a ``stronger'' norm, it appears to be an adequate substitute for $\mathcal D_\infty$ in the setting of sectors, and it has significant applications, as we shall see in Section \ref{norm_estimates}. The basic idea is a very simple change of variable in the $\mathcal D$-calculus. If $\Psi^s_A$ is the $\mathcal D_s^\infty$-calculus for a sectorial operator $A$, then one sets $\Upsilon_A(f):= \Psi^s_{A^\gamma}(f_{1/\gamma})$ for appropriate values of $\gamma$, determined by the sectoriality angle $\theta_A$ of $A$. This definition does not depend on the precise choice of $\gamma$, by Corollary \ref{fpD}. The definitions also agree for different $s>-1$ by Theorem \ref{Compatible}(ii), and we set $s=0$ for convenience. As we show below this eventually leads to a new calculus for Hardy-Sobolev algebras on sectors. Throughout this section, we assume that $A\in\Sect(\theta)$, $0<\theta<\psi<\pi$, and we let $\gamma:=\pi/(2\psi)$, and, as before, \[ M_\psi(A):=\sup_{z\in \Sigma_{\pi-\psi}}\,\|z(A+z)^{-1}\|. \] Then $A^\gamma\in \Sect(\pi/2-)$, and $\theta_{A^\gamma} =\gamma\theta_A \in (\theta_A/2,\pi/2)$ \cite[Proposition 3.1.2]{HaaseB}. We are particularly interested in cases where $\psi$ is close to $\theta$, so that $\mathcal{H}_\psi$ is as large as possible. \subsection{The operator $f(A)$ for $f\in \mathcal{H}_\psi$} Recall from Lemmas \ref{simple} and \ref{Dh1}(i) that if $f\in \mathcal{H}_\psi$ and $f_{1/\gamma}(z):=f(z^{1/\gamma})$ we have \[ f'_{1/\gamma}\in H^1(\mathbb{C}_{+}),\qquad f_{1/\gamma}(\infty)=f(\infty), \] and consequently $f_{1/\gamma}\in \mathcal{D}_0$. Together with Proposition \ref{FrepH} this motivates the following definition of the operator $f_{\mathcal H}(A)$ by means of the $\mathcal{D}$-calculus applied to $A^{\gamma}$: \begin{equation}\label{sigma_def} f_{\mathcal H}(A):=f_{1/\gamma}(\infty)-\frac{1}{\pi}\int_0^\infty\int_{-\infty}^\infty f'_{1/\gamma}(\alpha+i\beta) (A^{\gamma}+\alpha-i\beta)^{-1}\,d\beta\,d\alpha. \end{equation} The right-hand side of \eqref{sigma_def} converges in the uniform operator topology, and, by \eqref{estimateD} and \eqref{embedh}, \begin{equation}\label{A2Dop} \|f_{\mathcal H}(A)\|\le |f(\infty)|+\frac{M_{A^\gamma}}{\pi}\|f_{1/\gamma}\|_{\mathcal{D}_0,0} \le |f(\infty)|+ M_{A^\gamma} \|f'\|_{H^1(\psi)} \le M_{A^\gamma} \|f\|'_{\mathcal{H}_\psi}. \end{equation} If $f\in \mathcal{H}_\psi$ and $A$ is injective, then $f_{\operatorname{Hol}}(A)$ can also be defined using the holomorphic functional calculus and the Composition Rule within it \cite[Theorem 2.4.2]{HaaseB}: \[ f_{\operatorname{Hol}}(A)=f_{1/\gamma,\operatorname{Hol}}(A^\gamma) :=-\frac{1}{2\pi i}[A^\gamma(1+A^{\gamma})^{-2}]^{-1} \int_{\partial \Sigma_\omega}\,\frac{\lambda f(\lambda^{1/\gamma})}{(\lambda+1)^2}(\lambda-A^\gamma)^{-1}\,d\lambda, \] for $0<\theta<\omega<\psi$. The following proposition shows that our definition \eqref{sigma_def} of $f_{\mathcal H}(A)$ coincides with $f_{\operatorname{Hol}}(A)$, when $A$ is injective, and various other properties are easily deduced from the definition above and corresponding properties of the $\mathcal{D}$-calculus. \begin{prop} \label{Hprops} Let $f \in \mathcal{H}_\psi$, and $A \in \Sect(\theta)$, where $0 \le \theta < \psi < \pi$. \begin{enumerate}[\rm(i)] \item $f_\mathcal{H}(A)$ does not depend on the choice of $\psi$. \item If $\nu \in (0,1)$ and $g(z) = f(z^\nu)$, then $g_\mathcal{H}(A) = f_\mathcal{H}(A^\nu)$. \item If $f \in \mathcal{D}_\infty$ and $\theta < \pi/2$, then $f_\mathcal{H}(A) = f_\mathcal{D}(A)$. \item If $A$ is injective, then $f_\mathcal{H}(A) = f_{\operatorname{Hol}}(A)$. \item In the operator norm topology, $\lim_{\epsilon\to0+} f_\mathcal{H}(A+\epsilon) = f_\mathcal{H}(A)$. \end{enumerate} \end{prop} \begin{proof} Statements (i), (ii) and (iii) follow from Corollary \ref{fpD}(ii) and Lemma \ref{simple}. Statement (iv) follows from (iii) and Theorem \ref{Compatible}(i). Statement (v) follows from Proposition \ref{bounded_s}. An alternative direct proof of (iv) can be given as follows. We may assume that $f(\infty)=0$. Since $f_{1/\gamma}(z)=f(z^{1/\gamma})\in \mathcal{D}_0$, Corollary \ref{Repr} gives \[ f_{1/\gamma}(z)= -\frac{1}{\pi}\int_0^\infty \int_{-\infty}^\infty\frac{f_{1/\gamma}'(\alpha+i\beta)}{z+\alpha-i\beta}\,d\beta\,d\alpha,\qquad z\in \mathbb{C}_{+}. \] Using Fubini's theorem and some basic properties of the holomorphic functional calculus, we obtain \begin{align*} &\lefteqn{\null\hskip-5pt A^\gamma(1+A^\gamma)^{-2}f_{\operatorname{Hol}}(A)}\\ &=\frac{1}{\pi}\int_0^\infty \int_{-\infty}^\infty f_{1/\gamma}'(\alpha+i\beta) \left(\frac{1}{2\pi i} \int_{\partial \Sigma_\omega}\,\frac{\lambda(\lambda-A^\gamma)^{-1}}{(\lambda+1)^2(\lambda+\alpha-i\beta)}\,d\lambda\right) d\beta \,d\alpha\\ &= -A^\gamma(1+A^\gamma)^{-2}\left( \frac{1}{\pi}\int_0^\infty \int_{-\infty}^\infty f_{1/\gamma}'(\alpha+i\beta)(A^\gamma+\alpha-i\beta)^{-1} d\beta\, d\alpha\right)\\ &=A^\gamma(1+A^\gamma)^{-2}f_{\mathcal{H}}(A), \end{align*} and (iv) follows. \end{proof} Now we can formally define the $\mathcal H$-calculus. \begin{thm}\label{sigma_cal} Let $A \in \Sect(\theta)$, where $\theta \in (0,\pi/2)$. For any $\psi \in (\theta, \pi)$ the formula \eqref{sigma_def} defines a bounded Banach algebra homomorphism: \begin{eqnarray*} \Upsilon_A : \mathcal \mathcal{H}_\psi \mapsto L(X), \qquad \Upsilon_A(f)=f_\mathcal{H}(A). \end{eqnarray*} The homomorphism $\Upsilon_A$ satisfies $\Upsilon_A(r_\lambda)=(\lambda+A)^{-1}$ for all $\lambda \in \Sigma_{\pi-\psi}$, and it is the unique homomorphism with these properties. \end{thm} The homomorphism $\Upsilon_A$ will be called the \emph{$\mathcal H$-calculus} for $A$. \begin{proof} The boundedness of $\Upsilon_A$ follows from either \eqref{NewEs} or \eqref{A2Dop}. The homomorphism property is implied by Theorem \ref{Compat}. Indeed, employing the functional calculus $\Psi^0_A$ on $\mathcal D^\infty_0$ given by Corollary \ref{Compat}, one has \begin{align*} \Upsilon_A(fg) &= \Psi^0_A((fg)_{1/\alpha}) = \Psi^0_A(f_{1/\alpha}g_{1/\alpha}) \\ &= \Psi^0_A(f_{1/\alpha}) \Psi^0_A(g_{1/\alpha}) = \Upsilon_A(f)\Upsilon_A(g). \end{align*} The uniqueness follows from Theorem \ref{DM122}. \end{proof} \begin{remark} If $A$ is any operator for which there is an $\mathcal{H}_\psi$-calculus as in Theorem \ref{sigma_cal}, then $A \in \Sect(\theta)$ for some $\theta\in(0,\psi)$. This follows from \eqref{res_sigma_cal_intro}, and in combination with Theorem \ref{sigma_cal} this yields the proof of Theorem \ref{sigma_cal_intro}. If $\Upsilon_A(r_\lambda) = (\lambda+A)^{-1}$ for some $\lambda\in \Sigma_{\pi-\psi}$ then this holds for all $\lambda \in \Sigma_{\pi-\psi}$, by the resolvent identity. \end{remark} \subsection{The operator $\arccot(A^\gamma)$ and the $\arccot$-formula}\label{alternative} In this section, we derive an alternative to the formula \eqref{sigma_def} for the $\mathcal{H}$-calculus, in the form of an operator counterpart of Proposition \ref{L7.4} for scalar functions. In addition to its intrinsic interest, it helps us to compare our approach with the approach developed by Boyadzhiev \cite{Boy}, as we do at the end of this section. We introduce as an operator kernel the function \[ g(z) :=\arccot(z) = \frac{1}{2i} \log \left(\frac{z+i}{z-i}\right), \quad z \in \mathbb{C}_+, \] already considered in Example \ref{arc}. Note that $g \in \mathcal{D}_0$, $ g(\infty)=0$, $g'(z)=-(z^2+1)^{-1}$, and \eqref{ArcN} holds: \[ \arccot(z)=\frac{1}{\pi} \int_0^\infty\int_{-\infty}^\infty\frac{(z+\alpha-i\beta)^{-1}}{(\alpha+i\beta)^2+1}\,d\beta\,d\alpha\qquad z\in \mathbb{C}_{+}. \] Let $A \in \Sect(\theta)$, where $\theta \in [0,\pi)$. Let $\psi \in (\theta,\pi)$ and $\gamma=\pi/(2\psi)$. By the $\mathcal{D}$-calculus in \eqref{formulaD} and \eqref{estimateD}, \begin{equation}\label{defcot} \arccot(A^\gamma) := \arccot_{\mathcal{D}}(A^\gamma)=\frac{1}{\pi} \int_0^\infty\int_{-\infty}^\infty\frac{(A^\gamma+\alpha-i\beta)^{-1}}{(\alpha+i\beta)^2+1}\,d\beta\,d\alpha, \end{equation} where the integral converges in the operator norm, and \begin{equation}\label{NormA} \|\arccot(A^\gamma)\|\le\frac{M_{A^\gamma}}{\pi}\|\arccot z\|_{\mathcal{D}_0} < 3M_{A^\gamma},\;\; M_{A^\gamma}:=\sup_{z\in \mathbb{C}_{+}}\,\|z(A^\gamma+z)^{-1}\|. \end{equation} We will provide an alternative estimate for the operator $\arccot (A^\gamma/s^\gamma)$. Following (\ref{rew}), we may formally write \begin{align} \label{acint} \arccot_{{\rm int}}(A^{\gamma})&:=\frac{1}{4\pi}\int_0^\infty \log\left|\frac{1+t^\gamma}{1-t^\gamma}\right| \left((t-e^{i\psi}A)^{-1} +(t-e^{-i\psi}A)^{-1}\right)\,dt\\ &\null\hskip30pt +\frac{1}{4 i}\int_{|\lambda|=1,\,\arg\lambda\in (\psi,2\pi-\psi)}\, (\lambda-A)^{-1}\,d\lambda. \notag \end{align} \begin{lemma}\label{arc_bound} Let $A \in \Sect(\theta)$ and $\gamma = \pi/(2\psi)$, where $0 \le \theta < \psi < \pi$. The operator $\arccot_{{\rm int}}(A^{\gamma})$ is well-defined and \begin{equation}\label{NesT} \|\arccot_{{\rm int}}(A^{\gamma})\|\le M_\psi(A)\frac{\pi}{2}. \end{equation} \end{lemma} \begin{proof} Using (\ref{remA}), one notes that \begin{align*} \|\arccot(A^{\gamma})\|&\le \frac{M_\psi(A)}{2\pi}\int_0^\infty \log\left|\frac{1+t^\gamma}{1-t^\gamma}\right| \,\frac{dt}{t} +\frac{M_\psi(A)}{4}\int_{|\lambda|=1,\; \arg \lambda\in (\psi,2\pi-\psi)}\,\frac{|d\lambda|}{|\lambda|}\\ &=M_\psi(A)\left(\frac{\psi}{2}+ \frac{(\pi-\psi)}{2}\right) =M_\psi(A)\frac{\pi}{2}. \qedhere \end{align*} \end{proof} The next lemma shows that the formula $\arccot_{{\rm int}}(A^{\gamma})$ coincides with the definition of $\arccot (A^{\gamma})$ by the $\mathcal D$-calculus. When $A$ is injective, $\arccot(A^\gamma)$ is defined in the holomorphic functional calculus by \begin{equation}\label{archol} \arccot_{\operatorname{Hol}}(A^\gamma):=-\frac{1}{2\pi i}[A(1+A)^{-1}]^{-1} \int_{\partial\Sigma_\omega}\,\frac{\lambda\arccot(\lambda^\gamma)}{\lambda+1}(\lambda-A)^{-1}\,d\lambda, \end{equation} where $0<\theta<\omega<\psi<\pi$. \begin{lemma}\label{Ss2} Under the assumptions above, $\arccot_{{\rm int}}(A^\gamma)=\arccot_{\mathcal{D}}(A^\gamma)$. If $A$ is injective, then $\arccot_{{\rm int}}(A^{\gamma})=\arccot_{\operatorname{Hol}}(A^{\gamma})$. \end{lemma} \begin{proof} Assume first that $A$ is injective. Using \eqref{archol}, (\ref{rew}) and \eqref{defcot} we obtain \begin{align*} \lefteqn{A(1+A)^{-1} \arccot_{\rm{Hol}} (A^\gamma)} \\ &= - \frac{1}{2\pi i} \int_{\partial\Sigma_\omega}\,\frac{z\arccot(z^\gamma)}{z+1}(z-A)^{-1}\,dz\\ &=\frac{1}{4\pi}\int_0^\infty \log\left|\frac{1+t^\gamma}{1-t^\gamma}\right| \left(-\frac{1}{2\pi i} \int_{\partial\Sigma_\omega}\,\frac{z}{z+1}(z-A)^{-1}\, V_\psi(z, t)\,dz \right)\,\frac{dt}{t}\\ &\hskip30pt + \frac{1}{4i}\int_{|\lambda|=1,\,\arg \lambda\in (\psi,2\pi-\psi)}\,\left( -\frac{1}{2\pi i}\int_{\partial\Sigma_\omega} \frac{z (z-A)^{-1}}{(z+1)(\lambda-z)} dz \right) d \lambda \\ &=A(A+1)^{-1}\frac{1}{4\pi}\int_0^\infty \log\left|\frac{1+t^\gamma}{1-t^\gamma}\right| \left((t-e^{i\psi}A)^{-1}+(t-e^{-i\psi}A)^{-1}\right)\,dt\\ &\null\hskip30pt +A(A+1)^{-1}\frac{1}{4i}\int_{|\lambda|=1,\,\arg \lambda\in (\psi,2\pi-\psi)}\, (\lambda-A)^{-1}\, d \lambda\\ &=A(A+1)^{-1}\arccot_{{\rm int}}(A^{\gamma}). \end{align*} Thus the second statement holds. Moreover, $\arccot_{\operatorname{Hol}}(A^\gamma)=\arccot_{\mathcal{D}}(A^\gamma)$, by Theorem \ref{Compat}, and the first statement follows. If $A$ is not injective, we have $\arccot_{\rm int}((A+\epsilon)^\gamma) = \arccot_\mathcal{D}((A+\epsilon)^\gamma)$ from the case above. When $\epsilon\to0+$, the left-hand side converges in operator norm to $\arccot_{\rm int}(A^\gamma)$ by applying the dominated convergence theorem in \eqref{acint}, and the right-hand side converges to $\arccot_{\mathcal{D}}(A^\gamma)$ by Proposition \ref{bounded_s}. \end{proof} \begin{thm}\label{Sovp} Let $A \in \Sect(\theta)$ and $\gamma=\pi/(2\psi)$. If $f \in \mathcal{H}_\psi$, and $f_\psi$ is given by \eqref{fpsi}, then \begin{equation}\label{formula_a} f_{\mathcal H}(A)= f(\infty)-\frac{2}{\pi}\int_0^\infty f_\psi'(t) \,\arccot(A^\gamma/t^{\gamma})\,dt \end{equation} where the integral converges in the uniform operator topology, and \begin{equation}\label{NewEs1} \|f_{\mathcal H}(A)\ \le |f(\infty)| + \frac{M_\psi(A)}{2} \|f'\|_{H^1(\Sigma_\psi)} \le {M_\psi(A)} \|f\|_{\mathcal{H}_\psi}. \end{equation} Moreover, if $f(\infty)=0$, then \begin{equation}\label{NewEs2} \|f_{\mathcal H}(A)\| \le \frac{M_\psi(A)}{2} \|f\|_{\mathcal{H}_\psi}. \end{equation} \end{thm} \begin{proof} Let $\varphi \in (\theta,\psi)$ be fixed. Let $g(z)=\arccot(z)$ and $g_\gamma(z)=\arccot(z^\gamma)$. Since $g \in \mathcal{D}_0 \subset \mathcal{H}_{\gamma \varphi}$, $g_\gamma \in \mathcal{H}_\varphi$. Now \begin{align}\label{arccotes} \int_0^\infty |g'_\gamma(\rho e^{\pm i\varphi})|\,d\rho &= \gamma\int_0^\infty \rho^{\gamma-1}|g'(\rho^\gamma e^{\pm i\gamma\varphi})|\,d\rho \\ &=\int_0^\infty |g'(te^{\pm i\gamma\varphi})|\,dt =\int_0^\infty \frac{dt}{|1+t^2e^{\pm 2i\gamma\varphi}|} \notag \end{align} and \begin{align*} \int_0^\infty \frac{d\rho}{|1+\rho^2e^{2i\psi}|} &= 2 \int_1^\infty \frac{d\rho}{((\rho^2-1)^2 + 4\rho^2\cos^2\psi)^{1/2}} \\ &\le 2\sqrt{2}\int_1^\infty \frac{d\rho}{(\rho+\cos\psi)^2-(1+\cos^2\psi)}\\ &=\frac{\sqrt{2}}{(1+\cos^2\psi)^{1/2}} \log \left(\frac{1+\cos\psi+\sqrt{1+\cos^2\psi}}{1+\cos\psi-\sqrt{1+\cos^2\psi}}\right) \\ &\le \sqrt{2} \log\left(\frac{6}{\cos\psi}\right). \end{align*} Let $q_t(z) = \arccot(z^\gamma/t^\gamma), \, t>0$. By scale-invariance (Lemma \ref{simple}), $\|q_t\|_{\mathcal{H}_\varphi} = \|g_\gamma\|_{\mathcal{H}_\varphi}$. It follows from Lemma \ref{duren21} that $t \mapsto q_t$ is continuous from $(0,\infty)$ to $\mathcal{H}_\varphi$. Hence, in view of Proposition \ref{L7.4}, we have \begin{equation} \label{hbochner} f=f(\infty)-\frac{2}{\pi}\int_0^\infty f_\psi'(t)\, q_t \,dt, \end{equation} where the integral is understood as a Bochner integral in $\mathcal{H}_\varphi$. By applying the bounded operator $\Upsilon_A$ to both sides of \eqref{hbochner}, we obtain (\ref{formula_a}). The estimate \eqref{NewEs1} follows from \eqref{formula_a}, \eqref{NewEs2}, and Lemmas \ref{arc_bound} and \ref{Ss2}. \end{proof} \begin{rems}\label{rem_h} 1. If $M_\psi(A)=1$ in \eqref{NewEs1}, that is, if $-A$ generates a holomorphic $C_0$-semigroup which is contractive on $\Sigma_{(\pi/2)-\psi}$, then the $\mathcal H$-calculus is contractive. This seems to be a new feature which has not been present in constructions of other calculi in the literature. \noindent 2. An alternative to the estimate \eqref{NewEs1} is \begin{equation}\label{NewEs} \|f_{\mathcal H}(A)\|\le |f(\infty)|+ 3 M_{A^\gamma}\|f_\psi'\|_{L^1(\mathbb{R}_+)}. \end{equation} This is obtained from \eqref{formula_a}, using the estimates \eqref{estimateD} (with $s=0$) and \eqref{arccot_D} to obtain the estimate $\|\arccot(A^\gamma)\| \le 3\pi M_{A^\gamma}$. The constant $3$ is not optimal. It is possible to provide explicit bounds for $M_{A^\gamma}$ in terms of $M_A$. However we refrain from doing so in this paper, and we refer the interested reader to \cite[Propositions 5.1 and 5.2]{BGTad}. \end{rems} Finally in this section, we discuss the relations between \cite{Boy} and the present work. For $\psi \in (0,\pi)$, as in \cite{Boy}, let \begin{align*} k_\psi(t) =\frac{1}{\pi^2}\log\left|\coth\left(\frac{\pi t }{4 \psi}\right) \coth\left(\frac{\pi t}{4(\pi-\psi)}\right)\right|. \end{align*} Note that $k_\psi$ is an even function on $\mathbb R\setminus \{0\}$, and $\|k_\psi\|_{L^1(\mathbb R)}=1.$ For any $f \in L^\infty (\mathbb R_+),$ let \[ (f \circ k_\psi) (t):=\int_{0}^{\infty}f (s)k_\psi \left(\log (t/s) \right)\, \frac{ds}{s}, \qquad t > 0, \] and for $A \in \Sect(\theta)$, $\theta \in [0,\psi)$, define \begin{equation}\label{boya} W_\psi(A,t)=-\frac{A}{2}\left(e^{-i\psi} (A-e^{-i\psi}t)^{-2} + e^{i\psi}(A-e^{i\psi} t)^{-2}\right). \end{equation} It was proved in \cite[Theorem 3.1]{Boy} that if $A \in \Sect(\theta)$, $A$ has dense range, and \begin{equation}\label{absolute} \int_{0}^{\infty}|\langle W_{\psi}(A,t)x, x^*\rangle |\, dt <\infty, \qquad x \in X,\, x^* \in X^*, \end{equation} then $A$ admits a bounded $H^\infty(\Sigma_\psi)$-calculus given by \begin{equation}\label{formula} \langle f(A)x, x^*\rangle=\int_{0}^{\infty}\langle W_\psi(A,t)x, x^*\rangle (f_\psi \circ k_\psi)(t)\, dt, \quad x \in X,\, x^* \in X^*, \end{equation} where the integral converges absolutely (in the weak sense). Conversely, if $\psi \in (\theta,\pi)$ and $\varphi \in (\theta, \psi)$ are such that $A$ has a bounded $H^\infty(\Sigma_{\varphi})$-calculus, then \eqref{absolute} holds. (Note that in this situation $A$ has a bounded $H^\infty(\Sigma_{\psi})$-calculus given by \eqref{formula}, by the uniqueness of the calculus.) The formula \eqref {formula} is obtained in \cite{Boy} by rather involved Fourier analysis, and some technical details are omitted in \cite{Boy}. In \cite[Proposition 5.1]{Boy} it is observed that, if $f \in H^\infty(\Sigma_\psi)$ and is holomorphic in a larger sector, and $f'_\psi \in L^1(\mathbb R_+)$, $f'_\psi(\infty)=0$ (this assumption is not relevant), then one can integrate by parts and rewrite \eqref{formula} as \begin{equation}\label{boyad_f} f(A)=\int_0^\infty V_\psi(A,t) (f_\psi'\circ k_\psi)(t)\,\frac{dt}{t}, \end{equation} where \[ V_\psi(A,t)=-\frac{t}{2} \left(e^{-i\psi} (A-t e^{-i\psi})^{-1} +e^{i\psi}(A-t e^{i\psi})^{-1}\right), \] and the integral converges absolutely. This formally leads to the estimate \eqref{NewEs1}. While our reproducing $\arccot$-formula \eqref{formula_a} was inspired by \eqref{boyad_f}, it is not easy to put formal considerations in \cite{Boy} into the theory of functional calculi considered in this paper. One can relate \eqref{boyad_f} to \eqref{formula_a} and show that the formulas are essentially equivalent within the $\mathcal H$-calculus. This requires a number of technicalities and we intend to communicate them elsewhere. Here we note only that $f'_\psi \in L^1(\mathbb R_+)$ and $f \in H^\infty(\Sigma_\psi)$ (for $\psi=\pi/2$) do not imply that $f \in \mathcal{H}_\psi$ in general, as shown by an intricate example kindly communicated to us by A. Borichev. \section{Convergence Lemmas and Spectral Mapping Theorems}\label{SMT_sec} \subsection{Convergence Lemmas} Given a negative semigroup generator $A$, a Convergence Lemma for the holomorphic functional calculus is a useful result allowing one to deduce the convergence of $(f_k(A))_{k=1}^\infty$ to $f(A)$ from rather weak assumptions on convergence of $(f_k)_{k=1}^\infty$ to $f$; see \cite[Lemma 2.1]{Cowling}, \cite[Proposition 5.1.4]{HaaseB} and \cite[Theorem 3.1]{BHM}, for example. The following result is similar to a Convergence Lemma for the $\mathcal B$-calculus in \cite[Theorem 4.13 and Corollary 4.14]{BGT} (see also \cite[Section 8.1]{BGT2}). However, the different Convergence Lemmas deal with different classes of functions. To adjust the Convergence Lemma from \cite{BGT} to the current setting, we apply the change of variables method used in previous sections, and derive a variant of the Convergence Lemma for the $\mathcal D$-calculus. For any $\gamma \in (0,1)$ set formally \[ f_{k,\gamma}(z):=f_k(z^\gamma),\qquad g_\gamma(z):=g(z^\gamma). \] In the following result, $f(A)$ refers to the $\mathcal{D}$-calculus. \begin{thm}\label{convop_lemma} Let $A \in \Sect(\pi/2-)$. Let $s >-1$ and let $(f_k)_{k\ge 1} \subset \mathcal{D}_s$ be such that \begin{equation} \sup_{k \ge 1}\|f_k\|_{\mathcal{D}_s} <\infty, \end{equation} and there exists \[ f(z):=\lim_{k\to\infty}\,f_k(z),\qquad z\in \mathbb{C}_{+}. \] Let $g\in \mathcal{D}_s$ satisfy \[ g(0)=g(\infty)=0. \] Then \begin{equation}\label{conv_uni} \lim_{k\to\infty}\,\|(f(A)-f_k(A))g(A)\|=0. \end{equation} In particular, if $A$ has dense range, then \begin{equation}\label{conv_st} \lim_{k\to\infty}\,\|f(A)x-f_k(A)x\|=0, \end{equation} for all $x\in X$. \end{thm} \begin{proof} By assumption there exists $\theta \in (0,\pi/2)$ such that $A \in \Sect(\theta)$. By Corollary \ref{Fatou} we have $f \in \mathcal D_s$. Thus, without loss of generality, we can assume that $f\equiv 0$. Let $\gamma\in (1, \pi/(2\theta))$. Then by Corollary \ref{fpD}, \[ f_k(A)=f_{k,1/\gamma}(A^{\gamma}) \qquad \text{and} \qquad g(A)=g_{1/\gamma}(A^{\gamma}). \] Since $g_{1/\gamma} \in \mathcal{D}_0^\infty$, $f_{k,1/\gamma}\in \mathcal{D}_0^\infty$ and $f_{k,1/\gamma} g_{1/\gamma}\in \mathcal{D}_0^\infty$ (see Corollary \ref{gggC}), and the $\mathcal{D}$-calculus is an algebra homomorphism, we have \[ f_k(A)g(A)=f_{k,1/\gamma}(A^{\gamma})g_\gamma(A^{\gamma})=(f_{k,1/\gamma} g_{1/\gamma})(A^{\gamma}). \] Now \eqref{conv_uni} follows from Lemma \ref{48St2} and the continuity of the $\mathcal D$-calculus given by \eqref{epsd11}. Let $g(z)=z(1+z)^{-2}$ and note that $g \in \mathcal D_s$, and $g$ vanishes at zero and at infinity. If $A$ has dense range, then the range of $g(A)=A(1+A)^{-2}$ is dense as well (see \cite[Proposition 9.4]{KW}, for example). Since $\sup_{k \ge 1}\|f_k(A)\|<\infty$, \eqref{conv_uni} implies \eqref{conv_st}. \end{proof} In the following result, $f(A)$ refers to the $\mathcal{H}$-calculus. \begin{thm}\label{convop_h} Let $A \in \Sect(\theta),$ and let $\psi \in (\theta,\pi)$. Let $(f_k)_{k\ge 1} \subset \mathcal{H}_\psi$ be such that \begin{equation*} \sup_{k \ge 1}\|f_k\|_{\mathcal{H}_\psi} <\infty, \end{equation*} and there exists \[ f(z):=\lim_{k\to\infty}\,f_k(z),\qquad z\in \mathbb{C}_{+}. \] Let $g\in\mathcal{H}_\psi$ satisfy \[ g(0)=g(\infty)=0. \] Then \begin{equation*} \lim_{k\to\infty}\,\|(f(A)-f_k(A))g(A)\|=0. \end{equation*} In particular, if $A$ has dense range, then \begin{equation*} \lim_{k\to\infty}\,\|f(A)x-f_k(A)x\|=0, \end{equation*} for all $x\in X$. \end{thm} \begin{proof} The proof is very similar to Theorem \ref{convop_lemma}. Corollary \ref{Fatou} is replaced by Lemma \ref{FatouH}, the compatibility with fractional powers follows from the definitions and Proposition \ref{Hprops}(i), Corollary \ref{gggC} is replaced by Lemma \ref{simple}, Lemma \ref{48St2} is replaced by Lemma \ref{H11cl}, and \eqref{epsd11} is replaced by Corollary \ref{fpD}. \end{proof} \subsection{Spectral mapping theorems} Given a semigroup generator $-A$, a spectral mapping theorem for a functional calculus $\Xi_A$ signifies informally that $\Xi_A$ is associated to $A$ in a ``natural'' way. However, in general the spectral ``mapping'' theorem states only the inclusion $f(\sigma(A))\subset \sigma(\Xi_A(f))$. Equality may fail here even for functions such as $e^{-tz}$ and for rather simple operators $A$; see \cite[Section IV.3]{EN}, for example. While one may expect only the spectral inclusion as above, the equality $f(\sigma(A))\cup \{f(\infty)\} = \sigma(\Xi_A(f))\cup \{f(\infty)\}$ sometimes holds if $A$ inherits some properties of bounded operators such as strong resolvent estimates. Note that the spectral mapping theorem may not hold even for bounded operators if the functional calculus possesses only weak continuity properties, as discussed in \cite{Bercovici}. The statement below shows that the $\mathcal D$-calculus possesses the standard spectral mapping properties. It is similar to \cite[Theorem 4.17]{BGT}, with the addition of a statement about approximate eigenvalues. Recall that for $f \in \mathcal{D}_\infty$, its values $f(\infty)$ at infnity and $f(0)$ at $0$ are defined by \eqref{finf} and \eqref{fzero}. This convention is used below. \begin{thm} \label{SMT} Let $A \in \Sect(\pi/2-)$, $f \in\mathcal D_\infty$, and $\lambda \in \mathbb{C}$. \begin{enumerate} [\rm1.] \item If $x \in D(A)$ and $Ax = \lambda x$, then $f_{\mathcal D}(A)x = f(\lambda)x$. \item If $x^* \in D(A^*)$ and $A^*x^* = \lambda x^*$, then $f_{\mathcal D}(A)^*x^* = f(\lambda)x^*$. \item If $(x_n)_{n\ge1}$ are unit vectors in $D(A)$ and $\lim_{n\to\infty}\|Ax_n - \lambda x_n\| = 0$, then $\lim_{n\to\infty}\|f_{\mathcal{D}}(A)x_n - f(\lambda)x_n\| = 0$. \item One has $\sigma(f_{\mathcal D}(A)) \cup \{f(\infty)\} = f(\sigma(A)) \cup \{f(\infty)\}$. \end{enumerate} \end{thm} \begin{proof} The statements (1) and (2) are direct corollaries of \eqref{formulaD} and the reproducing formula for the $\mathcal D_s$-spaces given in Corollary \ref{Repr}. For (3), we use the $F$-product of the semigroup $(e^{-tA})_{t\ge0}$, as introduced in \cite{Dern}. Let $Y$ be the Banach space of all bounded sequences $\mathbf y := (y_n)_{n\ge1}$ in $X$ such that $\lim_{t\to0+} \|e^{-tA}y_n - y_n\| = 0$ uniformly in $n$, where $(e^{-tA})_{t\ge0}$ is the bounded holomorphic $C_0$-semigroup generated by $-A$. Let $Z$ be the closed subspace of $Y$ consisting of the sequences $\mathbf y$ such that $\lim_{n\to \infty} \|y_n\|=0$, and let $\widetilde Y = Y/Z$ and $Q: Y \to \widetilde Y$ be the quotient map. Then $(e^{-tA})_{t\ge0}$ induces a bounded holomorphic $C_0$-semigroup $(e^{-t\tilde A})_{t\ge0}$ on $\widetilde Y$, whose negative generator $\widetilde A$ is given by \[ D(\widetilde A) = \{Q(\mathbf y) : y_n \in D(A), (Ay_n) \in Y\}, \qquad \widetilde A(Q\mathbf y) = Q((Ay_n)). \] Then $\mathbf x := (x_n) \in Y$, $Q\mathbf x \in D(\widetilde A)$ and $\widetilde A Q(\mathbf x) = \lambda Q(\mathbf x)$. It follows from (1) that $f_{\mathcal{D}}(\widetilde A) Q\mathbf x = \lambda Q \mathbf x$. However it is very easy to see that $f_{\mathcal{D}}(\widetilde A) Q\mathbf x = Q((f_{\mathcal{D}}(A)x_n))$ (see \cite[Theorem 1.7(i)]{Dern}), and this establishes (3). To prove the spectral mapping theorem in (4), we follow the Banach algebra method used in \cite{BGT} for similar purposes and inspired by \cite[Section 16.5]{HP} and \cite[Section 2.2]{Dav80}. We may assume without loss of generality that $f(\infty)=0$. Let $\mathcal A$ be the bicommutant of $\{(z+A)^{-1} : -z \in \rho(A)\}$ in $L(X)$, so $\mathcal A$ is a commutative Banach algebra and the spectrum of $f_\mathcal{D}(A)$ in $\mathcal{A}$ coincides with the spectrum in $L(X)$. Observe that $\sigma (A) \subset \mathbb C_+\cup \{0\}$. Let $\chi$ be any character of $\mathcal{A}$. If $\chi((1+A)^{-1})=0$, then $\chi((z+A)^{-1}) = 0$ for all $z \in \mathbb{C}_+$, and hence $\chi(f(A)) = 0 = f(\infty)$. Otherwise, by the resolvent identity, $\chi((z+A)^{-1})=(z+\lambda)^{-1}$ for some $\lambda \in \sigma(A)$ and all $z \in \mathbb C_+$. Let $s>-1$ be such that $f \in \mathcal{D}_s$. Noting that the Stieltjes representation \eqref{Sti} converges in the uniform operator topology, we infer that $\chi((z+A)^{-(s+1)}) = (z+\lambda)^{-(s+1)}, z \in \mathbb C_+$. Applying $\chi$ to \eqref{formulaD} gives \[ \chi (f_{\mathcal{D}}(A))=- \frac{2^s}{\pi}\int_0^\infty \alpha^s\int_{-\infty}^\infty f'(\alpha+i\beta)(\lambda+\alpha-i\beta)^{-(s+1)}\,d\beta\,d\alpha, \] and then, by the reproducing formula \eqref{qnrep} for $\mathcal D_s$-functions (valid on $\mathbb C_+\cup\{0\}$), we obtain \[ \chi(f_{\mathcal{D}}(A)) = f(\lambda) \in f(\sigma(A)). \] Hence $\sigma(f_{\mathcal D}(A)) \cup \{0\} \subset f(\sigma(A)) \cup \{0\}$. To prove the opposite inclusion, note that if $\lambda \in \sigma (A)$ is fixed, then there is a character $\chi$ such that $\chi ((z+A)^{-1})=(z+\lambda)^{-1}$, so the above argument can be reversed yielding $\chi(f_{\mathcal{D}}(A))=f(\lambda)$, and thus finishing the proof. \end{proof} Next, using the same approach via Banach algebras, we prove the analogous spectral result for the $\mathcal H$-calculus. \begin{thm} Let $A \in \Sect(\theta)$ and $f \in \mathcal \mathcal{H}_\psi$ for some $\theta <\psi < \pi$, and let $\lambda \in \mathbb{C}$. \begin{enumerate} [\rm1.] \item If $x \in D(A)$ and $Ax = \lambda x$, then $f_{\mathcal H}(A)x = f(\lambda)x$. \item If $x^* \in D(A^*)$ and $A^*x^* = \lambda x^*$, then $f_{\mathcal H}(A)^*x^* = f(\lambda)x^*$. \item If $(x_n)_{n\ge1}$ are unit vectors in $D(A)$ and $\lim_{n\to\infty}\|Ax_n - \lambda x_n\| = 0$, then $\lim_{n\to\infty}\|f_{\mathcal{H}}(A)x_n - f(\lambda)x_n\| = 0$. \item One has $\sigma(f_{\mathcal H}(A)) \cup \{f(\infty)\} = f(\sigma(A)) \cup \{f(\infty)\}$. \end{enumerate} \end{thm} \begin{proof} The proofs of (1) and (2) are straightforward consequences of the reproducing formula \eqref{RepA1}. Moreover, (3) is deduced from (1) in the same way as in Theorem \ref{SMT}. The proof of (4) is similar to the corresponding proof in Theorem \ref{SMT}, based on the formula \eqref{sigma_def} which converges in the uniform operator topology. Let $\gamma=\pi/(2\psi)$. By the spectral mapping theorem for the holomorphic functional calculus \cite[Theorem 2.7.8]{HaaseB}, or \cite[Theorem 5.3.1]{MS}, one has \begin{equation}\label{alpha_spec} \sigma ((A^{\gamma}+t-i\beta)^{-1})=\{(\lambda^\gamma+t-i\beta)^{-1}:\lambda \in \sigma(A) \}\cup \{0\}, \quad t>0, \, \beta \in \mathbb R. \end{equation} As in the proof of Theorem \ref{SMT}, let $\mathcal A$ be the bicommutant of $\{(z+A)^{-1} : -z \in \rho(A)\}$ in $L(X)$. Then $f_{\mathcal H}(A) \in \mathcal{A}$ and the spectrum of $f_{\mathcal H}(A)$ in $\mathcal A$ coincides with the spectrum in $L(X)$. Let $\chi$ be any character of $\mathcal{A}$, and let $f \in \mathcal \mathcal{H}_\psi$ be such that $f(\infty)=0$. If $\chi((1+A^\gamma)^{-1})=0$, then, as above, $\chi((z+A^\gamma)^{-1}) = 0$ for all $z \in \mathbb{C}_+$, hence $\chi(f_\mathcal{H}(A)) = 0 = f(\infty)$. Otherwise $\chi ((z+A^{\gamma})^{-1})=(\lambda^{\gamma}+z)^{-1}$ for some $\lambda \in \sigma(A)$ and all $z \in \mathbb C_+$. Applying $\chi$ to \eqref{sigma_def} and using the representation \eqref{RepA1} for $\mathcal{H}_\psi$-functions, one gets \[ \chi (f_{\mathcal H}(A))= -\frac{1}{\pi}\int_0^\infty\int_{-\infty}^\infty f'_{1/\gamma}(t+i\beta) (\lambda^{\gamma}+t-i\beta)^{-1}\,d\beta\, dt=f(\lambda). \] Hence $\sigma (f_{\mathcal H}(A))\cup \{0\} \subset f(\sigma(A)) \cup\{0\}$. On the other hand, if $\lambda \in \sigma(A)$, then by \eqref{alpha_spec} there is a character $\chi$ such that $\chi ((z+A^{\gamma})^{-1}) =(z+\lambda^\gamma)^{-1}$. So using \eqref{RepA1} again, we infer that $f(\lambda) \in \sigma (f_{\mathcal H}(A))$. Combining the two paragraphs above yields (4). \end{proof} Our spectral mapping theorems differ from known spectral mapping theorems for the holomorphic functional calculus (see \cite{Haase_Sp} or \cite[Section 2.7]{HaaseB}) in at least three respects. We do not assume that $A$ is injective, we cover a wider class of functions including some with a mild singularity at zero (for example, $e^{-1/z}$), and our proofs are completely different. \section{Some applications to norm-estimates}\label{norm_estimates} In this section we apply directly the $\mathcal D$- and $\mathcal H$-calculi that we have constructed to obtain some operator norm-estimates. In particular, we obtain uniform bounds on the powers of Cayley transforms and on the semigroup generated by the inverse of a semigroup generator. We then compare the results to known estimates in the literature. We also revisit the theory of holomorphic $C_0$-semigroups and obtain several basic estimates along with some slight generalizations. \subsection{Norm-estimates via the $\mathcal D$-calculus} Let $A \in \Sect(\pi/2 -)$, and $V(A)$ be the Cayley transform $(A-I)(A+I)^{-1}$ of $A$. We now review several important estimates from the literature in the framework of the constructed $\mathcal{D}$- and $\mathcal{H}$-calculi. Recall that $-A$ is the generator of a bounded holomorphic semigroup $(e^{-tA})_{t\ge0}$. Let $e_t(z) = e^{-tz}, \, t \ge 0,\,z\in \mathbb{C}_+$. Then $e_t \in \mathcal{LM} \subset \mathcal{D}_\infty$ and \[ e^{-tA} = (e_t)_{\mathrm{HP}}(A) = (e_t)_{\mathcal{D}}(A). \] \begin{cor}\label{cayley_l} Let $A \in \Sect(\pi/2 -)$, so that \eqref{Aaa1} holds. \begin{enumerate}[{\rm(i)}] \item One has \[ \|V(A)^n\|\le 1+32(1+(\sqrt2\pi)^{-1})M_A^2, \qquad n \in \mathbb N. \] \item One has \[ \|e^{-tA}\|\le 2 M_A^2, \qquad t \ge 0. \] \item For every $\nu>0$ one has \[ \|A^\nu e^{-tA}\| \le 2^{\nu+2} t^{-\nu} \Gamma(\nu+1) M_A^{\lceil\nu\rceil+2}, \qquad t\ge0. \] \end{enumerate} If, in addition, the inverse $A^{-1}$ exists and is densely defined, then $A^{-1}$ generates a bounded holomorphic $C_0$-semigroup $(e^{-tA^{-1}})_{t \ge 0}$ satisfying \[ \|e^{-tA^{-1}}\|\le 1 + 2 M_A^2, \qquad t \ge 0, \] and, for every $\nu>0$, \[ \|A^{-\nu} e^{-tA^{-1}}\| \le 2^{\nu+2} t^{-\nu} \Gamma(\nu+1) M_A^{\lceil\nu\rceil+2}, \qquad t\ge0. \] \end{cor} \begin{proof} By Lemma \ref{cayley} and Theorem \ref{dcalculus}, for every $s >0$, \[ \|V(A)^n\|\le 1+ 2^{s+4} \pi^{-1} (B(s/2, 1/2)+2^{-s/2}) M_A^{\lceil{s}\rceil+1}, \qquad n \in \mathbb N. \] Setting $s=1$ we get the assertion (i). By Proposition \ref{BDs} or by Example \ref{dm} and Lemma \ref{leminv}(iii), the function $e^{-tz} \in \mathcal{D}_s$ for $s > 0$ and $t>0$, and by \eqref{frac} and \eqref{estimateD}, \[ \|e^{-tA}\|\le 2^s \pi^{-1}B(s/2, 1/2)M_A^{\lceil{s}\rceil+1}, \qquad t \ge 0. \] So the estimate (ii) follows by setting $s=1$ above. If $f_\nu(z):=z^\nu e^{-tz}$, $\nu>0$, then $f_\nu \in\mathcal{D}_s$ if and only if $s >\nu$, and in that case $\|f_\nu\|_{\mathcal{D}_s} = 2t^{-\nu} B((s-\nu)/2,1/2) \Gamma(\nu+1)$ (see Example \ref{dm}). Since $f_\nu$ has zero polynomial limits at zero and at infinity, $(f_{\nu})_{\mathcal D}(A)$ coincides with $A^{\nu}e^{-tA}$ as defined by the holomorphic functional calculus (see Remark \ref{coincid}). Using \eqref{frac}, it follows that, for every $s >\nu$, \[ \|A^\nu e^{-tA}\|\le \frac{2^{s+1}t^{-\nu}}{\pi} B\left(\frac{s-\nu}{2},\frac{1}{2}\right)\Gamma(\nu+1) M_A^{\lceil s\rceil+1}. \] Setting $s=\nu+1$, the first assertion in (iii) follows. The other two estimates are consequences of Lemma \ref{leminv}(i) and the estimates for $e^{-tA}$ and $A^\nu e^{-tA}$ obtained above. \end{proof} The results in Corollary \ref{cayley_l} are not new, and it serves as an ilustration of the utility of the $\mathcal{D}$-calculus. We have not aimed at finding the best possible estimates, but it seems that the $\mathcal{H}$-calculus provides bounds that are fairly precise whenever it is applicable. The power-boundedness of $V(A)$ was shown in \cite{Cr93} and \cite{Palencia}, using different methods. In \cite[Corollary 5.9]{BGT}, a weaker result was shown using the $\mathcal{B}$-calculus (so all operators satisfying \eqref{8.1}). Corollary \ref{cayley_l} shows how the $\mathcal{D}$- and $\mathcal{H}$-calculi can give a sharper estimate than the $\mathcal{B}$-calculus in the case of sectorial operators. Part (ii) above is one of many estimates for the bound on a bounded holomorphic semigroup in terms of its sectorial bound, and it is clearly not sharp. A careful estimation in \cite[Lemma 4.7]{BGT} of the bound obtained via the $\mathcal{B}$-calculus gave a bound of order $M_A \log M_A$ when $M_A$ is large. See also \cite[Theorem 5.2]{Sch1} where the result was established for the first time. Estimates of the form given in part (iii) have been known for a long time, but usually without showing the dependence on $M_A$. Next we consider estimates similar to Corollary \ref{cayley_l}(iii). In Lemma \ref{ders} and Theorem \ref{AnalF} $f(A)$ refers to the $\mathcal{D}$-calculus. \begin{lemma} \label{ders} Let $A\in \Sect(\pi/2-).$ If $f \in \mathcal{D}_s$, $s>-1$ and $n \in\mathbb{N}$, then \begin{equation}\label{formulaDD1} (z^nf^{(n)})(A)= C_{s,n} \int_0^\infty \alpha^s \int_{-\infty}^\infty f'(\alpha+i\beta)A^n (A+\alpha-i\beta)^{-(s+n+1)}\,d\beta\,d\alpha, \end{equation} where \[ C_{s,n} = {(-1)}^{n+1} 2^s \frac{\Gamma(s+n+1)}{\Gamma(s+1)}. \] \end{lemma} \begin{proof} By Corollary \ref{AlgDer} and the boundedness of the $\mathcal{D}_{s+n}$-calculus, \linebreak[4] $(z^nf^{(n)})(A)$ coincides with the derivative of order $n$ of the function $t \mapsto f(tA)$ evaluated at $t=1$. The formula \eqref{formulaD} for $f(tA) \in \mathcal D_s$ can be differentiated repeatedly with respect to $t$ by a standard method, and putting $t=1$ then gives the formula \eqref{formulaDD1}. \end{proof} \begin{thm}\label{AnalF} Let $A \in \Sect(\pi/2-)$. Let $f \in \mathcal{D}_\infty$, $n\in\mathbb{N}$, and assume that $f^{(k)} \in \mathcal{D}_\infty$ for $k=1,\dots,n$. Then \begin{equation} \label{anzn} (z^nf^{(n)})(A) = A^n f^{(n)}(A). \end{equation} Moreover, if $f \in \mathcal{D}_s, \, s>-1$, then \begin{equation}\label{tFG} \|t^n A^nf^{(n)}(tA)\|\le \frac{2^s\Gamma(s+n+1)}{\pi \Gamma(s+1)}(M_A+1)^n M_A^{\lceil{s}\rceil+1}\|f\|_{\mathcal{D}_s},\quad t>0. \end{equation} In particular, for $f(z)=e^{-z}\in \mathcal{D}_1$, \begin{equation}\label{tFGA} \|t^n A^n e^{-tA}\|\le 2(n+1)!(M_A+1)^n M_A^2,\qquad t>0,\quad n\in \mathbb{N}. \end{equation} \end{thm} \begin{proof} We will prove \eqref{anzn} by induction. First, assume that $f,f' \in \mathcal{D}_\infty$. Then \[ (1+A)^{-1} (zf')(A) + (1+A)^{-1}f'(A) = f'(A), \] This implies that $(zf')(A) = Af'(A)$. Now assume that $(z^kg^{(k)})(A) = A^kg^{(k)}(A)$ for $k=1,\dots,n$, for all functions $g$ such that $g^{(k)} \in \mathcal{D}_\infty$ for $k=0,\dots,n$. Let $f^{(k)} \in \mathcal{D}_\infty$ for $k=0,1,\dots,n+1$. Applying the base case ($k=1$) to the function $z^nf^{(n)}$ (noting that this function and its first derivative are in $\mathcal{D}_\infty$, by Corollary \ref{D00n}), and then the case $k=n$ to the function $f$, and the cases $k=n-1$ and $k=n$ to the function $f'$, we obtain \[ (z(z^nf^{(n)})')(A) = A (z^nf^{(n)})'(A), \] and hence \begin{align*} (z^{n+1}f^{(n+1)})(A) &= A \left(n(z^{n-1}f^{(n)}) + (z^nf^{(n+1)})\right)(A) - n (z^nf^{(n)})(A) \\ &= A^{n+1}f^{(n+1)}(A). \end{align*} This completes the proof of \eqref{anzn}. Since $M_A = M_{tA}$ for all $t>0$, it suffices to prove (\ref{tFG}) for $t=1$. From Lemmas \ref{fractional} and \ref{ders}, we obtain \begin{align*} \lefteqn{\null\hskip-15pt \|A^n\,f^{(n)}(A)\|}\\ &\le \frac{2^s\Gamma(s+n+1)}{\pi \Gamma(s+1)}(M_A+1)^n M_A^{\lceil{s}\rceil+1} \int_0^\infty\alpha^s\int_{-\infty}^\infty \frac{|f'(\alpha+i\beta)|}{|\alpha-i\beta|^{s+1}}\,d\beta\,d\alpha\\ &=\frac{2^s \Gamma(s+n+1)}{\pi\Gamma(s+1)}(M_A+1)^n M_A^{\lceil{s}\rceil+1}\|f\|_{\mathcal{D}_s}. \qedhere \end{align*} \end{proof} \begin{rem} In Theorem \ref{AnalF}, the assumption that $f \in \mathcal{D}_\infty$ and $f^{(k)} \in \mathcal{D}_\infty, k=1,2,\dots,n$, can be replaced by the assumption that $f \in \mathcal{D}_\infty$ and $f^{(n)} \in \mathcal{D}_\infty$, by using a result of Lyubich \cite{Lyu}. See Corollary \ref{lyubich}. \end{rem} \subsection{Norm-estimates via the $\mathcal{H}$-calculus} \label{Bernstein} Now we use the $\mathcal{H}$-calculus to provide a new proof that holomorphy of operator semigroups generated by $-A$ is preserved for subordinate semigroups generated by $-g(A)$ where $g$ is a Bernstein function. This was proved for the first time in \cite{GT15}. If $-A$ is the generator of a bounded $C_0$-semigroup $(e^{-tA})_{t \ge 0}$ on a Banach space $X$, and $g$ is a Bernstein function given by \eqref{bsfn}, then the operator \begin{equation}\label{DBer} g_0(A)x:=ax+bAx+\int_{(0, \infty)}\left(x-e^{-tA}x\right)\, d\mu(t), \quad x \in D(A), \end{equation} is closable, and $g(A)$ can be defined as the closure of $g_0(A)$. Thus $D(A)$ is a core for $g(A)$, and one can prove that $-g(A)$ generates a contraction $C_0$-semigroup on $X$. Several equivalent definitions of $g(A)$ are possible, and we refer the reader to \cite{Schill}, \cite{GHT} and \cite{GT15}. If $A$ is injective, then $g(A)$ is well-defined within the (extended) holomorphic functional calculus and is given by \eqref{DBer} as above; see Proposition \ref{Hprops} and \cite[Propositions 3.3 and 3.6]{GT15}. The next statement shows that the so-called semigroup subordination preserves the holomorphy of $C_0$-semigroups along with the holomorphy angles. It was one of the main results of \cite{GT15}, settling a question raised by Kishimoto and Robinson \cite{KR}. See also \cite{BGTad} and \cite{BGT} for generalizations and other proofs. \begin{thm}\label{BF1} Let $A \in \Sect(\theta)$, where $\theta\in [0,\pi/2)$, and let $g$ be a Bernstein function as in \eqref{bsfn}. Then $g(A) \in \Sect(\theta)$. More precisely, for all $\psi\in (\theta,\pi/2)$, $\varphi\in (\psi,\pi)$, and $\lambda\in{\Sigma}_{\pi-\varphi}$, \begin{equation}\label{bound_berns} \|\lambda (\lambda + g(A))^{-1}\| \le 2M_\psi (A)\left(\frac{1}{\sin(\min(\varphi,\pi/2))}+\frac{2}{\cos\psi \sin^2((\varphi-\psi)/2)}\right). \end{equation} \end{thm} \begin{proof} Let $\psi\in (\theta,\pi/2)$, $\varphi\in (\psi,\pi)$, and $\lambda\in{\Sigma}_{\pi-\varphi}$. If \[ f(z)=f(z;\lambda):=(\lambda+g(z))^{-1},\qquad z\in \Sigma_\psi, \] then by Corollary \ref{BF} we have $f \in \mathcal \mathcal{H}_\psi$. We shall show that $f_\mathcal{H}(A) = (\lambda+g(A))^{-1}$. It then follows from \eqref{NewEs1} and \eqref{bernst} that \eqref{bound_berns} holds. Since the choice of $\psi\in (\theta,\pi)$ and $\varphi\in (\psi,\pi)$ is arbitrary, this shows that the operator $g(A)$ is sectorial of angle $\theta$. If $A$ is injective, then $f(A)$ and $\lambda + g(A)$ are consistently defined in the holomorphic functional calculus, and therefore $f_{\mathcal H}(A) = f(A) = (\lambda+g(A))^{-1}$ (see \cite[Theorem 1.3.2f)]{HaaseB}). When $A$ is not injective, we follow the approach proposed in the proof of \cite[Theorem 4.8]{BGTad}. Since $A+\epsilon$ is invertible, we have \[ (\lambda+g(A+\epsilon))^{-1}=f_\mathcal{H}(A+\epsilon). \] By Proposition \ref{Hprops}(v), \begin{equation} \label{epsD} \lim_{\epsilon\to0+} \|f_\mathcal{H}(A+\epsilon) - f_\mathcal{H}(A)\|=0. \end{equation} Let $x \in D(A)$. Since $f_\mathcal{H}(A+\epsilon)$ commutes with $(1+A)^{-1}$, we have $f_\mathcal{H}(A+\epsilon)x\in D(A)$, and by \eqref{DBer}, \begin{gather} \label{first} x-(\lambda+g(A))f_\mathcal{H}(A+\epsilon)x =[g(A+\epsilon)-g(A)]f_\mathcal{H}(A+\epsilon)x\\ \null{\hskip30pt} =\epsilon bf_\mathcal{H}(A_\epsilon)x -\int_{(0,\infty)} (1-e^{-\epsilon t})e^{-tA} f_\mathcal{H}(A+\epsilon)x\,d\mu(t).\notag \end{gather} It follows from \eqref{epsD} that \[ C_\lambda := \sup_{\epsilon\in(0,1]} \|f_\mathcal{H}(A+\epsilon)\| < \infty, \] and hence \begin{align}\label{second} \|x - (\lambda+g(A))f_\mathcal{H}(A+\epsilon)x\| &\le \epsilon b C_\lambda\|x\| +C_\lambda K_A \int_{(0,\infty)} (1-e^{-\epsilon t})\,d\mu(t)\,\|x\| \\ &\to 0,\qquad \epsilon\to 0, \notag \end{align} where $K_A := \sup_{t>0} \|e^{-tA}\|$. Since $\lambda+g(A)$ is closed, it now follows that \[ (\lambda+g(A))f_\mathcal{H}(A)x=x,\quad x\in D(A). \] Since $D(A)$ is dense in $X$ and it is a core for $g(A)$, the boundedness of $f_\mathcal{H}(A)$ implies that \[ (\lambda+g(A))f_\mathcal{H}(A)x=x,\qquad x\in X. \] Then \[ f_\mathcal{H}(A)(\lambda+g(A))x=x,\qquad x\in D(A), \] and it follows that this holds for all $x \in X$. Thus, $f_\mathcal{H}(A)=(\lambda+g(A))^{-1}$, as required. \end{proof} \begin{rems} 1. A new feature of Theorem \ref{BF1} is an explicit sectoriality constant for $g(A)$, given by the right hand-side of \eqref{bound_berns}. This could be valuable when applying the result to families of sectorial operators. Thus \eqref{bound_berns} offers an improvement over similar estimates in \cite{BGTad}, \cite{BGT} and \cite{GT15}, where the sectoriality constants for $g(A)$ are rather implicit. \noindent 2. We take this opportunity to correct a parsing misprint in the proof \cite[Theorem 4.9]{BGTad}. One should replace $f(A)$ with $f(A)+z$ in the third and fourth displays on \cite[p.932]{BGTad} (see \eqref{first} and \eqref{second} above for similar formulas). \end{rems} Finally, as an illustration, we show how the holomorphy of $C_0$-semigroups generated by operators $-A^\gamma$ fits within the $\mathcal{H}$-calulus, and how estimates of similar type to Corollary \ref{cayley_l}(ii) can be obtained from the representation of the $\mathcal{H}$-calculus and the function $\arccot$, as in Theorem \ref{Sovp}. The following result is similar to \cite[Corollary 5.2]{Boy}, and a generalization of the main result in \cite{DeLa87} to non-integer $\gamma$. See also \cite[Remark 2, p.83]{Cr93}. \begin{cor} \label{better} Let $A \in \Sect(\theta), \theta\in (0,\pi)$, and let $\gamma = \psi/\theta \in (0,\pi/(2\theta))$. Then $(e^{-tA^\gamma})_{t\ge0}$ is a bounded holomorphic $C_0$-semigroup of angle $(\pi/2) - \gamma\theta$. More precisely, if $\psi \in (\theta,\pi)$ and $\lambda = |\lambda| e^{i\varphi} \in \Sigma_{(\pi/2)-\gamma\psi}$, then \begin{equation} \label{est_gamma} \|e^{-\lambda A^{\gamma}}\| \le \frac{1}{2}\left(\frac{1}{\cos(\gamma\psi+\varphi)}+\frac{1}{\cos(\gamma\psi-\varphi)}\right) M_{\gamma\psi}(A). \end{equation} \end{cor} \begin{proof} Let $\psi \in(\theta,\pi)$. Since the $\mathcal{H}$-calculus $\Upsilon_A$ is a homomorphism and $\Upsilon_A(e^{-\lambda z^\gamma}) = e^{- \lambda A^\gamma}$ for every $\lambda \in \Sigma_{(\pi/2)-\gamma\psi}$, $(e^{-\lambda A^\gamma})_{\lambda \in \Sigma_{(\pi/2)-\gamma\psi}}$ is an operator semigroup. By Theorem \ref{hardy2} and the argument in Section \ref{prelims}, the map $\lambda \mapsto e^{-\lambda z^\gamma}$ is holomorphic from $\Sigma_{(\pi/2)-\gamma\psi}$ to $\mathcal{H}_\psi$, so $\lambda \mapsto e^{-\lambda A^\gamma}$ is also holomorphic. The estimate \eqref{est_gamma} follows from Example \ref{hexs}(2) and \eqref{NewEs2}, and it shows boundedness of the semigroup on each relevant sector. \end{proof} Note that, for $\gamma=1$, Corollary \ref{better} provides a sharper bound than Corollary \ref{cayley_l}(ii). \section{Appendix: Shifts on $\mathcal{D}_s$ and $\mathcal{H}_\psi$} \label{shifts} The shift semigroups on the Besov space $\mathcal{B}$ had an important role in the study of the $\mathcal{B}$-calculus in \cite{BGT} and \cite{BGT2}. While the semigroups are not essential in this paper, we think they will be important for further research, and so we describe their properties on the spaces $\mathcal{D}_s$ and $\mathcal{H}_\psi$. In this appendix we prove that the shifts $(T(\tau))_{\tau\in\mathbb{C}_+}$ given by \[ (T(\tau)f)(z):=f(z+\tau),\quad z\in \mathbb{C}_{+},\quad \tau\in \mathbb{C}_{+}, \] form holomorphic $C_0$-semigroups on $\mathcal{D}_s$ for each $s>-1$, and on $\mathcal{H}_\psi$ for each $\psi \in(0,\pi)$. We consider first the space $\mathcal{D}_s$, and we begin by proving that the semigroup $(T(\tau))_{\tau \in \Sigma_\psi}$ of operators is uniformly bounded on $\mathcal{D}_s$, for each $s>-1$ and $\psi \in (0,\pi/2)$. \begin{thm}\label{21T} Let $s>-1$, $\psi \in (0,\pi/2)$ and $a=\tan\psi$. For all $\tau\in \Sigma_\psi$, we have \begin{equation}\label{Taa} \|T(\tau)f\|_{\mathcal{D}_s}\le C_{a,s}\|f\|_{\mathcal{D}_s},\quad f\in \mathcal{D}_s, \end{equation} where \begin{align} C_{a,s} &:= \frac{(s+1)2^{s}B((s+1)/2,1/2)}{\pi \cos\psi \cos^{s+2}\psi_a} +2^{s+1}, \label{const_a}\\ \psi_a &:=\arctan \big(a+\sqrt{1+a^2}\big). \end{align} \end{thm} \begin{proof} Let $\tau =t+iv \in \Sigma_\psi$ and $f \in \mathcal{D}_s$. We have \begin{align*} \|T(\tau)f\|_{\mathcal{D}_s}&\le |f(\infty)|+ \int_{\mathbb{C}_+} \frac{(\Re z)^s}{|z|^{s+1}} |f'(z+\tau)| \, dS(z) = |f(\infty)|+J(\tau), \end{align*} where $dS$ denotes area measure on $\mathbb{C}_+$ and \[ J(\tau):=\int_{\Re z \ge t} \frac{(\Re z-t)^s} {|z-\tau|^{s+1}} |f'(z)|\,dS(z). \] Let \begin{align*} W(\tau)&:= \left\{z\in \mathbb{C}: \Re z\ge t,\;|z-\tau| \le |\tau| \right\},\\ W_0(\tau):&= W(\tau)-\tau= \left\{z\in \mathbb{C}_{+}:\;|z| \le |\tau| \right\}. \end{align*} If $\Re z \ge t$ and $z \notin W(\tau)$, then $|z| \le |z-\tau|+|\tau| \le 2|z-\tau|$. Hence \begin{align} \label{JSo} J(\tau) &\le \int_{W(\tau)} \frac{(\Re z - t)^s}{|z-\tau|^{s+1}} |f'(z)|\,dS(z)\\ & \null \hskip30pt + 2^{s+1}\int_{\Re z \ge t} \frac{(\Re z)^s}{|z|^{s+1}} |f'(z)| \,dS(z) \notag \\ &\le \max_{z\in W(\tau)}\,|f'(z)| \int_{W_0(\tau)} \frac{(\Re z)^s}{|z|^s}\,dS(z) +2^{s+1}\|f'\|_{\mathcal{V}_s}.\notag \end{align} Moreover, \begin{equation}\label{sts} \int_{W_0(\tau)} \frac{(\Re z)^s}{|z|^{s+1}}\,dS(z) = \int_{-\pi/2}^{\pi/2}\int_0^{|\tau|} \cos^{s}\varphi \,d\rho\,d\varphi = |\tau| B((s+1)/2,1/2). \end{equation} For $z\in W(\tau)$, we also have $\Re z \ge t$ and \[ |\Im z| \le |v| + \sqrt{t^2+v^2} \le \left(a + \sqrt{1+a^2}\right)t. \] Hence $z\in \Sigma_{\psi_a}$ and $|z|\ge t$, so by Corollary \ref{Cangle}, \begin{equation}\label{der} \max_{z\in W(\tau)}\,|f'(z)| \le \max_{z\in \Sigma_{\psi_a},\,|z|\ge t}\,|f'(z)| \le \frac{(s+1)2^s}{\pi t\cos^{s+2}\psi_a}\|f'\|_{\mathcal{V}_s}. \end{equation} Inserting the estimates \eqref{sts} and \eqref{der} into (\ref{JSo}) and using $|\tau|\le t/\cos\psi$, we obtain \[ J(\tau)\le \left( \frac{(s+1)2^s B((s+1)/2,2)}{\pi\cos\psi \cos^{s+2}\psi_a} +2^{s+1}\right) \|f'\|_{\mathcal{V}_s}, \] and (\ref{Taa}) follows. \end{proof} \begin{cor}\label{SemDa} For any $s>-1$, the family $T:=(T(\tau))_{\tau\in \mathbb{C}_{+}}$ is a bounded holomorphic $C_0$-semigroup on $\mathcal{D}_s$ of angle $\pi/2$. The generator of the semigroup is $-A_{\mathcal{D}_s}$, where \[ D(A_{\mathcal{D}_s}) = \{f \in \mathcal{D}_s : f' \in \mathcal{D}_s\}, \qquad Af = - f'. \] \end{cor} \begin{proof} By Theorem \ref{21T}, $T$ is bounded on $\Sigma_\psi$ for each $\psi \in (0,\pi/2)$, and as noted in Remark \ref{remdr} the function $\lambda \mapsto r_\lambda$ is a holomorphic function from $\mathbb{C}_+$ to $\mathcal{D}_s$, so $\tau \mapsto T(\tau)r_\lambda$ is holomorphic. Since $\widetilde{\mathcal{R}}(\mathbb{C}_+)$ is dense in $\mathcal{D}_s$ (Theorem \ref{D00}), it follows that $T$ is strongly continuous on $\Sigma_\psi \cup \{0\}$, and moreover, for $f \in \mathcal{D}_s$, the map $\tau \mapsto T(\tau)f$ is holomorphic on $\mathbb{C}_+$. The proof of the statement about the generator is almost identical to the proof for the space $\mathcal{B}$ in \cite[Lemma 2.6]{BGT}. \end{proof} \begin{cor} \label{lyubich} Let $f \in \mathcal{D}_s$, $n \in \mathbb{N}$, and assume that $f^{(n)} \in \mathcal{D}_s$. Then $f^{(k)} \in \mathcal{D}_s$ for $k=1,2,\dots,n-1$. \end{cor} \begin{proof} Consider the operator $Ag = -g'$ on $\operatorname{Hol}(\mathbb{C}_+)$, and its part $A_{\mathcal{D}_s}$ in the subspace $\mathcal{D}_s$. The operators $A_{\mathcal{D}_s}+m, \, m=1,2,\dots,n$, are surjective on $\mathcal{D}_s$, and $g=0$ is the only solution in $\mathcal{D}_s$ to $\prod_{m=1}^n(A+m)g=0$. The statement follows from \cite[Theorem 1]{Lyu}. \end{proof} \begin{rem} The space $\mathcal{B}$ is invariant under vertical shifts: $f(z) \mapsto f(z+i\sigma)$ for $\sigma \in \mathbb{R}$. However the spaces $\mathcal{D}_s$ and $\mathcal{D}_s^\infty$ are not invariant under vertical shifts. See Example \ref{eac}. \end{rem} Now we will show that the shifts, $(T(\tau)f)(z) = f(z+\tau)$, also form a bounded holomorphic $C_0$-semigroup on $\mathcal{H}_\psi$ for $\psi\in(0,\pi)$. If $\psi>\pi/2$, the operators are defined for $\tau \in \Sigma_{\pi-\psi}$. For this aim, we will recall the Gabriel inequality for holomorphic functions. Let $\Omega$ be a bounded convex domain in $\mathbb{C}$ and let $\Gamma \subset \Omega$ be a Jordan curve wth convex interior. Then there exists a universal constant $K>0$ (not depending on $f, \Omega$ and $\Gamma$) such that, for all $f \in \operatorname{Hol}(\Omega) \cap C(\overline\Omega)$, \begin{equation}\label{gabriel} \int_{\Gamma} |f(z)|\, |dz| \le K \int_{\partial \Omega} |f(z)|\, |dz|. \end{equation} The inequality was conjectured by J. Littlewood and first proved by Gabriel in \cite[Theorem I]{Gabriel_35}. It is thoroughly discussed in \cite[Selected Seminars, 2, 4 and 5]{Beurling} and \cite[Section 5]{Granados} providing simpler proofs, more general versions and additional insights. It is known that $2 < K < 3.7$ (see \cite[p.457]{Beurling}, for example). \begin{thm}\label{Hsg} Let $\psi \in (0,\pi)$ and $\psi_0=\min\{\psi,\pi-\psi\}$. The family $T=(T(\tau))_{\tau \in \Sigma_{\psi_0}}$ is a bounded holomorphic $C_0$-semigroup on each of the spaces $H^1(\Sigma_\psi)$ and $\mathcal H_\psi$. \end{thm} \begin{proof} We will consider first the space $H^1(\Sigma_\psi)$ and show that the family of shifts on that space are uniformly. Then the results for both spaces, $H^1(\Sigma_\psi)$ and $\mathcal H_\psi$, follow quickly. Assume first that $\psi \in (0,\pi/2]$, and let $\psi' \in (0,\psi)$. Let $f \in H^1(\Sigma_\psi)$ and $\alpha:=\sin(\psi -\psi')$. By the mean value inequality, for any $r>0$ and $\varphi \in (-\psi',\psi')$ we have \begin{align}\label{limit1} r |f(re^{i\varphi})| &\le \frac{1}{\pi \alpha^2 r}\int_{|z-re^{i\varphi}|\le \alpha r} |f(z)| \, dS(z)\\ &\le \frac{2(1+\alpha)}{\pi \alpha} \int_{-\psi'}^{\psi'} \int_{(1-\alpha)r}^{(1+\alpha)r} |f(\rho e^{i\phi})|\, d\rho\,d\phi \notag\\ &\to 0, \end{align} as $r \to 0$ or $r \to \infty$, by the dominated convergence theorem. Now let $\tau \in \Sigma_\psi$ and take $\varphi$ such that $|\arg\tau| < \varphi < \psi$. Let $\Gamma_{\tau, \varphi}=\{\tau+te^{i\varphi}: t \ge 0\} \cup \{\tau+te^{-i\varphi} : t \ge 0 \}$. Let $\psi'\in(\varphi,\psi)$, and take $r_1$ and $r_2$ such that $0< r_1 <|\tau| < r_2$. We now apply Gabriel's inequality \eqref{gabriel} with \[ \Omega_{r_1,r_2} := \{z \in \mathbb{C}: \Re z > r_1 \cos\psi', |z| < r_2, |\arg z| < \psi'\} \] and \[ \Gamma_{\tau,\varphi,r_2} := \{z \in \Gamma_{\tau,\varphi} : |z| \le r_2\} \cup \{z \in \mathbb{C}: |z|=r_2, \, |\arg (z-\tau)| \le \varphi\}. \] We obtain \begin{align}\label{gabr} &\\ \lefteqn{K^{-1} \int_{\Gamma_{\tau,\varphi,r_2}} |f(z)|\, |dz| \le \int_{\partial\Omega_{r_1,r_2}} |f(z)|\, |dz|} \notag\\ &\le \int_{r_1}^{r_2}|f(re^{i\psi'})|\, dr +\int_{r_1}^{r_2}|f(re^{-i\psi'})|\, dr + 2 \pi r_2 \sup_{\varphi \in (-\psi', \psi')} |f(r_2 e^{i\varphi})| \notag \\ &\null\hskip20pt +2 r_1\sin\psi' \sup_{\varphi \in (-\psi', \psi')} \left|f\left(\frac{r_1\cos \psi'}{\cos \varphi} e^{i\varphi}\right)\right|,\notag \end{align} where $K>0$ is given by \eqref{gabriel}. By \eqref{limit1}, the last two terms converge to zero as $r_1 \to 0$ and $r_2 \to \infty$. From Theorem \ref{hardy0}(iii), we obtain that \[ K^{-1} \int_{\Gamma_{\tau,\varphi}} |f(z)|\, |dz| \le \|f\|_{H^1(\Sigma_{\psi'})}\le \|f\|_{H^1(\Sigma_\psi)}. \] Letting $\psi' \to \psi$ and using Fatou's Lemma, we infer that \begin{equation}\label{acute} K^{-1} \int_{\Gamma_{\tau,\varphi}} |f(z)|\, |dz| \le \|f\|_{H^1(\Sigma_{\psi})}. \end{equation} Now we consider the case when $\psi >\pi/2$, $\varphi<\pi-\psi$. Let $\Gamma^{\pm}_{\tau,\varphi}:= \{\tau+te^{\pm i\varphi}: t \ge 0\}$. If $\varphi=0$, then, by the previous case applied for $H^1(\Sigma_{\pi/2})$, and Theorem \ref{hardy0}(iii), \begin{equation}\label{acute0} \int_{\Gamma_{\tau,\varphi}} |f(z)|\, |dz| \le \|f\|_{H^1(\Sigma_{\pi/2})}\le \|f\|_{H^1(\Sigma_{\psi})}. \end{equation} For $\varphi \ne 0$, we consider two convex subsectors of $\Sigma_\psi$. Take $\varphi \in (0,\psi)$ and $\psi' \in (\varphi,\psi)$. Let $\tau \in \Sigma_{\pi-\psi}$, and assume without loss that $\tau \in \Sigma_\psi^+$. Let \[ \Omega^+ := \{z \in \mathbb{C}: \Re z > r_1 \cos\psi', |z| < r_2, 0 < \arg z < \psi'\}, \] and $r_1$ be small enough so that the lines $\Gamma^{\pm}_{\tau, \varphi}$ are contained in $\Sigma^{\pm}_\psi$, respectively. Observe that by the definition of the $H^1(\Sigma_\psi)$-norm, \[ 2 \int_{\Gamma_{0,0}} |f(z)|\, |dz| \le \|f\|_{H^1(\Sigma_\psi)}. \] Therefore, \begin{align}\label{nacute} K^{-1}& \int_{\Gamma_{\tau,\varphi}} |f(z)|\, |dz|= K^{-1} \int_{\Gamma^{+}_{\tau,\varphi}} |f(z)|\, |dz|+ K^{-1} \int_{\Gamma^{-}_{\tau,\varphi}} |f(z)|\, |dz|\\ \le& \|f\|_{H^1(\Sigma_\psi)} +2 \int_{\Gamma_{0,0}} |f(z)|\, |dz|\le 2 \|f\|_{H^1(\Sigma_\psi)}.\notag \end{align} Combining \eqref{acute}, \eqref{acute0} and \eqref{nacute}, it follows that for all $\theta \in (0, \psi_0)$, $\tau \in \Sigma_\theta$ and $\varphi \in [0,\psi),$ \begin{equation}\label{hardy1} \|T(\tau)f\|_{H^1(\Sigma_\psi)}\le 2K \|f\|_{H^{1}(\Sigma_\psi)}. \end{equation} If $f \in \mathcal H_\psi$, then the same argument yields \[ \|T(\tau)f'\|_{H^1(\Sigma_\psi)}\le 2K \|f'\|_{H^{1}(\Sigma_\psi)}, \] so that for all $\theta \in (0, \psi_0)$, $\tau \in \Sigma_\theta$ and $\varphi \in [0,\psi)$, \begin{equation}\label{hardy3} \|T(\tau)f\|_{\mathcal H_\psi}\le \max(1,2K) \|f\|_{\mathcal H_\psi}. \end{equation} A direct verification shows that $\|T(\tau)r_\lambda-r_\lambda\|_{\mathcal{H}_\psi} \to 0, \tau \to 0$, for every $\lambda \in \mathbb{C} \setminus \overline{\Sigma}_\psi$. It now follows from \eqref{hardy3} that $(T(\tau))_{\tau\in \Sigma_{\psi_0}}$ is a bounded $C_0$-semigroup on $\mathcal{H}_{\psi,0}$ and then on $\mathcal H_{\psi}$. The holomorphy of $(T(\tau))_{\tau\in \Sigma_{\psi_0}}$ on $\mathcal H_\psi$ follows from Theorem \ref{hardy2}(iii) and the method discussed in Section \ref{prelims}. The fact that $(T(\tau))_{\tau \in \Sigma_{\psi_0}}$ is a (sectorially) bounded holomorphic $C_0$-semigroup on $H^1(\Sigma_\psi)$ can be proved in the same way. The continity of the evaluation functionals follows from \eqref{limit1}. \end{proof} \begin{rem} Following a more conventional approach, one may try to prove Theorem \ref{Hsg} by reducing the estimates to the half-plane case and applying Carleson's embedding theorem for $H^1(\mathbb C_+)$. However, the technical details become rather cumbersome, so we prefer to use Gabriel's inequality allowing for a more transparent argument. \end{rem} \section{Appendix: The $\mathcal{D}$-calculus vs the HP- and the $\mathcal B$-calculi}\label{VS} It is natural to compare the strength of the $\mathcal D$-calculus with some other functional calculi, such as the recently constructed $\mathcal B$-calculus and the well-known HP-calculus. To show the advantages of our $\mathcal D$-calculus with respect to the $\mathcal B$-calculus and the HP-calculus, as an illustrative example, we consider the family of functions $\{f_n: n \ge 1\}$ given by \[ f_n(z)=\left(\frac{z-1}{z+1}\right)^n,\qquad z\in \mathbb{C}_{+},\quad n\in \mathbb{N}. \] This family is contained in $\mathcal{LM}$ (see \cite[Section 6]{BGT2}), and it arises naturally in the study of asymptotics for powers of Cayley transforms of semigroup generators. It is shown in \cite[Lemma 3.7]{BGT} and \cite[Lemmas 5.1 and 5.2]{BGT2} that \[ \|f_n\|_\mathcal{B} \asymp \log n \qquad \text{and} \qquad \|f_n\|_{{\rm HP}} \asymp n^{1/2}, \qquad n\to\infty. \] We will show that the $\mathcal D$-calculus provides sharper estimates for the corresponding operator functions. To this aim we need the next lemma. \begin{lemma}\label{cayley} For $s>0$, \[ \|f_n\|_{\mathcal{D}_s}\le 1+16\left(B(s/2,1/2)+2^{-s/2}\right),\qquad n\in \mathbb{N}. \] \end{lemma} \begin{proof} Let $s>0$ and $n \in \mathbb N$ be fixed. We have \[ f_n(\infty)=1,\quad f'_n(z)=2n\frac{(z-1)^{n-1}}{(z+1)^{n+1}}, \] and then \[ \|f_n\|_{\mathcal{D}_s}=1+8n \int_0^{\pi/2} \cos^s\psi\int_1^\infty \frac{g_n(\rho,\psi)\,d\rho}{\rho^2+2\rho\cos\psi+1}\,d\psi =1+8n J_n, \] where \[ g_n(\rho,\psi)=\left(\frac{\rho^2-2\rho\cos\psi+1}{\rho^2+2\rho\cos\psi+1}\right)^{(n-1)/2}. \] Let $J_n= J_{1, n} + J_{2,n}+J_{3,n}$, where \begin{align*} J_{1,n}&=\int_0^{\pi/2} \cos^s\psi\int_2^\infty \frac{g_n(\rho,\psi)\,d\rho}{\rho^2+2\rho\cos\psi+1}\,d\psi,\\ J_{2,n}&=\int_{\pi/4}^{\pi/2} \cos^s\psi\int_1^2\frac{g_n(\rho,\psi)\,d\rho}{\rho^2+2\rho\cos\psi+1}\,d\psi,\\ J_{3,n}&=\int_0^{\pi/4}\cos^s\psi\int_1^2\frac{g_n(\rho,\psi)\,d\rho}{\rho^2+2\rho\cos\psi+1}\,d\psi. \end{align*} We estimate each of the summands $J_{1, n}$, $J_{2,n}$, and $J_{3,n}$ separately. First, \begin{align*} \frac{\partial}{\partial \rho}\,g_{n+2}(\rho,\psi)&=2(n+1)\frac{(\rho^2-1)\cos\psi}{(\rho^2+2\rho\cos\psi+1)^2}g_{n}(\rho,\psi)\\ &\ge \frac{2(n+1)}{3} \frac{\cos\psi\,g_n(\rho,\psi)}{(\rho^2+2\rho\cos\psi+1)}>0,\quad \rho>2. \end{align*} Hence \begin{align*} J_{1,n}&\le \frac{3}{2(n+1)}\int_0^{\pi/2}\cos^{s-1}\psi \int_2^\infty \frac{\partial}{\partial \rho}\,g_{n+2}(\rho,\psi)\,d\rho\,d\psi\\ &\le \frac{3}{2(n+1)}\int_0^{\pi/2}\cos^{s-1}\psi\,d\psi<\frac{B(s/2,1/2)}{n}. \end{align*} Next, for all $\rho\in (1,2)$ and $\psi\in(\pi/4,\pi/2)$, \begin{align*} \frac{\partial}{\partial \psi}\,g_{n+2}(\rho,\psi)&=2(n+1)\frac{\rho(\rho^2+1)\sin\psi}{(\rho^2+2\rho\cos\psi+1)^2}g_{n}(\rho,\psi)\\ &\ge \frac{(n+1)}{2} \frac{\,g_n(\rho,\psi)}{(\rho^2+2\rho\cos\psi+1)}>0, \end{align*} so \begin{align*} J_{2,n}&\le \frac{2}{(n+1)} \int_1^2 \int_{\pi/4}^{\pi/2}\cos^s\psi \frac{\partial}{\partial \rho}\,g_{n+2}(\rho,\psi)\,d\psi\,d\rho\\ &\le \frac{2}{(n+1)} \int_{\pi/4}^{\pi/2}(\cos^s\psi)'\,d\psi =\frac{2}{2^{s/2}(n+1)}. \end{align*} Finally, since \begin{align*} \sup_{\rho\in (1,2),\,\psi\in (0,\pi/4)} g_n(\rho, \psi) &=g_n(2,\pi/4) =\left(\frac{5-2\sqrt{2}}{5+2\sqrt{2}}\right)^{(n-1)/2} <\frac{2}{n}, \end{align*} we have \[ J_{3,n}\le \frac{2}{n} \int_0^{\pi/4}\cos^s\psi\,d\psi <\frac{B(s/2,1/2)}{n}. \] Summing up the above estimates above, the assertion of the lemma follows. \end{proof} The following corollary showing sharpness of the $\mathcal D$-calculus is immediate. \begin{cor} For any $s>0$, \[ \frac{\|f_n\|_\mathcal{B}}{\|f_n\|_{\mathcal{D}_s}} \asymp \log n, \qquad \frac{\|f_n\|_{{\rm HP}}}{\|f_n\|_{\mathcal{D}_s}} \asymp n^{1/2}, \quad n\to\infty. \] \end{cor} \begin{rem} Curiously, for $s=0$ the asymptotics of $\|f_n\|_{\mathcal D_0}$ match those of $\|f_n\|_{\mathcal B}$, and one does not get any advantages using the $\mathcal D$-calculus in this case. Specifically, one can show that, for $n \in \mathbb{N}$, \begin{equation}\label{ApLog1} 1+2 \log (n+1)\le \|f_n\|_{\mathcal{D}_0}\le 8(4+\log(n+1)). \end{equation} \end{rem}

https://arxiv.org/abs/2302.08174

A Direttissimo Algorithm for Equidimensional Decomposition

We describe a recursive algorithm that decomposes an algebraic set into locally closed equidimensional sets, i.e. sets which each have irreducible components of the same dimension. At the core of this algorithm, we combine ideas from the theory of triangular sets, a.k.a. regular chains, with Gröbner bases to encode and work with locally closed algebraic sets. Equipped with this, our algorithm avoids projections of the algebraic sets that are decomposed and certain genericity assumptions frequently made when decomposing polynomial systems, such as assumptions about Noether position. This makes it produce fine decompositions on more structured systems where ensuring genericity assumptions often destroys the structure of the system at hand. Practical experiments demonstrate its efficiency compared to state-of-the-art implementations.

\section{Introduction} \label{sec:intro} \paragraph{Problem statement} Let $\mathbb{K}$ be an algebraically closed field, let $R = \mathbb{K}[x_1,\dotsc,x_n]$ be a polynomial ring and let~$f_1,\dotsc,f_c \in R$ be a polynomial system generating an ideal~$I\subseteq R$. The zero set~$X$ of the polynomials~$f_1,\dotsc,f_c$ in $\mathbb{K}$, decomposes uniquely as a union of irreducible algebraic sets such that none of them contains another. These are the \emph{irreducible components} of~$X$ and correspond to the \emph{minimal associated primes} of~$I$. The variety~$X$ is \emph{equidimensional} if all its irreducible components have the same dimension. It is clear that~$X$ always admits a decomposition~$X = Y_1 \cup \dotsb \cup Y_c$ where the~$Y_i$ are equidimensional algebraic sets. Given $f_1,\dots,f_c \in R$, we aim at computing such an \textit{equidimensional decomposition} of $X = V(f_1,\dots,f_c)$. Moreover, if the input polynomials have coefficients in a subfield~$\mathbb{L}\subseteq \mathbb{K}$, we want the components of the decomposition to be defined by polynomials over~$\mathbb{L}$. This problem finds natural applications in singularity analysis of sensor-based controllers of mechanism design \parencite[e.g.][and references therein]{pascualescudero:hal-03070525, garciafontan:hal-03499974}, in algorithms of real algebraic geometry \parencite[e.g.][]{AUBRY2002543, el2004properness} and real algebra \parencite{SYZ18, SAFEYELDIN2021259} as well as automated theorem proving and geometry \parencite[e.g.][]{wu2007mathematics, yang2001complete, yang1998automated, chen2013application}. \paragraph{Prior Works} The importance of this computational problem fostered a vast body of literature often also as an intermediate step towards primary decomposition of ideals or prime decomposition of varieties. Algorithms for equidimensional decomposition of algebraic sets can be classified along the data structures which they employ to represent (equidimensional) algebraic sets. There are two prominent strategies for equidimensional decomposition using Gröbner bases frequently implemented in computer algebra systems. The first one uses algebraic elimination techniques. It combines the knowledge of the dimension of the ideal generated by the input polynomials with the elimination theorem \parencite[Theorem~3.1.2]{CoLiOS07} to compute a description of the projection of the algebraic set under study on a well-suited affine linear subspace to deduce how to split the corresponding ideal \parencite{gianni1988, krick1991, caboara1997, decker1999, KALKBRENER1994365}. The projection of the equidimensional component of highest dimension (frequently called the \textit{equidimensional hull}) of the algebraic set under question will then be cut out by a hypersurface whose defining polynomial has degree equal to the degree of this equidimensional hull. As a consequence, such algorithms have the disadvantage that they need to manipulate polynomials of degree in the order of the Bezout bound of the input system. To circumvent this drawback, another set of methods, called \emph{direct methods} has been introduced by \textcite{eisenbud1992}. They rely on homological algebra to reduce the problem of equidimensional decomposition to the computation of syzygies which are then used to split the polynomial ideal under study, while avoiding projections. These algorithms often provide an intermediate step towards primary decomposition of ideals. For this problem, modular techniques and dedicated algorithms for the case where $\mathbb{K}$ is a finite field have been designed \parencite{NORO20041227, yokoyama2002prime, ishihara2022modular}. Another body of work uses {\em lazy} representations of algebraic sets. Frequently, the core idea is to exploit the fact that any equidimensional algebraic set is locally a complete intersection, i.e. an equidimensional algebraic set of codimension $c$ can be represented by the vanishing of $c$ polynomials on a dense Zariski open subset of itself. Hence, equidimensional algebraic sets can be understood as the Zariski closures of locally closed sets defined by polynomial equations and inequations. Taking this perspective, one additionally enforces these $c$ polynomial equations to form a {\em triangular set}. These have their origin in the Wu-Ritt characteristic sets \parencite{ritt1950a, wu1986, chou1990, wang1993, Gallo1991}. Triangularity is therein understood with respect to the variables of the underlying polynomial ring (i.e. in a sense analogous to the notion of triangular matrices in linear algebra). This triangular structure naturally also yields the equations of the algebraic set where the $c$ polynomials fail to define the algebraic set at hand and thus triangular sets have a description of the previously mentioned Zariski open subset attached to them in a natural way. Because of their triangular structure they allow the reduction of certain algorithmic challenges to a univariate problem. Of particular importance, especially in the realm of equdimensional decomposition, are certain special triangular sets called \textit{regular chains}, introduced by \textcite{kalkbrener1993, lu1994}. A regular chain models an \textit{unmixed} dimensional ideal and has good algorithmic properties with respect to the ideal it represents. A related frequently implemented algorithm was also given by \textcite{lazard1991}. Algorithms using regular chains are prominently part of the computer algebra system Maple \parencite{chen2007a, chen2012}. We refer to \textcite{wang2001, hubert2003} for introductions to the subject and to \textcite{aubry1999} for a theoretical account as to how certain different notions of triangular sets relate to each other. It should be noted that, as for methods based on Gr\"obner bases combined with algebraic elimination, these triangular encodings make use of polynomials whose degrees, in the worst case, can be as high as is the degree of the equidimensional components they do encode. Nonetheless, algorithms based on triangular representations can be quite well behaved compared to Gröbner basis techniques especially on certain sparse polynomial systems. Another data structure naturally encoding equidimensional algebraic sets is that of a \textit{geometric resolution} developed by \textcite{lecerf2000, lecerf2003}. A geometric resolution is a certain zero-dimensional parametrization of an algebraic set in Noether position. In our setting, these zero-dimensional parametrizations are used to encode generic points in the equidimensional components of the algebraic set under study (the numerical counterpart of this encoding is known as the notion of witness sets \parencite{sommese2005}, a notion that will be utilized in this paper as well). Under certain generically satisfied assumptions on the input, these can be combined with \textit{straight line programs} to obtain the best known complexity bounds for equidimensional decomposition. See also \parencite{jeronimo2002} for a related approach. To bypass the ``projection-degree'' problem, incremental ap\-proa\-ches have been investigated in combination with Gr\"obner bases algorithms. Incremental means here that they feed the decomposition algorithm with one input polynomial after another, in the same way as \textcite{lazard1991} or \textcite{lecerf2000}, for example, to identify when some polynomial is a zero divisor in the ring of polynomials quotiented by the ideal generated by the previous polynomials. \Textcite{moroz2008} combines Gr\"obner bases computations with representations of equidimensional algebraic sets by means of locally closed sets. In a previous work, we also investigated this approach by exploiting properties of signature-based Gr\"obner bases algorithms to enhance the detection and exploitation of zero divisors and compute the so-called nondegenerate locus of a polynomial system \parencite{eder2022}. \paragraph{This Work} In this work, we again take the incremental approach previously mentioned. As in the other incremental algorithms, the foundation of our algorithm is a decomposition algorithm to, given an equidimensional algebraic set $X$ and some $f\in R$, determine the \textit{locus of proper intersection} of $f$ on $X$, i.e. the set of points $p\in k^n$ such that $X\cap V(f)$ has dimension one less than $X$. This is then used to iterate over the input equations $f_1,\dots,f_r$. More precisely, one starts by decomposing $V(f_1,f_2)$ then uses the output to decompose $V(f_1,f_2,f_3)$ and so on. In contrast to a lot of other algorithms for equidimensional decomposition based on Gröbner bases we borrow from the theory of triangular sets and work with locally closed sets instead of polynomial ideals simlarly to \textcite{moroz2008}. In the iterative approach outlined above this turns out to have two benefits. First, it naturally removes from the output sets of our iterative algorithm certain embedded components that appear during the decomposition. To illustrate this consider the following example: \begin{example} Let $R:=\mathbb{Q}[x,y,z]$, $X:=V(xy), f:=xz$. To decompose $X\cap V(f)$ into equidimensional components one may start by decomposing $X = V(x) \cup V(y)$. Then one intersects these two components with $V(f)$ to obtain the equidimensional decomposition $X\cap V(f) = V(x) \cup V(y,xz)$. The latter set has the irreducible component $V(y,x)$ which is embedded in $V(x)$. If one instead splits into a {\em disjoint} union $X = V(x) \cup \left[ V(y)\setminus V(x) \right]$ and again intersects both components with $V(f)$ one obtains $X = V(x) \cup \left(V(y,z)\setminus V(x) \right)$ and the latter component no longer has the irreducible component $V(y,x)$. \end{example} Second, an iterative equidimensional decomposition algorithm may produce redundant components, which, if they are not deduplicated, may yield an exponential blow-up in the number of components: if one has decomposed $X = \bigcup_i X_i$ with the $X_i$ sharing a large number of irreducible components then decomposing each $X_i\cap V(f)$ to obtain a decomposition of $X\cap V(f)$ results in an even more redundant decomposition. Because we use locally closed sets to model our equidimensional sets we are enabled to enforce that every time we decompose a locally closed set the resulting output sets be pairwise set-theoretically disjoint. Our experiments indicate that this seems to enforce a sufficiently strong irredundancy between our components to avoid an exponential blow-up in the number of components. In this paper we provide two methods to work with the locally closed sets appearing in our algorithm: One method models them ``naively'' in the sense that we encode them by storing their defining equations and inequations and use Gröbner bases of their associated ideals to work with them algorithmically. The other method tries to avoid having to know a Gröbner basis for the ideal associated to a locally closed set as much as possible by storing instead a Gröbner basis for a \emph{witness set} of the locally closed set in question. Using Gröbner bases here with the graded reverse lexicographical ordering has the effect that, compared to algorithms using triangular sets, we are able to \begin{itemize} \item avoid computing projections of the algebraic sets to be decomposed and certain frequently made genericity assumptions such as ideals being in Noether position; \item obtain desciptions of these sets with lower degree polynomials. \end{itemize} Borrowing further from the theory of triangular sets we also adopt the heuristic that it is a good idea to decompose given algebraic sets as often and as finely as possible when working with them. This philosophy is baked into the recursive structure of our algorithms which exists so as to decompose a given locally closed set as much as possible given generating sets for certain saturation ideals. We implemented our algorithm in the computer algebra system \href{https://oscar.computeralgebra.de/}{\sc Oscar}{} \parencite{Oscar2023} using its interface to the library \href{https://msovle.lip6.fr/}{\tt msolve}{} \parencite{berthomieu2021} for all necessary Gröbner basis computations. Experimental results indicate that our algorithm is able to tackle polynomial systems which are out of reach of state-of-the art implementations of algorithms for equidimensional decomposition which are available in leading computer algebra systems. \section{Algorithms} \label{sec:algop} \subsection{Principles} To illustrate the basic principles behind our equidimensional decomposition algorithm, consider an equidimensional variety~$X$ in the affine space~$\affspace{\mathbb{K}}$. Let~$f \in R$. The variety~$X$ is partitioned into: \begin{enumerate} \item Points~$p$ where~$f$ is a non zero divisor locally at~$p$ (that is in the ring~$R_p/I(X) R_p$). The polynomial $f$ takes nonzero values in any open neighborhood of~$p$ in~$X$.\\ This defines an open subset~$X_\text{proper}$ of~$X$. \item Points~$p$ contained in an irreducible component of~$X$ on which~$f$ vanishes identically. \\ This defines a closed subset~$X_\text{improper}$ of~$X$. \end{enumerate} It is clear that~$X = X_\text{proper} \sqcup X_\text{improper}$ (where~$\sqcup$ denotes a disjoint union) and that~$X_\text{improper} \subseteq V(f)$, so that \[ X\cap V(f) = \left( X_\text{proper} \cap V(f) \right) \sqcup X_\text{improper}. \] By construction, the $ X_\text{proper} \cap V(f)$ is a \emph{proper} intersection: it is equidimensional of dimension~$\dim X -1$, or empty. As a union of irreducible components of~$X$, the closed set $X_\text{improper}$ is equidimensional, with the same dimension as the one of~$X$, unless it is empty. So we obtain an equidimensional decomposition of~$X\cap V(f)$. Given defining equations for $X$, this process can be applied iteratively to obtain an equidimensional decomposition of any affine algebraic variety. In our algorithm we apply the above idea without directly computing $X_{{text}{proper}}$ and $X_{\text{improper}}$. Let $I(X)\subset R$ be an ideal such that $V(I(X)) = X$. Further, we denote by $\satu{I(X)}{f}$ the saturation ideal of $I(X)$ by $f$. Recall that $V(\satu{I(X)}{f})$ is the Zariski closure of $X \setminus V(f)$ \parencite[Theorem~4.4.10]{CoLiOS07}. We look for an element~$g \in \satu{I(X)}{f} \setminus I(X)$. If there is none, this implies that~$X_\text{improper} = \varnothing$ so~$X \cap V(f)$ is equidimensional. If there is such a $g$, then we consider the following partition of~$X$: \begin{enumerate} \item the closed locus~$X_1$ of points~$p$ where~$g$ has nonzero values in any neighborhood of~$p$ in~$X$; \item the open locus~$X_2$ of points~$p$ where~$g$ is zero in some neighborhood of~$p$ in~$X$. \end{enumerate} These two sets are equidimensional. By construction, $fg$ vanishes identically on~$X$, so~$X_1 \subseteq X_\text{improper}$ and this gives the following decomposition of~$X$: \begin{equation} X = X_1 \sqcup \left( X_2\cap V(f) \right).\label{eq:2} \end{equation} The ideal of~$X_1$ is given by~$\satu{I(X)}{g}$. The term $X_2\cap V(f)$ may not be equidimensional but we may apply the above idea recursively: We again split $X_2$ along an element in $\satu{I(X_2)}{f}\setminus I(X_2)$ if it exists. This leads to Algorithm \emph{split}. The set~$X_2$ is not closed, this raises the need to deal not only with closed sets of the affine space, but more generally locally closed sets. We do so by partitioning them into special locally closed sets, more precisely into closed sets in the complement of a hypersurface in the affine space, which we call \emph{affine cells}. Concretely, suppose that $I(X_1) = \satu{I(X)}{g}$ is given by a finite generating set $H\sqcup \{h\}\subset R$. We then recursively decompose $X_2 = X\setminus V(H\sqcup \{h\})$ via \[ X_2 = X \setminus V(H \cup \left\{ h \right\}) = ( X\setminus V(h)) \sqcup \big( (X\setminus V(H)) \cap V(h) \big). \] The intersection with~$V(h)$ is computed with \emph{split} to ensure equidimensionality. Algorithm~\emph{remove} below performs these operations. Findally we obtain an equidimensional decomposition algorithm following an incremental strategy by repeated application of \emph{split}, see Algorithm~\emph{equidim}. The primitive operations we use to manipulate affine cells are presented next, while the proof of correctness and termination of the algorithms are in Section~\ref{sec:corterm}. \subsection{Primitives} \label{sec:primitives} \begin{definition} \label{def:hypcomp} An \emph{affine cell}~$X$ is a locally closed set of~$\mathbb{K}^n$ of the form~$Z\setminus V(g)$ where~$Z$ is an algebraic set and~$g\in R$. An affine cell~$X$ is \emph{equidimensional} if all the irreducible components of the Zariski closure~$\overline X$ have the same dimension. \end{definition} Regardless of the mode of representation of an affine cells, we assume that we can perform the following operations on any affine cell~$X$: \begin{enumerate}[(1)] \item\label{it:inter} Given~$f\in R$, compute the affine cell~$X\cap V(f)$; \item\label{it:complement} Given~$f\in R$, compute the affine cell~$X \setminus V(f)$; \end{enumerate} As often in effective algebraic geometry, algebraic sets are defined by ideals that are not always radical so our affine cells come with a distinguished ideal~$I(X)\subseteq R$ such that~$\overline X = V(I(X))$. The radical of~$I(X)$ is denoted~$\operatorname{rad} I(X)$. We assume that operations~\ref{it:inter} and~\ref{it:complement} satisfy $I(X) + \langle f\rangle \subseteq I( X\cap V(f) )$ and~$I(X) \subseteq I(X \setminus V(f))$. We assume further that we can perform the following operations on any affine cell~$X$: \begin{enumerate}[(1), resume] \item Given~$f\in R$, decide if~$f\in I(X)$; \item Compute a basis of $I(X)$, denoted~\fn{basis}(X). \end{enumerate} For example, we may represent an affine cell~$X$ by a pair~$(F, g)$, where~$F$ is a Gröbner basis of~$I(X)$, for some monomial ordering, and~$g$ a polynomial such that~$X = \overline{X}\setminus V(g)$ \parencite[see][for an introduction to Gröbner bases]{becker1993}. We denote~$X = V(F; g)$. For a set $F\subseteq R$ and an element $g\in R$, let~$\fn{sat}(F, g)$ denote a Gröbner basis of the saturation ideal~$\satu{\langle F \rangle}{g}$. Recall that \[ \satu{I}{g} \stackrel{\text{def}}= \left\{ f\in R \ \middle|\ \exists k\in \mathbb{N}, fg^k \in I \right\}.\] Using these primitive \fn{sat}, we can perform all the four operations above: \begin{enumerate}[(1)] \item $V(F; g) \cap V(f) = V( \fn{sat}(F \cup \left\{ f \right\}, g); g)$; \item $V(F; g) \setminus V(f) = V( \fn{sat}(F, f); fg)$; \item $f \in I(X)$ if and only if the normal form of~$f$ w.r.t.~$F$ is zero; \item $\fn{basis}(V(F; g)) = F$. \end{enumerate} \begin{remark} In Section \ref{sec:implem} we explain how to perform the above primitive operations on an affine cell $X$ using a notion called \textit{witness sets}, introduced for the purpose of equidimensional decomposition by \textcite{lecerf2003} under the name lifting fibers. This leads to a lazier representation of $X$, one where a Gröbner basis for $I(X)$ is not always required. \end{remark} \begin{algorithm}[H] \caption{Equidimensional decompositions} \label{alg:split} \raggedright \begin{description} \item[Input] An equidimensional affine cell~$X$, an element~$f\in R$ \item[Output] A partition of~$X\cap V(f)$ into equidimensional affine cells \end{description} \begin{pseudo} function \fn{split}(X, f)\\+ $G \gets \fn{basis}( X\setminus V(f))$\\ if $G \subseteq I(X)$ \label{line:testproper} \ct{can be replaced by~$G\subseteq \operatorname{rad} I(X)$}\\+ return $\left\{ X\cap V(f) \right\}$ \label{line:proper}\\- else\\+ $g\gets $ \tn{any element of~$G \setminus I(X)$} \label{line:pickg}\\ $H \gets \fn{basis}(X \setminus V(g))$\\ $\mathcal{D} \gets\left\{ X\cap V(H) \right\}$ \label{line:improper}\\ for $Y \in \fn{remove}(X \cap V(g), H)$ \label{line:addg}\\+ $\mathcal{D} \gets \mathcal{D} \cup \fn{split}(Y, f)$\label{line:reccall}\\- end\\ return $\mathcal{D}$\\- end\\- end \end{pseudo} \bigskip \begin{description}[labelsep=1em] \item[Input] An affine cell~$X$, a finite set~$H \subset R$ \item[Precondition] $X\setminus V(H)$ is equidimensional \item[Output] A partition of~$X\setminus V(H)$ into equidimensional affine cells \end{description} \begin{pseudo} function \fn{remove}(X, H)\\+ if $H = \varnothing$\\+ return $\varnothing$\\-\ else\\+ $h\gets $ \tn{any element of~$H$}\\ $\mathcal{D} \gets \left\{ X \setminus V(h) \right\}$\\ for $Y \in \fn{remove}(X, H\setminus \left\{ h \right\})$\\+ $\mathcal{D} \gets \mathcal{D} \cup \fn{split}(Y, h)$\\- end\\ return $\mathcal{D}$\\- end\\- end \end{pseudo} \bigskip \begin{description}[labelsep=1em] \item[Input] a finite set $F \subseteq R$ \item[Output] A partition of~$V(F)$ into equidimensional affine cells \end{description} \begin{pseudo} function \fn{equidim}(F)\\+ $\mathcal{D} \gets \left\{ V(\varnothing; 1) \right\}$ \ct{the full affine space}\\ for $f$ in $F$\\+ $\mathcal{D} \gets \bigcup_{X\in \mathcal{D}} \fn{split}(X, f)$\\- end\\ return $\mathcal{D}$\\- end \end{pseudo} \end{algorithm} \begin{example} \label{exp:split} To illustrate Algorithm \emph{split} we spell out how it behaves on the input $X := V(xy,zw)$ and $f:=xz$. Using the notation of Algorithm \emph{split} we find $G = \{y,w\}$. This is not contained in $I(X)$, so we may choose $g := y$ in line 6 of Algorithm \emph{split}. Then we find $H = \{x,zw\}$. Note that $X\setminus V(zw) = \emptyset$ and so Algorithm \emph{split} returns \begin{align*} &X\cap V(H) \text{ and }\text{\emph{split}}(\text{\emph{remove}}(X,H),f)\\ =\; &V(zw,x) \text{ and } \text{\emph{split}}(V(y,zw)\setminus V(x), xz). \end{align*} This second call to Algorithm \emph{split} finds $G = \{y,w\}$, again this set is not contained in $I(V(y,zw)\setminus V(x))$, and so we can choose $g := w$ in line 6. Then we find $H = \{z\}$ which this time yields \begin{align*} \text{\emph{split}}(V(y,zw)\setminus V(x), xz) = &V(y,z)\setminus V(x) \\ &\text{and } \text{\emph{split}}(V(y,w)\setminus V(xz),xz) \end{align*} The last call to split simply finds the empty set and so all in all we have obtained the decomposition \begin{align*} V(xy,zw,xz) = V(x,zw) \cup V(y,z)\setminus V(x). \end{align*} \end{example} \begin{remark} Example \ref{exp:split} illustrates the fact that Algorithm \emph{split} may split an algebraic set even if it is equidimensional. Heuristically, the finer the intermediate decomposition in Algorithm \emph{equidim} is, the computationally easier subsequent steps will be. \end{remark} \subsection{Correctness and Termination} \label{sec:corterm} When computing an interection of an equidimensional affine cell~$X$ with a hypersurface~$V(f)$, we distinguish two cases, depending on whether $V(f)$ intersects~$X$ properly or not. Lemma~\ref{lem:regint} deals with the first case, while Lemma~\ref{lem:irregular-step} deals with the second. \begin{lemma} \label{lem:regint} Let~$X$ be an equidimensional affine cell. Let~$f \in R$, such that $\satu{I(X)}{f} \subseteq \operatorname{rad} I(X)$. Then~$X \cap V(f)$ is empty or equidimensional with dimension~$\dim X - 1$. \end{lemma} \begin{proof}Let~$I = I(X)$. Suppose that $X \cap V(f)$ is not empty. By Krull's principal ideal theorem any minimal prime over $I + \langle f \rangle$ has codimension at most $\operatorname{codim}I + 1$. The condition~$\satu{I}{f} \subseteq \operatorname{rad} I$ means geometrically that~$X\subseteq \overline{X\setminus V(f)}$, so that $f$ has nonzero values in the neighborhood of any point in~$X$. So~$f$ is a not a zero divisor in~$R/I$. In particular, there is a regular sequence of length $\operatorname{codim}I + 1$ in $I + \langle f \rangle$. Since the polynomial ring $R$ is Cohen-Macaulay it follows that every minimal prime over $I + \langle f \rangle$ has at least codimension $\operatorname{codim} I + 1$. \end{proof} \begin{lemma}\label{lem:irregular-step} Let~$X$ be an equidimensional affine cell. Let~$f \in R$, let~$g \in \satu{I(X)}{f}$ and let~$I_g = \satu{I(X)}{g}$. Let~$X_1 = X\cap V(I_g)$ and~$X_2 = (X \cap V(g)) \setminus V(I_g)$. Then: \begin{enumerate}[(i)] \item \label{it:part0} $X = X_1\sqcup X_2$; \item \label{it:partition} $X \cap V(f) = X_1 \sqcup \left( X_2 \cap V(f) \right)$ ; \item \label{it:equidim1} $X_1$ is empty or equidimensional with~$\dim X_1 = \dim X$; \item \label{it:equidim2} $X_2$ is empty or equidimensional with~$\dim X_2 = \dim X$; \end{enumerate} \end{lemma} \begin{proof} Obviously $X = X_1 \sqcup (X \setminus V(I_g))$. As a set, $X_1$ is the union of the components of~$X$ on which~$g$ is not identically zero. In particular~$X \setminus V(I_g)$ is the set of points of~$X$ in a neighborhood of which~$g$ is identically zero. Therefore~$X \setminus V(I_g) \subseteq V(g)$, so we obtain \[ X \setminus V(I_g) = (X \cap V(g)) \setminus V(I_g), \] which gives~\ref{it:part0}. Next, we have~$I(X_1) = I(X) + I_g = \satu{I(X)}{g}$. Moreover~$f \in \operatorname{rad} I(X_1)$. Indeed, $g f^k \in I(X)$ for some~$k \geq 0$, by definition of~$g$, and therefore~$f \in \operatorname{rad} \left( I(X) : g \right) \subseteq \operatorname{rad}\satu{I(X)}{g}$. So~$X_1 \subseteq V(f)$. It follows that~$X \cap V(f) = X_1 \sqcup \left( X_2 \cap V(f) \right)$. This proves~\ref{it:partition}. Since~$X$ is equimensional, it follows that~$X_1$ (as a union of components of~$X$) is also equidimensional of same dimension, unless it is empty. This proves~\ref{it:equidim1}. As for~$X_2$, it is open in~$X$, so it inherits the equidimensionality and the dimension of~$X$, unless it is empty. This proves~\ref{it:equidim2}. \end{proof} We now prove correctness and termination of Algorithms \emph{split} and \emph{remove} with a mutual induction. On line~\ref{line:testproper}, the test~$G\subseteq I(X)$ can be replaced by~$G\subseteq \operatorname{rad} I(X)$, or any condition which holds when~$G\subseteq I(X)$ and doee not hold when~$G\not\subseteq \operatorname{rad} I(X)$, this does not affect correctness or termination. We will use this variant in Section~\ref{sec:implem}. \begin{theorem}\label{lem:term-correc-technical} For any affine cell~$X$: \begin{enumerate}[(i)] \item \label{it:split} If~$X$ is equidimensional, then for any~$f\in R$, the procedure~\emph{split} terminates on input~$X$ and~$f$ and outputs a partition of\/~$X\cap V(f)$ into equidimensional affine cells~$Y$ with~$I(X) \subseteq I(Y)$. \item \label{it:remove} For any finite set~$H\subset R$ such that~$X \setminus V(H)$ is equidimensional, the procedure~\emph{remove} terminates on input~$X$ and~$H$ and outputs a partition of\/~$X\cap V(H)$ into equidimensional affine cells~$Y$ with~$I(X) \subseteq I(Y)$; \end{enumerate} \end{theorem} \begin{proof} We proceed by Noetherian induction on~$I(X)$ and assume the statement holds for any affine cell $X'$ with~$I(X) \subsetneq I(X')$. We begin with \fn{split}. Let~$f \in R$ and let~$I_f = \satu{I(X)}{f}$. If~$I_f \subseteq I(X)$, then Lemma~\ref{lem:regint} applies and~$X\cap V(f)$ is equidimensional. So $\fn{split}(X, f)$ terminates and is correct in this case. Assume now that there is some~$g\in I_f \setminus I(X)$. Let~$I_g= \satu{I}{g}$. Lemma~\ref{lem:irregular-step} applies: an equidimensional decomposition of~$X\cap V(f)$ is given by~$X\cap V(I_g)$ and an equidimensional decomposition of~$\left( (X\cap V(g))\setminus V(I_g) \right)\cap V(f)$. Moreover~$(X\cap V(g)) \setminus V(I_g)$ is equidimensional. Since~$g\not\in I(X)$, we have~$I(X) \subsetneq I(X\cap V(g))$ so \fn{remove}(X\cap V(g), H) (using the notations of Algorithm \emph{split}, where~$H$ is a generating set of~$I_g$) is correct and terminates, by induction hypothesis. Moreover, it outputs affine cells~$Y$ such that~$I(X) \subsetneq I(X \cap V(g)) \subseteq I(Y)$. So the recursive calls \fn{split}(Y, f) are correct and terminate. As for \fn{remove}, let~$H \subset R$ finite such that~$X \setminus V(H)$ is equidimensional. If~$H = \varnothing$, then~\ref{it:remove} holds trivially. As for the case~$H\neq \varnothing$, let~$ h\in H$ and~$H' = H\setminus h$. Since~$V(H) = V(h) \cap V(H')$, we have \begin{equation} X \setminus V(H) = (X \setminus V(h)) \sqcup \left( (X \setminus V(H')) \cap V(h) \right).\label{eq:1} \end{equation} The set $X\setminus V(h)$ and $X\setminus V(H')$ are open in~$X\setminus V(H)$ so equidimensional (or empty). By induction on the cardinal of~$H$, we assume that \fn{remove}(X, H') is a partition of~$X \setminus V(H')$ into equidimensional affine cells, and that every cell $Y$ of this partition satisfies~$I(X) \subseteq I(Y)$. By~\ref{it:split}, the calls \fn{split}( Y, h) terminates and yield a partition of $(X \setminus V(H')) \cap V(h)$ into cells~$Y$ with~$I(X) \subseteq I(Y)$. Moreover the affine cell~$Y = X\setminus V(h)$ also satisfies~$I(X) \subseteq I(Y)$. By~\eqref{eq:1}, \fn{remove}(X, H) terminates too and is a partition of~$X \setminus V(H)$ into cells~$Y$ with~$I(X) \subset I(Y)$. \end{proof} \begin{corollary} Algorithm \emph{equidim} is correct and terminates. \end{corollary} \section{Implementation and experimental results} \label{sec:implem} \subsection{Implementation Details} \label{sec:algdet} In this section we give some implementation details and alternatives. In particular, we show a lazier data structure for affine cells which is able to delay some Gröbner basis computations at the cost of a Monte Carlo randomization. We have implemented both the method described in Section \ref{sec:algop} and the method described in this section. For either method, we will need an algorithm that, given generators for an ideal $I$ and an element $f\in R$, computes generators for the saturation $\satu{I}{p}$. Even for our lazy representation, this will still sometimes be needed to compute a Gröbner basis for the ideal $I(X)$, where $X$ is an affine cell. In the probabilistic setting, some saturations will be replaced by saturations of zero dimensional ideals. In our implementation we chose the standard method of performing saturations using Gröbner bases. To compute generators for $\satu{I}{p}$, fix a monomial order $\leq$ on $R[t]$ for a new variable $t$ such that $\leq$ eliminates $t$. Compute a Gröbner basis $G$ for the ideal $I + \langle tp - 1\rangle \subset R[t]$ w.r.t $\leq$. Then the elements in $G$ that do not contain the variable $t$ give a Gröbner basis of $\satu{I}{p}$ by the elimination theorem. Other saturation methods also exist such as the methods presented in \textcite{eder2022} or \textcite{berthomieu2022}. Randomization relies on intersecting with random linear subspaces of appropriate dimension to reduce to the zero-dimensional case. This idea is well known in symbolic computation \parencite{lecerf2003} and numerical algebraic geometry \parencite[e.g.]{bates2013} wherein these intersections of algebraic sets with random suitable random linear subspaces are known under the name \emph{witness sets}. \begin{proposition}\label{prop:proba} Let $X\subseteq \affspace{\mathbb{K}}$ be an equidimensional affine cell of dimension $d$ and let $f\in R$. Then, for a generic linear subspace $L\subset \affspace{\mathbb{K}}$ of codimension $d$ the following statements hold: \begin{enumerate} \item $f\in \operatorname{rad}{I(X)}$ if and only if $f \in \operatorname{rad}{I(X\cap L)}$. \item $I(X\setminus V(f)) \subseteq \operatorname{rad}{I(X)}$ if and only if $X \cap L \cap V(f) = \varnothing$. \end{enumerate} \end{proposition} \begin{proof} We always have~$\operatorname{rad} I(X) \subseteq \operatorname{rad} I(X \cap L)$. Conversely, assume that~$f\not\in \operatorname{rad} I(X)$. Let~$U = \left\{ p\in X\ \middle|\ f(p) \neq 0 \right\}$/ It is an open subset of~$X$ and it is non empty, by hypothesis. Since~$X$ is equidimensional, $U$ has dimension~$d$ and the intersection~$U\cap L$ is nonempty (because~$L$ is generic). Therefore~$f$ is nonzero on a nonempty subset of~$X\cap L$. In particular, $f \not \in \operatorname{rad} I(X\cap L)$. This proves the first point. For the second point note that $I(X\setminus V(f)) \subseteq \operatorname{rad}{I(X)}$ if and only if~$X$ and~$V(f)$ intersect properly, that is~$X\cap V(f)$ is equidimensional of dimension~$d-1$. The intersection of~$X\cap V(f)$ with the codimension~$d$ generic space~$L$ is empty if and only if the dimension of~$X\cap V(f)$ is less than~$d$. The proves the second point. \end{proof} In this setting, we represent an equidimensional affine cell~$X$ by a triple~$(F, G, W, d)$, where~$F$, $G$ and~$W$ are subsets of~$R$ and~$d$ is an integer such that~$\dim X = d$, $X = V(F) \setminus V(\prod_{g\in G} G)$ and~$W$ (stands for \emph{witness set}) is a Gröbner basis of~$I(X \cap L)$ for some generic linear subspace space~$L$ of~$\mathbb{K}^n$ of codimension~$d$. We denote~$X = V(F; G, W, d)$. In practice, $L$ will only be random and sufficient genericity will only hold with high probability (assuming that~$\mathbb{K}$ has enough elements). Given only $F$, $G$ and~$d$, we can compute a suitable set~$W$ by choosing a set~$J\subseteq R$ of~$d$ random linear forms and computing a Gröbner basis of~$(((F : g_1^\infty) : \dotsb): g_r^\infty)$, where~$G = \left\{ g_1,\dotsc,g_r \right\}$. This procedure is denoted \fn{witness}(F, G, d). The four primitive operations are perfomed as follows. For the intersection operation, we need some additional knowledge on the expected dimension of the output. Let~$X = V(F; G, W, d)$ be an equidimensional cell. \begin{enumerate}[(1)] \item \emph{[Proper intersection]} Given $f\in R$ such that~$X$ intersects~$V(f)$ properly, \[ X \cap V(f) = V\big(F' ; G, \fn{witness}(F', G, d-1), d-1\big), \] with~$F' = F \cup \left\{ f \right\}$; \item[(1')] \emph{[Purely improper intersection]} Given~$H\subset R$ such that~$X \cap V(H)$ is a union of components of~$X$, \[ X \cap V(H) = V\big(F \cup H ; G, \fn{gb}(W \cup H), d \big), \] where \fn{gb}(W\cup H) denotes a Gröbner basis of the ideal generated by~$X\cup H$; \item for~$f\in R$, $X\setminus V(f) = V(F; G\cup \left\{ f \right\}, \fn{sat}(H, f), d)$; \item $f\in \operatorname{rad} I(X)$ if and only if~$1 \in (W : f^\infty)$; \item $I(X)$ is computed by saturating~$\langle F\rangle$ successively by all the elements of~$G$. \end{enumerate} In the decomposition algorithm, we always know \emph{a priori} the kind of each intersection. The intersection on line~\ref{line:proper} of \fn{split} is proper, the intersection on line~\ref{line:improper} is purely improper. The one on line~\ref{line:addg} is more subtle. Indeed, the decomposition algorithm may produce here a nonequidimensional cell when considering~$X\cap V(g)$. With the notations of this algorithm, the cell~$X' = X\cap V(g)$ is only equidimensional outside of~$V(H)$ (of dimension~$\dim X$). This nonequidimensional cell will go through only one operation among the four primitives: $X'\setminus V(h)$ for some~$h\in H$. This operation restores equidimensionality. So we can mostly ignore this issue and compute the intersection $X\cap V(g)$ as a purely improper intersection, pretending that~$X\cap V(g)$ is equidimensional. In addition we obtain a fifth operation: a probabilistic algorithm to check if $X\cap V(f)$ is empty or equidimensional of dimension one less than $X$ (or, equivalently $\satu{I(X)}{f} \subseteq \operatorname{rad} I(X)$). This is given by Algorithm \emph{isProper}. Equipped with this algorithm, we can replace the if-condition in line~\ref{line:testproper} of Algorithm \emph{split} with $\fn{isProper}(X,f)$. Only if this is not satisfied we proceed to compute a Gröbner basis for $I(X\setminus V(f))$. \begin{algorithm}[] \caption{Proper intersection check} \label{alg:isreg} \raggedright \begin{description}[labelsep=1em] \item[Input] An equidimensional affine cell~$X$, an element~$f\in R$ \item[Output] \kw{true} if $X\cap V(f)$ is a proper intersection, \kw{false} otherwise \end{description} \begin{pseudo} function \fn{isProper}(X, f)\\+ $W \gets $ \tn{the withness set of~$X$}\\ $W' \gets$ \tn{a Gröbner basis of $\satu{\langle W\rangle}{f}$}\\ return $1 \in W'$ \end{pseudo} \end{algorithm} Lastly we want to note the following: In Algorithm \emph{split}, on input $X$ and $f$, we may have to compute $G := \fn{basis}(X\setminus V(f))$ but we use only one element in $G$ in line~\ref{line:pickg} of Algorithm \emph{split}. This situation can be improved by a simple caching mechanism: Note that in line~\ref{line:reccall} of Algorithm \emph{split} we call $\fn{split}(Y,f)$ with affine cells $Y$ satisfying $Y \subset X$. This certainly means $G = \fn{basis}(X\setminus V(f)) \subseteq \fn{basis}(Y \setminus V(f))$. Hence we may first try to pick an element from the already computed set $G$ in 6 of the call $\fn{split}(Y,f)$ before computing $\fn{basis}(Y\setminus V(f))$. \subsubsection{Rationale for the new Data Structure} \label{sec:rat} Always knowing a Gröbner basis for the affine cells appearing in Algorithm \emph{equidim} puts a large penalty on the cost of our algorithms. This is actually related to well-known observations on the complexity of Gröbner bases under some regularity assumptions. Indeed, for a regular sequence in strong Noether position, the cost of linear algebra steps needed to compute intermediate Gröbner bases in an incremental manner is higher than the final steps \parencite{BARDET201549}. Dimension dependent complexity bounds provide another confirmation of this behaviour \parencite{10.1145/3087604.3087624}. Using witness sets we can potentially avoid a lot of intermediate Gröbner basis computations in our algorithms. In our experience, for a large number of cases, using witness sets greatly improves the efficiency of our algorithm which is theoretically backed up by the previously mentioned complexity results. Furthermore, in the data structure for affine cells presented in the last subsection we store the definining inequation of our affine cells as a factorization. If one wants to saturate a polynomial ideal $I$ by an element $f\in R$ which is known to have a factorization $f = \prod_{g\in G}g$ given by a finite set $G$ then it is expected to be cheaper to saturate by the elements $g\in G$ one-by-one instead of saturating by $f$ directly using the above elimination method. This lowers the degrees of the polynomials involved. \subsubsection{A Better Version of \emph{remove}} \label{sec:rem} Furthermore, we encountered the following problem when implementing Algorithm \emph{remove} as presented in Section \ref{sec:algop}. In this algorithm~$H$ is a Gröbner basis so it tends to be very redundant (that is very far from being a minimal set of generators). So it often happens that there are two or more elements $h_1,h_2\in H$ such that for $X_1 := X\setminus V(h_1)$ and $X_2:=X\setminus V(h_2)$ we have $I(X_1) = I(X_2)$. Eventually the sets $X_1$ and $X_2$ become disjoint, since eventually $X_1$ is intersected with $V(h_1)$ or $X_2$ is intersected with $V(h_1)$, but Algorithm \emph{split} may have to split $X_1$ and $X_2$ before that happens. Since splitting an affine cell with Algorithm \emph{split} depends only on the underlying ideals we may then repeat the exact same operations on the level of ideals twice or more. This issue then compounds exponentially due to the recursive nature of our algorithms. We therefore modified Algorithm \emph{remove} to obtain disjoint equidimensional affine cells from $X_1$ and $X_2$ as fast as possible, resulting in Algorithm~\ref{alg:nznew}. Note that when we use witness sets this algorithm avoids knowing Gröbner bases for the ideals $I(X_i)$ until potentially line 12. \begin{algorithm}[] \caption{remove'} \label{alg:nznew} \raggedright \begin{description}[labelsep=1em] \item[Input] An affine cell $X$, a finite set $H\subset R$ \item[Output] A partition of $X\setminus V(H)$ into equidimensional affine cells \end{description} \begin{pseudo} function \fn{remove'}(X, H)\\+ $\mathcal{D} \gets \emptyset$\\ for $i$ from $1$ to $r$\\+ $X_i \gets X\setminus V(h_i)$\\ $H_i\gets \emptyset$\\ for $j$ from $1$ to $i-1$\\+ if \fn{isProper}(X_i, h_j)\\+ $X_i\gets X_i\cap V(h_j)$\\- else\\+ $H_i\gets H_i\cup \{h_j\}$\\- end \\- end \\ $\mathcal{D}_i\gets $ \tn{decomposition of $X_i\cap V(H_i)$}\\ \hspace{2cm}\tn{by repeated application of \emph{split}}\\ $\mathcal{D} \gets \mathcal{D} \cup \mathcal{D}_i$\\- end\\ return $\mathcal{D}$\\- end \end{pseudo} \end{algorithm} \subsection{Experimental Results} \label{sec:experi} In this section we give some experimental results. We compare the timings of our implementation of Algorithm \emph{equidim} and methods for equidimensional decomposition of algebraic sets available in various computer algebra systems in a table given below. Some of the timings are discussed in more detail in the next section. We compared to the following implementations: \begin{enumerate} \item The function \texttt{Triangularize} from the \texttt{RegularChains} library in Maple \parencite{lemaire2005}, which decomposes a polynomial system into regular chains, \item The function \texttt{equidimensional\_decomposition\_weak} in \href{https://oscar.computeralgebra.de/}{\sc Oscar}{} \parencite{Oscar2023} which is a wrapper around a corresponding \texttt{Singular} function \parencite{singular}. \item The Magma \parencite{bosma1997} functions\\ \texttt{EquidimensionalDecomposition} (corresponding to the column ``Magma'' in the table) and\\ \texttt{ProbablePrimeDecomposition} (corresponding to the column ``Magma (prime dec.)'' in the table) and \item The numerical polynomial systems solver Bertini \parencite{bates2013} which we ran on each system at hand by requesting a witness set decomposition into irreducible components with fixed precision set to Bertini's default value. \end{enumerate} The implementation of our algorithms is itself done in \href{https://oscar.computeralgebra.de/}{\sc Oscar}{} which is written in the programming language Julia \parencite{bezanson2017}. Its source code is available at \begin{center} https://github.com/RafaelDavidMohr/Decomp.jl \end{center} For all necessary Gröbner basis computations we employ the library \href{https://msovle.lip6.fr/}{\tt msolve}{} \parencite{berthomieu2021} for which \href{https://oscar.computeralgebra.de/}{\sc Oscar}{} offers an interface. Our suite of example systems is comprised as follows: \begin{enumerate} \item Cyclic$(8)$, coming from the classical Cyclic$(n)$ benchmark. \item The systems P4L 1 to 3 come from the perspective-four line problem in robotics, see \textcite{garciafontan:hal-03499974}. \item The systems C1 to C3 are certain jacobian ideals of single multivariate polynomials which define singular hypersurfaces. \item Ps$(n)$, encoding pseudo-singularities via polynomials $$f_1,\dots,f_{n-1},g_1,\dots,g_{n-1}$$ with $f_i\in \mathbb{K}[x_1, \ldots, x_{n-2}, z_1, z_2]$, $g_i\in \mathbb{K}[y_1, \ldots, y_{n-2}, z_1, z_2]$, the $f_i$ being chosen as a random dense quadrics, and $g_i$ chosen such that $g_i(x_1,\dots,x_{n-2},z_1,z_2) = f$, i.e. as a copy of $f_i$ in the variables $y_1,\dots,y_{n-2},z_1,z_2$. \item sos$(s, n)$, encoding the critical points of the restriction of the projection on the first coordinate to a hypersurface which is a sum of $s$ random dense quadrics in $\mathbb{K}[x_1, \ldots, x_n]$. $$f, \frac{\partial f}{\partial x_2}, \ldots, \frac{\partial f}{\partial x_n}, \quad f = \sum_{i=1}^s g_i^2.$$ \item sing$(n)$, encoding the critical points of the restriction of the projection on the first coordinate to a (generically singular) hypersurface which is defined by the resultant of two random dense quadrics $A, B$ in $\mathbb{K}[x_1, \ldots, x_{n+1}]$: $$f, \frac{\partial f}{\partial x_2}, \ldots, \frac{\partial f}{\partial x_n}, \quad f = \textrm{resultant}(A, B, x_{n+1}).$$ \item The Steiner polynomial system, coming from \textcite{breiding2020}. \item All remaining examples are part of the BPAS library \parencite{asadi2021}. The BPAS library offers an alternative to the \texttt{RegularChains} library in Maple with special emphasis on paralellism and it will be interesting to compare it to our algorithm in the future. \end{enumerate} To obtain the timings in the table below we almost exclusively used the witness set based data structure for affine cells. Every polynomial system was computed with in characteristic 65521 with the exception of Bertini which, as a numerical piece of software, computes over the complex numbers. Due to this difference a comparison between Bertini's and our timings needs to be considered carefully. We tried to indicate this in the table below by coloring the Bertini column in grey. All computations except for Magma were done on an single core of an Intel Xeon Gold 6244 CPU @ 3.60GHz. All Magma computations were done on a single core of an Intel Xeon E5-2690 @ 2.90GHz. We let every algorithm run for at least an hour or 50 times the time it took for the fastest algorithm to complete the system in question, whichever was bigger. Using the witness sets of our output we also did the following to compare to Bertini: We ran our algorithm in a large random prime charateristic. We then removed the embedded irreducible components from each of our output components and computed the degrees of the output components. This gives us the degree in each dimension of the algebraic set defined by the input. Whenever Bertini reports different degrees, we marked it in the respective column. Due to the randomly chosen large characteristic these degrees should be the same one obtains when considering the algebraic set in question over the complex numbers. In the second column of this table, we additionally provide the number of affine cells that Algorithm \emph{equidim} decomposed the respective system into. All timings in this table are given in seconds. Due to the way we measured the timings of Bertini we can only report them without any decimal places, rounded up. \subsection{Discussion of Experimental Results} \label{sec:disc} We provide here some further information about some of the examples and the behaviour of the different implementations on these examples compared below. Our algorithm, i.e. Algorithm \emph{equidim}, seems to behave best in comparison to the other implementations when the input system is dense in the sense that each of the input equations of the system in question involves most, or all, of the variables. This is the case for cyclic 8, the class of the Ps($\bullet$) systems, the class of the Sing($\bullet$) systems, the class of the sos($\bullet,\bullet$) systems and the Steiner polynomial system. On certain polynomial systems, where each input equation involves only a small subset of the variables, we were able to improve our timings by foregoing the witness set based data structure and instead running a deterministic version of our algorithm akin to the version in Section~\ref{sec:algop}. The improvement we thusly obtained can be explained by the fact that intersecting very sparse systems with random hyperplanes can ``destroy their sparsity'' and make certain Gröbner basis computations much harder. This was the case for the example Leykin-1: Here running the deterministic version improved our timing to 2.6 seconds. The Gonnet and dgp6 polynomial systems demonstrate that our algorithm is highly sensitive to the ordering of the input equations: By default we ran our implementation by iterating over the input equations degree by degree in Algorithm \emph{equidim}. With this ordering, our algorithm did not terminate within several hours of computation. When we changed this ordering on these two examples and sorted the input equations instead by length of support, our algorithm terminated in less than one second on these two examples. The system sys2874 can be attacked by both changing the order of the input equations to be ordered by length of support and by using the deterministic version of our algorithm: Doing this, the timing improved by several orders of magnitude to 0.26 seconds. We also remark that \href{https://oscar.computeralgebra.de/}{\sc Oscar}'s timings improved significantly on the examples sys2449, sys2297 and Leykin-1 (each to less than one second) if one decomposes the radicals of these systems instead of the systems themselves. For the examples KdV and sys2882 we seem to be bottlenecked by very difficult Gröbner basis computations and less by the inherent structure of our algorithm. Informal experiments where we tried to compute just a Gröbner basis for these systems using msolve suggest that even this is a highly non-trivial computation. For these two systems, techniques involving regular chains seem to be vastly superior over anything that involves Gröbner basis computations. All in all, these experiments illustrate that on a wide range of examples, our algorithm performs on average better than state-of-the-art implementations and can tackle some problems which were previously unrea \section*{Acknowledgements} The authors wish to thank Marc Moreno Maza for providing feedback on the BPAS library and its benchmark examples. This work has been supported by the Agence nationale de la recherche (ANR), grant agreements ANR-18-CE33-0011 (SESAME), ANR-19-CE40-0018 (Re Rerum Natura); by the joint ANR-Austrian Science Fund FWF grant agreements ANR-19-CE48-0015 (ECARP) and ANR-22-CE91-0007 (EAGLES); by the EOARD-AFOSR grant agreement FA8665-20-1-7029; by the DFG Sonderforschungsbereich TRR 195; the Forschungsinitiative Rheinland-Pfalz; and by the European Research Council (ERC) under the European Union’s Horizon Europe research and innovation programme, grant agreement 101040794 (10000~DIGITS). hable. \newpage \begin{table*}[p] \label{tbl:compare} \begin{adjustwidth}{-1.5cm}{-1cm} \begin{center} \scriptsize{\input{benchmark.tex}} \end{center} \end{adjustwidth} \bigskip Timings are in seconds, except otherwise indicated. The ratio with respect to the best time is given when the latter is over 1~second. \begin{description}[labelsep=1em] \item[\faHandPaperO] We made some minor preparation of the input (like reordering the input equations, or disabling the probabilistic representation of affine cells) to improve the timing. \item[\faClose] Bertini terminated the computation with an error. \item[\faWarning] The result given by Bertini is not consistent with our result in terms of degree/dimension. \end{description} \end{table*} \clearpage \printbibliography \end{document}

https://arxiv.org/abs/1506.08324

Homotopy theory and generalized dimension subgroups

Let $G$ be a group and $R,S,T$ its normal subgroups. There is a natural extension of the concept of commutator subgroup for the case of three subgroups $\|R,S,T\|$ as well as the natural extension of the symmetric product $\|\bf r,\bf s,\bf t\|$ for corresponding ideals $\bf r,\bf s, \bf t$ in the integral group ring $\mathbb Z[G]$. In this paper, it is shown that the generalized dimension subgroup $G\cap (1+\|\bf r,\bf s,\bf t\|)$ has exponent 2 modulo $\|R,S,T\|.$ The proof essentially uses homotopy theory. The considered generalized dimension quotient of exponent 2 is identified with a subgroup of the kernel of the Hurewicz homomorphism for the loop space over a homotopy colimit of classifying spaces.

\section{Introduction} Let $G$ be a group and $\mathbb Z[G]$ its integral group ring. Every two-sided ideal $\mathfrak a$ in the integral group ring $\mathbb Z[G] $ of a group $G$ determines a normal subgroup $$D(G, {\bf a}):=G \cap (1 + {\bf a})$$ of $G$. Such subgroups are called {\it generalized dimension subgroups}. The identification of generalized dimension subgroups is a fundamental problem in the theory of group rings. In general, given an ideal ${\bf a}$, the identification of $D(G,{\bf a})$ is very difficult, for a survey on the problems in this area see \cite{Gupta}, \cite{MikhailovPassi}. The idea that {\it the generalized dimension subgroups are related to the kernels of Hurewicz homomorphisms of certain spaces} was discussed in \cite{MikhailovPassi}, \cite{MPW}, however, in the cited sources, all application of homotopical methods to the problems of group rings were related to very special cases. In this paper, we apply homotopy theory for a purely group-theoretical result of a more general type, namely to the description of the exponent of generalized dimension quotient constructed for a triple of normal subgroups in any group $G$. Let $G$ be a group and $R,S$ its normal subgroups. Denote ${\bf r}=(R-1)\mathbb Z[G],\ {\bf s}=(S-1)\mathbb Z[G]$. It is proved in \cite{BD} that \begin{equation}\label{bd} D(G,{\bf r}{\bf s}+{\bf s}{\bf r})=[R,S]. \end{equation} The following question arises naturally: how one can generalize the result (\ref{bd}) to the case of three and more normal subgroups of $G$. Our main result is the following. \begin{theorem}\label{main} Let $G$ be a group and $R,S,T$ its normal subgroups. Denote $${\bf r}=(R-1)\mathbb Z[G],\ {\bf s}=(S-1)\mathbb Z[G],\ {\bf t}=(T-1)\mathbb Z[G]$$ and \begin{align*} & \|R,S,T\|:=[R,S\cap T][S,R\cap T][T,R\cap S]\\ & \|\bf r, \bf s,\bf t\|:=\bf r(\bf s\cap \bf t)+(\bf s\cap \bf t)\bf r+\bf s(\bf r\cap \bf t)+(\bf r\cap \bf t)\bf s+\bf t(\bf r\cap \bf s)+(\bf r\cap \bf s)\bf t. \end{align*} Then, for every $g\in D(G, \|{\bf r},{\bf s},{\bf t}\|)$, $g^2\in \|R,S,T\|$, i.e. the generalized dimension quotient \begin{equation}\label{dimension} \frac{D(G, \|{\bf r},{\bf s},{\bf t}\|)}{\|R,S,T\|} \end{equation} is a $\mathbb Z/2$-vector space. \end{theorem} The proof of theorem \ref{main} consists of the following steps. First we show that there exists a space $X$ such that there is a commutative diagram \begin{equation}\label{jq} \xymatrix@M.7em{\frac{R\cap S\cap T}{\|R,S,T\|}\ar@{->}[r]\ar@{=}[d] & \frac{\bf r\cap \bf s\cap \bf r}{\|\bf r,\bf s,\bf t\|}\ar@{->}[d]\\ \pi_2(\Omega X)\ar@{->}[r]^{h_2\Omega}& H_2(\Omega X)} \end{equation} where the lower horizontal map is the Hurewicz homomorphism. Secondly, we show that, for any space $X$, the kernel of the Hurewicz homomorphism $$ \Omega h_2: \pi_2(\Omega X)\to H_2(\Omega X) $$ is a 2-torsion subgroup of $\pi_2(\Omega X)=\pi_3(X)$. There are examples of groups with triples of subgroups such that the generalized dimension quotient \ref{dimension} is non-trivial. Let $F=F(a,b,c)$ be a free group with basis $\{a,b,c\}$. Consider the following normal subgroups of $F$: $$ R=\langle a^2,c\rangle^F,\ S=\langle a,bc^{-1}\rangle^F,\ T=\langle a,b\rangle^F. $$ Then, there exists the following natural commutative diagram $$ \xymatrix@M.7em{\mathbb Z/2\ar@{>->}[r] \ar@{=}[d] & \frac{R\cap S\cap T}{\|R,S,T\|}\ar@{->}[r]\ar@{=}[d] & \frac{\bf r\cap \bf s\cap \bf r}{\|\bf r,\bf s,\bf t\|}\ar@{=}[d]\\ \mathbb Z/2\ar@{>->}[r] & \pi_2(\Omega \Sigma \mathbb RP^2)\ar@{->}[r]^{h_2\Omega}& H_2(\Omega \Sigma \mathbb RP^2)} $$ This example and discussion of a generalization of the considered construction to the case of $>3$ normal subgroups is given in section \ref{lastsection}. Another application of homotopic methods is the following identification of the generalized dimension subgroup (see theorem \ref{modg}): \begin{equation}\label{bd11} D(G,{\bf rs+sr}+({\bf r}\cap {\bf s}){\bf t}+{\bf t}({\bf r}\cap {\bf s}))=[R,S][R\cap S,T]. \end{equation} This generalizes (\ref{bd}), indeed, (\ref{bd}) is equivalent to (\ref{bd11}) for $T=1$. The space $X$ from (\ref{jq}) is the homotopy colimit of the cubic diagram of eight classifying spaces $BG, B(G/R), B(G/S), B(G/T), B(G/RS), B(G/RT), B(G/ST)$. The left vertical isomorphism in (\ref{jq}) is proved in \cite{EM}. In section \ref{section3} we develop the theory of cubes of fibrations in the category of simplicial non-unital rings and correspondence between $n$-cubes of fibrations with crossed $n$-cubes of rings. We obtain ring-theoretical analogs of the result from \cite{ES}. Note that, in this paper, we do not consider the properties of universality of crossed $n$-cubes of rings. The universality property is needed for an explicit description of the homology groups $H_*(\Omega X)$ of homotopy colimits $X$ of classifying spaces (see the proof of theorem 1 in \cite{EM}). For the reason of this paper, namely, for an analysis of generalized dimension subgroups, only crossed properties of the diagrams of rings are enough and these properties are given in section \ref{section3}. \section{Hurewicz homomorphism} \subsection{Two lemmas about squares of abelian groups} First we state two lemmas about squares of abelian groups. These lemmas are advanced versions of well known statements (see \cite[Part 1, 6.2.6]{Faith}). We give them without a prove because it is standard. Authors learned these lemmas from non-formal discussions with Alexander Generalov. Consider a square of abelian groups $\mathcal S$ with induced homomorphisms on kernels and cokernels: $$ \xymatrix@-5mm{&& {\rm Ker}(g)\ar[dd]\ar[rr]^{\tilde f} && {\rm Ker}(g')\ar[dd]&& \\ &&&&&& \\ {\rm Ker}(f)\ar[rr]\ar[dd]^{\tilde g} && A \ar[rr]^f \ar[dd]^{g} && B\ar[dd]^{g'}\ar[rr] && {\rm Coker}(f)\ar[dd]^{\tilde g'}\\ && &\mathcal S & &&\\ {\rm Ker}(f')\ar[rr]&& C\ar[dd]\ar[rr]^{f'} && D\ar[dd]\ar[rr] && {\rm Coker}(f')\\ &&&&&& \\ && {\rm Coker}(g)\ar[rr]^{\tilde f'} && {\rm Coker}(g'), && } $$ and maps $A\to B\oplus C \to D$ given by $a\mapsto (f(a),-g(a))$ and $(b,c)\mapsto g'(b)+f'(c).$ \begin{lemma}\label{Lemma_pushout} The following statements are equivalent. \begin{enumerate} \item $\mathcal S$ is a pushout square. \item The sequence $A\to B\oplus C \to D \to 0$ is exact. \item $\tilde g'$ is an isomorphism and $\tilde g$ is an epimorphism. \item $\tilde f'$ is an isomorphism and $\tilde f$ is an epimorphism. \end{enumerate} \end{lemma} \begin{lemma}\label{Lemma_pullback} The following statements are equivalent. \begin{enumerate} \item $\mathcal S$ is a pullback square. \item The sequence $0\to A\to B\oplus C \to D$ is exact. \item $\tilde g$ is an isomorphism and $\tilde g'$ is a monomorphism. \item $\tilde f$ is an isomorphism and $\tilde f'$ is a monomorphism. \end{enumerate} \end{lemma} \subsection{Whitehead quadratic functor} For an abelian group $A$, the Whitehead group $\Gamma(A)$ is generated by symbols $\gamma(a),\ a\in A$ with the following relations \begin{align*} & \gamma(0)=0,\\ & \gamma(-a)=\gamma(a),\ a\in A\\ & \gamma(a+b+c)-\gamma(a+b)-\gamma(a+c)-\gamma(b+c)+\gamma(a)+\gamma(b)+\gamma(c)=0,\ a,b,c\in A. \end{align*} The correspondence $A\mapsto \Gamma(A)$ defines a quadratic functor in the category of abelian groups called the {\it Whitehead quadratic functor}. It has the following simple properties \begin{align*} & \Gamma(\mathbb Z/n)=\mathbb Z/(2n,n^2)\\ & A\otimes B={\rm Ker}\{\Gamma(A\oplus B)\to \Gamma(A)\oplus \Gamma(B)\}. \end{align*} There is a natural transformation of functors $\Gamma\to \otimes^2$ defined, for an abelian group $A$, as $\gamma(a)\mapsto a\otimes a,\ a\in A$. Define the functor $\Phi$ as a kernel of $\Gamma\to \otimes^2$. Then, for any abelian group $A$, there is a natural exact sequence $$ 0\to \Phi(A)\to \Gamma(A)\to A\otimes A\to \Lambda^2(A)\to 0 $$ where $\Lambda^2$ is the exterior square. One can easily check that, for any pair of abelian groups $A,B$, the (bi)natural map between the cross-effects of the functors $\Gamma$ and $\otimes^2$ $$ \Gamma(A|B)=A\otimes B \to \otimes^2(A|B)=A\otimes B\oplus B\otimes A $$ is a monomorphism. From this property together with the above description of the values of $\Gamma$ for cyclic groups follows that $\Phi$ is a 2-torsion functor, i.e. {\it for any $A$ and $a\in \Phi(A)$, $2a=0$ in $\Phi(A)$}. \subsection{Kernel of the Hurewicz homomorphism} \begin{proposition}\label{h2prop} For any connected space $X$, the kernel of the Hurewicz homomorphism $$ h_2\Omega: \pi_2(\Omega X)\to H_2(\Omega X) $$ is a 2-torsion subgroup of $\pi_2(\Omega X)=\pi_3(X).$ \end{proposition} Observe that, the statement about the third Hurewicz homomorphism obviously is not true without taking loops. For any odd prime $p$, the Moore space $P^3(p)$ has $\pi_3(P^3(p))=\mathbb Z/p$ and $H_3(P^3(p))=0$. \begin{proof} First consider the case of a simply-connected space $Y$. Let $GY$ be the simplicial Kan loop construction. The following diagram of fibrations $$ \xymatrix@M.7em{[GY,GY]\ar@{>->}[r] \ar@{->}[d] & GY \ar@{->>}[r] \ar@{->}[d] & (GY)_{ab}\ar@{->}[d]\\ \Delta^2(GY)\ar@{>->}[r] & \mathbb Z[GY]\ar@{->>}[r] & \mathbb Z[GY]/\Delta^2(GY)} $$ induces the commutative diagram of homotopy groups \begin{equation}\label{h4d} \xymatrix@M.7em{H_4(Y) \ar@{=}[d] \ar@{->}[r] & \pi_2([GY,GY])\ar@{->}[r] \ar@{->}[d] & \pi_2(\Omega Y)\ar@{->>}[r] \ar@{->}[d]^{h_2\Omega}& H_3(Y)\ar@{=}[d]\\ H_4(Y) \ar@{->}[r] & \pi_2(\Delta^2(GY))\ar@{->}[r] & H_2(\Omega Y) \ar@{->>}[r] & H_3(Y) } \end{equation} A simple analysis of connectivity of the simplicial groups $[GY, GY]$ and $\Delta^2(GY)$ shows that there are natural isomorphisms \begin{align*} & \pi_2([GY,GY])=\pi_2([GY,GY]/[[GY,GY],GY])=\pi_2(\Lambda^2((GY)_{ab})),\\ & \pi_2(\Delta^2(GY))=\pi_2(\Delta^2(GY)/\Delta^3(GY))=\pi_2((GY)_{ab}\otimes (GY)_{ab}). \end{align*} The derived functors of $\Lambda^2$ and $\otimes^2$ are well-known in a general situation (see, for example, \cite{BauesPira}). We obtain the following natural diagram $$ \xymatrix@M.7em{\pi_2([GY,GY])\ar@{->}[r]\ar@{=}[d] & \pi_2(\Delta^2(GY))\ar@{=}[d]\\ \Gamma(H_2(Y))\ar@{->}[r] & H_2(Y)\otimes H_2(Y)} $$ The left hand isomorphism in the last diagram is a reformulation of the result due to Whitehead \cite{W}, the right hand isomorphism follows from the Kunneth formula. Now lemmas \ref{Lemma_pushout} and \ref{Lemma_pullback} imply that, the diagram (\ref{h4d}) can be extended to the following diagram $$ \xymatrix@M.7em{ & \Phi(H_2(Y)) \ar@{>->}[d] \ar@{->>}[r] & K\ar@{>->}[d]\\ H_4(Y) \ar@{=}[d] \ar@{->}[r] & \Gamma(H_2(Y))\ar@{->}[d] \ar@{->}[r] & \pi_3(Y)\ar@{->>}[r]\ar@{->}[d]^{h_2\Omega} & H_3(Y)\ar@{=}[d] \\ H_4(Y) \ar@{->}[r] & H_2(Y)\otimes H_2(Y)\ar@{->>}[d] \ar@{->}[r] & H_2(\Omega Y) \ar@{->>}[d] \ar@{->>}[r] & H_3(Y)\\ & H_2(Y)\wedge H_2(Y)\ar@{=}[r] & H_2(\pi_1(\Omega Y))} $$ where the upper horizontal map is an epimorphism. Since the group $\Phi(H_2(Y))$ is 2-torsion, the kernel $K$ of the Hurewicz homomorphism also is 2-torsion and the needed statement is proved. Now consider the case of arbitrary connected space $X$. Consider its universal cover $\tilde X\to X$. The needed statement follows from the diagram $$ \xymatrix@M.7em{\pi_2(\Omega \tilde X)\ar@{=}[r] \ar@{->}[d]^{h_2\Omega} & \pi_2(\Omega X)\ar@{->}[d]^{h_2\Omega}\\ H_2(\Omega \tilde X)\ar@{>->}[r] & H_2(\Omega X)} $$ and the above proof of the statement for the simply-connected case. \end{proof} \section{Cubes of simplicial non-unital rings and their crossed cubes}\label{section3} \subsection{Cubes of fibrations and fibrant cubes} Set $\langle n\rangle=\{1,\dots, n\}.$ By a ring we assume a \underline{\bf non-unital} ring, and by a ring homomorphism we assume a non-unital ring homomorphism. Consider the category of simplicial rings (s.r.) as a model category, whose weak equivalences are weak equivalences of underlying simplicial sets and fibrations are level-wise surjective homomorphisms (see Ch.2 $\S4$ \cite{Quillen}). Then a fibration sequence in ${\sf sRng}$ is isomorphic to a sequence of the form \begin{equation}\label{eq_fibr} I\to R \twoheadrightarrow R/I, \end{equation} where $I$ is an ideal of the simplicial ring $R.$ Consider the ordered sets $\{0, 1\}$ and $\{-1, 0, 1\}$ as categories in a usual way. Let $\mathcal F:\{-1,0,1\}^n\to {\sf sRng}$ be a functor. For two disjoint subsets $\alpha,\beta\subseteq \langle n \rangle $ (i.e. $\alpha \cap \beta =\emptyset$) we put $$\mathcal F(\alpha,\beta)=\mathcal F(i_1,\dots,i_n),$$ where $\alpha=\{k\mid i_k=1\}$ and $\beta=\{k\mid i_k=-1\}.$ Then, if $\alpha'\supseteq \alpha$ and $\beta' \subseteq \beta,$ we have a map $$\mathcal F(\alpha,\beta)\longrightarrow \mathcal F(\alpha',\beta').$$ An {\bf $n$-cube of fibrations} of s.r. (see \cite[1.3]{Loday}) is a functor $\mathcal F:\{-1,0,1\}^n\to {\sf sRng}$ such that for any disjoint subsets $\alpha,\beta\subseteq \langle n \rangle$ and $k\in \langle n \rangle\setminus (\alpha\cup \beta)$ we have a fibration sequence $$\mathcal F(\alpha,\beta\cup\{k\})\to \mathcal F(\alpha,\beta)\twoheadrightarrow \mathcal F(\alpha\cup\{k\},\beta).$$ If $n=2$ it is a $3\times 3$ square whose rows and columns are fibration sequences: $$\xymatrix{ \mathcal F(\emptyset,\langle 2\rangle)\ar[r]\ar[d] & \mathcal F(\emptyset,\{2\})\ar[r]\ar[d] & \mathcal F(\{1\},\{2\})\ar[d]\\ \mathcal F(\emptyset,\{1\})\ar[r]\ar[d] & \mathcal F(\emptyset,\emptyset)\ar[r]\ar[d] & \mathcal F(\{1\},\emptyset)\ar[d] \\ \mathcal F(\{2\},\{1\})\ar[r] & \mathcal F(\{2\},\emptyset)\ar[r] & \mathcal F(\langle 2\rangle,\emptyset). }$$ An {\bf $n$-cube} of s.r. is a functor $\mathcal R:\{0,1\}^n\to {\sf sRng}.$ We set $$\mathcal R(\alpha)=\mathcal R(i_1,\dots,i_n),$$ where $\alpha=\{k\mid i_k=1\}.$ It is easy to see that $\{0,1\}^n$ is a direct category, and hence, it is a Reedy category. It follows that there is a natural model structure on the category of $n$-cubes of simplicial rings, called Reedy model structure \cite{Hovey}, \cite{Hirschhorn}. Weak equivalences of $n$-cubes in this model structure are defined level-wise. An $n$-cube $\mathcal R$ is {\bf fibrant} in this model structure if and only if the map $$\mathcal R(\alpha)\longrightarrow {\sf lim}_{\alpha'\supsetneq\: \alpha}\: \mathcal R(\alpha')$$ is a fibration of s.r. for any $\alpha\subseteq \langle n \rangle$. Let $\mathcal R$ be an $n$-cube of s.r. We consider the functor ${\sf E}_{\rm fib}(\mathcal R):\{-1,0,1\}^n\to {\sf sRng}$ given by \begin{equation}\label{eq_E_fib} {\sf E}_{\rm fib}(\mathcal R)(\alpha,\beta)={\rm Ker}\left( \mathcal R(\alpha)\to \prod_{i\in \beta} \mathcal R(\alpha\cup \{i\}) \right) \end{equation} with obvious morphisms. \begin{lemma}\label{Lemma_fibrant_cubes}Let $\mathcal R$ be an $n$-cube of s.r. Then the following statements are equivalent. \begin{enumerate} \item $\mathcal R$ is fibrant. \item $\mathcal R$ can be embedded into an $n$-cube of fibrations. \item ${\sf E}_{\rm fib}(\mathcal R)$ is an $n$-cube of fibrations. \end{enumerate} Moreover, if $\mathcal R$ is a fibrant $n$-cube of s.r., ${\sf E}_{\rm fib}(\mathcal R)$ is the unique (up to unique isomorphism that respects the embeddings) $n$-cube of fibrations to which $\mathcal R$ can be embedded. \end{lemma} \begin{proof} If $d\geq 0,$ $k\notin \alpha$ and $r\in \mathcal R(\alpha)_d,$ we denote by $r^k$ the image of $r$ in $\mathcal R(\alpha\cup \{k\})_d.$ Assume that $\alpha,\beta\subseteq \langle n \rangle$ are disjoint sets and $d\geq 0$. An {\bf $(\alpha,\beta)$-collection} is a collection $(r_i)\in \prod_{i\in \beta} \mathcal R(\alpha\cup \{i\})$ such that $r_i^j=r_j^i$ for any $i,j\in \beta.$ A {\bf lifting} of an $(\alpha,\beta)$-collection $(r_i)$ is an element $r\in \mathcal R(\alpha)_d$ such that $r_i=r^i.$ It is easy to see that the ring ${\sf lim}_{\alpha'\supsetneq\: \alpha}\: \mathcal R(\alpha')_d$ consist of $(\alpha,\alpha^{\rm c})$-collections, where $\alpha^{\sf c}=\langle n \rangle \setminus \alpha.$ Then $\mathcal R$ is fibrant if and only if for any $(\alpha,\alpha^{\sf c})$-collecion there exists a lifting. We claim that if $\mathcal R$ is fibrant then for any disjoint $\alpha,\beta$ and any $(\alpha,\beta)$-collection there exist a lifting. The prove is by induction on $|\langle n \rangle\setminus (\alpha\cup \beta)|.$ If it is equal to $0,$ we done. Assume that $(r_i)$ is an $(\alpha,\beta)$-collection. Consider any $j\in \langle n \rangle\setminus (\alpha\cup \beta)$ and the $(\alpha\cup \{j\},\beta)$-collection $(r^j_i).$ By induction hypothesis we have a lifting $r_j\in \mathcal R(\alpha\cup \{j\})$ of $(r_i^j).$ Then we get a $(\alpha,\alpha^{\sf c})$-collection $(r_i)_{i\in \alpha^{\sf c}},$ whose lifting is the lifting of the original $(\alpha,\beta)$-collection $(r_i).$ Therefore $\mathcal R$ is fibrant if and only if any $(\alpha,\beta)$-collection has a lifting. (1)$\Rightarrow$(3). Assume that $\mathcal R$ is fibrant and prove that ${\sf E}_{\rm fib}(\mathcal R)$ is an $n$-cube of fibrations. Consider the diagram with exact rows $$\xymatrix{0\ar[r] & {\sf E}_{\rm fib}(\mathcal R)(\alpha,\beta\cup\{k\})\ar[r]\ar[d] & \mathcal R(\alpha) \ar[r]\ar[d] & \prod_{i\in \beta\cup \{k\}}\mathcal R(\alpha\cup\{i\}) \ar[d]\\ 0 \ar[r] & {\sf E}_{\rm fib}(\mathcal R)(\alpha,\beta)\ar[r]\ar[d] & \mathcal R(\alpha) \ar[r]\ar[d] & \prod_{i\in \beta}\mathcal R(\alpha\cup\{i\}) \ar[d] \\ 0\ar[r] & {\sf E}_{\rm fib}(\mathcal R)(\alpha\cup\{k\},\beta)\ar[r] & \mathcal R(\alpha\cup\{k\}) \ar[r] & \prod_{i\in \beta}\mathcal R(\alpha\cup\{i,k\}). }$$ We have to prove that the left column is a short exact sequence. The only non-obvious thing is that ${\sf E}_{\rm fib}(\mathcal R)(\alpha,\beta)_d\to {\sf E}_{\rm fib}(\mathcal R)(\alpha\cup\{k\},\beta)_d $ is surjective. Consider any $r_k\in {\sf E}_{\rm fib}(\mathcal R)(\alpha\cup\{k\},\beta)_d$ and denote $r_i=0$ for $i\in \beta.$ Then $(r_i)$ is a $(\alpha,\beta\cup \{k\})$-collection, whose lifting is a preimage of $r_k.$ (3)$\Rightarrow$(1). Assume that ${\sf E}_{\rm fib}(\mathcal R)$ is an $n$-cube of fibrations and prove that $\mathcal R $ is fibrant. We need to prove that for any $(\alpha,\beta)$-collection there exists a lifting. Prove it by induction on $|\beta|.$ If $\beta=\emptyset,$ it is obvious. Assume that it holds for $\beta$ and prove it for $\beta'=\beta\cup\{k\},$ where $k\in \langle n\rangle\setminus (\alpha\cup \beta).$ Consider an $(\alpha,\beta\cup \{k\})$-collection $(r_i)_{i\in \beta\cup\{k\}}.$ By induction hypotheses its $(\alpha,\beta)$-subcollection $(r_i)_{i\in \beta}$ has a lifting $\bar r\in \mathcal R(\alpha)_d.$ Then $r_k-\bar r^k\in {\sf E}_{\rm fib}(\alpha\cup \{k\},\beta)_d.$ Since the map ${\sf E}_{\rm fib}(\mathcal R)(\alpha,\beta)_d\to {\sf E}_{\rm fib}(\mathcal R)(\alpha\cup\{k\},\beta)_d $ is surjective we get a preimage $\hat r\in {\sf E}_{\rm fib}(\mathcal R)(\alpha,\beta)_d$ such that $\hat r^k=r_k-\bar r^k.$ Then $r=\hat r+\bar r$ is a lifting of the $(\alpha,\beta\cup \{k\})$-collection $(r_i)_{i\in \beta\cup\{k\}}.$ (3)$\Rightarrow$(2). Obvious. (2)$\Rightarrow$(3). Assume that $\mathcal R$ is embedded into an $n$-cube of fibrations $\mathcal F.$ Replacing $\mathcal F$ by isomorphic one, we can assume that the fibres are identical embeddings. Prove that $\mathcal F(\alpha,\beta)={\sf E}_{\rm fib}(\alpha,\beta)$ by induction on $|\beta|.$ If $\beta=\emptyset,$ then $\mathcal F(\alpha,\emptyset)=\mathcal R(\alpha)={\sf E}_{\rm fib}(\alpha,\emptyset).$ Assume that it holds for $\beta$ and prove it for $\beta'=\alpha \cup \{k\}.$ By induction hypothesis we have a commutative diagram $$\xymatrix{&0\ar[d]&& \\ & \mathcal F(\alpha,\beta\cup\{k\})\ar[r]\ar[d] & \mathcal R(\alpha) \ar[r]\ar[d] & \prod_{i\in \beta\cup \{k\}}\mathcal R(\alpha\cup\{i\}) \ar[d]\\ 0 \ar[r] & {\sf E}_{\rm fib}(\mathcal R)(\alpha,\beta)\ar[r]\ar[d] & \mathcal R(\alpha) \ar[r]\ar[d] & \prod_{i\in \beta}\mathcal R(\alpha\cup\{i\}) \ar[d] \\ 0\ar[r] & {\sf E}_{\rm fib}(\mathcal R)(\alpha\cup\{k\},\beta)\ar[r] & \mathcal R(\alpha\cup\{k\}) \ar[r] & \prod_{i\in \beta}\mathcal R(\alpha\cup\{i,k\}), }$$ where the right column is a fibration sequence. Using that $\mathcal F(\alpha,\beta\cup\{k\})_d={\rm Ker}({\sf E}_{\rm fib}(\alpha,\beta)_d\to {\sf E}_{\rm fib}(\alpha\cup\{k\},\beta)_d)$ it is easy to deduce from the diagram that $\mathcal F(\alpha,\beta\cup\{k\})_d={\rm Ker}(\mathcal R(\alpha)_d\to \prod_{i\in \beta\cup \{k\}}\mathcal R(\alpha \cup \{i\})_d)={\sf E}_{\rm fib}(\mathcal R)(\alpha,\beta\cup \{k\})_d.$ \end{proof} \subsection{Cubes of fibrations and good tuples of ideals} An {\bf $n$-tuple of ideals} of a s.r. is a tuple $I=(R;I_1,\dots,I_n),$ where $R$ is a simplicial ring and $I_i$ are (simplicial) ideals of $R.$ For $\beta \subseteq \langle n \rangle$ we set $$I(\beta)=\bigcap_{i\in \beta} I_i.$$ An $n$-tuple of ideals $I$ is said to be {\bf good} if for any disjoint subsets $\alpha,\beta \subset \langle n \rangle$ and $k\in \langle n\rangle\setminus (\alpha \cup \beta)$ the following equality holds \begin{equation}\label{eq_good} I(\beta\cup \{k\})\cap \left(\sum_{i\in \alpha} I(\beta\cup \{i\})\right)=\sum_{i\in \alpha} I(\beta\cup \{k,i\}). \end{equation} It easy to check that any $2$-tuple of ideals is always good. But a $3$-tuple of ideals is good if and only if for any $i,j,k\in \{1,2,3\}$ the following holds $I_i\cap (I_j+I_k)=I_i\cap I_j+I_i\cap I_k.$ For an $n$-tuple of ideals $I=(R;I_1,\dots,I_n)$ we consider the functor ${\sf E}_{\rm idl}(I):\{-1,0,1\}^n\to {\sf sRng}$ given by \begin{equation}\label{eq_E_idl} {\sf E}_{\rm idl}(\mathcal I)(\alpha,\beta)=\frac{I(\beta)}{\sum_{i\in \alpha} I(\beta\cup \{i\})} \end{equation} with obvious morphisms. For example ${\sf E}_{\rm idl}(R;I,J)$ looks as follows: $$\xymatrix{ I\cap J\ar[r]\ar[d] & I\ar[r]\ar[d] & I/(I\cap J)\ar[d]\\ J\ar[r]\ar[d] & R\ar[r]\ar[d] & R/J\ar[d] \\ J/(I\cap J)\ar[r] & R/I\ar[r] & R/(I+J). }$$ Note that this definition can be rewritten in a way dual to \eqref{eq_E_fib}: ${\sf E}_{\rm idl}(I)(\alpha,\beta )={\rm Coker}\left(\coprod_{i\in \alpha} I(\beta \cup \{i\})\to I(\beta) \right).$ For an $n$-cube cube of fibrations $\mathcal F$ we consider an $n$-tuple of ideals \begin{equation}\label{eq_T_idl} {\sf T}_{\rm idl}(\mathcal F)=(R;I_1,\dots,I_n), \ \ R=\mathcal{F}(\emptyset,\emptyset),\ \ I_i={\rm Ker}(\mathcal F(\emptyset,\emptyset)\to \mathcal{F}(\{i\},\emptyset)). \end{equation} \begin{lemma}\label{Lemma_good_ideals}Let $I$ be an $n$-tuple of ideals of a s.r. Then the following statements are equivalent. \begin{enumerate} \item $I$ is good; \item $I={\sf T}_{\rm idl}(\mathcal F)$ for some $n$-cube of fibrations $\mathcal F$; \item ${\sf E}_{\rm idl}(I)$ is an $n$-cube of fibrations. \end{enumerate} Moreover, if $I$ is good, ${\sf E}_{\rm idl}(I)$ is the unique (up to unique isomorphism that respects the equalities) $n$-cube of fibrations such that $I={\sf T}_{\rm idl}({\sf E}_{\rm idl}(I))$. \end{lemma} \begin{proof} (1)$\Leftrightarrow$(3). Let $\alpha,\beta\subset \langle n \rangle$ be disjoint sets and $k\in \langle n \rangle \setminus (\alpha\cup \beta).$ Consider the sequence $$\frac{I(\beta\cup\{k\})}{\sum_{i\in \alpha}I(\beta\cup\{i,k\})}\longrightarrow \frac{I(\beta)}{\sum_{i\in \alpha}I(\beta\cup\{i\})}\longrightarrow \frac{I(\beta)}{\sum_{i\in \alpha\cup\{k\}}I(\beta\cup\{i\})}.$$ It is easy to see that the right hand map is en epimorphism and that it is exact in the middle term. Moreover, the left hand homomorphism is a monomorphism if and only if \eqref{eq_good} holds. (3)$\Rightarrow$(2). Follows from the equality ${\sf T}_{\rm idl}({\sf E}_{\rm idl}(I))=I.$ (2)$\Rightarrow$(3). Assume that $I={\sf T}_{\rm idl}(\mathcal F)$ for some $n$-cube of fibrations $\mathcal F.$ Replacing $\mathcal F$ by isomorphic one, we can assume that the fibres are identical embeddings. First we prove that $I(\beta)=\mathcal F(\emptyset,\beta).$ The prove is by induction on $|\beta|.$ If $|\beta|=0,1$ it is obvious. Assume that $|\beta|\geq 2$ and fix two distinct elements $k,l\in \beta.$ By induction hypothesis we have $\mathcal F(\emptyset,\beta\setminus \{k\})=I(\beta\setminus \{k\})$ and $\mathcal F(\emptyset,\beta\setminus \{l\})=I(\beta\setminus \{l\}).$ Consider the diagram $$\xymatrix{0\ar[r] & \mathcal F(\emptyset,\beta)\ar[r]\ar[d] & I(\beta\setminus \{k\})\ar[r]\ar[d] & \mathcal F(\{k\},\beta)\ar[d]\ar[r] & 0\\ 0\ar[r] &I(\beta\setminus \{l\})\ar[r] & I({\beta\setminus\{k,l\}})\ar[r] & \mathcal F(\{k\},{\beta\setminus\{k,l\}})\ar[r] & 0, \\ }$$ whose rows are short exact sequences. Since the left square consists of monomorphisms and the map on the cokernels $\mathcal F(\{k\},\beta)\to \mathcal F(\{k\},{\beta\setminus\{k,l\}})$ is a monomorphism, by Lemma \ref{Lemma_pullback} we get that the left square is a pulback square. Hence, $\mathcal F(\emptyset,\beta)=I(\beta\setminus \{k\}) \cap I(\beta\setminus \{l\})=I(\beta).$ So we have $\mathcal F(\emptyset,\beta)={\sf E}_{\rm idl}(\emptyset,\beta).$ Further we prove by induction on $ |\alpha|$ that there is a unique isomorphism $\mathcal F(\alpha,\beta)\cong {\sf E}_{\sf idl}(I)(\alpha,\beta)$ that satisfies commutation properties and lifts this equality for $\alpha=\emptyset$. Assume that this holds for $\alpha$ and prove it for $\alpha\cup\{k\}.$ By the induction hypothesis we have that $\mathcal{F}(\alpha,\beta\cup\{k\})=\frac{I(\beta\cup\{k\})}{\sum_{i\in \alpha}I(\beta\cup\{i,k\})}$ and $\mathcal{F}(\alpha,\beta)=\frac{I(\beta)}{\sum_{i\in \alpha}I(\beta\cup\{i\})}.$ It follows that there is a short exact sequence $$0\to \frac{I(\beta\cup\{k\})}{\sum_{i\in \alpha}I(\beta\cup\{i,k\})}\longrightarrow \frac{I(\beta)}{\sum_{i\in \alpha}I(\beta\cup\{i\})}\longrightarrow \mathcal F(\alpha\cup\{k\},\beta)\to 0,$$ which induces the required isomorphism $\mathcal F(\alpha,\beta)\cong \frac{I(\beta)}{\sum_{i\in \alpha\cup\{k\}}I(\beta\cup\{i\})}.$ \end{proof} \subsection{Three equivalent categories} Consider the truncation functor \begin{equation}\label{eq_T_fib} {\sf T}_{\rm fib}:\text{(}n\text{-cubes of fibrations of s.r.})\longrightarrow \text{(fibrant }n\text{-cubes of s.r.)} \end{equation} that induced by the embedding $\{0,1\}^n\subset \{-1,0,1\}^n.$ Lemma \ref{Lemma_fibrant_cubes} implies that this functor is well defined. \begin{proposition}The functors \begin{equation} \xymatrix{ \text{\rm (fibrant } n\text{\rm-cubes of s.r.)}\ar@<+10pt>[d]^{\sf E_{\rm fib}}_{\sim\hspace{1.5mm}} \\ \text{\rm(}n\text{\rm-cubes of fibrations of s.r.)}\ar@<+10pt>[d]^{{\sf T}_{\rm idl}}_{\sim\hspace{1.5mm}}\ar@<+10pt>[u]^{{\sf T}_{\rm fib}}\\ \text{\rm(good }n\text{\rm-tuples of ideals of s.r.)}\ar@<+10pt>[u]^{{\sf E}_{\rm idl}}\\ } \end{equation} given by \eqref{eq_E_fib}, \eqref{eq_E_idl},\eqref{eq_T_fib} and \eqref{eq_T_idl} define mutually invert equivalences of categories. \end{proposition} \begin{proof} The equalities ${\sf T}_{\rm fib}{\sf E}_{\rm fib}={\sf Id},$ ${\sf T}_{\rm idl}{\sf E}_{\rm idl}={\sf Id}$ are obvious. The isomorphisms ${\sf E}_{\rm fib}{\sf T}_{\rm idl}\cong {\sf Id}$ and ${\sf E}_{\rm fib}{\sf T}_{\rm idl}\cong {\sf Id}$ follow from Lemma \ref{Lemma_fibrant_cubes} and Lemma \ref{Lemma_good_ideals}. \end{proof} Consider the functor \begin{equation}\label{eq_fibre} {\sf Fibre}:\text{\rm (fibrant } n\text{\rm-cubes of s.r.)} \overset{\sim}\longrightarrow \text{\rm(good }n\text{\rm-tuples of ideals of s.r.)}, \end{equation} given by ${\sf Fibre}(\mathcal R)=(R;I_1,\dots, I_n),$ where $R=\mathcal R(\emptyset)$ and $I_i={\rm Ker}(\mathcal R(\emptyset)\to \mathcal R(\{i\})).$ \begin{corollary}The functor \eqref{eq_fibre} is an equivalence of categories. \end{corollary} \subsection{Crossed cubes of cubes of simplicial rings} Following Ellis \cite{Ellis} we define a {\bf crossed $n$-cube of rings} $\{R_\beta\}$ as a family of rings, where $\beta\subseteq \langle n \rangle$ together with homomorphisms $\mu_i:R_\beta \to R_{\beta\setminus \{i\}}$ and $h:R_\beta\otimes R_{\beta'} \to R_{\beta\cup \beta'}$ such that for $a,a'\in R_\beta,$ $b,b'\in R_{\beta'},$ $c\in R_{\beta''}$ and $i,j\in \langle n \rangle$ such that \begin{itemize} \item $\mu_i a=a$ if $i\notin \beta;$ \item $\mu_i\mu_ja=\mu_j\mu_ia;$ \item $\mu_ih(a\otimes b)=h(\mu_ia\otimes b)=h(a\otimes \mu_ib);$ \item $h(a\otimes b)=h(\mu_ia\otimes b)=h(a\otimes \mu_ib)$ if $i\in \beta\cap \beta';$ \item $h(a\otimes a')=aa';$ \end{itemize} with the assotiative property: \begin{itemize} \item $h(h(a\otimes b)\otimes c)=h(a\otimes h(b\otimes c)).$ \end{itemize} Morphisms of crossed $n$-cubes are defined obviously. Consider the functor \begin{equation}\label{eq_pi0} \pi_0: \text{(}n\text{-tuples of ideals of s.r.)} \xrightarrow{\ \ \ \ } \text{(crossed } n\text{-cubes of r.)} \end{equation} that sends $I$ to $\{R_\beta\},$ where $R_\beta=\pi_0I(\beta),$ $\mu_i=\pi_0(I(\beta)\hookrightarrow I(\beta\setminus \{i\}))$ and $h:R_{\beta}\otimes R_{\beta'} \to R_{\beta\cup \beta'}$ is the composition of the isomorphism $\pi_0I(\beta)\otimes \pi_0I({\beta'})\cong \pi_0(I(\beta)\otimes\mathcal I({\beta'}))$ and the map $\pi_0(I_{\beta} \otimes I_{\beta'}\to I_{\beta\cup {\beta'}}).$ It is easy to check that $\{R_\beta\}$ is a crossed $n$-cube of rings. In the Reedy model structure on the category of $n$-cubes there is a functorial fibrant replacement $$\gamma:\mathcal R\overset{\sim}\longrightarrow \overline{\mathcal R},$$ where $\overline{\mathcal R}$ is a fibrant $n$-cube and $\gamma$ is a weak equivalence and a cofibration. Then consider the functor \begin{equation}\label{pi}\Pi:\text{(}n\text{-cubes of s.r.)} \longrightarrow \text{(crossed }n\text{-cubes of rings)},\end{equation} given by $\Pi(\mathcal R)=\pi_0({\sf Fibre}(\overline{\mathcal R})),$ which is analogue of the one constructed in \cite{Brown_Loday} for simplicial rings. Analysing the definition of $\Pi$ we get the following. If $\mathcal R$ is an $n$-cube of s.r. we can embed it into a $n$-cube of {\bf homotopy} fibration sequences $\mathcal F$ in the {\bf homotopy category} of simplicial rings by taking homotopy fibres of all arrows, and then $$\Pi(\mathcal R)_\beta=\pi_0(\mathcal F(\emptyset,\beta)).$$ \section{Proof of theorem \ref{main}} \subsection{Proof of theorem \ref{main}} Now suppose that $R, S, T$ are normal subgroups of a group $G$. Let $X$ be a homotopy pushout of the following diagram of classifying spaces: \begin{equation}\label{pd}{\tiny { \xymatrix@M.7em{& & BG \ar@{->}[rr] \ar@{->}[ldd] \ar@{-}[dd] && B(G/R) \ar@{->}[ddd] \ar@{->}[ldd]\\ \\ & B(G/S) \ar@{->}[ddd] \ar@{->}[rr] & \ar@{->}[d] & B(G/RS) \ar@{->}[ddd]\\ & & B(G/T) \ar@{-}[r] \ar@{->}[ldd] & \ar@{->}[r] & B(G/RT) \ar@{->}[ldd]\\ & & \\ & B(G/ST) \ar@{->}[rr] & & X } }} \end{equation} There is a natural isomorphism of groups (see \cite{EM}): $$ \pi_1(X)\simeq G/RST. $$ Certain higher homotopy groups of $X$ are described in \cite{EM}. In particular, there is a natural isomorphism of $\pi_1(X)$-modules: \begin{equation}\label{ellis} \pi_3(X)\simeq \frac{R\cap S\cap T}{\|R,S,T\|} \end{equation} where the action of $\pi_1(X)\simeq G/RST$ on the right hand side of (\ref{ellis}) is viewed via conjugation in $G$. Recall the idea of the proof from \cite{EM}. Extend the above pushout to the cube of fibrations which have 27 spaces. The $\pi_1$ of the complement to the pushout in the cube of fibrations $$ {\tiny { \xymatrix@M.7em{& & \pi_1({\sf upper}\ {\sf corner}) \ar@{->}[rr] \ar@{->}[ldd] \ar@{-}[dd] && S\cap T \ar@{->}[ddd] \ar@{->}[ldd]\\ \\ & R\cap T \ar@{->}[ddd] \ar@{->}[rr] & \ar@{->}[d] & T \ar@{->}[ddd]\\ & & R\cap S \ar@{-}[r] \ar@{->}[ldd] & \ar@{->}[r] & S \ar@{->}[ldd]\\ & & \\ & R \ar@{->}[rr] & & G } }} $$ is a crossed cube of groups. Moreover, it is a {\it universal} crossed cube of groups (see \cite{EM} for definition and discussion of the universality). One can realize the pushout diagram as a diagram of simplicial groups. The functor of group rings ${\mathbb Z}[-]: {\sf groups}\to {\sf group}\ {\sf rings}$ sends pushouts to the pushouts in the category of simplicial rings. Extending this pushout diagram to a cube of homotopy fibration sequences in the category of simplicial rings and taking $\pi_0$ of the complement part as in (\ref{pi}) we obtain the crossed cube of rings $${\small { \xymatrix@M.7em{& & \pi_0({\sf upper}\ {\sf corner}) \ar@{->}[rr]^{\mu_1} \ar@{->}[ldd]^{\mu_3} \ar@{-}[dd]^{\mu_2} && {\bf s}\cap {\bf t} \ar@{->}[ddd] \ar@{->}[ldd]\\ \\ & {\bf r}\cap {\bf s} \ar@{->}[ddd] \ar@{->}[rr] & \ar@{->}[d] & {\bf s} \ar@{->}[ddd]\\ & & {\bf r}\cap {\bf t} \ar@{-}[r] \ar@{->}[ldd] & \ar@{->}[r] & {\bf t} \ar@{->}[ldd]\\ & & \\ & {\bf r} \ar@{->}[rr] & & \mathbb Z[G] } }} $$ Now we observe that $$ \|{\bf r},{\bf s},{\bf t}\|\subseteq {\rm Im}(\mu_i),\ i=1,2,3. $$ This follows from the properties of crossed cubes $$\mu_ih(a\otimes b)=h(\mu_ia\otimes b)=h(a\otimes \mu_ib),\ i=1,2,3$$ for $a\in {\bf r\cap \bf s}, b\in {\bf t}$ and other choices of ideals. Applying homotopy exact sequences of fibrations three times, and comparing them for simplicial groups and group rings, we obtain the needed commutative diagram $$ \xymatrix@M.7em{\frac{R\cap S\cap T}{\|R,S,T\|}\ar@{->}[r]\ar@{=}[d] & \frac{\bf r\cap \bf s\cap \bf r}{\|\bf r,\bf s,\bf t\|}\ar@{->}[d]\\ \pi_2(\Omega X)\ar@{->}[r]^{h_2\Omega}& H_2(\Omega X)} $$ which considered together with proposition \ref{h2prop} imply the needed statement. Theorem \ref{main} follows.\ \ $\Box$ \subsection{The case of two subgroups} Observe that, in the case of two normal subgroups $R,S$ in $G$ is much simpler than the above case. In this case, one has a square of fibrations (in the category of simplicial rings) $$ \xymatrix@M.7em{T\ar@{->}[r] \ar@{->}[d] \ar@{->}[r] & {\bf r} \ar@{->}[d] \ar@{->}[r] & fib_2\ar@{->}[d]\\ {\bf s}\ar@{->}[r]\ar@{->}[d] & \mathbb Z[G] \ar@{->}[r] \ar@{->}[d] & \mathbb Z[G/S]\ar@{->}[d] \\ fib_1\ar@{->}[r] & \mathbb Z[G/R] \ar@{->}[r] & \mathbb Z[X]} $$ such that $$ \xymatrix@M.7em{\pi_0(T) \ar@{->}[r]^{\mu_2} \ar@{->}[d]^{\mu_1} & {\bf r}\ar@{->}[d]\\ {\bf s}\ar@{->}[r] & \mathbb Z[G]} $$ is a crossed square of rings. Since $\mu_1h(a\otimes b)=ab,\ a\in {\bf s}, b\in {\bf r},$ and $\mu_1h(a\otimes b)=ab,\ a\in {\bf s},b\in {\bf r}$, ${\bf rs+sr}\subseteq {\rm Im}(\mu_1)$ and ${\bf rs+sr}\subseteq {\rm Im}(\mu_2)$. Comparing the picture for groups and group rings we conclude that there is a commutative diagram $$ \xymatrix@M.7em{\frac{R\cap S}{[R,S]}\ar@{=}[d]\ar@{>->}[r] & \frac{{\bf r}\cap {\bf s}}{\bf rs+sr}\ar@{->}[d]\\ \pi_1(\Omega X)\ar@{->}[r]^{h_1\Omega} & H_1(\Omega X)} $$ Since, for any connected space $X$, the Hurewicz homomorphism $\pi_1(\Omega X)\to H_1(\Omega X)$ is a monomorphism, we obtain the following identification of the generalized dimension subgroup $$ D(G, {\bf rs+sr})=[R,S]. $$ This gives a new proof of the result from \cite{BD}. This result can be generalized as follows. Let $T$ be a normal subgroup of subgroup of $G$ and ${\bf n}=(T-1)\mathbb Z[G]$. We obtain the following diagram $$ \xymatrix@M.7em{\frac{R\cap S}{[R,S][R\cap S,T]}\ar@{=}[d]\ar@{->}[r] & \frac{{\bf r}\cap {\bf s}}{\bf rs+sr+({\bf r}\cap {\bf s}){\bf t}+{\bf t}({\bf r}\cap {\bf s})}\ar@{->}[d]\\ \pi_1(\Omega X)_{TRS/RS}\ar@{->}[r]^{(h_1\Omega)_{TRS/RS}} & H_1(\Omega X)_{TRS/RS}.} $$ Here the group $TRS/RS$ is considered as a subgroup of $\pi_1(X)=G/RS.$ We show that $(h_1\Omega)_{TRS/RS}\colon \pi_1(\Omega X)_{TRS/RS}\longrightarrow H_1(\Omega X)_{TRS/RS}$ is a monomorphism for any normal subgroup $T$ of $G$. Let $G_*$ be a simplicial group such that the geometric realization $|G_*|$ is weakly homotopy equivalent to $\Omega X$. Observe that the action of $\pi_1(X)\cong\pi_0(G_*)$ is induced by the conjugation action of $G_0$ on $G_*$. Let $\tilde G_*$ be the path-connected component of $G_*$ containing the identity element. From the simplicial Postnikov system, there is a short exact sequence of simplicial groups $$ 1\longrightarrow \tilde G_*\longrightarrow G_*\longrightarrow \pi_0(G)\longrightarrow 1, $$ where $\pi_0(G)$ is the discrete simplicial group. Hence $$ G_*=\coprod_{g\in \pi_0(G)} g\tilde G_* $$ as a simplicial set. Let $\chi_h\colon G_*\to G_*, x\mapsto hxh^{-1}$ be the conjugation action of $h\in G_0$ on $G_*$. Then $\chi_h(g\tilde G_*)=hgh^{-1}\tilde G_*$ with $$ \chi_h(gx)=(hgh^{-1})(hxh^{-1}). $$ This implies that $$ H_k(\Omega X)\cong H_*(G_*)\cong H_k(\tilde G_*)\otimes \mathbb{Z}[\pi_0(G_*)] $$ as modules over $\mathbb{Z}[\pi_0(G_*)]$ for $k\geq0$, where $\mathbb{Z}[\pi_0(G_*)]$ acts diagonally on the tensor product $H_k(\tilde G_*)\otimes \mathbb{Z}[\pi_0(G_*)]$. It follows that $H_k(\Omega_0X)\cong H_k(\tilde G_*)$ is a $\mathbb{Z}[\pi_0(G_*)]\cong \mathbb{Z}[\pi_1(X)]$-equivariant summand of $H_k(\Omega X)$, where $\Omega_0X$ is the path-connected component of $X$ containing the basepoint. By taking $k=1$ with using the fact that $\pi_1(\Omega X)\cong H_1(\Omega_0X)$, we have $(h_1\Omega)_{TRS/RS}\colon \pi_1(\Omega X)_{TRS/RS}\longrightarrow H_1(\Omega X)_{TRS/RS}$ is a monomorphism for any normal subgroup $T$ of $G$. As a consequence, we obtain that \begin{equation} \frac{R\cap S}{[R,S][R\cap S,T]}\longrightarrow \frac{{\bf r}\cap {\bf s}}{\bf rs+sr+({\bf r}\cap {\bf s}){\bf t}+{\bf t}({\bf r}\cap {\bf s})} \end{equation} is a monomorphism for any normal subgroup $T\leq G$. We proved the following \begin{theorem}\label{modg} For a group $G$ and its normal subgroups $R,S,T$, $$ D(G,{\bf rs+sr}+({\bf r}\cap {\bf s}){\bf t}+{\bf t}({\bf r}\cap {\bf s}))=[R,S][R\cap S,T]. $$ \end{theorem} \section{Generalizations and examples}\label{lastsection} \subsection{Simplicial groups} Let $G$ be a group and $R_1,\dots, R_n,\ n\geq 2$ its normal subgroups. Denote $$ \|R_1,\dots, R_n\|:=\prod_{I\cup J=\{1,\dots,n\}, I\cap J=\emptyset}[\cap_{i\in I}R_i,\cap_{j\in J}R_j] $$ Similarly, for a collection $\mathbf{a}_0,\dots, \mathbf{a}_n$ of ideals in a ring $R$, denote $$ \| \mathbf{a}_1,\dots, \mathbf{a}_n\|:=\sum_{I\cup J=\{1,\dots,n\}, I\cap J=\emptyset}(\cap_{i\in I}\mathbf{a}_i)(\cap_{j\in J}\mathbf{a}_j)+(\cap_{j\in J}\mathbf{a}_j)(\cap_{i\in I}\mathbf{a}_i). $$ It is easy to check that for arbitrary ideals ${\bf a}, {\bf b}$ of $\mathbb Z[G]$ we have $D({\bf a})D({\bf b}) \subseteq D({\bf a}+{\bf b})$ and $[D({\bf a}),D({\bf b})]\subseteq D({\bf a}{\bf b}+{\bf b}{\bf a}).$ Indeed, the first inclusion is obvious, and the second follows from the equality $g^{-1}h^{-1}gh-1=g^{-1}h^{-1}((g-1)(h-1)-(h-1)(g-1))$ for arbitrary $g,h\in G.$ Therefore, for any $G$ and its normal subgroups $R_1,\dots, R_n$, $$ \|R_1,\dots, R_n\|\subseteq D(G,\| \mathbf{r}_1,\dots, \mathbf{r}_n\|), $$ where $\mathbf{r}_i=(R_i-1)\mathbb Z[G]$. One can ask how to identify the generalized dimension quotient in the case of $>3$ subgroups or at least to describe its exponent. Here is an approach for constructing of different examples. If $\mathcal G$ is a simplicial group, we denote by $N\mathcal G$ its Moore complex. Cycles of this complex we denote by $Z_n\mathcal G={\rm Ker}(d:N_n\mathcal G\to N_{n-1}\mathcal G),$ and boundaries by $B_n\mathcal G={\rm Im}(d:N_{n+1}\mathcal G\to N_n\mathcal G).$ Then $\pi_n\mathcal G=Z_n\mathcal G/B_n\mathcal G.$ The following theorem is Theorem B of \cite{Castiglioni_Ladra}. \begin{theorem}[{\cite{Castiglioni_Ladra}}]\label{Theorem_B} Let $\mathcal G$ be a simplicial group and $n\geq 1$ such that $\mathcal G_{n+1}$ is generated by degeneracies. Set $K_i:={\rm Ker}(d_i:\mathcal G_n\to \mathcal G_{n-1}).$ Then $$B_n\mathcal G=\|K_0,\dots, K_n\|,\hspace{1cm} \pi_n\mathcal G=\frac{\bigcap_{i=0}^n K_i}{\|K_0,\dots, K_n\|}.$$ \end{theorem} The following theorem is an analogue of the previous one for the case of simplicial rings. It follows from Theorem A of \cite{Castiglioni_Ladra}. \begin{theorem}[{\cite{Castiglioni_Ladra}}]\label{Theorem_A} Let $\mathcal R$ be a simplicial ring and $n\geq 1$ such that $\mathcal R_{n+1}$ is generated by degeneracies as a ring. Set ${\mathbf{k}}_i:={\rm Ker}(d_i:\mathcal R_n\to \mathcal R_{n-1}).$ Then $$B_n\mathcal R=\|\mathbf{k}_0,\dots,\mathbf{k}_n\|, \hspace{1cm} \pi_n\mathcal R=\frac{\bigcap_{i=0}^n {\bf k}_i}{\|\mathbf{k}_0,\dots,\mathbf{k}_n\|}.$$ \end{theorem} For a simplicial group $\mathcal G$, the Hurewicz homomorphism $h:\pi_n\mathcal G\to H_n\mathcal G$ for $n\geq 1$ is induced by the map $\mathcal G\to \mathbb Z[\mathcal G]$ given by $g\mapsto g-1.$ The following statement is a direct corollary of theorems \ref{Theorem_B} and \ref{Theorem_A}. \begin{proposition}\label{simple} Let $\mathcal G$ be a simplicial group and $n\geq 1$ such that $\mathcal G_{n+1}$ is generated by degeneracies. Set \begin{align*} & K_i:={\rm Ker}(d_i:\mathcal G_n\to \mathcal G_{n-1})\\ & {\bf k}_i:={\rm Ker}(d_i: \mathbb Z[\mathcal G_n] \to \mathbb Z[\mathcal G_{n-1}]). \end{align*} Then $$\frac{D(\mathcal G_n,\|\mathbf{k}_0,\dots,\mathbf{k}_n\|)}{\|K_0,\dots, K_n\|} ={\rm Ker}(h_n: \pi_n\mathcal G \longrightarrow H_n\mathcal G),$$ where $h_n$ is the $n$th Hurewicz homomorphism. \end{proposition} The generalized dimension subgroups as in proposition \ref{simple} were considered in \cite{MPW} for the case of simplicial Carlsson's constructions. The main example which we will consider here is the $p$-Moore space $P^3(p)=S^2\cup_{p}e^2$ for a prime $p\geq 2$. The lowest homotopy group of $P^3(p)$ which contains $\mathbb Z/p^2$-summand is $\pi_{2p-1}P^3(p)$. This was proved in \cite{CMN} for $p>3$, however, $\pi_3P^3(2)=\mathbb Z/4$, $\pi_5P^3(3)=\mathbb Z/9$. Since all homology groups $H_*(\Omega P^3(p))$ have exponent $p$, we have $$ \mathbb Z/p\subseteq {\rm Ker}\{h_{2p-2}:\pi_{2p-2}(\Omega P^3(p))\to H_{2p-2}(\Omega P^3(p))\}. $$ Taking $\mathcal G$ to be a simplicial model for $\Omega P^3(p)$ with $\mathcal G_3$ generated by degeneracies, we obtain the following example. Set $G=\mathcal G_{2p-2},\ K_i={\rm Ker} d_i: \mathcal G_{2p-2}\to \mathcal G_{2p-3},$ then, by proposition \ref{simple}, the generalized dimension quotient $$ \frac{D(G,\|{\bf k}_0,\dots, {\bf k}_{2p-2}\|)}{\|K_0,\dots, K_{2p-2}\|} $$ contains a subgroup $\mathbb Z/p$. In the case $p=2$, we can choose $$\mathcal G_1=F(\sigma),\ \mathcal G_2=F(a,b,c)$$ with the face maps $$ d_0:\begin{cases} a\mapsto \sigma^2\\ b\mapsto \sigma\\ c\mapsto 1\end{cases},\ d_1:\begin{cases} a\mapsto 1\\ b\mapsto \sigma\\ c\mapsto \sigma\end{cases},\ d_2: \begin{cases} a\mapsto 1\\ b\mapsto 1\\ c\mapsto \sigma\end{cases} $$ Taking the kernels $R={\rm Ker}(d_0),\ S={\rm Ker}(d_1),\ T={\rm Ker}(d_2)$, we obtain an example discussed in introduction, with $$ \frac{D(G, \|{\bf r},{\bf s},{\bf t}\|)}{\|R,S,T\|}\simeq \mathbb Z/2. $$ \noindent{\bf Conjecture.} For a prime $p$, any group $G$ and its normal subgroups $R_1,\dots, R_n,\ n\geq 2$, the quotient $$ \frac{D(G,\|{\bf r}_1,\dots, {\bf r}_n\|)}{\|R_1,\dots, R_n\|} $$ is $p$-torsion free provided $n<2p-1$. \subsection{Connectivity conditions} Let G be a group with normal subgroups $R_1,\dots, R_n,$ $n\geq 2$. Recall the connectivity condition from \cite{EM}. The $n$-tuple of normal subgroups $(R_1,\dots, R_n)$ is called {\it connected} if either $n\geq 2$ or $n\geq 3$ and for all subsets $I,J\subseteq \{1,\dots, n\},$ with $|I|\geq 2,$ $|J|\geq 1$, the following holds $$ \left(\bigcap_{i\in I} R_i\right)\left(\prod_{j\in J}R_j\right)=\bigcap_{i\in I}\left(R_i(\prod_{j\in J}R_j)\right) $$ Now consider the homotopy colimit $X$ of classifying spaces $B(G/\prod_{i\in I}R_i)$, where $I$ ranges over all proper subsets $I\subset \{1,\dots, n\}$. Then, if for any $i=1,\dots, n$, the $n-1$-tuple of normal subgroups $(R_1,\dots, \hat R_i,\dots, R_n)$ is connected, then $$ \pi_n(X)=\frac{R_1\cap \dots \cap R_n}{\|R_1,\dots, R_n\|}. $$ This is proved in \cite{EM}. The proof of theorem 1 from \cite{EM} together with results of section \ref{section3}, namely, together with the construction of the functor $\Pi$, imply the following \begin{proposition}\label{p12} If for any $i=1,\dots, n$, the $n-1$-tuple of normal subgroups $(R_1,\dots, \hat R_i,\dots, R_n)$ is connected, then there is a commutative diagram $$ \xymatrix@M.7em{\frac{R_1\cap \dots \cap R_n}{\|R_1,\dots, R_n\|}\ar@{->}[r]\ar@{=}[d] & \frac{{\bf r}_1\cap \dots \cap {\bf r}_n}{\|{\bf r}_1,\dots, {\bf r}_n\|}\ar@{->}[d]\\ \pi_{n-1}(\Omega X)\ar@{->}[r]^{h_{n-1}\Omega} & H_{n-1}(\Omega X)}. $$ \end{proposition} One can use proposition \ref{p12} for proving the above conjecture about the $p$-torsion in the generalized dimension quotient in some particular cases.

https://arxiv.org/abs/1602.06376

Movement of time-delayed hot spots in Euclidean space

We investigate the shape of the solution of the Cauchy problem for the damped wave equation. In particular, we study the existence, location and number of spatial maximizers of the solution. Studying the shape of the solution of the damped wave equation, we prepare a decomposed form of the solution into the heat part and the wave part. Moreover, as its another application, we give $L^p$-$L^q$ estimates of the solution.

\section{Introduction} Let $f$ and $g$ be real-valued smooth functions defined on $\mathbb{R}^n$. We consider the {\it damped wave equation} with initial data $(f,g)$, \begin{equation}\label{DWfg} \begin{cases} \displaystyle \( \frac{\partial^2}{\partial t^2} -\Delta + \frac{\partial}{\partial t} \) u(x,t) =0, &x \in \mathbb{R}^n ,\ t>0,\\ \displaystyle \( u, \frac{\partial u}{\partial t} \) (x,0)=(f, g)(x), &x \in \mathbb{R}^n . \end{cases} \end{equation} In the one-dimensional case, the damped wave equation is known as the {\it telegrapher's equation} introduced by Oliver Heaviside and describes the current and voltage in an electrical circuit with resistance and inductance. More generally, the equation \eqref{DWfg} is a model of the propagation of the wave with friction or resistance. The damped wave equation is also known as the {\it hyperbolic heat conduction equation} introduced in [Cat, Ch, L, MF, V]. The classical model of the heat equation admits the infinite speed of the propagation of heat conduction, which is physically inadmissible. Therefore, the equation \eqref{DWfg} was introduced to modify the model of heat conduction with finite speed of the propagation. In order to derive the equation \eqref{DWfg} as the hyperbolic heat conduction equation, along Li's framework in [L], let us consider the one-dimensional case as below: Let $v(x,t)$ be the temperature at a point $x \in \mathbb{R}$ and at time $t$; Let $q(x,t)$ be the heat flux at a point $x \in \mathbb{R}$ and at time $t$; Then, the heat balance law implies \begin{equation}\label{heat_balance} \frac{\partial v}{\partial t}(x,t) + \frac{\partial q}{\partial x} (x,t) =0,\ x \in \mathbb{R} ,\ t>0; \end{equation} From the time-delayed Fourier's law with a small enough positive parameter $\tau$ \begin{equation}\label{Fourier_delay} q(x,t+\tau ) \approx q(x,t) + \tau \frac{\partial q}{\partial t}(x,t) = -\frac{\partial v}{\partial x} (x,t) ,\ x \in \mathbb{R} ,\ t>0, \end{equation} instead of the usual Fourier's law \begin{equation}\label{Fourier} q(x,t)=-\frac{\partial v}{\partial x}(x,t) ,\ x \in \mathbb{R} ,\ t>0, \end{equation} we get the damped wave equation \begin{equation}\label{dwv} \( \tau \frac{\partial^2}{\partial t^2} -\frac{\partial^2}{\partial x^2} + \frac{\partial}{\partial t} \) v(x,t) =0 ,\ x\in \mathbb{R} ,\ t>0, \end{equation} with delay time $\tau$; The additional term $\tau \partial^2 v / \partial t^2$ brings the finite propagation speed property to the equation (\ref{dwv}); By the scale transformation \begin{equation}\label{scale_trans} u(x,t) = \frac{1}{\tau} v\( \sqrt{\tau}x ,\tau t \) , \end{equation} we obtain \begin{equation}\label{dwu} \( \frac{\partial^2}{\partial t^2} -\frac{\partial^2}{\partial x^2} + \frac{\partial}{\partial t} \) u(x,t) =0,\ x \in \mathbb{R} ,\ t>0. \end{equation} From such a background, as $t$ goes to infinity in (\ref{dwu}) (corresponding to the case where $\tau$ tends to $0^+$ in (\ref{dwv})), it is expected that the solution of the damped wave equation approaches to that of the (usual) heat equation, which is called the {\it diffusion phenomenon} and has been studied by many researchers ([FG, HO, MN, Nar, Nis, YM]). In this paper, we investigate the relation between the damped wave and heat equations in view of the study on the shapes of the solutions. Precisely, we give correspondence to Chavel and Karp's results in [CK]. Let us review their works as below: Let $P_n(t) \phi (x)$ be the unique bounded solution of the Cauchy problem for the heat equation with bounded initial datum $\phi$, that is, \begin{equation}\label{P} P_n(t) \phi (x)= \frac{1}{\( 4\pi t\)^{n/2}} \int_{\mathbb{R}^n} \exp \( -\frac{r^2}{4t} \) \phi (y) dy,\ x \in \mathbb{R}^n ,\ t>0,\ r=\lvert x-y \rvert ; \end{equation} When $\phi$ is a non-zero non-negative bounded function with compact support, they studied the behavior of the set of {\it hot spots} \begin{equation}\label{H} H_\phi (t) = \left\{ x \in \mathbb{R}^n \lvert P_n(t)\phi (x) = \max_{\xi \in \mathbb{R}^n} P_n (t) \phi (\xi ) \right\} \right. ; \end{equation} They showed that hot spots exist at each time $t$, that all of them are contained in the convex hull of the support of $\phi$ for any time $t$, and that the set $H_\phi (t)$ converges to the one-point set of the centroid (the center of mass) of $\phi$ as $t$ goes to infinity; Furthermore, calculating the Hessian of $P_n(t)\phi$, in [JS], Jimbo and Sakaguchi indicated that the set of hot spots $H_\phi (t)$ consists of one point after a large time $t$. There are many results on the study of hot spots besides [CK]. As examples, we introduce [JS, I1, I2, IK, FI1, FI2, S] as below: In [JS], Jimbo and Sakaguchi studied the large time behavior of hot spots in unbounded domains. For example, they considered the exterior domain of a ball with a radially symmetric initial datum. In this case, the explicit representation of solutions like \eqref{P} was still useful; In [I1, I2], Ishige studied the large time behavior of hot spots in the exterior domain of a ball without the radially symmetric assumption of [JS] by using the self-similar transformation and the eigenfunction expansion: In [IK], the similar approach to [I1, I2] was also applicable for the heat equation with a potential; In [FI1, FI2], Fujishima and Ishige studied the large time behavior of hot spots in $\mathbb{R}^n$ without the non-negativity of the initial datum $\phi$ and applied their investigation to the blow-up set of a semi-linear heat equation; In [S], the first author generalized the results in [CK, JS] in terms of a potential with a radially symmetric kernel. In view of the diffusion phenomena, it is expected that spatial maximizers of the solution of (\ref{DWfg}) have similar properties to hot spots shown in [CK, JS, S]. To this aim, when $f$ and $g$ are compactly supported, and when $h:= f+g$ is non-zero and non-negative, we study the behavior of the set of {\it time-delayed hot spots} \begin{equation}\label{dH} \mathcal{H}(t) = \left\{ x \in \mathbb{R}^n \lvert u(x,t) = \max_{\xi \in \mathbb{R}^n} u(\xi, t) \right\} \right. . \end{equation} Precisely, we show the following properties: \begin{enumerate} \item[(1)] After a large enough time, the set $\mathcal{H}(t)$ is contained in the convex hull of the support of $h$. Furthermore, for some small time, we give some examples of $(f,g)$ such that the set $\mathcal{H}(t)$ escapes from the convex hull of the support of $h$ (Theorem \ref{movement} (1), Examples \ref{ex1}, \ref{ex2}, \ref{ex3} and \ref{ex4}). \item[(2)] The set $\mathcal{H}(t)$ converges to the one-point set of the centroid of $h$ as $t$ goes to infinity (Theorem \ref{movement} (2)). \item[(3)] After a large enough time, the set $\mathcal{H}(t)$ consists of one point (Proposition \ref{uniqueness}). \end{enumerate} In order to understand the meaning of the above statements, we, for example, consider the case where $f$ is non-zero and non-negative, the maximum value of $f$ is greater than that of $h$, and the supports of $f$ and $h$ are separated. Since $u(x ,0) =f(x)$, for any sufficiently small $t>0$, all of the time-delayed hot spots are ``close'' to maximum points of $f$, that is, they are contained in the support of $f$ and not contained in the support of $h$. Roughly speaking, our main results claim that time-delayed hot spots move from the set of maximum points of $f$ to the centroid of $h$. The above properties of time-delayed hot spots are due to the decomposition of the solution operator (fundamental solution) $S_n(t)$ into the heat part and the wave part as \begin{equation}\label{decomp} S_n(t)g(x) = J_n(t) g(x) + e^{-t/2}\vect{W}_n(t)g(x). \end{equation} Here, $J_n(t)g(x)$ and $\vect{W}_n(t) g(x)$ are suitable functions behaving like as $P_n(t)g(x)$ and the solution of the free wave equation with initial datum $(0,g)$, respectively. The decomposition \eqref{decomp} was firstly discovered by Nishihara in [Nis] in the three-dimensional case and so-called the {\it Nishihara decomposition}. In this paper, we give the generalization of the Nishihara decomposition in higher dimensional cases to study the behavior of time-delayed hot spots. Moreover, as its another application, we slightly improve Narazaki's $L^p$-$L^q$ estimates given in [Nar]. As a by-product of the $L^p$-$L^q$ estimates, we obtain \begin{equation}\label{LpLq} \| u(\cdot ,t) - P_n(t)h \|_{L^{\infty}} \le C t^{-n/2-1} \( \| f \|_{L^1} + \|g\|_{L^1} + \| f \|_{W^{*,\infty}} + \|g\|_{W^{*,\infty}} \) . \end{equation} One may consider that the large time behavior of time-delayed hot spots can be easily investigated by combining the results in [CK] and the estimate \eqref{LpLq}. But the expectation is incorrect. This is because the difference between the values of $P_n(t)h$ in the convex hull of support of $h$ and its outside can have the order worse than $t^{-n/2-1}$. Hence the above estimate cannot exclude the possibility that time delayed hot spots escape from the convex hull of support of $h$ by the effect of the wave part. Therefore, we should obtain more precise information about the solution of the damped wave equation.\\ This paper is organized as follows. In Section 2, we give a representation of the solution of \eqref{DWfg} and its Nishihara decomposition. In Section 3, we give preliminary estimates for our investigation. In Section 4, we investigate the movement of time-delayed hot spots. In Section 5, we give $L^p$-$L^q$ estimates of the difference between the damped wave and the heat by using the Nishihara decomposition described in Section 2.\\ \noindent{\bf Notation.} For the end of this section, we explain our notation. \begin{itemize} \item The letter $C$ indicates the generic constant which may change from line to line. \item We denote the usual $L^p$ norm by $\|\cdot\|_{L^p}$, that is, \begin{equation} \| \phi \|_{L^p}= \begin{cases} \displaystyle \( \int_{\mathbb{R}^n} \lvert \phi (x) \rvert^p dx\)^{1/p} &( 1\leq p <\infty ),\\ \displaystyle \esssup_{x \in \mathbb{R}^n} \lvert \phi (x) \rvert &(p=\infty) . \end{cases} \end{equation} \item For a natural number $\ell$, we denote the Sobolev norm by $\| \cdot \|_{W^{\ell ,\infty}}$, that is, \begin{equation} \| \phi \|_{W^{\ell ,\infty}} =\sum_{|\alpha |\leq \ell} \| \partial^\alpha \phi \|_{L^\infty} . \end{equation} \item Let $B^n$ and $S^{n-1}$ be the $n$-dimensional unit closed ball and the $(n-1)$-dimensional unit sphere, respectively. \item For real numbers $a$ and $b$, and for two sets $X$ and $Y$ in $\mathbb{R}^n$, we use the notation (Minkowski sum) $aX+bY = \left\{ ax+by \lvert x \in X,\ y\in Y \right\} \right.$. In particular, we write $B_t^n(x) =tB^n +\{ x \}$ and $S_t^{n-1}(x) = tS^{n-1}+ \{ x \}$, that is, the $n$-dimensional closed ball with radius $t$ centered at $x$ and the $(n-1)$-dimensional sphere with radius $t$ centered at $x$, respectively. \item Let us denote by $CS(\phi )$ the convex hull of the support of a function $\phi$. \item Let $\sigma_n$ denote the $n$-dimensional Lebesgue surface measure. \item We understand that the letter $r$ is always used for $r=\lvert x-y \rvert$. \end{itemize} \noindent{\bf Acknowledgements.} The authors would like to express their deep gratitude to Professor Tatsuo Nishitani and Professor Jun O'Hara for giving them valuable comments. \section{Decomposition of the solution} \subsection{Solution formula of the damped wave equation} In this subsection, we prepare the explicit form of the solution of the Cauchy problem \eqref{DWfg}. For a smooth function $g$, let us denote by $S_n(t)g$ the solution of the Cauchy problem \begin{equation}\label{DWg} \begin{cases} \displaystyle \( \frac{\partial^2}{\partial t^2} -\Delta + \frac{\partial}{\partial t} \) u(x,t) =0, &x \in \mathbb{R}^n ,\ t>0,\\ \displaystyle \( u, \frac{\partial u}{\partial t} \) (x,0)=(0, g)(x), &x \in \mathbb{R}^n. \end{cases} \end{equation} The symbol $S_n(t)$ is called the solution operator of \eqref{DWg}. Let $I_{\nu}(s)$ be the modified Bessel function of order $\nu$, \begin{equation} I_{\nu}(s) =\sum_{j=0}^{\infty}\frac{1}{j!\Gamma (j+\nu+1)}\(\frac{s}{2}\)^{2j+\nu}. \end{equation} Put \begin{equation}\label{cn} c_n = \begin{cases} 1 &(n =1),\\ \( (n-2)!! \sigma_{n-1}\( S^{n-1} \) \)^{-1} = 2^{-(n+1)/2} \pi^{-(n-1)/2} &(n \in 2\mathbb{N} +1),\\ \( (n-1)!! \sigma_{n}\( S^n \) \)^{-1} = 2^{-(n+2)/2} \pi^{-n/2} &(n\in 2\mathbb{N}), \end{cases} \end{equation} where $(2\ell -1) !! = (2\ell -1 ) \cdot (2\ell -3 ) \cdot \cdots \cdot 3 \cdot 1$ and $(2\ell )!! = 2 \ell \cdot (2\ell -2) \cdot \cdots \cdot 4 \cdot 2$. \begin{prop}\label{propsol Let $g$ be a smooth function. For any natural number $n$, we have \[ S_n(t)g(x)= \begin{cases} \displaystyle \frac{e^{-t/2}}{2}\int_{x-t}^{x+t} I_0\(\frac{1}{2}\sqrt{t^2-r^2}\)g(y)dy &(n=1),\\ \displaystyle c_n e^{-t/2} \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-1)/2} \int_{B_t^n(x)}I_0\(\frac{1}{2}\sqrt{t^2-r^2}\)g(y)dy &(n\in 2 \mathbb{N} +1 ),\\ \displaystyle 2c_n e^{-t/2} \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-2)/2} \int_{B_t^n(x)}\frac{\cosh \( \frac{1}{2}\sqrt{t^2-r^2} \)}{\sqrt{t^2-r^2}} g(y)dy &(n\in 2\mathbb{N} ). \end{cases} \] \end{prop} \begin{proof The argument is due to the method of descent described in [CH]. It is well known that the solution of the free wave equation \[ \begin{cases} \displaystyle \( \frac{\partial^2}{\partial t^2}-\Delta \) w (x,t) =0,&(x,t)\in \mathbb{R}^n \times (0,\infty),\\ \displaystyle \( w, \frac{\partial w}{\partial t} \) (x,0)=(0,g)(x),&x\in\mathbb{R}^n \end{cases} \] is given by \[ W_n(t)g(x)= \begin{cases} \displaystyle \frac{1}{2}\int_{x-t}^{x+t}g(y)dy &(n=1),\\ \displaystyle c_n \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-3)/2} \(\frac{1}{t} \int_{S_t^{n-1}(x)} g(y) d\sigma_{n-1}(y) \) &(n\in 2\mathbb{N} +1),\\ \displaystyle 2c_n \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-2)/2} \(\int_{B_t^n(x)} \frac{1}{\sqrt{t^2-r^2}}g(y)dy\) &(n \in 2\mathbb{N} ). \end{cases} \] For a point $x=(x_1,\ldots,x_n) \in \mathbb{R}^n$ and a number $\xi \in \mathbb{R}$, we write $z=(x,\xi) \in \mathbb{R}^{n+1}$. Let $u(x,t)$ be the solution of \eqref{DWfg}, and $w(z,t) = \exp ( (\xi +t)/2) u(x,t)$. Then, $w(z,t)$ satisfies the $(n+1)$-dimensional free wave equation with initial datum $(w , \partial w/\partial t)(z,0)=( 0, e^{\xi/2}g(x) )$. We first consider the one-dimensional case. Using the fact \eqref{Bessel_1}, we have \begin{align*} u(x,t) &=e^{-(\xi +t)/2} w(z,t) \\ &=\frac{e^{-t/2}}{2\pi} \int_{B_t^2(z)} \frac{e^{\(y_2 -\xi \)/2}g\( y_1 \)}{\sqrt{t^2 -\lvert x- y_1 \rvert^2 -\lvert \xi - y_2 \rvert^2}} dy\\ &=\frac{e^{-t/2}}{2\pi} \int_{x-t}^{x+t} g\( y_1 \) \( \int_{-\sqrt{t^2 -\lvert x-y_1 \rvert^2}}^{\sqrt{t^2 -\lvert x-y_1 \rvert^2}} \frac{e^{\(y_2 -\xi \)/2}}{\sqrt{t^2 -\lvert x- y_1 \rvert^2 -\lvert \xi - y_2 \rvert^2}} dy_2\) dy_1\\ &=\frac{e^{-t/2}}{2} \int_{x-t}^{x+t} I_0 \( \frac{1}{2} \sqrt{t^2 -\lvert x -y_1 \rvert^2} \) g\( y_1 \) dy_1 . \end{align*} When $n$ is an odd number greater than one, as we have \begin{align*} &\int_{B_t^{n+1}(z)} \frac{e^{\(y_{n+1} -\xi \)/2}g\( y'\)}{\sqrt{t^2 -\lvert x -y' \rvert^2 -\lvert \xi - y_{n+1} \rvert^2 }}dy \\ & =\int_{B_t^n(x)} g\( y' \) \( \int_{-\sqrt{t^2-\lvert x-y'\rvert^2}}^{\sqrt{t^2-\lvert x-y'\rvert^2}} \frac{e^{\(y_{n+1} -\xi \)/2}}{\sqrt{t^2 -\lvert x -y' \rvert^2 -\lvert \xi - y_{n+1} \rvert^2 }}dy_{n+1} \)dy' \\ & =\int_{B_t^n(x)} I_0 \( \frac{1}{2} \sqrt{t^2 -\lvert x -y' \rvert^2} \) g\( y' \) dy' , \end{align*} where $y= ( y' ,y_{n+1} ) \in \mathbb{R}^n \times \mathbb{R}$, we obtain the conclusion. Finally, we consider even-dimensional cases. As we have \begin{align*} &\frac{1}{t} \int_{S_t^n(z)} e^{-\xi /2} g(y) d \sigma_n (y) \\ &=\frac{1}{t} \( \int_{S_t^n(z) \cap \left\{ y_{n+1} \geq \xi \right\}} +\int_{S_t^n(z) \cap \left\{ y_{n+1} \leq \xi \right\}} \) e^{-\xi /2} g(y) d\sigma_n (y) \\ &=\int_{B_t^n(x)} \( e^{\sqrt{t^2 -\lvert x-y'\rvert^2}/2} + e^{-\sqrt{t^2 -\lvert x-y'\rvert^2}/2} \) \frac{g\( y'\)}{\sqrt{t^2 -\lvert x-y'\rvert^2}} dy' \\ &=2\int_{B_t^n(x)} \frac{\cosh \( \frac{1}{2} \sqrt{t^2 -\lvert x-y'\rvert^2}\)}{\sqrt{t^2 -\lvert x-y'\rvert^2}} g\( y' \) dy', \end{align*} the proof is completed. \end{proof} \begin{ex}\label{exsol} {\rm We have the following formulas: \begin{align*} S_1(t)g(x)&=\frac{e^{-t/2}}{2}\int_{x-t}^{x+t} I_0\(\frac{1}{2}\sqrt{t^2-r^2}\)g(y)dy,\\ S_2(t)g(x)&=\frac{e^{-t/2}}{2\pi} \int_{B_t^2(x)}\frac{\cosh \( \frac{1}{2}\sqrt{t^2-r^2} \)}{\sqrt{t^2-r^2}} g(y)dy,\\ S_3(t)g(x)&=\frac{e^{-t/2}}{4\pi} \( \frac{1}{t}\frac{\partial}{\partial t} \) \int_{B_t^3(x)}I_0\( \frac{1}{2}\sqrt{t^2-r^2}\)g(y)dy. \end{align*} These were given in [CH] in the same manner as in Proposition \ref{propsol}.} \end{ex} \begin{rem}\label{Duhamel} {\rm Let $f$ and $g$ be smooth functions. Using the solution operator $S_n(t)$, we can express the solution of the Cauchy problem \eqref{DWfg} as \[ u(x,t)=S_n(t) \( f+g \) (x)+\frac{\partial}{\partial t} S_n(t)f(x). \] } \end{rem} \subsection{Decomposition of the fundamental solution} In this subsection, using Proposition \ref{propsol}, we give the Nishihara decomposed form of the solution operator $S_n(t)$. \begin{thm}\label{thmdecom Let $g$ be a smooth function. \begin{enumerate}[$(1)$] \item Let $n$ be an odd number greater than one. Put \begin{align*} k_\ell (s)&=\frac{1}{2^\ell}\sum_{j=0}^{\infty} \frac{1}{j!(j+\ell )!}\(\frac{s}{2}\)^{2j} =\frac{I_\ell (s)}{s^\ell},\\ J_n(t)g(x)&=\frac{c_ne^{-t/2}}{2^{n-1}} \int_{B_t^n(x)}k_{\frac{n-1}{2}}\(\frac{1}{2}\sqrt{t^2-r^2}\)g(y)dy, \\ \vect{W}_n(t)g(x)&=c_n \sum_{j=0}^{(n-3)/2} \frac{1}{8^j j!}\(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-3)/2-j} \(\frac{1}{t}\int_{S_t^{n-1}(x)}g(y)d\sigma_{n-1}(y) \) . \end{align*} Then, we have \[ S_n(t)g(x)=J_n(t)g(x)+e^{-t/2}\vect{W}_n(t)g(x) . \] Moreover, we have \[ k_\ell (s) =\frac{1}{s^\ell} \frac{e^s}{\sqrt{2\pi s}} \(1-\frac{(\ell -1/2)(\ell +1/2)}{2s}+O\( \frac{1}{s^2} \) \) \] as $s$ goes to infinity. \item Let $n$ be an even number. Put \begin{align*} k_\ell (s) &=\sum_{j=0}^{\infty} \frac{1}{ \( 2(j+\ell ) \) !!(2j+1)!!}s^{2j+1},\\ J_n(t)g(x) &= \frac{c_ne^{-t/2}}{2^{n-2}} \int_{B_t^n(x)}k_{\frac{n}{2}}\(\frac{1}{2}\sqrt{t^2-r^2}\)g(y)dy, \\ \vect{W}_n(t)g(x) &= 2 c_n \sum_{j=0}^{(n-2)/2}\frac{1}{8^j j!} \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-2)/2-j} \int_{B_t^n(x)}\frac{1}{\sqrt{t^2-r^2}}g(y)dy. \end{align*} Then, we have \[ S_n(t)g(x)=J_n(t)g(x)+e^{-t/2}\vect{W}_n(t)g(x) . \] Moreover, we have \[ k_\ell (s) =\frac{e^s}{2s^\ell}\(1-\frac{\ell (\ell -1)}{2s}+O\( \frac{1}{s^2} \)\) \] as $s$ goes to infinity. \end{enumerate} \end{thm} \begin{rem} \label{recursion} {\rm \begin{enumerate}[(1)] \item Let $n$ be an odd number greater than one. From the fact \eqref{Bessel_2}, the kernel $k_\ell (s)$ has the following properties: \[ k_{\ell +1} (s) = \frac{k_\ell'(s)}{s} ,\ k_\ell (0) = \frac{1}{2^\ell \ell !},\ k_\ell' (0) =0 . \] \item Let $n$ be an even number. In the proof of Theorem \ref{thmdecom}, we will show that the kernel $k_\ell (s)$ is defined by the following recursion: \[ k_1(s)=\frac{\cosh(s)-1}{s},\ k_\ell (s)=\frac{k_{\ell-1}' (s)-k_{\ell-1}' (0)}{s}. \] In particular, we have the following properties: \[ k_\ell (0)=0,\ k_\ell' (0)=\frac{1}{2^\ell \ell !}. \] \end{enumerate} } \end{rem} \begin{proof}[Proof of Theorem \ref{thmdecom} (1) The solution formula in Proposition \ref{propsol} implies \begin{align*} S_n(t)g(x) &=c_n e^{-t/2} \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-3)/2} \(\frac{1}{t}\int_{S_t^{n-1}(x)}g(y)d\sigma_{n-1}(y) \)\\ &\quad + c_n e^{-t/2} \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-3)/2} \int_{B_t^n(x)}\frac{I_1\(\frac{1}{2}\sqrt{t^2-r^2}\)}{2\sqrt{t^2-r^2}}g(y)dy. \end{align*} Note that the second term of the right-hand side can be written as \[ \frac{c_n e^{-t/2}}{4} \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-3)/2} \int_{B_t^n(x)}k_1\(\frac{1}{2}\sqrt{t^2-r^2}\)g(y)dy. \] Since we have \[ \frac{1}{t}\(\frac{\partial}{\partial t}k_1\(\frac{1}{2}\sqrt{t^2-r^2}\)\) =\frac{k_1' \(\frac{1}{2}\sqrt{t^2-r^2}\)}{2\sqrt{t^2-r^2}} =\frac{1}{4}k_2\(\frac{1}{2}\sqrt{t^2-r^2}\right), \] we obtain \begin{align*} &\frac{c_ne^{-t/2}}{4} \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-3)/2} \int_{B_t^n(x)}k_1\(\frac{1}{2}\sqrt{t^2-r^2}\)g(y)dy\\ &=\frac{c_ne^{-t/2}}{4} \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-5)/2} \(\frac{k_1(0)}{t}\int_{S_t^{n-1}(x)}g(y)d\sigma_{n-1}(y) \)\\ &\quad+ \frac{c_n e^{-t/2}}{4^2} \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-5)/2} \int_{B_t^n(x)}k_2\(\frac{1}{2}\sqrt{t^2-r^2}\)g(y)dy. \end{align*} Continuing this argument, we obtain the decomposed form of $S_n(t)g$. The asymptotic expansions are direct consequences from \eqref{Bessel_3} and $k_{\ell+1}(s)=k_\ell' (s)/s$. (2) By the solution formula in Proposition \ref{propsol}, we have \begin{align*} S_n(t)g(x) &=2c_n e^{-t/2} \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-2)/2} \int_{B_t^n(x)}\frac{1}{\sqrt{t^2-r^2}}g(y)dy\\ &\quad + 2c_n e^{-t/2} \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-2)/2} \int_{B_t^n(x)} \frac{\cosh \(\frac{1}{2}\sqrt{t^2-r^2}\)-1}{\sqrt{t^2-r^2}} g(y)dy.\\ &=: V_1+K_1. \end{align*} Here, we note that $K_1$ can be written as \begin{align*} K_1 &= c_n e^{-t/2} \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-2)/2} \int_{B_t^n(x)}k_1\(\frac{1}{2}\sqrt{t^2-r^2}\) g(y)dy\\ &= c_n e^{-t/2} \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-4)/2} \int_{B_t^n(x)} \frac{k_1' \(\frac{1}{2}\sqrt{t^2-r^2}\)}{2\sqrt{t^2-r^2}} g(y)dy\\ &= \frac{c_n e^{-t/2}}{2} k_1' (0) \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-4)/2} \int_{B_t^n(x)} \frac{1}{\sqrt{t^2-r^2}} g(y)dy\\ &\quad + \frac{c_n e^{-t/2}}{4} \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-4)/2} \int_{B_t^n(x)} k_2\(\frac{1}{2}\sqrt{t^2-r^2}\) g(y)dy\\ &=: V_2+K_2. \end{align*} In the same manner as in the calculation of $K_1$, we have \begin{align*} K_2 &= \frac{c_n e^{-t/2}}{2^3} k_2' (0) \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-6)/2} \int_{B_t^n(x)} \frac{1}{\sqrt{t^2-r^2}} g(y)dy\\ &\quad +\frac{c_n e^{-t/2}}{4^2} \(\frac{1}{t}\frac{\partial}{\partial t}\)^{(n-6)/2} \int_{B_t^n(x)} k_3\(\frac{1}{2}\sqrt{t^2-r^2}\) g(y)dy\\ &=: V_3+K_3. \end{align*} Continuing this argument, we obtain \[ V_j = \frac{c_n e^{-t/2}}{2^{2 j -3}} k_{j-1}' (0) \(\frac{1}{t}\frac{\partial}{\partial t}\)^{n/2-j} \int_{B_t^n(x)} \frac{1}{\sqrt{t^2-r^2}} g(y)dy \] and the decomposed form of $S_n(t) g$. The asymptotic expansions follow from \[ k_1(s)=\frac{e^s}{2s}+\frac{e^{-s}}{2s}-\frac{1}{s} \] and the recursion of $k_\ell (s)$. \end{proof} \begin{rem}\label{decom1} {\rm Let $g$ be a smooth function. Put \begin{align*} J_1(t)g(x)&=\frac{e^{-t/2}}{2}\int_{x-t}^{x+t} \( I_0\(\frac{1}{2}\sqrt{t^2-r^2}\)-1 \) g(y)dy,\\ \vect{W}_1(t)g(x) &=W_1(t)g(x) =\frac{1}{2}\int_{x-t}^{x+t}g(y)dy . \end{align*} Thanks to Proposition \ref{propsol}, we get the Nishihara decomposed form of $S_1(t)$ as \[ S_1(t)g(x) = J_1(t)g (x) +e^{-t/2} \vect{W}_1(t)g(x) \] which was given in [MN]. } \end{rem} \begin{ex}\label{decom23} {\rm We have the following formulas: \begin{align*} J_2(t)g(x)&=\frac{e^{-t/2}}{2\pi}\int_{B_t^2(x)} \frac{\cosh \( \frac{1}{2}\sqrt{t^2-r^2} \)-1}{\sqrt{t^2-r^2}} g(y)dy,\\ J_3(t)g(x)&=\frac{e^{-t/2}}{4\pi} \int_{B_t^3(x)} \frac{I_1\( \frac{1}{2}\sqrt{t^2-r^2} \)}{2\sqrt{t^2-r^2}} g(y)dy,\\ \vect{W}_2(t)g(x) &=W_2(t)g(x) =\frac{1}{2\pi}\int_{B_t^2(x)} \frac{1}{\sqrt{t^2-r^2}}g(y)dy,\\ \vect{W}_3(t)g(x) &=W_3(t)g(x) =\frac{1}{4\pi t}\int_{S_t^2 (x)}g(y)d\sigma_2 (y). \end{align*} These were given in [HO, Nis] using the explicit form of $S_n(t)g(x)$ given in Example \ref{exsol}. } \end{ex} In [HO, MN, Nis], it was shown that $J_n(t)g$ behaves like $P_n(t)g$ as $t$ goes to infinity. More precisely, in the case of $1\leq n \leq 3$, for $t>0$ and $1\le q\le p\le\infty$, the $L^p$-$L^q$ estimate \begin{equation}\label{estimate_low} \|J_n(t)g-P_n(t)g\|_{L^p}\leq Ct^{-\frac{n}{2}\(\frac{1}{q}-\frac{1}{p}\)-1} \| g \|_{L^q} \end{equation} was shown. The Nishihara decomposition and the estimate \eqref{estimate_low} imply the following properties of the damped wave equation (see [Nis, pp. 632]): \begin{itemize} \item The damped wave equation does not have the smoothing effect, and the singularity of the initial datum propagates along the light cone by the wave property. However, the strength of the singularity decays exponentially by the damping effect. \item If the initial datum is sufficiently smooth, then the damped wave equation may have the same properties as those to parabolic equations under some suitable situations. \end{itemize} For the case of $n\ge 4$, in [Nar], Narazaki proved a similar decomposition to Theorem \ref{thmdecom} in the Fourier space. However, the explicit form of the decomposition in the configuration space was not known. In section 5, using Theorem \ref{thmdecom}, we will give an estimate of the difference $S_n(t)g- P_n(t)g -e^{-t/2}\vect{W}_n(t)g$, which is slightly sharper than that of [Nar]. \subsection{Decomposition of the solution with general initial datum} In this subsection, we give the Nishihara decomposed form of the solution of \eqref{DWfg}. \begin{prop}\label{prop_de_fg} Let $f$ and $g$ be smooth functions, $h=f+g$, and $u$ the unique classical solution of the Cauchy problem \eqref{DWfg}. \begin{enumerate}[$(1)$] \item Let $n=1$. Put \begin{align*} \tilde{J}_1(t)f(x) &=\frac{e^{-t/2}}{4}\int_{x-t}^{x+t} \( \frac{t}{\sqrt{t^2-r^2}} I_1 \( \frac{1}{2}\sqrt{t^2-r^2} \) -I_0 \( \frac{1}{2}\sqrt{t^2-r^2} \) \) f(y) dy, \\ \widehat{\vect{W}}_1(t)f(x) &= \frac{1}{2} \( f(x+t) + f(x-t) \) ,\\ \widetilde{\vect{W}}_1(t;f,g) (x) &= \vect{W}_1(t) h(x) + \widehat{\vect{W}}_1(t)f(x) . \end{align*} Then, we have \[ \frac{\partial}{\partial t} S_1 (t) f(x) = \tilde{J}_1 (t) f(x) + e^{-t/2} \widehat{\vect{W}}_1(t) f(x). \] In other words, the solution $u(x,t)$ is expressed as \[ u(x,t) = J_1(t) h(x) + \tilde{J}_1(t) f( x ) + e^{-t/2}\widetilde{\vect{W}}_1(t;f,g)(x). \] \item Let $n$ be an odd number greater than one. Put \begin{align*} \tilde{J}_n(t) f(x) &= \frac{c_n e^{-t/2}}{2^{n+1}} \int_{B_t^n(x)} \( t k_{\frac{n+1}{2}}\(\frac{1}{2}\sqrt{t^2-r^2}\) -2k_{\frac{n-1}{2}} \( \frac{1}{2}\sqrt{t^2 -r^2} \) \) f(y)dy ,\\ \widehat{\vect{W}}_n(t)f(x) &= \frac{c_n}{2^{\frac{3(n-1)}{2}} \( \frac{n-1}{2} \) !} \int_{S_t^{n-1}(x)}f(y)d\sigma_{n-1}(y) ,\\ \widetilde{\vect{W}}_n(t; f, g)(x) &= \frac{1}{2}\vect{W}_n(t)f(x) + \vect{W}_n(t)g(x) + \widehat{\vect{W}}_n(t)f(x) + \frac{\partial}{\partial t} \vect{W}_n(t)f(x) . \end{align*} Then, we have \[ \frac{\partial}{\partial t} S_n(t) f(x) =\tilde{J}_n(t)f(x) + e^{-t/2} \widehat{\vect{W}}_n(t)f(x) -\frac{e^{-t/2}}{2} \vect{W}_n(t)f(x) + e^{-t/2} \frac{\partial}{\partial t} \vect{W}_n(t)f(x). \] In other words, the solution $u(x,t)$ is expressed as \[ u( x, t )=J_n(t) h (x) + \tilde{J}_n(t) f (x) + e^{-t/2} \widetilde{\vect{W}}_n(t; f, g)(x). \] \item Let $n$ be an even number. Put \begin{align*} \tilde{J}_n(t) f(x) &= \frac{c_n e^{-t/2}}{2^n} \int_{B_t^n(x)}\( t k_{\frac{n}{2}+1}\(\frac{1}{2}\sqrt{t^2-r^2}\) -2k_{\frac{n}{2}}\(\frac{1}{2}\sqrt{t^2-r^2}\)\)f(y)dy, \\ \widehat{\vect{W}}_n(t)f(x) &= \frac{c_n t}{2^{\frac{3n-2}{2}} \( \frac{n}{2} \) !} \int_{B_t^n(x)}\frac{1}{\sqrt{t^2-r^2}} f(y) dy, \\ \widetilde{\vect{W}}_n(t; f, g)(x) &= \frac{1}{2}\vect{W}_n(t)f(x) + \vect{W}_n(t)g(x) + \widehat{\vect{W}}_n(t)f(x) + \frac{\partial}{\partial t} \vect{W}_n(t)f(x) . \end{align*} Then, we have \[ \frac{\partial}{\partial t} S_n(t) f(x) =\tilde{J}_n(t)f(x) + e^{-t/2} \widehat{\vect{W}}_n(t)f(x) -\frac{e^{-t/2}}{2} \vect{W}_n(t)f(x) + e^{-t/2} \frac{\partial}{\partial t} \vect{W}_n(t)f(x). \] In other words, the solution $u(x,t)$ is expressed as \[ u( x, t )=J_n(t) h (x) + \tilde{J}_n(t) f (x) + e^{-t/2} \widetilde{\vect{W}}_n(t; f, g)(x). \] \end{enumerate} \end{prop} \begin{proof} (1) From Proposition \ref{propsol}, direct computation shows \begin{align*} \frac{\partial}{\partial t}S_1(t) f(x) &= \frac{e^{-t/2}}{4}\int_{x-t}^{x+t} \( \frac{t}{\sqrt{t^2-r^2}} I'_0 \( \frac{1}{2}\sqrt{t^2-r^2} \)-I_0 \( \frac{1}{2}\sqrt{t^2-r^2} \) \) f(y) dy\\ &\quad +\frac{e^{-t/2}}{2} \( f(x+t) + f(x-t) \) \\ &= \tilde{J}_1(t)f(x) + e^{-t/2}\widehat{\vect{W}}_1(t)f(x) . \end{align*} Here, we used $I_0'(s) = I_1(s)$ (see \eqref{Bessel_2}). From Remark \ref{Duhamel}, we get the conclusion. (2) From Remark \ref{recursion}, we have \begin{align*} \frac{\partial}{\partial t}J_n(t)f(x)&=\frac{c_n}{2^{n-1}} \int_{B_t^n(x)}\frac{\partial}{\partial t} \( e^{-t/2}k_{\frac{n-1}{2}}\(\frac{1}{2}\sqrt{t^2-r^2}\)\) f(y)dy\\ &\quad +\frac{c_n e^{-t/2}}{2^{\frac{3(n-1)}{2}} \( \frac{n-1}{2} \) !} \int_{S_t^{n-1}(x)}f(y)d\sigma_{n-1}(y)\\ &= \tilde{J}_n(t)g(x) + e^{-t/2}\widehat{\vect{W}}_n(t)g(x). \end{align*} From Remark \ref{Duhamel}, we get the conclusion. (3) From Remark \ref{recursion}, we have \begin{align*} \frac{\partial}{\partial t} J_n(t)g(x) &=\frac{c_n}{2^{n-2}} \int_{B_t^n(x)}\frac{\partial}{\partial t} \(e^{-t/2} k_{\frac{n}{2}}\(\frac{1}{2}\sqrt{t^2-r^2}\)\)g(y)dy \\ &= \tilde{J}_n(t)g(x) +e^{-t/2}\widehat{\vect{W}}_n(t)g(x). \end{align*} From Remark \ref{Duhamel}, we get the conclusion. \end{proof} \section{Preliminary estimates} In this section, we prepare some estimates which will be frequently used for studying the behavior of time-delayed hot spots. We postpone the proofs of the following lemmas until Appendix 1. This is because their proofs consist of tedious calculations. \begin{lem}\label{lem_est_Wn} Let $f$ and $g$ be smooth bounded functions defined on $\mathbb{R}^n$. \begin{enumerate}[$(1)$] \item For any $x \in \mathbb{R}$, $t >0$ and $\alpha = (\alpha_1 , \alpha_2 ) \in \mathbb{Z}_{\geq 0}^2$, we have the following inequalities: \begin{align*} \lvert \frac{\partial^{\lvert \alpha \rvert}}{\partial x^{\alpha_1} \partial t^{\alpha_2}}\vect{W}_1(t)g(x) \rvert & \le (1+t) \| g\|_{W^{\lvert \alpha \rvert, \infty}} ,\\ \lvert \frac{\partial^{\lvert \alpha \rvert}}{\partial x^{\alpha_1} \partial t^{\alpha_2}} \widehat{\vect{W}}_1(t)f(x) \rvert &\le \| f \|_{W^{\lvert \alpha \rvert,\infty}},\\ \lvert \frac{\partial^{\lvert \alpha \rvert}}{\partial x^{\alpha_1} \partial t^{\alpha_2}} \widetilde{\vect{W}}_1(t;f,g)(x) \rvert &\le 2 (1+t) \( \| f \|_{W^{\lvert \alpha \rvert , \infty}} + \| g \|_{W^{\lvert \alpha \rvert,\infty}} \) , \end{align*} where $\lvert \alpha \rvert=\alpha_1 +\alpha_2$. \item Let $n$ be an odd number greater than one. There exists a positive constant $C=C(n)$ such that, for any $x \in \mathbb{R}^n$, $t >0$ and $\alpha = (\alpha_1 ,\ldots , \alpha_{n+1}) \in \mathbb{Z}_{\geq 0}^{n+1}$, we have the following inequalities: \begin{align*} \lvert \frac{\partial^{\lvert \alpha \rvert}}{\partial x_1^{\alpha_1} \cdots \partial x_n^{\alpha_n} \partial t^{\alpha_{n+1}}}\vect{W}_n(t)g(x) \rvert &\le C (1+t)^{n-2} \| g\|_{W^{(n-3)/2+\lvert \alpha \rvert, \infty}} ,\\ \lvert \frac{\partial^{\lvert \alpha \rvert}}{\partial x_1^{\alpha_1} \cdots \partial x_n^{\alpha_n} \partial t^{\alpha_{n+1}}} \widehat{\vect{W}}_n(t)f(x) \rvert &\le C (1+t)^{n-1} \| f \|_{W^{\lvert \alpha \rvert,\infty}},\\ \lvert \frac{\partial^{\lvert \alpha \rvert}}{\partial x_1^{\alpha_1} \cdots \partial x_n^{\alpha_n} \partial t^{\alpha_{n+1}}} \widetilde{\vect{W}}_n(t;f,g)(x) \rvert &\le C (1+t)^{n-1} \( \| f \|_{W^{(n-1)/2+\lvert \alpha \rvert,\infty}} + \| g \|_{W^{(n-3)/2+\lvert \alpha \rvert,\infty}} \) , \end{align*} where $\lvert \alpha \rvert =\alpha_1 + \cdots + \alpha_{n+1}$. \item Let $n$ be an even number greater than one. There exists a positive constant $C=C(n)$ such that, for any $x \in \mathbb{R}^n$, $t >0$ and $\alpha = (\alpha_1 ,\ldots , \alpha_{n+1}) \in \mathbb{Z}_{\geq 0}^{n+1}$, we have the following inequalities: \begin{align*} \lvert \frac{\partial^{\lvert \alpha \rvert}}{\partial x_1^{\alpha_1} \cdots \partial x_n^{\alpha_n} \partial t^{\alpha_{n+1}}} \vect{W}_n(t)g(x) \rvert &\le C(1+t)^{n-1} \| g\|_{W^{n/2-1+\lvert \alpha \rvert, \infty}} ,\\ \lvert \frac{\partial^{\lvert \alpha \rvert}}{\partial x_1^{\alpha_1} \cdots \partial x_n^{\alpha_n} \partial t^{\alpha_{n+1}}} \widehat{\vect{W}}_n(t)f(x) \rvert &\le C (1+t)^n \| f \|_{W^{\lvert \alpha \rvert,\infty}},\\ \lvert \frac{\partial^{\lvert \alpha \rvert}}{\partial x_1^{\alpha_1} \cdots \partial x_n^{\alpha_n} \partial t^{\alpha_{n+1}}} \widetilde{\vect{W}}_n(t;f,g)(x) \rvert &\le C (1+t)^n \( \| f \|_{W^{n/2 +\lvert \alpha \rvert,\infty}} + \| g \|_{W^{n/2 -1+\lvert \alpha \rvert,\infty}} \) , \end{align*} where $\lvert \alpha \rvert =\alpha_1 + \cdots + \alpha_{n+1}$. \end{enumerate} \end{lem} \begin{lem}\label{list_J} Let $h:\mathbb{R}^n \to \mathbb{R}$ be a smooth function. \begin{enumerate}[$(1)$] \item Let $n=1$. We have the following identities: \begin{align*} \frac{\partial}{\partial x} J_1(t)h(x) &=-\frac{e^{-t/2}}{4} \int_{x-t}^{x+t} \frac{1}{\sqrt{t^2-r^2}} I_1 \( \frac{1}{2} \sqrt{t^2-r^2} \) h(y) (x-y) dy ,\\ \frac{\partial^2}{\partial x^2} J_1(t)h(x) &=\frac{te^{-t/2}}{8} \( h(x+t) +h(x-t) \) \\ &\quad -\frac{e^{-t/2}}{4} \int_{x-t}^{x+t} \frac{1}{\sqrt{t^2-r^2}} I_1 \( \frac{1}{2} \sqrt{t^2-r^2} \) h(y) dy \\ &\quad +\frac{e^{-t/2}}{8} \int_{x-t}^{x+t} \frac{1}{t^2-r^2} I_2 \( \frac{1}{2} \sqrt{t^2-r^2} \) h(y) (x-y)^2 dy . \end{align*} \item Let $n$ be an odd number greater than one. For any direction $\omega \in S^{n-1}$, we have the following identities: \begin{align*} \nabla J_n(t)h(x) &= \frac{c_ne^{-t/2}}{2^{n-1}} k_{\frac{n-1}{2}}(0) t^{n-1} \int_{S^{n-1}} h(x+t\theta ) \theta d\sigma_{n-1} (\theta ) \\ &\quad - \frac{c_ne^{-t/2}}{2^{n+1}} \int_{B^n_t(x)} k_{\frac{n+1}{2}} \( \frac{1}{2} \sqrt{t^2 -r^2} \) h(y) (x-y) dy ,\\ (\omega \cdot \nabla )^2 J_n(t) h(x) &= \frac{c_ne^{-t/2}}{2^{n-1}} k_{\frac{n-1}{2}}(0) t^{n-1} \int_{S^{n-1}} \omega \cdot \nabla h(x+t\theta ) \omega \cdot \theta d\sigma_{n-1} (\theta ) \\ &\quad + \frac{c_ne^{-t/2}}{2^{n+1}} k_{\frac{n+1}{2}}(0) t^n \int_{S^{n-1}} h(x+t\theta ) \( \omega \cdot \theta \)^2 d\sigma_{n-1} (\theta ) \\ &\quad - \frac{c_ne^{-t/2}}{2^{n+1}} \int_{B^n_t(x)} k_{\frac{n+1}{2}} \( \frac{1}{2} \sqrt{t^2 -r^2} \) h(y) dy \\ &\quad + \frac{c_ne^{-t/2}}{2^{n+3}} \int_{B^n_t(x)} k_{\frac{n+3}{2}} \( \frac{1}{2} \sqrt{t^2 -r^2} \) h(y) \( \omega \cdot (x-y) \)^2 dy. \end{align*} \item Let $n$ be an even number. For any direction $\omega \in S^{n-1}$, we have the following identities: \begin{align*} \nabla J_n(t)h(x) &= \frac{c_ne^{-t/2}}{2^{n-1}} k'_{\frac{n}{2}}(0) t^n \int_{B^n} \frac{1}{\sqrt{1-\lvert z \rvert^2}} h(x+tz ) z dz \\ &\quad - \frac{c_ne^{-t/2}}{2^n} \int_{B^n_t(x)} k_{\frac{n}{2}+1} \( \frac{1}{2} \sqrt{t^2 -r^2} \) h(y) (x-y) dy ,\\ (\omega \cdot \nabla )^2 J_n(t) h(x) &= \frac{c_n e^{-t/2}}{2^{n-1}} k'_{\frac{n}{2}}(0) t^n \int_{B^n} \frac{1}{\sqrt{1-\lvert z \rvert^2}} \omega \cdot \nabla h(x+tz) \omega \cdot z dz \\ &\quad + \frac{c_n e^{-t/2}}{2^{n+1}} k'_{\frac{n}{2}+1}(0) t^{n+1} \int_{B^n} \frac{1}{\sqrt{1-\lvert z \rvert^2}} h(x+tz ) \( \omega \cdot z \)^2 dz \\ &\quad - \frac{c_ne^{-t/2}}{2^n} \int_{B^n_t(x)} k_{\frac{n}{2}+1} \( \frac{1}{2} \sqrt{t^2 -r^2} \) h(y) dy \\ &\quad + \frac{c_ne^{-t/2}}{2^{n+2}} \int_{B^n_t(x)} k_{\frac{n}{2}+2} \( \frac{1}{2} \sqrt{t^2 -r^2} \) h(y) \( \omega \cdot (x-y) \)^2 dy. \end{align*} \end{enumerate} \end{lem} \begin{lem}\label{En} Let $h$ be a smooth function with compact support. Put \begin{align*} E_n(r,t)= \begin{cases} \displaystyle \frac{ e^{-t/2}}{4}\frac{1}{\sqrt{t^2-r^2}} I_1 \( \frac{1}{2}\sqrt{t^2-r^2} \) &(n=1),\\ \displaystyle \frac{c_n e^{-t/2}}{2^{n+1}} k_{\frac{n+1}{2}} \(\frac{1}{2}\sqrt{t^2-r^2}\) &(n\in 2\mathbb{N} +1),\\ \displaystyle \frac{c_n e^{-t/2}}{2^n} k_{\frac{n}{2}+1}\( \frac{1}{2}\sqrt{t^2-r^2} \) &(n\in 2\mathbb{N} ). \end{cases} \end{align*} If $x\in CS(h) + (t-d_h)B^n$ and $t \geq d_h$, then we have \[ \nabla J_n(t)h(x)=-\int_{B_t^n(x)}E_n(r,t)h(y) (x-y) dy . \] \end{lem} \begin{lem}\label{En_asymp} Let $\varphi :[0, +\infty ) \to [0,+\infty )$ be a non-decreasing function with $\varphi (0)=0$. \begin{enumerate}[$(1)$] \item If $\sqrt{t^2-\varphi(t)^2}$ diverges as $t$ goes to infinity, then we have \[ E_n \( \varphi (t),t \)=\frac{1}{2(4\pi)^{n/2} \( t^2-\varphi (t)^2 \)^{n/4+1/2}} \exp \( \frac{-t+\sqrt{t^2-\varphi (t)^2}}{2} \) \( 1+O\( \frac{1}{\sqrt{t^2-\varphi (t)^2}} \) \) \] as $t$ goes to infinity. \item If $\varphi(t)$ is of small order of $t$ as $t$ goes to infinity, then we have \[ E_n \( \varphi (t),t \) =\frac{1}{2(4\pi)^{n/2}t^{n/2+1}} \exp \( \frac{-t+\sqrt{t^2-\varphi (t)^2}}{2} \) \( 1+O\( \frac{1}{t} \) + O\( \frac{\varphi(t)^2}{t^2} \) \) \] as $t$ goes to infinity. \item If $\varphi (t)$ is of small order of $\sqrt{t}$, then we have \[ E_n \( \varphi(t),t \) = \frac{1}{2(4\pi)^{n/2}t^{n/2+1}} \( 1+ O \( \frac{1}{t} \) + O \( \frac{\varphi (t)^2}{t} \) \) \] as $t$ goes to infinity. \end{enumerate} \end{lem \begin{lem}\label{list_tildeJ} Let $f:\mathbb{R}^n \to \mathbb{R}$ be a smooth function. \begin{enumerate}[$(1)$] \item Let $n=1$. We have the following identities: \begin{align*} &\frac{\partial}{\partial x} \tilde{J}_1(t) f(x) \\ &= \frac{(t-4)e^{-t/2}}{16} \( f(x+t) -f(x-t) \) \\ &\quad - \frac{e^{-t/2}}{8} \int_{x-t}^{x+t} \frac{1}{\sqrt{t^2 -r^2}} \( \frac{t}{\sqrt{t^2-r^2}} I_2 \( \frac{1}{2} \sqrt{t^2-r^2} \) - I_1 \( \frac{1}{2} \sqrt{t^2-r^2} \) \) f(y) (x-y) dy,\\ &\frac{\partial^2}{\partial x^2} \tilde{J}_1(t) f(x) \\ &= \frac{(t-4)e^{-t/2}}{16} \( f' (x+t) -f' (x-t) \) \\ &\quad + \frac{\( t^2-8t \)e^{-t/2}}{256} \( f(x+t)+f(x-t) \) \\ &\quad - \frac{e^{-t/2}}{8} \int_{x-t}^{x+t} \frac{1}{\sqrt{t^2 -r^2}} \( \frac{t}{\sqrt{t^2-r^2}} I_2 \( \frac{1}{2} \sqrt{t^2-r^2} \) - I_1 \( \frac{1}{2} \sqrt{t^2-r^2} \) \) f(y) dy \\ &\quad +\frac{e^{-t/2}}{16} \int_{x-t}^{x+t} \frac{1}{\( t^2-r^2\)} \( \frac{t}{\sqrt{t^2-r^2}} I_3 \( \frac{1}{2} \sqrt{t^2-r^2} \) - I_2 \( \frac{1}{2} \sqrt{t^2-r^2} \) \) f(y) \( x-y \)^2 dy. \end{align*} \item Let $n$ be an odd number greater than one. For any direction $\omega \in S^{n-1}$, we have the following identities: \begin{align*} &\nabla \tilde{J}_n(t)f(x)\\ &=\frac{c_ne^{-t/2}}{2^{n+1}}\( tk_{\frac{n+1}{2}}(0) -2k_{\frac{n-1}{2}}(0) \) t^{n-1} \int_{S^{n-1}} f(x+t\theta ) \theta d\sigma_{n-1} (\theta ) \\ &\quad - \frac{c_n e^{-t/2}}{2^{n+3}} \int_{B^n_t (x)} \( t k_{\frac{n+3}{2}} \( \frac{1}{2} \sqrt{t^2-r^2} \) -2 k_{\frac{n+1}{2}} \( \frac{1}{2} \sqrt{t^2-r^2} \) \) f(y) (x-y) dy,\\ &\( \omega \cdot \nabla \)^2 \tilde{J}_n(t)f(x)\\ &=\frac{c_ne^{-t/2}}{2^{n+1}}\( tk_{\frac{n+1}{2}}(0) -2k_{\frac{n-1}{2}}(0) \) t^{n-1} \int_{S^{n-1}} \omega \cdot \nabla f(x+t\theta ) \omega \cdot \theta d\sigma_{n-1} (\theta ) \\ &\quad + \frac{c_ne^{-t/2}}{2^{n+3}}\( tk_{\frac{n+3}{2}}(0) -2k_{\frac{n+1}{2}}(0) \) t^n \int_{S^{n-1}} f(x+t\theta ) (\omega \cdot \theta )^2 d\sigma_{n-1} (\theta ) \\ &\quad - \frac{c_n e^{-t/2}}{2^{n+3}} \int_{B^n_t (x)} \( t k_{\frac{n+3}{2}} \( \frac{1}{2} \sqrt{t^2-r^2} \) -2 k_{\frac{n+1}{2}} \( \frac{1}{2} \sqrt{t^2-r^2} \) \) f(y) dy \\ &\quad + \frac{c_n e^{-t/2}}{2^{n+5}} \int_{B^n_t (x)} \( t k_{\frac{n+5}{2}} \( \frac{1}{2} \sqrt{t^2-r^2} \) -2 k_{\frac{n+3}{2}} \( \frac{1}{2} \sqrt{t^2-r^2} \) \) f(y) \( \omega \cdot (x-y ) \)^2 dy . \end{align*} \item Let $n$ be an even number. For any direction $\omega \in S^{n-1}$, we have the following identities: \begin{align*} &\nabla \tilde{J}_n(t)f(x)\\ &=\frac{c_ne^{-t/2}}{2^{n+1}}\( tk'_{\frac{n}{2}+1}(0) -2k'_{\frac{n}{2}}(0) \) t^n \int_{B^n} \frac{1}{\sqrt{1-\lvert z \rvert^2}} f(x+tz) z dz \\ &\quad - \frac{c_n e^{-t/2}}{2^{n+2}} \int_{B^n_t (x)} \( t k_{\frac{n}{2}+2} \( \frac{1}{2} \sqrt{t^2-r^2} \) -2 k_{\frac{n}{2}+1} \( \frac{1}{2} \sqrt{t^2-r^2} \) \) f(y) (x-y) dy,\\ &\( \omega \cdot \nabla \)^2 \tilde{J}_n(t)f(x)\\ &=\frac{c_ne^{-t/2}}{2^{n+1}}\( tk'_{\frac{n}{2}+1}(0) -2k'_{\frac{n}{2}}(0) \) t^n \int_{B^n} \frac{1}{\sqrt{1-\lvert z \rvert^2}} \omega \cdot \nabla f(x+tz ) \omega \cdot z dz \\ &\quad + \frac{c_ne^{-t/2}}{2^{n+3}}\( tk'_{\frac{n}{2}+3}(0) -2k'_{\frac{n}{2}+1}(0) \) t^{n+1} \int_{B^n} \frac{1}{\sqrt{1-\lvert z \rvert^2}} f(x+tz) \( \omega \cdot z \)^2 dz \\ &\quad - \frac{c_n e^{-t/2}}{2^{n+2}} \int_{B^n_t (x)} \( t k_{\frac{n}{2}+2} \( \frac{1}{2} \sqrt{t^2-r^2} \) -2 k_{\frac{n}{2}+1} \( \frac{1}{2} \sqrt{t^2-r^2} \) \) f(y) dy \\ &\quad + \frac{c_n e^{-t/2}}{2^{n+4}} \int_{B^n_t (x)} \( t k_{\frac{n}{2}+3} \( \frac{1}{2} \sqrt{t^2-r^2} \) -2 k_{\frac{n}{2}+2} \( \frac{1}{2} \sqrt{t^2-r^2} \) \) f(y) \( \omega \cdot (x-y ) \)^2 dy . \end{align*} \end{enumerate} \end{lem} \begin{lem}\label{lem_tildeJ} Let $f$ be a non-zero smooth function with compact support. \begin{enumerate}[$(1)$] \item There exists a positive constant $C=C(n)$ such that, for any $x \in \mathbb{R}^n$ and $t > 0$, we have \[ \lvert \tilde{J}_n(t) f(x) \rvert \le C(1+ t)^{-n/2-1} \| f \|_{L^1} . \] \item Let $\psi :[0, +\infty ) \to [0,+\infty )$ be a non-decreasing function with $\psi (0)=0$. Suppose that $\psi (t)$ is of small order of $\sqrt{t}$. Let \[ T_0 (\psi ) = \min \left\{ T>0 \lvert \forall t\geq T,\ t \geq \psi (t) +d_f \right\} \right. . \] There exists a positive constant $C=C(n,d_f ,\psi )$ such that, for any $x \in CS(f) +\psi (t) B^n$ and $t\geq T_0 (\psi )$, we have \[ \lvert \nabla \tilde{J}_n(t) f (x) \rvert \leq C \( 1+t \)^{-n/2 -3} \( 1+ t+ \psi (t)^2 \) \( 1+ \psi (t) \) \| f \|_{L^1} . \] \item Let $R$ be a positive constant. There exists a positive constant $C=C(n,d_f,R)$ such that, for any $x \in CS(f)+RB^n$, $\omega \in S^{n-1}$ and $t \geq R+d_f$, we have \[ \lvert \( \omega \cdot \nabla \)^2 \tilde{J}_n(t)f(x) \rvert \le C(1+t)^{-n/2-2} \| f \|_{L^1} . \] \end{enumerate} \end{lem} \section{Movement of the time-delayed hot spots} Let $u$ denote the classical solution of the Cauchy problem \eqref{DWfg}. In this section, we investigate the asymptotic behavior of spatial maximizers of the function $u(\cdot ,t) :\mathbb{R}^n \to \mathbb{R}$. \begin{notation}\label{assump} {\rm Let us list up our notation for this section. \begin{enumerate} \item[($fg$)] Let $f$ and $g$ be compactly supported smooth functions such that the sum of them $h:=f+g$ is non-zero and non-negative. \item[($\mathcal{H}$)] For a function $\phi :\mathbb{R}^n \to \mathbb{R}$, we denote by $\mathcal{M} (\phi )$ and $\mathcal{C} (\phi )$ the set of maximum and critical points of $\phi$, respectively: \[ \mathcal{M} (\phi )= \left\{ x \in \mathbb{R}^n \lvert \phi (x) = \max_{\xi \in \mathbb{R}^n} \phi (\xi ) \right\} \right. ,\ \mathcal{C} (\phi )= \left. \left\{ x \in \( \supp \phi \)^\circ \rvert \nabla \phi (x)=0 \right\} . \] In particular, we write \[ \mathcal{H}(t)=\mathcal{M} \( u(\cdot,t) \) \] and call a point $p \in \mathcal{H}(t)$ a {\it time-delayed hot spot} at time $t$. We remark that $\mathcal{H}(t)$ is always contained in $\mathcal{C} \( u(\cdot, t) \)$. \item[($m$)] Under the condition $(fg)$, we investigate the distance between time-delayed hot spots and the centroid (center of mass) of $h$, \[ \displaystyle m_h =\left. \int_{\mathbb{R}^n}h (y) y dy \right/ \int_{\mathbb{R}^n}h (y)dy. \] We remark that the centroid $m_h$ is in the interior of the convex hull of $\supp h$. \item[$(\delta)$] Let $K$ and $L$ be convex bodies in $\mathbb{R}^n$. Let \[ \delta (K,L) = \sup_{\eta \in L} \dist (\eta ,K). \] We remark that the parallel body $K+\delta (K,L)B^n$ contains the convex body $L$. \item[($\psi$)] Let $\psi :[0,+\infty ) \to [0,+\infty )$ be a non-decreasing function such that $\psi (0)= 0$ and $\psi (t)$ is of small order of $\sqrt{t}$ as $t$ goes to infinity. \item[($\varphi$)] Let $\varphi :[0,+\infty ) \to [0,+\infty )$ be a non-decreasing function such that $\varphi (0)= 0$ and $\varphi (t)$ is of small order of $t$ as $t$ goes to infinity. \item[($T_0$)] Let $f$ and $g$ be as in $(fg)$. For a non-decreasing function $\phi :[0,+\infty ) \to[0,+\infty )$, let \begin{align*} T_0( \phi ) &= T_0 \( \phi ; d_h, \delta \( CS(h) ,CS(f) \) , d_f \) \\ &= \min \left\{ T\geq 0 \lvert \forall t \geq T,\ t \geq \phi (t) + \max \left\{ d_h ,\ \delta \( CS(h) ,CS(f) \) +d_f \right\} \right\} \right. . \end{align*} We remark that, if $t \geq T_0 (\phi)$, then, for any $x \in CS(h) +\phi (t) B^n$, the ball $B_t^n (x)$ contains the union of $CS(h)$ and $CS(f)$. \end{enumerate} } \end{notation} \subsection{Asymptotic behavior of time-delayed hot spots} In this subsection, we are interested in the behavior of $\mathcal{C} ( u(\cdot, t) )$ and $\mathcal{H}(t)$. As we mentioned in the introduction, in [CK], Chavel and Karp showed that, for each $t$, the non-empty set $H_\phi (t) := \mathcal{M} \( P_n (t)\phi \)$ is contained in the convex hull of the support of $\phi$, and that \begin{equation} \sup \left\{ \lvert x-m_\phi \rvert \lvert x\in H_\phi (t) \right\} \right. =O \( \frac{1}{t} \) \end{equation} as $t$ goes to infinity. Let us show that similar results hold for the damped wave equation \eqref{DWfg}. \begin{rem} {\rm Under the condition $(fg)$ in Assumption and Notation \ref{assump}, for each $t>0$, the support of $u(\cdot ,t)$ is compact. More precisely, the support of $u(\cdot ,t)$ is contained in the union of two parallel bodies $CS(h) + t B^n$ and $CS(f) +tB^n$. Hence we always have a time-delayed hot spot. } \end{rem} \begin{lem}\label{hp1} We use Notation \ref{assump}. There exist a positive constant $C$ and a time $T \geq T_0 ( \psi )$ such that, for any $t \geq T$, we have \begin{align*} & \sup \left\{ \lvert x - m_h \rvert \lvert x\in \mathcal{C}\(u\( \cdot, t \)\) \cap \( CS(h) +\psi (t) B^n \) \right\} \right. \\ &\leq C \( \frac{1+\psi (t)^2}{t} + \frac{1+\psi (t)}{t} \frac{\| f\|_{L^1}}{\| h \|_{L^1}} + e^{-t/2}t^{3n/2} \frac{\| f\|_{W^{*,\infty}} + \| g \|_{W^{*,\infty}}}{\| h\|_{L^1}} \) . \end{align*} \end{lem} \begin{proof} We give a proof for even dimensional cases. The other cases go parallel. Let $x$ be a point in $\mathcal{C}( u( \cdot, t )) \cap (CS(h) + \psi (t) B^n)$. We remark that, from Proposition \ref{prop_de_fg} and Lemma \ref{En}, we have \[ x= \left. \( \nabla \tilde{J}_n(t) f(x) + e^{-t/2} \nabla \widetilde{\vect{W}}_n(t;f,g)(x) + \int_{\mathbb{R}^n} E_n(r,t)h(y)ydy \) \right/ \int_{\mathbb{R}^n} E_n(r,t)h(y)dy . \] Since the function $E_n(\cdot ,t)$ is strictly decreasing, we obtain \begin{align*} &2(4\pi)^{n/2}t^{n/2+1} E_n \( \psi (t)+d_h ,t \) \| h \|_{L^1} \lvert x-m_h \rvert \\ &\leq 2(4\pi)^{n/2}t^{n/2+1} \lvert \( \int_{\mathbb{R}^n} E_n (r,t) h(y) dy \) \( x -m_h \) \rvert \\ &\leq 2(4\pi)^{n/2}t^{n/2+1} \( \lvert \int_{\mathbb{R}^n} E_n (r,t) h(y) \( y -m_h \) dy \rvert + \lvert \nabla \tilde{J}_n(t) f(x) + e^{-t/2} \nabla \widetilde{\vect{W}}_n(t;f,g)(x) \rvert \) . \end{align*} Applying the third assertion in Lemma \ref{En_asymp} to the first term, there exists a constant $C$ such that, for any sufficiently large $t$, we have \[ 2(4\pi)^{n/2}t^{n/2+1} \lvert \int_{\mathbb{R}^n} E_n (r,t) h(y) \( y -m_h \) dy \rvert \leq C \frac{1+\psi (t)^2}{t} \| h \|_{L^1} . \] Applying Lemmas \ref{lem_est_Wn} and \ref{lem_tildeJ} to the second term, there exists a constant $C$ such that, for any sufficiently large $t$, we have \begin{align*} &2(4\pi)^{n/2}t^{n/2+1} \lvert \nabla \tilde{J}_n(t) f(x) + e^{-t/2} \nabla \widetilde{\vect{W}}_n(t;f,g)(x) \rvert \\ &\leq C \( \frac{1+\psi (t)}{t} \| f\|_{L^1} + e^{-t/2}t^{3n/2} \( \| f\|_{W^{n/2+1,\infty}} + \| g \|_{W^{n/2,\infty}} \) \) \end{align*} From the third assertion in Lemma \ref{En_asymp}, the function $2(4\pi)^{n/2}t^{n/2+1} E_n \( \psi (t)+d_h ,t \)$ is bounded from below with respect to $t$. Hence we obtain the conclusion. \end{proof} \begin{cor}\label{location_critical_small} We use Notation \ref{assump}. \begin{enumerate}[$(1)$] \item There exists a time $T \geq T_0 (\psi )$ such that, for any $t\geq T$, the intersection $\mathcal{C}(u(\cdot,t)) \cap ( CS(h) + \psi (t) B^n )$ is contained in the convex hull of the support of $h$. \item We have \[ \sup \left\{ \lvert x - m_h \rvert \lvert x\in \mathcal{C}\(u\( \cdot, t \)\) \cap \( CS(h) + \psi (t) B^n \) \right\} \right. = O \( \frac{1}{t} \) \] as $t$ goes to infinity. \end{enumerate} \end{cor} \begin{proof} (1) Since the centroid of $h$ is in the interior of the convex hull of $\supp h$, Lemma \ref{hp1} guarantees the conclusion. (2) From the first assertion, after a large time, we have \[ \sup \left\{ \lvert x - m_h \rvert \lvert x\in \mathcal{C}\(u\( \cdot, t \)\) \cap \( CS(h) + \psi (t) B^n \) \right\} \right. =\sup \left\{ \lvert x - m_h \rvert \lvert x\in \mathcal{C}\(u\( \cdot, t \)\) \cap CS(h) \right\} \right. . \] Applying Lemma \ref{hp1} to the case of $\psi =0$, we get the conclusion. \end{proof} \begin{lem}\label{nocritical We use Notation \ref{assump}. Suppose that the function $\psi (t)$ diverges as $t$ goes to infinity. If $\varphi (t) \geq \psi (t)$, then there exists a time $T \geq T_0 (\varphi )$ such that, for any $t \geq T$, the gradient of $u(\cdot ,t)$ does not vanish on the region $( CS(h) + \varphi (t) B^n ) \setminus (CS(h) + \psi (t) B^n )$. \end{lem} \begin{proof We give a proof for even dimensional cases. The other cases go parallel. From Proposition \ref{prop_de_fg}, the solution $u(x,t)$ is expressed as \[ u(x,t) = J_n(t) h(x) + \tilde{J}_n(t)f (x) + e^{-t/2} \widetilde{\vect{W}}_n(t;f,g)(x) . \] We remark that, from Lemma \ref{lem_est_Wn}, we have \[ \lvert \nabla \widetilde{\vect{W}}_n(t;f,g)(x) \rvert \leq C (1+t)^n \( \| f \|_{W^{n/2+1, \infty}} + \| g \|_{W^{n/2, \infty}} \). \] From Lemma \ref{list_J} and \ref{list_tildeJ}, we have \begin{align*} &\nabla J_n(t) h(x) + \nabla \tilde{J}_n(t) f(x) \\ &=-\frac{c_n e^{-t/2}}{2^{n+2}} \int_{B^n_t(x)} \left[ tk_{\frac{n}{2}+2} \( \frac{1}{2} \sqrt{t^2 -r^2} \) f(y) + k_{\frac{n}{2}+1} \( \frac{1}{2} \sqrt{t^2 -r^2} \) \( 4h(y) -2f(y) \) \right] (x-y) dy. \end{align*} Using the asymptotic expansions of $k_\ell(s)$ in Theorem \ref{thmdecom}, we have the expansion \begin{align*} \nabla J_n(t) h(x) + \nabla \tilde{J}_n(t) f(x) &=-\frac{c_n e^{-t/2}}{2^{n/2+1}} \int_{B^n_t(x)} \frac{1}{\( t^2 -r^2 \)^{n/4+1/2}} \exp \( \frac{\sqrt{t^2-r^2}}{2} \) \( 1+ O \( \frac{1}{\sqrt{t^2 -r^2}} \) \) \\ &\quad \times \( 2h(y) + \( \frac{t}{\sqrt{t^2 -r^2}} -1 \) f(y) \) (x-y) dy \end{align*} as $t$ goes to infinity. In order to complete the proof, we use the contradiction argument. For any natural number $N \geq T_0 (\varphi )$, we assume the existence of $t_N \geq N$ such that the function $u(\cdot ,t_N )$ has a critical point $x_N$ in the region $( CS(h) + \varphi (t) B^n ) \setminus (CS(h) + \psi (t) B^n )$. Let $r_N = \vert x_N -y \vert$. Since the unit sphere $S^{n-1}$ is compact, we may assume that the sequence $(x_N -m_h)/ \vert x_N -m_h\vert$ converges to a direction $\omega$ as $N$ goes to infinity. Then, we have \begin{align*} 0&= -\frac{2^{n/2+1}}{c_n} \exp \( \frac{t_N - \sqrt{t_N^2 -\lvert x_N -m_h \rvert^2}}{2} \) \frac{t_N^{n/2+1}}{\lvert x_N -m_h \rvert} \nabla u\( x_N ,t_N \) \\ &=\int_{B^n_{t_N} \( x_N \)} \exp \( \frac{\sqrt{t_N^2 -r_N^2} -\sqrt{t_N^2 -\lvert x_N -m_h \rvert^2}}{2} \) \frac{t_N^{n/2+1}}{\( t_N^2 -r_N^2 \)^{n/4+1/2}} \( 1+ O \( \frac{1}{\sqrt{t_N^2 -r_N^2}} \) \) \\ &\quad \times \( 2h(y) + \( \frac{t_N}{\sqrt{t_N^2 -r_N^2}} -1 \) f(y) \) \frac{x_N-y}{\lvert x_N -m_h \rvert} dy\\ &\quad -\frac{2^{n/2+1}}{c_n} \exp \( -\frac{\sqrt{t_N^2 -\lvert x_N -m_h \rvert^2}}{2} \) \frac{t_N^{n/2+1}}{\lvert x_N -m_h \rvert} \nabla \widetilde{\vect{W}}_n \( t_N ;f,g \) \( x_N \) \\ &\to \( 2 \int_{\mathbb{R}^n} h(y) dy \) \omega \end{align*} as $N$ goes to infinity, which contradicts to the non-negativity of $h$. \end{proof} \begin{cor}\label{location_critical We use Notation \ref{assump}. \begin{enumerate}[$(1)$] \item There exists a time $T \geq T_0 ( \varphi )$ such that, for any $t \geq T$, the intersection $\mathcal{C} ( u(\cdot ,t) ) \cap (CS(h) + \varphi (t)B^n )$ is contained in the convex hull of the support of $h$. \item We have \[ \sup \left\{ \lvert x-m_h\rvert \lvert x \in \mathcal{C} \( u(\cdot,t) \) \cap \( CS(h) +\varphi (t) B^n \) \right\} \right. = O \( \frac{1}{t} \) \] as $t$ goes infinity. \end{enumerate} \end{cor} \begin{proof} Corollary \ref{location_critical_small} and Lemma \ref{nocritical} guarantee the conclusion. \end{proof} \begin{lem}\label{expdecay We use Notation \ref{assump}. There exists a positive constant $C$ such that, for any $x \notin CS(h) + \varphi (t) B^n$ and $t>0$, we have \[ \lvert u(x,t) \rvert \le C \exp \( -\frac{\varphi (t)^2}{4t} \) \( \| h \|_{L^1} + \| f \|_{L^1} + \| f \|_{W^{* ,\infty}} + \| g \|_{W^{*, \infty}} \) . \] \end{lem} \begin{proof We give a proof for even dimensional cases. The other cases go parallel. From Proposition \ref{prop_de_fg}, we have \[ u(x,t) = J_n(t) h(x) + \tilde{J}_n(t)f (x) + e^{-t/2} \widetilde{\vect{W}}_n(t;f,g) (x) . \] Since the function $k_\ell$ is strictly increasing, for any $x \notin CS(h) + \varphi (t) B^n$, $y \in \supp h$ and $t >0$, we have \[ e^{-t/2} k_{\frac{n}{2}} \( \frac{1}{2} \sqrt{t^2 -r^2} \) \leq e^{-t/2} k_{\frac{n}{2}} \( \frac{1}{2} \sqrt{t^2 -\varphi (t)^2} \) . \] Using the asymptotic expansion of $k_\ell$ in Theorem \ref{thmdecom}, for any $t>0$, we have \[ e^{-t/2} k_{\frac{n}{2}} \( \frac{1}{2} \sqrt{t^2 -\varphi (t)^2} \) \leq C \exp \( \frac{-t+\sqrt{t^2 -\varphi (t)^2}}{2} \) \leq C \exp \( -\frac{\varphi (t)^2}{4t} \) . \] Hence, from Lemma \ref{list_J}, we can take a positive constant $C$ such that, for any $x \notin CS(h) + \varphi (t) B^n$ and $t>0$, we have \[ \lvert J_n (t) h (x) \rvert \leq C \exp \( -\frac{\varphi (t)^2}{4t} \) \| h\|_{L^1}. \] In the same manner, from Lemma \ref{list_tildeJ}, we have \[ \lvert \tilde{J}_n (t) f (x) \rvert \leq C \exp \( -\frac{\varphi (t)^2}{4t} \) \| f\|_{L^1}. \] Combining these estimates and Lemma \ref{lem_est_Wn}, we get the conclusion. \end{proof} \begin{lem}\label{estbelow We use Notation \ref{assump}. There exist a positive constant $C$ and a time $T \geq d_h$ such that, for any $t\ge T$, we have \[ \inf_{x\in CS(h)} u(x,t) \ge Ct^{-n/2} \| h \|_{L^\infty}. \] \end{lem} \begin{proof} We give a proof for even dimensional cases. The other cases go parallel. We remark that, from Lemmas \ref{lem_est_Wn} and \ref{lem_tildeJ}, we have the following estimates: \[ \lvert \widetilde{\vect{W}}_n(t;f,g)(x) \rvert \le C (1+t)^{n} \( \| f\|_{W^{n/2,\infty}} + \|g \|_{W^{n/2-1 ,\infty}} \) ,\ \lvert \tilde{J}_n(t) f(x) \rvert \le C (1+t)^{-n/2-1} \| f \|_{L^1}. \] Let us estimate the function $J_n(t) h(x)$. Since $h$ is non-negative, there is a point $\eta \in CS(h)$ such that, for any $y \in B_\rho^n(\eta )$, $h(y) \geq \| h \|_{L^\infty} /2$. Using the asymptotic expansion of $k_{n/2} (s)$ in Theorem \ref{thmdecom}, we have \begin{align*} &J_n(t)h(x) \\ &\ge \frac{c_n \| h \|_{L^{\infty}}}{2^{n-1}} e^{-t/2} \int_{B_{\rho}^n(\eta )} k_{\frac{n}{2}} \( \frac{1}{2}\sqrt{t^2-r^2} \)dy \\ &= \frac{c_n \| h \|_{L^{\infty}} }{2^{n/2-2}} \int_{B_{\rho}^n(\eta )} t^{-n/2} \( 1+ O \( \frac{1}{t^2} \) \) \( 1 + O\( \frac{1}{t} \) \) \( 1 + O\( \frac{1}{t} \) \) dy\\ &\ge C t^{-n/2} \| h \|_{L^\infty} \end{align*} for any sufficiently large $t$. Hence, for any sufficiently large $t$, we obtain \begin{align*} \lvert u(x,t) \rvert &\geq \lvert J_n(t) h(x) \rvert - \lvert \tilde{J}_n(t) f(x) \rvert -e^{-t/2} \lvert \widetilde{\vect{W}}_n(t;f,g)(x) \rvert \\ &\geq Ct^{-n/2}\| h \|_{L^\infty} -Ct^{-n/2-1}\| f \|_{L^1} -Ce^{-t/2} t^n \( \| f\|_{W^{n/2,\infty}} + \|g \|_{W^{n/2-1 ,\infty}} \) \\ &\geq Ct^{-n/2} \| h \|_{L^\infty} , \end{align*} which completes the proof. \end{proof} \begin{cor}\label{location_hotspot} We use Notation \ref{assump}. Suppose that $\exp ( -\varphi (t)^2 /(4t))$ is of small order of $t^{-n/2}$ as $t$ goes infinity. There exists a constant $T \geq T_0 (\varphi )$ such that, for any $t \geq T$, all of the time-delayed hot spots are contained in the parallel body $CS(h) + \varphi (t) B^n$. \end{cor} \begin{thm}\label{movement} Let $f$ and $g$ be as in $(fg)$ in Notation \ref{assump}. \begin{enumerate}[$(1)$] \item There exists a time $T \geq \max \{ d_h ,\ \delta (CS(h) , CS(f) )+ d_f \}$ such that, for any $t\geq T$, all of the time-delayed hot spots at time $t$ are contained in the convex hull of $h$. \item We have \[ \sup \left\{ \lvert x - m_h \rvert \lvert x \in \mathcal{H}(t) \right\} \right. = O \( \frac{1}{t} \) \] as $t$ goes to infinity. \end{enumerate} \end{thm} \begin{proof} We take a function $\varphi$ as in $(\varphi)$ in Notation \ref{assump} such that $\exp ( -\varphi (t)^2 /(4t))$ is of small order of $t^{-n/2}$ as $t$ goes to infinity. Then, Corollary \ref{location_hotspot} guarantees that all of the time-delayed hot spots are contained in the parallel body $CS(h)+\varphi (t)B^n$ after a large time. Hence, Corollary \ref{location_critical} implies the conclusion. \end{proof} \begin{rem} {\rm Let $g$ be a non-zero non-negative smooth function with compact support. If $n=1$ and $f=0$, then, for any $t\geq 0$, all of the time-delayed hot spots are contained in the convex hull of $\supp h= \supp g$. In other words, in this case, we can take $T=0$ in the first assertion of Theorem \ref{movement}. } \end{rem} \begin{proof} Fix a point $x$ in the complement of the convex hull of $\supp g$. Let $x'$ be the point that gives the distance between $x$ and the convex hull of $\supp g$. We have $\vert x-y \vert > \vert x' -y \vert$ for any $y$ in $\supp g$, and $[ x-t, x+t] \cap \supp g$ is contained in $[ x'-t, x'+t] \cap \supp g$. Hence the strictly increasing behavior of $I_0$ implies \begin{align*} S_1 (t) g(x) &= \frac{e^{-t/2}}{2} \int_{x-t}^{x+t} I_0 \( \frac{1}{2} \sqrt{t^2 -\lvert x-y \rvert^2} \) g(y) dy \\ &\leq \frac{e^{-t/2}}{2} \int_{x'-t}^{x'+t} I_0 \( \frac{1}{2} \sqrt{t^2 -\lvert x-y \rvert^2} \) g(y) dy \\ &< \frac{e^{-t/2}}{2} \int_{x'-t}^{x'+t} I_0 \( \frac{1}{2} \sqrt{t^2 -\lvert x'-y \rvert^2} \) g(y) dy \\ &=S_1 (t) g \( x' \) , \end{align*} which completes the proof. \end{proof} \subsection{Uniqueness of a time-delayed hot spot} In [JS], Jimbo and Sakaguchi showed that the set of hot spots $H_g(t)$ consists of one point for sufficiently large $t$. For the damped wave equation, let us show the corresponding result to [JS]. \begin{lem}\label{concavity We use Notation \ref{assump}. There exists a time $T \geq \max \{ d_h, \delta (CS (f) ,CS (h)) +d_f \}$ such that, for any $t\geq T$, the function $u(\cdot ,t)$ becomes strictly concave on the convex hull of the support of $h$. \end{lem} \begin{proof}\normalfon Let us give a proof for even dimensional cases. The other cases go parallel. In view of Proposition \ref{prop_de_fg}, we estimate the second derivatives of $J_n(t)h$, $\tilde{J}_n(t)f$ and $e^{-t/2} \widetilde{\vect{W}}_n(t;f,g)$. We fix a a point $x \in CS(h)$ and a direction $\omega \in S^{n-1}$. From the asymptotic expansion of $k_\ell$ in Theorem \ref{thmdecom}, there exists a positive constant $C$ such that, for any $y \in \supp h$ and $t \geq d_h$, we have \begin{align*} &e^{-t/2}\left[ -4k_{\frac{n}{2}+1} \( \frac{1}{2} \sqrt{t^2 -r^2} \) + k_{\frac{n}{2} +2} \( \frac{1}{2} \sqrt{t^2 -r^2} \) \( \omega \cdot (x-y ) \)^2 \right] \\ &= \frac{2^{n/2+2}}{\( t^2-r^2 \)^{n/4+1/2}} \exp \( \frac{-t+\sqrt{t^2 -r^2}}{2} \) \( -1 + O\( \frac{1}{t} \) \) \\ &\leq -C t^{-n/2-1} . \end{align*} Hence, from Lemma \ref{list_J}, there exists a positive constant $C$ such that, for any $x \in CS(h)$, $\omega \in S^{n-1}$ and $t \geq d_h$, we have \[ \( \omega \cdot \nabla \)^2 J_n (t) h(x) \leq -C t^{-n/2-1} \| h \|_{L^1}. \] In the same manner, from Lemma \ref{list_tildeJ}, we can obtain the existence of a positive constant $C$ such that, for any $x \in CS(h)\subset CS(f)+ \delta (CS(f) ,CS(h) ) B^n$, $\omega \in S^{n-1}$ and $t \geq \delta (CS(f) ,CS(h)) +d_f$, we have \[ \lvert \( \omega \cdot \nabla \)^2 \tilde{J}_n(t) f(x) \rvert \leq Ct^{-n/2 -2} \| f\|_{L^1}. \] On the other hand, from Lemma \ref{lem_est_Wn}, there exists a positive constant $C$ such that, for any $x \in CS(h)$, $\omega \in S^{n-1}$ and $t>0$, we have \[ e^{-t/2} \lvert \( \omega \cdot \nabla \)^2 \widetilde{\vect{W}}_n(t; f,g) (x) \rvert \leq Ce^{-t/2}(1+t)^{n} \( \| f \|_{W^{n/2+2,\infty}} + \| g \|_{W^{n/2+1 ,\infty}} \). \] Hence we obtain the strict concavity of $u(\cdot ,t)$ on $CS(h)$ for any sufficiently large $t$. \end{proof} \begin{prop}\label{uniqueness We use Notation \ref{assump}. There exists a time $T \geq \max \{ T_0 (\varphi), \delta (CS (f) ,CS (h)) +d_f \}$ such that, for any $t\ge T$, the set of critical points of $u(\cdot ,t)$ contained in the parallel body $CS(h) + \varphi (t) B^n$ consists of one-point. \end{prop} \begin{proof} From Corollary \ref{location_critical}, it is sufficient to show the strict concavity of the function $u(\cdot ,t)$ on the convex hull of the support of $h$. Hence Lemma \ref{concavity} guarantees the uniqueness of a critical point of $u(\cdot ,t)$. \end{proof} \subsection{Wave effect of the damped wave in view of time-delayed hot spots} In this subsection, we investigate the wave properties of the damped wave equation in view of the movement of time-delayed hot spots. We give some examples of initial data $(f,g)$ which allow time-delayed hot spots to escape from the convex hull of the support of $h:=f+g$ for some small time. \begin{ex}\label{ex1} {\rm Let $n=1$. We consider the equation \eqref{DWfg} with $g=0$. By Example \ref{decom1} and Proposition \ref{prop_de_fg}, we have \begin{align*} u(x,t) &=S_1(t)f(x)+ \tilde{J}_1(t) f(x) +e^{-t/2} \widehat{\vect{W}}_n(t) f(x) \\ &=\frac{e^{-t/2}}{4} \int_{x-t}^{x+t} \( \frac{t}{\sqrt{t^2 -r^2}} I_1 \( \frac{1}{2} \sqrt{t^2 -r^2} \) +I_0 \( \frac{1}{2} \sqrt{t^2 -r^2} \) \) f(y) dy \\ &\quad +\frac{e^{-t/2}}{2}\( f(x+t)+f(x-t)\) \end{align*} Let us give an example such that if the initial datum $f$ has a sufficiently large maximum value and a constant $L^1$ norm, then, for some $t$, time-delayed hot spots escape from the convex hull of the support of $f$. Let $\rho$ be a non-negative smooth function satisfying $\supp \rho = [-2, 2]$, $\| \rho \|_{L^1}=1$ and $\rho(y)\ge \| \rho \|_{L^{\infty}}/2$ for $-1 \leq y \leq 1$. For example, if we normalize the function \[ \tilde{\rho}(y) = \begin{cases} \displaystyle \exp \( -\frac{1}{4-y^2} \) &(-2 \leq y \leq 2),\\ 0 &({\rm otherwise}), \end{cases} \] then the normalized function $\tilde{\rho}/\| \tilde{\rho}\|_{L^1}$ satisfies the conditions of $\rho$. We define $f_\varepsilon (y)=\rho(y/\varepsilon)/\varepsilon$ with a small parameter $\varepsilon>0$. Then we have $f_\varepsilon (x+t)\geq \|\rho\|_{L^{\infty}}/(2\varepsilon )$ for $-t-\varepsilon \leq x \leq -t+\varepsilon$ and $f_\varepsilon (x-t) \geq \|\rho\|_{L^{\infty}}/(2\varepsilon)$ for $t-\varepsilon \leq x \leq t+\varepsilon$. On the other hand, noting $\| f_\varepsilon \|_{L^1}=1$, we can choose a constant $C$ independent of $\varepsilon$ such that we have \[ \| S_1 (t)f+ \tilde{J}_1(t) f \|_{L^\infty} \leq C e^{-t/2}\( I_0 \(\frac{t}{2} \) + (1+t) \( 1+ I_1 \( \frac{t}{2} \) \) \) . \] Using the facts $I_0(0)=1$ and $I_1(0)=0$, we can take the parameter $\varepsilon$ sufficiently small so that there is a time $t\geq 4\varepsilon$ satisfying the inequality \[ \frac{e^{-t/2}}{2}\frac{\| f_\varepsilon \|_{L^\infty}}{2} =\frac{e^{-t/2}}{2}\frac{\| \rho \|_{L^\infty}}{2\varepsilon} > Ce^{-t/2} \( I_0\(\frac{t}{2} \)+ (1+t) \( 1+ I_1 \( \frac{t}{2}\) \)\) . \] Hence if $t\geq 4\varepsilon$ satisfies the above inequality and $x \in [t-\varepsilon, t+\varepsilon]\cup [-t-\varepsilon, -t+\varepsilon]$ then we have $u(x,t) > u(\xi ,t)$ for any $\xi \in \supp f_\varepsilon$. } \end{ex} \begin{ex}\label{ex2} {\rm Let $n=2$. We consider the damped wave equation \eqref{DWfg} with $f=0$. By Proposition \ref{propsol}, we have \[ u(x,t)=S_2(t)g(x) =\frac{e^{-t/2}}{2\pi}\int_{B_t^2(x)} \frac{\cosh(\frac{1}{2}\sqrt{t^2-r^2})}{\sqrt{t^2-r^2}}g(y)dy. \] Let us show that, if we choose a clever initial datum $g$, then, for some $t$, $S_2(t)g$ has a (non-trivial) critical point in the complement of the convex hull of $\supp g$. Let $s_*$ be the unique critical point of the function $\cosh(s/2)/s$. Direct computation shows $2<s_*<3$. Fix a small positive parameter $\varepsilon$ with $2 \varepsilon < s_*$. Then, we have $s_* < (s_*^2 +4\varepsilon^2)/(4\varepsilon)$. Let $g_\varepsilon$ be a non-zero non-negative radially symmetric smooth function with support $B_{\varepsilon}^2(0)$. Let us show that, for any $s_* \leq t \leq (s_*^2 +4\varepsilon^2)/(4\varepsilon)$, the function $S_2(t)g$ has a critical point in the complement of the disk of radius $\sqrt{t^2 -s_*^2}+\varepsilon$ centered at origin. Fix a point $x$ with $\vert x \vert = \sqrt{t^2 -s_*^2}+\varepsilon$. We remark that, for any $y \in B_\varepsilon^2 (0)$, we have the following inequalities: \[ \lvert x-y \rvert < t ,\ \sqrt{t^2 - \lvert x -y \rvert^2} \leq s_* . \] Let \[ \delta = \frac{t- \sqrt{t^2 -s_*^2}-2\varepsilon}{2} \frac{1}{\sqrt{t^2 -s_*^2} +\varepsilon}. \] Then, for the point $x':= (1+\delta )x$ and any $y \in B_\varepsilon^2(0)$, we have the following inequalities: \[ \lvert x'-y \rvert <t,\ \sqrt{t^2 - \lvert x' -y \rvert^2} \leq s_*. \] Therefore, we have \begin{align*} S_2(t)g_\varepsilon (x)&=\frac{e^{-t/2}}{2\pi}\int_{B_t^2(x)} \frac{\cosh \( \frac{1}{2}\sqrt{t^2-|x-y|^2} \)}{\sqrt{t^2-|x-y|^2}}g_\varepsilon(y)dy\\ &< \frac{e^{-t/2}}{2\pi}\int_{B_t^2(x')} \frac{\cosh \( \frac{1}{2}\sqrt{t^2-\lvert x' -y\rvert^2} \)}{\sqrt{t^2-\lvert x'-y\rvert^2}}g_\varepsilon (y)dy\\ &=S_2(t)g_\varepsilon \( x'\) . \end{align*} Thanks to the compactness of the support of $S_2(t)g_\varepsilon$, for each direction $\omega \in S^1$, we get the existence of a maximal point of the function \[ \( \sqrt{t^2 -s_*^2}+\varepsilon ,t+\varepsilon \) \ni \rho \mapsto S_2(t) g_\varepsilon (\rho \omega ) \in \mathbb{R}. \] Since the function $g_\varepsilon$ is radially symmetric, the function $S_2(t) g_\varepsilon$ so is, and we obtain the existence of a critical point of $S_2(t) g_\varepsilon$ in the complement of the disk of radius $\sqrt{t^2 -s_*^2}+\varepsilon$ centered at origin. } \end{ex} \begin{ex}\label{ex3} {\rm Let $n=2$. We consider the damped wave equation \eqref{DWfg} with $f=0$ again. Let $g$ be a non-zero non-negative smooth function with compact support. Suppose $2 d_g <s_*$. Let us show that, for any $2d_g \leq t \leq s_*$, there exists a point $x$ in the complement of $CS(g)$ such that, for any point $\xi \in CS( g)$, we have $S_2(t) g(\xi ) < S_2 (t) g(x)$. In other words, if $g$ has a small support so that $d_g <s_* /2$, then, for any $2d_g \leq t \leq s_*$, time-delayed hot spots escape from the convex hull of the support of $g$. Fix an arbitrary time $2d_g \leq t \leq s_*$. We can choose a point $x \in CS( g)^c$ which satisfies the following conditions: \[ \max_{y\in \supp g} \lvert x-y \rvert =t,\ \min_{y \in \supp g} \lvert x-y \rvert \geq t-d_g . \] For such a point $x$, any $\xi \in CS( g)$ and $y \in \supp g$, the assumption of $t$ implies \[ 0 \leq \sqrt{t^2 -\lvert x-y \rvert^2} \leq \sqrt{\(2t-d_g\)d_g} \leq \sqrt{t^2 -d_g^2} \leq \sqrt{t^2 -\lvert \xi -y \rvert^2} \leq t, \] and the strictly decreasing behavior of $\cosh(s/2)/s$ for $0<s<s_*$ implies \[ \frac{\cosh \( \frac{1}{2} \sqrt{t^2 -\lvert x-y \rvert^2}\)}{\sqrt{t^2 -\lvert x-y \rvert^2}} \geq \frac{\cosh \( \frac{1}{2} \sqrt{t^2 -\lvert \xi-y \rvert^2}\)}{\sqrt{t^2 -\lvert \xi-y \rvert^2}}. \] Hence we obtain $S_2(t) g(x) > S_2(t) g(\xi )$ for any $\xi \in CS( g)$. } \end{ex} \begin{ex}\label{ex4 {\rm Let $n=3$. We consider the damped wave equation \eqref{DWfg} with $f=0$. By Example \ref{decom23}, we have \[ S_3(t)g =J_3(t)g +e^{-t/2}W_3(t)g . \] Let us give an example such that if the initial datum $g$ has a sufficiently large maximum value and a constant $L^1$ norm, then, for some $t$, hot spots escape from the convex hull of the support of $g$. Let $\rho$ be non-zero non-negative smooth function satisfying $\supp \rho = B_2^3(0)$, $\| \rho \|_{L^1}=1$ and $\rho (y)\ge \| \rho \|_{L^\infty}/2$ on the unit ball $B^3$. We define $g_\varepsilon (y)=\rho \( y/\varepsilon \) /\varepsilon^3$ with a small parameter $\varepsilon >0$. If $t>2\varepsilon$ and $x \in S_t^2(0)$, then we have \[ \sigma_2\( S_t^2(x) \cap B^3_{\varepsilon}(0) \) =\pi \varepsilon^2, \] and then, we get \[ e^{-t/2}W_3(t)g_\varepsilon(x) =\frac{e^{-t/2}}{4\pi t}\int_{S_t^2(x)}g_\varepsilon (y)d\sigma_2(y) \ge \frac{\varepsilon^2 e^{-t/2}}{8t}\| g_\varepsilon\|_{L^\infty} =\frac{e^{-t/2}}{8\varepsilon t} \| \rho \|_{L^\infty}. \] On the other hand, as we will see in \eqref{4_lem2_eq1}, $J_3(t)g_{\varepsilon}$ is estimated by \[ \| J_3(t)g_\varepsilon \|_{L^\infty} \leq C(1+t)^{-3/2}\|g_\varepsilon \|_{L^1} =C(1+t)^{-3/2}, \] where $C$ is independent of $\varepsilon$. We can take the parameter $\varepsilon$ sufficiently small so that there is a time $t\geq 4\varepsilon$ satisfying the inequality \[ \frac{e^{-t/2}}{8\varepsilon t}\| \rho \|_{L^\infty} >C(1+t)^{-3/2} . \] If $t \geq 4\varepsilon$ satisfies the above inequality, then, for any $x \in S_t^2(0)$ and $\xi \in \supp g_\varepsilon =B_{2\varepsilon}^3(0)$, we have \[ S_3(t) g_\varepsilon (x) \geq e^{-t/2}W_3(t)g_\varepsilon (x) > J_3(t)g_\varepsilon (\xi ) = S_3(t) g_\varepsilon (\xi), \] that is, time-delayed hot spots are not in (the convex hull of) the support of $g_\varepsilon$. } \end{ex} \section{Application of the Nishihara decomposition: {\boldmath $L^p$-$L^q$} estimates} In this section, as an application of Theorem \ref{thmdecom}, we give $L^p$-$L^q$ estimates for the solution of the damped wave equation \eqref{DWfg}. In [HO, MN, Nis], when $n\leq 3$, the following $L^p$-$L^q$ estimates were shown: \begin{equation} \label{4_LpLq} \left\| u(\cdot ,t)-P_n(t) \( f+g \) -e^{-t/2}\widetilde{\vect{W}}_n(t;f,g)\right\|_{L^p} \leq Ct^{-\frac{n}{2} \( \frac{1}{q}-\frac{1}{p} \) -1} \( \|f\|_{L^q}+\|g\|_{L^q} \), \end{equation} where $t>0$ and $1\leq q\leq p\leq \infty$. In [Nar], when $n\ge 4$, Narazaki showed the following estimates: \begin{align} \label{4_Na} \left\|\mathcal{F}^{-1}\left[ \(\hat{u}(\cdot ,t)-\hat{v}(\cdot ,t)\) \chi \right] \right\|_{L^p} \le C \( 1+t \)^{-\frac{n}{2} \( \frac{1}{q}-\frac{1}{p} \) -1+\varepsilon} \( \|f\|_{L^q}+\|g\|_{L^q} \), \end{align} where $1\leq q\leq p\leq \infty$, $\varepsilon$ is an arbitrary small positive number, $C=C(n,p,q,\varepsilon)$ is a positive constant, $\chi$ is a compactly supported radially symmetric smooth function satisfying $\chi=1$ near the origin, $v(x,t)=P_n(t)(f+g)(x)$, $\hat{u}$ and $\hat{v}$ denote the Fourier transform of $u$ and $v$, respectively, and $\mathcal{F}^{-1}$ is the inverse Fourier transform. Moreover, in the case where $1<q<p<\infty$, $(p,q)=(2,2)$ or $(p,q)=(\infty, 1)$, we may take $\varepsilon=0$, that is, we have \begin{equation} \label{Nar_est2} \left\| \mathcal{F}^{-1}\left[ \( 1-\chi \) \( \hat{u}(\cdot, t) -e^{-t/2} \( \vect{M}_0(\cdot ,t )\hat{f}(\cdot ,t)+\vect{M}_1(\cdot ,t) \hat{g}(\cdot ,t) \) \) \right] \right\|_{L^p} \leq Ce^{-\delta t} \|g\|_{L^q} \end{equation} for some $\delta>0$, where $1<q\leq p<\infty$, $C=C(n,p,q)$ is a positive constant, and \begin{align} \nonumber \vect{M}_1(\xi ,t) &= \frac{1}{\sqrt{|\xi|^2-1/4}}\( \sin \( t|\xi| \) \sum_{0\le k<(n-1)/4}\frac{(-1)^k}{(2k)!}t^{2k}\Theta(\xi)^{2k} \right. \\ &\quad-\left. \cos \( t|\xi|\) \sum_{0\le k<(n-3)/4}\frac{(-1)^k}{(2k+1)!} t^{2k+1}\Theta(\xi)^{2k+1}\),\\ \nonumber \vect{M}_0(\xi ,t) &=\cos \( t|\xi| \) \sum_{0\le k<(n+1)/4}\frac{(-1)^k}{(2k)!}t^{2k}\Theta(\xi)^{2k}\\ &\quad +\sin \( t|\xi| \) \sum_{0\le k<(n-1)/4}\frac{(-1)^k}{(2k+1)!} t^{2k+1}\Theta(\xi)^{2k+1}+\frac{1}{2}\vect{M}_1(\xi ,t) \end{align} with $\Theta(\xi)=|\xi|-\sqrt{|\xi|^2-1/4}$. The aim of this section is to remove the $\varepsilon$ in the estimate \eqref{4_Na} and the restriction $q\neq 1$ and $p \neq \infty$. \begin{thm}\label{4_thm2 Let $1\leq q \leq p \leq \infty$. Assume that the initial data $f$ and $g$ are $L^q$-integrable smooth functions. Let $u$ be the solution to \eqref{DWfg}. There exists a positive constant $C$ such that, for any $t>0$, we have \[ \left\| u(\cdot, t)-P_n(t) \( f+g \) -e^{-t/2}\widetilde{\vect{W}}_n \( t;f,g\) \right\|_{L^p} \leq Ct^{-\frac{n}{2} \( \frac{1}{q}-\frac{1}{p} \) -1} \( \|f\|_{L^q}+\|g\|_{L^q} \) . \] \end{thm} The proof of this theorem is almost same as in [Nis] (see also [INZ]). We note that the one-dimensional case has already been proved by Marcati and Nishihara in [MN], and we give a proof only for the case $n\ge 2$. By Proposition \ref{prop_de_fg}, we have \[ u(\cdot ,t)-P_n(t) \( f+g \) -e^{-t/2}\widetilde{\vect{W}}_n \( t;f,g\) =J_n(t) \( f+g \) -P_n(t) \( f+g \) +\tilde{J}_n(t)f. \] Therefore, the proof is reduced to the following estimates: \begin{lem}\label{4_lem2 Let $1\le q\le p\le \infty$, and $g$ an $L^q$-integrable smooth function. There exists a constant $C>0$ such that, for any $t >0$, we have the following inequalities: \begin{align} \label{4_lem2_eq1} \left\| J_n(t)g \right\|_{L^p} &\le C \( 1+t \)^{-\frac{n}{2}\(\frac{1}{q}-\frac{1}{p}\)}\|g\|_{L^q},\\ \label{4_lem2_eq2} \left\| \tilde{J}_n(t)g \right\|_{L^p} &\le C(1+t)^{-\frac{n}{2}\(\frac{1}{q}-\frac{1}{p}\)-1}\|g\|_{L^q} ,\\ \label{4_lem2_eq3} \left\| J_n(t)g-P_n(t)g \right\|_{L^p} &\leq Ct^{-\frac{n}{2}\(\frac{1}{q}-\frac{1}{p}\)-1}\|g\|_{L^q} . \end{align} \end{lem} \begin{proof} Let us give a proof for higher odd dimensional cases. Even dimensional cases go parallel. We first show \eqref{4_lem2_eq1} and \eqref{4_lem2_eq3}. We assume $t\ge 1$ and write $\tilde{c}_n = 2^{-(n-1)}c_n$. For a constant $0 < \varepsilon <1/2$, put \begin{align*} X_1&=\int_{t^{(1+\varepsilon)/2}B^n(x)} \( \tilde{c}_n e^{-t/2} k_{\frac{n-1}{2}} \( \frac{1}{2} \sqrt{t^2-r^2} \) -\frac{e^{-r^2/(4t)}}{(4\pi t)^{n/2}} \)g(y)dy,\\ X_2&=\int_{B_t^n(x)\setminus t^{(1+\varepsilon)/2}B^n(x)} \( \tilde{c}_n e^{-t/2} k_{\frac{n-1}{2}} \(\frac{1}{2}\sqrt{t^2-r^2}\) -\frac{e^{-r^2/(4t)}}{(4\pi t)^{n/2}} \)g(y)dy,\\ X_3&=\int_{B_t^n(x)^c}\frac{e^{-r^2/(4t)}}{(4\pi t)^{n/2}}g(y)dy. \end{align*} Then we have \[ J_n(t)g(x) -P_n(t)g (x) =X_1+X_2+X_3. \] By the Hausdorff-Young inequality ([GGS, p. 142]), we estimate the integral $X_3$ as \[ \|X_3\|_{L^p} \le \(\int_{B_t^n(0)^c}\frac{e^{-\rho|y|^2/(4t)}}{(4\pi t)^{\rho n/2}}dy\)^{1/\rho} \|g\|_{L^q} \le e^{-t/8}\|g\|_{L^q}, \] where $\rho$ is determined by the relation $1/q-1/p=1-1/\rho$. In the same manner as in the above estimate, we can obtain \[ \|X_2\|_{L^p} \le e^{-ct^\varepsilon}\|g\|_{L^q} \] with some constant $c>0$. Let us estimate the integral $X_1$. By the asymptotic expansion in Theorem \ref{thmdecom}, we have \begin{align*} &\tilde{c}_n e^{-t/2} k_{\frac{n-1}{2}}\(\frac{1}{2}\sqrt{t^2-r^2}\)\\ &=\frac{1}{(4\pi)^{n/2}}\frac{1}{\( t^2-r^2 \)^{n/4}} \exp\( \frac{-t+ \sqrt{t^2-r^2}}{2} \) \(1-\frac{n(n-2)}{4\sqrt{t^2-r^2}}+O\(\frac{1}{t^2-r^2}\)\). \end{align*} Therefore, we obtain \[ X_1=\frac{1}{(4\pi t)^{n/2}} \int_{t^{(1+\varepsilon)/2}B^n(x)}e^{-r^2/(4t)}F(r,t)g(y)dy, \] where \[ F(r,t)=\exp\( \frac{r^2}{4t}+\frac{-t+\sqrt{t^2-r^2}}{2} \) \(\frac{t}{\sqrt{t^2-r^2}}\)^{n/2} \(1-\frac{n(n-2)}{4\sqrt{t^2-r^2}} +O\(\frac{1}{t^2-r^2}\)\) -1. \] Hence we have \[ \|X_1\|_{L^p}\le \frac{C}{t^{n/2}} \(\int_{t^{(1+\varepsilon)/2}B^n} e^{-\rho r^2/(4t)} \lvert F \( |y|,t \) \rvert^{\rho} dy\)^{1/\rho} \|g\|_{L^q} \] with $1/q-1/p=1-1/\rho$. Asymptotic expansions \eqref{app_taylor1} and \eqref{app_taylor3} imply \begin{align*} F \( |y|,t \) &=\(1+\frac{1}{t}O\(\frac{|y|^4}{t^2}\)\) \(1+\frac{1}{t}O\(\frac{|y|^2}{t}\)\)^{n/2} \(1+ O \( \frac{1}{t} \) +\frac{1}{t} O\( \frac{\lvert y \rvert^2}{t}\) \)-1\\ &=\frac{1}{t}O\(1+\frac{|y|^2}{t}+\cdots+\(\frac{|y|^2}{t}\)^N\) \end{align*} for some large integer $N$. Consequently, we obtain \begin{align*} \|X_1\|_{L^p}&\le \frac{C}{t^{n/2+1}} \(\int_{t^{(1+\varepsilon)/2}B^n(0)}e^{-\rho |y|^2/(4t)} \(1+\frac{|y|^2}{t}+\cdots+\(\frac{|y|^2}{t}\)^N\)^{\rho} dy\)^{1/\rho} \|g\|_{L^q}\\ &\le \frac{C}{t^{n/2+1}}t^{n/(2\rho)} \(\int_{\mathbb{R}^n}e^{-\rho |z|^2} \( 1+|z|^2+\cdots+|z|^{2N} \)^{\rho} dz \)^{1/\rho} \|g\|_{L^q}\\ &\le Ct^{-\frac{n}{2} \( \frac{1}{q}-\frac{1}{p} \) -1}\|g\|_{L^q}, \end{align*} which implies the estimate \eqref{4_lem2_eq3} for $t\ge 1$. Moreover, we recall the well-known fact \[ \|P_n(t)g\|_{L^p}\le Ct^{-\frac{n}{2}(\frac{1}{q}-\frac{1}{p})}\|g\|_{L^q},\ t>0 \] (see [GGS, p. 8]). Using this fact, we have \[ \left\| J_n(t)g \right\|_{L^p}\leq \left\| J_n(t)g-P_n(t)g \right\|_{L^p} + \left\| P_n(t)g \right\|_{L^p} \le Ct^{-\frac{n}{2} \( \frac{1}{q}-\frac{1}{p} \) }\|g\|_{L^q}, \] which implies \eqref{4_lem2_eq1} for $t\ge 1$. The estimates \eqref{4_lem2_eq1} and \eqref{4_lem2_eq3} for $0\le t<1$ are easy, and we omit the proof. Next, we show the estimate \eqref{4_lem2_eq2}. We assume $t\ge 1$. For a constant $0<\varepsilon <1/2$, put \begin{align*} X_4&=\tilde{c}_n\int_{B_t^n(x)\setminus t^{(1+\varepsilon)/2}B^n(x)} \frac{\partial}{\partial t}\(e^{-t/2}k_{\frac{n-1}{2}}\(\frac{1}{2}\sqrt{t^2-r^2}\)\) g(y)dy,\\ X_5&=\tilde{c}_n\int_{t^{(1+\varepsilon)/2}B^n(x)} \frac{\partial}{\partial t} \( e^{-t/2}k_{\frac{n-1}{2}} \( \frac{1}{2}\sqrt{t^2-r^2}\)\)g(y)dy. \end{align*} Then we have \[ \tilde{J}_n(t)g(x) =X_4+X_5. \] Since $k_{\ell+1}(s)=k'_{\ell}(s)/s$ leads to \begin{align*} \frac{\partial}{\partial t} \( e^{-t/2}k_{\frac{n-1}{2}}\(\frac{1}{2}\sqrt{t^2-r^2}\)\) &=e^{-t/2}\left[ -\frac{1}{2}k_{\frac{n-1}{2}}\(\frac{1}{2}\sqrt{t^2-r^2}\) +\frac{t}{2\sqrt{t^2-r^2}} k_{\frac{n-1}{2}}' \(\frac{1}{2}\sqrt{t^2-r^2}\)\right] \\ &=e^{-t/2}\left[ -\frac{1}{2}k_{\frac{n-1}{2}}\(\frac{1}{2}\sqrt{t^2-r^2}\) +\frac{t}{4}k_{\frac{n+1}{2}}\(\frac{1}{2}\sqrt{t^2-r^2}\) \right], \end{align*} in the same manner as the estimate of $X_2$, we can obtain \[ \|X_4\|_{L^p}\leq Ce^{-ct^\varepsilon}\|g\|_{L^q} \] with some constant $c>0$. Now we turn to the estimate for $X_5$. By using the asymptotic expansion of $k_\ell$, \eqref{app_taylor1} and \eqref{app_taylor2} again, we have \begin{align*} &\frac{\partial}{\partial t}\(e^{-t/2}k_{\frac{n-1}{2}}\(\frac{1}{2}\sqrt{t^2-r^2}\)\)\\ &=\frac{2^{(n-3)/2}}{\sqrt{\pi}} e^{-r^2/(4t)} \exp\( \frac{r^2}{4t}+\frac{-t+\sqrt{t^2-r^2}}{2} \) (t^2-r^2)^{-n/4}\(\frac{t}{\sqrt{t^2-r^2}}-1\) \( 1+O\( \frac{1}{t} \) \) \\ &=\frac{2^{(n-3)/2}}{\sqrt{\pi}} e^{-r^2/(4t)} \( 1+\frac{1}{t}O\(\frac{r^2}{t}\)\) t^{-n/2}\( 1+\frac{1}{t}O\(\frac{r^2}{t}\)\)^{n/2} \frac{1}{t}O\(\frac{r^2}{t}\) \(1+O\(\frac{1}{t}\)\)\\ &\leq Ct^{-n/2-1}e^{-r^2(4t)} \( 1+\frac{r^2}{t}+\cdots+\(\frac{r^2}{t}\)^N\) \end{align*} on the ball $t^{(1+\varepsilon)/2}B^n(x)$ with some large integer $N$. Consequently, we obtain \begin{align*} \|X_5\|_{L^p}&\le Ct^{-\frac{n}{2}-1} \(\int_{t^{(1+\varepsilon)/2}B^n}e^{-\rho |y|^2/(4t)} \(1+\frac{|y|^2}{t}+\cdots+\(\frac{|y|^2}{t}\)^N\)^{\rho} dy \)^{1/\rho} \|g\|_{L^q}\\ &\le Ct^{-\frac{n}{2}-1-\frac{n}{2\rho}} \(\int_{\mathbb{R}^n} e^{-\rho |z|^2/4} \( 1+|z|^2+\cdots+|z|^{2N} \)^{\rho} dz \)^{1/\rho} \|g\|_{L^q}\\ &\le Ct^{-\frac{n}{2}\(\frac{1}{q}-\frac{1}{p}\)-1}\|g\|_{L^q} \end{align*} with $1/q-1/p=1-1/\rho$, which implies \eqref{4_lem2_eq2} for $t\ge 1$. The estimate \eqref{4_lem2_eq2} for $0\le t<1$ is easy, and we omit the proof. \end{proof} \section{Appendices} \subsection{Proofs of preliminary estimates} \begin{proof}[Proof of Lemma \ref{lem_est_Wn} Let us give a proof for even dimensional cases. The other cases go parallel. Changing the variable as $y=x+tz$ with $z \in B^n$, we have \[ \int_{B_t^n(x)} \frac{1}{\sqrt{t^2-r^2}} g(y) dy = t^{n-1} \int_{B^n} \frac{1}{\sqrt{1-\lvert z \rvert^2}} g(x+tz) dz. \] When we estimate the function \[ \vect{W}_n(t)g(x) =2c_n \sum_{j=0}^{(n-2)/2} \frac{1}{8^j j !}\( \frac{1}{t} \frac{\partial}{\partial t} \)^{(n-2)/2-j} \( t^{n-1} \int_{B^n} \frac{1}{\sqrt{1-\lvert z \rvert^2}} g(x+tz) dz \) , \] the worst term with respect to the growth order of $t$ is given by $j= (n-2)/2$. We can bound it above by $C(1+t)^{n-1} \| g \|_{L^\infty}$. Furthermore, we can bound the term of $j =0$ above by $C (1+t)^{n/2} \| g \|_{W^{n/2-1 ,\infty}}$. The other terms are bounded above by these two quantities (up to a constant multiple). Hence we obtain the estimate for $\vect{W}_n(t)g(x)$. From Proposition \ref{prop_de_fg}, we have \[ \widehat{\vect{W}}_n(t) f(x) = \frac{c_n t^n}{2^{\frac{3n-2}{2}}\(\frac{n}{2} \) !} \int_{B^n} \frac{1}{\sqrt{1-\lvert z \rvert^2}} f(x+tz) dz, \] which implies the estimate for $\widehat{\vect{W}}_n(t) f(x)$. Also, we have \[ \widetilde{\vect{W}}_n(t;f,g) (x) = \frac{1}{2} \vect{W}_n (t) f(x) + \vect{W}_n (t) g(x) + \widehat{\vect{W}}_n(t) f(x) + \frac{\partial}{\partial t} \vect{W}_n(t)f(x). \] Combining the above estimates for $\vect{W}_n(t)g(x)$ and $\widehat{\vect{W}}_n(t) f(x)$, we obtain the conclusion. \end{proof \begin{proof}[Proof of Lemma \ref{list_J}] Using Remark \ref{recursion}, integration by parts implies the identities. \end{proof} \begin{proof}[Proof of Lemma \ref{En}] If $\dist (x, CS (h) ) \leq t-d_h$ and $t\geq d_h$, then the intersection $S^{n-1}_t(x) \cap CS(h)$ is a null set with respect to the $(n-1)$-dimensional spherical Lebesgue measure. Hence Lemma \ref{list_J} guarantees the conclusion. \end{proof} \begin{proof}[Proof of Lemma \ref{En_asymp}] (1) We give a proof for even dimensional cases. The other cases go parallel. We remark that, from the definition of $c_n$ \eqref{cn}, we have \[ E_n (r,t) = \frac{e^{-t/2}}{2^{3n/2+1}\pi^{n/2}} k_{\frac{n}{2}+1}\( \frac{1}{2}\sqrt{t^2-r^2} \) . \] From Theorem \ref{thmdecom}, we have \[ k_{\frac{n}{2}+1} \( \frac{1}{2} \sqrt{t^2 -\varphi (t)^2} \) = \frac{1}{2} \( \frac{2}{\sqrt{t^2 -\varphi (t)^2}} \)^{n/2+1} \exp \( \frac{\sqrt{t^2 -\varphi (t)^2}}{2} \) \( 1+ O \( \frac{1}{\sqrt{t^2 -\varphi (t)^2}} \) \) \] as $t$ goes to infinity, which implies the conclusion. (2) Applying the fact \eqref{app_taylor1} to the first assertion, we obtain the conclusion. (3) Applying the fact \eqref{app_taylor2} to the second assertion, we obtain the conclusion. \end{proof} \begin{proof}[Proof of Lemma \ref{list_tildeJ}] Using Remark \ref{recursion}, integration by parts implies the identities. \end{proof} \begin{proof}[Proof of Lemma \ref{lem_tildeJ}] (1) This is a direct consequence of \eqref{4_lem2_eq2} in Lemma \ref{4_lem2}. (2) We give a proof for even dimensional cases. The other cases go parallel. From the asymptotic expansion in Theorem \ref{thmdecom}, we have \begin{align*} &e^{-t/2} \( t k_{\frac{n}{2}+2} \( \frac{1}{2} \sqrt{t^2 -r^2} \) -2 k_{\frac{n}{2}+1} \( \frac{1}{2} \sqrt{t^2 -r^2} \) \) \\ &=\frac{2^{n/2+1}}{\( t^2 -r^2 \)^{n/4+1/2}} \exp \( \frac{-t+\sqrt{t^2 -r^2}}{2} \) \\ &\quad \times \left[ \frac{t}{\sqrt{t^2 -r^2}} \( 1+ O \( \frac{1}{\sqrt{t^2-r^2}} \) \) - \( 1+ O \( \frac{1}{\sqrt{t^2-r^2}} \) \) \right] . \end{align*} Using the facts \eqref{app_taylor1} and \eqref{app_taylor2}, the above expansion coincides with \[ 2^{n/2+1} t^{-n/2-1} \( O \( \frac{1}{t} \) + O \( \frac{\psi (t)^2}{t^2} \) \) , \] which implies the conclusion. (3) Using the asymptotic expansion in Theorem \ref{thmdecom}, in the same manner as in the second assertion, we obtain the conclusion. \end{proof} \subsection{Frequently used Taylor's expansions} Let us list up frequently used Taylor's expansions: \begin{itemize} \item As $s$ tends to zero, we have \begin{equation} \label{app_taylor} (1-s^2)^\alpha = 1 - \alpha s^2 + \frac{\alpha (\alpha-1)}{2} s^4 + O \( s^6 \) . \end{equation} \item If a function $\varphi (t)$ is of small order of $t$ as $t$ goes to infinity, then we have \begin{equation} \label{app_taylor1} \( t^2-\varphi (t)^2 \)^\alpha = t^{2\alpha} \( 1 -\alpha \( \frac{\varphi (t)}{t} \)^2 + \frac{\alpha (\alpha-1)}{2} \( \frac{\varphi (t)}{t} \)^4 + O\( \( \frac{\varphi (t)}{t} \)^6 \) \) \end{equation} as $t$ goes to infinity. \item If a function $\varphi (t)$ is of small order of $\sqrt{t}$ as $t$ goes to infinity, then, as $t$ goes to infinity, we have the following expansions: \begin{align} \label{app_taylor2} &\exp\( \frac{-t+\sqrt{t^2-\varphi (t)^2}}{2} \) = 1 - \frac{\varphi (t)^2}{4t} + O\( \frac{\varphi (t)^4}{t^2} \) ,\\ \label{app_taylor3} &\exp\( \frac{\varphi(t)^2}{4t}+\frac{-t+\sqrt{t^2-\varphi (t)^2}}{2} \) = 1 + \frac{1}{t}O\( \frac{\varphi (t)^4}{t^2} \). \end{align} \end{itemize} \subsection{Properties of modified Bessel functions} In this section, we collect some properties of the modified Bessel functions \begin{equation} \label{Bessel} I_{\nu}(s)=\sum_{j=0}^\infty \frac{1}{j ! \Gamma (j+\nu+1)} \(\frac{s}{2}\)^{2j+\nu} \end{equation} used in this paper from [NO]: \begin{itemize} \item For a positive constant $a$, we have \begin{equation} \label{Bessel_1} \int_{-a}^a\frac{e^{s/2}}{\sqrt{a^2-s^2}}ds=\pi I_0\(\frac{a}{2}\). \end{equation} \item Direct computation shows the following recursion: \begin{equation} \label{Bessel_2} I_0' (s)=I_1(s), \ I_1'(s)=I_0(s)-\frac{1}{s}I_1(s), \ \frac{1}{s} \frac{d}{ds}\( \frac{I_\ell(s)}{s^\ell} \) =\frac{I_{\ell+1}(s)}{s^{\ell +1}}. \end{equation} \item The modified Bessel function $I_\nu (s)$ has the expansion \begin{align} \nonumber I_\nu (s)=\frac{e^s}{\sqrt{2\pi s}} &\( 1-\frac{(\nu-1/2)(\nu+1/2)}{2s} +\frac{(\nu-1/2)(\nu-3/2)(\nu+3/2)(\nu+1/2)}{2!2^2s^2} \right. \\ &\quad \left. -\cdots+(-1)^\ell \frac{1}{\ell !2^\ell s^\ell} \prod_{j=1}^{\ell} \( \nu-(j-1/2) \) \( \nu+(j-1/2)\) +O\( \frac{1}{s^{\ell+1}} \) \) \label{Bessel_3} \end{align} as $s$ goes to infinity. \end{itemize}

https://arxiv.org/abs/1709.09149

The center of the reflection equation algebra via quantum minors

We give simple formulas for the elements $c_k$ appearing in a quantum Cayley-Hamilton formula for the reflection equation algebra (REA) associated to the quantum group $U_q(\mathfrak{gl}_N)$, answering a question of Kolb and Stokman. The $c_k$'s are certain canonical generators of the center of the REA, and hence of $U_q(\mathfrak{gl}_N)$ itself; they have been described by Reshetikhin using graphical calculus, by Nazarov-Tarasov using quantum Yangians, and by Gurevich, Pyatov and Saponov using quantum Schur functions; however no explicit formulas for these elements were previously known.As byproducts, we prove a quantum Girard-Newton identity relating the $c_k$'s to the so-called quantum power traces, and we give a new presentation for the quantum group $U_q(\mathfrak{gl}_N)$, as a localization of the REA along certain principal minors.

\section{Introduction} \label{sec:introduction} Attached to an $N\times N$ matrix $A$ is its \emph{characteristic polynomial}, $$p_A(t) = t^N - c_1\cdot t^{N-1} + \cdots + (-1)^N c_N.$$ Here $c_1$ and $c_N$ are the trace and determinant of $A$; more generally each $c_k$ is the sum of all principal $k\times k$ minors of $A$, and can be expressed as the $k$th elementary symmetric function in the eigenvalues of $A$. Taken together, the functions $c_1,\ldots, c_N$ generate the algebra $\mathcal{O}(Mat_N)^{GL_N}$ of conjugation-invariant polynomial functions on the variety of $N\times N$ matrices. The Cayley-Hamilton theorem states that $p_A(A)= 0$ identically, i.e. \begin{equation} \label{eqn:CH} A^N - c_1\cdot A^{N-1} + \cdots + (-1)^N c_N = 0. \end{equation} In addition to the coefficients $c_k$ of the characteristic polynomial, there are the power traces $s_k=tr(A^k)$, for $k=1,\ldots, N$; these can be expressed as the power sum symmetric functions of the eigenvalues of $A$. The elementary and power sum generators are related via the Girard-Newton identities, \begin{equation} \label{eqn:NI} kc_k = \sum_{i=1}^k (-1)^{i-1}s_ic_{k-i}, \end{equation} which allow us to write either set of generators uniquely in terms of the others. The present paper is concerned with the center of the \emph{reflection equation algebra} (REA), denoted $\REA$, associated to the quantized universal enveloping algebra $U_q(\gl)$, and with the analog of identities \eqref{eqn:CH} and \eqref{eqn:NI}. The REA is generated by the entries, $a^i_j$, of an $N\times N$ matrix; the commutation relations amongst the $a^i_j$ are encoded in the so-called reflection equation relations, see Equation \eqref{eqn-RE}. The REA defines a flat (PBW) $q$-deformation of the coordinate algebra $\mathcal{O}(Mat_N)$, along the \emph{Semenov-Tian-Shansky} Poisson bracket \cite{M06}. The reflection equation first arose in the context of mathematical physics, specifically integrable systems related to factorizable scattering on a half-line with reflecting wall. The RE prescribes the consistency relations for collisions with the reflecting wall, analogously to the QYBE for collisions on the interior (see \cite{K96,S92,C84,KS93}). Somewhat independently, the REA played a central role in the braided Hopf algebra perspective on quantum groups pioneered by Majid \cite{M91,LM94,M95} and the quantization of conjugacy classes by Donin and Mudrov \cite{DM02, M07}. More recently, the REA associated to $U_q(\gl)$ arose as the Hochschild homology/horizontal trace of the category $\Rep_q(GL_N)$ of representations of the quantum group \cite{BZBJ15,BZBJ16}, and in the formulation of a $K$-theoretic geometric Satake theorem \cite{CK15}. Finally, we note that the reflection equation appears in the context of co-ideal subalgebras \cite{N96,MRS03}, in particular the reflection equation algebra is an important source of co-ideal sub-algebras admitting a universal $K$-matrix \cite{KS09,K17,JM11} making them into braided module categories for $\Rep_q(GL_N)$. The conjugation action of $GL_N$ on $Mat_N$ also $q$-deforms to the REA, making it into a $U_q(\gl)$-equivariant algebra. The Joseph-Letzter-Rosso isomorphism \cite{JL92, R90thesis, R90} gives an algebra embedding of the REA into the locally finite sub-algebra of $U_q(\gl)$. When $q$ is not a root of unity, the so-called Harish-Chandra isomorphism proved in \cite{JL92} (see also \cite[Proposition 3.6]{KS09}) identifies the center of the REA with its invariant subalgebra; this in turn is isomorphic abstractly to a polynomial ring in $N$ commuting generators. While these isomorphisms describe the center of $U_q(\gl)$ and $\REA$ abstractly, they do not yield any concrete formulas for generators, nor any quantum analog of the Cayley-Hamilton and Girard-Newton identities. The combinatorial problem of giving such formulas has a long history. In Section \ref{sec-related} we review: the diagrammatic techniques of Reshetikhin \cite{R89,B98}, the $q$-Schur function approach of \cite{GPS97,PS96,IOP99}, the quantum Yangian approach of \cite{NT94}, and finally the parallel works of \cite{DL03} and \cite{Z98} for the FRT algebra. However, what is still missing from all these developments is any closed-form, PBW-ordered expression for a system of generators of the center of the REA, and in particular for those appearing as coefficients in a quantum Cayley-Hamilton identity. Remarkably, there is to date no such expression even for the quantum determinant in $\REA$ (except for $N=2$, see Example~\ref{exm:ck-n=2and3}), which must appear when giving a generators-and-relations presentation of $\REAG$ and $\mathcal{O}_q(SL_N)$. This issue has been highlighted in \cite{KS09}, Remark 5.10, and \cite{J14}, Remark 7.9 (see also \cite[Section~4]{DL05} for a related question about explicit computations in the REA). This state of affairs is rectified in this paper: we produce simple, explicit and PBW-ordered expressions for each of the central generators $c_1,\ldots c_N$ appearing in a quantum Cayley-Hamilton identity, and we identify our elements with the prior formulations given in Section \ref{sec-related}. \subsection{Statement of Results} \label{sec:statement-results} We will need the following notation: for each $N\in\mathbb{N}$, we have the set $\SetN = \{1,\ldots,N\}$, we have the collection ${\SetN \choose k}$ of its $k$-element subsets, and we have the symmetric group $S_N$ of permutations of $\SetN$. For each $I\in {\SetN \choose k}$, we have the subgroup $\Sym(I)\subset S_N$, consisting of all elements which fix the complement $I^\comp$ of $I$ in $\SetN$, and we have the \emph{weight}, $\wt(I) = \sum_{i\in I} i.$ Finally, for each permutation $\sigma\in S_N$, we have its \emph{length} and its \emph{exceedance}: \begin{align*}\lng(\sigma) &= \textrm{ the number of pairs $i<j\in\SetN\times \SetN$ such that $\sigma(i)>\sigma(j)$},\\ e(\sigma) &= \textrm{ the number of $i\in \SetN$ such that $\sigma(i)>i$}.\end{align*} For a set \( I \in {[N]\choose k} \) we will index its elements in ascending order \( i_1 < i_2 < \ldots < i_k \). Finally, recall that the generators of $\REA$ organize naturally into the $N\times N$ matrix, $$A = \sum_{i,j}a^i_jE^j_i.$$ In the following discussion we regard $A$ as an element in $Mat_N(\REA)$. With this notation in hand, we can state our main result: \begin{Theorem}[see Corollaries~\ref{cor:cs-are-DL-invars} and~\ref{cor:qCH}]\label{main-thm} The center of $\mathcal{O}_q(Mat_N)$ is freely generated by the elements: $$c_k = \!\!\sum_{I \in {\SetN \choose k}}q^{-2\wt(I)}\!\!\sum_{\sigma\in Sym(I)} (-q)^{\ell(\sigma)}\cdot q^{e(\sigma)}\cdot a^{i_1}_{\sigma(i_1)}\cdots a^{i_k}_{\sigma(i_k)},$$ for $k=1,\ldots, N$. Moreover, the elements $c_k$ satisfy the following \emph{quantum Cayley Hamilton} identity: \begin{equation} \sum_{k=0}^N(-q^2)^{N-k}c_{N-k}\cdot A^k =0 \label{eqn:QCH}. \end{equation} \end{Theorem} \begin{Example} \label{exm:ck-n=2and3} In the case $N=2$, our formulas (up to a choice of normalization) appeared already in \cite{M91} and \cite{KS92}: \begin{equation*} c_1 = q^{-2}a^1_1 + q^{-4} a^2_2, \qquad c_2 = q^{-6}\left(a^1_1a^2_2 - q^2a^1_2a^2_1\right). \end{equation*} Already for $N=3$, our formulas for $c_2$ and $c_3$ appear to be new. They read: \begin{align*} c_1 &= q^{-2}a^1_1+q^{-4}a^2_2 + q^{-6}a^3_3\\ c_2 &= q^{-6}(a^1_1a^2_2-q^2a^1_2a^2_1) + q^{-8}(a^1_1a^3_3 - q^4a^1_3a^3_1) + q^{-10} ( a^2_2a^3_3 - q^2a^2_3a^3_2),\\ c_3 &= q^{-12}\left(a^1_1a^2_2a^3_3 - q^2 a^1_1a^2_3a^3_2 -q^2 a^1_2a^2_1a^3_3 -q^4 a^1_3a^2_2a^3_1+ q^4 a^1_2a^2_3a^3_1 +q^3a^1_3a^2_1a^3_2\right). \end{align*} These should be contrasted with superficially similar formulas for the conjugation co-invariants for the FRT algebra given in Example \ref{DLinv-n=2and3}. \end{Example} \begin{Remark} Note that we have an obvious isomorphism $\Sym(I) \cong S_{|I|}$, however the notion of length appearing in the formula above is that inherited from $S_N$, not from $S_{|I|}$. The first instance of this is the coefficient $q^4$ of $a^1_3a^3_1$ above. \end{Remark} \subsection{Relations to other works}\label{sec-related} Our formulas for $c_k$ are motivated by a number of well-known descriptions for the center of the REA and/or $U_q(\gl)$. Let us now outline these antecedents. Our starting point is in fact Reshetikhin's description \cite{R89} (see also \cite{B98}), which uses the graphical calculus for the REA, as championed independently by Lyubashenko-Majid \cite{LM94}. Reshetikhin's canonical generators $c_k$ are defined as the ``quantum traces'' of the irreducible representations $V(\omega_i)$, for $i=1,\ldots N$. Here, by quantum trace, we mean the composite, $$c_k:\mathbf{1} \xrightarrow{\coev} V(\omega_i)\otimes V(\omega_i)^* \xrightarrow{\sigma} V(\omega_i)^* \otimes V(\omega_i)\hookrightarrow \REA.$$ See Figure~\ref{fig-ckid} for a depiction of $c_k$ in the graphical calculus, and Section~\ref{sec:graph-calc} for a discussion of the graphical caluclus for $\REA$ more generally. The explicit formulas we produce in Theorem~\ref{main-thm} are more precisely formulas for Reshetikhin's central elements $c_k$. The essential challenge here is that while Reshetikhin's generators have a simple and elegant formulation in the 3-dimensional graphical calculus, translating these into ``1-dimensional'' (i.e, algebraic) formulas is highly non-trivial. The first appearance of a quantum Cayley-Hamilton identity in the context of quantum groups appears in the work of Nazarov and Tarasov \cite{NT94}. Their construction involves realizing $U_q(\gl)$ as a quotient of the quantum Yangian, and establishing the required identities in the quantum Yangian. The quantum Cayley-Hamtilon identity was systematically revisited in work of Gurevich, Pyatov, and Saponov. In \cite{GPS97} two systems of central generators $s_k$ and $\sigma_k$ are proposed as $q$-deformations of Schur functions, namely the power sum and elementary symmetric functions respectively. Perhaps the simpler of these are the quantum power traces, $s_k=tr_q(A^k)$, which are $q$-deformations of the classical power traces. Here $tr_q$ of a matrix $B$ is defined as the weighted sum $\sum_{i}b^i_iq^{-2i}$ of its diagonal entries. The $\sigma_k$ are introduced so as to satisfy a quantum Newton identity with respect to the $s_k$'s. Such an identity uniquely determines the $\sigma_k$, because the $s_k$'s freely generate the center. However, they do not propose explicit formulas for the generators $\sigma_k$, such as those of Theorem \ref{main-thm}. In the same paper, the authors further establish a quantum Cayley-Hamilton identity, precisely \eqref{eqn:QCH} up to matching conventions for scalar factors. In Theorem~\ref{thm:qNewton-id}, we prove a similar ``quantum Newton identity'': \begin{equation} [k]_q s_k = \sum_{j=1}^{k} \frac{c_{j-1}}{[j-1]_q!}s_{k-j}=0,\label{eqn:QNI} \end{equation} relating Reshetikhin's invariants to the quantum power traces. Our proof uses the graphical calculus and the finite Hecke algebra. This formula has several consequences. First, it allows us to identify Reshetikhin's central generators $c_k$ with Gurevich-Pyatov-Saponov's generators $\sigma_k$. Secondly, it allows us to transport the quantum Cayley Hamilton identity from \cite{GPS97} to Reshetikhin's generators, and hence to see that our expilcit central elements satisfy a quantum Cayley-Hamilton identity. Finally, let us note that similar, but simpler, formulas to those we obtain for the $c_k$ were discovered in \cite{DL03} for the co-invariants of the so-called FRT algebra (See Example \ref{ex:q-minors}; these feature also in \cite{AY11}). The FRT algebra is a quantization of $Mat_N$ along the \emph{Sklyanin} Poisson bracket; it becomes isomorphic to the restricted dual Hopf algebra to $U_q(\gl)$, upon inverting its quantum determinant. The co-invariants in the FRT algebra are not central however (in fact the center is generated by the quantum determinant only), and the analog of the quantum Cayley-Hamilton theorem for them (see \cite{Z98}) does not have a straightforward interpretation as matrix multiplication as for the REA. There is a process, with many names -- transmutation, equivariantization, twisting -- which produces the REA from the FRT algebra. The FRT algebra is simpler in many respects than the REA, so we perform computations later in the paper by first identifying Reshetikhin's generators $c_k$ as the twist of Domokos-Lenagan's co-invariants, and then computing the effect of the twisting procedure on these. We can summarize these relationships as follows: \begin{Theorem}[see Theorem \ref{thm:qNewton-id}, Corollary \ref{cor:cs-are-DL-invars}] The central generators $c_k$ asserted in Theorem \ref{main-thm} coincide with: Reshetikin's central elements $c_k$, the twist of Domokos-Lenagan's central elements $\DLinv{k}$, and (up to a scalar) Gurevich-Pyatov-Saponov's central elements $\sigma_k$. \end{Theorem} An interesting and related problem is the \( q \)-analogue of the question asked in \cite[Section~3.2]{MR02}, where Molev and Ragoucy construct central elements of certain reflection equation algebras appearing as subalgebras of the Yangian. \subsection{Quantization of conjugacy classes} The works \cite{DM02, M07} of Donin and Mudrov produce explicit quantizations of semi-simple conjugacy classes in $GL_N$. These are obtained by quotenting the REA by quantum Cayley-Hamilton relations, specialized at a given central character according to the eigenvalues of the classical orbit. We hope that the explicit reformulation of the QCH identity in Theorem \ref{main-thm} can further clarify these quantizations and their properties. As an illustrative example, let us consider the quantization of the unipotent cone in $GL_N$. These are all matrices whose characteristic polynomial coincides with that of the identity matrix. For this we consider the co-unit $\epsilon:\REAG\to\CC$, which sends $a^i_j$ to 1 if $i=j$ and to $0$ otherwise. Restricing to the central elements $c_k$, an immediate computation gives that $$\epsilon(c_k) = \sum_{I\in {[N] \choose k}} q^{-2\wt(I)},$$ and hence the resulting quantum Cayley-Hamilton identity becomes: $$(A-q^{-2})(A-q^{-4})\cdots (A-q^{-2N})=0.$$ Interestingly, while the unipotent cone consists of generically non-semisimple matrices, we see that the defining equation for its natural quantization is in fact multiplicity free. \subsection{A new presentation of $U_q(\gl)$} \label{sec:new-pres-u_qgl} As an application, we can give a new presentation of the quantum group $U_q(\gl)$ as a finite degree extension of a localization of $\REAG$, along certain quantum minors which we introduce herein. Since the relations of $\REAG$ are completely encoded in the reflection equation algebra, this is entirely independent of the Serre presentation. It is more closely related to the presentation of $U_q(\gl)$ via so-called $L$-operators (see \cite[Section 8.5]{KS97}). Since the presentation follows easily from our main result, we outline it here in the introduction. Let us first recall the Joseph-Letzter-Rosso isomorphism \cite{JL92, R90}, which allows us to identify $\REAG$ with the subalgebra of locally finite vectors inside $U_q(\mathfrak{gl}_N)$ (see also \cite[Proposition 10.34]{KS97} and \cite[Proposition 2.8]{KS09}). In particular, using this isomorphism, we can just as well regard the central elements $c_k$ in Theorem~\ref{main-thm} (together with the inverse of $c_N$) as generators of the center of $U_q(\mathfrak{gl}_N)$ (in fact this how they were first interpeted by Reshetikhin \cite{R89, B98}). Taking this point further, we can obtain a new presentation of $U_q(\mathfrak{gl}_N)$. We recall the generators $K_i$ in the presentation of $U_q(\mathfrak{gl}_N)$ given in \cite[Section 6.1]{KS97}. We have the following expression relating $c_N$ to the $K_i$'s: \begin{equation}c_N = (K_1\ldots K_N)^2.\label{eqn-cNasKs}\end{equation} It follows\footnote{The statement for $U_q(\mathfrak{gl}_N)$ must be extracted from the corresponding statement for $U_q(\mathfrak{sl}_N)$, but this is straightforward.} from \cite{JL92} (see \cite[Theorem 6.33]{KS97} for an exposition) that while the $q$-commuting element $$K^{-2\rho} = (K_1^{1-N}K_2^{2-N}\ldots K_{N-1})^2\in U_q(\gl)$$ is a locally finite vector for the quantum adjoint action, the element $$K^{\rho} = K_1^{N-1}K_2^{N-2}\ldots K_{N-1},$$ is not, and that furthermore all of $U_q(\mathfrak{gl}_N)$ is obtained from its locally finite part by adjoining $K^{\rho}$. In order to find an expression for $K^\rho$ in the generators of the reflection equation algebra, we need to recall the standard inclusions $U_q(\mathfrak{gl}_{N-1})\hookrightarrow U_q(\mathfrak{gl}_N)$, which can be made compatible with the Joseph-Letzter-Rosso isomorphism. We encode this as follows. For $k=1,\ldots, N$ we let $A_{\geq k}$ denote $(N-k+1)\times(N-k+1)$ matrix, $$A_{\geq k} = \left(\begin{array}{cccc} a^{k}_{k} & \cdots & a^{k}_N\\ \vdots & \ddots & \vdots\\a^N_{k} & \cdots & a^N_N\end{array}\right).$$ In particular, $A_{\geq 1}=A$, and $A_{\geq N} = (a^N_N)$. Inspecting~(\ref{eq:REA-explicit-rels}), one sees that the entries of the matrix $A_{\geq k}$ form a subalgebra of $\mathcal{O}_q(GL_N)$ isomorphic to $\mathcal{O}_q(GL_{N-k+1})$. This is in fact compatible with the inclusion of $U_q(\mathfrak{gl}_{N-k+1})$ into $U_q(\mathfrak{gl}_N)$, upon identifying $\mathcal{O}_q$ in each case with the locally finite vectors. For an $N\times N$ matrix $A$, we denote by $\detq(A)$ the formula $c_N$ above in the entries of $A$. The next proposition follows easily from the above considerations and Equation \eqref{eqn-cNasKs}. \begin{Proposition}\label{dets-and-Ks} For each $k=1,\ldots, N$, we have: $$\detq(A_{\geq k}) = (K_{k}\cdots K_N)^2.$$ \end{Proposition} Now we simply observe that, by Proposition \ref{dets-and-Ks}, we have $$K^{-2\rho} = \detq(A_{\geq 1})^{(1-N)}\cdot \detq(A_{\geq 2}) \cdots \detq(A_{\geq N}),$$ hence it follows: \begin{Corollary} We have an algebra isomorphism, $$U_q(\gl)\cong \REA[\detq(A_{\geq 1})^{-\frac12}, \ldots, \detq(A_{\geq N})^{-\frac12}].$$ \end{Corollary} This presentation reflects the following well-known fact: while $U_q(\gl)$ is often regarded informally as a deformation of the universal enveloping algebra $U(\gl)$, it is more precisely a deformation of the Tits $2^r$-fold cover of the big Bruhat cell $Bw_0B$ inside the group $G$. The big cell $Bw_0B$ is the complement to the union of hypersurfaces defined by the classical limit of the $\detq(A_{>k})$'s. Denoting the Tits cover by $\widetilde{Bw_0B}$ and $\REAG^\circ = \REA[\detq(A_{>0})^{-1}, \ldots, \detq(A_{>N-1})^{-1}]$, we summarize the situation with the following diagram: \[ \begin{tikzcd} \widetilde{Bw_0B} \arrow[two heads,r] \arrow[squiggly,d] & Bw_0B \arrow[hook,r] \arrow[squiggly,d] & G \arrow[squiggly,d] \\ U_q(\gl) & \REAG^\circ \arrow[hook',l] & \REAG^\circ \arrow[hook',l] \end{tikzcd} \] where the downward wavy arrows designate ``pass to coordinate algebras and quantize''. \subsection{Outline of paper} \label{sec:outline-paper} In Section~\ref{sec:preliminaries}, we recall the necessary preliminaries about the reflection equation algebra: its presentation, how to obtain it from the FRT algebra by the twist procedure, and finally we recall the graphical calculus for the REA, in Section~\ref{sec:graph-calc}. We recall some basic facts about the finite Hecke algebra in~\ref{sec:finite-hecke-algebra}, and use it to introduce Reshetikhin's central generators in~\ref{sec:resh-centr-elem}. In Section~\ref{sec:quant-girard-newton}, we prove the quantum Girard-Newton identities for Reshetikhin's generators, in the process we relate them to the generators $\sigma_k$ of \cite{PS96}, and hence are able import the quantum Cayley-Hamilton identity from \cite{GPS97} to our setting. In Section~\ref{sec:minors-combinatorics}, we introduce some combinatorics: we recall in~\ref{sec:permutations} the notions of length and exceedance of a permutation, and then generalize these notions from permutations to general bijections between subsets of a fixed set. We define the quantum minors which appear in the formulas for $c_k$, and prove a simple row expansion formula for them. In Section~\ref{sec:expansion-cliques} we introduce the notion of an expansion clique: this abstracts the set of indices appearing inductively in a row expansion formula for quantum determinants, for use in the proof of our main result. In Section~\ref{sec:main-theorem}, we prove our main result, i.e. formulas for the central generators $c_k$: this proceeds by an elaborate but elementary induction on the collection of expansion cliques, using the row expansion formula. \subsection{MAGMA code} In the process of writing this paper, the authors developed MAGMA \cite{MAGMA} code for working with the reflection equation algebra. The code is publicly available at \href{http://www.github.com/noaham/REA}{http://www.github.com/noaham/REA}. The authors would be happy to help an interested reader to use the code. \subsection{Acknowledgements} \label{sec:acknowledgements} We would like to thank Sammy Black, Pavel Etingof and Masahiro Namiki for helpful discussions at an early stage of the project, and Tom Lenagan for helpful comments on a preliminary version of the paper. Most of the key identities used in the paper were found using extensive computer experimentation in MAGMA. We thank the Simons Foundation for providing site licenses to US institutions, in particular UCLA, so that we could carry out this experimentation. The work of the first author is supported by European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement no. 637618). A portion this work was completed while the second author was a guest at the Max Plank Institute for Mathematics in Bonn. \section{Preliminaries} \label{sec:preliminaries} Here we collect the necessary background on the REA and FRT algebras and the diagrammatic calculus for the REA. We recall Reshetikhin's central elements \cite{R89}, following the exposition of \cite{B98}. Let \( R \) denote the quantum R-matrix for the defining representation \( V \) of \( \gl \) (we follow the conventions in \cite{KS97}): \begin{equation*}\label{eqn-Rmat} R = q\sum_{i}E_i^i\otimes E_i^i + \sum_{i\neq j}E_i^i\otimes E_j^j+(q-q^{-1})\sum_{i>j}E_i^j\otimes E_j^i \end{equation*} (an operator on \( V \otimes V \)), and let us denote by $R_{21} = \tau \cdot R \cdot \tau$, where $\tau$ denotes the vector flip $\tau(v\otimes w) = w\otimes v$. Here $E^i_j$ denotes the elementary matrix, $E^i_j\cdot v_k = v_j,$ if $i=k$, $0$ otherwise. For an element \( M \in \End(V\ox V) \), we denote by \( M^{t_2} \in \End(V\ox V) \) the operator obtained by transposing on the second factor. Let \( \tilde{R} = ((R^{t_2})^{-1})^{t_2} \), then, as is easily verified, \begin{equation*}\label{eqn-Rmat} \tilde{R} = q^{-1}\sum_{i}E_i^i\otimes E_i^i + \sum_{i\neq j}E_i^i\otimes E_j^j-(q-q^{-1})\sum_{i>j}q^{2(i-j)}E_i^j\otimes E_j^i. \end{equation*} \begin{Definition}[See \cite{KS97}, Example 10.18] The \emph{reflection equation} algebra \( \REA \) has generators \( a^i_j \) for \( i,j \in [N] \). The relations amongst the generators are the $N^2\times N^2$ entries of the matrix equation, \begin{equation}\label{eqn-RE} R_{21}A_1R_{12}A_2 = A_2R_{21}A_1R_{12}, \end{equation} where we collect into a matrix \( A = (a^i_j) = \sum a^i_j\otimes E^j_i\), and we denote $A_1=A\otimes\id$, and $A_2=\id\otimes A$. \end{Definition} We will need to make use of the explicit form of these relations, which we can write as follows. Let $\theta$ denote the Heaviside theta function, i.e. \( \theta(x) = 0 \) if \( x \le 0 \) and \( \theta(x) = 1 \) if \( x > 0 \) and let \( \delta_{ij} \) denote the Kronecker delta function, so $\delta_{ij}=1$ if $i=j$, zero else. Applying Equation \eqref{eqn-Rmat}, we have that Equation \eqref{eqn-RE} is equivalent to the list of relations (c.f. \cite[Section~3]{DL05}), \begin{equation}\label{eq:REA-explicit-rels} \begin{aligned} &\text{when } j<i: &a^i_m a^j_n &= q^{\delta_{i,n} + \delta_{n,m} - \delta_{m,j} } a^j_n a^i_m \\ &&&\phantom{=} + \theta(n-m) q^{\delta_{i,m} - \delta_{j,m}}(q-q^{-1}) a^j_m a^i_n \\ &&&\phantom{=} + \delta_{i,n}(q-q^{-1})q^{\delta_{n,m} - \delta_{j,m}} \sum_{p > i} a^j_p a^p_m \\ &&&\phantom{=} + \delta_{i,m}\theta(n-m)(q-q^{-1})^2 \sum_{p > i} a^j_p a^p_n \\ &&&\phantom{=} - \delta_{j,m} q^{-1}(q-q^{-1}) \sum_{p > j} a^i_p a^p_n \\ &\text{when } j=i \text{ and } n<m : \qquad &a^i_m a^i_n &= q^{\delta_{i,n} -\delta_{i,m} - 1 } a^i_n a^i_m \\ &&&\phantom{=} + \delta_{i,n} q^{-1}(q-q^{-1}) \sum_{p > i} a^i_p a^p_m \\ &&&\phantom{=} - \delta_{i,m} q^{-1}(q-q^{-1}) \sum_{p > i} a^i_p a^p_n. \end{aligned} \end{equation} \begin{Definition}[See \cite{KS97}, Section 9.1] The FRT algebra, \( \FRT \) has generators \( x^i_j \) for \( i,j \in [N] \). The relations amongst the generators are the $N^2\times N^2$ entries of the matrix equation, \begin{equation}\label{eqn-FRT}RX_1X_2=X_2X_1R\end{equation} where we collect into a matrix \( X = (a^i_j) = \sum x^i_j\otimes E^j_i\), and we denote $X_1=X\otimes\id$, and $X_2=\id\otimes X$. \end{Definition} Applying Equation \eqref{eqn-Rmat}, we have that Equation \eqref{eqn-FRT} is equivalent to the list of relations, \begin{equation} \label{eq:FRTrel} \begin{aligned} x^i_k x^i_l &= q x^i_l x^i_k \\ x^i_k x^j_k &= q x^j_k x^i_k \\ x^i_l x^j_k &= x^j_k x^i_l \\ x^i_k x^j_l - x^j_l x^i_k &= (q-q^{-1}) x^i_l x^j_k \end{aligned} \end{equation} for any \( 1 \le i < j \le n \) and \( 1 \le k < l \le n \). \begin{Remark} A key distinction between \( \FRT \) and $\REA$ is that the relations of $\FRT$ are ``local'' in the sense that any \( 2\times 2 \) submatrix of \( X = (x^i_j) \) generates a subalgebra isomorphic to $\widetilde{\mathcal{O}_q}(Mat_2)$. This is essentially because the $R$ matrix appears only linearly in the relations \eqref{eqn-FRT}, but quadratically in \eqref{eqn-RE}, and it considerably simplifies computations in the FRT algebra versus the REA. \end{Remark} \subsection{The PBW bases for the RE and FRT algebras} \label{sec:pbw-bases} Suppose we have an algebra \( A \) given in the form \( A = T(V)/R \), for some finite dimensional vector space \( V \) and a two sided ideal \( R \). A totally ordered basis \( \set{x_1 < x_2 < \ldots < x_r} \) of \( V \) is said to be a \emph{PBW generating set} if the set of ordered monomials \begin{equation*} \setc{ x_{i_1}x_{i_2}\cdots x_{i_k}}{ 1 \le i_1 \le i_2 \le \ldots \le i_k \le r, k \in \mathbb{N}} \end{equation*} form a basis of \( A \) which will be called a \emph{PBW basis}. The generating set \( \set{a^i_j} \) for \( \REA \) is given the lexiographic total order where \( a^i_j < a^k_l \) is \( i < k \) or if \( i=k \) and \( j<l \). The generating set \( \set{x^i_j} \) for \( \FRT \) is given the same total order. For example \( a^2_4 < a^3_2 \) and \( a^2_2 < a^2_3 \) and the monomial \( a^1_3a^3_2a^3_3a^4_1 \) is ordered. \begin{Proposition} \label{prp:PBW-algebras} The generating sets \( \set{a^i_j} \) and \( \set{x^i_j} \) are PBW generating sets for the algebras \( \REA \) and \( \FRT \) respectively. \end{Proposition} \begin{proof} This is well-known and can be proved directly, see \cite[Proposition~3.1]{DL05} for a careful proof; or alternatively as a consequence of the quantum Peter-Weyl theorem, see for example \cite[Section~11.5.4]{KS97}. \end{proof} \subsection{The twist construction} \label{sec:twist-mult} We recall here the ``twist'' construction, for relating the algebras $\FRT$ and $\REA$. We follow the exposition of \cite[Section 10.3]{KS97}, and its categorical reformulation in \cite[Section 3]{JM11}. For the history of the construction, see \cite{M95}. To begin, let us recall that the FRT algebra carries commuting $U_q(\gl)$-actions of left and right translation, which are compatible with the co-product in such a way that the left translation action is equivariant for $\Delta^{op}$, rather than $\Delta$. Hence, the FRT algebra is naturally an algebra in the category $$\Rep_q(\GL)^{op}\bt\Rep_q(\GL),$$ where $op$ denotes the opposite tensor product, $V\ot^{op} W=W\ot V$. The tensor product functor, $$T=(T,\id\bt\sigma\bt\id):\Rep_q(\GL)\bt \Rep_q(\GL)\to \Rep_q(\GL),$$ and the identity functor, $$\widetilde{Id}=(Id,\sigma): \Rep_q(\GL)^{op}\to\Rep_q(\GL),$$ each are canonically endowed with tensor structures using the braiding in the appropriate place. Hence we may compose these to obtain a tensor functor, $$\Rep_q(\GL)^{op}\bt\Rep_q(\GL) \xrightarrow{\widetilde{Id}\bt Id} \Rep_q(\GL)\bt\Rep_q(\GL) \xrightarrow{T} \Rep_q(\GL).$$ Hence, given any algebra in $\Rep_q(\GL)^{op}\bt\Rep_q(\GL)$, its image in $\Rep_q(\GL)$ is again an algebra. Because the functors $T$ and $Id$ are both compatible with the forgetful functors to vector spaces, this procedure does not change the underlying vector space, but rather it modifies the multiplication according to the appearances of $\sigma$ in the tensor structures. It is a straightforward computation with $R$-matrices to see the algebra in $\Rep_q(\GL)$ obtained from $\FRT$ in this way is canonically isomorphic to $\REA$. As a particular consequence, we have: \begin{Proposition}[{See \cite[Example 10.18]{KS97} and \cite[Example~1.3.2(c)]{VV10}}] \label{prp:iso} We have a unique isomorphism of vector spaces, \[ \Phi\map{\REA}{\FRT}, \] defined on the generators by \[ a^i_j \mapsto x^i_j, \] and extended to arbitrary elements via the cocycle relation, \begin{equation} \Phi(a b) = m_{FRT}(\mathcal{R}_{13}\mathcal{R}_{23}(a \otimes b)). \label{eq:mult-relate}\end{equation} We will denote the inverse of \( \Phi \) by \( \Psi = \Phi^{-1}\map{\FRT}{\REA} \) \end{Proposition} \begin{Lemma}[See \cite{KS97}, Example 10.18] \label{lem:expanded-rea-mult} The expression in~(\ref{eq:mult-relate}) is given by \begin{align*} \Phi(a^i_j a^k_l) &= R^{in}_{st} x^s_m \tilde{R}^{mk}_{jn} x^t_l \\ &= \begin{cases} x^i_j x^k_l &\text{if } i < k, j \neq k \\ q x^i_j x^k_l &\text{if } i=k, j \neq k \\ q^{-1}x^i_j x^j_l + (q^{-1} - q) \sum_{m > j} q^{-2(m-j)}x^i_m x^m_l &\text{if } i < j, j=k. \end{cases} \end{align*} and the inverse \( \Psi \) is given explicitly by \( \Psi(x^i_jx^k_l) = (R^{-1})^{ik}_{su}a^s_tR^{tu}_{jv}a^v_l \). \end{Lemma} We note in passing that the construction of each of the FRT and RE algebras, as well as the twist construction, can be be understood most conceptually as an instance of the Co-End construction, given originally in this context by Lyubashenko \cite{L95} and Majid \cite{LM94, M95}. We refer to \cite{JM11} for more details. \subsection{Graphical calculus and Reshetikhin's central generators} \label{sec:graph-calc} In this section we briefly recall the graphical calculus for morphisms in a rigid braided tensor category, referring to \cite{K95} for more details. We then recall the constructions of Majid which construct the REA as a braided Hopf algebra, with all the structure maps given as simple morphisms in the graphical calculus. Finally, we use this language to recall a construction Reshetikhin's central generators, adapted to the REA. We refer to \cite{B98} for a careful exposition of Reshetikhin's elements. Let $\mathbf{1}$ denote the unit object in a rigid braided tensor category $\mathcal{A}$. For any objects $V,W$ of a rigid braided tensor category, we have the following basic morphisms, of evaluation $\ev:V^*\ot V\to \mathbf{1}$, coevaluation $\coev:\mathbf{1}\to V\ot V^*$, and braiding $\sigma_{V,W}:V\ot W\to W\ot V$. These are depicted in the graphical calculus in Figure \ref{fig-graphcalc}. \begin{figure}[h] \centering \begin{tikzpicture}[xscale=0.75,baseline={(current bounding box.center)}] \node[anchor=north] (A) at (0,0) {\( V \)}; \node[anchor=north] (B) at (-1,0) {\( V^* \)}; \begin{knot}[draft mode=strands] \strand (A) .. controls +(0,\y) and +(0,\y) .. (B); \end{knot} \end{tikzpicture} \qquad \begin{tikzpicture}[xscale=0.75,baseline={(current bounding box.center)}] \node[anchor=north] (A) at (0,0) {\( V^* \)}; \node[anchor=north] (B) at (-1,0) {\( V \)}; \begin{knot}[draft mode=strands] \strand (A) .. controls +(0,-\y) and +(0,-\y) .. (B); \end{knot} \end{tikzpicture} \qquad \begin{tikzpicture}[xscale=0.75,baseline={(current bounding box.center)}] \node[anchor=north] (A) at (0,0) {\( V \)}; \node[anchor=north] (B) at (1,0) {\( W \)}; \node[anchor=south] (C) at (0,1) {\( W \)}; \node[anchor=south] (D) at (1,1) {\( V \)}; \begin{knot}[] \strand (B) .. controls +(-\x,\y) and +(\x,-\y) .. (C); \strand (A) .. controls +(\x,\y) and +(-\x,-\y) .. (D); \end{knot} \end{tikzpicture} \qquad \begin{tikzpicture}[xscale=1,baseline={(current bounding box.center)}] \node[anchor=north] (A) at (0,0) {\( {^*V} \)}; \node[anchor=south] (B) at (0,2) {\( V^{*} \)}; \draw (A) to [out=90, in=315] (-0.3,1) to [out=135, in=90, looseness=1.7] (-0.8,1); \draw[white,line width=4pt] ($(-0.3,1)+(-2pt,2pt)$) to ($(-0.3,1)+(2pt,-2pt)$); \draw (-0.8,1) to [out=270, in=225, looseness=1.7] (-0.3,1) to [out=45, in=270] (B); \end{tikzpicture} = \begin{tikzpicture}[xscale=0.75,baseline={(current bounding box.center)}] \node[anchor=north] (A) at (0,0) {\( {^*V} \)}; \node[anchor=south] (B) at (0,2) {\( V^{*} \)}; \draw (A) to coordinate[pos=0.5] (Circ) (B); \draw[fill=white, xscale=1.33] (Circ) circle (1.5pt); \end{tikzpicture} \caption{The evaluation, coevaluation, braiding morphism, and loop-de-loop depicted in the graphical calculus. The loop-de-loop is abbreviated by a bead as indicated.} \label{fig-graphcalc} \end{figure} In addition, we will need the ``loop-de-loop'' morphism $l_V: {^*V}\to V^{*}$, defined as the composition, $$l_V = (\ev_{^*V}\otimes \id_{V^*})\circ(\id_V \otimes\sigma_{V^*,{^*V}})\circ(\coev_V\otimes\id_{^*V}),$$ and depicted in the graphical calculus in Figure \ref{fig-graphcalc}. To reduce clutter in diagrams, we will abbreviate this morphism by a bead, as indicated. Following \cite{M95}, \cite{LM94}, \cite{M91}, the quantum coordinate algebra $\REAG$ arises as an algebra of (braided) matrix coefficients for the quantum group $U_q(\gl)$. This equips $\REAG$, and hence $\REA$, with a number of structures we may exploit in our computations to come. Firstly, we recall that, for any finite dimensional (not necessarily irreducible) representation $W$ of $U_q(\gl)$, we have a \emph{canonical} homomorphism of $U_q(\gl)$-modules, $$\iota_W: W^*\otimes W\to \REAG,$$ and if $W$ is a polynomial representation, then the image of $\iota_W$ lies in $\REA\subset \REAG$, recalling that $\REAG$ is obtained from $\REA$ by inverting the matrix coefficient of the quantum determinant representation. Hence for a finite-dimensional polynomial representation $W$ of $U_q(\gl)$, let us denote by $C_{W}$ denote the image of $W^*\otimes W$ under the canonical $U_q(\gl)$-module homomorphism $\iota_W$. The REA is a \emph{braided bi-algebra} in $\Rep_q(GL_N)$. To specify the bialgebra structure maps on $\REA$, it suffices to specify them on each subspace $C_W$. The multiplication $m$, co-multiplication $\Delta$, and co-unit $\epsilon$ morphisms are depicted in the graphical calculus, in Figure \ref{fig-bialg}. The unit is just the identification $\mathbf{1}=C_{\mathbf{1}}$. We use the notational convention \( m^{(k)}:A^{\ot k} \longrightarrow A \) and \( \Delta^{(k)}:A\longrightarrow A^{\ot k} \) for the iterated product and coproduct. In particular \( m = m^{(2)} \) and \( \Delta = \Delta^{(2)} \). \begin{figure}[h] \centering \begin{tikzpicture}[xscale=0.75] \node[anchor=north] (A) at (0,0) {\( V^* \)}; \node[anchor=north] (B) at (1,0) {\( V \)}; \node[anchor=north] (C) at (2,0) {\( W^* \)}; \node[anchor=north] (D) at (3,0) {\( W \)}; \node[anchor=south] (Au) at (0,1) {\( W^* \)}; \node[anchor=south] (Bu) at (1,1) {\( V^* \)}; \node[anchor=south] (Cu) at (2,1) {\( V \)}; \node[anchor=south] (Du) at (3,1) {\( W \)}; \begin{knot}[] \flipcrossings{1,2} \strand (A) .. controls +(\x,\y) and +(-\x,-\y) .. (Bu); \strand (B) .. controls +(\x,\y) and +(-\x,-\y) .. (Cu); \strand (C) .. controls +(-\x,\y) and +(\x,-\y) .. (Au); \strand (D) .. controls +(0,\y) and +(0,-\y) .. (Du); \end{knot} \end{tikzpicture} \qquad \begin{tikzpicture}[xscale=0.6] \node[anchor=north] (A) at (1,0) {\( V^* \)}; \node[anchor=north] (B) at (2,0) {\( V \)}; \node[anchor=south] (Au) at (0,1) {\( V^* \)}; \node[anchor=south] (Bu) at (1,1) {\( V \)}; \node[anchor=south] (Cu) at (2,1) {\( V^* \)}; \node[anchor=south] (Du) at (3,1) {\( V \)}; \begin{knot}[] \strand (A) .. controls +(-\x,\y) and +(\x,-\y) .. (Au); \strand (B) .. controls +(\x,\y) and +(-\x,-\y) .. (Du); \strand (Bu) .. controls +(0,-\y) and +(0,-\y) .. (Cu); \end{knot} \end{tikzpicture} \qquad \begin{tikzpicture}[xscale=0.75] \node[anchor=north] (A) at (0,0) {\( V \)}; \node[anchor=north] (B) at (-1,0) {\( V^* \)}; \begin{knot}[draft mode=strands] \strand (A) .. controls +(0,\y) and +(0,\y) .. (B); \end{knot} \end{tikzpicture} \caption{The bi-algebra structure maps of multiplication, co-multiplication and co-unit in the $\REA$ are depicted in the graphical calculus.}\label{fig-bialg} \end{figure} The matrix $$ A = \sum_{i,j} E^i_j \ot a^j_i \in Mat_N(\CC) \otimes \REA \cong Mat_N(\REA)$$ defines an invariant element of \( \REA \ot V \ot V^* \), and thus may be identified with a homomorphism \(A\in \Hom_{U_q(\gl)}(C_V, \REA) \). Note that we abuse notation and use $A$ for both its matrix and the associated homomorphism. It is clear from the construction that $$A=\iota_V: C_V \to\REA,$$ Likewise, $A^k$ corresponds both to the $k$th power of the matrix $A$, and to the morphism $$A^k:C_V\xrightarrow{\iota_V} \REA \xrightarrow{\Delta^{(k)}} \REA^{\ot k} \xrightarrow{m^{(k)}} \REA,$$ where $\Delta^{(k)}$ and $m^{(k)}$ denote the $(k-1)$th iterated co-product, and product, respectively. For example, the matrix, $$A^2 = \sum_{i,j} E^i_j \ot \sum_k a^j_ka^k_i = \sum_{i,j} E^i_j \ot m \circ \Delta(a^i_j)$$ corresponds to the morphism depicted in Figure \ref{fig-A2}. \begin{figure}[h] \centering \[ A^2 = \begin{tikzpicture}[xscale=0.6, baseline={([yshift=-0.5ex]current bounding box.center)}] \node[anchor=north,gray] at (-0.8,0.5) {\( \Delta \)}; \node[anchor=south,gray] at (-0.8,1.5) {\( m \)}; \node[anchor=north] (A) at (1,0) {\( V^* \)}; \node[anchor=north] (B) at (2,0) {\( V \)}; \node[] (Au) at (0,1) {\( V^* \)}; \node[] (Bu) at (1,1) {\( V \)}; \node[] (Cu) at (2,1) {\( V^* \)}; \node[] (Du) at (3,1) {\( V \)}; \node[anchor=south] (Auu) at (0,2) {\( V^* \)}; \node[anchor=south] (Buu) at (1,2) {\( V^* \)}; \node[anchor=south] (Cuu) at (2,2) {\( V \)}; \node[anchor=south] (Duu) at (3,2) {\( V \)}; \begin{knot}[end tolerance=0em] \flipcrossings{1,2} \strand (A) .. controls +(-\x,\y) and +(\x,-\y) .. (Au); \strand (B) .. controls +(\x,\y) and +(-\x,-\y) .. (Du); \strand (Bu) .. controls +(0,-\y) and +(0,-\y) .. (Cu); \strand (Au) .. controls +(\x,\y) and +(-\x,-\y) .. (Buu); \strand (Bu) .. controls +(\x,\y) and +(-\x,-\y) .. (Cuu); \strand (Cu) .. controls +(-\x,\y) and +(\x,-\y) .. (Auu); \strand (Du) .. controls +(0,\y) and +(0,-\y) .. (Duu); \end{knot} \end{tikzpicture} = \begin{tikzpicture}[xscale=0.6, baseline={([yshift=-0.5ex]current bounding box.center)}] \node[anchor=north] (A) at (1,0) {\( V^* \)}; \node[anchor=north] (B) at (2,0) {\( V \)}; \node[anchor=south] (Au) at (0,1) {\( V^* \)}; \node[anchor=south] (Bu) at (1,1) {\( V^* \)}; \node[anchor=south] (Cu) at (2,1) {\( V \)}; \node[anchor=south] (Du) at (3,1) {\( V \)}; \begin{knot}[end tolerance=0em] \strand[] (Au) to [out=280, in=170] (0.5,0.7) to [out=350, in=260] coordinate[pos=0.4] (Circ) (Cu); \strand (B) .. controls +(\x,\y) and +(-\x,-\y) .. (Du); \strand (Bu) to [out=260, in=70] (0.5,0.7) to [out=250, in=100] (A); \end{knot} \draw[fill=white,xscale=1.666] (Circ) circle (1.5pt); \end{tikzpicture} \] \caption{The morphism $A^2:C_V\to \REA$ depicted here in the graphical calculus is a stand-in for the operation of squaring the matrix $A$.}\label{fig-A2} \end{figure} \begin{Lemma}\label{lem-Ak} The morphism $A^k: C_V \to C_{V^{\ot k}}\subseteq \REA$ is as depicted in Figure \ref{fig-Ak}. \end{Lemma} \begin{figure}[h] \[ A^k = \begin{tikzpicture}[xscale=0.8, baseline={([yshift=-0.5ex]current bounding box.center)}] \node[anchor=north] (A) at (1,0) {\( V^* \)}; \node[anchor=north] (B) at (2,0) {\( V \)}; \node[anchor=south] (Au) at (0,2) {\( V^* \)}; \node[anchor=south] (Bu) at (1,2) {\( V^* \)}; \node[anchor=south] (Cu) at (1.9,2) {\( V \)}; \node[anchor=south] (Du) at (2.8,2) {\( V \)}; \begin{knot}[end tolerance=0em, clip radius=8pt] \strand[double distance=12pt] ($(Au)+(0.25,-0.25)$) to [out=270, in=140] (0.5,1.5) to [out=320, in=180] (1.35,0.7) to [out=0, in=270] coordinate[pos=0] (Circ) ($(Cu)+(0.35,-0.25)$); \strand ($(Du) + (-0.0,-0.25)$) .. controls +(0,-1.7*\y) and +(0,+\y) .. (B); \strand ($(Bu)+(-0.2,-0.25)$) to [out=260, in=30] (0.3,1.5) to [out=210, in=135] (0.5,0.65) to [out=315, in=90] (A); \end{knot} \draw[double distance=4pt] ($(Au)+(0.25,-0.25)$) to [out=270, in=140] (0.5,1.5) to [out=320, in=180] (1.35,0.7) to [out=0, in=270] coordinate[pos=0] (Circ) ($(Cu)+(0.35,-0.25)$); \draw[fill=white,xscale=1.25] ($(Circ)+(0,-6pt)$) circle (1.5pt); \draw[fill=white,xscale=1.25] ($(Circ)+(0,-2pt)$) circle (1.5pt); \draw[fill=white,xscale=1.25] ($(Circ)+(0,2pt)$) circle (1.5pt); \draw[fill=white,xscale=1.25] ($(Circ)+(0,6pt)$) circle (1.5pt); \node[anchor=north] at ($(Cu)+(0.4,0)$) {\( \ldots \)}; \node[anchor=north] at ($(Au)+(0.3,0)$) {\( \ldots \)}; \begin{scope}[label box] \draw (-0.1,1.85) rectangle (1.1,1.2); \node[anchor=east] at (-0.1,1.525) {\( T_{(k\ldots 1)} \)}; \end{scope} \end{tikzpicture} \] \caption{The formula for the morphism $A^k$ asserted in Lemma \ref{lem-Ak} is depicted here in the graphical calculus.}\label{fig-Ak} \end{figure} \begin{proof} The proof of this is an easy induction starting from $k=2$, with the induction step illustrated in Figure \ref{fig-Akpf}. \begin{figure}[h] \[ A^{k+1} = \begin{tikzpicture}[xscale=0.8, baseline={([yshift=-0.5ex]current bounding box.center)}] \node[anchor=north] (Ab) at (1,-1) {\( V^* \)}; \node[anchor=north] (Bb) at (2,-1) {\( V \)}; \coordinate[anchor=north] (A) at (1,0); \coordinate[anchor=north] (B) at (2,0); \node[anchor=south] (Au) at (0,2) {}; \node[anchor=south] (Bu) at (1,2) {}; \node[anchor=south] (Cu) at (1.9,2) {}; \node[anchor=south] (Du) at (2.8,2) {}; \node[anchor=south] (Auu) at (0,2.5) {\( V^* \)}; \node[anchor=south] (Buu) at (1,2.5) {\( V^* \)}; \node[anchor=south] (Cuu) at (1.9,2.5) {\( V \)}; \node[anchor=south] (Duu) at (2.8,2.5) {\( V \)}; \node[anchor=south] (eAuu) at (-0.6,2.5) {\( V^* \)}; \node[anchor=south] (eDuu) at (3.3,2.5) {\( V \)}; \begin{knot}[end tolerance=0em, clip radius=8pt] \strand[double distance=12pt] ($(Auu)+(0.25,-0.25)$) to [out=270, in=90] ($(Au)+(0.25,-0.25)$) to [out=270, in=140] (0.5,1.5) to [out=320, in=180] (1.35,0.7) to [out=0, in=270] coordinate[pos=0] (Circ) ($(Cu)+(0.35,-0.25)$) to [out=90, in=270] ($(Cuu)+(0.35,-0.25)$); \strand ($(Duu) + (-0.0,-0.25)$) to [out=270, in=90] ($(Du) + (-0.0,-0.25)$) .. controls +(0,-1.7*\y) and +(0,+\y) .. (B); \strand ($(Buu) + (-0.2,-0.25)$) to [out=270, in=90] ($(Bu)+(-0.2,-0.25)$) to [out=260, in=30] (0.3,1.5) to [out=210, in=135] (0.5,0.65) to [out=315, in=90] (A) to (Ab); \strand (eDuu) to [out=270, in=90] ($(eDuu)+(0,-2.6)$) to [out=270, in=90, looseness=0.9] (Bb); \end{knot} \draw[double distance=4pt] ($(Auu)+(0.25,-0.25)$) to [out=270, in=90] ($(Au)+(0.25,-0.25)$) to [out=270, in=140] (0.5,1.5) to [out=320, in=180] (1.35,0.7) to [out=0, in=270] coordinate[pos=0] (Circ) ($(Cu)+(0.35,-0.25)$) to [out=90, in=270] ($(Cuu)+(0.35,-0.25)$); \draw[fill=white,xscale=1.25] ($(Circ)+(0,-6pt)$) circle (1.5pt); \draw[fill=white,xscale=1.25] ($(Circ)+(0,-2pt)$) circle (1.5pt); \draw[fill=white,xscale=1.25] ($(Circ)+(0,2pt)$) circle (1.5pt); \draw[fill=white,xscale=1.25] ($(Circ)+(0,6pt)$) circle (1.5pt); \draw[line width=3.5pt, white] (B) to [out=270, in=270, looseness=1.5] ($(eDuu)+(-0.2,-2.3)$) to ($(eDuu)+(-0.2,-0.9)$) to [out=90, in=270, looseness=0.4] (eAuu); \draw (B) to [out=270, in=270, looseness=1.5] ($(eDuu)+(-0.2,-2.3)$) to ($(eDuu)+(-0.2,-0.9)$) to [out=90, in=270, looseness=0.4] (eAuu); \node[anchor=north] at ($(Cuu)+(0.4,0)$) {\( \ldots \)}; \node[anchor=north] at ($(Auu)+(0.3,0)$) {\( \ldots \)}; \begin{scope}[label box] \draw (-0.7,2.4) rectangle (3.45,1.95); \node[anchor=east] at (-0.7,2.165) {\( m \)}; \draw (-0.15,1.85) rectangle (2.95,0.2); \node[anchor=east] at (-0.15,1.025) {\( A^k \)}; \draw (0.8,0.1) rectangle (3.45,-0.8); \node[anchor=east] at (0.8,-0.35) {\( \Delta \)}; \end{scope} \end{tikzpicture} = \begin{tikzpicture}[xscale=0.8, baseline={([yshift=-0.5ex]current bounding box.center)}] \node[anchor=north] (A) at (1,0) {\( V^* \)}; \node[anchor=north] (B) at (2,0) {\( V \)}; \node[anchor=south] (Au) at (0,2) {\( V^* \)}; \node[anchor=south] (Bu) at (1,2) {\( V^* \)}; \node[anchor=south] (Cu) at (1.9,2) {\( V \)}; \node[anchor=south] (Du) at (2.8,2) {\( V \)}; \begin{knot}[end tolerance=0em, clip radius=8pt] \strand[double distance=12pt] ($(Au)+(0.25,-0.25)$) to [out=270, in=140] (0.5,1.5) to [out=320, in=180] (1.35,0.7) to [out=0, in=270] coordinate[pos=0] (Circ) ($(Cu)+(0.35,-0.25)$); \strand ($(Du) + (-0.0,-0.25)$) .. controls +(0,-1.7*\y) and +(0,+\y) .. (B); \strand ($(Bu)+(-0.2,-0.25)$) to [out=260, in=30] (0.3,1.5) to [out=210, in=135] (0.5,0.65) to [out=315, in=90] (A); \end{knot} \draw[double distance=4pt] ($(Au)+(0.25,-0.25)$) to [out=270, in=140] (0.5,1.5) to [out=320, in=180] (1.35,0.7) to [out=0, in=270] coordinate[pos=0] (Circ) ($(Cu)+(0.35,-0.25)$); \draw[fill=white,xscale=1.25] ($(Circ)+(0,-6pt)$) circle (1.5pt); \draw[fill=white,xscale=1.25] ($(Circ)+(0,-2pt)$) circle (1.5pt); \draw[fill=white,xscale=1.25] ($(Circ)+(0,2pt)$) circle (1.5pt); \draw[fill=white,xscale=1.25] ($(Circ)+(0,6pt)$) circle (1.5pt); \node[anchor=north] at ($(Cu)+(0.4,0)$) {\( \ldots \)}; \node[anchor=north] at ($(Au)+(0.3,0)$) {\( \ldots \)}; \begin{scope}[label box] \draw (-0.1,1.85) rectangle (1.1,1.2); \node[anchor=east] at (-0.1,1.525) {\( T_{(k+1\ldots 1)} \)}; \end{scope} \end{tikzpicture} \] \caption{The morphism $A^{k+1}$ is by definition the composition of $m\circ A^k\circ\Delta$. Inductively assuming our formula for $A^k$, we obtain the formula for $A^{k+1}$ by elementary isotopies of braids.}\label{fig-Akpf} \end{figure} \end{proof} \subsection{The finite Hecke algebra} \label{sec:finite-hecke-algebra} In order to formulate Reshetikhin's central elements, we need to recall some basic facts about the finite Hecke algebra. \begin{Definition} The finite Hecke algebra, $H_q(n)$ is generated by the symbols $$T_1,\ldots, T_{n-1},$$ subject to the relations $$T_iT_j=T_jT_i \textrm{ if $|i-j|\geq 2$}, \qquad T_iT_{i+1}T_i = T_{i+1}T_iT_{i+1} \textrm{ for $i=1,\ldots n-1$}$$ $$ T_i-T_{i}^{-1} = q-q^{-1}, \textrm{ for $i=1,\ldots, n-1$}.$$ \end{Definition} \begin{Definition} Given an element $\sigma\in S_n$, we denote by $T_\sigma$ its canonical lift to $H_q(n)$, obtained by writing $\sigma=s_{i_1}\ldots s_{i_\ell}$ as a reduced word in the generators $s_i$ and lifting those to $T_\sigma = T_{i_1}\ldots T_{i_\ell}$. \end{Definition} The element $T_\sigma$ depends only on $\sigma$, not on the reduced word presentation of $\sigma$ chosen. In particular, given a cycle of the form $(k\ldots i)$, for $i<k$ we denote by $$T_{(k\ldots i)} = T_{i}\cdots T_{k-1}$$ the lift obtained in this way. The braid $T_{(k\ldots i)}$ is depicted in Figure \ref{fig-Ak}. \begin{Definition} We define elements $\omega_k, \overline{\omega}_k\in H_q(n)$ as: $$\omega_k = \sum_{\sigma\in S_k} (-q)^{-\ell(\sigma)} T_\sigma \in H_q(k),\qquad \overline{\omega}_k = \frac{\omega_k}{[k]_q!}.$$ where we denote the quantum integers, and the quantum factorial, by: $$[k]_q=1+q^{-2} + \ldots + q^{-2(k-1)}=\frac{1-q^{-2k}}{1-q^{-2}}, \qquad [k]_q! = [k]_q[k-1]_q\cdots [2]_q[1]_q.$$ \end{Definition} For $k<n$, we abuse notation and regard $\omega_k,\overline{\omega}_k$ as an element of $H_q(n)$ via the obvious inclusion $H_q(k)\subset H_q(n)$, sending $T_i$ to $T_i$. \begin{Remark} We note that $\omega_k$ is only a weak idempotent, as $\omega_k\omega_k = [k]_q! \omega_k$, while its normalization $\overline{\omega}_k$ is an idempotent.\end{Remark} Recall that, for $k<N$, quantum Schur-Weyl duality \cite[Section~8.6.3]{KS97} gives an isomorphism $$\End(V^{\ot k})\cong H_q(k),$$ sending the generators $T_i$ to the braiding on the $i$th and $(i+1)$st components. In particular, the fundamental representations of $U_q(\mathfrak{gl}_N)$ arise as sub-modules of $V^{\ot k}$ obtained by projecting along the weak idempotents $\omega_k$. We use this formulation in the next section. \subsection{Reshetikhin's central elements} \label{sec:resh-centr-elem} Finally, let us turn to the construction of Reshetikhin's central elements $c_k$. For any representation $W$ in $\Rep_q(GL_N)$, we define $\tr_q(W) = \iota_W\circ\sigma_{W,W^*}\circ\coev_W(1)$. This is depicted in Figure \ref{fig-trqA}. \begin{figure}[h] \centering \begin{tikzpicture}[xscale=0.75,baseline={([yshift=-0.5ex]current bounding box.center)}] \node[anchor=south] (A) at (0,0) {\( W \)}; \node[anchor=south] (B) at (-1,0) {\( W^* \)}; \coordinate (xs) at (-0.5,-0.5); \coordinate (min) at (-0.5,-0.8); \draw (A) to [out=270, in=45] (xs) to [out=225, in=180, looseness=1.7] (min); \draw[white,line width=4pt] ($(xs)+(-2pt,2pt)$) to ($(xs)+(2pt,-2pt)$); \draw (min) to [out=0, in=315, looseness=1.7] (xs) to [out=135, in=270] (B); \end{tikzpicture} = \begin{tikzpicture}[xscale=0.75,baseline={([yshift=-0.5ex]current bounding box.center)}] \node[anchor=south] (A) at (0,0) {\( W \)}; \node[anchor=south] (B) at (-1,0) {\( W^* \)}; \begin{knot}[draft mode=strands] \strand (A) .. controls +(0,-\y) and +(0,-\y) .. coordinate[pos=0.5] (Circ) (B); \end{knot} \draw[fill=white, xscale=1.33] (Circ) circle (1.5pt); \end{tikzpicture} \caption{The morphism $tr_q(W)$ defines a canonical invariant in $\REAG$, for every $W\in\Rep_q(GL_N)$.}\label{fig-trqA} \end{figure} \begin{Definition} Reshetikhin's canonical central elements are: $$c_k = \tr_q(\Lambda_q^k(V)) = \tr_q(L(\omega_k)),$$ i.e. the central invariant corresponding to the irreducible representation of highest weight $\omega_k$, equivalently the $k$th exterior power of the defining representation. \end{Definition} \begin{Lemma}\label{lem-ckid} We have the identities for Reshetikhin's elements depicted in Figure~\ref{fig-ckid}.\end{Lemma} \begin{proof} The first equality is the definition of $c_k$. The second equality follows from the construction of $\Lambda^k(V)$ as a submodule of $V^{\ot k}$, determined as the image of the projector $\omega_k$ in the finite Hecke algebra, which under Schur-Weyl duality identifies with endomorphisms of $V^{\ot k}$. The third equality follows by decomposing the $k$-fold loop-de-loop into a k-fold inverse double-braiding, composed with the individual loop-de-loops on each strand. The $k$-fold inverse double-braiding has $2\cdot {k \choose 2}$ inverse crossings, each of which simplifies to $-q$ upon meeting the projector $\pi$. \end{proof} \begin{figure}[h] \centering \[ c_k = \begin{tikzpicture}[xscale=1,yscale=1,baseline={([yshift=-0.5ex]current bounding box.center)}] \node[anchor=south] (A) at (0,0) {\( \Lambda^k (V) \)}; \node[anchor=south] (B) at (-1.5,0) {\( \Lambda^k (V)^* \)}; \coordinate (xs) at (-0.75,-0.75); \coordinate (min) at (-0.75,-1.2); \draw (A) to [out=270, in=45] (xs) to [out=225, in=180, looseness=1.7] (min); \draw[white,line width=4pt] ($(xs)+(-3pt,3pt)$) to ($(xs)+(3pt,-3pt)$); \draw (min) to [out=0, in=315, looseness=1.7] (xs) to [out=135, in=270] (B); \end{tikzpicture} = \begin{tikzpicture}[xscale=1,yscale=1,baseline={([yshift=-0.5ex]current bounding box.center)}] \node[anchor=south] (A) at (0,0.5) {\( V^{\otimes k} \)}; \node[anchor=south] (B) at (-1.5,0.5) {\( (V^{\otimes k})^* \)}; \coordinate (xs) at (-0.75,-1); \coordinate (min) at (-0.75,-2); \draw[double distance=9pt] (A) to ($(A)+(0,-0.5)$) to [out=270, in=45] (xs) to [out=225, in=180, looseness=1.7] (min); \draw[double distance=3pt] (A) to ($(A)+(0,-0.5)$) to [out=270, in=45] (xs) to [out=225, in=180, looseness=1.7] (min); \draw[white,line width=14pt] ($(xs)+(-20pt,20pt)$) to ($(xs)+(20pt,-20pt)$); \draw[double distance=9pt] (min) to [out=0, in=315, looseness=1.7] (xs) to [out=135, in=270] ($(B)+(0,-0.5)$) to (B); \draw[double distance=3pt] (min) to [out=0, in=315, looseness=1.7] (xs) to [out=135, in=270] ($(B)+(0,-0.5)$) to (B); \draw[fill=white] ($(A)+(-8pt,-0.4)$) rectangle ($(A)+(8pt,-0.8)$); \node at ($(A)+(0,-0.6)$) {\( \omega_k \)}; \end{tikzpicture} = \;\; q^{2\binom{k}{2}} \cdot \!\!\!\!\! \begin{tikzpicture}[xscale=1,yscale=1,baseline={([yshift=-0.5ex]current bounding box.center)}] \node[anchor=south] (A) at (0,0.5) {\( V^{\otimes k} \)}; \node[anchor=south] (B) at (-1.5,0.5) {\( (V^{\otimes k})^* \)}; \coordinate (min) at (-0.75,-0.6); \draw[double distance=9pt] (A) to ($(A)+(0,-0.5)$) to [out=270, in=0, looseness=1.0] (min) to [out=180, in=270, looseness=1.0] ($(B)+(0,-0.5)$) to (B); \draw[double distance=3pt] (A) to ($(A)+(0,-0.5)$) to [out=270, in=0, looseness=1.0] (min) to [out=180, in=270, looseness=1.0] ($(B)+(0,-0.5)$) to (B); \draw[fill=white] ($(A)+(-7.5pt,-0.35)$) rectangle ($(A)+(7.5pt,-0.7)$); \node at ($(A)+(0,-0.525)$) {\( \omega_k \)}; \draw[fill=white] ($(min)+(0,-4.5pt)$) circle (1.25pt); \draw[fill=white] ($(min)+(0,-1.5pt)$) circle (1.25pt); \draw[fill=white] ($(min)+(0,1.5pt)$) circle (1.25pt); \draw[fill=white] ($(min)+(0,4.5pt)$) circle (1.25pt); \end{tikzpicture} \] \caption{The canonical central elements are the quantum traces of the $k$th exterior power of the defining representation, which may be embedded into $C_{V^{\ot k}}$ via the corresponding projector $\omega_k$. See Lemma \ref{lem-ckid}.}\label{fig-ckid} \end{figure} \begin{Theorem}[{\cite[Theorem 2]{B98}}] The elements $c_1,\ldots, c_N$ generate the center of $\REA$, and together with the inverse of $c_N$ they generate the center of $\REAG$. \end{Theorem} \begin{Remark} We note in passing that the appearance of weight functions $\wt(I)$ in our formulas can ultimately be traced back to the appearance of loop-de-loop's appearing in the right-hand side of Figure \ref{fig-ckid}. It is well-known that the loop-de-loop on the defining representation acts diagonally on the weight basis with powers $1,q^{-2},\cdots, q^{2-N}$. \end{Remark} \section{Quantum Girard-Newton identities} \label{sec:quant-girard-newton} In this section, we prove a quantum Newton identity relating the canonical generators $c_k$ to the quantum trace generators $s_k$. As a corollary, we identify Reshetkhin's canonical generators $c_k$ with (a scalar multiple of) the generators $\sigma_k$ of \cite{PS96,GPS97}. As a corollary, we can apply the quantum Cayley-Hamilton identity from \cite{GPS97} to the $c_k$'s, as asserted in Theorem~\ref{main-thm}. \begin{Definition}\label{def:alphak} We define a linear map, $$\alpha_k: \End_{U_q(\gl)}(V^{\ot k}) \to \REA,$$ by sending $f\in \End(V^{\ot k})$ to the composition, $$\alpha_k(f) = (l_{V}^{\ot k}\ot f)\circ\coev_{(^*V)^{\ot k}},$$ depicted in Figure \ref{fig:alphak}. \end{Definition} \begin{figure}[h] \centering \[ \alpha_k(f) = \begin{tikzpicture}[xscale=1,yscale=1,baseline={([yshift=-0.5ex]current bounding box.center)}] \node[anchor=south] (A) at (0,0.5) {\( V^{\otimes k} \)}; \node[anchor=south] (B) at (-1.5,0.5) {\( (V^{\otimes k})^* \)}; \coordinate (min) at (-0.75,-0.6); \draw[double distance=9pt] (A) to ($(A)+(0,-0.5)$) to [out=270, in=0, looseness=1.0] (min) to [out=180, in=270, looseness=1.0] ($(B)+(0,-0.5)$) to (B); \draw[double distance=3pt] (A) to ($(A)+(0,-0.5)$) to [out=270, in=0, looseness=1.0] (min) to [out=180, in=270, looseness=1.0] ($(B)+(0,-0.5)$) to (B); \draw[fill=white] ($(A)+(-7.5pt,-0.35)$) rectangle ($(A)+(7.5pt,-0.75)$); \node at ($(A)+(0,-0.55)$) {\( f \)}; \draw[fill=white] ($(min)+(0,-4.5pt)$) circle (1.25pt); \draw[fill=white] ($(min)+(0,-1.5pt)$) circle (1.25pt); \draw[fill=white] ($(min)+(0,1.5pt)$) circle (1.25pt); \draw[fill=white] ($(min)+(0,4.5pt)$) circle (1.25pt); \end{tikzpicture} \] \caption{The linear map $\alpha_k(f)$ of Definition \ref{def:alphak} is depicted in the graphical calulus.} \label{fig:alphak} \end{figure} \begin{Proposition} The representation, \begin{align*}\rho: H_q(n) &\to \End(V^{\otimes n}),\\ T_i &\mapsto (\sigma_{V,V})_{i,i+1},\end{align*} descends through $\alpha_{n}$ to a linear map, $$\HH(H_q(n))\to C_{V^{\otimes n}}.$$ \end{Proposition} \begin{proof} Amongst the defining linear relations of $C_{V^{\otimes n}}$ are the relations $$\phi^*f\otimes v = f\otimes\phi(v),$$ for any $\phi:V^{\otimes n}\to V^{\otimes n}$, in particular for the braiding morphisms. Under $\alpha_{n}$, these identify the operation of post-composing and pre-composing with the map $\phi$. Hence $\alpha_{n}\circ \rho$ respects the defining linear relations, $ab-ba$, of $\HH$. \end{proof} \begin{Proposition}\label{QNI-prop} We have the following identity in $\HH(H_q(k))$: \begin{align*}\omega_k &= \omega_{k-1} + (-q^{-1})[k-1]_qT_{k-1}\omega_{k-1}. \end{align*} \end{Proposition} \begin{proof} By definition, we have: $$\omega_k = \omega_{k-1} + \sum_{i=1}^{k-1} (-q)^{k-i}T_{(k\ldots i)} \omega_{k-1}.$$ However, in $\HH$, we can pull each $T_j$, with $j<k-1$, appearing in the product $T_{(k\cdots i)} = T_i\cdots T_{k-1}$ around to the right, where it multiplies against $\omega_{k-1}$ to give a further factor of $(-q^{-1})$. The claimed formula follows immediately. \end{proof} \begin{Corollary}\label{cor-HeckeQNI} The following identity holds in $H_q(n)$: \begin{align*}[k]_q\overline{\omega}_k &= \sum_{j=1}^{k} (-q^{-1})^{k-j-1}\overline{\omega}_{j-1}T_{(j\cdots k)} \end{align*} \end{Corollary} \begin{proof} This is a simple induction starting from Proposition \ref{QNI-prop}.\end{proof} \begin{Theorem} \label{thm:qNewton-id} We have the following ``quantum Newton identity'': \begin{equation} [k]_q s_k = \sum_{j=1}^{k} \frac{c_{j-1}}{[j-1]_q!}s_{k-j}. \end{equation} \end{Theorem} \begin{proof} The theorem follows from Corollary \ref{cor-HeckeQNI}, as soon as we establish the following identity: \begin{equation}\alpha_{k} (\omega_{j-1}T_{(k\ldots j)}) = \alpha_{k}(\omega_{j-1})\alpha_{k}(T_{(k\ldots j)}) = c_{j-1}s_{k-j}.\label{eqn-alphaidentity}\end{equation} First, we recall the identity for $A^k$ given in Lemma \ref{lem-Ak}. Note that pre-composing this identity with with $\tr_q: \mathbf{1} \to C_V$ immediately yields: $$s_k=\tr_q(A^k) = \alpha_{k}(T_{(k\ldots 1)}).$$ On the other hand, Reshetikhin's central elements $c_k$ were defined the as the quantum trace of the $k$th fundamental representation, hence we have: $$c_k = \tr_q(\Lambda^k(V)) = \alpha_{k}(\omega_k) .$$ While $\alpha_{k}$ is not an algebra homomorphism, we note that when two braids $a$ and $b$ involve disjoint strands $1,\ldots, i$ and $i+1,\ldots i+j$, then we have that $\alpha_{i+j}(ab) = \alpha_i(a)\alpha_j(b)$. This is depicted in Figure \ref{fig-ctimesp}. \begin{figure}[h] \centering \[ s_{k-j}c_{j-1} = \begin{tikzpicture}[xscale=1,yscale=1,baseline={([yshift=-0.5ex]current bounding box.center)}] \def1.5{1.5} \def1.5{1.5} \def7.5pt{7.5pt} \coordinate (A) at (0,0); \coordinate (B) at ($(A)+(1.5,0)$); \coordinate (C) at ($(B)+(1.5,0)$); \coordinate (D) at ($(C)+(1.5,0)$); \coordinate (min1) at ($0.5*(A)+0.5*(B)+(0,-2*1.5)$); \coordinate (min2) at ($0.5*(C)+0.5*(D)+(0,-2*1.5)$); \coordinate (exit) at (-0.15,-2.5); \coordinate (dirc) at (-0.2,1); \draw ($(B)+(7.5pt,0)$) .. controls +(0,-1.05) and +(0,0.6) .. ($(A)+(0,-1.5)+(7.5pt,0)$) .. controls +(0,-0.6) and +($0.5*(dirc)$) .. (exit) .. controls +($-0.75*(dirc)$) and +(-0.1,0) .. ($(min1)+(0,-7.5pt)$) .. controls +(1,0) and +(0,-0.6) .. ($(B)+(0,-1.5)+(7.5pt,0)$) .. controls +(0,0.6) and +(0,-1.05) .. ($(C)+(7.5pt,0)$); \draw[white,line width=12.5pt] (B) to [out=270, in=90] ($(A)+(0,-1.5)$) to [out=270, in=180] (min1) to [out=0, in=270] ($(B)+(0,-1.5)$) to [out=90, in=270] (C); \draw[double distance=9pt] (B) to [out=270, in=90] ($(A)+(0,-1.5)$) to [out=270, in=180] (min1) to [out=0, in=270] ($(B)+(0,-1.5)$) to [out=90, in=270] (C); \draw[double distance=3pt] (B) to [out=270, in=90] ($(A)+(0,-1.5)$) to [out=270, in=180] (min1) to [out=0, in=270] ($(B)+(0,-1.5)$) to [out=90, in=270] (C); \draw[white,line width=13pt] (A) to[out=270, in=90] ($(C)+(0,-1.5)$) to[out=270, in=180] (min2) to[out=0, in=270] ($(D)+(0,-1.5)$) to[out=90, in=270] (D); \draw[double distance=9pt] (A) to[out=270, in=90] ($(C)+(0,-1.5)$) to[out=270, in=180] (min2) to[out=0, in=270] ($(D)+(0,-1.5)$) to[out=90, in=270] (D); \draw[double distance=3pt] (A) to[out=270, in=90] ($(C)+(0,-1.5)$) to[out=270, in=180] (min2) to[out=0, in=270] ($(D)+(0,-1.5)$) to[out=90, in=270] (D); \coordinate (rect) at ($(D)+(0,-1.5)+(0,-0.25)$); \draw[fill=white] ($(rect)+(-12pt,0.2)$) rectangle ($(rect)+(11pt,-0.2)$); \node at (rect) {\( \omega_{j-1} \)}; \draw[fill=white] ($(min1)+(0,-4.5pt)$) circle (1.25pt); \draw[fill=white] ($(min1)+(0,-1.5pt)$) circle (1.25pt); \draw[fill=white] ($(min1)+(0,1.5pt)$) circle (1.25pt); \draw[fill=white] ($(min1)+(0,4.5pt)$) circle (1.25pt); \draw[fill=white] ($(min1)+(0,-7.5pt)$) circle (1.25pt); \draw[fill=white] ($(min2)+(0,-4.5pt)$) circle (1.25pt); \draw[fill=white] ($(min2)+(0,-1.5pt)$) circle (1.25pt); \draw[fill=white] ($(min2)+(0,1.5pt)$) circle (1.25pt); \draw[fill=white] ($(min2)+(0,4.5pt)$) circle (1.25pt); \begin{scope}[label box] \coordinate (mrecttl) at ($(A)+(-7.5pt,-3pt)$); \coordinate (mrectbr) at ($(D)+(7.5pt,-1.5)+(0,6pt)$); \draw (mrecttl) rectangle (mrectbr); \gettikzxy{(mrectbr)}{\mrectbrx}{\mrectbry} \gettikzxy{(mrecttl)}{\mrecttlx}{\mrecttly} \node[anchor=east] at ($0.5*(mrecttl)+0.5*(\mrecttlx,\mrectbry)$) {\( m \)}; \coordinate (precttl) at ($(A)+(0,-1.5)+(-7.5pt,+3pt)$); \coordinate (prectbr) at ($(B)+(10.5pt,-2.4*1.5)+(0,6pt)$); \draw (precttl) rectangle (prectbr); \gettikzxy{(prectbr)}{\prectbrx}{\prectbry} \gettikzxy{(precttl)}{\precttlx}{\precttly} \node[anchor=east] at ($0.5*(precttl)+0.5*(\precttlx,\prectbry)$) {\( p_{k-j} \)}; \coordinate (crecttl) at ($(C)+(0,-1.5)+(-7.5pt,+3pt)$); \coordinate (crectbr) at ($(D)+(13.5pt,-2.4*1.5)+(0,6pt)$); \draw (crecttl) rectangle (crectbr); \gettikzxy{(crectbr)}{\crectbrx}{\crectbry} \gettikzxy{(crecttl)}{\crecttlx}{\crecttly} \node[anchor=west] at ($0.5*(crectbr)+0.5*(\crectbrx,\crecttly)$) {\( c_{j-1} \)}; \end{scope} \end{tikzpicture} = \alpha_k(T_{(k\ldots j)}\omega_{j-1}) \] \caption{Because $T_{(k\cdots j)}$ and $\omega_{j-1}$ involve disjoint strands, we have the identity $\alpha_k(T_{(k\cdots j)}\omega_{j-1}) = \alpha_{k-j}(T_{(k\cdots j)})\alpha_{j-1}(\omega_{j-1})$}\label{fig-ctimesp} \end{figure} Hence Equation \eqref{eqn-alphaidentity} follows, and the theorem is proved. \end{proof} As a consequence, we have that Reshetikhin's central generators $c_k$ coincide with Gurevich, Pyatov and Saponov's generators $\sigma_k$ (up to an inconsequential scaling), since both sets of generators are uniquely determined by the fact that they satisfy a quantum Newton identity with respect to quantum power traces, as in the theorem. Since the quantum Cayley-Hamilton theorem was proved already in \cite{GPS97} for the $\sigma_k$'s, we immediately obtain the analogous quantum Cayley-Hamilton identity asserted in the introduction, for the $c_k$'s. \begin{Corollary} \label{cor:qCH} The central elements \( c_k \) satisfy the the following quantum Cayley Hamilton identity \begin{equation} \sum_{k=0}^N(-q^2)^{N-k}c_{N-k}\cdot A^k =0. \end{equation} \end{Corollary} \begin{Remark} It is possible to use similar techniques as above to directly obtain the quantum Cayley-Hamilton formula involving Reshetikhin's generators. However, as it is already proved in \cite{GPS97} that the QCH identities follows from the quantum Girard-Newton identities, we omit such a discussion here. \end{Remark} \section{Quantum minors for the FRT and the REA} \label{sec:minors-combinatorics} In this section we define three of our primary objects of study: the quantum minors for the FRT and the REA algebra, and the truncated minors in the REA. We recall a well-known row expansion formula for quantum minors in the FRT, and we formulate an analogous one for the truncated quantum minors. \subsection{Permutation statistics} \label{sec:permutations} First we recall the classical definitions for permutations in \( \SN \), that is, bijections from \( [N] \) to itself. The \emph{length} \( \lng(\sigma) \) of a permutation \( \sigma \in \SN \) is the number of inversions, that is \[ \lng(\sigma) = \# \setc{(i,j) \in [N]^2}{i < j \text{ and } \sigma(i) > \sigma(j)}. \] We also recall the definition of the \emph{exceedance} \( e(\sigma) \) of a permutation \( \sigma \in \SN \) as \[ e(\sigma) = \# \setc{i \in [N]}{\sigma(i) > i}. \] We extend the definitions of the above statistics to bijections to more general bijections. We fix subsets \( I,J \subset [N] \) such that \( \# I = \# J \), and an \emph{auxiliary set} \( U \subset (I\cup J)^\comp \) (the complement of \( I\cup J \) in \( [N] \)). Given a bijection \( \tau\map{I}{J} \) let \( \hat{\tau}\map{I\cup U}{J \cup U} \) be the bijection defined by \begin{equation*} \hat{\tau}(i) = \begin{cases} \tau(i) &\text{if } i \in I, \\ i &\text{otherwise}. \end{cases} \end{equation*} \begin{Definition} \label{def:length-exceedance} The \emph{exceedance}, \( e(\tau) \) and the \emph{length}, \( \lng_U(\tau) \), of a bijection \( \tau\map{I}{J} \) are defined by \begin{align*} e(\tau) &= \# \setc{a \in I}{\tau(a) > a}, \text{ and } \\ \lng_U(\tau) &= \# \setc{(a,b) \in (I\cup U) \times (I \cup U)}{a<b \text{ and } \hat{\tau}(a) > \hat{\tau}(b)}. \end{align*} \end{Definition} Note that the exceedance does not depend on the auxiliary set \( U \). In the case \( I = J \) and \( U = (I\cup J)^\comp \) this recovers the usual length function of \( \tau \in S)N \). In this case we will drop the subscript \( U \). See Example \ref{ex:q-minors} for examples computing $e(\tau)$ and $\lng_U(\tau)$ for various bijections $\tau$. \subsection{Domokos-Lenagan minors $\DLmin$ and their twists $\Tmin$, $\PTmin$} \label{sec:DL-minors} Let \( I \subset [N] \) be a subset of \( k \) elements, and let \( \Sym(I) \) be the symmetric group permuting the elements of \( I \). Let us denote by \( \tau_I \) the unique order preserving bijection \( \tau_I:I \rightarrow [k] \), and for a permutation \( \sigma \in S_I \) denote \( \sigma^\circ = \tau_I \sigma \tau_I^{-1} \in S_k \). By construction, we have \[ \lng(\sigma^\circ) = \# \setc{(i,j) \in I\times I}{i < j \text{ and } \sigma(i) > \sigma(j)}, \] i.e. $\lng(\sigma^\circ)$ ignores elements of $[N]$ which are not in $I$. Recall that \( \wt(I) = \sum_{i\in I} i \). If \( I \subset [N] \) is a set with \( k \) elements, we will set \( I = \set{i_1,i_2,\ldots,i_k} \) where \( i_1 < i_2 < \ldots < i_k \). \begin{Definition} \label{def:dl-minors} For subsets \( I,J \subset [N] \) such that \( \# I = \# J = k \), we define: \begin{enumerate} \item The \emph{Domokos-Lenagan minor} \cite[Section 7]{DL03} \begin{equation} \label{eq:DL-minor} \DLmin(I,J) = q^{-\wt(I)-\wt(J)}\sum_{\sigma \in \Sym(J)} (-q)^{\lng(\sigma^\circ)}x^{i_1}_{\sigma(j_1)} x^{i_2}_{\sigma(j_2)} \cdots x^{i_k}_{\sigma(j_k)}\in \FRT, \end{equation} where \( i_m = \tau_I^{-1}(m) \) and \( j_m = \tau_J^{-1}(m) \). \item For \( U \subset (I \cup J)^\comp \), the \emph{truncated minor}, \begin{equation*} \PTmin_U(I,J) = q^{-\wt(I)-\wt(J)}\sum_{\tau:I\rightarrow J} (-q)^{\lng_U(\tau)} q^{e(\tau)} a^{i_1}_{\tau(i_1)} a^{i_2}_{\tau(i_2)} \cdots a^{i_k}_{\tau(i_k)} \in\REA, \end{equation*} where the sum is taken over all bijections \( I \longrightarrow J \). \item The \emph{quantum minor}, \begin{equation*} \Tmin(I,J) = \Psi(\DLmin(I,J))\in\REA. \end{equation*} \end{enumerate} \end{Definition} \begin{Example}\label{ex:q-minors} In the case \( N=4 \) and $U=(I\cup J)^\comp$, we have: \begin{align*} \DLmin(\set{1,3},\set{3,4}) &= q^{-11} \Big( x^1_3x^3_4 - q x^1_4 x^3_3 \Big), \\ \PTmin_U(\set{1,3},\set{3,4}) &= q^{-11} \Big( -q^3 a^1_3a^3_4 + q^3 a^1_4 a^3_3 \Big),\\ \Tmin(\set{1,3},\set{3,4}) &= q^{-11} \Big( q a^1_3 a^3_4 - q a^1_4 a^3_3 + (q-q^{-1}) a^1_4a^4_4 \Big) \end{align*} \begin{align*} &\DLmin(\set{1,2,4},\set{1,2,3}) \\ &\phantom{===}= q^{-13} \Big( x^1_1 x^2_2 x^4_3 - q x^1_2 x^2_1 x^4_3 -q x^1_1 x^2_3 x^4_2 -q^3 x^1_3 x^2_2 x^4_1 +q^2 x^1_2 x^2_3 x^4_1 +q^2 x^1_3 x^2_1 x^4_2 \Big),\\ &\PTmin_U(\set{1,2,4},\set{1,2,3}) \\ &\phantom{===}= q^{-13} \Big( a^1_1 a^2_2 a^4_3 - q^2 a^1_2 a^2_1 a^4_3 -q^2 a^1_1 a^2_3 a^4_2 -q^4 a^1_3 a^2_2 a^4_1 +q^4 a^1_2 a^2_3 a^4_1 +q^3 a^1_3 a^2_1 a^4_2 \Big),\\ &\Tmin(\set{1,2,4},\set{1,2,3}) \\ &\phantom{===}= q^{-13} \Big( a^1_1 a^2_2 a^4_3 - q^2 a^1_2 a^2_1 a^4_3 -q a^1_1 a^2_3 a^4_2 -q^3 a^1_3 a^2_2 a^4_1 +q^3 a^1_2 a^2_3 a^4_1 +q^2 a^1_3 a^2_1 a^4_2 \\ &\phantom{==== q^{-13} \Big(}- (q-q^{-1})qa^1_3a^3_1a^4_3 + (q-q^{-1})q^2 a^1_3a^3_3a^4_1 \Big) \end{align*} The formulas above illustrate several key features. While the Domokos-Lenagan minors and the truncated minors can be expressed as a combinatorial sum over bijections, the quantum minors cannot. Similarly, while Propositions \ref{prp:DL-row-exp} and \ref{prp:row-exp-PTmins} below assert a row expansion formula for the Domokos-Lenagan minors and the truncated minors, the quantum minors do not admit such an obvious expansion. In addition to extra summands appearing in the quantum minors, the coefficients of the common summands also disagree, owing to the differing notion of length appearing in their construction. It is therefore rather remarkable that these discrepancies telescope away in the sums computed in Theorem \ref{thm:clique-sums-thm}.\end{Example} \begin{Theorem}[\cite{DL03}, Proposition 7.2] The elements,\label{thm:DLinv} $$\DLinv{k} = \sum_{I\in {[N] \choose k}} \DLmin(I,I),$$ are co-invariants for the adjoint $\FRT$-coaction on itself. \end{Theorem} \begin{Example}\label{DLinv-n=2and3} Let us list the Domokos-Lenagan FRT co-invariants alongside the corresponding REA invariants, when $N=2$ and $3$ for comparison: \begin{equation*} \begin{aligned} N&=2: \\ \DLinv{1} &= q^{-2}x^1_1 + q^{-4} x^2_2, \qquad \DLinv{2} = q^{-6}\left(x^1_1x^2_2 - qx^1_2x^2_1\right).\\ c_1 &= q^{-2}a^1_1 + q^{-4} a^2_2, \qquad c_2 = q^{-6}\left(a^1_1a^2_2 - q^2a^1_2a^2_1\right).\\ \\ N&=3: \\ \DLinv{1} &= q^{-2}x^1_1+q^{-4}x^2_2 + q^{-6}x^3_3,\\ c_1 &= q^{-2}a^1_1+q^{-4}a^2_2 + q^{-6}a^3_3,\\ \DLinv{2} &= q^{-6}(x^1_1x^2_2-q x^1_2x^2_1) + q^{-8}(x^1_1x^3_3 - q x^1_3x^3_1) + q^{-10} ( x^2_2x^3_3 - q x^2_3x^3_2),\\ c_2 &= q^{-6}(a^1_1a^2_2-q^2a^1_2a^2_1) + q^{-8}(a^1_1a^3_3 - q^4a^1_3a^3_1) + q^{-10} ( a^2_2a^3_3 - q^2a^2_3a^3_2),\\ \DLinv{3} &= q^{-12}\left(x^1_1x^2_2x^3_3 - q x^1_1x^2_3x^3_2 - q x^1_2x^2_1x^3_3 - q^3 x^1_3x^2_2x^3_1 + q^2 x^1_2x^2_3x^3_1 + q^2 x^1_3x^2_1x^3_2\right),\\ c_3 &= q^{-12}\left(a^1_1a^2_2a^3_3 - q^2 a^1_1a^2_3a^3_2 -q^2 a^1_2a^2_1a^3_3 -q^4 a^1_3a^2_2a^3_1 + q^4 a^1_2a^2_3a^3_1 +q^3a^1_3a^2_1a^3_2\right). \end{aligned} \end{equation*} The formulas for $\DLinv{k}$ and $c_k$ are essentially the same when $k=1$, but differ in every other case due to the exceedance contribution and the refined notion of length for truncated minors; the simplest case where both can be seen is in comparing $\DLinv{2}$ and $c_2$ when $N=3$. The middle minor has a different relative weighting to the others.\end{Example} \begin{Proposition} \label{prp:twist-maps-dl-resh} The twist isomorphism $\Psi$ maps each Domokos-Lenagan invariant $\DLinv{k}$ to a constant scalar multiple of Reshetikhin's canonical central elements $c_k$. \end{Proposition} \begin{proof} This follows by the fact that $\DLinv{k}$ and $\Phi(c_k)$ are uniquely determined (up to a scalar) as the unique elements of degree $k$ lying in a component $\Lambda^k(V)^*\bt\Lambda^k(V)$, for the $U_q(\gl)$-bimodule action on $\FRT$. \end{proof} In Corollary \ref{cor:cs-are-DL-invars}, we show that the scalar multiple is one, i.e. we show that $\Psi(\DLinv{k})=c_k$. \begin{Remark} While our strategy to prove Theorem \ref{main-thm} involves replacing the quantum minors with the better-behaved truncated versions, it may be interesting nevertheless to find explicit formulas for the quantum minors. For instance, the ideal formed by the quantum minors of a fixed degree $k$ will define a quantization of the variety of matrices with rank less than or equal to $k-1$. \end{Remark} \subsection{Row expansion formulas} \label{sec:row-expans-form} A useful feature of the Domokos-Lenagan minors and the truncated minors is that they admit an inductive definition via row expansion formulas, mimicking those for classical matrices. We will employ the following notation. For a set $X$, a subset \( S \subset X \) and and element \( a \in S \) we will denote by \( S_a \) the set \( S - \set{a} \). If \( b \in X \) then \( S^b \) will denote \( S \cup \set{b} \). We will also stack this notation so that, for instance, \( S_a^{b,c} \) will denote the set \( (S - \set{a})\cup\set{b,c} \). \begin{Proposition} \label{prp:DL-row-exp} The following row expansion formula holds, \begin{equation*} \DLmin(I,J) = \sum_{m=k}^1 (-q)^{m-1}q^{-i_1-j_m} x^{i_1}_{j_m} \DLmin(I_{i_1}, J_{j_m}). \end{equation*} \end{Proposition} \begin{proof} The formula follows from the fact that if \( \sigma \in \Sym(J) \) such that \( \sigma(j_1)=j_m \) then \( \lng(\sigma^\circ) = \lng(\tau^\circ)+m-1 \) where \( \tau \in \Sym(J_{j_m}) \) is the restriction of \( \sigma \). \end{proof} To state the analogous row expansion formula for the truncated quantum minors, we require the following lemma, whose proof follows easily from the definitions. \begin{Lemma} \label{lem:lenth-exceedance} Let \( \tau\map{I}{J} \) be a bijection such that \( \tau(i_1) = j_m \). Let \( \tau_m \map{I_{i_1}}{J_{j_m}} \) be the bijection obtained by restricting \( \tau \). We have the following additive properties for the length and exceedance of \( \tau \), \begin{enumerate} \item \( e(\tau) - e(\tau_m) = \theta(j_m - i_1) \), \item \( \lng_U(\tau) - \lng_{U}(\tau_m) = m-1 + \gamma^{IJ}_U(m), \) \end{enumerate} where \( \gamma_U^{IJ}(m) = \# \setc{a \in U}{i_1 < a < j_m \text{ or } j_m < a < i_1} \). \end{Lemma} With this in hand, we can state the following \begin{Proposition} \label{prp:row-exp-PTmins} We have the following row expansion formula for the truncated minors: \begin{equation*} \PTmin_U(I, J) = \sum_{m=1}^k (-q)^{m-1+\gamma_U^{IJ}(m)} q^{\theta(j_m-i_1)-i_1-j_m} a^{i_1}_{j_m} \PTmin_{U}(I_{i_1}, J_{j_m}). \end{equation*} \end{Proposition} \begin{proof} This follows directly from Lemma~\ref{lem:lenth-exceedance}. \end{proof} \begin{Example} \label{exm:row-exp-example} Let \( U=\emptyset \). The row expansion formula for \( \PTmin_U(\set{1,3,4},\set{2,3,4}) \) yields: \begin{align*} \PTmin_\emptyset(\set{1,2,4},\set{1,2,3}) &= q^{-2}a^1_1\PTmin_{\emptyset}(\set{2,4},\set{2,3}) - q^{-2}a^1_2\PTmin_{\emptyset}(\set{2,4},\set{1,3}) \\ &\phantom{=}+ q^{-2}a^1_3\PTmin_{\emptyset}(\set{2,4},\set{1,2}) \end{align*} which is easily verified using Example~\ref{ex:q-minors} and the fact that \begin{align*} \PTmin_{\emptyset}(\set{2,4},\set{2,3}) &= q^{-11} \Big( a^2_2a^4_3 - q^2a^2_3a^4_2 \Big) \\ \PTmin_{\emptyset}(\set{2,4},\set{1,3}) &= q^{-10} \Big( a^2_1a^4_3 - q^2a^2_3a^4_1 \Big) \\ \PTmin_{\emptyset}(\set{2,4},\set{1,2}) &= q^{-9} \Big( a^2_1a^4_2 - q a^2_2a^4_1 \Big). \end{align*} \end{Example} \section{Expansion cliques} \label{sec:expansion-cliques} Modelled on the row expansion of minors in the previous section, we introduce a combinatorial device called an \emph{expansion clique}. An expansion clique $Cl_k(I,J)$ depends on two subsets $I,J\subset [N]$ of equal cardinality: roughly speaking, the expansion clique $Cl_k(I,J)$ is the set of all complementary indices $I'$ and $J'$, such that (a scalar multiple of) \( a^I_J\PTmin_U(I',J') \) appears as a summand upon performing iterated row expansion to the minor \( \PTmin_U(I\cup I', J\cup J') \). Here \( a^I_J = a^{i_1}_{j_1}\ldots a^{i_{\#I}}_{j_{\#J}} \). \begin{Definition} Fix an integer $k \leq N$, and sets $I,J$ of size $m<k$. The \emph{expansion clique}, $Cl_k(I,J)$, is the set, $$Cl_k(I,J) = \left\{ (I', J') \,\,,\,\, \begin{array}{c} I' \subset \{i_m+1,\ldots N\}\\ J' \subset \{1,\ldots, N\}\end{array} \,\, \left| \,\,\begin{array}{c} I\cup I' = J\cup J',\\ |I'|=|J'|=k-m\end{array}\right. \right\}.$$ \end{Definition} The following lemma will be useful in the proof of the main theorem. \begin{Lemma} \label{lem:bijection-cliques} For any pair of sets \( I,J \), the map \begin{equation*} \beta : \bigsqcup_{m=1}^{k-\# I} Cl_k(I,J) \longrightarrow \bigsqcup_{s,t} Cl_k(I^s,J^t) \end{equation*} given by sending \( (I',J')_m \) to \( (I'_{i'_1},J'_{j'_m}) \in Cl_k(I^{i'_1} | J^{j'_m}) \) is a bijection. Here we label elements of the domain \( (I'J')_m \) as pairs indexed by an integer to indicate which member of the disjoint union they belong to. The \( s \) and \( t \) in the second disjoint union range over \( s \in [i_{\# I}+1,N] \) and \( t \in [N] - J \). \end{Lemma} \begin{proof} The inverse is given by sending \( (I'',J'') \in Cl_k(I^s|J^t) \) to \( (I''\cup s, J''\cup t)_m \) where \( m \) is the index of \( t \) in \( J'' \cup t \). \end{proof} Here we collect a number of identities that will be useful later. The reader may wish to skip ahead to Section \ref{sec:main-theorem} to see where these identities are used. Fix a pair of subsets \( I,J \subset [N] \) such that \( \# I = \# J \le k \). Fix also \( s > i_{\# I} \) and \( t \in J^\comp \), and \( (I'',J'') \in Cl_k(I^s,J^t) \) and an auxiliary set \( U \subset (I^s \cup I'')^\comp \). We will use \( [s,t] \) to denote the interval \( [s,t] \) or \( [t,s] \) depending on whether \( s\le t \) or \( t\le s \). We use \( \ind_I(i) \) to denote the index of \( i \) in \( I \), i.e. \( \ind(i_e) = e \) and we will adopt the convention that if \( r \notin I \) then \( \ind_I(r) := \ind_{I^r}(r) \). We shall consider the quantities: \begin{align} X_{IJ}(I'',J'',s,t) &= \ind_{J''}(t) - 1 + \gamma_U^{(I'')^s, (J'')^t}(\ind_{J''}(t)) \label{eq:X-expression}\\ Y_{IJ}(I'',J'',s,t,r) &= X_{IJ}(s,r) + \lng_{U_{sr}}(\tau_{(I'')_t^r J''}) + N((I'')_t^r,r,t),\label{eq:Y-expression} \end{align} where \( N(K,r,t) = \# K \cap (r,t) \). \begin{Lemma} \label{lem:X-expression} We have the following identity: \[X_{IJ}(I'',J'',s,t)= \# \left( s,t \right) \cap (U \cup I \cup I'') + 1 + \ind_{I}(s) - \ind_{J}(t) - 2 \#I \cap [s,t]. \] In particular, it is independent of \( I'' \) and \( J'' \), so we will henceforth abbreviate $X_{IJ}(s,t)=X_{IJ}(I'',J'',s,t)$. \end{Lemma} \begin{proof} The proof will use the key fact that in the interval between \( s \) and \( t \), either only elements of \( I \) occur (and not \( I'' \)) or only elements of \( I'' \) occur (and not elements of \( I \)), depending on whether \( s \le t \) or \( t\le s \). First consider the case when \( s \le t \). In this case, since every element of \( (I\cup I'')\cap[s,t] \) is actually contained in \( I'' \) (and not \( I \)), we can write \begin{align*} \ind_{J''}(t) -1 &= \# \setc{j \in J''}{j< t} \\ &= \# \setc{j \in J^t\cup J''}{j < t} - \#\setc{j \in J}{j<t} \\ &= \# \setc{i \in I^s\cup I''}{i < t} - \#\setc{j \in J}{j<t} \\ &= \# \setc{i \in I^s}{i \le s} + \#\setc{i \in I''}{s<i<t} - \#\setc{j \in J}{j<t} \\ &= \ind_{I}(s) - \ind_{J}(t)+1 + \#\setc{i \in I''}{s<i<t}. \end{align*} On the other hand, note that the smallest element of \( (I'')^s \) is \( s \) so \begin{align*} \gamma_U^{(I'')^s,(J'')^t}(\ind_{J''}(t)) &= \# (s,t)\cap U. \end{align*} Taking the sum gives the desired result (noting that \( \#I \cap [s,t] = 0 \) when \( s \le t \)). In the case \( t \le s \), we proceed similarly. So \begin{align*} \ind_{J''}(t) -1 &= \# \setc{i \in I^s\cup I''}{i < t} - \#\setc{j \in J}{j<t} \\ &= \# \setc{i \in I^s}{i \le s} - \#\setc{i \in I}{t \le i \le s} - \#\setc{j \in J}{j<t} \\ &= \ind_{I}(s) - \ind_{J}(t)+1 - \#\setc{i \in I}{t \le i \le s}, \end{align*} and again \( \gamma_U^{(I'')^s,(J'')^t}(\ind_{J''}(t)) = \# (s,t)\cap U \) so taking the sum again gives the desired result. \end{proof} For any two subsets \( I,J \subset [N] \) such that \( \# I = \# J \), let \( \tau_{IJ} \) be the unique, order preserving bijection \( I \longrightarrow J \). \begin{Lemma} \label{lem:telescoping-condition} Assume that \( s < t \). If \( r < r' \) are adjacent elements of \( (s,t] - I'' \) then \[ Y_{IJ}(I'',J'',s,t,r') - Y_{IJ}(I'',J'',s,t,r) = 2. \] \end{Lemma} \begin{proof} We can use the expression from Lemma~\ref{lem:X-expression} to see that \begin{align*} X_{IJ}(s,r')-X_{IJ}(s,r) &= \#(s,r')\cap (U \cup I \cup I'') +\ind_{I}(s)-\ind_{J}(r') \\ &\phantom{=}- \left( \#(s,r)\cap (U \cup I \cup I'')+\ind_{I}(s)-\ind_{J}(r) \right) \\ &= \#[r,r')\cap (U \cup I \cup I'') - \left( \ind_{J}(r') - \ind_{J}(r) \right) \\ &= \# I'' \cap [r,r'] +1 - \# J \cap [r,r']. \end{align*} The last equality is due to the fact that \( r \) and \( r' \) are adjacent in \( (s,t] - I'' \), thus every element of \( (r,r') \) must be in \( I'' \). Clearly \[ N(I'',r',t) - N(I'',r,t) = -N(I'',r,r') = -\#I'' \cap [r,r']. \] Finally, we will show that \[ \lng_{U}(\tau_{(I'')_t^{r'} J''}) - \lng_{U}(\tau_{(I'')_t^r J''}) = \# J \cap [r,r'] + 1 \] which will mean we are done. To do this, we imagine the sets \( I,I'' \) depicted by circles and crosses (similarly the sets \( J, J'' \)) and strings connecting \( I'' \) and \( J'' \) to depict \( \tau_{(I'')^r_t, J''} \) as in Figure~\ref{fig:bijection--tauUrt}. We also depict with red lines, the elements of \( U \). We call these \emph{loose strings}. \begin{figure} \centering \begin{tikzpicture}[inner sep=0.5pt] \node (i1) at (1,2) {\( \circ \)}; \node (i2) at (2,2) {\( \circ \)}; \node (i3) at (3,2) {\( \circ \)}; \node (i4) at (4,2) {\( \circ \)}; \node (i5) at (5,2) {\( \circ \)}; \node (s) at (6,2) {\( \star \)}; \node (ip1) at (7,2) {\( \times \)}; \node (ip2) at (8,2) {\( \times \)}; \node (ip3) at (9,2) {\( \times \)}; \node (ip4) at (10,2) {\( \times \)}; \node (ip5) at (11,2) {\( \times \)}; \node (ip6) at (12,2) {\( \times \)}; \node (jp1) at (1,0) {\( \times \)}; \node (j1) at (2,0) {\( \circ \)}; \node (jp2) at (3,0) {\( \times \)}; \node (jp3) at (4,0) {\( \times \)}; \node (j2) at (5,0) {\( \circ \)}; \node (j3) at (6,0) {\( \circ \)}; \node (j4) at (7,0) {\( \circ \)}; \node (jp4) at (8,0) {\( \times \)}; \node (jp5) at (9,0) {\( \times \)}; \node (t) at (10,0) {\( \star \)}; \node (jp6) at (11,0) {\( \times \)}; \node (j5) at (12,0) {\( \circ \)}; \node[anchor=south] at (6,2.2) {\( s \)}; \node[anchor=south] at (10,2.2) {\( t \)}; \node[anchor=south] at (7.67,2.2) {\( r \)}; \node[anchor=south] at (9.5,2.2) {\( r' \)}; \node (u1) at (1.5,2) {\( \cdot \)}; \node (l1) at (1.5,0) {\( \cdot \)}; \node (u2) at (3.5,2) {\( \cdot \)}; \node (l2) at (3.5,0) {\( \cdot \)}; \node (u3) at (5.33,2) {\( \cdot \)}; \node (l3) at (5.33,0) {\( \cdot \)}; \node (u4) at (5.67,2) {\( \cdot \)}; \node (l4) at (5.67,0) {\( \cdot \)}; \node (u5) at (6.33,2) {\( \cdot \)}; \node (l5) at (6.33,0) {\( \cdot \)}; \node (u6) at (7.33,2) {\( \cdot \)}; \node (l6) at (7.33,0) {\( \cdot \)}; \node (u7) at (7.67,2) {\( \cdot \)}; \node (l7) at (7.67,0) {\( \cdot \)}; \node (u8) at (9.5,2) {\( \cdot \)}; \node (l8) at (9.5,0) {\( \cdot \)}; \node (u9) at (10.5,2) {\( \cdot \)}; \node (l9) at (10.5,0) {\( \cdot \)}; \draw (ip1) to (jp1); \draw (u7) to (jp2); \draw (ip2) to (jp3); \draw (ip3) to (jp4); \draw (ip5) to (jp5); \draw (ip6) to (jp6); \draw[red] (u1) to (l1); \draw[red] (u2) to (l2); \draw[red] (u3) to (l3); \draw[red] (u4) to (l4); \draw[red] (u5) to (l5); \draw[red] (u6) to (l6); \draw[red,dashed] (u7) to (l7); \draw[red] (u8) to (l8); \draw[red] (u9) to (l9); \draw[red] (ip4) to (t); \end{tikzpicture} \caption{The bijection \( \tau_{(I'')^r_t, J''} \). \( \circ \in I \) or \( J \), and \( \times \in I'' \) or \( J'' \).} \label{fig:bijection--tauUrt} \end{figure} Since \( \tau_{(I'')^r_t,J''} \) is order preserving, its length is exactly the number of times black strings cross loose strings. In the example in Figure~\ref{fig:bijection--tauUrt}, the length is \( 16 \). We can imagine changing the diagram to the diagram for the bijection \( \tau_{(I'')^{r'}_t,J''} \) by moving the top of the \( r \)-string and the top of each \( I''\cap[r,r'] \)-sting, up one available position (with the last string moving to \( r' \)). We only need to account for crossings gained and lost. Since \( r \) and \( r' \) are adjacent in \( (s,t] - I'' \), these considerations simplify considerably: crossings are only lost by removing \( r' \) as a loose string and crossings are only gained by adding \( r \) as a loose string. The number of strings crossing \( r \) (i.e. the number of crossings gained) is \[ \ind_{J''}(r) - \ind_{I''}(r) \] and the number of strings crossing \( r' \) (crossings lost) is \[ \ind_{J''}(r') - 1 - \ind_{I''}(r') . \] The difference of these two quantities is \[ \# I'' \cap (r,r'] - \# J'' \cap (r,r'] + 1, \] so all that is left to notice is that \( \# I'' \cap (r,r'] - \# J'' \cap (r,r'] = \# J \cap [r,r'] \). \end{proof} \begin{Lemma} \label{lem:Y-min-r} Assume that \( s < t \). Let \( r \) be the minimal element of \( (s,t]-I'' \). Then \begin{equation*} Y_{IJ}(I'',J'',s,t,r) = \lng_{U}(\tau_{(I'')^s (J'')^t}) + \ind_{J''}(t)-1. \end{equation*} \end{Lemma} \begin{proof} Recall the definitions \eqref{eq:X-expression}, \eqref{eq:Y-expression}: \begin{align*} X_{IJ}(s,r) &= \ind_{J''}(r) - 1 + \gamma_U^{(I'')^s,(J'')^r}(\ind_{J''}(r)),\\ Y_{IJ}(I'',J'',s,t,r) &= X_{IJ}(s,r) + \lng_{U_{sr}}(\tau_{(I'')^r_t J''}) + N(I'',r,t). \end{align*} Note that since \( r \) is chosen to be minimal, \( \gamma_U^{(I'')^s,(J'')^r}(\ind_{J''}(r)) = 0 \). Thus the claim in the Lemma reduces to proving \begin{align*} \lng_{U}(\tau_{(I'')^s (J'')^t}) - \lng_{U}(\tau_{(I'')^r_t J''}) &= N(I'',r,t) + \ind_{J''}(r) - \ind_{J''}(t)\\ & = N(I'',r,t) - N(J'',r,t). \end{align*} We will start with the diagram for \( \tau_{(I'')^r_t J''} \) (see the example in Figure~\ref{fig:bijection--tau-r-min}) and describe a process by which to transform it into the diagram for \( \tau_{(I'')^s (J'')^t} \). At each step we will account for any crossings gained or lost and thus arrive at an expression for the difference of the lengths. \begin{figure} \centering \begin{tikzpicture}[inner sep=0.5pt] \node (i1) at (1,2) {\( \circ \)}; \node (i2) at (2,2) {\( \circ \)}; \node (i3) at (3,2) {\( \circ \)}; \node (i4) at (4,2) {\( \circ \)}; \node (i5) at (5,2) {\( \circ \)}; \node (s) at (6,2) {\( \star \)}; \node (ip1) at (7,2) {\( \times \)}; \node (ip2) at (8,2) {\( \times \)}; \node (ip3) at (9,2) {\( \times \)}; \node (ip4) at (10,2) {\( \times \)}; \node (ip5) at (11,2) {\( \times \)}; \node (ip6) at (12,2) {\( \times \)}; \node (jp1) at (1,0) {\( \times \)}; \node (j1) at (2,0) {\( \circ \)}; \node (jp2) at (3,0) {\( \times \)}; \node (j2) at (4,0) {\( \circ \)}; \node (jp3) at (5,0) {\( \times \)}; \node (jp4) at (6,0) {\( \times \)}; \node (j3) at (7,0) {\( \circ \)}; \node (jp5) at (8,0) {\( \times \)}; \node (t) at (9,0) {\( \star \)}; \node (j4) at (10,0) {\( \circ \)}; \node (jp6) at (11,0) {\( \times \)}; \node (j5) at (12,0) {\( \circ \)}; \node[anchor=south] at (6,2.2) {\( s \)}; \node[anchor=south] at (9,2.2) {\( t \)}; \node[anchor=south] at (7.33,2.2) {\( r \)}; \node (u1) at (1.5,2) {\( \cdot \)}; \node (l1) at (1.5,0) {\( \cdot \)}; \node (u2) at (3.33,2) {\( \cdot \)}; \node (l2) at (3.33,0) {\( \cdot \)}; \node (u3) at (3.67,2) {\( \cdot \)}; \node (l3) at (3.67,0) {\( \cdot \)}; \node (u4) at (5.5,2) {\( \cdot \)}; \node (l4) at (5.5,0) {\( \cdot \)}; \node (u5) at (7.33,2) {\( \cdot \)}; \node (l5) at (7.33,0) {\( \cdot \)}; \node (u6) at (7.67,2) {\( \cdot \)}; \node (l6) at (7.67,0) {\( \cdot \)}; \node (u8) at (9.33,2) {\( \cdot \)}; \node (l8) at (9.33,0) {\( \cdot \)}; \node (u9) at (9.67,2) {\( \cdot \)}; \node (l9) at (9.67,0) {\( \cdot \)}; \node (u10) at (11.5,2) {\( \cdot \)}; \node (l10) at (11.5,0) {\( \cdot \)}; \draw (ip1) to (jp1); \draw (u5) to (jp2); \draw (ip2) to (jp3); \draw (ip4) to (jp4); \draw (ip5) to (jp5); \draw (ip6) to (jp6); \draw[red] (u1) to (l1); \draw[red] (u2) to (l2); \draw[red] (u3) to (l3); \draw[red] (u4) to (l4); \draw[red,dashed] (u5) to (l5); \draw[red] (u6) to (l6); \draw[red] (u7) to (l7); \draw[red] (u8) to (l8); \draw[red] (u9) to (l9); \draw[red] (u10) to (l10); \draw[red] (ip3) to (t); \end{tikzpicture} \caption{The bijection \( \tau_{(I'')^r_t, J''} \) for \( r \) minimal.} \label{fig:bijection--tau-r-min} \end{figure} The first step is to move the strings starting at positions \( i''_1, i''_2, \ldots, r \in (I'')^r \) one position to the left. Since \( r \) is minimal, no crossings are lost, and no new crossings are gained. The second step will be to add in the loose string at position \( r \). This creates as many new crossings as there are strings crossing \( r \), i.e. there will be \begin{equation} \label{eq:gains-1} \ind_{J''}(r) - 1 - \ind_{I''}(s) \end{equation} new crossings. Thirdly, if \( m = \ind_{J''}(t) \) and \( e = \ind_{I''}(t) \), we take the \( i''_e \)-string, the \( i''_{e+1} \)-string, up to the \( i''_{m-1} \)-string, and move them all one position to the left. This reduces the number of crossings by \begin{equation} \label{eq:losses-1} \sum_{k=e}^{m-1} \# U \cap (i''_{k-1},i''_k) = \# U \cap (t,i''_{m-1}) - (m-e-1). \end{equation} The fourth step will be to introduce a string starting at \( i''_{m-1} \) and ending at \( t \). This introduces crossings equal to the number of loose strings between \( t \) and \( i''_{m-1} \), i.e. exactly the number given in~(\ref{eq:losses-1}). Thus these contributions cancel out. Lastly, we delete the loose string at position \( t \). This results in a loss of crossings equal to the number of strings crossing \( t \), i.e. there will be \begin{equation} \label{eq:losses-2} \ind_{J''}(t)-1 - \ind_{I''}(t) \end{equation} crossings gained. The total number of crossings gained is thus given by the difference of (\ref{eq:gains-1}) and (\ref{eq:losses-2}). Noting that \begin{align*} N(I'',r,t) &= \ind_{I''}(t) - \ind_{I''}(r), \text{ and } \\ N(J'',r,t) &= \ind_{J''}(t) - \ind_{J''}(r), \end{align*} the claim is proved. \end{proof} \section{The formula for the invariants $c_k$} \label{sec:main-theorem} In this section we prove our main results, by an elaborate induction using the row expansion formulas for the Domokos-Lenagan and truncated minors. They key observation is that while the truncated minors only approximate the image under the twisting isomorphism $\Psi$, the symmetric sums appearing in the definition of the invariants coincide whether we use the truncated or the quantum minors. This is captured in the following theorem. \begin{Theorem} \label{thm:clique-sums-thm} For any \( 1 \le k \le N \), and any pair \( I,J \subset \brak{N} \). We have \begin{equation} \label{eq:main-identity} \sum_{I',J'} \PTmin_{(I\cup I')^\comp}(I',J') = \sum_{I',J'} (-q)^{\lng_{(I\cup I')^\comp}(\tau_{I'J'})} \Tmin(I',J'), \end{equation} where both sums are over \( (I',J') \in Cl_k(I,J) \). \end{Theorem} \begin{proof} Recall that \( \Tmin(I',J') = \Psi\left( \DLmin(I',J') \right) \). We will prove the identity by inducting on \( k-m \). When \( k=m \), the clique \( Cl_k(I,J) \) is either empty and the result is vacuous, or \( I=J \) and the clique contains only the pair of empty subsets \( (\emptyset, \emptyset) \). In this case \( \DLmin(\emptyset,\emptyset) = 1 \) and so \( \PTmin_{I^\comp}(\emptyset,\emptyset) = 1 = \Tmin(\emptyset,\emptyset) \). When \( k > m \), we will in fact prove \[ \sum_{I',J'} \Phi(\PTmin_{(I\cup I')^\comp}(I',J')) = \sum_{I',J'} (-q)^{\lng_{(I\cup I')^\comp}(\tau_{I'J'})} \DLmin(I',J'). \] Since \( \Phi \) is an isomorphism (of vector spaces), this is equivalent to~(\ref{eq:main-identity}). However, the PBW ordering relations in the FRT algebra are much simpler, so it is convenient to check the identity there. To begin, we use the row expansion formula from Proposition~\ref{prp:row-exp-PTmins} to express the left hand side as \begin{align*} \sum_{I',J'} \Phi&\left(\PTmin_{(I\cup I')^\comp}(I',J')\right) \\ &\phantom{=}= \sum_{I', J'} \sum_{m=1}^{k-\# I} q^{\theta(j_m'-i_1')-i'_1-j'_m}(-q)^{m-1+\gamma_{(I\cup I')^\comp}^{I'J'}(m)} \Phi\left(a^{i'_1}_{j'_m} \PTmin_{(I\cup I')^\comp}(I'_{i'_1}, J'_{j'_m})\right). \end{align*} Both sums are taken over \( I',J' \in Cl_k(I,J) \). The terms in the inner summation on the RHS are indexed by the domain of the bijection \( \beta \) in Lemma~\ref{lem:bijection-cliques}. We can use this bijection to re-index the sum and obtain the following expression for the RHS above: \begin{align*} &\sum_{I',J'} \Phi\left(\PTmin_{(I\cup I')^\comp}(I',J')\right) \\ &\phantom{}= \sum_{s,t}\sum_{I'', J''} q^{- s - t + \theta(t-s)} (-q)^{\ind_{J''}(t) -1 + \gamma_{(I^s\cup I'')^\comp}^{(I'')^s, (J'')^t}(\ind_{J''}(t))} \Phi\left( a^{s}_{t} \PTmin_{{(I^s\cup I'')^\comp}}(I'', J'') \right). \end{align*} The first sum is taken over all pairs \( s,t \) so that \( s \in [i_{\# I}+1, N] \) and \( t \in [N] - J \). Here we recognize the exponent of $(-q)$ as the quantity, \[ X_{IJ}(s,t) = \ind_{J''}(t) - 1 + \gamma_{(I^s\cup I'')^\comp}^{(I'')^s,(J'')^t}(\ind_{J''}(t)), \] which was defined in \eqref{eq:X-expression}. {\it A priori} this quantity depends on \( I'',J'' \), however Lemma~\ref{lem:X-expression} tells us this is independent of \( I'' \) and \( J'' \). Thus we can take this factor and the \( a^s_t \) out of the first summand to obtain \begin{align*} \sum_{I',J'} \Phi&\left(\PTmin_{(I\cup I')^\comp}(I',J')\right) \\ &\phantom{=======}= \sum_{s,t}q^{-s-t+\theta(t-s)}(-q)^{X_{IJ}(s,t)} \Phi\left( a^s_t \sum_{I'', J''} \PTmin_{(I^s\cup I'')^\comp}(I'', J'') \right). \end{align*} We can now apply the induction hypothesis to obtain \begin{align*} \sum_{I',J'} \Phi&\left(\PTmin_{(I\cup I')^\comp}(I',J')\right) \\ &\phantom{=}= \sum_{s,t}q^{-s-t+\theta(t-s)}(-q)^{X_{IJ}(s,t)} \Phi\left( a^s_t \sum_{I'', J''} (-q)^{\lng_{(I^s\cup I'')^\comp}(\tau_{I'',J''})} \Tmin(I'', J'') \right) \\ &\phantom{=}= \sum_{s,t}q^{-s-t+\theta(t-s)}(-q)^{X_{IJ}(s,t)} \sum_{I'', J''} (-q)^{\lng_{(I\cup I')^\comp}(\tau_{I'',J''})} \Phi\left( a^s_t\Tmin(I'', J'') \right). \end{align*} We now require the following: \begin{Lemma}\label{lem-aijtimesDL} Let \( I,J \subset [N] \) be subsets of the same size. Assume that \( i \in [N] \) is smaller than any element of \( I \). If $j\not\in I$, we have $\Phi(a^i_j\Tmin(I,J)) = x^i_j\DLmin(I,J)$. If $j\in I$ we have: \begin{align*}\Phi(a^i_j\Tmin(I,J)) &= q^{-1}x^i_j\DLmin(I,J) + (q^{-1}-q)\sum_{k>j} q^{j-k}(-q)^{N(I,j,k)}x^i_k\DLmin(I^k_j,J)\end{align*} where \( N(I,j,k) = \#I \cap (j,k) \). \end{Lemma} \begin{proof} If \( j \notin I \), then only the first case in Lemma~\ref{lem:expanded-rea-mult} applies and the formula \( \Phi(a^i_j\Tmin(I,J)) = x^i_j \DLmin(I,J) \) follows. If \( j \in I \) then a combination of cases (1) and (3) in Lemma~\ref{lem:expanded-rea-mult} imply the formula. \end{proof} Applying the calculation from Lemma~\ref{lem-aijtimesDL} we arrive at the expression \begin{align*} \sum_{I',J'} \Phi&\left(\PTmin_{(I\cup I')^\comp}(I',J')\right) \\ &\phantom{==}= \sum_{s \ge t} q^{-s-t+\theta(t-s)}(-q)^{X_{IJ}(s,t)} \sum_{I'', J''} (-q)^{\lng_{{(I^s\cup I'')^\comp}}(\tau_{I'',J''})} x^{s}_{t} \DLmin(I'', J'') \\ &\phantom{===}+ \sum_{s < t} q^{-s-t+\theta(t-s)}(-q)^{X_{IJ}(s,t)} \sum_{I'', J''} (-q)^{\lng_{{(I^s\cup I'')^\comp}}(\tau_{I'',J''})} \\ &\phantom{===}\cdot \left( q^{-1} x^{s}_{t} \DLmin(I'', J'') + (q^{-1}-q) \sum_{r>t} q^{t-r}(-q)^{N(I'',t,r)} x^s_r \DLmin((I'')^r_t,J'') \right). \end{align*} The next step will be to swap the roles of \( r \) and \( t \) in order to reorganize the sum. The aim is to collect together the terms \( x^s_t \DLmin(I'',J'') \). The resulting expression is \begin{align*} \sum_{I',J'} \Phi\left(\PTmin_{(I\cup I')^\comp}(I',J')\right) &=\sum_{s \ge t} \sum_{I'', J''} q^{-s - t} (-q)^{X_{IJ}(s,t) + \lng_{(I^s\cup I'')^\comp}(\tau_{I'',J''})} x^{s}_{t} \DLmin(I'', J'') \\ + &\sum_{s < t} \sum_{I'', J''} \bigg[ q^{-s - t} (-q)^{X_{IJ}(s,t)+\lng_{(I^s\cup I'')^\comp}(\tau_{I'',J''})}x^{s}_{t} \DLmin(I'', J'') \\ &+ (1-q^2) \sum_{\substack{r \in (s,t) \\ r \notin I''}} q^{-s - t)} (-q)^{Y_{IJ}(I'',J'',s,t,r)}x^{s}_{t} \DLmin(I'', J'') \bigg], \end{align*} where we recall from \eqref{eq:Y-expression} that \begin{equation*} Y_{IJ}(I'',J'',s,t,r) = X_{IJ}(s,r)+\lng_{(I^s\cup I'')^\comp}(\tau_{(I'')^r_t,J''}) + N((I'')^r_t,r,t). \end{equation*} Lemma~\ref{lem:telescoping-condition} shows that the sum indexed by \( r \) telescopes and we arrive at the expression \begin{align*} \sum_{I',J'} \Phi\left(\PTmin_{(I\cup I')^\comp}(I',J')\right) &= \sum_{s \ge t} \sum_{I'', J''} q^{-s - t} (-q)^{X_{IJ}(s,t)+ \lng_{(I^s\cup I'')^\comp}(\tau_{I'',J''})} x^{s}_{t} \DLmin(I'', J'') \\ + &\sum_{s < t} \sum_{I'', J''} q^{-s - t} (-q)^{Y_{IJ}(I'',J'',s,t,r)}x^{s}_{t} \DLmin(I'', J'') \end{align*} where \( r \) is now the minimal element of \( (s,t] - I'' \). By the definition of \( \gamma \), as long as \( s \ge t \), we have \[ \lng_{(I^s\cup I'')^\comp}(\tau_{(I'')^s(J'')^t}) = \lng_{{(I^s\cup I'')^\comp}}(\tau_{I'' J''}) + \gamma_{(I^s\cup I'')^\comp}^{(I'')^s (J'')^t}(\ind_{J''}(t)) \] Thus using Lemma~\ref{lem:Y-min-r}, the above expression simplifies to \begin{align*} \sum_{I',J'} \Phi&\left(\PTmin_{(I\cup I')^\comp}(I',J')\right) \\ &\phantom{=======}= \sum_{s,t} \sum_{I'', J''} q^{-s -t} (-q)^{\lng_{(I^s\cup I'')^\comp}(\tau_{(I'')^s,(J'')^t})+\ind_{J''}(t)-1} x^{s}_{t} \DLmin(I'', J''). \end{align*} This allows use to use the bijection in Lemma~\ref{lem:bijection-cliques} in reverse and obtain \begin{equation*} \sum_{I',J'} \Phi\left(\PTmin_{(I\cup I')^\comp}(I',J')\right) =\sum_{m=1}^{k-\# I}\sum_{I', J'} q^{-i'_1-j'_m}(-q)^{\lng_{(I\cup I')^\comp}(\tau_{I' J'})+m-1} x^{i'_1}_{j'_m} \DLmin(I'_{i'_1},J'_{j'_m}) \end{equation*} which we recognise as the row expansion formula from Proposition~\ref{prp:DL-row-exp} applied to \begin{equation*} \sum_{I',J'} (-q)^{\lng_{(I\cup I')^\comp}(\tau_{I'J'})} \DLmin(I',J'), \end{equation*} and the proof is complete. \end{proof} Our claimed formula for Reshetikhin's invariants $c_k$ is now a special case of the above, when $I=J=\emptyset$: \begin{Corollary} \label{cor:cs-are-DL-invars} Reshetikhin's elements \( c_k \) are given by the PBW ordered expressions \begin{equation*} c_k = \sum_{I \in {\SetN \choose k}} \PTmin_{I^\comp}(I,I) = \sum_{I \in {\SetN \choose k}}q^{-2\wt(I)}\!\!\sum_{\sigma\in Sym(I)} (-q)^{\ell(\sigma)}\cdot q^{e(\sigma)}\cdot a^{i_1}_{\sigma(i_1)}\cdots a^{i_k}_{\sigma(i_k)}. \end{equation*} \end{Corollary} \begin{proof} By Proposition~\ref{prp:twist-maps-dl-resh}, $c_k$ is a scalar multiple of \( \Psi(\DLinv{k}) = \sum_{I} \Tmin(I,I) \). We apply Theorem~\ref{thm:clique-sums-thm} with \( I=J=\emptyset \). Then \( Cl_k(I,J) \) is the set of all pairs \( (I',I') \) where \( I' \) is a \( k \)-element subset of \( [N] \). Thus the left hand side of \ref{eq:main-identity} is \( \sum_{I} \PTmin_{I^\comp}(I,I) \). Note that \( \tau_{I'I'}=\id \) so the right hand side of (\ref{eq:main-identity}) is precisely \( \Psi(\DLinv{k}) \). All that remains is to check the scalar, and for this we can compare a single monomial. It is clear from Lemma \ref{lem:expanded-rea-mult} that $\Psi(x^1_1\cdots x^k_k)=a^1_1\cdots a^k_k$, hence the required scalar multiple must be one. \end{proof} \printbibliography \end{document}

https://arxiv.org/abs/1612.01402

On the existence of weak subgame perfect equilibria

We study multi-player turn-based games played on (potentially infinite) directed graphs. An outcome is assigned to every play of the game. Each player has a preference relation on the set of outcomes which allows him to compare plays. We focus on the recently introduced notion of weak subgame perfect equilibrium (weak SPE). This is a variant of the classical notion of SPE, where players who deviate can only use strategies deviating from their initial strategy in a finite number of histories. Having an SPE in a game implies having a weak SPE but the contrary is generally false.We propose general conditions on the structure of the game graph and on the preference relations of the players that guarantee the existence of a weak SPE, that additionally is finite-memory. From this general result, we derive two large classes of games for which there always exists a weak SPE: (i) the games with a finite-range outcome function, and (ii) the games with a finite underlying graph and a prefix-independent outcome function. For the second class, we identify conditions on the preference relations that guarantee memoryless strategies for the weak SPE.

\section{Introduction} Subgame-perfect equilibria (SPEs) are a core solution concept for sequential games. For infinite duration games, they only exist in restricted cases, though. A weaker variant of SPE, \emph{weak SPE} was recently introduced in~\cite{BBMR15}. While an SPE must be resistant to any unilateral deviation of one player, a weak SPE must be resistant to such deviations where the deviating strategy differs from the original one on a \emph{finite number} of histories only, or, equivalently, a \emph{single} history. The latter class of deviating strategies is a well-known notion that for instance appears in the proof of Kuhn's theorem~\cite{kuhn53} with the one-step deviation property. There are games for which there exists a weak SPE but no SPE~\cite{BBMR15,SV03}. The notion of weak SPE is important for several reasons (more details are given in the related work discussed below). First, for the large class of games with upper-semicontinuous payoff functions and for games played on finite trees, the notions of SPE and weak SPE are equivalent. Second, it is a central technical ingredient used to reason on SPEs as shown in~\cite{BBMR15} and~\cite{Flesch10}. Third, being immune to strategies that finitely deviate from the initial strategy profile may be sufficient from the perspective of synthesis (see more below). In this paper, we provide the following contributions. First, we identify a general class of games played on potentially infinite graphs and prove that these games always admit weak SPE (Theorem~\ref{thm:generalgraph}). The proof of our result has an algorithmic flavour, and proceeds via transfinite induction. The weak SPEs we construct require only finite memory to execute, meaning that the prescribed action at any history depends only on the current vertex and on the state of some finite automaton. Second, starting from this general existence result, we prove the existence of a weak SPE: \begin{itemize} \item for games on infinite trees with a \emph{finite} number of outcomes (Theorem~\ref{thm:infinitetree}, reproving a result from \cite{Flesch10}); \item for games with a \emph{finite} underlying graph and a \emph{prefix-independent} outcome function (Theorem~\ref{thm:finitegraph}). \end{itemize} \noindent Additionally, in the second result, we identify conditions on the players' outcome preferences that guarantee the existence of a weak SPE composed of \emph{uniform memoryless} strategies only (Theorem~\ref{thm:uniform}). \paragraph{{\bf Related work}} The concept of SPE has been first introduced and studied by the game theory community. In~\cite{kuhn53}, Kuhn proves the existence of SPEs in games played on finite trees. This result has been generalized in several ways. All games with a continuous real-valued outcome function and a finitely branching tree always have an SPE~\cite{Roux14} (the special case with finitely many players is first established in~\cite{Fudenberg83}). In~\cite{Flesch10} (resp. \cite{Purves11}), the authors prove that there always exists an SPE for games with a finite number of players and with a real-valued outcome function that is upper-semicontinuous (resp. lower-semicontinuous) and has finite range. The result of \cite{Purves11} is extended to an infinite number of players in~\cite{Flesch17}. In~\cite{Roux14}, it is proved using Borel determinacy that all two-player games with antagonistic preferences over finitely many outcomes and a Borel-measurable outcome function have an SPE. In~\cite{Roux15}, Le Roux shows that all games where the preferences over finitely many outcomes are free of some ``bad pattern'' and the outcome function is $\Delta^0_2$ measurable (a low level in the Borel hierarchy) have an SPE. In part of the aforementioned works, the equivalence between SPEs and weak SPEs is implicitly used as a proof technique: in a finite setting in~\cite{kuhn53}, in a continuous setting in \cite{Fudenberg83}, and in a lower-semicontinuous setting in \cite{Flesch10}. In the latter reference, it is implicitly proven that all games with a finite range real-valued outcome function always have a weak SPE (which appears to be an SPE when the outcome function is additionally lower-semicontinuous). We obtain this result here as a consequence of a more general theorem, with a proof of a more algorithmic nature. The concept of SPE and other solution concepts for multi-player non zero-sum games have been considered recently by the theoretical computer community, see~\cite{BrenguierCHPRRS16} for a survey. The existence of SPEs (and thus weak SPEs) is established in~\cite{Ummels06} for games played on graphs by a finite number of players and with Borel Boolean objectives. In~\cite{BBMR15}, weak SPEs are introduced as a technical tool for showing the existence of SPEs in quantitative reachability games played on finite weighted graphs. An algorithm is also provided for the construction of a (finite-memory) weak SPE that appears to be an SPE for this particular class of games. In this paper, we give several existence results that are orthogonal to the results obtained in \cite{BBMR15} as they are concerned with possibly infinite graphs or prefix-independent outcome functions. Other refinements of Nash equilibria (NE) are studied. Let us mention the secure equilibria for two players first introduced in~\cite{CHJ06} and then used for reactive synthesis in~\cite{ChatterjeeH07}. These equilibria are generalized to multiple players in~\cite{Depril14} or to quantitative objectives in~\cite{BMR14}, see also a variant called Doomsday equilibrium in \cite{Chatterjee0FR14}. Like NEs, they are subject to possible non-credible threats. Other alternatives to NE are provided by the notion of admissible strategy introduced in~\cite{Berwanger07}, with computational aspects studied in~\cite{BrenguierRS14}, and potential for synthesis studied in~\cite{BrenguierRS15}. Note that these notions are free, like (weak) SPEs, of non-credible threats. Finally, in \cite{KupfermanPV16}, the authors introduce the notion of cooperative and non-cooperative rational synthesis as a general framework where rationality can be specified by either NE, or SPE, or the notion of dominating strategies. In all cases except~\cite{BMR14} and~\cite{Depril14}, the proposed solution concepts are not guaranteed to exist, hence results concern mostly algorithmic techniques to decide their existence, instead of general conditions for existence as in this paper. \paragraph{{\bf Applications to reactive synthesis}} Games played on graphs have a large number of applications in theoretical computer science. One particularly important application is \emph{reactive synthesis}~\cite{PR89}, i.e. the design of a controller that guarantees a good behavior of a reactive system evolving in a possibly hostile environment. One classical model proposed for the synthesis problem is the notion of \emph{two-player zero-sum game played on a graph}. One player is the reactive system and the other one is the environment; the vertices of the graph model their possible states and the edges model their possible actions. Interactions between the players generate an infinite play in the graph which model behaviors of the system within its environment. As one cannot assume cooperation of the environment, the objectives of the two players are considered to be opposite. Constructing a controller for the system then means devising a \emph{winning strategy} for the player modeling it. Reality is often more subtle and the environment is usually not fully adversarial as it has its own objective, meaning that the game should be non zero-sum. Moreover instead of two players, we could consider the more general situation of several players modeling different interacting systems/environments each of them with its own objective. This has lead to an exploration of a variety of solution concepts for sequential games from the perspective of theoretical computer science (see survey \cite{bruyere-survey}). Weak SPE have the benefit of allowing less unreasonable threats than Nash equilibria, but existing in more cases than SPE. We can even imagine ruling out infinite deviations by letting a meta-agent punish every one-shot deviation with a (low) fixed probability. A player using an infinitely-deviating strategy will thus be punished by the meta-agent with probability one. Protocols like BitTorrent use similar ideas: every deviant user is temporarily denied suitable bandwidth (see Chapter \textit{Bandwidth Trading as Incentive} in \cite{BitTorrent} for details). \paragraph{{\bf Structure of the paper}} In Section~\ref{sec:prelim}, we recall the useful notions of game, strategy and weak SPE. In Section~\ref{sec:general}, we present our general conditions that guarantee the existence of a weak SPE. From this general existence result, we derive two large classes of games with a weak SPE: games with a finite-range outcome function in Section~\ref{sec:first}, and games with a finite underlying graph and a prefix-independent outcome function in Section~\ref{sec:second}. In Section \ref{sec:counterexample} we provide an example of a game without weak SPE demonstrating limitations to possible extensions of our main theorem. An extended abstract omitting most proofs has appeared as \cite{wSPE-fossacs}. \section{Preliminaries} \label{sec:prelim} In this section, we recall the useful notions of game, strategy, and weak subgame perfect equilibrium. We illustrate these notions with examples. \subsection{Games} We consider multi-player turn-based games such that an outcome is assigned to every play. Each player has a preference relation on the set of outcomes which allows him to compare plays. \begin{definition} A \emph{game} is a tuple $G = (\Pi, V, (V_i)_{i \in \Pi}, E, O, \mu, (\prec_i)_{i \in \Pi})$ where: \begin{itemize} \item $\Pi $ is a set of players, \item $V$ is a set of vertices and $E \subseteq V \times V$ is a set of edges, such that w.l.o.g. each vertex has at least one outgoing edge, \item $(V_i)_{i \in \Pi}$ is a partition of $V$ such that $V_i$ is the set of vertices controlled by player $i \in \Pi$, \item $O$ is a set of outcomes and $\mu : V^\omega \to O$ is an outcome function, \item $\prec_i$ $\subseteq O \times O$ is a preference relation for player $i \in \Pi$. \end{itemize} \end{definition} In this definition the underlying graph $(V,E)$ can be infinite (that is, of arbitrarily cardinality), as well as the set $\Pi$ of players and the set $O$ of outcomes. A \emph{play} of $G$ is an infinite (countable) sequence $\rho = \rho_0 \rho_1 \ldots \in V^\omega$ of vertices such that $(\rho_i, \rho_{i + 1}) \in E$ for all $i \in \mathbb N$. \emph{Histories} of $G$ are finite sequences $h = h_0 \ldots h_n \in V^+$ defined in the same way. We often use notation $hv$ to mention the last vertex $v \in V$ of the history. Usually histories are non empty, but in specific situations it will be useful to consider the empty history $\epsilon$. The set of plays is denoted by $Plays$ and the set of histories (ending with a vertex in $V_i$) by $Hist$ (resp. by $Hist_i$).\footnote{Indexing $Plays_G$ or $Hist_G$ with $G$ allows to recall the related game $G$.} A \emph{prefix} (resp. \emph{suffix}) of a play $\rho = \rho_0\rho_1 \ldots$ is a finite sequence $\rho_{\leq n} = \rho_0 \dots \rho_n$ (resp. infinite sequence $\rho_{\geq n} = \rho_n \rho_{n+1} \ldots$). We use notation $h < \rho$ when a history $h$ is prefix of a play $\rho$. When an initial vertex $v_0 \in V$ is fixed, we call $(G, v_0)$ an \emph{initialized} game. In this case, plays and histories are supposed to start in $v_0$, and we use notations $Plays(v_0)$ and $Hist(v_0)$. In this article, we often \emph{unravel} the graph of the game $(G, v_0)$ from the initial vertex $v_0$, which yields an infinite tree rooted at $v_0$. The outcome function assigns an outcome $\mu(\rho) \in O$ to each play $\rho \in V^\omega$. It is \emph{prefix-independent} if $\mu(h \rho) = \mu(\rho)$ for all histories $h$ and play $\rho$. A \emph{preference} relation $\prec_i$ $\subseteq O \times O$ is an irreflexive and transitive binary relation. It allows for player~$i$ to compare two plays $\rho, \rho' \in V^\omega$ with respect to their outcome: $\mu(\rho) \prec_i \mu(\rho')$ means that player~$i$ prefers $\rho'$ to $\rho$. In this paper we restrict to \emph{linear} preferences. (It is w.l.o.g. since the preference properties that we use are preserved by linear extension). We write $o \preceq_i o'$ when $o \prec_i o'$ or $o = o'$; notice that $o \nprec_i o'$ if and only if $o' \preceq_i o$. We sometimes use notation $\prec_v$ instead of $\prec_i$ when vertex $v \in V_i$ is controlled by player~$i$. \begin{example} \label{ex:classical} Let us mention some classical classes of games where the set of outcomes~$O$ is a subset of $(\mathbb R \cup \{+\infty, -\infty \})^\Pi$, and for all player~$i \in \Pi$, $\prec_i$ is the usual ordering $<$ on $\mathbb R \cup \{+\infty, -\infty \}$ on the outcome $i$-th components. In other words, each player~$i$ has a real-valued payoff function $\mu_i : Plays \to \mathbb R \cup \{+\infty, -\infty \}$. The outcome function of the game is then equal to $\mu = (\mu_i)_{i \in \Pi}$, and for all $i \in \Pi$, $\mu(\rho) \prec_i \mu(\rho')$ whenever $\mu_i(\rho) < \mu_i(\rho')$. Games with \emph{Boolean} objectives are such that $\mu_i : Plays \to \{0,1\}$ where $1$ (resp. $0$) means that the play is won (resp. lost) by player~$i$. Classical objectives are Borel objectives including $\omega$-regular objectives, like reachability, B\"uchi, parity, aso~\cite{GU08}. Prefix-independence of $\mu_i$ holds in the case of B\"uchi and parity objectives, but not for reachability objective. We have \emph{quantitative} objectives when $\mu_i : Plays \to \mathbb R \cup \{+\infty, -\infty \}$ replaces $\mu_i : Plays \to \{0,1\}$. Usually, such a $\mu_i$ is defined from a weight function $w_i : E \to \mathbb R$ that assigns a weight to each edge. Classical examples of $\mu_i$ are \emph{limsup} and \emph{mean-payoff} functions~\cite{LaurentDoyen}, that is\footnote{The limit inferior can be used instead of the limit superior.}, \begin{itemize} \item \emph{limsup}: $\mu_i(\rho) = \limsup_{k \to \infty} w_i(\rho_k,\rho_{k+1})$ \item \emph{mean-payoff}: $\mu_i(\rho) = \limsup_{n \to \infty} \sum_{k=0}^{n} \frac{w_i(\rho_k,\rho_{k+1})}{n}$ \end{itemize} \end{example} \subsection{Strategies Let $(G, v_0)$ be an initialized game. A \emph{strategy} $\sigma$ for player~$i$ in $(G,v_0)$ is a function $\sigma: Hist_i(v_0) \to V$ assigning to each history $hv \in Hist_i(v_0)$ a vertex $v' = \sigma(hv)$ such that $(v, v') \in E$. A strategy $\sigma$ of player $i$ is \emph{positional} if it only depends on the last vertex of the history, i.e. $\sigma(hv) = \sigma(v)$ for all $hv \in Hist_i(v_0)$. It is a \emph{finite-memory} strategy if it can be encoded by a deterministic \emph{Moore machine} ${\cal M} = (M, m_0, \alpha_U, \alpha_N)$ where $M$ is a finite set of states (the memory of the strategy), $m_0 \in M$ is an initial memory state, $\alpha_U : M \times V \rightarrow M$ is an update function, and $\alpha_N : M \times V_i \rightarrow V$ is a next-move function.\footnote{Moore machines are usually defined for finite sets $V$ of vertices. We here allow infinite sets $V$.} Such a machine defines a strategy $\sigma$ such that $\sigma(hv) = \alpha_N(\widehat{\alpha}_U(m_0,h),v)$ for all histories $hv \in Hist_i(v_0)$, where $\widehat{\alpha}_U$ extends $\alpha_U$ to histories as expected. The \emph{memory size} of $\sigma$ is then the size $|M|$ of $\cal M$. In particular $\sigma$ is positional when it has memory size one. The previous definitions of (positional, finite-memory) strategy are given for an initialized game $(G, v_0)$. We call \emph{uniform} every positional strategy $\sigma$ of player~$i$ defined for all $hv \in Hist_i$ (instead of $Hist_i(v_0)$), that is, when $\sigma$ is a positional strategy in all initialized games $(G,v)$, $v \in V$. A play $\rho$ is \emph{consistent} with a strategy $\sigma$ of player~$i$ if $\rho_{n+1} = \sigma(\rho_{\leq n})$ for all $n$ such that $\rho_n \in V_i$. A \emph{strategy profile} is a tuple $\bar\sigma = (\sigma_i)_{i \in \Pi}$ of strategies, where each $\sigma_i$ is a strategy of player~$i$. It is called \emph{positional} (resp. \emph{finite-memory with memory size bounded by $c$}, \emph{uniform}) if all $\sigma_i$, $i \in \Pi$, are positional (resp. finite-memory with memory size bounded by $c$, uniform). Given an initial vertex $v_0$, such a strategy profile determines a unique play of $(G, v_0)$ that is consistent with all the strategies. This play induced by $\bar\sigma$ in $(G,v_0)$ is denoted by $\out{\bar\sigma}_{v_0}$ and we say that $\bar\sigma$ has outcome $\mu(\out{\bar\sigma}_{v_0})$. Let $\bar \sigma$ be a strategy profile. When all players stick to their own strategy except player $i$ that shifts from $\sigma_i$ to $\sigma'_i$, we denote by $(\sigma'_i, \bar \sigma_{-i})$ the derived strategy profile, and by $\out{\sigma'_i, \bar \sigma_{-i}}_{v_0}$ the induced play in $(G, v_0)$. We say that $\sigma'_i$ is a \emph{deviating} strategy from $\sigma_i$. When $\sigma_i$ and $\sigma'_i$ only differ on a finite number of histories (resp. on $v_0$), we say that $\sigma'_i$ is a \emph{finitely-deviating} (resp. \emph{one-shot deviating}) strategy from $\sigma_i$. One-shot deviating strategies is a well-known notion that for instance appears in the proof of Kuhn's theorem~\cite{kuhn53} with the one-step deviation property. Finitely-deviating strategies have been introduced in \cite{BBMR15}. \subsection{Variants of subgame perfect equilibria} In this section we recall the notion of subgame perfect equilibrium (SPE) and its variants. Let us first recall the classical notion of Nash equilibrium (NE). Informally, a strategy profile $\bar\sigma$ in an initialized game $(G,v_0)$ is an NE if no player has an incentive to deviate (with respect to his preference relation), if the other players stick to their strategies. \begin{definition} Given an initialized game $(G, v_0)$, a strategy profile $\bar \sigma = (\sigma_i)_{i \in \Pi}$ of $(ùG,v_0)$ is a \emph{Nash equilibrium} if for all players $i \in \Pi$, for all strategies $\sigma'_i$ of player~$i$, we have $\mu(\out{\bar \sigma}_{v_0}) \nprec_i \mu(\out{\sigma'_i, \bar \sigma_{-i}}_{v_0})$. \end{definition} When $\mu(\out{\bar \sigma}_{v_0}) \prec_i \mu(\out{\sigma'_i, \bar \sigma_{-i}}_{v_0})$, we say that $\sigma'_i$ is a \emph{profitable deviation} for player~$i$ w.r.t. $\bar\sigma$. The notion of subgame perfect equilibrium is a refinement of NE. In order to define it, we need to introduce the following concepts. Given a game $G = (\Pi, V, (V_i)_{i \in \Pi}, E, \mu, (\prec_i)_{i \in \Pi})$ and a history $h \in Hist$, we denote by $\Sub{G}{h}$ the game $(\Pi, V, (V_i)_{i \in \Pi}, E, \Sub{\mu}{h}, (\prec_i)_{i \in \Pi})$ where $\Sub{\mu}{h}(\rho) = \mu(h\rho)$ for all plays of $\Sub{G}{h}$\footnote{In this article, we will always use notation $\mu(h\rho)$ instead of $\Sub{\mu}{h}(\rho)$.}, and we say that $\Sub{G}{h}$ is a \emph{subgame} of $G$. Given an initialized game $(G, v_0)$ and a history $hv \in Hist(v_0)$, the initialized game $(\Sub{G}{h}, v)$ is called the subgame of $(G, v_0)$ with history $hv$. In particular $(G, v_0)$ is a subgame of itself with history $hv_0$ such that $h = \epsilon$. Given a strategy $\sigma$ of player $i$ in $(G,v_0)$, the strategy $\Sub{\sigma}{h}$ in $(\Sub{G}{h}, v)$ is defined as $\Sub{\sigma}{h}(h') = \sigma(hh')$ for all histories $h' \in Hist_i(v)$. Given a strategy profile $\bar \sigma$ in $(G,v_0)$, we use notation $\Sub{\bar \sigma}{h}$ for $(\Sub{\sigma_i}{h})_{i \in \Pi}$, and $\out{\Sub{\bar \sigma}{h}}_{v}$ is the play induced by $\Sub{\bar \sigma}{h}$ in the subgame $(\Sub{G}{h}, v)$. We can now recall the classical notion of subgame perfect equilibrium: an SPE is a strategy profile in an initialized game that induces an NE in each of its subgames. Two variants of SPE, called weak SPE and very weak SPE, are proposed in~\cite{BBMR15} such that no player has an incentive to deviate in any subgame using finitely deviating strategies and one-shot deviating strategies respectively (instead of any deviating strategy). \begin{definition} Given an initialized game $(G, v_0)$, a strategy profile $\bar \sigma$ of $(G,v_0)$ is a \emph{(weak, very weak resp.) subgame perfect equilibrium} if for all histories $hv \in Hist(v_0)$, for all players $i \in \Pi$, for all (finitely, one-shot resp.) deviating strategies $\sigma'_i$ from $\Sub{\sigma_i}{h}$ of player~$i$ in the subgame $(\Sub{G}{h}, v)$, we have $\mu(\out{\Sub{\bar \sigma}{h}}_{v}) \nprec_i \mu(\out{\sigma'_i, \bar\sigma_{-i | h}}_{v})$. \end{definition} Trivially, every SPE is a weak SPE, and every weak SPE is a very weak SPE. \begin{proposition}[\cite{BBMR15}] \label{prop:weak-veryweak} Let $\bar \sigma$ be a strategy profile in $(G, v_0)$. Then $\bar \sigma$ is a weak SPE iff $\bar \sigma$ is a very weak SPE. There exists an initialized game $(G, v_0)$ with a weak SPE but no SPE. \end{proposition} \begin{figure}[ht!] \begin{center} \begin{tikzpicture}[initial text=,auto, node distance=2cm, shorten >=1pt] \node[state, scale=0.6] (1) {$v_0$}; \node[state, rectangle, scale=0.6] (2) [right=of 1] {$v_1$}; \node[state, scale=0.6] (3) [left=of 1] {$v_2$}; \node[state, scale=0.6] (4) [right=of 2] {$v_3$}; \path[->] (1) edge [bend right=25, thick, black] node[below, scale=0.7, black] {} (2) edge node[above, scale=0.7] {} (3) (2) edge [bend left=-25] node[above, scale=0.7] {} (1) edge [thick, black] node[above, scale=0.7, black] {} (4) (3) edge [loop above, thick, black] node[midway, scale=0.7, black] {} () (4) edge [loop above, thick, black] node[midway, scale=0.7, black] {} (); \end{tikzpicture} \end{center} \caption{A initialized game $(G,v_0)$ with a (very) weak SPE and no SPE.} \label{fig:gameNoSPE} \end{figure} \begin{example}[\cite{BBMR15}] \label{ex:contrex} Consider the two-player game $(G, v_0)$ in Figure~\ref{fig:gameNoSPE} such that player~$1$ (resp. player~$2$) controls vertices $v_0, v_2, v_3$ (resp. vertex $v_1$). The set $O$ of outcomes is equal to $\{o_1, o_2, o_3 \}$, and the outcome function is prefix-independent such that $\mu((v_0v_1)^\omega) = o_1$, $\mu(v_2^\omega) = o_2$, and $\mu(v_3^\omega) = o_3$. The preference relation for player~$1$ (resp. player~$2$) is $o_1 \prec_1 o_2 \prec_1 o_3$ (resp. $o_2 \prec_2 o_3 \prec_2 o_1$). It is known that this game has no SPE~\cite{SV03}. Nevertheless the positional strategy profile $\bar \sigma$ depicted with thick edges is a very weak SPE, and thus a weak SPE by Proposition~\ref{prop:weak-veryweak}. Let us give some explanation. Due to the simple form of the game, only two cases are to be treated. Consider first the subgame $(\Sub{G}{h}, v_0)$ with $h \in (v_0v_1)^\ast$, and the one-shot deviating strategy $\sigma'_1$ from $\Sub{\sigma_1}{h}$ such that $\sigma'_1(v_0) = v_2$. Then $\out{\Sub{\bar\sigma}{h}}_{v_0} = v_0v_1v_3^\omega$ and $\out{\sigma'_1,\Sub{\sigma_2}{h}}_{v_0} = v_0v_2^\omega$ with respective outcomes $o_3$ and $o_2$, showing that $\sigma'_1$ is not a profitable deviation for player~$1$ in $(\Sub{G}{h}, v_0)$. Now in the subgame $(\Sub{G}{h}, v_1)$ with $h \in (v_0v_1)^\ast v_0$, the one-shot deviating strategy from $\Sub{\sigma_2}{h}$ such that $\sigma'_2(v_1) = v_0$ is not profitable for player~$2$ in $(\Sub{G}{h}, v_1)$ because $\out{\Sub{\bar\sigma}{h}}_{v_1} = v_1v_3^\omega$ and $\out{\Sub{\sigma_1}{h},\sigma'_2}_{v_1} = v_1v_0v_1v_3^\omega$ with the same outcome $o_3$. Notice that $\bar \sigma$ is not an SPE. Indeed the strategy $\sigma'_2$ such that $\sigma'_2(hv_1) = v_0$ for all $h$, is infinitely deviating from $\sigma_2$, and is a profitable deviation for player~$2$ in $(G, v_0)$ since $\out{\sigma_1,\sigma'_2}_{v_0} = (v_0v_1)^\omega$ with outcome~$o_1$. \end{example} \section{General conditions for the existence of weak SPEs} \label{sec:general} In this section, we propose general conditions to guarantee the existence of weak SPEs. In the next sections, from this result, we will derive two interesting large families of games always having a weak SPE. \begin{theorem} \label{thm:generalgraph} Let $(G,v_0)$ be an initialized game with a subset $L \subseteq V$ of vertices called \emph{leaves} with only one outgoing edge $(l,l)$ for all $l \in L$. Suppose that: \begin{enumerate} \item for all $v \in V$, there exists a play $\rho = hl^{\omega}$ for some $h \in Hist(v)$ and $l \in L$, \item for all plays $\rho = hl^\omega$ with $h \in Hist(v)$ and $l \in L$, $\mu(\rho) = \mu(l^\omega)$, \item the set of outcomes $O_L = \{\mu(l^{\omega}) \mid l \in L\}$ is finite. \end{enumerate} Then there always exists a weak SPE $\bar \sigma$ in $(G, v_0)$. Moreover, $\bar \sigma$ is finite-memory with memory size bounded by $|O_L|$. \end{theorem} Let us comment the hypotheses. The first condition means that from each vertex $v$ of the game there is a leaf reachable from $v$; in particular $L$ is not empty. The second condition expresses prefix-independence of the outcome function restricted to plays eventually looping in a leaf $l \in L$. The last condition means that even if there is an infinite number of leaves, the set of outcomes assigned by $\mu$ to plays eventually looping in $L$ is finite. The next example describes a family of games satisfying the conditions of Theorem~\ref{thm:generalgraph}. \begin{example} \label{ex:Gn} For each natural number $n \geq 3$, we build a game $G_n$ with $n$ players, $2n$ vertices, $3n$ edges, and $n+1$ outcomes. The set of players is $\Pi = \{1,2, \ldots, n\}$ and the set of vertices is $V = \{v_1,\ldots,v_n, l_1, \ldots l_n \}$ such that $V_i = \{v_i, l_i\}$ for all $i \in \Pi$. The edges are $(v_1, v_2), (v_2,v_3), \ldots, (v_n,v_1)$, and $(v_i, l_i), (l_i,l_i)$ for all $i \in \Pi$. The game $G_4$ is depicted in Figure~\ref{fig:G4}. The set $O$ of outcomes is equal to $\{o_1, \ldots, o_n, \bot \}$, and the outcome function is prefix-independent such that $\mu((v_1v_2 \ldots v_n)^\omega) = \bot$ and $\mu(l_i^\omega) = o_i$ for all $i \in \Pi$. Each player $i$ has a preference relation $\prec_i$ satisfying $\bot \prec_i o_{i-1} \prec_i o_i \prec_i o_j$ for all $j \in \Pi \setminus \{i-1,i\}$ (with the convention that $o_0 = o_n$). \begin{figure} \centering \begin{tikzpicture}[shorten >=1pt,node distance=2cm, auto] \node[state, scale=0.6] (a1) {$v_1$}; \node[state, scale=0.6] (a2) [right of = a1] {$v_2$}; \node[state, scale=0.6] (a3) [below of = a2] {$v_3$}; \node[state, scale=0.6] (a4) [left of = a3] {$v_4$}; \node[state, scale=0.6] (x1) [above left of = a1]{$l_1$}; \node[state, scale=0.6] (x2) [above right of = a2]{$l_2$}; \node[state, scale=0.6] (x3) [below right of = a3]{$l_3$}; \node[state, scale=0.6] (x4) [below left of = a4]{$l_4$}; \path[->] (a1) edge node {} (a2) (a2) edge node {} (a3) (a3) edge node {} (a4) (a4) edge node {} (a1) (a1) edge node {} (x1) (a2) edge node {} (x2) (a3) edge node {} (x3) (a4) edge node {} (x4) (x1) edge [loop above] node {} () (x2) edge [loop above] node {} () (x3) edge [loop below] node {} () (x4) edge [loop below] node {} (); \end{tikzpicture} \caption{Game $G_4$} \label{fig:G4} \end{figure} Each game $(G_n,v_1)$ satisfies the hypotheses of Theorem~\ref{thm:generalgraph} with $L = \{l_1,\ldots,l_n\}$ and thus has a finite-memory weak SPE. Such a strategy profile $\bar \sigma$ is depicted in Figure~\ref{fig:profileG4} for $n = 4$ (see the thick edges on the unravelling of $G_4$ from the initial vertex $v_1$) and can be easily generalized to every $n \geq 3$. One verifies that this profile is a very weak SPE, and thus a weak SPE by Proposition~\ref{prop:weak-veryweak}. For all $i \in \Pi$, the strategy $\sigma_i$ of player~$i$ is finite-memory with a memory size equal to $n-1$. Intuitively, along $(v_1 \ldots v_n)^\omega$, player~$i$ repeatedly produces one move $(v_i,l_i)$ followed by $n-2$ moves $(v_i,v_{i+1})$. Hence the memory states of the Moore machine for $\sigma_i$ are counters from $1$ to $n-1$. The Moore machine for $\sigma_1$ in the game $(G_4,v_1)$ is depicted in Figure~\ref{fig:MooreG4} (with $M = \{1,2,3\}$, $m_0 = 1$, and the update and next-move functions indicated by the edges). \begin{figure} \centering \begin{tikzpicture}[shorten >=1pt,node distance=.9cm, auto] \node (a11) {$v_1$}; \node (a21) [right of = a11] {$v_2$}; \node (a31) [right of = a21] {$v_3$}; \node (a41) [right of = a31] {$v_4$}; \node (a12) [right of = a41] {$v_1$}; \node (a22) [right of = a12] {$v_2$}; \node (a32) [right of = a22] {$v_3$}; \node (a42) [right of = a32] {$v_4$}; \node (a13) [right of = a42] {$v_1$}; \node (a23) [right of = a13] {$v_2$}; \node (a33) [right of = a23] {$v_3$}; \node (a43) [right of = a33] {$v_4$}; \node (a14) [right of = a43] {$v_1$}; \node (inf) [right of = a14] {}; \node (x11) [below of = a11]{$l_1$}; \node (x21) [below of = a21] {$l_2$}; \node (x31) [below of = a31] {$l_3$}; \node (x41) [below of = a41] {$l_4$}; \node (x12) [below of = a12] {$l_1$}; \node (x22) [below of = a22] {$l_2$}; \node (x32) [below of = a32] {$l_3$}; \node (x42) [below of = a42] {$l_4$}; \node (x13) [below of = a13] {$l_1$}; \node (x23) [below of = a23] {$l_2$}; \node (x33) [below of = a33] {$l_3$}; \node (x43) [below of = a43] {$l_4$}; \node (x14) [below of = a14] {$l_1$}; \draw[->] (a11) edge node {} (a21); \draw[->] (a21) edge [thick] node {} (a31); \draw[->] (a31) edge [thick] node {} (a41); \draw[->] (a41) edge node {} (a12); \draw[->] (a12) edge [thick] node {} (a22) ; \draw[->] (a22) edge [thick] node {} (a32); \draw[->] (a32) edge node {} (a42); \draw[->] (a42) edge [thick]node {} (a13); \draw[->] (a13) edge [thick]node {} (a23); \draw[->] (a23) edge node {} (a33); \draw[->] (a33) edge [thick]node {} (a43); \draw[->] (a43) edge [thick] node {} (a14); \draw[dashed] (a14) edge node {} (inf); \draw[->] (a11) edge [thick] node {} (x11); \draw[->] (a21) edge node {} (x21); \draw[->] (a31) edge node {} (x31); \draw[->] (a41) edge [thick] node {} (x41); \draw[->] (a12) edge node {} (x12); \draw[->] (a22) edge node {} (x22); \draw[->] (a32) edge [thick] node {} (x32); \draw[->] (a42) edge node {} (x42); \draw[->] (a13) edge node {} (x13); \draw[->] (a23) edge [thick] node {} (x23); \draw[->] (a33) edge node {} (x33); \draw[->] (a43) edge node {} (x43); \draw[->] (a14) edge [thick] node {} (x14); \end{tikzpicture} \caption{Weak SPE in $(G_4,v_1)$} \label{fig:profileG4} \end{figure} \begin{figure} \centering \begin{tikzpicture}[initial text=, auto, shorten >=1pt,node distance=2cm, auto] \node[state, scale=0.6] (1) {$1$}; \node[state, scale=0.6] (2) [right=2.5cm of 1] {$2$}; \node (fictif) [right=1.25cm of 1] {}; \node[state, scale=0.6] (3) [below=1cm of fictif] {$3$}; \path[->] (1) edge node [scale=0.7] {$v_1/l_1$} (2) (2) edge node [scale=0.7] {$v_1/v_2$} (3) (3) edge node [scale=0.7] {$v_1/v_2$} (1) (1) edge [loop above] node [scale=0.7] {$v_2,v_3,v_4$} () (2) edge [loop above] node [scale=0.7] {$v_2,v_3,v_4$} () (3) edge [loop below] node [scale=0.7] {$v_2,v_3,v_4$} (); \end{tikzpicture} \caption{The Moore machine for $\sigma_1$} \label{fig:MooreG4} \end{figure} \end{example} Let us now proceed to the proof of Theorem~\ref{thm:generalgraph}. Recall that it is enough to prove the existence of a very weak SPE by Proposition~\ref{prop:weak-veryweak}. The proof idea is the following one. Initially, for each vertex $v$, we accept all plays $\rho = hl^{\omega}$ with $h \in Hist(v)$ and $l \in L$ as \emph{potential} plays induced by a very weak SPE in the initialized game $(G,v)$. We thus label each $v$ by the set of outcomes $\mu(l^\omega)$ for such leaves $l$ (recall that $\mu(\rho) = \mu(l^\omega)$ by the second condition of Theorem~\ref{thm:generalgraph}). Notice that this labeling is finite (resp. not empty) by the third (resp. first) condition of the theorem. Step after step, we are going to remove some outcomes from the vertex labelings by a \emph{\textit{Remove}} operation followed by an \emph{\textit{Adjust}} operation. The \textit{Remove}\ operation removes an outcome $o$ from the labeling of a given vertex $v$ when there exists an edge $(v,v')$ for which $o \prec_v o'$ for all outcomes $o'$ that label $v'$. Indeed $o$ cannot be the outcome of a play induced by a very weak SPE since the player who controls $v$ will choose the move $(v,v')$ to get a preferable outcome $o'$. Now it may happen that for another vertex $u$ having $o$ in its labeling, all potential plays induced by a very weak SPE from $u$ with outcome $o$ necessarily cross vertex $v$. As $o$ has been removed from the labeling of $v$, these potential plays do no longer survive and $o$ will also be removed from the labeling of $u$ by the \textit{Adjust}\ operation. Repeatedly applying these two operations converge to a fixpoint for which we will prove non-emptiness (this is the difficult part of the proof, non-emptiness will be obtained by maintaining three invariants, see Lemma~\ref{lem:fixpoint}). From this fixpoint, for each vertex $v$ and each outcome $o$ of the resulting labeling of $v$, there exists a play $\rho_{v,o}= hl^{\omega}$ with outcome $o$ for some $h \in Hist(v)$ and $l \in L$. We can thus build a very weak SPE $\bar \sigma$ in $(G,v_0)$ as follows. The construction of $\bar \sigma$ is done step by step: \emph{(i)} initially $\bar \sigma$ is partially defined such that $\out{\bar \sigma}_{v_0} = \rho_{v_0,o_0}$ for some $o_0$; \emph{(ii)} then in the subgame $(\Sub{G}{h},v)$ such that $\out{\Sub{\bar \sigma}{h}}_{v} = \rho_{v,o}$, if the player who controls $v$ chooses the move $(v,v')$ in a one-shot deviation, then there exists $\rho_{v',o'}$ such that $o \nprec_v o'$ by definition of the fixpoint, and we thus extend the construction of $\bar \sigma$ such that $\out{\Sub{\bar \sigma}{hv}}_{v'} = \rho_{v',o'}$. Let us now go into the details of the proof. For each $l \in L$, we denote by $o_l$ the outcome $\mu(l^{\omega})$. Recall that for all $\rho = hl^{\omega}$ we have $\mu(\rho) = o_l$ by the second hypothesis of the theorem. For each $v \in V$, we denote by $Succ(v)$ the set of successors of $v$ distinct from $v$, that is, the vertices $v' \neq v$ such that $(v,v') \in E$. Notice that the leaves $l$ are the vertices with only one outgoing edge $(l,l)$. Thus, by definition, $Succ(v) = \emptyset$ for all $v \in L$ and $Succ(v) \neq \emptyset$ for all $v \in V \setminus L$. The labeling $\lambda_{\alpha}(v)$ of the vertices $v$ of $G$ by subsets of $O_L$ is an inductive process on the ordinal $\alpha$. Initially (step $\alpha = 0$), each $v \in V$ is labeled by: $$\lambda_0(v) = \{o_l \in O_L \mid \mbox{there exists a play } hl^{\omega} \mbox{ with } h \in Hist(v) \mbox{ and } l \in L\}.$$ (In particular $\lambda_0(l) = \{o_l\}$ for all $l \in L$). By the first hypothesis of the theorem, $\lambda_0(v) \neq \emptyset$. Let us introduce some additional terminology. At step $\alpha$, when there is a path\footnote{By path, we mean a finite path} $\pi$ from $v$ to $v'$ in $G$, we say that $\pi$ is \emph{$(o,\alpha)$-labeled} if $o \in \lambda_{\alpha}(u)$ for all the vertices $u$ of $\pi$. Thus initially, we have a $(o_l,0)$-labeled path from $v$ to $l$ for each $o_l \in \lambda_0(v)$. For $v \in V$, let $$m_{\alpha}(v) = \mbox{$\max_{\prec_{v}}$} \{ \mbox{$\min_{\prec_{v}}$} \lambda_{\alpha}(v') \mid v' \in Succ(v)\}$$ with the convention that $m_\alpha(v) = \top$ if $Succ(v) = \emptyset$ or if $\lambda_{\alpha}(v') = \emptyset$ for all $v' \in Succ(v)$.\footnote{We suppose that $o \prec_v \top$ for all $o \in O_L$.} When $m_\alpha(v) \neq \top$, we says that $v' \in Succ(v)$ \emph{realizes} $m_{\alpha}(v)$ if $m_{\alpha}(v) =\min_{\prec_{v}} \lambda_{\alpha}(v')$. Notice that even if $Succ(v)$ could be infinite, there are finitely many sets $\lambda_{\alpha}(v')$ since $O_L$ is finite. This justifies our use of $\max_{\prec_{v}}$ and $\min_{\prec_{v}}$ operators in the definition of $m_{\alpha}(v)$. We alternate between applying \textit{Remove}\ and \textit{Adjust}\ to the current labeling. More formally, we define the labeling $\lambda_\alpha$ inductively\footnote{Note that our definition as written makes non-deterministic choices. This is immaterial for our purposes, but could be determinized by demanding a well-ordering of the vertex set and the outcomes.}. In the following, $\gamma$ is always assumed to be a limit ordinal and $n$ to be a natural number. \begin{itemize} \item {\bf Defining $\lambda_{\gamma + 2n + 1}$ via \textit{Remove}\ operation} Let $\alpha := \gamma + 2n + 1$. Test if for some $v \in V$, there exist $o \in \lambda_{\alpha - 1}(v)$ and $v' \in Succ(v)$ such that $$o \prec_v o', \mbox{ for all } o' \in \lambda_{\alpha - 1}(v').$$ If such a $v$ exists, then $\lambda_{\alpha}(v) = \lambda_{\alpha - 1}(v) \setminus \{o\}$, and $\lambda_{\alpha}(u) = \lambda_{\alpha - 1}(u)$ for the other vertices $u \neq v$. Otherwise $\lambda_{\alpha}(u) = \lambda_{\alpha - 1}(u)$ for all $u \in V$. \item {\bf Defining $\lambda_{\gamma + 2n + 2}$ via \textit{Adjust}\ operation} Let $\alpha := \gamma + 2n + 2$. Suppose that $\lambda_{\alpha - 1}(v) = \lambda_{\alpha - 2}(v) \setminus \{o\}$ at the previous step. For all $u \in V$ such that $o \in \lambda_{\alpha - 1}(u)$, test if there exists a $(o,\alpha - 1)$-labeled path from $u$ to some $l \in L$. If yes, then $\lambda_{\alpha}(u) = \lambda_{\alpha - 1}(u)$, otherwise $\lambda_{\alpha}(u) = \lambda_{\alpha - 1}(u) \setminus \{o\}$. For all $u \in V$ such that $o \not\in \lambda_{\alpha-1}(u)$, let $\lambda_{\alpha}(u) = \lambda_{\alpha-1}(u)$. Suppose that $\lambda_{\alpha-1}(v) = \lambda_{\alpha-2}(v)$ for all $v \in V$ at the previous step, then $\lambda_{\alpha}(v) = \lambda_{\alpha-1}(v)$ for all $v \in V$. \item {\bf Defining $\lambda_\gamma$ via intersection} Let $\lambda_{\gamma}(v) = \cap_{\beta < \gamma} \lambda_{\beta}(v)$ for all $v \in V$. \end{itemize} For each $v$, the sequence $(\lambda_\alpha(v))_{\alpha}$ is nonincreasing (w.r.t.~set inclusion), and thus the sequence $(m_\alpha(v))_{\alpha}$ is nondecreasing (w.r.t.~$\prec_v$). Moreover, the sequence $(\lambda_\alpha)_\alpha$ is nonincreasing w.r.t.~pointwise set inclusion. Thus, there exists some ordinal $\alpha^*$ such that $\lambda_{\alpha^*} = \lambda_\beta$ for all $\beta > \alpha^*$. By inspecting the definition, we see that it suffices to check that $\lambda_{\alpha^*} = \lambda_{\alpha^*+1} = \lambda_{\alpha^*+2}$ in order to see that $\alpha^*$ is a fixed point. If $V$ is finite, such a fixed point is reached after at most $2 |O_L| \cdot |V|$ steps. The central challenge is to show that this fixed point is non-empty in each component. Notice that for all leaves $l \in L$ and all steps $\alpha$, we have $\lambda_\alpha(l) = \{o_l\}$. \begin{lemma} \label{lem:fixpoint} There exists an ordinal $\alpha^*$ such that $$\lambda_{\alpha^*}(v) = \lambda_{\alpha^*+1}(v) = \lambda_{\alpha^*+2}(v) \mbox{ for all } v \in V.$$ Moreover, $\lambda_{\alpha^*}(v) \neq \emptyset$ for all $v \in V$. \end{lemma} To be able to prove that $\lambda_{\alpha^*}(v) \neq \emptyset$, we introduce three invariants for which we will prove that they are initially true (Lemma~\ref{lemma:invinit}) and remain true after each step~$\alpha$ (Lemmata \ref{lemma:invremove},\ref{lemma:invadjust},\ref{lemma:invlimit}). The non emptiness of $\lambda_{\alpha^*}(v)$ will follow from the second invariant. \begin{description} \item[INV1] For $v \in V$, we have for all $v' \in Succ(v)$ that $$\{ o \in \lambda_{\alpha}(v') \mid m_{\alpha}(v) \preceq_{v} o \} \subseteq \lambda_{\alpha}(v).$$ In particular, when $m_{\alpha}(v) \neq \top$, for each $v'$ that realizes $m_{\alpha}(v)$, we have \begin{eqnarray}\label{a(v)} \lambda_{\alpha}(v') \subseteq \lambda_{\alpha}(v). \end{eqnarray} \item[INV2] For $v \in V$, $\lambda_{\alpha}(v) \neq \emptyset$. \item[INV3] For $v \in V$, there exists a path from $v$ to some $l \in L$ such that for all vertices $u$ in this path, $\lambda_{\alpha}(u) \subseteq \lambda_{\alpha}(v)$. \end{description} \begin{lemma} \label{lemma:invinit} All three invariants are true for $\lambda_0$. \end{lemma} \begin{proof} Consider $v \in V$ at the initial step $\alpha = 0$. By hypothesis there is a path from $v$ to some $l \in L$. Thus $\lambda_{\alpha}(v) \neq \emptyset$ and INV2 is true. Moreover, for all $v' \in Succ(v)$, we have $\lambda_{\alpha}(v') \subseteq \lambda_{\alpha}(v)$ by the initial labeling, and thus INV1 and INV3 are also true. \qed\end{proof} \begin{lemma} \label{lemma:invremove} All three invariants are preserved by \textit{Remove}. \end{lemma} \begin{proof} Consider some $\alpha = \gamma + 2n$ for limit ordinal $\gamma$ and $n \in \mathbb{N}$ such that all invariants hold for $\lambda_\alpha$. If $\lambda_{\alpha+1} = \lambda_{\alpha}$, then trivially, all invariants hold for $\lambda_{\alpha+1}$. Otherwise there exist $v$ and $o$ such that $\lambda_{\alpha+1}(v) = \lambda_{\alpha}(v) \setminus \{o\}$ and $\lambda_{\alpha+1}(u) = \lambda_{\alpha}(u)$ for all $u \neq v$. In particular $v \notin L$. For all $u \in V$, we have $m_{\alpha}(u) \preceq_{u} m_{\alpha +1}(u)$, with the particular case $m_{\alpha}(v) = m_{\alpha+1}(v)$. \begin{itemize} \item {\bf \textit{Remove}\ cannot violate INV1}. We first consider $u \in V$ such that $u \neq v$. For all $u' \in Succ(u)$, we have $$\begin{array}{llll} \{o' \in \lambda_{\alpha+1}(u') \mid m_{\alpha+1}(u) \preceq_u o' \}\\ \subseteq \{o' \in \lambda_{\alpha}(u') \mid m_{\alpha}(u) \preceq_u o' \} & \mbox{since $\lambda_{\alpha+1}(u') \subseteq \lambda_{\alpha}(u')$} \\ & \mbox{and $m_{\alpha}(u) \preceq_{u} m_{\alpha +1}(u)$,}\\ \subseteq \lambda_{\alpha}(u) & \mbox{by INV1 at step $\alpha$,} \\ = \lambda_{\alpha+1}(u) & \mbox{as $u \neq v$.} \end{array}$$ Let us turn to vertex $v$. As $o \prec_v m_{\alpha}(v)$, the previous inclusions can be modified as follows. For all $v' \in Succ(v)$, we now have $\{o' \in \lambda_{\alpha+1}(v') \mid m_{\alpha+1}(v) \preceq_v o' \} \subseteq \{o' \in \lambda_{\alpha}(v') \mid m_{\alpha}(v) \preceq_v o' \} \subseteq \lambda_{\alpha}(v) \setminus \{o\} = \lambda_{\alpha+1}(v)$. \item {\bf \textit{Remove}\ cannot violate INV2}. We only have to show that $\lambda_{\alpha + 1}(v) \neq \emptyset$. As $Succ(v) \neq \emptyset$\footnote{Recall that $v \not\in L$, and that $Succ(v) \neq \emptyset$ for all $v \in V \setminus L$.} and by INV2, we have $m_{\alpha}(v) \neq \top$. Hence there exists $v' \in Succ(v)$ that realizes $m_{\alpha}(v) = m_{\alpha+1}(v)$. By INV1 and in particular $(\ref{a(v)})$ at step $\alpha +1$, we thus have $\lambda_{\alpha+1}(v') \subseteq \lambda_{\alpha+1}(v)$. As $\lambda_{\alpha+1}(v') = \lambda_{\alpha}(v') \neq \emptyset$, it follows that $\lambda_{\alpha+1}(v) \neq \emptyset$. \item {\bf \textit{Remove}\ cannot violate INV3}. We first consider $u \neq v$. By INV3, there exists a path $\pi$ from $u$ to some $l \in L$ such that $\lambda_{\alpha}(w) \subseteq \lambda_{\alpha}(u)$ for all vertices $w$ in this path. We can keep the path $\pi$ at step $\alpha+1$ since $\lambda_{\alpha+1}(w) \subseteq \lambda_{\alpha}(w)$ for all $w$ in $\pi$ and $\lambda_{\alpha+1}(u) = \lambda_{\alpha}(u)$. We now consider vertex $v$. Consider again $v' \in Succ(v)$ that realizes $m_{\alpha+1}(v)$. By $(\ref{a(v)})$, $\lambda_{\alpha+1}(v') \subseteq \lambda_{\alpha+1}(v)$. We know that there exists a path $\pi$ from $v'$ to some $l \in L$ such that $\lambda_{\alpha}(w) \subseteq \lambda_{\alpha}(v')$ for all $w$ in $\pi$. This path $\pi$ augmented with the edge $(v,v')$ is the required path for INV3 at step $\alpha+1$ because for all $w$ in $\pi$, we have $\lambda_{\alpha+1}(w) \subseteq \lambda_{\alpha}(w) \subseteq \lambda_{\alpha}(v') = \lambda_{\alpha+1}(v') \subseteq \lambda_{\alpha+1}(v)$. \end{itemize} \qed\end{proof} \begin{lemma} \label{lemma:invadjust} All three invariants are preserved by \textit{Adjust}. \end{lemma} \begin{proof} Let all three invariants hold for $\alpha = \gamma + 2n + 1$ for limit ordinal $\gamma$ and $n \in \mathbb{N}$. Then the preceding step was a \textit{Remove}\ step. If $\lambda_{\alpha} = \lambda_{\alpha-1}$, then $\lambda_{\alpha+1} = \lambda_\alpha$. Otherwise, there are $v_0 \in V$ and an outcome $o$ such that $\lambda_{\alpha}(v_0) = \lambda_{\alpha-1}(v_0) \setminus \{o\}$ and $\lambda_\alpha(u) = \lambda_{\alpha-1}(u)$ for all $u \neq v_0$. For all $v \in V$, either $\lambda_{\alpha+1}(v) = \lambda_{\alpha}(v)$ or $\lambda_{\alpha+1}(v) = \lambda_{\alpha}(v) \setminus \{o\}$, and $m_{\alpha}(v) \preceq_v m_{\alpha+1}(v)$. Consider $v \in V$ such that $o \notin \lambda_{\alpha+1}(v)$ and $o \in \lambda_{\alpha}(v)$. Then \begin{eqnarray} \label{son} \forall v' \in Succ(v),~ o \not\in \lambda_{\alpha+1}(v') \end{eqnarray} Otherwise if $o \in \lambda_{\alpha+1}(v')$ for some $v' \in Succ(v)$, this means that $o$ has not been removed from $\lambda_{\alpha}(v')$, i.e., there exists a $(o,\alpha)$-labeled path from $v'$ to some $l \in L$, and thus also from $v$ to $l$ by using the edge $(v,v')$. This is in contradiction with $o$ being removed from $\lambda_{\alpha}(v)$. \begin{itemize} \item {\bf \textit{Adjust}\ cannot violate INV1}. We first consider $v \in V$ such that $\lambda_{\alpha+1}(v) = \lambda_{\alpha}(v)$. As done for INV1 and \textit{Remove}, we have for all $v' \in Succ(v)$ that $\{o' \in \lambda_{\alpha+1}(v') \mid m_{\alpha+1}(v) \preceq_v o' \} \subseteq \{o' \in \lambda_{\alpha}(v') \mid m_{\alpha}(v) \preceq_v o' \} \subseteq \lambda_{\alpha}(v) = \lambda_{\alpha+1}(v)$. We now consider $v \in V$ such that $\lambda_{\alpha+1}(v) \neq \lambda_{\alpha}(v)$. Let $v' \in Succ(v)$. From $(\ref{son})$, we have $\{o' \in \lambda_{\alpha+1}(v') \mid m_{\alpha+1}(v) \preceq_v o' \} \subseteq \{o' \in \lambda_{\alpha}(v') \mid m_{\alpha}(v) \preceq_v o' \} \setminus \{o \} \subseteq \lambda_{\alpha}(v) \setminus\{o\} = \lambda_{\alpha+1}(v)$. \item {\bf \textit{Adjust}\ cannot violate INV2}. Assume that for some $v \in V$, $\lambda_{\alpha+1}(v) = \emptyset$, that is, $\lambda_{\alpha}(v) = \{o\}$. By INV3, there exists a path $\pi$ from $v$ to some $l \in L$ such that $\lambda_{\alpha}(u) \subseteq \lambda_{\alpha}(v)$ for all $u$ in $\pi$. From $\lambda_{\alpha}(v) = \{o\}$ and $\lambda_{\alpha}(u) \neq \emptyset$ (by INV2), we get $\lambda_{\alpha}(u) = \{o\}$ for all such $u$. Therefore, the path $\pi$ from $v$ to $l$ is $(o,\alpha)$-labeled and $o$ cannot be removed from $\lambda_{\alpha}(v)$, showing that $\lambda_{\alpha+1}(v) \neq \emptyset$. \item {\bf \textit{Adjust}\ cannot violate INV3}. Let $v \in V$ and by INV3 take a path $u_1 \ldots u_n$ from $v = u_1$ to some $l = u_n$ with $l \in L$ such that $\lambda_{\alpha}(u_i) \subseteq \lambda_{\alpha}(v)$ for all $i$. Either this path is still valid at step $\alpha+1$, or there exists a smallest $i$ such that $o \in \lambda_{\alpha+1}(u_i) = \lambda_{\alpha}(u_i)$, but $o \in \lambda_{\alpha}(v)$ and $o \not\in \lambda_{\alpha+1}(v)$. By minimality of $i$, $o \not\in \lambda_{\alpha+1}(u_j)$ for all $j \leq i-1$. By the contraposition of (\ref{son}) with $u_{i-1}$ and $u_{i}$, knowing that $o \not\in \lambda_{\alpha+1}(u_{i-1})$, it follows that $o \not\in \lambda_{\alpha}(u_{i-1})$. By INV3 there is a path $\pi$ from $u_{i-1}$ to some $l' \in L$ such that for all $w$ in $\pi$, $\lambda_{\alpha}(w) \subseteq \lambda_{\alpha}(u_{i-1})$ $(\subseteq \lambda_{\alpha}(v))$. Notice that $o \not\in \lambda_{\alpha}(w)$ for all these $w$ since $o \not\in \lambda_{\alpha}(u_{i-1})$. The path $\pi'$ obtained by concatenating $u_1 \ldots u_{i-1}$ with $\pi$ is the required path from $v$ for INV3 at step $\alpha +1$. Indeed for all $w'$ in $\pi'$, we have seen that $\lambda_{\alpha}(w') \subseteq \lambda_{\alpha}(v)$ and $o \notin \lambda_{\alpha+1}(w')$. Thus $\lambda_{\alpha+1}(w') \subseteq \lambda_{\alpha}(v) \setminus \{o\}= \lambda_{\alpha+1}(v)$. \end{itemize} \qed\end{proof} \begin{lemma} \label{lemma:invlimit} If all three invariants are true for each $\lambda_\beta$, $\beta < \alpha$, $\alpha$ a limit ordinal, then they are true for $\lambda_\alpha$. \end{lemma} \begin{proof} Let $\alpha$ be a limit ordinal, and suppose that the three invariants are true for each ordinal $\beta < \alpha$. Given $v \in V$, as the set $\lambda_{\beta}(v)$ is finite\footnote{This is the place in the proof where finiteness of the number of outcomes is used in a crucial way.} and the sequence $(\lambda_{\beta}(v))_{\beta < \alpha}$ is nonincreasing, there exists some $\gamma < \alpha$ such that $\lambda_{\beta}(v) = \lambda_{\gamma}(v)$ for all $\beta$, $\gamma \leq \beta < \alpha$. Therefore \begin{eqnarray} \label{eq:gamma} \lambda_\alpha(v) = \cap_{\beta < \alpha} \lambda_{\beta}(v) = \lambda_{\gamma}(v). \end{eqnarray} It immediately follows that INV2 holds at step $\alpha$. To show that INV3 also holds, consider a path $\pi$ from $v$ to some $l \in L$ such that $\lambda_{\gamma}(u) \subseteq \lambda_{\gamma}(v)$ for all $u$ in $\pi$ (by INV3 at step $\gamma$). We can take this path $\pi$ for INV3 at step $\alpha$ since for all these $u$, we have $\lambda_{\alpha}(u) \subseteq \lambda_{\gamma}(u) \subseteq \lambda_{\gamma}(v) = \lambda_{\alpha}(v)$. Finally, the first invariant remains true at step $\alpha$ because for all $v' \in Succ(v)$, we have $$\begin{array}{llll} \{ o \in \lambda_{\alpha}(v') \mid m_{\alpha}(v) \preceq_{v} o \} \\ \subseteq \{ o \in \lambda_{\gamma}(v') \mid m_{\gamma}(v) \preceq_{v} o \} & \mbox{since $\lambda_{\alpha}(v') \subseteq \lambda_{\gamma}(v')$ and $m_{\gamma}(v) \preceq_v m_{\alpha}(v)$,} \\ \subseteq \lambda_{\gamma}(v) & \mbox{by INV1 at step $\gamma$,} \\ = \lambda_{\alpha}(v) & \mbox{by (\ref{eq:gamma}).} \end{array}$$ \qed\end{proof} To get Theorem~\ref{thm:generalgraph}, it remains to explain how to build a finite-memory weak SPE $\bar \sigma$ from the fixed point provided by Lemma~\ref{lem:fixpoint}. \begin{proof}[of Theorem~\ref{thm:generalgraph}] By Lemma~\ref{lem:fixpoint}, we have a fixed point of \textit{Remove}\ and \textit{Adjust}\ such that that $\lambda_{\alpha^*}(v) \neq \emptyset$ for all $v \in V$. Since $\lambda_{\alpha^*}$ is unchanged by \textit{Adjust}, for all $o \in \lambda_{\alpha^*}(v)$, there is a $(o,\alpha^*)$-labeled path $\pi$ from $v$ to some $l \in L$ with $o_l = o$. We denote by $\rho_{v,o}$ the play $\pi l^\omega$: \begin{eqnarray} \label{eq:rhovl} \rho_{v,o} = \pi l^\omega. \end{eqnarray} (*) Recall that $\mu(\rho_{v,o}) = o_l$, and have in mind that $o_l \in \lambda_{\alpha^*}(u)$ for all vertices $u$ in $\rho_{v,o}$. The construction of $\bar \sigma$ will be done step by step thanks to a progressive labeling of the histories by outcomes in $O_L$ and by using the plays $\rho_{v,o}$. This labeling $\kappa : Hist(v_0) \rightarrow O_L$ will allow to recover from history $hv$ the outcome $o$ of the play $\out{\Sub{\bar \sigma}{h}}_{v}$ induced by $\bar \sigma$ in the subgame $(\Sub{G}{h},v)$. We start with history $v_0$ and any $o_0 \in \lambda_{\alpha^*}(v_0)$. Consider $\rho_{v_0,o_0}$ as in (\ref{eq:rhovl}). The strategy profile $\bar \sigma$ is partially built such that $\out{\bar \sigma}_{v_0} = \rho_{v_0,o_0}$. The non empty prefixes $g$ of $\rho_{v_0,o_0}$ are all labeled with $\kappa(g) = o_0$. At the following steps, we consider a history $h'v'$ that is not yet labeled, but such that $h' = hv$ has already been labeled by $\kappa(hv) = o$. The labeling of $hv$ by $o$ means that $\bar \sigma$ has already been built to produce the play $\out{\Sub{\bar \sigma}{h}}_{v}$ with outcome $o$ in the subgame $(\Sub{G}{h},v)$, such that $\out{\Sub{\bar \sigma}{h}}_{v}$ is suffix of $\rho_{u,o}$ from some $u$. By (*) we have $o \in \lambda_{\alpha^*}(v)$. As $\lambda_{\alpha^*}$ is invariant under \textit{Remove}\ (noting $o \in \lambda_{\alpha^*}(v)$ and $v' \in Succ(v)$), there exists $o' \in \lambda_{\alpha^*}(v')$ such that \begin{eqnarray} \label{eq:chosen} o \nprec_v o'. \end{eqnarray} With $\rho_{v',o'}$ as in (\ref{eq:rhovl}), we then extend the construction of $\bar \sigma$ such that $\out{\Sub{\bar \sigma}{h'}}_{v'} = \rho_{v',o'}$, and for each non empty prefix $g$ of $\rho_{v',o'}$, we label $h'g$ by $\kappa(h'g) = o'$ (notice that the prefixes of $h'$ have already been labeled by choice of $h'$). This process is iterated to complete the construction of $\bar \sigma$. Let us show that the constructed profile $\bar \sigma$ is a very weak SPE in $(G,v_0)$. Consider a history $h' = hv \in Hist(v_0)$ with $v \in V_i$, and a one-shot deviating strategy $\sigma'_i$ from $\Sub{\sigma_i}{h}$ in the subgame $(\Sub{G}{h},v)$. Let $v'$ be such that $\sigma'_i(v) = v'$. By definition of $\bar \sigma$, we have $\kappa(hv) = o$ and $\kappa(h'v') = o'$ such that (\ref{eq:chosen}) holds. Let $\rho = \out{\Sub{\bar \sigma}{h}}_v$ and $\rho' = \out{\Sub{\bar \sigma}{h'}}_{v'}$. Then $o = \mu(h\rho)$ and $o' = \mu(hv \rho')$ by (*). By (\ref{eq:chosen}), $\sigma'_i$ is not a profitable deviation for player $i$. Hence $\bar \sigma$ is a very weak SPE and thus a weak SPE by Proposition~\ref{prop:weak-veryweak}. It remains to prove that $\bar \sigma$ is finite-memory by correctly choosing the plays $\rho_{v,o}$ of (\ref{eq:rhovl}). Fix $o \in O_L$ and consider the set $U_o$ of vertices $v$ such that $o \in \lambda_{\alpha^*}(v)$. Then we choose the plays $\rho_{v,o} = \pi l^\omega$ for all $v \in U_o$, such that the set of associated finite paths $\pi l$ forms a tree. Therefore having $o$ in memory, the required Moore machine can produce positionally each $\rho_{v,o}$ with $v \in U_o$. Hence its set $M$ of states is equal to $O_L$. \qed\end{proof} The next corollary is an easy consequence of Theorem~\ref{thm:generalgraph}. Under the same conditions except perhaps the second one, and when the underlying graph of $G$ is a tree, it guarantees the existence of a weak SPE that is positional. \begin{corollary} \label{cor:generaltree} Let $(G,v_0)$ be an initialized game with a subset $L \subseteq V$ of leaves\footnote{The existence of leaves $l$ with a unique outgoing edge $(l,l)$ is abusive since the graph is a tree: it should be understood as a unique infinite play from $l$.} such that the underlying graph is a tree rooted at $v_0$. If $(G,v_0)$ satisfies the first and third conditions of Theorem~\ref{thm:generalgraph}, then there exists a positional weak SPE in $(G, v_0)$. \end{corollary} \begin{proof} If the second condition of Theorem~\ref{thm:generalgraph} is not satisfied, we replace the outcome function $\mu$ by a new function $\mu'$ defined as follows. For all plays $l^\omega$, with $l \in L$, there is a unique path $\pi$ from $v_0$ to $l$ as the underlying graph is a tree. For all suffixes $\rho$ of $\pi l^\omega$, we let $\mu'(\rho) = \mu(\pi l^\omega)$. For all the remaining plays $\rho$, we let $\mu'(\rho) = \mu(\rho)$. With the new function $\mu'$, the game $(G,v_0)$ now satisfies all the conditions of Theorem~\ref{thm:generalgraph} and has thus a weak SPE $\bar \sigma$ with respect to $\mu'$. It is easy to see that $\bar \sigma$ is also a weak SPE with respect to $\mu$. Notice that this profile is necessarily positional as the underlying graph is a tree. \qed\end{proof} In the next two sections, we present two large families of games for which there always exists a weak SPE. We will explain how these results are obtained from Theorem~\ref{thm:generalgraph} and its Corollary~\ref{cor:generaltree}. Before that, we demonstrate the argument establishing Theorem~\ref{thm:generalgraph} on the game $G_4$ as introduced in Example \ref{ex:Gn}. \begin{example} Let us describe the inductive process for the game $G_4$ of Figure~\ref{fig:G4} (Page \pageref{fig:G4}). For all $i \in \Pi$ and all steps $\alpha$, we have $\lambda_{\alpha}(l_i) = \{o_i\}$. Table~\ref{tab:steps} indicates the different steps until reaching $\alpha^*$ for the vertices $v_i$, $i \in \Pi$, with $O_L = \{o_1,o_2,o_3,o_4\}$. For instance, at step~$1$, \textit{Remove}\ removes $o_4$ from $\lambda_\alpha(v_1)$ because $o_4 \prec_1 o'$ for all $o' \in \lambda_\alpha(l_1) = \{o_1\}$. At step~2, \textit{Adjust}\ removes no outcome. For $v = v_1$ and $o \in \lambda_\alpha(v_1)$, the plays $\rho_{v,o}$ are: $$\rho_{v_1,o_1} = v_1 l_1^{\omega}, \quad \rho_{v_1,o_2} = v_1 v_2 l_2^{\omega}, \quad \rho_{v_1,o_3} = v_1 v_2 v_3 l_3^{\omega}.$$ The other vertices $v \neq v_1$ have similar plays $\rho_{v,o}$. \begin{table} \begin{center} $\begin{array}{|c|c|c|c|c|} \hline \alpha & \lambda_\alpha(v_1) & \lambda_\alpha(v_2) & \lambda_\alpha(v_3) & \lambda_\alpha(v_4) \\ \hline 0 & O_L & O_L & O_L & O_L \\ 1 & O_L\setminus \{o_4\} & O_L & O_L & O_L \\ 2 & O_L\setminus \{o_4\} & O_L & O_L & O_L \\ 3 & O_L\setminus \{o_4\} & O_L \setminus \{o_1\} & O_L & O_L \\ 4 & O_L\setminus \{o_4\} & O_L \setminus \{o_1\} & O_L & O_L \\ 5 & O_L\setminus \{o_4\} & O_L \setminus \{o_1\} & O_L \setminus \{o_2\} & O_L \\ 6 & O_L\setminus \{o_4\} & O_L \setminus \{o_1\} & O_L \setminus \{o_2\} & O_L \\ 7 & O_L \setminus \{o_4\} & O_L \setminus \{o_1\} & O_L \setminus \{o_2\} & O_L \setminus \{o_3\} \\ \alpha^* = 8 & O_L \setminus \{o_4\} & O_L \setminus \{o_1\} & O_L \setminus \{o_2\} & O_L \setminus \{o_3\} \\ \hline \end{array}$ \end{center} \caption{The different steps until reaching a fixed point for game $G_4$} \label{tab:steps} \end{table} In the case of game $(G_4,v_1)$, the construction of a weak SPE $\bar \sigma$, as described in the previous proof, leads to the strategy profile of Figure~\ref{fig:profileG4}. Indeed, the construction of $\bar \sigma$ begins with history $v_1$ and $\rho_{v_1,o_1} = v_1 l_1^\omega$. At the next step, we consider history $v_1v_2$ and $\rho_{v_2,o_4} = v_2v_3v_4l_4^\omega$ such that $o_1 \nprec_1 o_4$, aso. Notice that the previous proof states a memory size equal to~$4$ for $\bar \sigma$ whereas Figure~\ref{fig:MooreG4} depicts a Moore machine for $\bar \sigma$ with a better memory size equal to~$3$. \end{example} \section{First application} \label{sec:first} In this section, we begin with the first application of the results of the previous section (more particularly Corollary~\ref{cor:generaltree}): when an initialized game has an outcome function with finite range, then it always has a weak SPE. \begin{theorem} \label{thm:infinitetree} Let $(G,v_0)$ be an initialized game such that the outcome function has finite range. Then there exists a weak SPE in $(G,v_0)$. \end{theorem} Let us comment this theorem. \emph{(i)} Kuhn's theorem~\cite{kuhn53} states that there always exist an SPE in initialized games played on a \emph{finite tree} (notice that in this particular case, the existence of a weak SPE is equivalent to the existence of an SPE). Theorem~\ref{thm:infinitetree} can be seen as a generalization of Kuhn's theorem: if we keep the outcome set finite, all initialized games (regardless of the underlying graph and the player set) have weak SPE. \emph{(ii)} The next theorem is proved in~\cite{Flesch10} for outcome functions $\mu = (\mu_i)_{i \in \Pi}$ as presented in Example~\ref{ex:classical} and has strong relationship with Theorem~\ref{thm:infinitetree}. Recall that a payoff function $\mu_i : Plays \to \mathbb R$ is \emph{lower-semicontinuous} if whenever a sequence of plays $(\rho_n)_{n \in \mathbb N}$ converges to a play $\rho = \lim_{n \rightarrow \infty} \rho_n$, then $\liminf_{n \rightarrow \infty} \mu_i(\rho_n) \geq \mu_i(\rho)$. \begin{theorem}[\cite{Flesch10}] Let $(G,v_0)$ be an initialized game with a finite set $\Pi$ of players and an outcome function $\mu = (\mu_i)_{i \in \Pi}$ such that each $\mu_i : Plays \to \mathbb R$ has finite range and is lower-semicontinuous. Then there exists an SPE in $(G,v_0)$. \end{theorem} \noindent As every weak SPE is an SPE in the case of lower-semicontinuous payoff functions $\mu_i$~\cite{BBMR15}, we recover the previous result with our Theorem~\ref{thm:infinitetree}. Even if it is not explicitly mentioned in~\cite{Flesch10}, a close look at the details of the proof shows that the authors first show the existence of a weak SPE (without the hypothesis of lower-semicontinuity) and then show that it is indeed an SPE (thanks to this hypothesis). The first part of their proof could be replaced by ours, which is simpler: we remove outcomes from the sets $\lambda_{\alpha}(v)$ (see the proof of Theorem~\ref{thm:generalgraph}) whereas plays are removed in the inductive process of \cite{Flesch10}. \subsection{Intermediate results} The proofs of Theorem~\ref{thm:infinitetree} in this section and Theorem~\ref{thm:finitegraph} in the next section require several intermediate results that we now describe. We begin with the next lemma where the set $\mu^{-1}(\{o\})$, with $o \in O$, is said to be \emph{dense in $(G,v_0)$} if for all $h \in Hist(v_0)$, there exists $\rho$ such that $h\rho$ is a play with outcome $\mu(h\rho) = o$. \begin{lemma} \label{lem:dense} Let $(G,v_0)$ be an initialized game. If for some $o \in O$, the set $\mu^{-1}(\{o\})$ is dense in $(G,v_0)$, then there exists a weak SPE with outcome $o$ in $(G,v_0)$. \end{lemma} \begin{proof} The construction of a very\footnote{As already done before, we apply Proposition~\ref{prop:weak-veryweak}. It will be the case in the sequel of the article without mentioning anymore this proposition.} weak SPE $\bar \sigma$ is done step by step thanks to a progressive marking of the histories $hv \in Hist(v_0)$. Let us give the construction of $\bar \sigma$. Initially, for history $v_0$, we know by density that there exists $\rho_0 \in Plays(v_0)$ with outcome $o$. We partially construct $\bar \sigma$ such that it produces $\rho_0$, and we mark each non empty prefix of $\rho_0$. Then we consider a shortest unmarked history $hv$, and we choose some $\rho \in Plays(v)$ such that $\mu(h\rho) = o$ (this is possible by density). We continue the construction of $\bar \sigma$ such that it produces the play $\rho$ in $(\Sub{G}{h},v)$, and for each non empty prefix $g$ of $\rho$, we mark $hg$ (notice that the prefixes of $h$ have already been marked by choice of $h$), and so on. In this way, we get a strategy profile $\bar \sigma$ in $(G,v_0)$ that is a weak SPE because in each subgame $(\Sub{G}{h},v)$, the play $\rho$ induced by $\Sub{\bar \sigma}{h}$ has outcome $\mu(h\rho) = o$ and each one-shot deviating strategy in $(\Sub{G}{h},v)$ leads to a play with outcome $o$. \qed\end{proof} Lemma~\ref{lem:dense} leads to the next two corollaries. The first one states the existence of a uniform weak SPE in each initialized game $(G,v)$, $v \in V$, when the underlying graph of $G$ is strongly connected and the outcome function is prefix-independent. This corollary will provide a first step towards Theorem~\ref{thm:finitegraph} presented in Section~\ref{sec:second}; it is already interesting on its own right. \begin{corollary} \label{cor:SCC} Let $G$ be a game such that the underlying graph is strongly connected and the outcome function $\mu$ is prefix-independent. \begin{itemize} \item Then for all realizable outcomes $o$ such that $o = \mu(\rho)$ with $\rho \in Plays(v_0)$, there exists a weak SPE with outcome $o$ in $(G,v_0)$. \item Moreover, there exists a uniform strategy profile $\bar \sigma$ and an outcome $o$ such that for all $v \in V$ taken as initial vertex, $\bar \sigma$ is a weak SPE in $(G,v)$ with outcome~$o$. \end{itemize} \end{corollary} \begin{proof} For the first statement, take $\rho \in Plays(v_0)$ such that $o = \mu(\rho)$. By Lemma~\ref{lem:dense}, it is enough to show that $\mu^{-1}(\{o\})$ is dense in $(G,v_0)$ to get a weak SPE in $(G,v_0)$. For all $hv \in Hist(v_0)$, there exists a path $\pi v_0$ from $v$ to $v_0$ as the underlying graph is strongly connected. The play $h\pi\rho$ has outcome equal to $\mu(\rho) = o$ since $\mu$ is prefix-independent. Hence $\mu^{-1}(\{o\})$ is dense. To get the second statement, we need to go further by exhibiting a uniform weak SPE with the same outcome $o$ independently of the initial vertex $v$. Take any simple cycle $\pi_0v_0$ from $v_0$ to $v_0$. Such a cycle exists since the underlying graph is strongly connected. Let $\rho = \pi_0^\omega$ and $o = \mu(\rho)$ be its outcome. We partially construct a positional strategy profile $\bar \sigma$ that produces $\pi_0^{\omega}$ (recall that $\pi_0$ is simple). Let $U$ be the set of vertices that belong to $\pi_0$. Then extend the construction of $\bar \sigma$ to all $v \in V \setminus U$ in a way to reach $U$ (i.e. the cycle $\pi_0$) positionally. We then get the required uniform strategy profile $\bar \sigma$ with outcome~$o$. \qed\end{proof} The second corollary is a generalization of the previous one. It still guarantees the existence of a uniform weak SPE in all games $(G,v)$, $v \in V$, for graphs that are not necessarily strongly connected but have bottom strongly connected components all containing a play induced by a simple cycle and with the same outcome. This result will be useful in the proof of Theorem~\ref{thm:uniform} in Section~\ref{sec:second}. \begin{corollary} \label{cor:SCCwithsamepayoff} Let $G$ be a game such that the underlying graph is finite and the outcome function $\mu$ is prefix-independent. Suppose that there exists an outcome $o$ such that in each bottom strongly connected component $C$ of $G$, one can find a play $\rho_C \in Plays(v)$ for some $v \in C$ such that $\mu(\rho_C) = o$ and $\rho_C$ is induced by a simple cycle. Then there exists a uniform weak SPE with outcome $o$ in $(G,v)$, for all $v \in V$. \end{corollary} \begin{proof} Let ${\cal C}$ be the set of bottom strongly connected components of $G$. The construction of the strategy profile $\bar \sigma$ is very close to the one proposed in the previous proof. We partially construct $\bar \sigma$ in a way to produce each $\rho_C$. This is possible positionally since each $\rho_C$ is induced by a simple cycle. Let $U$ be the set of vertices that belong to $\cup_{C \in {\cal C}}\rho_C$. Then extend the construction of $\bar \sigma$ to all $v \in V \setminus U$ in a way to reach $U$ positionally. This is possible by definition of $\cal C$. The resulting strategy profile $\bar \sigma$ is uniform and is a weak SPE in each $(G,v)$, $v \in V$, such that $\mu(\out{\bar \sigma}_{v}) = o$. Indeed each $\rho_C$ has outcome $o$ and $\mu$ is prefix-independent. \qed\end{proof} We end with a last lemma which indicates how to combine different weak SPEs into one weak SPE. It will be used in the proofs of Theorems~\ref{thm:infinitetree} and~\ref{thm:finitegraph}. \begin{lemma} \label{lem:cutting} Consider an initialized game $(G,v_0)$ and a set of vertices $L \subseteq V$ such that for all $hl \in Hist(v_0)$ with $l \in L$, the subgame $(\Sub{G}{h},l)$ has a weak SPE with outcome $o_{hl}$. Consider another initialized game $(G',v_0)$ obtained from $(G,v_0)$ \begin{itemize} \item by replacing all edges $(l,v) \in E$ by one edge $(l,l)$, for all $l \in L$, \item and with outcome function $\mu'$ such that for all $\rho' \in Plays_{G'}(v_0)$, $\mu'(\rho') = o_{hl}$ if $\rho' = hl^\omega$ with $l \in L$ and $\mu'(\rho') = \mu(\rho')$ otherwise. \end{itemize} If $(G',v_0)$ has a weak SPE, then $(G,v_0)$ has also a weak SPE. \end{lemma} \begin{proof} Denote by $\bar \sigma^{hl}$ the weak SPE in each $(\Sub{G}{h},l)$, and by $\bar \sigma'$ the weak SPE in $(G',v_0)$. We then build a strategy profile $\bar \tau$ in $(G,v_0)$ as follows. For player~$i \in \Pi$ and history $hv \in Hist_i(v_0)$: \begin{itemize} \item if no vertex of $L$ occurs in $hv$, then $\tau_i(hv) = \sigma'_i(hv)$; \item otherwise, decompose $hv$ as $h_1h_2v$ such that the first occurrence of a vertex $l \in L$ is the first vertex of $h_2$. Then $\tau_i(hv) = \sigma^{h_1l}_i(h_2v)$. \end{itemize} Hence in the first case, $\tau_i$ mimics $\sigma'_i$ in the game $(G',v_0)$, and in the second case, $\tau_i$ mimics $\sigma^{h_1l}$ in the subgame $(\Sub{G}{h_1},l)$. Let us show that $\bar \tau$ is a weak SPE in $(G,v_0)$. Consider any subgame $(\Sub{G}{h},v)$ such that $v \in V_i$, and any one-shot deviation strategy $\tau'_i$ of player~$i$ from $\Sub{\bar \tau}{h}$. Either no vertex of $L$ occurs in $hv$, and $\tau'_i$ is not profitable for player~$i$ because $\bar \sigma'$ is a weak SPE in $(G',v_0)$ and by definition of $\mu'$. Or $h = h_1h_2v$ such that the first occurrence of a vertex $l \in L$ is the first vertex of $h_2$, and again $\tau'_i$ is not profitable because $\bar \sigma^{h_1l}$ is a weak SPE in the subgame $(\Sub{G}{h_1},l)$. \qed\end{proof} \subsection{Proof of Theorem~\ref{thm:infinitetree}} Now that we have established all useful intermediate results for this section and the next one, we can finally proceed to the proof of Theorem~\ref{thm:infinitetree}. W.l.o.g. we can suppose that the underlying graph of $G$ is a tree rooted at $v_0$ (by unraveling this graph from $v_0$). We first show how to transform a game played on an infinite tree to a game satisfying Conditions 1 and 2 from Theorem~\ref{thm:generalgraph} while reflecting weak SPE. In the following lemma we write $h \sqsubseteq l$ to denote that $h$ is a prefix of $l$, and denote by $\mathrm{cl} (A)$ the topological closure of $A$. \begin{lemma} \label{lemma:bairecategory} Consider a game played on an infinite tree $C^\omega$ with countable outcome set $O$ and outcome function $\mu : C^\omega \to O$. There exists a prefix-free set $L \subseteq C^*$ of \emph{leaves} and an assignment $\Theta : L \to O$ such that \begin{enumerate} \item For each $h \in C^*$ there exists some $l \in L$ with $h \sqsubseteq l$ or $l \sqsubseteq h$. \item For each $l \in L$ we find that $\mu^{-1}(\{\Theta(l)\})$ is dense in $lC^\omega$. \end{enumerate} \end{lemma} \begin{proof} By iterative use of the Baire Category Theorem. We go through all $h \in C^*$ in some order, add elements to $L$ and extend $\Theta$. Let $h \in C^*$ be the current candidate. If we do not yet have added $l$ to $L$ with $l \sqsubseteq h$ or $h \sqsubseteq l$, then consider that $hC^\omega = hC^\omega \cap \bigcup_{o \in O} \mu^{-1}(\{o\})$. As $O$ is countable, the Baire Category Theorem implies that some $\mu^{-1}(\{o_0\})$ is somewhere dense, i.e.~that there exists some $l \sqsupseteq h$ such that $lC^\omega \subseteq \mathrm{cl} \left ( \mu^{-1}(\{o_0\}) \right )$. We add $l$ to $L$ and set $\Theta(l) = o_0$. Then we proceed to the next $h$. In the limit, we have constructed $L$ and $\Theta$ as desired. To see that $L$ is prefix-free, assume that there are $l_1, l_2 \in L$ with $l_1 \sqsubset l_2$. If $l_1$ was added first, and $l_2$ was added when dealing with the history $h$, then $h \sqsubseteq l_2$. But as prefixes of a given history are linearly ordered, either $h \sqsubseteq l_1$ or $l_1 \sqsubseteq h$ follows. Thus, we would not have added $l_2$ when dealing with $h$. If $l_2$ was added first, and then $l_1$ when dealing with $h$, then we would find that $h \sqsubseteq l_1 \sqsubseteq l_2$, thus $h \sqsubseteq l_2$, thus we would not have added $l_2$. Hence, $L$ is prefix-free. \qed\end{proof} \begin{proof}[of Theorem~\ref{thm:infinitetree}] Instead of reasoning with the underlying graph of $G$, we work w.l.o.g. with its unraveling from the initial vertex $v_0$. We can apply Lemma \ref{lemma:bairecategory} to transform the game. For each leaf, we can apply Lemma~\ref{lem:dense} to obtain a weak SPE in the corresponding subgame. Together, the criteria of Lemma~\ref{lem:cutting} are satisfied. The implication of Lemma~\ref{lem:cutting} is true by Corollary~\ref{cor:generaltree}, and the conclusion yields the desired statement. \qed\end{proof} \section{Second application} \label{sec:second} In this section, we present a second large family of games with a weak SPE, as another application of the general results of Section~\ref{sec:general} (more particularly Theorem~\ref{thm:generalgraph}). This family is constituted with all games with a finite underlying graph and a prefix-independent outcome function. \begin{theorem} \label{thm:finitegraph} Let $(G,v_0)$ be an initialized game such that the underlying graph is finite and the outcome function is prefix-independent. Then there exists a weak SPE in $(G,v_0)$. \end{theorem} Let us comment this theorem. \emph{(i)} It guarantees the existence of a weak SPE for classical games with \emph{quantitative} objectives as presented in Example~\ref{ex:classical}, such that their outcome function is prefix-independent. This is the case of \emph{limsup} and \emph{mean-payoff} payoff functions (and their limit inferior counterparts). Recall that Example~\ref{ex:contrex} (see also Figure~\ref{fig:gameNoSPE}) provides a game with no SPE, where the payoff functions $\mu_i$ can be seen as either \emph{limsup} or \emph{mean-payoff} (or their limit inferior counterparts). \emph{(ii)} Later in this section, we will show that under the hypotheses of Theorem~\ref{thm:finitegraph}, there always exists a weak SPE that is \emph{finite-memory} (Corollary~\ref{cor:finite-mem}), and we will study in which cases it can be \emph{positional} or even \emph{uniform} (Theorem~\ref{thm:uniform}). \emph{(iii)} The families of games of Theorems~\ref{thm:infinitetree} and~\ref{thm:finitegraph} are incomparable: Boolean reachability games are in the first family but not in the second one, and mean-payoff games are in the second family but not in the first one. \subsection{Proof of Theorem~\ref{thm:finitegraph}} The proof of Theorem~\ref{thm:finitegraph} follows the same structure as for Theorem~\ref{thm:infinitetree}. The idea is to apply Lemma~\ref{lem:cutting} where $L$ is equal to the union of the bottom strongly connected components of the graph of $G$. The weak SPEs required by Lemma~\ref{lem:cutting} exist on the subgames $(\Sub{G}{h},l)$ with $l \in L$ by Corollary~\ref{cor:SCC}, and on the game $(G',v_0)$ thanks to Theorem~\ref{thm:generalgraph}. \begin{proof}[of Theorem~\ref{thm:finitegraph}] Let ${\cal C}$ be the set of bottom strongly connected components of the finite graph of $G$. By Corollary~\ref{cor:SCC}, for all $C \in {\cal C}$, there exist a uniform strategy profile $\bar \sigma_C$ and a outcome $o_C$ such that $\bar \sigma_C$ is a weak SPE with outcome $o_C$ in each $(G,v)$ with $v \in C$. Notice that as $\mu$ is prefix-independent, $\bar \sigma_C$ is also a weak SPE with outcome $o_C$ in all subgames $(\Sub{G}{h},v)$ with $hv \in Hist(v_0)$ and $v \in C$. If the initial vertex $v_0$ belongs to some $C \in {\cal C}$, then $\bar \sigma_C$ is the required weak SPE in $(G,v_0)$ (it is clearly finite-memory as it is uniform). From now on we suppose that $v_0 \not\in C$ for all $C \in {\cal C}$. We consider the graph $(G',v_0)$ constructed from $(G,v_0)$ as described in Lemma~\ref{lem:cutting} with $L = \cup_{C \in {\cal C}}C$. This graph satisfies all the hypotheses of Theorem~\ref{thm:generalgraph}. The set $L$ of leaves is the one used for Lemma~\ref{lem:cutting}. The first hypothesis holds because $L$ is the union of the bottom strongly connected components of $G$. The second hypothesis holds because $\mu$ is prefix-independent. The third hypothesis holds because $V$ is finite. Therefore, $(G',v_0)$ has a weak SPE $\bar \sigma'$ by Theorem~\ref{thm:generalgraph}. By the existence of the previous strategy profiles $\bar \sigma'$ and $\bar \sigma_C$, $C \in {\cal C}$, it follows by Lemma~\ref{lem:cutting} that there exists a weak SPE $\bar \tau$ in $(G,v_0)$. \qed\end{proof} \subsection{Finite-memory weak SPE} \label{subsec:finite-mem} We here make the statement of Theorem~\ref{thm:finitegraph} more precise by guaranteeing the existence of a weak SPE with finite-memory. \begin{corollary} \label{cor:finite-mem} Let $(G,v_0)$ be an initialized game such that the underlying graph is finite and the outcome function is prefix-independent. Then there exists a finite-memory weak SPE in $(G,v_0)$ with memory size bounded by the number of bottom strongly connected components of the graph. Moreover, a memory size linear in the number of bottom components is necessary. \end{corollary} \begin{proof} In the proof of Theorem~\ref{thm:finitegraph}, we have constructed a weak SPE $\bar \tau$. Let us show that $\bar \tau$ is a finite-memory strategy profile with memory size bounded by $|{\cal C}|$. Let us first come back to the construction of $\bar \tau$ given in the proof of Lemma~\ref{lem:cutting}. Consider player~$i \in \Pi$ and history $hv \in Hist_i(v_0)$. If no vertex of $L$ occurs in $hv$, then $\tau_i(hv) = \sigma'_i(hv)$. Otherwise, decompose $hv$ as $h_1h_2v$ such that the first occurrence of a vertex $l \in C \subseteq L$ is the first vertex of $h_2$, then \begin{eqnarray} \label{eq:l-C} \tau_i(hv) = \sigma_{C,i}(v). \end{eqnarray} Notice that in (\ref{eq:l-C}) $\tau_i(hv)$ only depends on $C$, and not on $l \in C$, since $\bar \sigma_C$ is uniform. Now let us recall the construction of $\bar \sigma'$ with a memory size $|L|$ given in the proof of Theorem~\ref{thm:generalgraph}, and in particular to equation (\ref{eq:rhovl}). In $(G',v_0)$ the plays $\rho_{v,o} = \pi l^\omega$ can be produced positionally while keeping $l \in L$ in memory. Therefore by $(\ref{eq:l-C})$ and as $\bar \sigma_C$ is uniform, it follows that the memory size of $\bar \tau$ can be reduced from $|L|$ to $|{\cal C}|$. Let us now prove that there exist games with a finite set $V$ and a prefix-independent function $\mu$, that require a memory size in $O(|{\cal C}|)$ for their weak SPEs. To this end, we come back to the family of games $G_n$ of Example~\ref{ex:Gn} with $n$ bottom strongly connected components. Consider the unravelling of $G_n$ from the initial vertex $v_1$ as depicted in Figure~\ref{fig:profileG4} and let us study the form of any weak SPE $\bar \sigma$ in $(G_n,v_1)$. In all subgames $(\Sub{G_n}{h},v_i)$, the induced play cannot be $(v_iv_{i+1} \ldots v_{i-1})^\omega$ with outcome $\bot$ since each player would have a profitable one-shot deviation. W.l.o.g let us suppose that $\sigma_1(v_1) = l_1$ (player~$1$ decides to move from $v_1$ to $l_1$ at the root of the unravelling, as in Figure~\ref{fig:profileG4}). Then the outcome of the play $\rho$ induced by $\Sub{\bar\sigma}{v_1}$ in the subgame $(\Sub{G_n}{v_1},v_2)$ is necessarily $o_1$ or $o_n$, otherwise player~$1$ would have a profitable one-shot deviation in $(G_n,v_0)$ (recall that $o_1 \prec_1 o_j$ for all $j \in \Pi \setminus \{1,n\}$). The first case $o_1$ cannot occur otherwise player~$2$ would have a profitable one-shot deviation in $(\Sub{G_n}{v_1},v_2)$ (recall that $o_1 \prec_2 o_2$). With similar arguments one can verify that the induced play $\rho$ is necessarily equal to $v_2v_3 \ldots v_nl_n^\omega$ with outcome $o_n$ (as in Figure~\ref{fig:profileG4}). We can repeat the same reasoning for the play induced by $\Sub{\bar\sigma}{v_1v_2 \cdots v_n}$ in the subgame $(\Sub{G_n}{v_1v_2 \cdots v_n},v_1)$ which must be equal to $v_1v_2 \ldots v_{n-1}l_{n-1}^\omega$ with outcome $o_{n-1}$, aso. Hence all weak SPEs of $(G_n,v_1)$ have the form of the one described in Figure~\ref{fig:profileG4} and they have finite memory of size $n-1$ as explained previously in Example~\ref{ex:Gn} (see also Figure~\ref{fig:MooreG4}). Let us show that such a weak SPE $\bar \sigma$ cannot have a memory size $< n-1$. Assume the contrary: wlog consider the previous weak SPE $\bar \sigma$ (as in Figure~\ref{fig:profileG4}) and in particular a Moore machine ${\cal M} = (M,m_0,\alpha_U,\alpha_N)$ encoding $\sigma_1$ such that $|M| < n-1$. Let $h_jv_1$, $j \in \{0, \ldots, n-1\}$ be consecutive histories, with $h_j = (v_1v_2 \cdots v_n)^j$. On one hand, we have $\sigma_1(h_jv_1) = \alpha_N(\widehat{\alpha}_U(m_0,h_j),v_1)$ for all $j$. On the other hand, $\sigma_1(h_0v_1) = \sigma_1(h_{n-1}v_1) = l_1$ and $\sigma_1(h_jv_1) = v_2$ for all $j \in \{1, \ldots, n-2\}$. Therefore there exists $j_1, j_2 \in \{1, \ldots, n-2\}, j_1 \neq j_2$, such that the associated memory state is identical, i.e, $\widehat{\alpha}_U(m_0,h_{j_1}) = \widehat{\alpha}_U(m_0,h_{j_2})$. Thus $\cal M$ enters into a cycle while reading the prefixes of $(v_1v_2 \cdots v_n)^\omega$. This means that $\cal M$ defines $\sigma_1(hv) = v_2$ for all histories $h$ of which $h_1$ is prefix, in contradiction with $\sigma_1(h_{n-1}v_1) = l_1$. \qed\end{proof} \subsection{Positional weak SPE} \label{subsec:positional} In the previous section, Corollary~\ref{cor:finite-mem} guarantees the existence of a finite-memory weak SPE for games with a finite underlying graph and a prefix-independent outcome function. In this section, we identify conditions on the preference relations of the players, as expressed in the next lemma, that guarantee the existence of a \emph{uniform} weak SPE (see Theorem~\ref{thm:uniform}). \begin{lemma}[Lemma 4 of \cite{Roux15}] \label{lem:killer} Let $O$ be a non empty set of outcomes. Let $\prec_i$ be a preference relation over $O$, for all $i \in \Pi$. The following assertions are equivalent. \begin{itemize} \item For all $i, i' \in \Pi$ and all $o, p, q \in O$, we have $\neg(o \prec_i p \prec_i q \wedge q \prec_{i'} o \prec_{i'} p)$. \item There exist a partition $\{O_k \}_{k \in K}$ of $O$ and a linear order $<$ over $K$ such that \begin{itemize} \item $k < k'$ implies $o \prec_i o'$ for all $i \in \Pi$, $o \in O_k$ and $o' \in O_{k'}$, \item $\Sub{\prec_i}{O_k} = \Sub{\prec_{i'}}{O_k}$ or $\Sub{\prec_i}{O_k} = (\Sub{\prec_{i'}}{O_k})^{-1}$ for all $i, i' \in \Pi$. \end{itemize} \end{itemize} \end{lemma} In the previous lemma, we call each set $O_k$ a \emph{layer}. The second assertion states that $(i)$ if $k < k'$ then all outcomes in $O_{k'}$ are preferred to all outcomes in $O_k$ by all players, and $(ii)$ inside a layer, any two players have either the same preference relations or the inverse preference relations. When a set of outcomes satisfies the conditions of Lemma~\ref{lem:killer}, we say that it is \emph{layered}. In~\cite{Roux15}, the author characterizes the preference relations that always yield SPE in games with outcome functions in the Hausdorff difference hierarchy of the open sets. One condition is that the set of outcomes is layered. \begin{theorem} \label{thm:uniform} Let $G$ be a game with a finite underlying graph and such that the outcome function is prefix-independent with a layered set $O$ of outcomes. Then there exists a uniform weak SPE in $(G,v)$, for all $v \in V$. \end{theorem} \begin{example} \label{ex:opt} Remember the class $G_n$ of games, $n \geq 3$, of Example~\ref{ex:Gn}, such that $O = \{o_1, \ldots, o_n, \bot \}$ and each player $i$ has a preference relation $\prec_i$ satisfying $\bot \prec_i o_{i-1} \prec_i o_i \prec_i o_j$ for all $j \in \Pi \setminus \{i-1,i\}$. This set of outcomes is not layered because the first assertion of Lemma~\ref{lem:killer} is not satisfied. Indeed we have $$o_2 \prec_3 o_3 \prec_3 o_1 \wedge o_1 \prec_2 o_2 \prec_2 o_3.$$ Recall that in the proof of Corollary~\ref{cor:finite-mem} we have shown that all weak SPEs of $G_n$ require a memory size in $O(n)$. Hence the hypothesis of Theorem~\ref{thm:uniform} about the preference relations is not completely dispensable. \end{example} Let us proceed to the proof of Theorem~\ref{thm:uniform}. Let ${\cal C}$ be the set of the bottom strongly connected components of the finite underlying graph of $G$. For each $C \in {\cal C}$, we fix a play $\rho_C \in Plays(v)$ for some $v \in C$ induced by a simple cycle. The set $O_{\cal C} = \{o_C \mid o_C = \mu(\rho_C), C \in {\cal C} \}$ is finite. It is layered by hypothesis with a finite partition into layers $\{O_k \}_{k \in K}$. The proof of Theorem~\ref{thm:uniform} is by induction on the number of layers and uses the next lemma dealing with one layer. \begin{lemma} \label{lem:onelayer} Suppose that $|K| = 1$, then there exists a uniform strategy profile~$\bar \sigma$ that is a weak SPE in each $(G,v)$, $v \in V$, such that $\mu(\out{\bar \sigma}_v) = o_C$ for some $C \in {\cal C}$. \end{lemma} The proof of this lemma is by induction on $|O_{\cal C}|$. The case of only one outcome is solved by Corollary~\ref{cor:SCCwithsamepayoff}. When they are several outcomes in $O_{\cal C}$, we will show how to decompose $G$ into two subgames $G'$ and $G''$ such that the bottom strongly connected component of $G'$ (resp. $G''$) are those components $C \in \cal C$ of $G$ such that $o_C = o$ for some $o$ (resp. $o_C \in O_{\cal C} \setminus \{o\}$). By Corollary~\ref{cor:SCCwithsamepayoff} for $G'$ and by induction hypothesis for $G''$, we will get two uniform weak SPEs that can be merged to get a uniform weak SPE for $G$. \begin{proof}[of Lemma~\ref{lem:onelayer}] The proof is by induction on $|O_{\cal C}|$. We solve the basic case $|O_{\cal C}| = 1$ by Corollary~\ref{cor:SCCwithsamepayoff}. Suppose that $|O_{\cal C}| = n > 1$. By Lemma~\ref{lem:killer}, we have $\prec_i = \prec_{i'}$ or $\prec_i = \prec_{i'}^{-1}$ for all $i, i' \in \Pi$. We can thus merge the players into two \emph{meta-players} ${\cal P}_1$ and ${\cal P}_2$ with their respective preference relations $\prec_1$, $\prec_2$ on $O_{\cal C}$ satisfying $o_1 \prec_1 o_2 \prec_1 \ldots \prec_1 o_n$ and $o_n \prec_2 o_{n-1} \prec_2 \ldots \prec_2 o_1$. Notice that ${\cal P}_2$ could not exist. For the sequel, we need the classical concept of \emph{attractor} of $U \subseteq V$ for ${\cal P}_1$~\cite{2001automata}: it is the set ${Attr}_1(U)$ composed of all $v \in V$ from which ${\cal P}_1$ can force, against ${\cal P}_2$, to reach $U$. More precisely, ${Attr}_1(U)$ is constructed by induction as follows: ${Attr}_1(U) = \cup_{k \geq 0} X_k$ such that \begin{eqnarray*} \label{eq:attr} X_0 &=& U, \\ X_{k+1} &=& X_k \cup \{v \in V \mid v \mbox{ is controlled by ${\cal P}_1$ and } \exists (v,v') \in E, v' \in X_k \} \\ && ~~~~\cup \{v \in V \mid v \mbox{ is controlled by ${\cal P}_2$ and } \forall (v,v') \in E, v' \in X_k \}. \end{eqnarray*} Let ${\cal C}' = \{C \in {\cal C} \mid o_C = o_n\}$ and ${\cal C}'' = {\cal C} \setminus {\cal C}'$. We construct a subset $V'$ of $V$ as follows: \begin{enumerate} \item Initially $V' \gets \cup \{C \mid C \in {\cal C}'\}$ \item $V' \gets {Attr}_1(V')$. Let $\cal D$ be the set of bottom strongly connected components of $\Sub{G}{V \setminus V'}$ \item If $\cal D$ contains components not in ${\cal C}''$, then add all of them to $V'$ and goto~2, else stop \end{enumerate} At the end of the process, we get two sets $V'$ and $V'' = V \setminus V'$, and the related subgames $G'$ and $G''$ respectively induced by $V'$ and $V''$. Let us prove by induction on the three steps that (*) for all $v \in V'$, there is a path from $v$ to some $C \in {\cal C}'$. To this end, we denote $W = {Attr}_1(V')$ at step~2 and $T = W \bigcup \cup\{D \in {\cal D} \mid D \not\in {\cal C}\}$ at step~3. After step~1, (*) is true (with the empty path from $v$ to $v$). It is also the case after step~2, since by definition of the attractor, there is a path from $v \in W = {Attr}_1(V')$ to some $v' \in V'$ for which there is a path to some $C \in {\cal C}'$ by induction hypothesis. Consider now $v \in D$ such that $D \in \cal D$ is added to $W$ in step 3. As $D$ does not belong to ${\cal C}''$ and $D$ is a bottom component of $\Sub{G}{V \setminus W}$, then there must exist a path from $v \in D$ to some $C \in {\cal C}'$ and (*) holds. By construction each $C \in {\cal C}'$ (resp. $C \in {\cal C}''$) is a bottom strongly connected component of $G'$ (resp. $G''$). Let us prove that neither $G'$ nor $G''$ contain other bottom components. Assume the contrary and let $v$ be a vertex belonging to such a bottom component $D$. By step 3 of the previous process, $v$ cannot belong to $V''$. By (*), $v$ cannot belong to $V'$. Therefore the set of bottom strongly connected components of $G'$ and $G''$ is equal to ${\cal C}$. By Corollary~\ref{cor:SCCwithsamepayoff} for $G'$ and by induction hypothesis for $G''$, there exist two uniform strategy profiles $\bar \sigma'$ and $\bar \sigma''$ respectively on $G'$ and $G''$ such that $\bar \sigma'$ (resp. $\bar \sigma''$) is a weak SPE in each $(G',v')$, $v' \in V'$ (in each $(G'',v'')$, $v'' \in V''$). Moreover $\mu(\out{\bar \sigma'}_{v'}) = o_n$ and $\mu(\out{\bar \sigma''}_{v''}) \in P_{\cal C} \setminus \{o_n\}$. The required uniform strategy profile $\bar \sigma$ on $G$ is built such that $\Sub{\bar \sigma}{V'} = \bar \sigma'$ and $\Sub{\bar \sigma}{V''} = \bar \sigma''$. Let us show that it is a weak SPE in all $(G,v)$, $v \in V$. Consider first a subgame $(\Sub{G}{h},v')$ such that $\out{\Sub{\bar \sigma}{h}}_{v'}$ is a play in $G'$ and a one-shot deviating strategy using an edge $(v',v'')$ with $v' \in V'$ and $v'' \in V''$. By step 2 (i.e. by definition of the attractor), $v'$ belongs to ${\cal P}_1$ who has no incentive to use $(v',v'')$ since the deviating play goes to $G''$ for which ${\cal P}_1$ receives an outcome $o_m$ such that $o_m \prec_1 o_n$. Consider next a subgame $(\Sub{G}{h},v'')$ such that $\out{\Sub{\bar \sigma}{h}}_{v''}$ is a play in $G''$ and a one-shot deviating strategy using an edge $(v'',v')$ with $v' \in V'$ and $v'' \in V''$. By step 2, $v''$ now belongs to ${\cal P}_2$ who has no incentive to use $(v'',v')$ since he will receive an outcome $o_m$ such that $o_n \prec_2 o_m$. \qed\end{proof} We can now proceed to the proof of Theorem~\ref{thm:uniform}, which is by induction on the number of layers of $O$. The case of one layer is treated in Lemma~\ref{lem:onelayer}. In case of several layers, we show in the proof how to decompose $G$ into two subgames $G'$ and $G''$ such that there is only one layer in $G'$ and less layers in $G''$ than in $G$. From the two uniform weak SPEs obtained for $G'$ by Lemma~\ref{lem:onelayer} and for $G''$ by induction hypothesis, we construct the required uniform weak SPE for $G$. \begin{proof}[of Theorem~\ref{thm:uniform}] We will prove the theorem by induction on the number of layers and additionally show that for all $v \in V$, $\mu(\out{\bar \sigma}_v) = o_C$ for some $C \in {\cal C}$. Let $O' \subseteq O_{\cal C}$ be the highest layer of $O_{\cal C}$ (with respect to the linear order $<$ over $K$). If $O' = O_{\cal C}$, then there is only one layer and the required uniform strategy profile follows from Lemma~\ref{lem:onelayer}. If $O' \subset O_{\cal C}$, we define $V' \subset V$ composed of all vertices $v$ for which there exists a path from $v$ to some component $C \in \cal C$ such that $o_C \in O'$ (in particular $V'$ includes all such components), and we let $V'' = V \setminus V'$. We obtain two subgames $G'$ and $G''$ respectively induced by $V'$ and $V''$. By construction of $V'$, one easily checks that the union of the bottom strongly connected components of $G'$ and $G''$ is equal to ${\cal C}$. Hence, $G'$ has only one layer (equal to $O'$) and $G''$ has one layer less than $G$. It follows (by Lemma~\ref{lem:onelayer} and by induction hypothesis) the existence of two strategy profiles $\bar \sigma'$ and $\bar \sigma''$ respectively on $G'$ and $G''$: $\bar \sigma'$ is a uniform weak SPE in each $(G',v')$, $v' \in V'$, such that $\mu(\out{\bar \sigma'}_{v'}) \in O'$, and $\bar \sigma''$ is a uniform weak SPE in each $(G'',v'')$, $v'' \in V''$, such that $\mu(\out{\bar \sigma''}_{v''}) \in O \setminus O'$. The required strategy profile $\bar \sigma$ on $G$ is built such that $\Sub{\bar \sigma}{V'} = \bar \sigma'$ and $\Sub{\bar \sigma}{V''} = \bar \sigma''$. As in the proof of Lemma~\ref{lem:onelayer}, we consider crossing edges between $G'$ and $G''$. By construction, there is no edge $(v'',v')$ with $v' \in V'$ and $v'' \in V''$ showing that a play starting in $G''$ remains in $G''$. On the contrary, there exist edges $(v',v'')$ with $v' \in V'$ and $v'' \in V''$, but no player has an incentive to use them in a one-shot deviating strategy since the resulting outcome is in a layer smaller than $O'$. Therefore, $\bar \sigma$ is a weak SPE in each $(G,v)$. \qed\end{proof} \section{A counterexample for countably many players and outcomes} \label{sec:counterexample} We proceed to give an example of a game without weak SPE. It shows that the requirement of only finitely many leaf-outcomes is not dispensable in Theorem~\ref{thm:generalgraph} or Theorem~\ref{thm:infinitetree}. In \cite[Section 4.3]{Flesch10} there is an example of a game in extensive form with countably many players, uncountably many outcomes, preference heights $3$, but without weak SPE. Our example is similar, but with only countably many outcomes, one single proper infinite play (i.e.~not ending in a leaf), and preferences of height $3$. \begin{example} We consider the initialized game $(G,v_0)$ of Figure~\ref{fig:counterex}. The set of players is $\mathbb N$. The player $i$ acts at most once, at the vertex $v_i$, and can either enter the leaf $l_i$ or move onwards to $v_{i+1}$. The play starts with player $0$ at $v_0$. The outcome attached to reaching $l_i$ is denoted by $2^{i}1^\omega$, the outcome attached to the infinite path $v_0v_1v_2\ldots$ is denoted by $0^\omega$. The preferences of player $i$ are given by $p \prec_i q$ iff $p(i) < q(i)$. The game $(G,v_0)$ has no SPE. To prove this statement, it is enough to show that there is no very weak SPE by Proposition~\ref{prop:weak-veryweak} and since every player only acts one. Assume by contradiction that there exists a very weak SPE $\bar \sigma$. In each subgame $(\Sub{G}{h},v_i)$, the play induced by $\bar\sigma$ cannot be the one with outcome $0^\omega$. Otherwise player $i$ has a profitable one-shot deviating strategy by moving to leaf $l_i$ (by increasing his payoff from 0 to 1). Therefore, for all $k$, there exists a player $i \geq k$ who moves to leaf $l_{i}$. Let $i$ be the first such player. It follows that in $(\Sub{G}{h},v_i)$, he can increase his payoff from 1 to 2 by moving to $v_{i+1}$ instead to $l_i$, contradiction. \begin{figure} \centering \begin{tikzpicture}[level distance=8mm] \node{$v_0$}[sibling distance=14mm] child{node{$1^\omega$}[sibling distance=10mm]} child{node{$v_1$}[sibling distance=14mm] child{node{$2\cdot 1^\omega$}} child{node{$v_2$} [sibling distance=14mm] edge from parent[solid] child{node{$22\cdot 1^\omega$} edge from parent[solid]} child{node{$v_3$} [sibling distance=14mm] edge from parent[solid] child{node{$222\cdot 1^\omega$} edge from parent[solid]} child{node{} edge from parent[dashed] child{node{}edge from parent[draw=none]} child{node{$0^\omega$} edge from parent[dashed]} } } } }; \end{tikzpicture} \caption{A game with no weak SPE} \label{fig:counterex} \end{figure} \end{example} \bibliographystyle{plain}

https://arxiv.org/abs/1903.04214

How far away must forced letters be so that squares are still avoidable?

We describe a new non-constructive technique to show that squares are avoidable by an infinite word even if we force some letters from the alphabet to appear at certain occurrences. We show that as long as forced positions are at distance at least 19 (resp. 3, resp. 2) from each other then we can avoid squares over 3 letters (resp. 4 letters, resp. 6 or more letters). We can also deduce exponential lower bounds on the number of solutions. For our main Theorem to be applicable, we need to check the existence of some languages and we explain how to verify that they exist with a computer. We hope that this technique could be applied to other avoidability questions where the good approach seems to be non-constructive (e.g., the Thue-list coloring number of the infinite path).

\section{Introduction} A square is a word of the form $uu$ where $u$ is a non empty word. We say that a word is square-free (or avoids squares) if none of its factors is a square. For instance, $hotshots$ is a square while $minimize$ is square-free. In 1906, Thue showed that there are arbitrarily long ternary words avoiding squares \cite{Thue06}. This result is often regarded as the starting point of combinatorics on words, and the generalizations of this particular question received a lot of attention. The authors of \cite{mainquestionpandd} study three such questions asked by Harju \cite{HARJU2018}. They also introduced a stronger version of the third problem. \begin{Problem}[{\cite[Problem 4]{mainquestionpandd}}] Let $p\ge2$ be an integer and let $v=v_1v_2v_3\ldots$ be any infinite ternary word. Does there exist an infinite ternary square-free word $w=w_1w_2w_3\ldots$ such that for all $i$, $w_{p\cdot i}=v_i$? \end{Problem} They give a partial solution to this question and they show that the answer is yes for any $v$ if $p\ge30$. In fact, they showed something slightly stronger. Let $d(\Sigma)$ be the smallest integer such that for all $v\in\Sigma^\omega$ and for all sequence of indices $(p_i)_{1\le i}$ such that $\forall i$, $p_{i+1}-p_i\ge d(\Sigma)$, there is an infinite square-free word $u\in\Sigma^\omega$ such that $v_i= u_{p_i}$. They showed that $6\le d(\{0,1,2\})\le30$. Moreover, the fact that squares are avoidable over 3 letters can be used to show that $d(\{0,1,2,3\})\le 7$. We show that $7\le d(\{0,1,2\})\le19$, $d(\{0,1,2,3\})=3$, $d(\{0,1,2,...,k\})=2$ for $k\ge6$. The main theorem of this paper gives sufficient conditions for the existence of square-free languages that fulfill some constraints. Kolpakov showed that there are more than $1.30125^n$ square-free words of length $n$ over a ternary alphabet using a new non-constructive technique \cite{Kolpakov2007}. One of the ideas behind Kolpakov's result is roughly to approximate (using a computer) the language of square-free words by the language of words avoiding squares of period less than $l$ for large $l$, and to show that we do not lose too many words if we remove the larger squares from this language. We use a similar idea in this paper. We also use ideas from the power series method (see for instance \cite{BELL20071295,doublepat}) even if we do not explicitly manipulate any power series. It seems to be a good approach to show that the Thue-list number of paths is $3$ (see \cite{CZERWINSKI2007453,doi:10.1002/rsa.20411} for definitions and conjectures on this topic) or to tackle other problems that might require a non-constructive approach. This paper is organized as follows. We start by fixing some notations in Section \ref{secdef}. In Section \ref{secidea}, we give a weaker version of Theorem \ref{mainTh} to present the ideas of the theorem without some of the technicalities. Then in Section \ref{secentr}, we give the proof of Theorem \ref{mainTh}, our main theorem. In Section \ref{secfindingL}, we explain how to verify with a computer the existence of some languages that are required to apply Theorem \ref{mainTh}. Finally, in Section \ref{secconc}, we use Theorem \ref{mainTh} to bound the values of $d$ for different alphabet sizes. \section{Definitions and notations}\label{secdef} We denote the set of non-negative integer (resp. positive integers) by $\mathbb{N}_0$ (resp. $\mathbb{N}_{>0}$). For any word $w\in\Sigma^*$, we denote the $i$th letter of $w$ by $w_i$ and the length of $w$ by $|w|$. Then for any $w\in\Sigma^*$, $w=w_1w_2\ldots w_{|w|}$. For any set of non-empty words $W$, we let $W^*$ (resp. $W^\omega$) be the set of words obtained by catenation of finitely many (resp. infinitely many) elements of $W$. A \emph{language} over an alphabet is a set of finite words over this alphabet. We use the convention that $\prod_{x\in\emptyset}x = 1$ and $\max_{x\in\emptyset}x = 0$ (we could use $-\infty$ for the second one, but it is slightly less convenient for the implementation). A \emph{partial word over $\Sigma$} is a (possibly infinite) word over the alphabet $\Sigma\cup\{\diamond\}$. For any partial word $\mu\in (\Sigma\cup\{\diamond\})^*\cup (\Sigma\cup\{\diamond\})^\omega$ and word $v\in\Sigma^*\cup \Sigma^\omega$, we say that \emph{$v$ is compatible with $\mu$} if $|v|\le|\mu|$ and $\mu_i\not=\diamond \implies \mu_i=v_i$ for all $i$ such that $v_i$ and $\mu_i$ are defined. We denote by $S(\mu)$ the set of square-free words that are compatible with the partial word $\mu$. \section{Idea of Theorem \ref{mainTh}}\label{secidea} The main Theorem of this paper is Theorem \ref{mainTh}. The main idea of this theorem is that if a language avoids short squares and is large enough then it contains square-free words of any length. The statement and proof of this theorem are rather difficult to follow so we give in this section a version of the Theorem for the case where the set $W$ is a singleton $\{w\}$. We hope that this helps to convey the ideas of the proof of Theorem \ref{mainTh}. This is in fact really similar to the ideas of \cite{doublepat}, but instead of building the word letter by letter, we construct it factor by factor. For that we fix one size of a factor and look at the number of words whose length corresponds to multiples of this size. \begin{theorem}\label{easy} Let $\Sigma$ be an alphabet, $w\in(\Sigma\cup\{\diamond\})^*$ be a finite partial word and $p\ge2|w|$ such that $|w|$ divides $p$. Suppose that there are $C\in\mathbb{N}_{>0}$ and $L$ a language such that: \begin{enumerate}[(I)] \item $\varepsilon\in L$. \item For all $u\in L$, $u$ avoids squares of period less than $p$. \item For any $u\in L$ there are at least $C$ different words $v\in\Sigma^{|w|}$ compatible with $w$ such that $uv\in L$. \item \label{condeq}There exists $x\in]0,1[$ such that: \begin{equation*} C\left(1-\frac{x^{\frac{p}{|w|}-1}|w|^2 }{1-x}\right) \ge x^{-1} \end{equation*} \end{enumerate} Then $S(w^\omega)$ is infinite. \end{theorem} \begin{proof} Let $\mu=w^\omega$. Let $L(\mu)$ be a set of words from $L$ that are compatible with $\mu$ such that, for any $u\in L(\mu)$ of length divisible by $|w|$, there are exactly $C$ different words $v\in\Sigma^*$ compatible with $w$ with $uv\in L(\mu)$. Conditions (I),(II) and (III) imply that such a set can be obtained by removing words from $L$. For all non-negative $i$, let $s_{i}= \left|S(\mu)\cap L(\mu)\cap \{u\in\Sigma^*: |u|=i|w|\}\right|$ be the number of square-free words of $L(\mu)$ of length $i|w|$. We will show by induction on $i$ that for all positive $i$, $s_{i+1}\ge x^{-1}s_{i}$. Let $n$ be a positive integer such that: \begin{equation} \forall 0\le i<n,\, s_{i+1}\ge x^{-1}s_{i}\tag{IH1}\label{IH11} \end{equation} By definition of $L(\mu)$, for any word $w$ of $S(\mu)\cap L(\mu)$ there are exactly $C$ different factors $v$ of length $|w|$ such that $wv$ is in $L(\mu)$. Let $F$ be the set of words in $L(\mu)\setminus S(\mu)$ of length $(n+1)|w|$ whose prefix of length $n|w|$ is in $S(\mu)\cap L(\mu)$. Then by definition: \begin{equation}\label{eqrecs1} s_{n+1}\ge Cs_{n}-|F|. \end{equation} In order to bound $|F|$, let us introduce for all $i<n+1$, $F_i=\{uvvy\in F: |w|(i-1)<|uv|\le i|w|, |y|<|w|\}$. That is, $F_i$ is the set of words of $F$ that contain a square whose midpoint (the middle of the square) is located between the positions $(i-1)|w|$ and $i|w|$ in the word. Clearly $|F|\le \sum_{i=1}^{n} |F_i|$, so our next task is to compute bounds on $|F_i|$ for all $i$. \begin{lemma}\label{sizefi1} We have the following inequalities: \begin{itemize} \item for all $i> n+1-\frac{p}{|w|}$, $|F_i|=0$, \item for all $i \le n+1-\frac{p}{|w|}$, $|F_{i}|\le s_{i}C|w|^2$. \end{itemize} \end{lemma} \begin{proof} If $i> n+1-\frac{p}{|w|}$, then $(i-1)|w|+p\ge (n+1)|w|$. Since, $L$ does not contain squares of period less than $p$, $F_i=\emptyset$. Now, let $i \le n+1-\frac{p}{|w|}$. For any $i$ and $z\in S(\mu)\cap L(\mu)\cap \{u\in\Sigma^*: |u|=i|w|\}$ let $F_i(z)$ be the set of words of $F_i$ that admit $z$ as a prefix. By definition of $F_i$, any word $F_i(z)$ contains a square whose second half starts in position $a+1$ and ends in position $b$ where $(i-1)|w|< a\le i|w|$ and $ n|w|< b\le (n+1)|w|$. Given $z$, $a$ and $b$, we know the first half of the square and thus the word is known at least up to position $n|w|$. By definition of $L(\mu)$ there are at most $C$ possible values for the remaining $|w|$ letters. By summing over all the values of $a$ and $b$ one gets: $|F_{i}(z)|\le C|w|^2$. By summing over all the values of $z$, we finally get $|F_{i}|\le s_{i}C|w|^2$ \end{proof} Now, by \eqref{IH11}, for all $i$, $|F_i|\le C|w|^2x^{n-i}s_n$ and thus: \begin{align*} |F|\le& \sum_{i=1}^{n} |F_i|\le \sum_{i=1}^{n+1-\frac{p}{|w|}} C|w|^2x^{n-i}s_n\le C|w|^2s_n \sum_{i=\frac{p}{|w|}-1}^{\infty} x^{i}\\ |F|\le& C|w|^2s_n \frac{x^{\frac{p}{|w|}-1}}{1-x} \end{align*} We can use this bound in inequality (\ref{eqrecs1}) and we get: \begin{align*} s_{n+1}&\ge Cs_{n}-C|w|^2s_n \frac{x^{\frac{p}{|w|}-1}}{1-x}\\ s_{n+1}&\ge s_nC\left(1-\frac{x^{\frac{p}{|w|}-1}|w|^2 }{1-x}\right)\\ s_{n+1}&\ge x^{-1}s_n\text{\ \ \ (By Theorem hypothesis \ref{condeq})} \end{align*} This concludes the proof that for all positive $i$, $s_{i+1}\ge x^{-1}s_{i}$. Since $s_0=1$, we deduce that $s_i$ is unbounded and thus $S(w^\omega)$ is infinite \end{proof} \section{The main theorem}\label{secentr} This section is devoted to the proof of the main Theorem. As already mentioned the ideas of the proof are the same as for the proof of Theorem \ref{easy}. However, this is more technical because $W$ is not a singleton anymore. Moreover, we want the equivalent of condition \eqref{condeq} to be as general as possible and for that, we need to bound the size of $|F|$ as tightly as possible. Thus the equivalent of Lemma \ref{sizefi1} (Lemma \ref{sizefi2}) is much more technical and we delay its proof to a later subsection. \begin{theorem}\label{mainTh} Let $\Sigma$ be an alphabet, $W\subseteq(\Sigma\cup\{\diamond\})^*$ be a finite set of finite partial words, $p\ge2\max\{|w|: w\in W\}$ be an integer. Suppose that there is a language $L$ and a function $f:\mathbb{N}_{>0}\rightarrow \mathbb{N}_{>0}$ such that: \begin{enumerate}[(I)] \item $\varepsilon\in L$. \item For any $u\in L$ and $w\in W$ there are at least $f(|w|)$ different words $v\in\Sigma^{|w|}$ compatible with $w$ and such that $uv\in L$. \item For all $u\in L$, $u$ avoids squares of period less than $p$. \item \label{condeq2} For all $u,v\in W$ and integer $1\le i\le|v|$, let $$\alpha(|u|,|v|)=\sum_{m=1}^{|u|} \sum_{j=0}^{\left\lfloor\frac{|v|-1}{m}\right\rfloor}\min\left\{f(|v|),(|\Sigma|-1)^{|v|-1-jm}\right\}$$ and $$\alpha'(i,|v|) =\sum_{m=0}^{i-1} \min\left\{f(|v|),(|\Sigma|-1)^{m}\right\}.$$ There exist $x_1,x_2,\ldots, x_{\max\{|w|: w\in W\}}\in]0,1[$ and $\beta:\{0,\ldots,p\}\rightarrow [0,1]$ solution of the following system: \begin{equation*} \left\{ \begin{array}{l} \forall w\in W,\\ f(|w|) - \max\limits_{\substack{u,v\in W \\1\le r\le |v|}}\left\{\beta\left(r+p-|w|-|v|\right)\left(\alpha'(r,|w|)+\frac{x_{|v|}\alpha(|u|,|w|)}{1-x_{|u|}}\right)\right\} \ge x_{|w|}^{-1}\\ \forall j\le p, \beta(j)=\max\left\{\prod\limits_{i\in\{|u|:u\in W\}} x_i^{n_i} \Bigg| \begin{array}{c} \forall i\in\{|u|:u\in W\} , n_i\in \mathbb{N}_0,\\ \text{and}\\ \sum\limits_{i\in\{|u|:u\in W\}} i \cdot n_i= j \end{array}\right\}\\ \end{array} \right. \end{equation*} \end{enumerate} Then for any infinite partial word $\mu\in W^\omega$, $S(\mu)$ is infinite \end{theorem} \begin{proof} Let $\mu\in W^\omega$ and $(\mu_i)_{i\in\mathbb{N}_{>0}}\in W^{\mathbb{N}_{>0}}$ be a sequence of elements of $W$ such that $\mu=\mu_1\mu_2\mu_3\ldots$. For any integer $i$, let $l(i)= |\mu_1\ldots \mu_i|$. Let $L(\mu)$ be a set of words from $L$ that are compatible with $\mu$ such that, for any $u\in L(\mu)$ of length $|\mu_1\ldots \mu_j|$, there are exactly $f(|\mu_{j+1}|)$ different words $v\in\Sigma^*$ compatible with $\mu_{j+1}$ with $uv\in L(\mu)$. That is, we remove words from $L$ in order to replace the ``at least $f(|\mu_{j+1}|)$'' by ``exactly $f(|\mu_{j+1}|)$''. For all non-negative $i$, let $s_{i}= \left|S(\mu)\cap L(\mu)\cap \{v\in\Sigma^*: |v|=i\}\right|$ be the number of square-free words of $L(\mu)$ of length $i$. We will show by induction on $i$ that for all positive $i$, $s_{l(i+1)}\ge x_{|\mu_{i+1}|}^{-1}s_{l(i)}$. Let $n$ be a positive integer such that: \begin{equation} \forall 0\le i<n,\, s_{l(i+1)}\ge x_{|\mu_{i+1}|}^{-1}s_{l(i)}\tag{IH1}\label{IH1} \end{equation} By definition of $L(\mu)$, for any word $w$ of $S(\mu)\cap L(\mu)$ of length $l(n)$ there are exactly $f(|\mu_{n+1}|)$ different factors $v$ of length $|\mu_{n+1}|$ such that $wv$ is in $L(\mu)$. Let $F$ be the set of words in $L(\mu)\setminus S(\mu)$ of length $l(n+1)$ whose prefix of length $l(n)$ is in $S(\mu)\cap L(\mu)$. Then by definition: \begin{equation}\label{eqrecs} s_{l(n+1)}\ge f(|\mu_{n+1}|)s_{l(n)}-|F|. \end{equation} In order to bound $|F|$, let us introduce for all $i<n+1$, $F_i=\{uvvw\in F: l(i-1)<|uv|\le l(i), |w|<|\mu_{n+1}|\}$. That is, $F_i$ is the set of words of $F$ that contain a square whose midpoint (the middle of the square) is located between the positions $|\mu_1\ldots \mu_i|$ and $|\mu_1\ldots \mu_{i-1}|$ in the word. Clearly $|F|\le \sum_{i=1}^{n} |F_i|$, so our next task is to compute the values of $|F_i|$ for all i. Let $d$ be the smallest integer such that $|\mu_{d+1}\ldots \mu_{n+1}|\le p$ and let $r=|\mu_{d}\mu_{d+1}\ldots \mu_{n+1}|-p$. Remark that $r>0$. \begin{restatable}{lemma}{sizefii} \label{sizefi2} We have the following inequalities: \begin{itemize} \item for all $i> d$, $|F_i|=0$, \item $|F_{d}|\le s_{l(d)}\alpha'(r,|\mu_{n+1}|)$, \item for all $i \le d$, $|F_{i}|\le s_{l(i)}\alpha(|\mu_{i}|,|\mu_{n+1}|)$. \end{itemize} \end{restatable} The proof of this Lemma is not really informative and is mostly a rather technical counting argument, so we moved it to Section \ref{secprooflemma}. We can use the bounds on the sizes of the $F_i$s to bound $|F|$: \begin{lemma}\label{sizeofF} We have $$|F|\le s_{l(d)}\max_{u\in W}\left\{\alpha'(r,|\mu_{n+1}|)+\frac{x_{|\mu_d|}\alpha(|u|,|\mu_{n+1}|)}{1-x_{|u|}}\right\}.$$ \end{lemma} \begin{proof} First, let us show by induction on $i$ that for all $0\le i<d$: \begin{equation*}\sum_{ j=0}^{i} s_{l(j)}\alpha(|\mu_j|,|\mu_{n+1}|)\le s_{l(i)}\max_{u\in W}\left\{\frac{\alpha(|u|,|\mu_{n+1}|)}{1-x_{|u|}}\right\}.\tag{IH2}\label{IH2} \end{equation*} Let us first show that this is true with $i=0$, using the fact that $\mu_0\in]0,1[$. $$s_{l(0)}\alpha(|\mu_0|,|\mu_{n+1}|)\le s_{l(0)}\frac{\alpha(|\mu_0|,|\mu_{n+1}|)}{1-x_{|\mu_0|}}\le s_{l(0)}\max_{u\in W}\left\{\frac{\alpha(|u|,|\mu_{n+1}|)}{1-x_{|u|}}\right\}. $$ Now, let $i+1$ be an integer such that \eqref{IH2} is true for $i$. \begin{align*} &\sum_{ j=0}^{i+1} s_{l(j)}\alpha(|\mu_j|,|\mu_{n+1}|)\le s_{l(i+1)}\alpha(|\mu_{i+1}|,|\mu_{n+1}|)+s_{l(i)}\max_{u\in W}\left\{\frac{\alpha(|u|,|\mu_{n+1}|)}{1-x_{|u|}}\right\}\\ &\le s_{l(i+1)}\left(\alpha(|\mu_{i+1}|,|\mu_{n+1}|)+x_{|\mu_{i+1}|}\max_{u\in W}\left\{\frac{\alpha(|u|,|\mu_{n+1}|)}{1-x_{|u|}}\right\}\right) \text{ (By \eqref{IH1})}\\ &\le s_{l(i+1)}\max_{u\in W}\left\{\frac{\alpha(|u|,|\mu_{n+1}|)}{1-x_{|u|}}\right\} \end{align*} Thus equation \eqref{IH2} is true for all $i\le d$ and in particular for $i=d-1$ and we get: \begin{align*} |F|&\le|F_d|+ s_{l(d-1)}\max_{u\in W}\left\{\frac{\alpha(|u|,|\mu_{n+1}|)}{1-x_{|u|}}\right\}\\ |F|&\le s_{l(d)}\alpha'(r,|\mu_{n+1}|)+ s_{l(d)}x_{|\mu_{d}|} \max_{u\in W}\left\{\frac{\alpha(|u|,|\mu_{n+1}|)}{1-x_{|u|}}\right\}\\ |F|&\le s_{l(d)}\max_{u\in W}\left\{\alpha'(r,|\mu_{n+1}|)+\frac{x_{|\mu_{d}|}\alpha(|u|,|\mu_{n+1}|)}{1-x_{|u|}}\right\} \end{align*} This concludes the proof of this Lemma. \end{proof} By induction hypothesis \eqref{IH1} $s_{l(d)}\le s_{l(n)}\prod_{i=d+1}^{n}x_{|\mu_{j}|}$. Let us bound the product on the right hand side: \begin{align*} \prod_{i=d+1}^{n}x_{|\mu_{j}|}&\le\max\left\{\prod\limits_{i\in\{|u|:u\in W\}} x_i^{n_i} \Bigg| \begin{array}{c} \forall i\in\{|u|:u\in W\} , n_i\in \mathbb{N}_{0}\\ \text{and}\\ \sum\limits_{i\in\{|u|:u\in W\}} i \cdot n_i= l(n+1)-l(d)-\mu_{n+1} \end{array}\right\}\\ \prod_{i=d+1}^{n}x_{|\mu_{j}|}&\le\beta\left(l(n+1)-l(d)-\mu_{n+1}\right)\\ \prod_{i=d+1}^{n}x_{|\mu_{j}|}&\le\beta\left(r+p-\mu_{n+1}-\mu_{d}\right) \end{align*} Now, using this equation with Lemma \ref{sizeofF} gives $$|F|\le s_{l(n)}\beta\left(r+p-\mu_{n+1}-\mu_{d}\right)\max_{u\in W}\left\{\alpha'(r,|\mu_{n+1}|)+\frac{x_{|\mu_{d}|}\alpha(|u|,|\mu_{n+1}|)}{1-x_{|u|}}\right\}$$ Now recall that $r=|\mu_{d}\mu_{d+1}\ldots \mu_{n+1}|-p$ and thus by definition of $d$, $1\le r\le\mu_d$. We deduce: $$|F|\le s_{l(n)}\max\limits_{\substack{u,v\in W \\1\le r\le |v|}}\left\{\beta\left(r+p-\mu_{n+1}-|v|\right)\left(\alpha'(r,|\mu_{n+1}|)+\frac{x_{|v|}\alpha(|u|,|\mu_{n+1}|)}{1-x_{|u|}}\right)\right\}$$ We can finally replace $|F|$ by this bound in inequality (\ref{eqrecs}) and we get: \begin{align*} &s_{l(n+1)}\ge s_{l(n)}\Bigg(f(|\mu_{n+1}|) -\\ &\max\limits_{\substack{u,v\in W \\r\in\{1,\ldots, |v|\}}}\left\{\beta\left(r+p-\mu_{n+1}-|v|\right)\left(\alpha'(r,|\mu_{n+1}|)+\frac{x_{|v|}\alpha(|u|,|\mu_{n+1}|)}{1-x_{|u|}}\right)\right\}\Bigg)\\ &s_{l(n+1)}\ge s_{l(n)}x_{|\mu_{n+1}|}^{-1} \text{\ \ \ (By Theorem hypothesis \eqref{condeq2})} \end{align*} Moreover $s_0=1$ and thus for all $i$, $s_{|\mu_1\ldots\mu_i|}\ge\prod_{j=1}^i x_{|\mu_j|}^{-1}$. For all $j$, $x_{|\mu_j|}^{-1}>1$, so we conclude that $S(\mu)$ is infinite. \end{proof} Remark that Theorem \ref{mainTh} is far from sharp. One could improve the bounds given by Lemma \ref{sizefi2}. This could be done by lowering $\alpha$ and $\alpha'$ or by introducing a third coefficient $\alpha''$ for the second non-empty $F_i$. However, we were not able to obtain significant improvement that were worth the additional technicalities. In Section \ref{secfindingL} we explain how to verify with a computer that there exists a language $L$ that satistfies conditions (I),(II) and (III). We also need a way to verify condition (IV). In order to compute $\beta$, we can use that $\beta(0)=1$ and, for all $j\in \left\{1,\ldots, p\right\}$, $\beta(j)=\max\left\{x_{|u|}\beta(j-|u|): u\in W, |u|\le j\right\}$. Thus given the values of the $x_i$ one can compute $\beta$ using a dynamic algorithm and all the rest is straight forward to compute. Thus it is easy to verify with a computer whether or not a given set of values of $x_i$ is a solution. We provide a C++ program that takes as input $|\Sigma|$, $k$, $p$, $f$ and $x_1,\ldots, x_k$ and verifies whether this is a solution of the equations of condition (IV). \subsection{Proof of Lemma \ref{sizefi2}}\label{secprooflemma} This subsection is dedicated to the proof of Lemma \ref{sizefi2}. Remark that the statement and proof are not self-contained since some of the notations are defined in the proof of Theorem \ref{mainTh}. \sizefii* \begin{proof} If $i> d$ then by definition $|\mu_{i}\ldots \mu_{n+1}|\le p$. Moreover, $L$ does not contain squares of period less than $p$ and thus $F_i=\emptyset$. Now, let $i\le d$. By definition, any word from $F_i$ can be written $uvvy$ with $l(i-1)<|uv|\le l(i), |y|<|\mu_{n+1}|$. For any $i$ and $z\in S(\mu)\cap L(\mu)\cap \{u\in\Sigma^*: |u|=l(i)\}$, let $F_i(z)$ be the set of words of $F_i$ that admit $z$ as a prefix. Clearly $F_i=\sum_{z\in S(\mu)\cap L(\mu)\cap \{u\in\Sigma^*: |u|=l(i)\}}F_i(z)$. Let $a,b$ (resp. $a',b'$) be integers such that there is an element $z'\in F_i(z)$ that contains a square starting at $a$ (resp. $a'$) and of period $b$ (resp. $b'$) with $l(i-1)+1<a'+b'=a+b\le |z|+1$ and $a>a'$. Because of the square in $z'$, we know that for all $0\le j\le |z|-a-b$, $z_{a+j}=z_{a+b+j}=z_{a'+b'+j}=z_{a'+j}$. If $a\le a'+|z|-a-b+1$ then $z$ contains a square which is not possible. Hence \begin{align} a&> a'+|z|-a-b+1\nonumber\\ a+b+b'&> a'+b+b'+|z|-a-b+1\nonumber\\ a'+2b'&> a+2b+|z|-a-b+1\ \text{ since }a+b=a'+b'\label{squaresamemiddle} \end{align} Let $u\in F_i(z)$ be a word that contains a square starting at $a$ and of period $b$ then we know its suffix of size $a+2b-1>l(n)$. Thus there are at most $f(|\mu_{n+1}|)$ possibilities, moreover since the size of the unknown suffix is $l(n+1)+1-a-2b<p$, it is square-free and there are at most $(|\Sigma|-1)^{l(n+1)+1-a-2b}$ possibilities. Thus for a fixed $z$ and value of $a+b$, the number of ways to add a suffix to $z$ to obtain an element of $F_i$ that contains a square of period $b$ starting at $a$ is: \begin{align*} &\max_{ \substack{j\in \mathbb{N}_0,\, l(n)+2\le p_1,\ldots, p_j\le l(n+1)+1,\\ \forall i<j,\, p_{i+1}>p_{i}+|z|-a-b+1 }} \left\{\sum_{i=1}^{j}\min\left\{f(|\mu_{n+1}|),(|\Sigma|-1)^{l(n+1)+1-p_i}\right\} \right\}\\ &= \max_ {\substack{j\in \mathbb{N}_0,\, 0\le p_1,\ldots, p_j\le |\mu_{n+1}|-1,\\ \forall i<j,\, p_{i+1}<p_{i}+|z|-a-b+1}} \left\{\sum_{i=1}^{j}\min\left\{f(|\mu_{n+1}|),(|\Sigma|-1)^{|\mu_{n+1}|-1-p_i}\right\}\right\} \end{align*} But since $\min\left\{f(|\mu_{n+1}|),(|\Sigma|-1)^{|\mu_{n+1}|-1-p_i}\right\}$ is a non-increasing function in $p_i$, the maximum is reached when we pack the $p_i$ as much as we can on the lowest values of $p_i$. Thus this quantity is in fact equal to: \begin{equation*} \sum_{j=0}^{\left\lfloor\frac{|\mu_{n+1}|-1}{|z|-a-b+2}\right\rfloor}\min\left\{f(|\mu_{n+1}|),(|\Sigma|-1)^{|\mu_{n+1}|-1-j(|z|-a-b+2)}\right\} \end{equation*} Now we can sum over all the values of $a+b$ and we get: \begin{align} |F_i(z)|&\le \sum_{a+b=|z|+2-|\mu_i|}^{|z|+1} \sum_{j=0}^{\left\lfloor\frac{|\mu_{n+1}|-1}{|z|-a-b+2}\right\rfloor}\min\left\{f(|\mu_{n+1}|),(|\Sigma|-1)^{|\mu_{n+1}|-1-j(|z|-a-b+2)}\right\}\nonumber\\ &\le \sum_{a+b=0}^{|\mu_{i}|-1} \left(\sum_{j=0}^{\left\lfloor\frac{|\mu_{n+1}|-1}{|\mu_i|-a-b}\right\rfloor}\min\left\{f(|\mu_{n+1}|),(|\Sigma|-1)^{|\mu_{n+1}|-1-j(|\mu_i|-a-b)}\right\}\right)\nonumber\\ &\le \sum_{m=1}^{|\mu_{i}|}\left( \sum_{j=0}^{\left\lfloor\frac{|\mu_{n+1}|-1}{m}\right\rfloor}\min\left\{f(|\mu_{n+1}|),(|\Sigma|-1)^{|\mu_{n+1}|-1-jm}\right\}\right)\nonumber\\ |F_i(z)|&\le \alpha(|\mu_{i}|,|\mu_{n+1}|)\nonumber \end{align} Summing over all the possible $z$ yields $|F_i|\le s_{l(i)}\alpha(|\mu_{i}|,|\mu_{n+1}|)$. The remaining case is $i=d$. Once again, let $a,b$ (resp. $a',b'$) be integers such that there is an element of $F_d(z)$ that contains a square starting at $a$ (resp. $a'$) and of period $b$ (resp. $b'$) with $l(d-1)+1<a'+b'=a+b\le l(n+1)+1-p$ and $a>a'$. By definition of $d$, $ l(n+1)- l(d)\le p$. We can use equation \eqref{squaresamemiddle} again and we get: $$a'+2b'> b+|z|+1\ge p+|z|+1 \ge l(n+1)+1$$ This is a contradiction with the fact that $a'+2b'\le l(n+1)+1$. Thus given the value of $a+b$ there is at most one possible value for $a$ and $b$. The number of ways for a fixed $z$ and value of $a+b$ to complete $z$ with a suffix into an element of $F_d$ is at most: \begin{align*} &\max\left\{\min\left\{f(|\mu_{n+1}|),(|\Sigma|-1)^{l(n+1)+1-s}\right\}: \begin{array}{l} s\in\mathbb{N}_{>0},s\le l(n+1)+1,\\ a+b+p\le s,\\ \end{array}\right\}\\ &\le \min\left\{f(|\mu_{n+1}|),(|\Sigma|-1)^{l(n+1)+1-a-b-p}\right\} \end{align*} Then by summing over all the possible values of $a+b$, we get: $$ |F_d(z)|\le \sum_{a+b=l(d-1)+2}^{l(n+1)+1-p} \min\left\{f(|\mu_{n+1}|),(|\Sigma|-1)^{l(n+1)+1-a-b-p}\right\}$$ We can use the variable substitution $m=l(n+1)+1-a-b-p$ and remark that $l(n+1)+1-p-(l(d-1)+2)=|\mu_{d}\mu_{d+1}\ldots \mu_{n+1}|-p-1=r-1$ and we get: $$ |F_d(z)|\le \sum_{m=0}^{r-1} \min\left\{f(|\mu_{n+1}|),(|\Sigma|-1)^{m}\right\}\le \alpha'(r,|\mu_{n+1}|)$$ We conclude that $|F_{d}|\le s_{l(d)}\alpha'(r,|\mu_{n+1}|)$ by summing over all $z$ \end{proof} \section{Finding a set \texorpdfstring{$L$}{L} that satisfies Theorem \ref{mainTh}}\label{secfindingL} In this section, we explain how to verify the existence of a language that fulfills conditions (I),(II) and (III) of Theorem \ref{mainTh}. We consider some particular directed labeled graphs: $G(V,A)$ is a set $V$ of vertices together with a set $A\subseteq(V\times V\times\Sigma)$ of labeled arcs. For any $u,v\in V$ and $a\in\Sigma$, $(u,v,a)\in A$ is an arc from $u$ to $v$ with label $a$. These graphs could also be seen as finite state machines where all the states are initial and final. The Rauzy graph of length $n$ of a factorial language $L$ over $\Sigma$ is the graph $G(V, A)$ where $V = L\cap \Sigma^n$ and $E=\{(au,ub, b): aub\in L, a,b\in\Sigma \}$. For any graph $G(V,A)$ and any set $X\in V$, we denote by $G[X]$ the subgraph induced by $X$. Let $R_p(\Sigma)$ be the Rauzy graph of length $2p-3$ of the square-free words over $\Sigma$. Remark, that the factors of length $2p-2$ of any walk on $R_p(\Sigma)$ correspond to edges of $R_p(\Sigma)$ and by definition they are square free. Thus, the sequence of labels of any walk on $R_p(\Sigma)$ avoids squares of period less than $p$, but can contain longer squares. We let $S_p(\Sigma)$ be the set of words that contains no square of period less than $p$ (from the previous remark $S_p(\Sigma)$ can also be seen as the set of walks on $R_p(\Sigma)$). As an illustration, we give $R_3(\{0,1,2\})$ in Fig. \ref{fullrauzy} without the arc labels. \begin{figure}[ht] \begin{center} \includegraphics[scale=0.8]{example_Rauzy-eps-converted-to.pdf} \end{center} \caption{The Rauzy graph $R_3(\{0,1,2\})$.\label{fullrauzy}} \end{figure} For this Section, we abuse the notation and allow ourself to identify words and sequences. For any graph $G(V,A)$ and partial word $w\in(\Sigma\cup\{\diamond\})^*$, we define inductively for any integer $i\in\{0,\ldots, |w|\}$ and vertex $v\in V$: $$p_{i,w,G}(v)=\left\{\begin{array}{ccl} 1 &&\text{if }i=0,\\ \sum\limits_{\substack{(v,u,a)\in A\\a\in \Sigma}}p_{i-1,w,G}(u) &&\text{if }w_{|w|+1-i}=\diamond,\\ \sum\limits_{\substack{(v,u,w_{|w|+1-i})\in A}}p_{i-1,w,G}(u) &&\text{otherwise.} \end{array}\right. $$ Intuitively, $p_{i,w,G}(v)$ gives the number of walks of length $i$ starting from $v$ that are compatible with the $i$ last letters of $w$. Indeed, there is one walk of length $0$ and we always take the transition that is labeled by the current letter of $w$ and any transition if this letter is $\diamond$. Remark that in the third case, there are in fact either $0$ or $1$ summands in the sum. \begin{lemma}\label{firstRauzy} Let $W\subseteq(\Sigma\cup\{\diamond\})^*$, $G(V,A)=R_p(\Sigma)$, $f:\mathbb{N}_{>0}\rightarrow\mathbb{N}_{>0}$ and a non-empty set $X\subseteq V$. If for all $v\in X$ and $w\in W$, $p_{|w|,w,G[X]}(v)\ge f(|w|)$, then there exists a language $L$ such that: \begin{itemize} \item $\varepsilon\in L$, \item for all $u\in L$, $u$ avoids squares of period less than $p$, \item for any $u\in L$ and $w\in W$ there are at least $f(|w|)$ different words $v\in\Sigma^{|w|}$ compatible with $w$ such that $uv\in L$. \end{itemize} \end{lemma} \begin{proof} Let $L$ be the set of sequences of labels that correspond to a walk in $G[X]$. By definition, the two first conditions on $L$ are fulfilled. Let $u\in L$ and $w\in W$. If $|u|\ge2p-3$ we let $u'$ be the suffix of length $2p-3$ of $u$. Otherwise, we let $u'\in L$ such that $|u'|=2p-3$ and $w$ is a suffix of $u'$ (there is such an element in $L$). Each walk of length $|w|$ starting in $u'$ gives a unique sequence of labels $u''$ such that $uu''$ contains no square of period less than $p$. We easily deduce, by induction on $i$, that for all $v$ the number of walks of length $i$ starting at $v$ that are compatible with $w_{|w|-i+1}w_{|w|-i+2}\ldots w_{|w|}$ is at least $p_{i,w,G[X]}(v)$. So, in particular, the number of walks of length $|w|$ starting at $u'$ and compatible with $w$ is at least $p_{|w|,w,G[X]}(u')\ge f(|w|)$. This concludes the proof. \end{proof} In fact, we need something stronger because for the values of $p$ that we use the graphs $R_p$ are too big to fit in a computer. We can exploit symmetries of $R_p(\Sigma)$ to work on a smaller equivalent graph. For any square-free word $w\in\Sigma^*$, we let $\Psi(w)$ be the shortest suffix of $w$ such that for all $i\in\{\lceil\frac{|\Psi(w)|}{2}\rceil+1,\ldots,p-1\}$ there exists $k\in\{0,1,\ldots, |\Psi(w)|-i-1\}$ with $\Psi(w)_{|\Psi(w)|-k}\not=\Psi(w)_{|\Psi(w)|-k-i}$. If $|w|=2p-3$, then $w$ is such a suffix of itself (since $\{\lceil\frac{|w|}{2}\rceil+1,\ldots,p-1\}$ is empty) and thus there is always a shortest suffix. For instance, with $p=5$ we have $\Psi(0210120)=10120$. For any letter $\alpha$, the word $0210120\alpha$ is square free if and only if $10120\alpha$ is square free. Indeed, a new square is necessarily a suffix of $0210120\alpha$, it is enough to look at the two letters in bold in $\mathbf{1}012\mathbf{0}\alpha$ to deduce that there is no square of length $4$ and all the other possible suffix of even length of $0210120\alpha$ are also suffixes of $10120\alpha$. In fact, for any $w$ of size $2p-3$, the word $wa$ avoids squares if and only if $\Psi(w)a$ avoids squares (this is proven in the next lemma) and this is the main motivation behind the definition of $\Psi$ (in particular, words with the same image by $\Psi$ can be extended in the same way). For any graph $G(V,A)$, let $\Psi(G)(\Psi(V),A')$ be the graph such that $A'=\Psi(A)=\{(\Psi(a),\Psi(b),c): (a,b,c)\in A\}$. The next lemma tells us that we only need to consider the walks on $\Psi(G)$ instead of the walks on $G$. \begin{lemma}\label{lemmaforf} Let $p$ be a positive integer and $w\in(\Sigma\cup\{\diamond\})^*$, $G(V,A)=R_p(\Sigma)$ and $X\subseteq \Psi(V)$. Let $\Psi^{-1}(X)=\{x\in V: \Psi(x)\in X\}$. Then for all $v\in \Psi^{-1}(X)$, $p_{|w|,w,G[\Psi^{-1}(X)]}(v)= p_{|w|,w,\Psi(G)[X]}(\Psi(v))$. \end{lemma} \begin{proof} Let us first show that for any $a\in\Sigma$ and $v\in S_p(\Sigma)$ with $|v|=2p-3$ if $\Psi(v)a\in S_p(\Sigma)$ then $va\in S_p(\Sigma)$. Let us show that under these assumptions, for any $i$, $va$ avoids squares of period $i$. Since $v$ is square free, we only need to show that no suffix of $va$ is a square. We have to distinguish between two cases: \begin{itemize} \item $2i\le|\Psi(v)|+1$. Suppose for the sake of contradiction that there is a square of period $i$ in $va$. We deduce that the suffix of length $|\Psi(v)|+1$ of $va$ contains a square of period $i$. That is, $\Psi(v)a$ contains a square of period $i$ which is a contradiction. \item $2i\ge|\Psi(v)|+2$. Since $i$ is an integer we get $i\ge\lceil\frac{|\Psi(v)|}{2}\rceil+1$. Moreover $|va|=2p-2$ and thus $i\le p-1$. Thus by definition of $\Psi(v)$, there exists $k\in\{0,1,\ldots, |\Psi(v)|-i-1\}$ such that $\Psi(v)_{|\Psi(v)|-k}\not=\Psi(v)_{|\Psi(v)|-k-i}$. Thus there is $k\in\{1,\ldots, |\Psi(v)a|-i-1\}$ such that $(\Psi(v)a)_{|\Psi(v)a|-k}\not=(\Psi(v)a)_{|\Psi(v)a|-k-i}$. Remark, that $|\Psi(v)a|-i-1=|\Psi(v)|-i\le 2i-2-i= i-2$. We conclude that there is $k\in\{0,\ldots, i-2\}$ such that $(va)_{|va|-k}\not=(va)_{|va|-k-i}$. This implies that the suffix of length $2i$ of $va$ is not a square of period $i$. \end{itemize} We deduce that for any $a\in\Sigma$ and $v\in S_p(\Sigma)$ with $|v|=2p-3$ if $\Psi(v)a\in S_p(\Sigma)$ then $va\in S_p(\Sigma)$. Let $u\in V$, $v\in\Psi(V)$ and $a\in\Sigma$ such that $(\Psi(u),v,a)\in \Psi(A)$. By definition of $\Psi(A)$ this implies that there is $(u',v',a)\in A$ with $\Psi(u')=\Psi(u)$ and $\Psi(v')=v$. Thus $u'a\in S_p(\Sigma)$ and $\Psi(u)a=\Psi(u')a\in S_p(\Sigma)$. From the previous paragraph, it implies that $ua$ is square-free. Let us show that $\Psi(u_2u_3\ldots u_{|u|}a)=v$. By definition, for all $i\in\{\lceil\frac{|\Psi(u')|}{2}\rceil+1,\ldots,p-1\}$ there exists $k\in\{0,1,\ldots, |\Psi(u')|-i-1\}$ with $\Psi(u')_{|\Psi(u')|-k}\not=\Psi(u')_{|\Psi(u')|-k-i}$. We easily deduce that for all $i\in\{\lceil\frac{|\Psi(u')a|}{2}\rceil+1,\ldots,p-1\}$ there exists $k\in\{0,1,\ldots, |\Psi(u'a)|-i-1\}$ with $\Psi(u'a)_{|\Psi(u'a)|-k}\not=\Psi(u'a)_{|\Psi(u'a)|-k-i}$. This implies that $v=\Psi(v')$ is a suffix of $\Psi(u')a$. Since $\Psi(u')a= \Psi(u)a$, we deduce that $v$ is also a suffix of $u_2u_3\ldots u_{|u|}a$. Since $v=\Psi(v')$ and $v$ is a suffix of $u_2u_3\ldots u_{|u|}a$, we get that $\Psi(u_2u_3\ldots u_{|u|}a)=v$. We showed that if there are $u\in V$, $v\in\Psi(V)$ and $a\in\Sigma$ such that $(\Psi(u),v,a)\in \Psi(A)$, then there exists $v''$ such that $(u,v'',a)\in A$ and $\Psi(v'')=v$. We deduce that for all $u\in V$, $\{(\Psi(u),\Psi(v),a)\in \Psi(A)\}\subseteq \{(\Psi(u),\Psi(v),a): (u,v,a)\in A\}$. The other inclusion is clear from the definition of $\Psi(A)$ and we get for all $u\in V$ : \begin{equation}\label{eqfA} \{(\Psi(u),\Psi(v),a)\in \Psi(A)\}= \{(\Psi(u),\Psi(v),a): (u,v,a)\in A\} \end{equation} By definition of $R_p(\Sigma)$, for any $u$ there is at most one outgoing arc for every label in the set of the right. Since the two sets are equals, we deduce that every vertex of the graph $\Psi(G)$ has at most one outgoing arc for any label. Intuitively, \eqref{eqfA} implies (by induction on the length of the walk) that for any $u,v\in V$ the set of labeled walks from $u$ to $v$ in $G$ is equal to the set of labeled walks from $\Psi(u)$ to $\Psi(v)$ in $\Psi(G)$. We are now ready to show by induction on $i$ that for all $i\in\{0,\ldots, |w|\}$ and $v\in \Psi^{-1}(X)$, $p_{i,w,G[\Psi^{-1}(X)]}(v)= p_{i,w,\Psi(G)[X]}(\Psi(v))$. By definition $p_{0,w,G[\Psi^{-1}(X)]}(v)=1= p_{0,w,\Psi(G)[X]}(\Psi(v))$. Let $n$ be a positive integer such that for all $v\in\Psi^{-1}(X)$, \begin{equation}\tag{IH}\label{IHend} \forall i<n,\, p_{i,w,G[\Psi^{-1}(X)]}(v)= p_{i,w,\Psi(G)[X]}(\Psi(v)). \end{equation} Then, for all $v\in \Psi^{-1}(X)$, if $w_{|w|+1-i}\not=\diamond$ we get: \begin{align*} p_{i,w,G[\Psi^{-1}(X)]}(v)&=\sum\limits_{\substack{(v,u,w_{|w|+1-i})\in A\\u\in \Psi^{-1}(X)}}p_{i-1,w,G[\Psi^{-1}(X)]}(u)\\ p_{i,w,G[\Psi^{-1}(X)]}(v)&=\sum\limits_{\substack{(v,u,w_{|w|+1-i})\in A\\\Psi(u)\in X}}p_{i-1,w,\Psi(G)[X]}(\Psi(u))\text{\ \ \ (From \eqref{IHend})}\\ p_{i,w,G[\Psi^{-1}(X)]}(v)&=\sum\limits_{\substack{(\Psi(v),\Psi(u),w_{|w|+1-i})\in \Psi(A)\\\Psi(u)\in X}}p_{i-1,w,\Psi(G)[X]}(\Psi(u)) \text{\ \ \ (From \eqref{eqfA})}\\ p_{i,w,G[\Psi^{-1}(X)]}(v)&=\sum\limits_{\substack{(\Psi(v),u,w_{|w|+1-i})\in \Psi(A)\\u\in X}}p_{i-1,w,\Psi(G)[X]}(u)\\ p_{i,w,G[\Psi^{-1}(X)]}(v)&=p_{i,w,\Psi(G)[X]}(\Psi(v)) \end{align*} The case where $w_{|w|+1-i}=\diamond$ is similar. \end{proof} Using Lemma \ref{firstRauzy} together with Lemma \ref{lemmaforf}, we get the following lemma: \begin{lemma}\label{lemmagraphf} Let $W\subseteq(\Sigma\cup\{\diamond\})^*$, $G(V,A)=R_p(\Sigma)$, $f:\mathbb{N}_{>0}\rightarrow\mathbb{N}_{>0}$ and $X\subseteq \Psi(V)$ be a non-empty set. If for all $v\in X$ and $w\in W$, $p_{|w|,w,\Psi(G)[X]}(v)\ge f(|w|)$, then there exists a language $L$ such that: \begin{itemize} \item $\varepsilon\in L$, \item for all $u\in L$, $u$ avoids squares of period less than $p$, \item for any $u\in L$ and $w\in W$ there are at least $f(|w|)$ different words $v\in\Sigma^{|w|}$ compatible with $w$ and such that $uv\in L$. \end{itemize} \end{lemma} The graph $\Psi(R_p(\Sigma))$ is much smaller than $R_p(\Sigma)$ and we can use a computer to check the conditions of this lemma for the values of $p$ that we used. One should first find the graph $\Psi(R_p(\Sigma))$. The following fact allows us to easily compute the set of vertices of $\Psi(R_p(\Sigma))$ without computing $R_p(\Sigma)$: \begin{lemma}\label{compgfr} Let $w\in S_p(\Sigma)$. Then $w \in \Psi(S_p(\Sigma)\cap\Sigma^{2p-3})$ if and only if $w$ is the smallest non-empty suffix of $w$ such that $\Psi(w)=w$. \end{lemma} Moreover, given a graph $G$, the definition of $p_{|w|,w,G}$ gives a trivial dynamic algorithm that computes $p_{|w|,w,G}$ in time $O(|\Sigma|\cdot |w|\cdot|G|)$. Starting with $X=\Psi(R_p(\Sigma))$ and inductively removing from $X$ all the vertices for which $p_{|w|,w,\Psi(R_p(\Sigma))[X]}< f(|w|)$ gives the largest subgraph that meets the conditions of Lemma \ref{lemmagraphf}. As long as this subgraph is not empty one can then apply Lemma \ref{lemmagraphf}. Algorithm \ref{algosubgraph} computes the largest subgraph of $\Psi(G)$ with the required property. \begin{algorithm}[ht] \SetKwInOut{Input}{Input} \SetKwInOut{Output}{Output} \Input{The graph $\Psi(G)$, the set $W$} \Output{The largest set $X\subseteq \Psi(V)$ such that for all $v\in X$ and $w\in W$, $p_{|w|,w,\Psi(G)[X]}(v)\ge f(|w|)$} $X=\Psi(V)$\; $todo := true$\; \While{todo}{ $todo := false$\; \ForEach{$w\in W$}{ compute $p_{|w|,w,\Psi(G)[X]}$\; $X':=\{v\in X: p_{|w|,w,\Psi(G)[X]}(v)\ge f(|w|)\}$\; \If{$X\not=X'$}{ $X:=X'$\; $todo := true$\; } } } \KwRet $X$\; \caption{How to compute the subgraph of $\Psi(G)$.\label{algosubgraph}} \end{algorithm} \section{Application of Theorem \ref{mainTh}}\label{secconc} In this section we apply Theorem \ref{mainTh}. We provide a C++ implementation of Algorithm \ref{algosubgraph} that verifies the existence of the language that fulfill conditions (I),(II) and (III) from Theorem \ref{mainTh}. Condition (IV) can be easily verified (as long as solutions are given) and we also provide a C++ code to do that. \footnote{The codes can be found in the ancillary files of \url{https://arxiv.org/abs/1903.04214}} \begin{theorem} For any alphabet $\Sigma$, let $d(\Sigma)$ be the smallest integer such that for all $v\in\Sigma^\omega$ and for all sequence $(p_i)_{1\le i}$ such that $\forall i$, $p_{i+1}-p_i\ge d(\Sigma)$, there is an infinite square-free word $u\in\Sigma^\omega$ such that $v_i= u_{p_i}$. Then: \begin{itemize} \item $7\le d(\{0,1,2\})\le19$, \item $d(\{0,1,2,3\})=3$, \item $2\le d(\{0,1,2,3,4\})\le3$, \item if $|\Sigma|\ge6$, $d(\Sigma)=2$. \end{itemize} \end{theorem} \begin{proof} For any alphabet $\Sigma$, we have $d(\Sigma)\ge2$. Moreover, $d$ is a decreasing function of the size of the alphabet. Thus the third statement can easily be deduced from the second one. We show the remaining statements independently of each others. We will show the upper bounds using Algorithm \ref{algosubgraph}, Lemma \ref{lemmagraphf} and Theorem \ref{mainTh}. The lower bounds are verified by exhaustive search.\\ \item If $\mathbf{|\Sigma|\ge6,\, d(\Sigma)=2}$: Let $\Sigma = \{0,1,2,3,4,5\}$ and $W=\{{\diamond}\}\cup\{{\diamond} a{\diamond} : a\in\Sigma\}\cup\{{\diamond} a{\diamond} b: a,b\in\Sigma\}.$ We can use Algorithm \ref{algosubgraph} to check that we can apply Lemma \ref{lemmagraphf} with $f(1)=3$, $f(3)=f(4)=6$, $p=12$. Thus conditions (I),(II) and (III) of Theorem \ref{mainTh} are fullfilled. We can check with a computer that condition \eqref{condeq2} of Theorem \ref{mainTh} is also fullfilled with $x_1=\frac{2}{5}$ and $x_3=x_4=\frac{1}{4}$. This implies that for any $\mu\in W^\omega$ there are infinite square-free words over $\Sigma$ compatible with $\mu$. Since $\{{\diamond} ^i a : i\ge1, a\in \Sigma \}^\omega\subseteq W^\omega$, we deduce that for any $\mu\in \{{\diamond} ^i a : i\ge1, a\in \Sigma \}^\omega$ there are infinite square-free words over $\Sigma$ compatible with $\mu$. We get $d(\{0,1,2,4,5,6\})\le2$.\\ \item $\mathbf{d(\{0,1,2,3\})=3}$: Let $w=(0{\diamond}1{\diamond}2{\diamond}3{\diamond})^\omega$. An exhaustive search confirms that there are only $636$ square-free words over $\{0,1,2,3\}$ compatible with $w$. Thus $d(\{0,1,2,3\})\ge3$. Let $\Sigma = \{0,1,2,3\}$ and $W=\{{\diamond}\}\cup\{{\diamond}{\diamond} a{\diamond}: a\in\Sigma\} \cup\{{\diamond}{\diamond}a{\diamond}{\diamond}b: a,b\in\Sigma\}.$ We can use Algorithm \ref{algosubgraph} to check that we can apply Lemma \ref{lemmagraphf} with $f(1)=2$, $f(4)=5$ and $f(6)=8$, $p=18$. We can then apply Theorem \ref{mainTh} with $x_1=\frac{11}{20}$, $x_4=\frac{1}{4}$ and $x_6=\frac{1}{5}$ and we deduce that for any $\mu\in W^\omega$ there are infinite square-free words over $\Sigma$ compatible with $\mu$. Moreover, $\{\diamond ^i a : i\ge2, a\in \Sigma \}^\omega\subseteq W^\omega.$ We deduce that for any $\mu\in \{\diamond ^i a : i\ge2, a\in \Sigma \}^\omega$ there are infinite square-free words over $\Sigma$ compatible with $\mu$. Thus $d(\{0,1,2,3\})\le3$.\\ \item $\mathbf{7\le d(\{0,1,2\})\le 19}$: Let $w=(0{\diamond}^51{\diamond}^52{\diamond}^5)^\omega$. An exhaustive search confirms that there are only $4281$ square-free words over $\{0,1,2\}$ compatible with $w$. Thus $d(\{0,1,2\})\ge7$. Let $\Sigma = \{0,1,2\}$ and $W=\{\diamond^9\}\cup\{\diamond^i a: i\in\{18,\ldots, 26\}, a\in\Sigma\}.$ We can use Algorithm \ref{algosubgraph} to check that we can apply Lemma \ref{lemmagraphf} with $p=61$ and the values of $f$ given in Table \ref{tabfx}. Thus conditions (I),(II) and (III) of Theorem \ref{mainTh} are fullfilled. \begin{table}[htb] \center \begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|} \hline $|w|$&9&19&20&21&22&23&24&25&26&27\rule{0pt}{12pt}\\[2px]\hline $f(|w|)$&4&19&22&28&36&50&63&88&118&148\rule{0pt}{12pt}\\[2px]\hline $x_{|w|}$&$\frac{27}{100}$ &$\frac{7}{100}$ &$\frac{13}{200}$ &$\frac{11}{200}$ &$\frac{9}{200}$ &$\frac{1}{25}$&$\frac{3}{100}$ &$\frac{1}{40}$ &$\frac{1}{40}$ &$\frac{1}{50}$\rule{0pt}{14pt}\\[4px] \hline \end{tabular} \caption{The values of $f(|w|)$ and $x_{|w|}$ for the computation of $d(\{0,1,2\})$} \label{tabfx} \end{table} We can also check that the values of $x_{|w|}$ given in Table~\ref{tabfx} fulfill condition \eqref{condeq2} of Theorem \ref{mainTh}. We deduce that for any $\mu\in W^\omega$ there are infinite square-free words over $\Sigma$ compatible with $\mu$. Moreover, $\{\diamond ^i a : i\ge18, a\in \Sigma \}^\omega\subseteq W^\omega.$ We deduce that for any $\mu\in \{\diamond ^i a : i\ge18, a\in \Sigma \}^\omega$ there are infinite square-free words over $\Sigma$ compatible with $\mu$. Thus $d(\{0,1,2\})\le19$ \end{proof} The three applications of Algorithm \ref{algosubgraph} require between 30 and 100GB of RAM (and around 5 hours of computations). We had to optimize the way strings are stored in memory in order to be able to compute the graphs for large enough values of $p$. The rest of the computations (finding the solution to the system and the exhaustive search) easily run on a laptop in a few milliseconds. Remark that we showed something slightly stronger since the results would still hold if an adversary was to tell us at every choice of letter only the next $5$ forced letters with their positions (that is, we know the next element of $W$). Experimental computations suggest that $d(\{0,1,2\})$ is closer to 7 than to $19$ and that $d(\{0,1,2,3,4\})=2$. \subsection*{Acknowledgement} Computational resources have been provided by the Consortium des \'Equipements de Calcul Intensif (C\'ECI), funded by the Fonds de la Recherche Scientifique de Belgique (F.R.S.-FNRS) under Grant No. 2.5020.11 and by the Walloon Region

https://arxiv.org/abs/1901.01691

Dimension of invariant measures for affine iterated function systems

Let $\{S_i\}_{i\in \Lambda}$ be a finite contracting affine iterated function system (IFS) on ${\Bbb R}^d$. Let $(\Sigma,\sigma)$ denote the two-sided full shift over the alphabet $\Lambda$, and $\pi:\Sigma\to {\Bbb R}^d$ be the coding map associated with the IFS. We prove that the projection of an ergodic $\sigma$-invariant measure on $\Sigma$ under $\pi$ is always exact dimensional, and its Hausdorff dimension satisfies a Ledrappier-Young type formula. Furthermore, the result extends to average contracting affine IFSs. This completes several previous results and answers a folklore open question in the community of fractals. Some applications are given to the dimension of self-affine sets and measures.

\section{Introduction}\label{S-1} \subsection{Motivation and the main result} Let ${\Bbb R}^{d\times d}$ denote the set of real $d\times d$ matrices. By an {\it affine iterated function system} (affine IFS) on ${\Bbb R}^d$ we mean a finite family ${\mathcal S}=\{S_j\}_{j\in \Lambda}$ of affine mappings from ${\Bbb R}^d$ to ${\Bbb R}^d$, taking the form \begin{equation} \label{e-form} S_j(x)=M_jx+a_j,\qquad j\in \Lambda, \end{equation} where $M_j\in {\Bbb R}^{d\times d}$ and $a_j\in {\Bbb R}^d$. Here, in contrast to the usual definition of affine IFS, we don't assume that $M_j$ are invertible or contracting (in the sense that $\|M_j\|<1$ where $\|\cdot\|$ is the matrix operator norm). We say that ${\mathcal S}$ is {\it contracting} if all $M_j$ are contracting. It is well-known that if $\mathcal S$ is contracting, there exists a unique non-empty compact set $K\subset {\Bbb R}^d$ such that $$ K=\bigcup_{j\in \Lambda} S_j(K). $$ We call $K$ the {\em self-affine set} generated by ${\mathcal S}$. In particular, if all the maps in ${\mathcal S}$ are contracting similitudes, we call $K$ a {\em self-similar set}. Let $(\Sigma,\sigma)$ be the two-sided full shift over the alphabet $\Lambda$, i.e.~$\Sigma=\Lambda^{\Bbb Z}$ and $\sigma:\Sigma\to \Sigma$ is the left shift map. Endow $\Sigma$ with the discrete product topology and let ${\mathcal M}_\sigma(\Sigma)$ denote the space of $\sigma$-invariant Borel probability measures on $\Sigma$. \begin{de} \label{de-1.1} Let $m\in {\mathcal M}_\sigma(\Sigma)$. An affine IFS ${\mathcal S}=\{M_jx+a_j\}_{j\in \Lambda}$ is said to be average contracting with respect to $m$ if, for $m$-a.e.~$x=(x_n)_{n=-\infty}^\infty\in\Sigma$, the top Lyapunov exponent $\lambda(x)$ defined by \begin{equation*} \label{e-lyapunov} \lambda(x)=\lim_{n\to \infty} \frac{1}{n}\log \|M_{x_{0}}\cdots M_{x_{n-1}}\| \end{equation*} is strictly negative. \end{de} We remark that the above limit in defining $\lambda(x)$ exists and takes values in $[-\infty, \infty)$ for $m$-a.e.~$x$. This follows from the Furstenberg-Kesten theorem \cite{FurstenbergKesten1960} or Kingman's sub-additive ergodic theorem \cite{Kingman1968}. Now let $m\in {\mathcal M}_\sigma(\Sigma)$ and suppose that ${\mathcal S}$ is average contracting with respect to $m$. The canonical coding map $\pi: \Sigma\to {\Bbb R}^d$, given by \begin{equation} \label{e-pi1.4} \begin{split} \pi(x)&=\lim_{n\to \infty} S_{x_0}\circ S_{x_1}\circ \cdots\circ S_{x_{n}}(0)\\ &=\lim_{n\to \infty} \left(a_{x_0}+M_{x_0}a_{x_1}+\cdots +M_{x_{0}}\cdots M_{x_{n-1}}a_{x_n}\right), \end{split} \end{equation} is well-defined on $\Sigma$ up to a set of zero $m$-measure (\cite{Brandt1986, BougerolPicard1992}); see Section \ref{S-3} for a self-contained proof. The push-forward measure $m\circ \pi^{-1}$, given by $$ m\circ \pi^{-1}(F)=m(\pi^{-1}(F)) \quad \mbox{ for any Borel set }F\subset{\Bbb R}^d, $$ is called an {\it invariant measure} or {\it stationary measure} for ${\mathcal S}$. When $m$ is ergodic, $m\circ \pi^{-1}$ is called an {\it ergodic invariant measure} for $\mathcal S$. Moreover if $m$ is a Bernoulli product measure, $m\circ \pi^{-1}$ is called a {\it self-affine measure} generated by ${\mathcal S}$; if in addition, ${\mathcal S}$ consists of similarities, then $m\circ \pi^{-1}$ is called a {\em self-similar measure}. The main purpose of this paper is to study the dimension of invariant measures for affine IFSs. Recall that for a probability measure $\eta$ on a metric space $X$, the {\it local upper and lower dimensions} of $\eta$ at $x\in X$ are defined respectively by $$\overline{\dim}_{\rm loc}(\eta, x)=\limsup_{r\to 0}\frac{\log \eta (B(x,r))}{\log r},\quad \underline{\dim}_{\rm loc}(\eta, x)=\liminf_{r\to 0}\frac{\log \eta (B(x,r))}{\log r},$$ where $B(x,r)$ stands for the closed ball of radius $r$ centered at $x$. If $$\overline{\dim}_{\rm loc}(\eta, x)=\underline{\dim}_{\rm loc} (\eta, x),$$ the common value is denoted as $\dim_{\rm loc}(\eta,x)$ and is called the {\it local dimension} of $\eta$ at $x$. We say that $\eta$ is {\it exact dimensional} if there exists a constant $C$ such that the local dimension $\dim_{\rm loc}(\eta, x)$ exists and equals $C$ for $\eta$-a.e.~$x\in X$. It is well-known that if $\eta$ is an exact dimensional measure in ${\Bbb R}^d$, the Hausdorff and packing dimensions of $\eta$ coincide and are equal to the involved constant $C$, and so are some other notions of dimension (e.g.~ entropy dimension); see \cite{Young1982, Falconer1997}. Recall that the Hausdorff and packing dimensions of $\eta$ are defined by \begin{align*} {\dim}_{\rm H}\eta &=\inf \{\dim_{\rm H}F:\; \eta(F)>0\mbox{ and $F$ is a Borel set}\},\\ {\dim}_{\rm P}\eta &=\inf \{\dim_{\rm P}F:\; \eta({\Bbb R}^d\setminus F)=0\mbox{ and $F$ is a Borel set}\}, \end{align*} where $\dim_{\rm H}F, \dim_{\rm P}F$ stand for the Hausdorff and packing dimensions of $F$, respectively (cf.~\cite{Falconer2003}). A folklore open problem in fractal geometry asks whether every ergodic invariant measure for an affine IFS is exact dimensional. As the main result of this paper, we give the following confirmative answer. \begin{thm} \label{thm-1.0} Let ${\mathcal S}=\{M_jx+a_j\}_{j\in \Lambda}$ be an affine IFS on ${\Bbb R}^d$ and $m\in {\mathcal M}_\sigma(\Sigma)$. Suppose that ${\mathcal S}$ is average contracting with respect to $m$. Let $\mu=m\circ \pi^{-1}$. Then \begin{itemize} \item[(i)] $\dim_{\rm loc}(\mu, x)$ exists for $\mu$-a.e.~$x\in {\Bbb R}^d$. \item[(ii)] Assume furthermore that $m$ is ergodic. Then $\mu$ is exact dimensional and $\dim_{\rm H}\mu$ satisfies a Ledrappier-Young type dimension formula. \end{itemize} \end{thm} The precise dimension formula of $\mu$ and some of its applications will be given in Sections~\ref{S-DF}-\ref{S-SC}. Below we first give some background information about the above study. The problem of the existence of local dimensions has a long history in smooth dynamical systems, as well as in the study of IFSs. It is of great importance in dimension theory of dynamical systems and fractal geometry. In \cite{Young1982}, Young proved that an ergodic hyperbolic measure invariant under a $C^{2}$ surface diffeomorphism is always exact dimensional. (Here by hyperbolic one means that the measure has no zero Lyapunov exponent.) For a hyperbolic measure $\mu$ in high-dimensional $C^{2}$ systems, Ledrappier and Young \cite{LedrappierYoung1985} proved the existence of $\delta^u$ and $\delta^s$, the local dimensions along stable and unstable local manifolds, respectively, and the upper local dimension of $\mu$ is bounded by the sum of $\delta^u$ and $\delta^s$; moreover they obtained a formula for $\delta^u$ and $\delta^s$ in terms of conditional entropies and Lyapunov exponents, which nowadays is called ``Ledrappier-Young formula". Eckmann and Ruelle \cite{EckmannRuelle1985} indicated that it is unknown whether the local dimension of $\mu$ is equal to the sum of $\delta^u$ and $\delta^s$ if $\mu$ is a hyperbolic measure. Then the problem was referred as Eckmann-Ruelle conjecture, and was finally answered confirmatively by Barreira, Pesin and Schmeling in 1999 for $C^{1+\alpha}$ diffeomorphisms \cite{BarreiraPesinSchmeling1999}. Later, the result of exact dimensionality was further extended by Qian and Xie \cite{QianXie2008} and Shu \cite{Shu2010} to $C^2$ expanding endomorphisms and $C^2$ non-degenerate endomorphisms, respectively. For the study of IFSs, it is well-known that if ${\mathcal S}$ is a contractive IFS consisting of similarity maps, or more generally, a contracting $C^1$ conformal IFS, then under an additional separation condition (the so-called open set condition), the push-forward measure of any ergodic invariant measure under the coding map is exact dimensional with dimension given by the classical entropy divided by the Lyapunov exponent (cf.~\cite{Bedford1991,Hutchinson1981,Patzschke1997}). The result essentially follows from the Shannon-McMillan-Breiman theorem in entropy theory. However, the problem becomes much more complicated without assuming the open set condition. In \cite{FengHu2009}, by introducing a notion of projection entropy and adopting some ideas from \cite{LedrappierYoung1985}, Feng and Hu proved that for any contracting $C^1$ conformal IFS, the push-forward measure of every ergodic invariant measure under the coding map is exact dimensional, with dimension given by the projection entropy divided by the Lyapunov exponent. Later this result was further extended to some random self-similar measures \cite{FalconerJin2014, SagliettiShmerkinSolomyak2018} and push-forward measures of ergodic invariant measures for some random conformal IFSs \cite{MihailescuUrbanski2016}. It is worth pointing out that the exact dimensionality of overlapping self-similar measures was first claimed by Ledrappier; nevertheless no proof has been written out (cf.~\cite[p.~1619]{PeresSolomyak2000}). This property was also conjectured later by Fan, Lau and Rao in \cite{FanLauRao2002}. The first result for affine IFSs is due to Bedford \cite{Bedford1984} and McMullen \cite{McMullen1984}, who independently calculated the Hausdorff and box-counting dimensions of a special class of planar self-affine sets (which are now called Bedford-McMullen carpets) and showed that they are usually different. McMullen \cite{McMullen1984} also implicitly proved the exact dimensionality of self-affine measures on the Bedford-McMullen carpets, and calculated the precise value of the dimension. Later, Gatzouras and Lalley \cite{GatzourasLalley1992} and Bara\'{n}ski \cite{Baranski2007} obtained similar results for a class of more general carpet-like self-affine sets in the plane. In \cite{KenyonPeres1996}, Kenyon and Peres extended Bedford and McMullen's result to higher dimensional self-affine carpets, and moreover, they proved the exact dimensionality and gave a Ledrappier-Young type dimension formula for arbitrary ergodic invariant measure on these carpets. For more related results on carpet-like self-affine sets, see the survey paper \cite{Falconer2013}. In \cite{FengHu2009}, Feng and Hu proved that for each contracting invertible affine IFS in ${\Bbb R}^d$, Theorem \ref{thm-1.0} holds under an additional assumption that the linear parts of the IFS commute (i.e. $M_iM_j=M_jM_i$). It remained open whether this additional assumption could be removed. Very recently, B\'{a}r\'{a}ny and K\"{a}enm\"{a}ki \cite{BaranyKaenmaki2015} made a substantial progress. They proved that for contracting invertible affine IFSs, every planar self-affine measure (more generally, every self-affine measure in ${\Bbb R}^d$ having $d$ distinct Lyapunov exponents) is exact dimensional, and moreover, under certain domination condition on the linear parts $\{M_j\}$, the push-down of every quasi-Bernoulli measure on the self-affine set is exact dimensional, with dimension given by a Ledrappier-Young type formula. Some other partial results were also obtained in \cite{Barany2015, Rapaport2015, FraserJordanJurga2017}. Along another direction, it is proved that for a given ergodic $m\in {\mathcal M}_{\sigma}(\Sigma)$, $m\circ \pi^{-1}$ is exact dimensional for ``almost all'' contracting invertible affine IFSs satisfying $\|M_j\|<1/2$ (\cite{JordanPollicottSimon2007, Jordan2011, Rossi2014}); however, the result does not apply to any concrete case. Theorem \ref{thm-1.0} finally gives a full confirmative answer to the problem of the existence of local dimensions in the context of affine IFSs. It completes the aforementioned previous works on the problem. Exact dimensionality and Ledrappier-Young type dimension formula play a significant role in many of the recent advances in dimension theory of self-affine sets and measures (see e.g.~\cite{Barany2015, BaranyHochmanRapaport2017, BaranyKaenmaki2015, BKK2018, BaranyRams2018, BaranyRamsSimon2016, BRS2018, BarralFeng2013, DasSimmons2017, FalconerKempton2018, HochmanSolomyak2017, MorrisShmerkin2018, PrzytyckiUrbanski1989,Rapaport2015}). In the remaining part of this section, we will present some applications of Theorem \ref{thm-1.0} along the line of this development. The proof of Theorem \ref{thm-1.0} is based on some ideas from the work of Ledrappier and Young \cite{LedrappierYoung1985}. It also adopts and extends some ideas used in \cite{FengHu2009, BaranyKaenmaki2015, QianXie2008} for the construction of measurable partitions and the density estimates of associated conditional measures. Since our construction of measurable partitions is much different from these works (see Remark \ref{rem-tt}), and the IFSs in consideration may be non-invertible and non-contractive, many estimates of conditional measures need to be rebuilt or re-justified. A key part of our arguments is on the estimation of the so-called ``transverse dimension'' of these conditional measures, where significant efforts are made to handle the situation when the linear parts of the IFS do not satisfy any domination condition (in such case the angles of Oseledets subspaces maybe substantially small). Our strategy is to build an induced dynamics so that we are able to focus on the trajectories where the angles of Oseledets subspaces are significantly large. \subsection{Dimension formulas} \label{S-DF} Throughout this subsection, under the assumptions of Theorem \ref{thm-1.0}, we further assume that $m$ is ergodic. We are going to present certain dimension formulas for $\mu=m\circ \pi^{-1}$ and related conditional measures. First notice that in this ergodic case, the condition (3) in Definition~\ref{de-1.1} is equivalent to \begin{equation} \label{e-1.3} \lambda:=\lim_{n\to \infty}\frac{1}{n}\int \log \|M_{x_0}\cdots M_{x_{n-1}}\|\; d m(x)<0. \end{equation} By Oseledets' multiplicative ergodic theorem \cite{Oseledets1968}, there exist an integer $1\leq s\leq d$, numbers $\lambda=\lambda_1>\cdots>\lambda_s\geq -\infty$, positive integers $k_1, \ldots, k_s$ with $\sum_{i=1}^sk_i=d$, and measurable linear subspaces $$ {\Bbb R}^d=V_x^0\supsetneq V_x^1\supsetneq\cdots\supsetneq V_x^s=\{0\}, \quad x\in \Sigma, $$ such that for $m$-a.e.~$x=(x_n)_{n=-\infty}^\infty$, \begin{itemize} \item[(i)] $M_{x_{-1}}V_x^i\subset V^i_{\sigma^{-1}x}$; \item[(ii)] $\dim V_x^i=\sum_{j=i+1}^s k_j$; \item[(iii)] $\lim_{n\to \infty} \frac{1}{n}\log \|M_{x_{-n}}\cdots M_{x_{-1}}v\|=\lambda_{i+1}$ for $v\in V_{x}^i\backslash V_{x}^{i+1}$. \end{itemize} The numbers $\lambda_1,\ldots, \lambda_s$ are called the {\em Lyapunov exponents} of $(M_j)_{j\in \Lambda}$ with respect to $m$, and $k_i$ the {\em multiplicity} of $\lambda_i$, $i=1,\ldots,s$. Recall that $\pi(x)$ is well-defined for $m$-a.e.~$x$. Hence there exists a Borel set $\Sigma'\subset\Sigma$ with $\sigma(\Sigma')=\Sigma'$ and $m(\Sigma')=1$ such that $\pi$ is well-defined on $\Sigma'$ and the above properties (i)-(iii) hold for $x\in \Sigma'$. We remark that these linear subspaces $V_x^i$ only depend on $i$ and $x^{-}:=(x_j)_{j=-\infty}^{-1}$ since by (i)-(iii), one has $$ V_{x}^i=\left\{v\in {\Bbb R}^d: \; \lim_{n\to \infty} \frac{1}{n}\log \|M_{x_{-n}}\cdots M_{x_{-1}}v\|\leq \lambda_{i+1}\right\}. $$ Using this property, we construct a family of measurable partitions $\xi_0,\ldots, \xi_s$ of $\Sigma'$ as follows: $$ \xi_i(x):=\{y\in \Sigma':\; y^{-}=x^{-}, \; \pi y-\pi x\in V_x^i\}, $$ here $\xi_i(x)$ is the $\xi_i$-atom that contains $x$ (see Sections~\ref{Sect-con} and \ref{S-4} for the details). Moreover, let \begin{equation} \label{e-pp} {\mathcal P}=\{[j]_0\cap \Sigma':\; j\in \Lambda\} \end{equation} be the canonical partition of $\Sigma'$, where $[j]_0:=\{x=(x_n)_{n=-\infty}^\infty\in \Sigma:\; x_0=j\}$. Define \begin{equation} \label{e-hi1} h_i=H_m({\mathcal P}|\widehat{\xi_i}),\quad i=0,\ldots, s, \end{equation} where $H_m(\cdot|\cdot)$ stands for the conditional entropy and $\widehat{\xi_i}$ is the $\sigma$-algebra generated by $\xi_i$ (see Section~\ref{Sect-ent} for the definitions). Now we are ready to present the dimension formula for $m\circ \pi^{-1}$. \begin{thm} \label{thm-1.1} Let ${\mathcal S}=\{M_jx+a_j\}_{j\in \Lambda}$ be an affine IFS on ${\Bbb R}^d$ and $m$ be an ergodic $\sigma$-invariant measure on $\Sigma$. Suppose that ${\mathcal S}$ is average contracting with respect to $m$. Let $\mu=m\circ \pi^{-1}$. Then \begin{equation} \label{ly-dim} \dim_{\rm H}\mu=\sum_{i=0}^{s-1}\frac{h_{i+1}-h_i}{\lambda_{i+1}}, \end{equation} where $h_i$ are defined as in \eqref{e-hi1}. \end{thm} Next we give similar dimension formulas for certain conditional measures associated with $m$. For $i=0,\ldots, s$, let $\{m_x^{\xi_i}\}$ be the system of conditional measures of $m$ associated with the partition $\xi_i$ (cf.~Section~\ref{Sect-con}). For a linear subspace $W$ of ${\Bbb R}^d$, let $W^\perp$ denote the orthogonal complement of $W$ in ${\Bbb R}^d$, and let $P_W:\; {\Bbb R}^d\to W$ denote the orthogonal projection from ${\Bbb R}^d$ to $W$. \begin{thm} \label{thm-1.2} Under the assumptions of Theorem \ref{thm-1.1}, for any $0\leq i<j\leq s$ and $m$-a.e.~$x\in \Sigma'$, the measures $m_x^{\xi_i}\circ \pi^{-1}$, $m_x^{\xi_i}\circ \pi^{-1}\circ (P_{{(V_x^j)}^\perp})^{-1}$ are exact dimensional with \begin{align} &\dim_{\rm H}(m_x^{\xi_i}\circ \pi^{-1})=\sum_{\ell=i}^{s-1}\frac{h_{\ell+1}-h_\ell}{\lambda_{\ell+1}}, \label{e-con-dim}\\ &\dim_{\rm H}\left(m_x^{\xi_i}\circ \pi^{-1}\circ (P_{{(V_x^j)}^\perp})^{-1}\right)=\sum_{\ell=i}^{j-1}\frac{h_{\ell+1}-h_{\ell}}{\lambda_{\ell+1}}, \label{e-con-proj-dim} \end{align} Moreover, for $m$-a.e.~$x\in \Sigma'$ and any $1\leq j\leq s$, \begin{equation}\label{e-proj-dim} \dim_{\rm loc}\left(m\circ \pi^{-1}\circ (P_{{(V_x^j)}^\perp})^{-1}, P_{{(V_x^j)}^\perp}( \pi x)\right)=\sum_{\ell=0}^{j-1}\frac{h_{\ell+1}-h_{\ell}}{\lambda_{\ell+1}}. \end{equation} \end{thm} From the above theorem, we can deduce certain dimension conservation property for the measures $m^{\xi_0}_x\circ \pi^{-1}$ and $\mu$. To state the result, let $G(d,k)$ denote the Grassmannian manifold of $k$-dimensional linear subspaces of ${\Bbb R}^d$. For a Borel probability measure $\eta$ on ${\Bbb R}^d$ and $W\in G(d,k)$, let $\{\eta_{W,z}=\eta^{\zeta_W}_z\}_{z\in {\Bbb R}^d}$ denote the the system of conditional measures of $\eta$ associated with the measurable partition $\zeta_W$ given by $$ \zeta_W=\{W+a:\; a\in W^\perp\}. $$ These conditional measures are also called the {\it slicing measures} of $\eta$ along the subspace $W$ (cf. \cite[\S10.1]{Mattila1995}). Following Furstenberg \cite{Furstenberg2008}, we give the following. \begin{de} $\eta$ is said to be dimension conserving with respect to the projection $P_{W^\perp}$, if $$ \dim_{\rm H}\eta =\dim_{\rm H} \eta_{W, z}+\dim_{\rm H}\left( \eta\circ (P_{W^\perp})^{-1}\right) $$ for $\eta$-a.e.~$z\in {\Bbb R}^d$. \end{de} For $i\in \{0, \ldots, s-1\}$, define $\Pi_i:\Sigma'\to G(d, \sum_{j=i+1}^sk_j)$ by \begin{equation} \label{e-Pi} \Pi_i(x)= V_x^i. \end{equation} The push-forward measures $m\circ(\Pi_i)^{-1}$, $i=1,\ldots, s-1$, are called the {\em Furstenberg measures} or {\em Furstenberg-Oseledets measures} associated with $(M_j)_{j\in \Lambda}$ and $m$. An ergodic measure $\nu\in {\mathcal M}_\sigma(\Sigma)$ is said to be {\it quasi-Bernoulli} if there exists a positive constant $C$ such that $$C^{-1}\nu([I]_0)\nu([J]_0)\leq \nu([IJ]_0)\leq C\nu([I]_0) \nu([J]_0)$$ for any finite words $I$, $J$ over $\Lambda$, where $$[I]_0:=\{x\in \Sigma:\; x_j=i_j \mbox{ for }0\leq j\leq n-1\}$$ for $I=i_0\ldots i_{n-1}$. Similarly, we say that $\nu$ is {\it sub-multiplicative} if there exists a positive constant $C$ such that $\nu([IJ]_0)\leq C\nu([I]_0) \nu([J]_0)$ for any finite words $I$, $J$ over $\Lambda$. \begin{thm} \label{cor-1.0} Under the assumptions of Theorem \ref{thm-1.1}, we further assume that $s\geq 2$. Let $i\in \{1,\ldots, s-1\}$. Then the following statements hold. \begin{itemize} \item[(i)] For $m$-a.e.~$x\in \Sigma'$, $m_x^{\xi_0}\circ \pi^{-1}$ is dimension conserving with respect to $P_{{(V_x^i)}^\perp}$ and moreover, the projected measure $(m_x^{\xi_0}\circ \pi^{-1})\circ P_{{(V_x^i)}^\perp}^{-1}$ is exact dimensional, and so are the slicing measures $(m_x^{\xi_0}\circ \pi^{-1})_{V_x^i, y}$ for $(m_x^{\xi_0}\circ \pi^{-1})$-a.e.~$y$. \item[(ii)] Assume that $m$ is quasi-Bernoulli. Then for $m\circ (\Pi_i)^{-1}$-a.e.~$W$, $\mu$ is dimension conserving with respect to $P_{W^\perp}$, and moreover, the associated projected measure and almost all slicing measures are exact dimensional. \item[(iii)] Assume that $m$ is sub-multiplicative. Then for $m\circ (\Pi_i)^{-1}$-a.e.~$W$, there exists a subset $A_W$ of~${\Bbb R}^d$ with $\mu(A_W)>0$ such that for every $z\in A_W$, \begin{equation*} \begin{split} \dim_{\rm loc} (\mu_{W,z}, z)&=\sum_{\ell=i}^{s-1}\frac{h_{\ell+1}-h_\ell}{\lambda_{\ell+1}},\\ \dim_{\rm loc} (\mu\circ P_{W^\perp}^{-1}, P_{W^\perp}(z))&=\sum_{\ell=0}^{i-1}\frac{h_{\ell+1}-h_\ell}{\lambda_{\ell+1}}, \end{split} \end{equation*} and so, $\dim_{\rm H}\mu=\dim_{\rm loc} (\mu_{W,z}, z)+\dim_{\rm loc} (\mu\circ P_{W^\perp}^{-1}, P_{W^\perp}(z)).$ When $m$ is quasi-Bernoulli then one can take the set $A_W$ such that $\mu(A_W)=1$. \end{itemize} \end{thm} We remark that part (ii) of Theorem \ref{cor-1.0} was previously proved in \cite{BaranyKaenmaki2015} under the assumptions that ${\mathcal S}$ is contracting,\ invertible and its linear parts satisfy certain domination condition. According to part (iii) of the theorem, when $m$ is sub-multiplicative, $\mu$ partially satisfies dimension conservation. It is unknown whether part (ii) always holds when $m$ is only assumed to be ergodic. However, as is illustrated in the following theorem, this is true in the special case that the linear parts of the IFS commute. \begin{thm} \label{thm-1.7'} Let ${\mathcal S}=\{M_jx+a_j\}_{j\in \Lambda}$ be an affine IFS on ${\Bbb R}^d$, average contracting with respect to an ergodic $m\in {\mathcal M}_\sigma(\Sigma)$. Let $\mu=m\circ \pi^{-1}$. Assume $s\geq 2$ and in addition that $M_jM_{j'}=M_{j'}M_j$ for $j,j'\in \Lambda$. Then for $i\in \{1,\ldots, s-1\}$, $V_x^i$ is constant $m$-a.e., denoted by $W_i$, moreover, $\mu$ is dimension conserving with respect to $P_{(W_i)^\perp}$. \end{thm} It is worth pointing out that if $\mu$ is a contracting self-similar measure in ${\Bbb R}^d$ with a finite rotation group, then for each proper subspace $W$ of ${\Bbb R}^d$, $\mu$ is dimension conserving with respect to $P_W$. The result is due to Falconer and Jin \cite{FalconerJin2014}. Under an additional assumption of the strong separation condition, this result can be alternatively derived from a general result of Furstenberg (cf.~\cite[Theorem 3.1]{Furstenberg2008}). We remark that this dimension conservation property also extends to ergodic invariant measures for rotation-free self-similar IFSs (see Remark \ref{rem-6.3}). Furthermore, we remark that Theorem \ref{cor-1.0}-\ref{thm-1.7'} can be applied to analyze slices and projections of certain self-affine sets (see Remark \ref{rem-6.4}). \subsection{Semi-continuity of dimension and applications} \label{S-SC} Here we present a semi-continuity result on the dimension of ergodic invariant measures for affine IFSs and give its application to the dimension of self-affine sets. Again let $\mathcal S=\{M_ix+a_i\}_{i\in \Lambda}$ be an affine IFS on ${\Bbb R}^d$, average contracting with respect to an ergodic invariant measure $m$ on $\Sigma$. Write ${\bf a}=(a_i)_{i\in \Lambda}$. To emphasize the dependence on ${\bf a}$, let $\pi_{\bf a}$ be the coding map associated to $\mathcal S$ and let $h_{i,{\bf a}}$ ($i=1,\ldots, s$) be the conditional entropies of $m$ defined in \eqref{e-hi1}. Then we have the following. \begin{thm} \label{thm-1.3} \begin{itemize} \item[(1)] The mapping ${\bf a}\mapsto h_{i,{\bf a}}$ is upper semi-continuous for each $i\in \{1,\ldots, s\}$. \item[(2)] Moreover, the mapping ${\bf a}\mapsto \dim_{\rm H} \left(m\circ (\pi_{{\bf a}})^{-1}\right)$ is lower semi-continuous. \end{itemize} \end{thm} Part (1) of the above result was first proved by Rapaport \cite[Lemma 8]{Rapaport2015} in the case when $m$ is a Bernoulli product measure and ${\mathcal S}$ is invertible and contracting. Part (2) was shown by Hochman and Shmerkin \cite{HochmanShmerkin2012} for a special class of self-similar measures on ${\Bbb R}$. In Remark \ref{rem-8.1} we give a further extension of Theorem \ref{thm-1.3}. Below we present an application of Theorem \ref{thm-1.3} to the dimension of self-affine sets and associated stationary measures. For this purpose, in the remaining part of this subsection we assume that $\|M_j\|<1$ for $j\in \Lambda$ and write ${\bf M}=(M_j)_{j\in \Lambda}$. Let $K({\bf M},{\bf a})$ be the self-affine set generated by the IFS ${\mathcal S}=\{M_jx+a_j\}_{j\in \Lambda}$. In 1988, Falconer \cite{Falconer1988} introduce a quantity associated to ${\bf M}$, nowadays usually called the {\em affinity dimension} $\dim_{\rm AFF}({\bf M})$, which is always an upper bound for the upper box-counting dimension of $K({\bf M},{\bf a})$, and such that when $\|M_j\|<1/2$ for all $j$, then for ${\mathcal L}^{d|\Lambda|}$-a.e.~${\bf a}$, $\dim_{\rm H}K({\bf M},{\bf a})=\min (d, \dim_{\rm AFF}({\bf M}))$. In fact, Falconer proved this with $1/3$ as the upper bound on the norms; it was subsequently shown by Solomyak \cite{Solomyak1998} that $1/2$ suffices. The analogue of affinity dimension for measures is the Lyapunov dimension, which we denote $\dim_{\rm LY}(m, {\bf M})$; see Section~\ref{S-7} for its definition. In \cite{JordanPollicottSimon2007}, Jordan, Pollicott and Simon proved that the Lyapunov dimension $\dim_{\rm LY}(m, {\bf M})$ is always an upper bound for the Hausdorff dimension of $m\circ (\pi_{{\bf a}})^{-1}$, and moreover when $\|M_j\|<1/2$ for all $j$, then for ${\mathcal L}^{d|\Lambda|}$-a.e.~${\bf a}$, $\dim_{\rm H} m\circ (\pi_{{\bf a}})^{-1}=\min(d, \dim_{\rm LY}(m,{\bf M}))$. Recall a set in a topological space is said to be of {\em first category} if it can be written as the countable union of nowhere dense subsets. As an application of Theorem \ref{thm-1.3}, we get the following result. \begin{thm} \label{cor-1.7} Suppose that $\|M_j\|<1/2$ for $j\in \Lambda$. Then the following hold. \begin{itemize} \item[(i)] For every ergodic $\sigma$-invariant measure $m$ on $\Sigma$, the exceptional set $$ \left\{{\bf a}\in {\Bbb R}^{d|\Lambda|}:\; \dim_{\rm H} \left(m\circ (\pi_{{\bf a}})^{-1}\right)\neq \min(d, \dim_{\rm LY}(m, {\bf M}))\right\} $$ is of first category in ${\Bbb R}^{d|\Lambda|}$. \item[(ii)] The exceptional set $$ \left\{{\bf a}\in {\Bbb R}^{d|\Lambda|}:\; \dim_{\rm H} K({\bf M}, {\bf a})\neq \min(d, \dim_{\rm AFF}({\bf M}))\right\} $$ is of first category in ${\Bbb R}^{d|\Lambda|}$. \end{itemize} \end{thm} The above result says that these exceptional sets are also small in a topological sense. A fundamental and challenging question is to specify those translation vectors not lying in the exception sets. Significant progresses have been made recently in \cite{Barany2015,FalconerKempton2018,BaranyKaenmaki2015, Rapaport2015}, showing that under certain additional assumptions, the Hausdorff and Lyapunov dimensions of a self-affine measure (or more generally, the push-forward of a quasi-Bernoulli measure) coincide if the involved Furstenberg measures have enough large dimension. In next theorem we will drop off some redundant assumptions used in these works and further extend the result to the push-forward measures of ergodic sub-multiplicative measures. Recall that for a Borel probability measure $\eta$ on a metric space, its {\it upper Hausdorff dimension} $\dim_{\rm H}^*\eta$ is the smallest Hausdorff dimension of a Borel set $F$ of $\eta$ measure $1$. Set $d_0=0$ and $d_\ell=k_1+\cdots+k_\ell$ for $1\leq \ell\leq s$. \begin{thm} \label{thm-1.4} Let $\mathcal S=\{M_jx+a_j\}_{j\in \Lambda}$ be a contracting affine IFS on ${\Bbb R}^d$ satisfying the strong separation condition and $m\in \mathcal M_\sigma(\Sigma)$ be ergodic and sub-multiplicative. Let $i$ be the unique element in $\{1,\ldots, s\}$ so that $d_{i-1} \leq \dim_{\rm LY} (m, {\bf M})<d_i$. Then \begin{equation} \label{e-tz} \dim_{\rm H} m\circ \pi^{-1}=\dim_{\rm LY}(m, {\bf M}) \end{equation} provided one of the following conditions holds: \begin{itemize} \item[(a)] $s=1$. \item[(b)] $i=s>1$, $\lambda_s\neq -\infty$ and \begin{equation} \label{e-condLY} \dim_{\rm H}^*\left(m\circ (\Pi_{s-1})^{-1}\right) + \dim_{\rm LY} (m, {\bf M})\geq d_{s-1} (d-d_{s-1}+1). \end{equation} \item[(c)] $1\leq i \leq s-1$, and \begin{align} &\dim_{\rm H}^*\left(m\circ (\Pi_{i})^{-1}\right)-\dim_{\rm LY}(m, {\bf M})\geq d_{i}(d-d_i-1), \label{e-al2}\\ &\dim_{\rm H}^*\left(m\circ (\Pi_{i-1})^{-1}\right) + \dim_{\rm H} m\circ \pi^{-1} \geq d_{i-1}(d-d_{i-1}+1). \label{e-al1} \end{align} \end{itemize} \end{thm} The conditions (b), (c) in the above theorem were introduced in \cite{Rapaport2015} and \cite{BaranyHochmanRapaport2017}, respectively, in slightly stronger forms. For a contracting invertible affine IFS, Rapaport \cite{Rapaport2015} proved the implication (b)$\Rightarrow$ \eqref{e-tz} in the case when $m$ is Bernoulli and $(M_j)_{j\in \Lambda}$ satisfies an irreducibility assumption; whilst B\'{a}r\'{a}ny and K\"aenm\"aki \cite{BaranyHochmanRapaport2017} proved \eqref{e-tz} under the assumptions that the conditions \eqref{e-al2}-\eqref{e-al1} hold for all $i\in \{1,\ldots, s-1\}$, $m$ is Bernoulli and $d=2$, or $m$ is quasi-Bernoulli and $\{M_j\}_{j\in \Lambda}$ satisfies a domination condition. We remark that \eqref{e-al1} always holds whenever $i=1$, since $d_0=0$. It is worth pointing out that for every affine IFS $\mathcal S=\{M_jx+a_j\}_{j\in \Lambda}$ on ${\Bbb R}^d$, there exists at least one ergodic $m\in \mathcal M_\sigma(\Sigma)$, called {\em K\"aenm\"aki measure}, so that $\dim_{\rm LY}(m, {\bf M})=\dim_{\rm AFF}({\bf M})$. This was first proved by K\"aenm\"aki \cite{Kaenmaki2004} in the case when $\mathcal S$ is invertible, and it extends to the general case by the sub-additive thermodynamic formalism \cite{CaoFengHuang2008}. Very recently, Bochi and Morris \cite{BochiMorris2018} showed that whenever $\mathcal S$ is invertible, each K\"aenm\"aki measure is sub-multiplicative. Hence for an invertible ${\mathcal S}$ satisfying the strong separation condition, if one of the conditions (a)-(c) in Theorem \ref{thm-1.4} fulfills for some K\"aenm\"aki measure $m$, then $\dim_{\rm H} K({\bf M}, {\bf a})=\dim_{\rm AFF}({\bf M})=\dim_{\rm H}m\circ \pi^{-1}$. To check the conditions (b)-(c) in Theorem \ref{thm-1.4}, one needs to estimate the (upper) Hausdorff dimension of Furstenberg measures $m\circ (\Pi_{i})^{-1}$. So far there have been only a few dimensional results on these measures. In the case $d=2$, Hochman and Solomyak \cite{HochmanSolomyak2017} calculated the Hausdorff dimension of Furstenberg measures for Bernoulli $m$ under some mild assumptions. In \cite[Sect.~2.4]{BRS2018}, B\'ar\'any, Rams and Simon determined the Hausdorff dimension of Furstenberg measures for some special triangular affine IFSs in ${\Bbb R}^d$, in which $m$ could be any ergodic measure. We remark that the conditions of Theorem \ref{thm-1.4} might not be sharp. Very recently, B\'ar\'any, Hochman and Rapaport \cite{BaranyHochmanRapaport2017} made a significant progress in dimension theory of affine IFSs, showing that the Hausdorff and affinity dimensions of a planar self-affine set coincide under the strong separation condition and certain irreducibility assumption; and similarly, the Hausdorff and Lyapunov dimensions of a planar self-affine measure coincide under the same assumptions. \subsection{Organization of the paper} The paper is organized as follows. In Section~\ref{S-2}, we provide some density results about conditional measures, and present a version of Oseledets's multiplicative ergodic theorem due to Froyland et al.~\cite{FroylandLloydQuas2010}. In Section~\ref{S-3}, we give some auxiliary results on the coding maps for average contracting affine IFSs. In Section~\ref{S-4}, we construct a finite family of measurable partitions of $\Sigma$ for a given average contracting affine IFS and give some necessary properties. In Section~\ref{S-5} we prove an inequality for the transverse dimensions of the conditional measures associated with these measurable partitions. In Section~\ref{S-6}, we prove Theorems \ref{thm-1.0}-\ref{thm-1.2}, \ref{cor-1.0}-\ref{thm-1.7'}. In Section~\ref{S-7}, we give some properties of Lyapunov dimension. In Section~\ref{S-8}, we prove Theorems \ref{thm-1.3}-\ref{thm-1.4}. \section{Preliminaries} \label{S-2} In this section, we first give the definitions and some properties of conditional entropies and conditional measures. Then we collect some known facts about induced transformations and derive a useful result (Lemma \ref{lem-inderg}). In the end, we present a version of Oseledets' multiplicative ergodic theorem due to Froyland et al.~\cite{FroylandLloydQuas2010}. \subsection{Conditional information and entropy} \label{Sect-ent} Let $(X,{\mathcal B},m)$ be a probability space. For a sub-$\sigma$-algebra ${\mathcal A}$ of ${\mathcal B}$ and $f\in L^1(X,{\mathcal B},m)$, we denote by ${\bf E}_m(f|{\mathcal A})$ the the {\it conditional expectation of $f$ given ${\mathcal A}$}. For a countable ${\mathcal B}$-measurable partition $\xi$ of $X$, we denote by ${\bf I}_m(\xi|{\mathcal A})$ the {\it conditional information of $\xi$ given ${\mathcal A}$}, which is given by the formula \begin{equation} \label{e-1.2} {\bf I}_m(\xi|{\mathcal A})=-\sum_{A\in \xi}\chi_A\log {\bf E}_m(\chi_A |{\mathcal A}), \end{equation} where $\chi_A$ is the characteristic function on $A$. The {\it conditional entropy of $\xi$ given ${\mathcal A}$}, written as $H_m(\xi|{\mathcal A})$, is defined by the formula \begin{equation*} H_m(\xi|{\mathcal A})=\int {\bf I}_m(\xi|{\mathcal A})\; dm. \end{equation*} (See e.g. \cite{Parry1981, Walters1982} for more details.) The above information and entropy are unconditional when ${\mathcal A}={{\mathcal N}}$, the trivial $\sigma$-algebra consisting of sets of measure zero and one, and in this case we write \begin{equation*} {\bf I}_m(\xi|{{\mathcal N}})=:{\bf I}_\nu(\xi)\quad\mbox{and}\quad H_m(\xi|{{\mathcal N}})=:H_m(\xi). \end{equation*} For a countable partition $\xi$, we use $\widehat{\xi}$ to denote the $\sigma$-algebra generated by $\xi$. If $\xi_1$,\ldots, $\xi_n$ are countable partitions, then $\xi_1\vee \cdots \vee \xi_n=\bigvee_{i=1}^n \xi_i$ denotes the partition consists of the sets $A_1\cap\cdots \cap A_n$ with $A_i\in \xi_i$. Similarly for $\sigma$-algebras ${\mathcal A}_1$, ${\mathcal A}_2$,\dots, ${\mathcal A}_1\vee {\mathcal A}_2\vee \cdots$ or $\bigvee_{i} {\mathcal A}_i$ denotes the $\sigma$-algebra generated by $\bigcup_i {\mathcal A}_i$. In the following lemma, we list some basic properties of the (conditional) information and entropy. \begin{lem}[cf. \cite{Parry1981}] \label{lem-par} Let $T$ be a measure-preserving transformation of a separable probability space $(X,{\mathcal B},m)$. Let $\xi,\eta$ be two countable Borel partitions of $X$ with $H_m(\xi)<\infty$, $H_m(\eta)<\infty$, and ${\mathcal A}$ a sub-$\sigma$-algebra of ${\mathcal B}$. Then we have \begin{itemize} \item[(i)] ${\bf E}_m(f|{\mathcal A})\circ T={\bf E}_m(f\circ T|T^{-1}{\mathcal A})$ for $f\in L^1(X, {\mathcal B}, m)$. \item[(ii)] ${\bf I}_{m}(\xi|{\mathcal A})\circ T={\bf I}_m(T^{-1}\xi|T^{-1}{\mathcal A})$. \item[(iii)] ${\bf I}_m(\xi\vee \eta|{\mathcal A})={\bf I}_m(\xi|{\mathcal A})+{\bf I}_m(\eta|\widehat{\xi}\vee {\mathcal A})$. \item[(iv)] $H_m(\xi\vee \eta|{\mathcal A})=H_m(\xi|{\mathcal A})+H_m(\eta|\widehat{\xi}\vee {\mathcal A})$. \item[(v)] If ${\mathcal A}_1\subset {\mathcal A}_2\subset\cdots $ is an increasing sequence of sub-$\sigma$-algebras with ${\mathcal A}_n\uparrow {\mathcal A}$, then $\sup_n{\bf I}_m(\xi|{\mathcal A}_n)\in L^1$, and ${\bf I}_m(\xi|{\mathcal A}_n)$ converges almost everywhere and in $L^1$ to ${\bf I}_m(\xi|{\mathcal A})$. In particular, $\lim_{n\to \infty} H_m(\xi|{\mathcal A}_n)=H_m(\xi|{\mathcal A})$. \end{itemize} \end{lem} Below we present an additional property of the conditional expectation. \begin{lem}[{\cite[Lemma 3.10]{FengHu2009}}] \label{lem-2.10} Let $(X,{\mathcal B},m)$ be a probability space and ${\mathcal A}$ a sub-$\sigma$-algebra of ${\mathcal B}$. Let $A\in {\mathcal B}$ with $m(A)>0$. Then $$ {\bf E}_m(\chi_A|{\mathcal A})(x)>0 $$ for $m$-a.e.~$x\in A$. \end{lem} The following lemma is a variant of Maker's ergodic theorem (\cite{Maker1940}). \begin{lem}[\cite{Mane1987}, Corollary 1.6, p. 96] \label{lem-3.15} Let $T$ be a measure-preserving transformation of a probability space $(X,{\mathcal B},m)$. Let $g_k\in L^1(X, {\mathcal B}, m)$ be a sequence that converges almost everywhere and in $L^1$ to $g\in L^1(X, {\mathcal B}, m)$. Then $$ \lim_{k\to +\infty}\frac{1}{k}\sum_{j=0}^{k-1}g_{k-j}(T^jx)={\bf E}_m(g|{\mathcal I})(x) $$ almost everywhere and in $L^1$. \end{lem} \subsection{Conditional measures} \label{Sect-con} Here we give a brief introduction to Rohlin's theory of Lebesgue spaces, measurable partitions and conditional measures. The reader is referred to \cite{Rohlin1949, Parry1969, EinsiedlerWard2011} for more details. A probability space $(X, {\mathcal B}, m)$ is called a {\it Lebesgue space} if it is isomorphic to a probability space which is the union of $[0,s]$ ($0\leq s\leq 1$) with Lebesgue measure and a finite or countable number of atoms. Now let $(X, {\mathcal B},m)$ be a Lebesgue space. A {\it measurable partition} $\eta$ of $X$ is a partition of $X$ such that, up to a set of measure zero, the quotient space $X/\eta$ is separated by a countable number of measurable sets $\{B_i\}$. The quotient space $X/\eta$ with its inherited probability space structure, written as $(X_\eta, {\mathcal B}_\eta, m_\eta)$, is again a Lebesgue space. Also, any measurable partition $\eta$ determines a sub-$\sigma$-algebra of ${\mathcal B}$, denoted by $\widehat{\eta}$, whose elements are unions of elements of $\eta$. Conversely, any sub-$\sigma$-algebra ${\mathcal B}'$ of ${\mathcal B}$ is also countably generated, say by $\{B_i'\}$, and therefore all the sets of the form $\cap A_i$, where $A_i=B_i^\prime$ or its complement, form a measurable partition. In particular, ${\mathcal B}$ itself is corresponding to a partition into single points. An important property of Lebesgue spaces and measurable partitions is the following. \begin{thm}[Rohlin \cite{Rohlin1949}] \label{thm-2.1} Let $\eta$ be a measurable partition of a Lebesgue space $(X, {\mathcal B}, m)$. Then, for every $x$ in a set of full $m$-measure, there is a probability measure $m^\eta_x$ defined on $\eta(x)$, the element of $\eta$ containing $x$. These measures are uniquely characterized (up to sets of $m$-measure $0$) by the following properties: if $A\subset X$ is a measurable set, then $x\mapsto m^\eta_x(A)$ is $\widehat{\eta}$-measurable and $m(A)=\int m^\eta_x(A)d m(x)$. These properties imply that for any $f\in L^1(X,{\mathcal B}, m)$, $m_x^\eta(f)={\bf E}_m(f|\widehat{\eta})(x)$ for $m$-a.e.~$x$, and $m(f)=\int {\bf E}_m(f|\widehat{\eta})dm$. \end{thm} The family of measures $\{m^\eta_x\}$ in the above theorem is called the {\it canonical system of conditional measures associated with $\eta$}. Throughout the remaining part of this subsection, we assume that $(X,{\mathcal B},m)$ is a Lebesgue space. Suppose that $Y$ is a complete separable metric space and $\pi: X\to Y$ is a ${\mathcal B}$-measurable map. Denote $\gamma:={\mathcal B}(Y)$, the Borel-$\sigma$-algebra on $Y$. According to Rohlin's theory (cf.~\cite[Section~2.5]{Rohlin1949}, \cite[Chapter IV]{Parry1969}), the mapping $\pi$ induces a measurable partition \begin{equation} \label{e-xi1} \xi=\{\pi^{-1}(y):\; y\in Y\} \end{equation} of $X$ with $\widehat{\xi}=\pi^{-1}\gamma\mbox{ (mod 0)}$, and $(X_\xi, {\mathcal B}_\xi, m_\xi)$ is isomorphic $(\mbox{mod 0})$ to $(Y, \gamma, m\circ \pi^{-1})$. The system of conditional measures $\{m^\xi_x\}$ is also called the {\it disintegration of $m$ with respect to $\pi$}. For $y\in Y$, we use $B(y,r)$ to denote the closed ball in $Y$ of radius $r$ centered at $y$. Moreover we write for $x\in X$, \begin{equation} \label{e-ball} B^\pi(x,r)=\pi^{-1} B(\pi x,r). \end{equation} Furthermore, we say that $Y$ is a {\it Besicovitch space} if $Y$ is a complete separable metric space and the Besicovich covering lemma (see e.g.~\cite{Mattila1995}) holds in $Y$. Besicovich spaces include, for instance, Euclidean spaces, compact finite-dimensional Riemannian manifolds and complete separable ultrametric spaces. \begin{lem} \label{lem-2.4} Let $\pi: X\to Y$ be a measurable mapping from a Lebesgue space $(X, {\mathcal B}, m)$ to a Besicovitch space $Y$. Let $\eta$ be a measurable partition of $X$. Then the following properties hold. \begin{itemize} \item[(1)] Let $A\in {\mathcal B}$. Then for $m$-a.e.~$x\in X$, \begin{equation*} \lim_{r\to 0}\frac{ m_x^\eta(B^\pi(x,r)\cap A)}{ m_x^\eta(B^\pi(x,r))}={\bf E}_m(\chi_A|\hat{\eta}\vee \pi^{-1}\gamma)(x). \end{equation*} \item[(2)] Let $\alpha$ be a finite or countable measurable partition of $X$. Then for $m$-a.e.~$x\in X$, \begin{equation*} \lim_{r\to 0}\log \frac {m^\eta_x\left(B^\pi(x,r)\cap \alpha(x)\right)} {m^\eta_x\left(B^\pi(x,r)\right)} =-{\bf I}_m \left(\alpha|\hat{\eta}\vee\pi^{-1}\gamma\right)(x). \end{equation*} Furthermore, set \begin{equation*} g(x)=-\inf_{r>0}\log\frac {m^\eta_x\left(B^\pi(x,r)\cap \alpha(x)\right)} {m^\eta_x\left(B^\pi(x,r)\right)} \end{equation*} and assume $H_m(\alpha)<\infty$. Then $g\geq 0$ and $g\in L^1(X,{{\mathcal B}},m)$. \end{itemize} \end{lem} \begin{proof} These properties have been proved in \cite[Lemma 3.3, Proposition 3.5]{FengHu2009} in the case when $Y={\Bbb R}^d$. The proofs there remain valid for the general case when $Y$ is a Besicovitch space. \end{proof} \begin{rem}In the above lemma, we have ${\bf E}_m(\chi_A|\hat{\eta}\vee \pi^{-1}\gamma)={\bf E}_m(\chi_A|\hat{\eta}\vee \hat{\xi})$ and ${\bf I}_m \left(\alpha|\hat{\eta}\vee\pi^{-1}\gamma\right)={\bf I}_m \left(\alpha|\hat{\eta}\vee\hat{\xi}\right)$ $m$-a.e., where $\xi$ is given by \eqref{e-xi1}. This is because $\widehat{\xi}=\pi^{-1}\gamma\mbox{ (mod 0)}$. \end{rem} \begin{de} Two probability measures $m_1$ and $m_2$ on a measurable space $(X, {\mathcal B})$ are said to be strongly equivalent if there exists a positive constant $C$ such that $C^{-1}m_1(A)\leq m_2(A)\leq C m_1(A)$ for all $A\in {\mathcal B}$. \end{de} \begin{lem} \label{lem-2.8} Let $\pi: X\to Y$ be a measurable mapping from a Lebesgue space $(X, {\mathcal B}, m_1)$ to a Besicovitch space $Y$. Let $\xi$ be the measurable partition of $X$ given in \eqref{e-xi1}. Suppose $m_2$ is another probability measure on $(X,{\mathcal B})$ strongly equivalent to $m_1$. Then for $m_1$-a.e.~$x$, $(m_1)^\xi_x$ and $(m_2)^\xi_x$ are strongly equivalent. \end{lem} \begin{proof} Since $m_1$ and $m_2$ are strongly equivalent, there exists a positive constant $C$ such that $C^{-1} m_1(B)\leq m_2(B)\leq C m_1(B)$ for all $B\in {\mathcal B}$. Pick a countable subset ${\mathcal B}'$ of ${\mathcal B}$ such that $\sigma({\mathcal B}')={\mathcal B}$, where $\sigma({\mathcal B}')$ stands for the $\sigma$-algebra generated by ${\mathcal B}'$. Applying Lemma \ref{lem-2.4}(1) (in which we take $\eta={\mathcal N}$) to $m_1$ and $m_2$ yields that for $m_1$-a.e.~$x$, $$ (m_i)^\xi_x(A)= \lim_{r\to 0}\frac{ m_i(B^\pi(x,r)\cap A)}{ m_i(B^\pi(x,r))} \qquad (i\in \{1,2\},\; A\in {\mathcal B}'). $$ It implies that for $m_1$-a.e $x$, \begin{equation} \label{ee-1} C^{-2} (m_1)^\xi_x(A)\leq (m_2)^\xi_x(A)\leq C^2(m_1)^\xi_x(A) \end{equation} for all $A\in {\mathcal B}'$. Since $\sigma({\mathcal B}')={\mathcal B}$, for $m_1$-a.e.~$x$, \eqref{ee-1} holds for all $A\in {\mathcal B}$. This completes the proof of the lemma. \end{proof} \begin{lem} \label{lem-2.9} Let $\pi: X\to Y$ be a measurable mapping from a Lebesgue space $(X, {\mathcal B}, m)$ to a Besicovitch space $Y$. Let $\xi$ be the measurable partition of $X$ given in \eqref{e-xi1}. Suppose $A\in {\mathcal B}$ with $m(A)>0$ and let $m_A$ be the probability measure given by $m_A(E)=m(A\cap E)/m(A)$ for $E\in {\mathcal B}$. Then for $m$-a.e.~$x\in A$, $(m_A)^\xi_x=(m^\xi_x)_A$, that is, $$ (m_A)^\xi_x(E)=\frac{m^\xi_x(A\cap E)}{m_x^\xi(A)} \qquad \mbox{ for all } E\in {\mathcal B}. $$ \end{lem} \begin{proof} Pick a countable subset ${\mathcal B}'$ of ${\mathcal B}$ such that $\sigma({\mathcal B}')={\mathcal B}$. Applying Lemma \ref{lem-2.4}(1) (in which we take $\eta={\mathcal N}$) to $m_A$ and $m$ yields that for $m$-a.e.~$x\in A$, for all $E\in {\mathcal B}'$, \begin{align*} (m_A)^\xi_x(E) &= \lim_{r\to 0}\frac{ m_A(B^\pi(x,r)\cap E)}{ m_A(B^\pi(x,r))}\\ &= \lim_{r\to 0}\frac{ m(B^\pi(x,r)\cap E\cap A)}{ m(B^\pi(x,r)\cap A)}\\ &= \lim_{r\to 0}\frac{ m(B^\pi(x,r)\cap E\cap A)/ m(B^\pi(x,r))}{ m(B^\pi(x,r)\cap A)/ m(B^\pi(x,r))}\\ &=m^\xi_x(A\cap E)/m_x^\xi(A). \end{align*} Since $\sigma({\mathcal B}')={\mathcal B}$, for $m$-a.e.~$x\in A$ the equality $(m_A)^\xi_x(E)=m^\xi_x(A\cap E)/m_x^\xi(A)$ holds for all $E\in {\mathcal B}$. This completes the proof of the lemma.\end{proof} \subsection{Induced transformations} \label{Sect-2.3} Let $(X,{\mathcal B}, m, T)$ be an invertible measure-preserving system. Fix $N\in {\Bbb N}$ and $F\in {\mathcal B}$ with $m(F)>0$. By the Poincar\'{e} recurrence theorem, the {\it first return map to $F$ associated with $T^N$}, defined by $$r_F(x)=\inf\{n\geq 1:\; T^{Nn}(x)\in F\}, $$ exists almost everywhere. The map $T_F:\; F\to F$ defined almost everywhere by $$ T_F(x)=T^{Nr_F(x)}(x) $$ is called the {\it transformation induced by $T^N$ on the set $F$}. For $n\geq 1$, set $F_n=\{x\in F:\; r_F(x)=n\}$. Write $${\mathcal B}|_F:=\{B\cap F:\; B\in {\mathcal B}\}, \quad m_F:=\frac{1}{m(F)}m|_F,$$ where $m|_F$ stands for the restriction of $m$ on $F$, that is, $m|_F(B)=m(B\cap F)$ for $B\in {\mathcal B}$. The following result is well-known (see e.g. \cite[pp.~61-63]{EinsiedlerWard2011} and \cite[pp.~257-258]{Petersen1983} for a proof). \begin{lem} \label{lem-induce} \begin{itemize} \item[(i)] The induced transformation $T_F$ is a measure-preserving transformation on the space $(F, {\mathcal B}|_F, m_F)$. \item[(ii)] The family of sets $\{T^{Nj} F_n\}_{n\geq 1, \;0\leq j\leq n-1}$ are disjoint, and hence $$\sum_{n=1}^\infty n\;m(F_n)\leq 1.$$ \item[(iii)] $\displaystyle -\sum_{n=1}^\infty m(F_n)\log m(F_n)<\infty$. \end{itemize} \end{lem} Set ${\mathcal I}=\{B\in {\mathcal B}:\; T^{-1}(B)=B\}$ and ${\mathcal I}_F:=\{B\in {\mathcal B}|_F:\; (T_F)^{-1}(B)=B\}$. The following result will be needed in the proof of Theorem \ref{thm-1.0}. \begin{lem} \label{lem-inderg} Let $g\in L^1(X, {\mathcal B},m)$. Set $G(x)=\sum_{j=0}^{Nr_F(x)-1} g(T^jx)$ for $x\in F$. Then $G\in L^1(F, {\mathcal B}|_F, m_F)$. Moreover, \begin{equation} \label{e-identity} N {\bf E}_m(g|{\mathcal I})(x)=\frac{{\bf E}_{m_F}(G|{\mathcal I}_F)(x)}{{\bf E}_{m_F}(r_F|{\mathcal I}_F)(x)} \end{equation} for $m$-a.e.~$x\in F$. \end{lem} \begin{proof} First notice that \begin{equation*} \begin{split} \int_F |G| \; dm_F & = \frac{1}{m(F)} \sum_{n=1}^\infty \int_{F_n} |G|\;dm\\ & \leq \frac{1}{m(F)}\sum_{n=1}^\infty\sum_{p=0}^{Nn-1} \int_{F_n} |g\circ T^p|\; dm\\ &= \frac{1}{m(F)} \sum_{n=1}^\infty\sum_{p=0}^{Nn-1} \int_{T^{p} F_n} |g|\; dm \quad \mbox{(since $T$ is invertible and preserves $m$)} \\ &= \frac{1}{m(F)} \sum_{n=1}^\infty\sum_{k=0}^{N-1}\sum_{j=0}^{n-1} \int_{T^{Nj+k} F_n} |g|\; dm\\ &= \frac{1}{m(F)} \sum_{k=0}^{N-1}\sum_{n=1}^\infty\sum_{j=0}^{n-1} \int_{T^{Nj+k} F_n} |g|\; dm\\ &\leq \frac{N}{m(F)}\int_{X} |g|\; dm, \end{split} \end{equation*} where in the last inequality we have used the fact that for any $k$, the sets in the collection $\{T^{Nj+k} F_n:\; n\in {\Bbb N}, 0\leq j\leq n-1\}$ are disjoint (see Lemma \ref{lem-induce}(ii)). Hence $G\in L^1(m_F)$. Below we prove \eqref{e-identity}. Consider the sequence of integer-valued functions $(n_k(x))_{k=0}^\infty$, which are defined on $F$ almost everywhere by $n_0(x)=0$, and $$ n_k(x)=\sum_{j=0}^{k-1}r_F(T_F^jx)\quad \mbox{ for } k\geq 1, $$ where $T_F^j:=(T_F)^j$. Clearly, $n_k(x)\geq k$ and $T_F^k(x)=T^{N n_k(x)}(x)$. Hence, \begin{equation*} \begin{split} \sum_{j=0}^{k-1}G(T_F^jx)&=\sum_{j=0}^{k-1}\sum_{p=0}^{Nr_F(T_F^jx)-1} g(T^p(T_F^jx))\\ &=\sum_{j=0}^{k-1}\sum_{p=0}^{Nr_F(T^j_Fx)-1} g(T^{Nn_j(x)+p}x)\\ &=\sum_{j=0}^{k-1}\sum_{\ell=Nn_j(x)}^{Nn_{j+1}(x)-1} g(T^\ell x)\\ &=\sum_{i=0}^{Nn_k(x)-1} g(T^ix). \end{split} \end{equation*} By the Birkhoff ergodic theorem, we have \begin{equation} \label{e-Gg} \begin{split} \lim_{k\to +\infty}\frac{1}{n_k(x)}\sum_{j=0}^{k-1}G(T_F^jx) &=\lim_{k\to +\infty}\frac{1}{n_k(x)}\sum_{i=0}^{Nn_k(x)-1} g(T^ix)\\ &=N {\bf E}_m(g|{\mathcal I})(x) \end{split} \end{equation} for $m$-a.e.~$x\in F$. Applying the Birkhoff ergodic theorem again, we have \begin{align*} &\lim_{k\to +\infty}\frac{1}{k}\sum_{j=0}^{k-1}G(T_F^jx)={\bf E}_{m_F}(G|{\mathcal I}_F)(x) \quad \mbox{ and } \\ &\lim_{k\to +\infty}\frac{n_k(x)}{k}=\lim_{k\to +\infty}\frac{1}{k}\sum_{j=0}^{k-1}r_F(T_F^jx)={\bf E}_{m_F}(r_F|{\mathcal I}_F)(x) \end{align*} \mbox{ for $m$-a.e.~$x\in F$}. Here we have used the fact that $r_F\in L^1(F, {\mathcal B}|_F, m_F)$, which follows directly from Lemma \ref{lem-induce}(ii). Taking quotient we get $$ \lim_{k\to +\infty}\frac{1}{n_k(x)}\sum_{j=0}^{k-1}G(T_F^jx)\\ ={\bf E}_{m_F}(G|{\mathcal I}_F)(x)/{\bf E}_{m_F}(r_F|{\mathcal I}_F)(x) $$ for $m$-a.e.~$x\in F$. Combining this with \eqref{e-Gg} yields \eqref{e-identity}. \end{proof} \subsection{Oseledets' multiplicative ergodic theorem} \label{Sect-ose} Recall that the angle $\measuredangle (x,y)\in [0,\pi/2]$ between two vectors $x,y\in {\Bbb R}^d\backslash\{0\}$ is defined by $$ \sin \measuredangle (x,y)=\frac{(\|x\|^2\|y\|^2-\langle x, y\rangle^2)^{1/2}}{\|x\|\|y\|}, $$ where $\langle\cdot,\cdot\rangle$ is the standard inner product in ${\Bbb R}^d$. Similarly the angle between linear subspaces $U,V$ of ${\Bbb R}^d$ is defined by $$ \sin \measuredangle(U, V)=\inf_{x\in U\backslash\{0\},\; y\in V\backslash\{0\}}\sin \measuredangle (x,y). $$ Let $T$ be an invertible measure-preserving transformation of the Lebesgue space $(X,{\mathcal B}, m)$. We will require the following version of Oseledets' multiplicative ergodic theorem, due to Froyland et al. \cite[Theorem 4.1]{FroylandLloydQuas2010}: \begin{thm} \label{thm-2} Let $M: X\to {\Bbb R}^{d\times d}$ be a measurable function such that $$\int \log^+\|M(x)\|\; dm(x)<\infty.$$ Then there exists a measurable set $X'\subseteq X$ with $T(X')=X'$ and $m(X')=1$, such that for each $x\in X'$, there are positive integers $s(x), k_1(x),\ldots,k_{s(x)}(x)$ with $k_1(x)+\cdots+ k_{s(x)}(x)=d$, numbers $\lambda_1(x)>\cdots>\lambda_{s(x)}(x)\geq -\infty$ and a splitting ${\Bbb R}^d=E_x^1\oplus \cdots \oplus E_x^{s(x)}$ so that the following hold. \begin{itemize} \item[(i)] $\dim E_x^i=k_i(x)$. \item[(ii)] $M(x) E_x^i\subseteq E^i_{T x}$ (with equality if $\lambda_i(x)>-\infty$). \item[(iii)] For $1\leq i\leq s(x)$ and $v\in E_x^i\backslash\{0\}$, \begin{equation*} \lim_{n\to \infty} \frac{1}{n} \log \|M(T^{n-1}x)\cdots M(x)v\|=\lambda_i(x), \end{equation*} with uniform convergence on any compact subset of $E_x^i\backslash\{0\}$. \item[(iv)] For $1\leq i\leq s(x)$, \begin{equation*} \begin{split} &\lim_{n\to \infty} \frac{1}{n} \max_{v\in E^i_{T^{-n}x},\; \|v||=1} \log \|M(T^{-1}x)\cdots M(T^{-n}x)v\|\\ &=\lim_{n\to \infty} \frac{1}{n} \min_{v\in E^i_{T^{-n}x},\; \|v||=1} \log \|M(T^{-1}x)\cdots M(T^{-n}x)v\|=\lambda_i(x). \end{split} \end{equation*} \item[(v)] $\displaystyle\lim_{n\to \pm \infty} \frac{1}{n} \log \measuredangle ( \oplus_{i\in I}E^i_{T^nx}, \; \oplus_{j\in J} E^{j}_{T^n x})=0$ whenever $I\cap J=\emptyset$, \item[(vi)] The function $s:X'\to {\Bbb N}$ is measurable and $T$-invariant. \item[(vii)] The mappings $x\mapsto \lambda_i(x), E^i_x, k_i(x)$ are measurable on $\{x: s(x)\geq i\}$, and $\lambda_i(Tx)=\lambda_i(x)$, $k_i(Tx)=k_i(x)$. \end{itemize} \end{thm} \begin{rem} \label{rem-2.10} {\rm \begin{itemize} \item[(1)] Theorem \ref{thm-2} is only stated in \cite{FroylandLloydQuas2010} for the case when $m$ is ergodic. It extends directly to the general case by using ergodic decomposition. When $M(x)$ is invertible for all $x$ this is the classic Oseledets' multiplicative ergodic theorem, but we emphasize that the above is valid even in the non-invertible case (in which case the usual statements of Oseledets' theorem only provide a flag and not a splitting). \item[(2)] The uniform convergence in part (iii) of Theorem \ref{thm-2} is not stated in \cite{FroylandLloydQuas2010}. However it is well-known when $A$ takes values in $GL({\Bbb R},d)$, and the argument works also in the general case of ${\Bbb R}^{d\times d}$-valued cocycles. See e.g. \cite[p.~1111]{FengShmerkin2014} for a sketched proof. Part (iv) of Theorem \ref{thm-2} was only implicitly included in the proof of \cite[Theorem 4.1]{FroylandLloydQuas2010}. \item[(3)] The numbers $\lambda_1(x),\ldots, \lambda_{s(x)}(x)$ are called the {\em Lyapunov exponents} of $M$ at $x$ with respect to $m$. The number $k_i(x)$ is called the {\em multiplicity} of $\lambda_i(x)$. Moreover, $\{(\lambda_i(x), k_i(x))\}_{1\leq i\leq s(x)}$ is called the {\em Lyapunov spectrum} of $(M, m)$ over $X'$. \item[(4)] The decomposition $\bigoplus_{i=1}^{s(x)}E_x^i$ is called the {\em Oseledets splitting} of ${\Bbb R}^d$, and $E_x^i$, $1\leq i\leq s(x)$, are called the {\em Oseledets subspaces}. \end{itemize} } \end{rem} \section{Canonical coding maps for average contracting affine IFSs} \label{S-3} \ In this section, we prove the following proposition, which will be used in the proof of our main result. \begin{pro} \label{pro-3.1} Let ${\mathcal S}=\{S_j(x)=M_jx+a_j\}_{j\in \Lambda}$ be an affine IFS on ${\Bbb R}^d$ and $m\in {\mathcal M}_\sigma(\Sigma)$. Suppose that ${\mathcal S}$ is average contracting with respect to $m$. Let $\pi: \Sigma\to {\Bbb R}^d$ be given by \eqref{e-pi1.4}. Then there exists a Borel set $E\subset \Sigma$ with $\sigma(E)=E$ and $m(E)=1$ such that for any $x=(x_n)_{n=-\infty}^\infty\in E$, \begin{itemize} \item[(i)] $\pi(x)$ is well-defined, i.e.~the limit in defining $\pi(x)$ in \eqref{e-pi1.4} exists and is finite. \item[(ii)] $S_{x_0}(\pi\sigma x)=\pi(x)$. \item[(iii)] $ \lim_{n\to \infty} \frac{1}{n} \log^+ \| \pi(\sigma^n x)\|=0 $, where $\log^+z=\max\{0, \log z\}$. \end{itemize} \end{pro} Part (i) of the above proposition was first proved by Brandt \cite{Brandt1986} in the special case when $m$ is a Bernoulli product measure, and it was then extended by Bougerol and Picard \cite{BougerolPicard1992} to the general case when $m$ is ergodic. For the convenience of the reader, we shall provide a self-contained proof of part (i). Before proving Proposition \ref{pro-3.1}, we shall first prove the following auxiliary result, which is a variant of Proposition 2.1 in \cite{FengKaenmaki2011}. \begin{pro} \label{lem-2.1} Let\, $T:\;X\to X$ be an ergodic measure-preserving transformation on a probability space $(X,{\mathcal B}, m)$. Let $\{f_n\}_{n=1}^\infty$ be a sequence of non-negative measurable functions on $X$ such that $\log^+f_1\in L^1(m)$ and \begin{equation} \label{e-pro} f_{n+k}(x)\leq f_n(x)f_k(T^n x) \end{equation} for all $n,k\in {\Bbb N}$ and $x\in X$. Set $\lambda=\lim_{n\to \infty}({1}/{n})\int \log f_n \; dm$. Then for any $\epsilon>0$, the following properties hold: \begin{itemize} \item[(i)] If $\lambda \neq -\infty$, then for $m$-a.e.~ $x\in X$, there exists a positive integer $n_0(x)$ such that \begin{equation} \label{e-2.2} |\log f_n(T^k x)-n\lambda|\leq (n+k)\epsilon \end{equation} for all $n\geq n_0(x)$ and $k\in {\Bbb N}$. \item[(ii)] If $\lambda=-\infty$, then for any $N>0$ and $m$-a.e.~$x\in X$, there exists a positive integer $n_0(x)$ such that \begin{equation} \label{e-2.5} \log f_n(T^k x)\leq -Nn+ (n+k)\epsilon \end{equation} for all $n\geq n_0(x)$ and $k\in {\Bbb N}$. \end{itemize} \end{pro} \begin{proof} Here we modify the arguments of \cite[Proposition 2.1]{FengKaenmaki2011}. We only prove (i). The proof of (ii) is similar. Assume that $\lambda\in {\Bbb R}$. Let $\epsilon>0$ and take $0<\delta<\epsilon/5$. By the Kingman's sub-additive ergodic theorem, for $m$-a.e.~$x\in X$ there exists $n_0(x)$ such that $$ |\log f_n(x)-n\lambda |\leq n\delta $$ for all $n \geq n_0(x)$ and, $$ |\log f_k(x)-k\lambda|\leq (n_0(x)+k)\delta $$ for all $k\in {\Bbb N}$. Hence by \eqref{e-pro}, for $n\geq n_0(x)$ and $k \in {\Bbb N}$, \begin{equation} \label{e-t} \begin{split} \log f_n(T^k x)&\geq \log f_{n+k}(x) - \log f_k(x) \\ &\geq (n+k)(\lambda-\delta)-k\lambda-(n_0(x)+k)\delta \\ &\geq n\lambda-2(n+k)\delta\geq n\lambda-(n+k)\epsilon. \end{split} \end{equation} To see the opposite inequality, take $\ell$ large enough such that $|\beta-\lambda|<\delta$, where $$ \beta:=\frac{1}{\ell}\int \log f_\ell \; dm. $$ Applying the Birkhoff ergodic theorem to the functions $\log f_j$ ($j=1,\ldots, 2\ell$), we obtain \begin{equation} \label{e-(13)} \lim_{p\to \infty} \frac{1}{p}\log f_j(T^px)=0 \quad \mbox{ for $1\leq j\leq 2\ell$ and $m$-a.e.~$x$}. \end{equation} Let $n\geq 2\ell$, and $x\in X$. Write $n=q\ell +s$ with $\ell\leq s\leq 2\ell-1$. By sub-multiplicativity, we have $$ f_n(x)\leq f_j(x) \left(\prod_{p=0}^{q-1} f_\ell(T^{p\ell+j}x)\right) f_{s-j}(T^{q\ell+j}x), \quad j=0, 1, \ldots, \ell-1, $$ where we take the convention that $f_0\equiv 1$. Taking product of these inequalities yields $$ (f_n(x))^\ell \leq \left(\prod_{j=0}^{\ell-1} f_j(x)\right) \left(\prod_{p=0}^{q\ell -1} f_\ell(T^{p}x)\right) \left(\prod_{j=0}^{\ell-1}f_{s-j}(T^{q\ell +j}x)\right), $$ so for $k\geq 0$, $$ (f_n(T^kx))^\ell\leq \left(\prod_{j=0}^{\ell-1} f_j(T^k x)\right) \left(\prod_{p=k}^{q\ell +k-1} f_\ell(T^{p}x)\right) \left(\prod_{j=0}^{\ell-1}f_{s-j}(T^{q\ell +k+j}x)\right). $$ Taking logarithm and dividing both sides by $\ell$ we have \begin{equation} \label{e-(14)} \log (f_n(T^kx))\leq \left(\sum_{i=0}^{n+k-s-1} \frac{1}{\ell} \log f_\ell(T^{i}x)\right) - \left(\sum_{i=0}^{k-1} \frac{1}{\ell} \log f_\ell(T^{i}x)\right)+\Lambda_1+\Lambda_2, \end{equation} where $\Lambda_1:=\sum_{j=0}^{\ell-1} \frac{1}{\ell} \log f_j(T^k x)$, $\Lambda_2:=\sum_{j=0}^{\ell-1} \frac{1}{\ell} \log f_{s-j}(T^{q\ell +k+j}x)$. Applying the Birkhoff ergodic theorem to the function $\frac{1}{\ell}\log f_\ell$, and combining it with \eqref{e-(13)}-\eqref{e-(14)}, we see that for $m$-a.e.~$x\in X$, there exists an integer $\tilde{n}_0(x)$ such that \begin{align*} \log f_n(T^k x) &\leq (n+k)(\beta+\delta) - k(\beta-\delta)+k\delta +(n+k)\delta+\tilde{n}_0(x)\delta \\ &\leq n\beta+(2n+4k+\tilde{n}_0(x))\delta\leq n\lambda+5(n+k)\delta\\ &\leq n \lambda +(n+k)\epsilon \end{align*} for all $n\geq \tilde{n}_0(x)$ and $k\in {\Bbb N}$. This together with (\ref{e-t}) yields (\ref{e-2.2}). \end{proof} As a direct corollary of Proposition \ref{lem-2.1}, we have the following. \begin{cor} \label{cor-2.1} Under the assumptions of Proposition \ref{lem-2.1}, for any $\epsilon, N>0$ and for $m$-a.e.~$x\in X$, there exists $c(x)>0$ such that $$|f_n(T^k x)|\leq c(x) \exp(n \max\{\lambda, -N\})\exp ((n+k)\epsilon)$$ for all $n,k\in {\Bbb N}$. \end{cor} \begin{proof}[Proof of Proposition \ref{pro-3.1}] Without loss of generality we may assume that $m$ is ergodic, since the general case can be proved by considering the ergodic decomposition of $m$. Set $f_n(x)=\| M_{x_0} \cdots M_{x_{n-1}}\|$ for $x\in \Sigma$ and $n\geq 1$. Let $f_0(x)\equiv 1$ for convention. Since ${\mathcal S}$ is average contracting with respect to $m$, we have $$ \lim_{n\to \infty} \frac{1}{n}\int \log f_n \;dm=:\lambda<0. $$ Let $0<\epsilon<-\lambda/3$. Applying Corollary \ref{cor-2.1} to $\{f_n\}$ and the shift map $\sigma:\; \Sigma\to \Sigma$ (in which we take $N=2\epsilon$), we see that for $m$-a.e.~$x$, there exists $c(x)>0$ such that $$ f_n(\sigma^k x)\leq c(x) e^{-2n\epsilon} e^{(n+k)\epsilon} $$ for any $n\geq 1$ and $k\geq 0$. It follows that for $m$-a.e.~$x$, \begin{align*} \sum_{n=0}^\infty \|M_{x_{k}}\cdots M_{x_{k+n-1}} a_{x_{k+n}}\| &\leq (\max_i\|a_i\|) \sum_{n=0}^\infty f_n(\sigma^k x) \\ & \leq (\max_i\|a_i\|) c(x) \sum_{n=0}^\infty e^{-2n\epsilon} e^{(n+k)\epsilon}\\ &=(\max_i\|a_i\|) c(x) (1-e^{-\epsilon})^{-1} e^{k\epsilon} \end{align*} for all $k\geq 0$. It follows that for $m$-a.e.~$x$, $\pi(\sigma^kx)$ is well-defined and $\|\pi(\sigma^kx)\|\leq (\max_i\|a_i\|) c(x) (1-e^{-\epsilon})^{-1} e^{k\epsilon}$ for all $k\geq 0$. That is enough to conclude the proposition. \end{proof} \section{Measurable partitions associated with affine IFSs} \label{S-4} Let ${\mathcal S}=\{M_j x+a_j\}_{j\in \Lambda}$ be an affine IFS on ${\Bbb R}^d$ and $m\in {\mathcal M}_\sigma(\Sigma)$. Suppose that ${\mathcal S}$ is average contracting with respect to $m$. In this section, under an additional assumption formulated later in \eqref{e-assump}, we construct a finite family of measurable partitions of $\Sigma$ and give some properties of these partitions and the corresponding conditional measures of $m$. Define $M:\;\Sigma\to {\Bbb R}^{d\times d}$ by $$ M(x)=M_{x_{-1}},\quad x=(x_n)_{n=-\infty}^\infty. $$ Applying Theorem \ref{thm-2} to the measure-preserving system $(\Sigma, \sigma^{-1}, m)$ and the matrix cocycle $M$, we get a measurable $\Sigma'\subset \Sigma$ with $\sigma(\Sigma')=\Sigma'$ and $m(\Sigma')=1$, so that the Lyapunov spectrum $$ \{(\lambda_i(x), k_i(x))\}_{1\leq i\leq s(x)} $$ and the Oseledets splitting $$ {\Bbb R}^d=E_x^1\oplus \cdots \oplus E_x^{s(x)} $$ are well-defined for $x\in \Sigma'$ (cf. Remark~\ref{rem-2.10}). In this case, for any $x\in \Sigma'$ and $1\leq i\leq s(x)$, \begin{equation} \label{e-neg} \lim_{n\to \infty} \frac{1}{n} \log \|M_{x_{-n}}\cdots M_{x_{-1}}v\|=\lambda_i(x) \quad \mbox{ for $v\in E_x^i\backslash\{0\}$}, \end{equation} with uniform convergence on any compact subset of $E_x^i\backslash\{0\}$, \begin{equation} \label{e-e?} \begin{split} &\lim_{n\to \infty} \frac{1}{n} \max_{v\in E^i_{\sigma^nx},\; \|v||=1} \log \|M_{x_0}\cdots M_{x_{n-1}}v\|\\ &\mbox{}\quad =\lim_{n\to \infty} \frac{1}{n} \min_{v\in E^i_{\sigma^n x},\; \|v||=1} \log \|M_{x_0}\cdots M_{x_{n-1}}v\|=\lambda_i(x), \end{split} \end{equation} and \begin{equation} \label{e-e4.3} \limsup_{n\to \infty} \frac{1}{n} \max_{v\in \oplus_{j=i}^{s(x)}E^j_{\sigma^nx},\; \|v||=1} \log \|M_{x_0}\cdots M_{x_{n-1}}v\|\leq \lambda_i(x). \end{equation} In addition, by Proposition \ref{pro-3.1} we may assume that the coding map $\pi$ is well-defined on $\Sigma'$ and that \begin{equation} \label{e-e4.1} \lim_{n\to \infty}\frac{1}{n} \log^+ \|\pi(\sigma^n x)\|=0 \quad \mbox{ for } x\in \Sigma'. \end{equation} Define for $x\in \Sigma'$, \begin{equation} V_x^i:=\oplus_{j=i+1}^{s(x)} E_x^j \quad \mbox{ for } i=0,\ldots,s(x)-1, \quad \mbox{ and } \quad V_x^{s(x)}:=\{0\}. \end{equation} By \eqref{e-neg}, we have \begin{equation} \label{e-v1} V_x^i=\left\{v\in {\Bbb R}^d:\; \limsup_{n\to \infty} \frac{1}{n} \log \|M_{x_{-n}}\cdots M_{x_{-1}}v\|\leq \lambda_{i+1}(x)\right\} \end{equation} for $x\in \Sigma'$, $i=0,\ldots, s(x)-1$. For $x=(x_j)_{j=-\infty}^\infty\in \Sigma$, we write $x^{-}=(x_j)_{j=-\infty}^{-1}$. The following simple fact is our starting point in constructing measurable partitions of $\Sigma'$. \begin{lem} \label{lem-3.1} Let $x,y\in \Sigma'$ with $x^{-}=y^{-}$. Then $s(x)=s(y)$ and $\lambda_i(x)=\lambda_i(y)$ for $1\leq i\leq s(x)$. Moreover, $V_x^{i}=V_y^{i}$ for $0\leq i\leq s(x)$. \end{lem} \begin{proof} For $x\in \Sigma'$ and $v\in {\Bbb R}^d\setminus \{0\}$, define $$ \lambda(x,v):=\lim_{n\to \infty} \frac{1}{n}\log \|M_{x_{-n}}\cdots M_{x_{-1}}v\|. $$ By \eqref{e-neg}, the above limit always exists and takes values in $\{\lambda_i(x):\; 1\leq i\leq s(x)\}$. Clearly $ \lambda(x,v)$ only depends on $v$ and $x^{-}$. Hence for $x, y\in \Sigma'$ with $x^-=y^-$, we have $s(x)=s(y)$ and $\lambda_i(x)=\lambda_i(y)$ for $1\leq i\leq s(x)$; by \eqref{e-v1} we also have $V_x^i=V_y^i$ for $1\leq i\leq s(x)$. This completes the proof of the lemma. \end{proof} In the remaining part of this section, we always make the following assumption: \begin{equation} \label{e-assump} \mbox{$s(x), k_1(x),\ldots, k_{s(x)}(x)$ are constant for $m$-a.e.~$x\in \Sigma'$}. \end{equation} Here we don't make the stronger assumption that $m$ is ergodic. Let us write these constants as $s, k_1,\ldots, k_s$. Below we construct a finite family of measurable partitions $\xi_0,\ldots, \xi_s$ of $\Sigma'$. Let $\xi_0$ be the partition of $\Sigma'$ so that the $\xi_0$-atom containing $x=(x_j)_{j=-\infty}^{+\infty}\in \Sigma'$ is given by $$ \xi_0(x)=\{y=(y_j)_{j=-\infty}^\infty\in \Sigma':\; y_j=x_j \mbox{ for }j\leq -1\}. $$ By Lemma \ref{lem-3.1}, $V_y^i=V_x^i$ for any $y\in \xi_0(x)$ and $i\in \{0,1,\ldots, s\}$. Similarly, for $i\in \{1,\ldots, s\}$, we define the partition $\xi_i$ of $\Sigma'$ by $$ \xi_i(x)=\{y=(y_j)_{j=-\infty}^\infty\in \xi_0(x):\; \pi y-\pi x \in V_x^i\},\qquad x\in \Sigma'. $$ \begin{lem} $\xi_0,\ldots, \xi_s$ are measurable partitions of $(\Sigma',{\mathcal B}(\Sigma'), m)$. \end{lem} \begin{proof} By Rohlin theory (cf. \cite[Section~2.5]{Rohlin1949}, \cite[Chapter IV]{Parry1969}), it is enough to show that for every $i\in \{0,1,\ldots, s\}$, one can construct a measurable mapping $\pi_i$ from $\Sigma'$ to a complete separable metric space $Y_i$ such that $\xi_i$ is induced by $\pi_i$, in the sense that $\xi_i=\{\pi_i^{-1}(y):\; y\in Y_i\}$. Below we construct such mappings $\pi_i$. Let $\Sigma^-:=\{(x_n)_{n=-\infty}^{-1}:\; x_n\in \Lambda\mbox{ for all }n\leq -1\}$ and endow it with a suitable metric compatible to the discrete product topology. For $j\in \{0,\ldots, d\}$, the set of all $j$-dimensional affine subspaces in ${\Bbb R}^d$ forms a closed smooth manifold, which is called the $(d, j)$-{\it affine Grassmannian} and is denoted by ${\rm Graff}(d, j)$. Set $Y_i=\Sigma^-\times {\rm Graff}(d, k_{i+1}+\cdots+k_s)$ for $i\in \{0,\ldots, s-1\}$ and $Y_s=\Sigma^-\times {\Bbb R}^d$. Define $\pi_i:\;\Sigma'\to Y_i$ ($i=0,1,\ldots, s$) by $$ x\mapsto (x^-, V_x^i+\pi x). $$ It is readily checked that for each $i$, $\pi_i$ is measurable and $\xi_i$ is induced by $\pi_i$. \end{proof} \begin{rem} \label{rem-tt} The above construction of the measurable partitions $\xi_0,\cdots, \xi_s$ is different from that built in the previous work of \cite{FengHu2009, BaranyKaenmaki2015}. In \cite{FengHu2009}, the partitions were made on the one-sided shift space due to the simple structure of Oseledets splitting subspaces. In \cite{BaranyKaenmaki2015}, the partitions were made on the product space of the self-affine set and the flag manifolds. \end{rem} Let ${\mathcal P}$ be the canonical partition of $\Sigma'$ given in \eqref{e-pp}. For $n\in {\Bbb N}$, set $${\mathcal P}_0^{n-1}=\bigvee_{j=0}^{n-1} \sigma^{-j}{\mathcal P},$$ where $\vee$ stands for the join of partitions (cf. \cite{Parry1981}). For convenience, write \begin{equation} \label{e-qn} Q_{n,\epsilon}:=\left\{x\in \Sigma':\; \|\pi \sigma^j x\|\leq (1/2) e^{j\epsilon/2} \mbox{ for all }j\geq n\right\} \end{equation} for $n\in{\Bbb N}$ and $\epsilon>0$. Below we give several lemmas to further illustrate the properties of $\xi_i$ and the associated conditional measures. \begin{lem} \label{lem-3.4} \begin{itemize} \item[(1)] For $x\in \Sigma'$, $i\in \{0,\ldots, s\}$ and $n\in {\Bbb N}$, $$ \xi_i(x)\cap {\mathcal P}_{0}^{n-1} (x)= \sigma^{-n}(\xi_i(\sigma^n x)). $$ As a consequence, $\xi_i\vee {\mathcal P}_{0}^{n-1}=(\sigma^{-n}\xi_i)\vee {\mathcal P}_{0}^{n-1}=\sigma^{-n}\xi_i$. \item[(2)] Let $x\in \Sigma'$ and $\epsilon>0$. Then for $i\in \{0,\ldots, s-1\}$, \begin{equation} \label{e-l2} Q_{n,\epsilon}\cap \xi_i(x)\cap {\mathcal P}_{0}^{n-1} (x)\subset \left \{\begin{array}{ll} B^\pi(x, e^{n (\lambda_{i+1}(x)+2\epsilon)}) & \mbox{ if }\lambda_{i+1}(x)\neq -\infty\\ B^\pi(x, e^{-n/\epsilon}) & \mbox{ if }\lambda_{i+1}(x)=-\infty \end{array} \right. \end{equation} when $n$ is large enough, here $B^\pi(x,r)$ is defined as in \eqref{e-ball}. Moreover, \begin{equation} \label{e-l3} Q_{n,\epsilon}\cap {\mathcal P}_{0}^{n-1} (x)\subset\left \{\begin{array}{ll} B^\pi(x, e^{n (\lambda_{1}(x)+2\epsilon)}) & \mbox{ if }\lambda_{1}(x)\neq -\infty\\ B^\pi(x, e^{-n/\epsilon}) & \mbox{ if }\lambda_{1}(x)=-\infty \end{array} \right. \end{equation} when $n$ is large enough. \end{itemize} \end{lem} \begin{proof} We first prove (1). Let $x=(x_j)_{j=-\infty}^\infty\in \Sigma'$, $i\in \{0,\ldots, s\}$ and $n\in {\Bbb N}$. We only prove that $\xi_i(x)\cap {\mathcal P}_{0}^{n-1} (x)\subset\sigma^{-n}(\xi_i(\sigma^n x))$. The proof of the other direction is similar. Let $y=(y_j)_{j=-\infty}^\infty \in \xi_i(x)\cap {\mathcal P}_{0}^{n-1} (x)$. Then $\pi y-\pi x\in V_x^i$ and $y_j=x_j$ for $j\leq n-1$. By Proposition \ref{pro-3.1}(ii), \begin{equation} \label{e-pi} \begin{split} \pi y-\pi x & =S_{y_0\ldots y_{n-1}}(\pi \sigma^n y)-S_{x_0\ldots x_{n-1}}(\pi \sigma^n x)\\ &=S_{x_0\ldots x_{n-1}}(\pi \sigma^n y)-S_{x_0\ldots x_{n-1}}(\pi \sigma^n x)\\ &=M_{x_0\ldots x_{n-1}}(\pi \sigma^n y-\pi \sigma^n x),\\ \end{split} \end{equation} here and afterwards we write $M_{i_1\ldots i_n}$ for $M_{i_1}\cdots M_{i_n}$. Since $\pi y-\pi x\in V^i_x$, by \eqref{e-pi} and \eqref{e-v1} we have \begin{equation*} \begin{split} \limsup_{k\to \infty}&\frac{1}{n+k} \log \|M_{x_{-k}\ldots x_{-1}x_0\ldots x_{n-1}}(\pi \sigma^n y-\pi \sigma^n x)\|\\ =&\limsup_{k\to \infty}\frac{1}{n+k}\log \|M_{x_{-k}\ldots x_{-1}}(\pi y-\pi x)\| \leq \lambda_{i+1}(x)=\lambda_{i+1}(\sigma^nx). \end{split} \end{equation*} Applying \eqref{e-v1} to $V^i_{\sigma^nx}$ gives $\pi \sigma^n y-\pi \sigma^n x\in V_{\sigma^nx}^i$. In the meantime, since $y_j=x_j$ for $j\leq n-1$, we have also $\sigma^n y\in \xi_0(\sigma^n x)$. Therefore $y\in \sigma^{-n}(\xi_i(\sigma^n x))$. This proves $\xi_i(x)\cap {\mathcal P}_{0}^{n-1} (x)\subset\sigma^{-n}(\xi_i(\sigma^n x))$. Next we prove (2). Let $x\in \Sigma'$, $i\in \{0,\ldots, s-1\}$ and $\epsilon>0$. By \eqref{e-e4.3} and \eqref{e-e4.1}, there exists $n_0$ such that for any $n\geq n_0$, \begin{equation} \label{e-l1} \max_{v\in V_{\sigma^nx}^i,\; \|v\|=1} \|M_{x_0\ldots x_{n-1}}v\|\leq \left\{ \begin{array}{ll} e^{n(\lambda_{i+1}(x)+{\epsilon})} & \mbox{ if } \lambda_{i+1}(x)\neq -\infty\\ e^{-2n/\epsilon} & \mbox{ if } \lambda_{i+1}(x)= -\infty \end{array} \right. \end{equation} and \begin{equation} \label{e-pisigma} \| \pi \sigma^n x\|\leq \frac{1}{2}e^{n\epsilon/2}. \end{equation} Now let $n\geq n_0$ and $y\in Q_{n,\epsilon} \cap \xi_i(x)\cap {\mathcal P}_{0}^{n-1} (x)$. Then $\|\pi\sigma^n y\|\leq (1/2) e^{n\epsilon/2}$, $y^{-}=x^{-}$, $\pi y-\pi x\in V_x^i$ and furthermore by (1), $\pi\sigma^ny-\pi\sigma^nx\in V_{\sigma^nx}^i$. By \eqref{e-pi}-\eqref{e-pisigma}, \begin{align*} {\|\pi y-\pi x\|} &= \|M_{x_0\ldots x_{n-1}}(\pi\sigma^n y-\pi \sigma^n x)\|\\ &\leq \left( \max_{v\in V_{\sigma^nx}^i,\; \|v\|=1} \|M_{x_0\ldots x_{n-1}}v\| \right) \|\pi \sigma^n y-\pi \sigma^n x\|\\ & \leq \left\{ \begin{array}{ll} e^{n(\lambda_{i+1}(x)+2{\epsilon})} & \mbox{ if } \lambda_{i+1}(x)\neq -\infty\\ e^{-n/\epsilon} & \mbox{ if } \lambda_{i+1}(x)= -\infty \end{array} \right. . \end{align*} This proves \eqref{e-l2}. Moreover, since $V_x^0={\Bbb R}^d$, the above argument for the case $i=0$ actually proves \eqref{e-l3}. \end{proof} Recall that for a measurable partition $\eta$ of $\Sigma'$, $\{m_x^\eta\}$ stands for the canonical system of conditional measures associated with $\eta$ (cf. Section~\ref{Sect-con}). \begin{lem} \label{lem-3.6} Let $i\in \{0,1,\ldots, s\}$. Then for $m$-a.e.~$x\in \Sigma'$, the following hold. \begin{itemize} \item[(1)] $ m_x^{\sigma^{-n}\xi_i}(A) =m_{\sigma^n x}^{\xi_i}(\sigma^n A)$ for any $n\in {\Bbb N}$ and measurable $A\subset \Sigma'$. \medskip \item[(2)] $\displaystyle m_x^{\sigma^{-n} \xi_i}(A)= \frac {m_x^{\xi_i}(A\cap {\mathcal P}_0^{n-1}(x))} {m_x^{\xi_i}({\mathcal P}_0^{n-1}(x))} $ for any $n\in {\Bbb N}$ and measurable $A\subset \Sigma'$. \medskip \item[(3)] $\displaystyle \frac {m_x^{\xi_i} (\sigma^{-n} A\cap {\mathcal P}_0^{n-1}(x))} {m_{\sigma^n x}^{\xi_i}(A)}=m_x^{\xi_i}({\mathcal P}_0^{n-1}(x)) $ for any $n\in {\Bbb N}$ and measurable $A\subset \Sigma'$. \end{itemize} \end{lem} \begin{proof} All the results follow from the $\sigma$-invariance of $m$ and the uniqueness of conditional measures. For the reader's convenience, we include below the detailed arguments. To see (1), fix $n\in {\Bbb N}$ and define a family of probability measures $\{\mu_x\}_{x\in \Sigma'}$ such that $\mu_x$ is supported on $(\sigma^{-n}\xi_i)(x)=\sigma^{-n}(\xi_i(\sigma^n x))$ and satisfies $$ \mu_x(A)=m_{\sigma^n x}^{\xi_i}(\sigma^n A) \quad \mbox{ for any measurable } A\subset \Sigma'. $$ Then by Theorem \ref{thm-2.1}, for every measurable $A\subset \Sigma'$ and $m$-a.e.~$x$, \begin{align*} \mu_x(A)&={\bf E}_m(\chi_{\sigma^n A}|\widehat{\xi_i})(\sigma^n x) \\ &={\bf E}_m(\chi_{\sigma^n A}\circ \sigma^n |\sigma^{-n} \widehat{\xi_i})( x) \qquad \mbox{ (by Lemma \ref{lem-par}(i))} \\ &={\bf E}_m(\chi_{ A} |\sigma^{-n} \widehat{\xi_i})( x). \end{align*} It follows that $x\mapsto \mu_x(A)$ is $\sigma^{-n}\widehat{\xi_i}$-measurable and $m(A)=\int \mu_x(A) dm(x)$. Therefore, $\{\mu_x\}$ is a canonical system of conditional measures associated with $\sigma^{-n}\xi_i$. By the uniqueness of conditional measures, we have $\mu_x=m^{\sigma^{-n}\xi_i}_x$ for $m$-a.e.~$x$. This proves (1). To see (2), let $n\in {\Bbb N}$ and notice that $\sigma^{-n} \xi_i=\xi_i\vee {\mathcal P}_{0}^{n-1}$ by Lemma \ref{lem-3.4}(1). Similar to the proof of (1), we define a family of probability measures $\{\nu_x\}_{x\in \Sigma'}$ such that $\nu_x$ is supported on $(\sigma^{-n}\xi_i)(x)=\xi_i(x)\cap {\mathcal P}_0^{n-1}(x)$ and satisfies $$ \nu_x(A)= \frac {m_x^{\xi_i}(A\cap {\mathcal P}_0^{n-1}(x))} {m_x^{\xi_i}({\mathcal P}_0^{n-1}(x))} \quad \mbox{ for any measurable } A\subset \Sigma'. $$ Then by Theorem \ref{thm-2.1}, for every measurable $A\subset \Sigma'$ and $m$-a.e.~$x$, \begin{align} \label{e-chi} \nu_x(A)&=\sum_{B\in {\mathcal P}_0^{n-1}} \chi_B(x)\cdot h_B(x), \end{align} where $h_B:={\bf E}_m(\chi_{A\cap B}|\widehat{\xi_i})/{\bf E}_m(\chi_{ B}|\widehat{\xi_i})$. Since $h_B$ is $\widehat{\xi_i}$-measurable, the mapping $x\mapsto \nu_x(A)$ is $\widehat{\xi_i}\vee \widehat{{\mathcal P}_0^{n-1}}$-measurable (i.e. $\sigma^{-n}\widehat{\xi_i}$-measurable). Moreover by \eqref{e-chi}, \begin{align*} \int \nu_x(A) \;dm(x)&= \sum_{B\in {\mathcal P}_0^{n-1}}\int \chi_B h_B \;dm\\ &=\sum_{B\in {\mathcal P}_0^{n-1}}\int {\bf E}_m(\chi_B h_B|\widehat{\xi_i}) \;dm\\ &=\sum_{B\in {\mathcal P}_0^{n-1}}\int {\bf E}_m(\chi_B|\widehat{\xi_i} ) h_B \;dm\\ &=\sum_{B\in {\mathcal P}_0^{n-1}}\int {\bf E}_m( \chi_{A\cap B}|\widehat{\xi_i}) \;dm\\ &=\sum_{B\in {\mathcal P}_0^{n-1}}m(A\cap B)=m(A). \end{align*} Hence the family $\{\nu_x\}$ is a canonical system of conditional measures associated with $\sigma^{-n}\xi_i$, and so (2) follows by the uniqueness of conditional measures. Finally we prove (3). By (1), we have $$\displaystyle m_{\sigma^n x}^{\xi_i}(A)= m_{\sigma^n x}^{\xi_i}(\sigma^n(\sigma^{-n} A))=m_{x}^{\sigma^{-n} \xi_i}(\sigma^{-n} A).$$ Applying (2) to $\sigma^{-n} A$ (instead of $A$) yields that $$ m_{\sigma^n x}^{\xi_i}(A)=m_{x}^{\sigma^{-n} \xi_i}(\sigma^{-n} A)=\frac {m_x^{\xi_i} (\sigma^{-n} A\cap {\mathcal P}_0^{n-1}(x))} {m_x^{\xi_i}({\mathcal P}_0^{n-1}(x))}, $$ which implies (3). \end{proof} Now for $i\in \{0,1,\ldots, s\}$, define \begin{equation} \label{e-hi} h_i(x)={\bf E}_m(f_i|{\mathcal I})(x), \quad x\in \Sigma', \end{equation} where $f_i:={\bf I}_m({\mathcal P}|\widehat{\xi_i})$ and ${\mathcal I}=\{A\in {\mathcal B}(\Sigma'):\; \sigma^{-1}A=A\}$. By Lemma \ref{lem-par}(v), $f_i\geq 0$ a.e.~and $f_i\in L^1$. It follows that $h_i\geq 0$ a.e.~and $h_i\in L^1$. \begin{lem} \label{lem-3.5} Let $i\in \{0,1,\ldots, s\}$. Then for $m$-a.e.~$x\in \Sigma'$, \begin{equation} \label{e-h1} \begin{split} &\log m_x^{\xi_i}({\mathcal P}_{0}^{n-1}(x))=-\sum_{j=0}^{n-1}{\bf I}_m({\mathcal P} |\widehat{\xi}_i)(\sigma^jx)\quad\mbox{ and }\\ &-\lim_{n\to \infty} \frac{1}{n}\log m_x^{\xi_i}({\mathcal P}_{0}^{n-1}(x))=h_i(x). \end{split} \end{equation} Furthermore, \begin{equation} \label{e-h2} -\lim_{n\to \infty} \frac{1}{n}\log m({\mathcal P}_{0}^{n-1}(x))= h_0(x) \quad \mbox{ for $m$-a.e. }x\in \Sigma'. \end{equation} \end{lem} \begin{proof} Let $i\in \{0,1,\ldots, s\}$. By Theorem \ref{thm-2.1}, $$m_x^{\xi_i}({\mathcal P}_{0}^{n-1}(x))=\sum_{A\in {\mathcal P}_{0}^{n-1}}\chi_A(x)m_x^{\xi_i}(A)=\sum_{A\in {\mathcal P}_{0}^{n-1}}\chi_A(x){\bf E}_m(\chi_A|\widehat{\xi_i})(x)$$ and hence $-\log m_x^{\xi_i}({\mathcal P}_{0}^{n-1}(x))={\bf I}_m({\mathcal P}_0^{n-1}|\widehat{\xi_i})(x)$ for $m$-a.e.~$x$. By Lemma \ref{lem-par}, \begin{align*} {\bf I}_m({\mathcal P}_0^{n-1}|\widehat{\xi_i})&={\bf I}_m({\mathcal P}|\widehat{\xi_i})+{\bf I}_m\left(\bigvee_{j=1}^{n-1}\sigma^{-j} {\mathcal P}\big|\widehat{\xi}_i\vee \widehat{{\mathcal P}}\right)\\ &={\bf I}_m({\mathcal P}|\widehat{\xi}_i)+{\bf I}_m\left(\bigvee_{j=1}^{n-1}\sigma^{-j} {\mathcal P}\big|\sigma^{-1} \widehat{\xi}_i \right)\qquad (\mbox{by Lemma \ref{lem-3.4}(1)})\\ &={\bf I}_m({\mathcal P}|\widehat{\xi}_i)+{\bf I}_m({\mathcal P}_{0}^{n-2} |\widehat{\xi}_i)\circ \sigma \qquad \qquad (\mbox{by Lemma \ref{lem-par}(ii)}). \end{align*} Therefore by induction we have \begin{equation} \label{e-xi} {\bf I}_m({\mathcal P}_0^{n-1}|\widehat{\xi}_i)=\sum_{j=0}^{n-1}{\bf I}_m({\mathcal P} |\widehat{\xi}_i)\circ \sigma^j. \end{equation} Now \eqref{e-h1} follows from \eqref{e-xi} and the Birkhoff ergodic theorem. To see \eqref{e-h2}, applying the Shannon-McMillian-Breiman theorem (see e.g. \cite[p.~39] {Parry1981}) to the transformations $\sigma$ and $\sigma^{-1}$ respectively, we have the following convergences (pointwise and in $L^1$): \begin{equation} \label{e-h4} \begin{split} -\lim_{n\to +\infty} \frac{1}{n}\log m({\mathcal P}_{0}^{n-1}(x))&={\bf E}_m(g_1|{\mathcal I})(x),\\ -\lim_{n\to +\infty} \frac{1}{n}\log m({\mathcal P}_{-(n-1)}^{0}(x))&={\bf E}_m(g_2|{\mathcal I})(x), \end{split} \end{equation} where $g_1:={\bf I}_m({\mathcal P}|\bigvee_{j=1}^\infty \sigma^{-j}\widehat{\mathcal P})$, $g_2:={\bf I}_m({\mathcal P}|\bigvee_{j=1}^\infty \sigma^{j}\widehat{\mathcal P})$. Noticing that $\widehat \xi_0=\bigvee_{j=1}^\infty \sigma^{j}\widehat{\mathcal P}$, we have $g_2={\bf I}_m({\mathcal P}|\widehat \xi_0)=f_0$ and so ${\bf E}_m(g_2|{\mathcal I})=h_0$. To prove \eqref{e-h2}, by \eqref{e-h4} it suffices to show that \begin{equation} \label{e-h3} {\bf E}_m(g_1|{\mathcal I})(x)={\bf E}_m(g_2|{\mathcal I})(x)\quad \mbox{ for $m$-a.e.~$x$}. \end{equation} To see \eqref{e-h3} first observe that for $x\in \Sigma'$, ${\mathcal P}_{-(n-1)}^{0}(\sigma^n x)=\sigma^n({\mathcal P}_0^{n-1}( x))$ and hence $m({\mathcal P}_{-(n-1)}^{0}(\sigma^n x))=m(\sigma^n({\mathcal P}_0^{n-1}( x)))=m({\mathcal P}_0^{n-1}( x))$. For any $B\in {\mathcal I}$, we have \begin{align*} \int_B & \log m({\mathcal P}_{-(n-1)}^{0}(x)) \; dm(x)\\ &=\int \chi_B(x) \log m({\mathcal P}_{-(n-1)}^{0}(x)) \; dm(x)\\ &=\int \chi_B(\sigma^n x) \log m({\mathcal P}_{-(n-1)}^{0}(\sigma^n x)) \; dm(x) \qquad(\mbox{by the $\sigma$-invariance of $m$}) \\ &=\int \chi_B(\sigma^nx) \log m({\mathcal P}_0^{n-1}(x)) \; dm(x)\\ &=\int \chi_B(x) \log m({\mathcal P}_0^{n-1}(x)) \; dm(x) \qquad\qquad(\mbox{by $\chi_B=\chi_B\circ \sigma^n$ as $B\in {\mathcal I}$}) \\ &=\int_B \log m({\mathcal P}_0^{n-1}(x)) \; dm(x). \end{align*} Dividing both sides by $n$, letting $n\to \infty$ and applying \eqref{e-h4}, we have $$\int_B {\bf E}_m(g_1|{\mathcal I})\; dm= \int_B {\bf E}_m(g_2|{\mathcal I})\; dm\quad \mbox{ for all } B\in {\mathcal I}.$$ Therefore ${\bf E}_m(g_1|{\mathcal I})= {\bf E}_m(g_2|{\mathcal I})$ almost everywhere. This completes the proof of the lemma. \end{proof} Below we give an interesting corollary of Lemma \ref{lem-3.5}, although we will not use it in the rest part of the paper. \begin{cor} Let $i\in \{0,1,\ldots, s\}$. Then $h_i=0$ a.e.~if and only if $m_x^{\xi_i}=\delta_x$ (i.e.~$m_x^{\xi_i}(\{x\})=1$) for $m$-a.e.~$x\in \Sigma'$. \end{cor} \begin{proof} By Lemma \ref{lem-par}(v), $f_i:={\bf I}_m({\mathcal P}|\widehat{\xi_i})\geq 0$ a.e.~ and $f_i\in L^1$. Hence by \eqref{e-hi}, $h_i=0$ a.e.~if and only if $f_i=0$ a.e. However according to the first equality in \eqref{e-h1}, the condition $f_i=0$ a.e.~ implies that for $m$-a.e.~$x$, $m_x^{\xi_i}(\mathcal P_0^{n-1}(x))=1$ for every $n\geq 1$ and hence $$m_x^{\xi_i}(\{x\})=m_x^{\xi_i}\left(\xi_0(x)\cap {\mathcal P}_0^{\infty}(x)\right)=m_x^{\xi_i}\left( {\mathcal P}_0^{\infty}(x)\right)=1,$$ using the fact that $m_x^{\xi_i}$ is supported on $\xi_i(x)\subset \xi_0(x)$. Conversely, by the first equality in \eqref{e-h1} (applied to $n=1$), we obtain that $f_i(x)=-\log m_x^{\xi_i}({\mathcal P}(x))$; hence the condition $$m_x^{\xi_i}(\{x\})=1 \mbox{ a.e.}$$ implies that $f_i=0$ a.e. This completes the proof of the corollary. \end{proof} We end the section by the following. \begin{lem} \label{lem-4.6} Let $\epsilon>0$ and define $Q_{n,\epsilon}$ as in \eqref{e-qn} for $n\in {\Bbb N}$. Then for $m$-a.e.~$x\in \Sigma'$, $$ \lim_{n\to \infty} \frac{m_x^{\xi_i}(Q_{n,\epsilon}\cap {\mathcal P}_0^{n-1}(x))}{m_x^{\xi_i}({\mathcal P}_0^{n-1}(x))}=1 \qquad (i=0,1,\ldots, s) $$ and $$ \lim_{n\to \infty} \frac{m(Q_{n,\epsilon}\cap {\mathcal P}_0^{n-1}(x))}{m({\mathcal P}_0^{n-1}(x))}=1. $$ \end{lem} \begin{proof} The equalities follow from the Borel density lemma and the facts that the sequence $(Q_{n,\epsilon})$ of sets is monotone increasing as $n$ increases, and $\bigcup_n Q_{n,\epsilon}$ is of full $m$-measure by Proposition \ref{pro-3.1}(iii). \end{proof} \section{Transverse dimensions} \label{S-5} In this section, we prove an inequality for the transverse dimensions of the conditional measures that we constructed in Section~\ref{S-4}. Recall that ${\mathcal S}$ is an affine IFS on ${\Bbb R}^d$ of the form \eqref{e-form}, average contracting with respect to some $m\in {\mathcal M}_\sigma(\Sigma)$. Let $\pi$ be the associated coding map. Suppose that $H_m({\mathcal P})<\infty$, where ${\mathcal P}$ is defined as in \eqref{e-pp}. Let $M:\;\Sigma\to {\Bbb R}^{d\times d}$ be the matrix cocycle given by $M(x)=M_{x_{-1}}$, and $\{(\lambda_i(x), k_i(x))\}_{1\leq i\leq s(x),\; x\in \Sigma'}$ the Lyapunov spectrum for $M$ with respect to the transformation $\sigma^{-1}$. Suppose that \eqref{e-assump} holds, i.e. there exist $s, k_1,\ldots, k_s$ so that $s(x)=s$, $k_i(x)=k_i$ ($i=1,\ldots, s$) for $m$-a.e.~$x\in \Sigma'$. Let $\oplus_{i=1}^s E_x^i$ be the Oseledets splitting of ${\Bbb R}^d$, and $\{0\}=V_x^s\subset \cdots \subset V^0_x={\Bbb R}^d$ the associated filtration. Let $\xi_0,\xi_1,\ldots, \xi_s$ be the measurable partitions of $\Sigma'$ that we constructed in Section~\ref{S-4}. For $x\in \Sigma'$ and $r>0$, set \begin{equation} \label{gamma} {\Gamma_i}(x,r)=\{y\in \Sigma':\; \mbox{dist}(\pi y+V_x^i, \; \pi x+V_x^i)\leq r\}, \quad i=1,\ldots, s \end{equation} and define $$ \vartheta_{i-1}(x)=\liminf_{r\to 0}\frac{\log m^{\xi_{i-1}}_x({\Gamma_{i}}(x,r))}{\log r}, \quad i=1,\ldots, s. $$ We call $\vartheta_0$,\ldots, $\vartheta_{s-1}$ the {\it transverse dimensions} of $m$. Intuitively we may view $\vartheta_i(x)$ as the dimension of $m$ along the direction $E_x^{i+1}$. The main result of this section is the following, which plays a key role in the proof of Theorem \ref{thm-1.0}. \begin{pro} \label{pro-4.1} For $m$-a.e.~$x\in \Sigma'$, $$ \vartheta_{i-1}(x)\geq \frac{h_{i}(x)-h_{i-1}(x)}{\lambda_{i}(x)}, \quad i=1,\ldots, s, $$ where $h_i$ are defined as in \eqref{e-hi}. \end{pro} This result can be viewed as an analogue of Proposition 11.2 in \cite{LedrappierYoung1985}. A stronger version of the result, with the inequality being replaced by the equality, was proved earlier in \cite[Theorem 6.2]{FengHu2009}, \cite[Theorem 3.3]{Barany2015}, and \cite[Propositions 5.3 and 7.3]{BaranyKaenmaki2015} under various additional assumptions. The proof of Proposition \ref{pro-4.1} is quite long and delicate. Besides extending some ideas from the previous works \cite{LedrappierYoung1985, FengHu2009, BaranyKaenmaki2015}, we need to employ certain new strategy as well. We first introduce some notation and give several lemmas. For $x\in \Sigma'$, $i\in \{1,\ldots, s\}$ and $r>0$, set $$ B^i_x(r)=\{v\in E^i_x:\; \|v\|\leq r\} $$ and $$ T_i(x,r)=\{y\in \xi_{i-1}(x):\; \pi y-\pi x\in V_x^i\oplus B^i_x(r)\}. $$ \begin{lem} \label{lem-5.2} Let $x\in \Sigma'$, $i\in \{1,\ldots, s\}$, $n\in {\Bbb N}$ and $r>0$. For $$0\leq a\leq \min_{v\in E_{\sigma^n x}^i,\; \|v\|=1} \|M_{x_0\cdots x_{n-1}}v\|,$$ we have \begin{equation} \label{e-i2} T_i(x, a r)\cap {\mathcal P}_0^{n-1}(x)\subset \sigma^{-n} T_{i}(\sigma^nx, r). \end{equation} \end{lem} \begin{proof} Let $y\in T_i(x, ar)\cap {\mathcal P}^{n-1}_0(x)$. By definition, \begin{align} & y \in \xi_{i-1}(x)\cap {\mathcal P}^{n-1}_0(x) \quad \mbox{ and } \label{e-i3}\\ & \pi y - \pi x\in V_{x}^i\oplus B^i_{x}(ar). \label{e-i4} \end{align} By \eqref{e-i3} and Lemma \ref{lem-3.4}(1), $y\in \sigma^{-n}(\xi_{i-1}(\sigma^n x))$. Moreover since $y\in {\mathcal P}_{0}^{n-1}(x)$, by \eqref{e-pi}, \begin{equation} \label{e-i5}\pi y-\pi x = M_{x_0\ldots x_{n-1}}(\pi \sigma^n y-\pi\sigma^n x). \end{equation} Since $y\in \xi_{i-1}(x)$, by definition $\pi y-\pi x\in V_{x}^{i-1}=V_x^i\oplus E_x^i$. Applying \eqref{e-v1} to $V_x^{i-1}$ yields $$\limsup_{k\to \infty}\frac{1}{k}\log \|M_{x_{-k}\ldots x_{-1}}(\pi y-\pi x)\|\leq \lambda_i(x)=\lambda_i(\sigma^nx),$$ where the last equality follows from Theorem \ref{thm-2}(vii). Hence by \eqref{e-i5}, $$\limsup_{k\to \infty}\frac{1}{n+k}\log \|M_{x_{-k}\ldots x_{-1}x_0\ldots x_{n-1}}(\pi \sigma^n y-\pi \sigma^n x)\|\leq \lambda_i(\sigma^nx).$$ Applying \eqref{e-v1} to $V_{\sigma^nx}^{i-1}$ gives $\pi \sigma^n y-\pi\sigma^n x\in V_{\sigma^n x}^{i-1}=V_{\sigma^nx}^i\oplus E_{\sigma^n x}^i$. Write \begin{align*} \pi y-\pi x &=v_1+ w_1 \quad \mbox{ with $v_1\in V_{x}^i$ and $w_1\in E_x^i$},\\ \pi \sigma^n y-\pi \sigma^n x& =v_2+ w_2 \quad \mbox{ with $v_2\in V_{\sigma^n x}^i$ and $w_2\in E_{\sigma^nx}^i$}. \end{align*} By \eqref{e-i4}, $w_1\in B^i_{x}(ar)$ and hence $\|w_1\|\leq ar$. Since $M_{x_0\cdots x_{n-1}}V_{\sigma^nx}^i\subset V_x^i$ and $M_{x_0\cdots x_{n-1}}E_{\sigma^nx}^i\subset E_x^i$, by \eqref{e-i5} we see that $w_1=M_{x_0\ldots x_{n-1}}w_2$ and so $$ ar\geq \|w_1\|=\|M_{x_0\ldots x_{n-1}}w_2\|\geq a\|w_2\|. $$ It follows that $\|w_2\|\leq r$. Hence $\pi \sigma^n y-\pi\sigma^n x\in V_{\sigma^nx}^i\oplus B_{\sigma^nx}^i(r)$. This together with $y\in \sigma^{-n}(\xi_{i-1}(\sigma^n x))$ yields that $y\in \sigma^{-n}T_i(\sigma^n x, r)$. Therefore $$T_i(x, ar)\cap {\mathcal P}_0^{n-1}(x)\subset \sigma^{-n}T_i(\sigma^n x, r)$$ and we are done. \end{proof} Let $\theta(x)$ denote the smallest angle between the Oseledets subspaces, i.e. $$ \theta(x)=\min_{I\cap J=\emptyset} \measuredangle \Big( \oplus_{i\in I}E^i_{x},\; \oplus_{j\in J}E^j_{x}\Big). $$ We have the following. \begin{lem} \label{lem-4.2} For $x\in \Sigma'$, $i\in \{1,\ldots, s\}$ and $r>0$, $$T_{i}(x,r)\subset \xi_{i-1}(x)\cap {\Gamma_i}(x,r)\subset T_i(x, r/\sin\theta(x)).$$ \end{lem} \begin{proof} We first prove that $T_{i}(x,r)\subset \xi_{i-1}(x)\cap {\Gamma_i}(x,r)$. Let $y\in T_i(x,r)$. Then by definition, $y\in \xi_{i-1}(x)$ and $\pi y-\pi x=v+w$ for some $v\in V^i_x$, $w\in E^i_x$ with $\|w\|\leq r$, which implies that $$\mbox{dist}(\pi y+V_x^i, \; \pi x+V_x^i)\leq \|w\|\leq r.$$ Hence $y\in \xi_{i-1}(x)\cap {\Gamma_i}(x,r)$. This proves the relation $T_{i}(x,r)\subset \xi_{i-1}(x)\cap {\Gamma_i}(x,r)$. Next we prove that $\xi_{i-1}(x)\cap {\Gamma_i}(x,r)\subset T_i(x, r/\sin(\theta(x)))$. Let $U_x^i:=V^{i-1}_x\ominus V^i_x$ denote the orthogonal complement of $V_x^i$ in $V^{i-1}_x$. Let $z\in \xi_{i-1}(x)\cap {\Gamma_i}(x,r)$. Then $\pi z-\pi x\in V_x^{i-1}$ and $\mbox{dist}(\pi z+V_x^i, \pi x +V_x^i)\leq r$. Hence $\pi z-\pi x=v+u$ for some $v\in V_x^i$ and $u\in U_x^i$ with $\|u\|\leq r$. Since $v+u\in V_x^{i-1}=V_x^i\oplus E_x^i$, $v+u=v_1+w_1$ for some $v_1\in V^i_x$ and $w_1\in E_x^i$. Notice that $w_1= (v-v_1)+u$ with $u\perp (v-v_1)$. We have $$ \|w_1\|=\frac{\|u\|}{\sin \measuredangle(w_1, v-v_1)}\leq \frac{\|u\|}{\sin \theta(x)}\leq \frac{r}{\sin \theta(x)}. $$ Thus $\pi z-\pi x=v_1+w_1$, where $v_1\in V_x^i$ and $w_1\in E_x^i$ with $\|w_1\|\leq r/\sin \theta(x)$. Therefore, $z\in T_i(x, r/\sin \theta(x))$ and we are done. \end{proof} Now we turn back to the proof of Proposition \ref{pro-4.1}. Clearly, to prove the proposition it is sufficient to show that for any $\epsilon>0$, there exists $F(\epsilon)\subset \Sigma'$ so that \begin{equation} \label{e-import} \vartheta_{i-1}(x)\geq \frac{h_{i-1}(x)-h_{i}(x)}{-\lambda_{i}(x)+\epsilon}\quad \mbox{ for $m$-a.e.~$x\in F(\epsilon)$ and }i\in \{1,\ldots, s\}. \end{equation} and $\lim_{\epsilon\to 0}m(F(\epsilon))=1$. Here and afterwards in this section, we may assume that $\lambda_s\neq -\infty$ a.e.; since Proposition \ref{pro-4.1} holds automatically when $i=s$ and $\lambda_s(x)=-\infty$. We first construct $F(\epsilon)$ for $\epsilon>0$. Set \begin{equation} \label{e-t1} F_0(\epsilon):=\{x\in \Sigma':\;\sin \theta(x)>\epsilon\}. \end{equation} By \eqref{e-e?}, there exist a large integer $N(\epsilon)$ and a Borel set $F(\epsilon)\subset F_0(\epsilon)$ with $m(F(\epsilon))>(1-\epsilon)m(F_0(\epsilon))$ so that for $i\in \{1,\ldots, s\}$, \begin{equation}\label{e-t0} \|M_{x_0\cdots x_{n-1}}v\| \geq \epsilon^{-1} e^{n(\lambda_i(x)-\epsilon)}\|v\|\end{equation} for $x\in F(\epsilon)$, $n\geq N(\epsilon)$ and $v\in E_{\sigma^nx}^i$. Clearly, $m(F(\epsilon))\to 1$ as $\epsilon\to 0$. In the remaining part of this section we prove \eqref{e-import} for the constructed $F(\cdot)$. From now on, we fix $\epsilon>0$ and write simply $F=F(\epsilon)$ and $N=N(\epsilon)$. Let $\sigma_F: F\to F$ be the transformation induced by $\sigma^N$ on the set $F$ (cf. Section~\ref{Sect-2.3}). That is, $\sigma_F(x)=\sigma^{Nr_F(x)}(x)$, where $$ r_F(x):=\inf\{ n\geq 1: \; \sigma^{nN}x\in F\}. $$ The map $\sigma_F$ is well-defined on $F$ up to a set of zero $m$-measure. Let $m_F$ be the Borel probability measure on $F$ defined by $$m_F(D)=\frac{m(F\cap D)}{m(F)} \quad \mbox{ for any Borel set }D\subset F.$$ Recall that $m_F$ is $\sigma_F$-invariant. For $x\in F$, set \begin{equation} \label{e-rho} \begin{split} \ell (x)&=N r_F(x)\qquad \mbox{ and } \\ {\rho}(i, x)&=e^{\ell (x) (\lambda_i(x)-\epsilon)}, \qquad i=1,\ldots, s. \end{split} \end{equation} Then we have \begin{lem} For $x\in F$, $i\in \{1,\ldots, s\}$ and $r>0$, \begin{align} \xi_{i-1}(x)\cap {\Gamma_{i}}\left(x,\; {\rho}(i, x) r\right)&\cap {\mathcal P}_0^{\ell (x)-1}(x) \subset \sigma_F^{-1} \left({\Gamma_i}\left(\sigma_F x,r\right) \cap \xi_{i-1}(\sigma_Fx)\right). \label{e-t2} \end{align} \end{lem} \begin{proof} Fix $x\in F$, $i\in \{1,\ldots, s\}$ and $r>0$. Set $a=\epsilon^{-1}{\rho}(i, x)$. By \eqref{e-t0}, \begin{equation} \label{e-ttd} a=\epsilon^{-1}e^{\ell (x)(\lambda_i(x)-\epsilon)}\leq \inf\{\|M^{\ell (x)}(x)v\|:\; v\in E_{\sigma^nx}^i,\; \|v\|=1\}, \end{equation} where $M^n(x):=M_{x_0\cdots x_{n-1}}$. Observe that \begin{equation*} \begin{split} \xi_{i-1}&(x)\cap {\Gamma_{i}}\left(x,\; {\rho}(i, x) r\right)\\ & \subset T_i\left(x, {\rho}(i, x) r/\sin\theta(x)\right) \qquad (\mbox{by Lemma \ref{lem-4.2}})\\ & \subset T_i\left(x, \epsilon^{-1}{\rho}(i, x) r\right)\qquad\quad (\mbox{since $\sin \theta(x)\geq \epsilon$})\\ &= T_i\left(x, a r\right).\\ \end{split} \end{equation*} Hence \begin{equation*} \begin{split} \xi_{i-1}&(x)\cap {\Gamma_{i}}\left(x,\; {\rho}(i, x) r\right) \cap {\mathcal P}_0^{\ell (x)-1}(x)\\ & \subset T_i\left(x, a r\right)\cap {\mathcal P}_0^{\ell (x)-1}(x)\\ & \subset \sigma^{-\ell (x)} T_i\left(\sigma^{\ell (x)} x, r\right)\qquad\qquad (\mbox{by \eqref{e-ttd} and Lemma \ref{lem-5.2}})\\ &=\sigma_F^{-1} T_i(\sigma_F x, r)\\ &\subset\sigma_F^{-1} ({\Gamma_i}(\sigma_F x, r)\cap \xi_{i-1}(\sigma_Fx)) \qquad (\mbox{by Lemma \ref{lem-4.2}}). \end{split} \end{equation*} This completes the proof of the lemma. \end{proof} Now write \begin{equation} \label{e-l32} F_n:=\{x\in F:\; r_F(x)=n\},\qquad n=1,2,\ldots. \end{equation} Recall that $\{m_x^{\xi_i}\}$ is the canonical system of conditional measures associated with $\xi_i$, $i=0,\ldots, s$. The following result is an induced version of Lemma \ref{lem-2.4}. \begin{pro} \label{pro-4.5} Let $i\in \{1,\ldots, s\}$. Then for $m$-a.e.~$x\in F$, \begin{equation}\label{e-4.13} \begin{split} \lim_{r\to 0}\log \frac {m^{\xi_{i-1}}_x\left({\Gamma_i}(x,r)\cap {\mathcal P}_{0}^{\ell (x)-1}(x)\right)} {m^{\xi_{i-1}}_x\left({\Gamma_i}(x,r)\right)} &=-\sum_{k=1}^\infty \chi_{F_k}(x)\sum_{j=0}^{kN-1} {\bf I}_m({\mathcal P}|\widehat{\xi_i})(\sigma^j x). \end{split} \end{equation} Furthermore, set \begin{equation}\label{e-2.5''} g(x)=-\inf_{r>0} \log \frac {m^{\xi_{i-1}}_x\left({\Gamma_i}(x,r)\cap {\mathcal P}_{0}^{\ell (x)-1}(x)\right)} {m^{\xi_{i-1}}_x\left({\Gamma_i}(x,r)\right)}. \end{equation} Then $g\geq 0$ and $g\in L^1(F,{{\mathcal B}}|_F,m_F)$. \end{pro} \begin{proof} Fix $i\in \{1,\ldots, s\}$. Write $d_i=\sum_{j=i+1}^s{k_j}$. Define $\phi_i: \Sigma'\to Y_i:=G(d, d_i)\times {\Bbb R}^d$ by $$ \phi_i(x)=\left(V_x^i, \; P_{(V_x^i)^\perp}(\pi x)\right). $$ Then $\phi_i$ is measurable. Moreover, \begin{equation} \xi_i(x)=\{y\in \xi_0(x):\; \phi_i(y)=\phi_i(x)\},\quad x\in \Sigma'. \end{equation} Endow $Y_i$ with the following product metric $\rho_i$: $$ \rho_i\left((V, a), (W, b)\right)=\max\{\|P_V-P_W\|, \|a-b\|\}. $$ It is not hard to see that $Y_i$ is a Besicovitch space. For $x\in \Sigma'$ and $r>0$, set $$ B^{\phi_i}(x,r):=\{y\in \Sigma':\; \rho_i(\phi_iy, \phi_i x)\leq r\}. $$ Then by definition, \begin{equation} \label{e-phigamma} \xi_0(x)\cap B^{\phi_i}(x,r)=\xi_0(x)\cap \Gamma_i(x,r),\qquad x\in \Sigma',\; r>0. \end{equation} Hence for $x\in F$ and $r>0$, \begin{equation} \label{e-4.14} \begin{split} \frac {m^{\xi_{i-1}}_x\left({\Gamma_i}(x,r)\cap {\mathcal P}_{0}^{\ell (x)-1}(x)\right)} {m^{\xi_{i-1}}_x\left({\Gamma_i}(x,r)\right)}&=\frac {m^{\xi_{i-1}}_x\left(\xi_0(x)\cap {\Gamma_i}(x,r)\cap {\mathcal P}_{0}^{\ell (x)-1}(x)\right)} {m^{\xi_{i-1}}_x\left(\xi_0(x)\cap {\Gamma_i}(x,r)\right)}\\ &=\frac {m^{\xi_{i-1}}_x\left(\xi_0(x)\cap B^{\phi_i}(x,r)\cap {\mathcal P}_{0}^{\ell (x)-1}(x)\right)} {m^{\xi_{i-1}}_x\left(\xi_0(x)\cap B^{\phi_i}(x,r)\right)}\\ &=\frac {m^{\xi_{i-1}}_x\left(B^{\phi_i}(x,r)\cap {\mathcal P}_{0}^{\ell (x)-1}(x)\right)} {m^{\xi_{i-1}}_x\left(B^{\phi_i}(x,r)\right)}\\ &= \sum_{k=1}^\infty\sum_{A\in {\mathcal P}_{0}^{kN-1}} \chi_{ F_k\cap A}(x) \frac{m_x^{\xi_{i-1}}\left(B^{\phi_i}(x,r)\cap A\right)}{m_x^{\xi_{i-1}}\left(B^{\phi_i}(x,r)\right)}. \end{split} \end{equation} By \eqref{e-4.14} and applying Lemma \ref{lem-2.4}(1) to $\phi_i: \Sigma'\to Y_i$, we have for $m$-a.e.~$x\in F$, \begin{equation*} \begin{split} \lim_{r\to 0} & \log \frac {m^{\xi_{i-1}}_x\left({\Gamma_i}(x,r)\cap {\mathcal P}_{0}^{\ell (x)-1}(x)\right)} {m^{\xi_{i-1}}_x\left({\Gamma_i}(x,r)\right)}\\ & =\sum_{k=1}^\infty\sum_{A\in {\mathcal P}_{0}^{kN-1}} \chi_{A\cap F_k}(x)\log {\bf E}_m\left(\chi_A|\widehat{\xi_{i-1}} \vee \phi_{i}^{-1}{\mathcal B}(Y_i)\right)(x)\\ & =\sum_{k=1}^\infty\sum_{A\in {\mathcal P}_{0}^{kN-1}} \chi_{A\cap F_k}(x)\log {\bf E}_m\left(\chi_A|\widehat{\xi_{i}} \right)(x) \\ & =\sum_{k=1}^\infty \chi_{F_k}(x) \sum_{A\in {\mathcal P}_{0}^{kN-1}} \chi_{A}(x)\log {\bf E}_m\left(\chi_A|\widehat{\xi_{i}} \right)(x)\\ &=-\sum_{k=1}^\infty \chi_{F_k}(x) {\bf I}_m({\mathcal P}_0^{kN-1} |\widehat{\xi_i})(x)\\ &=-\sum_{k=1}^\infty \chi_{F_k}(x)\sum_{j=0}^{kN-1} {\bf I}_m({\mathcal P}|\widehat{\xi_i})(\sigma^j x) \qquad\mbox{(by \eqref{e-xi})}. \end{split} \end{equation*} This proves \eqref{e-4.13}. Next we prove that $g\in L^1(m_F)$. We mainly follow the arguments in \cite[Lemma 3.3 and Proposition 3.5]{FengHu2009}. By Theorem \ref{thm-2.1}, for any given $C\in \xi_{i-1}$, the conditional measures $m^{\xi_{i-1}}_x$ ($x\in C$) represent the same measure supported on $C$, which we rewrite as $m_C$. Fix $C\in \xi_{i-1}$, $k\in {\Bbb N}$ and $A\in {\mathcal P}_{0}^{kN-1}$. We define measures $\mu_C$ and $\nu_C$ on $Y_i$ by $\mu_C(E)=m_C(\phi_i^{-1}E\cap A)$ and $\nu_C(E)=m_C(\phi_i^{-1}E)$ for all $E\in {\mathcal B}(Y_i)$. By the Hardy-Littlewood maximal inequality (see, e.g.~Theorem 7.4 in \cite{Rudin1987}), there exists a positive constant $a$ (which depends on $Y_i$) such that $$ \mu_C\left\{z\in Y_i:\; \inf_{r>0}\frac{\mu_C(B(z,r))}{\nu_C(B(z,r))}<u \right\}\leq a u\qquad (u>0). $$ Hence for any $u>0$, $$ m_C\left(\ \left\{ x\in \Sigma':\; \inf_{r>0} \frac{m_C\left(B^{\phi_i}(x,r)\cap A\right)} {m_C\left(B^{\phi_i}(x,r)\right)}<u \right\} \cap A \right) \leq a u. $$ Integrating $C$ over $\xi_{i-1}$, we obtain $$ m\left(\ \left\{ x\in \Sigma':\; \inf_{r>0} \frac{m^{\xi_{i-1}}_x\left(B^{\phi_i}(x,r)\cap A\right)} {m^{\xi_{i-1}}_x\left(B^{\phi_i}(x,r)\right)}<u \right\} \cap A \right) \leq a u. $$ Write $\displaystyle g^A(x)=\inf_{r>0}\frac {m^{\xi_{i-1}}_x\left(B^{\phi_i}(x,r)\cap A\right)} {m^{\xi_{i-1}}_x\left(B^{\phi_i}(x,r) \right)}$. Then the above inequality can be rewritten as \begin{equation} \label{e-alambda} m(A\cap \{g^A<u\})\leq a u. \end{equation} Note that by (\ref{e-2.5''}) and \eqref{e-4.14}, $g(x)=-\sum_{k=1}^\infty \sum_{A\in {\mathcal P}_{0}^{kN-1} }\chi_{F_k\cap A} (x)\log g^A(x)$. Since $g$ is non-negative, \begin{eqnarray*} \int g\; dm &=&\int_0^\infty m\{g>t\}\;dt\\ &=& \int_0^\infty\sum_{k=1}^\infty \sum_{ A\in {\mathcal P}_{0}^{kN-1} } m(F_k\cap A\cap \{g^A<e^{-t}\})\; dt \\ &\leq& \sum_{k=1}^\infty \sum_{ A\in {\mathcal P}_{0}^{kN-1} } \int_0^\infty \min \{m(F_k\cap A), a e^{-t}\}\;dt \qquad\quad (\mbox{by \eqref{e-alambda}})\\ &\leq& \sum_{k=1}^\infty \sum_{ A\in {\mathcal P}_{0}^{kN-1} } \left(-m(F_k\cap A)\log m(F_k\cap A)+ m(F_k\cap A) (1+\log a) \right)\\ &\leq& 1+\log a + \sum_{k=1}^\infty \sum_{ A\in {\mathcal P}_{0}^{kN-1} } \left(-m(F_k\cap A)\log m(F_k\cap A) \right)\\ &\leq& 1+\log a+ \sum_{k=1}^\infty m(F_k) \left[ \left(\sum_{ A\in {\mathcal P}_{0}^{kN-1} } -\frac{m(F_k\cap A)}{m(F_k)} \log \frac{m(F_k\cap A)}{m(F_k)}\right)\right.\\ &\mbox{}&\qquad \qquad\qquad \qquad \qquad \qquad \qquad \left. + \log \frac{1}{m(F_k)} \right] \\ &\leq & 1+\log a+ \sum_{k=1}^\infty m(F_k) \left(kN\log (\#\Lambda) + \log \frac{1}{m(F_k)}\right)\\ &<&\infty \qquad \qquad \mbox{(by Lemma \ref{lem-induce}(ii)-(iii))}. \end{eqnarray*} This finishes the proof of the proposition. \end{proof} Finally we are ready to prove \eqref{e-import}, the last step in the proof of Proposition \ref{pro-4.1}. \begin{proof}[Proof of \eqref{e-import}] Fix $\epsilon>0$ and write $F=F(\epsilon)$. Let $i\in \{1,\ldots, s\}$. For $x\in F$ and $n\in {\Bbb N}$, define $$ {\rho}_n(i,x)=\prod_{k=0}^{n-1}{\rho}(i,\sigma_F^kx), $$ where $\sigma_F^k:=(\sigma_F)^k$, and ${\rho}(i,x)=e^{\ell (x)(\lambda_i(x)-\epsilon)}$ (as defined in \eqref{e-rho}). Moreover, write \begin{align*} H_n(x)&:=\log \frac{m_x^{\xi_{i-1}}\left({\Gamma_{i}}\left(x, {\rho}_n(i,x)\right) \right)} {m_{\sigma_Fx}^{\xi_{i-1}} \left( {\Gamma_{i}}\left(\sigma_F x, {\rho}_{n-1}(i,\sigma_F x)\right) \right)},\\ G_n(x)&:=\log \frac{m_x^{\xi_{i-1}}\left({\Gamma_{i}}\left(x, {\rho}_n(i,x)\right)\cap {\mathcal P}_0^{\ell (x)-1}(x) \right)} {m_{x}^{\xi_{i-1}} \left({\Gamma_{i}}\left( x, {\rho}_{n}(i, x)\right) \right)}. \end{align*} Then for $m$-a.e.~$x\in F$, \begin{align*} H_n(x)+G_n(x)&=\log \frac{m_x^{\xi_{i-1}}\left({\Gamma_{i}}\left(x, {\rho}_n(i,x)\right)\cap {\mathcal P}_0^{\ell (x)-1}(x) \right)} {m_{\sigma_Fx}^{\xi_{i-1}} \left( {\Gamma_{i}}\left(\sigma_F x, {\rho}_{n-1}(i,\sigma_F x)\right) \right)}\\ &=\log \frac{m_x^{\xi_{i-1}}\left(\xi_{i-1}(x)\cap {\Gamma_{i}}\left(x, {\rho}_n(i,x)\right)\cap {\mathcal P}_0^{\ell (x)-1}(x) \right)} {m_{\sigma_Fx}^{\xi_{i-1}} \left( {\Gamma_{i}}\left(\sigma_F x, {\rho}_{n-1}(i,\sigma_F x)\right) \right)}\\ &\leq \log \frac{m_x^{\xi_{i-1}}\left( \sigma_F^{-1} {\Gamma_{i}} \left(\sigma_F x, {\rho}_{n-1}(i,\sigma_F x)\right)\cap {\mathcal P}_0^{\ell (x)-1}(x) \right)} {m_{\sigma_Fx}^{\xi_{i-1}}\left( {\Gamma_{i}}\left(\sigma_F x, {\rho}_{n-1}(i,\sigma_F x)\right) \right)} \quad (\mbox{by \eqref{e-t2}})\\ &=\log m^{\xi_{i-1}}_x({\mathcal P}_0^{\ell (x)-1}(x)) \qquad \qquad \qquad \qquad (\mbox{by Lemma \ref{lem-3.6}(3)}) \\ &=-\sum_{k=1}^\infty \chi_{F_k}(x) \sum_{j=0}^{kN-1} {\bf I}_m({\mathcal P}|\widehat{\xi_{i-1}})(\sigma^j x)=:Q_{i-1}(x) \qquad (\mbox{by \eqref{e-h1}}), \end{align*} that is, $H_n(x)+G_n(x)\leq Q_{i-1}(x)$. Therefore for $m$-a.e.~$x\in F$, \begin{align*}-\log m_x^{\xi_{i-1}}\left({\Gamma_{i}}(x, {\rho}_n(i, x)) \right)&= -\left(\sum_{j=0}^{n-1}H_{n-j}(\sigma^j_Fx)\right)- \log m^{\xi_{i-1}}_{\sigma^n_F x}\left({\Gamma_{i}}(\sigma^n_F x, 1)\right)\\ &\geq -\sum_{j=0}^{n-1}H_{n-j}(\sigma^j_Fx)\\ &\geq \sum_{j=0}^{n-1}\left(G_{n-j}(\sigma^j_Fx)-Q_{i-1}(\sigma_F^jx)\right),\\ \end{align*} and thus \begin{eqnarray*}\label{e-5.8} \frac{-\log m_x^{\xi_{i-1}}\left(\Gamma_{i}(x, {\rho}_n(i,x)) \right)}{n}&\geq& \frac{1}{n}\sum_{j=0}^{n-1}\left(G_{n-j}(\sigma^j_F x)- Q_{i-1}(\sigma^j_Fx)\right). \end{eqnarray*} Notice that by Proposition \ref{pro-4.5}, when $n\to +\infty$, \begin{eqnarray*} G_n\to Q_{i}:=-\sum_{k=1}^\infty\chi_{F_k}\sum_{j=0}^{kN-1} {\bf I}_m({\mathcal P}|\widehat{\xi_{i}})\circ \sigma^j \end{eqnarray*} pointwise and in $L^1$. By Lemma \ref{lem-3.15}, for $m$-a.e.~$x\in F$, \begin{eqnarray*} \liminf_{n\to \infty}\frac{-\log m_x^{\xi_{i-1}}\left(\Gamma_{i}\left(x, {\rho}_n(i, x)\right) \right)}{n}&\geq& {\bf E}_{m_F}((Q_{i}-Q_{i-1})|{\mathcal I}_F)(x), \end{eqnarray*} where ${\mathcal I}_F:=\{B\in {\mathcal B}|_F:\; \sigma_F^{-1}(B)=B\}$. In the meantime, by the Birkhoff ergodic theorem, \begin{equation*} \begin{split} \lim_{n\to \infty} \frac{-1}{n}\log ({\rho}_n(i, x)) &= {\bf E}_{m_F}(N(-\lambda_i+\epsilon)r_F|{\mathcal I}_F)(x)\\ &=N(-\lambda_i(x)+\epsilon){\bf E}_{m_F}(r_F|{\mathcal I}_F)(x) \quad \mbox{$m_F$-a.e.}, \end{split} \end{equation*} where we use the fact that $\lambda_i$ is $\sigma$-invariant and thus $\sigma_F$-invariant. Hence for $m$-a.e.~$x\in F$, \begin{eqnarray*} \liminf_{r\to 0}\frac{\log m_x^{\xi_{i-1}}\left(\Gamma_{i}(x, r) \right)} {\log r} &=&\liminf_{n\to \infty}\frac{\log m_x^{\xi_{i-1}}\left(\Gamma_{i}\left(x, {\rho}_n(i, x)\right) \right)} {\log ({\rho}_n(i, x))}\\ &\geq & \frac{ {\bf E}_{m_F}((Q_{i}-Q_{i-1})|{\mathcal I}_F)(x)}{N(-\lambda_i(x)+\epsilon){\bf E}_{m_F}(r_F|{\mathcal I}_F)(x)}\\ &=& \frac{{\bf E}_m\left({\bf I}_m({\mathcal P}|\widehat{\xi_{i-1}})-{\bf I}_m({\mathcal P}|\widehat{\xi_{i}})\right)(x)}{-\lambda_i(x)+\epsilon}\quad\mbox{(by Lemma \ref{lem-inderg})}\\ &=& \frac{h_{i-1}(x)-h_{i}(x)}{-\lambda_i(x)+\epsilon}. \end{eqnarray*} That is, \eqref{e-import} holds. This completes the proof of Proposition \ref{pro-4.1}. \end{proof} \section{Local dimensions of invariant measures for affine IFSs} \label{S-6} In this section, we prove Theorems \ref{thm-1.0}-\ref{thm-1.2} and \ref{cor-1.0}-\ref{thm-1.7'}. Let $M:\; \Sigma\to {\Bbb R}^{d\times d}$ be the matrix-valued function defined by $$M(x)=M_{x_{-1}},\quad x=(x_j)_{j=-\infty}^{+\infty}.$$ Let $m\in {\mathcal M}_\sigma(\Sigma)$. Let $${\Bbb R}^d=\oplus_{i=1}^{s(x)}E_x^i\qquad (x\in \Sigma')$$ be the Oseledets splittings of ${\Bbb R}^d$ associated with $(\Sigma, \sigma^{-1},m)$ and $M$ (see Section~\ref{S-4}), and $0>\lambda_1(x)>\cdots>\lambda_{s(x)}(x)\geq -\infty$ the corresponding Lyapunov exponents. Below we prove parts (i) and (ii) of Theorem \ref{thm-1.0} separately. \begin{proof}[Proof of Theorem \ref{thm-1.0}(i)] In the beginning we assume that the condition \eqref{e-assump} holds, that is, for all $x\in \Sigma'$, $$s(x)=s, \quad \mbox{ and }\quad \dim E_x^i=k_i \mbox{ for }i=1, \ldots, s.$$ (Just keep in mind that we don't assume that $m$ is ergodic at this moment.) Write $\displaystyle V_x^i=\oplus_{j={i+1}}^s E_x^j$ for $i=0,\ldots, s-1$, and $V_x^s=\{0\}$. Clearly $$ \{0\}=V_x^{s}\subset V_x^{s-1}\subset\cdots\subset V_x^0={\Bbb R}^d.$$ Let $\xi_0,\xi_1,\ldots, \xi_s$ be the measurable partitions of $\Sigma'$ constructed as in Section~\ref{S-4}. Furthermore, we set \begin{equation} \label{e-conv} \xi_{-1} =\{\Sigma',\emptyset\}\qquad \mbox{and} \quad \lambda_{0}(x)=\lambda_1(x) \mbox{ for $x\in \Sigma'$}. \end{equation} Clearly $\xi_{-1}(x)=\Sigma'$ for any $x\in \Sigma'$. By Lemma \ref{lem-3.4}(2), we have \begin{equation} \label{e-t301} Q_{n,\epsilon}\cap \xi_i(x) \cap {\mathcal P}_0^n(x)\subset B^\pi(x, e^{n (\lambda_{i+1}(x)+\epsilon)}),\quad i=-1, 0,\ldots, s-1 \end{equation} when $n$ is large enough. Here $B^\pi(x,r)$ is defined as in \eqref{e-ball}. For $i=-1, 0,\ldots, s$, let $\{m_x^{\xi_i}\}$ be the canonical system of conditional measures associated with $\xi_i$. By the definition of $\xi_{-1}$, we see that $m_x^{\xi_{-1}}= m$ for any $x\in \Sigma'$. For $x\in \Sigma'$ and $i\in \{0,1,\ldots, s\}$, let $h_i(x)$ be defined as in \eqref{e-hi}. Write for convention that $$h_{-1}(x)=h_0(x).$$ According to Lemmas \ref{lem-3.5}-\ref{lem-4.6}, \begin{equation} \label{e-h} \lim_{n\to \infty} \frac{-\log m_x^{\xi_i}(Q_{n,\epsilon} \cap {\mathcal P}_0^n(x))}{n+1}=h_i(x) \quad \mbox{for $m$-a.e.$~x\in \Sigma'$},\qquad i=-1,0,\ldots, s. \end{equation} For $x\in \Sigma'$ and $r>0$, let $\Gamma_i(x,r)$ be defined as in \eqref{gamma}, that is, $$ {\Gamma_i}(x,r)=\{y\in \Sigma':\; \mbox{dist}(\pi y+V_x^i, \; \pi x+V_x^i)\leq r\}, \quad i=1,\ldots, s. $$ Write for convention that $$ {\Gamma_0}(x,r)=\Sigma'. $$ It is easy to see that for $i=0,1, \ldots, s$, \begin{equation} \label{e-5.4} {\Gamma_i}(x,r)=\{y\in \Sigma':\; \|P_{({V_x^i})^\perp}(\pi y-\pi x)\|\leq r\}, \end{equation} where $({V_x^i})^\perp$ stands for the orthogonal complement of the space $V_x^i$ in ${\Bbb R}^d$, and $P_{W}$ is the orthogonal projection from ${\Bbb R}^d$ to $W$. Moreover, define \begin{equation} \vartheta_i(x)=\liminf_{r\to 0}\frac{\log m^{\xi_i}_x({\Gamma_{i+1}}(x,r))}{\log r}, \qquad i=-1,0,\ldots, s-1. \end{equation} Clearly $\vartheta_{-1}(x)=0$ for every $x\in \Sigma'$ since ${\Gamma_0}(x,r)=\Sigma'$. Combining this with Proposition \ref{pro-4.1} yields \begin{equation} \label{e-h'} \vartheta_i(x)\geq \frac{h_{i+1}(x)-h_{i}(x)}{\lambda_{i+1}(x)}\qquad (i=-1, 0, \ldots, s-1) \end{equation} for $m$-a.e.~$x\in \Sigma'$. For $i=-1, 0,\ldots, s$ and $x\in \Sigma'$, define $$ \overline{\delta}_i(x)=\limsup_{r\to 0}\frac{\log m^{\xi_i}_x(B^\pi(x,r))}{\log r},\quad \underline{\delta}_i(x)=\liminf_{r\to 0}\frac{\log m^{\xi_i}_x(B^\pi(x,r))}{\log r}. $$ We claim that for $m$-a.e.~$x\in \Sigma'$, \begin{itemize} \item[(C1)] $\overline{\delta}_{s}(x)=\underline{\delta}_{s}(x)=0$. \item[(C2)] $\displaystyle \frac{h_{i+1}(x)-h_{i}(x)}{\lambda_{i+1}(x)}\geq \overline{\delta}_i(x)-\overline{\delta}_{i+1}(x)$ for $i=-1, 0, \ldots, s-1$. \item[(C3)] $\underline{\delta}_{i+1}(x)+\vartheta_i(x)\leq \underline{\delta}_i(x)$ for $i=-1,0,\ldots, s-1$. \end{itemize} It is easy to see that (C1)-(C3) together with (\ref{e-h'}) force inductively that for $m$-a.e.~$x\in \Sigma'$, \begin{eqnarray} \vartheta_i(x)&=& \frac{h_{i+1}(x)-h_{i}(x)}{\lambda_{i+1}(x)} \quad \mbox{ for } i=s-1, \ldots, 0, -1,\label{e-h'1}\\ \underline{\delta}_i(x)&=&\overline{\delta}_i(x) \quad \mbox{ for } i=s, s-1,\ldots,0,-1 \nonumber \end{eqnarray} (we write the common value as $\delta_i(x)$), and furthermore \begin{equation} \label{e-key11} \delta_{i}(x)=\sum_{j=i}^{s-1} \vartheta_j(x)= \sum_{j=i}^{s-1} \frac{h_{j+1}(x)-h_{j}(x) }{\lambda_{j+1}(x)} \end{equation} for $i=-1, 0,\ldots,s$. In particular, \begin{equation} \label{e-key1} \dim_{\rm loc}(m\circ\pi^{-1},\pi x)=\delta_{-1}(x)=\delta_0(x)=\sum_{i=0}^{s-1} \vartheta_i(x)= \sum_{i=0}^{s-1} \frac{h_{i+1}(x)-h_{i}(x) }{\lambda_{i+1}(x)} \end{equation} for $m$-a.e.~$x$, which proves Theorem \ref{thm-1.0}(i) under the additional assumption \eqref{e-assump}. In the following we prove (C1)-(C3) respectively. \bigskip \begin{proof}[Proof of (C1)] Since $\xi_s(x)=\pi^{-1}(\pi x)\cap \xi_0(x)\subset B^\pi(x,r)$ for any $x\in \Sigma'$ and $r>0$, we have $$m^{\xi_s}_x(B^\pi(x,r))=m^{\xi_s}_x(\xi_s(x))=1$$ for all $x\in \Sigma'$. Thus $\overline{\delta}_s(x)=\underline{\delta}_s(x)=0$ for all $x\in \Sigma'$. \end{proof} \bigskip \begin{proof}[Proof of (C2)] We give a proof by contradiction, which is modified from \cite[\S10.2]{LedrappierYoung1985} and the proof of \cite[Theorem 2.11]{FengHu2009}. Assume that (C2) is not true. Then there exists $i\in \{-1,0,\ldots, s-1\}$ such that $$ \frac{h_{i+1}(x)-h_i(x)}{\lambda_{i+1}(x)}<\overline{\delta}_i(x)-\overline{\delta}_{i+1}(x) $$ on a set $U=U_i\subset \Sigma'$ with positive measure. Fix such $i$. Removing a suitable subset from $U$ if necessary, we may assume that one of the following holds: (a) $\lambda_{i+1}(x)\neq -\infty$ for all $x\in U$; or (b) $\lambda_{i+1}(x)=-\infty$ and $\overline{\delta}_i(x)>\overline{\delta}_{i+1}(x)$ for all $x\in U$. Notice that (b) can not occur unless $i=s-1$, since $\lambda_{i+1}(x)\neq -\infty$ for $i<s-1$. Now we first assume that the scenario (a) occurs. Then there exist $\alpha>0$ and real numbers $h_i,h_{i+1},\lambda_{i+1}, \overline{\delta}_i,\overline{\delta}_{i+1}$ with $\lambda_{i+1}<0$ such that \begin{equation} \label{e-ass} \frac{h_{i+1}-h_i}{\lambda_{i+1}}< \overline{\delta}_i-\overline{\delta}_{i+1}-\alpha \end{equation} and for any $\epsilon>0$, there exists $B_\epsilon\subset U$ with $m(B_\epsilon)>0$ so that for $x\in B_\epsilon$, $$|h_i(x)-h_i|<\epsilon/2, \quad |h_{i+1}(x)-h_{i+1}|<\epsilon/2, \quad \lambda_{i+1}(x)<\lambda_{i+1}+\epsilon/2$$ and $$\overline{\delta}_i(x)\geq \overline{\delta}_i-\epsilon/2, \quad \overline{\delta}_{i+1}(x)<\overline{\delta}_{i+1}+\epsilon/2.$$ Fix $\epsilon\in (0, -\lambda_{i+1}/3)$. There exists $n_0\colon B_\epsilon\to {\Bbb N}$ such that for $m$-a.e.~$x\in B_\epsilon$ and $n>n_0(x)$, we have \begin{itemize} \item[(1)] $\displaystyle \frac{\log m^{\xi_{i+1}}_x\left(B^\pi(x,e^{n(\lambda_{i+1}+2\epsilon)})\right)} {n(\lambda_{i+1}+2\epsilon)}< \overline{\delta}_{i+1}+\epsilon;$ \item[(2)] $\displaystyle -\frac{1}{n} \log m^{\xi_{i+1}}_x({\mathcal P}_{0}^n(x))> h_{i+1}-\epsilon$\qquad (by (\ref{e-h1})); \item[(3)] $\displaystyle Q_{n,\epsilon} \cap \xi_i(x)\cap {\mathcal P}_0^n(x)\subset B^\pi(x,e^{n(\lambda_{i+1}+2\epsilon)})$ \qquad (by (\ref{e-t301})); \item[(4)] $\displaystyle -\frac{1}{n}\log m^{\xi_i}_x(Q_{n,\epsilon} \cap {\mathcal P}_0^n(x))< h_i+\epsilon$\qquad (by (\ref{e-h})). \end{itemize} Take $N_0$ such that $$ \Delta:=\{x\in B_\epsilon\colon n_0(x)\leq N_0\}$$ has the positive measure. By Lemma \ref{lem-2.4}(1) and Lemma \ref{lem-2.10}, there exist $c>0$ and $\Delta'\subset \Delta$ with $m(\Delta')>0$ such that for $x\in \Delta'$, there exists $n=n(x)\geq N_0$ such that \begin{itemize} \item[(5)] $\displaystyle \frac{m^{\xi_{i+1}}_x (L\cap \Delta)}{m^{\xi_{i+1}}_x(L)}> c$, where $$ L:=B^\pi(x,e^{n(\lambda_{i+1}+2\epsilon)}); $$ \item[(6)] $\displaystyle \frac{\log m^{\xi_i}_x\left(B^\pi(x,2e^{n(\lambda_{i+1}+2\epsilon)})\right)} {n(\lambda_{i+1}+2\epsilon)}> \overline{\delta}_{i}-\epsilon;$ \item[(7)] $\displaystyle\frac{\log (1/c)}{n}<\epsilon$. \end{itemize} Take $x\in \Delta'$ such that (1)--(7) are satisfied with $n=n(x)$. Write $C=\xi_{i+1}(x)$ and $C'=\xi_i(x)$. Then by (5) and (1), $$m_x^{\xi_{i+1}}(L\cap \Delta )\geq cm_x^{\xi_{i+1}}(L)\geq ce^{n(\lambda_{i+1}+2\epsilon)(\overline{\delta}_{i+1}+\epsilon)}.$$ But for each $y\in L\cap \Delta$, by (2), $m_y^{\xi_{i+1}}({\mathcal P}_0^n(y))< e^{-n(h_{i+1}-\epsilon)}$. It follows that the number of distinct ${\mathcal P}_0^n$-atoms intersecting $C\cap L\cap \Delta$ is larger than $$m_x^{\xi_{i+1}}(L\cap \Delta) e^{n(h_{i+1}-\epsilon)}.$$ However each such a ${\mathcal P}_0^n$-atom, say ${\mathcal P}_0^n(y)$, intersects $C'\cap L\cap \Delta$. This implies that $Q_{n,\epsilon} \cap C'\cap {\mathcal P}_0^n(y)$ is contained in $C'\cap B^\pi(x,2e^{n(\lambda_{i+1}+2\epsilon)})$. To see this implication, let $z\in {\mathcal P}_0^n(y)\cap C^\prime\cap L\cap \Delta$; since $z\in L\cap \Delta$, we have $d(\pi z, \pi x)\leq e^{n(\lambda_{i+1}+2\epsilon)}$ and thus \begin{align*} Q_{n,\epsilon}\cap C'\cap{\mathcal P}_0^n(y)&=Q_{n,\epsilon}\cap \xi_i(z)\cap {\mathcal P}_0^n(z)\\ &\subset B^\pi(z,e^{n(\lambda_{i+1}+2\epsilon)})\quad (\mbox{by (3)})\\ &\subset B^\pi(x,2e^{n(\lambda_{i+1}+2\epsilon)}), \end{align*} so $Q_{n,\epsilon}\cap C'\cap{\mathcal P}_0^n(y)\subset C'\cap B^\pi(x,2e^{n(\lambda_{i+1}+2\epsilon)})$, as desired. In the meantime, by (4), $m^{\xi_i}_x(Q_{n,\epsilon}\cap {\mathcal P}_0^n(y))\geq e^{-n(h_i+\epsilon)}$. (To see it, picking $w\in {\mathcal P}_0^n(y)\cap C'\cap L\cap \Delta$, we have $\xi_i(x)=\xi_i(w)$ and thus $m^{\xi_i}_x(Q_{n,\epsilon}\cap {\mathcal P}_0^n(y))=m_w^{\xi_i}(Q_{n,\epsilon}\cap {\mathcal P}_0^n(w))\geq e^{-n(h_i+\epsilon)}$.) Hence \begin{eqnarray*} m^{\xi_i}_{x}(B^\pi(x,2e^{n(\lambda_{i+1}+2\epsilon)})) &\geq & \# \{ \mbox{${\mathcal P}_0^n$-atoms intersecting } C'\cap L\cap\Delta\} \cdot e^{-n(h_i+\epsilon)}\\ &\geq & m_x^{\xi_{i+1}}(L\cap \Delta) e^{n(h_{i+1}-\epsilon)}e^{-n(h_i+\epsilon)}\\ &\geq & ce^{n(\lambda_{i+1}+2\epsilon) (\overline{\delta}_{i+1}+\epsilon)}e^{n(h_{i+1}-\epsilon)}e^{-n(h_i+\epsilon)}. \end{eqnarray*} Combining the above inequality with (6) yields \begin{equation} \label{e-eq1} \begin{split} \mbox{}&(\lambda_{i+1}+2\epsilon)(\overline{\delta}_{i}-\epsilon)\\ & \quad \geq (\lambda_{i+1}+2\epsilon)(\overline{\delta}_{i+1}+\epsilon)+\frac{\log c}{n}+h_{i+1}-h_{i}-2\epsilon\\ & \quad \geq (\lambda_{i+1}+2\epsilon)(\overline{\delta}_{i+1}+\epsilon)+h_{i+1}-h_{i}-3\epsilon. \end{split} \end{equation} Taking $\epsilon\to 0$ yields $h_{i+1}-h_i\leq \lambda_{i+1} (\overline{\delta}_i-\overline{\delta}_{i+1})$, which leads to a contradiction with (\ref{e-ass}) (keep in mind that $\lambda_{i+1}<0$). Next we assume that the scenario (b) occurs, that is, $\lambda_{i+1}(x)=-\infty$ and $\overline{\delta}_i(x)>\overline{\delta}_{i+1}(x)$ for all $x\in U$. In this case, $i=s-1$, and thus by (C1), $\overline{\delta}_{i+1}(x)=\overline{\delta}_{i+1}:=0$ for all $x\in U$. Hence there exist real numbers $h_i, h_{i+1}, \overline{\delta}_i$ with $ \overline{\delta}_i>0$, so that for any $\epsilon>0$, there exists $B_\epsilon\subset U$ with $m(B_\epsilon)>0$ such that for $x\in B_\epsilon$, $$|h_i(x)-h_i|<\epsilon/2, \quad |h_{i+1}(x)-h_{i+1}|<\epsilon/2, \quad \overline{\delta}_i(x)\geq \overline{\delta}_i-\epsilon/2.$$ Set $\lambda_{i+1}=(-1/\epsilon)-2\epsilon$. Then an argument similar to that for the scenario (a) shows that the previous estimates (1)-(7) hold, and moreover, the inequality \eqref{e-eq1} still holds. Taking $\epsilon\to 0$ gives $\overline{\delta}_i\leq \overline{\delta}_{i+1}=0$, leads to a contradiction with $\overline{\delta}_i>0$. \end{proof} \bigskip \begin{proof}[Proof of (C3)] Here we give a proof by contradiction, following the lines of the proof of \cite[Theorem 2.11]{FengHu2009}, in which the arguments were adapted from the original proof of \cite[Lemma 11.3.1]{LedrappierYoung1985}. Assume that (C3) is not true. Then there exists $i\in \{-1,0,\ldots, s-1\}$ such that $ \underline{\delta}_{i+1}(x)+\vartheta_i(x)>\underline{\delta}_i(x) $ on a subset of $\Sigma'$ with positive measure. Hence there exist $\beta>0$ and real numbers $\underline{\delta}_i,\underline{\delta}_{i+1}, \vartheta_i$ such that \begin{equation} \label{e-as1} \underline{\delta}_{i+1}+\vartheta_i>\underline{\delta}_i+\beta, \end{equation} and for any $\epsilon>0$, there exists $A_\epsilon\subset \Sigma'$ with $m(A_\epsilon)>0$ so that for $x\in A_\epsilon$, \begin{equation}\label{e-as2} |\underline{\delta}_i(x)-\underline{\delta}_i|<\epsilon/2, \quad |\underline{\delta}_{i+1}(x)-\underline{\delta}_{i+1}|<\epsilon/2, \quad |\vartheta_i(x)-\vartheta_{i}|<\epsilon/2. \end{equation} Let $0<\epsilon<\beta/4$. Find $N_1$ and a set $A_\epsilon'\subset A_\epsilon$ with $m(A_\epsilon')>0$ such that \begin{equation} \label{e-6.2} m_x^{\xi_{i+1}}\left(B^\pi(x,2e^{-n})\right)\leq e^{-n(\underline{\delta}_{i+1}-\epsilon)}\qquad \mbox{ for }\; x\in A_\epsilon' \mbox{ and }n>N_1. \end{equation} By Lemma \ref{lem-2.4}(1) and Lemma \ref{lem-2.10}, we can find $c>0$ and $A_\epsilon''\subset A_\epsilon'$ with $m(A_\epsilon'')>0$ and $N_2>N_1$ such that for all $x\in A_\epsilon''$ and $n\geq N_2$, $$ \frac{m_x^{\xi_i}(A_\epsilon'\cap B^\pi(x,e^{-n}))}{m_x^{\xi_i}(B^\pi(x,e^{-n}))}>c. $$ For $x\in A_\epsilon''$ and $n\geq N_2$, we have \begin{equation} \label{e-6.3} \begin{split} m_x^{\xi_i}(B^\pi(x,e^{-n}))&\leq c^{-1}m_x^{\xi_i}(A_\epsilon'\cap B^\pi(x,e^{-n}))\\ &= c^{-1}\int m_y^{\xi_{i+1}}(A_\epsilon'\cap B^\pi(x,e^{-n}))\; dm^{\xi_i}_x(y)\\ &= c^{-1}\int_{{\Gamma_{i+1}}(x,e^{-n})} m_y^{\xi_{i+1}}(A_\epsilon'\cap B^\pi(x,e^{-n}))\; dm^{\xi_i}_x(y),\\ \end{split} \end{equation} where in the last equality, we use the fact that $y\in {\Gamma_{i+1}}(x,e^{-n})$, if $y\in \xi_i(x)$ and $\xi_{i+1}(y)\cap A_\epsilon'\cap B^\pi(x,e^{-n})\neq \emptyset$. To see this fact, let $y\in \xi_i(x)$ such that $\xi_{i+1}(y)\cap A_\epsilon'\cap B^\pi(x,e^{-n})\neq \emptyset$. Take $w\in \xi_{i+1}(y)\cap A_\epsilon'\cap B^\pi(x,e^{-n})$. Then $\pi w-\pi y\in V_{y}^{i+1}$, $\|\pi w-\pi x\|\leq e^{-n}$ and $w^-=y^{-}=x^{-}$ which implies $V_{w}^{i+1}=V_{y}^{i+1}=V_{x}^{i+1}$. Hence $$ \mbox{dist}(\pi y+V_x^{i+1}, \pi x+V_{x}^{i+1})= \mbox{dist}(\pi w+V_x^{i+1}, \pi x+V_{x}^{i+1})\leq \|\pi w-\pi x\|\leq e^{-n}, $$ and thus $y\in \Gamma_{i+1}(x,e^{-n})$. This completes the proof of the fact. In the above argument, since $\|\pi w-\pi x\|\leq e^{-n}$, we have $A_\epsilon'\cap B^\pi(x,e^{-n}) \subset B^\pi(w,2e^{-n})$ and thus \begin{eqnarray*} m^{\xi_{i+1}}_y(A_\epsilon'\cap B^\pi(x,e^{-n}))&=&m^{\xi_{i+1}}_w(A_\epsilon'\cap B^\pi(x,e^{-n}))\\ &\leq& m^{\xi_{i+1}}_w(B^\pi(w,2e^{-n}))\\ &\leq& e^{-n(\underline{\delta}_{i+1}-\epsilon)} \qquad \mbox{(by \eqref{e-6.2})}. \end{eqnarray*} Combining the above inequality with (\ref{e-6.3}) yields $$ m_x^{\xi_i}(B^\pi(x,e^{-n})) \leq c^{-1}e^{-n(\underline{\delta}_{i+1}-\epsilon)} m^{\xi_i}_x({\Gamma_{i+1}}(x,e^{-n}))\qquad (x\in A_\epsilon'',\; n\geq N_2). $$ Letting $n\to \infty$, we obtain $\underline{\delta}_i(x)\geq \underline{\delta}_{i+1}-\epsilon+\vartheta_i(x)$ for $x\in A_\epsilon''$. Combining this with (\ref{e-as2}) yields $$ \underline{\delta}_i\geq \underline{\delta}_{i+1}+\vartheta_i-4\epsilon\geq \underline{\delta}_{i+1}+\vartheta_i-\beta, $$ which contradicts (\ref{e-as1}). \end{proof} So far we have proved Theorem \ref{thm-1.0}(i) under the additional assumption \eqref{e-assump}. Now we consider the general case that the integer functions $s(x)$ and $\dim V_x^i$, $1\leq i\leq s(x)$, may not be constant over $\Sigma'$. In such case, by Theorem \ref{thm-2} there exists a finite Borel partition $$\Sigma'=\bigsqcup_{j=1}^k \Sigma_j$$ of $\Sigma'$ so that for each $j$, $\Sigma_j$ is $\sigma$-invariant, and $s(x)$ and $\dim V_x^i$ are constant restricted on $\Sigma_j$. Ignore those indices $j$ with $m(\Sigma_j)=0$. We define probability measures $m_j$ by $$m_j=\frac{m|_{\Sigma_j}}{m(\Sigma_j)}.$$ Then $m_j\in {\mathcal M}_\sigma(\Sigma_j)$. Since now \eqref{e-assump} holds for $m_j$ (in which $\Sigma'$ is replaced by $\Sigma_j$), we see that \eqref{e-key1} holds when replacing $m$ by $m_j$. In particular, the local dimension $\dim_{\rm loc}(m_j\circ {\pi}^{-1}, \pi x)$ exists for $m_j$-a.e.~$x\in \Sigma_j$. Equivalently, \begin{equation} \label{e-end1} \lim_{r\to 0}\frac{\log m(\Sigma_j\cap B^\pi(x,r))}{\log r} \quad \mbox{ exists for $m$-a.e.~$x\in \Sigma_j$}. \end{equation} By Lemma \ref{lem-2.4}(1) and Lemma \ref{lem-2.10}, for $m$-a.e.~$x\in \Sigma_j$, the following $$ \lim_{r\to 0} \frac{m(\Sigma_j\cap B^\pi(x,r))}{m(B^\pi(x,r))} $$ exists and takes positive value. This together with \eqref{e-end1} yields that the local dimension $\dim_{\rm loc} (m\circ\pi^{-1},\pi x)$ exists for $m$-a.e.~$x\in \Sigma_j$. Since $j$ is arbitrarily taken, $\dim_{\rm loc} (m\circ\pi^{-1},\pi x)$ exists for $m$-a.e.~$x\in \Sigma'$. This completes the proof of Theorem \ref{thm-1.0}(i). \end{proof} \begin{proof}[Proof of Theorems \ref{thm-1.0}(ii) and \ref{thm-1.1}] Since now $m$ is assumed to be ergodic, the condition \eqref{e-assump} holds and the functions $\lambda_i(x)$, $h_i(x)$ ($i=-1,\ldots, s$) considered in the proof of Theorem \ref{thm-1.0}(i) are all constant, which we denote by $\lambda_i$, $h_i$ respectively. The formula \eqref{ly-dim} just follows from \eqref{e-key1}. \end{proof} \begin{proof}[Proof of Theorem \ref{thm-1.2}] It is based on the proof of Theorem \ref{thm-1.0}. To see \eqref{e-con-dim}, let $i\in \{0,\ldots, s-1\}$. By \eqref{e-key11}, for $m$-a.e.~$x\in \Sigma'$, \begin{equation*} \dim_{\rm loc}(m_x^{\xi_i}\circ \pi^{-1}, \pi x)=\delta_i=\sum_{k=i}^{s-1}\frac{h_{k+1}-h_{k}}{\lambda_{k+1}}. \end{equation*} Equivalently, for $m$-a.e.~$x\in \Sigma'$ and $m^{\xi_i}_x$-a.e.~$y\in \xi_i(x)$, \begin{equation*} \dim_{\rm loc}(m_x^{\xi_i}\circ \pi^{-1}, \pi y)=\delta_i=\sum_{k=i}^{s-1}\frac{h_{k+1}-h_{k}}{\lambda_{k+1}}. \end{equation*} Hence for $m$-a.e.~$x\in \Sigma'$, $m_x^{\xi_i}\circ \pi^{-1}$ is exact dimensional with dimension given by \eqref{e-con-dim}. Next we prove \eqref{e-con-proj-dim} and \eqref{e-proj-dim}. By \eqref{e-h'1}, for $m$-a.e.~$x\in \Sigma'$, \begin{equation} \label{e-h''} \vartheta_i(x)= \vartheta_i:= \frac{h_{i+1}-h_{i}}{\lambda_{i+1}} \quad \mbox{ for } i=-1, 0, \ldots, s-1. \end{equation} Let $\Gamma_i(x,r)$ ($x\in \Sigma')$, $0\leq i\leq s$, be defined as in \eqref{e-5.4}. Fix $j\in \{1,\ldots, s\}$. For $i=-1,0,\ldots, j$ and $x\in \Sigma'$, define $$ \overline{\gamma}_{i, j}(x)=\limsup_{r\to 0}\frac{\log m^{\xi_i}_x({\Gamma_j}(x,r))}{\log r},\quad \underline{\gamma}_{i,j}(x)=\liminf_{r\to 0}\frac{\log m^{\xi_i}_x({\Gamma_j}(x,r))}{\log r}. $$ We claim that \begin{equation} \label{e-claim1} \xi_i(x)\cap {\Gamma_j}(x,r)= \xi_i(x)\cap g^{-1}(B(gx, r)), \end{equation} where $g:\; \xi_i(x)\to {(V_x^j)}^\perp$ is defined by $y\mapsto P_{{(V_x^j)}^\perp}(\pi y)$. To see this, let $y\in \xi_i(x)\cap {\Gamma_j}(x,r)$. Then $\mbox{dist} (\pi y+V_x^j, \pi x+ V_x^j)\leq r$, equivalently, $\| gy-gx\|\leq r$; hence $y\in g^{-1}(B(gx, r))$. This proves the direction $\xi_i(x)\cap {\Gamma_j}(x,r)\subset \xi_i(x)\cap g^{-1}(B(gx, r))$. The other direction can be proved similarly. This completes the proof of \eqref{e-claim1}. Now due to \eqref{e-claim1}, we have $m^{\xi_i}_x({\Gamma_j}(x,r))=m^{\xi_i}_x(g^{-1}(B(gx, r)))$, and so \begin{equation} \label{e-fact} \begin{split} \overline{\gamma}_{i,j}(x)&=\overline{\dim}_{\rm loc}\left(m^{\xi_i}_x\circ \pi^{-1}\circ (P_{{(V_x^j)}^\perp})^{-1}, \; P_{{(V_x^j)}^\perp} (\pi x)\right), \\ \underline{\gamma}_{i,j}(x)&=\underline{\dim}_{\rm loc}\left(m^{\xi_i}_x\circ \pi^{-1}\circ (P_{{(V_x^j)}^\perp})^{-1}, \; P_{{(V_x^j)}^\perp} (\pi x)\right). \end{split} \end{equation} We claim that for $m$-a.e.~$x\in \Sigma'$, the following properties hold: \begin{itemize} \item[(D1)] $\overline{\gamma}_{j,j}(x)=\underline{\gamma}_{j,j}(x)=0$. \item[(D2)] $h_{i}-h_{i+1}\geq -\lambda_{i+1} (\overline{\gamma}_{i,j}(x)-\overline{\gamma}_{i+1,j}(x))$ for $i=-1, 0, \ldots, j-1$. \item[(D3)] $\underline{\gamma}_{i+1,j}(x)+\vartheta_i\leq \underline{\gamma}_{i,j}(x)$ for $i=-1,0,\ldots, j-1$. \end{itemize} Clearly (D1)-(D3) together with (\ref{e-h''}) force that for $m$-a.e.~$x\in \Sigma'$, $$\underline{\gamma}_{i,j}(x)=\overline{\gamma}_{i,j}(x) \quad \mbox{ for $i=j, j-1,\ldots, 0, -1$},$$ (we write the common value as $\gamma_i(x)$), and furthermore \begin{align} \gamma_{-1,j}(x)&=\sum_{k=0}^{j-1} \vartheta_k= \sum_{k=0}^{j-1} \frac{h_{k+1}-h_{k} }{\lambda_{k+1}} \quad \mbox{ and } \label{e-key1*} \\ \gamma_{i,j}(x)&=\sum_{k=i}^{j-1} \frac{h_{k+1}-h_{k} }{\lambda_{k+1}} \quad \mbox{ for } i\in \{0,1,\ldots, j-1\}. \label{e-key1_*} \end{align} Now \eqref{e-proj-dim} just follows from \eqref{e-key1*} and the fact \eqref{e-fact}. To see \eqref{e-con-proj-dim}, let $i\in \{0,\ldots, j-1\}$. By \eqref{e-key1_*} and \eqref{e-fact}, we have for $m$-a.e.~$x\in \Sigma'$ and $m_x^{\xi_i}$-a.e.~$y\in \xi_i(x)$, $$ \dim_{\rm loc}\left(m^{\xi_i}_x\circ \pi^{-1}\circ (P_{{(V_x^j)}^\perp})^{-1}, \; P_{{(V_x^j)}^\perp} (\pi y)\right)=\gamma_{i,j}(x)=\sum_{k=i}^{j-1} \frac{h_{k+1}-h_{k} }{\lambda_{k+1}}, $$ where we use the fact that $V_y^i=V_x^i$ for $y\in \xi_i(x)$, due to $y\in \xi_0(x)$ (see Lemma \ref{lem-3.1}). As a consequence, for $m$-a.e.~$x\in \Sigma'$, $m^{\xi_i}_x\circ \pi^{-1}\circ (P_{{(V_x^j)}^\perp})^{-1}$ is exact dimensional and \eqref{e-con-proj-dim} holds. To complete the proof of Theorem \ref{thm-1.2}, in the following we prove (D1)-(D3) respectively. \medskip By the definition of $\xi_j$, for $x\in \Sigma'$ and $y\in \xi_j(x)$, we have $\pi y-\pi x\in V_x^j$ and thus $\pi y +V_x^j=\pi x +V_x^j$. It follows that $y\in {\Gamma_j}(x,r)$. Hence $\xi_j(x)\subset \Gamma_j(x,r)$ and thus $m^j_x({\Gamma_j}(x,r))=1$ for $x\in \Sigma'$ and any $r> 0$. Hence $\overline{\gamma}_{j,j}(x)=\underline{\gamma}_{j,j}(x)=0$ for all $x\in \Sigma'$. This proves (D1). The proofs of (D2) and (D3) are almost identical to that of (C2) and (C3), respectively. Indeed we only need to modify the proofs of (C2) and (C3) slightly. More precisely, among other minor adjustments, we may simply replace the terms $\delta_i$, $\delta_{i+1}$, $B^\pi(x, e^{n(\lambda_{i+1}+2\epsilon)})$, $B^\pi(x, 2e^{n(\lambda_{i+1}+2\epsilon)})$, $B^\pi(x, e^{-n})$ therein by $\gamma_{i,j}$, $\gamma_{i+1,j}$, ${\Gamma_j}(x, e^{n(\lambda_{i+1}+2\epsilon)})$, ${\Gamma_j}(x, 2e^{n(\lambda_{i+1}+2\epsilon)})$, and ${\Gamma_j}(x, e^{-n})$ respectively. This completes the proof of Theorem \ref{thm-1.2}. \end{proof} As a corollary of Theorem \ref{thm-1.2}, we have \begin{cor} \label{cor-6.1} Under the assumptions of Theorem \ref{thm-1.1}, for $i\in \{0,\ldots, s-1\}$ and $m$-a.e.~$x\in \Sigma'$, \begin{equation} \label{e-theta_i} \vartheta_{i}(x)=\lim_{r\to 0}\frac{\log m_{x}^{\xi_{i}}(\Gamma_{i+1}(x,r))}{\log r}=\frac{h_{i+1}-{h_{i}}}{\lambda_{i+1}}\leq k_{i+1}. \end{equation} \end{cor} \begin{proof} Fix $i\in \{0,\ldots, s-1\}$. As is proved in Theorem \ref{thm-1.2}, for $m$-a.e.~$x\in \Sigma'$, \begin{eqnarray*} \lim_{r\to 0}\frac{\log m_{x}^{\xi_{i}}(\Gamma_{i+1}(x,r))}{\log r}=\gamma_{i,i+1}(x) =\frac{h_{i+1}-{h_{i}}}{\lambda_{i+1}}. \end{eqnarray*} To see \eqref{e-theta_i} it remains to prove that $\frac{h_{i+1}-{h_{i}}}{\lambda_{i+1}}\leq k_{i+1}$. By Theorem \ref{thm-1.2}, for $m$-a.e.~$x\in \Sigma'$, the measure $\eta_x:=m^{\xi_i}_x\circ \pi^{-1}\circ (P_{{(V_x^{i+1})}^\perp})^{-1}$ is exact dimensional with dimension $\frac{h_{i+1}-{h_{i}}}{\lambda_{i+1}}$. However, $\eta_x$ is supported on the affine subspace $\pi x+ (V_x^{i}\ominus V_x^{i+1})$ of dimension $k_{i+1}$, where $V_x^{i}\ominus V_x^{i+1}$ stands for the orthogonal complement of $V_x^{i+1}$ in $V_x^{i}$. Hence $\dim_{\rm H}\eta_x\leq k_{i+1}$, and so, $\frac{h_{i+1}-{h_{i}}}{\lambda_{i+1}}\leq k_{i+1}$. \end{proof} \begin{lem} \label{lem-quasi} \begin{itemize} \item[(i)] Let $m\in {\mathcal M}_\sigma(\Sigma)$ be quasi-Bernoulli. Then for $m$-a.e.~$x\in \Sigma$, $m_x^{\xi_0}\circ \pi^{-1}$ is strongly equivalent to $m\circ \pi^{-1}$. \item[(ii)] Let $m\in {\mathcal M}_\sigma(\Sigma)$ be sub-multiplicative. Then for $m$-a.e.~$x\in \Sigma$, $m_x^{\xi_0}\circ \pi^{-1}$ is absolutely continuous with respect to $m\circ \pi^{-1}$. \end{itemize} \end{lem} \begin{proof} We first prove (i). Since $m$ is quasi-Bernoulli, by definition there exists a positive constant $C$ such that $$ C^{-1}m([I]_0)m([J]_0)\leq m([IJ]_0)\leq C m([I]_0)m([J]_0) $$ for all finite words $I, J$ over $\Lambda$. Below we show that for $m$-a.e.~$x\in \Sigma$, \begin{equation} \label{e-quasi} C^{-1}m([I]_0)\leq m_x^{\xi_0}([I]_0)\leq C m([I]_0) \end{equation} for all finite words $I$ over $\Lambda$. This is enough to conclude the strong equivalence between $m_x^{\xi_0}\circ \pi^{-1}$ and $m\circ \pi^{-1}$, since $\pi x$ only depends on $x^+:=(x_n)_{n=0}^\infty$. To see \eqref{e-quasi}, note that the measurable partition $\xi_0$ is induced by the mapping $\tau:\; \Sigma \to \Sigma^{-}$, $x\mapsto x^{-}=(x_n)_{-\infty}^{-1}$. That is, $\xi_0(x)=\{y\in \Sigma:\; \tau y=\tau x\}$ for every $x$. Applying Lemma \ref{lem-2.4}(1) to $\tau:\Sigma\to \Sigma^{-}$ yields that for $m$-a.e.~$x$, \begin{equation} \label{e-inequality} m^{\xi_0}_x([I]_0)= {\bf E}_m(\chi_{[I]_0}|\tau^{-1}({\mathcal B}(\Sigma^{-})))(x)=\lim_{n\to \infty} \frac{m([x_{-n}\ldots x_{-1}I]_0)}{m([x_{-n}\ldots x_{-1}]_0)} \end{equation} for all finite words $I$ over $\Lambda$. \eqref{e-quasi} is then obtained from the quasi-Bernoulli property of $m$. Next we prove (ii). Here $m$ is assumed to be sub-multiplicative and we only have the one-sided inequality $m([IJ]_0)\leq C m([I]_0)m([J]_0)$. However this is enough to derive from \eqref{e-inequality} that for $m$-a.e.~$x$, $m^{\xi_0}_x([I]_0)\leq Cm([I]_0)$ for all finite words $I$ over $\Lambda$. As a consequence, $m^{\xi_0}_x\circ \pi^{-1}$ is absolutely continuous with respect to $m\circ \pi^{-1}$, with a uniformly bounded Radon-Nikodym derivative. \end{proof} \begin{proof}[Proof of Theorem \ref{cor-1.0}] We first prove (i). Fix $i\in \{1,\ldots, s-1\}$. By Theorem \ref{thm-1.2}, for $m$-a.e.~$x\in \Sigma'$, $m_x^{\xi_i}\circ \pi^{-1}$ and $m_x^{\xi_0}\circ \pi^{-1}\circ (P_{({V_x^i})^\perp})^{-1}$ are exact dimensional with $$ \dim_{\rm H}(m_x^{\xi_i}\circ \pi^{-1})=\sum_{k=i}^{s-1}\frac{h_{k+1}-h_k}{\lambda_{k+1}} $$ and $$ \dim_{\rm H} \left(m_x^{\xi_0}\circ \pi^{-1}\circ (P_{({V_x^i})^\perp})^{-1}\right)=\sum_{k=0}^{i-1}\frac{h_{k+1}-h_k}{\lambda_{k+1}}, $$ hence by Theorem \ref{thm-1.1}, \begin{equation} \label{e-sum} \dim_{\rm H}(m_x^{\xi_i}\circ \pi^{-1})+\dim_{\rm H}\left(m_x^{\xi_0}\circ \pi^{-1}\circ (P_{({V_x^i})^\perp})^{-1}\right)=\dim_{\rm H}\left(m_x^{\xi_0}\circ \pi^{-1}\right). \end{equation} Next let $x\in \Sigma'$ and write $W=V_x^i$, $\nu=m^{\xi_0}_x$, $\eta=\nu\circ \pi^{-1}$. Notice that $\nu$ is supported on $\xi_0(x)$. Consider the measurable partition $\zeta$ of ${\Bbb R}^d$ given by $$ \zeta:=\{W+a:\; a\in W^{\perp}\}. $$ Set $\pi^{-1}\zeta:=\{\xi_0(x)\cap \pi^{-1}(W+a):\; a\in W^{\perp}\}$. Then $\pi^{-1}\zeta$ is a measurable partition of $\xi_0(x)$. Let $\{\nu_y^{\pi^{-1}\zeta}\}_{y\in \xi_0(x)}$ be the system of conditional measures of $\nu$ associated with $\pi^{-1}\zeta$, and $\{\eta_z^\zeta\}_{z\in {\Bbb R}^d}$ the system of conditional measures of $\eta$ associated with $\zeta$. Write $\eta_{W,z}:=\eta_z^\zeta$. By the uniqueness of conditional measures, we have for $\nu$-a.e.~$y$, \begin{equation} \label{e-cond} \nu_y^{\pi^{-1}\zeta} \circ \pi^{-1}=\eta_{W, \pi y}. \end{equation} Notice also that for $y\in \xi_0(x)$, the atom $(\pi^{-1}\zeta)(y)$ is nothing but $\xi_i(y)$. Hence we have $\nu_y^{\pi^{-1}\zeta}=m^{\xi_i}_y$ for $m$-a.e.~$x$ and $m_x^{\xi_0}$-a.e.~$y$. This combining with \eqref{e-cond} gives \begin{equation} \label{e-comp} m^{\xi_i}_x\circ \pi^{-1}= (m^{\xi_0}_x\circ \pi^{-1} )_{V_x^i, \pi x} \end{equation} for $m$-a.e.~$x$. Plugging the above equality into \eqref{e-sum}, we see that $m^{\xi_0}_x\circ \pi^{-1}$ satisfies dimension conservation along $V_x^i$. This proves (i). Now we turn to the proof of (ii). Suppose that $m$ is quasi-Bernoulli. By Lemma \ref{lem-quasi}(i), for $m$-a.e.~$x\in \Sigma'$, $m_x^{\xi_0}\circ \pi^{-1}$ is strongly equivalent to $\mu=m\circ \pi^{-1}$; as a consequence, $m^{\xi_0}_x\circ \pi^{-1}\circ (P_{({V_x^i})^\perp})^{-1}$ is strongly equivalent to $\mu\circ (P_{({V_x^i})^\perp})^{-1}$. It follows that for $m$-a.e.~$x\in \Sigma'$, $\mu\circ (P_{({V_x^i})^\perp})^{-1}$ is exact dimensional with dimension $\sum_{k=0}^{i-1} \frac{h_{k+1}-h_k}{\lambda_{k+1}}$. Equivalently, for $m\circ (\Pi_i)^{-1}$-a.e.~$W$, $\mu\circ (P_{W^\perp})^{-1}$ is exact dimensional with dimension $\sum_{k=0}^{i-1} \frac{h_{k+1}-h_k}{\lambda_{k+1}}$. Again since $m^{\xi_0}_x\circ \pi^{-1}$ is strongly equivalent to $\mu$ for $m$-a.e.~$x$, applying Lemma \ref{lem-2.8} to the orthogonal projection $P_{(V^i_x)^\perp}:\; {\Bbb R}^d\to (V^i_x)^\perp$, we see that $m$-a.e.~$x$, $\mu_{V_x^i, \pi x}$ is equivalent to $(m^{\xi_0}_x\circ \pi^{-1} )_{V_x^i, \pi x}=m^{\xi_i}_x\circ \pi^{-1}$, and so $\mu_{V_x^i, \pi x}$ is exact dimensional with dimension $\sum_{k=s-1}^{i} \frac{h_{k+1}-h_k}{\lambda_{k+1}}$. Equivalently, for $m\circ (\Pi_i)^{-1}$-a.e.~$W$ and $\mu$-a.e.~$z$, $\mu_{W, z}$ is exact dimensional with dimension $\sum_{k=i}^{s-1} \frac{h_{k+1}-h_k}{\lambda_{k+1}}$. Recall that we have proved that for $m\circ (\Pi_i)^{-1}$-a.e.~$W$, $\mu\circ (P_{W^\perp})^{-1}$ is exact dimensional with dimension $\sum_{k=0}^{i-1} \frac{h_{k+1}-h_k}{\lambda_{k+1}}$. This is enough to conclude (ii). Finally, we prove (iii). Suppose that $m$ is sub-multiplicative. By Lemma \ref{lem-quasi}(ii), for $m$-a.e.~$x\in \Sigma'$, $m_x^{\xi_0}\circ \pi^{-1}$ is absolutely continuous with respect to $\mu$. Hence there exists $H\subset \Sigma'$ with full $m$-measure such that for any $x\in H$, there exists a Borel set $F_x\subset {\Bbb R}^d$ with positive $\mu$-measure such that $(m_x^{\xi_0}\circ \pi^{-1})_{F_x}$ is strongly equivalent to $\mu_{F_x}$, where $\nu_A$ stands for the probability measure defined by $\displaystyle\nu_A(\cdot)={\nu(A\cap \cdot)}/{\nu (A)}$. As is proved in part (ii), when $m$ is quasi-Bernoulli, we can take $F_x=\Sigma$. Now fix $x\in H$ and $i\in \{1,\ldots, s-1\}$. Set $W=V^i_x$ and write for convenience \begin{align*} \eta:=m_x^{\xi_0}\circ \pi^{-1},\quad \eta':=(m_x^{\xi_0}\circ \pi^{-1})_{F_x},\quad \mu':=\mu_{F_x}. \end{align*} Applying Lemma \ref{lem-2.9} to the projection $P_{W^\perp}:{\Bbb R}^d\to {\Bbb R}^d$ and using the Borel density lemma, we see that for $\mu$-a.e.~$z\in F_x$ (equivalently for $\eta$-a.e.~$z\in F_x$), \begin{equation} \label{e-ff1} \begin{split} \dim_{\rm loc}((\eta')_{W,z}, z) &= \dim_{\rm loc}(\eta_{W,z}, z), \\ \dim_{\rm loc}((\mu')_{W,z}, z) &= \dim_{\rm loc}(\mu_{W,z}, z), \\ \dim_{\rm loc}(\eta'\circ (P_{{W^\perp}})^{-1}, P_{W^\perp}(z)) &= \dim_{\rm loc}(\eta\circ (P_{{W^\perp}})^{-1}, P_{W^\perp}(z)),\\ \dim_{\rm loc}(\mu'\circ (P_{{W^\perp}})^{-1}, P_{W^\perp}(z)) &= \dim_{\rm loc}(\mu\circ (P_{{W^\perp}})^{-1}, P_{W^\perp}(z)). \end{split} \end{equation} Since $\eta'$ and $\mu'$ are strongly equivalent, by Lemma \ref{lem-2.8}, for $\mu$-a.e.~$z\in F_x$, \begin{align*} \dim_{\rm loc}((\eta')_{W,z}, z) &= \dim_{\rm loc}((\mu')_{W,z}, z), \\ \dim_{\rm loc}(\eta'\circ (P_{{W^\perp}})^{-1}, P_{W^\perp}(z)) &= \dim_{\rm loc}(\mu'\circ (P_{{W^\perp}})^{-1}, P_{W^\perp}(z)). \end{align*} Combining the above equalities with \eqref{e-ff1} yields that for $\mu$-a.e.~$z\in F_x$, \begin{align*} \dim_{\rm loc}(\mu_{W,z}, z) &= \dim_{\rm loc}(\eta_{W,z}, z), \\ \dim_{\rm loc}(\mu\circ (P_{{W^\perp}})^{-1}, P_{W^\perp}(z)) &= \dim_{\rm loc}(\eta\circ (P_{{W^\perp}})^{-1}, P_{W^\perp}(z)). \end{align*} Now (iii) follows from (i). This completes the proof of the theorem. \end{proof} \begin{proof}[Proof of Theorem \ref{thm-1.7'}] Here we only give a sketched proof. It is based on \cite[Theorem 2.11]{FengHu2009} and its proof. Since the linear parts $M_j$ of ${\mathcal S}$ commute, ${\Bbb R}^d$ can be decomposed into the direct sum $T_1\oplus \cdots \oplus T_\ell$ of some subspaces with dimensions $q_1,\ldots, q_\ell$, so that for each pair $(j,p)\in \Lambda\times\{1,\ldots, \ell\}$, $M_jT_p\subset T_p$ and $M_j$ is ``weakly conformal'' on $T_p$ in the sense that there exists $a_{j,p}\geq 0$ so that $\lim_{n\to \infty}\|M_j^nv\|^{1/n}=a_{j,p}$ for $v\in T_{p}\backslash\{0\}$. Hence under a suitable coordinate change, ${\mathcal S}$ can be written as the direct product of some ``weakly conformal'' affine IFSs ${\mathcal S_1}$, \ldots, ${\mathcal S_\ell}$ on ${\Bbb R}^{q_1}$,\ldots, ${\Bbb R}^{q_\ell}$ (cf. \cite[Definition 2.10]{FengHu2009}). Set $\overline{\lambda}_p=\sum_{j\in \Lambda} m([j]_0) \log a_{j,p}$ for $p=1,\ldots, \ell$. Permutating ${\mathcal S_j}$'s if necessary, we may assume that $$ \overline{\lambda}_1\geq \cdots \geq \overline{\lambda}_\ell. $$ For $p\in \{1,\ldots, \ell\}$ and let $\tau_p$ be the orthogonal projection from ${\Bbb R}^d$ to $Y_p:={\Bbb R}^{q_1}\times \cdots \times {\Bbb R}^{q_p}$, and let $m^{\zeta_p}_x$ be the conditional measure of $m$ associated with the measurable partition $\{\pi^{-1}\circ \tau_p^{-1}(y):\ y\in Y_p\}$ of $\Sigma$. It is implicitly proved in \cite[Theorem 2.11]{FengHu2009} that there exist $h_m(\sigma)=\overline{h}_0\geq \overline{h}_1\geq \cdots\geq \overline{h}_\ell\geq 0$ such that for $m$-a.e.~$x\in \Sigma$ and $p\in \{1,\ldots, \ell-1\}$, the measure $m^{\zeta_p}_x\circ \pi^{-1}$ is exact dimensional with dimension $\sum_{j=p}^{\ell-1}\frac{\overline{h}_{j+1}-\overline{h}_j}{\overline{\lambda}_{j+1}}$, and moreover, $\mu=m\circ \pi^{-1}$ is exact dimensional with dimension $\sum_{j=0}^{\ell-1}\frac{\overline{h}_{j+1}-\overline{h}_j}{\overline{\lambda}_{j+1}}$. (We remark that this is only proved in \cite{FengHu2009} in the case when ${\mathcal S}$ is invertible and contracting. But it can be extended to the general case like Theorem \ref{cor-1.0}.) Applying this result to the IFS ${\mathcal S}_1\times\cdots\times {\mathcal S}_{p}$ gives that $\mu\circ \tau_p^{-1}$ is exact dimensional with dimension $\sum_{j=0}^{p-1}\frac{\overline{h}_{j+1}-\overline{h}_j}{\overline{\lambda}_{j+1}}$. Set $\mu=m\circ \pi^{-1}$. Let $\{\mu_{Y_p^\perp,z}\}$ denote the system of conditional measures of $\mu$ associated with the measurable partition $\{\tau_p^{-1}(y):\; y\in Y_p\}$ of ${\Bbb R}^d$. Similar to the proof of \eqref{e-comp}, we can show that for $m$-a.e.~$x\in \Sigma$, $\mu_{Y_p^\perp, \pi x}=m^{\zeta_p}_x\circ \pi^{-1}$. It follows that $\mu$ is dimension conserving with respect to the projection $\tau_p$. Moreover, $\mu_{Y_p^\perp, z}$ is exact dimensional for $\mu$-a.e~$z$. Now let $1\leq p_1<\cdots<p_{s'}=\ell$ be those integers so that $$\overline{\lambda}_1=\cdots=\overline{\lambda}_{p_1}>\overline{\lambda}_{p_1+1}=\cdots=\overline{\lambda}_{p_2}>\cdots> \overline{\lambda}_{p_{s'-1}+1}=\cdots=\overline{\lambda}_{p_{s'}}.$$ It is readily checked that $s=s'$, $\lambda_i=\overline{\lambda}_{p_i}$ and $V_x^i=W_i:=Y_{p_i}^\perp$ for $1\leq i\leq s$ and $m$-a.e~$x$. In particular, $P_{(W_i)^\perp}=\tau_{p_i}$ for $i=1,\ldots, s-1$. Hence $\mu$ is dimension conserving with respect to the projections $P_{(W_i)^\perp}$, $i=1,\ldots, s-1$. \end{proof} \begin{rem} \label{rem-6.3} {\rm The proof of Theorem \ref{thm-1.7'} implies the following result: Let ${\mathcal S}=\{S_j(x)=r_jx+a_j\}_{j\in \Lambda}$ be a self-similar IFS on ${\Bbb R}^d$ with $r_j>0$, average contracting with respect to an ergodic $m\in {\mathcal M}_\sigma(\Sigma)$. Then for any proper subspace $W$ of ${\Bbb R}^d$, $m\circ \pi^{-1}$ is dimension conserving with respect to $P_W$. This generalizes the result in \cite{FalconerJin2014, Furstenberg2008}. To see it, let $p=\dim W$ and let $v_1,\ldots, v_d$ be an orthonormal basis of ${\Bbb R}^d$ such that ${\rm span}(v_1,\ldots, v_p)=W$. Then one can check that ${\mathcal S}$ can be written as the product ${\mathcal S}_1 \times\cdots \times {\mathcal S}_d$ of some one-dimensional IFSs on $X_1,\ldots, X_d$, where $X_i={\rm span}(v_i)$, and moreover $\overline{\lambda}_1=\cdots = \overline{\lambda}_d$. Now the desired dimension conservation property follows from the proof of Theorem \ref{thm-1.7'}. } \end{rem} \section{Lyapunov dimension} \label{S-7} Throughout this section, let $m$ be an ergodic $\sigma$-invariant measure on $\Sigma$ and ${\bf M}=(M_j)_{j\in \Lambda}$ be a tuple of $d\times d$ real matrices satisfying $$\lambda({\bf M},m):=\lim_{n\to \infty}\frac{1}{n}\int \log \|M_{x_0}\cdots M_{x_{n-1}}\| \; dm(x)<0.$$ Let $\mathcal S=\{S_i(x)=M_jx+a_j\}_{j\in \Lambda}$ be an affine IFS on ${\Bbb R}^d$. Let $\{(\lambda_i, k_i)\}_{1\leq i\leq s}$ be the Lyapunov spectrum of ${\bf M}$ with respect to $(\Sigma, \sigma^{-1}, m)$. Set $$ L_0=0 \quad \mbox{ and }\quad L_i=-\sum_{\ell=1}^{i} \lambda_\ell k_\ell\;\mbox{ for }\; i=1,\ldots, s. $$ Clearly $L_0<L_1<\cdots<L_s$. Following \cite{JordanPollicottSimon2007}, we give the following. \begin{de} \label{de-LY} The Lyapunov dimension of $m$ with respect to ${\bf M}$, denoted as $\dim_{\rm LY} (m, {\bf M})$, is defined to be $$ \left\{ \begin{array}{ll} \displaystyle \left(\sum_{\ell=0}^{j-1}k_\ell\right)+\frac{ h_m(\sigma)-L_{j-1}}{(-\lambda_{j})} & \mbox{ if } L_{j-1}\leq h_m(\sigma) < L_j \mbox{ for some }j\in \{1,\ldots, s\},\\ &\\ \displaystyle \frac{d \;h_m(\sigma)}{L_s} & \mbox{ if } h_m(\sigma) \geq L_s. \end{array} \right. $$ \end{de} Let $\pi$ be the coding map associated with ${\mathcal S}$. Recall that $h_i$, $0\leq i\leq s$, are the conditional entropies of $m$ defined in \eqref{e-hi}, and $h_0=h_m(\sigma)$. The following result says that the Lyapunov dimension of $m$ is always an upper bound for the Hausdorff dimension of $m\circ \pi^{-1}$. This result was first proved in \cite{JordanPollicottSimon2007} under a stronger assumption that $\|M_j\|<1$ for all $j$. \begin{pro} \label{pro-7.2} $\dim_{\rm H} m\circ \pi^{-1} \leq \min\{d, \dim_{\rm LY}(m, {\bf M})\}$. Moreover, the equality holds if and only if one of the following holds: \begin{itemize} \item[(1)] $h_m(\sigma) \geq L_s$, and $h_i=h_m(\sigma)-L_i$ for all $i\in \{1,\ldots, s\}$. \item[(2)] $h_m(\sigma)\in [L_{j-1}, L_j)$ for some $j\in \{1,\ldots, s\}$, and $$ h_i=\left\{ \begin{array}{ll} h_m(\sigma)-L_i & \mbox{ if }1\leq i\leq j-1,\\ 0 & \mbox{ if }j\leq i\leq s. \end{array} \right. $$ \end{itemize} \end{pro} \begin{proof} Since $\lambda({\bf M}, m)<0$, the IFS $\mathcal S$ is average contracting with repect to $m$. By Theorem \ref{thm-1.1}, $\dim_{\rm H} m\circ \pi^{-1}=\sum_{i=0}^{s-1} \frac{h_{i+1}-h_i}{\lambda_{i+1}}$. Recall that $$0>\lambda({\bf M}, m)=\lambda_1>\cdots>\lambda_s\geq -\infty,$$ and $$h_m(\sigma)=h_0\geq h_1\geq \cdots \geq h_s\geq 0.$$ Moreover by Corollary \ref{cor-6.1}, $h_{i}-h_{i+1}\leq (-\lambda_{i+1})k_{i+1}$ for each $i$. Hence $\dim_{\rm H} m\circ \pi^{-1}$ is bounded above by $$ \Delta:=\max\left\{\sum_{i=0}^{s-1}\frac{x_{i+1}-x_i}{\lambda_{i+1}}:\; h_m(\sigma)=x_0\geq \cdots \geq x_s\geq 0, \; \frac{x_{i+1}-x_{i}}{\lambda_{i+1}}\leq k_{i+1} \mbox{ for all } i \right\}. $$ Now it is readily checked that the following hold: (a) if $h_m(\sigma) \geq L_s$, then $\Delta=d$ and the maximum in defining $\Delta$ is attained uniquely at $(x_1,\ldots, x_s)$ where $x_i=h_m(\sigma)-L_i$ for $0\leq i\leq s$; (b) if $h_m(\sigma)\in [L_{j-1}, L_j)$ for some $j\in \{1,\ldots, s\}$, then $$\Delta=\left(\sum_{\ell=0}^{j-1}k_\ell\right)+\frac{ h_m(\sigma)-L_{j-1}}{(-\lambda_{j})}, $$ and the maximum is attained uniquely at $(x_1,\ldots, x_s)$ where $x_i=h_m(\sigma)-L_i$ for $i\leq j-1$ and $0$ for $i\geq j$. As a consequence, the results of the proposition hold. \end{proof} \begin{rem} \label{rem-q1} By Proposition \ref{pro-7.2}, if $\dim_{\rm H} m\circ \pi^{-1}= \min\{d, \dim_{\rm LY}(m, {\bf M})\}$, then $$ \sum_{ \ell=1}^{j}\frac{h_\ell-h_{\ell-1}}{\lambda_\ell}=\min\{k_1+\cdots +k_j, \dim_{\rm H} m\circ \pi^{-1}\} \quad \mbox{ for }j=1,\ldots, s. $$ This result was partially proved in \cite[Corollay 2.7]{BaranyKaenmaki2015}. \end{rem} \begin{pro} \label{pro-7.3} Suppose that ${\mathcal S}$ is contracting and satisfies the strong separation condition. Then the following statements hold. \begin{itemize} \item [(i)] $h_s=0$, $h_m(\sigma)<L_s$ and $\dim_{\rm LY}(m,{\bf M})<d$. \item[(ii)] Let $j$ be the unique element in $\{1,\ldots, s\}$ so that $L_{j-1}\leq h_m(\sigma)<L_j$. Then $\dim_{\rm H} m\circ \pi^{-1}= \dim_{\rm LY} (m, {\bf M})$ if and only if \begin{equation} \label{e-equiv} \sum_{ \ell=1}^{j-1}\frac{h_\ell-h_{\ell-1}}{\lambda_\ell}=d_{j-1}, \qquad \sum_{\ell=j+1}^s\frac{h_\ell-h_{\ell-1}}{\lambda_\ell}=0, \end{equation} where $d_0:=0$ and $d_i:=k_1+\cdots+k_i$ for $1\leq i\leq s$. \end{itemize} \end{pro} \begin{proof} (i) We first claim that $h_s=0$. Since $\mathcal S$ satisfies the strong separation condition, $\xi_s(x)=\{x\}$ for each $x\in \Sigma'$. Thus $\widehat \xi_s={\mathcal B}(\Sigma')$ and hence $h_s=H_m({\mathcal P}|\widehat{\xi_s})=0$. Next we prove that $h_m(\sigma)<L_s$. Clearly this is true if $L_s=\infty$ (equivalently, if $\lambda_s=-\infty$). Below we assume that $\lambda_s>-\infty$. Let $K$ denote the self-affine set generated by ${\mathcal S}$. For $\delta>0$ let $K_\delta$ be the closed $\delta$-neighborhood of $K$, i.e. $K_\delta=\{z:\; d(z, K)\leq \delta\}$. Since $\mathcal S$ satisfies the strong separation condition, we can pick a small $\delta$ such that $S_i(K_\delta)$ ($i\in \Lambda$) are disjoint subsets of the interior of $K_\delta$ and hence ${\mathcal L}^d(K_\delta)>\sum_{i\in \Lambda} {\mathcal L}^d(S_i(K_\delta))$. It follows that $\rho:=\sum_{i\in \Lambda} |\det(M_i)|<1$. Since $m$ is ergodic $\sigma$-invariant, by \cite[Lemma 3.2]{FengShmerkin2014} and the Shannon-McMillan-Breiman theorem, for $m$-a.e.~$x\in \Sigma$, \begin{equation} \label{e-det} \lim_{n\to \infty} \frac{\log |\det(M_{x_0\ldots x_{n-1}})|}{n}=-L_s, \quad \lim_{n\to \infty} \frac{\log m([x_0\ldots x_{n-1}]_0)}{n}=-h_m(\sigma). \end{equation} For $\epsilon>0$ and $n\in {\Bbb N}$, let $\Lambda_{n,\epsilon}$ denote the set of words $I$ of length $n$ over the alphabet $\Lambda$ such that $$|\det(M_I)|\geq e^{-nL_s-n\epsilon},\quad m([I]_0) \leq e^{-nh_m (\sigma)+n\epsilon}.$$ By \eqref{e-det}, $\lim_{n\to \infty} \sum_{I\in \Lambda_{n,\epsilon}}m([I]_0)=1$. Notice that \begin{align*} \rho^n &=\sum_{I\in \Lambda^n}|\det(M_I)|\geq \sum_{I\in \Lambda_{n,\epsilon}} |\det(M_I)|\\ &\geq \sum_{I\in \Lambda_{n,\epsilon}} e^{-nL_s-n\epsilon} \frac{m([I]_0)}{e^{-nh_m (\sigma)+n\epsilon}}\\ &=e^{-n(L_s-h_m(\sigma)-2\epsilon)}\cdot \left( \sum_{I\in \Lambda_{n,\epsilon}}m([I]_0)\right). \end{align*} Letting $n\to \infty$ and $\epsilon\to 0$, we obtain the desired inequality $h_m(\sigma)\leq L_s+\log \rho<L_s$. Now the inequality $\dim_{\rm LY}(m, {\bf M})<d$ follows directly from Definition \ref{de-LY}. This proves (i). Finally we prove (ii). Since $h_s=0$ and $0\leq h_{\ell-1}-h_\ell\leq (-\lambda_\ell) k_\ell$ for each $\ell$ by Corollary \ref{cor-6.1}, we see that \eqref{e-equiv} holds if and only if $h_{\ell-1}-h_\ell=(-\lambda_\ell) k_\ell$ for $1\leq \ell\leq j-1$ and $h_{\ell}=0$ for $j\leq \ell\leq s$. By Proposition \ref{pro-7.2}, this is equivalent to that $\dim_{\rm H} m\circ \pi^{-1}=\dim_{\rm LY}(m, {\bf M})$. \end{proof} \begin{rem} \label{rem-6.4} {\rm Theorems \ref{cor-1.0} (resp.~Theorem \ref{thm-1.7'}) can be applied to estimate the dimension of slices and projections of certain self-affine sets. To see it, let $K$ a self-affine sets generated by a contracting affine IFS $\{S_j=M_jx+a_j\}_{j\in \Lambda}$ on ${\Bbb R}^d$. Suppose that there exists an ergodic $m\in \mathcal M_\sigma(\Sigma)$ so that \begin{equation} \label{e-as1} \dim_{\rm H}m\circ \pi^{-1}=\dim_{\rm H}K. \end{equation} Follow the notation in Theorem \ref{cor-1.0} and assume $s\geq 2$. Since the slicing measures $(m^{\xi_0}_x\circ \pi^{-1})_{V_x^i,y}$ are supported on the slices $K\cap (V_x^i+y)$, by using Theorem \ref{cor-1.0}(i) and a general inequality in Theorem 2.10.25 of Federer \cite{Federer1969}, we obtain that for $i\in \{1,\ldots, s-1\}$ and $m$-a.e.~$x$, $$ \dim_{\rm H} K\cap (V_x^i+y)=\sum_{\ell=i}^{s-1}\frac{h_{\ell+1}-h_\ell}{\lambda_{\ell+1}}\quad \mbox{ for $(m^{\xi_0}_x\circ \pi^{-1})\circ P^{-1}_{(V_x^i)^\perp}$-a.e.~$y\in (V_x^i)^\perp$} $$ and \begin{equation} \label{e-r1} \dim_{\rm H}\left\{y\in P_{(V_x^i)^\perp}(K):\; \dim_{\rm H} K\cap (V_x^i+y)=\sum_{\ell=i}^{s-1}\frac{h_{\ell+1}-h_\ell}{\lambda_{\ell+1}} \right\}=\sum_{\ell=0}^{i-1}\frac{h_{\ell+1}-h_\ell}{\lambda_{\ell+1}}. \end{equation} If in addition to the assumption \eqref{e-as1}, we further assume that $$\dim_{\rm H} m\circ \pi^{-1}=\dim_{\rm LY}(m, {\bf M}),$$ then \begin{equation} \label{e-r2} \dim_{\rm H} P_{(V_x^i)^\perp}(K)=\min\{\dim (V_x^i)^\perp, \dim_HK\} \quad \mbox{ for $m$-a.e.~$x$}. \end{equation} Indeed by Remark \ref{rem-q1}, the sum in the right-hand side of \eqref{e-r1} is equal to $$\min\{(\dim(V_x^i)^\perp), \dim_{\rm H} m\circ \pi^{-1}\},$$ and hence equal to $\min\{\dim(V_x^i)^\perp, \dim_{\rm H} K\}$. Now \eqref{e-r2} follows from \eqref{e-r1}. } \end{rem} \section{Semi-continuity of entropies and dimensions} \label{S-8} In this section, we prove Theorems \ref{thm-1.3}-\ref{thm-1.4}. Set \begin{equation} \label{e-fx} f(x)=\sum_{n=1}^\infty\|M_{x_0}\cdots M_{x_{n-1}}\|\quad \mbox{ for $x\in \Sigma$}. \end{equation} \begin{lem} \label{lem-7.1} Let $\eta$ be a Borel probability measure on $\Sigma$ with $\eta(\{f=\infty\})=0$. Then $\eta\circ \pi_{{\bf a}}^{-1}$ depends continuously on ${\bf a}$, in the sense that $\eta\circ \pi_{{\bf a}_n}^{-1}$ converges to $\eta\circ \pi_{{\bf a}}^{-1}$ weakly when ${\bf a}_n$ converges to ${\bf a}$. \end{lem} \begin{proof} Since $f(x)<\infty$ for $\eta$-a.e.~$x\in \Sigma$, $\pi_{{\bf a}}(x)$ is well-defined for every ${\bf a}\in {\Bbb R}^{d|\Lambda|}$ and moreover, \begin{equation} \label{e-pif} \|\pi_{{\bf a}}(x)-\pi_{{\bf b}}(x)\|\leq f(x)\|{\bf a}-{\bf b}\|. \end{equation} For $N\in {\Bbb N}$, set $A_N:=\{x: f(x)<N\}$. Let $({\bf a}_n)\subset {\Bbb R}^{d|\Lambda|}$ so that $\lim_{n\to \infty}{\bf a}_n={\bf a}$. For convenience, write $\nu_n=\eta\circ \pi_{{\bf a}_n}^{-1}$ and $\nu=\eta\circ \pi_{{\bf a}}^{-1}$. To show that $\nu_n$ converges weakly to $\nu$, by the Portmanteau theorem, it suffices to show that $\limsup_{n\to \infty} \nu_n(F)\leq \nu(F)$ for any compact set $F\subset {\Bbb R}^d$. Now fix a compact set $F\subset {\Bbb R}^d$. Let $\epsilon>0$. Take a small $r>0$ so that $\nu(V_r(F))\leq \nu(F)+\epsilon$, where $V_r(F)$ stands for the $r$-neighborhood of $F$. Take a large $N$ so that $\eta(\Sigma \setminus A_N)<\epsilon$. Pick $n_0$ so that $\|{\bf a}_n-{\bf a}\|<r/N$ when $n\geq n_0$. By \eqref{e-pif}, for $x\in A_N$ and $n\geq n_0$ we have $\|\pi_{{\bf a_n}}(x)-\pi_{{\bf a}}(x)\|\leq N\|{\bf a_n}-{\bf a}\|<r$. Hence $A_N\cap \pi_{{\bf a}_n}^{-1}(F)\subset A_N\cap \pi_{{\bf a}}^{-1}(V_r(F))$ for $n\geq n_0$. It follows that for $n\geq n_0$, \begin{align*} \nu_n(F)=\eta(\pi_{{\bf a}_n}^{-1}(F))&\leq \eta(\Sigma\setminus A_N)+ \eta(A_N\cap \pi_{{\bf a}_n}^{-1}(F))\\ &\leq \epsilon+ \eta(A_N\cap \pi_{{\bf a}}^{-1}(V_r(F)))\\ & \leq \epsilon+ \nu(V_r(F))\leq \nu(F)+2\epsilon. \end{align*} Hence $\limsup_{n\to \infty} \nu_n(F)\leq \nu(F)+2\epsilon$. Letting $\epsilon\to 0$ gives $\limsup_{n\to \infty} \nu_n(F)\leq \nu(F)$, as desired. \end{proof} \begin{proof}[Proof of Theorem \ref{thm-1.3}] We first prove part (1) of the theorem. This is done by extending an idea of Rapaport \cite[Lemma 8]{Rapaport2015}. It is implicitly proved in Proposition \ref{pro-3.1} that $m(\{f=\infty\})=0$, where $f$ is defined as in \eqref{e-fx}. Let $i\in \{1,\ldots, s\}$ and write $\xi_{i,{\bf a}}$ for $\xi_i$ so as to emphasize its dependence on ${\bf a}$. Since $$ 0=m(\{f=\infty\})=\int m^{\xi_{i,{\bf a}}}_x (\{f=\infty\}) d m(x), $$ the set $\Delta_{{\bf a}}:=\left\{x\in \Sigma': \; m^{\xi_{i,{\bf a}}}_x (\{f=\infty\})=0\right\}$ has full $m$-measure. Noticing that $\xi_0$ is independent of ${\bf a}$, and $\xi_{i,{\bf a}}$ is a refinement of $\xi_0$ (i.e. any set in $\xi_{i,{\bf a}}$ is a subset of an element in $\xi_0$), we have \begin{align*} h_{i, {\bf a}}&=H_m({\mathcal P}|\xi_{i,{\bf a}})\\ &=\int -\log m_x^{\xi_{i,{\bf a}}}({\mathcal P}(x))\;dm(x)\\ &=\int \int -\log m_y^{\xi_{i,{\bf a}}}({\mathcal P}(y))\;dm_x^{\xi_0}(y)\; dm(x)\\ &=\int H_{m_x^{\xi_0}}({\mathcal P}|\xi_{i,{\bf a}})\; dm(x). \end{align*} Fix ${\bf a}_0\in {\Bbb R}^{d|\Lambda|}$. In what follows we show that $h_{i, {\bf a}}$ is upper semi-continuous in ${\bf a}$ at ${\bf a}_0$. Since $\Delta_{{\bf a}_0}$ has full $m$-measure, $h_{i, {\bf a}}=\int_{\Delta_{{\bf a}_0}} H_{m_x^{\xi_0}}({\mathcal P}|\xi_{i,{\bf a}})\; dm(x)$. Hence it is sufficient to show that ${\bf a}\mapsto H_{m_x^{\xi_0}}({\mathcal P}|\xi_{i,{\bf a}})$ is upper semi-continuous at ${\bf a}_0$ for every $x\in \Delta_{{\bf a}_0}$. For this purpose, fix $x\in \Delta_{{\bf a}_0}$ and write $C=\xi_0(x)$, $W=V_x^i$ and $m_C=m_x^{\xi_0}$. Then by the definition of $\xi_{i,{\bf a}}$, $$H_{m_x^{\xi_0}}({\mathcal P}|\xi_{i,{\bf a}})=H_{m_C}({\mathcal P}|\pi_{\bf a}^{-1}\circ P_{W^\perp}^{-1}({\mathcal B}(W^\perp))).$$ Following the proof of \cite[Lemma 8.5]{Walters1982} or \cite[Lemma 8]{Rapaport2015} with minor changes, we can construct a sequence $(\beta_n)$ of finite Borel partitions of $W^\perp$ such that (i) $\sigma(\beta_n)\uparrow {\mathcal B}(W^{\perp})$ and (ii) $m_C\circ \pi^{-1}_{{\bf a}_0}(P_{W^\perp}^{-1}(\partial B))=0$ for any $B\in \bigcup_n \beta_n$. Since $\sigma(\beta_n)\uparrow {\mathcal B}(W^{\perp})$, \begin{align*} H_{m_C}({\mathcal P}|\pi_{\bf a}^{-1}\circ P_{W^\perp}^{-1}({\mathcal B}(W^{\perp}))&=\lim_{n\to \infty}H_{m_C}({\mathcal P}|\pi_{\bf a}^{-1}\circ P_{W^\perp}^{-1}(\sigma(\beta_n))\\ &=\lim_{n\to \infty}\sum_{A\in {\mathcal P}}\sum_{B\in \beta_n} u\left((m_C|_A)\circ \pi_{\bf a}^{-1} (P_{W^\perp}^{-1}(B))\right), \end{align*} where $u(z):=-z\log z$. Since $x\in \Delta_{{\bf a}_0}$, $m_C(\{f=\infty\})=0$. By Lemma \ref{lem-7.1}, the measures $m_C\circ \pi_{\bf a}^{-1}$ and $(m_C|_{A})\circ \pi_{\bf a}^{-1}$ ($A\in {\mathcal P}$) depend continuously on ${\bf a}$; and so do $m_C\circ \pi_{\bf a}^{-1}\circ P_{W^{\perp}}^{-1}$ and $(m_C|_{A})\circ \pi_{\bf a}^{-1}\circ P_{W^{\perp}}^{-1}$. It follows that, as functions of ${\bf a}$, $u\left(m_C\circ \pi_{\bf a}^{-1} (P_{W^\perp}^{-1}(B))\right)$ and $u\left((m_C|_{A})\circ \pi_{\bf a}^{-1} (P_{W^\perp}^{-1}(B))\right)$ ($A\in {\mathcal P}$) are continuous at ${\bf a}_0$, and so is $H_{m_C}({\mathcal P}|\pi_{\bf a}^{-1}\circ P_{W^\perp}^{-1}(\sigma(\beta_n))$. Hence ${\bf a}\mapsto H_{m_C}({\mathcal P}|\pi_{\bf a}^{-1}\circ P_{W^\perp}^{-1}({\mathcal B}(W^{\perp}))$ is upper semi-continuous at ${\bf a}_0$, as desired. This proves the upper semi-continuity of $h_{i,{\bf a}}$. Next we prove the lower semi-continuity of the mapping ${\bf a}\mapsto \dim_{{\rm H}}(m\circ \pi_{\bf a}^{-1})$. By Theorem \ref{thm-1.1}, we have \begin{equation} \label{e-uppc} \dim_{\rm H}(m\circ \pi_{\bf a}^{-1})=\sum_{i=0}^{s} t_i h_{i, {\bf a}}, \end{equation} where $t_0=-\frac{1}{\lambda_1}$ and $t_i= \frac{1}{\lambda_i}- \frac{1} {\lambda_{i+1}} $ for $i=1,\ldots, s$, with convention $\lambda_{s+1}:=-\infty$. Notice that $t_0>0$, $t_i\leq 0$ for $1\leq i\leq s$ and moreover, $h_{0, {\bf a}}\equiv h_\sigma(m)$. By part (1), $h_{1, {\bf a}}, \ldots, h_{s, {\bf a}}$ are upper semi-continuous in ${\bf a}$. Hence by \eqref{e-uppc}, $\dim_{\rm H}(m\circ \pi_{\bf a}^{-1})$ is lower semi-continuous in ${\bf a}$. \end{proof} \begin{rem} \label{rem-8.1} Theorem \ref{thm-1.3} can be further extended. For given $m$ and ${\bf M}=(M_j)_{j\in \Lambda}$, let $\mathcal S_{{\bf r}, {\bf a}}$ denote the IFS $\{r_jM_jx+a_j\}_{j\in \Lambda}$ where ${\bf r}=(r_j)_{j\in \Lambda}\in ({\Bbb R}\backslash \{0\})^\Lambda$ so that $S_{{\bf r}, {\bf a}}$ is average contracting with respect to $m$. Notice that the Oseledets subspaces with respect to $m$ and $(r_jM_j)_{j\in \Lambda}$ are independent of ${\bf r}$. A slight modification of the above proof establishes the upper semi-continuity of $({\bf r}, {\bf a})\mapsto h_{i,{\bf r}, {\bf a}}$ and the lower semi-continuity of $({\bf r}, {\bf a})\mapsto \dim_{\rm H}(m\circ \pi_{{\bf r}, {\bf a}}^{-1})$. \end{rem} \begin{proof}[Proof of Theorem \ref{cor-1.7}] We first prove (i). Let $m$ be an ergodic $\sigma$-invariant measure $m$ on $\Sigma$. For $n\in {\Bbb N}$, set $$ \Omega_n:=\left\{{\bf a} \in {\Bbb R}^{d|\Lambda|}:\; \dim_{\rm H} (m\circ \pi_{\bf a}^{-1})\leq \min(d, \dim_{\rm LY} (m, {\bf M}))-\frac{1}{n}\right\}. $$ Since $\dim_{\rm H} (m\circ \pi_{\bf a}^{-1})$ is lower semi-continuous in ${\bf a}$ by Theorem \ref{thm-1.3}, $\Omega_n$ is closed for each $n$. Meanwhile, it was proved in \cite{JordanPollicottSimon2007} that $\dim_{\rm H} (m\circ \pi_{\bf a}^{-1})=\min(d, \dim_{\rm LY} (m, {\bf M}))$ for $\mathcal L^{d|\Lambda|}$-a.e.~${\bf a}$. Hence for each $n$, $\Omega_n$ is a closed set of zero Lebesgue measure, so it is nowhere dense. This is enough to conclude (i). Next we prove (ii). It was shown by K\"aenm\"aki \cite{Kaenmaki2004} that there exists an ergodic $\sigma$-invariant measure $\eta$ on $\Sigma$ such that $\dim_{\rm LY} (\eta, {\bf M})=\dim_{\rm AFF} ({\bf M})$. Fix such $\eta$. Note that for each ${\bf a}$, $$\dim_{\rm H} (\eta\circ \pi_{\bf a}^{-1})\leq \dim_{\rm H} K({\bf M}, {\bf a})\leq \min(d, \dim_{\rm AFF} ({\bf M})).$$ It implies that \begin{align*} &\left\{{\bf a}\in {\Bbb R}^{d|\Lambda|}:\; \dim_{\rm H} K({\bf M}, {\bf a})\neq \min(d, \dim_{\rm AFF}({\bf M}))\right\}\\ &\mbox{}\quad\subset \left\{{\bf a}\in {\Bbb R}^{d|\Lambda|}:\; \dim_{\rm H} (\eta\circ \pi_{\bf a}^{-1})\neq \min(d, \dim_{\rm LY} (\eta, {\bf M})))\right\}. \end{align*} Now (ii) follows from (i). \end{proof} To prove Theorem \ref{thm-1.4} we need the following. \begin{lem}[{\cite[Lemma 22]{Rapaport2015}}] \label{lem-rap} Let $\mu$ be a probability Borel measure on ${\Bbb R}^d$ and $1\leq k< d$. Then the following statements hold. \begin{itemize} \item[(i)] If $\dim_{\rm H}\mu\leq k$ then for $0<t\leq \dim_{\rm H}\mu$, $$ \dim_{\rm H} \{W\in G(d, k):\; \dim_{\rm H} \mu\circ (P_W)^{-1} <t\}\leq k(d-k-1)+t. $$ \item[(ii)] If $\dim_{\rm H}\mu\geq k$ then for $\dim_{\rm H}\mu-k(d-k)<t\leq k$, $$ \dim_{\rm H} \{W\in G(d, k):\; \dim_{\rm H} \mu\circ (P_W)^{-1} <t\}\leq k(d-k)+t-\dim_{\rm H}\mu. $$ \end{itemize} \end{lem} \begin{proof}[Proof of Theorem \ref{thm-1.4}] The proof is mainly adapted from \cite{Rapaport2015}. For the convenience of the reader, we include the details. Write $\mu=m\circ \pi^{-1}$. Since ${\mathcal S}$ satisfies the strong separation condition, by Proposition \ref{pro-7.3} we have $h_s=0$, $h_m(\sigma)<\sum_{\ell=1}^s (-\lambda_\ell) k_\ell$ and $\dim_{\rm LY}(m, {\bf M})<d$. Let $i$ be the unique element in $\{1,\ldots, s\}$ so that $d_{i-1} \leq \dim_{\rm LY} (m, {\bf M})<d_i$. (Recall that $d_0=0$ and $d_j=k_1+\cdots+k_j$ for $j\geq 1$.) By Definition~\ref{de-LY}, we have $h_m(\sigma)\in [L_{i-1}, L_i)$ where $L_0:=0$ and $L_j:=-\sum_{\ell=1}^i\lambda_\ell k_\ell$ for $j\geq 1$. Below we prove the equality $\dim_{\rm H}\mu=\dim_{\rm LY} (m, {\bf M})$ under the assumption that one of the scenarios (a), (b), (c) occurs. We first consider the scenario (a). In this case, $s=1$ and by Theorem \ref{thm-1.1}, $$\dim_{\rm H} \mu=\frac{h_1-h_0}{\lambda_1}=-\frac{h_m(\sigma)}{\lambda_1}= \dim_{\rm LY}(m, {\bf M}).$$ Next we consider the scenario (b). In this case, $i=s$ and so $h_m(\sigma)\in [L_{s-1}, L_s)$. To show that $\dim_{\rm H}m\circ \pi^{-1}=\dim_{\rm LY}(m, {\bf M})$, it suffices to show that \begin{equation} \label{e-LYdim} h_{s-1}= h_m(\sigma)+k_1\lambda_1+\cdots +k_{s-1}\lambda_{s-1}. \end{equation} Indeed if \eqref{e-LYdim} holds, then $h_0-h_{s-1}=\sum_{\ell=1}^{s-1} (-\lambda_\ell)k_\ell$, which forces that $h_{j-1}-h_j=(-\lambda_j)k_j$ for $1\leq j\leq s-1$ (recalling that $h_{j-1}-h_j\leq (-\lambda_j)k_j$ for all $j$ by Corollary \ref{cor-6.1}); hence $$\sum_{\ell=1}^{s-1}\frac{h_\ell-h_{\ell-1}}{\lambda_\ell}=d_{s-1},$$ so \eqref{e-equiv} holds for $j=s-1$, then by Proposition \ref{pro-7.3}, we obtain that $\dim_{\rm H}\mu=\dim_{\rm LY}(m, {\bf M})$. To show \eqref{e-LYdim} we first prove that \begin{equation} \label{e-LYdim'} h_{s-1}\geq h_m(\sigma)+k_1\lambda_1+\cdots +k_{s-1}\lambda_{s-1}. \end{equation} To see this, replacing ${\mathcal S}$ by one of its iterations if necessary, we may assume that $\|M_j\|<1/2$ for all $j\in \Lambda$. By Theorem 1.9 in \cite{JordanPollicottSimon2007}, for ${\mathcal L}^{d|\Lambda|}$-a.e.~${\bf a}\in {\Bbb R}^{d|\Lambda|}$, $\dim_{\rm H} m\circ (\pi_{{\bf a}})^{-1}=\dim_{\rm LY}(m, {\bf M})$. Hence by Proposition \ref{pro-7.2}, for ${\mathcal L}^{d|\Lambda|}$-a.e.~${\bf a}\in {\Bbb R}^{d|\Lambda|}$, $h_{s-1,{\bf a}}=h_m(\sigma)+k_1\lambda_1+\cdots +k_{s-1}\lambda_{s-1}.$ Since $h_{s-1,{\bf a}}$ is upper semi-continuous in {{\bf a}} by Theorem \ref{thm-1.3}, it follows that $h_{s-1,{\bf a}}\geq h_m(\sigma)+k_1\lambda_1+\cdots +k_{s-1}\lambda_{s-1}$ for all ${\bf a}\in {\Bbb R}^{d|\Lambda|}$. This proves \eqref{e-LYdim'}. Now suppose on the contrary that \eqref{e-LYdim} does not hold. Then by \eqref{e-LYdim'}, there exists $\delta>0$ such that $h_{s-1}= h_m(\sigma)+k_1\lambda_1+\cdots +k_{s-1}\lambda_{s-1}+\delta$. By Theorem \ref{cor-1.0} (iii), for $m\circ (\Pi_{s-1})^{-1}$-a.e.~$W\in G(d, d-d_{s-1})$, \begin{align*} \dim_{\rm H} \mu\circ (P_{W^\perp})^{-1}&\leq \sum_{\ell=1}^{s-1} \frac{h_\ell-h_{\ell-1}}{\lambda_{\ell}}\\ &=\dim_{\rm H}\mu-\frac{h_s-h_{s-1}}{\lambda_s}\\ &=\dim_{\rm H}\mu-\frac{h_{s-1}}{(-\lambda_s)}\\ &= \dim_{\rm H}\mu +d_{s-1}-\dim_{\rm LY}(m, {\bf M})-\delta/(-\lambda_s), \end{align*} where in the last equality, we use the fact that $$\dim_{\rm LY}(m, {\bf M})=d_{s-1}+\frac{h_0-L_{s-1}}{(-\lambda_s)}=d_{s-1}+\frac{h_{s-1}-\delta}{(-\lambda_s)}.$$ Let ${\bf Y}$ denote the set of $W\in G(d, d-d_{s-1})$ such that $$\dim_{\rm H} \mu\circ (P_{W^\perp})^{-1}\leq \dim_{\rm H}\mu +d_{s-1}-\dim_{\rm LY}(m, {\bf M})-{\delta}/{(-\lambda_s)}.$$ Then $m\circ (\Pi_{s-1})^{-1}({\bf Y})=1$, so by \eqref{e-condLY}, \begin{equation} \label{e-lower} \begin{split}\dim_{\rm H} {\bf Y}&\geq \dim_{\rm H}^*m\circ (\Pi_{s-1})^{-1}\\ &\geq d_{s-1}(d-d_{s-1})+d_{s-1}-\dim_{\rm LY}(m, {\bf M}). \end{split} \end{equation} On the other hand, we can get an upper bound estimate for $\dim_{\rm H} {\bf Y}$ by using Lemma \ref{lem-rap}. Indeed, if $\dim_{\rm H}\mu\leq d_{s-1}$, then by Lemma \ref{lem-rap}(i) applied to $k=d_{s-1}$ and $t=\dim_{\rm H}\mu +d_{s-1}-\dim_{\rm LY}(m, {\bf M})-{\delta}/{(-\lambda_s)}$, we see that \begin{align*} \dim_{\rm H} {\bf Y}&\leq d_{s-1}(d-d_{s-1})+\dim_{\rm H}\mu-\dim_{\rm LY}(m, {\bf M})-{\delta}/{(-\lambda_s)}\\ &\leq d_{s-1}(d-d_{s-1})+d_{s-1}-\dim_{\rm LY}(m, {\bf M})-{\delta}/{(-\lambda_s)}; \end{align*} Conversely if $\dim_{\rm H}\mu>d_{s-1}$, then by Lemma \ref{lem-rap}(ii) applied to $k=d_{s-1}$ and $t=\dim_{\rm H}\mu +d_{s-1}-\dim_{\rm LY}(m, {\bf M})-{\delta}/{(-\lambda_s)}$, we get the same upper bound for $\dim_{\rm H} {\bf Y}$, which contradicts with \eqref{e-lower}. This proves \eqref{e-LYdim}. Finally we consider the scenario (c). In this case, $h_m(\sigma)\in [L_{i-1}, L_i)$. Clearly the assumptions \eqref{e-al2}-\eqref{e-al1} imply that $$d_{i-1}\leq \dim_{\rm H}\mu \leq \dim_{\rm LY}(m, {\bf M}) \leq d_i.$$ To prove $\dim_{\rm H}\mu=\dim_{\rm LY}(m, {\bf M})$, by Proposition \ref{pro-7.3} it suffices to prove that $\sum_{\ell=1}^{i-1} \frac{h_\ell-h_{\ell-1}}{\lambda_\ell}=d_{i-1}$ and $\sum_{\ell=i+1}^s \frac{h_\ell-h_{\ell-1}}{\lambda_\ell}=0$. As $d_0=0$, the first equality holds automatically when $i=1$. Now we first prove that $\sum_{\ell=1}^{i-1} \frac{h_\ell-h_{\ell-1}}{\lambda_\ell}=d_{i-1}$. To avoid triviality, we assume that $i\geq 2$. For $n\in {\Bbb N}$, let ${\bf X}_n$ denote the set of $W\in G(d, d-d_{i-1})$ so that $\dim_{\rm H}\mu\circ (P_{W^\perp})^{-1}\leq d_{i-1}-{1}/{n}$. By Lemma \ref{lem-rap}(ii) applied to $k=d_{i-1}$ and $t=d_{i-1}-1/n$, \begin{equation*} \begin{split} \dim_{\rm H} {\bf X}_n &\leq d_{i-1}(d-d_{i-1})+d_{i-1}-({1}/{n})-\dim_{\rm H}\mu\\ &< \dim_{\rm H}^*m\circ (\Pi_{i-1})^{-1}\qquad(\mbox{by \eqref{e-al1}}). \end{split} \end{equation*} It follows that $m\circ (\Pi_{i-1})^{-1}(X_n)<1$ and hence $\dim_{\rm H}\mu\circ (P_{W^\perp})^{-1}> d_{i-1}- {1}/{n}$ on a set of positive $m\circ (\Pi_{i-1})^{-1}$-measure. However by Theorem \ref{cor-1.0}(iii), \begin{equation} \label{e-915} \dim_{\rm H}\mu\circ (P_{W^\perp})^{-1}\leq \sum_{\ell=1}^{i-1} \frac{h_\ell-h_{\ell-1}}{\lambda_\ell} \quad \mbox{ for $m\circ (\Pi_{i-1})^{-1}$-a.e.~$W$}. \end{equation} It follows that $\sum_{\ell=1}^{i-1} \frac{h_\ell-h_{\ell-1}}{\lambda_\ell}\geq d_{i-1}-{1}/{n}$. As $n$ is arbitrary, we obtain that $\sum_{\ell=1}^{i-1} \frac{h_\ell-h_{\ell-1}}{\lambda_\ell}\geq d_{i-1}$. Since $h_{\ell-1}-h_\ell\leq (-\lambda_\ell)k_\ell$ for each $\ell$ by Corollary \ref{cor-6.1}, we have $\sum_{\ell=1}^{i-1} \frac{h_\ell-h_{\ell-1}}{\lambda_\ell}= d_{i-1}$, as desired. Next we prove that $\sum_{\ell=i+1}^{s} \frac{h_\ell-h_{\ell-1}}{\lambda_\ell}= 0$. For $n\in {\Bbb N}$, let ${\bf Z}_n$ denote the set of $W\in G(d, d-d_{i})$ so that $\dim_{\rm H}\mu\circ (P_{W^\perp})^{-1}\leq \dim_{\rm H}\mu- {1}/{n}$. By Lemma \ref{lem-rap}(i) applied to $k=d_i$ and $t=\dim_{\rm H}\mu-{1}/{n}$, \begin{equation*} \begin{split} \dim_{\rm H} {\bf Z}_n &\leq d_{i}(d-d_{i})-d_{i}+\dim_{\rm H}\mu-({1}/{n})\\ &\leq d_{i}(d-d_{i})-d_{i}+ \dim_{\rm LY}(m, {\bf M})-({1}/{n})\\ &< \dim_{\rm H}^*m\circ (\Pi_{i})^{-1}\qquad(\mbox{by \eqref{e-al2}}). \end{split} \end{equation*} Hence $m\circ (\Pi_{i-1})^{-1}({\bf Z}_n)<1$ and so $\dim_{\rm H}\mu\circ (P_{W^\perp})^{-1}> \dim_{\rm H}\mu- {1}/{n}$ on a set of positive $m\circ (\Pi_{i})^{-1}$-measure. This combining with \eqref{e-915} (in which we replace $i-1$ by $i$) yields that $\sum_{\ell=1}^{i} \frac{h_\ell-h_{\ell-1}}{\lambda_\ell}\geq \dim_{\rm H}\mu-{1}/{n}$. Letting $n\to \infty$ gives $\sum_{\ell=1}^{i} \frac{h_\ell-h_{\ell-1}}{\lambda_\ell}\geq \dim_{\rm H}\mu$, which, together with \eqref{ly-dim}, implies that $\sum_{\ell=i+1}^{s} \frac{h_\ell-h_{\ell-1}}{\lambda_\ell}= 0$. This completes the proof of the theorem. \end{proof} \medskip {\noindent \bf Acknowledgements}. The author is indebted to Julien Barral, Xiong Jin and Fran\c{c}ois Ledrappier for some helpful comments, and to Yufeng Wu for catching many typos. The work was partially supported by HKRGC GRF grants (projects CUHK14301218, CUHK14302415) and the Direct Grant for Research in CUHK.

https://arxiv.org/abs/solv-int/9811006

Canonical variables for multiphase solutions of the KP equation

The KP equation has a large family of quasiperiodic multiphase solutions. These solutions can be expressed in terms of Riemann-theta functions. In this paper, a finite-dimensional canonical Hamiltonian system depending on a finite number of parameters is given for the description of each such solution. The Hamiltonian systems are completely integrable in the sense of Liouville. In effect, this provides a solution of the initial-value problem for the theta-function solutions. Some consequences of this approach are discussed.

\section{Introduction} In 1970, Kadomtsev and Petviashvili \cite{kadom} derived two equations as generalizations of the Korteweg-de Vries (KdV) equation to two spatial dimensions: \renewcommand{\theequation}{\mbox{KP}} \begin{equation}\la{kp} \left(-4 u_t+6 u u_x+u_{xxx}\right)_x+3 \sigma^2 u_{yy}=0, \end{equation} \setcounter{equation}{0} \renewcommand{\theequation}{\mbox{\arabic{equation}}} \noindent where $\sigma^2=\pm 1$ and the subscripts denote differentiation. Depending on the physical situation, one derives the equation either with $\sigma^2=-1$ or $\sigma^2=+1$. The resulting partial differential equations are referred to as (KP1) and (KP2) respectively. For real-valued solutions, the two equations have different physical meaning and different properties \cite{kadom}. Nevertheless, for some purposes the sign of $\sigma^2$ is irrelevant and we equate $\sigma \equiv 1$. In this case, we simply refer to ``the KP equation'' or to ``(KP)''. This amounts to working with (KP2). If necessary, (KP1) is obtained through a complex scaling of the $y$-variable. The KP equation can be written as the compatibility condition of two linear equations for an auxiliary wave function $\Psi$ \cite{dryuma, shabat}: \setcounter{saveeqn}{\value{equation} \begin{eqnarray}\la{linear1} \sigma \Psi_y&=&\Psi_{xx}+u \Psi,\\\la{linear2} \Psi_t&=&\Psi_{xxx}+\frac{3}{2} u \Psi_x+\frac{3}{4} \left(u_x+w\right) \Psi. \end{eqnarray} \setcounter{equation}{\value{saveeqn} Expressing the compatibility of \rf{linear1} and \rf{linear2}, $\Psi_{yt}\equiv \Psi_{ty}$, and eliminating $w$ results in \rf{kp}, if we assume that a complete basis for the wave function $\Psi$ exists. Note that if the KP solution is independent of $y$, the above Lax pair (\ref{linear1}, \ref{linear2}) reduces to the Lax pair for (KdV) \cite{ggkm} by simple separation of variables of the wave function $\Psi$. The two equations \rf{linear1}, \rf{linear2} are only two equations of an infinite hierarchy of linear evolution equations for the wave function $\Psi$ with respect to {\em higher-order time variables} $t_n$ \cite{sato}. We refer to this collection of linear flows as the {\em linear KP hierachy}: \begin{equation}\la{lkphier} \Psi_{t_k}=A_k \Psi,~~\mbox{for~}k=1,2,3,\ldots \end{equation} \noindent with $A_k$ a linear differential operator in $x$ of order $k$ with (usually) nonconstant coefficients. We refer to the $t_k$, $k=1,2,3,4, \ldots$ as higher-order time variables. A consequence of our definition of the KP hierarchy given in Section 3 is that $t_1$ can be identified with $x$. Furthermore, $y$ and $t$ are related to $t_2$ and $t_3$ respectively. By expressing the compatibility of these different linear flows, $\Psi_{t_{k_1}t_{k_2}}=\Psi_{t_{k_2}t_{k_1}}$, and assuming the existence of a complete basis for $\Psi$, we obtain an infinite number of nonlinear partial differential equations for the evolution of $u$, $w$ (and other functions referred to as potentials) with respect to the $t_k$ \cite{sato}, called the {\em KP hierarchy}: \begin{equation}\la{kphier} \pp{A_{k_1}}{t_{k_2}}-\pp{A_{k_2}}{t_{k_1}}=\left[A_{k_2},A_{k_1}\right], \end{equation} \noindent where $\left[A,B\right]\equiv AB-BA$. The linear KP hierarchy \rf{lkphier} and the KP hierarchy \rf{kphier} are the fundamental ingredients for the methods presented in this paper. A large family of quasiperiodic solutions of the KP equation was found by Krichever \cite{krich1, kricheverintegration}. Each of these solutions has a finite number of phases. They are given by \setcounter{saveeqn}{\value{equation} \begin{eqnarray}\la{reconstruction} u&=&\tilde{u}+2 \partial_x^2 \ln \Theta_g(\phi_1, \phi_2, \ldots, \phi_g | \mbf{B}),\\\la{phases} \phi_j&=&k_j x+l_j y+\omega_j t+\phi_{0j},~~\mbox{for~}j=1,2,\ldots, g, \end{eqnarray} \setcounter{equation}{\value{saveeqn} \noindent for some constants $\tilde{u}, k_j, l_j, \omega_j$ and $\phi_{0j}$, $j=1, 2,\ldots, g$. $\Theta_g$ is a {\em Riemann theta function} of {\em genus} g, parametrized by a $g \times g$ {\em Riemann matrix} $B$: \begin{equation}\la{thetafunction} \Theta_g(\mbf{\phi}| \mbf{B})\equiv \sum_{\mbf{m}\in \mbf{Z}^g} \exp\left(\frac{1}{2} \mbf{m} \cdot \mbf{B} \cdot \mbf{m}+i \mbf{m} \cdot \mbf{\phi}\right). \end{equation} \noindent Here $\mbf{\phi}\equiv(\phi_1, \ldots, \phi_g)$. The vector $\mbf{m}$ runs over all $g$-dimensional vectors with integer components. The Riemann matrix $\mbf{B}$ is a symmetric $g \times g$ matrix with negative definite real part. Whenever the matrix $\mbf{B}$ is obtained from a compact, connected Riemann surface in a standard way \cite{dub}, \rf{reconstruction} defines a solution of the KP equation \cite{krich1, kricheverintegration}. In what follows, the dependence of the theta function $\Theta_g$ on the Riemann matrix $\mbf{B}$ and the index $g$ denoting the number of phases will be surpressed for notational simplicity. By construction, $\Theta(\mbf{\phi})$ is periodic in each component of $\mbf{\phi}$. Hence the restriction to the linear winding \rf{phases} makes \rf{reconstruction} by definition a quasiperiodic function in $x$ or $y$ or $t$. A solution of the form \rf{reconstruction} is said to have genus $g$. In the terminology of Krichever and Novikov \cite{kricheverrank, krichrank}, all solutions of the form \rf{reconstruction} have {\em rank 1}. The problem addressed in this paper is to find $u(x,y,t)$ such that: \begin{equation}\la{problem} \left\{ \begin{array}{l} (-4 u_t+6 u u_x+u_{xxx})_x+3 \sigma^2 u_{yy}=0\\ u(x,y,0)=\mbox{rank 1, finite-genus solution of KP,}\\ \phantom{u(x,y,0)=}~\mbox{evaluated at~}t=0. \end{array} \right. \end{equation} \noindent The initial data are purposely not written in the form \rf{reconstruction}. The problem is not only to determine the frequencies $\omega_j$, for $j=1,2,\ldots, g$, but also to fix the other parameters $g, \mbf{B}, \tilde{u}, k_j, l_j, \phi_{0j}$, for $j=1,2,\ldots, g$: the acceptable class of initial data consists of all finite-genus solutions of the KP equations evaluated at a fixed time, with an unspecified but finite number of phases. A solution to this problem was offered in \cite{ds1}, where a seven-step algorithm was presented. This algorithm was a mixture of new ideas and known work by Krichever \cite{kricheverchap, krich1, kricheverintegration} and Previato \cite{previato}. The main idea of the algorithm is to provide the ingredients required for Krichever's inverse procedure for the reconstruction of a finite-genus solution of the KP equation \cite{krich1, kricheverintegration}, {\em i.e.} a compact connected Riemann surface and a divisor on this surface. In the present paper, a different algorithm to solve the problem \rf{problem} is presented. This algorithm shares some steps with that in \cite{ds1}. However, in contrast to the first algorithm, the second algorithm does not work towards Krichever's inverse procedure \cite{krich1, kricheverintegration}. The main idea here is to examine the structure of a set of ordinary differential equations obtained in step 5 of \cite{ds1}. In this paper, we show the following: \begin{itemize} \item The rank 1, finite-genus solutions of the KP equation are governed by a system of ordinary differential equations. This system is constructed explicitly. \item This system of ordinary differential equations is Lagrangian. \item With some care, the Lagrangian equations are written as a Hamiltonian system of ordinary differential equations in $x$. \item This Hamiltonian system of ordinary differential equations is completely integrable in the sense of Liouville \cite{arnold}. A complete set of conserved quantities in involution under the Poisson bracket is explicitly constructed. \item From these conserved quantities, one easily constructs completely integrable Hamiltonian systems of ordinary differential equations describing the evolution of the given initial condition of \rf{problem} under any of the higher-order time variables $t_k$, including $k=1,2,3$. This provides a solution of \rf{problem}. \end{itemize} In the next section, it is shown how the information listed above is used in an algorithm to solve problem \rf{problem}. As with the algorithm in \cite{ds1}, most of the steps of the algorithm in this paper are due to others. The work of Bogoyavlenskii and Novikov \cite{bogoyavlenskii} provided the main inspiration: the algorithm presented here is a generalization to the KP equation of their approach to solve problem \rf{problem} for the KdV equation. The work by Gel'fand and Dikii \cite{gd1}, Veselov \cite{veselov}, Adler \cite{adler} and Dikii \cite{dickey} was used to execute some of the steps of the algorithm. Although all of the above authors have considered parts of the problem considered here, to the best of our knowledge a complete solution of problem \rf{problem} using ordinary differential equations to solve the posed initial-value problem was never given. The algorithm presented here offers an alternative approach to that in \cite{ds1}. There are however some natural consequences of the new approach that do not appear from the approach in \cite{ds1}. These include \begin{itemize} \item {\bf Canonical variables} for the rank 1, finite-genus solutions of the KP equation. Any rank 1, finite-genus solution of the KP equation satisfies a Hamiltonian system of ordinary differential equations. The Poisson bracket on the phase space of this Hamiltonian system is the canonical Poisson bracket, resulting in a description of the rank 1, finite-genus solutions of the KP equation in terms of canonical (Darboux) variables. \item {\bf Conserved Quantities} for rank 1, finite-genus solutions \rf{reconstruction} of the KP equation. The Hamiltonian system of ordinary differential equations is completely integrable in the sense of Liouville. A sufficient number of conserved quantities is constructed explicitly. These conserved quantities are mutually in involution under the canonical Poisson bracket. \item {\bf Parameter count} of the theta-function solutions \rf{reconstruction} of the KP equation. It is known \cite{dub} that a generic\footnote{for a precise definition, see section \ref{sec:algorithm}} solution of genus $g$ of the KP equation is characterized by $4g+1$ independent parameters. We reconfirm this result, and extend it, by providing a parameter count for nongeneric solutions of the KP equation as well. Furthermore, the parameters naturally divide in two classes: parameters with ``dynamical significance'' and other parameters. The dynamically significant parameters are the initial conditions of the Hamiltonian system of ordinary differential equations describing the solution. The other parameters foliate the phase space of the Hamiltonian system. \item {\bf Minimal characterization of the initial data}. The approach presented here demonstrates that a finite-genus solution $u(x,y,t)$ of the KP equation is completely determined by a one-dimensional slice of the initial condition $u(x,y=0,t=0)$. In other words, it suffices to specify the initial data of \rf{problem} at a single $y$-value, say at $y=0$. \end{itemize} Krichever \cite{krich3} proposed another method to solve an initial-value problem for the KP2 equation with initial data that are spatially periodic (in both $x$ and $y$). Krichever's method is not restricted to initial data of finite genus, hence it is in that sense more general than the algorithm presented here. On the other hand, the methods of this paper require no restriction to periodic initial data. \section{Overview of the algorithm}\la{sec:algorithm} A solution for problem \rf{problem} is obtained using a seven-step algorithm. In this section, an overview of this algorithm is given, along with references to earlier work. \begin{enumerate} \item {\bf Determine the genus of the initial data} \cite{ds1} Let us rewrite \rf{phases} in the form \begin{equation}\la{newphases} \phi_j=\mbf{\kappa_j} \cdot \mbf{x}+\omega_j t+\phi_{0j}, ~~~~j=1,2,\ldots, g, \end{equation} \noindent with $\mbf{\kappa_j}=(k_j,l_j)$ and $\mbf{x}=(x,y)$. If all wave vectors $\mbf{\kappa_j}$ are incommensurable, {\em i.e.,} if there is no relationship \begin{equation}\la{commensurable} \sum_{i=1}^g n_i \mbf{\kappa_i}=0 \end{equation} for integers $n_i$ not all zero, then a two-dimensional Fourier transform of the initial data resolves the vectors $\mbf{\kappa_j}$. Because the initial data contain only a finite number of phases, the Fourier transform is necessarily discrete; {\em i.e.,} it consists of isolated spikes. Since the condition \rf{commensurable} almost never holds, we can almost always find the genus of the initial condition by counting the number of spikes in the Fourier transform, modulo harmonics. If condition \rf{commensurable} holds, then the prescribed method finds only a lower bound on the genus of the initial data. The method fails especially dramatically in one important special case: if the initial data are spatially periodic, then \rf{commensurable} holds automatically for any two wave vectors $\mbf{\kappa_i}$ and $\mbf{\kappa_j}$. The lower bound obtained in this case for the number of phases is 1. This problem was already pointed out in \cite{ds1}. If the initial data are spatially periodic, it is most convenient to impose that the genus of the initial data also be given, as part of the initial data. The method of Fourier transform to determine the genus of the initial condition has been used in \cite{hmss, currysegur}. \item {\bf Determine two stationary flows of the KP hierarchy: find $(r,n)$} Mulase \cite{mulase} and later Shiota \cite{shiota} showed that a rank 1, finite-genus solution of the KP equation \rf{reconstruction} is a simultaneous solution to all flows of the KP hierarchy by using \rf{reconstruction} with \begin{equation}\la{phaseshier} \phi_j=\sum_{i=1}^\infty k_{j,i} t_i, \end{equation} \noindent for $j=1,2,\ldots, g$, instead of \rf{phases}. Mulase and Shiota demonstrated that the corresponding rank 1, finite-genus solutions are stationary with respect to all but a finite number of the higher-order times in the KP hierarchy. A rank 1, finite-genus solution of the KP equation is said to be stationary with respect to $t_k$ if \begin{equation}\la{stat} \sum_{i=1}^k d_{i} \pp{u}{t_i}=0, \end{equation} \noindent with all the $d_{i}$ constant and $d_{k}=1$. The algorithm presented here requires the knowledge of two independent higher-order times of the KP hierarchy $t_r$ and $t_n$, such that $u$ is stationary with respect to both $t_r$ and $t_n$. First, $r$ is the minimal $k$ for which \rf{stat} holds for $k=r$. For this $r$, $n$ corresponds to the lowest-order higher-order time $t_n$, such that the $t_n$-flow is independent of the $t_r$-flow and \rf{stat} holds for $k=n$. In \cite{ds1}, a recipe was presented to find $(r,n)$, given the genus $g$ of the initial data. Actually, a finite number of pairs $(r,n)$ is determined for any given $g$. As we will see in step 4, each one of the resulting pairs $(r,n)$ gives rise to a set of ordinary differential equations, one of which the initial condition necessarily satisfies. The pairs $(r,n)$ for which the initial condition does not satisfy the differential equations need to be rejected. Hence, only at step 4 do we nail down a pair of stationary flows of the KP hierarchy for the given initial data. Here, at step 2, the numbers of pairs $(r,n)$ is reduced to a finite number. For initial data with $g$ phases, the following constraints on $(r,n)$ are known \cite{ds1}: \begin{itemize} \item All values of $r$ with $2 \leq r \leq g+1$ are allowed. \item For each $r$, let $n_j(r)$ be the $j$-th integer greater than $r$ that is coprime with $r$. The lowest $(g-r+2)$ of these integers are possible values of $n$. \item Exclude from the list of pairs $(r,n)$ obtained above the values of $n$ for which \linebreak $(r-1)(n-1)/2<g$. \item The remaining pairs $(r,n)$ are all possible for genus $g$. \end{itemize} \newpage \noindent {\bf Remarks}\la{remrem} \begin{enumerate} \item When $r=2$, the only possibility for $n$ is $n=2g+1$. This is the case corresponding to one-dimensional solutions. \item A solution of the KP equation of genus $g$ is called {\em generic} if the first $g$ vectors $\mbf{k_i}=(k_{1,i},k_{2,i}, \ldots, k_{g,i})$ are linearly independent. For a generic solution of genus $g$ of the KP equation, $(r,n)=(g+1, g+2)$. \item We have the following confusing situation: to find a generic genus $g$ KP solution, we need $(r,n)=(g+1,g+2)$. This choice leads to a Hamiltonian system of ordinary differential equations (in step 5). However, a generic solution of genus $g$ is not a typical solution of this Hamiltonian system ({\em i.e.,} it does not depend on a maximal number of parameters for this system). Typical solutions of the Hamiltonian system are nongeneric solutions of higher genus. Since these higher-genus solutions depend on more parameters, one must search carefully to find the generic genus $g$ solutions among them. This is further discussed in Sections \ref{sec:reductions} and \ref{sec:parameters}. \end{enumerate} \item {\bf Impose the $r$-reduction} For a given value of $r$, obtained from step 2, we impose on the KP hierarchy \rf{kphier} the reduction that the KP solution is independent of $t_r$. Hence, the coefficients of $A_k$, for all k are independent of $t_r$. Following Gel'fand and Dikii \cite{gd1}, Adler \cite{adler} and Strampp and Oevel \cite{strampp}, this allows us to rewrite the {\em $r$-reduced KP hierarchy} as an infinite ($r\times r$ matrix) hierarchy of partial differential equations, each with one space dimension, and all of them mutually commuting. In other words, \rf{kphier} is replaced by a hierarchy of the form \begin{equation}\la{11hier} \pp{B}{t_k}=\left[B_k,B\right], ~~~r~\mbox{does not divide}~k. \end{equation} The matrices $B$ and $B_k$ contain $r-1$ unknown functions. The higher-order time variables of the hierarchy in \rf{11hier} are inherited from the KP hierarchy \rf{kphier}. Only $t_r$ and the higher-order times of the form $t_{(ir)}$ for integer $i$ do not appear any more. In particular $t_1\equiv x$. Each equation of the hierarchy \rf{11hier} is Hamiltonian, as is shown in Section \ref{sec:rred}, where the details of the $r$-reduction are given. \item {\bf Impose the $n$-reduction} After imposing stationarity of the KP solution with respect to $t_r$, we now impose stationarity of the KP solution with respect to $t_n$ as well. Imposing the $n$-reduction in addition to the $r$-reduction leads to the {\em $(r,n)$-reduced KP equation}. The $(r,n)$-reduced KP equation is a system of $(r-1)$ ordinary differential equations in $x$ for $(r-1)$ unknown functions $\mbf{u}$. Again, following Gel'fand and Dikii \cite{gd1}, Adler \cite{adler} and Strampp and Oevel \cite{strampp}, we write the $(r,n)$-reduced KP equation in Lagrangian form: \begin{equation}\la{lagode} \dd{{\cal L}}{\mbf{u}}=\mbf{0}, \end{equation} \noindent where $\delta{{\cal L}}/\delta{\mbf{u}}$ denotes the variational derivative of ${\cal L}$ with respect to a certain vector function $\mbf{u}=(f_1, f_2, \ldots, f_{r-1})$, which is explicitly determined in terms of the solution of (KP): \begin{equation}\la{vardervec} \dd{{\cal L}}{\mbf{u}}\equiv\left(\dd{{\cal L}}{f_1}, \dd{{\cal L}}{f_2}, \ldots, \dd{{\cal L}}{f_{r-1}}\right)^T \end{equation} \noindent and for any function $f$, the {\em variational derivative} of ${\cal L}$ with respect to $f$ is defined as \begin{equation}\la{varder} \dd{}{f}{\cal L}(u,u_x, u_{xx}, \ldots)\equiv \sum_{k\geq 0} (-1)^k \ppn{k}{}{x} \pp{{\cal L}}{f^{(k)}}. \end{equation} \noindent Here, $f^{(k)}$ denotes the $k$-th derivative of $f$ with respect to $x$. Equations \rf{lagode} are a set of ordinary differential equations that the initial condition needs to satisfy. This constitutes a test on the validity of the pair $(r,n)$, chosen after step 2. The details of imposing the $n$-reduction in addition to the $r$-reduction are found in Section \ref{sec:nred} \item {\bf The Ostrogradskii transformation, canonical variables and the Hamiltonian system} In Section \ref{sec:ostro}, the Lagrangian system of ordinary differential equations in $x$ is transformed to a Hamiltonian system of ordinary differential equations in $x$ with canonical variables. Since the Lagrangian ${\cal L}$ depends on more than the first derivatives of $\mbf{u}$, an extension of the Legendre transformation is needed. This is the {\em Ostrogradskii transformation} \cite{ostro, whittaker}, defined in Section \ref{sec:ostro}. It defines {\em canonical variables} $\mbf{q}$ and $\mbf{p}$ in terms of the Lagrangian variables $\mbf{u}$: \begin{equation}\la{ostrosimple} \mbf{q}=\mbf{q}(\mbf{u}, \mbf{u}_x, \mbf{u}_{xx}, \ldots), ~~~ \mbf{p}=\mbf{p}(\mbf{u}, \mbf{u}_x, \mbf{u}_{xx}, \ldots). \end{equation} The Lagrangian ${\cal L}$ is called {\em nonsingular} \cite{krupkova, dfn2} if the Ostrogradskii transformation is invertible, {\em i.e.,} if the transformation \rf{ostrosimple} can be solved for the Lagrangian variables $\mbf{u}$ and their derivatives in terms of the canonical variables $\mbf{q}$ and $\mbf{p}$. If the Lagrangian is nonsingular, the Euler-Lagrange equations corresponding to the Lagrangian ${\cal L}$ are equivalent to the Hamiltonian system \begin{equation}\la{hamsyssimple} \pp{\mbf{q}}{x}=\pp{H}{\mbf{p}},~~~ \pp{\mbf{p}}{x}=-\pp{H}{\mbf{q}} \end{equation} \noindent where the Hamiltonian $H$ is determined explicitly in terms of the Lagrangian. If both $r$ and $n$ are odd, the Lagrangian ${\cal L}$ is shown to be singular in Section \ref{sec:sing}. Nevertheless, the dynamics in terms of the Lagrangian variables is still well-posed, as shown by Veselov \cite{veselov}. In Section \ref{sec:sing}, the singular Lagrangians are further investigated. We indicate how one might be able to avoid dealing with singular Lagrangians: a simple invertible transformation on the Lagrangian variables should be able to transform the singular Lagrangian into a nonsingular one. Otherwise, one can always resort to the more general methods of Krupkova \cite{krupkova} or to the theory of constrained Hamiltonian systems \cite{dirac}. \item {\bf Complete integrability of the Hamiltonian system} The Hamiltonian system \rf{hamsyssimple} is shown to be {\em completely integrable in the sense of Liouville} in Section \ref{sec:comp}. If the dimension of the vectors $\mbf{q}$ and $\mbf{p}$ is $N$, the Hamiltonian system is $2N$-dimensional. A set of $N$ functionally independent conserved quantities $T_{k}$ is constructed. Generalizing the work of Bogoyavlenskii and Novikov \cite{bogoyavlenskii}, these conserved quantities are shown to be mutually {\em in involution}, {\em i.e.,} \begin{equation}\la{involsimple} \left\{T_{k},T_{l} \right\}\equiv 0, \end{equation} \noindent where $\{f,g\}$ denotes the {\em Poisson bracket} of the functions $f$ and $g$: \begin{equation}\la{pb} \left\{f,g\right\}\equiv \pp{f}{\mbf{q}} \pp{g}{\mbf{p}}- \pp{f}{\mbf{p}} \pp{g}{\mbf{q}}. \end{equation} A consequence of proving the involutivity of the conserved quantities $T_{k}$ is that $T_k=-H_k$, where $H_k$ is the Hamiltonian describing the evolution of the canonical variables along the higher-order time variable $t_k$: \begin{equation}\la{hamsysksimple} \pp{\mbf{q}}{t_k}=\pp{H_k}{\mbf{p}},~~~ \pp{\mbf{p}}{t_k}=-\pp{H_k}{\mbf{q}}. \end{equation} The canonical variables are related to the dependent variable $u$ of the KP equation. Hence, we have constructed a set of ordinary differential Hamiltonian systems, each one of which describes the evolution of a rank 1, finite-genus solution of the KP equation according to a different higher-order time variable. Since all these Hamiltonian systems are $2N$-dimensional and share a common set of $N$ functionally independent conserved quantities $T_k$, mutually in involution, they are all completely integrable in the sense of Liouville. \item {\bf Solve the Hamiltonian system; reconstruct the solution of the KP equation} The final step of the algorithm is to integrate explicitly the Hamiltonian systems obtained in the previous step. From Liouville's theorem \cite{arnold} it is known that the Hamiltonian equations of motion can be solved in terms of quadratures. This last step is not executed in this paper. For the KdV equation it can be found in \cite{dickey}. Some partial results for the KP equation are also discussed there. \end{enumerate} \section{The KP hierarchy}\la{sec:hier} In this section, the KP hierarchy is redefined, using the terminology of Gel'fand and Dikii \cite{gd1}, Adler \cite{adler} and others. More specifically, the notation of Strampp and Oevel \cite{strampp} is used. Consider the pseudo-differential operator \begin{equation}\la{pseudo1} L=\partial+u_2\partial^{-1}+u_3 \partial^{-2}+u_4 \partial^{-3}+\ldots=\sum_{j=-\infty}^{1}u_{1-j}\partial^j, \end{equation} \noindent with $u_0\equiv 1$, $u_1\equiv 0$. We have used the notation $\partial=\partial_x$. The coefficients $u_j$ can be functions of $x$. The $u_j$ are referred to as {\em potentials}. This term is also used for any other set of functions, related to the $u_j$ by an invertible transformation. \vspace*{12pt} \noindent {\bf Remark}\la{u1rem} In order to compare with the results in \cite{ds1}, we need $u_1\neq 0$, but constant, extending the definition of the pseudo-differential operator $L$. Although this changes some of the formulas in this section, the added results for the KP equation are minor. In \cite{dickey} and \cite{decthesis}, it is shown that this amounts to assigning a fixed value to the constant $\tilde{u}$ in \rf{reconstruction}. In the remainder of the paper, we assume $u_1\equiv 0$, unless stated otherwise. \vspace*{12pt} The action of the operator $\partial^j$ is defined by the generalized Leibniz rule: \begin{equation}\la{leibniz} \partial^j f=\sum_{i=0}^\infty \binomial{j}{i} f^{(i)} \partial^{j-i}, \end{equation} \noindent where $f$ is a function, $f^{(i)}$ is its $i$-th derivative with respect to $x$, and the binomial coefficients are defined as \begin{equation}\la{binomial} \binomial{j}{i}=\frac{j(j-1)\cdots (j-i+1)}{i!}, ~~\mbox{for}~i>0, ~~\mbox{and}~\binomial{j}{0}=1. \end{equation} \noindent Note that this definition makes sense for negative integers $j$. For non-negative integers $j$, \rf{leibniz} is a finite sum. Otherwise, \rf{leibniz} results in an infinite series. Next, consider positive integer powers of the pseudo-differential operator $L$: \begin{eqnarray}\nonumber L^r&=&\left(\partial+u_2\partial^{-1}+u_3\partial^{-2}+u_4 \partial^{-3}+\ldots\right)^r\\\la{def} &=&\sum_{j=-\infty}^r \alpha_j(r)\partial^j=\sum_{j=-\infty}^r \partial^j \beta_j(r). \end{eqnarray} \noindent The last two equalities define the functions $\alpha_j(r)$ and $\beta_j(r)$, for $j \leq r$, $r>0$. These are in general functions of $x$. One has \begin{equation}\la{initialization} \alpha_1(1)=1, \alpha_0(1)=0, \alpha_j(1)=u_{j+3}, ~~~~\mbox{for}~~j=-1, -2, -3, \ldots, \end{equation} \noindent and \begin{equation} \alpha_r(r)=1, ~~\beta_r(r)=1, ~~~\mbox{and}~~~\alpha_{r-1}(r)=0,~~ \beta_{r-1}(r)=0. \end{equation} \noindent Clearly the functions $\alpha_{j}(r)$ and $\beta_j(r)$ are related. Using \rf{leibniz}, we get from \rf{def} that \begin{equation}\la{triangular} \alpha_j(r)=\sum_{k=0}^{r-j}\binomial{j+k}{k} \beta_{j+k}^{(k)}(r). \end{equation} \noindent This triangular system can be solved to obtain the functions $\beta_j(r)$ in terms of the functions $\alpha_j(r)$, if so desired. Note in particular that $\alpha_{-1}(r)=\beta_{-1}(r)$, since the binomial coefficient \rf{binomial} vanishes for positive $j$ less than $i$. The functions $\alpha_j(r)$ can be determined explicitly in terms of the potentials $(u_2, u_3, \ldots)$. A convenient way to do this is to use a recursion relationship obtained from $L^r=LL^{r-1}$: \begin{equation}\la{recursion} \alpha_j(r)=\alpha_{j-1}(r-1)+\pp{}{x}\alpha_j(r-1)+u_{r-j}+\sum_{k=j-r+3}^{-1} u_{1-k} \sum_{m=j}^{k+r-3}\binomial{k}{m-j}\alpha_{m-k}^{(m-j)}(r-1). \end{equation} \noindent It is possible to obtain an explicit formula which expresses $\alpha_j(r)$ in terms of only the potentials $(u_2, u_3 \ldots)$ \cite{decthesis}, but such a formula is not as practical as the recursion relationship \rf{recursion}. However, the following result will be used. It extends a result of Date $et~al$ \cite{date}: \begin{equation}\la{linearpart} \alpha_j(r)=\sum_{k=1}^{r-j-1}\binomial{r}{k} \ppn{k-1}{}{x} u_{r-j-k+1}+\hat{\alpha}_j(r), \end{equation} \noindent and $\hat{\alpha}_j(r)$ is a differential polynomial in $(u_2,u_3, \ldots, u_{r-j-2})$ containing only nonlinear terms. This follows easily from \rf{recursion}. The differential part (including the purely multiplicative term) of the operator $L^r$ is denoted by $L^r_+$: \begin{equation}\la{positive} L^r_+=\sum_{j=0}^r \alpha_j(r)\partial^j=\sum_{j=0}^r \partial^j \beta_j(r). \end{equation} \noindent Observe from \rf{triangular} that the purely differential part of $L^r$ is independent of the representation \rf{def} used for $L^r$. This is also true for \begin{equation}\la{negative} L^r_-=L^r-L^r_+=\sum_{j=-\infty}^{-1} \alpha_j(r)\partial^j=\sum_{j=-\infty}^{-1} \partial^j \beta_j(r). \end{equation} Having introduced the above notation, the KP hierarchy is expressed quite easily. Consider the linear evolution equation for the wave function $\Psi$ \begin{equation}\la{laxpsi} \pp{\Psi}{t_r}=L^r_+ \Psi, ~~~~\mbox{for}~r=1,2,\ldots. \end{equation} \noindent This is the {\em linear KP hierarchy} ({\em cf.} \rf{lkphier}). For $r=1$, this equation given $\Psi_{t_1}=\Psi_x$, hence the identification $t_1\equiv x$. Assuming completeness of states, the {\em KP hierarchy} is obtained from the compatibility of the equations in \rf{laxpsi} ({\em cf.} \rf{kphier}): \begin{equation}\la{laxkp} \frac{\partial^2 \Psi}{\partial_{t_{r_1}} \partial_{t_{r2}}}= \frac{\partial^2 \Psi}{\partial_{t_{r_2}} \partial_{t_{r1}}}\Rightarrow \pp{L_+^{r_1}}{t_{r_2}}-\pp{L_+^{r_2}}{t_{r_1}}=\left[L_+^{r_2},L_+^{r_1}\right]. \end{equation} \noindent These equations determine how the potentials depend on the higher-order time variables $t_r$ so that the equations \rf{laxpsi} are compatible. Again assuming completeness of states, equations \rf{laxkp} can also be obtained from the compatibility of the following sequence of Lax-like equations \begin{equation}\la{lax} \pp{L}{t_r}=\left[L_+^r,L\right]=\left[L,L_-^r\right], ~~r\geq 1. \end{equation} \noindent The last equality is a consequence of $L^r=L^r_++L_-^r$. Introducing the KP hierarchy as in \rf{lax} is equivalent to the approach used in \cite{ds1}. Below, we use that \rf{laxpsi} is essentially equivalent to \rf{lax}. Our approach consists mainly of rewriting \rf{lax} and its reductions. Each time we increase $r$ by one, another potential appears in $L^r_+=L^r_+(u_2, u_3, \ldots, u_r)$. Furthermore, $u_r$ appears only in $\alpha_0(r)$: $\alpha_{0}(r)=r u_r+\tilde{\alpha}_0(r;u_2, u_3, \ldots, u_{r-1})$, as is seen from \rf{recursion}. As a consequence, there is a one-to-one correspondence between the potentials $u, w_1, w_2, \ldots$ appearing in the KP hierarchy as it is defined in \cite{ds1} and the set of potentials $u_2, u_3, u_4,\ldots$ appearing in \rf{pseudo1}. As an example, consider equations (\ref{linear1}, \ref{linear2}). These are contained in the formulation of the KP hierarchy given here: writing out \rf{laxpsi} for $r=2$ and $r=3$ and equating coefficients with \rf{linear1} and \rf{linear2} respectively gives $u_2=u/2$ and $u_3=w/4-u_x/4$. The explicit form of the Lax equations \rf{lax} is needed later on. We have \cite{strampp} \begin{equation}\la{laxexp} \pp{u_i}{t_r}=\sum_{j=1}^{i} M_{i,j}\beta_{-j}(r), ~~~~i=0,1,2, \ldots, \end{equation} \noindent where the differential operator $M_{i,j}$ is given by \begin{equation}\la{opera} M_{i,j}=\sum_{k=0}^{i-j}\left(\binomial{1-j-k}{i-j-k}u_k \partial^{i-j-k}-\binomial{-j}{i-j-k}\partial^{i-j-k}u_k\right). \end{equation} \noindent Here and in what follows, the contribution of a sum is assumed to be zero if its upper limit is less than its lower limit, as happens in \rf{opera} when $i=0$. Note that this immediately gives $\partial u_0/\partial t_r=0$ and $\partial u_1/\partial t_r=0$, for all $r$, as expected. Furthermore, $M_{i,i}=0, M_{i,i-1}=\partial$. The differential equations \rf{laxexp} determine a first-order system for the $t_r$ evolution of the infinite-dimensional vector of potentials $(u_2, u_3, u_4, \ldots)$. \section{Impose the $r$-reduction}\la{sec:rred} Next we obtain a closed first-order system of partial differential equations for finitely many of the potentials, by imposing an {\em $r$-reduction}. This is the first reduction step in our scheme towards our goal of finding a set of ordinary differential equations describing the rank 1, finite-genus solutions of (KP). The $r$-reduction of the operator $L$ is obtained by imposing that the $r$-th power of $L$ is purely differential: \begin{equation}\la{rred} L^r=L_+^r ~~\mbox{or}~~L^r_-=0~\Rightarrow~\beta_k(r)\equiv 0~\mbox{for}~k<0~\Rightarrow~\alpha_k(r)\equiv 0~\mbox{for}~k<0. \end{equation} \noindent Notice that the $r$-reduction implies immediately that all potentials are independent of $t_r$, from \rf{laxexp} or \rf{lax}. The $r$-reduction determines the potentials $u_{r+1}, u_{r+2}, u_{r+3}, \ldots$ as differential polynomials of the potentials $u_2, u_3, \ldots, u_{r}$. This is a consequence of the triangular structure of the system relating the potentials $(u_2, u_3, u_4, \ldots)$ to the potentials $(\alpha_{r-2}(r), \alpha_{r-3}(r), \ldots)$. \vspace*{12pt} \noindent {\bf Remark}\la{p:rem2} If we impose an $r$-reduction, for some positive integer number $r$, then we have automatically achieved an $rk$ reduction, for any positive integer $k$. If $L^r$ is purely differential, then so is $L^{rk}=(L^r)^k$. \vspace*{12pt} Under $r$-reduction, the infinite system of evolution equations in \rf{laxexp} reduces to a finite number of equations for the independent potentials $(u_2, u_3, \ldots, u_r)$. We write this finite system in Hamiltonian form. First, we write the system in matrix-operator form. The matrices involved are now finite-dimensional, as there are only $r-1$ independent potentials. For a given $n$, define the $(r-1)$-dimensional vectors \begin{equation}\la{capu} \mbf{U}(r)=\left( \begin{array}{c} u_2\\u_3\\u_4\\\vdots\\u_{r} \end{array} \right), ~~~ \mbf{\beta}(r,n)=\left( \begin{array}{c} \beta_{-1}(n)\\\beta_{-2}(n)\\\beta_{-3}(n)\\\vdots\\\beta_{-r+1}(n) \end{array} \right), \end{equation} \noindent and the operator-valued matrix \begin{equation}\la{opermat} \mbf{M}(r)=\left( \begin{array}{ccccc} M_{2,1}&0&0&\cdots&0\\ M_{3,1}&M_{3,2}&0&\cdots&0\\ \vdots&\vdots&\vdots&\ddots&\vdots\\ M_{r-1,1}&M_{r-1,2}&M_{r-1,3}&\cdots&0\\ M_{r,1}&M_{r,2}&M_{r,3}&\cdots&M_{r,r-1} \end{array} \right), \end{equation} \noindent with the operators $M_{i,j}$ defined by \rf{opera}. Then under $r$-reduction \rf{laxexp} can be written as \begin{equation}\la{explicitlaxrred} \pp{\mbf{U}(r)}{t_n}=\mbf{M}(r)\mbf{\beta}(r,n). \end{equation} In order to write the system of equations \rf{explicitlaxrred} in Hamiltonian form, we first introduce new coordinates on the phase space of the system. Define \begin{equation}\la{capalpha} \mbf{\alpha}(r)=\left( \begin{array}{c} \alpha_0(r)\\\alpha_{1}(r)\\\vdots\\\alpha_{r-2}(r) \end{array} \right). \end{equation} \noindent The Jacobian matrix of the transformation from the coordinates $\mbf{U}(r)\rightarrow \mbf{\alpha}(r)$ is the Fr\'{e}chet derivative of the transformation (which depends also on the spatial derivatives of the original coordinates, see for instance \rf{recursion}). The Jacobian for such a transformation is an operator-valued matrix, whose action on an arbitrary vector of functions $\mbf{v}=(v_2, v_3, \ldots, v_n)^T$ is given by \begin{eqnarray}\nonumber \mbf{D}(r) \mbf{v}&=&\pp{\mbf{\alpha}(r)}{\mbf{U}(r)} \mbf{v}\\\la{jacobian} &=&\left.\pp{}{\epsilon}\left( \begin{array}{c} \alpha_0(r)(u_2+\epsilon v_2, u_3+\epsilon v_3, \ldots, u_r+\epsilon v_r)\\ \alpha_1(r)(u_2+\epsilon v_2, u_3+\epsilon v_3, \ldots, u_r+\epsilon v_r)\\ \vdots\\ \alpha_{r-2}(r)(u_2+\epsilon v_2, u_3+\epsilon v_3, \ldots, u_r+\epsilon v_r)\\ \end{array} \right)\right|_{\epsilon=0}. \end{eqnarray} \noindent This Jacobian matrix is upper-triangular. This is a direct consequence of the triangular structure of the system relating the potentials $u_j$ to the potentials $\alpha_j(r)$, $j=2,3,\ldots, r$. We rewrite the equations \rf{explicitlaxrred} in terms of the coordinates $\mbf{\alpha}(r)$: \begin{equation}\la{first} \pp{\mbf{\alpha}(r)}{t_n}=\pp{\mbf{\alpha}(r)}{\mbf{U}(r)} \pp{\mbf{U}(r)}{t_n}=\mbf{D}(r) \mbf{M}(r) \mbf{\beta}(r,n)=\mbf{J}(r) \mbf{\beta}(r,n), \end{equation} \noindent where we introduce the operator-valued matrix $\mbf{J}(r)=\mbf{D}(r) \mbf{M}(r)$. Note that $\mbf{J}(r)$ is always upper-triangular. This follows from the upper-triangular structure of $\mbf{D}(r)$ and of the lower-triangular structure of $\mbf{M}(r)$. Next we rewrite the vector $\mbf{\beta}(r,n)$. We use an identity from the calculus of exterior derivatives for pseudo-differential operators \cite{manin1}: \begin{equation}\la{manin1} d \beta_{-1}(r+n)=\frac{r+n}{r} \sum_{j=-1-n}^{r-2} \beta_{-1-j}(n) d \alpha_{j}(r), \end{equation} \noindent which gives \begin{equation}\la{manin2} \beta_{-j}(n)=\frac{r}{r+n}\dd{\beta_{-1}(r+n)}{\alpha_{j-1}(r)}, \end{equation} \noindent where $\delta/\delta \alpha_{j-1}(r)$ is the {\em variational derivative} with respect to $\alpha_{j-1}(r)$. Hence \begin{equation}\la{maninfinal} \mbf{\beta}(r,n)=\frac{r}{r+n} \dd{}{\mbf{\alpha}(r)} \beta_{-1}(r+n). \end{equation} \noindent Equations \rf{first} become \begin{equation}\la{second} \pp{\mbf{\alpha}(r)}{t_n}=\frac{r}{r+n}\mbf{J}(r) \dd{}{\mbf{\alpha}(r)} \beta_{-1}(r+n). \end{equation} \noindent This set of equations is Hamiltonian \cite{strampp}, with Hamiltonian \begin{equation}\la{hamil} H(r,n)=\frac{r}{r+n} \beta_{-1}(r+n)=\frac{r}{r+n} \alpha_{-1}(r+n). \end{equation} \noindent (We have used the observation that $\alpha_{-1}(r)=\beta_{-1}(r)$.) It suffices to prove that the operator $J(r)$ is Hamiltonian \cite{anton}, $i.e.$, that the operator $\mbf{J}(r)$ defines a Poisson bracket. This Poisson bracket is given by \cite{strampp} \begin{equation}\la{pbpde} \{S,T\}=\left(\dd{S}{\mbf{\alpha}(r)}\right)^T \mbf{J}(r) \left(\dd{T}{\mbf{\alpha}(r)}\right). \end{equation} Denote by ${\cal H}$ the quotient space of all smooth functionals of the potentials in $\mbf{\alpha}(r)$, modulo total derivatives with respect to $x$. \noindent For \rf{pbpde} to define a Poisson bracket on ${\cal H}\times {\cal H}$, we need three properties: \begin{enumerate} \item {\bf bilinearity:} This is obvious. \item {\bf skew-symmetry:} This is less obvious. Notice that the functions appearing in the bracket \rf{pbpde} appear only through their variational derivatives. Hence, these functions are only defined up to total derivatives with respect to $x$, $i.e.$, they are elements of $\cal H$. The skew-symmetry of the Poisson bracket \rf{pbpde} operating on ${\cal H}\times {\cal H} $ is then easily obtained by integration by parts. \item {\bf Jacobi identity:} This is also not obvious. The proof can be found in \cite{strampp}. There it is shown that the above bracket is essentially the bracket Adler defines in \cite{adler}. The proof of the Jacobi identity then reduces to Adler's proof. \end{enumerate} The Hamiltonian system of PDE's \rf{second} describes a whole hierarchy of Hamiltonian PDEs for a fixed $r$. All the members of the hierarchy have the same Poisson structure with different Hamiltonians: for each $n$ coprime with $r$, a different system of Hamiltonian partial differential equations is obtained, describing the $t_n$-evolution of the potentials. Note that the first member of every hierarchy is trivial. From the first flow of \rf{lax}, we get \begin{equation}\la{trivial} \pp{L}{t_1}=\left[L_+,L\right]=\left[\partial,L\right]=\partial L-L \partial=\pp{L}{x}+L \partial-L \partial=\pp{L}{x}. \end{equation} \noindent which is the same for all $r$-reductions. Hence, the first member of every hierarchy is $\partial \mbf{\alpha}(r)/\partial t_1=\partial \mbf{\alpha}(r)/\partial x$. For example, choosing $r=2$ results in the KdV hierarchy with Poisson operator $J(2)=2 \partial$. Choosing $r=3$ gives the Boussinesq hierarchy \cite{zak,mckean1} with Poisson operator \begin{equation}\la{operjbous} J(3)=\left(\begin{array}{cc}0 & 3 \partial\\3 \partial & 0\end{array}\right). \end{equation} Some remarks are in order about the Hamiltonian PDE's \begin{equation}\la{hamsys} \pp{\mbf{\alpha}(r)}{t_n}=\mbf{J}(r) \dd{H(r,n)}{\mbf{\alpha}(r)}. \end{equation} \noindent {\bf Remarks} \begin{description} \item[(a)~] Note that in contrast with the Poisson brackets in \cite{gardner, zak, mckean1}, the bracket \rf{pbpde} is local, $i.e.$, it does not involve integrations. This is an immediate consequence of working in the quotient space ${\cal H}$. The Poisson bracket \rf{pbpde} for $r=2$ and $r=3$ are the integrands of the brackets introduced in \cite{gardner} and \cite{zak, mckean1} respectively. \item[(b)~] Bogoyavlenskii and Novikov \cite{bogoyavlenskii} considered only the Korteweg-deVries equation. As a consequence, none of the algebraic machinery of this and the previous section was required for their approach. Their starting point was the KdV equivalent of \rf{hamsys}. This same starting point is used if one considers any other integrable partial differential equation which has only one spatial dimension, such as the nonlinear Schr\"{o}dinger equation, the modified Korteweg-deVries equation, etc. By starting with a one-dimensional partial differential equation, the first step (imposing the $r$-reduction) is skipped. It is for this step that the algebraic language of the previous two sections is required. \item[(c)~] An infinite number of conserved quantities exists for each member of the hierarchy given by \rf{hamsys}. This is a necessary condition for the integrability of these partial differential equations. Adler \cite{adler} showed that the different members of the $r$-reduced hierarchy define mutually commuting flows. The infinite set of Hamiltonians $\{H(r,n): n\geq 1\}$ is a set of conserved densities for every member of the hierarchy. That different members of the hierarchy \rf{hamsys} commute is expressed as the commutation of their respective Hamiltonians under the Poisson bracket \rf{pbpde}: \begin{equation}\la{pbcommute} \left\{H(r,k_1),H(r,k_2)\right\}=0, \end{equation} for a fixed $r$ and all $k_1, k_2$. Denote the solution operator of the $n$-th member of the hierarchy \rf{hamsys} by $\mbf{K}_n(t_n)$. In other words, given initial conditions $\mbf{\alpha}(r)(x,t_n=0)$, the solution $\mbf{\alpha}(r)(x,t_n)$ for any $t_n$ is written as \begin{equation} \mbf{\alpha}(r)(x,t_n)=\mbf{K}_n(t_n) \mbf{\alpha}(r)(x,0). \end{equation} \noindent Adler's statement \cite{adler} that different flows in the hierarchy \rf{hamsys} commute is then expressed as \begin{equation} \mbf{K}_n(t_n) \mbf{K}_m(t_m)=\mbf{K}_m(t_m) \mbf{K}_n(t_n). \end{equation} \item[(d)~] The Hamiltonian operator $\mbf{J}(r)$ is usually degenerate, $i.e.$, its kernel is not empty. Adler \cite{adler} showed that the variational derivatives of the elements of the set \linebreak $\{\alpha_{-1}(r+n): -r+1 \leq n\leq -1\}$ are all annihilated by $\mbf{J}(r)$. In other words, these elements are Casimir functionals for the flows generated by the $r$-reduction. It is easy to see from the triangular form of $\mbf{J}(r)$ that the dimension of its kernel is exactly $r-1$. This implies that the set of Casimir functionals found by Adler forms a complete basis for the kernel of $\mbf{J}(r)$ (see also \cite{veselov}). \end{description} \section{Impose the $n$-reduction}\la{sec:nred} Next, we consider stationary solutions of the system \rf{hamsys}, for the $n$ value determined in step 2. Hence, from \rf{stat}, \begin{equation}\la{stathamsys} \sum_{k=1}^n d_k \mbf{J}(r) \dd{H(r,k)}{\mbf{\alpha}(r)}=0 ~\Rightarrow~\mbf{J}(r) \dd{}{\mbf{\alpha}(r)}\sum_{k=1}^n d_k H(r,k)=0, \end{equation} \noindent with $d_n=1$. Furthermore, without loss of generality, $d_{n-1}$ can be equated to 0, as was shown in \cite{krich4, ds1}. The following theorem was first proved for the KdV equation by Lax \cite{lax1} and Novikov \cite{novikov}. \begin{theo}\la{theo:statpoint} The set of stationary solutions with respect to $t_n$ is invariant under the action of any of the other higher-order flows. \end{theo} \noindent {\bf Proof} Consider the hierarchy of mutually commuting Hamiltonian systems \begin{equation}\la{hamsysgen} \pp{\mbf{\alpha}(r)}{\tilde{t}_n}=\mbf{J}(r)\dd{}{\mbf{\alpha}(r)}\left(\sum_{k=1}^{n} d_k H(r,k)\right). \end{equation} Denote the solution operator of the $n$-th member of this hierarchy as $\tilde{\mbf{K}}_n(\tilde{t}_n)$. Clearly these solution operators commute with the solution operators of the higher-order flows, $\mbf{K}_m(t_m)$: \begin{equation}\la{comgen} \tilde{\mbf{K}}_n(\tilde{t}_n) \mbf{K}_m(t_m)= \mbf{K}_m(t_m) \tilde{\mbf{K}}_n(\tilde{t}_n). \end{equation} \noindent A stationary solution with respect to $t_n$ is a fixed point of the operator $\tilde{\mbf{K}}_n(\tilde{t}_n)$:\linebreak $\tilde{\mbf{K}}_n(\tilde{t}_n) \mbf{\alpha}(r)(x)=\mbf{\alpha}(r)(x)$. Hence \begin{eqnarray} \mbf{K}_m(t_m)\tilde{\mbf{K}}_n(\tilde{t}_n) \mbf{\alpha}(r)(x)&=& \mbf{K}_m(t_m)\mbf{\alpha}(r)(x)\\ \Rightarrow~~\tilde{\mbf{K}}_n(\tilde{t}_n)\mbf{K}_m(t_m) \mbf{\alpha}(r)(x)&=& \mbf{K}_m(t_m)\mbf{\alpha}(r)(x), \end{eqnarray} \noindent since the two operators commute. Hence $\mbf{K}_m(t_m)\mbf{\alpha}(r)(x)$ is a fixed point of $\tilde{\mbf{K}}_n(\tilde{t}_n)$ and hence a stationary solution with respect to $t_n$. \hspace*{\fill}$\rule{3mm}{3mm}$ Let us examine the structure of the equations \rf{stathamsys}, determining the stationary solutions with respect to $t_n$. From \rf{stathamsys}, $\dd{}{\mbf{\alpha}(r)}\sum_{k=1}^n d_k H(r,k)$ is in the kernel of the Poisson operator $\mbf{J}(r)$. Hence it is a linear combination of the Casimir functionals: \begin{equation}\la{cashamsys} \dd{}{\mbf{\alpha}(r)}\sum_{k=1}^n d_k H(r,k)+\sum_{k=1}^{r-1}h_k \frac{r}{k} \dd{\alpha_{-1}(k)}{\mbf{\alpha}(r)}=0; \end{equation} \noindent the coefficient of the Casimir functional $\alpha_{-1}(r)$ has been written as $h_k r/k$ for convenience. Equation \rf{cashamsys} is a system of Euler-Lagrange equations, with Lagrangian depending on the $r-1$ potentials $\mbf{\alpha}(r)$ and their derivatives: \begin{eqnarray}\nonumber {\cal L}(r,n)&=&H(r,n)+\sum_{k=1}^{n-2}d_k H(r,k)+\sum_{k=1}^{r-1} h_k \frac{r}{k} \alpha_{-1}(k)\\\la{lagrangian} &=&H(r,n)+\sum_{k=1}^{n-2}d_k H(r,k)+\sum_{k=1}^{r-1} h_k H(r,k-r), \end{eqnarray} \noindent since $d_{n-1}\equiv 0$. The last term in this equation is a slight abuse of notation. It is to be interpreted using \rf{hamil}. The set of $r-1$ Euler-Lagrange equations \rf{cashamsys} \begin{equation}\la{el} \dd{{\cal L}(r,n)}{\mbf{\alpha}(r)}=0 \end{equation} \noindent will be referred to as the {\em $(r,n)$-th (stationary) KP equation}. This system of Euler-Lagrange equations is extremely important: it is a finite-dimensional system of {\em ordinary differential equations} describing how solutions of the $(r,n)$-th KP equation depend on $x$. At this point, the study of rank 1, finite-genus solutions of (KP) is immensely simplified: it is reduced to the study of a finite-dimensional set of ordinary differential equations \rf{el} that are derived from one scalar quantity, the Lagrangian \rf{lagrangian}. The remainder of this paper examines the special structure of the set of equations \rf{el}. \vspace*{12pt} \noindent {\bf Remarks} \begin{description} \item[(a)~] As pointed out before, the first flow of the KP hierarchy defines $t_1$ to be $x$. This first flow imposes no constraints on the $x$-dependence of the potentials. After one imposes the $r$- and $n$-reductions this $x$-dependence is determined by the Lagrangian system \rf{el}. \item[(b)~] The Euler-Lagrange equations are a {\em minimal} set of differential equations the potentials in $\mbf{\alpha}(r)$ have to satisfy to make the $t_r$-flow and the $t_n$-flow stationary. In step 5 of \cite{ds1}, a system of differential equations was proposed which the initial conditions of the KP equation needs to satisfy. Because of the way the $r$- and $n$-reductions were performed, it was not clear in \cite{ds1} that these differential equations are in fact {\em ordinary} differential equations. Furthermore, the differential equations obtained in \cite{ds1} were not necessarily functionally independent. The equations \rf{el} are functionally independent, as was shown by Veselov \cite{veselov}. \item[(c)~] Since the order of imposing the $r$-reduction and the $n$-reduction can be reversed (in \cite{ds1} they were executed simultaneously), the remark on page~\pageref{p:rem2} can be repeated with $n$ instead of $r$: if we impose an $n$-reduction, we automatically achieve an $nk$ reduction for any positive integer $k$. Imposing an $n$-reduction implies that $L^n$ is purely differential. In that case $L^{nk}=(L^n)^k$ is also purely differential. \item[(d)~] At this point, {\em Krichever's criterion \cite{krichcom}} surfaces\footnote{There is some discussion in the literature about the necessity of this criterion. See for instance \cite{kasman}}: for a finite-genus solution of rank 1 of the KP equation, $r$ and $n$ need to be coprime. Non-coprime $r$ and $n$ result in higher-rank solutions. We show next that to determine a rank 1, finite-genus solution completely, non-coprime $r$ and $n$ are not allowed. Imposing the $r$- and $n$-reductions amounts (up to including lower order flows) to imposing that both $L^r$ and $L^n$ are purely differential operators. If $r$ and $n$ are not coprime, let $r=k\hat{r}$ and $n=k \hat{n}$, for integers $k, \hat{r}$ and $\hat{n}$. So, $r$ and $n$ have the common factor $k$. If $L^k=L^k_+$ ($i.e.$, $L^k$ is purely differential) then the solution is stationary with respect to $t_k$. Since $L^r=(L^k)^{\hat{r}}$ and $L^n=(L^k)^{\hat{n}}$ are purely differential, the solution is trivially stationary with respect to $t_r$ and $t_n$. Thus, imposing stationarity with respect to $t_k$ implies stationarity with respect to $t_r$ and $t_n$. Imposing stationarity with respect to only one higher order flow $t_k$ however, does not provide enough information for the determination of the solution using our methods. Therefore, $r$ and $n$ are required to be coprime. \end{description} \section{The explicit dependence of the Lagrangian on the potentials and their derivatives}\la{sec:exp} We want to examine the explicit functional dependence of the Lagrangian ${\cal L}(r,n)$ on the potentials $\mbf{\alpha}(r)=(\alpha_0(r), \alpha_1(r), \ldots, \alpha_{r-2}(r))^T$ and their derivatives. We are especially interested in the order of the highest derivative of $\alpha_i(r)$, for $i=0, 1, \ldots, r-2$. This information is necessary in order to carry out the generalization of the Legendre transformation in Section \ref{sec:ostro}. \vspace*{12pt} \noindent {\bf Definition:} {\bf The weight of an operator f(x), W[f(x)]}, is defined to be an integer-valued functional with the following properties: \begin{enumerate} \item $W[f g]=W[f]+W[g]$, and $W[f^N]=N W[f]$ for integer N. \item If $W[f]=W[g]$ then $W[f \pm g]=W[f]=W[g]$. \item $W[L]=1$. \end{enumerate} \vspace*{12pt} The usefulness of this definition is connected with the scaling symmetry of the KP equation. This symmetry is shared by the whole KP hierarchy through its definition using the operator $L$. Introducing the weight function turns the algebra of operators used here into a so-called graded algebra \cite{dickey}. Essentially, the weight function introduced here is identical with the `rank' introduced by Miura, Gardner and Kruskal in \cite{mgk}. We use the name `weight' because `rank' has a very different meaning in KP theory \cite{nmpz, krichcom}. Using the defining properties of the weight function, we calculate the weight of some quantities we have used: \vspace*{12pt} \noindent {\bf Examples} \begin{itemize} \item Since $W[L]=1$, any term in $L$ has weight 1. In particular $W[\partial]=1$. \item Hence $W[u_k \partial^{-k+1}]$ $=$ $1 \Rightarrow W[u_k]+W[\partial^{-k+1}]=1 \Rightarrow$ \newline $W[u_k]+(-k+1)W[\partial]=1 \Rightarrow W[u_k]=k$. \item $W[L^r]=r$, hence $W[\alpha_k(r) \partial^k]=r~\Rightarrow$~ $W[\alpha_k(r)]+W[\partial^k]=r~\Rightarrow~W[\alpha_k(r)]=r-k$. Analogously $W[\beta_{k}(r)]=r-k$. \item $W[\partial/\partial t_r]=W[L^r]=r$. \item $W[H(r,n)]=r+n+1$, from \rf{hamil}. Then also, $W[{\cal L}(r,n)]=r+n+1$. \item $W[d_k]=n-k$ and $W[h_k]=r+n-k$. \end{itemize} \vspace*{12pt} We now use the weight function to calculate how the Lagrangian depends on the $r-1$ potentials in $\alpha(r)$ and their derivatives. For $j=2, 3, \ldots, r$, let us denote by $N_j$ the highest order of differentiation of the potential $\alpha_{r-j}(r)$ in the Lagrangian \rf{lagrangian}. We say a term in the Lagrangian is of degree $M$ if it contains $M$ factors (counting multiplicities) that are linear in one of the potentials or in one of their derivatives. The Lagrangian has linear terms ($i.e.$, terms of degree one), quadratic terms ($i.e.$, terms of degree two), cubic terms ( $i.e.$, terms of degree three), terms of degree four and so on. Clearly the linear terms can not depend on the derivatives of the potentials: such a term would be a total derivative and would be disregarded. All terms in the Lagrangian have the same weight, as a consequence of the scaling invariance of the KP hierarchy. In order to find the highest derivative of a potential, we need only consider the quadratic terms of the Lagrangians. All higher degree terms have contributions from more than one other potential. The nontrivial potential with the lowest weight is $\alpha_{r-2}(r)$: $W[\alpha_{r-2}(r)]=2$. Every time a potential appears in a term of the Lagrangian, the number of $x$-derivatives of the other potentials in that term decreases by at least 2. Since the linear terms cannot contain derivatives, it follows that the highest derivatives of the potentials are found in the quadratic terms of the Lagrangian. Similarly, every time one of the coefficients $h_k$, for $k=1, 2, \ldots, r-1$ or $d_j$ for $j=1,2,\ldots, n-2$ appears in a term of the Lagrangian, the number of $x$-derivatives in that term has to decrease by the weight of the coefficient $h_k$ or $d_j$ involved. It follows that the highest derivatives of the potentials are found in the quadratic terms of the Lagrangian, not containing any of these coefficients, $i.e.$, in the quadratic terms of $H(r,n)$. We use these observations to find how the Lagrangian depends on the highest derivatives of the potentials, $i.e.$, to find $N_j$, for $j=2,3,\ldots, r$. \begin{theo}\la{theo:lagdep} The order of differentiation in the Lagrangian ${\cal L}(r,n)$ of $\alpha_{r-2}(r)$ is $[(r+n-3)/2]$, $i.e.$, $N_2=[(r+n-3)/2]$. The order of differentiation in the Lagrangian ${\cal L}(r,n)$ of any other potential $\alpha_{r-i}(r)$ is $[(r+n-2i+2)/2]$, $i.e.$, $N_i=[(r+n-2i+2)/2]$ for $i=3,4,\ldots, r$. The square brackets denote the integer part of the expression inside. \end{theo} \noindent {\bf Proof} The proof is a tedious check on all the possibilities of combinations of derivatives of the potentials appearing in the Lagrangian ${\cal L}(r,n)$. We start with a few specific cases before examining the general case. \vspace*{0.5cm} \noindent{\em Dependence of ${\cal L}(r,n)$ on $\alpha_{r-2}(r)$}\vspace{0.5cm} We consider terms of the form $\aa{r-2}{k_1} \aa{r-j}{k_2}$. We want to find the maximal allowable value for $k_1$, $i.e.$, for the highest derivative of $\alpha_{r-2}(r)$ appearing in the Lagrangian. We have \begin{displaymath} W[\aa{r-2}{k_1} \aa{r-j}{k_2}]=k_1+k_2+2+j=r+n+1=W[{\cal L}(r,n)], \end{displaymath} \noindent hence \begin{displaymath} k_1+k_2=r+n-1-j. \end{displaymath} Only values of $k_1, k_2$ with $|k_1-k_2|\leq 1$ need to be considered in this case. Other values are reduced to these cases using integration by parts. If $j=2$, then necessarily $k_1=k_2$. Otherwise the term we are considering is a total derivative, equivalent to $([\aa{r-2}{k_2}]^2/2)'$. In this case we find \begin{displaymath} k_1=\frac{r+n-3}{2}. \end{displaymath} \noindent This value is not necessarily an integer. If it is, when $r+n-3$ is even, it would be the maximum value for $k_1$. Otherwise, this term does not appear in ${\cal L}(r,n)$, so we consider next $j=3$. If $k_1=k_2+1$, we find $k_1=(r+n-5)/2$. If this is an integer, than so is $(r+n-3)/2$, hence this does not raise the order of differentiation with which $\alpha_{r-2}(r)$ appears. On the other hand, if $k_1=k_2$, we find \begin{displaymath} k_1=\frac{r+n-4}{2}. \end{displaymath} \noindent Either $(r+n-3)/2$ or $(r+n-4)/2$ is guaranteed to be an integer. This integer is the maximal order of differentiation with which $\alpha_{r-2}(r)$ appears in the Lagrangian: hence \begin{equation}\la{N2} N_2=\left[\frac{r+n-3}{2}\right], \end{equation} \noindent where the square brackets denote the integer part of the expression inside. If $r+n-3$ is even, this results in the first value we obtained for $k_1$. If $r+n-3$ is odd, we find the second expression. \vspace*{0.5cm} \noindent{\em Dependence of ${\cal L}(r,n)$ on $\alpha_{r-3}(r)$}\vspace{0.5cm} Next consider terms of the form $\aa{r-3}{k_1} \aa{r-j}{k_2}$. We want to find the maximal allowable value for $k_1$, $i.e.$, $N_3$. We have \begin{displaymath} W[\aa{r-3}{k_1} \aa{r-j}{k_2}]=k_1+k_2+3+j=r+n+1=W[{\cal L}(r,n)], \end{displaymath} \noindent or \begin{displaymath} k_1+k_2=r+n-2-j. \end{displaymath} If $j=2$, then for the case $k_1=k_2$, we find \begin{displaymath} k_1=\frac{r+n-4}{2}. \end{displaymath} \noindent In the other case, $k_2=k_1+1$ (we can always write the Lagrangian such that the potentials with the lowest weight have the higher order of differentiation), we obtain \begin{displaymath} k_1=\frac{r+n-5}{2}. \end{displaymath} \noindent In this case, $k_2=(r+n-3)/2$, which corresponds to $N_2$ (if $r+n-3$ is even), found in the previous section. The analysis for $j>2$ does not increase the possible values of $k_1$. Either $(r+n-4)$ or $(r+n-5)$ is guaranteed to be even, so \begin{equation}\la{N3} N_3=\left[\frac{r+n-4}{2}\right]. \end{equation} \vspace*{0.5cm} \noindent{\em Dependence of ${\cal L}(r,n)$ on $\alpha_{r-4}(r)$}\vspace{0.5cm} Consider terms of the form $\aa{r-4}{k_1} \aa{r-j}{k_2}$. We want to find the maximal allowable value for $k_1$, $i.e.$, $N_4$. We have \begin{displaymath} W[\aa{r-4}{k_1} \aa{r-j}{k_2}]=k_1+k_2+4+j=r+n+1=W[{\cal L}(r,n)], \end{displaymath} \noindent or \begin{displaymath} k_1+k_2=r+n-3-j. \end{displaymath} Consider the case when $j=2$. If $k_1=k_2$, then $k_1=(r+n-5)/2$ and $k_2=(r+n-5)/2=(r+n-3)/2-1$. If $r+n-5$ is an integer, then so is $r+n-3$. In this case we can use integration by parts to decrease $k_1$ by 1 and increase $k_2$ by 1. Therefore this possibility is not allowed. If $k_2=k_1+1$, then we obtain \begin{displaymath} k_1=\frac{r+n-6}{2} ~~\mbox{and}~~ k_2=\frac{r+n-4}{2}. \end{displaymath} \noindent This possibility is allowed. Also, if we let $k_2=k_1+1$, we find \begin{displaymath} k_1=\frac{r+n-7}{2} ~~\mbox{and}~~ k_2=\frac{r+n-3}{2}. \end{displaymath} \noindent This possibility is also allowed. Examining the other possibilities for $j, k_2$ does not improve the allowed values for $k_1$, therefore \begin{equation}\la{N4} N_4=\left[\frac{r+n-6}{2}\right]. \end{equation} \noindent{\em Dependence of ${\cal L}(r,n)$ on $\alpha_{r-i}(r)$, $3\leq i \leq r$}\vspace{0.5cm} We now state the general case. Using arguments as above (we want potentials with lower weight to have a higher order of differentiation in each term of the Lagrangian), we have \begin{displaymath} W[\aa{r-i}{k_1} \aa{r-j}{k_2}]=k_1+k_2+i+j=r+n+1=W[{\cal L}(r,n)], \end{displaymath} \noindent or \begin{displaymath} k_1+k_2=r+n+1-i-j. \end{displaymath} Consider the case when $j=2$. Then if $k_2=k_1+m$, \begin{displaymath} k_1=\frac{r+n-1-i-m}{2} ~~\mbox{and}~~ k_2=\frac{r+n-1-i+m}{2}. \end{displaymath} \noindent Using the above argument, we obtain an allowed possibility if either $k_2=(n+r-3)/2$ or $k_2=(n+r-4)/2$. This gives two possible values for $m$: \begin{displaymath} m=i-2 ~~ \mbox{or} ~~ m=i-3. \end{displaymath} \noindent These respectively give \begin{displaymath} k_1=\frac{r+n-2i+1}{2} ~~\mbox{and}~~ k_2=\frac{r+n-2i+2}{2}. \end{displaymath} The other possibilities for $j$ give no additional information, hence \begin{equation}\la{Ni} N_i=\left[\frac{n+r-2i+2}{2}\right]. \end{equation} This formula is valid for all $i$, except $i=2$, the first value. \hspace*{\fill}$\rule{3mm}{3mm}$ Table \ref{table1} gives an overview of the possibilities. \settowidth{\mylength}{$\left[\frac{r+n-3}{2}\right]$} \settoheight{\myheight}{$\left[\frac{r+n-3}{2}\right]$} \addtolength{\myheight}{8pt} \begin{table}[htb] \begin{center} \caption{\bf The order of differentiation with which the potentials appear in the Lagrangian \la{table1}} \vspace*{0.2in} \begin{tabular}{|c|c|} \hline $i$ & $N_i$ \\ \hline\hline \pb{2} & $\left[\frac{r+n-3}{2}\right]$ \\ \hline \pb{3} & $\left[\frac{r+n-4}{2}\right]$ \\ \hline \pb{$\vdots$} & $\vdots$ \\ \hline \pb{$j$} & $\left[\frac{r+n-2 j+2}{2}\right]$ \\ \hline \pb{$\vdots$} & $\vdots$ \\ \hline \pb{$r$} & $\left[\frac{n-r+2}{2}\right]$\\ \hline \end{tabular} \end{center} \end{table} \vspace*{12pt} \noindent{\bf Remark} In \cite{ds1}, it was argued that a generic rank 1, genus $g$ solution of the KP equation corresponds to a solution of the $(r,n)$-th KP equation with $r=g+1$ and $n=g+2$. The dependence of the Lagrangian ${\cal L}(r,n)={\cal L}(g+1,g+2)$ on the potentials for this case is found from Table \ref{table1} by using these values for $r$ and $n$. It follows that in the generic case, the potential $\alpha_{r-j}(r)$ appears with $g-j+2$ derivatives, for $j=2,3,\ldots, g+1$. \section{The Ostrogradskii transformation, canonical variables and the Hamiltonian system}\la{sec:ostro} If we have a Lagrangian system, where the Lagrangian only depends on the potentials and their first derivatives, then under certain conditions we can use the Legendre transformation \cite{arnold} to write the Lagrangian system in first-order form as a Hamiltonian system with canonical variables $(\mbf{q},\mbf{p})$. Here the variables $\mbf{q}$ are the potentials appearing in the Lagrangian and all of their derivatives, except the highest derivative. Their {\em conjugate variables} $\mbf{p}$ are defined as the partial derivatives of the Lagrangian with respect to the $\mbf{q}$ variables \cite{arnold}. The Lagrangian system \rf{el} we constructed from the KP hierarchy, assuming two of its flows are stationary, depends on more than just the first derivatives of the potentials $\mbf{\alpha}(r) = (\alpha_{0}(r), \alpha_1(r), \ldots, $ $\alpha_{r-2}(r))^T$. The Legendre transformation is generalized to write the Lagrangian system \rf{el} in Hamiltonian form. This is achieved by Ostrogradskii's theorem, given later in this section. Consider the Ostrogradskii transformation (see \cite{dfn2, whittaker} for a simpler version) \begin{equation}\la{ostrotrans} q_{ij}=\ppn{j-1}{}{x}\alpha_{r-i-1}(r), ~~p_{ij}=\dd{{\cal L}(r,n)}{\aa{r-i-1}{j}} \end{equation} \noindent for $i=1,2, \ldots, r-1$ and $j=1, 2, \ldots, N_{i+1}$. Note that when all the $N_j=1$, for $j=2,3,\ldots, r$ ($i.e.$, when the Lagrangian only depends on first derivatives), then the Ostrogradskii transformation \rf{ostrotrans} reduces to the Legendre transformation. Using the definition of the variational derivative, we establish the recurrence relations \begin{equation}\la{ostrorecur} p_{ij}=\pp{{\cal L}(r,n)}{\aa{r-i-1}{j}}-\pp{p_{i(j+1)}}{x}, ~~\mbox{for} ~~ j=1,2,\ldots, N_{i+1}. \end{equation} Here we have defined $p_{i(N_{i+1}+1)}$ to be zero. These recurrence relations will be useful later on. Furthermore, from the definition of the Ostrogradskii transformation \rf{ostrotrans}, we obtain \begin{equation}\la{ostroweight} W[q_{ij}]=i+j~~\mbox{and}~~W[p_{ij}]=r+n-(i+j). \end{equation} \noindent Weight relationships such as these provide a quick check for the validity of large expressions, such as the ones encountered below. Many typographical errors are easily avoided by checking that only terms with the same weight are added. The Lagrangian needs to fulfill a nonsingularity requirement for the Ostrogradskii transformation to be invertible: \vspace*{12pt} \noindent {\bf Definition \cite{dfn2}:} The Lagrangian ${\cal L}(r,n)$ is {\bf (strongly) nonsingular} if the Ostrogradskii transformation \rf{ostrotrans} can be solved in the form \begin{displaymath} \aa{r-i}{j}=\aa{r-i}{j}(\mbf{q},\mbf{p}), ~~\mbox{for}~i=2,3,\ldots, r ~~\mbox{and}~j=1,2,\ldots, 2 N_i-1. \end{displaymath} \noindent Here $\mbf{q}$ and $\mbf{p}$ denote the vector of all variables $q_{ij}$ and $p_{ij}$ respectively. In other words, the Ostrogradskii transformation \rf{ostrotrans} is invertible if and only if the Lagrangian is nonsingular. Then the Euler-Lagrange equation \rf{el} can be written in first-order form using the variables $\mbf{q}$ and $\mbf{p}$. Define the vector $\mbf{X}=(\aa{r-2}{N_2}, \aa{r-3}{N_3}, \ldots, \aa{0}{N_r})^T$. This is the vector containing the highest derivatives of the potentials. We already know that the highest derivatives of the potentials are found in the quadratic terms of the Lagrangian. The Lagrangian is conveniently written in the following form: \begin{equation}\la{quadraticform} {\cal L}(r,n)=\frac{1}{2} \mbf{X}^T \mbf{{\cal G}}(r,n) \mbf{X}+\mbf{{\cal A}}^T (r,n) \mbf{X}+ \tilde{\mbf{{\cal L}}}(r,n), \end{equation} \noindent with $\mbf{{\cal G}}(r,n)$, $\mbf{{\cal A}}(r,n)$, $\tilde{{\cal L}}(r,n)$ independent of $\mbf{X}$. $\mbf{{\cal G}}(r,n)$ is a constant symmetric $(r-1) \times (r-1)$ matrix. $\mbf{{\cal A}}(r,n)$ is an $(r-1)$-dimensional vector.In the classical case of the Legendre transformation $\mbf{{\cal G}}(r,n)$ can be regarded as either a metric tensor or as the inverse of a mass tensor \cite{arnold}. The following theorem generalizes a well-known result for the Legendre transformation. \begin{theo}\la{prop:sing} The Lagrangian ${\cal L}(r,n)$ is nonsingular if and only if the matrix $\mbf{{\cal G}}(r,n)$ in \rf{quadraticform} is nonsingular. \end{theo} \noindent{\bf Proof} The proof is an extension of the proof in the case when the Lagrangian depends on only one potential \cite{dfn2}. We demonstrate that under the assumption that $\mbf{{\cal G}}(r,n)$ is nonsingular, the Ostrogradskii transformation \rf{ostrotrans} is invertible. Then by definition the Lagrangian is nonsingular. First note that the variables $\mbf{q}$ are expressed in terms of the potential and its derivatives, by their definition \rf{ostrotrans}. Furthermore, the Lagrangian ${\cal L}(r,n)$ is a function of only $\mbf{q}$ and $\mbf{X}$: it follows from the Ostrogradksii transformation \rf{ostrotrans} that all derivatives of the potentials in the Lagrangian are $\mbf{q}$-variables, except the highest derivative of the potentials. These are the components of $\mbf{X}$. We now construct the vector \begin{eqnarray*} \mbf{P}&=&\left( \begin{array}{cccc} p_{1N_2},& p_{2N_3},& \cdots,& p_{(r-1)N_r} \end{array} \right)^T\\ &=& \left( \dd{{\cal L}(r,n)}{\aa{r-2}{N_2}}, \dd{{\cal L}(r,n)}{\aa{r-3}{N_3}}, \cdots, \dd{{\cal L}(r,n)}{\aa{0}{N_r}} \right)^T\\&=& \dd{{\cal L}(r,n)}{\mbf{X}}. \end{eqnarray*} \noindent Since $\mbf{X}$ contains the highest derivatives of the potentials, by definition of the variational derivative we get \begin{eqnarray}\la{labeliguess} \left( \begin{array}{c} p_{1N_2}\\ p_{2N_3}\\ \vdots\\ p_{(r-1)N_r} \end{array} \right)=\pp{{\cal L}(r,n)}{\mbf{X}}=\mbf{{\cal G}}(r,n) \mbf{X}+ \mbf{{\cal A}}(r,n). \end{eqnarray} \noindent Now we solve \rf{labeliguess} for $\mbf{X}$, since by assumption $\mbf{{\cal G}}(r,n)$ is nonsingular. We denote $\mbf{{\cal G}}^{-1}(r,n)=\mbf{{\cal M}}(r,n)$. Since $\mbf{{\cal G}}(r,n)$ is symmetric, so is $\mbf{{\cal M}}(r,n)$. \begin{eqnarray}\nonumber \mbf{X}&=&\left( \begin{array}{c} \aa{r-2}{N_2}\\ \aa{r-3}{N_3}\\ \vdots\\ \aa{0}{N_r} \end{array} \right)=\mbf{\cal M}(r,n) \left( \begin{array}{c} p_{1N_2}\\ p_{2N_3}\\ \vdots\\ p_{(r-1)N_r} \end{array} \right)-\mbf{\cal M}(r,n)\mbf{\cal A}(r,n)\\\la{xintermsofp} &=&\mbf{\cal M}(r,n)\left(\mbf{P}-\mbf{\cal A}(r,n)\right). \end{eqnarray} We have expressed $\mbf{X}$ in terms of the coordinates $\mbf{q}$ and $\mbf{p}$. We want to do the same with its derivatives. Now consider the following set of Ostrogradskii equations, using the recurrence relations \rf{ostrorecur} \begin{eqnarray*} \left( \begin{array}{c} p_{1(N_2-1)}\\ p_{2(N_3-1)}\\ \vdots\\ p_{(r-1)(N_r-1)} \end{array} \right)&=& \left( \begin{array}{c} \pp{{\cal L}(r,n)}{\aa{r-2}{N_2-1}}-\pp{p_{1N_2}}{x}\\ \pp{{\cal L}(r,n)}{\aa{r-3}{N_3-1}}-\pp{p_{2N_3}}{x}\\ \vdots\\ \pp{{\cal L}(r,n)}{\aa{0}{N_r-1}}-\pp{p_{(r-1)N_r}}{x} \end{array} \right)\\&=& \left( \begin{array}{c} \pp{{\cal L}(r,n)}{\aa{r-2}{N_2-1}}\\ \pp{{\cal L}(r,n)}{\aa{r-3}{N_3-1}}\\ \vdots\\ \pp{{\cal L}(r,n)}{\aa{0}{N_r-1}} \end{array} \right)-\pp{}{x} \left( \begin{array}{c} p_{1N_2}\\ p_{2N_3}\\ \vdots\\ p_{(r-1)N_r} \end{array} \right). \end{eqnarray*} \noindent Note that the first term depends only on $\mbf{q}$ and $\mbf{X}$. The last term is the derivative of $\mbf{\cal G}(r,n) \mbf{X}+\mbf{\cal A}(r,n)$. Since $\mbf{\cal A}(r,n)$ depends only on $\mbf{q}$, the derivative of $\mbf{\cal A}(r,n)$ depends only on $\mbf{q}$ and on $\mbf{X}$. Since $\mbf{\cal G}(r,n)$ is constant, we can solve this relationship for $\mbf{X}'$: \begin{displaymath} \mbf{X}'=\mbf{\cal M}(r,n) \left( \begin{array}{c} \pp{{\cal L}(r,n)}{\aa{r-2}{N_2-1}}\\ \pp{{\cal L}(r,n)}{\aa{r-3}{N_3-1}}\\ \vdots\\ \pp{{\cal L}(r,n)}{\aa{0}{N_r-1}} \end{array} \right)-\mbf{\cal M}(r,n) \left( \begin{array}{c} p_{1(N_2-1)}\\ p_{2(N_3-1)}\\ \vdots\\ p_{(r-1)(N_r-1)} \end{array} \right)-\mbf{\cal M}(r,n)\pp{\mbf{\cal A}(r,n)}{x}, \end{displaymath} \noindent and we have expressed $\mbf{X}'$ in terms of $\mbf{q}, \mbf{p}$ and $\mbf{X}$. Since in the previous steps, $\mbf{X}$ was expressed in terms of $\mbf{q}$ and $\mbf{p}$, this proves that $\mbf{X}'$ is expressible in terms of $\mbf{q}$ and $\mbf{p}$. Continued use of the recursion relation \rf{ostrorecur} allows us to do the same with higher derivatives of $\mbf{X}$, which are needed as the Euler-Lagrange equations \rf{el} depend on twice as many derivatives of the potentials as the Lagrangian. This proves that the Ostrogradskii transformation \rf{ostrotrans} can be inverted. Hence the Lagrangian is nonsingular if $\mbf{\cal G}(r,n)$ is nonsingular. The converse statement is clearly also true: if the matrix $\mbf{\cal G}(r,n)$ is singular, then the Ostrogradksii transformation is not invertible (step 1 in the proof fails, since \rf{labeliguess} cannot be solved for $\mbf{X}$). Hence the Lagrangian is singular if the matrix $\mbf{\cal G}(r,n)$ is singular. \hspace*{\fill}$\rule{3mm}{3mm}$ \vspace*{12pt} In sharp contrast to the Korteweg-deVries hierarchy \cite{nmpz}, the KP hierarchy contains both singular and nonsingular Lagrangians. This is an indication that the set of potentials $\mbf{\alpha}(r)=(\alpha_{0}(r), \alpha_1(r), \ldots, \alpha_{r-2}(r))^T$ is not a good set of variables to describe the dynamics of the system. These points are further explained in Section \ref{sec:sing}. Denote by \begin{equation}\la{count} N=\sum_{i=2}^r N_i=N_2+N_3+\ldots+N_r. \end{equation} We have the following theorem: \begin{theo}\la{ostrotheo}{\bf (Ostrogradskii \cite{dfn2, ostro})} If the Lagrangian ${\cal L}(r,n)$ is nonsingular, then the first-order system obtained by rewriting the Euler-Lagrange equations in terms of the new variables $\mbf{q}$ and $\mbf{p}$ is Hamiltonian with Hamiltonian \begin{equation}\la{ostrohamil} H(\mbf{q},\mbf{p})=\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}} p_{ij} \aa{r-i-1}{j}- {\cal L}(r,n). \end{equation} Here the inverse Ostrogradskii transformation is to be used in order to express the Hamiltonian in terms of $\mbf{q}$ and $\mbf{p}$ only. The Euler-Lagrange equations are equivalent to the $N$-dimensional Hamiltonian system \begin{equation}\la{ostrodynsys} \pp{q_{ij}}{x}=\pp{H}{p_{ij}},~~~\pp{p_{ij}}{x}=-\pp{H}{q_{ij}}, \end{equation} for $i=1,2,\ldots, r-1$ and $j=1,2,\ldots, N_{i+1}$. \end{theo} \noindent{\bf Proof} The proof is identical to the proof in \cite{dfn2}, except that more than one potential appears in the Lagrangian. This slight modification does not change the proof in any fundamental way. \vspace*{12pt} In other words, the variables $q_{ij}$ and $p_{ij}$ are canonically conjugate variables under the classical symplectic structure where the Poisson bracket is given by \begin{equation}\la{cpb} \{f,g\}=\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}}\left(\pp{f}{q_{ij}}\pp{g}{p_{ij}}- \pp{f}{p_{ij}}\pp{g}{q_{ij}}\right)=(\mbf{\nabla} f)^T J (\mbf{\nabla} g), \end{equation} \noindent and \begin{displaymath}\la{J} \mbf{J}=\left( \begin{array}{cc} \mbf{0}_N&\mbf{I}_N\\ -\mbf{I}_N & \mbf{0}_N \end{array} \right), \end{displaymath} \noindent where $\mbf{0}_N$ is the $N \times N$-null matrix, $\mbf{I}_N$ is the $N\times N$-identity matrix, and \begin{eqnarray}\nonumber \mbf{\nabla} f&=&\left( \pp{f}{q_{11}}, \ldots, \pp{f}{q_{1N_2}}, \pp{f}{q_{21}}, \ldots, \pp{f}{q_{2N_3}}, \ldots, \pp{f}{q_{(r-1)N_r}}\right.,\\\la{nabla} &&\left.\pp{f}{p_{11}}, \ldots, \pp{f}{p_{1N_2}}, \pp{f}{p_{21}}, \ldots, \pp{f}{p_{2N_3}}, \ldots, \pp{f}{p_{(r-1)N_r}} \right)^T \end{eqnarray} \noindent is a $2N$-dimensional vector. \begin{theo} The Hamiltonian in \rf{ostrohamil} can be rewritten in the form \begin{eqnarray}\nonumber H(\mbf{q},\mbf{p})&=& \frac{1}{2} \left(\mbf{P}-\mbf{{\cal A}}(r,n)(\mbf{q})\right)^T \mbf{{\cal M}}(r,n) \left(\mbf{P}-\mbf{{\cal A}}(r,n)(\mbf{q})\right) +\\\la{goodform} &&\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}-1} p_{ij} q_{i(j+1)}-\tilde{\cal L}(r,n)(\mbf{q}), \end{eqnarray} \noindent with $\mbf{P}=(p_{1N_2}, p_{2N_3}, \ldots, p_{(r-1)N_r})^T$. \end{theo} \noindent{\bf Proof} (using \rf{ostrotrans}, \rf{quadraticform} and \rf{xintermsofp}) \begin{eqnarray*} H(\mbf{q},\mbf{p})&\hspace*{-4pt}=\hspace*{-4pt}&\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}} p_{ij} \aa{r-i-1}{j}- {\cal L}(r,n)\\ &\hspace*{-4pt}=\hspace*{-4pt}&\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}-1}p_{ij} \aa{r-i-1}{j}+\sum_{i=1}^{r-1} p_{iN_{i+1}}\aa{r-i-1}{N_{i+1}}- \frac{1}{2}\mbf{X}^T \mbf{\cal G}(r,n) \mbf{X}-\mbf{\cal A}^T(r,n)(\mbf{q}) \mbf{X}-\\&&\tilde{\cal L}(r,n)(\mbf{q})\\ &\hspace*{-4pt}=\hspace*{-4pt}&\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}-1}p_{ij} q_{i(j+1)}+ \mbf{P}^T \mbf{X} -\frac{1}{2}\mbf{X}^T \mbf{\cal G}(r,n) \mbf{X}- \mbf{\cal A}^T(r,n)(\mbf{q}) \mbf{X}-\tilde{\cal L}(r,n)(\mbf{q})\\ &\hspace*{-4pt}=\hspace*{-4pt}&\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}-1}p_{ij} q_{i(j+1)}+ \mbf{P}^T \mbf{\cal M}(r,n) \left(\mbf{P}-\mbf{\cal A}(r,n)(\mbf{q})\right)-\\ &&\frac{1}{2}\left(\mbf{P}-\mbf{\cal A}(r,n)(\mbf{q})\right)^T \mbf{\cal M}^T(r,n) \mbf{\cal G}(r,n)\mbf{\cal M}(r,n) \left(\mbf{P}-\mbf{\cal A}(r,n)(\mbf{q}) \right)-\\ &&\mbf{\cal A}^T(r,n)(\mbf{q}) \mbf{\cal M}(r,n) \left(\mbf{P}- \mbf{\cal A}(r,n)(\mbf{q})\right)- \tilde{\cal L}(r,n)(\mbf{q})\\ &\hspace*{-4pt}=\hspace*{-4pt}&\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}-1}p_{ij} q_{i(j+1)}+ \frac{1}{2}\left(\mbf{P}-\mbf{\cal A}(r,n)(\mbf{q})\right)^T \mbf{\cal M}(r,n) \left(\mbf{P}-\mbf{\cal A}(r,n)(\mbf{q})\right)-\tilde{\cal L}(r,n)(\mbf{q})\\ &\hspace*{-4pt}=\hspace*{-4pt}&\frac{1}{2} \left(\mbf{P}-\mbf{\cal A}(r,n)(\mbf{q})\right)^T \mbf{\cal M}(r,n) \left(\mbf{P}-\mbf{\cal A}(r,n)(\mbf{q})\right) +\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}-1} p_{ij} q_{i(j+1)}-\tilde{\cal L}(r,n)(\mbf{q}). \end{eqnarray*}\hspace*{\fill}$\rule{3mm}{3mm}$ \vspace*{12pt} \noindent Note that if the middle term were missing from \rf{goodform}, the Hamiltonian would be of the form: Kinetic Energy plus Potential Energy. Such Hamiltonians are called natural \cite{arnold}. (Note that $\tilde{\cal L}(r,n)$ plays the role of minus the potential energy by its definition \rf{quadraticform}.) However, the term natural is conventionally only used if the mass tensor $\mbf{{\cal M}}(r,n)$ is positive definite. This is not the case here, as the examples will illustrate. The next theorem is well known \cite{dfn2} for a Lagrangian depending on one potential. \begin{theo}\la{prop2} \begin{equation}\la{compare} \pp{H}{x}=-\mbf{\alpha}'^T(r) \dd{{\cal L}(r,n)}{\mbf{\alpha}(r)}. \end{equation} \end{theo} \noindent{\bf Proof} The proof is by direct calculation. \begin{eqnarray*} \pp{H}{x}&=&\pp{}{x}\left( \sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}} p_{ij} \aa{r-i-1}{j}-{\cal L}(r,n) \right)\\ &=&\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}}\left[ \pp{p_{ij}}{x}\aa{r-i-1}{j}+p_{ij}\pp{\aa{r-i-1}{j}}{x} \right]-\pp{{\cal L}(r,n)}{x}\\ &=&\sum_{i=1}^{r-1}\pp{p_{i1}}{x}\pp{\alpha_{r-i-1}(r)}{x}+ \sum_{i=1}^{r-1}\sum_{j=2}^{N_{i+1}}\pp{p_{ij}}{x}\aa{r-i-1}{j}+\\&& \sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}}p_{ij} \aa{r-i-1}{j+1}- \sum_{i=1}^{r-1}\sum_{j=0}^{N_{i+1}}\pp{{\cal L}(r,n)}{\aa{r-i-1}{j}} \aa{r-i-1}{j+1}\\ &=&\sum_{i=1}^{r-1}\pp{p_{i1}}{x}\pp{\alpha_{r-i-1}(r)}{x}+ \sum_{i=1}^{r-1}\sum_{j=2}^{N_{i+1}}\pp{p_{ij}}{x}\aa{r-i-1}{j}+ \sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}}p_{ij} \aa{r-i-1}{j+1}-\\&& \sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}}\pp{{\cal L}(r,n)}{\aa{r-i-1}{j}} \aa{r-i-1}{j+1}-\sum_{i=1}^{r-1}\pp{{\cal L}(r,n)}{\alpha_{r-i-1}(r)} \pp{\alpha_{r-i-1}(r)}{x}\\ &\since{ostrorecur}&\sum_{i=1}^{r-1}\left[ \pp{p_{i1}}{x}-\pp{{\cal L}(r,n)}{\alpha_{r-i-1}(r)} \right] \pp{\alpha_{r-i-1}(r)}{x}+ \sum_{i=1}^{r-1}\sum_{j=2}^{N_{i+1}}\left[ \pp{{\cal L}(r,n)}{\aa{r-i-1}{j-1}}-p_{i(j-1)} \right] \aa{r-i-1}{j}+\\ &&\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}}p_{ij} \aa{r-i-1}{j+1}- \sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}}\pp{{\cal L}(r,n)}{\aa{r-i-1}{j}} \aa{r-i-1}{j+1}\\ &\since{ostrotrans}&-\sum_{i=1}^{r-1} \left[ \pp{{\cal L}(r,n)}{\alpha_{r-i-1}(r)}-\pp{}{x}\dd{{\cal L}(r,n)}{\alpha_{r-i-1}'(r)} \right]\pp{\alpha_{r-i-1}(r)}{x}+\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}-1} \pp{{\cal L}(r,n)}{\aa{r-i-1}{j}} \aa{r-i-1}{j+1} -\\&&\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}-1}p_{ij}\aa{r-i-1}{j+1}+ \sum_{i=1}^{r-1} \sum_{j=1}^{N_{i+1}-1} p_{ij} \aa{r-i-1}{j+1}+\sum_{i=1}^{r-1}p_{iN_{i+1}}\aa{r-i-1}{N_{i+1}+1}-\\ &&\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}-1}\pp{{\cal L}(r,n)}{\aa{r-i-1}{j}} \aa{r-i-1}{j+1}- \sum_{i=1}^{r-1}\pp{{\cal L}(r,n)}{\aa{r-i-1}{N_{i+1}}} \aa{r-i-1}{N_{i+1}+1}\\ &=&-\sum_{i=1}^{r-1}\dd{{\cal L}(r,n)}{\alpha_{r-i-1}(r)} \pp{\alpha_{r-i-1}(r)}{x}+\sum_{i=1}^{r-1}\left[ p_{iN_{i+1}}-\pp{{\cal L}(r,n)}{\aa{r-i-1}{N_{i+1}}} \right]\aa{r-i-1}{N_{i+1}+1}\\ &=&-\mbf{\alpha}'^T(r) \dd{{\cal L}(r,n)}{\mbf{\alpha}(r)}. \end{eqnarray*}\hspace*{\fill}$\rule{3mm}{3mm}$ The simplest consequence of Theorem \ref{prop2} is that the Hamiltonian is conserved along trajectories of the system ($i.e.$, where the Euler-Lagrange equations \rf{el} are satisfied). But this result is a direct consequence of Ostrogradskii's theorem: since the resulting canonical Hamiltonian system is autonomous, the Hamiltonian is conserved. However, Theorem \ref{prop2} will be useful when we construct more conserved quantities in the next section. It shows how the Hamiltonian fits in with the other conserved quantities. \section{Complete integrability of the Hamiltonian system}\la{sec:comp} Denote \begin{equation} S(r,n)=\{H(r,k)| k=1,2, \ldots, \mbox{with} ~k~\mbox{not an integer multiple of}~r~\mbox{or}~n\}. \end{equation} We know that in the quotient space where the Poisson bracket \rf{pb} is defined, using \rf{pbcommute} we have \begin{equation}\la{commute} \left\{H(r,k_1), H(r,k_2)\right\}=\left(\dd{H(r,k_1)}{\mbf{\alpha}(r)}\right)^T \mbf{J}(r) \left(\dd{H(r,k_2)}{\mbf{\alpha}(r)}\right)=0. \end{equation} \noindent In particular, in the quotient space ($i.e.$, up to total derivatives) \begin{equation}\la{commute2} \left\{{\cal L}(r,n),H(r,k)\right\}=0, \end{equation} \noindent for $H(r,k)\in S(r,n)$. In other words, the Poisson bracket of these two quantities is a total derivative, \begin{equation}\la{ostroconsalmost} \left\{{\cal L}(r,n),H(r,k)\right\}=\pp{T_k}{x}. \end{equation} Along the trajectories of the Euler-Lagrange equations, $i.e.$, along the trajectories of the Hamiltonian system \rf{ostrodynsys}, the quantities $T_k$ are conserved. A list of conserved quantities for the system \rf{ostrodynsys} is hence obtained from \begin{equation}\la{ostrocons} T_k(\mbf{q},\mbf{p})=\int \left(\dd{{\cal L}(r,n)}{\mbf{\alpha}(r)}\right)^T \mbf{J}(r) \left(\dd{H(r,k)}{\mbf{\alpha}(r)}\right) dx, \end{equation} \noindent where $k$ is not an integer multiple of $r$ or $n$. The inverse Ostrogradskii transformation has to be used to express the right-hand side of this equation in terms of $\mbf{q}$ and $\mbf{p}$. Note that \rf{ostrocons} is analogous to the expression for the conserved quantities in \cite{bogoyavlenskii}. From the previous section, we know the Hamiltonian \rf{ostrohamil} is a conserved quantity for the system \rf{ostrodynsys}. The theorem below shows that up to a sign, the Hamiltonian is the first member of the newly constructed list of conserved quantities, \rf{ostrocons}. \begin{theo} \begin{equation}\la{cons1} H(\mbf{q},\mbf{p})=-T_1(\mbf{q},\mbf{p}) \end{equation} \end{theo} \noindent{\bf Proof} From Theorem \ref{prop2}, \begin{displaymath} \pp{H}{x}=-\mbf{\alpha}'^T(r) \dd{{\cal L}(r,n)}{\mbf{\alpha}(r)}= -\left(\dd{{\cal L}(r,n)}{\mbf{\alpha}(r)}\right)^T \pp{\mbf{\alpha}(r)}{x}. \end{displaymath} But the first flow of each hierarchy is the $x$-flow, therefore \begin{eqnarray*} \pp{H}{x}&=&-\left(\dd{{\cal L}(r,n)}{\mbf{\alpha}(r)}\right)^T \mbf{J}(r) \pp{H(r,1)}{\mbf{\alpha}(r)}\\ &=&-\pp{T_1}{x}, \end{eqnarray*} \noindent from which we obtain the theorem. \hspace*{\fill}$\rule{3mm}{3mm}$ \vspace*{12pt} The $N$-dimensional Hamiltonian system \rf{ostrodynsys} and a list of conserved quantities \rf{ostrocons} for it have been constructed. Complete integrability in the sense of Liouville \cite{arnold} can be concluded under the following conditions: \begin{itemize} \item In the phase space spanned by the independent coordinates $\mbf{q}$ and $\mbf{p}$ there are $N$ nontrivial functionally independent conserved quantities. \item These $N$ conserved quantities are mutually in involution with respect to the Poisson bracket \rf{cpb}, $i.e.$, their mutual Poisson brackets vanish. \end{itemize} The list of conserved quantities \rf{ostrocons} for $k$ not an integer multiple of $r$ or $n$ is infinite. It is clearly impossible for all of these conserved quantities to be functionally independent: a dynamical system in a $2 N$-dimensional phase space has at most $2N$ independent conserved quantities. In particular, a $2N$-dimensional autonomous Hamiltonian system has at most $N$ independent integrals of the motion that are mutually in involution. Below it is shown that all the conserved quantities $T_k(\mbf{q},\mbf{p})$ are mutually in involution; therefore at most $N$ of them are functionally independent. We wait until Theorem \ref{theo:atthispoint} to show that exactly $N$ different $T_k(\mbf{q},\mbf{p})$ are nontrivial and functionally independent. In the rest of this section, it is proved that all the $T_{k}(\mbf{q},\mbf{p})$ are in involution. This is not done by direct inspection of the mutual Poisson brackets. Instead, we follow the approach of Bogoyavlenskii and Novikov \cite{bogoyavlenskii}: we know from Adler \cite{adler} that all flows of the hierarchy \rf{hamsys} commute. If we denote by $X_{t_k}$ the vector field that evolves the phase space variables in the direction $t_k$, then the fact that all the flows in \rf{hamsys} commute can be equivalently stated as the mutual commutation of the vector fields $X_{t_k}$. Consider the Hamiltonian system with canonical variables $(\mbf{q},\mbf{p})$, Poisson bracket \rf{cpb} and Hamiltonian $H_k(\mbf{q},\mbf{p})=-T_k(\mbf{q},\mbf{p})$. We show below that the vector field of this system, $X_{H_k}$, is the restriction of $X_{t_k}$ to the phase space $(\mbf{q},\mbf{p})$ of the finite-dimensional Hamiltonian system. So, the different $t_k$-flows commute, even when they are restricted to the phase space consisting of the $(\mbf{q},\mbf{p})$ variables. In particular the $t_k$- and the $t_1$-flow ($i.e.$, the $x$-flow) commute. Hence we have a family of mutually commuting Hamiltonian systems, all with the same Poisson bracket \rf{cpb}. In \cite{arnold}, it is proved that then the family of Hamiltonians have mutually vanishing Poisson brackets. As a consequence, the $T_k(\mbf{q},\mbf{p})$ are mutually in involution and the system \rf{ostrodynsys} is completely integrable in the sense of Liouville, which is what we set out to prove. We remark however, that this way of proving complete integrability also provides us with $N$-dimensional Hamiltonian systems for the evolution of the phase space variables, not only in $x$, but in all `time' variables $t_k$. These different Hamiltonian systems have a common list of conserved quantities in involution. Hence each is completely integrable in the sense of Liouville. This will be spelt out in more detail once we finish proving that all the $T_k(\mbf{q}, \mbf{p})$ are in involution. The following lemma is used in the proof of the Bogoyavlenskii-Novikov theorem. \begin{lemma}\la{lemmalemma} \begin{eqnarray}\la{lemma1} \frac{\partial^2 H}{\partial p_{ij} \partial q_{i(j+1)}}&=&1, ~~~~~\mbox{for}~j\leq N_{i+1}-1\\\la{lemma2} \frac{\partial^2 H}{\partial p_{ij} \partial q_{is}}&=&0, ~~~~~\mbox{for}~s\neq j+1, j\leq N_{i+1}-1\\\la{lemma3} \frac{\partial^2 H}{\partial p_{i_1 j_1} \partial q_{i_2 j_2}}&=&0,~~~~~\mbox{for}~i_1\neq i_2 \end{eqnarray} \end{lemma} \noindent{\bf Proof} We use the form \rf{goodform} of the Hamiltonian. We get \begin{eqnarray*} \pp{H}{p_{ij}}=q_{i(j+1)},~~~~~\mbox{for}~j\leq N_{i+1}-1, \end{eqnarray*} from which \rf{lemma1} and \rf{lemma2} easily follow. Also, \rf{lemma3} follows from this if $j_1\leq N_{i_1+1}-1$ and ${j_2 \leq~N_{i_2+1}-1}$. For other values of $j_1, j_2$, \rf{lemma3} follows immediately from \rf{goodform}.\hspace*{\fill}$\rule{3mm}{3mm}$ The following theorem generalizes the fundamental idea of Bogoyavlenskii and Novikov \cite{bogoyavlenskii}: \begin{theo}{\bf (Bogoyavlenskii-Novikov)} On the solution space of the $(r,n)$-th stationary KP equation, the action of the $k$-th time variable $t_k$ can be written as an $N$-dimensional Hamiltonian system with Hamiltonian $H_k=-T_k$ and the same canonical variables as determined by Ostrogradksii's theorem for the $(r,n)$-th stationary KP equation. \end{theo} \newpage \noindent{\bf Proof} The proof consists of four steps: \begin{enumerate} \item Prove that for $i=1, 2, \ldots, r-1$ \begin{equation}\la{step1} \pp{q_{i1}}{t_k}=\pp{H_k}{p_{i1}}. \end{equation} \item This is an induction step. Assuming the validity of step 1, prove that for $i=1,2,\ldots, r-1$ and $j=1,2,\ldots, N_{i+1}-1$ \begin{equation}\la{step2} \pp{q_{i(j+1)}}{t_k}=\pp{H_k}{p_{i(j+1)}}. \end{equation} \item Prove that for $i=1, 2, \ldots, r-1$ \begin{equation}\la{step3} \pp{p_{iN_{i+1}}}{t_k}=-\pp{H_k}{q_{iN_{i+1}}}. \end{equation} \item The last step is a `backwards' induction step: assuming step 3 is valid, show that for $i=1,2,\ldots, r-1$ and $j=N_{i+1}+1, \ldots, 2,1$ \begin{equation}\la{step4} \pp{p_{i(j-1)}}{t_k}=-\pp{H_k}{q_{i(j-1)}}. \end{equation} \end{enumerate} \noindent{\bf Proof of step 1:} During the course of this step, the index $i$ can attain any value from $1,2,\ldots, r-1$. Using the definition of the variational derivative, \begin{eqnarray*} \dd{{\cal L}(r,n)}{\alpha_{r-i-1}(r)}&=&\pp{{\cal L}(r,n)}{\alpha_{r-i-1}(r)}-\frac{\partial}{\partial x} \dd{{\cal L}(r,n)}{\alpha_{r-i-1}'(r)}\\ &=&\pp{{\cal L}(r,n)}{\alpha_{r-i-1}(r)}-\pp{p_{i1}}{x}. \end{eqnarray*} \noindent This shows that the Euler-Lagrange equations are essentially the equations of motion for the variables $p_{i1}, ~i=1,2,\ldots, r-1$. The other equations of motion in Ostrogradskii's theorem are merely a consequence of the way the new variables in the Ostrogradskii transformation are introduced. On the other hand, from the definition of the Hamiltonian \rf{ostrohamil}, \begin{eqnarray*} \pp{H}{q_{i1}}&=&-\pp{{\cal L}(r,n)}{q_{i1}}\\ &=&-\pp{{\cal L}(r,n)}{\alpha_{r-i-1}(r)}. \end{eqnarray*} Combining the two results, \begin{equation}\la{step11} \dd{{\cal L}(r,n)}{\alpha_{r-i-1}(r)}=-\pp{H}{q_{i1}}-\pp{p_{i1}}{x}. \end{equation} We expand $\partial T_k/\partial x$ in two different ways: \begin{eqnarray*} \pp{T_k}{x}&=&\left(\dd{{\cal L}(r,n)}{\mbf{\alpha}(r)}\right)^T \mbf{J}(r) \left(\dd{H(r,k)}{\mbf{\alpha}(r)}\right)\\ &=&\sum_{i=1}^{r-1}\sum_{j=1}^{r-1}\dd{{\cal L}(r,n)}{\alpha_{i-1}(r)} J_{ij}(r)\dd{H(r,k)}{\alpha_{j-1}(r)}\\ &\since{step11}&-\sum_{i=1}^{r-1}\sum_{j=1}^{r-1}\left(\pp{H}{q_{(r-i)1}}+ \pp{p_{(r-i)1}}{x}\right)J_{ij}(r)\dd{H(r,k)}{\alpha_{j-1}(r)}, \end{eqnarray*} \noindent and \begin{eqnarray*} \pp{T_k}{x}&=&\sum_{i=1}^{r-1} \sum_{j=1}^{N_{i+1}}\left(\pp{T_k}{q_{ij}} \pp{q_{ij}}{x}+\pp{T_k}{p_{ij}}\pp{p_{ij}}{x}\right). \end{eqnarray*} As long as we do not impose the Euler-Lagrange equations, the derivatives of the phase space variables are independent variations. Hence their coefficients in both expressions for $\partial T_k/\partial x$ are equal. Expressing this equality for the coefficient of $\partial p_{i1}/ \partial x$ gives \begin{eqnarray*} \pp{T_k}{p_{i1}}&=&-\sum_{j=1}^{r-1}\left(J_{(r-i)j}(r) \dd{H(r,k)}{\alpha_{j-1}(r)}\right)\\ &\since{hamsys}&-\pp{\alpha_{r-i-1}(r)}{t_k}\\ &\since{ostrotrans}&-\pp{q_{i1}}{t_k}, \end{eqnarray*} \noindent or \begin{displaymath} \pp{q_{i1}}{t_k}=\pp{H_k}{p_{i1}}, \end{displaymath} \noindent which we needed to prove.\hspace*{\fill}$\rule{3mm}{3mm}$ \noindent{\bf Proof of step 2:} Assume \begin{displaymath} \pp{q_{i\tilde{j}}}{t_k}=\pp{H_k}{p_{i\tilde{j}}}, \end{displaymath} \noindent for $\tilde{j}=1,2,\ldots, j$. Then \begin{eqnarray*} \pp{q_{i(j+1)}}{t_k}&\since{ostrotrans}&\pp{\aa{r-i-1}{j}}{t_k}\\ &\since{ostrotrans}&\pp{}{x}\pp{q_{ij}}{t_k}\\ &=&\pp{}{x}\pp{H_k}{p_{ij}}, \end{eqnarray*} \noindent since the $x$-flow and the $t_k$-flow commute. In the last step the induction hypothesis is used. For any function of the variable $x$, \begin{eqnarray}\nonumber \pp{f}{x}&=&\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}}\left(\pp{f}{q_{ij}} \pp{q_{ij}}{x}+\pp{f}{p_{ij}}\pp{p_{ij}}{x}\right)\\\nonumber &=&\sum_{i=1}^{r-1}\sum_{j=1}^{N_{i+1}}\left(\pp{f}{q_{ij}} \pp{H}{p_{ij}}-\pp{f}{p_{ij}}\pp{H}{q_{ij}}\right)\\\la{step21} &=&\{f,H\}, \end{eqnarray} \noindent by definition of the Poisson bracket. With this well-known result, the above becomes \begin{eqnarray*} \pp{q_{i(j+1)}}{t_k}&=&\left\{\pp{H_k}{p_{ij}}, H\right\}\\ &\since{cpb}&\left\{\left\{q_{ij},H_k\right\},H\right\}\\ &=&-\left\{\left\{H_k,H\right\},q_{ij}\right\}- \left\{\left\{H,q_{ij}\right\},H_k\right\}\\ &\since{step21}&-\left\{\pp{H_k}{x},q_{ij}\right\}+ \left\{\left\{q_{ij},H\right\},H_k\right\}\\ &=&\left\{\left\{q_{ij},H\right\},H_k\right\}\\ &\since{step21}&\left\{\pp{q_{ij}}{x},H_k\right\}\\ &\since{ostrotrans}&\left\{q_{i(j+1)},H_k\right\}\\ &\since{cpb}&\pp{H_k}{p_{i(j+1)}}, \end{eqnarray*} \noindent where we have used the fact that the Poisson bracket \rf{cpb} satisfies the Jacobi identity. This is what we needed to prove.\hspace*{\fill}$\rule{3mm}{3mm}$ \noindent{\bf Proof of step 3:} From the commutativity of the $x$-flow and the $t_k$-flow, \begin{eqnarray}\nonumber \pp{}{x}\pp{q_{iN_{i+1}}}{t_k}&=&\pp{}{t_k}\pp{q_{iN_{i+1}}}{x}\\\la{step31} \Rightarrow~~~~~~~~\pp{}{x}\pp{H_k}{p_{iN_{i+1}}}&=&\pp{}{t_k} \pp{H}{p_{iN_{i+1}}}, \end{eqnarray} \noindent using steps 1 and 2 of the proof. We examine the left-hand side of this equation separately. \begin{eqnarray*} \mbox{Left-hand side}&=&\pp{}{x} \pp{H_k}{p_{iN_{i+1}}}\\ &\since{step21}&\left\{\pp{H_k}{p_{}iN_{i+1}},H\right\}\\ &\since{cpb}&-\left\{\left\{H_k,q_{iN_{i+1}}\right\},H\right\}\\ &=&\left\{\left\{q_{iN_{i+1}},H\right\},H_k\right\}+ \left\{\left\{H, H_k\right\},q_{iN_{i+1}}\right\}\\ &\since{step21}&\left\{\pp{q_{iN_{i+1}}}{x},H_k\right\}-\left\{ \pp{H_k}{x},q_{iN_{i+1}}\right\}\\ &=&\left\{\pp{H}{p_{iN_{i+1}}},H_k\right\}, \end{eqnarray*} \noindent again using the Jacobi identity. The factor $\partial H/\partial p_{iN_{i+1}}$ now appears both on the left and on the right of our equation. From \rf{goodform} \begin{eqnarray*} \hspace*{-12pt}&&\pp{H}{p_{iN_{i+1}}}\\ \hspace*{-12pt}&&=\pp{}{p_{iN_{i+1}}}\left(\frac{1}{2} \left(\mbf{P}-\mbf{\cal A}(r,n)(\mbf{q})\right)^T \mbf{\cal M}(r,n) \left(\mbf{P}-\mbf{\cal A}(r,n)(\mbf{q})\right) \right)\\ \hspace*{-12pt}&&=\pp{}{p_{iN_{i+1}}}\left(\frac{1}{2}\sum_{i=1}^{r-1} \sum_{j=1}^{r-1} {\cal M}_{ij}(r,n) \left(p_{iN_{i+1}}-{\cal A}_i(r,n)(\mbf{q})\right) \left(p_{jN_{j+1}}-{\cal A}_j(r,n)(\mbf{q})\right)\right)\\ \hspace*{-12pt}&&=\sum_{j=1}^{r-1}{\cal M}_{ij}(r,n) \left(p_{jN_{j+1}}-{\cal A}_j(r,n)(\mbf{q}) \right). \end{eqnarray*} Using this result in \rf{step31}, \begin{eqnarray*} &&\pp{}{t_k}\sum_{j=1}^{r-1} {\cal M}_{ij}(r,n)\left( p_{jN_{j+1}}-{\cal A}_j(r,n)(\mbf{q}) \right)\\ &&=\left\{\sum_{j=1}^{r-1} {\cal M}_{ij}(r,n) \left(p_{jN_{j+1}}-{\cal A}_j(r,n)(\mbf{q})\right), H_{k}\right\}\\ &&=\sum_{j=1}^{r-1} {\cal M}_{ij}(r,n)\left\{\left(p_{jN_{j+1}} -{\cal A}_j(r,n)(\mbf{q})\right), H_{k}\right\}\\ &&=-\sum_{j=1}^{r-1} {\cal M}_{ij}(r,n)\left(\pp{H_k}{q_{jN_{j+1}}}+ \left\{{\cal A}_j(r,n)(\mbf{q}), H_k\right\} \right). \end{eqnarray*} \noindent Multiplying both sides of this equation by ${\cal G}_{si}(r,n)$ and summing over $i$ from $1$ to $r-1$, this becomes \begin{eqnarray*} \pp{p_{sN_{s+1}}}{t_k}-\pp{{\cal A}_s(r,n)(\mbf{q}),}{t_k}= -\pp{H_k}{q_{sN_{s+1}}}-\left\{{\cal A}_s(r,n)(\mbf{q}), H_k\right\}. \end{eqnarray*} Since ${\cal A}_s(r,n)(\mbf{q})$ depends only on $\mbf{q}$, it follows from step 2 that the second term on the left-hand side is equal to the second term on the right-hand side. Hence \begin{displaymath} \pp{p_{sN_{s+1}}}{t_k}=-\pp{H_k}{q_{sN_{s+1}}}, \end{displaymath} \noindent for $s=1,2,\ldots, r-1$. This terminates the proof of step 3. \hspace*{\fill}$\rule{3mm}{3mm}$ Note that it is necessary for the Lagrangian ${\cal L}(r,n)$ to be nonsingular, as we are using the matrix ${\cal M}(r,n)$. \noindent{\bf Proof of step 4:} Assume \begin{displaymath} \pp{p_{i\tilde{j}}}{t_k}=-\pp{H_k}{q_{i\tilde{j}}}, \end{displaymath} \noindent for $\tilde{j}=N_{i+1}, N_{i+1}-1, \ldots, j$. We have \begin{eqnarray*} \pp{}{t_k}\pp{p_{ij}}{x}&=&\pp{}{x}\pp{p_{ij}}{t_k}\\ &\since{step21}&\left\{\pp{p_{ij}}{t_k},H\right\}\\ &=&-\left\{\pp{H_k}{q_{ij}},H\right\}\\ &\since{cpb}&\left\{\left\{p_{ij},H_k\right\},H\right\}\\ &=&-\left\{\left\{H_k, H\right\},p_{ij}\right\}- \left\{\left\{H, p_{ij}\right\},H_k\right\}\\ &\since{step21}&-\left\{\pp{H_k}{x}, H\right\}+\left\{\pp{p_{ij}}{x}, H_k\right\}\\ &=&-\left\{\pp{H}{q_{ij}}, H_k\right\}\\ &\since{cpb}& -\sum_{\gamma=1}^{r-1}\sum_{\delta=1}^{N_{\gamma+1}}\left( \frac{\partial^2 H}{\partial q_{ij}\partial q_{\gamma \delta}}\pp{H_k}{p_{\gamma \delta}}-\frac{\partial^2 H}{\partial q_{ij} \partial p_{\gamma \delta}}\pp{H_k}{q_{\gamma\delta}}\right)\\ &=&-\sum_{\gamma=1}^{r-1}\sum_{\delta=1}^{N_{\gamma+1}} \frac{\partial^2 H}{\partial q_{ij}\partial q_{\gamma \delta}}\pp{H_k}{p_{\gamma \delta}}+ \sum_{\gamma=1}^{r-1}\sum_{\delta=1}^{N_{\gamma+1}}\frac{\partial^2 H}{\partial q_{ij} \partial p_{\gamma \delta}}\pp{H_k}{q_{\gamma\delta}}\\ &\since{lemma3}&-\sum_{\gamma=1}^{r-1}\sum_{\delta=1}^{N_{\gamma+1}} \frac{\partial^2 H}{\partial q_{ij}\partial q_{\gamma \delta}}\pp{H_k}{p_{\gamma \delta}}+\sum_{\delta=1}^{N_{i+1}} \frac{\partial^2 H}{\partial q_{ij} \partial p_{i \delta}} \pp{H_k}{q_{i\delta}}\\ &\since{lemma1}&-\sum_{\gamma=1}^{r-1}\sum_{\delta=1}^{N_{\gamma+1}} \frac{\partial^2 H}{\partial q_{ij}\partial q_{\gamma \delta}}\pp{H_k}{p_{\gamma \delta}}+\frac{\partial^2 H}{\partial q_{ij} \partial p_{i(j-1)}} \pp{H_k}{q_{i(j-1)}}+\\&& \frac{\partial^2 H}{\partial q_{ij} \partial p_{iN_{i+1}}} \pp{H_k}{q_{iN_{i+1}}}\\ &\since{lemma2}&-\sum_{\gamma=1}^{r-1}\sum_{\delta=1}^{N_{\gamma+1}} \frac{\partial^2 H}{\partial q_{ij}\partial q_{\gamma \delta}}\pp{H_k}{p_{\gamma \delta}}+\pp{H_k}{q_{i(j-1)}}. \end{eqnarray*} The left-hand side of this equation can be expressed another way as well: \begin{eqnarray*} \pp{}{t_k}\pp{p_{ij}}{x}&=&\pp{}{t_k}\left(-\pp{H}{q_{ij}}\right)\\ &=&-\sum_{\gamma=1}^{r-1}\sum_{\delta=1}^{N_{\gamma+1}}\left(\frac{\partial^2 H}{\partial q_{ij} \partial q_{\gamma \delta}}\pp{q_{\gamma \delta}}{t_k}+\frac{\partial^2 H}{\partial q_{ij} \partial p_{\gamma \delta}} \pp{p_{\gamma \delta}}{t_k}\right)\\ &=&-\sum_{\gamma=1}^{r-1}\sum_{\delta=1}^{N_{\gamma+1}}\left(\frac{\partial^2 H}{\partial q_{ij} \partial q_{\gamma \delta}}\pp{H_k}{p_{\gamma \delta}} +\frac{\partial^2 H}{\partial q_{ij} \partial p_{\gamma \delta}} \pp{p_{\gamma \delta}}{t_k}\right)\\ &=&-\sum_{\gamma=1}^{r-1}\sum_{\delta=1}^{N_{\gamma+1}}\frac{\partial^2 H}{\partial q_{ij} \partial q_{\gamma \delta}}\pp{H_k}{p_{\gamma \delta}}-\pp{p_{i(j-1)}}{t_k}, \end{eqnarray*} \noindent where the second term has been simplified using lemma \ref{lemmalemma}, as before. Comparing the two right-hand sides, the double-summed term drops out and one finds \begin{displaymath} \pp{p_{i(j-1)}}{t_k}=-\pp{H_k}{q_{i(j-1)}}. \end{displaymath} This completes the proof of step 4 and hence of the Bogoyavlenskii-Novikov theorem. \hspace*{\fill}$\rule{3mm}{3mm}$ \vspace*{12pt} Let us recapitulate the most important results of the last few sections. \begin{itemize} \item The $(r,n)$-th stationary KP equation can be written as an $N$-dimensional Hamiltonian system in $x$, given by Ostrogradskii's theorem \rf{ostrodynsys}. \item This Hamiltonian system is completely integrable in the sense of Liouville. $N$ independent conserved quantities in involution $T_k(\mbf{q},\mbf{p})$ can be constructed explicitly. \item These conserved quantities can be interpreted as Hamiltonians: The $t_k$-flow induces on the solution space of the $(r,n)$-th KP equation an evolution which is Hamiltonian with Hamiltonian $H_k=-T_k$. This Hamiltonian system shares its phase space variables, symplectic structure and conserved quantities with the $x$-system. The $t_k$-evolution of the phase space variables is given by \begin{equation}\la{tkdynsys} \pp{q_{ij}}{t_k}=\pp{H_k}{p_{ij}},~~~~\pp{p_{ij}}{t_k}=-\pp{H_k}{q_{ij}}, \end{equation} \noindent for $i=1,2,\ldots, r-1$ and $j=1,2,\ldots, N_{i+1}$. The Hamiltonian is given by \begin{equation}\la{tkham} H_k(\mbf{q},\mbf{p})=-T_{k}(\mbf{q},\mbf{p}). \end{equation} This only gives non-trivial results only if $k$ is not a multiple of $r$ and $n$. Strictly speaking however, this is not required for the proof of the Bogoyavlenskii-Novikov theorem. \end{itemize} At this point, we have established enough results to argue that $N$ of the conserved quantities $T_k(r,n)$ are nontrivial and functionally independent: \begin{theo}\la{theo:atthispoint} In the phase space spanned by the independent coordinates $\mbf{q}$ and $\mbf{p}$ there are $N$ nontrivial functionally independent conserved quantities \end{theo} \noindent{\bf Proof} In the previous theorem, we have shown that the conserved quantity $T_k(\mbf{q},\mbf{p})$ is minus the Hamiltonian for the evolution of the phase space variables $(\mbf{q},\mbf{p})$ in the $t_k$-direction. So, if $T_k(\mbf{q},\mbf{p})$ is trivial ($i.e.$, $T_k(\mbf{q},\mbf{p})$=0 on the entire phase space), then so is the dependence of the solution of the $(r,n)$-th KP equation, and conversely. A typical solution of the $(r,n)$-th KP equation is a solution of genus $N$ of the KP equation (see \cite{ds1}) of the form \begin{equation}\la{compkpsol} u=u_0+2\partial_x^2 \ln \Theta(\sum_{j=1}^\infty \mbf{K}_j t_j). \end{equation} \noindent Here all the $\mbf{K}_j$ are $N$-dimensional vectors. If the conserved quantity $T_k(\mbf{q}, \mbf{p})$ is functionally dependent on any of the other $T_j(\mbf{q}, \mbf{p})$, $j<k$, then the vectorfield $X_{H_k}$ is a linear combination of the vectorfields $X_{H_j}, j<k$. Hence the vector $\mbf{K}_k$ is a linear combination of the vectors $\mbf{K}_j, j<k$. If $\mbf{K}_k$, is linearly dependent on the vectors with lower indices, we can use a linear transformation to obtain a solution of the form \rf{compkpsol} which depends on $t_1, t_2, \ldots, t_{k-1}$, but is independent of $t_k$ (for instance, this is possible for $t_r$ and $t_n$). If $\mbf{K}_k$ is independent of the vectors $\mbf{K}_j$, with $j<l$, then the solution depends on $t_k$ in a nontrivial manner. In this case, the conserved quantity $T_k(\mbf{q},\mbf{p})$ has to be nontrivial and functionally independent of $T_j$, for $j<k$. A linear space of dimension $N$ is spanned by $N$ linearly independent vectors. Hence a typical solution of the $(r,n)$-th KP equation has $N$ nontrivial functional independent conserved quantities $T_k(\mbf{q},\mbf{p})$. \hspace*{\fill}$\rule{3mm}{3mm}$ \vspace*{12pt} \noindent{\bf Remark} One often convenient way to integrate an integrable Hamiltonian system explicitly is analogous to the method of forward and inverse scattering, but restricted to systems of ordinary differential equations \cite{wojo}. To invoke this method, a Lax representation \cite{wojo} for the system of Hamiltonian equations is required. Such a representation was obtained in step 5 of the algorithm presented in \cite{ds1}. \vspace*{12pt} In what follows, only the Hamiltonian system in $x$ is considered. Any conclusions reached are however also valid for the Hamiltonian systems in any of the `time' variables $t_k$. \section{Examples} In this section, the abstract formalism of the previous sections is illustrated using concrete examples, by assigning concrete values to $r$ and $n$. The simplest cases are discussed: $r=2$ and various values for $n$ (the KdV hierarchy); $r=3$ and various values for $n$ (the Boussinesq hierarchy). A special case of $r=3$ gives rise to stationary three-phase solutions of the KP equation, namely for $n=4$. This case is illustrated in more detail than the other examples. It is the easiest example not covered by the work of Bogoyavlenskii and Novikov \cite{bogoyavlenskii}. \vspace*{12pt} \noindent {\bf (a) The KdV hierarchy: one-dimensional solutions of the KP equation} \vspace*{12pt} The KdV hierarchy is obtained from the KP hierarchy by imposing the reduction $r=2$, hence \begin{equation}\la{kdvrred} L^2_+=L^2. \end{equation} \noindent Since $L^2_+=\partial^2+2 u_2$, $u_2$ is the only independent potential with the other potentials determined in terms of it. From $L^2_-=0$ \begin{equation}\la{triangkdv} u_3=-\frac{u_2'}{2},~u_4=\frac{u_2''}{4}-\frac{u_2^2}{2}, ~u_5=\frac{3}{2} u_2 u_2' -\frac{u_2'''}{8}, ~etc. \end{equation} \noindent Other ingredients needed for \rf{explicitlaxrred} are: $M(2)=M_{2,1}=\partial$ and $\beta(2,n)=\beta_{-1}(n)=\alpha_{-1}(n)$. Hence the KdV hierarchy has the form \begin{equation}\la{kdvhiernonham} \pp{u_2}{t_n}=\pp{}{x} \alpha_{-1}(n). \end{equation} \noindent In order to write this in Hamiltonian form, we use the potential $\alpha_0(2)=2 u_2=u$. Then $D(2)=2$, by \rf{jacobian}. Hence the Poisson structure of the KdV hierarchy is given by $J(2)=D(2)M(2)=2\partial$. Using \rf{manin2} with $j=1$ and $r=2$, \begin{equation}\la{kdvhiermanin} \alpha_{-1}(2)=\beta_{-1}(2)=\frac{2}{2+n} \dd{\alpha_{-1}(2+n)}{u} \end{equation} we can recast the KdV hierarchy in its familiar Hamiltonian form \cite{gardner}: \begin{equation}\la{kdvhierhamsys} \pp{u}{t_n}=2 \pp{}{x}\dd{H(2,n)}{u}, \end{equation} \noindent with $H(2,n)=2 \alpha_{-1}(2+n)/(2+n)$. If the factor 2 is absorbed into the definition of the Hamiltonian, then this form of the KdV hierarchy is identical to that introduced by Gardner \cite{gardner}. Note that immediately all even flows ({\em i.e.,} $t_{n}=t_{2k}$, for $k$ a positive integer) are trivial because $2+2k$ is not coprime with $r=2$, so $H(2,n)=\alpha_{-1}(2+2k)/(1+k)\equiv 0$. We write out some nontrivial flows explicitly: \begin{description} \item[(i)~~] $n=1$: $H(2,1)=u^2/4$ and \begin{equation} \pp{u}{t_1}=u_x, \end{equation} as expected. \item[(ii)~] $n=3$: $H(2,3)=u^3/8-u_x^2/16$ and \begin{equation} \pp{u}{t_3}=\frac{1}{4}\left(6 u u_x+u_{xxx}\right), \end{equation} the KdV equation. \item[(iii)] $n=5$: $H(2,5)=5 u^4/64-5 u u_x^2/32+u_{xx}^2/64$ and \begin{equation}\la{kdv5} \pp{u}{t_5}=\frac{1}{16}\left(30 u^2 u_x+20 u_x u_{xx}+10 u u_{xxx}+u_{5x} \right), \end{equation} the 5th-order KdV equation. \end{description} There is only one Casimir functional for the KdV hierarchy, namely $H(2,-1)=2 \alpha_{-1}(1)=2 u_2=u$. It is easily verified that this is indeed a Casimir functional for the Poisson structure $J(2)=2 \partial$: $J(2)(\delta H(2,-1)/\delta u)=2 \partial (\delta u/\delta u)=2 \partial (1)=0$. Imposing the $n$-reduction, the Lagrangian ${\cal L}(2,n)$ has the form \begin{equation}\la{kdvlag} {\cal L}(2,n)=H(2,n)+\sum_{k=1}^{n-2}d_{k} H(2,k)+h_1 u. \end{equation} This Lagrangian was first obtained by Bogoyavlenskii and Novikov \cite{bogoyavlenskii}. From Table 1, $N_2=[(n-1)/2]=(n-1)/2$, since $n$ is odd. Necessarily, the Lagrangian has the form \begin{equation}\la{kdvhierlagform} {\cal L}(2,n)=\frac{1}{2} a \left(u^{((n-1)/2)}\right)^2+\hat{\cal L}(2,n), \end{equation} for some nonzero constant $a$ and $\hat{\cal L}(2,n)$ independent of $u^{((n-1)/2)}$. The Lagrangian is always nonsingular. Because the case $r=2$ was covered by Bogoyavlenskii and Novikov, no examples of the Ostrogradksii transformation and the resulting Hamiltonian system of ordinary differential equations will be given. A typical solution of the $(2,n)$-th KP equation, {\em i.e.,} a stationary solution of the $n$-KdV equation, depends on $N=N_2=(n-1)/2$ phases. These solutions are also one-dimensional since they are independent of $y=t_2$. \vspace*{12pt} \noindent {\bf (b) The Boussinesq hierarchy: stationary solutions of the KP equation} \vspace*{12pt} The Boussinesq hierarchy is obtained from the KP hierarchy by imposing the reduction $r=3$, hence \begin{equation}\la{bousrred} L^3_+=L^3. \end{equation} \noindent Since $L^3_+=\partial^3+3 u_2\partial+3 u_2'+3 u_3$, $u_2$ and $u_3$ are the only independent potentials with the other potentials determined in terms of these two. From $L^3_-=0$ \begin{equation}\la{triangbous} u_4=-u_3'-u_2^2-\frac{u_2''}{3}, ~u_5=-2 u_2 u_3+2 u_2 u_2'+\frac{2}{3} u_3''+\frac{u_2'''}{3}, ~etc. \end{equation} \noindent Furthermore, \begin{equation} \mbf{U}(3)=\left(\begin{array}{c}u_2\\u_3\end{array}\right), ~~ \mbf{\beta}(3,n)=\left(\begin{array}{c}\beta_{-1}(n)\\\beta_{-2}(n)\end{array}\right), ~~ \mbf{M}(3)=\left(\begin{array}{cc}\partial & 0\\ -\partial^2 & \partial\end{array}\right), \end{equation} \noindent so that the Boussinesq hierarchy is \begin{equation}\la{bousshiernonham} \pp{\mbf{U}(3)}{t_n}=\mbf{M}(3)\mbf{\beta}(3,n) ~\Rightarrow~\left\{ \begin{array}{rcl} \displaystyle{\pp{u_2}{t_n}}&\displaystyle{=}&\displaystyle{\pp{\beta_{-1}(n)}{x}}\\ \displaystyle{\pp{u_3}{t_n}}&\displaystyle{=}&\displaystyle{-\ppn{2}{\beta_{-1}(n)}{x}+\pp{\beta_{-2}(n)}{x}} \end{array} \right. . \end{equation} \noindent In order to write this in Hamiltonian form, we use the potentials $\alpha_1(3)=3 u_2=u$ and $\alpha_{0}(3)=3 u_2'+3 u_3=v$, where $u$ and $v$ are introduced for notational simplicity. Then \begin{equation} \mbf{\alpha}(3)=\left( \begin{array}{c} v\\u \end{array} \right), ~ \mbf{D}(3)=\left( \begin{array}{cc} 3\partial & 3\\3&0 \end{array} \right),~ \mbf{\beta}(3,n)=\frac{3}{3+n}\dd{}{\mbf{\alpha}(3)}\alpha_{-1}(3+n). \end{equation} \noindent Hence the Poisson structure of the Boussinesq hierarchy is \begin{equation}\la{poissonbous} J(3)=D(3)M(3)=3 \left(\begin{array}{cc}0 & \partial\\ \partial & 0\end{array}\right). \end{equation} \noindent The Boussinesq hierarchy is written in Hamiltonian form as: \begin{equation}\la{boushierhamsys} \pp{}{t_n}\left(\begin{array}{c}v\\u\end{array}\right)=3 \left(\begin{array}{cc}0 & \partial\\ \partial & 0\end{array}\right) \left(\begin{array}{c} \displaystyle{\dd{H(3,n)}{v}}\\ \displaystyle{\dd{H(3,n)}{u}} \end{array}\right)=3 \pp{}{x} \left( \begin{array}{c} \displaystyle{\dd{H(3,n)}{u}}\\ \displaystyle{\dd{H(3,n)}{v}} \end{array} \right) , \end{equation} \noindent with $H(3,n)=3 \alpha_{-1}(3+n)/(3+n)$. Up to a factor 3, this form of the Boussinesq hierarchy is identical to the one introduced by McKean \cite{mckean1}. We write some flows explicitly: \begin{description} \item[(i)~~] $n=1$: $H(3,1)=uv/3$ and \begin{equation} \left\{ \begin{array}{rcl} \displaystyle{v_{t_1}}&=&\displaystyle{v_x}\\ \displaystyle{u_{t_1}}&=&\displaystyle{u_x} \end{array} \right., \end{equation} as expected. \item[(ii)~] $n=2$: $H(3,2)=-u^3/27+u_x^2/9+v^2/3-v u_x/3$ and \begin{equation} \left\{ \begin{array}{rcl} \displaystyle{v_{t_2}}&=&\displaystyle{-\frac{2}{3}u u_x-\frac{2}{3}u_{xxx}+v_{xx}}\\ \displaystyle{u_{t_2}}&=&\displaystyle{2 v_x-u_{xx}} \end{array} \right.. \end{equation} Elimination of $v$ from these two equations gives the Boussinesq equation, \begin{equation}\la{bouss} u_{t_2t_2}+\frac{1}{3}u_{xxxx}+\frac{2}{3}\left(u^2\right)_{xx}=0. \end{equation} \item[(iii)] $n=4$: $H(3,4)=-u_{xx}^2/27+v_x u_{xx}/9-v_{x}^2/9+u u_x^2/9-2 u v u_x/9-u^4/81+2 u v^2/9$ and \begin{equation}\la{bouss4} \left\{ \begin{array}{rcl} \displaystyle{v_{t_4}}&=&\displaystyle{-\frac{4}{9}u^2 u_x+\frac{4}{3}v v_x-\frac{4}{3}u_x u_{xx}-\frac{2}{3}u u_{xxx}+\frac{2}{3}u_x v_x+\frac{2}{3}u v_{xx}-\frac{2}{9} u_{5x}+\frac{2}{3} v_{xxxx}}\\ &&\\ \displaystyle{u_{t_4}}&=&\displaystyle{-\frac{2}{3}u_x^2-\frac{2}{3}u u_{xx}+\frac{4}{3}v u_x-\frac{2}{3}u_{xxxx}+\frac{2}{3}v_{xxx}} \end{array} \right.. \end{equation} This is the next member of the Boussinesq hierarchy. \end{description} There are two Casimir functionals for the Boussinesq hierarchy, namely $H(3,-1)=3 \alpha_{-1}(2)/2=3 (u_{2}'+u_3)/2=v/2$ and $H(3,-2)=3 \alpha_{-1}(1)=3 u_2=u$. For convenience $u$ and $v$ are used as Casimir functionals below. Imposing the $n$-reduction, the Lagrangian ${\cal L}(3,n)$ has the form \begin{equation}\la{bouslag} {\cal L}(3,n)=H(3,n)+\sum_{k=1}^{n-2}d_{k} H(3,k)+h_1 u+h_2 v . \end{equation} Theorem \ref{theo:sing}, given in the next section, shows that this Lagrangian is always nonsingular. A typical solution of the $(3,n)$-th KP equation, {\em i.e.,} a stationary solution of the $n$-Boussinesq equation, depends on $N=N_2+N_3=n-1$ phases. These solutions are stationary solutions of the KP equation, since they are independent of $t=t_3$. \vspace*{12pt} \noindent {\bf (c) Stationary 3-phase solutions of the KP equation} \vspace*{12pt} Consider the $r=3$, $n=4$ reduction. The Lagrangian is \begin{eqnarray}\nonumber {\cal L}(3,4)&=&-\frac{u_{xx}^2}{27}+\frac{v_x u_{xx}}{9}-\frac{v_{x}^2}{9}+\frac{u u_x^2}{9}-\frac{2 u vu_x}{9}- \frac{u^4}{81}+\frac{2 u v^2}{9}+\\\la{lag34} &&d_2\left(-\frac{u^3}{27}+\frac{u_x^2}{9}+\frac{v^2}{3}-\frac{vu_x}{3}\right)+ d_1 \frac{uv}{3}+h_1 u+h_2 v. \end{eqnarray} The Ostrogradskii transformation \rf{ostrotrans} for this Lagrangian is \begin{eqnarray}\nonumber q_{11}=u,&&p_{11}=\dd{{\cal L}(3,4)}{u_x}=\frac{u_{xxx}}{54}+\frac{uu_x}{9}+\frac{d_2 u_x}{18}+\frac{d_1u}{6}+\frac{h_2}{2},\\\la{ostrotrans34} q_{12}=u_x,&&p_{12}=\dd{{\cal L}}{u_{xx}}=-\frac{2 u_{xx}}{27}+\frac{v_x}{9},\\\nonumber q_{21}=v,&&p_{21}=\dd{{\cal L}(3,4)}{v_x}=\frac{u_{xx}}{9}-\frac{2v_x}{9}. \end{eqnarray} \noindent In the definition of $p_{11}$, the Euler-Lagrange equations have been used to eliminate $v_{xx}$. The Ostrogradskii transformation can be inverted: \begin{eqnarray}\nonumber &u=q_{11},~~ u_x=q_{12},~~ v=q_{21},&\\\la{invostro34} &u_{xx}=-54 p_{12}-27 p_{21},~~v_x=-27 p_{12}-18p_{21},&\\\nonumber & u_{xxx}=54p_{11}-6 q_{11}q_{12}-3 d_2 q_{12}-9d_1q_{11}-27 h_2.& \end{eqnarray} \noindent Using the inverse Ostrogradskii transformation, the Hamiltonian corresponding to the Lagrangian \rf{lag34} is \begin{eqnarray}\nonumber H&=&-27p_{12}^2-27 p_{12} p_{21}-9 p_{21}^2+p_{11} q_{12}+\frac{q_{11}^4}{81}-\frac{q_{11} q_{12}^2}{9}-\frac{2 q_{21}^2 q_{11}}{9}+\frac{2 q_{11} q_{12} q_{21}}{9}+\\\la{ham34} &&d_2\left(\frac{q_{11}^3}{27}-\frac{q_{12}^2}{9}-\frac{q_{21}^2}{3}+ \frac{q_{12}q_{21}}{3}\right)-d_1 \frac{q_{11} q_{21}}{3}-h_1 q_{11}-h_2 q_{21}. \end{eqnarray} \noindent For simplicity the constants $d_1, d_2, h_1$ and $h_2$ are equated to zero in the remainder of this example. Using \rf{ostrocons}, three conserved quantities are computed for the Hamiltonian system generated by \rf{ham34}: \begin{eqnarray}\la{t134} T_1&=&-H\\\nonumber T_2&=&3 \int \left(\dd{{\cal L}(3,4)}{v}\pp{}{x}\dd{H(3,2)}{u}+ \dd{{\cal L}(3,4)}{u}\pp{}{x}\dd{H(3,2)}{v}\right) dx\\\nonumber &=&\frac{4 q_{12}q_{11}^3}{81}-\frac{8 q_{21} q_{11}^3}{81}-2 p_{12}q_{12}q_{11}-\frac{4p_{21} q_{12} q_{11}}{3}+4 p_{12}q_{11}q_{21}+2 p_{21} q_{21}q_{11}-\frac{q_{12}^3}{27}+\\\la{t234} &&\frac{q_{21}^3}{27}-\frac{2q_{12}q_{21}^2}{9}+9 p_{11}p_{21}+\frac{4q_{12}^2q_{21}}{27}\\\nonumber T_5&=&3 \int \left(\dd{{\cal L}(3,4)}{v}\pp{}{x}\dd{H(3,5)}{u}+ \dd{{\cal L}(3,4)}{u}\pp{}{x}\dd{H(3,5)}{v}\right) dx\\\nonumber &=&-\frac{2 q_{11}^6}{729}+\frac{8p_{12}q_{11}^4}{27}+\frac{4 p_{21} q_{11}^4}{27}-\frac{q_{12}^2 q_{11}^3}{243}+3 p_{11}p_{21}q_{21}+\frac{2q_{21}^2 q_{11}^3}{243}-\frac{2 q_{12} q_{21} q_{11}^3}{243}-9 p_{12}^2q_{11}^2-\\\nonumber &&2 p_{21}^2q_{11}^2-9 p_{12} p_{21} q_{11}^2+\frac{p_{11} q_{12} q_{11}^2}{3}-3 p_{11}^2 q_{11}+\frac{p_{21} q_{12}^2 q_{11}}{9}+\frac{q_{21}^4}{27}+54 p_{12}^3-\frac{2 q_{12} q_{21}^3}{27}+\\\la{t534} &&27 p_{12} p_{21}^2+\frac{4 q_{12}^2 q_{21}^2}{81}+81 p_{12}^2 p_{21}-3 p_{11}p_{12} q_{12}-3 p_{11}p_{21}q_{12}-\frac{q_{12}^3q_{21}}{81}- \frac{2 p_{21} q_{12} q_{21} q_{11}}{9}. \end{eqnarray} \noindent Since $r=3$ and $n=4$, the conserved quantities $T_3$ and $T_4$ are trivial. It is easy to check by direct computation that these three conserved quantities are in involution. Furthermore, the dependence of the solution of the KP equation on $t_2$ ({\em i.e.,} y) is governed by the Hamiltonian system generated by $H_2=-T_2$. The same statement holds for $t_5$ and $H_5=-T_5$. We will return to this example in Section \ref{sec:reductions} \vspace*{12pt} \section{Singular and nonsingular Lagrangians}\la{sec:sing} We have shown that the Hamiltonian system in $x$ \rf{ostrodynsys} is completely integrable in the sense of Liouville when Ostrogradksii's theorem applies, $i.e.$, when the Lagrangian ${\cal L}(r,n)$ is nonsingular. This has immediate consequences for the structure of the solutions. On a compact component of phase space, almost all solutions are quasi-periodic with $N$ phases. These phases move linearly on an $N$-dimensional torus \cite{arnold}. Such a torus is topologically equivalent to a compact component of \begin{equation}\la{torus} \Lambda(r,n)=\left\{T_k(\mbf{q},\mbf{p})={\cal T}_k, k \in\Omega(r,n)\right\} \end{equation} \noindent where $\Omega(r,n)=\{\mbox{first N values of}~k, \mbox{not integer multiples of $r$ or $n$}\}$. The constants ${\cal T}_k$ are determined by the initial conditions. The torus $\Lambda(r,n)$ is shared by all $t_k$-flows. The only difference between these different flows from this point of view is the linear motion on the torus. This linear motion determines the $N$ frequencies with respect to the variable $t_k$. We know from \cite{ds1} that a typical solution of the $(r,n)$-th KP equation has $g_{max}=(r-1)(n-1)/2$ phases. In the proof of Theorem \ref{theo:sing}, it is shown that $N=g_{max}$ if one of $r$, $n$ is even and the other one is odd. The case of both $r$ and $n$ even is not allowed, since $r$ and $n$ need to be coprime. On the other hand, if both $r$ and $n$ are odd, $N=g_{max}$ only if $r=3$. Otherwise $N>g_{max}$ and the Ostrogradskii transformation introduces more variables than are needed to span the phase space. The transformation is then not invertible and the Lagrangian ${\cal L}(r,n)$ is nonsingular. This situation does not occur in the work of Bogoyavlenskii and Novikov \cite{bogoyavlenskii}, because for the KdV hierarchy $r=2$. From the results in this section it follows that some rank 1, finite-genus solutions (namely $r>3$ and odd, $n>r$ and odd), give rise to interesting examples of singular Lagrangians. \begin{theo}\la{theo:sing} The Lagrangian ${\cal L}(r,n)$ is singular if and only if $r$ and $n$ are both odd and $r>3$. For all other cases of $(r,n)$, $N=g_{max}$. \end{theo} \noindent {\bf Proof} First note that $\left[R/2\right]+\left[(R+1)/2\right]=R$, for all integers $R$. Using Table \ref{table1} we can rewrite \begin{eqnarray}\nonumber N&=&\sum_{j=2}^{r}N_j\\\nonumber &=&N_2+N_3+\sum_{j=4}^{r}N_j\\\la{rewrite} &=&r+n-4+\sum_{j=4}^{r}\left[\frac{n+r-2 j+2}{2}\right]. \end{eqnarray} In the calculations below, we use that $n$ can always be chosen to be greater than $r$. The calculations are slightly different depending on whether $r$ and/or $n$ are even or odd. Since $r$ and $n$ cannot both be even, there are three cases. \begin{enumerate} \item $r$ is even and $n$ is odd. We write $r=2R$, $n=2M+1$. Use \rf{rewrite}, \begin{eqnarray*} N&=&r+n-4+\sum_{j=4}^{2R}\left(M+R-j+1\right)\\ &=&r+n-4+2RM-3M-2R+3\\ &=&\frac{(r-1)(n-1)}{2}\\ &=&g_{max}. \end{eqnarray*} \item $r$ is odd and $n$ is even. We write $r=2R+1$, $n=2M$. The calculations are similar to those in the previous case. \begin{eqnarray*} N&=&r+n-4+\sum_{j=4}^{2R+1}(M+R-j+1)\\ &=&\frac{(r-1)(n-1)}{2}\\ &=&g_{max}. \end{eqnarray*} \item $r$ is odd and $n$ is odd. We write $r=2R+1$, $n=2M+1$. In this case, the result is quite surprising. \begin{eqnarray*} N&=&r+n-4+\sum_{j=4}^{2R+1}\left(R+M-j+2\right)\\ &=&r+n-4+\frac{(R+M-2)(R+M-1)}{2}-\frac{(M-R+1)(M-R)}{2}\\ &=&\frac{(r-1)(n-1)}{2}+\frac{r-3}{2}\\ &=&g_{max}+\frac{r-3}{2}. \end{eqnarray*} \noindent Hence, in this case, $N\neq g_{max}$. So, if $r$ and $n$ are both odd and $r > 3$, the dimension of the torus $\Lambda(r,n)$ in \rf{torus} is seemingly greater than the maximal number of phases of a solution of the $(r,n)$-th KP equation, according to \cite{ds1}. This dimension exceeds the maximal genus by the amount of $(r-3)/2$, which is a positive integer when $r$ is odd and greater than three. This is an indication that the assumptions necessary for Ostrogradskii's theorem have been violated. Hence $(r,n ~\mbox{both odd}, r>3)$ is a sufficient condition for the Lagrangian ${\cal L}(r,n)$ to be singular. That this condition is necessary for ${\cal L}(r,n)$ to be singular follows from the results of Veselov \cite{veselov}. There it is demonstrated that the dimension of the phase space of the Euler-Lagrange equations \rf{el} is always equal to $2 g_{max}$, which is the desired dimension of the phase space of the Hamiltonian system \rf{hamsys}. In such a case the Lagrangian is only singular if the Ostrogradskii transformation introduces more variables than the dimension of the phase space \cite{krupkova}, which only happens when $r,n$ are noth odd and $r>3$. Hence this condition is also necessary for the singularity of the Lagrangian. \hspace*{\fill}$\rule{3mm}{3mm}$ \end{enumerate} Note that in the generic case of \cite{ds1}, $r=g+1$, $n=g+2$, we always have $g_{max}=N$, since this automatically puts one in case 1 or case 2. \vspace*{12pt} \noindent {\bf Example} \vspace*{8pt} The smallest odd value possible for $r$ is $r=3$. In this case, the count of the number of phases is still right. This corresponds to the Boussinesq hierarchy. The smallest values of $r$ and $n$ where the count of the number of phases is wrong occurs when $(r,n)=(5,7)$. This example is discussed below. In this example, the Lagrangian ${\cal L}(5,7)$ expressed in the $4$ variables $\mbf{\alpha}(5)=(\alpha_0(5),\alpha_1(5),\alpha_2(5), \alpha_3(5))^T$ is singular. It is shown below how one deals with the singular Lagrangian case. As it turns out, a relatively straightforward transformation reduces the Lagrangian ${\cal L}(5,7)$ to a nonsingular Lagrangian, expressed in the transformed variables. The Ostrogradskii transformation is applicable to this Lagrangian and it results in a completely integrable system with $N=g_{max}$. For simplicity, denote $u=\alpha_{3}(5)$, $v=\alpha_2(5)$, $w=\alpha_1(5)$ and $z=\alpha_0(5)$. The Lagrangian is \begin{eqnarray}\nonumber {\cal L}(5,7)&=&-\frac{2u_{xxxx}}{25}\left(z-\frac{5}{2}w_x+\frac{9}{5} v_{xx} \right)_{xx} +\frac{w_{xx}v_{xxxx}}{5}-\frac{7 z_x v_{xxxx}}{25}-\frac{7 z_{xx} w_{xx}}{25}+\\\la{lag57} &&\tilde{\cal L}(u,u_x,u_{xx}, u_{xxx},v,v_x, v_{xx},v_{xxx}, w, w_x, w_{xx}, z, z_x), \end{eqnarray} \noindent where all dependence on the vector $\mbf{X}=(u_{xxxx}, v_{xxxx}, w_{xxx}, z_{xx})^T$ is explicit. We have \begin{equation}\la{ganda} \mbf{\cal G}(5,7)=\left(\begin{array}{cccc} 0&-18/125&1/5&-2/25\\ -18/125&0&0&0\\ 1/5&0&0&0\\ -2/25&0&0&0 \end{array} \right),~\mbf{\cal A}(5,7)=\left(\begin{array}{c} 0\\w_{xx}/5-7z_x/25\\0\\-7 w_{xx}/25 \end{array} \right). \end{equation} \noindent Clearly $\mbf{\cal G}(5,7)$ is singular. If canonical variables $\mbf{q}$ and $\mbf{p}$ were introduced using the Ostrogradskii transformation on this Lagrangian, this would lead to $2N=2(N_2+N_3+N_4+N_5)=26$ phase space variables. The dimension of the phase space can be obtained from the Euler-Lagrange equations \cite{veselov} corresponding to ${\cal L}(5,7)$ and is 24. Since the corank of the matrix $\mbf{\cal G}(5,7)$ is 2, it follows from the Ostrogradskii transformation \rf{ostrotrans} that two linear combinations between $p_{14},p_{24}, p_{33}$ and $p_{42}$ exist. These are also linear in $\mbf{q}$ because $\mbf{\cal A}(5,7)$ is linear in $\mbf{q}$: \begin{equation}\la{lindep} p_{24}=-\frac{2}{5}p_{33}+\frac{7}{25}q_{33},~~~p_{42}=-\frac{18}{25}p_{33}+ \frac{1}{5}q_{33}-\frac{7}{25}q_{42}. \end{equation} \noindent At this point the theory of constrained Hamiltonian systems can be used \cite{dirac}. Another possibility is to use the general methods of Krupkova \cite{krupkova} for singular Lagrangians. However, it is possible to transform the potentials to a new set of potentials such that the Lagrangian is nonsingular when expressed in the new potentials. Motivated by the form of the Lagrangian, let \begin{equation}\la{nonsingtrans} \hat{u}=u,~~ \hat{v}=v,~~ \hat{w}=w,~~ \hat{z}=z-\frac{5}{2}w_x+\frac{9}{5}v_{xx}. \end{equation} \noindent This transformation is clearly invertible. In the new variables, the Lagrangian is \begin{eqnarray}\nonumber {\cal L}(5,7)= -\frac{2}{25}\hat{u}_{xxxx}\hat{z}_{xx}+\frac{1}{250}\hat{v}_{xxx}\hat{w}_{xx}- \frac{7}{25}\hat{v}_{xxxx}\hat{z}_x-\frac{7}{25}\hat{w}_{xx}\hat{}_{xx}+\\ \la{lag57transformed} \hat{\cal L}(\hat{u},\hat{u}_x,\hat{u}_{xx}, \hat{u}_{xxx},\hat{v},\hat{v}_x, \hat{v}_{xx},\hat{v}_{xxx}, \hat{w}, \hat{w}_x, \hat{w}_{xx}, \hat{z}, \hat{z}_x), \end{eqnarray} \noindent up to total derivatives. Define a new vector $\hat{\mbf{X}}=(\hat{u}_{xxxx}, \hat{v}_{xxxx}, \hat{w}_{xx}, \hat{z}_{xx})^T$. The Lagrangian is \begin{equation}\la{newlag} {\cal L}(5,7)=\frac{1}{2}\hat{\mbf{X}}^T \hat{\mbf{\cal G}}(5,7) \hat{\mbf{X}}+\hat{\mbf{\cal A}}(5,7) \hat{\mbf{X}}+\hat{\cal L}(5,7), \end{equation} \noindent with \begin{equation}\la{newg} \hat{\mbf{\cal G}}(5,7)=\left(\begin{array}{cccc} 0&0&0&-2/25\\ 0&0&1/250&0\\ 0&1/250&0&-7/25\\ -2/25&0&-7/25&0 \end{array} \right), \end{equation} \noindent which is nonsingular, hence by Theorem \ref{prop:sing} the Lagrangian is nonsingular. The Ostrogradskii transformation acting on the transformed Lagrangian \rf{newlag} introduces 24 canonical variables, as many variables as the dimension of the phase space \cite{veselov}. \vspace*{12pt} It is not clear if the approach of this example always works, $i.e.$, for other values of $r$ and $n$, both odd. There is no proof and the problem of dealing with a singular Lagrangian for values of $(r,n)$ other than $(5,7)$ describing the $(r,n)$-th KP equation may require the general methods alluded to above. \section{Autonomous symmetry reductions of the Hamiltonian system}\la{sec:reductions} As discussed in the remarks on page \pageref{remrem}, not all solutions of the $(r,n)$-th KP equation have the same number of phases. A generic genus $g$ solution of the KP equation is usually a very non-typical solution of the $(r,n)$-th KP equation, with $r=g+1$ and $n=g+2$. A solution of the $(r,n)$-th KP equation is completely determined by the phase-space variables $q$ and $p$. In the $2N$-dimensional phase space coordinatized by $\mbf{q}$ and $\mbf{p}$, for a given initial condition, the solution evolves on the torus $\Lambda(r,n)$ determined by \rf{torus}. Usually ($i.e.$, for almost all initial data in the class of solutions of the $(r,n)$-th KP equation), this torus is $N$-dimensional \cite{arnold}. However, special solutions correspond to lower-dimensional tori. For example, suppose that $N=2$. Then almost all solutions evolve on a two-dimensional torus: for almost all values of ${\cal T}_1$ and ${\cal T}_2$, the surface $\Lambda$ determined by $T_1={\cal T}_1$ and $T_2={\cal T}_2$ is topologically equivalent to a two-dimensional torus, like the one drawn in Figure \ref{fig1}. For special values of ${\cal T}_1$ and ${\cal T}_2$, however, this torus is degenerate and the resulting `surface' is only a one-dimensional torus, $i.e.$, a circle $C$. This is drawn in Figure \ref{fig1}. \begin{figure}[htb] \centerline{\psfig{file=torus.eps,width=3in}} \caption{\label{fig1} {\bf Usually solutions evolve on 2-D tori. Special solutions correspond to lower-dimensional tori.}} \end{figure} To a certain degree this scenario is the typical one that occurs. For almost all values of the constants $\{{\cal T}_k, ~k\in \Omega(r,n)\}$, the torus $\Lambda(r,n)$ in \rf{torus} is $N$-dimensional. For a special set of values $\{{\cal T}_k, ~k\in \Omega(r,n)\}$, the torus is $(N-1)$-dimensional, corresponding to a solution with $(N-1)$ phases. For an even more limited class of constants $\{{\cal T}_k, ~k\in \Omega(r,n)\}$, the torus $\Lambda(r,n)$ is $(N-2)$-dimensional, corresponding to solutions with $(N-2)$ phases, $etc$. The conditions on the set of constants $\{{\cal T}_k, ~k\in \Omega(r,n)\}$ can be expressed in terms of the variables $\mbf{q}$ and $\mbf{p}$. When these variables satisfy certain constraints (to be derived below), the dimension of the torus $\Lambda(r,n)$ decreases. We have argued above that the conserved quantities $T_k$ for $k \in \Omega(r,n)$ are functionally independent quantities if the $\mbf{q}$ and $\mbf{p}$ variables are considered independent variables. The only way for the torus $\Lambda(r,n)$ to be less than $N$-dimensional is if the conserved quantities are $not$ functionally independent. This is only possible if there are functional relationships between the variables $(\mbf{q},\mbf{p})$. The constraints on the variables $\mbf{q}$ and $\mbf{p}$ are obtained as follows: \begin{enumerate} \item Require that the conserved quantities $T_k$ for $k \in \Omega(r,n)$ be functionally dependent. This is expressed as \begin{equation}\la{funcdep} \mbox{rank}\left( \nabla T_{i_1}~\nabla T_{i_2}~\ldots~\nabla T_{i_N} \right)=g<N, \end{equation} where the indices $i_k$ are the elements of $\Omega(r,n)$ and $\nabla T_{i_k}$ is defined in \rf{nabla}. In this case, $N-g$ of the conserved quantities are functionally dependent on the remaining $g$ functionally independent conserved quantities. Without loss of generality, we assume that the first $g$ conserved quantities remain functionally independent, whereas the last $N-g$ conserved quantities are functionally dependent: \begin{equation}\la{firstg} T_{i_k}=F_{i_k}(T_{i_1}, T_{i_2}, \ldots, T_{i_g}), \end{equation} \noindent for $k=g+1, g+2, \ldots, N$. \item In this case, there are only $g$ functionally independent conserved quantities $T_{i_k}$, $k=1,2,\ldots, g$. The manifold \rf{torus} reduces to \begin{equation}\la{reducedtorus} \Lambda_g(r,n)=\{T_{i_k}={\cal T}_{k}, ~k=1,2,\ldots,g\}, \end{equation} which is topologically equivalent to a $g$-dimensional torus. This $g$-dimensional torus is parametrizable by $g$ phase variables moving linearly on the torus. In other words, if the evolution of a solution of the $(r,n)$-th KP equation is restricted to this $g$-dimensional torus, it has only $g$ phases. Since this solution is also a solution of the KP equation and it has rank 1 and $g$ phases, it must be a genus $g$, rank 1 solution of the KP equation. \item Equations \rf{funcdep} results in a number of conditions on the variables $q$ and $p$: \begin{equation}\la{pqcon} K_j(\mbf{q},\mbf{p})=0, ~~~~j=1, 2, \ldots, m, \end{equation} \noindent where $m$ is the number of conditions. If these conditions are satisfied for the `initial' conditions $\mbf{q}(0)$ and $\mbf{p}(0)$ then they are automatically satisfied for all other $x$-values, $i.e.$, the conditions on the variables $\mbf{q}$ and $\mbf{p}$ are invariant under the $x$-flow. This is easy to see: The conditions \rf{pqcon} are equivalent to the conditions \rf{funcdep} which are equivalent to the conditions \rf{firstg}, which only involve the conserved quantities. Clearly \rf{firstg} are invariant conditions. The conditions on the variables $\mbf{q}$ and $\mbf{p}$ \rf{pqcon} are polynomial in the $\mbf{q}$ and $\mbf{p}$ variables, since all the entries of the matrix on the left-hand side of \rf{funcdep} are polynomial in these variables. In practice, the conditions on the variables $\mbf{q}$ and $\mbf{p}$ \rf{funcdep} can be written as combinations of simpler conditions, \begin{equation}\la{simplercon} K_j=\sum_{k=1}^{m_g} Q_{j,k}(\mbf{q},\mbf{p}) P_k(\mbf{q},\mbf{p}), ~~~~j=1, 2, \ldots, m. \end{equation} \noindent Here both $Q_{j,k}(\mbf{q},\mbf{p})$ and $P_k(\mbf{q},\mbf{p})$ are polynomials in $q$ and $p$. If $P_{k}(\mbf{q},\mbf{p})=0$, for $k=1,2, \ldots, m_g$ then clearly the conditions \rf{pqcon} are satisfied. Clearly the decomposition \rf{simplercon} is not unique. The factors $P_k$ of a given decomposition are not necessarily invariant under the $x$-flow. In order to find a minimal ($i.e.$, smallest number of elements) set of conditions on the $\mbf{q}$ and $\mbf{p}$ variables, the invariance of such factors needs to be tested separately. Since the conditions \rf{funcdep} are invariant, as argued above, such a minimal set of invariant factors is guaranteed to exist. The existence of a {\em minimal} decomposition is essentially a restatement of Hilbert's Basis theorem \cite{abhyankar}. Below, we prove that the number of elements in this set is $m_g=2(N-g)$. Once this minimal set of conditions has been found, \rf{pqcon} can be replaced by the conditions \begin{equation}\la{pqmincon} P_k(\mbf{q},\mbf{p})=0, ~~~k=1,2,\ldots, m_g. \end{equation} \noindent The invariance of the factors $P_k(\mbf{q},\mbf{p})$ is easily tested. It is necessary and sufficient that \cite{olver} \begin{equation}\la{inv} \left\{P_k(\mbf{q},\mbf{p}),H\right\}=0,~~~~~~\mbox{for~} k=1,2,\ldots, m_g, \end{equation} \noindent on the solution manifold $P_k(\mbf{q},\mbf{p})=0$, for $k=1,2,\ldots, m_g$. \item The conditions on the variables $\mbf{q}$ and $\mbf{p}$ \rf{pqmincon} are autonomous, since the conditions do not depend explicitly on $x$. The conditions \rf{pqmincon} determine {\em autonomous invariant symmetry reductions} of the Hamiltonian system \rf{ostrodynsys}. \end{enumerate} \begin{theo}\cite{dullin}\la{dull} In order for a solution of the $(r,n)$-th KP equation to have genus $g$ instead of $N$, $2(N-g)$ conditions need to be imposed on the variables $\mbf{q}$ and $\mbf{p}$, $i.e.$, $m_{g}=2(N-g)$. \end{theo} \noindent{\bf Proof} By Ostrogradskii's theorem, a typical solution of the $(r,n)$-th KP equation resides in the $2N$-dimensional phase space with coordinates $\mbf{q}$ and $\mbf{p}$. The existence of $N$ conserved quantities $T_{i_k}$, for $k=1,2,\ldots, N$ guarantees that the motion starting from any initial condition evolves on a torus determined by the $N$ conditions $T_{i_k}={\cal T}_k$, for $k=1,2,\ldots, N$. Hence this torus is a hypersurface of codimension $N$, or of dimension $2N-N=N$. If we impose that the rank of the matrix $\left(\nabla T_{i_1}, \nabla T_{i_2}, \ldots, \nabla T_{i_N}\right)$ is $N-1$ in order to obtain genus $N-1$ solutions, then the motion of the solutions is restricted to an $(N-1)$-dimensional torus. An $(N-1)$-dimensional hypersurface in a $2N$-dimensional phase space is determined by $N+1$ equations. $N-1$ equations are provided by the relations $T_{i_k}={\cal T}_k$, for $k=1,2,\ldots, N-1$. Hence another two conditions are required on the coordinates $\mbf{q}$ and $\mbf{p}$. The proof of the theorem is now easily obtained by repeating this argument $N-g$ times. \hspace*{\fill}$\rule{3mm}{3mm}$ \vspace*{12pt} \noindent {\bf Remarks} \begin{description} \item[(a)~] The fact that $m_{N-1}=2$ is not easily seen from \rf{funcdep}. For \rf{funcdep} to be satisfied with $g=N-1$, the determinants of all $N\times N$ minors need to be zero. This seemingly results in $\binomial{2N}{N}$ $N$-dimensional minors of which $2N-1$ are functionally independent. The determinant of all these minors are decomposable in terms of two polynomials $P_1(\mbf{q},\mbf{p})$ and $P_2(\mbf{q},\mbf{p})$, which are both invariant under the $x$-flow. This is explicitly worked out in the example below, for $N=3$ and $g=2$. \item[(b)~] It should be mentioned that in \cite{ds1}, some ideas were given to find conditions on nontypical solutions of the $(r,n)$-th KP equation. Those ideas were correct, but seemingly hard to implement, as is seen in the example discussed there. This same example is discussed below. The algorithm offerered in this section to find the determining conditions on the nontypical solutions of the $(r,n)$-th KP equation is much more efficient. \end{description} \noindent {\bf Example: Generic genus 2 solutions of the KP equation} \vspace*{6pt} \noindent A generic solution of genus 2 of the KP equation is a solution of the $(3,4)$-th KP equation. The Hamiltonian system for this case was discussed on page \pageref{ham34}. The Hamiltonian system \rf{ostrodynsys} corresponding to the Hamiltonian \rf{ham34} is three-dimensional, hence a typical solution of this system executes linear motion on a 3-torus, topologically equivalent to \begin{equation}\la{3torus} \Lambda(3,4)=\left\{T_1={\cal T}_1,~T_2={\cal T}_2, ~T_5={\cal T}_5 ~\right\}. \end{equation} \noindent Such a solution has three phases and is hence a rank 1, genus 3 solution of the KP equation. Nevertheless, the generic rank 1, genus 2 solutions of the KP equation are obtained as solutions of the $(3,4)$-th KP equation. These solutions are special, nontypical solutions of the $(3,4)$-th KP equation. They are obtained by imposing an autonomous invariant symmetry reduction on the Hamiltonian system \rf{ostrodynsys}, as outlined in this section. The condition \rf{funcdep} for a solution of the $(3,4)$-th KP equation to have genus 2 is \begin{equation}\la{g2con} \mbox{rank}\left( \begin{array}{cccccc} \displaystyle{\pp{T_1}{q_{11}}} & \displaystyle{\pp{T_1}{q_{12}}} & \displaystyle{\pp{T_1}{q_{21}}} & \displaystyle{\pp{T_1}{p_{11}}} & \displaystyle{\pp{T_1}{p_{12}}} & \displaystyle{\pp{T_1}{p_{21}}}\\ \vspace{-5pt}\\ \displaystyle{\pp{T_2}{q_{11}}} & \displaystyle{\pp{T_2}{q_{12}}} & \displaystyle{\pp{T_2}{q_{21}}}& \displaystyle{\pp{T_2}{p_{11}}} & \displaystyle{\pp{T_2}{p_{12}}} & \displaystyle{\pp{T_2}{p_{21}}}\\ \vspace{-5pt}\\ \displaystyle{\pp{T_3}{q_{11}}} & \displaystyle{\pp{T_3}{q_{12}}} & \displaystyle{\pp{T_3}{q_{21}}}& \displaystyle{\pp{T_3}{p_{11}}} & \displaystyle{\pp{T_3}{p_{12}}} & \displaystyle{\pp{T_3}{p_{21}}}\\ \end{array} \right)=2. \end{equation} From Theorem \ref{dull}, it follows that \rf{g2con} is equivalent to two invariant conditions on the variables $q_{11}$, $q_{12}$, $q_{21}$, $p_{11}$, $p_{12}$ and $p_{21}$. These conditions are readily found in this specific case \cite{decthesis}. The expressions for these conditions are quite long, so we do not repeat them here. Let us denote them by \begin{equation}\la{g2cond} P_1(\mbf{q}, \mbf{p})=0,~~p_2(\mbf{q}, \mbf{p})=0. \end{equation} There is a more geometrical way to look at the conditions \rf{g2con}. Consider the three-dimensional space spanned by the conserved quantities $T_1, T_2$ and $T_5$. If the conditions \rf{g2cond} are satisfied, there is a functional relationship between the three conserved quantities: \begin{equation}\la{funcrelcons} \Omega: f(T_1,T_2,T_5)=0, \end{equation} \noindent which represents a surface in the space spanned by $T_1, T_2$ and $T_5$. By solving the conditions \rf{g2cond} for two of the phase space variables ($p_{11}$ and $p_{12}$ respectively, for instance), and substituting the result in the form of the conserved quantities \rf{t134}, \rf{t234} and \rf{t534}, a parametric representation of the surface $\Omega$ is obtained: \begin{equation}\la{g2parametric} \Omega:\left\{ \begin{array}{rcl} T_1&=&T_1(q_{11}, q_{12}, q_{21}, p_{21}),\\ T_2&=&T_2(q_{11}, q_{12}, q_{21}, p_{21}),\\ T_5&=&T_3(q_{11}, q_{12}, q_{21}, p_{21}). \end{array} \right. \end{equation} \noindent Apparently, two too many parameters are present in this set of equations, since the parametric representation of a surface should only contain two parameters. However, the existence of a functional relationship \rf{funcrelcons} guarantees that these parameters appear only in two different combinations such that there are essentially two parameters present in \rf{g2parametric}. The most convenient way to plot the resulting surface is to equate two of the `parameters' $q_{11}$, $q_{12}$, $q_{21}$ and $p_{21}$ to zero, while letting the remaining two vary. In Figure \ref{fig2}, two different views of the surface $\Omega$ are given. \begin{figure} \begin{tabular}{cc} \psfig{file=surf1.ps,width=3in} & \psfig{file=surf2.ps,width=3in}\\ (a) & (b)\\ \end{tabular} \caption{\label{fig2} {\bf The surface $\Omega$ from two different view points. The cusp of the surface is at the origin. Figure (b) shows the same surface as Figure (a), but rotated around the $T_5$-axis by 180 degrees. This figure was obtained using \rf{g2parametric} with $q_{21}=0=p_{21}$.}} \end{figure} Every point in the space spanned by $T_1$, $T_2$ and $T_5$ corresponds to a three-dimensional torus, on which the motion of the solutions of the $(3,4)$-th KP equation takes place. A point on the surface $\Omega$ corresponds to a degenerate three-dimensional torus, on which there are in essence only two independent directions, analogous to the idea demonstrated in Figure \ref{fig1}. In other words, points on the surface $\Omega$ correspond to two-dimensional tori and to genus two solutions of the KP equation. These genus two solutions are the generic rank 1, genus 2 solutions of the KP equation, as argued above. Note that a more drastic reduction to genus one solutions is possible. These solutions correspond to points on the singular curves on the surface $\Omega$. As the genus one solutions are more easily obtained from the $r=2$ case, this case is not essential. \section{Parameter count}\la{sec:parameters} The previous sections put us in a position to count easily the parameters that determine a solution of the $(r,n)$-th KP equation. The first column of table \ref{table2} lists the parameters that determine a solution of the $(r,n)$-th KP equation. They are the `initial' values variables $\mbf{q}(0)$ and $\mbf{p}(0)$ (not only for the $x$-flow, but any other $t_k$-flow), the constants $h_k$ that are the coefficients of the Casimir functionals in \rf{lagrangian}, the constants $d_k$ which are the coefficients of the lower-order flows in \rf{lagrangian} and the constant $u_1$, discussed in the remark on page \pageref{u1rem}. \settowidth{\mylength}{$\displaystyle{N=\frac{(r-1)(n-1)}{2}}$} \settoheight{\myheight}{\framebox{$\displaystyle{N=\frac{(r-1)(n-1)}{2}}$}} \addtolength{\myheight}{8pt} \begin{table} \begin{center} \caption{\bf The number of parameters determining the solutions of the $(r,n)$-th KP equation.\la{table2}} \vspace*{0.2in} \begin{tabular}{|c|c|c|c|} \hline & typical solution of & generic genus $g$ & non-typical solution of \\ &the $(r,n)$-th KP equation& solution of (KP)& the $(r,n)$-th KP equation\\ \hline\hline \pb{$\mbf{q}(0)$} & $\displaystyle{N=\frac{(r-1)(n-1)}{2}}$ & $g$ & $g$\\ \hline \pb{$\mbf{p}(0)$} & $\displaystyle{N=\frac{(r-1)(n-1)}{2}}$ & $g$ & $g$\\ \hline\hline \pb{$h_k$} & $r-1$ & $g$ & $r-1$\\ \hline \pb{$d_k$} & $n-2$ & $g$ & $n-2$\\ \hline \pb{$u_1$} & 1 & 1 & 1\\ \hline\hline \pb{total \#} & $rn-1$ & $4g+1$ & $2g+r+n-2$\\ \hline \end{tabular} \end{center} \end{table} How many of each of these parameters determine a typical solution of the $(r,n)$-th KP equation is indicated in the second column. A typical solution of the $(r,n)$-th KP equation has $N=(r-1)(n-1)/2$ variables $\mbf{q}$ and $N$ variables $\mbf{p}$. Each of these is determined by its initial conditions $\mbf{q}(0)$ and $\mbf{p}(0)$. For a typical solution of the $(r,n)$-th KP equation, $N$ is also the genus of the solution. Any solution of the $(r,n)$-th KP equation is determined by $r-1$ Casimir functionals (see \rf{lagrangian}). Also from \rf{lagrangian}, it follows that $n-2$ lower-order flows are included, accounting for the coefficients $d_k$. With the addition of the constant $u_1$, this results in a total number of $rn-1$ parameters that determine a typical solution of the $(r,n)$-th KP equation. The third column expresses the number of parameters in terms of the genus of the solution, for a generic genus $g$ solution of the KP equation. For such a solution $r=g+1$ and $n=g+2$ \cite{ds1}. Furthermore, the Hamiltonian system \rf{ostrodynsys} reduces significantly such that there are exactly $g$ variables $\mbf{q}$ and $\mbf{p}$. The total number of parameters adds up to $4g+1$. This is a known result, implied for instance in \cite{dub}. Not every nontypical solution of the $(r,n)$-th KP equation is generic. For such solutions, the number of variables $\mbf{q}$ is equal to the genus $g$ of the solution, as is the number of variables $\mbf{p}$. These results are given in the last column of Table \ref{table2}. There is an important distinction between the different types of parameters in table \ref{table2}. The entries in the top two rows have dynamical significance: they are initial values for the variables $\mbf{q}$ and $\mbf{p}$. The Hamiltonian system \rf{ostrodynsys} is a dynamical system for the determination of the variables $\mbf{q}$ and $\mbf{p}$. The other parameters merely show up as parameters in this Hamiltonian system. This distinction between two kinds of parameters, dynamical and nondynamical, is to our knowledge new. \section{Minimal Characterization of the initial data} A rank 1, finite-genus solution of the KP equation is completely determined by a solution of an $(r,n)$-th KP equation, for a certain $r$ and $n$. The $(r,n)$-th KP equation is given by the Euler-Lagrange equation \rf{el}. This is a set of $(r-1)$ ordinary differential equations in $x$. Various quantities appear in this system of ordinary differential equations: $(r-1)$ potentials $\mbf{\alpha}(r)$ and their derivatives with respect to $x$, $(r-1)$ constants $h_k$, $(n-2)$ constants $d_k$ and one constant potential $u_1$. Next we argue that the knowledge of the initial condition for KP, $u(x,y,t=0)$ along one direction (say the $x$-axis) is sufficient to determine the corresponding rank 1, finite genus solution for all $x$, $y$ and $t$. \begin{enumerate} \item If the initial condition is specified along the $x$-axis for $y=0$, all potentials and their derivatives can be found at any point on the $x$-axis (Note that a rank 1, finite-genus solution is analytic in all its independent variables). This is done in the following way: The Euler-Lagrange equations and their derivatives with respect to $x$ determine $algebraic$ conditions on the potentials and their derivatives, as well as the unknown parameters $h_k$, $d_k$ and $u_1$. In these conditions, $u$ and its derivatives with respect to $x$ are known, by assumption. Hence taking more derivatives of the Euler-Lagrange equations \rf{el} with respect to $x$ adds more conditions than unknowns. Taking enough derivatives of the Euler-Lagrange equations, we obtain a set of polynomial equations for the unknown potentials and their derivatives, as well as the parameters $h_k$, $d_k$ and $u_1$. \item We have already argued that the knowledge of the $x$-dependence completely determines the dependence of the solution on any higher-order time variable: the Hamiltonians determining the dependence of the solution on $t_k$ are conserved quantities for the $x$-evolution of the solution, $H_k=-T_k$. Furthermore, the initial conditions $\mbf{q}(0)$, $\mbf{p}(0)$ for $t_k$ are identical to the initial conditions for the $x$-evolution at $x=0$, $y=0$. \end{enumerate} This shows that it suffices to specify the rank 1, finite-genus initial condition along one direction: $u(x, y=0, t=0)$. It is clear that the above argument can be repeated if the initial condition is specified along any of the higher-order flows. This is specifically relevant for $t_2~(y)$ and $t_3~(t)$. \vspace*{12pt} \noindent {\bf Remarks} The procedure for finding the parameters $h_k$, $d_k$ and $u_1$ given above is not very practical. A large number of derivatives of the potential are required and a large polynomial system needs to be solved to find the values of the parameters. \section*{Acknowledgements} The author acknowledges useful discussions with O. I. Bogoyavlenskii and A. P. Veselov. H. Segur is thanked for his continued support and the invaluable interactions with the author on this subject. This work was carried out at the University of Colorado and supported by NSF grant DMS-9731097. \bibliographystyle{unsrt}

https://arxiv.org/abs/2012.01697

General Behaviour of P-Values Under the Null and Alternative

Hypothesis testing results often rely on simple, yet important assumptions about the behaviour of the distribution of p-values under the null and the alternative. We examine tests for one dimensional parameters of interest that converge to a normal distribution, possibly in the presence of nuisance parameters, and characterize the distribution of the p-values using techniques from the higher order asymptotics literature. We show that commonly held beliefs regarding the distribution of p-values are misleading when the variance and location of the test statistic are not well-calibrated or when the higher order cumulants of the test statistic are not negligible. Corrected tests are proposed and are shown to perform better than their first order counterparts in certain settings.

\section{Introduction} Statistical evidence against a hypothesis often relies on the asymptotic normality of a test statistic, as in the case of the commonly used Wald or score tests. Many authors ignore the asymptotic nature of the argument and assume that in finite samples the distribution of the test statistic is indeed normal. This perfunctory approach generates misleading beliefs about the $p$-value distribution, such as i) the distribution of the $p$-values under the null is exactly uniform or that ii) the cumulative distribution function (henceforth, cdf) of the $p$-value under the alternative is concave. However, there are important exceptions from these rules, e.g. discrete tests are not normally distributed in any finite sample settings, so that the distribution of the $p$-values under the null is certainly not uniform. Similarly, it is not obvious that the cdf is concave under the alternative as we will illustrate with some examples. Testing procedures aimed at controlling the family-wise error rate (FWER) or the false discovery rate (FDR, see \cite{BH}) typically assume that i) or ii) holds. In \cite{Cao2013}, the authors examine the optimality of FDR control procedures when i) or ii) are violated and provide alternative conditions to maintain said optimality. Clearly, a more precise characterization of the $p$-value distribution that accounts for the approximation error is pivotal in controlling the occurrence of false discoveries. Complicating matters even further, the issue of calibrating the location and variance of the test statistic is often overlooked, particularly under the alternative. Under the alternative the test statistic can be improperly re-scaled since often the variance of the test statistic is obtained under the null. While under the null, the test statistic may not have zero mean and may also not be correctly standardized, thus making the standard Gaussian approximation suspect. The problem of biases in the variance and expectation is aggravated in the presence of a large number of nuisance parameters. For instance, while it has been demonstrated in \cite{DiCicio} that the profile score statistic has a location and variance bias under the null, in Section \ref{subsec:score} we show that the variance bias can persist under the alternative. These concerns motivate us to perform a systematic study of the $p$-value distribution in the presence of information or location biases under the null and alternative, while accounting for the approximation error resulting from the use of asymptotic arguments. We explore how certain asymptotically non-vanishing and vanishing biases in the variance and location of the test statistic can occur in finite samples, violating the assumptions generally placed on the null and alternative distributions of the $p$-values. We study both continuous and discrete distributions supported on lattices. In doing so we include all approximation errors, including those induced by discreteness, to fully characterize the behaviour of the distributions of $p$-values under the null and alternative. This work extends the results of \cite{hung}, who studied the distribution of $p$-values under alternative assuming the test statistic is normally distributed, to a broader framework. We focus on univariate test statistics for a one dimensional parameter of interest based on sums of independent random variables, possibly in the presence of a large number of nuisance parameters. These types of test statistics are commonly used to infer the significance of individual coefficients in most regression models. The results of the paper are in the same vein as those found in \cite{hall2013bootstrap} and \cite[\S~3]{kolassa1994series}, whose objective was the coverage properties of confidence intervals. We expand their results to the $p$-values, motivated by the multitude of scientific investigations that rely on the $p$-value distribution rather than confidence intervals. We begin with a simple example illustrating how the standard assumptions on the null and alternative distributions of the $p$-values can be violated in practice. \begin{example} We wish to test the null hypothesis $H_0: \beta = 0.01 $ against the alternative $H_1: \beta = 0.01/1.05$, where $\beta$ is the rate parameter of a gamma distribution, based on 750 observations $x_1, \cdots, x_{750}$, assuming that the shape parameter is known to be $\alpha = 0.01$. From the central limit theorem, we know that the test statistic \begin{align*} S_n = \sqrt{n} \left(\frac{\bar{X} - 1}{\sigma} \right) \rightarrow N(0, 1), \end{align*} so we are able to obtain a two-sided $p$-value based on the standard normal distribution. We plot the histograms of the $p$-values obtained under the null and alternative in Figure \ref{fig:test}. The plots are obtained by simulation using 100,000 replications. We see on Figure \ref{fig:test} that the distribution of the $p$-values obtained from the simulations does not adhere to its expected behaviour under the null or the alternative. The upper left plot in Figure \ref{fig:test} shows a marked departure from the $U(0,1)$ distribution expected from the null. Thus, a typical rejection rule which assumes uniformity of the $p$-value distribution under the null will not provide type I error control for certain choices of $\alpha$. For example, if we desire a $10^{-4}$ significance level, we obtain a type I error approximately equal to $1.579*10^{-3}$, which is fifteen times higher than the nominal level. Under a local alternative, the upper right plot in Figure 1 shows that the $p$-value distribution may not be stochastically smaller than a $U(0,1)$. The resulting lack of concavity of the distribution $p$-value under the alternative can violate the typical assumption that the false negative rate is strictly decrease and the FDR is increasing in the nominal control level $\alpha$ in the multiple testing setting; see \cite{Cao2013}. Note that the cause for this poor calibration is not the low sample size. \end{example} \begin{figure}[ht] \centering \subfloat{\includegraphics[width = 2in]{null_1_1.png}} \subfloat{\includegraphics[width = 2in]{alt_1_105.png}}\\ \subfloat{\includegraphics[width = 2in]{gamma_saddle}} \subfloat{\includegraphics[width = 2in]{gamma_saddle_alt}} \caption{Distribution of $p$-values under $H_0$ and $H_1$ in Example 1. {\it Upper left}: $p$-values obtained under $H_0$ from the normal approximation. {\it Upper right}: $p$-values obtained under $H_1$ from the normal approximation. {\it Lower left and right}: corrected $p$-values for the null and alternative, respectively, using the saddlepoint approximation that is introduced in Section 3. The number of samples is 750. The upper left panel clearly does not exhibit uniformity and upper right panel's distribution does not appear to have a concave cdf. We plot the theoretical prediction from Theorem 1 in blue for the upper left and upper right panels.} \label{fig:test} \end{figure} To assure the reader that the above example is not a singular aberration, we present in Figure \ref{fig:lung} the histogram of over 13 million $p$-values from the genome-wide association study of lung cancer generated from the UK Biobank data \citep{biobank}. These $p$-values are produced by the Neale Lab \citep{neale}, based off 45,637 participants and 13,791,467 SNPs; SNPs with minor allele frequency less than 0.1\% and INFO scores less than 0.8 were excluded from the analysis. We note that the histogram exhibits a similar behaviour to the one seen in Example 1, i.e., the distribution of $p$-values exhibits a secondary mode that is far from zero. \begin{figure}[ht] \centering \includegraphics[width=8cm, height = 5cm]{lung_cancer} \caption{Empirical p-value distribution based on a genome wide association study ($n=45,637$) of lung cancer. } \label{fig:lung} \end{figure} Figure \ref{shapes} briefly summarizes the shapes that the density of the $p$-value distribution might take for two-tailed tests, based on the results in Theorems 1 and 2 that are introduced in Section 2. The descriptions in Table 1 verbalise the various mathematical conditions that can lead to the four shapes in Figure \ref{shapes}. In practice it is possible to have combinations of the shapes listed in Figure \ref{shapes}, as the observed test statistics may not be identically distributed and can be drawn from a mixture of the null and alternatives hypotheses. \begin{figure}[H] \centering \subfloat[Shape 1]{\includegraphics[width = 2in]{shape_1.png}} \subfloat[Shape 2]{\includegraphics[width = 2in]{shape_2.png}}\\ \subfloat[Shape 3]{\includegraphics[width = 2in]{shape_3.png}} \subfloat[Shape 4]{\includegraphics[width = 2in]{shape_4.png}} \caption{General chart of the behaviour of $p$-values under the null and alternative for a two tailed test. Shapes 1 to 3 were obtained from simulating from the null and alternative from Example 1 with different parameters $\alpha$ and $\beta$. Shape 4 was obtained when using a misspecified variance, as detailed in Section 2.2. }\label{shapes} \end{figure} \begin{table}[h] \caption{Description of the test statistic's characteristics and the resulting shapes (as shown in Figure \ref{shapes}) of the $p$-value distribution under the null or alternative hypothesis.} \fbox \begin{tabular}{ | m{3em} | m{6cm}| m{6cm} |} \hline Shape & Null & Alternative \\ \hline 1 & The typical uniform shape. & Possible if effect size is small. \\ \hline 2 & Possible if variance is misspecified, underestimated. & Typical behaviour. \\ \hline 3 & Possible if test statistic has large higher order cumulants, see Example 1. & Possible if effect size is small and the higher order cumulants are large, see Example 1. \\ \hline 4 & Possible if variance is misspecified, overestimated, see Example 3. & Possible if variance is misspecified, overestimated and the effect size is small, see Example 3. \\ \hline \end{tabular}} \end{table} Section 2 contains the main theoretical results of this paper, Theorems 1 and 2, which characterize the distribution of $p$-values under the null and alternative. Section 2.1 examines the $p$-value distribution resulting from the score test, while Section 2.2 studies specific examples. Section 3 provides numerical results and considers some remedies aimed at calibrating the $p$-value distribution. Section 4 closes the paper with a discussion of the implications of our results and some recommendations to practitioners. \section{Distribution of $p$-values under Non-Normality} All theoretical details and proofs, as well as a brief introduction of the concepts needed for the proof of Theorems 1 and 2, are deferred to the Supplementary Materials. We consider the case where the test statistic, $S_n$, can be discrete and may also have a non-zero mean under the null and a non-unit variance under the null or alternative. We assume that $\psi$ is a one dimensional parameter of interest and $\lambda$ is a vector of nuisance parameters. Without loss of generality, let the statistic $S_n$ either be used to test the null hypothesis $H_0: \psi = \psi_0$ for a two-sided test or $H_0: \psi \geq \psi_0 $ for a one-sided test. All results are given in terms of the cdf. Theorems 1 and 2 deal with the case where $S_n$'s distribution is continuous and discrete respectively. We first consider the case where the statistic $S_n$ admits a density. We typically assume that $S_n$ has been appropriately calibrated such that $\mathrm{E}[S_n]= 0$ under the null, and $\mathrm{E}[S_n] \neq 0$ under the alternative hypothesis. We let $p(S_n)$ denote the $p$-value obtained from the test statistic $S_n$. However, as discussed in the introduction, the mean of $S_n$ may not be exactly $0$ under the null due to a location bias. We also would expect that the variance of the test statistic should be 1 under the null and alternative, which may not be the case for all test statistics however; see Example 3. The location bias complicates the precise determination of whether $S_n$'s distribution should be considered under the null or the alternative. However, note that Theorem 1 statement is applicable under both the null and alternative, since its conclusion depends only on the expectation, variance, and the other cumulants of the statistic $S_n$, regardless of the true hypothesis. We first introduce some notations: \begin{itemize} \item[(i)] $Z_p$ is the $p$-th quantile of a standard normal distribution. \item[(ii)] $\rho_{n,i}$ is the $i$-th order standardized cumulant of $S_n$, and $\rho_n$ is a vector containing all cumulants. \item[(iii)] $\phi(x)$ is the standard normal density. \end{itemize} \begin{theorem}\label{th:cont_approx} Let $X_1,\ldots, X_n$ be a sequence of continuous, independent random variables. Set $S_n= \sqrt{n} (\bar{X}_n - a_n)/b_n$ where $\bar{X}_n = n^{-1}\sum_{i = 1}^n X_i$, and let $\{a_n\}_{n\ge 0}$, $\{b_n\}_{n\ge 0}$ be two sequences of real numbers. Let $\mathrm{E}[S_n]= \mu_n$, ${\mathrm{Var}}(S_n) = v_n^2$, and $\rho_n$ denote the cumulants of $(S_n - \mu_n)/v_n$. Then the CDF of the one-sided $p$-value is \begin{align} \mathbb{P} (p(S_n) < t) = \Phi\left(\frac{Z_t - \mu_n}{v_n}\right) -E_2\left(\frac{Z_t - \mu_n}{v_n}, \rho_{n} \right) + O\left(n^{-3/2}\right), \end{align} and the CDF of the two-sided $p$-value is: \begin{align} \mathbb{P} (p(S_n) < t) &= 1 + \Phi\left(\frac{Z_{t/2} - \mu_n}{v_n} \right) - \Phi\left(\frac{-Z_{t/2} - \mu_n}{v_n}\right)\nonumber \\ &+E_2\left(\frac{Z_{t/2} - \mu_n}{v_n} , \rho_n\right) - E_2\left( \frac{-Z_{t/2} - \mu_n}{v_n}, \rho_n\right) + O\left(n^{-3/2} \right), \label{eq:cont_one} \end{align} where, \begin{align} E_2(t, \rho_n) = -\phi(t)\Big\lbrace \frac{\rho_{n,3} H_2(t)}{6} + \frac{\rho_{n,4} H_3(t)}{24} + \frac{(\rho_{n,3} )^2 H_5(t)}{72} \Big\rbrace, \label{eq:cont_two} \end{align} and $H_j(t)$ denotes the j-th Hermite polynomial. \end{theorem} \begin{remark} The $j$-th order Hermite polynomial is a polynomial of $j$-th degree defined through the differentiation of a standard normal density. A table of the Hermite polynomials is given in the Supplementary Materials. \end{remark} \begin{remark} Should an approximation to the probability density of the $p$-value distribution be desired, it can be obtained from differentiating Equations (\ref{eq:cont_one}) and (\ref{eq:cont_two}). \end{remark} In general $E_2 = O(1/n^{1/2})$, however in the the case that $\mu_n = 0$, $E_2(Z_{t/2}/ v_n , \rho_n) - E_2( -Z_{t/2}/v_n, \rho_n) = O(1/n)$ for two-sided tests due to cancellations which occur in the difference of the odd Hermite polynomials. We refer to terms in $E_2(t, \rho_n)$ as the higher order terms. Therefore, supposing $\mu_n = 0$ under the null, meaning the sequence $a_n = \mathbb{E}[\bar{X}_n]$, we obtain the following corollary: \begin{corollary} Assume the setting and notation from Theorem \ref{th:cont_approx} and suppose that under the null we have $\mathrm{E}[S_n]= 0$, and $\mathrm{Var}(S_n) = 1$. The CDF of the distribution of the $p$-values for a one-sided test under the null is \begin{align} \mathbb{P} \left(p(S_n) < t \right) = t + O\left( n^{-1/2}\right), \end{align} and the CDF of the distribution of the $p$-values for two-sided test under the null is \begin{align} \mathbb{P} \left(p(S_n) < t \right) &= t + O\left(n^{-1} \right). \end{align} \end{corollary} Corollary 1 shows that the two-sided test is preferable unless there is a scientific motivation for using the one-sided test. The case when $S_n$ has a discrete distribution supported on a lattice is covered in Theorem \ref{th-discrete}. \begin{theorem} \label{th-discrete} Let $X_1, \cdots, X_n $ be a sequence of independent discrete random variables where $X_i$ has mean $m_i$. Suppose that $X_i - m_i$ is supported on a lattice of the form $c + j\cdot d$, for $j \in \mathbb{Z}$ and for all $1\le i \le n$. Assume $d$ is the largest number for which this property holds. Set $S_n= \sqrt{n} (\bar{X}_n - a_n)/b_n$, where $\bar{X}_n = n^{-1}\sum_{i = 1}^n X_i$, $\mathrm{E}[S_n]= \mu_n$, $\mathrm{Var}(S_n) = v_n^2$, $\rho_n$ as the cumulants of $(S_n - \mu_n)/v_n$ and $d_n = d \: v_n /(\sqrt{n}b_n) $. Then the CDF of the one-sided $p$-value is \begin{align*} \mathbb{P} (p(S_n) < t) &= \Phi\left( \frac{Z_t - \mu_n}{v_n} \right) +E_2 \left( \frac{ Z_t - \mu_n}{v_n}, \rho_n \right) + C_2\left(\frac{ Z_t - \mu_n}{v_n}, \rho_n \right) + O\left(n^{-3/2}\right), \end{align*} and the CDF of the two-sided $p$-value is \begin{align*} \mathbb{P} (p(S_n) < t) &=1 + \Phi\left( \frac{Z_{t/2} - \mu_n}{v_n} \right) - \Phi\left(\frac{-Z_{t/2} - \mu_n}{v_n} \right) +E_2\left(\frac{Z_{t/2} - \mu_n}{v_n}, \rho_n \right) \\ &- E_2 \left(\frac{-Z_{t/2} - \mu_n}{v_n}, \rho_n \right) +C_2 \left(\frac{Z_{t/2} - \mu_n}{v_n}, \rho_n \right) - C_2\left(\frac{-Z_{t/2} - \mu_n}{v_n}, \rho_n \right) +O \left(n^{-3/2} \right), \end{align*} where, \begin{align*} C_2(t, \rho_n) = - d_n Q_1\Big( \frac{ t - \sqrt{n}c}{d_n}\Big) \Big( 1 + \frac{\rho_{n,3} H_3(t)}{6} \Big) + \frac{ d_n^2}{2} Q_2\Big( \frac{ t - \sqrt{n}c}{d_n}\Big), \end{align*} and $Q_j(t)$ are periodic polynomials with a period of $1$. On $[0,1)$, they are defined by \begin{align*} Q_1(t) = t - \frac{1}{2}, \quad Q_2(t) = t^2 - 2t +\frac{1}{6}, \end{align*} and $E_2(t, \rho_n)$ and $H_j(t)$ are defined as in Theorem 1. \end{theorem} \begin{corollary} Assume the setting and notation from Theorem \ref{th-discrete} and suppose that under the null $\mathrm{E}[S_n]= 0$, and $Var(S_n) = 1$. Then the $p$-values obtained from one or two-sided tests satisfy \begin{align} \mathbb{P} (p(S_n) < t) = t + O\left(n^{-1/2}\right). \end{align} \end{corollary} Note that the convergence is slower by a factor of $n^{-1/2}$ compared to the continuous case for a two-sided test. This is due to the jumps in the CDF which are of order $O(n^{-1/2})$. \begin{remark} Under the alternative, the $p$-value distribution depends on the effect size, $\mu_n$, as well as the magnitude of the higher order cumulants. However, for large values of $\mu_n$ the impact of the higher order terms will be negligible, as $E_2$ is a product of an exponential function and a polynomial function which decays to 0 asymptotically in $\mu_n$. We explore this further in Example 4. \end{remark} \begin{remark} When performing multiple hypothesis testing corrections, the $p$-values of interest are often extremely small. Therefore from Corollary 1 and 2, we see that a large amount of samples is needed to guarantee the level of accuracy required since the approximation error is additive. \end{remark} \subsection{An Application of the Main Theorems: The Score Test} \label{subsec:score} We examine the broadly used score test statistic, also known as the Rao statistic. The popularity of the score statistic is due to its computational efficiency and ease of implementation. In the presence of nuisance parameters, the score statistic is defined through the profile likelihood. Suppose that the observations $y_i$'s are independent then \[ l_\text{pro}(\psi) = \sup_{\lambda} l(\psi, \lambda; Y) = l(\psi, \hat\lambda_\psi; Y), \] where $\hat\lambda_\psi$ denotes the constrained maximum likelihood estimator. The score statistic is defined as: \[ S_n(\psi_0) = \frac{l_{\text{pro}}^\prime(\psi_0)}{ \lbrace l^{\prime\prime}_{\text{pro} } (\psi_0) \rbrace^{1/2}}= \sum_{i = 1}^n \frac{ \frac{d}{d\psi} l(\psi, \hat\lambda_\psi; y_i)}{ \left\lbrace\sum_{i = 1} \frac{d^2}{d\psi^2} l(\psi, \hat\lambda_\psi; y_i)\right\rbrace^{1/2}} \xrightarrow{D} N(0,1), \] under the usual regularity assumptions. Due to the form of $S_n$, we may apply Theorem 1 or 2. The presence of nuisance parameters induces a bias in the mean and variance of the score statistic; see \cite{profile_bias} and \cite{DiCicio}. Thus, it is not the case that the mean of the score statistic is 0 and the variance is 1 under the null, as the profile likelihood does not behave like a genuine likelihood and does not satisfy the Bartlett identities. In general this problem is compounded if the number of nuisance parameters is increased, as we illustrate below. We only discuss the location bias, since the formulas for the information or variance bias are much more involved and compromise the simplicity of the arguments. From \cite{profile_bias}, the bias of the profile score under the null is: \begin{align} \mathbb{E}\lbrace l^\prime_{\text{pro}}(\psi_0) \rbrace &= \alpha_n + O\left( n^{-1} \right), \end{align} where the term $\alpha_n = O(1)$. The form of $\alpha_n$ is given in the Supplementary Materials. To estimate the effect of the dimension of the nuisance parameter on the size of the bias, we use a similar argument as \cite{laplace}, in which they count the number of nested summations that depends on $k$, the number of parameters in the model, to estimate the rate of growth of a function in $k$. From the expression of $\alpha_n$ given in the Supplementary Materials, we obtain at most 4 nested summations which depends on $k$ therefore the bias of the profile score is of order $O(k^4)$ in the worst case scenario. The rather large location bias can be impactful as it may induce a perceived significance when $k$ is large, an example using Weibull regression is given in Section 3.3. A similar argument can be applied to the information bias; see \cite{DiCicio} for a comprehensive discussion on the form of these biases. The information bias for the score statistic can also be highly influential under the alternative. In that case, the expected value of the score statistic is non-zero, which is desirable, but the variance of the statistic $S_n$ can be either over- or underestimated. Since the true parameter value is not $\psi_0$, there is no guarantee that ${l_{\text{pro}}^{\prime\prime}(\psi_0)}$ gives the correct standardization. If the estimated variance is larger than the true variance of the score, then it is possible to obtain Shape 4 in Figure \ref{shapes}, which violates the concavity assumption for $p$-value distribution's CDF under the alternative. Further, if we assume that under the null the $p$-value distribution is uniform, then this also violates the monotonicity assumption required by \cite{Cao2013} for the optimality of FDR control. Example \ref{ex:score_glm} illustrates this phenomenon using the score test in a generalised linear model. \begin{example} \label{ex:score_glm} Assume the following regression model based on the linear exponential family where the density of the observations $y_1, \cdots, y_n$ are independent and follows \[ h(y_i| \beta, X_i) = \exp\lbrace a(X_i\beta) y_i+ b(X_i\beta) + D(y_i) \rbrace, \] where $X_i$ is a vector of covariates associated with each $y_i$, and $\beta = (\beta_0, \beta_1, \cdots, \beta_k)$ is a vector of regression coefficients. Let $f(X\beta) = E[y|X]$ denote the mean function. \cite{score_reg} studied the score statistic for testing the global null $\beta_1 = \beta_2= \cdots = \beta_k = 0$ and linked the resulting statistic to linear regression. A similar analysis can be performed for different hypothesis, such as inference for a parameter of interest in the presence of nuisance parameters to produce a more general result whose derivation is consigned to the Supplementary Materials. The resulting score statistic takes the form \[S_n = \lbrace y - f(X\hat\beta_{\text{null}}) \rbrace^\top W X \left\lbrace X^\top D X \right\rbrace^{-1} X^\top W \lbrace y - f(X\hat\beta_{\text{null}} ) \rbrace \xrightarrow[]{D} \chi^2_{q}, \] where $f(X\hat\beta_{\text{null}})$ is a vector whose $i$-th entry is $f(X_i\hat\beta_{\text{null}})$, $q$ is the number of constraints in the null hypothesis, and $\hat\beta_{\text{null}}$ denotes the constrained maximum likelihood estimate under the null. W and D are square diagonal matrices of dimension $n$ whose entries are $[W]_{ii} = a^\prime(X_i\hat\beta_{\text{null}})$ and $[D]_{ii} = a^\prime(X_i\hat\beta_{\text{null}}) f^\prime(X_i\hat\beta_{\text{null}}) $ for $i = 1, \dots, n$. Using a suitable change of variable, the statistic $S_n$ can be related to weighted linear regression. In the common case where we wish to test for $\beta_j = 0$, the score statistic can be re-written in the form: \[S_n = [ (X^\top D X)^{-1} ]_{jj}^{1/2} \sum_{i = 1}^n a^\prime(X_i \hat\beta_{\text{null}}) x_{ij} \lbrace y_i - f(X_i \hat\beta_{\text{null}}) \rbrace \xrightarrow{D} N(0,1), \] under the null. Under the alternative we may write \[S_n = (1 + c_n )\tilde{S}_n + d_n, \] where $\tilde{S}_n$ converges in distribution to a standard normal. The scaling factor $c_n = O(1)$ is an information bias and $d_n$ plays the role of the effect size and will increase to infinity as the number of samples increases. However, if $\beta_j \approx 0 $ then $d_n$ can be quite small, meaning that the effect of the scaling factor $c_n$ can be consequential. For an example of this see Example 3, where the effect size is not large enough to offset the scaling factor. \end{example} \begin{remark} Although the likelihood ratio statistic can be written as a summation of independent random variables, the limiting distribution of the likelihood ratio test is a gamma random variable, therefore Theorem 1 or 2 are not directly applicable. It may be possible to modify the baseline density used in the Edgeworth expansions to obtain a result based on Laguerre polynomials. This can also be useful when examining the asymptotic behaviour of test statistics for testing vector parameters of interest, as these test statistics often have a gamma distributed limiting distribution. \end{remark} \begin{remark} The bias issue discussed within this section is also present for the Wald test statistics, even if it can not be represented by a summation of independent random variables. It is rarely the case that the maximum likelihood estimate is unbiased, and the same applies for the estimate of the variance of the maximum likelihood estimate. Generally the problem worsens as the number of nuisance parameters increases. \end{remark} \subsection{Numerical Examples of Application of the Main Theorems} We illustrate the results of the main results with some numerical examples to demonstrate how various problems in the distributions of the $p$-value can occur. We first examine a discrete case where the statistic $S_n$ does not admit a density. We note that when the $E_2$ term is negligible, our results on the distribution of the $p$-values coincide with those obtained by \cite{hung} when the exact normality of the test statistic holds. On the contrary, when the additional terms are not negligible or the variance is incorrectly specified, the behaviour of the distribution of $p$-values can be quite different. The exact size of the difference depends on the behaviour of the Hermite polynomials, the higher order cumulants and the variance. We consider the following examples in order to illustrate some of the ramifications. \begin{example}\label{ex:linkage} Consider a simple linkage analysis of sibling pairs who share the same trait of interest, a common problem in statistical genetics. The underlying principle is that genes that are responsible for the trait are expected to be over-shared between relatives, while the null hypothesis states that the trait similarity does not impact allele sharing, i.e. independence between the trait and gene. The problematic distribution of $p$-values in this example is caused by the discrete nature of the problem along with a misspecified variance under the alternative. Since the offsprings are from the same parents, under the null we would expect the number of shared alleles to be either 0, 1 or 2 with probability $\theta_{null} = (p_0, p_1, p_2) = (0.25,0.5,0.25)$ based on Mendel's first law of segregation. However, under the alternative we can expect the sharing level to be higher than expected. Assume that we have $n$ affected sibling pairs. Let $x_i$ be the number alleles shared amongst the $i$-th affected sibling pair. Then under the null $E[x_i] = 1$, $Var[x_i] = 0.5$ and we let $y_i = (x_i - 1)/\sqrt{0.5}$. We consider the following well known non-parametric linkage test \begin{align*} S_n = \sum_{i = 1}^n \frac{y_i}{\sqrt{n}} \sim N(0,1), \end{align*} see \cite{laird2010fundamentals}. The above can be compared to a score test as only the information under the null was used. Under the alternative, the distribution of the test can be misspecified, since a different distribution of allele sharing will yield a different variance. Consider the simple example when the distribution of the numbers of shared alleles follows a multinomial distribution with $\theta_{alt1} = (0.09,0.8,0.11)$. The variance of this distribution is $0.4 < 0.5$. Yet another alternative in which $\theta_{alt2} = (0.29,0.4,0.31)$ yields the variance $0.6 > 0.5$; in both case there is oversharing. We include visualizations of the $p$-value distribution under the two alternatives in Figure \ref{fig:linkage}. Theorem 2 is used to produce the approximation given by the blue curve, due to the discrete nature of the test statistic. \begin{figure} \centering \subfloat[Score test, $\theta = \theta_{alt1}$]{\includegraphics[width = 2in]{linkage_1.png}} \subfloat[Score test, $\theta = \theta_{alt2}$ ]{\includegraphics[width = 2in]{linkage_2.png}} \\ \subfloat[Wald test, $\theta = \theta_{alt1}$]{\includegraphics[width = 2in]{linkage_wald_2.png}} \subfloat[Wald test, $\theta = \theta_{alt2}$ ]{\includegraphics[width = 2in]{linkage_wald_1.png}} \caption{Plots for Example 3, examining the behaviour of the $p$-value distribution for non-parametric linkage analysis for the score test (upper panel) and the Wald test (lower panel). The simulation is performed with $n =400$, and $100,000$ replications. Samples of sibling pairs are generated from a multinomial distribution with $\theta_{alt1} = (0.09,0.8, 0.11) $ for the two plots on the left panel, and $\theta_{alt2} = (0.29,0.4,0.31)$ for the two plots on the right panel. For the score test, both histograms have spikes due to the discrete nature of the problem. The discrete version of the Edgeworth approximation, plotted in blue, is used as the test statistic is supported on a lattice. The histograms of the $p$-values obtained form the Wald test look much better than their score test counterparts. } \label{fig:linkage} \end{figure} \end{example} In this case the problem can be resolved by considering a Wald type test where the variance is calculated from the maximum likelihood estimate $\hat{\theta} = ( \#(x_i = 0)/n, \#(x_i = 1)/n, \#(x_i = 2)/n )$, and use: \[S_n^\prime = \sum_{i = 1}^n \frac{x_i - 1}{\sqrt{n\widehat{\text{var}}(x_i)}}. \] We plot the results of applying the Wald test in Figure \ref{fig:linkage}. The solution is quite simple in this case, but in more complex models it is more computationally expensive to calculate the variance estimate under the alternative. \\ \noindent\textbf{Example 1 revisited.} The abnormal distribution of $p$-values in this scenario is caused by a large numerical value of $\rho_{n,3}$ and $\rho_{n,4}$. Going back to Example 1, we look at the theoretically predicted behaviour of the $p$-values under the null and alternative. Figure \ref{fig:test} shows the histograms of the empirical $p$-values obtained by simulation versus the theoretical prediction given in Theorem 1, shown as the blue curve. Without accounting for the higher order terms in the expansion we would have expected the null distribution to be uniform, however, using Theorem 1, we obtain a much more accurate description of the $p$-value distribution. In the bottom panel of Figure \ref{fig:test} we also show a corrected version of the $p$-values approximation using the the saddlepoint approximation which will be introduced in Section 3. The estimation of small p-values based on the standard normal approximation can be drastically optimistic. We report in Table \ref{tb:deltas} the differences between the exact and the approximate $p$-value obtained from Example 1 for the 5 smallest $p$-values. The smallest $p$-values from the normal approximation are not on the same scale as the exact $p$-values, the smallest approximate $p$-value being five-fold times smaller than its exact counterpart. In contrast, the $p$-values produced by the saddlepoint approximation are very close to the exact ones. \begin{table}[b] \caption{ \label{tb:deltas} Table of $p$-values obtained from Example 1 under the null. The exact $p$-values are obtained from the density of the gamma distribution, the approximate $p$-values are obtained from the normal approximation.} \centering \begin{tabular}{rrrrr} \hline ID & rank & $p$-value exact & $p$-value approx. & $p$-val saddlepoint \\ \hline 60326 & 1 & 1.04E-05 & 1.04E-10 & 1.04E-05 \\ 91132 & 2 & 1.46E-05 & 3.06E-10 & 1.47E-05 \\ 83407 & 3 & 2.12E-05 & 9.66E-10 & 2.12E-05 \\ 97470 & 4 & 3.31E-05 & 3.75E-09 & 3.32E-05 \\ 2573 & 5 & 3.80E-05 & 5.66E-09 & 3.81E-05 \\ \hline \end{tabular} \end{table} \begin{example} We examine the influence of the effect size $\mu_n$ on the distribution of the $p$-values under the alternative using the same set-up as in Example 1. In our simulations we increase the effect size $\mu_n$ by changing the value of $\beta$, while keeping $\alpha$ fixed. The results are displayed in Figure \ref{fig:altnernaives}. \begin{figure}[H] \centering \subfloat[Subfigure 1 list of figures text][Small effect size ]{\includegraphics[width = 1.75in]{alt_1025.png}} \subfloat[Subfigure 2 list of figures text][Medium effect size ]{\includegraphics[width = 1.75in]{alt_105.png}} \subfloat[Subfigure 1 list of figures text][Large effect size ]{\includegraphics[width = 1.75in]{alt_1075.png}} \\ \subfloat[Subfigure 1 list of figures text][Small effect size ]{\includegraphics[width = 1.75in]{alt_1025_s.png}} \subfloat[Subfigure 2 list of figures text][Medium effect size ]{\includegraphics[width = 1.75in]{alt_105_s.png}} \subfloat[Subfigure 1 list of figures text][Large effect size ]{\includegraphics[width = 1.75in]{alt_1075_s.png}} \caption{Distribution of the approximated $p$-values (top panel) and the corrected $p$-values (bottom panel), under three different alternatives with $\alpha_1 =\alpha_2 = \alpha_3 =0.01$ and $\beta_2 = 0.01/1.025$, $\beta_1 = 0.01/1.05$ and $\beta_3 = 0.01/1.1$, from left to right. By increasing the effect size, the approximate $p$-values starts to behave in an expected manner. While the corrected $p$-values obtained by using the saddlepoint approximation is well behaved for all effect sizes.} \label{fig:altnernaives} \end{figure} As discussed in Remark 3, for large effect sizes $\mu_n$, the distribution of $p$-values generated from the test statistic follows the expected trend, where there is a concentration of $p$-values around $0$ and the density decreases in a monotone fashion to 1. Conversely, should $\mu_n$ be small then the behaviour under the alternative can be quite different from what we would expect, as illustrated by the top-right plot in Figure \ref{fig:altnernaives}. \end{example} \section{Additional Examples and Possible Remedies} We provide additional examples of problematic $p$-value distributions, and we explore some possible remedies based on high order asymptotics. We also provide additional examples of problematic $p$-value distributions. A commonly used tool for higher order asymptotics is the saddlepoint approximation, which is a density approximation that can be integrated to obtain tail probabilities, e.g. $p$-values. For a good survey of the saddlepoint approximation and its applications in statistics, we refer the reader to \cite{reid1988} or for a more technical reference, we suggest \cite{jensen1995saddlepoint} or \cite{kolassa1994series}. The saddlepoint approximation can be most easily obtained for a sum or average of independent random variables, $X_1, \dots, X_n$. The density approximation then results in an approximation of the cumulative distribution through a tail integration argument, \begin{align} P(\bar{X} < s) = \Phi(r_s)\lbrace 1 + O(n^{-1}) \rbrace, \label{accuracy_saddle} \end{align} where $r_s$ is a quantity constructed from the saddlepoint and the cumulants of the distribution of the $X_i$'s. This can be used for conditional inference in generalized linear models by approximating the distribution of the sufficient statistics in a exponential family model; see \cite{davison}. Another more broadly applicable tail approximation is the normal approximation to the $r^\star$ statistic \citep{barndorff1989asymptotic}, which is obtained by adding a correction factor to $r$, the likelihood root. It can be used in regression settings for inference on a scalar parameter of interest. Let $r =\text{sign}(\hat\psi) [2 \lbrace l_{\text{pro}}(\hat\psi) - l_{\text{pro}}(\psi_0) \rbrace]^{1/2}$ denote the likelihood root, and in what follows the quantity $Q$ varies depending on the model. \begin{align*} P(r < s) = \Phi\left\lbrace r + \frac{1}{r} \log\left(\frac{Q}{r}\right)\right\rbrace \left\lbrace 1 +O\left( n^{-3/2} \right) \right\rbrace. \end{align*} Using the above, we also obtain an improved approximation to the true distribution of the likelihood root. For a discussion of $r^\star$ see \cite{reid_wald}. The proposed methods require two model fits, one under the alternative and one under the null in order to obtain $r$, contrary to the score test. The methods listed here are by no means comprehensive since there are a variety of other candidates which may be of use, such as the often applied Firth correction \citep{firth} or other forms of bias correction obtainable by adjusting the score equation \citep{kosmidis2020mean}. \subsection{The Gamma example} We apply the saddlepoint approximation to Example 1 and display the results in Figure \ref{fig:test}. Considering the null $H_0: \alpha = \beta = 0.01$ (the two plots on the left panel), there is a spike around 0 for $p$-values obtained using the CLT (top left plot). In contrast, we see a marked improvement of the overall behaviour of the $p$-value distribution after the proposed correction (bottom left plot). \subsection{Logistic Regression in Genetic Association Studies} We apply the normal approximation to $r^\star$ to a simulated genome-wide association study to further illustrate the practical use of the proposed correction. We consider a logistic regression model that links the probability of an individual suffering from a disease to that individual's single nucleotide polymorphism (SNP), a genetic ordinal variable coded as 0, 1 or 2, and other covariates such as age and sex. Formally, let the disease status of the individual be $Y_i$, which is either $0$ (individual is healthy) or $1$ (individual is sick) and $\pi_i = E[Y_i]$ denote the probability of individual $i$ having the disease and let $X_{i, s}$ denote the genetic covariate of interest of the $i$-th individual, while $X_{i,j}, j = 1, 2$, $j\ne s$ are the other covariates. The regression model is: \begin{align*} \text{logit}(\pi_i) = X^i_s \beta_s + \sum_{j = 1}^2 X^i_j \beta_{j} + \beta_0. \end{align*} We consider the difficult case where the disease is uncommon in the population and the SNPs of interest are rare, i.e. most observed values of $X_{i,s}$ are 0. It is known that in this situation the single-SNP test performs poorly, and pooled analyses of multiple SNPs have been proposed \citep{pooled}. However for the purpose of this study, we assume that the individual SNPs are of interest. We consider a simulated example to demonstrate the effectiveness of the correction. We generate a sample of 3,000 individuals, their genetic variable $X_s$ are simulated from a $Binomial(2, 0.025)$, a binary variable $X_1$ from a $Binomial(1,0.5)$ and finally $X_2$ from a $N(20, 1)$. We let $\beta_0 = -3.5$, $\beta_s = 0$, $\beta_1 = 0.02$ and $\beta_2 = 0.02$. With this set of parameters we would expect on average $ \approx 4.6\%$ of the cohort to be in the diseased group, based on the expected value of the covariates, i.e. approximately 137 participants with $Y_i = 1$. For each replication of the simulation, we re-generate the labels from the logistic model. Figure \ref{fig:6} shows that the correction works well under the null. \begin{figure}[ht] \centering \subfloat{\includegraphics[width=2.75in]{Wald_null}} \subfloat{\includegraphics[width=2.75in]{cond_inf_null}} \caption{Empirical distribution of the null $p$-values from a logistic regression association study of SNPs with with low minor allele frequency and a low number of diseased individuals. Left histogram displays the $p$-values from the Wald test under the null. The right histogram displays the $p$-values histogram obtained from $r^\star$ under the null. } \label{fig:6} \end{figure} This example suggests that the usefulness of the proposed higher order corrections is not limited to small sample scenarios, as note by \cite{zhou2018efficiently} who used the saddlepoint approximation in case control studies with extreme sample imbalance. Naively we would expect that with 3,000 participants, of which 137 are in the diseased group, the Wald test should behave correctly. However, the skewed distribution of the SNP values severely reduces the accuracy of the test. The use of $r^\star$ corrects the distribution of the $p$-values as shown in Figure \ref{fig:6} (right plot) where the distribution of the $p$-values under the null ($\beta_s=0$) is approximately Unif(0, 1) as expected. In the example above it is clear that even though we have 3,000 individuals, of which 137 are affected by the disease, the standard approximation performs very poorly. This seems to suggest that in our particular example, the effective sample size is lower than 137 for the diseased group. Next we consider a simple regression with a single genetic covariate in order to illustrate the loss in information resulting from the sparsity of the minor allele. We use the available Fisher information about the parameter of interest as a measure of effective sample size. The standard deviation of the parameter of interest obtained from the inverse information matrix is \begin{align*} var(\hat\beta_s) &= \frac{\sum_{i = 1}^{n} \hat{P_i} (1 - \hat{P_i})}{\sum_{i = 1}^{n} \hat{P_i} (1 - \hat{P_i})\sum_{i = 1}^{n} x_i^2 \hat{P_i} (1 - \hat{P_i}) - (\sum_{i = 1}^{n} x_i \hat{P_i} (1 - \hat{P_i}) )^2},\\ &\approx \frac{1}{\sum_{i = 1}^{n} x_{i,s} \hat{P_i} (1 - \hat{P_i})}, \end{align*} where $\hat{P}_i$ is the predicted probability of an individual being diseased and the approximation is valid under the assumption that the allele frequency is low enough such that we observe very few 1's and almost no 2's. The information about the parameter $\beta_s$ is increasing in terms of $x_i \hat{P_i}(1 - \hat{P_i})$. It is apparent that the rate of increase in information is limited by the sparsity of the rare allele. In order to have more information about the parameter, we would need to observe more individuals who have the rare allele, i.e. $X_i \ne 0$. \subsection{Logistic Regression - Data from the 1000 Genome Project} We consider an additional logistic regression example as this type of model is broadly used in statistical genetics. Using phase 3 data from the 1000 Genome project \citep{1000genome}, we construct an artificial observational study in order to study how these approximations behave on real genome-wide genetic data. In our simulations, we take the 2504 individuals within the database and assign the $i$-th individual a label of $0$ or $1$ based on the following logistic model, where $\pi_i = P(Y_i = 1)$: \[ \text{logit}(\pi_i) = \sum_{j = 1}^4 X^i_j \beta_{j} + \beta_\text{Sex}*I( \text{Sex}_i = \text{male}) + \beta_0, \] where $\text{Sex}_i$ is the biological sex of the $i$-th individual. Four other covariates are included, where $X^i_j$ are independent for all $i, j$ and follow a standard normal distribution. The model coefficients are set to \[(\beta_0, \beta_1, \beta_2, \beta_3, \beta_4, \beta_{\text{Sex}}) = (-3.25, 0.025, -0.025, 0.025, -0.03, 0.1).\] Once we assign a label to the $i$-th individual we keep it fixed throughout the simulation. We then fit a logistic model using the SNPs for which the minor allele frequency is at least $1\%$ on chromosome $10$, and ethnicity as additional covariates. We use the Wald test, and $r^\star$, but do not consider the cases where perfect separation occurs, as both methods considered here cannot deal with this issue. We plot some of the results for the Wald test and $r^\star$. We focus on rare variants with MAF $\leq 2.5\%$ and semi-common variants with $2.5\% <$ MAF $\leq 10\%$, as the remaining common variants are expected to behave well. In total $160,580$ SNPs fall into the rare variant category while $176,350$ SNPs fall into the semi-common variant category. \begin{figure}[h] \centering \subfloat[Subfigure 1 list of figures text][Rare variants, Wald]{\includegraphics[width =2in]{wald_low_MAF_10.png}} \subfloat[Subfigure 1 list of figures text][Rare variants, $r^\star$]{\includegraphics[width = 2in]{rstar_low_MAF_10.png}}\\ \subfloat[Subfigure 1 list of figures text][Semi-common variants, Wald ]{\includegraphics[width = 2in]{wald_mid_MAF_10.png}} \subfloat[Subfigure 1 list of figures text][Semi-common variants, $r^\star$ ]{\includegraphics[width = 2in]{rstar_mid_MAF_10.png}} \caption{Distribution of $p$-values for Wald test and $r^\star$. The null distribution was simulated by using sex and four other randomly generated covariates. We fit a logistic regression model using SNPs from chromosome 10, with $160,580$ being rare ($\text{MAF}\leq 2.5\%$ ) and $176,350$ semi-common variants ($2.5\% < \text{MAF}\leq 10\%$ ).} \label{altnernaives} \end{figure} As expected, the two tests behave better for semi-common SNPs than rare SNPs (bottom vs. top panel of Figure 7), producing $p$-values that more closely follow the Unif(0,1) distribution. Among the two tests, the proposed $r^\star$ method clearly out-performs the traditional Wald test. However, this application also points out the limitation of $r^\star$ as the correction for rare variants is not sufficient (top right plot), and further improvement of the method in this case is of future interest. \subsection{Weibull survival regression} Consider an example where there is a large number of nuisance parameters, leading to an inconsistent estimate of the variance. We examine a Weibull survival regression model in which all of the regression coefficients, except the intercept, are set to $0$ by simulating $y_i \sim \text{Weibull}(1,2)$, independently of any covariate. We set the number of observations, $n$ to 200 and the number of covariates to $50$, and generated the covariates as IID standard Gaussian, and test for whether the first (non-intercept) regression coefficient is 0. We perform 10,000 replications and plot the histogram of the $p$-values, and compare the Wald test to the $r^\star$ correction. \begin{figure}[H] \centering \subfloat{\includegraphics[width=0.35\linewidth]{weibull.png}} \subfloat{ \includegraphics[width=0.35\linewidth]{weibull_r.png}} \caption{On the left, a histogram of the $p$-values produced by the Wald test for $\beta_1 = 0$ under the null with $n = 200$ and $p = 50$ with no censoring. On the right, a histogram of the $p$-value obtained from the $r^\star$ correction. 10,000 replications were performed.} \label{regression} \end{figure} In Figure \ref{regression} we see a high concentration of $p$-values around $0$ for the Wald test, leading to increased I error. The corrective procedure brings the distribution under the null much closer to uniformity. We see that naively adding more and more information into the model while trying to perform inference on a one dimensional parameter of interest is problematic as it creates a perceived significance of the parameter of interest under the null. \section{Discussion and Conclusion} We characterize the distribution of $p$-values when the test statistic is not well approximated by a normal distribution by using additional information contained in the higher order cumulants of the distribution of the test statistic. We also demonstrate that there are issues beyond failure to converge to normality in the that the expectation and variance of the test statistics can be misspecified, and these issues can persist even in large sample settings. In doing so we have extended the previous work done by \cite{hung} to greater generality, examining the score test in exponential models in the presence of nuisance parameters. We also examine some possible remedies for making the $p$-value distribution adhere more closely to their usual required behaviour such as uniformity under the null or concavity of the CDF under the alternative. These assumptions are very important to justify the usage of current FWER and FDR procedures. The proposed remedies may not solve all problems relating to the $p$-value distribution in the finite sample settings, but they do at least partially correct some of the flaws. We suggest the use of the proposed saddlepoint approximation or the normal approximation to $r^\star$ in practice, because a) the exact distribution of a test statistic is often unknown, b) the usual CLT approximation may not be adequate, and c) the high order methods are easy to implement. This will ensure a closer adherence to the assumptions usually needed to conduct corrective procedures used in FWER control or FDR control. \section*{Acknowledgement} The first author would like to thank Nancy Reid, Michele Lambardi di San Miniato and Arvind Shrivats for the help and support they provided. We also thank the Natural Sciences and Engineering Research Council, the Vector Institute and the Ontario government for their funding and support. \newpage

https://arxiv.org/abs/1505.05116

Data-driven Distributionally Robust Optimization Using the Wasserstein Metric: Performance Guarantees and Tractable Reformulations

We consider stochastic programs where the distribution of the uncertain parameters is only observable through a finite training dataset. Using the Wasserstein metric, we construct a ball in the space of (multivariate and non-discrete) probability distributions centered at the uniform distribution on the training samples, and we seek decisions that perform best in view of the worst-case distribution within this Wasserstein ball. The state-of-the-art methods for solving the resulting distributionally robust optimization problems rely on global optimization techniques, which quickly become computationally excruciating. In this paper we demonstrate that, under mild assumptions, the distributionally robust optimization problems over Wasserstein balls can in fact be reformulated as finite convex programs---in many interesting cases even as tractable linear programs. Leveraging recent measure concentration results, we also show that their solutions enjoy powerful finite-sample performance guarantees. Our theoretical results are exemplified in mean-risk portfolio optimization as well as uncertainty quantification.

\section{Introduction} \label{sec:introduction} Stochastic programming is a powerful modeling paradigm for optimization under uncertainty. The goal of a generic single-stage stochastic program is to find a decision $x\in \mathbb{R}^n$ that minimizes an expected cost~$\mathds{E}^\mathds{P}[h(x,\xi)]$, where the expectation is taken with respect to the distribution $\mathds{P}$ of the \change{continuous random vector} $\xi\in\mathbb{R}^m$. However, classical stochastic programming is challenged by the large-scale decision problems encountered in today's increasingly interconnected world. First, the distribution $\mathds{P}$ is never observable but must be inferred from data. However, if we calibrate a stochastic program to a given dataset and evaluate its optimal decision on a different dataset, then the resulting out-of-sample performance is often disappointing---even if the two datasets are generated from the same distribution. This phenomenon is termed the {\em optimizer's curse} and is reminiscent of overfitting effects in statistics \cite{ref:SmiRob-06}. Second, in order to evaluate the objective function of a stochastic program for a fixed decision $x$, we need to compute a multivariate integral, which is \#P-hard even if $h(x,\xi)$ constitutes the positive part of an affine function, while $\xi$ is uniformly distributed on the unit hypercube \cite[Corollary~1]{ref:GraKunWie-15}. Distributionally robust optimization is an alternative modeling paradigm, where the objective is to find a decision $x$ that minimizes the {\em worst-case} expected cost $\sup_{\mathds{Q} \in \mathcal P} \mathds{E}^\mathds{Q} [ h(x,\xi)]$. Here, the worst-case is taken over an ambiguity set $\mathcal P$, that is, a family of distributions characterized through certain known properties of the unknown data-generating distribution $\mathds{P}$. Distributionally robust optimization problems have been studied since Scarf's seminal treatise on the ambiguity-averse newsvendor problem in 1958 \cite{ref:Scarf-58}, but the field has gained thrust only with the advent of modern robust optimization techniques in the last decade \cite{ref:BenElNem-09,ref:BertSim-04}. Distributionally robust optimization has the following striking benefits. First, adopting a worst-case approach regularizes the optimization problem and thereby mitigates the optimizer's curse characteristic for stochastic programming. Second, distributionally robust models are often tractable even though the corresponding stochastic models \change{with fixed continuous distributions} are \#P-hard. Thus, surprisingly, distributionally robust models may offer better decisions than their stochastic counterparts and are also easier to solve; \change{see {\em e.g.}~\cite{ref:Del-09}}. The ambiguity set $\mathcal P$ is a key ingredient of any distributionally robust optimization model. A good ambiguity set should be rich enough to contain the true data-generating distribution with high confidence. On the other hand, the ambiguity set should be small enough to exclude pathological distributions, which would incentivize overly conservative decisions. The ambiguity set should also be easy to parameterize from data, and---ideally---it should facilitate a tractable reformulation of the distributionally robust optimization problem as a structured mathematical program that can be solved with off-the-shelf optimization software. \change{Distributionally robust optimization models where $\xi$ has finitely many realizations are reviewed in~\cite{ref:BenHer-13, ref:BertSAA-14, ref:PosHertMel-14}. This paper focuses exclusively on situations where $\xi$ can have a continuum of realizations. In this setting, } the existing literature on distributionally robust optimization has primarily studied moment ambiguity sets $\mathcal P$, which contain all distributions that satisfy certain moment constraints, see for example \cite{ref:DelYe-10,ref:GohSim-10,ref:WieKuhSim-14} and the references therein. An attractive alternative is to define $\mathcal P$ as a ball in the space of probability distributions by using a probability distance function such as the Prohorov metric~\cite{ref:ErdoIyen-06}, the Kullback-Leibler divergence \cite{ref:JiaGua-14, ref:HuHong-13}, or the Wasserstein metric~\cite{ref:PflWoz-07,ref:Woz-12} etc. Such metric-based ambiguity sets contain all distributions that are close to a {\em nominal} or {\em most likely} distribution with respect to the prescribed probability metric. By adjusting the radius of the ambiguity set, the modeler can thus control the degree of conservatism of the underlying optimization problem. If the radius drops to zero, then the ambiguity set shrinks to a singleton that contains only the nominal distribution, in which case the distributionally robust problem reduces to an ambiguity-free stochastic program. In this paper we study distributionally robust optimization problems with a {\em Wasserstein ambiguity set} centered at the uniform distribution $\wh{\PP}_N$ on $N$ independent and identically distributed training samples. The Wasserstein distance of two distributions $\mathds{Q}_1$ and $\mathds{Q}_2$ can be viewed as the minimum transportation cost for moving the probability mass from $\mathds{Q}_1$ to $\mathds{Q}_2$, and the Wasserstein ambiguity set contains all (continuous or discrete) distributions that are sufficiently close to the (discrete) empirical distribution $\wh{\PP}_N$ with respect to the Wasserstein metric. Modern measure concentration results from statistics guarantee that the unknown data-generating distribution $\mathds{P}$ belongs to the Wasserstein ambiguity set around $\wh{\PP}_N$ with confidence $1-\beta$ if its radius is a sublinearly growing function of $\log(1/\beta)/N$ \cite{ref:Boll-07, ref:FouGui-14}. The optimal value of the distributionally robust problem thus provides an upper confidence bound on the achievable out-of-sample cost. While Wasserstein ambiguity sets offer powerful out-of-sample performance guarantees and enable the decision maker to control the model's conservativeness, moment-based ambiguity sets appear to display better tractability properties. Specifically, there is growing evidence that distributionally robust models with moment ambiguity sets are more tractable than the corresponding stochastic models \change{because the intractable high-dimensional integrals in the objective function are replaced with tractable (generalized) moment problems} \cite{ref:DelYe-10,ref:GohSim-10,ref:WieKuhSim-14}. In contrast, distributionally robust models with Wasserstein ambiguity sets are believed to be harder than their stochastic counterparts \cite{ref:PflPich-14}. Indeed, the state-of-the-art method for computing the worst-case expectation over a Wasserstein ambiguity set~$\mathcal P$ relies on global optimization techniques. Exploiting the fact that the extreme points of $\mathcal P$ are discrete distributions with a fixed number of atoms~\cite{ref:Woz-12}, one may reformulate the original worst-case expectation problem as a finite-dimensional non-convex program, which can be solved via ``difference of convex programming" methods, see \cite{ref:Woz-12} or \cite[Section~7.1]{ref:PflPich-14}. However, the computational effort is reported to be considerable, and there is no guarantee to find the global optimum. \change{Nevertheless, tractability results are available for special cases. Specifically, the worst case of a convex law-invariant risk measure with respect to a Wasserstein ambiguity set $\mathcal P$ reduces to the sum of the nominal risk and a regularization term whenever $h(x,\xi)$ is affine in $\xi$ and $\mathcal P$ does not include any support constraints \cite{ref:Woz-14}. Moreover, while this paper was under review we became aware of the PhD thesis~\cite{ref:Zhao-14}, which reformulates a distributionally robust two-stage unit commitment problem over a Wasserstein ambiguity set as a semi-infinite linear program, which is subsequently solved using a Benders decomposition algorithm.} In this paper we demonstrate that the worst-case expectation over a Wasserstein ambiguity set can in fact be computed efficiently via convex optimization techniques for numerous loss functions of practical interest. Furthermore, we propose an efficient procedure for constructing an extremal distribution that attains the worst-case expectation---provided that such a distribution exists. Otherwise, we construct a sequence of distributions that attain the worst-case expectation asymptotically. As a by-product, our analysis shows that many interesting distributionally robust optimization problems with Wasserstein ambiguity sets can be solved in polynomial time. We also investigate the out-of-sample performance of the resulting optimal decisions---both theoretically and experimentally---and analyze its dependence on the number of training samples. We highlight the following main contributions of this paper. \begin{itemize} \item We prove that the worst-case expectation of an uncertain loss $\ell(\xi)$ over a Wasserstein ambiguity set coincides with the optimal value of a finite-dimensional convex program if $\ell(\xi)$ constitutes a pointwise maximum of finitely many concave functions. Generalizations to convex functions or to sums of maxima of concave functions are also discussed. We conclude that worst-case expectations can be computed efficiently to high precision via modern convex optimization algorithms. \item We describe a supplementary finite-dimensional convex program whose optimal (near-optimal) solutions can be used to construct exact (approximate) extremal distributions for the infinite-dimensional worst-case expectation problem. \item We show that the worst-case expectation reduces to the optimal value of an explicit linear program if the $1$-norm or the $\infty$-norm is used in the definition of the Wasserstein metric and if $\ell(\xi)$ belongs to any of the following function classes: (1) a pointwise maximum or minimum of affine functions; (2) the indicator function of a closed polytope or the indicator function of the complement of an open polytope; (3) the optimal value of a parametric linear program whose cost or right-hand side coefficients depend linearly on~$\xi$. \item Using recent measure concentration results from statistics, we demonstrate that the optimal value of a distributionally robust optimization problem over a Wasserstein ambiguity set provides an upper confidence bound on the out-of-sample cost of the worst-case optimal decision. We validate this theoretical performance guarantee in numerical tests. \end{itemize} If the uncertain parameter vector $\xi$ is confined to a fixed finite subset of $\mathbb{R}^m$, then the worst-case expectation problems over Wasserstein ambiguity sets simplify substantially and can often be reformulated as tractable conic programs by leveraging ideas from robust optimization. An elegant second-order conic reformulation has been discovered, for instance, in the context of distributionally robust regression analysis \cite{ref:MehZhan-14}, and a comprehensive list of tractable reformulations of distributionally robust risk constraints for various risk measures is provided in \cite{ref:PosHertMel-14}. Our paper extends these tractability results to the practically relevant case where $\xi$ has uncountably many possible realizations---without resorting to space tessellation or discretization techniques that are prone to the curse of dimensionality. When $\ell(\xi)$ is linear and the distribution of $\xi$ ranges over a Wasserstein ambiguity set without support constraints, one can derive a concise closed-form expression for the worst-case risk of $\ell(\xi)$ for various convex risk measures \cite{ref:Woz-14}. However, these analytical solutions come at the expense of a loss of generality. We believe that the results of this paper may pave the way towards an efficient computational procedure for evaluating the worst-case risk of $\ell(\xi)$ in more general settings where the loss function may be non-linear and $\xi$ may be subject to support constraints. Among all metric-based ambiguity sets studied to date, the Kullback-Leibler ambiguity set has attracted most attention from the robust optimization community. It has first been used in financial portfolio optimization to capture the distributional uncertainty of asset returns with a Gaussian nominal distribution \cite{ref:ElOksOus-03}. Subsequent work has focused on Kullback-Leibler ambiguity sets for discrete distributions with a fixed support, which offer additional modeling flexibility without sacrificing computational tractability \cite{ref:Cal-07,ref:BenHer-13}. It is also known that distributionally robust chance constraints involving a generic Kullback-Leibler ambiguity set are equivalent to the respective classical chance constraints under the nominal distribution but with a rescaled violation probability \cite{ref:JiaGua-14,ref:HuHongSo-13}. Moreover, closed-form counterparts of distributionally robust expectation constraints with Kullback-Leibler ambiguity sets have been derived in \cite{ref:HuHong-13}. \change{However, Kullback-Leibler ambiguity sets typically fail to represent confidence sets for the unknown distribution $\mathds{P}$. To see this, assume that $\mathds{P}$ is absolutely continuous with respect to the Lebesgue measure and that the ambiguity set is centered at the discrete empirical distribution $\wh{\PP}_N$. Then, any distribution in a Kullback-Leibler ambiguity set around $\wh{\PP}_N$ must assign positive probability mass to each training sample. As $\mathds{P}$ has a density function, it must therefore reside outside of the Kullback-Leibler ambiguity set irrespective of the training samples. Thus, Kullback-Leibler ambiguity sets around $\wh{\PP}_N$ contain $\mathds{P}$ with probability~0. In contrast, Wasserstein ambiguity sets centered at $\wh{\PP}_N$ contain discrete as well as continuous distributions and, if properly calibrated, represent meaningful confidence sets for~$\mathds{P}$. We will exploit this property in Section~\ref{sec:wass} to derive finite-sample guarantees. A comparison and critical assessment of various metric-based ambiguity sets is provided in~\cite{ref:Shap-15}. Specifically, it is shown that worst-case expectations over Kullback-Leibler and other divergence-based ambiguity sets are law invariant. In contrast, worst-case expectations over Wasserstein ambiguity sets are not. The law invariance can be exploited to evaluate worst-case expectations via the sample average approximation.} The models proposed in this paper fall within the scope of data-driven distributionally robust optimization \cite{ref:ErdoIyen-06,ref:ChehWeb-10, ref:BertSAA-14,ref:HanKuh-13}. Closest in spirit to our work is the robust sample average approximation \cite{ref:BertSAA-14}, which seeks decisions that are robust with respect to the ambiguity set of all distributions that pass a prescribed statistical hypothesis test. Indeed, the distributions within the Wasserstein ambiguity set could be viewed as those that pass a multivariate goodness-of-fit test in light of the available training samples. \change{This amounts to interpreting the Wasserstein distance between the empirical distribution $\wh{\PP}_N$ and a given hypothesis $\mathds{Q}$ as a test statistic and the radius of the Wasserstein ambiguity set as a threshold that needs to be chosen in view of the test's desired significance level $\beta$. The Wasserstein distance has already been used in tests for normality~\cite{ref:Barr-99} and to devise nonparametric homogeneity tests~\cite{ref:Ram15}.} The rest of the paper proceeds as follows. Section~\ref{sec:prob} sketches a generic framework for data-driven distributionally robust optimization, while Section~\ref{sec:wass} introduces our specific approach based on Wasserstein ambiguity sets and establishes its out-of-sample performance guarantees. In Section~\ref{sec:dist} we demonstrate that many worst-case expectation problems over Wasserstein ambiguity sets can be reduced to finite-dimensional convex programs, and we develop a systematic procedure for constructing worst-case distributions. Explicit linear programming reformulations of distributionally robust single and two-stage stochastic programs as well as uncertainty quantification problems are derived in Section~\ref{sec:cases}. Section~\ref{sec:exten} extends the scope of the basic approach to broader classes of objective functions, and Section~\ref{sec:num} reports on numerical results. \paragraph{\bf Notation} We denote by $\mathbb{R}_+$ the non-negative and by $\overline {\R} \coloneqq \mathbb{R} \cup \{-\infty,\infty \}$ the extended reals. Throughout this paper, we adopt the conventions of extended arithmetics, whereby $\infty\cdot0 = 0\cdot\infty = {0 / 0 } = 0$ and $\infty - \infty = -\infty + \infty = 1/0 = \infty$. The inner product of two vectors $a,b\in\mathbb{R}^m$ is denoted by $\inner{a}{b} \coloneqq a^\intercal b$. Given a norm $\|\cdot\|$ on $\mathbb{R}^m$, the dual norm is defined through $\|z\|_* \coloneqq \sup_{\|\xi\|\le 1} \inner{z}{\xi}$. A function $f:\mathbb{R}^m\rightarrow \overline {\R}$ is proper if $f(\xi)<+\infty$ for at least one $\xi$ and $f(\xi)>-\infty$ for every $\xi$ in $\mathbb{R}^m$. The conjugate of $f$ is defined as $f^*(z) \coloneqq \sup_{\xi \in \mathbb{R}^m} \inner{z}{\xi} - f(\xi)$. Note that conjugacy preserves properness. For a set $\Xi\subseteq \mathbb{R}^m$, the indicator function $\ind{\Xi}$ is defined through $\ind{\Xi}(\xi)=1$ if $\xi\in \Xi$; $=0$ otherwise. Similarly, the characteristic function $\chi_\Xi$ is defined via $\chi_\Xi(\xi)=0$ if $\xi\in \Xi$; $=\infty$ otherwise. The support function of $\Xi$ is defined as $\sigma_{\Xi}(z) \coloneqq \sup_{\xi \in \Xi} \inner{z}{\xi}$. It coincides with the conjugate of $\chi_\Xi$. We denote by $\dir{\xi}$ the Dirac distribution concentrating unit mass at $\xi\in\mathbb{R}^m$. The product of two probability distributions $\mathds{P}_1$ and $\mathds{P}_2$ on $\Xi_1$ and $\Xi_2$, respectively, is the distribution $\mathds{P}_1\otimes\mathds{P}_2 $ on $\Xi_1\times \Xi_2$. The $N$-fold product of a distribution $\mathds{P}$ on $\Xi$ is denoted by $\mathds{P}^N$, which represents a distribution on the Cartesian product space $\Xi^N$. \change{Finally, we set the expectation of $\ell:\Xi\rightarrow\overline {\R}$ under $\mathds{P}$ to $\mathds{E}^\mathds{P}[\ell(\xi)] = \mathds{E}^\mathds{P}\big[\max\{\ell(\xi),0\}\big] + \mathds{E}^\mathds{P}\big[\min\{\ell(\xi),0\}\big]$, which is well-defined by the conventions of extended arithmetics. } \section{Data-Driven Stochastic Programming} \label{sec:prob} Consider the stochastic program \begin{align} \label{Ex-true} J^\star \coloneqq \inf_{x \in \mathbb{X}} \left\{ \mathds{E}^\mathds{P} \big[ h(x,\xi) \big] = \int_{\Xi} h(x,\xi)\, \mathds{P}(\mathrm{d} \xi)\right\} \end{align} with feasible set $\mathbb{X} \subseteq \mathbb{R}^n$, uncertainty set $\Xi\subseteq \mathbb{R}^m$ and loss function $h : \mathbb{R}^n \times \mathbb{R}^m \rightarrow \overline {\R}$. The loss function depends both on the decision vector $x\in\mathbb{R}^n$ and the random vector $\xi\in\mathbb{R}^m$, whose distribution $\mathds{P}$ is supported on $\Xi$. Problem~\eqref{Ex-true} can be viewed as the first-stage problem of a two-stage stochastic program, where $h(x,\xi)$ represents the optimal value of a subordinate second-stage problem \cite{ref:Shap&Dent&Rusz}. Alternatively, problem~\eqref{Ex-true} may also be interpreted as a generic learning problem in the spirit of \cite{ref:Vapnik}. Unfortunately, in most situations of practical interest, the distribution $\mathds{P}$ is not precisely known, and therefore we miss essential information to solve problem \eqref{Ex-true} {\em exactly}. However, $\mathds{P}$ is often partially observable through a finite set of $N$ independent samples, {\em e.g.}, past realizations of the random vector $\xi$. We denote the training dataset comprising these samples by $\wh{\Xi}_N \coloneqq \{\wh{\xi}_i\}_{i\le N} \subseteq \Xi$. We emphasize that---before its revelation---the dataset $\wh{\Xi}_N$ can be viewed as a random object governed by the distribution $\mathds{P}^N$ supported on $\Xi^N$. A {\em data-driven solution} for problem \eqref{Ex-true} is a feasible decision $\wh{x}_N} %{x_{\rm DD}^\star \in \mathbb{X}$ that is constructed from the training dataset $\wh{\Xi}_N$. Throughout this paper, we notationally suppress the dependence of $\wh{x}_N} %{x_{\rm DD}^\star$ on the training samples in order to avoid clutter. Instead, we reserve the superscript `\,$\widehat{~}$\,' for objects that depend on the training data and thus constitute random objects governed by the product distribution $\mathds{P}^N$. The {\em out-of-sample performance} of $\wh{x}_N} %{x_{\rm DD}^\star$ is defined as $\mathds{E}^\mathds{P} \big[ h(\wh{x}_N} %{x_{\rm DD}^\star,\xi) \big]$ and can thus be viewed as the expected cost of $\wh{x}_N} %{x_{\rm DD}^\star$ under a new sample $\xi$ that is independent of the training dataset. As $\mathds{P}$ is unknown, however, the exact out-of-sample performance cannot be evaluated in practice, and the best we can hope for is to establish {\em performance guarantees} in the form of tight bounds. The feasibility of $\wh{x}_N} %{x_{\rm DD}^\star$ in~\eqref{Ex-true} implies $J^\star\leq \mathds{E}^\mathds{P} \big[ h(\wh{x}_N} %{x_{\rm DD}^\star,\xi) \big]$, but this lower bound is again of limited use as $J^\star$ is unknown and as our primary concern is to bound the costs from above. Thus, we seek data-driven solutions $\wh{x}_N} %{x_{\rm DD}^\star$ with performance guarantees of the type \begin{align} \label{out-of-sample} \mathds{P}^N\Big\{ \wh{\Xi}_N ~:~ \mathds{E}^\mathds{P} \big[ h(\wh{x}_N} %{x_{\rm DD}^\star,\xi) \big] \leq \wh{J}_N} %{J_{\rm DD}^\star \Big\}\geq 1-\beta, \end{align} where $\wh{J}_N} %{J_{\rm DD}^\star$ constitutes an upper bound that may depend on the training dataset, and $\beta\in (0,1)$ is a \emph{significance parameter} with respect to the distribution $\mathds{P}^N$, which governs both $\wh{x}_N} %{x_{\rm DD}^\star$ and $\wh{J}_N} %{J_{\rm DD}^\star$. \change{Hereafter we refer to $\wh{J}_N} %{J_{\rm DD}^\star$ as a {\em certificate} for the out-of-sample performance of $\wh{x}_N} %{x_{\rm DD}^\star$ and to $1-\beta$ as its {\em reliability}. Our ideal goal is to find a data-driven solution with the lowest possible out-of-sample performance. This is impossible, however, as $\mathds{P}$ is unknown, and the out-of-sample performance cannot be computed. We thus pursue a more modest but achievable goal, that is, to find a data-driven solution with a low certificate and a high reliability.} A natural approach to generate data-driven solutions $\wh{x}_N} %{x_{\rm DD}^\star$ is to approximate $\mathds{P}$ with the discrete empirical probability distribution \begin{align} \label{Pem} \wh{\PP}_N \coloneqq {1 \over N} \sum_{i = 1}^{N} \dir{\wh{\xi}_i}, \end{align} that is, the uniform distribution on $\wh{\Xi}_N$. This amounts to approximating the original stochastic program~\eqref{Ex-true} with the \emph{sample-average approximation} (SAA) problem \begin{align} \label{Ex_emp} \wh{J}_{\rm SAA} \coloneqq \inf_{x \in \mathbb{X}} \left\{ \mathds{E}^{\wh{\PP}_N} \big[ h(x,\xi) \big] = {1 \over N} \sum_{i = 1}^{N} h(x, \wh{\xi}_i)\right\} . \end{align} If the feasible set $\mathbb{X}$ is compact and the loss function is uniformly continuous in $x$ across all $\xi\in\Xi$, then the optimal value and optimal solutions of the SAA problem \eqref{Ex_emp} converge almost surely to their counterparts of the true problem \eqref{Ex-true} as $N$ tends to infinity \cite[Theorem~5.3]{ref:Shap&Dent&Rusz}. Even though finite sample performance guarantees of the type \eqref{out-of-sample} can be obtained under additional assumptions such as Lipschitz continuity of the loss function (see {\em e.g.},~\cite[Theorem~1]{ref:Shap:SAA-05}), the SAA problem has been conceived primarily for situations where the distribution $\mathds{P}$ is known and additional samples can be acquired cheaply via random number generation. However, the optimal solutions of the SAA problem tend to display a poor out-of-sample performance in situations where $N$ is small and where the acquisition of additional samples would be costly. In this paper we address problem \eqref{Ex-true} with an alternative approach that explicitly accounts for our ignorance of the true data-generating distribution $\mathds{P}$, and that offers attractive performance guarantees even when the acquisition of additional samples from $\mathds{P}$ is impossible or expensive. Specifically, we use $\wh{\Xi}_N$ to design an ambiguity set $\wh{\mathcal{P}}_N$ containing all distributions that could have generated the training samples with high confidence. \change{ This ambiguity set enables us to define the certificate $\wh{J}_N} %{J_{\rm DD}^\star$ as the optimal value of a distributionally robust optimization problem that minimize the {\em worst-case} expected cost.} \begin{align} \label{DRO} \wh{J}_N} %{J_{\rm DD}^\star \coloneqq \inf\limits_{x \in \mathbb{X}} \sup\limits_{\mathds{Q} \in \wh{\mathcal{P}}_N} \mathds{E}^\mathds{Q} \big[ h(x,\xi) \big] \end{align} \change{Following \cite{ref:PflWoz-07}, we construct $\wh{\mathcal{P}}_N$ as a ball around the empirical distribution~\eqref{Pem} with respect to the Wasserstein metric. In the remainder of the paper we will demonstrate that the optimal value $\wh{J}_N} %{J_{\rm DD}^\star$ as well as any optimal solution $\wh{x}_N} %{x_{\rm DD}^\star$ (if it exists) of the distributionally robust problem~\eqref{DRO} satisfy the following conditions.} \begin{enumerate}[label=(\roman*), itemsep = 1mm, topsep = 1mm] \item \label{cond-conf} {\bf Finite sample guarantee:} \change{For a carefully chosen size of the ambiguity set, the certificate $\wh{J}_N} %{J_{\rm DD}^\star$ provides a $1-\beta$ confidence bound of the type~\eqref{out-of-sample} on the out-of-sample performance of $\wh{x}_N} %{x_{\rm DD}^\star$.} \item \label{cond-asym} {\bf Asymptotic consistency:} \change{As $N$ tends to infinity, the certificate $\wh{J}_N} %{J_{\rm DD}^\star$ and the data-driven solution $\wh{x}_N} %{x_{\rm DD}^\star$ converge---in a sense to be made precise below---to the optimal value $J^\star$ and an optimizer $x^\star$ of the stochastic program~\eqref{Ex-true}, respectively.} \item \label{cond-trac} {\bf Tractability: } \change{For many loss functions $h(x,\xi)$ and sets $\mathbb{X}$, the distributionally robust problem \eqref{DRO} is computationally tractable and admits a reformulation reminiscent of the SAA problem~\eqref{Ex_emp}.} \end{enumerate} \change{Conditions~\ref{cond-conf}--\ref{cond-trac} have been identified in~\cite{ref:BertSAA-14} as desirable properties of data-driven solutions for stochastic programs. Precise statements of these conditions will be provided in the remainder. } In Section~\ref{sec:wass} we will use the Wasserstein metric to construct ambiguity sets of the type $\wh{\mathcal{P}}_N$ satisfying the conditions \ref{cond-conf} and \ref{cond-asym}. In Section~\ref{sec:dist}, we will demonstrate that these ambiguity sets also fulfill the tractability condition~\ref{cond-trac}. \change{We see this last result as the main contribution of this paper because the state-of-the-art method for solving distributionally robust problems over Wasserstein ambiguity sets relies on global optimization algorithms~\cite{ref:PflPich-14}.} \section{Wasserstein Metric and Measure Concentration} \label{sec:wass} \change{ Probability metrics represent distance functions on the space of probability distributions. One of the most widely used examples is the Wasserstein metric, which is defined on the space $\mathcal{M}(\Xi)$ of all probability distributions $\mathds{Q}$ supported on $\Xi$ with $\mathds{E}^\mathds{Q}\big[\|\xi\|\big] = \int_\Xi \|\xi\| \,\mathds{Q}(\mathrm{d}\xi)<\infty$. \begin{Def}[Wasserstein metric \cite{ref:KantRub-58}] \label{def:wass} The Wasserstein metric $d_{\rm W} : \mathcal{M}(\Xi)\times \mathcal{M}(\Xi)\rightarrow\mathbb{R}$ is defined via \begin{align*} \Wass{\mathds{Q}_1}{\mathds{Q}_2} \coloneqq \inf \left\{ \int_{\Xi^2} \| \xi_1 - \xi_2 \|\, \Pi(\mathrm{d} \xi_1, \mathrm{d} \xi_2) ~: \begin{array}{l}\mbox{$\Pi$ is a joint distribution of $\xi_1$ and $\xi_2$} \\ \mbox{with marginals $\mathds{Q}_1$ and $\mathds{Q}_2$, respectively}\! \end{array}\right\} \end{align*} for all distributions $\mathds{Q}_1,\mathds{Q}_2\in \mathcal{M}(\Xi)$, where $\|\cdot\|$ represents an arbitrary norm on $\mathbb{R}^m$. \end{Def} The decision variable $\Pi$ can be viewed as a \emph{transportation plan} for moving a mass distribution described by $\mathds{Q}_1$ to another one described by $\mathds{Q}_2$. Thus, the Wasserstein distance between $\mathds{Q}_1$ and $\mathds{Q}_2$ represents the cost of an optimal mass transportation plan, where the norm $\|\cdot\|$ encodes the transportation costs. We remark that a generalized $p$-Wasserstein metric for $p\geq 1$ is obtained by setting the transportation cost between $\xi_1$ and $\xi_2$ to $\|\xi_1-\xi_2\|^p$. In this paper, however, we focus exclusively on the $1$-Wasserstein metric of Definition~\ref{def:wass}, which is sometimes also referred to as the Kantorovich metric. We will sometimes also need the following dual representation of the Wasserstein metric. \begin{Thm}[Kantorovich-Rubinstein~\cite{ref:KantRub-58}] \label{thm:KantorovichRubinstein} For any distributions $\mathds{Q}_1, \mathds{Q}_2\in\mathcal M(\Xi)$ we have \begin{align*} \Wass{\mathds{Q}_1}{\mathds{Q}_2} = \sup_{f \in \mathcal{L}} \Big\{ \int_{\Xi} f(\xi) \,\mathds{Q}_1(\mathrm{d} \xi) - \int_{\Xi} f(\xi)\, \mathds{Q}_2(\mathrm{d} \xi)\Big\}, \end{align*} where $\mathcal{L}$ denotes the space of all Lipschitz functions with $|f(\xi)-f(\xi')|\leq \|\xi-\xi'\|$ for all $\xi,\xi'\in\Xi$. \end{Thm} Kantorovich and Rubinstein~\cite{ref:KantRub-58} originally established this result for distributions with bounded support. A modern proof for unbounded distributions is due to Villani~\cite[Remark~6.5, p.~107]{ref:Villani}. The optimization problems in Definition~\ref{def:wass} and Theorem~\ref{thm:KantorovichRubinstein}, which provide two equivalent characterizations of the Wasserstein metric, constitute a primal-dual pair of infinite-dimensional linear programs. The dual representation implies that two distributions $\mathds{Q}_1$ and $\mathds{Q}_2$ are close to each other with respect to the Wasserstein metric if and only if all functions with uniformly bounded slopes have similar integrals under $\mathds{Q}_1$ and $\mathds{Q}_2$. Theorem~\ref{thm:KantorovichRubinstein} also demonstrates that the Wasserstein metric is a special instance of an integral probability metric (see {\em e.g.} \cite{ref:Mull-97}) and that its generating function class coincides with a family of Lipschitz continuous functions. } In the remainder we will examine the ambiguity set \begin{align} \label{eq:wasserstein-ball} \ball{\wh{\PP}_N}{\varepsilon} \coloneqq \left\{ \mathds{Q} \in \mathcal{M}(\Xi) ~:~ \Wass{\wh{\PP}_N}{\mathds{Q}} \le \varepsilon\right \}, \end{align} which can be viewed as the Wasserstein ball of radius $\varepsilon$ centered at the empirical distribution~$\wh{\PP}_N$. Under a common light tail assumption on the unknown data-generating distribution $\mathds{P}$, this ambiguity set offers attractive performance guarantees in the spirit of Section~\ref{sec:prob}. \begin{As}[Light-tailed distribution] \label{a:exp} There exists an exponent $a > 1$ such that \begin{align*} A \coloneqq \mathds{E}^\mathds{P}\big[ \exp(\|\xi\|^a) \big] = \int_{\Xi} \exp(\|\xi\|^a)\,\mathds{P}(\mathrm{d} \xi) < \infty. \end{align*} \end{As} Assumption~\ref{a:exp} essentially requires the tail of the distribution $\mathds{P}$ to decay at an exponential rate. Note that this assumption trivially holds if $\Xi$ is compact. \change{Heavy-tailed distributions that fail to meet Assumption~\ref{a:exp} are difficult to handle even in the context of the classical sample average approximation. Indeed, under a heavy-tailed distribution the sample average of the loss corresponding to any fixed decision $x \in \mathbb{X}$ may not even converge to the expected loss; see {\em e.g.}~\cite{ref:Brown-15, ref:Catoni-12}.} The following modern measure concentration result provides the basis for establishing powerful finite sample guarantees. \change { \begin{Thm}[Measure concentration {\cite[Theorem 2]{ref:FouGui-14}}] \label{thm:concentration} If Assumption~\ref{a:exp} holds, we have \begin{align} \label{concentration} \mathds{P}^N \Big\{ \Wass{\mathds{P}}{\wh{\PP}_N} \ge \varepsilon \Big \} \le \left\{ \begin{array}{ll} c_1 \exp\big({-c_2N\varepsilon^{\max\{m,2\}}}\big) & \text{if } \varepsilon \le 1, \\ c_1 \exp\big({-c_2N\varepsilon^a}\big) & \text{if } \varepsilon > 1,\end{array}\right. \end{align} for all $N \ge 1$, $m \neq 2$, and $\varepsilon>0$, where $c_1, c_2$ are positive constants that only depend on $a$, $A$, and $m$.\footnote{\change{A similar but slightly more complicated inequality also holds for for the special case $m = 2$; see \cite[Theorem 2]{ref:FouGui-14} for details.}} \end{Thm} Theorem~\ref{thm:concentration} provides an a priori estimate of the probability that the unknown data-generating distribution~$\mathds{P}$ resides outside of the Wasserstein ball $\ball{\wh{\PP}_N}{\varepsilon}$. Thus, we can use Theorem~\ref{thm:concentration} to estimate the radius of the smallest Wasserstein ball that contains $\mathds{P}$ with confidence $1-\beta$ for some prescribed $\beta \in (0,1)$. Indeed, equating the right-hand side of \eqref{concentration} to $\beta$ and solving for $\varepsilon$ yields \begin{align} \label{eps_N} \varepsilon_N(\beta) \coloneqq \left\{ \begin{array}{ll} \Big({\log (c_1\beta^{-1}) \over c_2N} \Big)^{1/{\max\{m,2\}}} & \text{if } N \ge {\log(c_1\beta^{-1}) \over c_2}, \\ \Big({\log (c_1\beta^{-1}) \over c_2N} \Big)^{1/a} & \text{if } N < {\log(c_1\beta^{-1}) \over c_2}. \end{array}\right. \end{align} \begin{Thm}[Finite sample guarantee] \label{thm:fin} Suppose that Assumption~\ref{a:exp} holds and that $\beta\in (0,1)$. Assume also that $\wh{J}_N} %{J_{\rm DD}^\star$ and $\wh{x}_N} %{x_{\rm DD}^\star$ represent the optimal value and an optimizer of the distributionally robust program~\eqref{DRO} with ambiguity set $\wh{\mathcal{P}}_N = \ball{\wh{\PP}_N}{\varepsilon_N(\beta)}$. Then, the finite sample guarantee \eqref{out-of-sample} holds. \end{Thm} \begin{proof} The claim follows immediately from Theorem~\ref{thm:concentration}, which ensures via the definition of $\varepsilon_N(\beta)$ in~\eqref{eps_N} that $\mathds{P}^N \{ \mathds{P}\in \ball{\wh{\PP}_N}{\varepsilon_N(\beta)} \} \ge 1-\beta$. Thus, $\mathds{E}^\mathds{P} [ h(\wh{x}_N} %{x_{\rm DD}^\star,\xi)] \leq \sup_{\mathds{Q}\in\wh{\mathcal{P}}_N}\mathds{E}^\mathds{Q} [ h(\wh{x}_N} %{x_{\rm DD}^\star,\xi)] = \wh{J}_N} %{J_{\rm DD}^\star$ with probability $1-\beta$. \end{proof} It is clear from~\eqref{eps_N} that for any fixed $\beta>0$, the radius $ \varepsilon_N(\beta)$ tends to $0$ as $N$ increases. Moreover, one can show that if $\beta_N$ converges to zero at a carefully chosen rate, then the solution of the distributionally robust optimization problem~\eqref{DRO} with ambiguity set $\wh{\mathcal{P}}_N = \ball{\wh{\PP}_N}{\varepsilon_N(\beta_N)}$ converges to the solution of the original stochastic program~\eqref{Ex-true} as $N$ tends to infinity. The following theorem formalizes this statement. \begin{Thm}[Asymptotic consistency] \label{thm:convergence} Suppose that Assumption~\ref{a:exp} holds and that $\beta_N\in(0,1)$, $N \in \mathbb{N}$, satisfies $\sum_{N=1}^\infty\beta_N<\infty$ and $\lim_{N\rightarrow\infty}\varepsilon_N(\beta_N)=0$.\footnote{\change{A possible choice is $\beta_N = \exp(-\sqrt{N})$.}} Assume also that $\wh{J}_N} %{J_{\rm DD}^\star$ and $\wh{x}_N} %{x_{\rm DD}^\star$ represent the optimal value and an optimizer of the distributionally robust program~\eqref{DRO} with ambiguity set $\wh{\mathcal{P}}_N = \ball{\wh{\PP}_N}{\varepsilon_N(\beta_N)}$, $N\in\mathbb{N}$. \begin{enumerate}[label=(\roman*), itemsep = 1mm, topsep = 1mm] \item \label{thm:J-asy} If $h(x,\xi)$ is upper semicontinuous in $\xi$ and there exists $L\geq 0$ with $|h(x,\xi)|\leq L(1+\|\xi\|)$ for all $x\in\mathbb{X}$ and $\xi\in\Xi$, then $\mathds{P}^\infty$-almost surely we have $\wh{J}_N} %{J_{\rm DD}^\star \downarrow J^\star$ as $N \rightarrow \infty$ where $J^\star$ is the optimal value of~\eqref{Ex-true}. \item \label{thm:x-asy} If the assumptions of assertion~\ref{thm:J-asy} hold, $\mathbb{X}$ is closed, and $h(x,\xi)$ is lower semicontinuous in $x$ for every $\xi\in\Xi$, then any accumulation point of $\{\wh{x}_N} %{x_{\rm DD}^\star\}_{N \in \mathbb{N}}$ is $\mathds{P}^\infty$-almost surely an optimal solution for~\eqref{Ex-true}. \end{enumerate} \end{Thm} The proof of Theorem~\ref{thm:convergence} will rely on the following technical lemma. \begin{Lem}[Convergence of distributions] \label{lem:asy} If Assumption~\ref{a:exp} holds and $\beta_N\in(0,1)$, $N \in \mathbb{N}$, satisfies $\sum_{N=1}^\infty\beta_N<\infty$ and $\lim_{N\rightarrow\infty}\varepsilon_N(\beta_N)=0$, then, any sequence $\widehat\mathds{Q}_N \in \ball{\wh{\PP}_N}{\varepsilon_N(\beta_N)}$, $N\in\mathbb{N}$, where $\widehat \mathds{Q}_N$ may depend on the training data, converges under the Wasserstein metric (and thus weakly) to $\mathds{P}$ almost surely with respect to $\mathds{P}^\infty$, that is, \[\mathds{P}^{\infty} \left\{ \lim_{N \rightarrow \infty} \Wass{\mathds{P}}{\widehat\mathds{Q}_N} = 0 \right\} = 1.\] \end{Lem} \begin{proof} As $\widehat\mathds{Q}_N \in \ball{\wh{\PP}_N}{\delta_N}$, the triangle inequality for the Wasserstein metric ensures that \begin{align*} \Wass{\mathds{P}}{\widehat\mathds{Q}_N } \le \Wass{\mathds{P}}{\wh{\PP}_N} + \Wass{\wh{\PP}_N}{\widehat\mathds{Q}_N} \le \Wass{\mathds{P}}{\wh{\PP}_N} + \varepsilon_N(\beta_N). \end{align*} Moreover, Theorem~\ref{thm:concentration} implies that $\mathds{P}^N \{ \Wass{\mathds{P}}{\wh{\PP}_N} \le \varepsilon_N(\beta_N)\}\geq 1-\beta_N$, and thus we have $\mathds{P}^N \{ \Wass{\mathds{P}}{\widehat\mathds{Q}_N } \leq 2\varepsilon_N(\beta_N) \} \ge 1-\beta_N$. As $\sum_{N=1}^\infty\beta_N<\infty$, the Borel-Cantelli Lemma \cite[Theorem 2.18]{ref:Kallenberg-97} further implies that \[ \mathds{P}^{\infty} \left\{ \Wass{\mathds{P}}{\widehat\mathds{Q}_N} \le \varepsilon_N(\beta_N) ~ \text{for all sufficiently large } N \right\} = 1. \] Finally, as $\lim_{N\uparrow \infty}\varepsilon_N(\beta_N)=0$, we conclude that $\lim_{N\uparrow\infty}\Wass{\mathds{P}}{\widehat\mathds{Q}_N} =0$ almost surely. Note that convergence with respect to the Wasserstein metric implies weak convergence~\cite{ref:Bois-11}. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:convergence}] As $\widehat x_N\in\mathbb{X}$, we have $J^\star \le \mathds{E}^\mathds{P}[h(\widehat x_N,\xi)]$. Moreover, Theorem~\ref{thm:fin} implies that \[ \mathds{P}^N \left\{ J^\star\le\mathds{E}^\mathds{P}[h(\wh{x}_N} %{x_{\rm DD}^\star,\xi)] \le \wh{J}_N} %{J_{\rm DD}^\star \right\} \ge \mathds{P}^N \left\{ \mathds{P}\in \ball{\wh{\PP}_N}{\varepsilon_N(\beta_N)} \right\} \ge 1-\beta_N, \] for all $N \in \mathbb{N}$. As $\sum_{N=1}^\infty\beta_N<\infty$, the Borel-Cantelli Lemma further implies that \begin{equation*} \mathds{P}^{\infty} \left\{ J^\star \le \mathds{E}^{\mathds{P}}[h(\wh{x}_N} %{x_{\rm DD}^\star,\xi)] \le \wh{J}_N} %{J_{\rm DD}^\star ~ \text{for all sufficiently large }N \right\} = 1. \end{equation*} To prove assertion~\ref{thm:J-asy}, it thus remains to be shown that $\limsup_{N \rightarrow \infty}\wh{J}_N} %{J_{\rm DD}^\star \le J^\star$ with probability $1$. As $h(x,\xi)$ is upper semicontinuous and grows at most linearly in $\xi$, there exists a non-increasing sequence of functions $h_k(x,\xi)$, $k\in\mathbb{N}$, such that $h(x,\xi)=\lim_{k\rightarrow \infty} h_k(x,\xi)$, and $h_k(x,\xi)$ is Lipschitz continuous in $\xi$ for any fixed $x\in\mathbb{X}$ and $k\in\mathbb{N}$ with Lipschitz constant $L_k\geq 0$; see Lemma~\ref{lem:p.w.app} in the appendix. Next, choose any $\delta>0$, fix a $\delta$-optimal decision $x_\delta \in \mathbb{X}$ for \eqref{Ex-true} with $\mathds{E}^\mathds{P}[h(x_\delta,\xi)]\leq J^\star+\delta$, and for every $N\in\mathbb{N}$ let $\widehat \mathds{Q}_N \in \wh{\mathcal{P}}_N$ be a $\delta$-optimal distribution corresponding to $x_\delta$ with \[ \sup_{\mathds{Q} \in \wh{\mathcal{P}}_N}\mathds{E}^{\mathds{Q}}[h(x_\delta,\xi)] \le \mathds{E}^{\mathds{Q}_N}[h(x_\delta,\xi)] + \delta. \] Then, we have \begin{align*} \limsup_{N\rightarrow\infty}\wh{J}_N} %{J_{\rm DD}^\star \leq \limsup_{N \rightarrow \infty} \sup_{\mathds{Q} \in \wh{\mathcal{P}}_N}\mathds{E}^{\mathds{Q}}[h(x_\delta,\xi)] & \le \limsup_{N \rightarrow \infty} \mathds{E}^{\widehat \mathds{Q}_N} [h(x_\delta,\xi)] + \delta \\ & \le \lim_{k \rightarrow\infty} \limsup_{N \rightarrow \infty} \mathds{E}^{\widehat \mathds{Q}_N}[h_k(x_\delta,\xi)] + \delta \\ & \le \lim_{k \rightarrow\infty} \limsup_{N \rightarrow \infty} \left( \mathds{E}^{\mathds{P}}[h_k(x_\delta,\xi)] + L_k\, \Wass{\mathds{P}}{\widehat \mathds{Q}_N} \right) +\delta \\ & = \lim_{k \rightarrow \infty} \mathds{E}^{\mathds{P}}[h_k(x_\delta,\xi)] + \delta, \quad \mathds{P}^\infty\text{-almost surely}\\ &= \mathds{E}^{\mathds{P}}[h(x_\delta,\xi)] + \delta \leq J^\star+2\delta, \end{align*} where the second inequality holds because $h_k(x,\xi)$ converges from above to $h(x,\xi)$, and the third inequality follows from Theorem~\ref{thm:KantorovichRubinstein}. Moreover, the almost sure equality holds due to Lemma~\ref{lem:asy}, and the last equality follows from the Monotone Convergence Theorem \cite[Theorem 5.5]{ref:Lang-93}, which applies because $|\mathds{E}^{\mathds{P}}[h_k(x_\delta,\xi)]| < \infty$. Indeed, recall that $\mathds{P}$ has an exponentially decaying tail due to Assumption~\ref{a:exp} and that $h_k(x_\delta,\xi)$ is Lipschitz continuous in~$\xi$. As $\delta>0$ was chosen arbitrarily, we thus conclude that $\limsup_{N \rightarrow \infty}\wh{J}_N} %{J_{\rm DD}^\star \le J^\star$. To prove assertion~\ref{thm:x-asy}, fix an arbitrary realization of the stochastic process $\{\wh{\xi}_N\}_{N \in\mathbb{N}}$ such that $J^\star = \lim_{N \rightarrow \infty} \wh{J}_N} %{J_{\rm DD}^\star$ and $J^\star \le \mathds{E}^{\mathds{P}}[h(\wh{x}_N} %{x_{\rm DD}^\star,\xi)] \le \wh{J}_N} %{J_{\rm DD}^\star$ for all sufficiently large $N$. From the proof of assertion~\ref{thm:J-asy} we know that these two conditions are satisfied $\mathds{P}^\infty$-almost surely. Using these assumptions, one easily verifies that \begin{align} \label{pf:Jy} \liminf_{N \rightarrow \infty} \mathds{E}^{\mathds{P}}[h(\widehat x_{N},\xi)] \le \lim_{N \rightarrow \infty} \wh{J}_N} %{J_{\rm DD}^\star= J^\star. \end{align} Next, let $x^\star$ be an accumulation point of the sequence $\{\wh{x}_N} %{x_{\rm DD}^\star\}_{N \in\mathbb{N}}$, and note that $x^\star\in\mathbb{X}$ as $\mathbb{X}$ is closed. By passing to a subsequence, if necessary, we may assume without loss of generality that $x^\star = \lim_{N\rightarrow \infty}\wh{x}_N} %{x_{\rm DD}^\star$. Thus, \begin{align*} J^\star \le \mathds{E}^{\mathds{P}}[h(x^\star,\xi)] & \le \mathds{E}^{\mathds{P}}[\liminf_{N \rightarrow \infty} h(\widehat x_{N},\xi)] \le \liminf_{N \rightarrow \infty} \mathds{E}^{\mathds{P}}[h(\widehat x_{N},\xi)] \le J^\star, \end{align*} where the first inequality exploits that $x^\star \in \mathbb{X}$, the second inequality follows from the lower semicontinuity of $h(x,\xi)$ in $x$, the third inequality holds due to Fatou's lemma (which applies because $h(x,\xi)$ grows at most linearly in $\xi$), and the last inequality follows from~\eqref{pf:Jy}. Therefore, we have $\mathds{E}^{\mathds{P}}[h(x^\star,\xi)] = J^\star$. \end{proof} In the following we show that all assumptions of Theorem~\ref{thm:convergence} are necessary for asymptotic convergence, that is, relaxing any of these conditions can invalidate the convergence result. \begin{Ex}[Necessity of regularity conditions] \hfill \begin{enumerate}[itemsep = 1mm, topsep = 1mm] \item {\em Upper semicontinuity of $\xi \mapsto h(x,\xi)$ in Theorem~\ref{thm:convergence} \ref{thm:J-asy}:} \label{Ex:usc} \\ Set $\Xi = [0,1]$, $\mathds{P} = \dir{0}$ and $h(x,\xi) = \ind{(0,1]}(\xi)$, whereby $J^\star = 0$. As $\mathds{P}$ concentrates unit mass at $0$, we have $\wh{\PP}_N=\dir{0}=\mathds{P}$ irrespective of $N\in\mathbb{N}$. For any $\varepsilon > 0$, the Dirac distribution $\dir{\varepsilon}$ thus resides within the Wasserstein ball $\ball{\wh{\PP}_N}{\varepsilon}$. Hence, $\wh{J}_N} %{J_{\rm DD}^\star$ fails to converge to $J^\star$ for $\varepsilon\rightarrow 0$ because \begin{align*} \wh{J}_N} %{J_{\rm DD}^\star \ge \mathds{E}^{\dir{\varepsilon}} [h(x,\xi)] = h(x, \varepsilon) = 1,\quad \forall \varepsilon>0. \end{align*} \item {\em Linear growth of $\xi \mapsto h(x,\xi)$ in Theorem~\ref{thm:convergence} \ref{thm:J-asy}:} \label{Ex:growth} \\ Set $\Xi = \mathbb{R}$, $\mathds{P} = \dir{0}$ and $h(x,\xi) = \xi^2$, which implies that $J^\star=0$. Note that for any $\rho>\varepsilon$, the two-point distribution $\mathds{Q}_\rho = (1-\tfrac{\varepsilon}{\rho})\dir{0}+\tfrac{\varepsilon}{\rho}\dir{\rho}$ is contained in the Wasserstein ball $\ball{\wh{\PP}_N}{\varepsilon}$ of radius $\varepsilon >0$. Hence, $\wh{J}_N} %{J_{\rm DD}^\star$ fails to converge to $J^\star$ for $\varepsilon\rightarrow 0$ because \begin{align*} \wh{J}_N} %{J_{\rm DD}^\star \ge \, \sup_{\rho > \varepsilon} \,\mathds{E}^{\mathds{Q}_\rho} [h(x,\xi)] = \sup_{\rho > \varepsilon} \, \varepsilon \rho = \infty, \quad \forall \varepsilon>0. \end{align*} \item {\em Lower semicontinuity of $x \mapsto h(x,\xi)$ in Theorem~\ref{thm:convergence} \ref{thm:x-asy}:} \label{Ex:lsc}\\ Set $\mathbb{X} = [0,1]$ and $h(x,\xi) = \ind{[0.5,1]}(x)$, whereby $J^\star=0$ irrespective of $\mathds{P}$. As the objective is independent of $\xi$, the distributionally robust optimization problem~\eqref{DRO} is equivalent to~\eqref{Ex-true}. Then, $\widehat x_N = \tfrac{N-1}{2N}$ is a sequence of minimizers for~\eqref{DRO} whose accumulation point $x^\star = \tfrac{1}{2}$ fails to be optimal in~\eqref{Ex-true}. \end{enumerate} \end{Ex} Theorems~\ref{thm:fin} and \ref{thm:convergence} indicate that a careful a priori design of the Wasserstein ball results in attractive finite sample and asymptotic guarantees for the distributionally robust solutions. In practice, however, setting the Wasserstein radius to $\varepsilon_N(\beta)$ yields over-conservative solutions for the following reasons: \begin{itemize} \item Even though the constants $c_1$ and $c_2$ in \eqref{eps_N} can be computed based on the proof of Theorem~2 in~\cite{ref:FouGui-14}, the resulting Wasserstein ball is larger than necessary, {\em i.e.}, $\mathds{P}\notin \ball{\wh{\PP}_N}{\varepsilon_N(\beta)}$ with probability $\ll \beta$. \item Even if $\mathds{P}\notin \ball{\wh{\PP}_N}{\varepsilon_N(\beta)}$, the optimal value $\wh{J}_N} %{J_{\rm DD}^\star$ of \eqref{DRO} may still provide an upper bound on $J^\star$. \item The formula for $\varepsilon_N(\beta)$ in \eqref{eps_N} is independent of the training data. Allowing for random Wasserstein radii, however, results in a more efficient use of the available training data. \end{itemize} While Theorems~\ref{thm:fin} and \ref{thm:convergence} provide strong theoretical justification for using Wasserstein ambiguity sets, in practice, it is prudent to calibrate the Wasserstein radius via bootstrapping or cross-validation instead of using the conservative a priori bound $\varepsilon_N(\beta)$; see Section~\ref{sec:simulation} for further details. A similar approach has been advocated in \cite{ref:BertSAA-14} to determine the sizes of ambiguity sets that are constructed via goodness-of-fit tests. } So far we have seen that the Wasserstein metric allows us to construct ambiguity sets with favorable asymptotic and finite sample guarantees. In the remainder of the paper we will further demonstrate that the distributionally robust optimization problem \eqref{DRO} with a Wasserstein ambiguity set \eqref{eq:wasserstein-ball} is not significantly harder to solve than the corresponding SAA problem \eqref{Ex_emp}. \section{Solving Worst-Case Expectation Problems} \label{sec:dist} We now demonstrate that the inner worst-case expectation problem in \eqref{DRO} over the Wasserstein ambiguity set \eqref{eq:wasserstein-ball} can be reformulated as a finite convex program for many loss functions $h(x,\xi)$ of practical interest. For ease of notation, throughout this section we suppress the dependence on the decision variable $x$. Thus, we examine a generic worst-case expectation problem \begin{align} \label{dist-rob-Ex} \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{E}^\mathds{Q} \big[ \ell(\xi) \big] \end{align} involving a decision-{\em in}dependent loss function $\ell(\xi) \coloneqq \max_{k \le K}\ell_k(\xi)$, which is defined as the pointwise maximum of more elementary measurable functions $\ell_k:\mathbb{R}^m \rightarrow \overline {\R}$, $k\leq K$. The focus on loss functions representable as pointwise maxima is non-restrictive unless we impose some structure on the functions $\ell_k$. Many tractability results in the remainder of this paper are predicated on the following convexity assumption. \begin{As}[Convexity] \label{a:ell} The uncertainty set $\Xi\subseteq \mathbb{R}^m$ is convex and closed, and the negative constituent functions $-\ell_k$ are proper, convex, and lower semicontinuous for all $k\leq K$. Moreover, we assume that $\ell_k$ is not identically $-\infty$ on $\Xi$ for all $\le K$. \end{As} Assumption~\ref{a:ell} essentially stipulates that $\ell(\xi)$ can be written as a maximum of concave functions. As we will showcase in Section~\ref{sec:cases}, this mild restriction does not sacrifice much modeling power. Moreover, generalizations of this setting will be discussed in Section~\ref{sec:exten}. We proceed as follows. Subsection~\ref{subsec:worst-case} addresses the reduction of \eqref{dist-rob-Ex} to a finite convex program, while Subsection~\ref{subsec:ext-distr} describes a technique for constructing worst-case distributions. \subsection{Reduction to a Finite Convex Program} \label{subsec:worst-case} The worst-case expectation problem~\eqref{dist-rob-Ex} constitutes an infinite-dimensional optimization problem over probability distributions and thus appears to be intractable. \change{However, we will now demonstrate that \eqref{dist-rob-Ex} can be re-expressed as a finite-dimensional convex program by leveraging tools from robust optimization}. \begin{Thm}[Convex reduction] \label{thm:dist-rob-opt} If the convexity Assumption \ref{a:ell} holds, then for any $\varepsilon \ge0 $ the worst-case expectation~\eqref{dist-rob-Ex} equals the optimal value of the finite convex program \begin{align} \label{eq:thm-dual:2} \left\{ \begin{array}{clll} \inf\limits_{\lambda,s_i, z_{ik},\nu_{ik}} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} s_i && \\ \normalfont \text{s.t.} & [-\ell_k]^*(z_{ik} - \nu_{ik}) + \sigma_{\Xi}(\nu_{ik}) - \inner{z_{ik}}{\wh{\xi}_i} \le s_i & \forall i \le N, & \forall k \le K \\ & \|z_{ik}\|_* \le \lambda &\forall i \le N, & \forall k \le K. \end{array} \right. \end{align} \end{Thm} \change{ Recall that $[-\ell_k]^*(z_{ik} - \nu_{ik})$ denotes the conjugate of $-\ell_k$ evaluated at $z_{ik} - \nu_{ik}$ and $\|z_{ik}\|_*$ the dual norm of $z_{ik}$. Moreover, $\chi_\Xi$ represents the characteristic function of $\Xi$ and $\sigma_\Xi$ its conjugate, that is, the support function of $\Xi$. } \begin{proof}[Proof of Theorem~\ref{thm:dist-rob-opt}] \change{By using Definition~\ref{def:wass} we can re-express the worst-case expectation~\eqref{dist-rob-Ex} as} \begin{align*} \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{E}^\mathds{Q} \big[ \ell(\xi) \big] &= \left\{ \begin{array}{cl} \sup\limits_{\Pi,\mathds{Q}} & \int_{\Xi} \ell(\xi) \, \mathds{Q}(\mathrm{d} \xi) \\ \normalfont \text{s.t.} & \int_{\Xi^2} \|\xi -\xi'\| \, \Pi(\mathrm{d} \xi, \mathrm{d} \xi') \le \varepsilon\\[1ex] & \left\{ \begin{array}{l} \mbox{$\Pi$ is a joint distribution of $\xi$ and $\xi'$}\\ \mbox{with marginals $\mathds{Q}$ and $\wh{\PP}_N$, respectively} \end{array}\right. \end{array} \right.\\ & = \left\{ \begin{array}{cl} \sup\limits_{\mathds{Q}_i \in \mathcal{M}(\Xi)} & {1 \over N}\sum\limits_{i = 1}^{N} \int_{\Xi} \ell(\xi) \, \mathds{Q}_i(\mathrm{d} \xi) \\ \normalfont \text{s.t.} & {1 \over N}\sum\limits_{i = 1}^{N} \int_{\Xi} \|\xi -\wh{\xi}_i\| \, \mathds{Q}_i(\mathrm{d} \xi) \le \varepsilon. \end{array} \right. \end{align*} The second equality follows from the law of total probability, which asserts that any joint probability distribution $\Pi$ of $\xi$ and $\xi'$ can be constructed from the marginal distribution $\wh{\PP}_N$ of $\xi'$ and the conditional distributions $\mathds{Q}_i$ of $\xi$ given $\xi'=\wh{\xi}_i$, $i\leq N$, that is, we may write $\Pi = {1 \over N}\sum_{i = 1}^{N} \dir{\wh{\xi}_i}\otimes \mathds{Q}_i$. The resulting optimization problem represents a generalized moment problem in the distributions $\mathds{Q}_i$, $i\leq N$. Using a standard duality argument, we obtain \begin{subequations} \label{eq:pf:thm-dual} \begin{align} \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{E}^\mathds{Q} \big[ \ell(\xi) \big] &= \sup\limits_{\mathds{Q}_i \in \mathcal{M}(\Xi)} \inf\limits_{\lambda \ge 0} {1 \over N}\sum\limits_{i = 1}^{N} \int_{\Xi} \ell(\xi)\, \mathds{Q}_i(\mathrm{d} \xi) + \lambda \Big( \varepsilon - {1 \over N}\sum\limits_{i = 1}^{N} \int_{\Xi} \|\xi -\wh{\xi}_i\|\, \mathds{Q}_i(\mathrm{d} \xi) \Big) \notag \\ & \le \inf\limits_{\lambda \ge 0} \sup\limits_{\mathds{Q}_i \in \mathcal{M}(\Xi)} \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} \int_{\Xi} \left( \ell(\xi) - \lambda \|\xi -\wh{\xi}_i\|\right)\mathds{Q}_i(\mathrm{d} \xi) \label{eq:1} \\ & = \inf\limits_{\lambda \ge 0} \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} \sup_{\xi \in \Xi} \left (\ell(\xi) - \lambda \|\xi - \wh{\xi}_i\| \right ), \label{eq:1-} \end{align} where \change{\eqref{eq:1} follows from the max-min inequality}, and \eqref{eq:1-} follows from the fact that $\mathcal{M}(\Xi)$ contains all the Dirac distributions supported on $\Xi$. Introducing epigraphical auxiliary variables $s_i$, $i\leq N$, allows us to reformulate \eqref{eq:1-} as \begin{align} & \left\{ \begin{array}{clll} \inf\limits_{\lambda, s_i} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} s_i &&\\ \normalfont \text{s.t.} & \sup\limits_{\xi \in \Xi} \Big(\ell(\xi) - \lambda \|\xi - \wh{\xi}_i\| \Big)\le s_i & \forall i\le N&\\ & \lambda \ge 0 && \end{array} \right. \label{eq:2-} \\ =~& \left\{ \begin{array}{clll} \inf\limits_{\lambda, s_i} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} s_i &&\\ \normalfont \text{s.t.} & \sup\limits_{\xi \in \Xi} \Big(\ell_k(\xi) - \max\limits_{\|z_{ik}\|_* \le \lambda} \inner{z_{ik}}{\xi - \wh{\xi}_i} \Big)\le s_i &\forall i\le N,& \forall k\le K \\ & \lambda \ge 0 && \end{array} \right. \label{eq:2} \\ \le~& \left\{ \begin{array}{clll} \inf\limits_{\lambda, s_i} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} s_i &&\\ \normalfont \text{s.t.} & \min\limits_{\|z_{ik}\|_* \le \lambda} \sup\limits_{\xi \in \Xi} \Big(\ell_k(\xi) - \inner{z_{ik}}{\xi - \wh{\xi}_i} \Big)\le s_i &\forall i\le N,& \forall k\le K \\ & \lambda \ge 0. && \end{array} \right. \label{eq:3} \end{align} Equality~\eqref{eq:2} exploits the definition of the dual norm and the decomposability of $\ell(\xi)$ into its constituents $\ell_k(\xi)$, $k\le K$. Interchanging the maximization over $z_{ik}$ with the minus sign (thereby converting the maximization to a minimization) and then with the maximization over $\xi$ leads to a restriction of the feasible set of \eqref{eq:2}. The resulting upper bound \eqref{eq:3} can be re-expressed as \begin{align} & \notag \left\{ \begin{array}{clll} \inf\limits_{\lambda, s_i,z_{ik}} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} s_i &&\\ \normalfont \text{s.t.} & \sup\limits_{\xi \in \Xi} \Big(\ell_k(\xi) - \inner{z_{ik}}{\xi} \Big) + \inner{z_{ik}}{\wh{\xi}_i}\le s_i &\forall i\le N,& \forall k\le K \\ & \|z_{ik}\|_* \le \lambda & \forall i\le N, & \forall k\le K \end{array} \right. \\ = ~& \left\{ \begin{array}{clll} \inf\limits_{\lambda, s_i,z_{ik}} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} s_i &&\\ \normalfont \text{s.t.} & [-\ell_k + \indI{\Xi}]^*(z_{ik}) - \inner{z_{ik}}{\wh{\xi}_i}\le s_i &\forall i\le N,& \forall k\le K \\ & \|z_{ik}\|_* \le \lambda & \forall i\le N, & \forall k\le K, \end{array} \right. \label{eq:4} \end{align} \end{subequations} where \eqref{eq:4} follows from the definition of conjugacy, our conventions of extended arithmetic, and the substitution of $z_{ik}$ with $-z_{ik}$. Note that \eqref{eq:4} is already a finite convex program. Next, we show that Assumption~\ref{a:ell} reduces the inequalities \eqref{eq:1} and \eqref{eq:3} to equalities. Under Assumption~\ref{a:ell}, the inequality \eqref{eq:1} is in fact an equality for any $\varepsilon > 0$ by virtue of an extended version of {a} well-known strong duality result for moment problems \cite[Proposition~3.4]{ref:Shap-dual-01}. One can show that \eqref{eq:1} continues to hold as an equality even for $\varepsilon = 0$, in which case the Wasserstein ambiguity set \eqref{eq:wasserstein-ball} reduces to the singleton $\{\wh{\PP}_N\}$, while \eqref{dist-rob-Ex} reduces to the sample average $\frac{1}{N}\sum_{i=1}^N \ell(\wh{\xi}_i)$. Indeed, for $\varepsilon=0$ the variable $\lambda$ in \eqref{eq:1-} can be increased indefinitely at no penalty. As $\ell(\xi)$ constitutes a pointwise maximum of upper semicontinuous concave functions, an elementary but tedious argument shows that \eqref{eq:1-} converges to the sample average $\frac{1}{N}\sum_{i=1}^N \ell(\wh{\xi}_i)$ as $\lambda$ tends to infinity. The inequality \eqref{eq:3} also reduces to an equality under Assumption \ref{a:ell} thanks to the classical minimax theorem \cite[Proposition 5.5.4]{ref:Bert-09}, which applies because the set $\{z_{ik} \in \mathbb{R}^m : \|z_{ik}\|_* \le \lambda\}$ is compact for any finite $\lambda\geq 0$. Thus, the optimal values of \eqref{dist-rob-Ex} and \eqref{eq:4} coincide. Assumption~\ref{a:ell} further implies that the function $-\ell_k+\indI{\Xi}$ is proper, convex and lower semicontinuous. Properness holds because $\ell_k$ is not identically $-\infty$ on $\Xi$. By \cite[Theorem~11.23(a), p.~493]{ref:Rockafellar-10}, its conjugate essentially coincides with the \emph{epi-addition} (also known as \emph{inf-convolution}) of the conjugates of the functions $-\ell_k$ and $\sigma_{\Xi}$. Thus, \begin{align*} \change{[-\ell_k + \indI{\Xi}]^*(z_{ik})} & = \inf_{\nu_{ik}} \Big([-\ell_k]^*(z_{ik} - \nu_{ik}) + [\indI{\Xi}]^*(\nu_{ik}) \Big) \\ & = \cl\Big[\inf_{\nu_{ik}} \Big([-\ell_k]^*(z_{ik} - \nu_{ik}) + \sigma_{\Xi}(\nu_{ik}) \Big)\Big], \end{align*} where $\cl[\cdot]$ denotes the closure operator that maps any function to its largest lower semicontinuous minorant. As $\cl[f(\xi)]\leq 0$ if and only if $f(\xi)\leq 0$ for any function $f$, we may conclude that \eqref{eq:4} is indeed equivalent to \eqref{eq:thm-dual:2} under Assumption~\ref{a:ell}. \end{proof} Note that the semi-infinite inequality in \eqref{eq:2-} generalizes the nonlinear uncertain constraints studied in \cite{ref:BenHerVial-15} because it involves an additional norm term and as the loss function $\ell(\xi)$ is not necessarily concave under Assumption~\ref{a:ell}. As in \cite{ref:BenHerVial-15}, however, the semi-infinite constraint admits a robust counterpart that involves the conjugate of the loss function and the support function of the uncertainty set. From the proof of Theorem~\ref{thm:dist-rob-opt} it is immediately clear that the worst-case expectation~\eqref{dist-rob-Ex} is conservatively approximated by the optimal value of the finite convex program \eqref{eq:4} even if Assumption~\ref{a:ell} fails to hold. In this case the sum $-\ell_k + \indI{\Xi}$ in \eqref{eq:4} must be evaluated under our conventions of extended arithmetics, whereby $\infty - \infty = \infty$. These observations are formalized in the following corollary. \begin{Cor}[Approximate convex reduction] \label{cor:approx} For any $\varepsilon \ge0$, the worst-case expectation~\eqref{dist-rob-Ex} is smaller or equal to the optimal value of the finite convex program~\eqref{eq:4}. \end{Cor} \subsection{Extremal Distributions} \label{subsec:ext-distr} Stress test experiments are instrumental to assess the quality of candidate decisions in stochastic optimization. Meaningful stress tests require a good understanding of the extremal distributions from within the Wasserstein ball that achieve the worst-case expectation \eqref{dist-rob-Ex} for various loss functions. We now show that such extremal distributions can be constructed systematically from the solution of a convex program akin to~\eqref{eq:thm-dual:2}. \begin{Thm}[Worst-case distributions] \label{thm:dual-dual} If Assumption \ref{a:ell} holds, then the worst-case expectation~\eqref{dist-rob-Ex} coincides with the optimal value of the finite convex program \begin{align} \label{dual-dual} \left\{ \begin{array}{clll} \sup\limits_{\alpha_{ik}, q_{ik}} & {1 \over N} \sum\limits_{i = 1}^{N} \sum\limits_{k = 1}^{K} \alpha_{ik}\ell_k\big( \wh{\xi}_i - {q_{ik} \over \alpha_{ik}}\big) \\ \normalfont \text{s.t.} & {1 \over N}\sum\limits_{i =1}^{N} \sum\limits_{k =1}^{K} \|q_{ik}\| \le \varepsilon \\ & \sum\limits_{k = 1}^{K} \alpha_{ik} = 1 &\forall i \le N\\ & \alpha_{ik} \ge 0 &\forall i \le N, \quad \forall k\le K \\ & \wh{\xi}_i - {q_{ik} \over \alpha_{ik}} \in \Xi &\forall i \le N, \quad \forall k\le K \end{array} \right. \end{align} irrespective of $\varepsilon \ge 0$. Let $\big\{\alpha_{ik}(r), q_{ik}(r)\big\}_{r \in \mathbb{N}}$ be a sequence of feasible decisions whose objective values converge to the supremum of \eqref{dual-dual}. Then, the discrete probability distributions $$\mathds{Q}_r \coloneqq {1 \over N}\sum_{i = 1}^{N}\sum_{k = 1}^{K} \alpha_{ik}(r)\dir{\xi_{ik}(r)} \qquad \mbox{with}\qquad \xi_{ik}(r) \coloneqq \wh{\xi}_i - {q_{ik}(r) \over \alpha_{ik}(r)}$$ belong to the Wasserstein ball $\ball{\wh{\PP}_N}{\varepsilon}$ and attain the supremum of \eqref{dist-rob-Ex} asymptotically, i.e., \begin{align*} \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{E}^\mathds{Q} \big[ \ell(\xi) \big] = \lim\limits_{r \rightarrow \infty} \mathds{E}^{\mathds{Q}_r} \big[ \ell(\xi) \big] = \lim\limits_{k \rightarrow \infty} {1 \over N} \sum\limits_{i = 1}^{N} \sum\limits_{k = 1}^{K} \alpha_{ik}(r)\ell\big(\xi_{ik}(r)\big). \end{align*} \end{Thm} We highlight that all fractions in \eqref{dual-dual} must again be evaluated under our conventions of extended arithmetics. Specifically, if $\alpha_{ik}=0$ and $q_{ik}\neq 0$, then $q_{ik}/\alpha_{ik}$ has at least one component equal to $+\infty$ or $-\infty$, which implies that $\wh{\xi}_i - q_{ik}/\alpha_{ik}\notin \Xi$. In contrast, if $\alpha_{ik}=0$ and $q_{ik}= 0$, then $\wh{\xi}_i - q_{ik} / \alpha_{ik}=\wh{\xi}_i \in \Xi$. Moreover, the $ik$-th component in the objective function of \eqref{dual-dual} evaluates to $0$ whenever $\alpha_{ik} =0$ regardless of $q_{ik}$. The proof of Theorem \ref{thm:dual-dual} is based on the following technical lemma. \begin{Lem} \label{lem:perspective} Define $F: \mathbb{R}^m \times \mathbb{R}_{+} \rightarrow \overline {\R}$ through $F(q,\alpha) = \inf_{z \in \mathbb{R}^m} \inner{z}{q - \alpha \widehat{\xi}} + \alpha f^*(z)$ for some proper, convex, and lower semicontinuous function $f:\mathbb{R}^m\rightarrow \overline {\R}$ and reference point $\widehat{\xi}\in\mathbb{R}^m$. Then, $F$ coincides with the (extended) perspective function of the mapping $q \mapsto -f(\widehat \xi - q)$, that is, \begin{align*} F(q, \alpha) = \left\{\begin{array}{cc} - \alpha f \big(\widehat{\xi} - q /\alpha\big)& \text{if }\alpha > 0, \\ -\indI{\{0\}}(q) & \text{if }\alpha = 0. \end{array} \right. \end{align*} \end{Lem} \begin{proof} By construction, we have $F(q,0) = \inf_{z \in \mathbb{R}^m} \inner{z}{q} = - \indI{\{0\}}(q)$. For $\alpha > 0$, on the other hand, the definition of conjugacy implies that \begin{align*} F(q,\alpha) = -[\alpha f^*]^*(\alpha\widehat{\xi} - q) = - \alpha [f^*]^* \big(\widehat{\xi} - {q / \alpha}\big). \end{align*} The claim then follows because $[f^*]^* = f$ for any proper, convex, and lower semicontinuous function $f$ \cite[Proposition~1.6.1(c)]{ref:Bert-09}. Additional information on perspective functions can be found in \cite[Section~2.2.3, p.~39]{ref:Boyd}. \end{proof} \begin{proof}[Proof of Theorem \ref{thm:dual-dual}] By Theorem \ref{thm:dist-rob-opt}, which applies under Assumption~\ref{a:ell}, the worst-case expectation~\eqref{dist-rob-Ex} coincides with the optimal value of the convex program~\eqref{eq:thm-dual:2}. From the proof of Theorem \ref{thm:dist-rob-opt} we know that \eqref{eq:thm-dual:2} is equivalent to \eqref{eq:4}. \begin{subequations} The Lagrangian dual of \eqref{eq:4} is given by \begin{align} \left\{ \begin{array}{clll} \sup\limits_{\beta_{ik}, \alpha_{ik}} &\inf\limits_{\lambda, s_i, z_{ik}} \lambda \varepsilon + \sum\limits_{i = 1}^{N} \Big[ {s_i \over N} + &\hspace{-3mm}\sum\limits_{k = 1}^{K} \big[\beta_{ik} \big(\|z_{ik}\|_* -\lambda \big) + \alpha_{ik}\big( [-\ell_k + \indI{\Xi}]^*(z_{ik}) - \inner{z_{ik}}{\wh{\xi}_i} - s_i\big)\big]\Big] \\ \normalfont \text{s.t.} & \alpha_{ik} \ge 0& \forall i \le N, \quad \forall k\le K \\ & \beta_{ik} \ge 0 & \forall i \le N, \quad \forall k\le K, \end{array} \right. \notag \end{align} where the products of dual variables and constraint functions in the objective are evaluated under the standard convention $0 \cdot \infty = 0$. Strong duality holds since the function $[-\ell_k+\indI{\Xi}]^*$ is proper, convex, and lower semicontinuous under Assumption~\ref{a:ell} and because this function appears in a constraint of \eqref{eq:4} whose right-hand side is a free decision variable. By explicitly carrying out the minimization over $\lambda$ and $s_i$, one can show that the above dual problem is equivalent to \begin{align} \left\{ \begin{array}{clll} \sup\limits_{\beta_{ik}, \alpha_{ik}} & \inf\limits_{z_{ik}} ~ \sum\limits_{i = 1}^{N} \sum\limits_{k = 1}^{K} \beta_{ik} \|z_{ik}\|_* + &\hspace{-3mm}\alpha_{ik}[-\ell_k+ \indI{\Xi}]^*(z_{ik}) - \alpha_{ik}\inner{z_{ik}}{\wh{\xi}_i} \\ \normalfont \text{s.t.} & \sum\limits_{i =1}^{N} \sum\limits_{k = 1}^{K} \beta_{ik} = \varepsilon \\ & \sum\limits_{k = 1}^{K} \alpha_{ik} = {1 \over N} &\forall i \le N\\ & \alpha_{ik} \ge 0 & \forall i \le N, \quad \forall k\le K \\ & \beta_{ik} \ge 0 & \forall i \le N, \quad \forall k\le K. \end{array} \right. \label{eq:pf:dual-dual:1} \end{align} By using the definition of the dual norm, \eqref{eq:pf:dual-dual:1} can be re-expressed as \begin{align} & \left\{ \begin{array}{clll} \sup\limits_{\beta_{ik}, \alpha_{ik}}& \inf\limits_{z_{ik}} \sum\limits_{i = 1}^{N} \sum\limits_{k = 1}^{K} \max\limits_{\|q_{ik}\| \le \beta_{ik}}\inner{z_{ik}}{q_{ik}} + &\hspace{-3mm} \alpha_{ik}[-\ell_k + \indI{\Xi}]^*(z_{ik}) - \alpha_{ik}\inner{z_{ik}}{\wh{\xi}_i} \Big] \\ \normalfont \text{s.t.} & \sum\limits_{i =1}^{N} \sum\limits_{k =1}^{K}\beta_{ik} = \varepsilon \\ & \sum\limits_{k = 1}^{K} \alpha_{ik} = {1 \over N} &\forall i \le N\\ & \alpha_{ik} \ge 0 & \forall i \le N, \quad \forall k\le K \\ & \beta_{ik} \ge 0 & \forall i \le N, \quad \forall k\le K \end{array} \right. \\ =~& \left\{ \begin{array}{clll} \sup\limits_{\beta_{ik},\alpha_{ik}}& \max\limits_{\|q_{ik}\| \le \beta_{ik}} \inf\limits_{z_{ik}} \sum\limits_{i = 1}^{N} \sum\limits_{k =1}^{K} \inner{z_{ik}}{q_{ik}} + &\hspace{-3mm}\alpha_{ik}[-\ell_k+ \indI{\Xi}]^*(z_{ik}) - \alpha_{ik}\inner{z_{ik}}{\wh{\xi}_i} \\ \normalfont \text{s.t.} & \sum\limits_{i =1}^{N} \sum\limits_{k =1}^{K} \beta_{ik} = \varepsilon \\ & \sum\limits_{k = 1}^{K} \alpha_{ik} = {1 \over N} &\forall i \le N\\ & \alpha_{ik} \ge 0 & \forall i \le N, \quad \forall k\le K \\ & \beta_{ik} \ge 0 & \forall i \le N, \quad \forall k\le K, \end{array} \right. \label{eq:pf:dual-dual:2a} \end{align} where \eqref{eq:pf:dual-dual:2a} follows from the classical minimax theorem and the fact that the $q_{ik}$ variables range over a non-empty and compact feasible set for any finite $\varepsilon$; see \cite[Proposition 5.5.4]{ref:Bert-09}. Eliminating the $\beta_{ik}$ variables and using Lemma~\ref{lem:perspective} allows us to reformulate \eqref{eq:pf:dual-dual:2a} as \begin{align} &\left\{ \begin{array}{clll} \sup\limits_{\alpha_{ik}, q_{ik}} & \inf\limits_{z_{ik}}~ \sum\limits_{i = 1}^{N} \sum\limits_{k =1}^{K} \inner{z_{ik}}{q_{ik} - \alpha_{ik}\wh{\xi}_i} + &\hspace{-3mm} \alpha_{ik}[-\ell_k+ \indI{\Xi}]^*(z_{ik}) \\ \normalfont \text{s.t.} & \sum\limits_{i =1}^{N} \sum\limits_{k =1}^{K} \|q_{ik}\| \le \varepsilon \\ & \sum\limits_{k = 1}^{K} \alpha_{ik} = {1 \over N} &\forall i \le N\\ & \alpha_{ik} \ge 0 &\forall i \le N,\quad \forall k\le K \end{array} \right. \label{eq:pf:dual-dual:2b}\\ =~ & \left\{ \begin{array}{clll} \sup\limits_{\alpha_{ik}, q_{ik}} & \sum\limits_{i = 1}^{N} \sum\limits_{k = 1}^{K} - \alpha_{ik} \Big(-\ell_k\big(\wh{\xi}_i - {q_{ik} \over \alpha_{ik}}\big) + &\hspace{-3mm} \indI{\Xi} \big(\wh{\xi}_i - {q_{ik} \over \alpha_{ik}}\big) \Big)\ind{\{\alpha_{ik}>0\}} - \indI{\{0\}}(q_{ik})\ind{\{\alpha_{ik} = 0\}} \\ \normalfont \text{s.t.} & \sum\limits_{i =1}^{N} \sum\limits_{k =1}^{K} \|q_{ik}\| \le \varepsilon \\ & \sum\limits_{k = 1}^{K} \alpha_{ik} = {1 \over N} &\forall i \le N\\ & \alpha_{ik} \ge 0 &\forall i \le N, \quad \forall k\le K. \end{array} \right. \label{eq:pf:dual-dual:3} \end{align} Our conventions of extended arithmetics imply that the $ik$-th term in the objective function of problem~\eqref{eq:pf:dual-dual:3} simplifies to \begin{equation} \label{nasty_alpha} \alpha_{ik} \ell_k\big(\wh{\xi}_i - {q_{ik} \over \alpha_{ik}}\big) - \indI{\Xi}\big(\wh{\xi}_i - {q_{ik} \over \alpha_{ik}}\big). \end{equation} Indeed, for $\alpha_{ik}>0$, this identity trivially holds. For $\alpha_{ik}=0$, on the other hand, the $ik$-th objective term in \eqref{eq:pf:dual-dual:3} reduces to $- \indI{\{0\}}(q_{ik})$. Moreover, the first term in \eqref{nasty_alpha} vanishes whenever $\alpha_{ik} = 0$ regardless of $q_{ik}$, and the second term in \eqref{nasty_alpha} evaluates to 0 if $q_{ik}=0$ (as $0/0=0$ and $\wh{\xi}_i \in \Xi$) and to $-\infty$ if $q_{ik}\neq 0$ (as $q_{ik}/0$ has at least one infinite component, implying that $\wh{\xi}_i+q_{ik}/0\notin \Xi$). Therefore, \eqref{nasty_alpha} also reduces to $- \indI{\{0\}}(q_{ik})$ when $\alpha_{ik}=0$. This proves that the $ik$-th objective term in \eqref{eq:pf:dual-dual:3} coincides with \eqref{nasty_alpha}. Substituting \eqref{nasty_alpha} into \eqref{eq:pf:dual-dual:3} and re-expressing $- \indI{\Xi}\big(\wh{\xi}_i - {q_{ik} \over \alpha_{ik}}\big)$ in terms of an explicit hard constraint yields \begin{align} \left\{ \begin{array}{clll} \sup\limits_{\alpha_{ik}, q_{ik}} & \sum\limits_{i = 1}^{N} \sum\limits_{k = 1}^{K} \alpha_{ik} \ell_k\big(\wh{\xi}_i - {q_{ik} \over \alpha_{ik}}\big) \\ \normalfont \text{s.t.} & \sum\limits_{i =1}^{N} \sum\limits_{k =1}^{K} \|q_{ik}\| \le \varepsilon \\ & \sum\limits_{k = 1}^{K} \alpha_{ik} = {1 \over N} &\forall i \le N\\ & \alpha_{ik} \ge 0 &\forall i \le N, \quad \forall k\le K \\ & \wh{\xi}_i - {q_{ik} \over \alpha_{ik}} \in \Xi&\forall i \le N, \quad \forall k\le K. \end{array}\right. \label{eq:pf:dual-dual:4} \end{align} Finally, replacing $\big\{\alpha_{ik}, q_{ik}\big\}$ with ${1 \over N}\big\{\alpha_{ik}, q_{ik}\big\}$ shows that \eqref{eq:pf:dual-dual:4} is equivalent to \eqref{dual-dual}. This completes the first part of the proof. As for the second claim, let $\{\alpha_{ik}(r), q_{ik}(r)\}_{r \in \mathbb{N}}$ be a sequence of feasible solutions that attains the supremum in~\eqref{dual-dual}, and set $\xi_{ik}(r) \coloneqq \wh{\xi}_i - {q_{ik}(r) \over \alpha_{ik}(r)}\in\Xi$. Then, the discrete distribution $$\Pi_r \coloneqq {1 \over N}\sum_{i = 1}^{N}\sum_{k = 1}^{K} \alpha_{ik}(r)\dir{\big(\xi_{ik}(r), \wh{\xi}_i\big)}$$ has the distribution $\mathds{Q}_r$ defined in the theorem statement and the empirical distribution $\wh{\PP}_N$ as marginals. \change{By the definition of the Wasserstein metric, $\Pi_r$ represents a feasible mass transportation plan that provides an upper bound on the distance between $\wh{\PP}_N$ and $\mathds{Q}_r$; see Definition~\ref{def:wass}. Thus, we have } \begin{align*} \Wass{\mathds{Q}_r}{\wh{\PP}_N} &\le \int_{\Xi^2} \|\xi - \xi'\| \, \Pi_r(\mathrm{d} \xi, \mathrm{d} \xi') = {1 \over N} \sum\limits_{i = 1}^{N} \sum\limits_{k= 1}^{K} \alpha_{ik}(r) \big\| \xi_{ik}(r) - \wh{\xi}_i \big\| \change{= {1 \over N} \sum\limits_{i = 1}^{N} \sum\limits_{k= 1}^{K} \big\| q_{ik}(r)\big\| \le \varepsilon,} \end{align*} where the last inequality follows readily from the feasibility of $q_{ik}(r)$ in \eqref{dual-dual}. We conclude that \begin{align*} \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}}\mathds{E}^{\mathds{Q}}\big[\ell(\xi)\big] &\ge \limsup_{k \rightarrow \infty} \mathds{E}^{\mathds{Q}_r}\big[\ell(\xi)\big] = \limsup_{k \rightarrow \infty} {1 \over N} \sum\limits_{i = 1}^{N}\sum\limits_{k = 1}^{K}\alpha_{ik}(r) \ell\big(\xi_{ik}(r) \big)\\ & \ge \limsup_{k \rightarrow \infty} {1 \over N} \sum\limits_{i = 1}^{N}\sum\limits_{k = 1}^{K}\alpha_{ik}(r) \ell_k\big(\xi_{ik}(r) \big) = \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}}\mathds{E}^{\mathds{Q}}\big[\ell(\xi)\big], \end{align*} \end{subequations} where the first inequality holds as $\mathds{Q}_r \in \ball{\wh{\PP}_N}{\varepsilon}$ for all $k \in \mathbb{N}$, and the second inequality uses the trivial estimate $\ell \geq \ell_k$ for all $k\le K$. The last equality follows from the construction of $\alpha_{ik}(r)$ and $\xi_{ik}(r)$ and the fact that \eqref{dual-dual} coincides with the worst-case expectation~\eqref{dist-rob-Ex}. \end{proof} In the rest of this section we discuss some notable properties of the convex program \eqref{dual-dual}. In the \emph{ambiguity-free} limit, that is, when the radius of the Wasserstein ball is set to zero, then the optimal value of the convex program~\eqref{dual-dual} reduces to the expected loss under the empirical distribution. Indeed, for $\varepsilon = 0$ all $q_{ik}$ variables are forced to zero, and $\alpha_{ik}$ enters the objective only through $\sum_{k=1}^K \alpha_{ik}={1\over N}$. Thus, the objective function of \eqref{dual-dual} simplifies to $ \mathds{E}^{\wh{\PP}_N}[\ell(\xi)]$. We further emphasize that it is not possible to guarantee the existence of a worst-case distribution that attains the supremum in \eqref{dist-rob-Ex}. In general, as shown in Theorem~\ref{thm:dual-dual}, we can only construct a sequence of distributions that attains the supremum asymptotically. The following example discusses an instance of \eqref{dist-rob-Ex} that admits no worst-case distribution. \begin{Ex}[Non-existence of a worst-case distribution] \label{ex:worst-case} Assume that $\Xi = \mathbb{R}$, $N = 1$, $\wh{\xi}_1 = 0$, $K = 2$, $\ell_1(\xi) =0$ and $\ell_2(\xi) = \xi - 1$. In this case we have $\wh{\PP}_N=\dir{\{0\}}$, and problem \eqref{dual-dual} reduces to \begin{align*} \sup\limits_{\mathds{Q} \in \ball{\dir{0}}{\varepsilon}} \mathds{E}^\mathds{Q} \big[ \ell(\xi) \big] = \left\{ \begin{array}{clll} \sup\limits_{\alpha_{1j}, q_{1j}} & - q_{12} - \alpha_{12} \\ \normalfont \text{s.t.} & |q_{11}| + |q_{12}| \le \varepsilon \\ & \alpha_{11} + \alpha_{12} = 1 \\ & \change{\alpha_{11} \ge 0, \quad \alpha_{12} \ge 0.} \end{array} \right. \end{align*} The supremum on the right-hand side amounts to $\varepsilon$ and is attained, for instance, by the sequence $\alpha_{11}(r) = 1 - {1 \over k}$, $\alpha_{12}(r) = {1 \over k}$, $q_{11}(r) = 0$, $q_{12}(r) = - \varepsilon$ for $k\in\mathbb N$. Define $$\mathds{Q}_r = \alpha_{11}(r)\, \dir{\xi_{11}(r)}+\alpha_{12}(r)\, \dir{\xi_{12}(r)},$$ with $\xi_{11}(r) = \wh{\xi}_1 - {q_{11}(r) \over \alpha_{11}(r)}=0,$ and $\xi_{12}(r) = \wh{\xi}_1 - {q_{12}(r) \over \alpha_{12}(r)}=\varepsilon k$. By Theorem \ref{thm:dual-dual}, the two-point distributions $\mathds{Q}_r$ reside within the Wasserstein ball of radius $\varepsilon$ around $\dir{0}$ and asymptotically attain the supremum in the worst-case expectation problem. However, this sequence has no weak limit as $\xi_{12}(r) = \varepsilon k$ tends to infinity, see Figure~\ref{fig:Ex}. In fact, no single distribution can attain the worst-case expectation. Assume for the sake of contradiction that there exists $\mathds{Q}^\star\in \ball{\dir{0}}{\varepsilon}$ with $\mathds{E}^{\mathds{Q}^\star}[\ell(\xi)]=\varepsilon$. Then, we find $\varepsilon= \mathds{E}^{\mathds{Q}^\star}[\ell(\xi)]< \mathds{E}^{\mathds{Q}^\star}[|\xi|]\leq \varepsilon$, where the strict inequality follows from the relation $\ell(\xi)<|\xi|$ for all $\xi\neq 0$ and the observation that $\mathds{Q}^\star\neq\dir{0}$, \change{while the second inequality follows from Theorem~\ref{thm:KantorovichRubinstein}.} Thus, $\mathds{Q}^\star$ does not exist. \end{Ex} \begin{figure}[t!] \begin{center} \includegraphics[scale = 0.8]{Ex} \end{center} \caption{Example of a worst-case expectation problem without a worst-case distribution} \label{fig:Ex} \end{figure} The existence of a worst-case distribution can, however, be guaranteed in some special cases. \begin{Cor}[Existence of a worst-case distribution] \label{cor:worst-case} Suppose that Assumption \ref{a:ell} holds. If the uncertainty set $\Xi$ is compact or the loss function is concave (i.e., $K=1$), then the sequence $\{\alpha_{ik}(r), \xi_{ik}(r)\}_{r \in \mathbb{N}}$ constructed in Theorem~\ref{thm:dual-dual} has an accumulation point $\{\alpha^\star_{ik}, \xi^\star_{ik}\}$, and $$\mathds{Q}^\star \coloneqq {1 \over N}\sum_{i = 1}^{N}\sum_{k = 1}^{K} \alpha^\star_{ik}\dir{\xi^\star_{ik}}$$ is a worst-case distribution achieving the supremum in \eqref{dist-rob-Ex}. \end{Cor} \begin{proof} If $\Xi$ is compact, then the sequence $\{\alpha_{ik}(r), \xi_{ik}(r)\}_{r \in \mathbb{N}}$ has a converging subsequence with limit $\{\alpha^\star_{ik},\xi^\star_{ik}\}$. Similarly, if $K = 1$, then $\alpha_{i1} = 1$ for all $i\le N$, in which case \eqref{dual-dual} reduces to a convex optimization problem with an upper semicontinuous objective function over a compact feasible set. Hence, its supremum is attained at a point $\{\alpha^\star_{ik},\xi^\star_{ik}\}$. In both cases, Theorem~\ref{thm:dual-dual} guarantees that the distribution $\mathds{Q}^\star$ implied by $\{\alpha^\star_{ik},\xi^\star_{ik}\}$ achieves the supremum in \eqref{dist-rob-Ex}. \end{proof} The worst-case distribution of Corollary~\ref{cor:worst-case} is discrete, and its atoms $\xi^\star_{ik}$ reside in the neighborhood of the given data points $\wh{\xi}_i$. By the constraints of problem~\eqref{dual-dual}, the probability-weighted cumulative distance between the atoms and the respective data points amounts to $$\sum_{i=1}^N\sum_{k=1}^K \alpha_{ik}\| \xi^\star_{ik}-\wh{\xi}_i \| = \sum_{i=1}^N\sum_{k=1}^K \|q_{ik}\|\leq \varepsilon,$$ which is bounded above by the radius of the Wasserstein ball. The fact that the worst-case distribution $\mathds{Q}^\star$ (if it exists) is supported outside of $\wh{\Xi}_N$ is a key feature distinguishing the Wasserstein ball from the ambiguity sets induced by other probability metrics such as the total variation distance or the Kullback-Leibler divergence; see Figure~\ref{fig:balls}. Thus, the worst-case expectation criterion based on Wasserstein balls advocated in this paper should appeal to decision makers who wish to immunize their optimization problems against perturbations of the data points. \begin{figure*} [t] \centering \subfigure[Empirical distribution on a training dataset with $N = 2$ samples]{\label{fig:emp} \includegraphics[width=0.3\columnwidth]{emp}} \hspace{2mm} \subfigure[A representative discrete distribution in the total variation or the Kullback-Leiber ball]{\label{fig:TV-KL} \includegraphics[width=0.31\columnwidth]{TV-KL}} \hspace{2mm} \subfigure[A representative discrete distribution in the Wasserstein ball]{\label{fig:wass} \includegraphics[width=0.31\columnwidth]{wass}} \caption{Representative distributions in balls centered at $\wh{\PP}_N$ induced by different metrics} \label{fig:balls} \end{figure*} \begin{Rem}[Weak coupling] We highlight that the convex program \eqref{dual-dual} is amenable to decomposition and parallelization techniques as the decision variables associated with different sample points are only coupled through the norm constraint. We expect the resulting scenario decomposition to offer a substantial speedup of the solution times for problems involving large datasets. Efficient decomposition algorithms that could be used for solving the convex program \eqref{dual-dual} are described, for example, in \cite{ref:NeaBoyd-14} and \cite[Chapter 4]{ref:Bert-15}. \end{Rem} \section{Special Loss Functions} \label{sec:cases} We now demonstrate that the convex optimization problems \eqref{eq:thm-dual:2} and \eqref{dual-dual} reduce to computationally tractable conic programs for several loss functions of practical interest. \subsection{Piecewise Affine Loss Functions} We first investigate the worst-case expectations of convex and concave piecewise affine loss functions, which arise, for example, in option pricing \cite{ref:BertPop-00}, risk management \cite{ref:NatSimUich-10} and in generic two-stage stochastic programming~\cite{ref:BertDoaVinNat-10}. Moreover, piecewise affine functions frequently serve as approximations of {\em smooth} convex or concave loss functions. \begin{Cor}[Piecewise affine loss functions] \label{cor:affine} Suppose that the uncertainty set is a polytope, that is, $\Xi = \{ \xi \in \mathbb{R}^m : C \xi \le d \}$ where $C$ is a matrix and $d$ a vector of appropriate dimensions. Moreover, consider the affine functions $a_k(\xi) \coloneqq \inner{a_{k}}{\xi} + b_{k}$ for all $k\le K$. \begin{subequations} \label{affine-opt} \begin{itemize} \item[(i)] If $\ell(\xi)= \max_{k\le K}a_k(\xi)$, then the worst-case expectation \eqref{dist-rob-Ex} evaluates to \begin{align} \label{affine-max} \left\{ \begin{array}{clll} \inf\limits_{\lambda,s_i, \gamma_{ik}} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} s_i \\ \normalfont \text{s.t.} & b_k +\inner{a_k}{\wh{\xi}_i}+ \inner{\gamma_{ik}}{d-C\wh{\xi}_i} \le s_i & \forall i \le N, & \forall k \le K\\ & \|C^\intercal\gamma_{ik} - a_{k}\|_* \le \lambda & \forall i \le N, & \forall k \le K \\ & \gamma_{ik} \ge 0& \forall i \le N, & \forall k \le K . \end{array}\right. \end{align} \item[(ii)] If $\ell(\xi)= \min_{k\le K}a_k(\xi)$, then the worst-case expectation \eqref{dist-rob-Ex} evaluates to \begin{align} \label{affine-min} \left\{\begin{array}{clll} \inf\limits_{\lambda,s_i, \gamma_{i},\theta_{i}} \hspace{-1ex}& \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} s_i \\ \normalfont \text{s.t.} &\inner{\theta_i}{b+ A\wh{\xi}_i}+\inner{\gamma_{i}}{d- C\wh{\xi}_i} \le s_i & \forall i \le N\\ & \|C^\intercal \gamma_i-A^\intercal\theta_i\|_* \le \lambda & \forall i \le N \\ & \inner{\theta_{i}}{e} = 1 & \forall i \le N\\ & \gamma_{i}\ge 0& \forall i \le N\\ & \theta_{i} \ge 0& \forall i \le N, \end{array}\right. \end{align} where $A$ is the matrix with rows $a^\intercal_k$, $k\le K$, $b$ is the column vector with entries $b_k$, $k\le K$, and $e$ is the vector of all ones. \end{itemize} \end{subequations} \begin{proof} Assertion~(i) is an immediate consequence of Theorem~\ref{thm:dist-rob-opt}, which applies because $\ell(x)$ is the pointwise maximum of the affine functions $\ell_k(\xi)= a_k(\xi)$, $k\le K$, and thus Assumption~\ref{a:ell} holds for $J= K$. By definition of {the} conjugacy operator, we have \begin{align*} [-\ell_k]^*(z) =[-a_k]^*(z) = \sup\limits_{\xi} \inner{z}{\xi} + \inner{a_k}{\xi} + b_k=\left\{\begin{array}{cl} b_k & \text{if }z=-a_{k}, \\ \infty & \text{else,} \end{array} \right. \end{align*} and \begin{align*} \sigma_\Xi(\nu) = \left\{ \begin{array}{cl} \sup\limits_{\xi} & \inner{\nu}{\xi} \\ \normalfont \text{s.t.} & C \xi \le d \end{array}\right. = \left\{ \begin{array}{cl} \inf\limits_{\gamma\ge 0} & \inner{\gamma}{d} \\ \normalfont \text{s.t.} & C^\intercal \gamma= \nu, \end{array} \right. \end{align*} where the last equality follows from strong duality, which holds as the uncertainty set is non-empty. Assertion~(i) then follows by substituting the above expressions into \eqref{eq:thm-dual:2}. Assertion~(ii) also follows directly from Theorem~\ref{thm:dist-rob-opt} because $\ell(\xi)=\ell_1(\xi)= \min_{k\le K}a_j(\xi)$ is concave and thus satisfies Assumption~\ref{a:ell} for $J=1$. In this setting, we find \begin{align*} [-\ell]^*(z) &= \sup\limits_{\xi} \inner{z}{\xi} + \min_{k\le K}\Big\{ \inner{a_k}{\xi} + b_k\Big\} = \left\{ \begin{array}{cl} \sup\limits_{\xi,\tau} & \inner{z}{\xi} +\tau \\ \normalfont \text{s.t.} & A{\xi} + b \geq \tau e \end{array}\right. = \left\{ \begin{array}{cl} \inf\limits_{\theta \geq 0} & \inner{\theta}{b} \\ \normalfont \text{s.t.} & A^\intercal \theta = -z \\ & \inner{\theta}{e} = 1 \end{array} \right. \end{align*} where the last equality follows again from strong linear programming duality, which holds since the primal maximization problem is feasible. Assertion~(ii) then follows by substituting $[-\ell]^*$ as well as the formula for $\sigma_\Xi$ from the proof of assertion~(i) into \eqref{eq:thm-dual:2}. \end{proof} \end{Cor} As a consistency check, we ascertain that in the {\em ambiguity-free limit}, the optimal value of \eqref{affine-max} reduces to the expectation of $\max_{k\le K}a_k(\xi)$ under the empirical distribution. Indeed, for $\varepsilon = 0$, the variable $\lambda$ can be set to any positive value at no penalty. For this reason and because all training samples must belong to the uncertainty set ({\em i.e.}, $d-C\wh{\xi}_i\geq 0$ for all $i\le N$), it is optimal to set $\gamma_{ik}=0$. This in turn implies that $s_i= \max_{k\le K}a_k(\wh{\xi}_i)$ at optimality, in which case $\frac{1}{N}\sum_{i=1}^Ns_i$ represents the sample average of the convex loss function at hand. An analogous argument shows that, for $\varepsilon=0$, the optimal value of \eqref{affine-min} reduces to the expectation of $\min_{k\le K}a_k(\xi)$ under the empirical distribution. As before, $\lambda$ can be increased at no penalty. Thus, we conclude that $\gamma_i=0$ and \[ s_i=\min\limits_{\theta_i\geq 0}\left\{\inner{\theta_i}{b+ A\wh{\xi}_i}:\inner{\theta_{i}}{e} = 1 \right\} = \min_{k\le K}a_k(\wh{\xi}_i) \] at optimality, in which case $\frac{1}{N}\sum_{i=1}^Ns_i$ is the sample average of the given concave loss function. \subsection{Uncertainty Quantification} \label{subsec:UQ} A problem of great practical interest is to ascertain whether a physical, economic or engineering system with an uncertain state $\xi$ satisfies a number of safety constraints with high probability. In the following we denote by $\set{A}$ the set of states in which the system is safe. Our goal is to quantify the probability of the event $\xi\in\set A$ ($\xi\notin\set A$) under an ambiguous state distribution that is only indirectly observable through a finite training dataset. More precisely, we aim to calculate the {\em worst-case} probability of the system being {\em unsafe}, {\em i.e.}, \begin{subequations} \label{UQ} \begin{align} \label{worst-case-prob} \sup_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{Q} \left[ \xi\notin \set{A}\right], \end{align} as well as the {\em best-case} probability of the system being {\em safe}, that is, \begin{align} \label{best-case-prob} \sup_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{Q} \left[ \xi\in \set{A}\right]. \end{align} \end{subequations} \change{ \begin{Rem}[Data-dependent sets] The set $\set A$ may even depend on the samples $\wh{\xi}_1,\ldots,\wh{\xi}_N$, in which case $\set A$ is renamed as $\widehat{\set A}$. If the Wasserstein radius $\varepsilon$ is set to $\varepsilon_N(\beta)$, then we have $\mathds{P}\in \ball{\wh{\PP}_N}{\varepsilon}$ with probability $1-\beta$, implying that \eqref{worst-case-prob} and \eqref{best-case-prob} still provide $1-\beta$ confidence bounds on $\mathds{P}[\xi\notin\widehat{\set A}]$ and $\mathds{P}[\xi\in\widehat{\set A}]$, respectively. \end{Rem}} \begin{Cor}[Uncertainty quantification] \label{cor:chance} Suppose that the uncertainty set is a polytope of the form $\Xi = \{ \xi \in \mathbb{R}^m : C \xi \le d \}$ as in Corollary~\ref{cor:affine}. \begin{subequations} \label{chance} \begin{itemize} \item[(i)] If $\set A = \{\xi \in \mathbb{R}^m: A\xi < b\}$ is an open polytope and the halfspace $\big\{\xi:\inner{a_k}{\xi}\geq b_k \big\}$ has a nonempty intersection with $\Xi$ for any $k\le K$, where $a_k$ is the $k$-th row of the matrix $A$ and $b_k$ is the $k$-th entry of the vector $b$, then the worst-case probability \eqref{worst-case-prob} is given by \begin{align} \label{chance-worst} \left\{ \begin{array}{clll} \inf\limits_{\lambda,s_i, \gamma_{ik},\theta_{ik}} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} s_i \\ \normalfont \text{s.t.} &1-\theta_{ik}\big(b_k-\inner{a_k}{\wh{\xi}_i} \big) +\inner{\gamma_{ik}}{d-C\wh{\xi}_i} \le s_i & \forall i \le N, & \forall k \le K\\ & \|a_k\theta_{ik}-C^\intercal\gamma_{ik}\|_* \le \lambda & \forall i \le N, & \forall k \le K \\ & \gamma_{ik}\ge 0& \forall i \le N, & \forall k \le K\\ & \theta_{ik} \ge 0& \forall i \le N, & \forall k \le K\\ & s_i \ge 0 & \forall i \le N. \end{array}\right. \end{align} \item[(ii)] If $\set A = \{\xi \in \mathbb{R}^m : A\xi \le b\}$ is a closed polytope that has a nonempty intersection with $\Xi$, then the best-case probability \eqref{best-case-prob} is given by \begin{align} \label{chance-best} \left\{\begin{array}{clll} \inf\limits_{\lambda,s_i, \gamma_i, \theta_i} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} s_i \vspace{1mm} \\ \normalfont \text{s.t.} & 1+\inner{\theta_i}{b - A\wh{\xi}_i} + \inner{\gamma_{i}}{d - C\wh{\xi}_i} \le s_i & \forall i \le N \\ & \|A^\intercal \theta_i+C^\intercal \gamma_{i}\|_* \le \lambda & \forall i \le N \\ & \gamma_i \ge 0 & \forall i \le N\\ & \theta_{i} \ge 0 & \forall i \le N\\ & s_i\ge 0 & \forall i \le N. \end{array}\right. \end{align} \end{itemize} \end{subequations} \begin{proof} {The uncertainty quantification problems \eqref{worst-case-prob} and \eqref{best-case-prob} can be interpreted as instances of \eqref{dist-rob-Ex} with loss functions $\ell = 1 - \ind{\set A}$ and $\ell = \ind{\set A}$, respectively. In order to be able to apply Theorem~\ref{thm:dist-rob-opt}, we should represent these loss functions as finite maxima of concave functions as shown in Figure \ref{fig:ind}.} Formally, assertion~(i) follows from Theorem~\ref{thm:dist-rob-opt} for a loss function with $K+1$ pieces if we use the following definitions. For every $k\le K$ we define $$\ell_{k}(\xi) = \left\{\begin{array}{cl} 1 & \text{if }\inner{a_k}{\xi} \ge b_k, \\ -\infty & \text{otherwise.} \end{array}\right.$$ Moreover, we define $\ell_{K+1}(\xi) = 0$. {As illustrated in Figure~\ref{fig:ind:out}}, we thus have $\ell(\xi)=\max_{k\le K+1} \ell_k(\xi)= 1 - \ind{\set A}(\xi)$ and \[ \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{Q} \left[ \xi\notin \set{A}\right]~= \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{E}^\mathds{Q} \left[\ell(\xi)\right]. \] Assumption~\ref{a:ell} holds due to the postulated properties of $\set A$ and $\Xi$. In order to apply Theorem~\ref{thm:dist-rob-opt}, we must determine the support function $\sigma_\Xi$, which is already known from Corollary~\ref{cor:affine}, as well as the conjugate functions of $-\ell_k$, $k\le K+1$. A standard duality argument yields \begin{align*} [-\ell_k]^*(z) & = \left\{ \begin{array}{cl} \sup\limits_{\xi} & \inner{z}{\xi} + 1 \\ \normalfont \text{s.t.} & \inner{a_k}{\xi}\geq b_k \end{array}\right. = \left\{ \begin{array}{cl} \inf\limits_{\theta \ge 0} & 1 - b_k\theta \\ \normalfont \text{s.t.} & a_k \theta =-z, \end{array}\right. \end{align*} for all $k\le K$. Moreover, we have {$[-\ell_{K+1}]^* = 0$ if $\xi=0$; $=\infty$ otherwise}. Assertion~(ii) then follows by substituting the formulas for $[-\ell_k]^*$, $k\le K+1$, and $\sigma_\Xi$ into \eqref{eq:thm-dual:2}. Assertion~(ii) follows from Theorem~\ref{thm:dist-rob-opt} by setting $K= 2$, $\ell_1(\xi) = 1-\chi_{\set A}(\xi)$ and $\ell_2(\xi) = 0$. As illustrated in {Figure~\ref{fig:ind:in}}, this implies that $\ell(\xi)=\max\{\ell_1(\xi),\ell_2(\xi)\}=\ind{\set A}(\xi)$ and \[ \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{Q} \left[ \xi\in \set{A}\right]~= \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{E}^\mathds{Q} \left[\ell(\xi)\right]. \] Assumption~\ref{a:ell} holds by our assumptions on $\set A$ and $\Xi$. In order to apply Theorem~\ref{thm:dist-rob-opt}, we thus have to determine the support function $\sigma_\Xi$, which was already calculated in Corollary~\ref{cor:affine}, and the conjugate functions of $-\ell_1$ and $-\ell_2$. By the definition of the conjugacy operator, we find { \begin{align*} [-\ell_1]^*(z) &= \sup_{\xi \in \set{A}} \inner{z}{\xi} + 1 = \left\{ \begin{array}{cl} \sup\limits_{\xi} & \inner{z}{\xi} + 1 \\ \normalfont \text{s.t.} & A\xi \le b \end{array}\right. = \left\{ \begin{array}{cl} \inf\limits_{\theta_k \ge 0} & \inner{\theta}{b} + 1 \\ \normalfont \text{s.t.} & A^\intercal \theta = z \end{array}\right. \end{align*} }where the last equality follows from strong linear programming duality, which holds as the safe set is non-empty. Similarly, we find {$[-\ell_{2}]^* = 0$ if $\xi=0$; $=\infty$ otherwise}. Assertion~(ii) then follows by substituting the above expressions into \eqref{eq:thm-dual:2}. \end{proof} \end{Cor} \begin{figure}[t!] \centering \subfigure[Indicator function of the unsafe set]{\label{fig:ind:out}\includegraphics[scale = 0.55]{ind-out}} \qquad \subfigure[Indicator function of the safe set]{\label{fig:ind:in}\includegraphics[scale = 0.55]{ind-in}} \caption{Representing the indicator function of {a convex set and its complement as a pointwise maximum of concave functions}} \label{fig:ind} \end{figure} In the \emph{ambiguity-free limit} ({\em i.e.}, for $\varepsilon = 0$) the optimal value of \eqref{chance-worst} reduces to the fraction of training samples residing outside of the open polytope $\set A=\{\xi:A\xi <b\}$. Indeed, in this case the variable $\lambda$ can be set to any positive value at no penalty. For this reason and because all training samples belong to the uncertainty set ({\em i.e.}, $d-C\wh{\xi}_i\geq 0$ for all $i\le N$), it is optimal to set $\gamma_{ik}=0$. If the $i$-th training sample belongs to $\set A$ ({\em i.e.}, $b_k-\inner{a_k}{\wh{\xi}_i}> 0$ for all $k\le K$), then $\theta_{ik}\geq1/(b_k-\inner{a_k}{\wh{\xi}_i})$ for all $k\le K$ and $s_i=0$ at optimality. Conversely, if the $i$-th training sample belongs to the complement of $\set A$, ({\em i.e.}, $b_k-\inner{a_k}{\wh{\xi}_i}\leq 0$ for some $k\le K$), then $\theta_{ik}=0$ for some $k\le K$ and $s_i=1$ at optimality. Thus, $\sum_{i=1}^Ns_i$ coincides with the number of training samples outside of $\set A$ at optimality. An analogous argument shows that, for $\varepsilon=0$, the optimal value of \eqref{chance-best} reduces to the fraction of training samples residing inside of the closed polytope $\set A=\{\xi:A\xi \leq b\}$. \subsection{Two-Stage Stochastic Programming} A major challenge in linear two-stage stochastic programming is to evaluate the expected recourse costs, which are only implicitly defined as the optimal value of a linear program whose coefficients depend linearly on the uncertain problem parameters \cite[Section~2.1]{ref:Shap&Dent&Rusz}. The following corollary shows how we can evaluate the worst-case expectation of the recourse costs with respect to an ambiguous parameter distribution that is only observable through a finite training dataset. For ease of notation and without loss of generality, we suppress here any dependence on the first-stage decisions. \begin{Cor}[Two-stage stochastic programming] \label{cor:2stage} Suppose that the uncertainty set is a polytope of the form $\Xi = \{ \xi \in \mathbb{R}^m : C \xi \le d \}$ as in Corollaries~\ref{cor:affine} and \ref{cor:chance}. \begin{subequations} \label{2stage} \begin{itemize} \item[(i)] If $\ell(\xi) =\inf_{y} \left\{ \inner{y}{Q\xi} : Wy\ge h \right\}$ is the optimal value of a parametric linear program with objective uncertainty, and if the feasible set $\{y:Wy\ge h\}$ is non-empty and compact, then the worst-case expectation \eqref{dist-rob-Ex} is given by \begin{align} \label{objective-uncertainty} \left\{ \begin{array}{clll} \inf\limits_{\lambda,s_i, \gamma_i, y_i} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} s_i \vspace{1mm} \\ \normalfont \text{s.t.} & \inner{y_i}{Q\wh{\xi}_i} + \inner{\gamma_{i}}{d - C\wh{\xi}_i} \le s_i & \forall i \le N \\ & Wy_i\ge h & \forall i \le N\\ & \|Q^\intercal y_i-C^\intercal \gamma_{i}\|_* \le \lambda & \forall i \le N \\ & \gamma_i \ge 0 & \forall i \le N. \end{array}\right. \end{align} \item[(ii)] If $\ell(\xi) =\inf_{y} \left\{ \inner{q}{y} : Wy \ge H\xi + h \right\}$ is the optimal value of a parametric linear program with right-hand side uncertainty, and if the dual feasible set $\{\theta\ge 0:W^\intercal\theta=q\}$ is non-empty and compact with vertices $v_k$, $k\le K$, {then} the worst-case expectation \eqref{dist-rob-Ex} is given~by \begin{align} \label{rhs-uncertainty} \left\{ \begin{array}{clll} \inf\limits_{\lambda,s_i, \gamma_{ik}} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} s_i \\ \normalfont \text{s.t.} & \inner{v_k}{h} + \inner{H^\intercal v_k}{\wh{\xi}_i}+ \inner{\gamma_{ik}}{d-C\wh{\xi}_i} \le s_i & \forall i \le N, & \forall k \le K\\ & \|C^\intercal\gamma_{ik}-a_{ik}\|_* \le \lambda & \forall i \le N, & \forall k \le K \\ & \gamma_{ik} \ge 0& \forall i \le N, & \forall k \le K. \end{array}\right. \end{align} \end{itemize} \end{subequations} \begin{proof} Assertion~(i) follows directly from Theorem~\ref{thm:dist-rob-opt} because $\ell(\xi)$ is concave as an infimum of linear functions in $\xi$. Indeed, the compactness of the feasible set $\{y: Wy\ge h\}$ ensures that Assumption~\ref{a:ell} holds for $K=1$. In this setting, we find \begin{align*} [-\ell]^*(z)& = \sup\limits_{\xi} \left\{ \inner{z}{\xi} + \inf\limits_{y} \left\{ \inner{y}{Q\xi} : Wy\ge h\right\}\right\} \\ &= \inf\limits_{y}\left\{ \sup\limits_{\xi}\left\{ \inner{z+Q^\intercal y}{\xi} \right\} : Wy\ge h \right\}\\ &= \left\{\begin{array}{cl} 0 & \text{if there exists $y$ with } Q^\intercal y=-z \text{ and }Wy\ge h,\\ \infty & \text{otherwise,} \end{array} \right. \end{align*} where the second equality follows from the classical minimax theorem \cite[Proposition 5.5.4]{ref:Bert-09}, which applies because $\{y: Wy\ge h\}$ is compact. Assertion~(i) then follows by substituting $[-\ell]^*$ as well as the formula for $\sigma_\Xi$ from Corollary~\ref{cor:affine} into \eqref{eq:thm-dual:2}. Assertion~(ii) relies on the following reformulation of the loss function, \begin{align*} \ell(\xi) &= \left\{ \begin{array}{cl} \inf\limits_{y} & \inner{q}{y} \\ \normalfont \text{s.t.} & Wy \ge H\xi + h \end{array}\right. = \left\{ \begin{array}{cl} \sup\limits_{\theta\geq 0} & \inner{\theta}{H\xi + h} \\ \normalfont \text{s.t.} & W^\intercal \theta = q \end{array}\right. = \max\limits_{k\le K} \inner{v_k}{H\xi + h} \\ &= \max\limits_{k\le K} \inner{H^\intercal v_k}{\xi} + \inner{v_k}{h}, \end{align*} where the first equality holds due to strong linear programming duality, which applies as the dual feasible set is non-empty. The second equality exploits the elementary observation that the optimal value of a linear program with non-empty, compact feasible set is always adopted at a vertex. As we managed to express $\ell(\xi)$ as a pointwise maximum of linear functions, assertion~(ii) follows immediately from Corllary~\ref{cor:affine}~(i). \end{proof} \end{Cor} As expected, in the {\em ambiguity-free limit}, problem~\eqref{objective-uncertainty} reduces to a standard SAA problem. Indeed, for $\varepsilon=0$, the variable $\lambda$ can be made large at no penalty, and thus $\gamma_i=0$ and $s_i=\inner{y_i}{Q\wh{\xi}_i}$ at optimality. In this case, problem~\eqref{objective-uncertainty} is equivalent to \begin{align*} \inf\limits_{y_i} \left\{ {1 \over N}\sum\limits_{i = 1}^{N} \inner{y_i}{Q\wh{\xi}_i} : Wy_i\ge h \quad\forall i \le N\right\}. \end{align*} Similarly, one can verify that for $\varepsilon=0$, \eqref{rhs-uncertainty} reduces to the SAA problem \begin{align*} \inf\limits_{y_i} \left\{ {1 \over N}\sum\limits_{i = 1}^{N} \inner{y_i}{q} : Wy_i\ge H\wh{\xi}_i \quad\forall i \le N\right\}. \end{align*} We close this section with a remark on the computational complexity of all the {convex optimization problems derived} in this section. \begin{Rem}[Computational tractability] \label{rem:comp} \hfil \begin{itemize} \item If the Wasserstein metric is defined in terms of the $1$-norm (i.e., $\|\xi\|=\sum_{k=1}^m|\xi_k|$) or the $\infty$-norm (i.e., $\|\xi\|=\max_{k\le m}|\xi_k|$), then the optimization problems \eqref{affine-max}, \eqref{affine-min}, \eqref{chance-worst}, \eqref{chance-best}, \eqref{objective-uncertainty} and \eqref{rhs-uncertainty} all reduce to linear programs whose sizes scale with the number $N$ of data points and the number $J$ of affine pieces of the underlying loss functions. \item Except for the two-stage stochastic program with right-hand side uncertainty in \eqref{rhs-uncertainty}, the resulting linear programs scale polynomially in the problem description and are therefore computationally tractable. As the number of vertices $v_k$, $k\le K$, of the polytope $\{\theta\ge 0:W^\intercal\theta=q\}$ may be exponential in the number of its facets, however, the linear program \eqref{rhs-uncertainty} has generically exponential size. \item \change{Inspecting~\eqref{affine-max}, one easily verifies that the distributionally robust optimization problem~\eqref{DRO} reduces to a finite convex program if $\mathbb{X}$ is convex and $h(x,\xi)= \max_{k\le K} \inner{a_{k}(x)}{\xi} + b_{k}(x)$, while the gradients $a_{k}(x)$ and the intercepts $b_{k}(x)$ depend linearly on $x$. Similarly, \eqref{DRO} can be reformulated as a finite convex program if $\mathbb{X}$ is convex and $h(x,\xi)=\inf_{y} \left\{ \inner{y}{Q\xi} : Wy\ge h(x) \right\}$ or $h(x,\xi)=\inf_{y} \left\{ \inner{q}{y} : Wy \ge H(x)\xi + h(x) \right\}$, while the right hand side coefficients $h(x)$ and $H(x)$ depend linearly on $x$; see \eqref{objective-uncertainty} and \eqref{rhs-uncertainty}, respectively. In contrast, problems \eqref{affine-min}, \eqref{chance-worst} and \eqref{chance-best} result in non-convex optimization problems when their data depends on $x$.} \item We emphasize that the computational complexity of all convex programs examined in this section is independent of the radius $\varepsilon$ of the Wasserstein ball. \end{itemize} \end{Rem} \section{Tractable Extensions} \label{sec:exten} We now demonstrate that through minor modifications of the proofs, Theorems~\ref{thm:dist-rob-opt} and \ref{thm:dual-dual} extend to worst-case expectation problems involving even richer classes of loss functions. First, we investigate problems where the uncertainty can be viewed as a stochastic process and where the loss function is additively separable. Next, we study problems whose loss functions are convex in the uncertain variables {and are therefore} not necessarily representable as finite maxima of concave functions as postulated by Assumption~\ref{a:ell}. \subsection{Stochastic Processes with a Separable Cost} Consider a variant of the worst-case expectation problem~\eqref{dist-rob-Ex}, where the uncertain parameters can be interpreted as a stochastic process {$\xi = \big(\xi_1,\ldots,\xi_T\big)$}, and assume that $\xi_t \in \Xi_t$, where $ \Xi_t \subseteq \mathbb{R}^m$ is non-empty and closed for any $t\le T$. Moreover, assume that the loss function is additively separable with respect to the temporal structure of $\xi$, that is, \begin{align} \label{proc_cost} \ell(\xi) \coloneqq \sum\limits_{t = 1}^{T} \max_{k\le K}\ell_{tk} \big(\xi_t\big), \end{align} where $\ell_{tk}:\mathbb{R}^m\rightarrow \overline {\R}$ is a measurable function for any $k\le K$ and $t\le T$. Such loss functions appear, for instance, in open-loop stochastic optimal control or in multi-item newsvendor problems. Consider a process norm $\normT{\xi} = \sum_{t = 1}^{T} \|\xi_t\|$ associated with the base norm $\|\cdot\|$ on $\mathbb{R}^m$, and assume that its induced metric is the one used in the definition of the Wasserstein distance. Note that if $\|\cdot\|$ is the 1-norm on $\mathbb{R}^m$, then $\normT{\cdot}$ reduces to the 1-norm on $\mathbb{R}^{mT}$. By interchanging summation and maximization, the loss function~\eqref{proc_cost} can be re-expressed as \begin{align*} \ell(\xi)= \max_{k_t \le K} \sum\limits_{t = 1}^{T} \ell_{tk_t} \big(\xi_t \big), \end{align*} where the maximum runs over all $K^T$ combinations of $k_1,\ldots, k_T\le K$. Under this representation, Theorem~\ref{thm:dist-rob-opt} remains applicable. However, the resulting convex optimization problem would involve $\mathcal O(K^T)$ decision variables and constraints, indicating that an efficient solution may not be available. Fortunately, this deficiency can be overcome by modifying Theorem~\ref{thm:dist-rob-opt}. \begin{Thm}[Convex reduction for separable loss functions] \label{thm:dist-rob-opt-separable} Assume that the loss function $\ell$ is of the form \eqref{proc_cost}, and the Wasserstein ball is defined through the process norm $\normT{\cdot}$. Then, for any $\varepsilon \ge0 $, the worst-case expectation~\eqref{dist-rob-Ex} is smaller or equal to the optimal value of the finite convex program \begin{align} \label{eq:thm-dual_process:1} \left\{ \begin{array}{cllll} \inf\limits_{\lambda, s_{ti}, z_{tik}, \nu_{tik}} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} \sum\limits_{t = 1}^{T} s_{ti} && \\ \normalfont \text{s.t.} & [-\ell_{tk}]^*\big(z_{tik} - \nu_{tik}\big) + \sigma_{\Xi_t}(\nu_{tik}) - \inner{z_{tik}}{\wh{\xi}_{ti}} \le s_{ti} & \forall i \le N, & \forall k\le K, & \forall t \le T,\\ & \|z_{tik}\|_* \le \lambda &\forall i \le N, & \forall k\le K, & \forall t \le T. \end{array}\right. \end{align} If $\Xi_t$ and $\{\ell_{tk}\}_{k\le K}$ satisfy the convexity Assumption~\ref{a:ell} for every $t\le T$, then the worst-case expectation~\eqref{dist-rob-Ex} coincides exactly with the optimal value of problem~\eqref{eq:thm-dual_process:1}. \end{Thm} \begin{proof} Up until equation~\eqref{eq:2}, the proof of Theorem~\ref{thm:dist-rob-opt-separable} parallels that of Theorem~\ref{thm:dist-rob-opt}. Starting from \eqref{eq:2}, we then have \begin{align} \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} & \mathds{E}^\mathds{Q} \big[ \ell(\xi) \big] = \inf\limits_{\lambda \ge 0}~ \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} \sup_{\xi} \left (\ell(\xi) - \lambda \normT{\xi - \wh{\xi}_i} \right ) \notag \\ & = \inf\limits_{\lambda \ge 0} ~ \lambda \varepsilon + {1 \over N} \sum\limits_{i = 1}^{N} \sum\limits_{t = 1}^{T} \sup_{\xi_t \in \Xi_t} \left (\max_{k\le K} \ell_{tk}\big(\xi_t \big) - \lambda \big \|\xi_t - \wh{\xi}_{ti}\big\| \right ) ,\notag \end{align} where the interchange of the summation and the maximization is facilitated by the separability of the overall loss function. Introducing epigraphical auxiliary variables yields \begin{align} &\left\{ \begin{array}{cllll} \inf\limits_{\lambda, s_{ti}} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} \sum\limits_{t=1}^{T} s_{ti} \\ \normalfont \text{s.t.} & \sup\limits_{\xi_t\in \Xi_t} \Big(\ell_{tk}\big(\xi_t\big) - \lambda \big\|\xi_t - \wh{\xi}_{ti} \big\| \Big)\le s_{ti} & \forall i\le N, ~ \forall k\le K, ~ \forall t \le T \\ & \lambda \ge 0 \end{array} \right. \notag \\ \le& \left\{ \begin{array}{cllll} \inf\limits_{\lambda, s_{ti},z_{tik}} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} \sum\limits_{t = 1}^{T}s_{ti} \\ \normalfont \text{s.t.} & \sup\limits_{\xi_t\in \Xi_t} \Big(\ell_{tk}\big(\xi_t \big) - \inner{z_{tik}}{\xi_t} \Big) + \inner{z_{tik}}{\wh{\xi}_{ti}}\le s_{ti} &\forall i\le N, ~\forall k\le K, ~\forall t \le T \\ & \|z_{tik}\|_* \le \lambda & \forall i\le N,~ \forall k\le K, ~ \forall t \le T \end{array} \right. \notag \\ = & \left\{ \begin{array}{cllll} \inf\limits_{\lambda, s_{ti},z_{tik}} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} \sum\limits_{t = 1}^{T} s_{ti} \\ \normalfont \text{s.t.} & [-\ell_{tk} + \indI{\Xi_t}]^*\big(-z_{tik}\big) + \inner{z_{tik}}{\wh{\xi}_{ti}}\le s_{ti} &\forall i\le N,~ \forall k\le K,~ \forall t \le T \\ & \|z_{tik}\|_* \le \lambda & \forall i\le N, ~ \forall k\le K, ~ \forall t \le T, \end{array} \right. \notag \end{align} where the inequality is justified in a similar manner as the one in \eqref{eq:3}, and it holds as an equality provided that $\Xi_t$ and $\{\ell_{tk}\}_{k\le K}$ satisfy Assumption~\ref{a:ell} for all $t \le T$. {Finally, by \cite[Theorem~11.23(a), p.~493]{ref:Rockafellar-10}, the conjugate of $-\ell_{tk} + \indI{\Xi_t}$ can be replaced by the inf-convolution of the conjugates of $-\ell_{tk}$ and $\indI{\Xi_t}$. This completes the proof.} \end{proof} Note that the convex program~\eqref{eq:thm-dual_process:1} involves only $\mathcal{O}(NKT)$ decision variables and constraints. Moreover, if $\ell_{tk}$ is affine for every $t\le T$ and $k\le K$, while $\|\cdot\|$ represents the $1$-norm or the $\infty$-norm on $\mathbb{R}^m$, then \eqref{eq:thm-dual_process:1} reduces to a tractable linear program (see also Remark~\ref{rem:comp}). A natural generalization of Theorem \ref{thm:dual-dual} further allows us to characterize the extremal distributions of the worst-case expectation problem~\eqref{dist-rob-Ex} with a separable loss function of the form \eqref{proc_cost}. \begin{Thm}[Worst-case distributions for separable loss functions] \label{prop:process-dual-dual-separable} Assume that the loss function $\ell$ is of the form \eqref{proc_cost}, and the Wasserstein ball is defined through the process norm $\normT{\cdot}$. If $\Xi_t$ and $\{\ell_{tk}\}_{k\le K}$ satisfy Assumption~\ref{a:ell} for all $t \le T$, then the worst-case expectation~\eqref{dist-rob-Ex} coincides with the optimal value of the finite convex program \begin{align} \label{dual-dual-separable} \left\{ \begin{array}{clll} \sup\limits_{\alpha_{tik}, q_{tik}} & {1 \over N} \sum\limits_{i = 1}^{N} \sum\limits_{k = 1}^{K} \sum\limits_{t=1}^{T} \alpha_{tik}\ell_{tk}\Big( \wh{\xi}_{ti} - {q_{tik} \over \alpha_{tik}}\Big) \\ \normalfont \text{s.t.} & {1 \over N}\sum\limits_{i =1}^{N} \sum\limits_{k =1}^{K} \sum\limits_{t=1}^{T} \|q_{tik}\| \le \varepsilon\\ & \sum\limits_{k = 1}^{K} \alpha_{tik} = 1 &\forall i \le N, \quad \forall t \le T\\ & \alpha_{tik} \ge 0 &\forall i \le N, \quad \forall t \le T, \quad \forall k\le K \\ & \wh{\xi}_{ti} - {q_{tik} \over \alpha_{tik}} \in \Xi_t &\forall i \le N, \quad \forall t \le T, \quad \forall k\le K \end{array} \right. \end{align} irrespective of $\varepsilon \ge0 $. Let $\big\{\alpha_{tik}(r), q_{tik}(r)\big\}_{r \in \mathbb{N}}$ be a sequence {of} feasible decisions whose objective values converge to the supremum of \eqref{dual-dual-separable}. Then, the discrete (product) probability distributions \begin{align*} \mathds{Q}_r \coloneqq {1 \over N} \sum_{i = 1}^{N} \bigotimes_{t=1}^T \Big(\sum_{k = 1}^{K} \alpha_{tik}(r) \dir{\xi_{tik}(r)}\Big) \quad\mbox{with}\quad \xi_{tik}(r) \coloneqq \wh{\xi}_{ti} - {q_{tik}(r) \over \alpha_{tik}(r)} \end{align*} belong to the Wasserstein ball $\ball{\wh{\PP}_N}{\varepsilon}$ and attain the supremum of \eqref{dist-rob-Ex} asymptotically, i.e., \begin{align*} \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{E}^\mathds{Q} \big[ \ell(\xi) \big] = \lim\limits_{r \rightarrow \infty} \mathds{E}^{\mathds{Q}_r} \big[ \ell(\xi) \big] = \lim\limits_{r \rightarrow \infty} {1 \over N} \sum\limits_{i = 1}^{N} \sum\limits_{k=1}^{K} \sum\limits_{t=1}^{T} \alpha_{tik}(r) \ell_{tk}\big(\xi_{tik}(r)\big) . \end{align*} \end{Thm} \begin{proof} As in the proof of Theorem~\ref{thm:dual-dual}, the claim follows by dualizing the convex program~\eqref{eq:thm-dual_process:1}. Details are omitted for brevity of exposition. \end{proof} We emphasize that the distributions $\mathds{Q}_r$ from Theorem~\ref{prop:process-dual-dual-separable} can be constructed efficiently by solving a convex program of polynomial size even though they have $NK^T$ discretization points. \subsection{Convex Loss Functions} Consider now another variant of the worst-case expectation problem~\eqref{dist-rob-Ex}, where the loss function $\ell$ is proper, convex and lower semicontinuous. Unless $\ell$ is piecewise affine, we cannot represent such a loss function as a pointwise maximum of finitely many concave functions, and thus Theorem~\ref{thm:dist-rob-opt} may only provide a loose upper bound on the worst-case expectation~\eqref{dist-rob-Ex}. The following theorem provides an alternative upper bound that admits new insights into distributionally robust optimization with Wasserstein balls and becomes exact for $\Xi=\mathbb{R}^m$. \begin{Thm}[Convex reduction for convex loss functions] \label{thm:convex} Assume that the loss function $\ell$ is proper, convex, and lower semicontinuous, and define $\kappa \coloneqq \sup\big\{ \|\theta\|_* : \ell^*(\theta) < \infty \big \}$. Then, for any $\varepsilon \ge0 $, the worst-case expectation~\eqref{dist-rob-Ex} is smaller or equal to \begin{align} \label{kappa_bound} \kappa \varepsilon + {1 \over N}\sum_{i = 1}^{N} \ell(\wh{\xi}_i). \end{align} If $\Xi=\mathbb{R}^m$, then the worst-case expectation~\eqref{dist-rob-Ex} coincides exactly with \eqref{kappa_bound}. \end{Thm} \begin{Rem}[Radius of effective domain] \label{rem:kappa} The parameter $\kappa$ can be viewed as the radius of the smallest ball containing the effective domain of the conjugate function $\ell^*$ in terms of the dual norm. By the standard conventions of extended arithmetic, the term $\kappa\varepsilon$ in \eqref{kappa_bound} is interpreted as $0$ if $\kappa=\infty$ and $\varepsilon=0$. \end{Rem} \begin{proof} Equation~\eqref{eq:1-} in the proof of Theorem \ref{thm:dist-rob-opt} implies that \begin{align} \label{convex_loss} \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{E}^\mathds{Q} \big[ \ell(\xi) \big] = \inf\limits_{\lambda \ge 0} ~\lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} \sup_{\xi \in \Xi} \left (\ell(\xi) - \lambda \|\xi - \wh{\xi}_i\| \right ) \end{align} for every $\varepsilon > 0$. As $\ell$ is proper, convex, and lower semicontinuous, it coincides with its bi-conjugate function $\ell^{**}$, see {\em e.g.}\ \cite[Proposition 1.6.1(c)]{ref:Bert-09}. Thus, we may write \begin{align*} \ell(\xi) = \sup_{\theta \in \Theta} \inner{\theta}{\xi} - \ell^*(\theta), \end{align*} where $\Theta \coloneqq \{\theta\in\mathbb{R}^m : \ell^*(\theta) < \infty\}$ denotes the effective domain of the conjugate function $\ell^*$. Using this dual representation of $\ell$ in conjunction with the definition of the dual norm, we find \begin{align*} \sup\limits_{\xi\in \Xi} \Big(\ell(\xi) - \lambda \|\xi-\wh{\xi}_i\|\Big) &= \sup\limits_{\xi\in \Xi}~ \sup\limits_{\theta\in \Theta} \Big(\inner{\theta}{\xi} - \ell^*(\theta) - \lambda \|\xi-\wh{\xi}_i\|\Big) \\ & = \sup\limits_{\xi\in \Xi}~ \sup\limits_{\theta\in \Theta} \inf\limits_{\|z\|_* \le \lambda}\Big(\inner{\theta}{\xi} - \ell^*(\theta) + \inner{z}{\xi} - \inner{z}{\wh{\xi}_i}\Big). \end{align*} The classical minimax theorem \cite[Proposition 5.5.4]{ref:Bert-09} then allows us to interchange the maximization over $\xi$ with the maximization over $\theta$ and the minimization over $z$ to obtain \begin{align} \notag \sup\limits_{\xi\in \Xi} \Big(\ell(\xi) - \lambda \|\xi-\wh{\xi}_i\|\Big) & = \sup\limits_{\theta\in \Theta} \inf\limits_{\|z\|_* \le \lambda} \sup\limits_{\xi\in \Xi}\Big(\inner{\theta + z}{\xi} - \ell^*(\theta) - \inner{z}{\wh{\xi}_i}\Big) \\ \label{sigma} & = \sup\limits_{\theta\in \Theta} \inf\limits_{\|z\|_* \le \lambda} \sigma_{\Xi}(\theta + z) - \ell^*(\theta) - \inner{z}{\wh{\xi}_i}. \end{align} Recall that $\sigma_\Xi$ denotes the support function of $\Xi$. It seems that there is no simple exact reformulation of \eqref{sigma} for arbitrary convex uncertainty sets $\Xi$. Interchanging the maximization over $\theta$ with the minimization over $z$ in \eqref{sigma} would lead to the conservative upper bound {of Corollary \ref{cor:approx}}. Here, however, we employ an alternative approximation. By definition of the support function, we have $\sigma_\Xi\leq \sigma_{\mathbb{R}^m} = \indI{\{0\}}$. Replacing $\sigma_\Xi$ with $ \indI{\{0\}}$ in \eqref{sigma} thus results in the conservative approximation \begin{align} \label{eq:ell-convex} \sup\limits_{\xi\in \mathbb{R}^m} \Big(\ell(\xi) - \lambda \|\xi-\wh{\xi}_i\|\Big) &\leq \left \{ \begin{array}{cl} \ell(\wh{\xi}_i) & \mbox{if }\sup \big\{\|\theta\|_* : \theta \in \Theta \big \} \le \lambda, \\ \infty & \mbox{otherwise.} \end{array} \right. \end{align} The inequality \eqref{kappa_bound} then follows readily by substituting \eqref{eq:ell-convex} into \eqref{convex_loss} and using the definition of $\kappa$ in the theorem statement. For $\Xi=\mathbb{R}^m$ we have $\sigma_\Xi= \indI{\{0\}}$, and thus the upper bound \eqref{kappa_bound} becomes exact. Finally, if $\varepsilon=0$, then \eqref{dist-rob-Ex} trivially coincides with \eqref{kappa_bound} under our conventions of extended arithmetic. Thus, the claim follows. \end{proof} Theorem~\ref{thm:convex} asserts that for $\Xi=\mathbb{R}^m$, the worst-case expectation \eqref{dist-rob-Ex} of a convex loss function reduces the sample average of the loss adjusted by the simple correction term $\kappa\varepsilon$. The following proposition highlights that $\kappa$ can be interpreted as a measure of maximum steepness of the loss function. This interpretation has intuitive appeal in view of Definition~\ref{def:wass}. \begin{Prop}[Steepness of the loss function] \label{prop:kappa} Let $\kappa$ be defined as in Theorem~\ref{thm:convex}. \begin{enumerate}[label=(\roman*), itemsep = 1mm, topsep = 1mm] \item \label{itm:lem:upper} If $\ell$ is $\ol{L}$-Lipschitz continuous, i.e., if there exists $\xi' \in \mathbb{R}^m$ such that $\ell(\xi) - \ell(\xi') \le \ol{L}\|\xi-\xi'\|$ for all $\xi \in \mathbb{R}^m$, then~$\kappa \le \ol{L}$. \item \label{itm:lem:lower} If $\ell$ majorizes an affine function, i.e., if there exists $\theta\in \mathbb{R}^m$ with $\|\theta\|_*=:\ul L$ and $\xi' \in \mathbb{R}^m$ such that $\ell(\xi) - \ell(\xi') \ge \inner{\theta}{\xi-\xi'}$ for all $\xi \in \mathbb{R}^m$, then $\kappa \ge \ul{L} $. \end{enumerate} \end{Prop} \begin{proof} The proof follows directly from the definition of conjugacy. As for \ref{itm:lem:upper}, we have \begin{align*} \ell^*(\theta) = \sup_{\xi \in \mathbb{R}^m} \inner{\theta}{\xi} - \ell(\xi)~& \ge \sup_{\xi \in \mathbb{R}^m} \inner{\theta}{\xi} - \ol{L} \|\xi -\xi'\| -\ell(\xi')\\ & = \sup_{\xi \in \mathbb{R}^m} \inf_{\|z\|_*\le \ol{L}} \inner{\theta}{\xi} - \inner{z}{\xi -\xi'}- \ell(\xi'), \end{align*} where the last equality follows from the definition of the dual norm. Applying the minimax theorem \cite[Proposition 5.5.4]{ref:Bert-09} and explicitly carrying out the maximization over $\xi$ yields \begin{align*} \ell^*(\theta) \ge \left\{ \begin{array}{cl} \inner{\theta}{\xi'}-\ell(\xi') & \mbox{if } \|\theta\|_* \le \ol{L}, \\ \infty & \mbox{otherwise.} \end{array}\right. \end{align*} Consequently, $\ell^*(\theta)$ is infinite for all $\theta$ with $\|\theta\|_*> \ol L$, which readily implies that the $\|\cdot\|_*$-ball of radius $\ol L$ contains the effective domain of $\ell^*$. Thus, $\kappa \le \ol{L}$. As for \ref{itm:lem:lower}, we have \begin{align*} \ell^*(\theta) = \sup_{\xi \in \mathbb{R}^m} \inner{\theta}{\xi} - \ell(\xi) & \le \sup_{\xi \in \mathbb{R}^m} \inner{\theta}{\xi} - \inner{z}{\xi-\xi'} - \ell(\xi') \\ & = \sigma_{\mathbb{R}^m}(\theta - z)+ \inner{z}{\xi'} - \ell(\xi'), \end{align*} which implies that $\ell^*(\theta) \le \inner{\theta}{\xi'} - \ell(\xi') < \infty$. Thus, $\theta$ belongs to the effective domain of $\ell^*$. We then conclude that $\kappa \ge \|\theta\|_* = \ul{L}$. \end{proof} \change{ \begin{Rem}[Consistent formulations] If $\Xi=\mathbb{R}^m$ and the loss function is given by $\ell(\xi) = \max_{k \le K}\{\inner{a_{k}}{\xi} + b_{k}\}$, then both Corollary~\ref{cor:affine} and Theorem~\ref{thm:convex} offer an exact reformulation of the worst-case expectation~\eqref{dist-rob-Ex} in terms of a finite-dimensional convex program. On the one hand, Corollary~\ref{cor:affine} implies that \eqref{dist-rob-Ex} is equivalent~to \begin{align*} \left\{ \begin{array}{clll} \min\limits_{\lambda} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} \ell(\wh{\xi}_i)\\ \normalfont \text{s.t.} & \|a_k\|_* \le \lambda & \forall k \le K, \end{array} \right. \end{align*} which is obtained by setting $C=0$ and $d=0$ in~\eqref{affine-max}. At optimality we have $\lambda^\star=\max_{k\le K} \|a_k\|_*$, which corresponds to the (best) Lipschitz constant of $\ell(\xi)$ with respect to the norm~$\|\cdot\|$. On the other hand, Theorem~\ref{thm:convex} implies that \eqref{dist-rob-Ex} is equivalent to \eqref{kappa_bound} with $\kappa=\lambda^\star$. Thus, Corollary~\ref{cor:affine} and Theorem~\ref{thm:convex} are consistent. \end{Rem}} \change{ \begin{Rem}[$\varepsilon$-insensitive optimizers\footnote{\change{We are indepted to Vishal Gupta who has brought this interesting observation to our attention.}}] Consider a loss function $h(x,\xi)$ that is convex in $\xi$, and assume that $\Xi=\mathbb{R}^m$. In this case Theorem~\ref{thm:convex} remains valid, but the steepness parameter $\kappa(x)$ may depend on $x$. For loss functions whose Lipschitz modulus with respect to $\xi$ is independent of $x$ (e.g., the newsvendor loss), however, $\kappa(x)$ is constant. In this case the distributionally robust optimization problem~\eqref{DRO} and the SAA problem~\eqref{Ex_emp} share the same minimizers irrespective of the Wasserstein radius~$\varepsilon$. This phenomenon could explain why the SAA solutions tend to display a surprisingly strong out-of-sample performance in these problems. \end{Rem}} \section{Numerical Results} \label{sec:num} We validate the theoretical results of this paper in the context of a stylized portfolio selection problem. The subsequent simulation experiments are designed to provide additional insights into the performance guarantees of the proposed distributionally robust optimization scheme. \subsection{Mean-Risk Portfolio Optimization} Consider a capital market consisting of $m$ assets whose yearly returns are captured by the random vector $\xi = [\xi_1, \ldots, \xi_m]^\intercal$. If short-selling is forbidden, a portfolio is encoded by a vector of percentage weights $x=[x_1,\ldots,x_m]^\intercal$ ranging over the probability simplex $\mathbb{X}=\{x\in\mathbb R^m_+: \sum_{i=1}^{m}x_i = 1\}$. As portfolio $x$ invests a percentage $x_i$ of the available capital in asset $i$ for each $i=1,\ldots,m$, its return amounts to $\inner{x}{\xi}$. In the remainder we aim to solve the single-stage stochastic program \begin{align} J^\star = \inf_{x\in\mathbb{X}} \bigg\{\mathds{E}^{\mathds{P}}\big[-\inner{x}{\xi}\big] + \rho\, \mathds{P}\text{-}{\rm CVaR}_{\alpha}\big(-\inner{x}{\xi}\big) \bigg\}, \label{portfolio} \end{align} which minimizes a weighted sum of the mean and the conditional value-at-risk (CVaR) of the portfolio loss $-\inner{x}{\xi}$, where $\alpha\in (0,1]$ is referred to as the confidence level of the CVaR, and $\rho\in\mathbb{R}_+$ quantifies the investor's risk-aversion. Intuitively, the CVaR at level $\alpha$ represents the average of the $\alpha\times 100\%$ worst (highest) portfolio losses under the distribution $\mathds{P}$. Replacing the CVaR in the above expression with its formal definition \cite{ref:Rock&Yry-00}, we obtain \begin{align*} J^\star & = \inf_{x\in\mathbb{X}}\bigg\{\mathds{E}^{\mathds{P}}\big[-\inner{x}{\xi}\big] + \rho\,\inf_{\tau\in\mathbb{R}} \mathds{E}^\mathds{P} \Big[ \tau + {1\over \alpha} \max\big\{ -\inner{x}{\xi} - \tau, 0 \big \}\Big] \bigg\} \\ & = \inf_{x\in\mathbb{X}, \tau\in\mathbb{R}} \mathds{E}^\mathds{P} \Big[ \max_{k\le K} \, a_k\inner{ x}{\xi}+b_k \tau \Big], \end{align*} where $K=2$, $a_1= -1$, $a_2= -1-\frac{\rho}{\alpha}$, $b_1=\rho$ and $b_2= \rho(1-\frac{1}{\alpha})$. An investor who is unaware of the distribution $\mathds{P}$ but has observed a dataset $\wh{\Xi}_N$ of $N$ historical samples from $\mathds{P}$ and knows that the support of $\mathds{P}$ is contained in $\Xi=\{\xi\in \mathbb{R}^m:C\xi\leq d\}$ might solve the distributionally robust counterpart of \eqref{portfolio} with respect to the Wasserstein ambiguity set $\ball{\wh{\PP}_N}{\varepsilon}$, that is, \begin{align*} \change{\wh{J}_N} %{J_{\rm DD}^\star(\varepsilon)} \coloneqq \inf_{x\in\mathbb{X}, \tau\in\mathbb{R}} \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{E}^\mathds{Q} \Big[ \max_{k\le K} \, a_k \inner{ x}{\xi}+b_k \tau \Big], \end{align*} \change{where we make the dependence on the Wasserstein radius $\varepsilon$ explicit.} By Corollary~\ref{cor:affine} we know that \begin{align} \label{dro:portfolio} \change{\wh{J}_N} %{J_{\rm DD}^\star (\varepsilon)} = \left\{ \begin{array}{lclll} & \inf\limits_{x,\tau,\lambda,s_i, \gamma_{ik}} & \lambda \varepsilon + {1 \over N}\sum\limits_{i = 1}^{N} s_i \\ & \normalfont \text{s.t.} & x\in\mathbb{X}\\ & & b_k\tau +a_k\inner{x}{\wh{\xi}_i}+ \inner{\gamma_{ik}}{d-C\wh{\xi}_i} \le s_i & \forall i \le N, & \forall k \le K\\ & & \|C^\intercal\gamma_{ik} - a_{k}x\|_* \le \lambda & \forall i \le N, & \forall k \le K \\ & & \gamma_{ik} \ge 0& \forall i \le N, & \forall k \le K. \end{array} \right. \end{align} Before proceeding with the numerical analysis of this problem, we provide some analytical insights into its optimal solutions when there is significant ambiguity. In what follows we keep the training data set fixed and let $\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon)$ be an optimal distributionally robust portfolio corresponding to the Wasserstein ambiguity set of radius $\varepsilon$. We will now show that, for natural choices of the ambiguity set, $\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon)$ converges to the equally weighted portfolio $\frac{1}{m}e$ as $\varepsilon$ tends to infinity, where $e \coloneqq (1,\ldots,1)^\intercal$. The optimality of the equally weighted portfolio under high ambiguity has first been demonstrated in \cite{ref:PfluRichWoz-11} using analytical methods. We identify this result here as an immediate consequence of Theorem~\ref{thm:dist-rob-opt}, which is primarily a computational result. For any non-empty set $S\subseteq \mathbb{R}^m$ we denote by $\mbox{recc}(S) \coloneqq \{y\in\mathbb{R}^m:x+\lambda y\in S~\forall x\in S, ~\forall \lambda\geq 0\}$ the recession cone and by $S^\circ \coloneqq \{y\in \mathbb{R}^m:\inner{y}{x}\leq 0~\forall x\in S\}$ the polar cone of $S$. \begin{Lem} \label{polarrecession} If $\{\varepsilon_k\}_{k\in\mathbb N}\subset \mathbb{R}_+$ tends to infinity, then any accumulation point $x^\star$ of $\big\{\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon_k)\big\}_{k\in\mathbb N}$ is a portfolio that has minimum distance to $(\mbox{\em recc}(\Xi))^\circ$ with respect to $\|\cdot\|_*$. \end{Lem} \begin{proof} Note first that $\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon_k)$, $k\in\mathbb N$, and $x^\star$ exist because $\mathbb{X}$ is compact. For large Wasserstein radii $\varepsilon$, the term $\lambda\varepsilon$ dominates the objective function of problem~\eqref{dro:portfolio}. Using standard epi-convergence results \cite[Section~7.E]{ref:Rockafellar-10}, one can thus show that \begin{align*} x^\star\, & \in \arg\min_{x\in\mathbb{X}}~ \min_{\gamma_{ik}\geq 0}~ \max_{i\le N,\, k\le K} \|C^\intercal\gamma_{ik} - a_{k}x\|_*\\ & = \arg\min_{x\in\mathbb{X}}~ \max_{i\le N,\, k\le K}~ \min_{\gamma\geq 0} ~ \|C^\intercal\gamma + |a_{k}|\,x\|_* \\ & = \arg\min_{x\in\mathbb{X}} ~ \min_{\gamma\geq 0} ~ \|C^\intercal\gamma + x\|_* ~ \max_{k\le K} |a_k| \\ & = \arg\min_{x\in\mathbb{X}} ~ \min_{\gamma\geq 0} ~ \|C^\intercal\gamma + x\|_*, \end{align*} where the first equality follows from the fact that $a_k<0$ for all $k\le K$, the second equality uses the substitution $\gamma\rightarrow \gamma |a_k|$, and the last equality holds because the set of minimizers of an optimization problem is not affected by a positive scaling of the objective function. Thus, $x^\star$ is the portfolio nearest to the cone $\mathcal C=\{C^\intercal\gamma:\gamma\geq 0\}$. The claim now follows as the polar cone \begin{align*} \mathcal C^\circ \coloneqq \{y\in\mathbb{R}^m:y^\intercal x\leq 0~\forall x\in\mathcal C\}= \{y\in\mathbb{R}^m:y^\intercal C^\intercal \gamma \leq 0~\forall \gamma\geq 0\}= \{y\in\mathbb{R}^m: Cy\geq 0\} \end{align*} is readily recognized as the recession cone of $\Xi$ and as $\mathcal C=(\mathcal C^\circ)^\circ$. \end{proof} \begin{Prop}[Equally weighted portfolio] \label{prop:1/n} Assume that the Wasserstein metric is defined in terms of the $p$-norm in the uncertainty space for some $p\in[1,\infty)$. If $\{\varepsilon_k\}_{k\in\mathbb N}\subset \mathbb{R}_+$ tends to infinity, then $\big\{\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon_k)\big\}_{k\in\mathbb N}$ converges to the equally weighted portfolio $x^\star=\frac{1}{m}e$ provided that the uncertainty set is given by \begin{itemize} \item[(i)] the entire space, i.e., $\Xi=\mathbb{R}^m$, or \item[(ii)] the nonnegative orthant shifted by $-e$, i.e., $\Xi=\{\xi\in\mathbb{R}^m:\xi \geq-e\}$, which captures the idea that no asset can lose more than $100\%$ of its value. \end{itemize} \end{Prop} \begin{proof} (i) One easily verifies from the definitions that $(\mbox{recc}(\Xi))^\circ=\{0\}$. Moreover, we have $\|\cdot\|_*=\|\cdot\|_q$ where $\frac{1}{p}+\frac{1}{q}=1$. As $p\in [1,\infty)$, we conclude that $q\in (1,\infty]$, and thus the unique nearest portfolio to $(\mbox{recc}(\Xi))^\circ$ with respect to $\|\cdot\|_*$ is $x^\star=\frac{1}{m}e$. The claim then follows from Lemma \ref{polarrecession}. Assertion~(ii) follows in a similar manner from the observation that $(\mbox{recc}(\Xi))^\circ$ is now the non-positive orthant. \end{proof} \change{With some extra effort one can show that for every $p\in[1,\infty)$ there is a threshold $\bar \varepsilon>0$ with $\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon)=x^\star$ for all $\varepsilon\geq \bar \varepsilon$, see~\cite[Proposition~3]{ref:PfluRichWoz-11}. Moreover, for $p\in\{1,2\}$ the threshold $\bar \varepsilon$ is known analytically.} \subsection{Simulation Results: Portfolio Optimization} \label{sec:simulation} \change{Our experiments are based on a market with $m=10$ assets considered in~\cite[Section~7.5]{ref:BertSAA-14}.} In view of the capital asset pricing model we may assume that the return $\xi_i$ is decomposable into a systematic risk factor $\psi\sim \mathcal N(0,2\%)$ common to all assets and an unsystematic or idiosyncratic risk factor $\zeta_i\sim\mathcal N(i\times 3\%, i\times 2.5\%)$ specific to asset $i$. Thus, we set $\xi_i=\psi+\zeta_i$, where $\psi$ and the idiosyncratic risk factors $\zeta_i$, $i=1,\ldots,m$, constitute independent normal random variables. By construction, assets with higher indices promise higher mean returns at a higher risk. Note that the given moments of the risk factors completely determine the distribution $\mathds{P}$ of~$\xi$. This distribution has support $\Xi=\mathbb{R}^m$ and satisfies Assumption~\ref{a:exp} for the tail exponent $a=1$, say. We also set $\alpha=20\%$ and $\rho=10$ in all numerical experiments, and we use the $1$-norm to measure distances in the uncertainty space. Thus, $\|\cdot\|_*$ is the $\infty$-norm, whereby \eqref{dro:portfolio} reduces to a linear program. \begin{figure*} [t] \centering \subfigure[$N=30$ training samples]{\label{fig:x:1} \includegraphics[width=0.31\columnwidth]{N30x-eps-converted-to}} \hspace{0mm} \subfigure[$N=300$ training samples]{\label{fig:x:2} \includegraphics[width=0.31\columnwidth]{N300x-eps-converted-to}} \hspace{0mm} \subfigure[$N=3000$ training samples]{\label{fig:x:3} \includegraphics[width=0.31\columnwidth]{N3000x-eps-converted-to}} \hspace{0mm} \caption{Optimal portfolio composition as a function of the Wasserstein radius~$\varepsilon$ averaged over 200 simulations; the portfolio weights are depicted in ascending order, {\em i.e.}, the weight of asset~1 at the bottom (dark blue area) and that of asset~10 at the top (dark red area)} \label{fig:x} \end{figure*} \change{ \subsubsection{\bf Impact of the Wasserstein Radius} In the first experiment we investigate the impact of the Wasserstein radius $\varepsilon$ on the optimal distributionally robust portfolios and their out-of-sample performance. We solve problem \eqref{dro:portfolio} using training datasets of cardinality $N \in \{30, 300, 3000\}$. Figure~\ref{fig:x} visualizes the corresponding optimal portfolio weights $\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon)$ as a function of $\varepsilon$, averaged over $200$ independent simulation runs. Our numerical results confirm the theoretical insight of Proposition~\ref{prop:1/n} that the optimal distributionally robust portfolios converge to the equally weighted portfolio as the Wasserstein radius $\varepsilon$ increases; see also~\cite{ref:PfluRichWoz-11}. The out-of-sample performance \begin{align*} J\big(\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon)\big) \coloneqq \mathds{E}^{\mathds{P}}\big[-\inner{\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon)}{\xi}\big] + \rho\, \mathds{P}\text{-}{\rm CVaR}_{\alpha}\big(-\inner{\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon)}{\xi}\big) \end{align*} of any fixed distributionally robust portfolio $\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon)$ can be computed analytically as $\mathds{P}$ constitutes a normal distribution by design, see, {\em e.g.}, \cite[p.~29]{ref:Rock&Yry-00}. Figure~\ref{fig:perf:eps} shows the tubes between the 20\% and 80\% quantiles (shaded areas) and the means (solid lines) of the out-of-sample performance $J\big(\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon)\big)$ as a function of $\varepsilon$---estimated using $200$ independent simulation runs. We observe that the out-of-sample performance improves (decreases) up to a critical Wasserstein radius $\varepsilon_{\rm crit}$ and then deteriorates (increases). This stylized fact was observed consistently across all of simulations and provides an empirical justification for adopting a distributionally robust approach. Figure~\ref{fig:perf:eps} also visualizes the reliability of the performance guarantees offered by our distributionally robust portfolio model. Specifically, the dashed lines represent the empirical probability of the event $J\big(\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon)\big) \le \wh{J}_N} %{J_{\rm DD}^\star(\varepsilon)$ with respect to $200$ independent training datasets. We find that the reliability is nondecreasing in $\varepsilon$. This observation has intuitive appeal because $\wh{J}_N} %{J_{\rm DD}^\star(\varepsilon) \ge J(\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon))$ whenever $\mathds{P}\in \ball{\wh{\PP}_N}{\varepsilon}$, and the latter event becomes increasingly likely as $\varepsilon$ grows. Figure~\ref{fig:perf:eps} also indicates that the certificate guarantee sharply rises towards 1 near the critical Wasserstein radius $\varepsilon_{\rm crit}$. Hence, the out-of-sample performance of the distributionally robust portfolios improves as long as the reliability of the performance guarantee is noticeably smaller than 1 and deteriorates when it saturates at 1. Even though this observation was made consistently across all simulations, we were unable to validate it theoretically. \begin{figure*}[t] \centering \subfigure[$N=30$ training samples]{\label{fig:perf:1} \includegraphics[width=0.32\columnwidth]{N30av}} \subfigure[$N=300$ training samples]{\label{fig:perf:2} \includegraphics[width=0.32\columnwidth]{N300av}} \subfigure[$N=3000$ training samples]{\label{fig:perf:3} \includegraphics[width=0.32\columnwidth]{N3000av}} \change{\caption{Out-of-sample performance $J(\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon)) $ (left axis, solid line and shaded area) and reliability $\mathds{P}^N[J(\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon)) \le \wh{J}_N} %{J_{\rm DD}^\star(\varepsilon)]$ (right axis, dashed line) as a function of the Wasserstein radius $\varepsilon$ and estimated on the basis of 200 simulations}} \label{fig:perf:eps} \end{figure*} \subsubsection{\bf Impact of the Sample Size} \label{subsub:sim:N} Different Wasserstein radii $\varepsilon$ may result in robust portfolios $\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon)$ with vastly different out-of-sample performance $J(\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon))$. Ideally, one should select the radius $\widehat \varepsilon_N^{\rm \; opt}$ that minimizes $J(\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon))$ over all $\varepsilon\geq 0$; note that $\widehat \varepsilon_N^{\rm \; opt}$ inherits the dependence on the training data from $J(\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon))$. As the true distribution $\mathds{P}$ is unknown, however, it is impossible to evaluate and minimize $J(\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon))$. In practice, the best we can hope for is to approximate $\widehat \varepsilon_N^{\rm \; opt}$ using the training data. Statistics offers several methods to accomplish this goal: \begin{itemize} \item {\em Holdout method:} Partition $\wh{\xi}_1,\ldots,\wh{\xi}_N$ into a training dataset of size $N_T$ and a validation dataset of size $N_V=N-N_T$. Using only the training dataset, solve~\eqref{dro:portfolio} for a large but finite number of candidate radii $\varepsilon$ to obtain $\widehat x_{N_T}(\varepsilon)$. Use the validation dataset to estimate the out-of-sample performance of $\widehat x_{N_T}(\varepsilon)$ via the sample average approximation. Set $\widehat \varepsilon_N^{\rm \; hm}$ to any $\varepsilon$ that minimizes this quantity. Report $\wh{x}_N} %{x_{\rm DD}^\star^{\rm \; hm}=\widehat x_{N_T}(\widehat \varepsilon_N^{\rm \; hm})$ as the data-driven solution and $\wh{J}_N} %{J_{\rm DD}^\star^{\rm \; hm}=\widehat J_{N_T}(\widehat \varepsilon_N^{\rm \; hm})$ as the corresponding certificate. \item {\em $k$-fold cross validation:} Partition $\wh{\xi}_1,\ldots,\wh{\xi}_N$ into $k$ subsets, and run the holdout method $k$ times. In each run, use exactly one subset as the validation dataset and merge the remaining $k-1$ subsets to a training dataset. Set $\widehat \varepsilon_N^{\rm \; cv}$ to the average of the Wasserstein radii obtained from the $k$ holdout runs. Resolve~\eqref{dro:portfolio} with $\varepsilon=\widehat \varepsilon_N^{\rm \; cv}$ using all $N$ samples, and report $\wh{x}_N} %{x_{\rm DD}^\star^{\rm \; cv}=\wh{x}_N} %{x_{\rm DD}^\star(\widehat \varepsilon_N^{\rm \; cv})$ as the data-driven solution and $\wh{J}_N} %{J_{\rm DD}^\star^{\rm \; cv}=\wh{J}_N} %{J_{\rm DD}^\star(\widehat \varepsilon_N^{\rm \; cv})$ as the corresponding certificate. \end{itemize} The holdout method is computationally cheaper, but cross validation has superior statistical properties. There are several other methods to estimate the best Wassertein radius $\widehat \varepsilon_N^{\rm \; opt}$. By construction, however, no method can provide a radius $\widehat \varepsilon_N$ such that $\wh{x}_N} %{x_{\rm DD}^\star(\widehat \varepsilon_N)$ has a better out-of-sample performance than $\wh{x}_N} %{x_{\rm DD}^\star(\widehat \varepsilon_N^{\rm \; opt})$. In all experiments we compare the distributionally robust approach based on the Wasserstein ambiguity set with the classical sample average approximation (SAA) and with a state-of-the-art data-driven distributionally robust approach, where the ambiguity set is defined via a linear-convex ordering (LCX)-based goodness-of-fit test \cite[Section~3.3.2]{ref:BertSAA-14}. The size of the LCX ambiguity set is determined by a single parameter, which should be tuned to optimize the out-of-sample performance. While the best parameter value is unavailable, it can again be estimated using the holdout method or via cross validation. To our best knowledge, the LCX approach represents the only existing data-driven distributionally robust approach for {\em continuous} uncertainty spaces that enjoys strong finite-sample guarantees, asymptotic consistency as well as computational tractability.\footnote{\change{Much like worst-case expectations over Wasserstein balls, worst-case expectations over LCX ambiguity sets can be reformulated as finite convex programs whenever the underlying loss function represents a pointwise maximum of $K$ concave component functions. Unlike problem~\eqref{eq:thm-dual:2} in Theorem~\ref{thm:dist-rob-opt}, however, the resulting convex program scales exponentially with $K$.}} \begin{figure*} [t] \centering \subfigure[Holdout method]{\label{fig:perf-val} \includegraphics[width=0.31\columnwidth]{perf_val}} \hspace{1mm} \subfigure[Holdout method]{\label{fig:cert-val} \includegraphics[width=0.31\columnwidth]{cert_val}} \hspace{1mm} \subfigure[Holdout method]{\label{fig:beta-val} \includegraphics[width=0.31\columnwidth]{beta_val}} \hspace{1mm} \subfigure[$k$-fold cross validation]{\label{fig:perf-k} \includegraphics[width=0.31\columnwidth]{perf_k}} \hspace{1mm} \subfigure[$k$-fold cross validation]{\label{fig:cert-k} \includegraphics[width=0.31\columnwidth]{cert_k}} \hspace{1mm} \subfigure[$k$-fold cross validation]{\label{fig:beta-k} \includegraphics[width=0.31\columnwidth]{beta_k}} \hspace{1mm} \subfigure[Best size]{\label{fig:perf-best} \includegraphics[width=0.31\columnwidth]{perf_best}} \hspace{1mm} \subfigure[Best size]{\label{fig:cert-best} \includegraphics[width=0.31\columnwidth]{cert_best}} \hspace{1mm} \subfigure[Best size]{\label{fig:beta-best} \includegraphics[width=0.31\columnwidth]{beta_best}} \change{\caption{Out-of-sample performance $J(\wh{x}_N} %{x_{\rm DD}^\star)$, certificate $\wh{J}_N} %{J_{\rm DD}^\star$, and certificate reliability $\mathds{P}^N\big[J(\wh{x}_N} %{x_{\rm DD}^\star) \le \wh{J}_N} %{J_{\rm DD}^\star\big]$ for the SAA, LCX and Wasserstein solutions as a function of $N$}} \label{fig:perf:N} \end{figure*} In Figures~\ref{fig:perf-val}--\ref{fig:beta-val} the sizes of the (LCX and Wasserstein) ambiguity sets are determined via the holdout method, where $80\%$ of the data are used for training and $20\%$ for validation. Figure~\ref{fig:perf-val} visualizes the tube between the $20\%$ and $80\%$ quantiles (shaded areas) as well as the mean value (solid lines) of the out-of-sample performance $J(\wh{x}_N} %{x_{\rm DD}^\star)$ as a function of the sample size $N$ and based on 200 independent simulation runs, where $\wh{x}_N} %{x_{\rm DD}^\star$ is set to the minimizer of the SAA (blue), LCX (purple) and Wasserstein (green) problems, respectively. The constant dashed line represents the optimal value $J^\star$ of the original stochastic program~\eqref{Ex-true}, which is computed through an SAA problem with $N = 10^6$ samples. We observe that the Wasserstein solutions tend to be superior to the SAA and LCX solutions in terms of out-of-sample performance. Figure~\ref{fig:cert-val} shows the optimal values $\wh{J}_N} %{J_{\rm DD}^\star$ of the SAA, LCX and Wasserstein problems, where the sizes of the ambiguity sets are chosen via the holdout method. Unlike Figure~\ref{fig:perf-val}, Figure~\ref{fig:cert-val} thus reports {\em in-sample} estimates of the achievable portfolio performance. As expected, the SAA approach is over-optimistic due to the optimizer's curse, while the LCX and Wasserstein approaches err on the side of caution. All three methods are known to enjoy asymptotic consistency, which is in agreement with all in-sample and out-of-sample results. Figure~\ref{fig:beta-val} visualizes the reliability of the different performance certificates, that is, the empirical probability of the event $J(\wh{x}_N} %{x_{\rm DD}^\star) \le \wh{J}_N} %{J_{\rm DD}^\star$ evaluated over 200 independent simulation runs. Here, $\wh{x}_N} %{x_{\rm DD}^\star$ represents either an optimal portfolio of the SAA, LCX or Wasserstein problems, while $\wh{J}_N} %{J_{\rm DD}^\star$ denotes the corresponding optimal value. The optimal SAA portfolios display a disappointing out-of-sample performance relative to the optimistically biased mimimum of the SAA problem---particularly when the training data is scarce. In contrast, the out-of-sample performance of the optimal LCX and Wasserstein portfolios often undershoots~$\wh{J}_N} %{J_{\rm DD}^\star$. Figures~\ref{fig:perf-k}--\ref{fig:beta-k} show the same graphs as Figures~\ref{fig:perf-val}--\ref{fig:beta-val}, but now the sizes of the ambiguity sets are determined via $k$-fold cross validation with $k=5$. In this case, the out-of-sample performance of both distributionally robust methods improves slightly, while the corresponding certificates and their reliabilities increase significantly with respect to the na\"ive holdout method. However, these improvements come at the expense of a $k$-fold increase in the computational cost. One could think of numerous other statistical methods to select the size of the Wasserstein ambiguity set. As discussed above, however, if the ultimate goal is to minimize the out-of-sample performance of $\wh{x}_N} %{x_{\rm DD}^\star(\varepsilon)$, then the best possible choice is $\varepsilon=\widehat \varepsilon_N^{\rm \; opt}$. Similarly, one can construct a size parameter for the LCX ambiguity set that leads to the best possible out-of-sample performance of any LCX solution. We emphasize that these optimal Wasserstein radii and LCX size parameters are not available in practice as their computation would require infinitely many validation samples. In simulated experiments, however, they can be computed to high accuracy. Figures~\ref{fig:perf-best}--\ref{fig:beta-best} show the same graphs as Figures~\ref{fig:perf-val}--\ref{fig:beta-val} for optimally sized ambiguity sets. Thus, no implementable statistical method can achieve a better out-of-sample performance. \begin{figure*} [t] \centering \includegraphics[width=.43\columnwidth]{eps-eps-converted-to} \change{\caption{ Optimal Wasserstein radius $\widehat\varepsilon_N^{\rm\; opt}$ and its estimates $\widehat\varepsilon_N^{\rm\; hm}$ and $\widehat\varepsilon_N^{\rm\; cv}$ obtained via the holdout method and $k$-fold cross validation, respectively (averaged over 200 simulations)}} \label{fig:eps:N} \end{figure*} Finally, Figure~\ref{fig:eps:N} depicts the best possible Wasserstein radius $\widehat\varepsilon_N^{\rm \; opt}$ as well as the Wasserstein radii $\widehat\varepsilon_N^{\rm \; hm}$ and $\widehat\varepsilon_N^{\rm \; cv}$ obtained by the holdout method and via $k$-fold cross validation, respectively, as a function of the sample size $N$. All results are averaged across 200 independent simulation runs. As expected from Theorem~\ref{thm:convergence}, all three Wasserstein radii tend to zero as $N$ increases. Moreover, the convergence rate is approximately equal to $N^{-\frac{1}{2}}$. This rate is likely to be optimal. Indeed, if $\mathbb{X}$ is a singleton, then every quantile of the sample average estimator of $J^\star$ converges to $J^\star$ at rate $N^{-\frac{1}{2}}$ due to the central limit theorem. Thus, if $\widehat \varepsilon_N= o(N^{-\frac{1}{2}})$, then $\wh{J}_N} %{J_{\rm DD}^\star$ also converges to $J^\star$ at leading order $N^{-\frac{1}{2}}$ by Theorem~\ref{thm:convex}, which applies as the loss function is convex. This indicates that the a~priori rate $N^{-\frac{1}{m}}$ suggested by Theorem~\ref{thm:concentration} is too pessimistic in~practice. } \subsection{Simulation Results: Uncertainty Quantification} \label{sec:uqsimulation} Investors often wish to determine the probability that a given portfolio will outperform various benchmark indices or assets. Our results on uncertainty quantification developed in Section~\ref{subsec:UQ} enable us to compute this probability in a meaningful way---solely on the basis of the training dataset. \change{ Assume for example that we wish to quantify the probability that any data-driven portfolio $\wh{x}_N} %{x_{\rm DD}^\star$ outperforms the three most risky assets in the market {\em jointly}. Thus, we should compute the probability of the closed polytope \begin{align*} \widehat{\set{A}} = \Big\{\xi \in \mathbb{R}^m ~:~ \inner{\wh{x}_N} %{x_{\rm DD}^\star}{\xi}\geq \xi_i ~ \forall i=8,9,10 \Big \}. \end{align*} As the true distribution $\mathds{P}$ is unknown, the probability $\mathds{P}[\xi \in \widehat{\set A}]$ cannot be evaluated exactly. Note that $\widehat{\set A}$ as well as $\mathds{P}[\xi \in \widehat{\set A}]$ constitute random objects that depend on $\wh{x}_N} %{x_{\rm DD}^\star$ and thus on the training data. Using the {\em same} training dataset that was used to compute $\wh{x}_N} %{x_{\rm DD}^\star$, however, we may estimate $\mathds{P}[\xi \in \widehat{\set A}]$ from above and below by \begin{align*} \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{Q} \left[ \xi\in \widehat{\set A}\right]\qquad \text{and}\qquad \inf\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{Q} \left[ \xi\in \widehat{\set A}\right] = 1 - \sup\limits_{\mathds{Q} \in \ball{\wh{\PP}_N}{\varepsilon}} \mathds{Q} \left[ \xi\notin \widehat{\set A}\right], \end{align*} respectively. The upper confidence bound can be computed by solving the linear program~\eqref{chance-worst}. Replacing $\widehat{\set A}$ with its interior in the lower confidence bound leads to another (potentially weaker) lower bound that can be computed by solving the linear program~\eqref{chance-best}. We denote these computable bounds by $\wh{J}_N} %{J_{\rm DD}^\star^+(\varepsilon)$ and $\wh{J}_N} %{J_{\rm DD}^\star^-(\varepsilon)$, respectively. In all subsequent experiments $\wh{x}_N} %{x_{\rm DD}^\star$ is set to a solution of the distributionally robust program~\eqref{dro:portfolio} calibrated via $k$-fold cross validation as described in Section~\ref{subsub:sim:N}. \subsubsection{\bf Impact of the Wasserstein Radius}\label{subsubsec:sim:UQ_eps} \begin{figure}[t!] \centering \subfigure[$N=30$]{\label{fig:UQ:30} \includegraphics[width=0.43\columnwidth]{UQ30}} \quad \subfigure[$N=300$]{\label{fig:UQ:300} \includegraphics[width=0.43\columnwidth]{UQ300}} \change{\caption{Excess $\wh{J}_N} %{J_{\rm DD}^\star^+(\varepsilon)- \mathds{P}[\widehat{\set A}]$ and shortfall $\wh{J}_N} %{J_{\rm DD}^\star^-(\varepsilon)- \mathds{P}[\widehat{\set A}]$ (solid lines, left axis) as well as reliability $\mathds{P}^N[\wh{J}_N} %{J_{\rm DD}^\star^-(\varepsilon) \leq \mathds{P}[\widehat{\set A}] \leq \wh{J}_N} %{J_{\rm DD}^\star^+(\varepsilon)]$ (dashed lines, right axis) as a function of $\varepsilon$}} \label{fig:UQ_radius} \end{figure} As $\wh{J}_N} %{J_{\rm DD}^\star^+(\varepsilon)$ and $\wh{J}_N} %{J_{\rm DD}^\star^-(\varepsilon)$ estimate a random target $\mathds{P}[\widehat{\set A}]$, it makes sense to filter out the randomness of the target and to study only the differences $\wh{J}_N} %{J_{\rm DD}^\star^+(\varepsilon)- \mathds{P}[\widehat{\set A}]$ and $\wh{J}_N} %{J_{\rm DD}^\star^-(\varepsilon)- \mathds{P}[\widehat{\set A}]$. Figures~\ref{fig:UQ:30} and~\ref{fig:UQ:300} visualize the empirical mean (solid lines) as well as the tube between the empirical 20\% and 80\% quantiles (shaded areas) of these differences as a function of the Wasserstein radius $\varepsilon$, based on 200 training datasets of cardinality $N = 30$ and $N=300$, respectively. Figure~\ref{fig:UQ_radius} also shows the empirical reliability of the bounds (dashed lines), that is, the empirical probability of the event $\wh{J}_N} %{J_{\rm DD}^\star^-(\varepsilon) \leq \mathds{P}[\widehat{\set A}] \leq \wh{J}_N} %{J_{\rm DD}^\star^+(\varepsilon)$. Note that the reliability drops to 0 for $\varepsilon=0$, in which case both $\wh{J}_N} %{J_{\rm DD}^\star^+(0)$ and $\wh{J}_N} %{J_{\rm DD}^\star^-(0)$ coincide with the SAA estimator for $\mathds{P}[\widehat{\set A}]$. Moreover, at $\varepsilon=0$ the set $\widehat{\set A}$ is constructed from the SAA portfolio $\wh{x}_N} %{x_{\rm DD}^\star$, whose performance is overestimated on the training dataset. Thus, the SAA estimator for $\mathds{P}[\widehat{\set A}]$, which is evaluated using the same training dataset, is positively biased. For $\varepsilon>0$, finally, the reliability increases as the shaded confidence intervals move away from 0. \subsubsection{\bf Impact of the Sample Size}\label{subsubsec:sim:UQ_N} We propose a variant of the $k$-fold cross validation procedure for selecting $\varepsilon$ in uncertainty quantification. Partition $\wh{\xi}_1,\ldots,\wh{\xi}_N$ into $k$ subsets and repeat the following holdout method $k$ times. Select one of the subsets as the validation set of size $N_V$ and merge the remaining $k-1$ subsets to a training dataset of size $N_T=N-N_V$. Use the validation set to compute the SAA estimator of $\mathds{P}[\widehat{\set A}]$, and use the training dataset to compute $\widehat J_{N_T}^+(\varepsilon)$ for a large but finite number of candidate radii $\varepsilon$. Set $\widehat \varepsilon_N^{\; \rm hm}$ to the smallest candidate radius for which the SAA estimator of $\mathds{P}[\widehat{\set A}]$ is not larger than $\widehat J_{N_T}^+(\varepsilon)$. Next, set $\widehat \varepsilon_N^{\rm \; cv}$ to the average of the Wasserstein radii obtained from the $k$ holdout runs, and report $\wh{J}_N} %{J_{\rm DD}^\star^+=\widehat J_{N}^+(\widehat \varepsilon_N^{\rm \; cv})$ as the data-driven upper bound on $\mathds{P}[\widehat{\set A}]$. The data-driven lower bound $\wh{J}_N} %{J_{\rm DD}^\star^-$ is constructed analogously in the obvious way. Figure~\ref{fig:UQ_N} visualizes the empirical means (solid lines) as well as the tubes between the empirical 20\% and 80\% quantiles (shaded areas) of $\wh{J}_N} %{J_{\rm DD}^\star^+-\mathds{P}[\widehat{\set A}]$ and $\wh{J}_N} %{J_{\rm DD}^\star^--\mathds{P}[\widehat{\set A}]$ as a function of the sample size $N$, based on 300 independent training datasets. As expected, the confidence intervals shrink and converge to $0$ as $N$ increases. We emphasize that $\wh{J}_N} %{J_{\rm DD}^\star^+$ and $\wh{J}_N} %{J_{\rm DD}^\star^-$ are computed solely on the basis of $N$ training samples, whereas the computation of $\mathds{P}[\widehat{\set A}]$ necessitates a much larger dataset, particularly if $\widehat{\set A}$ constitutes a rare event. \begin{figure*}[t!] \centering \change{ \subfigure[Excess $\wh{J}_N} %{J_{\rm DD}^\star^+-\mathds{P}{[}\widehat{\set A}{]}$ and shortfall $\wh{J}_N} %{J_{\rm DD}^\star^--\mathds{P}{[}\widehat{\set A}{]}$ of the data-driven confidence bounds for $\mathds{P}{[}\widehat{\set A}{]}$]{\label{fig:UQ_N} \includegraphics[width=0.465\columnwidth]{UQ_N}} \quad \subfigure[Data-driven Wasserstein radius $\widehat\varepsilon_N^{\rm \; cv}$ obtained via $k$-fold cross validation]{\label{fig:UQ_eps} \includegraphics[width=0.49\columnwidth]{UQ_eps-eps-converted-to}} \caption{Dependence of the confidence bounds and the Wasserstein radius on $N$}} \label{fig:UQ_learning} \end{figure*} Figure~\ref{fig:UQ_eps} shows the Wasserstein radius $\widehat\varepsilon_N^{\rm \; cv}$ obtained via $k$-fold cross validation (both for $\wh{J}_N} %{J_{\rm DD}^\star^+$ and $\wh{J}_N} %{J_{\rm DD}^\star^-$). As usual, all results are averaged across 300 independent simulation runs. A comparison with Figure~\ref{fig:eps:N} reveals that the data-driven Wasserstein radii in uncertainty quantification display a similar but faster polynomial decay than in portfolio optimization. We conjecture that this is due to the absence of decisions, which implies that uncertainty quantification is less susceptible to the optimizer's curse. Thus, nature ({\em i.e.}, the fictitious adversary choosing the distribution in the ambiguity set) only has to compensate for noise but not for bias. A smaller Wasserstein radius seems to be sufficient for this purpose. } \paragraph*{\bf Acknowledgments} We thank Soroosh Shafieezadeh Abadeh for helping us with the numerical experiments. The authors are grateful to Vishal Gupta, Ruiwei Jiang and Nathan Kallus for their valuable comments. This research was supported by the Swiss National Science Foundation under Grant BSCGI0\_157733. \section{\@startsection{section}{1}% \def\subsection{\@startsection{subsection}{2}% \z@{.5\linespacing\@plus.7\linespacing}{.5\linespacing}% {\normalfont\large\bfseries}} \def\subsubsection{\@startsection{subsubsection}{3}% \z@{.5\linespacing\@plus.7\linespacing}{.5\linespacing}% {\normalfont\itshape}} \usepackage[usenames, dvipsnames]{color} \definecolor{darkblue}{rgb}{0.0, 0.0, 0.45} \usepackage[colorlinks = true, raiselinks = true, linkcolor = darkblue, citecolor = Mahogany, urlcolor = ForestGreen, pdfauthor = {Peyman Mohajerin Esfahani}, pdftitle = {}, pdfkeywords = {}, pdfsubject = {}, plainpages = false, bookmarks =false]{hyperref} \usepackage{dsfont,amssymb,amsmath,subfigure, graphicx,enumitem} \usepackage{amsfonts,dsfont,mathtools, mathrsfs,amsthm} \usepackage{algorithm,algorithmicx,fancyhdr,epstopdf} \usepackage{wasysym,MnSymbol} \allowdisplaybreaks \date{\today}

https://arxiv.org/abs/1501.03030

Classifying finite-dimensional C*-algebras by posets of their commutative C*-subalgebras

We consider the functor C that to a unital C*-algebra A assigns the partial order set C(A) of its commutative C*-subalgebras ordered by inclusion. We investigate how some C*-algebraic properties translate under the action of C to order-theoretical properties. In particular, we show that A is finite dimensional if and only C(A) satisfies certain chain conditions. We eventually show that if A and B are C*-algebras such that A is finite dimensional and C(A) and C(B) are order isomorphic, then A and B must be *-isomorphic.

\section*{Introduction} Given a C*-algebra $A$ with unit $1_A$, let $\CC(A)$ be the set of all commutative unital C*-subalgebras $C$ of $A$ such that $1_A\in C$. Then $\CC(A)$ becomes a poset\footnote{We refer to the appendix for an overview of the necessary concepts in order theory.} if we order it by inclusion. Now, one could consider the following question: is it possible to recover the structure of $A$ as a C*-algebra from the poset $\CC(A)$? More precisely, if $A$ and $B$ are C*-algebras such that $\CC(A)$ and $\CC(B)$ are isomorphic as posets, can we find an *-isomorphism between $A$ and $B$? Apart from its mathematical relevance, this problem is of considerable importance for the so called \emph{quantum toposophy} program, where one tries to describe quantum mechanics in terms of topos theory (see e.g., \cite{BI}, \cite{DI}, \cite{HLS}, \cite{Wolters}). In this program, the central objects of research are the topoi $\mathrm{Sets}^{\CC(A)}$ and $\mathrm{Sets}^{\CC(A)^\op}$. The motivation behind at least \cite{HLS} and \cite{Wolters} is Niels Bohr's doctrine of classical concepts, which, roughly speaking, states that a measurement provides a ``classical snapshot of quantum reality''. Mathematically, this corresponds with an element of $\CC(A)$, and knowledge of all classical snapshots should provide a picture of quantum reality, that is as complete as (humanly) possible. The possiblility of reconstructing $A$ from $\CC(A)$ would assure the soundness of this doctrine. For commutative C*-algebras, Mendivil showed that the answer to the question is affirmative (see \cite{Mendivil2}). It turns out that more can be said about the commutative case. As we will see, $\CC$ turns out to be a functor from unital C*-algebras to posets. Hence if $f:A\to B$ is a *-isomorphism, then $\CC(f):\CC(A)\to\CC(B)$ is an order isomorphism. In \cite{Hamhalter}, Hamhalter gave not only a different proof of Mendivil's statement that an order isomorphism $\psi:\CC(A)\to\CC(B)$, with $A$ and $B$ commutative C*-algebras, yields an *-isomorphism $f:A\to B$, but he showed as well that this *-isomorphism can be constructed in such a way that $\psi=\CC(f)$. Moreover, he proved that as long as $A$ is not two dimensional, there is only one *-isomorphism that induces $\psi$ in this way. For non-commutative C*-algebras however, the answer is negative, since Connes \cite{Connes} showed the existence of a C*-algebra $A_c$ (actually even a von Neumann algebra) that is not isomorphic to its opposite algebra $A_c^\op$. Here the opposite algebra is the C*-algebra with the same underlying topological vector space, but with multiplication defined by $(a,b)\mapsto ba$, where $(a,b)\mapsto ab$ denotes the original multiplication. Since $\CC(A)$ is always isomorphic to $\CC(A^\op)$ as poset, for each C*-algebra $A$, the existence of Connes' C*-algebra $A_c$ shows that the order structure of $\CC(A)$ is not always enough in order to reconstruct $A$. More recent counterexamples can be found in \cite{Phillips} and \cite{PV}. Nevertheless, there are still problems one could study. For instance, in \cite{DH} D\"oring and Harding consider a functor similar to $\CC$, namely the functor $\V$ assigning to a von Neumann algebra $M$ the poset $\V(M)$ of its commutative von Neumann subalgebras, and prove that one can reconstruct the \emph{Jordan} structure, i.e., the anticommutator $(a,b)\mapsto ab+ba$, of $M$ from $\V(M)$. Similarly, in \cite{Hamhalter}, it is shown that if $\CC(A)$ and $\CC(B)$ are order isomorphic, then there exists a quasi-linear Jordan isomorphism between $A_{\mathrm{sa}}$ and $B_{\mathrm{sa}}$, the sets of self-adjoint elements of $A$ and $B$, respectively. Here quasi-linear means linear with respect to elements that commute. In \cite{Hamhalter2}, it is even shown that this quasi-linear Jordan isomorphism is linear when $A$ and $B$ are \emph{AW*-algebras}. Moreover, one could replace $\CC(A)$ by a structure with stronger properties. An example of such a structure is an \emph{active lattice}, defined in \cite{HR}, where Heunen and Reyes also show that this structure is strong enough to determine AW*-algebras completely. In this paper, however, we will only study $\CC$ as a functor and as an invariant for C*-algebras. We shall examine some of its properties, and, as our main result, we shall prove that if $A$ is finite dimensional, $\CC(A)$ has enough structure to determine $A$ up to *-isomorphism. The way this result is proven considerably differs from the methods of Hamhalter and Mendivil in the commutative case, since we cannot use Gel'fand duality and the ensuing topological methods as they do. Our approach also differs from Hamhalter's methodes in the sense that if $\psi:\CC(A)\to\CC(B)$ is an order isomorphism with $A$ finite dimensional, we do not construct a *-isomorphism $f:A\to B$ such that $\CC(f)=\psi$, since it will turn out that this is not always possible. Instead, we look at order-theoretical invariants in $\CC(A)$ that reflect the C*-algebraic invariants in $A$ determinining $A$ up to *-isomorphism. The first step is to find order-theoretical properties of $\CC(A)$ that corresponds to $A$ being finite dimensional. These properties turn out to be chain conditions on $\CC(A)$. The next step is to identify $Z(A)$, the center of $A$, as element of $\CC(A)$. This turns out to be the meet of all maximal elements of $\CC(A)$. Finally, we consider the interval $[Z(A),M]$ in $\CC(A)$, where $M$ is any maximal element. This interval is a lattice, which factors into a product of directly indecomposable lattices. The length of the maximal chains in these lattices corresponds exactly to the dimensions of the matrix algebra factors of $A$, since by the Artin-Wedderburn Theorem $A$ factors in a unique way into a product of matrix algebras. \section{C*-subalgebras of a commutative C*-algebra} \begin{convention} All C*-algebras in this article are assumed to be unital. \end{convention} \begin{definition}\label{def:subalgebraequivalencerelation} Let $A$ be a C*-subalgebra of $C(X)$. Then we define an equivalence relation $\sim_A$ on $X$ by $x\sim_A y$ if and only if $f(x)=f(y)$ for each $f\in A$. We denote the equivalence class of $x$ under this equivalence relation by $[x]_A$. \end{definition} \begin{lemma}\cite[Proposition 5.1.3]{Weaver}\label{lem:subalgofCXisquotientofX} A C*-algebra $A$ is a C*-subalgebra of $C(X)$ if and only if there is a compact Hausdorff space $Y$ and a continuous surjection $q:X\to Y$ such that $C_q[C(Y)]=A$. The map $C_q:C(Y)\to C(X)$, $f\mapsto f\circ q$ is an isometric *-homomorphism in this case, and $Y$ and $X/\sim_A$ are homeomorphic. \end{lemma} \begin{definition}\label{def:idealalgebra} Let $X$ be compact Hausdorff and let $K\subseteq X$ be closed. Then we denote the C*-subalgebra $$\{f\in C(X):f\mathrm{\ is\ constant\ on\ }K\}$$ by $C_K$. \end{definition} \begin{lemma}\label{lem:subalgebraisintersectionofidealalgebras}\cite[Proposition 2.2]{Hamhalter} Let $A$ be a C*-subalgebra of $C(X)$. Then $A=\bigcap_{x\in X}C_{[x]_A}$, where $[x]_A$ is interpreted as a subset of $X$ (rather than as a point in $X/\sim_A$). \end{lemma} \begin{proof} By Lemma \ref{lem:subalgofCXisquotientofX}, we have $A=C_q[C(Y)]$ with $Y=X/\sim_A$ and $q:X\to Y$ the quotient map. So each $f\in A$ is of the form $g\circ q$ for some $g\in C(Y)$, meaning that $f$ is constant on $[x]_A$ for each $x\in X$. Conversely, let $f$ be constant on $[x]_A$ for each $x\in X$. Define $g:Y\to X$ by $g([x]_A)=f(x)$. Then $g$ is well defined, since $f$ is constant on $[x]_A$, and $f=g\circ q$. Moreover, let $U\subseteq\C$ be open. Then $g^{-1}[U]$ is open in $Y$ if and only if $q^{-1}[g^{-1}[U]]$ is open in $X$ by definition of the quotient topology. Since $f=g\circ q$, we find $g^{-1}[U]$ is open in $Y$ if and only if $f^{-1}[U]$ is open in $X$. Since $f$ is continuous, we find that $g^{-1}[U]$ is open, so $g\in C(X)$. Hence $f$ lifts to a function on $C(Y)$, so $f\in A$. Thus \begin{eqnarray*} A & = & \{f\in C(X):f\ \mathrm{is\ constant\ on\ }[x]_A\ \forall x\in X\}\\ & = & \bigcap_{x\in X}\{f\in C(X):f\ \mathrm{is\ constant\ on\ }[x]_A\}\\ & = & \bigcap_{x\in X}C_{[x]_A}. \end{eqnarray*} \end{proof} \section{The functor $\CC$} \begin{definition} Let $A$ be a C*-algebra with unit $1_A$. We denote the set of its commutative C*-subalgebras containing $1_A$ by $\CC(A)$. \end{definition} If we denote the category of unital C*-algebras with unital *-homomorphisms as morphisms by $\mathbf{uCStar}$ and the category of posets with order morphisms as morphisms by $\mathbf{Poset}$, then $$\CC:\mathbf{uCStar}\to\mathbf{Poset}$$ can be made into a functor \cite[Proposition 5.3.3]{Heunen}. \begin{lemma}\label{lem:CAisinvariantforA} $\CC:\mathbf{uCStar}\to\mathbf{Poset}$ becomes a functor if for *-homomorphisms $f:A\to B$ between C*-algebras $A$ and $B$ we define $\CC(f):\CC(A)\to\CC(B)$ by $C\mapsto f[C]$. \end{lemma} \begin{proof} Let $C\in\CC(A)$. Then the restriction of $f$ to $C$ is a *-homomorphism with codomain $B$. It follows from the First Isomorphism Theorem for C*-algebras (see for instance \cite[Theorem 3.1.6]{Murphy}) that $f[C]$ is a C*-subalgebra of $B$. Since $f$ is multiplicative, it follows that $f[C]$ is commutative. Clearly $f[C]$ is a *-subalgebra of $B$, so $f[C]\in\CC(B)$. Moreover, we have $f[C]\subseteq f[D]$ if $C\subseteq D$, so $\CC(f)$ is an order morphism. If $f:A\to B$ and $g:B\to D$ are *-homomorphisms, then $\CC(g\circ f)(C)=g\circ f[C]=g[f[C]]=\CC(g)\circ\CC(f)$, and if $I_A:A\to A$ is the identity morphism, then $\CC(I_A)=1_{\CC(A)}$, the identity morphism of $\CC(A)$. Thus $\CC$ is indeed a functor. \end{proof} \begin{lemma}\label{lem:CAismeetsemilattice} Let $A$ be a C*-algebra. Then $\CC(A)$ has all non-empty meets, where the meet is given by the intersection operator. In particular, $\CC(A)$ is a meet-semilattice, and has a least element $$\C1_A=\{\lambda 1_A:\lambda\in\C\}.$$ \end{lemma} \begin{proof} Elementary. \end{proof} \begin{lemma}\cite[Proposition 14]{BH}\label{lem:CAlatticewhenAcommutative} Let $A$ be a C*-algebra. Then the following statements are equivalent. \begin{enumerate} \item[(i)] $A$ is commutative; \item[(ii)] $\CC(A)$ is bounded; \item[(iii)] $\CC(A)$ is a complete lattice. \end{enumerate} \end{lemma} \begin{proof} This follows immediately from the observation that $A\in\CC(A)$ if and only if $A$ is commutative, and the fact that if $\CC(A)$ has a greatest element, it has all meets, so it must be a complete lattice. \end{proof} The next proposition is originally due to Spitters \cite{Spitters}. A similar statement for the functor $\V$ assigning to a von Neumann algebra $M$ the poset $\V(M)$ of its commutative von Neumann subalgebras can be found in \cite{DRSB}. \begin{proposition}\label{prop:CAisdcpo} Let $A$ be a C*-algebra. Then $\CC(A)$ is a dcpo, where $\bigvee\D=\overline{\bigcup\D}$ for each directed $\D\subseteq\CC(A)$, and where $\bigvee\D=\bigcup\D$ if $A$ is finite dimensional. \end{proposition} \begin{proof} Let $\D\subseteq\CC(A)$ be a directed subset. Let $S=\bigcup\D$. We show that $S$ is a commutative *-algebra. Let $x,y\in S$ and $\lambda,\mu\in\C$, there are $D_1,D_2\in\D$ such that $x\in D_1$ and $y\in D_2$. Since $\D$ is directed, there is some $D_3\in\D$ such that $D_1,D_2\subseteq D_3$. Hence $x,y\in D_3$, whence $\lambda x+\mu y,x^*,xy\in D_3$, and since $D_3$ is commutative, we obtain $xy=yx$. Since $D_3\subseteq S$, it follows that $S$ is a commutative *-subalgebra of $A$. Now, $\overline{S}$ is a commutative C*-subalgebra of $A$, which is the smallest commutative C*-subalgebra of $A$ containing every element of $\D$, hence it is the join of $\D$. If $A$ is finite dimensional, all subspaces of $A$ are closed, hence $\bigvee\D=\overline{\bigcup\D}=\bigcup\D$. \end{proof} If $A$ is not finite dimensional, we do not necessarily have the equality $\bigvee\D=\bigcup\D$ for each directed $\D\subseteq\CC(A)$. For instance, let $A=C([0,1])$ and $$\D=\{C_{[0,1/n]}:n\in\N\}.$$ Then $\bigcup\D$ consists of functions in $C([0,1])$ that are all constant on some neighborhood of $0$. Hence $f\notin\bigcup\D$, where $f:[0,1]\to\C$ is defined by $f(x)=x$. However, one can easily show that $\bigcup\D$ is a unital *-subalgebra of $C([0,1])$ that separates all points of $[0,1]$, so $\overline{\bigcup\D}=C([0,1])$ by the Stone-Weierstrass Theorem. Thus $\bigcup\D\neq\bigvee\D$. \begin{proposition}\label{prop:Coff} Let $f:A\to B$ be a *-homomorphism. Then \begin{enumerate} \item[(i)] If $\CC(f)$ is surjective, then $f$ is surjective; \item[(ii)] If $f$ is injective, then $\CC(f)$ is an order embedding such that $$\down\CC(f)[\CC(A)]=\CC(f)[\CC(A)]$$ and \begin{equation}\label{eq:Cfpreservesintersections} \CC(f)\left(\bigcap_{i\in I}C_i\right)=\bigcap_{i\in I}\CC(f)(C_i) \end{equation} for each non-empty subset $\{C_i\}_{i\in I}\subseteq\CC(A)$. If $\{D_j\}_{j\in J}\subseteq\CC(A)$ is a subset such that $\bigvee_{j\in J}D_j$ exists, then $\bigvee_{j\in J}\CC(f)(D_j)$ exists and \begin{equation}\label{eq:Cfpreservesjoins} \CC(f)\left(\bigvee_{j\in J}D_j\right)=\bigvee_{j\in J}\CC(f)(D_j). \end{equation} Moreover $\CC(f)$ has an upper adjoint $\CC(f)_*:\CC(B)\to\CC(A)$ given by $D\mapsto f^{-1}[D]$ such that $$\CC(f)_*\circ\CC(f)=1_{\CC(A)};$$ \item[(iii)] If $f$ is a *-isomorphism, then $\CC(f)$ is an order isomorphism. \end{enumerate} \end{proposition} \begin{proof}\ \begin{enumerate} \item[(i)] Assume that $\CC(f)$ is surjective. Let $b\in B$, and let $b_1=\frac{b+b^*}{2}$ and $b_2=\frac{b-b^*}{2i}$. Then $b=b_1+ib_2$, and $b_1$ and $b_2$ are self-adjoint elements of $B$. It follows that $C^*(b_i,1_B)\in\CC(B)$ for each $i=1,2$, hence by the surjectivity of $\CC(f)$, there are $C_1,C_2\in\CC(A)$ such that $\CC(f)(C_i)=C^*(b_i,1_B)$. Since $\CC(f)(C_i)=f[C_i]$, this means that there are $a_1\in C_1$ and $a_2\in C_2$ such that $f(a_i)=b_i$. Let $a=a_1+i a_2$. Then $f(a)=b$, hence $f$ is surjective. \item[(ii)] Assume that $f$ is injective. We first show that $\CC(f)$ has an upper adjoint $\CC(f)_*$. Let $D\in\CC(B)$ and $x,y\in f^{-1}[D]$. Then $f(x),f(y)\in D$, so $$f(xy-yx)=f(x)f(y)-f(y)f(x)=0.$$ By the injectivity of $f$ it follows that $xy=yx$, so $f^{-1}[D]$ is a commutative *-subalgebra of $A$, which is closed since $f$ is continuous and $D$ is closed. Moreover, since $f(1_A)=1_B$, and $1_B\in D$, it follows that $1_A\in f^{-1}[D]$. So $D\mapsto f^{-1}[D]$ is a well-defined map $\CC(B)\to\CC(A)$. Moreover, this map is clearly inclusion preserving, and since $f[C]\subseteq D$ if and only if $C\subseteq f^{-1}[D]$ for any two subsets $C$ and $D$ of $A$ and $B$, respectively, we find that $D\mapsto f^{-1}[D]$ is indeed the upper adjoint of $\CC(f)$. Let $\{C_i\}_{i\in I}$ be a non-empty collection of elements of $\CC(A)$. We always have $f[\bigcap_{i\in I}C_i]\subseteq\bigcap_{i\in I}f[C_i]$. Now, let $x\in\bigcap_{i\in I}f[C_i]$. Then for each $i\in I$ there is an $c_i\in C_i$ such that $x=f(c_i)$. Hence for each $i,j\in I$, we have $f(c_i)=f(c_j)$. By injectivity of $f$ it follows that $c_i=c_j$, so $x=f(c)$ with $c\in\bigcap_{i\in I} C_i$ equal to $c_i$ for each $i\in I$. We conclude that $f[\bigcap_{i\in I}C_i]=\bigcap_{i\in I}f[C_i]$, which is exactly (\ref{eq:Cfpreservesintersections}). Let $\{D_j\}_{j\in J}$ be a collection of elements of $\CC(A)$ such that $\bigvee_{j\in J}D_j$ exists. Then $\CC(f)(D_k)\leq\CC(f)\left(\bigvee_{j\in J}D_j\right)$ for each $k\in J$. Let $E\in\CC(B)$ such that $\CC(f)(D_k)\leq E$ for each $k\in J$. Since $\CC(f)$ has an upper adjoint $\CC(f)_*$, we find that $D_k\leq\CC(f)_*(E)$ for each $k\in J$, so $\bigvee_{j\in J}D_j\leq\CC(f)_*(E)$. Again using the adjunction, we find $\CC(f)\left(\bigvee_{j\in J}D_j\right)\leq E$. We conclude that $\CC(f)\left(\bigvee_{j\in J}D_j\right)$ is the join of $\{\CC(f)(D_j):j\in J\}$, and it follows automatically that (\ref{eq:Cfpreservesjoins}) holds. By injectivity of $f$, we find $$\CC(f)_*\circ\CC(f)(C)=f^{-1}[f[C]]=C$$ for each $C\in\CC(A)$, hence $\CC(f)_*\circ\CC(f)=1_{\CC(A)}$. Let $C\in\down\CC(f)[\CC(A)]$. Hence there is some $D\in\CC(A)$ such that $C\subseteq\CC(f)(D)$. Then $$\CC(f)_*(C)\subseteq \CC(f)_*\circ\CC(f)(D)=D.$$ Now, $$\CC(f)\circ\CC(f)_*(C)=f[f^{-1}[C]]=C\cap f[A]=C,$$ since $C\subseteq\CC(f)(D)=f[D]\subseteq f[A]$. Thus $C=\CC(f)(E)$ where $E=\CC(f)_*(C)$, hence $C\in\CC(f)[\CC(A)]$. We conclude that $$\down\CC(f)[\CC(A)]=\CC(f)[\CC(A)].$$ Finally, we show that $\CC(f)$ is an order embedding. Let $C_1$ and $C_2$ elements of $\CC(A)$. Since $\CC(f)$ is an order morphism, we have $\CC(f)(C_1)\subseteq\CC(f)(C_2)$ if $C_1\subseteq C_2$. Assume now that $\CC(f)(C_1)\subseteq\CC(f)(C_2)$. In other words, we have $f[C_1]\subseteq f[C_2]$. By the injectivity of $f$, we have $f^{-1}[f[C]]=C$ for each $C\in\CC(A)$. Since $f^{-1}$ preserves inclusions, this implies $C_1\subseteq C_2$. Hence we obtain $\CC(f)(C_1)\subseteq\CC(f)(C_2)$ if and only if $C_1\subseteq C_2$. In other words, $\CC(f)$ is an order embedding. \item[(iii)] This follows directly from the functoriality of $\CC$ and the fact that $f$ has an inverse. \end{enumerate} \end{proof} \section{Finite-dimensional C*-algebras} In this section we shall prove the following theorem. \begin{theorem}\label{thm:fdmaintheorem} Let $A$ be a finite-dimensional C*-algebra and $B$ any C*-algebra such that $\CC(A)\cong\CC(B)$. Then $A\cong B$. \end{theorem} The proof of the theorem relies on the fact that every finite-dimensional C*-algebra is *-isomorphic to a finite sum of matrix algebras. \begin{theorem}[Artin-Wedderburn]\index{Artin-Wedderburn Theorem} Let $A$ be a finite-dimensional C*-algebra. Then there are $k,n_1,\ldots,n_k\in\N$ such that $$A\cong\bigoplus_{i=1}^k\MM_{n_i}(\C).$$ The number $k$ is unique, whereas the numbers $n_1,\ldots,n_k$ are unique up to permutation. \end{theorem} \begin{proof} \cite[Theorem I.11.2]{Takesaki1} \end{proof} The strategy for proving Theorem \ref{thm:fdmaintheorem} is the following. First, we find an order-theoretic property of $\CC(A)$ that corresponds with the finite-dimensionality of $A$. Then we find a method of retrieving the numbers $k,n_1,\ldots,n_k$ from $\CC(A)$. \begin{definition} Let $\CC$ be a poset. Then $\CC$ is called \emph{Artinian}\index{Artinian poset} if it satisfies one of the following equivalent conditions: \begin{enumerate} \item[1)] Every non-empty subset of $\CC$ contains a minimal element; \item[2)] All non-empty filtered subsets of $\CC$ have a least element; \item[3)] $\CC$ satisfies the \emph{descending chain condition}\index{descending chain condition}: if we have a sequence of elements $C_1\geq C_2\geq \ldots$ in $\CC$, i.e., a countable descending chain, then the sequence stabilizes, i.e., there is an $n\in\N$ such that $C_k=C_{n}$ for all $k>n$. \end{enumerate} \end{definition} In order to see that these conditions are equivalent, assume that (1) holds and let $\F$ a non-empty filtered subset of $\CC$. Then $\F$ must have a minimal element $C$. Now, if $F\in\F$, then there must be an $G\in\F$ such that $G\leq C,F$. Since $C$ is minimal, it follows that $G=C$, so $C\leq F$, whence $C$ is the least element of $\F$. For (2) implies (3), let $C_1\geq C_2\geq C_3\geq \ldots$ be a descending chain. Then $\F=\{C_i\}_{i\in\N}$ is clearly a directed subset, so it has a least element, say $C_n$. So we must have $C_k=C_n$ for all $k>n$, hence $\CC$ satisfies (3). Finally, we show by contraposition that (3) follows from (1). So assume that $\CC$ does not satisfy the descending chain condition. Hence we can construct a sequence $C_1\geq C_2\geq \ldots$ that does not terminate. The set \mbox{$\F=\{C_n:n\in\N\}$} is then a non-empty subset of $\CC$ without a minimal element. Thus $\CC$ does not satisfy (1). There exists also a notion dual to the notion of an Artinian poset. \begin{definition} Let $\CC$ be a poset. Then $\CC$ is called \emph{Noetherian} if is satisfies one of the following equivalent conditions: \begin{enumerate} \item[1)] Every non-empty subset contains a maximal element; \item[2)] All non-empty directed subsets of $\CC$ have a greatest element; \item[3)] $\CC$ satisfies the \emph{ascending chain condition}\index{ascending chain condition}: if we have a sequence of elements $C_1\leq C_2\leq \ldots$ in $\CC$, i.e., a countable ascending chain, then the sequence stabilizes, i.e., there is an $n\in\N$ such that $C_k=C_{n}$ for all $k>n$. \end{enumerate} \end{definition} The following proposition can be found in \cite{Just} as Theorem 4.21. \begin{proposition}[Principle of Artinian induction]\index{Artinian induction}\label{prop:ArtinianInduction} Let $\CC$ be an Artinian poset and $\PP$ a property such that: \begin{enumerate} \item $\PP(C)$ is true for each minimal $C\in \CC$ \emph{(induction basis)}; \item $\PP(B)$ is true for all $B<C$ implies that $\PP(C)$ is true \emph{(induction step)}. \end{enumerate} Then $\PP(C)$ is true for each $C\in \CC$. \end{proposition} \begin{proof} Assume that $\F=\{C\in \CC:\PP(C)$ is not true$\}$ is non-empty. Since $\CC$ is Artinian, this means that $\F$ has a minimal element $C$. Hence $\PP(B)$ is true for all elements $B<C$, so $\PP(C)$ is true by the induction step, contradicting the definition of $\F$. \end{proof} \begin{definition}\label{def:gradedposet} Let $\CC$ be a poset. Then $\CC$ is called \emph{graded}\index{graded poset} if one can define a function $d:\CC\to\N$, called a \emph{rank function}\index{rank function} such that: \begin{enumerate} \item[(i)] $d(C)=1$ for each $C\in\min\CC$; \item[(ii)] $d(C_1)<d(C_2)$ for each $C_1,C_2\in\CC$ such that $C_1<C_2$; \item[(iii)] $C_2$ is a cover of $C_1$ if and only if $C_1\leq C_2$ and $d(C_2)=d(C_1)+1$ for each $C_1,C_2\in\CC$. \end{enumerate} \end{definition} There is no standard definition of a graded poset. For instance, in \cite{Roman} condition (i) is dropped and $\Z$ is taken as codomain of rank functions. On the other hand, \cite{KRY} assumes condition (i), but not condition (ii). For our purposes, it is convenient to combine both definitions. The next three lemmas are now easy to prove. \begin{lemma}\label{lem:gradedimpliesartinian} Let $\CC$ be a graded poset with rank funcion $d:\CC\to\N$. Then $\CC$ is Artinian. If the range of $d$ is bounded from above, $\CC$ is Noetherian as well. \end{lemma} \begin{proof} Let $\F\subseteq\CC$ a non-empty filtered subset. Then $d[\F]\subseteq\N$ is non-empty, so it contains a least element $n$. Let $F_1,F_2\in\F$ such that $d(F_1)=d(F_2)=n$. Since $\F$ is filtered, there is an $F\in\F$ such that $F\leq F_1,F_2$. Assume that $F\neq F_1$. Then $F<F_1$, so $d(F)<d(F_1)=n$ contradicting the minimality of $n$. Hence we must have $F=F_1$ and in a similar way, we find that $F=F_2$. So there is a unique element $F\in\F$ such that $d(F)=n$. Let $F'\in\F$. Since $\F$ is filtered, there is some $F''\in\F$ such that $F''\leq F,F'$. Again if $F''<F$, we find $d(F'')<n$ contradicting the minimality of $n$, so $F''=F$. It follows that $F\leq F'$, so $F$ is the least element of $\F$. We conclude that $\CC$ is Artinian. Now assume that $d[\CC]\subseteq\N$ has an upper bound, then the proof that $\CC$ is Noetherian follows in an analogous way. \end{proof} \begin{lemma}\label{lem:uniquenessrankfunction} Let $\CC$ be a graded poset. Then its rank function $d:\CC\to\N$ is unique. \end{lemma} \begin{proof} By Lemma \ref{lem:gradedimpliesartinian}, $\CC$ is Artinian. Hence every non-empty subset has a minimal element, and in particular $\min\CC\neq\emptyset$. Assume that $g:\CC\to\N$ is a rank function. By definition of a rank function we have $g(M)=d(M)=1$ for each $M\in\min\CC$. Let $C\in\CC$ such that $C\notin\min\CC$. Assume that $d(B)=g(B)$ for each $B<C$. The set $\{d(B):B<C\}\subseteq\N$ is non-empty and bounded by $d(C)$, hence it must have a maximum $n$. As a consequence, there is some $B<C$ such that $d(B)=n$. Assume that $n+1\neq d(C)$. Then $C$ does not cover $B$, hence there is some $B'\in\CC$ such that $B<B'<C$. Thus $n=d(B)<d(B')$ contradicting the maximality of $n$. We conclude that $d(B)+1=d(C)$, so $C$ covers some $B$. Hence we obtain $$d(C)=d(B)+1=g(B)+1=g(C),$$ so $d=g$ by Artinian induction. \end{proof} \begin{lemma}\label{lem:gradedposetunderorderiso} Let $\phi:\CC\to\D$ be an order isomorphism between graded posets $\CC$ and $\D$ with rank functions $d_\CC$ and $d_\D$, respectively. Then $d_\CC=d_\D\circ\phi$. \end{lemma} \begin{proof} Let $d=d_\D\circ\phi$. We check that $d$ is a rank function on $\CC$. Let $C\in\min\CC$, then $\phi(C)\in\min\D$, so $d(C)=d_\D\circ\phi(C)=1$. Let $C_1,C_2\in\CC$ such that $C_1<C_2$. Then $\phi(C_1)\leq\phi(C_2)$, and since $\phi$ is injective, we cannot have equality. Hence $\phi(C_1)<\phi(C_2)$, so $$d(C_1)=d_\D\circ\phi(C_1)<d_\D\circ\phi(C_2)=d(C_2).$$ Finally, let $C_1,C_2\in\CC$. Clearly, $C_2$ covers $C_1$ if and only if $\phi(C_2)$ covers $\phi(C_1)$ if and only if $d_2(\phi(C_2))=d_2(\phi(C_1))+1$ and $C_1\leq C_2$ if and only if $d(C_2)=d(C_1)+1$ and $C_1\leq C_2$. Thus $d$ is a rank function on $\CC$ and by Lemma \ref{lem:uniquenessrankfunction}, we obtain $d_1=d$. \end{proof} \begin{lemma}\label{lem:rankfunctiononCA} Let $A$ be a finite-dimensional C*-algebra. Then $\CC(A)$ is graded with a rank function $\dim:\CC(A)\to\N$ assigning to each element $C\in\CC(A)$ its dimension. Moreover, the range of $\dim$ is bounded from above. \end{lemma} \begin{proof} Trivial. \end{proof} Combining Lemmas \ref{lem:gradedimpliesartinian} and \ref{lem:rankfunctiononCA}, we find that if $A$ is finite dimensional, then $\CC(A)$ is both Artinian and Noetherian. We shall prove that the converse holds as well. \begin{lemma}\label{lem:maximalelementsofCA} Let $A$ be a C*-algebra. Then every element of $\CC(A)$ is contained in a maximal element of $\CC(A)$. In particular, the set $\max\CC(A)$ is non-empty. \end{lemma} \begin{proof} Let $C\in\CC(A)$ and let $\mathcal{S}=\{D\in\CC(A):C\subseteq D\}$. Then $\mathcal{S}$ is non-empty, so if $\CC(A)$ is Noetherian, we immediately find that $\mathcal{S}$ contains a maximal element. If $\CC(A)$ is not Noetherian, we need Zorn's Lemma. Let $\D=\{D_i\}_{i\in I}$ be a chain in $\mathcal{S}$. Then $\D$ is certainly directed, hence it must have a join $\bigvee\D$ by Proposition \ref{prop:CAisdcpo}. Since $\bigvee\D$ clearly contains $C$, we have $\bigvee\D\in\mathcal{S}$. So for every chain $\CC(A)$, there is an upper bound for the chain in $\mathcal{S}$. By Zorn's Lemma it follows that $\mathcal{S}$ contains a maximal element $M$. Now, $M$ must also be maximal in $\CC(A)$, since if there is some $A\in\CC(A)$ such that $M\subseteq A$, then $C\subseteq A$, so $A\in\mathcal{S}$. By the maximality of $M$ with respect to $\mathcal{S}$, it follows that $A$ must be equal to $M$. \end{proof} \begin{proposition}\label{prop:finitedimensionalmaximalcommutativesubalgebras} Let $A$ be a C*-algebra and $M$ a maximal commutative *-subalgebra. If $M$ is finite dimensional, then $A$ must be finite dimensional as well. \end{proposition} This statement can be found in \cite{KR1} as Exercise 4.12. The solution of this exercise can be found in \cite{KR3}. In particular, when $A=\MM_{n}(\C)$, there is a nice characterization of the maximal commutative C*-subalgebras of $\CC(A)$. \begin{lemma}\label{lem:dimensionofmaximalsubalgebraofmatrixalgebra} Let $A=\MM_n(\C)$ and $M\in\max\CC(A)$. Then $M$ is $n$-dimensional and there is some $u\in \mathrm{SU}(n)$ such that $$M=\{udu^*:d\in D_n\},$$ where $D_n$ is the commutative C*-subalgebra of $A$ consisting of all diagonal matrices. \end{lemma} \begin{proof} See \cite[Example 5.3.5]{Heunen}. \end{proof} \begin{definition} Let $X$ be a topological space with topology $\O(X)$. Then $X$ is called \emph{Noetherian}\index{Noetherian topological space} if the poset $\O(X)$ ordered by inclusion is Noetherian. \end{definition} \begin{lemma}\cite[Exercise I.1.7]{Hartshorne}\label{lem:Noetheriancharacterization} Let $X$ be a topological space. If $X$ is Noetherian and Hausdorff, then $X$ must be finite. \end{lemma} \begin{proof} First we show that every subset of $X$ is compact. So if $Y\subseteq X$, let $\mathcal U$ be a cover of $Y$. Let $\mathcal V$ be the set of all finite unions of elements of $\mathcal U$. Then $\mathcal V$ covers $Y$ as well, and moreover, $\mathcal V$ is directed. Since $\O(X)$ is Noetherian, $\mathcal V$ has a greatest element $V$. Now, $V$ contains every element of $\mathcal{V}$ and since $\mathcal{V}$ covers $Y$, we find that $V$ must contain $Y$. It follows from the definition of $\mathcal V$ that $V$ can be written as a finite union of elements of $\mathcal U$, so $\mathcal U$ has a finite subcover. Thus $Y$ is compact. Now let $x\in X$. Then $X\setminus\{x\}$ is compact, hence closed. Hence $\{x\}$ is open, and it follows that $X$ is discrete. Since $X$ itself is compact, $X$ must be finite. \end{proof} \begin{proposition}\label{prop:ArtinianandNoetherianisfinitedimensional} Let $A$ be a C*-algebra. Then $A$ is finite dimensional if and only if $\CC(A)$ is Artinian if and only if $\CC(A)$ is Noetherian. \end{proposition} \begin{proof} Assume that $A$ is finite dimensional. By Lemma \ref{lem:rankfunctiononCA}, $\CC(A)$ has a rank function whose range is bounded from above. By Lemma \ref{lem:gradedimpliesartinian}, $\CC(A)$ is both Artinian and Noetherian. Assume that $A$ is not finite dimensional. By Lemma \ref{lem:maximalelementsofCA}, $\CC(A)$ has a maximal element $M$. By Proposition \ref{prop:finitedimensionalmaximalcommutativesubalgebras}, it follows that $M$ cannot be finite dimensional. Since $M$ is a unital commutative C*-algebra, the Gel'fand-Naimark Theorem assures that $M=C(X)$ for some compact Hausdorff space $X$, which must have an infinite number of points since $A$ is infinite dimensional. We construct a descending chain in $\CC(A)$ as follows. First choose a countable subset $\{x_1,x_2,x_3,\ldots\}$ of $X$. Let $$C_n=\{f\in C(X):f(x_1)=\ldots=f(x_n)\}$$ for each $n\in\N$. Clearly, we have $C_1\supseteq C_2\supseteq C_3\supseteq\ldots$. Assume that $i<j$. Then $\{x_1,\ldots,x_i\}$ and $\{x_j\}$ are disjoint closed sets, hence Urysohn's Lemma assures the existence of some $f\in C(X)$ such that $f[\{x_1,\ldots,x_i\}]=\{1\}$ and $f(x_j)=0$. Clearly, $f\in C_i$, but $f\notin C_j$. This shows that $C_i\neq C_j$, so the chain is descending, but it never stabilizes. We construct an ascending chain in $\CC(A)$ as follows. First we notice that since $X$ is infinite and Hausdorff, Lemma \ref{lem:Noetheriancharacterization} implies that $X$ is not Noetherian. So there is an ascending chain $O_1\subseteq O_2\subseteq\ldots$ of open subsets of $X$ that does not stabilize. For each $i\in\N$, let $F_i=X\setminus O_i$. Then $F_1\supseteq F_2\supseteq\ldots$ is a descending chain of closed subsets of $X$, which does not stabilize. For each $i\in I$ let $C_i=C_{F_i}$. Then $C_i$ is a C*-subalgebra of $C(X)$ and if $i\leq j$, we have $F_i\supseteq F_j$, so $C_i\subseteq C_j$. Moreover, if $i<j$ and $F_i\neq F_j$, then there is some $x\in F_i$ such that $x\notin F_j$. By Urysohn's Lemma, there is an $f\in C(X)$ such that $f(x)=0$ and $f(y)=1$ for each $y\in F_j$. Hence $f\in C_j$, but $f\notin C_i$. It follows that $C_i\neq C_j$, so $C_1\subseteq C_2\subseteq\ldots$ is an ascending chain that does not stabilize. Thus $\CC(A)$ contains an ascending chain as well as a descending chain, neither of which stabilizes. Hence $\CC(A)$ can be neither Noetherian nor Artinian. \end{proof} Recall that the center of a C*-algebra $A$ is the set $$\{x\in A:xy=yx\ \forall y\in A\},$$ which is usually denoted by $Z(A)$. \begin{lemma}\label{lem:centerisintersectionmaximalcommutatives} Let $A$ be a C*-algebra. Then $Z(A)$ is a commutative C*-subalgebra. Moreover, $Z(A)$ is the intersection of all maximal commutative C*-subalgebras of $A$. \end{lemma} \begin{proof} For each $y\in A$, consider the map $f_y:A\to A$ given by the assignment $x\mapsto xy-yx$. Clearly this is continuous and linear, so $\ker f_y$ is a closed linear subspace of $A$. Hence $Z(A)=\bigcap_{y\in A}\ker f_y$ is a closed linear subspace as well. If $x,y\in Z(A)$ and $z\in A$, then $$xyz=xzy=zxy,$$ so $xy\in Z(A)$. Moreover, $x^*z=(z^*x)^*=(xz^*)^*=zx^*$, so $x^*\in Z(A)$. Clearly $xy=yx$, and $1_A\in Z(A)$, hence $Z(A)\in\CC(A)$. Let $x\in\bigcap\max\CC(A)$, i.e., $x\in M$ for each maximal $M\in\CC(A)$. Let $y\in A$. Then $y$ can be written as a linear combination of two self-adjoint elements $a_1,a_2$. If $a\in A$ is self-adjoint, then $C^*(a,1)$ is a commutative C*-subalgebra of $A$ containing $a$. By Lemma \ref{lem:maximalelementsofCA}, it follows that there are $M_1,M_2\in\max\CC(A)$ such that $a_i\in M_i$ for $i=1,2$. Since $x\in M_1,M_2$, it follows that $x$ commutes with both $a_1$ and $a_2$. Hence $x$ commutes with $y$, so $x\in Z(A)$. Thus $\bigcap\max\CC(A)\subseteq Z(A)$. Now assume that $x\in Z(A)$. Since $x$ commutes with all elements of $A$, it commutes in particular with $x^*$. Hence $x$ is normal. We have $x^*\in Z(A)$ as well, for $Z(A)$ is a *-subalgebra of $A$. Let $M\in\max\CC(A)$. Then $M\cup\{x,x^*\}$ is a set of mutually commuting elements, which is *-closed and contains $1_A$. It follows that $C^*(M\cup\{x,x^*\})$, the C*-subalgebra of $A$ generated by $M\cup\{x,x^*\}$, is commutative. Since $M$ is maximal, $C^*(M\cup\{x,x^*\})$ must be equal to $M$. As a consequence, $x\in M$, so we find that $x$ is contained in every maximal commutative C*-subalgebra of $A$. Hence $Z(A)\subseteq\bigcap\max\CC(A)$. \end{proof} \begin{lemma}\label{lem:centerandproducts} Let $A_1,\ldots, A_n$ be C*-algebras. Then $$ Z\left(\bigoplus_{i=1}^nA_i\right) = \bigoplus_{i=1}^nZ(A_i).$$ \end{lemma} \begin{proof} This follows directly from the fact that multiplication on $\bigoplus_{i=1}^nA_i$ is calcultated coordinatewisely. \end{proof} \begin{lemma}\label{lem:upcenterandproducts} Let $A_1,\ldots, A_n$ be C*-algebras. Let $A=\bigoplus_{i=1}^nA_i$ and $C\in\CC(A)$ such that $Z(A)\subseteq C$. Then there are $C_i\in\CC(A_i)$ such that $Z(A_i)\subseteq C_i$ and $C=\bigoplus_{i=1}^nC_i$. \end{lemma} \begin{proof} Let $p_i:A\to A_i$ be the projection on the $i$-th factor. Then we obtain an order morphism $\CC(p_i):\CC(A)\to\CC(A_i)$. Let $C_i=\CC(p_i)(C)$, or equivalently, $C_i=p_i[C]$. Then $$Z(A_i)=p_i\left[\bigoplus_{i=1}^nZ(A_i)\right]=p_i[Z(A)]\subseteq p_i[C]=C_i$$ for each $i\in I$. Let $c\in C$, then $p_i(c)\in C_i$ for each $i=1,\ldots,n$, so $$c=p_1(c)\oplus\ldots\oplus p_n(c).$$ Thus $c\in\bigoplus_{i=1}^nC_i$, hence $C\subseteq\bigoplus_{i=1}^nC_i$. Let $c_1\oplus\ldots \oplus c_n\in\bigoplus_{i=1}^nC_i$. This means that for each $i=1,\ldots,n$ there is a $d^i\in C$ such that $p_i(d^i)=c_i$. Here $d^i=d_1^i\oplus\ldots\oplus d_n^i$, with $d_j^i\in A_j$, and in particular we have $d^i_i=c_i$. For each $j=1,\ldots,n$, let $e^j\in A$ be the element $e^j_1\oplus\ldots\oplus e^j_n$, with $$e^j_i=\begin{cases} 1_{A_i} & i=j;\\ 0_{A_i} & i\neq j. \end{cases}$$ Here $1_{A_i}$ and $0_{A_i}$ denote the unit and the zero of $A_i$, respectively. Since $1_{A_i},0_{A_i}\in Z(A_i)$, Lemma \ref{lem:centerandproducts} assures that $e^j\in Z(A)$. Since $Z(A)\subseteq C$, we find that $e^j\in C$. It follows that $f^j=e^jd^j\in C$. Here $f^j=f^j_1\oplus\ldots\oplus f^j_n$ with $$f^j_i=\begin{cases} c_i & i=j;\\ 0_{A_i} & i\neq j. \end{cases}$$ Now, $$f^1+\ldots +f^n=c_1\oplus\ldots\oplus c_n=c,$$ and since $f^j\in C$, it follows that $c\in C$. So $C=\bigoplus_{i=1}^nC_i$. \end{proof} \begin{proposition}\label{prop:embeddingofCACBintoCAplusB} Let $A_1,\ldots,A_n$ be C*-algebras and let $A=\bigoplus_{i=1}^nA_i$. Define the map $\iota:\prod_{i=1}^n\CC(A_i)\to\CC(A)$ by $\langle C_1,\ldots, C_n\rangle\mapsto C_1\oplus\ldots\oplus C_n$, and let $\pi:\CC(A)\mapsto\prod_{i=1}^n\CC(A_i)$ be the map $\CC(p_1)\times\ldots\times\CC(p_n)$, where $p_i:A\to A_i$ denotes the projection on the $i$-th factor. Thus $\pi$ maps $C\in\CC(A)$ to $\langle p_1[C],\ldots,p_n[C]\rangle$. Then: \begin{enumerate} \item[(i)] $\iota$ is an embedding of posets; \item[(ii)] $\pi$ is surjective; \item[(iii)] $\pi\circ\iota=1_{\prod_{i=1}^n\CC(A_i)}$ and $1_{\CC(A)}\leq \iota\circ\pi$; \item[(iv)] the restriction of $\iota$ to a map $\prod_{i=1}^n\CC(A_i)\to\up Z(A)$ is an order isomorphism with inverse $\pi$. \end{enumerate} \end{proposition} \begin{proof} For each $i=1,\ldots,n$, let $C_i\in\CC(A_i)$. Then $C_1\oplus\ldots\oplus C_n$ is clearly a commutative C*-algebra of $A$, which is unital since $$1_{A}=1_{A_1}\oplus\ldots\oplus 1_{A_n}.$$ Hence the image of $\iota$ lies in $\CC(A)$, so $\iota$ is well defined. Furthermore, we remark that $\CC(p_i):\CC(A)\to\CC(A_i)$ is an order morphism by Lemma \ref{lem:CAisinvariantforA}. \begin{enumerate} \item[(i)] Let $\langle C_1,\ldots, C_n\rangle$ and $\langle D_1,\ldots, D_n\rangle$ be elements of $\prod_{i=1}^n\CC(A_i)$. Then $\langle C_1,\ldots,C_n\rangle\leq \langle D_1,\ldots,D_n\rangle$ implies $C_i\subseteq D_i$ for each $i=1,\ldots, n$. Hence $C_1\oplus\ldots\oplus C_n\subseteq D_1\oplus\ldots\oplus D_n$, which says exactly that $$\iota(\langle C_1,\ldots, C_n\rangle)\subseteq\iota(\langle D_1,\ldots, D_n\rangle).$$ Conversely, if $\iota(\langle C_1,\ldots, C_n\rangle)\subseteq\iota(\langle D_1,\ldots, D_n\rangle)$, we have $$C_1\oplus\ldots\oplus C_n\subseteq D_1\oplus\ldots\oplus D_n.$$ If we let act $\CC(p_i)$ on both sides of this inclusion, we obtain the inclusion $C_i\subseteq D_i$ for each $i=1,\ldots,n$. Hence $$\langle C_1,\ldots, C_n\rangle\leq\langle D_1,\ldots,D_n\rangle.$$ Thus $\iota$ is an embedding of posets. \item[(ii)] Let $\langle C_1,\ldots,C_n\rangle\in\prod_{i=1}^n\CC(A_i)$. If $C=C_1\oplus\ldots\oplus C_n$, then $C\in\CC(A)$ and \begin{eqnarray*} \pi(C) & = & \langle\CC(p_1)(C),\ldots,\CC(p_n)(C)\rangle=\langle p_1[C],\ldots,p_n[C]\rangle\\ & = & \langle C_1,\ldots,C_n\rangle. \end{eqnarray*} \item[(iii)] Let $\langle C_1,\ldots, C_n\rangle\in\prod_{i=1}^n\CC(A_i)$. Let $C=\iota(\langle C_1,\ldots, C_n\rangle)$. Then $C=C_1\oplus\ldots\oplus C_n$, and by the calculation in (ii), we obtain $\pi(C)=\langle C_1,\ldots, C_n\rangle$. Hence $\pi\circ\iota=1_{\prod_{i=1}^n\CC(A_i)}$. Let $C\in\CC(A)$. Then $$\iota\circ\pi(C)=\iota(\langle p_1[C],\ldots,p_n[C]\rangle)=p_1[C]\oplus\ldots\oplus p_n[C].$$ Let $c\in C$. Since $C\subseteq A$, and $A=\bigoplus_{i=1}^nA_i$, we have $$c=c_1\oplus\ldots\oplus c_n,$$ with $c_i\in A_i$ for each $i=1,\ldots,n$. Hence $c_i=p_i(c)$, and we find $c=p_1(c)\oplus\ldots\oplus p_n(c)$, so $c\in p_1[C]\oplus\ldots\oplus p_n[C]$. But $$p_1[C]\oplus\ldots\oplus p_n[C]=\iota(\langle p_1[C],\ldots, p_n[C]\rangle)=\iota\circ\pi(C).$$ Hence $c\in\iota\circ\pi(C)$, so $C\subseteq\iota\circ\pi(C)$. We conclude that the inequality $1_{\CC(A)}\leq\iota\circ\pi$ holds. \item[(iv)] In order to show that $\iota$ restricts to an order isomorphism $$\prod_{i=1}^n\up Z(A_i)\to\up Z(A)$$ with inverse $\pi$, it is enough to show that $\iota\circ\pi(C)=C$ for each $C\in\up Z(A)$. Then the statement follows directly from the equality in (iii). So let $C\in\CC(A)$ such that $Z(A)\subseteq C$. Then Lemma \ref{lem:upcenterandproducts} assures that there are $C_i\in\CC(A_i)$ for each $i=1,\ldots,n$ such that $C=C_1\oplus\ldots\oplus C_n$ and $Z(A_i)\subseteq C_i$ for each $i=1,\ldots,n$. Then $p_i[C]=C_i$, hence \begin{eqnarray*} \iota\circ\pi(C) &= &\iota(\langle p_1[C],\ldots,p_n[C]\rangle)=\iota(\langle C_1,\ldots, C_n\rangle)\\ & = & C_1\oplus\ldots\oplus C_n=C. \end{eqnarray*} \end{enumerate} \end{proof} \begin{proposition}\label{prop:intervalZAkommaM} Let $A=\bigoplus_{i=1}^k\MM_{n_i}(\C)$, where $k,n_1,\ldots,n_k\in\N$. Then $[Z(A),M]\cong\prod_{i=1}^k\CC(\C^{n_i})$ for each $M\in\max\CC(A)$. \end{proposition} \begin{proof} By Proposition \ref{prop:embeddingofCACBintoCAplusB}, there is an order morphism $$\pi:\CC(A)\to \prod_{i=1}^j\CC(\MM_{n_i}(\C))$$ whose restriction to $\up Z(A)$ is an order isomorphism with inverse $\iota$. Let $M\in\max\CC(A)$. By Lemma \ref{lem:centerisintersectionmaximalcommutatives} it follows that $Z(A)\subseteq M$, so $M\in\up Z(A)$, hence $M\in\max\up Z(A)$. Since $\pi$ is an order isomorphism, it follows that $\pi(M)$ is a maximal element of $\prod_{i=1}^k\CC(\MM_{n_i}(\C))$. Clearly there are $M_{n_i}\in\max\CC(\MM_{n_i}(\C))$ for each $i=1,\ldots,k$ such that $\pi(M)= \langle M_{n_1},\ldots, M_{n_k}\rangle$. It follows that $\down \pi(M)=\down M_{n_1}\times\ldots\times\down M_{n_k}$. Since $\iota$ is the inverse of $\pi$ and has codomain $\up Z(A)$, we find that $$\iota[\down\pi(M)]=\down\iota\circ\pi(M)\cap\up Z(A)=\down M\cap\up Z(A)=[Z(A),M].$$ Hence the restriction $\iota:\down M_{n_1}\times\ldots\times\down M_{n_k}\to[Z(A),M]$ is an order isomorphism. Notice that all maximal elements of $\CC(\MM_{n_i}(\C))$ are *-isomorphic by Lemma \ref{lem:dimensionofmaximalsubalgebraofmatrixalgebra}. More specificaly, $M_{n_i}\in\max\CC(\MM_{m_i}(\C))$ is *-isomorphic to $D_{n_i}$. Since $$D_{n_i}=\{\mathrm{diag}(\lambda_1,\ldots,\lambda_{n_i}):\lambda_1,\ldots,\lambda_{n_i}\in\C\},$$ we find that $M_{n_i}$ is *-isomorphic to $\C^{n_i}$. Hence there is an embedding $$f:\C^{n_i}\to \MM_{n_i}(\C)$$ such that $f[C^{n_i}]=M_{n_i}$. By Proposition \ref{prop:Coff}, we find that $$\CC(f):\CC(\C^{n_i})\to\CC(\MM_{n_i}(\C))$$ is an order embedding with image $\down M_{n_i}$. Thus there exists an order isomorphism between $\CC(\C^{n_i})$ and $\down M_{n_i}$ in $\CC(\MM_{n_i}(\C))$. Hence $[Z(A),M]\cong\prod_{i=1}^k\CC(\C^{n_i})$. \end{proof} \begin{definition}\cite[III.8]{Birkhoff} Let $\CC$ be a lattice. Then $\CC$ is called \emph{directly indecomposable}\index{directly indecomposable lattice} if $\CC\cong\CC_1\times\CC_2$ for some lattices $\CC_1,\CC_2$ implies that either $\CC_1=\mathbf{1}$ and $\CC_2=\CC$ or $\CC_1=\CC$ and $\CC_2=\mathbf{1}$, where $\mathbf 1$ denotes the one-point poset. \end{definition} The following result is also known as Hashimoto's Theorem. \begin{theorem}\label{thm:Hashimoto} Let $\CC$ be a lattice with a least element $0$. If there are two direct decompositions of $\CC$ \begin{eqnarray*} \CC & = & \A_1\times\ldots\times \A_n;\\ \CC & = & \B_1\times\ldots\times\B_m, \end{eqnarray*} where $n,m\in\N$, then there are lattices $\CC_{ij}$, $i=1,\ldots,n$ and $j=1,\ldots,m$ such that \begin{eqnarray*} \A_{i} & = & \CC_{i1}\times\ldots\times\CC_{im};\\ \B_j & = & \CC_{1j}\times\ldots\times\CC_{nj}. \end{eqnarray*} \end{theorem} \begin{proof} \cite[Theorem III.4.2]{Gratzer} \end{proof} \begin{corollary}\label{cor:Hashimoto} Let $\A_1\times\ldots\times\A_n=\B_1\times\ldots\times\B_m$, where $n,m\in\N$ and the $\A_i$ and $\B_i$ are directly indecomposable lattices. Then $n=m$, and there is some permutation $$\pi:\{1,\ldots,n\}\to\{1,\ldots,n\}$$ such that $A_i\cong B_{\pi(i)}$ for each $i=1,\ldots,n$. \end{corollary} \begin{definition} Let $\CC$ be a bounded lattice and $C\in\CC$. Then $D\in\CC$ is called a \emph{complement}\index{complement in a lattice} of $C$ if $C\wedge D=0$ and $C\vee D=1$. \end{definition} The next Proposition is actually an application of \cite[Theorem III.4.1]{Gratzer}. \begin{proposition}\label{prop:directlyindecomposableboundedposet} Let $\CC$ be a bounded lattice. If $0$ and $1$ are the only elements of $\CC$ with a unique complement (namely each other), then $\CC$ is directly indecomposable. \end{proposition} \begin{proof} Let $D\in\CC$ be a complement of $1$. Then $D=1\wedge D=0$. If $D$ is a complement of $0$, then $D=0\vee D=1$. Thus $0$ and $1$ are each other's unique complement. Now assume that there exists an order isomorphism $$\phi:\CC_1\times\CC_2\to\CC,$$ where $\CC_1$ and $\CC_2$ are bounded lattices not equal to $\mathbf 1$. This last condition implies that $$\langle 1,1\rangle\neq\langle 1,0\rangle\neq\langle 0,0\rangle.$$ Let $C=\phi(\langle 1,0\rangle)$. Then it follows that $1\neq C\neq 0$. Clearly $\langle 0,1\rangle$ is a complement of $\langle 1,0\rangle$ in $\CC_1\times\CC_2$, but it is also unique. Let $\langle D_1,D_2\rangle$ be a complement of $\langle 0,1\rangle$. Since meets and joins are calculated componentwise, we find \begin{eqnarray*} \langle D_1,0\rangle & = & \langle D_1\wedge 1,D_2\wedge 0\rangle=\langle D_1,D_2\rangle\wedge \langle 1,0\rangle =\langle 0,0\rangle;\\ \langle 1,D_2\rangle & = & \langle D_1\vee 1,D_2\vee 0\rangle=\langle D_1,D_2\rangle\vee \langle 1,0\rangle =\langle 1,1\rangle, \end{eqnarray*} hence $D_1=0$, $D_2=1$. Thus $\langle D_1,D_2\rangle=\langle 0,1\rangle$, whence $\langle 1,0\rangle$ indeed has a unique complement. Since $\phi$ is an order isomorphism, $\phi$ preserves meets and joins, hence $C$ has a complement $D=\phi(\langle 0,1)\rangle$. Now assume that $C$ has another complement $D'$. Since $\phi$ is an order isomorphism, it follows that $\phi^{-1}(D')$ is a complement of $\langle 1,0\rangle$, and by uniqueness of this complement, we obtain $\phi^{-1}(D')=\langle 0,1\rangle$. We find that $$D=\phi(\langle 0,1\rangle)=\phi\circ\phi^{-1}(D')=D'.$$ The statement follows now by contraposition. \end{proof} \begin{proposition}\label{prop:CAisdirectlyindecomposablewhenAfinitedimcomCalg} Let $A$ be a commutative finite-dimensional C*-algebra. Then $\CC(A)$ is a directly indecomposable lattice. \end{proposition} \begin{proof} By Lemma \ref{lem:CAlatticewhenAcommutative}, $\CC(A)$ is a bounded lattice. Let $X$ be the spectrum of $A$. If $X$ is a singleton set, then $A$ is one-dimensional, hence we have $\CC(A)=\{\C1_A\}$, so $\CC(A)=\mathbf 1$, the one-point lattice, and there is nothing to prove. If $X$ is a two-point set, then $\CC(A)=\{A,\C1_A\}$. So $\CC(A)$ contains no other elements than a greatest and a least one, and is therefore certainly directly indecomposable. Assume that $X$ has at least three points. Let $B\in\CC(A)$, assumed not equal to $\C1_A$ or $A$. By Lemma \ref{lem:subalgebraisintersectionofidealalgebras}, we have $B=\bigcap_{x\in X}C_{[x]_B}$. Since $X$ is finite, it follows that $X/\sim_B$ is finite as well. Notice that we cannot have $[x]_B=\{x\}$ for all $x\in X$, otherwise $B=C(X)=A$. Neither can $X/\sim_B$ be a singleton set, since otherwise $B=\C1_A$. For each element $[x]_B$ in $X/\sim_B$, choose a representative $x$. Let $K$ be the set of representatives. Notice that $K$ is not a singleton set, since $X/\sim_B$ contains at least two elements. Also notice that $K$ is not unique, since there is at least one $[x]_B\in X/\sim_B$ containing two or more points. Since $X$ is discrete, it follows that $K$ is closed. Let $f\in B\cap C_K$ and let $x,y\in X$ be points such that $x\neq y$. If $[x]_B=[y]_B$, then $f(x)=f(y)$. If $[x]_B\neq [y]_B$, then there are $x',y'\in K$ such that $x'\in [x]_B$ and $y'\in [y]_B$. Since $f\in C_K$, we find that $f(x')=f(y')$. Since $f\in B$, we obtain $f(x)=f(x')$ and $f(y)=f(y')$. Combining all equalities gives $f(x)=f(y)$. So in all cases, $f(x)=f(y)$. So $f$ must be constant, and we conclude that $B\cap C_K=\C1_A$. Since $\CC(A)$ is a lattice, $B\vee C_K$ exists. Let $f\in C(X)$. Define the map $g:X\to\C$ by $g(x)=f(k)$ if $x\in[k]_B$, where $k\in K$. Notice that is well defined, since $K$ is a collection of representatives. Moreover, since $X$ is discrete, $g$ is continuous, so $g\in C(X)$. By definition, we have $g\in \bigcap_{x\in X}C_{[x]_B}$, so $g\in B$. Let $h=f-g$. Then $h\in C(X)$, and if $k\in K$, we find $h(k)=f(k)-g(k)=0$, so $h$ is constant on $K$. We conclude that $f=g+h$ with $g\in B$ and $h\in C_K$. Hence $A=C(X)=B\vee C_K$. We find that $C_K$ is a complement of $B$. However, $K$ is not unique, and therefore neither is $C_K$. We conclude that $A$ and $\C1_A$ are the only elements with a unique complement, so $\CC(A)$ is indirectly indecomposable. \end{proof} The proof of this proposition is based on the proof of the directly indecomposability of partition lattices in \cite{Sachs}. More can be said about $\CC(A)$ when $A$ is a commutative C*-algebra of dimension $n$, namely that $\CC(A)$ is order isomorphic to the lattice of partitions of the set $\{1,\ldots,n\}$. We refer to \cite{Heunen2} for a complete characterization of $\CC(A)$ when $A$ is a commutative finite-dimensional C*-algebra. We are now ready to prove the main result of this section. \begin{proof}[Proof of Theorem \ref{thm:fdmaintheorem}] Let $A$ be a finite-dimensional C*-algebra, and $B$ a C*-algebra. Let $\phi:\CC(A)\to\CC(B)$ an order isomorphism. By Proposition \ref{prop:ArtinianandNoetherianisfinitedimensional}, $\CC(A)$ is Noetherian, and so $\CC(B)$ must be Noetherian as well. Hence Proposition \ref{prop:ArtinianandNoetherianisfinitedimensional} implies that $B$ is finite dimensional. It follows from Lemma \ref{lem:rankfunctiononCA} that both $\CC(A)$ and $\CC(B)$ have a rank function assigning to each element its dimension. By Lemma \ref{lem:uniquenessrankfunction} the rank function is unique, hence it follows from Lemma \ref{lem:gradedposetunderorderiso} that $\dim(\phi(C))=\dim(C)$ for each $C\in\CC(A)$. Therefore, we can reconstruct the dimensions of elements of $\CC(A)$ and $\CC(B)$, and the dimension is preverved by $\phi$. By the Artin-Wedderburn Theorem, there are unique $k,k'\in\N$ and unique $\{n_i\}_{i=1}^n,\{n_i'\}_{i=1}^{k'}$ with $n_i,n'_i\in\N$ such that \begin{eqnarray*} A & \cong & \bigoplus_{i=1}^k\MM_{n_i}(\C);\\ B & \cong & \bigoplus_{i=1}^{k'}\MM_{n'_i}(\C). \end{eqnarray*} Without loss of generality, we may assume that the $n_i$ and $n'_i$ form an descending (but not necessarily strictly descending) finite sequence. By Lemma \ref{lem:centerisintersectionmaximalcommutatives}, we have the equalities $Z(A)=\bigcap\max\CC(A)$ and $\bigcap\max\CC(B)=Z(B)$. Since the intersection is the meet operation in $\CC(A)$ and $\CC(B)$, and order isomorphism preserve both meets and maximal elements, we find that $\phi(Z(A))=Z(B)$, so $\dim(Z(A))=\dim(Z(B))$. Using Lemma \ref{lem:centerandproducts}, we find that $Z(A)=\bigoplus_{i=1}^nZ(\MM_{n_i}(\C))$, and since the dimension of the center of a matrix algebra is $1$, we find that $\dim Z(A)=k$. In the same way, we find that $\dim Z(B)=k'$, so we must have $k=k'$. Let $M\in\max\CC(A)$. Then $\phi(M)$ is a maximal element of $\CC(B)$, and since $\phi(Z(A))=Z(B)$, we find that $\phi$ restricts to an order isomorphism $[Z(A),M]\to [Z(B),\phi(M)]$. By Proposition \ref{prop:intervalZAkommaM}, we obtain an order isomorphism $$\prod_{i=1}^k\CC(\C^{n_i})\cong\prod_{i=1}^k\CC(\C^{n'_i}).$$ It is possible that for some $i$ we have $n_i=1$, in which case we have $\CC(\C^{n_i})=\mathbf{1}$. Since we assumed that $\{n_i\}_{i=1}^n$ is a descending sequence, there is a greatest number $r$ below $k$ such that $n_r\neq 1$. Likewise, let $s$ be the greatest number such that $n'_{s}\neq 1$. Then we obtain an order isomorphism $$\prod_{i=1}^{r}\CC(\C^{n_i})\cong\prod_{i=1}^{s}\CC(\C^{n'_i}).$$ By Proposition \ref{prop:CAisdirectlyindecomposablewhenAfinitedimcomCalg} and Corollary \ref{cor:Hashimoto}, we now find $r=s$, and there is a permutation $\pi:\{1,\ldots,r\}\to\{1,\ldots,r\}$ such that $\CC(\C^{n_i})\cong\CC(\C^{n'_{\pi(i)}})$ for each $i\in\{1,\ldots,r\}$. Let $\psi_i:\CC(\C^{n_i})\to\CC(\C^{n'_{\pi(i)}})$ be the accompanying order isomorphism. Lemma \ref{lem:rankfunctiononCA} assures that the function assigning to each element of $\CC(\C^{n_i})$ its dimension is a rank function, and similarly the dimension function is a rank function for $\CC(\C^{n'_{\pi(i)}})$. By Lemma \ref{lem:gradedposetunderorderiso}, we find that $\dim(C)=\dim(\psi_i(C))$ for each $C\in\CC(\C^{n_{i}})$. Hence $$n_i=\dim(\C^{n_i})=\dim\left(\psi_i\left(\C^{n_i}\right)\right)=\dim\left(\C^{n'_{\pi(i)}}\right)=n'_{\pi(i)},$$ where the fact that order isomorphisms map greatest elements to greatest elements is used in the third equality. By definition of $r$, we must have $n_i=n_i'=1$ for all $i\geq r$. Hence we can extend $\pi$ to a permutation $\{1,\ldots,k\}\to\{1,\ldots, k\}$ by setting $\pi(i)=i$ for each $i\geq r$. Hence $k=k'$ and $\{n_1,\ldots,n_k\}$ and $\{n_1',\ldots,n'_k\}$ are the same sets up to permutation. We conclude that $A$ and $B$ must be *-isomorphic. \end{proof} We note that since the class of all finite-dimensional C*-algebras and the class of all finite-dimensional von Neumann algebras are the same, a similar statement holds for the functor $\V$ assigning to a von Neumann algebra $M$ the poset $\V(M)$ of its commutative von Neumann subalgebras. Thus if $M$ and $N$ are von Neumann algebras such that $M$ is finite-dimensional, then $\V(M)\cong\V(N)$ implies $M\cong N$. If $A$ is a finite-dimensional C*-algebra and $B$ is a C*-algebra such that there is an order isomorphism $\phi:\CC(A)\to\CC(B)$, then it might be the case that even though $A$ and $B$ are *-isomorphic, we have $\phi=\CC(f)$ for more than one *-isomorphism $f:A\to B$. For instance, let $A=B=\C^2$. Let $f:\C^2\to\C^2$ be given by $f(\langle a,b\rangle)=\langle b,a\rangle$. Then both $\CC(f)=\CC(1_{\C^2})$. It might even be the case that $\phi\neq\CC(f)$ for each *-isomorphism $f:A\to B$. For instance, let $A=B=\MM_{2}(\C)$. Then $$\CC(\MM_2(\C))=\{\C1_{\MM_{2}(\C)}\}\cup\{uD_2u^*:u\in\mathrm{SU}(2)\},$$ where $D_2=\{\mathrm{diag}(\lambda_1,\lambda_2):\lambda_1,\lambda_2\in\C\}$. Furthermore, one can show that each *-isomorphism $f:\MM_2(\C)\to\MM_2(\C)$ is of the form $a\mapsto uau^{-1}$ for some $u\in \mathrm{U}(2)$ \cite[Theorem 4.27]{AS1}. Hence $\CC(f):\CC(\MM_2(\C))\to\CC(\MM_2(\C))$ is given by $C\mapsto uCu^*$ for some $u\in\mathrm{U}(2)$. Choose $v\in\mathrm{U}(2)$ such that $D_2\neq vD_2v^*$, and let $$\phi:\CC(\MM_2(\C))\to\CC(\MM_2(\C))$$ be defined by $\phi(D_2)=vD_2v^*$, $\phi(vD_2v^*)=D_2$, and $\phi(C)=C$ for all other $C\in\CC(\MM_2(\C))$. Then $\phi$ is clearly an order isomorphism. However, $\phi\neq\CC(f)$ for each *-isomorphism\\ $f:\MM_2(\C)\to \MM_2(\C)$. \section*{Outlook} We have shown that $\CC(A)$ is a complete invariant for finite-dimensional C*-algebras, whereas Mendivil and Hamhalter showed that $\CC(A)$ completely determine commutative C*-algebras. The question is whether there are more classes of C*-algebras which can be classified by $\CC(A)$. An interesting class might be that of \emph{AF-algebras}, i.e., C*-algebras $A$ that can be approximated by finite-dimensional C*-algebras. Usually one considers only separable AF-algebras, which are C*-algebras $A$ such that $A=\overline{\bigcup_{i=1}^\infty A_i}$, where $A_1\subseteq A_2\subseteq\ldots$ is an ascending chain of finite-dimensional C*-subalgebras of $A$. It is well known that this class of AF-algebras can be classified by Bratteli diagrams \cite{BratteliInductiveLimits} and by K-theory \cite{Elliott}. One could also look at C*-algebras $A$ such that $A=\overline{\bigcup\D}$ for some directed set $\D$ consisting of finite-dimensional C*-subalgebras of $A$. In this case $A$ need not be separable, and therefore C*-algebras of these form are called \emph{non-separable} AF-algebras. It turns out that neither Bratteli diagrams nor K-theory can completely classify this class of C*-algebras \cite{FK}, \cite{Katsura}. However, as one might have noticed, the framework of $\CC(A)$ might be suitable in order to classify non-separable AF-algebras if one compares the definition of non-separable AF-algebras with the content of Proposition \ref{prop:CAisdcpo}. If this is indeed the case, then $\CC(A)$ might be an interesting alternative for K-theory. Since $\CC(A)$ is a dcpo, and domain theory (see for instance \cite{CLD}) deals with various properties of dcpos, a first step might be the study of the domain-theoretical properties of $\CC(A)$. For details, we refer to \cite{HL}. It might be interesting to compare the domain-theoretical properties of $\CC(A)$ with those of $\V(M)$, the poset $\V(M)$ of commutative von Neumann subalgebras of a von Neumann algebra $M$. For the von Neumann case, we refer to \cite{DRSB}. It might be interesting to look at non-unital C*-algebras as well. The reason why we did not consider non-unital C*-algebras lies within quantum toposophy, from which this research evolved. In quantum toposophy one is forced to work constructively; and whereas constructive Gel'fand duality for unital commutative C*-algebras holds (see for instance \cite{BM} and \cite{CS}), it was not known yet whether the non-unital version holds as well. However, Henry recently proved a non-unital version of constructive Gel'fand duality \cite{Henry}, which suggests that non-unital C*-algebras can be incorporated within quantum toposophy as well. In the non-unital case one could proceed as follows. If $\mathbf{CStar}$ denotes the category of C*-algebras with *-homomorphisms as morphisms, we can define the functor $\CC_0:\mathbf{CStar}\to\mathbf{Poset}$ as follows. Given a C*-algebra $A$, we denote the poset of commutative C*-algebras by $\CC_0(A)$, and if $f:A\to B$ is a *-homomorphism, $\CC_0(f):\CC_0(A)\to\CC_0(B)$ is defined by $C\mapsto f[C]$. The functor $\CC_0$ shares some properties with $\CC$, for instance Proposition \ref{prop:Coff} holds as well if we replace $\CC$ by $\CC_0$. It is even the case that we can describe injectivity of a *-homomorphism $f:A\to B$ completely in order theoretic properties of $\CC_0(f)$. This is possible, since $\z$, the C*-algebra consisting of only one element $0$, is always an element of $\CC_0(A)$. Hence $f:A\to B$ is injective if and only if $\CC_0(f):\CC_0(A)\to\CC_0(B)$ has an upper adjoint $\CC_0(f)_*:\CC_0(B)\to\CC_0(A)$ such that $\CC_0(f)_*(\z)=\z$. The latter equality translates to $f^{-1}[\{0\}]=\{0\}$, which exactly states that $f$ is injective. We expect that Theorem \ref{thm:fdmaintheorem} holds as well if we replace $\CC$ by $\CC_0$. Some minor details in the proofs must be adjusted, but we expect that most lemmas still hold, since each finite-dimensional C*-algebra $A$ is automatically unital, hence $\CC(A)$ can be regarded as subposet of $\CC_0(A)$. However, it might be difficult to prove a non-unital version of Mendivil and Hamhalter's theorem to the effect that $\CC_0(A)$ determines a commutative C*-algebra $A$ up to *-isomorphism, since it is desirable that we can identify C*-ideals of $A$ as elements of $\CC_0(A)$ in order to reconstruct $A$, and it is not clear how to make this identification. This is already visible if we consider $\CC_0(\C^2)=\{\z,C_1,C_2,C_3,\C^2\}$, where \begin{eqnarray*} C_1 & = & \{\langle \mu,0\rangle:\mu\in\C\},\\ C_2 & = & \{\langle 0,\nu\rangle:\nu\in\C\};\\ C_3 & = & \{\langle\lambda,\lambda\rangle:\lambda\in\C\}. \end{eqnarray*} The least element and the greatest element of $\CC_0(\C^2)$ are $\z$ and $\C^2$, respectively, and $C_1,C_2,C_3$ are mutually incomparable. Here $C_1$ and $C_2$ are the only elements that correspond to ideals of $\C^2$, but it is not possible to distinguish them from $C_3$ in an order theoretical way. Thus $\CC_0(A)$ has some advantages as well as disadvantages with respect to $\CC(A)$. If $A$ is unital, it could be useful to consider both posets at the same time. In this case, $\CC(A)$ can be considered a subposet of $\CC_0(A)$. It might be interesting to remark that in quantum toposophy, a pair $(\CC,\D)$ of a poset $\CC$ and a subposet $\D$ of $\CC$ exactly corresponds to a \emph{site} $(\CC,J)$, i.e., a poset $\CC$ equipped with a \emph{Grothendieck topology}, such that the category $\mathrm{Sh}(\CC,J)$ of $J$-sheaves is equivalent to $\mathrm{Sets}^{\D^\op}$. Hence if $A$ is unital, then the pair $(\CC_0(A),\CC(A))$ corresponds to a site $(\CC_0(A),J)$ such that $\mathrm{Sh}(\CC_0(A),J)\cong\mathrm{Sets}^{\CC(A)^\op}$. Since one usually studies the topos $\mathrm{Sets}^{\CC(A)^\op}$, it follows that one can integrate $\CC_0(A)$ in an elegant way in the usual framework of quantum toposophy. For details on Grothendieck topologies and sheaves on posets, we refer to \cite{Me}.

https://arxiv.org/abs/0710.3206

Generalized Jacquet modules of parabolic induction

In this paper we study a generalization of the Jacquet module of a parabolic induction and construct a filtration on it. The successive quotient of the filtration is written by using the twisting functor.

\section{Introduction} The Jacquet module of a representation of a semisimple (or reductive) Lie group is introduced by Casselman~\cite{MR562655}. One of the motivation of considering the Jacquet module is to investigate homomorphisms to principal series representations, which is an important invariant of a representation. One of the powerful tools to study the Jacquet module of a parabolic induction is the Bruhat filtration~\cite{MR1767896}. This is a filtration on the Jacquet module defined from the Bruhat decomposition. Casselman-Hecht-Milicic~\cite{MR1767896} use the Bruhat filtration to determine the dimension of the (moderate-growth) Whittaker model of a principal series representation (another proof for Kostant's result). In this paper, we study the Bruhat filtration and show that the successive quotient is described using the twisting functor defined by Arkhipov~\cite{MR2074588}. If a principal series representation has the unique Langlands quotient, then the successive quotient is the induction from the Jacquet module of a smaller group (a Levi part of a parabolic subgroup). However, in a general case, it becomes ``twisted'' induction, which has the same character as that of an induced representation but has a different module structure. Moreover, we investigate its generalization, this is related to the Whittaker model. In \cite{MR562655}, Casselman suggested to generalize the notion of the Jacquet module. For this generalized Jacquet module, we can also define a Bruhat filtration and the successive quotient of the resulting filtration is described in terms of the generalized twisting functor. This result gives a strategy to determine all Whittaker models of a parabolic induction. To determine it, it is sufficient to study the successive quotients and extensions of the filtration. In a special case, we can carry out these steps. Now we state our results precisely. Let $G$ be a connected semisimple Lie group, $G = KA_0N_0$ an Iwasawa decomposition and $P_0 = M_0A_0N_0$ a minimal parabolic subgroup and its Langlands decomposition. As usual, the complexification of the Lie algebra is denoted by the corresponding German letter (for example, $\mathfrak{g} = \Lie(G)\otimes_\mathbb{R}\mathbb{C}$). Fix a character $\eta$ of $N_0$. Then for a representation $V$ of $G$, the generalized Jacquet modules $J'_\eta(V)$ and $J^*_\eta(V)$ are defined as follows. \begin{defn}\label{defn:Jacquet modules in Introduction} Let $V$ be a finite-length moderate growth Fr\'echet representation of $G$ (See Casselman~\cite[pp.~391]{MR1013462}). We define $\mathfrak{g}$-modules $J'_\eta(V)$ and $J^*_\eta(V)$ by \begin{align*} J'_\eta(V) & = \left\{v\in V'\Bigm| \begin{array}{l} \text{For some $k$ and for all $X\in\mathfrak{n}_0$,}\\ \text{$(X - \eta(X))^kv = 0$} \end{array} \right\},\\ J^*_\eta(V) & = \left\{v\in (V_{\textnormal{$K$-finite}})^*\Bigm| \begin{array}{l} \text{For some $k$ and for all $X\in\mathfrak{n}_0$,}\\ \text{$(X - \eta(X))^kv = 0$} \end{array} \right\}, \end{align*} where $V'$ is the continuous dual of $V$. \end{defn} Let $W$ be the little Weyl group of $G$ and take $w\in W$. Then the generalized twisting functor $T_{w,\eta}$ is defined as follows. Let $\overline{\mathfrak{n}_0}$ be the nilradical of the opposite parabolic subalgebra to $\mathfrak{p}_0$ and $e_1,\dots,e_l$ be a basis of $\Ad(w)\overline{\mathfrak{n}}_0\cap \mathfrak{n}_0$ such that each $e_i$ is a root vector with respect to $\mathfrak{h}$ where $\mathfrak{h}$ is a Cartan subalgebra of $\mathfrak{g}$ which contains $\mathfrak{a}_0$. Moreover, we choose $e_i$ such that $\bigoplus_{i\le j - 1}\mathbb{C} e_i$ is an ideal of $\bigoplus_{i\le j}\mathbb{C} e_i$ for all $j$. Let $U(\mathfrak{g})$ be the universal enveloping algebra of $\mathfrak{g}$ and $U(\mathfrak{g})_{e_i - \eta(e_i)}$ the localization of $U(\mathfrak{g})$ by a multiplicative set $\{(e_i - \eta(e_i))^n\mid n\in\mathbb{Z}_{>0}\}$. Put $S_{w,\eta} = (U(\mathfrak{g})_{e_i - \eta(e_i)}/U(\mathfrak{g}))\otimes_{U(\mathfrak{g})}\dotsb\otimes_{U(\mathfrak{g})}(U(\mathfrak{g})_{e_l - \eta(e_l)}/U(\mathfrak{g}))$. Then $S_{w,\eta}$ is a $\mathfrak{g}$-bimodule. The twisting functor $T_{w,\eta}$ is defined by $T_{w,\eta}V = S_{w,\eta}\otimes_{U(\mathfrak{g})}(wV)$ where $wV$ is a representation twisted by $w$ (i.e., $Xv = \Ad(w)^{-1}(X)\cdot v$ for $X\in \mathfrak{g}$ and $v\in wV$ where dot means the original action). Let $P$ be a parabolic subgroup containing $A_0N_0$ and take a Langlands decomposition $P = MAN$ such that $A_0\supset A$. Define $\rho_0\in\mathfrak{a}_0^*$ by $\rho_0(H) = (1/2)\Tr \ad(H)|_{\mathfrak{n}_0}$. Let $\rho$ be a restriction of $\rho_0$ on $\mathfrak{a}$. An element of $\mathfrak{a}^*$ corresponds to a character of $A$. We denote the corresponding character to $\lambda + \rho$ by $e^{\lambda + \rho}$ for $\lambda\in\mathfrak{a}^*$. Then for an irreducible representation $\sigma$ of $M$ and $\lambda\in\mathfrak{a}^*$, the parabolic induction $\Ind_P^G(\sigma\otimes e^{\lambda + \rho})$ is defined. Let $W_M$ be the little Weyl group of $M$. Define a subset $W(M)$ of $W$ by $W(M) = \{w\in W\mid \text{for all positive restricted root $\alpha$ of $M$, $w(\alpha)$ is positive}\}$. Then $W(M)$ is a complete representatives of $W/W_M$ and parameterizes $N_0$-orbits in $G/P$. For $w\in W$, fix a lift in $G$ and denote it by the same letter $w$. Enumerate $W(M) = \{w_1,\dots,w_r\}$ so that $\bigcup_{j\le i}N_0w_jP/P$ is a closed subset of $G/P$. Using the $C^\infty$-realization of a parabolic induction, we can regard an element of $J'_\eta(\Ind_P^G(\sigma\otimes e^{\lambda + \rho}))$ as a distribution on $G/P$. Then the Bruhat filtration $I_i\subset J'_\eta(\Ind_P^G(\sigma\otimes e^{\lambda + \rho}))$ is defined by \[ I_i = \left\{x\in J'_\eta(\Ind_P^G(\sigma\otimes e^{\lambda + \rho}))\biggm| \supp x \subset \bigcup_{j\le i}N_0w_jP\right\}. \] Since $w_i\in W(M)$, we have $\Ad(w_i)(\mathfrak{m}\cap \mathfrak{n}_0)\subset \mathfrak{n}_0$. Hence we can define a character $w_i^{-1}\eta$ of $\mathfrak{m}\cap \mathfrak{n}_0$ by $(w_i^{-1}\eta)(X) = \eta(\Ad(w_i)X)$. Using this character, we can define an $\mathfrak{m}\oplus\mathfrak{a}$-module $J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda + \rho})$. Then we have the following theorem. \begin{thm}[Theorem~\ref{thm:succ quot is I'_i}, Theorem~\ref{thm:structure of I_i/I_{i - 1}}]\label{thm:main theorem} The filtration $\{I_i\}$ has the following properties. \begin{enumerate} \item If the character $\eta$ is not unitary, then $J'_\eta(\Ind_P^G(\sigma\otimes e^{\lambda+\rho})) = 0$. \item Assume that $\eta$ is unitary. The module $I_i/I_{i - 1}$ is nonzero if and only if $\eta$ is trivial on $w_iNw_i^{-1}\cap N_0$ and $J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})\ne 0$. \item If $I_i/I_{i - 1} \ne 0$ then $I_i/I_{i - 1} \simeq T_{w_i,\eta}(U(\mathfrak{g})\otimes_{U(\mathfrak{p})}J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho}))$ where $\mathfrak{n}$ acts $J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$ trivially. \end{enumerate} \end{thm} Under the assumptions that $P$ is a minimal parabolic subgroup, $\sigma$ is the trivial representation, $\Ind_P^G(\sigma\otimes e^{\lambda + \rho})$ has the unique Langlands quotient and $\eta$ is the trivial representation, this theorem is proved in \cite{abe-2006}. The proof we give in \cite{abe-2006} is algebraic, while we give an analytic and geometric proof in this paper. For a module $J^*_\eta(\Ind_P^G(\sigma\otimes e^{\lambda + \rho}))$, we have the following theorem. We define two functors. For a $U(\mathfrak{g})$-module $V$, put $C(V) = ((V^*)_{\text{$\mathfrak{h}$-finite}})^*$ and $\Gamma_\eta(V) = \{v\in V\mid \text{for some $k$ and for all $X\in\mathfrak{n}_0$, $(X - \eta(X))^kv = 0$}\}$. \begin{thm}[Theorem~\ref{thm:stucture of J^*(I(sigma,lambda))}]\label{thm:main theorem2} There exists a filtration $0 = \widetilde{I_0}\subset \widetilde{I_1}\subset\cdots\subset\widetilde{I_r} = J^*_\eta(\Ind_P^G(\sigma\otimes\lambda))$ such that $\widetilde{I_i}/\widetilde{I_{i - 1}}\simeq \Gamma_\eta(C(T_{w_i}(U(\mathfrak{g})\otimes_{U(\mathfrak{p})}J^*(\sigma\otimes e^{\lambda+\rho}))))$ where $\mathfrak{n}$ acts $J^*(\sigma\otimes e^{\lambda+\rho})$ trivially. \end{thm} We state an application. The space of Whittaker vectors $\Wh_\eta(D)$ is defined by $\Wh_\eta(D) = \{x\in D\mid \text{$(X - \eta(X))x = 0$ for all $X\in \mathfrak{n}_0$}\}$ for a $U(\mathfrak{g})$-module $D$. If $V$ is a moderate-growth Fr\'echet representation of $G$, an element of $\Wh_\eta(V')$ corresponds to a moderate-growth homomorphism $V\to \Ind_{N_0}^G\eta$ and an element of $\Wh_\eta((V_{\text{$K$-finite}})^*)$ corresponds to an algebraic homomorphism $V_{\text{$K$-finite}}\to \Ind_{N_0}^G\eta$. In particular, when $\eta$ is the trivial representation, these correspond to homomorphisms to principal series representations. Let $\Sigma$ (resp.\ $\Sigma_M$) be the restricted root system for $(G,A_0)$ (resp.\ $(M,M\cap A_0)$), $\Sigma^+$ a positive system of $\Sigma$ corresponding to $N_0$ and $\Pi\subset\Sigma$ the set of simple roots determined by $\Sigma^+$. Put $\Sigma_M^+ = \Sigma_M\cap \Sigma^+$. Let $\widetilde{W}$ (resp.\ $\widetilde{W_M}$) be the (complex) Weyl group of $\mathfrak{g}$ (resp.\ $\mathfrak{m}$). Let $\widetilde{\mu}\in\mathfrak{h}^*$ be the infinitesimal character of $\sigma$. Let $\Delta$ be the root system of $(\mathfrak{g},\mathfrak{h})$. Put $\Sigma_\eta^+ = (\sum_{\eta|_{\mathfrak{g}_\beta} \ne 0,\ \beta\in\Pi}\mathbb{Z}\beta)\cap \Sigma^+$. Fix a $W$-invariant bilinear form $\langle\cdot,\cdot\rangle$ of $\mathfrak{a}_0$. Using the direct decompositions $(\mathfrak{m}\cap\mathfrak{a}_0)^*\oplus \mathfrak{a}^* = \mathfrak{a}_0^*$ and $\mathfrak{a}_0^*\oplus(\mathfrak{h}\cap \mathfrak{m}_0)^* = \mathfrak{h}^*$, we regard $\mathfrak{a}^*\subset \mathfrak{a}_0^*\subset \mathfrak{h}^*$. Recall that $\nu\in(\mathfrak{m}\cap\mathfrak{a}_0)^*$ is called an exponent of $\sigma$ if $\nu + \rho_0|_{\mathfrak{m}\cap\mathfrak{a}_0}$ is an $(\mathfrak{m}\cap \mathfrak{a}_0)$-weight of $\sigma/(\mathfrak{m}\cap \mathfrak{n}_0)\sigma$. We prove the following theorem. \begin{thm}[Theorem~\ref{thm:dimension Whittaker vectors}, Theorem~\ref{thm:dimension Whittaker vectors, algebraic}]\label{thm:main theorem3, dimension of the Whittaker vectors} For $\lambda\in\mathfrak{a}^*$ and an irreducible representation $\sigma$ of $M$, the following formulae hold. \begin{enumerate} \item Assume that for all $w\in W$ such that $\eta|_{wNw^{-1}\cap N_0} = 1$, the following two conditions hold: \begin{enumerate} \item For each exponent $\nu$ of $\sigma$ and $\alpha\in \Sigma^+\setminus w^{-1}(\Sigma^+_M\cup\Sigma_\eta^+)$, we have $2\langle\alpha,\lambda+\nu\rangle/\lvert\alpha\rvert^2\not\in\mathbb{Z}_{\le 0}$. \item For all $\widetilde{w}\in\widetilde{W}$, we have $\lambda - \widetilde{w}(\lambda + \widetilde{\mu})|_\mathfrak{a}\notin \mathbb{Z}_{\le 0}((\Sigma^+\setminus \Sigma_M^+)\cap w^{-1}\Sigma^+)|_\mathfrak{a}\setminus\{0\}$. \end{enumerate} Then we have \begin{multline*} \dim\Wh_\eta((\Ind_P^G(\sigma\otimes e^{\lambda+\rho}))')\\ = \sum_{w\in W(M),\ \eta|_{wNw^{-1}\cap N_0} = 1}\dim \Wh_{w^{-1}\eta}(\sigma'). \end{multline*} \item Assume that for all $\widetilde{w}\in\widetilde{W}\setminus\widetilde{W_M}$ we have $(\lambda + \widetilde{\mu}) - \widetilde{w}(\lambda + \widetilde{\mu}) \not\in\mathbb{Z}\Delta$. Then we have \begin{multline*} \dim\Wh_\eta((\Ind_P^G(\sigma\otimes e^{\lambda+\rho})_{\text{\normalfont $K$-finite}})^*)\\ = \sum_{w\in W(M)}\dim \Wh_{w^{-1}\eta}((\sigma_{\text{\normalfont $M\cap K$-finite}})^*). \end{multline*} \end{enumerate} \end{thm} In the case that $\sigma$ is finite-dimensional, we have the following theorem, which have been announced by T. Oshima (cf.\ his talk at National University of Singapore, January 11, 2006). Let $\Delta_M$ be the root system for $(\mathfrak{m}\oplus\mathfrak{a},\mathfrak{h})$ and take a positive system $\Delta_M^+$ compatible with $\Sigma^+_M$. Put $\widetilde{\rho_M} = (1/2)\sum_{\alpha\in\Delta_M^+}\alpha$. For subsets $\Theta_1,\Theta_2$ of $\Pi$, put $\Sigma_{\Theta_i} = \mathbb{Z}\Theta_i\cap \Sigma$, $W(\Theta_i) = \{w\in W\mid w(\Theta_i)\subset \Sigma^+\}$, $W_{\Theta_i}$ the Weyl group of $\Sigma_{\Theta_i}$ and $W(\Theta_1,\Theta_2) = \{w\in W(\Theta_1)\cap W(\Theta_2)^{-1}\mid w(\Sigma_{\Theta_1})\cap \Sigma_{\Theta_2} = \emptyset\}$. The parabolic subgroup $P$ defines a subset of $\Pi$. We denote this set by $\Theta$. \begin{thm}\label{thm:main theorem4, dimension of the Whittaker vectors, finite-dimensional case} Assume that $\sigma$ is an irreducible finite-dimensional representation with highest weight $\widetilde{\nu}$. Let $\dim_M(\lambda+\widetilde{\nu})$ be the dimension of a finite-dimensional irreducible representation of $M_0A_0$ with highest weight $\lambda+\widetilde{\nu}$. \begin{enumerate} \item Let $\widetilde{\nu}$ be the highest weight of $\sigma$. Assume that for all $w\in W$ such that $\eta|_{wN_0w^{-1}\cap N_0} = 1$ the following two conditions hold: \begin{enumerate} \item For all $\alpha\in \Sigma^+\setminus w^{-1}(\Sigma^+_M\cup\Sigma_\eta^+)$ we have $2\langle\alpha,\lambda+w_0\widetilde{\nu}\rangle/\lvert\alpha\rvert^2\not\in\mathbb{Z}_{\le 0}$. \item For all $\widetilde{w}\in\widetilde{W}$ we have $\lambda - \widetilde{w}(\lambda + \widetilde{\nu} + \widetilde{\rho_M})|_\mathfrak{a}\notin \mathbb{Z}_{\le 0}((\Sigma^+\setminus \Sigma_M^+)\cap w^{-1}\Sigma^+)|_\mathfrak{a}\setminus\{0\}$. \end{enumerate} Then we have \[ \dim \Wh_\eta(I(\sigma,\lambda)') = \# W(\supp\eta,\Theta)\times(\dim_M(\lambda+\widetilde{\nu})) \] \item Assume that for all $\widetilde{w}\in\widetilde{W}\setminus\widetilde{W_M}$, $(\lambda+\widetilde{\nu}) - \widetilde{w}(\lambda+\widetilde{\nu}) \not\in\Delta$. Then we have \begin{multline*} \dim\Wh_\eta((I(\sigma,\lambda)_{\text{\normalfont $K$-finite}})^*) \\= \# W(\supp\eta,\Theta)\times \#W_{\supp\eta}\times(\dim_M(\lambda+\widetilde{\nu})) \end{multline*} \end{enumerate} \end{thm} We summarize the content of this paper. In \S\ref{sec:Parabolic induction and Bruhat filtration}, we introduce the Bruhat filtration. From \S\ref{sec:Parabolic induction and Bruhat filtration} to \S\ref{sec:The module I_i/I_i-1} we study the module $J'_\eta(\Ind_P^G(\sigma\otimes\lambda))$. In \S\ref{sec:vanishing theorem} we prove the successive quotient is zero under some conditions. The structure of the successive quotients is investigated in \S\ref{sec:Analytic continuation}. We give the definition and properties of the generalized twisting functor in \S\ref{sec:Twisting functors} and, in \S\ref{sec:The module I_i/I_i-1} we reveal the relation between the twisting functor and the successive quotient. We complete the proof of Theorem~\ref{thm:main theorem} in this section. Theorem~\ref{thm:main theorem2} is proved in \S\ref{sec:the module J^*_eta(I(sigma,lambda))}. In \S\ref{sec:Whittaker vectors}, the dimension of the space of Whittaker vectors is determined and Theorem~\ref{thm:main theorem3, dimension of the Whittaker vectors} and Theorem~\ref{thm:main theorem4, dimension of the Whittaker vectors, finite-dimensional case} are proved. \subsection*{Acknowledgments} The author is grateful to his advisor Hisayosi Matumoto for his advice and support. He is supported by the Japan Society for the Promotion of Science Research Fellowships for Young Scientists. \subsection*{List of Symbols} \listofsymbols \subsection*{Notation} Throughout this paper we use the following notation. As usual we denote the ring of integers, the set of non-negative integers, the set of positive integers, the real number field and the complex number field by $\mathbb{Z},\mathbb{Z}_{\ge 0},\mathbb{Z}_{> 0},\mathbb{R}$ and $\mathbb{C}$, respectively. Let $G$ be a connected semisimple Lie group and $\mathfrak{g}$ the complexification of its Lie algebra. Fix a Cartan involution $\theta$ of $G$ and denote its derivation by the same letter $\theta$. Let $\mathfrak{g} = \mathfrak{k}\oplus \mathfrak{s}$ be the decomposition of $\mathfrak{g}$ into the $+1$ and $-1$ eigenspaces for $\theta$. Set $K = \{g\in G\mid \theta(g) = g\}$. Let $P_0 = M_0A_0N_0$ be a minimal parabolic subgroup and its Langlands decomposition such that $M_0\subset K$ and $\Lie(A_0)\subset \mathfrak{s}$. Denote the complexification of the Lie algebra of $P_0,M_0,A_0,N_0$ by $\mathfrak{p}_0,\mathfrak{m}_0,\mathfrak{a}_0,\mathfrak{n}_0$, respectively. Take a parabolic subgroup $P$ which contains $P_0$ and denote its Langlands decomposition by $P = MAN$. Here we assume $A\subset A_0$. Let $\mathfrak{p},\mathfrak{m},\mathfrak{a},\mathfrak{n}$ be the complexification of the Lie algebra of $P,M,A,N$. Put $\overline{P_0} = \theta(P_0)$, $\overline{N_0} = \theta(N_0)$, $\overline{P} = \theta(P)$, $\overline{N} = \theta(N)$, $\overline{\mathfrak{p}_0} = \theta(\mathfrak{p}_0)$, $\overline{\mathfrak{n}_0} = \theta(\mathfrak{n}_0)$, $\overline{\mathfrak{p}} = \theta(\mathfrak{p})$ and $\overline{\mathfrak{n}} = \theta(\mathfrak{n})$. In general, we denote the dual space $\Hom_\mathbb{C}(V,\mathbb{C})$ of a $\mathbb{C}$-vector space $V$ by $V^*$. Let $\Sigma\subset\mathfrak{a}_0^*$ be the restricted root system for $(\mathfrak{g},\mathfrak{a}_0)$ and $\mathfrak{g}_\alpha$ the root space for $\alpha\in\Sigma$. Then $\sum_{\alpha\in\Sigma}\mathbb{R}\alpha$ is a real form of $\mathfrak{a}_0^*$. We denote the real part of $\lambda\in\mathfrak{a}_0^*$ with respect to this real form by $\re\lambda$ and the imaginary part by $\im\lambda$. Let $\Sigma^+$ be the positive system determined by $\mathfrak{n}_0$. Put $\rho_0 = \sum_{\alpha\in\Sigma^+}(\dim \mathfrak{g}_\alpha/2)\alpha$ and $\rho = \rho_0|_{\mathfrak{a}}$. The positive system $\Sigma^+$ determines the set of simple roots $\Pi$. Fix a totally order of $\sum_{\alpha\in\Sigma} \mathbb{R}\alpha$ such that the following conditions hold: (1) If $\alpha > \beta$ and $\gamma \in\sum_{\alpha\in\Sigma} \mathbb{R}\alpha$ then $\alpha + \gamma > \beta + \gamma$. (2) If $\alpha > 0$ and $c$ is a positive real number then $c\alpha > 0$. (3) For all $\alpha \in\Sigma^+$ we have $\alpha > 0$. Write $W$ for the little Weyl group for $(\mathfrak{g},\mathfrak{a}_0)$, $e$ for the unit element of $W$ and $w_0$ for the longest element of $W$. For $w\in W$, we fix a representative in $N_K(\mathfrak{a})$ and denote it by the same letter $w$. Let $\mathfrak{t}_0$ be a Cartan subalgebra of $\mathfrak{m}_0$ and $T_0$ the corresponding Cartan subgroup of $M_0$. Then $\mathfrak{h} = \mathfrak{t}_0\oplus\mathfrak{a}_0$ is a Cartan subalgebra of $\mathfrak{g}$. Let $\Delta$ be the root system for $(\mathfrak{g},\mathfrak{h})$ and take a positive system $\Delta^+$ compatible with $\Sigma^+$, i.e., if $\alpha\in\Delta^+$ satisfies that $\alpha|_{\mathfrak{a}_0} \ne 0$ then $\alpha|_{\mathfrak{a}_0}\in \Sigma^+$. Let $\mathfrak{g}^\mathfrak{h}_\alpha$ be the root space of $\alpha\in\Delta$ and $\widetilde{W}$ the Weyl group of $\Delta$. Put $\widetilde{\rho} = (1/2)\sum_{\alpha\in\Delta^+}\alpha$. By the decompositions $(\mathfrak{m}\cap \mathfrak{a}_0)^*\oplus\mathfrak{a}^* = \mathfrak{a}_0^*$ and $\mathfrak{t}_0^*\oplus\mathfrak{a}_0^* = \mathfrak{h}^*$, we always regard $\mathfrak{a}^*\subset\mathfrak{a}_0^*\subset\mathfrak{h}^*$. We use the same notation for $M$, i.e., $\Sigma_M$ be the restricted root system of $M$, $\Sigma_M^+ = \Sigma_M\cap \Sigma^+$, $W_M$ the little Weyl group of $M$, $\Delta_M$ the root system of $M$, $\Delta_M^+ = \Delta_M\cap \Delta^+$, $\widetilde{W_M}$ the Weyl group of $M$ and $w_{M,0}$ the longest element of $W_M$. We can define an anti-isomorphism of $U(\mathfrak{g})$ by $X\mapsto -X$ for $X\in \mathfrak{g}$. We write this anti-isomorphism by $u\mapsto \check{u}$. For a $\mathfrak{g}$-module $V$ and $g\in G$, we define a $\mathfrak{g}$-module $gV$ as follows: The representation space is $V$ and the action of $X\in\mathfrak{g}$ is $X\cdot v = (\Ad(g)^{-1}X)v$ for $v\in gV$. For $\xi = (\xi_1,\dots,\xi_l)\in\mathbb{Z}^l$, put $\lvert\xi\rvert = \xi_1 + \dots + \xi_l$. \section{Parabolic induction and the Bruhat filtration}\label{sec:Parabolic induction and Bruhat filtration} Fix a character $\eta$ of $\mathfrak{n}_0$ and put $\supp_G \eta = \supp\eta = \{\alpha\in\Pi\mid \eta|_{\mathfrak{g}_\alpha} \ne 0\}$\newsym{$\supp_G\eta = \supp\eta$}. The character $\eta$ is called \emph{non-degenerate} if $\supp\eta = \Pi$. We denote the character of $N_0$ whose differential is $\eta$ by the same letter $\eta$. \begin{defn}\label{defn:Jacquet modules} Let $V$ be a finite-length moderate growth Fr\'echet representation of $G$ (See Casselman~\cite[pp.~391]{MR1013462}). We define $\mathfrak{g}$-modules $J'_\eta(V)$ and $J^*_\eta(V)$ by \begin{align*} J'_\eta(V) & = \left\{v\in V'\Bigm| \begin{array}{l} \text{For some $k$ and for all $X\in\mathfrak{n}_0$,}\\ \text{$(X - \eta(X))^kv = 0$} \end{array} \right\},\\ J^*_\eta(V) & = \left\{v\in (V_{\textnormal{$K$-finite}})^*\Bigm| \begin{array}{l} \text{For some $k$ and for all $X\in\mathfrak{n}_0$,}\\ \text{$(X - \eta(X))^kv = 0$} \end{array} \right\}, \end{align*} where $V'$ is the continuous dual of $V$. \end{defn} Put $J'(V) = J'_0(V)$ and $J^*(V) = J^*_0(V)$ where $0$ is the trivial representation of $\mathfrak{n}_0$. The module $J^*(V)$ is the (dual of) \emph{Jacquet module} defined by Casselman~\cite{MR562655}. By the automatic continuation theorem~\cite[Theorem~4.8]{MR727854}, we have $J'(V) = J^*(V)$. The correspondence $V\mapsto J'_\eta(V)$ and $V\mapsto J^*_\eta(V)$ are functors from the category of $G$-modules to the category of $\mathfrak{g}$-modules. In this section, we study the module $J'_\eta(V)$ for a parabolic induction $V$. An element of $\mathfrak{a}^*$ is identified with a character of $A$. We denote the character of $A$ corresponding to $\lambda + \rho$ by $e^{\lambda + \rho}$ where $\lambda\in\mathfrak{a}^*$. For an irreducible moderate growth Fr\'echet representation $\sigma$ of $M$ and $\lambda\in\mathfrak{a}^*$, put \[ I(\sigma,\lambda) = C^\infty\mathchar`-\Ind_P^G(\sigma\otimes e^{\lambda + \rho}). \] (For a moderate growth Fr\'echet representation, see Casselman~\cite{MR1013462}.) The representation $I(\sigma,\lambda)$ has a natural structure of a moderate growth Fr\'echet representation. \newsym{$I(\sigma,\lambda)$} Denote its continuous dual by $I(\sigma,\lambda)'$. Let $\mathcal{L}$ be a vector bundle on $G/P$ attached to the representation $\sigma\otimes e^{\lambda+\rho}$ and $\mathcal{L}'$ be the continuous dual vector bundle of $\mathcal{L}$. \newsym{$\mathcal{L}$} \begin{rem}\label{rem:identify func on flag and G} A $C^\infty$-section of $\mathcal{L}$ corresponds to a $\sigma$-valued $C^\infty$-function $f$ on $G$ such that $f(gman) = \sigma(m)^{-1}e^{-(\lambda+\rho)(\log a)}f(g)$ for $g\in G$, $m\in M$, $a\in A$, $n \in N$. In particular a $C^\infty$-function on $G/P$ corresponds to a right $P$-invariant $C^\infty$-function on $G$. We use this identification throughout this paper. \end{rem} We use the notation in Appendix~\ref{sec:C^infty-function with values in Frechet space}. We can regard $J'_\eta(I(\sigma,\lambda))$ as a subspace of $\mathcal{D}'(G/P,\mathcal{L})$. Set $W(M) = \{w\in W\mid w(\Sigma_M^+)\subset \Sigma^+\}$.\newsym{$W(M)$}\newsym{$r$} Then it is known that the multiplication map $W(M)\times W_M\to W$ is bijective~\cite[Proposition~5.13]{MR0142696}. By the Bruhat decomposition, we have \[ G/P = \bigsqcup_{w\in W(M)}N_0wP/P. \] (Recall that we fix a representative of $w\in W$, see Notation.) Enumerate $W(M) = \{w_1,\dots,w_r\}$ so that $\bigcup_{j \le i}N_0w_jP/P$ is a closed subset of $G/P$ for each $i$. Then we can define a submodule $I_i$ of $J'_\eta(I(\sigma,\lambda))$ by \[ I_i = \left\{x\in J'_\eta(I(\sigma,\lambda))\Biggm| \supp x\subset \bigcup_{j \le i}N_0w_jP/P\right\}.\newsym{$I_i$} \] The filtration $\{I_i\}$ is called the Bruhat filtration~\cite{MR1767896}. In the rest of this section, we study the module $I_i/I_{i - 1}$. Put $U_i = w_i\overline{N}P/P$ and $O_i = N_0w_iP/P$. The subset $U_i$ is an open subset of $G/P$ containing $O_i$ and $U_i\cap O_j = \emptyset$ if $j < i$.\newsym{$U_i$}\newsym{$O_i$} Hence, the restriction map $\Res_i\colon I_i \to \mathcal{D}(U_i,\mathcal{L})$ induces an injective map $\Res_i\colon I_i/I_{i - 1}\to \mathcal{D}(U_i,\mathcal{L})$.\newsym{$\Res_i$} Moreover, $\im\Res_i \subset \mathcal{T}_{O_i}(U_i,\mathcal{L})$. We have $\mathcal{T}_{O_i}(U_i,\mathcal{L}) = U(\Ad(w_i)\overline{\mathfrak{n}}\cap\overline{\mathfrak{n}})\otimes_\mathbb{C}\mathcal{T}(O_i,\mathcal{L}|_{O_i})$ by Proposition~\ref{prop:structure theorem of tempered distributions whose support is contained in submanifold}. Notice that by a map $n\mapsto nw_iP/P$ we have isomorphisms $w_i\overline{N}w_i^{-1}\simeq U_i$ and $w_i\overline{N}w_i^{-1}\cap N_0\simeq O_i$. Fix a Haar measure on $w_i\overline{N}w_i^{-1}\cap N_0$. Then we can define $\delta_i\in \mathcal{D}'(O_i,\mathcal{L}|_{O_i})$ by \[ \langle\delta_i,\varphi\rangle = \int_{w_i\overline{N}w_i^{-1}\cap N_0}\varphi(nw_i)dn. \] for $\varphi\in C_c^\infty(O_i,\mathcal{L}|_{O_i})$.\newsym{$\delta_i$} By the exponential map $\Ad(w_i)\Lie(\overline{N})\to w_i\overline{N}w_i^{-1}$ and diffeomorphism $w_i\overline{N}w_i^{-1}\simeq U_i$, $U_i$ has a vector space structure and $O_i$ is a subspace of $U_i$. Let $\mathcal{P}(O_i)$ be the ring of polynomials on $O_i$ (cf.\ Appendix~\ref{subsec:Distributions on a nilpotent Lie group} or \cite{MR1070979}).\newsym{$\mathcal{P}(O_i)$} Define a $C^\infty$-function $\eta_i$ on $O_i$ by $\eta_i(nw_iP/P) = \eta(n)$ for $n\in w_i\overline{N}w_i^{-1}\cap N_0$.\newsym{$\eta_i$} For a $C^\infty$-function $f$ on $O_i$ and $u'\in \sigma'$, we define $f\otimes u'\in C^\infty(O_i,\sigma')$ by $(f\otimes u')(x) = f(x)u'$. Since $w_i\in W(M)$, $\Ad(w_i)(\mathfrak{m}\cap \mathfrak{n}_0) \subset \mathfrak{n}_0$. Hence we can define a character $w_i^{-1}\eta$ of $\mathfrak{m}\cap\mathfrak{n}_0$ by $(w_i^{-1}\eta)(X) = \eta(\Ad(w_i)X)$. Using this character, we can define the Jacquet module $J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda + \rho})$ of $MA$-representation $\sigma\otimes e^{\lambda + \rho}$. This is an $\mathfrak{m}\oplus\mathfrak{a}$-module. Put \[ I'_i = \left\{\sum_{k = 1}^l T_k(((f_k\eta_i^{-1})\otimes u_k')\delta_i)\biggm| \begin{array}{ll} T_k\in U(\Ad(w_i)\overline{\mathfrak{n}}\cap \overline{\mathfrak{n}}),& f_k\in \mathcal{P}(O_i),\\ u_k'\in J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho}) \end{array} \right\}. \] The space $I'_i$ is a $U(\mathfrak{g})$-submodule of $\mathcal{D}'(U_i,\mathcal{L})$. Our aim is to prove that if $i$ satisfies some conditions then $I_i/I_{i - 1} \simeq I'_i$. \begin{lem}\label{lem:bracket of n and bar_n} Let $E_1,\dots,E_n$ be a basis of $\Ad(w_i)\overline{\mathfrak{n}}\cap \overline{\mathfrak{n}_0}$ such that each $E_s$ is a restricted root vector for some root (say $\alpha_s$) and $F\in(\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0)\oplus\Ad(w_i)(\mathfrak{m}\cap\mathfrak{n}_0)$. (Notice that $\Ad(w_i)(\mathfrak{m}\cap \mathfrak{n}_0) = \Ad(w_i)\mathfrak{m}\cap \mathfrak{n}_0$ since $w_i\in W(M)$, so $(\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0)\oplus\Ad(w_i)(\mathfrak{m}\cap\mathfrak{n}_0) = \Ad(w_i)(\overline{\mathfrak{n}}\oplus\mathfrak{m})\cap \mathfrak{n}_0$ is a subalgebra of $\mathfrak{g}$.) For $\xi = (\xi_1,\xi_2,\dots,\xi_n)\in\mathbb{Z}_{\ge 0}^n$, set $E^\xi = E_1^{\xi_1}E_2^{\xi_2}\dotsm E_n^{\xi_n}$. Then for all $c\in\mathbb{C}$ we have \begin{multline*} [(F - c)^k,E^\xi] \in \left(\sum_{\xi'\in A(\xi)}\mathbb{C} E^{\xi'}\right) U((\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0)\oplus\Ad(w_i)(\mathfrak{m}\cap\mathfrak{n}_0)) \\\subset U(\Ad(w_i)(\overline{\mathfrak{n}}\oplus(\mathfrak{m}\cap \mathfrak{n}_0))) \end{multline*} where $A(\xi) = \{\xi'\in\mathbb{Z}_{\ge 0}^n\mid \text{$\lvert\xi'\rvert < \lvert\xi\rvert$, or $\lvert\xi'\rvert = \lvert\xi\rvert$ and $\sum \xi'_i\alpha_i < \sum\xi_i\alpha_i$}\}$.\end{lem} \begin{proof} We may assume $k = 1$. We prove the lemma by induction on $\lvert\xi\rvert$. We have \[ [F - c,E^\xi] = [F,E^\xi] = \sum_{s = 1}^n\sum_{l = 0}^{\xi_s - 1}E_1^{\xi_1}\dotsm E_{s - 1}^{\xi_{s - 1}} E_s^l [F,E_s] E_s^{\xi_s - l - 1} E_{s + 1}^{\xi_{s + 1}}\dotsm E_n^{\xi_n}. \] Hence, it is sufficient to prove \begin{multline*} E_1^{\xi_1}\dotsm E_{s - 1}^{\xi_{s - 1}} E_s^l [F,E_s] E_s^{\xi_s - l - 1} E_{s + 1}^{\xi_{s + 1}}\dotsm E_n^{\xi_n}\\ \in \left(\sum_{\xi'\in A(\xi)}\mathbb{C} E^{\xi'}\right) U((\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0)\oplus\Ad(w_i)(\mathfrak{m}\cap\mathfrak{n}_0)). \end{multline*} We may assume that $F$ is a restricted root vector. If $[F,E_s]\in \Ad(w_i)\overline{\mathfrak{n}}\cap \overline{\mathfrak{n}_0}$ then the claim hold. Assume that $[F,E_s]\in (\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0)\oplus\Ad(w_i)(\mathfrak{m}\cap\mathfrak{n}_0)$. Put $\xi^{(1)} = (\xi_1,\dots,\xi_{s - 1},l,0,\dots,0)\in\mathbb{Z}^n$ and $\xi^{(2)} = (0,\dots,0,\xi_s - l - 1,\xi_{s + 1},\dots,\xi_n)\in\mathbb{Z}^n$. Using inductive hypothesis, we have \begin{align*} & E^{\xi^{(1)}}\bigl[ [F,E_s],E^{\xi^{(2)}}\bigr] \\ & \in E^{\xi^{(1)}}\left(\sum_{\xi'\in A(\xi^{(2)})}\mathbb{C} E^{\xi'}\right)U((\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0)\oplus\Ad(w_i)(\mathfrak{m}\cap\mathfrak{n}_0))\\ & \subset \left(\sum_{\xi'\in A(\xi^{(1)} + \xi^{(2)})}\mathbb{C} E^{\xi'}\right)U((\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0)\oplus\Ad(w_i)(\mathfrak{m}\cap\mathfrak{n}_0))\\ & \subset \left(\sum_{\xi'\in A(\xi)}\mathbb{C} E^{\xi'}\right)U((\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0)\oplus\Ad(w_i)(\mathfrak{m}\cap\mathfrak{n}_0)) \end{align*} On the other hand, we have \begin{multline*} E^{\xi^{(1)}}E^{\xi^{(2)}}[F,E_s]\in \left(\sum_{\lvert\xi'\rvert \le \lvert\xi^{(1)} +\xi^{(2)}\rvert}\mathbb{C} E^{\xi'}\right)[F,E_s]\\ \subset \left(\sum_{\lvert\xi'\rvert \le \lvert\xi^{(1)} +\xi^{(2)}\rvert}\mathbb{C} E^{\xi'}\right)(\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0)\oplus\Ad(w_i)(\mathfrak{m}\cap\mathfrak{n}_0). \end{multline*} Since $\lvert\xi^{(1)} + \xi^{(2)}\rvert = \lvert\xi\rvert - 1 < \lvert\xi\rvert$, we get the lemma. \end{proof} Let $X$ be an element of the normalizer of $\Ad(w_i)\overline{\mathfrak{n}}\cap\mathfrak{n}_0$ in $\mathfrak{g}$. For $f\in C^\infty(O_i)$ we define $D_i(X)f\in C^\infty(O_i)$ by \[ (D_i(X)f)(nw_i) = \left.\frac{d}{dt}f(\exp(-tX)n\exp(tX)w_i)\right|_{t = 0} \] where $n \in w_i\overline{N}w_i^{-1}\cap N_0$.\newsym{$D_i(X)$}\label{symbol:D_i} \begin{lem}\label{lem:no delta part} Fix $f\in C^\infty(O_i)$, $u'\in (\sigma\otimes e^{\lambda+\rho})'$ and $X\in\mathfrak{g}$. \begin{enumerate} \item If $X\in \mathfrak{a}_0$, then $X$ normalizes $\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0$ and we have \begin{multline*} X((f\otimes u')\delta_i) = ((D_i(X)f)\otimes u')\delta_i + (f\otimes((\Ad(w_i)^{-1}X)u'))\delta_i\\ + (w_i\rho_0 - \rho_0)(X)(f\otimes u')\delta_i. \end{multline*} \item If $X\in \Ad(w_i)(\mathfrak{m}\cap \mathfrak{n}_0)$ or $X\in\mathfrak{m}_0$, then $X$ normalizes $\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0$ and we have \[ X((f\otimes u')\delta_i) = ((D_i(X)f)\otimes u')\delta_i + (((\Ad(w_i)^{-1}X)u')\otimes f) \delta_i. \] \end{enumerate} \end{lem} \begin{proof} Let $X$ be as in the lemma. First we prove that $X$ normalizes $\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0$. If $X\in \mathfrak{m}_0 + \mathfrak{a}_0$, then $X$ normalizes each restricted root space. Hence, $X$ normalizes $\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0$. If $X\in \Ad(w_i)(\mathfrak{m}\cap \mathfrak{n}_0)$, then $X\in \mathfrak{n}_0$ since $w_i\in W(M)$. Hence, $X$ normalizes $\mathfrak{n}_0$. Since $\mathfrak{m}$ normalizes $\overline{\mathfrak{n}}$, $X$ normalizes $\Ad(w_i)\overline{\mathfrak{n}}$. Put $g_t = \exp(tX)$. Take $\varphi\in C_c^\infty(U_i,\mathcal{L})$ and we regard $\varphi$ as a $\sigma$-valued $C^\infty$-function on $w_i\overline{N}P$ (Remark~\ref{rem:identify func on flag and G}). Since $w_ig_tw_i^{-1}\in P$, we have $\varphi(xw_ig_tw_i^{-1}) = \sigma(w_ig_tw_i^{-1})^{-1}\varphi(x)$. Put $D(t) = \lvert\det(\Ad(g_t)^{-1}|_{\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0})\rvert$. Then \begin{align*} & \langle X((f\otimes u')\delta_i),\varphi\rangle = \langle (f\otimes u')\delta_i ,-X\varphi\rangle\\ & = \frac{d}{dt}\left.\int_{w_i\overline{N}w_i^{-1}\cap N_0}u'(\varphi(g_tnw_i))f(nw_i)dn\right|_{t = 0}\\ & = \frac{d}{dt}\left.\int_{w_i\overline{N}w_i^{-1}\cap N_0}u'(\varphi((g_tng_t^{-1})w_i(w_i^{-1}g_tw_i)))f(nw_i)dn\right|_{t = 0}\\ & = \frac{d}{dt}\left.\int_{w_i\overline{N}w_i^{-1}\cap N_0}u'(\sigma(w_i^{-1}g_tw_i)^{-1}\varphi((g_tng_t^{-1})w_i))f(nw_i)dn\right|_{t = 0}\\ & = \frac{d}{dt}\left.\int_{w_i\overline{N}w_i^{-1}\cap N_0}u'(\sigma(w_i^{-1}g_tw_i)^{-1}\varphi(nw_i))f(g_t^{-1}ng_tw_i)D(t) dn\right|_{t = 0}\\ & = \frac{d}{dt}\left.\int_{w_i\overline{N}w_i^{-1}\cap N_0}((w_i^{-1}g_tw_i)u')(\varphi(nw_i))f(g_t^{-1}ng_tw_i)D(t) dn\right|_{t = 0} \end{align*} This implies \begin{multline*} X((f\otimes u')\delta_i) = ((D_i(X)f)\otimes u')\delta_i + (f\otimes((\Ad(w_i)^{-1}X)u'))\delta_i\\ + \left.\frac{d}{dt}\lvert\det(\Ad(g_t)^{-1}|_{\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}})\rvert\right|_{t = 0}((f\otimes u')\delta_i) \end{multline*} (1) Assume that $X\in \mathfrak{a}_0$. Since $w_i\in W(M)$, we have $w_i\overline{N}w_i^{-1}\cap N_0 = w_i\overline{N}_0w_i^{-1}\cap N_0$. This implies that $\det(\Ad(g_t)^{-1}|_{\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0}) = e^{t(w_i\rho_0 - \rho_0)(X)}$. (2) First assume that $X\in\mathfrak{m}_0$. Since $g\mapsto \det(\Ad(g)^{-1}|_{\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0})$ is $1$-dimensional representation, it is unitary since $M_0$ is compact. Hence we have $\lvert\det(\Ad(g_t)^{-1}|_{\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0})\rvert = 1$. Next assume that $X\in (\Ad(w_i)\mathfrak{m}\cap \mathfrak{n}_0)$. Then $\ad(X)$ is nilpotent. Hence, $\Ad(g_t) - 1$ is nilpotent. This implies $\det(\Ad(g_t)^{-1}|_{\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0}) = 1$. \end{proof} \begin{lem}\label{lem:succ quot is sub of I'} Let $x\in\mathcal{T}_{O_i}(U_i,\mathcal{L})$. Assume that for all $X\in \Ad(w_i)\overline{\mathfrak{p}}\cap \mathfrak{n}_0$ there exists a positive integer $k$ such that $(X - \eta(X))^kx = 0$. Then $x\in I_i'$. In particular we have $\im\Res_i\subset I_i'$. \end{lem} \begin{proof} Let $E_s$ and $\alpha_s$ be as in Lemma~\ref{lem:bracket of n and bar_n}. For $\xi = (\xi_1,\xi_2,\dots,\xi_n)\in\mathbb{Z}_{\ge 0}^n$, set $E^\xi = E_1^{\xi_1}E_2^{\xi_2}\dots E_n^{\xi_n}$. Since $x\in\mathcal{T}_{O_i}(U_i,\mathcal{L}) = U(\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0)\otimes \mathcal{T}(O_i,\mathcal{L})$, there exist $x_\xi\in \mathcal{T}(O_i,\mathcal{L})$ such that $x = \sum_\xi E^\xi x_\xi$ (finite sum). First we prove $x_\xi\in (\mathcal{P}(O_i)\eta_i^{-1}\otimes(\sigma\otimes e^{\lambda+\rho})')\delta_i$ by backward induction on the lexicological order of $(\lvert\xi\rvert,\sum_s \xi_s\alpha_s)$. Fix a nonzero element $F\in\Ad(w_i)\overline{\mathfrak{n}}\cap\mathfrak{n}_0$. Then $(F - \eta(F))^kx = \sum_\xi [(F - \eta(F))^k,E^\xi] (x_\xi) + \sum_\xi E^\xi((F - \eta(F))^k x_\xi)$. Assume that $(F - \eta(F))^kx = 0$. Define the set $A(\xi)$ as in Lemma~\ref{lem:bracket of n and bar_n}. By Lemma~\ref{lem:bracket of n and bar_n}, we have \begin{multline*} \sum_\xi E^\xi ((F - \eta(F))^kx_\xi) = -\sum_\xi [(F - \eta(F))^k,E^\xi](x_\xi)\\ \in \sum_\xi \left(\sum_{\xi'\in A(\xi)}\mathbb{C} E^{\xi'}\right) U((\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n})\oplus\Ad(w_i)(\mathfrak{m}\cap\mathfrak{n}_0))(x_{\xi}). \end{multline*} Put $B(\xi) = \{\xi'\mid \text{$\lvert\xi'\rvert > \lvert\xi\rvert$ or $\lvert\xi'\rvert = \lvert\xi\rvert$ and $\sum \xi'_s\alpha_s > \sum \xi_s\alpha_s$} \}$. Notice that $U((\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n})\oplus\Ad(w_i)(\mathfrak{m}\cap\mathfrak{n}_0))(x_{\xi})\in \mathcal{T}(O_i,\mathcal{L})$. Since $\mathcal{T}_{O_i}(U_i,\mathcal{L}) = U(\Ad(w_i)\overline{\mathfrak{n}}\cap\overline{\mathfrak{n}})\otimes_\mathbb{C}\mathcal{T}(O_i,\mathcal{L}|_{O_i})$, we have \[ (F - \eta(F))^kx_\xi \in \sum_{\xi'\in B(\xi)}U((\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n})\oplus\Ad(w_i)(\mathfrak{m}\cap\mathfrak{n}_0))(x_{\xi'}). \] By inductive hypothesis, $x_{\xi'}\in (\mathcal{P}(O_i)\eta_i^{-1}\otimes (\sigma\otimes e^{\lambda+\rho})')\delta_i$ for all $\xi'\in B(\xi)$. Hence we have $(F - \eta(F))^kx_{\xi}\in (\mathcal{P}(O_i)\eta_i^{-1}\otimes (\sigma\otimes e^{\lambda+\rho})')\delta_i$. Therefore $x_{\xi}\in(\mathcal{P}(O_i)\eta_i^{-1}\otimes(\sigma\otimes e^{\lambda+\rho})')\delta_i$ by Corollary~\ref{cor:polynomial by some power of n}. Hence, we can write $x = \sum_{\xi}E^\xi \sum_l (f_{\xi,l}\eta_i^{-1}\otimes u_{\xi,l}')\delta_i$ (finite sum), where $f_{\xi,l}\in \mathcal{P}(O_i)$ and $u_{\xi,l}'\in (\sigma\otimes e^{\lambda+\rho})'$. Moreover, we can assume that $f_{\xi,l}$ is an $\mathfrak{a}_0$-weight vector with respect to $D_i$ and $\{f_{\xi,l}\}_l$ is lineally independent for each $\xi$. We prove $u_{\xi,l}'\in J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$. Take $F\in \mathfrak{n}_0\cap \mathfrak{m}$. By Lemma~\ref{lem:no delta part}, we have \begin{multline*} (\Ad(w_i)F - \eta(\Ad(w_i)F))^kx \\ = \sum_{\xi,l}[(\Ad(w_i)F - \eta(\Ad(w_i)F))^k,E^\xi]((f_{\xi,l}\eta_i^{-1}\otimes u_{xi,l}')\delta_i)\\ + \sum_{\xi}E^\xi\sum_{p = 0}^k\binom{k}{p}(((D_i(\Ad(w_i)F))^{k - p}(f_{\xi,l})\eta_i^{-1})\otimes\\(F - \eta(\Ad(w_i)F))^p(u_{\xi,l}'))\delta_i. \end{multline*} Now we prove $u_{\xi,l}'\in J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$ by backward induction on the lexicological order of $(\lvert\xi\rvert,\sum \xi_s\alpha_s,-\wt f_{\xi,l})$ where $\wt f_{\xi,l}$ is an $\mathfrak{a}_0$-weight of $f_{\xi,l}$ with respect to $D_i$. Take $k$ such that $(\Ad(w_i)F - \eta(\Ad(w_i)F))^k x = 0$. Then we have \begin{multline*} f_{\xi,l}\otimes (F - \eta(\Ad(w_i)F))^k(u'_{\xi,l})\delta_i\\ \in \sum_{\eta\in B(\xi),l}U((\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0)\oplus\Ad(w_i)(\mathfrak{m}\cap\mathfrak{n}_0))((f_{\eta,l}\eta_i^{-1}\otimes u_{\eta,l}')\delta_i)\\ + \sum_{\wt f_{\eta,l'} < \wt f_{\xi,l}}\sum_p(((D_i(\Ad(w_i)F))^pf_{\eta,l'}\eta_i^{-1})\otimes(U(\mathbb{C} F)u_{\eta,l'}'))\delta_i. \end{multline*} By inductive hypothesis, we have $(F - \eta(F))^ku_{\xi,l}'\in J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$. This implies that $u_{\xi,l}'\in J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$. \end{proof} In fact, we have $\im\Res_i = I'_i$ under some conditions. This is proved in Section~\ref{sec:Analytic continuation}. \section{Vanishing theorem}\label{sec:vanishing theorem} In this section, we fix $i \in \{1,2,\dots,r\}$ and a basis $\{e_1,e_2,\dots,e_l\}$ of $\Ad(w_i)\overline{\mathfrak{n}}\cap\mathfrak{n}_0$. Here we assume that each $e_i$ is a restricted root vector and denote its root by $\alpha_i$. By the decomposition \begin{multline*} N_0/[N_0,N_0] \simeq ((w_i\overline{P}w_i^{-1}\cap N_0)/(w_i\overline{P}w_i^{-1}\cap [N_0,N_0])) \\\times ((w_iNw_i^{-1}\cap N_0)/(w_iNw_i^{-1}\cap [N_0,N_0])) \end{multline*} where $[\cdot,\cdot]$ is the commutator group, we can define a character $\eta'$ of $N_0$ by $\eta'(n) = \eta(n)$ for $n\in w_i\overline{P}w_i^{-1}\cap N_0$ and $\eta'(n) = 1$ for $n\in w_iNw_i^{-1}\cap N_0$. \begin{lem}\label{lem:acts nilp} Let $X\in \mathfrak{n}_0$. Then for all $x\in I_i'$ there exists a positive integer $k$ such that $(X - \eta'(X))^kx = 0$. \end{lem} To prove this lemma, we prepare some notation. For $X\in \Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0$, we define a differential operator $R_i'(X)$ on $O_i$ by \[ (R'_i(X)\varphi)(nw_iP/P) = \left.\frac{d}{dt}\varphi(n\exp(tX)w_iP/P)\right|_{t = 0} \] where $n\in w_i\overline{N}w_i^{-1}\cap N_0$.\newsym{$R'_i(X)$} (Recall that $w_i\overline{N}w_i^{-1}\cap N_0\simeq O_i$ by the map $n\mapsto nw_iP/P$.) For $X\in \mathfrak{g}$, we define a differential operator $\widetilde{R}_i(X)$ on $w_i\overline{N}P$ by the same way, i.e., for a $C^\infty$-function $\varphi$ on $w_i\overline{N}P$, put \[ (\widetilde{R}_i(X)\varphi)(pw_i) = \left.\frac{d}{dt}\varphi(p\exp(tX)w_i)\right|_{t = 0} \] for $p \in w_i\overline{N}Pw_i^{-1}$.\newsym{$\widetilde{R}_i(X)$} Notice that even if $\varphi$ is right $P$-invariant, $\widetilde{R}_i(X)\varphi$ is not right $P$-invariant in general. Since $R'_i$ (resp.\ $\widetilde{R}_i$) is a Lie algebra homomorphism, we can define a differential operator $R'_i(T)$ (resp.\ $\widetilde{R}_i(T)$) for $T\in U(\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0)$ (resp.\ $T\in U(\mathfrak{g})$) as usual. For $T\in U(\mathfrak{g})$, $f\in C^\infty(O_i)$ and $u'\in (\sigma\otimes(\lambda + \rho))'$, we define $\delta_i(T,f,u')\in \mathcal{D}'_{O_i}(U_i,\mathcal{L})$\newsym{$\delta_i(T,f,u')$} by \[ \langle \delta_i(T,f,u'),\varphi\rangle = \int_{w_i\overline{N}w_i^{-1}\cap N_0} f(nw_i)u'((\widetilde{R}_i(T)\varphi)(nw_i))dn \] where $\varphi\in C^\infty_c(U_i,\mathcal{L})$ and we regard $\varphi$ as a function on $w_i\overline{N}P$ (Remark~\ref{rem:identify func on flag and G}). The following lemma is easy to prove. \begin{lem}\label{lem:fundamental properties of delta_i} We have the following properties. \begin{enumerate} \item For $X\in \Ad(w_i)\overline{\mathfrak{n}}\cap\mathfrak{n}_0$, $\delta_i(XT,f,u') = \delta_i(T,R'_i(-X)(f),u')$. \item For $X\in \Ad(w_i)\mathfrak{p}$, $\delta_i(TX,f,u') = \delta_i(T,f,\Ad(w_i)^{-1}Xu')$. \item The map $C^\infty(O_i)\otimes_{U(\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n})} U(\mathfrak{g})\otimes_{U(\Ad(w_i)\mathfrak{p})} w_i(\sigma\otimes e^{\lambda+\rho})'\to \mathcal{D}'(U_i,O_i,\mathcal{L})$ defined by $f\otimes T\otimes u'\mapsto \delta_i(T,f,u')$ is injective. \end{enumerate} \end{lem} \begin{lem}\label{lem:left2right} Let $\{e_i\}$ be a basis of $\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0$ such that $e_i$ is a restricted root vector, $\alpha_i$ the restricted root for $e_i$, $T,T'\in U(\mathfrak{g})$, $f\in C^\infty(O_i)$ and $u'\in (\sigma\otimes e^{\lambda + \rho})'$. Then we have \begin{multline*} T\delta_i(T',f,u') \\= \sum_{(k_1,\dots,k_l)\in \mathbb{Z}_{\ge 0}^l}\delta_i\left((\ad(e_l)^{k_l}\dotsm \ad(e_1)^{k_1}T)T',f\prod_{s = 1}^l\frac{(-x_s)^{k_s}}{k_s!},u'\right), \end{multline*} where $x_i$ is a polynomial on $O_i$ given by $\exp(a_1e_1)\dotsm \exp(a_le_l)w_iP/P\mapsto a_i$. (Notice that the right hand side is a finite sum since $\ad(e_i)$ is nilpotent.) \end{lem} \begin{proof} We remark that by a map $(a_1,\dots,a_l)\mapsto \exp(a_1e_1)\dotsm \exp(a_le_l)$, we have a diffeomorphism $\mathbb{R}^l\simeq w_i\overline{N}w_i^{-1}\cap N_0$ and a Haar measure of $w_i\overline{N}w_i^{-1}\cap N_0$ corresponds to the Euclidean measure of $\mathbb{R}^l$. Take $\varphi\in C_c^\infty(w_i\overline{N}P,\sigma\otimes e^{\lambda+\rho})$. Put $n(a) = \exp(a_1e_1)\dotsm \exp(a_le_l)$ for $a = (a_1,\dots,a_l)$. Recall the definition of $\check{T}$ from Notation. For $T\in \mathfrak{g}$, we have \begin{align*} &\langle T\delta_i(T',f,u'),\varphi\rangle\\ & = \int_{\mathbb{R}^l}u'((\check{T}\widetilde{R}_i(T')\varphi)(n(a)w_i))f(n(a)w_i)da\\ & = \left.\frac{d}{dt}\int_{\mathbb{R}^l}u'(\widetilde{R}_i(T')\varphi)(\exp(tT)n(a)w_i))f(n(a)w_i)da\right|_{t = 0}\\ & = \left.\frac{d}{dt}\int_{\mathbb{R}^l}u'((\widetilde{R}_i(T')\varphi)(n(a)\exp(t\Ad(n(a))^{-1}T)w_i))f(n(a)w_i)da\right|_{t = 0}. \end{align*} The formula \begin{align*} \Ad(n(a))^{-1}T & = e^{-\ad(a_le_l)}\dotsm e^{-\ad(a_1e_1)}T\\ & = \sum_{(k_1,\dots,k_l)\in \mathbb{Z}_{\ge 0}^l}\frac{(-a_1)^{k_1}}{k_1!}\dotsm \frac{(-a_l)^{k_l}}{k_l!}\ad(e_l)^{k_l}\dotsm \ad(e_1)^{k_1}T \end{align*} gives the lemma. \end{proof} For $\mathbf{k} = (k_1,\dots,k_l)$, we denote an operator $\ad(e_l)^{k_l}\dotsm \ad(e_1)^{k_1}$ on $\mathfrak{g}$ by $\ad(e)^{\mathbf{k}}$ and a function $((-x_1)^{k_1}/k_1!)\dotsm ((-x_l)^{k_l}/k_l!)\in\mathcal{P}(O_i)$ by $f_\mathbf{k}$. \begin{lem}\label{lem:diff vanish} Let $\mathbf{k} = (k_1,\dots,k_l)\in\mathbb{Z}_{\ge 0}^l$ and $X\in \mathfrak{n}_0$. Assume that $\ad(e)^\mathbf{k}X \in \Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0$. Then we have $R'_i(\ad(e)^\mathbf{k}X)f_\mathbf{k} = 0$. \end{lem} \begin{proof} We may assume that $X$ is a restricted root vector and denote its restricted root by $\alpha$. We consider an $\mathfrak{a}_0$-weight with respect to $D_i$. An $\mathfrak{a}_0$-weight of $f_\mathbf{k}$ is $-\sum_s k_s\alpha_s$. This implies that $R'_i(\ad(e)^\mathbf{k}X)f_\mathbf{k}$ has an $\mathfrak{a}_0$-weight $\alpha$. However, $\mathcal{P}(O_i)$ has a decomposition into the direct sum of $\mathfrak{a}_0$-weight spaces and its weight belongs to $\{\sum_{\beta\in\Sigma^+}b_\beta\beta\mid b_\beta\in\mathbb{Z}_{\le 0}\}$. Hence, we have $R'_i(\ad(e)^\mathbf{k}X)f_\mathbf{k} = 0$. \end{proof} For $f\in\mathcal{P}(O_i)$ and $X\in\mathfrak{n}_0$ we define $L_X(f)$\newsym{$L_X$} by \[ L_X(f)(nw_i) = \left.\frac{d}{dt}f(\exp(-tX)nw_i)\right|_{t = 0}. \] \begin{lem}\label{lem:caluculation of Xdelta(1,f,u)} Let $X\in \mathfrak{n}_0$ be a restricted root vector. For $f\in\mathcal{P}(O_i)$ and $u'\in J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$, we have \begin{multline*} (X - \eta'(X))\delta_i(1,f\eta_i^{-1},u') = \delta_i(1,L_X(f)\eta_i^{-1},u') \\ + \sum_{\ad(e)^\mathbf{k}X\in \Ad(w_i)\mathfrak{n}_0\cap \mathfrak{n}_0}\delta_i(1,ff_{\mathbf{k}}\eta_i^{-1},(\Ad(w_i)^{-1}(\ad(e)^{\mathbf{k}}X) - \eta'(\ad(e)^{\mathbf{k}}X))u'). \end{multline*} (Again the sum of the right hand side is finite.) \end{lem} \begin{proof} We have \[ X\delta_i(1,f\eta_i^{-1},u') = \sum_{\mathbf{k}\in\mathbb{Z}_{\ge 0}^l}\delta_i(\ad(e)^\mathbf{k}X,ff_\mathbf{k}\eta_i^{-1},u'). \] by Lemma~\ref{lem:left2right}. Since $\ad(e)^\mathbf{k}X$ belongs to $\mathfrak{n}_0$ and is a restricted root vector, we have either $\ad(e)^\mathbf{k}X\in \Ad(w_i)\overline{\mathfrak{n}_0}\cap \mathfrak{n}_0$ or $\ad(e)^\mathbf{k}X\in \Ad(w_i)\mathfrak{n}_0\cap \mathfrak{n}_0$. Recall that $\Ad(w_i)\overline{\mathfrak{n}_0}\cap \mathfrak{n}_0 = \Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0$ since $w_i\in W(M)$. Assume that $\ad(e)^\mathbf{k}X\in \Ad(w_i)\overline{\mathfrak{n}}\cap\mathfrak{n}_0$. By the definition of $\eta_i$ and $\eta'$, we have $R'_i(-\ad(e)^\mathbf{k}X)(\eta_i^{-1}) = \eta(\ad(e)^\mathbf{k}X)\eta_i^{-1} = \eta'(\ad(e)^\mathbf{k}X)\eta_i^{-1}$. Hence, using Lemma~\ref{lem:diff vanish}, \begin{align*} &\delta_i(\ad(e)^\mathbf{k}X,ff_\mathbf{k}\eta_i^{-1},u') \\ &= \delta_i(1,R'_i(-\ad(e)^\mathbf{k}X)(ff_\mathbf{k}\eta_i^{-1}),u')\\ &= \delta_i(1,R'_i(-\ad(e)^\mathbf{k}X)(f)f_\mathbf{k}\eta_i^{-1},u') + \eta'(\ad(e)^\mathbf{k}X)\delta_i(1,ff_\mathbf{k}\eta_i^{-1},u'). \end{align*} Next assume that $\ad(e)^{\mathbf{k}}X\in \Ad(w_i)\mathfrak{n}_0\cap \mathfrak{n}_0$. For $h\in\mathcal{P}(O_i)$, define $\widetilde{h}\in\mathcal{P}(U_i)$ by $\widetilde{h}(nn_0w_iP) = h(nw_iP)$ for $n\in w_i\overline{N}w_i^{-1}\cap N_0$ and $n_0\in w_i\overline{N}w_i^{-1}\cap \overline{N_0}$. Then we have $\widetilde{R'_i(Y)h} = \widetilde{R}_i(Y)\widetilde{h}$ for all $Y\in \Ad(w_i)\overline{\mathfrak{n}_0}\cap \mathfrak{n}_0$. Since $\widetilde{f}(pnw_i) = \widetilde{f}(pw_i)$ for $p\in w_i\overline{N}Pw_i^{-1}$ and $n\in w_iN_0w_i^{-1}\cap N_0$, we have $\widetilde{R}_i(-\ad(e)^{\mathbf{k}}X)(\widetilde{f}) = 0$. Hence we have \begin{multline*} \delta_i(\ad(e)^\mathbf{k}X,ff_\mathbf{k}\eta_i^{-1},u') = \delta_i(1,ff_\mathbf{k}\eta_i^{-1},\Ad(w_i)^{-1}(\ad(e)^{\mathbf{k}}X)u')\\ = \delta_i(1,R'_i(-\ad(e)^\mathbf{k}X)(\widetilde{f})|_{O_i}f_\mathbf{k}\eta_i^{-1},u') + \\ \delta_i(1,ff_\mathbf{k}\eta_i^{-1},\Ad(w_i)^{-1}(\ad(e)^{\mathbf{k}}X)u'). \end{multline*} By the same calculation as the proof of Lemma~\ref{lem:left2right}, we have \[ \widetilde{L_X(f)} = L_X(\widetilde{f}) = \sum_{\mathbf{k}\in\mathbb{Z}_{\ge 0}^l}\widetilde{R}_i(-\ad(e)^\mathbf{k}X)(\widetilde{f})\widetilde{f_\mathbf{k}}. \] Hence \begin{align*} \sum_{\mathbf{k}\in\mathbb{Z}_{\ge 0}^l}\delta_i(1,\widetilde{R}_i(-\ad(e)^\mathbf{k}X)(\widetilde{f})|_{O_i}f_\mathbf{k}\eta_i^{-1},u') &= \delta_i(1,\widetilde{L_X(f)}|_{O_i}\eta_i^{-1},u')\\ & = \delta_i(1,L_X(f)\eta_i^{-1},u'). \end{align*} These imply that \begin{multline*} (X - \eta'(X))\delta_i(1,f\eta_i^{-1},u') = \delta_i(1,L_X(f)\eta_i^{-1},u')\\ + \sum_{\ad(e)^\mathbf{k}X\in \Ad(w_i)\mathfrak{n}_0\cap \mathfrak{n}_0}\delta_i(1,ff_\mathbf{k}\eta_i^{-1},\Ad(w_i)^{-1}(\ad(e)^{\mathbf{k}}X)u')\\ + \sum_{\ad(e)^\mathbf{k}X\in \Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0}\eta'(\ad(e)^\mathbf{k}X)\delta_i(1,ff_\mathbf{k}\eta_i^{-1},u')\\ - \eta'(X)\delta_i(1,f\eta_i^{-1},u'). \end{multline*} Since $\eta'$ is a character, if $\mathbf{k}\ne (0,\dots,0)$ then $\eta'(\ad(e)^\mathbf{k}X) = 0$. Hence we have \[ \sum_{\mathbf{k}\in\mathbb{Z}_{\ge 0}^n}\eta'(\ad(e)^\mathbf{k}X)\delta_i(1,ff_\mathbf{k}\eta_i^{-1},u') = \eta'(X)\delta_i(1,f\eta_i^{-1},u'). \] This implies \begin{multline*} \left(\sum_{\ad(e)^\mathbf{k}X\in \Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0}\eta'(\ad(e)^\mathbf{k}X)\delta_i(1,ff_\mathbf{k}\eta_i^{-1},u')\right) - \eta'(X)\delta_i(1,f\eta_i^{-1},u')\\ = \sum_{\ad(e)^\mathbf{k}X\in\Ad(w_i)\mathfrak{n}_0\cap \mathfrak{n}_0}\eta'(\ad(e)^\mathbf{k})\delta_i(1,ff_\mathbf{k}\eta_i^{-1},u'). \end{multline*} We get the lemma. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:acts nilp}] Since $\ad(\mathfrak{n}_0)$ acts $\mathfrak{g}$ nilpotently, the subspace \[ \{x\in I'_i\mid \text{for some $k$ and for all $X\in \mathfrak{n}_0$, $(X - \eta(X))^kx = 0$}\} \] is $\mathfrak{g}$-stable. Hence we may assume that $x = ((f\eta_i^{-1})\otimes u')\delta_i = \delta_i(1,f\eta_i^{-1},u')$ for some $f\in\mathcal{P}(O_i)$ and $u'\in J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$. Set $V = U(\Ad(w_i)^{-1}\mathfrak{n}_0\cap \mathfrak{n}_0)u'$ where $\mathfrak{n}$ acts $J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$ trivially. Then $V$ is finite-dimensional. By applying Engel's theorem for $V\otimes (-w_i^{-1}\eta')$, there exists a filtration $0 = V_0 \subset V_1\subset\cdots\subset V_p = V$ such that $(V_s/V_{s - 1})\otimes (-w_i^{-1}\eta'|_{\Ad(w_i)^{-1}\mathfrak{n}_0\cap\mathfrak{n}_0})$ is the trivial representation of $\Ad(w_i)^{-1}\mathfrak{n}_0\cap\mathfrak{n}_0$. Then we have $V_s/V_{s - 1}\simeq w_i^{-1}\eta'|_{\Ad(w_i)^{-1}\mathfrak{n}_0\cap\mathfrak{n}_0}$ for all $s = 1,2,\dots,p$. We prove the lemma by induction on $p = \dim V$. We may assume that $X$ is a restricted root vector. By Lemma~\ref{lem:caluculation of Xdelta(1,f,u)}, we have \begin{multline*} (X - \eta'(X))\delta_i(1,f\eta_i^{-1},u')\in \delta_i(1,L_X(f)\eta_i^{-1},u')\\ + \sum_{h\in\mathcal{P}(O_i),\ v'\in V_{p - 1}}\delta_i(1,h\eta_i^{-1},v'). \end{multline*} Since $f$ is a polynomial, there exists a positive integer $c$ such that $(L_X)^c(f) = 0$. Then $(X - \eta'(X))^c\delta_i(1,f\eta_i^{-1},u') \in \sum_{h\in\mathcal{P}(O_i), v'\in V_{p - 1}}\delta_i(1,h\eta_i^{-1},v')$. By inductive hypothesis the lemma is proved. \end{proof} From the lemma, we get the following vanishing theorem. Recall that we define the character $w_i^{-1}\eta$ of $\mathfrak{m}\cap \mathfrak{n}_0$ by $(w_i^{-1}\eta)(X) = \eta(\Ad(w_i)X)$. \begin{lem}\label{lem:vanishing lemma} Assume that $I_i/I_{i - 1} \ne 0$. Then the following conditions hold. \begin{enumerate} \item The character $\eta$ is unitary. \item The character $\eta$ is zero on $\Ad(w_i)\mathfrak{n}\cap\mathfrak{n}_0$. \item The module $J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$ is not zero. \end{enumerate} \end{lem} \begin{proof} (2) By Lemma~\ref{lem:acts nilp} and the definition of $J'_\eta$, if $I_i/I_{i - 1} \ne 0$ then $\eta = \eta'$. By the definition of $\eta'$, $\eta = \eta'$ is equivalent to $\eta|_{\Ad(w_i)\mathfrak{n}\cap\mathfrak{n}_0} = 0$. (3) This is clear from Lemma~\ref{lem:succ quot is sub of I'}. (1) It is sufficient to prove that if $\eta$ is not unitary then $J'_\eta(V) = 0$ for all irreducible representation $V$ of $G$. By Casselman's subrepresentation theorem, $V$ is a subrepresentation of some principal series representation. Since $J'_\eta$ is an exact functor, we may assume $V$ is a principal series representation $\Ind_{P_0}^G(\sigma_0\otimes e^{\lambda_0 + \rho_0})$. Take the Bruhat filtration $\{I_i\}$ of $J'_\eta(V)$. We prove $I_i/I_{i - 1} = 0$ for all $i$. By (2), if $\eta$ is non-trivial on $w_iN_0w_i^{-1}\cap N_0$ then $I_i/I_{i - 1} = 0$. Hence we may assume that $\eta$ is not unitary on $w_i\overline{N_0}w_i^{-1}\cap N_0$. In this case, an nonzero element of $I_i'$ is not tempered. Hence $I_i/I_{i - 1} = 0$. \end{proof} \begin{rem} In the next section it is proved that the conditions of Lemma~\ref{lem:vanishing lemma} is also sufficient (Theorem~\ref{thm:succ quot is I'_i}). \end{rem} \begin{defn}[Whittaker vectors]\label{defn:Whittaker vectors} Let $V$ be a $U(\mathfrak{n}_0)$-module. We define a vector space $\Wh_\eta(V)$\newsym{$\Wh_\eta(V)$} by \[ \Wh_\eta(V) = \{v\in V\mid \text{for all $X\in\mathfrak{n}_0$ we have $Xv = \eta(X)v$}\}. \] An element of $\Wh_\eta(V)$ is called a \emph{Whittaker vector}. \end{defn} \begin{lem}\label{lem:Whittaker vector, no differential part} Assume that $\eta|_{\Ad(w_i)\mathfrak{n}\cap \mathfrak{n}_0} = 0$. Then we have \begin{multline*} \Wh_\eta\left(\left\{\sum_s(f_s\eta_i^{-1}\otimes u'_s)\delta_i\mid f_s\in\mathcal{P}(O_i),\ u'_s\in J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})\right\}\right) \\ = \{(\eta_i^{-1}\otimes u')\delta_i\mid u'_s\in \Wh_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})\}. \end{multline*} \end{lem} \begin{proof} By the assumption, we have $\eta = \eta'$. Hence the right hand side is a subspace of the left hand side by Lemma~\ref{lem:caluculation of Xdelta(1,f,u)}. Take $x = \sum_s (f_s\eta_i^{-1}\otimes u'_s) = \sum_s \delta(1,f_s\eta_i^{-1},u'_s)\in \Wh_\eta(I'_i)$. We assume that $\{u'_s\}$ is linearly independent. Take $X\in \Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0$. Since $\ad(e)^\mathbf{k}X\in \Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0$ for all $\mathbf{k}\in\mathbb{Z}_{\ge 0}^l$, we have $\sum_s\delta_i(1,L_X(f_s)\eta_i^{-1},u'_s) = 0$ by Lemma~\ref{lem:caluculation of Xdelta(1,f,u)}. Hence $L_X(f_s) = 0$. This implies $f_s\in \mathbb{C}$. From the above argument, $x = \delta(1,\eta_i^{-1},u')$ for some $u'\in J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$. Take $X\in \Ad(w_i)\mathfrak{m}\cap \mathfrak{n}_0$. By Lemma~\ref{lem:caluculation of Xdelta(1,f,u)}, we have \[ \delta_i(1,\eta_i^{-1},(\Ad(w_i)^{-1}X - \eta(X))u')\in \sum_{\mathbf{k}\ne 0,\ u_\mathbf{k}\in J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})}\delta_i(1,f_\mathbf{k}\eta_i^{-1},u_{\mathbf{k}}). \] If $\mathbf{k}\ne 0$ then the degree of $f_\mathbf{k}$ is greater than $0$. So the left hand side must be $0$. Hence we have $(\Ad(w_i)^{-1}X - \eta(X))u' = 0$. We have the lemma. \end{proof} The following lemma is well-known, but we give a proof for the readers (cf.\ Casselman-Hecht-Milicic~\cite{MR1767896}, Yamashita~\cite{MR849220}). \begin{lem}\label{lem:Whittaker vectors in non-degenerate case} Assume that $\supp\eta = \Pi$. Let $x\in \Wh_\eta(I(\sigma,\lambda)')$. Then there exists $u'\in \Wh_{w_r^{-1}\eta}((\sigma\otimes e^{\lambda+\rho})')$ such that $x = (\eta_r^{-1}\otimes u')\delta_r$. \end{lem} Recall that $r = \# W(M) = \#(W/W_M)$. \begin{proof} Assume that $i < r$. Then $w_iw_{M,0}$ is not the longest element of $W$. There exists a simple root $\alpha\in\Pi$ such that $s_\alpha w_iw_{M,0} > w_iw_{M,0}$. This means that $w_iw_{M,0}\Sigma^+\cap \Sigma^+ = s_\alpha(s_\alpha w_iw_{M,0}\Sigma^+\cap \Sigma^+)\cup \{\alpha\}$. The left hand side is $w_i(\Sigma^+\setminus\Sigma_M^+)\cap \Sigma^+$. Hence, $\eta$ is not trivial on $\Ad(w_i)\mathfrak{n}\cap\mathfrak{n}_0$. By Lemma~\ref{lem:vanishing lemma}, $I_i/I_{i - 1} = 0$. This implies that $J'_\eta(I(\sigma,\lambda))\subset I_r'$. There exists a polynomial $f_s\in\mathcal{P}(X_r)$ and $u_s'\in J'_{w_r^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$ such that $x = \sum_s((f_s\eta_r^{-1})\otimes u'_s)\delta_r$. By Lemma~\ref{lem:Whittaker vector, no differential part}, we have the lemma. \end{proof} \section{Analytic continuation}\label{sec:Analytic continuation} The aim of this section is to prove that $\im\Res_i = I'_i$ if $I_i/I_{i - 1} \ne 0$. Let $P_\eta$ be the parabolic subgroup corresponding to $\supp\eta \subset \Pi$ containing $P_0$ and $P_\eta = M_\eta A_\eta N_\eta$ its Langlands decomposition such that $A_\eta\subset A_0$.\newsym{$P_\eta = M_\eta A_\eta N_\eta$} Denote the complexification of the Lie algebra of $P_\eta$, $M_\eta$, $A_\eta$, $N_\eta$ by $\mathfrak{p}_\eta$, $\mathfrak{m}_\eta$, $\mathfrak{a}_\eta$, $\mathfrak{n}_\eta$, respectively.\newsym{$\mathfrak{p}_\eta = \mathfrak{m}_\eta\oplus\mathfrak{a}_\eta\oplus\mathfrak{n}_\eta$} Put $\mathfrak{l}_\eta = \mathfrak{m}_\eta\oplus\mathfrak{a}_\eta$, $\overline{N_\eta} = \theta(N_\eta)$ and $\overline{\mathfrak{n}_\eta} = \theta(\mathfrak{n}_\eta)$.\newsym{$\mathfrak{l}_\eta$}\newsym{$\overline{N_\eta}$}\newsym{$\overline{\mathfrak{n}_\eta}$} Set $\Sigma^+_\eta = \{\sum_{\alpha\in\supp\eta}n_\alpha\alpha\in\Sigma^+\mid n_\alpha\in\mathbb{Z}_{\ge 0}\}$ and $\Sigma^-_\eta = -\Sigma^+_\eta$.\newsym{$\Sigma^+_\eta,\Sigma^-_\eta$} The same notation will be used for $M$ with suffix $M$, i.e., $P_{M,\eta} = M_{M,\eta}A_{M,\eta}N_{M,\eta}$ is the parabolic subgroup of $M$ containing $M\cap P_0$ corresponding to $\supp\eta\cap\Sigma_M^+$, $\mathfrak{p}_{\mathfrak{m},\eta} = \mathfrak{m}_{\mathfrak{m},\eta}\oplus\mathfrak{a}_{\mathfrak{m},\eta}\oplus\mathfrak{n}_{\mathfrak{m},\eta}$ is a complexification of the Lie algebra of $P_{M,\eta} = M_{M,\eta}A_{M,\eta}N_{M,\eta}$.% \newsym{$P_{M,\eta} = M_{M,\eta}A_{M,\eta}N_{M,\eta}$}\newsym{$\mathfrak{p}_{\mathfrak{m},\eta} = \mathfrak{m}_{\mathfrak{m},\eta}\oplus\mathfrak{a}_{\mathfrak{m},\eta}\oplus\mathfrak{n}_{\mathfrak{m},\eta}$} For $w\in W$, there is an open dense subset $w\overline{N}P/P$ of $G/P$ and it is diffeomorphic to $\overline{N}$. Then for $w,w'\in W$, there exists a map $\Phi_{w,w'}$\newsym{$\Phi_{w,w'}$} from some open dense subset $U\subset\overline{N}$ to $\overline{N}$ such that $w\overline{n}P/P = w'\Phi_{w,w'}(\overline{n})P/P$ for $\overline{n}\in U$. The map $\Phi_{w,w'}$ is a rational function. Since the exponential map $\exp\colon\Lie(\overline{N})\to \overline{N}$ is diffeomorphism, $\overline{N}$ has a structure of a vector space. \begin{lem}\label{lem:property of e^rho} \begin{enumerate} \item The map $\overline{N}\to \mathbb{R}$ defined by $\overline{n}\mapsto e^{8\rho(H(\overline{n}))}$ is a polynomial. \item For all $\overline{n}\in\overline{N}$ we have $e^{8\rho(H(\overline{n}))} \ge 1$. \item Take $H_0\in \mathfrak{a}$ such that $\alpha(H_0) = -1$ for all $\alpha\in \Pi\setminus\Sigma_M$. There exists a continuous function $Q(\overline{n})\ge 0$ on $\overline{N}$ such that the following conditions hold: (a) The function $Q$ vanishes only at the unit element. (b) $e^{8\rho(H(\overline{n}))} \ge Q(\overline{n})$. (c) $Q(\exp(tH_0)\overline{n}\exp(-tH_0)) \ge e^{8t}Q(\overline{n})$ for $t\in \mathbb{R}_{>0}$ and $\overline{n}\in\overline{N}$. \end{enumerate} \end{lem} \begin{proof} By Knapp~\cite[Proposition~7.19]{MR1880691}, there exists an irreducible finite-dimensional $V_{4\rho}$ of $\mathfrak{g}$ with the highest weight $4\rho\in \mathfrak{a}_0^*\subset \mathfrak{h}^*$. Let $v_{4\rho}\in V_{4\rho}$ be a highest weight vector and $v_{-4\rho}^*\in V_{4\rho}^*$ the lowest weight vector of $V_{4\rho}^*$. Then $\mathbb{C} v_{4\rho}$ is a $1$-dimensional unitary representation of $M$. Take $\overline{n}\in \overline{N}$ and decompose $\overline{n} = kan$ where $k\in K$, $a\in A_0$ and $n\in N_0$. First we prove (1). We have $\theta(\overline{n})^{-1}\overline{n} = \theta(n)^{-1}a^2n$. Hence \begin{align*} \langle \theta(\overline{n})^{-1}\overline{n}v_{4\rho},v^*_{-4\rho}\rangle & = \langle \theta(n)^{-1}a^2nv_{4\rho},v^*_{-4\rho}\rangle\\ & = \langle a^2nv_{4\rho},\theta(n)v^*_{-4\rho}\rangle\\ & = e^{8\rho(H(\overline{n}))}\langle v_{4\rho},v^*_{-4\rho}\rangle. \end{align*} The left hand side is a polynomial. Next we prove (2) and (3). Fix a compact real form of $\mathfrak{g}$ containing $\Lie(K)$ and take an inner product on $V_{4\rho}$ which is invariant under this compact real form. We normalize an inner product $||\cdot||$ so that $||v_{4\rho}|| = 1$. Then we have $||\overline{n}v_{4\rho}|| = ||kanv_{4\rho}|| = ||av_{4\rho}|| = e^{4\rho(H(\overline{n}))}||v_{4\rho}|| = e^{4\rho(H(\overline{n}))}$. For $\nu\in\mathfrak{h}^*$ let $Q_\nu(\overline{n})\in V_{4\rho}$ be the $\nu$-weight vector such that $\overline{n}v_{4\rho} = \sum_\nu Q_{\nu}(\overline{n})$. Then we have $e^{8\rho(H(\overline{n}))} = \sum_\nu ||Q_\nu(\overline{n})||^2$. Since $Q_{4\rho}(\overline{n}) = v_{4\rho}$, we have $e^{8\rho(H(\overline{n}))} \ge 1$. Put $Q(\overline{n}) = \sum_{w\in W(M)\setminus\{e\}}||Q_{4w\rho}(\overline{n})||^2$. Assume that $\overline{n} \ne e$. Then there exist $w\in W(M)\setminus\{e\}$, $m'\in M$, $a'\in A$, $n'\in N$ and $\overline{n}'\in \overline{N}$ such that $\overline{n} = w\overline{n}'m'a'n'$. Let $v_{-4w\rho}^*\in V_{4\rho}^*$ be a weight vector with $\mathfrak{h}$-weight $-4w\rho$. Then we have \begin{multline*} ||Q_{4w\rho}(\overline{n})|| = \lvert\langle \overline{n}v_{4\rho},v_{-4w\rho}^*\rangle\rvert = \lvert\langle w\overline{n}'m'a'n'v_{4\rho},v_{-4w\rho}^*\rangle\rvert\\ = \lvert\langle a'v_{4\rho},w^{-1}v^*_{-4w\rho}\rangle\rvert = e^{4\rho(\log a')}\lvert \langle v_{4\rho},w^{-1}v_{-4w\rho}^*\rangle\rvert \ne 0. \end{multline*} Hence, if $\overline{n}\in \overline{N}\setminus\{e\}$ then $Q(\overline{n})\ne 0$. Let $t$ be a positive real number. Using $Q_\nu(\exp(tH_0)\overline{n}\exp(-tH_0)) = e^{t(\nu - 4\rho)(H_0)}Q_\nu(\overline{n})$, we have \[ Q(\exp(tH_0)\overline{n}\exp(-tH_0)) = \sum_{w\in W(M)\setminus\{e\}} e^{8t(w\rho - \rho)(H_0)}\lvert Q_{4w\rho_0}(\overline{n})\rvert^2. \] Since $(w\rho - \rho)(H_0) \ge 1$ for $w\in W(M)\setminus\{e\}$, we get the lemma. \end{proof} \begin{rem}\label{rem:asymptotic at infinity of e^rho} The condition Lemma~\ref{lem:property of e^rho} (3) implies that $\lim_{\overline{n}\to \infty}Q(\overline{n}) = \infty$. The proof is the following. Take $H_0$ as in Lemma~\ref{lem:property of e^rho}. Let $\{e_1,\dots,e_l\}$ be a basis of $\overline{\mathfrak{n}}$. Here, we assume that each $e_i$ is a restricted root vector and denote its root by $\alpha_i$. Any $\overline{n}\in \overline{N}$ can be written as $\overline{n} = \exp(\sum_{i = 1}^l a_ie_i)$ where $a_i\in \mathbb{R}$. Put $r(\overline{n}) = \sum_{i = 1}^l\lvert a_i\rvert^{-1/\alpha_i(H_0)}$. Set $C = \min_{r(\overline{n}) = 1}Q(\overline{n})$. Since $Q(\overline{n}) > 0$ if $\overline{n}$ is not the unit element, $C > 0$. Then we have $Q(\overline{n})\ge Cr(\overline{n})^8$ if $r(\overline{n}) > 1$. If $\overline{n}\to \infty$ then $r(\overline{n})\to\infty$. Hence, $Q(\overline{n})\to\infty$. \end{rem} \begin{lem}\label{lem:extension of polynomials} Let $f$ be a polynomial on $\overline{N}$. There exists a positive integer $k$ and a $C^\infty$-function $h$ on $G/P$ such that $h(w_i\overline{n}P/P) = e^{-k\rho(H(\overline{n}))}f(\overline{n})$ for all $\overline{n}\in\overline{N}$. \end{lem} \begin{proof} By Lemma~\ref{lem:property of e^rho} and Remark~\ref{rem:asymptotic at infinity of e^rho}, we can choose a positive integer $C$ such that $e^{-8C\rho(H(\overline{n}))}f(\overline{n}) \to 0$ when $\overline{n}\to \infty$. Let $\widetilde{f}$ be a function on $U_i$ defined by $\widetilde{f}(w_i\overline{n}P/P) = e^{-8C\rho(H(\overline{n}))}f(\overline{n})$ for $\overline{n}\in \overline{N}$. We prove that $\widetilde{f}$ can be extended to $G/P$. Take $w\in W(M)$. Then $\widetilde{f}$ is defined in a subset of $w\overline{N}P/P$. Using a diffeomorphism $\overline{N}\simeq w\overline{N}P/P$, $\widetilde{f}$ defines a rational function $\widetilde{f}\circ \Phi_{w_i,w}$ defined on an open dense subset of $\overline{N}$. By the condition of $C$, the function $\widetilde{f}\circ \Phi_{w_i,w}$ has no pole. Hence, $\widetilde{f}$ defines a $C^\infty$-function on $w\overline{N}P/P$. Since $\bigcup_{w\in W(M)}w\overline{N}P/P = G/P$, the lemma follows. \end{proof} Define $\kappa\colon G\to K$ and $H\colon G\to \Lie(A_0)$ by $g\in \kappa(g)\exp H(g) N_0$.\newsym{$\kappa$}\newsym{$H$} Recall that for a representation $V$ of $\mathfrak{g}$, $\nu\in \mathfrak{a}_0^*$ is called an exponent of $V$ if $\nu + \rho_0|_{\mathfrak{m}\cap\mathfrak{a}_0}$ is an $\mathfrak{a}_0$-weight of $V/\mathfrak{n}_0V$. \begin{prop}\label{prop:convergence and continuation} Let $\varphi$ be a $\sigma$-valued function on $K$ which satisfies $\varphi(km) = \sigma(m)^{-1}\varphi(k)$ for all $k\in K$ and $m\in M\cap K$. We define $\varphi_\lambda\in I(\sigma,\lambda)$ by $\varphi_\lambda(kman) = e^{-(\lambda + \rho)(\log a)}\sigma(m)^{-1}\varphi(k)$ for $k\in K$, $m\in M$, $a\in A$ and $n\in N$. For $u'\in J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$ and $f\in\mathcal{P}(O_i)$, put $I_{f,u'}(\varphi_\lambda) = \int_{w_i\overline{N}w_i^{-1}\cap N_0}u'(\varphi_\lambda(nw_i))\eta(n)^{-1}f(nw_i)dn$. (If $\supp\varphi\subset K\cap w_i\overline{N}P$ then the integral converges.) \begin{enumerate} \item If $\langle\alpha,\re\lambda\rangle$ is sufficiently large for each $\alpha\in\Sigma^+\setminus\Sigma^+_M$ then the integral $I_{f,u'}(\varphi_\lambda)$ absolutely converges. \item As a function of $\lambda$, the integral $I_{f,u'}(\varphi_\lambda)$ has a meromorphic continuation to $\mathfrak{a}^*$. \item If $\supp\eta = \Pi$ and $i = r$ then $I_{f,u'}(\varphi_\lambda)$ is holomorphic for all $\lambda\in\mathfrak{a}^*$. \item Let $\nu$ be an exponent of $\sigma$ and $u'\in \Wh_{w_i^{-1}\eta}((\sigma\otimes e^{\lambda+\rho})')$. If $2\langle\alpha,\lambda+\nu\rangle/\lvert\alpha\rvert^2\not\in \mathbb{Z}_{\le 0}$ for all $\alpha\in \Sigma^+\setminus w_i^{-1}(\Sigma^+\cup\Sigma^-_{\eta})$ then $I_{1,u'}(\varphi_\mu)$ is holomorphic at $\mu = \lambda$. \end{enumerate} \end{prop} \begin{proof} First we prove (1). If $f = 1$ then this is a well-known result. For a general $f$, extends $f$ to a function on $w_i\overline{N}P/P$ by $f(w_inn') = f(w_in)$ for $n\in w_i\overline{N}w_i^{-1}\cap N_0$ and $n'\in w_i\overline{N}w_i^{-1}\cap \overline{N_0}$. Then by Lemma~\ref{lem:extension of polynomials} there exists a positive number $C$ such that $\overline{n}\mapsto e^{-C\rho(H(\overline{n}))}f(w_i\overline{n})$ extends to a function $h$ on $G/P$. Since \[ I_{f,u'}(\varphi_\lambda) = \int_{w_i\overline{N}w_i^{-1}\cap N_0}u'(\varphi(\kappa(nw))e^{-(\lambda+\rho)(H(nw_r))}f(nw_r)\eta(n)^{-1}dn, \] we have $I_{f,u'}(\varphi_\lambda) = I_{1,u'}((\varphi h)_{\lambda - C\rho})$. We prove (3). By dualizing Casselman's subrepresentation theorem, there exist a representation $\sigma_0$ of $M_0$ and $\lambda_0\in\mathfrak{a}_0^*$ such that $\sigma$ is a quotient of $\Ind_{M\cap P_0}^{M}(\sigma_0\otimes e^{\lambda_0})$. Then we may regard $u'\in J'_{w_r^{-1}\eta}(\Ind_{M\cap P_0}^M(\sigma_0\otimes e^{\lambda_0}))$. By the proof of Lemma~\ref{lem:Whittaker vectors in non-degenerate case}, there exist a polynomial $f_0$ on $(M\cap N_0)w_{M,0}(M\cap P_0)/(M\cap P_0)$ and $u'_0\in (\sigma_0\otimes e^{\lambda_0})'$ such that $u'$ is given by \[ \varphi_0\mapsto \int_{M\cap N_0}u'_0(\varphi_0(n_0w_{M,0}))f_0(n_0w_{M,0})\eta(n_0)^{-1}dn_0 \] Let $\pi\colon \Ind_{P_0}^G(\sigma_0\otimes e^{\lambda + \lambda_0 + \rho})\to I(\sigma,\lambda)$ be the map induced from the quotient map $\Ind_{M\cap P_0}^{M}(\sigma_0\otimes e^{\lambda_0})\to \sigma$. Take $\widetilde{\varphi}\colon K\to \sigma_0$ which satisfies $\widetilde{\varphi}(km) = \sigma_0^{-1}(m)\widetilde{\varphi}(k)\ (k\in K,\ m\in M_0)$ and $\pi(\widetilde{\varphi}_{\lambda + \lambda_0}) = \varphi_\lambda$. Define a polynomial $\widetilde{f}\in \mathcal{P}(w_iw_{M,0}\overline{N_0}P_0/P_0)$ by \[ \widetilde{f}(w_iw_{M,0}nn_0P_0/P_0) = f(w_inP/P)f_0(w_{M,0}n_0(M\cap P_0)/(M\cap P_0)) \] for $n\in \overline{N}$ and $n_0\in M\cap \overline{N_0}$. (Notice that $w_{M,0}(M\cap \overline{N_0}) = (M\cap N_0)w_{M,0}$.) Then we have \begin{multline*} I_{f,u'}(\varphi_\lambda)\\ = \int_{w_iw_{M,0}\overline{N_0}(w_iw_{M,0})^{-1}\cap N_0}u_0'(\widetilde{\varphi}(nw_iw_{M,0}))\widetilde{f}(w_iw_{M,0}nP_0/P_0)\eta(n)^{-1}dn. \end{multline*} Hence, we may assume that $P$ is minimal. By the same argument in (1), we may assume $f = 1$. If $f = 1$ then this integral is known as a Jacquet integral and the analytic continuation is well-known~\cite{MR0271275}. We prove (2) and (4). By the same argument in (1), we may assume that $f = 1$. Take $w'\in W_{M_\eta}$ and $w''\in W(M_\eta)^{-1}$ such that $w_i = w'w''$. Then we have $w_i\overline{N}w_i^{-1}\cap N_0 = (w'\overline{N_0}(w')^{-1}\cap N_0)w'(w''\overline{N_0}(w'')^{-1}\cap N_0)(w')^{-1}$. The condition $w'\in W_{M_\eta}$ implies that $w'(\Sigma^+\setminus\Sigma^+_\eta) = \Sigma^+\setminus\Sigma^+_\eta$. Hence, $\supp\eta \cap w'\Sigma^+ = \supp\eta\cap w'\Sigma_\eta^+$. This implies \begin{multline*} \supp\eta\cap w'(w''\Sigma^-\cap \Sigma^+) = \supp\eta\cap w_i\Sigma^-\cap w'\Sigma^+\\ = \supp\eta\cap w_i\Sigma^-\cap w_i(w'')^{-1}\Sigma_\eta^+\subset \supp\eta\cap w_i\Sigma^-\cap w_i\Sigma^+ = \emptyset, \end{multline*} i.e., $\eta$ is trivial on $w'(w''\overline{N_0}(w'')^{-1}\cap N_0)(w')^{-1}$. Hence, we have \[ I_{1,u'}(\varphi) = \int_{w'\overline{N_0}(w')^{-1}\cap N_0}\int_{w''\overline{N_0}(w'')^{-1}\cap N_0}u'(\varphi(n_1w'n_2w''))\eta(n_1)^{-1}dn_2dn_1. \] Put $P' = (w''P(w'')^{-1}\cap M_\eta)N_\eta$. By the definition of $W(M_\eta)$, we have $w''N_0(w'')^{-1}\supset N_0\cap M_\eta$, this implies that $P'$ (resp.\ $w''P(w'')^{-1}\cap M_\eta$) is a parabolic subgroup of $G$ (resp.\ $M_\eta$). Define a $G$-module homomorphism $A(\sigma,\lambda)\colon I(\sigma,\lambda)\to \Ind_{P'}^G(w''(\sigma)\otimes e^{w''\lambda+\rho})$ by \[ (A(\sigma,\lambda)\varphi)(x) = \int_{w''\overline{N_0}(w'')^{-1}\cap N_0}\varphi(xnw'')dn. \] By a result of Knapp and Stein~\cite{MR582703}, this homomorphism has a meromorphic continuation. We have \[ I_{1,u'}(\varphi) = \int_{w'\overline{N_0}(w')^{-1}\cap N_0}u'((A(\sigma,\lambda)\varphi)(nw'))\eta(n)^{-1}dn. \] Notice that $w'\overline{N_0}(w')^{-1}\cap N_0\subset M_\eta$. Hence we get (2) by (3). To prove (4), we calculate $(w'')^{-1}\Sigma^-\cap\Sigma^+$. Since $(w'')^{-1}\in W(M_\eta)$, we have $(w'')^{-1}\Sigma_\eta^-\subset \Sigma^-$. Hence $(w'')^{-1}\Sigma_\eta^-\cap \Sigma^+ = \emptyset$. Then \begin{align*} (w'')^{-1}\Sigma^-\cap \Sigma^+ & = (w'')^{-1}(\Sigma^-\setminus\Sigma_\eta^-)\cap \Sigma^+\\ & = (w'')^{-1}(w')^{-1}(\Sigma^-\setminus\Sigma_\eta^-)\cap\Sigma^+\\ & = w_i^{-1}(\Sigma^-\setminus\Sigma_\eta^-)\cap\Sigma^+\\ & = \Sigma^+\setminus w_i^{-1}(\Sigma^+\cup\Sigma_\eta^-). \end{align*} Hence we have $2\langle\alpha,\lambda+\nu\rangle/\lvert\alpha\rvert^2\not\in\mathbb{Z}_{\ge 0}$ for all $\alpha\in(w'')^{-1}\Sigma^-\cap\Sigma^+$. By an argument of Knapp and Stein~\cite{MR582703}, $A(\sigma,\mu)$ is holomorphic at $\mu = \lambda$ if $\lambda$ satisfies the conditions of (4). Hence we get (4). \end{proof} In the rest of this section, we denote the Bruhat filtration $I_i\subset J'(I(\sigma,\lambda))$ by $I_i(\lambda)$. The following result is a corollary of Proposition~\ref{prop:convergence and continuation}. \begin{lem}\label{lem:meromorphic extension} Let $x\in I_i'$. Then there exists a distribution $x_t\in I_i(\lambda + t\rho)$ with meromorphic parameter $t$ such that $x_t|_{U_i}$ is a distribution with holomorphic parameter $t$ and $(x|_{U_i})|_{t = 0} = x$. Moreover, if $Tx = 0$ for $T\in U(\mathfrak{g})$, then $Tx_t = 0$. \end{lem} Let $C^\infty(K,\sigma)$ be the space of $\sigma$-valued $C^\infty$-functions. For $X\in \mathfrak{g}$ and $\lambda\in\mathfrak{a}^*$, we define an operator $D(X,\lambda)$ on $C^\infty(K,\sigma)$ as follows. For $\varphi\in C^\infty(K,\sigma)$, \begin{multline*} (D(X,\lambda)\varphi)(k)\\ = \frac{d}{dt}\left.(\sigma\otimes e^{\lambda+\rho})(\exp (-H(\exp(-tX)k)))\varphi(\kappa(\exp(-tX)k))\right|_{t = 0}.\newsym{$D(X,\lambda)$} \end{multline*} If we regard $I(\sigma,\lambda)$ as a subspace of $C^\infty(K,\sigma)$, $(X\varphi)(k) = (D(X,\lambda)\varphi)(k)$ for $\varphi\in I(\sigma,\lambda)$. It is easy to see that for some $D_1$ and $D_2$ we have $D(X,\lambda + t\rho) = D_1 + tD_2$ for all $t\in\mathbb{C}$. \begin{lem}\label{lem:holomorphic extension} Assume that the conditions of Lemma~\ref{lem:vanishing lemma} (1)--(3) hold. For $x\in I'_i$ there exists a distribution $x_t\in I_i(\lambda + t\rho)$ with holomorphic parameter $t$ defined near $t = 0$ such that $x_0 = x$ on $U_i$. \end{lem} \begin{proof} First we remark that $\eta = \eta'$ in Lemma~\ref{lem:acts nilp} by the condition (2) of Lemma~\ref{lem:vanishing lemma}. We prove by induction on $i$. If $i = 1$, then $x\in I'_1$. Take a distribution $x_t\in I_1(\lambda + t\rho)$ as in Lemma~\ref{lem:meromorphic extension}. Then $x_t|_{U_1}$ is holomorphic with respect to the parameter $t$. Since $\supp x_t\subset X_1$, $x_t|_{(G/P)\setminus X_1}$ is holomorphic with respect to the parameter $t$. Hence $x_t$ is holomorphic with respect to the parameter $t$ on $U_1\cup ((G/P)\setminus X_1) = G/P$. We have the lemma. Assume that $i > 1$. First we prove the following claim: for $y\in I_{i - 1}$, there exists a distribution $y_t\in I_{i - 1}(\lambda + t\rho)$ with holomorphic parameter $t$ defined near $t = 0$ such that $y_0 = y$. Using inductive hypothesis to $y|_{U_{i - 1}}$, there exists a distribution $y_t^{(i - 1)}\in I_{i - 1}(\lambda + t\rho)$ with holomorphic parameter $t$ defined near $t = 0$ such that $y_0^{(i - 1)} = y$ on $U_{i - 1}$. Since the supports of both sides are contained in $\bigcup_{j\le i - 1}N_0w_jP/P$, we have $y_0^{(i - 1)} = y$ on $\bigcup_{j\ge i - 1}N_0w_jP/P$. Using inductive hypothesis to $(y - y_0^{(i - 1)})|_{U_{i - 2}}$, there exists a distribution $y_t^{(i - 2)}\in I_{i - 2}(\lambda + t\rho)$ with holomorphic parameter $t$ defined near $t = 0$ such that $y_0^{(i - 2)} = y - y_0^{(i - 1)}$ on $U_{i - 2}$. Since the supports of both sides are contained in $\bigcup_{j\le i - 2}N_0w_jP/P$, we have $y_0^{(i - 1)} + y_0^{(i - 2)} = y$ on $\bigcup_{j\ge i - 2}N_0w_jP/P$. Iterating this argument, for $j = 1,\dots,i - 1$ there exists a distribution $y_t^{(j)}\in I_j(\lambda + t\rho)$ with holomorphic parameter $t$ defined near $t = 0$ such that $y = y_0^{(1)} + \dots + y_0^{(i - 1)}$. Hence we get the claim. Now we prove the lemma. By Lemma~\ref{lem:meromorphic extension}, there exists a distribution $x'_t\in I_i(\lambda + t\rho)$ with meromorphic parameter $t$ such that $x'_t|_{U_i}$ is holomorphic and $(x'_t|_{U_i})|_{t = 0} = x$. Let $x'_t = \sum_{s = -p}^\infty x^{(s)}t^s$ be the Laurent series of $x'_t$. Now we prove the following claim: if there exists a distribution $x'_t = \sum_{s = -p}^\infty x^{(s)}t^s\in I_i(\lambda + t\rho)$ with meromorphic parameter $t$ defined near $t = 0$ such that $x'_t|_{U_i}$ is holomorphic and $(x'_t|_{U_i})|_{t = 0} = x$, then there exists $x_t\in I(\lambda)$ with holomorphic parameter $t$ defined near $t = 0$ such that $x_0|_{U_i} = x$. We prove the claim by induction on $p$. If $p = 0$, we have nothing to prove. Assume $p > 0$. Take $E\in \mathfrak{n}_0$ and define differential operators $E_0$ and $E_1$ by $D(E,\lambda + t\rho) = E_0 + tE_1$. By Lemma~\ref{lem:acts nilp}, there exists a positive integer $k$ such that $(E_0 + tE_1 - \eta(E))^k x'_t = 0$. Hence, we have $(E_0 - \eta(E))^kx^{(-p)} = 0$. Since $x_t|_{U_i}$ is holomorphic, we have $\supp x^{(-p)}\subset \bigcup_{j < i}N_0w_jP/P$. Hence we have $x^{(-p)}\in I_{i - 1}$. By the claim stated in the third paragraph of this proof, there exists $x''_t\in I_{i - 1}(\lambda + t\rho)$ with holomorphic parameter $t$ defined near $t = 0$ such that $x''_0 = x^{(-p)}$. Using inductive hypothesis for $x_t' - t^{-p}x''_t$, we get the claim and the claim implies the lemma. \end{proof} \begin{thm}\label{thm:succ quot is I'_i} \begin{enumerate} \item The module $I_i/I_{i - 1}$ is non-zero if and only if the conditions of Lemma~\ref{lem:vanishing lemma} (1)--(3) hold. \item If $I_i/I_{i - 1} \ne 0$ then we have $I_i/I_{i - 1} \simeq I'_i$. \end{enumerate} \end{thm} \begin{proof} Assume that the conditions of Lemma~\ref{lem:vanishing lemma} (1)--(3) hold. We prove that the restriction map $\Res_i\colon I_i\to I'_i$ is surjective. For $x\in I'_i$, take $x_t\in I_i(\lambda + t\rho)$ as in Lemma~\ref{lem:holomorphic extension}. Then we have $\Res_i(x_0) = (x_0)|_{U_i} = x$. Hence $\Res_i$ is surjective. \end{proof} \section{Twisting functors}\label{sec:Twisting functors} Arkhipov defined the \emph{twisting functor} for $\widetilde{w}\in\widetilde{W}$~\cite{MR2074588}. In this section, we define a modification of the twisting functor. Let $\mathfrak{g}^\mathfrak{h}_\alpha$ be the root space of $\alpha\in\Delta$.\newsym{$\mathfrak{g}^\mathfrak{h}_\alpha$} Set $\mathfrak{u}_0 = \bigoplus_{\alpha\in \Delta^+}\mathfrak{g}^\mathfrak{h}_\alpha$\newsym{$\mathfrak{u}_0$}, $\overline{\mathfrak{u}_0} = \bigoplus_{\alpha\in \Delta^+}\mathfrak{g}^\mathfrak{h}_{-\alpha}$\newsym{$\overline{\mathfrak{u}_0}$} and $\mathfrak{u}_{0,{\widetilde{w}}} = \Ad(\widetilde{w})\overline{\mathfrak{u}_0}\cap \mathfrak{u}_0$.\newsym{$\mathfrak{u}_{0,{\widetilde{w}}}$} Let $\psi$ be a character of $\mathfrak{u}_{0,{\widetilde{w}}}$. Put $S_{\widetilde{w},\psi} = U(\mathfrak{g})\otimes_{U(\mathfrak{u}_{0,{\widetilde{w}}})}((U(\mathfrak{u}_{0,{\widetilde{w}}})^*)_{\text{$\mathfrak{h}$-finite}}\otimes_\mathbb{C}\psi)$.\newsym{$S_{\widetilde{w},\psi}$} This is a right $U(\mathfrak{u}_{0,{\widetilde{w}}})$-module and left $U(\mathfrak{g})$-module. We define a $U(\mathfrak{g})$-bimodule structure on $S_{\widetilde{w},\psi}$ in the following way. Let $\{e_1,\dots,e_l\}$ be a basis of $\mathfrak{u}_{0,{\widetilde{w}}}$ such that each $e_i$ is a root vector and $\bigoplus_{s\le t - 1}\mathbb{C} e_s$ is an ideal of $\bigoplus_{s\le t}\mathbb{C} e_s$ for each $t = 1,2,\dots,l$. Notice that a multiplicative set $\{(e_k - \psi(e_k))^n\mid n\in\mathbb{Z}_{\ge 0}\}$ satisfies the Ore condition for $k = 1,2,\dots,l$. Then we can consider the localization of $U(\mathfrak{g})$ by $\{(e_k - \psi(e_k))^n\mid n\in\mathbb{Z}_{\ge 0}\}$. We denote the resulting algebra by $U(\mathfrak{g})_{e_k - \psi(e_k)}$. Put $S_{e_k - \psi(e_k)} = U(\mathfrak{g})_{e_k - \psi(e_k)}/U(\mathfrak{g})$. Then $S_{e_k - \psi(e_k)}$ is a $U(\mathfrak{g})$-bimodule.\newsym{$S_{e_k - \eta(e_k)}$} \begin{prop}\label{prop:1-dim decomposition of S_w} As a right $U(\mathfrak{u}_{0,w})$-module and left $U(\mathfrak{g})$-module, we have $S_{\widetilde{w},\psi} \simeq S_{e_1 - \psi(e_1)}\otimes_{U(\mathfrak{g})}S_{e_2 - \psi(e_2)}\otimes_{U(\mathfrak{g})}\dots\otimes_{U(\mathfrak{g})}S_{e_l - \psi(e_l)}$. Moreover, the $U(\mathfrak{g})$-bimodule structure induced from this isomorphism is independent of a choice of $e_i$. \end{prop} The proof of this proposition is similar to that of Arkhipov~\cite[Thoerem~2.1.6]{MR2074588}. We omit it. An element of the right hand side is written as a sum of a form $(e_1 - \eta(e_1))^{-(k_1 + 1)}\otimes\dots\otimes (e_l - \eta(e_l))^{-(k_l + 1)}T$ for $T\in U(\mathfrak{g})$. We denote this element by $(e_1 - \eta(e_1))^{-(k_1 + 1)}\dotsm (e_l - \eta(e_l))^{-(k_l + 1)}T$ for short. Proposition~\ref{prop:1-dim decomposition of S_w} gives the $U(\mathfrak{g})$-bimodule structure of $S_{\widetilde{w},\psi}$. For a $U(\mathfrak{g})$-module $V$, we define a $U(\mathfrak{g})$-module $T_{\widetilde{w},\psi}V$ by $T_{\widetilde{w},\psi}V = S_{\widetilde{w},\psi}\otimes_{U(\mathfrak{g})}(\widetilde{w}V)$. (Recall that $\widetilde{w}V$ is a $\mathfrak{g}$-module twisted by $\widetilde{w}$. See Notation.) This gives the twisting functor $T_{\widetilde{w},\psi}$.\newsym{$T_{\widetilde{w},\psi}$} If $\psi$ is the trivial representation, $T_{\widetilde{w},\psi}$ is the twisting functor defined by Arkhipov. We put $T_{\widetilde{w}} = T_{\widetilde{w},0}$ where $0$ is the trivial representation. The restriction map gives a map $N_K(\mathfrak{h})/Z_K(\mathfrak{h})\to W$ and its kernel is isomorphic to $N_{M_0}(\mathfrak{t}_0)/Z_{M_0}(\mathfrak{t}_0)$ (Recall that $\mathfrak{t}_0$ is a Cartan subalgebra of $\mathfrak{m}_0$). The last group is isomorphic to $\widetilde{W_{M_0}}$. \begin{lem}\label{lem:good lift for W} Let $w\in W$. Then there exists $\iota(w)\in N_K(\mathfrak{h})$ such that $\Ad(\iota(w))|_{\mathfrak{a}_0} = w$ and $\Ad(\iota(w))(\Delta_{M_0}^+) = \Delta_{M_0}^+$. \end{lem} \begin{proof} Since $W\simeq N_K(\mathfrak{a}_0)/Z_K(\mathfrak{a}_{0})$, there exists $k\in N_K(\mathfrak{a}_{0})$ such that $\Ad(k)|_{\mathfrak{a}_0} = w$. Then $k$ normalizes $M_0$. Hence, there exists $m\in M_0$ such that $km$ normalizes $T_0$. This implies $km\in N_K(A_0T_0)$. Take $w'\in N_{M_0}(\mathfrak{t}_0)$ such that $\Ad(kmw')(\Delta^+_{M_0}) = \Delta^+_{M_0}$ and put $\iota(w) = kmw'$. Then $\iota(w)$ satisfies the conditions of the lemma. \end{proof} The map $\iota$ gives an injective map $W\to N_K(\mathfrak{h})/Z_K(\mathfrak{h})$. Since the group $N_K(\mathfrak{h})/Z_K(\mathfrak{h})$ can be regard as a subgroup of $\widetilde{W}$, we can regard $W$ as a subgroup of $\widetilde{W}$. Hence, we can define the twisting functor $T_{w,\psi}$ for $w\in W$ and the character $\psi$ of $\Ad(w)\overline{\mathfrak{n}_0}\cap \mathfrak{n}_0$. For a simplicity, we write $w$ instead of $\iota(w)$. (We regard $W$ as a subgroup of $\widetilde{W}$ by $\iota$.) \begin{prop}\label{prop:decomposition of T_ww'} Let $w,w'\in W$ and $\psi$ a character of $\Ad(ww')\overline{\mathfrak{n}_0}\cap \mathfrak{n}_0$. Assume that $\ell(w) + \ell(w') = \ell(ww')$ where $\ell(w)$ is the length of $w\in W$. Then we have $T_{w,\psi}T_{w',w^{-1}\psi} = T_{ww',\psi}$. \end{prop} \begin{proof} By the assumption, we have $\Sigma^+\cap ww'\Sigma^- = (\Sigma^+\cap w\Sigma^-)\cup w(\Sigma^+\cap w'\Sigma^-)$. Put $\Delta_0^\pm = \Delta^\pm\setminus\Delta_{M_0}^\pm$. Then we have $\Delta_0^+\cap ww'\Delta_0^- = (\Delta_0^+\cap w\Delta_0^-)\cup w(\Delta_0^+\cap w'\Delta_0^-)$. Since $w\Delta_{M_0}^\pm = \Delta_{M_0}^\pm$, we have $\Delta_0^+\cap w\Delta_0^- = \Delta^+\cap w\Delta^-$. Hence, $\Delta^+\cap ww'\Delta^- = (\Delta^+\cap w\Delta^-)\cup w(\Delta^+\cap w'\Delta^-)$. This implies that $\widetilde{\ell}(w) + \widetilde{\ell}(w') = \widetilde{\ell}(ww')$ where $\widetilde{\ell}(w)$ is the length of $w$ as an element of $\widetilde{W}$. Hence, the proposition follows from the construction of the twisting functor (See Andersen and Lauritzen~\cite[Remark~6.1 (ii)]{MR1985191}). \end{proof} \begin{lem}\label{lem:Xe^-k} Let $e$ be a nilpotent element of $\mathfrak{g}$, $X\in \mathfrak{g}$ and $k\in\mathbb{Z}_{\ge 0}$. For $c\in \mathbb{C}$ we have the following equation in $U(\mathfrak{g})_{e - c}$. \[ X(e - c)^{-(k + 1)} = \sum_{n = 0}^\infty \binom{n + k}{k}(e - c)^{-(n + k + 1)}\ad (e)^n(X). \] \end{lem} \begin{proof} We prove the lemma by induction on $k$. If $k = 0$, then the lemma is well-known. Assume that $k > 0$. Then we have \begin{align*} X(e - c)^{-(k + 1)} & = \sum_{k_0 = 0}^\infty (e - c)^{-(k_0 + 1)}\ad(e)^{k_0}(X) (e - c)^{-k}\\ & = \sum_{k_0 = 0}^\infty \sum_{k_1 = 0}^\infty \binom{k_1 + k - 1}{k - 1}(e - c)^{-(k_0 + k_1 + k + 1)}\ad(e)^{k_0 + k_1}(X)\\ & = \sum_{n = 0}^\infty \sum_{l' = 0}^n \binom{l' + k - 1}{k - 1}(e - c)^{-(n + k + 1)}\ad(e)^n(X)\\ & = \sum_{n = 0}^\infty \binom{n + k}{k}(e - c)^{-(n + k + 1)}\ad(e)^n(X). \end{align*} This proves the lemma. \end{proof} \section{The module $I_i/I_{i - 1}$}\label{sec:The module I_i/I_i-1} Put $J_i = U(\mathfrak{g})\otimes_{U(\mathfrak{p})}J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$\newsym{$J_i$}, where $\mathfrak{n}$ acts $J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$ trivially. In this section, we prove the following theorem. \begin{thm}\label{thm:structure of I_i/I_{i - 1}} Assume that $I_i/I_{i - 1} \ne 0$. Then we have $I_i/I_{i - 1} \simeq T_{w_i,\eta}J_i$. \end{thm} Notice that $\mathfrak{u}_{0,{w_i}} = \Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0$ since $w_i(\Delta_{M}^+) \subset \Delta^+$. In this section fix $i\in\{1,\dots,l\}$ and a basis $\{e_1,e_2,\dots,e_l\}$ of $\mathfrak{u}_{0,w_i}$ such that each vector $e_i$ is a root vector and $\bigoplus_{s\le t-1}\mathbb{C} e_s$ is an ideal of $\bigoplus_{s\le t}\mathbb{C} e_s$. Let $\alpha_s$ be the restricted root with respect to $e_s$. As in Section~\ref{sec:vanishing theorem}, for $\mathbf{k} = (k_1,\dots,k_l)\in\mathbb{Z}_{\ge 0}^l$ we denote $\ad(e_l)^{k_l}\dotsm \ad(e_1)^{k_1}$ by $\ad(e)^\mathbf{k}$ and $((-x_1)^{k_1}/k_1!)\dotsm((-x_l)^{k_l}/k_l!)$ by $f_\mathbf{k}$. \begin{lem}\label{lem:left2right for I'_i} We have \[ I'_i = \left\{\sum_{s = 1}^t \delta_i(T_s,f_s\eta_i^{-1},u'_s)\Bigm| \begin{array}{ll} T_s\in U(\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0),& f_s\in \mathcal{P}(O_i),\\ u'_s\in J'_{w_i^{-1}\eta}(\sigma\otimes(\lambda + \rho)) \end{array} \right\}. \] \end{lem} \begin{proof} By Lemma~\ref{lem:left2right}, we have \[ T((f\otimes u')\delta_i) = \sum_{\mathbf{k}\in\mathbb{Z}_{\ge 0}^l}\delta_i(\ad(e)^\mathbf{k}T,ff_\mathbf{k},u') \] for $T\in U(\mathfrak{g})$, $f\in \mathcal{P}(O_i)\eta_i^{-1}$ and $u'\in \sigma'$. Hence, the left hand side is a subset of the right hand side. Define $f_\mathbf{k}'\in\mathcal{P}(O_i)$ by $f_\mathbf{k}' = (x_1^{k_1}/k_1!)\dotsm(x_l^{k_l}/k_l!)$. By the similar calculation of Lemma~\ref{lem:left2right}, we have \[ \delta_i(T,f,u') = \sum_{\mathbf{k}\in\mathbb{Z}_{\ge 0}^l}(\ad(e)^{\mathbf{k}}T)(((ff_\mathbf{k}')\otimes u')\delta_i). \] This implies that the right hand side is contained in the left hand side. \end{proof} By the definition of the twisting functor and Poincar\"e-Birkhoff-Witt theorem, we have the following lemma. For $\mathbf{k} = (k_1,\dots,k_l)\in\mathbb{Z}^l$ put $(e - \eta(e))^\mathbf{k} = (e_1 - \eta(e_1))^{k_1}\dotsm (e_l - \eta(e_l))^{k_l}\in S_{w,\eta}$. Set $\mathbf{1} = (1,\dots,1)\in\mathbb{Z}^l$. \begin{lem}\label{lem:induction + twisting} Let $V$ be a $\mathfrak{p}$-module. Then we have a $\mathbb{C}$-vector space isomorphism \begin{multline*} \left(\sum_{\mathbf{k}\in\mathbb{Z}_{\ge 0}^l}\mathbb{C} (e - \eta(e))^{-(\mathbf{k} + \mathbf{1})}\right)\otimes_{U(\Ad(w_i)\overline{\mathfrak{n}}\cap\mathfrak{n}_0)}U(\mathfrak{g})\otimes_{U(\Ad(w_i)\mathfrak{p})}w_iV\\ \simeq T_{w_i}(U(\mathfrak{g})\otimes_{U(\mathfrak{p})}V) \end{multline*} given by $E\otimes T\otimes v\mapsto ET\otimes (1\otimes v)$. (Notice that $ET\in S_{w_i,0}$.) \end{lem} \begin{proof}[Proof of Theorem~\ref{thm:structure of I_i/I_{i - 1}}] By Lemma~\ref{lem:left2right for I'_i}, we have an isomorphism as a vector space, \[ I'_i \simeq \mathcal{P}(O_i) \otimes_{U(\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0)} U(\mathfrak{g}) \otimes_{U(\Ad(w_i)\mathfrak{p})} w_iJ'_{w_i^{-1}\eta}(\sigma\otimes(\lambda + \rho)) \] given by $\delta_i(T,f,u')\mapsto f\otimes T\otimes u'$. Notice that $\mathfrak{u}_{0,w_i} = \Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0$ since $w_i\in W(M)$. By Lemma~\ref{lem:induction + twisting}, we have \begin{multline*} T_{w_i,\eta}(J_i) \simeq \left(\sum_{\mathbf{k}\in\mathbb{Z}_{\ge 0}^l} \mathbb{C} (e - \eta(e))^{-(\mathbf{k}+\mathbf{1})}\right) \otimes_{U(\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0)} U(\mathfrak{g})\\ \otimes_{U(\Ad(w_i)\mathfrak{p})} w_iJ'_{w_i^{-1}\eta}(\sigma\otimes(\lambda + \rho)). \end{multline*} Here we remark $\sum_{\mathbf{k}\in\mathbb{Z}_{\ge 0}^l} \mathbb{C} (e - \eta(e))^{-(\mathbf{k}+\mathbf{1})}$ is an $\Ad(w_i)\overline{\mathfrak{n}}\cap\mathfrak{n}_0$-stable subspace of $S_{w_i,\eta}$. Hence, we can define a $\mathbb{C}$-vector space isomorphism $\Phi\colon T_{w_i,\eta}(J_i)\to I'_i$ by \[ \Phi((e - \eta(e))^{-(\mathbf{k}+\mathbf{1})}\otimes T\otimes u') = \delta_i(T,f_\mathbf{k}\eta_i^{-1},u'). \] We prove that $\Phi$ is a $\mathfrak{g}$-homomorphism. Fix $X\in \mathfrak{g}$. We prove that \[ \Phi(X((e - \eta(e))^{-(\mathbf{k}+\mathbf{1})}\otimes T\otimes u')) = X\Phi((e - \eta(e))^{-(\mathbf{k}+\mathbf{1})}\otimes T\otimes u'). \] By Lemma~\ref{lem:Xe^-k}, we have \begin{multline*} X((e - \eta(e))^{-(\mathbf{k}+\mathbf{1})}\otimes T\otimes u')\\ = \sum_{p_s\ge 0}\binom{p_1 + k_1}{k_1}\dotsm \binom{p_l + k_l}{k_l}(e - \eta(e))^{-(\mathbf{k}+\mathbf{p}+\mathbf{1})}\otimes(\ad(e)^\mathbf{p}X)T\otimes u'. \end{multline*} where $\mathbf{p} = (p_1,\dots,p_l)$. Hence, we have \begin{multline*} \Phi(X((e - \eta(e))^{-(\mathbf{k}+\mathbf{1})}\otimes T\otimes u'))\\ = \sum_{p_s\ge 0}\delta_i\left((\ad(e)^\mathbf{p}X)T,\left(\frac{(-x_1)^{k_1 + p_1}}{k_1!p_1!}\dotsm\frac{(-x_l)^{k_l + p_l}}{k_l!p_l!}\right)\eta_i^{-1}, u'\right). \end{multline*} By Lemma~\ref{lem:left2right}, we have \begin{align*} X\Phi((e - \eta(e))^{-(\mathbf{k} + 1)}\otimes T\otimes u') & = X\delta_i(T,f_\mathbf{k}\eta_i^{-1},u')\\ & = \sum_{\mathbf{p}\in\mathbb{Z}_{\ge 0}^l}\delta_i((\ad(e)^{\mathbf{p}}X)T,f_\mathbf{k}f_\mathbf{p}\eta_i^{-1},u'). \end{align*} Hence, we have the theorem. \end{proof} \section{The module $J^*_\eta(I(\sigma,\lambda))$}\label{sec:the module J^*_eta(I(sigma,lambda))} Now we investigate a module $J^*_\eta(I(\sigma,\lambda))$. For a finite-length Fr\'echet representation $V$ of $G$, put $J(V) = (\varprojlim_{k\to \infty}(V_{\text{$K$-finite}}/\mathfrak{n}_0^kV_{\text{$K$-finite}}))_{\text{$\mathfrak{a}$-finite}}$\newsym{$J(V)$}. This is also called the Jacquet module of V~\cite{MR562655}. Define a category $\mathcal{O}'_{P_0}$\newsym{$\mathcal{O}'_{P_0}$} by the full subcategory of finitely generated $\mathfrak{g}$-modules consisting an object $V$ satisfying the following conditions. \begin{enumerate} \item The action of $\mathfrak{p}_0$ is locally finite. (In particular, the action of $\mathfrak{n}_0$ is locally nilpotent.) \item The module $V$ is $Z(\mathfrak{g})$-finite. \item The group $M_0$ acts on $V$ and its differential coincides with the action of $\mathfrak{m}_0\subset \mathfrak{g}$. \item For $\nu\in\mathfrak{a}_0^*$ let $V_\nu$ be the generalized $\mathfrak{a}_0$-weight space with weight $\nu$. Then $V = \bigoplus_{\nu\in\mathfrak{a}_0^*}V_\nu$ and $\dim V_\nu < \infty$. \end{enumerate} We define the category $\mathcal{O}_{\overline{P_0}}'$ similarly.\newsym{$\mathcal{O}'_{\overline{P_0}}$} Then for a finite-length Fr\'echet representation $V$ of $G$ we have $J(V)\in\mathcal{O}_{\overline{P_0}}'$ and $J^*(V)\in\mathcal{O}_{P_0}'$. For a $U(\mathfrak{g})$-module $V$, put $D'(V) = (V^*)_{\text{$\mathfrak{h}$-finite}}$ and $C(V) = (D'(V))^*$.\newsym{$D'(V)$}\newsym{$C(V)$} Denote a full-subcategory of $\mathfrak{g}$-modules consisting finitely-generated and locally $\mathfrak{h}\oplus\mathfrak{u}$-finite modules by $\mathcal{O}'$\newsym{$\mathcal{O}'$}. If $V$ is an object of the category $\mathcal{O}'$ then $D'D'(V) \simeq V$. The relation between $J^*$ and $J$ is as follows. \begin{prop}\label{prop:relation J^* and J} Let $V$ be a finite-length Fr\'echet representation of $G$. Then we have $J^*(V) \simeq D'(J(V))$. \end{prop} The character $\eta\colon \mathfrak{n}_0\to \mathbb{C}$ defines an algebra homomorphism $U(\mathfrak{n}_0)\to \mathbb{C}$ by the universality of the universal enveloping algebra. Let $\Ker\eta$ be the kernel of this algebra homomorphism and put $\Gamma_\eta(V) = \{v\in V\mid \text{for some $k$, $(\Ker\eta)^kv = 0$}\}$.\newsym{$\Gamma_\eta(V)$} First we prove the following proposition. \begin{prop}\label{prop:relation J^*_eta and J_eta} Let $V$ be a finite-length Fr\'echet representation of $G$. Then we have $J^*_\eta(V) \simeq \Gamma_\eta(J(V)^*)$. \end{prop} \begin{proof} Recall that $\mathfrak{p}_\eta = \mathfrak{m}_\eta\oplus\mathfrak{a}_\eta\oplus\mathfrak{n}_\eta$ is the complexification of the Lie algebra of the parabolic subgroup corresponding to $\supp\eta$ (Section~\ref{sec:Analytic continuation}). If $\supp\eta = \Pi$, this proposition is proved by Matumoto~\cite[Theorem~4.9.2]{MR1047117}. Put $I = V_{\text{$K$-finite}}$. Let $\eta_0\colon U(\mathfrak{m}\cap \mathfrak{n}_0)\to \mathbb{C}$ be the restriction of $\eta$ on $U(\mathfrak{m}\cap \mathfrak{n}_0)$. Then we have \[ J^*_\eta(V) = \varinjlim_{k,l}(I/\mathfrak{n}_\eta^l(\Ker\eta_0)^kI)^* = \varinjlim_{k,l}((I/\mathfrak{n}_\eta^lI)/(\Ker\eta_0)^k(I/\mathfrak{n}_\eta^lI))^*. \] For a $U(\mathfrak{g})$-module $V_0$, put $G(V_0) = (\varprojlim_k V_0/\mathfrak{n}_0^kV_0)_{\text{$\mathfrak{a}$-finite}}$. For a $U(\mathfrak{m}_\eta\oplus\mathfrak{a}_\eta)$-module $V_1$, put $G_{M_\eta}(V_1) = (\varprojlim_k V_1/(\mathfrak{m}_\eta\cap\mathfrak{n}_0)^kV_1)_{\textrm{$(\mathfrak{m}\cap\mathfrak{a}_0)$-finite}}$. Since $I/\mathfrak{n}_\eta^lI$ is a Harish-Chandra module of $\mathfrak{m}_\eta\oplus\mathfrak{a}_\eta$, $J^*_{\eta_0}(I/\mathfrak{n}_\eta^lI) = \Gamma_{\eta_0}(G_{M_\eta}(I/\mathfrak{n}_\eta^lI)^*)$ by the result of Matumoto. Taking a subspace annihilated by $(\Ker\eta_0)^k$, we have \[ ((I/\mathfrak{n}_\eta^lI)/(\Ker\eta_0)^k(I/\mathfrak{n}_\eta^lI))^* = (G_{M_\eta}(I/\mathfrak{n}_\eta^lI)/(\Ker\eta_0)^kG_{M_\eta}(I/\mathfrak{n}_\eta^lI))^*. \] Since $I$ is a finitely-generated $U(\mathfrak{n}_0)$-module, the left hand side is finite-dimensional. Hence, we have \[ (I/\mathfrak{n}_\eta^lI)/(\Ker\eta_0)^k(I/\mathfrak{n}_\eta^lI) = G_{M_\eta}(I/\mathfrak{n}_\eta^lI)/(\Ker\eta_0)^kG_{M_\eta}(I/\mathfrak{n}_\eta^lI). \] It is sufficient to prove that $G_{M_\eta}(I/\mathfrak{n}_\eta^lI) = G(I)/\mathfrak{n}_\eta^lG(I)$. We have \[ (I/\mathfrak{n}_\eta^lI)/(\mathfrak{m}_\eta\cap\mathfrak{n}_0)^k(I/\mathfrak{n}_\eta^lI) = I/(\mathfrak{m}_\eta\cap\mathfrak{n}_0)^k\mathfrak{n}_\eta^lI = G(I)/(\mathfrak{m}_\eta\cap\mathfrak{n}_0)^k\mathfrak{n}_\eta^lG(I). \] Taking the projective limit we have $G_{M_\eta}(I/\mathfrak{n}_\eta^lI) = G_{M_\eta}(G(I)/\mathfrak{n}_\eta^lG(I))$. Since $G(I)/\mathfrak{n}_\eta^lG(I)\in \mathcal{O}'_{M_\eta\cap\overline{P_0}}$ we have $G_{M_\eta}(G(I)/\mathfrak{n}_\eta^lG(I)) = G(I)/\mathfrak{n}_\eta^lG(I)$. \end{proof} Combining Theorem~\ref{thm:structure of I_i/I_{i - 1}}, Proposition~\ref{prop:relation J^*_eta and J_eta} and the automatic continuation theorem~\cite[Theorem~4.8]{MR727854}, we have the following theorem. \begin{thm}\label{thm:stucture of J^*(I(sigma,lambda))} There exists a filtration $0 = \widetilde{I_1}\subset\cdots \subset \widetilde{I_r} = J^*_\eta(I(\sigma,\lambda))$ such that $\widetilde{I_i}/\widetilde{I_{i - 1}} \simeq \Gamma_\eta(C(T_{w_i}(U(\mathfrak{g})\otimes_{U(\mathfrak{p})} J^*(\sigma\otimes e^{\lambda+\rho}))))$.\newsym{$\widetilde{I_i}$} \end{thm} \section{Whittaker vectors}\label{sec:Whittaker vectors} In this section we study Whittaker vectors of $I(\sigma,\lambda)'$ and $(I(\sigma,\lambda)_{\text{$K$-finite}})^*$ (Definition~\ref{defn:Whittaker vectors}). For $i$ such that $I_i/I_{i - 1}\ne 0$, we define some maps as follows. Let $\gamma_1$ be the first projection with respect to the decomposition $U(\mathfrak{g}) = U(\mathfrak{l}_\eta)\oplus(\overline{\mathfrak{n}_\eta}U(\mathfrak{g}) + U(\mathfrak{g})\mathfrak{n}_\eta)$. Notice that by Lemma~\ref{lem:vanishing lemma} if $I_i/I_{i - 1}\ne 0$ then we have $\mathfrak{l}_\eta\cap \Ad(w_i)\overline{\mathfrak{n}}\subset \mathfrak{n}_0$. Define $\gamma_2$ by the first projection with respect to the decomposition $U(\mathfrak{l}_\eta) = U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{p}) \oplus U(\mathfrak{l}_\eta)\Ker\eta|_{\mathfrak{l}_\eta\cap \Ad(w_i)\overline{\mathfrak{n}}}$. Let $\gamma_3$ be the first projection with respect to the decomposition $U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{p}) = U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{l})\oplus (\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{n})U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{p})$. Finally define $\gamma_4$ by the first projection with respect to the decomposition $U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{l}) = U(\mathfrak{h})\oplus((\overline{\mathfrak{u}_0}\cap\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{l})U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{l}) + U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{l})(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{l} \cap \mathfrak{u}_0))$. Then the restriction of $\gamma_4\circ\gamma_3\circ\gamma_2\circ\gamma_1$ on $Z(\mathfrak{g})$ is the (non-shifted) Harish-Chandra homomorphism. If $x\in \Wh_\eta(I_i/I_{i - 1})$ then $Tx = \gamma_2\gamma_1(T)x$ for $T\in Z(\mathfrak{g})$. \begin{align*} \gamma_1\colon & U(\mathfrak{g}) = U(\mathfrak{l}_\eta)\oplus(\overline{\mathfrak{n}_\eta}U(\mathfrak{g}) + U(\mathfrak{g})\mathfrak{n}_\eta)\to U(\mathfrak{l}_\eta),\\ \gamma_2\colon & U(\mathfrak{l}_\eta) = U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{p}) \oplus U(\mathfrak{l}_\eta)\Ker\eta|_{\mathfrak{l}_\eta\cap \Ad(w_i)\overline{\mathfrak{n}}}\to U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{p}),\\ \gamma_3\colon & U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{p}) = U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{l})\oplus (\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{n})U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{p}) \tag*{$\to U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{l})$},\\ \gamma_4\colon & U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{l}) = U(\mathfrak{h})\oplus((\overline{\mathfrak{u}_0}\cap\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{l})U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{l}) \tag*{$+ U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{l})(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{l} \cap \mathfrak{u}_0))\to U(\mathfrak{h}).$} \end{align*} \newsym{$\gamma_1,\gamma_2,\gamma_3,\gamma_4$} \begin{lem}\label{lem:lemma of infinitesimal character} Let $V$ be a $U(\mathfrak{g})$-module with an infinitesimal character $\widetilde{\lambda}$, $\chi$ a character of $Z(\mathfrak{g})$ such that $z\in Z(\mathfrak{g})$ acts by $\chi(z)$ on $V$. Take a nonzero element $v\in V$ such that $(\gamma_3\gamma_2\gamma_1(z) - \chi(z))v = 0$. Moreover, assume that there exists $\mu\in\mathfrak{a}^*$ such that $Hv = (w_i\mu+\rho_0)(H)v$ for all $H\in \Ad(w_i)\mathfrak{a}$. Then there exists $\widetilde{w}\in \widetilde{W}$ such that $\widetilde{w}\widetilde{\lambda}|_\mathfrak{a} = \mu$. \end{lem} \begin{proof} Put $Z = (\gamma_3\gamma_2\gamma_1(Z(\mathfrak{g}))U(\Ad(w_i)\mathfrak{a}))$. By the assumption, there exists a character $\chi_0$ of $Z$ such that $zv = \chi_0(z)v$ for all $z\in Z$. By a theorem of Harsh-Chandra, $\gamma_4|_Z$ is injective and $\gamma_4(Z)\subset U(\mathfrak{h})$ is finite. Hence there exists an element $\widetilde{\lambda_1}\in \mathfrak{h}^*$ such that $\widetilde{\lambda_1}\circ \gamma_4 = \chi_0$ where we denote the algebra homomorphism $U(\mathfrak{h})\to \mathbb{C}$ induced from $\widetilde{\lambda_1}$ by the same letter $\widetilde{\lambda_1}$. Since $V$ has an infinitesimal character $\widetilde{\lambda}$, we have $\widetilde{\lambda_1}\in \widetilde{W}\widetilde{\lambda} + \widetilde{\rho}$. Since $\gamma_4$ is trivial on $U(\Ad(w_i)\mathfrak{a})$, $\widetilde{\lambda_1}|_{\Ad(w_i)\mathfrak{a}} = (w_i\mu + \rho_0)|_{\Ad(w_i)\mathfrak{a}}$. The restriction of $\widetilde{\rho}$ to $\mathfrak{a}_0$ is $\rho_0$. Hence $\widetilde{\rho}|_{\Ad(w_i)\mathfrak{a}} = \rho_0|_{\Ad(w_i)\mathfrak{a}}$. Then for some $\widetilde{w}\in\widetilde{W}$ we have $w_i\mu|_{\Ad(w_i)\mathfrak{a}} = \widetilde{w}\widetilde{\lambda}|_{\Ad(w_i)\mathfrak{a}}$. We get the lemma. \end{proof} \begin{lem}\label{lem:to outside delta_i} Let $X_1,\dots,X_n\in \mathfrak{g}$, $f_1\in C^\infty(O_i)$, $f_2\in C^\infty(U_i)$, $u'\in (\sigma\otimes e^{\lambda + \rho})'$. Assume that $\widetilde{R'_i}(X_s)(f_2) = 0$ for all $s = 1,\dots,n$. Then we have \[ \delta_i(X_1\dotsm X_n,f_1f_2,u') = \delta_i(X_1\dotsm X_n,f_1,u')f_2. \] \end{lem} \begin{proof} Put $T = X_1\dotsm X_n$. By the assumption and Leibniz's rule, we have \[ f_2(nw_i)(\widetilde{R_i}(T)\varphi)(nw_i) = (\widetilde{R_i}(T)(\varphi f_2))(nw_i) \] Hence, by the definition, for $\varphi\in C^\infty_c(U_i,\mathcal{L})$, we have \begin{align*} & \langle \delta_i(T,f_1f_2,u'),\varphi\rangle\\ & = \int_{w_i\overline{N}w_i^{-1}\cap N_0}f_1(nw_i)f_2(nw_i)(u'(\widetilde{R}_i(T)\varphi)(nw_i))dn\\ & = \int_{w_i\overline{N}w_i^{-1}\cap N_0}f_1(nw_i)(u'(\widetilde{R}_i(T)(\varphi f_2))(nw_i))dn\\ & = \langle\delta_i(T,f_1,u'),f_2\varphi\rangle\\ & = \langle\delta_i(T,f_1,u')f_2,\varphi\rangle. \end{align*} We get the lemma. \end{proof} \begin{lem}\label{lem:property of V(nu)} For $\nu\in \mathfrak{a}^*$ put \[ V(\nu) = \left\{\sum_s \delta_i(S_s,h_s\eta_i^{-1},v_s')\Biggm| \begin{array}{l} S_s\in U(\Ad(w_i)\overline{\mathfrak{n}}\cap \overline{\mathfrak{n}_0}),\ h_s\in\mathcal{P}(O_i),\\ v'_s\in J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho}),\\ w_i^{-1}(\wt h_s + \wt S_s)|_\mathfrak{a} = \nu \end{array} \right\}. \] Here, $\wt h_s$ is an $\mathfrak{a}_0$-weight of $h_s$ with respect to $D_i$ (see page~\pageref{symbol:D_i}) and $\wt S_s$ is an $\mathfrak{a}_0$-weight of $S_s$ with respect to the adjoint action. Define $\widetilde{\eta_i}\in C^\infty(U_i)$ by $\widetilde{\eta_i}(nn_0w_iP/P) = \eta_i(n)$ for $n\in w_i\overline{N}w_i^{-1}\cap N_0$ and $n_0\in w_i\overline{N}w_i^{-1}\cap \overline{N_0}$. \begin{enumerate} \item Let $X\in U(\mathfrak{l}_\eta\cap \Ad(w_i)\mathfrak{p})$. Assume that $X$ is an $\mathfrak{a}_0$-weight vector. For $\delta_i(T,f\eta_i^{-1},u')\in V(\nu)$, we have \[ X\delta_i(T,f\eta_i^{-1},u') - (X\delta_i(T,f\eta_i^{-1},u'))\widetilde{\eta_i}^{-1}\in\sum_{\nu' > \nu}V(\nu' + w_i^{-1}\wt T|_\mathfrak{a}). \] here, $\wt T$ is an $\mathfrak{a}_0$-weight of $T$ with respect to the adjoint action. \item For $\delta_i(S_s,h_s\eta_i^{-1},v'_s)\in V(\nu)$, we have \[ \sum_s \delta_i(S_s,h_s,v'_s)\widetilde{\eta_i}^{-1}\not\in \sum_{\nu' > \nu}V(\nu'). \] \end{enumerate} \end{lem} \begin{proof} (1) Fix a basis $\{e_1,e_2,\dots,e_l\}$ of $\mathfrak{u}_{0,w_i}$ such that each vector $e_i$ is a root vector and $\bigoplus_{s\le t-1}\mathbb{C} e_s$ is an ideal of $\bigoplus_{s\le t}\mathbb{C} e_s$. Let $\alpha_s$ be the restricted root of $e_s$. As in Section~\ref{sec:vanishing theorem}, for $\mathbf{k} = (k_1,\dots,k_l)\in\mathbb{Z}_{\ge 0}^l$ we denote $\ad(e_l)^{k_l}\dotsm \ad(e_1)^{k_1}$ by $\ad(e)^\mathbf{k}$ and $((-x_1)^{k_1}/k_1!)\dotsm((-x_l)^{k_l}/k_l!)$ by $f_\mathbf{k}$. By Lemma~\ref{lem:left2right}, \[ X\delta_i(T,f\eta_i^{-1},u') = \sum_{\mathbf{k}\in\mathbb{Z}_{\ge 0}^l}\delta_i((\ad(e)^\mathbf{k}X)T,ff_\mathbf{k}\eta_i^{-1},u'). \] Take $a_\mathbf{k}^{(p)}\in U(\Ad(w_i)\overline{\mathfrak{n}}\cap\mathfrak{n}_0)$, $b_\mathbf{k}^{(p)}\in U(\Ad(w_i)\overline{\mathfrak{n}}\cap \overline{\mathfrak{n}_0})$ and $c_\mathbf{k}^{(p)}\in U(\Ad(w_i)\mathfrak{p})$ such that $(\ad(e)^\mathbf{k}X)T = \sum_p a_\mathbf{k}^{(p)}b_\mathbf{k}^{(p)}c_\mathbf{k}^{(p)}$ and $\wt((\ad(e)^\mathbf{k}X)T) = \wt a_\mathbf{k}^{(p)} + \wt b_\mathbf{k}^{(p)} + \wt c_\mathbf{k}^{(p)}$. Then \begin{align*} &\delta_i((\ad(e)^\mathbf{k}X)T,ff_\mathbf{k}\eta_i^{-1},u')\\ & = \sum_p \delta_i(a_\mathbf{k}^{(p)}b_\mathbf{k}^{(p)}c_\mathbf{k}^{(p)},ff_\mathbf{k}\eta_i^{-1},u')\\ & = \sum_p \delta_i(b_\mathbf{k}^{(p)},R'_i(-a_\mathbf{k}^{(p)})(ff_\mathbf{k}\eta_i^{-1}),\Ad(w_i)^{-1}(c_\mathbf{k}^{(p)})u') \end{align*} By the Leibniz rule, there exists a subset $\mathcal{A}^{(p)}_\mathbf{k} \subset \{(a',a'')\in U(\Ad(w_i)\overline{\mathfrak{n}}\cap\mathfrak{n}_0)^2\mid \wt a' + \wt a'' = \wt a_\mathbf{k}^{(p)},\ a''\not\in\mathbb{C}\}$ such that \begin{align*} & \delta_i(b_\mathbf{k}^{(p)},R'_i(-a_\mathbf{k}^{(p)})(ff_\mathbf{k}\eta_i^{-1}) - R'_i(-a_\mathbf{k}^{(p)})(ff_\mathbf{k})\eta_i^{-1},\Ad(w_i)^{-1}(c_\mathbf{k}^{(p)})u')\\ & = \sum_{(a',a'')\in\mathcal{A}^{(p)}_\mathbf{k}}\delta_i(b_\mathbf{k}^{(p)},R'_i(a')(ff_\mathbf{k})R'_i(a'')(\eta_i^{-1}),\Ad(w_i)^{-1}c_\mathbf{k}^{(p)}u')\\ & = \sum_{(a',a'')\in\mathcal{A}^{(p)}_\mathbf{k}}-\eta(a'')\delta_i(b_\mathbf{k}^{(p)},R'_i(a')(ff_\mathbf{k})\eta_i^{-1},\Ad(w_i)^{-1}c_\mathbf{k}^{(p)}u') \end{align*} By the Poincar\'e-Birkhoff-Witt theorem, we have a direct decomposition $U(\Ad(w_i)\mathfrak{p}) = U(\Ad(w_i)\mathfrak{p})(\Ad(w_i)\mathfrak{n})\oplus U(\Ad(w_i)\mathfrak{l})$. Hence we may assume that $c_\mathbf{k}^{(p)} \in U(\Ad(w_i)\mathfrak{p})(\Ad(w_i)\mathfrak{n})$ or $c_\mathbf{k}^{(p)} \in U(\Ad(w_i)\mathfrak{l})$. If $c_\mathbf{k}^{(p)} \in U(\Ad(w_i)\mathfrak{p})(\Ad(w_i)\mathfrak{n})$ then this sum is equal to $0$. If $c_\mathbf{k}^{(p)} \in U(\Ad(w_i)\mathfrak{l})$ then $w_i^{-1}\wt c_\mathbf{k}^{(p)}|_{\mathfrak{a}} = 0$. Hence, \begin{align*} &w_i^{-1}(\wt b_\mathbf{k}^{(p)} + \wt(R'_i(a')ff_\mathbf{k}))|_{\mathfrak{a}}\\ & = w_i^{-1}(\wt c_\mathbf{k}^{(p)} + \wt b_\mathbf{k}^{(p)} + \wt a' + \wt f + \wt f_\mathbf{k})|_{\mathfrak{a}}\\ & = w_i^{-1}(\wt ((\ad(e)^{\mathbf{k}}X)T) + \wt f + \wt f_\mathbf{k} - \wt a'')|_{\mathfrak{a}}\\ & = w_i^{-1}(\wt X + \wt T + \wt f - \wt a'')|_{\mathfrak{a}}\\ & = \nu + w_i^{-1}(\wt X - \wt a'')|_{\mathfrak{a}} > \nu + w_i^{-1}\wt X|_{\mathfrak{a}}. \end{align*} So we have \begin{multline*} \delta_i(b_\mathbf{k}^{(p)},R'_i(-a_\mathbf{k}^{(p)})(ff_\mathbf{k}\eta_i^{-1}) - R'_i(-a_\mathbf{k}^{(p)})(ff_\mathbf{k})\eta_i^{-1},\Ad(w_i)^{-1}(c_\mathbf{k}^{(p)})u')\\ \in \sum_{\nu' > \nu}V(\nu' + w_i^{-1}\wt X|_\mathfrak{a}). \end{multline*} By the definition of $\widetilde{\eta_i}$, we have $\widetilde{R_i}(X')\widetilde{\eta_i} = 0$ for $X'\in \Ad(w_i)\overline{\mathfrak{n}}\cap \overline{\mathfrak{n}_0}$. Hence by Lemma~\ref{lem:to outside delta_i}, we have \begin{multline*} \delta_i(b_\mathbf{k}^{(p)},R'_i(-a_\mathbf{k}^{(p)})(ff_\mathbf{k})\eta_i^{-1},\Ad(w_i)^{-1}(c_\mathbf{k}^{(p)})u')\\ = \delta_i(b_\mathbf{k}^{(p)},R'_i(-a_\mathbf{k}^{(p)})(ff_\mathbf{k}),\Ad(w_i)^{-1}(c_\mathbf{k}^{(p)})u')\widetilde{\eta_i}^{-1} \end{multline*} Hence, we have \begin{align*} &\sum_{\mathbf{k},p}\delta_i(b_\mathbf{k}^{(p)},R'_i(-a_\mathbf{k}^{(p)})(ff_\mathbf{k})\eta_i^{-1},\Ad(w_i)^{-1}(c_\mathbf{k}^{(p)})u')\\ & = \sum_{\mathbf{k},p}\delta_i(b_\mathbf{k}^{(p)},R'_i(-a_\mathbf{k}^{(p)})(ff_\mathbf{k}),\Ad(w_i)^{-1}(c_\mathbf{k}^{(p)})u')\widetilde{\eta_i}^{-1}\\ & = \sum_{\mathbf{k},p}\delta_i(a_\mathbf{k}^{(p)}b_\mathbf{k}^{(p)}c_\mathbf{k}^{(p)},(ff_\mathbf{k}),u')\widetilde{\eta_i}^{-1}\\ & = (X\delta_i(T,f,u'))\widetilde{\eta_i}^{-1}. \end{align*} We get (1). (2) Take $T = 1$ in (1). Then we have \[ \delta_i(T,f\eta_i^{-1},u') - (\delta_i(T,f\eta_i^{-1},u'))\widetilde{\eta_i}^{-1}\in\sum_{\nu' > \nu}V(\nu'). \] Hence if \[ \sum_s \delta_i(S_s,h_s,v'_s)\widetilde{\eta_i}^{-1}\in \sum_{\nu' > \nu}V(\nu') \] then \[ \delta_i(T,f\eta_i^{-1},u')\in \sum_{\nu' > \nu}V(\nu') \] However, by Lemma~\ref{lem:fundamental properties of delta_i} (3), we have $V(\nu)\cap \sum_{\nu'\ne \nu}V(\nu') = 0$. This is a contradiction. \end{proof} \begin{prop}\label{prop:Whittaker vectors in a Bruaht cell} Let $\widetilde{\mu}\in (\mathfrak{h}\cap \mathfrak{m})^*$ be an infinitesimal character of $\sigma$. Assume that $I_i/I_{i - 1}\ne 0$ and for all $\widetilde{w}\in \widetilde{W}$, \[ \lambda - \widetilde{w}(\lambda + \widetilde{\mu})|_\mathfrak{a}\not\in \mathbb{Z}_{\le 0}((\Sigma^+\setminus\Sigma_M^+)\cap w_i^{-1}\Sigma^+)|_\mathfrak{a}\setminus\{0\}. \] Then \[ \Wh_\eta(I_i') = \{(\eta_i^{-1}\otimes u')\delta_i\mid u'\in \Wh_{w_i^{-1}\eta}((\sigma\otimes e^{\lambda+\rho})')\}. \] \end{prop} \begin{proof} Let $x = \sum_s\delta_i(T_s,f_s\eta_i^{-1},u_s')$ be an element of $\Wh_\eta(I_i')$ where $T_s\in U(\Ad(w_i)\overline{\mathfrak{n}}\cap \overline{\mathfrak{n}_0})$, $f_s\in\mathcal{P}(O_i)$ and $u'_s\in J'_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$. For $X\in \Ad(w_i)\overline{\mathfrak{n}}\cap\mathfrak{n}_0$, we have $(X - \eta(X))x = \sum_s \delta_i(T_s,(L_X - \eta(X))(f_s\eta_i^{-1}),u'_s) = \sum_s \delta_i(T_s,L_X(f_s)\eta_i^{-1},u'_s)$ by Lemma~\ref{lem:caluculation of Xdelta(1,f,u)}. Hence, we may assume $f_s = 1$. Let $z\in Z(\mathfrak{g})$. Since $J'_\eta(I(\sigma,\lambda))$ has an infinitesimal character $-(\lambda+\widetilde{\mu})$, $I_i'$ has the same character. Let $\chi(z)$ be a complex number such that $z$ acts by $\chi(z)$ on $I'_i$. Take $T_s$ and $u'_s$ such that $T_s$ are $\mathfrak{a}_0$-weight vectors and lineally independent. Let $\nu = \min\{w_i^{-1}\wt T_s|_{\mathfrak{a}}\}_s$. Then by Lemma~\ref{lem:property of V(nu)} (1), we have \begin{multline*} \chi(z)x = zx = \gamma_2\gamma_1(z)x \\\in \left(\gamma_3\gamma_2\gamma_1(z)\sum_{w_o^{-1}\wt T_s|_{\mathfrak{a}} = \nu}\delta_i(T_s,1,u_s')\right)\widetilde{\eta_i}^{-1} + \sum_{\nu' > \nu}V(\nu'). \end{multline*} By Lemma~\ref{lem:property of V(nu)} (1) ($T = 1$), we have \[ x \in \sum_{w_i^{-1}\wt T_s|_{\mathfrak{a}} = \nu}\delta_i(T_s,1,u'_s)\widetilde{\eta_i}^{-1} + \sum_{\nu' > \nu}V(\nu'). \] Hence we have \[ \left((\chi(z) - \gamma_3\gamma_2\gamma_1(z))\left(\sum_{w_i^{-1}\wt T_s|_{\mathfrak{a}} = \nu}\delta_i(T_s,1,u'_s)\right)\right)\widetilde{\eta_i}^{-1}\in \sum_{\nu' > \nu}V(\nu'). \] By Lemma~\ref{lem:property of V(nu)} (2), we have $(\chi(z) - \gamma_3\gamma_2\gamma_1(z))\delta_i(T_s,1,u_s') = 0$ for all $s$ such that $w_i^{-1}\wt T_s|_{\mathfrak{a}} = \nu$. By the same calculation as that of the proof of Lemma~\ref{lem:no delta part}, $H\delta_i(T_s,1_u,u_s') = (-w_i\lambda + \wt T_s + \rho_0)(H)\delta_i(T_s,1_u,u_s')$ for $H\in \Ad(w_i)\mathfrak{a}$. By Lemma~\ref{lem:lemma of infinitesimal character}, there exists a $\widetilde{w}\in\widetilde{W}$ such that $-\widetilde{w}(\lambda+\widetilde{\mu})|_{\Ad(w_i)\mathfrak{a}} = -w_i\lambda + \wt T_s$. Then $\lambda - w_i^{-1}\widetilde{w}(\lambda+\widetilde{\mu})|_\mathfrak{a} = w_i^{-1}\wt T_s|_\mathfrak{a}\in\mathbb{Z}_{\le 0}((\Sigma^+\setminus\Sigma_M^+)\cap w_i^{-1}\Sigma^+)|_\mathfrak{a}$. By the assumption, $\wt T_s = 0$, i.e., $T_s\in \mathbb{C}$. Hence, we may assume that $x$ has a form $x = \delta_i(1,\eta_i^{-1},u') + \sum_{s\ge 2} \delta_i(T_s,\eta_i^{-1},u_s')$ where $\wt T_s \ne 0$ for all $s\ge 2$. Take $X\in \mathfrak{n}_0\cap \Ad(w_i)\mathfrak{m}$. Then by Lemma~\ref{lem:caluculation of Xdelta(1,f,u)} and the above claim, \[ 0 = (X - \eta(X))x \in \delta_i(1,\eta_i^{-1},(\Ad(w_i)^{-1}X - \eta(X))u') + \sum_{\nu' > 0}V(\nu'). \] By Lemma~\ref{lem:property of V(nu)}, we have $\delta_i(1,\eta_i^{-1},(\Ad(w_i)^{-1}X - \eta(X))u') = 0$. Hence we have $u'\in \Wh_{w_i^{-1}\eta}((\sigma\otimes e^{\lambda + \rho})')$. This implies that $x - \delta_i(1,\eta_i^{-1},u')\in \Wh_\eta(I_i')$. If $x - \delta_i(1,\eta_i^{-1},u')\ne 0$, then by the above argument, we have $\min\{w_i^{-1}\wt T_s|_\mathfrak{a}\}_{s\ge 2} = 0$. This is a contradiction. \end{proof} \begin{thm}\label{thm:dimension Whittaker vectors} Assume that for all $w\in W(M)$ such that $\eta|_{wNw^{-1}\cap N_0} = 1$ the following two conditions hold: \begin{enumerate} \renewcommand*{\labelenumi}{(\alph{enumi})} \item For each exponent $\nu$ of $\sigma$ and $\alpha\in \Sigma^+\setminus w^{-1}(\Sigma^+\cup\Sigma_\eta^-)$, we have $2\langle\alpha,\lambda+\nu\rangle/\lvert\alpha\rvert^2\not\in\mathbb{Z}_{\le 0}$. \item For all $\widetilde{w}\in\widetilde{W}$ we have $\lambda - \widetilde{w}(\lambda + \widetilde{\mu})|_\mathfrak{a}\notin \mathbb{Z}_{\le 0}((\Sigma^+\setminus \Sigma_M^+)\cap w^{-1}\Sigma^+)|_\mathfrak{a}\setminus\{0\}$ where $\widetilde{\mu}$ is an infinitesimal character of $\sigma$. \end{enumerate} Moreover, assume that $\eta$ is unitary. Then we have \[ \dim\Wh_\eta(I(\sigma,\lambda)') = \sum_{w\in W(M),\ w(\Sigma^+\setminus\Sigma^+_M)\cap \supp\eta = \emptyset}\dim \Wh_{w^{-1}\eta}((\sigma\otimes e^{\lambda+\rho})'). \] \end{thm} \begin{proof} By the exact sequence $0\to I_{i - 1} \to I_i\to I_i/I_{i - 1}\to 0$, we have $0\to \Wh_\eta(I_{i - 1}) \to \Wh_\eta(I_i)\to \Wh_\eta(I_i/I_{i - 1})$. By Lemma~\ref{prop:Whittaker vectors in a Bruaht cell}, it is sufficient to prove that the last map $\Wh_\eta(I_i)\to \Wh_\eta(I_i/I_{i - 1})$ is surjective. Take $x\in \Wh_\eta(I'_i)\simeq \Wh_\eta(I_i/I_{i - 1})$. Then $x$ is $(\eta_i\otimes u')\delta_i$ for some $u'\in \Wh_{w_i^{-1}\eta}(\sigma\otimes e^{\lambda+\rho})$. By Lemma~\ref{lem:meromorphic extension}, there exists a distribution $x_t\in I_i(\lambda + t\rho)$ with meromorphic parameter $t$ such that $x_t|_{U_i}$ is holomorphic and $(x_t|_{U_i})|_{t = 0} = x$. Moreover, $(X - \eta(X))x_t = 0$ for $X\in\mathfrak{n}_0$. By Proposition~\ref{prop:convergence and continuation} and the condition (a), the distribution $x_t$ is holomorphic at $t = 0$. Hence $x_0|_{U_i} = x$. The map $\Wh_\eta(I_i)\to \Wh_\eta(I_i/I_{i - 1})$ is surjective. \end{proof} Next we consider the module $\Wh_\eta((I(\sigma,\lambda)_{\text{$K$-finite}})^*)$. Take a filtration $\widetilde{I_i}\subset J^*_\eta(I(\sigma,\lambda))$ as in Theorem~\ref{thm:stucture of J^*(I(sigma,lambda))}. \begin{lem}\label{lem:two step calc for Whittaker vector} Let $V$ be an object of the category $\mathcal{O}'$. Then we have $C(H^0(\mathfrak{n}_\eta,V)) = H^0(\mathfrak{n}_\eta,C(V))$ where $H^0(\mathfrak{n}_\eta,V) = \{v\in V\mid \mathfrak{n}_\eta v = 0\}$ is the $0$-th $\mathfrak{n}_\eta$-cohomology. \end{lem} \begin{proof} We get the lemma by the following equation. \begin{multline*} H^0(\mathfrak{n}_\eta,C(V)) = H^0(\mathfrak{n}_\eta,D'(V)^*) = (D'(V)/\mathfrak{n}_\eta D'(V))^*\\ = CD'(D'(V)/\mathfrak{n}_\eta D'(V)) = C(H^0(\mathfrak{n}_\eta,D'(V)^*)_{\text{$\mathfrak{h}$-finite}})\\ = C(H^0(\mathfrak{n}_\eta,D'D'(V))) = C(H^0(\mathfrak{n}_\eta,V)). \end{multline*} \end{proof} \begin{lem}\label{lem:relation in S_w} Let $e_1,\dots, e_l$ be a basis of $\Ad(w_i)\overline{\mathfrak{n}}\cap\mathfrak{n}_0$ such that each $e_s$ is a root vector and $\bigoplus_{s\le t - 1}\mathbb{C} e_s$ is an ideal of $\bigoplus_{s\le t}\mathbb{C} e_s$. In $S_{w_i,0}$ where $0$ is the trivial representation, we have the following formulae. \begin{enumerate} \item For all $t = 1,\dots,l$, \begin{multline*} e_t(e_1^{-1}\dotsm e_{t - 1}^{-1}e_t^{-(k_t + 1)}\dotsm e_l^{-(k_l + 1)})\\ = e_1^{-1}\dotsm e_{t - 1}^{-1}e_t^{-k_t}e_{t + 1}^{-(k_{t + 1} + 1)}\dotsm e_l^{-(k_l + 1)} \end{multline*} \item Fix $t \in \{1,\dots,l\}$ such that $e_t\in \mathfrak{n}_\eta$. Assume that $k_s = 0$ for all $s < t$ such that $e_s\in \mathfrak{n}_\eta$. Then \begin{multline*} e_t(e_1^{-(k_1 + 1)}\dotsm e_l^{-(k_l + 1)})\\ = e_1^{-(k_1 + 1)}\dots e_{t - 1}^{-(k_{t - 1} + 1)}e_t^{-k_t}e_{t + 1}^{-(k_t + 1)}\dots e_l^{-(k_l + 1)}. \end{multline*} \item $X(e_1^{-1}\dotsm e_l^{-1}) = (e_1^{-1}\dotsm e_l^{-1})X$ for $X\in \Ad(w_i)\mathfrak{m}\cap \mathfrak{n}_0$. \end{enumerate} \end{lem} \begin{proof} Let $\alpha_s$ be a restricted root corresponding to $e_s$. (1) It is sufficient to prove $e_t(e_1^{-1}\dotsm e_{t - 1}^{-1}) = (e_1^{-1}\dotsm e_{t - 1}^{-1})e_t$ in $S_{e_1}\otimes_{U(\mathfrak{g})}\dots \otimes_{U(\mathfrak{g})}S_{e_{t - 1}}$. Since $\bigoplus_{s = 1}^{t - 1}\mathbb{C} e_s$ is an ideal of $\bigoplus_{s = 1}^t \mathbb{C} e_s$, we have \[ e_t(e_1^{-1}\dotsm e_{t - 1}^{-1}) - (e_1^{-1}\dotsm e_{t - 1}^{-1})e_t\in \bigoplus_{k_s\ge 0}\mathbb{C} e_1^{-(k_1 + 1)}\dotsm e_{t - 1}^{-(k_{t - 1} + 1)}. \] An $\mathfrak{a}_0$-weight of the left hand side is $-\alpha_1 - \dots - \alpha_{t - 1} + \alpha_t$. However, the set of $\mathfrak{a}_0$-weights of the right hand side is $\{-(k_1 + 1)\alpha_1 - \dots - (k_{t - 1} + 1)\alpha_{t - 1}\mid k_s\in\mathbb{Z}_{\ge 0}\}$. Hence each $\mathfrak{a}_0$-weight appearing in the right hand side is less than that of the left hand side. This implies $e_t(e_1^{-1}\dots e_{t - 1}^{-1}) - (e_1^{-1}\dots e_{t - 1}^{-1})e_t = 0$. (2) We prove $e_t(e_1^{-(k_1 + 1)}\dotsm e_{t - 1}^{-(k_{t - 1} + 1)}) = (e_1^{-(k_1 + 1)}\dots e_{t - 1}^{-(k_{t - 1} + 1)})e_t$ in $S_{e_1}\otimes_{U(\mathfrak{g})}\dots \otimes_{U(\mathfrak{g})}S_{e_{t - 1}}$. As in the proof of (1), we have \begin{multline*} e_t(e_1^{-(k_1 + 1)}\dotsm e_{t - 1}^{-(k_{t - 1} + 1)}) - (e_1^{-(k_1 + 1)}\dotsm e_{t - 1}^{-(k_{t - 1} + 1)})e_t\\\in \bigoplus_{k_s\ge 0}\mathbb{C} e_1^{-(k_1 + 1)}\dotsm e_{t - 1}^{-(k_{t - 1} + 1)}. \end{multline*} An $\mathfrak{a}_\eta$-weight of the left hand side is $\sum_{e_s\in \mathfrak{n}_\eta,\ s < t}-\alpha_s + \alpha_t$. However, the set of $\mathfrak{a}_\eta$-weights of the right hand side is $\{\sum_{e_s\in\mathfrak{n}_\eta,\ s < t}-(k_s + 1)\alpha_s\mid k_s\in\mathbb{Z}_{\ge 0}\}$. Hence each $\mathfrak{a}_\eta$-weight appearing in the right hand side is less than that of the left hand side. This implies the lemma. (3) We may assume $X$ is a restricted root vector. Let $\alpha$ be a restricted root of $X$. Since $X$ normalizes $\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0$, we have \[ X(e_1^{-1}\dotsm e_l^{-1}) - (e_1^{-1}\dotsm e_l^{-1})X\in \bigoplus_{k_s\ge 0}\mathbb{C} e_1^{-(k_1 + 1)}\dotsm e_l^{-(k_l + 1)}. \] Then $X(e_1^{-1}\dotsm e_l^{-1}) - (e_1^{-1}\dotsm e_l^{-1})X$ has an $\mathfrak{a}_0$-weight $-(\alpha_1 + \dots + \alpha_s) + \alpha$. However, $e_1^{-(k_1 + 1)}\dotsm e_l^{-(k_l + 1)}$ has a $\mathfrak{a}_0$-weight $-((k_1 + 1)\alpha_1 + \dots + (k_l + 1)\alpha_l) < -(\alpha_1 + \dots + \alpha_s) + \alpha$. Hence $X(e_1^{-1}\dotsm e_l^{-1}) - (e_1^{-1}\dotsm e_l^{-1})X = 0$. \end{proof} \begin{lem}\label{lem:e^{-1}-part} Let $e_1,\dots, e_l$ be a basis of $\Ad(w_i)\overline{\mathfrak{n}}\cap\mathfrak{n}_0$ such that $e_s$ is a root vector and $\bigoplus_{s\le t - 1}\mathbb{C} e_s$ is an ideal of $\bigoplus_{s\le t}\mathbb{C} e_s$. Let $V$ be a $U(\mathfrak{m}\oplus\mathfrak{a})$-representation. Regard $V$ as a $\mathfrak{p}$-representation by $\mathfrak{n}V = 0$. By Lemma~\ref{lem:induction + twisting}, we have $T_{w_i}(U(\mathfrak{g})\otimes_{U(\mathfrak{p})}V) \simeq (\bigoplus_{k_s\ge 0}\mathbb{C} e_1^{-(k_1 + 1)}\dotsm e_l^{-(k_l + 1)})\otimes U(\Ad(w_i)\overline{\mathfrak{n}}\cap \overline{\mathfrak{n}_0})\otimes w_iV$. Then we have $\{v\in e_1^{-1}\dotsm e_l^{-1}\otimes 1\otimes w_iV\mid \mathfrak{n}_\eta v = 0\} = e_1^{-1}\dotsm e_l^{-1}\otimes 1\otimes H^0(\Ad(w_i)\mathfrak{m}\cap \mathfrak{n}_\eta,w_iV)$. \end{lem} \begin{proof} Take $v = e_1^{-1}\dotsm e_l^{-1}\otimes 1\otimes v_0\in H^0(\mathfrak{n}_\eta,T_{w_i}(U(\mathfrak{g})\otimes_{U(\mathfrak{p})}V)$. Then for $X\in \Ad(w_i)\mathfrak{m}\cap \mathfrak{n}_\eta$ we have $X(e_1^{-1}\dotsm e_l^{-1}\otimes 1\otimes v_0) = 0$. By Lemma~\ref{lem:relation in S_w}, we have $e_1^{-1}\dotsm e_l^{-1}\otimes 1\otimes Xv_0 = 0$. Hence $Xv_0 = 0$. \end{proof} By the definition of the Harish-Chandra homomorphism, we get the following lemma. \begin{lem}\label{lem:infinitesimal character of a space of partial highest weight vectors} Let $\mathfrak{q}$ be a parabolic subalgebra of $\mathfrak{g}$ containing $\mathfrak{h}\oplus\mathfrak{u}_0$. Take a Levi decomposition $\mathfrak{l}\oplus\mathfrak{u}_\mathfrak{q}$ of $\mathfrak{q}$ such that $\mathfrak{h}\subset \mathfrak{l}$. Let $\widetilde{W_\mathfrak{l}}\subset \widetilde{W}$ be the Weyl group of $\mathfrak{l}$, $V$ an $\mathfrak{l}$-module with an infinitesimal character $\widetilde{\mu}$. Put $V' = H^0(\mathfrak{u}_\mathfrak{q},V)$ and $\widetilde{\rho_{\mathfrak{u}_\mathfrak{q}}}(H) = (1/2)\Tr \ad(H)|_{\mathfrak{u}_\mathfrak{q}}$ for $H\in \mathfrak{h}$. Then $V'$ is $\mathfrak{l}$-stable and $V' = \bigoplus_{\widetilde{w}\in \widetilde{W_\mathfrak{l}}\backslash\widetilde{W}}(V')_{[\widetilde{w}\widetilde{\mu} - \widetilde{\rho_{\mathfrak{u}_\mathfrak{q}}}]}$ where $(V')_{[\widetilde{w}\widetilde{\mu} - \widetilde{\rho_{\mathfrak{u}_\mathfrak{q}}}]}$ is the maximal $\mathfrak{l}$-submodule which has an infinitesimal character $\widetilde{w}\widetilde{\mu} - \widetilde{\rho_{\mathfrak{u}_\mathfrak{q}}}$. In particular, for an $\mathfrak{l}$-submodule $V''$ of $V'$, a highest weight of $V'/V''$ belongs to $\{\widetilde{w}\widetilde{\mu} - \widetilde{\rho}\mid \widetilde{w}\in\widetilde{W}\}$. \end{lem} The following lemma is well-known. \begin{lem}\label{lem:distribution of weight} Let $V\in \mathcal{O}'$. Assume that $V$ has an infinitesimal character $\widetilde{\lambda}\in \mathfrak{h}^*$. Then a $\mathfrak{h}$-weight appearing in $V$ is contained in $\{\widetilde{w}\widetilde{\lambda} - \widetilde{\rho} - \alpha\mid \widetilde{w}\in\widetilde{W},\ \alpha\in\mathbb{Z}_{\ge 0}\Delta^+\}$. \end{lem} Now we determine the dimension of the space of Whittaker vectors of $\widetilde{I_i}/\widetilde{I_{i - 1}}$ under some conditions. \begin{lem}\label{lem:Whittaker vectors in a Bruhat cell, algebraic} Let $\widetilde{\mu}$ be an infinitesimal character of $\sigma$. Assume that for all $\widetilde{w}\in \widetilde{W}\setminus \widetilde{W_M}$, $(\lambda+\widetilde{\mu}) - \widetilde{w}(\lambda+\widetilde{\mu})\not\in \mathbb{Z}\Delta$. Then we have $\dim\Wh_\eta(\widetilde{I_i}/\widetilde{I_{i - 1}}) = \dim \Wh_{w_i^{-1}\eta}((\sigma_{\text{\normalfont $M\cap K$-finite}})^*)$. \end{lem} \begin{proof} Put $V = T_{w_i}(U(\mathfrak{g})\otimes_{U(\mathfrak{p})}J^*(\sigma\otimes e^{\lambda+\rho}))$. By Theorem~\ref{thm:stucture of J^*(I(sigma,lambda))}, we have $\Wh_\eta(\widetilde{I_i}/\widetilde{I_{i - 1}}) = \Wh_\eta(C(V))$. Let $e_1,\dots, e_l$ be a basis of $\Ad(w_i)\overline{\mathfrak{n}}\cap\mathfrak{n}_0$ such that $\bigoplus_{s\le t - 1}\mathbb{C} e_s$ is an ideal of $\bigoplus_{s\le t}\mathbb{C} e_s$. Moreover, assume that each $e_i$ is a root vector. For $\mathbf{k} = (k_1,\dots,k_l)\in \mathbb{Z}^l$, put $e^\mathbf{k} = e_1^{k_1}\dotsm e_l^{k_l}$. Set $\mathbf{1} = (1,\dots,1)\in\mathbb{Z}^l$. Then we have \[ V = \bigoplus_{k\in\mathbb{Z}_{\ge 0}^l}\mathbb{C} e^{-(\mathbf{k} + \mathbf{1})}\otimes U(\Ad(w_i)\overline{\mathfrak{n}}\cap\overline{\mathfrak{n}_0})\otimes w_iJ^*(\sigma\otimes e^{\lambda+\rho}). \] Put \[ V' = \bigoplus_{\mathbf{k}\in \mathcal{A}}e^{-(\mathbf{k} + \mathbf{1})}\otimes U(\Ad(w_i)\overline{\mathfrak{n}}\cap\overline{\mathfrak{n}_0}\cap \mathfrak{m}_\eta)\otimes H^0(\mathfrak{m}\cap\mathfrak{n}_\eta,w_iJ^*(\sigma\otimes e^{\lambda+\rho})) \] where $\mathcal{A} = \{(k_1,\dots,k_l)\in \mathbb{Z}^l_{\ge 0}\mid \text{if $e_i\in \mathfrak{n}_\eta$ then $k_i = 0$}\}$. It is easy to see that $V'$ is an $\mathfrak{m}_\eta\oplus \mathfrak{a}_\eta$-stable and $V'\subset H^0(\mathfrak{n}_\eta,V)$. We prove that $V' = H^0(\mathfrak{n}_\eta,V)$. To prove $V' = H^0(\mathfrak{n}_\eta,V)$, it is sufficient to prove that there exists no highest weight vector in $H^0(\mathfrak{n}_\eta,V)/ V'$. Let $v\in H^0(\mathfrak{n}_\eta,V)$ such that $(\mathfrak{m}_\eta\cap \mathfrak{u})v \in V'$. First, we prove that $v\in e^{-\mathbf{1}}\otimes U(\Ad(w_i)\overline{\mathfrak{n}}\cap \overline{\mathfrak{n}_0})\otimes J^*(\sigma\otimes e^{\lambda + \rho}) + V'$. Take $y_\mathbf{k}\in U(\Ad(w_i)\overline{\mathfrak{n}}\cap \overline{\mathfrak{n}_0})\otimes J^*(\sigma\otimes e^{\lambda + \rho})$ such that $v = \sum_{\mathbf{k}} e^{-(\mathbf{k} + \mathbf{1})}\otimes y_\mathbf{k}$. We prove that if $k_t\ne 0$ and $e_t\in \mathfrak{n}_\eta$ then $y_\mathbf{k} = 0$ by induction on $t$ where $\mathbf{k} = (k_1,\dots,k_l)$. Put $\mathbf{1}_t = (\delta_{st})_{1\le s\le l}\in \mathbb{Z}^l$ ($\delta_{st}$ is Kronecker's delta). By inductive hypothesis, for $s < t$ such that $e_s\in\mathfrak{n}_\eta$, if $y_\mathbf{k}\ne 0$ then $k_s = 0$. By Lemma~\ref{lem:relation in S_w} (2), we have $e_tv = \sum_{\mathbf{k}\in\mathbb{Z}_{\ge 0}^l}e^{-(\mathbf{k} + \mathbf{1}) + \mathbf{1}_t}\otimes y_\mathbf{k}$. Since $v\in H^0(\mathfrak{n}_\eta,V)$, we have $e_tv = 0$. Hence if $e^{-(\mathbf{k} + \mathbf{1}) + \mathbf{1}_t}\ne 0$ then $y_\mathbf{k} = 0$. Since $e^{-(\mathbf{k} + \mathbf{1}) + \mathbf{1}_t} = 0$ if and only if $k_t = 0$, $k_t \ne 0$ implies $y_\mathbf{k} = 0$. We prove that if $k_t\ne 0$ then $e^{-(\mathbf{k} + \mathbf{1})}\otimes y_\mathbf{k}\in V'$ by induction on $t$. If $e_t\in \mathfrak{n}_\eta$ then this claim is already proved. We may assume that $e_t\in \mathfrak{m}_\eta$. Hence $e_tV'\subset V'$. By inductive hypothesis, if $k_s\ne 0$ for some $s < t$ then $e^{-(\mathbf{k} + \mathbf{1})}\otimes y_\mathbf{k} \in V'$. Then we have $e_tv\in \sum_{\mathbf{k}\in\mathbb{Z}_{\ge 0}^l}e^{-(\mathbf{k} + \mathbf{1}) + \mathbf{1}_t}\otimes y_\mathbf{k} + V'$ by Lemma~\ref{lem:relation in S_w} (1). Since $e_tv\in V'$, we have $\sum_{\mathbf{k}\in\mathbb{Z}_{\ge 0}^l}e^{-(\mathbf{k} + \mathbf{1}) + \mathbf{1}_t}\otimes y_\mathbf{k} \in V'$. By the definition of $V'$, if $e^{-(\mathbf{k} + \mathbf{1}) + \mathbf{1}_t}\ne 0$ then $e^{-(\mathbf{k} + \mathbf{1})}\otimes y_\mathbf{k}\in V'$. Hence we get the claim. We may assume that $v$ is a weight vector with respect to $\mathfrak{h}$. We can take $\widetilde{w}\in \widetilde{W}$ such that $-\widetilde{w}(\lambda+\widetilde{\mu}) - \widetilde{\rho}$ is a $\mathfrak{h}$-weight of $v$ by Lemma~\ref{lem:infinitesimal character of a space of partial highest weight vectors}. Put $\widetilde{\rho_M} = \sum_{\alpha\in\Delta_M^+}(1/2)\alpha$. Since $J^*(\sigma\otimes e^{\lambda + \rho})$ has an infinitesimal character $-(\lambda + \widetilde{\mu} + \rho)$, a $\mathfrak{h}$-weight appearing in $J^*(\sigma\otimes e^{\lambda + \rho})$ is contained in $\{-\widetilde{w}(\lambda + \widetilde{\mu} + \rho) - \widetilde{\rho_M} + \alpha\mid \widetilde{w}\in \widetilde{W_M},\ \alpha\in\mathbb{Z}\Delta_M\}$ by Lemma~\ref{lem:distribution of weight}. Since $-\rho\in\mathfrak{a}^*$, we have $\widetilde{w}\rho = \rho$ for $\widetilde{w}\in \widetilde{W_M}$. Hence we have $- \widetilde{w}\rho - \widetilde{\rho_M} = -\rho - \widetilde{\rho_M} = -\widetilde{\rho}$. Notice that $w_i\widetilde{\rho} - \widetilde{\rho}\in \mathbb{Z}\Delta$. Therefore a $\mathfrak{h}$-weight appearing in $V$ is contained in \begin{multline*} -w_i\widetilde{W_M}(\lambda + \widetilde{\mu}) - w_i\widetilde{\rho} + w_i\mathbb{Z}\Delta_M + \mathbb{Z}_{\ge 0}(w_i\Delta^-\cap \Delta^-) - \mathbb{Z}_{\ge 1}(w_i\Delta^-\cap \Delta^+)\\ \subset -w_i\widetilde{W_M}(\lambda + \widetilde{\mu}) - \widetilde{\rho} + \mathbb{Z}\Delta. \end{multline*} This implies that for some $\widetilde{w'}\in\widetilde{W_M}$, we have $\widetilde{w}(\lambda + \widetilde{\mu}) - w_i\widetilde{w'}(\lambda +\widetilde{\mu})\in\mathbb{Z}\Delta$. By the assumption we have $\widetilde{w}\in w_i\widetilde{W_M}$. This implies $(\wt v)(\Ad(w_i)H) = -(\lambda(H) + w_i^{-1}\widetilde{\rho}(H))$ for all $H\in \mathfrak{a}$ where $\wt v$ is a $\mathfrak{h}$-weight of $v$. Take $T_p\in U(\Ad(w_i)\overline{\mathfrak{n}}\cap \overline{\mathfrak{n}_0})$ and $x_p\in w_iJ^*(\sigma\otimes e^{\lambda + \rho})$ such that $v \in \sum_p e^{-\mathbf{1}}\otimes T_p\otimes x_p + V'$. We may assume that $T_p$ (resp.\ $x_p$) is a $\mathfrak{h}$-weight vector with respect to the adjoint action (resp.\ the action induced from $\sigma\otimes e^{\lambda + \rho}$). We denote its $\mathfrak{h}$-weight by $\wt T_p$ and $\wt x_p$. Fix $H\in \mathfrak{a}$. Then $\alpha(H) = 0$ for all $\alpha\in\Delta_M$. Since $\wt x_p\in -w_i(\widetilde{W_M}(\lambda + \widetilde{\mu}) + \widetilde{\rho} + \mathbb{Z}\Delta_M)$, $(\wt x_p)(\Ad(w_i)H) = -(\lambda + \widetilde{\rho})(H)$. Hence \begin{align*} &(\wt v)(\Ad(w_i)H)\\ & = (\wt(e^{-\mathbf{1}}) + \wt(T_p) + \wt(x_p))(\Ad(w_i)(H))\\ & = (\wt(e^{-\mathbf{1}})(\Ad(w_i)H) + (\wt T_p)(\Ad(w_i)H) - (\lambda + \widetilde{\rho})(H)\\ & = (\wt(e^{-\mathbf{1}})(\Ad(w_i)H) + (\wt T_p)(\Ad(w_i)H) - (\lambda + \widetilde{\rho})(H). \end{align*} We calculate $\wt(e^{-\mathbf{1}})(\Ad(w_i)H)$. By the definition, $\wt(e^{-\mathbf{1}})(\Ad(w_i)H) = \Tr\ad(\Ad(w_i)H)|_{\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0}$. Since we have $\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0 = \Ad(w_i)\overline{\mathfrak{n}_0}\cap \mathfrak{n}_0$, we have \begin{multline*} \Tr\ad(\Ad(w_i)H)|_{\Ad(w_i)\overline{\mathfrak{n}}\cap \mathfrak{n}_0} = \Tr\ad(\Ad(w_i)H)|_{\Ad(w_i)\overline{\mathfrak{n}_0}\cap \mathfrak{n}_0}\\ ~ \Tr\ad(H)|_{\Ad(w_i)^{-1}\mathfrak{n}_0\cap \overline{\mathfrak{n}_0}} = (-\widetilde{\rho} + w_i^{-1}\widetilde{\rho})(H). \end{multline*} Hence we get \[ (\wt v)(\Ad(w_i)H) = (\wt T_p)(\Ad(w_i)H) - (\lambda + w_i^{-1}\widetilde{\rho})(H). \] We have already proved that $(\wt v)(\Ad(w_i)H) = -(\lambda + w_i^{-1}\widetilde{\rho})(H)$. Therefore we get $(\wt T_p)(\Ad(w_i)H) = 0$ for all $H\in\mathfrak{a}$. Since $T_p\in U(\Ad(w_i)\mathfrak{n})$, this implies $T_p\in \mathbb{C}$, i.e., there exist $v'\in e_1^{-1}\dotsm e_l^{-1}\otimes 1\otimes w_iJ^*(\sigma\otimes e^{\lambda+\rho})$ and $v''\in V'$ such that $v = v' + v''$. Therefore $\mathfrak{n}_\eta(v') = \mathfrak{n}_\eta(v - v'') = 0$. Hence, $v'\in V'$ by Lemma~\ref{lem:e^{-1}-part}. Therefore $H^0(\mathfrak{n}_\eta,V) = V'$. For an $\mathfrak{m}_0\oplus\mathfrak{a}_0$-module $\tau$ and a subalgebra $\mathfrak{c}$ of $\mathfrak{g}$ containing $\mathfrak{m}_0\oplus\mathfrak{a}_0$, put $M_{\mathfrak{c}}(\tau) = U(\mathfrak{c})\otimes_{U(\mathfrak{c}\cap\overline{\mathfrak{p}_0})}(\tau\otimes\rho')$ where $\overline{\mathfrak{n}_0}\cap \mathfrak{c}$ acts on $\tau$ trivially and $\rho'(H) = (\Tr(\ad(H)|_{\mathfrak{c}\cap\overline{\mathfrak{n}_0}}))/2$ for $H\in\mathfrak{a}_0$. For $\widetilde{\lambda}\in \mathfrak{h}^*$ such that $\widetilde{\lambda}|_{\mathfrak{m}_0}$ is regular dominant integral, let $\sigma_{M_0A_0,\widetilde{\lambda}}$ be the finite-dimensional representation of $M_0A_0$ with an infinitesimal character $\widetilde{\lambda}$. Let $\ch M$ be the character of $M$. We can take integers $c_{\widetilde{\lambda}}$ such that \begin{multline*} \ch D' H^0(\mathfrak{n}_\eta\cap \Ad(w_i)\mathfrak{m},w_iJ^*(\sigma\otimes e^{\lambda+\rho}))\\ = \sum_{\widetilde{\lambda}}c_{\widetilde{\lambda}}\ch M_{(\mathfrak{m}_\eta\cap\Ad(w_i)\mathfrak{m}) + \mathfrak{a}_0}(\sigma_{M_0A_0,\widetilde{\lambda}}) \end{multline*} Then we have $\ch D'V' = \sum_{\widetilde{\lambda}}c_{\widetilde{\lambda}}\ch M_{\mathfrak{m}_\eta\oplus\mathfrak{a}_\eta}(\sigma_{M_0A_0,\widetilde{\lambda}})$. The functor $X\mapsto \Wh_{\eta|_{\mathfrak{m}_\eta\cap\mathfrak{n}_0}}(X^*)$ is an exact functor by a result of Lynch~\cite{lynch-whittaker}. Hence, we have $\dim\Wh_{\eta|_{\mathfrak{m}_\eta\cap\mathfrak{n}_0}}(C(V')) = \sum_{\widetilde{\lambda}}c_{\widetilde{\lambda}}\dim\Wh_{\eta|_{\mathfrak{m}_\eta\cap\mathfrak{n}_0}}(M_{\mathfrak{m}_\eta\oplus\mathfrak{a}_\eta}(\sigma_{M_0A_0,\widetilde{\lambda}})^*)$. Lynch also proves $\dim\Wh_{\eta|_{\mathfrak{m}_\eta\cap\mathfrak{n}_0}}(M_{\mathfrak{m}_\eta}(\sigma_{M_0A_0,\widetilde{\lambda}})^*) = \dim\sigma_{M_0A_0,\widetilde{\lambda}}$. Therefore, by Lemma~\ref{lem:two step calc for Whittaker vector}, we have $\dim \Wh_\eta(\widetilde{I_i}/\widetilde{I_{i - 1}}) = \dim\Wh_{\eta|_{\mathfrak{m}_\eta\cap\mathfrak{n}_0}}(C(V')) = \sum_{\widetilde{\lambda}} c_{\widetilde{\lambda}}\dim\sigma_{M_0A_0,\widetilde{\lambda}}$. By the same argument we have \begin{align*} &\sum_{\widetilde{\lambda}} c_{\widetilde{\lambda}}\dim\sigma_{M_0A_0,\widetilde{\lambda}}\\ & = \sum_{\widetilde{\lambda}} c_{\widetilde{\lambda}}\dim \Wh_{\eta|_{\mathfrak{m}_\eta\cap\Ad(w_i)\mathfrak{m}\cap\mathfrak{n}_0}}(M_{(\mathfrak{m}_\eta\cap\Ad(w_i)\mathfrak{m}) + \mathfrak{a}_0}(\sigma_{M_0A_0,\widetilde{\lambda}})^*) \\ & = \dim\Wh_{\eta|_{\mathfrak{m}_\eta\cap\Ad(w_i)\mathfrak{m}\cap\mathfrak{n}_0}}(CH^0(\mathfrak{n}_\eta\cap \Ad(w_i)\mathfrak{m},w_iJ^*(\sigma\otimes e^{\lambda+\rho})))\\ & = \dim \Wh_{\eta|_{\Ad(w_i)\mathfrak{m}\cap\mathfrak{n}_0}}(C(w_iJ^*(\sigma\otimes e^{\lambda+\rho})))\\ & = \dim\Wh_{w_i^{-1}\eta}(C(J^*(\sigma\otimes e^{\lambda+\rho})))\\ & = \dim \Wh_{w_i^{-1}\eta}((\sigma_{\text{$M\cap K$-finite}})^*). \end{align*} This implies the lemma. \end{proof} \begin{thm}\label{thm:dimension Whittaker vectors, algebraic} Let $\widetilde{\mu}$ be an infinitesimal character of $\sigma$. Assume that for all $\widetilde{w}\in \widetilde{W}\setminus \widetilde{W_M}$, $(\lambda+\widetilde{\mu}) - \widetilde{w}(\lambda+\widetilde{\mu})\not\in \mathbb{Z}\Delta$. Then we have \[ \dim\Wh_\eta((I(\sigma,\lambda)_{\text{\normalfont $K$-finite}})^*) = \sum_{w\in W(M)}\dim \Wh_{w^{-1}\eta}((\sigma_{\text{\normalfont $M\cap K$-finite}})^*). \] \end{thm} \begin{proof} Since a $\mathfrak{h}$-weight appearing in $T_{w_i}(U(\mathfrak{g})\otimes_{U(\mathfrak{p})}J^*(\sigma\otimes e^{\lambda+\rho}))$ belongs to $\{-w_i\widetilde{w}(\lambda+\widetilde{\mu}) - \widetilde{\rho} + \alpha\mid \widetilde{w}\in\widetilde{W_M},\ \alpha\in\Delta\}$, the exact sequence $0\to I_{i - 1} \to I_i\to T_{w_i}(U(\mathfrak{g})\otimes_{U(\mathfrak{p})}J^*(\sigma\otimes e^{\lambda+\rho}))\to 0$ splits. Hence, we have $J^*_\eta(I(\sigma,\lambda)) = \bigoplus_i\Gamma_\eta(C(T_{w_i}(U(\mathfrak{g})\otimes_{U(\mathfrak{p})}J^*(\sigma\otimes e^{\lambda+\rho}))))$. Therefore the theorem follows from Lemma~\ref{lem:Whittaker vectors in a Bruhat cell, algebraic}. \end{proof} Finally we study the case of $\sigma$ is finite-dimensional. If $\sigma$ is finite-dimensional, then $\mathfrak{m}\cap\mathfrak{n}_0$ acts on $\sigma$ nilpotently. Hence $\Wh_{w_i^{-1}\eta}(\sigma^*)\ne 0$ if and only if $w_i^{-1}\eta = 0$ on $\mathfrak{m}\cap\mathfrak{n}_0$. \begin{defn} Let $\Theta,\Theta_1,\Theta_2$ be subsets of $\Pi$. \begin{enumerate} \item Put $W(\Theta) = \{w\in W\mid w(\Theta)\subset \Sigma^+\}$ and $\Sigma_\Theta = \mathbb{Z}\Theta\cap \Sigma$.\newsym{$W(\Theta)$} \item Put $W(\Theta_1,\Theta_2) = \{w\in W(\Theta_1)\cap W(\Theta_2)^{-1}\mid w(\Sigma_{\Theta_1})\cap \Sigma_{\Theta_2} = \emptyset\}$.\newsym{$W(\Theta_1,\Theta_2)$} \item Let $W_\Theta$ be the Weyl group of $\Sigma_\Theta$.\newsym{$W_\Theta$} \end{enumerate} \end{defn} \begin{lem}\label{lem:coparison oshima} Let $\Theta$ be a subset of $\Pi$ corresponding to $P$. \begin{enumerate} \item We have $\#W(\supp\eta,\Theta) = \#\{w\in W(M)\mid w(\Sigma^+)\cap \Sigma^+_\eta = \emptyset\}$. \item We have $\#W(\supp\eta,\Theta)\times\# W_{\supp\eta} = \#\{w\in W(M)\mid \supp\eta\cap w(\Sigma_M^+) = \emptyset\}$. \end{enumerate} \end{lem} \begin{proof} (1) Put $\mathcal{W} = \{w\in W(M)\mid w(\Sigma^+)\cap \Sigma^+_\eta = \emptyset\}$. Let $w_{\eta,0}$ be the longest Weyl element of $W_{M_\eta}$. We prove that the map $\mathcal{W}\to W(\supp\eta,\Theta)$ defined by $w\mapsto (w_{\eta,0}w)^{-1}$ is well-defined and bijective. First we prove that the map is well-defined. Let $w\in\mathcal{W}$. The equation $w(\Sigma^+)\cap \Sigma_\eta^+ = \emptyset$ implies that $(w_{\eta,0}w)^{-1}(\Sigma^+_\eta)\subset \Sigma^+$. Hence, $(w_{\eta,0}w)^{-1}\in W(\supp\eta)$. Moreover, $w(\Sigma_M^+)\subset \Sigma^+$ and $w(\Sigma^+)\cap \Sigma_\eta^+ = \emptyset$ imply that $w(\Sigma_M^+)\subset \Sigma^+\cap (\Sigma\setminus\Sigma^+_\eta) = \Sigma^+\setminus\Sigma^+_\eta$. Hence, $(w_{\eta,0}w)(\Sigma_M^+)\subset \Sigma^+\setminus\Sigma^+_\eta\subset \Sigma^+$. We have $(w_{\eta,0}w)^{-1}\in W(\Theta)^{-1}$. Finally $w(\Sigma_M^+)\subset \Sigma^+\setminus\Sigma_\eta^+$ implies $w(\Sigma)\subset \Sigma\setminus\Sigma_\eta$. Hence we have $(w_{\eta,0}w)^{-1}\Sigma_\eta\cap\Sigma_M = w^{-1}\Sigma_\eta\cap \Sigma_M = \emptyset$. Assume that $(w_{\eta,0}w)^{-1}\in W(\supp\eta,\Theta)$. Then $(w_{\eta,0}w)^{-1}(\Sigma^+_\eta)\subset\Sigma^+$ implies that $w(\Sigma^+)\cap \Sigma^+_\eta=\emptyset$. Since $(w_{\eta,0}w)^{-1}\Sigma_\eta\cap\Sigma_M = \emptyset$ we have $w(\Sigma_M)\cap \Sigma_\eta = \emptyset$. By $(w_{\eta,0}w)(\Sigma_M^+)\subset\Sigma^+$ and $w(\Sigma^+)\cap \Sigma_\eta^+ = \emptyset$, we have $w(\Sigma_M^+)\subset ((\Sigma^+\setminus\Sigma_\eta^+)\cup \Sigma_\eta^-)\cap (\Sigma\setminus\Sigma^-_\eta) = (\Sigma^+\setminus\Sigma^+_\eta)$. Consequently we have $w\in W(M)$. (2) Put $\mathcal{W} = \{w\in W(M)\mid \supp\eta\cap w(\Sigma_M^+) = \emptyset\}$. Define the map $\varphi\colon W(\supp\eta,\Theta)\times W_{\supp\eta} \to \mathcal{W}$ by $(w_1,w_2)\mapsto w_2w_1^{-1}$. This map is injective since $W(\supp\eta,\Theta)\subset W(\supp\eta)$. We prove that $\varphi$ is well-defined and surjective. Since $w_1^{-1}(\Sigma_M^+) = w_1^{-1}(\Sigma_M^+)\cap \Sigma^+\subset \Sigma^+\setminus \Sigma_\eta^+$, we have $w_2w_1^{-1}(\Sigma_M^+)\subset \Sigma^+\setminus\Sigma_\eta^+$. Hence, $\varphi$ is well-defined. Next let $w\in\mathcal{W}$. Let $w_1\in W(\supp\eta)^{-1}$ and $w_2\in W_{\supp\eta}$ such that $w = w_2w_1^{-1}$. Then $w_1^{-1}(\Sigma_M^+) = w_2^{-1}w(\Sigma_M^+)\subset w_2^{-1}(\Sigma^+\setminus\Sigma_\eta^+) = \Sigma^+\setminus\Sigma^+_\eta$. This implies $w_1\in W(\supp\eta,\Theta)$. \end{proof} \begin{lem}\label{lem:whittaker vectors of finite-dimensional representation} Assume that $\sigma$ is irreducible and finite-dimensional. Let $\widetilde{\mu}$ be the highest weight of $\sigma$ and $V$ the irreducible finite-dimensional representation of $M_0A_0$ with highest weight $\lambda+\widetilde{\mu}$. Then we have $\sigma/(\mathfrak{m}\cap\mathfrak{n}_0)\sigma\simeq V$ as an $M_0A_0$-module. In particular, $\dim\Wh_0(\sigma') = \dim V$. \end{lem} \begin{proof} We prove that $\Wh_0(\sigma^*) \simeq V^*$. Let $\widetilde{w}_{M,0}$ be the longest element of $\widetilde{W_M}$. Then both sides have a highest weight $-\widetilde{w}_{M,0}(\widetilde{\mu} + \lambda)$ and the space of highest weight vectors are $1$-dimensional. \end{proof} As an Corollary of Theorem~\ref{thm:dimension Whittaker vectors} and Theorem~\ref{thm:dimension Whittaker vectors, algebraic}, we have the following theorem announced by T. Oshima. Define $\widetilde{\rho_M}\in\mathfrak{h}^*$ by $\widetilde{\rho_M} = (1/2)\sum_{\alpha\in\Delta_M^+}\alpha$. \begin{thm}\label{thm:Whittaker vectors, finite-dimensional case} Assume that $\sigma$ is the irreducible finite-dimensional representation with highest weight $\widetilde{\nu}$. Let $\dim_M(\lambda+\widetilde{\nu})$ be a dimension of the finite-dimensional irreducible representation of $M_0A_0$ with highest weight $\lambda+\widetilde{\nu}$. \begin{enumerate} \item Assume that for all $w\in W$ such that $\eta|_{wN_0w^{-1}\cap N_0} = 1$ the following two conditions hold: \begin{enumerate} \item For all $\alpha\in \Sigma^+\setminus w^{-1}(\Sigma^+_M\cup\Sigma_\eta^+)$ we have $2\langle\alpha,\lambda+w_0\widetilde{\nu}\rangle/\lvert\alpha\rvert^2\not\in\mathbb{Z}_{\le 0}$. \item For all $\widetilde{w}\in\widetilde{W}$ we have $\lambda - \widetilde{w}(\lambda + \widetilde{\nu} + \widetilde{\rho_M})|_\mathfrak{a}\notin \mathbb{Z}_{\le 0}((\Sigma^+\setminus \Sigma_M^+)\cap w^{-1}\Sigma^+)|_\mathfrak{a}\setminus\{0\}$. \end{enumerate} Then we have \[ \dim \Wh_\eta(I(\sigma,\lambda)') = \# W(\supp\eta,\Theta)\times(\dim_M(\lambda+\widetilde{\nu})) \] \item Assume that for all $\widetilde{w}\in\widetilde{W}\setminus\widetilde{W_M}$, $(\lambda+\widetilde{\nu}) - \widetilde{w}(\lambda+\widetilde{\nu}) \not\in\Delta$. Then we have \begin{multline*} \dim\Wh_\eta((I(\sigma,\lambda)_{\text{\normalfont $K$-finite}})^*)\\ = \# W(\supp\eta,\Theta)\times \#W_{\supp\eta}\times(\dim_M(\lambda+\widetilde{\nu})) \end{multline*} \end{enumerate} \end{thm}

https://arxiv.org/abs/1009.0222

A further note on the inverse nodal problem and Ambarzumyan problem for the p-Laplacian

In this note, we extend some results in a previous paper on the inverse nodal problem and Ambarzumyan problem for the p-Laplacian to periodic or anti-periodic boundary conditions, and to L^1 potentials.

\section{Introduction} \setcounter{equation}{0} \hskip0.25in Recently, we studied the $p$-Laplacian with $C^{1}$-potentials and solved the inverse nodal problem and Ambarzumyan problem for Dirichlet boundary conditions \cite{LLW08}. In this note, we want to extend the results to periodic or anti-periodic boundary conditions, and to $L^1$ potentials. \par Consider the equation \begin{equation} -\left(y'^{(p-1)}\right)'=(p-1)(\la -q(x))y^{(p-1)}\ ,\label{eq1.1} \end{equation} where $f^{(p-1)}=|f|^{p-1} {\rm sgn} f$. Assume that $q(1+x)=q(x)$ for $x\in\mathbb{R}$, then (\ref{eq1.1}) can be coupled with periodic or anti-periodic boundary conditions respectively: \begin{equation} y(0)=y(1)\ ,\quad y'(0)=y'(1)\qquad \label{eq1.2}\ \ \ \end{equation} or \begin{equation} y(0)=-y(1)\ ,\quad y'(0)=-y'(1).\qquad \label{eq1.3} \end{equation} When $p=2$, the above is the classical Hill's equation. It follows from Floquet theory that there are countably many interlacing periodic and anti-periodic eigenvalues of Hill's operator. However, Floquet theory does not work for the case $p\neq 2$. In 2001, Zhang \cite{Zhang2001} studied the properties of eigenvalues for $p>1$ with $L^{1}$-potentials. He applied the rotation number function to define the minimal eigenvalue $\underline{\lambda}_n(q)$ and the maximal eigenvalue $\overline{\lambda}_n(q)$ corresponding to eigenfunctions having $n$ zeros in $[0,1)$, respectively. These numbers $\underline{\lambda}_n(q)$ and $\overline{\lambda}_n(q)$ are called rotational periodic eigenvalues and satisfy \begin{enumerate} \item[(i)] If $n\in \mathbb{N}\cup \{0\}$ is even, then $\underline{\lambda}_n(q)$ and $\overline{\lambda}_n(q)$ are eigenvalues of (\ref{eq1.1}) and (\ref{eq1.2}); if $n\in \mathbb{N}$ is odd, then $\underline{\lambda}_n(q)$ and $\overline{\lambda}_n(q)$ are eigenvalues of (\ref{eq1.1}) and (\ref{eq1.3}). \item[(ii)] $\overline{\lambda}_0(q)<\underline{\lambda}_1(q)\leq\overline{\lambda}_1(q)<\underline{\lambda}_2(q) \leq \overline{\lambda}_2(q)< \cdots \cdots$. \end{enumerate} Although the above properties are very similar to the linear case, it should be mentioned that the case for the $p$-Laplacian is much more complicated. For example, for the periodic or anti-periodic boundary conditions, there may exist an infinite sequence of variational eigenvalues and non-variational eigenvalues (\cite{BR2008}). In the same paper, the authors also showed that the minimal periodic eigenvalue is simple and variational, while the minimal anti-periodic eigenvalue is variational but may be not simple. \par In 2008, Brown and Eastham \cite{BE08} derived a sharp asymptotic expansion of eigenvalues of the $p$-Laplacian with locally integrable and absolutely continuous $(r-1)$ derivative potentials respectively. Below is a version of their theorem for periodic eigenvalues of the $p$-Laplacian (\ref{eq1.1}), (\ref{eq1.2}). \newtheorem{th1.1}{Theorem}[section] \begin{th1.1} \label{th1.1}(\cite[Theorem 3.1]{BE08}) Let $q$ be $1$-periodic and locally integrable in $(-\infty, \infty)$. Then the rotationally periodic eigenvalue $\la_{2n}=\underline{\lambda}_{2n}$, or $\overline{\lambda}_{2n}$ satisfies \begin{equation} \label{eq1.4} \lambda_{2n}^{1/p}=2n\widehat{\pi}+\frac{1}{p(2n\widehat{\pi})^{p-1}} \int_{0}^{1}q(t)dt+o(\frac{1}{n^{p-1}}). \end{equation} \end{th1.1} By a similar argument, the asymptotic expansion of the anti-periodic eigenvalue ${\lambda}_{2n-1}=\underline{\lambda}_{2n-1}$ or $\overline{\lambda}_{2n-1}$, which corresponds to the anti-periodic eigenfunction with $2n-1$ zeros in $[0,1)$, satisfies \begin{equation} \label{eq1.5} \lambda_{2n-1}^{1/p}=(2n-1)\widehat{\pi}+\frac{1}{p((2n-1)\widehat{\pi})^{p-1}} \int_{0}^{1}q(t)dt+o(\frac{1}{n^{p-1}}) .\end{equation} \hskip0.25in The inverse nodal problem is the problem of understanding the potential function through its nodal data. In 2006, some of us (C.-L.) \cite{CL062} studied Hill's equation. We first made a translation of the interval by the first nodal length so that the periodic problem is reduced to a Dirichlet problem, and then solved the uniqueness, reconstruction and stability problems using the nodal set of periodic eigenfunctions. \par We denote by $\{ x_i^{(n)}\}_{i=0}^{n-1}$ the zeros of the eigenfunction corresponding to $\la_{n}$, and define the nodal length $\ell_{i}^{(n)}=x_{i+1}^{(n)}-x_{i}^{(n)}$ and $j=j_{n}(x)=\max\{i :\ x_{i}^{(n)}\leq x \}$. Our main theorem is as follows. \newtheorem{th1.2}[th1.1]{Theorem} \begin{th1.2} \label{th1.2} Let $q\in L^{1}(0,1)$ be $1$-periodic. Define $F_n(x)$ as the following: \begin{enumerate} \item[(a)]For periodic boundary condition, let $$F_{2n}(x)=p(2n\widehat{\pi})^p[(2n)\ell_j^{(2n)}-1]+\int_0^1q(t)dt,$$ \item[(b)]For the anti-periodic boundary condition, let $$F_{2n-1}(x)=p((2n-1)\widehat{\pi})^p[(2n-1)\ell_j^{(2n-1)}-1]+\int_0^1q(t)dt.$$ \end{enumerate} Then both $\{F_{2n}\}$ and $\{F_{2n-1}\}$ converges to $q$ pointwisely a.e. and in $L^{1}(0,1)$. \end{th1.2} Thus either one of the sequences $\{F_{2n}\}/ \{F_{2n-1}\}$ will give the reconstruction formula for $q$. Note that here $q\in L^1(0,1)$. Furthermore, the map between the nodal space and the set of admissible potentials are homeomorphic after a partition (cf.\cite{LLW08}). The same idea also works for linear separated boundary value problems with integrable potentials. \par Using the eigenvalue asymptotics above, the Ambarzumyan problems for the periodic and anti-periodic boundary conditions can also be solved. \newtheorem{th1.3}[th1.1]{Theorem} \begin{th1.3} \label{th1.3}Let $q\in L^{1}(0,1)$ be periodic of period $1$. \begin{enumerate} \item[(a)] If the spectrum of periodic eigenvalues problem (\ref{eq1.1}), (\ref{eq1.2}) contains $\{(2n\widehat{\pi})^p:n\in \mathbb{N}\cup \{0\}\}$ and $0$ is the least eigenvalue, then $q=0$ on $[0,1]$. \item[(b)] If the spectrum of anti-periodic eigenvalue problem (\ref{eq1.1}), (\ref{eq1.3}) contains $\{((2n-1)\widehat{\pi})^p:n\in \mathbb{N}\}$; $\widehat{\pi}^p$ is the least eigenvalue and $\int_0^1q(t)(S_p(\widehat{\pi}t)S_p'(\widehat{\pi}t)^{(p-1)})'dt=0$, then $q=0$ on $[0,1]$. \end{enumerate} \end{th1.3} In section 2, we shall apply Theorem \ref{th1.1} to study on periodic and anti-periodic boundary conditions. In section 3, we shall deal with the case of linear separated boundary conditions. \par The stability issue of the inverse nodal problem with $L^1$ potentials associated with perodic/antiperiodic as well as linear separated boundary conditions can also be proved. The proof goes in the same manner as in \cite{LLW08} and is so omitted. \section{ Proof of main results} \setcounter{equation}{0} \hskip0.25in Fix $p>1$ and assume that $q=0$ and $\lambda=1$. Then (\ref{eq1.1}) becomes $$-(y'^{(p-1)})'=(p-1)y^{(p-1)}.$$ Let $S_p$ be the solution satisfying the initial conditions $\displaystyle S_p(0)=0$, $S_p'(0)=1$. It is well known that $S_p$ and its derivative $S_p'$ are periodic functions on $\mathbb{R}$ with period $2\widehat\pi$, where $ {\widehat \pi}=\frac{2\pi}{p\sin(\frac{\pi}{p})}$. The two functions also satisfy the following identities (cf.\ \cite{BE08,LLW08}). \newtheorem{th2.1}{Lemma}[section] \begin{th2.1} \label{th2.1} \begin{enumerate} \item[(a)] $|S_p(x)|^p+|S_p'(x)|^p=1$ for any $x\in \mathbb{R}$; \item[(b)] $(S_pS_p'^{(p-1)})'=|S_p'|^p-(p-1)|S_p|^p=1-p|S_p|^p=(1-p)+p|S_p'|^p$\ . \end{enumerate} \end{th2.1} Next we define a generalized Pr\"ufer substitution using $S_p$ and $S_p'$: \begin{equation} y(x)=r (x)S_p(\lambda ^{1/p}\theta (x)), \ \ y'(x)=\lambda ^{1/p} r(x)S_p'(\lambda ^{1/p}\theta (x))\ . \label{eq2.1} \end{equation} By Lemma \ref{th2.1}, one obtains (\cite{LLW08}) \begin{equation} \theta '(x)=1-\frac {q}{\lambda}|S_p(\lambda ^{1/p}\theta (x))|^p\ . \label{eq2.2} \end{equation} \newtheorem{th3.1}[th2.1]{Theorem} \begin{th3.1} \label{th3.1} In the periodic/antiperiodic eigenvalue problem, if $q\in L^{1}(0,1)$ be periodic of period $1$, then $$ q(x)=\lim _{n\rightarrow \infty}p\lambda _n\left( \frac{\lambda _n^{1/p}\ell_j^{(n)}}{\widehat{\pi}}-1\right)\ , $$ pointwisely a.e. and in $L^{1}(0,1)$, where $j=j_n(x)=\max\{k:x_k^{(n)}\leq x\}$. \end{th3.1} The proof below works for both even and odd $n$'s, i.e.\ for both periodic and antiperiodic problems. Some of the arguments above are motivated by \cite{CW07}. See also \cite{LSY1999}. \begin{proof} First, integrating (\ref{eq2.2}) from $x_{k}^{(n)}$ to $x_{k+1}^{(n)}$ with $\la=\la_n$, we have \begin{eqnarray*}\frac{\widehat{\pi}}{\lambda_{n}^{1/p}}&=&\ell_{k}^{(n)}-\int_{x_{k}^{(n)}}^{x_{k+1}^{(n)}}\frac{q(t)}{\la_{n}}|S_p(\lambda_{n}^{1/p}\theta(t))|^{p}dt\ ,\\ &=&\ell_{k}^{(n)}-\frac{1}{p\la_{n}}\int_{x_{k}^{(n)}}^{x_{k+1}^{(n)}}q(t)dt-\frac{1}{\la_{n}}\int_{x_{k}^{(n)}}^{x_{k+1}^{(n)}}q(t)(|S_p(\lambda_{n}^{1/p}\theta(t))|^{p}-\frac{1}{p})dt\ . \end{eqnarray*} Hence, \begin{equation} \ell_{k}^{(n)}=\frac{\widehat{\pi}}{\lambda_{n}^{1/p}} +\frac{1}{p\la_{n}}\int_{x_{k}^{(n)}}^{x_{k+1}^{(n)}}q(t)dt+\frac{1}{\la_{n}}\int_{x_{k}^{(n)}}^{ x_{k+1}^{(n)}}q(t)(|S_p(\lambda_{n}^{1/p}\theta(t))|^{p}-\frac{1}{p})dt\ . \end{equation} and \begin{equation} p\lambda_n\left( \frac{\lambda _n^{1/p}\ell_k^{(n)}}{\widehat{\pi}}-1\right)=\frac{\lambda _n^{1/p}}{ \widehat{\pi}}\int_{x_{k}^{(n)}}^{x_{k+1}^{(n)}}q(t)dt+\frac{p\lambda _n^{1/p}}{\widehat{\pi}} \int_{x_{k}^{(n)}}^{x_{k+1}^{(n)}}q(t)(|S_p(\lambda_{n}^{1/p}\theta(t))|^{p}-\frac{1}{p})dt\ \label{eq2.3} .\end{equation} Now, for $x\in (0,1)$, let $j=j_{n}(x)=\max\{k : x_{k}^{(n)}\leq x \}$. Then $x\in [x_{j}^{(n)},x_{j+1}^{(n)})$ and, for large $n$, $$[x_{j}^{(n)},x_{j+1}^{(n)}) \subset B(x,\frac{2\widehat{\pi}}{\lambda_{n}^{1/p}})\ ,$$ where $B(t,\varepsilon)$ is the open ball centering $t$ with radius $\varepsilon$. That is, the sequence of intervals $\{[x_{j}^{(n)},x_{j+1}^{(n)}) : n \mbox{ is sufficiently large}\}$ shrinks to $x$ nicely (cf. Rudin \cite[p.140]{Rudin}). Since $q\in L^{1}(0,1)$ and $\frac{\lambda _n^{1/p}\ell_{k}^{(n)}}{\widehat{\pi}}=1+o(1)$, we have $$h_{n}(x)\equiv\frac{\lambda _n^{1/p}}{\widehat{\pi}}\int_{x_{j}^{(n)}}^{x_{j+1}^{(n)}}q(t)dt = \frac{\lambda _n^{1/p}\ell_{j}^{(n)}}{\widehat{\pi}} \frac{1}{\ell_{j}^{(n)}}\int_{x_{j}^{(n)}}^{x_{j+1}^{(n)}}q(t)dt $$ converges to $q(x)$ pointwisely a.e. $x\in (0,1)$. Furthermore, since $$|h_{n}(x)|\leq \frac{\lambda _n^{1/p}}{\widehat{\pi}}\int_{x_{j}^{(n)}}^{x_{j+1}^{(n)}}|q(t)|dt\equiv g_{n}(x)\ ,$$ and $$\int_{0}^{1}g_{n}(t)dt=\sum_{k=0}^{n-1}\frac{\lambda _n^{1/p}\ell_{k}^{(n)}}{\widehat{\pi}}\int_{x_{k}^{(n)}}^{x_{k+1}^{(n)}}|q(t)|dt=(1+o(1))\|q\|_{1}\ , $$ we have $ h_{n}(t)\to q(t)$ in $L^{1}(0,1)$ by Lebesgue dominated convergence theorem. On the other hand, let $q_{k,n}\equiv\frac{1}{\ell_{k}^{(n)}}\int_{x_{k}^{(n)}}^{x_{k+1}^{(n)}}q(t)dt$. Then $q_{j,n}$ converges to $q$ pointwisely a.e. $x\in (0,1)$. Let $\phi_{n}(t)=|S_p(\lambda_{n}^{1/p}\theta(t))|^{p}-\frac{1}{p}$. Then \begin{eqnarray*} T_{n}(x)&\equiv&\frac{p\lambda _n^{1/p}}{\widehat{\pi}}\int_{x_{j}^{(n)}}^{x_{j+1}^{(n)}}q(t)\phi_{n}(t) dt\ ,\\ &=&\frac{p\lambda _n^{1/p}}{\widehat{\pi}}\int_{x_{j}^{(n)}}^{x_{j+1}^{(n)}}(q(t)-q_{j,n})\phi_{n}(t)dt +\frac{p\lambda _n^{1/p}}{\widehat{\pi}}\int_{x_{j}^{(n)}}^{x_{j+1}^{(n)}}q_{j,n}\phi_{n}(t)dt\ ,\\ &\equiv&A_{n}+B_{n}\ . \end{eqnarray*} By Lemma \ref{th2.1}(b) and (\ref{eq2.2}), \begin{eqnarray*}B_{n} &=&\frac{p\lambda _n^{1/p}q_{j,n}}{\widehat{\pi}}\int_{x_{j}^{(n)}}^{x_{j+1}^{(n)}}\left( |S_p(\lambda_{n}^{1/p}\theta(t))|^{p}-\frac{1}{p} \right)\left(\theta'(t)+ \frac{q(t)}{\lambda_{n}}|S_p(\lambda_{n}^{1/p}\theta (t))|^p\, \right)\, dt ,\\ &=&-\left.\frac{p q_{j,n}}{\widehat{\pi}}S_p(\la_n^{1/p}\th(t)) S_p'(\la_{n}^{1/p}\th(t))^{(p-1)}\right|_{x_{j}^{(n)}}^{x_{j+1}^{(n)}} +O(\la_{n}^{-1+1/p})\ ,\\ &=&O(\la_{n}^{-1+1/p})\ . \end{eqnarray*} Also, \begin{eqnarray*}|A_{n}|&\leq& \frac{p\lambda_n^{1/p}}{\widehat{\pi}}\int_{x_{j}^{(n)}}^{x_{j+1}^{(n)}} |q(t)-q_{j,n}|||S_p(\lambda_{n}^{1/p}\theta(t))|^{p}-\frac{1}{p}|dt\ ,\\ &\leq& \frac{(p-1)\lambda_n^{1/p}}{\widehat{\pi}}\int_{x_{j}^{(n)}}^{x_{j+1}^{(n)}}|q(t)-q_{j,n}|dt \ , \end{eqnarray*} which converges to $0$ pointwisely a.e.\ $x\in (0,1)$ because the sequence of intervals $\{[x_{j}^{(n)},x_{j+1}^{(n)}) : n \mbox{ is sufficiently large}\}$ shrinks to $x$ nicely. We conclude that $T_{n}(x)\to 0$ a.e. $x\in (0,1)$. Finally, applying Lebesgue dominated convergence theorem as above, $T_{n}(x)\to 0$ in $L^{1}(0,1)$. Hence the left hand side of (\ref{eq2.3}) converges to $q$ pointwisely a.e. and in $L^{1}(0,1)$. \end{proof} \vskip0.1in \begin{proof}[Proof of Theorem \ref{th1.2}] \hfill By the eigenvalue estimates (\ref{eq1.4}) and (\ref{eq1.5}), we have \begin{equation} p\lambda_{2n}(\frac{\lambda_{2n}^{1/p}\ell_{j_{2n}(x)}^{(2n)}}{\widehat{\pi}}-1) =p(2n\widehat{\pi})^p(2n\ell_j^{(2n)}-1)+2n\ell_{j_{2n}(x)}^{(2n)}\int_0^1q(t)dt+o(1)\ .\label{eq3.3} \end{equation} Hence by Theorem \ref{th3.1} and the fact that $2n\ell_j^{(2n)}=1+o(1)$, $$ F_{2n}(x)\equiv p(2n\widehat{\pi})^p(2n\ell_j^{(2n)}-1)+\int_0^1q(t)dt$$ also converges to $q$ pointwisely a.e. and in $L^{1}(0,1)$. The proof for (b) is the same. \end{proof} \begin{proof}[Proof of Theorem \ref{th1.3}] \hfill Here we only give the proof of (b). First, since all anti-periodic eigenvalues include $\{((2n-1)\widehat{\pi})^p:n\in \mathbb{N}\}$, we have, by (\ref{eq1.5}), $\int_0^1q(t)dt=0$. Moreover, $S_p(\widehat{\pi}x)$ satisfies anti-periodic boundary conditions. So by Lemma \ref{th2.1}(b), $$ \int_0^1|S_p'(\widehat{\pi}t)|^pdt-\frac{p-1}{p} = \int_0^1q(t)|S_p(\widehat{\pi}t)|^pdt= \int_0^1|S_p(\widehat{\pi}t)|^pdt-\frac{1}{p}=0\ . $$ Hence, by the variational principle, we have $$ \widehat{\pi}^p=\lambda_1 \leq \frac{\int_0^1\widehat{\pi}^p|S_p'(\widehat{\pi}t)|^pdt+(p-1)\int_0^1q(t)|S_p(\widehat{\pi}t)|^pdt} {(p-1)\int_0^1|S_p(\widehat{\pi}t)|^pdt}= \widehat{\pi}^p\ . $$ This implies $S_p(\widehat{\pi}x)$ is the first eigenfunction. Therefore $q=0$ on $[0,1]$. \end{proof} \section{Linear separated boundary conditions} \setcounter{equation}{0} \hskip0.25in Consider the one-dimensional $p$-Laplacian with linear separated boundary conditions \begin{equation} \left\{ \begin{array}{l} y(0)S_p'(\al)+y'(0)S_p(\al) = 0\\ y(1) S_p'(\be)+y'(1)S_p(\be)= 0 \end{array} \right.\ ,\label{eq4.1} \end{equation} where $\al,\be\in [0,\widehat{\pi})$. Letting $\la_n$ be the $n$th eigenvalue whose associated eigenfunction has exactly $n-1$ zeros in $(0,1)$, the generalized phase $\th_n$ as given in (\ref{eq2.2}) satisfies \begin{equation} \th_n(0)=\frac{-1}{\la_n^{1/p}}\wct^{-1}(-\frac{\wct(\al)}{\la_n^{1/p}});\qquad \th_n(1)=\frac{1}{\la_n^{1/p}}\left( n\widehat{\pi}- \wct^{-1}(-\frac{\wct(\be)}{\la_n^{1/p}})\right)\ ,\label{eq4.8} \end{equation} where the function $\ct(\ga):=\frac{S_p(\ga)}{S_p'(\ga)}$ is an analogue of cotangent function, while $\wct(\ga):=\ct(\ga)$ if $\ga\neq 0$; and $\wct(\ga):=0$ otherwise. Also $\wct^{-1}$ stands for the inverse of $\wct$, taking values only in $[0,\widehat{\pi})$. \par Let $\phi_{n}(x)=|S_p(\lambda_{n}^{1/p}\th_n(x)|^p-\frac{1}{p}$, where . Below we shall state a general Riemann-Lebesgue lemma, which shows that $\int_0^1\!\phi_n g\rightarrow 0$ for any $g\in L^1(0,1)$, when $\la_n$'s are associated with a certain linear separated boundary conditions. In the case of periodic boundary conditions, Brown and Eastham \cite{BE08} used a Fourier series expansion of $\phi_{n}$ where $\phi_{n}(\la_n^{1/p}\th_n(x))\approx \phi_{n}(\al+2n\widehat{\pi} x)$ and apply Plancherel Theorem to show convergence. \newtheorem{th2.2}{Lemma}[section] \begin{th2.2} \label{th2.2} Let $f_{n}$ be uniformly bounded and integrable on $(0,1)$. Suppose for each $n$, there exists a partition $\{ x_0^n=0<x_1^n<\cdots <x_n^n=1\}$ such that $\Delta x_k^n=o(1)$, and $F_k^n(x):=\int_{x_k^n}^x f_{n}(t)\, dt$ satisfies $F_k^n(x)=O(\frac{1}{n})$ for $x\in (x_{k}^{n},x_{k+1}^{n})$ and $F_k^n(x_{k+1}^n)=o(\frac{1}{n})$ uniformly in $k=1,\ldots,n-2$, as $n\to\infty$. Then for any $g\in L^1(0,1)$, $\int_0^1 gf_{n}\rightarrow 0$ as $n\to\infty$. \end{th2.2} \begin{proof} Take any $\ep>0$, there is a $C^1$ function $\tilde{g}$ on $[0,1]$ such that $\int_0^1|\tilde{g}-g|<\ep$. Let $|f_{n}|, |\tilde{g}|\leq M$. Then $$ \int_0^1 g f_{n} =\int_0^1 (g-\tilde{g})f_{n} +\int_0^1 \tilde{g} f_{n}, $$ where $|\int_0^1 (g-\tilde{g})f_{n}|\leq M\ep$. Also $$ \int_0^1 \tilde{g}f_{n}=\sum_{k=0}^{n-1}\int_{x_k^n}^{x_{k+1}^n} \tilde{g}f_{n}=\sum_{k=1}^{n-2} \left(\tilde{g}(x_{k+1}^n)F(x_{k+1}^n)-\int_{x_k^n}^{x_{k+1}^n} \tilde{g}' F_k^n \right)+o(1), $$ where $$ |\int_{x_k^n}^{x_{k+1}^n} \tilde{g}' F_k^n|=O(\frac{1}{n})\, \int_{x_k^n}^{x_{k+1}^n} |\tilde{g}'|=o(\frac{1}{n}). $$ Therefore $\int_0^1 \tilde{g}f_{n}=o(1)$ as $n\to\infty$. \end{proof} \newtheorem{th2.3}[th2.2]{Corollary} \begin{th2.3} \label{th2.3} Consider the $p$-Laplacian (\ref{eq1.1}) with boundary conditions (\ref{eq4.1}). Define $\phi_{n}(x)=|S_p(\lambda_{n}^{1/p}\th_n(x))|^p-\frac{1}{p}$, then for any $g\in L^1(0,1)$, $\int_0^1 \phi_{n}g\rightarrow 0$. \end{th2.3} \begin{proof} Since $\th_n(0)$ and $\th_n(1)$ are as given in (\ref{eq4.8}), $\phi_{n}$ is uniformly bounded on $[0,1]$. Take $x_k^n$ be such that $\th(x_k^n)=\frac{k\widehat{\pi}}{\lambda_{n}^{1/p}}$. Also by integrating the phase equation (\ref{eq2.2}), $\la_{n}^{1/p}=O(n)$, and $$ \Delta x_n=O(\frac{1}{\la_{n}^{1/p}})=O(\frac{1}{n}). $$ Hence by Lemma \ref{th2.1}(b) and (\ref{th2.2}), we have for $k=1,\ldots,n-2$, \begin{eqnarray*} \int_{x_k^n}^{x_{k+1}^n} \phi_{n}(x)\, dx &=& \frac{-1}{p\la_n^{1/p}}\int_{x_k^n}^{x_{k+1}^n} \frac{1}{\th'_n(x)}\, \frac{d}{dx}\left[ S_p(\la_n^{1/p}\th_n(x))S_p'(\la_n^{1/p}\th_n(x))^{(p-1)}\right]\, dx\ ,\\ &=& \frac{-1}{p\la_n^{1/p}} \left[ S_p(\la_n^{1/p}\th_n(x))S_p'(\la_n^{1/p}\th_n(x))^{(p-1)}\right]_{x_k^n}^{x_{k+1}^n} +O(\frac{1}{\la_{n}})\ ,\\ &=&O(\frac{1}{\la_{n}})=o(\frac{1}{n})\ , \end{eqnarray*} since $S_p(k\widehat{\pi})=0$. It is also clear that $\int_{x_k^n}^{x} \phi_{n}(x)\, dx =O(\frac{1}{n})$. Thus we may apply Lemma~\ref{th2.2} to complete the proof. \end{proof} \newtheorem{th4.1}[th2.2]{Theorem} \begin{th4.1} \label{th4.1} When $q\in L^1(0,1)$, the eigenvalues $\lambda_n$ of the Dirichlet $p$-Laplacian (\ref{eq1.1}) satisfies, as $n\rightarrow \infty$, \begin{equation} \lambda _n^{1/p} = n{\widehat \pi} + \frac{1}{p(n{\widehat \pi})^{p-1}} \int_0^1 q(t)dt+o(\frac{1}{n^{p-1}})\ .\label{eq4.3} \end{equation} Furthermore, $F_n$ converges to $q$ pointwisely and in $L^1(0,1)$, where $$F_n(x):=p(n\widehat{\pi} )^p(n\ell_j^{(n)}-1)+\int _0^1 q(t)\, dt. $$ \end{th4.1} \begin{proof} Integrating (\ref{eq2.2}) from $0$ to $1$, we have \begin{eqnarray*} \la_{n}^{1/p}&=&n\widehat{\pi} +\frac{1}{p\la_{n}^{1-1/p}}\int_{0}^{1}q(t)|S_p(\la_{n}^{1/p}\theta (t))|^{p}dt\ ,\\ &=& n\widehat{\pi}+\frac{1}{p\la_{n}^{1-1/p}}\int_{0}^{1}q(t)dt +\frac{1}{p\la_{n}^{1-1/p}}\int_{0}^{1}q(t)(|S_p(\la_{n}^{1/p}\theta (t))|^{p}-\frac{1}{p})dt\ . \end{eqnarray*} Then by Corollary \ref{th2.3}, we have $$ \int_{0}^{1}q(t)(|S_p(\la_{n}^{1/p}\theta (t))|^{p}-\frac{1}{p})dt=o(1)\ , $$ for any $q\in L^{1}(0,1)$. Hence (\ref{eq4.3}) holds. Furthermore, by Theorem \ref{th3.1}, we can obtain the reconstruction formula with pointwise and $L^{1}$ convergence. \end{proof} \noindent {\bf Remark.} In the same way, the Ambarzumyan Theorems for Neumann as well as Dirichlet boundary conditions as given in \cite[Theorems 1.3 and 5.1]{LLW08} can also be extended to work for $L^1$ potentials. On the other hand, for general linear separated boundary problems (\ref{eq4.1}), \begin{equation} \lambda_n^{1/p}=n_{\al\be}\widehat{\pi}+\frac{(\wct(\be))^{(p-1)}-(\wct(\al))^{(p-1)}}{(n_{\al\be}\widehat{\pi})^{p-1}} +\frac{1}{p(n_{\al\be}\widehat{\pi})^{p-1}}\int _0^1q(x)\, dx+o(\frac{1}{n^{p-1}}), \label{eq4.11} \end{equation} where $$ n_{\al\be}=\left\{ \begin{array}{ll} n & \mbox{if }\al=\be=0\\ n-1/2 & \mbox{if }\al>0=\be \mbox{ or } \be>0=\al \\ n-1 & \al,\be>0 \end{array} \right. $$ This is because, after an integration of (\ref{eq2.2}), \begin{equation} \th_n(1)-\th_n(0)=1-\frac{1}{\la_n}\int_0^1 q(x)|S_p(\la_n^{1/p}\th(x))|^p\, dx+o(\frac{1}{\la_n}).\label{eq4.9} \end{equation} By (\ref{eq4.8}), if $\al=0$, then $\th_n(0)=0$. Similarly $\th_n(1)=0$ if $\beta=0$. Now, let $y=CT_{p}^{-1}(x)$. Then $x=CT_{p}(y)$ and hence $$y'= -\frac{1/x^{2}}{1+\frac{1}{|x|^{p}}} =\frac{-|x|^{p-2}}{1+|x|^{p}}=-|x|^{p-2}(1+O(|x|^{p}), $$ when $|x|$ is sufficiently small. Since $y(0)=\frac{\wpi}{2}$, we have $$y(x)=\frac{\wpi}{2} - \frac{x^{(p-1)}}{p-1}+O(x^{2p-1})\ .$$ Therefore, when $n$ is sufficiently large, $$ \th_n(0)=\frac{\wpi}{2\la_n^{1/p}}+\frac{(\ct(\al))^{(p-1)}}{(p-1)\la_n^{(p-1)/p}}+O(\la_{n}^{\frac{1-2p}{p}}). $$ Similarly, when $\be\neq 0$, $$ \th_n(1)=\frac{(n-\frac{1}{2})\wpi}{\la_n^{1/p}}+\frac{(\ct(\be))^{(p-1)}}{(p-1)\la_n^{(p-1)/p}}+O(\la_{n}^{\frac{1-2p}{p}}). $$ Hence (\ref{eq4.11}) is valid. Furthermore, $F_n$ converges to $q$ pointwisely and in $L^1(0,1)$, where $$ F_n(x):=p(n_{\al\be}\widehat{\pi})^p \left[ (n_{\al\be}+\frac{(\wct(\be))^{(p-1)}-(\wct(\al))^{(p-1)}}{(n_{\al\be}\widehat{\pi})^{p-1}} )\ell_j^{(n)}-1\right]+\int _0^1 q(t)\, dt. $$ \section*{Acknowledgments} \hskip0.25in The authors are supported in part by National Science Council, Taiwan under contract numbers NSC 98-2115-M-110-006, NSC 97-2115-M-005-MY2 and NSC 97-2115-M-022-001.

https://arxiv.org/abs/2302.08254

Optimal uniform bounds for competing variational elliptic systems with variable coefficients

Let $\Omega \subset \mathbb{R}^N$ be an open set. In this work we consider solutions of the following gradient elliptic system \[ -\text{div}(A(x)\nabla u_{i,\beta}) = f_i(x,u_{i,\beta}) + a(x)\beta |u_{i, \beta}|^{\gamma -1}u_{i, \beta}\mathop{\sum_{j=1}^l}_{j\neq i} |u_{j, \beta}|^{\gamma + 1}, \] for $i=1,\ldots, l$. We work in the competitive case, namely $\beta<0$. Under suitable assumptions on $A$, $a$, $f_i$ and on the exponent $\gamma$, we prove that uniform $L^\infty$-bounds on families of positive solutions $\{u_\beta\}_{\beta<0}=\{(u_{1,\beta},\ldots, u_{l,\beta})\}_{\beta<0}$ imply uniform Lipschitz bounds (which are optimal).One of the main points in the proof are suitable generalizations of Almgren's and Alt-Caffarelli-Friedman's monotonicity formulas for solutions of such systems. Our work generalizes previous results, where the case $A(x)=Id$ (i.e. the operator is the Laplacian) was treated.

\section{Introduction} \label{chapter:introduction} \subsection{Statement of the main result} Let $\Omega \subset \mathbb{R}^N$ be an open set, $N\geq 1$. Consider $u_\beta = (u_{1,\beta},...,u_{l,\beta})$, a solution of the variational system of equations given by: \begin{align} \label{equation} &-\div(A(x)\nabla u_{i,\beta}) = f_i(x,u_{i,\beta}) + a(x)\beta |u_{i, \beta}|^{\gamma -1}u_{i, \beta} \mathop{\sum_{j=1}^l}_{j\neq i} |u_{j, \beta}|^{\gamma + 1} \end{align} for all $i=1,...,l$, where $\beta < 0$, $\gamma \geq 1$ and $x \in \Omega$. Under natural assumptions on $A$, $a$ and $f_i$, in this paper we obtain uniform optimal bounds in $\beta$ for classes of solutions $\{u_\beta\}_{\beta<0}$. \smallbreak More precisely, we make the following assumptions. For the matrix $A(x)$: \begin{itemize} \item[\textbf{(A1)}] There exists $\theta >0$ such that: \begin{equation*} \langle A(x) \xi , \xi\rangle> \theta |\xi|^2 \quad \quad \forall x \in \Omega, \xi \in \mathbb{R}^N. \end{equation*} \item[\textbf{(A2)}] $A(\cdot) \in C^{0,1}(\Omega, \text{Sym}^{N\times N})$, and \begin{align*} \sup_{x \in \Omega}\|A(x)\| \leq M, \quad\quad \sup_{x \in \Omega}\|DA(x)\| \leq M. \end{align*} \end{itemize} For the functions $f_i$, we assume that: \begin{itemize} \item[\textbf{(F)}] $f_i(\cdot,\cdot) \in C(\Omega\times \mathbb{R})$, and \end{itemize} \begin{equation*} \sup_{x \in \Omega}|f_i(x,s)| = O(s),\quad \text{ as $s \rightarrow 0$ for all $i=1,...,l$. } \end{equation*} Finally, we make the following assumption on the function $a(x)$: \begin{itemize} \item[\textbf{(a)}] $a(\cdot) \in C^1(\Omega)$ and there exists $\delta>0$ such that: \begin{equation*} a(x)>\delta>0 \quad \quad \forall x \in \Omega. \end{equation*} \end{itemize} Our main result reads as follows. \begin{thm} \label{DesiredTheorem} Under the previous assumptions on $A$, $a$, $f_i$, assume moreover that $\frac{\gamma N}{\gamma + 1} < 2$. Let $\{u_\beta\}_{\beta < 0}$ be a family of positive solutions to the system \eqref{equation} such that \begin{equation} \label{boundedness} \text{there exists $m>0$ such that } \sup_{\beta <0} \|u_\beta\|_{L^\infty(\Omega)} \leq m. \end{equation} Then, given $K \Subset \Omega$, there exists a constant $C>0$ such that \begin{equation}\label{eq:Lipbounds} \sup_{\beta <0} \|\nabla u_{i,\beta}\|_{L^\infty(K)}\leq C. \qquad \text{ for all $i \in \{1,...,l\}$.} \end{equation} \end{thm} \noindent To see the dependencies of the constant $C$ appearing in \eqref{eq:Lipbounds}, see Remark \ref{rem:dependence_of_constant} below. For a direct consequence in the framework of elliptic systems in Riemannian manifolds, see Corollary \ref{cor_Riem}. \smallbreak In the remainder of this introduction, we give background for this result, explain its proof and provide the structure of the paper. \subsection{Background} Systems of type \eqref{equation} have been widely considered in the literature in the case $A(x)=Id$ and $a(x)=1$, when the system reads \begin{equation}\label{equation:Laplacian} -\Delta u_{i,\beta} = f_i(x,u_{i,\beta}) + \beta |u_{i, \beta}|^{\gamma -1}u_{i, \beta}\mathop{\sum_{j=1}^l}_{j\neq i} |u_{j, \beta}|^{\gamma + 1}. \end{equation} From a physical point of view, these systems arise naturally when looking for standing wave solutions of associated systems of Gross-Pitaevskii/nonlinear Schr\"odinger equations. The later model important phenomena in Nonlinear Optics \cite{AkAn} and Bose-Einstein condensation \cite{Rogel_Salazar_2013, Timmermans}. In the models, the solutions are the corresponding condensate amplitudes, the term $f_i(x,u_{i,\beta})$ regulates self-interactions within the same component, while $\beta$ expresses the strength and the type of interaction between different components $i$ and $j$. When $\beta>0$ this represents cooperation, while $\beta<0$ represents competition. In the important case $f_i(x,u_{i,\beta})=|u_{i,\beta}|^{2\gamma}u_{i,\beta}$, starting from \cite{LinWei}, there is a vast literature regarding existence, multiplicity and classification of solutions to \eqref{equation:Laplacian}; we simply refer to the papers \cite{BartschDancerWang, ClappPistoia2022, ClappSzulkin2019, CorreiaJDE2016,CorreiaNA2016,DancerWeiWeth, Mandel,OliveiraTavares,PengWangWang2019,Soave,SoaveTavares,TianWang, WeiWu} (in the subcritical case) and \cite{ChenLinZou,ChenZouARMA2012, ChenZou2, ClappPistoia2018,PengPengWang2016,TavaresYou,TavaresYouZou, YinZou} (in the critical case) for more details and to check other references. \smallbreak In many situations (see for instance \cite{BartschDancerWang,ChenZouARMA2012,ChenZou2,ClappPistoia2018,DancerWeiWeth, TavaresYou,TavaresYouZou, TianWang}), one can build, using variational methods, families of solutions $\{u_\beta\}_{\beta<0}$ which have uniform bounds in $L^\infty(\Omega)$, namely that satisfy \eqref{boundedness}. It is therefore a natural question to understand what is the asymptotic behaviour of such solutions as $\beta\to -\infty$, what are the optimal bounds, and how to characterize the limiting profiles. This was done in \cite{uniformHolderBoundsHugoTerraciniNoris} for $\gamma=1$ (see also \cite{CaffarelliLin,CLLL}). Using the same strategy, the general case of system \eqref{equation:Laplacian} was done in the survey paper \cite{SOAVE2016388}. \begin{theorem_a}[{{\cite[Theorems 1.2 \& 1.5]{SOAVE2016388}}}]\label{theoremA} Take $\gamma>0$ and $f_i$ satisfying \textbf{(F)}. Let $\{u_\beta\}_{\beta<0}$ be a family of solutions of \eqref{equation:Laplacian} satisfying the uniform $L^\infty$--bound \eqref{boundedness}. Then, for every $K \Subset \Omega$ and $\alpha\in (0,1)$, there exists $C>0$ such that \begin{equation}\label{eq:Holder1} \sup_{\beta<0}\| u_\beta\|_{C^{0,\alpha}(K)}\leq C. \end{equation} In particular, there exists a limiting function $u=(u_1,\ldots, u_l)$, where each $u_i$ is Lipschitz continuous in $\Omega$, such that, up to a subsequence, \begin{enumerate} \item $u_\beta \to u$ strongly in $H^1_{loc}(\Omega)\cap C_{loc}^{0,\alpha}(\Omega)$ for every $0<\alpha<1$ and, for every compact $K \Subset \Omega$, we have \[ \beta \int_{K} |u_{i,\beta}|^{p+1} |u_{j,\beta}|^{p+1} \to 0\qquad \text{ as $\beta \to \infty$, whenever $i\neq j$;} \] \item $u_i u_j\equiv 0$ whenever $i\neq j$, and $-\Delta u_i=f_{i}(x,u)$ in the open set $\{|u_i|>0\}$. \end{enumerate} \end{theorem_a} Moreover, for any limiting profile $u=(u_1,\ldots,u_l)$ as in the previous theorem, by \cite[Theorems 1.1 \& 8.1]{HugoTerraciniWeakReflectionLaw}, one deduces the structure of the free boundary $\{u=0\}$: it is, up to a set of Hausdorff dimension at most $N-2$, a regular hypersurface. Theorefore, by this regularity result, Theorem A-(2) and Hopf's lemma, one concludes that Lipschitz regularity is optimal for the limiting profiles $u$ (the gradient has a jump on the regular part of the free boundary $\{u=0\}$). The next natural question is whether one can obtain uniform Lipschitz bounds for $L^\infty$-bounded sequences of solutions $\{u_\beta\}_{\beta<0}$. This was positively answered by Soave and Zilio in \cite{SoaveZilio}. \begin{theorem_b}[{{\cite[Theorem 1.3]{SoaveZilio}}}]\label{theoremB} Let $\gamma,N\geq 1$ be such that $\frac{\gamma N}{\gamma + 1} \leq 2$, and take $f_i$ satisfying \textbf{(F)}. Let $\{u_\beta\}_\beta$ be a family of solutions of \eqref{equation:Laplacian} satisfying the uniform $L^\infty$--bound \eqref{boundedness}. Then, for every $K \Subset \Omega$, there exists $C>0$ such that \begin{equation}\label{eq:Holder2} \sup_{\beta<0} \| u_\beta\|_{C^{0,1}(K)}\leq C. \end{equation} \end{theorem_b} Therefore, our Theorem \ref{DesiredTheorem} is an extension of this result to the framework of systems of type \eqref{equation}. We explain in Subsection \ref{subsec:structure} which are the main difficulties one faces when passing from the case of the Laplacian operator to a divergence operator with variable coefficients. \smallbreak Observe that passing from H\"older to Lipschitz bounds is a nontrivial task. The proofs of H\"older bounds are based, among other things, on the fact that there exist no harmonic functions, apart from the constants, which have bounded $\alpha$-H\"older seminorm in $\mathbb{R}^N$ for some $\alpha \in ]0,1[$. The proof of \eqref{eq:Holder1} proceeds by contradiction and by performing a blowup argument close to the region where one does not have a bound. One then reaches a contradiction, in the end, by studying all possible cases for the \emph{limiting profiles}, excluding them using Liouville type results like the one just stated (within this process, an Almgren's monotonicity formula is proved for the limits). However, this type of proof does not translate to the Lipschitz setting. In \cite{SoaveZilio}, in order to prove \eqref{eq:Holder2}, a contradiction is not obtained at the limit of the blowups, but instead along the blowup sequence; for this, the authors combine in a very nice way an Almgren and an Alt-Caffarelli-Friedman monotonicity formula for rescaled solutions of system \eqref{equation:Laplacian}. \smallbreak Let us also point out that our work is also a natural follow up of the following theorem, obtained in \cite{HugoHolderVariable}, concerning $\alpha$-H\"older bounds ($\alpha \in ]0,1[$) for solutions of the system \eqref{equation}. \begin{theorem_c}[{\cite[Theorem B.1]{HugoHolderVariable}}] \label{theoremForHolder} If $\{u_\beta\}_{\beta <0}$ is a family of solutions of \eqref{equation} satisfying the conditions \textbf{(A1)}, \textbf{(A2)}, \textbf{(F)}, \textbf{(a)} and \eqref{boundedness} then, for each $\alpha \in ]0,1[$ and $K \Subset \Omega$, there exists a constant $C$ such that $$ \sup_{\beta < 0} \| u_{\beta}\|_{C^{0,\alpha}(K)} \leq C. $$ \end{theorem_c} \begin{remark} A generalization of this result also easily follows using the same proof as in \cite{HugoHolderVariable}. Let, for each $\beta<0$, $A_{\beta}(\cdot) \in C^\infty(\Omega, \text{Sym}^{N\times N})$ be a matrix satisfying conditions \textbf{(A1)} and \textbf{(A2)} uniformly in $\beta$, and $f_{i,\beta} \in C^\infty(\Omega,\mathbb{R})$ satisfying \eqref{boundForF} uniformly. Take $v_\beta = (v_{1,\beta},...,v_{l,\beta})$ a positive solution of the system \begin{equation*} -\div( A_\beta(x) \nabla v_{i,\beta} ) = f_{i,\beta} (x,v_{i,\beta}) + a(x) \beta \sum_{j\neq i} |v_{i,\beta}|^{\gamma-1} v_{i,\beta} |v_{j,\beta}|^{\gamma+1} \end{equation*} satisfying $\|v_\beta\|_{L^\infty(\Omega)}<m$ for some $m>0$, Then, for each $\alpha \in ]0,1[$ and $K \Subset \Omega$, there exists $C>0$ depending on $\alpha, K,\gamma,N$ and the constants in the conditions \textbf{(A1)},\textbf{(A2)},\textbf{(F)},\textbf{(a)} such that $$ \sup_{\beta<0} \|u_\beta\|_{C^{0,\alpha}(K)} \leq C. $$ \end{remark} As a matter of fact, as we discuss next, these results that give uniform H\"older bounds are absolutely essential in our arguments to prove Theorem \ref{DesiredTheorem}. They will, in particular, be used in Section \ref{chapter:resultsChap4} to prove an Alt-Caffarelli-Friedman type monotonicity formula (see Lemma \ref{sphereLemma}). This is a key difference between our approach and the one used in \cite{SoaveZilio}. \vspace{2mm} \subsection{Structure of the paper and proof strategy}\label{subsec:structure} We now give a brief description of the structure of this work and of the proof of Theorem \ref{DesiredTheorem}. The proof follows the blowup argument and the scheme found in \cite{SoaveZilio} (where, we recall, Theorem B is proved, which corresponds to Theorem \ref{DesiredTheorem} with $A(x) = Id$, $a(x)$). It relies on two monotonicity formulas, the Almgren monotonicity formula and the Alt-Caffarelli-Friedman monotonicity formula that are proved for blowup sequences (and not only for the blowup limits). Apart from the natural technical issues that arise from the fact that we have a more complicated operator, the main difficulty in our case is how to generalize appropriately these monotonicity formulas to our setting of divergence type operators with variable coefficients. \smallbreak As stated before, the proof is based on a \emph{contradiction argument} using a normalized blowup sequence. It is a blowup done along the points $x_n$ where $\max_{j=1,...,l}|\nabla u_{j,\beta_n}|$ attains its maximum, for a sequence $\beta_n \rightarrow -\infty$. The normalization is done in such a way that the new sequence has bounded Lipschitz seminorm. Section \ref{chapter:background} is devoted to analyzing this blowup sequence and its properties. The results found are generalizations of arguments in \cite{SoaveZilio}, with adaptations for the variable coefficient case. \smallbreak Section \ref{chapter:implementation} is devoted to a generalization of the Almgren monotonicity formula for the variable coefficients case. In the context of limits $\beta\to -\infty$ of solutions to systems \eqref{equation:Laplacian} or to limits of blowup sequences, this formula has been used for instance in \cite{CaffarelliLin,uniformHolderBoundsHugoTerraciniNoris, SOAVE2016388, HugoTerraciniWeakReflectionLaw}; see \cite[Appendixes B \& C]{HugoHolderVariable} for the case of system \eqref{equation}. In the later case, a crucial point is to perform a change of variables, changing (locally) the operator to become a perturbation of the Laplacian (see \eqref{tildeU} below, which is inspired by the previous works \cite{Mariana2,GAROFALO2014682,Kukavica,SoaveWeth}). Here, we generalize \cite{SoaveZilio,SphereDensityAlt} (which deal with the Laplacian case) and prove an Almgren monotonicity formula for blowup \emph{sequences} associated to \eqref{equation}, see Theorem \ref{almgrenMonotonicity} below. Since our objective is to obtain a monotonicity formula for the blowup sequence, and not the limit, there are extra terms that have to be considered. In the article \cite{SoaveZilio}, these terms are circumvented by taking the dimension to satisfy $\frac{\gamma N}{\gamma + 1} \leq 2$ (which implies $N\leq 4$). In our case, we can only obtain a monotonicity formula when the inequality is strict, that is $\frac{\gamma N}{\gamma + 1} < 2$ (which gives $N\leq 3$). This is due to extra terms coming from the variable coefficients, and it is the only place in the paper where the restriction is needed (see the proof of Theorem \ref{almgrenMonotonicity} for the details, in particular inequality \eqref{eq:RESTRICTION}). \smallbreak Section \ref{chapter:resultsChap4} is where we prove Theorem \ref{AltCaffMonotonicity}, which is a generalization of the Alt-Caffarelli-Friedman formula found in \cite[Theorem 3.14]{SoaveZilio}. This is where our work differs the most from previous proofs, and it is one of the main contributions of our paper. Regarding this topic, there are two main problems in working with operators with variable coefficients. Firstly, recall that the core of the proof of the classical Alt-Caffarelli-Friedman formula \cite{FriedmanSphere} is a result about a spectral optimal partition problem on the sphere, which says that: \begin{equation} \label{minimizationOnSphereBasic123} \min \{ \gamma(\lambda_1(\Omega_1)) + \gamma(\lambda_1(\Omega_2)) : \Omega_1, \Omega_2 \subset \partial B_1, \, \Omega_1\cap \Omega_2 = \emptyset \} \geq 2, \end{equation} where $\lambda_1(\Omega)$ is the first Dirichlet eigenvalue of the Laplace-Beltrami operator on the sphere, $\Delta_\theta$, of the set $\Omega$, and $\gamma(t) = \sqrt{(\frac{N-2}{2})^2 + t} - \frac{N-2}{2}$. The proof of \cite[Theorem 3.14]{SoaveZilio} relies on a lower bound of a certain functional defined on the sphere, which is similar to the one found in \cite[Lemma 4.2]{SphereDensityAlt}, but with extra terms to account for the remaining terms in equation \eqref{equation:Laplacian}. Since in these papers (or in \eqref{minimizationOnSphereBasic123}) the functionals are related to the Laplacian, the proofs use a symmetrization argument which simplifies the procedure. This is not possible in our case due to the variable coefficients in our equations. The result in our work, in the form of Lemma \ref{sphereLemma}, even though has the same structure of \cite{SoaveZilio, SphereDensityAlt}, obtains similar bounds through very different approaches. In particular, due to the lack of symmetrization, we cannot conclude the minimizing functions are uniformly Lipschitz, and to circumvent this we use Theorem C to obtain uniform H\"older bounds and make nontrivial use of the equation in a way it is enough for our purposes. Secondly, the other main idea of the classical Alt-Caffarelli-Friedman formula is that (in dimension $N\geq 3$), $|y|^{2-N}$ is a fundamental solution of the Laplacian, that is $-\Delta (|y|^{2-N}) = C\delta$, for some $C$ depending on the dimension $N$. In our case, we are dealing (after a change of variables) with an operator $-\div(\tilde A_n(x)\nabla (\cdot))$, where $\tilde A_n(y) \sim Id$ for $y$ close to the origin. The idea is to approximate this operator by $-\Delta(|y|^{2-N})$ plus an ``error'' term, and then use Almgren's monotonicity formula to bound this error term. This allows an estimate like Lemma \ref{derivativeLemma}, and then a generalization of \eqref{minimizationOnSphereBasic123} in the form of Lemma \ref{sphereLemma}, which is the core of the proof of the monotonicity formula. \smallbreak Section \ref{sec:Lip} contains the proof of Theorem \ref{DesiredTheorem}. In this section, more refined properties of the blowup sequence related to the Almgren monotonicity formula are studied. One also shows that there exists a radius $R_0>0$ such that two components of the limit of blowup sequences are nontrivial in $B_{R_0}$. This nontriviality is used to show that the Alt-Caffarelli-Friedman monotonicity formula can't go to zero in the limit. This is where the Alt-Caffarelli-Friedman and Almgren's monotonicity formulas are combined to obtain a contradiction on the blowup sequence, concluding the proof. Here we follow \cite[Section 4]{SoaveZilio}, but adjustments for the variable coefficients case are (again) needed. \smallbreak Section \ref{sec:conditions} is devoted to proving that the conditions of the Alt-Caffarelli-Friedman formula of Section \ref{chapter:resultsChap4} are satisfied for the blowup sequence. This is based on the characterization of certain limits of certain blowups and blowdowns. For this, we make use of some of the theorems from Section \ref{chapter:implementation}, in particular Almgren's monotonicity formula. In the appendices, we present important results that are used throughout the paper. Appendix \ref{chapter:sphereDivergence} shows a relation between the divergence operator on the $N-1$ dimensional sphere and the divergence in $\mathbb{R}^{N-1}$ through a stereographic projection. Appendix \ref{appendix:classG} makes a quick overview of results for functions that belong to the class $\mathcal{G}(\Omega)$ introduced in \cite{HugoTerraciniWeakReflectionLaw}; this is a set which has a strong relation with blowups of competitive systems. Appendix \ref{chapter:mult1PointsApp} is where one states (based on \cite{MULT1DANCERWANG}) that for limits $\lim_{\beta}u_\beta = v \in \mathcal{G}(\Omega)$ of competition systems like \eqref{equation}, when $\beta \rightarrow -\infty$, and points $x_0$ where $v(x_0) = 0$, then for every neighborhood $V_{x_0}$ of $x_0$ we must have two nontrivial components of $v = (v_1,...,v_l)$. Finally, in Appendix \ref{chapter:AuxResultsApp} we collect other results. \smallbreak We conclude this introduction with one remark and an immediate corollary of Theorem \ref{DesiredTheorem} \begin{remark}\label{rem:dependence_of_constant} We notice that, by \textbf{(F)}, given $m>0$ there exists $d>0$ such that for all $i \in \{1,...,l\}$: $$ \sup_{x\in \Omega, s \in [-m,m]} |f_i(x,s)| \leq d|s|. $$ Combining the observation above with \eqref{boundedness} we obtain the existence of $d >0$ such that \begin{equation} \tag{\textbf{Fd}}\label{boundForF} |f_i(x,u_{i,\beta}(x))| \leq d|u_{i,\beta}(x)| \qquad \text{for all $i \in \{1,...,l\}$.} \end{equation} The constant $C$ in Theorem \ref{DesiredTheorem} depends on the dimension $N$, the exponent $\gamma$, the compact $K$, the ellipticity constant $\theta$ in \textbf{(A1)}, and the upper bounds $M,m$ from \textbf{(M)} and \eqref{boundedness}. \end{remark} Our main result, Theorem \ref{DesiredTheorem}, has a direct correspondence with systems with the Laplace-Beltrami operator defined on Riemannian manifolds. \begin{cor}\label{cor_Riem} Let $(\mathcal{M},g)$ be a $C^1$ Riemannian manifold, and consider $\{u_\beta\}_{\beta < 0}$ a family of positive solutions of the system \begin{equation} \label{systemRiem} -\Delta_g u= f_i(x,u_{i,\beta}) + \beta |u_{i, \beta}|^{\gamma -1}u_{i, \beta} \mathop{\sum_{j=1}^l}_{j\neq i} |u_{j, \beta}|^{\gamma + 1}\qquad \text{ in } \mathcal{M} \end{equation} for $i=1,\ldots, l$, under the assumption \textbf{(F)} for $f_i$, $N,\gamma\geq 1$. Assume moreover that $\frac{\gamma N}{\gamma + 1} < 2$, and that the sequence of solutions is uniformly bounded in $L^\infty$--norm \eqref{boundedness}. Then, given $K \Subset \mathcal{M}$, there exists a constant $C>0$ such that $$ \sup_{\beta <0} \|\nabla u_{i,\beta}\|_{L^\infty(K)}\leq C\qquad \text{ for all $i \in \{1,...,l\}$.} $$ \end{cor} \noindent Indeed, when using local coordinates in a small neighborhood of each point, \eqref{systemRiem} turns into \eqref{equation}, where $A(x)$ and $a(x)$ contain information about the metric $g$. \section{Contradiction argument and blowup sequences} \label{chapter:background} Let us suppose, without loss of generality, that $B_3 \subset \Omega$. Within this section, we work with dimension $N\geq 1$, for $A$ satisfying assumptions \textbf{(A1)} and \textbf{(A1)}, $f$ satisfying \textbf{(F)} and $a$ satisfying \textbf{(a)}, we take $u_\beta$ to be a family of solutions of the system of equations \eqref{equation} that has a uniform $L^\infty(B_3)$ bound: for some $m>0$, $$ \sup_{\beta<0} \|u_{\beta}\|_{L^\infty(B_3)} \leq m. $$ Our goal is to show uniform Lipschitz bounds in $B_1$. Assume, by contradiction, that there exists a sequence $\beta_n \rightarrow - \infty$ such that: \begin{equation*} \sup_{i = 1,...,l}\|\nabla u_{i,\beta_n}\|_{L^\infty(B_1)} \to \infty \end{equation*} In the spirit of \cite{HugoHolderVariable, SOAVE2016388, SoaveZilio, SphereDensityAlt}, since we want to localize the argument, we introduce a smooth cut-off function $0 \leq \eta(x) \leq 1$ such that $\eta(x) = 1$ for $x \in B_1$ and $\eta(x) = 0$ for $x \in \mathbb{R}^N\setminus B_2$ and observe that the contradiction assumption yields \begin{equation} \label{eq:contradictionassumption} L_n := \sup_{i=1,...,l} \sup_{x \in \overline{B_2}} |\nabla (\eta u_{i_n,\beta_n})| \rightarrow \infty \end{equation} as $n \rightarrow \infty$. For each $n \in \mathbb{N}$, there is a point $x_n \in \overline{B}_2$ such that $L_n = |\nabla (\eta u_{i,\beta_n})(x_n)|$ for some $i_n$. We can assume, without loss of generality (by possibly extracting another subsequence and relabelling components), that: \begin{itemize} \item $L_n=|\nabla (\eta u_{1,\beta_n})(x_n)|$, i.e., the maximum is always attained at the first coordinate; \item $x_n \to x_\infty\in \overline{B}_2$. \end{itemize} We define: \begin{equation} \label{tildeU} \tilde{u}_{i,\beta_{n}}(x) := u_{i,\beta_n}(x_n+A(x_n)^{\frac{1}{2}}x). \end{equation} and consider the matrix function given by: \begin{equation} \label{tildeMatrix} \tilde{A}_n(x) := A(x_n)^{-\frac{1}{2}} A(x_n + A(x_n)^{\frac{1}{2}}x) A(x_n)^{-\frac{1}{2}},\qquad \text{ which is such that } \qquad \tilde{A}_n(0)=Id \end{equation} \begin{lemma} \label{lemmaForMatrixTilde} We have: \begin{equation} \label{tildeEquation} -\div(\tilde{A}_n(x)\nabla \tilde{u}_{i,\beta_n}) = f_i(x_n+A(x_n)^{\frac{1}{2}}x,\tilde{u}_{i,\beta_n}) + a(x_n+A(x_n)^{\frac{1}{2}}x) \mathop{\sum_{j=1}^l}_{j\neq i}\beta_n |\tilde{u}_{j, \beta_n}|^{\gamma + 1} |\tilde{u}_{i, \beta_n}|^{\gamma -1}\tilde{u}_{i, \beta_n}, \end{equation} for $x \in A(x_n)^{-\frac{1}{2}}(B_3 - x_n)$, where $\tilde{A}_n$ is the matrix in \eqref{tildeMatrix}. Moreover: \begin{enumerate} \item $B_{{1}/{M^\frac{1}{2}}} \subset A(x_n)^{-\frac{1}{2}}(B_3 - x_n)$, \item $\langle \tilde{A}_n(x)\xi, \xi \rangle \geq \frac{\theta}{M} |\xi|^2$ for every $\xi \in \mathbb{R}^N$, \item there exists $C = C(M, \theta)>0$ such that $\|D\tilde{A}_n\|_{L^\infty} \leq C$ and $\|\tilde{A}_n\|_{L^\infty} \leq C$ . \end{enumerate} \end{lemma} \begin{proof} Equation \eqref{tildeEquation} follows from a straightforward computation. For (1), we start by observing that, by \textbf{(A2)}, we have $|A(x_n)^\frac{1}{2}\xi |^2 =\langle A(x_n)\xi, \xi \rangle \leq M |\xi|^2$ for every $\xi\in {\mathbb R}^N$. Hence, since $x_n\in \overline{B}_2$, we have (1) and $|\xi|^2\leq M|A(x_n)^{-\frac{1}{2}}\xi|^2$. From this last fact and using also \textbf{(A1)}: \[ \langle \tilde{A}_n(x)\xi, \xi \rangle = \langle A(x_n + A(x_n)^{\frac{1}{2}}x) A(x_n)^{-\frac{1}{2}}\xi,A(x_n)^{-\frac{1}{2}} \xi \rangle \geq \theta|A(x_n)^{-\frac{1}{2}}\xi|^2 \geq \frac{\theta}{M}|\xi|^2, \] which is (2). Finally, by hypothesis \textbf{(A2)}, we have: $$ | \partial_{x_k}({A}_{ij}(x_n + A(x_n)^{\frac{1}{2}}x))| = | \langle \nabla A_{ij}(x_n + A(x_n)^{\frac{1}{2}}x), A(x_n)^{\frac{1}{2}}e_k \rangle | \leq \kappa \|DA\|_{L^\infty(\Omega)} \cdot \|A(x_n)^\frac{1}{2}\| \leq \kappa M^\frac{3}{2}, $$ where $\{e_j\}_{j=1,...,N}$ is the canonical basis for $\mathbb{R}^N$. This, combined with the fact that $\|A(x_n)^{-\frac{1}{2}}\|_{L^\infty} \leq \theta^{-\frac{1}{2}}$ and $\tilde A_n(0)=Id$, implies (3). \end{proof} We present a result regarding the limit of $\tilde{u}_{\beta_n}$. This result will be a bit lateral for now, but will be necessary in the proof of Lemma \ref{radGoToZero} below. \begin{lemma} \label{remarkOnStuff} The sequence $\{\tilde{u}_{\beta_n}\}$ is bounded in $C^{0,\alpha}(B_{1/(2M^\frac{1}{2})})$ and there exists $\tilde{u}_\infty \in C^{0,\alpha}(B_{1/(2M^\frac{1}{2})})\cap H^1(B_{1/(2M^\frac{1}{2})})$ such that, up to a subsequence, $\tilde{u}_{\beta_n} \rightarrow \tilde{u}_\infty$ in $C^{0,\alpha}(B_{1/(2M^\frac{1}{2})})\cap H^1(B_{1/(2M^\frac{1}{2})})$ for every $\alpha\in (0,1)$. Moreover, $\tilde{u}_\infty(0) = 0$. \end{lemma} \begin{proof} By results in the first sentence follow directly from Theorem C and reasoning exaclty as in \cite[Theorem 1.4]{uniformHolderBoundsHugoTerraciniNoris} for the strong convergence in $H^1$. Now, if $\tilde{u}_\infty(0)> 0$, then there exist $\epsilon,r C>0$, $i \in \{1,...,l\}$ such that $\tilde{u}_{i,\beta_n}(x)\geq \epsilon$ for $x \in B_r$ and $n$ large enough. We now have two cases. If $i\neq 1$, then by the equation of $\tilde{u}_{1,\beta_n}$ we have $$ -\div(\tilde{A}_n(x)\nabla \tilde{u}_{1,\beta_n}) \leq - |M_n| \delta \epsilon^{\gamma} \tilde{u}_{1,\beta_n}^{\gamma} + dm \qquad \forall x \in B_{r} $$ and so, by Lemma \ref{estimateLemma}-(1), we conclude the existence of $c>0$ such that $|M_n|\tilde{u}_{1,\beta_n}^{\gamma}(x)<c$ for $x \in B_{r/2}$ and so (going back to the equation of $\tilde{u}_{1,\beta_n}$), $|\div(\tilde{A}_n(x)\nabla \tilde{u}_{1,\beta_n})|$ is uniformly bounded in $n$. By elliptic regularity theory, we must have that, up to a subsequence, $\tilde{u}_{1,\beta_n} \rightarrow \tilde{u}_{1,\infty}$ in $C^1(B_{r/2})$, in contradiction with the fact that the gradient $|\nabla \tilde{u}_{1,\beta_n}(0)|$ blows up. On the other hand, if $i = 1$, then for all $j \neq 1$ we can apply the argument above to conclude that there exists $c>0$ such that $|M_n| \tilde{u}_{j,n}(x) < c$ for all $x \in B_{r/2}$, from which we conclude a uniform bound for $|\div(\tilde{A}_n(x)\nabla \tilde{u}_{1,\beta_n})|$. This, once again, leads to a contradiction. \end{proof} Consider now the blowup sequences given by: \begin{align} &v_{i,n}(x) := \eta(x_n) \frac{\tilde{u}_{i,\beta_n}(r_nx)}{L_n r_n} \label{blowUp}\\ & \overline{v}_{i,n}(x) := \frac{(\eta u_{i, \beta_n})(x_n + r_n A(x_n)^{\frac{1}{2}} x)}{L_n r_n},\nonumber \end{align} where we consider each function defined in: \begin{equation} \label{Domain} \Omega_n := \frac{A(x_n)^{-\frac{1}{2}}(B_3 - x_n)}{r_n}, \quad \text{ where} \quad 0< r_n := \sum_{i=1}^l\frac{(\eta u_{i,\beta_n})(x_n)}{L_n}\to 0. \end{equation} The fact that $r_n\to 0$ is a consequence of the bound $r_n\leq lm/L_n \to 0$, using the uniform boundedness from \eqref{boundedness} and the contradiction assumption \eqref{eq:contradictionassumption}. On the other hand, to show that $r_n>0$, notice that, since $|\nabla(\eta u_{1,\beta_n})(x_n)| = L_n>0$, we must have $(\eta u_{1,\beta_n})(x_n)>0$, otherwise around the point $x_n$ the function $(\eta u_{1,\beta_n})$ would take negative values, which is a contradiction. With this choice of $r_n$, we have the normalization \begin{equation} \label{rnDefinition} \sum_{i=1}^l\overline{v}_{i,n}(0) = \sum_{i=1}^l\frac{(\eta u_{i, \beta_n})(x_n)}{L_n r_n} = 1. \end{equation} Define \begin{align*} &A_n(y) = \tilde{A}_n(r_n y) = A(x_n)^{-\frac{1}{2}} A(x_n + r_nA(x_n)^{\frac{1}{2}}y) A(x_n)^{-\frac{1}{2}},\\ &a_n(y) = a(x_n + r_nA(x_n)^{\frac{1}{2}}y),\\ &f_{i,n}(y,t) = \frac{r_n \eta(x_n)}{L_n}f_i(x_n + r_nA(x_n)^{\frac{1}{2}} y, \frac{L_n r_n}{\eta(x_n)} t). \end{align*} \begin{lemma} \label{lemmaMatrixBlowLimit} We have $A_n \in C^\infty(\Omega_n, \text{Sym}^{N\times N})$ and: \begin{align} \label{ellipticConstantForBlowUp} \begin{split} \|A_n(y) - Id\| \leq Cr_n |y|, \quad \|DA_n\|_{L^\infty(B_r)}\leq Cr_n, \quad \langle A_n(y)\xi,\xi \rangle\geq \frac{\theta}{M}|\xi|^2. \end{split} \end{align} Also $a_n(\cdot) \rightarrow a(x_\infty)$ locally uniformly in $\Omega_n$, $f_{i,n}(x,v_{i,n}(\cdot)) \rightarrow 0$ uniformly in $\Omega_n$, and $|f_{i,n}(x,v_{i,n}(x))|\leq dr_n^2|v_{i,n}(x)|$. \end{lemma} \begin{proof}The first three inequalities follow by Lemma \ref{lemmaForMatrixTilde} and by \textbf{(A2)}. Let $K \Subset \Omega_n$ be a compact set. Using \textbf{(a)}, we obtain the existence of $C>0$ also such that: $$ |a_n(y)-a(x_n)| \leq C r_n|y| $$ for all $y \in K$. Thus $a_n \rightarrow a(x_\infty) = \lim_n a(x_n)$ for each compact set in $\Omega_n$. For the last information about $f_n$, using \eqref{boundForF} we obtain: \begin{align*} |f_{i,n}(x,v_{i,n}(x))| = & |\frac{r_n \eta(x_n)}{L_n}f(x_n + r_nA(x_n)^{\frac{1}{2}} x, u_{i,n}(x_n + r_nA(x_n)^{\frac{1}{2}} x))|\\ \leq & d\frac{r_n \eta(x_n)}{L_n}|u_{i,n}(x_n + r_nA(x_n)^{\frac{1}{2}} x)| = dr_n^2 |v_{i,n}(x)|; \qedhere \end{align*} the uniform convergence to 0 is now a consequence of the fact that $r_n\to 0$, $L_n\to 0$ and the uniform bounds for $u_\beta$. \end{proof} \begin{prop} \label{ListProp} The following are satisfied: \begin{enumerate} \item We have that $ B_{1/(M^{\frac{1}{2}} r_n)} \subset \Omega_n, $ so $\Omega_n$ exhaust the whole $\mathbb{R}^N$ as $n\to \infty$. \item The sequence $v_{i,n}$ satisfies the equation: \begin{equation}\label{eq:system_rescaled} -\div(A_n(y)\nabla v_{i,n}) = f_{i,n}(y, v_{i,n}) + M_na_n(y)\mathop{\sum_{j=1}^l}_{j\neq i} |v_{j,n}|^{\gamma+1} |v_{i,n}|^{\gamma-1}v_{i,n}(y) \qquad \forall y \in \Omega_n \end{equation} with $M_n := \beta_n(\frac{ L_n}{\eta(x_n)})^{2\gamma} r_n^{2\gamma+2}$. \item The sequence $\overline{v}_n$ has a uniform Lipschitz bound: \begin{equation}\label{eq:boundedgradient} \sup_{i=1,...,l} \sup_{x \in \Omega_n} |\nabla \overline{v}_{i,n}| = \sup_{i=1,...,l}\, \mathop{\sup_{x,y \in \Omega_n}}_{x\neq y}\frac{|\overline{v}_{i,n}(x)-\overline{v}_{i,n}(y)|}{|x-y|} \leq M^{1/2}, \end{equation} for some $C>0$. Also, \begin{equation}\label{bounds_nabla_at_origin} |\nabla \overline{v}_{1,n}(0)| \geq \theta^\frac{1}{2}\quad \text{ and } \lim \inf |\nabla v_{1,n}(0)| \geq \theta^{\frac{1}{2}}. \end{equation} \item There exists $v = (v_1,...,v_l) \in C(\mathbb{R}^N,\mathbb{R}^l)$ such that $v_n, \overline{v}_n \rightarrow v$ in $L^\infty_{loc}(\mathbb{R}^N)$. \item Moreover, $v_{n} \rightarrow v$ in $H^1_{loc}(\mathbb{R}^N)$ and, for any $r$, there exists $C_r$ such that: \begin{equation}\label{eq:uniform_nonvariational} \int_{B_r} |M_n|a_n(y) \mathop{\sum_{j=1}^l}_{j\neq i} |v_{j,n}|^{\gamma+1}|v_{i,n}|^{\gamma-1}v_{i,n}dy \leq C_r \end{equation} for every $i \in \{1,...,l\}$. In particular, if $M_n \rightarrow -\infty$, then $v_{i,n}\cdot v_{j,n} \rightarrow 0$ whenever $j \neq i$. \end{enumerate} \end{prop} \begin{proof} Items (1) and (2) follows directly from the definitions and Lemma \ref{lemmaForMatrixTilde}. \noindent \textbf{Proof of (3).} By using \textbf{(A2)} and the fact that $\sup_{x \in B_2}|\nabla(\eta u_{i,\beta_n}(x))|\leq L_n$ for all $i \in \{1,...,l\}$ and $\eta=0$ on ${\mathbb R}^N\setminus B_2$: \begin{align*} |\nabla \overline{v}_{i,n}(y)| &= \frac{ |A(x_n)^{\frac{1}{2}}\nabla(\eta u_{i,\beta_{n}})(x_n+r_nA_n(x_n)^{\frac{1}{2}}y)| }{ L_n }\leq M^{1/2} \quad \quad \forall i = 1,...,l. \end{align*} On the other hand, we have: \begin{align*} |\nabla \overline{v}_{1,n}(0)| &= \frac{ |A(x_n)^{\frac{1}{2}}\nabla(\eta u_{1,\beta_{n}})(x_n)| }{ L_n }= \frac{1}{L_n}\langle A(x_n)^\frac{1}{2}\nabla(\eta u_{1,\beta_n} )(x_n),A(x_n)^\frac{1}{2}\nabla(\eta u_{1,\beta_n} )(x_n) \rangle^{\frac{1}{2}} =\\ &= \frac{1}{L_n}\langle A(x_n)\nabla(\eta u_{1,\beta_n} )(x_n),\nabla(\eta u_{1,\beta_n} )(x_n) \rangle^{\frac{1}{2}} \geq \theta^{\frac{1}{2}}\frac{ |\nabla(\eta u_{1,\beta_n} )(x_n)| } { L_n } = \theta^{\frac{1}{2}}, \end{align*} by using hypothesis \textbf{(A1)}. Moreover, $$ \nabla \overline{v}_{1,n}(0) = \frac{u_{1,\beta_n}(x_n) A(x_n)^{\frac{1}{2}}\nabla \eta(x_n)}{L_n} + \frac{A(x_n)^{\frac{1}{2}}\nabla u_{1,\beta_n}(x_n) \eta(x_n)}{L_n} = O(\frac{1}{L_n}) + \nabla v_{1,n}(0) $$ (using this time conditions \textbf{(A2)} and \eqref{boundedness}), concluding the result. \vspace{2mm} \noindent \textbf{Proof of (4).} Since the sequence $\bar{v}_n$ is bounded at $0$ (recall \eqref{rnDefinition}) and has bounded gradient (recall \eqref{eq:boundedgradient}), by Ascoli-Arzela's Theorem one can find $v \in C(\mathbb{R}^N)$ such that $\bar v_n \rightarrow v$ uniformly over any compact set (up to a subsequence). Now, for any $K \subset \mathbb{R}^N$ compact and $y \in K$, we have: $$ |\overline{v}_{i,n}(y) - v_{i,n}(y)| \leq \sup_{y \in K} \frac{ u_{i,\beta_n}(x_n+r_nA(x_n)^{\frac{1}{2}}y) }{ L_n }\frac{ |\eta(x_n) - \eta(x_n+r_nA(x_n)^{\frac{1}{2}}y)| }{r_n} \leq \frac{C}{L_n} \rightarrow 0 $$ we also conclude that $v_{i,n} \rightarrow v_i$ uniformly over any compact set in $\mathbb{R}^N$, for all $i \in \{1,...l\}$. \vspace{2mm} \noindent \textbf{Proof of (5).} For $r>0$ fixed, test the equation of $v_{i,n}$ against a function $\phi \in C_c^\infty(B_{2r})$ satisfying $0 \leq \phi\leq 1$ and $\phi(x) = 1$ for $x \in B_r$. Thus, for large $n$, there exists $C_r$ depending only on $r$ such that \begin{align*} &\mathop{\sum_{j=1}^l}_{j\neq i} \int_{B_r} a_n(y)|M_n| |v_{j,n}|^{\gamma+1}|v_{i,n}|^{\gamma-1}v_{i,n}dy \leq \mathop{\sum_{j=1}^l}_{j\neq i} -\int_{B_{2r}} a_n(y) M_n |v_{j,n}|^{\gamma+1}|v_{i,n}|^{\gamma-1}v_{i,n}\phi\\ &= \large|\int_{B_{2r}} f_{i,n}(y,v_{i,n})\phi + \langle A_n(y)\nabla v_{i,n}, \nabla \phi \rangle \large|= \large| \int_{B_{2r}} f_{i,n}(y,v_{i,n})\phi - v_{i,n} \div( A_n(y) \nabla \phi ) dy \large|\\ &\leq |B_{2r}|\cdot \|v_{i,n}\|_{L^\infty(B_{2r})} \large(d + C\|DA_n\|_{L^\infty(B_{2r})} \|\phi\|_{C^1(B_{2r})}+ C\|A_n\|_{L^\infty(B_{2r})} \|\phi\|_{C^2(B_{2r})} \large) \leq C_r, \end{align*} where we have used Lemma \ref{lemmaMatrixBlowLimit}, the fact that $v_{i,n}$ is uniformly bounded over compact sets (item (4)), and the fact that $\|DA_n\|_{L^\infty(B_{2r})}$ and $\|A_n\|_{L^\infty(B_{2r})}$ are uniformly bounded. This yields \eqref{eq:uniform_nonvariational}. To prove the that $v_n\to v$ in $H^1_{loc}({\mathbb R}^N)$, we test this time the equation of $v_{i,n}$ against $v_{i,n}\phi$. Using the ellipticity constant for $A_n$ given by \eqref{ellipticConstantForBlowUp}, we obtain: \begin{align} \frac{\theta}{M} &\int_{B_r}|\nabla v_{i,n}(x)|^2dx \leq \int_{B_r} \langle A_n(x) \nabla v_{i,n}, \nabla v_{i,n} \rangle \leq \left|\int_{B_{2r}}\phi(x)\langle A_n(x) \nabla v_{i,n}, \nabla v_{i,n}\rangle dx\right| \label{boundOnDerivativeProp2}\\ =& \left| \int_{B_{2r}} - \langle A_n(x)\nabla v_{i,n}, \nabla\phi \rangle v_{i,n} f_{i,n}(x,v_{i,n}) v_{i,n}\phi + a_n(x)M_n\sum_{j\neq i} |v_{j,n}|^{\gamma+1}|v_{j,n}|^{\gamma-1} v_{i,n}^2\phi \right|\nonumber \\ =& \left| \int_{B_{2r}} \frac{1}{2} v_{i,n} \div\left( A_n(x)\nabla\phi\right) v_{i,n}+ f_{i,n}(x,v_{i,n}) v_{i,n}\phi + a_n(x)M_n\mathop{\sum_{j=1}^l}_{j\neq i} |v_{j,n}|^{\gamma+1}|v_{j,n}|^{\gamma-1}v_{i,n}^2\phi \right|,\nonumber \end{align} and this is uniformly bounded: the first two terms since $v_{i,n}$ and $\div\left( A_n(x)\nabla\phi\right)$ are uniformly bounded in $B_{2r}$, and the last one by \eqref{eq:uniform_nonvariational}. In particular, $v_n$ is bounded in $H^1_{loc}({\mathbb R}^N)$ and $v_n \rightharpoonup v_i$ weakly. Now given $K>0$, we consider the set: $E_K = \large\{ s \in [0,r]: \sup_{i, n} \int_{\partial B_s} |\nabla v_{i,n}|^2 d\sigma(y) > K \large\} $. Since there exists $C>0$ such that for all $n$: \begin{align*} C \geq& \int_{B_r} |\nabla v_{i,n}|^2dx = \int_{0}^r \int_{\partial B_s} |\nabla v_{i,n}|^2 d\sigma(x) dr, \end{align*} then this implies that $|E_K| \leq \frac{C}{K}$. Thus by taking $K$ large enough such that $|E_K| < \epsilon$, then we can choose a slightly smaller radius $r' \in [r-\epsilon, r]$ such that: \begin{align*} \int_{\partial B_{r'}} |\nabla v_{i,n}|^2dx \leq K \end{align*} for all $n \in \mathbb{N}$. For the rest of the proof, we will still call $r$ this slightly smaller radius. Now we test the equation for $v_{i,n}$ against $v_{i,n}-v_i$ in $B_r$, obtaining \begin{align*} &\Big|\int_{B_r}\langle A_n(y) \nabla v_{i,n}, \nabla (v_{i,n}-v_{i}) \rangle \left| = \Big|\int_{\partial B_r} \langle A_n(y) \nu_y, \nabla v_{i,n} \rangle (v_{i,n} - v_{i}) d\sigma \right.\\ &\qquad \left. + \int_{B_r} f_{i,n}(y,v_{i,n})(v_{i,n}-v_{i}) + M_n a_n(y)\sum_{j\neq i}|v_j|^{\gamma+1} |v_i|^\gamma (v_{i,n}-v_{i})\, dy \right|\\ \leq& \left( M \int_{\partial B_r} |\nabla v_{i,n}| d\sigma + \int_{B_r} |f_{i,n}(y,v_{i,n})| + |M_n| a_n(y) \sum_{j\neq i} |v_{i,n}|^{\gamma+1}|v_{i,n}|^\gamma\, dy \right) \|v_{i,n}-v_{i}\|_{L^{\infty}(B_r)} \rightarrow 0 \end{align*} since all the terms are bounded and $\|v_{i,n}-v_{i}\|_{L^\infty(B_r)} \rightarrow 0$. On the other hand, by the weak covergence of $v_{i,n}$ we have: $$\int_{B_r}\langle \nabla v_{i}, \nabla(v_i-v_{i,n}) \rangle dy \rightarrow 0,\qquad \text{ and so }\qquad \int_{B_r}\langle A_n(y) \nabla v_i, \nabla (v_i-v_{i,n}) \rangle dy \rightarrow 0.$$ We conclude: $$ \theta\int_{B_r} \large| \nabla (v_i - v_{i,n})|^2 dy \leq \int_{B_r} \langle A_n(y)\nabla (v_i-v_{i,n}), \nabla (v_i-v_{i,n})\rangle dy \rightarrow 0 $$ therefore $v_{i,n}$ converges strongly in $H^1(B_r)$ to $v_i$. This, combined with \eqref{eq:uniform_nonvariational} and the lower bound of the function $a_n$ (cf. \textbf{(a)}), $$ \sum_{j\neq i}^l\int_{B_r} \delta |v_{i}|^{\gamma+1}|v_{i}|^\gamma dx \leq \lim \sum_{j\neq i}^l\int_{B_r} a_n(x) |v_{i,n}|^{\gamma+1}|v_{i,n}|^\gamma(x)dx \leq \lim \frac{C_r}{|M_n| }=0. $$ In conclusion, $v_i v_j = 0$ whenever $i \neq j$. \end{proof} \begin{lemma} \label{nonTrivial} Let $v=(v_1,\ldots, v_l)$ be the limit of the sequences $(v_n)$ and $(\bar v_n)$, provided by Proposition \ref{ListProp}-(4). Then the first component, $v_1$, is nonconstant. \end{lemma} \begin{proof} We split the proof in two cases, according to the asymptotic behaviour of $M_n$. \smallbreak \noindent Case 1. If $M_n$ is bounded, then the right hand side of \eqref{eq:system_rescaled} is also bounded. Then, by elliptic regularity theory, that $v_{1,n}\to v_1$ in $C^{1,\alpha}_{loc}$, for every $\alpha\in (0,1)$. Thus, by \eqref{bounds_nabla_at_origin}: $$ 0 < \theta^{\frac{1}{2}}\leq \lim \inf|\nabla v_{1,n}(0)| = |\nabla v_1(0)|, $$ and so $v_1$ is nonconstant. Case 2. If $|M_n| \rightarrow \infty$, then we know the limit $v=(v_1,\ldots, v_k)$ has segregated components by Proposition \ref{ListProp}-\emph{6}, that is, $v_iv_j = 0$ whenever $i\neq j$. Since $ \bar v_n\to v$ uniformly in $B_r$, passing to the limit the normalization condition \eqref{rnDefinition} we have $$\sum_{j=1}^l {v}_j(0) = 1$$ Therefore we either have $v_1(0) = 1$ or $v_1(0)= 0$. \smallbreak First we suppose $v_1(0) = 0$ (and we check that this leads to a contradiction). Then there exists $h \in \{2,...,l\}$ such that $v_h(0) = 1$. By the uniform convergence $v_n\to v$ in $B_r$, we must have: \begin{equation} \label{lowBoundSomewhere} v_{h,n}(0)\geq \frac{7}{8} \qquad \text{ for $n$ large enough.} \end{equation} Using Proposition \ref{ListProp}-(3), there exists $C>0$ such that $\|\nabla \overline{v}_n\|_{L^\infty(B_r)} \leq C$. Therefore, since $v_n-\bar v_n\to 0$ locally uniformaly, for $x \in B_{\frac{1}{2C}}$ we have $$|v_{h,n}(x) - v_{h,n}(0)| \leq |v_{h,n}(x) - \overline{v}_{h,n}(x)| + |\overline{v}_{h,n}(x) - \overline{v}_{h,n}(0)| \leq o(1) + C|x| \leq o(1)+ \frac{1}{2}$$ Therefore \[ v_{h,n}(x) \geq \frac{1}{8}\qquad \text{ for $x \in B_{\frac{1}{2C}}$,\ $n$ large enough, } \] For every $\epsilon>0$, by Lemma \ref{lemmaMatrixBlowLimit} we also have $$ \sup_{x\in B_{\frac{1}{2}}}|f_{i,n}(x,v_{i,n}(x))| < \epsilon \qquad \text{ for $n$ large}; $$ combining this with $a_n(y)\geq \delta>0$ and inequality \eqref{lowBoundSomewhere}, and remembering that $M_n<0$, we are able to conclude that $v_{1,n}$ satisfies the inequalities: \begin{equation} \label{equationForLemmaBound} -\div\left( A_n(x)\nabla v_{1,n} \right) \leq - \delta \left( \frac{7}{8} \right)^{\gamma+1} |M_n|v_{1,n}^\gamma + \epsilon, \end{equation} \begin{equation} \label{equationForGoodBound} |\div(A_n(x)\nabla v_{1,n})| \leq \delta \left( \frac{7}{8} \right)^{\gamma+1} |M_n|v_{1,n}^\gamma + \epsilon. \end{equation} Using equation \eqref{equationForLemmaBound} together with Lemma \ref{estimateLemma}-(1), we obtain the existence of $c>0$ such that $|M_n|v_{1,n}^\gamma(x)< c$ for $x \in B_{\frac{1}{4C}}$. Plugging this information in \eqref{equationForGoodBound} , we conclude that $|\div(A_n(x)\nabla v_{1,n}(x))|<C$ for all $n \in \mathbb{N}$. From the uniform ellipticity of $A_n(x)$ and by elliptic regularity theory, we conclude that (up to a subsequence) $v_{1,n} \rightarrow v_1$ in $C^1(B_{\frac{1}{4C}})$, and once again by \eqref{bounds_nabla_at_origin} we have: $$ |\nabla v_1(0)| = \lim_n |\nabla v_{1,n}(0)| \geq \theta^\frac{1}{2}>0, $$ which is a contradiction since $v_1$ is positive and $v_1(0) = 0$. Thus we must have $v_1(0) = 1$. Since $v_1(0) = 1$, we can reason similarly to the previous paragraph, showing via Lemma \ref{estimateLemma}-(1) that $|M_n| v_{j,n}^\gamma(x)< C$ for $x \in B_{\frac{1}{4C}}$ and $j \neq 1$. From this we conclude the boundedness of $|\div(A_n(x)\nabla v_{1,n})|$, so we can take the convergence $v_{n,1}\to v_1$ in $C^1(B_{\frac{1}{4C}})$ and $$ |\nabla v_{1}(0)|=\lim_n|\nabla v_{1,n}(0)| \geq \theta^\frac{1}{2}>0, $$ concluding the fact that $v_1$ is nonconstant. \end{proof} Now we make a concluding proposition of this section, which summarizes what is known about the limiting profiles $v$. \begin{prop} \label{concludeProp} Let $v=(v_1,\ldots, v_l)\in H^1_{loc}({\mathbb R}^N)\cap C({\mathbb R}^N)$ be the limit of the sequences $(v_n)$ and $(\bar v_n)$. Then \begin{equation}\label{eq:Lipschitzbound_limit} \max_{i=1,...,l}\mathop{\sup_{x\neq y}}_{x,y\in\mathbb{R}^N}\frac{|v_i(x)-v_i(y)|}{|x-y|}\leq C, \end{equation} and $v_1$ is nonconstant. Actually, $v$ may only have at most another nontrivial component, say $v_2$. These two components satisfy \begin{equation*} v_1(0)+v_2(0)=1, \qquad |\nabla v_1(0)|\geq \theta^\frac{1}{2}>0, \end{equation*} and furthermore \begin{enumerate} \item If $M_n \rightarrow -\infty$, then both $v_1$ and $v_2$ are subharmonic in $\mathbb{R}^N$ and: \begin{equation} \label{unboundedEquation} \begin{cases} -\Delta v_1 = 0 & \text{in } \, \{v_1>0\}\\ -\Delta v_2 = 0 & \text{in } \, \{v_2>0\}\\ v_1\cdot v_2 = 0 & \text{in } \, \mathbb{R}^N\\ v_1, v_2 \geq 0 & \text{in } \, \mathbb{R}^N \end{cases} \end{equation} \item If $M_n$ is bounded then there exists $M_\infty <0$ such that up to a subsequence $M_n \rightarrow M_\infty < 0$ and: \begin{equation} \label{boundedEquation} \begin{cases} -\Delta v_1 = M_\infty v_1^\gamma v_2^{\gamma+1} & \text{in } \, \mathbb{R}^N\\ -\Delta v_2 = M_\infty v_2^\gamma v_1^{\gamma+1} & \text{in } \, \mathbb{R}^N\\ v_1, v_2 \geq 0 & \text{in } \, \mathbb{R}^N \end{cases} \end{equation} \end{enumerate} \end{prop} \begin{proof} The first observation is just a consequence of Proposition \ref{ListProp}. To show that the limit can only have at most two nontrivial components and that these satisfy either \eqref{unboundedEquation} or \eqref{boundedEquation}, we divide the discussion in two cases. \noindent \textbf{Case 1.} Suppose that $M_n \rightarrow -\infty$. Then, by Proposition \ref{ListProp}-(5), we know that $v_{i,n}\cdot v_{j,n} \rightarrow 0$ locally uniformly for all $j \neq i$, and so also $v_i\cdot v_j=0$ whenever $j \neq i$. Fix $i$ and let $\phi \in C_c^\infty(\{v_i>0\})$, $K:= \text{supp}(\phi)$ and $\delta>0$ such that $v_i|_{K}>\delta$. Exactly as we did in the proof of Case 2 in Lemma \ref{nonTrivial}, we can use Lemma \ref{estimateLemma}-(1) and show that we have that $|M_n\|v_{j,n}|^\gamma \leq C$ in $K$, for every each $j \neq i$ (just use the uniform boundedness of $v_{n}$ over this compact set and take a cover by balls). In particular, $|M_n\|v_{j,n}|^{\gamma+1}|v_{i,n}(x)|^\gamma \rightarrow 0$ uniformly in $K$ Thus, using the equation for $v_n$ in \eqref{eq:system_rescaled} and the fact that $A_n(y) \to Id$ and $f_{i,n}(y,v_{i,n}(y)) \to 0$ over compact sets (recall Lemma \ref{lemmaMatrixBlowLimit}), we obtain: \begin{align*} \int_{\{v_i>0\}} \langle \nabla \phi,\nabla v_{i}\rangle dy &= \lim_n \int_{K} \langle A_n(y)\nabla \phi, \nabla v_{i,n}\rangle dy\\ &= \lim_n \int_{K} \left( f_{i,n}(y,v_{i,n})\phi(y) + M_na_n(y) \sum_{j\neq i}^l|v_{j,n}|^{\gamma+1}|v_{i,n}|^\gamma \phi(y) \right)dy=0, \end{align*} In conclusion, we have that: $$ -\Delta v_i = 0\qquad \text{ in $\{v_i>0\}$.} $$ Since $v_i \in C(\mathbb{R}^N)$ are zero outside this set we conclude that the $v_i$ are subharmonic in $\mathbb{R}^N$. Using Lemma \ref{subHarmonicNonTrivial}, since we have the bound \eqref{eq:Lipschitzbound_limit}, we conclude that this system has at most two nontrivial components. \noindent \textbf{Case 2.} In case $M_n \rightarrow M_\infty$ then using the fact that $f_{i,n}(y,v_{i,n}(y)) \rightarrow 0$ and that $a_n(x) \rightarrow a(x_\infty)>0$ uniformly over compact sets we conclude that: $$ -\Delta v_i = M_\infty a(x_\infty)\sum_{j\neq i} v_{i}^{\gamma}v_{j}^{\gamma+1}, $$ since, by elliptic regularity, $v_{i,n}$ converges in $C^{1,\alpha}_{loc}(\mathbb{R}^N)$. If $M_\infty = 0$, this would imply that the functions $v_i$ are harmonic and nonnegative which is a contradiction, thus $M_\infty < 0$. Using Lemma \ref{equationNonTrivial}, we conclude that $v$ only has two nontrivial components. \end{proof} Proposition \ref{concludeProp}, by itself, does not give any contradiction related to the limiting profiles $v$. We need to add more information about these limits; for this, in Section \ref{chapter:implementation} we explore Almgren's monotonicity formulas (a consequence of the variational structure of the approximating system) and in Sections \ref{chapter:resultsChap4} we show a Alt-Caffarelli-Friedman type monotonicity formula. \section{Almgren's Monotonicity Formula} \label{chapter:implementation} In this section we will deduce an Almgren's Monotonicity formula. The formula will be stated for vector functions $u_\beta = (u_{1,\beta},...,u_{l,\beta})$, whose components are positive solutions of the system: \begin{equation} \label{AlmgrenEq} -\div(A(x)\nabla u_{i,\beta})= f_i(x,u_{i,\beta})+a(x)\beta \mathop{\sum^l_{i,j=1}}_{j\neq i} u_{j,\beta}^{\gamma+1}u_{i,\beta}^\gamma \qquad i=1,...,l \end{equation} satisfying conditions \textbf{(A1)}, \textbf{(A2)}, \textbf{(F)} and \eqref{boundedness} where $\beta <0$. We also assume that the matrix $A$ satisfies: \begin{equation*} A(0)=Id. \end{equation*} Thus, around zero, the operator $v \mapsto \div(A(x)\nabla v)$ is just a perturbation of the usual Laplacian. The main purpose of this section is to prove Theorem \ref{almgrenMonotonicity} below. \smallbreak We define: $$ \mu(x) := \langle A(x)\frac{x}{|x|}, \frac{x}{|x|} \rangle \geq \theta \qquad \forall x \in \mathbb{R}^N\setminus \{0\}, $$ The following lemma is taken from \cite{HugoHolderVariable} (which, in turn, is based on \cite{GAROFALO2014682}). \begin{lemma}[{{\cite[Lemma C.2]{HugoHolderVariable}}}] \label{listLemma} There exists a constant $C$ and radius $0<\tilde{r}< d(0, \partial \Omega)$, depending only on the dimension $N$ and $\theta, M$ (bounds from conditions \textbf{(A1)} and \textbf{(A2)}) such that, for $|x| < \tilde{r}$, we have: \begin{enumerate} \item $\|A(x) - Id\| \leq C|x|,$ \item $|\mu(x)-1| \leq C|x|,$ \item $|\frac{1}{\mu(x)} - 1| \leq C|x|,$. \item $|\frac{1}{\mu^2(x)}-1| \leq \frac{C}{1-C|x|^2}|x|,$ \item $|\nabla \mu(x)| \leq C,$ \item $|\div(A(x)\nabla |x|)-\frac{N-1}{|x|}|\leq C,$ \item $|\div(\frac{A(x)x}{\mu(x)}) - N| \leq C|x|.$ \end{enumerate} \end{lemma} Define the following: \begin{align} E_{i,\beta}(u_\beta,r)&= \frac{1}{r^{N-2}}\int_{B_r}\langle A(x)\nabla u_{i,\beta}, \nabla u_{i,\beta} \rangle - f_i(x,u_{i,\beta})u_{i,\beta} - \mathop{\sum_{j=1}^l}_{j\neq i} a(x)\beta |u_{j,\beta}|^{\gamma +1} |u_{i,\beta}|^{\gamma+1} \Big)dx \nonumber \\ &= \frac{1}{r^{N-2}} \int_{\partial B_r} u_{i, \beta}\langle A(x) \nabla u_{i,\beta}, \nu\rangle d\sigma(x)\ \text{ for each $i=1,\ldots, l$, and} \label{ETerm}\\ E_\beta(u_\beta,r)&:=\sum_{i=1}^l E_{i,\beta}(u_\beta,r). \nonumber \end{align} where, to obtain \eqref{ETerm}, we have tested the $i$-th equation in \eqref{AlmgrenEq} by $u_{i,\beta}$ and integrated by parts. We also define \begin{align*} H_{i,\beta}(u_\beta,r) := \frac{1}{r^{N-1}}\int_{\partial B_r}\mu(x)|u_{i,\beta}|^2d\sigma(x), \ i=1,\ldots, l,\quad \text{ and } \quad H_\beta(u_\beta,r) := \sum_{i=1}^l H_{i,\beta}(u_\beta,r), \end{align*} and, whenever $H_\beta(u_\beta,r)\neq 0$, consider the \emph{Almgren's quotient}: \begin{align} \label{Neq} N_\beta(u_\beta,r):= \frac{ E_\beta(u_\beta,r) }{ H_\beta(u_\beta,r) }. \end{align} \begin{lemma} \label{derivOfH} There exists $\overline{r},C>0$ depending only on the dimension $N$, and the constants $M$ and $\theta$ from conditions \textbf{(A1)} and \textbf{(A2)}, such that: \[ \Big| H_{i,\beta}'(u_\beta,r)-\frac{2}{r}E_{i,\beta}(u_\beta,r) \Big| \leq CH_{i,\beta}(u_\beta,r),\quad \text{ and }\quad r\mapsto H_{i,\beta}(u_\beta,r)e^{Cr} \text{ is monotone nondecreasing} \] for all $r \in ]0,\overline{r}[$, $\beta <0$. In particular, by summing up in $i$, $$\Big| H_{\beta}'(u_\beta,r)-\frac{2}{r}E_{\beta}(u_\beta,r) \Big| \leq CH_{\beta}(u_\beta,r),\quad \text{ and }\quad r\mapsto H_{\beta}(u_\beta,r)e^{Cr} \text{ is monotone nondecreasing} $$\end{lemma} \begin{proof} Recalling \eqref{ETerm}, the proof of the bounds follows exactly as the one of \cite[Lemma C.5]{HugoHolderVariable}. Regarding the monotonicity, we compute the derivative of $H_{i,\beta}(u_\beta, r)$: \begin{align*} H_{i,\beta}'(u_\beta,r) &\geq \frac{2}{r}E_\beta(u_\beta,r) -CH_{i,\beta}(u_\beta,r) \\ &\geq \frac{2}{r^{N-1}}\int_{B_r}\left( \langle A(x) \nabla u_{i,\beta} , \nabla u_{i,\beta}\rangle - f_i(x,u_{i,\beta})u_{i,\beta} \right) dx - CH_{i,\beta}(u_\beta,r) \\ &\geq \int_{B_r}\Bigg(\frac{2\theta}{r^{N-1}}|\nabla u_{i,\beta}|^2 - \frac{2d r}{r^{N}}u_{i,\beta}^2\Bigg)dx- CH_{i,\beta}(u_\beta,r)\\ &= 2\theta \int_{B_r}\Bigg(\frac{1}{r^{N-1}} |\nabla u_{i,\beta}|^2 - \frac{d r}{\theta r^{N}}u_{i,\beta}^2\Bigg)dx - CH_\beta(u_\beta,r). \end{align*} By using Poincar\'e's inequality (Lemma \ref{PoincareIneq}) we know that, if $dr/\theta<N-1$, then: $$\int_{B_r} \left( \frac{1}{r^{N-1}}|\nabla u_{i,\beta}|^2 - \frac {d r} {\theta r^{N}} u_{i,\beta}^2 \right) dx \geq - \frac{1}{r^{N-1}}\int_{\partial B_r} u_{i,\beta}^2d\sigma(x) \geq - \frac{1}{\theta r^{N-1}} \int_{\partial B_r} \mu(x)u_{i,\beta}^2 d\sigma(x)$$ since $\mu(x)\geq \theta$. With this, we conclude: $$H_{i,\beta}'(u_\beta,r)\geq 2\theta\Big(-\frac{1}{\theta} H_{i,\beta}(u_\beta,r)\Big) - CH_{i,\beta}(u_\beta,r)\geq -\tilde{C}H_{i,\beta}(u_\beta,r),$$ from which the monotonicity of $\mapsto H_{i,\beta}(u_\beta,r)e^{\tilde Cr}$ easily follows. \end{proof} We now define the following vector field: $$Z(x) = \frac{A(x)x}{\mu(x)}$$ Observe that $Z(x)\sim x$ for $x\sim 0$, since $A(0) = Id$ and $\mu(x) \rightarrow 1$. Using \textbf{(A1)}, \textbf{(A2)} and Lemma \ref{listLemma}-\emph{7}, we have the existence of $\overline{r}$ and $C>0$ such that: \begin{equation}\label{eq:bounds_Z} |Z(x)|\leq C|x|, \qquad |\div(Z(x)-x)|\leq C|x|, \qquad \forall x: |x|\leq \overline{r}. \end{equation} The following lemma is written in the Einstein notation, so summations are suppressed. This means that any repeated indices over any pair of variables having that index are supposed to be summed over. \begin{lemma} \label{PozohaevLemma} For $r>0$ such that $B_r \subset \Omega$, if $A(x) = A = (a_{ij})$ we have that: \begin{align*}& r\int_{\partial B_r} \langle A(x)\nabla u_{i,\beta}, \nabla u_{i,\beta} \rangle = \int_{B_r} \div(Z) \langle A\nabla u_{i,\beta}, \nabla u_{i,\beta} \rangle + 2\int_{B_r} f_{i}(x,u_{i,\beta}) \langle\nabla u_{i,\beta}, Z \rangle\\ +& 2\int_{\partial B_r} \langle Z, \nabla u_{i,\beta} \rangle \langle A \nabla u_{i,\beta} , \nu\rangle + \int_{B_r} \langle Z, \nabla a_{hl}\rangle \frac{\partial u_{i,\beta}}{\partial x_h}\frac{\partial u_{i,\beta}}{\partial x_l} - 2\int_{B_r} a_{hl}\frac{\partial Z_j}{\partial x_h} \frac{\partial u_{i,\beta}}{\partial x_j}\frac{\partial u_{i,\beta}}{\partial x_l}\\ +& 2\sum_{j<i} \int_{B_r}\left( -\frac{\div(Z)}{\gamma+1} -\frac{\langle Z, \nabla a(x) \rangle}{a(x)(\gamma+1)} \right) a(x)\beta|u_{j,\beta}|^{\gamma+1}|u_{i,\beta}|^{\gamma+1} + 2 \sum_{j<i} \int_{\partial B_r}\frac{\langle Z, \nu \rangle}{\gamma +1} a(x)\beta |u_{j,\beta}|^{\gamma+1}|u_{i,\beta}|^{\gamma+1}. \end{align*} \end{lemma} \begin{proof} This proof follows exactly as the one of \cite[Lemma C.6]{HugoHolderVariable} (without considering the singular limit right at the end of the proof). \end{proof} Now define the quantity: $$\tilde{E}_\beta(u_\beta,r) := \sum_{i=1}^l\frac{1}{r^{N-2}}\int_{B_r}\langle A(x)\nabla u_{i,\beta}, \nabla u_{i,\beta} \rangle.$$ whose derivative we compute in the following lemma. \begin{lemma} \label{estimateOnVariable} We have: \begin{align} \tilde{E}_\beta'(u_\beta,r) =&\sum_{i=1}^l\Bigg( \frac{2}{r^{N-2}} \int_{\partial B_r} \frac{\langle A(x) \nabla u_{i,\beta} , \nu_x \rangle^2}{\mu(x)} d\sigma(x) + \frac{2}{r^{N-1}}\int_{B_r} f_{i}(x,u_{i,\beta})\langle\nabla u_{i,\beta}, Z \rangle \nonumber \\ &+ \frac{1}{r^{N-1}}\int_{B_r}\langle Z, \nabla a_{hl} \rangle \frac{\partial u_{i,\beta}}{\partial x_h}\frac{\partial u_{i,\beta}}{\partial x_l} - \frac{2}{r^{N-1}}\int_{B_r} a_{hl}\frac{\partial (Z_j-x_j)}{\partial x_h} \frac{\partial u_{i,\beta}}{\partial x_j}\frac{\partial u_{i,\beta}}{\partial x_l} \nonumber \\ &+ \frac{1}{r^{N-1}}\int_{B_r}\div(Z-x)\langle A \nabla u_{i,\beta} , \nabla u_{i,\beta}\rangle \Bigg) \nonumber \\ &+ \mathop{\sum_{i,j=1}^l}_{j<i} \frac{2}{r^{N-1}(\gamma+1)}\int_{\partial B_r}\langle Z(x), \nu_x \rangle{a(x)\beta}|u_{j,\beta}|^{\gamma+1}|u_{i,\beta}|^{\gamma+1} d\sigma(x) \nonumber \\ &+\frac{2}{r^{N-1}} \mathop{\sum_{i,j=1} ^l}_{j<i} \int_{B_r}\left(-\frac{\div(Z)}{\gamma+1} - \frac{\langle Z(x), \nabla a(x) \rangle}{a(x)(\gamma+1)} \right) a(x)\beta|u_{j,\beta}|^{\gamma+1}|u_{i,\beta}|^{\gamma+1}. \label{KineticEquation} \end{align} Moreover, there also exist $C$ and $\overline{r}$ such that: \begin{align} \Big|\sum_{i=1}^l\Big( \frac{1}{r^{N-1}} \int_{B_r} \langle Z, \nabla a_{hl}\frac{\partial u_{i,\beta}}{\partial x_h}\frac{\partial u_{i,\beta}}{\partial x_l} \rangle - \frac{2}{r^{N-1}}&\int_{B_r} a_{hl}\frac{\partial (Z_j-x_j)}{\partial x_h} \frac{\partial u_{i,\beta}}{\partial x_j} \frac{\partial u_{i,\beta}}{\partial x_l} \nonumber \\ + \frac{1}{r^{N-1}} &\int_{B_r}\div(Z-x) \langle A \nabla u_{i,\beta} , \nabla u_{i,\beta}\rangle \Big) \Big| \leq C\tilde{E}(u_\beta,r) \label{weirdBound} \end{align} for all $r \in ]0,\overline{r}[$ and $\beta < 0$. \end{lemma} \begin{proof} We have \begin{align} \tilde{E}_\beta'(u_\beta, r) &= \frac{2-N}{r}\tilde {E}_\beta(u_\beta, r) + \frac{1}{r^{N-2}} \left(\int_{B_r} \langle A\nabla u_{i,\beta}, \nabla u_{i,\beta}\rangle\right)' \nonumber \\ &= -\frac{(N-2)}{r^{N-1}}\int_{B_r} \langle A\nabla u_{i,\beta},\nabla u_{i,\beta}\rangle + \frac{1}{r^{N-2}}\left(\int_{B_r} \langle A\nabla u_{i,\beta}, \nabla u_{i,\beta}\rangle\right)'.\label{eq:original_derivative} \end{align} In order to compute the derivative of the last term, we define: \begin{align*} Q_\beta(u_\beta, r) :=& \mathop{\sum_{i,j=1}^l}_{j<i} \frac{2}{r^{N-1}(\gamma+1)} \int_{\partial B_r} \langle Z, \nu \rangle {a(x)\beta}|u_{j,\beta}|^{\gamma+1}|u_{i,\beta}|^{\gamma+1} d\sigma(x)+\\ &+\frac{2}{r^{N-1}} \mathop{\sum_{i,j=1}^l}_{j<i} \int_{B_r}\left(-\frac{\div(Z)}{\gamma+1} - \frac{\langle Z, \nabla a(x) \rangle}{a(x)(\gamma+1)} \right) a(x)\beta|u_{j,\beta}|^{\gamma+1}|u_{i,\beta}|^{\gamma+1}. \end{align*} From Lemma \ref{PozohaevLemma}, $\div(Z) =N+\div(Z-x)$ and using: \[ \frac{2}{r}\int_{\partial B_r} \langle Z,\nabla u_{i,\beta}\rangle \langle A\nabla u_{i,\beta},\nu\rangle=2\int_{\partial B_r}\frac{\langle A\nabla u_{i,\beta},\nu\rangle^2}{\mu},\quad \frac{2}{r}\int_{B_r} a_{hl} \frac{\partial x_j}{\partial x_h} \frac{\partial u_{i,\beta}}{\partial x_j} \frac{\partial u_{i,\beta}}{\partial x_l}= \frac{2}{r} \int_{B_r} \langle A \nabla u_{i,\beta}, \nabla u_{i,\beta} \rangle \] and , we have \begin{align*} \frac{d}{dr} & \int_{B_r} \langle A\nabla u_{i,\beta}, \nabla u_{i,\beta}\rangle= \int_{\partial B_r} \langle A(x)\nabla u_{i,\beta}, \nabla u_{i,\beta}\rangle \\ =& \frac{1}{r}\int_{B_r} \div(Z)\langle A\nabla u_{i,\beta}, \nabla u_{i,\beta} \rangle+ \frac{2}{r}\int_{B_r} f_{i}(x,u_{i,\beta})\langle\nabla u_{i,\beta}, Z \rangle+ \frac{2}{r}\int_{\partial B_r}\langle Z, \nabla u_{i,\beta} \rangle \langle A \nabla u_{i,\beta} , \nu\rangle\\ & + \frac{1}{r} \int_{B_r} \langle Z, \nabla a_{hl}\rangle \frac{\partial u_{i,\beta}}{\partial x_h}\frac{\partial u_{i,\beta}}{\partial x_l}- \frac{2}{r}\int_{B_r} a_{hl}\frac{\partial Z_j}{\partial x_h} \frac{\partial u_{i,\beta}}{\partial x_j}\frac{\partial u_{i,\beta}}{\partial x_l} + Q_\beta(u_\beta,r)r^{N-2}+ 2\int_{\partial B_r}\frac{\langle A\nabla u_{i,\beta},\nu\rangle^2}{\mu} \\ =&\frac{1}{r}\int_{B_r} \div(x + Z -x)\langle A(x)\nabla u_{i,\beta}, \nabla u_{i,\beta} \rangle + \frac{2}{r}\int_{B_r} f_{i}(x,u_{i,\beta}) \langle\nabla u_{i,\beta}, Z \rangle\\ & + \frac{1}{r} \int_{B_r} \langle Z, \nabla a_{hl}\rangle \frac{\partial u_{i,\beta}}{\partial x_h}\frac{\partial u_{i,\beta}}{\partial x_l} - \frac{2}{r}\int_{B_r} a_{hl}\frac{\partial (Z_j-x_j+x_j)}{\partial x_h} \frac{\partial u_{i,\beta}}{\partial x_j}\frac{\partial u_{i,\beta}}{\partial x_l} + Q_\beta(u_\beta,r)r^{N-2} \\ =& \frac{N-2}{r}\int_{B_r} \langle A\nabla u_{i,\beta},\nabla u_{i,\beta}\rangle +\frac{2}{r}\int_{B_r} f_i(x,u_{i,\beta})\langle \nabla u_{i,\beta},Z\rangle \\ &+2\int_{\partial B_r}\frac{\langle A\nabla u_{i,\beta},\nu\rangle^2}{\mu}+\frac{1}{r}\int_{B_r} \div (Z-x) \langle A\nabla u_{i,\beta},\nabla u_{i,\beta}\rangle +\frac{1}{r}\int_{B_r} \langle Z,\nabla a_{hl}\rangle\frac{\partial u_{i,\beta}}{\partial x_h} \frac{\partial u_{i,\beta}}{\partial x_l}\\ &-\frac{2}{r}\int_{B_r}a_{hl}\frac{\partial (Z_j-x_j)}{\partial x_h}\frac{\partial u_{i,\beta}}{\partial x_j} \frac{\partial u_{i,\beta}}{\partial x_l} + Q_\beta(u_\beta, r)r^{N-2}. \end{align*} Going back to \eqref{eq:original_derivative}, we conclude that identity \eqref{KineticEquation} is true. It remains to prove \eqref{weirdBound}. We bound each term individually. By \textbf{(A2)} and \eqref{eq:bounds_Z}, $$ \left| \frac{1}{r^{N-1}} \int_{B_r} \langle Z,\nabla a_{h,l} \rangle \frac{\partial u_{i,\beta}}{\partial x_h} \frac{\partial u_{i,\beta}}{\partial x_l} \right| \leq \frac{C}{r^{N-2}} \int_{B_r} |\nabla u_{i,\beta}|^2 \leq \frac{C}{\theta r^{N-2}} \tilde{E}_\beta(u_\beta, r). $$ Moreover, $\div(Z(x)-N)| \leq C|x|$ and \begin{align*} |\frac{\partial}{\partial x_k} (Z_j(x)- x_j)| &= |\frac{\partial}{\partial x_k} \left( \sum_{h=1}^N \frac{a_{jh}(x)x_h}{\mu(x)} - x_j \right)|\\ &=| \sum_{h=1}^N\left( \frac{\partial a_{jh}(x)}{\partial x_k} \frac{x_h}{\mu(x)} + \frac{a_{jh}(x)\delta_{hk}}{\mu(x)} + \frac{a_{jh}(x)x_h}{\mu^2(x)} \frac{\partial \mu(x)}{\partial x_l} \right) - \delta_{jk}|\\ &\leq C|x|+ | \frac{a_{jk}(x)}{\mu(x)} -\delta_{jk} | \leq C|x| + |\frac{a_{jk}(x) - \delta_{jk}}{\mu(x)} | + | \frac{\delta_{jk}}{\mu(x)} - \delta_{jk} |\leq C'|x|, \end{align*} where we used Lemma \ref{listLemma}-\emph{1.,3.,5.} and $\mu(x)\geq \theta >0$. With this, we conclude the desired bound: \begin{multline*} \Big| \frac{1}{r^{N-1}} \int_{B_r} \langle Z,\nabla a_{hl}\rangle \frac{\partial u_{i,\beta}}{\partial x_h} \frac{\partial u_{i,\beta}}{\partial x_l} + \frac{1}{r^{N-1}} \int_{B_r} \div (Z-x) \langle A\nabla u_{i,\beta},\nabla u_{i,\beta}\rangle\\ - \frac{2}{r^{N-1}} \int_{B_r} a_{hl} \frac{\partial u_{i,\beta}}{\partial x_l} \frac{\partial (Z_j-x_j)}{\partial x_h} \frac{\partial u_{i,\beta}}{\partial x_j} \Big| \leq C\frac{1}{r^{N-2}} \int_{B_r} \langle A\nabla u_{i,\beta},\nabla u_{i,\beta}\rangle = C\tilde{E}_\beta(u_{\beta},r), \end{multline*}which completes the proof. \end{proof} \begin{thm} \label{almgrenMonotonicity} Let $u_\beta$ be a positive solution of \eqref{AlmgrenEq}, under $\beta <0$ \textbf{(A1)}, \textbf{(A2)}, \textbf{(F)} and \eqref{boundedness}. Assume moreover that \begin{equation*} A(0)=Id. \end{equation*} If $N\in \mathbb{N}$, $\gamma\geq 1$ and $\frac{\gamma N}{\gamma + 1} < 2$, then there exist constants $\overline{r}$ and $C>0$, such that: \begin{equation*} \label{NisIncreasingStuff1} N_\beta(u_{\beta},r)+1\geq 0,\qquad H_\beta(u_\beta,r) > 0,\qquad N_\beta'(u_{\beta},r) \geq -C(N(u_{\beta},r)+1) \end{equation*} In particular, \begin{equation} \label{Nmonotonicity} (N_\beta(u_\beta,r)+1)e^{Cr} \end{equation} is monotone nondecreasing for every $r \in ]0,\overline{r}[$, $\beta<0$. The constants $C$ and $\bar r$ depend only on the ellipticity constant $\theta>0$ of $A$, the upper bound $M>0$ of $\|DA\|_\infty$, the dimension $N$, and the uniform bound $m>0$ from \eqref{boundedness} and the constant $d$ from \eqref{boundForF}. \end{thm} \begin{proof} \textbf{Step 1.} We show $E_\beta(u_\beta,r) + H_\beta(u_\beta,r) \geq 0$, which is equivalent to $N_\beta(u_{\beta},r)+1\geq 0$. Indeed, since $a(x)>0$, $\beta < 0$, \eqref{boundForF} and (which implies, in particular, that $\mu(y)\geq \theta$) we have \begin{align*} E_\beta(&u_\beta,r) + H_\beta(u_\beta,r) = \sum_{i=1}^l\Bigg[\frac{1}{r^{N-2}}\int_{B_r} \left(\langle A(x) \nabla u_{i,\beta} , \nabla u_{i,\beta} \rangle - f_i(x,u_{i,\beta})u_{i,\beta} - a(x)\beta \sum_{j\neq i} u_{j,\beta}^{\gamma+1}u_{i,\beta}^{\gamma+1}\right)\\ &\qquad\qquad+ \frac{1}{r^{N-1}} \int_{\partial B_r} \mu(x) u_{i,\beta}^2d\sigma(x)\Bigg] \\ \geq&\sum_{i=1}^l \Bigg[\int_{B_r}\left(\frac{1}{r^{N-2}} \langle A(x) \nabla u_{i,\beta} \nabla u_{i,\beta} \rangle -\frac{dr^2}{r^N} u_{i,\beta}^2\right)+ \frac{1}{r^{N-1}}\int_{\partial B_r} \mu(x) u_{i,\beta}^2d\sigma\Bigg] \\ \geq& \sum_{i=1}^l\Bigg[\int_{B_r}\left( \frac{\theta}{r^{N-2}} |\nabla u_{i,\beta}|^2 -\frac{dr^2}{r^N}u_{i,\beta}^2\right) + \frac{\theta}{r^{N-1}}\int_{\partial B_r} u_{i,\beta}^2d\sigma \Bigg] \geq \sum_{i=1}^l \left [ \int_{B_r} \frac{\theta(N-1)-dr^2}{r^N} u_{i,\beta}^2 \right], \end{align*} where we have used Poincar\'e's inequality (Lemma \ref{PoincareIneq}). The claim now follows by choosing $r$ small enough so that $dr^2 < \theta (N-1)$. \noindent \textbf{Step 2.} We show equation $N_\beta'(u_{\beta},r) \geq -C(N(u_{\beta},r)+1)$ whenever $r$ is small $H_\beta(u_\beta,r) \neq 0$ (which, in particular, shows \eqref{Nmonotonicity} for such $r$'s). First of all, by Lemma \ref{derivOfH}, \[ H_\beta'(u_\beta,r) = \frac{2E_\beta(u_\beta,r)}{r} + O(1)H_\beta(u_\beta,r), \] As for $E_\beta(u_\beta,r)$, recalling that \begin{align*} E_\beta(u_\beta,r) &= \tilde{E}_\beta(u_\beta,r) - \frac{1}{r^{N-2}}\int_{B_r}f_i(x,u_{i,\beta})u_{i,\beta} - \frac{1}{r^{N-2}}\int_{B_r}2\mathop{\sum_{j=1}^l}_{j<i} a(x)\beta |u_{j,\beta}|^{\gamma +1} |u_{i,\beta}|^{\gamma+1}dx \end{align*} we have, by Lemma \ref{estimateOnVariable}, \begin{align} E_\beta'(u_\beta,r) =& \tilde{E}_\beta'(u_\beta,r) + \sum_{i=1}^l\Bigg[\frac{(N-2)}{r^{N-1}}\int_{B_r} f_i(x,u_{i,\beta})u_{i,\beta} dx - \frac{1}{r^{N-2}}\int_{\partial B_r} f_i(x,u_{i,\beta})u_{i,\beta}d\sigma(x)\nonumber\\ &-\sum_{j<i}\Bigg(\frac{(4-2N)}{r^{N-1}}\int_{B_r}a(x)\beta|u_{j,\beta}|^{\gamma+1}|u_{i,\beta}|^{\gamma +1 } dx- \frac{2}{r^{N-2}}\int_{\partial B_r} a(x)\beta|u_{j,\beta}|^{\gamma+1}|u_{i,\beta}|^{\gamma +1 }d\sigma(x)\Bigg)\Bigg] \nonumber\\ =& O(1)\tilde E_\beta(u_\beta,r) + R_\beta(u_\beta, r)+\sum_{i=1}^l \Bigg[\frac{2}{r^{N-1}} \int_{\partial B_r}\frac{\langle A(x)\nabla u_{i,\beta}, \nu_x\rangle^2}{\mu(x)}d\sigma(x) \nonumber \\ &-\sum_{j<i}\Bigg(\frac{1}{r^{N-1}}\int_{B_r}\left(4-2N +\frac{2\div(Z)}{\gamma+1} - 2\frac{\langle Z(x), \nabla a(x)\rangle}{a(x)(\gamma+1)}\right)a(x)\beta|u_{j,\beta}|^{\gamma+1}|u_{i,\beta}|^{\gamma +1 }dx \nonumber\\ &- \frac{1}{r^{N-2}}\int_{\partial B_r}\left(2+\frac{2\langle Z(x), \nu_x \rangle}{r(\gamma+1)}\right) a(x)\beta|u_{j,\beta}|^{r(\gamma+1)}|u_{i,\beta}|^{\gamma +1 }d\sigma(x)\Bigg)\Bigg], \label{derivativeOfEeq123} \end{align} where \begin{align*} R_\beta(u_\beta,r) :=& \sum_{i=1}^l\Bigg[\frac{2}{r^{N-1}} \int_{B_r} \left(f_i(x,u_{i,\beta})\langle Z(x),\nabla u_{i,\beta} \rangle+ \frac{(N-2)}{r^{N-1}}f_i(x,u_{i,\beta})u_{i,\beta}\right)\\ &- \frac{1}{r^{N-2}}\int_{\partial B_r}f_i(x,u_{i,\beta})u_{i,\beta}d\sigma\Bigg]. \end{align*} and $O(1)$ is a bounded function that comes from \eqref{weirdBound}. Now, we check that the terms in \eqref{derivativeOfEeq123} with $\beta$ are all nonnegative. Indeed, using the fact that $Z(0) = 0$, $\div(Z) = N + O(r)$ and condition \textbf{(a)}, we obtain $\frac{\langle Z(x), \nabla a(x) \rangle}{a(x)(\gamma+1)} = O(r)$, and so, for $r$ small, we have \begin{equation}\label{eq:RESTRICTION} 4-2N + \frac{2\div(Z)}{\gamma+1} + \frac{2\langle Z(x), \nabla a(x)\rangle}{a(x)(\gamma+1)} = 2(2-\frac{\gamma N}{\gamma+1}) + O(r) > 0 \end{equation} since, by assumptin, $\frac{\gamma N}{\gamma + 1} < 2$. Also, on $\partial B_r$: $$ 2-\frac{\langle Z(x), \nu_x\rangle}{r(\gamma+1)} = 2 - \frac {\langle A(x)\frac{x}{|x|}, \frac{x}{|x|}\rangle} {\mu(x)(\gamma + 1)} = 2 - \frac{1}{\gamma+1}> 0.$$ Therefore, since $\beta < 0$, we conclude from \eqref{derivativeOfEeq123} that: \begin{equation} \label{KineticBoundFromBellow} E_\beta'(u_\beta,r) \geq \frac{2}{r^{N-1}} \sum_{i=1}^l \int_{\partial B_r} \frac{\langle A(x)\nabla u_{i,\beta}, \nu_x \rangle^2}{\mu(x)}d\sigma(x)+ O(1)\tilde{E}_\beta(u_\beta,r) + R_\beta(u_\beta, r). \end{equation} Next, we estimate $O(1)\tilde{E}_\beta(u_\beta,r)+R_\beta(u_\beta, r)$. Using \eqref{boundForF} (which implies that $|f(x,u_{i,\beta})|\leq d|u_{i,\beta}|$), the bound $|Z(x)|\leq C|x|$ for $|x|\leq \overline{r}$, the ellipticity condition \textbf{(A1)} and choosing $r<\overline{r}<1$ (which implies $r^{N}<r^{N-1}$), then: \begin{align} |O(1)\tilde{E}_\beta(u_\beta,r)&+R_\beta(u_\beta, r)| \leq O(1)\sum_{i=1}^l\left[\int_{B_r}\left(\frac{1}{r^{N-2}}\langle A(x)\nabla u_{i,\beta} , \nabla u_{i,\beta} \rangle + \frac{1}{r^{N}} u_{i,\beta}^2 \right)dx+ \frac{1}{r^{N-1}} \int_{\partial B_r} u_{i,\beta}^2 \right]\nonumber \\ & \leq O(1) \left[ E_\beta(u_\beta,r) + \sum_{i=1}^l \int_{B_r}f_i(x,u_{i,\beta})u_{i,\beta} dx + \frac{1}{r^{N-1}}\sum_{i=1}^l \int_{\partial B_r} u_{i,\beta}^2 \right] \nonumber\\ & \leq O(1) \left[ E_\beta(u_\beta,r) + H_\beta(u_\beta,r)+ \frac{d r^2}{r^{N}}\sum_{i=1}^l\int_{B_r}u_{i,\beta}^2dx.\right] \label{remainderFinalBound}. \end{align} On the other hand, using Poincar\'e's inequality (Lemma \ref{PoincareIneq}) and reasoning as above, \begin{align*} \frac{1}{r^{N}}\sum_{i=1}^l \int_{B_r} u_{i,\beta}^2(x) dx &\leq \sum_{i=1}^l\frac{1}{N-1}\left( \frac{1}{r^{N-2}} \int_{B_r} |\nabla u_{i,\beta}(x)|^2 dx + \frac{1}{r^{N-1}} \int_{\partial B_r} u_{i,\beta}^2(x)d\sigma(x) \right)\\ &\leq O(1)\Big(E_\beta(u_\beta,r) + H_\beta(u_\beta,r) + \frac{r^2d}{r^N} \int_{\partial B_r} u_{i,\beta}^2 dx \Big). \end{align*} Thus for $r$ small enough we conclude that: \begin{equation} \label{BoundOfMass} \frac{1}{r^{N}}\sum_{i=1}^l \int_{B_r} u_{i,\beta}^2(x) dx \leq O(1) \Big(E_\beta(u_\beta,r) + H_\beta(u_\beta,r) \Big). \end{equation} In conclusion, by combining \eqref{KineticBoundFromBellow}, \eqref{remainderFinalBound} and \eqref{BoundOfMass} we have: $$ E_\beta'(u_\beta,r) \geq \frac{2}{r^{N-1}} \sum_{i=1}^l \int_{\partial B_r}\frac{\langle A(x)\nabla u_{i,\beta}, \nu_x \rangle^2}{\mu(x)}d\sigma(x)+ O(1)\Big(E_\beta(u_\beta,r) + H_\beta(u_\beta,r)\Big). $$ Using Holder's inequality: \begin{align*} N_\beta'(u_\beta,r) &=\frac{E_\beta'(u_\beta,r)H_\beta(u_\beta,r) - E_\beta(u_\beta,r)H_\beta'(u_\beta,r)}{H_\beta^2(u_\beta,r)}\\ &\geq \frac{2}{H_\beta^2(u_\beta,r)r^{2N-3}} \Bigg[ \Big( \sum_{i=1}^l \int_{\partial B_r} \frac{\langle A(x) \nabla u_{i,\beta} , \nu \rangle^2}{\mu(x)}d\sigma \Big) \Big( \sum_{i=1}^l \int_{\partial B_r } \mu(x) u_{i,\beta}^2 d\sigma \Big)\\ &- \Big( \sum_{i=1}^l \int_{\partial B_r} u_{i,\beta} \langle A(x) \nabla u_{i,\beta} , \nu \rangle d\sigma \Big)^2 \Bigg]\\ & + \frac{1}{H_\beta^2(u_\beta,r)}\Bigg[ O(1) H_\beta(u_\beta,r) \Big( E_\beta(u_\beta,r) + H_\beta(u_\beta,r) \Big) + O(1)E_\beta(u_\beta,r)H_\beta(u_\beta,r) \Bigg] \\ &\geq \frac{1}{H_\beta^2(u_\beta,r)} \Bigg[ O(1)H_\beta(u_\beta,r) \Big( E_\beta(u_\beta,r) + H_\beta(u_\beta,r) \Big) + O(1)E_\beta(u_\beta,r)H_\beta(u_\beta,r) \Bigg] \\ &\geq -C\Big(\frac{E_\beta(u_\beta,r)}{H_\beta(u_\beta,r)} + 1\Big) = -C\Big(N_\beta(u_\beta,r)+1\Big), \end{align*} where in the last inequalities we used the fact (proved in Step 1.) that $E_\beta(u_\beta,r) + H_\beta(u_\beta,r)$ is positive. Thus we conclude $N_\beta'(u_\beta,r)\geq -C\left(N_\beta(u_\beta,r)+1\right)$ whenever $r$ is small and $H_\beta(u_\beta,r) \neq 0$. \noindent \textbf{Step 3.} There exists $\overline{r}>0$ small enough such that $H_\beta(u_\beta, r) \neq 0$ for $r \in ]0,\overline{r}[$. Indeed, we have $H_\beta'(u_\beta,r)=a_\beta(r)H_\beta(u_\beta,r)$, where (by Lemma \ref{derivOfH}) $a_\beta(r)=\frac{2}{r}N_\beta(u_\beta,r)+O(1)$. Then, by the existence and uniqueness theorem for this ODE, and since $u_\beta>0$, we have $H_\beta(u_\beta,r)>0$ for sufficiently small $r>0$. \end{proof} \begin{remark} The restriction $\frac{\gamma N}{\gamma+1}<2$ in Theorem \ref{DesiredTheorem} comes only from the proof of the previous theorem, namely from the necessity of having the inequality \[ 4-2N + \frac{2\div(Z)}{\gamma+1} + \frac{2\langle Z(x), \nabla a(x)\rangle}{a(x)(\gamma+1)} = 2(2-\frac{\gamma N}{\gamma+1}) + O(r) > 0 \] for small $r>0$. \end{remark} We conclude this section with the following result. \begin{lemma} \label{doublingLemma} Under the assumptions of Theorem \ref{almgrenMonotonicity}, there exists $C>0$ \begin{enumerate} \item If there exists $\tilde{r}$ and $R$ such that $N_\beta(u_\beta,r)\leq \lambda$ for all $0 \leq \tilde{r} \leq r \leq R \leq \overline{r}$, then: $$ r \mapsto \frac{H_\beta(u_\beta,r)}{r^{2\lambda}}e^{-Cr} $$ is monotone nonincreasing for $\tilde{r} \leq r\leq R$. \item If there exists $\tilde{r}$ and $R$ such that $N_\beta(u_\beta,r)\geq \gamma$ for all $0 \leq \tilde{r} \leq r \leq R \leq \overline{r}$, then: $$ r \mapsto \frac{H_\beta(u_\beta,r)}{r^{2\gamma}}e^{Cr}. $$ is monotone nondecreasing \end{enumerate} \end{lemma} \begin{proof} (1) We know by Lemma \ref{derivOfH} that there exists $C$ such that: \begin{align*} \frac{d}{dr}\log\left( H_\beta(u_\beta,r) \right) \leq \frac{2N_\beta(u_\beta,r)}{r} + C \leq \frac{2\lambda}{r} + C \end{align*} from this the result follows. The proof of (2) is similar. \end{proof} \section{An Alt-Caffarelli-Friedman type monotonicity formula ($N\geq 3$)} \label{chapter:resultsChap4} The purpose of this section is to prove an Alt-Caffarelli-Friedman type monotonicity formula, see Theorem \ref{AltCaffMonotonicity} below. For simplicity, we focus on systems of type \eqref{equation} with $\gamma = 1$; however, it should be clear how to adapt our proofs to the general case $\gamma\geq 1$, see Remark \ref{rem_finalSection4} for the details. Let $u_n = (u_{1,n}, u_{2,n},...,u_{l,n})$ be a nonnegative solution of the system: \begin{equation} \label{altCafEq} -\div(A_n(y)\nabla u_{i,n}) = f_{i,n}(y,u_{i,n}) + M_n a_n(y)\sum_{j\neq i} |u_{j,n}|^{2}u_{i,n} \qquad y \in \Omega_n \end{equation} where $M_n < 0$. We take $N\geq 3$ through this section, see Remark \ref{rem:dimension} below on why this is not restrictive. Before stating the monotonicity formula, we need some preparations. \noindent \textbf{Notation. } Within this section, we let $ \nabla u_{i,n}(y)$ denote the usual gradient in $\mathbb{R}^N$, while the gradient on a sphere is denoted by $$ \nabla_{\theta}u_{i,n}(y) := \pi_{T_y (\partial B_{|y|})} ( \nabla u_{i,n} (y) ), $$ where $\pi_{T_y (\partial B_{|y|})}$ is the projection of $\nabla u_{i,n}(y)$ onto $T_y\partial(B_{|y|}(0))$. Given $y \neq 0$, the vector $\nu_y := \frac{y}{|y|}$ is the exterior normal of the sphere $\partial(B_{|y|}(0))$ at $y$. We also define: $$\mu_n(y) := \langle A_n(y)\nu_y,\nu_y\rangle =\left\langle A_n(y) \frac{y}{|y|}, \frac{y}{|y|} \right\rangle \qquad \forall y \neq 0.$$ Fix the following objects: \begin{itemize} \item $c_n$ a sequence such that $c_n \rightarrow 0$, \item $R_n>1$ a sequence of radius such that $B_{R_n}\subset \Omega_n$, \item $\epsilon_n > 0$ a sequence of positive numbers, \item $\lambda, w > 0$ positive numbers, \end{itemize} and assume the following conditions: \begin{itemize} \item[($h_0$)]$\sup_{y \in B_r(0)}\|A_n(y)-Id\| \leq c_nr$ and $\sup_{y \in B_r(0)}\|DA_n(y)\| \leq c_n$ for all $r \in ]1,R_n[$;\\ \item[($h_1$)] $\epsilon_n R_n^2 \leq (\frac{N-2}{2})^2-\delta$ for some $\delta>0$;\\ \item[($h_2$)] $|f_{i,n}(y,u_{i,n})| \leq \frac{1}{2}\epsilon_n \mu_n(y) |u_{i,n}|$;\\ \item[($h_3$)] $ \displaystyle \frac{1}{\lambda} \leq \frac { \int_{\partial B_r} \mu_n(y) u_{1,n}^2 d\sigma(y) } { \int_{\partial B_r} \mu_n(y) u_{2,n}^2d\sigma(y) } \leq \lambda $ and $ \displaystyle \frac{ 1}{r^{N-1}}\int_{\partial B_r}\mu_n(y)u_{i,n}^2d\sigma(y)\geq w$ for all $r \in ]1,R_n[$ and $i=1,2$;\\ \item[($h_4$)] $c_nR_n \rightarrow 0$;\\ \item[($h_5$)] There exists $C>0$ such that, for all $s,r \in [0,R_n]$ satisfying $s\leq r$ and $i\in \{1,...,l\}$, we have \[ \frac{1}{s^{N-1}}\int_{\partial B_s} u_{i,n}^2(y)d\sigma(y) \leq \frac{C}{r^{N-1}}\int_{\partial B_r}u_{i,n}^2(y)d\sigma(y).\] \end{itemize} \begin{lemma} \label{divClaim} Under ($h0$)--($h5$), there exists $\alpha > 0$, independent of $n$ and $r$, such that: \begin{equation*} \label{alphaEquationFTW} \int_{B_r}u_{1,n}^2\div(A_n(y)\nabla|y|^{2-N})dy \leq \frac {\alpha c_n(N-2)}{r^{N-2}} \int_{\partial B_r}\mu_n(y)u_{1,n}^2d\sigma(y). \end{equation*} \end{lemma} \begin{proof} We know that $\Delta |y|^{2-N} = -C_N\delta$. Thus, we have: \begin{align*} \int_{B_r} u_{1,n}^2\div(A_n(y)\nabla |y|^{2-N})dy &= -C_Nu_{1,n}^2(0) + \int_{B_r} u_{1,n}^2\div\big((A_n(y)-Id\big)\nabla |y|^{2-N})dy \\ &\leq \int_{B_r} u_{1,n}^2 \div\big((A_n(y)-Id)\nabla|y|^{2-N}\big) dy. \end{align*} To estimate the integral above, we observe that there exists $\tilde{C}>0$ such that: \begin{align*} |\div ( &(A_n(y)-Id) \nabla |y|^{2-N}) | \leq\left| \sum_{i=1}^N \frac{\partial}{\partial y_i}\left((A_n(y)-Id) \nabla |y|^{2-N}\right)_i\right| \\ &=\left| \sum_{i=1}^N\left( \frac{\partial}{\partial y_i} (A_n(y)-Id) \nabla |y|^{2-N}\right)_i+ \left( (A_n(y)-Id) \frac{\partial}{\partial y_i} \nabla |y|^{2-N} \right)_i \right| \leq \frac{\tilde{C}c_n }{|y|^{N-1}}. \end{align*} since $\|A_n(y)-Id\|_{L^\infty(B_r)}\leq c_n|y|$ and $\|DA_n(y)\|_{L^\infty(B_r)} \leq c_n$, by ($h0$). Thus, we obtain \begin{align*} \left| \int_{B_r} u_{1,n}^2\div((A_n(y)-Id)\nabla|y|^{2-N}) dy \right| &\leq \tilde{C}c_n \int_{B_r}\frac{u_{1,n}^2}{|y|^{N-1}}dy = \tilde{C}c_n\int_0^r \int_{\partial B_s} \frac{u_{1,n}^2}{s^{N-1}}d\sigma(y)ds\\ &\leq \tilde{C} c_n \int_{0}^r \int_{\partial B_s} \frac{u_{1,n}^2}{r^{N-1}} d\sigma(y) ds \leq \tilde C C c_nr \int_{\partial B_r} \frac{u_{1,n}^2}{r^{N-1}}d\sigma(y). \end{align*} where we have used hypothesis ($h_5$). Now, since $|\mu_n(y) -1| \leq c_n$, then $1 \leq 2\mu_n(y)$ and so: $$ \int_{B_r}u_{1,n}^2\div(A_n(y)\nabla|y|^{2-N})dy \leq 2\tilde C C (c_nr)\int_{\partial B_r}\mu_n(y)\frac{u_{1,n}^2}{r^{N-1}}d\sigma(y), $$ and we can choose $\alpha = \frac{2\tilde{C} C}{N-2}$. \end{proof} We consider, for the first two components $(u_{1,n},u_{2,n})$ of the solution of \eqref{altCafEq}, the expressions: \begin{align*} J_{1,n}(r)&:= \int_{B_r}\left(\langle A_n(y)\nabla u_{1,n}, \nabla u_{1,n} \rangle - M_n a_n(y) |u_{1,n}|^2|u_{2,n}|^2 - u_{1,n}f_{1,n}(y,u_{1,n})\right)|y|^{2-N} dy,\\ J_{2,n}(r)&:= \int_{B_r}\left(\langle A_n(y)\nabla u_{2,n}, \nabla u_{2,n} \rangle - M_n a_n(y) |u_{1,n}|^{2}|u_{2,n}|^2- u_{2,n}f_{1,n}(y,u_{2,n})\right)|y|^{2-N}dy. \end{align*} For $\alpha>0$ as in Lemma \ref{divClaim}, we define \begin{align} \Lambda_{1,n}(r)&:= \frac{\displaystyle r^2\int_{\partial B_r} \langle A_n(y) \nabla_{\theta} u_{1,n}, \nabla_{\theta} u_{1,n} \rangle - \frac{\displaystyle \langle A_n \nabla_{\theta}u_{1,n}, \nu_y \rangle^2}{\mu_n(y)} - M_na_n(y)|u_{2,n}|^{2} |u_{1,n}|^2 -u_{1,n}f_{1,n}(y,u_{1,n})}{\displaystyle \int_{\partial B_r}(1+\alpha rc_n)u_{1,n}^2(y)\mu_n(y)d\sigma(y)} \label{LambdaAltCaf1}\\ \Lambda_{2,n}(r)&:= \frac{\displaystyle r^2\int_{\partial B_r} \langle A_n(y) \nabla_{\theta} u_{2,n}, \nabla_{\theta} u_{2,n} \rangle - \frac{\langle A_n \nabla_{\theta}u_{2,n}, \nu_y \rangle^2}{\mu_n(y)} - M_na_n(y)|u_{1,n}|^{2} |u_{2,n}|^2 -u_{2,n}f_{2,n}(y,u_{2,n})}{\displaystyle \int_{\partial B_r}(1+\alpha rc_n)u_{2,n}^2(y)\mu_n(y)d\sigma(y)}.\nonumber \end{align} In addition to ($h0$)--($h5$), we assume \begin{itemize} \item[($h_6$)] $J_{i,n}(r)>0$ and $\Lambda_{i,n}(r)>0$ for every $r \in ]1, R_n[$;\\ \end{itemize} The main objective of this section is to prove the following theorem: \begin{thm} \label{AltCaffMonotonicity} Let $u_n=(u_{1,n},...,u_{l,n})$ be a nonnegative solution of equation \eqref{altCafEq}, and that $A_n$ satisfy \textbf{(A1)} and \textbf{(A2)} for some $\theta>0$, $M>0$. Assume that ($h_0$)-($h_6$) hold true. Then, for any $0<\eta<\frac{1}{4}$, there exists a positive constant $C= C((R_n),(c_n), (\epsilon_n), \lambda, w, N,\eta)$, such that: $$r \mapsto \frac{J_{1,n}(r)J_{2,n}(r)}{r^4} e^{ -C|M_n|^{-\eta} r^{-2\eta} + C\epsilon_nr^2 + C c_n r } $$ is monotone nondecreasing for $r \in ]1,R_n[$. \end{thm} \begin{remark}\label{eq:differencesfromLaplace} This result is inspired by \cite[Theorem 3.14]{SoaveZilio}, which deals with a system with the Laplace operator. Comparing the hypothesis of this reference with ours, besides the technical changes, the main difference is condition ($h_5$). This condition was essential in the proof of Lemma \ref{divClaim}; this key bound is straightforward in case the operator ir the Laplacian, since $-\Delta(\frac{1}{|y|^{N-2}}) = C\delta$; in our case, we need ($h_5$) when approximating the fundamental solution of the operator: $-\div(A_n(x)\nabla (\cdot))$ by the fundamental solution of the Laplacian operator, recalling also that $A_n(y) \sim Id$ for $y$ close to zero. Later on we will apply Theorem \ref{AltCaffMonotonicity} to the blowup sequence $(v_n)$ introduced in Section \ref{chapter:background}, and condition ($h5$) will be a consequence of the Almgren's-monotonicity formula, see Section \ref{sec:conditions} below for the details. \end{remark} \begin{remark} In condition ($h4$) we assume that $r \in ]1,R_n[$, but we notice that the lower bound $1$ can be replaced with any other positive. \end{remark} The rest of the section is devoted to the proof of Theorem \ref{AltCaffMonotonicity}, which we divide into several lemmas. Before we start, we define: \begin{equation} \label{GammaCaffareliDef} \gamma(t):= \sqrt{(\frac{N-2}{2})^2 + t} - \frac{N-2}{2}, \end{equation} a natural quantity within this context (see \cite{FriedmanSphere}), which satisfies $\gamma(t)^2 + (N-2)\gamma(t) = t$. The following lemma clarifies the definition of $\Lambda_{i,n}(r)$, its relation with $J_{i,n}(r)$ and the need of Lemma \ref{divClaim}. \begin{lemma} \label{derivativeLemma} Let $u = (u_{1,n},...,u_{l,n})$ be a positive solution of \eqref{altCafEq}, assume hypothesis ($h_0$)-($h_6$), and \textbf{(A1)}. Then, for $i=1,2$, \begin{equation} \label{AltCaffJequation} J_{i,n}(r) \leq \frac{r}{2\gamma(\Lambda_i(r))} \int_{\partial B_r }\left(\langle A_n(y)\nabla u_{i,n}, \nabla u_{i,n} \rangle - M_n a_n(y) |u_{1,n}|^{2}|u_{2,n}|^2 - u_{i,n}f_{i,n}(y,u_{i,n})\right)|y|^{2-N}. \end{equation} \end{lemma} \begin{proof} We prove the statement for $i=1$. \noindent \textbf{Step 1.} Check that \begin{equation}\label{AltCaffJequation_aux} J_{1,n}(r) = \int_{B_r}\frac{1}{2} u_{1,n}^2\div( A_n(y) \nabla |y|^{2-N} ) + \int_{\partial B_r} \left(\frac{1}{r^{N-2}} u_{1,n}\langle A_n(y)\nabla u_{1,n}, \nu_y \rangle+ \frac{N-2}{2 r^{N-1}} \mu_n(y) u_{1,n}^2\right). \end{equation} First, we test the equation \eqref{altCafEq} for $u_{1,n}$ by $u_{1,n}(y)|y|^{2-N}$, obtaining \begin{align*} \int_{B_r} & \langle A_n(y) \nabla u_{1,n} , \nabla u_{1,n} \rangle |y|^{2-N} dy = -\int_{B_r} u_{1,n}(y) \langle A_n(y)\nabla u_{1,n}, \nabla |y|^{2-N}\rangle dy \\ &+ \int_{B_r} \left( \sum_{i\neq 1}^l M_n a_n(y) |u_{i,n}|^{2}|u_{1,n}|^2 + u_{1,n}f_{1,n}(y,u_{1,n}) \right) |y|^{2-N}dy + \frac{1}{r^{N-2}} \int_{\partial B_r} u_{1,n} \langle A_n(y)\nabla u_{1,n}, \nu_y \rangle . \end{align*} Using the above equation and the fact that $M_n<0$, $a(y)>0$ and the fact that $\nabla (|y|^{2-N}) = \frac{(2-N) y}{r^N} = \frac{(2 - N)}{r^{N-1}} \nu_y$ for $y \in \partial B_r$ , we conclude that: \begin{align*} J_{1,n}(r) &\leq \int_{B_r} \left(\langle A_n(y)\nabla u_{1,n}, \nabla u_{1,n} \rangle - \sum_{i \neq 1}^lM_n a_n(y) |u_{i,n}|^{2}|u_{1,n}|^2 - u_{1,n}f_{1,n}(y,u_{1,n})\right)|y|^{2-N}dy \\ &= -\int_{B_r}u_{1,n}\langle A_n(y)\nabla u_{1,n}, \nabla |y|^{2-N}\rangle dy + \frac{1}{r^{N-2}}\int_{\partial B_r} u_{1,n}\langle A_n(y)\nabla u_{1,n}, \nu_y \rangle d\sigma(y)\\ &= -\frac{1}{2}\int_{B_r}\langle \nabla( u_{1,n}^2), A_n(y) \nabla |y|^{2-N}\rangle dy+ \frac{1}{r^{N-2}}\int_{\partial B_r} u_{1,n}\langle A_n(y)\nabla u_{1,n}, \nu_y \rangle d\sigma(y)\\ &= \int_{B_r}\frac{1}{2} u_{1,n}^2\div( A_n(y) \nabla |y|^{2-N} ) + \frac{1}{r^{N-2}}\int_{\partial B_r} u_{1,n}\langle A_n(y)\nabla u_{1,n}, \nu_y \rangle d\sigma(y)\\ &\quad \quad+ \frac{N-2}{2 r^{N-1}} \int_{\partial B_r} \langle \nu_y, A_n(y) \nu_y \rangle u_{1,n}^2 d\sigma(y) \\ &= \int_{B_r}\frac{1}{2} u_{1,n}^2 \div( A_n(y) \nabla |y|^{2-N} ) dy + \frac{1}{r^{N-2}} \int_{\partial B_r} u_{1,n} \langle A_n(y)\nabla u_{1,n}, \nu_y \rangle d\sigma(y), \end{align*} which implies \eqref{AltCaffJequation_aux}, by recalling that $\mu_n(y) = \langle A_n(y) \nu_y, \nu_y \rangle$. \noindent \textbf{Step 2.} Conclusion of inequality \eqref{AltCaffJequation}. Using Lemma \ref{divClaim} in inequality \eqref{AltCaffJequation_aux}, we obtain \begin{equation} \label{randomEqJ} J_{1,n}(r) \leq \frac{1}{r^{N-2}} \int_{\partial B_r} u_{1,n} \langle A_n(y)\nabla u_{1,n}, \nu_y \rangle d\sigma (y) + \frac{N-2}{2 r^{N-1}} \int_{\partial B_r} (1+\alpha c_n r) \mu_n(y) u_{1,n}^2 d\sigma(y). \end{equation} Now, by Young's inequality, one obtains: \begin{align*} &\int_{\partial B_r} u_{1,n} \langle A_n(y)\nabla u_{1,n}, \nu_y \rangle d\sigma(y) \\ \leq& \frac{\gamma(\Lambda_{1,n}(r))}{2r}\int_{\partial B_r}u_{1,n}^2\mu_n(y)d\sigma(y)+\frac{r}{2\gamma(\Lambda_{1,n}(r))}\int_{\partial B_r}\frac{\langle A_n(y)\nabla u_{1,n}, \nu_y \rangle^2}{\mu_n(y)}d\sigma(y) \\ \leq& \frac{\gamma(\Lambda_{1,n}(r))}{2r} \int_{\partial B_r} (1+\alpha c_nr) u_{1,n}^2 \mu_n(y) d\sigma(y) + \frac{r}{2\gamma(\Lambda_{1,n}(r))} \int_{\partial B_r} \frac{ \langle A_n(y)\nabla u_{1,n}, \nu_y \rangle^2 }{ \mu_n(y) } d\sigma(y). \end{align*} Applying this inequality to equation \eqref{randomEqJ}: \begin{align*} J_{1,n}(r) \leq&\frac{1}{2r^{N-1}\gamma(\Lambda_{1,n}(r))} \Bigg[ \left(\gamma(\Lambda_{1,n}(r))^2 + (N-2)\gamma(\Lambda_{1,n}(r)) \right) \int_{\partial B_r}(1+\alpha c_nr)u_{1,n}^2\mu_n(y)d\sigma(y)\\ &+ r^2\int_{\partial B_r}\frac{\langle A_n(y)\nabla u_{1,n}, \nu_y \rangle^2}{\mu_n(y)}d\sigma(y)\Bigg]. \end{align*} Using the fact that $\gamma(\Lambda_{1,n}(r))^2 + (N-2)\gamma(\Lambda_{1,n}(r)) = \Lambda_{1,n}(r)$ then by the definition given in \eqref{LambdaAltCaf1} we obtain: \begin{align} J_{1,n}(r) \leq& \frac{r^2}{2r^{N-1} \gamma(\Lambda_{1,n}(r))} \Bigg( \int_{\partial B_r} \left( \langle A_n(y) \nabla_{\theta} u_{1,n}, \nabla_{\theta} u_{1,n} \rangle - \frac{\langle A_n(y) \nabla_{\theta}u_{1,n}, \nu_y \rangle^2}{\mu_n(y)} + \frac{ \langle A_n(y)\nabla u_{1,n}, \nu_y \rangle^2 }{ \mu_n(y) } \right)d\sigma(y) \nonumber \\ &\quad\quad+ \int_{\partial B_r}\left( - M_na_n(y)|u_{1,n}|^{2} |u_{2,n}|^{2} -u_{1,n}f_{1,n}(y,u_{1,n}) \right) d\sigma(y) \Bigg). \label{anothRandomEqJ} \end{align} Now, to compute $\frac{\langle A_n(y)\nabla u_{1,n}(y), \nu_y \rangle^2}{\mu_n(y)}$, we use the following auxiliary equations: $$ \nabla u_{1,n} = \nabla_{\theta} u_{1,n} + (\partial_{\nu} u_{1,n}) \nu_y, \qquad \langle A_n(y)\nabla u_{1,n}, \nu_y \rangle = \langle A_n(y) \nabla_\theta u_{1,n}, \nu_y \rangle + \mu_n(y)(\partial_\nu u_{1,n}). $$ With them, we obtain: $$ \frac{\langle A_n(y)\nabla u_{1,n}, \nu_y \rangle^2}{\mu_n(y)} = \frac{ \langle A_n(y) \nabla_\theta u_{1,n}, \nu_y \rangle^2 } { \mu_n(y) } + 2(\partial_\nu u_{1,n}) \langle A_n(y) \nabla_\theta u_{1,n}, \nu_y \rangle + (\partial_\nu u_{1,n})^2 \mu_n(y). $$ We also notice: \begin{align*} \langle A_n(y)\nabla u_{1,n}, \nabla u_{1,n} \rangle =& \langle A_n(y)\nabla_{\theta} u_{1,n}, \nabla_{\theta} u_{1,n} \rangle+ 2\langle A_n(y)\nabla_{\theta}u_{1,n}, \nu_y \rangle (\partial_\nu u_{1,n}) + (\partial_\nu u_{1,n})^2\mu_n(y), \end{align*} and from this we conclude that: $$ \langle A_n(y)\nabla u_{1,n}, \nabla u_{1,n} \rangle = \langle A_n(y) \nabla_{\theta} u_{1,n}, \nabla_{\theta} u_{1,n} \rangle - \frac{\langle A_n(y) \nabla_{\theta}u_{1,n}, \nu_y \rangle^2}{\mu_n(y)} + \frac{\langle A_n(y)\nabla u_{1,n}, \nu_y \rangle^2}{\mu_n(y)}.$$ Thus, applying this equality to equation \eqref{anothRandomEqJ}, we conclude \eqref{AltCaffJequation}, as wanted. \end{proof} Before we proceed, it is important at this point to simplify the notation of $\Lambda_{i,n}(r)$. We rewrite one of the terms as follows: \begin{align*} \Big\langle A_n(y) \nabla_{\theta} u_{1,n}, \nabla_{\theta} u_{1,n} \Big\rangle &- \frac{\langle A_n(y) \nabla_{\theta}u_{1,n}, \nu_y \rangle^2}{\mu_n(y)}\\ =& \Big\langle A_n(y) \nabla_{\theta} u_{1,n} , \nabla_{\theta} u_{1,n} \Big\rangle - \Big\langle \frac{\langle A_n(y) \nu_y, \nabla_{\theta}u_{1,n} \rangle}{\mu_n(y)}A_n(y) \nu_y, \nabla_{\theta}u_{1,n} \Big\rangle \\ =& \Big\langle A_n(y) \nabla_{\theta} u_{1,n} -\frac{\langle A_n(y) \nu_y, \nabla_{\theta}u_{1,n} \rangle}{\mu_n(y)}A_n(y) \nu_y, \nabla_{\theta} u_{1,n} \Big\rangle. \end{align*} \begin{definition}\label{definition:operatorB} For $y \neq 0$, we define the operator $B_n(y)$ by: $$ B_n(y) v := \left( A_n(y) v - \frac{\langle A_n(y) \nu_y, v \rangle}{\mu_n(y)}A_n(y) \nu_y\right) $$ and write, for $i,j\in \{1,2\}$, $i\neq j$, \begin{equation}\label{eq:equivalent_LAMBDA} \Lambda_{i,n}(r) = \frac{r^2\int_{\partial B_r} \langle B_n(y) \nabla_{\theta} u_{i,n}, \nabla_{\theta} u_{i,n} \rangle - M_na_n(y)|u_{i,n}|^{2} |u_{j,n}|^2 + u_{i,n}f_{i,n}(y,u_{i,n})d\sigma(y)}{\int_{\partial B_r}(1+\alpha(rc_n))\mu_n(y)u_{i,n}^2d\sigma(y)} \end{equation} \end{definition} \noindent It is straightforward to see that $B_n(y)$ is a symmetric operator. Moreover, for all $\xi \in \mathbb{R}^N$, we have that $B_n(y)\xi \in T_y (\partial B_{|y|})$. This is the case since: \begin{align*} \langle B_n(y) \xi, \nu_y \rangle = \langle A_n(y) \xi - \frac{\langle A_n(y)\nu_y, \xi \rangle}{\mu_n(y)} A_n(y) \nu_y, \nu_y \rangle = \langle A_n(y) \xi, \nu_y \rangle - \frac{\langle A_n(y) \nu_y, \xi \rangle \mu_n(y)}{\mu_n(y)} = 0. \end{align*} The following result shows what is the ellipticity constant of $B_n$, and provides a bound for its distance to the identity operator. \begin{lemma} \label{MatrixBoundsOnB} Suppose that \textbf{(A1)} is satisfied for the sequence $A_n$: \begin{equation} \label{ellipticityCaf1} \langle A_n(y) \xi, \xi \rangle \geq \theta |\xi|^2 \qquad \forall \xi \in \mathbb{R}^N, \end{equation} and ($h0$) and ($h4$) hold true. Then there exist $C,\tilde{\theta}>0$ depending only on $\theta$, $M$ and $N$ such that, for all $y \in B_{R_n}$, \begin{equation} \label{ineq2AltCafMatrixB} \langle B_n(y)\xi, \xi \rangle \geq \tilde{\theta}|\xi|^2 \quad \forall \xi \in T_y \left( \partial B_{|y|} \right), \quad \text{ and } \quad \| Id|_{T_{y}\partial B_{|y|}} - B_n(y) \| \leq C \|Id_{\mathbb{R}^N}-A_n(y)\|. \end{equation} \end{lemma} \begin{proof} Given $\xi \in T_y (\partial B_{|y|})$, we obtain: \begin{align*} |(B_n(y) - &Id_{T_y (\partial B_{|y|})} ) \xi| = | ( A_n(y)\xi - \frac { \langle A_n(y) \nu_y, \xi \rangle } { \mu_n(y) } ) - \xi | = | (A_n(y) - Id)\xi - \frac { \langle \nu_y, A_n(y)\xi \rangle } { \mu_n(y) } |\\ &\leq |(A_n(y)-Id)\xi| + |\frac{\langle \nu_y, A_n(y)\xi \rangle}{\mu_n(y)}| = |(A_n(y)-Id)\xi| + |\frac{\langle \nu_y, (A_n(y)-Id)\xi \rangle}{\mu_n(y)}|\\ &\leq \sqrt{N} \|A_n(y)-Id\|\cdot|\xi| + \frac{\sqrt{N}}{\mu_n(y)} \|A_n(y)-Id\|\cdot|\xi| \leq \sqrt{N} (1 + \frac{1}{\theta}) \|A_n(y)-Id\| \cdot |\xi|, \end{align*} where we used $\langle \xi, \nu_y \rangle = 0$, and the ellipticity constant of \eqref{ellipticityCaf1} to obtain $\frac{1}{\mu_n(y)} \leq \frac{1}{\theta}$. Taking $C =\sqrt{N}(1 + \frac{1}{\theta})$ concludes the second statement in \eqref{ineq2AltCafMatrixB}. Regarding the first statement in \eqref{ineq2AltCafMatrixB}, given $\xi \in T_y(\partial B_{|y|})$: \begin{align*} \langle B_n(y) \xi, \xi \rangle&= \Big\langle A_n(y)\xi - \frac{ \langle A_n(y)\nu_y , \xi \rangle A_n(y)\nu_y } { \mu_n(y) } , \xi \Big\rangle= \langle A_n(y)\xi,\xi \rangle - \frac{\langle A_n(y)\nu_y,\xi\rangle^2}{\mu_n(y)}\\ &= \langle A_n(y)\xi,\xi \rangle - \frac{\langle (A_n(y)-Id)\nu_y,\xi\rangle^2}{\mu_n(y)} \geq \theta |\xi|^2 -\frac{\langle (A_n(y)-Id)\nu_y,\xi\rangle^2}{\mu_n(y)} \\ &\geq \theta |\xi|^2 - \frac{\sqrt{N}}{\theta} \|A_n(y)-Id\|\cdot |\xi|^2\geq \theta |\xi|^2 - \frac{\sqrt{N}}{\theta} c_nR_n |\xi|^2\geq \frac{\theta}{2}|\xi|^2, \end{align*} where we used $(h_0)$ and $(h_5)$ in the last inequalities. \end{proof} Given $\lambda_n > 0$ and a sequence $\tilde{c}_n \rightarrow 0$, define, for each $n$ the subspace of $(H^1(\partial B_1))^2$ given by: \begin{equation*} H_{\lambda_n,\tilde{c}_n} = \Big\{ (u,v) \in (H^1(\partial B_1))^2: \int_{\partial B_1}(1+\alpha \tilde{c}_n)\mu_n(y)u^2dy = 1 ;\,\, \int_{\partial B_1}(1+\alpha \tilde{c}_n)\mu_n(y)v^2 dy = \lambda_n \Big\}. \end{equation*} A fundamental result in the proof of the classical Alt-Caffarelli-Friedman's monotonicity formula \cite{FriedmanSphere} is the following Friedman-Hayman inequality \cite{FriedlanHayman}: \begin{equation}\label{eq:FH_ineq} \gamma\left( \frac{\int_{\partial B_1} |\nabla_\theta f|^2 d\sigma(y)} {\int_{\partial B_1}f^2d\sigma(y)} \right) + \gamma\left( \frac{\int_{\partial B_1} |\nabla_\theta g|^2 d\sigma(y)}{\int_{\partial B_1} g^2d\sigma(y)} \right)\geq 2,\ \ \text{ for every $f,g\in H^1(\partial B_1)$ with $fg\equiv 0$,} \end{equation} where $\gamma$ is defined in \eqref{GammaCaffareliDef}. Inspired by \cite[Lemma 4.2]{SphereDensityAlt} and \cite[Lemma 3.10]{SoaveZilio}, we prove the following result. \begin{lemma} \label{sphereLemma} Fix $\overline{\lambda}>1$ and let $\delta$, $\tilde{c}$, $\tilde{M}$, $\tilde{\theta}$ be positive constants. Then, for every $0 <\eta < \frac{1}{4}$, there exists $C = C(N, \overline{\lambda}, \tilde{c}, \delta, \eta, \alpha, \tilde{\theta}, \tilde{M})>0$ such that, whenever: \begin{equation}\label{MatrixSphereBoundsCaffareli0} \frac{1}{\overline{\lambda}} < \lambda_n < \overline{\lambda}, \qquad 0 \leq \epsilon_n < (\frac{N-2}{2})^2 - \delta, \qquad \tilde{c}_n \to 0, \ \tilde{c}_n \leq \tilde{c}, \qquad k_n\geq0, \end{equation} and $\tilde{B}_n$ is a sequence of symmetric operators such that \begin{equation} \label{MatrixSphereBoundsCaffareli} \tilde{\theta} |\xi|^2\leq \langle \tilde{B}_n(y)\xi,\xi \rangle \quad \forall \xi \in T_y (\partial B_1), \qquad \sup_{y \in \partial B_1} \|D\tilde{B}_n(y)\| \leq \tilde{M}, \qquad \sup_{y \in \partial B_1} \| \tilde{B}_n(y) - Id|_{T_y \partial B_1} \| \leq \tilde{M} \tilde{c}_n, \end{equation} we have: \begin{multline*} \min_{(u,v)\in H_{\lambda_n,\tilde{c}_n}} \gamma\left( \int_{\partial B_1} \langle \tilde{B}_n(y)\nabla_{\theta} u, \nabla_{\theta} u \rangle + k_nu^2v^2-\epsilon_n \right) +\gamma\left( \frac{\int_{\partial B_1} \langle \tilde{B}_n(y)\nabla_{\theta} v, \nabla_{\theta} v \rangle + k_nu^2v^2-\epsilon_n \lambda_n}{\lambda_n} \right)\\ \geq 2 -C(\epsilon_n + k_{n}^{-\eta} + \tilde{c}_n). \end{multline*} \end{lemma} This proof is substantially harder than the one in the case where the operator is the Laplacian. Indeed, in order to obtain \eqref{eq:FH_ineq}, or to obtain \cite[Lemma 4.2]{SphereDensityAlt}, \cite[Lemma 3.10]{SoaveZilio}, an important part of the argument is to symmetrize the solutions of the underlying minimization problem; this, in particular, allows to conclude directly that the sequence of minimizers is Lipschitz continuous, and that its level sets are circles. The proof in our case is harder and required new ideas: since we are dealing with an operator with variable coefficients, we cannot use a symmetrization argument; instead, we rely on Theorem C, from which we obtain uniform H\"older bounds. We now prove Lemma \ref{sphereLemma}, after which we are able to conclude the proof of Theorem \ref{AltCaffMonotonicity}. \begin{proof}[Proof of Lemma \ref{sphereLemma}] We see that the minimization problem in the theorem is equivalent to minimizing in $H_{1,\tilde c_n}$ (by replacing $v$ with $\frac{v}{\sqrt{\lambda_n}}$): \begin{equation} \label{GammaMinimization} \min_{H_{1,\tilde{c}_n}} \gamma\left( \int_{\partial B_1} \langle \tilde{B}_n(y)\nabla_{\theta} u, \nabla_{\theta} u \rangle + k_n \lambda_n u^2v^2-\epsilon_n \right) + \gamma\left( {\int_{\partial B_1} \langle \tilde{B}_n(y)\nabla_{\theta} v, \nabla_{\theta} v \rangle + k_nu^2v^2-\epsilon_n} \right). \end{equation} The direct method of Calculus of Variations yields the existence of a minimizer $(u_n, v_n)$ of \eqref{GammaMinimization} satisfying: \begin{equation*}\label{GammaMinimization_norm} \int_{\partial B_1}(1+\alpha \tilde{c}_n)\mu_n(y)u_n^2d\sigma(y) = 1, \qquad \int_{\partial B_1}(1+\alpha \tilde{c}_n)\mu_n(y)v_n^2d\sigma(y) = 1. \end{equation*} By eventually replacing $(u,v)$ by $(|u|,|v|)$, we may assume without loss of generality that the minimizer $(u_n,v_n)$ is nonnegative. (notice that $ \int_{\partial B_1}\langle \tilde B_n(y)\nabla_\theta u, \nabla_\theta u \rangle d\sigma(y) = \int_{\partial B_1}\langle \tilde{B}_n(y) \nabla_\theta |u|, \nabla_\theta |u| \rangle d\sigma(y)$). Let: \begin{align} x_n &= \int_{\partial B_1} \left( \langle \tilde{B}_n(y)\nabla_{\theta} u_n, \nabla_{\theta} u_n \rangle + k_n \lambda_n u_n^2v_n^2 \right)d\sigma(y) -\epsilon_n, \label{def_of_x_n}\\ y_n &= {\int_{\partial B_1} \left( \langle \tilde{B}_n(y)\nabla_{\theta} v_n, \nabla_{\theta} v_n \rangle + k_nu_n^2v_n^2 \right) d\sigma(y)-\epsilon_n}.\nonumber \end{align} There exist two Lagrange multipliers $\sigma_{1,n}$ and $\sigma_{2,n}$ such that: \begin{equation} \label{SphereEq} \begin{cases} -\div_{\partial B_1}(\tilde{B}_n(y)\nabla_{\theta} u_n) = -k_n(\lambda_n + \frac{\gamma'(y_n)}{\gamma'(x_n)})u_nv_n^2 + \frac{\sigma_{1,n}}{\gamma'(x_n)}(1+ \alpha \tilde{c}_n)\mu_n(y) u_n\\ -\div_{\partial B_1}(\tilde{B}_n(y)\nabla_{\theta} v_n) = -k_n(1 + \frac{\lambda_n\gamma'(x_n)}{\gamma'(y_n)})v_nu_n^2 + \frac{\sigma_{2,n}}{\gamma'(y_n)}(1+\alpha\tilde{c}_n)\mu_n(y)v_n \end{cases} \end{equation} We divide the proof in several steps. \noindent \textbf{Step 1.} There exists $C = C(N, \tilde{M}, \tilde{\theta}, \delta, \overline{\lambda}, \tilde{c})>0$ such that $\|u_n\|_{H^1(\partial B_1)},\|v_n\|_{H^1(\partial B_1)}\leq C$, $\frac{1}{C}\leq \gamma'(x_n),\gamma'(y_n)\leq C$ and $0 \leq \sigma_{1,n}, \sigma_{2,n}<C$. \vspace{2mm} Fix $\phi = \phi^+ - \phi^- \in H^1(\partial B_1)$ such that $\|\phi^+\|_{H^1(\partial B_1)}^2 = \|\phi^-\|_{H^1(\partial B_1)}^2 = 1$. Then, since $(u_n, v_n)$ is the minimizer for \eqref{GammaMinimization}, and $\phi^+\cdot \phi^- = 0$, we have: \begin{multline*} \gamma\left( \int_{\partial B_1} \left( \langle \tilde{B}_n(y)\nabla_{\theta} u_n, \nabla_{\theta} u_n \rangle + k_n \lambda_n u_n^2v_n^2 \right) - \epsilon_n \right) + \gamma\left( { \int_{\partial B_1} \left(\langle \tilde{B}_n(y)\nabla_{\theta} v_n, \nabla_{\theta} v_n \rangle + k_nu_n^2v_n^2\right) -\epsilon_n } \right)\\ \leq \gamma \Big( \frac{ \int_{\partial B_1} \langle \tilde{B}_n(y) \nabla_{\theta} \phi^+ , \nabla_\theta \phi^+ \rangle - \epsilon_n (1+\alpha \tilde{c}_n) \mu_n(y) (\phi^+)^2 } { \int_{\partial B_1} (1+\alpha \tilde{c}_n)\mu_n(y) (\phi^+)^2 } \Big) \\ + \gamma \Big( \frac{ \int_{\partial B_1} \langle \tilde{B}_n(y) \nabla_{\theta} \phi^- , \nabla_\theta \phi^- \rangle - \epsilon_n (1+\alpha \tilde{c}_n) \mu_n(y) (\phi^-)^2 } { \int_{\partial B_1} (1+\alpha \tilde{c}_n) \mu_n(y) (\phi^-)^2 } \Big). \end{multline*} Therefore, we need a uniform bound on the right-hand-side of the previous inequality; we just bound the term involving $\phi^+$, since the computations for the other term are analogous. We use equations \eqref{MatrixSphereBoundsCaffareli0} and \eqref{MatrixSphereBoundsCaffareli} to conclude \begin{align*} \int_{\partial B_1} \left( \langle \tilde{B}_n(y) \nabla_{\theta} \phi^+ , \nabla_\theta \phi^+ \rangle - \epsilon_n (1+\alpha \tilde{c}) \mu_n(y) \phi^+ \right) &\leq \|\tilde{B}_n(y)\| \cdot \|\phi^+\|^2_{H^1(\partial B_1)} \leq (1+\tilde{M}\tilde c),\\ \int_{\partial B_1} (1+\alpha \tilde{c}_n)\mu_n(y) (\phi^+)^2 &\geq \int_{\partial B_1} \mu_n(y)(\phi^+)^2 \geq \tilde{\theta}\int_{\partial B_1}(\phi^+)^2 > 0. \end{align*} Thus, since $\gamma$ is monotone, there exists $C = C(N, \tilde{c}, \tilde{\theta}, \tilde{M}) = 2\gamma (\frac{(1+\tilde{M}\tilde c)}{\tilde \theta}) >0$ such that $ \gamma(x_n)+\gamma(y_n) \leq C $ for all $n \in \mathbb{N}$. Since $\epsilon_n < (\frac{N-2}{2})^2 - \delta$, and $x_n,y_n\geq -\epsilon_n$, then: $ \delta - (\frac{N-2}{2})^2 \leq x_n, y_n. $ Thus, due to the expression of $\gamma$ in \eqref{GammaCaffareliDef}, there exists $C = C(N, \tilde{c}, \tilde{\theta}, \delta)$ large enough such that: $$ \frac{1}{C} \leq \gamma'(x_n)\leq C, \quad \frac{1}{C} \leq \gamma'(y_n) \leq C, \quad \int_{\partial B_1} \Big( \langle \tilde{B}_n(y) \nabla_\theta u_n , \nabla_\theta u_n \rangle + \lambda_n k_n u_n^2 v_n^2 \Big) d\sigma(y) \leq C. $$ In particular, this implies that each one of the two terms in the last inequality are bounded; using the uniform ellipticity of $\tilde{B}_n(y)$, we obtain that $u_n$ and $v_n$ are uniformly bounded in $H^1(\partial B_1)$ with a bound depending again on $\tilde{\theta}$. Finally, as for the Lagrage's multiplier's, there exists $C = C(N,\tilde{c}, \tilde{\theta}, \delta, \overline{\lambda}, \tilde{M})$ such that $$ 0 \leq \sigma_{1,n} = \gamma'(x_n) \int_{\partial B_1} \Big( \langle \tilde{B}_n(y)\nabla_{\theta}u_n, \nabla_{\theta}u_n \rangle + k_n(\lambda_n+\frac{\gamma'(y_n)}{\gamma'(x_n)})u_n^2v_n^2 \Big) d\sigma(y) \leq C, $$ and similarly for $\sigma_{2,n}$. \noindent \textbf{Step 2.} We check that there exists $C = C(N, \tilde{c}, \tilde{\theta}, \delta, \overline{\lambda}, \alpha, \tilde{M})$ such that $\|u_n\|_{L^\infty(\partial B_1)}, \|v_n\|_{L^\infty(\partial B_1)} \leq C$ and moreover that, given $0 <\beta < 1$, there exists $D = D(N, \tilde{c}, \tilde{\theta}, \delta, \overline{\lambda}, \beta, \tilde M, \alpha)$ such that $(u_nv_n)^{\frac{1}{2}+\frac{1}{2\beta}}(x) \leq D k_n^{-\frac{1}{2}}$ for all $x \in \partial B_1$. In particular: \begin{align*} &x \in \{v_n-u_n\leq 0\} \implies v_n(z)\leq D^\frac{1}{2}k_n^{-\frac{\beta}{2(\beta+1)}},\qquad x \in \{v_n-u_n\geq 0\} \implies u_n(z)\leq D^\frac{1}{2}k_n^{-\frac{\beta}{2(\beta+1)}}. \end{align*} \vspace{2mm} To first show the uniform $L^\infty(\partial B_1)$ bound, we notice that the functions $u_n,v_n$ are nonnegative, and $k_n\geq 0$, so by equation \eqref{SphereEq} we have $$-\div_{\partial B_1} \left(\tilde{B_n}(y)\nabla_{\theta} u_n\right) \leq \frac{C\sigma_{1,n}}{\gamma'(x_n)}u_n, \quad -\div_{\partial B_1}\left(\tilde{B_n}(y)\nabla_{\theta} v_n\right) \leq \frac{C\sigma_{2,n}}{\gamma'(y_n)}v_n. $$ By Step 1, the sequences $\frac{\sigma_{1,n}}{\gamma'(x_n)}$, $\frac{\sigma_{2,n}}{\gamma'(y_n)}$ are uniformly bounded and the functions $u_n,v_n$ are uniformly bounded in $H^1(\partial B_1)$. Then, by a Brezis-Kato-type argument (see for instance \cite[Appendix B.2 B.3]{Struwe}), using the uniform ellipticity of $\tilde{B}_n$, we obtain a uniform $L^\infty(\partial B_1)$ bound on $u_n$ and $v_n$. Next, to prove the second part of Step 2, we suppose by contradiction that there exists a sequence of points $z_n \in \partial B_1$ such that: \begin{equation} \label{blowUpContradiction} k_n^{\frac{1}{2}} \left( v_n^{1+\frac{1}{\beta}}(z_n) u_n^{1+\frac{1}{\beta}}(z_n) \right)^{\frac{1}{2}} \rightarrow \infty. \end{equation} By the uniform boundedness of $(u_n,v_n)$ we have, since $\beta>0$, \begin{equation} \label{kGoToInfinity} k_n \rightarrow \infty ,\qquad k_n \left( v_n(z_n) u_n(z_n) \right) \rightarrow \infty. \end{equation} For each point $z_n \in \partial B_1$, we consider the parametrization of the sphere $\partial B_1$, $\phi_n: \mathbb{R}^{N-1} \rightarrow \partial B_1/\{-z_n\}$ given by the stereographic projection from that point, thus $\phi_n(0) = z_n$. Now fix the sequence $a_n$ for which there exists $C$ such that: $$ \frac{1}{C} \leq a_n := \sqrt{ \frac{ \lambda_n + \frac{ \gamma'(y_n) }{ \gamma'(x_n) } }{ 1 + \lambda_n \frac{ \gamma'(x_n) }{ \gamma'(y_n) } } } = \sqrt { \frac{ \gamma'(y_n) }{ \gamma'(x_n) } } \leq C $$ and take the functions $\tilde{v}_n, \tilde{u}_n: \mathbb{R}^{N-1} \rightarrow \mathbb{R}$: $$\tilde{u}_n(z) = u_n(\phi_n(z)), \qquad \tilde{v}_n(z) = a_nv_n(\phi_n(z)).$$ This change of variables leads to the equation: \begin{equation} \label{SemiPlaneSemiSphereEq} \begin{cases} -\div_{\partial B_1}\left( \tilde{B}_n(y)\nabla_\theta u_n \right)_{y = \phi_n(z)} = -k_n(1 + \frac{\lambda_n \gamma'(x_n)}{\gamma'(y_n)})\tilde{u}_n\tilde{v}_n^2 + \frac{\sigma_{1,n}(1+\alpha\tilde{c}_n)}{\gamma'(x_n)}\mu_n(\phi_n(z))\tilde{u}_n \\ -a_n\div_{\partial B_1}\left( \tilde{B}_n(y)\nabla_\theta v_n \right)_{y=\phi_n(z)} = -k_n(1 + \frac{\lambda_n\gamma'(x_n)}{\gamma'(y_n)})\tilde{v}_n\tilde{u}_n^2 + \frac{\sigma_{2,n}(1+\alpha\tilde{c}_n)}{\gamma'(y_n)}\mu_n(\phi_n(z))\tilde{v}_n \end{cases} \end{equation} By Proposition \ref{DivergenceSphereProp}, in appendix, we know that: $$ \div_{\partial B_1} \left( (\tilde{B}_n(y)\nabla_\theta u_n \right)_{y=\phi_n(z)} = (1+|z|^2)^{N-1} \div_{\mathbb{R}^{N-1}} \left( \frac{1}{4(1+|z|^2)^{N-3}} M_n(z) \nabla_{\mathbb{R}^{N-1}} \tilde{u}_n \right) $$ where $M_n(z) = (d\phi_n)^{-1}_{\phi(z)}\tilde{B}_n(\phi_n(z))(d\phi_n)_z$. For simplicity of notation, we define $\tilde{M}_n(z) := \frac{1}{4(1+|z|^2)^{N-3}}M_n(z)$ and $g(z) = (1+|z|^2)^{N-1}$. This allows us to rewrite equation \eqref{SemiPlaneSemiSphereEq} as: \begin{equation} \label{planeEq} \begin{cases} -\div_{\mathbb{R}^{N-1}} \left( \tilde{M}_n(z) \nabla_{\mathbb{R}^{N-1}} \tilde{u} \right) = -\frac{k_n}{g(z)} (1 + \frac{\lambda_n \gamma'(x_n)}{\gamma'(y_n)})\tilde{u}_n\tilde{v}_n^2 + \frac {\sigma_{1,n}(1+\alpha\tilde{c}_n)} {g(z)\gamma'(x_n)} \mu_n(\phi_n(z))\tilde{u}_n \\ -\div_{\mathbb{R}^{N-1}}\left( \tilde{M}_n(z)\nabla_{\mathbb{R}^{N-1}}\tilde{v}_n \right) = - \frac{k_n}{g(z)} (1 + \frac{\lambda_n\gamma'(x_n)}{\gamma'(y_n)})\tilde{v}_n\tilde{u}_n^2 + \frac { \sigma_{2,n}(1+\alpha\tilde{c}_n) }{ g(z)\gamma'(y_n) } \mu_n(\phi_n(z))\tilde{v}_n. \end{cases} \end{equation} By assumption, $\tilde{B}_n(y)$ has ellipticity constant $\tilde{\theta}$, and $\|D\tilde{B}_n\| \leq \tilde{M}$. Moreover, by Proposition \ref{DivergenceSphereProp} we know that, given the compact set $K = B_2(0) \subset \mathbb{R}^{N-1}$, there exists a constant $\tilde{C} = C(\tilde{M}, K)$ such that: $$ \langle \tilde{M}_n(y) \xi, \xi \rangle_{\mathbb{R}^{N-1}} \geq \frac{1}{4\cdot 5^{N-3}} \theta \langle \xi, \xi \rangle_{\mathbb{R}^{N-1}} \quad \forall \xi,y \in \mathbb{R}^{N-1}, \qquad \|D \tilde{M}_n\| \leq \tilde{C}, $$ and $$ \frac {1} {5^{N-1}} \leq \frac{1}{g(z)} \leq 1 \quad \forall z \in \mathbb{R}^{N-1}. $$ Since $(\tilde{u}_n, \tilde{v}_n)$ are uniformly bounded in $L^\infty$, satisfies the system \eqref{planeEq}, and $\tilde{M}_n$ are uniformly elliptic over $B_2(0) \subset \mathbb{R}^{N-1}$, we are under the assumptions of Theorem C. Therefore, for each $0<\beta<1$, there exists a constant $C_\beta = C(\beta, B_2(0), \theta, M, N, \tilde{\theta}, \tilde{M})$ such that \begin{equation*} \label{AltCaffHolderBound1} \|\tilde{u}_n\|_{C^{0,\beta}(B_1(0))}, \|\tilde{v}_n\|_{C^{0,\beta}(B_1(0))} \leq C_\beta. \end{equation*} Define: \begin{equation*} t_n := \tilde{u}_n(0) + \tilde{v}_n(0); \end{equation*} we claim that both $\tilde{u}_n(0) \rightarrow 0$ and $\tilde{v}_n(0) \rightarrow 0$. We suppose, in view of a contradiction, that $\tilde{v}_n(0)\geq \overline{\delta}>0$ for all $n$. Then, by uniform convergence and boundedness of H\"older norms, there exists a radius $R>0$ small enough such that $$ \inf_{x\in B_{2R}(0)}\tilde{v}_n(x) \geq \frac{\delta}{2}\qquad \text{ for all $n$}. $$ From this, we conclude the differential inequality \begin{align*} -\div_{\mathbb{R}^{N-1}} ( \tilde{M}_n(z) \nabla \tilde{u}_n ) &= -\frac{k_n}{g(z)} (1 + \frac{\lambda_n \gamma'(x_n)}{\gamma'(y_n)})\tilde{u}_n\tilde{v}_n^2 + \frac {\sigma_{1,n}(1+\alpha\tilde{c}_n)} {g(z)\gamma'(x_n)} \mu_n(\phi_n(z))\tilde{u}_n \\ &\leq \big( - \frac{k_n}{g(z)} \frac{\delta^2}{4} + \frac{\sigma_{1,n}(1+\alpha\tilde{c}_n)} {\gamma'(x_n)} \mu_n(\phi_n(z)) \big) \tilde{u}_n\\ & \leq \big( -\frac{k_n}{5^{N-1}} \frac{\delta^2}{4} + C^2(1+\alpha \tilde c)(1+\tilde c)^\frac{1}{2} \big) \tilde{u}_n \qquad \forall z \in B_{2R}(0)\subset \mathbb{R}^{N-1}. \end{align*} From \eqref{kGoToInfinity} and since $k_n \rightarrow \infty$, for $n$ large enough we have: $$ -\frac{k_n}{5^{N-1}} \frac{\delta^2}{4} + C^2(1+\alpha \tilde c)(1+\tilde c)^\frac{1}{2} \leq -\frac{k_n} {5^{N-1}} \frac{\delta^2}{8}, $$ and so \begin{equation} \label{divIneqAltCaff1} -\div_{\mathbb{R}^{N-1}} ( \tilde{M}_n(x) \nabla \tilde{u}_n ) \leq -\frac{k_n}{5^{N-1}} \frac{\delta^2}{8} \tilde{u}_n. \end{equation} Thus, by Lemma \ref{estimateLemma}-(2) and inequality \eqref{divIneqAltCaff1}, there exists $C$ and $c_2 = c_2(\beta, B_1(0), \theta, \tilde{M}, N, \tilde{\theta}, \tilde{c})$ such that $$ \sup_{x \in B_{R}(0)}\tilde{u}_n(x) \leq Ce^{-c_2R\sqrt{k_n\delta^2}}, $$ and since $k_n \rightarrow \infty$ we have $0\leq k_n\tilde{u}_n(0) \leq Ck_ne^{-c_2R\sqrt{k_n\delta^2}} \rightarrow 0. $. This contradicts the fact that: $ C k_n\tilde{u}_n(0) \geq k_n \tilde{v}_n(0) \tilde{u}_n(0) \to \infty$ coming from \eqref{blowUpContradiction} and the uniform boundedness $\|\tilde{v}_n\|_{L^\infty(\partial B_1)} \leq C$. Thus we have that $\tilde{v}_n(0) \rightarrow 0$, and similarly $\tilde{u}_n(0) \rightarrow 0$, thus $t_n = \tilde{v}_n(0) + \tilde{u}_n(0) \rightarrow 0$, as claimed. Now define the functions $(\overline{u}_n, \overline{v}_n)$ and the matrix $\overline{M}_n(z)$ by: \begin{equation*} \label{definitionOfOverlines} \overline{u}_n(z) = \frac{1}{t_n}\tilde{u}_n(t_n^{\frac{1}{\beta}}z), \qquad \overline{v}_n(z) = \frac{1}{t_n}\tilde{v}_n(t_n^{\frac{1}{\beta}}z), \qquad \overline{M}_n(z) = \tilde{M}_n(t_n^\frac{1}{\beta}z) . \end{equation*} From \eqref{planeEq}, we have: \begin{equation} \label{overlineEqAltCaff} \begin{cases} -\div_{\mathbb{R}^{N-1}} \left( \overline{M}_n(z) \nabla_{\mathbb{R}^{N-1}} \overline{u}_n \right) = -t_n^{2+\frac{2}{\beta}} \frac{k_n}{g(t_n^\frac{1}{\beta}z)} (1 + \frac{\lambda_n \gamma'(x_n)}{\gamma'(y_n)})\overline{u}_n\overline{v}_n^2 + t_n^\frac{2}{\beta} \frac {\sigma_{1,n}(1+\alpha\tilde{c}_n)} {g(t_n^\frac{1}{\beta}z)\gamma'(x_n)} \mu_n(\phi_n(t_n^\frac{1}{\beta}z))\overline{u}_n \\ -\div_{\mathbb{R}^{N-1}} \left( \overline{M}_n(z) \nabla_{\mathbb{R}^{N-1}} \overline{v}_n \right) = -t_n^{2+\frac{2}{\beta}} \frac{k_n}{g(t_n^\frac{1}{\beta}z)} (1 + \frac{\lambda_n \gamma'(x_n)}{\gamma'(y_n)})\overline{v}_n\overline{u}_n^2 + t_n^\frac{2}{\beta} \frac {\sigma_{2,n}(1+\alpha\tilde{c}_n)} {g(t_n^\frac{1}{\beta}z)\gamma'(x_n)} \mu_n(\phi_n(t_n^\frac{1}{\beta}z))\overline{v}_n. \end{cases} \end{equation} Moreover, the functions $\overline{u}_n, \overline{v}_n$ are $\beta$-H\"older, with constant $C_\beta$ in the set $B_{t_n^{-1/\beta}}(0)$, we have \begin{equation} \label{normalizationAtZero} \overline{u}_n(0)+\overline{v}_n(0) = 1, \end{equation} Since $t_n \rightarrow 0$, ovserve that $B_1(0) \subset B_{t_n^{-1/\beta}}$. The uniform H\"older bounds and the boundedness of the functions at $0$ by \eqref{normalizationAtZero} imply the existence of $C_\infty = 1 + C_\beta$ such that: \begin{equation*} \label{LinfinityBoundAltCaffareli} \|\overline{u}_n\|_{L^\infty(B_1(0))} \leq C_\infty,\qquad \|\overline{v}_n\|_{L^\infty(B_1(0))} \leq C_\infty, \end{equation*} and by Ascoli-Arzel\'a's Theorem there exists $(\overline{u}_\infty, \overline{v}_\infty) \in C^0(B_1(0))$ such that, up to a subsequence, $(\overline{u}_n,\overline{v}_n) \rightarrow (\overline{u}_\infty,\overline{v}_\infty)$ in $C^0(B_1(0))$. Since $\overline{u}_\infty(0)+\overline{v}_\infty(0) = 1$, we may assume without loss of generality that $\overline{u}_\infty(0) \geq \frac{1}{2}$. Then, there exists a $0 <\delta < 1$ and $\overline{n}$ large enough such that: $$\overline{u}_n(x)\geq \frac{1}{4} \qquad \forall n > \overline{n}, \ x \in B_{2\delta}(0).$$ Notice also that $\tilde{u}_n(0)\tilde{v}_n(0) \leq \tilde{u}_n^2(0)+\tilde{v}_n^2(0) + 2\tilde{u}_n(0) \tilde{v}_n(0) = t_n^2$, and so: $$k_n\tilde{u}_n^{1+\frac{1}{\beta}}(0)\tilde{v}_n^{1+\frac{1}{\beta}}(0) \leq k_n t_n^{2+\frac{2}{\beta}}.$$ Since, by the contradiction hypothesis \eqref{blowUpContradiction}, we have $k_n\tilde{u}_n^{1+\frac{1}{\beta}}(0)\tilde{v}_n^{1+\frac{1}{\beta}}(0) \rightarrow \infty$, then $k_nt_n^{2+\frac{2}{\beta}} \rightarrow \infty$. By using equation \eqref{overlineEqAltCaff}, that $k_nt_n^{2+\frac{2}{\beta}} \rightarrow \infty$ and $t_n^\frac{2}{\beta} \frac {\sigma_{1,n}(1+\alpha\tilde{c}_n)} {g(t_n^\frac{1}{\beta}z)\gamma'(x_n)} \mu_n(\phi_n(t_n^\frac{1}{\beta}z))$ is bounded in $B_1$, we conclude there exists a constant $C>0$ such that: \begin{align} - \div_{\mathbb{R}^{N-1}}& ( \overline{M}_n(z) \nabla_{\theta} \overline{v}_n ) = \overline{v}_n\left( -t_n^{2+\frac{2}{\beta}} \frac{k_n}{g(t_n^\frac{1}{\beta}z)} (1 + \frac{\lambda_n \gamma'(x_n)}{\gamma'(y_n)})\overline{u}_n^2 + t_n^\frac{2}{\beta} \frac {\sigma_{2,n}(1+\alpha\tilde{c}_n)} {g(t_n^\frac{1}{\beta}z)\gamma'(x_n)} \mu_n(\phi_n(t_n^\frac{1}{\beta}z)) \right) \nonumber \\ &\leq \overline{v}_n\left( -t_n^{2+\frac{2}{\beta}} \frac{k_n}{5^{N-1}} (\frac{1}{4})^2 + \frac{t_n^{\frac{2}{\beta}} \sigma_{2,n}(1+\alpha \tilde c) (1+\tilde c)^\frac{1}{2}}{\gamma'(y_n)} \right) \leq -Ct_n^{2+\frac{2}{\beta}}k_n\overline{v}_n \qquad \forall z \in B_{2\delta}. \label{overlineInequalityAltCaff} \end{align} Using once again Lemma \ref{estimateLemma}-(2) and inequality \eqref{overlineInequalityAltCaff}, there exist constants $C_1$ and $C_2$ depending on the ellipticity constant $\tilde \theta$, and the bounds on the norms of $\overline{M}_n$ such that: \begin{equation} \label{overlineInequalityCoolLemaAltCaff} \overline{v}_n(0) \leq C_1C_\infty e^{-C_2\delta t_n^{1+\frac{1}{\beta}}k_n^{\frac{1}{2}}}. \end{equation} Multiplying inequality \eqref{overlineInequalityCoolLemaAltCaff} by $\overline{u}_n(0)$ and bounding it by $C_\infty$, we obtain: \begin{equation} \label{overlineInequalityCoolLemaAltCaff2} \overline{u}_n(0)\overline{v}_n(0) \leq C_1C_\infty^2 e^{-C_2\delta t_n^{1+\frac{1}{\beta}}k_n^{\frac{1}{2}}}. \end{equation} Raising both sides of \eqref{overlineInequalityCoolLemaAltCaff2} to the power $\frac{1}{2}+\frac{1}{2\beta}$ and taking $\tilde{C}_1 = C_1^{\frac{1}{2}+\frac{1}{2\beta}}C_\infty^{1+\frac{1}{\beta}}$ and $\tilde{C}_2 = C_2(\frac{1}{2}+\frac{1}{2\beta})$ we get: \begin{equation} \label{overlineInequalityCoolLemaAltCaff3} \overline{u}_n^{\frac{1}{2}+\frac{1}{2\beta}}(0)\overline{v}_n^{\frac{1}{2}+\frac{1}{2\beta}}(0) \leq \tilde{C}_1 e^{-\tilde{C}_2t_n^{1+\frac{1}{\beta}}k_n^{\frac{1}{2}}}. \end{equation} Multiplying \eqref{overlineInequalityCoolLemaAltCaff3} by $k_n^{\frac{1}{2}}t_n^{1+\frac{1}{\beta}}$ we get: \begin{equation} \label{finalIneqAltCaff5} k_n^{\frac{1}{2}}\tilde{u}_n^{\frac{1}{2}+\frac{1}{2\beta}}(0)\tilde{v}_n^{\frac{1}{2}+\frac{1}{2\beta}}(0) = k_n^\frac{1}{2} t_n^{1+\frac{1}{\beta}}\overline{u}_n^{\frac{1}{2} + \frac{1}{2\beta}}(0)\overline{v}_n^{\frac{1}{2}+\frac{1}{2\beta}}(0) \leq k_n^{\frac{1}{2}}t_n^{1+\frac{1}{\beta}}\tilde{C}_1 e^{-\tilde{C}_2\delta t_n^{1+\frac{1}{\beta}}k_n^{\frac{1}{2}}} \end{equation} Since $ t_n^{1+\frac{1}{\beta}}k_n^{\frac{1}{2}} \rightarrow \infty $ we know that $ k_n^{\frac{1}{2}} t_n^{1+\frac{1}{\beta}}\tilde{C}_1 e^{-\tilde{C}_2\delta t_n^{1+\frac{1}{\beta}}k_n^{\frac{1}{2}}} \rightarrow 0 $ and, by the contradiction hypothesis \eqref{blowUpContradiction} we have $ k_n^{\frac{1}{2}}\tilde{u}_n^{\frac{1}{2}+\frac{1}{2\beta}}(0)\tilde{v}_n^{\frac{1}{2}+\frac{1}{2\beta}}(0) \rightarrow \infty $ in contradiction with inequality \eqref{finalIneqAltCaff5} concluding the proof of this step. Thus, there exists $D = D(N, \tilde{c}, \tilde{\theta}, \delta, \overline{\lambda}, \beta)>0$ such that $ u_nv_n \leq Dk_n^{-\frac{\beta}{\beta+1}}. $ In particular, we have: \begin{equation*} \label{boundOnEquality} x \in \{v_n-u_n\leq 0\} \implies v_n(x)\leq D^\frac{1}{2}k_n^{-\frac{\beta}{2(\beta+1)}},\quad x \in \{v_n-u_n\geq 0\} \implies u_n(x)\leq D^\frac{1}{2}k_n^{-\frac{\beta}{2(\beta+1)}}. \end{equation*} \noindent \textbf{Step 3.} We claim that there exists $C = C(N, \overline{\lambda}, \tilde{c}, \delta, \beta, \alpha, \tilde{\theta}, \tilde{M})$ such that \begin{equation} \label{squareGradientEstimate} \int_{\{v_n> u_n\}}|\nabla_\theta u_n|^2d\sigma(y) \leq Ck_n^{-\frac{\beta}{2(\beta+1)}}, \qquad \int_{\{u_n> v_n\}}|\nabla_\theta v_n|^2 d\sigma(y) \leq Ck_n^{-\frac{\beta}{2(\beta+1)}} \end{equation} and \begin{equation} \label{innerProductEstimate} \int_{\{u_n> v_n\}} \langle B_n(y) \nabla_\theta u_n, \nabla_\theta v_n \rangle d\sigma(y) \leq Ck_n^{-\frac{\beta}{2(\beta+1)}} ,\qquad \int_{\{v_n> u_n\}} \langle B_n(y) \nabla_\theta u_n, \nabla_\theta v_n \rangle d\sigma(y) \leq Ck_n^{-\frac{\beta}{2(\beta+1)}}. \end{equation} \vspace{2mm} To show this, we fix from now on $n \in \mathbb{N}$, and we consider $ \epsilon>0 $ such that if $ u_n(x)-v_n(x) = \epsilon. $ Then: $$ v_n(x)\leq Ck_n^{-\frac{\beta}{2(\beta+1)}},\qquad u_n(x) \leq Ck_n^{-\frac{\beta}{2(\beta+1)}}. $$ This is possible because, if $u_n(x)-v_n(x) = \epsilon$, then by Step 2 $v_n(x) \leq D^\frac{1}{2}k_n^{-\frac{\beta}{2(\beta+1)}}$ and $ u_n^2(x)-\epsilon u_n(x) \leq Dk_n^{-\frac{\beta}{(\beta+1)}}. $ Thus, by taking $\epsilon \leq \frac{Dk_n^{-\frac{\beta}{\beta+1}}}{2(1+\|u_n\|_{L^\infty(\partial B_1)})}$, we obtain $u_n(x) \leq 2^\frac{1}{2}D^\frac{1}{2}k_n^{-\frac{\beta}{2(\beta+1)}}$. Similarly, we obtain $v_n(x) \leq 2^\frac{1}{2}D^\frac{1}{2}k_n^{-\frac{\beta}{2(\beta+1)}}$ for $x \in \{u_n-v_n = \epsilon\}$. In particular, we conclude the following statements: \begin{align} \label{setImplicationsAltCaff} &x \in \{u_n-v_n\leq \epsilon\} \Longrightarrow u_n(x) \leq Ck_n^{-\frac{\beta}{2(\beta+1)}},\qquad x \in \{u_n-v_n\geq\epsilon\} \Longrightarrow v_n(x) \leq Ck_n^{-\frac{\beta}{2(\beta+1)}} \end{align} By Morse-Sard's theorem (see for instance the version in \cite[Lemma 2.96]{AmbrosioFusco}) we can also suppose that $ \epsilon $ is such that the set $ \{u_n-v_n=\epsilon\} $ is an $N-2$ dimensional submanifold in the sphere $\partial B_1$. Now we integrate equation \eqref{SphereEq} for $u_n$ in the subset $ \{u_n-v_n \geq \epsilon\} $, and using the divergence theorem one obtains: \begin{align} -\int_{\{u_n-v_n=\epsilon \}} &\langle\tilde{B}_n(y) \nabla_\theta u_n, (\nu_{\epsilon}^1)_{y}\rangle_{\partial B_1} d\mathcal{H}^{N-2} \nonumber \\ & = \int_{\{u_n-v_n\geq \epsilon\}} \left( -k_n( \lambda_n + \frac{\gamma'(y_n)}{\gamma'(x_n)}) u_nv_n^2 + \frac{\sigma_{1,n}(1+\alpha \tilde{c}_n)}{\gamma'(x_n)}\mu_n(y) u_n \right) d\sigma(y) , \label{unEquationAltCaffSphere} \end{align} where $ (\nu_{\epsilon}^1)_y \in T_y \partial B_1 $ is the exterior normal to the set $ \{u_n-v_n\geq \epsilon\} $. Integrating equation \eqref{SphereEq} in all of $\partial B_1$, we obtain: \begin{equation*} \int_{\partial B_1} k_n(\lambda_n + \frac{\gamma'(y_n)}{\gamma'(x_n)})u_nv_n^2 d\sigma(y) = \int_{\partial B_1} \frac{\sigma_{1,n}(1+\alpha \tilde{c}_n)}{\gamma'(x_n)}\mu_n(y) u_n d\sigma(y). \end{equation*} Thus, by Step 1, we conclude that there exists $C>0$ such that: \begin{equation} \label{boundOnInteraction} \int_{\partial B_1} k_n v_n^2u_nd\sigma(y) \leq C\qquad \int_{\partial B_1} k_n u_n^2v_n d\sigma(y) \leq C\qquad \text{ for all $n \in \mathbb{N}$.} \end{equation} With \eqref{boundOnInteraction}, we conclude that the right-hand-side of \eqref{unEquationAltCaffSphere} is uniformly bounded in $n$ from above and below thus there exists $C>0$ such that: \begin{equation} \label{boundOnBoundaryIntegral1} \Big| \int_{\{u_n-v_n=\epsilon \}} \langle\tilde{B}_n(y)\nabla_\theta u_n, (\nu_{\epsilon}^1)_y\rangle d\mathcal{H}^{N-2} \Big| \leq C. \end{equation} We can do the same with $v_n$: \begin{equation} \label{boundOnBoundaryIntegral2} \Big| \int_{\{u_n-v_n=\epsilon\}} \langle \tilde{B}_n(y)\nabla_\theta v_n, (\nu_{\epsilon}^1)_y\rangle d\mathcal{H}^{N-2} \Big| \leq C. \end{equation} Now we multiply the equation \eqref{SphereEq} by $u_n$ and integrate in the set $ \{u_n-v_n \leq \epsilon\} $. Then, by \eqref{setImplicationsAltCaff} for $x \in \{u_n-v_n\leq \epsilon\}$, we have $u_n(x) \leq Ck_n^{-\frac{\beta}{2(\beta+1)}}$, and so: \begin{multline*} \Big| \int_{\{u_n-v_n\leq \epsilon\}} \langle \tilde{B}_n(y)\nabla_\theta u_n, \nabla_\theta u_n \rangle d\sigma(y) - \int_{\{u_n-v_n=\epsilon\}} \langle \tilde{B}_n(y) \nabla_\theta u_n, (\nu_{\epsilon}^1)_y \rangle u_n(y) d\mathcal{H}^{N-2} \Big| \\ = \Big| \int_{\{u_n-v_n\leq \epsilon\}} \left( -k_n(\lambda_n + \frac{\gamma'(y_n)}{\gamma'(x_n)})u_n^2v_n^2 + \frac{\sigma_{1,n}}{\gamma'(x_n)}u_n^2 \right) d\sigma(y) \Big|\leq Ck_n^{-\frac{\beta}{2(\beta+1)}} \end{multline*} (since $-(\nu_{\epsilon}^1)$ is the exterior normal to $\{u_n-v_n \leq \epsilon\}$), the right hand side is bounded by $ Ck_n^{-\frac{\beta}{2(\beta+1)}} $ by using equation \eqref{setImplicationsAltCaff} and \eqref{boundOnInteraction}. We can also do the same for $v_n$ by integrating in $\{u_n-v_n\geq \epsilon\}$ and we obtain the following bounds: $$ \Big| \int_{\{u_n-v_n\leq \epsilon\}} \langle \tilde{B}_n(y)\nabla_\theta u_n, \nabla_\theta u_n \rangle d\sigma(y) - \int_{\{u_n-v_n=\epsilon\}} \langle \tilde{B}_n(y) \nabla_\theta u_n, (\nu_{\epsilon}^1)_y \rangle u_n d\mathcal{H}^{N-2} \Big| \leq Ck_n^{-\frac{\beta}{2(\beta+1)}}, $$ $$ \Big| \int_{\{u_n-v_n\geq \epsilon\}} \langle \tilde{B}_n(y)\nabla_\theta v_n, \nabla_\theta v_n \rangle d\sigma(y) + \int_{\{u_n-v_n=\epsilon\}} \langle \tilde{B}_n(y) \nabla_\theta v_n, (\nu_{\epsilon}^1)_y \rangle v_n d\mathcal{H}^{N-2} \Big| \leq Ck_n^{-\frac{\beta}{2(\beta+1)}}. $$ The bound for $v_n$ has an inverted sign for the integral in $\{u_n-v_n=\epsilon\}$ since the exterior normal to $\{u_n-v_n\geq \epsilon\}$ is simply $-(\nu_{\epsilon}^1)$, because $(\nu_{\epsilon}^1)$ is the exterior normal to $\{u_n-v_n\leq \epsilon\}$. Summing up both equations we obtain that: \begin{align} &\Big| \int_{\{u_n-v_n\leq \epsilon\}} \langle \tilde{B}_n(y)\nabla_{\theta} u_n, \nabla_{\theta} u_n \rangle d\sigma(y) + \int_{\{u_n-v_n\geq \epsilon\}} \langle \tilde{B}_n(y)\nabla_{\theta} v_n, \nabla_{\theta} v_n \rangle d\sigma(y) \nonumber \\ &+ \int_{\{u_n-v_n=\epsilon\}} \left( \langle \tilde{B}_n(y) (\nabla_{\theta} v_n - \nabla_{\theta} u_n), (\nu_{\epsilon}^1)_y \rangle u_n - \epsilon\langle \nabla \tilde{B}_n(y)\nabla_\theta v_n, (\nu_{\epsilon}^1)_y \rangle \right) d\mathcal{H}^{N-2} \Big| \leq Ck_n^{-\frac{\beta}{2(\beta+1)}}. \label{essentialEquationaltCaff3} \end{align} Now we make the observation that $$ \nabla_{\theta}\left( v_n-u_n \right) = \big|\nabla_\theta \left( v_n-u_n \right)\big| (\nu^1_{\epsilon})_y, $$ since $ \nu_{\epsilon}^1 $ is the normal exterior to the level set $ \{u_n-v_n\geq\epsilon\} $ . We conclude: $$ \langle \tilde{B}_n(y) \nabla_{\theta} (v_n - u_n), (\nu_{\epsilon}^1)_y \rangle = \big| \nabla_\theta \big(v_n-u_n\big) \big| \langle\tilde{B}_n(y) (\nu^1_{\epsilon})_y, (\nu^1_{\epsilon})_y \rangle \geq 0. $$ Thus, using the fact that the integrand below has a sign, we know that: \begin{align} &\Big| \int_{\{u_n-v_n=\epsilon\}} \langle \tilde{B}_n(y) (\nabla_{\theta} v_n - \nabla_{\theta} u_n), (\nu_{\epsilon}^1)_y \rangle u_nd\mathcal{H}^{N-2} \Big| \nonumber \\ &\leq \|u_n\|_{L^\infty(\{u_n-v_n=\epsilon\})} \Big| \int_{\{u_n-v_n=\epsilon\}} \langle \tilde{B}_n(y) (\nabla_{\theta} v_n - \nabla_{\theta} u_n), (\nu_{\epsilon}^1)_y \rangle d\mathcal{H}^{N-2} \Big| \leq Ck_n^{-\frac{\beta}{2(\beta+1)}} \label{diferenceLineIntegral1}\end{align} where we have used equations \eqref{setImplicationsAltCaff}, \eqref{boundOnBoundaryIntegral1} and \eqref{boundOnBoundaryIntegral2}. Similarly, we have that: \begin{equation} \label{lineIntegral2} \Big| \int_{\{u_n-v_n=\epsilon\}} \epsilon\langle \tilde{B}_n(y)\nabla_\theta v_n,(\nu_{\epsilon}^1)_y \rangle d\mathcal{H}^{N-2} \Big| \leq C\epsilon. \end{equation} From equations \eqref{diferenceLineIntegral1}, \eqref{lineIntegral2} and \eqref{essentialEquationaltCaff3} we obtain: $$ \int_{\{u_n-v_n\geq \epsilon\}} \langle \tilde{B}_n(y)\nabla_{\theta} u_n, \nabla_{\theta} u_n \rangle d\sigma(y) + \int_{\{u_n-v_n\leq \epsilon\}} \langle \tilde{B}_n(y)\nabla_{\theta} v_n, \nabla_{\theta} v_n \rangle d\sigma(y) \leq Ck_n^{-\frac{\beta}{2(\beta+1)}} + C\epsilon. $$ By making $\epsilon$ go to zero we conclude that: $$ \int_{\{u_n-v_n>0\}} \langle \tilde{B}_n(y)\nabla_{\theta}u_n, \nabla_{\theta}u_n \rangle d\sigma(y) + \int_{\{u_n-v_n<0\}} \langle \tilde{B}_n(y)\nabla_{\theta}v_n, \nabla_{\theta}v_n\rangle d\sigma(y) \leq Ck_n^{-\frac{\beta}{2(\beta+1)}}. $$ This shows \eqref{squareGradientEstimate}. Finally, by multiplying the equation for $v_n$ in \eqref{SphereEq} by $\min\{u_n,v_n\}$ and integrating in the entire sphere, we obtain an estimate of the form: $$ \Big| \int_{\{u_n<v_n\}} \langle B_n(y) \nabla_\theta u_n, \nabla_\theta u_n \rangle d\sigma(y) + \int_{\{v_n\leq u_n\}} \langle B_n(y) \nabla_\theta u_n, \nabla_\theta v_n \rangle d\sigma(y) \Big| \leq Ck_n^{-\frac{\beta}{2(\beta+1)}}, $$ thus using the first estimate given by \eqref{squareGradientEstimate} we can find $C = C(N, \overline{\lambda}, \tilde{c}, \delta, \beta, \alpha, \tilde{\theta}, \tilde{M})$ such that $$ \Big| \int_{\{v_n\leq u_n\}} \langle B_n(y) \nabla_\theta u_n, \nabla_\theta v_n \rangle d\sigma(y) \Big| \leq Ck_n^{-\frac{\beta}{2(\beta+1)}}, $$ which is the first estimate in \eqref{innerProductEstimate}. To obtain the second one we proceed in an analogous way, this time multiplying the equation for $v_n$ in \eqref{SphereEq} by $\min\{u_n,v_n\}$. \noindent \textbf{Step 4. } Conclusion of the proof of the lemma. Take the functions $ f_n = (u_n-v_n)^+, g_n = (u_n-v_n)^- \in H^1(\partial B_1) $. Using the classical Friedman-Hayman inequality in the sphere \eqref{eq:FH_ineq}, we obtain: $$ 2 \leq \gamma\left( \frac{\int_{\partial B_1} |\nabla_\theta f_n|^2 d\sigma(y)} {\int_{\partial B_1}f_n^2d\sigma(y)} \right) + \gamma\left( \frac{\int_{\partial B_1} |\nabla_\theta g_n|^2 d\sigma(y)}{\int_{\partial B_1} g_n^2d\sigma(y)} \right). $$ We compute the $L^2(\partial B_1)$ norm of the gradient $\nabla_\theta f_n$: from $ \sup_{y \in \partial B_1}\|\tilde{B}_n(y)|_{T_y \partial B_1}- Id|_{T_y \partial B_1}\| \leq \tilde{M}\tilde{c}_n $, the uniform boundedness of $f_n$, and the estimates \eqref{innerProductEstimate}, \eqref{squareGradientEstimate} proved in Step 3, \begin{align*} \int_{\partial B_1} &|\nabla_\theta f_n|^2 d\sigma(y)\leq \int_{\partial B_1}\langle\tilde{B}_n(y)\nabla_\theta f_n, \nabla_\theta f_n\rangle d\sigma(y)+ C\tilde{c}_n \\ &=\int_{\{u_n>v_n\}} \left(\langle \tilde{B}_n(y)\nabla_\theta u_n, \nabla_\theta u_n \rangle + \langle \tilde{B}_n(y)\nabla_\theta v_n,\nabla_\theta v_n \rangle - 2\langle \tilde{B}_n(y)\nabla_\theta u_n, \nabla_\theta v_n \rangle\right)d\sigma(y) + C\tilde{c}_n\\ &\leq \int_{\partial B_1}\langle \tilde{B}_n(y)\nabla_\theta u_n, \nabla_\theta u_n \rangle d\sigma(y) + C\tilde{c}_n+ Ck_n^{-\frac{\beta}{2(\beta+1)}} \\ &\leq \left(\int_{\partial B_1}\left(\langle \tilde{B}_n(y)\nabla_\theta u_n, \nabla_\theta u_n \rangle + k_nu_n^2v_n^2\right)d\sigma(y)- \epsilon_n\right) + C \left(\epsilon_n+k_n^{-\frac{\beta}{2(\beta+1)}} + \tilde{c}_n\right)\\ &\leq x_n + C\left(\epsilon_n+k_n^{-\frac{\beta}{2(\beta+1)}} + \tilde{c}_n\right), \end{align*} where we recall from \eqref{def_of_x_n} the definition of $x_n$. On the other hand, by \label{GammaMinimization_norm}, \begin{align*} 1&= \int_{\partial B_1}(1+\alpha \tilde{c}_n)\mu_n(y)u_n^2d\sigma(y) \leq \int_{\partial B_1} u_n^2+C\tilde c_n = \int_{\{u_n>v_n\}}u_n^2+\int_{\{v_n>u_n\}}u_n^2+C\tilde c_n\\ &=\int_{\partial B_1} f_n^2 + \int_{\{u_n>v_n\}} (2u_nv_n-v_n^2) + \int_{\{v_n>u_n\}}u_n^2+C\tilde c_n \leq \int_{\partial B_1} f_n^2 + C \tilde c_n+ Ck_n^{-\frac{\beta}{\beta+1}}, \end{align*} where we used estimates from Step 2. Then $$ \frac {\int_{\partial B_1} |\nabla_\theta f_n|^2 d\sigma(y) } {\int_{\partial B_1} |f_n|^2 d\sigma(y) } \leq \frac{x_n + C(\epsilon_n + \tilde{c}_n + k_n^\frac{-\beta}{2(\beta+1)})}{1-C\tilde{c}_n-Ck_n^{-\frac{\beta}{\beta+1}}} \leq x_n + C' (\epsilon_n + \tilde{c}_n + k_n^\frac{-\beta}{2(\beta+1)}). $$ Using moreover the monotonicity and concavity of $\gamma$, we have: $$ 2 \leq \gamma\left( \frac{\int_{\partial B_1} |\nabla_\theta f_n|^2 d\sigma(y)} {\int_{\partial B_1}f_n^2 d\sigma(y)} \right) + \gamma\left( \frac{\int_{\partial B_1} |\nabla_\theta g_n|^2 d\sigma(y)}{\int_{\partial B_1} g_n^2d\sigma(y)} \right) \leq \gamma(x_n) + \gamma(y_n) + C\left( \epsilon_n + k_n^{-\frac{\beta}{2(\beta+1)}} + \tilde{c}_n \right).$$ Thus, we obtain the desired bound: $$ 2-C(\epsilon_n+k_n^{-\frac{\beta}{2(\beta+1)}} + \tilde{c}_n) \leq \gamma(x_n) + \gamma(y_n).$$ Observing that an arbitrary choice of $\beta \in ]0,1[$ yields an arbitrary choice of $\frac{\beta}{2(\beta+1)} \in ]0,\frac{1}{4}[$, we conclude the proof of the lemma. \end{proof} We are now able to prove the Alt-Caffarelli-Friedman- type monotonicity formula. \begin{proof}[Proof of Theorem \ref{AltCaffMonotonicity}] We start by computing the derivative of $ \log\left( \frac{J_{1,n}(r)J_{2,n}(r)}{r^4} \right) $: \begin{multline*} \frac{d}{dr}\log\left( \frac{J_{1,n}(r)J_{2,n}(r)}{r^4} \right) = -\frac{4}{r}\\ + \frac{\int_{\partial B_r}\left( \langle A_n(y) \nabla u_{1,n}, \nabla u_{1,n} \rangle - M_na_n(y)|u_{1,n}|^2|u_{2,n}|^2 - u_{1,n}f_{1,n}(y,u_{1,n})|y|^{2-N}d\sigma(y) \right) } {\int_{B_r}\left( \langle A_n(y) \nabla u_{1,n}, \nabla u_{1,n} \rangle - M_na_n(y)|u_{1,n}|^2|u_{2,n}|^2 - u_{1,n}f_{1,n}(y,u_{1,n})|y|^{2-N}dy \right)}\\ + \frac{\int_{\partial B_r}\left( \langle A_n(y) \nabla u_{2,n}, \nabla u_{1,n} \rangle - M_na_n(y)|u_{2,n}|^2|u_{1,n}|^2 - u_{2,n}f_{2,n}(y,u_{2,n})|y|^{2-N}d\sigma(y) \right) } {\int_{B_r}\left( \langle A_n(y) \nabla u_{2,n}, \nabla u_{2,n} \rangle - M_na_n(y)|u_{2,n}|^2|u_{1,n}|^2 - u_{2,n}f_{2,n}(y,u_{2,n})|y|^{2-N}dy \right)}. \end{multline*} Since the hypothesis for Lemma \ref{derivativeLemma} are satisfied, we obtain: \begin{equation} \label{OkEquationCaff} \frac{d}{dr}\log\left( \frac{J_{1,n}(r)J_{2,n}(r)}{r^4} \right) \geq -\frac{2}{r}\left( 2 - \gamma\left( \Lambda_{1,n}(r) \right) - \gamma\left( \Lambda_{2,n}(r) \right) \right) \end{equation} Recalling the definition of the operator $B_n$ from Definition \ref{definition:operatorB} and the equivalent formulation of $\Lambda_{1,n}$ written in \eqref{eq:equivalent_LAMBDA}, we can use the fact that $M_n < 0$, hypothesis ($h_2$) and the monotonicity of $\gamma$ to conclude: $$ \gamma(\Lambda_{1,n}(r)) \geq \gamma\left( \frac{ r^2\int_{\partial B_r} \left( \langle B_n(y) \nabla_{\theta} u_{1,n}, \nabla_{\theta} u_{1,n} \rangle - M_na_n(y)|u_{1,n}|^{2} |u_{2,n}|^2 - (1+\alpha(rc_n))\epsilon_n \mu_n(y)u_{1,n}^2 \right)} { \int_{\partial B_r} (1+\alpha(rc_n))\mu_n(y)u_{1,n}^2 } \right). $$ In order to rewrite the above as integrals over $\partial B_1$, we consider the change of variables given by: \begin{equation} \label{changeOfVarAltCaff} u_{i,n,r}(z) = \frac{u_{i,n}(rz)} {\sqrt{\frac{1}{r^{N-1}}\int _{\partial B_r}(1+\alpha(rc_n))\mu_n(y)u_{1,n}^2}} \end{equation} and denote, for convenience, $$ d_{n,r} = \frac{1}{r^{N-1}}\int_{\partial B_r}(1+\alpha(rc_n))\mu_n(y) u_{1,n}^2d\sigma(y)\quad \text{ and }\quad m_{n,r} = (1+\alpha(rc_n)). $$ Therefore, \begin{multline*} \frac{ r^2\int_{\partial B_r} \langle B_n(y) \nabla_{\theta} u_{1,n}, \nabla_{\theta} u_{1,n} \rangle - M_na_n(y)|u_{2,n}|^{2} |u_{1,n}|^2-\epsilon_nm_{n,r} \mu_n(y)u_{1,n}^2 d} {\int_{\partial B_r} m_{n,r}\mu_n(y)u_{1,n}^2}\\ = {\int_{\partial B_1} \left( \langle B_n(rz) \nabla_\theta u_{1,n,r}, \nabla_{\theta} u_{1,n,r} \rangle - d_{n,r}M_nr^2a_n(rz)|u_{2,n,r}|^2|u_{1,n,r}|^2 - r^2\epsilon_nm_{n,r}\mu_n(rz)u_{1,n,r}^2 \right)} \end{multline*} and \begin{multline*} \frac{ r^2\int_{\partial B_r} \langle B_n(y) \nabla_{\theta} u_{2,n}, \nabla_{\theta} u_{2,n} \rangle - M_na_n(y)|u_{2,n}|^{2} |u_{1,n}|^2 - \epsilon_nm_{n,r} \mu_n(y)u_{2,n}^2 } { \int_{\partial B_r} m_{n,r}\mu_n(y)u_{2,n}^2 }\\ = \frac{ {\int_{\partial B_1} \left( \langle B_n(rz) \nabla_\theta u_{2,n,r}, \nabla_{\theta} u_{2,n,r} \rangle - d_{n,r}M_nr^2a_n(rz)|u_{2,n,r}|^2|u_{1,n,r}|^2 - r^2\epsilon_nm_{n,r}\mu_n(rz)u_{2,n,r}^2 \right)} } { \int_{\partial B_1} m_{n,r}\mu_n(y)u_{2,n,r}^2 }. \end{multline*} Thus, by hypotheses ($h_0$), ($h_1$), and Lemma \ref{MatrixBoundsOnB}, we know that: $$ \sup_{y \in \partial B_1}\|B_n(ry)|_{T_y \partial B_1}-Id_{T_y \partial B_1}\| \leq Cc_nr, $$ while by $(h_3)$ we know that there exists $w>0$ such that: \begin{equation*} \label{inequalityDnAltCaff} d_{n,r} = \frac{1}{r^{N-1}}\int_{\partial B_r}(1+\alpha c_nr)\mu_n(y)u_{1,n}^2 \geq w. \end{equation*} Combining this with the monotonicity of $\gamma$, and since $M_n<0$, we conclude: \begin{multline*} \gamma\left( \int_{\partial B_1} \left( \langle B_n(rz) \nabla_\theta u_{1,n,r}, \nabla_{\theta} u_{1,n,r} \rangle - d_{n,r}M_nr^2a_n(rz)|u_{2,n,r}|^2|u_{1,n,r}|^2 - r^2 \epsilon_nm_{n,r}\mu_n(rz)u_{1,n,r}^2 \right) \right)\\ +\gamma\left( \frac{ {\int_{\partial B_1} \left( \langle B_n(rz) \nabla_\theta u_{2,n,r}, \nabla_{\theta} u_{2,n,r} \rangle - d_{n,r}M_nr^2a_n(rz)|u_{2,n,r}|^2|u_{1,n,r}|^2 - r^2 \epsilon_nm_{n,r}\mu_n(rz)u_{2,n,r}^2 \right)} } { \int_{\partial B_1} m_{n,r}\mu_n(rz)u_{2,n,r}^2 } \right)\\ \geq \gamma\left( \int_{\partial B_1} \left( \langle B_n(rz) \nabla_\theta u_{1,n,r}, \nabla_{\theta} u_{1,n,r} \rangle - wM_nr^2a_n(rz)|u_{2,n,r}|^2|u_{1,n,r}|^2 - r^2 \epsilon_nm_{n,r}\mu_n(rz)u_{1,n,r}^2 \right) \right)\\ +\gamma\left( \frac{ {\int_{\partial B_1} \left( \langle B_n(rz) \nabla_\theta u_{2,n,r}, \nabla_{\theta} u_{2,n,r} \rangle - wM_nr^2a_n(rz)|u_{2,n,r}|^2|u_{1,n,r}|^2 - r^2 \epsilon_nm_{n,r}\mu_n(rz)u_{2,n,r}^2 \right)} } { \int_{\partial B_1} m_{n,r}\mu_n(rz)u_{2,n,r}^2 } \right). \end{multline*} By the change of variables \eqref{changeOfVarAltCaff}, we have: $$ \int_{\partial B_1} m_{n,r}\mu_n(rz)u_{1,n,r}d\sigma(z) = 1, $$ and, by \eqref{changeOfVarAltCaff} and hypothesis $(h_3)$: $$ \frac{1}{\lambda} \leq \int_{\partial B_1} m_{n,r}\mu_n(rz)u_{2,n,r}^2d\sigma(z) = \frac { \int_{\partial B_r} \mu_n(y)u_{2,n}^2d\sigma(y) } { \int_{\partial B_r} \mu_n(y)u_{1,n}^2d\sigma(y) } \leq \lambda. $$ We can now apply Lemma \ref{sphereLemma} with $\tilde B_n(z) = B_n(rz)$, $\tilde{c}_n = c_nr$, $\lambda_n = \int_{\partial B_1} m_{n,r}\mu_n(rz)u_{2,n,r}^2d\sigma(z)$, $\tilde{\epsilon}_n = \epsilon_n r^2<\epsilon_nR_n<(\frac{N-2}{2})^2-\delta$ and $k_n = wM_nr^2\min(a_n)$, concluding the existence of $C>0$ such that, for $\eta \in ] 0, \frac{1}{4}[$: \begin{multline*} \gamma(\Lambda_{1,n}(r)) + \gamma(\Lambda_{2,n}(r)) \nonumber\\ \geq \gamma\left( \int_{\partial B_1} \left( \langle B_n(rz) \nabla_\theta u_{1,n,r}, \nabla_{\theta} u_{1,n,r} \rangle - wM_nr^2a_n(rz)|u_{2,n,r}|^2|u_{1,n,r}|^2 - r^2 \epsilon_n m_{n,r}\mu_n(rz)u_{1,n,r}^2 \right) \right) \nonumber\\ +\gamma\left( \frac{ \int_{\partial B_1} \left( \langle B_n(rz) \nabla_\theta u_{2,n,r}, \nabla_{\theta} u_{2,n,r} \rangle - wM_nr^2a_n(rz)|u_{2,n,r}|^2|u_{1,n,r}|^2 - r^2\epsilon_nm_{n,r}\mu_n(rz)u_{2,n,r}^2 \right)} { \int_{\partial B_1}m_{n,r}\mu_n(rz) u_{2,n,r} } \right) \nonumber\\ \geq 2 - C\left( - |M_n|^{-\eta} r^{-2\eta} - \epsilon_nr^2 - c_nr \right) \end{multline*} Thus, combining this inequality with \eqref{OkEquationCaff}, we see that: \begin{align*} \frac{d}{dr}\left(\log\left( \frac{J_{1,n}(r)J_{2,n}(r)}{r^4} \right) \right) &\geq - \frac{2}{r}\left(2 - C\left( - M_n^{-\eta} r^{-2\eta} - \epsilon_nr^2 - c_nr \right) - 2 \right) \\ &= -C\left( |M_n|^{-\eta} r^{-2\eta-1} - \epsilon_nr - c_n \right), \end{align*} and the proof is finished. \end{proof} \begin{remark}\label{rem_finalSection4} In case $\gamma>1$, a result like this also holds true. The only necessary changes are in the definitions of $J_{i,n}$ and $\Lambda_{i,n}$, where the terms $M_n a_n(y) |u_{1,n}|^2|u_{2,n}|^2$ should be replaced by $M_n a_n(y) |u_{1,n}|^{\gamma+1}|u_{2,n}|^{\gamma+1}$, and in the proof of Lemma \ref{sphereLemma}, whenever Lemma \ref{estimateLemma}-(2) is used, one should use instead Lemma \ref{estimateLemma}-(1). \end{remark} \begin{remark}\label{rem:dimension} For this section we will consider the dimension $N\geq 3$, since we need it for the classical Alt-Caffarelli-Friedman formula. Thus, to obtain Theorem \ref{DesiredTheorem} for dimensions $N\leq 2$, if $u_\beta(x)$, $x\in \mathbb{R}^N$, is a solution to the system \eqref{equation}, we consider the new vector solution $\tilde{u}_\beta(x,y) =u_\beta(x) $ with $x\in \mathbb{R}^N$, $y\in \mathbb{R}^{3-N}$, obtaining a system in dimension $N=3$. This new system will still be of type \eqref{equation}, and one can apply the Alt-Caffarelli-Friedman type formula. \end{remark} \section{Interior Lipschitz bounds}\label{sec:Lip} In this section, we conclude the proof of Theorem \ref{DesiredTheorem}. As observed in Remark \ref{rem:dimension}, we just need to consider the case $N\geq 3$ (in particular, the results of the previous sections are true). In Section \ref{chapter:background}, under the contradiction assumption that $\{u_{\beta_n}\}$ was not uniformly Lipschitz, we introduced in \eqref{blowUp} a blowup sequence $\{v_n\}$, solution to \eqref{eq:system_rescaled}. This sequence was defined in such a way that it has bounded Lipschitz-seminorm, it concentrates at a point where the gradient blowsup, and solves a system where the differential operator is a perturbation of the Laplacian close to the blowup point. In that section we arrived at Proposition \ref{concludeProp}: this sequence converges locally uniformly to a limiting profile $v=(v_1,\ldots, v_l)$, of which at most $v_1$ and $v_2$ are nontrivial. Much more information is necessary to arrive at a contradiction. In the previous two sections, we proved Almgren and Alt-Caffareli-Friedman-type monotonicity formulas; we will now apply them to the sequence $\{v_n\}$ to achieve the desired contradiction. Here we follow the structure of \cite[Section 4]{SoaveZilio}, with the necessary modifications that arise from the fact that we are dealing with a system with divergence type operators with variable coefficients. We recall the sequence of functions $\tilde{u}_{i,\beta_n}>0$ defined in \eqref{tildeU} by: $$ \tilde{u}_{i,\beta_{n}}(x) = u_{i,\beta_n}(x_n+A(x_n)^{\frac{1}{2}}x). $$ This sequence, by Lemma \ref{lemmaForMatrixTilde}, satisfies \begin{equation}\label{eq:systems_for_tildeu_rep} -\div(\tilde{A}_n(x)\nabla \tilde{u}_{i,\beta_n}) = f_i(x_n+A(x_n)^{\frac{1}{2}}x,\tilde{u}_{i,\beta_n}) + a(x_n+A(x_n)^{\frac{1}{2}}x) \mathop{\sum_{j=1}^l}_{j\neq i}\beta_n |\tilde{u}_{j, \beta_n}|^{\gamma + 1} |\tilde{u}_{i, \beta_n}|^{\gamma -1}\tilde{u}_{i, \beta_n}, \end{equation} where $\tilde{A}_n(x) = A(x_n)^{-\frac{1}{2}}A(x_n +A(x_n)^{\frac{1}{2}}x)A(x_n)^{-\frac{1}{2}}$ is such that $\tilde{A}_n(0)=Id$. By hypotheses \eqref{boundedness} and \eqref{boundForF}, there exist $m>0$ and $d>0$ such that: $$ \max_{i=1,...,l}\|\tilde{u}_{i,\beta_n}\|_{L^\infty(B_{1/M^{\frac{1}{2}}})}<m,\qquad \max_{i=1,...,l}\sup_{y\in[0,m]}f_i(x_n+A(x_n)^\frac{1}{2}x , y)\leq d |y| $$ for all $n \in \mathbb{N}$, where $M$ is as in \textbf{(A2)}. Also by Lemma \ref{lemmaForMatrixTilde} there exists $C>0$ such that for all $n \in \mathbb{N}$ we have: $$ \langle \tilde{A}_n(x)\xi, \xi \rangle \geq \frac{\theta}{M}|\xi|^2, \qquad \|D\tilde{A}_n\|_{L^\infty(B_{1/M^{\frac{1}{2}}})} \leq C, \qquad \|\tilde{A}_n\|_{L^\infty(B_{1/M^{\frac{1}{2}}})} \leq C. $$ Moreover, at the point $0$ we have $\tilde{A}(0) = Id$. We also define $\tilde{\mu}_n(y) = \langle \tilde{A}_n(y)\frac{y}{|y|},\frac{y}{|y|} \rangle$. We may, therefore, apply all the results of Section \ref{chapter:implementation} to the sequence $\{\tilde{u}_{\beta_n}\}$. In particular, Lemma \ref{derivOfH} and Theorem \ref{almgrenMonotonicity} imply the following. \begin{prop} \label{MonotonicityForu} Let $\gamma \geq 1$ be such that $\frac{\gamma N}{\gamma+1}<2$. Then there exists $\tilde{r}$ and $\tilde{C}>0$ such that, for every $n \in \mathbb{N}$, the functions: $$(N_{\beta_n}(\tilde{u}_{\beta_n},r) + 1)e^{\tilde{C}r} \quad \text{ and } \quad H_{i,\beta_n}(\tilde{u}_{\beta_n},r)e^{\tilde{C}e}= \frac{1}{r^{N-1}} \int_{\partial B_r} \tilde{\mu}_n(y) \tilde{u}_{i,\beta_n}^2 d\sigma(y) e^{\tilde{C}r} $$ are monotone nondecreasing for $r \in ]0,\tilde{r}[$ and all $i \in \{1,...,l\}$. We recall that $N_{\beta_n}$ is defined in \eqref{Neq}. \end{prop} To ease notation, from now on in this section we omit the lower index $\beta_n$ in the functions $N_{\beta_n}(\tilde{u}_{\beta_n},r)$. Now we introduce the quantity given by: \begin{equation} \label{BigRdefinition} R_{\beta_n} := \sup \left\{r \in ]0,\tilde{r}[: (N(\tilde{u}_{\beta_n},r) + 1)e^{\tilde{C}r}<2-r \right\}. \end{equation} \begin{lemma} We have $R_{\beta_n}>0$ for all $n \in \mathbb{N}$. \end{lemma} \begin{proof} Fix $n \in \mathbb{N}$. Since $\tilde{u}_{i,\beta_n}$ is positive and of class $C^1$, there exist $\delta,\epsilon,C>0$ such that $\delta <u_{i,\beta_n}(x)<m$ and $|\nabla \tilde{u}_{i,\beta_n}(x)| < C$ whenever $|x|<\epsilon$. With this, for $r < \epsilon$ we conclude: \begin{align*} |N(\tilde{u}_{\beta_n},r)| =\left| \frac {\frac{1}{r^{N-2}} \sum_ {i=1}^l \int_{\partial B_r} \tilde{u}_{i, \beta_n}\langle \tilde{A}_n(x) \nabla \tilde{u}_{i,\beta_n}, \nu_x\rangle d\sigma(x) }{ \frac{1}{r^{N-1}}\sum_{i=1}^l\int_{\partial B_r}\tilde{\mu}_n(x)|\tilde{u}_{i,\beta_n}|^2d\sigma(x) } \right| \leq \left(\frac{m MC l}{\delta^2 \tilde{\theta}}\right)r, \end{align*} and so $N(\tilde{u}_{\beta_n},r) \rightarrow 0$ as $r \rightarrow 0$; this implies that $R_{\beta_n}>0$. \end{proof} \begin{lemma} \label{radGoToZero} $R_{\beta_n} \rightarrow 0$ as $n \rightarrow \infty$. \end{lemma} \begin{proof} Recall from Lemma \ref{remarkOnStuff} that (up to a subsequence) $\tilde{u}_{\beta_n} \rightarrow \tilde{u}_\infty$ in $C^{0,\alpha}(B_{1/(2M^\frac{1}{2})})\cap H^1(B_{1/(2M^\frac{1}{2})})$ for every $\alpha\in (0,1)$, and that $\tilde{u}_\infty(0) = 0$. Moreover, by using Lemma \ref{lemmaForMatrixTilde} and Ascoli Arzela's Theorem, there exists $\tilde{A}(\cdot) \in C^1(B_{1/(2M^\frac{1}{2})},\text{Sym}^{N\times N})$ such that $\tilde{A}_n(x) \rightarrow \tilde{A}(x)$ uniformly for $x\in B_{1/(2M^\frac{1}{2})}$. Let $\tilde{\mu}(x) = \langle \tilde{A}(x)\frac{x}{|x|}, \frac{x}{|x|} \rangle$, and consider the Almgren's quotient associated to \eqref{eq:systems_for_tildeu_rep}, namely \begin{align*} E(\tilde{u}_\infty,r) &= \frac{1}{r^{N-2}}\sum_{i=1}^l \int_{B_r} \left( \langle \tilde{A}(x)\nabla \tilde{u}_{i,\infty}, \nabla \tilde{u}_{i,\infty} \rangle - f_i(x_\infty+A(x_\infty)^\frac{1}{2}x,\tilde{u}_{i,\infty})\tilde{u}_{i,\infty} \right)dx,\\ H(\tilde{u}_\infty,r) &= \frac{1}{r^{N-1}}\sum_{i=1}^l\int_{\partial B_r} \tilde{u}_{i,\infty}^2\tilde{\mu}(x)d\sigma(x),\qquad N(\tilde{u}_\infty,r) = \frac{E(\tilde{u}_\infty,r)}{H(\tilde{u}_\infty,r)}. \end{align*} We divide the rest of the proof in two steps. \noindent \textbf{Step 1.} $\lim_{r \rightarrow 0^+}N(\tilde{u}_\infty,r) \geq 1$. To prove this step, we notice that, because of the convergence of $\tilde{u}_{i,\beta_n} \rightarrow \tilde{u}_\infty$ in the spaces $C(B_{M^{-\frac{1}{2}}})\cap H^1(B_{M^{-\frac{1}{2}}})$, we have that, for each $r \in ]0,\tilde{r}[$: \begin{equation} \label{EandHlimits} \lim_n E(\tilde{u}_{\beta_n},r) = E(\tilde{u}_{\infty},r), \qquad \lim_n H(\tilde{u}_{\beta_n},r) = H(\tilde{u}_\infty,r). \end{equation} Notice that the convergence of $E(\tilde{u}_{\beta_n},r)$ also comes from the fact that $ \beta_n \sum_{j\neq i} \int_{B_r} \tilde{a}_n(x) | \tilde{u}_{i,\beta_n} |^2 | \tilde{u}_{j,\beta_n} |^2 dx \rightarrow 0 $ (see for instance \cite[Theorem 1.4]{uniformHolderBoundsHugoTerraciniNoris}). A direct computation (see for instance \cite[Lemma C.5]{HugoHolderVariable} for the details) yields \begin{align*} H'(\tilde{u}_{\beta_n},r) &= \frac{1-N}{r}H(\tilde{u}_{\beta_n},r) + \frac{2}{r}E(\tilde{u}_{\beta_n},r) + \sum_{i=1}^l \frac{1}{r^{N-1}}\int_{\partial B_r} \tilde{u}_{i,\beta_n}^2\div(\tilde{A}_n(x)\nabla|x|) d\sigma(x) \end{align*} and so, passing to the limit and using \eqref{EandHlimits}: $$ H'(\tilde{u}_{\beta_n},r) \rightarrow \frac{1-N}{r}H(\tilde{u}_{\infty},r) + \frac{2}{r}E(\tilde{u}_{\infty},r) + \sum_{i=1}^l \frac{1}{r^{N-1}}\int_{\partial B_r} \tilde{u}_{i,\infty}^2\div(\tilde{A}(x)\nabla|x|)d\sigma(x). $$ On the other hand: \begin{align*} H(\tilde{u}_\infty,s_2) - H(\tilde{u}_\infty, s_1) &= \lim_n H(\tilde{u}_{\beta_n},s_2) - H(\tilde{u}_{\beta_n}, s_1) = \lim_n \int_{s_1}^{s_2} H'(\tilde{u}_{\beta_n},r) dr\\ &= \int_{s_1}^{s_2} \left( \frac{1-N}{r}H(\tilde{u}_{\infty},r) + \frac{2}{r}E(\tilde{u}_{\infty},r) + \sum_{i=1}^l \frac{1}{r^{N-1}} \int_{\partial B_r} \tilde{u}_{\infty,i}^2\div(\tilde{A}(x)\nabla|x|) d\sigma(x) \right), \end{align*} and from this we conclude that $\lim_n H'(\tilde{u}_{\beta_n},r) \to H'(\tilde{u}_\infty,r)$. Thus, applying these convergences, by passing to the limit the results in Lemma \ref{derivOfH}, we have that there exists $C>0$ such that: \begin{equation} \label{derivHBound} \Big| H'(\tilde{u}_\infty,r)-\frac{2}{r}E(\tilde{u}_\infty,r) \Big| \leq CH(\tilde{u}_\infty,r). \end{equation} Notice also that $\Big(N(\tilde{u}_{\beta_n},r) + 1\Big) e^{\tilde{C}r}$ is monotone nondecreasing, thus so is $(N(\tilde{u}_\infty,r)+1)e^{\tilde{C}r}$. We now suppose by contradiction that $\lim_{r \rightarrow 0^+} N(\tilde{u}_{\infty}, r) < 1$. Then there exists $\delta>0$ such that \begin{equation} \label{contradictionBoundOnN} N(\tilde{u}_{\infty}, r) < 1-\delta \qquad \forall r \in ]0,\overline{r}[. \end{equation} Now by using equation \eqref{derivHBound} and \eqref{contradictionBoundOnN}, for $r < \overline{r}$, we have: \begin{align} \frac{d}{dr}\log(H(\tilde{u}_\infty,r)) &= \frac{H'(\tilde{u}_\infty,r)} {H(\tilde{u}_\infty,r)} \leq \frac{2}{r} \frac{E(\tilde{u}_\infty, r)} {H(\tilde{u}_\infty, r)} + C = \frac{2N(\tilde{u}_\infty, r)}{r} + C \leq \frac{2(1-\delta)}{r} + C. \label{difEqOfHinfinity} \end{align} Integrating \eqref{difEqOfHinfinity} from $r \in ]0,\overline{r}[$ up to $\overline{r}$, we obtain: $$ \frac{ H(\tilde{u}_\infty,\overline{r}) }{ H(\tilde{u}_\infty,r) } \leq \Big(\frac{\overline{r}}{r}\Big)^{2(1-\delta)} e^{C(\overline{r}-r)}, $$ thus we conclude that there exists $c := \frac{H(\tilde{u}_\infty,\overline{r})}{\overline{r}^{2(1-\delta)}e^{C\overline{r}}}>0$ such that: $$ cr^{2(1-\delta)}< H(\tilde{u}_\infty,r) \qquad \forall r \in ]0,\overline{r}[. $$ On the other hand, since $\tilde{u}_\infty$ is bounded in $C^{0,\alpha}(B_{\overline{r}})$ for all $\alpha \in ]0,1[$, and since $\tilde{u}_\infty(0)=0$, there exists $C_\alpha>0$ such that: $$ |\tilde{u}_\infty(x)| = |\tilde{u}_\infty(x) - \tilde{u}_\infty(0)| \leq C_\alpha |x|^\alpha. $$ From this we then have the following bound: \begin{align*} H(\tilde{u}_\infty, r) &= \frac{1}{r^{N-1}}\sum_{i=1}^l\int_{\partial B_r} \tilde{u}_{i,\infty}^2\mu(x)d\sigma(x) \leq C_\alpha |\partial B_1| \|D\tilde{A}\|_{ L^\infty(B_{\overline{r}}) } r^{2\alpha} \end{align*} Therefore, for $C = |\partial B_1| \cdot C_\alpha \cdot \|D\tilde{A}\|_{ L^\infty(B_{\overline{r}}) }$, we have that: $$ cr^{2(1-\delta)}<H(\tilde{u}_\infty,r)\leq Cr^{2\alpha} \qquad \forall r \in ]0,\overline{r}[ $$ which is a contradiction for $r$ small, by choosing $2\alpha > 2(1-\delta)$ . So we conclude that $\lim_{r\rightarrow 0}N(\tilde{u}_\infty,r) \geq 1$. \noindent \textbf{Step 2. } $R_{\beta_n} \rightarrow 0$. Since $r\mapsto (N(\tilde{u}_{\beta_n},r)+1)e^{\tilde{C}r}$ is a continuous monotone nondecreasing function, converging pointwisely to the continuous function$r\mapsto (N(\tilde{u}_{\infty}, r)+1)e^{\tilde{C}r}$ in $]0,\tilde{r}[$, then the convergence is actually uniform over any compact subset in $]0,\tilde{r}]$ (see for example \cite[Lemma 4.3]{SoaveZilio}). We suppose by contradiction that $R_{\beta_n} \rightarrow R_\infty>0$. Then, using the definition \eqref{BigRdefinition} and the the uniform convergence: \begin{align*} 2 > 2-R_\infty &= \lim_n (2-R_{\beta_n}) \geq \lim_{n} (N(\tilde{u}_{\beta_n}, R_{\beta_n}) + 1)e^{\tilde{C}R_{\beta_n}} \\ &= (N(\tilde{u}_{\infty},R_\infty)+1)e^{\tilde{C}R_{\infty}}\geq (N(\tilde{u}_\infty, 0^+)+1)\geq 2, \end{align*} which is a contradiction. In the last inequality we used Step 1, $N(\tilde{u}_\infty, 0^+)\geq 1$. \end{proof} Now we apply Almgren's monotonicity formula to the blowup sequence $\{v_n\}$ given by (\ref{blowUp}). \begin{lemma} \label{MonotonicityForBlowup} Given the constants $\tilde{r}$ and $\tilde{C}>0$ of Lemma \ref{MonotonicityForu} for every $n$, let $\mu_n(x) = \langle A_n(x)\frac{x}{|x|}, \frac{x}{|x|}\rangle$. Then the functions: $$ r\mapsto\left(N(v_n,r)+1\right)e^{\tilde{C}r_n r}, \qquad r \mapsto H_{i}(v_n,r)e^{\tilde C r_n r}= \big( \frac{1}{r^{N-1}} \int_{\partial B_r} \mu_n(x) v_{i,n}^2 d\sigma(x) \big) e^{\tilde{C}r_nr}, $$ for all $i \in \{1,...,l\}$, are monotone nondecreasing for $r \in ]0,\frac{\tilde{r}}{r_n}[$. \end{lemma} \begin{proof} Recalling from \eqref{blowUp} that $v_n(x)=\frac{\eta(x_n)}{L_nr_n} \tilde{u}_{\beta_n}(r_n x)$, we have \[ E(v_n,r) = \frac{\eta^2(x_n)}{L_n^2 r_n^2}E(\tilde{u}_{\beta_n}, r_nr), \quad H_{i}(v_n,r)= \frac{\eta^2(x_n)}{L_n^2r_n^2}H_i(\tilde{u}_{\beta_n},r) \text{ and } N(v_n,r) = N(\tilde{u}_{\beta_n},r_nr). \] The claim follows as a direct application of Lemma \ref{MonotonicityForu}. \end{proof} Let $v_1 = \lim_n v_{1,n}$ and $v_2 = \lim_n v_{2,n}$ be the limits given by Proposition \ref{concludeProp} (the only two possible limiting components of the blowup sequence $v_n$). Next, we prove that \emph{both} $v_1$ and $v_2$ are nonconstant. \begin{lemma} \label{nonTriviality} Let $\theta$ be the constant from hypothesis \textbf{(A1)}. Then there exists $C = C(\theta, N)>0$, independent of $n$, such that: \begin{equation}\label{nonTriviality_eq} \frac{1}{r^{N-1}}\int_{\partial B_r}\mu_n(y) v_{i,n}^2d\sigma(y) \geq C \end{equation} for every $r \in [2N/\theta^\frac{1}{2},\frac{\tilde{r}}{r_n}]$ and $i=1,2$. In particular, both $v_1$ and $v_2$ are nonconstant in $B_r$ for every $r \in [2N/\theta^\frac{1}{2},\frac{\tilde{r}}{r_n}]$. \end{lemma} \begin{remark} The appearance of the constant $2N/\theta^{\frac{1}{2}}$ is directly related with Lemma \ref{harmonicLemma} in appendix: for a harmonic function $u$ such that $u(0) = 1$ and $|\nabla u (0)|\geq \theta^\frac{1}{2}$, such lemma implies that $u$ necessarily changes sign in $B_{2N/\theta^{\frac{1}{2}}}$. This fact is used in the following proof. \end{remark} \begin{proof}[Proof of Lemma \ref{nonTriviality}] By the monotonicity formula Lemma \ref{MonotonicityForBlowup}, we know that: $$ \frac{1}{r^{N-1}}\int_{\partial B_r}\mu_n(y) v_{i,n}^2 d\sigma(y) \geq \left( \frac{1}{(2N/\theta^\frac{1}{2})^{N-1}} \int_{\partial B_{2N/\theta^\frac{1}{2}}} \mu_n(y) v_{i,n}^2 d\sigma(y) \right) e^{\tilde{C}r_n(2N/\theta^\frac{1}{2}-\frac{\tilde{r}}{r_n})} $$ and so we only need to show that there exists $C>0$ such that $\int_{\partial B_{2N/\theta^\frac{1}{2}}}\mu_n(y) v_{i,n}^2d\sigma(y)>C$ for all $n \in \mathbb{N}$. We divide according to the asymptotic behaviour of $M_n$ (recall Proposition \ref{concludeProp}). \noindent \textbf{Case (i)} Assume that $M_n$ is bounded. By Proposition \ref{concludeProp} we know that $v_n \rightarrow v$ in $H^1_{loc}(\mathbb{R}^N)\cap C_{loc}(\mathbb{R}^N)$ and that $v_1,v_2$ are nonnegative functions satisfying the system \eqref{boundedEquation} and in particular are subharmonic in $\mathbb{R}^N$; moreover, $v_1(0)+v_2(0) = 1$. From this last fact, we may assume without loss of generality that there exists $C\geq \frac{1}{2}$ such that $v_1(0)\geq C>0$. Using the subharmonicity of $v_1$, we obtain: $$ \int_{\partial B_{2N/\theta^\frac{1}{2}} } v_1^2 d\sigma(y) > |\partial B_{2N/\theta^\frac{1}{2}}| v_1^2(0) = C>0, $$ and so, since $v_n \rightarrow v$ in $C_{loc}(\mathbb{R}^N)$, for $n$ large enough we know that: $$ \int_{\partial B_{2N/\theta^\frac{1}{2}}} v_{1,n}^2 d\sigma(y) \geq \frac{1}{2}C. $$ Now we suppose by contradiction that: \begin{equation}\label{eq_anothercontradassumption} \int_{\partial B_{2N/\theta^\frac{1}{2}}}v_{2,n}^2 d\sigma(y) \rightarrow 0. \end{equation} Using the subharmonicity of $v_2$, we conclude that $ v_2(x) = 0$ for $x \in B_{2N/\theta^\frac{1}{2}}$. Since $\gamma\geq 1$, $M_n<0$ and $v_2(x) = 0$ for all $x \in B_{2N/\theta^\frac{1}{2}}$, by the strong maximum principle we conclude that $v_2(x) = 0$ for all $x \in \mathbb{R}^N$. Going back to the system \eqref{boundedEquation}, then we conclude that $v_1$ is a nonnegative, nontrivial harmonic function in $\mathbb{R}^N$, which is a contradiction. Therefore \eqref{nonTriviality_eq} holds true. Function $v_1$ is nonconstant since $|\nabla v_1(0)|\geq \theta^{\frac{1}{2}}$, while $v_2$ is nonconstant by \eqref{boundedEquation}. \noindent \textbf{Case (ii)} If $M_n \rightarrow - \infty$, we know that $v_n \rightarrow v = (v_1,...,v_l)$ and $v_1, v_2$ satisfy the system \eqref{unboundedEquation}. Doing the same proof by contradiction as in Case (i), assume that $v_1(0)\geq C>0$ and that \eqref{eq_anothercontradassumption} holds true; then, as before, we conclude that: \begin{equation} \label{V2IsZero} v_2(y) = 0 \qquad \forall y \in B_{2N/\theta^\frac{1}{2}}. \end{equation} Now if $v_1(x_0) = 0$ for some $x_0 \in B_{2N/\theta^\frac{1}{2}}$, then $x_0$ is a zero of the limit profile $v = (v_1,...,v_l)$. The function $v$ satisfies the hypothesis of Theorem \ref{mult2Theoremyeah} in appendix, thus we have that $x_0$ is a zero of multiplicity at least 2. Since $v_1$ and $v_2$ are the only nontrivial components of $v$, that would imply there exists $y \in B_{2N/\theta^\frac{1}{2}}$ such that $v_2(y) > 0$, which is a contradiction with \eqref{V2IsZero}. Thus: \begin{equation*} v_1(y) > 0 \qquad \forall y \in B_{2N/\theta^\frac{1}{2}} \end{equation*} and, by the system \eqref{unboundedEquation}, this implies that $v_1$ is harmonic in the set $B_{2N/\theta^\frac{1}{2}}$. Now $v_1$ is nonnegative harmonic and satisfies $v_1(0) = 1$ and $|\nabla v_1(0)|\geq \theta^\frac{1}{2}$. A harmonic function with these properties by Lemma \ref{harmonicLemma} changes sign in $B_{2N/\theta^{\frac{1}{2}}}$, in contradiction with the fact that $v_1$ is positive in $B_{2N/\theta^{\frac{1}{2}}}$. In conclusion, we have deduced \eqref{nonTriviality_eq}. From \eqref{unboundedEquation} we know that $v_1v_2 = 0$, and using this fact we conclude that both $v_1$ and $v_2$ are nonconstant. \end{proof} We now introduce the following quantity: \begin{equation} \label{defOfBlowUpRadius} \overline{r}_n := \frac{R_{\beta_n}}{r_n} = \sup \left\{ r \in ]0, \frac{\tilde{r}}{r_n}[: \left( N(v_n,r) +1 \right)e^{\tilde{C}r_nr} \leq 2-rr_n \right\}. \end{equation} By Lemma \ref{radGoToZero} we know that: \begin{equation} \label{productLimit} \overline{r}_nr_n = R_{\beta_n} \rightarrow 0. \end{equation} The term $\overline{r}_n$ is a kind of threshold between sublinear (see for instance Lemma \ref{blowDownSequence}) and superlinear behavior for $v_n$. We refer to \cite[p. 640]{SOAVE2016388} for more insights. Recall that $\{v_n\}$ satisfies an Almgren monotonicity formula for $r\in ]0,\frac{\tilde r}{r_n}[$ (Lemma \ref{MonotonicityForBlowup}). For $r \in [\overline{r}_n, \frac{\tilde{r}}{r_n}]$, the function \[ \frac{ E(v_n,r) + H(v_n,r) }{ r^2 } \] is almost monotone (Lemma \ref{squaredMonotonicity} below). If $\overline{r}_n$ is bounded, it is not hard to obtain a contradiction (see the proof of Lemma \ref{radBlowUp}). Instead, if $\overline{r}_n\to \infty, $, then for $r \in [\overline{r}_n, \frac{\tilde{r}}{r_n}]$, we will see that $\{v_n\}$ satisfies an Alt-Caffarelli-Friedman type monotonicity formula (Lemma \ref{caffMonot}). All this information is combined to provide a contradiction also in this situation.. \begin{lemma} \label{sharpEstimate} Let $C$ be the constant from Lemma \ref{derivOfH}. Then: $$\Big| \frac{d}{dr}H(v_n,r) - \frac{2}{r}E(v_n,r) \Big| \leq C r_nH(v_n,r)\quad \forall r \in]0,\frac{\tilde{r}}{r_n}[$$ \end{lemma} \begin{proof} Since $v_n = \tilde{u}_{\beta_n}(r_nx)$ is a scaling, we just use Lemma \ref{derivOfH} and the identities of the proof of Lemma \ref{MonotonicityForBlowup} to conclude that: \begin{align*} \Big| \frac{d}{dr}&H(v_n,r) - \frac{2}{r}E(v_n,r) \Big| = \frac{\eta^2(x_n)} {L_n^2 r_n^2} \Big| \frac{d}{dr} \big( H(\tilde{u}_{\beta_n}, rr_n) \big) - \frac{2}{r}E(\tilde{u}_{\beta_n},rr_n) \Big|\\ & = \frac{\eta^2(x_n)} {L_n^2 r_n^2} \Big| r_n H'(\tilde{u}_{\beta_n}, rr_n) - \frac{2}{r r_n}r_n E(\tilde{u}_{\beta_n},rr_n) \Big| \leq C r_n \frac{\eta^2(x_n)} {L_n^2 r_n^2} \Big| H(\tilde{u}_{\beta_n}, rr_n) \Big| = C r_n H(v_n,r).\qedhere \end{align*} \end{proof} \begin{lemma} \label{doublingDownRefined}. There exists $C>0$ such that: \begin{enumerate} \item If there exists $\tilde{r}$ and $R$ such that $N(v_n,r)\leq d$ for all $0 \leq \tilde{r} \leq r \leq R \leq \frac{\tilde{r}}{r_n}$, then: $$ r \mapsto \frac{H(v_n,r)}{r^{2d}}e^{-Cr_nr}\quad \text{ is monotone nonincreasing for $r \in ]\tilde{r}, R[$.} $$ \item If there exists $\tilde{r}$ and $R$ such that $N(v_n,r)\geq \gamma$ for all $0 \leq \tilde{r} \leq r \leq R \leq \frac{\tilde{r}}{r_n}$, then: $$ r \mapsto \frac{H(v_n,r)}{r^{2\gamma}}e^{Cr_nr}\quad \text{ is monotone nondecreasing for $r \in ]\tilde{r},R[$.} $$ \end{enumerate} \end{lemma} \begin{proof} Follows from the proof of Lemma \ref{doublingLemma}, using the same scaling argument as in Lemma \ref{sharpEstimate}. \end{proof} We now define an auxiliary function that will be used in the next lemma. Given the constant $C$ of Lemma \ref{sharpEstimate} and $\tilde{C}$ of Lemma \ref{MonotonicityForBlowup} we define: \begin{equation*} \label{DefOfPhiAuxilliary} \varphi_n(r):= 2 \int_{\overline{r}_n}^{r} \left( 2\frac{ e^{-\tilde{C}r_nt}-1 }{ t } - \frac{ r_n\overline{r}_ne^{-\tilde{C}r_nt} }{ t } - \frac{Cr_n}{2} \right) dt . \end{equation*} \begin{remark} \label{uniformBoundOnPhi} We notice that this sequence is uniformly bounded in $L^\infty([0,\frac{\tilde{r}}{r_n}])$: there exists $\tilde{K}$ such that: \begin{align*} \varphi_n(r) &\leq 2 \int_{\overline{r}_n}^{r} \left( 2\frac{1-e^{-\tilde{C}r_nt}}{t} + \frac{ r_n \overline{r}_n e^{-\tilde{C}r_nt} }{ t } + \frac{ Cr_n }{ 2 } \right) dt \leq 2\left( \int_{\overline{r}_n}^{\frac{\tilde{r}}{r_n}} 2 \tilde{C} r_n dt+ \overline{r}_nr_n \int_{\overline{r}_n}^{\frac{\tilde{r}}{r_n}}\frac{dt}{t} + \frac{C\tilde{r}}{2}\right)\\ &\leq 4\tilde{C}\tilde{r} + 2r_n\overline{r}_n\left(|\log(\tilde{r})|+|\log(\overline{r}_nr_n)|\right)\leq \tilde{K}, \end{align*} since, by \eqref{productLimit}, we have $r_n\overline{r}_n \rightarrow 0$ and thus $r_n\overline{r}_n|\log(\overline{r}_nr_n)| \rightarrow 0$. \end{remark} \begin{lemma} \label{squaredMonotonicity} Let $\tilde{C}>0$ be the constant from Lemma \ref{MonotonicityForBlowup}. For every $n \in \mathbb{N}$, the function: $$ r \mapsto \frac{ E(v_n,r) + H(v_n,r) }{ r^2 } e^{\tilde{C}r_nr-\varphi_n(r)} $$ is monotone nondecreasing for $r \in [\overline{r}_n, \frac{\tilde{r}}{r_n}]$. \end{lemma} \begin{proof} By the definition of $\overline{r}_n$ in \eqref{defOfBlowUpRadius}, we know that: $$ (N(v_n,\overline{r}_n)+1)e^{\tilde{C}r_n\overline{r}_n} = 2-\overline{r}_nr_n. $$ Thus, for $r\in [\overline{r}_n, \frac{\tilde{r}}{r_n}]$, using the monotonicity of $(N(v_n,r)+1)e^{\tilde{C}r_nr}$ from Lemma \ref{MonotonicityForBlowup}, we obtain: $$ (N(v_n,r)+1)e^{\tilde{C}r_nr} \geq \left( N(v_n,\overline{r}_n)+1 \right)e^{\tilde{C}r_n\overline{r}_n} = 2-\overline{r}_nr_n $$ and $$ N(v_n,r)-1 \geq 2\left( e^{-\tilde{C}r_nr}-1 \right) - r_n\overline{r}_ne^{-\tilde{C}r_nr}. $$ Now, using Lemma \ref{sharpEstimate}, there exists $C>0$ independent of $n$ such that $H'(v_n,r)\geq \frac{2}{r}E(v_n,r)-Cr_nH(v_n,r)$, so: \begin{multline*} \frac{d}{dr}\log\left( \frac{ H(v_n,r) } { r^2 } \right) = \frac{H'(v_n,r)}{H(v_n,r)} - \frac{2}{r} \geq \frac{2}{r}\left( \frac{E(v_n,r)}{H(v_n,r)}-1 \right) - Cr_n\\ = \frac{2}{r}\left( N(v_n,r) - 1 \right) - Cr_n \geq \frac{4}{r}\left( e^{-\tilde{C}r_nr}-1 \right) - \frac{2}{r}r_n\overline{r}_ne^{-\tilde{C}r_nr} - Cr_n = \varphi'_n(r) \end{multline*} for $r \in [\overline{r}_n, \frac{\tilde{r}}{r_n}]$. This is equivalent to: $$ \frac{d}{dr} \left( \frac{H(v_n,r)}{r^2} \right) - \varphi_n'(r)\frac{H(v_n,r)}{r^2} \geq 0 $$ and integrating we deduce that $$ r\mapsto \frac{H(v_n,r)}{r^2}e^{-\varphi_n(r)} $$ is monotone nondecreasing for $r \in [\overline{r}_n, \frac{\tilde{r}}{r_n}]$. To conclude the proof we observe that, by Lemma \ref{MonotonicityForBlowup} and the above observations, for $r \in [\overline{r}_n, \frac{\tilde{r}}{r_n}]$: \begin{multline*} \frac{d}{dr}\log\left( \frac{E(v_n,r)+H(v_n,r)}{r^2}e^{\tilde{C}r_nr-\varphi_n(r)} \right) = \frac{d}{dr}\log\left( (N(v_n,r)+1)e^{\tilde{C}r_nr} \frac{H(v_n,r)}{r^2}e^{-\varphi_n(t)} \right)\\ = \frac{d}{dr}\log\left( (N(v_n,r)+1)e^{\tilde{C}r_nr} \right) + \frac{d}{dr}\log\left( \frac{H(v_n,r)}{r^2}e^{-\varphi_n(t)} \right) \geq 0\qedhere \end{multline*} \end{proof} \begin{lemma} \label{radBlowUp} It holds that $\overline{r}_n \rightarrow \infty$ as $n \rightarrow \infty$. \end{lemma} \begin{proof} We suppose by contradiction that there exists $\overline{r}$ such that, up to a subsequence, $\overline{r}_n \leq \overline{r}$. We know by Proposition \ref{concludeProp} that $v_n \rightarrow v$ in $C_{loc}(\mathbb{R}^n)$ and $H^1_{loc}(\mathbb{R}^n)$. We claim that the limit $v$ satisfies $E(v,r) \geq 0$. We divide the proof of this claim in two cases, according to Proposition \ref{concludeProp}. in case $M_n \rightarrow -\infty$, then: $$ \lim_n E(v_n,r) = E(v,r) = \frac{1}{r^{N-2}} \sum_{i=1}^l \int_{B_r} |\nabla v_i|^2 \geq 0, $$ while, if $M_n \rightarrow M_\infty < 0$, then: $$ \lim_n E(v_n,r) = E(v,r) = \frac{1}{r^{N-2}}\sum_{i=1}^l\left( \int_{B_r} |\nabla v_i|^2 - M_\infty\sum_{j\neq i} v_i^\gamma v_j^{\gamma+1} \right) \geq 0. $$ Then, for all $r \in [\overline{r}+1, \frac{\tilde{r}}{r_n}]$, using the monotonicity formula from Lemma \ref{squaredMonotonicity}, we obtain: \begin{align} 0 &\leq \frac{E(v,r) + H(v,r)}{r^2} = \lim_{n \rightarrow \infty} \frac{E(v_n,r) + H(v_n,r)}{r^2}\leq \lim_{n \rightarrow \infty} \frac{E(v_n,r) + H(v_n,r)}{r^2} e^{\tilde{C} r_nr - \varphi_n(r)} e^{\varphi_n(r)- \tilde{C}r_nr } \nonumber \\ &\leq \lim_{n \rightarrow \infty} r_n^2\frac{E(v_n,\frac{\tilde{r}}{r_n})+H(v_n,\frac{\tilde{r}}{r_n})}{\tilde{r}^2} \sup_{r\in[\overline{r}_n,\frac{\tilde{r}}{r_n}]} e^{\tilde{C}(\tilde{r}-r_nr)+\varphi_n(r)-\varphi_n(\frac{\tilde{r}}{r_n})} \nonumber \\ &\leq\lim_{n \rightarrow \infty} C' r_n^2\frac{E(v_n,\frac{\tilde{r}}{r_n}) + H(v_n,\frac{\tilde{r}}{r_n})}{\tilde{r}^2} = \lim_{n \rightarrow \infty} C'\eta^2(x_n)\frac{E(\tilde{u}_{\beta_n}, \tilde{r}) + H(\tilde{u}_{\beta_n}, \tilde{r})}{L_n^2\tilde{r}^2}, \label{eq:another_chain_of_inequalities_} \end{align} where we used the bound $e^{\tilde{C}(\tilde{r}-r_nr)+\varphi_n(r)-\varphi_n(\frac{\tilde{r}}{r_n})}\leq e^{2C\tilde{r}+2\tilde{K}} = C'$ (by Remark \ref{uniformBoundOnPhi}, $\varphi_n$ is uniformly bounded by $\tilde{K}$ in $[0,\frac{\tilde{r}}{r_n}]$). Now both $E(\tilde{u}_{\beta_n}, \tilde{r})$ and $H(\tilde{u}_{\beta_n}, \tilde{r})$ are uniformly bounded (for $E(\tilde{u}_{\beta_n},\tilde{r})$ proceed as in equation \eqref{boundOnDerivativeProp2} of Proposition \ref{ListProp}, using Lemma \ref{lemmaForMatrixTilde}). Thus, since $L_n \rightarrow \infty$, the last term in the chain of inequalities \eqref{eq:another_chain_of_inequalities_} to $0$, and so we conclude that $v(x)= 0$ for all $x \in B_{\overline{r}+1}(0)$, in contradiction with Lemma \ref{nonTriviality}. \end{proof} Up to this point, we just applied the results of Section \ref{chapter:implementation} (Almgren's monotonicity formula) to the blow up sequence $\{v_n\}$. It is now time to use the results of Section \ref{chapter:resultsChap4} (Alt-Caffarelli-Friedman-type monotonicity formula) to $\{v_n\}$ in the interval $r \in [2N/\theta^\frac{1}{2},\frac{\overline{r}_n}{3}]$. Given the functionals: $$ J_{1,n}(r) := \int_{B_r}\left( \langle A_n(y)\nabla v_{1,n} , \nabla v_{1,n} \rangle - M_na_n(y)v_{1,n}^2v_{2,n}^2 - v_{1,n}f_{1,n}(y,v_{1,n}) \right)|y|^{2-N}dy $$ $$ J_{2,n}(r) := \int_{B_r}\left( \langle A_n(y)\nabla v_{2,n}, \nabla v_{2,n} \rangle - M_na_n(y)v_{1,n}^2v_{2,n}^2 - v_{2,n}f_{2,n}(y,v_{2,n}) \right)|y|^{2-N}dy, $$ we also define: $$J_n(r) = \frac{J_{1,n}(r)J_{2,n}(r)}{r^4}.$$ \begin{lemma} \label{caffMonot} There exists $c,C>0$, independent of $n$, such that for every $0<\eta<\frac{1}{4}$ we have $$r\mapsto J_n(r)e^{-C|M_n|^{-\eta}r^{-2\eta} + Cr_n^2r^2 + Cr_nr}$$ is monotone nondecreasing in the interval $[2N/\theta^{-\frac{1}{2}}, \overline{r}_n/3]$ and $ J_n(2N/\theta^\frac{1}{2}) = \frac { J_{1,n}(2N/\theta^\frac{1}{2}) J_{2,n}(2N/\theta^\frac{1}{2}) } { (2N/\theta^\frac{1}{2})^4 } \geq c$. \end{lemma} The proof of this lemma consists in showing that the sequence $\{v_n\}$ satisfies conditions $(h0)$--$(h6)$ from Section \ref{chapter:resultsChap4} in the interval $[2N/\theta^{-\frac{1}{2}}, \overline{r}_n/3]$. From this, Lemma \ref{caffMonot} is a direct consequence of Theorem \ref{AltCaffMonotonicity}. However, the proof of such conditions is a delicate and long process. In order not to break the pace of this section, we leave it for later (it will be the content of Section \ref{sec:conditions} below). Instead, assuming the validity of Lemma \ref{caffMonot}, we immediatly pass to the proof of the main result of this paper. As previously done in Section \ref{chapter:resultsChap4}, for simplicity we focus on the case of $\gamma = 1$, remembering that the proof for general $\gamma$ follows from using Lemma \ref{estimateLemma}-(1) whenever we use Lemma \ref{estimateLemma}-(2). \begin{proof}[Proof of Theorem \ref{DesiredTheorem}] The claim that the following chain of inequalities hold true: \begin{align} \label{firstOfTheLast} 0 < c \leq J_n(2N/\theta^\frac{1}{2}) &\leq CJ_n(\frac{\overline{r}_n}{3}) \leq C\left( \frac{E(v_n,\overline{r}_n) + H(v_n,\overline{r}_n)}{\overline{r}_n^2} + o_n(1) \right)^2\\ &\leq C\left( r_n^2\frac{E(v_n,\frac{\tilde{r}}{r_n}) + H(v_n,\frac{\tilde{r}}{r_n})}{\tilde{r}^2} + o_n(1) \right)^2 \label{LatsOfTheLast} \end{align} where $o_n(1)\rightarrow 0$. A contradiction follows as soon as we prove this claim; indeed, using the same arguments as in Lemma \ref{radBlowUp}, we have that: $$ 0 < c \leq C\left( r_n^2\frac{E(v_n,\frac{\tilde{r}}{r_n}) + H(v_n,\frac{\tilde{r}}{r_n})}{\tilde{r}^2} + o_n(1) \right)^2 = C\left( \eta^2(x_n)\frac {E(\tilde{u}_{\beta_n},\frac{\tilde{r}}{r_n}) + H(\tilde{u}_{\beta_n},\frac{\tilde{r}}{r_n}) } { L_n^2\tilde{r}^2 } + o_n(1) \right)^2 \rightarrow 0, $$ which results in contradiction, and the theorem is proved. Now we prove the claim. The first two inequalities of \eqref{firstOfTheLast} follow from Lemma \ref{caffMonot}, since: $$ 0 < c \leq J_n(2N/\theta^\frac{1}{2}) \leq J_n(\frac{\overline{r}_n}{3}) e^{ -C|M_n|^{-\eta} \left( (\frac{\overline{r}_n}{3})^{-2\eta} - ( 2N/\theta^\frac{1}{2} )^{-2\eta} \right) + Cr_n^2 \left( ( \frac {\overline{r}_n} {3} )^2 - (2N/\theta^\frac{1}{2})^2 \right) + Cr_n \left( \frac {\overline{r}_n} {3} - (2N/\theta^\frac{1}{2}) \right) }. $$ By equation \eqref{productLimit} we have that $r_n \overline{r}_n \rightarrow 0$ and, by \eqref{Domain}, $r_n \rightarrow 0$. Furthermore, since by Proposition \ref{concludeProp} we either have $|M_n| \rightarrow \infty$ or $|M_n| \rightarrow |M_\infty|>0$, then this implies that $|M_n|^{-2\eta}(2N/\theta^\frac{1}{2})^{-2\eta}$ is bounded in $n$. With these observations, we conclude that: $$ e^{ -C|M_n|^{-2\eta} \left( \frac{\overline{r}_n}{3}^{-2\eta} - ( 2N/\theta^\frac{1}{2} )^{-2\eta} \right) + Cr_n^2 \left( \frac {\overline{r}_n} {3}^2 - (2N/\theta^\frac{1}{2})^2 \right) + Cr_n \left( \frac {\overline{r}_n} {3} - (2N/\theta^\frac{1}{2}) \right) } $$ is bounded in $n$, and so the second inequality of \eqref{firstOfTheLast} follows. The last inequality \eqref{LatsOfTheLast} follows from Lemma \ref{squaredMonotonicity}. The only thing left to prove is the middle inequality \begin{equation} \label{IneqProof} J_n(\frac{\overline{r}_n}{3}) \leq C\left( \frac{E(v_n,\overline{r}_n) + H(v_n,\overline{r}_n)}{\overline{r}_n^2} + o_n(1) \right)^2 \end{equation} and we proceed to prove it. First, we notice that: \begin{align*} \frac{1}{\overline{r}_n^2} \int_{B_{\overline{r}_n}/3} &\left( \langle A_n(y)\nabla v_{1,n},\nabla v_{1,n} \rangle - M_na_n(y)v_{1,n}^2v_{2,n}^2 - v_{1,n}f_{1,n}(y,v_{1,n}) \right) |y|^{2-N}dy \\ \leq& \frac{1}{\overline{r}_n^2}\int_{B_{\overline{r}_n}}\left( \langle A_n(y)\nabla v_{1,n},\nabla v_{1,n} \rangle - M_na_n(y)v_{1,n}^2v_{2,n}^2 - v_{1,n}f_{1,n}(y,v_{1,n}) \right) |y|^{2-N}dy \\ &+ \frac{1}{\overline{r}_n^2}\int_{B_{\overline{r}_n}\setminus B_{\overline{r}_n/3}} v_{1,n}f_{1,n}(y,v_{1,n})|y|^{2-N}dy. \end{align*} We divide the proof of \eqref{IneqProof} in two steps. \noindent \textbf{Step 1.} We show that: \begin{equation} \label{vanishTerm} \lim_{n \rightarrow \infty} \Big| \frac{1}{\overline{r}_n^2}\int_{B_{\overline{r}_n}-B_{\overline{r}_n/3}} v_{1,n}f_{1,n}(y,v_{1,n})|y|^{2-N}dy \Big| = 0. \end{equation} Indeed, using the bound $|f_{1,n}(x,v_{1,n})|\leq dr_n^2|v_{1,n}|$, obtained in Lemma \ref{lemmaMatrixBlowLimit}, we have that: \begin{multline*} \Big| \frac{1}{\overline{r}_n^2} \int_{B_{\overline{r}_n}(0)-B_{\overline{r}_n/3}(0)} v_{1,n} f_{1,n}(y,v_{1,n}) |y|^{2-N}dy \Big| \leq \frac{d r_n^2}{\overline{r}_n^2}\int_{B_{\overline{r}_n}(0)-B_{\overline{r}_n/3}(0)}v^2_{1,n}|y|^{2-N}dy\\ = \frac{d\eta^2(x_n)}{L_n^2\overline{r}_n^2r_n^2} \int_{B_{\overline{r}_nr_n}(0)-B_{\overline{r}_nr_n/3}(0)} \frac{\tilde{u}_{1,\beta_n}^2}{|y-x_n|^{N-2}}dy \leq \frac{ Cm^2 }{ L^2_n\overline{r}_n^Nr_n^N }\int_{B_{\overline{r}_nr_n}(0)}1dy \leq \frac{C}{L_n^2} \rightarrow 0, \end{multline*} where we use the uniform $L^\infty$-bound, $|\tilde{u}_{1,\beta_n}| \leq m$. \noindent \textbf{Step 2.} Show that: $$ \frac{1}{\overline{r}_n^2} \int_{B_{\overline{r}_n}(0)}\left( \langle A_n(y)\nabla v_{1,n}, \nabla v_{1,n} \rangle - M_na_n(y)v_{1,n}^2v_{2,n}^2 - v_{1,n}f_{1,n}(y,v_{1,n}) \right)|y|^{2-N} dy $$ $$ \leq C\frac{ E(v_n,\overline{r}_n) + H(v_n,\overline{r}_n)}{\overline{r}_n^2}. $$ To prove this, we use equation \eqref{randomEqJ} to conclude that there exists $\alpha>0$ such that: \begin{align} J_{1,n}(r_n) &=\int_{B_{\overline{r}_n}} \left( \langle A_n(y)\nabla v_{1,n}, \nabla v_{1,n} \rangle - M_na_n(y)v_{1,n}^2v_{2,n}^2 - v_{1,n} f_{1,n}(y,v_{1,n}) \right)|y|^{2-N} \nonumber \\ &\leq \frac{1}{\overline{r}_{n}^{N-2}} \int_{B_{\overline{r}_n}}\left( \langle A_n(y)\nabla v_{1,n}, \nabla v_{1,n} \rangle - M_na_n(y)v_{1,n}^2v_{2,n}^2 - v_{1,n} f_{1,n}(y,v_{1,n}) \right)dy \nonumber \\ &\qquad+ \frac{(N-2)(1+\alpha r_n\overline{r}_n)}{2\overline{r}_n^{N-1}}\int_{\partial B_{\overline{r}_n}}\mu_n(y)v_{1,n}^2d\sigma(y). \label{ineqFinal} \end{align} Now we notice that for all $n$ and $i\in\{1,...,l\}$, since $M_n<0$ and $a_n(y)>0$, we have that: \begin{multline*} \frac{1}{\overline{r}_n^{N-2}} \int_{B_{\overline{r}_n}}\left( \langle A_n(y)\nabla v_{i,n}, \nabla v_{i,n} \rangle - M_na_n(y)v_{i,n}^2 \mathop{\sum_{j=1}^l}_{j\neq i} v_{j,n}^2 - v_{i,n}f_{i,n}(y,v_{i,n}) \right) dy \\ \qquad+ \frac{(N-2)(1+\alpha r_n\overline{r}_n)}{2\overline{r}_n^{N-1}}\int_{\partial B_{\overline{r}_n}}\mu_n(y)v_{i,n}^2d\sigma(y)\\ \geq \frac{\eta^2(x_n)}{L_n^2r_n^2}\left( \int_{B_{{r_n\overline{r}_n}}} \left( \frac{1}{(r_n\overline{r}_n)^{N-2}} \langle \tilde{A}_n(y)\nabla \tilde{u}_{i,\beta_n}, \nabla \tilde{u}_{i,\beta_n} \rangle - \frac{d(\overline{r}_nr_n)^2}{(\overline{r}_nr_n)^N} \tilde{u}_{i,\beta_n}^2 \right) dy \right.\\ \qquad+ \left. \frac{(N-2)(1+\alpha r_n\overline{r}_n)}{2(\overline{r}_nr_n)^{N-1}} \int_{\partial B_{\overline{r}_nr_n}} \mu_n(\frac{y}{r_n}) \tilde{u}_{i,\beta_n}^2 d\sigma(y) \right)\\ \geq \frac{\eta^2(x_n)}{L_n^2r_n^2}\left( \int_{B_{{r_n\overline{r}_n}}} \left( \frac{1}{(r_n\overline{r}_n)^{N-2}} \frac{ \theta }{ M } |\nabla \tilde{u}_{i,\beta_n}|^2 - \frac{d(\overline{r}_nr_n)^2}{(\overline{r}_nr_n)^N} \tilde{u}_{i,\beta_n}^2 \right) dy \right.\\ \qquad \left. + \frac{ (N-2)(1+\alpha r_n\overline{r}_n)(1-Cr_n\overline{r}_n) }{ 2(\overline{r}_nr_n)^{N-1} } \int_{ \partial B_{\overline{r}_nr_n} } \tilde{u}_{i,\beta_n}^2 d\sigma(y) \right). \label{starEquation123} \end{multline*} In the first inequality we made a change of variables using the definition of $v_{i,n}$ in \eqref{blowUp} and the inequality of $f_{i,n}$ from Lemma \ref{lemmaMatrixBlowLimit}. In the last inequality, we used the ellipticity constant for $\tilde{A}_n(y) = A_n(\frac{y}{r_n})$ given in Lemma \ref{lemmaForMatrixTilde} and that: \begin{align} \mu_n(\frac{y}{r_n}) =& \langle A_n(\frac{y}{r_n}) \frac{y}{|y|} , \frac{y}{|y|} \rangle = 1 + \langle ( I - A_n(\frac{y}{r_n}) ) \frac{y}{|y|} , \frac{y}{|y|} \rangle= 1 + \langle ( I - \tilde{A}_n(y) ) \frac{y}{|y|} , \frac{y}{|y|} \rangle \geq (1-Cr_n\overline{r}_n) \end{align} for $y \in B_{r_n\overline{r}_n}$. Now, using Poincar\'e's inequality (Lemma \ref{PoincareIneq}) we have that \eqref{starEquation123} above is larger than or equal to: \begin{multline*} \frac{\eta^2(x_n)}{L_n^2r_n^2} \Bigg[ \frac{1}{(\overline{r}_nr_n)^{N-2}} \left( \frac{\theta}{M} - \frac{ d(r_n\overline{r}_n)^2 }{ N-1 } \right) \int_{B_{r_n\overline{r}_n}} |\nabla \tilde{u}_{i,\beta_n}|^2 dy \\ + \left( \frac{ (N-2)(1+\alpha r_n\overline{r}_n)(1-Cr_n\overline{r}_n) }{ 2 } - \frac{d(r_n\overline{r}_n)^2}{N-1} \right) \frac{1}{(\overline{r}_nr_n)^{N-1}} \int_{ \partial B_{\overline{r}_nr_n}(x_n) } \tilde{u}_{i,\beta_n}^2 d\sigma(y) \Bigg]. \end{multline*} Since $r_n\overline{r}_n \rightarrow 0$ the above is positive for $n$ large enough, thus we can assume that: \begin{align} &\frac{1}{\overline{r}_n^{N-2}} \int_{B_{\overline{r}_n}}\left( \langle A_n(y)\nabla v_{i,n}, \nabla v_{i,n} \rangle - M_na_n(y)v_{i,n}^2 \mathop{\sum_{j=1}^l}_{j\neq i} v_{j,n}^2 - v_{i,n}f_{i,n}(y,v_{i,n}) \right)dy \nonumber \\ &\qquad+ \frac{(N-2)(1+\alpha r_n\overline{r}_n)}{2\overline{r}_n^{N-1}}\int_{\partial B_{\overline{r}_n}}\mu_n(y)v_{i,n}^2(y)d\sigma(y) \geq 0 \label{ineqForAll} \end{align} for all $i \in \{1,...,l\}$ and large $n$. Now, coming back to equation \eqref{ineqFinal}, using inequality \eqref{ineqForAll} for $i=2,...,l$ we obtain: \begin{align*} & \int_{B_{\overline{r}_n}}\left( \langle A_n(y)\nabla v_{1,n}, \nabla v_{1,n} \rangle - M_na_n(y)v_{1,n}^2v_{2,n}^2 - v_{1,n}f_{1,n}(y,v_{1,n}) \right)|y|^{2-N} dy\\ &\leq \sum_{i=1}^l\left[ \frac{1}{\overline{r}_n^{N-2}} \int_{B_{\overline{r}_n}}\Big( \langle A_n(y)\nabla v_{i,n}, \nabla v_{i,n} \rangle - M_na_n(y)v_{i,n}^2 \mathop{\sum_{j=1}^l}_{j\neq i} v_{j,n}^2 - v_{i,n}f_{i,n}(y,v_{i,n}) \Big)dy \right.\\ \qquad & \qquad+ \frac{(N-2)(1+\alpha r_n\overline{r}_n)}{2\overline{r}_n^{N-1}}\int_{\partial B_{\overline{r}_n}}\mu_n(y)v_{i,n}^2d\sigma(y) \Bigg]\\ &\leq E(v_n,\overline{r}_n) + \frac{(N-2)(1+\alpha r_n\overline{r}_n)}{2}H(v_n,\overline{r}_n)\\ &\leq C\left(E(v_n,\overline{r}_n) + H(v_n,\overline{r}_n)\right). \end{align*} \noindent \textbf{Step 3.} $J_n(\frac{\overline{r}_n}{3})\leq C\left( \frac{E(v_n,\overline{r}_n) + H(v_n,\overline{r}_n)}{\overline{r}_n^2} + o_n(1) \right)^2.$ We can also do the same calculation for $v_{2,n}$ obtaining a vanishing term as in \eqref{vanishTerm} and inequalities similar to the ones from Step 1. and Step 2, and so we have: $$ J_{1,n} (\frac{\overline{r}_n}{3}) \leq C\left( \frac{E(v_n,\overline{r}_n) + H(v_n,\overline{r}_n)}{\overline{r}_n^2} + o_n(1) \right), \qquad J_{2,n} (\frac{\overline{r}_n}{3}) \leq C\left( \frac{E(v_n,\overline{r}_n) + H(v_n,\overline{r}_n)}{\overline{r}_n^2} + o_n(1) \right). $$ Taking the product of these inequalities we obtain the desired inequality \eqref{IneqProof}, concluding the proof. \end{proof} \section{Conditions for the Alt-Caffarelli-Friedman-type monotonicity formula: proof of Lemma \ref{caffMonot}}\label{sec:conditions} In this section we prove that Lemma \ref{caffMonot} is true (which was already used in the previous section to show Theorem \ref{DesiredTheorem}). This is a consequence of showing that conditions $(h_0)$-$(h_6)$ are satisfied by the blowup sequence $v_n$ defined in \eqref{blowUp}, to which we apply the Alt-Caffarelli-Friedman-type monotonicity formula, Theorem \ref{AltCaffMonotonicity}. Similarly to that section, for simplicity of notation we restrict our attention to the case $\gamma=1$. Remembering the main quantities at the beginning of Section \ref{chapter:resultsChap4}, we will set what some of them are in this context, while others will only be obtained through the proof of certain lemmas. We set \[ c_n := Cr_n,\qquad R_n := \frac{\overline{r}_n}{3}, \] where $r_n$ is defined in \eqref{Domain} and $\overline{r}_n$ in \eqref{defOfBlowUpRadius}. By Lemma \ref{radBlowUp}, for $n$ large enough, $R_n>1$. The constant $C$ is obtained from Lemma \ref{lemmaMatrixBlowLimit} in such a way that: $$ \sup_{y \in B_r(0)}\|A_n(y)-Id\|\leq Cr_n r. $$ With this, conditions ($h_0$) and ($h_4$) are automatically satisfied, since $r_n \overline{r}_n \rightarrow 0$. We now proceed to prove conditions $(h_1)$, $(h_2)$. With this we also define the quantity of Section \ref{chapter:resultsChap4} $$\epsilon_n := dr_n^2$$ \begin{lemma} \label{easyLemmaForConditions123} Provided $n$ is sufficiently large, there holds: $$ |f_{i,n}(x,v_{i,n})| \leq dr_n^2v_{i,n}, \qquad R_n^2\epsilon_n=\frac{dr_n^2\overline{r}_n^2}{9} \leq \big(\frac{N-2}{2}\big)^2-\delta, \qquad \text{ for small } \delta>0. $$ In particular, conditions $(h_1)$ and $(h_2)$ are satisfied. \end{lemma} \begin{proof} This is an easy consequence of Lemma \ref{lemmaMatrixBlowLimit} and the fact that $r_n\overline{r}_n \rightarrow 0$ as $n \rightarrow \infty$. \end{proof} The rest of the conditions are proved in different lemmas below. Condition $(h_3)$ is proved in Lemma \ref{sameWeightAltCafCond}, $(h_5)$ is proved in Lemma \ref{almgrenLemmaHypoithesisStuff} and $(h_6)$ is proved in Lemmas \ref{LambdasMoreThanZeroAltCafCond} and \ref{LastLemma}. Next, we are going to prove a couple of lemmas that will be the main tools for the rest of this section. In particular, Lemma \ref{blowDownSequence} concerns the characterization of certain blowdown-sequences. \begin{lemma} \label{doublingForAlt} There exists $\sigma \in ]0,1[$ such that \[ \text{ $\sigma \leq N(v_n,r) \leq 1$ for every $r \in [2N/\theta^\frac{1}{2},\overline{r}_n]$ for every $n$. } \]As a consequence, from Lemma \ref{doublingDownRefined}, there exists $C>0$ such that: $$ r \mapsto \frac{H(v_n, r)}{r^2}e^{-Crr_n}\quad \text{ is monotone nonincreasing for $r \in [2N/\theta^\frac{1}{2}, \overline{r}_n]$,} $$ and: $$ r \mapsto \frac{H(v_n,r)}{r^{2\sigma}}e^{Crr_n}\quad \text{ is monotone nondecreasing for $r \in [2N/\theta^\frac{1}{2},\overline{r}_n]$. } $$ \end{lemma} \begin{proof} From the Almgren monotonicity formula (Lemma \ref{MonotonicityForBlowup}) and the definition of $\overline{r}_n$ in \eqref{defOfBlowUpRadius}, we obtain: \begin{align*} N(v_n,r) +1 \leq \left( N(v_n,r) + 1 \right) e^{\tilde{C}r_nr} \leq \left( N(v_n,\overline{r}_n) + 1 \right)e^{\tilde{C}r_n\overline{r}_n} = 2 - r_n\overline{r}_n, \end{align*} so that $$ N(v_n,r) \leq (2-r_n\overline{r}_n) - 1 = 1-r_n\overline{r}_n $$ for all $r \in [0,\overline{r}_n]$. This gives the upper bound on $N(v_n,r)$. For the lower bound, we use again the Almgren monotonicity formula to conclude that: $$ \left( N(v_n,r) + 1 \right) e^{\tilde{C}r_nr} \geq \left( N(v_n,2N/\theta^{\frac{1}{2}}) + 1 \right) e^{\tilde{C}r_n2N/\theta^{\frac{1}{2}}} $$ for every $r \in [2N/\theta^\frac{1}{2},\overline{r}_n]$, which implies: \begin{equation} \label{eqinN} N(v_n,r) \geq \left( N(v_n,2N/\theta^{\frac{1}{2}}) + 1 \right) e^{-\tilde{C}r_nr} - 1. \end{equation} By Lemma \ref{nonTriviality}, the limits $v_{1}$ and $v_{2}$ are nonconstant in $B_{2N/\theta^\frac{1}{2}}$, thus there exists $\tilde{C}>0$ such that: $ \tilde{C} \leq \sum_{i=1}^l \int_{B_{2}} |\nabla v_i|^2 dx. $ Now, by Lemma \ref{lemmaMatrixBlowLimit}, we know that $f_{i,n}(x,v_{i,n})$ converges locally uniformly to zero, thus by the convergence of $v_n$ in $H^1(B_{2N/\theta^\frac{1}{2}})\cap C(\overline{B}_{2N/\theta^\frac{1}{2}})$ and that $M_n < 0$, $a_n(x)>0$ we obtain: \begin{align*} 0<\tilde{C} &\leq \sum_{i=1}^l \int_{B_{2}} |\nabla v_i|^2 dx \leq E(v,2N/\theta^\frac{1}{2})\\ &\leq \lim_n \sum_{i=1}^l \int_{B_{2N/\theta^{\frac{1}{2}}}}\left( \langle A_n(x)\nabla v_{i,n}, \nabla v_{i,n} \rangle - 2M_n\sum_{i<j}a_n(x)v_{i,n}^2v_{j,n}^2 \right)dx\leq \lim_n E(v_n,2N/\theta^\frac{1}{2}). \end{align*} Also by the local uniform convergence $v_n \rightarrow v$ we have $\alpha =H(v,2N/\theta^\frac{1}{2}) = \lim_n H(v_n,2N\theta^\frac{1}{2})$ thus: $N(v_n,2N/\theta^{\frac{1}{2}})\geq \frac{\tilde{C}}{2\alpha}>0$ for $n$ large enough. Since $r_nr \leq r_n \overline{r}_n \rightarrow 0$ as $n \rightarrow \infty$, coming back to equation \eqref{eqinN} we obtain: \[ N(v_n,r) \geq (\frac{\tilde{C}}{2\alpha}+1)e^{-\tilde{C}r_nr}-1 \geq \sigma > 0\quad \text{ for all $r \in [2N/\theta^\frac{1}{2},\overline{r}_n]$. } \qedhere \] \end{proof} We will now show that the limit of the blowdown sequence defined below in \eqref{linearLimit} behaves linearly in a ball $B_1$. For the next lemma, we recall that $\overline{r}_n \rightarrow \infty$ by Lemma \ref{radBlowUp}. \begin{lemma} \label{blowDownSequence} Let $(\rho_n)$ be a sequence such that $\rho_n \rightarrow \infty$ and $\rho_n \leq \frac{\overline{r}_n}{3}$. Then there exists $h,k \in \{1,...l\}$ and $\gamma_h,\gamma_k>0$ such that the blowdown sequence: \begin{equation} \label{linearLimit} \tilde{v}_{i,n}(x) := \frac{ v_{i,n}(\rho_n x) }{ \sqrt{H(v_n,\rho_n)} } \end{equation} converges in $H^1(B_1)\cap C(\overline{B_1})$, up to a rotation, to a function $\tilde{v}=(\tilde{v}_1,\ldots, \tilde{v}_l)$ defined by: $$ \tilde{v}_h(x) = \gamma_h x_1^+, \qquad \tilde{v}_k(x) = \gamma_k x_1^-, \qquad \tilde{v}_j(x) = 0, \qquad \forall j \neq h,k. $$ \end{lemma} \begin{proof} First we observe that the sequence $\tilde{v}_{n}$ satisfies the system: \begin{equation} \label{blowDownEq} -\div(A_n(\rho_n x)\nabla \tilde{v}_{i,n}) = \frac{ \rho_n^2 }{ \sqrt{H(v_n,\rho_n)} } f_{i,n}(\rho_n x,v_{i,n}(\rho_n x)) + \rho_n^2 H(v_n,\rho_n) M_n \tilde{v}_{i,n} \sum_{j \neq i} a_n(\rho_n x)\tilde{v}_{j,n}^2 \end{equation} in a set $B_3 \subset \tilde{\Omega}_n = \frac{\Omega_n}{\rho_n}$ (this follows from Proposition \ref{ListProp}, since $B_{1/M^\frac{1}{2}r_n} \subset \Omega_n$ and $r_n\rho_n\rightarrow 0$). Since $\rho_n \rightarrow \infty$, by Lemma \ref{nonTriviality} and by Proposition \ref{concludeProp} we know there exists $C>0$ small enough such that $H(v_n,\rho_n)\geq C$ and $M_n \leq -C<0$ for all $n$. With this we can conclude that the competition parameter in equation \eqref{blowDownEq} satisfies $\rho_n^2H(v_n,\rho_n)M_n\rightarrow -\infty$. Also, by Lemma \ref{lemmaMatrixBlowLimit}, it follows that: \begin{align*} -\div(A_n(\rho_n x)\nabla \tilde{v}_{i,n}) &\leq \frac{ \rho_n^2 }{ \sqrt{H(v_n,\rho_n)} } f_{i,n}(\rho_nx, v_{i,n}(\rho_n x))\leq \frac{ \rho_n^2 }{ \sqrt{H(v_n,\rho_n)} } dr_n^2v_{i,n}(\rho_n x) \leq d(\rho_n r_n)^2 \tilde{v}_{i,n}(x) \end{align*} in $B_3$. By the ellipticity of $A_n(\rho_n x)$ and a Brezis-Krato-type argument (see for instance \cite[Appendix B.2, B.3]{Struwe}), if we show that $\tilde{v}_{i,n}$ has a uniform bound in $H^1(B_3)$ then it follows that the sequence $\tilde{v}_{i,n}$ has a uniform bound in $L^\infty(B_2)$. Now by Lemma \ref{doublingForAlt} we know that: \begin{equation} \label{upBoundOnN1236} N(\tilde{v}_n,\rho) = N(v_n,\rho \rho_n) \leq 1 \end{equation} for every $0\leq \rho\leq 3$. Thus, by Lemma \ref{doublingForAlt}, for all $1 \leq \rho \leq 3$ there exists $C>0$ such that: \begin{align*} H(\tilde{v}_n,\rho) &= \frac{1}{\rho^{N-1}} \sum_{i=1}^l \int_{\partial B_\rho} \langle A_n(\rho_ny) \frac{y}{|y|} , \frac{y}{|y|} \rangle \tilde{v}_{i,n}^2 d\sigma(y)= \frac{1}{(\rho\rho_n)^{N-1}H(v_n,\rho_n)} \sum_{i=1}^l \int_{\partial B_{\rho\rho_n}} \mu_n(y) v_{i,n}^2 d\sigma(y)\\ &= \frac{ H(v_n,\rho_n\rho) }{ H(v_n,\rho_n) } \leq e^{-C(r_n\rho_n\rho-r_n\rho_n)}\rho^2. \end{align*} Therefore, since $r_n\rho_n \leq r_n \overline{r}_n \rightarrow 0$, $$ E(\tilde{v}_n,3) = N(\tilde{v}_n,3) H(\tilde{v}_n,3) \leq 9e^{-3Cr_n\rho_n}. $$ With this upper bound for the energy we are able to show the upper bound for the $H^1(B_3)$ norm. Indeed, using the ellipticity constant for $A_n$ and the property for $f_i$ from Lemma \ref{lemmaMatrixBlowLimit}, and Poincar\'e's inequality (Lemma \ref{PoincareIneq}): \begin{align*} E(\tilde{v}_n,3) &\geq \sum_{i=1}^l \frac{1}{3^{N-2}} \int_{B_3} \left( \langle A_n(\rho_n x)\nabla \tilde{v}_{i,n} , \nabla \tilde{v}_{i,n} \rangle + \frac{\rho_n^2}{\sqrt{H(v_n,\rho_n)}} f_{i,n}(\rho_nx, v_{i,n}(\rho_nx)) \tilde{v}_{i,n} \right) dx\\ &\geq \sum_{i=1}^l \int_{B_3} \left( \frac{\theta}{M 3^{N-2}} |\nabla \tilde{v}_{i,n}|^2 - \frac{d3^2(\rho_n r_n)^2}{3^{N}} \tilde{v}^2_{i,n} \right) dx\\ &\geq \sum_{i=1}^l \int_{B_3} \frac{\theta}{M 3^{N-2}} |\nabla \tilde{v}_{i,n}|^2 dx - \frac{d3^2(\rho_n r_n)^2}{(N-1)} \left( \int_{B_3} \frac{1}{3^{N-2}} |\nabla \tilde{v}_{i,n}|^2 dx + \int_{\partial B_3} \frac{1}{3^{N-1}} \tilde{v}_{i,n}^2 d\sigma(x) \right) \\ &\geq \frac{1}{3^{N-2}} (\frac{\theta}{M}-\frac{d3^2(\rho_n r_n)^2}{(N-1)}) \int_{B_3} \sum_{i=1}^l |\nabla \tilde{v}_{i,n}|^2 dx - \frac{d3^2M}{\theta(N-1)}(\rho_n r_n)^2 H(\tilde{v}_n,3)\\ &\geq C\int_{B_3}\sum_{i=1}^l|\nabla \tilde{v}_{i,n}|^2dx - o_n(1) \end{align*} by the observations above and since $\rho_nr_n\to 0$. This gives the desired $H^1(B_3)$ bound, and so $\tilde{v}_n$ is uniformly bounded in $L^\infty(B_2)$. With this, using Proposition \ref{propOfBlowUpInNiceSpace} in appendix and $\rho_n \rightarrow \infty$, we conclude that there exists $\tilde{v} \in C(B_{\frac{3}{2}}) \cap H^1(B_{\frac{3}{2}})$ such that $\tilde{v}_n \rightarrow \tilde{v}$ in both $C(B_{\frac{3}{2}}) \cap H^1(B_{\frac{3}{2}})$. Moreover, we have that $\tilde{v} \in \mathcal{G}(B_{\frac{3}{2}})$ (see Definition \ref{GspaceDef} below). By Lemma \ref{lemmaMatrixBlowLimit}, we have: $$ \Big| \frac{\rho_n^2}{\sqrt{H(v_n,\rho_n)}} f_{i,n}(\rho_n x, v_{i,n}(\rho_nx)) \Big| \leq C(\rho_n r_n)^2 \|\tilde{v}_{i,n}\|_{L^\infty(B_2)} \rightarrow 0 $$ By a proof like the one of Proposition \ref{concludeProp}, since $\|I-A_n(\rho_n y)\| \leq Cr_n\rho_n|y| \rightarrow 0$ and $\rho_n^2 H(v_n,\rho_n) M_n \rightarrow -\infty$, we obtain: $$ \Delta \tilde{v}_i(x) = 0\quad \text{ for $x \in \{\tilde{v}_i>0\}$.} $$ Moreover, we have that $0 \in \{\tilde{v}=0\}$, since the sequence $(v_n(0))$ is bounded, while by Lemma \ref{doublingForAlt}, $$ H(v_n,\rho_n) \geq \frac{ H(v_n,2) }{ 4^\sigma } \rho_n^{2\sigma} e^{Cr_n(2-\rho_n)} \rightarrow +\infty $$ as $n \rightarrow \infty$ since $r_n\rho_n \rightarrow 0$. This shows that $\tilde{v}(0) = \lim \tilde{v}_n(0) =\lim \frac{v_n(0)}{\sqrt{H(v_n,\rho_n)}} = 0$. Using Proposition \ref{theoremForNandFreeStuff}, since $\tilde{v}(0) = 0$, that is $0 \in \{\tilde{v} = 0\}$, we have that $1\leq N(\tilde{v},0^+)$. Also, since $\tilde{v} \in \mathcal{G}(B_{\frac{3}{2}})$, by Proposition \ref{homogenousSegregatedLemma}, we conclude that the function $N(\tilde{v},r)$ is monotone increasing for $r \in [0,\frac{3}{2}]$. Also by equation \eqref{upBoundOnN1236} we have that $N(\tilde{v},r)\leq 1$ for all $r \in [0,\frac{3}{2}]$. From this we conclude the chain of inequalities $$ 1 \leq N(\tilde{v},0^+) \leq N(\tilde{v},r) \leq 1 $$ for all $r \in ]0,\frac{3}{2}[$. This implies that $N(\tilde{v},r)$ is constant equal to $1$, thus by Proposition \ref{homogenousSegregatedLemma} we conclude that $\tilde{v}$ is a homogenous function of degree $1$ at zero. By Theorem \ref{mult2Theoremyeah} we conclude that there must exist two nontrivial components of $\tilde{v}$ around zero, since it is the limit of $\tilde{v}_n$, solutions of competition systems. By Lemma \ref{harmonic2compsystemCharacter} there must exist indices $h,k \in \{1,...,l\}$ and constants $\gamma_h, \gamma_k>0$ such that up to a rotation: \[ \tilde{v}_h(x) = \gamma_h x_1^+ ,\qquad \tilde{v}_k(x) = \gamma_k x_1^- ,\qquad \tilde{v}_j(x) = 0 \quad \forall j \neq h,k. \qedhere \] \end{proof} \begin{lemma} \label{sameWeightAltCafCond} There exists $\lambda > 0$ independent of $n$ such that: $$ \frac{1}{\lambda} \leq \frac{ \int_{\partial B_r} \mu_n(y)v_{1,n}^2 d\sigma(y) }{ \int_{\partial B_r} \mu_n(y)v_{2,n}^2 d\sigma(y) } \leq \lambda $$ for every $2N/\theta^{\frac{1}{2}} \leq r \leq \frac{\overline{r}_n}{3}$. On the contrary for $j = 3,...,N$ we have: $$ \sup_{r \in [2N/\theta^\frac{1}{2},\overline{r}_n/3]} \frac{ \int_{\partial B_r} \mu_n(y) v_{j,n}^2 d\sigma(y) }{ \int_{\partial B_r} \mu_n(y) v_{1,n}^2 d\sigma(y) } \rightarrow 0 $$ as $n \rightarrow \infty$. This combined with Lemma \ref{nonTriviality} shows hypothesis $(h_3)$ of Section \ref{chapter:resultsChap4}. \end{lemma} \begin{proof} For this we use Lemma \ref{simplexLemma} from the appendix. Given the sequence $\{v_n\}$, we consider the auxiliary functions: $$ g_{i,n}(\rho) := \begin{cases} \frac{1}{H(v_n,\rho \overline{r}_n/3)} \frac{1}{(\rho \overline{r}_n/3)^{N-1}} \int_{\partial B_{\rho \overline{r}_n/3}} \mu_n(y) v_{i,n}^2 d\sigma(y) & \text{ for } \frac{6N/\theta^\frac{1}{2}}{\overline{r}_n} \leq \rho \leq 1\\ \frac{1}{H(v_n,2N/\theta^{\frac{1}{2}})} \frac{1}{(2N/\theta^\frac{1}{2})^{N-1}} \int_{ \partial B_{2N/\theta^\frac{1}{2}} } \mu_n(y) v_{i,n}^2 d\sigma(y) & \text{ for } 0\leq \rho \leq \frac {6N/\theta^\frac{1}{2}} {\overline{r}_n}. \end{cases} $$ The proof is finished once we have proved the assumptions of Lemma \ref{simplexLemma}, that is $\lim_n \dist(g_n([0,1]),\Sigma_{2,l}) = 0$, where $ \Sigma_{2,l} := \Big\{ x \in \mathbb{R}^l: \exists i,j \in \{1,...l\},\ i\neq j,\ \text{ such that } x_h = 0 \quad \forall h \neq i,j \Big\}. $ By Lemma \ref{nonTriviality}, we must have that the indices for the nontrivial components should be $i=1,2$. By construction, each $g_{i,n}$ is continuous, $g_{i,n} \geq 0$, and $\sum_{i=1}^lg_{i,n}(x) = 1$ for all $x \in [0,1]$. We divide the proof into two steps: \noindent \textbf{(i)} First, we prove that there exists $\epsilon \in ]0,1[$ such that $g_{i,n}(x)\leq 1-\epsilon$ for all $x \in [0,1]$, $n \in \mathbb{N}$, $i \in \{1,...,l\}$. By contradiction, we assume there exists an index $i \in \{1,...,l\}$ and a sequence $s_n \in [0,1]$ such that: \begin{equation} \label{contradHypNonTriv} g_{i,n}(s_n) \rightarrow 1, \qquad g_{j,n}(s_n) \rightarrow 0 \qquad i,j \in \{1,...,l\} \text{ and } i \neq j. \end{equation} By Lemma \ref{nonTriviality} and the local uniform convergence $v_n \rightarrow v$, we conclude that $s_n\overline{r}_n \rightarrow \infty$. Indeed, were this not true and \eqref{contradHypNonTriv} would not be possible, since if $s_n \overline{r}_n \rightarrow \tilde{r}$ then: $$ \lim_n g_{i,n}(s_n) = \int_{\partial B_{\tilde{r}/3}} \frac{ \mu_n(y) v_{i}^2 } { H(v,\tilde{r}/3) (\tilde{r}/3)^{N-1} } d\sigma(y) >0 $$ for $i=1,2$ in case $\tilde{r}\geq 6N/\theta^\frac{1}{2}$, while in case $\tilde{r}< {6N}/{\theta^\frac{1}{2}}$ we also conclude that $$ \lim_n g_{i,n}(s_n) = \int_{\partial B_{2N/\theta^\frac{1}{2}}} \frac{\mu_n(y) v_{i}^2 } { H(v,2N/\theta^\frac{1}{2}) (2N/\theta^\frac{1}{2})^{N-1} } d\sigma(y) >0\qquad \text{for $i=1,2$.} $$ We consider the blowdown sequence given by \begin{equation}\label{blowdownapplied} \tilde{v}_{i,n}(x) := \frac{ v_{i,n}(s_n\overline{r}_nx/3) }{ \sqrt{H(v_n,s_n\overline{r}_n/3)} }, \end{equation} to which we apply Lemma \ref{blowDownSequence} with $\rho_n:=s_n\overline{r}_n/3$, concluding that the uniform limit of $\tilde{v}_n$ contains two two nontrivial components, in contradiction with (\ref{contradHypNonTriv}). \noindent \textbf{(ii)} Now we prove $\lim_{n} \dist(g_n([0,1]), \Sigma_{2,l}) = 0$. We assume by contradiction that there exists $\epsilon>0$ and three different indices $i,j,k$ and a sequence $s_n \in ]0,1[$ such that up, to a subsequence, \begin{align} \label{contrad2NonTriv} g_{i,n}(s_n)\geq \epsilon , \quad\quad g_{j,n}(s_n)\geq \epsilon , \quad\quad g_{k,n}(s_n)\geq \epsilon. \end{align} Again, we must have $s_n\overline{r}_n \rightarrow \infty$, otherwise, since by Proposition \ref{concludeProp} the limit $v = \lim v_n$ has a maximum of two nontrivial components, we would have $ \lim_n g_{i,n}(s_n) =0$ for $i \neq 1,2$. Exactly as before, considering again the blowdown sequence \eqref{blowdownapplied} and since $s_n\overline{r}_n/3 \rightarrow \infty$, we apply Lemma \ref{blowDownSequence} to conclude that the uniform limit of $v_n$ contains exactly two nontrivial components, in contradiction with equation \eqref{contrad2NonTriv}. \end{proof} \begin{lemma} \label{LambdasMoreThanZeroAltCafCond} There exists $C>0$ independent of $n$ such that: $$ \Lambda_{1,n}(r), \\ \Lambda_{2,n}(r)\geq C $$ for $r \in [2N/\theta^\frac{1}{2},\frac{\overline{r}_n}{3}]$. In particular the second condition of $(h_6)$ holds true. \end{lemma} \begin{proof} By contradiction, we assume there exists $\rho_n \in [2N/\theta^\frac{1}{2},\overline{r}_n/3]$ such that $\lim_n\Lambda_{1,n}(\rho_n) \leq 0$, that is: \begin{equation} \label{LambdaContradiction} \lim_n \rho_n^2\frac{ \int_{\partial B_{\rho_n}} \left( \langle B_n(x)\nabla_{\theta}v_{1,n}, \nabla_{\theta} v_{1,n} \rangle - M_na_n(x)v_{2,n}^2 v_{1,n}^2 + v_{1,n}f_{1,n}(x,v_{1,n}) \right) d\sigma(x) }{ \int_{\partial B_{\rho_n}} (1+\alpha \rho_n r_n)\mu_n(x) v_{1,n}^2 d\sigma(x) } \leq 0. \end{equation} We either have that $\rho_n$ is bounded or $\rho_n \rightarrow \infty$. \noindent \textbf{(i)} If $\rho_n \rightarrow \infty$, then we consider the scaled blowdown sequence: $$ \tilde{v}_{i,n}(x) := \frac{ v_{i,n}(\rho_nx) }{ \sqrt{H(v_n,\rho_n)} } $$ where $\rho_n \leq \overline{r}_n/3$. From Lemma \ref{blowDownSequence} we know that $\tilde{v}_n \rightarrow \tilde{v}$ uniformly, such that (up to a rotation): $$ \tilde{v}_i = \gamma_i x_1^+ ,\,\, \tilde{v}_j = \gamma_j x_1^-,\,\, \tilde{v}_k = 0 $$ for all $k \neq i,j$ and $\gamma_i, \gamma_j > 0$. Due to Lemma \ref{nonTriviality} we conclude that $i = 1$ and $j =2$. The idea is to turn the uniform convergence of $\tilde{v}_{1,n} \rightarrow \tilde{v}_1$ into a $C^{1,\alpha}$ convergence for $0<\alpha<1$ in a set away from the free boundary given by $\{\gamma_1x_1 > 2\tilde{\delta}\}$ for some $\tilde{\delta} > 0$. We take $\tilde{\delta}$ sufficiently small so that $\{\gamma_1x_1 > 2\tilde{\delta}\} \cap \partial B_1 \neq \emptyset$. If $x_0 \in B_2(0) \cap \{\gamma_1 x_1 >\tilde{\delta}\}$ and $\rho>0$ is small enough so that $B_\rho(x_0) \subset \{\gamma_1x_1>\tilde{\delta}/2\}$, then by uniform convergence of $\tilde{v}_{1,n}$ to $\tilde{v}_1$: \begin{equation} \label{lowerBoundForBlowDown} \tilde{v}_{1,n}\geq \frac{\tilde{\delta}}{4}>0 \qquad \text{ in } B_2\cap\{\gamma_1x_1>\tilde{\delta}/2\}. \end{equation} Equation \eqref{lowerBoundForBlowDown}, equation \eqref{blowDownEq} and \textbf{(a)} -- which provides $a_n(x)\geq \delta$ --, imply the following inequality for $j\neq 1 = i$: \begin{align} -\div &(A_n(\rho_n x)\nabla \tilde{v}_{j,n}) \leq \left( d\rho_n^2r_n^2 - 2 Ca_n (\rho_n x) H(v_n,\rho_n)\rho_n^2 |M_n| \right) \tilde{v}_{j,n}(x) \nonumber \\ & \leq \left( d\rho_n^2r_n^2 - 2 C \delta H(v_n,\rho_n)\rho_n^2|M_n| \right) \tilde{v}_{j,n}(x) \leq -C \delta H(v_n,\rho_n)\rho_n^2|M_n|\tilde{v}_{j,n}(x), \label{boundOnDivergenceAltCafConditions} \end{align} since $\rho_nr_n \rightarrow 0$ and $H(v_n,\rho_n)\rho_n^2 M_n \rightarrow -\infty$ using Proposition \ref{concludeProp} and Lemma \ref{nonTriviality}. Applying Lemma \ref{estimateLemma}-1 to \eqref{boundOnDivergenceAltCafConditions}, we conclude the uniform bound: $$ |H(v_n,\rho_n)\rho_n^2M_n\tilde{v}_{j,n}(x_0)| \leq C \quad \text{ for every $x_0 \in B_{2-\rho}(0) \cap \{\gamma_1x_1>2\tilde{\delta}\}$}, $$ which implies the uniform boundedness of $\div(A_n(\rho_n x)\nabla \tilde{v}_{1,n})$. This together with the uniform convergence $\tilde{v}_{1,n} \rightarrow \tilde{v}_1$ implies that it also converges in $C^{1,\alpha}(B_{2-\rho}(0) \cap \{\gamma_1x_1>2\tilde{\delta}\})$ for all $0< \alpha <1$ by standard elliptic estimates. Now we reach contradiction since, using equation \eqref{LambdaContradiction}, $\rho_n\rightarrow \infty$ and $M_n \leq 0$, we have: \begin{align*} 0 &\geq \lim_n \rho_n^2\frac{ \int_{\partial B_{\rho_n}} \left( \langle B_n(x)\nabla_{\theta}v_{1,n}, \nabla_{\theta} v_{1,n} \rangle - M_na_n(x)v_{2,n}^2 v_{1,n}^2 + v_{1,n}f_{1,n}(x,v_{1,n}) \right) d\sigma(x) }{ \int_{\partial B_{\rho_n}} (1+\alpha \rho_n r_n) \mu_n(x) v_{1,n}^2 d\sigma(x) } \\ &\geq \lim_n\frac{ \int_{\partial B_1} \langle B_n(\rho_n x) \nabla_\theta \tilde{v}_{1,n}, \nabla_\theta \tilde{v}_{1,n} \rangle d\sigma(x) }{ \int_{\partial B_{1}}(1+\alpha \rho_nr_n)\mu_n(x) \tilde{v}_{1,n}^2 d\sigma(x) } - \lim_n \rho_n^2 \frac{ \int_{\partial B_{\rho_n}} v_{1,n}f_{1,n}(x,v_{1,n}) d\sigma(x) }{ \int_{\partial B_{\rho_n}}(1+\alpha \rho_nr_n)\mu_n(x)v_{1,n}^2 d\sigma(x) }\\ & \geq \frac{ \int_{\partial B_1\cap \{\gamma_1x_1>2\tilde{\delta}\}} |\nabla_\theta\big(\gamma_1 x_1\big)|^2 d\sigma(x) }{ \int_{\partial B_1}\big(\gamma_1x_1\big)^2 d\sigma(x) } -2d\rho_n^2r_n^2\geq C > 0, \end{align*} where in the second to last inequality we used the $C^{1,\alpha}$--convergence of $\tilde{v}_n$, the fact that $B_n(\rho_n x) \rightarrow I$ uniformly over compact sets, and the bound for $f_{i,n}$ given by Lemma \ref{lemmaMatrixBlowLimit}. \noindent \textbf{(ii)} In the case where $\rho_n$ is bounded, there exists $\overline{\rho}$ such that $\rho_n \rightarrow \overline{\rho}$. If $M_n \rightarrow -\infty$ then $v_n \rightarrow v$ where $v$ satisfies the system \eqref{unboundedEquation} of Proposition \ref{concludeProp}. Similarly to above we have that $$ \lim_n \Big| \rho_n^2 \frac{ \int_{\partial B_{\rho_n}} v_{1,n}f_{1,n}(x,v_{1,n}) d\sigma(x) }{ \int_{\partial B_{\rho_n}}(1+\alpha \rho_nr_n)\mu_n(x)v_{1,n}^2 d\sigma(x) } \Big| \leq 2d\rho_n^2r_n^2 \rightarrow 0 $$ and also, $$ \frac{ -\int_{\partial B_{\overline{\rho}}} M_na_n(x)v_{2,n}^2 v_{1,n}^2 d\sigma(x) }{ \int_{\partial B_{\overline{\rho}}} (1+\alpha \rho_n r_n)\mu_n(x) v_{1,n}^2 d\sigma(x) } \geq 0. $$ Thus we must have that: \begin{equation} \label{contradictComplication123} \lim_n \frac{ \int_{\partial B_{\overline{\rho}}} \langle B_n(x) \nabla_\theta v_{1,n}, \nabla_\theta v_{1,n} \rangle d\sigma(x) }{ \int_{\partial B_{\overline{\rho}}} (1+\alpha \rho_nr_n) \mu_n(x) v_{1,n}^2 d\sigma(x) } = 0 \end{equation} By an argument similar to the one above in \textbf{(i)} we can conclude $C^{1,\alpha}$ convergence in sets where $\{v_1>0\}$, thus: $$ \lim_n \frac{ \int_{\partial B_{\rho_n}} \langle B_n( x) \nabla_\theta v_{1,n}, \nabla_\theta v_{1,n} \rangle d\sigma(x) }{ \int_{\partial B_{\rho_n}} (1+\alpha \rho_nr_n) \mu_n(x) \tilde{v}_{1,n}^2 d\sigma(x) } \geq \frac{ \int_{\partial B_{\overline{\rho}}\cap \{v_1>0\}} |\nabla_\theta v_1|^2 d\sigma(x) } { \int_{\partial B_{\overline{\rho}}} v_1^2 d\sigma(x) }. $$ By Lemma \ref{nonTriviality} we must have that $\int_{\partial B_{\overline{\rho}}}v_1^2d\sigma(x) > 0$ and $\int_{\partial B_{\overline{\rho}}}v_2^2d\sigma(x)>0$ and also $v_1\cdot v_2 = 0$. This implies that the set $\{v_1>0\}$ is non-empty and that $|\nabla_\theta v_1|$ must be different from zero since otherwise $v_1$ would be constant different from zero in $B_{\overline{\rho}}$ and since $\int_{B_{\overline{\rho}}}v_2^2d\sigma(x) > 0$ and $v_1\cdot v_2 = 0$ this can't happen. Thus: $$ \frac{ \int_{\partial B_{\overline{\rho}}\cap \{v_1>0\}} |\nabla_\theta v_1|^2 d\sigma(x) } { \int_{\partial B_{\overline{\rho}}} v_1^2 d\sigma(x) }>0, $$ in contradiction with \eqref{contradictComplication123}. On the other hand, if $M_n$ is bounded, then $v_n \rightarrow v$ in $C^{1,\alpha}_{loc}(\mathbb{R}^N)$. By Lemma \ref{nonTriviality} we know that both $v_{1}$ and $v_2$ are nonnegative nontrivial, and by the strong maximum principle we have $v_1, v_2 > 0$ in $\mathbb{R}^N$. This implies that: $$ \frac{ -\int_{\partial B_{\overline{\rho}}} M_na_n(x)v_{2,n}^2 v_{1,n}^2 d\sigma(x) }{ \int_{\partial B_{\overline{\rho}}} (1+\alpha \rho_n r_n)\mu_n(x) v_{1,n}^2 d\sigma(x) } \geq C > 0 $$ and this allows us to reach contradiction. \end{proof} \begin{lemma} \label{almgrenLemmaHypoithesisStuff} There exists $C$ such that, for $r,s \in ]0,R_n[ = ]0,\frac{\overline{r}_n}{3}[$ such that $r \leq s$, then \begin{equation*} \frac{1}{r^{N-1}} \int_{\partial B_r} v_{i,n}^2 d\sigma(y) \leq \frac{C}{s^{N-1}} \int_{\partial B_s} v_{i,n}^2 d\sigma(y). \end{equation*} In particular, this proves $(h_5)$. \end{lemma} \begin{proof} Using Lemma \ref{MonotonicityForBlowup} there exists $\tilde{C}>0$ such that for each $i \in \{1,...,l\}$, the function: $$ r \mapsto \left( \frac{1}{r^{N-1}} \int_{\partial B_r} \mu_n(y) v_{i,n}^2 d\sigma(y) \right) e^{\tilde{C}r_nr} $$ is monotone nondecreasing for $r \in ]0,R_n[\subset ]0, \frac{\tilde{r}}{r_n}[$. Using the matrix bounds from Lemma \ref{lemmaMatrixBlowLimit} and $r_n\overline{r}_n\rightarrow 0$, we conclude $\frac{\theta}{M} \leq \mu_n(y) \leq Cr_nR_n\leq C'$. Thus, given $r,s \in ]0, R_n[$ and $r < s$, we conclude: \begin{align*} \frac{1}{r^{N-1}} \int_{\partial B_r} v_{i,n}^2 d\sigma(y) &\leq \frac{1}{\theta} \frac{1}{r^{N-1}} \int_{\partial B_r} \mu_n(y) v_{i,n}^2 d\sigma(y) \leq \frac{1}{\theta} \left( \frac{1}{s^{N-1}} \int_{\partial B_s} \mu_n(y) v_{i,n}^2 d\sigma(y) \right) e^{\tilde{C}r_n(s-r)}\\ &\leq \frac{M}{\theta} \left( \frac{1}{s^{N-1}} \int_{\partial B_s} v_{i,n}^2 d\sigma(y) \right) e^{\tilde{C}r_n\overline{r}_n/3} \leq C \left( \frac{1}{s^{N-1}} \int_{\partial B_s} v_{i,n}^2 d\sigma(y) \right) \end{align*} since $r_n\overline{r}_n \rightarrow 0$, taking $C = \sup_{n}\frac{M}{\theta}e^{\tilde{C}r_n\overline{r}_n}$. \end{proof} It remains to show that also $J_{1,n}(r)$ and $J_{2,n}(r)$ are positive in the whole range $[2N/\theta^\frac{1}{2},\frac{\overline{r}_n}{3}]$, which is condition ($h_6$). \begin{lemma} \label{LastLemma} We have that: $$ J_{i,n}(r) > 0 \qquad \forall r \in [2N/\theta^\frac{1}{2}, \frac{\overline{r}_n}{3}], $$ for all $n \in \mathbb{N}$ and $i=1,2$. In particular, this together with Lemma \ref{LambdasMoreThanZeroAltCafCond} implies that $(h_6)$ holds true. Also there exists $c>0$ such that $J_n(2N/\theta^\frac{1}{2}) = \frac { J_{1,n}(2N/\theta^\frac{1}{2}) J_{2,n}(2N/\theta^\frac{1}{2}) } {(2N/\theta^\frac{1}{2})^4} > c. $ \end{lemma} \begin{proof} First of all, there exists $\overline{C}>0$ such that $J_{i,n}(r)\geq \overline{C}$ for every $r \in [2N/\theta^\frac{1}{2}, 10N/\theta^\frac{1}{2}]$ and $i=1,2$. This is a consequence of $v_{i,n} \rightarrow v_i$ in $C(B_{10N/\theta^\frac{1}{2}})\cap H^1(B_{10N/\theta^\frac{1}{2}})$, and $v_i$ (in particular $v_1$) is nonconstant in $B_{10N/\theta^\frac{1}{2}}$, and that $f_{i,n}(x,v_{i,n}) \rightarrow 0$ uniformly in $B_{10N/\theta^\frac{1}{2}}$ by Proposition \ref{ListProp} and $M_n$. This also proves the last part of the statement of the lemma. Define: $$ s_n := \sup \Big\{ s \in ]2N/\theta^\frac{1}{2}, \overline{r}_n/3[: J_{i,n}(r) > 0 , \text{ for every } r \in ]2N/\theta^\frac{1}{2},s[ \Big\}. $$ We wish to prove that $s_n = \overline{r}_n/3$. Using Lemmas \ref{easyLemmaForConditions123}, \ref{sameWeightAltCafCond} and \ref{LambdasMoreThanZeroAltCafCond}, the definition of $s_n$, the definitions of the constants $\epsilon_n = dr_n^2$ and $c_n = Cr_n$, all conditions $(h_0)$-$(h_6)$ of Section \ref{chapter:resultsChap4} are satisfied in the interval $]2N/\theta^\frac{1}{2},s_n[$, so by Theorem \ref{AltCaffMonotonicity} there exists $0<\eta<1$ and $C>0$ such: $$ r \mapsto \frac {J_{1,n}(r)J_{2,n}(r)} {r^4} e^{ -C|M_n|^{-\eta} r^{-2\eta} + Cr_n^2r^2 + Cr_nr } $$ is monotone nondecreasing for $r \in ]2N/\theta^\frac{1}{2}, s_n[$. Thus this implies that for all $s \in ]0,s_n[$ we have: \begin{align*} J_{1,n}(r)J_{2,n}(r) &= r^4J_n(r)\geq r^4 J_n(2N/\theta^\frac{1}{2}) e^{ -C|M_n|^{-\eta} + Cr_n^2 + Cr_n + C|M_n|^{-\eta} r^{-2\eta} - C r_n^2 r^2 - Cr_nr }\\ &\geq \overline{C}^2 e^{-C|M_n|^{-\eta} - C r_n^2 r^2 - Cr_nr}, \end{align*} for all $r \in [2N/\theta^\frac{1}{2}, R_n]$ we have $r_nr \rightarrow 0$ and by Proposition \ref{concludeProp} there exists $\epsilon>0$ such that $|M_n|\geq \epsilon$, thus $|M_n|^{-\frac{\gamma}{2(\gamma+1)}} \leq \epsilon^{-\frac{\gamma}{2(\gamma+1)}}$. We conclude there exists $\tilde{c}>0$ such that $e^{-C|M_n|^{-\eta} - C r_n^2 r^2 - Cr_nr} \geq C, $ and so: \begin{equation} \label{finalEquationForChapter} J_{1,n}(r)J_{2,n}(r) \geq \overline{C}^2C > 0. \end{equation} By continuity of $J_{i,n}(r)$, using \eqref{finalEquationForChapter}, we have that $J_{i,n}(r)>0$ for all $r \in [0,s_n]$. This implies that there exists $\tilde{\epsilon}>0$ such that for $s_n$ there exists $J_{1,n}(s_n)>\tilde{\epsilon}$ and $J_{2,n}(s_n)>\tilde{\epsilon}$, and by the continuity of $J_{i,n}$ we conclude that for $\tilde{\delta}>0$ small enough we have $J_{1,n}(s_n+\tilde{\delta})>0$ and $J_{2,n}(s_n+\tilde{\delta})>0$ in contradiction with the definition of $s_n$ in case $s_n < R_n$. Thus we conclude that $s_n = R_n = \frac{r_n}{3}$ \end{proof} \begin{proof}[Conclusion of the proof of Lemma \ref{caffMonot}] With all the conditions $(h_0)$-$(h_6)$ satisfied by the sequence $\{v_n\}$ in the interval $[2N/\theta^\frac{1}{2},\overline{r}_n/3]$, we can apply Theorem \ref{AltCaffMonotonicity} to $\{v_n\}$, conclude the validity of Lemma \ref{caffMonot}. \end{proof}

https://arxiv.org/abs/1204.5521

An integrable multicomponent quad equation and its Lagrangian formulation

We present a hierarchy of discrete systems whose first members are the lattice modified Korteweg-de Vries equation, and the lattice modified Boussinesq equation. The N-th member in the hierarchy is an N-component system defined on an elementary plaquette in the 2-dimensional lattice. The system is multidimensionally consistent and a Lagrangian which respects this feature, i.e., which has the desirable closure property, is obtained.

\section{Introduction} What we now call the lattice modified Korteweg-de Vries (MKdV) equation appeared, up to a point-transformation, as early as 1894 in the work of Bianchi \cite{Bia1894} as the permutability condition for B\"acklund transformations of the sine-Gordon equation. It appeared again as an integrable partial difference equation via Hirota's method in \cite{Hir1977}, and it was derived via the direct linearization scheme in \cite{NijQuiCap1983}. To describe the lattice MKdV in the modern setting, let $v=v(l,m)$ be the dependent variable depending on position $(l,m)$ in a 2-dimensional lattice, where $l$ and $m$ shift by integer values, and let $p,q$ be parameters associated with the $l,m$-lattice directions respectively. Then the lattice MKdV equation reads \begin{equation}\label{MKdV} p(v\hat{v}-\tilde{v}\hat{\tilde{v}}) = q(v\tilde{v}-\hat{v}\hat{\tilde{v}}), \end{equation} where ``$\;\tilde{}\;$'' denotes a shift in the $l$-direction, and ``$\;\hat{}\;$'' denotes a shift in the $m$-direction. A generalized direct linearization scheme proposed in \cite{NijPapCapQui1992} led to a higher order analogue of the lattice MKdV equation, which can be identified as a lattice version of the modified Boussinesq (MBSQ) equation of \cite{QuiNijCap1982}. This is a scalar equation on a stencil of 9 lattice points, given by \begin{equation}\label{scalarMBSQ} \biggl(\frac{p^2\hat{\tilde{v}}-q^2\hat{\hat{v}}}{p\hat{\hat{v}}-q\hat{\tilde{v}}}\biggr)\frac{\hat{\hat{\tilde{v}}}}{\hat{v}}-\biggl(\frac{p^2\tilde{\tilde{v}}-q^2\hat{\tilde{v}}}{p\hat{\tilde{v}}-q\tilde{\tilde{v}}}\biggr)\frac{\hat{\tilde{\tilde{v}}}}{\tilde{v}} = p\biggl(\frac{v}{\tilde{v}}-\frac{\hat{\hat{\tilde{v}}}}{\hat{\hat{\tilde{\tilde{v}}}}}\biggr)-q\biggl(\frac{v}{\hat{v}}-\frac{\hat{\tilde{\tilde{v}}}}{\hat{\hat{\tilde{\tilde{v}}}}}\biggr). \end{equation} A more convenient form of this equation appeared in \cite{Nij1999}, specifically a coupled system of equations in two dependent variables $v,w$ defined on an elementary plaquette: \begin{equation}\label{MBSQ} \frac{p\hat{v}-q\tilde{v}}{\hat{\tilde{v}}} = \frac{p\tilde{v}\hat{w}-q\hat{v}\tilde{w}}{v\hat{\tilde{w}}} = \frac{p\tilde{w}-q\hat{w}}{w}. \end{equation} The scalar equation \eqref{scalarMBSQ} can be derived from \eqref{MBSQ} by eliminating one or other of the variables. As shown in \cite{Nij1999}, in either form this is a discrete analogue of the potential modified Boussinesq equation \cite{QuiNijCap1982}. Symmetries and conservation laws for the two-component form of the equation were derived in \cite{XenNij2012}. The lattice Gel'fand-Dikii (GD) hierarchy, that is the system given in \cite{NijPapCapQui1992} which goes to the continuous GD hierarchy under suitable limits\footnote{In the 1970s members of the Russian school, including Gel'fand and Dikii, studied higher order spectral problems generalizing aspects of the Korteweg-de Vries equation (c.f. e.g. \cite{GelDik1976ii,Man1979}); hence these types of systems have been referred to as Gel'fand-Dikii hierarchies.}, is a multicomponent system whose first members are the lattice potential KdV equation and a lattice version of the Boussinesq equation. A lattice `modified' GD hierarchy was also proposed in the Appendix of \cite{NijPapCapQui1992}, the first two members of which, (\ref{MKdV}) and (\ref{scalarMBSQ}), were given explicitly, but its higher members were left implicit due to the difficulty of writing corresponding scalar equations in explicit form on increasing stencil size. In this paper we study a hierarchy of systems whose first two members are \eqref{MKdV} and \eqref{MBSQ}. Different to the direct linearization approach of \cite{NijPapCapQui1992} it will be obtained by a reduction procedure from the discrete Kadomtsev-Petviashvili (KP) equation of Hirota \cite{Hir1981}, and in fact the precise relationship between the system studied here and those of \cite{NijPapCapQui1992} is not yet known beyond the first two cases. Like the first cases (\ref{MKdV}) and (\ref{MBSQ}), each system is defined on an elementary plaquette of the 2-dimensional lattice (i.e., it is a quad equation); in particular each is consistent in multidimensions and possesses a Lax representation. Our main objectives are to present the derivation of this hierarchy as well as a Lagrangian formulation that respects its consistency property. The new kind of variational principle for multidimensionally consistent systems proposed in \cite{LobNij2009} relies on a closure property of the associated Lagrangian, allowing it to be interpreted as a closed form. A Lagrangian with this property was given in \cite{LobNij2010} for the generic member of the lattice GD hierarchy. We give here a Lagrangian for the generic member of the hierarchy in this paper, for which there is a clear connection to that of the discrete bilinear KP equation, whose Lagrangian was given in \cite{LobNijQui2009}. In Section \ref{Section_system} we derive the $N$-component lattice system, give the Lax pair, describe its multidimensional consistency and give some important symmetries. In Section \ref{Section_Lagrangian} we present the Lagrangian structure, and in Section \ref{Section_closure} we demonstrate that this Lagrangian obeys a closure relation on solutions of the system. Section \ref{Section_conclusion} contains some additional remarks. \section{The system}\label{Section_system} It was observed in \cite{Atk2008} that the one-variable and two-variable equations (\ref{MKdV}) and (\ref{MBSQ}) can be obtained by a natural reduction procedure from the discrete KP equation of Hirota \cite{Hir1981}. Generalization of this reduction procedure then led to a quadrilateral lattice equation in $N$ variables $v_1,\ldots,v_N$ that can be written as \begin{equation}\label{discreteMGD} \frac{p\hat{v}_{n-1}\tilde{v}_{n}-q\tilde{v}_{n-1}\hat{v}_{n}}{\hat{\tilde{v}}_{n-1}v_{n}} = \frac{p\hat{v}_{n}\tilde{v}_{n+1}-q\tilde{v}_{n}\hat{v}_{n+1}}{\hat{\tilde{v}}_{n}v_{n+1}}, \;\;\;\;\;\; n\in\{1,\dots,N\}, \end{equation} where by convention\footnote{This convention is a convenient notational device which avoids the need to separate equations such as (\ref{discreteMGD}) into sub cases $n=1$, $n\in\{2,\ldots,N-1\}$ and $n=N$.} $v_0=v_{N+1}=1$. This system, which contains (\ref{MKdV}) and (\ref{MBSQ}) as cases $N=1$ and $N=2$, is the main object studied in the present paper. \subsection{Reduction from Hirota's discrete KP} We begin by giving the reduction procedure from which (\ref{discreteMGD}) may be obtained. The starting point is the Hirota discrete KP equation \cite{Hir1981} \begin{equation}\label{notHirotaMiwa} \bar{\tau}\hat{\tilde{\tau}} = p\hat{\tau}\tilde{\bar{\tau}}-q\tilde{\tau}\hat{\bar{\tau}}, \end{equation} where $\tau=\tau(l,m,n)$, $\tilde{\tau}=\tau(l+1,m,n)$, $\hat{\tau}=\tau(l,m+1,n)$, $\bar{\tau}=\tau(l,m,n+1)$ etc. are values of the dependent variable $\tau$ as a function of independent variables $l,m,n\in\mathbb{Z}$. In this equation $p,q\in\mathbb{C}\setminus\{0\}$ are parameters, they are often taken equal to 1, which up to a non-autonomous change of variables is without loss of generality. Such change of variables will however not be possible after the reduction procedure. From the consistency property of (\ref{notHirotaMiwa}) there can be obtained a B\"acklund transformation as follows: \begin{subequations} \label{hbt} \begin{eqnarray} &&\bar{\tau}\tilde{\sigma} = p\tilde{\bar{\tau}}\sigma - \lambda\tilde{\tau}\bar{\sigma},\\ &&\bar{\tau}\hat{\sigma} = q\hat{\bar{\tau}}\sigma - \lambda\hat{\tau}\bar{\sigma}. \end{eqnarray} \end{subequations} Here, if $\tau=\tau(l,m,n)$ is a solution of (\ref{notHirotaMiwa}) then the coupled system (\ref{hbt}) for $\sigma$ is consistent and determines a function $\sigma=\sigma(l,m,n)$ also satisfying (\ref{notHirotaMiwa}). This is a weak B\"acklund transformation\footnote{This terminology was probably first used by McCarthy \cite{McC1978} to distinguish the situation where an equation is strictly sufficient for compatibility of the B\"acklund system, as opposed to a strong B\"acklund transformation, where an equation emerges as both necessary and sufficient.}, in fact the equation emerging as compatibility constraint for (\ref{hbt}) is \begin{equation}\label{10pteqn} \frac{p\hat{\tau}\tilde{\bar{\tau}}-q\tilde{\tau}\hat{\bar{\tau}}}{\bar{\tau}\hat{\tilde{\tau}}} = \biggl[\frac{p\hat{\tau}\tilde{\bar{\tau}}-q\tilde{\tau}\hat{\bar{\tau}}}{\bar{\tau}\hat{\tilde{\tau}}}\biggr]^{-}, \end{equation} and it can be directly verified that (\ref{hbt}) is a B\"acklund transformation, in the strong sense, for this equation. The equation \eqref{10pteqn} lies on a 10-point stencil on the three-dimensional lattice, by integration it can be seen as a non-autonomous version of (\ref{notHirotaMiwa}). The equation (\ref{discreteMGD}) governs solutions of (\ref{10pteqn}) satisfying the following additional constraints \begin{subequations} \label{constraints} \begin{eqnarray} \label{constraint1} &&\tau(l,m,n+N+1)=\tau(l,m,n), \quad l,m,n\in\mathbb{Z},\\ \label{constraint2} &&\tau(l,m,0)=1, \quad l,m\in\mathbb{Z}, \end{eqnarray} \end{subequations} which can be seen by identifying \begin{equation} v_{n}(l,m):=\tau(l,m,n), \quad l,m\in\mathbb{Z}, \quad n \in\{1,\ldots, N\}. \end{equation} The first constraint (\ref{constraint1}) is simply an imposed periodicity, the second constraint (\ref{constraint2}) is not so usual, it consists of some particular boundary data. \subsection{Integrability} The original B\"acklund transformation (\ref{hbt}) of equation (\ref{10pteqn}) does not preserve the constraints (\ref{constraints}) but it is consistent with the first constraint (\ref{constraint1}). That is, we may write \begin{equation} \sigma(l,m,n+N+1)=\sigma(l,m,n), \quad l,m,n\in\mathbb{Z}\label{constraint3} \end{equation} without breaking the consistency. The further observation that this B\"acklund transformation is already linear in $\sigma$ makes it, through this reduction procedure, a Lax pair for (\ref{discreteMGD}). In other words defining \begin{equation} \phi_{n}(l,m):=\sigma(l,m,n), \quad l,m\in\mathbb{Z}, \quad n \in\{0,\ldots, N\}, \end{equation} and imposing (\ref{constraints}) and (\ref{constraint3}) reduces (\ref{hbt}) to the system \begin{equation} \label{Lax} \tilde{\Phi} = L_{p}(\boldsymbol{v},\tilde{\boldsymbol{v}})\Phi,\quad \hat{\Phi} = L_{q}(\boldsymbol{v},\hat{\boldsymbol{v}})\Phi, \end{equation} where \begin{equation}\Phi := [\phi_0,\phi_1,\ldots,\phi_N]^T, \quad \boldsymbol{v}:=(v_1,\ldots,v_N),\label{notation}\end{equation} and \begin{equation} L_{p}(\boldsymbol{v},\tilde{\boldsymbol{v}}) = \left(\begin{array}{cccccc} p\tilde{v}_{1}/v_{1} & -\lambda/v_{1} & 0 & \cdots & \;\;\;\;\;\;\cdots & 0\\ 0 & p\tilde{v}_{2}/v_{2} & -\lambda\tilde{v}_{1}/v_{2} & 0 & \;\;\;\;\;\;\cdots & 0\\ & & & & & \vdots\\ \vdots & & \;\;\;\;\;\;\ddots & \;\;\;\;\;\;\ddots & \;\;\;\;\;\;\ddots & \vdots\\ & & & & & 0\\ 0 & \cdots & \cdots & 0 & p\tilde{v}_{N}/v_{N} & -\lambda\tilde{v}_{N-1}/v_{N}\\ -\lambda\tilde{v}_{N} & 0 & \cdots & & 0 & p \end{array}\right) \end{equation} is the $(N+1)$-square Lax matrix. The consistency condition of (\ref{Lax}) \begin{equation} L_{p}(\hat{\boldsymbol{v}},\hat{\tilde{\boldsymbol{v}}})L_{q}(\boldsymbol{v},\hat{\boldsymbol{v}})-L_{q}(\tilde{\boldsymbol{v}},\hat{\tilde{\boldsymbol{v}}})L_{p}(\boldsymbol{v},\tilde{\boldsymbol{v}})={\boldsymbol 0} \end{equation} gives the system \eqref{discreteMGD}. The multidimensional consistency of (\ref{discreteMGD}), like the Lax pair, is basically inherited from (\ref{notHirotaMiwa}). For this property, the vector $\boldsymbol{v}=\boldsymbol{v}(l,m)$ on $\mathbb{Z}^2$ is considered instead as a function on the $d$-dimensional regular lattice $\boldsymbol{v}=\boldsymbol{v}(l_1,l_2,\ldots,l_d)$ satisfying system (\ref{discreteMGD}) in all pairs of directions, where parameters $p,q$ appearing in (\ref{discreteMGD}) associated with the original $l$ and $m$ directions are to be replaced with pairs of parameters taken from the larger set $p_1,\ldots,p_d$ associated with the corresponding $l_1,\ldots,l_d$ directions. The consistency of this system can be verified directly by calculation. \subsection{Point symmetries}\label{Section_Symmetry} The system (\ref{discreteMGD}) is invariant under the scaling symmetry \begin{equation} (v_1,\ldots,v_N)\rightarrow(\mu_1 v_1,\ldots, \mu_{N} v_{N}), \qquad \mu_1,\ldots,\mu_N \in\mathbb{C}\setminus\{0\},\label{sym1} \end{equation} as well as the less obvious transformation \begin{equation} (v_1,v_2,\ldots,v_{N-1},v_{N})\rightarrow (v_2/v_1,v_3/v_1,\ldots,v_N/v_{1},1/v_1).\label{sym2} \end{equation} The latter transformation here generates a group of $N$ similar symmetries of (\ref{discreteMGD}). Such autonomous point symmetries allow an equation to take on alternative guises via a non-autonomous point transformation that preserves the autonomous nature of the equation, and also preserves its multidimensional consistency property (see for instance \cite{BobSur2002,Atk2009}). Explicitly, if $S_1$ and $S_2$ are (commuting) transformations of the vector $\boldsymbol{v}=(v_1,\ldots,v_N)$ taken from the group generated by (\ref{sym1}) and (\ref{sym2}), and $\boldsymbol{v}$ is governed by (\ref{discreteMGD}), then the equation governing $\boldsymbol{v}'=[S_1^l\cdot S_2^m](\boldsymbol{v})$ is both autonomous and multidimensionally consistent. The autonomous nature of the equation is preserved due to the fact the $S_i$ are symmetries, whilst the natural extension to multidimensions of the transformation: $[S_1^l\cdot S_2^m] \rightarrow [S_1^{l_1}\cdot S_2^{l_2} \cdot S_3^{l_3}\cdot \ldots ]$, demonstrates also that the consistency is preserved. This is an important part of the transformation theory of multi-component systems such as (\ref{discreteMGD}) because invariants of such transformations are not presently known beyond the multi-affine class \cite{AdlBobSur2009a}. The less obvious symmetry (\ref{sym2}) is included here because to distinguish between genuinely different quad equations in the multi-component class, it is necessary to be aware of all such transformations. \section{Lagrangian structure}\label{Section_Lagrangian} For discrete systems where each lattice direction is on an equal footing it appears that the Lagrangian is a more natural object than the Hamiltonian, given that it does not single out any one direction. Therefore an important feature of the multicomponent quad equation \eqref{discreteMGD} is its Lagrangian structure, which as we will see is inherited from that of Hirota's discrete KP equation \eqref{notHirotaMiwa}. The more remarkable feature is its natural interplay with the multidimensional consistency of the model, which will be explained in the following section (Section \ref{Section_closure}). A Lagrangian for the system \eqref{discreteMGD} is \begin{eqnarray}\label{DiscreteMGDLagrangian} \mathcal{L}_{pq} & = & \sum_{i=0}^{N}\biggl\{{\rm Li}_{2}\biggl(\frac{q\tilde{v}_{i}\hat{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}}\biggr) + \frac{1}{4}\biggl(\ln\biggl(-\frac{q\tilde{v}_{i}\hat{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}}\biggr)\biggr)^2 + \frac{1}{2}\bigl(\ln{\hat{v}_{i}}\ln{\tilde{v}_{i+1}}-\ln{\tilde{v}_{i}}\ln{\hat{v}_{i+1}}\bigr)\nonumber\\ && \;\;\;\;\;\;\;\;\; + \ln\biggl(\frac{v_{i+1}}{v_{i}}\biggr)\ln\biggl(\frac{\tilde{v}_{i}}{\hat{v}_{i}}\biggr)\biggr\}, \end{eqnarray} which through the discrete Euler-Lagrange equations gives 2 copies of the equations. The function ${\rm Li}_{2}$ is the dilogarithm function \begin{equation}\label{Li2} {\rm Li}_2(z) = -\int^z_0{\frac{\ln(1-z)}{z}dz}, \end{equation} which appears in many areas of physics, such as quantum electrodynamics (e.g. vacuum polarization) and electrical network problems; it also comes up in connection with algebraic K-theory, representation theory of infinite dimensional algebras, and combinatorics \cite{Kir1995}. By discrete Euler-Lagrange equations we mean in this case \begin{equation} \frac{\delta\mathcal{L}_{pq}}{\delta v_{i}}\equiv \frac{\partial\mathcal{L}_{pq}}{\partial v_{i}}+T_{p}^{-1}\frac{\partial\mathcal{L}_{pq}}{\partial \tilde{v}_{i}}+T_{q}^{-1}\frac{\partial\mathcal{L}_{pq}}{\partial \hat{v}_{i}} = 0, \;\;\;\; i\in\{1,\dots,N\}, \end{equation} where $T_{p}$ is a shift operator in the lattice direction associated with parameter $p$, and similarly $T_{q}$ is a shift operator in the lattice direction associated with parameter $q$. Then $T_{p}^{-1},T_{q}^{-1}$ are simply backwards shifts in the lattice. The discrete Euler-Lagrange equations arise from extremizing the action \begin{eqnarray}\label{DiscreteMGDaction} S & = & \sum_{l}\sum_{m}\sum_{i=0}^{N}\biggl\{{\rm Li}_{2}\biggl(\frac{q\tilde{v}_{i}\hat{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}}\biggr) + \frac{1}{4}\biggl(\ln\biggl(-\frac{q\tilde{v}_{i}\hat{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}}\biggr)\biggr)^2 + \frac{1}{2}\bigl(\ln{\hat{v}_{i}}\ln{\tilde{v}_{i+1}}-\ln{\tilde{v}_{i}}\ln{\hat{v}_{i+1}}\bigr)\nonumber\\ && \;\;\;\;\;\;\;\;\; + \ln\biggl(\frac{v_{i+1}}{v_{i}}\biggr)\ln\biggl(\frac{\tilde{v}_{i}}{\hat{v}_{i}}\biggr)\biggr\}, \end{eqnarray} where the first two sums over $l$ and $m$ are understood to be taken over the whole 2-dimensional lattice. In fact, rather than the Lagrangian by itself, it is more instructive to look at the action, since it is on this level that we can easily see we are dealing with a reduction of the bilinear discrete KP equation \begin{equation}\label{HM} A_{pq}\bar{\tau}\hat{\tilde{\tau}}+A_{qr}\tilde{\tau}\hat{\bar{\tau}}+A_{rp}\hat{\tau}\tilde{\bar{\tau}}=0; \end{equation} here the dependent variable is $\tau$, and the $A_{pq}$ are antisymmetric constants (so that $A_{qp}=-A_{pq}$) associated with the $p,q$ lattice directions. Making the identification $A_{pr}/A_{pq}=p$, $A_{qr}/A_{pq}=q$ and $\tau=u$, so that the equation is no longer covariant due to the singled-out lattice direction, brings us to equation \eqref{notHirotaMiwa}. A Lagrangian for \eqref{HM} was presented in \cite{LobNijQui2009}, which, adapted for the non-covariant equation \eqref{notHirotaMiwa}, is equivalent to \begin{eqnarray} S_{KP} & = & \sum_{l}\sum_{m}\sum_{n}\biggl\{{\rm Li}_{2}\biggl(\frac{q\tilde{\tau}\hat{\bar{\tau}}}{p\hat{\tau}\tilde{\bar{\tau}}}\biggr) + \frac{1}{4}\biggl(\ln\biggl(-\frac{q\tilde{\tau}\hat{\bar{\tau}}}{p\hat{\tau}\tilde{\bar{\tau}}}\biggr)\biggr)^2 + \frac{1}{2}\bigl(\ln{\hat{\tau}}\ln{\tilde{\bar{\tau}}}-\ln{\tilde{\tau}}\ln{\hat{\bar{\tau}}}\bigr)\nonumber\\ && \;\;\;\;\;\;\;\;\; + \ln\biggl(\frac{\bar{\tau}}{\tau}\biggr)\ln\biggl(\frac{\tilde{\tau}}{\hat{\tau}}\biggr)\biggr\}. \end{eqnarray} It is clear that identifying $\tau=v_{i}$, $\bar{\tau}=v_{i+1}$, and restricting the sum over $n$, we get the action \eqref{DiscreteMGDaction} of the reduction. \section{Closure property}\label{Section_closure} The Lagrangians for many integrable systems can be interpreted as closed forms; they obey a closure relation on solutions when consistently embedded in a higher-dimensional space. In a sense one can consider this to be multidimensional consistency on the level of the Lagrangian. It allows for a variational principle which includes the geometry of the space of independent variables, as well as that of the dependent variables; seeking extrema of the action under variations with respect to all of the variables, dependent and independent, leads to constraints which specify not only the equations of motion, but also to an extent the Lagrangians themselves. In the case of 2-dimensional systems, we have Lagrangian 2-forms, for instance Lagrangian functions which are evaluated on 2-dimensional plaquettes. For multidimensionally consistent systems we can suppose the action to be defined on a surface in an arbitrary number of dimensions, not just two, and this action will be a sum of the Lagrangian function evaluated on every plaquette in that surface. Demanding zero variation with respect to the surface implies a closure relation on the Lagrangian, while variation with respect to the dependent variable gives equations of motion. Examples can be found in both discrete and continuous cases of 1-dimensional systems \cite{YooLobNij2011,YooNij2011}, a large class of 2-dimensional systems \cite{LobNij2009,LobNij2010,XenNijLob2010}, and also the 3-dimensional discrete bilinear KP equation \cite{LobNijQui2009}. In \cite{LobNij2010} it was demonstrated that the Lagrangian for an arbitrary member of the lattice GD hierarchy obeys the closure relation \begin{equation}\label{closure} \Delta_{p}\mathcal{L}_{qr}+\Delta_{q}\mathcal{L}_{rp}+\Delta_{r}\mathcal{L}_{pq} = 0, \end{equation} on solutions to the lattice KP equation (the lattice KP equation is a consequence of the equations which constitute the lattice GD hierarchy itself). Here the operator $\Delta_{p}$ is a difference operator in the $p$-direction of the lattice; if we again employ $T_{p}$ to denote the shift operator in the $p$-direction, and write the identity operator as $id$, then $\Delta_{p}=T_{p}-id$. To demonstrate the Lagrangian \eqref{DiscreteMGDLagrangian} satisfies the closure relation \eqref{closure}, we need to invoke a 5-term identity for the dilogarithm function \begin{eqnarray}\label{5pt} {\rm Li}_2(s) + {\rm Li}_2(t) - {\rm Li}_2(st)& = & {\rm Li}_2\biggl(\frac{s-st}{1-st}\biggr)+{\rm Li}_2\biggl(\frac{t-st}{1-st}\biggr)\nonumber\\ && +\ln\biggl(\frac{1-s}{1-st}\biggr)\ln\biggl(\frac{1-t}{1-st}\biggr), \end{eqnarray} which holds for $s,t,st\neq 1$, up to imaginary constant terms, and also a 2-term identity which flips the argument \begin{equation}\label{flip} {\rm Li}_2(s)+{\rm Li}_2\biggl(\frac{1}{s}\biggr) = -\frac{1}{2}\bigl(\ln(-s)\bigr)^2-\frac{\pi^2}{6}, \end{equation} for $s\neq 0$. Further identities and other information about the dilogarithm function can be found in \cite{Lew1981}. Observe that \setlength{\unitlength}{1cm} \begin{picture}(10,9)(0,0) \put(0,9){\parbox[t]{6cm}{ \begin{eqnarray} \Gamma & \equiv & \bar{\mathcal{L}}_{pq}+\tilde{\mathcal{L}}_{qr}+\hat{\mathcal{L}}_{rp}-\mathcal{L}_{pq}-\mathcal{L}_{qr}-\mathcal{L}_{rp}\nonumber\\ & = & \sum_{i=0}^{N}\biggl\{\;{\rm Li}_{2}\biggl(\frac{q\tilde{\bar{v}}_{i}\hat{\bar{v}}_{i+1}}{p\hat{\bar{v}}_{i}\tilde{\bar{v}}_{i+1}}\biggr)\;\; +{\rm Li}_{2}\biggl(\frac{r\hat{\tilde{v}}_{i}\tilde{\bar{v}}_{i+1}}{q\tilde{\bar{v}}_{i}\hat{\tilde{v}}_{i+1}}\biggr) +{\rm Li}_{2}\biggl(\frac{p\hat{\bar{v}}_{i}\hat{\tilde{v}}_{i+1}}{r\hat{\tilde{v}}_{i}\hat{\bar{v}}_{i+1}}\biggr)\nonumber\\ && \;\;\;\;\;\;\; -{\rm Li}_{2}\biggl(\frac{q\tilde{v}_{i}\hat{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}}\biggr)\;\; -{\rm Li}_{2}\biggl(\frac{r\hat{v}_{i}\bar{v}_{i+1}}{q\bar{v}_{i}\hat{v}_{i+1}}\biggr) -{\rm Li}_{2}\biggl(\frac{p\bar{v}_{i}\tilde{v}_{i+1}}{r\tilde{v}_{i}\bar{v}_{i+1}}\biggr)\nonumber\\ && \;\;\;\;\;\;\; + \frac{1}{4}\biggl(\ln\biggl(-\frac{q\tilde{\bar{v}}_{i}\hat{\bar{v}}_{i+1}}{p\hat{\bar{v}}_{i}\tilde{\bar{v}}_{i+1}}\biggr)\biggr)^2 + \frac{1}{4}\biggl(\ln\biggl(-\frac{r\hat{\tilde{v}}_{i}\tilde{\bar{v}}_{i+1}}{q\tilde{\bar{v}}_{i}\hat{\tilde{v}}_{i+1}}\biggr)\biggr)^2 + \frac{1}{4}\biggl(\ln\biggl(-\frac{p\hat{\bar{v}}_{i}\hat{\tilde{v}}_{i+1}}{r\hat{\tilde{v}}_{i}\hat{\bar{v}}_{i+1}}\biggr)\biggr)^2\nonumber\\ && \;\;\;\;\;\;\; - \frac{1}{4}\biggl(\ln\biggl(-\frac{q\tilde{v}_{i}\hat{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}}\biggr)\biggr)^2 - \frac{1}{4}\biggl(\ln\biggl(-\frac{r\hat{v}_{i}\bar{v}_{i+1}}{q\bar{v}_{i}\hat{v}_{i+1}}\biggr)\biggr)^2 - \frac{1}{4}\biggl(\ln\biggl(-\frac{p\bar{v}_{i}\tilde{v}_{i+1}}{r\tilde{v}_{i}\bar{v}_{i+1}}\biggr)\biggr)^2\nonumber\\ && \;\;\;\;\;\;\; + \frac{1}{2}\bigl(\ln{\hat{\bar{v}}_{i}}\ln{\tilde{\bar{v}}_{i+1}}-\ln{\tilde{\bar{v}}_{i}}\ln{\hat{\bar{v}}_{i+1}}\bigr) + \frac{1}{2}\bigl(\ln{\tilde{\bar{v}}_{i}}\ln{\hat{\tilde{v}}_{i+1}}-\ln{\hat{\tilde{v}}_{i}}\ln{\tilde{\bar{v}}_{i+1}}\bigr)\nonumber\\ && \;\;\;\;\;\;\; + \frac{1}{2}\bigl(\ln{\hat{\tilde{v}}_{i}}\ln{\hat{\bar{v}}_{i+1}}-\ln{\hat{\bar{v}}_{i}}\ln{\hat{\tilde{v}}_{i+1}}\bigr) - \frac{1}{2}\bigl(\ln{\hat{v}_{i}}\ln{\tilde{v}_{i+1}}-\ln{\tilde{v}_{i}}\ln{\hat{v}_{i+1}}\bigr)\nonumber\\ && \;\;\;\;\;\;\; - \frac{1}{2}\bigl(\ln{\bar{v}_{i}}\ln{\hat{v}_{i+1}}-\ln{\hat{v}_{i}}\ln{\bar{v}_{i+1}}\bigr) - \frac{1}{2}\bigl(\ln{\tilde{v}_{i}}\ln{\bar{v}_{i+1}}-\ln{\bar{v}_{i}}\ln{\tilde{v}_{i+1}}\bigr)\nonumber\\ && \;\;\;\;\;\;\; + \ln\biggl(\frac{\bar{v}_{i+1}}{\bar{v}_{i}}\biggr)\ln\biggl(\frac{\tilde{\bar{v}}_{i}}{\hat{\bar{v}}_{i}}\biggr) + \ln\biggl(\frac{\tilde{v}_{i+1}}{\tilde{v}_{i}}\biggr)\ln\biggl(\frac{\hat{\tilde{v}}_{i}}{\tilde{\bar{v}}_{i}}\biggr) + \ln\biggl(\frac{\hat{v}_{i+1}}{\hat{v}_{i}}\biggr)\ln\biggl(\frac{\hat{\bar{v}}_{i}}{\hat{\tilde{v}}_{i}}\biggr)\nonumber\\ && \;\;\;\;\;\;\; - \ln\biggl(\frac{v_{i+1}}{v_{i}}\biggr)\ln\biggl(\frac{\tilde{v}_{i}}{\hat{v}_{i}}\biggr) - \ln\biggl(\frac{v_{i+1}}{v_{i}}\biggr)\ln\biggl(\frac{\hat{v}_{i}}{\bar{v}_{i}}\biggr) - \ln\biggl(\frac{v_{i+1}}{v_{i}}\biggr)\ln\biggl(\frac{\bar{v}_{i}}{\tilde{v}_{i}}\biggr)\biggr\}. \end{eqnarray} }} \put(1.9,7.2){\dashbox{0.1}(2.1,0.9)}\put(4.1,7.2){\framebox(5,0.9)} \put(1.7,6.2){\dashbox{0.1}(2.4,0.8)}\put(4.2,6.2){\framebox(5,0.8)} \end{picture} Using the 2-term dilogarithm identity \eqref{flip} on the terms in the dashed-line boxes, and the 5-point identity \eqref{5pt} on the terms in the solid-line boxes, this becomes: \begin{eqnarray} \Gamma & = & \sum_{i=0}^{N}\biggl\{{\rm Li}_{2}\biggl(-\frac{p\hat{\bar{v}}_{i}}{r\hat{\tilde{v}}_{i}}\cdot \frac{q\tilde{\bar{v}}_{i}\hat{\tilde{v}}_{i+1}-r\hat{\tilde{v}}_{i}\tilde{\bar{v}}_{i+1}}{p\hat{\bar{v}}_{i}\tilde{\bar{v}}_{i+1}-q\tilde{\bar{v}}_{i}\hat{\bar{v}}_{i+1}}\biggr) +{\rm Li}_{2}\biggl(-\frac{\tilde{\bar{v}}_{i+1}}{\hat{\tilde{v}}_{i+1}}\cdot \frac{r\hat{\tilde{v}}_{i}\hat{\bar{v}}_{i+1}-p\hat{\bar{v}}_{i}\hat{\tilde{v}}_{i+1}}{p\hat{\bar{v}}_{i}\tilde{\bar{v}}_{i+1}-q\tilde{\bar{v}}_{i}\hat{\bar{v}}_{i+1}}\biggr)\nonumber\\ && \;\;\;\;\;\;\; -{\rm Li}_{2}\biggl(-\frac{p\tilde{v}_{i+1}}{r\bar{v}_{i+1}}\cdot \frac{q\bar{v}_{i}\hat{v}_{i+1}-r\hat{v}_{i}\bar{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}-q\tilde{v}_{i}\hat{v}_{i+1}}\biggr) -{\rm Li}_{2}\biggl(-\frac{\hat{v}_{i}}{\bar{v}_{i}}\cdot \frac{r\tilde{v}_{i}\bar{v}_{i+1}-p\bar{v}_{i}\tilde{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}-q\tilde{v}_{i}\hat{v}_{i+1}}\biggr)\nonumber\\ && \;\;\;\;\;\;\; + \ln\biggl(-\frac{\hat{\bar{v}}_{i+1}}{\hat{\tilde{v}}_{i+1}}\cdot \frac{q\tilde{\bar{v}}_{i}\hat{\tilde{v}}_{i+1}-r\hat{\tilde{v}}_{i}\tilde{\bar{v}}_{i+1}}{p\hat{\bar{v}}_{i}\tilde{\bar{v}}_{i+1}-q\tilde{\bar{v}}_{i}\hat{\bar{v}}_{i+1}}\biggr) \ln\biggl(-\frac{q\tilde{\bar{v}}_{i}}{r\hat{\tilde{v}}_{i}}\cdot \frac{r\hat{\tilde{v}}_{i}\hat{\bar{v}}_{i+1}-p\hat{\bar{v}}_{i}\hat{\tilde{v}}_{i+1}}{p\hat{\bar{v}}_{i}\tilde{\bar{v}}_{i+1}-q\tilde{\bar{v}}_{i}\hat{\bar{v}}_{i+1}}\biggr)\nonumber\\ && \;\;\;\;\;\;\; - \ln\biggl(-\frac{\tilde{v}_{i}}{\bar{v}_{i}}\cdot \frac{q\bar{v}_{i}\hat{v}_{i+1}-r\hat{v}_{i}\bar{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}-q\tilde{v}_{i}\hat{v}_{i+1}}\biggr) \ln\biggl(-\frac{q\hat{v}_{i+1}}{r\bar{v}_{i+1}}\cdot \frac{r\tilde{v}_{i}\bar{v}_{i+1}-p\bar{v}_{i}\tilde{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}-q\tilde{v}_{i}\hat{v}_{i+1}}\biggr)\nonumber\\ && \;\;\;\;\;\;\; - \frac{1}{4}\biggl(\ln\biggl(-\frac{q\tilde{\bar{v}}_{i}\hat{\bar{v}}_{i+1}}{p\hat{\bar{v}}_{i}\tilde{\bar{v}}_{i+1}}\biggr)\biggr)^2 + \frac{1}{4}\biggl(\ln\biggl(-\frac{r\hat{\tilde{v}}_{i}\tilde{\bar{v}}_{i+1}}{q\tilde{\bar{v}}_{i}\hat{\tilde{v}}_{i+1}}\biggr)\biggr)^2 + \frac{1}{4}\biggl(\ln\biggl(-\frac{p\hat{\bar{v}}_{i}\hat{\tilde{v}}_{i+1}}{r\hat{\tilde{v}}_{i}\hat{\bar{v}}_{i+1}}\biggr)\biggr)^2\nonumber\\ && \;\;\;\;\;\;\; + \frac{1}{4}\biggl(\ln\biggl(-\frac{q\tilde{v}_{i}\hat{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}}\biggr)\biggr)^2 - \frac{1}{4}\biggl(\ln\biggl(-\frac{r\hat{v}_{i}\bar{v}_{i+1}}{q\bar{v}_{i}\hat{v}_{i+1}}\biggr)\biggr)^2 - \frac{1}{4}\biggl(\ln\biggl(-\frac{p\bar{v}_{i}\tilde{v}_{i+1}}{r\tilde{v}_{i}\bar{v}_{i+1}}\biggr)\biggr)^2\nonumber\\ && \;\;\;\;\;\;\; + \frac{1}{2}\bigl(\ln{\hat{\bar{v}}_{i}}\ln{\tilde{\bar{v}}_{i+1}}-\ln{\tilde{\bar{v}}_{i}}\ln{\hat{\bar{v}}_{i+1}}\bigr) + \frac{1}{2}\bigl(\ln{\tilde{\bar{v}}_{i}}\ln{\hat{\tilde{v}}_{i+1}}-\ln{\hat{\tilde{v}}_{i}}\ln{\tilde{\bar{v}}_{i+1}}\bigr)\nonumber\\ && \;\;\;\;\;\;\; + \frac{1}{2}\bigl(\ln{\hat{\tilde{v}}_{i}}\ln{\hat{\bar{v}}_{i+1}}-\ln{\hat{\bar{v}}_{i}}\ln{\hat{\tilde{v}}_{i+1}}\bigr) - \frac{1}{2}\bigl(\ln{\hat{v}_{i}}\ln{\tilde{v}_{i+1}}-\ln{\tilde{v}_{i}}\ln{\hat{v}_{i+1}}\bigr)\nonumber\\ && \;\;\;\;\;\;\; - \frac{1}{2}\bigl(\ln{\bar{v}_{i}}\ln{\hat{v}_{i+1}}-\ln{\hat{v}_{i}}\ln{\bar{v}_{i+1}}\bigr) - \frac{1}{2}\bigl(\ln{\tilde{v}_{i}}\ln{\bar{v}_{i+1}}-\ln{\bar{v}_{i}}\ln{\tilde{v}_{i+1}}\bigr)\nonumber\\ && \;\;\;\;\;\;\; + \ln\biggl(\frac{\bar{v}_{i+1}}{\bar{v}_{i}}\biggr)\ln\biggl(\frac{\tilde{\bar{v}}_{i}}{\hat{\bar{v}}_{i}}\biggr) + \ln\biggl(\frac{\tilde{v}_{i+1}}{\tilde{v}_{i}}\biggr)\ln\biggl(\frac{\hat{\tilde{v}}_{i}}{\tilde{\bar{v}}_{i}}\biggr) + \ln\biggl(\frac{\hat{v}_{i+1}}{\hat{v}_{i}}\biggr)\ln\biggl(\frac{\hat{\bar{v}}_{i}}{\hat{\tilde{v}}_{i}}\biggr)\biggr\}. \end{eqnarray} From the equations, we have that \begin{subequations} \begin{eqnarray} \hat{\tilde{v}}_{i} & = & \frac{p\hat{v}_{i}\tilde{v}_{i+1}-q\tilde{v}_{i}\hat{v}_{i+1}}{p\tilde{v}_{1}-q\hat{v}_{1}}\cdot\frac{v_{1}}{v_{i+1}},\\ \hat{\bar{v}}_{i} & = & \frac{q\bar{v}_{i}\hat{v}_{i+1}-r\hat{v}_{i}\bar{v}_{i+1}}{q\hat{v}_{1}-r\bar{v}_{1}}\cdot\frac{v_{1}}{v_{i+1}},\\ \tilde{\bar{v}}_{i} & = & \frac{r\tilde{v}_{i}\bar{v}_{i+1}-p\bar{v}_{i}\tilde{v}_{i+1}}{r\bar{v}_{1}-p\tilde{v}_{1}}\cdot\frac{v_{1}}{v_{i+1}}, \end{eqnarray} \end{subequations} for $i\in\{0,\dots,N\}$, and of course $v_{0}=v_{N+1}=1$ (so for $i=1$ these are trivial identities). Identifying also $v_{N+2}=v_{1}$, the equations hold trivially for $i=N+1$. Substituting these in gives \begin{eqnarray} \Gamma & = & \sum_{i=0}^{N}\biggl\{{\rm Li}_{2}\biggl(-\frac{p\tilde{v}_{i+1}}{r\bar{v}_{i+1}}\cdot \frac{q\bar{v}_{i}\hat{v}_{i+1}-r\hat{v}_{i}\bar{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}-q\tilde{v}_{i}\hat{v}_{i+1}}\biggr) +{\rm Li}_{2}\biggl(-\frac{\hat{v}_{i+1}}{\bar{v}_{i+1}}\cdot \frac{r\tilde{v}_{i+1}\bar{v}_{i+2}-p\bar{v}_{i+1}\tilde{v}_{i+2}}{p\hat{v}_{i+1}\tilde{v}_{i+2}-q\tilde{v}_{i+1}\hat{v}_{i+2}}\biggr)\nonumber\\ && \;\;\;\;\;\;\; -{\rm Li}_{2}\biggl(-\frac{p\tilde{v}_{i+1}}{r\bar{v}_{i+1}}\cdot \frac{q\bar{v}_{i}\hat{v}_{i+1}-r\hat{v}_{i}\bar{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}-q\tilde{v}_{i}\hat{v}_{i+1}}\biggr) -{\rm Li}_{2}\biggl(-\frac{\hat{v}_{i}}{\bar{v}_{i}}\cdot \frac{r\tilde{v}_{i}\bar{v}_{i+1}-p\bar{v}_{i}\tilde{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}-q\tilde{v}_{i}\hat{v}_{i+1}}\biggr)\nonumber\\ && \;\;\;\;\;\;\; + \ln\biggl(-\frac{\tilde{v}_{i+1}}{\bar{v}_{i+1}}\cdot \frac{q\bar{v}_{i}\hat{v}_{i+1}-r\hat{v}_{i}\bar{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}-q\tilde{v}_{i}\hat{v}_{i+1}}\biggr) \ln\biggl(-\frac{q\hat{v}_{i+1}}{r\bar{v}_{i+1}}\cdot \frac{r\tilde{v}_{i+1}\bar{v}_{i+2}-p\bar{v}_{i+1}\tilde{v}_{i+2}}{p\hat{v}_{i+1}\tilde{v}_{i+2}-q\tilde{v}_{i+1}\hat{v}_{i+2}}\biggr)\nonumber\\ && \;\;\;\;\;\;\; - \ln\biggl(-\frac{\tilde{v}_{i}}{\bar{v}_{i}}\cdot \frac{q\bar{v}_{i}\hat{v}_{i+1}-r\hat{v}_{i}\bar{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}-q\tilde{v}_{i}\hat{v}_{i+1}}\biggr) \ln\biggl(-\frac{q\hat{v}_{i+1}}{r\bar{v}_{i+1}}\cdot \frac{r\tilde{v}_{i}\bar{v}_{i+1}-p\bar{v}_{i}\tilde{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}-q\tilde{v}_{i}\hat{v}_{i+1}}\biggr)\nonumber\\ && \;\;\;\;\;\;\; - \frac{1}{4}\biggl(\ln\biggl(-\frac{q\tilde{\bar{v}}_{i}\hat{\bar{v}}_{i+1}}{p\hat{\bar{v}}_{i}\tilde{\bar{v}}_{i+1}}\biggr)\biggr)^2 + \frac{1}{4}\biggl(\ln\biggl(-\frac{r\hat{\tilde{v}}_{i}\tilde{\bar{v}}_{i+1}}{q\tilde{\bar{v}}_{i}\hat{\tilde{v}}_{i+1}}\biggr)\biggr)^2 + \frac{1}{4}\biggl(\ln\biggl(-\frac{p\hat{\bar{v}}_{i}\hat{\tilde{v}}_{i+1}}{r\hat{\tilde{v}}_{i}\hat{\bar{v}}_{i+1}}\biggr)\biggr)^2\nonumber\\ && \;\;\;\;\;\;\; + \frac{1}{4}\biggl(\ln\biggl(-\frac{q\tilde{v}_{i}\hat{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}}\biggr)\biggr)^2 - \frac{1}{4}\biggl(\ln\biggl(-\frac{r\hat{v}_{i}\bar{v}_{i+1}}{q\bar{v}_{i}\hat{v}_{i+1}}\biggr)\biggr)^2 - \frac{1}{4}\biggl(\ln\biggl(-\frac{p\bar{v}_{i}\tilde{v}_{i+1}}{r\tilde{v}_{i}\bar{v}_{i+1}}\biggr)\biggr)^2\nonumber\\ && \;\;\;\;\;\;\; + \frac{1}{2}\bigl(\ln{\hat{\bar{v}}_{i}}\ln{\tilde{\bar{v}}_{i+1}}-\ln{\tilde{\bar{v}}_{i}}\ln{\hat{\bar{v}}_{i+1}}\bigr) + \frac{1}{2}\bigl(\ln{\tilde{\bar{v}}_{i}}\ln{\hat{\tilde{v}}_{i+1}}-\ln{\hat{\tilde{v}}_{i}}\ln{\tilde{\bar{v}}_{i+1}}\bigr)\nonumber\\ && \;\;\;\;\;\;\; + \frac{1}{2}\bigl(\ln{\hat{\tilde{v}}_{i}}\ln{\hat{\bar{v}}_{i+1}}-\ln{\hat{\bar{v}}_{i}}\ln{\hat{\tilde{v}}_{i+1}}\bigr) - \frac{1}{2}\bigl(\ln{\hat{v}_{i}}\ln{\tilde{v}_{i+1}}-\ln{\tilde{v}_{i}}\ln{\hat{v}_{i+1}}\bigr)\nonumber\\ && \;\;\;\;\;\;\; - \frac{1}{2}\bigl(\ln{\bar{v}_{i}}\ln{\hat{v}_{i+1}}-\ln{\hat{v}_{i}}\ln{\bar{v}_{i+1}}\bigr) - \frac{1}{2}\bigl(\ln{\tilde{v}_{i}}\ln{\bar{v}_{i+1}}-\ln{\bar{v}_{i}}\ln{\tilde{v}_{i+1}}\bigr)\nonumber\\ && \;\;\;\;\;\;\; + \ln\biggl(\frac{\bar{v}_{i+1}}{\bar{v}_{i}}\biggr)\ln\biggl(\frac{\tilde{\bar{v}}_{i}}{\hat{\bar{v}}_{i}}\biggr) + \ln\biggl(\frac{\tilde{v}_{i+1}}{\tilde{v}_{i}}\biggr)\ln\biggl(\frac{\hat{\tilde{v}}_{i}}{\tilde{\bar{v}}_{i}}\biggr) + \ln\biggl(\frac{\hat{v}_{i+1}}{\hat{v}_{i}}\biggr)\ln\biggl(\frac{\hat{\bar{v}}_{i}}{\hat{\tilde{v}}_{i}}\biggr)\biggr\}. \end{eqnarray} Clearly some of these terms will cancel, reducing to \begin{eqnarray} \Gamma & = & \sum_{i=0}^{N}\frac{1}{2}\biggl\{ 2\biggl[{\rm Li}_{2}\biggl(-\frac{\hat{v}_{i+1}}{\bar{v}_{i+1}}\cdot \frac{r\tilde{v}_{i+1}\bar{v}_{i+2}-p\bar{v}_{i+1}\tilde{v}_{i+2}}{p\hat{v}_{i+1}\tilde{v}_{i+2}-q\tilde{v}_{i+1}\hat{v}_{i+2}}\biggr) -{\rm Li}_{2}\biggl(-\frac{\hat{v}_{i}}{\bar{v}_{i}}\cdot \frac{r\tilde{v}_{i}\bar{v}_{i+1}-p\bar{v}_{i}\tilde{v}_{i+1}}{p\hat{v}_{i}\tilde{v}_{i+1}-q\tilde{v}_{i}\hat{v}_{i+1}}\biggr) \biggr]\nonumber\\ && \;\;\;\;\;\;\; + 2\ln\biggl(-\frac{r\bar{v}_{1}-p\tilde{v}_{1}}{p\tilde{v}_{1}-q\hat{v}_{1}}\biggr) \biggl[\ln\biggl(\frac{\hat{\bar{v}}_{i+1}\tilde{v}_{i+1}}{\hat{\tilde{v}}_{i+1}\bar{v}_{i+1}}\biggr) -\ln\biggl(\frac{\hat{\bar{v}}_{i}\tilde{v}_{i}}{\hat{\tilde{v}}_{i}\bar{v}_{i}}\biggr)\biggr]\nonumber\\ && \;\;\;\;\;\;\; - \ln{p}\biggl[\ln\biggl(\frac{\tilde{\bar{v}}_{i+1}\hat{v}_{i+1}}{\hat{\tilde{v}}_{i+1}\bar{v}_{i+1}}\biggr) -\ln\biggl(\frac{\tilde{\bar{v}}_{i}\hat{v}_{i}}{\hat{\tilde{v}}_{i}\bar{v}_{i}}\biggr)\biggr] - \ln{q}\biggl[\ln\biggl(\frac{\hat{\bar{v}}_{i+1}\tilde{v}_{i+1}}{\hat{\tilde{v}}_{i+1}\bar{v}_{i+1}}\biggr) -\ln\biggl(\frac{\hat{\bar{v}}_{i}\tilde{v}_{i}}{\hat{\tilde{v}}_{i}\bar{v}_{i}}\biggr)\biggr]\nonumber\\ && \;\;\;\;\;\;\; + \ln{r}\bigl[\ln\bigl(\hat{\bar{v}}_{i+1}\tilde{\bar{v}}_{i+1}\tilde{v}_{i+1}\hat{v}_{i+1}\bigr) -\ln\bigl(\hat{\bar{v}}_{i}\tilde{\bar{v}}_{i}\tilde{v}_{i}\hat{v}_{i}\bigr)\bigr] - 2\ln{r}\bigl[\ln\bigl(\hat{\tilde{v}}_{i+1}\bar{v}_{i+1}\bigr) -\ln\bigl(\hat{\tilde{v}}_{i}\bar{v}_{i}\bigr)\bigr]\nonumber\\ && \;\;\;\;\;\;\; +\bigl[(\ln{\hat{\tilde{v}}_{i+1}})^2-(\ln{\hat{\tilde{v}}_{i}})^2\bigr]-\bigl[\ln{\tilde{\bar{v}}_{i+1}}\ln{\hat{\tilde{v}}_{i+1}}-\ln{\hat{\tilde{v}}_{i}}\ln{\tilde{\bar{v}}_{i}}\bigr] +\bigl[\ln{\hat{\bar{v}}_{i+1}}\ln{\tilde{\bar{v}}_{i+1}}-\ln{\tilde{\bar{v}}_{i}}\ln{\hat{\bar{v}}_{i}}\bigr]\nonumber\\ && \;\;\;\;\;\;\; +2\bigl[\ln{\bar{v}_{i+1}}\ln{\hat{\tilde{v}}_{i+1}}-\ln{\hat{\tilde{v}}_{i}}\ln{\bar{v}_{i}}\bigr] -2\bigl[\ln{\bar{v}_{i+1}}\ln{\hat{\bar{v}}_{i+1}}-\ln{\bar{v}_{i}}\ln{\hat{\bar{v}}_{i}}\bigr]\nonumber\\ && \;\;\;\;\;\;\; +2\bigl[\ln{\hat{v}_{i+1}}\ln{\hat{\bar{v}}_{i+1}}-\ln{\hat{v}_{i}}\ln{\hat{\bar{v}}_{i}}\bigr] -2\bigl[\ln{\hat{v}_{i+1}}\ln{\hat{\tilde{v}}_{i+1}}-\ln{\hat{v}_{i}}\ln{\hat{\tilde{v}}_{i}}\bigr]\nonumber\\ && \;\;\;\;\;\;\; -\bigl[\ln{\tilde{v}_{i+1}}\ln{\bar{v}_{i+1}}-\ln{\bar{v}_{i}}\ln{\tilde{v}_{i}}\bigr]+\bigl[\ln{\hat{v}_{i+1}}\ln{\tilde{v}_{i+1}}-\ln{\tilde{v}_{i}}\ln{\hat{v}_{i}}\bigr] +\bigl[(\ln{\bar{v}_{i+1}})^2-(\ln{\bar{v}_{i}})^2\bigr]\nonumber\\ && \;\;\;\;\;\;\; -\bigl[\ln{\bar{v}_{i+1}}\ln{\hat{v}_{i+1}}-\ln{\hat{v}_{i}}\ln{\bar{v}_{i}}\bigr] -\bigl[\ln{\hat{\tilde{v}}_{i+1}}\ln{\hat{\bar{v}}_{i+1}}-\ln{\hat{\bar{v}}_{i}}\ln{\hat{\tilde{v}}_{i}}\bigr]\biggr\}. \end{eqnarray} On summation, all of these terms will cancel out, leaving zero. \section{Further remarks}\label{Section_conclusion} Multi-component quad equations generalizing the one derived in \cite{NijPapCapQui1992}, as well as the one given here, are of course of interest. The multidimensional consistency was used as integrability criterion to obtain generalizations of the lattice Boussinesq equation in \cite{Hie2011}, and identification of these systems with a generalized dispersion relation within the direct linearization framework \cite{ZhaZhaNij2011} opens up a potentially fruitful avenue in this direction. As indicated in Section \ref{Section_Symmetry}, the transformation theory of such systems is quite involved even before non-local B\"acklund-type transformations are considered. There is also a hierarchy presented in \cite{Atk2008} with the lattice Schwarzian KdV and lattice Schwarzian Boussinesq equations as the lowest two members; this system was obtained directly in relation to the modified hierarchy (\ref{discreteMGD}) by generalizing a B\"acklund transformation known in the case $N=1$ since \cite{NijRamGraOht2001}: \begin{subequations} \begin{eqnarray} \tilde{z}_{n}-z_{n} & = & \frac{1}{p}\frac{v_{n+1}\tilde{v}_{n-1}}{v_{n}\tilde{v}_{n}},\\ \hat{z}_{n}-z_{n} & = & \frac{1}{q}\frac{v_{n+1}\hat{v}_{n-1}}{v_{n}\hat{v}_{n}}, \end{eqnarray} \end{subequations} where $n\in\{1,\dots,N\}$ and by our convention $v_0=v_{N+1}=1$ as usual. It can readily be seen that this system is compatible in the $z_{n}$ provided the $v_{n}$ satisfy \eqref{discreteMGD}, and similarly it is compatible in the $v_{n}$ provided the $z_{n}$ satisfy the system \begin{subequations} \begin{eqnarray} \frac{\hat{z}_{i+1}-z_{i+1}}{\tilde{z}_{i+1}-z_{i+1}} & = & \frac{\hat{\tilde{z}}_{i}-\tilde{z}_{i}}{\hat{\tilde{z}}_{i}-\hat{z}_{i}}, \;\;\;\;\;\; i\in\{1,\dots,N-1\},\\ \frac{p^{N+1}}{q^{N+1}}\prod_{i=1}^{N}\frac{\tilde{z}_{i}-z_{i}}{\hat{z}_{i}-z_{i}} & = & \frac{\hat{\tilde{z}}_{N}-\tilde{z}_{N}}{\hat{\tilde{z}}_{N}-\hat{z}_{N}}. \end{eqnarray} \end{subequations} Unlike for the modified hierarchy from which it is derived, the expected Lagrangian formulation of this multidimensionally consistent Schwarzian hierarchy is presently not known. The Lagrangian formulation of the multicomponent quad equation in this paper contributes to the mounting evidence that the closure relation and the associated least-action principle are key properties of discrete integrable systems (and indeed of the corresponding continuous systems as well, cf. \cite{XenNijLob2010}). An important motivating factor to study such a structure is its potential physical significance; in particular the Lagrangian multi-form structure constitutes a possible departure point for the quantization of such integrable models along the line of a path integral formulation. \section*{Acknowledgments} JA was supported by the Australian Research Council (ARC) Discovery Grant \#DP110104151. SBL was supported by the ARC Discovery Grant \#DP110100077. This grant also partially supported FWN during a recent visit to La Trobe University (Melbourne) where this work was undertaken. He currently holds a Royal Society/Leverhulme Trust Senior Research Fellowship (2011/12).

https://arxiv.org/abs/1210.2012

Complete monotonicity, completely monotonic degree, integral representations, and an inequality related to the exponential, trigamma, and modified Bessel functions

In the paper, the authors verify the complete monotonicity of the difference $e^{1/t}-\psi'(t)$ on $(0,\infty)$, compute the completely monotonic degree and establish integral representations of the remainder of the Laurent series expansion of $e^{1/z}$, and derive an inequality which gives a lower bound for the first order modified Bessel function of the first kind.

\section{Introduction} In~\cite[Lemma~2]{Yang-Fan-2008-Dec-simp.tex}, the inequality \begin{equation}\label{e-1-t-1} \psi'(t)<e^{1/t}-1 \end{equation} on $(0,\infty)$ was obtained and applied, where $\psi(t)$ stands for the digamma function which may be defined by the logarithmic derivative \begin{equation*} \psi(t)=[\ln\Gamma(t)]'=\frac{\Gamma'(t)}{\Gamma(t)} \end{equation*} and $\Gamma(t)$ is the classical Euler gamma function which may be defined for $\Re z>0$ by \begin{equation* \Gamma(z)=\int^\infty_0t^{z-1} e^{-t}\td t. \end{equation*} The derivatives $\psi'(z)$ and $\psi''(z)$ of $\psi(z)$ are respectively called the tri- and tetra-gamma functions. As a whole, the derivatives $\psi^{(k)}(z)$ for $k\in\{0\}\cup\mathbb{N}$ are called the polygamma functions. \par The first aim of this paper is to generalize the inequality~\eqref{e-1-t-1} to complete monotonicity of a difference between both sides of~\eqref{e-1-t-1}. \par Our first result can be formulated as Theorem~\ref{CM-Exp-thm} below. \begin{thm}\label{CM-Exp-thm} The function \begin{equation}\label{alpha-exp=psi-eq} h(t)=e^{1/t}-\psi'(t) \end{equation} is completely monotonic, that is, $(-1)^{k-1}h^{(k-1)}(t)\ge0$ for $k\in\mathbb{N}$, on $(0,\infty)$ and \begin{equation}\label{h(t)-limit=1} \lim_{t\to\infty}h(t)=1. \end{equation} \end{thm} Recently, the notion ``completely monotonic degree'' was introduced in~\cite{psi-proper-fraction-degree-two.tex}, which may be regarded as a slight but essential modification of~\cite[Definition~1.5]{Koumandos-Pedersen-09-JMAA}. \begin{dfn}[{\cite[Definition~1]{psi-proper-fraction-degree-two.tex}}]\label{x-degree-dfn} Let $f(t)$ be a function defined on $(0,\infty)$ and have derivatives of all orders. A number $r\in\mathbb{R}\cup\{\pm\infty\}$ is said to be the completely monotonic degree of $f(t)$ with respect to $t\in(0,\infty)$ if $t^rf(t)$ is a completely monotonic function on $(0,\infty)$ but $t^{r+\varepsilon}f(t)$ is not for any positive number $\varepsilon>0$. \end{dfn} For convenience, the notation \begin{equation} \cmdeg{t}[f(t)] \end{equation} was designed in~\cite[p.~9890]{psi-proper-fraction-degree-two.tex} to denote the completely monotonic degree $r$ of $f(t)$ with respect to $t\in(0,\infty)$. \par It was pointed out in~\cite[p.~9890]{psi-proper-fraction-degree-two.tex} that the degrees of completely monotonic functions on $(0,\infty)$ are at least zero and that if a function $f(t)$ on $(0,\infty)$ has a nonnegative completely monotonic degree then it must be a completely monotonic function on $(0,\infty)$. Equivalently speaking, a function defined on $(0,\infty)$ is completely monotonic if and only if its completely monotonic degree is not negative. \par The second aim of this paper is to compute the completely monotonic degree and to establish integral representations of the remainder of the Laurent series expansion of the exponential function $e^{1/z}$. \par Our second result may be stated as the following theorem. \begin{thm}\label{exp=k=degree=k+1-int-thm} For $k\in\mathbb{N}\cup\{0\}$ and $z\ne0$, let \begin{equation}\label{exp=k=sum-eq-degree=k+1} H_k(z)=e^{1/z}-\sum_{m=0}^k\frac{1}{m!}\frac1{z^m}. \end{equation} \begin{enumerate} \item The completely monotonic degree of $H_k(t)$ on $(0,\infty)$ meets \begin{equation}\label{H-k(t)-degree} \cmdeg{t}[H_k(t)]=k+1. \end{equation} \item For $\Re z>0$, the function $H_k(z)$ has the integral representation \begin{equation}\label{exp=k=degree=k+1-int} H_k(z)=\frac1{k!(k+1)!}\int_0^\infty {}_1F_2(1;k+1,k+2;t)t^k e^{-zt}\td t, \end{equation} where the hypergeometric series \begin{equation} {}_pF_q(a_1,\dotsc,a_p;b_1,\dotsc,b_q;x)=\sum_{n=0}^\infty\frac{(a_1)_n\dotsm(a_p)_n} {(b_1)_n\dotsm(b_q)_n}\frac{x^n}{n!} \end{equation} for $b_i\notin\{0,-1,-2,\dotsc\}$ and the shifted factorial $(a)_0=1$ and \begin{equation} (a)_n=a(a+1)\dotsm(a+n-1) \end{equation} for $n>0$ and any real or complex number $a$. \item For $\Re z>0$, the function $H_k(z)$ has the integral representation \begin{equation}\label{exp=k=degree=k+1-int-bes} H_k(z)=\frac1{z^{k+1}}\int_0^\infty \frac{I_{k+2}\bigl(2 \sqrt{t}\,\bigr)}{t^{(k+2)/2}} e^{-zt}\td t, \end{equation} where \begin{equation}\label{I=nu(z)-eq} I_\nu(z)= \sum_{k=0}^\infty\frac1{k!\Gamma(\nu+k+1)}\biggl(\frac{z}2\biggr)^{2k+\nu} \end{equation} for $\nu\in\mathbb{R}$ and $z\in\mathbb{C}$ is the modified Bessel function of the first kind. \end{enumerate} \end{thm} As an application of Theorems~\ref{CM-Exp-thm} and~\ref{exp=k=degree=k+1-int-thm}, the following inequality for the first order modified Bessel function of the first kind $I_1$ may be derived. \begin{thm}\label{Bessel-2-ineq-thm} For $t>0$, we have \begin{equation}\label{I=1-exp-ineq} I_1(t)>\frac{(t/2)^3}{1-e^{-(t/2)^2}}. \end{equation} \end{thm} \section{Proof of Theorem~\ref{CM-Exp-thm}} From the well known formula \begin{equation}\label{polygamma} \psi^{(n)}(z)=(-1)^{n+1}\int_0^{\infty}\frac{u^n}{1-e^{-u}}e^{-zu}\td u \end{equation} for $\Re z>0$ and $n\in\mathbb{N}$, see~\cite[p.~260, 6.4.1]{abram}, it is ready that $\lim_{t\to\infty}\psi^{(n)}(t)=0$ for $n\in\mathbb{N}$. So, the limit~\eqref{h(t)-limit=1} may be deduced immediately and, by \begin{equation}\label{exp-frac1x-expans} \bigl(e^{1/t}\bigr)^{(i)}=(-1)^ie^{1/t}\frac1{t^{2i}}\sum_{k=0}^{i-1}a_{i,k}t^{k} \end{equation} for $i\in\mathbb{N}$ and $t\ne0$, where \begin{equation} a_{i,k}=\binom{i}{k}\binom{i-1}{k}{k!} \end{equation} for $0\le k\le i-1$, in~\cite[Theorem~2.1]{exp-reciprocal-cm-IJOPCM.tex}, \begin{equation}\label{h(t)-i-der-to0} h^{(i)}(t)=\bigl(e^{1/t}\bigr)^{(i)}-\psi^{(i+1)}(t) =(-1)^ie^{1/t}\sum_{k=0}^{i-1}\frac{a_{i,k}}{t^{2i-k}}-\psi^{(i+1)}(t)\to0 \end{equation} for $i\in\mathbb{N}$ as $t\to\infty$. \par Utilizing the recurrence formula \begin{equation}\label{abram-6.4.7} \psi^{(n)}(z+1)=\psi^{(n)}(z)+(-1)^n\frac{n!}{z^{n+1}} \end{equation} in~\cite[p.~260, 6.4.7]{abram} and calculating reveal \begin{align*} h(t+1)-h(t)&=e^{1/(t+1)}-e^{1/t}+\psi'(t)-\psi'(t+1)\\ &=e^{1/(t+1)}-e^{1/t}+\frac1{t^2}\\ &=\frac1{t^2}+\sum_{k=0}^\infty\frac1{(k+1)!}\biggl[\frac1{(t+1)^{k+1}}-\frac1{t^{k+1}}\biggr], \end{align*} \begin{equation*} [h(t+1)-h(t)]^{(i)}=(-1)^i\frac{(i+1)!}{t^{i+2}} +\sum_{k=0}^\infty\frac{(-1)^i(i+k)!}{(k+1)!k!} \biggl[\frac1{(t+1)^{i+k+1}}-\frac1{t^{i+k+1}}\biggr], \end{equation*} and \begin{align*} (-1)^i[h(t+1)-h(t)]^{(i)}&=\frac{(i+1)!}{t^{i+2}} +\sum_{k=0}^\infty\frac{(i+k)!}{(k+1)!k!} \biggl[\frac1{(t+1)^{i+k+1}}-\frac1{t^{i+k+1}}\biggr]\\ &<\frac{(i+1)!}{t^{i+2}} +\sum_{k=0}^2\frac{(i+k)!}{(k+1)!k!} \biggl[\frac1{(t+1)^{i+k+1}}-\frac1{t^{i+k+1}}\biggr]\\ &=\frac{i!}{12t^{i+3}(t+1)^{i+3}}f_i(t), \end{align*} where \begin{align*} f_i(t)&=6(i+1)t(t+1)\bigl[(t+1)^{i+2}+t^{i+2}\bigr] -12 t^2(t+1)^2\bigl[(t+1)^{i+1}-t^{i+1}\bigr]\\ &\quad-(i+1)(i+2)\bigl[(t+1)^{i+3}-t^{i+3}\bigr]\\ &=6(i+1)t(t+1)\Biggl[\sum_{\ell=0}^{i+2}\binom{i+2}{\ell}t^{\ell}+t^{i+2}\Biggr] -12 t^2(t+1)^2\sum_{\ell=0}^{i}\binom{i+1}{\ell}t^{\ell}\\ &\quad-(i+1)(i+2)\sum_{\ell=0}^{i+2}\binom{i+3}{\ell}t^{\ell}\\ &=\frac{(i-1)(i+4)(i+5)}{2}\biggl[\frac{(2-i)(i+3)}{3}t-i\biggr]t^2-i(i+1)(i+5)t\\ &\quad-\sum_{\ell=4}^{i}\biggl[(i+1)(i+2)\binom{i+3}{\ell}-6(i+1)\binom{i+3}{\ell-1} +12\binom{i+3}{\ell-2}\biggr]t^\ell\\ &=\frac{(i-1)(i+4)(i+5)}{2}\biggl[\frac{(2-i)(i+3)}{3}t-i\biggr]t^2-i(i+1)(i+5)t\\ &\quad-(1+i)(2+i)-(i+4)(i+5)\sum_{\ell=4}^{i} \frac{(i-\ell+1)(i-\ell+2)}{\ell(i-\ell+5)}\binom{i+3}{\ell-1}t^\ell \end{align*} and an empty sum is understood to be nil. As a result, the function $f_i(t)$ is negative and \begin{equation}\label{(-1)=i[h(t+1)-h(t)]=(i)} (-1)^i[h(t+1)-h(t)]^{(i)}=(-1)^i[h(t+1)]^{(i)}-(-1)^i[h(t)]^{(i)}<0 \end{equation} for all $i\ge0$ and $t\in(0,\infty)$. Hence, by consecutive recursion and~\eqref{h(t)-i-der-to0}, \begin{multline*} (-1)^i[h(t)]^{(i)}\ge(-1)^i[h(t+1)]^{(i)}\ge(-1)^i[h(t+2)]^{(i)}\ge \dotsm\\* \ge(-1)^i[h(t+k)]^{(i)}\ge(-1)^i\lim_{k\to\infty}[h(t+k)]^{(i)}=0 \end{multline*} for $i\in\mathbb{N}$ and $t\in(0,\infty)$. This implies that the function $h(t)$ is decreasing on $(0,\infty)$. Combining this monotonicity with~\eqref{h(t)-limit=1} gives $h(t)>1$ on $(0,\infty)$. In conclusion, by definition, the function $h(t)$ is completely monotonic on $(0,\infty)$. The proof of Theorem~\ref{CM-Exp-thm} is complete. \section{Proof of Theorem~\ref{exp=k=degree=k+1-int-thm}} It is general knowledge that the exponential function $e^{1/z}$ for $z\in\mathbb{C}$ with $z\ne0$ can be expanded into the Laurent series \begin{equation}\label{exp-reciproc-series} e^{1/z}=\sum_{m=0}^\infty\frac1{m!}\frac1{z^m},\quad z\ne0. \end{equation} Therefore, it is clear that \begin{equation}\label{exp-residue-term} H_k(z)=\sum_{m=k+1}^\infty\frac{1}{m!}\frac1{z^m},\quad z\ne0 \end{equation} and $x^{k+1}H_k(x)$ is completely monotonic on $(0,\infty)$. That is, \begin{equation}\label{H-k(t)-degree>k+1} \cmdeg{t}[H_k(t)]\ge k+1. \end{equation} Since, for any $\varepsilon>0$, the function \begin{equation*} x^{k+1+\varepsilon}H_k(x)=x^\varepsilon\sum_{m=0}^\infty\frac{1}{(m+k+1)!}\frac1{x^m} \end{equation*} tends to $\infty$ as $x\to\infty$, we see that for any $\varepsilon>0$ the function $x^{k+1+\varepsilon}H_k(x)$ is not completely monotonic on $(0,\infty)$. That is, \begin{equation}\label{H-k(t)-degree<k+1} \cmdeg{t}[H_k(t)]\le k+1. \end{equation} Combining~\eqref{H-k(t)-degree>k+1} and~\eqref{H-k(t)-degree<k+1} leads to~\eqref{H-k(t)-degree}. \par For $\Re z>0$ and $\Re k>0$, it was listed in~\cite[p.~255, 6.1.1]{abram} that \begin{equation* \Gamma(z)=k^z\int_0^\infty t^{z-1}e^{-kt}\td t. \end{equation*} This formula can be rearranged as \begin{equation}\label{Gamma(z)=k-z-int-rearr} \frac1{z^w}=\frac1{\Gamma(w)}\int_0^\infty t^{w-1}e^{-zt}\td t \end{equation} for $\Re z>0$ and $\Re w>0$. Substituting the formula~\eqref{Gamma(z)=k-z-int-rearr} into~\eqref{exp-residue-term} yields \begin{align*} H_k(z)&=\int_0^\infty \Biggl[\sum_{m=k+1}^\infty\frac{1}{m!}\frac1{\Gamma(m)}t^{m-1}\Biggr]e^{-zt}\td t\\ &=\frac1{k!(k+1)!}\int_0^\infty {}_1F_2(1;k+1,k+2;t)t^k e^{-zt}\td t. \end{align*} The integral representation~\eqref{exp=k=degree=k+1-int} follows. \par The function $H_k(z)$ can be rewritten as \begin{align*} H_k(z)&=\frac1{z^{k+1}}\sum_{m=0}^\infty\frac{1}{(k+1+m)!}\frac1{z^m}\\ &=\frac1{z^{k+1}}\int_0^\infty \Biggl[\sum_{m=0}^\infty\frac{1}{(k+1+m)!} \frac1{\Gamma(m)}t^{m-1}\Biggr] e^{-zt}\td t\\ &=\frac1{z^{k+1}}\int_0^\infty \frac{I_{k+2}\bigl(2 \sqrt{t}\,\bigr)}{t^{(k+2)/2}} e^{-zt}\td t. \end{align*} The integral representation~\eqref{exp=k=degree=k+1-int-bes} follows. Theorem~\ref{exp=k=degree=k+1-int-thm} is thus proved. \section{Proof of Theorem~\ref{Bessel-2-ineq-thm}} When $k=0$, the integral representations~\eqref{exp=k=degree=k+1-int} and~\eqref{exp=k=degree=k+1-int-bes} become \begin{equation}\label{open-answer-1} e^{1/z}=1+\int_0^\infty \frac{I_1\bigl(2\sqrt{t}\,\bigr)}{\sqrt{t}\,} e^{-zt}\td t \end{equation} and \begin{equation}\label{open-answer-2} e^{1/z}=1+\frac1{z}\int_0^\infty \frac{I_{2}\bigl(2 \sqrt{t}\,\bigr)}{t} e^{-zt}\td t \end{equation} for $\Re z>0$. Hence, by~\eqref{polygamma} for $n=1$, the function $h(z)$ defined by~\eqref{alpha-exp=psi-eq} has the following integral representation \begin{equation}\label{h(t)-int-rep-eq} h(z)=1+\int_0^\infty \biggl[\frac{I_1\bigl(2\sqrt{u}\,\bigr)}{\sqrt{u}\,} -\frac{u}{1-e^{-u}}\biggr]e^{-zu}\td u \end{equation} and \begin{equation}\label{h(t)-int-rep-eq-der} (-1)^kh^{(k)}(t)=\int_0^\infty \biggl[\frac{I_1\bigl(2\sqrt{u}\,\bigr)}{\sqrt{u}\,} -\frac{u}{1-e^{-u}}\biggr]u^ke^{-tu}\td u \end{equation} for $k\ge1$ are completely monotonic on $(0,\infty)$. The famous Hausdorff-Bernstein-Widder Theorem~\cite[p.~161, Theorem~12b]{widder} states that a necessary and sufficient condition that $f(x)$ should be completely monotonic for $0<x<\infty$ is that \begin{equation} \label{berstein-1} f(x)=\int_0^\infty e^{-xt}\td\alpha(t), \end{equation} where $\alpha(t)$ is non-decreasing and the integral converges for $0<x<\infty$. Consequently, the function in the bracket of~\eqref{h(t)-int-rep-eq-der} is not less than zero, that is, \begin{equation*} \frac{I_1\bigl(2\sqrt{u}\,\bigr)}{\sqrt{u}\,} \ge\frac{u}{1-e^{-u}} \end{equation*} in which replacing $2\sqrt{u}\,$ by $t$ yields the inequality~\eqref{I=1-exp-ineq}. The proof of Theorem~\ref{Bessel-2-ineq-thm} is complete. \begin{rem} The integral representations~\eqref{open-answer-1} and~\eqref{open-answer-2} supply answers to an open problem posed in~\cite[p.~127, Section~4]{exp-reciprocal-cm-IJOPCM.tex}. \end{rem} \begin{rem} This paper is a corrected and extended version of the preprint~\cite{simp-exp-degree.tex}. \end{rem}

https://arxiv.org/abs/2110.00692

Decomposing a graph into subgraphs with small components

The component size of a graph is the maximum number of edges in any connected component of the graph. Given a graph $G$ and two integers $k$ and $c$, $(k,c)$-Decomposition is the problem of deciding whether $G$ admits an edge partition into $k$ subgraphs with component size at most $c$. We prove that for any fixed $k \ge 2$ and $c \ge 2$, $(k,c)$-Decomposition is NP-complete in bipartite graphs. Also, when both $k$ and $c$ are part of the input, $(k,c)$-Decomposition is NP-complete even in trees. Moreover, $(k,c)$-Decomposition in trees is W[1]-hard with parameter $k$, and is FPT with parameter $c$. In addition, we present approximation algorithms for decomposing a tree either into the minimum number of subgraphs with component size at most $c$, or into $k$ subgraphs minimizing the maximum component size. En route to these results, we also obtain a fixed-parameter algorithm for Bin Packing with the bin capacity as parameter.

\section{Introduction} Edge coloring is the problem of assigning a color to each edge of a graph such that no two adjacent edges have the same color. Any edge coloring of a graph with maximum degree $\Delta$ requires at least $\Delta$ colors. Vizing's theorem states that $\Delta + 1$ colors suffice. The minimum number of colors in any edge coloring of a graph is its \emph{chromatic index}. Deciding whether the chromatic index of a graph with maximum degree $\Delta$ is $\Delta$ or $\Delta+1$ is generally a hard computational problem, although a constructive proof of Vizing's theorem~\cite{MG92} implies a polynomial-time approximation algorithm with additive error at most $1$. Holyer~\cite{Ho81} proved that edge coloring cubic graphs with three colors, where $\Delta=3$, is already NP-complete. Extending this result, Leven and Galil~\cite{LG83} proved that edge coloring $\Delta$-regular graphs with $\Delta$ colors is NP-complete for any $\Delta \ge 3$. Cai and Ellis~\cite{CE91} further proved that edge coloring $\Delta$-regular line graphs of bipartite graphs with $\Delta$ colors is NP-complete for any odd $\Delta \ge 3$. On the other hand, it is known that $\Delta$ colors always suffice for edge coloring bipartite graphs; see~\cite[Proposition 18.1.3]{Ca94} for a simple proof. Moreover, edge coloring bipartite graphs (or bipartite multigraphs) admits an efficient polynomial-time exact algorithm, even when $\Delta$ is part of the input~\cite{COS01}. Define the \emph{component size} of a graph as the maximum number of edges in any connected component of the graph. In this paper, we study the following problem that generalizes edge coloring: \begin{definition} Given a graph $G$ and two integers $k \ge 2$ and $c \ge 1$, \textsc{$(k,c)$-Decomposition}\ is the problem of deciding whether $G$ admits an edge partition into $k$ subgraphs with component size at most $c$. \end{definition} We show that graph decomposition becomes hard, even in bipartite graphs, when the upper bound on the component size of the resulting subgraphs is relaxed from $1$ (as in edge coloring) to any fixed $c \ge 2$: \begin{theorem}\label{thm:k2c2} For any fixed $k \ge 2$ and $c \ge 2$, \textsc{$(k,c)$-Decomposition}\ is NP-complete in bipartite graphs. \end{theorem} The \emph{arboricity} of a graph is the minimum number of parts in an edge partition of the graph such that each part induces a forest~\cite{Na64}. The trees in the forests may be restricted to various subclasses of trees, leading to several related concepts of arboricity. For example, the \emph{linear arboricity} (respectively, \emph{star arboricity}) of a graph is the minimum number of parts in an edge partition of the graph such that each part induces a disjoint union of paths (respectively, stars). While the arboricity of any graph can be computed in polynomial time~\cite{GW92}, the arboricities with various restrictions on the trees often turn out to be hard to compute~\cite{Ji18}. Our proof of Theorem~\ref{thm:k2c2} shows that arboricity and star arboricity with bounded component size are also hard to compute: \begin{corollary}\label{cor:arboricity} For any fixed $k \ge 2$ and $c \ge 2$, deciding whether a bipartite graph admits an edge partition into $k$ forests with component size at most $c$ is NP-complete. \end{corollary} \begin{corollary}\label{cor:stararboricity} For any fixed $k \ge 2$ and $c \ge 2$, deciding whether a bipartite graph admits an edge partition into $k$ forests of stars with component size at most $c$ is NP-complete. \end{corollary} The \emph{linear $c$-arboricity} of a graph is the minimum number of parts in an edge partition of the graph into \emph{linear $c$-forests}, i.e., disjoint unions of paths of length at most $c$~\cite{BFHP84}. Note that linear $1$-arboricity is simply chromatic index. Thus the aforementioned results on edge coloring~\cite{Ho81,LG83,CE91} imply that determining the linear $1$-arboricity of a graph is hard. Bermond, Fouquet, Habib, and Peroche~\cite{BFHP84} proved that deciding whether a cubic graph has linear $3$-arboricity at most $2$ is NP-complete, and conjectured that determining the linear $c$-arboricity of a graph is hard for all $c$. Since linear $2$-arboricity is equivalent to arboricity and star arboricity with component size at most $2$, the $c = 2$ case of this conjecture is confirmed by our Corollary~\ref{cor:arboricity} and Corollary~\ref{cor:stararboricity}: \begin{corollary}\label{cor:linear2} For any fixed $k \ge 2$, deciding whether a bipartite graph has linear $2$-arboricity $k$ is NP-complete. \end{corollary} For two graphs $G$ and $H$, an \emph{$H$-decomposition} of $G$ is an edge partition of $G$ into subgraphs isomorphic to $H$. For any fixed graph $H$, the problem of deciding whether an input graph $G$ admits an $H$-decomposition is NP-complete when $H$ has a component of at least three edges~\cite{DT97}, and is polynomially solvable when every component of $H$ has at most two edges~\cite{BL09}, in contrast to Corollary~\ref{cor:linear2}. \bigskip We established in Theorem~\ref{thm:k2c2} that \textsc{$(k,c)$-Decomposition}\ is NP-complete in bipartite graphs. It is natural to ask whether the problem remains hard in simpler graphs. Our next theorem characterizes the complexity of \textsc{$(k,c)$-Decomposition}\ in trees: \begin{theorem}\label{thm:tree} When both $k$ and $c$ are part of the input, \textsc{$(k,c)$-Decomposition}\ in trees is NP-complete, is W[1]-hard with parameter $k$, and is FPT with parameter $c$. \end{theorem} Theorem~\ref{thm:tree} implies that for any fixed $c$, \textsc{$(k,c)$-Decomposition}\ in trees can be solved in polynomial time. For the related arboricity problem of deciding whether a tree admits an edge partition into $k$ linear $c$-forests, Chang, Chen, Fu, and Huang~\cite{CCFH00} presented an algorithm that runs in polynomial time when $c$ is fixed. \bigskip The decision problem \textsc{$(k,c)$-Decomposition}\ may be extended to two different optimization problems: \begin{definition} Given a graph $G$ and an integer $c \ge 1$, \textsc{Minimum-Color Decomposition}\ is the problem of decomposing $G$ into the minimum number of subgraphs with component size at most $c$. \end{definition} \begin{definition} Given a graph $G$ and an integer $k \ge 2$, \textsc{Minimum-Size Decomposition}\ is the problem of decomposing $G$ into $k$ subgraphs minimizing the maximum component size. \end{definition} Theorem~\ref{thm:k2c2} implies that \textsc{Minimum-Color Decomposition}\ for any fixed $c \ge 2$ and \textsc{Minimum-Size Decomposition}\ for any fixed $k \ge 2$ are APX-hard and inapproximable within a factor of $3/2$ in bipartite graphs. The previous results on edge coloring~\cite{Ho81,LG83,CE91} imply that \textsc{Minimum-Color Decomposition}\ for $c = 1$ is APX-hard and inapproximable within a factor of $4/3$ in general graphs. In the next two theorems, we show that both minimization problems can be approximated very well in trees: \begin{theorem}\label{thm:mcd} There is a linear-time approximation algorithm for \textsc{Minimum-Color Decomposition}\ in trees that finds an edge partition of any tree $T$ into at most $k^* + 1$ subgraphs with component size at most $c$, where $k^*$ is the smallest number such that $T$ can be decomposed into $k^*$ subgraphs with component size at most $c$. \end{theorem} \begin{theorem}\label{thm:msd} There is a polynomial-time approximation scheme for \textsc{Minimum-Size Decomposition}\ in trees that finds an edge partition of any tree $T$ into $k$ subgraphs with component size at most $(1+\epsilon)c^*$, where $c^*$ is the smallest number such that $T$ can be decomposed into $k$ subgraphs with component size at most $c^*$. \end{theorem} Kleinberg, Motwani, Raghavan, and Venkatasubramanian~\cite{KMRV97} introduced \textsc{$(k,c)$-Fragmented Coloring}, the problem of vertex partitioning a graph into $k$ induced subgraphs such that the number of vertices in each component of each subgraph is at most $c$. Our problem \textsc{$(k,c)$-Decomposition}\ is equivalent to \textsc{$(k,c)$-Fragmented Coloring}\ in line graphs, since edge partitioning a graph $G$ is the same as vertex partitioning the line graph $L(G)$. The problem \textsc{$(k,c)$-Fragmented Coloring}\ in general graphs is easily shown to be NP-hard for any $k \ge 2$ and $c \ge 2$, because the graph property of at most $c$ vertices in each component is additive and induced-hereditary, while vertex partitioning into fixed additive induced-hereditary properties is NP-hard in general~\cite{Fa04}, except the special case of bipartite testing, i.e., vertex partitioning into two independent sets. Also, the problem is trivial in bipartite graphs (and hence in trees), since every bipartite graph has a vertex partition into two induced subgraphs in which every component has only one vertex. \bigskip Our proofs of Theorem~\ref{thm:tree}, Theorem~\ref{thm:mcd}, and Theorem~\ref{thm:msd} are based on a close relationship between \textsc{$(k,c)$-Decomposition}\ in trees and the classical problem of \textsc{Bin Packing}: \begin{definition}\label{def:bp} Given $n$ items of integer weights $w_i$, $1 \le i \le n$, and $k$ bins of integer capacity $c$, \textsc{Bin Packing}\ is the problem of deciding whether the $n$ items can be packed into the $k$ bins, that is, whether the set of $n$ items can be partitioned into $k$ subsets, such that the total weight of the items in each subset is at most $c$. \textsc{Unary Bin Packing}\ is the version of \textsc{Bin Packing}\ where all integers in the problem instance are encoded in unary. \end{definition} It is well-known that \textsc{Bin Packing}\ is strongly NP-hard~\cite{GJ79}, and hence \textsc{Unary Bin Packing}\ is NP-hard, when both $k$ and $c$ are part of the input. In terms of parameterized complexity, Jansen et~al.~\cite{JKMS13} proved that \textsc{Unary Bin Packing}, even with the condition $\sum_{i=1}^n w_i = k c$, is already W[1]-hard with parameter $k$. In the following theorem, we obtain a fixed-parameter algorithm for \textsc{Bin Packing}\ with parameter $c$: \begin{theorem}\label{thm:fpt} \textsc{Bin Packing}\ admits an exact algorithm running in $2^{O(c^{3/2})}N^{O(1)}$ time, where $N$ is the size of the problem instance encoded in binary. \end{theorem} Hochbaum and Shmoys~\cite{HS87} introduced the notion of \emph{dual approximation algorithms} that approximate the feasibility of a problem rather than its optimality. The existence of dual approximation algorithms for \textsc{Bin Packing}, similar to the one in the following theorem, are known~\cite[Sections 9.3.2.1 and 9.3.2.2]{Ho97}. We obtain a simple alternative via Theorem~\ref{thm:fpt}. \begin{theorem}\label{thm:dual} \textsc{Bin Packing}\ admits a dual approximation algorithm that, given $n$ items and $k$ bins of capacity $c$, and given $0 < \epsilon \le 1$, either decides that the $n$ items can be packed into $k$ enlarged bins of capacity $(1+\epsilon)c$ and returns yes, or decides that the $n$ items cannot be packed into $k$ bins of capacity $c$ and returns no, in $2^{O(\epsilon^{-3})}N^{O(1)}$ time, where $N$ is the size of the problem instance encoded in binary. \end{theorem} Throughout the paper, the \emph{size} of a graph is the number of edges in it. We may refer to an edge partition of a graph into $k$ subgraphs and the corresponding $k$-color assignment to the edges interchangeably. \section{Hardness of decomposing bipartite graphs} In this section we prove Theorem~\ref{thm:k2c2}, Corollary~\ref{cor:arboricity}, and Corollary~\ref{cor:stararboricity}. Fix any $k \ge 2$ and $c \ge 2$. The problem \textsc{$(k,c)$-Decomposition}\ and the related arboricity problems are clearly in NP\@. We prove their NP-hardness in bipartite graphs in the following. \subsection{Building blocks $H_i$} The building blocks of our construction are a series of bipartite graphs $H_i = H_i(k,c)$ for $i \ge 0$. $H_0$ is the star $K_{1,kc}$ with one edge designated as an \emph{outlet}. For $i > 0$, $H_i$ has $c^i$ outlets and is constructed recursively as follows. Take $k-1$ disjoint copies of $H_{i-1}$ and $c^{i-1}$ disjoint copies of $K_{1,c}$. Attach a distinct outlet from each copy of $H_{i-1}$ to the center of each copy of $K_{1,c}$. Then designate the edges of all copies of $K_{1,c}$ in the resulting graph $H_i$ as its outlets. Refer to Figures~\ref{fig:k2c3} and~\ref{fig:k3c2} for some examples. It is easy to verify that $H_i$ is bipartite. Since it includes $(k-1)^i$ copies of $K_{1,kc}$ and $\sum_{j = 1}^i c^{j-1}(k-1)^{i - j}$ copies of $K_{1,c}$, its size is at most \begin{equation}\label{eq:H} (k-1)^i kc + \sum_{j = 1}^i c^j(k-1)^{i - j} \le (k-1)^i kc + (c+k)^i \le (c+k)^i(kc + 1) \le (c+k)^{i+2}. \end{equation} \begin{figure}[htbp] \centering\includegraphics{k2c3} \caption{$H_0$, $H_1$, and $H_2$ for $k = 2$ and $c = 3$.} \label{fig:k2c3} \end{figure} \begin{figure}[htbp] \centering\includegraphics{k3c2} \caption{$H_2$ for $k = 3$ and $c = 2$.} \label{fig:k3c2} \end{figure} \begin{lemma}\label{lem:H} $H_i$ admits an edge partition into $k$ subgraphs with component size at most $c$. Moreover, in any edge partition of $H_i$ into $k$ subgraphs with component size at most $c$, each component must be a star $K_{1,c}$, and the $c^i$ outlets must belong to the same subgraph. \end{lemma} \begin{proof} We prove the lemma by induction on $i$. For the base case when $i = 0$, the lemma clearly holds for $H_0 = K_{1,kc}$. Now let $i \ge 1$ for the inductive step. We first show that $H_i$ admits an edge partition into $k$ subgraphs with component size at most $c$. By the induction hypothesis, $H_{i-1}$ admits an edge partition into $k$ subgraphs in which each component is a star $K_{1,c}$, and the $c^{i-1}$ outlets are in the same subgraph. Consider the $k$-color assignment to the edges of $H_{i-1}$ corresponding to this edge partition. Then by rotation of colors, there is an edge partition of the union of the $k-1$ copies of $H_{i-1}$ in $H_i$, such that the outlets from different copies have different colors. Thus the outlets from the $k-1$ copies of $H_{i-1}$ have $k-1$ different colors. Assign the remaining color to all edges of the $c^{i-1}$ disjoint copies of $K_{1,c}$ in $H_i$. This color assignment corresponds to an edge partition of $H_i$ into $k$ subgraphs in which each component is a star $K_{1,c}$, and the $c^i$ outlets are all in the same subgraph. We next show that in any edge partition of $H_i$ into $k$ subgraphs with component size at most $c$, each component must be a star $K_{1,c}$, and the $c^i$ outlets must be in the same subgraph. Consider an arbitrary edge partition of $H_i$ into $k$ subgraphs with at most $c$ edges in each component, and the corresponding $k$-color assignment to the edges of $H_i$. Then in the derived edge partition of each copy of $H_{i-1}$ in $H_i$, each component is a star $K_{1,c}$, and all $c^{i-1}$ outlets are in the same subgraph, by the induction hypothesis. Moreover, since each component in each copy of $H_{i-1}$ already has the maximum allowable size $c$, the outlets from different copies of $H_{i-1}$ must have different colors to avoid forming larger components. Then all edges of the $c^{i-1}$ disjoint copies of $K_{1,c}$ in $H_i$ must have the only remaining color. \end{proof} \subsection{Reduction for $k = 2$} Our proof for the $k = 2$ case of Theorem~\ref{thm:k2c2}, Corollary~\ref{cor:arboricity}, and Corollary~\ref{cor:stararboricity} is based on a reduction from the NP-complete problem \textsc{2-Colorability of 3-Uniform Hypergraphs}~\cite{Lo73}. A \emph{$q$-uniform hypergraph} consists of a set $V$ of vertices and a set $E$ of hyperedges, where each hyperedge in $E$ is a subset of exactly $q$ vertices in $V$. A hypergraph is \emph{$k$-colorable} if its vertices can be colored with $k$ colors such that in each hyperedge, not all $q$ vertices have the same color. Given a $3$-uniform hypergraph $G$, we will construct a bipartite graph $G_2$ with maximum degree $2c$ such that $G$ is $2$-colorable if and only if $G_2$ admits an edge partition into $2$ subgraphs with component size at most $c$. Let $\Delta$ be the maximum degree of $G$, that is, the maximum number of hyperedges in which a vertex is contained. Let $j$ be the smallest integer such that $c^j \ge (c-1)\Delta$. Include in $G_2$ a distinct copy of $H_j$ for each vertex in $G$, and a distinct copy of $K_{1,c+1}$ for each hyperedge in $G$. For each hyperedge in $G$, which consists of three vertices, take $c+1$ distinct outlets from the corresponding three copies of $H_j$, with at least one outlet from each copy, then attach the $c+1$ outlets to the $c+1$ leaves of the copy of $K_{1,c+1}$ corresponding to the hyperedge, with one outlet to each leaf. This completes the construction of $G_2$. Recall that $H_j$ has $c^j$ outlets. Since $c^j \ge (c-1)\Delta$, each copy of $H_j$ can contribute at least $c-1$ distinct outlets to each adjacent copy of $K_{1,c+1}$. On the other hand, of the three copies of $H_j$ adjacent to each copy of $K_{1,c+1}$, each must contribute at least $1$ and hence at most $c+1 - 2 = c-1$ of the $c+1$ outlets. Thus there are enough outlets from each copy of $H_j$ to the adjacent copies of $K_{1,c+1}$. By our choice of $j$, we have $c^{j-1} < (c-1)\Delta$ and hence $c^j < c^2 \Delta$. Recall~\eqref{eq:H} that the size of $H_j$ is at most $(c+k)^{j+2}$. Since $k = 2$ and $c \ge 2$, we have $$ (c+k)^{j+2} \le (c^2)^{j+2} = c^4 (c^j)^2 < c^4 (c^2 \Delta)^2 = c^8 \Delta^2. $$ The reduction is clearly polynomial. The following lemma completes the proof for the $k = 2$ case of Theorem~\ref{thm:k2c2}, Corollary~\ref{cor:arboricity} and Corollary~\ref{cor:stararboricity}: \begin{lemma}\label{lem:k2} $G$ is $2$-colorable if and only if $G_2$ admits an edge partition into two subgraphs with component size at most $c$, if and only if $G_2$ admits an edge partition into two forests with component size at most $c$, if and only if $G_2$ admits an edge partition into two forests of stars with component size at most $c$. \end{lemma} \begin{proof} It suffices to prove a cycle of four implications: \begin{enumerate}\setlength\itemsep{0pt} \item if $G$ is $2$-colorable, then $G_2$ admits an edge partition into two forests of stars with component size at most $c$, \item if $G_2$ admits an edge partition into two forests of stars with component size at most $c$, then $G_2$ admits an edge partition into two forests with component size at most $c$, \item if $G_2$ admits an edge partition into two forests with component size at most $c$, then $G_2$ admits an edge partition into two subgraphs with component size at most $c$, \item if $G_2$ admits an edge partition into two subgraphs with component size at most $c$, then $G$ is $2$-colorable. \end{enumerate} We first prove implication 1. Suppose that there is a $2$-coloring of the vertices of $G$ such that in each hyperedge, not all three vertices have the same color. Color the edges in each copy of $H_j$ in $G_2$ with two colors to partition them into two subgraphs as in Lemma~\ref{lem:H}, such that all $c^j$ outlets in each copy of $H_j$ have the same color as the corresponding vertex in $G$. Next color the edges in each copy of $K_{1,c+1}$ in $G_2$ corresponding to a hyperedge in $G$, such that each edge of $K_{1,c+1}$ has a color different (note that there are only two colors available) from the adjacent outlet. Since in each hyperedge of $G$ not all three vertices have the same color, and since the $c+1$ outlets adjacent to the $c+1$ edges of $K_{1,c+1}$ include at least one outlet from each of the three copies of $H_j$, it follows that not all $c+1$ edges have the same color. Thus we obtain an edge partition of $G_2$ into two subgraphs with component size at most $c$. Note that the two subgraphs are indeed two forests of stars. Implications 2 and 3 are trivial. We next prove implication 4. Suppose that $G_2$ admits an edge partition into two subgraphs with component size at most $c$. Color the edges of $G_2$ with two colors according to this partition. Then it follows by Lemma~\ref{lem:H} that in each copy of $H_j$ in $G_2$, each component is a star with $c$ edges, and all outlets have the same color. Assign this color to the corresponding vertex in $G$. Since each outlet of $H_j$ is already in a component of the maximum allowable size $c$, each edge of $K_{1,c+1}$ must have a color different from its adjacent outlet. In each copy of $K_{1,c+1}$, not all $c+1$ edges can have the same color under the constraint of component size at most $c$. Correspondingly, in each hyperedge of $G$, not all three vertices have the same color either. Thus $G$ is $2$-colorable. \end{proof} \subsection{Reduction for $k \ge 3$} Our proof for the $k \ge 3$ case of Theorem~\ref{thm:k2c2}, Corollary~\ref{cor:arboricity}, and Corollary~\ref{cor:stararboricity} is based on a reduction from the NP-complete problem \textsc{$k$-Colorability}, which asks whether a given graph $G$ admits a vertex coloring with $k$ colors. Maffray and Preissmann~\cite{MP96} proved that for any $k \ge 3$, \textsc{$k$-Colorability} is NP-hard even in triangle-free graphs. Emden-Weinert, Hougardy, and Kreuter~\cite{EHK98} proved that for any $k \ge 3$, \textsc{$k$-Colorability} is NP-hard in graphs with maximum degree $k + \lceil\sqrt{k}\,\rceil - 1$. Note that \textsc{$k$-Colorability} for $k = 2$ is just bipartite testing which is well-known to be solvable in polynomial time. Thus we had to prove the $k = 2$ case by reduction from a different problem. Let $G$ be an input graph for \textsc{$k$-Colorability}. Construct $G_k$ as follows. Let $j$ be the smallest integer such that $c^j$ is at least the maximum degree of $G$. For each vertex in $G$, include in $G_k$ a distinct copy of $H_j$. For each edge in $G$, which consists of two vertices, take a distinct outlet from the copy of $H_j$ for each vertex, then join the two ends of the two outlets into one vertex. It is easy to verify that $G_k$ is bipartite, and the reduction is polynomial. Moreover, by a similar argument as in the proof of Lemma~\ref{lem:k2}, $G$ is $k$-colorable if and only if $G_k$ admits an edge partition into $k$ subgraphs with component size at most $c$, if and only if $G_k$ admits an edge partition into $k$ forests with component size at most $c$, if and only if $G_k$ admits an edge partition into $k$ forests of stars with component size at most $c$. \bigskip This completes the proof of Theorem~\ref{thm:k2c2}, Corollary~\ref{cor:arboricity}, and Corollary~\ref{cor:stararboricity}. \section{Fixed-parameter algorithm for bin packing} In this section we prove Theorem~\ref{thm:fpt} by obtaining a fixed-parameter algorithm for \textsc{Bin Packing}\ with parameter $c$. Recall Definition~\ref{def:bp} that the input to \textsc{Bin Packing}\ consists of $n$ items of integer weights $w_i$, $1 \le i \le n$, and $k$ bins of integer capacity $c$. The problem has a solution only if $\max\{ w_i \mid 1 \le i \le n\} \le c$ and $\sum_{i=1}^n w_i \le k c$, which we assume. Let $N$ be the size of the problem instance encoded in binary. Then $N = \Theta(n + \sum_{i=1}^n \log w_i + \log k + \log c)$. In particular, we have $\log kc = O(N)$. For $1 \le w \le c$, let $a_w$ be the number of items with weight $w$. In $N^{O(1)}$ time, we can compute $a_w$ for all $w$. Then $\sum_{w=1}^c w\,a_w = \sum_{i=1}^n w_i \le k c$. With the multiplicities $a_w$ for $1 \le w \le c$, we no longer need $w_i$ for $1 \le i \le n$. Henceforth, we work on the $c$ integers $a_w$ only. Without loss of generality, we increase $a_1$ until $\sum_{w=1}^c w\,a_w = k c$. Note that each $a_w$ has magnitude at most $kc$, where $\log kc = O(N)$. It remains to decide whether the items with multiplicities $a_w$ can be partitioned into $k$ subsets, each of total weight exactly $c$. Consider the two vectors $\boldsymbol{w} = (1,2,\ldots,c)$ and $\boldsymbol{a} = (a_1,a_2,\ldots,a_c)$ in $\mathbb{Z}_{\ge0}^c$. Then the condition $\sum_{w=1}^c w\,a_w = kc$ becomes $\boldsymbol{w}\cdot\boldsymbol{a} = kc$. For $b \ge 0$, let $$ \P_b = \{\, \boldsymbol{x} \in \mathbb{Z}_{\ge 0}^c \mid \boldsymbol{w}\cdot \boldsymbol{x} = b \,\}. $$ Then $\P_c$ is the set of vectors corresponding to all feasible \emph{patterns} of item multiplicities for a single bin, and we can reformulate the problem as deciding whether there exists a multiset of vectors in $\P_c$ whose sum is $\boldsymbol{a}$. Recall that the \emph{partition number} $p(n)$ for $n \ge 0$ is the number of multisets of positive integers summing up to $n$. In particular, $p(0) = 1$. Then, by definition, the size of $\P_c$ is exactly $p(c)$. Moreover, the size of $\P_b$ is at most $p(b)$ for all $b \ge 0$. It is known that $p(n) = 2^{O(\sqrt{n})}$~\cite[Theorem 15.7]{LW01}. Let $d(c) \ge 0$ be the smallest integer such that $2^d > \sum_{h=0}^d p(hc)$ for all $d > d(c)$. Since $\sum_{h=0}^d p(hc) = 2^{O(\sqrt{d c})}$, we have $d(c) = O(c)$. Let $\boldsymbol{v}_i$, $1 \le i \le p(c)$, be the vectors in $\P_c$. For any index set $I \subseteq \{ 1,\ldots, p(c) \}$, the vector $\sum_{i\in I} \lambda_i \boldsymbol{v}_i$, where $\lambda_i \in \mathbb{Z}_{\ge 0}$ for all $i \in I$, is called an \emph{integer conical combination} of vectors in $\P_c$. Consider the \emph{integer conical hull} of $\P_c$ consisting of all integer conical combinations of vectors in $\P_c$: $$ \mathrm{int.cone}(\P_c) = \bigg\{\, \sum_{i=1}^{p(c)} \lambda_i \boldsymbol{v}_i \;\Big\vert\; \lambda_i \in \mathbb{Z}_{\ge 0} \textrm{ for } 1 \le i \le p(c) \bigg\}. $$ We prove the following lemma using a common technique for Carath\'eodory bounds; see for example~\cite[Lemma 3]{ES06}. \begin{lemma}\label{lem:cone} Every vector in $\mathrm{int.cone}(\P_c)$ can be represented as the integer conical combination of at most $d(c)$ distinct vectors in $\P_c$. \end{lemma} \begin{proof} Let $I$ be any set of $d$ distinct indices from $\{ 1,\ldots, p(c) \}$, where $0 \le d \le p(c)$. Then $I$ has $2^d$ distinct subsets. Denote by $|H|$ the size of a set $H$. For each subset $H \subseteq I$, where $0 \le |H| \le d$, the vector $\boldsymbol{x} = \sum_{i\in H} \boldsymbol{v}_i$ satisfies $$ \boldsymbol{w}\cdot\boldsymbol{x} = \boldsymbol{w}\cdot \sum_{i\in H} \boldsymbol{v}_i = \sum_{i\in H} \boldsymbol{w}\cdot \boldsymbol{v}_i = \sum_{i\in H} c = |H|\,c, $$ and hence $\boldsymbol{x} \in \P_{hc}$ with $h = |H|$. Thus $\boldsymbol{x}$ is one of the vectors in $\cup_{h=0}^d \P_{hc}$. Consider any vector $\boldsymbol{u} = \sum_{i\in I} \lambda_i \boldsymbol{v}_i$, where $\lambda_i \in \mathbb{Z}_{> 0}$ for all $i\in I$. Suppose that $d > d(c)$. Then by our choice of $d(c)$, we have $2^d > \sum_{h=0}^d p(hc)$. Note that the number of vectors in $\cup_{h=0}^d \P_{hc}$ is at most $\sum_{h=0}^d p(hc)$. Thus the number of distinct subsets of $I$ exceeds the number of vectors in $\cup_{h=0}^d \P_{hc}$. By the pigeonhole principle, there are two distinct subsets $A$ and $B$ of $I$ such that $\sum_{i\in A} \boldsymbol{v}_i = \sum_{i\in B} \boldsymbol{v}_i$. Then $A' = A\setminus B$ and $B' = B\setminus A$ are two distinct and disjoint subsets of $I$ such that $\sum_{i\in A'} \boldsymbol{v}_i = \sum_{i\in B'} \boldsymbol{v}_i$. Assume without loss of generality that $A' \neq \emptyset$. Let $\lambda = \min\{\, \lambda_i \mid i \in A' \,\}$. We can rewrite $\boldsymbol{u}$ as follows: \begin{align*} \sum_{i\in I} \lambda_i \boldsymbol{v}_i &= \sum_{i\in I\setminus(A'\cup B')} \lambda_i \boldsymbol{v}_i + \sum_{i\in A'} \lambda_i \boldsymbol{v}_i + \sum_{i\in B'} \lambda_i \boldsymbol{v}_i\\ &= \sum_{i\in I\setminus(A'\cup B')} \lambda_i \boldsymbol{v}_i + \sum_{i\in A'} (\lambda_i - \lambda) \boldsymbol{v}_i + \sum_{i\in B'} (\lambda_i + \lambda) \boldsymbol{v}_i = \sum_{i\in I} \mu_i \boldsymbol{v}_i, \end{align*} where $$ \mu_i = \left\{ \begin{array}{ll} \lambda_i &\textrm{for } i\in I\setminus(A'\cup B')\\ \lambda_i - \lambda &\textrm{for } i\in A'\\ \lambda_i + \lambda &\textrm{for } i\in B'. \end{array} \right. $$ Note that $\mu_i \in \mathbb{Z}_{\ge 0}$ for all $i\in I$, and that $\mu_i = 0$ for at least one index $i \in A' \subseteq I$. Thus $\boldsymbol{u}$ is expressed as the integer conical combination of at most $d - 1$ distinct vectors in $\P_c$. By repeating the above argument, we can eventually obtain a representation of $\boldsymbol{u}$ as the integer conical combination of at most $d(c)$ distinct vectors in $\P_c$. \end{proof} If there is a multiset of vectors in $\P_c$ summing up to $\boldsymbol{a}$, then $\boldsymbol{a}$ is an integer conical combination of vectors in $\P_c$, and hence is included in $\mathrm{int.cone}(\P_c)$, and hence by Lemma~\ref{lem:cone} can be represented as an integer conical combination of $d \le d(c)$ distinct vectors in $\P_c$. Thus to decide whether there exists a multiset of vectors in $\P_c$ whose sum is $\boldsymbol{a}$, we can enumerate all subsets of $\{1,\ldots,p(c)\}$ of size $d \le d(c)$, then for each subset $I$, check whether $\sum_{i\in I} \lambda_i \boldsymbol{v}_i = \boldsymbol{a}$ has a solution with $\lambda_i \in \mathbb{Z}_{\ge 0}$ for all $i\in I$. The checking step reduces to an integer linear program, with $O(c)$ linear constraints on at most $d(c) = O(c)$ variables $\lambda_i$, and with integer coefficients at most $kc$, where $\log kc = O(N)$. We now analyze the running time of our algorithm. Recall that converting the item weights $w_i$ to multiplicities $a_w$ takes $N^{O(1)}$ time. Finding the $p(c)$ vectors in $\P_c$, $\boldsymbol{v}_i$ for $1 \le i \le p(c)$, takes $c^{O(c)}$ time. The number of subsets of $\{1,\ldots,p(c)\}$ of size $d \le d(c)$ is at most $(p(c))^{d(c)} = (2^{O(\sqrt{c})})^{O(c)} = 2^{O(c^{3/2})}$. Solving an integer linear program with $n$ variables and $m$ constraints and with maximum coefficient $\Delta$ takes $n^{O(n)} m^{O(1)} (\log\Delta)^{O(1)}$ time\footnote{See discussion at \url{https://cstheory.stackexchange.com/questions/16530} for more refined estimates.}~\cite{Le83,Ka87}. In our setting, this is $(O(c))^{O(c)} (O(c))^{O(1)} (O(N))^{O(1)} = c^{O(c)}N^{O(1)}$. Thus the total running time is $N^{O(1)} + c^{O(c)} + 2^{O(c^{3/2})}\cdot c^{O(c)}N^{O(1)}$, which simplifies to $2^{O(c^{3/2})}N^{O(1)}$, since $c^{O(c)} = 2^{O(c\log c)}$. This completes the proof of Theorem~\ref{thm:fpt}. \newpage \section{Dual approximation algorithm for bin packing} In this section we prove Theorem~\ref{thm:dual}. Fix any $0 < \epsilon \le 1$. Separate the $n$ items into two groups: \emph{small} items of weight at most $\epsilon c$, and \emph{large} items of weight greater than $\epsilon c$. Let $c' = \lfloor 2\epsilon^{-2} \rfloor$. Note that $c' \cdot \epsilon^2 c / 2 \le c$. For each large item of weight $w_i > \epsilon c$, let $w'_i$ be the largest integer such that $w'_i\cdot \epsilon^2 c / 2 \le w_i$. Then $(w'_i + 1)\cdot \epsilon^2 c / 2 > w_i$. The algorithm works as follows. If the total weight of the $n$ items exceeds $kc$, then clearly the $n$ items cannot be packed into $k$ bins of capacity $c$, so return no. Otherwise, construct a set of scaled items including one item of weight $w'_i$ for each large item of weight $w_i$. Use the fixed-parameter algorithm in Theorem~\ref{thm:fpt} to decide whether the set of at most $n$ scaled items $w'_i$ can be packed into $k$ scaled bins of capacity $c'$, then return the answer accordingly. \begin{lemma} If the scaled items can be packed into $k$ scaled bins of capacity $c'$, then the $n$ items can be packed into $k$ enlarged bins of capacity $(1+\epsilon)c$. \end{lemma} \begin{proof} For each large item of weight $w_i > \epsilon c$, the corresponding scaled item has weight $$ w'_i > \frac{w_i}{\epsilon^2 c / 2} - 1 > \frac{\epsilon c}{\epsilon^2 c / 2} - 1 = 2\epsilon^{-1} - 1 \ge \epsilon^{-1}. $$ Thus each scaled bin of capacity $c' = \lfloor 2\epsilon^{-2} \rfloor$ can hold at most $\lfloor 2\epsilon^{-2} \rfloor / \epsilon^{-1} \le 2\epsilon^{-1}$ scaled items. For the scaled items in each scaled bin, the sum of their weights $w'_i$ is at most $c'$, and hence the sum of $w'_i\cdot \epsilon^2 c / 2$ is at most $c'\cdot \epsilon^2 c / 2 \le c$. Recall that $w'_i\cdot \epsilon^2 c / 2 \le w_i < (w'_i + 1)\cdot \epsilon^2 c / 2$, and hence $w_i - w'_i\cdot \epsilon^2 c / 2 < \epsilon^2 c / 2$. Corresponding to the at most $2\epsilon^{-1}$ scaled items in each scaled bin, the total weight of the large items can exceed $c$ by at most $2\epsilon^{-1} \cdot \epsilon^2 c / 2= \epsilon c$, and hence is at most $(1 + \epsilon)c$. Thus we can pack all large items into $k$ enlarged bins of capacity $(1 + \epsilon)c$. Since the total weight of both large and small items is at most $kc$, at least one of the $k$ enlarged bins is filled to at most $c$ of its capacity $(1+\epsilon)c$, and hence can always accommodate one more small item of weight at most $\epsilon c$. Thus we can pack all small items into the $k$ enlarged bins after the large items. \end{proof} \begin{lemma} If the scaled items cannot be packed into $k$ scaled bins of capacity $c'$, then the $n$ items cannot be packed into $k$ bins of capacity $c$. \end{lemma} \begin{proof} We prove the contrapositive. Suppose that the $n$ items of total weight at most $kc$ can be packed into $k$ bins of capacity $c$. Then by discarding the small items and rounding down the weights $w_i$ of the large items to integer multiples of $\epsilon^2 c / 2$, the rounded items of weights $w'_i \cdot \epsilon^2 c / 2$ can be packed into $k$ bins of capacity $c = 2\epsilon^{-2} \cdot \epsilon^2 c / 2$, and hence the scaled items of integer weights $w'_i$ can be packed into $k$ scaled bins of capacity $c' = \lfloor 2\epsilon^{-2} \rfloor$. \end{proof} With at most $n$ scaled items and capacity $c' = O(\epsilon^{-2})$ for the $k$ scaled bins, it follows by Theorem~\ref{thm:fpt} that the running time of our dual approximation algorithm is $2^{O(\epsilon^{-3})}N^{O(1)}$, where $N$ is the size of the problem instance encoded in binary. This completes the proof of Theorem~\ref{thm:dual}. \section{Parameterized complexity of decomposing trees} In this section we prove Theorem~\ref{thm:tree}. The problem \textsc{$(k,c)$-Decomposition}\ is clearly in NP\@. In the following, we prove that when both $k$ and $c$ are part of the input, \textsc{$(k,c)$-Decomposition}\ in trees is not only NP-hard but also W[1]-hard with parameter $k$, and is FPT with parameter $c$. \subsection{NP-hardness and W[1]-hardness with parameter $k$} We give a polynomial reduction from \textsc{Unary Bin Packing}\ with $k$ bins to \textsc{$(k,c)$-Decomposition}\ in trees. Recall that \textsc{Unary Bin Packing}\ is W[1]-hard with parameter $k$~\cite{JKMS13}. Given $n$ items of weights $w_i$, $1 \le i \le n$, and $k$ bins of capacity $c$, we construct a tree $T$ with $n(k-1)c + \sum_{i=1}^n w_i$ edges as follows. For each item of weight $w_i$, make a star with $(k-1)c + w_i$ edges. Then select an arbitrary leaf from each star, and merge the $n$ leaves selected from the $n$ stars into one center vertex, denoted by $v$. We claim that the $n$ items can be packed into $k$ bins of capacity $c$ if and only if the tree $T$ can be decomposed into $k$ subgraphs with component size at most $c$. We first prove the direct implication of the claim. Suppose that the $n$ items can be packed into $k$ bins of capacity $c$. We color the edges of the tree $T$ with $k$ colors as follows. First color the $n$ edges incident to the center vertex $v$ with $k$ colors corresponding to the $n$ items in the $k$ bins. Then for each star with $(k-1)c + w_i$ edges, color $w_i - 1$ additional edges with the same color as the edge incident to the center vertex $v$, and color the remaining $(k-1)c$ edges with the other $k-1$ colors, $c$ edges of each color. Then the edges of each color induce a forest with component size at most $c$. We next prove the reverse implication of the claim. Suppose that the tree $T$ can be decomposed into $k$ subgraphs with component size at most $c$. Color the edges of $T$ with $k$ colors corresponding to the $k$ subgraphs. Then in the star with $(k-1)c + w_i$ edges corresponding to the item of weight $w_i$, the edge incident to the center vertex $v$ must have the same color as at least $w_i - 1$ other edges, because the other $k-1$ colors can accommodate at most $(k-1)c$ edges. Note that for each of the $k$ colors, the edges of this color that are incident to $v$, as well as the other edges of the same color in their respective stars, are all connected, and form one component of size at most $c$. Put the $n$ items into $k$ bins corresponding to the colors of the $n$ edges incident to $v$. Then the items in each bin must have total weight at most $c$ too. \subsection{FPT with parameter $c$} We present a fixed-parameter algorithm for \textsc{$(k,c)$-Decomposition}\ in trees with parameter $c$. Let $T$ be a tree with $n$ vertices and $n - 1$ edges, rooted at an arbitrary vertex of degree $1$. For each edge $e$ between a vertex $v$ and its parent in $T$, denote by $T_e$ the subtree of $T$ rooted at $v$ plus the edge $e$, and denote by $s(e)$ the minimum $s \ge 1$ such that $T_e$ admits an edge partition into $k$ subgraphs with component size at most $c$, under the constraint that $e$ is in a component of size at most $s$. If no such $s$ exists, let $s(e) = c + 1$. Then $T$ admits an edge partition into $k$ subgraphs with component size at most $c$ if and only if $s(e) \in [1,c]$ for all edges $e$. We can compute $s(e)$ by dynamic programming, following a post-order traversal of the rooted tree $T$. For each edge $e$ incident to a leaf, we clearly have $s(e) = 1$. For each edge $e$ between an internal node $v$ and its parent, and for a candidate size $s \ge 1$, we can check whether $s(e) \le s$ as follows. Let $m_v$ be the number of edges incident to $v$. Construct $m_v$ items corresponding to the $m_v$ edges. Each edge $f$ between $v$ and a child corresponds to an item of weight $s(f)$, which has been computed following the post-order traversal. The edge $e$ between $v$ and its parent corresponds to an item of weight $c-s + 1$, where the surplus of $c-s$ in addition to $1$ for the edge $e$ accounts for the difference between the desired component size $s$ for $e$ and the upper bound $c$. Then $T_e$ admits an edge partition into $k$ subgraphs with component size at most $c$, where $e$ is in a component of size at most $s$, if and only if the $m_v$ items can be packed into $k$ bins of capacity $c$. By Theorem~\ref{thm:fpt}, there is a fixed-parameter algorithm for \textsc{Bin Packing}\ that decides whether the $m_v$ items thus constructed can be packed into $k$ bins of capacity $c$ in $2^{O(c^{3/2})}N_v^{O(1)}$ time, where $N_v = O(m_v(1 + \log kc))$ is the encoding size of the \textsc{Bin Packing}\ instance with $m_v$ items. By one invocation of this fixed-parameter algorithm with $s = c$, we either find that $s(e) > c$ and abort the traversal of $T$, or find that $s(e) \le c$. In the latter case, we can then determine the exact value of $s(e)$ in $[1,c]$ by a binary search, using the fixed-parameter algorithm as a decision procedure. The total running time of the algorithm is $$ n^{O(1)} + \sum_v \big( O(1 + \log c) \cdot 2^{O(c^{3/2})}N_v^{O(1)} \big) = 2^{O(c^{3/2})}n^{O(1)}. $$ Thus \textsc{$(k,c)$-Decomposition}\ in trees is FPT with parameter $c$. This completes the proof of Theorem~\ref{thm:tree}. \section{Approximation algorithms for decomposing trees} In this section we prove Theorem~\ref{thm:mcd} and Theorem~\ref{thm:msd}. Let $T$ be a tree with $n$ vertices, $n-1$ edges, and maximum degree $\Delta < n$, rooted at an arbitrary vertex of degree $1$. \subsection{Approximating \textsc{Minimum-Color Decomposition}} We first present a linear-time approximation algorithm for \textsc{Minimum-Color Decomposition}. Perform a pre-order traversal of the rooted tree $T$. The root is incident to only one edge, which is colored arbitrarily. At each vertex other than the root, color the at most $\Delta-1$ edges to its children with at most $\lceil \frac{\Delta - 1}c \rceil$ colors (at most $c$ edges of each color), all different from the color of the edge from its parent. It is straightforward to verify that the edges of each color induce a forest with component size at most $c$, and moreover each component is a star with at most $c$ edges. Let $k' = \lceil \frac{\Delta}c \rceil$ and $k'' = \lceil \frac{\Delta - 1}c \rceil + 1$. The minimum number of subgraphs with component size at most $c$ into which $T$ can be decomposed is at least $k'$. On the other hand, our algorithm colors the edges of $T$ with $k''$ colors, where $k'' \le k' + 1$. This completes the proof of Theorem~\ref{thm:mcd}. \subsection{Approximating \textsc{Minimum-Size Decomposition}} We next present a polynomial-time approximation scheme for \textsc{Minimum-Size Decomposition}. We first note that our approximation algorithm for \textsc{Minimum-Color Decomposition}\ can be slightly modified to yield a $2$-approximation for \textsc{Minimum-Size Decomposition}\ as follows. During the pre-order traversal of the rooted tree $T$, at each vertex other than the root, color the at most $\Delta-1$ edges to its children with at most $k-1$ colors (at most $\lceil \frac{\Delta - 1}{k-1} \rceil$ edges of each color), all different from the color of the edge from its parent. Let $c' = \lceil \frac{\Delta}k \rceil$ and $c'' = \lceil \frac{\Delta - 1}{k-1} \rceil$. In any edge partition of $T$ into $k$ subgraphs, there is a subgraph with component size at least $c'$. On the other hand, our algorithm colors the edges of $T$ with $k$ colors such that the edges of each color induce a forest with component size at most $c''$, where $$ c'' = \left\lceil \frac{\Delta - 1}{k-1} \right\rceil = \left\lceil \frac{k}{k - 1} \cdot \frac{\Delta - 1}k \right\rceil \le \left\lceil \frac{k}{k - 1} \right\rceil \cdot \left\lceil \frac{\Delta - 1}k \right\rceil \le 2\left\lceil \frac{\Delta}k \right\rceil = 2c'. $$ Now fix any $0 < \epsilon \le 1$. To obtain a $(1+\epsilon)$-approximation of the optimal component size $c^*$, it suffices to perform a binary search in the integer range $(c' - 1, c'']$, using a subroutine that decides either $c^* > c$ or $c^* \le (1+\epsilon)c$ for any candidate component size $c$. When the search range is reduced to $(l, h]$ with $h - l = 1$, we must have $c^* \ge h$, and the subroutine with $c = h$ decomposes the tree into $k$ subgraphs with component size at most $(1+\epsilon)\,h$. \bigskip The subroutine on a candidate component size $c$ is similar to our fixed-parameter algorithm with parameter $c$ in Theorem~\ref{thm:tree}. As before, for each edge $e$ between a vertex $v$ and its parent in the rooted tree $T$, denote by $T_e$ the subtree of $T$ rooted at $v$ plus the edge $e$, and denote by $s(e)$ the minimum $s \ge 1$ such that $T_e$ admits an edge partition into $k$ subgraphs with component size at most $c$, under the constraint that $e$ is in a component of size at most $s$. If no such $s$ exists, let $s(e) = c + 1$. By definition, if $s(e) \le c$, then $s(e)$ is the unique integer $s \in [1,c]$ that satisfies the following two conditions: \begin{enumerate}\setlength\itemsep{0pt} \item[A.] $T_e$ admits an edge partition into $k$ subgraphs with component size at most $c$, where $e$ is in a component of size at most $s$. \item[B.] $T_e$ does not admit any edge partition into $k$ subgraphs with component size at most $c$, where $e$ is in a component of size at most $s - 1$. \end{enumerate} Instead of computing $s(e)$ directly, the subroutine tries to find a \emph{feasible} value $s\in[1,c]$ that satisfies the following two conditions, and records it in $\tilde s(e)$: \begin{enumerate}\setlength\itemsep{0pt} \item[C.] $T_e$ admits an edge partition into $k$ subgraphs with component size at most $(1+\epsilon)c$, where $e$ is in a component of size at most $s$. \item[B.] $T_e$ does not admit any edge partition into $k$ subgraphs with component size at most $c$, where $e$ is in a component of size at most $s - 1$. \end{enumerate} The subroutine computes $\tilde s(e)$ by dynamic programming, following a post-order traversal of the rooted tree $T$. For each edge $e$ incident to a leaf, set $\tilde s(e) \gets 1$. For each edge $e$ between an internal node $v$ and its parent, check whether there exists a feasible value $s \in [1,c]$ for $\tilde s(e)$ that satisfies conditions C and B, as follows. Let $m_v$ be the number of edges incident to $v$. Construct $m_v$ items corresponding to the $m_v$ edges. Each edge $f$ between $v$ and a child corresponds to an item of weight $\tilde s(f)$. The edge $e$ between $v$ and its parent corresponds to an item of weight $c-s + 1$. By Theorem~\ref{thm:dual}, there is a dual approximation algorithm for \textsc{Bin Packing}\ that, given the $m_v$ items and $k$ bins of capacity $c$, either decides that the $m_v$ items can be packed into $k$ enlarged bins of capacity $(1+\epsilon)c$ and returns yes, or decides that the $m_v$ items cannot be packed into $k$ bins of capacity $c$ and returns no. First call this dual approximation algorithm with $s = c$. If its answer is no, then abort the traversal of $T$, and declare that $\hl{c^* > c}$. Otherwise, find a feasible value $s \in [1,c]$ for $\tilde s(e)$ that satisfies both conditions C and B by a binary search, using the dual approximation algorithm as a decision procedure. If the traversal of $T$ finishes successfully, declare that $\hl{c^* \le (1+\epsilon)c}$. \bigskip The correctness of the subroutine can be verified by induction following the post-order traversal. In particular, for all edges $e$ such that a feasible value $s \in [1,c]$ is found for $\tilde s(e)$, we have $\tilde s(e) \le s(e)$ because they satisfy the same condition B, while condition C is a relaxation of condition A. Also observe that the answers of the dual approximation algorithm have the following two implications: \begin{itemize}\setlength\itemsep{0pt} \item If the $m_v$ items can be packed into $k$ enlarged bins of capacity $(1+\epsilon)c$, then $T_e$ admits an edge partition into $k$ subgraphs with component size at most $(1+\epsilon)c$, where $e$ is in a component of size at most $s$. \item If the $m_v$ items cannot be packed into $k$ bins of capacity $c$, then $T_e$ does not admit any edge partition into $k$ subgraphs with component size at most $c$, where $e$ is in a component of size at most $s$. \end{itemize} By Theorem~\ref{thm:dual}, the running time of the dual approximation algorithm on the $m_v$ items is $2^{O(\epsilon^{-3})}N_v^{O(1)}$, where $N_v = O(m_v(1 + \log kc))$ is the encoding size of the problem instance. The total running time of the algorithm, including the nested binary searches on $c^* \in (c'-1,c'']$ and on $s \in [1,c]$, is $$ n^{O(1)} + O(\log(\Delta/k))\Big( n^{O(1)} + \sum_v \big( O(1 + \log c) \cdot 2^{O(\epsilon^{-3})}N_v^{O(1)} \big) \Big) = 2^{O(\epsilon^{-3})}n^{O(1)}. $$ Thus we have a polynomial-time approximation scheme for \textsc{Minimum-Size Decomposition}. This completes the proof of Theorem~\ref{thm:msd}. \section{Concluding remark} In our study of graph decomposition problems in this paper, the focus is on bipartite graphs and trees. It may be interesting to investigate the complexities of these problems on other important graph classes too.

https://arxiv.org/abs/2005.00144

An Efficient Noisy Binary Search in Graphs via Median Approximation

Consider a generalization of the classical binary search problem in linearly sorted data to the graph-theoretic setting. The goal is to design an adaptive query algorithm, called a strategy, that identifies an initially unknown target vertex in a graph by asking queries. Each query is conducted as follows: the strategy selects a vertex $q$ and receives a reply $v$: if $q$ is the target, then $v=q$, and if $q$ is not the target, then $v$ is a neighbor of $q$ that lies on a shortest path to the target. Furthermore, there is a noise parameter $0\leq p<\frac{1}{2}$, which means that each reply can be incorrect with probability $p$. The optimization criterion to be minimized is the overall number of queries asked by the strategy, called the query complexity. The query complexity is well understood to be $O(\varepsilon^{-2}\log n)$ for general graphs, where $n$ is the order of the graph and $\varepsilon=\frac{1}{2}-p$. However, implementing such a strategy is computationally expensive, with each query requiring possibly $O(n^2)$ operations.In this work we propose two efficient strategies that keep the optimal query complexity. The first strategy achieves the overall complexity of $O(\varepsilon^{-1}n\log n)$ per a single query. The second strategy is dedicated to graphs of small diameter $D$ and maximum degree $\Delta$ and has the average complexity of $O(n+\varepsilon^{-2}D\Delta\log n)$ per query. We stress out that we develop an algorithmic tool of graph median approximation that is of independent interest: the median can be efficiently approximated by finding a vertex minimizing the sum of distances to a randomly sampled vertex subset of size $O(\varepsilon^{-2}\log n)$.

\section{Introduction} \label{sec:intro} Our research problems originate in the classical ``twenty questions game'' proposed by R\'{e}nyi~\cite{Renyi61} and Ulam~\cite{Ulam76}. The classical problem of binary search with erroneous comparisons received a considerable attention and optimal query complexity algorithms are known, see e.g. \cite{Ben-OrH08,BorgstromK93,DhagatGW92,FeigeRPU94,RivestMKWS80} for asymptotically best results. The binary search in linearly ordered data can be re-casted as a search on a path, where each query selects a vertex $q$ and reply gives whether the target element is $q$, or is to the left or to the right of $q$. This leads to the graph search problem introduced first for trees by Onak and Parys in \cite{OnakP06} and then recently for general graphs by Emamjomeh-Zadeh et~al. in \cite{Emamjomeh-Zadeh:2015aa}. We recall a following formal statement. \setlist[description]{leftmargin=\parindent,labelindent=\parindent} \begin{description} \item[Problem formulation.] Consider an arbitrary simple graph $G$ whose one vertex $v^{*}$ is marked as the \emph{target}. The target is unknown to the query algorithm. Each query points to a vertex $q$, and a \emph{correct} reply does the following: if $v^{*}=q$, then the reply returns $q$, and if $v^{*}\neq q$, then the reply returns a neighbor of $q$ that belongs to a shortest path from $q$ to $v^{*}$, breaking ties arbitrarily. We further assume that some replies can be incorrect: each query receives an erroneous reply (independently) with some fixed probability $0\leq p<\frac{1}{2}$ (the value of the \emph{noise parameter} $p$ is known to the algorithm). The goal is to design an algorithm, also called a \emph{strategy} performing as few queries as possible. \end{description} Typically in the applications of the adaptive query problems the main concern is the number of queries to be performed, i.e., their \emph{query complexity}. This is due to the fact that the queries usually model a time consuming and complex event like making a software check to verify whether it contains a malfunctioning piece of code, c.f. Ben-Asher et al. \cite{Ben-AsherFN99}, or asking users for some sort of feedback c.f. Emamjomeh-Zadeh and Kempe \cite{Emamjomeh-ZadehK17}. However, as a second measure the computational complexity comes into play and it is of practical interest to resolve the question of having an adaptive query algorithm that keeps an optimal query complexity and optimizes the computational cost as a second criterion. This may be especially useful in cases when queries are fast, like communication events over a noisy channel. The asymptotics of the query complexity is quite well understood to be roughly $\frac{\log n}{1-H(p)}=\mathcal{O}(\varepsilon^{-2}\log n)$ (c.f. \cite{DereniowskiTUW19,Emamjomeh-Zadeh:2015aa}), where $n$ is the order of the graph, $\varepsilon=\frac{1}{2}-p$, and $H(p)=-p\log_2 p-(1-p)\log_2(1-p)$ is the entropy. Thus, it is of theoretical and practical interest to know what is the optimal complexity of computing each particular query. This leads us to a general statement of the type of solution we seek. \begin{description} \item[Research question.] How much the computational complexity of an adaptive graph query algorithm can be improved without worsening the query complexity? \end{description} In this work we make the following assumption: a \emph{distance oracle} is available to the algorithm and it gives the graph distance between any pair of vertices. This is dictated by the observation that the computation of multiple-pair shortest paths throughout the search would dominate the computational complexity. On the other hand, we note that this is only used to resolve (multiple times) the following for a query: given a vertex $q$, its neighbor $v$ and an arbitrary vertex $u$, does $v$ lie on a shortest path from $q$ to $u$? Thus, some weaker oracles can be assumed instead. We further comment on this assumption in the next section. \subsection{Motivation} \label{sec:motivation} To sketch potential practical scenarios of using graph queries we mention a set of examples given in \cite{Emamjomeh-ZadehK17}. These examples are anchored in the field of machine learning, and since they have the same flavor with respect as how graphs are used, we refer to one of them. Consider a situation in which a \emph{system} wants to learn a clustering by asking queries. Each query presents a potential clustering to a user and if this is not the target clustering, then as a response the user either points two clusters that should be merged or points one cluster that should be split (but does not say how to split it). Thus, the goal is to construct a query algorithm to be used by the system. It turns out that learning the clustering can be done by asking queries on a graph: each vertex $v$ corresponds to a clustering and a reply of the user for $v$ will be aligned with one of the edges incident to $v$. In other words, the reply can be associated with an edge outgoing from $v$ that lies on a shortest path to the desired target clustering. We emphasize some properties of this approach. First, the fact that the reply indeed reveals the shortest path to the target is an important property of the underlying graph used by the algorithm and thus the graph needs to be carefully defined to satisfy it. Second, the user is not aware of the fact that such a graph-theoretic approach is used, as only a series of proposed clustering is presented. Third, this approach is resilient to errors on the user side: the graph query algorithms easily handle the facts that some replies can be incorrect (the user may make a mistake, or may not be willing to reveal the truth). It has been shown \cite{Emamjomeh-ZadehK17} that in a similar way one can approach the problems of learning a classifier or learning a ranking. From the standpoint of complexity we can approach such scenarios in two ways. First, one can derive an algorithm that specifically targets a particular application. More precisely, if one considers one of the above applications, then it may turn out that e.g. it is not necessary to construct the entire graph but instead reconstruct only what is necessary to perform each query. The second way is the general approach taken in this work: to consider the underlying graph as an abstract data structure out of the context of how it is used in particular applications. We note that examples like the ones mentioned above reveal that some applications may be burdened by the fact that the underlying graph is large, in which case the computational complexity, or local search procedures may be more crucial. We finally comment on our assumption that a shortest path oracle is provided to the algorithm. In the machine learning applications \cite{Emamjomeh-ZadehK17}, the graphs may be constructed in such a way that knowing which objects represent two vertices is sufficient to conclude the distance between them, i.e., a low-complexity distance oracle can be indeed implemented. This can be seen as a special case of a general approach to achieve distance oracles in practice through the so called distance-labeling schemes (c.f. Gavoille et al. \cite{DBLP:journals/jal/GavoillePPR04} and for practical approaches, c.f. Abraham et al. and Kosowski and Viennot \cite{AbrahamDFGW16,KosowskiV17}). We finally note that having the exact distances between vertices is crucial for this problem: if the distance oracle is allowed to provide even just a $1$-additive approximation of the exact distance, then each query algorithm needs to perform $\Omega(n)$ queries for some graphs c.f. Deligkas et~al. \cite{DeligkasMS19}. We note that the distance oracle access can be replaced with a multi-source distance computation (e.g. using BFS), at the cost of replacing some of the $\mathcal{O}(n)$ factors in the cost functions with $\mathcal{O}(m)$. Alternatively, a popular assumption borrowed from computational geometry is that we operate on a metric space with a metric (distance) function given. \subsection{Our Results and Techniques} \label{sec:our-results} For a query on a vertex $q$ with a reply $v$, we say that a vertex $u$ is \emph{consistent} with the reply if $q=v=u$, or $q\neq v$ but $v$ lies on a shortest path between $u$ and $q$; the set of all such consistent vertices $u$ is denoted by $N(q,v)$. Our method is based on a multiplicative weight update (MWU): the algorithm keeps the weights $\omega(v)$ for all vertices $v$, starting with a uniform assignment. The weight is representing the likelihood that a vertex is the target, although we point out that formally this is not a probability distribution. In MWU, the weight of each vertex that is not consistent with a reply is divided by an appropriately fixed constant $\Gamma$ that depends on $\varepsilon=\frac{1}{2}-p$. To keep the query complexity low, it is required that the queried vertex $q$ fulfills a measure of `centrality' in a graph in the sense that a query to such a central vertex results in an adequate decrease in the total weight. This is a graph-theoretic analogue of the `central' element comparison in the classical binary search. Two functions that have been used in the literature \cite{DeligkasMS19,DereniowskiTUW19,Emamjomeh-ZadehK17} to formalize this are \[\Phi(v) = \sum_{u \in V} d(u,v) \cdot \omega(u),\qquad\qquad\text{and}\qquad\qquad\Lambda(v) = \max_{u \in N(v)} \omega(N(v,u)),\] where $N(v)$ is the set of neighbors of $v$ in the graph, and $d(u,v)$ is the distance between $u$ and $v$. For brevity, $\omega(S)=\sum_{u\in S}\omega(u)$ for any $S\subseteq V$, and $\omega=\omega(V)$. \begin{definition} A vertex $q = \arg \min_{v \in V} \Phi(v)$ is called a \emph{median}. \end{definition} We note a fundamental bisection property of a median: \begin{lemma}[c.f. \cite{Emamjomeh-Zadeh:2015aa} section 2] \label{lem:folklore} If $q$ is a median, then $\Lambda(q) \le \omega(V)/2$. \end{lemma} Such property is key for building efficient binary-search algorithms in graphs, see \cite{DereniowskiTUW19,Emamjomeh-Zadeh:2015aa}: e.g., for the noiseless case, repeatedly querying a median of $X$, where $X \subseteq V$ is the subset of vertices that still can be a target, results in a strategy guaranteeing at most $\log_2 n$ queries. A disadvantage of using median is that it is computationally costly to find. Moreover, using its multiplicative approximation, that is, through a function $\Phi'$ such that $\Phi'(q) = (1 \pm \varepsilon')\Phi(q)$ for any constant $\varepsilon'>0$, blows up the strategy length exponentially \cite{DeligkasMS19} and thus this approach is not suitable. On the other hand, approximating $\Lambda$-minimizer is feasible, as noted also by \cite{DeligkasMS19}. Hence, we work towards a method of efficient median approximation through $\Lambda$ minimization. We believe that this algorithmic approach is of independent interest and can be used in different graph-theoretic problems. Interestingly, it turns out that we do not even need a multiplicative approximation of a $\Lambda$-minimizer but we only need that $\Lambda(q)$ is at most roughly half of the total weight. This is potentially usable in algorithms using generally understood graph bisection. (For an example of using such balanced separators for somewhat related search with persistent errors see e.g. Boczkowski et~al. \cite{BoczkowskiKR18}.) Formally, motivated by Lemma~\ref{lem:folklore}, we relax the notion of the median to the following. \begin{definition} \label{def:Lambda} We say that a vertex $q^*$ is \emph{$\delta$-close} to a median, for some $\delta>0$, when $\Lambda(q^*) \leq \left( \frac{1}{2}+\delta \right) \cdot \omega.$ \end{definition} To work-around the fact that $\Lambda$ is not efficient from the algorithmic standpoint, we introduce the following relaxation of $\Phi$: \[\Phi^{*}(q)=\sum_{v\in S}d(q,v),\] where $S$ is a random sample of vertices with probability distribution proportional to $\omega$. We can now formulate our main contribution in terms of new algorithmic tools: \begin{description} \item[Median approximation.] The relaxation of $\Phi$ to $\Phi^{*}$ provides, with high probability, a sufficient approximation of the median vertex in a graph. \end{description} We formalize this statement in the following way. Consider a sample size $s=\frac{8\ln n}{\delta^2}$, where $n$ is the number of vertices of the graph. This allows us to say how to approximate the median efficiently through a local condition: \begin{theorem} \label{lem:local-minimum} Let $q$ be a vertex such that for each $v\in N(q)$ it holds $\Phi^{*}(q) \leq \Phi^{*}(v) + \delta s$. Then, with high probability at least $1-n^{-3}$, the vertex $q$ is $\delta$-close to a median. \end{theorem} As a consequence, we obtain: \begin{corollary} \label{cor:median-equivalence} Let $q^* = \arg \min_{v \in V} \Phi^{*}(v)$. Then, the vertex $q^*$ is $\delta$-close to a median with high probability at least $1-n^{-3}$. \end{corollary} Returning to our search problem, these are enough to both find the right query vertex in each step, keep the strategy length low, and have a centrality measure that is efficient in terms of computational complexity. This leads us to the following theorem that is based on MWU with some appropriately fixed scaling factor $\Gamma$. \begin{restatable}{theorem}{complexityA} \label{thm:complexityA} Let $p = \frac{1}{2} - \varepsilon$ be the noise parameter for some $0 < \varepsilon \le \frac12$. There exists an adaptive query algorithm that after asking $\tau= \mathcal{O}(\frac{\log n}{\varepsilon^2})$ queries returns the target correctly with high probability. The computational complexity of the algorithm is $\mathcal{O}(\frac{n \log n}{\varepsilon})$ per query. \end{restatable} The algorithm behind the theorem iterates over the entire vertex set to find a $\Phi^{*}$-minimizer. We can refine this algorithm for graphs of low maximum degree $\Delta$ and diameter $D$. For that we use a local search whose direct application requires `visiting' $D\Delta$ vertices to get to a $\Phi^{*}$-minimizer. However, we introduce two ideas to speed it up. The first one is adding another approximation layer on top of $\Phi^{*}$: it is not necessary to find the exact $\Phi^{*}$-minimizer but its approximation, which we do as follows. Whenever the local search moves from one vertex $u$ to its neighbor $v$ and the improvement from $\Phi^{*}(u)$ to $\Phi^{*}(v)$ is sufficiently small, then $v$ will do for the next query. The second one is to start the local search from the vertex queried in the previous step. These two ideas combined lead to the second main result. \begin{restatable}{theorem}{complexityB} \label{thm:complexityB} Let $p = \frac{1}{2} - \varepsilon$ for some $0 < \varepsilon \le \frac12$. There exists an adaptive query algorithm that after asking $\tau= \mathcal{O}(\frac{\log n}{\varepsilon^2})$ queries returns the target correctly with high probability. The average computational complexity per query is $\mathcal{O}(n+D \Delta \frac{\log n}{\varepsilon^2})$ for graphs with diameter $D$ and maximum degree $\Delta$. \end{restatable} \subsection{Related Work} \label{sec:related-work} Median computation is one of the fundamental ways of finding central vertices of the graph, with huge impact on practical research \cite{Bavelas_1950,Beauchamp_1965,freeman1978centrality,hakimi1964optimum,Sabidussi_1966,tansel1983state}. A significant amount of research has been devoted to efficient algorithms for finding medians of networks \cite{Ostresh_1978,DBLP:conf/dis/TabataNK12,DBLP:journals/ieicet/TabataNK17} or approximating the notion \cite{DBLP:journals/siamjo/CantoneCFP05,DBLP:journals/tcs/Chang12}. We note the seminal work of Indyk~\cite{DBLP:conf/stoc/Indyk99} which includes $1+\varepsilon$ approximation to $1$-median in time $\mathcal{O}(n/\varepsilon^5)$ in metric spaces -- we note that the form of approximation there differs from ours, although the very-high level technique of using random sampling is common. Chechik et~al. in \cite{DBLP:conf/approx/ChechikCK15} use (non-uniform) random sampling to answer queries on sum of distances to the queried vertices in graphs. We also refer the reader to some recent work on the median computation in median graphs, see Beneteau~et~al. \cite{BeneteauCCV19} and references therein. More related centrality measures of a graph are discussed in \cite{AbboudGW15,DBLP:conf/soda/AbboudWW16,DBLP:conf/soda/Cabello17} in the context of fine-grained complexity, showing e.g. that efficient computation of a median vertex (in edge-weighted graphs) is equivalent under subcubic reductions to computation of All-Pairs Shortest Paths. Substantial amount of research has been done on searching in sorted data (i.e., paths), which included investigations for fixed number of errors \cite{Aigner96,RivestMKWS80}, optimal strategies for arbitrary number of errors and various error models, including linearly bounded \cite{DhagatGW92}, prefix-bounded \cite{BorgstromK93} and noisy/probabilistic \cite{Ben-OrH08,KarpK07}. Also, a lot of research has been done on how different types of queries influence the search process --- see \cite{DaganFGM17} for a recent work and references therein. The mostly studied comparison queries for paths have been extended to graphs in two ways. First one is a generalization to partial orders \cite{Ben-AsherFN99,LamY01}, although this does not further generalize well for arbitrary graphs \cite{Dereniowski08}. It is worth noting that a lot of work has been devoted to the computational complexity of finding error-less strategies \cite{DereniowskiKUZ17,LamY01,MozesOW08}. The second extension is by using the vertex queries studied in this work, for which much less is known in terms of complexity. It is worth to mention that the problem becomes equivalent to the vertex ranking problem for trees \cite{Schaffer89}, but not for general graphs (see also \cite{OnakP06}). Similarly as in the case of the classical binary search, the graph structure guarantees that there always exists a vertex that adequately partitions the search space in the absence of errors \cite{Emamjomeh-Zadeh:2015aa}. The problem becomes much more challenging as this is no longer the case when errors are present. A centrality measure that works well for finding the right vertex to be queried is a median used in \cite{DereniowskiTUW19,Emamjomeh-Zadeh:2015aa}. However, as shown in \cite{DeligkasMS19}, the median is sensitive to approximations in the following way. When the algorithm decides to query a $(1+\varepsilon')$-approximation $v$ of the median (minimizer of $\Phi'$ which is $1+\varepsilon'$ approximation of $\Phi$), then some graphs require $\mathcal{O}(\sqrt{n})$ queries, where the approximation is understood as $\Phi(v)\leq(1+\varepsilon')\min_{u\in V}\Phi(u)$. This results holds for the error-less case. Furthermore, the authors introduce in~\cite{DeligkasMS19} the potential $\Lambda$ (denoted by $\varGamma$ therein) and prove, also for the error-less case, that it guarantees $\frac{\log_2 n}{1-\log_2(1+\varepsilon)} \approx (1+\varepsilon) \log_2n$ queries, when in each step a $(1+\varepsilon)$-approximation of the $\Lambda$-minimizer is queried. However, this issue has been considered from a theoretical perspective and no optimization considerations have been made. In particular, it was left open as to how to reduce the query complexity at an expense of working with such approximations. This, and the consideration of the noise are two our main improvements with respect to \cite{DeligkasMS19}. We also stress out that our definition of $\delta$-closeness to a median differs from $(1+\varepsilon)$-approximations in the sense that our definition is much less strict: a vertex $q^*$ that is $\delta$-close to a median may have the property that $\Lambda(q^*)$ significantly deviates from $\min_{u\in V}\Lambda(u)$. Some complexity considerations have been touched in \cite{Emamjomeh-ZadehK17}, from the perspective of targeting specific machine learning applications, where already the above-mentioned $\Lambda$-minimizer has been used. To make the statements form that work comparable to our results, we have two distinguish two input size measures that apply. In \cite{Emamjomeh-ZadehK17}, for a particular application an input consists of a specific machine learning instance, and denote its size by $\tilde{n}$. In order to find a solution for this instance, a graph $G$ of size $n$ is constructed and an adaptive query algorithm is being run on this graph. It is assumed that $\log_2n$ is polynomial in $\tilde{n}$. The diameter $D$ and maximum degree $\Delta$ of $G$ are both assumed in \cite{Emamjomeh-ZadehK17} to be polylogarithmic in $\tilde{n}$. A local search is used to find a vertex that approximates the $\Lambda$-minimizer. For that, in each step a sampling is used for the approximation purposes: for each vertex $v$ along the local search, all its neighbors $u$ are tested for finding an approximation $\Lambda$, giving the complexity of $\mathcal{O}(D\Delta)=\mathcal{O}(m)$, where $m$ is the number of edges of $G$. It is concluded that the overall complexity of performing a single query is $\mathcal{O}(D\Delta \textrm{poly}(\log n,\frac{1}{\varepsilon}))=\mathcal{O}(m \cdot \textrm{poly}(\log n,\frac{1}{\varepsilon}))$. \subsection{Outline} We proceed in the paper as follows. Section~\ref{sec:linearly-bounded} provides a `template' strategy in which we simply query a vertex that is $\delta$-close to a median. The strategy length is there fixed carefully to meet the tail bounds on the error probability. Then, in Section~\ref{sec:guarantee}, we prove that our sample size is enough to ensure high success probability. Section~\ref{sec:sampling-complexity} observes that the overall complexity of the algorithm can be reduced by avoiding recasting the entire sample in each step: it is enough to replace only a small fraction of the current sample when going from one step of the strategy to the next. We then combine these tools to prove our main theorems in Section~\ref{sec:proof}, where for Theorem~\ref{thm:complexityB} we additionally make several observations on speeding-up the classical local search in a graph. \section{The Generic Strategy} \label{sec:linearly-bounded} As an intermediate convenient step, we recall the following adversarial error model: given a constant $r$, if the strategy length is $\tau$, then it is guaranteed that at most $r\cdot\tau$ errors occurred throughout the search (their distribution may be arbitrary). We set our parameters as follows: let $\eta = \varepsilon/2$, $r = \frac{1}{2} - \eta$, and assume without loss of generality that $\eta < 1/8$. Let $\delta = \eta/4$. With these parameters, we provide Algorithm \textsc{\ref{alg:linearly-bounded-search}} that runs the multiplicative weight update with $\Gamma=\frac{1}{1-4\eta}$ for $\tau=\frac{10\log_2n}{\eta^2}$ steps. Then we prove (cf. Lemma~\ref{lem:linearly-bounded}) that this strategy length is sufficient for correct target detection in this error model. We write $\omega_t$ to denote the vertex weight in a step $t$. (So, $\omega_0$ is the initial uniform weight assignment.) \begin{figure}[htb] \begin{center} \begin{minipage}{.8\linewidth} \begin{algorithm}[H] \SetAlgoRefName{LB-Search} \caption{Always query a $\delta$-close vertex to a median.} \label{alg:linearly-bounded-search} $\omega(v) \gets \frac{1}{n}$ and $\liecount{v}\gets 0$ for each $v\in V$\; \For{$\tau=10 \frac{\log_2 n}{\eta^2}$ steps} { Let $q$ be any vertex that is $\delta$-close to a median\label{ln:almost-median}\; Query the vertex $q$\; \For{each vertex $u$ not compatible with the answer} { $\omega(u) \gets \omega(u) / \Gamma$, where $\Gamma=\frac{1}{1-4\eta}$\; $\liecount{u} \gets \liecount{u}+1$ } } \Return the vertex $v$ with the smallest $\liecount{v}$ \end{algorithm} \end{minipage} \end{center} \end{figure} \begin{lemma} \label{lem:linearly-bounded} If during the execution of Algorithm \textsc{\ref{alg:linearly-bounded-search}} over total $\tau$ queries there were at most $r\cdot \tau$ errors, then the algorithm outputs the target. \end{lemma} \begin{proof} If a vertex $v$ at step $t$ satisfies $\omega_t(v) > (\frac{1}{2} + \delta) \omega_t$, then we say that $v$ is \emph{heavy} at step $t$. We aim at proving that the overall weight decreases multiplicatively either by at least $(1-\eta)^2$ or $\frac{\Gamma+1}{2\Gamma}$ per step. In the absence of a heavy vertex we get the first bound, and it is an immediate consequence of the Equation~\eqref{eq:potdroperror} below. If we get a heavy vertex at some point, none of these bounds may be true in this particular step (this phenomenon is inherent to the graph query model itself) but we show below that the second one holds in an amortized way (cf. Lemma~\ref{lem:heavy}). If at step $t$ there is no heavy vertex, then \begin{equation} \label{eq:potdroperror} \omega_{t+1} \le \left(\frac12+\delta + \frac{\frac12 - \delta}{\Gamma}\right) \omega_{t} = \left( 1 - 2 \eta + 4 \eta \delta\right)\omega_t = (1-\eta)^2 \omega_t. \end{equation} Assume otherwise that there is vertex $v$ that is heavy at step $t$. \begin{lemma} \label{lem:almost-median-heavy} If at any step $t$ there is a heavy vertex $v$, then $v$ is the only $\delta$-close to a median vertex at this step. \end{lemma} \begin{proof} For any $u\neq v$, we have that $\Lambda(u)\geq\omega_t(v)>(\frac{1}{2}+\delta)\omega_t$, i.e., $u$ is not $\delta$-close to a median. On the other hand, $\Lambda(v)\leq\omega_t(V\setminus\{v\})<(\frac{1}{2}-\delta)\omega_t$, i.e., $v$ is $\delta$-close to a median. \end{proof} The above lemma implies that if some $v$ is heavy then it will be queried in this particular step. The next lemma calculates the overall potential drop in a series of steps in which some vertex is heavy. \begin{lemma} \label{lem:heavy} Consider the maximal consecutive segment of steps $\mathcal{I}$ where some $q$ is heavy. That is, we pick $t_1,t_2$ such that $q$ is heavy in all steps $t \in \mathcal{I}=\{t_1, \ldots, t_2-1\}$ and is not heavy in steps $t_1-1$ and $t_2$. Then, $\omega_{t_2} \le \left(\frac{\Gamma+1}{2\Gamma}\right)^{t_2-t_1} \omega_{t_1}.$ \end{lemma} \begin{proof} First note that, by Lemma~\ref{lem:almost-median-heavy}, $q$ is queried in each step in $\mathcal{I}$. For a query on $q$, we say that a reply $v$ is a \emph{yes-answer} if $v=q$, and otherwise it is a \emph{no-answer}. Denote by $a$ and $b$ the number of yes- and no-answers in $\mathcal{I}$, respectively. Note that $a+b = t_2-t_1$. Moreover, \begin{align*} \omega_{t_2}(q) &= \left(\frac{1}{\Gamma}\right)^b \omega_{t_1}(q), &\text{and}&\qquad &\omega_{t_2}(V\setminus\{q\})\le \left(\frac{1}{\Gamma}\right)^a \omega_{t_1}(V\setminus\{q\}). \end{align*} The vertex $q$ being heavy at $t_1$ implies $\frac{\omega_{t_1}(q)}{\omega_{t_1}(V\setminus\{q\})} > \frac{1/2+\delta}{1/2 - \delta}$ and similarly $q$ not being heavy at $t_2$ implies $\frac{\omega_{t_2}(q)}{\omega_{t_2}(V\setminus\{q\})} \le \frac{1/2+\delta}{1/2 - \delta}$. Combining the equality and three inequalities above, we obtain $b \ge a$. We assume without loss of generality that all the yes-answers were given before all the no-answers in the range $\mathcal{I}$. Indeed, we observe that rearranging these answers does not change the state of the algorithm at step $t_2$, and $q$ remains heavy for all of the $\mathcal{I}$. We have then, for the all of the $a$ yes-answers and first $a$ no-answers, the following: \begin{equation} \label{eq:fist-2a-steps} \omega_{t_1+2a} \le \left(\frac{1}{\Gamma}\right)^a \omega_{t_1} = \left(\frac{1}{\sqrt{\Gamma}}\right)^{2a} \omega_{t_1} \le \left(\frac{\Gamma+1}{2\Gamma}\right)^{2a} \omega_{t_1}, \end{equation} where the first inequality is due to the fact that each of the $a$ pairs (a pair understood as a no-answer and a yes-answer) scales down each vertex by at least a factor of $\Gamma$, while in the last inequality we have used $1/\sqrt{\Gamma} \le (\Gamma+1)/(2\Gamma)$. For the remaining $b-a$ steps of $\mathcal{I}$, the weight of $q$ decreases by a factor of $\Gamma$. Thus, for each $t\in\{t_1+2a,\ldots,t_2-1\}$, using that $q$ is heavy in step $t$: \[\omega_{t+1} \leq \omega_t(V\setminus\{q\})+\frac{\omega_t(q)}{\Gamma} \leq \omega_t\left(\frac{1}{2}-\delta\right)+ \frac{\omega_t}{\Gamma} \left(\frac{1}{2} + \delta\right) \leq\frac{\omega_t}{2}+\frac{\omega_t}{2\Gamma}=\frac{\Gamma+1}{2\Gamma}\cdot\omega_t.\] Thus, $\omega_{t_2} \le \left(\frac{\Gamma+1}{2\Gamma}\right)^{b-a} \omega_{t_1+2a},$ which together with \eqref{eq:fist-2a-steps} completes the proof of Lemma~\ref{lem:heavy}. \end{proof} Let $q$ be the target, and $u$ be the output of Algorithm \textsc{\ref{alg:linearly-bounded-search}}. Assume w.l.o.g. that the algorithm run for $\tau' \ge \tau$ steps. Since \[\tau' \ge 10 \frac{\log_2 n}{ \eta^2} \ge \frac{\log_2 n}{r\log_2(1-4\eta) - 2\log_2(1-\eta) },\] where the inequality follows from $(1/2-\eta) \cdot \log_2(1-4\eta) - 2\log_2(1-\eta) \ge \frac{1}{10} \eta^2$ when $0 \le \eta \le \frac18$, we obtain a bound \begin{equation} \label{eq:eta-bound} (1-4\eta)^{r\tau'} \ge (1-\eta)^{2\tau'} \cdot n. \end{equation} We assume that the algorithm outputs an incorrect vertex $u$, and show that it leads to a contradiction. We consider the state of the weights after $\tau'$ steps. We consider two cases. \begin{enumerate} \item There is no heavy vertex after $\tau'$ steps. We observe that the starting weight satisfies $\omega_0 = 1$, and by the bound on the number of errors accumulated on target vertex $v^{*}$ (it cannot be more than $r\tau'$), we have $\omega_{\tau'} \ge \omega_{\tau'}(v^{*})+\omega_{\tau'}(u) > \frac{1}{n} \left( \frac{1}{\Gamma} \right)^{r\tau'}.$ By Equation~\eqref{eq:potdroperror} and Lemma~\ref{lem:heavy}, we know that every step contributed at least a factor $(1-\eta)^2$ or $(\Gamma+1)/(2\Gamma) = (1-2\eta)$ multiplicatively to the total weight. Thus, by~\eqref{eq:eta-bound}, $\omega_{\tau'} \le (1-\eta)^{2\tau'} \omega_0 \le \frac{1}{n} (1-4\eta)^{r\tau'} = \frac{1}{n} \left( \frac{1}{\Gamma} \right)^{r\tau'},$ which leads to a contradiction. \item Returned vertex $u$ is heavy after $\tau'$ steps. We append at the end of the strategy a virtual sequence of $k$ identical query-answers: algorithm queries $u$, and receives an no-answer pointing towards $q$. Here, $k$ is chosen to be minimal such that after $\tau'+k$ steps $u$ is no longer heavy (it exists, since each such query increases $\ell_u$ by 1, and leaves $\ell_q$ unchanged). However, at the end of $\tau'+k$ round $\ell_u$ is minimal (possibly not necessarily uniquely minimal). We note that appending those $k$ steps did not increase the total number of errors from the answerer, and all of the queries were asked to a heavy vertex $u$. This reduces this case to the previous one, with increased value of $\tau'$. \qedhere \end{enumerate} \end{proof} We now transit from the adversarial search to the noisy setting. This is done by using Algorithm~\textsc{\ref{alg:linearly-bounded-search}} as a black box with $\eta$ being fixed appropriately. Recall that $p = \frac{1}{2} - \varepsilon$, and we will use the following dependence of $\eta$ on $\varepsilon$ (note that by taking $\eta$ smaller than $\varepsilon$ we accommodate the necessary tail bound in the lemma below, i.e., we ensure that the event of having more than $r\tau$ errors is sufficiently unlikely). \begin{lemma} \label{lem:noise-success-probability} Run Algorithm~\textsc{\ref{alg:linearly-bounded-search}} with $r = \frac12 - \eta$, where $\eta = \varepsilon/2$. If an answer to each query was erroneous with probability at most $p$, independently, then the algorithm outputs the target vertex with a high probability of at least $1-n^{-3}$. \end{lemma} \begin{proof} Recall $\tau = 10\frac{\log_2 n}{\eta^2}$ in Algorithm~\textsc{\ref{alg:linearly-bounded-search}}. Denote by $L$ the overall number of errors that have occurred during the execution of the algorithm. The expected number of errors is $p \cdot \tau$. By the Hoeffding inequality, $$\text{Pr}[L \ge r \cdot \tau ] \le \exp\left(-2 \tau (r-p)^2 \right) = \exp( -20 \log_2 n) \le n^{-3}.$$ Thus with high probability number of errors is bounded so that we can apply Lemma~\ref{lem:linearly-bounded} (which in itself gives a deterministic guarantee). \end{proof} \section{Sampling Guarantees} \label{sec:guarantee} To take the `random sampling' counterparts of $\Phi$ and $\Lambda$, consider a $S = \{m_1,\ldots,m_s\}$ to be a multiset of $s$ vertices sampled from $V$ with repetitions, with sampling probabilities $p(v) \sim \omega(v)$. That is, for each $m_i$, we have $\Pr(m_i = v) = \frac{\omega(v)}{\omega}$ and choices made for $m_i$ are fully independent. To such an $S$ we refer as a \emph{random sample}. We then define the following potentials $$\Phi^{*}(v) = \sum_{u \in S} d(u,v)\qquad\qquad\text{and}\qquad\qquad\Lambda^{*}(v) = \max_{u \in N(v)} |S \cap N(u,v)| ,$$ where the intersection of a multiset $S$ with some set $X \subset V$ is defined as a multiset $S \cap X= \{ m_i \colon i \in \{ 1 ,\ldots, s \} \wedge m_i \in X \}$. We note a specific detail regarding these functions -- we will prove and use the fact that in order to find a vertex that is $\delta$-close to a median (a vertex we need to query), it is enough to pick an approximation of the $\Phi^{*}$-minimizer. This is slightly counterintuitive, since $\delta$-closeness is defined in terms of $\Lambda$ which has a similar meaning to $\Lambda^{*}$. However, the subtlety here is due to a complexity issue --- it is easier to recompute the $\Phi^{*}$ upon updating the sample $S$. We denote $s = |S|$ and assume in the rest of the paper that $s = \frac{8\ln n}{\delta^2}$. In this section we prove that this choice of $s$ is sufficient, and then Section~\ref{sec:sampling-complexity} deals with the complexity issues of the sampling method. The following is shown in the appendix: \begin{restatable}{lemma-restate}{samplinglambda} \label{lem:sampling-lambda} For any $v$, there is $\frac{\Lambda(v)}{\omega} \le \frac{\Lambda^{*}(v)}{s}+\delta/2$ with a high probability at least $1 - n^{-3}$. \end{restatable} \begin{proof} Consider any neighbor $u$ of $v$. Denote by $X_i$ the indicator variable that $m_i \in N(v,u)$. Observe that $|S \cap N(v,u)| = \sum_i X_i$ and $\mathbb{E}[X_i] = \frac{\omega(N(v,u))}{\omega}$ and so $\mathbb{E}[ \sum_i X_i ] = s \cdot \frac{\omega(N(v,u))}{\omega}$. By a standard application of Hoeffding bound there is \[\Pr\left(\mathbb{E}[\sum X_i] - (\sum_i X_i) \ge \frac{s\delta}2\right) \le e^{-2s(\delta/2)^2} = \frac{1}{n^4}.\] So \[\frac{\omega(N(v,u))}{\omega} \le \frac{|S \cap N(v,u)|}{s} + \delta/2\] holds with probability at least $1 - n^{-4}$. Taking union bound over at most $n$ neighbors $v$, we have that with probability at least $1 - n^{-3}$ the following hold \begin{align*} \Lambda(v) &= \max_{u \in N(v)} \omega(N(v,u))\\ &\le \max_{u \in N(v)} \left(\frac{|S \cap N(v,u)|}{s}+\delta/2 \right) \cdot \omega\\ &= \left(\frac{\Lambda^{*}(v)}{s}+\delta/2\right) \cdot \omega. \end{align*} \end{proof} \begin{lemma} \label{lem:potdrop} Let $q$ be a vertex such that $\forall_{v \in N(q)} \Phi^{*}(q) \le \Phi^{*}(v) + \delta s$. Then, $\Lambda^{*}(q) \le \frac{s(1 + \delta)}{2}$. \end{lemma} \begin{proof} To see that suppose, towards a contradiction, that $\Lambda^{*}(q) > \frac{s(1 + \delta)}{2}$, i.e. there is $v \in N(q)$, such that $|S \cap N(q, v)| > \frac{s(1 + \delta)}{2}$. Denote $A^-=S\cap N(q,v)$ and $A^+=S\setminus A^-$. Using $\sum_{u\in A^-}d(v,u) = \sum_{u\in A^-}d(q,u) - |A^-|$ and $\sum_{u\in A^+}d(v,u) \leq \sum_{u\in A^+}d(q,u) + |A^+|$, we get \[\Phi^{*}(v) \leq |A^+| - |A^-| + \sum_{u\inS}d(q,u) < \frac{s(1 - \delta)}{2} - \frac{s(1 + \delta)}{2} + \Phi^{*}(q) = \Phi^{*}(q) - \delta s \leq \Phi^{*}(v),\] which yields a contradiction. \end{proof} Hence, we can prove Theorem~\ref{lem:local-minimum}: Combining Lemma~\ref{lem:potdrop} and Lemma~\ref{lem:sampling-lambda}, $$\Lambda(q) \le \left(\frac{\Lambda^{*}(q)}{s}+\delta/2\right) \cdot \omega \le \left(\frac12+\delta/2 + \delta/2\right) \cdot \omega$$ with probability at least $1-n^{-3}$. \section{Maintaining the Sample} \label{sec:sampling-complexity} We now discuss the complexity of maintaining the sample $S$ upon the vertex weight updates. Given a sample set $S_t$ at step $t$, the next sample $S_{t+1}$ is computed by a call to Algorithm~\textsc{\ref{alg:resampling}} below. \begin{figure}[htb] \begin{center} \begin{minipage}{.8\linewidth} \begin{algorithm}[H] \SetAlgoRefName{Resampling} \caption{Update of the sample after step $t$.} \label{alg:resampling} \ForEach{$x_t\inS_t$} { \If{$x_t$ is consistent with the reply in step $t$} { $x_{t+1} \gets x_t$ } \Else { \If{with probability $1/\Gamma$} { let $x_{t+1} \gets x_t$ } \Else { $x_{t+1}$ is drawn randomly from $V$ with distribution proportional to the weights $\omega_{t+1}$ } } Insert $x_{t+1}$ to $S_{t+1}$ } \end{algorithm} \end{minipage} \end{center} \end{figure} The correctness of Algorithm~\textsc{\ref{alg:resampling}} is given by Lemma~\ref{lem:resampling}. Its proof follows the cases in the pseudo-code to show that both the vertices that remain in the sample and the new ones meet the probability requirements for a random sample. \begin{restatable}{lemma-restate}{resampling} \label{lem:resampling} Suppose that in Algorithm~\textsc{\ref{alg:linearly-bounded-search}}, after each weight update the current random sample $S$ is recalculated by a call to Algorithm \textsc{\ref{alg:resampling}}. Then, with high probability at least $1 - n^{-3}$, at most $2\varepsilon |S|$ resampling operations occur at each step. \end{restatable} \begin{proof} Recall that after querying a vertex $q$ at step $t$ and receiving an answer $v$, the weights are updated as follows. For each $u \in V$: \begin{itemize} \item if $u \in N(q,v)$, then $\omega_{t+1}(u) \gets \omega_t(u)$, \item if $u \not\in N(q,v)$, then $\omega_{t+1}(u) \gets \omega_t(u)/\Gamma$, \end{itemize} where we recall that $\Gamma = \frac{1}{1-4\eta}$. Consider a vertex $x_t$. Assume the for every $u \in V$, $\Pr(x_t = u) = \frac{\omega_t(u)}{\omega_t}$. We have two cases: \begin{enumerate} \item If $u \in N(q,v)$, then \begin{align*}\Pr(x_{t+1} = u) &= \Pr(x_t = u) + \Pr(x_t \not\in N(q,v)) \cdot \left(1 - \frac{1}{\Gamma}\right) \cdot \Pr(x_{t+1} \text{ is sampled as } u)\\ &= \frac{\omega_t(u)}{\omega_t} + \frac{\omega_t(V \setminus N(q,v))}{\omega_t} \left(1 - \frac{1}{\Gamma}\right) \frac{\omega_{t+1}(u)}{\omega_{t+1}}\\ &= \frac{\omega_{t+1}(u)}{\omega_t} \left(1+ \frac{\omega_t(V \setminus N(q,v))}{\omega_{t+1}} \left(1 - \frac{1}{\Gamma}\right)\right)\\ & = \frac{\omega_{t+1}(u)}{\omega_t} \cdot \frac{\omega_{t+1} + \omega_t(V \setminus N(q,v))\left(1 - \frac{1}{\Gamma}\right)}{\omega_{t+1}}\\ &= \frac{\omega_{t+1}(u)}{\omega_t} \cdot \frac{\omega_{t}(N(q,v)) + \omega_t(V \setminus N(q,v))\frac{1}{\Gamma}+ \omega_t(V \setminus N(q,v))\left(1 - \frac{1}{\Gamma}\right)}{\omega_{t+1}}\\ &= \frac{\omega_{t+1}(u)}{\omega_t} \cdot \frac{\omega_t}{\omega_{t+1}} = \frac{\omega_{t+1}(u)}{\omega_{t+1}}. \end{align*} \item Otherwise, if $u \not\in N(q,v)$, then \begin{align*} \Pr(x_{t+1} = u) &= \Pr(x_t = u) \cdot \frac{1}{\Gamma} + \Pr(x_t \not\in N(q,v)) \cdot \left(1 - \frac{1}{\Gamma}\right) \cdot \Pr(x_{t+1} \text{ is sampled as } u)\\ &= \frac{\omega_t(u)}{\omega_t} \frac{1}{\Gamma} + \frac{\omega_t(V \setminus N(q,v))}{\omega_t} \left(1 - \frac{1}{\Gamma}\right) \frac{\omega_{t+1}(u)}{\omega_{t+1}}\\ &= \frac{\omega_{t+1}(u)}{\omega_t} \left(1+ \frac{\omega_t(V \setminus N(q,v))}{\omega_{t+1}} \left(1 - \frac{1}{\Gamma}\right)\right)\\ &= \frac{\omega_{t+1}(u)}{\omega_{t+1}}. \end{align*} \end{enumerate} This proves that probabilities for each sample are maintained between steps. We now bound the actual number of resampling operations necessary. Observe that each element of $S_t$ is re-sampled with probability at most $4 \eta = 2 \varepsilon$. Let $K$ denote number of re-sampled vertices. $\mathbb{E}[K] = s \cdot 2\varepsilon$, and then by Chernoff bound $\Pr[K > 4s\varepsilon] \le \exp(- 2s\varepsilon/3) = \exp(- \frac{1024 \ln n}{3\varepsilon}) \le n^{-3}.$ \end{proof} We comment on the computational complexity of sampling according to a distribution. \begin{observation} \label{obs:sampl} Sampling $K$ vertices according to distribution $\omega$ can be done in $\mathcal{O}(n + K \log K)$ operations. \end{observation} The time for sampling is $\mathcal{O}(K \log K)$ for generating sorted list of $K$ real-values picked uniformly at random from $[0,1]$, and $\mathcal{O}(n)$ for linear scan of all of the weights from $\omega$. \section{Proofs of the Main Theorems} \label{sec:proof} \complexityA* \begin{proof} First, assume without loss of generality that $\frac{\log n}{\varepsilon^2} < n^2$, as otherwise the claimed one-step complexity is $\varepsilon^{-1}n\log n=\Omega(n^2)$. This can be met by an algorithm that at each step queries a median vertex, see \cite{Emamjomeh-Zadeh:2015aa}. Run Algorithm \textsc{\ref{alg:linearly-bounded-search}} that performs $\tau=\mathcal{O}(\frac{\log n}{\varepsilon^2})$ queries by Lemma~\ref{lem:noise-success-probability}. The algorithm maintains a sample $S_t$ at each step $t$ by using Algorithm~\textsc{\ref{alg:resampling}}. By Corollary~\ref{cor:median-equivalence}, the probability that each step of the algorithm indeed uses a vertex that is $\delta$-close to a median is $1-n^{-3}$. After each query, the algorithm updates the weights in time $\mathcal{O}(n)$, and $\mathcal{O}(\varepsilon s)$ vertices are re-sampled by Lemma~\ref{lem:resampling}, for the cost of $\mathcal{O}(n + \varepsilon s \log n)$ which is subsumed by other terms. Thus the cost of maintaining the values of $\Phi^{*}$ is $\mathcal{O}(\varepsilon s)$ per vertex, or $\mathcal{O}(n \varepsilon s)$ in total, which is the dominant cost for the algorithm, with the update being performed as: \[\Phi^{*}(v) \gets \Phi^{*}(v) - \sum_{u \in S_{t+1}\setminus S_t} d(u,v) + \sum_{u \in S_{t}\setminus S_{t+1}} d(u,v).\] Taking a union bound over all steps, we obtain the high success probability $1 - \mathcal{O}(n^{-1})$. \end{proof} Now we turn out attention to the proof of Theorem~\ref{thm:complexityB}, where a local search is used. This is a natural approach that gives an improvement for low-degree low-diameter graphs. The two `twists' that we add are early termination (see the pseudo-code shown as Algorithm~\textsc{\ref{alg:local-search}}) and resuming from the vertex $v$ that is the output of the previous execution of the local search (which is used in the proof of Theorem~\ref{thm:complexityB}). The former allows us to directly bound the number of iterations; cf. Observation~\ref{obs:iterations}. \begin{figure}[htb] \begin{center} \begin{minipage}{.8\linewidth} \begin{algorithm}[H] \SetAlgoRefName{Local-Search} \caption{Find a local median starting from an input vertex $v$} \label{alg:local-search} \While{true} { $q = \arg \min_{u \in N(v)} \Phi^{*}(u)$\; \If{$\Phi^{*}(q) > \Phi^{*}(v) - \delta s$} { \Return{$v$} } \Else { $v = q$ } } \end{algorithm} \end{minipage} \end{center} \end{figure} \begin{observation} \label{obs:iterations} If Algorithm~\textsc{\ref{alg:local-search}} run with an input vertex $v$ returns a vertex $v'$, then the number of iterations is upper-bounded by $1+\frac{\Phi^{*}(v)-\Phi^{*}(v')}{\delta s}$. \end{observation} \complexityB* \begin{proof} First, w.l.o.g. assume that $\log n /\varepsilon^2 < n$, by the same reasoning as in the proof of Theorem~\ref{thm:complexityA}. By Lemma~\ref{lem:noise-success-probability}, Algorithm \textsc{\ref{alg:linearly-bounded-search}} that performs $\tau=\mathcal{O}(\frac{\log n}{\varepsilon^2})$ queries. We consider the following modification to Algorithm~\textsc{\ref{alg:linearly-bounded-search}}. As before, the algorithm updates weights in time $\mathcal{O}(n)$ and maintains a sample $S_t$ at each step $t$ (by using Algorithm~\textsc{\ref{alg:resampling}}) in time $\mathcal{O}(n + \varepsilon s \log n)$ which is subsumed by other terms. However, instead of choosing a vertex that is $\delta$-close to a median in line~\ref{ln:almost-median}, the updated algorithm runs Algorithm~\textsc{\ref{alg:local-search}} with the previously queried vertex as an input, and sets the output vertex to be the vertex $q$ to be queried. In other words, at each step $t$, it uses Algorithm~\textsc{\ref{alg:local-search}} with input $v_{t-1}$ which returns $v_t$, and queries $q=v_t$. The algorithm initializes $v_0$ arbitrarily. By Lemma~\ref{lem:local-minimum}, $v_t$ is $\delta$-close to a median. By Observation~\ref{obs:iterations}, we bound the total number of iterations $K$ done by Algorithm~\textsc{\ref{alg:local-search}} by \begin{align*} K &\le \sum_{t=0}^{\tau-1} \left(1 + \frac{\Phi^{*}_{t+1}(v_t) - \Phi^{*}_{t+1}(v_{t+1})}{\delta s} \right)\\ &= \tau + \frac{\Phi^{*}_1(v_0) + \sum_{t=1}^{\tau-1} (\Phi^{*}_{t+1}(v_t) -\Phi^{*}_t(v_t)) - \Phi^{*}_{\tau}(v_{\tau})}{\delta s}\\ &\le \tau + \frac{s D + 2\tau s \varepsilon D}{\delta s} = \mathcal{O}(D \tau), \end{align*} where we used that $\Phi^{*}_{t+1}(v_t) -\Phi^{*}_t(v_t) \le 2s\varepsilon D$ holds with high probability by Lemma~\ref{lem:resampling}. Each iteration in Algorithm~\textsc{\ref{alg:local-search}} has complexity $\mathcal{O}(\Delta s)$ making the total complexity of the algorithm to be $\mathcal{O}(\tau (n + D \Delta s))$. \end{proof} \section{Open Problems} \label{sec:summary} Having an algorithm that keeps an optimal query complexity and obtains a low computational complexity, one can ask what are the possible tradeoffs between the two? Another question is how much further the computational complexity can be decreased? Also, are there any possible lower bounds that can reveal the limits of what is not achievable in the context of these problems? Regarding the centrality measures we consider, we propose an efficient median approximation. Motivated by this, another question is what are other possible vertex-functions that may allow for further improvements, e.g. in the complexity? \clearpage

https://arxiv.org/abs/1404.1618

Some results on minimum skew zero forcing sets, and skew zero forcing number

Let $G$ be a graph, and $Z$ a subset of its vertices, which we color black, while the remaining are colored white. We define the skew color change rule as follows: if $u$ is a vertex of $G$, and exactly one of its neighbors $v$, is white, then change the color of $v$ to black. A set $Z$ is a skew zero forcing set for $G$ if the application of the skew color change rule (as many times as necessary) will result in all the vertices in $G$ colored black. A set $Z$ is a minimum skew zero forcing set for $G$ if it is a skew zero forcing set for $G$ of least cardinality. The skew zero forcing number $\sZ (G)$ is the minimum of $|Z|$ over all skew zero forcing sets $Z$ for $G$.In this paper we discuss graphs that have extreme skew zero forcing number. We characterize complete multipartite graphs in terms of $\sZ (G)$. We note relations between minimum skew zero forcing sets and matchings in some bipartite graphs, and in unicyclic graphs. We establish that the elements in the set of minimum skew zero forcing sets in certain bipartite graphs are the bases of a matroid.

\section{Introduction} A {\em graph} is a pair $G = (V_G, E_G)$, where $V_G$ is the (finite, nonempty) set of vertices of $G$ and $E_G$ is the set of edges, where an edge is a two-element subset of vertices. The {\em complete graph} on $n$ vertices is denoted $K_n$. An {\em induced subgraph} of $G$ is a subgraph obtained from $G$ by deleting a vertex $v$, or a number of vertices $S$, and we write $G-v$ or $G-S$, respectively. If $\{u, v \} \in E_G$ the vertices $u$ and $v$ are said to be {\em adjacent}, they are also said to be {\em neighbors}. The set $N(v)$, consisting of all the neighbors of $v$, is called the {\em open neighborhood of $v$} (it does not include $v$), the set $N[v] = N(v) \cup \{v \}$ is the {\em closed neighborhood of $v$}. The {\em degree of a vertex $v \in V_G$}, denoted by $\deg_G (v)$, is the number of edges adjacent to $v$. The minimum (respectively, maximum) degree in a graph $G$ is denoted $\delta (G)$ (respectively, $\Delta (G)$). A subset $S \subseteq V_G$ is called {\em independent} if no two vertices in $S$ are adjacent. A graph $G$ is {\em connected} if each pair of vertices in $V_G$ belongs to a path. A vertex $v \in V_G$ is a {\em cut-vertex} if the induced graph $G-v$ is not connected. We say that $G$ is the {\em vertex sum} of two graphs $G_1$ and $G_2$, and write $G_1 \bigoplus_v G_2$ if $v$ is a cut-vertex of $G$, $V_{G_1} \cap V_{G_2} = \{v \}$, and $E_{G_1} \cap E_{G_2} = \emptyset$. A graph with no cut-vertices is said to be {\em nonseparable}. A {\em matching} in a graph $G$ is a set of edges $M = \{ \{ i_1, j_1\}, \{ i_2, j_2 \}, \dots, \{ i_k, j_k \} \} \subseteq E_G$, such that no endpoints are shared. The vertices that determine the edges in $M$ are called {\em $M$-saturated} vertices, all other vertices in $V_G$ are called {\em $M$-unsaturated} vertices. A {\em perfect matching} in a graph $G$ is a matching that saturates all vertices of $G$. A {\em maximum matching} in a graph $G$ is a matching of maximum order among all matchings in $G$. The {\em matching number} of a graph $G$, denoted by $\operatorname{match}(G)$, is the number of edges in a maximum matching. An even cycle in a graph $G$ is called {\em $M$-alternating} if it alternates between edges in $M$ and edges not in $M$. A matching $M$ in a graph $G$ is {\em uniquely restricted} if $G$ does not contain an $M$-alternating cycle. A graph $G$ is {\em $k$-partite} if $V_G$ can be expressed as the union of $k$ (possibly empty) independent sets, and is denoted $K_{n_1,n_2,\dots,n_k}, k \ge 2, n_i \ge 1, i = 1, 2, \dots, k$. A {\em tree} is a connected graph $T$, with $\left| E_T \right| = \left| V_T-1 \right|$, trees are {\em 2-partite}, also known as {\em bipartite}. Although many of the results presented here are valid for some finite fields, we assume throughout this paper that ${\mathbb F}$ is an infinite field. A matrix $A \in {\mathbb F}^{n \times n}$ is {\em skew-symmetric} if $A^T = -A$. For an $n \times n$ skew-symmetric matrix $A$, the {\em graph of $A$}, denoted ${\mathcal G}(A)$, is the graph with vertices $\{v_1, . . . , v_n \}$ and edges $\{ \{v_i, v_j \} : a_{ij} \ne 0, 1 \le i < j \le n \}$. Let $\mathcal{S}^-({\mathbb F}, G) = \{ A \in {\mathbb F}^{n \times n} : A^T = -A, {\mathcal G}(A) = G \}$ be the set of skew-symmetric matrices over the field ${\mathbb F}$ described by a graph $G$. The {\em minimum skew rank of a graph $G$ over the field ${\mathbb F}$ } is defined as $\operatorname{mr}^-({\mathbb F}, G) = \min \{ {\rm rank}(A) : A \in \mathcal{S}^-({\mathbb F}, G) \}$, the {\em maximum skew nullity of $G$ over the field ${\mathbb F}$} is defined as $\operatorname{M}^-({\mathbb F}, G) = \max \{\operatorname{nullity} (A) : A \in \mathcal{S}^-({\mathbb F}, G) \}$, and the {\em maximum skew rank of $G$ over the field ${\mathbb F}$} as $\sMR({\mathbb F}, G) = \max \{ {\rm rank}(A) : A \in \mathcal{S}^-({\mathbb F}, G) \}$. Clearly $\operatorname{mr}^-({\mathbb F}, G) + \operatorname{M}^-({\mathbb F}, G) = |G|$, but note that, since a skew symmetric matrix has even rank, $\sMR ({\mathbb F}, G) \le |G|$. For a graph $G$, select $Z \subseteq V_G$, color all vertices in $Z$ black, and all others white. Next apply the {\em skew color change rule}: if $u \in V_G$ ($u$ any color), and exactly one of its neighbors $v$, is white, then change the color of $v$ to black (we say $u$ forces $v$ black). Continue to apply the skew color change rule until no more changes are possible. A {\em skew zero forcing set} for a graph $G$ is a subset $Z$ of $V_G$, such that, if initially the vertices in $Z$ are colored black and the remaining vertices are colored white, the skew color change rule forces all the vertices in $V_G$ black. A {\em minimum skew zero forcing set} for a graph $G$ is a skew zero forcing set of minimum order among all skew zero forcing sets for $G$. The {\em skew zero forcing number $\operatorname{Z}^-(G)$} is the minimum of $|Z|$ over all skew zero forcing sets $Z \subseteq V_G$. \section{Preliminary results}\label{prelim} The parameter $Z(G)$ was introduced in~\cite{AIM}, while the parameter $\operatorname{Z}^-(G)$, was introduced in~\cite{10IMA}. \begin{prop}\label{tree-known} \begin{enumerate} \item\label{induced}\cite[Observation~1.7]{10IMA} If $H$ is an induced subgraph of $G$, then $\operatorname{mr}^- ({\mathbb F}, H) \le \operatorname{mr}^- ({\mathbb F}, G)$. \item\label{2.810IMA}~\cite[Proposition 3.5]{10IMA} For any graph $G$, $\operatorname{M}^- ({\mathbb F}, G) \le \operatorname{Z}^- (G)$ and $\operatorname{mr}^- ({\mathbb F}, G) \ge |G| - \operatorname{Z}^- (G)$. \item\label{M=Z}~\cite[Proposition 4.2]{AIM} For any tree $T$, $M ({\mathbb F}, T) = Z (T)$, and hence $\operatorname{mr} ({\mathbb F}, T) = |T| - \operatorname{M} ({\mathbb F}, T) = |T| - Z (T)$. \end{enumerate} \end{prop} \begin{thm}\label{knowngen} \begin{enumerate} \item\label{smr=2}\cite[Theorem~2.1]{10IMA} Let $G$ be a connected graph with $|G| \ge 2$, then $\operatorname{mr}^- ({\mathbb F}, G) = 2$ if and only if $G = K_{n_1,n_2,\dots,n_s}, s \ge 2, n_i \ge 1, i = 1, 2, \dots, s$. \item\label{mr=match}~\cite[Theorem~2.5]{10IMA} For a graph $G$, $\sMR ({\mathbb F}, G) = 2 \operatorname{match} (G)$, and every even rank between $\operatorname{mr}^- ({\mathbb F}, G)$ and $\sMR ({\mathbb F}, G)$ is realized by a matrix in $\mathcal{S}^-({\mathbb F}, G)$. \item\label{upm}~\cite[Theorem~2.6]{10IMA} For a graph $G$, $\operatorname{mr}^-({\mathbb F}, G) = |G| = \sMR({\mathbb F}, G)$ if and only if $G$ has a unique perfect matching. \item\label{smrtree}\cite[Theorem~2.8]{10IMA} If $T$ is a tree, then $\operatorname{mr}^-({\mathbb F}, T) = 2 \operatorname{match}(T) = \sMR({\mathbb F}, T)$. \item\label{cover}\cite[Proposition 3.3]{10IMA}. Let ${\mathbb F}$ be a field and $G = \cup_{i=1}^k \, G_i$ be a graph. Suppose that for all $i \ne j$, $G_i$ and $G_j$ have no edges in common, then $\operatorname{mr}^- ({\mathbb F}, G) \le \sum_{i=1}^k \, \operatorname{mr}^- ( {\mathbb F}, G_i)$. \end{enumerate} \end{thm} \section{Graphs with extreme skew zero forcing number}\label{extreme} It is a fact that for any graph $G$, $0 \le \operatorname{Z}^- (G) \le |G|$. If a graph has isolated vertices, those vertices must belong to all skew zero forcing sets for the graph. Thus, without loss of generality, we assume that graphs have no isolated vertices. Also, some of the results presented here are valid for graphs that are disconnected, we specifically note when a graph must be connected. \begin{rem}\label{zsub} If $G$ is a graph, $v \in V_G$, and $Z$ a minimum skew zero forcing set for $G-v$, then $Z \cup \{v \}$ is a skew zero forcing set for $G$, so $\operatorname{Z}^-(G) \le \operatorname{Z}^-(G - v) +1$. \end{rem} From Proposition~\ref{tree-known}, and Theorem~\ref{knowngen}, we obtain the inequalities \begin{equation}\label{ineq} |G| - \operatorname{Z}^- (G) \le \operatorname{mr}^- ({\mathbb F}, G) \le \sMR ({\mathbb F}, G) = 2 \operatorname{match} (G) \le |G|. \end{equation} The following are derived using the inequalities in Equation~\ref{ineq}, and the definition of skew zero forcing set. \begin{obs}\label{smallz} \begin{enumerate} \item\label{z=0} If $\operatorname{Z}^- (G) = 0$, then there is a vertex, $v \in V_G$, such that $\deg_G (v) = 1$. \item\label{z=0upm} If $\operatorname{Z}^- (G) = 0$, then $|G|$ is even, $\operatorname{mr}^- ({\mathbb F}, G) = |G|$ and $G$ has a unique perfect matching. \item\label{z=1even} If $\operatorname{Z}^- (G) = 1$ and $|G|$ is even, then $\operatorname{mr}^- ({\mathbb F}, G) = |G|$ and $G$ has a unique perfect matching. \item\label{z=1odd} If $\operatorname{Z}^- (G) = 1$ and $|G|$ is odd, then $\operatorname{mr}^- ({\mathbb F}, G) = |G| - 1.$ \item\label{z=2odd} If $\operatorname{Z}^- (G) = 2$ and $|G|$ is odd, then $\operatorname{mr}^- ({\mathbb F}, G) = |G| - 1.$ \item\label{z=2even} If $\operatorname{Z}^- (G) = 2$ and $|G|$ is even, then either $\operatorname{mr}^- ({\mathbb F}, G) = |G|$ and $G$ has a unique perfect matching, or $\operatorname{mr}^- ({\mathbb F}, G) = |G| -2$. \item (\cite[Observation 1.6]{10IMA} $\operatorname{Z}^- (G) = |G|$ if and only if $G$ consists only of isolated vertices. \end{enumerate} \end{obs} It is clear that the converses of Items~\ref{z=0}--\ref{z=2even} in Proposition~\ref{smallz} are not true, the graphs in Figures~\ref{upmsz0}--\ref{upmsz2}, and Item~\ref{tri} in Observation~\ref{special}, also illustrate this. For the graph $G_2$ in Figure~\ref{upmsz1} (which is a cactus graph, and also a block-clique graph), $\operatorname{mr}^- ({\mathbb F}, G) = |G| > |G| - 1 = |G| - \operatorname{Z}^- (G)$. \eol{.125} \begin{figure}[h!] \begin{center} \scalebox{.4}{\includegraphics{upm}} \caption{Graph with $\operatorname{Z}^- (G_1) = 0 = |G_1|-6$.}\label{upmsz0} \end{center} \end{figure} \begin{figure}[h!] \begin{center} \scalebox{.4}{\includegraphics{upmz1}} \caption{Graphs with $\operatorname{Z}^- (G_i) = 1 = |G_2| - 5 = |G_3| - 4$.}\label{upmsz1} \end{center} \end{figure} \begin{figure}[h!] \begin{center} \scalebox{.4}{\includegraphics{upmz2}} \caption{Graphs with $\operatorname{Z}^- (G_i) = 2 = |G_4| -10 = |G_5| -2 = |G_6| - 7$.}\label{upmsz2} \end{center} \end{figure} Note that if $G$ is one of the graphs $K_2$, $K_3$, or $K_{2,1}$, then $\operatorname{Z}^- (G) = |G| - 2$. We now show that this equation characterizes all complete multipartite graphs. The proof given below involves the use of $\operatorname{mr}^- ({\mathbb F}, G)$, but one can easily construct a field independent proof. \begin{thm}\label{sz=g-2} A connected graph $G$ is a complete multipartite graph $K_{n_1, n_2, \dots, n_s}$, $s \ge 2, n_i \ge 1$ if and only if $\operatorname{Z}^-(G) = |G|-2$. \end{thm} \begin{proof} Let $G = K_{n_1, n_2, \dots, n_s}$, $s \ge 2, n_i \ge 1$, with $|G| \ge 4$. From Item~\ref{smr=2} in Theorem~\ref{knowngen}, $\operatorname{mr}^- ({\mathbb F}, G) = 2$, hence from Equation~\ref{ineq}, $|G| - 2 \le \operatorname{Z}^- (G)$. Pick $u, v \in V_G$ adjacent (in different partite classes), then $Z = V_G - \{u, v \}$ is a skew zero forcing set for $G$, and $\operatorname{Z}^-(G) \le |G| - 2$. It follows that $\operatorname{Z}^-(G) = |G| - 2$. Conversely, if $G$ is connected, but not $K_{n_1, n_2, \dots, n_s}$, $s \ge 2, n_i \ge 1$, then $G$ has an induced $P_4$ or an induced paw (\cite[Remark 2.2]{10IMA}). If $v_1, v_2, v_3, v_4$ induce a $P_4$, or the paw, then $Z = V_G - \{v_1, v_2, v_3, v_4 \}$ is a skew zero forcing set for $G$, so $\operatorname{Z}^-(G) \le |G| - 4 \ne |G| - 2$. \end{proof} \begin{cor}\label{g-mr} If $G$ is a connected graph, then either $\operatorname{Z}^- (G) = |G| -2$ or $\operatorname{Z}^- (G) \le |G| -4$. There are no connected graphs for which $\operatorname{Z}^- (G) = | G|-1$, and no connected graphs for which $\operatorname{Z}^- (G) = |G|-3$. \end{cor} \begin{rem}\label{ord8} One can verify, directly, that for connected graphs $G$ (pictured in~\cite{98RW}), with $4 \le |G| \le 6$, $\operatorname{Z}^-(G) = |G| - 4$ if and only if $\operatorname{mr}^- ({\mathbb F}, G) = 4$. Henceforth we assume $|G| \ge 7$. \end{rem} \begin{prop}\label{cutv} If $G$ is connected, has a cut-vertex, and $\operatorname{Z}^- (G) = |G| - 4$, then $\operatorname{mr}^- ({\mathbb F}, G) = 4$. \end{prop} \begin{proof} Let $G$ be connected, and $v \in V_G$ be a cut-vertex. Let $G_1$ be the connected subgraph of $G$ induced by the vertices of one of the components of $G-v$ and $v$, and $G_2$ be the connected subgraph of $G$ induced by $(V_G - V_{G_1}) \cup \{ v\}$, so that $G= G_1 \bigoplus_v G_2$. If $\operatorname{Z}^- (G_1) = |G_1| - 2$, and $\operatorname{Z}^- (G_2) = |G_2| - 2$, then $G$ is the vertex sum of two complete multipartite graphs, and in this case $\operatorname{mr}^- ({\mathbb F}, G) = 4$. The two other possibilities that arise from Corollary~\ref{g-mr} do not allow $\operatorname{Z}^- (G) = |G| - 4$. Let $Z_1$ and $Z_2$ be minimum skew zero forcing sets for $G_1$ and $G_2$, respectively. If $\operatorname{Z}^- (G_1) = |G_1| - 2$, $\operatorname{Z}^- (G_2) \le |G_2| - 4$, and $Z_1 \ne \emptyset$, then from the proof of Theorem~\ref{sz=g-2}, we can take $v \in Z_1$, thus $Z_1 \cup Z_2$ is a skew zero forcing set for $G$. If $\operatorname{Z}^- (G_i) \le |G_i| - 4, i = 1, 2$, then $Z_1 \cup Z_2 \cup \{ v \}$ is a skew zero forcing set for $G$. In both cases, $\operatorname{Z}^- (G) < |G| - 4$. \end{proof} \begin{prop}\label{induce6} If $G$ is a connected graph, and $H$ is a connected induced subgraph of $G$ of order 6 that has a unique perfect matching, then $\operatorname{Z}^- (G ) < |G| -4$. \end{prop} \begin{proof} Figure~\ref{upm1} shows the twenty connected graphs on six vertices that have a unique perfect matching. One of these is $G_2$, also pictured in Figure~\ref{upmsz1}, and satisfies $\operatorname{Z}^- (G_2 ) = 1$, all others have $\operatorname{Z}^- (H ) = 0$. If $H=G_2$, and $u \in V_{G_2}$, then $(V_G-V_H) \cup \{ u \}$ is a skew zero forcing set of $G$; if $H \ne G_2$, then $V_G - V_H$ is a skew zero forcing set for $G$. Thus, $\operatorname{Z}^- (G ) \le |G| -5 < |G| - 4$. \end{proof} \begin{figure}[h!] \begin{center} \scalebox{.5}{\includegraphics{upm6}} \caption{Graphs on six vertices~(\cite{98RW}) with a unique perfect matching.}\label{upm1} \end{center} \end{figure} \begin{prop}\label{g-4a} Let ${\mathbb F}$ be a field, and $G$ a connected graph with $|G| \ge 4$. If $\operatorname{mr}^- ({\mathbb F}, G) =4$, then $\operatorname{Z}^-(G) = |G| - 4$. \end{prop} \begin{proof} If $\operatorname{mr}^- ({\mathbb F}, G) = 4$, then $\operatorname{Z}^- (G) \le |G| - 4$, so $4 \le |G| - \operatorname{Z}^- (G) \le \operatorname{mr}^- ({\mathbb F}, G) =4$. \end{proof} The following example provided by Sudipta Mallik and Bryan Shader, and constructed using their methods as in~\cite{13MS}, shows that the converse of Proposition~\ref{g-4a} is not true. \begin{defn}\cite[p. 3651]{13MS} A collection $\{ N_i : i \in \mathcal{I} \}$ of vectors is a minimally dependent set of vector if it is a linearly dependent set and for each $j \in \mathcal{I}, \{ N_i : i \ne j, i \in \mathcal{I} \}$ is a linearly independent set of vectors. \end{defn} \begin{ex} If $G =K_3 \times K_3$ is the graph with adjacency matrix and graph as in Figure~\ref{counterex}, \begin{figure}[h!] \begin{center} \scalebox{.5}{\includegraphics{counter}} \caption{The graph $K_3 \times K_3$.}\label{counterex} \end{center} \end{figure} then $\operatorname{Z}^- (G) = |G| -4$, and $\operatorname{mr}^- (G) \ge 6$. \end{ex} \begin{proof} Using the graph in Figure~\ref{counterex}, and the fact that $G$ is tripartite, it is not difficult to show that $\operatorname{Z}^-(G) > 4$, and that $Z= \{ 1,2,3,4,7 \}$ is a minimum skew zero forcing set for $G$. Hence $\operatorname{Z}^- (G) = 5 = |G|-4$, and $4 = |G| - \operatorname{Z}^-(G) \le \operatorname{mr}^- (G)$. Let $B \in \mathcal{S}^- (G)$, and assume columns $3i + 1, 3i + 2, 3i + 3$ are linearly independent for $i = 0, 1, 2$. Then from the zero-nonzero pattern of $B$ we observe that columns $3i + 1, 3i + 2, 3i + 3, 3i + 4, 3i + 5 \ (\mbox{mod } 9)$ are linearly independent, and since $B$ is skew symmetric, ${\rm rank} (B) \ge 6$. Assume now that columns $3i + 1, 3i + 2, 3i + 3$ of $B$ are linearly dependent for $i = 0, 1, 2$, and hence minimally linearly dependent. By Lemma 4.7 in~\cite{13MS}, the nullspace of $B$ contains vectors of the form \begin{equation} \left[ \begin{array}{c} a\\ b\\ c\\ 0\\ 0\\ 0\\ 0\\ 0\\ 0 \end{array} \right], \left[ \begin{array}{c} 0\\ 0\\ 0\\ d\\ e\\ f\\ 0\\ 0\\ 0 \end{array} \right], \mbox{ and } \left[ \begin{array}{c} 0\\ 0\\ 0\\ 0\\ 0\\ 0\\ g\\ h\\ k \end{array} \right],\end{equation}\label{vecnul} for some $a, b, c, d, e, f, g, h, k$, each of which is nonzero. Let $D = \mbox{diag} (a, b, c, d, e, f, g, h, k)$, hence $DBD \in \mathcal{S}^-(G)$, and ${\rm rank} (DBD) = {\rm rank} (B)$. Note that the nullspace of $DBD$ contains vectors as in Equation~\ref{vecnul} with $a = b = c = d = e = f = g = h = k = 1$. Direct calculations now show that $DBD$ has the form: $$\left[ \begin{array}{ccc} 0 & x S & y S\\ x S & 0 & z S\\ y S & z S & 0 \end{array} \right] = \left[ \begin{array}{ccc} 0 & x & y\\ x & 0 & z\\ y & z &0 \end{array} \right] \otimes S$$ for some nonzero $x, y$ and $z$. Since ${\rm rank} \left( \left[ \begin{array}{ccc} 0 & x & y\\ x & 0 & z\\ y & z &0 \end{array} \right] \right) = 3$, and ${\rm rank} (S) = 2$, it follows that ${\rm rank} (B) = {\rm rank} (DBD) = 6$. Hence $\operatorname{mr}^-(G) \ge 6$. \end{proof} \section{Bipartite graphs}\label{main} In this section we study the relation between certain matchings and skew zero forcing sets. Bipartite graphs provide a natural setting for this discussion. \begin{prop}\label{constrained} If $B$ is a bipartite graph, and $M$ a uniquely restricted matching in $B$, then the set of $M$-unsaturated vertices of $B$ is a skew zero forcing set for $B$. \end{prop} \begin{proof} Let $B$ be a bipartite graph, $M$ a uniquely restricted matching in $B$, and $H$ the connected subgraph of $B$ induced by the vertices in $M$ (If $H$ is not connected, the following process can be applied separately to each of the components of $H$). Suppose the vertices in the bipartition of $H$ are $u_1, \dots, u_r$ and $v_1, \dots, v_r$, $\{u_i, v_i \} \in M$, $\{u_i, v_j \} \notin E_H$ whenever $1 \le i < j \le r$. Let $Q = V_B - V_H$, and color the vertices in $Q$ black. Without loss of generality we may assume $\deg_H (v_r) = 1$, then we have the following sequence of forces $v_r \rightarrow u_r, v_{r-1} \rightarrow u_{r-1}, \dots, v_1 \rightarrow u_1, u_1 \rightarrow v_1, u_2 \rightarrow v_2, \dots, u_r \rightarrow v_r$. Thus $Q$ forms a skew zero forcing set for $B$. \end{proof} \begin{prop}\label{qlessz} Let $G$ be a graph, $M$ a matching in $G$, and $\operatorname{mr}^- ({\mathbb F}, G) \le 2|M|$. If the set of $M$-unsaturated vertices of $G$ is a skew zero forcing set for $G$, then it is a minimum skew zero forcing set for $G$. \end{prop} \begin{proof} Let ${\mathbb F}$ be a field, $M$ a matching in $G$, and $Q$ the set of $M$-unsaturated vertices. From Equation~\ref{ineq}, $|G| - \operatorname{Z}^- (G) \le \operatorname{mr}^- ({\mathbb F}, G) \le 2 |M|$, so $|Q| = |G| - 2 |M| \le \operatorname{Z}^- (G)$. Thus, if $Q$ is a zero forcing set for $G$, it is a minimum skew zero forcing set for $G$. \end{proof} \begin{cor}\label{z=q} Let $G$ be a graph, $M$ a matching in $G$, and $\operatorname{mr}^- ({\mathbb F}, G) \le 2|M|$. If the set of $M$-unsaturated vertices is a minimum skew zero forcing set for $G$, then $|G| - \operatorname{Z}^- (G) = \operatorname{mr}^- ({\mathbb F}, G)$. \end{cor} \begin{proof} If $Q$ denotes the set of $M$-unsaturated vertices, then $2 |M| = |G| - |Q| = |G| - \operatorname{Z}^- (G) \le \operatorname{mr}^- ({\mathbb F}, G) \le 2|M|$. \end{proof} \begin{cor}\label{bitt} If $B$ is a bipartite graph (in particular a tree, or cactus) with $\operatorname{mr}^- ({\mathbb F}, B) = \sMR ({\mathbb F}, B)$, then \begin{enumerate} \item\label{maxmr} There is a maximum matching in $B$ such that the set of $M$-unsaturated vertices is a minimum skew zero forcing set for $B$. \item\label{mrt=t-z} $\operatorname{Z}^- (B) = |B| - \operatorname{mr}^- ({\mathbb F}, B)$, and $\operatorname{M}^- ({\mathbb F}, B) = \operatorname{Z}^- (B)$. \item $\operatorname{Z}^- (B) =0$, if and only if $B$ has a unique perfect matching. \item $\operatorname{Z}^- (B) \le \frac{ \Delta (B) |B| - 2 |E_B|}{\Delta (B)}$. In particular, \begin{enumerate} \item if $T$ is a tree, $\operatorname{Z}^- (T) \le \frac{|T| ( \Delta (T) -2) + 2}{\Delta (T)}$, and this bound is sharp for paths and stars; \item if $U$ is a unicyclic, $\operatorname{Z}^- (U) \le \frac{|U| ( \Delta (U) -2)}{\Delta (U)}$. \end{enumerate} \end{enumerate} \end{cor} \begin{proof} \begin{enumerate} \item If $B$ is a bipartite graph with $\operatorname{mr}^- ({\mathbb F}, B) = \sMR ({\mathbb F}, B)$, then there must be a uniquely restricted maximum matching $M$, in $B$. Then use Proposition~\ref{constrained}, and Proposition~\ref{qlessz}. \item\label{m=z} This follows from Item~\ref{maxmr} above, and Corollary~\ref{z=q}. \item This follows from Item~\ref{upm} in Theorem~\ref{knowngen}, as well as Item~\ref{mrt=t-z} above. \item This follows from the fact that a bipartite graph $B$ has a matching of size at least $\frac{|E_B|}{\Delta (B)}$ (\cite[p. 108]{08HHM}). \end{enumerate} \end{proof} The results in Proposition~\ref{constrained} and Proposition~\ref{qlessz}, suggest a certain duality between maximum matchings and minimum skew zero forcing sets. We establish that a natural duality, via matroids, does indeed exist for some families of graphs. \begin{defn}{\bf The matching matroid and its dual.}~\cite[pp. 92--93]{86LP} If $G$ is a bipartite graph, the set $$\mu =\{ X \subseteq V_G \, : \, X \, \mbox{ is saturated by some matching} \}$$ is a matroid on $V_G$ with bases the sets of vertices saturated by maximum matchings in $G$. Its dual matroid $$\mu^* =\{ Q \subseteq V_G \, : \, Q \, \mbox{ is not saturated by some maximum matching} \}$$ on $V_G$ has bases $V_G - B_i$, where $B_i$ is a basis of $\mu$. \end{defn} \begin{thm}\label{matroid} {\bf The zero forcing matroid.} In a bipartite graph $B$, in which all maximum matchings are uniquely restricted, the elements in the set of minimum skew zero forcing sets are the bases of a matroid on the vertices of the corresponding graph, and this matroid is the dual of the matching matroid on $B$. \end{thm} \begin{thm}\label{treeall} If $B$ is a bipartite graph in which all maximum matchings are uniquely restricted, and $M$ a matching in $B$, then $M$ is a maximum matching in $B$ if and only if the set of $M$-unsaturated vertices of $B$ is a minimum skew zero forcing set for $B$. Alternatively, let $Z$ be a skew zero forcing set for $B$, then $Z$ is a minimum skew zero forcing set for $B$ if and only if $V_B - Z$ has a unique perfect matching which is a maximum matching in $B$. \end{thm} \begin{proof} Let $B$ be a a bipartite graph in which all maximum matchings are uniquely restricted, $M$ a maximum matching in $B$, and $Q$ the set of $M$-unsaturated vertices. Since $M$ is a uniquely restricted matching, from Proposition~\ref{constrained}, $Q$ is a minimum skew zero forcing set for $B$. Conversely, suppose that $Q$ is a minimum skew zero forcing set for $B$. From Theorem~\ref{matroid}, $V_B-Q$ has a unique perfect matching which is a maximum matching in $B$. We omit the alternate proof. \end{proof} In a tree all maximum matchings are uniquely restricted matchings, thus we have the following. \begin{cor}\label{tree} If $T$ is a tree, and $M$ a matching in $B$, then $M$ is a maximum matching in $T$ if and only if the set of $M$-unsaturated vertices of $T$ is a minimum skew zero forcing set for $T$. Alternatively, let $Z$ be a skew zero forcing set for $T$, then $Z$ is a minimum skew zero forcing set for $T$ if and only if $V_T - Z$ has a unique perfect matching which is a maximum matching in $T$. \end{cor} \section{Unicyclic Graphs} Results on the minimum skew rank of unicyclic graphs can be found in~\cite{11D}, explicitly: $\operatorname{mr}^- ({\mathbb F}, U) = \sMR({\mathbb F}, U)$ if the unique cycle is odd, or if the unique cycle is even and $U$ has a uniquely restricted maximum matching; $\operatorname{mr}^- ({\mathbb F}, U) = \sMR({\mathbb F}, U) - 2$ if the unique cycle is even and $U$ does not have a uniquely restricted maximum matching. \begin{prop}\label{uniodd} If $U$ is a unicyclic graph, then there exists a matching $M$, in $U$, such that the set of $M$-unsaturated vertices of $U$ is a minimum skew zero forcing set for $U$. \end{prop} \begin{proof} If the unique cycle has odd order, and $M$ is a maximum matching in $U$, we can construct a proof by induction to show that the set of $M$-unsaturated vertices is a minimum skew zero forcing set for $U$. The base cases follow from examples in~\cite{11D}; we omit the details of the proof. If the unique cycle has even order, and $\operatorname{mr}^- ({\mathbb F}, U) = \sMR({\mathbb F}, U)$, then the result follows from Item~\ref{maxmr} in Corollary~\ref{bitt}. If the unique cycle has even order, $\operatorname{mr}^- ({\mathbb F}, U) = \sMR({\mathbb F}, U)-2$, and $\widehat M$ is a maximum matching in $U$, then the cycle is $\widehat M$-alternating. If $U$ is a cycle, and $e $ is an edge in $\widehat M$, then $M = \widehat M-e$ is a uniquely restricted matching in $U$. If $U$ is not a cycle, then it has an induced subgraph $H$, consisting of the vertex sum of the cycle and a path of order 3, that is $H= C \bigoplus_{v_1} P_3$, where $P_3 = (\{v_1, v_2, v_3\}, \{\{v_1, v_2 \}, \{ v_2 , v_3 \} \})$, and $u$ is a neighbor of $v_1$ on the cycle. Thus, there exists a maximum matching $\widehat M$, in $U$, containing the edges $\{ u, v_1 \}$, and $\{ v_2, v_3 \}$ (if $v_2$ is $\widehat M$-unsaturated, then $(\widehat M - \{ \{ u, v_1 \} \} ) \cup \{ \{ v_1, v_2 \}\}$ is a uniquely restricted maximum matching). The matching $M = ( \widehat M - \{ \{ u, v_1 \}, \{ v_2, v_3 \} \} ) \cup \{ \{ v_1, v_2 \} \}$, is a uniquely restricted matching in $U$. In either case $M$ has order $\frac{\sMR({\mathbb F}, U) - 2}{2}$, and since $\operatorname{mr}^- (U) = \sMR({\mathbb F}, U) - 2$, it follows from Proposition~\ref{constrained} that the set of $M$-unsaturated vertices is a minimum skew zero forcing set for $U$. \end{proof} \begin{cor}\label{uni} If $U$ is a unicyclic graph, then $\operatorname{Z}^- (U) = |U| - \operatorname{mr}^- ({\mathbb F}, U)$. \end{cor} \section{Additional Examples} We conclude with several contrasting examples of graphs $G$, for which there is a matching $M$, in $G$, such that the set of $M$-unsaturated vertices is a minimum skew zero forcing set for $G$. Also, in Observation~\ref{special}, we list the skew zero forcing number of some special graphs. \begin{ex} The graph $G_7$, in Figure~\ref{cactus}, is a non-bipartite cactus, does not have a unique maximum matching, but has a maximum matching, $M_7$, of order 5, with no $M_7$-alternating cycle. The set of $M_7$-unsaturated vertices (in black) is a minimum skew zero forcing set for $G_7$, and from Item~\ref{z=2even} in Observation~\ref{smallz}, $|G_7| - \operatorname{Z}^- (G_7) = 10 = \operatorname{mr}^- ({\mathbb F}, G_7) \ne \sMR ({\mathbb F}, G_7) = 12$. \end{ex} \begin{figure}[h!] \begin{center} \scalebox{.4}{\includegraphics{cactus}} \caption{Non-bipartite graph with $\operatorname{Z}^- (G_7) = 2, |G_7| - 2 = \operatorname{mr}^- ({\mathbb F}, G_7)$.}\label{cactus} \end{center} \end{figure} \begin{ex} The graph $G_8$ (see~\cite[pp.6--7]{86LP}), in Figure~\ref{lovasz}, is bipartite, does not have a perfect matching, but has a uniquely restricted matching $M_8$, of cardinality 18 (it is easy to verify that all matchings in $G_8$ of cardinalities 20 and 19 are not uniquely restricted), the set of $M_8$-unsaturated vertices (in black) is a minimum skew zero forcing set for $G_8$, and with the aid of Mathematica one can verify that $|G_8| - \operatorname{Z}^- (G_8) = 42 - 6 = 36 = \operatorname{mr}^- ({\mathbb F}, G) \ne \sMR ({\mathbb F}, G)$. \end{ex} \begin{figure}[h!] \begin{center} \scalebox{.4}{\includegraphics{lovasz}} \caption{Bipartite graph with $\operatorname{Z}^- (G_8) = 6, |G_8| - 6 = \operatorname{mr}^- ({\mathbb F}, G_8)$.}\label{lovasz} \end{center} \end{figure} \begin{ex} The graph $G_4$, in Figure~\ref{upmsz2} is not bipartite, not unicyclic, and not a cactus, has a unique perfect matching, but also has a matching $M_4$ of order $5$, such that the set of $M_4$-unsaturated vertices is a minimum skew zero forcing set for $G_4$, and $|G_4| - \operatorname{Z}^- (G_4) = 10 < \operatorname{mr}^- ({\mathbb F}, G) = \sMR ({\mathbb F}, G)$. \end{ex} Below we list some graphs and their skew zero forcing numbers. We refer the reader to~\cite{10IMA} for the definitions of $W_n$, the wheel on $n$ vertices; $P_{m,k}$, the $m, k$-pineapple, with $m \ge 3, k \ge 1$; $Q_s$, the $s$th hypercube; $T_n$, the super-triangle; $H_s$ the $s$th half-graph; $N_s$, the necklace with $s$ diamonds; $G \circ H$, the corona of $G$ with $H$; $G \boxempty H$, the Cartesian product of $G$ and $H$. \begin{obs}\label{special} For the graphs $G$ in Items~\ref{path},~\ref{cycle},~\ref{wheel},~\ref{pineapple},~\ref{cube},~\ref{half},~\ref{corona}, and~\ref{pdp}, $\operatorname{Z}^-(G) = |G| - \operatorname{mr}^- ({\mathbb F}, G)$ (note that there might be restrictions on the field ${\mathbb F}$, see~\cite{10IMA}): \begin{enumerate} \item\label{path} $\operatorname{Z}^- (P_n) = \left\{ \begin{array}{lll} 0 & \mbox{if} & n \mbox{ is even},\\ 1 & \mbox{if} & n \mbox{ is odd};\\ \end{array} \right.$ \item\label{cycle} $\operatorname{Z}^- (C_n) = \left\{ \begin{array}{lll} 2 & \mbox{if} & n \mbox{ is even},\\ 1 & \mbox{if} & n \mbox{ is odd};\\ \end{array} \right.$ \item\label{wheel} $\operatorname{Z}^- (W_n) = \left\{ \begin{array}{lll} 2 & \mbox{if} & n \mbox{ is even},\\ 3 & \mbox{if} & n \mbox{ is odd};\\ \end{array} \right.$ \item\label{pineapple} $\operatorname{Z}^- (P_{m,k} ) = |P_{m,k} | - 4= m+k-4, m \ge 3, k \ge 1 $; \item\label{cube} $\operatorname{Z}^- (Q_s) = 2^{s-1}, s \ge 2$; \item\label{tri} $\operatorname{Z}^- (T_n) = n-1$; \item\label{half} $\operatorname{Z}^- (H_s ) = 0$; \item\label{neck} $\operatorname{Z}^- (N_s ) = s$, $\operatorname{Z}^-(N_s) = |N_s| - \operatorname{mr}^- ({\mathbb F}, N_s)$ if and only if $s=2$; \item\label{corona} $\operatorname{Z}^- (G \circ K_1 ) = 0$; \item\label{corona2} $\operatorname{Z}^- (C_t \circ K_s ) = st-3t+2, s \ge 2, \operatorname{Z}^-(C_t \circ K_s ) = |C_t \circ K_s | - \operatorname{mr}^- ({\mathbb F}, C_t \circ K_s )$ if and only if $s$ is even; \item\label{pdp} $\operatorname{Z}^- (P_s \square P_s ) = s$; \item\label{cdp} $\operatorname{Z}^- (K_3 \square P_2 ) = 2$, $\operatorname{Z}^-(K_3 \square P_2 ) = |K_3 \square P_2 | - \operatorname{mr}^- ({\mathbb F}, K_3 \square P_2 )$, and for $s \ge 3, t \ge 3$, $\operatorname{Z}^- (K_s \square P_t ) = s$, $\operatorname{Z}^-(K_s \square P_t ) = |K_s \square P_t | - \operatorname{mr}^- ({\mathbb F}, K_s \square P_t )$ if and only if $s$ is even, or both $s$ and $t$ are odd. \end{enumerate} \end{obs}

https://arxiv.org/abs/1301.2390

Completely Positive formulation of the Graph Isomorphism Problem

Given two graphs $G_1$ and $G_2$ on $n$ vertices each, we define a graph $G$ on vertex set $V_1\times V_2$ and the edge set as the union of edges of $G_1\times \bar{G_2}$, $\bar{G_1}\times G_2$, $\{(v,u'),(v,u"))(|u',u"\in V_2\}$ for each $v\in V_1$, and $\{((u',v),(u",v))|u',u"\in V_1\}$ for each $v\in V_2$. We consider the completely-positive Lovász $\vartheta$ function, i.e., $cp\vartheta$ function for $G$. We show that the function evaluates to $n$ whenever $G_1$ and $G_2$ are isomorphic and to less than $n-1/(4n^4)$ when non-isomorphic. Hence this function provides a test for graph isomorphism. We also provide some geometric insight into the feasible region of the completely positive program.

\section{Introduction} Let $G_1=(V,E_1),G_2=(V,E_2)$ be two simple undirected graphs, where $V$ is the set of vertices of cardinality $n$ and $E_1,E_2$ are the respective sets of edges. $G_1$ and $G_2$ are called isomorphic if there exists a bijection $\sigma:V\rightarrow V$ such that $(\sigma(x),\sigma(y))\in E_2$ if and only if $(x,y)\in E_1$. The graph isomorphism problem (GI) is the problem of determining if $G_1$ and $G_2$ are isomorphic. This problem although clearly in class \textbf{NP}, has not been known to be either in \textbf{P} or \textbf{NP-Complete} \cite{GJ}, except for certain graphs where it is known to have polynomial complexity \cite{BGM,Bod,FM,HW,Luks,Miller}. There has been evidence suggesting that GI is not likely to be \textbf{NP-Complete}. One of them being that its counting version is reducible to its decision version \cite{Mathon}. Moreover, if the problem were \textbf{NP-Complete}, then the polynomial time hierarchy would collapse to its second level \cite{Babai,BHZ,Schon}. A lot of research has therefore gone into determining the largest complexity class for which it can be shown that GI is hard \cite{JKMT,Toran}. The largest complexity class known to be reducible to GI is \textbf{DET} \cite{Toran}. The complexity aspects of GI are treated in much detail in \cite{AT,KST}. Apart from the obvious theoretical importance of determining its computational complexity, the graph isomorphism problem finds such diverse applications as chemical identification \cite{Sussen}, scene analysis \cite{ABBBP} and construction and enumeration of combinatorial configurations \cite{CM}. Several approaches to solve GI in polynomial time have been adopted. Among them is an approach to incrementally build an isomorphism between the graphs, \cite{RND}. Another approach has been to find a canonical labeling of the vertices of the two graphs, \cite{BES,BK,McKay}. A comprehensive list of all the approaches is difficult to present here. There are some survey papers on the work published on this problem, such as \cite{Fortin}. It was conjectured in \cite{Ramana} that the graph isomorphism problem can be reduced to a semidefinite feasibility problem. We make an attempt in this direction. Given two graphs $G_1$ and $G_2$ on $n$ vertices each, we consider the Lov\'asz $\vartheta$ function, for an $n^2$ vertex graph based on the two input graphs, with positive semidefinite condition replaced by completely positivity condition. We show that if the graphs are isomorphic then the function evaluates to $n$ and if non-isomorphic then it evaluates to a value less than $n-1/(4n^4)$. Hence this provides a test for GI. \section{Preliminaries} \subsection{Positive Semidefinite Matrices} An $m\times m$ symmetric matrix $M$ is said to be {\em positive semidefinite} if it can be expressed as $Q\cdot Q^T$ for some $m\times k$ matrix $Q$. If the row vectors of $Q$ are $v_1,\dots,v_m$, then we will call this set a vector-realization of $M$ in $k$-dimensional space. We will denote the corresponding matrix $M$ by ${\bf M}(v_1,\dots,v_m)$. It is easy to see that there is always a vector realization in $k= rank(M)$ dimensional space. If all entries of a positive semidefinite matrix $M$ are non-negative, then it is called a {\em doubly-non-negative} (DN) matrix. Further, if $M$ has a vector realization in which each component of each vector is non-negative (i.e., $M$ has a decomposition $Q\cdot Q^T$ where each entry of $Q$ is non-negative), then $M$ is called a {\em completely positive} (CP) matrix. We will call it a non-negative vector realization of the CP matrix $M$. Every principal submatrix of a DN (resp. CP) matrix is DN (resp. CP). It is easy to see that every CP matrix is a DN matrix. It is not necessary that every decomposition $Q\cdot Q^T$ of a CP matrix has all non-negative entries in $Q$. The smallest $k$, for which such an $m\times k$ matrix exists, is called the {\em cp-rank} of $M$. \begin{theorem}[\cite{Hannah}] \label{thm1} For any CP matrix $M$ of rank $r$, $cp$-$rank(M) \leq r(r+1)/2$. \end{theorem} A geometrical view of a CP matrix is that if $v_1,\dots,v_m$ is a non-negative vector realization of it, then these vectors belong to the {\em closed positive orthant} (`closed' in the sense of a polyhedron) of some orthogonal basis of the space. \subsection{United Vectors} Let $w$ be any fixed unit vector. Then for every unit vector $v$, we call $u=(w+v)/2$ a {\em united vector with respect to $w$}. We will drop the reference to $w$ when it is unambiguous. \begin{observation}\label{obs1} With respect to a fixed unit vector $w$,\\ (i) a vector $u$ is united if and only if $u\cdot w = u^2$,\\ (ii) if $u_1$ and $u_2$ are mutually orthogonal united vectors, then $u_1+u_2$ is also a united vector.\\ (iii) let $u_1,\dots,u_k$ be a set of pairwise orthogonal united vectors. This set is maximal (i.e., no new united vector can be added to it while preserving pairwise orthogonality) if and only if $w$ belongs to the subspace spanned by these vectors if and only if $\sum_i u_i = w$.\\ (iv) for any collection of pairwise orthogonal united vectors $u_1,\dots, u_j$, $w\cdot \sum_i u_i = \sum_i u_i^2 \leq 1$. Further, $\sum_i u_i^2 = 1$ if and only if the set is maximal, i.e., $\sum_i u_i = w$. \end{observation} \begin{lemma}\label{lem1} Let $u_1,\dots, u_k$ be pairwise orthogonal united vectors. Then matrices ${\bf M}(u_1,\dots,u_k)$ and ${\bf M}(u_1,\dots,u_k,w)$ are CP. \end{lemma} \begin{proof} Suppose $u_i\cdot w=a_i$ for $i=1,\dots,k$. Then there exists a coordinate system in $(k+1)$-dimensional space in which $u_1=(\sqrt{a_1},0,\dots,0)$, $u_2=(0,\sqrt{a_2},0,\dots,0)$ so on, and $w=(\sqrt{a_1},\sqrt{a_2},\dots,\sqrt{a_k},\sqrt{1-\sum_ia_i})$. Since all entries are non-negative reals, the claim is established. $\Box$ \end{proof} \section{Lov\'asz Theta Function \cite{LTF}} Given two graphs, each on $n$ vertices, $G_1=([n],E_1)$ and $G_2=([n],E_2)$. consider the semidefinite program SDP-LT given below, which computes a $\vartheta$-function. The variable matrix $Y$ is of size $(n^2+1)\times (n^2+1)$ with index set $\{ij | i,j\in [n]\} \cup \{\omega\}$. \begin{alignat}{2} \text{SDP-LT:}\quad\text{maximize } \quad & \sum_{i,j\in [n]}Y_{ij,ij}\ \notag\\ \text{subject to} \quad & Y\succeq 0 \ \\ & Y_{ij,kl}\geq 0\ &,\ & 1\leq i,j,k,l\leq n\\ & Y_{\omega,\omega}=1 \ \\ & Y_{ij,\omega}=Y_{ij,ij}\ &,\ & 1\leq i,j\leq n\\ & Y_{ij,ik}=0\ &,\ & 1\leq i,j,k\leq n,\ j\neq k\\ & Y_{ji,ki}=0\ &,\ & 1\leq i,j,k\leq n,\ j\neq k\\ & Y_{ij,kl}=0\ &,\ & (i,k)\in E_1 ,\ (j,l)\notin E_2\\ & Y_{ij,kl}=0\ &,\ & (i,k)\notin E_1 ,\ (j,l)\in E_2 \end{alignat} Let $Y$ be a solution of SDP-LT and let $\{u_{ij}|i,j\in[n]\} \cup \{w\}$ be a vector realization of $Y$. Then from conditions (3) and (4) every $u_{ij}$ is a united vector with respect to the unit vector $w$. Conditions (1) and (2) ensure that $Y$ is DN. Every solution matrix $Y$ of SDP-LT is $(n^2+1)\times (n^2+1)$ in size in which the last row and the last column are same as the diagonal. Hence from here onwards we will drop the last row and the last column and assume that $Y$ is an $n^2\times n^2$ matrix. Let $\{u_{ij}|i,j\in [n]\}\cup \{w\}$ be a vector realization of any solution $Y$ of SDP-LT. Consider the matrix \[ W=\left( \begin{array}{cccc} u_{11} & u_{12} & \dots & u_{1n}\\ u_{21} & u_{22} & \dots & u_{2n}\\ \vdots & \vdots & \ddots & \vdots\\ u_{n1} & u_{n2} & \dots & u_{nn} \end{array} \right)\] Each row and each column of this matrix is a set of pairwise orthogonal united vectors with respect to $w$. Hence from Observation \ref{obs1}(iv) the value of the objective function of SDP-LT, $\sum_{i,j\in[n]}u_{ij}^2$, is at most $n$. \vspace{1mm} \noindent {\bf Remark:} The graph for which SDP-LT is a Lov\'asz $\vartheta$ function, has a clique cover of size $n$ (the edges of condition (5)). Hence this also establishes that the function value is bounded above by $n$. \section{The Case of Isomorphic Graphs} Let us continue to assume that $\{u_{ij}|i,j\in [n]\}\cup \{w\}$ is a vector realization of an arbitrary solution $Y$ of SDP-LT. Suppose there exists any set of $n$ vectors $\{u_{i_1j_1},\dots, u_{i_nj_n}\}$ in which each pair has positive inner product. We will call such a set a {\em complete consistent set}. Observe that both $i_1,\dots,i_n$ and $j_1,\dots,j_n$ are permutations of $1,2,\dots,n$. Hence we can rearrange them as $\{u_{1\sigma(1)},u_{2\sigma(2)},\dots,\\ u_{n\sigma(n)}\}$. It is easy to see that in this case $\sigma$ is an isomorphism between $G_1$ and $G_2$. Consider the case when $\sigma$ is an isomorphism between $G_1$ and $G_2$. Consider a special solution of SDP-LT for this case: Let $w_0$ be some constant unit vector. Define $w=w_0$ and $u_{i\sigma(i)} = w$ for all $i$ and $u_{ij}=0$ whenever $j\neq \sigma(i)$. In this case the matrix $Y$ is $P^{[2]}_{\sigma}$ defined below. \begin{definition} For any permutation $\sigma \in S_n$ (the symmetric group), the $n^2\times n^2$ matrix $P^{[2]}_{\sigma}$ is defined by $[P^{[2]}_{\sigma}]_{ij,kl} = [P_{\sigma}]_{ij}\cdot [P_{\sigma}]_{kl}$, where $P_{\sigma}$ denotes the permutation matrix of $\sigma$. The convex hull of $\{P^{[2]}_{\sigma}| \sigma \textrm{ is a } G_1,G_2 \textrm{ isomorphism}\}$ will be denoted by ${\cal P}_{G_1,G_2}$. \end{definition} \begin{observation}\label{obs2} By construction $P^{[2]}_{\sigma}$ matrices are rank-1 positive semidefinite matrices. Since all entries of $P_{\sigma}$ are non-negative, $P^{[2]}_{\sigma}$ are CP. \end{observation} If $P^{[2]}_{\sigma}$ is a solution of SDP-LT, then $\sigma$ is an isomorphism because $(P^{[2]}_{\sigma})_{i\sigma(i),j\sigma(j)}\\ = 1$ for all $i,j$. Above discussion leads to the following lemma. \begin{lemma}\label{lem3} $P^{[2]}_{\sigma}$ is a CP solution of SDP-LT if and only if $\sigma$ is an isomorphism between $G_1$ and $G_2$. \end{lemma} The value of the objective function for $Y=P^{[2]}_{\sigma}$ is $\sum_{ij}[P^{[2]}_{\sigma}]_{i\sigma(i),i\sigma(i)} = n$. From the last statement of the previous section we have the following result. \begin{lemma}\label{lem4} The value of the objective function of SDP-LT is less than or equal to $n$. It reaches its maximum value $n$ when $G_1$ and $G_2$ are isomorphic. \end{lemma} \begin{lemma}\label{lem5} Let $Y=\sum_{\sigma \in I} a_{\sigma} P^{[2]}_{\sigma}$ is a solution of SDP-LT, where $a_{\sigma}>0$ for each $\sigma \in I$ and $\sum_{\sigma\in I}a_{\sigma}=1$. Then $P^{[2]}_{\sigma}$ is a solution of SDP-LT for each $\sigma \in I$. \end{lemma} \begin{proof} When $P^{[2]}_{\sigma}$ is extended to $(n^2+1)\times (n^2+1)$, then conditions (3) and (4) of SDP-LT are trivially satisfied. The extended matrix is equal to $Q\cdot Q^T$ where $(n^2+1) \times 1$ matrix $Q$ has first $n^2$ entries same as those of $P_{\sigma}$ (i.e., first $n^2$ entries is the vectorized $P_{\sigma}$) and the last entry is $1$. Hence it satisfies conditions (1) and (2). The last four conditions of the SDP are satisfied by the extended $P^{[2]}_{\sigma}$ because every zero condition satisfied by $Y$ is also satisfied by $P^{[2]}_{\sigma}$. Thus $P^{[2]}_{\sigma}$ is a solution for every $\sigma\in I$. $\Box$ \end{proof} From now on we consider SDP-LT with conditions (1) and (2) replaced with the condition that $Y\in\mathcal{C}^{*}$ where $\mathcal{C}^{*}$ is the cone of Completely Positive matrices. Let us call the modified program CP-LT and denote the function by $cp\vartheta$. Now we present the main result of this section. \begin{lemma}\label{lem6} The $cp\vartheta$ function value is $n$ if and only if $G_1$ and $G_2$ are isomorphic. Moreover, in this case the feasible region of CP-LT is equal to ${\cal P}_{G_1G_2}$. \end{lemma} \begin{proof} Consider a non-negative vector realization $\{u_{ij}|i,j\in[n]\}\cup \{w\}$ for a CP solution $Y$ and the corresponding matrix $W$ defined towards the end of Section 3. Since objective function attains value $n$, from Observation \ref{obs1} vectors of each row/column form a maximal set of pairwise orthogonal united vectors. Also from the same Observation each row and each column adds up to $w$. Assume that the vector realization is in an $N$-dimensional space. Consider the $r$-th component of the matrix, i.e., the matrix formed by the $r$-th component of each vector. Let us denote it by $D_r$. Each element of $D_r$ is non-negative and each row and each column adds up to $w_r$, the $r$-th component of $w$. Hence $D_r$ is $w_r$ times a doubly-stochastic matrix. But the vectors of the same row (resp. column) are orthogonal so exactly one entry is non-zero in each row (resp. column) if $w_r >0$. So $D_r = w_rP_{\sigma_r}$ for some permutation $\sigma_r$. We can express $W$ by $\sum_r w_rP_{\sigma_r}e_r$ where $e_r$ denotes the unit vector along the $r$-th axis. $Y_{ijkl}$ is the inner product of the vectors $u_{ij}$ and $u_{kl}$ which is $(\sum_r w_r(P_{\sigma_r})_{ij}e_r)\cdot (\sum_s w_s(P_{\sigma_s})_{kl}e_s) = \sum_r w_r^2 (P_{\sigma_r})_{ij})(P_{\sigma_r})_{kl}) = \sum_r w_r^2(P^{[2]}_{\sigma_r})_{ij,kl}$. Thus $Y = \sum_r w_r^2 P^{[2]}_{\sigma_r}$. Since $\sum_rw_r^2 = w^2 =1$, $Y$ is a convex combination of some of the $P^{[2]}_{\sigma}$. From Lemmas \ref{lem5} and \ref{lem3} each $\sigma_r$, with $w_r >0$, is an isomorphism between $G_1$ and $G_2$. Since $w$ is a unit vector, $w_r>0$ for at least one $r$. Hence $G_1$ and $G_2$ are isomorphic. Conversely from Lemma \ref{lem3} if $\sigma$ is an isomorphism, then $P^{[2]}_{\sigma}$ is a solution and its objective function value is $n$. When the $cp\vartheta$ function has value $n$, the above discussion implies that the feasible region is contained in ${\cal P}_{G_1G_2}$. Conversely, from Lemma \ref{lem3} and the fact that convex combination of CP solutions is also a CP solution, we deduce that ${\cal P}_{G_1G_2}$ is contained in the feasible region. $\Box$ \end{proof} \section{The Case of Non-Isomorphic Graphs} \begin{lemma}\label{lem7} Let $Y$ be a solution of CP-LT. If $\sum_{j}Y_{ij,ij} \geq 1 - 1/(4n^4)$ for each $i$, then $G_1$ and $G_2$ are isomorphic. \end{lemma} \begin{proof} Let $N$ denote the cp-rank of $Y$. So we have a non-negative vector realization of $Y$, $\{u_{ij}| i,j \in [n]\}\cup \{w\}$ in an $N$-dimensional space. So we have $u_{ij}\cdot u_{kl} = Y_{ij,kl}$ for all $i,j,k,l$ and there is an orthonormal basis of this space, $B = \{e_p| p\in [N]\}$, such that every $u_{ij}$ belongs to the closed positive orthant of this basis. From Theorem \ref{thm1} $N < n^4$ because the rank of $Y$ is at most $n^2$. Hence from the statement of this lemma $w\cdot \sum_ju_{ij} = \sum_ju_{ij}^2 = \sum_{ij}Y_{ij,ij}= > 1 - 1/(4N)$ for each $i$. Let $S_i=\{u_{i1},\dots,u_{in}\}$ for each $i$. Since it is a set of orthogonal united vectors with respect to $w$, $w.\sum_j u_{ij} = \sum_j u_{ij}^2 \leq 1$. Without loss of generality assume that $w\cdot e_1\geq w\cdot e_j$ for all $j$. So $w\cdot e_1 \geq 1/\sqrt{N}$ because $w$ is a unit vector. If every vector in $S_i$ is perpendicular to $e_1$, then $w \cdot \sum_ju_{ij}$ can be at most $|w-(w\cdot e_1)e_1|$, which is at most $(1-(w\cdot e_1)^2)^{1/2} \leq (1-1/N)^{1/2} \leq 1-1/(2N)$, contrary to the given fact. So there exists a vector $u_{ij_i}\in S_i$ such that $u_{ij_i}\cdot e_1>0$. Let there be a $k\neq j_i$ such that $u_{ik}\cdot e_1>0$. As all vectors are in the closed positive orthant, $u_{ij_i}\cdot u_{ik}\geq (u_{ij_i}\cdot e_1)(u_{ik}\cdot e_1) > 0$. This contradicts the fact that the vectors of $S_i$ are pairwise orthogonal. Hence we conclude that for each $i$ there exists a unique vector $u_{ij_i}\in S_i$ such that $u_{ij_i}\cdot e_1>0$. Thus $\sum_{j=1}^n u_{ij}\cdot e_1 = u_{ij_i}\cdot e_1$. Next we will show that $u_{ij_{i}}\cdot u_{kj_{k}} \geq 1/(16N^2)$ for all $i,k\in [n]$. For any $i$, from the given facts $1-1/(4N) < w\cdot \sum_ju_{ij} = (w\cdot e_1)(\sum_ju_{ij}\cdot e_1) + (w-(w\cdot e_1)e_1)\cdot (\sum_ju_{ij}-(\sum_ju_{ij}\cdot e_1)e_1)$. Since $(\sum_ju_{ij})^2 \leq 1$, $(w-(w\cdot e_1)e_1)\cdot (\sum_ju_{ij}- (\sum_ju_{ij}\cdot e_1)e_1) \leq |(w-(w\cdot e_1)e_1)| \leq 1-1/(2N)$. The last inequality has been established in the previous paragraph. So $(w\cdot e_1)(\sum_ju_{ij}\cdot e_1)\geq 1/(4N)$ for all $i$. Hence $u_{ij_i}\cdot e_1 = \sum_ju_{ij}\cdot e_1 \geq 1/(4N)$. All vectors of each $S_i$ are in the closed positive orthant hence $u_{ij_{i}}\cdot u_{kj_{k}} \geq (u_{ij_{i}}\cdot e_1)(u_{kj_{k}} \cdot e_1) \geq 1/(16N^2)$ for all $i,k\in [n]$. Thus the set $\{u_{1j_1},\dots,u_{nj_n}\}$ is pairwise non-orthogonal, and hence a complete consistent set. From the first paragraph of Section 4 we know that the permutation, $\sigma(i)=j_i$ for all $i$, is an isomorphism between $G_1$ and $G_2$. $\Box$ \end{proof} \begin{corollary} If $G_1$ and $G_2$ are non-isomorphic, then the value of the $cp\vartheta$-function of CP-LT must be less than $n- 1/(4n^4)$. \end{corollary} \section{Conclusion} We have seen that if $G_1$ and $G_2$ are isomorphic, then the $cp\vartheta$ function value is $n$. If the graphs are not isomorphic, then the function value remains less than $n-1/(4n^4)$. Moreover in isomorphic case the feasible region is exactly ${\cal P}_{G_1G_2}$. Since a completely positive program cannot be solved in polynomial time, the above is not a polynomial time test. We have the following theorem. \begin{theorem} The proposed completely positive $\vartheta$ function takes value $n$ when the graphs are isomorphic and it takes value less than $n-1/(4n^4)$ when the graphs are non-isomorphic. This gives a test for graph isomorphism. \end{theorem} \bibliographystyle{plain}

https://arxiv.org/abs/2106.03641

A Shape-Newton approach to the problem of covering with identical balls

The problem of covering a region of the plane with a fixed number of minimum-radius identical balls is studied in the present work. An explicit construction of bi-Lipschitz mappings is provided to model small perturbations of the union of balls. This allows us to obtain analytical expressions for first- and second-order derivatives using nonsmooth shape optimization techniques under appropriate regularity assumptions. Singular cases are also studied using asymptotic analysis. For the case of regions given by the union of disjoint convex polygons, algorithms based on Voronoi diagrams that do not rely on approximations are given to compute the derivatives. Extensive numerical experiments illustrate the capabilities and limitations of the introduced approach.

\section{Introduction} \label{intro} The problem of covering a region of the plane with a fixed number of minimum-radius identical balls is studied in the present work by expanding the nonsmooth shape optimization approach introduced in \cite{coveringfirst}. The main challenge in this previous work was the the first-order shape sensitivity analysis with respect to perturbations of the balls' centers and radii. Therefore, investigating the second-order shape sensitivity is a natural albeit challenging extension of \cite{coveringfirst}. Shape optimization is the study of optimization problems where the variable is a geometric object; see~\cite{MR2731611,MR3791463,MR1215733}. One of the key concepts in this discipline is the notion of \textit{shape derivative}, that measures the sensitivity of functions with respect to perturbations of the geometry. The theoretical study of second-order shape derivatives is a difficult topic in shape optimization. There exists an abundant literature on the shape Hessian in the smooth setting \cite{Dambrine2000,Dambrine2003,MR2731611,MR1215733}; while in the nonsmooth setting it is still an active research topic \cite{MR2148282,LAURAIN2020328}. Numerical methods based on second-order shape derivative are rarely used in shape optimization due to several difficulties. First of all, the second-order shape derivative is often difficult to compute and costly to implement numerically, especially when partial differential equations are involved. Second, the shape Hessian presents several theoretical issues, such as the {\it two norms-discrepancy} and lack of coercivity, that have been extensively studied in control problems; see \cite{Afraites2008,Dambrine2000} and the references therein. There exist only few attempts at defining numerical methods based on second-order information in shape optimization. In \cite{EH2005}, a regularized shape-Newton method is introduced to solve an inverse problem for star-shaped geometries. Second order preconditioning of the shape gradient has been used in \cite{Hintermller2004} for image segmentation and in \cite{doi:10.1080/10618569908940813,Schmidt2009} for aerodynamic optimization. Automatic shape differentiation has also been successfully employed to compute first- and second-order shape derivatives \cite{Ham2019,Schmidt2018}. We also observe that the numerical investigations using Newton-type algorithms \cite{EH2005,Hintermller2004} are set in a relatively smooth setting. In \cite{LAURAIN2020328}, the shape Hessian was calculated for nonsmooth geometries and polygons in a form that was convenient for numerical experiments, but no numerical investigations were performed. To the best of our knowledge, the present paper is the first attempt at designing and analyzing a shape-Newton algorithm in a genuinely nonsmooth setting. From a theoretical perspective, the main achievement of \cite{coveringfirst} was to build bi-Lipschitz transformations to model the geometry perturbations corresponding to covering with identical balls. In the present work, these transformations are key elements for the calculation of the second-order shape derivative, which, unlike the first-order shape derivative, differs from the expression that would be obtained in a smooth setting. Indeed, for the piecewise smooth shapes considered in the covering problem, various terms with a support at singular boundary points, typically circles intersection, appear in the shape Hessian. Due to the generality of the regions to be covered considered in~\cite{coveringfirst}, in the presented numerical experiments, the function that measures the covering and its first-order derivatives were approximated with discretization strategies that may by very time consuming if high precision is required. In the present work, by restricting the region to be covered to be the union of disjoint convex polygons, algorithms based on Voronoi diagrams to compute the covering function and its first- and second-order derivatives \textit{without} relying on approximations are given. The problem of covering a two-dimensional region with identical balls has already been considered in the literature. Covering equilateral triangles and squares was considered in~\cite{nurmelaT} and~\cite{nurmelaS}, respectively; while covering the union and difference of polygons was considered in~\cite{stoyan}. The covering of rectangles, triangles, squares and arbitrary regions was considered in~\cite{heppes}, \cite{melissenT1997}, \cite{melissenS1996} and~\cite{xavier2D}, respectively. However, the problem addressed in~\cite{xavier2D} actually consists of covering an arbitrary set of points, which is substantially different from the problem of covering an entire region. All of these papers approach the problem as an optimization problem. In~\cite{heppes,melissenT1997,melissenS1996} a simulated annealing approach with local search in which the centers of the balls are chosen as points on an adaptive mesh is considered. In~\cite{nurmelaT,nurmelaS}, a discrete rule is used to define the radius; while a BFGS method is used to solve subproblems in which the radius is fixed. A feasible direction method that requires solving a linear programming problem at each iteration was proposed in~\cite{stoyan}. None of the mentioned works addresses the problem in a unified way as a continuous optimization problem, nor do they present first- or second-order derivatives of the functions that define the problem. In~\cite{birgin}, the problem of covering an arbitrary region is modeled as a nonlinear semidefinite programming problem using convex algebraic geometry tools. The introduced model describes the covering problem without resorting to discretizations, but it depends on some polynomials of unknown degrees whose coefficients are difficult to compute, limiting the applicability of the method. The rest of this paper is organized as follows. Section~\ref{sec:shape_opt_covering} presents a formal definition of the problem, the formula for the first-order derivative introduced in~\cite{coveringfirst}, and the formula for the second-order derivative being introduced in the present work. Section~\ref{sec:D2G_proof} presents the derivation of the second-order derivatives for non-degenerate cases; while degenerate cases are considered in Section~\ref{sec:singularcases}. Algorithms based on Voronoi diagrams for the exact calculation of the covering function and its first- and second-order derivatives are introduced in Section~\ref{sec:calc_algorithms}. Extensive numerical experiments are given in Section~\ref{sec:num_exp}. Final considerations are given in Section~\ref{sec:conclusion}.\\ \noindent \textbf{Notation:} Given $x, y \in \mathds{R}^n$, $x \cdot y = x^\top y \in \mathds{R}$; while $x \otimes y = x y^\top \in \mathds{R}^{n \times n}$. The divergence of a sufficiently smooth vector field $\mathds{R}^2\ni (x,y)\mapsto V(x,y)=(V_1(x,y),V_2(x,y))\in\mathds{R}^2$ is defined by $\operatorname{div} V := \frac{\partial V_1}{\partial x} + \frac{\partial V_2}{\partial y}$, and its Jacobian matrix is denoted $DV$. Given an open set $S\in\mathds{R}^n$, $\overline{S}$ denotes its closure, $\partial S = \overline{S} \setminus S$ its boundary, and $\operatorname{Vol}(S)$ its volume. Let $B(x_i,r)$ denote an open ball with center $x_i\in\mathds{R}^2$ and radius $r$. For a sufficiently smooth set $S\subset\mathds{R}^2$, $\nu_S(z)$ denotes the unitary-norm outwards normal vector to $S$ at $z$ and $\tau_S(z)$ the unitary-norm tangent vector to $\partial S$ at~$z$ (pointing counter-clockwise). In the particular case $S = B(x_i,r)$ we use the simpler notation $\nu_i:=\nu_{B(x_i,r)}$ and $\tau_i := \tau_{B(x_i,r)}$, and we have $\nu_i(z) = (\cos \theta_z, \sin \theta_z)^\top$ and $\tau_i(z) = (-\sin \theta_z, \cos \theta_z)^\top$, where $\theta_z$ is the angular coordinate of $z-x_i$. For intersection points $z\in \partial S \cap B(x_i,r)$, we also use the notation $\nu_{-i}(z): = \nu_S(z)$. \section{The shape optimization problem}\label{sec:shape_opt_covering} Let $A\subset \mathds{R}^2$ and $\Omega(\boldsymbol{x},r) = \cup_{i=1}^m B(x_i,r)$ with $\boldsymbol{x}:= \{x_i\}_{i=1}^m$. We consider the problem of covering $A$ using a fixed number $m$ of identical balls $B(x_i, r)$ with minimum radius~$r$, i.e., we are looking for $(\boldsymbol{x}, r) \in \mathds{R}^{2m+1}$ such that $A \subset \Omega(\boldsymbol{x}, r)$ with minimum~$r$. The problem can be formulated as \begin{equation} \label{prob1} \Minimize_{(\boldsymbol{x},r) \in \mathds{R}^{2m+1}} \; r \; \mbox{ subject to } \; G(\boldsymbol{x},r) = 0, \end{equation} where \begin{align} \label{aquiG} G(\boldsymbol{x},r) &:= \operatorname{Vol}(A) - \operatorname{Vol}(A \cap \Omega(\boldsymbol{x},r)). \end{align} Note that $G(\boldsymbol{x},r) = 0$ if and only if $A\subset \Omega(\boldsymbol{x},r)$ up to a set of zero measure, i.e., when $ \Omega(\boldsymbol{x},r)$ covers $A$. The derivatives of~$G$ can be computed using techniques of shape calculus \cite{MR2731611,MR3791463,MR3535238,LAURAIN2020328,MR1215733}. In particular it was shown in~\cite{coveringfirst} that, under suitable assumptions, \begin{align}\label{gradG} \nabla G(\boldsymbol{x},r) = - \begin{pmatrix} \displaystyle \int_{\mathcal{A}_1} \nu_1(z) \, dz, \cdots, \displaystyle \int_{\mathcal{A}_m} \nu_m(z) \, dz, \displaystyle \int_{\partial \Omega(\boldsymbol{x},r)\cap A} \, dz \end{pmatrix}^\top, \end{align} where \begin{equation} \label{partition} \mathcal{A}_i = \partial B(x_i,r) \cap \partial \Omega(\boldsymbol{x},r) \cap A \end{equation} for $i=1,\dots,m$. In the present work, we show that \begin{equation} \label{derivadasegunda} \nabla^2 G (\boldsymbol{x},r) = \begin{pmatrix} \nabla^2_{\boldsymbol{x}} G (\boldsymbol{x},r) & \nabla^2_{\boldsymbol{x},r} G (\boldsymbol{x},r)\\ \nabla^2_{\boldsymbol{x},r} G (\boldsymbol{x},r)^\top & \nabla^2_r G (\boldsymbol{x},r) \end{pmatrix}, \end{equation} where $\nabla^2_{\boldsymbol{x}} G (\boldsymbol{x},r) \in \mathds{R}^{2m \times 2m}$, $\nabla^2_{\boldsymbol{x},r} G (\boldsymbol{x},r) \in \mathds{R}^{2m}$, and $\nabla^2_r G (\boldsymbol{x},r)= \partial_{r}^{2} G(\boldsymbol{x},r) \in \mathds{R}$ are described below. Their description is based on the fact that each set $\mathcal{A}_i$ can be represented by a finite number $m_i \geq 0$ of arcs of the circle $\partial B(x_i,r)$. Note that, since $(\cup_{i=1}^m \partial B(x_i,r)) \cap \partial \Omega(\boldsymbol{x},r) = \partial \Omega(\boldsymbol{x},r)$, by~\eqref{partition}, \begin{equation} \label{unionAi} \bigcup_{i=1}^m \mathcal{A}_i = \partial\Omega(\boldsymbol{x},r)\cap A, \end{equation} i.e., the union of all $\mathcal{A}_i$ represents a partition of $\partial\Omega(\boldsymbol{x},r)\cap A$; see Figure~\ref{fig:arcos}. Each arc in $\mathcal{A}_i$ can be represented by a pair of points $(v,w)$, named starting and ending points, in counter-clockwise direction, i.e., such that the angular coordinates $\theta_v$ and $\theta_w$ of $v-x_i$ and $w-x_i$, respectively, satisfy $\theta_v \in [0,2\pi)$ and $\theta_w \in (\theta_v,\theta_v+2\pi]$; see Figure~\ref{fig:sec2}. If $\mathcal{A}_i$ is not a full circle, we denote by $\mathbb{A}_i$ the set of pairs $(v,w)$ that represent the arcs in $\mathcal{A}_i$; otherwise, we define $\mathbb{A}_i=\emptyset$. In addition, if $\mathcal{A}_i$ is a full circle, then we set $\mathrm{Circle}(\mathbb{A}_i)$ equal to true; otherwise, we set $\mathrm{Circle}(\mathbb{A}_i)$ equal to false. We say a configuration $(\boldsymbol{x},r)$ is non-degenerate if, for every $i=1,\dots,m$, every $(v,w) \in \mathbb{A}_i$, and every $z \in \{v,w\}$, there exists one and only one $\nu_{-i}(z)$ and $\nu_{-i}(z) \cdot \tau_i(z) \neq 0$. A characterization of non-degenerate configurations, which satisfy Assumptions~\ref{a1} and~\ref{a3}, is given in the next section. \begin{figure}[ht!] \begin{center} \begin{tabular}{cc} \includegraphics{A_omega.mps} & \includegraphics{A_partialOmega.mps} \\ (a) & (b) \end{tabular} \end{center} \caption{(a) represents a region $A$ to be covered and an arbitrary configuration of balls $\Omega(\boldsymbol{x},r)$. (b) represents, in red, $\partial \Omega(\boldsymbol{x},r) \cap A$. Each $\mathcal{A}_i$ corresponds to the red arcs that intersect $\partial B(x_i,r)$. Note that, in this example, most sets $\mathcal{A}_i$ contain two or three maximal arcs; and there is only one set $\mathcal{A}_i$ with four maximal arcs.} \label{fig:arcos} \end{figure} Assuming~$(\boldsymbol{x},r)$ is non-degenerate, we have that $\nabla^2_r G(\boldsymbol{x},r)$ in~\eqref{derivadasegunda} is given by \begin{equation} \label{d2Grr} \nabla^2_r G(\boldsymbol{x},r) = - \frac{\mathrm{Per}(\partial\Omega(\boldsymbol{x},r)\cap A)}{r} - \sum_{i=1}^m \sum_{(v,w) \in \mathbb{A}_i} \left\llbracket \, \frac{|L(z)| - \nu_{-i}(z) \cdot \nu_i(z)}{\nu_{-i}(z) \cdot \tau_i(z)} \, \right\rrbracket_v^w, \end{equation} where, for an arbitrary expression $\Phi(z)$, $\llbracket \Phi(z) \rrbracket_v^w := \Phi(w) - \Phi(v)$, $\mathrm{Per}(S)$ denotes the perimeter of the set~$S$, and, for an extreme $z$ of an arc represented by $(v,w) \in \mathbb{A}_i$, $L(z)=\{ \ell \in \{1,\dots,m\} \setminus \{i\} \;|\; z \in \partial B(x_\ell,r) \}$. Matrix $\nabla^2_{\boldsymbol{x}} G (\boldsymbol{x},r)$ in~\eqref{derivadasegunda} is given by the $2 \times 2$ diagonal blocks \begin{equation} \label{d2xixiG} \partial^2_{x_i x_i} G(\boldsymbol{x},r) = \frac{1}{r} \int_{\mathcal{A}_i} - \nu_i(z) \otimes \nu_i(z) + \tau_i(z) \otimes \tau_i(z) \, dz + \sum_{(v,w) \in \mathbb{A}_i} \left\llbracket \, \frac{\nu_{-i}(z) \cdot \nu_i(z)}{\nu_{-i}(z) \cdot \tau_i(z)} \, \nu_i(z) \otimes \nu_i(z) \, \right\rrbracket_{v}^{w} \end{equation} and the $2 \times 2$ off-diagonal blocks \begin{equation} \label{d2Gxixj} \partial^2_{x_i x_\ell} G(\boldsymbol{x},r) = \sum_{v \in \mathcal{I}_{i\ell}} \frac{\nu_i(v) \otimes \nu_\ell(v)}{\nu_\ell(v) \cdot \tau_i(v)} - \sum_{w \in \mathcal{O}_{i\ell}} \frac{\nu_i(w) \otimes \nu_\ell(w)}{\nu_\ell(w) \cdot \tau_i(w)}, \end{equation} where ${\cal I}_{i\ell} = \{ v \in \partial B(x_\ell,r) \; | \; (v,\cdot) \in \mathbb{A}_i \}$ and $\mathcal{O}_{i\ell} = \{ w \in \partial B(x_\ell,r) \; | \; (\cdot,w) \in \mathbb{A}_i \}$. (Note that ${\cal I}_{i\ell}=\mathcal{O}_{i\ell}=\emptyset$ for all $\ell\neq i$ if $\mathbb{A}_i=\emptyset$.) Finally, array $\nabla^2_{\boldsymbol{x},r} G (\boldsymbol{x},r)$ in~\eqref{derivadasegunda} is given by the 2-dimensional arrays \begin{equation} \label{eq:781} \partial^2_{x_i r} G(\boldsymbol{x},r) = - \frac{1}{r} \int_{\mathcal{A}_i} \nu_i(z) \, dz + \sum_{(v,w) \in \mathbb{A}_i} \left\llbracket \, \frac{\nu_{-i}(z) \cdot \nu_i(z)}{\nu_{-i}(z) \cdot \tau_i(z)} \nu_i(z) - \sum_{\ell \in L(z)} \frac{\nu_i(z)}{\tau_i(z) \cdot \nu_{\ell}(z)} \, \right\rrbracket_v^w. \end{equation} \begin{figure} \centering \includegraphics{fig-secao2.mps} \caption{The set $\mathcal{A}_i = \partial B(x_i,r) \cap \Omega(\boldsymbol{x},r) \cap A$ is composed of two arcs (in red). If $z \in \partial B(x_i,r) \cap \partial B(x_{\ell},r)$ for some $\ell \neq i$, as for $z=w$, then $\nu_{-i}(z) = \nu_{\ell}(z)$, while if $z \in \partial B(x_i,r) \cap \partial A$, as for $z \in \{u,v\}$, then $\nu_{-i}(z) = \nu_{A}(z)$.} \label{fig:sec2} \end{figure} \section{Proof of second-order differentiability of \texorpdfstring{$G$}{G}} \label{sec:D2G_proof} In this section, we prove that the second-order derivatives of $G$, as defined in~\eqref{aquiG}, are given by (\ref{derivadasegunda}, \ref{d2Grr}, \ref{d2xixiG}, \ref{d2Gxixj}, \ref{eq:781}). In \cite{coveringfirst} we have built appropriate bi-Lipschitz mappings $T_t$ in order to use integration by substitution for the differentiation of $G(\boldsymbol{x}+t\delta\boldsymbol{x},r)$ and $G(\boldsymbol{x},r+t\delta r)$. Some of the more technical aspects of these constructions were related to the fact that $G(\boldsymbol{x},r)$ is an area functional, which required defining $T_t$ on $\Omega(\boldsymbol{x},r)\cap A$ and on $\partial(\Omega(\boldsymbol{x},r)\cap A)$. Since $\nabla G$ only involves boundary integrals that in addition can be decomposed into integrals on arcs, this facilitates the construction of the mappings $T_t$ required for the calculation of $\nabla^2 G (\boldsymbol{x},r)$, as $T_t$ only needs to be defined on $\partial \Omega(\boldsymbol{x},r)\cap A$. We consider two types of transformations for the shape sensitivity analysis. First, in the case of fixed radius and center perturbations one needs a mapping $T_t$ between the reference set $\partial\Omega(\boldsymbol{x},r)\cap A$ and the perturbed set $\partial\Omega(\boldsymbol{x} +t\delta \boldsymbol{x},r) \cap A$; see Theorem~\ref{thm:EB}. Second, in the case of fixed centers and radius perturbation one needs a mapping $T_t$ between the reference set $\partial\Omega(\boldsymbol{x}, r)\cap A$ and the perturbed set $\partial\Omega(\boldsymbol{x}, r+t\delta r)\cap A$; see Theorem~\ref{thm5}. The shape sensitivity analysis of $\nabla G$ is then achieved through integration by substitution using $T_t$. The construction of these mappings $T_t$ is similar to the constructions in \cite{coveringfirst}; however the results are presented in a different way as we need specific properties of $T_t$ to compute the derivatives of $\nabla G$. One of the main differences with respect to~\cite{coveringfirst} appears in Theorem~\ref{thm:EB}, where one considers a simultaneous perturbations of all the balls' center, which allows us to simplify the calculations of the Hessian of $G$. On the one hand, $T_t$ was used in \cite{coveringfirst} mainly to prove first-order shape differentiability and its unusual structure did not affect the expression of the first-order shape derivative, in the sense that a similar formula would have been obtained in a smooth setting. On the other hand, the expression of the second-order shape derivative of~$G$ at a nonsmooth reference domain $\Omega$ differs significantly from the expression that would be obtained for a smooth $\Omega$, as it involves terms with a support at singular boundary points of $\Omega$, and the particular structure of $T_t$ now plays an important role in the calculation of those singular terms. This can be understood by considering that, unlike the first-order derivative, the second-order shape derivative depends on the tangential component of $\left. \partial_t T_t\right|_{t= 0}$ on the boundary of the reference domain. In \cite{coveringfirst}, we have described detailed conditions to avoid degenerate situations and we also discussed various examples of such degeneracies and how they may affect the numerical algorithm. In the present paper we use the same conditions to prove second-order differentiability of $G$. To summarize, the main issues when studying the differentiability of $G$ arise when two balls are tangent or exactly superposed, when the boundaries of more than two balls intersect at the same point, or when $\Omega(\boldsymbol{x},r)$ and $A$ are not compatible in the sense of Definition \ref{def:compatible}. The role of Assumptions~\ref{a1} and \ref{a3} is to avoid these singular cases, which allows us to prove second-order differentiability of $G$. We emphasize that these assumptions only exclude a null-measure set of balls' configurations in $\mathds{R}^{2m+1}$, and in Section \ref{sec:singularcases} we show via the study of several singular cases that the second-order differentiability of~$G$ fails when these assumptions are not satisfied. \begin{assumption}\label{a1} The centers $\{x_i\}_{i=1}^m$ satisfy $\|x_i-x_j\|\notin \{0,2r\}$ for all $1\leq i,j\leq m$, $i\neq j$ and $\partial B(x_i,r)\cap \partial B(x_j,r) \cap \partial B(x_k,r) =\emptyset$ for all $1\leq i,j,k\leq m$ with $i,j,k$ pairwise distinct. \end{assumption} \begin{definition}\label{def:compatible} Let $\omega_1,\omega_2$ be open subsets of $\mathds{R}^2$. We call $\omega_1$ and $\omega_2$ {\it compatible} if $\omega_1\cap\omega_2\neq\emptyset$, $\omega_1$ and $\omega_2$ are Lipschitz domains, and the following conditions hold: (i) $\omega_1\cap\omega_2$ is a Lipschitz domain; (ii) $\partial\omega_1\cap \partial\omega_2$ is finite; (iii) $\partial\omega_1$ and $\partial\omega_2$ are locally smooth in a neighborhood of $\partial\omega_1\cap \partial\omega_2$; (iv) $\tau_1(x)\cdot \nu_2(x)\neq 0$ for all $x\in \partial\omega_1\cap \partial\omega_2$, where $\tau_1(x)$ is a tangent vector to $\partial\omega_1$ at $x$ and $\nu_2(x)$ is a normal vector to $\partial\omega_2$ at $x$. \end{definition} \begin{assumption}\label{a3} Sets $\Omega(\boldsymbol{x},r)$ and $A$ are compatible. \end{assumption} We observe that $\Omega(\boldsymbol{x},r)$ is Lipschitz under Assumption~\ref{a1}, and if, in addition, the intersection of $\partial\Omega(\boldsymbol{x},r)$ and $\partial A$ is empty, then Assumption~\ref{a3} holds. Hence, in this particular case we can drop Assumption~\ref{a3} in Theorems~\ref{thm:EB} and~\ref{thm5}. We also recall the following basic results, which are key ingredients for the calculation of the shape Hessian of $G$. \begin{thm}[Tangential divergence theorem]\label{lem:tandivpol} Let $\Gamma\subset\mathds{R}^2$ be a $C^k$ open curve, $k\geq 2$, with a parameterization $\gamma$, and denote $(v,w)$ the starting and ending points of $\Gamma$, respectively, with respect to $\gamma$. Let $\tau$ be the unitary-norm tangent vector to $\Gamma$, $\nu$ the unitary-norm normal vector to $\Gamma$, and $\mathcal{H}$ the mean curvature of $\Gamma$, with respect to the parameterization $\gamma$. Let $F\in W^{1,1}(\Gamma,\mathds{R}^2)\cap C^0(\overline{\Gamma},\mathds{R}^2)$, then we have $$ \int_{\Gamma} \operatorname{div}_\Gamma(F) = \int_{\Gamma} \mathcal{H} F\cdot \nu + F(w)\cdot \tau(w) - F(v) \cdot \tau(v) = \int_{\Gamma} \mathcal{H} F\cdot \nu + \llbracket F(z)\cdot \tau(z)\rrbracket_v^w,$$ where $\operatorname{div}_\Gamma(F):=\operatorname{div}(F) - DF\nu\cdot\nu$ is the tangential divergence of $F$ on $\Gamma$. \end{thm} \begin{proof} The result follows from \cite[\S~7.2]{MR756417} and \cite[Ch.~9, \S~5.5]{MR2731611}. \end{proof} \begin{lemma}[Integration by substitution for line integrals]\label{lem:intsub} Let $\Gamma\subset\mathds{R}^2$ be a $C^k$ open curve, $k\geq 2$, and $\nu$ a unitary-norm normal vector to $\Gamma$. Let $F\in C^0(\overline{\Gamma},\mathds{R}^2)$ and $T_t:\overline{\Gamma}\to T_t(\overline{\Gamma})$ be a bi-Lipschitz mapping. Then $$ \int_{T_t(\Gamma)} F(z)\, dz = \int_{\Gamma} F(T_t(z)) \omega_t, $$ where \begin{equation}\label{density:changevar} \omega_t(z) := \|M(z,t)\nu(z)\| \end{equation} and $M(z,t) := \det(DT_t(z)) DT_t(z)^{-\top}$ is the cofactor matrix of $DT_t(z)$. Furthermore, we have \begin{equation}\label{density:der} \partial_t\omega_t|_{t=0} = \operatorname{div}_\Gamma V \text{ with } V: = \partial_t T_t|_{t=0} \text{ on } \Gamma. \end{equation} \end{lemma} \begin{proof} See \cite[Prop.~5.4.3]{MR3791463}. \end{proof} \subsection{Construction of a perturbation field for center perturbations} Theorem \ref{thm:EB} below employs several ideas from \cite[Thm. 3.2 \& Thm. 3.6]{coveringfirst}. However, an important difference is that we consider simultaneous center perturbations for all balls instead of just one, which is more convenient for the calculation of $\nabla^2 G$. Theorem \ref{thm:EB} provides an appropriate mapping $T_t$ for the differentiation of $\partial_{x_i} G (\boldsymbol{x}+t\delta\boldsymbol{x},r)$ that will be used in Sections~\ref{sec:second_D_xiG} and~\ref{sec:derxixl} and for the differentiation of $\partial_{r} G (\boldsymbol{x}+t\delta\boldsymbol{x},r)$ in Section~\ref{sec:derxir}. \begin{thm}\label{thm:EB} Suppose that Assumptions~\ref{a1} and \ref{a3} hold. Then there exists $t_0>0$ such that for all $t\in [0,t_0]$ we have the following decomposition \begin{equation}\label{eq:dec_E} \partial \Omega(\boldsymbol{x}+t\delta\boldsymbol{x},r)\cap A = \bigcup_{k=1}^{\bar{k}} \mathcal{S}_k(t), \end{equation} where $\bar{k}$ is independent of $t$, $\mathcal{S}_k(t)$ are arcs parameterized by an angle aperture $[\theta_{k,v}(t), \theta_{k,w}(t) ]$, and $t\mapsto \theta_{k,v}(t)$, $t\mapsto \theta_{k,w}(t)$ are continuous functions on $[0,t_0]$. Also, for all $t\in [0,t_0]$ there exists a bi-Lipschitz mapping $T_t: \partial\Omega(\boldsymbol{x},r)\cap A\to \mathds{R}^2$ satisfying $T_t(\partial\Omega(\boldsymbol{x},r)\cap A) = \partial\Omega(\boldsymbol{x}+t\delta\boldsymbol{x},r)\cap A $ and $T_t(\mathcal{S}_k(0))=\mathcal{S}_k(t)$ for all $k=1,\dots,\bar{k}$. Furthermore, we have \begin{equation}\label{eq:Vtheta_arc} V := \left.\partial_t T_t\right|_{t= 0} =\delta x_i + \partial_t\xi(0,\theta) r \tau_i \text{ on } \mathcal{S}_k(0)\subset \partial B(x_i,r), \end{equation} where $\xi$ is defined in \eqref{114} and \begin{align}\label{eq:Vthetaa} V(z) & = \delta x_i - \frac{\nu_A(z)\cdot \delta x_i}{ \tau_i(z) \cdot \nu_A(z)}\tau_i(z) \quad \text{ if } z\in\partial B(x_i,r)\cap \partial A,\\ \label{eq:Vthetaa_2} V(z) & = \delta x_i - \frac{\nu_\ell(z)\cdot (\delta x_i-\delta x_\ell)}{ \tau_i(z) \cdot \nu_\ell(z)}\tau_i(z) \quad \text{ if } z\in\partial B(x_i,r)\cap \partial B(x_\ell,r), i\neq\ell. \end{align} \end{thm} \begin{proof} The decomposition \eqref{eq:dec_E} relies on Assumptions~\ref{a1} and \ref{a3} and is obtained in a similar way as in \cite[Thm.~3.2]{coveringfirst}. Therefore, in this proof we focus on the construction of the mapping $T_t$. We observe that each extremity of the arcs $\mathcal{S}_k(t)$ in the decomposition \eqref{eq:dec_E} is either a point belonging to $\partial B(x_i+t\delta x_i,r)\cap \partial A$ or a point in $\partial B(x_i+t\delta x_i,r)\cap\partial B(x_\ell+t\delta x_\ell,r)$. We first provide a general formula for the angle $\vartheta(t)$, in local polar coordinates with the pole $x_i+t\delta x_i$, describing an intersection point between the circle $\partial B(x_i+t\delta x_i,r)$ and $\partial A$. Let $z\in\partial B(x_i,r)\cap \partial A$ and $\phi$ be the oriented distance function to $A$, defined as $ \phi(x) := d(x,A) - d(x,A^c)$, where $d(x,A)$ is the distance from $x$ to the set $A$. Since $\Omega(\boldsymbol{x},r)$ and $A$ are compatible due to Assumption~\ref{a3}, it follows that $\partial A$ is locally smooth around the points $\partial B(x_i,r)\cap \partial A$, hence there exists a neighborhood $U_z$ of $z$ such that the restriction of $\phi$ to $U_z$ is smooth, $\phi(x) = 0$ and $\|\nabla \phi(x)\| = 1$ for all $x\in\partial A \cap U_z$. Let $(r,\theta_z)$ denote the polar coordinates of $z$, with the pole $x_i$. Introduce the function $$\psi(t,\vartheta) = \phi\left( x_i +t\delta x_i + r \begin{pmatrix} \cos\vartheta \\ \sin\vartheta \end{pmatrix}\right) .$$ We compute $$\partial_{\vartheta} \psi(0,\theta_z) = r\begin{pmatrix} -\sin\theta_z \\ \cos\theta_z \end{pmatrix} \cdot \nabla \phi\left( x_i + r \begin{pmatrix} \cos\theta_z \\ \sin\theta_z \end{pmatrix}\right) = r \tau_i(z) \cdot \nabla \phi(z).$$ Since $\Omega(\boldsymbol{x},r)$ and $A$ are compatible, $B(x_i,r)$ is not tangent to $\partial A$ and using $\|\nabla \phi(z)\| = 1$ we obtain $\tau_i(z) \cdot \nabla \phi(z) \neq 0$. Thus, we can apply the implicit function theorem and this yields the existence of a smooth function $[0,t_0]\ni t\mapsto \vartheta(t)$ with $\psi(t,\vartheta(t)) = 0$ and $\vartheta(0) = \theta_z$. We also compute, using $\nabla \phi(z) = \|\nabla \phi(z)\| \nu_A(z)$ since $\phi$ is the oriented distance function to $\partial A$, \begin{equation}\label{der_vartheta0} \vartheta'(0) = -\frac{\partial_t \psi(0, \vartheta(0))}{\partial_\vartheta \psi(0, \vartheta(0))} = - \frac{\nabla \phi(z)\cdot \delta x_i}{r \tau_i(z) \cdot \nabla \phi(z)} = - \frac{\nu_A(z)\cdot \delta x_i}{r \tau_i(z) \cdot \nu_A(z)}. \end{equation} We now consider the second case of an intersection point in $\partial B(x_i+t\delta x_i,r)\cap\partial B(x_j+t\delta x_\ell,r)$, $i\neq \ell$. Introduce the functions $$\psi(t,\vartheta) = \|\zeta(t,\vartheta)\|^2 - r^2 \ \text{ with }\ \zeta(t,\vartheta) = x_i +t\delta x_i - x_\ell -t\delta x_\ell + r \begin{pmatrix} \cos\vartheta \\ \sin\vartheta \end{pmatrix}.$$ Observe that $\vartheta \mapsto \zeta(t,\vartheta)$ is a parameterization of the circle $\partial B(x_i+t\delta x_i,r)$ in a coordinate system of center $x_\ell$, which means that the solutions of the equation $\psi(t,\vartheta)=0$ describe the intersections between $\partial B(x_i+t\delta x_i,r)$ and $\partial B(x_\ell+t\delta x_\ell,r)$. We compute $\partial_{\vartheta} \psi(0,\vartheta) = 2 \zeta(0,\vartheta) \cdot \partial_{\vartheta} \zeta(0,\vartheta) $ with \begin{align*} \zeta(0,\vartheta) &= x_i - x_\ell + r \begin{pmatrix} \cos\vartheta \\ \sin\vartheta \end{pmatrix} \quad \text{ and }\quad\partial_{\vartheta} \zeta(0,\vartheta) =r \begin{pmatrix} -\sin\vartheta \\ \cos\vartheta \end{pmatrix}. \end{align*} Now let $z\in\partial B(x_i,r)\cap\partial B(x_\ell,r)$ and let $\theta_z$ be the corresponding angle in a polar coordinate system with pole $x_i$. Since Assumption~\ref{a1} is satisfied, it is easy to see that \begin{equation*} \partial_{\vartheta} \psi(0,\theta_z) = 2 \zeta(0,\theta_z) \cdot \partial_{\vartheta} \zeta(0,\theta_z) \neq 0. \end{equation*} Hence, the implicit function theorem can be applied to $(t,\vartheta) \mapsto \psi(t,\vartheta)$ in a neighbourhood of $(0,\theta_z)$. This yields the existence, for $t_0$ sufficiently small, of a smooth function $t\mapsto\vartheta(t)$ in $[0,t_0]$ such that $\psi(t,\vartheta(t)) = 0$ in $[0,t_0]$ and $\vartheta(0) = \theta_z$. We also have the derivative \begin{equation*} \vartheta'(t) = -\frac{\partial_t \psi(t,\vartheta(t))}{\partial_{\vartheta} \psi(t,\vartheta(t))} = -\frac{ \zeta(t,\vartheta(t)) \cdot \partial_t \zeta(t,\vartheta(t)) }{ \zeta(t,\vartheta(t)) \cdot \partial_{\vartheta} \zeta(t,\vartheta(t)) }, \end{equation*} and in particular, using $\nu_i = (\cos\theta_z, \sin\theta_z)^\top$ and $\tau_i = (-\sin\theta_z, \cos\theta_z)^\top$, \begin{equation}\label{theta_prime_a_0} \vartheta'(0) = -\frac{ ( x_i - x_\ell + r \nu_i)\cdot (\delta x_i -\delta x_\ell)}{ ( x_i - x_\ell + r\nu_i)\cdot (r \tau_i)} = -\frac{ \nu_\ell\cdot (\delta x_i - \delta x_\ell)}{ r \nu_\ell\cdot \tau_i}. \end{equation} We are now ready to build the mapping $T_t$. Let $\mathcal{S}(t)\subset \partial B(x_i+t\delta x_i,r)$ be one of the arcs parameterized by the angle aperture $[\theta_v(t), \theta_w(t) ]$ in the decomposition~\eqref{eq:dec_E}; we have dropped the index $k$ for simplicity. Then, $\theta_v(t)$ and $\theta_w(t)$ are given by $\vartheta(t)$ with either $\theta_z=\theta_v(0)$ or $\theta_z=\theta_w(0)$, and $\vartheta(t)$ either corresponds to an intersection $\partial B(x_i+t\delta x_i,r)\cap \partial A$ or to an intersection $\partial B(x_i+t\delta x_i,r)\cap\partial B(x_\ell+t\delta x_\ell,r)$. Thus we define $T_t$ on the arc $\mathcal{S}(0)$ as \begin{equation}\label{113} T_t(x) := x_i + t\delta x_i +r \begin{pmatrix} \cos\xi(t,\theta) \\ \sin\xi(t,\theta) \end{pmatrix} \text{ with } x = x_i +r \begin{pmatrix} \cos\theta \\ \sin\theta \end{pmatrix} \in \mathcal{S}(0) , \end{equation} where \begin{equation}\label{114} \resizebox{0.9\textwidth}{!}{$ \xi(t,\theta) := \alpha(t) (\theta - \theta_w(0)) + \theta_w(t) \text{ for } (t,\theta)\in [0,t_0]\times [\theta_v(0),\theta_w(0)] \text{ and } \alpha(t) := \frac{\theta_w(t) - \theta_v(t)}{\theta_w(0) - \theta_v(0)}. $} \end{equation} The bi-Lipschitz property of $T_t$ on $\partial\Omega(\boldsymbol{x},r)\cap A$ is obtained as in the proof of \cite[Thm.~3.3]{coveringfirst}. Finally, differentiating in \eqref{113} with respect to $t$ and using $\xi(0,\theta)=\theta$ we get \eqref{eq:Vtheta_arc}. Then \eqref{114} yields $\xi(t,\theta_v(0)) = \theta_v(t)$, $\xi(t,\theta_w(0)) = \theta_w(t)$, $\partial_t\xi(0,\theta_v(0)) = \theta'_a(0)$, $ \partial_t\xi(0,\theta_w(0)) = \theta'_b(0)$, consequently using \eqref{der_vartheta0} we obtain \eqref{eq:Vthetaa} and using \eqref{theta_prime_a_0} we obtain \eqref{eq:Vthetaa_2}. \end{proof} \subsection{Construction of a perturbation field for radius perturbations}\label{sec:der_rr} Theorem~\ref{thm5} below relies on several ideas from \cite[Thm. 3.3 \& Thm. 3.8]{coveringfirst}, and provides an appropriate mapping $T_t$ for the differentiation of $\partial_{r} G (\boldsymbol{x},r+t\delta r)$ that will be used in Section~\ref{sec:second_D_rG_fast}. \begin{thm}\label{thm5} Suppose that Assumptions~\ref{a1} and \ref{a3} hold. Then there exists $t_0>0$ such that for all $t\in [0,t_0]$ we have the following decomposition \begin{equation}\label{eq:dec_Eradius} \partial \Omega(\boldsymbol{x},r+t\delta r)\cap A = \bigcup_{k=1}^{\bar{k}} \mathcal{S}_k(t), \end{equation} where $\bar{k}$ is independent of $t$, $\mathcal{S}_k(t)$ are arcs parameterized by an angle aperture $[\theta_{k,v}(t), \theta_{k,w}(t) ]$, and $t\mapsto \theta_{k,v}(t)$, $t\mapsto \theta_{k,w}(t)$ are continuous functions on $[0,t_0]$. Also, for all $t\in [0,t_0]$ there exists a bi-Lipschitz mapping $T_t: \partial\Omega(\boldsymbol{x},r)\cap A\to \mathds{R}^2$ satisfying $T_t(\Omega(\boldsymbol{x},r)\cap A) = \partial\Omega(\boldsymbol{x},r+t\delta r)\cap A $ and $T_t(\mathcal{S}_k(0))=\mathcal{S}_k(t)$ for all $k=1,\dots,\bar{k}$. In addition, we have \begin{equation}\label{eq:Vtheta_arc_radius} V := \left.\partial_t T_t\right|_{t= 0} =\delta r \nu_i + \partial_t\xi(0,\theta) r \tau_i \text{ on } \mathcal{S}_k(0)\subset \partial B(x_i,r), \end{equation} where $\xi$ is defined in \eqref{114} and \begin{align}\label{eq:Vthetaa_radius} V(z) & = \delta r \nu_i(z) - \delta r\frac{\nu_A(z)\cdot \nu_i(z)}{ \tau_i(z) \cdot \nu_A(z)}\tau_i(z) \quad \text{ if } z\in\partial B(x_i,r)\cap \partial A,\\ \label{eq:Vthetaa_2_radius} V(z) & = \delta r \nu_i(z) + \delta r \frac{ 1 - \nu_\ell(z)\cdot\nu_i(z)}{ \tau_i(z) \cdot \nu_\ell(z)}\tau_i(z) \quad \text{ if } z\in\partial B(x_i,r)\cap \partial B(x_\ell,r), i\neq\ell. \end{align} \end{thm} \begin{proof} The proof has the same structure as the proof of Theorem \ref{thm:EB}, i.e., we separate the two cases of a point belonging to $\partial B(x_i,r+t\delta r)\cap \partial A$ and a point in $\partial B(x_i,r+t\delta r)\cap\partial B(x_\ell,r+t\delta r)$. The decomposition \eqref{eq:dec_Eradius} relies on Assumptions~\ref{a1} and \ref{a3} and is obtained in a similar way as in \cite[Thm.~3.2]{coveringfirst}. First we consider the case of a point in $\partial B(x_i,r+t\delta r)\cap \partial A$. We provide a general formula for the angle $\vartheta(t)$, in local polar coordinates with pole $x_i$, describing such an intersection point. Let $z\in\partial B(x_i,r)\cap \partial A$ and $(r,\theta_z)$ denote the polar coordinates of $z$ with center $x_i$. Let $\phi$ be the oriented distance function to $A$ defined as in the proof of Theorem \ref{thm:EB}. Introduce the function $$\psi(t,\vartheta) = \phi\left( x_i + (r+t\delta r) \begin{pmatrix} \cos\vartheta \\ \sin\vartheta \end{pmatrix}\right) .$$ We compute $$\partial_{\vartheta} \psi(0,\theta_z) = r\begin{pmatrix} -\sin\theta_z \\ \cos\theta_z \end{pmatrix} \cdot \nabla \phi\left( x_i + r \begin{pmatrix} \cos\theta_z \\ \sin\theta_z \end{pmatrix}\right) = r \tau_i(z) \cdot \nabla \phi(z).$$ Since $\Omega(\boldsymbol{x},r)$ and $A$ are compatible due to Assumption~\ref{a3}, $B(x_i,r)$ is not tangent to $\partial A$ and using $\|\nabla \phi(z)\| = 1$ we obtain $\tau_i(z) \cdot \nabla \phi(z) \neq 0$. Thus, we can apply the implicit function theorem and this yields the existence of a smooth function $[0,t_0]\ni t\mapsto \vartheta(t)$ with $\psi(t,\vartheta(t)) = 0$ and $\vartheta(0) = \theta_z$. We also compute the derivative \begin{equation}\label{der_vartheta} \vartheta'(0) = -\frac{\partial_t \psi(0, \vartheta(0))}{\partial_\vartheta \psi(0, \vartheta(0))} = -\frac{\partial_t \psi(0, \vartheta(0))}{\partial_\vartheta \psi(0, \vartheta(0))} = - \frac{\delta r \nabla \phi(z)\cdot \nu_i(z)}{r \tau_i(z) \cdot \nabla \phi(z)} = - \frac{\delta r\nu_A(z)\cdot \nu_i(z)}{r \tau_i(z) \cdot \nu_A(z)}, \end{equation} where we have used $\nabla \phi(z) = \|\nabla \phi(z)\| \nu_A(z)$ since $\phi$ is the oriented distance function to $\partial A$. Now we provide a general formula for the angle $\vartheta(t)$, in local polar coordinates with pole $x_i$, describing an intersection point of two circles $\partial B(x_i,r + t\delta r)$ and $\partial B(x_\ell,r + t\delta r)$, $i\neq \ell$. Introduce $$\psi(t,\vartheta) = \|\zeta(t,\vartheta)\|^2 - (r+t\delta r)^2 \ \text{ with }\ \zeta(t,\vartheta) = x_i - x_\ell + (r + t\delta r) \begin{pmatrix} \cos\vartheta \\ \sin\vartheta \end{pmatrix}.$$ Observe that $\vartheta \mapsto \zeta(t,\vartheta)$ is a parameterization of the circle $\partial B(x_i,r + t\delta r)$ in a coordinate system of center $x_\ell$, which means that the solutions of $\psi(t,\vartheta)=0$ describe the intersections between $\partial B(x_i,r + t\delta r)$ and $\partial B(x_\ell,r + t\delta r)$. We compute $\partial_{\vartheta} \psi(0,\vartheta) = 2 \zeta(0,\vartheta) \cdot \partial_{\vartheta} \zeta(0,\vartheta) $ with \begin{align*} \zeta(0,\vartheta) &= x_i - x_\ell + r \begin{pmatrix} \cos\vartheta \\ \sin\vartheta \end{pmatrix} \quad \text{ and }\quad\partial_{\vartheta} \zeta(0,\vartheta) =r \begin{pmatrix} -\sin\vartheta \\ \cos\vartheta \end{pmatrix}. \end{align*} Now let $z\in\partial B(x_i,r)\cap\partial B(x_\ell,r)$ and let $\theta_z$ be the corresponding angle in a polar coordinate system with pole $x_i$. Since the conditions of Assumption~\ref{a1} hold, it is easy to see that \begin{equation*} \partial_{\vartheta} \psi(0,\theta_z) = 2 \zeta(0,\theta_z) \cdot \partial_{\vartheta} \zeta(0,\theta_z) \neq 0. \end{equation*} Hence, the implicit function theorem can be applied to $(t,\vartheta) \mapsto \psi(t,\vartheta)$ in a neighbourhood of $(0,\theta_z)$. This yields the existence, for $t_0$ sufficiently small, of a smooth function $t\mapsto\vartheta(t)$ in $[0,t_0]$ such that $\psi(t,\vartheta(t)) = 0$ in $[0,t_0]$ and $\vartheta(0) = \theta_z$. We also have the derivative \begin{equation*} \vartheta'(t) = -\frac{\partial_t \psi(t,\vartheta(t))}{\partial_{\vartheta} \psi(t,\vartheta(t))} = -\frac{ \zeta(t,\vartheta(t)) \cdot \partial_t \zeta(t,\vartheta(t)) - (r+t\delta r)\delta r}{ \zeta(t,\vartheta(t)) \cdot \partial_{\vartheta} \zeta(t,\vartheta(t)) }, \end{equation*} and in particular, using $\nu_i = (\cos\theta_z, \sin\theta_z)^\top$ and $\tau_i = (-\sin\theta_z, \cos\theta_z)^\top$, \begin{equation}\label{theta_prime_a_2} \vartheta'(0) = -\frac{ ( x_i - x_\ell + r \nu_i)\cdot (\delta r\nu_i) - r\delta r}{ ( x_i - x_\ell + r \nu_i)\cdot (r \tau_i)} = \frac{\delta r}{r} \left( \frac{1 - \nu_\ell\cdot \nu_i }{ \nu_\ell\cdot \tau_i} \right). \end{equation} We are now ready to build the mapping $T_t$. Let $\mathcal{S}(t)\subset \partial B(x_i,r+t\delta r)$ be one of the arcs parameterized by the angle aperture $[\theta_v(t), \theta_w(t) ]$ in the decomposition~\eqref{eq:dec_Eradius}; we have dropped the index $k$ for simplicity. Then, $\theta_v(t)$ and $\theta_w(t)$ are given by $\vartheta(t)$ with either $\theta_z=\theta_v(0)$ or $\theta_z=\theta_w(0)$, and $\vartheta(t)$ either corresponds to a point in $\partial B(x_i,r+t\delta r)\cap \partial A$ or to a point in $\partial B(x_i,r+t\delta r)\cap\partial B(x_\ell,r+t\delta r)$. Next, define $\xi(t,\theta)$ and $\alpha(t)$ as in \eqref{114}. Then, for $\theta\in [\theta_v(0),\theta_w(0)]$ we have $\xi(t,\theta) \in [\theta_v(t),\theta_w(t)]$ and $\xi(t,\theta)$ is a parameterization of $\mathcal{S}(t)$. A point $x\in \mathcal{S}(0)$ may be parameterized by \begin{equation}\label{eq:Tt_r} x = x_i +r \begin{pmatrix} \cos\theta \\ \sin\theta \end{pmatrix}, \ \text{ and define }\ \mathds{T}_t(\theta) := x_i + (r+t\delta r) \begin{pmatrix} \cos\xi(t,\theta) \\ \sin\xi(t,\theta) \end{pmatrix} . \end{equation} Writing $\xi(t,\theta) = \theta + \beta(t,\theta)$ with $\beta(t,\theta) := (\alpha(t) -1)(\theta -\theta_b(t))$, we observe that $$ \begin{pmatrix} \cos\xi(t,\theta) \\ \sin\xi(t,\theta) \end{pmatrix} = R(x_i, \beta(t,\theta)) \begin{pmatrix} \cos\theta \\ \sin\theta \end{pmatrix} = R(x_i, \beta(t,\theta)) \nu_i,$$ where $R(x_i, \beta(t,\theta))$ is a rotation matrix of center $x_i$ and angle $\beta(t,\theta)$. Also, thanks to $\theta_v(0)<\theta_w(0)<\theta_v(0)+ 2\pi$ and $\theta\in [\theta_v(0),\theta_w(0)]$, there exists a smooth bijection $\theta:\mathcal{A}\ni x\mapsto \theta(x)\in [\theta_v(0),\theta_w(0)]$. Thus, using \eqref{eq:Tt_r} we can define the mapping \begin{equation}\label{eq:Ttarc} T_t(x) := \mathds{T}_t(\theta(x)) = x - r \nu_i(x) + (r+t\delta r)R(x_i, \beta(t,\theta(x))) \nu_i(x) \text{ for all }x\in \mathcal{S}(0)\subset \partial B(x_i,r). \end{equation} The bi-Lipschitz property of $T_t$ on $\partial\Omega(\boldsymbol{x},r)\cap A$ can be obtained as in the proof of \cite[Thm.~3.3]{coveringfirst}. Finally, differentiating in \eqref{eq:Tt_r} with respect to $t$ and using $\xi(0,\theta)=\theta$ we get \eqref{eq:Vtheta_arc_radius}. Then \eqref{114} yields $\xi(t,\theta_a(0)) = \theta_a(t)$, $\xi(t,\theta_b(0)) = \theta_b(t)$, $\partial_t\xi(0,\theta_a(0)) = \theta'_a(0)$, $ \partial_t\xi(0,\theta_b(0)) = \theta'_b(0)$, consequently using \eqref{der_vartheta} we obtain \eqref{eq:Vthetaa_radius} and using \eqref{theta_prime_a_2} we obtain \eqref{eq:Vthetaa_2_radius}. \end{proof} \subsection{Second-order derivative of \texorpdfstring{$G$}{G} with respect to the radius}\label{sec:second_D_rG_fast} The first-order derivative of $G$ with respect to the radius is given by $$\partial_r G(\boldsymbol{x},r) = -\int_{\partial\Omega(\boldsymbol{x},r)\cap A} \, dz,$$ see \eqref{gradG} and \cite[\S3.3]{coveringfirst} for the detailed calculation. As in \cite{coveringfirst}, the calculation is achieved through integration by substitution using the mapping $T_t$ given by Theorem \ref{thm5}, which requires that Assumption~\ref{a1} and Assumption~\ref{a3} hold. According to Theorem \ref{thm5}, there exists a bi-Lipschitz mapping $T_t$ satisfying $T_t(\partial\Omega(\boldsymbol{x},r)\cap A) = \partial \Omega(\boldsymbol{x}, r+t\delta r)\cap A$, and this yields, using Lemma \ref{lem:intsub} on each arc of $\partial\Omega(\boldsymbol{x},r)\cap A$, \begin{align*} \partial_r G(\boldsymbol{x},r + t\delta r) &= - \int_{\partial\Omega(\boldsymbol{x},r+ t\delta r)\cap A}\, dz = - \int_{T_t(\partial\Omega(\boldsymbol{x},r)\cap A)} \, dz = - \int_{\partial\Omega(\boldsymbol{x},r)\cap A} \omega_t(z) \, dz. \end{align*} Thus, using Lemma \ref{lem:intsub} and the decomposition \eqref{unionAi}, we compute \begin{align*} \left.\frac{d}{dt} \partial_r G(\boldsymbol{x},r + t\delta r)\right|_{t= 0} & = - \int_{\partial\Omega(\boldsymbol{x},r)\cap A} \operatorname{div}_\Gamma V(z) \, dz = - \sum_{i=1}^m \int_{\mathcal{A}_i} \operatorname{div}_\Gamma V(z) \, dz. \end{align*} Applying Theorem \ref{lem:tandivpol} for each arc in $\mathcal{A}_i$, we obtain \begin{align}\label{eq:333} \left.\frac{d}{dt} \partial_r G(\boldsymbol{x},r + t\delta r)\right|_{t= 0} & = - \sum_{i=1}^m \int_{\mathcal{A}_i} \mathcal{H} V\cdot \nu_i \, dz - \sum_{i=1}^m \sum_{(v,w) \in \mathbb{A}_i} \llbracket V(z)\cdot \tau_i(z)\rrbracket_v^w. \end{align} To get a more explicit formula we need to determine $V(v),V(w)$ and $V\cdot \nu_i$ on $\mathcal{A}_i$. For this we apply Theorem \ref{thm5} to two different cases. On the one hand, if $v\in\partial B(x_i,r)\cap \partial B(x_\ell,r)$ for some $i\neq\ell$, then applying \eqref{eq:Vthetaa_2_radius} we obtain \begin{align}\label{zkin} V(v)\cdot \tau_i(v) = \delta r \frac{1 - \nu_\ell(v)\cdot \nu_i(v)}{\nu_\ell(v)\cdot \tau_i(v)}. \end{align} On the other hand, if $v\in\partial B(x_i,r)\cap \partial A$, then applying \eqref{eq:Vthetaa_radius} we get \begin{equation}\label{zkin2} V(v)\cdot \tau_i(v) = - \delta r\frac{\nu_A(v)\cdot \nu_i(v)}{\tau_i(v) \cdot \nu_A(v)}. \end{equation} Then, recalling that $L(z)=\{ \ell \in \{1,\dots,m\} \setminus \{i\} \;|\; z \in \partial B(x_\ell,r) \}$ for $z \in \{v,w\}$, and that $\nu_{-i}(z):=\nu_\ell(z)$ if $z\in\partial B(x_i,r)\cap \partial B(x_\ell,r)$, $\ell\neq i$, and $\nu_{-i}(z):=\nu_A(z)$ if $z\in\partial B(x_i,r)\cap \partial A$, we can merge \eqref{zkin} and \eqref{zkin2} into a unique formula: \begin{equation}\label{zkin3} V(v)\cdot \tau_i(v) = \delta r\frac{|L(v)| - \nu_{-i}(v)\cdot \nu_i(v)}{\nu_{-i}(v)\cdot\tau_i(v)}. \end{equation} In a similar way, we also obtain \begin{equation*} V(w)\cdot \tau_i(w) = \delta r\frac{|L(w)| - \nu_{-i}(w)\cdot \nu_i(w)}{\nu_{-i(w)}\cdot\tau_i(w)}. \end{equation*} Gathering these results we get \begin{align*} \left.\frac{d}{dt} \partial_r G(\boldsymbol{x},r + t\delta r)\right|_{t= 0} & = - \delta r\frac{\mathrm{Per}(\partial\Omega(\boldsymbol{x},r)\cap A)}{r} - \delta r \sum_{i=1}^m \sum_{(v,w) \in \mathbb{A}_i} \left\llbracket \, \frac{|L(z)| - \nu_{-i}(z) \cdot \nu_i(z)}{\nu_{-i}(z) \cdot \tau_i(z)} \, \right\rrbracket_v^w, \end{align*} where we have used $\mathcal{H} V\cdot \nu_i = \frac{\delta r}{r} \text{ on } \mathcal{A}_i\subset\partial B(x_i,r)$ due to \eqref{eq:Vtheta_arc_radius} and $\mathcal{H}=1/r$. Thus we have obtained \eqref{d2Grr}. \subsection{Second-order derivative of \texorpdfstring{$G$}{G} with respect to the centers}\label{sec:second_D_xiG} The first-order derivative of $G$ with respect to the center $x_i$ is given by $$\partial_{x_i} G(\boldsymbol{x},r) = - \int_{\mathcal{A}_i} \nu_i(z) \, dz.$$ see \eqref{gradG} and \cite[\S~3.4]{coveringfirst} for the detailed calculation. As in \cite{coveringfirst}, the calculation is achieved through integration by substitution using the mapping $T_t$ from Theorem \ref{thm:EB} with the specific perturbation $\delta \boldsymbol{x} = (0,\dots,0,\delta x_i, 0,\dots, 0)$, which requires that Assumption~\ref{a1} and Assumption~\ref{a3} hold. Using Lemma \ref{lem:intsub} yields \begin{align*} \partial_{x_i} G(\boldsymbol{x} + t\delta \boldsymbol{x},r) &= - \int_{\partial B(x_i+ t\delta x_i,r)\cap \partial\Omega(\boldsymbol{x}+ t\delta \boldsymbol{x},r)\cap A} \nu_t(z) \, dz = - \int_{T_t(\mathcal{A}_i)} \nu_t(z) \, dz\\ & = - \int_{\mathcal{A}_i} \nu_t(T_t(z)) \omega_t(z) \, dz, \end{align*} where $\nu_t$ is the outward unit normal vector to $\partial B(x_i+ t\delta x_i,r)\cap \partial\Omega(\boldsymbol{x}+ t\delta \boldsymbol{x},r)\cap A$ and $\omega_t$ is given by \eqref{density:changevar}. To obtain the derivative of $\partial_{x_i} G(\boldsymbol{x} + t\delta \boldsymbol{x},r) $ with respect to $t$ at $t=0$ we need the so-called {\it material derivative} of the normal vector given by $$\frac{d}{dt}\nu_t(T_t(z))|_{t=0} = - (D_\Gamma V)^\top \nu_i \text{ on } \mathcal{A}_i,$$ with $V := \partial_t T_t|_{t= 0}$; see [Walker, Lemma 5.5, page 99]. Here, $D_\Gamma V: = DV - (DV)\nu_i\otimes\nu_i$ denotes the tangential Jacobian of $V$ on $\mathcal{A}_i$. Then, using \eqref{density:der} we obtain \begin{align*} \left.\frac{d}{dt} \partial_{x_i} G(\boldsymbol{x} + t\delta \boldsymbol{x},r) \right|_{t= 0} & = - \int_{\mathcal{A}_i} - (D_\Gamma V)^\top \nu_i + \nu_i\operatorname{div}_\Gamma(V) \, dz. \end{align*} This expression can be further transformed using the following tensor relations: \begin{align} \label{tensor:rel}\operatorname{div}_\Gamma(\nu_i\otimes V) = \operatorname{div}_\Gamma (V)\nu_i + (D_\Gamma \nu_i)V \text{ and } \nabla_\Gamma (V\cdot\nu_i) = D_\Gamma \nu_i^\top V + D_\Gamma V^\top \nu_i \text{ on } \mathcal{A}_i. \end{align} We show that $D_\Gamma \nu_i^\top V = D_\Gamma \nu_i V$ on $\mathcal{A}_i$. Indeed, let $W\in\mathds{R}^2$ and denote $V_\tau$ and $W_\tau$ the tangential components of $V$ and $W$ on $\mathcal{A}_i$. Differentiating $\nu_i\cdot\nu_i = 1$ on $\mathcal{A}_i$ we get $(D_\Gamma \nu_i)^\top \nu_i = 0$ and then \begin{align*} (D_\Gamma \nu_i)^\top V \cdot W = (D_\Gamma \nu_i)^\top V_\tau \cdot W = (D_\Gamma \nu_i)^\top V_\tau \cdot W_\tau = (D_\Gamma \nu_i) V_\tau \cdot W_\tau, \end{align*} where we have used the well-known fact that the second fundamental form $(V_\tau,W_\tau)\mapsto (D_\Gamma \nu_i) V_\tau \cdot W_\tau$ is symmetric. Further, \begin{align}\label{sym:dgamma_nu} (D_\Gamma \nu_i)^\top V \cdot W = (D_\Gamma \nu_i)^\top W_\tau \cdot V_\tau = (D_\Gamma \nu_i)^\top W \cdot V = (D_\Gamma \nu_i) V \cdot W \text{ on } \mathcal{A}_i. \end{align} Now, using \eqref{tensor:rel}, \eqref{sym:dgamma_nu} we obtain \begin{align*} \left.\frac{d}{dt} \partial_{x_i} G(\boldsymbol{x} + t\delta \boldsymbol{x},r) \right|_{t= 0} & = - \int_{\mathcal{A}_i} \operatorname{div}_\Gamma(\nu_i\otimes V) - \nabla_\Gamma (V\cdot\nu_i)\, dz. \end{align*} Applying Theorem \ref{lem:tandivpol} to the integral of $\operatorname{div}_\Gamma(\nu_i\otimes V)$ on each arc in $\mathcal{A}_i$ we get \begin{align*} \left.\frac{d}{dt} \partial_{x_i} G(\boldsymbol{x} + t\delta \boldsymbol{x},r) \right|_{t= 0} & = - \int_{\mathcal{A}_i} \mathcal{H}(\nu_i\otimes V)\cdot \nu_i - \nabla_\Gamma (V\cdot\nu_i) \, dz - \sum_{(v,w) \in \mathbb{A}_i} \llbracket(\nu_i(z)\otimes V(z))\cdot \tau_i(z)\rrbracket_v^w, \end{align*} and then, using $\mathcal{H}=1/r$ on $\mathcal{A}_i$, \begin{align}\label{eq:577} \begin{split} \left.\frac{d}{dt} \partial_{x_i} G(\boldsymbol{x} + t\delta \boldsymbol{x},r) \right|_{t= 0} & = - \frac{1}{r} \int_{\mathcal{A}_i}(V\cdot \nu_i)\nu_i - r\nabla_\Gamma (V\cdot\nu_i) \, dz - \sum_{(v,w) \in \mathbb{A}_i} \llbracket (V(z)\cdot \tau_i(z))\nu_i(z)\rrbracket_v^w. \end{split} \end{align} Applying \eqref{eq:Vtheta_arc} yields $V\cdot \nu_i = \delta x_i\cdot \nu_i \text{ on } \mathcal{A}_i.$ Considering that $\delta x_\ell=0$ for $\ell\neq i$ since we use the specific perturbation $\delta \boldsymbol{x} = (0,\dots,0,\delta x_i, 0,\dots, 0)$, \eqref{eq:Vthetaa} and \eqref{eq:Vthetaa_2} actually provide the same formula in this particular case: \begin{align}\label{eq:vtau} \begin{split} V(z)\cdot \tau_i(z) &= \delta x_i\cdot \left(\tau_i - \frac{\nu_{-i}}{\tau_i \cdot \nu_{-i}}\right)(z)\\ & = - \left(\frac{\nu_{-i} \cdot \nu_i}{ \nu_{-i}\cdot \tau_i } \delta x_i\cdot \nu_i\right)(z) \text{ for }z\in \{v, w\} \text{ and } (v,w)\in\mathbb{A}_i. \end{split} \end{align} We also have $\nabla_\Gamma (V\cdot\nu_i) = \nabla_\Gamma (\delta x_i\cdot \nu) = (D_\Gamma\nu)^\top \delta x_i$ and $$ D_\Gamma\nu_i = D_\Gamma \begin{pmatrix} \cos\theta \\ \sin\theta\end{pmatrix} = \begin{pmatrix} \nabla_\Gamma (\cos\theta)^\top \\ \nabla_\Gamma (\sin\theta)^\top\end{pmatrix} = \frac{1}{r}\begin{pmatrix} \partial_\theta (\cos\theta)\tau_i^\top \\ \partial_\theta (\sin\theta)\tau_i^\top\end{pmatrix} = \frac{1}{r} \tau_i\otimes \tau_i \text{ on }\mathcal{A}_i.$$ Gathering these results and using $V\cdot \nu_i = \delta x_i\cdot \nu_i \text{ on } \mathcal{A}_i$ we get \begin{align*} \left.\frac{d}{dt} \partial_{x_i} G(\boldsymbol{x} + t\delta \boldsymbol{x},r) \right|_{t= 0} & = - \frac{1}{r} \int_{\mathcal{A}_i}(\delta x_i\cdot \nu_i)\nu_i - (\tau_i\otimes \tau_i)\delta x_i\, dz + \sum_{(v,w) \in \mathbb{A}_i} \left\llbracket\frac{\nu_{-i} \cdot \nu_i}{ \nu_{-i}\cdot \tau_i } \nu_i\otimes \nu_i\right\rrbracket_v^w \delta x_i, \end{align*} which yields \eqref{d2xixiG}. \subsection{Second order derivative with respect to \texorpdfstring{$x_i$}{xi} and \texorpdfstring{$x_\ell$}{xl} of \texorpdfstring{$G$}{G}}\label{sec:derxixl} As in Section \ref{sec:second_D_xiG} we use the mapping $T_t$ from Theorem \ref{thm:EB}, which requires that Assumptions~\ref{a1} and \ref{a3} hold, but now with the specific perturbation $\delta \boldsymbol{x} = (0,\dots,0,\delta x_\ell, 0,\dots, 0)$. This yields a transformation $T_t$ satisfying in particular $T_t(\mathcal{A}_i) = \partial B(x_i,r)\cap \partial\Omega(\boldsymbol{x}+ t\delta \boldsymbol{x},r)\cap A$. Then, using Lemma \ref{lem:intsub} we obtain \begin{align*} \partial_{x_i} G(\boldsymbol{x} + t\delta \boldsymbol{x},r) &= - \int_{\partial B(x_i,r)\cap \partial\Omega(\boldsymbol{x}+ t\delta \boldsymbol{x},r)\cap A} \nu_t(z) \, dz = - \int_{T_t(\mathcal{A}_i)} \nu_t(z) \, dz = - \int_{\mathcal{A}_i} \nu_t(T_t(z)) \omega_t(z) \, dz, \end{align*} where $\nu_t$ is the outward unit normal vector to $\partial B(x_i,r)\cap \partial\Omega(\boldsymbol{x}+ t\delta \boldsymbol{x},r)\cap A$ and $\omega_t$ is given by \eqref{density:changevar}. Applying \eqref{eq:Vtheta_arc} and considering that $\delta x_i=0$ since we are using the specific perturbation $\delta \boldsymbol{x} = (0,\dots,0,\delta x_\ell, 0,\dots, 0)$, we get \begin{equation}\label{eq:743} V\cdot \nu_i = 0 \text{ on } \mathcal{A}_i. \end{equation} Then, applying \eqref{eq:Vthetaa_2} with $\delta x_i = 0$ we get \begin{align}\label{eq:744} V(z)\cdot\tau_i(z) &= \frac{\delta x_\ell\cdot \nu_\ell(z)}{\tau_i(z)\cdot \nu_\ell(z)}\quad \text{ if } z\in \partial B(x_\ell,r)\cap\partial B(x_i,r), i\neq \ell. \end{align} Next, the derivative of $\partial_{x_i} G(\boldsymbol{x} + t\delta \boldsymbol{x},r)$ with respect to $t$ at $t=0$ is already calculated in \eqref{eq:577}, but the terms $(V\cdot \nu_i)\nu_i$ and $\nabla_\Gamma (V\cdot\nu_i)$ in \eqref{eq:577} vanish due to \eqref{eq:743}. We also observe that $V(z) = 0$ if $z\in\{v,w\}$ with $(v,w) \in \mathbb{A}_i$ and $z\notin \partial B(x_\ell,r)$. Finally, using ${\cal I}_{i\ell} = \{ v \in \partial B(x_\ell,r) \; | \; (v,\cdot) \in \mathbb{A}_i \}$, $\mathcal{O}_{i\ell} = \{ w \in \partial B(x_\ell,r) \; | \; (\cdot,w) \in \mathbb{A}_i \}$ and \eqref{eq:744} we get \begin{align*} \left.\frac{d}{dt} \partial_{x_i} G(\boldsymbol{x} + t\delta \boldsymbol{x},r) \right|_{t= 0} & = \sum_{v \in {\cal I}_{i\ell}} V(v)\cdot \tau_i(v)\nu_i(v) - \sum_{w \in {\cal O}_{i\ell}} V(w)\cdot \tau_i(w)\nu_i(w) \\ & = \left[ \sum_{v \in {\cal I}_{i\ell}} \frac{\nu_i(v)\otimes \nu_\ell(v)}{\nu_\ell(v)\cdot \tau_i(v)} -\sum_{w \in {\cal O}_{i\ell}} \frac{\nu_i(w)\otimes \nu_\ell(w)}{\nu_\ell(w)\cdot \tau_i(w)}\right] \delta x_\ell, \end{align*} which yields \eqref{d2Gxixj}. \subsection{Second order derivative with respect to \texorpdfstring{$x_i$}{xi} and \texorpdfstring{$r$}{r} of \texorpdfstring{$G$}{G}}\label{sec:derxir} In a similar way as in Sections \ref{sec:second_D_xiG} and \ref{sec:derxixl}, we use the mapping $T_t$ from Theorem \ref{thm:EB} with the specific perturbation $\delta \boldsymbol{x} = (0,\dots,0,\delta x_i, 0,\dots, 0)$. This yields, using Lemma \ref{lem:intsub}, \begin{align*} \partial_r G(\boldsymbol{x} + t\delta \boldsymbol{x},r) &= - \int_{\partial\Omega(\boldsymbol{x} + t\delta \boldsymbol{x},r)\cap A}\, dz = - \int_{T_t(\partial\Omega(\boldsymbol{x},r)\cap A)} \, dz = - \int_{\partial\Omega(\boldsymbol{x},r)\cap A} \omega_t(z) \, dz. \end{align*} Proceeding as in the calculation leading to \eqref{eq:333}, we get \begin{align}\label{eq:777} \begin{split} \left.\frac{d}{dt} \partial_r G(\boldsymbol{x} + t\delta \boldsymbol{x},r)\right|_{t= 0} & = - \sum_{\ell=1}^m \int_{\mathcal{A}_\ell} \mathcal{H} V\cdot \nu_\ell \, dz - \sum_{\ell=1}^m \sum_{(v,w) \in \mathbb{A}_\ell} \llbracket V(z)\cdot \tau_\ell(z)\rrbracket_v^w. \end{split} \end{align} Considering that $\delta x_\ell=0$ for $\ell\neq i$, since we use the specific perturbation $\delta \boldsymbol{x} = (0,\dots,0,\delta x_i, 0,\dots, 0)$, \eqref{eq:Vthetaa} and \eqref{eq:Vthetaa_2} actually provide the same formula in this particular case: \begin{align}\label{eq:772b} \begin{split} V(z)\cdot \tau_i(z) &= \delta x_i\cdot \left(\tau_i - \frac{\nu_{-i}}{\tau_i \cdot \nu_{-i}}\right)(z)\\ & = - \left(\frac{\nu_{-i} \cdot \nu_i}{ \nu_{-i}\cdot \tau_i } \delta x_i\cdot \nu_i\right)(z) \text{ for }z\in \{v, w\} \text{ and } (v,w)\in\mathbb{A}_i, \end{split} \end{align} and also \begin{align*} V(z)\cdot \tau_\ell(z) &= \delta x_i \cdot \left( \tau_\ell - \frac{\nu_\ell}{\tau_i \cdot \nu_\ell} (\tau_i\cdot\tau_\ell)\right)(z)\\ &= \delta x_i \cdot \left( \frac{\mu}{\tau_i \cdot \nu_\ell} \right)(z) \text{ if } z\in\partial B(x_i,r)\cap \partial B(x_\ell,r) \text{ and } \ell\neq i. \end{align*} with $\mu:= \tau_\ell (\tau_i\cdot \nu_\ell) - (\tau_i\cdot \tau_\ell)\nu_\ell$. This yields $\mu\cdot\tau_i = 0$ and $$ \mu\cdot \nu_i = (\tau_\ell\cdot \nu_i)(\tau_i\cdot \nu_\ell) - (\tau_i\cdot \tau_\ell)(\nu_\ell\cdot \nu_i) = - (\tau_i\cdot \nu_\ell)^2 - (\tau_i\cdot \tau_\ell)^2 = -1,$$ where we have used the geometric properties $\tau_\ell\cdot \nu_i = - \tau_i\cdot \nu_\ell$ and $\tau_i\cdot \tau_\ell = \nu_\ell\cdot \nu_i$. Thus $\mu = -\nu_i$ and we get \begin{align}\label{eq:774} V(z)\cdot \tau_\ell(z) = - \frac{\delta x_i \cdot \nu_i(z)}{\tau_i(z) \cdot \nu_\ell(z)} \text{ if } z\in\partial B(x_i,r)\cap \partial B(x_\ell,r) \text{ and } \ell\neq i. \end{align} In \eqref{eq:777}, we observe that $V(z)=0$ whenever $z\in\{v,w\}$ and $z\notin \partial B(x_i,r)$; this can be seen from \eqref{eq:Vthetaa}-\eqref{eq:Vthetaa_2} and the fact that we use the specific perturbation $\delta \boldsymbol{x} = (0,\dots,0,\delta x_i, 0,\dots, 0)$. Hence, recalling that $L(z)=\{ \ell \in \{1,\dots,m\} \setminus \{i\} \;|\; z \in \partial B(x_\ell,r) \}$, \begin{align*} \sum_{\ell=1}^m \sum_{(v,w) \in \mathbb{A}_\ell} \llbracket V(z)\cdot \tau_\ell(z)\rrbracket_v^w &= \sum_{(v,w) \in \mathbb{A}_i} \llbracket V(z)\cdot \tau_i(z)\rrbracket_v^w + \sum_{\substack{\ell=1 \\ \ell\neq i}}^m \sum_{(v,w) \in \mathbb{A}_\ell} \llbracket V(z)\cdot \tau_\ell(z)\rrbracket_v^w\\ &= \sum_{(v,w) \in \mathbb{A}_i} \llbracket V(z)\cdot \tau_i(z)\rrbracket_v^w - \sum_{(v,w) \in \mathbb{A}_i} \left\llbracket \sum_{\ell\in L(z)} V(z)\cdot \tau_\ell(z)\right\rrbracket_v^w. \end{align*} Note that the negative sign in front of the last sum is due to the fact that if an ending point of an arc in $\mathbb{A}_\ell$ belongs to some arc in $\mathbb{A}_i$, then it is a starting point for this arc in $\mathbb{A}_i$, and vice versa. Using \eqref{eq:Vtheta_arc} we have $V\cdot \nu_\ell \equiv 0$ on $\mathcal{A}_\ell$ for all $\ell\neq i$. Since $\mathcal{H} = 1/r$, we may write \eqref{eq:777} as \begin{align}\label{eq:778} \begin{split} \left.\frac{d}{dt} \partial_r G(\boldsymbol{x} + t\delta \boldsymbol{x},r)\right|_{t= 0} & = - \frac{1}{r}\int_{\mathcal{A}_i} V\cdot \nu_i \, dz - \sum_{(v,w) \in \mathbb{A}_i} \left\llbracket V(z)\cdot \tau_i(z) - \sum_{\ell\in L(z)} V(z)\cdot \tau_\ell(z)\right\rrbracket_v^w. \end{split} \end{align} Using \eqref{eq:Vtheta_arc} we get $V\cdot \nu_i = \delta x_i\cdot\nu_i$ on $\mathcal{A}_i$. Finally, using \eqref{eq:772b}-\eqref{eq:774} we get \begin{align*} \begin{split} \left.\frac{d}{dt} \partial_r G(\boldsymbol{x} + t\delta \boldsymbol{x},r)\right|_{t= 0} & = - \frac{1}{r} \int_{\mathcal{A}_i} \delta x_i\cdot\nu_i \, dz\\ &\quad + \sum_{(v,w) \in \mathbb{A}_i} \left\llbracket\frac{\nu_{-i}(z) \cdot \nu_i(z)}{ \nu_{-i}(z)\cdot \tau_i(z) } \delta x_i\cdot \nu_i(z) - \sum_{\ell\in L(z)} \frac{\delta x_i \cdot \nu_i(z)}{\tau_i(z) \cdot \nu_\ell(z)} \right\rrbracket_v^w , \end{split} \end{align*} which yields \eqref{eq:781}. \section{Analysis of singular cases}\label{sec:singularcases} The gradient $\nabla G$ and Hessian $\nabla^2 G$ were obtained under Assumptions~\ref{a1} and \ref{a3}, and in this section we investigate several singular cases where these assumptions are not satisfied. On the one hand, it is shown in \cite[\S~3.5]{coveringfirst} that $G$ is often differentiable even when Assumptions~\ref{a1} and \ref{a3} do not hold, and in the few cases where $G$ is not differentiable it is at least Gateaux semidifferentiable. On the other hand, $G$ is never twice differentiable in any of the singular geometric configurations studied in this section. Nevertheless, Gateaux semidifferentiability of the components of $\nabla G$ can often be proven. We recall here that $f:\mathds{R}^n\to\mathds{R}$ is Gateaux semidifferentiable at $x$ in the direction $v$ if $$ \lim_{t\searrow 0} \frac{f(x+tv) - f(x) }{t} \text{ exists in } \mathds{R}^n,$$ and that $f$ has a derivative in the direction $v$ at $x$ if $$ \lim_{t\to 0} \frac{f(x+tv) - f(x) }{t} \text{ exists in } \mathds{R}^n.$$ \begin{example} \label{examp1} Suppose $m=2$, $\Omega(\boldsymbol{x} +t\delta\boldsymbol{x},r)\subset A$ for all $t\in [0,t_0]$ and $t_0$ sufficiently small, and the two balls are tangent at $t=0$, i.e., $\| x_1 -x_2 \|= 2r$; thus Assumption \ref{a1} is not satisfied. Two cases need to be considered to compute the gradient of $G$. First, if $(x_1 -x_2)\cdot (\delta x_1 - \delta x_2) \geq 0$, then it is clear that $B(x_1 +t\delta x_1,r)\cap B(x_2+t\delta x_2,r) =\emptyset$ for all $t\in [0,t_0]$. Therefore $G(\boldsymbol{x}+t\delta\boldsymbol{x},r)=G(\boldsymbol{x},r)=\operatorname{Vol}(A) - 2\pi r^2$ for all $t\in [0,t_0]$, and $ \lim_{t\searrow 0} (G(\boldsymbol{x}+t\delta\boldsymbol{x},r) - G(\boldsymbol{x},r) )/t = 0$. Second, if $(x_1 -x_2)\cdot (\delta x_1 - \delta x_2) < 0$ then $B(x_1 +t\delta x_1,r)\cap B(x_2+t\delta x_2,r) \neq \emptyset$ for all $t \in (0,t_0]$. The intersection of $B(x_1 +t\delta x_1,r)$ and $B(x_2+t\delta x_2,r)$ form a symmetric lens whose area is given by $$a(t) = 2r^2 \arccos \left( d(t)/ 2r \right) -d(t) \left( r^2 - d(t)^2/4 \right)^{1/2},$$ where $d(t):= \|x_1 + t\delta x_1 - (x_2 +t\delta x_2)\|$. It is convenient to rewrite this expression as $$ a(t) = 2r^2 \arccos \left( (1-g(t))^{1/2} \right) - 2r^2\left( g(t) + g(t)^2 \right)^{1/2},$$ with $g(t):= -(2t (x_1 -x_2)\cdot (\delta x_1 - \delta x_2) + t^2 \| \delta x_1 - \delta x_2 \|^2)/(4r^2)$, $g(t)\geq 0$ for all $t\in [0,t_0]$ for $t_0$ small enough, $ d(t) =2r ( 1- g(t))^{1/2}$, and $g'(0) = - (x_1 -x_2)\cdot (\delta x_1 - \delta x_2)/(2r^2)$. After simplifications, we obtain $ a'(t) = 2r^2 \left( \frac{g(t)}{1 -g(t)} \right)^{1/2} g'(t),$ and in particular $a'(0) = 0$. This shows that $$ \lim_{t\searrow 0} \frac{G(\boldsymbol{x}+t\delta\boldsymbol{x},r) - G(\boldsymbol{x},r) }{t} = 0 \quad \text{ when } (x_1 -x_2)\cdot (\delta x_1 - \delta x_2) < 0.$$ Hence $\lim_{t\to 0} (G(\boldsymbol{x}+t\delta\boldsymbol{x},r) - G(\boldsymbol{x},r) )/t = 0$ for all $\delta\boldsymbol{x}\in\mathds{R}^4$. Proceeding in a similar way we can also show that $ \lim_{t\to 0} (G(\boldsymbol{x},r +t) - G(\boldsymbol{x},r) )/t = -4\pi r$. Thus $\nabla G(\boldsymbol{x},r) = (0,\dots,0,-4\pi r)^\top$ in the case $\| x_1 -x_2 \|= 2r$. (In \cite[Example~3.10]{coveringfirst} it is written $\nabla G(\boldsymbol{x},r) = (0,\dots,0,4\pi r)^\top$ where it should be written $\nabla G(\boldsymbol{x},r) = (0,\dots,0,-4\pi r)^\top$.) It is easy to check that formula \eqref{gradG} also gives $\nabla G(\boldsymbol{x},r) = (0,\dots,0,-4\pi r)^\top$ in this case. This indicates that, for the analyzed case, \eqref{gradG} is valid even without the satisfaction of Assumption~\ref{a1}. However, we had to use a different technique to prove that \eqref{gradG} holds, as $G(\boldsymbol{x}+t\delta\boldsymbol{x},r)$ takes different expressions depending on the sign of $(x_1 -x_2)\cdot (\delta x_1 - \delta x_2)$. We now study second-order differentiability of $G$. Let $f(t):=G(\boldsymbol{x}+t\delta\boldsymbol{x},r)$, then if $(x_1 -x_2)\cdot (\delta x_1 - \delta x_2) \geq 0$ we have $$f(t)=\operatorname{Vol}(A) - \operatorname{Vol}(B(x_1+t\delta x_1,r)) - \operatorname{Vol}(B(x_2+t\delta x_2,r)) = \operatorname{Vol}(A) -2\pi r^2 \text{ for all } t\in [0,t_0].$$ Thus in this case we get the right derivatives $f'(0)=\nabla_{\boldsymbol{x}} G(\boldsymbol{x},r)\cdot\delta\boldsymbol{x}=0$ and $f''(0)=\nabla^2_{\boldsymbol{x}} G(\boldsymbol{x},r)\delta\boldsymbol{x}\cdot\delta\boldsymbol{x} = 0$. In the case $(x_1 -x_2)\cdot (\delta x_1 - \delta x_2) \leq 0$ we get $$f(t)=\operatorname{Vol}(A) - \operatorname{Vol}(B(x_1+t\delta x_1,r)) - \operatorname{Vol}(B(x_2+t\delta x_2,r)) +a(t) \text{ for all } t\in [0,t_0],$$ and $f'(t)=\nabla_{\boldsymbol{x}} G(\boldsymbol{x}+t\delta\boldsymbol{x},r)\cdot\delta\boldsymbol{x}=a'(t)$, $f''(t)=\nabla^2_{\boldsymbol{x}} G(\boldsymbol{x}+t\delta\boldsymbol{x},r)\delta\boldsymbol{x}\cdot\delta\boldsymbol{x} = a''(t)$. The calculation yields $$ a''(t) = r^2 \left( \frac{g(t)}{1 -g(t)} \right)^{-1/2}\left( \frac{g'(t)}{1 -g(t)} + \frac{g(t)g'(t)}{(1 -g(t))^2}\right) g'(t) + 2r^2 \left(\frac{g(t)}{1 -g(t)} \right)^{1/2} g''(t),$$ and $g(0)=0$, $g'(0) = - (x_1 -x_2)\cdot (\delta x_1 - \delta x_2)/(2r^2)$, $g''(0)= - \| \delta x_1 - \delta x_2 \|^2/(2r^2)$. This shows that $|a''(t)|\to\infty$ as $t\to 0$ and consequently $|\nabla^2_{\boldsymbol{x}} G(\boldsymbol{x}+t\delta\boldsymbol{x},r)\delta\boldsymbol{x}\cdot\delta\boldsymbol{x}| \to\infty$ as $t\to 0$. This result is coherent with \eqref{d2xixiG} as we can show that $|\partial^2_{x_1x_1} G(\boldsymbol{x},r)u\cdot u| \to\infty$ as $\|x_1 - x_2\|\to 2r$ with $\|x_1 - x_2\|< 2r$ and $u=(1,0)^\top$. Thus $\nabla_{\boldsymbol{x}} G$ is not differentiable in this geometric configuration. Now we investigate the Gateaux semidifferentiability of $\partial_r G$ with respect to the radius. Let us introduce the notation $L(\rho) := \operatorname{Vol} \left(B(x_1,\rho)\cap B(x_2,\rho)\right)$ with $\rho>r>\|x_1-x_2\|/2.$ Then $L(\rho)$ is the area of a symmetric lens given by $$ L(\rho) = 2\rho^2 \arccos \left( \frac{r}{\rho} \right) - 2r\left( \rho^2 - r^2 \right)^{1/2},$$ and we have $G(\boldsymbol{x},\rho)=\operatorname{Vol}(A) - 2\pi \rho^2 +L(\rho) \text{ for } 2r>\rho\geq r$. Thus \begin{align*} \partial_\rho G(\boldsymbol{x},\rho) &= - 4\pi \rho +L'(\rho)\text{ and } \partial^2_{\rho\rho} G(\boldsymbol{x},\rho) = - 4\pi +L''(\rho) \text{ for } 2r>\rho> r, \end{align*} and we compute \begin{align*} L'(\rho) &= 4\rho \arccos \left( \frac{r}{\rho} \right),\qquad L''(\rho) = 4 \arccos \left( \frac{r}{\rho} \right) + \frac{4r}{(\rho^2 -r^2)^{1/2}} \text{ for } \rho>r. \end{align*} Then we observe that $L'(r) = 0$ and $\lim_{\rho\searrow r} L''(\rho) = +\infty$. Thus, if $\delta r>0$ then $$ \lim_{t\searrow 0} \frac{\partial_r G(\boldsymbol{x},r+t\delta r) - \partial_r G(\boldsymbol{x},r) }{t} = +\infty,$$ and $\partial_r G$ is not Gateaux semidifferentiable in direction $(0,0,\delta r)$ with $\delta r>0$. On the other hand, if $\delta r<0$ and $t>0$, then we have $G(\boldsymbol{x},r+t\delta r)=\operatorname{Vol}(A) - 2\pi (r+t\delta r)^2$, thus $\partial_r G$ is Gateaux semidifferentiable in direction $(0,0,\delta r)$ and $$ \lim_{t\searrow 0} \frac{\partial_r G(\boldsymbol{x},r+t\delta r) - \partial_r G(\boldsymbol{x},r) }{t} = -4\pi\delta r \text{ for }\delta r<0.$$ We conclude that $\partial_r G$ is not differentiable in this geometric configuration. \end{example} \begin{example} \label{examp2} Suppose $A$ is a square, $m=1$, $\Omega(\boldsymbol{x},r)\subset A$ and $\Omega(\boldsymbol{x},r)$ is tangent to $\partial A$ on the right side of the square but is not tangent to the other sides. Note that Assumption \ref{a3} is not satisfied as $\Omega(\boldsymbol{x},r)$ and $A$ are not compatible. Then $\delta\boldsymbol{x} = \delta x_1$ and for sufficiently small $t>0$ we have \begin{align*} G(\boldsymbol{x}+t\delta\boldsymbol{x},r) &=\operatorname{Vol}(A) - \pi r^2 \text{ if } \delta x_1 = (-1,0)^\top\\ G(\boldsymbol{x}+t\delta\boldsymbol{x},r) &=\operatorname{Vol}(A) - \pi r^2 + a(t) \text{ if } \delta x_1 = (1,0)^\top. \end{align*} with $$a(t) = r^2\arccos\left(\frac{r-t\delta x_1}{r}\right) -(r-t\delta x_1)g(t)^{1/2} $$ and $g(t)=(r^2 -(r-t\delta x_1)^2)$. We compute $a'(t) = 2\delta x_1 g(t)^{1/2}$, $a''(t) = \delta x_1 g(t)^{-1/2}g'(t)$, $g(0)=0$ and $g'(0)=2\delta x_1 r$. Thus $a'(0)=0$ and $a''(t)\to +\infty$ as $t\to 0$. It is clear that $G(\boldsymbol{x}+t\delta\boldsymbol{x},r)$ does not depend on the second component of $\delta x_1$ for sufficiently small $t>0$, thus we have shown that $G$ has a derivative in direction $(\delta\boldsymbol{x},0)$ for all $\delta\boldsymbol{x}\in\mathds{R}^2$ and that $\partial_{x_1} G(\boldsymbol{x},r)=0$, which gives the same value as \eqref{gradG} even though $\Omega(\boldsymbol{x},r)$ and $A$ are not compatible in this example. Since $G(\boldsymbol{x}+t\delta\boldsymbol{x},r)$ is constant for $\delta\boldsymbol{x} = (-1,0)^\top$, we have $$ \lim_{t\searrow 0} \frac{\partial_{x_1} G(\boldsymbol{x}+t\delta\boldsymbol{x},r) - \partial_{x_1} G(\boldsymbol{x},r) }{t} = (0,0)^\top \quad \text{ for } \delta\boldsymbol{x} = (-1,0)^\top,$$ thus $\partial_{x_1} G$ is Gateaux semidifferentiable in direction $(\delta\boldsymbol{x},0)$ with $\delta\boldsymbol{x}=\delta x_1 = (-1,0)^\top$. On the other hand for $\delta\boldsymbol{x} = (1,0)^\top$ we have $$ \lim_{t\searrow 0} \frac{( \partial_{x_1} G(\boldsymbol{x}+t\delta\boldsymbol{x},r) - \partial_{x_1} G(\boldsymbol{x},r) ) \cdot\delta\boldsymbol{x}}{t} = \lim_{t\searrow 0} a''(t) = +\infty \quad \text{ for } \delta\boldsymbol{x} = (1,0)^\top,$$ thus $\partial_{x_1} G$ is not Gateaux semidifferentiable in direction $(\delta\boldsymbol{x},0)$ with $\delta\boldsymbol{x}=\delta x_1 = (1,0)^\top$. This shows that $G$ is not twice differentiable in this geometric configuration. \end{example} \begin{example} \label{examp2b} Let $A = [0,2]^2$, $m=1$, $\Omega(\boldsymbol{x},r)=B(x_1,r)$ with $x_1 = (0,1/2)$ and $r=1/2$, then $\partial B(x_1,r)$ intersects $\partial A$ at a vertex, thus $A$ and $\Omega(\boldsymbol{x},r)$ are not compatible and Assumption~\ref{a3} is not satisfied. In the case of a horizontal translation $\delta x_1 = (1,0)^\top$ we symmetrize the square $A$ vertically by defining $A_s = A\cup [0,2]\times [-2,0]$. Defining $G_s(\boldsymbol{x},r) = \operatorname{Vol}(A_s) - \operatorname{Vol}(A_s \cap \Omega(\boldsymbol{x},r))$ we observe that $G_s(\boldsymbol{x}+t\delta\boldsymbol{x},r) = G(\boldsymbol{x}+t\delta\boldsymbol{x},r) + \operatorname{Vol}([0,2]\times [-2,0])$ for $\delta\boldsymbol{x} = \delta x_1 = (1,0)^\top$ and sufficiently small $t$, so that $G_s(\boldsymbol{x},r)$ and $G(\boldsymbol{x},r)$ have the same partial derivatives in direction $\delta x_1 = (1,0)^\top$ since $\operatorname{Vol}([0,2]\times [-2,0])$ is constant. Since $A_s$ and $\Omega(\boldsymbol{x},r)$ are compatible, this shows that $\partial_{x_1} G(\boldsymbol{x},r)$ has a derivative at $(\boldsymbol{x},r)$ in direction $(\delta\boldsymbol{x},0)$ with $\delta\boldsymbol{x}=\delta x_1 = (1,0)^\top$. In the case of a vertical translation $\delta x_1 = (0,\pm 1)^\top$ we symmetrize the square $A$ horizontally by defining $A_s = A\cup [-2,0]\times [0,2]$. Then $\Omega(\boldsymbol{x},r)$ is tangent to one side of $A_s$ and we can use the results of Example~\ref{examp2}. We conclude that $\partial_{x_1} G$ is Gateaux semidifferentiable in direction $(\delta\boldsymbol{x},0)$ with $\delta\boldsymbol{x}=\delta x_1 = (0,1)^\top$ but is not Gateaux semidifferentiable in direction $(\delta\boldsymbol{x},0)$ with $\delta\boldsymbol{x}=\delta x_1 = (0,-1)^\top$. This shows that $G$ is not twice differentiable in this geometric configuration. \end{example} \begin{example} \label{examp3} Let $m=3$ and $x_1,x_2,x_3$ be the vertices of an equilateral triangle. The circles $\partial B(x_1,r)$, $\partial B(x_2,r)$ and $\partial B(x_3,r)$ intersect at a single point exactly when $\|x_1 - x_2\| = \|x_1 - x_3\| = \|x_2 - x_3\| = \sqrt{3} r$ and Assumption \ref{a1} is not satisfied in this geometric configuration. For $\|x_1 - x_2\|> r> r_0$ with $r_0 := \|x_1 - x_2\| /\sqrt{3}$, the intersection $B(x_1,r)\cap B(x_2,r)\cap B(x_3,r)$ forms a shape called Reuleaux triangle, whose area is denoted by $R(r)$. Also, the intersection of two disks of identical radius creates a geometric figure called symmetric lens whose area is denoted by $L(r)$. Then it is easy to see that \begin{align} \label{reu1} G(\boldsymbol{x},r) &= \operatorname{Vol}(A) - \sum_{i=1}^3\operatorname{Vol}(B(x_i,r+t\delta r)) + 3L(r) \quad\text{ if } r< r_0,\\ \label{reu2} G(\boldsymbol{x},r) &= \operatorname{Vol}(A) - \sum_{i=1}^3\operatorname{Vol}(B(x_i,r+t\delta r)) + 3L(r) - 3R(r) \quad\text{ if } r\geq r_0. \end{align} An explicit calculation using trigonometry yields $$R(r) = \frac{\pi-\sqrt{3}}{2} s(r)^2 \text{ with } s(r) = \left( r^2 - \frac{3 r_0^2}{4}\right)^{1/2} -\frac{3 r_0}{2} +r, $$ and $ R'(r) = (\pi-\sqrt{3}) s(r)s'(r)$, $R''(r) = (\pi-\sqrt{3}) (s'(r)^2 + s(r)s''(r)) $ with \begin{align*} s'(r) &= r\left( r^2 - \frac{3 r_0^2}{4}\right)^{-1/2} +1,\quad s''(r) = \left( r^2 - \frac{3 r_0^2}{4}\right)^{-1/2} - r^2\left( r^2 - \frac{3 r_0^2}{4}\right)^{-3/2}. \end{align*} We also compute $s(r_0)=0$, $s'(r_0)=3$ and $s''(r_0)=-6/r_0$. Thus $R'(r_0)=0$ and $R''(r_0)=9(\pi-\sqrt{3})\neq 0$. Since $R'(r_0)=0$, (\ref{reu1},\ref{reu2}) shows that $G$ has a derivative in direction $(0,0,0,\delta r)$ at $r=r_0$ for any $\delta r\in\mathds{R}$, even though Assumption~\ref{a1} is not satisfied and we have $\partial_r G(\boldsymbol{x},r_0)=0$. On the other hand, since $R''(r_0)\neq 0$, $G$ is not twice differentiable at $r=r_0$ in view of (\ref{reu1},\ref{reu2}). However, (\ref{reu1},\ref{reu2}) shows that $\partial_r G$ is Gateaux semidifferentiable in both directions $(0,0,0,1)$ and $(0,0,0,-1)$ at $r_0$ with \begin{align*} \lim_{\delta r\to 0^-} (\partial_r G(\boldsymbol{x},r_0+\delta r) - \partial_r G(\boldsymbol{x},r_0))/\delta r &= -6\pi + 3L''(r_0),\\ \lim_{\delta r\to 0^+} (\partial_r G(\boldsymbol{x},r_0+\delta r) - \partial_r G(\boldsymbol{x},r_0))/\delta r &= -6\pi + 3L''(r_0) - 3R''(r_0).\\ \end{align*} \end{example} \section{Exact calculation of \texorpdfstring{$G$}{G} and its derivatives}\label{sec:calc_algorithms} In this section, we consider that $A = \cup_{j=1}^{p} A_j$ and $\{A_j\}_{j=1}^{p}$ are non-overlapping convex polygons. (If not available, such decomposition can be computed in $\mathcal{O}(e_A + \bar e_A)$, where $e_A$ is the number of vertices of $A$ and $\bar e_A$ is its number of notches; see, for example, \cite{keil} and the references therein.) The key ingredient for the exact computation of $G$, $\nabla G$, and $\nabla^2 G$ as stated in Section~\ref{sec:shape_opt_covering} is to consider partitions \begin{equation} \label{partitiondeverdade} A_j \cap \Omega(\boldsymbol{x},r) = \bigcup_{i \in \mathcal{K}_{A_j}} S_{ij}, \quad j=1,\dots,p, \end{equation} where $\mathcal{K}_{A_j} \subseteq \{ i \in \{1,\dots,m\} \;|\; B(x_i,r) \cap A_j \neq \emptyset\}$ for $j=1,\dots,p$ and each $S_{ij}$ is such that $\partial S_{ij}$ is a simple and convex curve given by the union of segments and arcs of the circle $\partial B(x_i,r)$. It is worth noticing that, since $A_1, A_2,\dots,A_p$ are disjoint, then $\{S_{ij}\}_{(i,j) \in \mathcal{K}}$ with $\mathcal{K} = \{ (i,j) \;|\; j \in \{1,\dots,p\} \mbox{ and } i \in \mathcal{K}_{A_j}\}$ is a partition of $A \cap \Omega(\boldsymbol{x},r)$, i.e., \begin{equation} \label{xicara} A \cap \Omega(\boldsymbol{x},r) = \bigcup_{(i,j) \in \mathcal{K}} S_{ij}, \end{equation} see Figure~\ref{fig:Sij}. Note that~8 out of the~10 balls intersect either $A_1$ or $A_2$ in Figure~\ref{fig:Sij}. Let us number the balls intersecting only $A_1$ from~1 to~5 and the balls intersecting only~$A_2$ from 8 to 10. Thus, balls 1 to~5 contribute to~\eqref{partitiondeverdade} with $S_{11}, S_{21}, \dots, S_{51}$, i.e., they contribute to the partition of $A_1$ only; while balls 8, 9, and~10 contribute with $S_{82}$, $S_{92}$, and $S_{10,2}$, i.e., they contribute to the partition of $A_2$ only. Balls 6~and~7 intersect both~$A_1$ and $A_2$ and contribute with $S_{61}$ and $S_{71}$ to the partition of $A_1$ and with $S_{62}$ and $S_{72}$ to the partition of $A_2$. Therefore we have $\mathcal{K}_{A_1}=\{1,2,3,4,5,6,7\}$, $\mathcal{K}_{A_2}=\{6,7,8,9,10\}$, and $\mathcal{K}=\{ (1,1), (2,1), (3,1), (4,1), (5,1), (6,1), (6,2), (7,1), (7,2), (8,2), (9,2), (10,2)\}$. In addition, for further use, we define $\mathcal{K}_{B_1}=\mathcal{K}_{B_2}=\mathcal{K}_{B_3}=\mathcal{K}_{B_4}=\mathcal{K}_{B_5}=\{1\}$, $\mathcal{K}_{B_6}=\mathcal{K}_{B_7}=\{1,2\}$, $\mathcal{K}_{B_8}=\mathcal{K}_{B_9}=\mathcal{K}_{B_{10}}=\{2\}$. \begin{figure}[ht!] \begin{center} \begin{tabular}{cc} \includegraphics{convpols_balls.mps} & \includegraphics{curv_pols.mps} \\ (a) & (b) \end{tabular} \end{center} \caption{(a) represents a region $A$ to be covered given by $A=\cup_{j=1}^p A_j$ with $p=2$ and an arbitrary configuration of balls $\Omega(\boldsymbol{x},r) = \cup_{i=1}^m B(x_i,r)$ with $m=10$. (b) represents the partitions of $A_1 \cap \Omega(\boldsymbol{x},r)$ and $A_2 \cap \Omega(\boldsymbol{x},r)$ defined in~\eqref{partitiondeverdade} that, together, as expressed in~\eqref{xicara}, represent a partition of $A \cap \Omega(\boldsymbol{x},r)$. In (b), the Voronoi diagram that allows the partitions to be computed is depicted.} \label{fig:Sij} \end{figure} The computation of the partitions in~\eqref{partitiondeverdade} is based on Voronoi diagrams. For a given $(\boldsymbol{x},r)$, we first compute the Voronoi diagram with cells $\{V_i\}_{i=1}^m$ associated with the balls centers $x_1,\dots,x_m$. Each cell $V_i$ is a (bounded or unbounded) polyhedron given by the points $y \in \mathds{R}^2$ such that $\|y-x_i\| = \min_{\{\ell=1,\dots,m\}} \{ \| y - x_\ell\| \}$. Then, for each $j=1,\dots,p$ and each $i=1,\dots,m$, we compute the convex polygons $W_{ij}=A_j \cap V_i$ and, in the sequence, $S_{ij} = W_{ij} \cap B(x_i,r)$. (Note that, by construction, $W_{ij} \cap B(x_i,r) = W_{ij} \cap \Omega(\boldsymbol{x},r)$; and so~\eqref{partitiondeverdade} and, in consequence, \eqref{xicara} hold.) In the construction process, we obtain the sets $\mathcal{K}_{A_j} = \{ i \in \{1,\dots,m\} \;|\; S_{ij} \neq \emptyset \}$, $\mathcal{K}_{B_i} = \{ j \in \{1,\dots,p\} \;|\; S_{ij} \neq \emptyset \}$, and $\mathcal{K}$ such that $(i,j) \in \mathcal{K}$ if and only if $S_{ij} \neq \emptyset$. Let $\mathcal{V}(S_{ij})$ be the set of vertices of $S_{ij}$, $\mathcal{A}(S_{ij}) = \partial S_{ij} \cap \partial B(x_i,r)$ the union of the arcs in $\partial S_{ij}$, and $\mathcal{E}(S_{ij}) = \partial S_{ij} \setminus \mathcal{A}(S_{ij})$ the union of the edges in $\partial S_{ij}$. Moreover, we associate with $\mathcal{E}(S_{ij})$ and $\mathcal{A}(S_{ij})$ the corresponding sets of maximal arcs $\mathbb{A}(S_{ij})$ and edges $\mathbb{E}(S_{ij})$. Strictly speaking, these are sets of pairs of points representing arcs and edges, respectively. Each edge is represented by a pair $[v,w]$ of vertices in counter-clockwise order and each arc is represented by a pair $(v,w)$ of vertices, in counter-clockwise order, that unequivocally determines two angles. For each vertex $z \in \mathcal{V}(S_{ij})$, we save whether $z \in \partial A$ or not. If $z \in \partial A$, then we save, whenever it exists, the unitary (Euclidean) norm outward normal vector to $\partial A$ at $z$, named $\nu_A(z)$. Additionally, for each vertex~$z \in {\cal V}(S_{ij})$, we save the set of indices~$L(z) \subseteq \{1,\dots, m\} \setminus \{i\}$ such that $z \in \partial B(x_\ell,r)$ for all $\ell \in L(z)$. Each set $\mathcal{A}_i = \partial B(x_i,r) \cap \partial \Omega(\boldsymbol{x},r) \cap A$ for $i=1,\dots,m$, defined in~\eqref{partition}, corresponds to the union of the arcs in $\partial S_{ij}$ for all $j \in \mathcal{K}_{B_i}$, i.e., it holds \begin{equation} \label{labendita} \mathcal{A}_i = \bigcup_{j \in \mathcal{K}_{B_i}} \mathcal{A}(S_{ij}) \end{equation} for $i=1,\dots,m$. It is worth noticing~\eqref{labendita} does not mean that every arc in $\mathbb{A}_i$ belongs to $\mathbb{A}(S_{ij})$ for some $j \in \mathcal{K}_{B_i}$ nor that $|\mathbb{A}_i|=\sum_{j \in \mathcal{K}_{B_i}} |\mathbb{A}(S_{ij})|$. Indeed, if $z$ is an extremity of an arc in $\mathbb{A}(S_{ij})$ then either $z \in \partial B(x_\ell,r)$ for some $\ell\neq i$ or $z \in \partial A_j$. In the case $z \in \partial A_j$, it may happen that $z \notin \partial A$, and consequently $z$ is not an extremity of an arc in $\mathbb{A}_i$. To construct $\mathbb{A}_i$, consecutive arcs (arcs with a extreme in common) in $\cup_{j \in \mathcal{K}_{B_i}} \mathbb{A}(S_{ij})$ must be merged into a single arc. So, what holds is that each arc in $\mathbb{A}_i$ belongs to $\mathbb{A}(S_{ij})$ for some $j \in \mathcal{K}_{B_i}$ or is the union of two or more consecutive arcs in $\cup_{j \in \mathcal{K}_{B_i}} \mathbb{A}(S_{ij})$. Thus $|\mathbb{A}_i| \leq \sum_{j \in \mathcal{K}_{B_i}} |\mathbb{A}(S_{ij})|$. The particular case $\mathcal{A}_i = \partial B(x_i,r)$ is considered separately; in this case, we set $\mathbb{A}_i=\emptyset$ and $\mathrm{Circle}(\mathbb{A}_i)$ equal to true. In a similar way, we also define \begin{equation} \label{labendita2} \mathcal{E}_i = \bigcup_{j \in \mathcal{K}_{B_i}} \mathcal{E}(S_{ij}), \end{equation} and the associate set $\mathbb{E}_i$ of pairs $[v,w]$ representing edges, for $i=1,\dots,m$. These sets of edges play a role in the computation of $G$ only. Thus, while the same principle of merging consecutive edges could be applied, it has no practical relevance because one way or the other, the same result is obtained. A second ingredient for the exact computation of $G$ and its derivatives are the parameterizations \begin{equation} \label{edgespar} t \mapsto \left( \begin{array}{c} x_{\cal E}(t) \\ y_{\cal E}(t) \end{array} \right) = v + t (w-v), \quad t \in [0,1], \end{equation} of each edge represented by $[v,w] \in \cup_{i=1}^m \mathbb{E}_i$; and the parameterizations \begin{equation} \label{arcspar} \theta \mapsto \left( \begin{array}{c} x_{\cal A}(\theta) \\ y_{\cal A}(\theta) \end{array} \right) = x_i + r \left( \begin{array}{c} \cos \theta \\ \sin \theta \end{array} \right), \quad \theta \in [\theta_v,\theta_w], \end{equation} of each arc represented by $(v,w) \in \mathbb{A}_i$ for $i=1,\dots,m$, where $\theta_v$ and $\theta_w$ are the angular coordinates of $v-x_i$ and $w-x_i$, respectively. We are now ready to compute $G$ and its derivatives. By~\eqref{aquiG}, \begin{equation} \label{alg1a} G(\boldsymbol{x},r) = \operatorname{Vol}(A) - \operatorname{Vol}(A \cap \Omega(\boldsymbol{x},r)) = \operatorname{Vol}(A) - \sum_{(i,j) \in \mathcal{K}} \operatorname{Vol}(S_{ij}). \end{equation} By Green's Theorem, \begin{equation} \label{alg1b} \operatorname{Vol}(S_{ij}) = \int_{S_{ij}} dxdy = \int_{\partial S_{ij}} x \, dy = \sum_{[v,w] \in \mathbb{E}(S_{ij})} \int_0^1 x_{\mathcal{E}}(t) \, dy_{\mathcal{E}}(t) + \sum_{(v,w) \in \mathbb{A}(S_{ij})} \int_{\theta_v}^{\theta_w} x_{\mathcal{A}}(\theta) \, dy_{\mathcal{A}}(\theta) \end{equation} for all $(i,j) \in \mathcal{K}$; while, by~\eqref{edgespar}, \begin{equation} \label{alg1c} \int_0^1 x_{\mathcal{E}}(t) \, dy_{\mathcal{E}}(t) = \frac{(v_1 + w_1)(w_2 - v_2)}{2} \end{equation} for all $[v,w] \in \mathbb{E}(S_{ij})$ and all $(i,j) \in \mathcal{K}$, and, by~\eqref{arcspar}, \begin{equation} \label{alg1d} \begin{array}{rcl} \displaystyle \int_{\theta_v}^{\theta_w} x_{\mathcal{A}}(\theta) \, dy_{\mathcal{A}}(\theta) &=& \displaystyle (x_{i})_1 \, r \, (\sin{\theta_w} - \sin{\theta_v}) \\[4mm] &+& \displaystyle \frac{r^2}{2} \left( \theta_w - \theta_v +\sin{\theta_w} \cos{\theta_w} - \sin{\theta_v} \cos{\theta_v} \right) \end{array} \end{equation} for all $(v,w) \in \mathbb{A}(S_{ij})$ and all $(i,j) \in \mathcal{K}$. The computation of~$G$ as defined in~\eqref{aquiG} using~(\ref{alg1a}--\ref{alg1d}) is summarized in Algorithm~\ref{G}. \begin{algorithm}[ht!] \caption[G]{\textsc{Computes $G(\boldsymbol{x},r)$.}} \label{G} \KwInput{$\operatorname{Vol}(A)$, $(\boldsymbol{x},r)$, and sets $\{\mathbb{E}_i,\mathbb{A}_i\}_{i=1}^m$.} \KwOutput{$G(\boldsymbol{x},r)$.} $\gamma \gets 0$ \\ \For{$i=1,\dots,m$}{ \If{$\mathrm{Circle}(\mathbb{A}_i)$}{ $\gamma \gets \gamma + \pi r^2$ } \Else{ \ForEach{$[v,w] \in \mathbb{E}_i$}{ $\gamma \gets \gamma+\frac{1}{2}(v_1 + w_1)(w_2 - v_2)$ } \ForEach{$(v,w) \in \mathbb{A}_i$}{ $\gamma \gets \gamma + {(x_i)}_1 r (\sin \theta_w - \sin \theta_v ) + \frac{r^2}{2} (\theta_w - \theta_v + \sin \theta_w \cos \theta_w - \sin \theta_v \cos \theta_v )$ } } } \Return $G = \operatorname{Vol}(A) - \gamma$ \end{algorithm} For computing~$\nabla G = (\partial_{x_1} G(\boldsymbol{x},r),\dots,\partial_{x_m} G(\boldsymbol{x},r),\partial_{r} G(\boldsymbol{x},r))^\top$, by~\eqref{gradG} and \eqref{partition}, we have that \begin{equation} \label{alg2a} \begin{array}{rcl} \partial_{x_i} G(\boldsymbol{x},r) &=& \displaystyle - \int_{\mathcal{A}_i} \nu_i(z) \, dz = \displaystyle - \sum_{(v,w) \in \mathbb{A}_i} \int_{\theta_v}^{\theta_w} r \, (\cos \theta, \sin \theta)^\top \, d \theta \\[4mm] &=& \displaystyle \sum_{(v,w) \in \mathbb{A}_i} r \, (\sin \theta_v - \sin \theta_w, \cos \theta_w - \cos \theta_v)^\top, \end{array} \end{equation} for $i=1,\dots,m$, and, by~\eqref{gradG} and \eqref{unionAi}, we have that \begin{equation} \label{alg2b} \partial_{r} G(\boldsymbol{x},r) = \displaystyle - \int_{\cup_{i=1}^m \mathcal{A}_i} dz = \displaystyle - \sum_{(v,w) \in \cup_{i=1}^m \mathbb{A}_i} \int_{\theta_v}^{\theta_w} r \, d \theta = \displaystyle - \sum_{(v,w) \in \cup_{i=1}^m \mathbb{A}_i} r \, (\theta_w - \theta_v). \end{equation} The computation of~$\nabla G$ as defined in~\eqref{gradG} using~(\ref{alg2a},\ref{alg2b}) is summarized in Algorithm~\ref{grad_G}. \begin{algorithm}[ht!] \caption[$\nabla G$]{\textsc{Computes $\nabla G(\boldsymbol{x},r)$.}} \label{grad_G} \KwInput{$(\boldsymbol{x},r)$, and sets $\{\mathbb{A}_i\}_{i=1}^m$.} \KwOutput{$\nabla G(\boldsymbol{x},r)$.} $g_r\gets0$ and $g_{x_i}\gets(0, 0)^\top$ for $i=1,\dots,m$.\\ \For{$i=1,\dots,m$}{ \If{$\mathrm{Circle}(\mathbb{A}_i)$}{ $g_r \gets g_r - 2 \pi r$ } \Else{ \ForEach{$(v,w) \in \mathbb{A}_i$}{ $g_r \gets g_r - r(\theta_w - \theta_v)$ \\ $g_{x_i} \gets g_{x_i} + r (\sin{\theta_v} - \sin{\theta_w}, \cos{\theta_w} - \cos{\theta_v})^\top$ } } } \textbf{return} $\nabla G(\boldsymbol{x},r) = (g_r, g_{x_1}^\top, \dots, g_{x_m}^\top)^\top$ \end{algorithm} For computing~$\nabla^2 G$, we use that, in~\eqref{d2Grr}, \begin{equation} \label{alg3a} - \frac{\mathrm{Per}(\partial\Omega(\boldsymbol{x},r)\cap A)}{r} = \sum_{(v,w) \in \cup_{i=1}^m \mathbb{A}_i} (\theta_v - \theta_w), \end{equation} in~\eqref{d2xixiG}, \begin{align} \label{alg3b} & \frac{1}{r} \displaystyle \int_{\mathcal{A}_i} - \nu_i(z) \otimes \nu_i(z) + \tau_i(z) \otimes \tau_i(z) \, dz \nonumber \\ &= \frac{1}{r} \sum_{(v,w) \in \mathbb{A}_i} r \int_{\theta_v}^{\theta_w} - \left( \begin{array}{cc} (\cos \theta)^2 & \sin \theta \cos \theta \\ \sin \theta \cos \theta & (\sin \theta)^2 \end{array} \right) + \left( \begin{array}{cc} (\sin \theta)^2 & -\sin \theta \cos \theta \\ -\sin \theta \cos \theta & (\cos \theta)^2 \end{array} \right) \, d \theta \\ &= \sum_{(v,w) \in \mathbb{A}_i} \left( \begin{array}{cc} \sin(\theta_v - \theta_w)\cos(\theta_v + \theta_w) & (\cos\theta_w)^2 - (\cos\theta_v)^2 \\ (\cos\theta_w)^2 - (\cos\theta_v)^2 & \sin(\theta_w - \theta_v)\cos(\theta_v + \theta_w) \end{array} \right), \nonumber \end{align} and, in~\eqref{eq:781}, \begin{equation} \label{alg3c} - \frac{1}{r} \int_{\mathcal{A}_i} \nu_i(z) \, dz\\[4mm] = \sum_{(v,w) \in \mathbb{A}_i} (\sin \theta_v - \sin \theta_w, \cos \theta_w - \cos \theta_v)^\top. \end{equation} Recall that in~(\ref{d2Grr},\ref{d2xixiG},\ref{eq:781}), for $z \in \partial B(x_i,r)$, $\nu_{-i}(z)$ represents the unitary-norm outwards normal vector to the set intersecting $\partial B(x_i,r)$ at $z$. If this set is $\partial A$, then $\nu_{-i}(z)=\nu_A(z)$. If this set is $\partial B(x_\ell,r)$ for some $\ell \in L(z)$, then $\nu_{-i}(z)=\nu_\ell(z)=(\cos \vartheta_z, \sin \vartheta_z)^\top$, where $\vartheta_z$ is the angular coordinate of $z-x_\ell$. With these definitions and substituting~(\ref{alg3a},\ref{alg3b},\ref{alg3c}) in~(\ref{d2Grr},\ref{d2xixiG},\ref{d2Gxixj},\ref{eq:781}), we arrive at Algorithm~\ref{hessian_G}. \begin{algorithm}[ht!] \caption[hessian]{\textsc{Computes $\nabla^2 G(\boldsymbol{x},r)$.}} \label{hessian_G} \KwInput{$(\boldsymbol{x},r)$ and sets $\{\mathbb{A}_i\}_{i=1}^m$.} \KwOutput{The lower triangle of $H = \nabla^2 G(\boldsymbol{x},r) \in \mathbb{R}^{2m+1,2m+1}$.} $H \gets 0$.\\ \For{$i=1,\dots,m$}{ \If{$\mathrm{Circle}(\mathbb{A}_i)$}{ $h_{2m+1, 2m+1} \gets h_{2m+1, 2m+1} - 2\pi$ } \Else{ \ForEach{$(v,w) \in \mathbb{A}_i$}{ \vspace{1mm} let $a \odot b$ mean $a \gets a + b$\\ $\scalebox{0.90}{$ \begin{array}{rcl} \left( \begin{array}{cc} h_{2i-1, 2i-1} & \\ h_{2i, 2i-1} & h_{2i, 2i} \end{array} \right) &\odot& \left( \begin{array}{cc} \sin(\theta_v - \theta_w)\cos(\theta_v + \theta_w) & \\ (\cos\theta_w)^2 - (\cos\theta_v)^2 & \sin(\theta_w - \theta_v)\cos(\theta_v + \theta_w) \end{array} \right) \\[4mm] (h_{2m+1, 2i-1}, h_{2m+1, 2i}) &\odot& (\sin \theta_v - \sin \theta_w, \cos \theta_w - \cos \theta_v) \\[2mm] h_{2m+1, 2m+1} &\odot& \theta_v - \theta_w \end{array} $}$\\ \For{$z \in \{v,w\}$}{ \vspace{1mm} \textbf{if} $z=v$ \textbf{then} let $a \odot b$ mean $a \gets a - b$ \textbf{else} let $a \odot b$ mean $a \gets a + b$\\[1mm] \If{$z \in \partial A$}{ \vspace{1mm} $\scalebox{0.90}{$ \begin{array}{l} \alpha \gets \left(-\nu_A(z) \cdot (\cos \theta_z, \sin \theta_z)^\top \right) / \left( \nu_A(z) \cdot (-\sin \theta_z, \cos \theta_z)^\top \right) \\[2mm] \begin{array}{rcl} \left( \begin{array}{cc} h_{2i-1, 2i-1} & \\ h_{2i, 2i-1} & h_{2i, 2i} \end{array} \right) &\odot& \alpha \left( \begin{array}{cc} (\cos\theta_z)^2 & \\ \sin \theta_z \cos \theta_z & (\sin\theta_z)^2 \end{array} \right) \\[4mm] (h_{2m+1, 2i-1}, h_{2m+1, 2i}) &\odot& \alpha(\cos \theta_z, \sin \theta_z) \\[2mm] h_{2m+1, 2m+1} &\odot& \alpha \end{array} \end{array} $}$ } \ForEach{$\ell(z) \in L(z)$}{ \vspace{1mm} $\scalebox{0.90}{$ \begin{array}{rcl} \left( \begin{array}{cc} h_{2i-1, 2i-1} & \\ h_{2i, 2i-1} & h_{2i, 2i} \end{array} \right) &\odot& \left( \begin{array}{cc} \cotan(\vartheta_z - \theta_z) (\cos\theta_z)^2 & \\ \cotan(\vartheta_z - \theta_z) \sin \theta_z \cos \theta_z & \cotan(\vartheta_z - \theta_z) (\sin\theta_z)^2 \end{array} \right) \\[4mm] h_{2m+1, 2i-1} &\odot& \cotan(\vartheta_z - \theta_z) \cos \theta_z - \cos \theta_z / \sin(\vartheta_z - \theta_z) \\[2mm] h_{2m+1, 2i} &\odot& \cotan(\vartheta_z - \theta_z) \sin \theta_z - \sin \theta_z / \sin(\vartheta_z - \theta_z) \\[2mm] h_{2m+1, 2m+1} &\odot& (\cos(\vartheta_z - \theta_z)-1) / \sin(\vartheta_z - \theta_z) \end{array} $}$\\ \If{${\ell(z)} > i$}{ $\scalebox{0.90}{$ \begin{array}{rcl} \left( \begin{array}{cc} h_{2\ell(z)-1, 2i-1} & h_{2\ell(z)-1, 2i}\\ h_{2\ell(z), 2i-1} & h_{2\ell(z), 2i} \end{array} \right) &\odot& -(\sin(\vartheta_z - \theta_z))^{-1} \left( \begin{array}{cc} \cos \theta_z \cos \vartheta_z & \sin \theta_z \cos \vartheta_z \\ \cos \theta_z \sin \vartheta_z & \sin \theta_z \sin \vartheta_z \end{array} \right) \end{array} $}$ } } } } } } \Return $H$ \end{algorithm} Algorithms~\ref{G}, \ref{grad_G}, and \ref{hessian_G} depend on the computation of sets $\mathbb{E}_i$ and $\mathbb{A}_i$ for $i=1,\dots,m$. Computing these sets requires (a) to compute the Voronoi diagram with cells $\{V_i\}_{i=1}^m$ associated with the balls' centers $x_1,\dots,x_m$ and (b) for each $i \in \{1,\dots,m\}$ and $j \in \{1,\dots,p\}$, to compute $W_{ij} = V_i \cap A_j$ and $S_{ij} = W_{ij} \cap B(x_i,r)$. Computing the Voronoi diagram, using for example Fortune's algorithm~\cite{fortune}, has known time complexity $\mathcal{O}(m\log{m})$~\cite[Lem.~7.9, p.158]{deberg}. Since the intersection between a two-dimensional polyhedron defined by~$a$ half-planes and a convex polygon with~$b$ sides can be computed in $\mathcal{O}(ab)$~\cite{horowitz1992}, all $W_{ij}$ can be computed in \begin{equation} \label{primc} \mathcal{O}(\sum_{i=1}^m{\sum_{j=1}^p}{e_{V_i} e_{A_j}}), \end{equation} where $e_{V_i}$ is the number of half-planes that define $V_i$, for $i=1,\dots,m$, and $e_{A_j}$ is the number of sides of each $A_j$, for $j=1,\dots,p$. However, it is also known~\cite[Thm.7.3, p.150]{deberg} that a Voronoi diagram generated by $m\geq3$ points has at most $3m-6$ edges; and since each edge is part of exactly two cells, we have that $\sum_{i=1}^me_{V_i}=\mathcal{O}(m)$. Thus, \eqref{primc} reduces to $\mathcal{O}(m \sum_{j=1}^p e_{A_j})$. By construction, it also holds that $\sum_{i=1}^m{|\mathbb{E}_i|}$ is $\mathcal{O}(m \sum_{j=1}^p e_{A_j})$. Finally, a simple inspection of Algorithm~\ref{inter_pol_circle}, used to compute $S_{ij} = W_{ij} \cap B(x_i,r)$, shows that the computational effort required to compute all $S_{ij}$, as well as $\sum_{i=1}^m{|\mathbb{A}_i|}$, are both $\mathcal{O}(m \sum_{j=1}^p e_{A_j})$. This implies that the worst-case time complexity of Algorithms~\ref{G}, \ref{grad_G}, and \ref{hessian_G} is $\mathcal{O}(m\log{m} + m \sum_{j=1}^p e_{A_j})$. \section{Numerical experiments} \label{sec:num_exp} In this section, we aim to illustrate the capabilities and limitations of the proposed approach. We implemented Algorithms~\ref{G}, \ref{grad_G}, and \ref{hessian_G} in Fortran~90. Given the balls' centers $\{x_i\}_{i=1}^m$, the Voronoi diagram is computed with subroutine \textsc{Dtris2} from Geompack~\cite{geompack} (available at \url{https://people.math.sc.edu/Burkardt/f_src/geompack2/geompack2.html}). In fact, \textsc{Dtris2} provides a Delaunay triangulation from which the Voronoi diagram is extracted. The intersection $W_{ij}$ of each Voronoi cell $V_i$ (that is a bounded or unbounded polyhedron) and each convex polygon~$A_j$ is computed with the Sutherland-Hodgman algorithm~\cite{sutherland1974}. For each convex polygon~$W_{ij}$, the intersection $S_{ij}$ with the ball $B(x_i,r)$ is computed with an adaptation of a single iteration of the Sutherland-Hodgman algorithm, detailed as Algorithm~\ref{inter_pol_circle} in Appendix~\ref{appen:alg4}. Problem~\eqref{prob1} is a nonlinear programming problem of the form \begin{equation} \label{prob2} \mbox{Minimize } f(\boldsymbol{x},r) := r \mbox{ subject to } G(\boldsymbol{x},r)=0 \mbox{ and } r \geq 0 \end{equation} that can be tackled with an Augmented Lagrangian (AL) approach~\cite{bmbook}. In the numerical experiments, we considered the AL method Algencan~\cite{abmstango,bmbook,bmcomper}. Algencan~4.0, implemented in Fortran~90 and available at~\url{http://www.ime.usp.br/~egbirgin/tango/}, was considered. Algencan is an AL method with safeguards that, at each iteration, solves a bound-constrained subproblem. Since, in the present work, second-order derivatives are available, subproblems are solved with an active-set Newton's method; see~\cite{bmgencan} and~\cite[Ch.9]{bmbook} for details. When Algencan is applied to problem~(\ref{prob2}), on success, it finds $(\boldsymbol{x}^\star,r^\star,\lambda^\star)$ with $r^\star > 0$ satisfying \begin{equation} \label{kkt} \| \nabla f(\boldsymbol{x}^\star,r^\star) + \lambda^\star \nabla G(\boldsymbol{x}^\star,r^\star) \|_{\infty} \leq \varepsilon_{\mathrm{opt}} \mbox{ and } \| G(\boldsymbol{x}^\star,r^\star) \|_{\infty} \leq \varepsilon_{\mathrm{feas}}, \end{equation} where $\varepsilon_{\mathrm{feas}} > 0$ and $\varepsilon_{\mathrm{opt}} > 0$ are given feasibility and optimality tolerances, respectively; i.e., it finds a point that approximately satisfies KKT conditions for problem~(\ref{prob2}). Following~\cite{coveringfirst}, in order to enhance the probability of finding an approximation to a global minimizer, a simple multistart strategy with random initial guesses is employed; see~\cite[\S5]{coveringfirst} for details. In the numerical experiments, we considered $\varepsilon_{\mathrm{feas}} = \varepsilon_{\mathrm{opt}} = 10^{-8}$. In the numerical experiments, we considered (i) a non-convex polygon with holes already considered in~\cite{stoyan}, (ii) a sketch of a map of America available from~\cite[\S13.2]{bmbook} and already considered in~\cite{coveringfirst}, (iii) an eight-pointed star, (iv) iteration two of the Minkowski island fractal, and (v) iteration three of the Ces\`aro fractal; see Figures~\ref{fig:stoyan}a--\ref{fig:cesaro}a. In Figures~\ref{fig:stoyan}b--\ref{fig:cesaro}b, the way in which the problems were partitioned into convex polygons is made explicit. Appendix~\ref{appen:probs} presents an explicit description of each problem by exhibiting the vertices of each convex polygon that compose the problem. Fortran source code of Algorithms~\ref{G}, \ref{grad_G}, \ref{hessian_G}, and~\ref{inter_pol_circle}, the source code of the considered problems, as well as the source code necessary to reproduce all numerical experiments, is available at \url{http://www.ime.usp.br/~egbirgin/}. All tests were conducted on a computer with an AMD Opteron 6376 processor and 256GB 1866 MHz DDR3 of RAM memory, running Debian GNU/Linux (version 9.13--stretch). Code was compiled by the GFortran compiler of GCC (version 6.3.0) with the -O3 optimization directive enabled. In the experiments, we covered the five considered regions with $m \in\{10, 20, \dots, 100 \}$ balls. For each problem and each considered value of $m$, the multistart strategy makes $10{,}000$ attempts, i.e.\ $10{,}000$ different random initial guesses are considered. Table~\ref{tab:results} and Figures~\ref{fig:stoyan}--\ref{fig:cesaro} show the results. In Table~\ref{tab:results}, $r^*$ represents the smallest obtained radius, $G(\boldsymbol{x}^*,r^*)$ corresponds to the value of $G$ at the obtained solution, and ``trial'' is the ordinal of the initial guess that yields the smallest radius. In addition, some performance metrics are also displayed in the remaining columns of the table. ``outit'' and ``innit'' correspond to the so called outer and inner iterations of the AL method, respectively, ``Alg.1'', ``Alg.2'', and ``Alg.3'' correspond to the number of calls to Algorithms~\ref{G}, \ref{grad_G}, and~\ref{hessian_G}, respectively, i.e.\ to the number of evaluations of~$G$, $\nabla G$, and $\nabla^2 G$ that were required in the optimization process, and ``CPU Time'' corresponds to the elapsed CPU time in seconds. All these performance metrics correspond to the trial that leads to the smallest radius for a given problem and a given number of balls~$m$. Thus, the whole process took approximately $10{,}000$ times this effort. Clearly, the overall cost can be reduced by reducing the number of trials. Figure~\ref{bestr} illustrates, for the ``non-convex with holes problem'' with $m \in \{10, 20, \dots, 100\}$, the best obtained radius as a function of the number of trials. The picture shows that, for all values of $m$, good quality local minimizers are found with less than 100 trials and that in the remaining 99\% additional trials only marginal improvements are obtained. \begin{figure}[ht!] \begin{center} \includegraphics{graph-rxtrials.pdf} \end{center} \caption{Best radius $r^*$ for $m \in \{10, 20, \dots, 100\}$ as a function of the number of trials in the multistart globalization strategy.} \label{bestr} \end{figure} As a whole, numerical experiments show that, by using second-order information, the AL method is able to find high-precision local solutions efficiently. It is worth noticing that, as shown in column $G(\boldsymbol{x}^*,r^*)$ of Table~\ref{tab:results}, using $\varepsilon_{\mathrm{feas}}=10^{-8}$ means that the area of the region~$A$ to be covered and the covered region $A \cap \Omega(\boldsymbol{x}^*,r^*)$ coincide in eight significant digits. Since, in the considered problems, the region with largest area has area equal to~16 (see the description of the problem in Appendix~\ref{appen:probs}), this means that reported solutions cover more than 99.999999\% of the region. This precision is in contrast with the relatively low-quality solutions obtained with the approximate procedure considered in~\cite{coveringfirst}. A scaled versions of the non-convex region with holes considered in the present work was also considered in~\cite{stoyan}, where radius $r^* = 16.6176655 / 150 \approx 0.110784446$ and $r^*= 14.07100757 / 150 \approx 0.09380671713$ for the cases with $m=30$ and $m=40$ were reported. A direct comparison is not possible, because the balls' centers and the covering's precision of these solutions was not reported in~\cite{stoyan}. Anyway, smaller radii were found for these two cases in the present work, namely, $r^* = 0.10944963099046681$ and $r^*=0.092110416532448419$, respectively. The region that represents a sketch of the map of America was also considered in~\cite{coveringfirst}. Solutions presented in~\cite{coveringfirst} are not comparable to the ones presented here. The latter are much more precise and can be found with much less effort. To put the practical performance of the current approach in perspective in relation to the practical performance of the method implemented in~\cite{coveringfirst}, consider the trivial configuration depicted in Figure~\ref{fig:trivial}. The configuration shows a square of side three with the bottom-left corner at the origin and two unitary-radius balls with centers $x_1=(0,3)^\top$ and $x_2=(1.2, 1.7)^\top$. The covered area can be computed analytically and it is given by $\mathrm{Vol}(A \cap \Omega(\boldsymbol{x},r)) = 5\pi/4 - 2\arccos(d/2) + d \sqrt{1 - (d/2)^2} \approx 3.781718647855564$, where $d=\|x_1-x_2\|$. Algorithm~\ref{G} computes this quantity up to the machine precision in $10^{-6}$ seconds of CPU time. Algorithm~1 from~\cite{coveringfirst}, devised to cover more general non-polygonal regions, approximates a covered area with precision $O(h)$ at cost $O(h^2)$ by partitioning a region $D$ that contains $A$ in small squares of side~$h$, where $h>0$ is a given parameter. In this specific trivial example, it takes 271.92 seconds of CPU time to compute the covered area with half of the machine precision using $h=10^{-5}$. (With $h=10^{-3}$ and $h=10^{-4}$, four and six correct decimal digits are obtained, by consuming 0.024 and 2.4 seconds of CPU time, respectively. Moreover, the cost is proportional to the area of $D$, which is as small as possible since we considered $D=A$.) So, in this trivial example, we showed that the approach proposed in the present work computes the covered area with twice the number of correct digits with a computational cost that is eight orders of magnitude smaller (i.e.,\ a hundred million times faster) than the cost of the approach proposed in~\cite{coveringfirst}, thus dramatically improving the computational efficiency. This, together with a similar state of things with respect to the computation of~$\nabla G$, plus the computation of $\nabla^2 G$ that is absent in~\cite{coveringfirst}, justify the much higher quality of the obtained results. \begin{figure}[ht!] \centering \includegraphics{config-comp-covfirst.mps} \caption{A trivial example that illustrates the comparison between the exact computation of~$G$ introduced in the current work and the approximate scheme considered in~\cite{coveringfirst}.} \label{fig:trivial} \end{figure} \begin{table}[ht!] \begin{center} \resizebox{0.9\textwidth}{!}{ \begin{tabular}{|c|c|c|c|r|r|r|r|r|r|c|} \hline & $m$ & $r^*$ & $G(\boldsymbol{x}^*,r^*)$ & \multicolumn{1}{c|}{trial} & \multicolumn{1}{c|}{outit} & \multicolumn{1}{c|}{innit} & \multicolumn{1}{c|}{Alg.1} & \multicolumn{1}{c|}{Alg.2} & \multicolumn{1}{c|}{Alg.3} & \multicolumn{1}{c|}{CPU Time} \\ \hline \hline \multirow{10}{*}{\rotatebox{90}{Non-convex with holes}} &10 & 1.9546630973359513e$-$01 & 5.2e$-$09 & 7078 & 23 & 154 & 538 & 388 & 384 & 0.33 \\ &20 & 1.3277721146997093e$-$01 & 4.2e$-$09 & 4580 & 21 & 123 & 426 & 345 & 333 & 0.56\\ &30 & 1.0944963099046681e$-$01 & 9.9e$-$09 & 7155 & 22 & 187 & 1154 & 413 & 407 & 1.48\\ &40 & 9.2110416532448419e$-$02 & 9.3e$-$09 & 8981 & 21 & 209 & 847 & 432 & 419 & 1.85\\ &50 & 8.2059696677895658e$-$02 & 9.0e$-$09 & 3176 & 21 & 218 & 937 & 450 & 428 & 2.57\\ &60 & 7.3972529936974535e$-$02 & 8.4e$-$09 & 7718 & 22 & 245 & 1750 & 484 & 465 & 4.54\\ &70 & 6.8954683287629770e$-$02 & 9.0e$-$09 & 2942 & 20 & 209 & 1228 & 421 & 409 & 4.35\\ &80 & 6.4065368587975027e$-$02 & 7.5e$-$09 & 8908 & 21 & 209 & 1366 & 419 & 419 & 5.69\\ &90 & 6.0345840506149377e$-$02 & 7.7e$-$09 & 3741 & 23 & 263 & 2595 & 500 & 493 & 9.71\\ &100 & 5.7226511303503126e$-$02 & 6.9e$-$09 & 2619 & 20 & 225 & 1390 & 448 & 425 & 5.45\\ \hline \multirow{10}{*}{\rotatebox{90}{Sketch of America map}} &10 & 1.1022680217297048e$-$01 & 6.2e$-$09 & 7191 & 22 & 226 & 1198 & 434 & 446 & 0.91\\ &20 & 7.0566193751253600e$-$02 & 4.2e$-$09 & 558 & 21 & 256 & 1541 & 455 & 466 & 2.14\\ &30 & 5.6728945376199408e$-$02 & 3.7e$-$09 & 3341 & 20 & 241 & 1451 & 428 & 441 & 3.36\\ &40 & 4.8479681841390981e$-$02 & 5.3e$-$09 & 7518 & 21 & 274 & 1227 & 506 & 484 & 4.18\\ &50 & 4.3079623896669902e$-$02 & 4.6e$-$09 & 9471 & 22 & 190 & 915 & 405 & 410 & 3.82\\ &60 & 3.8669223381267957e$-$02 & 9.0e$-$09 & 6539 & 22 & 328 & 2124 & 544 & 548 & 9.17\\ &70 & 3.5479536239229441e$-$02 & 9.3e$-$09 & 2774 & 20 & 290 & 1864 & 508 & 490 & 10.81\\ &80 & 3.3035213466515133e$-$02 & 3.7e$-$09 & 9176 & 23 & 281 & 1098 & 529 & 511 & 8.94\\ &90 & 3.1081859427563651e$-$02 & 9.4e$-$09 & 1815 & 20 & 296 & 967 & 528 & 496 & 11.20\\ &100 & 2.9185582405640495e$-$02 & 7.3e$-$09 & 2427 & 21 & 302 & 1271 & 525 & 512 & 10.55\\ \hline \multirow{10}{*}{\rotatebox{90}{Eight-pointed star}} &10 & 1.3040713549156926e$+$00 & 7.4e$-$09 & 2129 & 28 & 212 & 1405 & 471 & 492 & 0.40\\ &20 & 7.2447962534018184e$-$01 & 6.6e$-$09 & 1569 & 28 & 383 & 3437 & 682 & 663 & 1.71\\ &30 & 5.5386599521018731e$-$01 & 4.4e$-$09 & 9204 & 28 & 241 & 971 & 539 & 521 & 1.21\\ &40 & 4.6618323934219452e$-$01 & 4.1e$-$09 & 9298 & 28 & 312 & 1999 & 614 & 592 & 2.68\\ &50 & 4.1522639848076626e$-$01 & 3.7e$-$09 & 759 & 27 & 276 & 1974 & 572 & 546 & 3.51\\ &60 & 3.7211553871395336e$-$01 & 1.0e$-$08 & 8549 & 27 & 278 & 2235 & 568 & 541 & 4.74\\ &70 & 3.3883252892004639e$-$01 & 9.5e$-$09 & 3297 & 26 & 247 & 818 & 545 & 507 & 3.53\\ &80 & 3.1591211839929362e$-$01 & 8.9e$-$09 & 3712 & 26 & 266 & 1160 & 559 & 526 & 4.46\\ &90 & 2.9594385965306919e$-$01 & 8.8e$-$09 & 257 & 27 & 309 & 3063 & 613 & 579 & 10.15\\ &100 & 2.7907469799758938e$-$01 & 8.4e$-$09 & 8809 & 26 & 274 & 2596 & 540 & 533 & 7.12\\ \hline \multirow{10}{*}{\rotatebox{90}{Minkowski island fractal}} &10 & 9.9730787966959566e$-$01 & 5.8e$-$09 & 85 & 28 & 269 & 2287 & 490 & 549 & 0.79\\ &20 & 6.4157361024666815e$-$01 & 5.4e$-$09 & 114 & 29 & 233 & 943 & 533 & 523 & 1.03\\ &30 & 5.3259264476359935e$-$01 & 4.9e$-$09 & 9963 & 26 & 303 & 1059 & 587 & 563 & 1.74\\ &40 & 4.4275330752709730e$-$01 & 4.0e$-$09 & 9678 & 27 & 323 & 3034 & 587 & 593 & 4.22\\ &50 & 3.9534726462521569e$-$01 & 9.7e$-$09 & 3428 & 24 & 230 & 1155 & 479 & 470 & 2.82\\ &60 & 3.4918562471568843e$-$01 & 8.8e$-$09 & 8144 & 26 & 248 & 775 & 522 & 508 & 3.15\\ &70 & 3.2807457983514665e$-$01 & 9.3e$-$09 & 4385 & 27 & 278 & 2414 & 549 & 548 & 7.06\\ &80 & 3.1016946802464157e$-$01 & 9.6e$-$09 & 7306 & 27 & 298 & 3068 & 570 & 568 & 9.58\\ &90 & 2.9050989196451837e$-$01 & 8.6e$-$09 & 9902 & 26 & 346 & 3204 & 591 & 606 & 12.18\\ &100 & 2.7525512468934971e$-$01 & 9.0e$-$09 & 7719 & 26 & 377 & 2929 & 672 & 637 & 10.51\\ \hline \multirow{10}{*}{\rotatebox{90}{Cesàro fractal}} &10 & 2.1276864595120507e$-$01 & 5.5e$-$09 & 7054 & 22 & 180 & 1348 & 377 & 400 & 0.63\\ &20 & 1.3326878209070328e$-$01 & 3.8e$-$09 & 4870 & 23 & 278 & 1152 & 421 & 508 & 1.22\\ &30 & 1.0522163653090458e$-$01 & 4.1e$-$09 & 7850 & 23 & 245 & 1219 & 486 & 475 & 2.18\\ &40 & 9.3428035096055656e$-$02 & 9.4e$-$09 & 2646 & 21 & 193 & 871 & 424 & 403 & 2.43\\ &50 & 8.3314180748730718e$-$02 & 9.4e$-$09 & 1317 & 23 & 197 & 1322 & 441 & 427 & 3.68\\ &60 & 7.8415153849036370e$-$02 & 8.8e$-$12 & 9229 & 31 & 433 & 3704 & 674 & 743 & 10.83\\ &70 & 7.0460470988540802e$-$02 & 8.4e$-$09 & 5859 & 22 & 289 & 1698 & 544 & 509 & 6.96\\ &80 & 6.6110791995596219e$-$02 & 8.6e$-$09 & 7697 & 21 & 329 & 1694 & 581 & 539 & 9.38\\ &90 & 6.1956278506660224e$-$02 & 7.8e$-$09 & 5722 & 23 & 318 & 3185 & 568 & 548 & 15.12\\ &100 & 5.8465961897078852e$-$02 & 8.6e$-$09 & 3205 & 21 & 296 & 1729 & 548 & 506 & 8.32\\ \hline \end{tabular}} \end{center} \caption{Details of the obtained solutions and performance metrics of the application of Algencan to the five considered covering problems.} \label{tab:results} \end{table} \begin{figure}[ht!] \begin{center} \begin{tabular}{ccc} \includegraphics[scale=0.8]{region-stoyan.mps} & \includegraphics[scale=0.8]{stoyanpartition.mps} & \includegraphics[scale=0.8]{covering-stoyan10-10.mps} \\ (a) Region & (b) Partition & (c) $m=10$ \\ \includegraphics[scale=0.8]{covering-stoyan20-20.mps} & \includegraphics[scale=0.8]{covering-stoyan30-30.mps} & \includegraphics[scale=0.8]{covering-stoyan40-40.mps} \\ (d) $m=20$ & (e) $m=30$ & (f) $m=40$ \\ \includegraphics[scale=0.8]{covering-stoyan50-50.mps} & \includegraphics[scale=0.8]{covering-stoyan60-60.mps} & \includegraphics[scale=0.8]{covering-stoyan70-70.mps} \\ (g) $m=50$ & (h) $m=60$ & (i) $m=70$ \\ \includegraphics[scale=0.8]{covering-stoyan80-80.mps} & \includegraphics[scale=0.8]{covering-stoyan90-90.mps} & \includegraphics[scale=0.8]{covering-stoyan100-100.mps} \\ (j) $m=80$ & (k) $m=90$ & (l) $m=100$ \\ \end{tabular} \end{center} \caption{(a) Non-convex polygon with holes considered in~\cite{stoyan}, partitioned into $p=14$ convex polygons as depicted in (b). Pictures from (c) to (l) display the solutions found with $m \in \{10,\dots,100\}$.} \label{fig:stoyan} \end{figure} \begin{figure}[ht!] \begin{center} \begin{tabular}{cccc} \includegraphics[scale=0.6]{region-america.mps} & \includegraphics[scale=0.6]{americapartition.mps} & \includegraphics[scale=0.6]{covering-america10-10.mps} & \includegraphics[scale=0.6]{covering-america20-20.mps} \\ (a) Region & (b) Partition & (c) $m=10$ & (d) $m=20$ \\ \includegraphics[scale=0.6]{covering-america30-30.mps} & \includegraphics[scale=0.6]{covering-america40-40.mps} & \includegraphics[scale=0.6]{covering-america50-50.mps} & \includegraphics[scale=0.6]{covering-america60-60.mps} \\ (e) $m=30$ & (f) $m=40$ & (g) $m=50$ & (h) $m=60$ \\ \includegraphics[scale=0.6]{covering-america70-70.mps} & \includegraphics[scale=0.6]{covering-america80-80.mps} & \includegraphics[scale=0.6]{covering-america90-90.mps} & \includegraphics[scale=0.6]{covering-america100-100.mps} \\ (i) $m=70$ & (j) $m=80$ & (k) $m=90$ & (l) $m=100$ \\ \end{tabular} \end{center} \caption{(a) Sketch of America available from~\cite[\S13.2]{bmbook} and already considered in~\cite{coveringfirst}, partitioned into $p=34$ convex polygons as depicted in (b). Pictures from (c) to (l) display the solutions found with $m \in \{10,\dots,100\}$.} \label{fig:america} \end{figure} \begin{figure}[ht!] \begin{center} \begin{tabular}{ccc} \includegraphics[scale=0.75]{region-star.mps} & \includegraphics[scale=0.75]{starpartition.mps} & \includegraphics[scale=0.75]{covering-star10-10.mps} \\ (a) Region & (b) Partition & (c) $m=10$ \\ \includegraphics[scale=0.75]{covering-star20-20.mps} & \includegraphics[scale=0.75]{covering-star30-30.mps} & \includegraphics[scale=0.75]{covering-star40-40.mps} \\ (d) $m=20$ & (e) $m=30$ & (f) $m=40$ \\ \includegraphics[scale=0.75]{covering-star50-50.mps} & \includegraphics[scale=0.75]{covering-star60-60.mps} & \includegraphics[scale=0.75]{covering-star70-70.mps} \\ (g) $m=50$ & (h) $m=60$ & (i) $m=70$ \\ \includegraphics[scale=0.75]{covering-star80-80.mps} & \includegraphics[scale=0.75]{covering-star90-90.mps} & \includegraphics[scale=0.75]{covering-star100-100.mps} \\ (j) $m=80$ & (k) $m=90$ & (l) $m=100$ \\ \end{tabular} \end{center} \caption{(a) Eight-pointed star, partitioned into $p=9$ convex polygons as depicted in (b). Pictures from (c) to (l) display the solutions found with $m \in \{10,\dots,100\}$.} \label{fig:star} \end{figure} \begin{figure}[ht!] \begin{center} \begin{tabular}{ccc} \includegraphics[scale=0.9]{region-minkowski.mps} & \includegraphics[scale=0.9]{minkowskipartition.mps} & \includegraphics[scale=0.9]{covering-minkowski10-10.mps} \\ (a) Region & (b) Partition & (c) $m=10$ \\ \includegraphics[scale=0.9]{covering-minkowski20-20.mps} & \includegraphics[scale=0.9]{covering-minkowski30-30.mps} & \includegraphics[scale=0.9]{covering-minkowski40-40.mps} \\ (d) $m=20$ & (e) $m=30$ & (f) $m=40$ \\ \includegraphics[scale=0.9]{covering-minkowski50-50.mps} & \includegraphics[scale=0.9]{covering-minkowski60-60.mps} & \includegraphics[scale=0.9]{covering-minkowski70-70.mps} \\ (g) $m=50$ & (h) $m=60$ & (i) $m=70$ \\ \includegraphics[scale=0.9]{covering-minkowski80-80.mps} & \includegraphics[scale=0.9]{covering-minkowski90-90.mps} & \includegraphics[scale=0.9]{covering-minkowski100-100.mps} \\ (j) $m=80$ & (k) $m=90$ & (l) $m=100$ \\ \end{tabular} \end{center} \caption{(a) Minkowski island fractal, partitioned into $p=16$ convex polygons as depicted in (b). Pictures from (c) to (l) display the solutions found with $m \in \{10,\dots,100\}$.} \label{fig:minkowski} \end{figure} \begin{figure}[ht!] \begin{center} \begin{tabular}{ccc} \includegraphics[scale=0.8]{region-cesaro.mps} & \includegraphics[scale=0.8]{cesaropartition.mps} & \includegraphics[scale=0.8]{covering-cesaro10-10.mps} \\ (a) Region & (b) Partition & (c) $m=10$ \\ \includegraphics[scale=0.8]{covering-cesaro20-20.mps} & \includegraphics[scale=0.8]{covering-cesaro30-30.mps} & \includegraphics[scale=0.8]{covering-cesaro40-40.mps} \\ (d) $m=20$ & (e) $m=30$ & (f) $m=40$ \\ \includegraphics[scale=0.8]{covering-cesaro50-50.mps} & \includegraphics[scale=0.8]{covering-cesaro60-60.mps} & \includegraphics[scale=0.8]{covering-cesaro70-70.mps} \\ (g) $m=50$ & (h) $m=60$ & (i) $m=70$ \\ \includegraphics[scale=0.8]{covering-cesaro80-80.mps} & \includegraphics[scale=0.8]{covering-cesaro90-90.mps} & \includegraphics[scale=0.8]{covering-cesaro100-100.mps} \\ (j) $m=80$ & (k) $m=90$ & (l) $m=100$ \\ \end{tabular} \end{center} \caption{(a) Ces\`aro fractal, partitioned into $p=21$ convex polygons as depicted in (b). Pictures from (c) to (l) display the solutions found with $m \in \{10,\dots,100\}$.} \label{fig:cesaro} \end{figure} \section{Final considerations} \label{sec:conclusion} From the shape optimization perspective, the present work completes \cite{coveringfirst} with a second-order shape sensitivity analysis for nonsmooth domains defined as a union of balls intersected with the domain to be covered. The analysis of several singular cases seems to indicate that the assumptions used to derive $\nabla^2 G$ cannot be weakened. From the practical point of view, the exact calculation of $G$ and its first- and second-order derivatives represents the possibility, absent in~\cite{coveringfirst}, of solving very efficiently and with high accuracy, problems in which the area to be covered is given by a non-convex polygon. We now discuss potential extensions of our approach. Redefining $\Omega(\boldsymbol{x},\boldsymbol{r}):=\cup_{i=1}^m B(x_i,r_i)$ and $G(\boldsymbol{x},\boldsymbol{r}):=\operatorname{Vol}(A \setminus \Omega(\boldsymbol{x},\boldsymbol{r}))$, where $\boldsymbol{r}:=\{r_i\}_{i=1}^m$, expressions and algorithms to approximate $G(\boldsymbol{x},\boldsymbol{r})$, $\nabla G(\boldsymbol{x},\boldsymbol{r})$ and $\nabla^2 G(\boldsymbol{x},\boldsymbol{r})$ can be obtained with straightforward modifications to the introduced approach. From the practical point of view, underlying partitions that lead to exact calculations might be implemented using power diagrams~\cite{powerdia1,powerdia2}. We observe that formulae (\ref{derivadasegunda}, \ref{d2Grr}, \ref{d2xixiG}, \ref{d2Gxixj}, \ref{eq:781}) are valid for general sets $A$ satisfying Assumptions~\ref{a1} and~\ref{a3}, but the exact numerical computation of $G$, $\nabla G$ and $\nabla^2 G$ requires $A$ to be a union of non-overlapping convex polygons. The exact calculation of $\nabla G$ and $\nabla^2 G$ can actually be performed for any set $A$ such that the intersections of $\partial A$ with circles can be computed analytically. However, the possibilities of computing $G$ exactly are more restricted as this requires the computation of integrals on subsets of $\partial A$. In some specific cases, this calculation could be done exactly, for instance when $A$ is a union of balls. Nevertheless, in more general cases the integrals on subsets of $\partial A$ could be efficiently approximated with high accuracy. The case where $\Omega(\boldsymbol{x},r)$ is a union of objects with arbitrary (sufficiently smooth) shapes is challenging and would require a generalization of the techniques developed in \cite{coveringfirst} and in the present paper. A key idea of our construction of the mappings $T_t$, which is still valid for objects with arbitrary shapes, is that the value of $T_t$ at the intersection points of the objects' boundaries (or the intersections with $\partial A$) is fully determined by the motion of these singular points, whereas the value of $T_t$ at the regular points of $\partial\Omega(\boldsymbol{x},r)$ is underdetermined. When the objects are balls, this underdetermination is conveniently resolved using polar coordinates to extend $T_t$ to the regular parts of $\partial\Omega(\boldsymbol{x},r)$. In the case of arbitrary shaped-objects however, a more general construction is required. A generalization to three dimensions of the nonsmooth shape optimization techniques developed in \cite{coveringfirst} and in the present paper is conceivable but would also require a more general approach to build $T_t$. Another interesting direction for future investigations would be the application of these techniques for optimization problems involving partial differential equations. The calculation of the shape derivatives would depend on the specific partial differential equation, but the construction of the transformations $T_t$ would remain the same. \bibliographystyle{plain}

https://arxiv.org/abs/1012.0096

Isomorphisms of Algebraic Number Fields

Let $\mathbb{Q}(\alpha)$ and $\mathbb{Q}(\beta)$ be algebraic number fields. We describe a new method to find (if they exist) all isomorphisms, $\mathbb{Q}(\beta) \rightarrow \mathbb{Q}(\alpha)$. The algorithm is particularly efficient if the number of isomorphisms is one.

\section{Introduction} Let $\mathbb{Q}(\alpha)$ and $\mathbb{Q}(\beta)$ be two number fields, given by the minimal polynomials $f(x)=\sum_{i=0}^n{f_i x^i}$ and $g(x) = \sum_{i=0}^n {g_i x^i}$ of $\alpha$ and $\beta$ respectively. In this paper we give an algorithm to compute the isomorphisms $\mathbb{Q}(\beta) \rightarrow \mathbb{Q}(\alpha)$. Suppose there is an isomorphism then we have the following diagram of field extensions: $$ \xymatrix{ \mathbb{Q}(\beta) \ar@{-}[d]^{g(x)} \ar[r]^{\cong} &\mathbb{Q}(\alpha)\ar@{-}[d]^{f(x)}\\ \mathbb{Q} &\mathbb{Q}} $$ \par To represent such an isomorphism we need to give the image of $\beta$ in $\mathbb{Q}(\alpha)$, in other words, we need to give a root of $g(x)$ in $\mathbb{Q}(\alpha)$. \par We now describe two common methods of computing isomorphisms of number fields. \begin{enumerate} \item[Method I.] Field Isomorphism Using Polynomial Factorization [11, Algorithm 4.5.6] \begin{itemize} \item Find all roots of $g$ in $\mathbb{Q}(\alpha)$. Each corresponds to an isomorphism $\mathbb{Q}(\beta) \rightarrow \mathbb{Q}(\alpha)$. The roots can be found by factoring $g$ over $\mathbb{Q}(\alpha)$. \begin{enumerate} \item If done with Trager's method, one ends up factoring a polynomial in $\mathbb{Q}[x]$ of degree $n^2$. \item An alternative is Belabas' algorithm for factoring in $\mathbb{Q}(\alpha)$. \end{enumerate} \end{itemize} \item[Method II.] Field Isomorphism Using Linear Algebra [11, Algorithm 4.5.1/4.5.5] \begin{enumerate} \item Let $\alpha_1, \dots, \alpha_d$ be the roots of $f$ in $\mathbb{Q}_p$ (choose $p$ with $d>0$). \item Let $\beta_1, \dots, \beta_d$ be the roots of $g$ in $\mathbb{Q}_p$. \item If $\beta \mapsto h(\alpha)$ is an isomorphism, then $h(\alpha_1) = \beta_i$ for some $i\in \{1,\dots,d\}$. \item Run a loop $i=1,\dots,d$ and for each $i$, use LLL[9] techniques to check if there exists a polynomial $h(x)\in \mathbb{Q}[x]_{<n}$ for which $h(\alpha_1)=\beta_i$. \end{enumerate} \end{enumerate} \par Method I is often fast, but one can give examples where it becomes slow, e.g. for so-called Swinnerton-Dyer polynomials, the degree $n^2$ factoring leads to a lattice reduction in [VH 2002] of dimension approximately $n^2/2$. Method I(b) can be faster, but one can still produce examples where it becomes slow, e.g. [7]. For such examples method II is faster because the lattice reduction there has dimension approximately $n$. \par Our algorithm is similar to Method II. There can be $n$ distinct isomorphisms (in the Galois case) and in this case our algorithm is the same as Algorithm II. However, if there is only one isomorphism then we can save roughly a factor $d$. This is because we can do the LLL computation for all $\beta_i$ simultaneously. \section{Preliminaries} \par We let $\mathbb{Q}[x]_{<n}$ denote the polynomials over $\mathbb{Q}$ with degree less than $n$. \begin{defi} If $h(\alpha) \in \mathbb{Q}(\alpha)$ then the notation $h(x)$ is the element of $\mathbb{Q}[x]_{<n}$ that corresponds to $h(\alpha)$ under $x \mapsto \alpha$. \end{defi} \par Under an isomorphism $\beta$ will map to some $ h(\alpha) \in \mathbb{Q}(\alpha)$, \begin{equation} \beta \mapsto h(\alpha) = \sum_{i=0}^{n-1} {h_i \alpha^i} \end{equation}. \par A polynomial $h(x) \in \mathbb{Q}[x]_{<n}$ represents an isomorphism if and only if $h(\alpha)$ is a root of $g$, i.e. $g(h(\alpha)) = 0$. \par If $\mathbb{Q}(\beta)$ is isomorphic to $\mathbb{Q}(\alpha)$ then $g$ and $f$ have the same factorization pattern in $\mathbb{Q}_p[x]$ for every prime $p$. To simplify the factoring in $\mathbb{Q}_p[x]$ we restrict to good primes $p$, defined as: \begin{defi} A good prime $p$ is one that does not divide the leading coefficient of $f$ or $g$ and does not divide the discriminant of either $f$ or $g$. \end{defi} \begin{rmk} Both $f$ and $g$ are taken to be in $\mathbb{Z}[x]$. \end{rmk} \par For a good prime $p$ we can factor $f$ in $\mathbb{Q}_p[x]$ up to any desired $p$-adic precision by factoring in $\mathbb{F}_p[x]$, followed by Hensel Lifting [11, p. 137]. Likewise we can distinct-degree factor $f$ as: \begin{equation} \label{distdeg} f=F_1 F_2\dots F_m \text{ in } \mathbb{Q}_p[x] \end{equation} where $F_d$ is the product of all irreducible factors of $f$ in $\mathbb{Q}_p[x]$ of degree $d$ [11, Section 3.4.3]. \begin{defi} {\bf Sub-traces} Let $p$ be a prime and $d$ a positive integer. Then we define the $\mathbb{Q}$-linear map: $$ Tr^d _p(f,*) : \mathbb{Q}(\alpha) \rightarrow \mathbb{Q}_p$$ as follows. Let $h(x) \in \mathbb{Q}[x]_{<n}$, $h(\alpha) \in \mathbb{Q}(\alpha)$ and $F_d$ as above, then: $$ Tr^d _p (f,h(\alpha)) := \sum _{\stackrel{\gamma \in \overline{\mathbb{Q}}_p}{F_d(\gamma)=0}} {h(\gamma)} $$ We call these maps {\em sub-traces} because the sum is taken over a subset of the roots of $f$. \\ Likewise we define $Tr^d_p(g,*): \mathbb{Q}(\beta) \rightarrow \mathbb{Q}_p$. \end{defi} \begin{rmk} \label{trace} The map $Tr^d_p$ does not depend on the choice of the minimal polynomial $f$ that is used to represent the number field. In particular if $\beta \mapsto h(\alpha)$ is an isomorphism $\mathbb{Q}(\beta) \rightarrow \mathbb{Q}(\alpha)$ then $$Tr^d_p(g,\beta) = Tr^d_p (f,h(\alpha)) \text{ for every } p, d$$ \end{rmk} \begin{defi} We will now define two bases of $\mathbb{Q}(\alpha)$ that we will need. The first one is the standard basis, which is $\{ 1, \alpha, \alpha^2, \dots, \alpha^{n-1} \}$. The second will be called the { \em rational representation basis}, which is \\ $\{1/f'(\alpha), \alpha/f'(\alpha), \dots, \alpha^{n-1}/f'(\alpha) \} $. \end{defi} \par Rational representation can improve running time and complexity results, see [4]. This representation has also been used under various names, see [2, 4], and occurs naturally in algebraic number theory as a dual basis under the trace operator, see [2]. \par A basis for $\mathbb{Q}(\alpha)$ corresponds to a map $\rho: \mathbb{Q}^n \rightarrow \mathbb{Q}(\alpha)$. We use the rational representation basis, therefore $$\rho: (a_0, a_1, \dots, a_{n-1}) \mapsto \frac{1}{f'(\alpha)} \sum_{i=0}^{n-1} {a_i \alpha^i}.$$ \begin{defi} The inverse linear map $h(\alpha) \mapsto \vec{h}$, from $\mathbb{Q}(\alpha)$ to $\mathbb{Q}^n$ is as follows. Let $h(\alpha) = \sum^{n-1}_{i=0} {a_i \alpha^i} \in \mathbb{Q}(\alpha)$ and write $f'(\alpha) \cdot h(\alpha)$ as $\sum_{i=0}^{n-1} {b_i \alpha^i}$. Then define $\vec{h} := (b_0, b_1, \dots, b_{n-1}) \in \mathbb{Q}^n$. \end{defi} \begin{rmk} \label{reason} One of the advantages of rational representation is: by using the $b_i$ in $\vec{h}$ instead of the $a_i$, we have $\vec{h} \in \mathbb{Z}^n $ for every algebraic integer $h(\alpha)$, see lemma~\ref{AlgInt}. Moreover, as in [4] this also improves bounds (section 4). It is also better to use $g_n \vec{h}$ than simply using $h(\alpha)$ since $g_n \vec{h}$ will have integer components, by Corollary~\ref{coralgint}, which are easier to bound and are heuristically of smaller size [4, Section 6]. \end{rmk} For a polynomial $f(x)= \sum_{i=0}^n {f_i x^i}$ denote $$\|f(x)\| := \left( \sum_{i=0} ^n { |f_i|^2} \right) ^{1/2}.$$ \par Let $M(f)$ be the Mahler measure of $f$, $$M(f) := f_n \cdot \prod_{\stackrel{ f(\gamma)=0} { \gamma \in \mathbb{C}}} {\max{\{1, |\gamma| \}}}.$$ \section{Overview of the Algorithm} {\bf Goal}: To find all $g_n \vec{h} \in \mathbb{Z}^n$ for which $\beta \mapsto h(\alpha)$ defines an isomorphism $\mathbb{Q}(\beta) \rightarrow \mathbb{Q}(\alpha)$. \\ {\bf Idea}: The aim of the pre-processing algorithm in Section 5 is to find a sequence $$ \mathbb{Z}^n = L_0 \supseteq L_1 \supseteq L_2 \supseteq \dots \supseteq L_k$$ such that all $g_n \vec{h}$ are in each $L_i$. We can then use $L_k$ to speed up the computation of the isomorphism(s), especially when $\text{dim}(L_k)$ is small. The cost of computing $L_k$ is comparable to one iteration in Method II. \par In the algorithm we start with the lattice $\mathbb{Z}^n$ and then add the restrictions imposed by the condition that under an isomorphism, sub-traces $\mathbb{Q}(\alpha) \rightarrow \mathbb{Q}_p$ must correspond to sub-traces $\mathbb{Q}(\beta) \rightarrow \mathbb{Q}_p$. By doing this for several primes we are able to narrow down the possible isomorphisms. If dim($L_k$) $\leq 1$, this directly gives the isomorphism or shows that there is no isomorphism. If dim($L_k$) $>1$ then we switch to Method II, but starting with $L_k$. Thus we end up with $d$ lattice reductions of dimension dim($L_k$). In the worst case dim($L_k$) $ \approx n$, this costs the same as Method II. In the best case, dim($L_k$) $\leq 1$ and we save a factor $d$. \section{Bounding the length of $g_n \vec{h}$} \par To effectively carry out this algorithm we will need a good upper bound on the size of $g_n \vec{h}$. In this section we aim to find such a bound. \begin{defi} Let $\alpha_1, \dots, \alpha_n \in \mathbb{C}$ be the roots of $f$. Then using the basis $\{1,x,x^2, \dots, x^{n-1}\}$ of $\mathbb{C}[x]_{<n}$ and the standard basis $\{e_1, e_2, \dots, e_n\}$ for $\mathbb{C}^n$, the interpolation map $\mathbb{C}^n \rightarrow \mathbb{C}[x]_{<n}$ is given by: $$e_i \mapsto \frac{f(x)/(x-\alpha_i)}{f'(\alpha_i)}$$ This polynomial takes value $1$ at $x=\alpha_i$ and value $0$ at $x=\alpha_j$ ($i\neq j$). The inverse of the interpolation map is the evaluation map, which is given by the Vandermonde matrix: $$\begin{bmatrix} 1&\alpha_1&\alpha_1^2&\dots&\alpha_1^{n-1}\\ 1&\alpha_2&\alpha_2^2&\dots&\alpha_2^{n-1} \\ 1&\alpha_3&\alpha_3^2&\dots&\alpha_3^{n-1} \\ \vdots & \vdots &\ddots &\dots & \vdots \\ 1 & \alpha_n & \alpha_n^2 & \dots & \alpha_n^{n-1} \end{bmatrix}$$ \end{defi} \begin{lemma} \label{AlgInt} If $a \in \mathbb{Q}(\alpha)$ is an algebraic integer and $f(x)$ is the minimal polynomial for $\alpha$, then $f'(\alpha)\cdot a \in \mathbb{Z}[\alpha]$. \end{lemma} \begin{proof} Denote by $^{(i)}$ the complex embeddings of $\mathbb{Q}(\alpha)$. Then define $$ m(x) := \sum_{i=1}^n {a^{(i)} \frac{f(x)}{x-\alpha^{(i)}}}.$$ The coefficients of $m(x)$ are in $\mathbb{Q}$ since the polynomial is symmetric in the $\alpha^{(i)}$. But $m(x)$ is also a sum of polynomials all of whose entries are algebraic integers. Hence $m(x) \in \mathbb{Z}[x]$. Note that for $\alpha = \alpha^{(1)}$ we get $m(\alpha) = a f'(\alpha) \in \mathbb{Z}[\alpha]$. \end{proof} \begin{cor} \label{coralgint} Let $\beta \mapsto h(\alpha)$ be an isomorphism of $\mathbb{Q}(\beta)$ and $\mathbb{Q}(\alpha)$. Then $g_n h(\alpha)$ is an algebraic integer and hence $g_n \vec{h} \in \mathbb{Z}^n$. \end{cor} \begin{proof} Apply lemma~\ref{AlgInt} by letting $a = g_n h(\alpha)$ and recall that $g_n \vec{h}$ is comprised of the coefficients of $g_n f'(\alpha)h(\alpha)$ in the standard basis, each of which will be integers by lemma~\ref{AlgInt}. \end{proof} \begin{lemma} \label{inter} Let $P(x) = \sum_{i=1}^n{\beta_i \frac{f(x)}{x- \alpha_i}} \in \mathbb{Q}[x]_{<n}$, then $P(\alpha) = f'(\alpha) h(\alpha)$. \end{lemma} \begin{proof} If we evaluate $f'(x) h(x)$ at the roots of $f(x)$ and then interpolate we get: $$ \sum_{i=1}^n{\beta_i f'(\alpha_i) \frac{f(x)/(x-\alpha_i)}{f'(\alpha_i)} } = \sum_{i=1}^n{\beta_i \frac{f(x)}{x- \alpha_i}}.$$ Therefore $\sum_{i=1}^n{\beta_i \frac{f(x)}{x- \alpha_i}}$ will be the remainder of $f'(x) h(x)$ divided by $f(x)$, because they are of the same degree and coincide on the $n$ roots of $f(x)$. The lemma then follows from the fact that $\alpha$ is a root of $f(x)$. \end{proof} \par In order to bound $f'(\alpha)h(x)$ we will have to bound both $\frac{f(x)}{x-\alpha_i}$ and also $|\beta_i|$. We will use Corollary~\ref{grancor} to bound $\frac{f(x)}{x-\alpha_i}$ and since we know the $\beta_i$ up to a permutation (they are roots of $g(x)$), we can bound $\sum{|\beta_i|}$. \begin{thm} \label{granthm} If $f(x)$ and $\tilde{f}(x)$ are polynomials with complex coefficients, of degree $n$ and $d$ respectively, such that $\tilde{f}(x)$ divides $f(x)$ and $|f(0)|=|\tilde{f}(0)| \neq 0$, then \begin{equation} \label{gran} \|\tilde{f}(x)\| \leq \left( \sum_{j=0}^{n-d} { { \binom {d}{j} } ^2}\right)^{1/2} \|f(x)\|. \end{equation} \end{thm} \begin{proof} See Granville, [1]. \end{proof} \begin{cor} \label{grancor2} If $f(x)$ and $\tilde{f}(x)$ have the same leading coefficient and $f(0),\tilde{f}(0) \neq 0$ and $\tilde{f}(x)$ divides $f(x)$ then equation (\ref{gran}) holds. \end{cor} \begin{proof} Apply Theorem~\ref{granthm} to the reciprocals of $f$ and $\tilde{f}$. \end{proof} \begin{cor} \label{grancor} Let $P(x)$ be an irreducible polynomial (over $\mathbb{Q}$) of degree $n \geq 1$ and let $\{\gamma_1, \gamma_2, \dots, \gamma_n\}$ be its complex roots. Then $$\left\|{\frac{P(x)}{x-\gamma_i}}\right\| \leq n \|P(x)\|$$ \end{cor} \begin{proof} Take $\tilde{f}(x)= {\frac{P(x)}{x-\gamma_i}}$ and $f(x)=P(x)$ and apply Corollary~\ref{grancor2}. Then $$\left\|{\frac{P(x)}{x-\gamma_i}}\right\| \leq \left( \sum_{j=0}^{n-(n-1)} { { {n-1} \choose j}^2}\right)^{1/2} \|P(x)\| = \left( \sum_{j=0}^{1} { { {n-1} \choose j} ^2} \right)^{1/2} \|P(x)\| $$ $$= \left( { {n-1} \choose 0}^2 + { {n-1} \choose 1}^2 \right) ^{1/2} \|P(x)\| = (n^2 -2n+2)^{1/2} \|P(x)\| \leq n \|P(x)\|.$$ \end{proof} \begin{thm} \label{bound} Let $$S_{g(x)}:=\sum_{\stackrel{g(\beta)=0}{\beta \in \mathbb{C}}}{|\beta_i|}$$ then: \begin{equation} g_n \|\vec{h} \| \leq g_n n \left(S_{g(x)} \right) \|f(x)\| .\end{equation} There are several ways to bound $S_{g(x)}$:\\ 1) $S_{g(x)} \leq$ The degree of $g(x)$ times the rootbound described in [3].\\ 2) $S_{g(x)} \leq M(g)/lc(g) + (n-1)$, where the Mahler measure can be bounded by $\|g(x)\|$. \end{thm} \begin{proof} (of Equation (4)) $$ \| \vec{h} \| = \| P \| = \| \sum_{i=1}^n{\beta_i \frac{f(x)}{x- \alpha_i}} \| \leq n \|f(x)\| \sum_{i=1}^n{|\beta_i}| = n \|f(x)\| S_{g(x)} .$$ The first equality is by the definition of $\vec{h}$, the second by Lemma~\ref{inter} and the inequality by Corollary~\ref{grancor}. \end{proof} \section{The Algorithms} Here we give the algorithms for computing the isomorphisms between number fields. The {\em Pre-processing} algorithm reduces the lattice of possible isomorphisms and gives the explicit isomorphism if there is only one. The next algorithm, {\em FindIsomorphism}, calls the {\em Pre-processing} algorithm and uses the remaining lattice to check which maps on roots corresponds to an isomorphism. {\bf Algorithm: LLL-with-removals}[10] \\ \itshape {\bf Input} A matrix $A$ and a bound $b$. \\ {\bf Output} A set of LLL reduced row vectors where the last vector is removed if its Gram-Schmitt length is greater than $b$.\\ {\bf Algorithm: FindSuitablePrime} \\ \itshape {\bf Input} $(f(x), g(x), x, {\rm bp}, b, e)$, where {\rm bp} is the first prime to test, $b$ and $e$ determine the level to Hensel Lift to. \\ {\bf Output} $p, p^a, m, [ [F_{d_1},G_{d_1}], [F_{d_2},G_{d_2}], \dots [F_{d_m},G_{d_m}]]$, see equation ({\rm 2}) for notation.\\ {\bf Procedure} \\ \begin{enumerate} \item $p:= {\rm bp}$, counter$:=0$. \item Repeat (until the algorithm stops in Steps 2(d)ii, 2(f) or 2(j)). \begin{enumerate} \item $p:= \text{nextprime}(p)$ \item if $p |$ discriminant($f,x$) or $p | f_n$ then go to Step 2(a) \item if $p |$ discriminant($g,x$) or $p | g_n$ then go to Step 2(a) \item Distinct Degree Factor $f$ as $f \equiv F_{d_1} F_{d_2} \dots F_{d_m} \mod p$. \begin{enumerate} \item If $m=1$ then counter := counter $+1$. \item If counter $>25$ then print ``{\rm They appear to be Galois}'' and return 0,0,0,0. \item Return to Step 2(a). \end{enumerate} \item Distinct Degree Factor $g$ as $g \equiv G_{d'_1} G_{d'_2} \dots G_{d'_{m'}} \mod p$. \item If $m \neq m'$ or if the degrees of $F_i$ and $G_i$ do not match then return ``{\rm There is no isomorphism}''. \item Let $a := \left\lceil b^{e/10} 2^{e/4} \right\rceil$. \item Hensel lift $f \equiv F_{d_1} F_{d_2} \dots F_{d_m} \mod p^a$ and likewise for $g$. \item If deg$(F_1)>0$ then store $p$ for later use. \item Return $p, p^a, m, [ [F_{d_1},G_{d_1}], [F_{d_2},G_{d_2}], \dots [F_{d_m},G_{d_m}]]$ as output and stop. \end{enumerate} \end{enumerate} {\bf Algorithm: Pre-Processing} \\ \itshape {\bf Input} Polynomials $f(x)$ and $g(x)$. \\ {\bf Output} Either ``No isomorphism exists", a verified isomorphism, or a $\mathbb{Z}$-module which contains $(g_n \vec{h}, g_n)$ for every isomorphism $h$.\\ {\bf Remark}: This lattice is given as the row space of a matrix $C$.\\ {\bf Procedure} \\ \begin{enumerate} \item \emph{ Initialize} \begin{enumerate} \item $e:=n+1$. \item $C :=$ $(n+1)$ {\rm x} $(n+1)$ identity matrix. \item $p:=3$. \item $q:=0$. \item Let $\{ {\rm Base}_i\} \in \mathbb{Q}(\alpha)_{<n}, i=1\dots n$ be $ \{ \rho(1, 0, \dots, 0), \rho(0, 1, \dots, 0), \dots, \rho(0, 0, \dots, 1) \}$ with $\rho$ defined in Section 2. \end{enumerate} \item Let $S$ be an upper bound for $\sum_{\stackrel{g(\beta)=0}{\beta \in \mathbb{C}}}{|\beta_i|}$, e.g. (\ref{bound}.1) or (\ref{bound}.2). Our implementation uses (\ref{bound}.1). \item Let $b := n S \|f(x)\|$, be the bound described in Theorem~\ref{bound}. \item Repeat (until the algorithm stops in 4(b), 4(e) or 4(i)). \begin{enumerate} \item $q:= q+1$. \item $p, p^a, m, M_q :=$ {\rm FindSuitablePrime}$(f, g, x, p, b, e)$. \begin{enumerate} \item If $p=0$ then return $C$. \end{enumerate} \item Find $Tr_p^d(f,{\rm Base}_i)$ for $i =1 \dots n$ and $Tr_p^d(g,\beta)$ for each $d$ with deg($F_d$)$ >0$. The necessary $F_d,G_d$ are read from $M_q$. \item $A:=\left[ \begin{array}{cc} C & CT \\ 0 & P \end{array} \right] $, where $$P:= \left[ \begin{array}{ccc} p^a&& \\ &\ddots& \\ &&p^a \end{array} \right],$$ $$T:= \left[\begin{array}{ccc} Tr^{d_1}_{p_1}(f, {\rm Base}_1)&\dots&Tr^{d_m}_{p_1}(f, {\rm Base}_1) \\ Tr^{d_1}_{p_1}(f, {\rm Base}_2)&\dots&Tr^{d_m}_{p_1}(f, {\rm Base}_2)\\ \vdots&\dots&\vdots \\ Tr^{d_1}_{p_1}(f, {\rm Base}_n)&\dots&Tr^{d_m}_{p_1}(f, {\rm Base}_n)\\ Tr^{d_1}_{p_1}(g,\beta)&\dots& Tr^{d_m}_{p_1}(g,\beta) \end{array} \right]$$ the $d_1,\dots,d_m$ are as in Step 2(h) in Algorithm FindSuitablePrime. (Omitted entries are zero.) \item If $CT \equiv 0 \mod p^a$ then \begin{enumerate} \item counter $:=$ counter $+1$. \item If counter $< 10$ then {\rm Go to Step 4(a)} else return $C$ and stop. \end{enumerate} \item $L:=$ LLL-with-removals($A$, $b$). \item Let $C$ be the matrix with the first $n+1$ columns of $L$ and $B$ the remaining $m$ columns of $L$, so $L=[\begin{array}{cc} C&B \end{array}]$. \item if $B \neq 0$ then \begin{enumerate} \item $ B:= 10^{20} \cdot B$ \item $A := [\begin{array}{cc} C&B \end{array}]$ \item $L := \text{LLL-with-removals}(A,b)$, then go to Step 4(g). \end{enumerate} \item Let $e := $ number of rows of $C$. \begin{enumerate} \item if $e=0$ then output ``There is no isomorphism." \item if $e=1$ then let $C$ be $[V,v]$ with $V$ an $n$ dimensional vector, and let $h$ be the polynomial corresponding to $V/v$. \begin{enumerate} \item Let {\rm iso}:= $\frac{h(\alpha) g_n}{f'(\alpha)}$. \item If {\rm iso} satisfies $g$ then output ``{\rm iso} is the only isomorphism." \item If not then output ``There is no isomorphism." \end{enumerate} \item Else, go to Step 4a. \end{enumerate} \end{enumerate} \end{enumerate} \normalfont \begin{rmk} If we let $d$ be the number of isomorphisms ($\mathbb{Q}(\beta) \rightarrow \mathbb{Q}(\alpha)$) then just by looking at the input/output of the Pre-processing algorithm we see that: $$ \text{If } d \in \{0,1\} \text{ then the output is either } \left\{ \begin{array}{c} \text{all isomorphisms} \\ \text{a lattice} \end{array} \right. $$ $$ \text{If } d > 1 \text{ then the output is a lattice} $$ \end{rmk} \par In the next algorithm we use the lattice outputted from {\em Pre-Processing} to check all possible maps on the roots to see which are actual isomorphisms. This will find all isomorphisms from $\mathbb{Q}(\beta) \rightarrow \mathbb{Q}(\alpha)$. \par The following algorithm is described for (linear) roots of $f$ and $g$ in $\mathbb{Q}_p$ and can be extended to the roots of $F_i$ and $G_i$ instead. \begin{rmk} \label{useful} It should be noted that even if the Pre-processing Algorithm does not find the isomorphism(s), the LLL switches it performs will still contribute to the FindIsomorphism Algorithm. This is true for the same reason as in [11, pg 175]. \end{rmk} {\bf Algorithm: FindIsomorphism} \\ \itshape {\bf Input} Two polynomials, $f,g \in \mathbb{Z}[x]$ which are irreducible and of the same degree. \\ {\bf Output} The set of all isomorphisms from $\mathbb{Q}[x]/(f)$ to $\mathbb{Q}[x]/(g)$. \\ {\bf Procedure} \\ \begin{enumerate} \item $C$ := {\em Pre-Processing}($f(x)$ , $g(x)$, $x$). \item If Step 2(i) in Algorithm {\em FindSuitablePrime} (called from {\em Pre-Processing}) stored at least one prime, then choose one with smallest $\text{deg}(F_1)$. Otherwise keep calling Algorithm {\em FindSuitablePrime} until such a prime is found. \item Let $\alpha_1, \dots, \alpha_d$ be the roots of $F_1$ and Hensel lift them to $\mathbb{Z}/(p^a)$ with $a$ as in Algorithm {\em FindSuitablePrime}. Likewise let $\beta_1, \dots, \beta_d \in \mathbb{Z}/(p^a)$ be the roots of $G_1$. \item For $j$ from 1 to $d$ do: \begin{enumerate} \item Apply steps 4(d) through 4(i)ii of {\em Pre-Processing} using $$T:= \left[\begin{array}{c} {\rm Base}_1|_{\alpha = \alpha_j} \\ {\rm Base}_2|_{\alpha = \alpha_j}\\ \vdots \\ {\rm Base}_n|_{\alpha = \alpha_j}\\ \beta_1 \end{array} \right]$$ \item If $e > 1$ then \begin{enumerate} \item Hensel Lift the roots of $f$ and $g$ to twice the current $p$-adic precision, i.e. $p^{2a}$. \item Apply Step 4(a) with the more precise roots. \end{enumerate} \end{enumerate} \end{enumerate} \normalfont \subsection{Proofs of Termination and Validity} In this section we prove that the algorithms terminate and show that the algorithm does indeed produce all isomorphisms of the number fields $\mathbb{Q}(\beta)$ and $\mathbb{Q}(\alpha)$. \par First we cite a lemma which shows why we can use LLL with removal in our algorithm. \begin{lemma} \label{LLL} Let $\{ b_1, \dots, b_k \}$ be a basis for a lattice, $C$, and $\{ b_1^*, \dots , b_k^* \}$ the corresponding Gram-Schmitt orthogonalized basis for $C$. If $ \| b_k^* \| > B$ then a vector in $C$ with norm less than $B$ will be a $\mathbb{Z}$-linear combination of $\{ b_1, \dots, b_{k-1} \}$. \end{lemma} \begin{proof} This follows from the proof of Proposition 1.11 in [9], it is also stated as Lemma 2 in [10]. \end{proof} \begin{cor} Using LLL-with-removals on a lattice containing $g_n\vec{h}$ with the bound $b$, computed in Step 3 of {\em Pre-Processing}, does not remove $g_n\vec{h}$ from the lattice. \end{cor} \begin{proof} Using Lemma~\ref{LLL} and Theorem~\ref{bound} we know that removing final vectors with Gram-Schmitt length bigger than $b$ does not remove any of the $g_n \vec{h}$. \end{proof} \begin{lemma} The {\em Pre-Processing} Algorithm terminates. \end{lemma} \begin{proof} The steps of the {\em Pre-Processing} algorithm are known algorithms that terminate, the only one that is not immediate is Step 4(h). Step 4(h) terminates because each run increases the determinant of the lattice (Step 4(h)i) and any final (see Lemma~\ref{LLL}) vector with Gram-Schmitt length bigger than $b$ is removed, thus the number of vectors is monotonically decreasing and hence it can only be run a finite number of times. \end{proof} \begin{lemma} The {\em FindIsomorphism} Algorithm described above terminates. \end{lemma} \begin{proof} \par For Steps 1-3 it is clear why each will terminate. We show that Step 4 terminates by contradiction. \\ Suppose Step 4 never terminates (i.e. the lattice always has dimension $>$ 1) then it contains at least two vectors: $(h_1, e_1)$ and $(h_2,e_2)$. Let $H=h_1$ if $e_1=0$ or $H=e_1 h_2- e_2 h_1$ otherwise. Then $H(\alpha) \equiv 0 \mod p^a$. We get a contradiction when $p^a$ is larger than an upper bound for ${\rm Res}_x(H,f)$. An upper bound for $H$ can be obtained from equation $1.7$ in [9] and the fact that the last vector after LLL-with-removals has Gram-Schmitt length $\leq b$. \end{proof} \section{Heuristic estimate on the rank of $C$} Let $C \subseteq \mathbb{Z}^{n+1}$ be the output of the {\rm Pre-Processing} Algorithm. \begin{observation} In most (but not all) examples, dim($C$) is equal to $n- n/d + 1$. \end{observation} This means that {\rm Pre-Processing} is most effective when $d=1$. Though as pointed out in Remark~\ref{useful} the work done in {\rm Pre-Processing} reduces the amount left to do. \par Let $G$ be the Galois group of $f(x)$ and let $H_i$ be the stabilizer of $\alpha_i$ for $i\in \{1,2,\dots,n\}$, where the $\alpha_i$ are the roots of $f(x)$. \par Let $d$ be the number of $j$ such that $H_1 = H_j$, then $d$ is the number of automorphisms of $\mathbb{Q}(\alpha)$. If $\mathbb{Q}(\alpha)$ and $\mathbb{Q}(\beta)$ are isomorphic then $d$ will also be the number of isomorphisms from $\mathbb{Q}(\beta)$ to $\mathbb{Q}(\alpha)$. \begin{rmk} We view $G$, which as the Galois group acts on $\{\alpha_1,\alpha_2,\dots,\alpha_n\}$, as acting on the set $\{1,2,\dots,n\}$ in the most natural way. Hence we view $G$ as a subgroup of $S_n$, the symmetric group. \end{rmk} We will construct a partition matrix as follows. For each $\sigma \in G$, group together the cycles of the same length. Different group elements and cycle lengths will correspond to different rows. For each element of $G$ and for each cycle length in $\sigma$, construct one row of $P$ as follows: place a $1$ in the $i^{th}$ entry if $\alpha_i$ is in a cycle of that length. We call the resulting matrix $P$. \par For example for $\sigma_1 = (1)(2)(3)(456)$ and $\sigma_2 = (12)(3456)$ we would get the following partition matrix : $$ P=\begin{array}{c} \sigma_1 \text{ } l=1\\ \sigma_1 \text{ } l=3\\ \sigma_2 \text{ } l=2\\ \sigma_2 \text{ } l=4\\ \vdots \end{array}\begin{bmatrix} 1&1&1&0&0&0 \\ 0&0&0&1&1&1 \\ 1&1&0&0&0&0 \\ 0&0&1&1&1&1 \\ \vdots&\vdots&\vdots&\vdots&\vdots&\vdots \end{bmatrix}$$ \par Since there are $d$ automorphisms the number of distinct columns of $P$ will be $\leq n/d$, hence $\text{rank}(P) \leq n/d$ and thus $\text{Nullspace}(P) \geq n-n/d.$ \par This translates into an estimate on the rank of the lattice $C$ since it helps us bound $$V = \bigcap_{p,d} {Ker(Tr_p^d(f,*))}.$$ \par Nullspace($P$) corresponds to elements for which all sub-traces are zero, so dim(Nullspace($P$)) $\leq$ dim($V$). \par Since we used LLL-with-removals with cut off point $b$, if $V$ admits a basis whose norms are all smaller than $b$ then $V \subseteq \pi_{1\dots n}(C)$, where $\pi_{1\dots n}$ is the projection on the first $n$ coordinates. Therefore under that assumption $$\text{dim}(\pi_{1\dots n}(C) ) \geq \text{dim(Nullspace($P$))} \geq n- n/d.$$ This leads to our estimate: \begin{equation} \label{rank} \text{dim}(C) \approx n-n/d +1. \end{equation} \par For most polynomials taken from the database [5] our estimate is an equality. Peter Muller provided an infinite sequence of counter-examples for the case we were most interested in ($d=1$). For the first group in this sequence, the database [5] provides the following example: \\ $f := x^{14}+2x^{13}-5x^{12}-184x^{11}-314x^{10}+474x^9+1760x^8+1504x^7-400x^6-1478x^5-818x^4+73x^3+260x^2+121x+23,$ \\ which has one automorphism but the Pre-processing algorithm outputs a dimension $2$ lattice. \section{Computational Efficiency} We compare our algorithm implemented in Maple with other methods of finding isomorphisms. The best algorithm we know for factoring over number fields is give by Belabas in [6], which is implemented in Pari/Gp. Both implementations were run on a standard 2.2GHz processor. We tested them on the field extensions given by the following two degree 25 polynomials: \\ $f_1 := 2174026154062500000\,{x}^{25}-12927273797812500000\,{x}^{24}+ 44254465332187500000\,{x}^{23}-102418940816662500000\,{x}^{22}+ 180537842164766250000\,{x}^{21}-249634002590534050000\,{x}^{20}+ 292282923494920350000\,{x}^{19}-384197583430502150000\,{x}^{18}+ 815826517614521346000\,{x}^{17}-2131245874043847615600\,{x}^{16}+ 4352260622811059705104\,{x}^{15}-6463590834754261173232\,{x}^{14}+ 6920777688226436002712\,{x}^{13}-4525061881234027826296\,{x}^{12}+ 528408698276686662696\,{x}^{11}+2762117617850418790424\,{x}^{10}- 4343360968383689825174\,{x}^{9}+4191186502263628451150\,{x}^{8}- 2802452375464033976482\,{x}^{7}+1332292171242725153638\,{x}^{6}- 161285249796825311495\,{x}^{5}-429207332210687640181\,{x}^{4}+ 264147194777000152867\,{x}^{3}+6032198632961699729\,{x}^{2}- 42885793067858008650\,x+13774402803823804220$ and \\ $f_2 := -42885793067858008650\,x-13774402803823804220-161285249796825311495\,{ x}^{5}+429207332210687640181\,{x}^{4}+264147194777000152867\,{x}^{3}- 6032198632961699729\,{x}^{2}-1332292171242725153638\,{x}^{6}- 2802452375464033976482\,{x}^{7}-4191186502263628451150\,{x}^{8}- 4343360968383689825174\,{x}^{9}-2762117617850418790424\,{x}^{10}+ 528408698276686662696\,{x}^{11}+4525061881234027826296\,{x}^{12}+ 6920777688226436002712\,{x}^{13}+6463590834754261173232\,{x}^{14}+ 4352260622811059705104\,{x}^{15}+815826517614521346000\,{x}^{17}+ 384197583430502150000\,{x}^{18}+292282923494920350000\,{x}^{19}+ 249634002590534050000\,{x}^{20}+180537842164766250000\,{x}^{21}+ 102418940816662500000\,{x}^{22}+44254465332187500000\,{x}^{23}+ 2131245874043847615600\,{x}^{16}+12927273797812500000\,{x}^{24}+ 2174026154062500000\,{x}^{25}.$ \par These are field extensions with one isomorphism between them. Using Belabas' method we have a runtime of 11.69 seconds, which includes the indispensable operation of defining the number field, and with our algorithm we have a runtime of 2.97 seconds. \par We also tested this algorithm in the case where our heuristic estimate on the rank does not apply, namely on the first counter-example. Using Belabas' method we have a runtime of .091 seconds and with our algorithm we have a runtime of .797 seconds. \par Bearing in mind that our implementation is not optimized and is coded in Maple we expect it to be much faster when using a better implementation of LLL. \par We also tested them on a larger example, namely the degree 81 example located at [7], our algorithm found the isomorphism in 2326.581 seconds and the Belabas' implementation did not finish as it ran out of memory after trying for a few days. Therefore there are certainly advantages to using this algorithm, as there are examples where there is a significant reduction in the required runtime/resources. \section{Summary} Method II (from Section 1) can be described by the following procedure: first pick $p$ such that $f$ and $g$ have roots in $\mathbb{Q}_p$. Fix one root $\beta \in \mathbb{Q}_p$ of $g$, take all roots $\alpha_1, \dots, \alpha_d \in \mathbb{Q}_p$ of $f$. Then for each $\alpha_i$ use LLL to find $h_i \in \mathbb{Q}[x]$ (if it exists) with $h_i(\alpha_i) = \beta_i$. \par Our approach is similar, the difference is that we start with LLL reductions (obtained from sub-traces) that are valid for all $\alpha_i$. This way, a portion of the LLL computation to be done for each $\alpha_i$ is now shared. The time saved is then $(d-1)$ times the cost of the shared portion. This can be made rigorous by introducing a progress counter for LLL cost similar to [10]. \section{References} \begin{enumerate} \item Granville, A. ``Bounding the coefficients of a divisor of a given polynomial", Monatsh. Math. 109 (1990), 271-277. \item Conrad, Kieth. ``The different ideal". Expository papers/Lecture notes. Available at: \\ http://www.math.uconn.edu/$\sim$kconrad/blurbs/gradnumthy/different.pdf \item Monagan, M. B. ``A Heuristic Irreducibility Test for Univariate Polynomials", J. of Symbolic Comp., 13, No. 1, Academic Press (1992) 47-57. \item Dahan, X. and Schost, $\acute{E}$. 2004. ``Sharp estimates for triangular sets". In Proceedings of the 2004 international Symposium on Symbolic and Algebraic Computation (Santander, Spain, July 04 - 07, 2004). ISSAC '04. ACM, New York, NY, 103-110. \item Database by J$\ddot{u}$rgen Kl$\ddot{u}$ners and Gunter Malle , located at: \\ http://www.math.uni-duesseldorf.de/$\sim$klueners/minimum/minimum.html \item Belabas, Karim. ``A relative van Hoeij algorithm over number fields". J. Symbolic Computation, Vol. 37 (2004), no. 5, pp. 641-668. \item Website with implementations and Degree 81 examples:\\ http://www.math.fsu.edu/\ $\sim$vpal/Iso/ \item van Hoeij, Mark. ``Factoring Polynomials and the Knapsack Problem." J. Number Th. 95, 167-189, 2002 \item Lenstra, A. K.; Lenstra, H. W., Jr.; Lov\'{a}sz, L. ``Factoring polynomials with rational coefficients". Mathematische Annalen 261 (4), 515-534, 1982. \item M. van Hoeij and A. Novocin, `` Gradual sub-lattice reduction and a new complexity for factoring polynomials", accepted for proceedings of LATIN 2010. \item Cohen, Henri A Course in Computational Algebraic Number Theory, Graduate Texts in Mathematics 138, Springer-Verlag, 1993. \end{enumerate} Florida State University 211 Love Building, Tallahassee, Fl 32306-3027, USA\\ {\em E-mail address}: hoeij@math.fsu.edu\\ {\em E-mail address}: vpal@math.fsu.edu\\ \end{document}

https://arxiv.org/abs/2206.11700

Universal Neyman-Pearson Classification with a Known Hypothesis

We propose a universal classifier for binary Neyman-Pearson classification where null distribution is known while only a training sequence is available for the alternative distribution. The proposed classifier interpolates between Hoeffding's classifier and the likelihood ratio test and attains the same error probability prefactor as the likelihood ratio test, i.e., the same prefactor as if both distributions were known. In addition, like Hoeffding's universal hypothesis test, the proposed classifier is shown to attain the optimal error exponent tradeoff attained by the likelihood ratio test whenever the ratio of training to observation samples exceeds a certain value. We propose a lower bound and an upper bound to the training to observation ratio. In addition, we propose a sequential classifier that attains the optimal error exponent tradeoff.

\section{Introduction} The problem of deciding between multiple statistical descriptions of a given observation is one of the main problems in statistics and finds applications in a number of fields including engineering, signal processing, medicine and finance among others (see e.g. \cite{Lehmann}). Depending of whether or not the possible probability distributions of the observation are known, the problem is termed hypothesis testing or classification. In the case where there are only two possible distributions of the observation one typically refers to these decision problems as binary. When priors on the distributions are available, the problem is cast as Bayesian and the average probability of error determines the quality of the detection. In the absence of priors, the design of tests and classifiers minimizes one pairwise error probability while keeping the other upper bounded by a constant. This setting, proposed by Neyman and Pearson \cite{Neyman} has been adopted as the de-facto test design setting since it allows for both pairwise error probabilities being bounded, unlike average risk minimization. Upon observing a vector of $n$ observations, the hypothesis test that minimizes the above pairwise error probability tradeoff when the two possible distributions are known is given by the likelihood ratio test \cite{Neyman}. Instead, when decisions need to be made each time a new observation arrives, the sequential probability ratio test provides the best exponential decay of the error probability (or error exponent) \cite{Wald, Tanseq}. When one of the hypothesis distributions belongs to some class of distributions and only the class is known, Hoeffding proposed the generalized likelihood ratio test that attains the optimal error exponent \cite{Hoeffding}. When the distributions are not known, several classifiers such as logistic regression, support vector machines, naive Bayes, have been proposed in the literature \cite{VC, Shirani, Gaborpattern, CoverNN}. However, none of these classifiers ensures a guarantee on the type-\RNum{1} error probability resulting in the possibility of a large type-\RNum{1} errors. In the Neyman-Pearson setting, \cite{Tong} gives uniform bounds on type-\RNum{1} error probability for the plugin likelihood ratio test when both distributions are unknown, and only training sequences from both distributions are available. Gutman \cite{gutman1989asymptotically} proposed universal classifiers that guarantee a certain type-\RNum{1} error exponent under any probability distribution pair while achieving the lowest type-\RNum{2} error probability. Tan \cite{Tan} proved the second order optimality of Gutman's classifier. This paper considers the setting when the null hypothesis generating distribution is known while only a training sample is available for the alternative hypothesis. This can be the case in the unbalanced training sample size when a large number of training samples is available for the null hypothesis so that with high accuracy, the distribution can be estimated. In contrast, the alternative distribution cannot be accurately estimated due to small number of training samples. The scenario when the number of test and training samples is fixed can also be viewed as composite hypothesis testing with known null hypothesis and an additional training sequence. Hoeffding's test can achieve the optimal error exponent when the null hypothesis distribution is given. However, the prefactor attained by Hoeffding's test is only optimal for distributions defined over binary alphabets. In this paper, we propose a classifier that achieves the optimal prefactor using the additional training samples while attaining the optimal error exponent tradeoff whenever the training to observation sample ratio exceeds a certain value. The proposed classifier interpolates between the plugin likelihood ratio and Hoeffding's tests. We also study the case when the test and training samples are generated sequentially. There is no classifier similar to Hoeffding's in this scenario that can achieve the error exponent gain by taking test samples sequentially when only the null hypothesis is known. However, our proposed classifier can improve the achievable error exponent by using the additional training samples from the unknown distribution. This paper is structured as follows. Section \ref{sec:pre} introduces notation and reviews the preliminaries about the likelihood ratio test and Hoeffding's generalized likelihood ratio test. Section \ref{sec:fixedclas}, proposes the new classifier for a fixed sample size setting and shows the finite length improvements over Hoeffding's test. Section \ref{sec:steinclas} discusses the classification problem in the Stein regime and shows the optimality of Gutman's test in this regime. Section \ref{sec:seqclas}, proposes a sequential classifier achieving the highest error exponent tradeoff in the proposed setting. Proofs of the main results can be found in the Appendices. \section{Prelimenaries} \label{sec:pre} Consider the following binary classification problem where an observation $\xv=(x_1,\dotsc,x_n)$ is generated in an i.i.d. fashion from either of two possible distributions $P_0$ or $P_1$ defined on the probability simplex $\Pc(\Xc)$ with alphabet size $|\Xc|<\infty$. We assume that the distribution $P_0$ is known while only a sequence of training samples $\X=(X_1,\dotsc,X_k)\sim P_1^k$ generated in an i.i.d. fashion from $P_1$ is available; training and test sequences are sampled independently from each other. We also assume that both $P_0(x)>0, P_1(x)>0$ and $\frac{P_0(x)}{P_1(x)} \leq c$ for each $x\in\Xc$ for some positive $c$. Also we let $k$, the length of the training, be such that $k=\alpha n$ for some positive $\alpha$. The type of an $n$-length sequence $\yv$ is defined as $\hat{T}_{\yv}(a)=\frac{N(a|\yv)}{n}$, where $N(a|\yv)$ is the number of occurrences of symbol $a\in\Xc$ in sequence $\yv$. The types of the observation and training sequences $\xv, \X$ are denoted by $\Tx, \TX $ respectively. The set of all sequences of length $n$ with type $P$, denoted by $\mathcal{T}_{P}^n$, is called the type class. The set of types formed with length $n$ sequences on the simplex $\Pc(\Xc)$ is denoted as $\Pc_n(\Xc)$. Let $\phi(\X,\xv): \Xc^k \times \Xc^n \rightarrow \{0,1\}$ be a classifier that decides the distribution that generated the observation $\xv$ upon processing the training sequence $\X$. We consider deterministic classifiers $\phi$ that decide in favor of $P_0$ if $\xv\in \Ac_0(P_0,\X)$, where $ \Ac_0(P_0,\X)\subset \Xc^n$ is the decision region for the first hypothesis and is a function of $P_0$ and the training samples $\X$. We define $\Ac_1(P_0,\X)=\Xc^n \setminus \Ac_0$ to be the decision region for the second hypothesis. If we assume no prior knowledge on either distribution, the two possible pairwise error probabilities determine the performance of the classifier. Specifically, the type-\RNum{1} and type-\RNum{2} error probabilities are defined as \begin{align}\label{eq:e1} \epsilon_0 (\phi)& = \sum_{\X \in \Xc^k} P_1(\X) \sum_{\xv \in \Ac_1(P_0,\X)} P_0(\bx^n),\\ \epsilon_1 (\phi)&= \sum_{\X \in \Xc^k} P_1(\X) \sum_{\xv \in \Ac_0(P_0,\X)} P_1(\bx^n). \end{align} In the case where both distributions are known, the training sequence is not needed and the classifier becomes a hypothesis test. In this case, the classifier is said to be optimal whenever it achieves the optimal error probability tradeoff given by \begin{equation}\label{eq:trade} \min_{\phi: \epsilon_0 (\phi) \leq \xi} \epsilon_1 (\phi), \end{equation} where $\xi \in [0,1]$. It is well known that likelihood ratio test defined as \begin{equation} \phi^{\rm lrt}(\xv)= \mathbbm{1} \bigg \{ \frac{P_1^n(\bx)}{P_0^n(\bx)} \geq e^{n\gamma} \bigg\}, \end{equation} attains the optimal tradeoff \eqref{eq:trade} for every $\gamma$ \cite{Neyman}. The likelihood ratio test can also be expressed as a function of the type of the observation $\Tx$ as \cite{Cover,Csiszar} \begin{align}\label{eq:LRTtype} \phi^{\rm lrt}(\Tx)= \mathbbm{1} \big\{ D(\Tx\|P_0)-D(\Tx\|P_1) \geq \gamma \big\}. \end{align} where $D(P\|Q)= \sum_{\Xc} P(x) \log \frac{P(x)}{Q(x)}$ is the relative entropy between distributions $P$ and $Q$. The optimal error exponent tradeoff $(E_0,E_1)$ is defined as \begin{align}\label{eq:tradefix} E_1^*(E_0) \triangleq \sup \big\{E_1\in \mathbb{R}_{+}: \exists \phi , \exists n_0 \in \mathbb{Z}_+ \ \text{s.t.} \ \forall n>n_0 , \epsilon_0(\phi) \leq e^{-nE_0} \quad \text{and} \quad \epsilon_1(\phi) \leq e^{-nE_1}\big\}. \end{align} By using Sanov's Theorem \cite{Cover,Dembo}, the optimal error exponent tradeoff $(E_1,E_0)$, attained by the likelihood ratio is given by \begin{align} E_0(\phi^{\rm lrt})=\min_{Q \in \mathcal{Q}_0(\gamma)} D(Q\|P_0)\label{eq:min1},\\ E_1(\phi^{\rm lrt})=\min_{Q \in \mathcal{Q}_1(\gamma)} D(Q\|P_1)\label{eq:min2}, \end{align} where \begin{align} \mathcal{Q}_0(\gamma)&= \big\{Q\in \mathcal{P}(\Xc): D(Q\| P_0)-D(Q\|P_1) \geq \gamma \big\},\\ \mathcal{Q}_1(\gamma)&= \big\{Q\in \mathcal{P}(\Xc): D(Q\| P_0)-D(Q\|P_1) \leq \gamma \big\}. \end{align} Furtheremore, the minimizing distribution in \eqref{eq:min1}, \eqref{eq:min2} is the tilted distribution \begin{equation}\label{eq:tilted} Q_{\lambda^*}(x)= \frac{ P_{0}^{\lambda^*}(x) P_{1}^{1-\lambda^*}(x) } {\sum_{a \in \Xc } P_{0}^{1-\lambda^*}(a) P_{1}^{\lambda^*}(a) }, ~~~0\leq\lambda^* \leq 1, \end{equation} where $\lambda^*$ is the solution of $D(Q_{\lambda^*}\| P_0)-D(Q_{\lambda^*} \| P_1) = \gamma$. The classification problem described above with known $P_0$ and a training sequence from $P_1$, can also be viewed as the composite binary hypothesis problem where additional training sequence samples are given for the second hypotheses. In the case of a composite hypothesis testing problem where $P_0$ is given, and the other hypothesis is unrestricted to $\Pc(\Xc)$, Hoeffding proposed in \cite{Hoeffding} a generalized likelihood-ratio test given by \begin{align} \phi^{\rm glrt}(\xv)= \mathbbm{1} \big\{ D(\Tx\|P_0) > E_0 \big\}, \end{align} which attains the optimal error exponent tradeoff in \eqref{eq:tradefix}. By Sanov's theorem, the error exponent of Hoeffding's test is given by \begin{align} E_0(\phi^{\rm glrt})&=E_0,\\ E_1(\phi^{\rm glrt})&=\min_{\substack{Q\in \mathcal{P}(\Xc), \\ D(Q\| P_0) \leq E_0 }} D(Q\|P_1),\label{eq:min2H} \end{align} where the minimizing distributions in \eqref{eq:min2H} is given by \begin{equation}\label{eq:tiltedH} Q_{\mu^*}(x)= \frac{ P_{0}^{\frac{\mu^*}{1+\mu^*}}(x) P_{1}^{\frac{1}{1+\mu^*}}(x) } {\sum_{a \in \Xc } P_{0}^{\frac{\mu^*}{1+\mu^*}}(a) P_{1}^{\frac{1}{1+\mu^*}}(a) }, ~~~\mu^*\geq0, \end{equation} and $\mu^*$ is the solution to $D(Q_{\mu^*}\|P_0)=E_0$. Using a large deviations refinement \cite{Iltis,Albert}, the type-\RNum{1} error probability of the likelihood ratio test is \begin{equation}\label{eq:LRTsaddle} \epsilon_0(\phi^{\rm lrt}) = \frac{1}{\sqrt{n}} e^{-nE_0}\big (c+ o(1)\big), \end{equation} while, Hoeffding's test type-\RNum{1} error probability is given by \cite{Veeravalli,Iltis} \begin{equation}\label{eq:Hoefasymp} \epsilon_0(\phi^{\rm glrt}) =n^{\frac{|\Xc|-3}{2}} e^{-nE_0}\big (c'+ o(1)\big) \end{equation} where $c,c'$ are constants that only depend on $P_0, P_1$ and the corresponding test thresholds. As a result, when the number of observations is large, Hoeffding's test, although attaining the optimal error exponent tradeoff, suffers in exponential prefactor when compared to the likelihood ratio's $\frac{1}{\sqrt{n}}$ for observation alphabets such that $|\Xc|>2$. \section{Fixed Sample Sized Universal Classifier}\label{sec:fixedclas} In this section, we propose a classifier that interpolates between the likelihood ratio and Hoeffding's tests that attains a prefactor that is independent of the alphabet size and is equal to $\frac{1}{\sqrt{n}}$. In addition, we show that if the ratio of training samples to the number of test samples $\alpha$ exceeds a certain threshold, the proposed test also achieves the optimal error exponent tradeoff. Hoeffding's test can favor the second hypothesis for test sequences with types close to $P_0$ while far from $P_1$. Suppose we have a training sequence type $\TX$, we can relax the Hoeffding's test from a ball centered at $P_0$ to a hyperplane tangent to the Hoffding's test ball, directed towards the type of the training sequence. As we will see, this is precisely what enables the improvement in the prefactor of the type-\RNum{1} probability of error. We propose the following classifier \begin{align}\label{eq:clsfix} \cls(\Tx,\TX)=\mathbbm{1} \big\{ \beta D(\Tx\|\TX') - D(\Tx\|P_0)\ \leq \gamma(E_0,\TX') \big\}, \end{align} where $0 \leq \beta \leq 1$ controls how much the training weights in the decision, the threshold $\gamma(E_0,Q_1)$ is given by \begin{align} \gamma(E_0,Q_1)&= \beta \min_{\substack{Q\in \mathcal{P}(\Xc), \\ D(Q\| P_0) \leq E_0 } } D(Q\|Q_1) -E_0 \label{eq:threshcls}, \end{align} the perturbed training type $\TX'(a)$ is \begin{align}\label{eq:typeplus} \TX'(a)&=\big(1-\delta_n\big)\TX(a)+\frac{\delta_n}{|\Xc|}, \end{align} where, $\delta_n$ can be chosen as any function of the order $o(n^{-1})$. We add this small perturbation of the training type to avoid the cases where some of the alphabet symbols have not been observed in the training sequence. We define the decision regions of the proposed classifier by \begin{align} \Az&=\{Q: Q\in \Pc(\Xc), \cls(Q,\TX)=0 \}, \\ \Ao&=\{Q: Q\in \Pc(\Xc), \cls(Q,\TX)=1 \}. \end{align} Observe that when $\beta=0$ we recover Hoeffding's test while for $\beta=1$ the test is reminiscent of a likelihood ratio test where instead of $P_1$, we have the perturbed training type $\TX'(a)$. Intuitively, as long as we have enough training samples, the training type $\TX'(a)$ will be close to $P_1$ and we will attain the optimal error exponent tradeoff. This is indeed what will be shown next. Figure \ref{fig:classification} illustrates the proposed classifier for a realization of the training sequence when $\beta=1$. The proposed classifier becomes the plug-in likelihood ratio test, where the test's threshold has been adjusted by the training samples such that the resulting hyperplane of the likelihood ratio test is tangent to the relative entropy ball of radius $E_0$ centered at $P_0$. The type-\RNum{1} error exponent will be equal to $E_0$ for any realization of $\TX$. However, the type-\RNum{2} error exponent for a given training sample $E_1(\TX')$ is the projection of $P_1$ into the separating hyperplane determined by the test, which is the function of the training sequence and can vary accordingly. \begin{figure}[!h] \centering \begin{tikzpicture}[scale=0.8] \draw [line width=0.3mm] (5.5,1.2) -- (0,-5) -- (11,-5) --(5.5,1.2) ; \draw [line width=0.3mm, dashed] (3.1,-1.5) -- (4.5,-5) ; \node at (7.4,-4.5) {\small $D(\Tx\|P_0)-D(\Tx\|\TX') = \gamma(E_0,\TX') $}; \node[draw,circle,inner sep=1pt,fill] at (3,-3.5) {}; \node[draw,circle,inner sep=1pt,fill] at (6.5,-2.6) {}; \node at (2.7,-3.4) {\small $ P_0$}; \node at (6.8,-2.95) {\small $\TX'$}; \draw (2.8,-3.5) ellipse (1.05cm and 0.7cm); \node at (1,-4.7) {$\mathcal{P}({\Xc})$}; \draw [<-,>=stealth] (6.5,-2.6) -- (6.,-1.85); \draw [->,>=stealth] (3,-3.5) -- (3.2,-4.15); \node at (2.9,-3.9) {\small $E_0$}; \node[draw,circle,inner sep=1pt,fill] at (6,-1.85) {}; \draw [line width=0.25mm, gray]plot [smooth, tension=1.5] coordinates{(6,-1.85) (5,-2.1) (3.55,-2.7)} ; \draw [line width=0.25mm, gray]plot [smooth, tension=1.5] coordinates{(3,-3.5) (4.1,-3.2) (5.5,-2.8) (6.5,-2.6)} ; \node at (4.6,-1.75) {\small $ E_1(\TX')$}; \node at (6.1,-1.65) {\small $P_1$}; \node[draw,circle,inner sep=1pt,fill] at (3.57,-2.7) {}; \node[draw,circle,inner sep=1pt,fill] at (3.8,-3.3) {}; \end{tikzpicture} \caption{Proposed classifier with known distribution $P_0$ and a training sequence with type $\TX$.} \label{fig:classification} \end{figure} Next, we find a refined expression for the type-\RNum{1} error probability and show that the error probability prefactor is $O(\frac{1}{\sqrt{n}})$, i.e., of the same order of the prefactor achieved by the likelihood ratio test. \begin{theorem}\label{thm:E0 For every $P_0, P_1, 0<\beta \leq 1$ the type-\RNum{1} probability of error of the classifier $\cls$ defined in \eqref{eq:clsfix} can be expressed as \begin{equation}\label{eq:type0classifier} \epsilon_0(\cls)= \frac{1}{\sqrt{n}} e^{-nE_0}(c+ o(1)), \end{equation} where $c$ is a positive constant that only depends on the data distributions and $E_0$. \end{theorem} \begin{proof} The proof can be found in Appendix \ref{sec:proofth1}. \end{proof} Having established the improvement of the prefactor of the type-$\RNum{1}$ error probability, we are ready to study the type-\RNum{2} error exponent of our proposed classifier. \begin{theorem}\label{thm:E1} For every $P_0, P_1$ there exists a finite training to sample size ratio $\alpha^*_\beta$ such that for any $\alpha>\alpha^*_\beta$ we have \begin{equation} E_1(\cls)=E_1^*(E_0), \end{equation} that is, for any $E_0$, the proposed classifier achieves the optimal error exponent tradeoff defined in \eqref{eq:tradefix}. \end{theorem} \begin{proof} The proof can be found in Appendix \ref{sec:proofth2}. \end{proof} We have shown that the classifier proposed in \eqref{eq:clsfix} not only achieves the optimal error exponent tradeoff for $\alpha>\alpha^*_\beta$ but also achieves the same prefactor of the type-\RNum{1} error probability of the likelihood ratio test. This represents a significant improvement with respect to the Hoffding's universal test for observation alphabets $|\Xc|>2$. However, this might come as an expense of worse polynomial decay of type-\RNum{2} error probability. In the following result, by using the refinements of large deviation techniques, we show that this is not the case and that the proposed classifier achieves the same prefactor as the type-\RNum{2} error probability of the likelihood ratio test, establishing the optimality of the proposed classifier up to a constant. \begin{theorem}\label{thm:refinement} For every $P_0, P_1$, and $\alpha^*_\beta < \alpha $ the type-\RNum{2} probability of error of the classifier $\cls$ is \begin{equation}\label{eq:type1classifier} \epsilon_1(\cls)= \frac{1}{\sqrt{n}} e^{-nE_1(E_0,P_1)}(c+ o(1)), \end{equation} where $c$ is a positive constant depending only on data distributions and $E_0$. \end{theorem} \begin{proof} The proof can be found in Appendix \ref{sec:proofth3}. \end{proof} \begin{example}\label{ex:ex1} We present a numerical example to illustrate the performance of the proposed classifier in \eqref{eq:clsfix}. Consider two trinary distributions $P_0=[0.3,0.3,0.4], P_1=[0.35,0.35,0.3]$ and set $\alpha= 2$ and $E_0=0.005$. Each point in the figure is obtained by estimating the average error probability as follows. For each length of the test sequence $n$, we estimate the type-\RNum{1} and type-\RNum{2} error probabilities of the classifier in \eqref{eq:clsfix} as well as those of the likelihood ratio test in \eqref{eq:LRTtype} and Hoeffding's classifier over $5*10^7$ independent experiments. As it can be seen in Figures \ref{fig:e0}, the type-\RNum{1} error exponent of the likelihood ratio test and the proposed classifier are very close to each other and outperform the Hoeffding's test. In addition, we observe that the type-\RNum{2} error exponent of the proposed classifier is slightly worse than that of the likelihood ratio and Hoeffding's tests for small $n$; as $n$ increases, the proposed classifier achieves the optimal exponent. In addition, in order to clearly observe the effect of the prefactor, in Figures \ref{fig:e0asymp} we plot $\log\frac{\epsilon_i(\phi)}{\frac{1}{\sqrt n}e^{-nE_i}}$ for $i\in \{0,1 \}$ for the same classifiers. We first notice that in the case of $\log\frac{\epsilon_1(\phi)}{\frac{1}{\sqrt n}e^{-nE_1}}$, the three classifiers achieve the optimal $\frac{1}{\sqrt{n}}$ prefactor, though with different constants. We also observe that $\log\frac{\epsilon_0(\phi)}{\frac{1}{\sqrt n}e^{-nE_0}}$ converges to a constant for the likelihood ratio test and the new classifier, as predicted by our analysis. Instead, as \eqref{eq:Hoefasymp} suggests, Hoeffding's classifier fails to attain the optimal $\frac{1}{\sqrt{n}}$ prefactor for $\epsilon_0$. \end{example} \begin{figure}[htpb] \centering \input{e0exp.tex} \input{e1exp.tex} \caption{Type-\RNum{1} and type-\RNum{2} error exponents for the likelihood ratio test, Hoeffding's test and the proposed classifier. } \label{fig:e0} \end{figure} \begin{figure}[htpb] \centering \input{e0asymp.tex} \input{e1asymp.tex} \caption{$\log\frac{\epsilon_i(\phi)}{{\frac{1}{\sqrt{n}}}e^{-nE_i}}$ for $i\in \{0,1 \}$ and for the likelihood ratio test, Hoeffding's test and the proposed classifier. } \label{fig:e0asymp} \end{figure} Theorem \ref{thm:E1} shows that the classifier in \eqref{eq:clsfix} can achieve the optimal error exponent tradeoff if the training to sample ratio $\alpha$ is large enough. As it is evident by the proof of Theorem \ref{thm:E1}, the proposed classifier cannot achieve the optimal error exponent tradeoff for $\alpha$ close to zero. Therefore, it is desirable to find the smallest $\alpha^*_\beta$ such that the above theorem holds. The next results introduce lower and upper bounds to $\alpha^*_\beta$. \begin{theorem}\label{thm:lower} For every $P_0, P_1 \in \Pc(\Xc)$, and $E_0 > 0$, we have $ \underline{\alpha}_\beta \leq \alpha^*_\beta $ where \begin{equation} \underline{\alpha}_\beta=- \Lambda_{\min}(\Hm) , \end{equation} and $ \Lambda_{\min}$ is the smallest eigenvalue of the matrix \begin{equation}\label{eq:Hess} \Hm=\beta \eta^*_\beta \sqrt{\Jm} \Big [ \Qm+ \eta^*_1 \Vm +(1-\eta^*_1)\Wm -\Tm \Big ] \sqrt{\Jm}, \end{equation} where $\Wm=\wv \wv^T, \Vm=\vv \vv^T, \Qm=\qv \qv^T $ and \begin{align} \wv^T= \frac{1}{{\sqrt{ \Var_{Q_{\mu^*}} \Big (\log \frac{Q_{\mu^*}}{P_0} \Big ) }}} \bigg ( Q_{\mu^*}(1) {\log\Big ( \frac{Q_{\mu^*}(1)}{P_0(1)} \Big )-E_0 }, \ldots, Q_{\mu^*}(|\Xc|){\log\Big ( \frac{Q_{\mu^*}(|\Xc|)}{P_0(|\Xc|)} \Big )-E_0 } \Bigg), \end{align} \begin{align} \vv=\frac{1}{\sqrt{\Var{Q_{\eta^*_\beta} }(\Omega) }} \big ( Q_{\eta^*_\beta}(1)\Omega(1),\ldots,Q_{\eta^*_\beta}(|\Xc|)\Omega(|\Xc|) \big ), \end{align} \begin{align} \Omega(i)=\beta \log \frac{P_1(i)}{P_0(i)} +(1-\beta) \log \frac{Q_{\eta^*_\beta}(i)}{P_0(i)},\ \ i \in \Xc, \end{align} \begin{align} \qv= \big (Q_{\mu^*} (1),\ldots, Q_{\mu^*}(|\Xc|)\big), \end{align} \begin{align} \Tm= \diag (Q_{\mu^*} ), \quad \Jm=\diag \bigg(\frac{1}{P_1} \bigg) \label{eq:fisher} \end{align} where $Q_{\mu^*}$ is defined in \eqref{eq:tiltedH} when $Q_1=P_1$, and ${Q}_{\eta_\beta^*}$ defined in \eqref{eq:tiltbeta}, is the projection of $P_1$ onto $\Ac_0(P_1,\beta)$ which equals to $Q_{\mu^*}$. Furthermore, $\eta^*_\beta$ is the optimal Lagrange multiplier in \eqref{eq:lagrangeE} for $0<\beta\leq 1$. Moreover, for $|\Xc| \geq 6$ by further lower bounding $\underline{\alpha}_\beta$ we have \begin{align}\label{eq:lower} \beta \eta^*_\beta \Bigg [\frac{Q_{\eta^*_\beta}}{P_1} \Bigg ] _{(3)} \leq \underline{\alpha}_\beta, \end{align} where $\Big [\frac{Q_{\eta^*_\beta}}{P_1} \Big ]_{(3)}$ is the third largest value of $\frac{Q_{\eta^*_\beta(i)}}{P_1(i)}$ for $i \in \{1,\ldots, |\Xc|\}$. \end{theorem} \begin{proof} The proof can be found in Appendix \ref{sec:proofth5}. \end{proof} \begin{example}\label{exp:lower} Letting the distributions $P_0=\text{Bern}(0.3), P_1=\text{Bern}(0.4)$, we have set $E_0=0.005$. Figure \ref{fig:alpha_beta} shows the relation of the $\underline{\alpha}_\beta$ for $0<\beta<1$. As can be seen from the figure, $\underline{\alpha}_\beta$ is increasing in $\beta$ and is equal to zero for $\beta=0$ as expected, since for $\beta=0$ the proposed classifier is equal to universal Hoeffding's test and achieves the optimal type-\RNum{2} error exponent for every $P_1$. However, $\beta=0$, as discussed before, is the singularity point as the type-\RNum{1} error probability prefactor looses its independence from dimension. \end{example} \begin{figure}[h] \centering \input{alpha_beta.tex} \caption{Lower bound on the required training to test sample ratio $\underline{\alpha}_\beta$ for achieving the optimal error exponent trade-off with the proposed classifier for $0< \beta\leq 1$. } \label{fig:alpha_beta} \end{figure} We next find an upper bound on the optimal training to observation ratio $\alpha^*_\beta$. \begin{theorem}\label{thm:upper} For every $P_0, P_1 \in \Pc(\Xc)$, and $E_0 > 0$, we have that $\alpha^*_\beta \leq \bar{\alpha} $ where \begin{equation} \bar{\alpha}=\frac{\lambda^*(4+\lambda^*)(1+\kappa) }{(P_1^{\text{min}})^2}, \end{equation} $P_1^{\text{min}}\triangleq \min_{x\in\Xc} P_1(x)$, $\lambda^*$ is the optimal Lagrange multiplier in \eqref{eq:tilted}, and \begin{equation}\label{eq:upper} \kappa =\sqrt{\frac{E_1(E_0,P_1)}{\lambda^*(4+\lambda^*)} }. \end{equation} \end{theorem} \begin{proof} The proof can be found in Appendix \ref{sec:proofth7}. \end{proof} Note that the upper bound \eqref{eq:upper} and lower bound \eqref{eq:lower} to $\alpha^*_\beta$ suggests that as $P_1$ approaches the probability simplex boundaries, i.e., as $P_1^\text{min}$ approaches zero, the classification problem becomes more challenging, and we need more training samples to achieve the optimal exponents. For the classification problem in the Example \ref{exp:lower}, the upper bound in \eqref{eq:upper} gives $\bar{\alpha}=14.19$, i.e., if the ratio of training samples to the number of test samples exceeds $14.19$ the proposed test with $\beta=1$ achieves \eqref{eq:type0classifier}, \eqref{eq:type1classifier}. See \cite{viswanath} for general conditions between alphabets size and training sample size when both hypotheses are unknown. \section{Stein Regime Classification}\label{sec:steinclas} In this section, we will study the classification problem in the Stein regime. For the case where both hypotheses are known, the Stein regime is defined as the highest error exponent under one hypothesis when the error probability under the other hypothesis is at most some fixed $\epsilon$ (see e.g., \cite{Cover}) \begin{align}\label{eq:steindef} E_1^{(\epsilon)} \triangleq \sup \big \{E_1\in \mathbb{R}_{+}: \exists \phi , \exists n_0 \in \ZZ_+ \ \text{s.t.} \ \forall n>n_0 ~ \epsilon_0 (\phi)\leq \epsilon \quad \text{and} \quad \epsilon_1(\phi) \leq e^{-nE_1} \big \}. \end{align} The the optimal $E_1^{(\epsilon)}$, given by \cite{Cover} \begin{equation}\label{eq:Stein2trade} E_1^{(\epsilon)} (\phi^{\rm lrt}) = D(P_0\|P_1), \end{equation} can be achieved by setting the threshold in \eqref{eq:LRTtype} to be ${\gamma} = -D(P_0\|P_1)+\frac{c}{\sqrt{n}}$, where $c$ is a constant that depends on distributions $P_0, P_1$ and $\epsilon$. Similarly, our setting where $P_0$ is known and a training sequence of the second hypothesis is available, we can define the Stein exponent as the highest error exponent under one hypothesis when the error probability under the other hypothesis is at most some fixed $\epsilon$ for all probability distributions $\tilde{P}_1$, .i.e., \begin{equation}\label{eq:Stein1trade} E_1^{(\epsilon)} \triangleq \sup \{ E_1\in \mathbb{R}_{+}: \exists \phi , \exists n_0 \in \ZZ_+ \ \text{s.t.} \ \forall n>n_0, ~ \forall \tilde{P_1} \in \Pc(\Xc), \epsilon_0(\phi| P_0, \tilde{P_1})\leq \epsilon \ \text{and} \ \epsilon_1(\phi| P_0,P_1) \leq e^{-nE_1} \}, \end{equation} where $\epsilon_0(\phi| P_0, \tilde{P_1})$ is the error probability when the generating distributions are $P_0, \tilde{P}_1$, i.e., for all possible distributions $\tilde{P}_1$ the type-\RNum{1} probability of error is bounded by some $\epsilon$ and $E_1^{(\epsilon)}$ is the maximum achievable type-\RNum{2} error exponent under the actual distribution generating data $P_1$. Similarly, \begin{equation}\label{eq:Stein0} E_0^{(\epsilon)} \triangleq \sup \{ E_0\in \mathbb{R}_{+}: \exists \phi , \exists n_0 \in \ZZ_+ \ \text{s.t.} \ \forall n>n_0 ,~ \epsilon_0(\phi| P_0, P_1)\leq e^{-nE_0}, \ \ \epsilon_1(\phi| P_0,\tilde{P}_1) \leq \epsilon \ \ \forall \tilde{P_1} \in \Pc(\Xc) \}. \end{equation} In other words, we are interested in the best possible error exponent for one of the hypothesis while the probability of error under alternative hypothesis is some $\epsilon$ universally for any distribution $P_1$. \begin{theorem}\label{thm:Stein} Let $\epsilon \in (0,1) $, then for any probability distributions $P_0, P_1$, the Stein regime exponents are given by \begin{align} E_1^{(\epsilon)} &=D(P_0\|P_1),\label{eq:stein1}\\ E_0^{(\epsilon)} &=D_{\frac{\alpha}{1+\alpha}} (P_1\|P_0)\label{eq:stein2}, \end{align} where $D_\rho(P_1\|P_0) = \frac{1}{\rho-1} \log \sum_{x\in \Xc} P_1^\rho P_2^{1-\rho}$ is the R\'enyi divergence of order $\rho$. \end{theorem} \begin{proof} The proof can be found in Appendix \ref{sec:proofth41}. \end{proof} Observe that $E_1^{(\epsilon)}$ is equal to the Stein exponent for the likelihood ratio test where both distributions are known. However, since R\'enyi divergence is a non-decreasing function of its order and $\frac{\alpha}{1+\alpha} <1$ hence $E_0^{(\epsilon)}$ is strictly smaller than the Stein regime exponent achieved by the likelihood ratio test. \section{Sequential Classification with a Known Hypothesis}\label{sec:seqclas} In this section, we study sequential classification with a known hypothesis. When both hypotheses are known, the sequential probability ratio test can achieve higher exponents compared to the likelihood ratio test. When only one of the hypotheses is known, Hoeffding's test can achieve the best error exponent achieved by the likelihood ratio test in the fixed sample size scenario. However, there is no counterpart to Hoeffding's test in the sequential case when only one of the hypotheses is known, i.e., no classifier can achieve the same error exponent performance as the sequential probability ratio test \cite{Wald}. We propose a classifier inspired by the sequential probability ratio test and show that having training samples from the second hypothesis can improve the error exponent tradeoff compared to the fixed sample-sized classification. In the sequential setting, the number of samples is a random variable called the stopping time $\tau$, taking values in $\ZZ_{+}$. A sequential classifier is a pair $\Phi=(\phi: \Xc^\tau\times \Xc^{\alpha\tau} \rightarrow \{0,1\},\tau)$, where for every $n\geq 0$ the event $\{\tau\leq n\} \in \mathscr{F}_n$, and $\mathscr{F}_n$ is the sigma-algebra induced by random variables $\xv^n , \X^{\alpha n} $, i.e., $ \sigma(\xv^n , \X^{\alpha n})$. We also assume that at every stage, additional training and test samples are available to the classifier, such that $\alpha=\frac{k}{n}$ remains constant. Moreover, $\phi$ is a $\mathscr{F}_{\tau}$ measurable decision rule, i.e., the decision rule determined by casually observing the sequence $\xv^n, \X^{\alpha n}$. In other words, at each time instant, the test attempts to decide in favor of one of the hypotheses or chooses to take new samples from the source $P_1$ as well as new samples from the unknown data source. The two possible pairwise error probabilities that measure the performance of the test are defined as \begin{equation}\label{eq:errprob} \epsilon_0(\Phi)=\PP_0\big [\phi(\xv^\tau , \X^{\alpha \tau} ) \neq 0 \big] ~ \text{,} ~ \epsilon_1(\Phi)=\PP_1\big[\phi(\xv^\tau , \X^{\alpha \tau})\neq 1 \big], \end{equation} where the probabilities are over $P_0, P_1$, respectively. Similarly to the sequential hypothesis testing case we define the optimal error exponent as \begin{align}\label{eq:tradeseq1} E^*_1(E_0) \triangleq \sup \Big \{E_1\in \mathbb{R}^{+}&: \exists \Phi ,\ \exists \ n \in \ZZ_{+} \text{ s.t.} \ \mathbb{E}_{P_0} [\tau] \leq n, \mathbb{E}_{P_1} [\tau] \leq n, \ \epsilon_0(\Phi) \leq 2^{- n E_0} ~ \text{and} ~ \epsilon_1(\Phi) \leq 2^{- n E_1} \Big \} . \end{align} When both hypotheses are known, the sequential probability ratio test (SPRT) $\Phi=(\phi,\tau)$ proposed by Wald \cite{Wald1} achieves the optimal exponent tradeoff. The sequential probability ratio test is given by \begin{align} \tau=\inf \big\{t\geq1:S_t& \geq \gamma_0 \ \text{or} \ S_t \leq -\gamma_1\big\}, \label{eq:deftau} \end{align} where \begin{align}\label{eq:LLR} S_n=\sum_{i=1}^t \log \frac{P_0(x_i)}{P_1(x_i)}, \end{align} is the the accumulated log-likelihood ratio (LLR) of the observed sequence $\xv$ and the thresholds $\gamma_0, \gamma_1$ are two positive real numbers. Moreover, the test makes a decision according to the rule \begin{align} \phi(\Tx)= \begin{cases} 0& \text{if } S_\tau \geq \gamma_0 \\ 1 & \text{if } S_\tau \leq - \gamma_1.\\ \end{cases} \label{eq:defsprt} \end{align} It is shown in \cite{Wald1} that the above test attains the optimal error exponent tradeoff, i.e., as thresholds $\gamma_0, \gamma_1$ approach infinity, the test achieves the best error exponent trade-off in \eqref{eq:tradeseq1}. It is known that the error probabilities of sequential probability ratio test as a function of $\gamma_0$ and $\gamma_1$ are \cite{Wood} \begin{align} \epsilon_0 = c_0 \cdot e^{-\gamma_1 } \quad , \quad \epsilon_1 =c_1 \cdot e^{-\gamma_0 } , \end{align} as $\gamma_0, \gamma_1 \rightarrow \infty$ where $c_0, c_1$ are positive constants. Moreover, it can also be shown that \begin{align} \mathbb{E}_{P_0} [\tau] &= \frac{ \gamma_0}{D(P_0\|P_1)}(1+o(1)) \label{eq:SPRT1}, \\ \mathbb{E}_{P_1} [\tau] &= \frac{ \gamma_1}{D(P_1\|P_0)}(1+o(1)) \label{eq:SPRT2}. \end{align} Therefore, according to definition \eqref{eq:tradeseq1}, the optimal error exponent tradeoff is given by, \begin{equation} E_0 = D(P_1\|P_0) , ~ E_1= D(P_0\|P_1), \end{equation} where thresholds $\gamma_0, \gamma_1$ are chosen as \begin{equation}\label{eq:thresh} \gamma_0=n\big (D(P_0\|P_1)+o(1)\big),~ \gamma_1= n\big(D(P_1\|P_0)+o(1)\big). \end{equation} Hence, the sequential probability ratio test achieves the Stein regime error exponents achievable by the standard likelihood ratio test \cite{Neyman} simultaneously. For every fixed $n$. we propose the following sequential classifier \begin{align} \tau=\inf \big\{t\geq n:S_t(\Tx,\TX)& \geq \gzt\ \text{or} \ S_t(\Tx,\TX) \leq -\got \big\} , \label{eq:deftau} \end{align} where \begin{align}\label{eq:testseq} S_t(\Tx,\TX) =\sum_{i=1}^t \log \frac{P_0(x_i)}{\TX'(x_i)}, \end{align} is the accumulated log-likelihood ratio (LLR) using the plugin perturbed type of the training sequence evaluated at the observed sequence $\xv$, and \begin{equation} \TX'=(1-\delta_n) \TX +\frac{\delta_n}{|\Xc|} \end{equation} where $\delta_n= o(n^{-1})$ and $\gzt, \got$ are chosen as \begin{equation}\label{eq:thresh} \gzt=nD(P_0\|\TX')+ (4 |\Xc|+4) \log(t+1),~ \got= nD(\TX\|P_0)+( 4|\Xc|+4) \log(t+1), \end{equation} and the test makes a decision according to the rule \begin{align}\label{eq:seqclas} \phi(\Tx,\TX)= \begin{cases} 0& \text{if } S_\tau(\Tx,\TX) \geq \gzt \\ 1 & \text{if } S_\tau(\Tx,\TX) \leq -\got,\\ \end{cases} \end{align} where $\Tx, \TX$ are types of the test and training samples at the stopping time $\tau$. As can be seen from the above expressions, the proposed classifier is the plugin sequential probability ratio test, replacing $P_1$ by the perturbed training type $\TX'$. The next theorem gives a lower bound on the achievable error exponent tradeoff of the proposed sequential classifier. \begin{theorem}\label{thm:seq} For every $P_0, P_1$, there exists a training to observation ratio $\alpha^*_{\rm seq}$ such that for any $\alpha \geq \alpha^*_{\rm seq}$, the sequential classifier $\Phi^{\rm seq}=(\phi(\Tx,\TX'),\tau)$ defined in \eqref{eq:deftau}, \eqref{eq:seqclas} achieves \begin{equation} E_0(\Phi^{\rm seq})E_1(\Phi^{\rm seq})\geq D(P_0\|P_1)D_{\frac{\alpha}{1+\alpha}}(P_1\|P_0). \end{equation} Furthermore, the average stopping times of the classifier satisfy \begin{equation} \mathbb{E}_{P_0}[\tau]= \mathbb{E}_{P_1}[\tau]= n(1+o(1)). \end{equation} \end{theorem} \begin{proof} The proof can be found in Appendix \ref{sec:proofth51}. \end{proof} This theorem shows that similar to the hypothesis testing problem with known distributions, the proposed sequential classifier can achieve the Stein regime exponents simultaneously when only one of the distributions is known. \begin{example}\label{ex:ex1} In Figures \ref{fig:e0seq}, we present a numerical example to illustrate the performance of the proposed sequential classifier in $ \Phi^{\rm seq}$ with $\gzt=nD(P_0\|\TX'), \got= nD(\TX\|P_0)$. Consider two binary distributions $P_0=Bern(0.45)$, $Bern(0.55)$ and set $\alpha= 10$. Each point in the figures is obtained by estimating the average error probability as follows. For each length of the test sequence $n$, we estimate the type-\RNum{1} and type-\RNum{2} error probabilities of the sequential classifier in \eqref{eq:testseq} by generating a sample from the test source and $\alpha$ samples from $P_1$ until the test stops and makes a decision. We have plotted the $-\frac{1}{n} \log \epsilon_i$ for $i\in \{0,1 \}$. We can notice that the type-\RNum{1} error exponent converges to $D_{\frac{\alpha}{1+\alpha}}(P_1\|P_0)$ while the type-\RNum{2} error exponent tends to $D(P_0\|P_1)$ as Theorem \ref{thm:seq} suggests. \end{example} \begin{figure}[htpb] \centering \input{e0claseq.tex} \input{e1claseq.tex} \caption{Type-\RNum{1} and type-\RNum{2} error exponents for the proposed sequential classifier. } \label{fig:e0seq} \end{figure} \begin{theorem}\label{thm:converse} For every sequential classifier $\Phi^{\rm seq}=(\phi^{\rm seq}(\Tx,\TX), \tau)$, such that $\mathbb{E}_{P_0}[\tau] \leq n, \mathbb{E}_{P_1}[\tau]\leq n$, we have \begin{equation} \max_{\Phi^{\rm seq}} \min_{P_1\in \Pc(\Xc)} E_0(\Phi^{\rm seq})E_1(\Phi^{\rm seq})\leq D(P_0\|P_1)D_{\frac{\alpha}{1+\alpha}}(P_1\|P_0). \end{equation} \end{theorem} This suggests that the proposed sequential classifier is universal in the sense that it achieves the highest error exponent tradeoff over all possible classifiers and distributions $P_1$. \begin{proof} The proof can be found in Appendix \ref{sec:proofth52}. \end{proof} \newpage \appendices \section{Proof of Theorem \ref{thm:E0}} \label{sec:proofth1} First, we prove that for every realization of the training sequence $\TX$, the type-\RNum{1} error exponent is equal to $E_0$. Note that the solution to the optimization problem in \eqref{eq:threshcls} is a convex problem, and by the Karush-Kuhn-Tucker (KKT) conditions \cite{Boyd}, the minimizer is unique and is the tilted distribution of $P_0$ and $\TX'$, i.e., \begin{equation}\label{eq:tiltedmu} Q_{\mu^*}(x)= \frac{ P_{0}^{\frac{\mu^*}{1+\mu^*}}(x) {\TX'}^{\frac{1}{1+\mu^*}}(x) } {\sum_{a \in \Xc } P_{0}^{\frac{\mu^*}{1+\mu^*}}(a) {\TX'}^{\frac{1}{1+\mu^*}}(a) }, \end{equation} and $\mu^*$ is the solution to \begin{equation}\label{eq:KKTgamma} D(Q_{\mu^*}\| P_0)= E_0. \end{equation} Therefore, the classifier threshold in \eqref{eq:threshcls} can be written as \begin{equation}\label{eq:threshsol} \gamma(E_0,\TX')=\beta D(Q_{\mu^*}\| \TX')-D(Q_{\mu^*}\| P_0). \end{equation} Now by Sanov's theorem, we can find the type-\RNum{1} error exponent by solving \eqref{eq:min1} when $P_1$ replaced by $\TX'$ and $\gamma$ is replaced by \eqref{eq:threshsol}. This optimization problem is convex in $Q$ for every $\TX'$ when $\beta=1$, however for $\beta<1$ this is not the case and the KKT conditions are only necessary conditions. Spelling out the Lagrangian we obtain the KKT conditions \begin{align} L(Q,\lambda,\nu)&= D(Q\|P_0)+ \lambda \big(\beta D(Q\|\TX')- D(Q\|P_0)-\gamma(E_0,\TX')\big) +\nu \Big(\sum_{x\in\Xc} Q(x)-1\Big)\\ \frac{\partial L(Q,\lambda,\nu)}{\partial Q(x)}&= 1+\log \frac{Q(x)}{P_0(x)} +\lambda \Bigg(\beta+\beta \log \frac{Q(x)}{\TX'(x)}-1- \log \frac{Q(x)}{P_0(x)}\Bigg)+\nu. \end{align} Setting the derivative to zero, we get \begin{align}\label{eq:tiltedlambda} Q_{\beta , \lambda^*} (x)= \frac{ P_{0}^{\frac{1-\lambda^*}{1-\lambda^*+\lambda^* \beta}}(x) {\TX'}^{\frac{\lambda^*\beta}{1-\lambda^*+\lambda^* \beta}}(x) } {\sum_{a \in \Xc } P_{0}^{\frac{1-\lambda^*}{1-\lambda^*+\lambda^* \beta}}(a) {\TX'}^{\frac{\lambda^*\beta}{1-\lambda^*+\lambda^* \beta}}(a) }, ~~~0\leq\lambda^* , \end{align} and by complementary slackness condition \cite{Boyd} \begin{equation} D(Q_{\beta , \lambda^*}\| P_0)-\beta D(Q_{\beta, \lambda^*} \| \TX') = D(Q_{\mu^*}\| P_0)-\beta D(Q_{\mu^*}\| \TX'). \end{equation} Note that, $\lambda^*$ cannot be zero as that sets $Q_{\beta , \lambda^*}= P_0(x)$ which is invalid solution when $E_0>0$. It is easy to see that this equality is satisfied when $\frac{1-\lambda^*}{1-\lambda^*+\lambda^* \beta} =\frac{\mu^*}{1+\mu^*} $. It is also easy to see this is the unique solution as $\frac{1-\lambda^*}{1-\lambda^*+\lambda^* \beta} $ is strictly decreasing function in $\lambda^*$ and hence $D(Q_{\beta , \lambda^*}\|P_0)$ and $D(Q_{\beta , \lambda^*}\|\TX')$ are strictly increasing and strictly decreasing in $\lambda^*$ respectively. Therefore, we get \begin{equation}\label{eq:KKTgammaLRT} D(Q_{\beta, \lambda^*}\| P_0)= E_0. \end{equation} Next, by rewriting the type-\RNum{1} error probability as a function of the types of the training and observation sequences $\Tx,\TX$ we have \begin{align}\label{eq:e0saddle} \epsilon_0(\cls)= \sum_{\TX\in \Pc_k(\Xc)} P_1(\TX ) \sum_{\substack{ \Tx \in \Pc_n(\Xc), \\ \cls(\Tx,\TX) =1}} P_0(\Tx), \end{align} that is the type-\RNum{1} error exponent is equal to $E_0$ for every realization of training sequence $\TX$. As the hypothesis test is a type base test, and by letting $P_0(\mathcal{T}_{Q}^n) = n^{\frac{-|\Xc|+1}{2}} e^{-nD(Q\|P_0)}(c+o(1))$ \cite{Csiszarinfo}, we obtain \begin{equation}\label{eq:prefactore0} \epsilon_0(\cls)= \sum_{\TX \in \Pc_k(\Xc)} P_1(\TX) \sum_{ \Tx \in \Ao}n^{\frac{-|\Xc|+1}{2}} e^{-nD(\Tx \|P_0)}(c+o(1)). \end{equation} To find the polynomial decay of the error probability, we use the following theorems by \cite{Reeds, Iltis, Ney}, to approximate the summation by an integral and then use the saddle point approximation. We also use the shorthand notation $a_n \asymp b_n$ for any two positive real sequences such that $\log \frac{a_n}{b_n} = \Oc(1)$. \begin{theorem}\label{thm:reeds} \cite{Reeds} Suppose $\psi: \mathbb{R}^d \rightarrow \mathbb{R}$ is a Lipschitz continuous function and the open set $\Dc \subseteq \mathbb{R}^d $ has minimally smooth boundary. Then \begin{equation} \sum_{\wv \in \Dc \cap \Pc_k(\Xc) } e^{ -k\psi(\wv) } \asymp k^{d} \int_{\Dc} e^{ -k\psi(\wv) } \ d\wv. \end{equation} \end{theorem} The precise definition of minimally smooth boundary can be found in \cite{Reeds}; we only work with smooth boundaries which are also minimally smooth. Letting $\TX'$ be fixed, it can be shown that $\phi(Q)=D(Q\| P_0)-\beta D(Q \| \TX') $ is a smooth function, and hence we can conclude the boundary of the decision region $\Ac_0(\TX, \beta)$ is smooth since the graph of a smooth function is a smooth manifold \cite{Lee}. Furthermore, by the continuous differentiability of $\psi(Q)=D(Q\|P_0)$ in $Q$, the conditions of the Theorem \ref{thm:reeds} are satisfied and we can approximate the summation in \eqref{eq:prefactore0} by the integral \begin{equation}\label{eq:reeds} \sum_{ \wv \in \Ao \cap \Pc_n(\Xc) } e^{-nD(\wv \|P_0)} \asymp n^{|\Xc|-1} \int_{\Ao } e^{-nD(\wv \|P_0)} d\wv. \end{equation} Next we will use the following theorem which can be derived from \cite{Iltis} to approximate the integral. \begin{theorem}\cite{Iltis} For a set $\Gamma$ in $\mathbb{R}^d$ having smooth boundary with a unique dominating point $\nu$ where the large deviation rate function $I(x)$ is minimized with $I(\nu)=a$, let the level surface $S_a=\{x\in \mathbb{R}^d : I(x)=a \}$ be described locally by $x_d=f(x)$, where $x = (x_1,, x_{d-1})$, and let the region $\Gamma$ be bounded by the surface $x_d=g(x)$, where $g(x)$ is a three times differentiable function of $x$ in a neighbourhood of $0$. If the Hessians of $g$ and $f$ at $0$ satisfy $H_g(0)>H_s(0 )$ then \begin{align} \int_{\wv \in \Gamma} e^{-nD(\wv \|P_0)} d \wv =n^{-\frac{d+1}{2}} e^{-na} (d_0 + o( 1)), \label{eq:intsaddle} \end{align} where \end{theorem} Since the optimizing distribution solving Sanov's Theorem is unique and equal to $Q_{\lambda^*}$, the dominating point is unique. Hence, we only need to prove the the Hessian of the decision region $\Dc$ minus the Hessian of the level surface $\Sc_{E_0}=\{Q \in \Pc(\Xc) :D(Q\|P_0)\leq E_0 \} $ is positive definite matrix. Writing $Q_{\beta, \lambda^*}(|\Xc|) = 1- \sum_{x=1}^{|\Xc|-1} Q_{\beta , \lambda^*}(x)$ and taking the derivatives we have: \begin{align} \Hm_{\Dc}^{(|\Xc|-1) \times (|\Xc|-1)}&= -(1-\beta) \Bigg ( \diag \Big (\frac{1}{Q_{\beta , \lambda^*}} \Big )+\frac{ \onev \onev^T}{Q_{\beta , \lambda^*}(|\Xc|)} \Bigg ) ,\\ \Hm_{\Sc_{E_0}}^{(|\Xc|-1) \times (|\Xc|-1)}&=- \diag \Bigg (\frac{1}{Q_{\beta , \lambda^*}} \Bigg )-\frac{ \onev \onev^T}{Q_{\beta , \lambda^*}(|\Xc|)} , \end{align} and hence \begin{equation} \Hm_{\Sc_{E_0}}-\Hm_{\Dc}=\beta \diag \Bigg (\frac{1}{Q_{\beta , \lambda^*}} \Bigg ) + \beta \frac{ \onev \onev^T}{Q_{\beta , \lambda^*}(|\Xc|)} , \end{equation} which is positive for any $\beta >0$. Therefore, by \eqref{eq:prefactore0}, \eqref{eq:reeds}, \eqref{eq:intsaddle}, we have \begin{align} \epsilon_0(\cls)&= \sum_{\TX} P_1(\TX) n^{\frac{-|\Xc|+1}{2}} n^{|\Xc|-1} n^{-\frac{|\Xc|}{2}} e^{-nE_0}(c+o(1))\\ &= \sum_{\TX} P_1(\TX) \frac{1}{\sqrt{n}} e^{-nE_0}(c+o(1)). \end{align} Finally, since $c$ is finite, positive, and only depends on $P_0, P_1, E_0$, there exists a $\tilde c$ such that \begin{align} \epsilon_0(\cls)= &\frac{1}{\sqrt{n}} e^{-nE_0} (\tilde c+ o(1)\big), \end{align} which concludes the proof. \section{Proof of Theorem \ref{thm:E1}} \label{sec:proofth2} Using standard properties of the method of types \cite{Csiszarinfo}, the type-\RNum{2} probability of error can be written as \begin{align} \epsilon_1 (\cls)&= \sum_{\xv, \X: \cls(\Tx,\TX)=0 } P_1(\X) P_1(\xv)\\ &= \sum_{\substack{Q\in \Pc_n(\Xc) , Q_1 \in \Pc_k(\Xc),\\ \cls(Q,Q_1)=0 }} P_1(\mathcal{T}_{Q}^n) P_1(\mathcal{T}_{Q_1}^k) \label{eq:method}\\ &\leq \sum_{\substack{Q\in \Pc_n(\Xc) , Q_1 \in \Pc_k(\Xc),\\ \cls(Q,Q_1)=0 }} e^{-nD(Q\|P_1)-kD(Q_1\|P_1)}\\ &\leq (n+1)^{|\Xc|}(k+1)^{|\Xc|} 2^{-n \tilde{E}_1^{(n)}(\cls) } \end{align} where \begin{equation}\label{eq:exponent} \tilde{E}_1^{(n)}(\cls)= \min_{ \substack{ \cls(Q,Q_1)=0 \\ Q,Q_1 \in \Pc(\Xc) } }D(Q\|P_1)+\alpha D(Q_1\|P_1). \end{equation} It is worth observing that this optimization problem is non-convex. In addition, lower bounding \eqref{eq:method} yields \begin{align} \epsilon_1 (\cls) \geq (n+1)^{-|\Xc|} (k+1)^{-|\Xc|} e^{-n\tilde{E}_1^{(n)}(\cls) }. \end{align} Observe that the proposed test $\cls$ depends on $n$; hence if the limit of \eqref{eq:exponent} as $n$ goes to infinity exists, then the limit is the type-\RNum{2} error exponent. Using a change of variable, we can write \eqref{eq:exponent} as \begin{equation} \begin{aligned} \tilde{E}_1^{(n)}(\cls)=\min_{Q,Q_1} ~ & D(Q\|P_1)+\alpha D\Bigg(\frac{1}{1-\delta_n} Q_1 - \frac{\delta_n}{1-\delta_n} U \Big \|P_1\Bigg) \\ \textrm{s.t.} ~& \beta D(Q\|Q_1) -D(Q\|P_0) \geq \gamma(E_0,Q_1)\\ &Q \in \Pc(\Xc) \\ & Q_1 \in (1-\delta_n)\Pc(\Xc) + \delta_n U \end{aligned} \end{equation} where $U$ is the uniform distribution on $\Pc(\Xc)$. Now, by using a Taylor expansion of $D\big(\frac{1}{1-\delta_n} Q_1 - \frac{\delta_n}{1-\delta_n} U \|P_1\big)$ around $\delta_n=0$, we get \begin{equation} \begin{aligned} \tilde{E}_1^{(n)}(\cls)=\min_{Q,Q_1} \quad & D(Q\|P_1)+\alpha D(Q_1 \|P_1) + \Oc\Big(\frac{\delta_n}{1-\delta_n}\Big) \\ \textrm{s.t.} \quad &\beta D(Q\|Q_1) -D(Q\|P_0) \geq \gamma(E_0,Q_1)\\ &Q \in \Pc(\Xc) \\ & Q_1 \in (1-\delta_n)\Pc(\Xc) + \delta_n U \end{aligned} \end{equation} Therefore, we can approximate the type-\RNum{2} error exponent by \begin{equation}\label{eq:approxcontin} E^{(n)}_1(\cls) = \min_{\substack{\beta D(Q\|Q_1) -D(Q\|P_0) \geq \gamma(E_0,Q_1) \\ Q \in \Pc(\Xc), Q_1 \in \Pc_{\delta_n}(\Xc)}} D(Q\|P_1) +\alpha D(Q_1\|P_1) \end{equation} where \begin{equation} \Pc_{\delta_n}(\Xc) = \big\{Q: Q=(1-\delta_n)P + \delta_n U, P \in \Pc(\Xc) \big\}, \end{equation} and the approximation error is of order $\Oc\big(\frac{\delta_n}{1-\delta_n}\big)$. Letting $\delta_n=o(n^{-1})$ we have \begin{equation} \epsilon_1 (\cls) = e^{-n{E}_1^{(n)}(\cls)+o(1)} \asymp e^{-n{E}_1^{(n)}(\cls)}. \end{equation} Hence, the type-\RNum{2} error exponent can be calculated as the limit of \eqref{eq:approxcontin} when $n$ goes to infinity. We also define following optimization problems. For every $Q_1 \in \Pc_{\delta_n}(\Xc)$, let \begin{equation}\label{eq:firstopt} E_1(E_0,Q_1)=\min_{\substack{\beta D(Q\|Q_1) -D(Q\|P_0) \geq \gamma(E_0,Q_1) \\ Q \in \Pc(\Xc), } } D(Q\|P_1) , \end{equation} which is the error exponent when the type of the training sequence is $Q_1$. Also, let \begin{equation}\label{eq:secondopt} E_1^{(n)}(E_0,r)=\min_{\substack{D(Q_1\|P_1)\leq r \\ Q_1 \in \Pc_{\delta_n}(\Xc) } } E_1(E_0,Q_1). \end{equation} The latter optimization problem is the worst case achievable error exponent by $\cls$ if we know the training sequence type $Q_1$ is inside a relative entropy ball around the original distribution $P_1$ of radius $r$ \cite{BoroumandTran} . Using \eqref{eq:firstopt}, \eqref{eq:secondopt}, we can wirte \eqref{eq:approxcontin} as \begin{align} E_1^{(n)} (\cls)&= \min_{\substack{\beta D(Q\|Q_1) -D(Q\|P_0) \geq \gamma(E_0,Q_1), \\ D(Q_1\|P_1)=r~ , ~ r \geq 0 , \\ Q \in \Pc(\Xc) ~,~ Q_1 \in \Pc_{\delta_n}(\Xc) } } D(Q\|P_1) +\alpha r\\ &=\min_{\substack{ D(Q_1\|P_1) = r ~,~ r \geq 0 , \\Q_1 \in \Pc_{\delta_n}(\Xc)} } E_1(E_0,Q_1)+\alpha r \\ &=\min_{\substack{D(Q_1\|P_1) \leq r ~,~ r \geq 0 \\ Q_1 \in \Pc_{\delta_n}(\Xc)} } E_1(E_0,Q_1)+\alpha r \label{eq:concaveopt} \\ &=\min_{ r \geq 0 } E_1^{(n)}(E_0,r)+\alpha r, \label{eq:mismatch} \end{align} where in \eqref{eq:concaveopt} we used the KKT conditions to show that the minimizer $Q_1^*$ of \eqref{eq:concaveopt} should satisfy the inequality condition $D(Q_1\|P_1)\leq r$ with equality, i.e., $D(Q^*_1\|P_1) = r$. We will use \eqref{eq:mismatch} to analyze the behaviour of the non-convex problem in \eqref{eq:exponent}. We also define $r_{c}$ as \begin{equation} r_c=\min_{\substack{Q_1\in \Pc(\Xc) \\ \min_{i\in\Xc} Q_1(i)=0}} D(Q_1\|P_1), \end{equation} that is the minimum distance between $P_1$ and distributions on the boundary of the probability simplex $\Pc(\Xc)$. Since $P_1(i)$ is non zero for every $i$, therefore $r_{c}>0$, and we can lower bound $E_1^{(n)}(\cls)$ by \begin{align} E_1^{(n)} (\cls)&\geq \min\Big \{ \min_{0\leq r \leq r_c-\epsilon} E_1^{(n)}(E_0,r)+\alpha r, \alpha(r_c-\epsilon) \Big \}, \end{align} where $0<\epsilon<r_c$, and we used the fact that $E^{(n)}_1(E_0,r) \geq0$ for $r \geq 0$ by non-negativity of relative entropy. For $r\in[0,r_c-\epsilon]$, we also define $E_1(E_0,r)$ as \begin{align} E_1(E_0,r) =& \lim_{n\rightarrow \infty} E_1^{(n)}(E_0,r)\\ =& \lim_{n\rightarrow \infty} \min_{\substack{D(Q_1\|P_1)\leq r \\ Q_1 \in \Pc_n(\Xc) } } E_1(E_0,Q_1)\\ =& \min_{\substack{D(Q_1\|P_1)\leq r \\ Q_1 \in \Pc(\Xc) } } E_1(E_0,Q_1). \label{eq:limitE} \end{align} Next, we use the following lemma to find a lower bound on type-\RNum{2} error exponent. \begin{lemma}\label{lem:cont} Let $r \in[0,r_c)$, then for every $n \in \mathbb{Z}^{+}$, the exponent functions $E^{(n)}_1(E_0,r)$, and $E_1(E_0,r)$ are continuous in $r$. \end{lemma} The proof of Lemma \ref{lem:cont} relies on Berge's maximum theorem \cite{Berge} and we defer it to Appendix \ref{app:cont}. Taking the limit of $E_1^{(n)} (\cls)$ when $n$ goes to infinity, we have \begin{align} E_1 (\cls)&= \liminf_{n \rightarrow \infty} E_1^{(n)} (\cls) \nonumber\\ &\geq\liminf_{n \rightarrow \infty} \min\Big \{ \min_{0\leq r \leq r_c-\epsilon} E_1^{(n)}(E_0,r)+\alpha r, \alpha(r_c-\epsilon) \Big \}\\ &= \min\Big \{ \min_{0\leq r \leq r_c-\epsilon} E_1(E_0,r)+\alpha r, \alpha(r_c-\epsilon) \Big \} \label{eq:lim}, \end{align} In order to show \eqref{eq:lim}, by Lemma \ref{lem:cont}, $E^{(n)}_1(E_0,r)$, and $E_1(E_0,r)$ are continuous on the compact interval $[0,r_c-\epsilon]$. Also, since $E_1^{(n)}(E_0,r) \geq E_1^{(n+1)}(E_0,r)$ for every $r$ in the domain, by Dini's theorem the convergence of $E^{(n)}_1(E_0,r)$ is uniform \cite{Rudin}. Finally, since the convergence is uniform we can interchange the minimum and limit to get \eqref{eq:lim}. When $r=0$, we have $Q_1=P_1$ and $\gamma(E_0,P_1)=\beta E_1^*(E_0)-E_0$, resulting $E_1(E_0,r=0)=E_1^*(E_0)$ where $E_1^*(E_0)$ is the optimal error exponent tradeoff in \eqref{eq:tradefix}. Moreover, for $r_1\leq r_2$, we have $\mathcal{B}(P_1,r_1) \subseteq \mathcal{B}(P_1,r_2)$ where $\mathcal{B}(P,r)=\{ Q \in \Pc(\Xc): D(Q\|P) \leq r\}$. Hence $E_1(E_0,r)$ is a non-increasing function of $r$. Therefore, the sufficient condition to have $E_1(\cls)=E_1^*(E_0)$ is \begin{equation}\label{eq:cond} \min\Big \{ \min_{0\leq r \leq r_c-\epsilon} E_1(E_0,r)+\alpha r, \alpha(r_c-\epsilon) \Big \} =E_1(E_0,r=0). \end{equation} Equivalently, letting $\epsilon$ to go to zero, \eqref{eq:cond} can be written as the following two sufficient conditions \begin{align} E_1(E_0,r)+\alpha r &\geq E_1(E_0,r=0) ~~ \forall r: ~ 0 \leq r < r_c,\\ \alpha &> \frac{E_1(E_0,r=0)}{r_c}. \end{align} Hence, letting \begin{equation}\label{eq:optimalalpha} \alpha^* _\beta= \max \bigg \{\sup_{0\leq r < r_c} \frac{E_1(E_0,r=0)-E_1(E_0,r)}{r}, \frac{E_1^*(E_0)}{r_c} \bigg \} \end{equation} for any $\alpha > \alpha^*_\beta$, we have $E_1(\cls) =E_1^*(E_0)$. Finally, since $E_1(E_0,r)$ is non-increasing and continuous on $r\in [0,r_c)$, and $-\frac{\partial E_1(E_0,r)}{\partial r} \big|_{r=0} < \infty$ as it will be proved in Theorem \ref{thm:lower}, the supremum in \eqref{eq:optimalalpha} is finite and \begin{equation} 0<\alpha^*_\beta< \infty, \end{equation} and this concludes the proof. \section{Proof of Theorem \ref{thm:refinement}} \label{sec:proofth3} Using the method of types, the type-\RNum{2} probability of error can be written as \begin{align} \epsilon_1 (\cls)&= \sum_{\xv, \X: \cls(\Tx,\TX)=0 } P_1(\X) P_1(\xv)\\ &=\sum_{Q_1 \in \Pc_k(\Xc)\cap \Pc_{\delta_n} (\Xc) } P_1(\mathcal{T}_{Q_1}^k) \sum_{ \substack{Q \in \Pc_n(\Xc),\\ \cls(Q,Q_1) =0}} P_1(\mathcal{T}_{Q}^n) \\ &\asymp \sum_{Q_1 \in \Pc_k(\Xc) \cap \Pc_{\delta_n} (\Xc) } c k^{\frac{-|\Xc|+1}{2}} e^{-kD(Q_1\|P_1)} \sum_{ \substack{Q \in \Pc_n(\Xc),\\ \cls(Q,Q_1) =0}} P_1(\mathcal{T}_{Q}^n)\label{eq:errorprob1} \end{align} where in the last step we used \cite{Csiszarinfo} \begin{equation} P_1(\mathcal{T}_{Q_1}^k) \asymp k^{\frac{-|\Xc|+1}{2}} e^{-kD(Q_1\|P_1)}. \end{equation} We need to show that conditioned on the training sequence having a type $Q_1$, the conditional type-\RNum{2} error probability is \begin{equation}\label{eq:mismatchprefac} \sum_{ Q: \cls(Q,Q_1) =0} P_1(\mathcal{T}_{Q}^n)= \frac{1}{\sqrt{n}} e^{-nE_1(E_0,Q_1)} \big (c+ o(1)\big), \end{equation} when $E_1(E_0,Q_1)$ is positive, i.e, the error decay has a prefactor of $\frac{1}{\sqrt{n}}$. We first show that for any classifier with $0 \leq \beta \leq 1$ we can upper bound the type-\RNum{2} error probability by the classifier type-\RNum{2} error probability with $\beta=1$. \begin{lemma}\label{lem:inclusion} For every fixed $Q_1$ and for every $\beta_1\leq \beta_2$ we have $\Ac_0(Q_1,\beta_1) \subseteq \Ac_0(Q_1,\beta_2)$. \end{lemma} \begin{proof} Let $Q \in \Ac_0(Q_1,\beta_1)$, we need to show that for every such $Q$ , we have $Q \in \Ac_0(Q_1,\beta_2)$. We then have \begin{align} D(Q\|P_0) - \beta_1 D(Q\|Q_1) \leq D(Q_{\mu^*}\|P_0) - \beta_1 D(Q_{\mu^*}\|Q_1) \end{align} where $Q_{\mu^*}$ is the solution to optimization in $\gamma(E_0,Q_1)$ in \eqref{eq:tiltedmu} and is independent of $\beta$. Furthermore, \begin{align}\label{eq:betainclusion} D(Q\|P_0) - \beta_2 D(Q\|Q_1)\leq D(Q_{\mu^*}\|P_0)- \beta_2 D(Q_{\mu^*}\|Q_1) + (\beta_2-\beta_1) \Big ( D(Q_{\mu^*}\|Q_1) -D(Q\|Q_1)\Big ). \end{align} Similarly to the proof of Theorem \ref{thm:E0}, using the KKT conditions it can be shown that the projection of $Q_1$ to the set $\Ac(Q_1,\beta)$ also equals to $Q_{\mu^*}$, i.e., \begin{equation} Q_{\mu^*}=\argmin_{D(Q\|P_0) - \beta_1 D(Q\|Q_1) \leq \gamma(E_0,Q_1)} D(Q\|Q_1). \end{equation} Hence for every $Q\in \Ac_0(Q_1,\beta_1)$ we have \begin{equation} D(Q_{\mu^*}\|Q_1) \leq D(Q\|Q_1), \end{equation} and by \eqref{eq:betainclusion} \begin{align} D(Q\|P_0) - \beta_2 D(Q\|Q_1)\leq D(Q_{\mu^*}\|P_0)- \beta_2 D(Q_{\mu^*}\|Q_1), \end{align} i.e., $Q \in \Ac_0(Q_1,\beta_2)$ which concludes the proof. \end{proof} Therefore we can upper bound the error probability by setting $\beta=1$. Since for every fixed $Q_1$ the test is the Neyman-Pearson test using the mismatched distribution $Q_1$, the decision region boundary is characterised by a hyperplane and by \cite{Reeds} the conditional type-\RNum{2} error probability prefactor equals to $\frac{1}{\sqrt{n}}$ for every $Q_1$. Hence, substituting \eqref{eq:mismatchprefac} into \eqref{eq:errorprob1} we have \begin{align} \epsilon_1 (\cls) \leq \sum_{Q_1 \in \Pc_k(\Xc) \cap \Pc_{\delta_n} (\Xc) } \frac{k^{\frac{-|\Xc|+1}{2}}}{\sqrt{n}} e^{-kD(Q_1\|P_1)} e^{-nE_1(E_0,Q_1)} \big (c+ o(1)\big). \label{eq:errorprob1saddle} \end{align} Observe that one can also use the similar approach in the proof of Theorem \ref{thm:E0}, to show stronger result that for all $0<\beta \leq 1 $ the prefactor equals to $\frac{1}{\sqrt{n}}$. Also notice that by choosing some small enough $\delta_n>0$, and constraining $\TX \in \Pc_{\delta_n}(\Xc)$, the approximation error caused by substituting $\TX'$ by $\TX$ can be summarized in the $c+o(1)$ term in \eqref{eq:errorprob1saddle}, and we get \begin{align} \epsilon_1 (\cls)&\asymp n^{-\frac{|\Xc|}{2}} \sum_{Q_1 \in \Pc_k(\Xc) } e^{ -k\psi(Q_1) }, \label{eq:sumsaddle}\\ \psi(Q_1)&=\frac{1}{\alpha}E_1(E_0,Q_1) + D(Q_1\|P_1), \end{align} where we used the fact that $\alpha=\frac{k}{n}$ is fixed and independent of $n, k$. Next we approximate the summation by an integral using Theorem \ref{thm:reeds}. As shown in Theorem \ref{thm:E1} \begin{equation} P_1=\argmin_{Q_1: D(Q_1\|P_1) <r_c} \psi(Q_1) \end{equation} for $\alpha^* _\beta < \alpha$. We show that the terms in the summation \eqref{eq:sumsaddle} where $Q_1$ is in the vicinity of the $P_1$ dominate the summation. Let $\Dc=\Pc(\Xc)$ and $\Dc'=\Bc(P_1,\epsilon)$ where $ \epsilon = \frac{ s \log k}{k} $, $s>0$. We can split the summation \eqref{eq:sumsaddle} as \begin{align} \sum_{Q_1 \in \Dc \cap \Pc_k(\Xc) } e^{ -k\psi(Q_1) }& = \sum_{Q_1 \in \Dc' \cap \Pc_k(\Xc) } e^{ -k\psi(Q_1) } +\sum_{Q_1 \in (\Dc')^c \cap \Pc_k(\Xc) } e^{ -k\psi(Q_1) }. \label{eq:tworegions} \end{align} By Theorem \ref{thm:lower}, $\psi(Q_1)$ achieves its minimum at $Q_1=P_1$ for $\alpha>\alpha^*_\beta$, i.e., $\psi(Q_1)> \psi(P_1)$ for $Q_1\neq P_1$. Hence, by the mean value theorem and using a Taylor series expansion of $\psi(Q)$ around $Q_1=P_1$ for $ Q_1 \in (\Dc')^c$ we get \begin{equation} \psi(Q_1) = \psi(P_1) + \frac{1}{2}\thetav_{P_1}^{T} \tilde{\Hm}(Q_1) \thetav_{P_1} \end{equation} where $\thetav_{P_1}$ is the difference vector $Q_1-P_1$ defined in \eqref{eq:diffQ1}, and $\tilde{\Hm}(Q_1)$ is a positive definite matrix for every $Q_1 \in (\Dc')^c$. Also, we have used the result form Theorem \ref{thm:lower} that the first order term of the expansion is zero. Furthermore, note that by the condition $ Q_1 \in (\Dc')^c$ we have $\| \thetav_{P_1} \|_\infty^2 \geq O(\epsilon)$. Hence for large enough $n$ we have \begin{equation} \sum_{Q_1 \in (\Dc')^c \cap \Pc_k(\Xc) } e^{ -k\psi(Q_1)} \leq (k+1)^{|\Xc|} e^{-k\Big (\frac{1}{\alpha} E_1(E_0,P_1) +\frac{1}{2} \Lambda \frac{s \log k}{k}\Big )} \end{equation} where $\Lambda= \min_{Q_1\in(\Dc')^c } \Lambda_{\min}\Big(\tilde{\Hm}(Q_1)\Big )$, and $\Lambda_{\min}(.)$ is the smallest eigenvalue of a matrix. Note that $\Lambda$ is strictly positive since $\psi(Q_1)> \psi(P_1)$ for $Q_1\neq P_1$. Finally, by setting $s \geq \frac{2 } { \Lambda } (|\Xc|+\frac{3}{2}) $ we have \begin{equation} \sum_{Q_1 \in (\Dc')^c \cap \Pc_k(\Xc) } e^{ -k\psi(Q_1)} \leq n^{-\frac{1}{2}} e^{-n E_1(E_0,P_1) } O(n^{-1}). \end{equation} It remains to bound the first term in \eqref{eq:tworegions}. By the envelope theorem \cite{Segal} as shown in Lemma \ref{lem:HessianE}, we can conclude that $\psi(\wv): \mathbb{R}^{|\Xc|-1} \rightarrow \mathbb{R}$ is continuously differentiable and hence Lipschitz on $\Dc'$. Furthermore, $\Dc'$ is a $C^{\infty}$ manifold, and hence $\Dc'$ has a minimally smooth boundary. Therefore, by Theorem \ref{thm:reeds} we have \begin{align} \epsilon_1 (\cls)&\asymp n^{\frac{|\Xc|-2}{2}} \int_{\Dc'} e^{ -k\psi(\wv) } \ d\wv. \end{align} By using a Taylor series expansion of $\psi(\wv)$ around $\wv=P_1$, and Lemma \ref{lem:HessianE} we have \begin{align} \epsilon_1 (\cls)&\asymp n^{\frac{|\Xc|-2}{2}} \int_{\Dc'} e^{ -n \big(E_1(E_0,P_1) + \frac{1}{2}<(\wv-P_1),(\Hm_2+\alpha \Jm) (\wv-P_1)> +O\big (\epsilon^{\frac{3}{2}} \big) \big )} d(\wv-P_1). \end{align} Note that $\wv$ is bounded to the the $|\Xc|-1$ dimensional probability simplex as $\wv \in \Dc'$. Hence for every $\wv-P_1$ there exists a $|\Xc| \times |\Xc|$ rotation matrix $\Rm$ such that for every $\wv \in \Pc(\Xc) $ we have $\ev_{|\Xc|}^T \Rm (\wv-P_1) =0$, where $\ev_{|\Xc|}$ is the unit vector with $1$ in the $|\Xc|$'th coordinate and $0$ otherwise. Letting $\wv' = \Rm(\wv-P_1)$ we have \begin{align} \epsilon_1 (\cls)&\asymp n^{\frac{|\Xc|-2}{2}} \int_{\Dc' \cap \R^{|\Xc|-1}} e^{ -n \big(E_1(E_0,P_1) + \frac{1}{2 }\big<\Rm^T \wv',(\Hm_2+\alpha \Jm)\Rm^T \wv'\big> +O\big (\epsilon^{\frac{3}{2}} \big) \big )} d \Rm^T\wv', \end{align} where we have used $\Rm \Rm^T=\Id$. Furthermore, from the proof of Theorem \ref{thm:lower}, $\Hm_2+\alpha \Jm$ is strictly positive definite and hence full rank for $\alpha^*_\beta < \alpha$. Using a change of variables $\yv=\sqrt{n}\wv'$ we have $d\yv=n^{-\frac{|\Xc|-1}{2}} d\wv'$, and \begin{align} \epsilon_1 (\cls)&\asymp n^{-\frac{1}{2}} e^{ -n E_1(E_0,P_1)} \int_{\Dc'} e^{ - \frac{1}{2}\big<\Rm^T \yv,(\Hm_2+\alpha \Jm)\Rm^T \yv\big> +O\big (\sqrt{\frac{\log^3{n}}{n}}\big ) }d\yv, \end{align} where the integral is a Gaussian integral and equals to $c'+o(1)$ for a constant $c'$ and this concludes the proof. \section{Proof of Thereom \ref{thm:lower}} \label{sec:proofth5} From \eqref{eq:optimalalpha} we have \begin{equation} \underline{\alpha}_\beta\triangleq- \frac{\partial E_1(E_0,r)}{\partial r} \bigg|_{r=0} \leq \alpha^*_\beta. \end{equation} To find the derivative of $E_1(E_0,r)$ at $r=0$, we use a Taylor series expansion of $E_1(E_0,Q_1)$ around $Q_1=P_1$. We use the following lemmas in our proof. First lemma provides the first and second derivative of $\gamma (E_0,Q_1)$. \begin{lemma}\label{lem:gamsen} For every $P_0, P_1 \in \Pc(\Xc)$, the gradient of threshold function $\gamma(E_0,Q_1)$ is \begin{align}\label{eq:firstdergam} \frac{\partial \gamma (E_0,Q_1)}{\partial Q_1 (i)} = -\beta \frac{Q_{\mu^*}(i)}{Q_1(i)}, \end{align} Furthermore the Hessian matrix of $\gamma(E_0,Q_1)$ evaluated at $Q_1=P_1$ is \begin{align} \Hm_1 &= \Hm(Q_1) \Big |_{Q_1=P_1}, \\ \Hm(Q_1) &= \nabla_{Q_1}^2 \gamma(E_0,Q) \Big |_{Q_1=P_1} \\ &= \beta \Jm\Bigg (\frac{\mu^*}{1+\mu^*} \Tm+ \frac{1}{(1+\mu^*)} (\Qm+ \Wm ) \Bigg ) \Jm, \label{eq:Hessgam} \end{align} where $\Wm=\wv \wv^T, \Qm=\qv \qv^T $ and \begin{align} \wv^T= \frac{1}{{\sqrt{ \Var_{Q_{\mu^*}} \Big (\log \frac{Q_{\mu^*}}{P_0} \Big ) }}} \Bigg ( Q_{\mu^*}(1) {\log\Big ( \frac{Q_{\mu^*}(1)}{P_0(1)} \Big )-E_0 }, \ldots, Q_{\mu^*}(|\Xc|){\log\Big ( \frac{Q_{\mu^*}(|\Xc|)}{P_0(|\Xc|)} \Big )-E_0 } \Bigg), \end{align} \begin{align} \qv= \Big (Q_{\mu^*} (1),\ldots, Q_{\mu^*}(|\Xc|)\Big), \end{align} \begin{align} \Tm= \diag \Big(Q_{\mu^*} \Big), \quad \Jm=\diag \Bigg(\frac{1}{Q_1} \Bigg), \end{align} and $Q_{\mu^*}, \mu^*$ are defined as in \eqref{eq:tiltedmu} and \eqref{eq:KKTgamma}. \end{lemma} \begin{proof} The proof can be found in Appendix \ref{sec:prooflemd1}. \end{proof} Next using the derivatives of $\gamma (E_0,Q_1)$, and \eqref{eq:firstopt} we can find the derivatives of the exponent function $E_1(E_0,Q_1)$. \begin{lemma}\label{lem:HessianE} For every $P_0, P_1\in \Pc(\Xc)$, the gradient of the exponent function $E_1(E_0,Q_1)$ evaluated at $Q_1=P_1$ is \begin{align}\label{eq:zeroder} \frac{\partial E_1(E_0,Q_1)}{\partial Q_1(i)} \Bigg|_{Q_1=P_1} = 0. \end{align} Furthermore, the Hessian matrix evaluated at $Q_1=P_1$ is \begin{align}\label{eq:Hessian} \Hm_2 &\triangleq \nabla^2_{Q_1} E_1(E_0,Q_1) \Big |_{Q_1=P_1}\\ &= \beta \eta^*_\beta \Jm \Big [ \Qm + \eta^*_1 \Vm +(1-\eta^*_1)\Wm - \Tm \Big ] \Jm \end{align} where $\eta^*_\beta$ and $\eta^*_1$ are the optimal Lagrange multipliers in \eqref{eq:lagrangeE} when $0<\beta\leq 1$ and $\beta=1$ respectively, and $\Vm=\vv \vv^T$, \begin{align} \vv=\frac{1}{\sqrt{\Var{Q_{\eta^*_\beta} }(\Omega) }} \Big ( Q_{\eta^*_\beta}(1)\big ( \Omega(1) -(E_0- \beta E_1)\big ),\ldots,Q_{\eta^*_\beta}(|\Xc|)\big (\Omega(|\Xc|)-(E_0- \beta E_1)\big) \Big ), \end{align} \begin{align} \Omega(i)=\beta \log \frac{P_1(i)}{P_0(i)} +(1-\beta) \log \frac{Q_{\eta^*_\beta}(i)}{P_0(i)}, i \in \Xc. \end{align} \end{lemma} \begin{proof} The proof can be found in Appendix \ref{sec:prooflemd2}. \end{proof} Lemma \ref{lem:HessianE} shows that the gradient of $E_1(E_0,Q_1)$ respect to $Q_1$ is zero at $Q_1=P_1$ which is due to the choice of the threshold function $\gamma(E_0,Q_1)$. Note that in the case of the plugin classifier with the fixed threshold the gradient is non zero which results in $\frac{\partial E_1(E_0,r)}{\partial r} \bigg|_{r=0}= \infty$. Next by using a Taylor series expansion of $E_1(E_0,Q_1)$ and the relative entropy in \eqref{eq:limitE} we can get the following lemma. \begin{lemma}\label{lem:approx} For any $P_0, P_1 \in \Pc(\Xc)$, the approximation \begin{align}\label{eq:worstapproxopt} E_1(E_0,r) = E^*_1(E_0)+ \min_{\substack{ \frac{1}{2} \thetav_{P_1}^T \Jm \thetav_{P_1} \leq r \\ \onev^T\thetav_{P_1}=0 } } \frac{1}{2}\thetav_{P_1}^{T} \Hm_2 \thetav_{P_1} + o(r), \end{align} where \begin{align}\label{eq:diffQ1} \thetav_{P_1}&= \Big(Q_1(1)-P_1(1),\dotsc,Q_1(|\Xc|)-P_1(|\Xc|)\Big)^T, \end{align} holds. \end{lemma} \begin{proof} The proof can be found in Appendix \ref{sec:prooflemd3}. \end{proof} We can further simplify the convex optimization problem in \eqref{eq:worstapproxopt}, using the following lemma. \begin{lemma}\label{lem:optwo} For $\Jm, \Hm_2$ defined in \eqref{eq:fisher}, \eqref{eq:Hessian}, we have \begin{equation}\label{eq:simp} \min_{\substack{ \frac{1}{2} \thetav_{P_1}^T \Jm \thetav_{P_1} \leq r \\ \onev^T\thetav_{P_1}=0 } } \frac{1}{2}\thetav_{P_1}^{T} \Hm_2 \thetav_{P_1} = \min_{\substack{ \frac{1}{2} \psiv_{P_1}^T \psiv_{P_1} \leq r } } \frac{1}{2}\psiv_{P_1}^{T} \Hm \psiv_{P_1}, \end{equation} where \begin{equation} \Hm=\ \beta \eta^*_\beta \sqrt{\Jm} \Big [ \Qm+ \eta^*_1 \Vm +(1-\eta^*_1)\Wm -\Tm \Big ] \sqrt{\Jm}. \end{equation} \end{lemma} \begin{proof} The proof can be found in Appendix \ref{sec:prooflemd4}. \end{proof} Finally, by Lemma \ref{lem:approx} and \ref{lem:optwo} we only need to solve the optimization in \eqref{eq:simp}. Assuming $\Hm$ has negative eigenvalues which is the case for $|\Xc| \geq 4$ by Weyl's inequality \cite{Johnson}, it can be shown that at the optimal solution $\psiv_{P_1}^*$, the inequality constraint of optimization \eqref{eq:simp} is satisfied with equality, hence \begin{align} \min_{\substack{ \frac{1}{2} \psiv_{P_1}^T \psiv_{P_1} \leq r } } \frac{1}{2}\psiv_{P_1}^{T} \Hm \psiv_{P_1} &= r \min_{\substack{ \frac{1}{2} \psiv_{P_1}^T \psiv_{P_1} \leq r } } \frac{1}{2}\frac{\psiv_{P_1}^{T} \Hm \psiv_{P_1}}{\psiv_{P_1}^T \psiv_{P_1}} \\ &= \Lambda_{\min}(\Hm) r \end{align} where $\Lambda_{\min}(\Hm) $ is the smallest eigenvalue of $\Hm$. Taking the derivative of \eqref{eq:worstapproxopt} and setting $r=0$, we obtain \begin{equation} \underline{\alpha}_\beta= - \Lambda_{\min}(\Hm). \end{equation} Further upper bounding the $\Lambda_{\min}(\Hm)$ using Weyl's inequality \cite{Johnson}, we get \begin{align} -\Lambda_{\min}(\Hm) \geq \beta \eta^*_\beta \Bigg [\frac{Q_{\eta^*_\beta}}{P_1} \Bigg ] _{(3)} \end{align} where we have used the fact that $\Qm+ \eta^*_1 \Vm +(1-\eta^*_1)\Wm$ can only have three nonzero eigenvalues as $\Qm,\Vm,\Wm$ are rank one matrices. \section{Proof of Theorem \ref{thm:upper}} \label{sec:proofth7} By Lemma \ref{lem:inclusion} we can see that for every $\beta_1\leq \beta_2$ we have $\Ac_0(Q_1,\beta_1) \subseteq \Ac_0(Q_1,\beta_2)$, and hence $E_1(E_0,Q_1)$ is non-increasing in $\beta$. Therefore, to find an upper bound for $\alpha^*_\beta$ we can find an upper bound to $\alpha^*=\alpha^*_{\beta=1}$ and that would be an upper bound to any test with $\beta <1$. Hence we let $\beta=1$ in the rest of the proof. To upper bound $\alpha^*$ in \eqref{eq:optimalalpha} we first find a lower bound to $E_1(E_0,r)$. Rewriting $E_1(E_0,Q_1)$ in the dual form we have \cite{BoroumandTran} \begin{align} E_1(E_0,Q_1) =& \max_{\nu\geq 0} -\nu \gamma(E_0,Q_1) - \log \sum_{a \in \mathcal{X} } P_{0}^{\nu}(a) P_1(a) Q_{1}^{-\nu}(a). \end{align} Setting $\nu=\lambda^*$ where $\lambda^*$ is the optimal Lagrange multiplier in \eqref{eq:tilted} when $P_1$ is known in the test. We get \begin{align}\label{eq:lowerEQ} E_1(E_0,Q_1) \geq -\lambda^* \gamma(E_0,Q_1) - \log \sum_{a \in \mathcal{X} } P_{0}^{\lambda^*}(a) P_1(a) Q_{1}^{-\lambda^*}(a). \end{align} By the Taylor remainder theorem, for every $Q_1 \in \mathcal{B}(P_1,r)$ where $r<r_c$, we have \begin{align} E_1(E_0,Q_1) \geq E_1(E_0,P_1) +\frac{1}{2} \thetav_{P_1}^T \Sigm(\tilde{Q}_1) \thetav_{P_1}, \end{align} where \begin{align} \Sigm({Q}_1)&=-\lambda^* \Hm(Q_1) - \lambda^*(1+\lambda^*) \diag \Big(\frac{\hat{Q}_{\lambda^*}}{Q_1^2} \Big)+(\lambda^*)^2 \uv^T \uv,\\ \Hm(Q_1) & = \nabla_{Q_1}^2 \gamma(E_0,Q), \\ \uv&= \Bigg ( \frac{\hat{Q}_{\lambda^*}(1)}{Q_1(1)}, \ldots, \frac{\hat{Q}_{\lambda^*}(k)}{Q_1(k)} \Bigg)^T,\\ \hat{Q}_{\lambda^*}(x)&= \frac{ P_{0}^{\lambda^*}(x) P_1(x) Q_1^{-\lambda^*}(x) } {\sum_{a \in \Xc } P_{0}^{\lambda^*}(a) P_1(a) Q_1^{-\lambda^*}(a) }, \end{align} evaluated at some $\tilde{Q}_1 \in \mathcal{B}(P_1,r)$, and $\thetav_{P_1}$ is difference vector $Q_1-P_1$ defined in \eqref{eq:diffQ1}. $\Hm(Q_1)$ is the Hessian matrix of the threshold function in \eqref{eq:Hessgam}. Note that we used the fact that the gradient of RHS in \eqref{eq:lowerEQ} evaluated at $Q_1=P_1$ is equal to zero, hence we only need to control the second order term. Letting \begin{equation} \Lambda (\Sigm,r)=\min_{\tilde{Q}_1 \in \Bc (P_1,r) } \Lambda_{\text{min}} \big(\Sigm(\tilde{Q}_1)\big ), \end{equation} where $ \Lambda_{\text{min}} \big(\Sigm(\tilde{Q}_1)\big ) $ is the smallest eigenvalue of $ \Sigm(\tilde{Q}_1)$. We can further lower bound $E_1(E_0,Q_1)$ by \begin{align}\label{eq:lowerE} E_1(E_0,Q_1) \geq E_1(E_0,P_1) +\frac{\Lambda(\Sigm,r)}{2} \thetav_{P_1}^T \thetav_{P_1}, \end{align} Substituting \eqref{eq:lowerE} in \eqref{eq:limitE} we get \begin{equation}\label{eq:Elowerapprox} E_1(E_0,r) \geq \min_{\substack{D(Q_1\|P_1)\leq r \\ Q_1 \in \Pc(\Xc) } } E_1(E_0,P_1) +\frac{\Lambda(\Sigm,r)}{2} \thetav_{P_1}^T \thetav_{P_1}. \end{equation} Similarly to $E_1(E_0,Q_1)$, by using a Taylor series expansion of $D(Q_1\|P_1)$ around $Q_1=P_1$ and using the remainder theorem, for every $Q_1 \in \mathcal{B}(P_1,r)$ we have \begin{align} D(Q_1\|P_1) &\geq \frac{\Lambda(\Jm,r)}{2} \thetav_{P_1}^T \thetav_{P_1},\label{eq:lowerkl} \\ \Jm(\tilde{Q}_1)&=\diag\bigg( \frac{1}{\tilde{Q}_1(1)},\dotsc,\frac{1}{\tilde{Q}_1(|\Xc|)}\bigg). \end{align} Therefore, by \eqref{eq:Elowerapprox}, and \eqref{eq:lowerkl} we get \begin{equation} E_1(E_0,r) \geq E_1(E_0,P_1) +\frac{\Lambda(\Sigm,r)}{\Lambda(\Jm,r)}r. \end{equation} By Weyl's inequality \cite{Johnson} we have \begin{align} \Lambda(\Sigm,r)\geq& \min_{\tilde{Q}_1 \in \Bc (P_1,r) } -\lambda^* \Bigg( \frac{\mu^*}{1+\mu^*} \Bigg[\frac{Q_{\mu^*}}{\tilde{Q}_1^2} \Bigg]_{(1)}+ \frac{1}{(1+\mu^*)} \big(\|\wv_{\tilde{Q}_1}\|_2^2 +\|\qv_{\tilde{Q}_1}\|_2^2 \big) \Bigg) -\lambda^*(1+\lambda^*) \Bigg [\frac{Q_{\lambda^*}}{\tilde{Q}_1^2} \Bigg]_{(1)}\\ \geq & -\lambda^*(4+\lambda^* )\Bigg( \frac{1}{{Q}_{1}^{\text{min}}(r)} \Bigg ) ^2 \end{align} where \begin{equation} Q_1^{\text{min}}(r) =\min_{\substack{ \tilde{Q}_1 \in \Bc(P_1,r) \\ a\in \Xc }} \tilde{Q}_1(a), \end{equation} and we used $Q_{\lambda^*}(a), Q_{\mu^*}(a) \leq 1$ for all $a \in \Xc $ and $\mu^*\geq0$. Finally, by $\Lambda(\Jm,r) \geq 1$ we get \begin{equation} E_1(E_0,r) \geq E_1(E_0,P_1) -\lambda^*(4+\lambda^* )\bigg( \frac{1}{{Q}_{1}^{\text{min}}(r)} \bigg ) ^2 r. \end{equation} By the Pinsker's inequality \cite{Cover}, we have $\|\thetav_{P_1}\|_1 \leq \sqrt{2r}$, hence $Q_1^{\text{min}} \geq P_1^{\text{min}} -\frac{\sqrt{2r}}{2}$, \begin{equation}\label{eq:upperalpha} E_1(E_0,r) \geq E_1(E_0,P_1) -\lambda^*(4+\lambda^* )\Bigg( \frac{\sqrt{r}}{ P_1^{\text{min}} -\sqrt{r} } \Bigg ) ^2 . \end{equation} Since $E_1(E_0,r) \geq 0$ for all $r\geq0$ we can improve the lower bound to \begin{equation}\label{eq:upperalpha} E_1(E_0,r) \geq \Bigg \{ E_1(E_0,P_1) -\lambda^*(4+\lambda^* )\Bigg( \frac{\sqrt{r}}{ P_1^{\text{min}} -\sqrt{r} } \Bigg ) ^2 \Bigg \} \mathbbm{1} \{ r \leq \bar{r} \} . \end{equation} where \begin{equation} \bar{r}= \Bigg(\frac{\kappa P_1^{\text{min}}}{1+\kappa} \Bigg )^2. \end{equation} Next by \eqref{eq:optimalalpha}, \eqref{eq:upperalpha}, we have \begin{align} \alpha^*_\beta &\leq \max \Bigg \{ \max_{0 \leq r \leq r_c} \lambda^*(4+\lambda^* )\Bigg( \frac{\mathbbm{1} \{ r \leq \bar{r} \} }{ P_1^{\text{min}} -\sqrt{r} } \Bigg ) ^2 , \frac{E_1(E_0,P_1)}{r_c} \Bigg \} \\ &= \frac{ \lambda^*(4+\lambda^* ) }{ (P_1^{\text{min}} -\sqrt{\bar{r}} )^2 } \\ &= \frac{\lambda^*(4+\lambda^*)(1+\kappa) }{(P_1^{\text{min}})^2} \end{align} where we have dropped the second term in the maximization as it is derived by letting $E_1(E_0,r)$ to be a straight line from $E_1(E_0,r=0)$ at $r=0$ to $0$ at $r=r_c$ and the first term in the maximization is derived by the lower bound to $E_1(E_0,r)$, and $ \bar{r}< r_c $, hence the first term dominates. This concludes the proof. \section{Proof of Lemma \ref{lem:cont}} \label{app:cont} \label{sec:prooflemb1} We need to prove the the continuity of $E_1(E_0,r)$ on $r\in [0,r_c)$. To prove this, we will use the Berge's Maximum Theorem stated below \cite{Berge}. \begin{proposition}[Berge's Maximum Theorem] Let $X\subseteq \mathbb{R}^n$ and $ \Theta \subseteq \R^m$. Let $F: X\times \Theta \rightarrow \mathbb{R}$ be a continuous function, and let $\Gc : \Theta \rightarrow X$ be a compact valued and continuous correspondence. Then the maximum value function \begin{equation} V(\Theta)= \max_{X\in \Gc(\Theta)} F(X,\Theta) \end{equation} is well-defined and continuous, and the optimal policy correspondence \begin{equation} X^*(\Theta) = \{X \in \Gc(\Theta) | F(X,\Theta) = V(\Theta)\} \end{equation} is nonempty, compact valued, and upper hemicontinuous. \end{proposition} We first prove that correspondence $\Gc_1(Q_1) = \{Q \in \mathcal{P} (\Xc): D(Q\|P_0)-D(Q\|Q_1) \geq \gamma(E_0,Q_1)\}$ is a continuous correspondence for $Q_1 \in \mathcal{P}_{\delta} (\Xc)$. Then by Berge's theorem $E_1(E_0,Q_1)$ is continuous on $Q_1$ over $\mathcal{P}_{\delta} (\Xc)$. Next, by showing the continuity of the correspondence $\Gc_2(r) = \{Q_1 \in \mathcal{P}_{\delta} (\Xc): D(Q_1\|P_1) \leq r\} $, we conclude the continuity of $E_1(E_0,r)$ in $r$ by Berge's maximum theorem. So we only need to prove that $\Gc_1, \Gc_2$ are compact and continuous correspondences. Note that $Q$ and $Q_1$ are subsets of $\mathbb{R}^{|\Xc|}$ and also $\Gc_1(Q_1), \Gc_2(Q)$ are both closed and bounded, hence By Heine-Borel theorem \cite{Rudin} both $\Gc_1, \Gc_2$ are compact. To prove the continuity of a correspondence we need to prove the upper and lower hemicontinuity of the correspondence. \begin{definition} The compact valued correspondence $\Gc$ is upper hemicontinuous at $\theta \in \Theta$ if $\Gc(\theta)$ is nonempty and if, for every sequence $\{\theta^{(j)} \}$ with $\theta^{(j)} \rightarrow \theta$ and every sequence $\{X^{(j)} \}$ with $X^{(j)} \in \Gc(\theta^{(j)})$ for all $j$, there exists a convergent subsequence $\{X^{(j_k)}\}$ such that $X^{(j_k)} \rightarrow X \in \Gc(\theta)$. \end{definition} To prove upper hemicontinuity of $\Gc_1(Q_1)$, fix $Q_1 \in \mathcal{P}_{\delta} (\Xc)$ and assume for every $1\leq j$, $Q_1^{(j)} \in \mathcal{P}_{\delta} (\Xc)$ be a sequence converging to $Q_1$. Moreover, assume for every $j$ there exists a $Q^{(j)}$ such that $Q^{(j)} \in \Gc_1(Q_1^{(j)})$ for all $j$. By the definition of convergence in metric spaces, since $Q_1^{(j)} \rightarrow Q_1$, then there exists a closed bounded set $\mathcal{Q}_2\subseteq \mathcal{P}_{\delta}$, such that for all large enough $j$, we have $Q_1^{(j)} \in \mathcal{Q}_2 \subset \mathbb{R}^{|Xc|-1} $. Furthermore, since for every $Q_1^{(j)}$, $\Gc_1(Q_1^{(j)})$ is an intersection of a half space created by the classifier and the probability simplex, $Q^{(j)} \in \Gc_1(Q_1)$ will lie in a closed and bounded subset of $\mathbb{R}^{|\Xc|-1} $. Therefore, for large enough $j$, the tuple $(Q^{(j)},Q_1^{(j)})$ also lies in a closed and bounded subset of $\mathbb{R}^{2|\Xc|-2} $. Finally, by the Bolzano-Weierstrass theorem \cite{Rudin}, each bounded sequence in the Euclidean space $\mathbb{R}^{2|\Xc|-2} $ has a convergent subsequence $(Q^{(j_k)},Q_1^{(j_k)})$ with the limit point $(Q,Q_1)$. Finally, since each element of this convergent subsequence satisfies the constraint in $\Gc_1(Q_1)$, then by showing the continuity of $\gamma(E_0,Q_1)$ in $Q_1$, we can conclude that the limit point will also satisfy $Q \in \Gc_1(Q_1)$ and gives the upper hemicontinuity. To prove the continuity of $\gamma(E_0,Q_1)$, note that $D(P\|Q)$ is continuous in the pair for any $ P \in \mathcal{P}(\Xc)$ and $Q \in \mathcal{P}_{\delta}(\Xc)$ where $\delta >0$. Hence, by the Berge's theorem, $\gamma(E_0,Q_1)$ is continuous in $Q_1\in \Pc_{\delta}(\Xc)$, and the optimizer $Q_{\mu}(Q_1)$ is upper hemicontinuous. Also, since by \eqref{eq:tiltedmu}, $Q_{\mu}(Q_1)$ is single valued , we get that $Q_{\mu}(Q_1)$ is in fact continuous in $Q_1$. \begin{definition} The correspondence $\Gc$ is lower hemicontinuous at $\theta \in \Theta$ if $\Gc(\theta)$ is nonempty and if, for every $X \in \Gc(\theta) $ and every sequence $\{\theta^{(j)} \}$ such that $\theta^{(j)} \rightarrow \theta$, there is a $1\leq J$ and a sequence $\{ X^{(j)}\}$ such that $X^{(j)} \in \Gc(\theta^{(j)})$ for all $J \leq j$ and $X^{(j)} \rightarrow X$. \end{definition} To prove the lower hemicontinuity of $\Gc_1(Q_1)$, let $Q_1 \in \mathcal{P}(\Xc)$ and $Q \in \Gc_1(Q_1)$ be fixed. Also, let $Q_1^{(j)} \in \mathcal{P}(\Xc) $ be a sequence converging to $Q_1$. Next we construct the sequence $Q^{(j)}$ such that for every $j$, $Q^{(j)} \in \Gc_1(Q_1^{(j)})$ and $Q^{(j)} \rightarrow Q$. Note that for every $Q_1^{(j)}$, the correspondence is a half space characterized by the hyperplane \begin{equation} \Hc^{(j)}=\{Q\in\Pc(\Xc): \big(Q-Q_{\mu}(Q_1^{(j)})\big)^T \nv^{(j)}=0\}, \end{equation} where $\nv^{(j)}$ is \begin{equation} \nv^{(j)}= \Bigg(\log \frac{Q_{\mu}(Q_1^{(j)})(1)}{P_0(1)},\ldots, \log \frac{Q_{\mu}(Q_1^{(j)})(|\Xc|)}{P_0(|\Xc|)} \Bigg)^T. \end{equation} Also for every $Q^{(j)}$, define the line passing $P_0$ and intersecting $\Hc^{(j)}$ as \begin{align} L(Q^{(j)})&= \beta P_0 +(1-\beta) Q_{\dagger}^{(j)}, \beta \in [0,1],\\ Q_{\dagger}^{(j)}&= \{P: P=\beta P_0+(1-\beta)Q, \beta\in \mathbb{R} \} \cap \Hc^{j}. \end{align} Moreover, for every $Q$ let $Q=\beta^* P_0 +(1-\beta^*) Q_{\dagger}$. Finally, we can define the sequence $Q^{(j)}$ as \begin{align} Q^{(j)}=\beta^*P_0+ (1-\beta^*) Q_{\dagger}^{(j)}. \end{align} By our construction, it is clear that for every $j$, $Q^{(j)} \in \Gc_1(Q_1^{(j)})$. Moreover, by continuity of $Q_{\mu}(Q_1^{(j)})$, we conclude that $\Hc^{(j)} \rightarrow \Hc$ and consequently $Q_{\dagger}^{(j)} \rightarrow Q_{\dagger} $, and $ Q^{(j)} \rightarrow Q$. Next, we show the continuity of $\Gc(r)$. To show the upper hemicontinuity of the correspondence, we can use the same argument as the upper hemicontinuity of $\Gc_1(Q_1)$. So it only remains to prove the lower hemicontinuity. Let $ 0\leq r$ and $Q_1 \in \Pc_{\delta}(\Xc)$ be fixed. Also, let $0\leq r^{(j)}$ is a sequence converging to $r$. Construct $Q_1^{(j)}\in\Pc_{\delta}(\Xc)$ as \begin{equation} Q^{(j)}_1=\bigg( \frac{r^{(j)}}{r} \bigg) Q_1 + \bigg( 1- \frac{r^{(j)}}{r} \bigg) P_1. \end{equation} It is clear that $Q^{(j)}_1 \rightarrow Q_1$ as $r^{(j)} \rightarrow r$. Also by convexity of KL-divergence we get \begin{equation} D \big(Q^{(j)}_1 \| P_1 \big) \leq \frac{r^{(j)}}{r} D(Q_1\|P_1) \leq r^{(j)} \end{equation} where in the last inequality we used the fact that $Q_1\in \Gc(r)$. Therefore, $Q^{(j)}_1 \in \Gc\big(r^{(j)}\big)$, and that gives the lower hemicontinuity of $\Gc(r)$, and concludes the proof. \section{Proof of Lemma \ref{lem:gamsen}} \label{sec:prooflemd1} By the envelope theorem \cite{Segal}, we can find the partial derivative simply by taking the derivative of Lagrangian and evaluating at its optimal solution. Writing the Lagrangian we have \begin{align}\label{eq:lagrange} L(Q,Q_1, \mu, \nu)= D(Q\|Q_1) + \mu \big( D(Q\|P_0)-E_0\big ) + \nu \Big ( \sum_{x\in \Xc} Q(x)-1\Big ), \end{align} hence \begin{align} \frac{\partial \gamma}{\partial Q_1 (j)} &=\beta \frac{\partial L}{\partial Q_1(j)} \Bigg |_{Q=Q_{\mu^*}}\\ &=-\beta \frac{Q_{\mu^{*}}(j)}{Q_1(j)}. \end{align} Taking the second order partial derivative, we get \begin{align}\label{eq:secder} \frac{\partial^{2} \gamma}{\partial Q_1 (i)Q_1(j)} =\beta \Big (-\frac{1}{Q_1(j)}\frac{\partial Q_{\mu^{*}}(j)}{\partial Q_1(i)} + \frac{Q_{\mu^*}(i)}{Q_1^2(i)} \mathbbm{1}\{i=j\}\Big ). \end{align} Therefore, we need to find the sensitivity of $Q_{\mu^{*}}$ ,the optimal solution of optimization \eqref{eq:clsfix}, to local changes in $Q_1$. \cite{Castil} presents a general approach to find the partial derivative of primal and dual variables solution with respect to any parameter of the optimization problem. For ease of reference, we state this result. The result in \cite{Castil} is more general; however, we only state the version we need in our proof. \begin{proposition}\label{prop:sensi} Consider the following primal non-linear programming problem: \begin{equation} \min_{\substack{\xv: ~g(\xv,\av)\leq 0 \\ ~~ h(\xv,\av)=0 }} z=f(\xv,\av), \end{equation} where $f,g,h: \mathbb{R}^n \times \mathbb{R}^p \rightarrow \mathbb{R}$, and $f,h,g \in C^2$. Let $\mu, \nu$ be the Lagrange multipliers corresponding to inequality and equality constraint. Furthermore, assume at the optimal solution the constraints are active and $\mu^*\neq0$, then \begin{equation} \Bigg (\frac{\partial \xv}{\partial \av} , \frac{\partial \nu }{\partial \av}, \frac{\partial \mu}{\partial \av} , \frac{\partial z}{\partial a} \Bigg )^T= \Um^{-1} \Sm \end{equation} where \begin{equation} \Um= \begin{pmatrix} \Fm_{xx}& \Hm_x& \Gm_x & \bold{0} \\ \Hm_x ^{T} & \bold{0} & \bold{0} & \bold{0} \\ \Gm_x^{T} & \bold{0} & \bold{0} & \bold{0} \\ \Fm_x^{T} & \bold{0} & \bold{0} & -1 \\ \end{pmatrix} , ~~\Sm=- \begin{pmatrix} \Fm_{xa} & \\ \Hm_a ^{T} & \\ \Gm_a ^{T}&\\ \Fm_a^{T} & \end{pmatrix} \end{equation} and $\bold{0}$ is the matrix with entries equal to zero with the corresponding dimensions and the submatrices are defined as \begin{align} \Fm_{xx}&=\nabla_{xx} f(x^*,a) + \mu^* \nabla_{xx} g(x^*,a) +\nu^* \nabla_{xx} h(x^*,a), \nonumber \\ \Fm_{xa}&=\nabla_{xa} f(x^*,a) + \mu^* \nabla_{xa} g(x^*,a) +\nu^* \nabla_{xa} h(x^*,a), \nonumber \\ \end{align} where $(x^*,\mu^*,\nu^*)$ is the optimal primal and dual variable solutions. Moreover, $\Fm_x, \Fm_a, \Gm_x, \Gm_a, \Hm_x, \Hm_a$ are gradient of $f, g, h$ respect to $x, a$ evaluated at $x^*$ respectively. \end{proposition} Using this proposition we can find $\frac{\partial Q_{\mu^{*}}}{\partial Q_1} $ in equation \eqref{eq:secder}. Letting $x=Q$, and $a=Q_1$ we have \begin{align} &\Fm_{QQ}=(1+\mu^*)\diag \bigg ( \frac{1}{Q_{\mu^*}(1)},\ldots,\frac{1}{Q_{\mu^*}(|\Xc|)}\bigg),\\ &\Fm_{QQ_1}=-\diag \Big ( \frac{1}{Q_1(1)},\ldots,\frac{1}{Q_1(|\Xc|)}\Big),\\ &\Fm_{Q}= \Bigg ( 1+\log \frac{Q_{\mu^*}(1)}{Q_1(1)},\ldots, 1+\log \frac{Q_{\mu^*}(|\Xc|)}{Q_1(|\Xc|)} \Bigg)^T,\\ &\Fm_{Q_1}=\Bigg (-\frac{Q_{\mu^*}(1)}{Q_1(1)},\ldots,- \frac{Q_{\mu^*}(|\Xc|)}{Q_1(|\Xc|)} \Bigg)^T,\\ &\Gm_{Q}=\Bigg ( 1+\log \frac{Q_{\mu^*}(1)}{P_0(1)},\ldots, 1+\log \frac{Q_{\mu^*}(|\Xc|)}{P_0(|\Xc|)} \Bigg) ^T,\\ &\Gm_{Q_1}=\big ( 0,\ldots, 0 \big)^T,\\ &\Hm_{Q}=\big ( 1,\ldots, 1 \big)^T,\\ &\Hm_{Q_1}=\big ( 0,\ldots, 0 \big)^T. \end{align} Writing matrix $\Um$ as \begin{equation}\label{} \Um= \begin{pmatrix} \Am^{\tiny (|\Xc| \times |\Xc|)} & \Bm^{\tiny (|\Xc| \times 3)} \\ \Cm^{\tiny (3\times |\Xc|)} & \Dm^{\tiny (3\times3)} \end{pmatrix}, \end{equation} we have \begin{align} \Am=(1+\mu^*)\diag \bigg ( \frac{1}{Q_{\mu^*}(1)},\ldots,\frac{1}{Q_{\mu^*}(|\Xc|)}\bigg), \end{align} \begin{align} \Bm= \begin{pmatrix} 1 & 1+ \log \frac{Q_{\mu^*}(1)}{P_0(1)}& 0 \\ \vdots & \vdots & \vdots\\ 1& 1+ \log \frac{Q_{\mu^*}(|\Xc|)}{P_0(|\Xc|)} & 0\\ \end{pmatrix}, ~~ \Cm= \begin{pmatrix} 1 & \ldots & 1\\ 1+ \log \frac{Q_{\mu^*}(1)}{P_0(1)}& \ldots & 1+ \log \frac{Q_{\mu^*}(|\Xc|)}{P_0(|\Xc|)} \\ 1+ \log \frac{Q_{\mu^*}(1)}{Q_1(1)}& \ldots & 1+ \log \frac{Q_{\mu^*}(|\Xc|)}{Q_1(|\Xc|)} \end{pmatrix}, ~~\Dm= \begin{pmatrix} 0 & 0 & 0\\ 0 & 0 & 0\\ 0 & 0 & -1 \end{pmatrix}. \end{align} Also writing $\Sm$ as \begin{equation} \Sm= \begin{pmatrix} \Fm ^{\tiny(|\Xc|\times |\Xc|)}\\ \Km^{\tiny (3 \times |\Xc|)} \end{pmatrix}, \end{equation} we have \begin{align} &\Fm= \diag \bigg ( \frac{1}{Q_1(1)},\ldots,\frac{1}{Q_1(|\Xc|)}\bigg),\\ &\Km= \begin{pmatrix} 0& \ldots & 0\\ 0& \ldots & 0\\ \frac{Q_{\mu^*}(1)}{Q_1(1)} & \ldots & \frac{Q_{\mu^*}(|\Xc|)}{Q_1(|\Xc|)} \end{pmatrix}. \end{align} By the blockwise inversion formula, we have \begin{align}\label{eq:blockinv} &\Um^{-1}= \begin{pmatrix} \Am^{-1} + \Am^{-1} \Bm \Mm \Cm \Am^{-1} & -\Am^{-1} \Bm \Mm \\ - \Mm \Cm \Am^{-1} & \Mm \end{pmatrix}, \end{align} where $\Mm= \big( \Dm -\Cm\Am^{-1} \Bm \big)^{-1}$. Since we are only interested in $\frac{\partial Q_{\mu^{*}}}{\partial Q_1} $ it suffices to find the first $|\Xc|$ rows of $\Um^{-1}$. Applying the block inversion formula to find $\Mm$ we get \begin{align} \Bm \Mm = \frac{1+\mu^*}{\Var_{Q_{\mu^*}} \big (\log \frac{Q_{\mu^*}}{P_1} \big )} \footnotesize\begin{pmatrix} (1+E_0) \big ( \log \frac{Q_{\mu^*}(1)}{P_0(1)} -E_0 \big )- \Var_{Q_{\mu^*}} \big (\log \frac{Q_{\mu^*}}{P_0} \big ), & - \big ( \log \frac{Q_{\mu^*}(1)}{P_0(1)} -E_0 \big ),& 0\\ \vdots & \vdots & \vdots\\ (1+E_0) \big ( \log \frac{Q_{\mu^*}(|\Xc|)}{P_0(|\Xc|)} -E_0 \big )- \Var_{Q_{\mu^*}} \big (\log \frac{Q_{\mu^*}}{P_0} \big ), & -\big ( \log \frac{Q_{\mu^*}(|\Xc|)}{P_0(|\Xc|)} -E_0 \big ),& 0 \end{pmatrix} \end{align} where we used that at the optimal solution $D(Q_{\mu^*}\|P_0)=E_0$. Furthermore, we get \begin{align} \Am^{-1} \Bm \Mm \Cm \Am^{-1} = -\frac{1}{(1+\mu^*)} (\Qm+\Wm). \end{align} Finally, by the structure of $\Km$, and $\Bm \Mm$ we can easily see that \begin{align}\label{eq:Jacob} -\diag \Big(\frac{1}{Q_1}\Big) *\Big ( \frac{\partial {Q}_{\mu^{*}}}{\partial Q_1} \Big) \Bigg |_{Q_1=P_1} &=- \Fm\big (\Am^{-1} + \Am^{-1} \Bm \Mm \Cm \Am^{-1} \big ) \Fm\\ &=\frac{1}{1+\mu^*} \Jm \Bigg [ \Tm - (\Qm+\Wm) \Bigg]\Jm. \end{align} Substituting \eqref{eq:Jacob} into \eqref{eq:secder} we get the Hessian matrix in \eqref{eq:Hessgam}. \section{Proof of Lemma \ref{lem:HessianE}} \label{sec:prooflemd2} Similar to Lemma \ref{lem:gamsen} by writing the Lagrangian and using the envelope theorem \cite{Segal} we have \begin{align}\label{eq:lagrangeE} L(Q,Q_1, \eta, \nu)= D(Q\|P_1) + \eta \big( D(Q\|P_0)-\beta D(Q\|Q_1) + \gamma(E_0,Q_1) \big )+ \nu \Big ( \sum_{x\in \Xc} Q(x)-1\Big ), \end{align} hence \begin{equation} \frac{\partial E_1(E_0,Q_1)}{\partial Q_1 (j)} =\frac{\partial L}{\partial Q_1(j)} \Bigg |_{Q=\hat{Q}_{\eta^*}}= {\eta^*}\Bigg( \beta \frac{\hat{Q}_{{\eta^*}}(j)}{Q_1(j)} +\frac{\partial \gamma}{\partial Q_1 (j)} \Bigg),\label{eq:firstord} \end{equation} \begin{equation}\label{eq:tiltedMM1} \hat{Q}_{\eta^*}(x)= \frac{ P_1^{\frac{1}{1+\eta^*-\eta^* \beta}}(x) Q_{1}^{\frac{-\eta^* \beta }{1+\eta^*-\eta^* \beta}}(x) P_{0}^{\frac{\eta^*}{1+\eta^*-\eta^* \beta}}(x) } {\sum_{a \in \Xc } P_1^{\frac{1}{1+\eta^*-\eta^* \beta}}(a) Q_{1}^{\frac{-\eta^* \beta }{1+\eta^*-\eta^* \beta}}(a) P_{0}^{\frac{\eta^*}{1+\eta^*-\eta^* \beta}}(a) }. \end{equation} Letting $Q_1=P_1$, we get \begin{equation}\label{eq:tiltbeta} {Q}_{\eta^*_\beta}(x)= \frac{ P_{1}^{\frac{1-\eta^*_\beta \beta }{1+\eta^*_\beta-\eta^*_\beta \beta}}(x) P_{0}^{\frac{\eta^*_\beta}{1+\eta^*_\beta-\eta^*_\beta \beta}}(x) } {\sum_{a \in \Xc } P_{1}^{\frac{1-\eta^*_\beta \beta }{1+\eta^*_\beta-\eta^*_\beta \beta}}(a) P_{0}^{\frac{\eta^*_\beta}{1+\eta^*_\beta-\eta^*_\beta \beta}}(a) }. \end{equation} Furthermore, since $D(Q_{\eta^*_\beta}\|P_0)=E_0$, it is easy to see that, ${Q}_{\eta^*_\beta}= Q_{\lambda^*}$ in \eqref{eq:tilted}, and $\frac{\eta^*_\beta}{1+\eta^*_\beta-\eta^*_\beta \beta}=\lambda^*$. Also, by \eqref{eq:KKTgammaLRT} ${Q}_{\eta^*_\beta}= Q_{\mu^*}$, and finally from \eqref{eq:firstdergam} we conclude that \begin{align} \Big( \beta \frac{\hat{Q}_{\eta^*}(j)}{Q_1(j)} +\frac{\partial \gamma}{\partial Q_1 (j)} \Big) \bigg|_{Q_1=P_1}&=\beta \frac{{Q}_{\eta^*_\beta}(j)}{P_1(j)} -\beta \frac{Q_{\mu^*}(j)}{P_1(j)} \\ &=0, \end{align} which concludes \eqref{eq:zeroder}. Next, by taking the second derivative of \eqref{eq:firstord} we get \begin{align}\label{eq:secordE} \frac{1}{\beta} \frac{\partial^2 E_1(E_0,Q_1)}{\partial Q_1(i) Q_1(j)} &= \frac{\partial {\eta^*} }{\partial Q_1(i) } \bigg( \frac{\hat{Q}_{\eta^*}(j)}{Q_1(j)} + \frac{1}{\beta} \frac{\partial \gamma}{\partial Q_1 (j)} \bigg)\\ &+ {\eta^*} \bigg( \frac{1 }{Q_1(j)} \frac{\partial \hat{Q}_{\eta^*}(j)}{\partial Q_1(i)} - \frac{\hat{Q}_{\eta^*}(i)}{Q_1^2(i)} \mathbbm{1}\{i=j\} + \frac{1}{\beta} \frac{\partial^2 \gamma}{\partial Q_1(i) Q_1 (j)} \bigg). \end{align} Letting $Q_1=P_1$ the first term is zero, hence we only require to find $\frac{\partial \hat{Q}_{\eta^*}(j)}{\partial Q_1(i)} \Big|_{Q_1=P_1}$. For simplicity in calculations, we write all the expressions in term of the optimal Lagrange multipliers when $\beta=1$. Since for every $0<\beta \leq 1$ the optimizing distribution $Q$ when $Q_1=P_1$ is the tilted distribution of $P_0$, and $P_1$ and since for every such $Q$, the optimizing distribution should satisfy the condition \eqref{eq:KKTgammaLRT}, hence the tilted exponent is equal for every $\beta$ and by equating the exponents in \eqref{eq:tiltbeta} we can define \begin{equation} \rho \triangleq \frac{\eta^*_\beta}{\eta^*_1}=1+\eta^*_\beta-\beta \eta^*_\beta, \end{equation} where $\eta^*_\beta$ is the Lagrange multiplier for aribitrary $\beta$ and $\eta^*_1$ is the Lagrange multiplier when $\beta=1$. Using proposition \ref{prop:sensi} and letting $x=Q$, and $a=Q_1$, and setting $Q_1=P_1$, we have \begin{align} &\Fm_{QQ}^\beta= \rho \Fm_{QQ} , \quad \Fm_{QQ_1}^\beta= \rho \Fm_{QQ_1} , \\ & \Fm_{Q}^\beta = \Fm_{Q}, \quad \Fm_{Q_1}^\beta = \Bm_Q, \\ &\Gm_{Q}^\beta = \beta \Gm_{Q} + (1-\beta) \Fm_{Q} , \quad \Gm_{Q_1}^\beta =\Gm_{Q_1}, \\ &\Hm_{Q}^\beta=\Hm_{Q} \quad \Hm_{Q_1}^\beta=\Hm_{Q}, \end{align} where \begin{align} \Fm_{QQ}&= \diag \Bigg ( \frac{1}{Q_{\eta^*_1}(1)},\ldots,\frac{1}{Q_{\eta^*_1}(|\Xc|)}\Bigg),\\ \Fm_{QQ_1}&= \lambda^*_1 \diag \Big ( \frac{1}{P_1(1)},\ldots,\frac{1}{P_1(|\Xc|)}\Big),\\ \Fm_{Q}&= \Bigg ( 1+\log \frac{Q_{\eta^*_1}(1)}{P_1(1)},\ldots, 1+\log \frac{Q_{\eta^*_1}(|\Xc|)}{P_1(|\Xc|)} \Bigg)^T,\\ \Fm_{Q_1}&= \big ( 0,\ldots, 0 \big)^T, \\ \Bm_{Q}&=\Bigg ( 1+\log \frac{Q_{\eta^*_1}(1)}{P_0(1)},\ldots, 1+\log \frac{Q_{\eta^*_1}(|\Xc|)}{P_0(|\Xc|)} \Bigg)^T,\\ \Gm_{Q}&=\Bigg ( \log \frac{P_1(1)}{P_0(1)},\ldots, \log \frac{P_1(|\Xc|)}{P_0(|\Xc|)} \Bigg) ^T,\\ \Gm_{Q_1}&= \Bigg ( \bigg( \frac{\hat{Q}_{\eta^*_1}(1)}{Q_1(1)} +\frac{\partial \gamma}{\partial Q_1 (1)} \bigg) \Bigg|_{Q_1=P_1} ,\ldots, \bigg( \frac{\hat{Q}_{\eta^*_1}(|\Xc|)}{Q_1(|\Xc|)} +\frac{\partial \gamma}{\partial Q_1 (|\Xc|)} \bigg) \Bigg|_{Q_1=P_1} \Bigg)^T \\ & =\big ( 0,\ldots, 0 \big)^T,\\ \Hm_{Q}&=\big ( 1,\ldots, 1 \big)^T,\\ \Hm_{Q_1}&=\big ( 0,\ldots, 0 \big)^T. \end{align} Similar to the Lemma \ref{lem:gamsen} we can write $\Um$ as \begin{equation}\label{} \Um_\beta = \begin{pmatrix} \Am^{\tiny (|\Xc| \times |\Xc|)}_\beta & \Bm^{\tiny (|\Xc| \times 3)} _\beta \\ \Cm^{\tiny (3\times |\Xc|)}_\beta & \Dm^{\tiny (3\times3)}_\beta \end{pmatrix}, \end{equation} where \begin{align} \Am_\beta&=\rho \Am_1\\ \Bm_\beta&=\beta \Bm_1 + (1-\beta) \Em_0\\ \Cm_\beta &= \beta \Cm_1 +(1-\beta) \Lm_0\\ \Dm_\beta&=\Dm_1 \end{align} and \begin{align} \Am_1=\diag \Bigg ( \frac{1}{Q_{\eta^*_1}(1)},\ldots,\frac{1}{Q_{\eta^*_1}(|\Xc|)}\Bigg), \quad \Dm_1= \begin{pmatrix} 0 & 0 & 0\\ 0 & 0 & 0\\ 0 & 0 & -1 \end{pmatrix}, \end{align} \begin{align} \Bm_1= \begin{pmatrix} 1 & \log \frac{P_1(1)}{P_0(1)}& 0 \\ \vdots & \vdots & \vdots\\ 1& \log \frac{P_1(|\Xc|)}{P_0(|\Xc|)} & 0\\ \end{pmatrix}, \quad \Em_0= \begin{pmatrix} 1 &1+ \log \frac{Q_{\eta_1^*} (1)}{P_0(1)}& 0 \\ \vdots & \vdots & \vdots\\ 1& 1+\log \frac{Q_{\eta_1^*} }{P_0(|\Xc|)} & 0\\ \end{pmatrix}, \end{align} \begin{align} \Cm_1= \begin{pmatrix} 1 & \ldots & 1\\ \log \frac{P_1(1)}{P_0(1)}& \ldots & \log \frac{P_1(|\Xc|)}{P_0(|\Xc|)} \\ 1+ \log \frac{Q_{\eta_1^*}(1)}{P_1(1)}& \ldots & 1+ \log \frac{Q_{\eta_1^*}(|\Xc|)}{P_1(|\Xc|)} \end{pmatrix}, \quad \Lm_0= \begin{pmatrix} 1 & \ldots & 1\\ 1+ \log \frac{ Q_{\eta^*_1}(1) }{P_0(1)}& \ldots & 1+\log \frac{Q_{\eta^*_1}(1)}{P_0(|\Xc|)} \\ 1+ \log \frac{Q_{\eta_1^*}(1)}{P_1(1)}& \ldots & 1+ \log \frac{Q_{\eta_1^*}(|\Xc|)}{P_1(|\Xc|)} \end{pmatrix}, \end{align} Also writing $\Sm$ as \begin{equation} \Sm= \begin{pmatrix} -\Fm_{QQ_1}^\beta \\ \Km \end{pmatrix}, \end{equation} we have \begin{align} \Km^{\tiny (3 \times |\Xc|)}&= \begin{pmatrix} 0& \ldots & 0\\ 0& \ldots & 0\\ 0& \ldots & 0 \end{pmatrix}. \end{align} By block inversion formula in \eqref{eq:blockinv}, and similar arguments as in Lemma \ref{lem:gamsen} we have \begin{align} \Am_\beta ^{-1} \Bm_\beta \big( \Dm_\beta -\Cm_\beta \Am_\beta^{-1} \Bm_\beta \big)^{-1} \Cm_\beta \Am_\beta^{-1} = -\frac{1}{\rho} \big (\Tm^2 +\Vm \big ), \end{align} where $\Vm=\vv \vv^T$ and \begin{align} \vv=\frac{1}{\sqrt{\Var{Q_{\eta^*_\beta} }(\Omega) }} \Big ( Q_{\eta^*_\beta}(1)\Omega(1),\ldots,Q_{\eta^*_\beta}(|\Xc|)\Omega(|\Xc|) \Big ), \end{align} \begin{align} \Omega(i)=\beta \log \frac{P_1(i)}{P_0(i)} +(1-\beta) \log \frac{Q_{\eta^*_\beta}(i)}{P_0(i)}, i \in \Xc, \end{align} Finally, by the structure of $\Sm$, we can easily derive \begin{align}\label{eq:Jacob2} \diag \Big(\frac{1}{P_1}\Big) *\Big ( \frac{\partial \hat{Q}_{\eta^{*}}}{\partial Q_1} \Big) \Bigg |_{Q_1=P_1} &=\Jm \Big [ -\Tm + (\Tm^2+\Vm) \Big]\Jm. \end{align} Substituting \eqref{eq:Jacob2} and \eqref{eq:Hessgam} into \eqref{eq:secordE}, we get \begin{align} \Hm_2= \rho \eta^*_1 \Jm \Big [ -\Tm + \Qm+ \eta^*_1 \Vm +(1-\eta^*_1)\Wm \Big ] \Jm, \end{align} where we have used the identity $\eta_1^*=\frac{\mu^*}{1+\mu^*}$ derived by setting the tilted distribution exponents to be equal for when $Q_1=P_1$. \section{Proof of Lemma \ref{lem:approx}} \label{sec:prooflemd3} By applying a Taylor expansion to $E_1(E_0,Q_1)$ around $Q_1=P_1$ we obtain \begin{align}\label{eq:linearapprox1} E_1(E_0,Q_1)=E_1(E_0,P_1) + \thetav_{P_1}^{T} \nabla E_1(E_0,Q_1)\big|_{Q_1=P_1} + \frac{1}{2}\thetav_{P_1}^{T} \Hm_2 \thetav_{P_1} + o(\| \thetav_{P_2} \|_{\infty}^{2}). \end{align} The first term in the expansion is $E^*_1(E_0)$, also by Lemma \ref{lem:HessianE}, the gradient evaluated at $Q_1=P_1$ is zero. Further approximating the constraint in \eqref{eq:secondopt}, we get \begin{equation}\label{eq:KLapprox} D(Q_1 \| P_1 ) = \frac{1}{2} \thetav_{P_1}^T \Jm \thetav_{P_1} + o (\| \thetav_{P_1} \|^2_{\infty}). \end{equation} By substituting the expansions \eqref{eq:linearapprox1} and \eqref{eq:KLapprox} in \eqref{eq:limitE} we obtain \begin{align}\label{eq:approxworst} E_1 (E_0,r)= E^*_1(E_0)+ \min_{\substack{ \frac{1}{2} \thetav_{P_1}^T \Jm \thetav_{P_1}+ o (\| \thetav_{P_1} \|^2_{\infty}) \leq r \\ \onev^T\thetav_{P_1}=0 } } \frac{1}{2}\thetav_{P_1}^{T} \Hm_2 \thetav_{P_1} + o(\|\thetav_{P_1} \|_{\infty}^2). \end{align} To find the error of above approximation as a function of $r$, first note that we can take $o(\|\thetav_{P_1} \|_{\infty}^2)$ out of minimization and substitute it with $o(\|\thetav^*_{P_1}(r) \|_{\infty}^2)$ where $\thetav^*_{P_1} $ is the optimal solution of the minimization. Moreover, approximating the inequality constraint can result an error of $o(\sqrt{r})$ in $\|\thetav^*_{P_1}\|_{\infty}$. Therefore, by inequality constraint and the fact that it impose a constraint on the length of the vector $\thetav$ we have that $\|\thetav^*_{P_1}\|_{\infty} \leq c\sqrt{r}+o(\sqrt{r})$ where $c$ is independent from $r$ and only depends on $\Jm, \Hm_2$. This argument and \eqref{eq:approxworst} concludes \eqref{eq:worstapproxopt}. \section{Proof of Lemma \ref{lem:optwo}} \label{sec:prooflemd4} Since $\Jm$ is a diagonal matrix with non zero diagonal entries, we have \begin{align} \sqrt{\Jm} &=\diag\bigg( \frac{1}{\sqrt{P_1(1)}},\dotsc,\frac{1}{\sqrt{P_1({|\Xc|})}}\bigg). \end{align} Letting $\psiv_{P_1}= \sqrt{\Jm} \thetav_{P_1}$, we obtain \begin{align} \frac{1}{2}\thetav_{P_1}^{T} \Hm_2 \thetav_{P_1} &= \frac{1}{2}\psiv_{P_1}^{T} \Hm \psiv_{P_1}, \\ \frac{1}{2} \thetav_{P_1}^T \Jm \thetav_{P_1}&= \frac{1}{2} \psiv_{P_1}^T \psiv_{P_1},\\ \onev^T\thetav_{P_1}&= \onev^T \sqrt{\Jm}^{-1} \psiv_{P_1}. \end{align} Next we need to show that we can drop the equality constraint. Note that $\sigv_1 \defeq \frac{\sqrt{\Jm}^{-1} \onev}{\| \sqrt{\Jm}^{-1} \onev\|} $ is in the null space of $\Hm$, i.e., \begin{align} \Hm\sqrt{\Jm}^{-1} \onev = \beta \eta^*_\beta \sqrt{\Jm} \Big [ \Qm+ \eta^*_1 \Vm +(1-\eta^*_1)\Wm -\Tm \Big ] \onev=\zerov \end{align} since $\vv^T \onev=0, \wv^T \onev=0, \Tm \onev=\qv$. Now, assume the optimizer of the second optimization in \eqref{eq:simp} is of the form \begin{equation}\label{eq:sol} \psiv_{P_1}=\rho_1 \sigv_1 + \rho_2 \sigv_2, \end{equation} where $\sigv_1 \perp \sigv_2, \| \sigv_2\|=1$. Then \begin{align} \psiv_{P_1}^{T} \Hm \psiv_{P_1}&= (\rho_1\sigv_1 +\rho_2 \sigv_2)^{T} \Hm (\rho_1\sigv_1 +\rho_2 \sigv_2)\\ &= \rho_2^2 \sigv_2^T \Hm \sigv_2 \end{align} where we also used $ \Hm \sigv_1 =0, \Hm=\Hm^T$. Assuming $\Hm$ has negative eigenvalues (otherwise, both optimizations are equal to zero), the inequality constraint should be satisfied with equality. Then \begin{align} \psiv_{P_1}^{T} \Hm \psiv_{P_1} = \Big (2r-\rho_1^2 \Big) \sigv_2^T \Hm \sigv_2. \end{align} Therefore, to achieve the minimum, $\rho_1$ must be zero, which concludes the proof. \section{Proof of Theorem \ref{thm:Stein}} \label{sec:proofth41} By Theorem \ref{thm:E1}, it can be shown that by taking $E_0=\Theta({n}^{-1})$ there exists an $\alpha$ such that the proposed classifier \eqref{eq:clsfix} achieves $E_1=D(P_0\|P_1)$ which is equal to the Stein regime exponent of hypothesis testing with known distributions, and hence the \eqref{eq:Stein1trade} is equal to $D(P_0\|P_1)$. In fact, there is no need for a training sequence to achieve the \eqref{eq:Stein1trade}. Since Hoeffding's test achieves the optimal error exponent tradeoff only by having the distribution $P_0$, it is easy to see that the Hoeffding test achieves the Stein regime exponent for the unknown distribution for any $P_1$. However, to achieve the largest error exponent under $P_0$ such that it guarantees that for any $P_1$ the type-\RNum{2} probability of error is bounded by some $\epsilon \in (0,1)$, Hoeffding's and our proposed classifier are not universal, since for any choice of threshold $E_0>0$, the type-\RNum{2} probability of error for any distribution $P_1$ such that $D(P_1\|P_0)<E_0$ converges to one. We show that Gutman's universal test \cite{gutman1989asymptotically} achieves the largest type-\RNum{1} error exponent while the type-\RNum{2} probability of error is bounded away from one. Using the Gutman's test we obtain \begin{equation} D_{\alpha}^{\rm GJS} (\Tx\|\TX) \leq \frac{1}{2n} G^{-1}_{ |\Xc |-1}(\epsilon), \end{equation} where \begin{equation} D_{\alpha}^{\rm GJS} (Q\|P) =D\bigg(Q\bigg \| \frac{Q+\alpha P}{1+\alpha}\bigg) + \alpha D\bigg(P \bigg \|\frac{Q+\alpha P}{1+\alpha}\bigg), \end{equation} is the generalized Jensen-Shannon divergence and $ G^{-1}_a(.)$ is the inverse of the complementary CDF of a chi-squared random variable with $a$ degrees of freedom; \cite{Tan} shows that Gutman test with the chosen threshold achieves type-\RNum{2} error probability of $\epsilon$ for any $P_1$ and achieves the type-\RNum{1} error exponent $D_{\frac{\alpha}{1+\alpha}}(P_1\|P_0)$. We prove the converse showing that the achievable error exponent using Gutman's test is, in fact the Stein's regime exponent defined in \eqref{eq:Stein1trade}. Note that as $\alpha \rightarrow \infty $, i.e., when the number of training samples is much larger than the number of test samples, then the exponent $D_{\frac{\alpha}{1+\alpha}} (P_1\|P_0)$ converges to $D(P_1\|P_0)$ which is the Stein regime error exponent under when both distributions are known. We prove the asymptotic optimality of Gutman's test in this setting. Specifically, we show that for any test such that the type-\RNum{1} error exponent $E_0^{(\epsilon)}> D_{\frac{\alpha}{1+\alpha}}(P_1\|P_0) $, there exists a $P_1$ such that type-\RNum{2} probability of error of the test goes to one, i.e., for any test $\phi_n$ such that \begin{equation} \limsup_{n\rightarrow \infty} \epsilon_1 (\phi_n) \leq \epsilon \end{equation} for all $P_1\in \Pc(\Xc)$ then \begin{equation}\label{converse} \lim_{n\rightarrow \infty} E_0^{(\epsilon)}(\phi_n) \leq D_{\frac{\alpha}{1+\alpha}} (P_1\|P_0). \end{equation} First, by the lemma 6 in \cite{Tan}, any optimal test can be converted to a test based on samples type, and the error probabilities of such a test only change by a constant factor. Hence, the converted type-based test is asymptotically optimal, and we can limit the classifiers to type-based ones when studying the error exponent. Next, using the similar idea to \cite{Tan} we show that in order to have a type-\RNum{2} error probability bounded away form one for every $P_1$, the classifier necessarily needs to decide in favor of the second hypothesis when $\| \Tx-\TX\| \leq \delta$, for $\delta>0$, since under $P_1$ the types of the test and training sequence will converge to $P_1$. Also. we only look at deterministic test as it can be shown that randomizing the test cannot increase the error exponents. We have the following lemma. \begin{lemma}\label{lem:concentration} Let $\epsilon \in (0,1)$ and $\X^k, \xv^n$ be two independent i.i.d sequences, generated by distribution $P$. Then \begin{equation} P \Big ( \max_{a \in \mathcal{X}} \Big |\TX(a)- \Tx(a)\Big| \geq \epsilon \Big ) \leq 2 |\mathcal{X}|e^{-n \frac{\epsilon^2}{2}} +2 |\mathcal{X}|e^{-\alpha n \frac{\epsilon^2}{2}} . \end{equation} \end{lemma} \begin{proof} By the triangle inequality, union bound, and the Hoeffding's inequality \cite{gabor}, we have \begin{align} P\Big ( \max_{a \in \mathcal{X}} \big |\TX(a)- \Tx(a) \big | \geq \epsilon \Big ) &=P\big( \max_{a \in \mathcal{X}}\big |\TX(a)-P(a) + P(a)- \Tx(a) \big| \geq \epsilon \Big ) \\ &\leq P\Big ( \max_{a \in \mathcal{X}} \big|\TX(a)-P(a)\big| + \big| \Tx(a)-P(a)\big| \geq \epsilon \Big ) \\ &\leq P \Big( \max_{a \in \mathcal{X}} \big|\TX(a)-P(a)\big| + \max_{a \in \mathcal{X}} \big | \Tx(a)-P(a)\big| \geq \epsilon \Big ) \\ &\leq P \Big( \max_{a \in \mathcal{X}} \big|\TX(a)-P(a)\big| \geq \frac{\epsilon}{2} \cup \max_{a \in \mathcal{X}} \big | \Tx(a)-P(a)\big| \geq \frac{\epsilon}{2} \Big)\\ &\leq P \Big( \max_{a \in \mathcal{X}} \big |\TX(a)-P(a)\big| \geq \frac{\epsilon}{2} \Big ) + P\Big (\max_{a \in \mathcal{X}} \big | \Tx(a)-P(a)\big| \geq \frac{\epsilon}{2} \Big)\\ & \leq \sum_ {a \in \mathcal{X}} P \Big( \big |\TX(a)-P(a)\big | \geq \frac{\epsilon}{2} \Big )+ \sum_ {a \in \mathcal{X}} P\Big (\big | \Tx(a)-P(a)\big| \geq \frac{\epsilon}{2} \Big)\\ & \leq 2 |\mathcal{X}|e^{-2n (\frac{\epsilon}{2})^2} +2 |\mathcal{X}|e^{-2k (\frac{\epsilon}{2})^2} \label{hoef}, \end{align} which concludes the lemma. \end{proof} \begin{lemma}\label{lem:theconv} Let $(Q,Q_1) \in \Pc_{n}(\Xc)\times \Pc_k(\Xc) $ satisfy \begin{equation}\label{eq:lemmacondstein} \max_{a \in \mathcal{X}} \Big \{ \big |Q(a)-Q_1(a) \big | \Big \} \leq \sqrt{\frac{2}{(\alpha\wedge 1)n} \log \frac{4|\Xc|}{1-\epsilon}}. \end{equation} Then for any type based test $\phi_n(\Tx,\TX)$ such that for all distributions $\tilde{P}_1 \in \Pc(\Xc)$, \begin{equation} \epsilon_1\big(\phi_n(\Tx,\TX)\big) < \epsilon, \quad\epsilon \in(0,1), \end{equation} we have $\phi_n(Q,Q_1)=1$. \end{lemma} \begin{proof} We prove this lemma by contradiction. Assume there exists a type dependent test $\phi_n$ such that $\epsilon_1\big(\phi_n(\Tx,\TX)\big) < \epsilon $, and $\phi_n(Q,Q_1)=0$ for types $Q,Q_1$ satisfuing \eqref{eq:lemmacondstein}. We have also used the notation $a \wedge b = \min\{a,b\}$. Then the type-\RNum{2} probability of error for such test is \begin{align} \epsilon_1(\phi_n)&=P_1 ( \phi_n(\Tx,\TX)=0)\\ &\geq P_1( \phi_n(\Tx,\TX)=0 , \Tx=Q, \TX=Q_1). \end{align} Conditioning on training and test sequences, we have \begin{align} \epsilon_1\big(\phi_n)&\geq P_1 ( \phi_n(\Tx,\TX)=0 | \Tx=Q, \TX=Q_1\big) {P}_1 ( \Tx=Q, \TX=Q_1\big)\\ &={P}_1 ( \Tx=Q, \TX=Q_1\big)\\ &\geq \Big(1-\frac{(1-\epsilon)}{2}-\frac{(1-\epsilon)}{2} \Big )\\ &\geq \epsilon \end{align} where in the last step we used the Lemma \ref{lem:concentration}. Therefore, for any probability distribution $P_1$ the type-\RNum{2} error probability exceeds $\epsilon$ which contradicts the assumption. Hence any type based classifier that achieves $\epsilon_1(\phi_n)< \epsilon$ for all distributions of $\tilde{P}_1 \in \Pc(\Xc)$ , should satisfy $\phi(\Tx,\TX)=1$ for sufficiently close types $\Tx,\TX$. \end{proof} Finally by the method of types, we can lower bound the type-\RNum{1} probability of error as \begin{align} \epsilon_0&=P_0 (\phi_n(Q,Q_1)=1)\\ & \geq (n+1)^{-|\mathcal{X}|} (k+1)^{-|\mathcal{X}|} \sum_{\phi_n(Q,Q_1)=1 } e^{-n\big ( D(Q\|P_0) + \alpha D(Q_1\|P_1) \big ) }\\ & \geq (n+1)^{-|\mathcal{X}|} (k+1)^{-|\mathcal{X}|} e^{-n\big ( \min_{\phi_n(Q,Q_1)=1 } D(Q\|P_0) + \alpha D(Q_1\|P_1) \big ) }. \end{align} Therefore for any type based test the type-\RNum{1} error exponent is upper bounded by \begin{equation}\label{eq:convexp} E_0(\phi) \leq \liminf_{n\rightarrow \infty} \min_{\phi_n(Q,Q_1)=1 } D(Q\|P_0) + \alpha D(Q_1\|P_1). \end{equation} Now by Lemma \ref{lem:theconv} for any test with type-\RNum{2} error probability bounded away from one, we have \begin{align} E_0(\phi)& \leq \liminf_{n\rightarrow \infty} \min_{\max_{a \in \mathcal{X}} |Q(a)-Q_1(a)| \leq \sqrt{\frac{2}{(\alpha\wedge 1)n} \log \frac{4|\Xc|}{1-\epsilon}} } D(Q\|P_0) + \alpha D(Q_1\|P_1)\\ & = \lim_{n\rightarrow \infty} \min_{Q \in \Pc (\Xc) } D(Q\|P_0) + \alpha D(Q_1\|P_1) + o(1) \\ &= \min_{Q \in \Pc (\Xc)} D({Q}\|P_0) + \alpha D({Q}\|P_1)\\ &= D_{\frac{\alpha}{1+\alpha}} (P_1\|P_0) \end{align} where the last step follows from e.g. \cite{Harremos}. This concludes the proof. Observe that the proof can be easily generalized for the case where both probability distributions are unknown and only training samples from both are given, i.e., the largest type-\RNum{1} error exponent achievable when the type-\RNum{2} error probability is bounded away from one is also $D_{\frac{\alpha}{1+\alpha}} (P_1\|P_0)$, and Gutman's test can achieve it. \section{Proof of Theorem \ref{thm:seq}} \label{sec:proofth51} We first find the error probabilities as a function of thresholds, and then we find the average stopping time under each hypothesis. The type-\RNum{1} error probability can be upper bounded by \begin{align} \epsilon_0 \leq \sum_{t=n}^{\infty} \PP \Big [ t \big( D(\Tx \| {P}_0)-D(\Tx\| \TX') \big) \geq \got \Big ]. \end{align} By the method of types \cite{Cover}, we have \begin{align}\label{eq:uppersensamp} \epsilon_0 &\leq \sum_{t=n}^{\infty} \sum_{(Q,Q_1) \in {\mathcal{Q}_{01}(t)} \cap \Pc_t(\Xc) \times \Pc_{\alpha t}(\Xc) } e^{ -t \big(D(Q\| P_0)+\alpha D(Q_1\|P_1)\big) }\\ &\leq \sum_{t=n}^{\infty} (\alpha t +1)^{|\mathcal{X}|} (t+1)^{|\mathcal{X}|} e^{- \Etzo}, \end{align} where \begin{equation}\label{eq:Eadverse1} \Etzo=\min_{(Q,Q_1)\in \Qc_{01}(t)} t \big(D(Q\| P_0)+\alpha D(Q_1\|P_1)\big ), \end{equation} \begin{equation}\label{eq:set} {\mathcal{Q}_{01}}(t)=\Bigg \{(Q,Q_1): D(Q\| {P}_0)-D(Q\| Q_1') \geq \frac{\got}{t}, Q_1'=(1-\delta_n)Q_1+\delta_n U \Bigg \}. \end{equation} Similarly to the proof of the Theorem \ref{thm:E1}, we can expand all the exponents defined in this proof around $Q_1$ and by choosing $\delta_n=o(n^{-1})$ the error term of the expansion vanishes as $n$ goes to infinity and we can substitutde $Q_1'$ with $Q_1$ for $Q_1\in \Pc_{\delta_n}(\Xc)$. This is also true for all the exponent function we define in the rest of the proof. For every fixed $Q_1$ we can use the dual form of the optimization \eqref{eq:Eadverse1} over $Q$ to get \cite{BoroumandTran} \begin{align}\label{eq:dualE0} \Etzo=\min_{Q_1\in \Pc_{\delta_n}(\Xc)} \Big ( \max_{\lambda \geq 0} \got \lambda -t \log \sum_{x\in \Xc} P_0^{1-\lambda}(x) Q_1^{\lambda}(x) \Big )+\alpha t D(Q_1\|P_1)\Big ). \end{align} Substituting $\got=nD(Q_1\|P_0)+( 4|\Xc|+4)\log (t+1)$ and setting $\lambda=1$ to get a lower bound, we obtain \begin{align}\label{eq:Eadverse} \Etzo \geq( 4 |\Xc|+4)\log (t+1)+ n \min_{Q_1\in \Pc_{\delta_n}(\Xc)} D(Q_1\|P_0) +\frac{\alpha t}{n} D(Q_1\|P_1), \end{align} where for large enough $n$ \begin{align} \Etzo \geq ( 4|\Xc|+4)\log (t+1)+ nD_{\beta(t)}(P_1\|P_0), \label{eq:lowerE0} \end{align} and \begin{equation} \beta(t)=\frac{\frac{\alpha t}{n}}{1+\frac{\alpha t}{n}}. \end{equation} Furthermore as $\beta(t)$ is strictly increasing function in $t$ and $D_{\beta(t)}(P_1\|P_0)$ is a non-decreasing function in $\beta$ \cite{Harremos}, then for all $ t\geq n$ we have \begin{equation}\label{eq:lowerRenyi} D_{\frac{\alpha}{1+\alpha}}(P_1\|P_0)\leq D_{\beta(t)}(P_1\|P_0). \end{equation} Therefore, by \eqref{eq:lowerE0}, \eqref{eq:lowerRenyi} \begin{align} \epsilon_0 &\leq \sum_{t=n}^{\infty} (\alpha t +1)^{|\mathcal{X}|} (t+1)^{-3|\mathcal{X}|-4} e^{-n D_{\frac{\alpha}{1+\alpha}}(P_1\|P_0)}\\ &\leq c e^{- nD_{\frac{\alpha}{1+\alpha}}(P_1\|P_0) }, \end{align} where $c$ is a positive constant. Next, we find a lower bound to the type-\RNum{2} error exponent. Upper bounding the type-\RNum{2} error probability, we have \begin{align} \epsilon_1 \leq \sum_{t=n}^{\infty} \PP \Big [ & t( D(\Tx\| \TX')-D(\Tx \| {P}_0)) \geq \gzt \Big ]. \end{align} By the method of types, we have \begin{align} \epsilon_1 &\leq \sum_{t=n}^{\infty} \sum_{(Q,Q_1) \in {\mathcal{Q}_{10}} (t) \cap \Pc_t(\Xc) \times \Pc_{\alpha t }(\Xc) } e^{ -t \big(D(Q\| P_1)+\alpha D(Q_1\|P_1)\big) }\\ &\leq \sum_{t=n}^{\infty} (\alpha t+1)^{|\mathcal{X}|} (t+1)^{|\mathcal{X}|} e^{- E_{10}(t)}, \label{eq:uppersensamp} \end{align} where \begin{equation}\label{eq:Eadverse} E_{10}(t)=\min_{(Q,Q_1) \in \Qc_{10}(t)} t \big(D(Q\| P_1)+\alpha D(Q_1\|P_1)\big ), \end{equation} \begin{equation}\label{eq:set} {\mathcal{Q}_{10}}(\gamma)=\Bigg \{(Q,Q_1): D(Q\| Q_1')-D(Q\| {P}_0) \geq \frac{\gzt}{t} , Q_1'=(1-\delta_n)Q_1+\delta_n U \Bigg \}. \end{equation} We show that there exists a finite $\alpha^*_{\rm seq}$ such that for every $\alpha\geq \alpha^*_{\rm seq}$, the achievable type-\RNum{2} error exponent is lower bounded by $nD(P_0\|P_1)$. Similarly to the fixed sample sized case, for every $Q_1 \in \Pc_{\delta_n}(\Xc)$, let \begin{equation}\label{eq:E10Q} E_{10}(Q_1,t) =\min_{\substack{D(Q\|Q_1)-D(Q\|P_0) \geq \frac{\gzt}{t} \\ Q \in \Pc(\Xc) } } tD(Q\|P_1) , \end{equation} which is the error exponent when the type of the training sequence is $Q_1$. We also define \begin{equation} \Etro=\min_{\substack{D(Q_1\|P_1)\leq r \\ Q_1 \in \Pc_{\delta_n}(\Xc) } } \Etqo. \end{equation} Next, we will Taylor expand $\Etro $. It is sufficient to show that there exists a finite $\alpha$ such that \begin{equation}\label{eq:condseq} \inf_{\substack{n\leq t \\ t \in \NN}} \inf_{0\leq r \leq \frac{r_c}{2}} \Etro+\alpha tr \geq nD(P_1\|P_0)+ ( 2 |\Xc|+2)\log (t+1) . \end{equation} Equivalently, \eqref{eq:condseq} can be written as the following condition \begin{align}\label{eq:alseq} \Etro+\alpha tr \geq nD(P_1\|P_0)+ ( 2 |\Xc|+2)\log (t+1) ~~ \forall r: ~ 0\leq r\leq \frac{r_c}{2}, n\leq t, t \in \NN. \end{align} Using a Taylor series expansion of $\Etro$ around $r=0$ we have \cite{BoroumandTran} \begin{align} \Etro \geq E_{10}(r=0,t)+ z_1(t) \sqrt{r} \mathbbm{1} \{t\geq n+1\}+ h r \mathbbm{1} \{t=n\} \end{align} where \begin{align} z_1(t)&=\inf_{0\leq r\leq \frac{r_c}{2}} \frac{ \partial \Etro }{\partial \sqrt{r}} \\ &=\inf_{D(Q_1\|P_1) \leq \frac{r_c}{2} } -\sqrt{ \text{Var}_{P_1} \Bigg (\lambda^*(t) \frac{Q_{\lambda^*(t)}-\frac{n}{t}P_0}{Q_1} \Bigg )},\label{eq:firstterm}\\ h&=\frac{1}{2}\inf_{0\leq r \leq \frac{r_c}{2}} \frac{ \partial^2 E_{10}(r,t=n) }{\partial \sqrt{r} ^2}. \end{align} where $Q_{\lambda^*(t)}, \lambda^*(t)$ are the minimizing distribution and the Lagrange multiplier in \eqref{eq:E10Q}. We have used the fact that for $t=n$ the optimization problem $E_{10}(r,t=n)$ is the same as the fixed sample sized classifier and hence we can use the result of the Theorem \ref{thm:upper} to lower bound the exponent by $h r \mathbbm{1} \{t=n\}$ for some finite $h$. Also for every $n<t$, the Taylor expansion of $E_{10}(r,t)$ has a nonzero first order term, hence the expansion includes $\sqrt{r}$. Writing $E_{10}(r=0,t)$ in the dual form, we have the lower bound \begin{equation} E_{10}(r=0,t) \geq \max_{\frac{1}{2} \leq \lambda } \lambda \gamma -\log \sum_{x\in \Xc} P_0^{\lambda}(x) P_1^{1-\lambda}(x), \end{equation} where $\gamma=\frac{\gzt}{t}\Big |_{Q_1=P_1}$. Then $E_{10}(r=0,t)$ is convex as it is the supremum of linear functions in $\gamma$ \cite{Boyd}. Hence, we can lower bound $E_{10}(r=0,t) $ by expanding it around $\gamma=\frac{nD(P_0\|P_1)}{t}$ using the envelope theorem we get \cite{Segal} \begin{equation} E_{10}(r=0,t) \geq \tilde{E}_{10}(r=0,t) + ( 2 |\Xc|+2)\log (t+1) \end{equation} where \begin{equation}\label{eq:seqE1lower} \tilde{E}_{10}(r=0,t) =t\max_{\frac{1}{2} \leq \lambda } \lambda \frac{nD(P_0\|P_1)}{t} -\log \sum_{x\in \Xc} P_0^{\lambda}(x) P_1^{1-\lambda}(x). \end{equation} Further expanding $\tilde{E}_{10}(r=0,t)$ around $t=n$ we have \begin{align}\label{eq:E1apprx} \Etro &\geq nD(P_1\|P_0)+ ( 2 |\Xc|+2)\log (t+1) \nonumber \\ &+m(t-n) \mathbbm{1}\{ t > 2n\} +m(t-n)^2 \mathbbm{1}\{ t \leq 2n\}+ z_1(t) \sqrt{r} \mathbbm{1} \{t\geq n+1\}+ h r \mathbbm{1} \{t=n\} \end{align} where we have used the fact that $\frac{ \partial \tilde{E}_{10}(r=0,t) }{\partial t} \Big|_{t=n}=0$, and \begin{equation} m=\min \bigg \{ \inf_{2n< t} \frac{ \partial\tilde{E}_{10}(r=0,t) }{\partial t }, \frac{1}{2} \inf_{n\leq t \leq 2n} \frac{ \partial^2\tilde{E}_{10}(r=0,t) }{\partial t ^2} \bigg \}. \end{equation} We have expanded $\tilde{E}_{10}(r=0,t)$ as above since it behaves linearly as $t$ goes to infinity while quadratically for $t$ close to $n$. Using Proposition \ref{prop:sensi}, we have \begin{align} \frac{ \partial \tilde{E}_{10}(r=0,t) }{\partial t }&= -\log \sum_{x\in \Xc} P_0^{\tilde\lambda^*(t)}(x) P_1^{1-\tilde\lambda^*(t)}(x),\\ \frac{ \partial^2 \tilde{E}_{10}(r=0,t) }{\partial t ^2}&= \frac{ \Big (\sum Q^*(t) \log \frac{P_0}{P_1} \Big)^2 } {t\text{Var}_{P_1}\Big ( \log \frac{P_0}{P_1} \Big) }, \end{align} which are finite and strictly positive for any $2n< t$, $n\leq t \leq 2n$, respectivly, since $Q^*(t)$ is a probability distribution and $\frac{1}{2} \leq \tilde \lambda^*(t)< 1$ is the optimizer in \eqref{eq:seqE1lower}. From condition \eqref{eq:alseq} by substituting the approximation \eqref{eq:E1apprx}, we need that \begin{align}\label{eq:alcons} m\Big((t-n) \mathbbm{1}\{ t > 2n\} +(t-n)^2 \mathbbm{1}\{ t \leq 2n\} \Big) + z_1(t) \sqrt{r} \mathbbm{1} \{t\geq n+1\}+ h r \mathbbm{1} \{t=n\} + \alpha t r \geq 0 \end{align} for every $0\leq r, \ n\leq t, t\in \NN$. For $\alpha > \frac{|h|}{n}$, letting $r^*=\frac{z_1(t) \mathbbm{1} \{t\geq n+1\}}{2( \alpha t+h\mathbbm{1} \{t=n\} ) }$ to minimize the LHS of \eqref{eq:alcons} over $r$ we get the condition \begin{align} m\Big((t-n) \mathbbm{1}\{ t > 2n\} +(t-n)^2\mathbbm{1}\{ t \leq 2n\}\Big) -\frac{\rho(t)}{{4(\alpha t+h\mathbbm{1} \{t=n\} )}} \geq 0, \end{align} where \begin{equation} \rho(t)= {z^2_1(t) \mathbbm{1} \{t\geq n+1\}} \end{equation} and we have used that $0\leq z_1(t) < \infty$ by \eqref{eq:firstterm} since for every $Q_1$ satisfying $D(Q_1\|P_1)\leq \frac{r_c}{2}$, $\lambda^*(t)$ is finite and $Q_{\lambda^*(t)}$ is a probability distribution, hence the variance in \eqref{eq:firstterm} is finite. Moreover, $\rho(t)$ equals to zero at $t=n$ and it is finite for every $t\in \NN$ with finite limit as $t\rightarrow \infty$. Hence there exists a finite $c_1$ such that $\rho(t)\leq c_1\Big((t-n) \mathbbm{1}\{ t > 2n\} +(t-n)^2\mathbbm{1}\{ t \leq 2n\}\Big)$ for $t\in \NN$. Then, by further relaxing condition \eqref{eq:alcons}, we need \begin{align} \Big((t-n) \mathbbm{1}\{ t > 2n\} +(t-n)^2\mathbbm{1}\{ t \leq 2n\}\Big)\Big (m- \frac{c_1}{4\alpha t} \Big ) \geq 0, \end{align} where we have dropped the $h\mathbbm{1} \{t=n\}$ term as it is only nonzero for $t=n$ which sets $(t-n)$ and $(t-n)^2$ to zero. Therefore, if $\alpha >\frac{c_1}{4m}$, the sufficient condition is satisfied and \begin{equation}\label{eq:lowerE10} E_{10}(t) \geq nD(P_0\|P_1) + ( 2 |\Xc|+2)\log (t+1) \end{equation} for all $n\leq t, t \in \NN, 0 \leq r \leq \frac{r_c}{2} $. Therefore, for $\alpha>\max \Big \{ \frac{c_1}{4m}, |h|, \frac{2D(P_1\|P_0)}{r_c}+4|\Xc|+4 \Big \}$, by substituting \eqref{eq:lowerE10} in \eqref{eq:uppersensamp}, we get \begin{align}\label{eq:experror1} \epsilon_1 \leq c e^{- nD(P_0\|P_1) }, \end{align} where $c$ is a positive constant. Next, we find the average stopping times of the proposed sequential classifier. We first show the convergence of $\tau$ in probability under each hypothesis, and by proving its uniform integrability, we can conclude its convergence in the $L^1$ norm. The following lemma states that for every $n$ the classifier stops with the probability of one. \begin{lemma}\label{lem:finiteness}. Let $\tau_0, \tau_1$ be the the smallest time that the sequential classifier crosses threshold $\gzt$ or $\got$ respectively, i.e., \begin{align}\label{eq:tau1} \tau_{0}=\inf \{t\geq n: {S}_t \geq \gzt\}, \quad \quad \tau_{1}=\inf \{t\geq n: {S}_t \leq -\got\}. \end{align} Then for $i \in \{0,1\}$, $t\geq n$, \begin{equation}\label{eq:finiteupper} \PP_i[\tau_i > t] \leq c_i(t+1)^{d_i |\mathcal{X}|} e^{\xi_i n} e^{-t E_i}, \end{equation} where $E_i, \xi_i,c_i,d_i >0$ and finite. \end{lemma} \begin{proof} By the method of types the probability of passing the threshold $\gzt$ under the first hypothesis at a time after $t\geq n$ can be upper bounded by \begin{align} \PP_0[\tau_0 > t] &\leq \PP_0 \big[ {S}_{t} \leq \gzt \big ]\\ &= \PP_0 \bigg[ D(\Tx\|\TX')-D(\Tx\|P_0) \leq \frac{\gzt}{t} \bigg ] \\ &\leq \sum_{(Q,Q_1) \in {\mathcal{Q}_{00} \cap \Pc_t(\Xc)} \times \Pc_{\alpha t}(\Xc) } e^{ -t \big(D(Q\| P_0)+\alpha D(Q_1\|P_1)\big) }\\ &\leq (\alpha t +1)^{|\Xc|}(t+1)^{|\mathcal{X}|} e^{- \Etzz}, \label{eq:expt} \end{align} where \begin{equation} \Etzz=\min_{(Q,Q_1)\in \Qc_{00}(t)} t \big(D(Q\| P_0)+\alpha D(Q_1\|P_1)\big ), \end{equation} \begin{equation}\label{eq:set} {\mathcal{Q}_{00}}(t)=\bigg \{(Q,Q_1): D(Q\| {P}_0) - D(Q\| Q_1) \geq - \frac{\gzt}{t} \bigg \}. \end{equation} For every fix $Q_1$ we can use the dual form of the optimization over $Q$ to get \cite{BoroumandTran} \begin{align} \Etzz &=\min_{Q_1\in \Pc(\Xc)} \Big ( t\max_{ 0 \leq \lambda } -\frac{\gzt}{t} \lambda - \log \sum_{x\in \Xc} P_0^{1-\lambda}(x) Q_1^{\lambda}(x) \Big )+\alpha t D(Q_1\|P_1)\Big )\\ &\geq \min_{Q_1\in \Pc(\Xc)} \Big ( t\max_{ 0 \leq \lambda \leq 1} -\frac{\gzt}{t} \lambda - \log \sum_{x\in \Xc} P_0^{1-\lambda}(x) Q_1^{\lambda}(x) \Big )+\alpha t D(Q_1\|P_1)\Big ) \label{eq:dualE00} \end{align} Let \begin{align} E(\gamma)=\max_{0 \leq \lambda \leq 1 } \lambda \gamma -\log \sum_{x\in \Xc} P_0^{1-\lambda}(x) Q_1^{\lambda}(x) . \end{align} Then $E(\gamma)$ is convex as it is the supremum of linear functions in $\gamma$ \cite{Boyd}. Therefore letting the $E_0$ to be the nonzero minimum of the optimization, we have \begin{equation} E\Big(\frac{\gamma}{t}\Big) \geq E_0 +\frac{\partial E(\gamma)}{ \partial \gamma}\Bigg |_{\gamma=0} \frac{\gamma}{t}. \end{equation} Applying the expansion to \eqref{eq:dualE00} we get \begin{align} \Etzz& \geq t \min_{Q_1\in \Pc(\Xc)} \Big ( C(P_0\|Q_1)- \frac{n}{t} D(P_0\|Q_1)- (4 |\Xc|+4) \frac{\log(t+1)}{t} \Big)^{+} +\alpha D(Q_1\|P_1) \end{align} where we have used $\frac{\partial E(\gamma)}{ \partial \gamma} = \lambda^*$ with $\lambda^*\leq 1$ to be the Lagrange multiplier solving \eqref{eq:dualE00}, and lower bounding $\Etzz$ by setting $\lambda^*=1$. Moreover, $C(P_0\|Q_1)= \max_{0 \leq \lambda \leq 1} - \log \sum_{x\in \Xc} P_0^{1-\lambda}(x) Q_1^{\lambda}(x) $ is the Chernoff information. Further lower bounding $\Etzz$ we get \begin{align} \Etzz& \geq t \min_{Q_1\in \Pc(\Xc)} \Big ( C(P_0\|Q_1)- \frac{n}{t} D(P_0\|Q_1) \Big)^{+} \mathbbm{1} \{t \geq \zeta n \} - (4 |\Xc|+4) \frac{\log(t+1)}{t} +\alpha D(Q_1\|P_1)\\ & \geq - (4 |\Xc|+4) \log(t+1) + t \min_{Q_1\in \Pc(\Xc)} \Big ( C(P_0\|Q_1)- \frac{1}{\zeta} D(P_0\|Q_1) \Big)^{+} \mathbbm{1} \{t \geq \zeta n \} +\alpha D(Q_1\|P_1) \label{eq:lowerT0} \end{align} where $(x)^+=\max \{x,0 \}$. Choosing $\zeta > \frac{D(P_0\|P_1)}{C(P_0\|P_1)}$, for every $t\geq \zeta n$ the solution to the optimization \eqref{eq:lowerT0} is nonzero since $\alpha D(Q_1\|P_1)$ can be zero if and only if $Q_1=P_1$ while the first term $ \Big ( C(P_0\|Q_1)- \frac{1}{\zeta} D(P_0\|Q_1) \Big)^{+} \mathbbm{1} \{t \geq \zeta n \} $ is nonzero for that choice of $Q_1$. Therefore, we have \begin{align} \Etzz& \geq - (4 |\Xc|+4) \log(t+1) + E_{0}t\mathbbm{1} \{t \geq \zeta n \} \\ & \geq - (4 |\Xc|+4) \log(t+1) + E_{0} t(1-\zeta \frac{n}{t})\\ &=- (4 |\Xc|+4) \log(t+1) + E_{0}t - \xi_0 n \label{eq:lowerexpt} \end{align} where $\xi_0= E_{0} \zeta, E_0> 0$. Substituting \eqref{eq:lowerexpt} in \eqref{eq:expt} gives \eqref{eq:finiteupper}. To prove the result under the hypothesis $P_1$, let \begin{equation} \Etoo=\min_{(Q,Q_1)\in \Qc_{11}(t)} t \big(D(Q\| P_1)+\alpha D(Q_1\|P_1)\big ), \end{equation} \begin{equation} {\mathcal{Q}_{11}}(t)=\Big \{(Q,Q_1): D(Q\| Q_1) - D(Q\| P_0) \geq - \frac{\got}{t} \Big \}. \end{equation} Similar to the steps for hypothesis $P_0$, we have \begin{align} &\Etoo+ \lambda^*(4 |\Xc|+4) \log(t+1) \\ &\geq t \min_{Q_1\in \Pc(\Xc)} \Big( \max_{0 \leq \lambda } - \log \sum_{x\in \Xc} P_1(x) Q_1(x)^{-\lambda} P_0^{\lambda} \Big ) -\frac{n}{t} \lambda^* D(Q_1\|P_0) +\alpha D(Q_1\|P_1) \label{eq:lowerE1t} \\ & \geq t\min_{Q,Q_1: D(Q\|Q_1) \geq D(Q\|P_0)} D(Q\|P_1)+ \frac{\alpha}{2} D(Q_1\|P_1) + \min_{Q_1} \frac{\alpha}{2}tD(Q_1\|P_1) -n \lambda^* D(Q_1\|P_0) \\ & \geq t\min_{Q,Q_1: D(Q\|Q_1) \geq D(Q\|P_0)} D(Q\|P_1)+ \frac{\alpha}{2} D(Q_1\|P_1) + n\min_{Q_1} \frac{\alpha}{2}D(Q_1\|P_1) - \lambda^* D(Q_1\|P_0)\\ &=E_{1}t - \xi_1 n \end{align} where in the last step we used the fact that both optimizations are finite and $E_1 >0, \xi_1 >0$, and where $\lambda^*$ is the Lagrange multiplier solving the maximization in \eqref{eq:lowerE1t}. This concludes the proof. \end{proof} \begin{lemma}\label{lem:vanishing} For $i\in\{0,1\}$ \begin{equation} |\gamma_{i,n}(t+1)-\gamma_{i,n}(t)| \xrightarrow[]{a.s.} 0, \end{equation} as $t\rightarrow \infty$. \end{lemma} \begin{proof} We first show the lemma for $i=0$. By the triangle inequality and the definition of KL-divergence, we have \begin{align} |\gamma_{0,n}(t+1)-\gamma_{0,n}(t)| &\leq (4 |\Xc|+4) | \log(t+1) -\log(t)| +n \Bigg | \sum_{x\in\Xc} P_0(x)\log \frac{\TX'^{1:\alpha(t+1)}(x)}{\TX'^{1:\alpha t}(x)}\Bigg |\\ &\leq \frac{c}{t}+ n\Bigg | \sum_{x\in\Xc} P_0(x)\log \frac{\frac{ t}{t+1}\TX'^{1:\alpha t}(x) +\frac{1 }{t+1}\TX^{ \alpha t+1: \alpha(t+1)}(x) }{\TX'^{1:\alpha t}(x)}\Bigg |\\ &\leq \frac{c}{t}+ n\Bigg | \sum_{x\in\Xc} P_0(x)\log \frac{ t}{t+1} + \frac{1 }{t+1} \frac{\TX^{\alpha t+1: \alpha (t+1)} (x) }{\TX'^{1:\alpha t}(x)}\Bigg | \rightarrow 0 \end{align} where $\TX^{i:j}$ is the type of the sequence $(X_i,\ldots,X_j)$, and in the last step, the logarithm will be either zero or goes to zero as $t$ goes to infinity. Next, for $\gamma_{1,n}(t)$, by the triangle inequality and the $L_1$ bound on entropy \cite{Cover} we have \begin{align} |\gamma_{1,n}(t+1)-\gamma_{1,n}(t)| \leq &(4 |\Xc|+4)| \log(t+1) -\log(t)| +n \Big |H(\TX'^{ 1: \alpha (t+1)})- H(\TX'^{1: \alpha t}) \Big | \nonumber \\ & + n \bigg|\sum_{x\in\Xc} (\TX'^{ 1: \alpha (t+1)} (x) - \TX'^{1: \alpha t}(x) ) \log P_0(x) \bigg | \\ \leq& \frac{c}{t}-n \|\TX'^{ 1: \alpha (t+1)}- \TX'^{1: \alpha t} \|_1 \log \frac{\|\TX'^{ 1: \alpha (t+1)}- \TX'^{1: \alpha t} \|_1 }{|\Xc|} + c'\|\TX'^{ 1: \alpha (t+1)}- \TX'^{1: \alpha t} \|_1 \\ \leq &\frac{c}{t}+ \frac{n }{t+1} \log \frac{1 }{|\Xc| (t+1)}+\frac{c'}{t+1} \rightarrow 0, \end{align} where in the last step we have used \begin{align} \|\TX'^{ 1: \alpha (t+1)}- \TX'^{1: \alpha t} \|_1 \leq \frac{1}{t+1} \|\TX'^{ 1: \alpha t}- \TX'^{ \alpha t : \alpha (t+1)} \|_1 \leq \frac{1}{t+1}, \end{align} as $\TX'$ is a type of a training sequence, and $c,c'$ are positive constants. \end{proof} Now by the finiteness of $\tau_0$ for every $n$, and the definition of $\tau_0$, there exists a finite $\tau_0$ with probability one such that \begin{align}\label{eq:convergp} {S}_{\tau_0-1} < \gamma_{0,n}(\tau_0-1), ~~~ \gamma_{0,n}(\tau_0) \leq {S}_{\tau_0}~~~ \text{w.p}.1. \end{align} Furthermore, for every $\tau_0$ we have \begin{align} \frac{{S}_{\tau_0}}{\tau_0}= D(\Tx^{\tau_0} \| \TX^{\tau_0})- D(\Tx^{\tau_0}\|P_1)+o(1). \end{align} Also, since by design $\tau_0 \geq n$, using the WLLN, and the continuous mapping theorem, as $n \rightarrow \infty$, we get \begin{align}\label{eq:convergp1} \frac{{S}_{\tau_0}}{\tau_0} \xrightarrow[]{p} D(P_0\|{P}_1), \quad \frac{{S}_{\tau_0-1}}{\tau_0-1} \xrightarrow[]{p} D(P_0\|{P}_1). \end{align} Therefore, by Lemma \ref{lem:finiteness}, Lemma \ref{lem:vanishing}, and \eqref{eq:convergp}, \eqref{eq:convergp1} we can conclude that \begin{equation}\label{eq:convP1} \frac{\gamma_{0,n}(\tau_0)}{\tau_0} \xrightarrow[]{p} { D(P_0\|P_1)}, \end{equation} as $n \rightarrow \infty$. Also, we have \begin{equation}\label{eq:conveq} \frac{\gamma_{0,n}(\tau_0)}{\tau_0}= \frac{n}{\tau_0} D(P_0\|\TX') +\frac{\log (\tau_0+1)}{\tau_0}, \end{equation} and by assumtion $\max_{x\in\Xc} \frac{P_0(x)}{P_1(x)} \leq c$, $D(P_0\|\TX')$ is a consistent estimator of $D(P_0\|P_1)$ \cite{KL}, and hence \begin{equation}\label{eq:consistency} D(P_0\|\TX') \xrightarrow[]{p} D(P_0\|P_1). \end{equation} Finally by \eqref{eq:convP1}, \eqref{eq:conveq}, \eqref{eq:consistency}, and using continuous mapping theorem \cite{Resnick}, we have \begin{equation} \frac{\tau_0}{n}\xrightarrow[]{p} 1, \end{equation} as $n\rightarrow \infty$. To show the convergence in $L^1$ we only need to prove the uniform integrability of the sequence of random variables $\frac{\tau_0}{ n}$, where $\tau_0$ depends on $n$. Equivalently, we need to show that, \begin{equation}\label{eq:uniform} \lim_{t \rightarrow \infty} \sup_{n\geq 1} \mathbb{E}_{P_0} \Bigg [\frac{\tau_0}{ n} \mathds{1}\Big \{\frac{\tau_0}{ n} \geq t \Big \} \Bigg] =0. \end{equation} By \eqref{eq:finiteupper}, we can upper bound the given expectation in \eqref{eq:uniform} as \begin{align} \mathbb{E}_{P_0} \Big [\frac{\tau_0 }{n} \mathds{1}\big \{{\tau_0} \geq t n \big \} \Big] &= \frac{1}{n} \sum_{m=1}^{\infty} \PP_0\big [\tau_0-t n \geq m \big ]\\ & \leq \frac{1}{n} t e^{-n (t E_0-\xi_0)} \sum_{m=0 }^{\infty} c(m+tn+1)^{4|\mathcal{X}|} e^{-m E_0}. \end{align} Hence the expectation is vanishing as $t\rightarrow\infty$ for every $n$ giving the uniform integrability of $\frac{\tau_0}{ n}$, and hence convergence in $L^1$ \cite{Bill}, i.e, \begin{equation} \label{eq:convergT1} \lim_{n \rightarrow \infty} \mathbb{E}_{P_0} \Big [ \Big|\frac{\tau_0}{ n}- 1 \Big | \Big]=0. \end{equation} Finally, we prove the convergence of $\tau$. By \eqref{eq:experror1}, \eqref{eq:convP1} and the union bound, we obtain \begin{align} \PP_0 \Big[\Big|\frac{\tau}{ n}- 1\Big | \geq \epsilon \Big] & \leq \PP_0 \Big[\Big|\frac{\tau}{ n}- 1\Big | \geq \epsilon, {\phi}=0 \Big ]+ \PP_0[{\phi}=1 ]\\ &=\PP_0 \Big[\Big|\frac{\tau_0}{ n}- 1 \Big | \geq \epsilon\Big] + \epsilon_0, \end{align} which tends to $0$ as $n \rightarrow \infty$, establishing the convergence of $\frac{\tau}{ n}$ in probability. Now, using that $\tau\leq \tau_0$ we have \begin{align}\label{eq:upper1} \mathbb{E}_{P_0} \Big [\frac{\tau}{ n} \mathds{1}\Big \{\frac{\tau}{ n} \geq t \big \} \Big] \leq \mathbb{E}_{P_0} \Big [\frac{\tau_0}{ n} \mathds{1}\Big \{\frac{\tau_0}{ n} \geq t \Big \} \Big]. \end{align} Therefore, uniform integrability of $\tau_0$ gives the uniform integrability of $\tau$, and hence convergence in $L^1$ norm and also expectation of $\frac{\tau}{ n}$, which concludes the proof. \section{Proof of Theorem \ref{thm:converse}} \label{sec:proofth52} For the type-\RNum{2} error exponent, the converse for sequential hypothesis testing is applicable, i.e., for every sequential test with $\mathbb{E}_{P_0}[\tau] \leq n$, we have $E_1 \leq D(P_0\|P_1)$ \cite{Poly}. To find an upper bound to $E_0$ we use the following lemma. \begin{lemma}\label{lem:theconvseq} For any type based sequential test $\Phi^{\rm seq}$, let $\tau \in \Nc_1^\epsilon$, where $\Nc_1^\epsilon=\{1,...,t \}$ is the typical stopping time set such that \begin{equation} P_1(\tau \in \Nc_1^\epsilon)\geq 1-\frac{\epsilon}{2}. \end{equation} Also for every $t$, let $(\xv^t, \X^{\alpha t}) \in \Bc_t^\epsilon$ where \begin{equation} \Bc_t^\epsilon= \Bigg\{ (\xv^t, \X^{\alpha t}) : \max_{a \in \mathcal{X}} \Big \{ |\Tx(a)-\TX(a)| \Big \} \leq \sqrt{\frac{2}{(\alpha \wedge 1) t} \log \frac{8|\Xc|}{\epsilon}} \Bigg \}. \end{equation} Then for any type based test $(\phi(\Tx,\TX),\tau)$ such that for all distributions ${P}_1 \in \Pc(\Xc)$, \begin{equation} \epsilon_1\big(\phi(\xv^\tau,\X^{\alpha \tau})\big) < \epsilon, \quad\epsilon \in\Big (0,\frac{1}{2}\Big), \end{equation} we have \begin{equation}\label{eq:steinseq} \PP \Big [ \phi(\xv^\tau,\X^{\alpha \tau})=1, (\xv^\tau, \X^{\alpha \tau}) \in \Bc_\tau^\epsilon, \tau \in \Nc_1^\epsilon \Big] \geq 1-2\epsilon. \end{equation} \end{lemma} \begin{proof} We prove the lemma by contradiction. Assume \begin{equation} \PP \Big [ \phi(\xv^\tau,\X^{\alpha \tau})=1, (\xv^\tau, \X^{\alpha \tau}) \in \Bc_\tau^\epsilon, \tau \in \Nc_1^\epsilon \Big] < 1-2\epsilon. \end{equation} Then \begin{align} 2\epsilon &< \PP \Bigg [ \phi(\xv^\tau,\X^{\alpha \tau})\neq 1\cup (\xv^\tau, \X^{\alpha \tau}) \notin \Bc_\tau^\epsilon \cup \tau \notin \Nc_1^\epsilon \Big]\\ &\leq \PP \Big [ \phi(\xv^\tau,\X^{\alpha \tau})\neq 1 \Big] + \PP \Big [ (\xv^\tau, \X^{\alpha \tau}) \notin \Bc_\tau^\epsilon \Big]+ \PP \Big [ \tau \notin \Nc_1^\epsilon \Big]\\ &\leq \epsilon_1 + \frac{\epsilon}{4} +\frac{\epsilon}{4} +\frac{\epsilon}{2}. \end{align} where in the last step we used Lemma \ref{lem:concentration}. Hence, $\epsilon \leq \epsilon_1$ which is a contradiction and hence \eqref{eq:steinseq} holds. \end{proof} By Lemma \ref{lem:theconvseq} we can conclude that if the condition \eqref{eq:steinseq} does not hold there exists a distribution $P_1$ such the that the type-\RNum{2} error probability is bounded away from zero and hence the type-\RNum{2} error exponent of such test equals to zero. Therefore, by \eqref{eq:steinseq} we can lower bound the $\epsilon_0$ for any test with nonzero $E_1$ as \begin{align} \epsilon_0 =& \PP[\phi(\xv^\tau, \X^{\alpha \tau}) =1]\\ \geq &\PP \Big [ \phi(\xv^\tau,\X^{\alpha \tau})=1, (\xv^\tau, \X^{\alpha \tau}) \in \Bc_\tau^\epsilon, \tau \in \Nc_1^\epsilon \Big]\\ =& \PP \Big [ \phi(\xv^\tau,\X^{\alpha \tau})=1 \Big| (\xv^\tau, \X^{\alpha \tau}) \in \Bc_\tau^\epsilon, \tau \in \Nc_1 \Big] \PP \Big [ (\xv^\tau, \X^{\alpha \tau}) \in \Bc_\tau^\epsilon , \tau \in \Nc_1 \Big]. \end{align} By the previous lemma, we can lower bound the first probability by $1-2\epsilon$. Also, let $\epsilon$ to be sufficiently small, such that $n \in \Nc_1^\epsilon=\{1,...,N\}$, then by the method of types \begin{align} \epsilon_0 \geq& (1-2\epsilon) \sum_{t=1}^{N} \PP \Big [ (\xv^\tau, \X^{\alpha \tau}) \in \Bc_\tau^\epsilon \Big | \tau =t \Big] \PP_0 [ \tau=t]\\ \geq &(1-2\epsilon)\sum_{t=1}^{n} (t+1)^{-|\mathcal{X}|} (\alpha t+1)^{-|\mathcal{X}|} e^{-t \min_{(Q,Q_1) \in \Bc_t^\epsilon} D(Q\|P_0) + \alpha D(Q_1\|P_1) } \PP_0 [ \tau=t]\\ \geq &(1-2\epsilon)ce^{-n \min_{(Q,Q_1) \in \Bc_t^\epsilon} D(Q\|P_0) + \alpha D(Q_1\|P_1) } \PP_0 [ \tau\leq n]\\ \end{align} where $c$ is a positive constant. Now by $\mathbb{E}_{P_0}[\tau]\leq n$, the optimal test should stops by the time $n$ with a positive probability, i.e., $\PP_0 [ \tau\leq n]>0$, since otherwise, $\mathbb{E}_{P_0}[\tau] > n$. Finally, by letting $\epsilon \rightarrow 0, n \rightarrow \infty$, we have \begin{equation} E_0 \leq D_{\frac{\alpha}{1+\alpha}}(P_1\|P_0). \end{equation} Hence, for any sequential test with a finite stopping time, and the type-\RNum{2} error probability that is bounded away from one for every distribution $P_1$, the type-\RNum{1} error exponent is bounded by $D_{\frac{\alpha}{1+\alpha}}(P_1\|P_0)$, which concludes the proof. \bibliographystyle{ieeebib} \bibliographystyle{ieeetr}

https://arxiv.org/abs/1909.04651

Inviscid limit of vorticity distributions in Yudovich class

We prove that given initial data $\omega_0\in L^\infty(\mathbb{T}^2)$, forcing $g\in L^\infty(0,T; L^\infty(\mathbb{T}^2))$, and any $T>0$, the solutions $u^\nu$ of Navier-Stokes converge strongly in $L^\infty(0,T;W^{1,p}(\mathbb{T}^2))$ for any $p\in [1,\infty)$ to the unique Yudovich weak solution $u$ of the Euler equations. A consequence is that vorticity distribution functions converge to their inviscid counterparts. As a byproduct of the proof, we establish continuity of the Euler solution map for Yudovich solutions in the $L^p$ vorticity topology. The main tool in these proofs is a uniformly controlled loss of regularity property of the linear transport by Yudovich solutions. Our results provide a partial foundation for the Miller--Robert statistical equilibrium theory of vortices as it applies to slightly viscous fluids.

\section{Introduction } In this paper we discuss the connection between Yudovich solutions of the Euler equations \begin{equation}\label{eu} \partial_t \omega+ u\cdot \nabla \omega= g, \end{equation} with bounded forcing $g\in L^\infty(0,T;L^\infty(\mathbb{T}^2))$, and initial data \begin{equation} \omega(0)=\omega_0\in L^\infty(\mathbb{T}^2), \label{ideu} \end{equation} and the vanishing viscosity limit ($\lim_{\nu\to 0}$) of solutions of the Navier-Stokes equations, \begin{equation}\label{nse} \partial_t \omega^\nu+ u^\nu\cdot \nabla \omega^\nu= \nu \Delta \omega^\nu+ g, \end{equation} with initial data \begin{equation} \omega^{\nu}(0) = \omega^\nu_0\in L^{\infty}(\mathbb{T}^2), \label{idnse} \end{equation} and the same forcing $g$. We consider uniformly bounded initial data \begin{equation} \sup_{\nu>0}\|\omega^\nu_0\|_{L^{\infty}(\mathbb T^2)}\le \Omega_{0,\infty}<\infty. \label{supid} \end{equation} The solutions of (\ref{eu}), (\ref{ideu}), (\ref{nse}), (\ref{idnse}) are uniformly bounded in $L^{\infty}(\mathbb T^2)$: \begin{equation} \sup_{\nu\ge 0}\sup_{0\le t\le T}\|\omega^{\nu}(t)\|_{L^{\infty}(\mathbb T^2)} \le \Omega_{\infty} = \Omega_{0,\infty} + \int_0^T\|g(t)\|_{L^{\infty}(\mathbb T^2)}dt. \label{omegafty} \end{equation} This bound is valid in $\mathbb T^2$ or $\mathbb R^2$ but is not available if boundaries are present or in 3D. The bound will be used repeatedly below. We are interested in the small viscosity behavior of vorticity distribution function $\pi_{\omega^{\nu}(t)}({\rm d} y)$ defined by \begin{equation}\label{vordens} \int f(y) \pi_{\omega^\nu(t)}({\rm d} y) = \int f(\omega^\nu(t,x)) {\rm d} x, \end{equation} for all continuous functions (observables) $f$. If $\omega_0^\nu \to \omega_0$ we prove that the distributions convergence \begin{equation} \label{distconv} \pi_{\omega^\nu(t)}({\rm d} y)\ \ \xrightarrow[]{\nu \to 0} \ \ \pi_{\omega(t)}({\rm d} y) = \pi_{\omega_0}({\rm d} y), \end{equation} where the time invariance of the vorticity distribution function for the Euler equations follows from Lagrangian transport $\omega(t)= \omega_0\circ X_t^{-1}$ and volume preservation of the homeomorphism $A_t=X_t^{-1}$. The statement \eqref{distconv} is a consequence of the strong convergence of the vorticity in $L^\infty(0,T;L^p(\mathbb{T}^2))$ for all $p\in [1,\infty)$ and for any $T>0$. We prove this fact here, extending previous work for vortex patch solutions with smooth boundary \cite{CW1}, and removing additional assumptions on the Euler path \cite{CW2}. Implications of our result for equilibrium theories of decaying two dimensional turbulence \cite{Miller,Robert} are briefly discussed at the end of this paper. Our main result is the following. \begin{thm}\label{Thm} Let $\omega$ be the unique Yudovich weak solution of the Euler equations with initial data $\omega_0\in L^\infty(\mathbb{T}^2)$ and forcing $g\in L^\infty(0,T;L^\infty(\mathbb{T}^2))$. Let $\omega^\nu$ be the solution of the Navier-Stokes equation with the same forcing and initial data $\omega^\nu_0 \to \omega_0$ strongly in $L^2(\mathbb{T}^2)$. Then, for any $T>0$ and $p\in [1,\infty)$, the inviscid limit $\omega^\nu\to \omega$ holds strongly in $L^\infty(0,T; L^p(\mathbb{T}^2))$: \begin{equation} \lim_{\nu\to 0}\sup_{0\le t\le T}\|\omega^\nu(t)-\omega (t)\|_{L^p(\mathbb T^2)} = 0. \label{invp} \end{equation} Consequently, the distributions converge, \begin{equation} \lim_{\nu\to 0} \pi_{\omega^\nu(t)}({\rm d} y) = \pi_{\omega_0}({\rm d} y), \end{equation} for all $ t\in [0,T]$. \end{thm} \begin{rem} There are several senses in which this theorem is sharp. First, there can be no infinite time result as the Euler solution is conservative and the Navier-Stokes solution is dissipative. This is obvious if we consider the stationary solutions $\omega_0(x)=\sin(Nx)$ and $g=0$. Secondly, there can be no rate without additional regularity assumptions on $\omega_0$, as is the case for the heat equation. Thirdly, there can be no strong convergence in $L^\infty$ because $\omega_0$ may not be continuous while $\omega^\nu$ is smooth for any $t>0$. And, finally there can be no strong convergence for $p>1$ in domains with boundaries, if the boundary condition of the Navier-Stokes solutions is no slip, and the Euler solution has non-vanishing tangential velocity at the boundary, in other words, if there are boundary layers \cite{Kelliher}. \end{rem} \begin{rem} One implication of theorem \ref{Thm} is that the dissipation of convex functions of vorticity must vanish, \begin{equation}\label{enstdiss} \lim_{\nu \to 0} \nu \int_0^T \int_{\mathbb{T}^2} f''(\omega^\nu) |\nabla \omega^\nu|^2 {\rm d} x {\rm d} t = 0. \end{equation} In the special case when $f(x)= |x|^2/2$, the above is the enstrophy dissipation (palenstrophy). In fact, it was proved by Eyink that anomalous enstrophy dissipation requires that $\omega_0\notin L^2(\mathbb{T}^2)$ \cite{Eyink01,HML06}. The idea is that, if $\omega_0 \in L^2(\mathbb{T}^2)$, the enstrophy remains uniformly-in-$\nu$ bounded since it is non-increasing under the Navier-Stokes evolution. Applying the Aubin-Lions lemma yields weak convergence on subsequences to $\omega$, a weak solution of the Euler equations (possibly non-unique). Thus $\omega^\nu\to \omega$ in $C(0,T; w-L^2(\mathbb{T}^2))$. Moreover, for such initial data, all weak Euler solutions can be shown to be renormalized in the sense of DiPerna-Lions and hence conservative \cite{CS15}. Thus, by weak lower semi-continuity of the $L^2$ norm, the Navier-Stokes enstrophy balance implies also that norms converge and hence the convergence is strong in $L^2$, pointwise in time, i.e. $\omega^\nu(t)\to \omega$ in $L^2(\mathbb{T}^2)$ for each $t\in [0,T]$. In fact, whenever the vorticity converges weakly to a conservative weak Euler solution, one has strong convergence and there can be no anomaly. The convergence can be made uniform in time. This proof using compactness, however, inherently gives a qualitative statement and one cannot extract information about rates of convergence. On the other hand, our proof is quantitative. Specifically, given information on, say, the spectrum of the initial vorticity at high wavenumber, one can obtain a rate of convergence. One class of examples which we discuss in corollary \ref{cor} concerns vorticity in the space $\omega_0\in L^\infty \cap B^s_{p,\infty}$ for $s>0$. However, more generally, for any $\omega_0\in L^\infty$ our proof provides a computable rate of convergence depending only on $\omega_0$, independent of the particular subsequence $\{\nu_n\}_{n\geq 0}$. \end{rem} A corollary of the proof of theorem \ref{Thm} and lemma \ref{chem} is the continuity of the Yudovich solution map $\omega(t)= S_{t}(\omega_0)$ in the $L^p$ topology when restricted to fixed balls in $L^\infty$. \begin{cor}\label{contcor} Fix $T>0$, $\omega_0 \in L^\infty(\mathbb{T}^2)$ and ${\varepsilon}>0$. There exists $\delta:= \delta(\omega_0,T, {\varepsilon})$ so that for all $\omega \in L^\infty(\mathbb{T}^2)$, \begin{equation} \|\omega- \omega_0\|_{L^p(\mathbb T^2)}<\delta, \qquad \text{implies} \qquad \sup_{t\in[0,T]} \|S_t(\omega) - S_t(\omega_0)\|_{L^p(\mathbb T^2)} <{\varepsilon}. \end{equation} \end{cor} The proof of Theorem \ref{Thm} is based on the fact that linear transport by Yudovich solutions has a short time uniformly controlled loss of regularity: it maps bounded sets in $W^{1,p}, \; p>2$ to bounded sets in $H^1$, uniformly in viscosity. More precisely, we consider the Yudovich solutions $\omega (t)$ and $\omega^\nu(t)$ of the Euler and Navier-Stokes equations with initial data $\omega_0\in L^{\infty}$ and denote their corresponding velocities by $u(t)$ and, respectively, $u^\nu(t)$. We take a sequence of regularizations $\omega_{0,n}\in W^{1,\infty}$ of $\omega_0$ which is uniformly bounded in $W^{1,p}, \; p>2$ and is such that $\omega_{0,n}\rightarrow \omega_0$ strongly in $L^2$. We let $\omega_n(t)$ be the unique solutions of the linear transport problems \[\partial_t \omega_n+u\cdot\nabla\omega_n=0\] and respectively $\omega_n^\nu(t)$ of \[\partial_t \omega_n^\nu+u^\nu\cdot\nabla\omega_n^\nu=\nu\Delta\omega_n^\nu.\] On one hand, $\omega_n(t)$ remains close to $\omega(t)$ and $\omega_n^\nu(t)$ remains close to $\omega^\nu(t)$ in $L^p$ spaces because linear transport bu Yudovich velocities is clearly bounded in $L^p$. The essential additional ingredient we show is a controlled loss of regularity: $\omega_n(t)$ and $\omega_n^\nu(t)$ are bounded in $H^1$ on a short time interval by their initial norms in $W^{1,p}$, $p>2$. This uses the fact that $\nabla u$ and $\nabla u^\nu$ are exponentially integrable. The rest of the proof rests on these observations as well as energy estimates and a time splitting. In the direction of propagating regularity, we also prove the fact that if additional smoothness is assumed on the data then some degree of fractional smoothness in $L^p$ can be propagated uniformly in viscosity. We consider the unforced case $g=0$ and we fix initial data $\omega_0^\nu = \omega_0$ for simplicity, the natural extension being straightforward. \begin{prop}\label{propReg} Suppose $\omega_0\in (L^\infty\cap B^{s}_{p,\infty})(\mathbb{T}^2)$ for some $s>0$ and some $p\geq 1$. Then the solutions of the Navier-Stokes equations satisfy $\omega^\nu(t)\in(L^\infty\cap B^{s(t)}_{p,\infty})(\mathbb{T}^2)$ uniformly in $\nu$, where \[ s(t) = {s}\exp(-Ct\|\omega_0\|_{L^{\infty}(\mathbb T^2)}) \] for some universal constant $C>0$. \end{prop} The proof of Proposition \ref{propReg} relies on the fact that the velocity is log-Lipschitz uniformly in $\nu$ and shows that the exponential estimate with loss of \cite{BC94} holds uniformly in viscosity. Our proof uses the stochastic Lagrangian representation formula of \cite{CI08}: \begin{equation}\label{traj} {\rm d} X_t(x) = u^\nu(X_t(x),t){\rm d} t+ \sqrt{2\nu} \ {\rm d} W_t, \qquad X_0(x)=x, \end{equation} yielding the representation formula \begin{equation} \label{stochRep} \omega^\nu(t) = \mathbb{E}\left[\omega_0\circ A_t \right] \end{equation} where back-to-labels map is defined as $A_t= X_t^{-1}$. The noisy Lagrangian picture allows for a nearly direct application of the Theorems and proofs of \cite{BC94,BCD11} to the viscous case. We remark that the uniform Sobolev regularity can be established by similar arguments; if $\omega_0\in (L^\infty\cap W^{s,p})(\mathbb{T}^2)$ then $\omega^\nu(t)\in(L^\infty\cap W^{s(t),p})(\mathbb{T}^2)$ with uniformly bounded norms. The uniform regularity of Proposition \ref{propReg} is used to deduce \begin{cor}\label{cor} Let $\omega_0\in (L^\infty\cap B^{s}_{2,\infty})(\mathbb{T}^2)$ with $s>0$ and let $\omega$ and $\omega^\nu$ solve respectively (\ref{eu}) and (\ref{nse}), with the same initial data $\omega^\nu_0=\omega_0$. Then the $L^p$ convergence of vorticity, for any $p\in[1,\infty)$ and any finite time $T>0$, occurs at the rate \begin{equation} \sup_{t\in [0,T]}\|\omega^\nu(t)- \omega(t)\|_{L^p(\mathbb{T}^2)} \lesssim (\nu T)^{\frac{ s\exp(-2CT\|\omega_0\|_{\infty})}{p(1+ s \exp(-CT\|\omega_0\|_{\infty}))}- }, \end{equation} with the universal constant $C>0$ in Proposition \ref{propReg}. \end{cor} \begin{rem} Recently, the estimate with loss of \cite{BC94} was sharpened for fixed $p\in(1,\infty)$ in \cite{BN19} where it is shown that the propagated regularity decays inversely with time rather than exponentially, i.e. $ \tilde{s}(t) = {s}/{(1+Ct p s )}$ for some universal constant $C>0$. See Corollary 1.4 of \cite{BN19}. This improvement is accomplished by taking greater advantage of the uniform exponential integrability of the velocity gradient stated in Lemma \ref{explemm} below. The stochastic representation can also be used to show uniform boundedness of the vorticity in $\omega^\nu(t)\in(L^\infty\cap B^{\tilde{s}(t)}_{p,\infty})$ as was done in Proposition \ref{propReg}. We omit details here, which are straightforward extensions of the proofs of \cite{BN19}. This extension can lead to an improved rate in Corollary \ref{cor}. \end{rem} Corollary \ref{cor} applies in particular to the to inviscid limits of vortex patches with non-smooth boundary. Indeed, lemma 3.2 of \cite{CW2} shows that if $\omega_0= \chi_\Omega$ is the characteristic function of a bounded domain whose boundary has box-counting (fractal) dimension $D$ not larger than the dimension of space $d=2$, i.e. $d_F(\partial \Omega ):= D<2$, then $\omega_0\in B_{p,\infty}^{(2-D)/p}(\mathbb{T}^2)$. Proposition \ref{propReg} then shows that some degree of fractional Besov regularity of the solution $\omega^\nu(t)$ is retained uniformly in viscosity for any finite time $T<\infty$ and corollary \ref{cor} provides a rate depending only $D, T$ and $p$ at which the vanishing viscosity limit holds, removing therefore the need for the additional assumptions on the solution imposed in \cite{CW2}. \section{Proof } \begin{proof}[Proof of Theorem \ref{Thm}] It suffices to prove that \begin{equation} \lim_{\nu\to 0} \sup_{t\in [0,T]} \|\omega^\nu(t)-\omega(t)\|_{L^2(\mathbb{T}^2)}=0. \end{equation} Indeed, convergence in $L^p$ for any $p\in [2,\infty)$ then follows from interpolation and boundedness in $L^\infty$: \begin{align} \|\omega^\nu(t)-\omega(t)\|_{L^p(\mathbb{T}^2)} &\leq 2\Omega_{\infty}^{\frac{p-2}{p}}\|\omega^\nu(t)-\omega(t)\|_{L^2(\mathbb{T}^2)}^{\frac{2}{p}}. \end{align} In order to establish strong $L^\infty_tL^2_x$ convergence for arbitrary finite times $T$, it is enough to the convergence for a short time which depends only on a uniform $L^{\infty}$ bound on the initial vorticity: \begin{prop}\label{prop} Let $\omega$ and $\omega^\nu$ solve (\ref{eu}) and (\ref{nse}) respectively, with initial data (\ref{ideu}) and (\ref{idnse}). Assume that the Navier-Stokes initial data converge uniformly in $L^{2}(\mathbb T^2)$ \begin{equation} \lim_{\nu\to 0}\|\omega^\nu_0 - \omega_0\|_{L^2(\mathbb T^2)} = 0. \label{limidl2} \end{equation} Assume also that there exists a contant $\Omega_{\infty}$ such that the initial data are uniformly bounded in $L^{\infty}(\mathbb T^2)$: \begin{equation} \sup_{\nu>0}\|\omega_0^{\nu}\|_{L^{\infty}(\mathbb T^2)}\le \Omega_{\infty}. \label{unifid} \end{equation} Then there exists a constant $C_*$ such that the vanishing viscosity limit holds \begin{equation} \lim_{\nu\to 0} \sup_{t\in [0,T_*]} \|\omega^\nu(t)-\omega(t)\|_{L^2(\mathbb{T}^2)}=0 \end{equation} on the time interval $[0, T_*]$ where \begin{equation} T_*= (C_* \Omega_{\infty})^{-1}. \label{tstar} \end{equation} \end{prop} Once this proposition is established, the proof of theorem \ref{Thm} follows by dividing the time interval $[0,T]$ in subintervals $$ [0,T] = [0,T_*] \cup [T_*, 2T_*] \cup\cdots $$ where $T_*$ is determined from the uniform bound (\ref{omegafty}), and applying proposition \ref{prop} to each interval, with initial data $\omega(nT_*)$, and respectively $ \omega^\nu( nT_*)$. As there is no required rate of convergence for the initial data in proposition \ref{prop}, theorem \ref{Thm} follows. \begin{proof}[Proof of Proposition \ref{prop}] We introduce functions $\omega_\ell$ and $\omega_\ell^\nu$ which are the unique solutions of the following \emph{linear} problems. We fix $\ell>0$ and let \begin{align}\label{omell} \partial_t \omega_\ell + u\cdot \nabla \omega_\ell &=\varphi_\ell *g , \qquad\qquad\quad \ \ \ \omega_\ell(0)=\varphi_\ell *{\omega}_0,\\ \partial_t \omega_\ell^\nu+ u^\nu\cdot \nabla \omega_\ell^\nu &=\nu \Delta \omega_\ell^\nu+\varphi_\ell *g , \qquad \omega_\ell^\nu(0)=\varphi_\ell *{\omega}_0^\nu, \label{omellnu} \end{align} where $\varphi_\ell $ is a standard mollifier at scale $\ell$ and where $u$ and $u^\nu$ are respectively the unique solutions of Euler and Navier-Stokes equations. Note that the solutions to the linear problems \eqref{omell} and \eqref{omellnu} exist globally and are unique because the Yudovich velocity field $u$ is log-Lipshitz. We observe that we have \begin{align*} \|\omega^\nu(t)-\omega(t)\|_{L^2(\mathbb{T}^2)} &\leq \|\omega(t)-\omega_\ell(t)\|_{L^2(\mathbb{T}^2)} + \|\omega^\nu(t)-\omega^\nu_\ell(t)\|_{L^2(\mathbb{T}^2)} \\ &\qquad + \|\omega^\nu_\ell(t)-\omega_\ell(t)\|_{L^2(\mathbb{T}^2)}. \end{align*} Because the equations for $\omega_\ell,\omega_\ell^\nu$ and, respectively $\omega,\omega^\nu$ share the same incompressible velocities, we find \begin{align} \|\omega(t)-\omega_\ell(t)\|_{L^2(\mathbb{T}^2)}&\leq \|{\omega}_0-\varphi_\ell *{\omega}_0\|_{L^2(\mathbb{T}^2)} +\int_0^t \| g(s)- \varphi_\ell * g(s)\|_{L^2(\mathbb{T}^2)}{\rm d} s,\\ \|\omega^\nu(t)-\omega^\nu_\ell(t)\|_{L^2(\mathbb{T}^2)}&\leq \|{\omega}_0^\nu-\varphi_\ell *{\omega}_0^\nu\|_{L^2(\mathbb{T}^2)} +\int_0^t \| g(s)- \varphi_\ell * g(s)\|_{L^2(\mathbb{T}^2)}{\rm d} s. \end{align} As mollification can be removed strongly in $L^p$, the two terms in the right hand sides converge to zero as $\ell,\nu\to 0$, in any order. It remains to show that \begin{align} \label{convFixell} \lim_{\nu\to 0}\sup_{t\in [0,T_*]} \|\omega^\nu_\ell(t)-\omega_\ell(t)\|_{L^2(\mathbb{T}^2)}\to 0 \end{align} for fixed $\ell$. In order to establish this, we use two auxilliary results. The first one is a general statement about the Biot-Savart law in dimension two. \begin{lemma}\label{explemm} Let $\omega\in L^\infty(\mathbb{T}^2)$ and let $u$ be obtained from $\omega$ by the Biot-Savart law \[ u= K[\omega] = \nabla^\perp (-\Delta)^{-1} \omega. \] There exist constants $\gamma>0$ (nondimensional and $C_K$ (with units of area) such that \begin{equation}\label{expinte} \int_{\mathbb{T}^2} \exp\left\{ \beta |\nabla u(x)|\right\} {\rm d} x \leq C_K \end{equation} holds for any $\beta>0$ such that \begin{equation} \beta\|\omega\|_{L^{\infty}(\mathbb T^2)} \le \gamma. \label{betagamma} \end{equation} \end{lemma} \begin{proof}[Proof of Lemma \ref{explemm}] The bound \eqref{expinte} holds due to the fact that Calderon-Zygmund operators map $L^\infty$ to BMO \cite{Stein}, $\omega\in L^\infty \mapsto \nabla u = \nabla K[u] \in BMO$, and from the John-Nirenberg inequality \cite{Nirenberg} for BMO functions. We provide below a direct and elementary argument (modulo a fact about norms of singular intergal operators), for the sake of completeness. We recall that there exists a constant $C_*$ so that for all $p\geq 2$, \begin{equation}\label{BSbnd} \|\nabla K[v]\|_{L^p(\mathbb{T}^2)}=\|\nabla \otimes \nabla (-\Delta)^{-1} v\|_{L^p(\mathbb{T}^2)}\leq C_* p \|v\|_{L^p(\mathbb{T}^2)}. \end{equation} (See \cite{Stein}). The dependence of \eqref{BSbnd} on $p$ is the important point. Thus, \begin{align} \nonumber \int_{\mathbb{T}^2} e^{ \beta |\nabla u|} {\rm d} x&= \sum_{p=0 }^\infty \beta^p \frac{\|\nabla u\|^p_{L^p(\mathbb{T}^2)} }{p!} \leq \sum_{p=0 }^\infty \frac{ \left( C_* \beta \|\omega\|_{L^p(\mathbb{T}^2)}\right)^pp^p}{p!}\\ &\leq |\mathbb{T}^2| \sum_{p=0 }^\infty \frac{ \left( C_* \beta \|\omega\|_{L^\infty(\mathbb{T}^2)}\right)^pp^p}{p!}. \end{align} This is a convergent series provided $ C_* \beta \|\omega\|_{L^\infty(\mathbb{T}^2)}<1/e$. Indeed, this can be seen using Stirling's bound $n!\geq \sqrt{2\pi} n^{n+1/2} e^{-n} $ which yields \begin{equation} \sum_{p=0}^\infty \frac{c^pp^p}{p!} \leq 1+\sum_{p=1}^\infty \frac{p^{-1/2}}{\sqrt{2\pi} }(ce)^{p} \leq \frac{1}{1-ce}, \quad \text{provided}\quad c \in[0, 1/e) \end{equation} where $c:=C_* \beta \|\omega\|_{L^\infty(\mathbb{T}^2)}$. We may take thus \begin{equation} \gamma = (2C_*e)^{-1}, \quad C_K = 2\left |\mathbb T^2\right|. \label{gammack} \end{equation} The constant $\gamma$ depends on the Biot-Savart kernel and is nondimensional, the constant $C_K$ then is proportional to the area of the domain. \end{proof} The second auxilliary result concerns scalars transported and amplified by a velocity with bounded curl in two dimensions. \begin{lemma}\label{scalalemm} Let $u:=u(x,t)$ be divergence free and $\omega:=\nabla^\perp \cdot u \in L^\infty(0,T;L^\infty(\mathbb{T}^2))$ with \begin{equation} \sup_{0\le t\le T}\|\omega(t)\|_{L^{\infty}(\mathbb T^2)}\le \Omega_{\infty}. \label{omegfty} \end{equation} Consider a nonnegative scalar field $\theta:=\theta(x,t)$ satisfying the differential inequality \begin{equation}\label{scalareqn} \partial_t \theta + u \cdot \nabla \theta - \nu \Delta \theta \leq |\nabla u |\theta + f, \end{equation} with initial data $\theta|_{t=0}=\theta_0\in L^\infty(\mathbb{T}^2)$, and forcing $f\in L^\infty(0,T;L^\infty(\mathbb{T}^2))$. Let $\gamma>0$ be the constant from Lemma \ref{explemm}. Then, for any $p>1$ and the time $T(p)= \frac{\gamma(p-1)}{2p\Omega_{\infty}}$ it holds that \begin{equation} \sup_{t\in [0,T(p)]} \|\theta(t)\|_{L^2(\mathbb{T}^2)} \leq C_1 \|\theta_0\|_{L^{2p}(\mathbb{T}^2)}^{p} + C_2 \end{equation} for some constants $C_1, C_2$ depending only on $p$, $\Omega_{\infty}$ and $\|f\|_{L^\infty(0,T;L^\infty(\mathbb{T}^2))}$. \end{lemma} \begin{proof}[Proof of Lemma \ref{scalalemm}] Let $p:=p(t)$ with $p(0)=p_0$ and time dependence of $p(t)$ to be specified below. Consider \begin{align} \nonumber \frac{1}{2}\frac{{\rm d}}{{\rm d} t} \int_{\mathbb{T}^2} |\theta|^{2p(t)} {\rm d} x &= p'(t) \int_{\mathbb{T}^2} \ln|\theta| |\theta|^{2p(t)} {\rm d} x+ p(t)\int_{\mathbb{T}^2} |\theta|^{2p(t)-2} \theta \partial_t \theta {\rm d} x \\ \nonumber &\leq p'(t) \int_{\mathbb{T}^2} \ln|\theta| |\theta|^{2p(t)} {\rm d} x- p(t)\int_{\mathbb{T}^2} |\theta|^{2p(t)-2} \theta u \cdot \nabla \theta {\rm d} x \\ \nonumber &\quad + \nu p(t)\int_{\mathbb{T}^2} |\theta|^{2p(t)-2} \theta \Delta \theta {\rm d} x + p(t)\int_{\mathbb{T}^2} |\theta|^{2p(t)-2} |\nabla u| \theta^2 {\rm d} x \\ &\quad + p(t)\int_{\mathbb{T}^2} |\theta|^{2p(t)-2} \theta f {\rm d} x. \end{align} We now use the following facts \begin{align} \int_{\mathbb{T}^2} |\theta|^{2p-2} \theta f {\rm d} x &\leq C\|f\|_{L^\infty(0,T;L^\infty(\mathbb{T}^2))} \|\theta\|_{2p}^{2p-1},\\ p\int_{\mathbb{T}^2} |\theta|^{2p-2} \theta u \cdot \nabla \theta {\rm d} x&= \frac{1}{2} \int_{\mathbb{T}^2} u\cdot \nabla( |\theta|^{2p}) {\rm d} x =0,\\ \nu \int_{\mathbb{T}^2} |\theta|^{2p-2} \theta \Delta \theta {\rm d} x &= -\nu (2p-1) \int_{\mathbb{T}^2} |\theta|^{2p-2} |\nabla \theta|^2{\rm d} x \leq 0. \end{align} In the second equality we used the fact that the velocity is divergence free. Altogether we find thus \begin{align} \nonumber \frac{1}{2}\frac{{\rm d}}{{\rm d} t} \|\theta(t)\|_{2p(t)}^{2p(t)} {\rm d} x &\leq p'(t) \int_{\mathbb{T}^2} \ln|\theta| |\theta|^{2p(t)} {\rm d} x \\ &\qquad + p(t)\int_{\mathbb{T}^2} |\theta|^{2p(t)} |\nabla u| {\rm d} x + p(t) \|f\|_{L^\infty} \|\theta\|_{2p}^{2p-1}. \end{align} We now use the following elementary inequality: for $a\in \mathbb{R}$ and $b>0$, \begin{equation}\label{elemIq} ab \leq e^a+ b \ln b-b. \end{equation} In fact, we use only that $ab \leq e^a+ b \ln b$. The inequality \eqref{elemIq} is proved via calculus and follows because the Legendre transform of the convex function $b\ln b-b+1$ is $e^a-1$. Setting $a= \beta |\nabla u|$ and $b= \frac{1}{\beta} |\theta|^{2p}$, applying \eqref{elemIq} and Lemma \ref{explemm} we obtain \begin{align} \nonumber \frac{1}{2}\frac{{\rm d}}{{\rm d} t}\|\theta(t)\|_{2p(t)}^{2p(t)} &\leq p'(t) \int_{\mathbb{T}^2} \ln|\theta| |\theta|^{2p} {\rm d} x+ \frac{p(t)}{\beta} \int_{\mathbb{T}^2} \ln(\beta^{-1}|\theta|^{2p}) |\theta|^{2p} {\rm d} x\\ \nonumber &\qquad +p(t) \int_{\mathbb{T}^2} e^{ \beta |\nabla u|} {\rm d} x+ C p(t) \|f\|_{L^\infty} \|\theta\|_{2p}^{2p-1}\\ \nonumber &\leq\left( p'(t)+ \frac{2p(t)^2}{\beta}\right) \int_{\mathbb{T}^2} \ln|\theta| |\theta|^{2p} {\rm d} x + \frac{p(t)}{\beta}\ln (\beta^{-1})\|\theta(t)\|_{2p}^{2p}\\ &\qquad +p(t)C_K+ Cp(t) \|f\|_{L^\infty} \|\theta\|_{2p}^{2p-1}, \end{align} where $C_K$ is the constant from Lemma \ref{explemm} and $\beta = \frac{\gamma}{\Omega_{\infty}}$ depends on the bound for $\|\omega(t)\|_{L^{\infty}}$. We now choose $p$ to evolve according to \begin{equation}\label{peq} p'(t)=- 2\beta^{-1} p(t)^2, \ \ p(0)=p_0\quad \implies \quad p(t)= \frac{\beta p_0}{\beta + 2p_0 t}. \end{equation} Note that $p(t)$ is a positive monotonically decreasing function of $t$. Let the time $t_*$ defined by $t_*=T(p_0):=\beta(p_0-1)/2p_0$ be such that $p(t_*)=1$. Then $p(t)\in [1,p_0]$ for all $t\in [0,t_*]$. Note also from (\ref{peq}) that \[ \int_0^t p(s)ds = \log\left(\frac{p_0}{p(t)}\right)^{2\beta} = \log\left(1+ \frac{2p_0 t}{\beta}\right)^{\fr2\beta}. \] Defining $m(t)=\frac{1}{2} \|\theta(t)\|_{2p(t)}^{2p(t)}$ and using \eqref{peq} we have the differential inequality \begin{equation} m'(t) \leq p(t)(C_1 m(t) + C_2) \implies C_1 m(t)+ C_2 \leq (C_1 m_0 + C_2)\left (1+ \frac{2p_0 t}{\beta}\right)^{\frac{2C_1}{\beta}} \end{equation} with $C_1$ and $C_2$ depending on $\|f\|_{L^\infty(0,T; L^\infty(\mathbb{T}^2))}$, $p_0$, $C_K$ and $\beta$. Thus \[ m(t) \le m_0 \left(1+ \frac{2p_0 t}{\beta}\right)^{\frac{2C_1}{\beta}} + \frac{C_2}{C_1}\left[\left (1+ \frac{2p_0 t}{\beta}\right)^{\frac{2C_1}{\beta}} -1\right]. \] Note that $ {p_0}/{p(t)}= 1 + 2p_0\beta^{-1} t$ is increasing on $[0,t_*]$ from $1$ to ${p_0}/{p(t_*)}=p_0$. Consequently \begin{equation} \|\theta(t)\|_{2p(t)} \leq C_1\|\theta_0\|_{2p_0}^{p_0} + C_2 \end{equation} where the constants $C_1$ and $C_2$ have been redefined but the dependence on parameters is the same. As $p(t)\in[1,p_0]$ for all $t\in [0,t_*]$ we have that $\|\theta(t)\|_{2}\leq \|\theta(t)\|_{2p(t)}$ and we obtain \begin{equation} \sup_{t\in [0,t_*]} \|\theta(t)\|_{2} \leq C_1 \|\theta_0\|_{2p_0}^{p_0} + C_2, \end{equation} which completes the proof. \end{proof} A similar idea to our Lemma \ref{scalalemm} was used in \cite{EJ17}, Lemma 3. We apply our two lemmas to the two dimensional linearized Euler and Navier-Stokes equations to obtain uniform boundedness of vorticity gradients for short time. \begin{lemma}\label{lem} Fix $\ell>0$ and let $\omega_\ell$ and $\omega_\ell^\nu$ solve \eqref{omell} and \eqref{omellnu} respectively. Then there exists a constant $C_*$ and a constant $C_\ell<\infty$ depending only on $\ell$, the forcing norm $\|g\|_{L^\infty(0,T;L^\infty(\mathbb{T}^2))}$, and the uniform bound on solutions given in (\ref{omegafty}) such that for $T_*\le (C_* \Omega_{\infty})^{-1}$, we have that \begin{equation} \sup_{t\in [0,T_*]} \left(\|\omega_\ell(t)\|_{H^1} + \|\omega_\ell^\nu(t)\|_{H^1} \right)\leq C_\ell. \end{equation} \end{lemma} \begin{proof}[Proof of Lemma \ref{lem}] We focus on proving a viscosity independent bound for $ \|\omega_\ell^\nu(t)\|_{H^1}$. The proof for $\|\omega_\ell(t)\|_{H^1}$ is the same, setting $\nu=0$. We show that $|\nabla \omega_\ell^\nu|$ obeys \eqref{scalareqn}. Differentiating \eqref{omellnu}, we find \begin{equation} (\partial_t + u^\nu\cdot \nabla)\nabla \omega_\ell^\nu+\nabla u^\nu \cdot \nabla \omega_\ell^\nu =\nu \Delta (\nabla \omega_\ell^\nu)+ \nabla (\varphi_\ell *g). \end{equation} A standard computation shows that $|\nabla \omega_\ell^\nu|$ satisfies \begin{equation} (\partial_t + u^\nu\cdot \nabla-\nu \Delta) |\nabla \omega_\ell^\nu| \leq |\nabla u| |\nabla \omega_\ell^\nu|+ | \nabla (\varphi_\ell *g)| \end{equation} which is a particular case of the scalar inequality \eqref{scalareqn} with $\theta=|\nabla \omega_\ell^\nu|$, initial data $\theta_0=|\nabla (\varphi_\ell *{\omega}_0^\nu)|\in L^\infty(\mathbb{T}^2)$ and forcing $f= | \nabla (\varphi_\ell *g)|\in L^\infty(0,T;L^\infty(\mathbb{T}^2))$, as claimed. Applying Lemma \ref{scalalemm}, we find that for any $p>1$ (e.g. $p=2$) we have \begin{align}\nonumber \sup_{t\in [0,T_*]} \|\omega_\ell^\nu(t)\|_{H^1} &=C_1 \frac{1}{\ell^p}\left( \int_{\mathbb{T}^2} | {\omega}_0^\nu*(\nabla\varphi)_\ell |^{2p} {\rm d} x\right)^{1/2}+C_2\\ \label{finalbnd} & \lesssim C_\ell \|{\omega}_0^\nu\|_{L^\infty(\mathbb{T}^2)}^p\lesssim C_\ell \Omega_{\infty}^p. \end{align} The constant $C_\ell$ depends on $\Omega_{\infty}$. It diverges with the mollification scale $\ell$, through the prefactor ${\ell^{-p}}$ and through the dependence on $\|\nabla(\varphi_\ell * g)\|_{L^\infty} \lesssim \ell^{-1} \|g\|_{L^\infty}$. The important point however is that \eqref{finalbnd} holds uniformly in viscosity, completing the proof. \end{proof} We return now now to the proof of the main theorem. Using Lemma \ref{lem}, the difference energy obeys \begin{align} \nonumber \frac{{\rm d}}{{\rm d} t} \|\omega_\ell^\nu- \omega_\ell\|_{L^2(\mathbb{T}^2)}^2 &= -\int_{\mathbb{T}^2} (u^\nu-u) \cdot \nabla \omega^\nu_\ell (\omega_\ell^\nu- \omega_\ell) {\rm d} x \\ \nonumber &\qquad - \nu \int_{\mathbb{T}^2}|\nabla \omega_\ell^\nu|^2 {\rm d} x + \nu \int_{\mathbb{T}^2} \nabla \omega_\ell^\nu\cdot \nabla \omega_\ell {\rm d} x\\ \nonumber &\leq 4 \Omega \|u^\nu-u\|_{L^{2}} \| \nabla \omega_\ell^\nu\|_{L^{2}} + \nu \| \nabla \omega_\ell^\nu\|_{L^2} \| \nabla \omega_\ell\|_{L^2} \\ &\lesssim C_\ell \|u^\nu-u\|_{L^\infty(0,T;L^{2}(\mathbb{T}^2))} + \nu C_\ell ^2. \end{align} Integrating we find \begin{equation}\label{omdiff} \|\omega_\ell^\nu- \omega_\ell\|_{L^2}^2\lesssim \|\varphi_\ell*( {\omega}_0^\nu- {\omega}_0)\|_{L^2}^2+ C_\ell T \|u^\nu-u\|_{L^\infty(0,T;L^{2}(\mathbb{T}^2))} + \nu C_\ell ^2 T . \end{equation} To conclude the proof we must show that, at fixed $\ell>0$, we have $\lim_{\nu\to 0} \|\omega_\ell^\nu- \omega_\ell\|_{L^2(\mathbb{T}^2)}=0$. Recall that by our assumption \eqref{limidl2} we have that $\lim_{\nu \to 0} \|\omega^\nu_0 - \omega_0\|_{L^2(\mathbb{T}^2)}\to 0$. Thus we need only establish strong convergence of the velocity in $L^2(0,T;L^2(\mathbb{T}^2))$. If $g=0$ and $u^\nu_0=u_0$, this is a consequence of Theorem 1.4 of \cite{Chemin}. Below is a generalization of \cite{Chemin} which applies in our setting and is proved by a different argument. \begin{lemma}{\label{chem}} Let $\omega_0\in L^{\infty}(\mathbb T^2)$. There exist constants $U$, $\Omega_2$ and $K$ (see below (\ref{U}), (\ref{omegap}), (\ref{K})) depending on norms of the initial data and of the forcing such that the difference $v = u^{\nu}-u$ of velocities of solutions (\ref{eu}) and (\ref{nse}) obeys \begin{equation} \|v(t)\|_{L^2}^2 \le 3U^2K^{\frac{5(t-t_0)\Omega_{\infty}}{\gamma}} \left (\frac{\|v(t_0)\|_{L^2(\mathbb T^2)}^2}{U^2} + \gamma \frac{\Omega_2^2}{U^2\Omega_{\infty}}\nu \right)^{1-\frac{5(t-t_0)\Omega_{\infty}}{\gamma}} \label{vineq} \end{equation} for all $0\le t_0\le t $. By iterating the above, we obtain \begin{equation}\label{vineqLT} \|v(t)\|_{L^2}^2 \leq 20U^2K^{1-e^{-{10 t \Omega_{\infty}}/{\gamma}}} \left (\frac{\|v(0)\|_{L^2(\mathbb T^2)}^2}{U^2} + \gamma \frac{ \Omega_2^2}{U^2\Omega_{\infty}}\nu \right)^{e^{-\frac{10 t \Omega_{\infty}}{\gamma}}} \end{equation} provided that $\|v(0)\|_{L^2(\mathbb T^2)}^2+ {\gamma}\nu \Omega_2^2/{\Omega_{\infty}}\leq 9K U^2$. \end{lemma} \begin{rem}[Continuity of Solution Map] At zero viscosity, Lemma \ref{chem} establishes H\"{o}lder continuity of the Yudovich (velocity) solution map. Specifically, denoting $u_t:= S_t^v(u_0)$ and setting $\nu=0$, a consequence of Lemma \ref{chem} is that $\| S_t^v(u_0)- S_t^v(u_0')\|_{L^2(\mathbb{T}^2)} \leq C \|u_0-u_0'\|_{L^2(\mathbb{T}^2)}^{\alpha(t)}$ where $\alpha(t):= e^{-ct} $ and $c, C>0$ are appropriate constants. This fact is used to prove Corollary \ref{contcor}. {It is worth further remarking that the condition on the data $\|v(0)\|_{L^2(\mathbb T^2)}^2 \leq 9K U^2$ required for the above estimate to hold is $O(1)$ (data need not be taken very close).} \end{rem} \begin{proof}[Proof of Lemma \ref{chem}] The proof proceeds in two steps. \\ \vspace{-3mm} \noindent \textbf{Step 1: Short time bound.} The proof of the lemma starts from the equation obeyed by the difference $v$, \[ \partial_t v + u^{\nu}\cdot\nabla v + v\cdot\nabla u + \nabla p = \nu\Delta v + \nu \Delta u \] leading to the inequality \begin{equation} \frac{d}{dt}\|v\|_{L^2}^2 + \nu \|\nabla v\|^2_{L^2} \le \nu \|\nabla u\|^2_{L^2} + 2 \int |\nabla u| |v|^2{\rm d} x \label{env} \end{equation} which is a straightforward consequence of the equation, using just integration by parts. We use the bound $\Omega_{\infty}$ (\ref{omegafty}) for the vorticity of the Euler solution. We also use a bound for the $L^2$ norms \begin{equation} \sup_{0\le t\le T}\left(\|u^{\nu}(t)\|_{L^2(\mathbb T^2)} + \|u(t)\|_{L^2(\mathbb T^2)}\right)\le U, \label{U} \end{equation} which is easily obtained from energy balance. We use also bounds for $L^p$ norms of vorticity, \begin{equation} \Omega_p = \sup_{0\le t\le T}\|\omega(t)\|_{L^p(\mathbb T^2)} \le \Omega_\infty. \label{omegap} \end{equation} We split the integral \[ \int |\nabla u| |v|^2 {\rm d} x = \int_B |\nabla u| |v|^2{\rm d} x + \int_{\mathbb T^2\setminus B} |\nabla u| |v|^2{\rm d} x \] where \[ B = \{ x\left |\right.\; |v(x,t)| \ge MU\} \] with $M$ to be determined below. Although $B$ depends in general on time, it has small measure if $M$ is large, \[ \left | B\right | \le M^{-2}. \] The constant $M$ has dimensions of inverse length. We bound \begin{equation}\label{termbd} 2\int_B|\nabla u||v|^2{\rm d} x \leq 2 \| \nabla u\|_{L^2} \| v\|_{L^4}^2\leq 2 |B|^{\frac{1}{4}}\| \nabla u\|_{L^4} \|v(t)\|_{L^4}^2 \end{equation} where we used $\int_B|\nabla u|^2dx \le |B|^{\frac{1}{2}}\| \nabla u\|_{L^4}^2$. We now use the fact that we are in Yudovich class and Ladyzhenskaya inequality to deduce \[ \|v(t)\|_{L^4}^2 \le C\|v(t)\|_{L^2}[\|\omega_0\|_{L^2} + \|g\|_{L^1(0,T; L^2)}] \le CU\Omega_2 \] and we use also \[ \|\nabla u\|_{L^4}\le [C\|\omega_0\|_{L^4} + \|g\|_{L^1(0,T; L^4)}] = \Omega_4 \] to bound \eqref{termbd} by \begin{equation} 2\int_B|\nabla u||v|^2{\rm d} x \le CU\Omega_2\Omega_4M^{-\frac{1}{2}}, \label{smallb} \end{equation} We nondimensionalize by dividing by $U^2$ and we multiply by $\beta = {\gamma}/{\Omega_{\infty}}$. The quantity \begin{equation} y(t) = \frac{\|v(t)\|_{L^2{(\mathbb T^2)}}^2}{U^2} \label{y} \end{equation} obeys the inequality \begin{equation} \beta\frac{{\rm d} y}{{\rm d} t} \le \beta\nu\frac{\Omega_2^2}{U^2} + C\beta\Omega_4\frac{\Omega_2}{U}M^{-\frac{1}{2}} + 2\int_{\mathbb T^2\setminus B} \beta|\nabla u|\frac{|v|^2}{U^2}{\rm d} x. \label{intery} \end{equation} We write the term \begin{equation} 2\int_{\mathbb T^2\setminus B} \beta |\nabla u| |v|^2 U^{-2} {\rm d} x = 2\int_{\mathbb T^2\setminus B}(\beta|\nabla u| + \log\epsilon + \log\frac{1}{\epsilon})|v|^2 U^{-2} {\rm d} x\end{equation} with $\epsilon$ (with units of inverse area) to be determined below. We use the inequality (\ref{elemIq}) and Lemma \ref{explemm} with \[ a = \beta |\nabla u| + \log\epsilon, \qquad b= \frac{|v|^2}{U^2} \] to deduce \begin{equation} 2\int_{\mathbb T^2\setminus B} \beta |\nabla u| |v|^2 U^{-2}{\rm d} x\le 2\epsilon C_K + 2\log\frac{M^2}{\epsilon} y(t). \label{largeb} \end{equation} Inserting (\ref{largeb}) in (\ref{intery}) we obtain \begin{equation} \beta\frac{{\rm d} y}{{\rm d} t} \le F + \log\left(\frac{M^2}{\epsilon}\right )y(t) \label{yneqdt} \end{equation} with \begin{equation} F = \beta\nu\frac{\Omega_2^2}{U^2} + C\beta\Omega_4\frac{\Omega_2}{U}M^{-\frac{1}{2}} + 2\epsilon C_K. \label{f} \end{equation} Note that $F$ and $\frac{M^2}{\epsilon}$ are nondimensional. From (\ref{yneqdt}) we obtain immediately \begin{equation} y(t) \le \left(\frac{M^2}{\epsilon}\right)^{\frac{t-t_0}{\beta}} y(t_0) + \frac{F}{\log\left(\frac{M^2}{\epsilon}\right )}\left(\left(\frac{M^2}{\epsilon}\right)^{\frac{t-t_0}{\beta}} -1\right). \label{yneq} \end{equation} We choose $M$ such that \begin{equation} C\beta\Omega_4\frac{\Omega_2}{U}M^{-\frac{1}{2}} = \beta\nu\frac{\Omega_2^2}{U^2} + y(t_0) \label{Mchoice} \end{equation} and we choose $\epsilon$ such that \begin{equation} 2\epsilon C_K = \beta\nu\frac{\Omega_2^2}{U^2} + y(t_0). \label{epsilonchoice} \end{equation} These choices imply \begin{equation} F = 3\beta\nu\frac{\Omega_2^2}{U^2} + 2y(t_0). \label{fexplicit} \end{equation} Then we see that \begin{equation} \Gamma = \frac{M^2}{\epsilon} = 2C_K\left(C\beta\Omega_4\frac{\Omega_2}{U}\right)^4\times \left (\beta\nu\frac{\Omega_2^2}{U^2} + y(t_0)\right)^{-5}. \label{Gamma} \end{equation} Taking without loss of generality $\log \Gamma \ge 1$, we have from (\ref{yneq}) \begin{align}\nonumber y(t) &\le 3\left (y(t_0) + \beta \nu \frac{\Omega_2^2}{U^2}\right)\Gamma^{\frac{t-t_0}{\beta}}\\ & \le 3\left (y(t_0) + \beta \nu \frac{\Omega_2^2}{U^2}\right)^{1-\frac{5(t-t_0)}{\beta}}\times \left(2C_K\left(C\beta\Omega_4\frac{\Omega_2}{U}\right)^4\right)^{\frac{5(t-t_0)}{\beta}}. \label{ygamma} \end{align} Recalling that $\beta = {\gamma}/{\Omega_{\infty}}$ and denoting the nondimensional constant \begin{equation} K = 2C_K\left(C\beta\Omega_4\frac{\Omega_2}{U}\right)^4 \label{K} \end{equation} we established \begin{equation} \frac{\|v(t)\|^2}{U^2} \le 3 K^{\frac{5(t-t_0)\Omega_{\infty}}{\gamma}}\left (\frac{\|v(t_0)\|_{L^2(\mathbb T^2)}^2}{U^2} + \beta \nu \frac{\Omega_2^2}{U^2}\right)^{1-\frac{5(t-t_0)\Omega_{\infty}}{\gamma}}. \label{vbound} \end{equation} Thus, we established \eqref{vineq}. \\ \noindent \textbf{Step 2: Long time bound.} With \eqref{vineq} established, we now prove \eqref{vineqLT}. Let $c= {5\Omega_{\infty}}/{\gamma}$, $\Delta t= 1/2c$ and $t_i = t_{i-1}+\Delta t$ and $a_i = \|v(t_i)\|_{L^2}^2/U^2$ for $i\in \mathbb{N}$. Then \eqref{vineq} states \begin{equation} a_i \leq C_1 \left(a_{i-1} + C_2\nu \right)^{1/2}, \qquad i=1, 2, \dots \label{aineq} \end{equation} with $C_1= 3K^{\frac{5\Omega_{\infty}}{2c\gamma}}=3K^{\frac{1}{2}}$ and $C_2= \beta \frac{\Omega_2^2}{U^2}$. We set \begin{equation} \delta_n = \frac{a_i+ C_2\nu}{C_1^2} \label{deltan} \end{equation} and observe that (\ref{aineq}) is \begin{equation} \delta_n \le \sqrt{\delta_{n-1}} + \widetilde{\nu} \label{deltanineq} \end{equation} where \begin{equation} \widetilde{\nu} = \frac{C_2\nu}{C_1^2} \label{widetlidenu} \end{equation} is a nondimensional inverse Reynolds number. It follows then by induction that \begin{equation} \delta_n \le (\delta_0)^{2^{-n}} + \sum_{i=0}^{n-1}(\widetilde{\nu})^{2^{-i}}. \label{deltanind} \end{equation} Indeed, the induction step follows from \begin{equation} \delta_{n+1} \le \sqrt{\delta_n} + \widetilde{\nu} \label{indstep} \end{equation} and the subadditivity of $\lambda\mapsto \sqrt{\lambda}$. If \begin{equation} \widetilde{\nu} \le \frac{1}{\sqrt{5}-1} \label{widetildecond} \end{equation} then the iteration (\ref{deltanineq}) starting from $0<\delta_0< r$ where $r$ is the positive root of the equation $x^2-x-\widetilde{\nu} = 0$, remains in the interval $(0,r)$, and for any $n$, $\delta_n$ obeys (\ref{deltanind}). We observe that \begin{equation} \sum_{i=0}^{n-1}(\widetilde{\nu})^{2^{-i}} = (\widetilde{\nu})^{2^{-n+1}}\left (1 + \cdots + (\widetilde{\nu})^{2^{n-1}}\right)\le \frac{1}{1-\widetilde{\nu}}(\widetilde\nu)^{2^{-n+1}} \label{smallnutilde} \end{equation} and therefore (\ref{vineqLT}) follows from (\ref{deltanind}). We note that the iteration defined with equality in (\ref{deltanineq}) converges as $n\to\infty$ to $r$. {Fixing any $t>0$ and letting $n= \lceil t/\Delta t\rceil = \lceil 2ct\rceil = \lceil{10 t \Omega_{\infty}}/{\gamma}\rceil$ establishes the bound.} \end{proof} Due to assumption \eqref{limidl2} we have that $\lim_{\nu \to 0} \|u^\nu_0 - u_0\|_{L^2(\mathbb{T}^2)}\to 0$. Lemma \ref{chem} then allows us to conclude from \eqref{omdiff} that $\lim_{\nu \to 0}\sup_{t\in [0,T_*]} \|\omega_\ell^\nu- \omega_\ell\|_{L^2(\mathbb{T}^2)}\to0$ at fixed $\ell>0$ and the proof of proposition \ref{prop} is complete. \end{proof} With the Proposition proved, the proof of the strong convergence of the vorticity in $L^p$ statement in the theorem is established. To obtain convergence of the distribution functions, see Thm 3.6 in \cite{CW2}. \end{proof} \begin{proof}[Proof of Proposition \ref{propReg}] This proof makes use of the the stochastic Lagrangian representation for Navier--Stokes solutions \cite{CI08}, together with the uniform--in--$\nu$ boundedness of vorticity. In light of the Lagrangian representation \eqref{traj}, \eqref{stochRep}, the key ingredient of propagating some degree of fractional regularity on the vorticity is the (uniform) H\"{o}lder regularity of the inverse flow $A_t$. Since the diffusion coefficients on the additive noise on \eqref{traj} are spatially constant, it follows that the results of Chapter 3 of \cite{BCD11} hold realization-by-realization for the stochastic flow $X_t$ and its inverse $A_t$, uniformly in viscosity. This gives uniform bounds on the separation of two trajectories driven by the same realization of Brownian noise, independent of viscosity, thereby establishing spatial H\"{o}lder regularity of the flow. Although straightforward, we include a proof of this statement for completeness. \begin{prop}\label{flowreg} There exists a unique measure-preserving stochastic flow of homeomorphisms solving \eqref{traj}. This flow and the back-to-labels map are continuous flows $X, A$ which for all $t\in [0,T]$ are uniformly--in--$\nu$ of the class $C^{\alpha(t)}(\mathbb{T}^2)$ with $\alpha(t)= \exp(-Ct/\beta)$ with constants defined in \eqref{loglip}. \end{prop} \begin{proof}[Proof of Proposition \ref{flowreg}] We employ the log-Lipshitz property of $u^\nu$, i.e. there exists an absolute constant $C>2$ such that one has the following uniform-in-viscosity estimate \begin{equation}\label{loglip} |u^\nu(x,t)-u^\nu(y,t)|\leq \frac{C}{\beta} d(x,y) \ln\left(\frac{CC_K}{ d(x,y)^2}\right), \qquad \forall x,y\in \mathbb{T}^2, \end{equation} where $\beta$ and $C_K$ are the constants in Lemma \ref{explemm} which depend only on $\|\omega_0\|_{L^\infty}$. See Lemma A.1 of \cite{BN19}. Here $d(x,y) := \min \{ |x-y-k| \ : \ k\in \mathbb{Z}^d, \ |k|\leq 2\}$ is the geodesic distance on the torus upon the identification $\mathbb{T}^d=[0,1)^d$. Now, due to the spatial uniformity of the noise on the trajectories \begin{equation} {\rm d} X_t(x) = u(X_t(x),t){\rm d} t+\sqrt{2\nu}\ {\rm d} W_t , \qquad X_0(x)=x, \end{equation} we find that the difference has no martingale part and satisfies \begin{equation} {\rm d}\left(X_t(x)- X_t(y)\right) =\left( u^\nu(X_t(x),t)- u^\nu(X_t(y),t)\right){\rm d} t. \end{equation} Upon integration, we obtain the inequality \begin{align}\nonumber d\left(X_t(x),X_t(y)\right) &\leq d(x,y) +\frac{C}{\beta} \int_0^t d\left(X_s(x),X_s(y)\right) \ln\left(\frac{CC_K}{ d\left(X_s(x),X_s(y)\right)^2}\right){\rm d} s, \label{holdertraj} \end{align} for all $x,y\in \mathbb{T}^2$. The solution of this integro-inequality (with a possibly larger constant $C$) is \begin{equation}\label{ineqX} d\left(X_t(x),X_t(y)\right) \leq (C C_K)^{1+ e^{-Ct/\beta}} d(x,y)^{ e^{-Ct/\beta}} \quad \text{a.s.}. \end{equation} Since the bound holds almost surely, this says that the map $X_t(\cdot)$ is H\"{o}lder continuous $C^{\alpha(t)}(\mathbb{T}^2)$ with $\alpha(t)= e^{-Ct/\beta}$ as claimed. We remark that deterministic trajectories in a log-Lipshitz field satisfying \eqref{loglip} satisfy precisely the same upper bound \eqref{ineqX}. To obtain H\"{o}lder regularity of the back-to-labels map, it suffices to note that $A_t$ can be identified with the backwards flow $X_{t,0}$ which solves the following backward stochastic differential equation \begin{equation} \hat{{\rm d}} X_{t,s}(x) = u(X_{t,s}(x) ,s) {\rm d} s +\sqrt{2\nu} \ \hat{{\rm d}} \widehat{W}_s, \qquad X_{t,t}(x)=x, \end{equation} where the $ \hat{{\rm d}} $ indicates the the backward differential and $\widehat{W}_s=W_{t-s}- W_t$ is a Brownian motion adapted to the backward filtration $\hat{\mathcal{F}}_s^t:= \sigma\{ \widehat{W}_u, u\in [0,s]\}$. For a discussion of backward It$\bar{{\rm o}}$ equations, see e.g. \cite{Kunita}. With this identification, one finds as above that for any $t>0$ and all $s\in [0,t]$ one has \begin{equation}\label{ineqA} d\left(X_{t,s}(x),X_{t,s}(y)\right) \leq (C C_K)^{1+ e^{-C(t-s)/\beta}} d(x,y)^{ e^{-C(t-s)/\beta}} \quad \text{a.s.}. \end{equation} By setting $s=0$ we find that $A_t=X_{t,0}$ satisfies the same estimate \eqref{ineqX} as $X_t$ and therefore is H\"{o}lder continuous with the same exponentially decaying exponent. \end{proof} Proceeding forward to obtain uniform bounds we wish to make use of the representation formula \eqref{traj}, \eqref{stochRep}. This requires some regularity on the initial condition, so we replace $\omega_0\in L^\infty(\mathbb{T}^2)$ with a mollification of it $\omega_0*\varphi_\ell \in C^\infty (\mathbb{T}^2)$ for $\ell>0$. All the bounds will be manifestly independent of $\ell$ which can be taken to zero at the end, so we simplify the notation by writing ``$\omega_0$". We continue by following closely the proof of Theorem 3.32 of \cite{BCD11}. In particular, we introduce the space $F_p^s(\mathbb{T}^d)$ (which belongs to the family of Triebel--Lizorkin spaces $F_p^s= F_{p,\infty}^{s}$ provided $p>1$) that is comprised of measurable functions $f\in L^p(\mathbb{T}^d)$ which are finite in the seminorm \begin{align}\nonumber [f]_{F_p^s} := \inf_{g\in L^p(\mathbb{T}^d) } \Big\{ \|g\|_{L^p(\mathbb{T}^d)} \ : &\ \ |f(x)-f(y)|\leq d(x,y)^\alpha(g(x)+g(y)),\\ &\qquad\qquad\qquad\qquad \qquad \ \forall\ x,y\in \mathbb{T}^d\Big\}<\infty\label{Fseminorm} \end{align} where $d(x,y)$ is the distance function on the torus defined above. See Definition 3.30 of \cite{BCD11}. The key of the argument is to understand how composition with a (uniformly) H\"older continuous stochastic diffeomorphism provided by Prop. \ref{flowreg} operate on $F_p^s$. Using the stochastic representation \eqref{stochRep}, \begin{equation} \omega^\nu(t) = \mathbb{E}\left[\omega_0\circ A_t \right], \end{equation} Jensen's inequality, H\"older continuity of the back-to-labels map and the fact that $\omega_0\in F_p^s$ we have \begin{align}\nonumber \frac{|\omega^\nu(x,t)-\omega^\nu(y,t)|}{d(x,y)^{s\alpha}} &= \frac{|\mathbb{E}[\omega_0(A_t(x))-\omega_0(A_t(y)) ]|}{d(x,y)^{s\alpha}}\\ \nonumber &\leq \mathbb{E}\left[ \frac{|\omega_0(A_t(x))-\omega_0(A_t(y))|}{d(x,y)^{s\alpha}} \right] \\ \nonumber &= \mathbb{E}\left[ \frac{|\omega_0(A_t(x))-\omega_0(A_t(y))|}{d\left(A_t(x),A_t(y)\right)^s} \frac{d\left(A_t(x),A_t(y)\right)^s}{d(x,y)^{s\alpha}} \right]\\ \label{Finewt} &\leq \|A_t\|_{C^{\alpha}}^s \mathbb{E} [g(A_t(x))+g(A_t(y))] \end{align} for any $g\in L^p(\mathbb{T}^2)$, where we used that $\omega_0\in F_p^s$ together with the definition \eqref{Fseminorm}. Letting $ \tilde{g}(x):= \mathbb{E} [g(A_t(x))]$. Note that, since $A_t$ is measure preserving and Jensen's inequality, we have $\| \tilde{g}\|_{L^p(\mathbb{T}^2)}\leq \| {g}\|_{L^p(\mathbb{T}^2)}<\infty$. Thus $\tilde{g}\in L^p(\mathbb{T}^2)$ and it follows by linearity of the expectation that the the right-hand-side of \eqref{Finewt} is a $L^p$ function. This shows \begin{equation}\label{Fbnd} [\omega^\nu(t)]_{F_p^{s(t)}} \leq (C K)^{s(1+ e^{-Ct/\beta})} [\omega_0]_{F_p^{s}}, \qquad s(t)= s \exp(-Ct/\beta) \end{equation} where we used the explicit bound on H\"{o}lder norm computed in \eqref{ineqX}. The bound \eqref{Fbnd} holds uniformly in viscosity. In order to connect to some $B^{s}_{p,\infty}$ (which is a larger space) we need to use an embedding for the initial data \begin{equation} B^{s_3}_{p,\infty} \subset B^{s_2}_{p,1}\subset W^{s_2,p}\subset F^{s_1}_p \end{equation} with $s_3>s_2>s_1$. The proposition follows from Lemma 3.31 of \cite{BCD11}, which shows that the Triebel--Lizorkin spaces are continuously embedded in the Besov spaces, i.e. $F_p^s(\mathbb{T}^d) \hookrightarrow B_{p,\infty}^{s}(\mathbb{T}^d)$. \end{proof} \begin{proof}[Proof of Corollary \ref{cor}] We need the following elementary Lemma \begin{lemma}\label{lemma} For any $s>0$ and $f\in B_{2,\infty}^{s}(\mathbb{T}^d)$, the following inequality holds for all $0<s'<s$ \begin{equation}\label{besovineq} \|f\|_{L^2(\mathbb{T}^d)} \leq \|f\|_{H^{-1}(\mathbb{T}^d)}^{s'/(1+s')}\|f\|_{B^{s}_{2,\infty} (\mathbb{T}^d)}^{1/(1+s')}. \end{equation} \end{lemma} \begin{proof} First note that the interpolation inequality \begin{equation}\nonumber \|f\|_{L^2(\mathbb{T}^d)} \leq \|f\|_{H^{-1}(\mathbb{T}^d)}^{s'/(1+s')}\|f\|_{H^{s'} (\mathbb{T}^d)}^{1/(1+s')} \end{equation} which follows from Holder inequality and the Fourier definition of the Sobolev norm. The claim follows from the embedding $B_{p, q}^{s}(\mathbb{T}^d) \subset B_{p,q'}^{s'}(\mathbb{T}^d)$ for $s'<s$ and any $q',q$ (see \S 2.3.2 of \cite{Treibel}) and the identification $H^{s}:= B^{s}_{2,2}$. \end{proof} Proceeding with the proof, applying Lemma \ref{lemma} for all $t\in [0,T]$ we have \begin{align*} \| \omega^\nu(t) -\omega(t)\|_{L^2(\mathbb{T}^2)} &\leq \|\omega^\nu(t) -\omega(t)\|_{H^{-1}(\mathbb{T}^2)}^{\frac{s'}{1+s'} } \| \omega^\nu(t) -\omega(t)\|_{B_{2,\infty}^{s(t)}(\mathbb{T}^2)}^{\frac{1}{1+s'} } \\ &\lesssim \sup_{t\in [0,T]} \|u^\nu(t)-u(t)\|_{L^2(\mathbb{T}^2)}^{\frac{s(t)}{1+s(t)}- } \end{align*} for any $s'< s(t):= {s} \exp(-CT\|\omega_0\|_{\infty})$. In the above, we appealed to Proposition \ref{propReg} to establish uniform--in--$\nu$ boundedness of the solution $\omega^\nu$ in the space $L^\infty(0,t; B_{2,\infty}^{s(t)}(\mathbb{T}^2))$. We now use Lemma \ref{chem} to conclude \begin{align}\nonumber \|\omega^\nu(t) -\omega(t)\|_{L^p(\mathbb{T}^2)} &\leq \|\omega^\nu -\omega\|_{L^\infty(0,T;L^\infty(\mathbb{T}^2))}^{\frac{p-2}{p}}\|\omega^\nu(t) -\omega(t)\|_{L^2(\mathbb{T}^2)}^{\frac{2}{p}}\\ & \lesssim \sup_{t\in [0,T]} \|u^\nu(t)-u(t)\|_{L^2(\mathbb{T}^2)}^{\frac{2s(t)}{p(1+s(t))}- }\lesssim (\nu T)^{\frac{ s\exp(-2CT\|\omega_0\|_{\infty})}{p(1+ s \exp(-CT\|\omega_0\|_{\infty}))} -}. \end{align} This completes our proof. \end{proof} \begin{rem} The stochastic Lagrangian representation of the vorticity offers also an expression for the enstrophy dissipation as the variance of the (randomly sampled) initial data \begin{equation} \nu \int_0^t \int_{\mathbb{T}^2} |\nabla \omega^\nu (t',x)|^2 {\rm d} x {\rm d} t' = \frac{1}{2} \int_{\mathbb{T}^2}{\rm Var} \left[\omega_0^\nu(A_t(x)) \right] {\rm d} x. \end{equation} The above is a special case of the Lagrangian fluctuation dissipation relation for active scalars derived in \cite{DE17}. This relation is easily generalized to incorporate the effect of body forces. A consequence of our Theorem \ref{Thm} is that the enstrophy dissipation vanishes in the high Reynolds number limit, forcing also the variance to become zero. Thus, there is no ``spontaneous stochasticity" of Lagrangian trajectories in the vanishing viscosity limit for 2d Navier-Stokes with initial data in the Yudovich class. \end{rem} \section{Discussion} Predicting the long-time vortex structures in two-dimensional turbulence is of long standing interest, starting with the work on dynamics of point vortices by Onsager \cite{Onsager}. There have been a number of theories developed to this effect. We briefly review the celebrated mean-field theory of Miller \cite{Miller} and Robert \cite{Robert} to give context to our result. The idea is to describe an equilibrium configuration $\omega_{eq}$ satisfying \begin{equation} \omega(t) \ \ \xrightarrow[]{t \to \infty} \ \ \omega_{eq} \end{equation} in some sense. If $\omega(t)$ is an Euler path with bounded initial vorticity, then one has the information \begin{enumerate} \item conservation of energy: \begin{equation}\label{enercons} \|u(t)\|_{L^2(\mathbb{T}^2)} = \|u_0\|_{L^2(\mathbb{T}^2)}, \end{equation} \item conservation of vorticity ``casmirs": for any continuous $f$, \begin{equation}\label{ifinv} I_f:= \int_{\mathbb{T}^2} f(\omega(x,t)) {\rm d} x = \int_{\mathbb{T}^2} f(\omega_0(x)) {\rm d} x. \end{equation} \end{enumerate} For long-time limits of Euler flows, there is a natural candidate object to describe $ \omega_{eq}$. In particular, provided only $\omega_0\in L^\infty(\mathbb{T}^2)$, then $\omega(t)\in L^\infty(\mathbb{T}^2)$ is the unique solution of Euler \cite{Yudovich} and in the weak--$*$ sense \begin{equation}\label{Onsagertheory} \lim_{n\to\infty} \int_{\mathbb{T}^2} \varphi(x) \omega(x,t_n) {\rm d} x = \int_{\mathbb{T}^2} \varphi(x) \bar{\omega}(x){\rm d} x, \qquad \forall \varphi\in L^1(\mathbb{T}^2) \end{equation} for some $\bar{\omega}\in L^\infty(\mathbb{T}^2)$ and some subsequence $t_n\to \infty$ as $n\to \infty$. However, large oscillations can remain in this limit. In particular, the above convergence does not imply for all continuous functions $f$ that $f(\omega(x,t_n))$ converges to $f(\bar{\omega}(x))$ in the same sense, so it is not clear how the information \eqref{enercons} and \eqref{ifinv} can be retained and in what sense. On the other hand, the fundamental theorem of Young measures guarantees \begin{equation}\label{youngmeasurelim} \lim_{n\to\infty} \int_{\mathbb{T}^2} \varphi(x) f(\omega(x,t_n)) {\rm d} x= \int_{\mathbb{T}^2} \varphi(x) \int_{-M}^M f(y) \nu_x({\rm d} y) {\rm d} x \end{equation} with $M=\|\omega_0\|_{L^\infty(\mathbb{T}^2)}$. Note that, having introduced the Young measure $\nu_x({\rm d} y)$, the convergence \eqref{Onsagertheory} holds with \begin{equation} \label{barom} \bar{\omega}(x)= \int_{-M}^M y \nu_x({\rm d} y), \qquad \forall f\in C([-M,M]). \end{equation} Kraichnan developed a theory for the equilibrium distribution $ \bar{\omega}$ discarding most of the information on the casmirs, keeping only conservation of energy and enstrophy \cite{Kraichnan}. However, it was since recognized that invariants involving higher powers of vorticity should not be neglected on compact domains such as $\mathbb{T}^2$. In order to retain as much information about the Euler solution as possible, Miller \cite{Miller} and Robert \cite{Robert} independently suggested that the long-time vorticity distribution resulting from freely decaying two-dimensional turbulence is a Young measure of the form \begin{equation} \label{MRmeasure} \nu_x({\rm d} y) = \rho(x,y) {\rm d} y. \end{equation} These Young measures have the property that their marginal distribution is the (initial) vorticity distribution function \eqref{distconv}, which is left invariant under the Euler flow. Thus, if a measure \eqref{MRmeasure} with the above property can be constructed such that also the energy associated to $\bar{\omega}$ equals that of $\omega_0$, then the information on all ideal invariants is retained at the level of the predicted equilibrium distribution. Miller and Robert provide such a construction.\footnote{We remark that the Miller--Robert theory applies for any compact domain $\Omega\subset\mathbb{R}^2$ with smooth boundary, with the torus $\Omega=\mathbb{T}^2$ as a special case. It is worth noting that convergence of higher-order vorticity moments in the zero viscosity limit on domains with boundaries is -- in general -- false. In fact, if the Euler velocity $u$ is not identically zero along the boundary and no-slip Navier-Stokes solutions converge to these e.g. $u^\nu\rightharpoonup u$ weakly in $L^\infty(0,T;L^2(\Omega))$, then $\limsup_{\nu\to 0} \|\omega^\nu\|_{L^\infty(0,T;L^p(\Omega))}= \infty$ for all $p\in (1,\infty]$ (see Theorem 3.1 of \cite{Kelliher}). If weak convergence fails to hold, then by Kato's energy dissipation condition we know $\limsup_{\nu\to 0} \|\omega^\nu\|_{L^2(0,T;L^2(\Omega))}= \infty$. Thus, unless the Euler solution is identically zero on the boundary, higher moments of vorticity must diverge in the inviscid limit, presenting a great difficulty for the Miller--Robert theory as it applies to inviscid limits.} Specifically, by a Boltzmann counting argument, they showed that the entropy associated with a given density $\rho(x,y)$ of the Young measure has a specific form. Assuming ergodicity at long times, i.e. that the 2D Euler flow is sufficiently chaotic in phase space, they suggested to maximize this entropy subject to the above constraints. The prediction of the theory is the long-time distribution is \begin{equation} \rho(x,y) = \frac{\exp\left(\beta\left[y\bar{\psi}(x)+ \mu(y)\right]\right)}{\int_{-M}^M \exp\left(\beta\left[y\bar{\psi}(x)+ \mu(y)\right]\right){\rm d} y}, \end{equation} where the ``inverse temperature" $\beta$ and ``chemical potential" $\mu(y)$ are Lagrange multipliers to enforce energy conservation and the marginal density $\pi_{\omega_0}[{\rm d} y]$ respectively, and where the stream function $\bar{\psi}$ solves \begin{equation} \Delta \bar{\psi}(x) = \bar{\omega}= \frac{\int_{-M}^M y\exp\left(\beta\left[y\bar{\psi}(x)+ \mu(y)\right]\right){\rm d} y}{\int_{-M}^M \exp\left(\beta\left[y\bar{\psi}(x)+ \mu(y)\right]\right){\rm d} y}. \end{equation} Thus, the prediction is that the expected (average or coarsened) vorticity solves a very particular steady Euler equation $\omega=F(\psi)$ where $\psi$ is the stream function. The function $F$ depends on the distribution $\pi_{\omega_0}[{\rm d} y]$ and the energy $E_0$. It is important to remark that conservation individual casmirs may not survive as $t\to\infty$, but that according to this theory, at a given energy $E_0$, they are forever remembered at the level of the equilibrium distribution. Some numerical simulations have provided corroboratory evidence supporting this theory over competitive ones such as the Onsager-Joyce-Montgomery theory, at least in situations where $\omega_0$ is supported on a finite area \cite{Sommeria}. Whether or not the theory rigorously applies is open. There are two major questions remaining about the domain of applicability of the Miller--Robert theory. The first being whether or not 2D Euler possesses the requisite ergodicity properties to justify entropy maximization. The second, and the one that motivates the present study, is whether the theory should apply to 2D Navier-Stokes solutions at small viscosity. This is related to the issue of anomalies in ideally conserved quantities. For energy, there is no question since $E^\nu(t):=\frac{1}{2} \int_{\mathbb{T}^2} |u^\nu(t)|^2 {\rm d} x \xrightarrow[]{\nu \to 0} E_0$ for any finite time under the assumption that $\omega_0\in L^\infty(\mathbb{T}^2)$. On the other hand, it has not been clear that high-order ideal moments such as $I_n^\nu= \int_{\mathbb{T}^2} |\omega^\nu(t)|^n {\rm d} x$ for $n>2$ will be conserved in the limit of zero viscosity or if there will be an associated anomaly do to fine-scale mixing of the vorticity field. If they are not, it seems unlikely that these casmirs should be remembered at the level of the equilibrium distribution of vorticity. Our Theorem establishes that there can be no such anomalies of higher-order invariants on any finite time interval $[0,T]$ with $T$ arbitrarily large. Thus, it shows that the dependence of $F$ on viscosity is slow which provides a partial foundation for the Miller--Robert theory as it applies to slightly viscous fluids. \subsection*{Acknowledgments} We would like to thank Helena J. Nussenzveig Lopes for insightful comments. The research of PC was partially supported by NSF grant DMS-1713985. Research of TD was partially supported by NSF grant DMS-1703997. Research of TE was partially supported by NSF grant DMS-1817134. \bibliographystyle{cpam}

https://arxiv.org/abs/1507.04944

Forbidding induced even cycles in a graph: typical structure and counting

We determine, for all $k\geq 6$, the typical structure of graphs that do not contain an induced $2k$-cycle. This verifies a conjecture of Balogh and Butterfield. Surprisingly, the typical structure of such graphs is richer than that encountered in related results. The approach we take also yields an approximate result on the typical structure of graphs without an induced $8$-cycle or without an induced $10$-cycle.

\section{Introduction} \subsection{Background} The enumeration and description of the typical structure of graphs with given side constraints has become a successful and popular area at the interface of probabilistic, enumerative, and extremal combinatorics (see e.g.~\cite{BolSurv} for a survey of such work). For example, a by now classical result of Erd\H{o}s, Kleitman and Rothschild~\cite{EKR} shows that almost all triangle-free graphs are bipartite (given a fixed graph $H$, a graph is called $H$-\emph{free} if it does not contain $H$ as a not necessarily induced subgraph). This result was generalised to $K_k$-free graphs by Kolaitis, Pr\"omel and Rothschild~\cite{KPR}. There are now many precise results on the number and typical structure of $H$-free graphs and more generally graphs, hypergraphs and other combinatorial structures with a given (anti-)monotone property. Given a fixed graph $H$, a graph is called \emph{induced}-$H$-\emph{free} if it does not contain $H$ as an induced subgraph. Associated counting and structural questions are equally natural as in the non-induced case, but seem harder to solve. Thus much less is known about the typical structure and number of induced-$H$-free graphs than that of $H$-free graphs, though considerable work has been done in this area (see, e.g.~\cite{ABBM, BaBu, KMRS, PrSt2, PrSt3, PrSt4}). In particular, Pr\"omel and Steger~\cite{PrSt4} obtained an asymptotic counting result for the number of induced-$H$-free graphs on $n$ vertices, showing that the logarithm of this number is essentially determined by the so-called colouring number of $H$. This was generalised to arbitrary hereditary properties independently by Alekseev~\cite{Alek} as well as Bollob\'as and Thomason~\cite{BoTh}. Recent exciting developments in~\cite{BMS, SaTh} have opened up the opportunity to replace counting results by more precise results which identify the typical asymptotic structure. In this paper we determine the typical structure of induced-$C_{2k}$-free graphs (from which the corresponding asymptotic counting result follows immediately). The key difficulty we encounter is that the typical structure turns out to be more complex than encountered in previous results on forbidden induced subgraphs. This requires new ideas and a more intricate analysis when `excluding' classes of graphs which might be candidates for typical induced-$C_{2k}$-free graphs. \subsection{Graphs with forbidden induced cycles} Given a class of graphs $\mathcal{A}$, we let $\mathcal{A}_n$ denote the set of all graphs in $\mathcal{A}$ that have precisely $n$ vertices, and we say that \emph{almost all graphs in} $\mathcal{A}$ \emph{have property} $\mathcal{B}$ if $$\lim\limits_{ n\to \infty }\frac{|\{G\in \mathcal{A}_n: G \text{ has property } \mathcal{B}\}|}{|\mathcal{A}_n|} = 1.$$ Given graphs $H_1,\dots, H_m$, we say \emph{$G$ can be covered by $H_1,\dots, H_m$} if $V(G)$ admits a partition $A_1\cup \dots \cup A_m=V(G)$ such that $G[A_i]$ is isomorphic to $H_i$ for every $i\in \{1,\dots, m\}$. Pr\"omel and Steger proved in~\cite{PrSt2} that almost all induced-$C_4$-free graphs can be covered by a clique and an independent set, and in~\cite{PrSt1} characterised the structure of almost all induced-$C_5$-free graphs too. More recently, Balogh and Butterfield~\cite{BaBu} determined the typical structure of induced-$H$-free graphs for a wide class of graphs $H$. In particular they proved that almost all induced-$C_7$-free graphs can be covered by either three cliques or two cliques and an independent set, and that for $k\geq 4$ almost all induced-$C_{2k+1}$-free graphs can be covered by $k$ cliques. They also conjectured that for $k\geq 6$ almost all induced-$C_{2k}$-free graphs can be covered by $k-2$ cliques and a graph whose complement is a disjoint union of stars and triangles. Our main result completely verifies this conjecture. \begin{theorem}\label{main theorem} For $k\geq 6$, almost all induced-$C_{2k}$-free graphs can be covered by $k-2$ cliques and a graph whose complement is a disjoint union of stars and triangles. \end{theorem} Theorem~\ref{main theorem} together with the discussed results in~\cite{BaBu, EKR, PrSt1, PrSt2} implies that the typical structure of induced-$C_k$-free graphs is determined for every $k\in \mathbb{N}$ apart from $k\in \{6,8,10\}$. For the cases $k=8$ and $k=10$ the methods we use to prove Theorem~\ref{main theorem} allow us to also prove an approximate result on the typical structure of induced-$C_k$-free graphs. In order to state this result we require the following definitions. Given $\eta>0$ and graphs $G$ and $G'$ on the same vertex set, we say $G'$ is \emph{$\eta$-close to $G$} if $G'$ can be made into $G$ by changing (i.e. adding or deleting) at most $\eta |G|^2$ edges. We say a graph $G$ is a {\em sun} if either $G$ consists of a single vertex or $V(G)$ can be partitioned into sets $A,B$ such that $E(G)=\{uv : |\{u,v\}\cap B|\leq 1\}$. We call $A$ the {\em body} of the sun and $B$ the {\em side} of the sun. Note that all stars and cliques (including triangles) are suns, and that we consider a single vertex to be both a star of order one and a clique of order one. \begin{theorem}\label{secondary theorem}~ \begin{enumerate}[{\rm (i)}] \item For every $\eta>0$, almost all induced $C_{10}$-free graphs are $\eta$-close to graphs that can be covered by three cliques and a graph whose complement is a disjoint union of cliques. \item For every $\eta>0$, almost all induced $C_{8}$-free graphs are $\eta$-close to graphs that can be covered by two cliques and a graph whose complement is a disjoint union of suns. \end{enumerate} \end{theorem} We remark that in Theorems~\ref{main theorem} and~\ref{secondary theorem} we get exponential bounds on the proportion of induced-$C_{2k}$-free graphs that do not satisfy the relevant structural description. Our proofs also show that the $k-2$ cliques in the covering have size close to $n/(k-1)$ in Theorem~\ref{main theorem}, with analogous bounds in Theorem~\ref{secondary theorem}. Theorem~\ref{main theorem} also strengthens a result by Kang, McDiarmid, Reed and Scott~\cite{KMRS} showing that almost all induced-$C_{2k}$-free graphs have a linear sized homogeneous set. (Their results were motivated by the Erd\H{o}s-Hajnal conjecture, and actually apply to a large class of forbidden graphs $H$.) It would of course be interesting to determine the typical structure of induced-$C_6$-free graphs. \begin{question} What is the typical structure of induced-$C_6$-free graphs? \end{question} It seems likely that almost all induced-$C_6$-free graphs can be covered by one clique and one cograph, where a cograph is a graph not containing an induced copy of $P_4$. Another natural question is that of the typical structure of induced-$H$-free graphs of a given density. In particular, an intriguing question is whether their typical structure exhibits a non-trivial `phase transition' as found for triangle-free graphs~\cite{OPT} and more generally $K_r$-free graphs~\cite{BMSW}. \subsection{Overview of the paper} A key tool in our proofs is the recent hypergraph container approach, which was developed independently by Balogh, Morris and Samotij~\cite{BMS}, and Saxton and Thomason~\cite{SaTh}. Briefly, their result states that under suitable conditions on a uniform hypergraph $G$, there is a small collection $\mathcal{C}$ of small subsets (known as containers) of $V(G)$ such that every independent set of vertices in $G$ is a subset of some element of $\mathcal{C}$. The precise statement of the application used here is deferred until Section~\ref{sec: rough structure}. Given a graph $G$ and a set $A\subseteq V(G)$, we denote by $G[A]$ the graph induced on $G$ by $A$, and we denote the complement of $G$ by $\overline{G}$. For $k\in \mathbb{N}$ and a set $V$ of vertices we define an \emph{ordered $k$-partition of $V$} to be a $k$-partition of $V$ such that one partition class is labelled and the rest are unlabelled. If $Q$ is an ordered $k$-partition with labelled class $Q_0$ and unlabelled classes $Q_1,\dots, Q_{k-1}$ then we write $Q=(Q_0,\{Q_1,\dots, Q_{k-1}\})$. For $k\geq 4$, we say that a graph $G$ is a {\em $k$-template} if $V(G)$ has an ordered $(k-1)$-partition $Q=(Q_0 , \{Q_1 , \dots, Q_{k-2} \})$ such that $G[Q_i]$ is a clique for all $i\in [k-2]$ and one of the following holds. \begin{itemize} \item $k=4$ and $\overline{G}[Q_0]$ is a disjoint union of suns. \item $k=5$ and $\overline{G}[Q_0]$ is a disjoint union of stars and cliques. \item $k\geq 6$ and $\overline{G}[Q_0]$ is a disjoint union of stars and triangles. \end{itemize} Clearly every $k$-template is induced-$C_{2k}$-free. If $V(G)$ has such an ordered $(k-1)$-partition $Q$, we say that \emph{$G$ is a $k$-template on $Q$}, or \emph{$G$ has ordered $(k-1)$-partition $Q$}. If $Q'$ is the (unordered) $(k-1)$-partition with the same partition classes as $Q$, we may also say that \emph{$G$ is a $k$-template on $Q'$}. Thus Theorem~\ref{main theorem} can be reformulated as: $$\textit{`For $k\geq 6$, almost all induced $C_{2k}$-free graphs are $k$-templates.'}$$ Theorem~\ref{secondary theorem} can be similarly reformulated in terms of $4$- and $5$-templates. As mentioned earlier, the main difficulty in proving Theorem~\ref{main theorem} (compared to related results) is that typically $G[Q_0]$ is close to, but not quite, a complete graph. This makes it very difficult to rule out other similar classes of graphs as typical structures. To overcome this we use tools such as Ramsey's theorem to classify the graphs according to the neighbourhoods of certain vertices. More precisely, our approach to proving our main result is as follows. Firstly, in Section~\ref{sec: rough structure} we use the hypergraph containers result discussed above to show that almost all induced-$C_{2k}$-free graphs are close to being a $k$-template, for every $k\geq 4$ (see Lemma~\ref{lem: rough structure}). Note that Lemma~\ref{lem: rough structure} immediately implies Theorem~\ref{secondary theorem}. In Section~\ref{sec: number of templates} we prove upper and lower bounds on the number of $k$-templates on $n$ vertices (see Lemmas~\ref{size of template} and~\ref{the number of templates}). In Section~\ref{sec: set-up} we prove some preliminary results about graphs that are close to being a $k$-template. In Section~\ref{sec: main proof} we state a key result which is a version of Theorem~\ref{main theorem} with respect to a given ordered $(k-1)$-partition (see Lemma~\ref{induction conclusion}) and use it together with Lemma~\ref{the number of templates} to derive Theorem~\ref{main theorem}. The remainder of the paper is devoted to proving Lemma~\ref{induction conclusion} via an inductive argument, which we introduce at the end of Section~\ref{sec: main proof}. This argument involves partitioning the class of graphs considered in Lemma~\ref{induction conclusion} into three `bad' classes of graphs, and in each of Sections~\ref{sec: estimate 1}, \ref{sec: estimate 2} and~\ref{sec: estimate 3} we use Lemma~\ref{size of template} and the results in Section~\ref{sec: set-up} to prove an upper bound on the number of graphs in a different one of these classes (see Lemmas~\ref{F^1}, \ref{F^2_Q} and~\ref{F^3}). In particular, Lemmas~\ref{F^1} and~\ref{F^2_Q} already show that almost all induced-$C_{2k}$-free graphs are `extremely close' to being $k$-templates (see Proposition~\ref{beta}). Finally in Section~\ref{sec: final calculation} we use Lemmas~\ref{lem: rough structure}, \ref{F^1}, \ref{F^2_Q} and~\ref{F^3} to complete the inductive argument set up in Section~\ref{sec: main proof} and so prove Lemma~\ref{induction conclusion}. Before starting on any of this however, we lay out some notation and set out some useful tools in Section~\ref{sec: notation}, below. \section{Notation and tools}\label{sec: notation} Given a graph $G$, a vertex $x\in V(G)$, and an ordered $(k-1)$-partition $Q=(Q_0,\{Q_1,\dots,Q_{k-2}\})$ of $V(G)$, we let $N(x), \overline{N}(x)$ denote the set of neighbours and non-neighbours of $x$ in $G$, respectively. We also let $N_{Q_i}(x), \overline{N}_{Q_i}(x)$ denote the set of neighbours of $x$ in $Q_i$ and non-neighbours of $x$ in $Q_i$, respectively. We sometimes use the notation $d^i_{G,Q}(x)=|N_{Q_i}(x)|$ and $\overline{d}^i_{G,Q}(x)=|\overline{N}_{Q_i}(x)|$ when we want to emphasise which graph we are working with. For a set $A$ of vertices in $G$, we define $$N(A):= \bigcap_{v\in A} N(v), \hspace{0.6cm} \overline{N}(A):= \bigcap_{v\in A} \overline{N}(v),$$ $$N_{Q_i}(A):= \bigcap_{v\in A} N_{Q_i}(v), \hspace{0.6cm} \overline{N}_{Q_i}(A):= \bigcap_{v\in A} \overline{N}_{Q_i}(v).$$ If it generates no ambiguity, we may write $N_i(x), \overline{N}_{i}(x), N_i(A), \overline{N}_i(A)$ for $N_{Q_i}(x), \overline{N}_{Q_i}(x), N_{Q_i}(A),$ and $\overline{N}_{Q_i}(A)$ respectively. Given $A,B\subseteq V(G)$, we define $$N^*(A,B) := N(A)\cap \overline{N}(B), \hspace{0.6cm} N_i^*(A,B) := N_i(A)\cap \overline{N}_i(B).$$ In the case when $A$ and $B$ both have size one, containing vertices $a,b$ respectively, we may write $N^*(a,b)$ for $N^*(A,B)$ and $N_i^*(a,b)$ for $N_i^*(A,B)$. We say that a partition of vertices is \emph{balanced} if the sizes of any two partition classes differ by at most one. Given a $(k-1)$-partition $Q$ of $[n]$ with partition classes $Q_0,\dots,Q_{k-2}$, and a graph $G = (V,E)$ on vertex set $[n]$, and an edge or non-edge $e = uv$ with $u \in Q_i$ and $v \in Q_j$, we call $e$ {\em crossing} if $i\neq j$ and {\em internal} if $i = j$. We denote a path on $m$ vertices by $P_m$. Given a path $P=p_1\dots p_m$ and a sequence $A_1,\dots, A_m$ of sets of vertices, we say that $P$ \emph{has type} $A_1,\dots, A_m$ if $p_{\ell}\in A_{\ell}$ for every $\ell\in [m]$. We call a graph a \emph{linear forest} if it is a forest such that all components are paths or isolated vertices. Given $\ell, t\in \mathbb{N}$ we let $R_{\ell}(t)$ denote the $\ell$-colour Ramsey number for monochromatic $t$-cliques, i.e.~$R_{\ell}(t)$ is the smallest $N\in \mathbb{N}$ such that every $\ell$-colouring of the edges of $K_{N}$ yields a monochromatic copy of $K_t$. We define $$n_k:= \left\lceil \frac{n}{k-1} \right\rceil.$$ In a number of our proofs we shall use the following Chernoff bound. \begin{lemma}[Chernoff bound]\label{Chernoff Bounds} Let $X$ have binomial distribution and let $0<a\leq \mathbb{E}[X]$. Then \begin{enumerate}[{\rm (i)}] \item $P(X>\mathbb{E}[X]+a)\leq \exp \left( -\frac{a^2}{4\mathbb{E}[X]} \right)$. \item $P(X<\mathbb{E}[X]-a)\leq \exp \left( -\frac{a^2}{2\mathbb{E}[X]} \right)$. \end{enumerate} \end{lemma} Whenever this does not affect the argument, we assume all large numbers to be integers, so that we may sometimes omit floors and ceilings for the sake of clarity. In some proofs, given $a,b\in \mathbb{R}$ with $0<a,b<1$, we will use the notation $a\ll b$ to mean that we can find an increasing function $g$ for which all of the conditions in the proof are satisfied whenever $a\leq g(b)$. Throughout we write $\log x$ to mean $\log_2 x$. We define $\xi(p):= - 3 p(\log p )/2$. The following bounds will prove useful to us. For $n\geq 1$ and $3\log n /n\leq p \leq 10^{-11}$, \begin{equation}\label{entropy bound} \binom{n}{\leq pn}:= \sum\limits_{i=0}^{\left\lfloor pn \right\rfloor} \binom{n}{i} \leq pn \left( \frac{en}{pn} \right)^{pn}\leq 2^{\xi(p)n}, \end{equation} and \begin{equation}\label{entropy bound 1.5} \xi(p)\leq \frac{3}{2}p \left(\frac{1}{p} \right)^{1/8}\leq p^{3/4}. \end{equation} \section{Approximate structure of typical induced-$C_{2k}$-free graphs}\label{sec: rough structure} The main result of this section is Lemma~\ref{lem: rough structure}, which approximately determines the typical structure of induced-$C_{2k}$-free graphs. As mentioned earlier, we make use of a `container theorem' which reduces the proof of Lemma~\ref{lem: rough structure} to an extremal problem involving induced-$C_{2k}$-free graphs. More precisely, the argument is structured as follows. We first introduce a number of tools (see Subsection~\ref{subsec: rough 1}): a `Containers' theorem (Theorem~\ref{thm: containers}), a Stability theorem (Theorem~\ref{thm: basic stability}), and two Removal Lemmas (Theorem~\ref{thm: induced removal lemma}, Lemma~\ref{lem: multigraph removal lemma}). In Subsection~\ref{subsec: rough 2} we use Theorem~\ref{thm: basic stability} to derive a Stability result involving induced-$C_{2k}$-free graphs (Lemma~\ref{lem: specialised stability}). Similarly we use Theorem~\ref{thm: induced removal lemma} to derive another specialised version of the Removal Lemma (Lemma~\ref{lem: specialised removal lemma}). In Subsection~\ref{subsec: rough 3} we use Theorem~\ref{thm: containers} together with Lemmas~\ref{lem: multigraph removal lemma},~\ref{lem: specialised stability} and~\ref{lem: specialised removal lemma} to determine the approximate structure of typical induced-$C_{2k}$-free graphs. We denote the number of (labelled) induced-$C_{2k}$-free graphs on $n$ vertices by $F(n,k)$. \begin{lemma}\label{lem: rough structure} Let $k\geq 4$. For every $\eta >0$ there exists $\varepsilon >0$ such that the following holds for all sufficiently large $n$. All but at most $F(n,k) 2^{-\varepsilon n^2}$ induced-$C_{2k}$-free graphs on $n$ vertices can be made into a $k$-template by changing at most $\eta n^2$ edges. \end{lemma} Note that Lemma~\ref{lem: rough structure} immediately implies Theorem~\ref{secondary theorem}. \subsection{Tools: containers, stability and removal lemmas}\label{subsec: rough 1} The key tool in this section is Theorem~\ref{thm: containers}, which is an application of the more general theory of Hypergraph Containers developed in~\cite{BMS,SaTh}. We use the formulation of Theorem~1.5 in~\cite{SaTh}. We require the following definitions in order to state it. A \emph{$2$-coloured multigraph} $G$ on vertex set $[N]$ is a pair of edge sets $G_R, G_B\subseteq [N]^{(2)}$, which we call the red and blue edge sets respectively. If $H$ is a fixed graph on vertex set $[h]$, a copy of $H$ in $G$ is an injection $f:[h]\rightarrow[N]$ such that for every edge $uv$ of $H$, $f(u)f(v)\in G_R$, and for every non-edge $u'v'$ of $H$, $f(u')f(v')\in G_B$. We write $H\subseteq G$ if $G$ contains a copy of $H$, and we say that $G$ is \emph{$H$-free} if there are no copies of $H$ in $G$. We say that $G$ is \emph{complete} if $G_R \cup G_B = [N]^{(2)}$. We denote by $G^B$ the graph on vertex set $[N]$ and edge set $G_B$. \begin{theorem}\label{thm: containers} Let $H$ be a fixed graph with $h:= |V(H)|$. For every $\eps>0$, there exists $c>0$ such that for all sufficiently large $N$, there exists a collection $\mathcal{C}$ of complete $2$-coloured multigraphs on vertex set $[N]$ with the following properties. \begin{enumerate}[{\rm (a)}] \item For every graph $I$ on $[N]$ that contains no induced copy of $H$, there exists $G\in \mathcal{C}$ such that $I \subseteq G$. \item Every $G\in \mathcal{C}$ contains at most $\eps N^h$ copies of $H$. \item $\log |\mathcal{C}| \leq c N^{2 - (h-2)/(\binom{h}{2}-1)} \log N$. \end{enumerate} \end{theorem} Another tool that we will use is the following classical Stability theorem of Erd\H{o}s and Simonovits (see e.g.~\cite{Erd1, Erd2, Sim}). By $T_k(n)$ we denote the \emph{Tur\'an graph}, the largest complete $k$-partite graph on $n$ vertices, and we define $t_{k}(n):= e(T_k(n))$. Given a family $\mathcal{H}$ of fixed graphs, we say a graph $G$ is \emph{$\mathcal{H}$-free} if $G$ does not contain any $H\in \mathcal{H}$ as a (not necessarily induced) subgraph, and we say $G$ is \emph{induced-$\mathcal{H}$-free} if $G$ does not contain any $H\in \mathcal{H}$ as an induced subgraph. \begin{theorem}\label{thm: basic stability} Let $\mathcal{H}=\{H_1,\dots, H_{\ell}\}$ be a family of fixed graphs, and let $k:= \min_{1\leq i\leq \ell} \chi (H_i)$. For every $\delta>0$ there exists $\varepsilon>0$ such that the following holds for all sufficiently large $n$. If a graph $G$ on $n$ vertices is $\mathcal{H}$-free and $e(G)\geq t_{k-1}(n)-\varepsilon n^2$, then $G$ can be obtained from $T_{k-1}(n)$ by changing at most $\delta n^2$ edges. \end{theorem} The final tools that we introduce in this subsection are the following two Removal Lemmas. The first is an extension of the Induced Removal Lemma to families of forbidden graphs, and is due to Alon and Shapira~\cite{AlSh}. The original statement of this theorem also applies to infinite families of forbidden graphs, but the version for finite families is sufficient for our purposes. The second is a version of the Removal Lemma applicable to complete $2$-coloured multigraphs. The proof is similar to that of the standard Removal Lemma, so we omit it here; for details see~\cite{TownsendPhD}. For two sets $A,B$, we denote their symmetric difference by $A\triangle B$. For $2$-coloured multigraphs $G, G'$ on the same vertex set we define their distance by $\text{dist}(G, G'):= |G_R\triangle G'_R| + |G_B\triangle G'_B|$. \begin{theorem}\cite{AlSh}\label{thm: induced removal lemma} For every finite family of fixed graphs $\mathcal{H}$ and every $\delta >0$, there exists $\varepsilon>0$ such that the following holds for all sufficiently large $n$. If a graph $G$ on $n$ vertices contains at most $\varepsilon n^h$ induced copies of each graph $H\in \mathcal{H}$ on $h$ vertices, then $G$ can be made induced-$\mathcal{H}$-free by changing at most $\delta n^2$ edges. \end{theorem} \begin{lemma}\label{lem: multigraph removal lemma} For every fixed graph $H$ on $h$ vertices, and every $\delta>0$, there exists $\varepsilon>0$ such that the following holds for all sufficiently large $n$. If a complete $2$-coloured multigraph $G$ on vertex set $[n]$ contains at most $\varepsilon n^h$ copies of $H$, then there exists a complete $2$-coloured multigraph $G'$ on vertex set $[n]$ such that $G'$ is $H$-free and ${\rm dist}(G, G')\leq \delta n^2$. \end{lemma} \subsection{Stability and removal lemmas for even cycles}\label{subsec: rough 2} Suppose $H$ is a complete $2$-coloured multigraph on $m$ vertices with $H_R\cap H_B=\emptyset$. If $m=3$ and $|H_R|\leq 1$ we call $H$ a \emph{mostly blue triangle}. For $k\in \{4,5,6\}$, if $m=4$ and $|H_R|\geq 6-k$ and $H^B$ contains a copy of $P_4$ then we call $H$ a \emph{$k$-good tetrahedron}. The following technical proposition will be useful in proving Lemmas~\ref{lem: specialised stability} and~\ref{lem: specialised removal lemma}. \begin{proposition}\label{prop: C_2k in small multigraphs} Let $k\geq 4$ and let $G$ be a complete $2$-coloured multigraph on $2k$ vertices. If $G$ satisfies one of the following properties then $G$ contains a copy\COMMENT{Note that I do want to say copy, instead of induced copy, here and in similar instances throughout this section, since $G$ is a $2$-coloured multigraph, and this useage is consistent with the definition of $G$ containing a fixed graph that was given at the beginning of the section.} of $C_{2k}$. Below, $r_i$ always denotes a red edge. \begin{enumerate}[{\rm (E1)}] \item $G_R\triangle G_B$ is a set of at most $k$ disjoint (red or blue) edges. \item $G_R\triangle G_B$ is the edge set of two disjoint copies of a blue $K_k$. \item $G_R\triangle G_B$ is the edge set of a union of disjoint graphs $K_3^1, K_3^2, r_1,\dots, r_{k-3}$, where each $K_3^i$ is a mostly blue triangle\COMMENT{Note that (E3) is not sufficient to imply that $G$ contains a $C_{2k}$ in the case $k=3$ (in particular, there is no covering permutation if exactly one of the mostly blue triangles contains a red edge). This creates big problems when trying to emulate the proof of Lemma~\ref{lem: specialised removal lemma} for $k=3$. So extending the rough structure section to the $k=3$ case may be more effort than it's worth. - Tim}. \item $G_R\triangle G_B$ is the edge set of a union of disjoint graphs $K_4^1, r_1,\dots, r_{k-2}$, where $K_4^1$ is a $4$-good tetrahedron. \item $k\geq 5$ and $G_R\triangle G_B$ is the edge set of a union of disjoint graphs $K_4^1, r_1,\dots, r_{k-2}$, where $K_4^1$ is a $5$-good tetrahedron. \item $k\geq 6$ and $G_R\triangle G_B$ is the edge set of a union of disjoint graphs $K_4^1, r_1,\dots, r_{k-2}$, where $K_4^1$ is a $6$-good tetrahedron. \end{enumerate} \end{proposition} \begin{proof} Let $V(G)=\{v_1,\dots, v_{2k}\}$. Let $C=c_1\dots c_{2k}$ be a $2k$-cycle. Note that if there exists a permutation $\sigma$ of $[2k]$ such that for every edge $c_ic_j\in E(C)$ we have $v_{\sigma(i)}v_{\sigma(j)}\in G_R$ and such that for every non-edge $c_{i'}c_{j'}\notin E(C)$ we have $v_{\sigma(i')}v_{\sigma(j')}\in G_B$, then $v_{\sigma(1)}\dots v_{\sigma(2k)}$ is a copy of $C_{2k}$ in $G$. We call such a permutation $\sigma$ a \emph{covering permutation from $C$ to $G$}. For ease of reading, we will write a permutation $\sigma$ on $[2k]$ using the notation $\sigma = (\sigma(1), \dots, \sigma(2k))$. If $\sigma$ restricted to $\{m, m+1, \dots, 2k\}$ is the identity permutation, we may simply write $\sigma=\{\sigma(1),\dots, \sigma(m-1)\}$ instead. So for example if $\sigma=(1,3,4,2)$ is a covering permutation from $C$ to $G$, then $v_1v_3v_4v_2v_5 \dots v_{2k}$ is a copy of $C_{2k}$ in $G$. We now show that each of the properties (E1),$\dots$,(E6) imply that there exists a covering permutation from $C$ to $G$, and hence that $G$ contains a copy of $C_{2k}$. \begin{enumerate}[{\rm (E1)}] \item There exists $b,r\in \mathbb{N}\cup \{0\}$ with $b+r\leq k$ such that, by relabelling vertices if necessary, $G_B \backslash G_R =\{v_1v_2, \dots, v_{2b-1}v_{2b}\}$ and $G_R \backslash G_B=\{v_{2b+1}v_{2b+2}, \dots, v_{2(b+r) -1}v_{ 2(b+r)}\}$. Depending on the value of $b$ we find the following covering permutations $\sigma$ from $C$ to $G$, as required. \begin{itemize} \item If $b=0$ then $\sigma$ is the identity permutation. \item If $b=1$ then $\sigma=(1,3,4,2)$. \item If $b\geq 2$ then $\sigma=(1,3,\dots, 2b-1, 2,4, \dots, 2b).$ \end{itemize} \item Let $\{v_1,\dots, v_k\}, \{v_{k+1},\dots, v_{2k}\}$ be the respective vertex sets of the two copies of a blue $K_k$ in $G_R\triangle G_B$. Then $\sigma=(1,k+1,2,k+2,\dots,k,2k)$ is a covering permutation from $C$ to $G$, as required. \item Let $V(K_3^1)=\{v_1,v_2,v_3\}, V(K_3^2)=\{v_4,v_5,v_6\}$ and $V(r_i)=\{v_{2i+5}, v_{2i+6}\}$ for every $i\in [k-3]$. Depending on the colour of the edges in $K_3^1, K_3^2$ we find the following covering permutations $\sigma$ from $C$ to $G$, as required. \begin{itemize} \item If $K_3^1, K_3^2$ both contain no red edges, then $\sigma = (1,4,2,5,3,6)$. \item If $K_3^1$ contains exactly one red edge $v_1v_2$ and $K_3^2$ contains no red edges, then $\sigma = (4,1,2,5,3,6)$. \item If $K_3^1$ contains exactly one red edge $v_1v_2$ and $K_3^2$ contains exactly one red edge $v_5v_6$, then $\sigma = (1,2,4,3,5,6)$. \end{itemize} \item Let $V(K_4^1)=\{v_1,v_2,v_3,v_4\}$ and $V(r_i)=\{v_{2i+3}, v_{2i+4}\}$ for every $i\in [k-2]$. Depending on the configuration of red edges in $K_4^1$ we find the following covering permutations $\sigma$ from $C$ to $G$, as required. \begin{itemize} \item If $K_4^1$ contains exactly three red edges $v_1v_2, v_2v_3, v_3v_4$, then $\sigma$ is the identity permutation. \item If $K_4^1$ contains exactly two red edges $v_1v_2, v_2v_3$, then $\sigma = (1,2,3,5,6,4)$. \item If $K_4^1$ contains exactly two red edges $v_1v_2, v_3v_4$, then $\sigma = (1,2,5,6,3,4)$. \end{itemize} \item We may assume that $K_4^1$ contains exactly one red edge, since that is the only case not covered by (E4). Let $V(K_4^1)=\{v_1,v_2,v_3,v_4\}$ and $V(r_i)=\{v_{2i+3}, v_{2i+4}\}$ for every $i\in [k-2]$, and let $v_1v_2$ be the red edge in $K_4^1$. Then $\sigma = (1,2,5,6,3,7,8,4)$ is a covering permutation from $C$ to $G$, as required. \item We may assume that $K_4^1$ contains no red edges, since that is the only case not covered by (E5). Let $V(K_4^1)=\{v_1,v_2,v_3,v_4\}$ and $V(r_i)=\{v_{2i+3}, v_{2i+4}\}$ for every $i\in [k-2]$. Then $\sigma = (1,5,6,2,7,8,3,9,10,4)$ is a covering permutation from $C$ to $G$, as required. \end{enumerate} \end{proof} We now use Theorem~\ref{thm: basic stability} and Proposition~\ref{prop: C_2k in small multigraphs} to prove the following more specialised Stability result involving $C_{2k}$-free $2$-coloured multigraphs. \begin{lemma}\label{lem: specialised stability} Let $k\geq 4$. For every $\delta>0$ there exists $\varepsilon>0$ such that the following holds for all sufficiently large $n$. If a complete $2$-coloured multigraph $G$ on vertex set $[n]$ is $C_{2k}$-free and $|G_R\cap G_B|\geq t_{k-1}(n)-\varepsilon n^2$, then the graph $([n], G_R\cap G_B)$ can be obtained from $T_{k-1}(n)$ by changing at most $\delta n^2$ edges. \end{lemma} \begin{proof} Choose $n_0\in \mathbb{N}$ and $\varepsilon>0$ such that $1/n_0 \ll \varepsilon \ll \delta$. Let $n\geq n_0$. Since $G$ is $C_{2k}$-free, we know by Proposition~\ref{prop: C_2k in small multigraphs} that no $2k$ vertices of $G$ induce on $G$ a $2$-coloured multigraph $G'$ that satisfies (E1). So, since $G$ is complete, the graph $([n], G_R\cap G_B)$ must be $T_k(2k)$-free. Note that $\chi(T_k(2k))=k$. By Theorem~\ref{thm: basic stability}, this together with the fact that $|G_R\cap G_B|\geq t_{k-1}(n)-\varepsilon n^2$ implies that the graph $([n], G_R\cap G_B)$ can be obtained from $T_{k-1}(n)$ by changing at most $\delta n^2$ edges. \end{proof} The following proposition characterises the structure of graphs without $k$-good tetrahedrons. It will be useful in proving Lemma~\ref{lem: specialised removal lemma}. The proof is fairly straightforward so we give only a sketch of it here. \begin{proposition}\label{prop: k-good tetradehron free} Let $G$ be a $2$-coloured multigraph with $G_R\cap G_B=\emptyset$. \begin{enumerate}[{\rm (i)}] \item If $G$ does not contain a $6$-good tetrahedron then $G^B$ is a disjoint union of stars and triangles\COMMENT{Full Proof: Suppose a $2$-coloured multigraph $G$ is $6$-good tetrahedron-free. Then $G^B$ is $P_4$-free, by definition. So all paths in $G^B$ have length at most $2$. Let $H$ be a component of $G^B$ that is not a triangle. Then $H$ is a tree, since any $4$-cycle or longer in $H$ would contain a $P_4$ (as would a triangle with a pendant edge). Suppose for a contradiction that $H$ has $2$ vertices $u,v$ both of degree at least $2$. Find a path $P=\{u,u_1,\dots, u_m,v\}$ in $H$ (of length at least $1$). Note that if the path has length $1$ then $u=u_m$ and $v=u_{1}$. Note that since $H$ is a tree, $N(u)\backslash \{u_1\}\cap P=\emptyset$ and $N(v)\backslash \{u_m\}\cap P=\emptyset$. So since $|N(u)|, |N(v)|\geq 2$, there exist vertices $u', v'$ such that $u'\in N(u)\backslash P$ and $v'\in N(v)\backslash P$. Since $H$ is a tree, $u'\ne v'$, and $u'Pv'$ is a path of length at least $3$, which contradicts the fact that $H$ is $P_4$-free. So $H$ has at most one vertex of degree at least $2$, and so $H$ must be a star. So $G^B$ is a disjoint collection of stars and triangles, as required.}. \item If $G$ does not contain a $5$-good tetrahedron then $G^B$ is a disjoint union of stars and cliques\COMMENT{Full Proof: Suppose a $2$-coloured multigraph $G$ is $5$-good tetrahedron-free. Then the vertices of every copy of $P_4$ in $G^B$ induce a $K_4$, by definition. Let $H$ be a component of $G^B$ that is not a star or triangle. Since $P_4$-free graphs consist only of a disjoint union of stars and triangles, $H$ must contain a $P_4$, and hence a $K_4$. We now prove that $H$ is a clique by induction on $n=|H|$, for all $n\geq 4$. For the base case, $n=4$ so $H=K_4$ and we are done. Otherwise, suppose $n> 4$. Choose a $K_4$ in $H$ and choose $1$ vertex $v$ not in this $K_4$, and note that $H-v$ is not a triangle or a star (since it contains a $K_4$). So by the inductive hypothesis, $H-v$ is a clique. Suppose for a contradiction that there is a vertex $u$ in $H-v$ such that $uv\notin E(H)$. Then $v$ and $u$ together with a vertex in $H-v$ adjacent to $v$, and any other vertex in $H-v$ forms a $P_4$ but not a $K_4$. This is a contradiction, and hence $H$ is a clique. This completes the induction, and hence the proof.}. \item If $G$ does not contain a $4$-good tetrahedron then $G^B$ is a disjoint union of suns\COMMENT{Full Proof: Suppose a $2$-coloured multigraph $G$ is $4$-good tetrahedron-free. Then the vertices of every copy of $P_4$ in $G^B$ induce a $K_4$ or a $K_4^-$, by definition. Let $H$ be a component of $G^B$ that is not a star or clique. Since all graphs in which the vertices of every $P_4$ induce a $K_4$ consist only of a disjoint union of stars and cliques, $H$ must contain a $P_4$ that does not induce a $K_4$, and hence $H$ must contain an induced $K_4^-$. We now prove that $H$ is a sun by induction on $n=|H|$, for all $n\geq 4$. For the base case, $n=4$ so $H=K_4^-$ and we are done. Otherwise, suppose $n> 4$. Choose an induced $K_4^-$ in $H$ and choose $1$ vertex $v$ not in this $K_4^-$, and note that $H-v$ is not a star or a clique (since it contains an induced $K_4^-$). So by the inductive hypothesis, $H-v$ is a sun that is not a clique. Let $A$ be the sun's body, and $B$ it's side. If $v$ is adjacent to every vertex in $H-v$, then $H$ is a sun with $v$ in it's body, and we are done, so assume not. Suppose for a contradiction that there exists $v_0\in A$ and $v_1\in B$ with $v_0v\notin E(H)$ and $v_1v\in E(H)$. Since $H-v$ is not a clique, $|B|\geq 2$, and so there exists a vertex $v_2\in B$ such that $v_2 \ne v_1$. Then $vv_1v_0v_2$ is a $P_4$, but $v,v_1,v_0,v_2$ induce at least two red edges $vv_0,v_1v_2$, which is a contradiction. So at least one of the following hold: $N(v)\supseteq A$, $\overline{N(v)}\supseteq B$.\\ \noindent {\bf Case 1:} \emph{$N(v)\supseteq A$}\\ \noindent In this case, we claim that there cannot exist a vertex $v_1\in B$ with $v_1v\in E(H)$. Indeed, otherwise there also exists a vertex $v_2\in B$ with $v_2\ne v_1$ and $v_2v\notin E(H)$ (since $v$ is not adjacent to every vertex in $H-v$). But then for any vertex $u\in A$, we have that $vv_1uv_2$ is a $P_4$, but $v,v_1,u,v_2$ induce at least two red edges $vv_2, v_1v_2$, which is a contradiction. This proves the claim, so in this case we have that $N(v)=A$, and so $H$ is a sun with $v$ in it's side.\\ \noindent {\bf Case 2:} \emph{$\overline{N(v)}\supseteq B$}\\ \noindent In this case, we claim that there cannot exist a vertex $v_0\in A$ with $v_0v\notin E(H)$. Indeed, otherwise there also exists a vertex $v_3\in A$ with $v_3\ne v_0$ and $v_3v\in E(H)$ (since $v$ is adjacent to at least one vertex of $H$, since $H$ is a component). But then for any vertex $w\in B$, we have that $vv_3wv_0$ is a $P_4$, but $v,v_3,w,v_0$ induce at least two red edges $vv_0, vw$, which is a contradiction. This proves the claim, so in this case we have that $N(v)=A$, and so $H$ is a sun with $v$ in it's side.\\ This completes the induction, and hence the proof.}. \end{enumerate} \end{proposition} \begin{proof} (i) follows immediately from the fact that if $G$ is $6$-good tetrahedron-free then $G^B$ does not contain a $P_4$. To see (ii), note that if $G$ is $5$-good tetrahedron-free and $P$ is a copy of $P_4$ in $G^B$, then $G^B[V(P)]=K_4$. So every component $H$ of $G^B$ is either a star or a triangle or contains a $K_4$. But in the latter case it is easy to check that $H$ is actually a clique. It remains to prove (iii). If $G$ is $4$-good tetrahedron-free and $P$ is a copy of $P_4$ in $G^B$, then $G^B[V(P)]$ is either a $K_4$ or a copy of the graph $K_4^-$ obtained from $K_4$ by deleting one edge. So every component $H$ of $G^B$ is either a star or a clique or contains an induced copy of $K_4^-$. Using induction on $|H|$, it is not hard to show that in the latter case $H$ must be a sun. \end{proof} We now use Theorem~\ref{thm: induced removal lemma} together with Propositions~\ref{prop: C_2k in small multigraphs} and~\ref{prop: k-good tetradehron free} to prove the following more specialised Removal Lemma involving even cycles. \begin{lemma}\label{lem: specialised removal lemma} For every $k\geq 4$ and every $\delta>0$ there exists $\varepsilon>0$ such that the following holds for all sufficiently large $n$. Suppose $G$ is a complete $2$-coloured multigraph on $n$ vertices such that $G_R\cap G_B = E(T_{k-1}(n))$. Let $Q$ be the unique $(k-1)$-partition of the vertices of $G$ such that no partition class induces an edge in $G_R\cap G_B$. Suppose further that $G$ contains at most $\varepsilon n^{2k}$ copies of $C_{2k}$. Then there exists a $k$-template $T=(V(G), E^T)$ on $Q$ such that $|G_R\triangle E^T|\leq \delta n^2$. \end{lemma} \begin{proof} We first prove the lemma in the case $k\geq 6$. Choose $n_0\in \mathbb{N}$ and $\varepsilon, \gamma>0$ such that $1/n_0\ll \varepsilon \ll \gamma \ll \delta, 1/k$. Let $n\geq n_0$ and let $Q=(Q_1, \dots, Q_{k-1})$. Let $c:= \varepsilon^{1/3}$. We claim that for no two distinct $i,j\in [k-1]$ do $G[Q_i]$ and $G[Q_j]$ both contain at least $c n^k$ copies of a blue $K_k$. Indeed, if they do then there are at least $c^2 n^{2k} > \varepsilon n^{2k}$ sets of $2k$ vertices that each induce on $G$ a $2$-coloured multigraph $G'$ that satisfies (E2). By Proposition~\ref{prop: C_2k in small multigraphs} each such $G'$ contains a copy of $C_{2k}$. This contradicts the assumption that $G$ contains at most $\varepsilon n^{2k}$ copies of $C_{2k}$, which proves the claim. Thus there exists $J\subseteq [k-1]$ with $|J|\leq 1$ such that for all $i\in [k-1]$ with $i\notin J$, $G[Q_i]$ contains fewer than $c n^k$ copies of a blue $K_k$. Together with Theorem~\ref{thm: induced removal lemma} (applied to $G^B[Q_i]$) this implies that $G[Q_i]$ can be made free of blue cliques of size $k$ by changing the colour of at most $\gamma n^2$ edges inside $Q_i$. So by Tur\'an's Theorem, for all $i\in [k-1]$ with $i\notin J$, $G[Q_i]$ must have at least $$(k-1)\binom{n/(k-1)^2}{2}-2\gamma n^2 \geq \frac{n^2}{4(k-1)^3}$$ red edges. \vspace{0.3cm} \noindent {\bf Claim 1:} \emph{There is at most one index $i\in [k-1]$ such that $G[Q_i]$ contains at least $c n^3$ mostly blue triangles. Moreover, if there is such an index $i$ then $J\subseteq \{i\}$, and if there is no such index then $J=\emptyset$.} \noindent Indeed, suppose for a contradiction that there exist distinct $i,j\in [k-1]$ such that $Q_i, Q_j$ both contain at least $c n^3$ mostly blue triangles. Note that any class that contains at least $c n^k$ copies of a blue $K_k$ must contain at least $c n^3$ mostly blue triangles. So we may assume that $J\subseteq \{i,j\}$. Thus for every index $\ell \ne i,j$, $G[Q_{\ell}]$ contains at least $n^2/(4(k-1)^3)$ red edges. Thus there are at least $2\varepsilon n^{2k}$ sets of $2k$ vertices that each induce on $G$ a $2$-coloured multigraph $G'$ that satisfies (E3). (To see this, note that to choose such a set of $2k$ vertices we may choose, for both indices $i,j$, the vertices of any one of the at least $c n^3$ mostly blue triangles in $G[Q_i], G[Q_j]$ respectively, and then choose, for each index $\ell \ne i,j$, any one of the at least $n^2/(4(k-1)^3)$ red edges in $Q_{\ell}$.) By Proposition~\ref{prop: C_2k in small multigraphs} each such $G'$ contains a copy of $C_{2k}$. This contradicts the assumption that $G$ contains at most $\varepsilon n^{2k}$ copies of $C_{2k}$, which proves the claim. \vspace{0.3cm} Let $J'$ consist of the index $j_0\in [k-1]$ such that $G[Q_{j_0}]$ contains at least $c n^3$ mostly blue triangles, if such an index exists. Otherwise let $J':= \emptyset$. Thus $J\subseteq J'$. For all $i\in [k-1]$ with $i\notin J'$, Claim~1 together with Theorem~\ref{thm: induced removal lemma} (applied to $G^B[Q_i]$) implies that $G[Q_i]$ can be made free of mostly blue triangles by changing the colour of at most $\gamma n^2$ edges inside $Q_i$. This implies that the blue edges inside $Q_i$ after such a change form a matching. Hence $G[Q_i]$ contains at most $2\gamma n^2$ blue edges. If $J'=\emptyset$ then $G[Q_i]$ contains at most $2\gamma n^2$ blue edges for all $i\in [k-1]$, and hence $|G_B\backslash G_R|\leq \delta n^2$ (since $\gamma \ll \delta, 1/k$). In this case we are done by setting $T$ to be $K_n$. Otherwise, $J'=\{j_0\}$ and it suffices to show that the blue edges in $G[Q_{j_0}]$ can be made into the edge set of a disjoint collection of stars and triangles by changing the colour of at most $\gamma n^2$ edges inside $Q_{j_0}$, since then we are done by setting $T$ to be $K_n$ minus this disjoint collection of stars and triangles. \vspace{0.3cm} \noindent {\bf Claim 2(a):} \emph{$G[Q_{j_0}]$ contains fewer than $c n^4$ $6$-good tetrahedrons.} \noindent Indeed, otherwise there are at least $\varepsilon^{1/2} n^{2k}$ sets of $2k$ vertices that each induce on $G$ a $2$-coloured multigraph $G'$ that satisfies (E6). (To see this, note that to choose such a set of $2k$ vertices we may first choose the vertices of any one of the at least $c n^4$ $6$-good tetrahedrons, and then choose, for each other class $Q_i$, any one of the at least $n^2/(4(k-1)^3)$ red edges in $Q_i$.) By Proposition~\ref{prop: C_2k in small multigraphs} each such $G'$ contains a copy of $C_{2k}$. This contradicts the assumption that $G$ contains at most $\varepsilon n^{2k}$ copies of $C_{2k}$, which proves the claim. \vspace{0.3cm} Claim~2(a) together with Theorem~\ref{thm: induced removal lemma} (applied to $G^B[Q_{j_0}]$) implies that $G[Q_{j_0}]$ can be made free of $6$-good tetrahedrons by changing the colour of at most $\gamma n^2$ edges inside $Q_{j_0}$. Proposition~\ref{prop: k-good tetradehron free}{\rm (i)} implies that after such a change, all blue edges inside $Q_{j_0}$ form a disjoint collection of stars and triangles, as required. This completes the proof in the case $k\geq 6$. For the case $k=5$, the proof is almost identical to the case $k\geq 6$, except that instead of Claim~2(a) we prove the following weaker claim, which follows in a similar way. \vspace{0.3cm} \noindent {\bf Claim 2(b):} \emph{$G[Q_{j_0}]$ contains fewer than $c n^4$ $5$-good tetrahedrons.} \vspace{0.3cm} Claim~2(b) together with Theorem~\ref{thm: induced removal lemma} (applied to $G^B[Q_{j_0}]$) implies that $G[Q_{j_0}]$ can be made free of $5$-good tetrahedrons by changing the colour of at most $\gamma n^2$ edges inside $Q_{j_0}$. Proposition~\ref{prop: k-good tetradehron free}{\rm (ii)} implies that after such a change, all blue edges inside $Q_{j_0}$ form a disjoint collection of stars and cliques. We are now done by setting $T$ to be $K_n$ minus this disjoint collection of stars and cliques. For the case $k=4$, the proof is again almost identical to the case $k\geq 6$, except that instead of Claim~2(a) we prove the following even weaker claim, which follows in a similar way. \vspace{0.3cm} \noindent {\bf Claim 2(c):} \emph{$G[Q_{j_0}]$ contains fewer than $c n^4$ $4$-good tetrahedrons.} \vspace{0.3cm} Claim~2(c) together with Theorem~\ref{thm: induced removal lemma} (applied to $G^B[Q_{j_0}]$) implies that $G[Q_{j_0}]$ can be made free of $4$-good tetrahedrons by changing the colour of at most $\gamma n^2$ edges inside $Q_{j_0}$. Proposition~\ref{prop: k-good tetradehron free}{\rm (iii)} implies that after such a change, all blue edges inside $Q_{j_0}$ form a disjoint collection of suns. We are now done by setting $T$ to be $K_n$ minus this disjoint collection of suns. \end{proof} \subsection{Approximate structure of typical induced $C_{2k}$-free graphs}\label{subsec: rough 3} We are now in a position to prove the main result of this section. \removelastskip\penalty55\medskip\noindent{\bf Proof of Lemma~\ref{lem: rough structure}.} Choose $n_0\in \mathbb{N}$ and $\varepsilon, \delta, \gamma, \beta>0$ such that $1/n_0\ll \varepsilon \ll \delta \ll \gamma \ll \beta \ll \eta, 1/k$. Let $\varepsilon':= 2\varepsilon$ and $n\geq n_0$. First we claim that $F(n,k)\geq 2^{t_{k-1}(n)}$. To see this, first note that any graph $G$ that contains $\overline{T_{k-1}(n)}$ is induced-$C_{2k}$-free (since for any set of $2k$ vertices on $G$, $3$ of them must form a triangle). Moreover, there are precisely $2^{t_{k-1}(n)}$ such graphs for any given labelling of the vertices, which proves the claim. By Theorem~\ref{thm: containers} (with $C_{2k}, n$ and $\varepsilon'$ taking the roles of $H, N$ and $\varepsilon$ respectively) there is a collection $\mathcal{C}$ of complete $2$-coloured multigraphs on vertex set $[n]$ satisfying properties (a)--(c). In particular, by (a), every induced-$C_{2k}$-free graph on vertex set $[n]$ is contained in some $G\in \mathcal{C}$. Let $\mathcal{C}_1$ be the family of all those $G\in \mathcal{C}$ for which $|G_R\cap G_B|\geq t_{k-1}(n) - \varepsilon' n^2$. Then the number of (labelled) induced-$C_{2k}$-free graphs not contained in some $G\in \mathcal{C}_1$ is at most \[ |\mathcal{C}| \, 2^{t_{k-1}(n) - \varepsilon' n^2} \leq 2^{- \eps n^2} F(n,k), \] because $|\mathcal{C}|\leq 2^{n^{2-\varepsilon'}}$, by (c), and $F(n,k)\geq 2^{t_{k-1}(n)}$. We claim that for every $G\in \mathcal{C}_1$ there exists a complete $2$-coloured multigraph $\tilde{G}$ and a $k$-template $T$ on partition $Q=\{Q_0, Q_1,\dots, Q_{k-2}\}$ such that $$\tilde{G}_R\cap Q_i^{(2)}=E(T[Q_i]) \hspace{0.6cm} \text{and} \hspace{0.6cm} \tilde{G}_R\cap \tilde{G}_B\cap Q_i^{(2)}=\emptyset$$ for every $i\in \{0,1,\dots, k-2\}$, and $\text{dist}(G, \tilde{G})\leq \eta n^2$. (Note that this claim implies that every induced-$C_{2k}$-free graph contained in $G$ can be made into a $k$-template by changing a total of at most $\eta n^2$ edges within the vertex classes $Q_i$.) Indeed, by (b), each $G\in \mathcal{C}_1$ contains at most $\eps' n^{2k}$ copies of $C_{2k}$. Thus by Lemma~\ref{lem: multigraph removal lemma} there exists a complete $2$-coloured multigraph $G'$ on the same vertex set that is $C_{2k}$-free, such that $\text{dist}(G, G')\leq \delta n^2$. Then $|G'_R\cap G'_B|\geq t_{k-1}(n) - (\varepsilon'+\delta) n^2$. Thus by Lemma~\ref{lem: specialised stability} there exists a complete $2$-coloured multigraph $G''$ on the same vertex set, with $G''_R\cap G''_B = E(T_{k-1}(n))$ and such that $\text{dist}(G', G'')\leq \gamma n^2$. Note that $G''$ can contain at most $\gamma n^{2k}$ copies of $C_{2k}$, since $G'$ is $C_{2k}$-free. Let $Q=\{Q_0, Q_1, \dots, Q_{k-2}\}$ be the unique $(k-1)$-partition of $V(G'')$ such that no partition class induces an edge in $G''_R\cap G''_B$. Thus by Lemma~\ref{lem: specialised removal lemma}, there exists a $k$-template $T=(V(G), E^T)$ on $Q$ such that $|G''_R\triangle E^T|\leq \beta n^2$. Define $\tilde{G}$ to be the $2$-coloured multigraph with $\tilde{G}_R\cap \tilde{G}_B = G''_R\cap G''_B$ and $\tilde{G}_R\cap Q_i=E(T[Q_i])$ for every $i\in \{0,1,\dots, k-2\}$. Then $\text{dist}(G, \tilde{G})\leq (\delta+\gamma+\beta)n^2 \leq \eta n^2$, and $\tilde{G}$ satisfies the required properties. This proves the claim and thus the lemma. \endproof \section{The number of $k$-templates}\label{sec: number of templates} For $k\geq 4$ we denote the set of all $k$-templates on $n$ vertices by $T(n,k)$. Let $T_Q(n,k)$ denote the set of all $k$-templates on $n$ vertices for which $Q$ is an ordered $(k-1)$-partition. The aim of this section is to estimate $|T_Q(n,k)|$ and $|T(n,k)|$ (see Lemmas~\ref{size of template} and~\ref{the number of templates} respectively). Before we start with this we need to introduce some more notation. A \emph{$k$-sun} is defined as follows. \begin{itemize} \item If $k=4$, a $k$-sun is any sun (as defined in Section~\ref{subsec: rough 3}). \item If $k=5$, a $k$-sun is a star or a clique. \item If $k\geq 6$, a $k$-sun is a star or a triangle. \end{itemize} Note that the results of this section are only needed for Theorem~\ref{main theorem} (and not Theorem~\ref{secondary theorem}) and so we would only need to consider the case $k\geq 6$. However, including the cases $k=4,5$ makes little difference to the proofs, and are also interesting in their own right, so we work with all $k\geq 4$ throughout this section. Let $F_k(n)$ denote the set of all $n$-vertex graphs whose complement is a disjoint union of $k$-suns. Define $f_k(n):= |F_k(n)|$. A pair of vertices $x,y$ is called a {\em twin pair} if $N(x)\backslash \{y\} = N(y)\backslash \{x\}$. The following two lemmas give some estimates of the value of $f_k(n)$. Note that we do not make use of the upper bound in Lemma~\ref{f-estimate-1} anywhere in this paper, but we include it for its intrinsic interest. It would not be difficult to obtain more accurate bounds, though an asymptotic formula would probably require more work. \begin{lemma}\label{f-estimate-1} For all $n\in \mathbb{N}$ and $k\geq 4$, $$2^{n\log n-en\log\log n} \leq f_k(n) \leq 2^{n\log n - n \log\log n + n}.$$ \end{lemma} \begin{proof} Let $P(n)$ denote the number of partitions of an $n$ element set. It is well known (see e.g.~\cite{NGdB}) that $$2^{n\log n-en\log\log n} \leq P(n)\leq 2^{n\log n-n\log\log n}.$$ We will count the number $f_k(n)$ of graphs $G\in F_k(n)$. Note that $f_k(n)\geq P(n)$ follows by considering each partition class as the vertex set of a star in $\overline{G}$. This then immediately yields the lower bound in Lemma~\ref{f-estimate-1}. Now note that if we choose a partition of $[n]$ into the vertex sets of disjoint suns in $\overline{G}$ (for which there are at most $2^{n\log n-n\log\log n}$ choices), and then for every vertex choose whether the vertex will be in the body of its sun or side of its sun (for which there are a total of $2^n$ choices), we can generate every possible graph $G\in F_k(n)$ (note that some such graphs can be generated by multiple different choices). This yields the upper bound in Lemma~\ref{f-estimate-1}. \end{proof} \begin{lemma}\label{f-estimate-2} For $k \geq 4$ and $n>s\geq 10^7$, $$ s^{s/2} \leq \frac{f_k(n)}{f_k(n-s)} \hspace{0.6cm} \text{and} \hspace{0.6cm} \frac{f_k(n)}{f_k(n-1)} \leq n^2.$$ \end{lemma} \begin{proof} By Lemma \ref{f-estimate-1}, $f_k(n) \geq f_k(s) f_k(n-s) \geq 2^{s\log{s}-es\log\log{s}} f_k(n-s) \geq 2^{s\log s/2} f_k(n-s)$, which gives us the lower bound in the statement of the lemma. For the upper bound, note that every graph in $F_k(n)$ has a twin pair. For any twin pair $i,j \in [n]$ the number of graphs in $F_k(n)$ for which $i,j$ are twins is at most $2 f_k(n-1)$, since every such graph can be obtained from a graph in $F_k(n-1)$ on vertex set $[n]\setminus \{i\}$ by adding the vertex $i$ and choosing whether to add the edge $ij$ (note that all other edges incident to $i$ are prescribed, since $i,j$ are twins). Thus $$f_k(n) \leq \sum_{0<i\leq n-1} \sum_{i<j\leq n} 2f_k(n-1) \leq n^2 f_k(n-1),$$ as required. \end{proof} The following proposition can be proved by a simple but tedious calculation, which we omit here\COMMENT{Note that (i) is taken directly from~\cite{KOTZ}. The proof of (ii) is as follows.\\ \noindent{\bf Proof of (ii):} Without loss of generality we assume $T_{k-1}(n-s)$ to be such that its vertex classes are subsets of the corresponding vertex classes of $T_{k-1}(n)$. Let $S:=V(T_{k-1}(n))\backslash V(T_{k-1}(n-s))$. We may assume that $T_{k-1}(n)[S]=T_{k-1}(s)$. Then $$t_{k-1}(n)=e(T_{k-1}(n))=e(T_{k-1}(n-s))+\left(\sum_{v\in S} d_{T_{k-1}(n)}(v)\right) - e(T_{k-1}(n)[S]),$$ and \begin{itemize} \item $e(T_{k-1}(n-s))=t_{k-1}(n-s)$, \item $e(T_{k-1}(n)[S])=t_{k-1}(s)$, \item For every $v\in S$, $d_{T_{k-1}(n)}(v)\geq (k-2)n/(k-1)-(k-2)$. Indeed, every vertex class of $T_{k-1}(n)$ has size at least $\lfloor n/k-1 \rfloor\geq n/(k-1) -1$, and $v$ is adjacent to all vertices in every vertex class that is not its own (of which there are $k-2$). \end{itemize} Putting all this together yields the result.}. \begin{proposition}\label{omitted proposition} Let $k, n\in \mathbb{N}$ with $n\geq k\geq 2$ and let $0<s<n$. \begin{enumerate}[{\rm (i)}] \item Suppose $G$ is a $k$-partite graph on $n$ vertices in which some vertex class $A$ satisfies $|A-n/k|\geq s$. Then $$e(G)\leq t_{k}(n)-s\left( \frac{s}{2}-k\right).$$ \item $t_{k-1}(n)\geq t_{k-1}(n-s) + sn(k-2)/(k-1)-s(k-2)-t_{k-1}(s)$. \end{enumerate} \end{proposition} \begin{lemma}\label{size of template} Let $k\geq 4$. There exists $n_0\in \mathbb{N}$ such that for every $n\geq n_0$ and every ordered $(k-1)$-partition $Q$ of $[n]$, the number of $k$-templates on $Q$ satisfies $$|T_Q(n,k)|\leq 2^{6 (\log n)^2}2^{t_{k-1}(n)}f_k\left(n_k \right),$$ where we recall that $n_k:= \left\lceil n/(k-1) \right\rceil$. \end{lemma} \begin{proof} Denote the classes of $Q$ by $Q_0,Q_1,\dots, Q_{k-2}$ and let $b:= ||Q_0| - \lceil \frac{n}{k-1} \rceil|$. Then by Proposition~\ref{omitted proposition}(i) the number of $k$-templates on this partition is at most $$f_k(|Q_0|) 2^{\sum_{0\leq i<j\leq k-2} |Q_i||Q_j|} \leq f_k\left(n_k +b\right) 2^{t_{k-1}(n) - b\left(b/2-(k-1)\right)}.$$ Let $h(b):=f_k(n_k +b) 2^{t_{k-1}(n) - b(b/2-(k-1))}$. Then by Lemma \ref{f-estimate-2}, $$\frac{h(b+1)}{h(b)} \leq \left(\frac{n}{k-1}+b+2\right)^2 2^{-\left((2b+1)/2-(k-1)\right)}. $$ Thus $h(b)$ is a decreasing function for $b\geq 3\log n$. This together with Lemma \ref{f-estimate-2} gives us that the number of $k$-templates on $Q$ is at most \begin{align*} h(b) &\leq f_k\left(n_k + 3\log n\right) 2^{t_{k-1}(n)} \leq (n^2)^{3\log n} 2^{t_{k-1}(n)} f_k\left(n_k \right)\\ &= 2^{6 (\log n)^2}2^{t_{k-1}(n)}f_k\left(n_k \right), \end{align*} as required. \end{proof} We call a component of a graph \emph{non-trivial} if it contains at least $2$ vertices. The proof of Lemma~\ref{the number of templates} will make use of the following proposition. \begin{proposition}\label{prop: two components} Let $k\geq 4$. There exists $n_0\in \mathbb{N}$ such that the following holds for every $n\geq n_0$. Let $Q$ be a balanced ordered $(k-1)$-partition of $[n]$. The proportion of $k$-templates $G$ on $Q$ that are such that $\overline{G}[Q_0]$ has at most one non-trivial component is at most $2^{-n}$. \end{proposition} \begin{proof} Since $Q$ is balanced, the number of $k$-templates on $Q$ is at least $2^{t_{k-1}(n)}f_k(\lfloor \frac{n}{k-1}\rfloor )$. We can generate all possible edge sets for $G[Q_0]$ such that $\overline{G}[Q_0]$ has at most one non-trivial component in the following way. Note that for every such $G[Q_0]$, $\overline{G}[Q_0]$ contains at most one disjoint sun $S$ of order at least two. For every vertex in $Q_0$ we choose whether it will belong to the body of $S$, the side of $S$, or neither (for which there are a total of at most $3^n$ choices). Hence the number of $k$-templates $G$ on $Q$ that are such that $\overline{G}[Q_0]$ has at most one non-trivial component is at most $3^{n} 2^{t_{k-1}(n)}$. Since we have by Lemma~\ref{f-estimate-1} that $f_k(m)\geq 2^{m\log m-em\log\log m}$ for all $m\in \mathbb{N}$, the result follows (with some room to spare). \end{proof} The following trivial observation will be useful in the proof of Lemma~\ref{the number of templates}. \begin{equation}\label{induced four cycle observation} \textrm{If a graph $G$ is a disjoint union of suns then $G$ contains no induced $4$-cycles.} \end{equation} \begin{lemma}\label{the number of templates} For every $k\geq 4$ there exists $n_0\in \mathbb{N}$ such that the following holds for all $n\geq n_0$, where we recall that $n_k:= \left\lceil n/(k-1) \right\rceil$. The number of $k$-templates on vertex set $[n]$ satisfies $$|T(n,k)| \geq \frac{(k-1)^n}{2(k-2)! n^k}2^{t_{k-1}(n)}f_k\left(n_k \right).$$ \end{lemma} \begin{proof} Choose $n_0$ such that $1/n_0\ll 1/k$, and let $n\geq n_0$. Given a $k$-template $G$ on vertex set $[n]$ and an ordered $(k-1)$-partition $Q=(Q_0,\{Q_1,\dots,Q_{k-2}\})$ of $[n]$, we say that $G$ is $Q$-compatible if $G$ is a $k$-template on $Q$ and the following hold: \begin{enumerate} \item[$(\alpha)$] Whenever $\ell\leq 2k$ and $0\leq i\leq k-2$ and $v_1,v_2,\dots, v_{\ell} \in V(G)\setminus Q_i$, we have that $$|\overline{N}_{Q_i}(\{v_1,v_2,\dots,v_{\ell}\})| \geq \frac{n}{2^{\ell+1}(k-1)}.$$ \item[$(\beta)$] $\overline{G}[Q_0]$ has at least $2$ non-trivial components. \end{enumerate} \vspace{0.3cm} \noindent {\bf Claim 1:} \emph{Given a balanced ordered $(k-1)$-partition $Q=(Q_0,\{Q_1,\dots,Q_{k-2}\})$ of $[n]$, the number of $k$-templates $G$ on vertex set $[n]$ that are $Q$-compatible is at least $2^{t_{k-1}(n)-1}f_k(n_k)/n^2$.} \noindent Indeed, consider a random graph $G$ where for each potential crossing edge with respect to $Q$ we choose the edge to be present or not, each with probability $1/2$, independently; we let $G[Q_0]$ be one of the $f_k(|Q_0|)$ graphs in $F_k(|Q_0|)$, chosen uniformly at random; and we choose all edges to be present inside $Q_i$ for every $i>0$. So each $k$-template on $Q$ is equally likely to be generated. Note that the number of potential crossing edges with respect to $Q$ is $2^{t_{k-1}(n)}$. This together with Lemma~\ref{f-estimate-2} implies that the number of graphs in the probability space is at least $2^{t_{k-1}(n)}f_k(n_k)/n^2$. By Lemma~\ref{Chernoff Bounds}(ii) and Proposition~\ref{prop: two components} respectively, we have that at least half of all graphs $G$ in the probability space satisfy $(\alpha)$ and $(\beta)$, which proves the claim. \vspace{0.3cm} \vspace{0.3cm} \noindent {\bf Claim 2:} \emph{Given two balanced ordered $(k-1)$-partitions $Q=(Q_0,\{Q_1,\dots,Q_{k-2}\})$ and $Q'=(Q_0',\{Q_1',\dots,Q'_{k-2}\})$ of $[n]$, and a $k$-template $G$ on $[n]$ that is both $Q$-compatible and $Q'$-compatible, there exist $k$ vertices $u_0,v_0,v_1,\dots, v_{k-2}\in [n]$ that are such that $G[\{u_0,v_0,v_1,\dots, v_{k-2}\}]$ contains exactly one edge $u_0v_0$ and $u_0\in Q_0\cap Q'_0$ and $v_i\in Q_i\cap Q_i'$ for all $i\geq 0$.} \noindent To show this, we first choose a set of $2k$ vertices $U=\{u_{0,1}, w_{0,1}, u_{0,2}, w_{0,2}, u_1, w_1, \dots, u_{k-2}, w_{k-2}\}$ such that $u_{0,1}, w_{0,1}, u_{0,2}, w_{0,2}\in Q_0$ and $u_i, w_i\in Q_i$ for every $i>0$ and $$E(G[U])=\{u_{0,1}u_{0,2}, u_{0,2}w_{0,1}, w_{0,1}w_{0,2}, w_{0,2}u_{0,1}, u_1 w_1, \dots, u_{k-2} w_{k-2}\}.$$ This is possible since $G$ satisfies $(\alpha), (\beta)$ with respect to $Q$. Now if there exist distinct $i,j>0$ such that $u_i, w_i, u_j, w_j \in Q_0'$ then $\overline{G}[Q_0']$ contains the induced $4$-cycle $u_i u_j w_i w_j$, which by (\ref{induced four cycle observation}) contradicts the fact that $G$ is a $k$-template on $Q'$. So, by relabelling vertices if necessary, we may assume that $u_2,\dots, u_{k-2}\notin Q_0'$. If $u_{0,1},w_{0,1}\notin Q_0'$ then by the pigeon-hole principle there must exist $i>0$ such that $Q_i'$ contains at least $2$ elements of $\{u_{0,1},w_{0,1}, u_2,\dots, u_{k-2} \}$, contradicting the assumption that $G[Q_i']$ is a clique. So, by relabelling vertices if necessary, we may assume that $u_{0,1}\in Q_0'$, and similarly that $u_{0,2}\in Q_0'$. Now if $u_1, w_1 \in Q_0'$ then $\overline{G}[Q_0']$ contains the induced $4$-cycle $u_{0,1} u_1 u_{0,2} w_1$, which by (\ref{induced four cycle observation}) contradicts the fact that $G$ is a $k$-template on $Q'$. So, by relabelling vertices if necessary, we may assume that $u_1\notin Q_0'$, and thus $u_1,\dots, u_{k-2}\notin Q_0'$. Recall that for all $i>0$, $G[Q_i']$ is a clique, so $Q_i'$ can contain at most one vertex in $\{u_1,\dots, u_{k-2}\}$. Thus we may assume, by relabelling indices if necessary, that $u_{0,1}, u_{0,2}\in Q_0\cap Q_0'$ and $u_i\in Q_i\cap Q_i'$ for every $i>0$. So setting $u_0:= u_{0,1}, v_0:= u_{0,2}$ and $v_i:= u_i$ for all $i>0$ yields the required set of vertices. \vspace{0.3cm} \vspace{0.3cm} \noindent {\bf Claim 3:} \emph{If there exist balanced ordered $(k-1)$-partitions $Q=(Q_0,\{Q_1,\dots,Q_{k-2}\})$ and $Q'=(Q_0',\{Q_1',\dots,Q'_{k-2}\})$ of $[n]$, and a $k$-template $G$ on $[n]$ that is both $Q$-compatible and $Q'$-compatible, then $Q=Q'$.} \noindent Consider any $k$ vertices $u_0, v_0,\dots, v_{k-2}\in V(G)$ that are such that $G[\{u_0,v_0,v_1,\dots, v_{k-2}\}]$ contains exactly one edge $u_0v_0$ and $u_0\in Q_0\cap Q'_0$ and $v_i\in Q_i\cap Q_i'$ for all $i\geq 0$. Such vertices exist by Claim~2. For $i>0$ define $$\overline{N}_i := \overline{N}_{Q_i}( \{u_0,v_0,\dots, v_{k-2}\} \setminus \{v_i\}).$$ $$\overline{N}'_i := \overline{N}_{Q'_i}( \{u_0,v_0,\dots, v_{k-2}\} \setminus \{v_i\}).$$ Since both $\overline{N}_i$ and $\overline{N}'_i$ are subsets of the common non-neighbourhood of $\{u_0,v_0,v_1\dots, v_{k-2}\} \setminus \{v_i\}$, neither can intersect $Q_j$ or $Q'_j$ for $j\notin \{0,i\}$. Note that all vertices in $\overline{N}_i$ are adjacent. Thus $|\overline{N}_i \cap Q'_0| \leq 1$, since otherwise $\overline{G}[Q'_0]$ contains an induced $4$-cycle on $u_0,v_0$ together with $2$ vertices from $\overline{N}_i$, which by (\ref{induced four cycle observation}) contradicts the fact that $G$ is a $k$-template on $Q'$. Similarly, $|\overline{N}'_i \cap Q_0| \leq 1$. Define $$\overline{N}^{\dagger}_i :=(\overline{N}_i \cup \overline{N}'_i) \setminus (Q_0\cup Q'_0).$$ Then $\overline{N}^{\dagger}_i \subseteq Q_i\cap Q_i'$. Now we consider any vertex $w\in Q_0$. Since $G$ satisfies $(\alpha)$ with respect to $Q$, we have that for every $i> 0$, \begin{align}\label{eqn: counting templates} |\overline{N}_{Q'_i}(w)|&\geq |\overline{N}(w)\cap\overline{N}^{\dagger}_i | \geq |\overline{N}(w)\cap\overline{N}_i|-1 \\ &= |\overline{N}_{Q_i}( \{u_0,v_0,\dots, v_{k-2},w\}\setminus \{v_i\})| -1 \geq \frac{n}{2^{k+1}(k-1)} -1 \geq 1.\nonumber \end{align} Thus $w$ must belong to $Q'_0$, since $G[Q_i']$ is a clique for every $i> 0$. Hence $Q_0\subseteq Q'_0$. In the same way we can show that $Q'_0\subseteq Q_0$. Thus $Q_0=Q'_0$. Now we consider any vertex $w\in Q_j$, for $j> 0$. Since $G$ satisfies $(\alpha)$ with respect to $Q$, we have (similarly to (\ref{eqn: counting templates})) that for every $i\neq j$ with $i> 0$, $$|\overline{N}_{Q'_i}(w)|\geq |\overline{N}(w)\cap \overline{N}^{\dagger}_i | \geq 1.$$ Thus $w\in Q_0'\cup Q_j'$. Together with the fact that $Q_0= Q'_0$ this implies that $w\in Q'_j$. Thus $Q_j\subseteq Q'_j$ for all $j>0$. Hence $Q=Q'$, which proves the claim. \vspace{0.3cm} We now count the number of balanced ordered $(k-1)$-partitions. Since the vertex classes of a balanced ordered $(k-1)$-partition of $[n]$ have sizes $\lceil \frac{n}{k-1} \rceil,\lceil \frac{n-1}{k-1} \rceil, \dots, \lceil \frac{n-k+2}{k-1} \rceil$, the number of such $(k-1)$-partitions is \begin{equation*} \frac{1}{(k-2)!}{\binom{n}{\lceil \frac{n}{k-1} \rceil,\lceil \frac{n-1}{k-1} \rceil, \dots, \lceil \frac{n-k+2}{k-1} \rceil}}. \end{equation*} This together with Claims~1 and~3 implies that \begin{equation}\label{eqn: counting templates 2} |T(n,k)| \geq \frac{1}{2(k-2)!n^2}{\binom{n}{\lceil \frac{n}{k-1} \rceil,\lceil \frac{n-1}{k-1} \rceil, \dots, \lceil \frac{n-k+2}{k-1} \rceil}}2^{t_{k-1}(n)}f_k(n_k). \end{equation} Now note that if $a_1+\dots + a_{k-1} = n$, then ${\binom{n}{a_1,a_2,\dots, a_{k-1}}}$ is maximized by taking $a_j := \lceil \frac{n-j+1}{k-1}\rceil $ for every $j$. This implies that $$(k-1)^n = \sum_{a_1+\dots+a_{k-1}=n}{\binom{n}{a_1,a_2,\dots, a_{k-1}} } \leq n^{k-2} {\binom{n}{\lceil \frac{n}{k-1} \rceil,\lceil \frac{n-1}{k-1} \rceil, \dots, \lceil \frac{n-k+2}{k-1} \rceil}},$$ which together with (\ref{eqn: counting templates 2}) implies the result. \end{proof} \section{Properties of near-$k$-templates}\label{sec: set-up} In this section we collect some properties of graphs which are close to being $k$-templates. In particular, when $k\geq 6$, this means we consider graphs $G$ which have a vertex partition such that each vertex class induces on $G$ an almost complete graph. (As in the previous section, we will need the results of this section for our main results only for the case $k\geq 6$, but we prove the results for all $k\geq 4$ since it makes little difference to the proofs.) More formally, given $k\geq 4$, a graph $G$ on vertex set $[n]$, and an ordered $(k-1)$-partition $Q$ of $[n]$ we define $$h(Q,G):=\sum_{i=0}^{k-2} |E(\overline{G}[Q_i])|.$$ We say $Q$ is an {\em optimal ordered $(k-1)$-partition of $G$} if $h(Q,G)$ is the minimum value $h(Q',G)$ takes over all partitions $Q'$ of $[n]$. Note that if $h(Q,G)=0$ then $G$ is a $k$-template on $Q$, and that the following also holds. \begin{equation}\label{eqn: near templates} \textrm{If $k\geq 6$ then every $k$-template $G'$ on $Q$ satisfies $h(Q,G')\leq n$.} \end{equation} Note that (\ref{eqn: near templates}) does not hold for $k\in \{4,5\}$. We will require the following definitions in what follows. \begin{itemize} \item Recall that $F(n,k)$ denotes the set of all labelled induced-$C_{2k}$-free graphs on vertex set $[n]$. \item Given $n\in \mathbb{N}$, $k\geq 4$, and $\eta>0$, we define $F(n,k,\eta)\subseteq F(n,k)$ to be the set of all graphs in $F(n,k)$ such that $h(Q,G)\leq \eta n^2$ for some optimal ordered $(k-1)$-partition $Q$ of $G$. \item Given further an ordered $(k-1)$-partition $Q=(Q_0,\{Q_1,\dots, Q_{k-2}\})$ of $[n]$ we define $F_Q(n,k)\subseteq F(n,k)$ to be the set of all graphs in $F(n,k)$ for which $Q$ is an optimal ordered $(k-1)$-partition \item Similarly we define $F_Q(n,k,\eta)\subseteq F(n,k,\eta)$ to be the set of all graphs in $F(n,k,\eta)$ for which $Q$ is an optimal ordered $(k-1)$-partition. \end{itemize} Recall that, given a graph $G$ on vertex set $[n]$ and an index $i\in \{0,1,\dots, k-2\}$, we let $d^i_{G,Q}(x), \overline{d}^i_{G,Q}(x)$ denote the number of neighbours and non-neighbours of $x$ in $Q_i$, respectively. The following proposition follows immediately\COMMENT{{\bf Proof of Proposition~\ref{optimality}:} Suppose for a contradiction that there exists a vertex $x \in Q_i$ that satisfies $\overline{d}^j_{G,Q}(x) < \overline{d}^i_{G,Q}(x)$. Define an ordered $(k-1)$-partition $Q'=(Q'_0,\{Q'_1,\dots, Q'_{k-2}\})$ by $Q'_j:= Q_j\cup \{v\}, Q'_i:= Q_i\setminus \{x\}$ and $Q'_\ell:=Q_\ell$ for all $\ell\in \{0,1,\dots, k-2\}\backslash\{i,j\}$. Then $h(Q',G) = h(Q,G)+\overline{d}^j_{G,Q}(x)-\overline{d}^i_{G,Q}(x)< h(Q,G)$, which contradicts the optimality of $Q$ and hence completes the proof.} from the definition of optimality. \begin{proposition}\label{optimality} Let $k\geq 4$, let $\eta>0$, let $Q=(Q_0,\{Q_1,\dots, Q_{k-2}\})$ be an ordered $(k-1)$-partition of $[n]$, and let $G\in F_Q(n,k,\eta)$. For any two distinct indices $i,j\in \{0,1,\dots, k-2\}$ every vertex $x\in Q_i$ satisfies $\overline{d}^j_{G,Q}(x) \geq \overline{d}^i_{G,Q}(x)$. \end{proposition} Next we show that for most graphs which are close to being $k$-templates, the bipartite graphs between the partition classes are quasirandom. Given $k\geq 4$, $\nu=\nu(n)>0$ and an ordered $(k-1)$-partition $Q=(Q_0,\{Q_1,\dots, Q_{k-2}\})$ of $[n]$, we define the following properties that a graph on vertex set $[n]$ may satisfy with respect to $Q$. \begin{enumerate} \item[$({\rm F}1)_{\nu}$] If $U_i\subseteq Q_i$ and $U_j\subseteq Q_j$ with $|U_i||U_j|\geq \nu^2 n^2$ for distinct $0\leq i,j\leq k-2$, then $ \frac{1}{4} \leq \frac{|e(U_i,U_j)|}{|U_i||U_j|} \leq \frac{3}{4}$. \item[$({\rm F}2)_{\nu}$] $||Q_i|-\frac{n}{k-1}|\leq \nu n$ for every $0\leq i\leq k-2$. \end{enumerate} Given $\eta, \mu>0$ we define $F_Q(n,k,\eta,\mu)$ to be the set of all graphs in $F_Q(n,k,\eta)$ that satisfy $({\rm F}1)_{\mu}$ and $({\rm F}2)_{\mu}$ with respect to $Q$. \begin{lemma}\label{pre regular lemma} Let $n\geq k\geq 4$, let $0<\eta<1$, let $6k/n\leq \nu=\nu(n)\leq 1$, let $6\log n\leq m=m(n)\leq 10^{-11}n^2$, and let $Q=(Q_0,\{Q_1,\dots, Q_{k-2}\})$ be an ordered $(k-1)$-partition of $[n]$. Then the following hold. \begin{enumerate}[{\rm (i)}] \item The number of graphs $G$ in $F_Q(n,k,\eta)$ that fail to satisfy $({\rm F}1)_{\nu}$ with respect to $Q$ and that have at most $m$ internal non-edges is at most $2^{t_{k-1}(n) + \xi(m/n^2)n^2} 2^{2n+1} \exp(-\nu^2n^2/32)$. \item The number of graphs $G$ in $F_Q(n,k,\eta)$ that fail to satisfy $({\rm F}2)_{\nu}$ with respect to $Q$ and that have at most $m$ internal non-edges is at most $2^{t_{k-1}(n) + \xi(m/n^2)n^2} \exp( -\nu^2 n^2/6 )$. \end{enumerate} \end{lemma} \begin{proof} For both (i) and (ii) we consider constructing such a graph $G$. By (\ref{entropy bound}) there are at most $\binom{n^2}{\leq m} \leq 2^{\xi(m/n^2) n^2}$ choices for the internal edges of $G$. We first prove (i). For a given choice of internal edges, consider the random graph $H$ where for each possible crossing edge with respect to $Q$ we choose the edge to be present or not, with probability $1/2$, independently. Note that the total number of ways to choose the crossing edges is at most $2^{t_{k-1}(n)}$, and each possible configuration of crossing edges is equally likely. So an upper bound on the number of graphs $G \in F_Q(n,k,\eta)$ that fail to satisfy property $({\rm F}1)_{\nu}$ with respect to $Q$ and that have at most $m$ internal non-edges is \begin{equation}\label{F intermediate bound} 2^{t_{k-1}(n) + \xi(m/n^2)n^2 }\mathbb{P}(H \text{ fails to satisfy } ({\rm F}1)_{\nu} \text{ with respect to } Q). \end{equation} Note that the number of choices for $U_i\subseteq Q_i, U_j\subseteq Q_j$ with $|U_i||U_j|\geq \nu^2 n^2$ is at most $2^{2n}$ and that $\mathbb{E}(e(U_i,U_j)) = |U_i||U_j|/2 \geq \nu^2 n^2/2$. Hence by Lemma~\ref{Chernoff Bounds}, \begin{equation*} \mathbb{P}(H \text{ fails to satisfy } ({\rm F}1)_{\nu} \text{ with respect to } Q) \leq 2^{2n+1} \exp\left(-\frac{\nu^2 n^2}{32}\right). \end{equation*} This together with (\ref{F intermediate bound}) yields the result. We now prove (ii). If $||Q_i|-\frac{n}{k-1}|>\nu n$ for some $0\leq i\leq k-2$, then by Proposition~\ref{omitted proposition}(i) the number of crossing edges in $G$ is at most $$ t_{k-1}(n) - \frac{\nu^2n^2}{3}.$$ We can conclude that the number of $G \in F_Q(n,k,\eta)$ that fail to satisfy $({\rm F}2)_{\nu}$ with respect to $Q$ and that have at most $m$ internal non-edges is at most \begin{equation*} 2^{\xi(m/n^2)n^2} 2^{t_{k-1}(n) -\frac{\nu^2n^2}{3} }\leq 2^{t_{k-1}(n) + \xi(m/n^2)n^2}\exp\left( -\frac{\nu^2 n^2}{6} \right), \end{equation*} as required. \end{proof} We will apply the following special case of Lemma~\ref{pre regular lemma} in Section~\ref{sec: final calculation} in the proof of Lemma~\ref{induction conclusion}. \begin{corollary}\label{regular lemma} Let $k\geq 4$ and let $0<\eta,\mu<10^{-11}$ be such that $\mu^2 > 24\xi(\eta)$. There exists an integer $n_0 = n_0(\mu,k)$ such that for all $n\geq n_0$ and every ordered $(k-1)$-partition $Q$ of $[n]$, $$ |F_Q(n,k,\eta)\setminus F_Q(n,k,\eta,\mu)| \leq 2^{t_{k-1}(n) - \frac{\mu^2n^2}{100}}.$$ \end{corollary} \begin{proof} We choose $n_0$ such that $1/n_0 \ll \eta, \mu, 1/k$. Applying Lemma~\ref{pre regular lemma} with $\mu, \eta n^2$ playing the roles of $\nu, m$ respectively yields that $$ |F_Q(n,k,\eta)\setminus F_Q(n,k,\eta,\mu)| \leq 2^{t_{k-1}(n)+\xi(\eta)n^2}2^{2n+1} \left( e^{- \frac{\mu^2 n^2}{6} }+ e^{- \frac{\mu^2 n^2}{32} } \right) \leq 2^{t_{k-1}(n) - \frac{\mu^2 n^2}{100}},$$ as required. \end{proof} The next proposition follows immediately from~\cite[Lemma~2.22]{BaBu}. We will use it to find induced copies of $C_{2k}$. (Usually $T$ will be a suitable induced subgraph of $C_{2k}$ and the $A_i,B_i$ will be the intersection of (non-)neighbourhoods of vertices that we have already embedded.) \begin{proposition} \label{building} Let $n_0,k \in \mathbb{N}$ and $\eta,\mu >0$ be chosen such that $k\geq 4$ and $1/n_0\ll \eta \ll \mu \ll 1/k$. Then the following holds for all $n\in \mathbb{N}$ with $n\geq n_0$. Let $Q=(Q_0,\{Q_1,\dots, Q_{k-2}\})$ be an ordered $(k-1)$-partition of $[n]$ and suppose $G\in F_Q(n,k,\eta,\mu)$. Let $I\subseteq \{0,1,\dots, k-2\}$. For every $i\in I$ let $A_i,B_i\subseteq Q_i$ be disjoint with $|A_i|, |B_i| \geq \mu^{1/2} n$. Let $T$ be a $2|I|$-vertex graph with a perfect matching whose edges are $v_iu_i$ for every $i \in I$. Then there exists an injection $f:V(T) \rightarrow V(G)$ such that $f(v_i) \in A_i, f(u_i) \in B_i$ for every $i\in I$, and $f(V(T))$ induces on $G$ a copy of $T$. \end{proposition} Finally we show that if $G$ is close to being a $k$-template then removing a small number of vertices from $G$ does not alter its optimal ordered $(k-1)$-partition very much. Given $m,n\in \mathbb{N}$ and an ordered $(k-1)$-partition $Q$ of $[n]$, we define $\mathcal{P}(Q,m)$ to be the collection of all ordered $(k-1)$-partitions of $[n]$ that can be obtained from $Q$ by moving at most $m$ vertices between partition classes, and possibly choosing a different partition class to be the labelled one. Then it is easy to see that \begin{equation}\label{size mathcal{P}} |\mathcal{P}(Q,m)| \leq k {\binom{n}{m}} k^m \leq k\left(\frac{ekn}{m}\right)^m \leq 2^{m\log (ek^2 n/m)}. \end{equation} Given an ordered $(k-1)$-partition $Q$ of $[n]$ and a set $S\subseteq [n]$, let $Q-S$ denote the ordered $(k-1)$-partition (possibly with some empty classes) obtained from $Q$ by deleting all elements of $S$ from their partition classes. \begin{lemma} \label{pre new partition} Let $k\geq 4$, let $0<\eta,\mu\leq 1/k^3$, let $0<\nu=\nu(n)\leq 1/k^3$, and let $0\leq m=m(n)\leq n^2$ with $\nu^2 > 4m/n^2$ for all $n\in \mathbb{N}$. There exists $n_0\in \mathbb{N}$ such that the following holds for all $n\geq n_0$. Let $Q=(Q_0,\{Q_1,\dots, Q_{k-2}\})$ be an ordered $(k-1)$-partition of $[n]$ and let $S\subseteq [n]$ with $|S|\leq n/k^2$. Then for every $G\in F_Q(n,k,\eta,\mu)$ that satisfies $({\rm F}1)_{\nu}$ with respect to $Q$ and that has at most $m$ internal non-edges, every optimal ordered $(k-1)$-partition of $G - S$ is an element of $\mathcal{P}(Q-S,k^4\nu^2 n)$. \end{lemma} \begin{proof} Let $G\in F_Q(n,k,\eta, \mu)$ have at most $m$ internal non-edges and satisfy $({\rm F}1)_{\nu}$ with respect to $Q$, and let $Q'=(Q_0',\{Q_1',\dots, Q_{k-2}'\})$ be an optimal ordered $(k-1)$-partition of $G-S$. By optimality of $Q'$ it must be that $G-S$ has at most $m$ internal non-edges with respect to $Q'$. For every $i \in \{0,1,\dots,k-2\}$, since $|Q_i - S| \geq n/(k-1) - \mu n - n/k^2 \geq n/k$, the pigeon-hole principle implies that there exists $j \in \{0,1,\dots, k-2\}$ such that $|Q_i\cap Q_{j}'|\geq n/k^2$. We define a function $\sigma$ by setting $\sigma(i)$ to be an index in $\{0,1,\dots,k-2\}$ that satisfies $|Q_i\cap Q_{\sigma(i)}'|\geq n/k^2$, for every $i \in \{0,1,\dots,k-2\}$. Suppose for a contradiction that there exists $i'\in \{0,1,\dots, k-2\}$ with $i\ne i'$ such that $|Q_{i'}\cap Q_{\sigma(i)}'| \geq k^2\nu^2 n$. Then since $G$ satisfies $({\rm F}1)_{\nu}$ with respect to $Q$ we have that the number of internal non-edges in $G-S$ with respect to $Q'$ is at least $|Q_{i}\cap Q_{\sigma(i)}'||Q_{i'}\cap Q_{\sigma(i)}'|/4 \geq \nu^2 n^2/4>m$. This contradicts our previous observation that $G-S$ has at most $m$ internal non-edges with respect to $Q'$. Hence $\sigma$ is a permutation on $\{0,1,\dots,k-2\}$. Moreover $|Q_i \cap Q'_{j}| < k^2\nu^2 n$ for all $j\in \{0,1,\dots, k-2\}$ with $j\neq \sigma(i)$. Let $\mathcal{P}$ be the set of all ordered $(k-1)$-partitions of $[n]\backslash S$ for which such a permutation exists. So by the above we have that for every $G\in F_Q(n,k,\eta)$ that satisfies $({\rm F}1)_{\nu}$ with respect to $Q$ and that has at most $m$ internal non-edges, every optimal ordered $(k-1)$-partition of $G-S$ is an element of $\mathcal{P}$. So it remains to show that $\mathcal{P}\subseteq \mathcal{P}(Q-S,k^4\nu^2 n)$. This follows from the observation that every element of $\mathcal{P}$ can be obtained by starting with the (labelled) $(k-1)$-partition $Q_0\setminus S, Q_1 \setminus S,\dots, Q_{k-2} \setminus S$, applying a permutation of $\{0,1,\dots,k-2\}$ to the partition class labels, then for every ordered pair of partition classes moving at most $k^2\nu^2 n$ elements from the first partition class to the second, and finally unlabelling all but one of the resulting partition classes. \end{proof} The following is an immediate corollary of Lemma~\ref{pre new partition}, applied with $\mu, \eta n^2$ playing the roles of $\nu, m$, respectively. \begin{corollary}\label{new partition} Let $k\geq 4$ and $0<\eta,\mu< 1/k^3$ with $\mu^2 > 4\eta$. There exists $n_0\in \mathbb{N}$ such that the following holds for all $n\geq n_0$. Let $Q=(Q_0,\{Q_1,\dots, Q_{k-2}\})$ be an ordered $(k-1)$-partition of $[n]$ and let $S\subseteq [n]$ with $|S|\leq n/k^2$. Then for every $G\in F_Q(n,k,\eta,\mu)$, every optimal ordered $(k-1)$-partition of $G - S$ is an element of $\mathcal{P}(Q-S,k^4\mu^2 n)$. \end{corollary} \section{Derivation of Theorem~\ref{main theorem} from the main lemma}\label{sec: main proof} The following lemma is the key result in our proof of Theorem~\ref{main theorem}. Together with Lemma~\ref{size of template} it implies that, for $k\geq 6$, almost all induced-$C_{2k}$-free graphs $G$ with a given optimal ordered $(k-1)$-partition are $k$-templates\COMMENT{Note that I do mean Lemma~\ref{size of template} and not Lemma~\ref{the number of templates} here, since the sentence is about induced-$C_{2k}$-free graphs \emph{with a given optimal partition}}. Recall that $n_k:=\left\lceil n/(k-1) \right\rceil$, that $f_k(n)$ and $T_Q(n,k)$ were defined at the beginning of Section~\ref{sec: number of templates}, and that $F_Q(n,k)$ was defined at the beginning of Section~\ref{sec: set-up}. \begin{lemma}\label{induction conclusion} For every $n,k\in \mathbb{N}$ with $k\geq 6$ there exists $C\in \mathbb{N}$ such that the following holds. For every ordered $(k-1)$-partition $Q$ of $[n]$, \begin{align*} |F_Q(n,k)| &\leq |T_Q(n,k)|+ 5C 2^{-n^{\frac{1}{2k^2}}/3}f_k(n_k) 2^{t_{k-1}(n)}. \end{align*} \end{lemma} Lemma~\ref{induction conclusion} will be proved in the remaining sections of this paper. We will now use it to derive Theorem~\ref{main theorem}. \removelastskip\penalty55\medskip\noindent{\bf Proof of Theorem~\ref{main theorem}.} Let $n_0\in \mathbb{N}$ be as in Lemma~\ref{the number of templates}, let $C\in \mathbb{N}$ be as in Lemma~\ref{induction conclusion}, let $n_1\in \mathbb{N}$ satisfy $1/n_1 \ll 1/k$, let $n\in \mathbb{N}$ with $n\geq \max \{n_0, n_1\}$, and let $\mathcal{Q}$ be the set of all ordered $(k-1)$-partitions of $[n]$. Since $T(n,k)\subseteq F(n,k)$ and $T_Q(n,k)\subseteq F_Q(n,k)$ for every $Q\in \mathcal{Q}$, Lemma~\ref{induction conclusion} implies that \begin{align*} |F(n,k)|-|T(n,k)| &= |F(n,k)\backslash T(n,k)| \leq \sum_{Q\in \mathcal{Q}} |F_Q(n,k) \setminus T_Q(n,k)|\\ &= \sum_{Q\in \mathcal{Q}} \left( |F_Q(n,k)| - |T_Q(n,k)| \right)\leq 5C (k-1)^n 2^{-n^{\frac{1}{2k^2}}/3} f_k(n_k) 2^{t_{k-1}(n)}\\ &\leq C 2^{-n^{\frac{1}{2k^2}}/4} \frac{(k-1)^n}{2(k-2)! n^k} f_k(n_k) 2^{t_{k-1}(n)}. \end{align*} This together with Lemma~\ref{the number of templates} implies that $$|F(n,k)|-|T(n,k)| \leq C 2^{-n^{\frac{1}{2k^2}}/4} |T(n,k)|= o(|T(n,k)|),$$ where we use the little $o$ notation with respect to $n$. So $|F(n,k)| = (1+o(1))|T(n,k)|$, as required. \endproof Sections~\ref{sec: estimate 1}--\ref{sec: final calculation} are devoted to proving Lemma~\ref{induction conclusion} by an inductive argument. For the remainder of the paper we fix constants $C, k, n_0\in \mathbb{N}$ with $k\geq 6$ and $\varepsilon, \eta, \mu, \gamma, \beta, \alpha>0$ such that \begin{equation}\label{eqn: hierarchy} \frac{1}{C} \ll \frac{1}{n_0}\ll \varepsilon \ll \eta \ll \mu \ll \gamma \ll \beta \ll \alpha \ll \frac{1}{k}. \end{equation} We also set $M:= R_{2k-2}(\lceil \frac{1}{\gamma}\rceil) +1$, fix an arbitrary integer $n\geq n_0$, and fix an arbitrary ordered $(k-1)$-partition $Q=(Q_0, \{Q_1,\dots, Q_{k-2}\})$ of $[n]$. We make the following inductive assumption in Sections~\ref{sec: estimate 1}, \ref{sec: estimate 2} and~\ref{sec: estimate 3}: for every $n'\leq n-1$, and every ordered $(k-1)$-partition $Q'=(Q'_0, \{Q'_1,\dots, Q'_{k-2}\})$ of $[n']$, $$|F_{Q'}(n',k)\setminus T_Q(n',k)| \leq 5 C 2^{-(n')^{\frac{1}{2k^2}}/3} f_k\left(n'_k \right) 2^{t_{k-1}(n')}.$$ Note that this together with Lemma~\ref{size of template} implies that \begin{equation} \label{induct} |F_{Q'}(n',k)|\leq 6C 2^{6(\log n')^2} f_k\left( n'_k \right) 2^{t_{k-1}(n')}. \end{equation} We now give a number of definitions that will be used in the remaining sections. Given an index $i\in \{0,1,\dots, k-2\}$, we call a vertex $x$ of a graph $G$ {\em $i$-light} if at least one of the following holds. \begin{enumerate}[({A}1)] \item $d^i_{G,Q}(x)\leq \alpha n$. \item $\overline{d}^i_{G,Q}(x) \leq \alpha n$. \item There exists $z\in V(G)$ such that $|N^*_i(x,z)|+|N^*_i(z,x)| \leq \alpha n$. \end{enumerate} (Intuitively, the neighbourhood in $Q_i$ of an $i$-light vertex is `atypical', and this is unlikely to happen.) Given $\psi>0$ and an index $i\in \{0,1,\dots, k-2\}$, we call $\{x,y_1,y_2,y_3\}\subseteq V(G)$ a {\em $(k,x,i,\psi)$-configuration} if it satisfies the following. \begin{enumerate}[(C1)] \item $G[\{x,y_1,y_2,y_3\}]$ is a linear forest. \item $\overline{d}^{j}_{G,Q}(x) \geq 13\cdot 6^k \psi n$ for all $j\in \{0,1,\dots, k-2\} \setminus \{i\} $. \item There exists $i'\neq i$ such that $d^{j}_{G,Q}(x) \geq 13 \cdot 6^k \psi n$ for all $j\in\{0,1,\dots,k-2\} \setminus \{i,i'\}$. \item $\min\{ d^i_{G,Q}(y_j), \overline{d}^i_{G,Q}(y_j)\} \leq \psi^2 n$ for all $j\in [3]$. \end{enumerate} (Intuitively, (C1)--(C3) of the definition of $(k,x,i,\psi)$-configurations are useful for `building' induced copies of $C_{2k}$, so the existence of a $(k,x,i,\psi)$-configuration in an induced-$C_{2k}$-free graph $G$ severely constrains the choices for the remaining edge set of $G$. The bounds arising from this are still not sufficiently strong though; we also need (C4), which gives further constraints on the choices for the remaining edge set of $G$.) We partition $F_Q(n,k,\eta,\mu)$ into the sets $T_Q, F^1_Q, F^2_Q, F^3_Q$ defined as follows. \begin{enumerate}[(F1)] \item[(F0)] $T_Q:= T_Q(n,k)\cap F_Q(n,k,\eta,\mu)$. \item $F^1_Q \subseteq F_Q(n,k,\eta,\mu)\setminus T_Q$ is the set of all remaining graphs $G$ which satisfy one of the following. \begin{enumerate}[{\rm (i)}] \item $G$ contains a $(k,x,i,\psi)$-configuration for some $i\in \{0,1,\dots, k-2\}$, some $x\in V(G)$ and some $\psi \in \{\beta^{1/2}, \beta^2\}$. \item $G$ contains a vertex $x$ which is both $i$-light and $j$-light for some distinct indices $i, j\in \{0,1,\dots, k-2\}$. \end{enumerate} \item $F^2_Q\subseteq F_Q(n,k,\eta,\mu) \setminus (T_Q \cup F^1_Q)$ is the set of all remaining graphs that for some $i\in \{0,1,\dots, k-2\}$ contain a vertex $x\in Q_i$ that satisfies $\overline{d}^{i}_{G,Q}(x), d^{i}_{G,Q}(x) \geq \beta n$. \item $F^3_Q := F_Q(n,k,\eta,\mu) \setminus (T_Q \cup F^1_Q\cup F^2_Q)$ is the set of all remaining graphs. \end{enumerate} Sections~\ref{sec: estimate 1}, \ref{sec: estimate 2} and~\ref{sec: estimate 3} are devoted to proving upper bounds on $|F^1_Q|, |F^2_Q|$ and $|F^3_Q|$ respectively. As mentioned earlier, it turns out that $F^3_Q$ is the class of induced-$C_{2k}$-free graphs which are `extremely close' to being $k$-templates (see Proposition~\ref{beta}). In Section~\ref{sec: final calculation} we will use these bounds to complete the proof of Lemma~\ref{induction conclusion}. \section{Estimation of $|F^1_Q|$}\label{sec: estimate 1} To estimate $|F^1_Q|$ we will bound the number of graphs satisfying (F1)(i) and (F1)(ii) separately. The main difficulty is in estimating those satisfying (F1)(i), i.e. the ones containing a $(k,x,i,\psi)$-configuration. The idea here is that a $(k,x,i,\psi)$-configuration has many potential extensions into an induced copy of $C_{2k}$. More precisely, given a $(k,x,i,\psi)$-configuration $H$ we can find many disjoint `skeleton' graphs $L$ with the same number of components as $H$ such that $H\cup L$ is a linear forest on $2k$ vertices (i.e. $H\cup L$ has a potential extension into an induced $C_{2k}$). Thus each skeleton induces a restriction on further edges that can be added. Since the skeletons are disjoint we obtain many edge restrictions in total, and thus a good bound on the number of graphs containing a $(k,x,i,\psi)$-configuration. The next two propositions are used to formalise the notion of extendibility into an induced $C_{2k}$. (Roughly, in these propositions one can consider $L_1$ as a $(k,x,i,\psi)$-configuration and $L_2$ as an associated skeleton.) \begin{proposition}\label{incomplete} Let $c\geq 1$ and let $L_1, L_2$ be disjoint linear forests, each with exactly $c$ components, such that $|V(L_1)|+|V(L_2)|=2k$. Then there exists a set $E'$ of edges between $V(L_1)$ and $V(L_2)$ such that the graph $(V(L_1)\cup V(L_2),E'\cup E(L_1) \cup E(L_2))$ is isomorphic to $C_{2k}$. \end{proposition} The proof of Proposition~\ref{incomplete} is trivial, and is omitted\COMMENT{{\bf Proof of Proposition~\ref{incomplete}:} For every $j\in [2]$ we denote the components of $L_j$ by $P^1_j,\dots, P^c_j$. For every $i\in [c]$ and $j\in [2]$, if $|V(P^i_j)|>1$ we let the two endpoints of the path $P^i_j$ be denoted $s^i_j, t^i_j$. Otherwise $|V(P^i_j)|=1$ and we let $s^i_j = t^i_j$ be the unique vertex in $V(P^i_j)$. Then $E' := \{ t^1_1s^1_2, t^1_2s^2_1, t^2_1s^2_2, \dots, t^c_1s^c_2, t^c_2s^1_1 \}$ is a set of edges as required.}. Proposition~\ref{incomplete2} follows from an easy application of Proposition~\ref{incomplete}, and we give only a brief sketch of the proof\COMMENT{{\bf Proof of Proposition~\ref{incomplete2}:} We consider cases as follows.\\ \vspace{0.3cm} \noindent{\bf Case 1:} $d_{L_1}(x)=0$.\\ \noindent In this case note that the assumptions that $|V(L_1)|>1$ and $d_{L_1}(x)=0$ imply that $c>1$. Since $d_{L_1}(x) + d_{L_2}(x)=2$, $x$ has degree $2$ in $L_2$. So $L_1 - x$ and $L_2$ both have exactly $c-1\geq 1$ components. Thus applying Proposition~\ref{incomplete} to $L_1 - x$, $L_2$ yields a set $E'$ of edges between $V(L_1)\backslash\{x\}$ and $V(L_2)\backslash\{x\}$ such that the graph $(V(L_1)\cup V(L_2),E'\cup E(L_1) \cup E(L_2))$ is isomorphic to $C_{2k}$, as required.\\ \vspace{0.3cm} \noindent{\bf Case 2:} $d_{L_2}(x)=0$.\\ \noindent In this case recall the assumption that $L_1$ and $L_2 - x$ both have exactly $c$ components. Since $d_{L_1}(x) + d_{L_2}(x)=2$, $x$ has degree $2$ in $L_1$. So applying Proposition~\ref{incomplete} to $L_1$, $L_2 - x$ yields a set $E'$ of edges between $V(L_1)\backslash\{x\}$ and $V(L_2)\backslash\{x\}$ such that the graph $(V(L_1)\cup V(L_2),E'\cup E(L_1) \cup E(L_2))$ is isomorphic to $C_{2k}$, as required.\\ \vspace{0.3cm} \noindent{\bf Case 3:} $d_{L_1}(x)=d_{L_2}(x)=1$.\\ \noindent In this case note that $L_1$ and $L_2$ both have $c$ components. For every $j\in [2]$ we denote the components of $L_j$ by $P^1_j,\dots, P^c_j$. For every $i\in [c]$ and $j\in [2]$, if $|V(P^i_j)|>1$ we let the two endpoints of the path $P^i_j$ be denoted $s^i_j, t^i_j$. Otherwise $|V(P^i_j)|=1$ and we let $s^i_j = t^i_j$ be the unique vertex in $V(P^i_j)$. Without loss of generality we may assume that $t^1_1 = s^1_2 =x$. Then $E' := \{ t^1_2s^2_1, t^2_1s^2_2, \dots, t^c_1s^c_2, t^c_2s^1_1 \}$ is a set of edges between $V(L_1)\backslash\{x\}$ and $V(L_2)\backslash\{x\}$ such that the graph $(V(L_1)\cup V(L_2),E'\cup E(L_1) \cup E(L_2))$ is isomorphic to $C_{2k}$, as required.\\ \vspace{0.3cm} \noindent Since $d_{L_1}(x) + d_{L_2}(x)=2$, this covers all cases and hence completes the proof.}. \begin{proposition}\label{incomplete2} Let $c\geq 1$ and let $L_1, L_2$ be linear forests that satisfy the following. \begin{itemize} \item $V(L_1)\cap V(L_2)=\{x\}$. \item $|V(L_1)|, |V(L_2)| > 1$. \item $d_{L_1}(x) + d_{L_2}(x)=2$. \item $L_1$ and $L_2- \{x\}$ both have exactly $c$ components. \item $|V(L_1)\cup V(L_2)| = 2k$. \end{itemize} Then there exists a set $E'$ of edges between $V(L_1)\backslash\{x\}$ and $V(L_2)\backslash\{x\}$ such that the graph $(V(L_1)\cup V(L_2),E'\cup E(L_1) \cup E(L_2))$ is isomorphic to $C_{2k}$. \end{proposition} \begin{proof} If $d_{L_1}(x)=0$ we apply Proposition~\ref{incomplete} to $L_1-x, L_2$; if $d_{L_2}(x)=0$ we apply Proposition~\ref{incomplete} to $L_1, L_2-x$. If $d_{L_1}(x)=d_{L_2}(x)=1$ one can easily find $E'$ directly. \end{proof} \begin{lemma} \label{F^1} $|F^1_Q| \leq C 2^{- \frac{\beta^2 n}{14^k}} f_k(n_k) 2^{t_{k-1}(n)}$. \end{lemma} \begin{proof} Let $F^1_{Q,(i)}$ denote the set of all graphs in $F^1_Q$ that satisfy (F1)(i). Similarly let $F^1_{Q,(ii)}$ denote the set of all graphs in $F^1_Q$ that satisfy (F1)(ii). Clearly, \begin{equation}\label{eqn: F1 count} |F^1_Q|\leq |F^1_{Q,(i)}| + |F^1_{Q,(ii)}|. \end{equation} We will first estimate the number of graphs in $F^1_{Q,(i)}$. Any graph $G\in F^1_{Q,(i)}$ can be constructed as follows. We first choose $\psi\in \{\beta^2, \beta^{1/2}\}$, and then perform the following steps. \begin{itemize} \item We choose an index $i\in \{0,1,\dots, k-2\}$, a set of three (labelled) vertices $Y=\{y_1,y_2,y_3\}$ in $[n]$, a vertex $x\in [n]\backslash Y$, and a set $E$ of edges between these four vertices such that $Y\cup \{x\}$ spans a linear forest. Let $b_1$ denote the number of such choices. The choices in the next steps will be made such that $Y\cup \{x\}$ is a $(k,x,i,\psi)$-configuration in $G$. \item Next we choose the graph $G'$ on vertex set $[n]\backslash Y$ such that $G[[n]\backslash Y]=G'$. Let $b_2$ denote the number of possibilities for $G'$. \item Next we choose the set $E'$ of edges in $G$ between $Y$ and $Q_i\backslash (Y\cup \{x\})$ such that $E'$ is compatible with our previous choices. Let $b_3$ denote the number of possibilities for $E'$. \item Finally we choose the set $E''$ of edges in $G$ between $Y$ and $[n]\backslash (Q_i\cup Y \cup \{x\})$ such that $E''$ is compatible with our previous choices. Let $b_4$ denote the number of possibilities for $E''$. \end{itemize} Hence, \begin{equation}\label{eqn: F1i count} |F^1_{Q,(i)}|\leq 2\max\limits_{\psi\in \{\beta^2, \beta^{1/2}\}} \left\{b_1\cdot b_2\cdot b_3\cdot b_4\right\}. \end{equation} We then estimate the number of graphs in $F^1_{Q,(ii)}$. Any graph $G\in F^1_{Q,(ii)}$ can be constructed as follows. \begin{itemize} \item We first choose a single vertex $x$ from $[n]$ and distinct indices $i,j\in \{0,1,\dots, k-2\}$. Let $c_1$ denote the number of such choices. The choices in the next steps will be made such that $x$ is both $i$-light and $j$-light in $G$. \item Next we choose the graph $G'$ on vertex set $[n]\backslash \{x\}$ such that $G[[n]\backslash \{x\}]=G'$. Let $c_2$ denote the number of possibilities for $G'$. \item Next we choose the set $E$ of edges in $G$ between $\{x\}$ and $(Q_i\cup Q_j)\backslash \{x\}$ such that $E$ is compatible with our previous choices. Let $c_3$ denote the number of possibilities for $E$. \item Finally we choose the set $E'$ of edges in $G$ between $\{x\}$ and $[n]\backslash (Q_i\cup Q_j\cup \{x\})$. Let $c_4$ denote the number of possibilities for $E'$. \end{itemize} Hence, \begin{equation}\label{eqn: F1ii count} |F^1_{Q,(ii)}|\leq c_1\cdot c_2\cdot c_3\cdot c_4. \end{equation} The following series of claims will give upper bounds for the quantities $b_1,\dots, b_4, c_1, \dots, c_4$. Claims~1 and~5 are trivial, while the proof of Claim~6 is almost identical to that of Claim~2; we give proofs of Claims~2,3,4,7 and~8. \vspace{0.3cm} \noindent {\bf Claim 1:} $b_1\leq 2^6kn^4$. \vspace{0.3cm} \noindent {\bf Claim 2:} $b_2\leq C 2^{\mu^{1/2}n} f_k(n_k) 2^{t_{k-1}(n-3)}$. \noindent Indeed, note that for every graph $\tilde{G}\in F^1_{Q,(i)}$, Corollary~\ref{new partition} together with (\ref{size mathcal{P}}) implies that every optimal ordered $(k-1)$-partition of $\tilde{G}[[n]\backslash Y]$ is contained in some set $\mathcal{P}$ of size at most $2^{\mu n}$. Since $G[[n]\backslash Y]$ is clearly induced-$C_{2k}$-free, this together with (\ref{induct}) implies that \begin{align*} b_2 &\leq \sum_{Q'\in \mathcal{P}} |F_{Q'}(n-3,k)| \leq 6C 2^{\mu n} 2^{6(\log n)^2} f_k(\lceil (n-3)/(k-1) \rceil ) 2^{t_{k-1}(n-3)}\\ &\leq C 2^{\mu^{1/2}n} f_k(n_k) 2^{t_{k-1}(n-3)}, \end{align*} as required. \vspace{0.3cm} \noindent {\bf Claim 3:} $b_3\leq 2^{4\psi^{3/2} n}$. \noindent Indeed, for every graph $\tilde{G}\in F^1_{Q,(i)}$ for which $\{x,y_1,y_2,y_3\}$ is a $(k,x,i,\psi)$-configuration we have that $\min\{d^i_{\tilde{G},Q}(y_j), \overline{d}^i_{\tilde{G},Q}(y_j) \} \leq \psi^2 n$ for all $j\in [3]$. So $b_3\leq \prod_{j=1}^{3} h(j)$ where $h(j)$ denotes the number of possibilities for a set of edges between $\{y_j\}$ and $Q_i\backslash (Y\cup \{x\})$ such that either $d^i_{G,Q}(y_j)\leq \psi^2 n$ or $\overline{d}^i_{G,Q}(y_j)\leq \psi^2 n$. Note that by (\ref{entropy bound}), $h(j)\leq 2 {\binom{n}{\leq \psi^2 n}} \leq 2^{\xi(\psi^2) n+1}$. Hence, $$b_3\leq \prod\limits_{j=1}^{3} h(j)\leq ( 2^{\xi(\psi^2)n+1})^3 \stackrel{(\ref{entropy bound 1.5})}{\leq} 2^{4\psi^{3/2} n},$$ as required. \vspace{0.3cm} \noindent {\bf Claim 4:} $b_4\leq 2^{3(k-2)n/(k-1)}2^{\mu^{1/2} n} 2^{-\psi n/11^k}$. \noindent Indeed, first define $L$ to be the graph on vertex set $Y\cup \{x\}$ that satisfies $E(L)=E$. We say an induced subgraph $H$ of $G'-x$ is an $L$-\emph{compatible skeleton} if it satisfies the following. \begin{itemize} \item $|V(H)|=2k-4$. \item $G'[V(H)\cup \{x\}]$ is a linear forest. \item In $G'$, $x$ has $2-d_L(x)$ neighbours in $V(H)$. \item $L$ and $H$ have the same number of components. \end{itemize} Given an $L$-compatible skeleton $H$, note that Proposition~\ref{incomplete2}, applied with $L, G'[V(H)\cup \{x\}]$ playing the roles of $L_1, L_2$ respectively, implies that there exists a set $E_{L,H}$ of possible edges between $Y$ and $V(H)$ such that $(Y\cup \{x\}\cup V(H), E\cup E(H)\cup E_{L,H})$ is isomorphic to $C_{2k}$. We will show that there exist a large number of disjoint $L$-compatible skeletons in $G'-x$. Since there is a limited number of ways to choose edges between $Y$ and each of these $L$-compatible skeletons so as not to create an induced copy of $C_{2k}$, this will imply the claim. For every index $j\ne i$, let $N^1_j(x), N^2_j(x)\subseteq N_{Q_j}(x)$ be disjoint with $|N^1_j(x)|,|N^2_j(x)|\geq \lfloor \frac{1}{2}|N_{Q_j}(x)| \rfloor$. Similarly, let $\overline{N}^1_j(x), \overline{N}^2_j(x)\subseteq \overline{N}_{Q_j}(x)$ be disjoint with $|\overline{N}^1_j(x)|,|\overline{N}^2_j(x)|\geq \lfloor \frac{1}{2}|\overline{N}_{Q_j}(x)| \rfloor$. Note that we may assume that there exists an index $i'\in \{0,1,\dots, k-2\}\backslash \{i\}$ such that in $G'$, $|\overline{N}_{Q_j}(x)|\geq 12 \cdot 6^k \psi n$ for all $j\in \{0,1,\dots, k-2\}\backslash \{i\}$ and $|N_{Q_j}(x)| \geq 12 \cdot 6^k \psi n$ for all $j\in \{0,1,\dots, k-2\}\backslash \{i,i'\}$, since otherwise $\{x,y_1,y_2,y_3\}$ cannot be a $(k,x,i,\psi)$-configuration. Define $\ell_1,\dots,\ell_{k-2}$ such that $\{\ell_1,\dots, \ell_{k-2}\}=\{0,1,\dots, k-2\}\backslash \{i\}$ and $\ell_{k-2} = i'$. Thus the following hold. \begin{enumerate}[(a)] \item $|N^1_{\ell_{j}}(x)|,|N^2_{\ell_{j}}(x)|, |\overline{N}^1_{\ell_{j}}(x)|, |\overline{N}^2_{\ell_{j}}(x)|\geq 6 \cdot 6^k \psi n$ for all $j\in \{1,\dots, k-3\}$. \item $|\overline{N}^1_{\ell_{k-2}}(x)|, |\overline{N}^2_{\ell_{k-2}}(x)|\geq 6 \cdot 6^k \psi n$. \end{enumerate} We now show that $G'-x$ contains at least $5\cdot 6^k\psi n$ disjoint $L$-compatible skeletons. Define $t$ to be the number of components of $L$, and define $s:=d_L(x)$. Then $1\leq t\leq 4$ and $0\leq s\leq 2$. Note that $t+s\geq 2$, since a $4$-vertex linear forest with one component contains no isolated vertices. We consider two cases. In each case we will describe the length and type of $t$ path components, $P^1,\dots, P^t$, each with an even number of vertices. Proposition~\ref{building} (applied repeatedly) together with (a),(b) will then imply that $G'-x$ contains at least $5\cdot 6^k\psi n$ disjoint $L$-compatible skeletons, each consisting exactly of $t$ components isomorphic to $P^1,\dots, P^t$. (We can apply Proposition~\ref{building} here since in each case $P^1 \cup \dots \cup P^t$ will contain a perfect matching.) \begin{enumerate}[{\bf {Case} 1:}] \item \emph{$s=2$.} \begin{itemize} \item For $1\leq r\leq t-1$, $P^r$ is a $K_2$ of type $\overline{N}^1_{{\ell_r}}(x), \overline{N}^2_{{\ell_r}}(x)$. \item $P^{t}$ is a $P_{2k-2t-2}$ of type $ \overline{N}^1_{{\ell_t}}(x), \overline{N}^2_{{\ell_t}}(x), \overline{N}^1_{{\ell_{t+1}}}(x), \overline{N}^2_{{\ell_{t+1}}}(x),\dots, \overline{N}^1_{{\ell_{k-2}}}(x), \overline{N}^2_{{\ell_{k-2}}}(x)$. \end{itemize} \item \emph{Either $s=1$ or $s=0, t>1$.} \begin{itemize} \item For $1\leq r\leq 1-s$, $P^r$ is a $K_2$ of type $N^1_{{\ell_r}}(x), \overline{N}^1_{{\ell_r}}(x)$. \item $P^{2-s}$ is a $P_{2k-2t-2}$ of type $ N^1_{{\ell_{2-s}}}(x), \overline{N}^2_{{\ell_{2-s}}}(x), \overline{N}^1_{{\ell_{3-s}}}(x), \overline{N}^2_{{\ell_{3-s}}}(x),\dots, \overline{N}^1_{{\ell_{k-t-s}}}(x),\\ \overline{N}^2_{{\ell_{k-t-s}}}(x)$. \item For $k-t-s+1\leq r\leq k-2$, $P^r$ is a $K_2$ of type $\overline{N}^1_{{\ell_r}}(x), \overline{N}^2_{{\ell_r}}(x)$. \end{itemize} \end{enumerate} Since $t+s\geq 2$, this covers all cases. Now fix a set $SK$ of $5\cdot 6^k\psi n$ disjoint $L$-compatible skeletons in $G'-x$, and let $H\in SK$. Let $h_H$ denote the number of possibilities for a set $E^*$ of edges between $Y$ and $V(H)$. Note that such a set $E^*$ cannot equal $E_{L,H}$, since $G$ needs to be induced-$C_{2k}$-free. Thus $h_H\leq 2^{|Y||V(H)|}-1=2^{6(k-2)}-1$. Note that by $({\rm F}2)_{\mu}$ the number of vertices outside $Q_i$ that are not contained in some graph $H\in SK$ is at most $(k-2)n/(k-1) + \mu n - 10(k-2)6^k \psi n$. Hence, \begin{align*} b_4&\leq 2^{3(k-2)n/(k-1) - 30(k-2)6^k\psi n + 3\mu n}\prod\limits_{H\in SK} h_H \\ &\leq 2^{3(k-2)n/(k-1) - 30(k-2)6^k\psi n + 3\mu n} \left( 2^{6(k-2)}\left( 1-2^{-6(k-2)} \right) \right)^{5\cdot 6^k\psi n}\\ &\leq 2^{3(k-2)n/(k-1)}2^{3\mu n} e^{-5\cdot 6^k\psi n/(2^{6(k-2)})}\leq 2^{3(k-2)n/(k-1)}2^{\mu^{1/2} n} 2^{-\psi n/11^k}, \end{align*} as required. \vspace{0.3cm} \noindent {\bf Claim 5:} $c_1\leq k^2 n$. \vspace{0.3cm} \noindent {\bf Claim 6:} $c_2\leq C2^{\mu^{1/2}n} f(n_k) 2^{t_{k-1}(n-1)}$. \COMMENT{Indeed, note that for every graph $\tilde{G}\in F^1_{Q,(ii)}$, Corollary~\ref{new partition} together with (\ref{size mathcal{P}}) implies that every optimal ordered $(k-1)$-partition of $\tilde{G}[[n]\backslash \{x\}]$ is contained in some set $\mathcal{P}$ of size at most $2^{\mu n}$. Since $G[[n]\backslash \{x\}]$ is clearly induced-$C_{2k}$-free, this together with (\ref{induct}) implies that \begin{align*} c_2 &\leq \sum_{Q'\in \mathcal{P}} |F_{Q'}(n-1,k)| \leq 6C2^{\mu n} 2^{6(\log n)^2} f(\lceil (n-1)/(k-1) \rceil ) 2^{t_{k-1}(n-1)}\\ &\leq C2^{\mu^{1/2}n} f(n_k) 2^{t_{k-1}(n-1)},\end{align*} as required.} \vspace{0.3cm} \noindent {\bf Claim 7:} $c_3\leq 2^{7\xi(\alpha)n}$. \noindent Indeed, for every graph $\tilde{G}\in F^1_{Q,(ii)}$ for which $x$ is both $i$-light and $j$-light, we have that, for every $\ell\in \{i,j\}$, either $\min \{|N_{Q_{\ell}}(x)|, |\overline{N}_{Q_{\ell}}(x)|\}\leq \alpha n$ or else there exists a vertex $z\neq x$ such that $|N^*_{\ell}(x,z)| + |N^*_{\ell}(z,x)| \leq \alpha n$. For $\ell\in \{i,j\}$, let $h(\ell,1)$ denote the number possibilities for a set of edges in $G$ between $\{x\}$ and $Q_{\ell}\backslash \{x\}$ such that $\min \{|N_{Q_{\ell}}(x)|, |\overline{N}_{Q_{\ell}}(x)|\}\leq \alpha n$. Then $h(\ell,1)\leq 2\binom{n}{\leq \alpha n}\leq 2^{\xi(\alpha)n+1}$. For $\ell\in \{i,j\}$, let $h(\ell,2)$ denote the number possibilities for a set of edges between $\{x\}$ and $Q_{\ell}\backslash \{x\}$ such that there exists a vertex $z\neq x$ such that $|N^*_{\ell}(x,z)| + |N^*_{\ell}(z,x)| \leq \alpha n$. Then $h(\ell,2) \leq n{\binom{|N_{Q_{\ell}}(z)|}{\leq \alpha n}}{\binom{|\overline{N}_{Q_{\ell}}(z)|}{\leq \alpha n}}\leq 2^{3\xi(\alpha)n}$. Hence $$c_3\leq (h(i,1)+h(i,2))(h(j,1)+h(j,2))\leq (2^{\xi(\alpha)n+1}+ 2^{3\xi(\alpha)n})^2 \leq 2^{7\xi(\alpha)n},$$ as required. \vspace{0.3cm} \noindent {\bf Claim 8:} $c_4\leq 2^{ (k-3)n/(k-1) } 2^{2\mu n}$. \noindent Indeed, since the number of possible edges between $\{x\}$ and $[n]\backslash (Q_i\cup Q_j\cup \{x\})$ is at most $(k-3)n/(k-1) + 2\mu n$, we have that $c_4 \leq 2^{ (k-3)n/(k-1) + 2\mu n}$, as required. \vspace{0.3cm} \noindent Now (\ref{eqn: F1i count}) together with Claims~1--4 and Proposition~\ref{omitted proposition}(ii) implies that \begin{align}\label{eqn: F1i count two} &|F^1_{Q,(i)}|\\ \leq \hspace{0.2cm} &2\max\limits_{\psi\in \{\beta^2, \beta^{1/2}\}} \left\{2^6kn^4\cdot C 2^{\mu^{1/2}n} f_k\left( n_k \right) 2^{t_{k-1}(n-3)}\cdot 2^{4\psi^{3/2} n}\cdot 2^{\frac{3(k-2)n}{k-1}}2^{\mu^{1/2} n} 2^{-\frac{\psi n}{11^k}} \right\} \nonumber\\ \leq \hspace{0.2cm} &\max\limits_{\psi\in \{\beta^2, \beta^{1/2}\}} \left\{ Cf_k(n_k)2^{t_{k-1}(n-3)+\frac{3(k-2)n}{k-1}}2^{-\frac{\psi n}{12^k}} \right\} \leq Cf_k(n_k)2^{t_{k-1}(n)}2^{-\frac{\beta^2 n}{13^k}}.\nonumber \end{align} Similarly, (\ref{eqn: F1ii count}) together with Claims~5--8 and Proposition~\ref{omitted proposition}(ii) implies that \begin{align}\label{eqn: F1ii count two} |F^1_{Q,(ii)}|&\leq k^2 n\cdot C2^{\mu^{1/2}n} f_k\left( n_k \right) 2^{t_{k-1}(n-1)}\cdot 2^{7\xi(\alpha)n}\cdot 2^{ \frac{(k-3)n}{k-1} } 2^{2\mu n}\\ &\leq C2^{\mu^{1/3}n}f_k(n_k)2^{t_{k-1}(n-1) + \frac{(k-2)n}{k-1}}2^{-\frac{n}{k-1}}2^{7\xi(\alpha)n}\leq Cf_k(n_k)2^{t_{k-1}(n)}2^{-\frac{n}{k}}.\nonumber \end{align} Now (\ref{eqn: F1 count}) together with (\ref{eqn: F1i count two}) and (\ref{eqn: F1ii count two}) implies that $$|F^1_Q|\leq Cf_k(n_k)2^{t_{k-1}(n)}\left(2^{-\frac{\beta^2 n}{13^k}}+2^{-\frac{n}{k}}\right)\leq Cf_k(n_k)2^{t_{k-1}(n)}2^{-\frac{\beta^2 n}{14^k}},$$ as required. \end{proof} \section{Estimation of $|F^2_Q|$}\label{sec: estimate 2} Given $G\in F^2_Q\cup F^3_Q$ and $i\in \{0,1,\dots, k-2\}$, let $A^i_G := \{x \in Q_i: \overline{d}^i_{G,Q}(x),d^i_{G,Q}(x) \geq \beta n\}$. The key result of this section (Lemma~\ref{size of A^i_G}) states that $A^i_G$ has bounded size. To prepare for this, we will classify the pairs of vertices in $A^i_G$ according to their (non-)neighbourhood intersection pattern. The fact that $G\notin F^1_Q$ allows us to observe some restrictions on these patterns (see Propositions~\ref{prop: not j j' identical} and~\ref{asym-reg}). In the proof of Lemma~\ref{size of A^i_G} we use a Ramsey argument to restrict our view to one abundant type of pattern. This quickly leads to a contradiction if $|A^i_G|$ is large. Using the fact that $G\notin F^1_Q$ we show that the remainder of each class (i.e. $G[Q_i\backslash A^i_G]$) induces a very simple structure (Proposition~\ref{Q_i setminus A^i_G}). We translate this structural information into a sufficiently strong bound on the number of graphs in $F^2_G$, in Lemma~\ref{F^2_Q}. Let $\mathcal{L}$ denote the collection of all $4$-vertex linear forests. The following proposition is an analogue of Proposition~\ref{prop: k-good tetradehron free}(i) that can be applied to graphs rather than $2$-coloured multigraphs. It follows immediately from Proposition~\ref{prop: k-good tetradehron free}(i). \begin{proposition}\label{prop: L free} Let $G$ be a graph such that for every $H\in \mathcal{L}$, $G$ is induced $H$-free. Then $\overline{G}$ is a disjoint union of stars and triangles. \end{proposition} \begin{proposition}\label{Q_i setminus A^i_G} Let $G\in F^2_Q\cup F^3_Q$ and $i\in \{0,1,\dots, k-2\}$. Then $\overline{G}[Q_i\setminus A^i_G]$ is a disjoint union of stars and triangles. \end{proposition} \begin{proof} Suppose for a contradiction that $\overline{G}[Q_i\setminus A^i_G]$ is not a disjoint union of stars and triangles. Then Proposition~\ref{prop: L free} implies that $\overline{G}[Q_i\setminus A^i_G]$ contains an induced copy of a graph in $\mathcal{L}$, with vertex set $\{x,y_1,y_2,y_3\}$ say. We will show that $\{x,y_1,y_2,y_3\}$ is a $(k,x,i,\beta^{1/2})$-configuration, which contradicts the fact that $G\notin F^1_Q$. Note that $G[\{x,y_1,y_2,y_3\}]$ is a linear forest, and so $\{x,y_1,y_2,y_3\}$ satisfies (C1). By the definition of $A^i_G$ we have that $\min \{d^i_{G,Q}(y_j), \overline{d}^i_{G,Q}(y_j)\}\leq \beta n$ for all $j\in [3]$, and so $\{x,y_1,y_2,y_3\}$ satisfies (C4). Since $G\notin F^1_Q$, $x$ is $j$-light for at most one index $j\in\{0,1,\dots, k-2\}$. Since $x \in Q_i\setminus A^i_G$, $x$ is $i$-light. Thus for every $j\in \{0,1,\dots, k-2\}$ with $i\ne j$ we have that $x$ is not $j$-light, and hence $d^j_{G,Q}(x), \overline{d}^j_{G,Q}(x) > \alpha n > 13\cdot 6^k \cdot \beta^{1/2} n$, and so $\{x,y_1,y_2,y_3\}$ satisfies (C2) and (C3). Therefore $\{x,y_1,y_2,y_3\}$ is a $(k,x,i,\beta^{1/2})$-configuration, as required. \end{proof} The following definitions will be useful in order to show that $|A^i_G|$ is small. Suppose $S$ is a star or triangle. If $S$ is a star on at least three vertices, we call the unique vertex in $S$ of degree greater than one the \emph{centre} of $S$. Otherwise we call the vertex of $S$ with the smallest label the centre of $S$. Let $G\in F^2_Q\cup F^3_Q$ and $i,j\in \{0,1,\dots, k-2\}$ and let $x,y\in A^i_G$. \begin{itemize} \item We say $x,y$ are \emph{$j$-irregular} if $|\overline{N}_{j}(\{x,y\})| \leq \gamma n$. \item We say $x,y$ are \emph{$j$-asymmetric} if $|N^*_j(x,y)| +|N^*_j(y,x)| > 3\gamma n$ and either $|N^*_j(x,y)| \leq \gamma n$ or $|N^*_j(y,x)| \leq \gamma n$. \item We say $x,y$ are \emph{$j$-identical} if $|N^*_j(x,y)| +|N^*_j(y,x)| \leq 3\gamma n$. \end{itemize} Roughly speaking, if one of the above holds then the neighbourhoods of $x,y$ do not behave in a `random' like way (thus constraining the number of possibilities for choosing the neighbourhoods). The following statement follows immediately from the above definitions and the fact that $\gamma \ll \alpha$. \begin{equation}\label{identical-light} \text{If }x, y \text{ are } j \text{-identical then }x,y \text{ are both }j\text{-light.} \end{equation} \begin{proposition}\label{prop: not j j' identical} Let $G\in F^2_Q\cup F^3_Q$ and $i\in \{0,1,\dots, k-2\}$ and let $x,y \in A^i_G$. Then $x,y$ are $j$-identical for at most one index $j\in \{0,1,\dots, k-2\}$. \end{proposition} \begin{proof} Suppose $x,y$ are $j$-identical for some $j\in \{0,1,\dots, k-2\}$ and suppose $j'\in \{0,1,\dots, k-2\}$ with $j'\ne j$. It suffices to show that $x,y$ are not $j'$-identical. Note that $x$ is $j$-light by (\ref{identical-light}). Since $G\notin F^1_Q$, $x$ is $j''$-light for at most one index $j''\in \{0,1,\dots, k-2\}$. Thus $x$ is not $j'$-light, and hence by (\ref{identical-light}) $x,y$ are not $j'$-identical, as required. \end{proof} \begin{proposition} \label{asym-reg} Let $G\in F^2_Q\cup F^3_Q$ and $i\in \{0,1,\dots, k-2\}$ and let $x,y \in A^i_G$. Then there exists an index $j\in \{0,1,\dots, k-2\}$ such that $x,y$ are $j$-irregular or $j$-asymmetric (or both). \end{proposition} \begin{proof} Suppose for a contradiction that for every index $\ell\in \{0,1,\dots, k-2\}$, $x,y$ are neither $\ell$-irregular nor $\ell$-asymmetric. Since, by Proposition~\ref{prop: not j j' identical}, $x,y$ are $j$-identical for at most one index $j$, and $k\geq 6$, we may assume without loss of generality that $x,y$ are not $\ell$-identical for $\ell\in \{1,2,3\}$. We consider the following two cases. \vspace{0.3cm} \noindent {\bf Case 1:} \emph{$x,y$ are adjacent.} \noindent In this case we define sets $A_{\ell}, B_{\ell}$ for $\ell\in \{0,1,\dots, k-2\}$ as follows. We will use these sets to extend $x,y$ into an induced copy of $C_{2k}$. \begin{itemize} \item Let $A_1:= N^*_1(x,y)$ and $B_1 := \overline{N}_1(\{x,y\})$. \item Let $A_2:= N^*_2(y,x)$ and $B_2 := \overline{N}_2(\{x,y\})$. \item For every $\ell\in \{0,1,\dots, k-2\}\backslash \{1,2\}$, let $A_{\ell}, B_{\ell}\subseteq \overline{N}_{\ell}(\{x,y\})$ be disjoint and satisfy $|A_{\ell}|,|B_{\ell}|\geq \lfloor|\overline{N}_{\ell}(\{x,y\})|/2\rfloor$. \end{itemize} Since for every $\ell\in \{0,1,\dots, k-2\}$ $x,y$ are neither ${\ell}$-irregular nor ${\ell}$-asymmetric, and for every $\ell\in \{1,2\}$ $x,y$ are not $\ell$-identical, we have that $|A_{\ell}|,|B_{\ell}|\geq \gamma n/3$ for every $\ell\in \{0,1,\dots, k-2\}$. This together with Proposition~\ref{building} and the fact that $\mu \ll \gamma$ implies that there exists in $G$ an induced copy of $P_{2k-2}$ of type $A_1,B_1,A_0,B_0,A_3,B_3,\dots, A_{k-2},B_{k-2}, B_2,A_2$. By the definition of the sets $A_{\ell}, B_{\ell}$, the vertices of this $P_{2k-2}$ together with $x,y$ induce on $G$ a copy of $C_{2k}$. This contradicts the fact that $G\in F_Q(n,k)$. \vspace{0.3cm} \noindent {\bf Case 2:} \emph{$x,y$ are not adjacent.} \noindent In this case we define sets $A_{\ell}, B_{\ell}$ for $\ell\in \{0,1,\dots, k-2\}$ as follows. Similarly to the previous case, we will find an induced $C_{2k}$ which contains $x,y$ together with exactly one vertex from each of these sets. \begin{itemize} \item Let $A_1:= N^*_1(x,y)$ and $B_1 := N^*_1(y,x)$. \item Let $A_2:= N^*_2(x,y)$ and $B_2 := \overline{N}_2(\{x,y\})$. \item Let $A_3:= N^*_3(y,x)$ and $B_3 := \overline{N}_3(\{x,y\})$. \item For every $\ell\in \{0,1,\dots, k-2\}\backslash \{1,2,3\}$, let $A_{\ell}, B_{\ell}\subseteq \overline{N}_{\ell}(\{x,y\})$ be disjoint and satisfy $|A_{\ell}|,|B_{\ell}|\geq \lfloor|\overline{N}_{\ell}(\{x,y\})|/2\rfloor$. \end{itemize} Since for every $\ell\in \{0,1,\dots, k-2\}$ $x,y$ are neither ${\ell}$-irregular nor ${\ell}$-asymmetric, and for every $\ell\in \{1,2,3\}$ $x,y$ are not $\ell$-identical, we have that $|A_{\ell}|,|B_{\ell}|\geq \gamma n/3$ for every $\ell\in \{0,1,\dots, k-2\}$. As before, this together with Proposition~\ref{building} implies that there exists in $G$ an induced copy of the graph $H$ that consists of the following two components: \begin{itemize} \item One $P_{2k-4}$ of type $A_2,B_2,A_0,B_0,A_4,B_4,\dots, A_{k-2},B_{k-2}, B_3,A_3$. \item One $K_2$ of type $A_1,B_1$. \end{itemize} By the definition of the sets $A_{\ell}, B_{\ell}$, the vertices of $H$ together with $x,y$ induce on $G$ a copy of $C_{2k}$. This contradicts the fact that $G\in F_Q(n,k)$. \vspace{0.3cm} \noindent This covers all cases, and hence completes the proof. \end{proof} Recall from Section~\ref{sec: main proof} that $M:= R_{2k-2}(\lceil \frac{1}{\gamma}\rceil) +1$. \begin{lemma}\label{size of A^i_G} Let $G\in F^2_Q\cup F^3_Q$ and $i\in \{0,1,\dots, k-2\}$. Then $|A^i_G| < M$. \end{lemma} \begin{proof} Suppose for a contradiction that $|A^i_G| \geq M$. Consider an auxiliary complete graph $H_i$ with $V(H_i)= A^i_G$. We define a $(2k-2)$-edge-colouring $\mathcal{C}$ of $H_i$ with colours $\{a_0,b_0,a_1,b_1,\dots, a_{k-2},b_{k-2}\}$ as follows. \begin{itemize} \item For every $j\in \{0,1,\dots, k-2\}$, an edge $xy\in E(H)$ is coloured $a_j$ if $x,y$ are $j$-irregular and for every $j'\in \{0,1,\dots, k-2\}$ with $j'<j$, $x,y$ are not $j'$-irregular. \item An edge $xy\in E(H)$ that was not coloured in the previous step is coloured $b_j$ if $x,y$ are $j$-asymmetric, and for every $j'\in \{0,1,\dots, k-2\}$ with $j'<j$, $x,y$ are not $j'$-asymmetric. \end{itemize} Note that by Proposition~\ref{asym-reg}, every edge is coloured by a unique colour in $\mathcal{C}$. Now since $M > R_{2k-2}(\lceil 1/\gamma \rceil )$, $H_i$ contains a monochromatic clique of size at least $1/\gamma$. Let $X=\{x_1,x_2,\dots, x_{\lceil 1/\gamma \rceil}\}$ be the vertex set of such a monochromatic clique. We consider the following two cases. \vspace{0.3cm} \noindent {\bf Case 1:} \emph{$X$ has colour $a_j$ for some $j\in \{0,1,\dots, k-2\}$.} \noindent In this case every pair of vertices in $X$ is $j$-irregular, by definition of $\mathcal{C}$. Let $X':=\{x_1,x_2,\dots, x_{\lceil \beta/2\gamma \rceil} \}$ and suppose $z,z'\in X'$. By the definition of $j$-irregularity, $|\overline{N}_{j}(z)\cap\overline{N}_{j}(z')|\leq \gamma n$. Note also that $|\overline{N}_{j}(z)| \geq \beta n$ by Proposition~\ref{optimality} and the fact that $z\in A^i_G$. So by the inclusion-exclusion principle, \begin{align*} 2n/(k-1)&\geq n/(k-1) + \mu n \geq |Q_j| \geq \sum_{z\in X'} |\overline{N}_{j}(z)| - \sum_{\stackrel{z,z'\in X'}{z \ne z'}} |\overline{N}_{j}(z)\cap\overline{N}_{j}(z')|\\ &\geq \beta \lceil \beta/2\gamma \rceil n - \lceil \beta^2/(4\gamma^2)\rceil \gamma n \geq \beta^2n/5\gamma > 2n/(k-1), \end{align*} where the last inequality follows from the fact that $\gamma \ll \beta$. This is a contradiction. \vspace{0.3cm} \noindent {\bf Case 2:} \emph{$X$ has colour $b_j$ for some $j\in \{0,1,\dots, k-2\}$.} \noindent In this case every pair of vertices in $X$ is $j$-asymmetric, by definition of $\mathcal{C}$. Suppose $\ell,\ell'\in [\lceil 1/\gamma \rceil]$ are distinct. By the definition of $j$-asymmetry, exactly one of the following holds. \begin{enumerate}[(a)] \item $|N^*_j(x_\ell,x_{\ell'})| \leq \gamma n$ and $|N^*_j(x_{\ell'},x_{\ell})| > 2\gamma n$. \item $|N^*_j(x_{\ell'},x_{\ell})| \leq \gamma n$ and $|N^*_j(x_{\ell},x_{\ell'})| > 2\gamma n$. \end{enumerate} Consider the auxiliary tournament $T$ with $V(T)=X$ and $E(T)= \{ \overrightarrow{x_\ell x_{\ell'}} : \ell,\ell' \text{ satisfy (a)}\}$. By Redei's theorem every tournament contains a directed Hamilton path. So, by relabelling the indices if necessary, we may assume that $\overrightarrow{x_{\ell}x_{\ell+1}}\in E(T)$ for every $\ell\in [\lceil 1/\gamma \rceil-1]$. Thus for every $\ell\in [\lceil 1/\gamma \rceil-1]$, \begin{align*} |\overline{N}_{j}(x_{\ell+1})| = |\left(\overline{N}_{j}(x_{\ell}) \setminus N^*_j(x_{\ell+1},x_\ell) \right) \cup N^*_{j}(x_\ell,x_{\ell+1})|\leq |\overline{N}_{j}(x_{\ell})| -2\gamma n + \gamma n \leq |\overline{N}_{j}(x_{\ell})| -\gamma n. \end{align*} Hence, $$|\overline{N}_{j}(x_{\lceil 1/\gamma \rceil})| \leq |\overline{N}_{j}(x_{1})| - \left( \frac{1}{\gamma}-1\right) \cdot \gamma n \leq |Q_j| - (1-\gamma)n < 0,$$ which is a contradiction. \vspace{0.3cm} \noindent This covers all cases, and hence completes the proof. \end{proof} Suppose $G\in F^2_Q$ and $i\in \{0,1,\dots, k-2\}$. By Proposition~\ref{Q_i setminus A^i_G}, $\overline{G}[Q_i \setminus A^i_G]$ is a disjoint union of stars and triangles. Let $\mathcal{S}$ be the set of components of $\overline{G}[Q_i \setminus A^i_G]$ with the largest number of vertices. Let $S^{\diamond}$ be the component in $\mathcal{S}$ whose centre $c$ has the smallest label. Define $Y_i=Y_i(G,Q)$ to be the set of all isolated vertices in $\overline{G}[Q_i\setminus A^i_G]$ together with all vertices in $V(S^{\diamond})\backslash \{c\}$. \begin{lemma}\label{size of Y_i} Let $G\in F^2_Q$ and $i\in \{0,1,\dots, k-2\}$. Then $|Y_i| \geq 10 n/\log n$. \end{lemma} \begin{proof} Define $s:=\lceil 10 n/\log n \rceil$. Suppose for a contradiction that $|Y_i|< s$. Since $G \in F^2_Q$, there exists an index $i'\in \{0,1,\dots, k-2\}$ such that $|A^{i'}_G|>0$. Let $x\in A^{i'}_G$. The definition of $A^{i'}_G$ together with Proposition~\ref{optimality} implies that $|\overline{N}_{Q_{j}}(x)|\geq \beta n$ for every $j\in \{0,1,\dots, k-2\}$. This together with Lemma~\ref{size of A^i_G} implies that $|\overline{N}_{Q_i}(x)\setminus A^i_G| \geq \beta n - M > 2s$. Also, since $|Y_i|<s$, at most $s$ components in $\overline{G}[Q_i\setminus A^i_G]$ are isolated vertices and every component in $\overline{G}[Q_i\setminus A^i_G]$ has order at most $s$. Thus there are at least two non-trivial components $S,S'$ of $\overline{G}[Q_i\setminus A^i_G]$ that each contain a non-neighbour of $x$. Since $S$ is a non-trivial component of $\overline{G}[Q_i\setminus A^i_G]$ there exist vertices $y,y'\in S$ such that $xy,yy'\notin E(G[Q_i])$. Let $y''\in S'$ be such that $xy''\notin E(G[Q_i])$. Since $y''$ belongs to a different component of $\overline{G}[Q_i\setminus A^i_G]$ to $y$ and $y'$, it follows that $yy'',y'y''\in E(G[Q_i])$. Thus, \begin{equation}\label{linear forest} E(G[\{x,y,y',y''\}])\in\{ \{yy'',y'y''\}, \{xy',yy'',y'y''\} \}. \end{equation} \vspace{0.3cm} \noindent {\bf Claim:} \emph{$\{x,y,y',y''\}$ is a $(k,x,i,\beta^2)$-configuration.} \noindent Indeed, by (\ref{linear forest}), $G[\{x,y,y',y''\}]$ is a linear forest and so $\{x,y,y',y''\}$ satisfies (C1). As observed above, $\overline{d}^j_{G,Q}(x)\geq \beta n > 13\cdot 6^k \beta^2 n$ for every $j\in \{0,1,\dots, k-2\}$, and so $\{x,y,y',y''\}$ satisfies (C2). Since $G\notin F^1_Q$, there do not exist distinct $j,j'\in \{0,1,\dots, k-2\}$ such that $x$ is both $j$-light and $j'$-light. So there exists $j\in \{0,1,\dots, k-2\}$ such that for every $j'\in \{0,1,\dots, k-2\}$ with $j'\ne j$, $d^{j'}_{G,Q}(x)> \alpha n > 13\cdot 6^k \beta^2 n$, and so $\{x,y,y',y''\}$ satisfies (C3). Since $S,S'$ each contain at most $s$ vertices, $y,y',y''$ each have at most $s$ non-neighbours in $G[Q_i\setminus A^i_G]$. This together with Lemma~\ref{size of A^i_G} implies that $y,y',y''$ each have at most $s+M\leq \beta^4 n$ non-neighbours in $G[Q_i]$, and so $\{x,y,y',y''\}$ satisfies (C4). Hence $\{x,y,y',y''\}$ is a $(k,x,i,\beta^2)$-configuration, as required. \vspace{0.3cm} \noindent The above claim contradicts the fact that $G\notin F^1_Q$, and hence completes the proof. \end{proof} Lemma~\ref{size of Y_i} guarantees a large set of vertices in each class $Q_i$ (namely $Y_i$) with an extremely restricted (non-)neighbourhood. This is the key idea in our estimation of $|F^2_Q|$. \begin{lemma}\label{F^2_Q} $ |F^2_Q| \leq C 2^{-n}f_k(n_k) 2^{t_{k-1}(n)}$. \end{lemma} \begin{proof} Define $s:=\lceil 10 n/\log n \rceil$. Since by Lemma~\ref{size of Y_i} $|Y_i(G,Q)|\geq s$ for every graph $G\in F^2_Q$, any graph $G\in F^2_Q$ can be constructed as follows. \begin{itemize} \item First we choose sets $S_{\ell}\subseteq Q_{\ell}$ such that $|S_{\ell}|=s$, for every $\ell\in \{0,1,\dots, k-2\}$. Let $b_1$ denote the number of such choices. \item Next we choose the graph $G'$ on $[n]\backslash \bigcup_{\ell\in \{0,1,\dots, k-2\}} S_{\ell}$ such that $G[[n]\backslash \bigcup_{\ell\in \{0,1,\dots, k-2\}} S_{\ell}]=G'$. Let $b_2$ denote the number of possibilities for $G'$. \item Next we choose the set $E'$ of internal edges of $G$ that are incident to at least one vertex in $\bigcup_{\ell\in \{0,1,\dots, k-2\}} S_{\ell}$ in such a way that the resulting graph $G$ will satisfy $S_{\ell}\subseteq Y_{\ell}(G,Q)$ for every $\ell\in \{0,1,\dots, k-2\}$. Let $b_3$ denote the number of possibilities for $E'$. \item Finally we choose the set $E''$ of crossing edges of $G$ that are incident to at least one vertex in $\bigcup_{\ell\in \{0,1,\dots, k-2\}} S_{\ell}$. Let $b_4$ denote the number of possibilities for $E''$. \end{itemize} Hence, \begin{equation}\label{eqn: F2 count} |F^2_Q|\leq b_1\cdot b_2\cdot b_3\cdot b_4. \end{equation} The following series of claims will give upper bounds for the quantities $b_1,\dots, b_4$. The proof of Claim~2 is almost identical to that of Claim~2 in Lemma~\ref{F^1}; we give proofs of the others. \vspace{0.3cm} \noindent {\bf Claim 1:} $b_1\leq 2^n$. \noindent Indeed, $$b_1 \leq {\binom{n}{\left\lceil \frac{10n}{\log n}\right\rceil}}^{k-1} \leq \left( \left(\frac{e\log n}{10}\right)^{\frac{10n}{\log n}}\right)^{k-1} \leq 2^n,$$ as required. \vspace{0.3cm} \noindent {\bf Claim 2:} $b_2\leq C 2^{\mu^{1/2}n} f_k(\lceil n/(k-1)-s \rceil ) 2^{t_{k-1}(n-(k-1)s)}$. \COMMENT{Indeed, note that for every graph $\tilde{G}\in F^2_Q$, Corollary~\ref{new partition} together with (\ref{size mathcal{P}}) implies that every optimal ordered $(k-1)$-partition of $\tilde{G}[[n]\backslash \bigcup_{\ell\in \{0,1,\dots, k-2\}} S_{\ell}]$ is contained in some set $\mathcal{P}$ of size at most $2^{\mu n}$. Since $G[[n]\backslash \bigcup_{\ell\in \{0,1,\dots, k-2\}} S_{\ell}]$ is clearly induced-$C_{2k}$-free, this together with (\ref{induct}) implies that \begin{align*} b_2 &\leq \sum_{Q'\in \mathcal{P}} |F_{Q'}(n- (k-1)s,k)| \leq 6C 2^{\mu n}2^{6(\log n)^2} f_k(\lceil n/(k-1)-s \rceil ) 2^{t_{k-1}(n-(k-1)s)}\\ &\leq C 2^{\mu^{1/2}n} f_k(\lceil n/(k-1)-s \rceil ) 2^{t_{k-1}(n-(k-1)s)}, \end{align*} as required.} \vspace{0.3cm} \noindent {\bf Claim 3:} $b_3\leq 2^n$. \noindent Indeed, for every graph $G^*\in F^2_Q$ for which $S_{\ell}\subseteq Y_{\ell}(G^*,Q)$ for every $\ell\in \{0,1,\dots, k-2\}$, let $G^*_{B, \ell}:= \overline{G^*}[Q_{\ell}\backslash A^{\ell}_{G^*}]$. Then each $S_{\ell}$ consists of isolated vertices in $G^*_{B, \ell}$ as well as non-centre vertices of a single component $\tilde{C}$ of $G^*_{B, \ell}$. (Note that $\tilde{C}$ is a star or triangle in $G^*_{B, \ell}$, with some centre $u\in Q_{\ell}\setminus (A^{\ell}_{G^*}\cup S_{\ell})$.) By Lemma~\ref{size of A^i_G}, we also have that $|A^{\ell}_{G^*}|\leq M$. Hence $b_3\leq \prod_{\ell=0}^{k-2} \prod_{j=1}^3 h(\ell,j)$, where the quantities $h(\ell, j)$ are defined as follows. Let $h(\ell,1)$ denote the number of ways to choose a set $\tilde{A}^{\ell}\subseteq Q_{\ell}\backslash S_{\ell}$ of size at most $M$. (In what follows $\tilde{A}^{\ell}$ will play the role of $A^i_G$.) Then $h(\ell,1)\leq n^M$. Given such a set $\tilde{A}^{\ell}$, let $h(\ell,2)$ denote the number of ways to choose $\tilde{C}$. Then $h(\ell, 2)\leq n 2^{|S_{\ell}|}+n^3\leq n 2^{s+1}$. (Indeed, if $\tilde{C}$ is a star we have at most $n$ choices for the centre $u$, and for every vertex $v\in S_{\ell}$ we can choose whether $v$ is adjacent to $u$ or not; if $\tilde{C}$ is a triangle we have at most $n^3$ choices for its vertices.) Given a set $\tilde{A}^{\ell}$ as above, let $h(\ell,3)$ denote the number of possible sets of edges between $S_{\ell}$ and $\tilde{A}^{\ell}$. Then $h(\ell,3)\leq 2^{|S_{\ell}||\tilde{A}^{\ell}|}\leq 2^{sM}$. Hence $$b_3\leq \prod\limits_{\ell=0}^{k-2} \prod\limits_{j=1}^3 h(\ell,j)\leq (n^M\cdot n 2^{s+1}\cdot 2^{sM})^{k-1}\leq 2^n,$$ as required. \vspace{0.3cm} \noindent {\bf Claim 4:} $b_4\leq 2^{(k-2)sn-\binom{k-1}{2}s^2}$. \noindent Indeed, note that, for a fixed index $\ell\in \{0,1,\dots, k-2\}$, the number $h_{\ell}$ of possible crossing edges in $G$ that are incident to a vertex in $S_{\ell}$ is at most $s(n - |Q_\ell|)$. Also, the number of possible crossing edges in $G$ that are incident to two vertices in $\bigcup_{\ell\in \{0,1,\dots, k-2\}} S_{\ell}$ is exactly $\binom{k-1}{2}s^2$. Hence, $$b_4 \leq 2^{\sum_{\ell=0}^{k-2} h_{\ell}} 2^{-\binom{k-1}{2}s^2}\leq 2^{(k-2)sn-\binom{k-1}{2}s^2},$$ as required. \vspace{0.3cm} \noindent Note that $t_{k-1}(s(k-1)) = \binom{k-1}{2}s^2$ and that by Lemma \ref{f-estimate-2}, $f_k(n_k) \geq s^{s/2}f_k(\lceil n/(k-1)-s \rceil ) \geq 2^{4n} f_k(\lceil n/(k-1)-s \rceil )$. These observations together with (\ref{eqn: F2 count}), Claims~1--4 and Proposition~\ref{omitted proposition}(ii) imply that \begin{align*} |F^2_Q|&\leq 2^n \cdot C 2^{\mu^{1/2}n} f_k(\lceil n/(k-1)-s \rceil ) 2^{t_{k-1}(n-(k-1)s)} \cdot 2^n \cdot 2^{(k-2)sn-\binom{k-1}{2}s^2}\\ &\leq C 2^{3n}2^{-4n}f_k(n_k)2^{t_{k-1}(n-(k-1)s)+(k-2)sn-s(k-1)(k-2)-t_{k-1}(s(k-1))}\\ &\leq C 2^{-n} f_k(n_k) 2^{t_{k-1}(n)}, \end{align*} as required. \end{proof} \section{Estimation of $|F^3_Q|$}\label{sec: estimate 3} The information we have gained so far allows us to easily deduce that every $G\in F^3_Q$ is extremely close to being a $k$-template (see Proposition~\ref{beta}). One advantage of this is that it allows us to use more precise estimates when applying induction (see Corollary~\ref{new partition 2}). \begin{proposition}\label{beta} Let $G\in F^3_Q$ and $i\in \{0,1,\dots, k-2\}$. Then the following hold. \begin{enumerate}[{\rm (i)}] \item $\overline{G}[Q_i]$ is a disjoint union of stars and triangles. \item $G$ contains at most $n$ internal non-edges. \item Every vertex $x\in Q_i$ satisfies $\overline{d}^i_{G,Q}(x)< \beta n$. \end{enumerate} \end{proposition} \begin{proof} \begin{enumerate}[{\rm (i)}] \item Since $G\notin F^2_Q$, every vertex $x \in Q_i$ satisfies $\min \{ d^i_{G,Q}(x),\overline{d}^i_{G,Q}(x)\} < \beta n$. Thus $A^i_G=\emptyset$, and so by Proposition~\ref{Q_i setminus A^i_G}, $\overline{G}[Q_i]$ is a disjoint union of stars and triangles. \item This follows immediately from (i). \item Let $x\in Q_i$. Let us first show that $d^i_{G,Q}(x) \geq \beta n$. Suppose not. Then $\overline{d}^i_{G,Q}(x) = |Q_i|-d^i_{G,Q}(x)-1 > |Q_i| -\beta n -1 \geq n/(k-1)-\mu n - \beta n -1$. Thus for every $j\in \{0,1,\dots, k-2\}$ with $j\ne i$, Proposition~\ref{optimality} implies that \begin{align*} d^j_{G,Q}(x) &= |Q_j| - \overline{d}^j_{G,Q}(x) \leq |Q_j| - \overline{d}^i_{G,Q}(x) < \left(\frac{n}{k-1} + \mu n \right) - \left(\frac{n}{k-1}-\mu n - \beta n -1 \right)\\ &= \beta n + 2\mu n +1 < \alpha n, \end{align*} where the last inequality follows from the fact that $\mu, \beta \ll \alpha$. Thus $x$ is both $i$-light and $j$-light, which contradicts the fact that $G\notin F^1_Q$. Thus $d^i_{G,Q}(x) \geq \beta n$. This together with the fact (observed in the proof of (i), above) that $A^i_G=\emptyset$ implies that $\overline{d}^i_{G,Q}(x) < \beta n$, as required. \end{enumerate} \end{proof} Recall the definition of property $({\rm F}1)_{\nu}$ in Section~\ref{sec: set-up}. We define $T^*_Q(n,k) \subseteq F^3_Q$ to be the set of all (labelled) graphs in $F^3_Q$ that satisfy property $({\rm F}1)_{(40n\log n)^{1/2}/n}$ with respect to $Q$. Proposition~\ref{beta}(ii) together with Lemma~\ref{pre regular lemma}(i) applied with $(40n\log n)^{1/2}/n,n$ playing the roles of $\nu, m$ respectively implies that \begin{equation}\label{strong regular} |F^3_Q\setminus T^*_Q(n,k)| \leq 2^{t_{k-1}(n) - n\log n/5 }. \end{equation} So (\ref{strong regular}) allows us to restrict our attention to the class $T^*_Q(n,k)$. In particular, this allows us to apply property $({\rm F}1)_{\nu}$ to much smaller vertex sets than in the preceding sections. This in turn gives us a much better bound on the number of partitions that may arise after deleting a small number of vertices. More precisely, Lemma~\ref{pre new partition} applied with $(40n\log n)^{1/2}/n,n$ playing the roles of $\nu, m$ respectively implies the following result. Recall that $\mathcal{P}(Q,s)$ was defined before (\ref{size mathcal{P}}). \begin{corollary}\label{new partition 2} Let $S\subseteq [n]$ with $|S|\leq n/k^2$. Then for every $G\in T^*_Q(n,k)$, every optimal ordered $(k-1)$-partition of $G - S$ is an element of $\mathcal{P}(Q-S,40 k^4 \log n)$. \end{corollary} In order to estimate $|T_Q^*(n,k)|$ (and thus $|F^3_Q|$) we will further split $T_Q^*(n,k)$ into four classes $\mathcal{A}_1,\dots, \mathcal{A}_4$. To define these classes we require some further notation. We say that $G$ contains a {\em $(6,3)$-forest} with respect to $Q$ if there exist distinct indices $i,j\in \{0,1,\dots, k-2\}$ such that there exist six vertices in $Q_i\cup Q_j$ that induce on $G$ a linear forest with at most three components. A $(6,3)$-forest has potential extensions into an induced $C_{2k}$, so its existence in every $G\in \mathcal{A}_3$ (see below) constrains the possible edge sets for $G$ (and thus the number of choices for $G$). To obtain a significant constraint on the possible edge sets however, we first need to exclude the situations that arise in the classes $\mathcal{A}_1$ and $\mathcal{A}_2$, described below. These involve the structure of the stars of the complement graph inside the vertex classes, so to describe these classes of graphs recall that the centres of stars and triangles were defined before Proposition~\ref{prop: not j j' identical}. Given a graph $G\in F^3_Q$ and an index $i\in \{0,1,\dots, k-2\}$ we define the following sets. \begin{itemize} \item $C^i(G,Q)$ is the set of all centres of triangles and non-trivial stars in $\overline{G}[Q_i]$. \item $C^i_{high}(G,Q)$ is the set of all centres of stars in $\overline{G}[Q_i]$ of order at least $n^{1-\frac{1}{2 k^2}}/200k^2$. \item $B^i_{high}(G,Q)$ is the set of all vertices in $Q_i$ which have a non-neighbour in $C^i_{high}$. \item $C^i_{low}(G,Q)$ is the set of all centres of triangles and non-trivial stars in $\overline{G}[Q_i]$ of order less than $n^{1-\frac{1}{2 k^2}}/200k^2$. \item $B^i_{low}(G,Q)$ is the set of all vertices in $Q_i$ which have a non-neighbour in $C^i_{low}$. \item $C^i_{0}(G,Q)$ is the set of all isolated vertices in $\overline{G}[Q_i]$. \end{itemize} We may sometimes write $C^i$ for $C^i(G,Q)$ when the graph $G$ and ordered $(k-1)$-partition $Q$ we consider are clear from the context (and similarly for $C^i_{high}, B^i_{high}, C^i_{low}, B^i_{low}, C^i_0$). Note that Proposition~\ref{beta}(i) implies that $C^i_{high}, B^i_{high}, C^i_{low}, B^i_{low}, C^i_0$ form a partition of $Q_i$. Given a subset $B\subseteq B^i_{low}$, we denote by $C(B)$ the set of all vertices in $C^i_{low}$ that have a non-neighbour in $B$. We partition $T^*_Q(n,k)$ into the sets $\mathcal{A}_1, \dots, \mathcal{A}_4$ defined as follows. \begin{itemize} \item $\mathcal{A}_1$ is the set of all graphs $G \in T^*_Q(n,k)$ for which there exist distinct indices $i,j\in \{0,1,\dots, k-2\}$ such that $|B^{i}_{low}| \geq n/2k^2$ and there exist distinct vertices $y_1,y_2,y_3 \in Q_j$ that satisfy $|\overline{N}(\{y_1,y_2,y_3\}) \cap B^{i}_{low} | \leq n/200 k^2$. \item $\mathcal{A}_{2}$ is the set of all graphs $G\in T^*_Q(n,k) \setminus \mathcal{A}_1$ for which there exist distinct indices $i,j\in \{0,1,\dots, k-2\}$ such that $|B^{i}_{low}| \geq n/2k^2$ and there exist distinct vertices $y_1,y_2,y_3 \in Q_j$ with $y_1,y_2 \notin C^j(G,Q)$ that satisfy \begin{equation}\label{picture equation} C(\overline{N}(\{y_1,y_2,y_3\}) \cap B^{i}_{low}) \cap \overline{N}(\{y_1,y_2\}) = \emptyset. \end{equation} (See Figure $1$.) \begin{figure} \centering \begin{tikzpicture}[scale=0.6] \draw[line width=0.8pt] (0,0) ellipse (2 and 3); \draw[line width=0.8pt] (5,0) ellipse (2 and 3); \draw[line width=0.8pt] (1,0) ellipse (0.8 and 2); \draw[line width=0.8pt] (-0.6,1) ellipse (0.6 and 1); \filldraw[fill=black] (1,1.5) circle (2pt); \filldraw[fill=black] (1,0.5) circle (2pt); \filldraw[fill=black] (1,-0.5) circle (2pt); \filldraw[fill=black] (1,-1.5) circle (2pt); \filldraw[fill=black] (-0.6,1.5) circle (2pt); \filldraw[fill=black] (5,1.5) circle (2pt); \filldraw[fill=black] (5,0.5) circle (2pt); \filldraw[fill=black] (5,-0.5) circle (2pt); \filldraw[fill=black] (-0.6,-0.5) circle (2pt); \draw (1,-1.5) -- (-0.6,-0.5) -- (1,-0.5) -- (5,-0.5) -- (1,1.5) -- (5,0.5) -- (1,0.5) -- (-0.6,1.5) -- (1,1.5) -- (5,1.5); \draw (-0.6,1.5) to[bend left] (5,1.5); \node at (5.5,1.5) {$y_1$}; \node at (5.5,0.5) {$y_2$}; \node at (5.5,-0.5) {$y_3$}; \node at (-0.6,1.1) {$x$}; \node at (0,-3.5) {$Q_i$}; \node at (5,-3.5) {$Q_j$}; \node at (2.2,-2.4) {$B^i_{low}$}; \node at (1.6,3.6) {$C(\overline{N}(\{y_1,y_2,y_3\})\cap B^i_{low})$}; \draw[line width=0.2pt, ->, >= angle 60] (1.4, 3.2) -- (-0.4,2); \draw[line width=0.2pt, ->, >= angle 60] (1.6, -2.4) -- (1.2,-2); \node at (2.5,-4.2) {\textit{Figure 1: An illustration of $\overline{G}$ for $G\in \mathcal{A}_2$. Note that}}; \node at (2.5,-4.9) {\textit{{\rm (\ref{picture equation})} implies that at most one of $xy_1, xy_2$ is an edge in $\overline{G}$.}}; \end{tikzpicture} \end{figure} \item $\mathcal{A}_3$ is the set of all graphs $G \in T^*_Q(n,k) \setminus (\mathcal{A}_1 \cup \mathcal{A}_2)$ such that $G$ contains a $(6,3)$-forest with respect to $Q$. \item $\mathcal{A}_4 := T^*_Q(n,k) \setminus (\mathcal{A}_1 \cup \mathcal{A}_2 \cup \mathcal{A}_3)$ is the set of all remaining graphs. \end{itemize} We will estimate the sizes of $\mathcal{A}_1,\dots,\mathcal{A}_4$ separately. Lemma~\ref{lem: A_1} below gives a bound on $|\mathcal{A}_1|$. The idea of the proof of Lemma~\ref{lem: A_1} is that in this case the neighbourhoods of $y_1,y_2,y_3$ are `atypical', and hence a Chernoff estimate (see Claim~4) shows that graphs in $\mathcal{A}_1$ are rare. \begin{lemma}\label{lem: A_1} $|\mathcal{A}_1| \leq C 2^{-n/150k^2} f_k(n_k) 2^{t_{k-1}(n)}$. \end{lemma} \begin{proof} Any graph $G\in \mathcal{A}_1$ can be constructed as follows. \begin{itemize} \item First we choose distinct indices $i,j\in \{0,1,\dots, k-2\}$, distinct vertices $y_1,y_2,y_3\in Q_j$, and a set $E$ of edges between $y_1,y_2,y_3$. Let $b_1$ denote the number of such choices. The choices in the next steps will be made such that $G$ satisfies $|B^{i}_{low}| \geq n/2k^2$ and $|\overline{N}(\{y_1,y_2,y_3\}) \cap B^{i}_{low} | \leq n/200 k^2$. \item Next we choose the graph $G'$ on vertex set $[n]\backslash \{y_1,y_2,y_3\}$ such that $G[[n]\backslash \{y_1,y_2,y_3\}]=G'$. Let $b_2$ denote the number of possibilities for $G'$. \item Next we choose the set $E'$ of edges in $G$ between $\{y_1,y_2,y_3\}$ and $Q_j\backslash \{y_1,y_2,y_3\}$. Let $b_3$ denote the number of possibilities for $E'$. \item Finally we choose the set $E''$ of edges in $G$ between $\{y_1,y_2,y_3\}$ and $[n]\backslash Q_j$ such that $E''$ is compatible with our previous choices. Let $b_4$ denote the number of possibilities for $E''$. \end{itemize} Hence, \begin{equation}\label{eqn: A_1 count} |\mathcal{A}_1|\leq b_1\cdot b_2\cdot b_3\cdot b_4. \end{equation} The following series of claims will give upper bounds for the quantities $b_1,\dots, b_4$. Claim~1 is trivial; we give proofs of the others. \vspace{0.3cm} \noindent {\bf Claim 1:} $b_1\leq 2^3 k^2 n^3$. \vspace{0.3cm} \noindent {\bf Claim 2:} $b_2\leq C 2^{2(\log n)^3} f_k(n_k) 2^{t_{k-1}(n-3)}$. \noindent Indeed, note that for every graph $\tilde{G}\in \mathcal{A}_1$, Corollary~\ref{new partition 2} together with (\ref{size mathcal{P}}) implies that every optimal ordered $(k-1)$-partition of $\tilde{G}[[n]\backslash \{y_1,y_2,y_3\}]$ is contained in some set $\mathcal{P}$ of size at most $2^{(\log n)^3}$. Since $G[[n]\backslash \{y_1,y_2,y_3\}]$ is clearly induced-$C_{2k}$-free, this together with (\ref{induct}) implies that \begin{align*} b_2 &\leq \sum_{Q'\in \mathcal{P}} |F_{Q'}(n-3,k)| \leq 6C 2^{(\log n)^3}2^{6(\log n)^2} f_k(\lceil (n-3)/(k-1) \rceil ) 2^{t_{k-1}(n-3)}\\ &\leq C 2^{2(\log n)^3} f_k(n_k) 2^{t_{k-1}(n-3)}, \end{align*} as required. \vspace{0.3cm} \noindent {\bf Claim 3:} $b_3\leq 2^{3\xi(\beta) n}$. \noindent Indeed, for every graph $\tilde{G}\in \mathcal{A}_1$ and every $\ell \in [3]$, Proposition~\ref{beta}(iii) implies that $\overline{d}^j_{\tilde{G},Q}(y_{\ell})< \beta n$. Thus $$b_3\leq \binom{n}{\leq \beta n} ^3\leq 2^{3\xi(\beta) n},$$ as required. \vspace{0.3cm} \noindent {\bf Claim 4:} $b_4\leq 2^{3((k-2)n/(k-1)+\mu n)-n/128k^2}$. \noindent Consider the graph obtained by starting with the graph $([n], E(G')\cup E')$ and adding edges between $\{y_1,y_2,y_3\}$ and $[n]\backslash Q_j$ randomly, independently, with probability $1/2$. Note that the number of graphs that this process can generate is at most $2^{3((k-2)n/(k-1)+\mu n)}$, with each such graph equally likely to be generated. So an upper bound on $b_4$ is given by $$b_4\leq 2^{3((k-2)n/(k-1)+\mu n)}\mathbb{P}\left(|\overline{N}(\{y_1,y_2,y_3\})\cap B^i_{low}| \leq \frac{n}{200k^2} \right).$$ Since $G'$ was chosen such that $|B^i_{low}|\geq n/2k^2$, we have that $\mathbb{E}(|\overline{N}(\{y_1,y_2,y_3\})\cap B^i_{low}|)\geq n/16k^2$. So Lemma~\ref{Chernoff Bounds}(ii) implies that $$\mathbb{P}\left(|\overline{N}(\{y_1,y_2,y_3\})\cap B^i_{low}| \leq \frac{n}{200k^2} \right) \leq \exp \left( - \frac{n}{128k^2} \right)\leq 2^{- \frac{n}{128k^2}}.$$ Hence $b_4\leq 2^{3((k-2)n/(k-1)+\mu n)-n/128k^2}$, as required. \vspace{0.3cm} \noindent Now (\ref{eqn: A_1 count}) together with Claims~1--4 and Proposition~\ref{omitted proposition}(ii) implies that \begin{align*} |\mathcal{A}_1|&\leq 2^3 k^2 n^3 \cdot C 2^{2(\log n)^3} f_k(n_k) 2^{t_{k-1}(n-3)} \cdot 2^{3\xi(\beta) n} \cdot 2^{3((k-2)n/(k-1)+\mu n)-n/128k^2}\\ &\leq C 2^{-n/150k^2} f_k(n_k) 2^{t_{k-1}(n-3)+3(k-2)n/(k-1) - 3(k-2) - 3}\\ &\leq C 2^{-n/150k^2} f_k(n_k) 2^{t_{k-1}(n)}, \end{align*} as required. \end{proof} \begin{lemma}\label{lem: A_2} $|\mathcal{A}_2| \leq C 2^{-n^{1/2k^2}/3} f_k(n_k) 2^{t_{k-1}(n)}$. \end{lemma} \begin{proof} Note that for every $G\in \mathcal{A}_2$ and every $s\in \{0,1,\dots, k-2\}$ the definition of $C^s(G,Q)$ implies that $Q_s\backslash C^s(G,Q)\geq |Q_s|/2$. So any graph $G\in \mathcal{A}_2$ can be constructed as follows. We first choose $a\in \mathbb{N}$ such that $n/2k^2 \leq a\leq n$, and then perform the following steps. \begin{itemize} \item We choose distinct indices $i,j\in \{0,1,\dots, k-2\}$, a set $W=\{y_1,y_2\}\cup \{w_{\ell}^s: \ell\in [2], s\in \{0,1,\dots, k-2\}\backslash \{j\} \}$ of vertices satisfying $y_1,y_2\in Q_j$ and $w_1^s, w_2^s\in Q_s$ for every $s\in \{0,1,\dots, k-2\}\backslash \{j\}$, a vertex $y_3\in Q_j\backslash W$, and a set $E$ of edges between the vertices in $W\cup \{y_3\}$. Let $b_1$ denote the number of such choices. The choices in this step and the next steps will be made such that $y_1,y_2\notin C^j(G,Q)$ and $w_1^s, w_2^s\notin C^s(G,Q)$ for every $s\in \{0,1,\dots, k-2\}\backslash \{j\}$, and $|B^{i}_{low}|=a$ and $C(Y) \cap \overline{N}(\{y_1,y_2\}) = \emptyset$, where $Y:= \overline{N}(\{y_1,y_2,y_3\}) \cap B^{i}_{low}(G,Q)$. \item Next we choose the graph $G'$ on vertex set $[n]\backslash W$ such that $G[[n]\backslash W]=G'$. Let $b_2$ denote the number of possibilities for $G'$. \item Next we choose the set $E'$ of internal edges in $G$ with exactly one endpoint in $W$ such that $E'$ is compatible with our previous choices\COMMENT{Note that we are overcounting here, since we have already fixed the edges between $\{y_1, y_2\}$ and $y_3$, but this is fine.}. Let $b_3$ denote the number of possibilities for $E'$. \item Next we choose the set $E''$ of crossing edges in $G$ between $W$ and $B^{i}_{low}\backslash W$ such that $E''$ is compatible with our previous choices. Let $b_4$ denote the number of possibilities for $E''$. \item Finally we choose the set $E'''$ of crossing edges in $G$ between $W$ and $[n]\backslash (W\cup B^{i}_{low})$ such that $E'''$ is compatible with our previous choices. Let $b_5$ denote the number of possibilities for $E'''$. \end{itemize} Hence \begin{equation}\label{eqn: A_2 count} |\mathcal{A}_2|\leq n\max\limits_{n/2k^2 \leq a\leq n} \{b_1\cdot b_2\cdot b_3\cdot b_4\cdot b_5 \}. \end{equation} The main idea of the proof is that since $Y$ is large for $G\in \mathcal{A}_2$, it follows that $C(Y)$ is also large. So the assumption that every element of $C(Y)$ has at least one neighbour in $\{y_1,y_2\}$ places a significant restriction on the number of choices for $G$. The role of the $w^s_{\ell}$ is to `balance out' the vertex classes, i.e. in the proof of Claim~5 it will be useful that $W$ contains two vertices from each vertex class. The following series of claims will give upper bounds for the quantities $b_1,\dots, b_5$. Claims~1 and~4 are trivial, and the proof of Claim~2 proceeds in an almost identical way to that of Claim~2 in the proof of Lemma~\ref{lem: A_1}; we give proofs of Claims~3 and~5. \vspace{0.3cm} \noindent {\bf Claim 1:} $b_1\leq k^2 n^{2k-1} 2^{\binom{2k-1}{2}}$. \vspace{0.3cm} \noindent {\bf Claim 2:} $b_2\leq C 2^{2(\log n)^3} f_k(n_k) 2^{t_{k-1}(n-(2k-2))}$. \vspace{0.3cm} \noindent {\bf Claim 3:} $b_3\leq n^{4(k-1)}$. \noindent Indeed, for every graph $\tilde{G}\in \mathcal{A}_2$ such that $y_1,y_2\notin C^j(\tilde{G},Q)$ and $w_1^s, w_2^s\notin C^s(\tilde{G},Q)$ for every $s\in \{0,1,\dots, k-2\}\backslash \{j\}$, Proposition~\ref{beta}(i) implies that $\overline{d}^j_{\tilde{G},Q}(y_{\ell}), \overline{d}^s_{\tilde{G},Q}(w^s_{\ell}) \leq 2$ for every $\ell \in [2]$ and every $s\in \{0,1,\dots, k-2\}\backslash \{j\}$. Thus $$b_3\leq n^{2|W|}\leq n^{4(k-1)},$$ as required. \vspace{0.3cm} \noindent {\bf Claim 4:} $b_4\leq 2^{(2k-4)a}$. \vspace{0.3cm} \noindent {\bf Claim 5:} $b_5\leq 2^{(2k-4)\left(n-a \right)} 2^{-2n^{1/2k^2}/5}$. \noindent Indeed, suppose $G$ satisfies $C(Y) \cap \overline{N}(\{y_1,y_2\}) = \emptyset$. Since we choose $G$ such that $|B^i_{low}|=a\geq n/2k^2$, the fact that $G\notin \mathcal{A}_1$ implies that $|Y|>n/200k^2$. Now the definitions of $C^i_{low}, B^i_{low}$ imply that $$|C(Y)|\geq \frac{200k^2|Y|}{n^{1-1/2k^2}}\geq n^{1/2k^2}.$$ So since in $G$ every vertex in $C(Y)$ must have at least one neighbour in $\{y_1,y_2\}$, \begin{align}\label{eqn: b_5} b_5&\leq 2^{2\sum_{s\in \{0,1,\dots, k-2\}\backslash \{j\}} |[n]\backslash (Q_s\cup B^i_{low})|} 2^{2|[n]\backslash (Q_j\cup B^i_{low}\cup C(Y))|}3^{|C(Y)|}\\ &\leq 2^{(2k-4)\left(n-a \right)} 2^{-2n^{1/2k^2}/5},\nonumber \end{align} as required. The second inequality of (\ref{eqn: b_5}) is where it is important that $W$ contains two vertices from each vertex class. \vspace{0.3cm} \noindent Now (\ref{eqn: A_2 count}) together with Claims~1--5 and Proposition~\ref{omitted proposition}(ii) implies that \begin{align*} |\mathcal{A}_2|\leq \hspace{0.1cm} &n\cdot k^2 n^{2k-1} 2^{\binom{2k-1}{2}} \cdot C 2^{2(\log n)^3} f_k(n_k) 2^{t_{k-1}(n-(2k-2))}\\ &\cdot n^{4(k-1)} \cdot \max\limits_{n/2k^2 \leq a\leq n} \left\{ 2^{(2k-4)a} \cdot 2^{(2k-4)\left(n-a \right)} 2^{-2n^{1/2k^2}/5} \right\}\\ \leq \hspace{0.1cm} &C 2^{-n^{1/2k^2}/3} f_k(n_k)\cdot 2^{t_{k-1}(n-(2k-2))+(2k-2)(k-2)n/(k-1) - (2k-2)(k-2) - t_{k-1}(2k-2)}\\ \leq \hspace{0.1cm} &C 2^{-n^{1/2k^2}/3} f_k(n_k) 2^{t_{k-1}(n)}, \end{align*} as required. \end{proof} As mentioned earlier, a $(6,3)$-forest (with edge set $E$ say) is a useful building block for constructing many induced copies of $C_{2k}$. More precisely, in Lemma~\ref{lem: A_3} we will show that there are many `$E$-compatible' linear forests $H$, which play a similar role to that of the skeletons in the proof of Lemma~\ref{F^1}. Each such $E\cup E(H)$ gives us a non-trivial restriction on the remaining edge set, resulting in an adequate bound on $|\mathcal{A}_3|$. \begin{lemma}\label{lem: A_3} $|\mathcal{A}_3| \leq C 2^{-\frac{n}{2^{14k}}} f_k(n_k) 2^{t_{k-1}(n)}$. \end{lemma} \begin{proof} Any graph $G\in \mathcal{A}_3$ can be constructed as follows. \begin{itemize} \item First we choose distinct indices $i,j\in \{0,1,\dots, k-2\}$, a set $X\subseteq Q_i\cup Q_j$ of six vertices, and a set $E$ of edges between vertices in $X$ such that the graph $(X,E)$ is a linear forest with at most three components (so $E$ will be the edge set of a $(6,3)$-forest in $G$). Let $b_1$ denote the number of such choices. \item Next we choose a graph $G'$ on vertex set $[n]\backslash X$ such that $G[[n]\backslash X]=G'$. Let $b_2$ denote the number of possibilities for $G'$. \item Next we choose the set $E'$ of internal edges in $G$ with exactly one endpoint in $X$. Let $b_3$ denote the number of possibilities for $E'$. \item Finally we choose the set $E''$ of crossing edges in $G$ between $X$ and $[n]\backslash X$ such that $E''$ is compatible with our previous choices. Let $b_4$ denote the number of possibilities for $E''$. \end{itemize} Hence, \begin{equation}\label{eqn: A_3 count} |\mathcal{A}_3|\leq b_1\cdot b_2\cdot b_3\cdot b_4. \end{equation} The following series of claims will give upper bounds for the quantities $b_1,\dots, b_4$. Claim~1 is trivial, and the proofs of Claims~2 and~3 follow in an almost identical way to those of Claims~2 and~3 in the proof of Lemma~\ref{lem: A_1}, so we give only a proof of Claim~4. \vspace{0.3cm} \noindent {\bf Claim 1:} $b_1\leq 2^{15} k^2 n^6$. \vspace{0.3cm} \noindent {\bf Claim 2:} $b_2\leq C 2^{2(\log n)^3} f_k(n_k) 2^{t_{k-1}(n-6)}$. \vspace{0.3cm} \noindent {\bf Claim 3:} $b_3\leq 2^{6\xi(\beta) n}$. \vspace{0.3cm} \noindent {\bf Claim 4:} $b_4\leq 2^{\frac{6(k-2)n}{k-1}} 2^{\mu^{1/4}n} 2^{-\frac{n}{2^{13k}}}$. \noindent Indeed, we define an \emph{$E$-compatible forest} to be a linear forest $H$ on $2k-6$ vertices, with the same number of components as $(X,E)$, such that $V(H)\cap Q_s$ induces a clique on two vertices for every $s\in \{0,1,\dots, k-2\}\backslash \{i,j\}$. Note that an $E$-compatible forest exists since $2k-6\geq 2 \cdot 3$ and $(X,E)$ has at most three components. Moreover, an $E$-compatible forest contains a perfect matching, so Proposition~\ref{building} implies that for every graph $\tilde{G}\in \mathcal{A}_3$, the number of disjoint $E$-compatible forests in $\tilde{G}$ is at least $$\left\lfloor \frac{n/(k-1)-\mu n - 2 \mu^{1/2} n}{2} \right\rfloor \geq \frac{n}{2(k-1)}-3\mu^{1/2} n.$$ Hence $G'$ contains at least $n/2(k-1)-3\mu^{1/2} n$ disjoint $E$-compatible forests. Now fix a set $CF$ of $n/2(k-1)-3\mu^{1/2} n$ disjoint $E$-compatible forests in $G'$, and let $H\in CF$. Let $h_H$ denote the number of possibilities for a set $E^*$ of edges between $X$ and $V(H)$. By Proposition~\ref{incomplete} there exists at least one set $\tilde{E}$ of edges between $X$ and $V(H)$ such that the graph $(X\cup V(H), E\cup E(H)\cup \tilde{E})$ is isomorphic to $C_{2k}$. So since $G$ must be induced-$C_{2k}$-free, we must have that $E^*\ne\tilde{E}$, and hence $h_H\leq 2^{|X||V(H)|}-1=2^{12(k-3)}-1$. Note that the number of vertices outside $Q_i\cup Q_j$ that are not contained in some graph $H\in CF$ is at most $(k-3)n/(k-1) + 2\mu n - (2k-6)(n/2(k-1)-3\mu^{1/2} n)\leq 6k\mu^{1/2} n$. Hence, \begin{align*} b_4&\leq 2^{6\cdot \max \{|Q_i|, |Q_j|\}} 2^{6(6k\mu^{1/2} n)} \prod\limits_{H\in CF} h_H\\ &\leq 2^{6(n/(k-1)+\mu n)} 2^{6(6k\mu^{1/2} n)} \left( 2^{12(k-3)}\left( 1-2^{-12(k-3)} \right) \right)^{n/(2(k-1))-3\mu^{1/2} n}\\ &\leq 2^{\frac{6(k-2)n}{k-1}} 2^{40k\mu^{1/2} n} e^{-\frac{n/(2(k-1))}{2^{12(k-3)}}} \leq 2^{\frac{6(k-2)n}{k-1}} 2^{\mu^{1/4}n} 2^{-\frac{n}{2^{13k}}}, \end{align*} as required. \vspace{0.3cm} \noindent Now (\ref{eqn: A_3 count}) together with Claims~1--4 and Proposition~\ref{omitted proposition}(ii) implies that \begin{align*} |\mathcal{A}_3|&\leq 2^{15} k^2 n^6 \cdot C 2^{2(\log n)^3} f_k(n_k) 2^{t_{k-1}(n-6)} \cdot 2^{6\xi(\beta) n} \cdot 2^{\frac{6(k-2)n}{k-1}} 2^{\mu^{1/4}n} 2^{-\frac{n}{2^{13k}}}\\ &\leq C 2^{-\frac{n}{2^{14k}}} f_k(n_k) 2^{t_{k-1}(n-6)+6(k-2)n/(k-1) - 6(k-2) - t_{k-1}(6)}\\ &\leq C 2^{-\frac{n}{2^{14k}}} f_k(n_k) 2^{t_{k-1}(n)}, \end{align*} as required. \end{proof} The next proposition shows that for every $G \in \mathcal{A}_4$, the small stars and triangles in $\overline{G}[Q_0]$ do not cover too many vertices. \begin{proposition} \label{B_low} For every $G \in \mathcal{A}_4$ and index $i\in \{0,1,\dots, k-2\}$, $|B^i_{low}| < n/2k^2$. \end{proposition} \begin{proof} Suppose for a contradiction that there exists a graph $G \in \mathcal{A}_4$ such that $|B^i_{low}| \geq n/2k^2$ for some index $i\in \{0,1,\dots, k-2\}$. Since $G \in \mathcal{A}_4 \subseteq F^3_Q$, $G$ is not a $k$-template. This fact together with Proposition~\ref{beta}(i) implies that there exists an index $j\in \{0,1,\dots, k-2\}\backslash \{i\}$ and a non-edge $y_1y_3$ inside $Q_j$. At most one of $y_1,y_3$ can be in $C^j$ (by definition of $C^j$), and so without loss of generality we assume that $y_1 \notin C^j$. So Proposition~\ref{beta}(i) implies that $\overline{d}^j_{G,Q}(y_1)\leq 2$. This together with the observation that $|Q_j\backslash C^j|\geq |Q_j|/2$ (by definition of $C^j$) implies that there exists a vertex $y_2 \in Q^j\setminus C^j$ that is a neighbour of $y_1$. Define $Y:= \overline{N}(\{y_1,y_2,y_3\}) \cap B^{i}_{low}$. Since $|B^i_{low}| \geq n/2k^2$ and $G\notin \mathcal{A}_1$, $|Y|> n/200k^2$. Since $|B^i_{low}| \geq n/2k^2$ and $G\notin \mathcal{A}_2$, $C(Y)$ contains a vertex $x_3\in \overline{N}(\{y_1,y_2\})$. Since $x_3\in C(Y)$ there exists a vertex $x_1\in Y$ that is a non-neighbour of $x_3$. By Proposition~\ref{beta}(iii), $\overline{d}^i_{G,Q}(x_1), \overline{d}^i_{G,Q}(x_3) \leq \beta n$. So since $|Y| > n/200k^2 \geq 2\beta n$, there exists a vertex $x_2 \in Y\cap N(\{x_1,x_3\})$. Then $E(G[\{x_1,x_2,x_3,y_1,y_2,y_3\}]) = \{ x_1x_2, x_2x_3, y_1y_2\} \cup E'$ with $E'\subseteq \{y_2y_3, y_3x_3\}$. Thus the set $\{x_1,x_2,x_3,y_1,y_2,y_3\}\subseteq Q_i\cup Q_j$ induces on $G$ a linear forest with at most three components, and so $G$ contains a $(6,3)$-forest with respect to $Q$. This contradicts the fact that $G\notin \mathcal{A}_3$, and hence completes the proof. \end{proof} We now have sufficient information about the set $\mathcal{A}_4$ of remaining graphs to count them directly (i.e. $\mathcal{A}_4$ is the only class for which we do not use induction in our estimates). In particular, we now know that in $\overline{G}$ every vertex class is the union of triangles and stars, where crucially the number of triangles and small stars is not too large (see Proposition~\ref{B_low}). This allows us to show by a direct counting argument that $|\mathcal{A}_4|$ is negligible. \begin{lemma}\label{lem: A_4} $|\mathcal{A}_4| \leq 2^{-\frac{n\log n}{3k^2}} f_k(n_k) 2^{t_{k-1}(n)}$. \end{lemma} \begin{proof} Any graph $G\in \mathcal{A}_4$ can be constructed as follows. \begin{itemize} \item First we choose a partition of $Q_i$ into five sets, $C^i_h, B^i_h, C^i_\ell, B^i_\ell, C^i_z,$ for every $i\in \{0,1,\dots, k-2\}$. Let $b_1$ denote the number of such choices. \item Next we choose the set $E$ of crossing edges in $G$ with respect to $Q$. Let $b_2$ denote the number of possibilities for $E$. \item Finally we choose the set $E'$ of internal edges in $G$ with respect to $Q$ such that $G$ satisfies $C^i_h=C^i_{high}$, $B^i_h=B^i_{high}$, $C^i_\ell=C^i_{low}$, $B^i_\ell=B^i_{low}$, and $C^i_z=C^i_0$ for every $i\in \{0,1,\dots, k-2\}$. Let $b_3$ denote the number of possibilities for $E'$. \end{itemize} Hence \begin{equation}\label{eqn: A_4 count} |\mathcal{A}_4|\leq b_1\cdot b_2\cdot b_3. \end{equation} The following series of claims will give upper bounds for the quantities $b_1,b_2,b_3$. Claims~1 and~2 are trivial; we give only a proof of Claim~3. \vspace{0.3cm} \noindent {\bf Claim 1:} $b_1\leq 5^n$. \vspace{0.3cm} \noindent {\bf Claim 2:} $b_2\leq 2^{t_{k-1}(n)}$. \vspace{0.3cm} \noindent {\bf Claim 3:} $b_3\leq 2^{ \frac{(k-1/2)n\log n}{k^2}}$. \noindent For any given $i\in \{0,1,\dots, k-2\}$ and any vertex $x\in B^i_{high}$, the number of possibilities for the unique non-neighbour of $x$ in $C^i_{high}$ (namely the centre of the star in $\overline{G}$ containing $x$) is $|C^i_{high}|$. Now consider $x\in B^i_{low}$. Then $x$ has a unique non-neighbour $y$ in $C^i_{low}$, and has the possibility of either being part of a triangle in $\overline{G}$ or a star in $\overline{G}$. Note also that $|B^i_{low}|< n/2k^2$ by Proposition~\ref{B_low}, and that by definition of $C^i_{high}$, $$|C^i_{high}|\leq \frac{200k^2 n}{n^{1-1/2k^2}}\leq 200k^2 n^{1/2k^2}.$$ Hence, \begin{align*} b_3 &\leq \prod_{i=0}^{k-2} (2|C^i_{low}|)^{|B^i_{low}|} |C^i_{high}|^{|B^i_{high}|} \leq \prod_{i=0}^{k-2} n^{\frac{n}{2k^2}} (200k^2)^{n} (n^{\frac{1}{2k^2}})^n = 2^{ \frac{(k-1)n\log n}{k^2}} (200k^2)^{n(k-1)}\\ &\leq 2^{\frac{(k-1/2)n\log n}{k^2}}, \end{align*} as required. \vspace{0.3cm} \noindent Now (\ref{eqn: A_4 count}) together with Claims~1--3 and Lemma~\ref{f-estimate-1} implies that \begin{align*} |\mathcal{A}_4|&\leq 5^n \cdot 2^{t_{k-1}(n)} \cdot 2^{ \frac{(k-1/2)n\log n}{k^2}}\\ &\leq 5^n 2^{-\frac{n\log n}{2k^2}} 2^{n_k \log n_k - en_k \log \log n_k} 2^{en_k \log \log n_k} 2^{t_{k-1}(n)}\leq 2^{-\frac{n\log n}{3k^2}} f_k(n_k) 2^{t_{k-1}(n)}, \end{align*} as required. \end{proof} Recall that $F^3_Q= (F^3_Q\backslash T^*_Q(n,k))\cup \mathcal{A}_1 \cup \mathcal{A}_2 \cup \mathcal{A}_3 \cup \mathcal{A}_4$. The following bound on $|F^3_Q|$ follows immediately from this observation together with (\ref{strong regular}) and Lemmas~\ref{lem: A_1},~\ref{lem: A_2},~\ref{lem: A_3} and~\ref{lem: A_4}. \begin{lemma}\label{F^3} $|F^3_Q|\leq 2C2^{-n^{\frac{1}{2k^2}}/3}f(n_k) 2^{t_{k-1}(n)}$. \end{lemma} \section{Proof of Lemma~\ref{induction conclusion}}\label{sec: final calculation} \removelastskip\penalty55\medskip\noindent{\bf Proof of Lemma~\ref{induction conclusion}.} Recall from Section~\ref{sec: main proof} that we prove Lemma~\ref{induction conclusion} by induction on $n$ and that we choose constants satisfying (\ref{eqn: hierarchy}). The fact that $1/C \ll 1/n_0, 1/k$ implies that the statement of Lemma~\ref{induction conclusion} holds for all $n\leq n_0$. So suppose that $n>n_0$ and that the statement holds for all $n'< n$. Then we obtain the bounds in Lemmas~\ref{F^1}, \ref{F^2_Q} and~\ref{F^3}. These bounds together with the fact that $F_Q(n,k,\eta,\mu) = T_Q \cup F^1_Q \cup F^2_Q \cup F^3_Q$ and $T_Q\subseteq T_Q(n,k)$ imply that \begin{align*} |F_Q(n,k,\eta,\mu) \setminus T_Q(n,k)| &\leq C\left( 2^{-\beta^2 n/14^{k}} + 2^{-n} + 2\cdot 2^{-n^{\frac{1}{2k^2}}/3} \right) f_k(n_k) 2^{t_{k-1}(n)}\\ &\leq 3C 2^{-n^{\frac{1}{2k^2}}/3} f_k(n_k) 2^{t_{k-1}(n)}. \end{align*} This together with Corollary~\ref{regular lemma} implies that \begin{align}\label{eqn: F_Q(n,k,eta) size} |F_Q(n,k,\eta)\setminus T_Q(n,k)| &\leq |F_Q(n,k,\eta)\setminus F_Q(n,k,\eta,\mu)|+|F_Q(n,k,\eta,\mu)\setminus T_Q(n,k)| \\ & \leq \left(2^{- \frac{\mu^2n^2}{100}} + 3C 2^{-n^{\frac{1}{2k^2}}/3} f_k(n_k) \right) 2^{t_{k-1}(n)}\nonumber\\ &\leq 4C 2^{-n^{\frac{1}{2k^2}}/3} f_k(n_k) 2^{t_{k-1}(n)}.\nonumber \end{align} Note that Lemma~\ref{lem: rough structure} (applied with $\eta/2$ playing the role of $\eta$) together with (\ref{eqn: near templates}) implies that \begin{equation} \label{rough} |F(n,k)\setminus F(n,k,\eta)| \leq 2^{-\varepsilon n^2}|F(n,k,\eta)|. \end{equation} Let $\mathcal{Q}$ denote the set of all ordered $(k-1)$-partitions of $[n]$, and recall that our choice of $Q\in \mathcal{Q}$ was arbitrary. Now (\ref{eqn: F_Q(n,k,eta) size}) together with (\ref{rough}) and Lemma~\ref{size of template} implies that \begin{align*} |F(n,k)\setminus F(n,k,\eta)| &\leq 2^{-\varepsilon n^2} \sum_{Q'\in \mathcal{Q}} \left(|F_{Q'}(n,k,\eta)\backslash T_{Q'}(n,k)|+|T_{Q'}(n,k)|\right)\\ &\leq 2^{-\varepsilon n^2} (k-1)^n \left(4C 2^{-n^{\frac{1}{2k^2}}/3} + 2^{6 (\log n)^2} \right)f_k(n_k) 2^{t_{k-1}(n)}\\ &\leq C 2^{-\varepsilon n^2 /2 } f_k(n_k) 2^{t_{k-1}(n)}. \end{align*} Now this together with (\ref{eqn: F_Q(n,k,eta) size}) implies that \begin{align*} |F_Q(n,k)| &\leq |F_Q(n,k,\eta)| + |F(n,k)\setminus F(n,k,\eta)|\\ &\leq |T_Q(n,k)| + |F_Q(n,k,\eta)\setminus T_Q(n,k)| + |F(n,k)\setminus F(n,k,\eta)|\\ &\leq |T_Q(n,k)|+ \left( 4\cdot 2^{-n^{\frac{1}{2k^2}}/3} + 2^{-\epsilon n^2 /2} \right) C f_k(n_k) 2^{t_{k-1}(n)}\\ &\leq |T_Q(n,k)|+ 5C 2^{-n^{\frac{1}{2k^2}}/3}f_k(n_k) 2^{t_{k-1}(n)}, \end{align*} which completes the inductive step, and hence the proof. \endproof \section{Acknowledgement} We are grateful to Mihyun Kang for helpful remarks on the number of $k$-templates.

https://arxiv.org/abs/1710.09000

Control Problems with Vanishing Lie Bracket Arising from Complete Odd Circulant Evolutionary Games

We study an optimal control problem arising from a generalization of rock-paper-scissors in which the number of strategies may be selected from any positive odd number greater than 1 and in which the payoff to the winner is controlled by a control variable $\gamma$. Using the replicator dynamics as the equations of motion, we show that a quasi-linearization of the problem admits a special optimal control form in which explicit dynamics for the controller can be identified. We show that all optimal controls must satisfy a specific second order differential equation parameterized by the number of strategies in the game. We show that as the number of strategies increases, a limiting case admits a closed form for the open-loop optimal control. In performing our analysis we show necessary conditions on an optimal control problem that allow this analytic approach to function.

\section{Introduction} Cyclic competition occurs frequently in nature \cite{JB75,SL96,KRFB02,GHS03,KR04,KNS05,NHK11}. Cyclic competition in coral reef populations are studied in \cite{JB75}. Sinervo and Lively first characterized rock-paper-scissors like competition in lizards \cite{SL96}, while Gilg, Hanski and Sittler \cite{GHS03} study this behavior in rodents. Cyclic behavior in microbial populations is studied in \cite{KRFB02,KR04,KNS05,NHK11}. In classical and evolutionary game theory, cyclic dominance (e.g., matching pennies, rock-paper-scissors) games are commonly studied \cite{Mor94,Wei95}. Biologically speaking, in an idealized cyclic game, the absolute fitness measure (payoff) resulting from species interaction can be represented by a circulant matrix, in which row $k$ is a rotation of row $k-1$ for each $k$. Games with circulant payoff matrices have been studied extensively in evolutionary game theory \cite{Ze80,SSW80,HS98,HS03,DG09,HS00,SMJS14,GK16} and provide some of the most interesting behaviors \cite{SMJS14}. In early work, cyclic interaction is considered without explicit reference to games. Cyclic (chemo-biological) interactions are studied extensively in \cite{SSW78,HSSW80,SSW79} in which both competitive and cooperative behaviors are identified. Analysis of the replicator in which a circulant matrix emerges is studied in \cite{SSW80} as a result of cyclic mass interaction kinetics. Zeeman \cite{Ze80} made an early study of the dynamics of cyclic games showing that in rock-paper-scissors a degenerate Hopf bifurcation leads to the emergence of a non-linear center with no limit cycle possible in any 3 strategy game under the replicator dynamics. Since this early work, several authors have investigated various cyclic games and games characterized by circulant matrices. Among many other works: Hofbauer and Schlag \cite{HS00} consider imitation in cyclic games; Diekmann and Gils specifically study the cyclic replicator dynamics and focus on the properties of low-dimensional cyclic games \cite{DG09}; Ermentrout et al. consider a transition matrix evolutionary dynamic in which a limit cycle emerges in the rock-paper-scissors game \cite{EGB16}; and Griffin and Belmonte \cite{GB17} study a triple public goods game and show that is is diffeomorphic to generalized rock-paper-scissors. Each of these works focuses explicitly on classes of circulant games, while recent work by Grani{\'{c}} and Kerns \cite{GK16} characterizes the Nash equilibria of arbitrary circulant games, but does not focus on the evolutionary game context. There has also been extensive work on spatial games with circulant payoff matrices. Peltom\"aki and Alvara \cite{PA08} consider both 3 and 4 state rock-paper-scissors. Other papers consider rock-paper-scissors with variations on reaction rate \cite{HMT10} or study the basins of attraction \cite{SWYL10}. DeForest and Belmonte \cite{dB13} study a fitness gradient variation on the spatial replicator and show rock-paper-scissors can exhibit spatial chaos under these dynamics. Finally more recent work by Szczesny et. al \cite{SMR14} considers spiral formations in rock-paper-scissors. In this paper, we extend work in \cite{GB17} by studying an optimal control problem defined on a $N$-strategy ($N=3,5,7,\dots$) generalization of rock-paper-scissors. Odd cardinality interactions are interesting because they model specific biological cases \cite{SL96,GHS03,SSW78,HSSW80,SSW79}. Additionally, when $N$ is very large, these have the potential to model systems in which individuals a variety of individuals with strengths and weaknesses interact. In particular, the payoff matrix is (i) defined by the sum of two circulant matrices, and (ii) admits a single control parameter. Thus we consider the general class of control problems first studied in a specific case in \cite{GB17}. Our payoff matrix is inspired by the generalized rock-paper-scissors matrix defined in \cite{Wei95}. Since every pair of heterogeneous strategic interactions (e.g., rock vs. scissors) results in a non-zero payoff, we refer to this class of games as \textit{complete odd circulant games}. The major results of this paper are: \begin{enumerate} \item We show that the complete odd circulant games admit only a unique interior fixed point under the replicator dynamics. We also characterize the fixed points of the $N$-strategy complete odd circulant game in terms of the fixed points of the $M < N$ complete odd circulant games. \item We show that the replicator dynamics can be written as the sum of an uncontrolled component and a controlled component both of which have circulant Jacobian matrices. \item As a consequence, we completely characterize the stability of the interior fixed point and use this to define an optimal control problem with objective to drive the system trajectories toward this interior fixed point. \item We describe the properties of a general class of control problems with a vanishing Lie Bracket that will be used to analyze the control problem we define. This suggests interesting geometric and algebraic connections between evolutionary games and optimal control theory. As a by-product, we generalize recent control theoretic results in \cite{FG17}. \item We show that a quasi-linearization of the control problem (as done in \cite{GB17}) has special form admitting a complete characterization of the dynamics of the optimal control. We also derive a sufficient condition on control optimality and thus completely generalize the results in \cite{GB17} to arbitrary complete odd circulant games. \item As a part of the generalization, we find a limiting second order ordinary differential equation (as $N$ grows large) that the optimal control must obey and show that it has a natural closed form solution. \end{enumerate} The remainder of this paper is organized as follows: In Section \ref{sec:Prelim} we present preliminary results and notation. In Section \ref{sec:GeneralControl} we introduce the control problem of interest and study a general class of optimal control problems that will assist in the derivation of our main results. Our main results on control of complete odd circulant games are found in Section \ref{sec:CyclicControl}. Conclusions and future directions are presented in Section \ref{sec:Conclusion}. \section{Notation and Preliminary Results}\label{sec:Prelim} A \textit{circulant matrix} is a square matrix with form: \begin{displaymath} \mathbf{A} = \begin{bmatrix} a_0 & a_{n-1} & a_{n-2} & \cdots & a_{1}\\ a_1 & a_0 & a_{n-1} & \cdots & a_{2}\\ \vdots & \vdots & \vdots & \ddots & \vdots\\ a_{n-1} & a_{n-2} & a_{n-3} & \cdots & a_0 \end{bmatrix}. \end{displaymath} A circulant matrix is entirely characterized by its first row and all other rows are cyclic permutations of this first row. The set of $N \times N$ circulant matrices forms a commutative algebra, a fact that will be used frequently in this paper. Moreover, the eigenvalues of these matrices have special form. If $\mathbf{A}$ is an $N \times N$ circulant matrix and $\omega_0,\dots,\omega_{N-1}$ are the $N^\text{th}$ roots of unity, then eigenvalue $\lambda_j$ ($j=0,\dots,N-1$) is given by the expression: \begin{displaymath} \lambda_j = a_0 + a_{n-1}\omega_j + a_{n-2}\omega_j^2 + \cdots + a_{1}\omega_j^{N-1}. \end{displaymath} Further details on this class of matrices is available in \cite{D12}. Let: \begin{displaymath} \Delta_N = \left\{\mathbf{u} \in \mathbb{R}^N : \mathbf{1}^T\mathbf{u} = 1, \mathbf{u} \geq \mathbf{0}\right\} \end{displaymath} be the unit $N$-simplex embedded in $N$-dimensional Euclidean space. Here $\mathbf{1}$ is an appropriately sized vector of $1$'s and $\mathbf{0}$ is a zero vector. We consider a family of control problems defined on parametrized cyclic games with $N = 2n + 1$ strategies, where $n = 1,2,\dots,$. If $i \in \{1,\dots,N\}$ and $j \in \mathbb{Z}$, define: \begin{equation} \mu(k) = \begin{cases} k \mod N & \text{if $k \mod N \neq 0$},\\ N & \text{otherwise}. \end{cases} \end{equation} For the remainder of this paper, define $\mathbf{L}_N,\mathbf{M}_N \in \mathbb{R}^{N \times N}$ so that: \begin{equation} \mathbf{L}_{N_{ij}} = \begin{cases} (-1)^{\mu(N(i-1)+j-i)} & \text{if $i \neq j$},\\ 0 & \text{otherwise}. \end{cases} \end{equation} and \begin{equation} \mathbf{M}_{N_{ij}} = \begin{cases} 1 & \text{if $i \neq j$ and $(-1)^{\mu(N(i-1)+j-i)}=1 $},\\ 0 & \text{otherwise}. \end{cases} \end{equation} By way of example, we illustrate the matrices $\mathbf{L}_5$ and $\mathbf{M}_5$ for the 5-strategy cyclic game. \begin{displaymath} \mathbf{L}_5 = \begin{bmatrix} 0 & -1 & 1 & -1 & 1 \\ 1 & 0 & -1 & 1 & -1 \\ -1 & 1 & 0 & -1 & 1 \\ 1 & -1 & 1 & 0 & -1 \\ -1 & 1 & -1 & 1 & 0 \end{bmatrix}, \qquad \mathbf{M}_5 = \begin{bmatrix} 0 & 0 & 1 & 0 & 1 \\ 1 & 0 & 0 & 1 & 0 \\ 0 & 1 & 0 & 0 & 1 \\ 1 & 0 & 1 & 0 & 0 \\ 0 & 1 & 0 & 1 & 0 \end{bmatrix}. \end{displaymath} From a game-theoretic perspective, we can think of $\mathbf{L}_N$ as being the traditional payoff matrix of the complete cyclic game with $N$ strategies; e.g., the payoff matrix of rock-paper-scissors. On the other hand, $\mathbf{M}_N$ can be thought of an actuating matrix that will determine whether the interior fixed point of the complete cyclic game is stable or unstable, as we show in the sequel. In the remainder of this paper, we will consider the generalized cyclic game with $N$ strategies where $N = 3, 5, 7, \dots$, and we note that both $L_N$ and $M_N$ are circulant matrices. The payoff matrix for the generalized cyclic game with $N$ strategies and parameter $\gamma$ is: \begin{displaymath} \mathbf{A}_N(\gamma) = \mathbf{L}_N + \gamma\mathbf{M}_N. \end{displaymath} If $\gamma = 0$ and $N = 3$, then $\mathbf{A}_3(\gamma)$ is just the rock-paper-scissors matrix. Without loss of generality, we assume $\gamma > -1$. Otherwise, the natural winning precedence in the cyclic game is reversed. The matrix $\mathbf{M}_N$ is the adjacency matrix of an $N$ vertex circulant directed graph $\mathcal{G}(\mathbf{M}_N)$, whose edge direction determines the winning precedence between two strategies in the underlying cyclic game. This is illustrated in Figure \ref{fig:GameGraphs}. \begin{figure}[htbp] \centering \subfigure[Five Strategy Game]{\includegraphics[scale=0.65]{Figures/FiveStratGame.pdf}} \subfigure[Seven Strategy Game]{\includegraphics[scale=0.65]{Figures/SevenStratGame.pdf}} \caption{Five and seven strategy game represented as graphs.} \label{fig:GameGraphs} \end{figure} In particular, if $\mathbf{M}_{N_{ij}} = 1$, then Strategy $i$ defeats Strategy $j$ and yields payoff $1+\gamma$ in $\mathbf{A}_N(\gamma)$. In the control problem defined in the sequel, the replicator dynamics are the nonlinear equations of motion with control parameter $\gamma$: \begin{equation} \mathcal{S}_N =\left\{ \begin{aligned} &\dot{\mathbf{u}}_i = \mathbf{u}_i\left((\mathbf{e}_i - \mathbf{u})^T\mathbf{A}_N(\gamma)\mathbf{u}\right),\\ &\mathbf{u}(0) = \mathbf{u}_0. \end{aligned}\right. \label{eqn:RepDyn} \end{equation} Here $\mathbf{u} = \langle{u_1,\dots,u_N}\rangle$ is the vector denoting the proportion of the population playing each of the $N$ strategies. It is well known \cite{HS98,Wei95} that if $\mathbf{u}_0 \in \Delta_N$, then $\mathbf{u}(t)$ is confined to $\Delta_N$ for all time. For the remainder of this paper, we assume $\mathbf{u}_0 \in \Delta_N$. Assume $m,n \in \{1,2,\dots,\}$ and let $M = 2m+1$ and $N = 2n+1$. The point-to-set mapping $\pi_{M,N}:\Delta_M \to 2^{\partial\Delta_N}$ maps a point $\mathbf{u} \in \Delta_M$ to the set of points $\mathbf{v}$ in $\partial\Delta_N$ so that: \begin{enumerate} \item $M$ elements of the vector $\mathbf{v}$ consist of the elements of $\mathbf{u}$. The remaining $M-N$ elements are $0$. \item If the rows and columns of $\mathbf{L}_N$ (resp. $\mathbf{M}_N$) corresponding to the zero-entries in $\mathbf{v}$ are deleted to form the matrix $\tilde{\mathbf{L}}_N$ (resp. $\tilde{\mathbf{M}}_N$), then $\tilde{\mathbf{L}}_N = \mathbf{L}_M$ (resp. $\tilde{\mathbf{M}}_N = \mathbf{M}_M$. \end{enumerate} The second condition is equivalent to stating that the vertices corresponding to the non-zero strategies in $\mathbf{v}$ induce a sub-graph of $\mathcal{G}(\mathbf{M}_N)$ that is isomorphic to $\mathcal{G}(\mathbf{M}_M)$. \begin{lemma} Let $n \in \{1,2,\dots,\}$ and $N = 2n+1$, then: \begin{enumerate} \item $\mathcal{S}_N$ has among its fixed points $\mathbf{e}_i \in \Delta_N$ ($i=1,\dots,N$) and $\frac{1}{N}\mathbf{1} \in \mathrm{int}(\Delta_N)$. \item Furthermore, if $n > m \geq 1$, $M = 2m+1$, and $N = 2n+1$ and $\mathbf{u}^+$ is a fixed point of $\mathcal{S}_M$, then $\pi_{M,N}(\mathbf{u}^+)$ are fixed points of $\mathcal{S}_N$. \end{enumerate} \end{lemma} \begin{proof} To prove Statement 1, note first that pure strategies are always fixed points of the replicator dynamics \cite{Wei95}. To see that $\mathbf{u}^* = \frac{1}{N}\mathbf{1}$ is a fixed point, note that: \begin{gather*} \mathbf{L}_N\mathbf{u}^* = \mathbf{0},\\ \gamma\mathbf{M}_N\mathbf{u}^* = \gamma\frac{n}{N}\mathbf{1}. \end{gather*} Consequently: \begin{displaymath} \gamma\left(\mathbf{u}^*\right)^T\mathbf{M}_N\mathbf{u}^* = \gamma\frac{n}{N} = \gamma\mathbf{e}_i^T\mathbf{M}_N\mathbf{u}^*. \end{displaymath} The fact that $\mathbf{u}^* = \frac{1}{N}\mathbf{1}$ is a fixed point follows immediately. To prove Statement 2, suppose that $\mathbf{u} \in \pi_{M,N}(\mathbf{u}^+)$ for some fixed point $\mathbf{u}^+$ of $\mathcal{S}_M$. Let $\tilde{\mathbf{L}}_N$ and $\tilde{\mathbf{M}}_N$ be the square sub-matrices of $\mathbf{L}_N$ and $\mathbf{M}_N$ obtained by removing the rows and columns corresponding to the zero entries in $\mathbf{u}$. By assumption, it follows at once that: \begin{displaymath} (\mathbf{e}_i^T - \mathbf{u})^T\mathbf{A}_N(\gamma)\mathbf{u} = (\mathbf{e}_i^T - \mathbf{u}^+)^T\mathbf{A}_M(\gamma)\mathbf{u}^+ = 0. \end{displaymath} This completes the proof. \end{proof} \begin{corollary} The fixed point $\mathbf{u}^* = \frac{1}{N}\mathbf{1}$ is the unique interior fixed point for $\mathcal{S}_N$. \label{cor:uu} \end{corollary} \begin{proof} If $\mathbf{u}$ is any interior fixed point, then necessarily it must satisfy the equation: \begin{displaymath} \mathbf{A}_N(\gamma)\mathbf{u} = \mathbf{u}^T\mathbf{A}_N(\gamma)\mathbf{u} \cdot \mathbf{1}. \end{displaymath} Note that for any $\mathbf{u} \in \Delta_N$, $\mathbf{u}^T\mathbf{L}\mathbf{u} = 0$. Therefore, $\mathbf{u}$ must satisfy: \begin{displaymath} \mathbf{A}_N(\gamma)\mathbf{u} = \gamma\cdot \mathbf{u}^T\mathbf{M}_N\mathbf{u} \cdot \mathbf{1}. \end{displaymath} It is straight forward to compute: \begin{displaymath} \mathbf{u}^T\mathbf{M}_N\mathbf{u} = \sum_{i=1}^N\sum_{j > i} u_iu_j. \end{displaymath} We now proceed in cases. If $\gamma = 0$, then we solve: \begin{align*} \mathbf{L}_N\mathbf{u} & = \mathbf{0},\\ \mathbf{1}^T\mathbf{u} &= 1. \end{align*} This system has rank $n$ with $n$ unknowns and therefore admits a unique solution. Elementary row reduction shows that $\mathbf{u}^* = \frac{1}{N}\mathbf{1}$ is the unique fixed point in this case. On the other hand, if $\gamma \neq 0$, then let: \begin{displaymath} \alpha = \sum_{i=1}^N\sum_{j > i} u_iu_j. \end{displaymath} We can determine a relationship between the variables by solving: \begin{displaymath} \mathbf{A}_N(\gamma)\mathbf{u} = \alpha\mathbf{1}. \end{displaymath} Row-reduction on the system shows that when $\gamma \neq 0$ we obtain the relationship: \begin{equation} \mathbf{u} = \frac{\alpha}{n\gamma} \mathbf{1}. \end{equation} Consequently, $u_1 = u_2 = \cdots = u_N$ and necessarily $\mathbf{u}^* = \frac{1}{N}\mathbf{1}$ is again the unique interior fixed point. This completes the proof. \end{proof} For the remainder of this paper, assume $\mathbf{u}^* = \frac{1}{N}\mathbf{1}$ and let $\mathbf{F},\mathbf{G}:\Delta_N \rightarrow \mathbb{R}^n$ be defined component-wise as: \begin{align} \mathbf{F}_i(\mathbf{u}) &= u_i\left((\mathbf{e}_i - \mathbf{u})\mathbf{L}_N\mathbf{u}\right)\label{eqn:Fz},\\ \mathbf{G}_i(\mathbf{u}) &= u_i\left((\mathbf{e}_i - \mathbf{u})\mathbf{M}_N\mathbf{u}\right)\label{eqn:Gz}. \end{align} The replicator dynamics are then: \begin{displaymath} \dot{u}_i = \mathbf{F}(\mathbf{u}) + \gamma\mathbf{G}(u), \end{displaymath} which are the dynamics that will be used in the control problem of interest. We note that $\mathbf{F}$ and $\mathbf{G}$ are the functional imprints of the standard payoff matrix $\mathbf{L}_N$ and the actuation matrix $\mathbf{M}_N$ within the replicator framework. \begin{lemma} The Jacobian matrix of $\mathbf{F}$ evaluated at $\mathbf{u}^* = \frac{1}{N}\mathbf{1}$ is: \begin{displaymath} \mathbf{J} \triangleq D_\mathbf{u}\mathbf{F} = \frac{1}{N}\mathbf{L}_N. \end{displaymath} \label{lem:DF} \end{lemma} \begin{proof} We prove the result for Row 1 of $D_\mathbf{u}\mathbf{F}$. The remainder of the argument follows from the circulant structure of $\mathbf{L}_N$. We have: \begin{displaymath} \mathbf{F}_1(\mathbf{u}) = u_1\mathbf{e}_1^T\mathbf{L}_N\mathbf{u}, \end{displaymath} because $\mathbf{u}^T\mathbf{L}_N\mathbf{u} = 0$. Note: \begin{equation} u_1\mathbf{e}_1^T\mathbf{L}_N\mathbf{u} = u_1\left(\sum_{j=2}^{N}(-1)^{j-1}u_j\right). \label{eqn:Row1} \end{equation} Differentiating with respect to $u_1$ and evaluating at $u_1 = u_2 = \cdots = u_n = \tfrac{1}{N}$ yields: \begin{displaymath} \left[D_\mathbf{u}\mathbf{F}\right]_{1,1} = \frac{1}{N}\sum_{j=2}^{N}(-1)^{j-1} = 0, \end{displaymath} since $N$ is odd. Differentiating Expression \ref{eqn:Row1} with respect to $u_j$ and evaluating at $u_1 = u_2 = \cdots = u_n = \tfrac{1}{N}$ yields: \begin{displaymath} \left[D_\mathbf{u}\mathbf{F}\right]_{1,j} = \frac{(-1)^{j-1}}{N} = \frac{1}{N}\mathbf{L}_{N_{1,j}}. \end{displaymath} The result now follows from the fact that $\mathbf{L}_N$ is a circulant matrix. \end{proof} \begin{lemma} The Jacobian matrix of $\mathbf{G}$ evaluated at $\mathbf{u}^* = \frac{1}{N}\mathbf{1}$ is: \begin{displaymath} \mathbf{H} \triangleq D_\mathbf{u}\mathbf{G} = \frac{1}{N^2}\mathbf{M}_N - \frac{2n}{N^2}\left(\mathbf{1}_N - \mathbf{M}_N\right) = \frac{1}{N^2}\left(N\left(\mathbf{M}_N-\mathbf{1}_N\right) + \mathbf{1}_N\right), \end{displaymath} where $\mathbf{1}_N$ is an $N \times N$ matrix of $1$'s. \label{lem:DG} \end{lemma} \begin{proof} As in the proof of Lemma \ref{lem:DF}, we show the result for Row 1 of $D_\mathbf{u}\mathbf{G}$. The remainder of the argument follows from the circulant structure of $\mathbf{M}_N$. We have already noted in Corollary \ref{cor:uu} that: \begin{displaymath} \mathbf{u}^T\mathbf{M}_N\mathbf{u} = \sum_{i=1}^N\sum_{j > i} u_iu_j. \end{displaymath} We compute: \begin{displaymath} \mathbf{e}_1^T\mathbf{M}_N\mathbf{u} = \sum_{j=1}^{(N-1)/2} u_{2j+1}. \end{displaymath} Differentiate with respect to $k = 2j+1$ for $j \in \{1,\dots,n\}$, corresponding to a non-zero index in the first row of $\mathbf{M}_N$. We have: \begin{displaymath} \frac{\partial\mathbf{G}}{\partial u_k} = u_1\left(1 - \left(\sum_{j \neq k} u_j\right)\right). \end{displaymath} Evaluating at $u_1 = u_2 = \cdots = u_n = \tfrac{1}{N}$ we obtain: \begin{displaymath} \left[D_\mathbf{u}\mathbf{G}\right]_{1,k} = \frac{1}{N}\left(1 - \frac{N-1}{N}\right) = \frac{1}{N^2}, \end{displaymath} for $k=2j+1$ with $j \in \{1,\dots,n\}$. Differentiate now with respect to $u_1$ to obtain: \begin{equation} \frac{\partial\mathbf{G}}{\partial u_1} = \sum_{j=1}^{(N-1)/2} u_{2j+1} - 2\sum_{j = 2}^N u_1u_j - \sum_{i=2}^N\sum_{j > i}u_{i}u_{j}. \end{equation} Evaluating at $u_1 = u_2 = \cdots = u_n = \tfrac{1}{N}$ we obtain: \begin{multline*} \left[D_\mathbf{u}\mathbf{G}\right]_{1,1} = \frac{n}{N} - 2\frac{2n}{N^2} - \frac{1}{N^2}\left(\frac{1}{2}(N-2)(N-1)\right) = \\ \frac{-N^2-2nN+N-2}{N^2} = -\frac{2n}{N}, \end{multline*} when $2n+1$ is substituted for $N$ in the numerator. Finally, consider $k \neq 1$ and $k \neq 2j+1$ for $j \in \{1,\dots,n\}$. Differentiating with respect to $u_k$ we have: \begin{displaymath} \frac{\partial\mathbf{G}}{\partial u_k} = -u_1\left(\sum_{j \neq k}u_j\right). \end{displaymath} Evaluating at $u_1 = u_2 = \cdots = u_n = \tfrac{1}{N}$ we obtain: \begin{displaymath} \left[D_\mathbf{u}\mathbf{G}\right]_{1,k} = -\frac{1}{N^2}(N-1) = -\frac{2n}{N}. \end{displaymath} The result now follows from the fact that $\mathbf{M}_N$ is a circulant matrix. \end{proof} \begin{theorem} If $\gamma > 0$, then the fixed point $\mathbf{u}^* = \frac{1}{N}\mathbf{1}$ is asymptotically stable. If $\gamma < 0$, then the fixed point $\mathbf{u}^*$ is asymptotically unstable. \label{thm:GammaSign} \end{theorem} \begin{proof} From Lemmas \ref{lem:DF} and \ref{lem:DG}, we note that the Jacobian matrix of $\mathcal{S}_N$ at $\mathbf{u}^*$ has the following form: \begin{displaymath} \bm{\mathcal{J}} = \mathbf{J} + \gamma\mathbf{H} = \frac{1}{N}\mathbf{L}_N + \gamma\left(\frac{1}{N^2}\mathbf{M}_N - \frac{2n}{N^2}\left(\mathbf{1}_N - \mathbf{M}_N\right)\right). \end{displaymath} By its construction, it is a circulant matrix with first row given by: \begin{equation} \bm{\mathcal{J}}_{1j} = \begin{cases} - \gamma\frac{2n}{N^2} & \text{if $j = 1$},\\ -\frac{1}{N}-\gamma\frac{2n}{N^2} & \textit{if $j > 1$ and $j-1$ is odd},\\ \frac{1}{N} + \gamma\frac{1}{N^2} & \text{otherwise}. \end{cases} \label{eqn:JacobianRow1} \end{equation} Letting $\omega_j$ for $j = 0,\dots,N-1$ be the $N^\text{th}$ roots of unity,\footnote{For details see \cite{D12}}, we know that the $j^{th}$ eigenvalue of $\bm{\mathcal{J}}$ is: \begin{displaymath} \sum_{k=1}^N\bm{\mathcal{J}}_{1,k}\omega_j^{k-1}. \end{displaymath} It now remains to show that the sign of the real-part of $\lambda_j$ is entirely dependent on $\gamma$. The real part of the eigenvalue is given by: \begin{equation} \mathrm{Re}\left(\lambda_j\right) =\sum_{k=1}^N \bm{\mathcal{J}}_{1,k}\cos\left(\frac{2\pi j(k-1)}{N}\right). \label{eqn:EigReal} \end{equation} The first eigenvalue ($j=0$) is real and readily computed: \begin{displaymath} \lambda_0 = n\left(\gamma\frac{1}{N^2} - \gamma\frac{2n}{N^2} \right) - \gamma\frac{2n}{N^2} = -\gamma\frac{n(1+2n)}{N^2}. \end{displaymath} It is clear at once that the sign of this eigenvalue is entirely controlled by the sign of $\gamma$. For $j > 0$, note that the periodicity of the cosine function (and the fact that the roots of unity are the vertices of the regular unit $N$-gon) implies that the coefficient of $\bm{\mathcal{J}}_{1,k}$ is identical to the coefficient of $\bm{\mathcal{J}}_{1,N-(k-2)}$ if $2 \leq k \leq n+1$. From this fact and the Expression \ref{eqn:JacobianRow1}, the sum in Equation \ref{eqn:EigReal} becomes: \begin{displaymath} \mathrm{Re}\left(\lambda_j\right) = \gamma\left(-\frac{2n}{N^2}+\sum_{k=2}^{n+1}\cos\left(\frac{2\pi j(k-1)}{N}\right)\left(\frac{1}{N^2}-\frac{2n}{N^2} \right)\right). \end{displaymath} Factoring further we see: \begin{multline*} \mathrm{Re}\left(\lambda_j\right) = \frac{\gamma}{N^2}\left(-2n + \sum_{k=2}^{n+1}\cos\left(\frac{2\pi j(k-1)}{N}\right)\left(1-2n\right) \right) = \\ \frac{\gamma}{N^2}\left(-2n + \left(1-2n\right)\left( \sum_{k=2}^{n+1}\cos\left(\frac{2\pi j(k-1)}{N}\right) \right) \right). \end{multline*} The roots of unity are evenly distributed on the vertices of the unit $N$-gon in $\mathbb{C}$ and therefore the sum of the real parts must be zero. It follows that: \begin{displaymath} \sum_{k=2}^{n+1}\cos\left(\frac{2\pi j(k-1)}{N}\right) = -\frac{1}{2}. \end{displaymath} We now obtain an exact value for the real parts of the eigenvalues: \begin{equation} \mathrm{Re}\left(\lambda_j\right) = \frac{\gamma}{N^2}\left(-2n - \frac{1}{2}\left(1-2n\right)\right) = -\frac{\gamma}{N^2}\left(n+\frac{1}{2}\right). \end{equation} Thus we have proved that when $\gamma > 0$, then $\mathrm{Re}(\lambda_j) < 0$ for all $j$ and if $\gamma < 0$, then $\mathrm{Re}(\lambda_j) > 0$. The asymptotic stability (resp. instability) of the fixed point follows immediately. \end{proof} We illustrate the attractive interior point of the five and seven strategy cyclic games by projecting them into regular $N$-gons ($N = 5,7$), shown in Figure \ref{fig:Projection}. A pure strategy corresponds to a vertex, while a mixed strategy is located in the interior. Note, these are not true trajectories as would be seen in a classic ternary plot on the three rock-paper-scissors, but they are similarly representative. \begin{figure}[htbp] \centering \subfigure[Five Strategies]{\includegraphics[scale=0.5]{Figures/DrawKinkyPlot5Final.pdf}} \subfigure[Seven Strategies]{\includegraphics[scale=0.5]{Figures/DrawKinkyPlot7Final.pdf}} \caption{A figure illustrating the attraction of the unique interior fixed point of the cyclic game when $\gamma = 1$ assuming $5$ and $7$ strategies. The trajectories are projected into regular $N$-gons ($N=5,7$). The closer a trajectory is to a vertex, the closer it is to that pure strategy.} \label{fig:Projection} \end{figure} \section{The Control Problem and Some General Results}\label{sec:GeneralControl} We now state our control problem of interest: \begin{equation} \mathcal{C}_N = \left\{ \begin{aligned} \min \;\; & \int_0^{t_f}\;\frac{1}{2}\norm{\mathbf{u} - \mathbf{u}^*}^2 + \frac{r}{2}\gamma^2\;\;dt\\ s.t.\;\; &\dot{\mathbf{u}} = \mathbf{F}(\mathbf{u}) + \gamma\mathbf{G}(\mathbf{u}),\\ &\mathbf{u}(0) = \mathbf{u}_0. \end{aligned}\right. \label{eqn:MainControl} \end{equation} where $\mathbf{u}^* = \tfrac{1}{N}\mathbf{1}$. Such a problem arises naturally if we consider species interacting in a cyclic manner and $\gamma$ is a costly control mechanism by which an external manager may control species populations. In \cite{GB17}, $\gamma$ arises naturally as a tax in a public-goods game, which is shown to be diffeomorphic to a three stragegy cyclic game. As in \cite{GB17}, we will show that a quasi-linearization of this control problem has special structure. In this case, however, we show this special structure holds for all cyclic games (i.e., for all $N = 2n+1$). Furthermore, we discuss the limiting behavior of the control as $N$ grows large. To do this, we first consider a very general optimal control problem and obtain necessary conditions for simplifying the Euler-Lagrange necessary conditions. We then use these simplifications to generalize the results in \cite{GB17}. \subsection{Control Problems with One Control and Vanishing Lie Bracket} In the remainder of this section, the functions $\mathbf{F},\mathbf{G}:\mathbb{R}^n \rightarrow \mathbb{R}^n$ are arbitrary smooth functions, rather than the functions specific to the replicator dynamics for cyclic games given in Equations \ref{eqn:Fz} and \ref{eqn:Gz}, $\mathbf{x} \in \mathbb{R}^n$ is a state vector, and $\gamma$ is the control function to be determined. Consider the general optimal control problem with form: \begin{equation} \left\{ \begin{aligned} \min\;\; & \Psi(\mathbf{x}(t_f)) + \int_0^{t_f} F_0(\mathbf{x}) + \gamma G_0(\mathbf{x}) + \frac{r}{2} \gamma^2 \;\; dt\\ s.t.\;\; & \dot{\mathbf{x}} = \mathbf{F}(\mathbf{x}) + \gamma\mathbf{G}(\mathbf{x}),\\ & \mathbf{x}(0) = \mathbf{x}_0. \end{aligned} \right. \label{eqn:main} \end{equation} The functions $F_0,G_0:\mathbb{R}^n \rightarrow \mathbb{R}$ are smooth. Let $r > 0$, $t_f$ be the terminal time, and $F_0(\mathbf{x})$ be convex. Expression \ref{eqn:MainControl} has this structure, so we are simply considering a more general case of our problem of interest. The Euler-Lagrange necessary conditions for control are simple to derive for this problem and have an almost linear behavior. Note the Hamiltonian is: \begin{equation} \mathcal{H}(\mathbf{x},\gamma,\bm{\lambda}) = F_0(\mathbf{x}) +\gamma G_0(\mathbf{x}) + \frac{r}{2}\gamma^2 + \bm{\lambda}^T\mathbf{F}(\mathbf{x}) + \gamma \bm{\lambda}^T\mathbf{G}(\mathbf{x}). \label{eqn:Hamiltonian} \end{equation} The Hamiltonian is (strictly) convex in the control $\gamma$, and thus we propose the following: \begin{lemma} Any solution $\gamma^*$ to $\mathcal{H}_\gamma = 0$ satisfies the necessary conditions: \begin{enumerate} \item $\mathcal{H}_\gamma = 0$, and \item $\mathcal{H}_{\gamma\gamma} > 0$, the strong Legendre-Clebsch condition; \end{enumerate} therefore, it minimizes the Hamiltonian at all times. \hfill\qed \label{lem:NecessaryControl} \end{lemma} Deriving the optimal control by solving $\partial \mathcal{H}/\partial{\gamma} = 0$ for $u$ to obtain: \begin{equation} \gamma^* = -\frac{1}{r}\left(\bm{\lambda}^T\mathbf{G}(\mathbf{x})+G_0(\mathbf{x})\right). \label{eqn:ClosedFormCoState} \end{equation} The two conditions in Lemma \ref{lem:NecessaryControl}, along with the fact that $\mathbf{x}^*$ and $\lambda^*$ solve the resulting Euler-Lagrange two-point boundary value problem (see Expression \ref{eqn:EulerLagrange}), form the complete set of necessary conditions for the optimal control problem. Adding in the additional requirement that the corresponding matrix Riccati equation is bounded on $[0,t_f]$, these form sufficient conditions for a weak local minimal optimal controller \cite{BY65,J70}. We discuss this sufficient condition in the sequel. For simplicity, we refer to the optimal control as $\gamma$ (rather than $\gamma^*$) in the remainder of this paper and assume it is given by Equation \ref{eqn:ClosedFormCoState}. The adjoint dynamics are: \begin{equation} \dot{\bm{\lambda}}^T = -(\nabla_\mathbf{x}F_0(\mathbf{x}))^T - \gamma(\nabla_\mathbf{x}G_0(\mathbf{x}))^T - \bm{\lambda}^TD_\mathbf{x}\mathbf{F} - \gamma \bm{\lambda}^TD_\mathbf{x}\mathbf{G}, \label{eqn:costate} \end{equation} where $D_\mathbf{x}\mathbf{F}$ is the Jacobian (with respect to $\mathbf{x}$). Thus we have the Euler-Lagrange two-point boundary value problem: \begin{equation} \left\{ \begin{aligned} \dot{\mathbf{x}} &= \mathbf{F}(\mathbf{x}) + \gamma\mathbf{G}(\mathbf{x}),\\ \dot{\bm{\lambda}} &= -\nabla_\mathbf{x}F_0(\mathbf{x})-\gamma\nabla_\mathbf{x}G_0(\mathbf{x}) -(D_\mathbf{x}\mathbf{F})^T\bm{\lambda} - u(D_\mathbf{x}\mathbf{G})^T\bm{\lambda},\\ \mathbf{x}(0) &= \mathbf{x}_0,\\ \bm{\lambda}(t_f) &= \nabla_\mathbf{x}\Psi(\mathbf{x}[t_f]). \qquad \text{(Transverality Condition)} \end{aligned} \right. \label{eqn:EulerLagrange} \end{equation} \begin{proposition} If $u^*$ is an optimal control, then: \begin{equation} \gamma(t_f) = -\frac{1}{r}\left(\nabla_{\mathbf{x}}\Psi(\mathbf{x}(t_f))^T\mathbf{G}(\mathbf{x}(t_f)) + G_0(\mathbf{x}(t_f)) \right). \end{equation} \end{proposition} \begin{proof} This follows from the transversality condition. \end{proof} From Equation \ref{eqn:ClosedFormCoState}, note that: \begin{equation} r\dot{\gamma} = -\dot{\bm{\lambda}}^T\mathbf{G}(\mathbf{x}) - \bm{\lambda}^T(D_\mathbf{x}\mathbf{G})\dot{\mathbf{x}} - (\nabla_\mathbf{x}G_0)\dot{\mathbf{x}}. \end{equation} Then: \begin{multline} r\dot{\gamma} = \left((\nabla_\mathbf{x}F_0(\mathbf{x}))^T + \gamma(\nabla_\mathbf{x}G_0(\mathbf{x}))^T+ \bm{\lambda}^T(D_\mathbf{x}\mathbf{F}) + \gamma \bm{\lambda}^T(D_\mathbf{x}\mathbf{G})\right)\mathbf{G}(\mathbf{x}) - \\ \bm{\lambda}^T(D_\mathbf{x}\mathbf{G})\left(\mathbf{F}(\mathbf{x}) + \gamma\mathbf{G}(\mathbf{x}) \right) - (\nabla_\mathbf{x}G_0)\left(\mathbf{F}(\mathbf{x}) + \gamma\mathbf{G}(\mathbf{x}) \right). \end{multline} Simplifying we have: \begin{displaymath} r\dot{\gamma} = (\nabla_\mathbf{x}F_0(\mathbf{x}))^T\mathbf{G}(\mathbf{x}) - (\nabla_\mathbf{x}G_0(\mathbf{x}))^T\mathbf{F}(\mathbf{x}) + \bm{\lambda}^T\left((D_\mathbf{x}\mathbf{F})\mathbf{G}(\mathbf{x}) - (D_\mathbf{x}\mathbf{G})\mathbf{F}(x)\right). \end{displaymath} If the Lie Bracket vanishes, i.e.,: \begin{equation} [\mathbf{F},\mathbf{G}]=(D_\mathbf{x}\mathbf{F})\mathbf{G} - (D_\mathbf{x}\mathbf{G})\mathbf{F} = \mathbf{0}, \label{eqn:LieBracketVanish} \end{equation} then this simplifies to: \begin{equation} \dot{\gamma} = \frac{1}{r}\left( (\nabla_\mathbf{x}F_0(\mathbf{x}))^T\mathbf{G}(\mathbf{x}) - (\nabla_\mathbf{x}G_0(\mathbf{x}))^T\mathbf{F}(\mathbf{x}) \right), \end{equation} and all co-state variables are eliminated. We have shown the following: \begin{theorem} Consider the general optimal control problem given in Expression \ref{eqn:main}. If $[\mathbf{F},\mathbf{G}] = \mathbf{0}$ and $\gamma$ is an optimal control, then the pair $(\mathbf{x}(\gamma), \gamma)$ is a solution of the two point boundary value problem: \begin{equation} \left\{ \begin{aligned} \dot{\mathbf{x}} = &\mathbf{F}(\mathbf{x}) + \gamma\mathbf{G}(\mathbf{x}),\\ \dot{\gamma} = & \frac{1}{r}\left( (\nabla_\mathbf{x}F_0(\mathbf{x}))^T\mathbf{G}(\mathbf{x}) - (\nabla_\mathbf{x}G_0(\mathbf{x}))^T\mathbf{F}(\mathbf{x}) \right),\\ \mathbf{x}(0) =& \mathbf{x}_0,\\ \gamma(t_f) = & -\frac{1}{r}\left(\nabla_{\mathbf{x}}\Psi(\mathbf{x}(t_f))^T\mathbf{G}(\mathbf{x}(t_f)) + G_0(\mathbf{x}(t_f)) \right). \end{aligned}\right. \label{eqn:VectorThm} \end{equation}\hfill\qed \label{thm:main} \end{theorem} Geometrically, Equation \ref{eqn:LieBracketVanish} implies that the flows derived by the vector fields $\mathbf{F}$ and $\mathbf{G}$ commute locally. From a game-theoretic view, this means that locally evolutionary motion caused by competition in uncontrolled game commutes with evolutionary motion caused by the actuation payoffs on local space/time scales. As we see in the sequel, this is not true for actuated cyclic games, but is true for quasi-linear approximations of the evolutionary dynamics as in \cite{GB17}. It is worth noting that a differential equation for the control function is derived in \cite{B78}, without the assumption of the vanishing Lie Bracket. However, without this assumption the system does not simplify in as useful a way and, in fact, in \cite{B78} the relevant Lie Bracket is not considered. Note, in formulating Theorem \ref{thm:main}, we are assuming that solving the Euler-Lagrange equations will yield an optimal control. We can use the well known fact that a sufficient condition for optimality is the boundedness of the solution to the matrix Ricatti equation \cite{BY65,J70} to derive a complete necessary and sufficient condition for optimality of the control. Let: \begin{displaymath} \dot{\mathbf{x}} = \mathbf{F}(\mathbf{x}) + \gamma\mathbf{G}(\mathbf{x}) = \mathbf{f}(\mathbf{x},\gamma). \end{displaymath} Then the Matrix Ricatti equation is: \begin{equation} \left\{ \begin{aligned} -\dot{\mathbf{S}} &= D_\mathbf{xx}\mathcal{H} + (D_\mathbf{x}\mathbf{f})^T\mathbf{S} + \mathbf{S}(D_\mathbf{x}\mathbf{f}) - \\ & \hspace*{4em}\frac{1}{r}\left(D_{\gamma\mathbf{x}}\mathcal{H} + (\partial_\gamma\mathbf{f})^T\mathbf{S}\right)^T\left(D_{\gamma\mathbf{x}}\mathcal{H} + (\partial_\gamma\mathbf{f})^T\mathbf{S}\right),\\ \mathbf{S}(t_f) &= \nabla^2_\mathbf{x}\Psi(\mathbf{x}[t_f]). \end{aligned}\right. \label{eqn:matrixricatti} \end{equation} Here $D_\mathbf{xx}$ is the second differential operator with respect to the state and $\partial_\gamma$ is an ordinary partial derivative, since there is only one control variable. When taken together with Lemma \ref{lem:NecessaryControl}, the system of differential equations in Theorem \ref{thm:main} and the co-state dynamics, Equation \ref{eqn:costate}, we have a complete characterization of the necessary and sufficient conditions for the optimal control. This yields the corollary: \begin{corollary}[Corollary to Theorem \ref{thm:main}] Let $\mathbf{f}(\mathbf{x},\gamma) = \mathbf{F}(\mathbf{x}) + \gamma\mathbf{G}(\mathbf{x})$ in the optimal control problem in Expression \ref{eqn:main}, with Hamiltonian $\mathbf{H}(\mathbf{x},\gamma,\bm{\lambda})$. Any solution to the system of differential equations: \begin{equation} \left\{ \begin{aligned} \dot{\mathbf{x}} &= \mathbf{f}(\mathbf{x},\gamma),\\ \dot{\gamma} &= \frac{1}{r}\left( (\nabla_\mathbf{x}F_0(\mathbf{x}))^T\mathbf{G}(\mathbf{x}) - (\nabla_\mathbf{x}G_0(\mathbf{x}))^T\mathbf{F}(\mathbf{x}) \right),\\ \dot{\bm{\lambda}} &= -\nabla_\mathbf{x}F_0(\mathbf{x})-\gamma\nabla_\mathbf{x}G_0(\mathbf{x}) -(D_\mathbf{x}\mathbf{F})^T\bm{\lambda} - u(D_\mathbf{x}\mathbf{G})^T\bm{\lambda},\\ -\dot{\mathbf{S}} &= D_\mathbf{xx}\mathcal{H} + (D_\mathbf{x}\mathbf{f})^T\mathbf{S} + \mathbf{S}(D_\mathbf{x}\mathbf{f}) - \\ & \hspace*{4em}\frac{1}{r}\left(D_{\gamma\mathbf{x}}\mathcal{H} + (\partial_\gamma\mathbf{f})^T\mathbf{S}\right)^T\left(D_{\gamma\mathbf{x}}\mathcal{H} + (\partial_\gamma\mathbf{f})^T\mathbf{S}\right),\\ \mathbf{x}(0) &= \mathbf{x}_0,\\ \gamma(t_f) &= -\frac{1}{r}\left(\nabla_{\mathbf{x}}\Psi(\mathbf{x}(t_f))^T\mathbf{G}(\mathbf{x}(t_f)) + G_0(\mathbf{x}(t_f)) \right),\\ \bm{\lambda}(t_f) &= \nabla_x\Psi(\mathbf{x}[t_f]),\\ \mathbf{S}(t_f) &= \nabla^2_\mathbf{x}\Psi(\mathbf{x}[t_f]), \end{aligned}\right. \end{equation} in which $\gamma(t) = -\tfrac{1}{r}\left(\bm{\lambda}^T\mathbf{G}[\mathbf{x}] + G_0[\mathbf{x}]\right)$ and $\mathbf{S}$ is bounded for all $t \in [0,t_f]$ constitutes a weak local optimal solution for Expression \ref{eqn:main}. \label{cor:main} \end{corollary} We note that this is the general analog of Proposition 2 in \cite{FG17}, which is specialized to a one-dimensional control problem. In general, checking the boundedness of the solution to the Matrix Ricatti equation must be done numerically. In the sequel we develop a simpler test for optimality using Mangasarian's sufficiency condition; i.e., by checking that the Hamiltonian is jointly convex. Problem \ref{eqn:MainControl} ($\mathcal{C}_N$) does not satisfy the necessary condition that $[\mathbf{F},\mathbf{G}] = \mathbf{0}$. However, a quasi-linearization of the problem does satisfy this condition (as in \cite{GB17}). We now discuss a special case of Theorem \ref{thm:main} as well as extensions that apply to this quasi-linearized form. \subsection{The Quasi-Linear Case} \label{sec:linear} For the remainder of this section, let $\mathbf{J}$ and $\mathbf{H}$ be arbitrary matrices of appropriate size, rather than the Jacobian matrices derived in Lemmas \ref{lem:DF} and \ref{lem:DG}. In Problem \ref{eqn:main}, let: \begin{equation} \left\{ \begin{aligned} F_0(\mathbf{x}) &= \frac{1}{2}\mathbf{x}^T\mathbf{Q}\mathbf{x},\\ G_0(\mathbf{x}) &= 0,\\ \mathbf{F}(\mathbf{x}) &= \mathbf{J}\mathbf{x},\\ \mathbf{G}(\mathbf{x}) &= \mathbf{H}\mathbf{x}. \end{aligned} \right. \label{eqn:LinearConditions} \end{equation} where $\mathbf{Q}$ is a (symmetric) positive definite matrix of appropriate size. We will add additional criteria to $\mathbf{J}$ and $\mathbf{H}$ as we proceed. We refer to this as a quasi-linear case because the only non-linearity arises from the interaction of the state and control variables. The following Corollary is immediate from Theorem \ref{thm:main}: \begin{corollary} If $\mathbf{J}\mathbf{H} = \mathbf{H}\mathbf{J}$ and $\gamma^*$ is an optimal control, then the pair $(\mathbf{x}(\gamma^*), \gamma^*)$ is a solution of the two point boundary value problem: \begin{equation} \left\{ \begin{aligned} \dot{\mathbf{x}} = &\mathbf{J}\mathbf{x} + \gamma\mathbf{H}\mathbf{x},\\ \dot{\gamma} = & \frac{1}{r}\mathbf{x}^T\mathbf{Q}\mathbf{H}\mathbf{x},\\ \mathbf{x}(0) =& \mathbf{x}_0,\\ \gamma(t_f) = & -\frac{1}{r}\nabla_{\mathbf{x}}\Psi(\mathbf{x}(t_f))^T\mathbf{H}\mathbf{x}(t_f). \end{aligned}\right. \end{equation} \hfill\qed \label{cor:LinRes1} \end{corollary} The condition that $\mathbf{J}$ and $\mathbf{H}$ commute is exactly the statement that the Lie Bracket of the vector fields in the dynamics vanishes. Therefore, Theorem \ref{thm:main} can be applied to any linear quadratic control problem where the state equation satisfies this condition. We now derive some special results on $\ddot{\gamma}$ and the optimal control in this quasi-linear case. Let $\mathbf{K} \triangleq \mathbf{Q}\mathbf{H}$ and assume $\mathbf{J}\mathbf{H} = \mathbf{H}\mathbf{J}$. Computing the second derivative of $\gamma$ yields: \begin{multline*} r \ddot{\gamma} = \dot{\mathbf{x}}^T\mathbf{K}\mathbf{x} + \mathbf{x}^T\mathbf{K}\dot{\mathbf{x}} = \left(\mathbf{x}^T\mathbf{J}^T + \gamma\mathbf{x}^T\mathbf{H}^T\right)\mathbf{K}\mathbf{x} + \mathbf{x}^T\mathbf{K}\left(\mathbf{J}\mathbf{x} + \gamma \mathbf{H}\mathbf{x}\right) = \\ \mathbf{x}^T\left(\mathbf{J}^T\mathbf{K} + \mathbf{K}\mathbf{J} + \gamma\left( \mathbf{H}^T\mathbf{K} + \mathbf{K}\mathbf{H} \right) \right)\mathbf{x}. \end{multline*} To simplify this, we will add an additional assumption to $\mathbf{J}$; suppose that $\mathbf{J}^T = -\mathbf{J}$ (i.e., $\mathbf{J}$ is skew-symmetric) and $\mathbf{J}\mathbf{K} = \mathbf{K}\mathbf{J}$. Then: \begin{equation} r \frac{\ddot{\gamma}}{\gamma} = \mathbf{x}^T\left( \mathbf{H}^T\mathbf{K} + \mathbf{K}\mathbf{H} \right)\mathbf{x}. \end{equation} Before proceeding note that: \begin{align*} \frac{d}{dt}\left(\mathbf{x}^T\mathbf{Q}\mathbf{x}\right) = & \left(\mathbf{x}^T\mathbf{J}^T + \gamma\mathbf{x}^T\mathbf{H}^T \right)\mathbf{Q}\mathbf{x} + \mathbf{x}^T\mathbf{Q}\left(\mathbf{J}\mathbf{x} + \gamma\mathbf{H}\mathbf{x}\right) = \\ &\mathbf{x}^T\left(-\mathbf{J}\mathbf{Q} + \mathbf{Q}\mathbf{J}\right)\mathbf{x} + \gamma\mathbf{x}^T\left(\mathbf{H}^T\mathbf{Q} + \mathbf{Q}\mathbf{H}\right)\mathbf{x} =\\ &\mathbf{x}^T\left(-\mathbf{J}\mathbf{Q} + \mathbf{Q}\mathbf{J}\right)\mathbf{x} + \gamma \mathbf{x}^T\left(\mathbf{H}^T\mathbf{Q}^T + \mathbf{Q}\mathbf{H}\right)\mathbf{x} = \\ &\mathbf{x}^T\left(-\mathbf{J}\mathbf{Q} + \mathbf{Q}\mathbf{J}\right)\mathbf{x} + 2 \gamma \mathbf{x}^T\mathbf{K}\mathbf{x} = \\ &\mathbf{x}^T\left(-\mathbf{J}\mathbf{Q} + \mathbf{Q}\mathbf{J}\right)\mathbf{x} + 2 r\gamma \dot{\gamma}. \end{align*} Thus, we have the following proposition and its corollary: \begin{proposition} If $\mathbf{J} = -\mathbf{J}^T$, $\mathbf{J}\mathbf{H} = \mathbf{H}\mathbf{J}$ and $\mathbf{J}\mathbf{Q} = \mathbf{Q}\mathbf{J}$ and $\mathbf{K} \triangleq \mathbf{Q}\mathbf{H}$, then: \begin{enumerate} \item \begin{equation}\nonumber \mathbf{J}\mathbf{K} = \mathbf{K}\mathbf{J}, \end{equation} \item \begin{equation} \frac{d}{dt}\left(\mathbf{x}^T\mathbf{Q}\mathbf{x}\right) = 2r\gamma \dot{\gamma}, \end{equation} \item \begin{equation} r \frac{\ddot{\gamma}}{\gamma} = \mathbf{x}^T\left(\mathbf{H}^T\mathbf{K} + \mathbf{K}\mathbf{H}\right)\mathbf{x}. \end{equation} \end{enumerate} \label{prop:VectorProp} \end{proposition} \begin{corollary} For some constant $C$, \begin{equation} \mathbf{x}^T\mathbf{Q}\mathbf{x} = r\gamma^2 + C \end{equation} is the implicit closed-loop control, where $C$ must satisfy: \begin{equation} C = \mathbf{x}^T(t_f)\mathbf{Q}\mathbf{x}(t_f) - r\left(\frac{1}{r}\nabla_{\mathbf{x}}\Psi(\mathbf{x}(t_f))^T\mathbf{H}\mathbf{x}(t_f) \right)^2. \label{eqn:C2} \end{equation} Furthermore the optimal control $\gamma$ exists at time $t$ just in case: \begin{equation} \mathbf{x}^T(t)\mathbf{Q}\mathbf{x}(t) - C \geq 0. \end{equation} \label{cor:xQx} \end{corollary} \section{Application to Complete Odd Circulant Games}\label{sec:CyclicControl} We now return to the study of cyclic games with $N$ strategies and specifically to the control problem in Expression \ref{eqn:MainControl}. As noted already, we cannot apply Theorem \ref{thm:main} directly to Problem \ref{eqn:MainControl} because the appropriate Lie Bracket does not vanish. However, we can construct the quasi-linearized form of the problem. Let $\mathbf{x} = \mathbf{u} - \mathbf{u}^*$. The quasi-linearized problem is: \begin{equation} \tilde{\mathcal{C}}_N = \left\{ \begin{aligned} \min \;\; & \int_0^{t_f}\;\frac{1}{2}\norm{\mathbf{x}}^2 + \frac{r}{2}\gamma^2\;\;dt\\ s.t.\;\; &\dot{\mathbf{x}} = \mathbf{J}\mathbf{x} + \gamma\mathbf{H}\mathbf{x},\\ &\mathbf{x}(0) = \mathbf{x}_0. \end{aligned}\right. \label{eqn:MainControlQuasi} \end{equation} In Expression \ref{eqn:MainControlQuasi}, $\mathbf{J}$ and $\mathbf{H}$ are the Jacobian matrices of $\mathbf{F}(\mathbf{u})$ and $\mathbf{G}(\mathbf{u})$ as defined in Lemmas \ref{lem:DF} and \ref{lem:DG}. Problem \ref{eqn:MainControlQuasi} is an instance of the general control problem studied in Section \ref{sec:GeneralControl}. The following useful fact follows at once from Lemma \ref{lem:DF}. \begin{corollary} The Jacobian matrix $\mathbf{J}$ is skew-symmetric. \hfill\qed \label{cor:SkewSymmetric} \end{corollary} \begin{lemma} Let $\gamma$ be the optimal control for Problem \ref{eqn:MainControl}. Then: \begin{enumerate} \item The (open-loop) optimal control obeys the following differential equations: \begin{align} \dot{\gamma} &= \frac{1}{r}\mathbf{x}^T\mathbf{H}\mathbf{x},\; \gamma(t_f) = 0, \label{eqn:dotgamma}\\ r\frac{\ddot{\gamma}}{\gamma} &= \mathbf{x}^T\left(\mathbf{H}^T\mathbf{H} + \mathbf{H}\mathbf{H}\right)\mathbf{x}. \label{eqn:ddotgamma1} \end{align} \item The following identity holds: \begin{equation} \mathbf{x}^T\mathbf{x} = \norm{x}^2 = r\gamma^2 + C, \label{eqn:closedloop} \end{equation} where: \begin{displaymath} C = \norm{\mathbf{x}(t_f)}^2. \end{displaymath} \end{enumerate} \label{lem:MainControl1} \end{lemma} \begin{proof} Problem \ref{eqn:MainControlQuasi} is an instance of Problem \ref{eqn:main}, but with quasi-linear system dynamics and quadratic objective as given in the quasi-linear conditions in Expression \ref{eqn:LinearConditions}. In particular, Problem \ref{eqn:MainControlQuasi} sets $\mathbf{Q} = \mathbf{I}_N$. As a consequence the matrix $\mathbf{K} = \mathbf{Q}\mathbf{H} = \mathbf{H}$. From Corollary \ref{cor:SkewSymmetric}, we know $\mathbf{J}$ is skew-symmetric. Further, since the circulant matrices form a commutative algebra, we have $\mathbf{H}\mathbf{J} = \mathbf{J}\mathbf{H}$. The lemma follows at once from Corollary \ref{cor:LinRes1}, Proposition \ref{prop:VectorProp} and Corollary \ref{cor:xQx}. \end{proof} Expression \ref{eqn:closedloop} is the closed-loop control law for the controlled cyclic game. Furthermore, Equation \ref{eqn:dotgamma} allows us to determine some structural properties of $\gamma$. \begin{proposition} The matrix $\mathbf{H}$ is negative definite and therefore $\dot{\gamma} \leq 0$ for all $t$. \label{prop:DotGammaSign} \end{proposition} \begin{proof} Consider any vector $\mathbf{x} \in \mathbb{R}^N$. Then: \begin{multline*} \mathbf{x}^T\mathbf{H}\mathbf{x} = \frac{1}{N^2}\mathbf{x}^T\left(N\left(\mathbf{M}_N-\mathbf{1}_N\right) + \mathbf{1}_N\right)\mathbf{x}^T = \\ \frac{1}{N^2}\left(N\left(\sum_{i=1}^N\sum_{j>i}x_i x_j - \sum_{i=1}^N x_i^2 -2\sum_{i=1}^N\sum_{j > i}x_ix_j \right) + \sum_{i=1}^N x_i^2 + 2\sum_{i=1}^N\sum_{j > i}x_ix_j\right) = \\ -\frac{N-1}{N^2}\sum_{i=1}^N x_i^2 - \frac{N-2}{N^2} \sum_{i=1}^N\sum_{j > i}x_ix_j. \end{multline*} Let $\mathbf{S} \in \mathbb{R}^{N \times N}$ be the upper-triangular matrix defined as: \begin{displaymath} \mathbf{S}_{ij} = \begin{cases} -\frac{N-1}{N^2} & \text{if $i = j$},\\ -\frac{N-2}{N^2} & \text{otherwise}. \end{cases} \end{displaymath} Then: \begin{displaymath} \mathbf{x}^T\mathbf{H}\mathbf{x} = \mathbf{x}^T\mathbf{S}\mathbf{x} = -\frac{N-1}{N^2}\sum_{i=1}^N x_i^2 - \frac{N-2}{N^2} \sum_{i=1}^N\sum_{j > i}x_ix_j. \end{displaymath} The leading principal minors of $\mathbf{S}$ alternate in sign (the diagonal is entirely negative) and thus by Sylvester's criterion, $\mathbf{S}$ is negative definite. It follows at once that $\mathbf{x}^T\mathbf{H}\mathbf{x} < 0$ for all $\mathbf{x} \neq \mathbf{0}$ and thus $\mathbf{H}$ is negative definite. The fact that $\dot{\gamma} < 0$ now follows from Lemma \ref{lem:MainControl1}. \end{proof} \begin{corollary} The optimal control $\gamma$ is a decreasing function on $[0,t_f]$. \end{corollary} In addition to determining that $\gamma$ is decreasing, Problem \ref{eqn:MainControlQuasi} has further special structure, which allows us to understand the structure of the derived control $\gamma$ in greater detail and ultimately derive a closed-form approximation for large $N$. The derivation is similar to the one found in \cite{GB17} for a special case diffeomorphic to rock-paper-scissors. \begin{lemma} Let $\mathbf{H} = D_\mathbf{u}\mathbf{G}(\mathbf{u})$. Then for all $\mathbf{x} \in \mathbb{R}^N$: \begin{equation} \mathbf{x}^T\left(\mathbf{H}^T\mathbf{H} + \mathbf{H}\mathbf{H} + \mathbf{H}\right)\mathbf{x} = -\frac{n}{N^2}\left\lVert\mathbf{x}\right\rVert^2. \end{equation} \label{lem:5H} \end{lemma} \begin{proof} From Lemma \ref{lem:DG} we have: \begin{displaymath} \mathbf{H} = \frac{1}{N^2}\left(N\left(\mathbf{M}_N - \mathbf{1}_N\right) + \mathbf{1}_N\right). \end{displaymath} Let $\mathbf{R} = \mathbf{M}_N - \mathbf{1}_N$. The following computations are straight forward: \begin{align*} \mathbf{H}^T\mathbf{H} &= \frac{1}{N^4}\left(N^2\mathbf{R}^T\mathbf{R} + N\mathbf{R}^T\mathbf{1}_N + N\mathbf{1}^T_N\mathbf{R} + \mathbf{1}_N^T\mathbf{1}_N \right),\\ \mathbf{H}\mathbf{H} &= \frac{1}{N^4}\left(N^2\mathbf{R}\mathbf{R} + N\mathbf{R}\mathbf{1}_N + N\mathbf{1}_N\mathbf{R} + \mathbf{1}_N\mathbf{1}_N \right). \end{align*} Note that: \begin{displaymath} \mathbf{1}^T_N\mathbf{1}_N = \mathbf{1}_N\mathbf{1}_N = N\mathbf{1}_N. \end{displaymath} We may also compute: \begin{displaymath} \mathbf{1}_N\mathbf{M}_N = \mathbf{1}^T_N\mathbf{M}_N = n\mathbf{1}_N, \end{displaymath} because $M_N$ contains $n$ unit entries in each column (row). Consequently: \begin{displaymath} \left(\mathbf{M}_N^T\mathbf{1}_N\right)^T = \mathbf{1}^T\mathbf{M}_N = n\mathbf{1}_N, \end{displaymath} and therefore: \begin{displaymath} \mathbf{M}_N^T\mathbf{1}_N = n\mathbf{1}_N^T=n\mathbf{1}_N. \end{displaymath} Using this information, we compute: \begin{equation} \mathbf{R}^T\mathbf{1}_N = \mathbf{1}_N^T\mathbf{R} = \mathbf{1}_N\mathbf{R} = \mathbf{R}\mathbf{1}_N = -(n+1)\mathbf{1}_N. \label{eqn:R1} \end{equation} Thus: \begin{multline*} \mathbf{H}^T\mathbf{H} + \mathbf{H}\mathbf{H} = \frac{1}{N^4}\left( N^2\left(\mathbf{R}^T + \mathbf{R}\right)\mathbf{R} -4N(n+1)\mathbf{1}_N + 2N\mathbf{1}_N\right) = \\ \frac{1}{N^4}\left(N^2\left(\mathbf{R}^T + \mathbf{R}\right)\mathbf{R} - 2N\left(2(n+1)-1\right)\mathbf{1}_N \right) = \\ \frac{1}{N^2}\left(N^2\left(\mathbf{R}^T + \mathbf{R}\right)\mathbf{R} -2N^2\mathbf{1}_N\right) = \frac{1}{N^2}\left(\left(\mathbf{R}^T + \mathbf{R}\right)\mathbf{R} - 2\mathbf{1}_N\right). \end{multline*} Using the fact that $\mathbf{H} = (N\mathbf{R} + \mathbf{1}_N)/N^2$, we may write: \begin{displaymath} \mathbf{H}^T\mathbf{H} + \mathbf{H}\mathbf{H} + \mathbf{H} = \frac{1}{N^2}\left( \left(\mathbf{R}^T + \mathbf{R} + N\mathbf{I}_N\right)\mathbf{R} - \mathbf{1}_N \right). \end{displaymath} The circulant structure of $\mathbf{M}$ implies the identity: \begin{displaymath} \mathbf{M}^T + \mathbf{M} = \mathbf{1}_N - \mathbf{I}_N. \end{displaymath} Therefore: \begin{displaymath} \mathbf{R}^T + \mathbf{R} = \mathbf{1}_N - \mathbf{I}_N - 2\mathbf{1}_N = -\mathbf{1}_N - \mathbf{I}_N. \end{displaymath} Thus, using Equation \ref{eqn:R1} and the fact that $N - 1 = 2n$: \begin{multline*} \mathbf{H}^T\mathbf{H} + \mathbf{H}\mathbf{H} + \mathbf{H} = \frac{1}{N^2}\left(\left((N-1)\mathbf{I}_N - \mathbf{1}_N \right)\mathbf{R} - \mathbf{1}_N \right) = \\ \frac{1}{N^2}\left((N-1)(\mathbf{M}_N - \mathbf{1}_N) + (n+1)\mathbf{1}_N - \mathbf{1}_N \right) = \\ \frac{1}{N^2}\left(2n\mathbf{M}_N + -n\mathbf{1}_N\right) = \frac{n}{N^2}\left(2\mathbf{M}_N - \mathbf{1}_N\right). \end{multline*} Recall from Corollary \ref{cor:uu} that: \begin{displaymath} \mathbf{x}^T\mathbf{M}_N\mathbf{x} = \sum_{i=1}^N\sum_{j > i}x_ix_j. \end{displaymath} Furthermore, it is straight forward to compute: \begin{displaymath} \mathbf{x}^T\mathbf{1}_N\mathbf{x} = \sum_{i=1}^N x_i^2 + 2\sum_{i=1}^N\sum_{j > i}x_ix_j. \end{displaymath} Therefore: \begin{multline*} \mathbf{x}^T\left( \mathbf{H}^T\mathbf{H} + \mathbf{H}\mathbf{H} + \mathbf{H} \right)\mathbf{x} = \frac{n}{N^2}\mathbf{x}^T\left(2\mathbf{M}_N - \mathbf{1}_N\right)\mathbf{x} = \\ \frac{n}{N^2}\left(2\sum_{i=1}^N\sum_{j > i}x_ix_j - \sum_{i=1}^N x_i^2 - 2\sum_{i=1}^N\sum_{j > i}x_ix_j\right) = -\frac{n}{N^2}\left\lVert\mathbf{x}\right\rVert^2. \end{multline*} This completes the proof. \end{proof} \begin{theorem} If $\gamma$ is the open-loop optimal control for Problem \ref{eqn:MainControlQuasi}, then $\gamma$ satisfies the following second order differential equation: \begin{equation} \left\{ \begin{aligned} r\ddot{\gamma} + r\gamma\dot{\gamma} + \frac{n}{N^2}\gamma\left(r\gamma^2 + C\right) &= 0,\\ \gamma(t_f) &= 0,\\ \gamma'(0) &= \frac{1}{r}\mathbf{x}_0^T\mathbf{H}\mathbf{x}_0,\\ C &= \norm{x(t_f)}^2. \end{aligned} \right. \label{eqn:2ODE-Full} \end{equation} \end{theorem} \begin{proof} From Lemma \ref{lem:MainControl1} we have: \begin{displaymath} r\frac{\ddot{\gamma}}{\gamma} = \mathbf{x}^T\left(\mathbf{H}^T\mathbf{H} + \mathbf{H}\mathbf{H}\right)\mathbf{x}, \end{displaymath} and \begin{displaymath} r\dot{\gamma} = \mathbf{x}^T\mathbf{H}\mathbf{x}. \end{displaymath} Adding these together we obtain: \begin{displaymath} r\frac{\ddot{\gamma}}{\gamma} + r\dot{\gamma} = \mathbf{x}^T\left(\mathbf{H}^T\mathbf{H} + \mathbf{H}\mathbf{H} + \mathbf{H}\right)\mathbf{x}. \end{displaymath} Therefore by Lemma \ref{lem:5H}: \begin{displaymath} r\frac{\ddot{\gamma}}{\gamma} + \dot{\gamma} = -\frac{n}{N^2}\norm{x(t)}^2. \end{displaymath} From Lemma \ref{lem:MainControl1}, we have: \begin{displaymath} r\frac{\ddot{\gamma}}{\gamma} + \dot{\gamma} = -\frac{n}{N^2}\left(r\gamma^2 + C\right), \end{displaymath} where $C = \norm{\mathbf{x}(t_f)}^2$. Thus: \begin{displaymath} r\ddot{\gamma} + \gamma\dot{\gamma} + \frac{n}{N^2}\gamma\left(r\gamma^2 + C\right) = 0. \end{displaymath} The boundary conditions $\gamma(0) = 0$ and $\gamma'(0) = \mathbf{x}_0^T\mathbf{H}\mathbf{x}_0$ follows from Lemma \ref{lem:MainControl1}. \end{proof} \begin{corollary} For $N$ large, the open loop control $\gamma$ can be approximated by $\zeta$, a solution to the following two-point boundary value problem: \begin{equation} \left\{ \begin{aligned} r\ddot{\zeta} + r\zeta\dot{\zeta} = 0,\\ \zeta(t_f) &= 0,\\ \zeta'(0) &= \frac{1}{r}\mathbf{x}_0^T\mathbf{H}\mathbf{x}_0. \end{aligned} \right. \label{eqn:2ODE} \end{equation} \label{cor:2ODE} \end{corollary} In practice, we will show that $N \geq 5$ is sufficient for this approximation to be valid. \subsection{Closed Form Analysis of the Limiting Behavior} For simplicity, let $r = 1$ in the following analysis. Corollary \ref{cor:2ODE} can be made more useful by re-writing Equation \ref{eqn:2ODE} as a system of first order differential equations and examining the phase portrait (see Fig. \ref{fig:PhasePortrait}): \begin{equation} \begin{aligned} \dot{\zeta} &= v,\\ \dot{v} &= -\zeta v,\\ \zeta(t_f) &= 0,\\ v(0) &= \mathbf{x}_0^T\mathbf{H}\mathbf{x}_0. \end{aligned} \label{eqn:SysODE} \end{equation} \begin{figure}[htbp] \centering \includegraphics[scale=0.75]{Figures/StreamPlot.pdf} \caption{The phase portrait of the first order system representing the limiting behavior of the open loop control $\gamma$ and it's first derivative.} \label{fig:PhasePortrait} \end{figure} The phase portrait indicates a sharp behavioral change in the direction field when moving from the $v < 0$ half-plane to the $v > 0$ half-plane. For the half-plane where $v < 0$, $\gamma \geq 0$ necessarily by Theorem \ref{thm:GammaSign}. This is consistent with Proposition \ref{prop:DotGammaSign}. System \ref{eqn:SysODE} has a closed form solution\footnote{Derived with \textit{Mathematica}$^\text{\textsc{TM}}$.} with several branches. The relevant solutions on the interval $[0,t_f]$ are: \begin{align*} \zeta(t) &= \sqrt{2} \sqrt{\kappa \tanh ^2\left(\frac{\sqrt{\kappa } \left(t_f-t\right)}{\sqrt{2}}\right)},\\ \zeta'(t) &= -2 \sqrt{\kappa } \sqrt{\kappa \tanh ^2\left(\frac{\sqrt{\kappa } \left(t_f-t\right)}{\sqrt{2}}\right)} \csch\left(\sqrt{2} \sqrt{\kappa } \left(t_f-t\right)\right). \end{align*} Note, $\zeta \sim O(|\tanh(\cdot)|)$. Thus, the usual behavior of $\tanh$ is modified so that $\zeta$ is a decreasing function on $[0,t_f]$ and then an increasing function outside this range. As a consequence, this solution is only valid on the control domain of interest. In the closed form solution, $\kappa$ is a constant of integration that must be chosen so that $v(0) = \zeta'(0) = \mathbf{x}_0^T\mathbf{H}\mathbf{x}_0$. Finding a closed form expression for $\kappa$ is difficult. However as $N$ increases, $\mathbf{x}_0^T\mathbf{H}\mathbf{x}_0$ decreases in size because $\mathbf{H} \sim 1/N$ and $\norm{\mathbf{x}}$ is bounded, since $\mathbf{x}$ is just a translation of $\mathbf{u} \in \Delta_N$. For small values of $\mathbf{x}_0^T\mathbf{H}\mathbf{x}_0$, we expect $\kappa$ to be small because of the structure of $\zeta'(t)$. Furthermore, $\zeta'(0)$ can be approximated as: \begin{displaymath} \gamma'(0) \approx -\kappa + O(\kappa^2). \end{displaymath} Thus, setting $\kappa = -\mathbf{x}_0^T\mathbf{H}\mathbf{x}_0$ will give a reasonable approximation of the solution, as we illustrate in the examples below. \subsection{Sufficiency of the Euler-Lagrange Conditions}\label{sec:Sufficiency} Corollary \ref{cor:main} contains both necessary and sufficient conditions for the computed $\gamma(t)$ to be the optimal control. However, these conditions require the solution of the matrix Riccati equation. For the quasi-linearized optimal control problem on cyclic games, a simpler test can be constructed using Mangaserian's condition, which states that the Hessian of the Hamiltonian must be positive definite (i.e., jointly convex in state and control). For the optimal control in quasi-linearized cyclic games, the Hamiltonian of this optimal control problem is: \begin{displaymath} \mathcal{H}(\mathbf{x},\gamma,\bm{\lambda}) = \norm{\mathbf{x}}^2 + \frac{1}{r}\gamma^2 + \bm{\lambda}^T\mathbf{J}\mathbf{x} + \gamma\bm{\lambda}^T\mathbf{H}\mathbf{x}. \end{displaymath} This is a specialization of Equation \ref{eqn:Hamiltonian} to the quasi-linearized cyclic games problem. The Hessian of $\mathcal{H}$ is: \begin{displaymath} \mathfrak{H} = \begin{bmatrix} \mathbf{I}_N & \mathbf{H}^T\bm{\lambda}^T\\ \bm{\lambda}\mathbf{H}& r \end{bmatrix}. \end{displaymath} Here, $\bm{\lambda}$ is the co-state for the optimal control problem. \begin{theorem} If $r > \norm{\mathbf{H}^T\bm{\lambda}^T}^2$ for all $t \in [0,t_f]$, then the control derived in Lemma \ref{lem:MainControl1} is optimal. \label{thm:Sufficient} \end{theorem} \begin{proof} The Hessian matrix $\mathfrak{H}$ is positive definite if and only if it has a Cholesky decomposition, which then implies that $\mathcal{H}(\mathbf{x},\gamma,\bm{\lambda})$ is convex in both its state and control. Computing the Cholesky decomposition for $\mathfrak{H}$ we obtain: \begin{displaymath} \mathfrak{H} = \begin{bmatrix} \mathbf{I}_N & \mathbf{0}\\ \bm{\lambda}^T\mathbf{H} & \sqrt{r - \norm{\mathbf{H}^T\bm{\lambda}^T}}^2 \end{bmatrix}\begin{bmatrix} \mathbf{I}_N & \mathbf{H}^T\bm{\lambda}^T\\ \mathbf{0} & \sqrt{r - \norm{\mathbf{H}^T\bm{\lambda}^T}^2} \end{bmatrix}. \end{displaymath} This decomposition exists if and only if $r > \norm{\mathbf{H}^T\bm{\lambda}^T}^2$. The result follows immediately. \end{proof} As we demonstrate in the examples, this sufficient condition for optimality is precisely the one identified in \cite{GB17} for the triple public goods game, which was shown to be diffeomorphic to the cyclic game with three strategies (rock-paper-scissors). \subsection{Examples with $N = 3, 5, 7,9$} We study the derived optimal controls for the case when $N = 3, 5, 7, 9$ in both the fully non-linear optimal control problem and the quasi-linearized optimal control problem. In particular we observe similar structure to the optimal controls in all cases. For these examples, we set $t_f = 6$, except in the last case where we extend it to show an example where the sufficient condition for optimality is not satisfied. In Figure \ref{fig:3cyclic} we show the optimal control for both the non-linear and quasi-linearized optimal control problems. We set $r = 0.2$ and the initial state $\mathbf{u}_0 = (0.2333, 0.3333, 0.43333)$, which is close enough to the fixed point for the quasi-linearized approximation to be valid. We also demonstrate the equivalence between the solution to the quasi-linearized Euler-Lagrange equations and the second order differential equation (Eq. \ref{eqn:2ODE-Full}). We also show the approximation to the quasi-linearized control function that arises as a solution to Equation \ref{eqn:2ODE}. \begin{figure}[htbp] \centering \includegraphics[scale=0.24]{Figures/ControlPlot3-Double.pdf} \caption{The control function, its approximations and the $\norm{\mathbf{H}^T\bm{\lambda}^T}^2$, used in determining whether the solution to the necessary conditions are sufficient for an optimal control for the 3 strategy cyclic game (rock-paper-scissors).} \label{fig:3cyclic} \end{figure} In the 3 strategy case, the condition for optimality is: \begin{equation} r > \norm{\mathbf{H}^T\bm{\lambda}^T}^2 = \frac{1}{9}\norm{\bm{\lambda}}^2. \end{equation} This is not generally true, but it is equivalent to the sufficient condition derived in \cite{GB17} for the diffeomorphic triple public goods game. Thus, Theorem \ref{thm:Sufficient} generalizes the results from \cite{GB17}. In Figure \ref{fig:5cyclic} we show the relevant control plots for the 5 strategy cyclic game (rock-paper-scissors-Spock-lizard\footnote{Developed by Sam Kass. See \url{http://www.samkass.com/theories/RPSSL.html} or Episode 8, Season 2 of \textit{The Big Bang Theory}. Note, to properly organize the moves to produce a circulant matrix, the strategies should be ordered as rock-paper-scissors-Spock-lizard rather than rock-paper-scissors-lizard-Spock as they are on \textit{The Big Bang Theory}. Kass correctly organizes the strategies.}). We again use $r = 0.2$ and set $\mathbf{u}_0 = (0.1,0.3,0,0.1,0.3)$. \begin{figure}[htbp] \centering \includegraphics[scale=0.24]{Figures/ControlPlot5-Double.pdf} \caption{The control function, its approximations and the $\norm{\mathbf{H}^T\bm{\lambda}^T}^2$, used in determining whether the solution to the necessary conditions are sufficient for an optimal control for the 5 strategy cyclic game (rock-paper-scissors-Spock-lizard).} \label{fig:5cyclic} \end{figure} An interesting feature of the 5 strategy cyclic game is that the limiting approximation (Equation \ref{eqn:2ODE}) does not perform as well as it did for the 3 strategy game. As we see in Figures \ref{fig:7cyclic} and \ref{fig:9cyclic}, the approximation does improve (as we expect) as $N$ increases. This anomalous behavior may be a function of numerical instability or a property of the 5 strategy cyclic game. We do note that because of the properties of Equations \ref{eqn:2ODE-Full} and \ref{eqn:2ODE} (i.e, branching solutions), we did observe some numerical instability when simulating these systems. We note that in the 5 strategy case, the sufficient condition on optimality is satisfied. In Figures \ref{fig:7cyclic} and \ref{fig:9cyclic} we illustrate the optimal control for 7 and 9 strategy games. In these cases $r = 0.2$ again. To maintain feasibility of the starting solution, we set $\mathbf{u}_0$ by alternately adding and subtracting $0.05$ from the equilibrium, but kept the third strategy at proportion $1/N$ in both cases. Thus we assured $\mathbf{u}_0$ was in $\Delta_N$ in both cases. \begin{figure}[htbp] \centering \includegraphics[scale=0.24]{Figures/ControlPlot7-Double.pdf} \caption{The control function, its approximations and the $\norm{\mathbf{H}^T\bm{\lambda}^T}^2$, used in determining whether the solution to the necessary conditions are sufficient for an optimal control for the 7 strategy cyclic game.} \label{fig:7cyclic} \end{figure} \begin{figure}[htbp] \centering \includegraphics[scale=0.24]{Figures/ControlPlot9-Double.pdf} \caption{The control function, its approximations and the $\norm{\mathbf{H}^T\bm{\lambda}^T}^2$, used in determining whether the solution to the necessary conditions are sufficient for an optimal control for the 9 strategy cyclic game.} \label{fig:9cyclic} \end{figure} As we expect, as $N$ increases, the approximation in Equation \ref{eqn:2ODE} improves. It is also interesting to note that the general structure of the optimal control function is similar in all cases with the quasi-linearized control. It exhibits almost linear behavior, and the fully non-linear controller shows decreasing oscillation. That the optimal controller is a decreasing function is consistent with Proposition \ref{prop:DotGammaSign}. We can analyze the control problem even when the starting state is not \textit{near} the fixed point, which yields a case where the sufficient condition for optimality fails to hold. We again consider the case when $N = 3$ and extend the time horizon of control to $t_f = 15$. We start at the point $\mathbf{u}_0 = (0.8,0.1,0.1)$, which is not near the equilibrium point $\mathbf{u}^* = \tfrac{1}{3}\mathbf{1}$, thus reducing the accuracy of the quasi-linearized approximation. The objective of this example is to study both the extended time control horizon as well as the control that results when the starting point is further from the equilibrium point. We compute an optimal control using the fully non-linear Euler-Lagrange equations, the Euler-Lagrange equations for the quasi-linearized system, and the exact differential equation for $\gamma$ given in Lemma \ref{lem:MainControl1}. As demonstrated in Figure \ref{fig:Riccati}, the quasi-linearized control functions are identical to each other (as expected), and they are highly correlated to the control function for the fully non-linear system. The co-state, however does not satisfy the sufficient condition $r > \mathbf{H}^T\bm{\lambda}$. Here $r = 0.2$, as in the previous examples. \begin{figure}[htbp] \centering \subfigure[Control and Co-State Plot]{\includegraphics[scale=0.24]{Figures/ControlPlot3-DoubleLong.pdf}} \subfigure[Matrix Riccati Equation Solution Curves]{\includegraphics[scale=0.35]{Figures/RiccatiPlot.pdf}} \caption{A solution that violates the sufficient condition for optimality $r > \mathbf{H}^T\bm{\lambda}$ but can be shown to be optimal by appealing to the matrix Riccatti equation, which has bounded solutions for $t \in [0,t_f]$. We compare the solution to the ordinary controller derived from the ordinary Euler-Lagrange equations and the controller that is directly computed using Lemma \ref{lem:MainControl1}. Note they are identical as expected.} \label{fig:Riccati} \end{figure} We can still analyze this problem by solving the matrix Riccati equation as given in Theorem \ref{thm:main} to show that it is bounded, and thus the control identified for the quasi-linearized system is optimal. The matrix Riccati equation for this system is: \begin{align*} -\dot{\mathbf{S}} &= \mathbf{I}_3 + \left(\mathbf{J}+\gamma\mathbf{H}\right)^T\mathbf{S} + \mathbf{S}\left(\mathbf{J} + \gamma\mathbf{H}\right) - \\ & \hspace*{5em}\frac{1}{r}\left(\bm{\lambda}^T\mathbf{H}+\mathbf{x}^T\mathbf{H}^T\mathbf{S}\right)^T\left(\bm{\lambda}^T\mathbf{H}+\mathbf{x}^T\mathbf{H}^T\mathbf{S}\right),\\ \mathbf{S}(t_f) &= \mathbf{0}. \end{align*} This system of equations contains nine equations because $\mathbf{S}$ is $3 \times 3$. As shown in Figure \ref{fig:Riccati}, the solution curves for this equation are all bounded on the time interval of consideration. Furthermore, since the state and co-state necessarily satisfy the Euler-Lagrange equations, and the Hamiltonian is convex in $\gamma$ and $\gamma$ (and therefore maximizes the Hamiltonian at all times (see Lemma \ref{lem:NecessaryControl})), the control function must be (locally) optimal. Note that while the starting point in this example is not near the equilibrium point, the control computed for the quasi-linear approximation is still highly correlated to the control computed for the non-linear system. Thus, in this case quasi-linearization still provides a reasonable approximation to the optimal control in the non-linear system. \section{Final Discussion and Future Directions}\label{sec:Conclusion} In this paper we studied an optimal control problem arising from the class of complete, odd circulant games that generalize rock-paper-scissors. We used the replicator dynamics as the natural equations of motion in the optimal control problem. In particular, the control problem was to drive trajectories toward the unique interior fixed point of the replicator dynamics. We first studied the uncontrolled fixed points of the replicator. We then showed that this class of problems admits a natural quasi-linearization, and that this quasi-linearized optimal control problem has a open-loop optimal control satisfying a specific second order differential equation. Furthermore, we showed that as the number of strategies grows, this differential equation admits a closed form solution. Numerical comparisons showed that this limiting case provides a natural approximation for the optimal control. We also showed that even when the starting conditions for the optimal control are far from the interior Nash equilibrium, where quasi-linearization is performed, we still can use it to approximate the optimal control in the original problem. Unfortunately, the results presented in this paper do not apply to arbitrary circulant games or to cyclic games with a control parameter. In this context, a cyclic game is any circulant game in which the first row of the $\mathbf{L}_N$ matrix has form $(0,1,0,\dots,0,-1)$. Here the number of $0$'s between $1$ and $-1$ is $N-3$. The resulting matrix $\mathbf{M}_N$ with $1$'s corresponding to the $1'$ in $\mathbf{L}$ and $0$ elsewhere is the adjacency matrix of a cycle graph. Rock-paper-scissors is the only game that is both a cyclic and complete odd circulant graph (because the three-cycle is the complete graph on three vertices). Any circulant game whose payoff matrix can be written as $\mathbf{L}_N + \gamma\mathbf{M}_N$ will obey Equation \ref{eqn:dotgamma} because the circulant matrices form a commutative algebra. Beyond that, it is possible addition dynamics govern the optimal controls in these cases. This presents a logical area for further study. Another logical extension of this work is to study circulant games with an off-center interior equilibrium point, rather than $\mathbf{u}^* = \tfrac{1}{N}\mathbf{1}$. Doing so, however, should introduce additional control parameters. This would be an interesting extension as well since this paper considered only a single control parameter. It also will make the application more realistic since an individual may have varying degrees of control over each population. In addition to introducing additional controls, another natural extension of this work is to derive controllers that drive the system toward a non-interior equilibrium point. In particular it would be intriguing to study the problem of deliberately eliminating one or more species. Finally, studying more complex dynamics with control, like the mutator-replicator, may produce interesting and useful results. However, these may not admit the necessary conditions to allow the control mechanisms identified in this paper to be applied. \section*{Acknowledgement} CG was supported by the National Science Foundation under grant number CMMI-1463482. The authors thank Andrew Belmonte for his helpful comments and discussion.

https://arxiv.org/abs/0804.3662

Jensen Shannon divergence as a measure of the degree of entanglement

The notion of distance in Hilbert space is relevant in many scenarios. In particular, distances between quantum states play a central role in quantum information theory. An appropriate measure of distance is the quantum Jensen Shannon divergence (QJSD) between quantum states. Here we study this distance as a geometrical measure of entanglement and apply it to different families of states.

\section{Introduction} Discerning possible candidates for measuring distances between quantum states is a subject of perennial interest. Many of these measures were first defined as distances between probability distributions and subsequently employed as distance-measures in Hilbert space. Let ${\cal H}$ be the Hilbert space associated with a quantum system and let ${\cal S}$ be the set of all states, i.e. the set of self-adjoint, (semi)positive and trace-one operators. A frequently employed notion of distance between quantum states is the relative entropy, which is a natural extension to the realm of quantum mechanics of the Kullback-Leibler divergence. This quantity, however, is not useful for ascertaining the degree of purification of an arbitrary state with respect to a pure reference state. The relative entropy of an operator $\rho$, with respect to an operator $\sigma$, both belonging to $\cal{S}$, is given by \begin{equation} S(\rho\|\sigma)={\mathrm{tr}}[\rho(\log\rho-\log\sigma)],\label{relative-entropy} \end{equation} where $\log$ stands for logarithm in base two. $S(\rho\|\sigma)$ is nonnegative and vanishes if and only if $\rho = \sigma$, being nonsymmetric and unbounded. A particular and important requirement indicates that the relative entropy is well defined only when the support of $\sigma$ is equal to or larger than that of $\rho$. Otherwise, it is defined to be $+\infty$ \cite{Lindblad} (the support of an operator is the subspace spanned by its eigenvectors with non-zero eigenvalues). To overcome this restriction we have introduced a distance between elements of ${\cal S}$ that shares with the relative entropy several of their main properties but that is always well defined and bounded \cite{Majtey}. This distance is the quantum Jensen-Shannon divergence (QJSD), which is a quantum mechanical extension of the Jensen-Shannon divergence (JSD) introduced by Rao \cite{Rao} and Lin \cite{Lin} as a distance between probability distribution (for a detailed analysis of the properties of the JSD, see reference \cite{Majtey0}). Here we wish to investigate the ability of the QJSD to serve as a measure of the degree of entanglement. The structure of the paper is as follows. In the next Section we review the basic properties of the QJSD. In Section 3 we study its properties as an entanglement measure and we apply it to quantify the degree of entanglement of different families of two-qubit mixed states. Finally, some conclusions are drawn in Section 4. \section{The quantum Jensen-Shannon divergence} We define the QJSD in the fashion \cite{Majtey} \begin{equation} JS(\rho\|\sigma) = \frac{1}{2}\left[S\left(\rho\|\frac{\rho+\sigma}{2}\right)+ S\left(\sigma\|\frac{\rho+\sigma}{2}\right)\right], \label{relative-entropy} \end{equation} which can be also recast in terms of the von Neumann entropy $H_N(\rho) = - {\mathrm{tr}}(\rho \log \rho)$ as follows \begin{equation} JS(\rho\|\sigma) = H_N\left(\frac{\rho + \sigma}{2}\right)-\frac{1}{2} H_N(\rho) - \frac{1}{2} H_N(\sigma). \end{equation} This quantity has a lot to speak for, being positive, null iff $\rho = \sigma$, symmetric, bounded, and always well defined. In fact, the restriction imposed on the supports of $\rho$ and $\sigma$ for the relative entropy (\ref{relative-entropy}) is lifted for the QJSD, that possesses all the adequate properties of a proper distance between states in a Hilbert space. As stated above, in this work we attempt to study the QJSD as an entanglement measure which justifies listing the main QJSD properties. Most of these properties are discussed and proved in Ref.~\cite{Majtey}. The list reads \begin{itemize} \item[(i)] $JS(\rho\|\sigma)\geq 0$ with the equality iff $\rho=\sigma$ \item[(ii)] Unitary operations left JS invariant, i.e., $JS(U\rho U^\dag\|U \sigma U^\dag) = JS(\rho\|\sigma)$. \item[(iii)] $JS({\mathrm{tr}}_p\rho \|{\mathrm{tr}}_p\sigma) \leq JS(\rho \| \sigma)$ where ${\mathrm{tr}}_p$ is the partial trace. \item[(iv)] JS is jointly convex $JS(\sum_i \alpha_i \rho^{(i)}\|\sum_i \alpha_i \sigma^{(i)}) \leq \sum_i \alpha_i JS(\rho^{(i)}\|\sigma^{(i)})$, where the $\alpha_i$ are positive real numbers such that $\sum_i \alpha_i =1$. \item[(v)] $JS(\Phi\rho\|\Phi\sigma)\leq JS(\rho\|\sigma)$ for all positive mappings $\Phi$. \item[(vi)] For any set of orthogonal projectors $P_i$, such that $P_iP_j=\delta_{ij}P_i$, $JS(\sum P_i\rho P_i\|\sum P_i\sigma P_i)=\sum JS(P_i\rho P_i\| P_i\sigma P_i)$ \item[(vii)] $JS(\rho\otimes P_{\alpha}\|\sigma\otimes P_{\alpha})=JS(\rho\|\sigma)$ where $P_{\alpha}$ is any projector. \end{itemize} The two last properties, which have not been discussed before, are verified by the relative entropy \cite{Lindblad,Vedral98}, and inherited by the QJSD. In a recent work \cite{Triangular} the metric character of the QJSD has been discussed. There it was formally proved for pure states, and checked numerically for mixed ones by performing Monte Carlo simulations. We can thus assert that the square root of the QJSD verifies the triangle inequality, as it does the square root of the JSD. \section{The entanglement measure} Entanglement constitutes a physical resource that lies at the heart of quantum information processes \cite{Nielsen,LPS98,BEZ00}. Quantum teleportation, superdense coding, and quantum computation, are some of the most representative examples. Let us recall that a state of a composite quantum system is called entangled if it can not be expressed as a convex sum of factorizable pure states. Otherwise, the state is called separable. Nowadays a variety of measures are used to quantify the degree of entanglement. These include the entanglement of distillation, the relative entropy of entanglement, etc. The canonical measure of entanglement in a bipartite pure systems is the so-called entanglement of formation, which is a strictly monotonic function of the squared concurrence. For simplicity the entanglement of formation is frequently used as {\it the} measure of entanglement. However, for mixed state there exist several available measures. Before starting with the entanglement-characterization using the QJSD we study the structure of Hilbert space (HS) according to this distance measure. To do that we evaluate $d_{JS}=\sqrt{JS}$ between a given random generated state and the maximally mixed (MM) state $\rho_{MM}=I/N$, with $N$ the HS-dimension. In Fig. 1 we plot the probability distribution for finding an arbitrary state $\rho$ at a given distance from the $\rho_{MM}$. We find, as expected, that the mean value of such distance increases for higher Hilbert space-dimensions. Let us now enumerate the properties that any adequate measure of entanglement should satisfy \cite{Bengtsson}: \begin{itemize} \item[(i)] Discrimination: ${\cal{E}}(\rho)=0 \;\; iff\;\; \rho$ is separable. \item[(ii)] Monotonicity: the measure does not increase under local general measurements and classical communication, for every completely positive map $\Phi$, ${\cal{E}}(\Phi\rho)\leq{\cal{E}}(\rho)$ \item[(iii)] Convexity: ${\cal{E}}(x\rho + (1-x)\sigma)\leq x{\cal{E}}(\rho)+(1-x){\cal{E}}(\sigma)$, with $x \in [0,1]$. \end{itemize} As it was already stated, the main purpose of the present communication is to investigate the QJSD as a {\it geometrical} measure of entanglement. Following Vedral \textit{et al.}, \cite{Vedral97} we define an entanglement measure $\cal{E}(\rho)$ as the minimum QJSD from the state $\rho$ to the set ${\cal D}$ of the disentangled states. \begin{equation} {\cal{E}}_{JS}(\rho)=\min_{\sigma\in {\cal D}}JS(\rho,\sigma).\label{js-entanglement} \end{equation} The QJSD properties enunciated in the preceding section ensure that (\ref{js-entanglement}) fulfills the conditions for an adequate entanglement measure. For convenience's sake, we normalize the entanglement measure by a trivial re-scaling in order to adequately compare different quantities of interest, i.e., ${\cal{E}}(\rho)={\cal{E}}(\rho)/{\cal{E}}(|\psi^-\rangle\langle\psi^-|).$ It is notheworthy to stress that expression (\ref{js-entanglement}) has already been investigated for different distances, other than the JS-one. For example, Vedral and coworkers used to this effect the relative entropy (${\cal{E}}_{RE}(\rho)$) and the Bures metric (${\cal{E}}_B(\rho)$) \cite{Vedral97,Vedral98}. Also the Trace-distance and the Hilbert-Schmidt metric have received consideration for the purpose \cite{Eisert03,Witte99}. We shall below use as a reference the quantity ${\cal{E}}_B(\rho)$. \begin{figure}[h] \includegraphics[scale=0.9,angle=0]{fig1.eps} \caption{Probability distribution for finding a quantum state at a given distance from de maximally mixed one (for different Hilbert space dimensions).} \end{figure} \begin{figure} \includegraphics[scale=0.9,angle=0]{fig2.eps} \caption{Comparison between the concurrence, the JS- measure, and the measure ${\cal{E}}_{B}$ induced by the Bures metric. The task is accomplished for (a) Werner, (b) MEM, and (c) PDC states.} \end{figure} In order to investigate the behavior of the QJSD as an entanglement measure we consider now three well-known families of two-qubits mixed states, namely, \begin{itemize} \item Werner ones \cite{Werner}, \item maximally entangled mixed ones \cite{MEM} (MEM), and \item parametric down-conversion \cite{PDC} (PDC) states. \end{itemize} All of these are diagonal in the Bell basis. We perform an optimization procedure so as to find the separable state that lies in the closest proximity to the state of interest $\rho$. Such task is performed by following a simulated annealing minimization procedure, starting from the maximally mixed state $\rho_{MM}$. Appropriately perturbing $\rho_{MM}$ we look for the minimal QJSD between $\rho$ and some state in $\cal{D}$. In Fig.2 we depict, versus the linear entropy $S_L$ (the degree of impurity), i) the concurrence and the two normalized quantities: ii) our ${\cal{E}}_{JS}$ and iii) Vedral's ${\cal{E}}_{B}$ measure. This is done for the three families enumerated above. Notice that the concurrence and ${\cal{E}}_{B}$ are greater than ${\cal{E}}_{JS}$, except for the ``extreme" cases of maximally entangled or disentangled Bell states in two of the sub-figures. In the MEMs-instance this is not entirely so. Crossings are detected between ${\cal{E}}_{B}$ and the concurrence ${\cal{C}}$, although the ${\cal{E}}_{JS}-$curve lies always below the other two for all the three families, except for the extreme (trivial) situations. The behavior here described mimics that of the relative entropy vs. $S_L$. In that instance, the relative entropy-induced entanglement is strictly smaller than that of formation, as analytically demonstrated in \cite{Vedral98}. Our crucial point is the fact that the measure induced by the QJSD is easier to compute than that induced by the Bures metric. This is so because of the square root in the definition of Bures' metric, which forces cumbersome matrix-diagonalization and basis' changes in order to compute ${\cal{E}}_{B}$. Instead, so as to calculate the QJSD one i) only needs to solve the eingenvalue-equation for three matrices (eigenvectors are not necessary) and ii) the optimization procedure is much more efficient than in the Bures situation. \section{Conclusions} We have here advanced a new entanglement-measure ${\cal{E}}_{JS}$, based on the Jensen-Shannon distance. For three important families of states this new quantity behaves in rather similar fashion as the established entanglement-measure ${\cal{E}}_{B}$ of Vedral \textit{et al.}, with an important difference: the former is much more easy to compute than the later. \section*{Acknowledgments} This work was partially supported by the MEC grant FIS2005-02796 (Spain) and FEDER (EU) and by CONICET (Argentine Agency). AB and APM acknowledge support from MEC through FPU grant AP-2004-2962 and contract SB-2006-0165. PWL wants to thank SECyT-UNC (Argentina) for financial support. We thank J. Batlle and Prof. A. R. Plastino for using his Hilbert-Monte Carlo numerical program.

https://arxiv.org/abs/1502.01205

Centroid of triangles associated with a curve

Archimedes showed that the area between a parabola and any chord $AB$ on the parabola is four thirds of the area of triangle $\Delta ABP$, where P is the point on the parabola at which the tangent is parallel to the chord $AB$. Recently, this property of parabolas was proved to be a characteristic property of parabolas. With the aid of this characterization of parabolas, using centroid of triangles associated with a curve we present two conditions which are necessary and sufficient for a strictly locally convex curve in the plane to be a parabola.

\section{Introduction} \vskip 0.50cm We study strictly locally convex plane curves. Recall that a regular plane curve $X:I\rightarrow {\mathbb R}^{2}$ in the plane ${\mathbb R}^{2}$, where $I$ is an open interval, is called {\it convex} if, for all $s\in I$ the trace $X(I)$ of $X$ lies entirely on one side of the closed half-plane determined by the tangent line at $X(s)$ (\cite{dC}). A regular plane curve $X:I\rightarrow {\mathbb R}^{2}$ is called {\it locally convex} if, for each $s\in I$ there exists an open subinterval $I_0\subset I$ containing $s$ such that the curve $X|_{I_0}$ restricted to $I_0$ is a convex curve. \vskip 0.3cm Henceforth, we will say that a locally convex curve $X$ in the plane ${\mathbb R}^{2}$ is {\it strictly locally convex} if the curve is smooth (that is, of class $C^{(3)}$) and is of positive curvature $\kappa$ with respect to the unit normal $N$ pointing to the convex side. Hence, in this case we have $\kappa(s)=\left< X''(s), N(X(s))\right> >0$, where $X(s)$ is an arc-length parametrization of $X$. \vskip 0.3cm When $f:I\rightarrow {\mathbb R}$ is a smooth function defined on an open interval $I$, we will also say that $f$ is {\it strictly convex} if the graph of $f$ has positive curvature $\kappa$ with respect to the upward unit normal $N$. This condition is equivalent to the positivity of $f''(x)$ on $I$. \vskip 0.3cm Suppose that $X$ is a strictly locally convex curve in the plane ${\mathbb R}^{2}$ with the unit normal $N$ pointing to the convex side. For a fixed point $P \in X$, and for a sufficiently small $h>0$, we consider the line $\ell$ passing through $P+hN(P)$ which is parallel to the tangent line $t$ of $X$ at $P$ and the points $A$ and $B$ where the line $\ell$ intersects the curve $X$. We denote by $t_1$, $t_2$ the tangent lines of $X$ at the points $A,B$ and by $Q,A_1,B_1$ the intersection points $t_1\cap t_2$, $t_1\cap t$, $t_2\cap t$, respectively. We let $L_P(h)$, $g_P(h)$, $j_P(h)$ and $k_P(h)$ denote the length of the chord $AB$ , the distance from the centroid $G$ of the section of $X$ cut off by $\ell$ to the line $\ell$, the distance from the centroid $J$ of the triangle $\bigtriangleup QAB$ to the line $\ell$ and the distance from the centroid $K$ of the triangle $\bigtriangleup QA_1B_1$ to the line $\ell$, respectively. \vskip 0.3cm Now, we consider $S_P(h)$ and $T_P(h)$ defined by the area of the region bounded by the curve $X$ and chord $AB$, the area $|\bigtriangleup PAB|$ of triangle $\bigtriangleup PAB$, respectively. Then, obviously we have \begin{equation}\tag{1.1} \begin{aligned} T_P(h)=\frac{1}{2}hL_P(h) \end{aligned} \end{equation} and we get (\cite{KK4}) \begin{equation}\tag{1.2} \begin{aligned} \frac{d}{dh}S_P(h)=L_P(h). \end{aligned} \end{equation} \vskip 0.3cm It is well known that parabolas satisfy the following properties. \vskip 0.3cm \noindent {\bf Proposition 1.} Suppose that $X$ is an open part of a parabola. Then we have the following. \noindent 1) For arbitrary point $P\in X$ and sufficiently small $h>0$, $X$ satisfies \begin{equation}\tag{1.3} \begin{aligned} S_P(h)=\frac{4}{3}T_P(h). \end{aligned} \end{equation} \noindent 2) For arbitrary point $P\in X$ and sufficiently small $h>0$, $X$ satisfies \begin{equation}\tag{1.4} \begin{aligned} g_P(h)=\frac{2}{5}h. \end{aligned} \end{equation} \noindent 3) For arbitrary point $P\in X$ and sufficiently small $h>0$, $X$ satisfies \begin{equation}\tag{1.5} \begin{aligned} j_P(h)=\frac{2}{3}h. \end{aligned} \end{equation} \noindent 4) For arbitrary point $P\in X$ and sufficiently small $h>0$, $X$ satisfies \begin{equation}\tag{1.6} \begin{aligned} k_P(h)=\frac{4}{3}h. \end{aligned} \end{equation} \vskip 0.3cm \noindent {\bf Proof.} For a proof of 1), see \cite{S}. If we denote by $V$ the point where the parallel line $m$ through the point $P$ to the axis of $X$ meets the chord $AB$, then $V$ is the mid point of $AB$ and the point $Q$ is on the line $m$ with $PV=PQ$. This completes the proof of 2), 3) and 4). $\square$ \vskip 0.3cm Very recently, the first author of the present paper and Y. H. Kim showed that among strictly convex plane curves, the above area property (1.3) of parabolic sections characterize parabolas. More precisely, they proved as follows (\cite{KK4}). \vskip 0.3cm \noindent {\bf Proposition 2.} Let $X$ be a strictly convex curve in the plane ${\mathbb R}^{2}$. Then $X$ is a parabola if and only if it satisfies \vskip 0.3cm \noindent $(C):$ For a point $P$ on $X$ and a chord $AB$ of $X$ parallel to the tangent of $X$ at $P$, the area of the region bounded by the curve and $AB$ is $4/3$ times the area of triangle $\bigtriangleup ABP$, that is, $$ S_P(h)=\frac{4}{3}T_P(h). $$ \vskip 0.3cm Archimedes showed that parabolas satisfy (1.3) (\cite{S}). Actually, in \cite{KK4} the first author of the present paper with Y. H. Kim established five characterizations of parabolas, which are the converses of well-known properties of parabolas originally due to Archimedes (\cite{S}). For some properties and characterizations of parabolas with respect to the area of triangles associated with a curve, see \cite{D,KKKP,KS, Kr}. For the higher dimensional analogues of some results in \cite{KK4}, see \cite{KK2} and \cite{KK3}. \vskip 0.3cm \vskip 0.3cm In \cite{KKP}, using Proposition 1, D.-S. Kim et al. proved the following characterization theorem for parabolas with respect to the function $ g_P(h)$. \vskip 0.3cm \noindent {\bf Proposition 3.} Let $X$ be a strictly locally convex plane curve in the plane ${\mathbb R}^{2}$. For a fixed point $P$ on $X$ and a sufficiently small $h>0$, we denote by $\ell$ the parallel line through $P+hN(P)$ to the tangent $t$ of the curve $X$ at $P$. We let $g_P(h)$ the distance from the center $G$ of gravity of the section of $X$ cut off by $\ell$ to the line $\ell$. Then $X$ is an open part of a parabola if and only if it satisfies for a fixed point $P$ on $X$ and a sufficiently small $h>0$ $$ g_P(h)=\frac{2}{5}h. $$ \vskip 0.3cm In \cite{KKP}, the distance from the center $G$ of gravity of the section of $X$ cut off by $\ell$ to the tangent $t$ of $X$ at $P$ was denoted by $d_P(h)$. Hence, we see that $d_P(h)+g_P(h)=h$. \vskip 0.3cm In this article, we study whether the remaining properties of parabolas in Proposition 1 characterize parabolas. \vskip 0.3cm First of all, in Section 2 we prove the following: \vskip 0.3cm \noindent {\bf Theorem 4.} Suppose that $X$ denotes a strictly locally convex plane curve in the plane ${\mathbb R}^{2}$. For a fixed point $P$ on $X$ and a sufficiently small $h>0$, we denote by $\ell$ the parallel line through $P+hN(P)$ to the tangent $t$ of the curve $X$ at $P$. Then we have \begin{equation}\tag{1.7} \begin{aligned} \lim _{h\rightarrow 0}\frac{j_P(h)}{h}=\frac{2}{3}. \end{aligned} \end{equation} and \begin{equation}\tag{1.8} \begin{aligned} \lim _{h\rightarrow 0}\frac{k_P(h)}{h}=\frac{4}{3}. \end{aligned} \end{equation} \vskip 0.3cm Finally, with the aid of the characterization theorem of parabolas (Theorem 3 in \cite{KK4}), in Section 4 we prove the following. \vskip 0.3cm \noindent {\bf Theorem 5.} Suppose that $X$ denotes a strictly locally convex $C^{(3)}$ curve in the plane ${\mathbb R}^{2}$. Then the following are equivalent. \noindent 1) For all $P\in X$ and sufficiently small $h>0$, $X$ satisfies \begin{equation}\tag{1.9} \begin{aligned} j_P(h)=\lambda(P)h^{\mu(P)}, \end{aligned} \end{equation} where $\lambda(P)$ and $\mu(P)$ are some functions. \noindent 2) For all $P\in X$ and sufficiently small $h>0$, $X$ satisfies $$ j_P(h)=\frac{2}{3}h. $$ \noindent 3) $X$ is an open part of a parabola. \vskip 0.3cm For the function $k_P(h)$, the similar argument as in the proof of Theorem 5 yields the following. \vskip 0.3cm \noindent {\bf Theorem 6.} Suppose that $X$ denotes a strictly locally convex $C^{(3)}$ curve in the plane ${\mathbb R}^{2}$. Then the following are equivalent. \noindent 1) For all $P\in X$ and sufficiently small $h>0$, $X$ satisfies \begin{equation}\tag{1.10} \begin{aligned} k_P(h)=\lambda(P)h^{\mu(P)}, \end{aligned} \end{equation} where $\lambda(P)$ and $\mu(P)$ are some functions. \noindent 2) For all $P\in X$ and sufficiently small $h>0$, $X$ satisfies $$ k_P(h)=\frac{4}{3}h. $$ \noindent 3) $X$ is an open part of a parabola. \vskip 0.3cm \noindent {\bf Remark.} If we consider the distance $\delta_P(h)$ from the centroid of the triangle $\Delta PAB$ to the parallel line $\ell$ through the point $P+hN(P)$ to the tangent $t$ of a strictly locally convex curve $X$ at $P$, then $X$ always satisfies $\delta_P(h)=\frac{1}{3}h$ for all sufficiently small $h>0$. \vskip 0.3cm For some characterizations of parabolas or conic sections by properties of tangent lines, see \cite{KKa} and \cite{KKPj}. In \cite{KK1}, using curvature function $\kappa$ and support function $h$ of a plane curve, the second author of the present paper and Y. H. Kim gave a characterization of ellipses and hyperbolas centered at the origin. Among the graphs of functions, \'A. B\'enyi et al. proved some characterizations of parabolas (\cite{BSV1, BSV2}). In \cite{R}, B. Richmond and T. Richmond established a dozen necessary and sufficient conditions for the graph of a function to be a parabola by using elementary techniques. \vskip 0.3cm Throughout this article, all curves are of class $C^{(3)}$ and connected, unless otherwise mentioned. \vskip 0.50cm \section{Proof of Theorem 4} \vskip 0.5cm In this section, we prove Theorem 4. First of all, we need the following lemma (\cite{KK4}) which is useful in the proof of main theorems. \vskip 0.3cm \noindent {\bf Lemma 7.} Suppose that $X$ is a strictly locally convex $C^{(3)}$ curve in the plane ${\mathbb R}^{2}$ with the unit normal $N$ pointing to the convex side. Then we have \begin{equation}\tag{2.1} \begin{aligned} \lim_{h\rightarrow 0} \frac{1}{\sqrt{h}}L_P(h)= \frac{2\sqrt{2}}{\sqrt{\kappa(P)}}, \end{aligned} \end{equation} where $\kappa(P)$ is the curvature of $X$ at $P$ with respect to the unit normal $N$ pointing to the convex side. \vskip 0.3cm Now, we prove Theorem 4 as follows. Let us denote by $X$ a strictly locally convex $C^{(3)}$ curve in the Euclidean plane ${\mathbb R^2}$. We fix an arbitrary point $P$ on $X$. Then, we may take a coordinate system $(x,y)$ of ${\mathbb R}^{2}$ such that $P$ is the origin $(0,0)$ and $x$-axis is the tangent line $t$ of $X$ at $P$. Furthermore, we may regard $X$ to be locally the graph of a non-negative strictly convex function $f: {\mathbb R}\rightarrow {\mathbb R}$ with $f(0)=f'(0)=0$. Then $N$ is the upward unit normal. Since the curve $X$ is of class $C^{(3)}$, the Taylor's formula of $f(x)$ is given by \begin{equation}\tag{2.2} f(x)= ax^2 + f_3(x), \end{equation} where $2a=f''(0)$ and $f_3(x)$ is an $O(|x|^3)$ function. Noting that the curvature $\k$ of $X$ at $P$ is given by $\kappa(P)=f''(0)>0$, we see that $a$ is positive. For a sufficiently small $h>0$, the line $\ell$ through $P+hN(P)$ and orthogonal to $N(P)$ is given by $y=h$. We denote by $A(s,f(s))$ and $B(t,f(t))$ the points where the line $\ell:y=h$ meets the curve $X$ with $s<0<t$. Then we have $f(s)=f(t)=h$. The tangent lines $t_1$ and $t_2$ to $X$ at $A$ and $B$ intersect at the point $Q=(x_0(h),y_0(h))$ with \begin{equation}\tag{2.3} x_0(h)=\frac{tf'(t)-sf'(s)}{ f'(t)-f'(s)}, \end{equation} \begin{equation}\tag{2.4} y_0(h)=h+\frac{(t-s)f'(t)f'(s)}{ f'(t)-f'(s)}<0 \end{equation} and they meet the $x$-axis (the tangent to $X$ at $P$) at $B_1(s-h/f'(s),0)$ and $B_2(t-h/f'(t),0)$, respectively. Noting $L_P(h)=t-s$, one gets \begin{equation}\tag{2.5} j_P(h)=h-\frac{1}{3}\{2h+y_0(h)\}=-\frac{1}{3}\frac{L_P(h)f'(t)f'(s)}{ f'(t)-f'(s)} \end{equation} and \begin{equation}\tag{2.6} k_P(h)=h-\frac{1}{3}y_0(h)=\frac{2}{3}h-\frac{1}{3}\frac{L_P(h)f'(t)f'(s)}{ f'(t)-f'(s)}. \end{equation} Hence we obtain \begin{equation}\tag{2.7} \frac{j_P(h)}{h}=\frac{1}{3}\frac{L_P(h)}{\sqrt{h}}\frac{1}{\alpha_P(h)} \end{equation} and \begin{equation}\tag{2.8} \frac{k_P(h)}{h}=\frac{2}{3}+\frac{j_P(h)}{h}, \end{equation} where we use \begin{equation}\tag{2.9} \begin{aligned} \alpha_P(h)=\frac{(f'(s)-f'(t))}{f'(s)f'(t)}\sqrt{h}. \end{aligned} \end{equation} On the other hand, it follows from Lemma 5 in \cite{KS} that \begin{equation}\tag{2.10} \begin{aligned} \lim_{h\rightarrow0}\alpha_P(h)=\frac{\sqrt{2}}{\sqrt{\kappa(P)}}. \end{aligned} \end{equation} Together with (2.7) and Lemma 7, this shows that \begin{equation}\tag{2.11} \begin{aligned} \lim_{h\rightarrow0}\frac{j_P(h)}{h}=\frac{2}{3}, \end{aligned} \end{equation} and hence from (2.8) we also get \begin{equation}\tag{2.12} \begin{aligned} \lim_{h\rightarrow0}\frac{k_P(h)}{h}=\frac{4}{3}. \end{aligned} \end{equation} This completes the proof of Theorem 4. \vskip 0.3cm \vskip 0.5cm \section{Proofs of Theorems 5 and 6} \vskip 0.5cm In this section, in order to prove Theorem 5 we use the main result of \cite{KK4} (Theorem 3 in \cite{KK4}) and Theorem 4 stated in Section 1. First, we prove \vskip 0.3cm \noindent {\bf Lemma 8.} Suppose that $X$ is a strictly locally convex $C^{(3)}$ curve in the plane ${\mathbb R}^{2}$ with the unit normal $N$ pointing to the convex side. Then we have \begin{equation}\tag{3.1} \begin{aligned} \sqrt{h}\frac{d}{dh}L_P(h)=\alpha_P(h). \end{aligned} \end{equation} where $\alpha_P(h)$ is defined in (2.9). \vskip 0.3cm \noindent {\bf Proof.} Just as in the proof of Theorem 4 in Section 2, for an arbitrary point $P$ on $X$ we take a coordinate system $(x,y)$ of ${\mathbb R}^{2}$ so that (2.2) holds with $f(0)=f'(0)=0$ and $2a=f''(0)>0$. Then, for sufficiently small $h>0$, we put $f(t)=h$ with $t>0$ and we denote by $A(s(t),h)$ and $B(t,h)$ the points where the line $\ell:y=h$ meets the curve $X$ with $s=s(t)<0<t$. Then we have \begin{equation}\tag{3.2} \begin{aligned} f(s(t))=f(t)=h \end{aligned} \end{equation} and \begin{equation}\tag{3.3} \begin{aligned} L_P(h)=t-s(t). \end{aligned} \end{equation} Noting $h=f(t)$, one obtains from (3.3) that \begin{equation}\tag{3.4} \begin{aligned} \frac{d}{dh}L_P(h)=(1-s'(t))\frac{dt}{dh}=\frac{1-s'(t)}{f'(t)}. \end{aligned} \end{equation} Therefore, it follows from (3.2) that \begin{equation}\tag{3.5} \begin{aligned} \frac{d}{dh}L_P(h)=\frac{1}{f'(t)}-\frac{1}{f'(s(t))}=\frac{f'(s)-f'(t)}{f'(t)f'(s)}. \end{aligned} \end{equation} Together with (2.9), this completes the proof of Lemma 8. $\square$ \vskip 0.3cm It is obvious that any open part of parabolas satisfy 1) and 2) in Theorem 5. Conversely, suppose that $X$ is a strictly locally convex $C^{(3)}$ curve in the plane ${\mathbb R}^{2}$ which satisfies for all $P\in X$ and sufficiently small $h>0$ $$ j_P(h)=\lambda(P)h^{\mu(P)}, $$ where $\lambda(P)$ and $\mu(P)$ are some functions. Using Theorem 4, by letting $h\rightarrow 0$ we see that \begin{equation}\tag{3.6} \begin{aligned} \lim_{h\rightarrow 0}h^{\mu(P)-1}=\frac{2}{3}\frac{1}{\lambda(P)}, \end{aligned} \end{equation} which shows that $\mu(P)=1$ and $\lambda(P)=\frac{2}{3}$. Therefore, the curve $X$ satisfies for all $P\in X$ and sufficiently small $h>0$ $$ j_P(h)=\frac{2}{3}h. $$ Now, using Lemma 8 we get the following. \vskip 0.3cm \noindent {\bf Lemma 9.} Suppose that $X$ denotes a strictly locally convex $C^{(3)}$ curve in the plane ${\mathbb R}^{2}$ which satisfies (1.5) for all $P\in X$ and sufficiently small $h>0$. Then for all $P\in X$ and sufficiently small $h>0$ we have \begin{equation}\tag{3.7} L_P(h)=\frac{2\sqrt{2}}{\sqrt{\kappa(P)}}\sqrt{h}. \end{equation} \vskip 0.3cm \noindent {\bf Proof.} It follows from (1.5) and (2.7) that \begin{equation}\tag{3.8} L_P(h)=2\sqrt{h}\alpha_P(h). \end{equation} Together with Lemma 8, this yields \begin{equation}\tag{3.9} 2h\frac{d}{dh}L_P(h)=L_P(h). \end{equation} By integrating (3.9) with respect to $h$, we get for some constant $C=C(P)$ \begin{equation}\tag{3.10} L_P(h)=C\sqrt{h}. \end{equation} Thus, Lemma 7 completes the proof of Lemma 9. $\square$ \vskip 0.3cm \vskip 0.3cm Finally, we prove Theorem 5 as follows. \vskip 0.3cm It follows from (1.2) and $S_P(0)=0$ that by integrating (3.7) we get \begin{equation}\tag{3.11} S_P(h)=\frac{4\sqrt{2}}{3\sqrt{\kappa(P)}}h\sqrt{h}. \end{equation} Hence, together with (1.1) and (3.7), (3.11) yields that for all $P\in X$ and sufficiently small $h>0$ \begin{equation}\tag{3.12} S_P(h)=\frac{4}{3}T_P(h). \end{equation} Thus, it follows from Proposition 2 that $X$ is an open part of a parabola. This completes the proof of Theorem 5. \vskip 0.3cm In order to prove Theorem 6, with the help of (2.8) we may use the similar argument as in the proof of Theorem 5. \vskip 0.50cm \vskip 0.50cm

https://arxiv.org/abs/1103.4251

On exit time of stable processes

We study the exit time $\tau=\tau_{(0,\infty)}$ for 1-dimensional strictly stable processes and express its Laplace transform at $t^\alpha$ as the Laplace transform of a positive random variable with explicit density. Consequently, $\tau$ satisfies some multiplicative convolution relations. For some stable processes, e.g. for the symmetric $\frac23$-stable process, explicit formulas for the Laplace transform and the density of $\tau$ are obtained as an application.

\section{Introduction} Let $\alpha \in (0,2)$ and $(X_t, \mathbb{P}^x)$ be a strictly $\alpha$-stable process in $\mathbb{R}$ with characteristic function $$ \mathbb{E}^0 e^{i X_t z} = \exp\left[-t |z|^\alpha\left(1-i \beta \arctan \frac{\pi \alpha}{2}\sgn z\right)\right], $$ where $\beta \in [-1,1]$ and $\beta =0$ for $\alpha=1$. For any $D \subset \mathbb{R}$ let $$\tau_D = \inf\{t \ge 0 \colon X_t \not\in D\}$$ be the first exit time from $D$ of the process $X_t$. Throughout this article we shall consider the starting point $x>0$ and $$ \tau=\tau_{(0,\infty)}, $$ the exit time of $X_t$ from the positive half-line. The question of the first exit time from domains are basic for all stochastic processes. Surprisingly few exact formulas are known for stable processes. The only exceptions are Brownian motion, completely asymmetric stable processes with $\alpha >1$ (see \cite{Bing}, \cite{2006-EAK-sv}, \cite{TS-em}) and symmetric Cauchy process (\cite{1956-DAD-tams}, see also \cite{KKMS}). The quotient $\hat \tau/\tau$ was studied for independent $\tau$ and dual $\hat \tau$ in \cite{Doney2}. Some recent results on this problem in the completely asymmetric case were obtained by T. Simon in \cite{TS-em} and next were applied in \cite{TS-prep1} and \cite{TS-prep2}. On the other hand, some formulas were found by A. Kuznetsov in \cite{AK-ap}, however the final expressions are complicated. M. Kwa\'snicki in \cite{MK} gives an integral representation of the density of $\tau$ in the case of symmetric stable processes ($\beta =0$). In this article we study the exit time $\tau=\tau_{(0,\infty)}$ for 1-dimensional stable processes and give in Theorem \ref{thm:main} a new formula for its Laplace transform. It follows that $\tau$ satisfies some multiplicative convolution relations(Corollary \ref{cor:sploty}); in particular for $\alpha>1$ the exit time $\tau$ is the multiplicative convolution of a $1/\alpha$-stable subordinator with an explicitly given random variable $M_{\alpha,\rho}$. We generalize in this way the result of \cite{TS-em} for all stable processes. Applications of Theorem \ref{thm:main} are next given in the final part of the article. New explicit formulas for the Laplace transform and the density of $\tau$ are proven for the processes dual to those of the Doney's class ${C}_{1,1}$, in particular for the symmetric $\frac23$-stable process (Proposition \ref{cor:Lap} and Corollary \ref{cor:density}). Further applications of Theorem \ref{thm:main} will be presented in a forthcoming paper. The main tool to prove the results of this article is a series representation that we obtained in \cite{PG-TJ-aihp} for the logarithm of the bivariate Laplace exponent $\kappa(\eta,\theta)$ of the ascending ladder process built from the process $X_t$. This application of the series representation of $\ln \kappa$ was announced in \cite{PG-TJ-aihp}. It allows to determine explicitely in Proposition \ref{prop:St} the inverse Stieltjes transform of the function $1/\kappa(1,\theta)$. \section{ Stieltjes transform and Wiener-Hopf factors} In this part of the article we will exploit our series representation of $\kappa(1,\theta)$ from \cite{PG-TJ-aihp} in inverting a Stieltjes transform. Recall that if $ \mu$ is a positive Borel measure on $[0,\infty)$ then for any $x \in (0,\infty)$ the Stieltjes transform of $\mu$ is defined by \begin{equation}\label{eq:ST} \mathcal{S} \mu(\theta) = \int_0^\infty \frac{1}{\theta+x} d\mu(x) \end{equation} whenever the integral converges. According to \cite{CB-LNM}, a function $G$ on $(0,\infty)$ is of the form $ G(\theta)=a+\mathcal{S} \mu(\theta)$ for a positive measure $\mu$ and $ a\ge 0$ if and only if (S1) $G$ extends to a holomorphic function in the cut plane $\mathbb{C}\setminus \mathbb{R}_-$ (S2) $G(\theta)\ge 0$ for $\theta>0$ (S3) $\Im G(z)\le 0$ for $\Im z>0$.\\ Then the inverse Stieltjes transform is $$ \mathcal{S}^{-1}(G)(x)= -(1/\pi)\lim_{y\to 0^+} \Im G(-x+iy)\,, \qquad x>0\,, $$ where the limit, in general, is in vague sense and equals $\mu$. If $\mu$ is absolutely continuous with a continuous density, the limit is equal to the density of $\mu$ for all $x>0$ (\cite{Widder}).\\ Let $\alpha \in (0,2)$ and $(X_t, \mathbb{P}^x)$ be a strictly $\alpha$-stable process in $\mathbb{R}$. By $\kappa_{\alpha,\rho}(\eta,\theta)$ we denote the bivariate Laplace exponent of the ascending ladder process built from $X_t$. We normalize it requiring that $\kappa_{\alpha,\rho}(1,0)=1$. To simplify the notation we will write $\kappa(\eta,\theta)$ for a fixed pair $\alpha,\rho$ (or equivalently a fixed process $X_t$). By $\hat\kappa$ we denote the Laplace exponent for the dual process $\hat X_t=-X_t$. As usually we write the positivity coefficient $$ \rho=\mathbb{P}^0(X_t\ge 0)= \frac{1}{2}+ \frac{1}{\pi \alpha}\arctan\left(\beta \tan\frac{\pi \alpha}{2}\right). $$ \begin{prop}\label{prop:St} For $\rho \in (0, 1] \setminus \{ 1/\alpha \}$ we have \begin{equation}\label{eq:inv_kappa} \frac{\sin(\rho \alpha \pi)}{\pi} \int_0^\infty \frac{1}{x+\theta} \frac{x^{\alpha} \widehat{\kappa}(1,x)}{x^{2\alpha} +2x^\alpha\cos(\rho\alpha\pi) +1}\, dx = \frac{1}{\kappa(1,\theta)}. \end{equation} \end{prop} \begin{proof} Denote $G(\theta)=1/\kappa(1,\theta)$. The function $G(\theta) $ extends to a holomorphic function $h_1(z)$ on $\mathbb{C} \setminus \mathbb{R}_-$ (see \cite{2006-SF-ptrf},(i) p.205). Let $\mathcal{L}$ be the set of Liouville numbers. For $\theta \in (0,1)$ and $\alpha\not\in\mathcal{L} \cup \mathbb{Q}$ we have by \cite{PG-TJ-aihp} \begin{align*} G(\theta) = \exp\left( - \sum_{m=1}^\infty \frac{(-1)^{m+1} \theta^{m}\sin(\rho m \pi)}{m\sin(\frac{m\pi}{\alpha})} - \sum_{k=1}^\infty \frac{(-1)^{k+1} \theta^{\alpha k}\sin(\rho\alpha k\pi)}{k\sin(\alpha k\pi)}\right). \end{align*} The right hand side of the last formula may be extended to a holomorphic function $h_2(z)$ on $\{z \in \mathbb{C} \colon |z| < 1\}\setminus \mathbb{R}_-$ defining $ w^{\alpha} = \exp(\alpha {\rm Log} w) $ where $ {\rm Log} w=\ln|w|+i{\rm Arg} w$, ${\rm Arg} w\in(-\pi,\pi]$, is the principal value of the complex logarithm. We note that $h_1 = h_2$ on $(0,1)$, hence $h_1 = h_2$ on $\{z \in \mathbb{C} \colon |z| < 1\}\setminus \mathbb{R}_-$ and $h_2$ extends to a holomorphic function on $\mathbb{C} \setminus \mathbb{R}_-$, equal for $|z|>1$ to the holomorphic extension of $ \frac{1}{\kappa(1,\theta)}$ for $\theta>1$.\\ In the first part of the proof we will compute $$l(x)= -\frac1\pi \lim_{y\to 0^+ } \Im G(-x+iy)$$ for positive $x$. Denote by $h(z)$ the expression under exponential of $h_2$. Let us compute for $0<x<1$ $$ l(x)= -\frac1\pi \lim_{y\to 0^+ }\Im \exp(h(-x+iy)) = - \frac1\pi e^{\Re(w)} \sin({\Im(w)}), $$ where $$ w=- \sum_{m=1}^\infty \frac{(-1)^{m+1} (-x)^{m}\sin(\rho m \pi)}{m\sin(\frac{m\pi}{\alpha})} -\sum_{k=1}^\infty \frac{(-1)^{k+1} e^{i\alpha k\pi }x^{\alpha k}\sin(\rho\alpha k\pi)}{k\sin(\alpha k\pi)}. $$ The last limit is justified by a standard estimation argument, that implies that in a converging power series one can enter the limit under the series. Moreover, the same argument shows that when $0<x<1$, we have \begin{equation}\label{granica} l(x)= - \frac1\pi\lim_{w\to -x,\Im w>0} \Im G(w)= - \frac1\pi e^{\Re(w)} \sin({\Im(w)}). \end{equation} Now we evaluate \begin{align*} \Re(w)&=\sum_{m=1}^\infty \frac{x^{m}\sin(\rho m \pi)}{m\sin(\frac{m\pi}{\alpha})} -\sum_{k=1}^\infty \frac{(-1)^{k+1} \cos(\alpha k\pi) x^{\alpha k}\sin(\rho\alpha k\pi)}{k\sin(\alpha k\pi)},\\ \Im(w)&=- \sum_{k=1}^\infty \frac{(-1)^{k+1} \sin(\alpha k\pi) x^{\alpha k}\sin(\rho\alpha k\pi)}{k\sin(\alpha k\pi)}= \sum_{k=1}^\infty \frac{(-1)^{k} x^{\alpha k}\sin(\rho\alpha k\pi)}{k}. \end{align*} We will need the following formulas from \cite{2007-ISG-IMR-eap} \begin{equation}\label{eq:sumsin} \sum_{k=1}^\infty \frac{p^k \sin(k\varphi)}{k} = \arctan \frac{p \sin \varphi}{1 - p\cos \varphi}\,,\qquad \varphi \in (0,2\pi),\,p^2 \le 1. \end{equation} \begin{equation}\label{eq:sumcos} \sum_{k=1}^\infty \frac{p^k \cos(k\varphi)}{k} = -\frac{1}{2}\log (1 - 2 p \cos \varphi + p^2)\,,\qquad \varphi \in (0,2\pi),\,p^2 \le 1. \end{equation} Therefore applying a formula $\sin(\arctan u)=\frac{u}{\sqrt{1+u^2}}$ we get \begin{align*} \sin(\Im (w)) & =\sin( \arctan \frac{-x^\alpha \sin (\rho\alpha \pi)}{1 + x^\alpha\cos (\rho\alpha \pi)}) \\ & = \frac{-\frac{x^\alpha \sin (\rho\alpha \pi)}{1 + x^\alpha\cos (\rho\alpha \pi)}}{\sqrt{1 + \left(\frac{x^\alpha \sin (\rho\alpha \pi)}{1 + x^\alpha\cos (\rho\alpha \pi)}\right)^2}} = \frac{-x^\alpha \sin (\rho\alpha \pi)}{\sqrt{x^{2\alpha} + 2 x^\alpha\cos (\rho\alpha \pi) +1}}. \end{align*} Now we compute $\Re(w)$. By (\ref{eq:sumcos}) we get \begin{align*} \Re(w) &= \sum_{m=1}^\infty \frac{(-1)^{m+1}x^{m}\sin((1-\rho) m \pi)}{m\sin(\frac{m\pi}{\alpha})}\\ &+\sum_{k=1}^\infty \frac{(-1)^{k+1} x^{\alpha k}\sin((1-\rho)\alpha k\pi)}{k\sin(\alpha k\pi)} - \sum_{k=1}^\infty \frac{(-1)^{k+1} x^{\alpha k} \cos(\rho\alpha k\pi) }{k}\\ & = \log \widehat{\kappa}(1,x) - \frac{1}{2}\log (1 + 2 x^\alpha \cos(\rho\alpha\pi) + x^{2\alpha}). \end{align*} Hence \begin{align*} -\frac{1}{\pi}e^{\Re(w)}\sin(\Im(w)) &= \frac{\sin (\rho\alpha \pi)}{\pi} \frac{x^\alpha \widehat{\kappa}(1,x) }{x^{2\alpha} + 2 x^\alpha\cos (\rho\alpha \pi) +1}= l(x)>0. \end{align*} By \cite[Lemma 5]{PG-TJ-aihp} we have for $\theta >1$ $$ \kappa(1,\theta) = \theta^{\alpha\rho} \kappa(1,1/\theta) $$ and we use the same method and (\ref{granica}) to obtain \begin{equation}\label{granica2} l(x)= - \frac1\pi\lim_{w\to -x,\Im w>0} \Im G(w)= \frac{\sin (\rho\alpha \pi)}{\pi} \frac{x^\alpha \widehat{\kappa}(1,x) }{x^{2\alpha} + 2 x^\alpha\cos (\rho\alpha \pi) +1} \end{equation} for $x>0, x\not=1$. As the function $l(x)$ is continuous at $x=1$ and by \cite[p.205]{2006-SF-ptrf} the limit $\lim_{w\to -1,\Im w>0} \Im G(w)$ exists, it follows that the convergence in (\ref{granica2}) holds also for $x=1$. Let us now justify the fact that the function $G(\theta)=1/\kappa(1,\theta)$ is a Stieltjes transform of a positive measure $\mu$ on $\mathbb{R}^+$. We will check the conditions (S1-3) given in the beginning of this section. The function $G(\theta)$ is strictly positive for $\theta\in (0,\infty)$ and it extends to a holomorphic function $h_1(z)$ on $\mathbb{C} \setminus \mathbb{R}_-$. Thus the conditions (S1) and (S2) are verified. In order to justify (S3), we use the following property that we proved above: the harmonic function $-(1/\pi)\Im G$ extends continuously to the closed upper half-space $\{\Im z\ge 0\}$ and its boundary values on $\mathbb{R}$ are $ l(-x)>0$ when $x<0$ and $0$ for $x\ge 0$. Taking into account the fact that $\lim_{|z|\to\infty}G(z)=0$ (\cite[p.205]{2006-SF-ptrf}), the maximum principle(\cite[1.10]{ABR}) implies that $ \Im G(z)\le 0$ on $\{\Im z> 0\}$ and (S3) also holds. It follows that for a certain $a\ge 0$ we have $G(\theta)=a+\mathcal{S}(l)(\theta)$. Considering $\theta\to\infty$ we determine $a=0$. Finally consider any $\alpha \in (0,2]$. Since the Lebesgue measure of the set $\mathcal{L} \cup \mathbb{Q}$ is $0$ we can take a sequence $\alpha_n$ tending to $\alpha$. Passing to the limit we obtain (\ref{eq:inv_kappa}) for all $\alpha \in (0,2]$. \end{proof} {\bf Remark}. Other proofs of the fact that $1/\kappa(1,\theta)$ is the Stieltjes transform of a positive measure $\mu$ seem possible, using properties of Bernstein functions (\cite{2010-RS-RS-ZV-gsm}). We deduce immediately from Proposition \ref{prop:St} the following corollary. \begin{cor} For $\rho \in [0, 1) \setminus \{1- 1/\alpha \}$ we have \begin{equation}\label{eq:inv_hkappa} \frac{\sin((1-\rho) \alpha \pi)}{\pi} \int_0^\infty \frac{1}{x+\theta} \frac{x^{\alpha} \kappa(1,x)}{x^{2\alpha} +2x^\alpha\cos((1-\rho)\alpha\pi) +1}\, dx = \frac{1}{\widehat{\kappa}(1,\theta)}. \end{equation} \end{cor} \section{Laplace transform of $\tau$ and applications} The following theorem is the main result of the article. \begin{thm}\label{thm:main} Let $X_t$ be a non-spectrally positive strictly $\alpha$-stable process on $\mathbb{R}$. For any $t>0$ we have \begin{equation}\label{formula:main} \mathbb{E}^1 e^{-t \tau} = \frac{\sin((1-\rho)\alpha\pi)}{\pi} \int_0^\infty e^{-t^{1/\alpha}x} \frac{x^{\alpha-1} \kappa(1,x)}{x^{2\alpha} + 2x^{\alpha}\cos((1-\rho)\alpha\pi) +1}\,dx . \end{equation} \end{thm} {\bf Remark.} Observe that the only case excluded from the Theorem \ref{thm:main} is well known: when $X_t$ is a spectrally positive $\alpha$-stable process starting from $X_0=x$, $1<\alpha<2$, then $(\tau^x_{(0,\infty)})_{x>0}$ is a $1/\alpha$-stable subordinator and $\mathbb{E}^1 e^{-t \tau}= e^{-t^{1/\alpha}}$ (\cite{Bing} p.281). When $X_t$ is spectrally negative, the formula (\ref{formula:main}) was obtained recently by T.Simon(\cite{TS-em}). \begin{proof} We note that if $g(0) = 1$ then \begin{equation}\label{eq:impli} \int_0^\infty \frac{1}{x+t} f(x) dx = g(t)\quad \Longrightarrow \quad\int_0^\infty \frac{1}{x+t} \frac{f(x)}{x} dx = \frac{1}{t} - \frac{g(t)}{t}. \end{equation} Indeed $$ 1 = \int_0^\infty \frac{x+t}{x+t}\frac{f(x)}{x}dx = g(t) + \int_0^\infty \frac{t}{x+t} \frac{f(x)}{x} dx $$ From \cite{2006-EAK-sv} we know that \begin{equation}\label{eq:secondfactor} \int_0^\infty e^{-\theta y}\mathbb{E}^y e^{-\eta \tau}\,dy = \frac{1}{\theta} - \frac{\widehat{\kappa}(\eta,0)}{\theta\widehat{\kappa}(\eta,\theta)}, \end{equation} where $\tau = \tau_{(0,\infty)}$ and $\eta, \theta >0$. Putting $\eta = 1$ and applying (\ref{eq:impli}) to (\ref{eq:inv_hkappa}) we get \begin{align*} &\int_0^\infty e^{-\theta y}\mathbb{E}^y e^{-\tau}\,dy \\ = & \frac{\sin((1-\rho) \alpha \pi)}{\pi} \int_0^\infty \frac{1}{x+\theta} \frac{x^{\alpha-1} \kappa(1,x)}{x^{2\alpha} +2x^\alpha\cos((1-\rho)\alpha\pi) +1}\, dx \\ = & \frac{\sin((1-\rho) \alpha \pi)}{\pi} \int_0^\infty\int_0^\infty e^{-y(x+\theta)} \frac{x^{\alpha-1} \kappa(1,x)}{x^{2\alpha} +2x^\alpha\cos((1-\rho)\alpha\pi) +1}\, dy\, dx \\ = & \int_0^\infty e^{- \theta y} \left(\frac{\sin((1-\rho) \alpha \pi)}{\pi} \int_0^\infty e^{-yx} \frac{x^{\alpha-1} \kappa(1,x)}{x^{2\alpha} +2x^\alpha\cos((1-\rho)\alpha\pi) +1}\, dx \right) dy. \end{align*} Therefore $$ \mathbb{E}^y e^{-\tau}= \frac{\sin((1-\rho) \alpha \pi)}{\pi} \int_0^\infty e^{-yx} \frac{x^{\alpha-1} \kappa(1,x)}{x^{2\alpha} +2x^\alpha\cos((1-\rho)\alpha\pi) +1}\, dx $$ and assertion of the theorem follows from scaling property $\mathbb{E}^y e^{-\tau} = \mathbb{E}^1 e^{-y^\alpha \tau}$ of stable processes. \end{proof} For $\alpha \ge 1$ we immediately obtain from Theorem \ref{thm:main} a formula for the density $h= h_{\alpha,\rho}$ of $\tau$ under $\mathbb{P}^1$. \begin{cor}\label{cor:alpha>1} Let $\alpha \in (1,2)$. The density of $\tau=\tau_{(0,\infty)}$ under $\mathbb{P}^0$ is given by \begin{equation}\label{eq:alpha>1} h(s) = \frac{\sin((1-\rho)\alpha\pi)}{\pi}\int_0^\infty \eta_{1/\alpha}(x,s) \frac{x^{\alpha-1} \kappa(1,x)}{x^{2\alpha} + 2x^{\alpha}\cos((1-\rho)\alpha\pi) +1}\, dx, \end{equation} where $\eta_{\gamma}(t,x)$ is the transition density of $\gamma$-stable subordinator. \begin{proof} Since $\int_0^\infty e^{-xs} \eta_{\gamma}(t,x) dx = e^{-ts^\gamma}$ we obtain by Theorem \ref{thm:main} \begin{align*} &\int_0^\infty e^{-st} h_\alpha(s)\,ds \\ &= \frac{\sin((1-\rho)\alpha\pi)}{\pi}\int_0^\infty \int_0^\infty e^{-st}\eta_{1/\alpha}(x,s) \frac{x^{\alpha-1} \kappa(1,x)}{x^{2\alpha} + 2x^{\alpha}\cos((1-\rho)\alpha\pi) +1}\,ds\,dx\\ &=\int_0^\infty e^{-st} \frac{\sin((1-\rho)\alpha\pi)}{\pi}\int_0^\infty \eta_{1/\alpha}(x,s) \frac{x^{\alpha-1} \kappa(1,x)}{x^{2\alpha} + 2x^{\alpha}\cos((1-\rho)\alpha\pi) +1}\,dx\,ds \end{align*} and the assertion follows. \end{proof} \end{cor} \begin{cor}\label{cor:alpha=1} If $X_t$ is a Cauchy process on $\mathbb{R}$ ($\alpha=1, \rho=1/2$) then the density of the exit time $\tau=\tau_{(0,\infty)}$ under $\mathbb{P}^1$ is given by $$ h(x) =\frac{1}{\pi} \frac{\kappa(1,x)}{x^{2} +1}, \ x\ge 0. $$ \end{cor} \noindent{\bf Remark}. The above formula for $h_{1,1/2}$ was obtained previously by Darling in \cite{1956-DAD-tams} (see also \cite[(7.13)]{KKMS}).\\\\ For $\alpha\not =1$, Theorem \ref{thm:main} gives interesting multiplicative convolution relations verified by $\tau$. We present them in the following subsection. \subsection{Interpretation in terms of multiplicative convolutions} For a given strictly $\alpha$-stable process $X_t$ with $\rho \in [0,1) \setminus \{1-1/\alpha\}$ we define the following function on $\mathbb{R}^+$ $$ m_{\alpha,\rho}(x)=\frac{\sin((1-\rho)\alpha\pi)}{\pi\alpha} \frac{\kappa(1,x^{1/\alpha})}{x^{2} + 2x\cos((1-\rho)\alpha\pi) +1}, \ x\ge 0. $$ Observe that $m_{\alpha,\rho}(x)$ is a probability density on $\mathbb{R}^+$. This follows from the formula $$ \frac{\sin((1-\rho)\alpha\pi)}{\pi} \int_0^\infty \frac{x^{\alpha-1} \kappa(1,x)}{x^{2\alpha} + 2x^{\alpha}\cos((1-\rho)\alpha\pi) +1}\,dx=1 $$ obtained from (\ref{formula:main}) when $t\to 0$ and by a change of variables $x=y^{1/\alpha}$. Denote by $M_{\alpha,\rho}$ a positive random variable with density $m_{\alpha,\rho}$. The variable $M_{\alpha,1/\alpha}$, $1<\alpha<2$ appeared for the first time in \cite{TS-em} for the special case of a completely asymmetric $\alpha$-stable process, $1<\alpha<2$ (in our context a spectrally negative process), when $\kappa(1,x^{1/\alpha})=1+x^{1/\alpha}$. Let $\gamma\in(0,1)$ and $\eta_{\gamma}(t,x)$ be the transition density of $\gamma$-stable subordinator. Denote by $N(\gamma)$ a random variable with the density $\eta_{\gamma}(1,x)$, i.e. $\mathbb{E}\exp(-xN(\gamma))=e^{-x^\gamma}$, $x\ge 0$. Recall that if $Y$ and $Z$ are independent random variables on $(0,\infty)$ with densities $f$ and $g$ respectively, then the multiplicative convolution $Y \times Z^p$ is a random variable with the density \begin{equation}\label{eq:splot} \mathbb{P}[Y\times Z^p\in dt]= \int_0^\infty f(\frac{t}{u^p}) g(u) \frac{du}{u^p}. \end{equation} \begin{cor}\label{cor:sploty} (i) Let $1<\alpha <2$. Suppose that random variables $M_{\alpha,\rho}$ and $N(1/\alpha)$ are independent and that $X_0=1$. We have $$\tau \stackrel{d}{=} M_{\alpha,\rho}\times N(1/\alpha), $$ (ii) Let $0<\alpha<1$. Suppose that random variables $\tau$ and $N(\alpha)$ are independent. We have $$ \tau \times N(\alpha)^\alpha \stackrel{d}{=} M_{\alpha,\rho}. $$ \end{cor} \begin{proof} Part (i) follows immediately from Corollary \ref{cor:alpha>1}. In (\ref{eq:alpha>1}) we substitute $x^\alpha =u$ and use the scaling property $\eta_{1/\alpha}(u^{1/\alpha},s) = u^{-1}\eta_{1/\alpha}(1,su^{-1})$. In order to prove (ii), we use $e^{-ty}=\mathbb{E}[-(ty)^{1/\alpha} N(\alpha)]$ in the left-hand side of (\ref{formula:main}), we apply Fubini and change variables $x=y^{1/\alpha } u$. By unicity of the Laplace transform we get $$ \int_0^\infty h_\alpha(\frac{x^\alpha}{u^\alpha})\eta_{\alpha}(u)\frac{du}{u^\alpha}= \frac{\sin((1-\rho)\alpha\pi)}{\pi\alpha} \frac{\kappa(1,x)}{x^{2\alpha} + 2x^{\alpha}\cos((1-\rho)\alpha\pi) +1} $$ Replacing $x^\alpha$ by $x$ in the last formula and using (\ref{eq:splot}) ends the proof of (ii). \end{proof} \subsection{Application for the Doney's class $\hat C_{1,1}$ } Let $\alpha\in[\frac12,1)$ and $1-\rho = 1/\alpha-1$, i.e. we consider a process dual to a process from the class $C_{1,1}$ from R. Doney's article \cite{1987-RAD-ap}. We denote the class of such processes by $\hat C_{1,1}$. In this case we have by \cite{1987-RAD-ap} or \cite{PG-TJ-aihp} $$\kappa(1,x) = \frac{x^{2\alpha}-2x^\alpha \cos \alpha\pi +1}{1+x}.$$ We recall the definition of the confluent hypergeometric functions $_1F_1$ and $U$ (see \cite{1964-AS-book}) \begin{align*} &_1F_1(a,b,z) = \sum_{k=0}^\infty \frac{(a)_k z^k}{(b)_k k!}\,, \quad z \in \mathbb{C},\\ &U(a,b,z)\\ &= \frac{\pi}{\sin \pi b}\left(\frac{_1F_1(a,b,z)}{\Gamma(1+a-b)\Gamma(b)} - z^{1-b} \frac{_1F_1(1+a-b,2-b,z)}{\Gamma(a)\Gamma(2-b)} \right), \quad z \in \mathbb{C} \setminus \mathbb{R}_-, \end{align*} where $(a)_k = a(a+1)\ldots(a+k-1)$, $(a)_0=1$ is Pochhammer symbol. Using Theorem \ref{thm:main} we get the following formulas for Laplace and Stieltjes transforms of $\tau$ \begin{prop}\label{cor:Lap} If $\alpha\in[\frac12,1)$ and $X_t\in \hat C_{1,1}$ then \begin{align} (i) \quad &\mathbb{E}^1 e^{-\tau t}= \frac{\sin(\alpha\pi)}{\pi} \Gamma(\alpha)\Gamma(1-\alpha, t^{1/\alpha})e^{t^{1/\alpha}}, \label{eq:Lap1}\\ (ii) \quad &\mathbb{E}^1 \frac{1}{x+\tau} = \int_0^\infty e^{-u} \left(\frac{e^{u^{1/\alpha}x^{-1/\alpha}}}{x} - \frac{u^{1/\alpha-1}}{x^{1/\alpha}} \frac{_1F_1(1,2-\alpha,u^{1/\alpha}x^{-1/\alpha})}{\Gamma(2-b)} \right)du, \label{eq:Lap2} \end{align} where $\Gamma(a,z) = \int_z^\infty t^{a-1}e^{-t}\,dt$ is the incomplete Gamma function. \end{prop} \begin{proof} To prove $(i)$ we use simple transformations of integrals \begin{align} \mathbb{E}^1 e^{-\tau t} &= \frac{\sin(\alpha(1/\alpha-1)\pi)}{\pi} \int_0^\infty \frac{e^{-t^{1/\alpha}x} x^{\alpha-1} (x^{2\alpha}-2x^\alpha \cos \alpha\pi +1)}{(1+x)(x^{2\alpha} + 2x^{\alpha}\cos(\alpha(1/\alpha-1)\pi) +1)}\,dx \nonumber\\ & = \frac{\sin(\alpha\pi)}{\pi} \int_0^\infty \frac{e^{-t^{1/\alpha}x} x^{\alpha-1}}{1+x}\,dx \label{eq:C11}\\ & = \frac{\sin(\alpha\pi)}{\pi} \int_0^\infty\int_0^\infty e^{-t^{1/\alpha}x} e^{-s(x+1)}x^{\alpha-1}\,ds\,dx \nonumber\\ & = \frac{\sin(\alpha\pi)}{\pi} \int_0^\infty e^{-s} \frac{\Gamma(\alpha)}{(t^{1/\alpha}+s)^{\alpha}}\,ds = \frac{\sin(\alpha\pi)\Gamma(\alpha)}{\pi} \int_{t^{1/\alpha}}^\infty e^{t^{1/\alpha}-r} \frac{1}{r^{\alpha}}\,dr \nonumber\\ & = \frac{\sin(\alpha\pi)}{\pi} \Gamma(\alpha)\Gamma(1-\alpha, t^{1/\alpha})e^{t^{1/\alpha}}.\nonumber \end{align} To prove $(ii)$ we use the following integral representation of $U$ (see \cite{1964-AS-book}) $$ U(a,b,z) = \frac{1}{\Gamma(a)}\int_0^\infty e^{-zt} t^{a-1}(1+t)^{b-a-1}\,dt, \qquad \RE z >0. $$ Applying this to (\ref{eq:C11}) we get \begin{align*} \mathbb{E}^1 \frac{1}{x+\tau} &= \frac{\sin(\alpha\pi)}{\pi} \int_0^\infty\int_0^\infty e^{-xt} e^{-t^{1/\alpha}s} \frac{s^{\alpha-1}}{1+s}\,ds\,dt \\ &= \frac{\sin(\alpha\pi)}{\pi} \int_0^\infty\int_0^\infty e^{-u} e^{-u^{1/\alpha}x^{-1/\alpha}s} \frac{s^{\alpha-1}}{x(1+s)}\,ds\,du \\ & = \frac{\sin(\alpha\pi)}{\pi} \int_0^\infty e^{-u} \Gamma(\alpha) \frac{U(\alpha,\alpha, u^{1/\alpha}x^{-1/\alpha})}{x}\,du. \end{align*} Hence \begin{align*} \mathbb{E}^1 \frac{1}{x+\tau} &= \int_0^\infty \frac{e^{-u}}{x} \left(_1F_1(\alpha,\alpha,u^{1/\alpha}x^{-1/\alpha}) - \frac{u^{1/\alpha-1}}{x^{1/\alpha-1}} \frac{_1F_1(1,2-\alpha,u^{1/\alpha}x^{-1/\alpha})}{\Gamma(2-b)} \right)\,du\\ &= \int_0^\infty e^{-u} \left(\frac{e^{u^{1/\alpha}x^{-1/\alpha}}}{x} - \frac{u^{1/\alpha-1}}{x^{1/\alpha}} \frac{_1F_1(1,2-\alpha,u^{1/\alpha}x^{-1/\alpha})}{\Gamma(2-b)} \right)\,du. \end{align*} \end{proof} It is possible to invert Stieltjes transform in (\ref{eq:Lap2}) and the resulting density of $\tau$ is given by a series either in $x$ or in $1/x$ (cf. \cite{AK-ap}). For $\alpha=2/3$ the process $X_t$ is symmetric and the density of $\tau$ has a nice integral representation involving hypergeometric function $$ _1F_2(a; b,c; z) = \sum_{k=0}^\infty \frac{(a)_k z^k}{(b)_k(c)_k k!}, \qquad z \in \mathbb{C}. $$ \begin{cor}\label{cor:density} Let $X_t$ be a symmetric $\frac23$-stable process on $\mathbb{R}$ with $X_0=1$. Then the density of $\tau=\tau_{(0,\infty)}$ is given by the formula \begin{equation}\label{form23} \mathbb{P}^1(\tau \in dx)= \frac{1}{\pi} \int_0^\infty e^{-tx} \left(\sin(t^{3/2}) + \frac{t^{1/2}{_1F_2}(1;2/3,7/6;-t^3/4)}{\Gamma(4/3)} \right)\,dt \end{equation} and its Laplace transform is \begin{equation*} \mathbb{E}^1 e^{-t \tau} = \frac{\sqrt{3}}{2\pi}\Gamma(2/3)\Gamma(1/3, t^{3/2})e^{ t^{3/2}} \,, \qquad t >0. \end{equation*} \end{cor} \begin{proof} The second part is a direct consequence of (\ref{eq:Lap1}). By (\ref{eq:Lap2}) \begin{align*} \mathbb{E}^1 \frac{1}{x+\tau} &= \int_0^\infty e^{-u} \left(\frac{e^{u^{3/2}x^{-3/2}}}{x} - \frac{u^{1/2}}{x^{3/2}} \frac{_1F_1(1,4/3,u^{3/2}x^{-3/2})}{\Gamma(4/3)} \right)\,du. \end{align*} Inverting Stieltjes transform we get \begin{align*} &\mathbb{P}^1(\tau \in dx) \\ &= -\frac{1}{\pi} \IM \int_0^\infty e^{-u} \left(\frac{e^{u^{3/2}x^{-3/2}e^{-3i\pi/2}}}{-x} - \frac{u^{1/2}}{x^{3/2}e^{3i\pi/2}} \frac{_1F_1(1,4/3,u^{3/2}x^{-3/2}e^{-3i\pi/2})}{\Gamma(4/3)} \right)\,du \\ &= -\frac{1}{\pi} \IM \int_0^\infty e^{-u} \left(\frac{e^{u^{3/2}x^{-3/2}i}}{-x} - i\frac{u^{1/2}}{x^{3/2}} \frac{_1F_1(1,4/3,u^{3/2}x^{-3/2}i)}{\Gamma(4/3)} \right)\,du. \\ \end{align*} Since \begin{align*} \IM i _1F_1(1,4/3,y i) &= \RE \sum_{k=0}^\infty \frac{(1)_k(iy)^k}{k!(4/3)_k} = \sum_{k=0}^\infty \frac{(-y^2)^k}{(4/3)_{2k}}\\ &= \sum_{k=0}^\infty \frac{(1)_k(-y^2)^k}{k! (4/6)_k (7/6)_k 2^{2k}} = {_1F_2}(1; 2/3,7/6; -y^2/4) \end{align*} we finally get \begin{align*} &\mathbb{P}^1(\tau \in dx) \\ &= \frac{1}{\pi} \int_0^\infty e^{-u} \left(\frac{\sin(u^{3/2}x^{-3/2})}{x} + \frac{u^{1/2}}{x^{3/2}} \frac{_1F_2(1;2/3,7/6;-u^3x^{-3}/4)}{\Gamma(4/3)} \right)\,du \\ &= \frac{1}{\pi} \int_0^\infty e^{-tx} \left(\sin(t^{3/2}) + \frac{t^{1/2}{_1F_2}(1;2/3,7/6;-t^3/4)}{\Gamma(4/3)} \right)\,dt . \end{align*} \end{proof} {\bf Remark}. It is possible to obtain the formula (\ref{form23}) from the results of M. Kwa\'snicki \cite{MK}. However, in order to do this, one has to make several non-elementary transformations of integrals and our approach seems simpler than using \cite{MK}.\\ {\bf Acknowledgement}. We thank Christian Berg for discussions on the Stieltjes transform and Thomas Simon for helpful comments and bibliographical indications. \bibliographystyle{abbrv}

https://arxiv.org/abs/2109.08378

Maximum-Rate Optimization of Hybrid Intelligent Reflective Surface and Relay Systems

We consider a wireless communication system, where a transmitting source is assisted by both a reconfigurable intelligent reflecting surface (IRS) and a decode-and-forward half-duplex relay (hybrid IRS-relay scheme) to communicate with a destination receiver. All devices are equipped with multiple antennas, and transmissions occur in two stages. In stage 1, the source splits the transmit message into two sub-messages, transmitted to the destination and the relay, respectively, using block diagonalization to avoid interference. Both transmissions will benefit from the IRS. In stage 2, the relay re-encodes the received sub-message and forwards it (still through the IRS) to the destination. We optimize power allocations, beamformers, and configurations of the IRS in both stages, in order to maximize the achievable rate at the destination. We compare the proposed hybrid approach with other schemes (with/without relay and IRS), and confirm that high data rate is achieved for the hybrid scheme in case of optimal IRS configurations.

\section{Introduction} A reconfigurable \ac{irs} is a programmable metasurface that can alter the phase and amplitude of an impinging signal by dynamically adjusting the reflection coefficients of its elements. Recently, \acp{irs} have drawn enormous research interest as a promising technology for the \ac{6g} of cellular networks \cite{9387701}, due to their ability of controlling the wireless propagation environment. Before the advent of \acp{irs}, relays have been studied and used in cellular networks to increase coverage and improve the received signal quality. Among various solutions, \ac{df} relays are \ac{hd} devices that alternate two stages, one wherein they receive a message from the source, and a second wherein they re-encode the message and transmit it to the destination. The alternative use of \acp{irs} and relays has been widely investigated. In \cite{Bjornson19}, \acp{irs} and single-antenna \ac{df} relays are compared in terms of power consumption, whereas in \cite{Huang19} the energy efficiency of systems using \acp{irs} is compared to a system with multi-antenna \ac{af} relays. A comparison between \acp{irs} and \ac{df} \ac{hd}/\ac{fd} relays is presented in \cite{DiRenzo20}, proving that sufficiently large \acp{irs} yield higher spectral and energy efficiency than relay-aided systems. Nevertheless, due to the expensive deployment of \acp{irs}, {\em hybrid \ac{irs}-relay systems}, wherein both devices are jointly adopted, will be a cost-effective solution for the near future of smart electromagnetic environments. In \cite{Abdullah20}, the combination of a \ac{hd} \ac{df} relay and an \ac{irs} is investigated and tight upper bounds for the \ac{ar} are derived. A hybrid system with a \ac{fd} \ac{df} relay is studied in \cite{Abdullah21}, showing that the performance further improves, as long as the relay self-interference is low. However, both works consider source and destination equipped with a single antenna each. In \cite{9335947}, a system wherein an \ac{irs} assists both a relay and a destination (and the source has no direct link with either the relay and the destination) is considered, with source, relay, and destination again having all one antenna each. For a system with multiple relays, still in the presence of an \ac{irs}, the selection of one relay to assist communication between a source and a destination is solved by machine-learning in \cite{9344820}. In this paper, we consider a hybrid \ac{irs}-relay \ac{mimo} system, which generalizes the systems considered in \cite{Abdullah20,Abdullah21}, and \cite{9335947}, as we now assume that all devices are equipped with multiple antennas. Moreover, contrary to \cite{9335947}, we also consider the link between the source and the relay. The relay is \ac{hd} and operates in the \ac{df} mode. We propose a transmission protocol operating in two stages. In stage 1, the source splits the transmit message into two sub-messages, transmitted to the destination and the relay, respectively, using \acl{bd} to avoid interference. Both transmissions will benefit from the \ac{irs}. In stage 2, the relay re-encodes the received sub-message and forwards it (still through the \ac{irs}) to the destination. We optimize power allocations, beamformers, and configurations of the \ac{irs} in both stages, to maximize the \ac{ar} at the destination. In particular, we split the \ac{ar} optimization problem into two sub-problems, one for each stage, then coupled by the choice of the \ac{irs} configuration and the power split between the signal for the relay and the destination in stage 1. Lastly, we compare the proposed hybrid approach with other schemes (with/without relay and \ac{irs}), and confirm that a high data rate is achieved for the hybrid scheme in case of optimal \ac{irs} configuration. The rest of this paper is organized as follows. Section II describes transmission characteristics and the two-stage protocol. In Section~III we formalize the maximum-rate optimization problem and describe the alternating optimization solution. In Section~IV we discuss numerical results before the conclusions are taken in Section V. {\it Notation:} Scalars are denoted by italic letters, vectors and matrices by boldface lowercase and uppercase letters, respectively, and sets are denoted by calligraphic uppercase letters. $\diag(\bm{a})$ indicates a square diagonal matrix with the elements of $\bm{a}$ on the principal diagonal. $\bm{A}^H$ denotes the conjugate transpose of matrix $\bm{A}$. $\mathbb{E}[\cdot]$ denotes the statistical expectation. $\bm{I}_x$ is the identity matrix of size $x$. \section{System Model} \begin{figure} \centering \includestandalone[mode=buildnew, width=\columnwidth]{figures/systemmodel_SUMIMO} \caption{Two-stage \ac{irs}- and relay-assisted \ac{mimo} system. Solid arrows represent the S-D link in stage 1, dashed arrows the S-R link in stage 1, and dotted arrows the R-D link in stage 2.} \label{fig:system_model} \end{figure} We consider the narrowband single-user \ac{mimo} communication system shown in Fig.~\ref{fig:system_model}, wherein the transmission from a source (S) to a destination (D) is assisted by both a relay (R) and an \ac{irs} (I). We assume that S and R have maximum transmit powers $P_{\rm S}$ and $P_{\rm R}$, respectively. S, D, and R are equipped with \acp{ula} with $N_{\rm S}$, $N_{\rm D}$, and $N_{\rm R}$ antennas, respectively, whereas I is a \ac{upa} with $N_{\rm I}$ passive reflective elements. We denote with $\bm{H}_{\rm SI} \in \mathbb{C}^{N_{\rm I} \times N_{\rm S}}$ and $\bm{H}_{\rm SR} \in \mathbb{C}^{N_{\rm R} \times N_{\rm S}}$ the S-R and S-I channels, with $\bm{H}_{\rm RI} \in \mathbb{C}^{N_{\rm I} \times N_{\rm R}}$ and $\bm{H}_{\rm RD} \in \mathbb{C}^{N_{\rm D} \times N_{\rm R}}$ the R-I and R-D channels, and with $\bm{H}_{\rm IR} = \bm{H}_{\rm RI}^T$ and $\bm{H}_{\rm ID} \in \mathbb{C}^{N_{\rm D} \times N_{\rm I}}$ the I-R and I-D channels. We consider narrowband mmWave channels \cite{Rappaport2014}, each having $M$ \ac{nlos} components. Hence, channel matrix $\bm{H}_{{\rm XY}}$ between transmitter $\rm{X}$ and receiver $\rm{Y}$ is \begin{equation}\label{channelmatrix} \bm{H}_{\rm{XY}} = \frac{1}{\sqrt{M}}\sum_{m=1}^{M} g_m\rho(d)\bm{a}\left(\bm{\omega}_{{\rm X},m}\right)\bm{a}^H\left(\bm{\omega}_{{\rm Y},m}\right), \end{equation} where $g_m \sim \mathcal{CN}(0, 1)$ is the gain of the $m$-th path, $\rho(d)$ is the path loss attenuation factor, with $d$ being the distance between X and Y, $\bm{\omega}_{\cdot,m}=\left(\xi_{\cdot,m}, \psi_{\cdot,m}\right)$ is the vector of azimuth ($\xi_{\cdot,m}$) and elevation ($\psi_{\cdot,m}$) angles, and $\bm{a}\left(\bm{\omega}_{\cdot,m}\right)=\left(1,\ldots,e^{j\pi[x \sin(\psi_{\cdot,m})\cos(\xi_{\cdot,m})+y \sin(\psi_{\cdot,m})\sin(\xi_{\cdot,m})]},\ldots \right)^{T}$ is the array response vector for the $m$-th path, with $1 \leq x \leq N_{\rm{X}}-1$ and $1 \leq y \leq N_{\rm{Y}}-1$. We assume all devices operate in \ac{hd} and have perfect channel state information. \subsection{\ac{irs} Model} Each element of the \ac{irs} acts as an omnidirectional antenna element that captures and reflects signals, introducing an attenuation and a phase shift on the baseband-equivalent signal. Following the model of \cite{Abeywickrama20}, we denote with $\phi_n=A_n(\theta_n)e^{j\theta_n}$ the reflection coefficient of the $n$-th \ac{irs} element, where $\theta_n \in [-\pi,\pi)$ is the induced phase shift and $A_n^2(\theta_n) \in [0,1]$ is the corresponding power attenuation factor. Indicating with $\bm{x} \in {\mathbb C}^{1\times N_{\rm I}}$ the impinging signal on the \ac{irs}, the reflected signal $\bm{y} \in {\mathbb C}^{1\times N_{\rm I}}$ is $\bm{y} = \bm{\Phi}\bm{x}$, with $\bm{\Phi}=\diag(\phi_1,\ldots,\phi_{N_{\rm I}})$, which is the \ac{irs} reflection matrix, also denoted {\em \ac{irs} configuration}. We consider the realistic baseband-equivalent model of the \ac{irs} described in \cite{Abeywickrama20}, where \begin{equation} A_n(\theta_n) = (1-A_{\rm \min})\left(\frac{\sin{\theta_n-\zeta}+1}{2}\right)^\nu + A_{\rm \min}, \end{equation} with $A_{\rm \min} \geq 0$, $\zeta \geq 0$, and $\nu \geq 0$ being \ac{irs}-specific parameters, assumed to be identical for all \ac{irs} elements. The phase shifts $\theta_n$ are controllable, thus indirectly controlling also the attenuations. Moreover, since continuous-phase shifts are hardly implementable \cite{Tan2016}-\cite{Tan2018}, we assume that the phase shifts are chosen from a discrete set $\mathcal{F}_\theta = \left\{0, \frac{2\pi}{2^b},\ldots,\frac{2\pi(2^b-1)}{2^b} \right\}$, where $b>0$ is the \ac{irs} phase shift resolution, i.e., the number of bits employed to control the phase shifts. The source has full control of the phase shifts, which can be optimized together with beamforming. \subsection{Two-stage Communication Protocol} For a \ac{hd} \ac{df} relay, signal reception and transmission have to occur in two stages, here assumed to be of the same duration. \paragraph*{Stage 1} S splits the message into two sub-messages, and encodes/modulates them into the two signals $\bm{x}_{\rm SR}$ and $\bm{x}_{\rm SD}$, intended for R and D, respectively. The two signals are precoded with \ac{bd} precoders $\bm{B}_{\rm SR}$ and $\bm{B}_{\rm SD}$ before transmission, such that they are received only at the indented destination, without mutual interference. The signal transmitted by S is thus \begin{equation} \bm{s} = \bm{B}_{\rm SR} \bm{x}_{\rm SR} + \bm{B}_{\rm SD} \bm{x}_{\rm SD}. \end{equation} Then, for a given IRS configuration $\bm{\Phi}_1$, the received signals at R and D are, respectively, \begin{equation} \label{yR1} \bm{y}_{R,1} = \underbrace{(\bm{H}_{\rm SR}+ \bm{H}_{\rm IR}\bm{\Phi}_1\bm{H}_{\rm SI})\bm{B}_{\rm SR}}_{\Tilde{\bm{H}}_{\rm SR}(\bm{\Phi}_1)} \bm{x}_{\rm SR} + \bm{n}_{R,1}, \end{equation} \begin{equation} \label{yD1} \bm{y}_{D,1} = \underbrace{\bm{H}_{\rm ID}\bm{\Phi}_1\bm{H}_{\rm SI}\bm{B}_{\rm SD} }_{\Tilde{\bm{H}}_{\rm SD}(\bm{\Phi}_1)}\bm{x}_{\rm SD} + \bm{n}_{D,1}, \end{equation} where $\Tilde{\bm{H}}_{\rm SR}(\bm{\Phi}_1)$ ($\Tilde{\bm{H}}_{\rm SD}(\bm{\Phi}_1)$) is the S-R (S-D) equivalent channel matrix (we highlight their dependency on the \ac{irs} configuration), and $\bm{n}_{R,1} \sim \mathcal{CN}\left(0,\sigma^2\bm{I}_{N_{\rm R}}\right)$ ($\bm{n}_{D,1} \sim \mathcal{CN}\left(0,\sigma^2\bm{I}_{N_{\rm D}}\right)$) is the complex Gaussian noise vector at R (D). \paragraph*{Stage 2} S remains silent, while R decodes the sub-message received by S in stage 1 and re-encodes/re-modulates it into the signal $\bm{x}_{\rm RD}$. Then, R transmits $\bm{x}_{\rm RD}$ to D with the \ac{irs} using a new configuration $\bm{\Phi}_2$. D receives the signal vector \begin{equation} \label{yD2} \bm{y}_{D,2} = \underbrace{(\bm{H}_{\rm RD}+\bm{H}_{\rm ID}\bm{\Phi}_2\bm{H}_{\rm RI})}_{\Tilde{\bm{H}}_{\rm RD}(\bm{\Phi}_2)}\bm{x}_{\rm RD} + \bm{n}_{D,2}, \end{equation} where $\Tilde{\bm{H}}_{\rm RD}(\bm{\Phi}_2)$ is the R-D equivalent channel matrix, and $\bm{n}_{D,2} \sim \mathcal{CN}\left(0,\sigma^2\bm{I}_{N_{\rm D}}\right)$ is the complex Gaussian noise vectors at $D$. Note that, in both stages, the \ac{irs} configurations $\bm{\Phi}_1$ and $\bm{\Phi}_2$ are provided by S, which has full control of the phase shifts. \section{Maximum-Rate Problem} We now first derive the \ac{ar} and then, we formulate the problem of maximizing the \ac{ar}. \subsection{Achievable Rate} For the first stage, the transmit beamformers $\bm{B}_{\rm SD}$ and $\bm{B}_{\rm SR}$ are chosen such that $\bm{x}_{\rm SR}$ and $\bm{x}_{\rm SD}$ do not generate interference at D and R, respectively. To this end, \ac{bd} is applied (see \cite{Tomasin}), using in general a reduced set of streams for the two links. Let $\bm{H}_{\rm SD} = \bm{U}_{\rm SD}\bm{\Gamma}_{\rm SD} \bm{V}_{\rm SD}$ and the \ac{svd} of $\bm{H}_{\rm SD}$; a subset $\mathcal S_{\rm SD}$ of streams (corresponding to diagonal elements of $\bm{\Gamma}_{SD}$) is selected for transmission to D. The \ac{bd} beamformer for transmission to R is $\bm{B}_{\rm SR} = \bm{N}_{\rm SD}\bm{B}'_{\rm SR}$, where $\bm{N}_{\rm SD}$ collects the columns of $\bm{V}_{\rm SD}$ with indices not in the set $\mathcal S_{\rm SD}$, while $\bm{B}'_{\rm SR}$ is the capacity-achieving precoder for the resulting S-R channel. A similar procedure is applied for the definition of the S-D precoder $\bm{B}_{\rm SD}$, for which $\mathcal S_{\rm SR}$ streams are selected. We must also have $|\mathcal S_{\rm SR}| + |\mathcal S_{\rm SR}| \leq N_{\rm S}$. Lastly, $\bm{x}_{SD}$ and $\bm{x}_{SR}$, are zero-mean complex Gaussian vectors with independent entries of size $|\mathcal S_{\rm SD}|$ and $|\mathcal S_{\rm SR}|$. For the second stage, R applies capacity-achieving precoding, and $\bm{x}_{RD}$ zero-mean complex Gaussian vectors with independent entries of size $N_{\rm R}$. As a result, the S-D \ac{mimo} equivalent channel can be decomposed into $|\mathcal{S}_{\rm SD}|$ independent parallel \ac{awgn} channels with gains $\{\gamma_{\rm SD}(i)\}$.\footnote{$\gamma_{\rm SD}(i)$ is the $i$-th singular value of $\Tilde{\bm{H}}_{\rm SD}(\bm{\Phi}_1)\Tilde{\bm{H}}_{\rm SD}^H(\bm{\Phi}_1)$.} The capacity of the S-D channel is therefore \begin{equation} C_{\rm SD} = \sum_{i \in \mathcal{S}_{\rm{SD}}} \log_2\left[1+\gamma_{\rm SD}(i) \frac{P_{\rm SD}(i)}{\sigma^2}\right], \label{capa} \end{equation} where $P_{\rm SD}(i)$ is the power allocated to channel $i$. Similarly, the S-R and R-D channels can be decomposed into $|\mathcal{S}_{\rm SR}|$ and $|\mathcal{S}_{\rm RD}|$ parallel \ac{awgn} channels, with gains $\{\gamma_{\rm SR}(i)\}$ and $\{\gamma_{\rm RD}(i)\}$, respectively, and the S-R and R-D capacities $C_{\rm SR}$ and $C_{\rm RD}$ can be written as in \eqref{capa}, where subscript SD is replaced by subscripts SR and RD, respectively. The \ac{ar} of the considered two-stage scheme is therefore \begin{equation} C_{\rm HYB} = \frac{1}{2} (C_{\rm SD} + \min\{C_{\rm SR}, C_{\rm RD}\}), \label{CRSR} \end{equation} where the two stages requires twice the time of direct transmission, hence the factor 1/2. Note that for a transmission using only the relay, the \ac{ar} $C_{\rm relay}$ is still given by \eqref{CRSR}, with the \ac{irs} switched off ($A_n(\theta) = 0$, $\forall \theta$). A transmission using only the \ac{irs} can instead be performed in a single stage and the \ac{ar} is $C_{\rm \ac{irs}} = C_{\rm SD}$. In both cases, no \ac{bd} is needed. Note also that IRS- or relay-only transmissions occur if no streams are selected for the S-R or S-D links, i.e., if $|\mathcal{S}_{\rm SR}|=0$ or $|\mathcal{S}_{\rm SD}|=0$, respectively. \subsection{Optimization Problem} With this choice of beamformers, we are left with the problem of optimizing a) the transmit power, b) the \ac{irs} configurations in both stages, and c) the set of streams assigned to R and D in stage 1. The \ac{ar} maximization problem can be formalized as follows: \begin{subequations} \label{maxprobl} \begin{equation} \argmax_{\substack{\bm{\Phi}_1, \bm{\Phi}_2 \\\mathcal{S}_{\rm SD}, \mathcal{S}_{\rm SR}, \mathcal{S}_{\rm RD}\\\left\{P_{\rm SD}(i)\right\},\left\{P_{\rm SR}(j)\right\},\left\{P_{\rm RD}(k)\right\}}} \hspace{-1cm} \left( C_{\rm SD} + C_{\rm SR} \right), \label{target} \end{equation} ~\vspace{-0.4cm} \begin{alignat}{2} & \text{s.t.}\; & & \bm{\Phi}_k=\diag(\phi_{1,k},\ldots,\phi_{N_{\rm I},k}), \quad k=1,2, \label{Cphi1}\\ & & & \phi_{n,k}=A_{n,k}(\theta_{n,k})e^{j\theta_{n,k}}, \quad 1 \leq n \leq N_{\rm I}, \quad k=1,2, \label{con_phi}\\ & & & \theta_{n,k} \in \mathcal{F}_\theta, \quad 1 \leq n \leq N_{\rm I}, \quad k=1,2, \label{Cth2}\\ & & & \sum_{i \in \mathcal{S}_{\rm{SD}}} P_{\rm SD}(i) + \sum_{j \in \mathcal{S}_{\rm{SR}}} P_{\rm SR}(j) \leq P_{\rm S}, \label{CP0}\\ & & & \sum_{k\in \mathcal{S}_{\rm RD}} P_{\rm RD}(k) \leq P_{\rm R}, \label{CP2}\\ & & & \sum_{i \in \mathcal{S}_{\rm{SD}}} P_{\rm SD}(i) + \sum_{j \in \mathcal{S}_{\rm{SR}}}P_{\rm SR}(j) + \sum_{k \in \mathcal{S}_{\rm{RD}}} P_{\rm RD}(k) \leq P_{\rm max}, \label{CP1}\\ & & & C_{\rm SR} \leq C_{\rm RD} \label{consSRRD} \\ & & &\mathcal S_{\rm SD}, \mathcal S_{\rm SR} \in \{1, \ldots, N_{\rm S}\}, \label{constre1}\\ & & & 0<|\mathcal S_{\rm SD}| + |\mathcal S_{\rm SR}| \leq N_{\rm S}. \label{constre2} \end{alignat} \end{subequations} The minimum in \eqref{CRSR} is now reflected by constraint \eqref{consSRRD}. Constraints \eqref{Cphi1}-\eqref{Cth2} are related to the control of \ac{irs} phase shifts, and constraints \eqref{CP0} and \eqref{CP1} are power constraints at the devices, and we added the total power constraint $P_{\max}$. This constraint will make the comparison with schemes using only the \ac{irs} more fair, by imposing $P_{\max}$ the maximum power for S. Constraints \eqref{constre1}-\eqref{constre2} are relative to the stream assignment. \subsection{Alternating Optimization Solution} Notice that constraint \eqref{con_phi} makes the problem non-convex, thus we resort to an alternating optimization solution, where we optimize over the \ac{irs} configurations and stream sets, and for each considered configuration we optimize the transmission powers. For fixed \ac{irs} configurations and stream selections, the optimization problem \eqref{maxprobl} becomes \begin{equation} \argmax_{\substack{\left\{P_{\rm SD}(i)\right\},\left\{P_{\rm SR}(j)\right\},\left\{P_{\rm RD}(k)\right\}}} \hspace{-1cm} \left( C_{\rm SD} + C_{\rm SR} \right), \quad \mbox{s.t. } \eqref{CP0}-\eqref{consSRRD}. \label{optProb} \end{equation} Observe that, due to constraint \eqref{consSRRD}, the problem is still non-convex. However, the powers in stage 1 and stage 2 are coupled only through the constraint \eqref{CP1}. We can decouple the two problems by introducing the auxiliary variable $P_{\rm R, eff}$ such that \begin{equation} \sum_{k \in \mathcal{S}_{\rm{RD}}} P_{\rm RD}(k) = P_{\rm R, eff}, \end{equation} so that the power that can be effectively used by S is, from \eqref{CP0}, \eqref{CP2}, and \eqref{CP1}, as $ P_{\rm S, eff} = \min\{P_{\rm S}, P_{\rm max} - P_{\rm R, eff}\}$. With these new definitions, \eqref{optProb} can be split into the two (coupled) problems, for given $P_{\rm R, eff}$, \begin{subequations} \label{P1} \begin{alignat}{2} C_{\rm RD}^*= & \max_{\left\{P_{\rm RD}(k)\right\}} & \, & C_{\rm RD}, \quad \text{s.t.} \sum_{k \in \mathcal{S}_{\rm{RD}}} P_{\rm RD}(k) = P_{\rm R, eff}, \end{alignat} \end{subequations} and \begin{subequations}\label{opt1} \begin{equation} \argmax_{\substack{\left\{P_{\rm SD}(i)\right\},\left\{P_{\rm SR}(j)\right\}}} \frac{1}{2}\left( C_{\rm SD} + C_{\rm SR} \right), \end{equation} \begin{alignat}{2} & \text{s.t.} \;& & C_{\rm SR} \leq C_{\rm RD}^*, \label{capConstP2} \\ & & & \sum_{i \in \mathcal{S}_{\rm{SD}}} P_{\rm SD}(i) + \sum_{j \in \mathcal{S}_{\rm{SR}}} P_{\rm SR}(j) = P_{\rm S, eff}. \label{PSeffconst} \end{alignat} \label{probopt1} \end{subequations} Note that \eqref{P1} and \eqref{opt1} are convex optimization problems and, therefore, they can be solved in closed-form, as detailed in the next sub-section. Then, we need to optimize the \ac{irs} reflection coefficients $\bm{\Phi}_1$ and $\bm{\Phi}_2$, the stream sets $\mathcal{S}_{\rm SR}$, $\mathcal{S}_{\rm SD}$, and $\mathcal{S}_{\rm RD}$, and the auxiliary variable $P_{\rm R, eff}$, in what turns out to be a non-convex problem. Thus, we resort to the discrete \ac{ga} \cite{Holland1972}, which operates iteratively, solving sub-problems \eqref{opt1} and \eqref{P1} for given \ac{irs} configurations, power $P_{\rm R, eff}$, stream sets $\mathcal{S}_{\rm SR}$, $\mathcal{S}_{\rm SD}$, and $\mathcal{S}_{\rm RD}$ at each iteration. \subsection{Decoupled Problem Solution} \paragraph*{Solution of Problem \eqref{P1}} Since the capacity $C_{\rm SR}$ is upper bounded by $C_{\rm RD}^*$ from \eqref{capConstP2}, we first optimize the transmit powers $\left\{P_{\rm RD}(k)\right\}$ at R, given $P_{\rm R, eff}$. Indeed, \eqref{P1} can be solved via the standard waterfilling algorithm \cite{Tomasin} on channels with gains $\{\gamma_{\rm RD}(i)\}$ and total power $P_{\rm R, eff}$. \paragraph*{Solution of Problem \eqref{opt1}} The Lagrangian function of \eqref{opt1} is (with $\lambda_1$ and $\lambda_2$ multipliers) \begin{align} \mathcal{L} & = \left( C_{\rm SD} + C_{\rm SR} \right)\notag -\lambda_2 \left(C_{\rm SR} -C_{\rm RD}^* + s\right) \\ & - \lambda_1 \left(\sum_{i \in \mathcal{S}_{\rm{SD}}} P_{\rm SD}(i) + \sum_{j \in \mathcal{S}_{\rm{SR}}} P_{\rm SR}(j) - P_{\rm S, eff}\right), \end{align} where $s \geq 0$ is an additional slack variable. Setting to zero the derivative of the Lagrangian function, we obtain the following stationary points \begin{align} P_{\rm SD}(i) = \frac{1}{\ln(2) \lambda_1} - \frac{1}{\gamma_{\rm SD}(i)}, \; P_{\rm SR}(j) = \frac{1}{\ln(2) \lambda_1} - \frac{1}{\gamma_{\rm SR}(j)}, \end{align} with $\lambda_1$ such that \eqref{PSeffconst} is satisfied. Now, letting $s = C_{\rm RD}^* - C_{\rm SR}$, if $s \geq 0$ we have found the optimal solution. If instead $s < 0$, then we must assume $s=0$, i.e., the S-R rate in stage $1$ equals the R-D rate in stage $2$. Consequently, we allocate the minimum power that satisfies this constraint to the S-R link, while all the remaining power is assigned to the S-D link. Hence, we first solve the following problem \begin{equation} \argmin_{\left\{ P_{\rm SR}(j)\right\}} \sum_{j \in \mathcal{S}_{\rm{SR}}} P_{\rm SR}(j), \quad \mbox{\, s.t.\, } C_{\rm SR} =C_{\rm RD}^*, \label{subprob1} \end{equation} with the Lagrangian multipliers method, providing \begin{equation} P^*_{\rm SR}(j) = \left[\left(\frac{2^{C_{\rm RD}^*}}{\prod_{j\in \mathcal{S}_{\rm SR}} \gamma_{\rm SR}(j)}\right)^{\frac{1}{|\mathcal{S}_{\rm SR}|}} - \frac{1}{\gamma_{\rm SR}(j)}\right]^+, \end{equation} where $(x)^+ = x$ if $x \geq 0$, while $(x)^+ = 0$ otherwise. For the obtained optimal powers $P_{\rm SR}(j)^*$, we solve \begin{equation} \argmax_{\left\{P_{\rm SD}(i)\right\}} \, C_{\rm SD}, \quad \mbox{\, s.t.\, } \eqref{PSeffconst}, \label{subprob2} \end{equation} which is similar to \eqref{P1} and can be solved likewise. \section{Numerical Results} In this section, we assess the performance of the proposed protocol. S, R, D, and I have coordinates $(0,0,3)$, $(10,-10,3)$, $(20,0,1.5)$, and $(10,y_{\rm I},3)$~m, respectively (see Fig.~\ref{fig:system_model}), and $y_{\rm I}$ is a parameter to be set. We consider $M=2$ \ac{nlos} components for each mmWave link. S, R, and D are equipped with \acp{ula} of $N_{\rm S}=16$, $N_{\rm R}=8$, and $N_{\rm D}=4$ antennas, respectively, whereas the \ac{irs} is an \ac{upa} with $N_{\rm I}=36$ elements and parameters (see \cite{Abeywickrama20}) $A_{\rm min}=0.2$, $\zeta=0.43 \pi$, and $\nu=1.6$. Angles in the array response vector are chosen according to a uniform random distribution, in particular, $\psi_{\cdot, m} \sim \mathcal{U}[0, 2\pi)$ and $\xi_{{\rm I},m} \sim \mathcal{U}[0, \pi/2)$ for the \ac{irs}, while $\xi_{\cdot, m} = 0$ for other devices with \ac{ula}. The transmit \ac{snr} is $P_{\rm max}/\sigma^2 = 10$ ($10$ dB). The path loss term is modelled as $\rho(d) = K_0 (d/d_0)^{-\alpha/2}$, where $K_0 = \rho(d_0) = 0$ dB is the path loss at the reference distance $d_0=10$~m, and $\alpha=5.76$ is the path loss exponent \cite{Rappaport2013}. We compare five schemes: the proposed optimized hybrid \ac{irs}-relay scheme ({\tt Hyb. Opt.}), a hybrid scheme with random \ac{irs} configuration ({\tt Hyb. Rand.}), a scheme without relay and an optimized \ac{irs} ({\tt \ac{irs} Opt.}), a scheme with a random \ac{irs} ({\tt \ac{irs} Rand.}), and a scheme without \ac{irs} and a relay ({\tt Relay}). \begin{figure} \centering \includegraphics[width=0.95\columnwidth]{figures/b} \caption{\ac{ar} versus $b$, for $P_{\rm S}/P_{\rm max} = 0.5$, $N_I=36$, and $y_{\rm I}=20$~m.} \label{fig:f_of_b} \end{figure} Fig.~\ref{fig:f_of_b} shows the \ac{ar} as a function of the \ac{irs} phase shift resolution $b$ for $y_{\rm I}=20$~m and $P_{\rm S}/P_{\rm max} = 0.5$. For $b=0$ we consider a fixed \ac{irs} configuration with phase shifts $\theta_n = \pi$, $\forall n$, corresponding to the maximum value of $A(\cdot)$. For all schemes, the \ac{ar} saturates with just $b=1$ or $2$ bits per element, thus, as already observed in the literature, a very limited number of configurations are enough to achieve the gains provided by the \ac{irs}. In the following, we will consider $b=2$. The schemes with randomly configured \ac{irs} show a penalty for higher resolution, since configurations with lower gains $A(\cdot)$ are used. \begin{figure} \centering \includegraphics[width=0.95\columnwidth]{figures/PS} \caption{\ac{ar} versus $P_{\rm S}/P_{\rm max}$, for $y_{\rm I}=20$~m, $N_I=36$, and $b=2$.} \label{fig:f_of_P_S} \end{figure} Fig.~\ref{fig:f_of_P_S} shows the \ac{ar} as a function of the fractional available power at S, i.e., $P_{\rm S}/P_{\rm max}$ for $y_{\rm I}=20$~m. The {\tt Hyb. Opt.} scheme provides the highest \ac{ar} for all values of $P_{\rm S}/P_{\rm max}$. Still, for low $P_{\rm S}/P_{\rm max}$, the relay has a considerable fraction of power, thus the {\tt Relay} scheme is close to optimal. Instead, at high $P_{\rm S}/P_{\rm max}$, the constraint on $C_{\rm RD}$ limits the \ac{ar} at the relay, and the {\tt \ac{irs} Opt.} scheme attains higher performance. The {\tt \ac{irs} Rand.} scheme yields very poor performance, due to the absence of the relay and the random configuration of the \ac{irs}. \begin{figure} \centering \includegraphics[width=0.95\columnwidth]{figures/y_I} \caption{\ac{ar} versus $y_{\rm I}$, for $P_{\rm S}/P_{\rm max} = 0.5$, $N_I=36$, and $b=2$.} \label{fig:f_of_y_I} \end{figure} Fig.~\ref{fig:f_of_y_I} shows the \ac{ar} as a function of the \ac{irs} distance $y_{\rm I}$, when $P_{\rm S}/P_{\rm max} = 0.5$. For small $y_{\rm I}$ values, the \ac{irs} link is dominant with respect to the relay link, making the {\tt Hyb. Opt.} scheme transmit exclusively towards the \ac{irs}, thus avoiding the half-rate penalty of the two-stage protocol, and approaching the \ac{ar}. On the other hand, the \ac{irs} assistance becomes marginal as $y_{\rm I}$ grows, resulting in similar performance between {\tt Hyb.} and {\tt Relay} schemes. Finally, Fig.~\ref{fig:f_of_N_I} shows the \ac{ar} as a function of the number of reflecting elements $N_{\rm I}$, for $P_{\rm S}/P_{\rm max} = 0.5$ and $y_{\rm I} = 20$~m. As expected, due to the huge beamforming gain introduced by large \acp{irs}, the \ac{ar} grows with the number of reflecting elements. \begin{figure} \centering \includegraphics[width=0.95\columnwidth]{figures/N_I} \caption{\ac{ar} versus $N_{\rm I}$, for $P_{\rm S}/P_{\rm max} = 0.5$, $y_{\rm I} = 20$~m and $b=2$.} \label{fig:f_of_N_I} \end{figure} \section{Conclusions} In this paper, we considered an hybrid \ac{irs}-relay system, optimizing power allocation, \ac{irs} configurations, and stream sets to maximize the \ac{ar}. Numerical results showed that, in the considered scenarios, large phase-optimized \acp{irs} yield higher \acp{ar} than systems using only either the relay or the \ac{irs}. Indeed, the best performance is achieved by different uses of the relay and the \ac{irs} under different positions of the devices or power split among the source and the relay. This suggests that the proposed hybrid solution, which is able to switch among the various uses, is always advantageous. \bibliographystyle{IEEEtran}

https://arxiv.org/abs/1007.4127

Characteristic Classes and Integrable Systems for Simple Lie Groups

This paper is a continuation of our previous paper \cite{LOSZ}. For simple complex Lie groups with non-trivial center, i.e. classical simply-connected groups, $E_6$ and $E_7$ we consider elliptic Modified Calogero-Moser systems corresponding to the Higgs bundles with an arbitrary characteristic class. These systems are generalization of the classical Calogero-Moser (CM) systems related to a simple Lie groups and contain CM systems related to some (unbroken) subalgebras. For all algebras we construct a special basis, corresponding to non-trivial characteristic classes, the explicit forms of Lax operators and Hamiltonians.

\section{} command!!! \newcommand{\sect}[1]{\setcounter{equation}{0}\section{#1}} \renewcommand{\theequation}{\thesection.\arabic{equation}} \newtheorem{predl}{Proposition}[section] \newtheorem{defi}{Definition}[section] \newtheorem{rem}{Remark}[section] \newtheorem{cor}{Corollary}[section] \newtheorem{lem}{Lemma}[section] \newtheorem{theor}{Theorem}[section] \begin{flushright} ITEP-TH-24/10\\ \end{flushright} \vspace{10mm} \begin{center} {\Large{\bf Characteristic Classes and Integrable Systems for Simple Lie Groups.} }\\ \vspace{5mm} A.Levin \\ {\sf State University - Higher School of Economics, Department of Mathematics, } \\ {\sf 20 Myasnitskaya Ulitsa, Moscow, 101000, Russia } \\ {\em e-mail alevin@hse.ru}\\ M.Olshanetsky\\ { \sf Institute of Theoretical and Experimental Physics, Moscow, Russia} \\ {\em e-mail olshanet@itep.ru}\\ A.Smirnov\\ {\sf Institute of Theoretical and Experimental Physics, Moscow, Russia,}\\ {\em e-mail asmirnov@itep.ru}\\ A.Zotov \\ {\sf Institute of Theoretical and Experimental Physics, Moscow, Russia,}\\ {\em e-mail zotov@itep.ru}\\ \vspace{5mm} \end{center} \begin{abstract} This paper is a continuation of our previous paper \cite{LOSZ}. For simple complex Lie groups with non-trivial center i.e. classical simply-connected groups, $E_6$ and $E_7$ we consider elliptic Modified Calogero-Moser systems corresponding to the Higgs bundles with an arbitrary characteristic class. These systems are generalization of the classical Calogero-Moser (CM) systems related to a simple Lie groups and contain CM systems related to some (unbroken) subalgebras. For all algebras we construct a special basis, corresponding to non-trivial characteristic classes, the explicit forms of Lax operators and Hamiltonians. \end{abstract} \tableofcontents \section{Introduction} \setcounter{equation}{0} This paper is a continuation of our previous paper \cite{LOSZ}. Here we consider concrete implementation of the generic formulae for all simple groups with a non-trivial center. In particular, we find the structure of the unbroken Lie subalgebras $\ti\gg_0$ (see Table I in \cite{LOSZ}). We refer the formulae from \cite{LOSZ} with a number I. The information about roots, weights and so on was taken from \cite{Bou,Jac}. \vspace{0.3cm} \bigskip {\small {\bf Acknowledgments.}\\ The work was supported by grants RFBR-09-02-00393, RFBR-09-01-92437-KEa, NSh-3036.2008.2, RFBR-09-01-93106-NCNILa (A.Z. and A.S.), RFBR-09-02-93105-NCNILa (M.O.) and to the Federal Agency for Science and Innovations of Russian Federation under contract 14.740.11.0347. The work of A.Z. was also supported by the Dynasty fund and the President fund MK-1646.2011.1. A.L and M.O. are grateful for hospitality to the Max Planck Institute of Mathematics, Bonn, where the part of this work was done.} \section{$\SLN$ - the root system $A_{N-1}$} \setcounter{equation}{0} \subsection*{Roots and weights.} For the $A_n$, ($n=N-1$) root system we have two groups $\bar G=\SLN$, $G^{ad}=\PSLN$. Choose the Cartan subalgebra $\gH\subset\gg$ as an subalgebra of traceless diagonal matrices. Then $\gH$ can be identify with the hyperplane in $\mC^N$ $\gH=\{\,\bfx=(x_1,\ldots,x_N)\in\mC^N\,|\,\sum_{j=1}^Nx_j=0\,\}$. The simple roots $\Pi=\{\al_k\}$ $=\{\al_1=e_1-e_2\,\ldots,\al_{N-1}=e_{N-1}-e_N\}$ form a basis in the dual space $\gH^*$. Here $\{e_j\}$ $j=1,\ldots,N$ is a canonical basis in $\mC^N$. The vectors $e_j$ generate the set of roots $R=\{\,(e_j-e_k)\,,~j\neq k\,\}$ of type $A_{N-1}$. The minimal root is \beq{mra} \al_0=-\sum_{\al\in\Pi}\al_k=e_N-e_1\,. \eq It defines the extended Dynkin diagram \vspace{0.5cm} \begin{center} \unitlength 1mm \linethickness{0.6pt} \ifx\plotpoint\undefined\newsavebox{\plotpoint}\fi \begin{picture}(62.142,41.099)(0,0) \put(26.087,9.881){\texttt{Fig.1:} \large A$_{n}$ and action of $\la_{N-1}$} \put(31.008,21.968){\circle{2.531}} \put(38.997,22.074){\circle{2.322}} \put(61.07,22.074){\circle{2.144}} \put(46.039,36.894){\circle*{2.765}} \put(32.479,21.968){\line(1,0){5.3607}} \put(40.363,22.073){\line(1,0){5.781}} \put(55.814,21.968){\line(1,0){4.099}} \multiput(31.744,23.23)(.052555788,.0504702408){252}{\line(1,0){.052555788}} \multiput(47.3,36.263)(.0505344115,-.0513429621){260}{\line(0,-1){.0513429621}} \put(31.744,27.539){\vector(-1,-1){.07}} \multiput(40.994,36.474)(-.052258863,-.05047731){177}{\line(-1,0){.052258863}} \put(40.048,18.92){\vector(1,0){4.625}} \put(56.235,18.92){\vector(1,0){3.574}} \put(51.189,35.948){\vector(-1,1){.07}} \multiput(61.28,25.752)(-.050453556,.050979114){200}{\line(0,1){.050979114}} \put(30.903,15.767){$\alpha_1$} \put(38.996,15.977){$\alpha_2$} \put(61.07,16.082){$\alpha_n$} \put(46.039,41.099){$\alpha_0$} \put(31.008,18.71){\vector(1,0){7.042}} \end{picture} \end{center} Since the half-sum of positive roots is $\rho=\oh(N-1,N-3,\ldots,3-N,1-N)$ and $h=N$, $\,\ka= \frac{\rho}h=\f1{2N}(N-1,N-3,\ldots,3-N,1-N)$. For $\al=e_k-e_{k+a}$ the level (A.3.I) $f_\al=a$ and \beq{rhoa} \lan\ka,\al\ran=a/N \eq (see (A.14.I)). We identify $\gH^*$ and $\gH$ by means of the standard metric on $\mC^N$. Then the coroot system coincides with $R$, and the coroot lattice $Q^\vee$ coincides with $Q$ \beq{rl} Q=\{\sum m_je_j\,|\,m_j\in \mZ\,,~\sum m_j=0\}\,. \eq The fundamental weights $\varpi_k$, $(k=1,\ldots,N-1)$, dual to the basis of simple roots $\Pi^\vee\sim\Pi\,$ $\,(\varpi_k(\al^\vee_k)=\de_{kj})$, are \beq{fw1} \varpi_j=e_1+\ldots+e_j-\frac{j}{N}\sum_{l=1}^Ne_l\,,~~~ \left\{ \begin{array}{l} \varpi_1=(\frac{N-1}N,-\f1{N},\ldots,-\f1{N})\\ \varpi_2=(\frac{N-2}N,\frac{N-2}N,\ldots,-\frac{2}{N}) \\ \ldots \\ \varpi_{N-1}=(\frac{1}N,\frac{1}N,\ldots,\frac{1-N}{N}) \,. \\ \end{array} \right. \eq In the basis of simple roots the fundamental weights are $$ \varpi_k= \f1{N}[(N-k)\al_1+2(N-k)\al_2+\ldots(k-1)(N-k)\al_{k-1} $$ $$ +k(N-k)\al_k+k(N-k-1)\al_{k+1}+\ldots+k\al_{N-1}]\,. $$ Similar to the roots and coroots we identify the fundamental weights and the fundamental coweights. They generate the weight (coweight) lattice \beq{wl} P\subset\gH\,,~P=\{\sum_ln_l\varpi_l\,|\,n_l\in\mZ\}\,,~~{\rm or}~ P=\{\sum_{j=1}^Nm_je_j\,,~m_j\in \f1{N}\mZ\,,~m_j-m_k\in\mZ\}\,. \eq or \beq{gv} P=Q+\mZ\varpi_{N-1}\,,~~(\varpi_{N-1}=-e_N+\f1{N}\sum_{j=1}^Ne_j)\,. \eq \subsection*{Transition matrices.} The factor-group $P^\vee/Q^\vee\sim P/Q$ is isomorphic to the center $\clZ(\SLN)\sim\mu_N$. It is generated by $\zeta=\exp 2\pi i\varpi_{N-1}$. Following Proposition 3.1.I define $\la_{N-1}$ and its action on the extended Dynkin diagram. Consider the fundamental alcove (A.17.I). It follows from (\ref{mra}) that $$ C_{alc}= \{0,\varpi_1,\ldots,\varpi_{N-1}\}\,. $$ Thus, any fundamental weight generates a nontrivial $\La$. For $\xi=\varpi_{N-1}$ in (3.10.I) we find the corresponding transformation \beq{gac1} \la_{N-1}\,:\,e_j\to e_{j+1}\,,~(\al_k\to\al_{k+1})\,. \eq The action of $\la_{N-1}$ on $\Pi^{ext}$ is presented on Fig.1. It means that in the canonical basis in $\mC^N\,$ $\la_{N-1}$ acts as the permutation matrix \beq{lamb} \La=c\left( \begin{array}{ccccc} 0 & 1 & 0 & \ldots & 0 \\ 0 & 0 & 1 & \ldots & 0 \\ \vdots & \ddots & \ddots & \ddots & \vdots \\ 0 & 0 & & 0 & 1 \\ 1& 0 & \ldots & 0 & 0 \\ \end{array} \right)\,,~~~c=\left\{ \begin{array}{cc} 1\,,& N=2m+1\\ \exp\,\frac{\pi i}{2m} \,,& N=2m \end{array}\right. \,. \eq Evidently, that $\La\in\SLN$ and $\La^N\sim Id$. On the other hand \beq{clqu} \clQ=\bfe\,(\ka)=\di (\bfe\left(\frac{N-1}N\right),\bfe\left(\frac{N-3}N\right), \dots,\bfe\left(\frac{1-N}N\right))\,,~~(\bfe(x)=\exp\,(2\pi ix))\,. \eq Thus, $$ [\La,\clQ]=\zeta \,,~~\zeta=\om\cdot Id\,,~\om=\bfe\,(1/N)\,, $$ and $\zeta$ is an obstruction to lift a $\PSLN$-bundle to a $\SLN$-bundle. Since $Ker\,(\la_{N-1}-1)|_\gH=0$ $\ti\gH=\empty$ (see Proposition 3.1.I) the moduli space is empty set (see (3.29.I) and (3.32.I)). Consider another extreme case when $\xi$ lies in the root lattice $Q$. Then $\La=Id$ and $\gH\in Ker\La$. Since $[\La,\clQ]=Id$, the bundles have a trivial characteristic class. Consider an intermediate case and assume that $N$ has nontrivial divisors $N=pl$, $(l\neq 1,N)$. It can be found that \beq{gac} \varpi_{j}\,\to\,\la_j\,:\, e_k\to e_{k+N-j}\,. \eq Then there exists a sublattice $P_l= Q+\mZ\varpi_{N-p}\,$, $\,Q\subset P_l\subset P\,$, such that $P/P_l\sim\mu_l$. It follows from (\ref{gac}) that $\la_{N-p}\,:\,e_j\to e_{j+p}$. Therefore, $\La_{N-p}$ is equal the $p$-th degree of (\ref{lamb}) ($\La_{N-p}^l=Id$) For $\zeta_p=\bfe (\varpi_{N-p})=\om ^p Id\,$, \,($\om^N=1$) (3.3.I) takes the form \beq{co0} [\La_{N-p},\clQ]=\om^p\cdot Id_N\,, \eq where $\clQ$ is (\ref{clqu}). Define the factor group $G_l$ \beq{co1} 1\to\mu_l\to\SLN\to G_l\to 1\,. \eq It has the center $\clZ(G_l)\sim\mu_p=\mZ/p\mZ$. Therefore, $$ 1\to\mu_p\to G_l\to\PSLN\to 1\,. $$ The sublattice $P_l$ is isomorphic the group of cocharcters $t(G_l)$ (A.43.I)0 For the dual group $^LG_l=G_p$ we have the similar exact sequences, where the role of $l$ and $p$ are changed $$ 1\to\mu_p\to\SLN\to G_p\to 1\,, $$ \beq{co2} 1\to\mu_l\to G_p\to\PSLN\to 1\,. \eq It follows from (\ref{co1}) and (\ref{co2}) that the cocycle $H^2(\Si_\tau,\mu_l)$ is obstruction to lift $G_l$-bundles to $\SLN$ bundles or to lift $\PSLN$-bundles to $G_p=^LG_l$-bundles. \subsection*{Bases} \subsubsection*{Sin-basis.} Consider the case $\zeta=\exp(2\pi i\varpi_{N-1})$. Therefore $\La$ is (\ref{lamb}) and all it orbits in $R$ have the same length $N$. The number of orbits is $N-1$. In other words $\sharp R=N(N-1)$. Let us take the root subspaces $E_{1,a}$ $(a\neq 1)$ as representatives of orbits in the space $\gL$. Then the basis (5.4.I) in $\gL$ is $$ \gt^c_a=\f1{\sqrt{N}}\sum_{m=1}^{N}\om^{mc}E_{1+m,a+m}\,,~c=(0,\ldots,N-1)\,,~~\om^N=1\,. $$ In particular, $$ \gt^c_1=\f1{\sqrt{N}} \left( \begin{array}{ccccc} 0 & \om^c & 0 & \ldots & 0 \\ 0 & 0 & \om^{2c} & \ldots & 0 \\ \vdots & \ddots & \ddots & \ddots & \vdots \\ 0 & & & & \om^{(N-1)c} \\ 1 & 0 & \ldots & \ldots & 0 \\ \end{array} \right)\,. $$ There is only one orbit in the Cartan subalgebra $\gH$ under the action of $\la_{N-1}$ $$ e_1\to e_2\to e_3,\to\ldots,e_N,\,. $$ For the $A_{N-1}$ roots it is convenient to pass to the canonical over-complete basis $(e_1,e_2,\ldots,e_N)$. Then the basis (5.33.I) on $\gH$ is $$ \gh^c=\f1{\sqrt{N}}\sum_{m=0}^{N-1}\om^{mc}e_{m+1}=\f1{\sqrt{N}}\di(1,\om^c,\ldots,\om^{(N-1)c}) \,, ~(c=1,\ldots,N-1)\,. $$ Essentially, $(\gt^c_a,\gh^c)$ form a basis of the sin-algebra \cite{FFZ}. If $N$ is a primitive number then the center $\clZ(\SLN)=\mu_N$ has not nontrivial subgroups. Therefore, one can put in (3.9.I) $\xi\in P$. This case leads to the sin-basis. Another options is $\xi\in Q$. In this case $\la=Id$ and we come to the Chevalley basis (see Remark 5.1.I). \subsubsection*{Generalized Sin-Basis in $\SLN$.} Let $N=pl$ and $\xi=\varpi_{N-p}$ generates $\La_{N-p}=\La^p$. All orbits of $\mu_l$ have the length $l$. In the space of bases $\clE=\{E_{jk}\,,$ $j\neq k$ there are $p(N-1)$ orbits passing through the matrices \beq{sa} E_{s,s+a}\,,~~(s=1,\ldots ,p\,,\,a=1,\ldots,N\,,\,\,(mod\,N))\,. \eq The off-diagonal basis is \beq{atk} \gt^c_{s,k}=\f1{\sqrt{l}}\sum_{m=0}^{l-1}\om^{mpc}E_{s+mp,s+k+mp}\,,~~\om^N=1\,, \eq $$ c=0,\ldots,l-1\,,~s=1,\ldots ,p\,,\,k=1,\ldots,N-1\,,(mod\,N))\,. $$ The pairing in this basis (5.8.I) assumes the form \beq{stp} (\gt^{a}_{s_1,k},\gt^{b}_{s_2,j})= \de_{k,-j}\de^{(a,-b\,,mod\,l)}\de_{(s_1+k,s_2+rp)}\om^{rpb}\,,~~(s_1+k-s_2=rp)\,. \eq Let $(e_1,\ldots,e_N)$ be the canonical basis in $\gH$. There are $p$ orbits of length $l$ $$ \begin{array}{l} \ti{e}_1=e_1+e_{p+1}+\ldots+e_{(l-1)p+1} \\ \ti{e}_2=e_2+e_{p+2}+\ldots+e_{(l-1)p+2} \\ \ldots \\ \ti{e}_{p}=e_{p}+e_{2p}+\ldots+e_{lp} \,. \end{array} $$ Let $\ti\gH_0\subset\gH$ be a Cartan subalgebra $$ \ti\gH=\{\ti\bfu=\sum_{j=1}^pu_j\ti{e}_j\,|\,\sum_{j=1}^pu_j=0\}\,. $$ with the basis of simple coroots \beq{slco} \ti\Pi^\vee=\{\ti\al^\vee_k=\ti e_k-\ti e_{k+1}\}\,. \eq The simple roots $$ \ti\Pi=\{\ti\al_k=\f1{l}(\ti e_k-\ti e_{k+1})\} $$ along with $\ti\Pi^\vee$ generate $A_{p-1}$ type Cartan matrix $a_{jk}=\lan\ti\al^\vee_j,\ti\al_k\ran$. The simple coroots and the subspaces \beq{cbp} \ti E_{i,a}=\sum_{m=0}^{l-1}E_{i+mp,i+a+mp}\,,~~(1\leq i\leq p\,,~ a=\pm 1,\pm 2,\ldots\pm(p-1)) \eq form the Chevalley basis in the invariant subalgebra $\ti\gg_0={\bf\rm sl}(p,\mC)$ (see Proposition 5.1). It follows from (\ref{atk}) that the basis (5.29.I) in the space $V$ (5.30.I) takes the form \beq{abv} V=\{\gt^0_{s,a}=\f1{\sqrt{l}}\sum_{m=0}^{l-1}E_{s+mp,s+a+mp}\,, ~(1\leq s\leq p\,,~p\leq a\leq N)\}\,. \eq Together with $\ti\gg_0$, $V$ forms the invariant subalgebra $$ \gg_0=\ti\gg_0=\underbrace{{\bf\rm sl}(p,\mC) \oplus\ldots\oplus{\bf\rm sl}(p,\mC)}_{l} \oplus(\underbrace{\mC\oplus\ldots\oplus\mC)}_{l-1} $$ Its structure is obtained from $\Pi^{ext}$ by dropping $l$ roots $(\al_p,$ $\ldots,\al_{(l-1)p},\al_0)$. This procedure defines an automorphism of $\gg_{A_{N-1}}$ of order $l=N/p$. The Killing form in $\ti\gg_0$ is \beq{siap} (\ti e_k,\ti e_j)=l\de_{kj}\,,~~(\ti E_{i,a},\ti E_{j,b})=l\de_{a,-b}\de_{i+a,j\,,~(mod\,p)} \eq and commutation relations $$ [\ti e_k,\ti E_{i,a}]=(\de_{i,k}-\de_{i-k+a,0\,mod\,p})\ti E_{i,a}\,, $$ $$ [\ti{E}_{i,a},\ti{E}_{j,b}]= \de_{i+a,j\,mod\,p}\ti E_{i,a+b}-\de_{i,j+b\,mod\,p}\ti E_{j,a+b}\,,~(b\neq p-a)\,, $$ $$ [\ti E_{i,a}, \ti{E}_{i+a,p-a}]=\ti e_i-\ti e_j\,. $$ \bigskip The basis in $\gH$ is generated by $\ti e_j$ and by \beq{ahk} \gh_{j}^c=\f1{\sqrt{l}}\sum_{m=0}^{l-1}\om^{mpc}e_{j+mp}\,,~~ c=1,\ldots,l-1\,, ~~j=1,\ldots,p \eq with the pairing \beq{pahk} (\gh_{j}^{c_1},\gh_{k}^{c_2})=\de_{j,k}\de^{c_1,-c_2}\,. \eq The subspace $\gg_c$ in (5.1.I) is formed by the basis $\{\gh_{j}^c ~(\ref{ahk})\,,~~\gt^c_{s,a}~(\ref{atk})\,,~c\neq 0\}$. The general commutation relations in this basis \beq{scr} [\gt^{c_1}_{i,a},\gt^{c_2}_{j,b}]=\f1{\sqrt{l}} \left(\om^{(i-j+a)c_2}\de_{j,i+a\,(mod\,p)}\gt^{c_1+c_2}_{i,a+b}- \om^{(j-i+b)c_1}\de_{i,j+b\,(mod\,p)}\gt^{c_1+c_2}_{j,a+b}\right)\,, \eq \beq{scr1} [\ti e_{i},\gt^{c}_{i,a}]= \de_{j,i}\gt^{c}_{i,a}- \de_{i,j+b\,(mod\,p)}\gt^{c}_{j,b}\,, \eq \beq{scr2} [\gh^{c_1}_{i},\gt^{c_2}_{j,b}]=\f1{\sqrt{l}} \left(\om^{(i-j)c_2}\de_{j,i}\gt^{c_1+c_2}_{i,b}- \om^{(j-i+b)c_1}\de_{i,j+b\,(mod\,p)}\gt^{c_1+c_2}_{j,b}\right)\,, \eq \subsection*{The Lax operators and the Hamiltonians.} \subsubsection*{Trivial bundles} For trivial bundles $\xi\in Q$, $\La=Id$ and $\gH\subset Ker\La$. One can consider trivial $\SLN$ or $\PSLN$ bundles. They differ by their moduli spaces (3.21.I) or (3.22.I). At the case at hand they assume the following form. Let $$ C^+=\{\ti\bfu\in\mC^N\,|\,u_1\geq u_2\geq\ldots\geq u_N\} $$ be a positive Weyl chamber (A.8.I). Then (3.21.I) and (3.22.I) take the form \beq{csca} C^{SL} =\{\bfu\in C^+\,|\, u_j\sim u_j+\tau n_j+m_j\,,~n_j,m_j\in\mZ\,,~\sum_j n_j=\sum_j m_j=0\}\,. \eq \beq{cada} C^{PSL} =\{\bfu\in C^{SL}\,|\,n_j,m_j\in\f1{N}\mZ\,,~ n_j-n_k\in\mZ\,,~m_j-m_k\in\mZ~\}\,. \eq The Lax operator is the SL$(N)$, (or PSL$(N)$) and the Hamiltonian CM system are well-known \cite{GH,LX,Wo} $$ L^{CM}_{SL(N)}=\sum_{j=1}^N(v_j+S_j)e_j+\sum_{j\neq k}S_{jk}\phi(u_k-u_j,z)E_{jk}\,, $$ $$ H^{CM}_{SL(N)}=\oh\sum_{j=1}^Nv_j^2-\sum_{j\neq k}S_{jk}S_{kj}E_2(u_j-u_k)\,. $$ \subsubsection*{Nontrivial bundles} Define a moduli space of bundles with characteristic class $\om^p$ (\ref{co0}). Evidently,\\ $\ti\gH=Ker(\La_{N-p}-1)|_{\ti\gH_0}$. Let $Q_l\subset Q$ be an invariant coroot lattice in $\ti\gH_0\subset{\rm\bf sl}(p,\mC)$ $$ Q^\vee_l=\{\ga=\sum_{j=1}^pm_j\ti e_j\,,~~m_j\in\mZ\,,~\sum_{j=1}^pm_j=0\}\,, $$ generated by the simple coroots (\ref{slco}). It is an invariant sublattice of $Q^\vee$ with respect to the $\La_{N-p}$ action. The fundamental coweights $(\lan\ti\varpi^\vee_j,\ti\al_k\ran=\de_{jk})$ $$ \begin{array}{l} \ti\varpi^\vee_1=\frac{p-1}p\ti e_1-\f1{p}\ti e_2-\ldots \f1{p}\ti e_{p}\,,\\ \ti\varpi^\vee_2=\frac{p-2}p\ti e_1+\frac{p-2}{p}\ti e_2-\ldots -\frac{2}{p}\ti e_{p}\,, \\ \ldots\ldots\\ \ti\varpi^\vee_p=\frac{1}p\ti e_1+\f1{p}\ti e_2+\ldots \frac{1-p}{p}\ti e_{p}\, \end{array} $$ form a basis in the coweight lattice $$ P^\vee_l=\{\ga=\sum _{j=1}^pn_j\ti e_j\,,~~n_j\in\f1{p}\mZ\,,~~ \sum _{j=1}^pn_j=0\,,~~n_j-n_k\in\mZ\}\,. $$ It is an invariant sublattice of $P^\vee$. Let $\ti W$ is a permutation group of $\ti{e}_1,\ldots,\ti{e}_{p}$. It is a Weyl group of $\ti R(\ti\Pi)$. Define the semidirect products (3.28.I), (3.31.I) $$ \ti W_{BS}=\ti W\ltimes(\tau Q^\vee_l+Q^\vee_l)\,,~~~ \ti W^{ad}_{BS}=\ti W\ltimes(\tau P^\vee_l+P^\vee_l)\,. $$ The moduli space is defined as in (3.29.I), (3.31.I) \beq{mscl} C^{(l) \,sc}=\ti\gH/ \ti W_{BS}\,,~~~ C^{(l) \,ad}=\ti\gH/\ti W^{ad}_{BS}\,. \eq Thus, for $$ C^{(l) \,sc}\,:\, u_j\sim u_j+\tau m_j+n_j\,,~n_j,m_j\in\mZ\,,~ \sum_{j=1}^p n_j=\sum_{j=1}^p m_j=0\,, $$ and for $$ C^{(l) \,ad}\,:\, u_j\sim u_j+\tau m_j+n_j\,,~~~~n_j,m_j\in\f1{p}\mZ\,, $$ $$ \sum_{j=1}^p n_j=\sum_{j=1}^p m_j=0\,,~~~n_j-n_k\in\mZ\,,~~~m_j-m_k\in\mZ\,. $$ Define $\ti\bfu\in\gH$ such that \beq{u} u_{i}=u_s~{\rm if}~i=s+mp\,. \eq It means that $\ti\bfu\in\ti\gH_0$, and $\ti\bfu=\sum_{i=1}^pu_i\ti e_i$. The Lax operator $L=\ti L_0+L'_0+\sum_{a=1}^{l-1}L_a$ (6.16.I), (6.17.I), (6.18.I) takes the form, (see (\ref{rhoa})) $$ L_a(z)=\sum_{j=1}^pS^{a}_{j}\phi(\frac{a}l,z)\gh^a_{j}+ \sum_{i,k=1}^NS^{a}_{i,k}\bfe(zk/N) \phi(u_{i+k}-u_i+\tau k/N+\frac{a}l,z)\gt^a_{i,k}\,, $$ $$ \ti L_0(z)=\sum_{i=1}^p(v_{i}+ S_{i}E_1(z))\ti e_{i}+ \sum_{i,k}\ti S_{i,k}\bfe(zk/N) \phi(u_{i+k}-u_i+\tau k/N,z)\ti E_{i,k}\,, $$ \beq{sLM} L'_0(z)= \sum_{i,k}S^{'}_{i,k}\bfe(zk/N) \phi(u_{i+k}-u_i+\tau k/N,z)\gt^0_{i,k}\,. \eq To come to an integrable system impose the moment constraints $\ti S^{\ti\gH}_{i}=0$ $(i=1,\dots,p)$ and the gauge fixing constraints. The calculation of the quadratic Hamiltonian is based on the pairing relations (\ref{stp}), (\ref{pahk}), (\ref{siap}). Using (5.1.I) and (\ref{stp}) we come to the Hamiltonian $$ H=\ti H_0+H_0'+\sum_{k=1}^{[l/2]}H_k\,. $$ Here $\ti H_0$ is the sl$(p)$ Calogero-Moser Hamiltonian $$ \ti H_0=l\Bigl(\frac{1}2\sum_{i=1}^pv_i^2-\sum_{i=1}^p\sum_{k=1}^{p-1}\ti S_{i,k}\ti S_{i+k,-k}E_2(u_i-u_{i+k})\Bigr)\,. $$ For $H_a$ and $H_0'$ we have $$ H_0'=-\oh\sum_{i=1}^p\sum_{k=p+1}^NS^0_{i,k}S^{0}_{i+k+rp,-k} E_2(u_{i+k}-u_i+\tau k/N)\,, $$ $$ H_a=-\frac{1}2\sum_{s=1}^p\Bigl( S^a_{s}S^{-a}_{s}E_2(a/l)+\sum_{k=1}^N\om^{-aip} S^{a}_{s,k}S^{-a}_{s+k+rp,-k} E_2(u_{s+r}-u_s+a/l+\tau k/N)\Bigr)\,. $$ These Hamiltonians and the corresponding Lax operators were obtained in \cite{LZ}. The basis, used there is not the GS basis. The corresponding Hamiltonians describe $p$ interacting EA tops with inertia tensors depending on $\ti\bfu$. This interpretation is specific for $\SLN$ $(N=pl)$. In particular, if $\xi=\varpi_{N-1}$, $\ti\gH_0=\emptyset$ and we deal with the sin-basis. The corresponding bundle has no moduli ($p=1$, $l=N$). The invariant Hamiltonian $\ti H_0=0$ and $H_0'$, $H_a$ describe the so-called the Elliptic top \cite{KLO,RSTS}. \section{SO$(2n+1)$\,, Spin$(2n+1)\,$, $\,B_n$ root system.} \setcounter{equation}{0} \subsection*{Roots and weights.} For the Lie algebra $B_n$ the universal covering group $\bG$ is Spin$(2n+1)$ and $G^{ad}=$SO$(2n+1)$. The simple roots of $B_n$ are \beq{bsr1} \Pi_{B_n}=\{\al_j=e_j-e_{j+1}\,,~j=1,\dots,n-1\,,~\al_n=e_n\}\,., \eq and the minimal root is \beq{mbr1} \al_0=-e_1-e_2=-(\al_1+2\al_2+\ldots+2\al_n)\,. \eq \unitlength 1mm \linethickness{0.4pt} \ifx\plotpoint\undefined\newsavebox{\plotpoint}\fi \begin{picture}(64.96,44.893)(0,0) \put(26.087,9.881){\texttt{Fig.2:} \large B$_{n}$ and action of $\la_{1}$} \put(22,36.568){\circle{.42}} \put(22,36.717){\circle{2.102}} \put(22,24.825){\circle*{2.081}} \put(28.987,30.919){\circle{1.808}} \put(37.014,31.068){\circle{1.808}} \put(55.149,30.919){\circle{2.264}} \put(63.92,30.919){\circle{2.081}} \put(56.339,31.663){\line(1,0){6.838}} \put(56.636,30.473){\line(1,0){6.243}} \multiput(60.055,32.406)(.03344645,-.03344645){40}{\line(0,-1){.03344645}} \multiput(61.244,31.068)(-.03303353,-.03716272){36}{\line(0,-1){.03716272}} \put(22.892,35.974){\line(6,-5){5.351}} \multiput(30.027,30.919)(1.218937,.02973){5}{\line(1,0){1.218937}} \put(37.906,30.919){\line(1,0){3.716}} \multiput(50.839,31.068)(.624334,-.02973){5}{\line(1,0){.624334}} \multiput(28.095,30.325)(-.037976505,-.033636333){137}{\line(-1,0){.037976505}} \put(17.838,40.136){$\alpha_1$} \put(17.095,21.852){$\alpha_0$} \put(30.027,34.636){$\alpha_2$} \put(54.704,35.676){$\alpha_{l-1}$} \put(64.069,35.974){$\alpha_l$} \thicklines \put(21.554,28.244){\vector(0,-1){.07}}\put(21.554,33.298){\vector(0,1){.07}}\put(21.554,33.298){\line(0,-1){5.0541}} \end{picture} The root system $R_{B_n}$ contains $2n$ short roots $\pm e_j\,$ $(j=1,\dots n)$ and $2n(n-1)$ long roots $(e_j\pm e_k)$ $(j\neq k)$. The positive roots are \beq{prb1} R^+=\{(e_j\pm e_k)\,,~(j>k)\,,~~e_j\,,~~(j,k=1,\ldots,n)\}\,. \eq The Weyl group of $R_{B_n}$ is the semi-direct product of the permutation group $S_n$ acting on $e_j$ and the signs changing $e_j\to -e_j$ \beq{web} W_{B_n}=S_n\ltimes(\sum\mu_2)\,. \eq The coroot basis $H_\al$ in $\gH_{B_n}$ is formed by simple roots of the dual system \beq{src} \Pi^\vee_{B_n}=\Pi_{C_n}=\{\al_j=e_j-e_{j+1}\,,~j=1,\ldots,n-1\,,~~\al_n=2e_n\}\,. \eq $\Pi^\vee_{B_n}$ generates the coroot lattice \beq{bcrl3} Q_{B_n}^\vee=Q_{C_n}=\{\ga=\sum_{j=1}^nm_je_j\,|\,\sum_{j=1}^nm_j\,{\rm is~even}\}\,. \eq The half-sum of positive coroots is $\rho_{B_n}^\vee=ne_2+(n-1)e_2+\ldots+2e_{n-1}+e_n$, and since $h=2n$ \beq{iv1} \ka_{B_n}=\rho/h=\oh e_1+(\oh-\f1{2n})e_2+\ldots+\f1{n}e_{n-1}+\f1{2n}e_n)\,. \eq The dual to $\Pi_{B_n}$ the system of the fundamental coweights takes the form $$ \{\varpi_j^\vee=e_1+e_2+\ldots+e_j\,,~j=1,\ldots,n\}\,. $$ It generates the coweight lattice \beq{bcrl1} P^\vee_{B_n}=P_{C_n}=\{\ga=\sum_{j=1}^nm_je_j\,|\,m_j\in\mZ\}\,. \eq The factor group $P_{B_n}^\vee/Q_{B_n}^\vee\sim\mu_2$ is isomorphic to the center of $Spin(2n+1)$. The center is generated by the coweight $\varpi^\vee_1\,$ $(\zeta=\bfe\,(\varpi^\vee_1))$. It follows from (A.17.I) and (\ref{mbr1}) that the fundamental alcove has the vertices $$ C_{alc}=(0,\varpi^\vee_1,\oh\varpi^\vee_2,\ldots,\oh\varpi^\vee_n)\,. $$ We use this expression to find $\la_1$ corresponding to $\xi=\varpi^\vee_1$: (3.10.I) \beq{lab} \la_1:\,(e_1\to -e_1\,,e_j\to e_j\,,~(j=2,\ldots,n))\,. \eq It acts on $\Pi^{ext}_{B_n}$ as $$ \la_1:\,(\al_1\to\al_0\,,~\al_j\to\al_j\,,~(j=2,\ldots,n))\,. $$ It is clear that $\ti\gH=\gH_{B_{n-1}}$ and the moduli vector is $\ti\bfu=(u_2,\ldots,u_n)$. \subsection*{Bases.} \subsubsection*{The Chevalley basis} The Chevalley basis in $\gg_{B_n}$ is generated by the simple coroots $\Pi^\vee_{C_n}=\Pi_{B_n}$ (\ref{bsr1}) in $\gH_{B_n}$ and by the root subspaces. To describe the Chevalley basis for the classical groups we use their fundamental representations. For $\gg_{B_n}$ it is a fundamental representation $\pi_1$ corresponding to the weight $\varpi_1=e_1$. It has dimension $2n+1$. $\gg_{B_n}$ becomes the Lie algebra of matrices of order $2n+1$ satisfying the constraints $$ Zq+qZ^T=0\,, $$ where $q$ in the bilinear form $$ q=\left( \begin{array}{ccc} 0 & 0& J \\ 0&1&0\\ J &0& 0 \\ \end{array} \right)\,, $$ and $J$ is an $n$-th order anti-diagonal matrix \beq{15} J=\left( \begin{array}{cccc} & & & 1 \\ & & 1 & \\ & & & \\ 1& & & \\ \end{array} \right) \eq Then $Z$ takes the form \beq{31b} Z=\left( \begin{array}{ccc} A &\al& B \\ \ti\be^T&0&-\ti\al^T\\ C &-\be& -\ti A \\ \end{array} \right)\,,~~ B=\ti{B}\,,~C=\ti{C}\,,~~\ti{X}=JX^TJ\,, \eq $$ \al=(\al_1,\ldots,\al_n)\,,~~~\be=(\be_1,\ldots,\be_n)\,,~~~~\ti\al=\al J\,. $$ The basis in $\gH_{B_n}$ is generated by the canonical basis $(e_1,\ldots,e_n)$. Then the canonical basis in $\pi_1(\gH)$ is $$ \di(e_1,\ldots,e_n,0,-e_n,\ldots,-e_1)\,. $$ The root subspaces in $\pi_1$ are \beq{chbb} (e_j-e_k)\,,~j\neq k\to\gG^-_{jk}=(E_{j,k}\in A-E_{2n+2-k,2n+2-j}\in A\,, \ti A)\,, \eq $$ j<k\sim({\rm positive ~roots})\,,~~j>k\sim({\rm negative ~roots})\,, $$ $$ ( e_j+ e_k)\,,\to\gG^+_{jk}=(E_{j,k+n+1}-E_{n+1-k,2n+2-j})\in B\,,C \,, $$ $$ e_j\,,\to\gG^+_{j}=\sqrt{2}( E_{j,n+1}-E_{n+1.2n+2-j})\in\al \,,\,({\rm positive ~roots}) \,, $$ $$ -e_j\,,\to \gG^-_{j}=\sqrt{2}(E_{n+1,j}-E_{2n+2-j,n+1})\in\be\,,\, ({\rm negative ~roots}) \,. $$ The levels of positive roots are \beq{rlb1} f_{e_j-e_k}=k-j\,,~f_{e_j}=n-j+1\,,~f_{e_j+e_k}=2n-k-j+22n\,. \eq The Killing form normalized as in (A.24.I), (A.25.I) takes the form \beq{kfb1} (Z_1,Z_2)=\f1{2}\tr\,Z_1Z_2\,. \eq Then \beq{kfb} (\gG_{jk}^\pm,\gG_{il}^\pm)=\de_{ki}\de_{jl}\,,~~~(\gG_j^+,\gG_k^-)=2\de_{jk}\,. \eq \subsubsection*{The GS-basis} The Weyl transformation $\La_1$ in this basis is represented by the matrix $$ \La_1= \left( \begin{array}{ccc} 0 & 0 & 1 \\ 0 & -Id_{2n-1} & 0 \\ 1 & 0 & 0 \\ \end{array} \right)\,. $$ Taking into account its action on $Z$ (\ref{31b}) we find that $\Pi_1=\Pi_{B_{n-1}}$ and $\ti\Pi^\vee=\Pi^\vee_{B_{n-1}}$. Then $$ \gg_{B_{n}}=\gg_0+\gg_1\,,~~~\gg_0=\ti\gg_0+V\,,~~\ti\gg_0=\gg_{B_{n-1}}\,, $$ $$ \dim \,\gg_{B_{n-1}}=(n-1)(2n-1) \,,~~\dim \,V=2(n-1)+1\,,~~\dim \,\gg_1=2(n-1)+2\,, $$ where $V$ is the vector representation of $\ti\gg_0=\bfso(2n-1)$. The invariant subalgebra $\gg_0$ is isomorphic to $\bfso(2n)$. Its Dynkin diagram is obtained from $\Pi^{ext}_{B_n}$ by dropping out the root $\al_n$ \cite{Ka}. The space $V+\gg_1$ is represented by matrices of the form $$ \left( \begin{array}{ccccc} a & \vec x & \al & \vec y & 0 \\ (\vec z) J & 0 & 0 & 0 & -(\vec y)J \\ \be & 0 & 0 & 0 & -\al \\ (\vec w) J & 0 & 0 & 0 & -(\vec x) J \\ 0 & -\vec w & -\be & -\vec z & -a \\ \end{array} \right)\,. $$ Since $\ti\gg_0$ is generated by a trivial orbit $(l=1)$, the GS-basis is formed by the Chevalley basis in $\gg_{B_{n-1}}$ (\ref{chbb}) and the GS generators in $V+\gg_1$. $$ V=\left\{ \begin{array}{ll} \gt^0_{1,k}=\f1{\sqrt{2}}(E_{1,k}+E_{2n+1,k}-\ldots)\, &\gt^0_{1,n+1+k}=\f1{\sqrt{2}}(E_{1,n+1+k}+E_{2n+1,n+1+k}-\ldots)\,, \\ \gt^0_{j+n+1,1}=\f1{\sqrt{2}}(E_{j+n+1,1}+E_{j+n+1,2n+1}-\ldots)\,, & \gt^0_{j,1}=\f1{\sqrt{2}}(E_{j,1}+E_{j,2n+1}-\ldots)\,, \\ \gt^0_{n+1,1}=(E_{n+1,1}+E_{n+1,2n+1}-\ldots)\,, & \gt^0_{1,n+1}=(E_{1,n+1}+E_{2n+1,n+1}-\ldots)\,, \end{array} \right. $$ $$ \gg_1=\left\{ \begin{array}{ll} \gt^1_{1,k}=\f1{\sqrt{2}}(E_{1,k}-E_{2n+1,k}-\ldots)\, &\gt^1_{1,n+1+k}=\f1{\sqrt{2}}(E_{1,n+1+k}-E_{2n+1,n+1+k}-\ldots)\,, \\ \gt^1_{j+n+1,1}=\f1{\sqrt{2}}(E_{j+n+1,1}-E_{j+n+1,2n+1}-\ldots)\,, & \gt^1_{j,1}=\f1{\sqrt{2}}(E_{j,1}-E_{j,2n+1}-\ldots)\,, \\ \gt^1_{n+1,1}=\f1{\sqrt{2}}(E_{n+1,1}-E_{n+1,2n+1}-\ldots)\,, & \gt^1_{1,n+1}=\f1{\sqrt{2}}(E_{1,n+1}-E_{2n+1,n+1}-\ldots)\,,\\ \gh^1_1=\sqrt{2}\di(e_1,0\ldots,0-e_1) & \,. \end{array} \right. $$ From (5.9.I) (5.35.I) and (\ref{kfb1}) we find the dual basis \beq{dbb} \gT^a_{jk}=\gt^a_{kj}\,,~~j~{\rm or}~k\neq n+1\,,~~\gH^1_1=\gh^1_1\,,~~ \gT^a_{n+1,1}=\oh\gt^a_{1,n+1}\,,~~\gT^a_{1,n+1}=\oh\gt^a_{n+1,1}\,. \eq and \beq{bpa} (\gt^a_{1j},\gt^b_{i1})=\left\{ \begin{array}{cc} \de^{(ab)}\de_{(ji)} & i,j\neq n+1\,, \\ 2\de^{(ab)}\de_{(ji)} & i=n+1\,, \end{array} \right.~~~ (\gh^1_1,\gh^1_1)=4\,. \eq \subsection*{The Lax operators and the Hamiltonians} \subsubsection*{Trivial bundles.} For trivial bundles the moduli space is described by the vector $\bfu=(u_1,\ldots,u_n)$. For a trivial $\bar G=Spin(2n+1)$-bundles it means that \beq{csc1} \bfu\in\gH_{B_n}/W_{B_n}\ltimes(\tau Q^\vee\oplus Q^\vee)\,, \eq where $W_{B_n}$ is defined in (\ref{web}). For trivial $SO(2n+1)$-bundles we have \beq{csc2} \bfu\in\gH_{B_n}/W_{B_n}\ltimes(\tau P^\vee\oplus P^\vee)\,. \eq The dual variables $\bfv=(v_1,\ldots,v_n)\,$ $v_j\in\mC$ are the same in the both cases. The standard CM Lax operator in the Chevalley basis is \beq{tlb} L^{CM}_{B_n}(z)=\sum_{j=1}^n(v_j+S_{0,j}E_1(z))\gE_j+\sum_{j\neq k}S_{jk}\phi(u_j-u_k,z)\gG^-_{jk} \eq $$ +\sum_{j\neq k}S_{j,-k}\phi(u_j+u_k,z)\gG^+_{j,k}+ \sum_{j}S^{\pm}_j\phi(u_j,z)\gG^{\pm} _j\,, $$ where $\gE_j=\di(0, \ldots,1,0,\dots,0,-1,0,\ldots,0) $. The quadratic Hamiltonian after the reduction with respect to the Cartan subgroup takes the form (see (\ref{kfb})) \beq{bth} H^{CM}_{B_n}=\oh\sum_{j=1}^nv^2_j-\oh\sum_{j\neq k}((S_{jk}S_{kj}E_2(u_j-u_k)+ S_{j,-k}S_{k,-j}E_2(u_j+u_k))-\sum_{j}S^+_jS^-_jE_2(u_j,z)\,. \eq Thus, we have two types of the standard CM systems with the same Hamiltonians and different configuration spaces described by (\ref{csc1}) and (\ref{csc2}). \subsubsection*{Nontrivial bundles} The GS-basis is described above. Now $\ti\bfu=(u_2,u_3,\ldots,u_n)$ belongs to $\gH_{B_{n-1}}$. In fact, $\ti\bfu$ belongs to one of fundamental domains in $\gH_{B_{n-1}}$ (\ref{csc1}), (\ref{csc2}) under the action of the coweight and the coroot lattices. Taking into account (\ref{rlb1}) and $h=2n$ we find from the general prescription $$ L_1(z)=S^1_1\phi(\oh,z)\gh^1_1+S^1_{n+1,1}\bfe\,(\frac{z}2)\phi(\frac{1+\tau}2,z)\gt_{1,n+1}^1+ S^1_{1,n+1}\bfe\,(-\frac{z}2)\phi(\frac{1+\tau}2,z)\gt_{n+1,1}^1+ $$ $$ \sum_{k=2}^n\Bigl(S^1_{k,1}\bfe\,((\frac{k-1}{2n})z)\phi(\frac{(k-1)\tau}{2n}-u_k+\oh,z)\gt_{1,k}^1 +S^1_{1,k}\bfe\,((\frac{1-k}{2n})z)\phi(\frac{(1-k)\tau}{2n}+u_k+\oh,z)\gt_{k,1}^1 $$ $$ +S^1_{n+1+k,1}\bfe\,(\frac{2n+1-k}{2n}z)\phi(\frac{(2n+1-k)\tau}{2n}-u_k+\oh,z)\gt_{1,n+1+k}^1 $$ $$ +S^1_{1,n+1+k}\bfe\,(-\frac{2n+1-k}{2n}z)\phi(u_k-\frac{(2n+1-k)\tau}{2n}+\oh,z)\gt_{k,1}^1 \Bigr)\,, $$ $$ L'_0(z)=S'_{n+1,1}\bfe\,(\frac{z}2)\phi(\frac{\tau}2,z)\gt_{1,n+1}^0+ S'_{1,n+1}\bfe\,(-\frac{z}2)\phi(-\frac{\tau}2,z)\gt_{n+1,1}^0+ $$ $$ \sum_{k=2}^n\Bigl(S'_{k,1}\bfe\,((\frac{k-1}{2n})z)\phi(\frac{(k-1)\tau}{2n}-u_k,z)\gt_{1,k}^0 +S'_{1,k}\bfe\,((\frac{1-k}{2n})z)\phi(\frac{(1-k)\tau}{2n}+u_k,z)\gt_{k,1}^0 $$ $$ +S'_{n+1+k,1}\bfe\,(\frac{2n+1-k}{2n}z)\phi(\frac{(2n+1-k)\tau}{2n}-u_k,z)\gt_{1,n+1+k}^0 $$ $$ +S'_{1,n+1+k}\bfe\,(-\frac{2n+1-k}{2n}z)\phi(u_k-\frac{(2n+1-k)\tau}{2n},z)\gt_{k,1}^0 \Bigr)\,. $$ The Lax operator $\ti L_0(z)$ coincides with (\ref{tlb}) after the corresponding replacement of indices. Taking into account the pairing (\ref{bpa}) after the reduction we come to the Hamiltonian $$ H=H^{CM}_{B_{n-1}}+H'+H_1\,, $$ $$ -H'=S'_{n+1,1}S'_{1,n+1}E_2(\frac{\tau}2)+\oh\sum_{k=2}^n\Bigl(S'_{k,1}S'_{1,k} E_2(u_k-\frac{(k-1)\tau}{2n})+S'_{n+1+k,1}S'_{1,n+1+k}E_2(\frac{(2n+1-k)\tau}{2n}-u_k) \Bigr)\,, $$ $$ -H_1=S^1_{n+1,1}S^1_{1,n+1}E_2(\frac{1+\tau}2)+\oh (S^1_1)^2E_2(\oh)+ $$ $$ \oh\sum_{k=2}^n\Bigl(S^1_{k,1}S^1_{1,k} E_2(u_k-\frac{(k-1)\tau}{2n}-\oh)+S'_{n+1+k,1}S'_{1,n+1+k}E_2(\frac{(2n+1-k)\tau}{2n}-u_k-\oh) \Bigr)\,. $$ Again, we have two types of systems with the same Hamiltonians and different configuration spaces described by (\ref{csc1}) and (\ref{csc2}), where $u_1$ is omitted. \section{Sp$(n)$ - $C_n$ root system.} \setcounter{equation}{0} \subsection*{Roots, weights and bases.} The algebra Lie $\gg_{C_n}$ has rank $n$ and $\dim\,(\gg_{C_n})=2n^2+n$. The system of simple roots $\Pi_{C_n}$ is defined in (\ref{src}). The minimal root is \beq{mrcn} -\al_0=2e_1=2\sum_{j=1}^{n-1}\al_j+\al_n\,. \eq There are $2n$ long roots $2e_{\pm j}$ and $2n(n-1)$ short roots $\pm e_j\pm e_k$, $j\neq k$. The Weyl group $W_{C_n}$ of $R(C_n)$ coincides with $W_{B_n}$ (\ref{web}). The levels of positive roots (A.3.I) are \beq{rlb} f_{e_j-e_k}=k-j\,,~f_{2e_j}=2n-2j+1\,,~f_{e_j+e_k}=2n-k-j+1\,. \eq The simple coroots $\Pi^\vee_{C_n}=\Pi_{B_n}$ (\ref{bsr1}) generates the coroot lattice \beq{bcrl} Q_{C_n}^\vee=Q_{B_n}=\{\ga=\sum_{j=1}^nm_je_j\,|\,m_j\in\mZ\}\,. \eq Since$ \rho_{C_n}^\vee=(n-\oh)e_1+(n-\frac{3}2)e_2+\ldots+\frac{3}2e_{n-1}+\oh e_n$, and $h=2n$ \beq{icv} \ka_{C_n}=\rho/h=(\oh-\frac{1}{4n})e_1+(\oh-\frac{3}{4n})e_2+\ldots+ \frac{3}{4n}e_{n-1}+\frac{1}{4n}e_n\,. \eq The dual to $\Pi_{C_n}$ the system of the fundamental coweights takes the form $$ \{\varpi_j^\vee=e_1+e_2+\ldots+e_j\,,~j=1,\ldots,n-1\,,~ \varpi_n^\vee=\oh\sum_{j=1}^ne_j\}\,. $$ It generates the coweight lattice \beq{bcwl} P_{C_n}^\vee=\{\ga=\sum_{j=1}^n\mZ e_j+\mZ\bigl(\oh\sum_{j=1}^n e_j\bigr)\}\,. \eq The factor-group $P^\vee_{C_n}/Q^\vee_{C_n}\sim\mu_2$ is isomorphic to the center of $Sp(n)$. The center is generated by the coweight $\varpi^\vee_n\,$ $(\zeta=\bfe\,(\varpi^\vee_n))$. It follows from (A.17.I) and (\ref{mrcn}) that the fundamental alcove has the vertices $$ C_{alc}=(0,\oh\varpi^\vee_1,\ldots,\oh\varpi^\vee_{n-1},\varpi^\vee_{n})\,. $$ We use this expression to find $\la_n$ corresponding to $\xi=\varpi^\vee_n$ (3.10.I). $\la_n$ acts on the vertices of $ C_{alc}$ and, in this way, on $\varpi^\vee_j$, as $$ \la_n:\,(\varpi^\vee_1\leftrightarrow\varpi^\vee_{n-1}\,,\,\varpi^\vee_2\leftrightarrow\varpi^\vee_{n-2}\,, \ldots,\varpi^\vee_n\leftrightarrow 0 )\,. $$ Then its action on $\Pi^{ext}_{C_n}$ assumes the form $$ \la_n:\,(\al_0\leftrightarrow\al_{n}\,,\al_2\leftrightarrow\al_{n-2}\,,\ldots\al_n\leftrightarrow~\al_{0},)\,. $$ The action of $\la_n$ on the canonical basis $(e_1,e_2,\ldots, e_n)$ takes the form \beq{ccb} \la_n:\,(e_1\leftrightarrow -e_n\,,e_2\leftrightarrow -e_{n-1}\,,\ldots)\,. \eq If $\bfu=\sum_ju_je_j$, then $ \la_n:\,(u_1\leftrightarrow -u_n\,,u_2\leftrightarrow -u_{n-1}\,,\ldots)$. We define the invariant vector $\ti\bfu$ in the invariant basis $\ti e_1=e_1-e_n\,,\,\ti e_2=e_2-e_{n-1}\,,\ldots,\ti e_l=e_l-e_{l+1}$, $l=\Bigl[\frac{n}2\Bigr]$ \beq{modc} \ti\bfu =\sum_{j=1}^lu_j\ti e_j\,. \eq \begin{flushright} \unitlength 1mm \linethickness{0.4pt} \ifx\plotpoint\undefined\newsavebox{\plotpoint}\fi \begin{picture}(62.142,100.099)(0,0) \put(-74.426,46.563){\circle*{3.64}} \put(-72.426,47.813){\line(1,0){9.5}} \multiput(-72.426,45.563)(1.125,.03125){8}{\line(1,0){1.125}} \put(-63.426,45.813){\line(-1,0){.5}} \put(-61.676,46.563){\circle{3.162}} \put(-59.926,46.563){\line(1,0){10.5}} \put(-47.176,46.563){\circle{3.041}} \put(-6.176,47.313){\circle{3.202}} \multiput(-4.426,48.563)(-.03125,-.03125){8}{\line(0,-1){.03125}} \multiput(-4.676,48.313)(.03125,-.125){8}{\line(0,-1){.125}} \put(-4.426,47.313){\line(-1,0){.25}} \multiput(-4.676,47.313)(1.625,.03125){8}{\line(1,0){1.625}} \put(10.574,47.313){\circle{3.536}} \multiput(12.824,48.063)(1,.03125){8}{\line(1,0){1}} \put(20.824,48.313){\line(1,0){3.25}} \multiput(12.824,46.813)(1.34375,.03125){8}{\line(1,0){1.34375}} \put(25.324,47.563){\circle{3.041}} \multiput(-69.176,48.313)(.07222222,-.03333333){45}{\line(1,0){.07222222}} \multiput(-65.926,46.813)(-.05416667,-.03333333){60}{\line(-1,0){.05416667}} \multiput(15.574,47.563)(.07083333,.03333333){60}{\line(1,0){.07083333}} \multiput(16.324,47.313)(.09210526,-.03289474){38}{\line(1,0){.09210526}} \qbezier(-70.426,56.063)(-29.551,101.938)(17.824,54.313) \qbezier(-54.926,56.313)(-30.426,77.688)(1.074,57.563) \qbezier(-38.426,55.563)(-28.176,64.688)(-14.926,55.313) \put(-74.426,34.313){$\alpha_0$} \put(-61.426,35.313){$\alpha_1$} \put(10.074,37.563){$\alpha_{n-1}$} \put(24.824,39.063){$\alpha_n$} \multiput(-68.176,60.063)(-.03365385,-.05769231){52}{\line(0,-1){.05769231}} \multiput(-69.926,56.063)(.04104478,.03358209){67}{\line(1,0){.04104478}} \multiput(-52.926,59.313)(-.03333333,-.05){60}{\line(0,-1){.05}} \multiput(-54.926,56.313)(.07894737,.03289474){38}{\line(1,0){.07894737}} \multiput(-37.176,57.563)(-.03289474,-.04605263){38}{\line(0,-1){.04605263}} \multiput(-38.176,55.813)(.1086957,.0326087){23}{\line(1,0){.1086957}} \multiput(-18.676,56.813)(.07894737,-.03289474){38}{\line(1,0){.07894737}} \multiput(-17.176,58.563)(.03333333,-.05555556){45}{\line(0,-1){.05555556}} \multiput(-3.176,59.063)(.08552632,-.03289474){38}{\line(1,0){.08552632}} \multiput(.324,58.063)(-.03353659,.03658537){82}{\line(0,1){.03658537}} \multiput(14.574,56.563)(.04477612,-.03358209){67}{\line(1,0){.04477612}} \multiput(15.574,58.063)(.03333333,-.06666667){45}{\line(0,-1){.06666667}} \multiput(17.074,55.063)(-.0333333,.0333333){15}{\line(0,1){.0333333}} \put(-39.747,46.993){\line(1,0){1.8958}} \put(-35.955,46.993){\line(1,0){1.8958}} \put(-32.163,46.993){\line(1,0){1.8958}} \put(-28.372,46.993){\line(1,0){1.8958}} \put(-24.58,46.993){\line(1,0){1.8958}} \put(-20.788,46.993){\line(1,0){1.8958}} \put(-51.097,28.414){\makebox(0,0)[cc]{Fig.3 C$_n$ and $\lambda_n$ action}} \end{picture} \end{flushright} \subsection*{Bases.} \subsubsection*{The Chevalley basis} The Chevalley basis in $\gg_{C_n}$ is generated by the simple coroots $\Pi^\vee_{C_n}=\Pi_{B_n}$ (\ref{bsr1}) in $\gH$ and by the root subspaces. It is convenient to define the Chevalley basis using a fundamental representation $\pi_1$ corresponding to the weight $\varpi_1=e_1$. It has dimension $2n$. We define it as the Lie algebra of matrices $\{Z\}$ satisfying the constraints $$ Zq+qZ^T=0\,,~~Z\in\pi_1\,, $$ where $q$ in the bilinear form $$ q=\left( \begin{array}{cc} 0 & J \\ -J & 0 \\ \end{array} \right)\,, $$ and $J$ is an $n$-th order anti-diagonal matrix (\ref{15}). Then $Z$ takes the form \beq{31a} Z=\left( \begin{array}{cc} A & B \\ C & -\ti A \\ \end{array} \right)\,,~~ B=\ti{B}\,,~C=\ti{C}\,,~~\ti{X}=JX^TJ\,. \eq The basis in $\gH_{C_n}$ is generated by the canonical basis $(e_1,\ldots,e_n)$. Then the canonical basis in $\pi_1(\gH)$ is $$ \di(e_1,\ldots,e_n,-e_n,\ldots,-e_1)\,. $$ The Killing form is similar to the $B_n$ case (\ref{kfb1}) $(Z_1,Z_2)=\f1{2}\tr\,Z_1Z_2$. The root subspaces in $\pi_1$ are \beq{chbc} \begin{array}{cl} (e_j-e_k)&\to\gG_{jk}^-=(E_{j,k}\in A-E_{2n+1-k,2n+1-j}\in \ti A),,\\ ( e_j+ e_k) &\to\gG_{jk}^+=(E_{j,k+n}\in B+E_{n+1-k,2n+1-j}\in C)\,, \end{array} \eq $$ j<k~ \sim~({\rm positive ~roots})\,,~~~j>k~\sim~({\rm negative ~roots}) $$ $$ 2e_j\to \gG^+_j=E_{j,j+n}\in B\,,~({\rm positive ~roots}) \,, ~~ -2e_j\to \gG_j^-= E_{j+n,j}\in C ~({\rm negative ~roots})\,. $$ The Killing form in this basis is \beq{cbkfc} (\gG_{jk}^\pm,\gG_{il}^\pm)=\de_{ki}\de_{jl}\,,~~~ (\gG^+_k,\gG_j^-)=\oh\de_{jk}\,. \eq \subsubsection*{The GS-basis} The transformation $\la_n$ (\ref{ccb}) in $\pi_1$ takes the form \beq{lncn} \La_n= \left( \begin{array}{cc} 0 & i\,Id_n \\ i\,Id_n & 0 \\ \end{array} \right)\,,~~\la_n(Z)= \left( \begin{array}{cc} -\ti A & C \\ B & A \\ \end{array} \right)\, \eq Since $\la_n^2=Id$ \beq{cgr2} \gg_{C_n}=\gg_0+\gg_1\,, \eq where $$ \gg_0=\left\{ \left( \begin{array}{cc} X & Y \\ Y & X \\ \end{array} \right)\,\begin{array}{lc} | & \ti X=-X \\ | & \ti Y=Y \end{array} \right\}\,, ~~~ \gg_1=\left\{ \left( \begin{array}{cc} X & Y \\ -Y & -X \\ \end{array} \right)\,\begin{array}{lc} | & \ti X=X \\ | & \ti Y=Y \end{array} \right\}\,, $$ \beq{cgr3} \gg_0=\ti\gg_0+V\,,~~\ti\gg_0 =\left\{ \left( \begin{array}{cc} X & 0 \\ 0 & X \\ \end{array} \right)\,\right\}\,,~~V=\left\{ \left( \begin{array}{cc} 0 & Y \\ Y & 0 \\ \end{array} \right) \right\}\,, \eq $$ \dim\,\gg_{C_n}=n(2n+1)\,,~~\dim\,\ti\gg_0=\oh n(n-1)\,,~~\dim\,V=\oh n(n+1)\,,~~ \dim\,\gg_1=n(n+1)\,. $$ A type of $\ti\gg_0$ depends on a parity of $n$. Note, that $\Pi_1=A_{n-1}$ (\ref{dpi1}) is generated by roots $(\al_1,\ldots,\al_{n-1})$. Let $\la_n|_{\gH_{A_{n-1}}}=\ti\la$. We prove general Lemma about automorphisms of $\gg_{A_{n-1}}$ that will be applied to $C_n$ and $D_n$ algebras. Let $(e_1,\ldots,e_{n})$ be a canonical basis in $\gH_{A_{n-1}}$ and $E_{j,k},$ $\,(1\leq j<k\leq n)$ is the root basis $\gg_{A_{n-1}}$ =sl$(n,\mC)$. It follows from (\ref{ccb}) and (\ref{lncn}) that the action of $\ti\la$ on the Chevalley basis of $\gH_{A_{n-1}}$ takes the form \beq{laca} \ti\la\,:\, \left\{ \begin{array}{l} (e_1,e_2\ldots,e_{n})\,\to\,(-e_{n},-e_{n-1},\ldots,-e_1)\, \\ \,E_{jk}\,\to\,-E_{n-k+1,n-j+1}\,. \end{array} .\right. \eq \begin{lem} Under the action (\ref{laca}) $\gg_{A_{n-1}}$ is decomposed as $$ \gg_{A_{n-1}}=\ti\gg_0+\ti\gg_1\,, $$ where\\ $\bullet$ $$ \ti\gg_0=\left\{ \begin{array}{cc} \gg_{D_{\frac{n}2}} & n-{\rm even}\,, \\ \gg_{B_{\frac{n-1}2}} & n-{\rm odd}\,. \end{array} \right. $$ with the defined below Chevalley bases (\ref{bsoe}), (\ref{bsoo}).\\ $\bullet$ $\ti\gg_1$ is a space of traceless symmetric matrices of order $n$ $(\dim\,(\ti\gg_1)=\oh n(n+1)$\,, and $\ti\gg_0$ acts on $\ti\gg_1$ by commutators. \end{lem} \emph{Proof}\\ Let $X\in$sl$(n,\mC)$ be a traceless matrix of order $n$ and $$ \ti X=JX^TJ\,,~~ J_{ik}=\de_{i,n-1-k}\,. $$ It follows from (\ref{laca}) that $\ti\la(X)=-\ti X$ and the invariant subalgebra $$ \ti\gg_0=\gg_{A_{n-1}}+\ti\la(\gg_{A_{n-1}}) $$ is the algebra of anti-invariant matrices with respect to the symmetric form $J_{jk}$ $$ XJ+JX^T=0\,. $$ Thus, $\ti\gg_0$ is the Lie algebra of orthogonal matrices SO$(n,\mC)$ and we come to the first statement. Similarly, for $B\in\ti\gg_1$ \beq{B} YJ-JY^T=0\,. \eq It means that $\ti\gg_1$ is the space of complex symmetric matrices with respect to the secondary diagonal. In these terms the commutation relations in $A_{n-1}=$sl$(n,\mC)$ assume the form $$ [\ti\gg_0,\ti\gg_0]\subset\ti\gg_0\,,~~ [\ti\gg_0,\ti\gg_1]\subset\ti\gg_1\,,~~ [\ti\gg_1,\ti\gg_1]\subset\ti\gg_0\,. $$ We construct the Chevalley basis in the invariant subalgebras. The $\ti\la$ on the simple roots of $A_{n-1}$ is $$ \al_1\leftrightarrow\al_{n-1}\,,\al_2\leftrightarrow\al_{n-2}\,,\ldots, \left\{ \begin{array}{lc} \al_{\frac{n}2-1}\leftrightarrow\al_{\frac{n}2+1}& n \,{\rm even}\,, \\ \al_{\frac{n}2}\leftrightarrow\al_{\frac{n}2+1}& n \,{\rm even}\,, \\ \al_{\frac{n-1}2-1}\leftrightarrow\al_{\frac{n-1}2+2}& n \,{\rm odd}\,, \\ \al_{\frac{n-1}2}\leftrightarrow\al_{\frac{n-1}2+1}& n \,{\rm odd}\,, \\ \end{array} \right. $$ For $n=2l$ $\,\ti\gg_0=\gg_{D_l}={\bf so}(2l,\mC)$ the Chevalley basis is formed by the invariant coroots $\{\tial_j^\vee,$ $(j=1,\dots,l)\}$, constructed from $A_{2l-1}$ roots $\al_1,\ldots,\al_{2l-1}$, and the root spaces basis \beq{kc} \ti\Pi^\vee_{D_l}=\{\tial_1^\vee=\al_1+\al_{2l-1}\,,\ldots,\tial^\vee_{l-1}=\al_{l-1}+\al_{l+1}\,,\, \tial_l^\vee=\al_{l-1}+2\al_l+\al_{l+1}\}= \eq $$ \{\tial_1^\vee=e_1-e_2+e_{n-1}-e_n\,,\ldots\tial^\vee_{l-1}=e_{\frac{n}2-1}- e_{\frac{n}2}+e_{\frac{n}2+1}-e_{\frac{n}2+2}\,,\tial^\vee_{l}= e_{\frac{n}2-1}+e_{\frac{n}2}-e_{\frac{n}2+1}-e_{\frac{n}2+2}\}\,. $$ The simple roots dual to coroots are $\tial_j=\oh \tial_j^\vee$. The matrix $a_{jk}=\lan\tial_j,\tial_k^\vee\ran$ is the Cartan matrix $D_l$. The Chevalley generators corresponding to the simple roots are \beq{bsoe} \ti E_{\tial_j}=E_{j,j+1}-E_{n-j,n-j+1}\,,~ j<l\,, ~~~\ti E_{\tial_l}=E_{l-1,l+1}-E_{l,l+2}\,. \eq For $n=2l+1$ the Chevalley basis of the $\ti\gg_0=\gg_{B_l}={\bf so}(2l+1,\mC)$-algebra (see Section 9) takes the form \beq{cibb} \ti\Pi^\vee_{B_l}=\{\tial_1^\vee=\al_1+\al_{2l}\,,\ldots, \tial_{l-1}^\vee=\al_{l-1}+\al_{l+2}\,, \tial_l^\vee=2(\al_l+\al_{l+1})\}= \eq $$ \{\tial_1^\vee=e_1-e_2+e_{n-1}-e_n\,,\ldots\tial^\vee_{l-1}=e_{\frac{n-1}2-1}- e_{\frac{n-1}2}+e_{\frac{n-1}2+1}-e_{\frac{n-1}2+2}\,,\tial^\vee_{l}= 2(e_{\frac{n-1}2-1}-e_{\frac{n-1}2+2})\}\,. $$ \beq{bsoo} \ti E_{\tial_j}=E_{j,j+1}-E_{n-j,n-j+1}\,,~ j<l\,, ~~~\ti E_{\tial_l}=E_{l,l+1}-E_{l+1,l+2}\,,~ j=l\,. \eq The dual root system $$ \ti\Pi_{B_l}=\{\tial_1=\oh\tial_1^\vee\,,\ldots,\tial_{l-1}=\oh\tial_{n-1}^\vee\,, \tial_l=\f1{4}\tial_l^\vee\}\,. $$ leads to the $B_l$ Cartan matrix $a_{jk}=\lan\ti\al_j,\ti\al_k^\vee\ran$. $\Box$ \begin{rem} In this Lemma we have found the coroot basis in the invariant subalgebra $\ti \gH_0$ for the special cases $C_l$, $D_l$ . The expressions (\ref{kc} ), (\ref{cibb} ) for $\ti\al^\vee_l$ replace the general formula (5.26.I). \end{rem} \bigskip \emph{\textbf{Basis in $V$ and $\gg_1$}} We have constructed a basis in $\ti \gg_0$. Consider other component in (\ref{cgr2}), (\ref{cgr3}). The basis in $V=\{B\}$, where $B$ satisfies (\ref{B}), has the form \beq{BV} \gt^0_{j,k+n}=\f1{\sqrt{2}}(E_{j,k+n}+E_{n+j,k}+E_{n-k+1,2n-j+1}+E_{2n-k+1,n-j+1})\,,~(j\neq k)\,, \eq $$ \gt^0_{j,n+j}=\f1{\sqrt{2}}(E_{j,n+j}+E_{n+j,j}) \,. $$ It is easy to find that the invariant subalgebra $\gg_0=\ti\gg_0+V$ is isomorphic to ${\bf gl}(n)$. The form of $\gg_0$ is read off from from the extended Dynkin diagram for $\Pi^{ext}_{C_n}$ (Fig.3) by dropping out the roots $\al_0$ and $\al_n$ (see \cite{Ka}). The GS basis in $\gg_1$ is represented as $$ \gh^1_j=\f1{\sqrt{2}}(e_j+e_{n-j+1}- e_{2n-j+1}-e_{n+j})\,,~~(j=1,\dots\Bigl[\frac{n}2\Bigr])\,, $$ \beq{Bg1} \gt^1_{j,n+k}=\f1{\sqrt{2}}(E_{j,n+k}+E_{n-k+1,2n-j+1}-E_{n+j,k}-E_{2n-k+1,n-j+1})\,,~(j\neq k)\,, \eq $$ \gt^1_{j,k}=\f1{\sqrt{2}}(E_{j,k}+E_{n-k+1,n-j+1}-E_{n+j,n+k}-E_{2n-k+1,2n-j+1})\,,~~(j\neq k)\,, $$ $$ \gt^1_{j,n+j}=\f1{\sqrt{2}}(E_{j,n+j}-E_{n+j,j})\,. $$ In terms of the GS basis the Killing form is \beq{kfgsc} (\gt^a_{j,n+j},\gt^b_{k,n+k})=\frac{(-1)^a}{2}\de_{jk}\de^{a,b}\,,~~(\gt^a_{s,r+n},\gt^b_{j,k+n})= (-1)^a\de_{sk}\de_{rj}\de^{a,b}\,, \eq $$ (\gt^1_{s,r},\gt^1_{j,k})=\de_{sk}\de_{rj}\,,~~(\gh^1_j,\gh^1_k)=\de_{jk}\,. $$ \bigskip \subsection*{The Lax operators and the Hamiltonians.} \subsubsection*{Trivial bundles.} For trivial bundles the moduli space is described by the vector $\bfu=(u_1,\ldots,u_n)$. If $E$ is a $\bar G=Sp(n)$-bundle \beq{cscc1} \bfu\in\gH_{C_n}/W_{C_n}\ltimes(\tau Q_{C_n}^\vee\oplus Q_{C_n}^\vee)\,. \eq For trivial $Sp(n)/\mu_2$-bundles \beq{cscc2} \bfu\in\gH/W_{C_n}\ltimes(\tau P_{C_n}^\vee\oplus P_{C_n}^\vee)\,. \eq The dual variables $\bfv=(v_1,\ldots,v_n)\,$ $v_j\in\mC$ are the same in the both cases. The standard CM Lax operator in the Chevalley basis is \beq{tlc} L(z)=\sum_{j=1}^n(v_j+S_{0,j}E_1(z))e_j+\sum_{j\neq k}S_{j,k}\phi(u_j-u_k,z)\gG^-_{j,k} \eq $$ +\sum_{j<k}S_{j,k+n}\phi(u_j+u_k,z)\gG^+_{j,k} +\sum_{j>k}S_{j+n,k}\phi(-u_j-u_k,z)\gG^+_{j,k} $$ $$ \sum_{j}(S_{j,j+n}\phi(2u_j,z)\gG^-_{j}+S_{j+n,j}\phi(-2u_j,z)\gG^+_{j})\,. $$ The quadratic Hamiltonian after the reduction with respect to the Cartan subgroup takes the form (see (\ref{cbkfc})) $$ H^{CM}_{C_n}=\oh\sum_{j=1}^nv^2_j-\oh\sum_{j\neq k}((S_{j,k}S_{k,j}E_2(u_j-u_k)+ S_{j,k+n}S_{j+n,k}E_2(u_j+u_k))-\f1{4}\sum_{j}S_{j,j+n}S_{j+n,j}E_2(2u_j,z)\,. $$ Thus, we have two types of the standard CM systems with the same Hamiltonians and different configuration spaces described by (\ref{cscc1}) and (\ref{cscc2}). \subsubsection*{Nontrivial bundles} \emph{Moduli space} Consider a bundle with a characteristic class defined by $\zeta=\bfe\,(\varpi^\vee_n)$. The moduli space is defined by vectors $\ti\bfu\in\ti\gH_0$ (\ref{modc}). Let $n=2l+1$ be an odd number. As it was explained above, $\ti\gH_0$ is a Cartan subalgebra of $\gg_{B_l}$. The coroot lattice in $\ti\gH_0$ (\ref{bcrl}) \beq{crsb} \ti Q^\vee=\{\ga=\sum_j^lm_j\ti e_j\,,~\sum_j^lm_j~{\rm is~even}\} \eq is invariant sublattice in $Q^\vee_{C_n}$ (\ref{bcrl}). The coweight lattice in $\ti\gH_0$ (\ref{bcrl1}) $$ \ti P^\vee=\{\ga=\sum_{j=1}^lm_j\tie_j\,|\,m_j\in\mZ\}\,. $$ Let $\ti W=W_{B_l}$ be the Weyl group. A closer of the positive Weyl chamber for $\ti\bfu=(u_1,\ldots,u_l)$ is \beq{wcc} u_1\geq u_2\geq\ldots\geq u_l\geq 0\,. \eq There are two types of the moduli spaces defined as the quotient of $\ti\gH_0$ $$ \ti\gH_0/(\ti W\ltimes(\tau\ti Q^\vee+\ti Q^\vee)\,, ~~~ \ti\gH_0/(\ti W\ltimes(\tau\ti P^\vee+\ti P^\vee)\,. $$ For $n=2l$ the invariant subalgebra is $\gg_{D_l}$. The structure of lattices for this algebra will be described in Section 5. The invariant coroot sublattice is the same as above (\ref{crsb}). The form of the coweight sublattice depends on a parity of $l$. Let $l$ be odd. In this case $$ \ti P^\vee=\ti Q^\vee+\mZ(\oh\sum_j^l\ti e_j)\,. $$ There is an intermediate sublattice \beq{isl} \ti P_2^\vee=\{\ga=\sum_{j=1}^lm_j\tie_j\,|\,m_j\in\mZ\}\,. \eq Therefore, there are three types of the moduli spaces: $$ \ti\gH_0/(\ti W\ltimes(\tau\ti Q^\vee+\ti Q^\vee)\,, ~~~ \ti\gH_0/(\ti W\ltimes(\tau\ti P^\vee+\ti P^\vee)\,, ~~~ \ti\gH_0/(\ti W\ltimes(\tau\ti P_2^\vee+\ti P_2^\vee)\,. $$ For $l$ even there are two type of coweight lattices $$ \ti P^{L\vee}=\ti Q^\vee+\mZ(\oh\sum_j^l\ti e_j)\,, ~~~ \ti P^{R\vee}=\ti Q^\vee+\mZ\oh(\sum_j^{l-1}\ti e_j-\tie_l) $$ and the intermediate sublattice $\ti P_2^\vee$ (\ref{isl}). In this case there are four types of the moduli spaces. \bigskip \emph{Lax operator} Using the general prescription for Lax operators (6.15.I), (6.17.I), (6.18.I), the GS basis (\ref{BV}), (\ref{Bg1}), and (\ref{rlb}) we define $L$ in our case. The invariant operator $\ti L_0(z)$ is the Lax operator of the SO$(n)$ CM system (\ref{tlb}) and (\ref{trso}). $$ L_1(z)=\sum_{j=1}^n\Bigl(S_j\phi(\oh,z)\gh^1_j+ \oh S_{j,j+n}\bfe\,(\frac{2n+1-2j}{2n}z)\phi(\frac{(2n+1-2j)\tau}{2n}+\oh-2u_j,z) \gt^1_{j,j+n}\Bigr) $$ \beq{l1sp} +\sum_{j\neq k}^n\Bigl(S_{j,n+k}\bfe\,(\frac{2n+1-j-k}{2n}z) \phi(\frac{(2n+1-j-k)\tau}{2n}-u_j-u_k+\oh,z)\gt^1_{j,n+k}\Bigr. \eq $$ \Bigl. +S_{j,k}\bfe\,(\frac{k-j}{n}z) \phi(\frac{(k-j)\tau}{n}-u_j+u_k+\oh,z)\gt^1_{k,j}\Bigr)\,, $$ $$ L_0'(z)=\sum_{j\neq k}^nS_{j,k+n}^{'}\bfe\,(\frac{2n+1-j-k}{2n}z) \phi(\frac{(2n+1-j-k)\tau}{2n}-u_j-u_k,z)\gt^0_{j,k+n}\Bigr. $$ \beq{l0sp} +\oh\sum_{j=1}^nS_{j,n+j}^{'}\bfe\,(\frac{2n+1-2j}{2n}z) \phi(\frac{(2n+1-2j)\tau}{2n}-2u_j,z)\gt^0_{j,n+j}\,. \eq In these expressions the identification $u_j=-u_{n+1-j}$ as a result of $\la_n$ action is assumed. \bigskip \emph{Hamiltonian}\\ Due to the gradation the Hamiltonian contain three terms $$ H=\ti H_0+H'_0+H_1\,, $$ where $\ti H_0$ is the CM Hamiltonian related to SO$(n,\mC)$ (\ref{bth}), (\ref{dh}). To define $H'_0$ and $H_1$ one should take into account the Killing form (\ref{kfgsc}). Then from (\ref{l1sp}) and (\ref{l0sp}) we find $$ H_1= \oh\sum_{j=1}^n\Bigl(-S_{j}^2E_2(\oh)+ S^2_{j,j+n}E_2(\frac{(2n+1-2j)\tau}{2n}-2u_j+\oh) \Bigr) $$ $$ -\oh\sum_{j<k}^n\Bigl(-S_{j,k+n}S_{k,j+n}E_2(\frac{(2n+1-j-k)\tau}{2n}-u_j-u_k-\oh)+ S_{j,k}S_{k,j}E_2(\frac{(k-j)\tau}{n}-u_j+u_k-\oh)\Bigr)\,, $$ $$ H'_0=-\oh\sum_{j<k}^n\Bigl(S'_{j,k+n}S'_{k,j+n}E_2(\frac{(2n+1-j-k)\tau}{2n}-u_j-u_k) +\oh (S'_{j,j+n})^2E_2(\frac{(2n+1-2j)\tau}{2n}-2u_j)\Bigr)\,. $$ As above $u_j=-u_{n-j+1}$. \section{SO$(2n)$\,, Spin$(2n)\,$. General construction} \setcounter{equation}{0} \subsection*{Roots and weights.} The Lie algebra $D_n$ has dimension $2n^2-n$ and rank $n$. The universal covering group $\bG$ is Spin$(2n)$ and $G^{ad}=$SO$(2n)/\mu_2$. In terms of the canonical basis on $\gH_{D_n}$ $(e_1,e_2,\ldots,e_n)$ simple roots of $D_n$ are $\Pi_{D_n}=\{\al_j=e_j-e_{j+1}\,,~j=1,\dots,n-1\,,~\al_n=e_1+e_n\}$, and the minimal root is \beq{mbr} \al_0=-e_1-e_2=-(\al_1+2\al_2+\ldots+2\al_{n-2}+\al_{n-1}+\al_{n})\,. \eq The roots of system $R_{D_n}$ have the same length and $\sharp R=2n(n-1)$. The positive roots are \\ $R^+=\{(e_j\pm e_k)\,,~j,k=1,\ldots,n\,,~(j>k)\}$. The levels of positive roots are \beq{rld} f_{e_j-e_k}=k-j\,,~~f_{e_j+e_k}=2n-k-j\,. \eq The Weyl group $W_{D_n}$ of $R_{D_n}$ is the semidirect product of permutations $S_n$ and the sign changes \beq{wed} S_n\ltimes{\,\rm (the~sign~changes\,}e_j\to-e_j)\,,~(\prod_{j=1}^n(\pm 1)_j=1)\,. \eq The root and coroot systems coincide and $R_{D_n}$ generates the root lattice \beq{drl} Q_{D_n}=\{\ga=\sum_{j=1}^nm_je_j\,|\,\sum_{j=1}^nm_j\,{\rm is~even}\}\,. \eq The half-sum of positive $D_n$ coroots is $\rho_{D_n}=(n-1)e_1+(n-2)e_2+\ldots+2e_{n-2}+e_{n-1}$, and since the Coxeter number $h=2n-2$ \beq{iv} \ka=\rho/h=\oh e_1+(\oh-\f1{2(n-1)})e_2+\ldots+(\oh-\frac{j-1}{2n-2})e_{j}+\ldots\f1{2n-2}e_{n-1}\,. \eq The system of the fundamental weights takes the form $$ \varpi_j=e_1+e_2+\ldots+e_j\,,~j=1,\ldots,n-2\,, $$ $$ \varpi_{n-1}=\oh(e_1+\ldots+e_{n-1}-e_n)\,,~~ \varpi_{n}=\oh(e_1+\ldots+e_{n-1}+e_n)\,. $$ The weights $\varpi_1$, $\varpi_{n-1}$, and $\varpi_{n}$ are the highest weights of the vector $\underline{2n}$, right spinors $(\underline{2^{n-1})}^R$ and left spinors $(\underline{2^{n-1})}^L$ representations. We find from (A.17.I) and (\ref{mbr}) that the fundamental alcove has the vertices \beq{alcd} C_{alc}=(0,\varpi_1,\oh\varpi_2,\ldots,\oh\varpi_{n-2}, \varpi_{n-1}, \varpi_n)\,. \eq \subsection*{The Chevalley basis} The Chevalley basis is convenient to define using a fundamental representation $\pi_1$ corresponding to the weight $\varpi_1=e_1$. It has dimension $2n$. If the symmetric form is represented by the anti-diagonal matrix as for $B_n$ then $Z\in${\bf so}$(2n)$ takes the form \beq{31} Z= \left( \begin{array}{cc} A & B \\ C & -\ti A \\ \end{array} \right)\,,~~ B=-\ti{B}\,,~C=-\ti{C}\,,~~\ti{X}=JX^TJ\,. \eq The basis in $\gH_{D_n}$ is generated by the simple roots $\Pi_{D_n}$ in $\gH$ (or by the canonical basis $(e_1,\ldots,e_n)$). In $\pi_1$ the canonical basis in $\gH_{D_n}$ is $\di(e_1,\ldots,e_n,-e_n,\ldots,-e_1)$. The root subspaces in $\pi_1$ are \beq{chbd} \begin{array}{ccc} (e_j-e_k)\,,~j<k&\to\gG_{jk}^-=(E_{j,k}(\in A)-E_{2n+1-k,2n+1-j}(\in \ti A),,&({\rm positive ~roots})\\ (e_j-e_k)\,,~j>k&\to\gG_{jk}^-=(E_{j,k}(\in A)-E_{2n+1-k,2n+1-j}(\in \ti A)),,&({\rm negative ~roots})\,, \end{array} \eq $$ \begin{array}{ccc} ( e_j+ e_k)\,,~j<k &\to\gG_{jk}^+=(E_{j,k+n}-E_{n+1-k,2n+1-j})\in B &({\rm positive ~roots}) \,, \\ ( e_j+ e_k)\,,~j>k &\to\gG_{jk}^+= (E_{j+n,k}-E_{2n+1-k,n+1-j})\in C& ({\rm negative ~roots})\,. \end{array} $$ The Killing form (A.25.I) takes the form \beq{kfcd} (\gG_{jk}^\pm,\gG^\pm_{il})=\de_{jl}\de_{ik}\,. \eq \section{SO$(2n)$\,, Spin$(2n)\,$, $n=2l+1$.} \setcounter{equation}{0} \subsection*{Lattices and characteristic classes.} It follows from (\ref{drl}) that $$ \varpi_j\in Q(D_n)~{\rm for}~ 1<j<n-1\,,~~ \varpi_1\sim 2\varpi_n\,,~\varpi_{n-1}\sim 3\varpi_n\,,~ mod\,Q(D_n)\,. $$ It means that \beq{pdno} P(D_n)=Q(D_n)+\mZ\varpi_n\sim Q(D_n)+\mZ\varpi_{n-1}\,. \eq The weight lattice $P(D_n)$ contains a sublattice $P_2(D_n)$ of index two \beq{P2D} P_2(D_n)=Q(D_n)+\mZ \varpi_1=\{\sum_{j=1}^nm_je_j\,|\,m_j\in\mZ\}\,. \eq This lattice is self-dual $P_2(D_n)=^LP_2(D_n)$ and isomorphic to the group of cocharcters $P_2(D_n)=t(SO(2n))$, (A.43.I), where $SO(2n)=Spin(2n)/\mu_2$. On the other hand, the weight lattice $P(D_n)$ is dual to the root lattice $^LP(D_n)=Q(D_n)$. The center of $Spin(2n)$ for odd $n$ is $\clZ(\bar G)=P_{D_n}/Q_{D_n}\sim\mu_4$. The group element $\zeta=\bfe\,(\xi)$ for $\xi=\varpi_n$ generates $\mu_4$. Putting $\xi=\varpi_n$ and $C_{alc}$ (\ref{alcd}) in (3.10.I) we find $\la_n$ $$ \la_n\,:\,~ \left\{ \begin{array}{l} \varpi_n\to 0\to \varpi_{n-1}\to\varpi_1\to\varpi_n\,, \\ \varpi_j\to\varpi_{n-j}\,,~1<j<n-1\,. \end{array} \right\}\,, ~~~ \left\{ \begin{array}{l} \al_0\to \al_n\to \al_{1}\to\al_{n-1}\to\al_0\,,\\ \al_j\to\al_{n-j}\,,~~1<j<n-1\,. \end{array} \right\}\,. $$ In $\pi_1$ the transformation takes the form $$ \la_n= \left( \begin{array}{cccc} 0 & 1 & 0 & 0 \\ 0 & 0 & 0 & Id_{n-1} \\ Id_{n-1}& 0 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ \end{array} \right) $$ It acts on the basis elements in the space of $\underline{2n}$ representation as \beq{indo} \left( \begin{array}{cccccccccccc} 1\,,&2\,,&\ldots,&n-1\,,& n\,,&n+1\,,&n+2\,,&\ldots, &2n-1\,,& 2n \\ n\,,&n+2 \,,&\ldots,&2n-1&2n\,,&1\,,&2\,,&\ldots, &n-1\,,& n+1 \end{array} \right)\,. \eq In the canonical basis $(e_1,e_2,\ldots,e_n)$ $\la_n$ is represented by the matrix \beq{cbdl} \la_n\to \La_{jk}= \left\{ \begin{array}{c} \de_{j,n-k+1}\,,~j<n\\ -\de_{j,n-k+1}\,,~j=n\,. \end{array} \right. \eq This transformation is an element of the Weyl group $W(SO(2n))$. \vspace{0.5cm} \unitlength 1mm \linethickness{0.4pt} \ifx\plotpoint\undefined\newsavebox{\plotpoint}\fi \begin{picture}(59.806,33.967)(0,0) \put(23.636,23.784){\circle{1.808}} \put(30.771,19.473){\circle{2.081}} \put(37.906,19.77){\circle{2.081}} \put(51.88,19.77){\circle{1.808}} \put(58.866,23.635){\circle{1.88}} \put(58.718,15.757){\circle{1.808}} \put(23.636,15.905){\circle*{1.88}} \put(24.528,23.338){\circle*{.297}} \put(32.109,19.621){\circle*{.297}} \put(32.109,19.621){\line(1,0){4.757}} \put(39.096,19.77){\line(1,0){5.054}} \put(48.461,19.621){\line(1,0){.149}} \put(48.461,19.77){\line(1,0){2.378}} \multiput(57.826,23.189)(-.05679775,-.03340449){89}{\line(-1,0){.05679775}} \multiput(52.623,18.878)(.06937333,-.03369333){75}{\line(1,0){.06937333}} \multiput(29.879,18.73)(-.06937333,-.03369333){75}{\line(-1,0){.06937333}} \multiput(24.379,23.338)(.055175258,-.033721649){97}{\line(1,0){.055175258}} \put(20.812,26.905){\makebox(0,0)[cc]{$\alpha_1$}} \put(30.474,15.608){\makebox(0,0)[cc]{$\alpha_2$}} \put(49.65,16.203){\makebox(0,0)[cc]{$\alpha_{n-2}$}} \put(56.934,27.203){\makebox(0,0)[cc]{$\alpha_{n-1}$}} \put(55.15,12.784){\makebox(0,0)[cc]{$\alpha_n$}} \put(19.771,11.743){\makebox(0,0)[cc]{$\alpha_0$}} \put(50.541,11.595){\vector(4,1){.07}}\qbezier(28.541,12.041)(39.541,8.399)(50.541,11.595) \put(27.203,24.527){\vector(-1,0){.07}}\qbezier(56.933,19.027)(49.501,26.386)(27.203,24.527) \put(53.663,26.311){\vector(2,-1){.07}}\qbezier(26.163,27.054)(40.359,33.967)(53.663,26.311) \put(25.865,18.284){\vector(-3,-2){.07}}\qbezier(54.406,23.784)(35.082,24.156)(25.865,18.284) \put(38.649,5.054){\makebox(0,0)[cc]{Fig.4 $D_n$, $n$-odd, $\lambda_{n}$}} \end{picture} \vspace{0.5cm} In a similar way the fundamental weight $\varpi_{n-1}$ generates the Weyl transformation $$ \la_{n-1}\,:\, \left\{ \begin{array}{l} \varpi_n\to\varpi_1\to\varpi_{n-1}\to 0\to\varpi_n\,, \\ \varpi_j\to\varpi_{n-j}\,,~1<j< n-1\,. \end{array} \right. $$ Thus, $\la_{n-1}=\la_n^{-1}$ and we will not consider this case. \bigskip Consider a subgroup of order two of $\clZ(\bG)$ generated by $\xi=\varpi_1$. Acting as above we find $$ \la_1\,:\, \left\{ \begin{array}{l} \varpi_1\leftrightarrow 0\,, \\ \varpi_{n-1}\leftrightarrow\varpi_n\,,\\ \varpi_j\leftrightarrow\varpi_{j}\,,~1<j< n-1\,. \end{array} \right. $$ In terms of roots the action assumes the form \beq{ldr1} \la_1\,:\,~ \left\{ \begin{array}{l} \varpi_1\leftrightarrow 0\,, \\ \varpi_{n-1}\leftrightarrow\varpi_n\,,\\ \varpi_j\leftrightarrow\varpi_{j}\,,~1<j< n-1\,. \end{array} \right\}\,,~~~ \left\{ \begin{array}{l} \al_0\leftrightarrow \al_{1}\,,\\ \al_n\leftrightarrow \al_{n-1}\,,\\ \al_j\leftrightarrow\al_{j}\,,~~1<j\leq n\,. \end{array} \right\}. \eq Explicitly, in $\pi_1$ $$ \la_1= \left( \begin{array}{cccccc} 0 & & & & & 1 \\ & Id_{n-2} & & & 0 & \\ & & 0 & 1 & & \\ & & 1 & 0 & & \\ & 0 & & & Id_{n-2} & \\ 1 & & & & & 0 \\ \end{array} \right)\,, $$ or \beq{indo1} \left( \begin{array}{cccccccccccccc} 1\,,&2\,,&\ldots,&n-1\,,& n\,,&n+1\,,&n+2\,,&\ldots, &2n-1\,,& 2n \\ 2n\,,&2 \,,&\ldots,&n-1\,,& n+1\,,&n\,,&n+2\,,&\ldots, &2n-1\,,& 1 \end{array} \right)\,. \eq In the canonical basis $(e_1,e_2,\ldots,e_n)$ $\la_n$ is represented by the matrix \beq{cbdl1} \la_1\to \La_{jk}= \left\{ \begin{array}{c} -\de_{j,k}\,,~j=1,n\\ \de_{j,k}\,,~j\neq 1,n\,. \end{array} \right. \eq \subsection*{The GS-basis.} For $n$ odd \beq{gdo} \gg_{D_n}=\gg_0+\gg_1+\gg_2+\gg_3\,,~~(\la_n(\gg_k)=i^k\gg_k)\,, \eq where $$ \gg_0=\ti\gg_0+V\,. $$ In $\Pi^{ext}$ there are one orbit of length 4 and $\frac{n-3}2$ orbits of length 2. The former orbit passes through $\al_0$ and the latter orbits contain $\Pi_1=A_{n-3}$. Since $n-3$ is even it follows from Lemma 10.1 that $\ti\gg_0=\gg_{B_{\frac{n-3}2}}$. We will demonstrate it explicitly. The subspaces in (\ref{gdo}) can be read off from the $\la_n$-action. First we find that $$ \ti\gg_0=\left\{Z= \left( \begin{array}{ccccc} 0 & & & & 0 \\ & X & & 0 & \\ & & 0 & & \\ & 0& & X & \\ 0 & & & & 0 \\ \end{array} \right)\,,~~~X= A^{(n-2)}-\ti A^{(n-2)} \right\}\,, $$ where $A^{(n-2)}$ is a matrix of order $n-2$, $\ti A^{(n-2)}=J_{n-2}A^{(n-2)}J_{n-2}$. $$ \dim\,(\ti\gg_0)=\frac{(n-2)(n-3)}2\,. $$ It is easy to see that $\ti\gg_0$ has a type $B_{\frac{n-3}2}$. Namely, $\{X= A^{(n-2)}-\ti A^{(n-2)}\}={\bf so}(n-2)$. The invariant Cartan subalgebra $\ti\gH\subset\ti\gg_0$ has a basis \beq{hdno} \ti e_2=e_2-e_{n-1}\,,\, \ti e_3=e_3-e_{n-2}\,, \ldots, \ti e_{l}= e_{l}-e_{l+2}+e_{n+l}\,, ~~~(l=\frac{n-3}2)\,. \eq An arbitrary element from $\ti\gH_0$ has the form \beq{diu} \ti\bfu=\di\,(0,u_2,\dots,u_l,0,-u_l,\dots,-u_2,0,0,u_2,\dots,u_l,0,-u_l,\dots,-u_2,0)\,. \eq The space $V$ is generated by its GS basis. In following formulas $j=2,\ldots,n-1$ \beq{gt0d} \gt^0_{1j}=\f1{2}(E_{1,j}+E_{n,n+j}+E_{2n,j}+E_{n+1,n+j})-\ldots\,,~~(j=2,\ldots,n-1) \eq $$ \gt^0_{1,n}=\f1{2}(E_{1,n}+E_{n,2n}+E_{2n,n+1}+E_{n+1,1})-\dots\,, $$ $$ \gt^0_{j,1}=\f1{2}(E_{j,1}+E_{n+j,n}+E_{j,2n}+E_{n+j,n+1}) -\ldots\,,~~(j=2,\ldots,n-1)\,, $$ $$ \gt^0_{j,n+k}=\f1{\sq2}(E_{j,n+k}+E_{n+j,k})-E_{n-k+1,2n-j+1} -E_{2n-k+1,n-j+1}\,,~~(j,k=2,\ldots,n-1)\,, $$ where $\ldots$ means the antisymmetric part of $Z$ (\ref{31}). The latter generators form the adjoint representation of $\ti\gg_0={\bf so}(n-2)$. We have $$ \dim\,(V)=\oh(n-2)(n-3)+2\cdot (n-2)+1\,. $$ Let $\underline{(n-2)}$, $\underline{1}$ be a vector and a scalar representations of ${\bf so(n-2)}$. Then $\gg_0$ is decomposed as $$ \gg_0=(\ti\gg_0={\bf so(n-2)})+2\times\underline{(n-2)}+\underline{1}+ \underline{\oh(n-2)(n-3)}\,,~~\dim\,(\gg_0)=n^2-3n+3\,. $$ It is isomorphic to ${\bf so}(n-1)+{\bf so}(n-1)+\underline{1}$. This algebra is obtained from the extended Dynkin diagram by dropping two middle roots. This procedure generates the automorphism of order four \cite{Ka}. The Killing form in $V$ is defined by (5.8.I). We write down its nonzero components \beq{kfvdo} (\gt^0_{j_1,n+k_1},\gt^0_{j_2,n+k_2})=\de_{j_1,k_2}\de_{k_1,j_2}\,,~~ (\gt^0_{1,j},\gt^0_{j,1})=1\,,~~(\gt^0_{1,n},\gt^0_{1,n})=-1\,. \eq Formally, the last pairing vanishes. However, one can consider instead nontrivial pairing $(\gt^0_{1,n},\gt^0_{n,1})$. The basic element $\gt^0_{n,1}$ is defined by the equivalent orbit passing through $E_{n,1}$. The sign minus in the last pairing arises due to the antisymmetry of matrix $Z$. Consider the GS basis in $\gg_1$ \beq{gt1d} \gh^1_{1}=\f1{\sq2}(e_{1}+ie_{n})-\ldots\,, \eq $$ \gt^1_{1,j}=\f1{2}(E_{1,j}+iE_{n,n+j}-E_{2n,j}-iE_{n+1,n+j}) \ldots\,,~~(j=2,\ldots,n-1)\,, $$ $$ \gt^1_{1,n}=\f1{2}(E_{1,n}+iE_{n,2n}-E_{2n,n+11}-iE_{n+1,1})-\dots\,, $$ $$ \gt^1_{j,1}=\f1{2}(E_{j,1}+iE_{n+j,n}-E_{j,2n}-iE_{n+j,n+1}) -\ldots\,,~~(j=2,\ldots,n-1)\,, $$ $$ \dim\,\gg_1=2(n-2)+2=2n-2\,. $$ The GS basis in $\gg_2$ is $$ \gh^1_{1}=\f1{\sq2}(e_{1}-e_{n})-\ldots\,, $$ \beq{gt2d} \gt^2_{1j}=\f1{2}(E_{1,j}-E_{n,n+j}+E_{2n,j}-E_{n+1,n+j}) \ldots\,. \eq $$ \gt^2_{1n}=\f1{2}(E_{1,n}-E_{n,2n}+E_{2n,n+1}-E_{n+1,1})-\dots\,, $$ $$ \gt^2_{j,1}=\f1{2}(E_{j,1}-E_{n+j,n}+E_{j,2n}-E_{n+j,n+1}) -\ldots\,,~~(j=2,\ldots,n-1)\,, $$ $$ \gt^2_{j,n+k}=\f1{\sq2}(E_{j,n+k}-E_{n+j,k})-\ldots\,,~~(j,k=2,\ldots,n-1)\,, $$ $$ \gt^2_{j,k}=\f1{\sq2}(E_{j,k}+E_{n+1-k,n+1-j})-\ldots\,,~~(j,k=2,\ldots,n-1)\,. $$ $$ \dim\,\gg_2=2(n-2)+2+\frac{(n-2)(n-3)}2+\frac{(n-2)(n-1)}2=n^2-2n+1\,. $$ Finely, the GS basis in $\gg_3$ is \beq{gt3d} \gh^3_{1}=\f1{\sq2}(e_{1}-ie_{n})-\ldots\,, \eq $$ \gt^3_{1j}=\f1{2}(E_{1,j}-iE_{n,n+j}-E_{2n,j}+iE_{n+1,n+j}) \ldots\,,~~(j=2,\ldots,n-1)\,, $$ $$ \gt^3_{1n}=\f1{2}(E_{1n}-iE_{n,2n}-E_{2n,n+1}+iE_{n+1,1})-\dots\,, $$ $$ \gt^3_{j,1}=\f1{2}(E_{j,1}-iE_{n+j,n}-E_{j,2n}+iE_{n+j,n+1}) -\ldots\,,~~(j=2,\ldots,n-1)\,, $$ $$ \dim\,\gg_3=2(n-2)+2=2n-2\,. $$ It follows from (5.8.I) that there is a nontrivial pairing $(\gg_2,\gg_2)$ and $(\gg_1,\gg_3)$. For the basis we have \beq{kfvdok1} (\gh^1_{1},\gh^3_{1})=1\,,~~(\gt^1_{1,j},\gt^3_{j,1})=1 \,,~~(\gt^3_{1,j},\gt^1_{1,j})=1 \,,~~(\gt^1_{1,n},\gt^3_{1,n})=-1\,, \eq $$ (\gh^2_{1},\gh^2_{1})=1\,,~~(\gt^2_{1j},\gt^2_{j,1})=1 \,,~~(\gt^2_{1,n},\gt^2_{1,n})=-1\,, $$ \beq{kfvdok2} (\gt^2_{j_1,k_1},\gt^2_{j_2,k_2})=\de_{j_1,k_2}\de_{j_2,k_1}\,,~~ (\gt^2_{j_1,n+k_1},\gt^2_{j_2,n+k_2})=-\de_{j_1,k_2}\de_{j_2,k_1}\,. \eq \bigskip Consider the GS-basis corresponding to $\la_1$ (\ref{indo1}). In this case $$ \gg_{D_n}=\gg_0+\gg_1\,,~~(\la_1(\gg_k)=(-1)^k\gg_k)\,, $$ $$ \gg_0=\ti\gg_0+V\,,~~\ti\gg_0=\gg_{B_{n-2}}={\bf so}(2n-3)\,. $$ Then $Z\in\ti\gg_0$ if $$ Z= \left( \begin{array}{cccccc} 0 & 0 & 0 & 0 & 0 & 0 \\ 0 & A^{(n-2)} & a_{jn} &a_{jn} & B^{(n-2)} & 0 \\ 0 & a_{nj} & 0 & 0 & -a^T_{jn} & 0 \\ 0 & a_{nj} & 0 & 0 & -a^T_{jn} & 0 \\ & C^{(n-2)} & -a^T_{nj} & -a^T_{nj} & -\ti A^{(n-2)} & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 \\ \end{array} \right)\in {\bf so}(2n-3)\,. $$ Here $A^{(n-2)}$, $B^{(n-2)}$ and $C^{(n-2)}$ are matrices of order $n-2$, $\,\ti A^{(n-2)}=J(A^{(n-2)})^TJ$ and\\ $\ti B^{(n-2)}=-B^{(n-2)}$, $\,\ti C^{(n-2)}=-C^{(n-2)}$. The Cartan subalgebra $\ti\gH_0\sim\gH_{B_{n-2}}$ has the form \beq{csbn} \ti\gH_0=\{\ti\bfu=\di(0,u_2,\ldots,u_{n-1},0,0,-u_{n-1},\ldots,-u_2,0\}\,. \eq For $Z\in V$ we have $$ Z= \left( \begin{array}{cccccc} 0 & a_{1j} & a_{1n} & b_{11} & -a^T_{j1} & 0 \\ a_{j1} & 0 & 0 & 0 & 0 & a_{j1} \\ -a_{1n} & 0 & 0 & 0 & 0 & -b_{11} \\ -b_{11} & 0 & 0 &0 & 0 & -a_{1n} \\ -a^T_{1j} & 0 & 0 & 0 & 0 & -a^T_{1j} \\ 0 & a_{1j} & b_{11} & a_{1n} & -a^T_{j1} & 0 \\ \end{array} \right)\,, $$ $$ V=\underline{2(n-2)+1}+\underline{1}\,. $$ Here $\underline{2(n-2)+1}$ is a vector representation of $B_{n-2}=${\bf so$(2n-3)$}. The GS-basis in $V$ takes the form \beq{gt1d2} \gt^0_{1,n}=\f1{\sq2}(E_{1,n}+E_{2n,n+1})-\ldots\,,~~ \gt^0_{1,j}=\f1{\sq2}(E_{1,j}+E_{2n,j})-\ldots \,,~(j=2,\ldots,n-1)\,, \eq $$ \gt^0_{n,1}=\f1{\sq2}(E_{n,1}+E_{n+1,2n})-\ldots\,,~~ \gt^0_{j,1}=\f1{\sq2}(E_{j,1}+E_{j,2n})-\ldots\,,~(j=2,\ldots,n-1)\,. $$ As above, $\ldots$ means the antisymmetric partners of the generators. Thus $$ \dim\,\gg_0=\dim\,\gg_{B_{n-2}}+\dim\,V=2n^2-5n+4\,. $$ The invariant subalgebra $\gg_0$ is isomorphic to ${\bf so(2n-2)}+\underline{1}$. It is obtained from the extended Dynkin diagram for ${\bf so(2n)}$ by dropping two roots roots $\al_1$ and $\al_0$. This procedure provides the second order automorphism \cite{Ka} that we consider. Non-vanishing components of the Killing form on $V$ are \beq{kfd1} (\gt^0_{1,n},\gt^0_{n,1})=-1\,,~~ (\gt^0_{1,j},\gt^0_{j,1})=1\,. \eq Consider the GS-basis in $\gg_1$ \beq{gt1d1} \gh^1_{1}=e_{1}\,,~~\gh^1_{n}=e_{n}\,, \eq $$ \gt^1_{1,j}=\f1{\sq2}(E_{1,j}-E_{2n,j}) \ldots\,, ~~ \gt^1_{1,n}=\f1{\sq2}(E_{1,n}-E_{2n,n+1})-\ldots\,, $$ $$ \gt^1_{n,1}=\f1{\sq2}(E_{n,1}-E_{n+1,2n})-\ldots\,, ~~ \gt^1_{j,1}=\f1{\sq2}(E_{j,1}-E_{j,2n}) -\ldots\,, $$ The Killing form on this basis assumes the form \beq{kfd2} (\gh^1_{1},\gh^1_{1})=1\,,~~ (\gh^1_{n},\gh^1_{n})=1\,,~~ (\gt^1_{1,n},\gt^1_{n,1})=1\,,~ ~(\gt^1_{1,j},\gt^1_{j,1})=1\,. \eq \subsection*{The Lax operators and the Hamiltonians.} \subsubsection*{Trivial bundles.} The moduli space for the trivial $\bG=Spin(2n)$- bundles is the quotient $\gH_{D_n}/(W_{D_n}\ltimes(\tau Q(D_n)+Q(D_n)))$, where $W_{D_n}$ is defined in (\ref{wed}). It implies that \beq{tdbg} u_j\sim u_j+m_j+n_j\tau\,,~~n_j\,,m_j\in\mZ\,,~\sum_{j=1}^nm_j\,{\rm~and ~}~ \sum_{j=1}^nn_j\,{\rm~ is~even} \eq (see (\ref{drl})). For trivial $G^{ad}=SO(2n)/\mu_2$-bundles the moduli space is $\gH_{D_n}/(W\ltimes(\tau P(D_n)+P(D_n)))$, where $P(D_n)$ is the $D_n$-weight lattice (\ref{pdno}). In this case instead of the shifts (\ref{tdbg}) the moduli space is defined up to the shifts \beq{tdbg1} u_j\sim u_j+m_j+n_j\tau\,,~~n_j\,,m_j\in\mZ {~\rm or}~\in\oh+\mZ ,. \eq The intermediate situation arises for trivial $SO(2n)$-bundles. In this case $\bfu$ is defined up to the shifts \beq{tdbg2} u_j\sim u_j+\tau m_j+n_j\,,~~m_j,n_j\in\mZ\,. \eq For all $Spin(2n)$, $SO(2n)$ $SO(2n)/\mu_2$ trivial bundles we have the same Lax operator \beq{trso} L(z)=\sum_{j=1}^n(v_j+S_{0,j}E_1(z))e_{j}+\sum_{j\neq k}^n S_{k,j}\phi(u_j-u_k,z)\gG^-_{j,k} +\sum_{j\neq k}^nS_{k+n,j}\phi(u_j+u_k,z)\gG^+_{j,k+n}\,, \eq and the same CM Hamiltonian \beq{dh} H=\oh\sum_{j=1}^nv^2_j-\sum_{j\neq k}^n S_{j,k} S_{k,j}E_2(u_j-u_k) -\sum_{j\neq k}^nS_{j,k+n}S_{j+n,k}E_2(u_j+u_k)\,, \eq but different configuration spaces (\ref{tdbg}), (\ref{tdbg1}), (\ref{tdbg2}). \subsubsection*{Nontrivial bundles.} \textbf{The moduli space.}\\ Consider a bundle with a nontrivial characteristic class generated by the weight $\varpi_n$. The moduli space is a quotient of $\ti\gH_0=\{ \ti\bfu=\sum_{j=1}^lu_j\ti e_j\},~$ $l=\frac{n-3}2$, (see (\ref{diu})), where $\ti\gH_0$ is a Cartan subalgebra of the invariant algebra $\ti\gg_0=\gg_{B_l} ={\bf so}(n-3)$. Following (\ref{bcrl3}) and (\ref{bcrl1}) we construct invariant coroot and coweight sublattices $$ \ti Q^\vee(B_l)=\{\sum_{j=1}^lm_j\tie_j\,,~m_j\in\mZ\,,~\sum_{j=1}^lm_j-{\rm even}\}\,, ~~~ \ti P^\vee(B_l)=\{\sum_{j=1}^lm_j\tie_j\,,~m_j\in\mZ\}\,. $$ where $\{\tie_j\}$ is the invariant basis (\ref{hdno}). Therefore, there are two types of the moduli spaces. They are defined up to the shifts by a vector $\ti\bfu$ $$ \ti\bfu\sim\ti\bfu+\tau\ga_1+\ga_2\,,~~\ga_j\in \ti Q^\vee(B_l)\,,~{\rm or}~\ti P^\vee(B_l)\,. $$ The nontrivial $Spin_{2n}$, SO$(2n)$, SO$(2n)/\mu_2$ - bundles with the characteristic class $\varpi_n$ have the same moduli space as trivial $Spin_{n-2}$, SO$(n-2)$-bundles. Two fundamental domains describe these moduli spaces. Let us define the moduli space of bundles with nontrivial characteristic class generated by the weight $\varpi_1$. In this case $\Pi_1=\Pi_{D_{n-1}}$ and $\ti\Pi=\Pi_{B_{n-2}}$ $(\ti\gg_0={\bf so}(2n-3))$. Then $\ti\bfu\in\gH_{B_{n-2}}$ (\ref{csbn}). Two lattices $\ti Q^\vee(B_{n-2})$, $\ti P^\vee(B_{n-2})$ in $\gH_{B_{n-2}}$ (\ref{bcrl3}), (\ref{bcrl1}) defines two fundamental domains $$ \ti\bfu\sim\ti\bfu+\tau\ga_1+\ga_2\,,~~\ga_j\in \ti Q^\vee(B_{n-2})\,, ~{\rm or}~\ti P^\vee(B_{n-2})\,. $$ They are the moduli spaces of trivial $Spin_{2n-3}$, SO$(2n-3)$-bundles. \textbf{The Lax operators and Hamiltonians.}\\ In the GS-basis corresponding to the characteristic class $\varpi_n$ the Lax operator is decomposed as $$ L=\ti L_0+L'_0+L_1+L_2+L_3\,, $$ where $\ti L_0$ is a Lax operator corresponding to the trivial SO$(n-2)$ bundle (\ref{tlb}). The other components take the form $$ L'_0= S'_{1,n}\bfe\,(z/2)\phi(\oh,z)\gt^0_{1,n}+ $$ $$ \sum_{j=2}^{n-1}\Bigl( S'_{j,1}\bfe\,(\frac{j-1}{2n-2}z)\phi(\frac{j-1}{2n-2}\tau-u_j,z)\gt^0_{1,j} +S'_{1,j} \bfe\,(\frac{1-j}{2n-2}z)\phi(\frac{1-j}{2n-2}\tau+u_j,z)\gt^0_{j,1} \Bigr. $$ $$ \Bigl. +\sum_{k=1}^{n-2}S'_{k,n+j} \bfe\,(\frac{2n-k-j}{2n-2}z)\phi(\frac{2n-k-j}{2n-2}\tau-u_j-u_k,z)\gt^0_{j,k+n} \Bigr)\,. $$ Here $u_j$ are elements of the diagonal matrix (\ref{diu}) $u_i=-u_{n-i}$, $(i=2,\ldots,l)$. $$ L_{k=1,3}=S^{4-k}_{1}\phi(\frac{k}{4},z)\gh^k_{1}+S^{4-k}_{1,n}\bfe(zk/4)\phi(k/4,z)\gt^k_{1,n} $$ $$ +\sum_{j=2}^{n-1}\Bigl(S^{4-k}_{j,1}\bfe\,(\frac{j-1}{2n-2}z)\phi(\frac{j-1}{2n-2}\tau-u_j +\frac{k}{4},z)\gt^k_{1,j} +S^{4-k}_{1,j} \bfe\,(\frac{1-j}{2n-2}z)\phi(\frac{1-j}{2n-2}\tau+u_j-\frac{k}{4},z)\gt^k_{j,1} \Bigr)\,, $$ $$ L_2=S^2_{1,n}\bfe(z/2)\phi(1/2,z)\gt^2_{1,n}+S^2_{1}\bfe(z/2)\phi(1/2,z)\gh^2_{1}+ $$ $$ \sum_{j=2}^{n-1}\Bigl(S^2_{j,1}\bfe\,(\frac{j-1}{2n-2}z)\phi(\frac{(j-1)\tau}{2n-2}-u_j +\frac{1}{2},z)\gt^2_{1,j} +S^2_{1,j} \bfe\,(\frac{1-j}{2n-2}z)\phi(\frac{1-j}{2n-2}\tau+u_j-\frac{1}{2},z)\gt^2_{j,1}+ \Bigr. $$ $$ \Bigl. \sum_{m=2,m\neq j}^{n-1}S^2_{m,j+n} \bfe\,(\frac{2n-m-j}{2n-2}z)\phi(\frac{2n-m-j}{2n-2}\tau-u_j-u_m-\frac{1}{2},z)\gt^2_{j,m+n} $$ $$ +\sum_{m=2,m\neq j}^{n-1}S^2_{m,j} \bfe\,(\frac{m-j}{2n-2}z)\phi(\frac{(m-j)}{2n-2}\tau-u_j+u_m-\frac{1}{2},z)\gt^2_{j,m}\Bigr)\,. $$ After the symplectic reduction with respect to the Cartan subgroup $\ti{\cal H}$ we come to Hamiltonians of integrable systems $$ H=\ti H_0+H_0'+H_2\,, $$ where $\ti H_0$ is the Hamiltonian of $B_l$ CM system (\ref{bth}), $$ H_0'=\oh(L'_0(z),L'_0(z))|_{\rm const.~part}\,,~~ H_2=(L_1(z),L_3(z))+\oh(L_2(z),L_2(z))|_{\rm const.~part}\,. $$ From (\ref{kfvdo}) and (\ref{kfvdok2}) we find $$ -H_0'=\oh\Bigl(-(S'_{1,n})^2E_2(\oh)+\sum_{j=2}^{n-1} S'_{1,j}S'_{j,1}E_2(\frac{j-1}{2n-2}\tau-u_j) +\sum_{k\neq j}^{n-1}S'_{j,n+k}S'_{k,n+j}E_2(\frac{2n-k-j}{2n-2}\tau-u_j-u_k) \Bigr)\,, $$ $$ H_2=(S^1_1S^3_1-S^1_{1,n}S^3_{1,n})E_2(1/4)-\oh(S^2_{1,n})^2E_2(\oh) +\oh\sum_{j\neq m}^{n-1}S^2_{j,m}S^2_{m,j} E_2(\frac{m-j}{2n-2}\tau-u_j+u_m-\oh) $$ $$ +\sum_{j=2}^{n-1}(S^1_{j,1}S^3_{j,1}+S^3_{j,1}S^1_{j,1}) E_2(\frac{1-j}{2n-2}\tau+u_j-\f1{4})+\oh\sum_{j\neq m}^{n-1}(S^2_{j,m+n}S^2_{m+n,j} E_2(\frac{2n-m-j}{2n-2}\tau-u_j-u_m-\oh)\,. $$ \bigskip Consider systems corresponding to the characteristic class $\varpi_1$. In this case using the GS-basis (\ref{gt1d2}), (\ref{gt1d1}) we define $L=\ti L_0+L_0'+L_1\,,$ where $\ti L_0$ is the Lax operator of $B_{n-2}$ (\ref{tlb}), $$ L'_0(z)=S'_{n,1}\bfe\,(\frac{1}{2}z)\phi(\tau/2,z)\gt^0_{1,n}+ S'_{1,n}\bfe\,(-\frac{1}{2}z)\phi(\tau/2,z)\gt^0_{n,1}+ $$ $$ \sum_{j=2}^{n-1}\Bigl( S'_{1,j}\bfe\,(\frac{j-1}{2n-2}z)\phi(\frac{(j-1)\tau}{2n-2}+u_j,z)\gt^0_{1j} +S'_{j,1} \bfe\,(\frac{1-j}{2n-2}z)\phi(\frac{(1-j)\tau}{2n-2}-u_j,z)\gt^0_{j,1} \Bigr)\,, $$ $$ L_1(z)=(S^1_{1}\gh^1_{1}+ S^1_{n}\gh^1_{n})\phi(\frac{1}{2},z)+(S^1_{1,n}\gt^1_{n,1}+S^1_{n,1}\gt^1_{1,n}) \bfe\,(\oh z)\phi(\frac{1}{2}(1+\tau),z) $$ $$ +\sum_{j=2}^{n-1}\Bigl(S^k_{1,j}\bfe\,(\frac{j-1}{2n-2})\phi(\frac{(j-1)\tau}{2n-2}-u_j +\oh,z)\gt^1_{1j} +S^1_{j,1} \bfe\,(\frac{1-j}{2n-2})\phi(\frac{(1-j)\tau}{2n-2}+u_j+\oh,z)\gt^1_{j,1} \Bigr)\,. $$ Then we come to the Hamiltonians $\ti H_0=H_{B\frac{n-3}2}$ (\ref{bth}) and $$ H'_0=-\oh S'_{1,n}S'_{n,1}\wp(\tau/2)+\sum_{j=2}^{n-1}S'_{1,j}S'_{j,1} E_2(\frac{(j-1)\tau}{2n-2}-u_j)\,, $$ $$ -2H_1=((S^1_1)^2+(S^1_n)^2)E_2(\oh)+ \sum_{j=2}^{n-1}S^1_{1,j}S^1_{j,1}E_2(\frac{(j-1)\tau}{2n-2}+\oh-u_j) +S^1_{1,n}S^1_{n,1}E_2(\frac{1}{2}(1+\tau)) \,. $$ \section{SO$(2n)$\,, Spin$(2n)\,$, $n=2l$.} \setcounter{equation}{0} \subsection*{Lattices and characteristic classes} For $n=2l$ \beq{lfen} P(D_n)=Q(D_n)+\mZ\varpi_a+\mZ\varpi_{b}\,,~~a,b=(1,n-1,n)\,,~a\neq b\,. \eq The weight lattice $P(D_n)$ contains apart from $Q(D_n)$ three sublattices of index $2$ generated by these weights \beq{wlen} P^V(D_n)=Q(D_n)+\mZ\varpi_1\,, \eq $$ P^R(D_n)=Q(D_n)+\mZ\varpi_{n-1}\,,~~ P^L(D_n)=Q(D_n)+\mZ\varpi_{n}\,. $$ The sublattice $P^V(D_n)$ is self-dual. If $n=8l$ then $^L(P^R(D_n))=P^L(D_n)$. For $n=8l+4\,$ $\,^L(P^R(D_n))=P^R(D_n)$, and $^L(P^L(D_n))=P^L(D_n)$. In other words, we have the following hierarchy of the lattices $$ \begin{array}{ccccc} & &P & \\ & \swarrow & \downarrow & \searrow & \\ P^{R} & & P^V & & P^{L} \\ & \searrow & \downarrow & \swarrow & \\ & &Q & & \end{array} $$ The center of $Spin(2n)$ for even $n$ is $\clZ(\bar G)\sim\mu_2\times\mu_2$. The group element $\zeta=\bfe\,(\xi)$ for $\xi=\varpi_n$ generates one of the subgroups $\mu_2$. Putting $\xi=\varpi_n$ and $C_{alc}$ (\ref{alcd}) in (3.10.I) we find $\la_n$ $$ \la_n\,:\, \left\{ \begin{array}{l} 0\leftrightarrow \varpi_{n}\,,~~\varpi_1\leftrightarrow\varpi_{n-1}\,, \\ \varpi_j\leftrightarrow\varpi_{n-j}\,,~1<j<n-1\,. \end{array} \right. $$ In terms of roots the action assumes the form $\al_0\leftrightarrow\al_n \,,~\al_1\leftrightarrow \al_{n-1}\,,\ldots$ $\al_j\leftrightarrow\al_{n-j}\,,$\\ $1<j<n-1$. In the representation $\pi_1$ (\ref{31}) the $\la_n$ action takes the form $$ \la_n=\left( \begin{array}{cc} 0 & Id_n \\ Id_n & 0 \\ \end{array} \right)\,. $$ Its action on the indices of the basis $(e_1,e_2,\ldots,e_{2n})$ is \beq{soel} \la_n\,:\,\left( \begin{array}{cccccc} 1 & j & n & n+1 & n+j & 2n \\ n+1 & n+j & 2n & 1 & j & n \end{array} \right)\,. \eq In the canonical basis $(e_1,e_2,\ldots,e_n)$ $\la_n$ in $\gH_{D_{2l}}$ is represented by the matrix $$ \la_n\to E_{jk}=-\de_{j,n-k+1}\,. $$ \vspace{0.5cm} \unitlength 1mm \linethickness{0.4pt} \ifx\plotpoint\undefined\newsavebox{\plotpoint}\fi \begin{picture}(57.427,33.149)(0,0) \put(21.257,23.636){\circle{1.808}} \put(28.392,19.325){\circle{2.081}} \put(35.527,19.622){\circle{2.081}} \put(49.501,19.622){\circle{1.808}} \put(56.487,23.487){\circle{1.88}} \put(56.339,15.609){\circle{1.808}} \put(21.257,15.757){\circle*{1.88}} \put(22.149,23.19){\circle*{.297}} \put(29.73,19.473){\circle*{.297}} \put(29.73,19.473){\line(1,0){4.757}} \put(36.717,19.622){\line(1,0){5.054}} \put(46.082,19.473){\line(1,0){.149}} \put(46.082,19.622){\line(1,0){2.378}} \multiput(55.447,23.041)(-.05679775,-.03340449){89}{\line(-1,0){.05679775}} \multiput(50.244,18.73)(.06937333,-.03369333){75}{\line(1,0){.06937333}} \multiput(27.5,18.582)(-.06937333,-.03369333){75}{\line(-1,0){.06937333}} \multiput(22,23.19)(.055175258,-.033721649){97}{\line(1,0){.055175258}} \put(18.433,26.757){\makebox(0,0)[cc]{$\alpha_1$}} \put(28.095,15.46){\makebox(0,0)[cc]{$\alpha_2$}} \put(47.271,16.055){\makebox(0,0)[cc]{$\alpha_{n-2}$}} \put(54.555,27.055){\makebox(0,0)[cc]{$\alpha_{n-1}$}} \put(52.771,12.636){\makebox(0,0)[cc]{$\alpha_n$}} \put(17.392,11.595){\makebox(0,0)[cc]{$\alpha_0$}} \put(49.352,27.054){\vector(2,-1){.07}}\qbezier(23.933,26.757)(35.602,33.149)(49.352,27.054) \put(46.528,23.041){\vector(3,-1){.07}}\qbezier(30.027,23.041)(38.426,26.311)(46.528,23.041) \put(46.974,13.676){\vector(4,1){.07}}\qbezier(26.163,13.527)(37.163,10.777)(46.974,13.676) \put(35.974,5.351){\makebox(0,0)[cc]{Fig.5 $D_n$, $n$-even, $\lambda_n$}} \end{picture} \vspace{0.5cm} Let us take $\xi=\varpi_{n-1}$. Then the corresponding Weyl transformation $\la_{n-1}$ acts on $C_{alc}$ as $$ \la_{n-1}\,:\, \left\{ \begin{array}{l} 0\leftrightarrow\varpi_{n-1}\,,~~\varpi_1\leftrightarrow\varpi_{n}\,, \\ \varpi_j\leftrightarrow\varpi_{n-j}\,,~1<j<n-1\,, \end{array} \right. $$ or $\al_{n-1}\leftrightarrow \al_0\,,$ $ \al_{n}\leftrightarrow\al_1\,,$ $\al_j\leftrightarrow\al_{n-j}\,,$ $1<j<n-1$\,. \vspace{0.5cm} \unitlength 1mm \linethickness{0.4pt} \ifx\plotpoint\undefined\newsavebox{\plotpoint}\fi \begin{picture}(59.806,33.967)(0,0) \put(23.636,23.784){\circle{1.808}} \put(30.771,19.473){\circle{2.081}} \put(37.906,19.77){\circle{2.081}} \put(51.88,19.77){\circle{1.808}} \put(58.866,23.635){\circle{1.88}} \put(58.718,15.757){\circle{1.808}} \put(23.636,15.905){\circle*{1.88}} \put(24.528,23.338){\circle*{.297}} \put(32.109,19.621){\circle*{.297}} \put(32.109,19.621){\line(1,0){4.757}} \put(39.096,19.77){\line(1,0){5.054}} \put(48.461,19.621){\line(1,0){.149}} \put(48.461,19.77){\line(1,0){2.378}} \multiput(57.826,23.189)(-.05679775,-.03340449){89}{\line(-1,0){.05679775}} \multiput(52.623,18.878)(.06937333,-.03369333){75}{\line(1,0){.06937333}} \multiput(29.879,18.73)(-.06937333,-.03369333){75}{\line(-1,0){.06937333}} \multiput(24.379,23.338)(.055175258,-.033721649){97}{\line(1,0){.055175258}} \put(20.812,26.905){\makebox(0,0)[cc]{$\alpha_1$}} \put(30.474,15.608){\makebox(0,0)[cc]{$\alpha_2$}} \put(49.65,16.203){\makebox(0,0)[cc]{$\alpha_{n-2}$}} \put(56.934,27.203){\makebox(0,0)[cc]{$\alpha_{n-1}$}} \put(55.15,12.784){\makebox(0,0)[cc]{$\alpha_n$}} \put(19.771,11.743){\makebox(0,0)[cc]{$\alpha_0$}} \put(27.203,24.527){\vector(-1,0){.07}}\qbezier(56.933,19.027)(49.501,26.386)(27.203,24.527) \put(25.865,18.284){\vector(-3,-2){.07}}\qbezier(54.406,23.784)(35.082,24.156)(25.865,18.284) \put(38.649,5.054){\makebox(0,0)[cc]{Fig.6 $D_n$, $n$-even, $\lambda_{n-1}$}} \end{picture} In the representation $\pi_1$ (\ref{31}) the $\la_{n-1}$ action takes the form $$ \la_{n-1}= \left( \begin{array}{cccccc} 0 &0 & 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & Id_{n-2}&0 \\ 1 &0 & & 0 & 0 & 0 \\ 0 &0 & & 0 & 0 & 1 \\ 0& Id_{n-2} & 0 & 0& 0 & 0 \\ 0 & 0 & 0 & 1 & 0 & 0\\ \end{array} \right)\,. $$ It is conjugated to $\la_{n}$ by the matrix $$ \left(\begin{array}{cccc} Id_{n-1} & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 0 & Id_{n-1} \end{array} \right)\,. $$ We will not consider the corresponding GS basis. The case $\xi=\varpi^\vee_1$ was already considered (\ref{indo1}). In this case $\ti\gg_0={\bf so}(2n-3)$. \subsection*{The GS-basis.} The $\la_n$ action on $Z$ in $\pi_1$ takes the form $$ \la_n(Z)=\la_n \left( \begin{array}{cc} A & B \\ C & -\ti A \\ \end{array} \right)=\left( \begin{array}{cc} -\ti A & C \\ B & A \\ \end{array} \right)\,, $$ Since $\la_n^2=Id$ \beq{cgr} \gg_{D_n}=\gg_0+\gg_1\,, \eq where $$ \gg_0=\left\{ \left( \begin{array}{cc} X & Y \\ Y & X \\ \end{array} \right)\,\begin{array}{lc} | & \ti X=-X \\ | & \ti Y=-Y \end{array} \right\}\,, ~~~ \gg_1=\left\{ \left( \begin{array}{cc} X & Y \\ -Y & -X \\ \end{array} \right)\,\begin{array}{lc} | & \ti X=X \\ | & \ti Y=-Y \end{array} \right\}\,, $$ where \beq{cgr1} \gg_0=\ti\gg_0+V\,,~~\ti\gg_0 =\left\{ \left( \begin{array}{cc} X & 0 \\ 0 & X \\ \end{array} \right)\,\right\}\,,~~V=\left\{ \left( \begin{array}{cc} 0 & Y \\ Y & 0 \\ \end{array} \right) \right\}\,. \eq According to Lemma 9.1 $\ti\gg_0=\gg_{D_{\frac{n}2}}={\bf so}(n)$. The Cartan subalgebra $\ti\gH_0$ in $\gg_0$ has the basis \beq{bne} \begin{array}{l} \tie_1=\di(1,0,\ldots,0,-1,1,0,\ldots,0,-1)\,, \\ \tie_2=\di(0,1,\ldots,-1,0,0,1,\ldots,-1,0)\,, \\ \ldots\ldots,\\ \tie_l=\di(0,\ldots,1,-1,\dots,0,0,\ldots,1,-1,\dots,0)\,. \end{array} \eq It defines the moduli space of the SO$(2n,\mC)$-bundles with characteristic class corresponding to $\varpi_n$ \beq{csinv} \ti\gH_0=\{\ti\bfu=\sum_{j=1}^lu_j\tie_j\}\,. \eq From (\ref{soel}) the GS-basis in $V$ takes the form \beq{Vn} \gt^0_{j,n+k}=\f1{\sq2}(E_{j,n+k}+E_{n+j,k}-E_{n-k+1,2n-j+1} -E_{2n-k+1,n-j+1})\,,~~(j,k=1,\ldots,n\,,~j\neq k) \eq with the Killing form \beq{pvn} (\gt^0_{j_1,n+k_1},\gt^0_{j_2,n+k_2})=\de_{j_1,k_2}\de_{k_1,j_2}\,. \eq The GS-basis in $\gg_1$ is $$ \gh_{j}^1=\f1{\sq2}(e_{j}-e_{n+j}+e_{n-j+1}-e_{2n-j+1})\,,~~(j=1,\ldots,l)\,, $$ \beq{gs1e} \gt^1_{j,k}=\f1{\sq2}(E_{j,k}-E_{n+j,n+k}+E_{n-k+1,n-j+1}-E_{2n-k+1,2n-j+1})\,,~~(j,k=1,\ldots,l)\,, \eq $$ \gt^1_{j,n+k}=\f1{\sq2}(E_{j,n+k}-E_{n+j,k}-E_{n-k+1,2n-j+1} +E_{2n-k+1,n-j+1})\,,~~(j,k=1,\ldots,l\,,~j\neq k)\,, $$ with the Killing form \beq{pg1} (\gh_{j}^1,\gh_{k}^1)=\de_{j.k}\,,~~ (\gt^1_{j_1,k_1},\gt^1_{j_2,k_2})=\de_{j_1,k_2}\de_{j_2,k_1}\,, \eq $$ (\gt^1_{j_1,n+k_1},\gt^1_{j_2,n+k_2})=-\de_{j_1,k_2}\de_{j_2,k_1}\,. $$ \bigskip The GS-basis and the Lax operator related to the characteristic class defined by $\varpi_{n-1}$ are conjugated to the GS-basis and the Lax operator related to the characteristic class defined by $\varpi_{n}$. Therefore the corresponding Hamiltonians coincide. \subsection*{Lax operators and Hamiltonians.} Consider a bundle with nontrivial characteristic class generated by the weight $\varpi_n$. The moduli space is a quotient of $\ti\gH_0=\{ \ti\bfu=\sum_{j=1}^lu_j\ti e_j\},~$ $l=\frac{n}2$, where $\ti\gH_0$ is a Cartan subalgebra of the invariant algebra $\ti\gg_0=\gg_{D_l}$ (\ref{csinv}) For $l$ odd there are three types of invariant sublattices $\ti Q(D_l)$ (\ref{drl}), $\ti P(D_l)$ (\ref{pdno}) and $\ti P_2(D_l)$ (\ref{P2D}). They define three types of moduli spaces for these bundles as the fundamental domains \beq{lod} \ti\bfu\sim\ti\bfu+\ga_1\tau+\ga_2\,, ~~\ga_j\in\ti Q(D_l)\,,~{\rm or}~ \ti P(D_l)\,,~{\rm or}~\ti P_2(D_l)\,. \eq For $l$ even the invariant sublattices are $\ti Q(D_l)$, $\ti P^L(D_l)$, $\ti P^R(D_l)$ and $\ti P^V(D_l)$ (\ref{wlen}). Therefore, in this case there are four types of the moduli spaces. \beq{lev} \ti\bfu\sim\ti\bfu+\ga_1\tau+\ga_2\,, ~~\ga_j\in\ti Q(D_l)\,,~{\rm or}~ \ti P^L(D_l)\,,~{\rm or}~\ti P^R(D_l)~{\rm or}~\ti P^V(D_l)\,. \eq Consider the Lax operator for bundles with the characteristic class defined by $\varpi_n$ in the GS-basis (\ref{Vn}), (\ref{gs1e}) $L=\ti L_0+L_0'+L_1$. Here $\ti L_0$ is the Lax operator of $D_{\frac{n}2}$, $$ L'_0(z)= \sum_{j\neq k}^{n} S'_{k,n+j} \bfe\,(\frac{2n-j-k}{2n-2}z)\phi(\frac{2n-j-k}{2n-2}\tau-u_j-u_k,z)\gt^0_{j,n+k}\,, $$ $$ L_1(z)=\sum_{j=1}^{n}S^1_{j}\phi(\frac{1}{2},z)\gh^1_{j}+ \sum_{j\neq k}^{n} S^1_{j,k} \bfe\,(\frac{k-j}{2n-2}z)\phi(\frac{k-j}{2n-2}\tau-u_j+u_k+\oh,z)\gt^1_{k,j} $$ $$ -\sum_{j\neq k}^{n}S^1_{j,n+k}\bfe\,(\frac{2n-j-k}{2n-2}z) \phi(\frac{2n-j-k}{2n-2}\tau-u_j-u_k-\oh,z)\gt^1_{k,n+j}\,. $$ From (\ref{pvn}) and (\ref{pg1}) after the diagonal reduction we come to the Hamiltonians $$ H'_0=-\oh\sum_{j\neq k}^{n}S^1_{j,n+k}S^1_{k,n+j} E_2(\frac{2n-j-k}{2n-2}\tau-u_j-u_k)\,, $$ $$ H_1=-\oh\sum_{j=1}^n(S^1_{j})^2E_2(\frac{1}{2})-\oh\sum_{j\neq k}^{n}S^1_{j,k}S^1_{k,j}E_2(\frac{k-j}{2n-2}\tau-u_j+u_k-\oh) $$ $$ +\oh\sum_{j\neq k}^{n}S^1_{j,n+k}S^1_{n+k,j}E_2(\frac{2n-j-k}{2n-2}\tau-u_j-u_k-\oh)\,. $$ Note that in all expressions $u_j=-u_{n+1-j}$. Summarizing, the Hamiltonian $H^{CM}_{D_l}+H_0'+H_1$ describes the integrable systems corresponding to the bundles with characteristic classes defined by $\varpi_n$, or $\varpi_{n-1}$ with the moduli spaces (\ref{lod}) for $l$ odd, or (\ref{lev}) for $l$ even. \section{E$_6$\,.} \setcounter{equation}{0} \subsection*{Roots and weights} The Cartan subalgebra of the Lie algebra ${\bf e_6}$ is the space \beq{cse6} \gH({\bf e_6})=\{\bfu\in\mC^7\,|\, u_5+u_6+u_7=0\}\,. \eq The root system $R({\bf e_6})$ is related to the root system $R(\bas8)$ of $\bas8$ $$ R(\bas8)=\sum_{j\neq k}^4\pm e_j\pm e_k\,,~~\sharp\,(R(\bas8))=24\,. $$ Let $$ P^L=\{\varpi^L_a\}\,,~P^R=\{\varpi_a^R\}\,,~P^V=\{\varpi_a^V\}\,,~~(a=1,\ldots,8) $$ be the weights of the left (L) and right (R) spinor representations $\underline{8}^L$, $\underline{8}^R$ and the vector representation (V) $\underline{8}^V$ of $\bas8$. They are equal to the following combinations of the basic vectors \beq{we} \begin{array}{l} \varpi^L_a\to\oh\sum_{k=1}^4\pm e_k \,,~~{\rm even~number~of~negative~terms}\,,\\ \varpi^R_a\to\oh\sum_{k=1}^4\pm e_k \,,~~{\rm odd~number~of~negative~terms}\,, \\ \varpi^V_a\to\pm e_k \,. \end{array} \eq In these terms \beq{are6r} R({\bf e_6})=R(\bas8)\cup (P^L\pm\f1{\sqrt{2}}(e_5-e_7)) \cup (P^R\pm\f1{\sqrt{2}}(e_5-e_6))\cup (P^V\pm\f1{\sqrt{2}}(e_6-e_7))= \eq $$ R(\bas8)\cup \{ \al_{a,\pm}^L=(\varpi_a^L\pm\f1{\sqrt{2}}(e_5-e_7))\,, \al_{a,\pm}^R =(\varpi_a^R\pm\f1{\sqrt{2}}(e_5-e_6)) \,, \al_{a,\pm}^V=(\varpi_a^V\pm\f1{\sqrt{2}}(e_6-e_7))\}\,. $$ The systems of ${\bf e_6}$ roots is self-dual and the corresponding Dynkin diagram is simply-laced. Since $\sharp(P^L)=\sharp(P^R)=\sharp(P^V)=16$ the number of roots is equal $\sharp(R({\bf e_6}))=24+16+16+16=72$. We have from here $\dim\,{\bf e_6}={\rm rank}\,{\bf e_6}+\sharp(R({\bf e_6}))=78$. It follows from here that ${\bf so(8)}$ is subalgebra of ${\bf e_6}$. It can be found that \beq{e6d} {\bf e_6}={\bf so(8)}\oplus \clJ=\underline{28}+\underline{50}\,, \eq where $\clJ$ is a representation space of ${\bf so(8)}$. It is decomposed on irreducible components as \beq{clJ} \clJ=2\times\underline{1}+2\times\underline{8}^L+2\times\underline{8}^R+ 2\times\underline{8}^R\,. \eq Here two scalar representations complete the Cartan subalgebra $\gH(\bas8)$ to the Cartan subalgebra $\gH({\bf e_6})$ (\ref{cse6}). The basis of simple roots can be chosen as \beq{sr} \Pi=\left\{ \begin{array}{l} \al_1=\oh(e_4-e_3-e_2-e_1)+\f1{\sqrt{2}}(e_5-e_6)\,, \\ \al_2=e_3-e_4 \,,\\ \al_3=e_2-e_3\,, \\ \al_4=e_1-e_2\,, \\ \al_5=-e_1+ \f1{\sqrt{2}}(e_6-e_7)\,,\\ \al_6=e_4+e_3\,. \end{array} \right. \eq The subsystem of simple roots \beq{p1o8} \Pi_1=\{\al_2,\al_3,\al_4,\al_6\} \eq is a system of simple roots of subalgebra ${\bf so(8)}$. The Weyl group $W_{\bf e_6}$ of $R({\bf e_6})$ is generated by $W_{\bf so(8)}$ (\ref{wed}) and by the reflections $s_{\al_1}$, $s_{\al_5}$ \beq{wee6} W_{\bf e_6}=\{W_{\bf so(8)}\,,~(e_1\leftrightarrow-e_1\,,e_6\leftrightarrow e_7)\,, ~(e_j\leftrightarrow-e_j\,, \,j=1,\ldots,6)\}\,. \eq The defined below $\la_1$ (\ref{lae6}) is an element from $W_{\bf e_6}$ The simple roots (\ref{sr}) define the fundamental Weyl chamber $$ C=\{\bfu\in\gH\,|\,u_1>u_2>u_3>u_4>0\,,~u_5-u_6>\sqrt{2}\sum_{j=1}^4u_j\,, ~u_6-u_7>\sqrt{2}u_1\}\,. $$ The minimal root is \beq{mre6} \al_0=-\oh(e_4+e_3+e_2+e_1)-\f1{\sqrt{2}}(e_5-e_7) \eq $$ =-(\al_1+2\al_2+3\al_3+2\al_4+\al_5+2\al_6)\,. $$ Then \beq{wce6} C_{alc}=\{\bfu\in\gH~|~u_1>u_2>u_3>u_4>0\,, \eq $$ u_5-u_6>\sqrt{2}\sum_{j=1}^4u_j\,, ~u_6-u_7>\sqrt{2}u_1\,,~ \oh\sum_{j=1}^4u_j+\sqrt{2}(u_5-u_7)<1\}\,. $$ The subset of positive roots corresponding to $\Pi_{\bf e_6}$ assumes the form \beq{pre6} R^+({\bf e_6})=R^+(\bas8)\cup (P^L+\f1{\sqrt{2}}(e_5-e_7)) \cup (P^R+\f1{\sqrt{2}}(e_5-e_6))\cup (P^V+\f1{\sqrt{2}}(e_6-e_7))\,. \eq This data allows one to construct the Chevalley basis in ${\bf e_6}$. It follows from (\ref{sr}) that the root lattice of ${\bf e_6}$ is \beq{rle6} Q({\bf e_6})=Q(\bas8)+\mZ(-e_1+\f1{\sqrt2}(e_6-e_7))+ \mZ(\oh(e_4-e_3-e_2-e_1)+\f1{\sqrt2}(e_5-e_6))\,. \eq The fundamental weights dual to $\Pi_{\bf e_6}$ (\ref{sr}) are \beq{fwe6} \left\{ \begin{array}{l} \varpi_1 =\frac{\sqrt 2}{3}(2e_5-e_6-e_7)= \f1{3}(4\al_1+5\al_2+6\al_3+4\al_4+2\al_5+3\al_6),,\\ \varpi_2=\oh(e_1+e_2+e_3-e_4)+\f1{3\sqrt 2}(5e_5-e_6-4e_7)= \f1{3}(5\al_1+10\al_2+12\al_3+8\al_4+4\al_5+6\al_6)\,, \\ \varpi_3=e_1+e_2+\sqrt 2(e_5-e_7)= 2\al_1+4\al_2+6\al_3+4\al_4+2\al_5+3\al_6 \,,\\ \varpi_4=e_1+\f1{3\sqrt 2}(4e_5+e_6-5e_7)= \f1{3}(4\al_1+8\al_2+12\al_3+10\al_4+5\al_5+6\al_6)\,, \\ \varpi_5=\frac{\sqrt 2}{3}(e_5+e_6-2e_7)= \f1{3}(2\al_1+4\al_2+6\al_3+5\al_4+4\al_5+3\al_6)\,, \\ \varpi_6=\oh(e_1+e_2+e_3+e_4)+\f1{\sqrt 2}(e_5-e_7) =\al_1+2\al_2+3\al_3+2\al_4+\al_5+2\al_6 \,. \end{array} \right. \eq The vector $\rho_{\bf e_6}$ takes the form $\rho_{\bf e_6}=\rho_{\bf so(8)}+\frac{8}{\sqrt{2}}(e_5-e_7)=3e_1+2e_2+e_3+ \frac{8}{\sqrt{2}}(e_5-e_7)$ and since $h=12$ \beq{re6} \ka_{\bf e_6}=\f1{12}\rho_{\bf so(8)}+\frac{2}{3\sqrt{2}}(e_5-e_7)=\f1{4}e_1+\f1{6}e_2+\f1{12}e_3+ \frac{2}{3\sqrt{2}}(e_5-e_7)\,. \eq The weight lattice $P({\bf e_6})$ is generated by the root lattice $Q({\bf e_6})$ (\ref{wle6}) and one of the fundamental weights $\varpi_k\,$ $\,(k=1,2,4,5)$ \beq{wle6} P({\bf e_6})=Q({\bf e_6})+\mZ\varpi_1\,. \eq The factor group $P/Q$ being isomorphic to the center of the universal covering group $E_6$ is \beq{cent} P({\bf e_6})/Q({\bf e_6})\sim\mu_3\,. \eq It follows from (\ref{mre6}) that the fundamental alcove (\ref{wce6}) has the vertices $$ C_{alc}=(0,\varpi_1,\oh\varpi_2,\f1{3}\varpi_3,\oh\varpi_4,\varpi_5,\oh\varpi_6)\,. $$ The transformation $\la_1\in\G_{C_{alc}}$ generated by $\varpi_1$ acts on the extended Dynkin diagram as \beq{lae6} \la_1=\left\{ \begin{array}{l} \al_1\to \al_0 \,,\\ \al_2\to \al_6\,, \\ \al_3\to \al_3\,,\\ \al_4\to \al_2\,, \\ \al_5\to \al_1\,, \\ \al_6\to \al_4\,, \\ \al_0\to \al_5\,, \end{array} ~~~~ \begin{array}{l} e_1\to \oh(e_1+e_2+e_3-e_4)\,,\\ e_2\to \oh(e_1+e_2-e_3+e_4)\,,\\ e_3\to \oh(e_1-e_2+e_3+e_4)\,,\\ e_4\to \oh(e_1-e_2-e_3-e_4)\,,\\ e_5\to e_6\,,\\ e_6 \to e_7\,, \\ e_7 \to e_5\,. \end{array} \right. \eq \unitlength 1mm \linethickness{0.4pt} \ifx\plotpoint\undefined\newsavebox{\plotpoint}\fi \begin{picture}(60.684,68.413)(0,0) \put(22.804,37.477){\circle{3.182}} \put(33.057,37.3){\circle{2.55}} \put(41.012,37.653){\circle{2.475}} \put(50.205,37.3){\circle{2.828}} \put(59.397,37.477){\circle{2.574}} \put(41.012,27.931){\circle{2.5}} \put(24.042,37.3){\line(1,0){7.248}} \put(31.289,37.3){\line(1,0){.53}} \multiput(34.471,37.477)(.883883,.029463){6}{\line(1,0){.883883}} \multiput(42.25,37.653)(1.090123,-.029463){6}{\line(1,0){1.090123}} \multiput(51.972,37.477)(.972272,.029463){6}{\line(1,0){.972272}} \put(41.012,36.062){\line(0,-1){7.071}} \put(40.659,17.854){\circle*{2.574}} \put(40.659,26.693){\line(0,-1){7.248}} \put(58.867,42.426){\vector(3,-4){.07}}\qbezier(22.981,42.78)(39.333,68.413)(58.867,42.426) \put(49.321,43.841){\vector(3,-4){.07}}\qbezier(32.704,43.134)(38.891,56.922)(49.321,43.841) \put(44.371,29.345){\vector(-2,-1){.07}}\qbezier(50.558,33.234)(48.702,31.289)(44.371,29.345) \put(44.548,20.329){\vector(-3,-2){.07}}\qbezier(57.983,32.704)(55.154,26.517)(44.548,20.329) \put(33.057,34.648){\vector(-1,4){.07}}\qbezier(36.239,29.345)(33.941,30.229)(33.057,34.648) \put(24.042,32.35){\vector(-1,2){.07}}\qbezier(35.532,20.683)(28.549,23.335)(24.042,32.35) \put(20.153,40.835){\makebox(0,0)[cc]{$\alpha_1$}} \put(29.875,41.012){\makebox(0,0)[cc]{$\alpha_2$}} \put(39.421,41.366){\makebox(0,0)[cc]{$\alpha_3$}} \put(48.437,41.543){\makebox(0,0)[cc]{$\alpha_4$}} \put(57.276,41.189){\makebox(0,0)[cc]{$\alpha_5$}} \put(35.002,25.633){\makebox(0,0)[cc]{$\alpha_6$}} \put(31.997,17.678){\makebox(0,0)[cc]{$\alpha_0$}} \put(38.184,11.137){\makebox(0,0)[cc]{Fig.7 ${\bfe_6}$,$~\lambda_1$}} \end{picture} Note that being restricted on $\bas8$ $\,\la$ acts on the fundamental weights as $$ \varpi^V \to\varpi^R\to\varpi^L\,. $$ \subsection*{Chevalley basis} We start with the Chevalley basis in subalgebra ${\bf so(8)}$. It is generated by the canonical basis $(e_1,e_2,e_3,e_4)$ in $\gH(\bas8)$ and the root subspaces \beq{rso8} \begin{array}{ll} \al_{jk}=(e_j-e_k)\to E_{j,k}\,, ~~j\neq k \,,& \lan\ka_{\bf e_6},\al_{jk}\ran=\frac{j-k}{12} \,,\\ \al_{j,k+4}= (e_j+e_k)\to E_{j,k+4}\,,~~j\neq k \,, & \lan\ka_{\bf e_6},\al_{jk}\ran=\frac{j+k-8}{12} \,, \\ \al_{j+4,k}=(-e_j-e_k)\to E_{j+4,k}\,,~~j\neq k ,k+4\,,~~j\neq k \,, & \lan\ka_{\bf e_6},\al_{jk}\ran=\frac{-j-k+8}{12} \,, \,. \end{array} \eq The root subspaces in $R({\bf e_6})\setminus R(\bas8)$ are representations $2\times\underline{8}^L+2\times\underline{8}^R+2\times\underline{8}^R$ of $\bas8$ (\ref{clJ}) $$ \al_{a,\pm}^L=(\varpi_a^L\pm\f1{\sqrt{2}}(e_5-e_7))\to E_{a,\pm}^L \,,~~a=1,\ldots,8\,,~~ \lan\ka_{\bf e_6}, \al_{a,\pm}^L\ran=\pm\frac{2}3+\lan\ka_{\bf so(8)}, \varpi^L_{a}\ran\,, $$ \beq{sroe6} \al_{a,\pm}^R =(\varpi_a^R\pm\f1{\sqrt{2}}(e_5-e_6))\to E_{a,\pm}^R \,,~~a=1,\ldots,8\,, ~~ \lan\ka_{\bf e_6}, \al_{a,\pm}^R\ran=\pm\frac{1}3+\lan\ka_{\bf so(8)}, \varpi^R_{a}\ran\,, \eq $$ \al_{a,\pm}^V=(\varpi_a^V\pm\f1{\sqrt{2}}(e_6-e_7))\to E_{a,\pm}^V \,,~~a=1,\ldots,8 ~~ \lan\ka_{\bf e_6}, \al_{a,\pm}^V\ran=\pm\frac{1}3+\lan\ka_{\bf so(8)}, \varpi^V_{a}\ran\,. $$ It follows from (\ref{we}) that $-\varpi_a^{L,R,V}$ is again a spinor weight of the same type. We denote by $E_{a,\pm}^A$ the Chevalley generator related to $\varpi_a^A$. The Killing form in the Chevalley basis assumes the form $$ (e_j,e_k)=\de_{j,k}\,,~~~( E_{j,k},E_{l,m})=\de_{k,l}\de_{j,m}\,,~~(j,k=1,\ldots,4)\,, $$ \beq{kfc6} (E_{a,\pm}^A,E_{b,\mp}^B)=\de_{a,-b}\de^{A,B}\,, ~~A,B=L,R,V\,. \eq \subsection*{GS basis} Consider the Weyl action $\la_1$ (\ref{lae6}) on the root spaces of ${\bf e_6}$. It easy to see that $\la_1$ preserves ${\bf so(8)}$ and therefore $\clJ$ in (\ref{e6d}). \subsubsection*{GS basis in ${\bf so(8)}$.} Consider the action of $\la_1$ on ${\bf so(8)}$. Note first, that an orbit of $\la_1$ that does not contains the minimal root is $\Pi_1$ (\ref{p1o8}). It is a basis in the Cartan subalgebra $\gH(\bas8)$. Since $\la_1^3=1$ \beq{g2c} \gH(\bas8)=\ti\gH_0+\gH_1+\gH_2\,,~~~\la_1\,\gH_j=\om^j\gH_j\,,~~\om=\exp\,\frac{2\pi i}3\,. \eq Here $\ti\gH_0$ is a Cartan subalgebra in $\ti\gg_0$. The basis $\ti\Pi^\vee$ of simple coroots is \beq{srg2a} \ti\Pi^\vee=\{\ti\al^\vee_3=\sum_{k=0}^2\la^k\al_3=3\al_3= 3(e_2-e_3)\,,~~ \ti\al^\vee_2=\sum_{k=0}^2\la^k\al_2=(e_1-e_2+2e_3)\}\,. \eq The dual system is a system of simple roots of ${\bf g_2}$ \beq{srg2} \ti\Pi=\{\ti\al_3=\f1{3}(e_2-e_3)\,,~\ti\al_2=\f1{3}(e_1-e_2+2e_3)\}\,. \eq In this basis the positive ${\bf g_2}$ roots are \beq{rg2} \ti R^+({\bf g_2})=(\ti\Pi\,,~\f1{3}(2,1,1,0)\,,~\f1{3}(1,2,-1,0)\,,~\f1{3}(1,1,0,0)\,,~ \f1{3}(1,0,1,0))\,. \eq It is convenient to pass from the coroot basis $\ti\Pi^\vee$ in $\gH(\bag2)$ to the canonical basis $(e_1,e_2,e_3)$ \beq{cbg2} \ti\bfu=u_1e_1+u_2e_2+u_3e_3\,, ~~u_1=u_2+u_3\,,~~(u_4=0)\,. \eq According with (\ref{g2c}) the GS basis in $\gH(\bas8)$ is generated by (\ref{cbg2}) and by \beq{gh0} \left\{ \begin{array}{c} \gh^1_{\al_2}=\{\f1{\sqrt{3}}(\al_2+\om\al_4+\om^2\al_6)\} =\{\f1{\sqrt{3}}(\om,-\om,1+\om^2,-1+\om^{2})\}\,, \\ \gh^2_{\al_2}=\{\f1{\sqrt{3}}(\al_2+\om^2\al_4+\om\al_6)\}= \{\f1{\sqrt{3}}(\om^2,-\om^2,1+\om,-1+\om)\}\,. \end{array} \right. \eq With respect to the $\la_1$ $\bas8$ is decomposed as \beq{so8ex1} {\bf so(8)}={\bf g}_2\oplus\underline{7}\oplus\underline{7}'\,,~~ \la_1({\bf g}_2)={\bf g}_2\,,~\la_1(\underline{7})=\om\cdot\underline{7}\,, ~\la_1(\underline{7}')=\om^2\cdot\underline{7}'\,, \eq where $\underline{7}$, $\,\underline{7}'$ are fundamental representations of ${\bf g}_2$. According with the gradation we have the following gradation of the root spaces \vspace{2mm} $$ \begin{tabular}{|l|l|l|} \hline ${\bf g_2}$ & $\underline{7}$ & $\underline{7}'$ \\ \hline $E_{12}^0=E_{12}+E_{34}+E_{35}$ & $\gt_{12}^1=\f1{\sqrt{3}}(E_{12} +\om E_{34}+\om^2E_{35}) $&$\gt_{12}^2=\f1{\sqrt{3}}(E_{12} +\om^2 E_{34}+\om E_{35}) $\\ $E_{13}^0=E_{13}+E_{24}+E_ {25} $&$ \gt_{13}^1=\f1{\sqrt{3}}(E_{13}+\om E_{24}+\om^2E_ {25}) $&$ \gt_{13}^2=\f1{\sqrt{3}}(E_{13}+\om^2 E_{24}+\om E_ {25}) $ \\ $E_{14}^0=E_{14}+E_{15}+E_ {26} $&$ \gt_{14}^1=\f1{\sqrt{3}}(E_{14}+\om E_{15}+\om^2E_ {26} ) $&$ \gt_{14}^2=\f1{\sqrt{3}}(E_{14}+\om^2 E_{15}+\om E_ {26}) $\\ $E_{23}^0=E_{23}\,,~E_{16}^0=E_{16}\,,$ & $\gt^1_{(2,3,4),1}$\,, $~\gh^1_{\al_2}$ & $\gt^2_{(2,3,4),1}$\,, $~\gh^2_{\al_2}$ \\ $E^0_{17}=E_{17}$ & &\\ \hline \end{tabular} $$ \begin{center} \textbf{Table1.} GS basis in ${\bf so(8)}$. \end{center} \vspace{2mm} The roots of ${\bf so(8)}$ that parameterized the GS basis are $$ (e_1-e_m)\to \gt^a_{1,m}\,,~~(e_m-e_1)\to \gt^a_{m,1}\,. $$ The left column of table 5 contains seven positive root subspaces of ${\bf g_2}$. In particular, $E_{23}^0=\ti E_{\al_3}$ and $E_{12}^0=\ti E_{\al_2}$. The Killing form on ${\bf g}_2$ is the canonical form on $\gH({\bf g}_2)$, and for the root subspaces non-vanishing elements are \beq{kfg2r} (E_{12}^0,E_{21}^0)=(E_{13}^0,E_{31}^0)= (E_{14}^0,E_{41}^0)=3\,, \eq $$ (E_{23}^0,E_{32}^0)=(E_{16}^0,E_{61}^0)=(E^0_{17},E^0_{71})=1\,. $$ For rest generators of the GS basis in $\bas8$ we have \beq{kfg8} (\gh^1_{\al_2},\gh^2_{\al_2})=2\,,~~ (\gt^{k_1}_{jk},\gt^{k_2}_{lm})=\de^{(k_1+k_2,0)}\de_{j,m}\de_{k,l}\,. \eq \subsubsection*{GS basis in $\clJ$.} The basis in the Cartan part of $\clJ$ $$ \gH(\clJ)=\{u_5e_5+u_6e_6+u_7e_7\,,~~u_5+u_6+u_7=0\}\,. $$ is transformed as $\la_1(e_5)=e_{7}\,, ~\la_1(e_6)=e_{5}\,, ~\la_1(e_7)=e_{6}$. The basis $ E_{a,\pm}^A$ (\ref{sroe6}) is transformed as $$ \la_1\,:~~E_{a,+}^L\to E_{a,-}^V\to E_{a,-}^R\,,~~~ E_{a,-}^L\to E_{a,+}^V\to E_{a,+}^R\,. $$ There are 16 orbits of this type in $\clJ$. Thus, the GS basis in $\clJ$ takes the form \beq{gsJ} \gt^k_{a,+}=\f1{\sqrt{3}}( E_{a,+}^L+\om^k E_{a,-}^V+\om^{2k} E_{a,-}^R)\,, \eq $$ \gt^k_{a,-}=\f1{\sqrt{3}}(E_{a,-}^L+\om^k E_{a,+}^V+\om^{2k} E_{a,+}^R)\,, ~~(k=0,1,2)\,, $$ $$ \gh^k_5=\f1{\sqrt{3}}(e_5+\om^k e_6+\om^{2k} e_7)\,,~~(k=1,2)\,. $$ Here $\gt^0_{a,5,7}$, $\gt^0_{a,7,5}$ generate the GS basis in $V$ (5.29.I). The Killing form is \beq{kfg9} (\gt^{k_1}_{a,+},\gt^{k_2}_{b,-})=\de_{a,-b}\de^{(k_1+k_2,0)}\,,~~~ (\gh^{k_1}_5,\gh_5^{k_2})=\de^{(k_1+k_2,0)}\,. \eq \bigskip In summary, under the $\la_1$ action ${\bf e_6}$ is decomposed as \beq{mde} {\bf e_6}=\gg_0+\gg_1+\gg_2\,, \eq where $$ \gg_0={\bf g_2}+V\,, ~~V=\{\gt^0_{a,+}\,,~\gt^0_{a,-}\}\,,~~\dim\,V=16\,. $$ The $\la_1$-invariant subalgebra $\gg_0$ is isomorphic to $\bas8+2\times\underline{1}$. It is obtained from the ${\bf e_6}$ extended Dynkin diagram (Fig.7) by dropping out three vertices $\al_0$, $\al_1$ and $\al_5$ . The subspaces $\gg_1$ and $\gg_2$ have the GS basis of the form \beq{GSe6} \gg_1=\left\{\gt^1_{a,+}\,,~\gt^1_{a,-}\,,~\gh^1_{5}~ ~\gh^1_{\al_2}\,,~\gt^1_{1,(2,3,4)}\,,~\gt^1_{(2,3,4),1} \right\}\,, \eq $$ \gg_2=\{\gt^2_{a,+}\,,~\gt^2_{a,-}\,,~\gh^2_{5}\,,~ \gh^2_{\al_2}\,,~\gt^2_{1,(2,3,4)}\,,~\gt^2_{(2,3,4),1}\}\,. $$ In correspondence with (\ref{mde}) we have $$ \underline{78}=\underline{30}+\underline{24}+\underline{24}\,. $$ The form of the dual basis (5.9.I), (5.15.I) follows from (\ref{kfg8}), (\ref{kfg9}) \beq{dbe6} \gH^k_{\al_2}=\oh\gh_{\al_2}^{3-k}\,, ~~\gT^k_{jm}=\gt^{3-k}_{mj}\,,~~(k=1,2)\,, \eq $$ \gT^k_{a,+}=\gt^{3-k}_{-a,-}\,,~~\gT^k_{a,-}=\gt^{3-k}_{-a,+}\,,~~(k=1,2)\,, $$ $$ \gH_5^k=\gh^{3-k}_5\,,~~(k=1,2)\,. $$ \subsection*{Lax operators and Hamiltonians} \subsubsection*{Trivial bundles} The moduli space of trivial $\bG=E_6$ bundles is the quotient $$ \gH_{\bf e_6}/(W_{\bf e_6}\ltimes(\tau Q({\bf e_6})+Q({\bf e_6}))\,, $$ where $W_{\bf e_6}$ is defined in (\ref{wee6}) It means in particular that $\bfu\in\gH_{\bf e_6}$ is defined up to the shifts \beq{e6sc} \bfu\sim\bfu+\tau\ga_1+\ga_2\,,~~\ga_{1,2}\in Q(\bf e_6)\,. \eq where $Q({\bf e_6})$ is (\ref{rle6}). The moduli space of trivial $G^{ad}=E_6/\mu_3$ bundles is the quotient $$ \gH_{\bf e_6}/(W_{\bf e_6}\ltimes(\tau P({\bf e_6})+P(\bf e_6)))\,. $$ It means that \beq{e6ad} \bfu\sim\bfu+\tau\ga_1+\ga_2\,,~~\ga_{1,2}\in P(\bf e_6)\,, \eq and $P(\bf e_6)$ is (\ref{cent}). For trivial bundles the Lax operator takes the form \beq{lmtre} L^{CM}_{e_6}(z)=L^{CM}_{so(8)}+\sum_{j=5}^7(v_j+S_{0,j}E_1(z))e_j + \eq $$ \sum_{a=1}^8\Bigl( S^L_{a,+}\phi((\bfu,\varpi_a^L)+\f1{\sqrt{2}}(u_5-u_7),z)E^L_{a,+}+ S^L_{a,-}\phi((\bfu,\varpi_a^L)+\f1{\sqrt{2}}(u_7-u_5),z)E_{a,-}^L+\Bigr. $$ $$ S^R_{a,+}\phi((\bfu,\varpi_a^R)+\f1{\sqrt{2}}(u_5-u_6),z)E_{a,+}^R+ S^R_{a,-}\phi((\bfu,\varpi_a^R)+\f1{\sqrt{2}}(u_6-u_5),z)E_{a,-}^R+ $$ $$ \Bigl. S^V_{a,+}\phi((\bfu,\varpi_a^V)+\f1{\sqrt{2}}(u_6-u_7),z)E_{a,+}^V+ S^V_{a,-}\phi((\bfu,\varpi_a^V)+\f1{\sqrt{2}}(u_7-u_6),z)E_{a,-}^V \Bigr)\,. $$ \bigskip After symplectic reduction with respect to the action of the Cartan subgroup we come to integrable $E_6$ hierarchy with two types of moduli space (\ref{e6sc}), (\ref{e6ad}). The quadratic Hamiltonian is \beq{he6} H^{CM}_{e_6}=H^{CM}_{so(8)}+\oh\sum_{j=5}^7v_j^2+ \sum_{a=1}^8\Bigl(S^L_{a,+}S^L_{-a,-}E_2((\bfu,\varpi_a^L)+\f1{\sqrt{2}}(u_5-u_6)) \Bigr. \eq $$ \Bigl. S^R_{a,+}S^R_{-a,-}E_2((\bfu,\varpi_a^R)+\f1{\sqrt{2}}(u_5-u_6))+ S^V_{a,+}S^V_{-a,-}E_2((\bfu,\varpi_a^V)+\f1{\sqrt{2}}(u_6-u_7)) \Bigr)\,. $$ \subsubsection*{Nontrivial bundles} \textbf{The moduli space.} Now $\bfu=\ti\bfu\in\gH(\bag2)$ \beq{hg2} \ti\bfu=u_1e_1+u_2e_2+u_3e_3\,,~~u_1=u_2+u_3\,. \eq More exactly, $\ti\bfu$ belongs to fundamental domains with respect to action of affine Weyl group corresponding to ${\bag2}$. The Weyl group $W_{\bag2}$ is isomorphic to the dihedral group of order 12 with two generators $$ (u_1,u_2,u_3)\to (u_1,u_3,u_2)\,, $$ $$ \left( \begin{array}{c} u_1\\ u_2 \\ u_3 \\ \end{array} \right) \to \f1{3} \left( \begin{array}{rrr} 2 & 1 & -2 \\ 1 & 2 & 2 \\ -2 & 2 & -1\\ \end{array} \right) \left( \begin{array}{c} u_1\\ u_2 \\ u_3 \\ \end{array} \right)\,. $$ The invariant root sublattice $\ti Q^\vee=Q^\vee(\bag2)$ coincides with $\ti P^\vee=P^\vee(\bag2)$. From (\ref{srg2a}) we find $$ \ti Q^\vee(\bag2)=\{m_2e_1+(3m_3-m_2)e_2+(2m_2-3m_3)e_3\,|\,m_2,m_3\in\mZ\}\,. $$ Then the moduli space of nontrivial $E_6$ bundle is defined as \beq{mne6} \gH(\bag2)/(W_{\bag2}\ltimes(\tau \ti Q^\vee(\bag2)+\ti Q^\vee(\bag2)))\,. \eq In other words $$ u_1\sim u_1+\tau m_2+m_2'\,,~~u_2\sim u_2+\tau(3m_3-m_2)+3m'_3-m'_2 \,, $$ $$ u_3\sim u_3+\tau(2m_2-3m_3)+2m_2'-3m_3'\,,~~m_{2,3}\in\mZ\,. $$ \bigskip \textbf{The Lax operators} According with (\ref{mde}) $L(z)=\ti L_{0}(z)+L_0'(z)+L_1(z)+L_2(z)$. Here $\ti L_0(z)=L^{CM}_{\bag2}(z)$. Let $\ti\bfv=(v_1,v_2,v_3)$ $\,(v_1=v_2+v_3)$ be the Poisson dual vector to $\ti\bfu$ (\ref{hg2}). Following to (\ref{rg2}) and table 5 we find $$ L^{CM}_{\bag2}(z)=\sum_{j=1}^3(v_j+S_{0,j}E_1(z))e_j+ S^0_{12}\phi(\f1{3}(u_1-u_2+2u_3),z)E_{12}^0 + S^0_{21}\phi(\f1{3}(-u_1+u_2-2u_3),z)E_{21}^0 $$ $$ + S^0_{23}\phi(\f1{3}(u_2-u_3),z)E_{23}^0+S^0_{32}\phi(\f1{3}(u_3-u_2),z)E_{32}^0+ S^0_{16}\phi(\f1{3}(u_1+u_3),z)E_{16}^0+S^0_{61}\phi(\f1{3}(-u_1-u_3),z)E_{61}^0 $$ $$ +S^0_{17}\phi(\f1{3}(u_1+u_2),z)E_{17}^0+S^0_{71}\phi(\f1{3}(-u_1-u_2),z)E_{71}^0+ S^0_{14}\phi(\f1{3}(2u_1+u_2+u_3),z)E_{14}^0 $$ $$ +S^0_{41}\phi(\f1{3}(-2u_1-u_2-u_3),z)E_{41}^0+ S^0_{13}\phi(\f1{3}(u_1+2u_2-u_3),z)E_{13}^0+ S^0_{31}\phi(\f1{3}(-u_1-2u_2+u_3),z)E_{31}^0\,. $$ Define $\varphi^k_{\be}$ (6.14.I) $$ \varphi^k_{\be}(\ti\bfu,z)=\bfe\,\Bigl(\lan\ka_{\bf e_6},\be\ran z\Bigr) \phi(\lan\ti\bfu+\ka_{\bf e_6}\tau,\be\ran+k/3, z)\,,~~~(k=0,1,2)\,, $$ where $\be=\al_{jk}$ and $\lan\ka_{\bf e_6},\be\ran$ (\ref{rso8}), or $\be=\al_{(\pm a,\pm)}^L$ and $\lan\ka_{\bf e_6},\be\ran$ (\ref{sroe6}). Then following (6.15.I) and (6.17.I) we find $$ L'_0(z)=\sum_{a=1}^8\Bigl(S^0_{a,-}\varphi^0_{\al_{(-a,-)}^L}(-\ti\bfu,z) \gt^0_{a,+}+ S^0_{a,+}\varphi^0_{\al_{(a,+)}^L}(-\ti\bfu,z)\gt^0_{a,-}\Bigr)\,, $$ and for $k=1,2$ $$ L_k(z)=S_5^{3-k}\phi(k/3, z)\gh_5^{k}+S_{\al_2}^{3-k}\phi(k/3, z)\gh_{\al_2}^{k}+ \sum_{m=2}^4(S^{3-k}_{m,1}\varphi^k_{\al_{1,m}}(-\ti\bfu,z)\gt^k_{1,m}+ S^{3-k}_{1,m}\varphi^k_{\al_{m,1}}(-\ti\bfu,z)\gt^k_{m,1})+ $$ $$ \sum_{a=1}^8(S^{3-k}_{-a,-}\varphi^k_{\al_{(a,+)}^L}(-\ti\bfu,z)\gt^{k}_{a,+}+ S^{3-k}_{-a,+}\varphi^k_{\al_{(a,-)}^L}(-\ti\bfu,z)\gt^{k}_{a,-})\,. $$ After the symplectic reduction we come to the integrable system with quadratic Hamiltonian $H_{\bf e_6}=H^{CM}_{\bag2}+H_0'+H_1$, where $H_0'$ is defined by $\oh\tr(L_0^{'2})$, $H_1$ by $\tr(L_1L_2)$, and $$ H^{CM}_{\bag2}=\oh\sum_{j=1}^3v_j^2-3S^0_{12}S^0_{21}E_2(\f1{3}(u_1-u_2+2u_3))- S^0_{23}S^0_{32}E_2(\f1{3}(u_2-u_3)) $$ $$ - S^0_{16}S^0_{61}E_2(\f1{3}(u_1+u_3))- S^0_{17}S^0_{71}E_2(\f1{3}(u_1+u_2))-3S^0_{14}S^0_{41}E_2(\f1{3}(2u_1+u_2+u_3)) $$ $$ - 3S^0_{13}S^0_{31}E_2(\f1{3}(u_1+2u_2-u_3))\,, $$ $$ H_{0}'=- \sum_{a=1}^8S^L_{a,+}S^L_{-a,-}E_2(\varpi^L_a,\ti\bfu)\,, $$ $$ -H_1=(S_5^1S^2_5+2S^1_{\al_2}S^2_{\al_2})E_2(1/3)+ \sum_{m=2}^4(S^1_{m,1}S^2_{1,m}+ S^1_{1,m}S^2_{m,1}) E_2(u_1-u_m)+ $$ $$ \sum_{a=1}^8(S^1_{a,-}S^2_{-a,+}+S^1_{a,+}S^2_{-a,-})E_2(\varpi_a^L,\ti\bfu)\,. $$ \section{E$_7$\,.} \setcounter{equation}{0} \subsection*{Roots and weights} The Cartan subalgebra can be identified with the space $\gH_{\bf e_7}=\{\bfu\in\mC^7\}$. The root system $R({\bf e_7})$ is related to the root system $R({\bf e_6})$ (\ref{are6r}) as follows \beq{are7} R({\bf e_7})=R({\bf e_6})\cup \pm (\sqrt{2}e_{i+4}\,,~i=1,2,3) \eq $$ \cup ( P^R\pm\f1{\sqrt{2}}(e_5+e_6))\cup ( P^V\pm\f1{\sqrt{2}}(e_6+e_7)) \cup ( P^L\pm\f1{\sqrt{2}}(e_5+e_7))= $$ $$ R({\bf e_6})\cup\{\al^+_{i+4}\,,\al^{V,+}_{a,\pm}\,,\al^{R,+}_{a,\pm}\,, \al^{L,+}_{a,\pm}\}\,,~(a=1,\ldots,8\,,\,i=1,2,3) $$ (compare with (\ref{are6r})). Then $\sharp(R({\bf e_7}))=72+6+3\times 16=126$ and $\dim\,{\bf e_7}={\rm rank}\,{\bf e_7}+\sharp(R({\bf e_7}))=133$. The basis of simple roots can be chosen as \beq{sr7} \Pi=\left\{ \begin{array}{l} \al_1=\oh(e_4-e_3-e_2-e_1)+\f1{\sqrt{2}}(e_5-e_6)\,, \\ \al_2=e_3-e_4 \,,\\ \al_3=e_2-e_3\,, \\ \al_4=e_1-e_2\,, \\ \al_5=-e_1+ \f1{\sqrt{2}}(e_6-e_7)\,,\\ \al_6=e_4+e_3\,,\\ \al_7=\sqrt{2}e_7\,. \end{array} \right. \eq The subsystem of simple roots \beq{p17} \Pi_1=\{\al_i\,,~~i=1,\ldots 6\}=\Pi({\bf e_6}) \eq is a system of simple roots of subalgebra ${\bf e_6}$. It follows from (\ref{are7}) that the positive roots, corresponding to $\Pi$ is $$ R^+({\bf e_7})=R^+({\bf e_6})\cup (\sqrt{2}e_{i+4}\,,~i=1,2,3) $$ $$ \cup ( P^R+\f1{\sqrt{2}}(e_5+e_6))\cup ( P^V+\f1{\sqrt{2}}(e_6+e_7)) \cup ( P^L+\f1{\sqrt{2}}(e_5+e_7)) $$ Then the root lattice takes the form \beq{rle7} Q({\bf e_7})=Q({\bf e_6})+\mZ\sqrt{2}e_{7}+\mZ(-e_1+\f1{\sqrt{2}}(e_6+e_7)) +\mZ(\oh(e_1-e_2-e_3-e_4)+\f1{\sqrt{2}}(e_5+e_6))\,. \eq The simple roots (\ref{sr7}) define the fundamental Weyl chamber $C=\{\bfu\in\gH\,|\,\lan\bfu,\al\ran>0\,,~\al\in \Pi({\bf e_7})\}$. The minimal root is \beq{mre7} \al_0=-\sqrt{2}e_5=-(2\al_1+3\al_2+4\al_3+3\al_4+2\al_5+2\al_6+\al_7)\,. \eq The root system $R_{\bf e_7}$ is self-dual and the corresponding Dynkin diagram is simply-laced. The half-sum of the positive roots $\rho_{\bf e_7}=3e_1+2e_2+e_3+\f1{\sqrt{2}}(17e_5+9e_6+e_7)$ can be expressed in terms of roots of subalgebras ${\bf e_6}$ or $\bf f_4$ (see below) $$ \rho_{\bf e_7}=\rho_{\bae}+\frac{9}{\sqrt{2}}(e_5+e_6+e_7)= \rho_{\bf f_4}+\f1{4}(3e_1+e_2+e_3-e_4)+\frac{5}{2\sqrt{2}}(e_5-e_7)\,, $$ where $\rho_{\bf f_4}=\f1{4}(9,7,3,1,\frac{11}{2\sqrt{2}},0,-\frac{11}{2\sqrt{2}})$. The Coxeter number is equal to $h=18$. Then \beq{ka7} \ka_{\bf e_7}=\f1{18}\ka_{f_4}+\f1{72}(3e_1+e_2+e_3-e_4)+ \frac{5}{36\sqrt{2}}(e_5-e_7)\,,~~~(\ka_{\bf f_4}=\f1{18}\rho_{\bf f_4})\,. \eq We define the fundamental weights dual to $\Pi$ (\ref{sr7}) \beq{fwe7} \begin{array}{l} \varpi_1 =\sr2 e_5= 2\al_1+3\al_2+4\al_3+3\al_4+2\al_5+2\al_6+\al_7\,,\\ \varpi_2=\oh(e_1+e_2+e_3-e_4)+\f1{3\sqrt 2}(5e_5-e_6-4e_7)\\ =\oh(4\al_1+8\al_2+12\al_3+9\al_4+6\al_5+7\al_6+3\al_7)\,, \\ \varpi_3=e_1+e_2+\sqrt 2(e_5-e_7)= 3\al_1+6\al_2+8\al_3+6\al_4+4\al_5+4\al_6+2\al_7 \,,\\ \varpi_4=e_1+\f1{3\sqrt 2}(4e_5+e_6-5e_7)= 4\al_1+8\al_2+12\al_3+9\al_4+6\al_5+6\al_6+3\al_7\,, \\ \varpi_5=\sr2(e_5+e_6)= \f1{2}(6\al_1+12\al_2+18\al_3+15\al_4+10\al_5+9\al_6+5\al_7)\,, \\ \varpi_6=\oh(e_1+e_2+e_3+e_4)+\f1{\sqrt 2}(e_5-e_7)\\ =2\al_1+4\al_2+6\al_3+5\al_4+4\al_5+3\al_6+2\al_7 \,,\\ \varpi_7=\f1{\sr2}(e_5+e_6+e_7)=\ \oh(2\al_1+4\al_2+6\al_3+5\al_4+4\al_5+3\al_6+3\al_7)\,. \end{array} \eq It follows from (\ref{fwe7}) that \beq{wle7} P(\ble)=Q(\ble)+\mZ\varpi_7\,. \eq and the factor-group $P(\ble)/Q(\ble)$ is isomorphic $\mu_2$. The minimal root (\ref{mre7}) defines the vertices of the fundamental alcove\\ $C_{alc}=(0,\oh\varpi_1,\f1{3}\varpi_2,\f1{4}\varpi_3,\f1{3}\varpi_4, \oh\varpi_5,\oh\varpi_6,\varpi_7)$. Then $\varpi_7$ generates $\la_7$ \beq{lae7} \la_7=\left\{ \begin{array}{l} \al_1\to \al_5 \,,\\ \al_2\to \al_4\,, \\ \al_3\to \al_3\,,\\ \al_4\to \al_2\,, \\ \al_5\to \al_1\,, \\ \al_6\to \al_6\,, \\ \al_7\to \al_0\,, \\ \al_0\to \al_7\,, \end{array} ~~~~ \begin{array}{l} e_1\to \oh(e_1+e_2+e_3-e_4)\,,\\ e_2\to \oh(e_1+e_2-e_3+e_4)\,,\\ e_3\to \oh(e_1-e_2+e_3+e_4)\,,\\ e_4\to \oh(-e_1+e_2+e_3+e_4)\,,\\ e_5\to -e_7\,,\\ e_6 \to -e_6\,, \\ e_7 \to -e_5\,. \end{array} \right. \eq \unitlength 1mm \linethickness{0.4pt} \ifx\plotpoint\undefined\newsavebox{\plotpoint}\fi \begin{picture}(84.036,63.251)(0,0) \put(40.731,38.204){\circle*{2.081}} \put(47.866,38.204){\circle{1.88}} \put(54.852,38.204){\circle{1.88}} \put(61.988,38.352){\circle{1.904}} \put(68.826,38.204){\circle{1.808}} \put(76.109,38.054){\circle{1.904}} \put(83.095,37.906){\circle{1.88}} \put(61.987,31.068){\circle{2.081}} \multiput(41.771,37.906)(1.040552,.02973){5}{\line(1,0){1.040552}} \multiput(48.906,37.906)(1.010822,.02973){5}{\line(1,0){1.010822}} \multiput(56.041,37.906)(.981092,.02973){5}{\line(1,0){.981092}} \multiput(63.028,38.054)(.981092,-.02973){5}{\line(1,0){.981092}} \put(70.014,37.757){\line(1,0){4.905}} \put(77.298,37.757){\line(1,0){4.608}} \put(61.987,37.163){\line(0,-1){4.608}} \qbezier(40.73,43.852)(57.453,63.251)(81.609,44.892) \qbezier(74.92,44.744)(59.534,55.075)(47.419,44.298) \qbezier(54.852,44.298)(60.129,49.575)(67.487,43.852) \thicklines \qbezier(40.73,44)(42.142,46.676)(41.473,45.784) \qbezier(41.473,45.784)(40.582,42.886)(40.879,43.852) \qbezier(40.73,44.149)(42.96,45.041)(42.217,44.744) \multiput(46.973,44.149)(.03344631,.04459508){40}{\line(0,1){.04459508}} \multiput(47.271,44)(.0550557,.0330334){27}{\line(1,0){.0550557}} \multiput(54.555,44)(.0335662,.0527469){31}{\line(0,1){.0527469}} \multiput(54.852,44)(.0431565,.0335662){31}{\line(1,0){.0431565}} \multiput(65.703,44.446)(.1167966,-.0318536){14}{\line(1,0){.1167966}} \multiput(67.19,44.149)(-.0330334,.0440445){27}{\line(0,1){.0440445}} \multiput(72.987,45.338)(.0991002,-.0330334){18}{\line(1,0){.0991002}} \multiput(74.771,44.744)(-.0335662,.0431565){31}{\line(0,1){.0431565}} \multiput(79.528,45.933)(.057542,-.0335662){31}{\line(1,0){.057542}} \multiput(81.46,44.892)(-.03303339,.04542091){36}{\line(0,1){.04542091}} \put(38.352,34.338){\makebox(0,0)[cc]{$\alpha_0$}} \put(45.636,34.784){\makebox(0,0)[cc]{$\alpha_1$}} \put(53.217,35.23){\makebox(0,0)[cc]{$\alpha_2$}} \put(59.906,41.622){\makebox(0,0)[cc]{$\alpha_3$}} \put(60.798,27.352){\makebox(0,0)[cc]{$\alpha_6$}} \put(66.595,34.784){\makebox(0,0)[cc]{$\alpha_4$}} \put(74.474,34.933){\makebox(0,0)[cc]{$\alpha_5$}} \put(81.163,35.379){\makebox(0,0)[cc]{$\alpha_7$}} \put(60.947,21.406){\makebox(0,0)[cc]{Fig.8 $E_7$, $\lambda_7$}} \end{picture} \subsection*{Chevalley basis} The Lie algebra $\ble$ can be defined in terms of $\bae$ and its representations \beq{e7e6} \ble=\bae\oplus \clI\,,~~\clI=\underline{27}\oplus\underline{\overline{27}}\oplus\underline{1}\,. \eq Here $\underline{27}$ and $\underline{\overline{27}}$ are two fundamental representations of $\bae$ corresponding to the weights $\varpi_1$ and $\varpi_5$ (\ref{fwe6}). The scalar $\underline{1}$ corresponds to the generator $(e_5+e_6+e_7)$ in the Cartan subalgebra $\gH_{\ble}$. It allows us to use in $\gH(\ble)$ the unrestricted canonical basis $e_j$, $j=1,\ldots,7$. It follows from (\ref{are7}) that the rest generators correspond to new 54 root subspaces. Similarly to (\ref{sroe6}) they are \beq{sroe7} \begin{array}{ll} \al_{(a,\pm)}^{(L,+)}= (\varpi_a^L\pm\f1{\sqrt{2}}(e_5+e_7))\to E_{a,\pm}^{L,+} & a=1,\ldots,8\,, \\ \al_{(a,\pm)}^{(R,+)}= (\varpi_a^R\pm\f1{\sqrt{2}}(e_5+e_6))\to E_{a,\pm}^{R,+} \,, & a=1,\ldots,8\,, \\ \al_{(a,\pm)}^{(V,+)}= (\varpi_a^V\pm\f1{\sqrt{2}}(e_6+e_7))\to E_{a,\pm}^{R,+} \,, & a=1,\ldots,8\,, \\ \pm \al^{(+)}_j= \pm\sqrt{2}e_j\to E_{j,\pm}\,, & j=5,6,7\,. \end{array} \eq The positive root subspaces are $E_{a,+}^{A,+}$ and $E_{j,+}$. The new generators are orthogonal to the $\bae$ generators and (see (A.25.I)) \beq{kfc7} (E_{a,+}^{A,+},E_{b,-}^{B,+})=\de_{a,-b}\de^{A,B}\,, ~~A,B=L,R,V\,, \eq $$ (E_{j,\pm},E_{k,\mp})=\de_{jk}\,. $$ \subsection*{GS basis} Since $\la_7^2=1$, \beq{gse7} \ble=\gg_0\oplus\gg_1\,,~~\gg_0=\ti\gg_0+V\,. \eq It follows from (\ref{e7e6}) and Fig. 8 that $\gg_0\sim\bae+\underline{1}$ We will prove that $\ti\gg_0=\baf$. The root subsystem $\Pi_1$ that does not contain an orbit of $\la$ passing through $\al_0$ is $\Pi_{\bae}$ (\ref{p17}). It follows from (\ref{lae7}) that the $\la_7$-action on $\Pi_{\bae}$ takes the form $$ \ti\la_7=\left\{ \begin{array}{l} \al_1\to \al_5 \,,\\ \al_2\to \al_4\,, \\ \al_3\to \al_3\,,\\ \al_4\to \al_2\,, \\ \al_5\to \al_1\,, \\ \al_6\to \al_6\,, \\ \end{array} \right. $$ The set of orbits in $\Pi_1^\vee=\Pi_1$ is $$ \ti\Pi^\vee=\Pi^\vee_1/\mu_2=\Bigl(\ti\al_1^\vee=\al_6\,,~\ti\al_2^\vee=\al_3\,,~ \ti\al_3^\vee=\al_2+ \al_4\,,~\ti\al_4^\vee=\al_1 +\al_5\Bigr)= $$ \beq{cbe7} =\Bigl(e_4+e_3\,,e_2-e_3\,,e_1-e_2+e_3-e_4\,,\oh(e_4-e_3-e_2-3e_1)+\f1{\sqrt{2}}(e_5-e_7) \Bigr)\,. \eq It is the coroot basis in the invariant subalgebra $\ti\gH_0$. The dual root system \beq{srf4} \ti\al_1=e_3+e_4\,,~\ti\al_2=e_2-e_3\,,~\ti\al_3=\oh(e_1-e_2+e_3-e_4)\,, \eq $$ \ti\al_4=\frac{1}4(-3e_1-e_2-e_3+e_4)+\f1{2\sqrt{2}}(e_5-e_7)\,. $$ defines simple roots of type ${\bf f_4}$. Then $\ti\gH_0=\gH(\baf)$. \subsubsection*{GS basis in $\bae$ subalgebra} Under the $\ti\la_7$-action $\bae$ is decomposed as \beq{e6f4} \bae=\baf\oplus\underline{26}\,, \eq where $\underline{26}$ is a fundamental representation of $\baf$. In terms of (\ref{gse7}) $\ti\gg_0=\baf$ and $\underline{26}\subset\gg_1$. Let us describe (\ref{e6f4}) in terms of the familiar decomposition (\ref{e6d}) \beq{povt} \bae={\bf so(8)}\oplus\clJ=\underline{28}\oplus\underline{50} \eq taking into account the $\ti\la_7$-action. \bigskip \noindent \textbf{GS basis in ${\bf so(8)}$-component.}\\ The ${\bf so(8)}$ subalgebra is generated by the simple roots $(\al_2,\al_3,\al_4,\al_6)$. The $\ti\la_7$ invariant subsystem $(\ti\al_1,\ti\al_2,\ti\al_3)$ (\ref{srf4}) forms $B_3$ root subsystem. It implies that \beq{so8ex} {\bf so(8)}={\bf so(7)}\oplus\underline{7}\,,~~\underline{28}=\underline{21}\oplus\underline{7}\,, \eq $$ \ti\la_7({\bf so(7)})={\bf so(7)}\,,~\ti\la_7(\underline{7})=-\underline{7}\,, $$ where $\underline{7}$, is the fundamental representations of ${\bf so(7)}$. The 24 root subspaces of ${\bf so(8)}$ contains 18 root subspaces of ${\bf so(7)}$. Then according with the definition of ${\bf so(8)}$ root subspaces (\ref{rso8}) we define the GS-basis in ${\bf so(8)}$. \vspace{3mm} \begin{center} \begin{tabular}{|l|l|} \hline ${\bf so(7)}$ & $\underline{7}$ \\ \hline $E_{12}^0=E_{12}+E_{34}$ & $\gt_{12}^1=\f1{\sqrt{2}}(E_{12} - E_{34}) $\\ $E_{13}^0=E_{13}+E_{24}$&$ \gt_{13}^1=\f1{\sqrt{2}}(E_{13}- E_{24})$ \\ $E_{15}^0=E_{15}+E_{26}$&$ \gt_{15}^1=\f1{\sqrt{2}}(E_{15}- E_{26} ) $\\ $E_{23}^0=E_{23}\,,~E_{14}^0=E_{14}\,,$ & $\gt_{j,1}^1\,,~j=2,3,5$ \\ $E^0_{35}=E_{35}\,,~E_{25}^0=E_{25}$ & $\gh^1_{\al_2}$ (\ref{ghf4}) \\ $E^0_{16}=E_{16}\,,~E_{17}^0=E_{17}$ & \\ \hline \end{tabular} \\ \vspace{5mm} \textbf{Table 2.} GS basis in ${\bf so(8)}$. \end{center} The left column represents 9 positive root subspaces of ${\bf so(7)}$, and the right column represents the root subspaces and a Cartan element in $\underline{7}$. The ${\bf so(8)}$ roots that parameterized the GS-basis are \beq{re7gs} \begin{array}{ll} \al_{12}=(e_1-e_2)\to \gt_{12}^1\,, & -\al_{12}=(e_2-e_1)\to \gt_{21}^1\,, \\ \al_{13}= (e_1-e_3)\to \gt_{13}^1\,, & -\al_{13}= (e_3-e_1)\to \gt_{31}^1\,, \\ \al_{15}=(e_1+e_4)\to \gt_{15}^1\,, & -\al_{15}=(-e_1-e_4)\to \gt_{51}^1\,, \\ \end{array} \eq Consider the embedding of $\gH({\bf so(7)})$ in $\gH({\bf so(8)})$. Remember that the basis in $\gH({\bf so(7)})$ is $(\ti\al_k^\vee\,,~k=1,2,3)$. In addition we have the anti-invariant generator \beq{ghf4} \gh^1_{\al_2}=\f1{\sqrt{2}}(\al_2-\al_4)=\f1{\sqrt{2}}(-e_1+e_2+e_3-e_4)\,, \eq $$ \gH({\bf so(8)})=\gH({\bf so(7)})\oplus\mC\gh^1_{\al_2}\,. $$ In this way we have defined the $\ti\la_7$-action on ${\bf so(8)}$ component in (\ref{povt}). In correspondence with (\ref{e6f4}) we have $$ {\bf so(7)}\subset\baf\,,~~ \underline{7}\subset\underline{26}\,. $$ \bigskip \noindent \textbf{GS basis in $\clJ$-component}\\ Now consider the $\la$-action on the space $\clJ$ (\ref{povt}). It is represented by $48$ root subspaces (\ref{are6r}) and two elements from $\gH(\bae)$. Let us define the latter generators. They are \beq{ce6a} \gh^1_{\al_1}=\f1{\sqrt{2}}(\al_1-\al_5)=\f1{\sqrt{2}}\Bigl(\oh(e_1-e_2-e_3+e_4)+ \f1{\sqrt{2}}(e_5-2e_6+e_7)\Bigr) \eq and the already defined invariant generator $\ti\al_4^\vee=2\ti\al_4$ (\ref{srf4}) that completes $\gH({\bf so(7)})$ to $\gH(\baf)$. The $\ti\la_7$-action on the the root subspaces takes the form $$ \ti\la_7~:~\left\{ \begin{array}{c} (P^R\pm\f1{\sqrt{2}}(e_5-e_6))\leftrightarrow (P^V\pm\f1{\sqrt{2}}(e_6-e_7))\,, \\ (P^L\pm\f1{\sqrt{2}}(e_5-e_7))\leftrightarrow(P^L\pm\f1{\sqrt{2}}(e_5-e_7))\,. \end{array} \right. $$ Since $\ti\la_7\,:\,\varpi_4^{L}\to -\varpi_4^{L}=\varpi_8^{L}$, all roots $P^L\pm\f1{\sqrt{2}}(e_5-e_7)$ are fixed under the $\ti\la_7$-action, except $(\varpi_4^{L}=\oh(e_1-e_2-e_3+e_4))\pm\f1{\sqrt{2}}(e_5-e_7)$. Then from the $\ti\la_7$ action on the Chevalley basis in $\bae$ (\ref{sroe6}) we obtain generators of the GS basis \beq{te67} \begin{array}{cc} \gt^{R,k}_{a,\pm}=\f1{\sqrt{2}}(E^R_{a,\pm}+(-1)^kE^V_{a,\pm})\,, & k=0,1\,, \\ \gt^{L,k}_{4,\pm}=\f1{\sqrt{2}}(E^L_{4,\pm}+(-1)^kE^L_{8,\pm})\,, & k=0,1\,, \\ \gt^{L,0}_{a,\pm}=E^L_{a,\pm}\,, & ~a\neq 4,8\,. \end{array} \eq The $\mu_2$-gradation $\clJ=\clJ_0\oplus\clJ_1$ takes the form \beq{clj0} \clJ_0=\{\sqrt{2}\gt^{R,0}_{a,\pm}=E_{\ti\al^R_{a,\pm}}\,,~a=1,\ldots,8\,,~~ \sqrt{2}\gt^{L,0}_{a,\pm}=E_{\ti\al^L_{a,\pm}}\,,~a\neq 4,8\,,~~ \sqrt{2}\gt^{L,0}_{4,\pm}=E_{\ti\al^L_{4,\pm}}\,,~\ti\al_4^\vee\}\,, \eq \beq{clj1} \clJ_1=\{\gt^{R,1}_{a,\pm}\,,~a=1,\ldots,8\,,~~\gt^{L,1}_{4,\pm}\,,~\gh^1_{\al_1}\}\,, \eq $$ \dim\,\clJ_0=31\,,~~\dim\,\clJ_1=19\,. $$ Here $E_{\ti\al^R_{a,\pm}}$ are invariant root subspaces constructed from the root subspaces of $\bae$ (\ref{sroe6}). Finally, by comparing (\ref{e6f4}) and (\ref{povt}) we come to the decompositions \beq{f4d} \baf={\bf so(7)}\oplus\clJ_0\,,~~(52=21+31)\,, \eq \beq{26d} \underline{26}=\clJ_1\oplus\underline{7}\,,~~(26=19+7)\,, \eq where $\underline{7}$ is represented by the right column in Table 2. \subsubsection*{Chevalley basis in $\baf$} In what follows we need the Chevalley basis in $\baf$ (\ref{f4d}) in terms of the Chevalley basis in $\bae$. We pass to the canonical basis in $\gH(\baf)$. From (\ref{cbe7}) we find \beq{cf4} \gH(\baf)=\{\ti\bfu=u_1e_1+u_2e_2+u_3e_3+u_4e_4+ u_5e_5-u_5e_7 \}\,, \eq where \beq{fres} u_1-u_2-u_3+u_4=0\,. \eq In other terms it can be written in the form (see (\ref{f4d})) \beq{f4e7} \gH(\baf)=\gH({\bf so(7)})\oplus\mC(\frac{1}2(-3e_1-e_2-e_3+e_4)+\f1{\sqrt{2}}(e_5-e_7))\,. \eq It follows from (\ref{clj0}), (\ref{f4d}) that the $\baf$ roots are \beq{rf4} R(\baf)=R({\bf so(7)})\cup R(\clJ_0)\,, \eq $$ R^+(\clJ_0)=\{\ti\al^R_{a,+}=\f1{2}(\varpi_a^R+\varpi_a^V+\f1{\sqrt{2}}(e_5-e_7))\,,~ \ti\al^L_{a,+}=\oh(\varpi_a^L+\f1{\sqrt{2}}(e_5-e_7))\,,~(a\neq 4,8)\,,~\}\,, $$ $$ \ti\al_4\,,~(\ref{srf4})\,,~~R^+({\bf so(7)})=\Bigl\{\ti\al_j\,,~j=1,2,3,~(\ref{srf4})\,,\Bigr. $$ $$ \left. \oh(e_1+e_2-e_3-e_4)\,,~\oh(e_1+e_2+e_3+e_4)\,,~(e_1-e_4)\,,~(e_2+e_4)\,,~ (e_1+e_3)\,,~(e_1+e_2)\right\}\,. $$ The ${\bf so(7)}$ root subspaces are read of from Table 2, while the root subspaces in $\clJ_0$ are defined in (\ref{clj0}). On $\gH(\baf)$ with the basis $H_{\ti\al_j}=\ti\al_j^\vee$, $j=1,\ldots,4$ (\ref{cbe7}) the Killing form is defined by the $\baf$ Cartan matrix (A.24.I). On $\gL(\baf)$ the Killing form is is $$ (E^0_{23},E^0_{32})=(E^0_{14},E^0_{41})=(E^0_{35},E^0_{53})= (E^0_{25},E^0_{52})=(E^0_{16},E^0_{61})=(E^0_{17},E^0_{71})=1\,, $$ \beq{kff4} (E^0_{12},E^0_{12})=(E^0_{13},E^0_{31})=(E^0_{15},E^0_{51})=2\,, \eq $$ (\gt^{R,0}_{a,+},\gt^{R,0}_{b,-})=2\de_{a,-b}\,,~~ (\gt^{L,0}_{a,+},\gt^{L,0}_{b,-})=2\de_{a,-b}\,. $$ \subsubsection*{GS basis in $\clI$} Consider the $\ti\la_7$-action on the subspace $\clI$ in (\ref{e7e6}). 55 Chevalley generators in $\clI$ (\ref{sroe7}) form the GS basis similarly to (\ref{te67}). Then we come to the GS basis \beq{1te67} \begin{array}{l} \gt^{R,+,k}_{a,+}=\f1{\sqrt{2}}(E^{R,+}_{a,+}+(-1)^kE^{V,+}_{a,-})\,,~~k=0,1\,, \\ \gt^{R,+,k}_{a,-}=\f1{\sqrt{2}}(E^{R,+}_{a,-}+(-1)^kE^{V,+}_{a,+})\,,~~k=0,1\,, \\ \gt^{L,+,k}_{4,+}=\f1{\sqrt{2}}(E^{L,+}_{4,+}+(-1)^kE^{L,+}_{8,-})\,,~~k=0,1\,, \\ \gt^{L,+,k}_{4,-}=\f1{\sqrt{2}}(E^{L,+}_{4,-}+(-1)^kE^{L,+}_{8,+})\,,~~k=0,1\,, \\ \gt^{L,+,k}_{a,+}=\f1{\sqrt{2}}(E^{L,+}_{a,+}+(-1)^kE^{L,+}_{a,-})\,,~~a\neq 4,8\,,~~k=0,1\,, \\ \gt^{k}_{5,\pm}=\f1{\sqrt{2}}(E_{5,\pm}+(-1)^kE_{7,\mp})\,,~~k=0,1\,, \\ \gt^{k}_{6,+}=\f1{\sqrt{2}}(E_{6,+}+(-1)^kE_{6,-})~~k=0,1\,, \\ \gh^1_{e_5}=\f1{\sqrt{3}}(e_5+e_6+e_7)\,. \end{array} \eq The generators with $k=0$ form the GS basis in $V$ and with $k=1$ along with the GS basis in $\underline{26}$ (\ref{26d}) form basis in $\gg_1$. Thus, we come to the following GS basis $$ \ble=\baf\oplus V\oplus\gg_1\,, ~~ (\underline{133}=\underline{52}\oplus\underline{27}\oplus\underline{54})\,, $$ \beq{gsbe7} V=\{\gt^{R,+,0}_{a,\pm}\,,~\gt^{L,+,0}_{a,+}\,,\,(a\neq 4,8)\,,~~ \gt^{L,+,0}_{4,\pm}\,,~\gt^{0}_{5,\pm}\,,~\gt^{0}_{6,+}\}\,, \eq $$ \gg_1=\Bigl\{\gt^{R,1}_{a,\pm}\,,~\gt^{L,1}_{4,\pm}\,,~\gt^{R,+,1}_{a,\pm}\,,~\gt^{L,+,1}_{a,+}\,,~ (a\neq 4,8)\,,~\gt^{L,+,1}_{4,\pm}\,,\Bigr. $$ $$ \gt^1_{1j}\,,~\gt^1_{j1}\,,\,(j=2,3,5)\,,~\gt^{1}_{5,\pm}\,,~\gt^{1}_{6,+}\,,~ \gh_{\al_1}^1\,,~\gh_{\al_2}^1\,,~\gh^1_{e_5} \Bigr\}\,. $$ The Killing form on the GS generators takes the form $$ (\gt^{R,+,k_1}_{a,+},\gt^{R,+,k_2}_{b,-})=\de_{a,-b}\de^{k_1+k_2,0\,,~mod(2)}\,, ~~(\gt^{L,+,k_1}_{4,\pm},\gt^{L,+,k_2}_{4,\pm})=\de^{k_1+k_2,0\,,~mod(2)}\,, $$ $$ (\gt^{L,k_1}_{4,\pm},\gt^{L,k_2}_{4,\mp})=\de^{k_1+k_2,0\,,~mod(2)}\,,~~ (\gt^{R,k_1}_{a,\pm},\gt^{R,k_2}_{b,\mp})=\de_{a,-b}\de^{k_1+k_2,0\,,~mod(2)}\,, $$ \beq{kfgse7} (\gt^{L,+,k_1}_{a,+},\gt^{L,+,k_2}_{b,+})=\de_{a,-b}\de^{k_1+k_2,0\,,~mod(2)}\,,~ (a\neq 4,8)\,,~~(t^1_{1,j},t^1_{k,1})=\de_{jk}\,, \eq $$ (\gh^1_{e_6},\gh^1_{e_6})=1\,,~~~ (\gt^{k_1}_{5,+},\gt^{k_2}_{5,-})=\de^{k_1+k_2,0\,,~mod(2)}\,,~~~ (\gt^{k_1}_{6,+},\gt^{k_2}_{6,-})=\de^{k_1+k_2,0\,,~mod(2)}\,. $$ It allows us to define the dual basis (5.9.I) $$ \gT^{R,+,k}_{a,+}=\gt^{R,+,k}_{-a,-}\,,~~\gT^{R,+,k}_{a,-}=\gt^{R,+,k}_{-a,+}\,,~~ \gT^{L,+,k}_{4,\pm}=\gt^{L,+,k}_{4,\pm}\,, $$ $$ \gT^{R,k}_{a,+}=\gt^{R,k}_{-a,-}\,,~~ \gT^{L,+,k}_{4,\pm}=\gt^{L,+,k}_{4,\mp}\,, $$ \beq{dbe7} \gT^{L,+,k}_{a,\pm}=\gt^{L,+,k}_{-a,\mp}\,,~(a\neq 4,8)\,, ~~\gH^1_{e_5}=\gh^1_{e_5}=\f1{\sqrt{3}}(e_5+e_6+e_7)\,, \eq $$ \gT^{k}_{5,+}=\gt^{k}_{5,-}\,,~~~\gT^{k}_{6,+}=\gt^{k}_{6,-}\,, ~~\gT^1_{1,j}= t^1_{j,1}\,,~~ \gT^1_{j,1}=t^1_{1,j}\,. $$ It follows from (\ref{ghf4}) and (\ref{ce6a}) that the scalar product matrix $a_{j,k}=(\gh_{\al_j}^1,\gh_{\al_k}^1)$ for $(j,k=1,2)$ is the Cartan matrix $A_2$. We need the inverse matrix to define the dual basis (5.15.I) \beq{scpr7} \clA=\f1{3}\left( \begin{array}{cc} 2 & 1 \\ 1 & 2 \\ \end{array} \right)\,. \eq Thus, \beq{dbce7} \gH_{\al_1}^1=\frac{2}3\gh_{\al_1}^1+\f1{3}\gh_{\al_2}^1= \f1{3}(e_5-2e_6+e_7)\,, \eq \beq{dbce7a} \gH_{\al_2}^1=\frac{1}3\gh_{\al_1}^1+\frac{2}{3}\gh_{\al_2}^1= \f1{2\sqrt{2}}(-e_1+e_2-e_3-e_4)+\f1{6}(e_5-2e_6+e_7)\,. \eq The scalar product $(\gH_{\al_j}^1,\gH_{\al_k}^1)$ is defined by $\clA$ (\ref{scpr7}), while $\gH^1_{e_5}$ is orthogonal to them. \subsection*{Lax operators and Hamiltonians} \subsubsection*{Trivial bundles} The moduli space of trivial $\bG=E_7$ bundles is the quotient $$ \gH(\ble)/W_{\ble}\ltimes(\tau(Q(\ble)+Q(\ble))\,. $$ The moduli space of trivial $G^{ad}=E_7/\mu_2$ bundles is the quotient $$ \gH(\ble)/W_{\ble}\ltimes(\tau(P(\ble)+P(\ble))\,, $$ where $P(\ble)$ are defined by the basis (\ref{wle7}). For trivial bundles Lax operator takes the form $$ L^{CM}_{\ble}(z)=L^{CM}_{\bae}(z)+ \sum_{a=1}^8\sum_{A=L,R,V}S_{a,\pm}^{A,+}\phi((\bfu,\varpi_a^A)\pm\f1{\sqrt{2}}(u_j+u_k),z) E_{a,\pm}^{A,+} +\sum_{j=5,6,7} S_{j,\pm}\phi( \pm\sqrt{2}u_j,z) E_{j,\pm}\,. $$ Here we have the correspondence $L\to (j=5,7)$, $R\to (j=5,6)$, $V\to (j=6,7)$. In contrast with $\bae$ there is no restrictions $\sum_{k=1}^7 v_k=\sum_{k=1}^7 u_k=0$. The Hamiltonian after the symplectic reduction with respect to the $\clH(\ble)$ action assumes the form $$ H^{CM}_{\ble}=H^{CM}_{\bae}+\sum_{a=1}^4\sum_{A=L,R,V}S_{a,+}^{A,+}S_{a,-}^{A,+} E_2((\bfu,\varpi_a^A)+\f1{\sqrt{2}}(u_j+u_k))+ \sum_{j=5,6,7} S_{j,+}S_{j,-}E_2(\sqrt{2}u_j,z)\,. $$ \subsubsection*{Nontrivial bundles} \textbf{The moduli space.}\\ The moduli space are described by elements $\ti\bfu\in\gH(\baf)$ (\ref{cf4}). The Weyl group $W(\baf)$ is generated by reflections with respect the planes, orthogonal to simple roots (\ref{srf4}). The coroot lattice $\ti Q^\vee(\baf)$ is generated by the simple coroots (\ref{cbe7}). The moduli space of nontrivial $E_7$ bundles is defined as $$ \gH(\baf)/(W(\baf)\ltimes(\tau\ti Q^\vee(\baf)+\ti Q^\vee(\baf)))\,. $$ \bigskip \textbf{$F_4$ Calogero-Moser system.}\\ Represent the Lax operator in the form (6.7.I) $$ L(z)=\ti L_0(z)+L'(z)+L_1(z)\,, $$ where $\ti L_0(z)=L_{\baf}^{CM}(z)$. In defined below expression $\ti\bfv=(v_1,v_2,v_3,v_4,v_5)$ the momentum vector satisfies the same restriction as the vector $\ti\bfu$ (\ref{fres}). It follows from (\ref{f4d}) it takes the form $$ L_{\baf}^{CM}(z)=L_{\bf so(7)}^{CM}(z)+ \frac{1}4\Bigl(-3(v_1+S_{0,1}E_1(z))e_1-(v_2+S_{0,2}E_1(z))e_2-(v_3+S_{0,3}E_1(z))e_3 $$ $$ +(v_4+S_{0,4}E_1(z))e_4)+ \f1{2\sqrt{2}}((v_5+S_{0,5}E_1(z))e_5-(v_5+S_{0,5}E_1(z))e_7)\Bigr)+ $$ $$ +\sum_{a=1}^8\Bigl(S_{a,+}^{R,0}\phi((\ti\bfu,\varpi_a^R)+\f1{\sqrt{2}}u_5,z)\gt^{R,0}_{-a,-} +S_{-a,-}^{R,0}\phi(-(\ti\bfu,\varpi_a^R)-\f1{\sqrt{2}}u_5,z)\gt^{R,0}_{a,+}\Bigr) $$ $$ S_{4,+}^{L,0}\phi((\ti\bfu,\varpi_4^L)+\sqrt{2}u_5,z)\gt^{L,0}_{4,-}+ S_{4,-}^{L,0}\phi(-(\ti\bfu,\varpi_4^L)-\sqrt{2}u_5,z)\gt^{L,0}_{4,+}+ $$ $$ \sum_{a\neq 4}^8\Bigl(S_{a,+}^{L,0}\phi((\ti\bfu,\varpi_a^L)+\sqrt{2}u_5,z)\gt^{L,0}_{-a,-} +S_{-a,-}^{L,0}\phi(-(\ti\bfu,\varpi_a^L)-\sqrt{2}u_5,z)\gt^{R,0}_{a,+}\Bigr)\,. $$ We find the corresponding quadratic Hamiltonian $$ H_{\baf}^{CM}=H_{\bf so(7)}^{CM}+\oh\Bigl(v_1^2+v_2^2+v_3^2+v_4^2+2v_5^2\Bigr) $$ $$ +\sum_{a=1}^8S_{a,+}^{R,0}S_{-a,-}^{R,0}E_2((\ti\bfu,\varpi_a^R)+\f1{\sqrt{2}}u_5) +S_{4,+}^{L,0}S_{4,-}^{L,0}E_2((\ti\bfu,\varpi_4^L)+\sqrt{2}u_5) $$ $$ +\sum_{a\neq 4}^8S_{a,+}^{L,0}S_{-a,-}^{L,0}E_2((\ti\bfu,\varpi_a^L)+\sqrt{2}u_5)\,. $$ \bigskip \textbf{The Lax operators and Hamiltonians}\\ Define as in the general case (6.14.I) the function $$ \varphi^k_\be(\ti\bfu,z)=\bfe\,\Bigl(\lan\ka_{\bf e_7},\be\ran z\Bigr) \phi(\lan\ka_{\bf e_7}-\ti\bfu,\be\ran+k/2, z)\,,~~~(k=0,1)\,, $$ where $\ka_{\bf e_7}$ (\ref{ka7}) and $\be\in R({\bf e_7})$ generating the GS basis. The following $R({\bf e_7})$ roots define GS generators (\ref{gsbe7}) \beq{be7} \be= \left\{ \al_{(a,\pm)}^{(L,+)}\,,~~ \al_{(a,\pm)}^{(R,+)}\,, ~ ~\al_{1j}=e_1-e_j\,,~(j=2,3,5)\,,~~(\ref{sroe7})\,, \right. \eq $$ \left. \al_{(a,\pm)}^{L}\,,~ \al_{(a,\pm)}^{R}~(\ref{sroe6})\,, ~~ \pm \al^{(+)}_j,~(j=5,6,7)\,,~(\ref{re7gs})\,,~~ \right\}\,. $$ Then following (6.15.I) and (6.17.I) from (\ref{gsbe7}) we find $$ L_0'(z)=\sum_{a=1}^8S_{a,\pm}^{R,+,0}\varphi^0_{ \al_{(a,\pm)}^{(L,+)}}(\ti\bfu,z) \gt^{R,+,0}_{-a,\mp}+ \sum_{a\neq 4,8}S_{a,\pm}^{L,+,0}\varphi^0_{ \al_{(a,\pm)}^{(L,+)}}(\ti\bfu,z)\gt^{L,+,0}_{-a,\mp}+ $$ $$ S^{L,+,0}_{4,\pm}\varphi^0_{ \al_{(4,\pm)}^{(L,+)}}(\ti\bfu,z)\gt^{L,+,0}_{4,\pm}+ S^{0}_{5,\pm}\varphi^0_{ \pm\al_{(5)}^{(+)}}(\ti\bfu,z)\gt^{0}_{5,\mp}+ S^{0}_{6,\pm}\varphi^0_{ \pm\al_{(6)}^{(+)}}(\ti\bfu,z)\gt^{0}_{6,\mp}\,. $$ $$ L_1(z)=\Bigl(S^1_{\al_1}\gH_{\al_1}^1+S^1_{\al_2}\gH_{\al_2}^1 +S^1_{e_5}\gh^1_{e_5}\Bigr)\phi(\oh,z) $$ $$ \sum_{a=1}^8S^{R,1}_{a,\pm}\varphi^1_{ \al_{(a,\pm)}^{(R)}}(\ti\bfu,z)\gt^{R,1}_{a,\mp}+ S^{L,1}_{4,\pm}\varphi^1_{ \al_{(4,\pm)}^{(L)}}(\ti\bfu,z)\gt^{L,1}_{4,\mp} $$ $$ +\sum_{a=1}^8S_{a,\pm}^{R,+,1} \varphi^1_{\al_{(a,\pm)}^{(R,+)}}(\ti\bfu,z) \gt^{R,+,1}_{-a,\mp}+ \sum_{a\neq 4,8}S_{a,\pm}^{L,+,1}\varphi^1_{ \al_{(a,\pm)}^{(L,+)}}(\ti\bfu,z)\gt^{L,+,1}_{-a,\mp}+ $$ $$ \sum_{j=2,3,4,5}\Bigl(S^1_{1,j}\varphi^1_{ \al_{(1,j)}}(\ti\bfu,z)t^1_{j,1}+ S^1_{j,1}\varphi^1_{ \al_{(j,1)}}(\ti\bfu,z)t^1_{1,j}\Bigr)+ $$ $$ S^{L,+,1}_{4,\pm}\varphi^1_{ \al_{(4,\pm)}^{(L,+)}}(\ti\bfu,z)\gt^{L,+1}_{4,\mp}+ S^{1}_{5,\pm}\varphi^1_{\pm\al_{(5)}^{(+)}}(\ti\bfu,z)\gt^{1}_{5,\mp}+ S^{1}_{6,\pm}\varphi^1_{\pm\al_{(6)}^{(+)}}(\ti\bfu,z)\gt^{1}_{6,\mp}\,. $$ where for $\gH_{\al_1}^1$ and $\gH_{\al_2}^1$ see (\ref{dbce7}) and (\ref{dbce7a}). For the $E_7$ quadratic Hamiltonians we have $$ H_{\ble}=H^{CM}_{\baf}+H_0'+H_1\,, $$ where $H_0'$ comes from $\oh(L_0^{'2})$ and $H_1$ from $\oh(L_1^{2})$. To calculate the Hamiltonians we use (\ref{kfgse7}) and (\ref{scpr7}) for scalar products of the dual Cartan generators $\gH_{\al_j}^1$. Then $$ -H_0'=\sum_{a=1}^8S_{a,+}^{R,+,0}S_{-a,-}^{R,+,0}E_2(\lan \al_{(a,+)}^{(L,+)},\ti\bfu\ran)+ \sum_{a\neq 4,8}S_{a,+}^{L,+,0}S_{-a,-}^{L,+,0} E_2(\lan \al_{(a,+)}^{(L,+)},\ti\bfu\ran)+ $$ $$ S^{L,+,0}_{4,+}S^{L,+,0}_{4,-}E_2( \lan\al_{(4,+)}^{(L,+)},\ti\bfu\ran)+ S^{0}_{5,+}S^{0}_{5,-}E_2(\lan\al_{(5)}^{(+)},\ti\bfu\ran)+ S^{0}_{6,+}S^{0}_{6,-}E_2(\lan\al_{(6)}^{(+)},\ti\bfu\ran)\,. $$ $$ H_1=\frac{2}3\Bigl((S^1_{\al_1})^2+(S^1_{\al_2})^2+S^1_{\al_1}S^1_{\al_2}\Bigr)E_2\Bigl(\oh\Bigr) +(S^1_{e_5})^2E_2\Bigl(\oh\Bigr) $$ $$ +\sum_{a=1}^8S^{R,1}_{a,+}S^{R,1}_{-a,-}E_2(\lan \al_{(a,+)}^{(R)},\ti\bfu+\oh\ran)+ S^{L,1}_{4,+}S^{L,1}_{4,-}E_2(\lan\al_{(4,+)}^{(L)},\ti\bfu\ran+\oh) $$ $$ +\sum_{a=1}^8S_{a,+}^{R,+,1}S_{-a,-}^{R,+,1}E_2(\lan\al_{(a,+}^{(R,+)},\ti\bfu\ran+\oh) + \sum_{a\neq 4,8}S_{a,+}^{L,+,1}S_{-a,-}^{L,+,1}E_2(\lan\al_{(a,+)}^{(L,+)},\ti\bfu\ran+\oh) $$ $$ +\sum_{j=2,3,4,5}S^1_{1,j}S^1_{j,1}E_2(\lan \al_{(1,j)},\ti\bfu\ran+\oh) +S^{L,+,1}_{4,+}S^{L,+,1}_{4,-}E_2(\lan \al_{(4,+)}^{(L,+)},\ti\bfu\ran+\oh) $$ $$ + S^{1}_{5,+}S^{1}_{5,-}E_2(\lan\al_{(5)}^{(+)},\ti\bfu\ran+\oh)+ S^{1}_{6,+}S^{1}_{6,-}E_2(\lan\al_{(6)}^{(+)},\ti\bfu\ran+\oh) \,, $$ where $\ti\bfu$ is defined by (\ref{cf4}), (\ref{fres}). \small{

https://arxiv.org/abs/1907.10842

Using Boolean cumulants to study multiplication and anticommutators of free random variables

We study how Boolean cumulants can be used in order to address operations with freely independent random variables, particularly in connection to the $*$-distribution of the product of two selfadjoint freely independent random variables, and in connection to the distribution of the anticommutator of such random variables.

\section{Introduction} \subsection{Multiplication of free random variables, in terms of free cumulants.} $\ $ \noindent Let $( {\mathcal A} , \varphi )$ be a noncommutative probability space. It is known since the 90's (cf. \cite{NiSp-Vo1996}) how to handle the multiplication of two freely independent elements of $ {\mathcal A} $ in terms of free cumulants. More precisely, let $( \kappa_n : {\mathcal A} ^n \to {\mathbb C} )_{n=1}^{\infty}$ be the family of free cumulant functionals of $( {\mathcal A} , \varphi )$. If $a,b \in {\mathcal A} $ are freely independent, then the free cumulants of the product $ab$ are described by the formula \begin{equation} \label{eqn:11a} \kappa_n (ab, \ldots , ab) = \sum_{\pi \in NC(n)} \prod_{U \in \pi} \kappa_{|U|} (a, \ldots , a) \cdot \prod_{V \in \mbox{Kr} ( \pi )} \kappa_{|V|} (b, \ldots , b), \end{equation} where $NC(n)$ is the lattice of non-crossing partitions of $\{ 1, \ldots , n \}$, and $ \mbox{Kr} : NC(n) \to NC(n)$ is an important anti-automorphism of this lattice, called Kreweras complementation map. The formula (\ref{eqn:11a}) is very useful because it allows one to take advantage of many pleasant properties the lattices $NC(n)$ are known to have. In particular, upon re-writing (\ref{eqn:11a}) in terms of formal power series and upon doing suitable manipulations, one can use it (cf. \cite{NiSp1997}) to derive the multiplicativity of the well-known $S$-transform of Voiculescu \cite{Vo1987}. In view of how we will make our presentation of results below, it is worth mentioning here that the clearest proof of the formula (\ref{eqn:11a}) is made in 3 steps, as follows: \begin{equation} \label{eqn:11b} \left\{ \begin{array}{ll} \mbox{Step 1.} & \mbox{On the left-hand side of (\ref{eqn:11a}), use the formula} \\ & \mbox{$\ $ (with summation over $NC(2n)$) which describes} \\ & \mbox{$\ $ free cumulants with products as arguments.} \\ \mbox{Step 2.} & \mbox{Use the fact that, due to the freeness of $a$ from $b$,} \\ & \mbox{$\ $ all their mixed free cumulants vanish.} \\ \mbox{Step 3.} & \mbox{Perform a direct combinatorial analysis of the non-crossing} \\ & \mbox{$\ $ partitions in $NC(2n)$ which were not pruned in Step 2.} \end{array} \right. \end{equation} \vspace{6pt} Let us now upgrade to the framework where $( {\mathcal A} , \varphi )$ is a $*$-probability space, and where $a,b$ are two freely independent selfadjoint elements of $ {\mathcal A} $. Since $ab$ isn't generally selfadjoint, we now need to keep track of the joint moments, or equivalently of the joint free cumulants of $ab$ and $(ab)^{*} = ba$. That is, we now need to look at free cumulants of the form \begin{equation} \label{eqn:11c} \kappa_n \bigl( \, (ab)^{\varepsilon (1)}, \ldots , (ab)^{\varepsilon (n)} \, \bigr) , \ \mbox{ with $n \in {\mathbb N} $ and } \varepsilon = ( \varepsilon (1), \ldots , \varepsilon (n) ) \in \{ 1,* \}^n . \end{equation} The tools invoked in Steps 1 and 2 of (\ref{eqn:11b}) can still be used in connection to the cumulants from (\ref{eqn:11c}). But the combinatorial analysis in Step 3 (where some version of the Kreweras complementation map would be hoped to appear) becomes ad-hoc and does not seem to reveal a pattern -- this is seen on very simple examples, e.g. when doing the calculation which expresses $\kappa_3 ( ab, ab, (ab)^{*} )$ in terms of the $*$-free cumulants of $a$ and those of $b$. \vspace{10pt} \subsection{Use Boolean cumulants instead of free cumulants?} $\ $ \noindent For a noncommutative probability space $( {\mathcal A} , \varphi )$, one can also consider the family $( \beta_n : {\mathcal A} ^n \to {\mathbb C} )_{n=1}^{\infty}$ of Boolean cumulant functionals of $( {\mathcal A} , \varphi )$. Boolean cumulants are the analogue of free cumulants in the parallel (and simpler) world of Boolean probability. They are known to have some direct connections with free probability, particularly in connection to a development called ``Boolean Bercovici-Pata bijection''. An intriguing fact around this topic is that \begin{center} ``the Boolean Bercovici-Pata bijection preserves the structure behind multiplication of free random variables''. \end{center} A possible way to pitch this fact is as follows: when one describes the multiplication of two freely independent elements $a, b \in {\mathcal A} $ in terms of Boolean cumulants, the resulting formula has exactly the same structure as in (\ref{eqn:11a}): \begin{equation} \label{eqn:12a} \beta_n (ab, \ldots , ab) = \sum_{\pi \in NC(n)} \prod_{U \in \pi} \beta_{|U|} (a, \ldots , a) \cdot \prod_{V \in \mbox{Kr} ( \pi )} \beta_{|V|} (b, \ldots , b), \ \ \forall \, n \geq 1. \end{equation} The formula (\ref{eqn:12a}) was first found in \cite[Theorem 2']{BeNi2008}. It can be proved via a strategy with 3 steps parallel to the one described in (\ref{eqn:11b}). (See also \cite[Lemma 3.2]{PoWa2011} for a similar result stated in terms of the so-called ``c-free cumulants'', which relate at the same time to free and to Boolean cumulants.) The main point of the present paper is that, for Boolean cumulants, the strategy with 3 steps can be pushed to the framework where $( {\mathcal A} , \varphi )$ is a $*$-probability space and where we look at Boolean cumulants of the form \begin{equation} \label{eqn:12b} \beta_n \bigl( \, (ab)^{\varepsilon (1)}, \ldots , (ab)^{\varepsilon (n)} \, \bigr), \mbox{ with $n \in {\mathbb N} $ and } \varepsilon = ( \varepsilon (1), \ldots , \varepsilon (n) ) \in \{ 1,* \}^n . \end{equation} What makes this possible is the use of an alternative facet of Kreweras complementation, suggested by the work in \cite{BeNi2008}. This facet of Kreweras complementation is discussed in the next subsection. (We reiterate here that the possibility of using it is specific to Boolean cumulants, and -- as seen on very simple examples -- fails to work for free cumulants.) \vspace{10pt} \subsection{Kreweras complementation and VNRP property.} $\ $ \noindent The definition of the Kreweras complement for a partition $\pi \in NC(n)$, as originally given by G. Kreweras \cite{Kr1972}, goes via a maximality argument. One considers partitions (not necessarily non-crossing) of $\{ 1, \ldots , 2n \}$ of the form ``$ \pi^{(\mathrm{odd})} \sqcup \rho^{(\mathrm{even})} $'', which have a copy of $\pi$ placed on $\{ 1,3, \ldots , 2n-1 \}$ and a copy of some other partition $\rho \in NC(n)$ placed on $\{ 2,4, \ldots , 2n \}$. The Kreweras complement $ \mbox{Kr} _n ( \pi )$ is the maximal element, with respect to the reverse refinement order ``$\leq$'' on $NC(n)$, for the set $\{ \rho \in NC(n) \mid \pi^{(\mathrm{odd})} \sqcup \rho^{(\mathrm{even})} \in NC(2n) \}$. The paper \cite{BeNi2008} considered another partial order ``$\ll$'' on $NC(n)$, coarser than the reverse refinement order $\leq$, which is useful for studying connections between free cumulants and Boolean cumulants. In Proposition 6.10 of \cite{BeNi2008}, another maximality property related to Kreweras complements was noticed to hold: the partitions of the form $ \pi^{(\mathrm{odd})} \sqcup ( \mbox{Kr} _n (\pi) )^{\mathrm{(even)}}$ are the maximal elements with respect to $\ll$ for the set \begin{equation} \label{eqn:13a} \Bigl\{ \sigma \in NC(2n) \begin{array}{ll} \vline & \mbox{every block of $\sigma$ is contained either} \\ \vline & \mbox{in $\{1,3, \ldots , 2n-1 \}$ or in $\{ 2,4, \ldots , 2n \}$, } \\ \vline & \mbox{and $\sigma$ has exactly two outer blocks} \end{array} \Bigr\} ; \end{equation} moreover, for every $\sigma$ in the above set, there exists a unique $\pi \in NC(n)$ such that $\sigma \ll \pi^{(\mathrm{odd})} \sqcup ( \mbox{Kr} _n (\pi))^{\mathrm{(even)}}$. \noindent [The concept of outer block, and the related concept of depth for a block of a non-crossing partition are reviewed in Section 2 below. The requirement ``$\sigma$ has exactly two outer blocks'' in (\ref{eqn:13a}) is a minimality condition, since the block of $\sigma$ which contains the number $1$ and the block of $\sigma$ which contains the number $2n$ are always sure to be outer blocks.] \vspace{6pt} In the present paper we extend the result described above to a framework where considering parities is a special case of considering a colouring $c : \{ 1, \ldots , 2n \} \to \{ 1,2 \}$ (given by $c(i) = i ( \mbox{mod $2$})$ for $1 \leq i \leq 2n$). Upon examining the point of view of colourings, one finds that the property which isolates partitions of the form $ \pi^{(\mathrm{odd})} \sqcup ( \mbox{Kr} _n (\pi))^{\mathrm{(even)}}$ within the set (\ref{eqn:13a}) is a certain {\em vertical-no-repeat property}, or VNRP for short. It is straightforward how to define VNRP for a general colouring (Definition \ref{def:41} below). Given a partition $\sigma \in NC(m)$ and a colouring in $s$ colours $c : \{ 1, \ldots , m \} \to \{ 1, \ldots , s \}$ which is constant along the blocks of $\sigma$, the fact that $\sigma$ and $c$ have VNRP amounts to the requirement that \[ c ( \, \mathrm{Parent} _{\sigma} (V) \, ) \neq c(V), \ \ \mbox{ for every inner block $V$ of $\sigma$.} \] [Here we refer to the fairly intuitive fact that every inner block $V$ of a non-crossing partition $\sigma$ must have a ``parent-block'' $ \mathrm{Parent} _{\sigma} (V)$ into which it is nested. The precise definition of how this goes is reviewed in Section 2 below.] The key-property of VNRP, extending the considerations on Kreweras complements from the preceding paragraph, is then stated as follows. \begin{theorem} \label{thm:131} {\em (Key-property of VNRP.) } \noindent Let $m$ be in $ {\mathbb N} $, let $c : \{ 1, \ldots , m \} \to \{ 1, \ldots , s \}$ be a colouring, and consider the set of partitions \[ NC(m;c) := \{ \sigma \in NC(m) \mid \mbox{ $c$ is constant on every block of $\sigma$} \} . \] For every $\sigma \in NC(m;c)$ there exists a $\tau \in NC(m;c)$, uniquely determined, such that $\sigma \ll \tau$ and such that $\tau$ has the VNRP property \footnote{ The correct formulation here would be to say that $\tau$ and $c$ (together) have VNRP. We will occasionally replace this with saying that ``$\tau$ has VNRP with respect to $c$'', or that ``$c$ has VNRP with respect to $\tau$''.} with respect to $c$. \end{theorem} $\ $ \subsection{Free independence in terms of Boolean cumulants.} $\ $ \noindent An easy calculation based on Theorem \ref{thm:131} leads to a description of free independence in terms of Boolean cumulants, as follows. \begin{theorem} \label{thm:141} Let $( {\mathcal A} , \varphi )$ be a noncommutative probability space and let $( \beta_n : {\mathcal A} ^n \to {\mathbb C} )_{n=1}^{\infty}$ be the family of Boolean cumulant functionals associated to it. Let $ {\mathcal A} _1,\ldots, {\mathcal A} _s\subseteq {\mathcal A} $ be unital subalgebras. The following two statements are equivalent. \vspace{6pt} (1) $ {\mathcal A} _1,\ldots, {\mathcal A} _s$ are free with respect to $\varphi$. \vspace{6pt} (2) For every $n \in {\mathbb N} $, every colouring $c : \{ 1, \ldots , n \} \to \{ 1, \ldots , s \}$, and every $a_1 \in {\mathcal A} _{c(1)}, \ldots a_n \in {\mathcal A} _{c(n)}$, one has \begin{equation} \label{eqn:141a} \beta_n (a_1, \ldots , a_n) = \sum_{ \begin{array}{c} {\scriptstyle \pi \in NC(n;c) \ with} \\ {\scriptstyle unique \ outer \ block} \\ {\scriptstyle and \ with \ VNRP} \end{array} } \ \prod_{V \in \pi} \beta_{|V|} \bigl( (a_1, \ldots a_n) | V \bigr). \end{equation} \end{theorem} \begin{remark} \label{rem:142} (1) The essential part of Theorem \ref{thm:141} is the implication $(1) \Rightarrow (2)$, which gives an explicit formula for Boolean cumulants with free arguments. Once this is established, the converse $(2) \Rightarrow (1)$ follows by combining $(1) \Rightarrow (2)$ with a standard ``replica trick''. (2) Coming from the study of a very general notion of noncommutative independence, Proposition 4.30 of the recent paper \cite{JeLi2019} gives a description of free independence in terms of Boolean cumulants which (when considered with $ {\mathbb C} $ as field of scalars) is equivalent to the above Theorem \ref{thm:141}. To be precise, condition 2 in Proposition 4.30 of \cite{JeLi2019} is the moment formula which comes out when one performs an additional summation over interval partitions on both sides of (\ref{eqn:141a}) -- see Corollary \ref{cor:53} below. Conversely, the latter moment formula can be used to retrieve Equation (\ref{eqn:141a}), via an easy application of M\"obius inversion. (3) Equation (\ref{eqn:141a}) implies an amusing formula for the Boolean cumulants of the sum of two freely independent random variables. The structure of this formula cannot be as simple as what one gets by using free cumulants, but we present it nevertheless in Proposition \ref{prop:55} below, in anticipation of the similarly looking formula concerning free anticommutators (where the use of free cumulants does not provide a simpler alternative). \end{remark} \vspace{10pt} \subsection{Joint Boolean cumulants for $\mathbf{ab}$ and $\mathbf{(ab)^{*}}$, and free anticommutators.} $\ $ \noindent We now continue the thread from Section 1.2, concerning the joint Boolean cumulants of $ab$ and $(ab)^{*}$, where $a$ and $b$ are freely independent selfadjoint elements in a $*$-probability space. We will put into evidence some special sets of non-crossing partitions which appear in the explicit formula for a joint cumulant $ \beta_n \bigl( \, (ab)^{\varepsilon (1)}, \ldots , (ab)^{\varepsilon (n)} \, \bigr)$ as in (\ref{eqn:12b}) and then (upon doing a summation over $\varepsilon \in \{ 1,* \}^n$) in the explicit formula for the Boolean cumulants of a free anticommutator $ab + ba$. We will refer to these special non-crossing partitions by using the ad-hoc term of ``anticommutator-friendly'', due to how they appear in the formula for the Boolean cumulants of a free anticommutator, in Theorem \ref{thm:154} below. Their actual definition doesn't, however, require any knowledge of a free probabilistic framework, and is stated as follows. $\ $ \begin{definition} \label{def:151} Let $n$ be a positive integer, let $\sigma$ be a partition in $NC(2n)$, and let us consider the set $ \mathrm{OuterMax} ( \sigma ) := \{ \max (W) \mid \mbox{ $W$ is an outer block of $\sigma$} \}$. We will say that $\sigma$ is {\em anticommutator-friendly} when it satisfies the following two conditions. \vspace{6pt} \noindent (AC-Friendly1) $ \mathrm{OuterMax} ( \sigma ) \subseteq \{ 1,3, \ldots , 2n-1 \} \cup \{ 2n \}$. \vspace{6pt} \noindent (AC-Friendly2) For every $j \in \{ 1,3, \ldots , 2n-1 \} \setminus \mathrm{OuterMax} ( \sigma )$, one has $ \mathrm{depth} _{\sigma} (j) \neq \mathrm{depth} _{\sigma} (j+1)$, where ``$ \mathrm{depth} _{\sigma} (j)$'' stands for the depth of the block of $\sigma$ which contains the number $j$. \end{definition} \begin{notation-and-remark} \label{def:152} For every $n \in {\mathbb N} $ we will denote \[ NC_{\mathrm{ac-friendly}} (2n) := \{ \sigma \in NC(2n) \mid \mbox{ $\sigma$ is anticommutator-friendly} \} . \] For instance $ NC_{\mathrm{ac-friendly}} (2)$ consists of only one partition, $\{ \, \{ 1 \} , \{ 2 \} \} \in NC(2)$. (The partition $\sigma = \{ \, \{ 1,2 \} \, \}$ is not anticommutator-friendly because it has $ \mathrm{depth} _{\sigma} (1) = \mathrm{depth} _{\sigma} (2) = 0$.) The next such set, $ NC_{\mathrm{ac-friendly}} (4)$, consists of the 5 partitions which are depicted in Figure 1 below. The cardinalities of the sets of partitions $ NC_{\mathrm{ac-friendly}} (2n)$ are tractable, in the respect that their generating series satisfies an algebraic equation of order 4, which can be solved explicitly: \begin{equation} \label{eqn:15a} \sum_{n=1}^{\infty} | \, NC_{\mathrm{ac-friendly}} (2n) \, | z^n = \frac{1}{2} - \sqrt{ (1-8z) \frac{1-2z-\sqrt{1-8z}}{8z} }. \end{equation} \end{notation-and-remark} \vspace{0.5cm} \begin{minipage}{1\linewidth} \begin{center} \begin{tikzpicture} \draw[thick] (0,0) -- (0,0.25); \draw[thick] (2,0) -- (2,0.25); \draw[thick] (-1,0) to[out=90,in=90] node [sloped,above] {} (1,0); \node[below] (1) at (-1,0) {$1$}; \node[below] (2) at (0,0) {$2$}; \node[below] (3) at (1,0) {$3$}; \node[below] (4) at (2,0) {$4,$}; \draw[thick] (4,0) -- (4,0.25); \draw[thick] (7,0) -- (7,0.25); \draw[thick] (5,0) to[out=90,in=90] node [sloped,above] {} (6,0); \node[below] (1) at (4,0) {$1$}; \node[below] (2) at (5,0) {$2$}; \node[below] (3) at (6,0) {$3$}; \node[below] (4) at (7,0) {$4,$}; \draw[thick] (9,0) -- (9,0.25); \draw[thick] (11,0) -- (11,0.25); \draw[thick] (10,0) to[out=90,in=90] node [sloped,above] {} (12,0); \node[below] (7) at (9,0) {$1$}; \node[below] (8) at (10,0) {$2$}; \node[below] (9) at (11,0) {$3$}; \node[below] (10) at (12,0) {$4,$}; \end{tikzpicture} \end{center} \end{minipage} \begin{minipage}{1\linewidth} \begin{center} \begin{tikzpicture} \draw[thick] (2,0) -- (2,0.25); \draw[thick] (3,0) -- (3,0.25); \draw[thick] (1,0) to[out=90,in=90] node [sloped,above] {} (4,0); \node[below] (1) at (1,0) {$1$}; \node[below] (2) at (2,0) {$2$}; \node[below] (3) at (3,0) {$3$}; \node[below] (4) at (4,0) {$4,$}; \draw[thick] (7,0) to[out=90,in=90] node [sloped,above] {} (10,0); \draw[thick] (8,0) to[out=90,in=90] node [sloped,above] {} (9,0); \node[below] (7) at (7,0) {$1$}; \node[below] (8) at (8,0) {$2$}; \node[below] (9) at (9,0) {$3$}; \node[below] (10) at (10,0) {$4.$}; \end{tikzpicture} \end{center} \end{minipage} \begin{center} {\bf Figure 1.} {\em The 5 partitions in $ NC_{\mathrm{ac-friendly}} (4)$.} \end{center} $\ $ The formulas to be stated in Theorems \ref{thm:153} and \ref{thm:154} below will also refer to a ``canonical alternating colouring'' of the blocks of a non-crossing partition, which is described next. \begin{definition} \label{def:153} Let $m \in {\mathbb N} $ and $\sigma \in NC(m)$ be given. We will use the name {\em canonical alternating colouring} of $\sigma$ for the colouring $ \mathrm{calt} _{\sigma} : \{ 1, \ldots , m \} \to \{ 1,2 \}$ whose values on the blocks of $\sigma$ are determined by the following conditions: \vspace{6pt} \noindent (C-Alt1) Denoting by $W_1$ the block of $\sigma$ which contains the number $1$, one has $ \mathrm{calt} _{\sigma} (W_1) = 1$. \vspace{6pt} \noindent (C-Alt2) If $W$ and $W'$ are ``consecutive'' outer blocks of $\sigma$, with $\min (W') = 1 + \max (W)$, then $ \mathrm{calt} _{\sigma} (W') \neq \mathrm{calt} _{\sigma} (W)$. \vspace{6pt} \noindent (C-Alt3) If $V$ is an inner block of $\sigma$, then $ \mathrm{calt} _{\sigma} (V) \neq \mathrm{calt} _{\sigma} ( \, \mathrm{Parent} _{\sigma} (V) \, )$. \end{definition} $\ $ Note that if $\sigma \in NC(m)$ has a unique outer block, then $ \mathrm{calt} _{\sigma}$ simply follows the parities of the depths of blocks of $\sigma$. For a general $\sigma \in NC(m)$, $ \mathrm{calt} _{\sigma}$ first does an alternating colouring of the outer blocks of $\sigma$, going from left to right; then for every outer block $W$ of $\sigma$ one follows the vertical alternance idea in order to colour the blocks of $\sigma$ which are nested inside $W$. In order to state the explicit formula for a joint Boolean cumulant of the kind indicated in (\ref{eqn:12b}) of Section 1.2, there is one last observation we need to make, namely that: in the canonical alternating colouring of a partition $\sigma \in NC_{\mathrm{ac-friendly}} (2n)$ one can naturally read (encoded in the colouring) a tuple $\varepsilon \in \{ 1,* \}^n$, which will be denoted as ``$ \mathrm{oddtuple} ( \sigma )$'' -- see Notation \ref{def:64} below for the precise definition. We then have the following theorem. $\ $ \begin{theorem} \label{thm:153} Let $( {\mathcal A} , \varphi )$ be a $*$-probability space and let $( \beta_n : {\mathcal A} ^n \to {\mathbb C} )_{n=1}^{\infty}$ be the family of Boolean cumulant functionals associated to it. Consider two selfadjoint elements $a,b \in {\mathcal A} $ such that $a$ is freely independent from $b$, and consider the sequences of Boolean cumulants $( \beta_n (a) )_{n=1}^{\infty}$ and $( \beta_n (b) )_{n=1}^{\infty}$ of $a$ and of $b$ (where we use natural abbreviations such as $\beta_n (a) := \beta_n (a, \ldots , a)$, $n \in {\mathbb N} $). \vspace{6pt} (1) For $n \in {\mathbb N} $ and $\varepsilon = ( \varepsilon (1), \ldots , \varepsilon (n)) \in \{ 1,* \}^n$ such that $\varepsilon (1) = 1$, one has \begin{equation} \label{eqn:153a} \beta_n \bigl( \, (ab)^{\varepsilon (1)}, \ldots , (ab)^{\varepsilon (n)} \, \bigr) = \end{equation} \[ \sum_{ \begin{array}{c} {\scriptstyle \sigma \in NC_{\mathrm{ac-friendly}} (2n),} \\ {\scriptstyle such \ that} \\ {\scriptstyle \mathrm{oddtuple} ( \sigma ) = \varepsilon} \end{array} } \ \Bigl( \prod_{ \begin{array}{c} {\scriptstyle U \in \sigma , \ with} \\ {\scriptstyle \mathrm{calt} _{\sigma} (U) = 1} \end{array} } \ \beta_{|U|} (a) \Bigr) \cdot \Bigl( \prod_{ \begin{array}{c} {\scriptstyle V \in \sigma , \ with} \\ {\scriptstyle \mathrm{calt} _{\sigma} (V) = 2} \end{array} } \ \beta_{|V|} (b) \Bigr). \] \vspace{6pt} (2) Let $n \in {\mathbb N} $ and let $\varepsilon = ( \varepsilon (1), \ldots , \varepsilon (n)) \in \{ 1,* \}^n$ be such that $\varepsilon (1) = *$. Consider the complementary tuple $\varepsilon ' \in \{ 1,* \}^n$, uniquely determined by the requirement that $\varepsilon ' (i) \neq \varepsilon (i)$, for all $1 \leq i \leq n$. One has \begin{equation} \label{eqn:153b} \beta_n \bigl( \, (ab)^{\varepsilon (1)}, \ldots , (ab)^{\varepsilon (n)} \, \bigr) = \end{equation} \[ \sum_{ \begin{array}{c} {\scriptstyle \sigma \in NC_{\mathrm{ac-friendly}} (2n),} \\ {\scriptstyle such \ that} \\ {\scriptstyle \mathrm{oddtuple} ( \sigma ) = \varepsilon '} \end{array} } \ \Bigl( \prod_{ \begin{array}{c} {\scriptstyle U \in \sigma, \ with} \\ {\scriptstyle \mathrm{calt} _{\sigma} (U) = 1} \end{array} } \ \beta_{|U|} (b) \Bigr) \cdot \Bigl( \prod_{ \begin{array}{c} {\scriptstyle V \in \sigma , \ with} \\ {\scriptstyle \mathrm{calt} _{\sigma} (V) = 2} \end{array} } \ \beta_{|V|} (a) \Bigr). \] \end{theorem} $\ $ By summing over $\varepsilon \in \{ 1,* \}^n$ in Theorem \ref{thm:153}, we arrive to a formula for the Boolean cumulants of a free anticommutator. $\ $ \begin{theorem} \label{thm:154} Consider the same framework and notation as in Theorem \ref{thm:153}. Then, for every $n \in {\mathbb N} $, the $n$-th Boolean cumulant of $ab + ba$ is \[ \beta_n (ab+ba) = \sum_{\sigma \in NC_{\mathrm{ac-friendly}} (2n)} \ \Bigl( \prod_{ \begin{array}{c} {\scriptstyle U \in \sigma , } \\ {\scriptstyle \mathrm{calt} _{\sigma} (U) = 1} \end{array} } \ \beta_{|U|} (a) \Bigr) \cdot \Bigl( \prod_{ \begin{array}{c} {\scriptstyle V \in \sigma , } \\ {\scriptstyle \mathrm{calt} _{\sigma} (V) = 2} \end{array} } \ \beta_{|V|} (b) \Bigr) \] \begin{equation} \label{eqn:154a} + \ \sum_{\sigma \in NC_{\mathrm{ac-friendly}} (2n)} \ \Bigl( \prod_{ \begin{array}{c} {\scriptstyle U \in \sigma , } \\ {\scriptstyle \mathrm{calt} _{\sigma} (U) = 1} \end{array} } \ \beta_{|U|} (b) \Bigr) \cdot \Bigl( \prod_{ \begin{array}{c} {\scriptstyle V \in \sigma , } \\ {\scriptstyle \mathrm{calt} _{\sigma} (V) = 2} \end{array} } \ \beta_{|V|} (a) \Bigr) . \end{equation} \end{theorem} \begin{remark} \label{rem:155} (1) In Theorem \ref{thm:153} it should be noted that if we make $\varepsilon = (1,1, \ldots , 1)$, then what comes out is precisely the formula from \cite{BeNi2008} which was reviewed in Equation (\ref{eqn:12a}). A discussion of why this is the case appears in Remark \ref{rem:58} below. (2) In Theorem \ref{thm:154}, we note that formula (\ref{eqn:154a}) simplifies a lot when $a$ and $b$ have the same distribution. In this case, denoting by $( \lambda_n )_{n=1}^{\infty}$ the common sequence of Boolean cumulants of $a$ and of $b$, we find that the $n$-th Boolean cumulant of $ab+ba$ is \begin{equation} \label{eqn:155a} \beta_n (ab+ba) = 2 \cdot \sum_{\sigma \in NC_{\mathrm{ac-friendly}} (2n)} \ \prod_{V \in \sigma} \lambda_{|V|}. \end{equation} \end{remark} \begin{remark} \label{rem:156} In order to put things into perspective, we give here some background on the past work around the problem of the free anticommutator. A noteworthy fact to begin with is that this problem is vastly simplified when we make the additional hypothesis that $a$ and $b$ have symmetric distributions (that is, $\varphi (a^{2n-1}) = 0 =\varphi (b^{2n-1})$ for all $n \in {\mathbb N} $). In this case, the nonselfadjoint element $ab \in {\mathcal A} $ has a certain ``$R$-diagonal'' property (cf. Lecture 15 of \cite{NiSp2006}). From here it follows in particular that $ab + ba$ and $i(ab-ba)$ have the same distribution (due to radial symmetry displayed by $R$-diagonal elements); moreover, the common distribution of $ab + ba$ and $i(ab-ba)$ is tractable due to the very special form of joint free cumulants that an $R$-diagonal element and its adjoint are known to have. It is interesting that the combinatorial study of $i(ab-ba)$ remains tractable even when we drop the assumption of $a$ and $b$ having symmetric distributions, because one can follow (cf. \cite{NiSp1998}) how terms cancel in the expansions of the free cumulants $\kappa_n ( i(ab-ba))$. The situation is not at all the same concerning $ab + ba$, where there are no cancellations to be followed. Here the expansions for moments or cumulants just create some large summations, and the combinatorial line of attack goes via precise identification of the combinatorial structures which appear as index sets for these large summations. Another noteworthy possibility to be mentioned is via approaches that are plainly analytic in nature, and produce systems of equations which can in principle be used to calculate the Cauchy transform of $ab+ba$. Such a system of equations is proposed in \cite{Va2003}. Another possibility of proceeding on these lines is suggested by the linearization method championed in \cite{HeMaSp2018}. \end{remark} \vspace{10pt} \subsection{Equations with $\boldmath{\eta}$-series.} $\ $ \noindent The possibility of approaching the distribution of $ab + ba$ via a system of equations in (not necessarily convergent) power series can also be pursued in the framework of the present paper. Here we use the generating power series for Boolean cumulants, which are also known as {\em $\eta$-series}: for $a \in {\mathcal A} $, we put \begin{equation} \label{eqn:16a} \eta_a (z) := \sum_{n=1}^{\infty} \beta_n (a) z^n \in {\mathbb C} [[z]]. \end{equation} In Section 6 of the paper we make a detailed analysis of the $\eta$-series $\eta_{ab+ba}$ of a free anticommutator, and we come up with a system of equations which, when solved, leads to the explicit determination of this $\eta$-series. Our derivation of this system of equations is combinatorial in nature, and is intimately related to the study of recursions satisfied by ac-friendly non-crossing partitions. The best way to describe our system of equations leading to $\eta_{ab+ba}$ is in a $2 \times 2$ matrix form, where we make use of some auxiliary power series $f_{a,a}, f_{a,a^{*}}, f_{a^{*},a}, f_{a^{*},a^{*}}$ grouped \footnote{It is convenient to have the entries of $F_a$ indexed by symbols $a$ and $a^{*}$, even though the intended use of $F_a$ is when $a$ is selfadjoint. The rationale for this notation is given at the beginning of Section 6.1.} in a matrix \begin{equation} \label{eqn:16b} F_a =\begin{bmatrix} f_{a,a} & f_{a,a^*} \\ f_{a^*,a} & f_{a^*,a^*} \end{bmatrix}, \end{equation} and of some power series $f_{b,b}, \ldots , f_{b^{*},b^{*}}$ likewise grouped in a $2 \times 2$ matrix $F_b$. For illustration, in this Introduction we present the special case when $a$ and $b$ have the same distribution. In this case we only need to refer to the matrix $F_a$, and we have the theorem stated next. (In the case when $a$ and $b$ are not required to have the same distribution, we get a more involved system of equations, where we use both matrices $F_a$ and $F_b$. This is described in Theorem \ref{thm:61} below.) \begin{theorem} \label{thm:161} Let $( {\mathcal A} , \varphi )$ be a $*$-probability space, and let $a,b$ be selfadjoint elements of $ {\mathcal A} $ such that $a$ is free from $b$ and such that $a,b$ have the same distribution. (1) The matrix $F_a$ from Equation (\ref{eqn:16b}) is obtained by solving the matrix equation \begin{equation} \label{eqn:161a} F_a H_a = \eta_a (z H_a), \end{equation} where $\eta_a$ is the $\eta$-series of $a$ (as in (\ref{eqn:16a}), and \begin{align*} H_a & := \begin{bmatrix} f_{a^*,a^*} (1-f_{a,a^*})^{-1} & f_{a^*,a}+f_{a^*,a^*} (1-f_{a,a^*})^{-1} f_{a,a}\\ (1-f_{a,a^*})^{-1} & (1-f_{a,a^*})^{-1} f_{a,a} \end{bmatrix}. \end{align*} (2) The $\eta$-series of $ab+ba$ can be obtained from the entries of $F_a$ via the equation \begin{equation} \label{eqn:161b} \eta_{ab+ba}(z^2) = 2 \bigl( f_{a,a^*}(z)+ \frac{f_{a,a}(z)f_{a^*,a^*}(z)}{1-f_{a^*,a}(z)} \bigr). \end{equation} \end{theorem} \begin{remark} \label{rem:162} Very much in agreement with the discussion at the beginning of Remark \ref{rem:156}, the study of free anticommutators via equations in $\eta$-series also simplifies substantially in the case when $a$ and $b$ have symmetric distributions. In this case the matrices $F_a$ and $F_b$ mentioned above are sure to have some vanishing entries ($f_{a,a} = f_{a^{*}, a^{*}} = 0$ and $f_{b,b} = f_{b^{*}, b^{*}} = 0$), and the systems of equations that have to be solved become simpler, as shown in Proposition \ref{prop:64} and Corollary \ref{cor:65}. While the examples of free anticommutators of symmetric distributions are covered by the methods from \cite{NiSp1998}, it is nevertheless interesting to work out some examples of this kind and combine them with a use ``in reverse'' of Theorem \ref{thm:154}, in order to obtain corollaries about the enumeration of ac-friendly non-crossing partitions. For instance, in order to count the non-crossing partitions $\sigma \in NC_{\mathrm{ac-friendly}} (2n)$ with the property that all blocks $V \in \sigma$ have even cardinality, one uses elements $a,b \in {\mathcal A} $ which are freely independent and have distribution $\frac{1}{4} ( \delta_{- \sqrt{2}} + \delta_{\sqrt{2}} ) + \frac{1}{2} \delta_0$. The reason for choosing the latter distribution is that the common sequence $( \lambda_n )_{n=1}^{\infty}$ of Boolean cumulants for $a$ and $b$ simply has $\lambda_n = 1$ for $n$ even and $\lambda_n = 0$ for $n$ odd. In view of Remark \ref{rem:155}(2), the Boolean cumulant $\beta_n (ab+ba)$ is then equal to twice the cardinality we are interested to determine. Upon combining this with the explicit formula obtained for $\eta_{ab+ba}$, we can determine precisely what is the required cardinality, as explained in Example \ref{ex:69} and Corollary \ref{cor:610} below. \end{remark} \begin{example} \label{ex:163} In the framework of Theorem \ref{thm:161}, it is instructive to consider the simplest possible non-symmetric example, where both $a$ and $b$ have distribution $\frac{1}{2} ( \delta_0 + \delta_2 )$. \begin{center} \includegraphics[width=0.8\textwidth]{BinomZeroTwo} {\bf Figure 2.} {\em Plot of the density of distribution of $ab+ba$ for $a,b$ free and having distribution $\tfrac{1}{2}\delta_0+\tfrac{1}{2}\delta_2$, together with a histogram of eigenvalues of random matrix approximation.} \end{center} $\ $ The matrix equation from Theorem \ref{thm:161} is easy to solve in this example, and we end with an explicit formula for the $\eta$--series $\eta_{ab+ba} (z)$. This can be followed with a calculation of Cauchy transform and with a Stieltjes inversion, in order to concretely determine what is the law of $ab+ba$ -- we find an absolutely continuous distribution supported on the interval $[-1,8]$. The explicit formula for the density of this distribution and the calculations leading to it are presented in Proposition \ref{prop:611} below. Figure 2 shows the graph of the density $f(x)$ found for the law of $ab+ba$. For a check, Figure 2 also shows a histogram of empirical eigenvalues distribution for $AB+BA$ where $A$ is a diagonal $6000\times 6000$ matrix with half of diagonal entries equal $0$ and half equal 2, and $B=UAU^*$ where $U$ is a random unitary matrix. This example offers a very good illustration of how one gets to have different distributions for the free commutator and anticommutator -- indeed, the law of the commutator $i(ab-ba)$ is easily found to be the arcsine distribution on $[-2,2]$ (cf. Example \ref{ex:66}, and the discussion in the paragraph preceding Proposition \ref{prop:611}). We point out that this example has a combinatorial significance as well, and can be used (cf. Corollary \ref{cor:612}) to infer the formula indicated in Equation (\ref{eqn:15a}) for the generating series of cardinalities of sets $NC_{ac-friendly} (2n)$. \end{example} \vspace{10pt} \subsection{Organization of the paper.} $\ $ \noindent Besides the present Introduction, we have five other sections. After a review of background in Section 2, we discuss VNRP and prove Theorem \ref{thm:131} in Section 3. In Section 4 we discuss the applications of VNRP to free independence via Boolean cumulants. In Section 5 we prove the results about the joint Boolean cumulants of $ab$ and $(ab)^{*}$ and about the Boolean cumulants of the free anticommutator which were advertised in Section 1.5 above. Finally, in Section 6 we consider the conversion from Boolean cumulants to $\eta$-series, and prove the results that were advertised in Section 1.6 above. $\ $ \section{Background and Notation} In this section we review some background on set-partitions, and the two types of cumulants we want to work with. \vspace{6pt} \subsection{Nestings and depths for blocks of a non-crossing partition.} $\ $ \noindent We start by reviewing, for the sake of setting notation, the definition of the two basic types of set-partitions used in this paper, the {\em non-crossing partitions} and the {\em interval partitions}. \begin{definition} \label{def:21} (1) Let $n$ be a positive integer and let $\pi = \{ V_1 , \ldots , V_k \}$ be a partition of $\{ 1, \ldots ,n \}$; that is, $V_1 , \ldots , V_k$ are non-empty pairwise disjoint sets (called the {\em blocks} of $\pi$) with $V_1 \cup \cdots \cup V_k$ = $\{ 1, \ldots , n \}$. The number $k$ of blocks of $\pi$ will be denoted as $| \pi |$, and we will occasionally use the notation ``$V \in \pi$'' to mean that $V$ is one of $V_1, \ldots , V_k$. We say that $\pi \in NC(n)$ is an {\em interval partition} to mean that every block $V$ of $\pi$ is of the form $V = [i,j] \cap {\mathbb N} $ for some $1 \leq i \leq j \leq n$. We say that $\pi$ is a {\em non-crossing partition} to mean that for every $1 \leq i_1 < i_2 < i_3 < i_4 \leq n$ such that $i_1$ is in the same block with $i_3$ and $i_2$ is in the same block with $i_4$, it necessarily follows that all of $i_1, \ldots , i_4$ are in the same block of $\pi$. \vspace{6pt} (2) For every $n \in {\mathbb N} $, we denote by $ \mbox{Int} (n)$ the set of all interval partitions of $\{ 1, \ldots , n \}$, and we denote by $NC(n)$ the set of all non-crossing partitions of $\{ 1, \ldots , n \}$. \end{definition} \begin{remark} \label{rem:21} Clearly, one has $ \mbox{Int} (n) \subseteq NC(n)$ for all $n \in {\mathbb N} $. It is not hard to see that $| \mbox{Int} (n) | = 2^{n-1}$ and that $NC(n)$ is counted by the $n$-th Catalan number: \[ | NC(n) | = \mathrm{Cat} _n := \frac{ (2n)! }{ n! (n+1)!}, \ \ \forall \, n \in {\mathbb N} . \] For a more detailed introduction to the $NC(n)$'s, one can for instance consult Lectures 9 and 10 of \cite{NiSp2006}. \end{remark} \vspace{10pt} Given a non-crossing partition $\pi \in NC(n)$, it is convenient to formalize the notion of ``relative nesting'' for blocks of $\pi$, as follows. \begin{notation-and-remark} \label{def:22} Let $n$ be in $ {\mathbb N} $ and let $\pi$ be a partition in $NC(n)$. \vspace{6pt} (1) Let $V, W$ be blocks of $\pi$. We will write ``$V \stackrel{\mathrm{nest}}{\leq} W$'' to mean that we have the inequalities \[ \min (W) \leq \min (V) \mbox{ and } \max (W) \geq \max (V). \] We will write ``$V \stackrel{\mathrm{nest}}{<} W$'' to mean that $V \stackrel{\mathrm{nest}}{\leq} W$ and $V \neq W$. We will occasionally also use the notations $W \stackrel{\mathrm{nest}}{\geq} V$ instead of $V \stackrel{\mathrm{nest}}{\leq} W$ and $W \stackrel{\mathrm{nest}}{>} V$ instead of $V \stackrel{\mathrm{nest}}{<} W$. \vspace{6pt} (2) It is immediate that ``$ \stackrel{\mathrm{nest}}{\leq} $'' is a partial order relation on the set of blocks of $\pi$. A block $W \in \pi$ which is maximal with respect to $ \stackrel{\mathrm{nest}}{\leq} $ will be said to be an {\em outer block}. A block $V \in \pi$ which is not outer will be said to be an {\em inner block}. \end{notation-and-remark} \begin{remark-and-definition} \label{rem:23} Let $n$ be in $ {\mathbb N} $, let $\pi$ be a partition in $NC(n)$, and let $V$ be a block of $\pi$. It is easy to check that the set $\{ W \in \pi \mid W \stackrel{\mathrm{nest}}{\geq} V \}$ is totally ordered by $ \stackrel{\mathrm{nest}}{\leq} $. That is, we can write \begin{equation} \label{eqn:23a} \{ W \in \pi \mid W \stackrel{\mathrm{nest}}{\geq} V \} = \{ V_1, \ldots , V_k \} \end{equation} where $k \geq 1$ and $V_1 \stackrel{\mathrm{nest}}{<} V_2 \stackrel{\mathrm{nest}}{<} \cdots \stackrel{\mathrm{nest}}{<} V_k$. In (\ref{eqn:23a}) we note, in particular, that $V_1 = V$ and that $V_k$ is an outer block. The {\em depth} of $V$ in $\pi$ is defined as \[ \mathrm{depth} _{\pi} (V) := k-1 , \] with $k$ picked from Equation (\ref{eqn:23a}). If $k \geq 2$ (which is equivalent to saying that $ \mathrm{depth} _{\pi} (V) \neq 0$, or that $V$ is an inner block), then the block $V_2$ appearing in (\ref{eqn:23a}) is called the {\em parent-block for $V$}, and will be denoted as $ \mathrm{Parent} _{\pi} (V)$. The parent-block could be equivalently introduced via the requirement that \[ \left\{ \begin{array}{cl} \mbox{(i)} & V \stackrel{\mathrm{nest}}{<} \mathrm{Parent} _{\pi} (V) , \mbox{ and } \\ \mbox{(ii)} & \mbox{There is no block $V' \in \pi$ such that $V \stackrel{\mathrm{nest}}{<} V' \stackrel{\mathrm{nest}}{<} \mathrm{Parent} _{\pi} (V)$.} \end{array} \right. \] \end{remark-and-definition} \begin{remark} \label{rem:24} Let $n$ be in $ {\mathbb N} $ and let $\pi$ be a partition in $NC(n)$. (1) The notion of depth for the blocks of $\pi$ could also be defined recursively, by postulating that outer blocks have depth $0$ and by making the requirement that \[ \mathrm{depth} _{\pi} (V) = 1 + \mathrm{depth} _{\pi} \bigl( \, \mathrm{Parent} _{\pi} (V) \, \bigr) \ \mbox{ for every inner block $V \in \pi$.} \] (2) As mentioned in the Introduction, for an $i \in \{ 1, \ldots , n \}$ we will sometimes write ``$ \mathrm{depth} _{\pi} (i)$'' in order to refer to the depth of the block of $\pi$ which contains the number $i$. Thus $ \mathrm{depth} _{\pi}$ can be viewed as a special example of colouring of $\pi$ (a function from $\{ 1, \ldots , n \}$ to $ {\mathbb Z} $ which is constant along the blocks of $\pi$). \end{remark} \begin{remark} \label{rem:25} Let $n$ be in $ {\mathbb N} $ and let $\pi$ be a partition in $NC(n)$. It is easy to see that one can always list the set of outer blocks of $\pi$ as $\{ W_1, \ldots , W_{\ell} \}$ in such a way that \begin{equation} \label{eqn:25a} \left\{ \begin{array}{l} \min (W_1) = 1, \ \ \max (W_{\ell}) = n, \ \ \mbox{ and } \\ \min (W_{i+1} ) = 1 + \max (W_i) \ \ \mbox{ for every $1 \leq i < \ell$.} \end{array} \right. \end{equation} In the case when the number $\ell$ of outer blocks of $\pi$ is $\ell = 1$, the second condition in (\ref{eqn:25a}) is vacuous ($\pi$ has a unique outer block $W$, with $1,n \in W$). In the notation from (\ref{eqn:25a}): the interval partition \[ \overline{\pi} := \{ J_1, \ldots , J_{\ell} \} \mbox{ with $J_i := [ \min (W_i) , \max (W_i) ] \cap {\mathbb N} $, for $1 \leq i \leq \ell$} \] is sometimes called the {\em closure} of $\pi$ in the set of interval-partitions. Note that knowing what is $\overline{\pi}$ provides exactly the same information as knowing the set $ \mathrm{OuterMax} ( \pi )$ which was introduced in Definition \ref{def:151} of the Introduction. \end{remark} \vspace{10pt} \subsection{Review of free and of Boolean cumulant functionals.} $\ $ \noindent Let $( {\mathcal A} , \varphi )$ be a noncommutative probability space (in purely algebraic sense) -- that is, $ {\mathcal A} $ is a unital algebra over $ {\mathbb C} $ and $\varphi : {\mathcal A} \to {\mathbb C} $ a linear functional with $\varphi ( 1_{{ }_{ {\mathcal A} }} ) = 1$. In this subsection we briefly review the definition of the free and the Boolean cumulants of $( {\mathcal A} , \varphi )$. Before starting, we record a customary notation which will appear in the formulas for both types of cumulants: given an $n \in {\mathbb N} $, a tuple $( a_1, \ldots , a_n ) \in {\mathcal A} ^n$, and a non-empty subset $S = \{ i_1, \ldots , i_m \} \subseteq \{ 1, \ldots , n \}$ with $i_1 < \cdots < i_m$, we denote \begin{equation} \label{eqn:26a} ( a_1, \ldots , a_n ) \mid S := ( a_{i_1}, \ldots , a_{i_m} ) \in {\mathcal A} ^m. \end{equation} $\ $ \begin{definition} \label{def:26} Notations as above. \vspace{6pt} (1) The {\em free cumulants} associated to $( {\mathcal A} , \varphi )$ are the family of multilinear functionals $( \kappa_n : {\mathcal A} ^n \to {\mathbb C} )_{n=1}^{\infty}$ which is uniquely determined by the requirement that \begin{equation} \label{eqn:26b} \varphi (a_1 \cdots a_n) = \sum_{\pi \in NC(n)} \prod_{V \in \pi} \kappa_{|V|} ( \, (a_1, \ldots , a_n) \mid V \, ), \end{equation} holding for all $n \in {\mathbb N} $ and $a_1, \ldots , a_n \in {\mathcal A} $. \vspace{6pt} (2) The {\em Boolean cumulants} associated to $( {\mathcal A} , \varphi )$ are the family of multilinear functionals $( \beta_n : {\mathcal A} ^n \to {\mathbb C} )_{n=1}^{\infty}$ which is uniquely determined by the requirement that \begin{equation} \label{eqn:26c} \varphi (a_1 \cdots a_n) = \sum_{\pi \in \mbox{Int} (n)} \prod_{V \in \pi} \beta_{|V|} ( \, (a_1, \ldots , a_n) \mid V \, ), \end{equation} holding for all $n \in {\mathbb N} $ and $a_1, \ldots , a_n \in {\mathcal A} $. \end{definition} $\ $ \begin{remark} \label{rem:27} It is easy to see that the families of equations indicated in either (\ref{eqn:26b}) or (\ref{eqn:26c}) have unique solutions. Indeed, all that actually matters is that the index sets $ \mbox{Int} (n)$ and $NC(n)$ for the summations on the right-hand sides of these equations contain the partition, usually denoted as ``$1_n$'', of the set $\{ 1, \ldots , n \}$ into only one block. For instance in connection to (\ref{eqn:26b}): by separating the term indexed by $1_n$ on the right-hand side, this equation can be written as \begin{equation} \label{eqn:27a} \kappa_n (a_1, \ldots , a_n) = \varphi (a_1, \cdots , a_n) - \sum_{ \begin{array}{c} {\scriptstyle \pi \in NC(n),} \\ {\scriptstyle \pi \neq 1_n} \end{array} } \ \prod_{V \in \pi} \kappa_{|V|} ( \, (a_1, \ldots , a_n) \mid V \, ). \end{equation} Then (\ref{eqn:27a}) can be used as an explicit definition of the functional $\kappa_n$, under the assumption that explicit formulas for $\kappa_1, \ldots , \kappa_{n-1}$ have already been determined. A similar recursive argument holds in connection to solving the system of equations indicated in (\ref{eqn:26c}). It is in fact not difficult to write in a really explicit way some formulas giving the $\kappa_n$'s and the $\beta_n$'s in terms of $\varphi$. This is not needed in the present paper, so we only mention that the way to do it goes by using some standard elements of ``M\"obius inversion theory in a partially ordered set'' (as presented e.g. in Chapter 3 of the monograph \cite{St1997}). There also is a nice direct formula which expresses Boolean cumulants in terms of free cumulants, as follows. \end{remark} \begin{proposition} \label{prop:28} Let $( {\mathcal A} , \varphi )$ be a noncommutative probability space, and let $( \kappa_n : {\mathcal A} ^n \to {\mathbb C} )_{n=1}^{\infty}$ and $( \beta_n : {\mathcal A} ^n \to {\mathbb C} )_{n=1}^{\infty}$ be the free and respectively the Boolean cumulants associated to $( {\mathcal A} , \varphi )$. For every $n \in {\mathbb N} $ and $a_1, \ldots , a_n \in {\mathcal A} $ one has \begin{equation} \label{eqn:28a} \beta_n (a_1, \ldots , a_n) = \sum_{ \begin{array}{c} {\scriptstyle \pi \in NC(n) \ with} \\ {\scriptstyle unique \ outer \ block} \end{array} } \ \prod_{V \in \pi} \kappa_{|V|} ( \, (a_1, \ldots , a_n) \mid V \, ). \end{equation} \end{proposition} The proof of Proposition \ref{prop:28} can e.g. be obtained by an immediate re-phrasing of the argument proving Proposition 3.9 in \cite{BeNi2008}. \begin{notation-and-remark} \label{def:27b} Let $( {\mathcal A} , \varphi )$ be a noncommutative probability space, and let $( \kappa_n : {\mathcal A} ^n \to {\mathbb C} )_{n=1}^{\infty}$ and $( \beta_n : {\mathcal A} ^n \to {\mathbb C} )_{n=1}^{\infty}$ be the free and respectively the Boolean cumulants associated to $( {\mathcal A} , \varphi )$. Let $a \in {\mathcal A} $ be given. We will use the abbreviations \[ \beta_n (a) := \beta_n (a, \ldots , a) \mbox{ and } \kappa_n (a) := \kappa_n (a, \ldots , a), \ \ n \in {\mathbb N} . \] The power series \[ M_a (z) = \sum_{n=1}^{\infty} \varphi (a^n) z^n, \ R_a (z) = \sum_{n=1}^{\infty} \kappa_n (a) z^n \mbox{ and } \eta_a (z) = \sum_{n=1}^{\infty} \beta_n (a) z^n \] are called the {\em moment series}, the {\em R-transform} and respectively the {\em $\eta$-series} associated to $a$. In this paper, an important role is played by $\eta$-series. We note that, as an easy consequence of the formula (\ref{eqn:26c}) connecting moments to Boolean cumulants, one has a very simple relation between $\eta_a$ and $M_a$: \[ M_a (z) = \eta_a (z) / ( 1 - \eta_a (z) ), \mbox{ or equivalently, } \eta_a (z) = M_a (z) / ( 1 + M_a (z) ). \] \end{notation-and-remark} \vspace{10pt} \subsection{Cumulants with products as arguments.} $\ $ \noindent When working with cumulants of any kind, it is good to have an efficient formula for what happens when every argument of the cumulant is a product of elements of the underlying algebra. For free cumulants, this formula was put into evidence in \cite{KrSp2000}. We will need here the analogous fact for Boolean cumulants. In order to state this fact and to explain the analogy with \cite{KrSp2000}, we need to use the lattice structure (with respect to the partial order by reverse refinement) on $NC(n)$ and on $ \mbox{Int} (n)$, so we first do a brief review of this structure. $\ $ \begin{definition} \label{def:29} Let $n$ be a positive integer. \vspace{6pt} (1) On $NC(n)$ we consider the partial order by {\em reverse refinement}, where for $\pi, \rho \in NC(n)$ we put \begin{equation} \label{eqn:29a} ( \pi \leq \rho ) \ \stackrel{def}{\Longleftrightarrow} \ \Bigl( \mbox{ every block of $\rho$ is a union of blocks of $\pi$} \Bigr). \end{equation} The partially ordered set $( NC(n), \leq )$ turns out to be a lattice. That is, every $\pi_1 , \pi_2 \in NC(n)$ have a least common upper bound, denoted as $\pi_1 \vee \pi_2$, and have a greatest common lower bound, denoted as $\pi_1 \wedge \pi_2$. One refers to $\pi_1 \vee \pi_2$ and to $\pi_1 \wedge \pi_2$ as the {\em join} and respectively as the {\em meet} of $\pi_1$ and $\pi_2$ in $NC(n)$. We will use the notation $0_n$ for the partition of $\{ 1, \ldots , n \}$ into $n$ singleton blocks and the notation $1_n$ for the partition of $\{ 1, \ldots , n \}$ into one block. It is immediate that $0_n, 1_n \in NC(n)$ and that $0_n \leq \pi \leq 1_n$ for all $\pi \in NC(n)$. \vspace{6pt} (2) Consider the restriction of the partial order by reverse refinement from $NC(n)$ to $ \mbox{Int} (n)$. For $\pi_1, \pi_2 \in \mbox{Int} (n)$, the partitions $\pi_1 \vee \pi_2, \pi_1 \wedge \pi_2 \in NC(n)$ which were defined in (1) above turn out to still belong to $ \mbox{Int} (n)$. As a consequence, $( \mbox{Int} (n) , \leq )$ is a lattice as well, and for $\pi_1 , \pi_2 \in \mbox{Int} (n)$ there is no ambiguity in the meaning of what are $\pi_1 \vee \pi_2$ and $\pi_1 \wedge \pi_2$ (considering the join and meet of $\pi_1$ and $\pi_2$ in $ \mbox{Int} (n)$ gives the same result as when considering them in $NC(n)$). The special partitions $0_n$ and $1_n$ considered in (1) belong to $ \mbox{Int} (n)$, hence they also serve as minimum and maximum elements for the poset $( \mbox{Int} (n) , \leq )$. \end{definition} $\ $ The formula for Boolean cumulants with products as entries is then stated as follows. \begin{proposition} \label{prop:210} Let $( {\mathcal A} , \varphi )$ be a noncommutative probability space, and let $( \beta_n : {\mathcal A} ^n \to {\mathbb C} )_{n=1}^{\infty}$ be the family of Boolean cumulant functionals of $( {\mathcal A} , \varphi )$. Consider a Boolean cumulant of the form $\beta_m ( x_1, \ldots , x_m )$ where each of the elements $x_1, \ldots , x_m \in {\mathcal A} $ is written as a product: \[ x_1 = a_1 \cdots a_{i(1)}, x_2 = a_{i(1)+1} \cdots a_{i(2)}, \ldots , x_m = a_{i(m-1)+1} \cdots a_{i(m)}, \] where $1 \leq i(1) < i(2) < \cdots < i(m) =:n$ are some positive integers, and where $a_1, \ldots , a_n \in {\mathcal A} $. Then one has \begin{equation} \label{eqn:210a} \beta_m ( x_1, \ldots , x_m ) = \sum_{ \begin{array}{c} {\scriptstyle \pi \in \mbox{Int} (n) \ such} \\ {\scriptstyle that \ \pi \vee \sigma = 1_n} \end{array} \ } \ \prod_{V \in \pi} \beta_{|V|} ( ( a_1, \ldots , a_n) \mid V), \end{equation} with \begin{equation} \label{eqn:210b} \sigma := \{ \, \{ 1, \ldots , i(1) \} , \, \{ i(1) + 1, \ldots , i(2) \} , \ldots , \{ i(m-1) + 1, \ldots , i(m) \} \, \} \in \mbox{Int} (n). \end{equation} \end{proposition} \vspace{6pt} The proof of Proposition \ref{prop:210} is left as an exercise to the reader. A way to do it is by going over the development presented on pages 178-180 of \cite{NiSp2006} about free cumulants with products as arguments, and by replacing everywhere on those pages the occurrences of lattices of non-crossing partitions by occurrences of lattices of interval partitions. The statement of Proposition \ref{prop:210} should then emerge as the Boolean analogue of Theorem 11.12(2) of \cite{NiSp2006}. \vspace{6pt} \begin{remark} \label{rem:211} The lattice $ \mbox{Int} (n)$ is in fact a Boolean lattice. Indeed, one has a natural bijection which identifies $ \mbox{Int} (n)$ to the lattice of subsets of $\{ 1, \ldots , n-1 \}$, by sending a partition $\pi = \{ J_1, \ldots , J_k \} \in \mbox{Int} (n )$ to the set $\{ \max (J_1), \ldots , \max (J_k) \} \setminus \{ n \} \subseteq \{ 1, \ldots , n-1 \}$. By using this fact it is easy to see that, with $\sigma$ as defined in Equation (\ref{eqn:210b}), a partition $\pi \in \mbox{Int} (n)$ has \[ \bigl( \pi \vee \sigma = 1_n \bigr) \ \Leftrightarrow \Bigl( \begin{array}{c} \mbox{$i(p)$ and $i(p)+1$ belong to the same} \\ \mbox{block of $\pi$, for all $1 \leq p \leq m-1$} \end{array} \Bigr) . \] This allows for a somewhat more convenient re-phrasing of the join condition invoked on the right-hand side of Equation (\ref{eqn:210a}). \end{remark} $\ $ \section{The partial order $\ll$, VNRP, and the proof of Theorem \ref{thm:131}} \subsection{The partial order \boldmath{$\ll$} on $NC(n)$, and its upper ideals.} $\ $ \noindent In this paper we also make use of another partial order relation on $NC(n)$, coarser than reverse refinement, which is denoted as ``$\ll$''. The partial order $\ll$ has been used for some time in free probability (starting with \cite{BeNi2008}), in the description of relations between free and Boolean cumulants. $\ $ \begin{defrem} \label{def:31} {\em (The partial order ``$\ll$''.)} (1) For $\pi , \rho \in NC(n)$, we will write $\pi \ll \rho$ to mean that $\pi \leq \rho$ and that, in addition, for every block $W$ of $\rho$ there exists a block $V$ of $\pi$ such that $\min (W), \max (W) \in V$. (2) Since in this paper we have a lot of occurrences of the special case ``$\pi \ll 1_n$'', let us record the obvious fact that this simply amounts to requiring $\pi$ to have a unique outer block $W$, with $1,n \in W$. (3) It is immediate that $ \mbox{Int} (n)$ is precisely equal to the set of maximal elements of the poset $( NC(n), \ll )$. (4) A significant point about the partial order $\ll$ on $NC(n)$ is that we have a nice structure for its upper ideals, that is, for the sets of non-crossing partitions of the form \begin{equation} \label{eqn:31a} \{ \rho \in NC(n) \mid \rho \gg \pi \} , \ \mbox{ for a fixed $\pi \in NC(n)$.} \end{equation} This was noticed in Section 2 of \cite{BeNi2008}, but the discussion around the set (\ref{eqn:31a}) was mostly done in a proof (cf. proof of Proposition 2.13 in \cite{BeNi2008}), and it will be useful for our present purposes to spell that out in more detail. It turns out to be convenient to use a notion of ``projection map'' for blocks of a fixed $\pi \in NC(n)$, as introduced in the next definition. \end{defrem} $\ $ \begin{definition} \label{def:32} Let $n$ be in $ {\mathbb N} $ and let $\pi$ be a partition in $NC(n)$. A {\em block-projection} for $\pi$ is a map $\Phi : \pi \to \pi$ which has the following properties. (i) $\Phi$ is a projection map; that is, $\Phi \circ \Phi = \Phi$. (ii) If $A,B \in \pi$ and if $A \stackrel{\mathrm{nest}}{\leq} B$, then it follows that $\Phi (A) \stackrel{\mathrm{nest}}{\leq} \Phi (B)$. (iii) $A \stackrel{\mathrm{nest}}{\leq} \Phi (A)$ for all $A \in \pi$. \vspace{6pt} \noindent For such a $\Phi$ we will denote $ \mbox{Ran} ( \Phi ) := \{ B \in \pi \mid \exists \, A \in \pi \mbox{ such that } \Phi (A) = B \}$. Note that, due to the property (i) satisfied by $\Phi$, we can also write $ \mbox{Ran} ( \Phi ) = \{ B \in \pi \mid \Phi (B) = B \}$. \end{definition} \begin{remark} \label{rem:33} Let $n$ be in $ {\mathbb N} $ and let $\pi$ be a partition in $NC(n)$. (1) Let $\Phi : \pi \to \pi$ be a block-projection for $\pi$. Property (iii) satisfied by $\Phi$ implies that $ \mbox{Ran} ( \Phi )$ contains all the outer blocks of $\pi$. (2) Let $\Phi , \Psi : \pi \to \pi$ be block-projections for $\pi$, and suppose that $ \mbox{Ran} ( \Phi ) = \mbox{Ran} ( \Psi )$. Then $\Phi = \Psi$. Indeed, for every block $A \in \pi$ we can apply $\Psi$ to both sides of the relation $A \stackrel{\mathrm{nest}}{\leq} \Phi (A)$ to get that $\Psi (A) \stackrel{\mathrm{nest}}{\leq} \Psi ( \, \Phi (A) \, ) = \Phi (A)$, (where the latter equality holds because $\Phi (A) \in \mbox{Ran} ( \Phi ) = \mbox{Ran} ( \Psi )$, hence $\Phi (A)$ is fixed by $\Psi$). A symmetric argument gives that $\Phi (A) \stackrel{\mathrm{nest}}{\leq} \Psi (A)$, and it follows that $\Phi (A) = \Psi (A)$, as required. \end{remark} \begin{lemma} \label{lemma:34} Let $n$ be in $ {\mathbb N} $ and let $\pi$ be a partition in $NC(n)$. Let $ {\mathfrak M} $ be a subset of $\pi$ such that $ {\mathfrak M} $ contains all the outer blocks of $\pi$. Then there exists a block-projection $\Phi : \pi \to \pi$, uniquely determined, such that $ \mbox{Ran} ( \Phi ) = {\mathfrak M} $. \end{lemma} \begin{proof} Uniqueness of $\Phi$ follows from the preceding remark. In order to prove existence, we use the following prescription to define $\Phi$: \begin{equation} \label{eqn:34a} \left\{ \begin{array}{l} \mbox{if $A \in {\mathfrak M} $, then $\Phi (A) = A$;} \\ \mbox{if $A \in \pi \setminus {\mathfrak M} $, then $\Phi (A) = \Phi( \, \mathrm{Parent} _{\pi} (A) \, )$.} \\ \end{array} \right. \end{equation} The definition proposed via Equations (\ref{eqn:34a}) is consistent because if we start with any $A \in \pi$ and do iterations of the $ \mathrm{Parent} _{\pi}$ map on it, we will eventually have to find a block that belongs to $ {\mathfrak M} $. Or more precisely: if we start with $A \in \pi$ and we write explicitly $\{ B \in \pi \mid B \stackrel{\mathrm{nest}}{\geq} A \} = \{ B_1, \ldots , B_k \}$ in the way indicated in Remark \ref{def:22}(4), then Equations (\ref{eqn:34a}) define $\Phi (A) = B_j$ with $j := \min \Bigl\{ i \in \{ 1, \ldots , k \} \mid B_i \in {\mathfrak M} \Bigr\}$. \end{proof} \begin{remark} \label{rem:35} One has a natural construction of block-projection map $\Phi : \pi \to \pi$ which arises whenever we are given two partitions $\pi , \rho \in NC(n)$ such that $\pi \ll \rho$. Recall that, in this situation, for every block $X \in \rho$ there exists a block $B \in \pi$ such that \begin{equation} \label{eqn:35a} B \subseteq X \mbox{ and } \min (B) = \min (X), \ \max (B) = \max (X). \end{equation} We then define $\Phi : \pi \to \pi$ as follows: for every $A \in \pi$ we consider the (unique) block $X \in \rho$ such that $X \supseteq A$, and then we define $\Phi (A) := B$, where $B$ is as in (\ref{eqn:35a}). It is easy to check that the map $\Phi : \pi \to \pi$ defined in this way fulfills the conditions (i), (ii) and (iii) from Definition \ref{def:32}, hence is indeed a block-projection map for $\pi$. Let us record that, in the terminology introduced in \cite{BeNi2008}, a block $B$ as in (\ref{eqn:35a}) is said to be a {\em $\rho$-special} block of $\pi$. The block-projection $\Phi$ constructed above is characterized by the fact that $ \mbox{Ran} ( \Phi )$ is precisely the set of all $\rho$-special blocks of $\pi$. \end{remark} We now come to the main point concerning the set of partitions indicated in (\ref{eqn:31a}), namely that it is actually ``parametrized '' by the set of block-projection maps for $\pi$, where the parametrization is just the inverse of the natural construction indicated in Remark \ref{rem:35}. The formal statement of how this works is recorded in the next proposition. \begin{proposition} \label{prop:36} Let $n$ be in $ {\mathbb N} $, let $\pi$ be a partition in $NC(n)$, and let $\Phi : \pi \to \pi$ be a block-projection map. Then there exists $\rho \in NC(n)$, uniquely determined, such that $\rho \gg \pi$ and such that $\Phi$ is obtained from $\pi$ and $\rho$ by using the recipe described in Remark \ref{rem:35}. If we list the range of $\Phi$ as $ \mbox{Ran} ( \Phi ) =: \{ B_1, \ldots, B_p \}$, then the partition $\rho$ can be described explicitly as $\rho = \{ X_1, \ldots , X_p \}$, where \begin{equation} \label{eqn:36a} X_j = \cup_{A \in \Phi^{-1} (B_j)} \ A, \ \ 1 \leq j \leq p. \end{equation} \end{proposition} The proof of Proposition \ref{prop:36} amounts essentially to reproducing the proof of Proposition 2.13 from \cite{BeNi2008}, but where we work with $\Phi$ itself rather than writing all the arguments in terms of the set of blocks $ \mbox{Ran} ( \Phi )$. We note that in view of Lemma \ref{lemma:34}, the parametrization of $\{ \rho \in NC(n) \mid \rho \gg \pi \}$ in terms of block-projections for $\pi$ can also be viewed as a parametrization in terms of subsets of $\pi$ which contain all the outer blocks -- this is, actually, what was observed in Proposition 2.13 of \cite{BeNi2008} and in the proof of that proposition. We conclude the discussion about $\ll$ with an observation that will be needed in the next subsection. This observation does not depend on Proposition \ref{prop:36}, it is just a direct consequence of how the notion of ``$\rho$-special block of $\pi$'', is defined. It goes as follows. \begin{lemma} \label{lemma:37} Let $n$ be in $ {\mathbb N} $, and let $\pi , \rho \in NC(n)$ be such that $\pi \ll \rho$. Let $A$ be a $\rho$-special block of $\pi$ which is not outer, and let $B = \mathrm{Parent} _{\pi} (A) \in \pi$. Let $X,Y$ be the blocks of $\rho$ determined by the requirements that $X \supseteq A$ and $Y \supseteq B$. Then $Y = \mathrm{Parent} _{\rho} (X)$. \hfill $\square$ \end{lemma} The proof of Lemma \ref{lemma:37} is done by an elementary argument, directly from the definitions of the notions involved. (One must keep in mind, of course, that the hypothesis ``$A$ is $\rho$-special'' means, by definition, that $\min (A) = \min (X)$ and $\max (A) = \max (X)$.) \vspace{10pt} \subsection{VNRP, and the proof of Theorem \ref{thm:131}.} $\ $ \noindent Throughout this whole subsection we fix the data used in the statement of Theorem \ref{thm:131}. That is, we fix two positive integers $m$ and $s$, and a function $c : \{ 1, \ldots , m \} \to \{ 1, \ldots , s \}$. (We think of $c$ as of a ``colouring of $\{ 1, \ldots , m \}$ in $s$ colours''.) We will denote by $NC(m;c)$ the subset of $NC(m)$ defined by \[ NC(m;c) := \{ \sigma \in NC(m) \mid \mbox{ $c$ is constant on every block of $\sigma$} \} . \] For $\sigma \in NC(m;c)$ and $A \in \sigma$, we will use the notation $c(A)$ for the common value $c(a) \in \{ 1, \ldots , s \}$ taken by $c$ on all $a \in A$. Note that on $NC(m;c)$ we have two partial order relations ``$\leq$'' (reverse refinement) and ``$\ll$'', induced from $NC(n)$. The definition of VNRP goes as follows. \begin{definition} \label{def:41} A partition $\sigma \in NC(m;c)$ will be said to have the {\em vertical no-repeat property} with respect to $c$ when the following happens: for every inner block $A \in \sigma$, one has \[ c( \, \mathrm{Parent} _{\sigma} (A) \, ) \neq c(A). \] As already done in the Introduction, we will refer to the vertical no-repeat property by using the acronym ``VNRP''. (Note that if the number of colours $s$ would happen to be $s=2$, VNRP could also go under the name of ``vertical alternance property''.) \end{definition} $\ $ Our goal for the section is to prove Theorem \ref{thm:131} stated in the Introduction. In order to do so, we start with an adjustment of Proposition \ref{prop:36} to the present framework which uses coloured partitions from $NC(m;c)$. \begin{proposition} \label{prop:Y3} Let $\sigma$ be a partition in $NC(m;c)$. Let $\Phi : \sigma \to \sigma$ be a block-projection map, and let $\rho \in NC(m)$ be the partition with $\rho \gg \sigma$ which is parametrized by $\Phi$ in Proposition \ref{prop:36}. We have that: \begin{equation} \label{eqn:Y3a} \Bigl( \rho \in NC(m;c) \Bigr) \ \Leftrightarrow \ \Bigl( c( \Phi (A) ) = c(A), \ \forall \, A \in \sigma \Bigr) . \end{equation} \end{proposition} \begin{proof} ``$\Rightarrow$'' From the concrete description of $\rho$ provided by Equation (\ref{eqn:36a}) of Proposition \ref{prop:36}, it is clear that for every $A \in \sigma$, the blocks $A$ and $\Phi (A)$ are contained in the same block $X$ of $\rho$. The latter fact implies in particular that $c(A) = c(X) = c( \Phi (A) )$. The condition on $\Phi$ listed on the right-hand side of (\ref{eqn:Y3a}) thus follows. \vspace{6pt} ``$\Leftarrow$'' Let $X$ be a block of $\rho$, and consider the block $B$ of $\sigma$ such that $\min (X), \max (X) \in B$. The explicit description of $\rho$ provided by Equation (\ref{eqn:36a}) in Proposition \ref{prop:36} tells us that \begin{equation} \label{eqn:Y3b} X = \cup_{A \in \Phi^{-1} (B)} \ A. \end{equation} Denoting $c(B) = s_o \in \{ 1, \ldots , s \}$, we see that $c(A) = s_o$ for every $A$ appearing in the union from (\ref{eqn:Y3b}) -- indeed, one has $\Phi (A) = B$, hence the hypothesis $c(A) = c( \Phi (A) )$ gives $c(A) = c(B) = s_o$. This makes it clear that $c$ is constantly equal to $s_o$ on $X$, and completes the verification that $\rho \in NC(n;c)$. \end{proof} \begin{corollary} \label{cor:Y4} Let $\sigma, \rho$ be partitions in $NC(m;c)$ such that $\sigma \ll \rho$. Let $\Phi : \sigma \to \sigma$ be the block-projection map which corresponds to $\rho$ in Remark \ref{rem:35}. One has \begin{equation} \label{eqn:Y4a} \mbox{Ran} ( \Phi ) \supseteq \{ A \in \sigma \mid A \mbox{ is inner and } c( \mathrm{Parent} _{\sigma} (A) ) \neq c(A) \} . \end{equation} \end{corollary} \begin{proof} We prove the reverse inclusion for the complements of the sets indicated in in (\ref{eqn:Y4a}). That is: we pick $A_o \in \sigma \setminus \mbox{Ran} ( \Phi )$, and we prove that $A_o$ belongs to the complement of the set on the right-hand side of (\ref{eqn:Y4a}). The condition $A_o \not\in \mbox{Ran} ( \Phi )$ implies in particular that $A_o$ is inner, so what we have to prove is the equality $c( \mathrm{Parent} _{\sigma} (A_o) ) = c(A_o)$. We denote $ \mathrm{Parent} _{\sigma} (A_o) = A_1$. It is easy to see (directly from the properties of $\Phi$ listed in Definition \ref{def:32}) that the assumption $A_o \not\in \mbox{Ran} ( \Phi )$ (which is equivalent to $\Phi ( A_o ) \neq A_o$, hence to $A_o \stackrel{\mathrm{nest}}{<} \Phi (A_o)$) entails the equality $\Phi (A_o) = \Phi (A_1)$. On the other hand, the assumption that $\rho \in NC(m;c)$ entails, via Proposition \ref{prop:Y3}, the equalities $c(A_o) = c( \Phi (A_o) )$ and $c(A_1) = c( \Phi (A_1) )$. By putting all these things together we find that $c(A_o) = c( \Phi (A_o) )$ = $c( \Phi (A_1) ) = c(A_1)$. Hence $c( \mathrm{Parent} _{\sigma} (A_o) ) = c(A_o)$, as required. \end{proof} $\ $ The next proposition addresses the uniqueness part in the statement of Theorem \ref{thm:131}, by giving an explicit description of the set of $\tau$-special blocks of $\sigma$, for the partition $\tau \gg \sigma $ which is needed in the conclusion of the theorem. \begin{proposition} \label{prop:Y5} Let $\sigma$ be a partition in $NC(m;c)$. Suppose that $\tau \in NC(m;c)$ has VNRP, and is such that $\tau \gg \sigma$. Then the set of $\tau$-special blocks of $\sigma$ is equal to \begin{equation} \label{eqn:Y5a} \{ A \in \sigma \mid A \mbox{ is outer} \} \cup \{ A \in \sigma \mid A \mbox{ is inner and } c( \mathrm{Parent} _{\pi} (A) ) \neq c(A) \} . \end{equation} \end{proposition} \begin{proof} Let $\Phi : \sigma \to \sigma$ be the block-projection map which parametrizes $\tau$ in the way described in Proposition \ref{prop:36}. The set of $\tau$-special blocks of $\sigma$ is thus the same as $ \mbox{Ran} ( \Phi )$. Corollary \ref{cor:Y4} then assures us that the set of $\tau$-special blocks of $\sigma$ contains all blocks $A \in \sigma$ such that $A$ is inner and $c ( \mathrm{Parent} _{\pi} (A) ) \neq c(A)$. Since $ \mbox{Ran} ( \Phi )$ is also sure to contain all the outer blocks of $\sigma$ (Remark \ref{rem:33}(1)), it follows that $ \mbox{Ran} ( \Phi )$ must contain the set of blocks indicated in formula (\ref{eqn:Y5a}). Let us assume, for contradiction, that $ \mbox{Ran} ( \Phi )$ is strictly larger than the set from (\ref{eqn:Y5a}), i.e. that it contains a block $V \in \sigma$ which is inner and has $c( \mathrm{Parent} _{\sigma} (V) ) = c(V)$. This $V$ is a $\tau$-special block of $\sigma$ (since it is in $ \mbox{Ran} ( \Phi )$), hence we can use Lemma \ref{lemma:37} with $A := V$ and $B := \mathrm{Parent} _{\sigma} (V)$. Denoting by $X,Y$ the blocks of $\tau$ which contain $A$ and $B$, respectively, we get from Lemma \ref{lemma:37} that $ \mathrm{Parent} _{\tau} (X) = Y$. Now, we must have $c(A) = c(X)$ and $c(B) = c(Y)$ (since $A \subseteq X$ and $B \subseteq Y$), so from $c(A) = c(B)$ it follows that $c(X) = c(Y)$. This contradicts the VNRP of $\tau$, and concludes the proof. \end{proof} For the existence part in Theorem \ref{thm:131}, one could go by showing directly that the ``candidate for $\tau$'' suggested by Proposition \ref{prop:Y5} does indeed the required job. We leave this approach as an exercise to the interested reader, and we just invoke here a simple maximality argument. \begin{lemma} \label{lem:46} Consider the partial order given by $\ll$ on $NC(m;c)$. Every maximal element of $( NC(m;c) , \ll )$ has the VNRP property. \end{lemma} \begin{proof} We prove the contrapositive: if $\pi \in NC(m;c)$ does not have VNRP, then it cannot be a maximal element with respect to $\ll$. Indeed, let us pick a $\pi \in NC(m;c)$ without VNRP, and let $V, V'$ be blocks of $\pi$ such that $V' = \mbox{Parent}_{\pi} (V)$ but $V, V'$ have the same colour. Let $\rho$ be the partition of $\{1, \ldots , n \}$ which is obtained out of $\pi$ by joining together the blocks $V$ and $V'$. Then $\rho \in NC(m;c)$ and $\pi \ll \rho$ (this is a slight modification of Lemma 6.4.3 from \cite{BeNi2008}), showing in particular that $\pi$ is not maximal with respect to $\ll$. \end{proof} $\ $ \begin{remark} \label{rem:46} We leave it as an exercise to the reader to check that the converse of Lemma \ref{lem:46} is true as well: the maximal elements of $( NC(m;c), \ll )$ are precisely the partitions which have VNRP with respect to the colouring $c$. \end{remark} $\ $ \begin{ad-hoc} {\bf Proof of Theorem \ref{thm:131}.} We fix a partition $\sigma \in NC(m;c)$ and we have to prove that there exists a $\tau \in NC(m;c)$, uniquely determined, such that $\sigma \ll \tau$ and such that $\tau$ has VNRP with respect to $c$. And indeed: the uniqueness of $\tau$ with the required properties follows from Proposition \ref{prop:Y5}. On the other hand, the existence of $\tau$ follows from Lemma \ref{lem:46} and the fact that $\sigma$ must have a majorant which is maximal with respect to $\ll$. \hfill $\square$ \end{ad-hoc} $\ $ \section{Free independence in terms of Boolean cumulants, via VNRP} \label{sec:Lemmas} Based on Theorem \ref{thm:131}, one gets the characterization of free independence in terms of Boolean cumulants which was announced in Theorem \ref{thm:141} from the Introduction. $\ $ \begin{ad-hoc} {\bf Proof of Theorem \ref{thm:141}.} Recall that in this theorem we are given a noncommutative probability space $( {\mathcal A} , \varphi )$ and some unital subalgebras $ {\mathcal A} _1,\ldots, {\mathcal A} _s\subseteq {\mathcal A} $, and we have to prove the equivalence of two statements $(1)$ and $(2)$ concerning $ {\mathcal A} _1, \ldots, {\mathcal A} _s$. Throughout the proof we use the notation $( \beta_n : {\mathcal A} ^n \to {\mathbb C} )_{n=1}^{\infty}$ and respectively $( \kappa_n : {\mathcal A} ^n \to {\mathbb C} )_{n=1}^{\infty}$ for the families of Boolean and respectively free cumulant functionals of $( {\mathcal A} , \varphi )$. \vspace{6pt} {\em Proof that $(1) \Rightarrow (2)$.} Here we know that $ {\mathcal A} _1,\ldots, {\mathcal A} _s$ are freely independent with respect to $\varphi$. We consider an $n \in {\mathbb N} $, a colouring $c : \{ 1, \ldots , n \} \to \{ 1, \ldots , s \}$, and some elements $a_1 \in {\mathcal A} _{c(1)}, \ldots a_n \in {\mathcal A} _{c(n)}$, and we have to prove that \begin{equation} \label{eqn:51a} \beta_n (a_1, \ldots , a_n) = \sum_{ \begin{array}{c} {\scriptstyle \pi \in NC(n;c),} \ {\scriptstyle \pi \ll 1_n ,} \\ {\scriptstyle \pi \ has \ VNRP} \end{array} } \ \prod_{V \in \pi} \beta_{|V|} ( (a_1, \ldots , a_n) \mid V ). \end{equation} To that end, we write \begin{align*} \beta_n (a_1, \ldots , a_n) & = \sum_{\pi \in NC(n), \ \pi \ll 1_n} \ \prod_{V \in \pi} \kappa_{|V|} ( (a_1, \ldots , a_n) \mid V) \\ & = \sum_{ \pi \in NC(n,c), \ \pi \ll 1_n} \ \prod_{V \in \pi} \kappa_{|V|} ( (a_1, \ldots , a_n) \mid V), \end{align*} where at the first equality sign we used the formula expressing Boolean cumulants in terms of free cumulants, and at the second equality sign we used the vanishing of mixed free cumulants with entries from the free subalgebras $ {\mathcal A} _1, \ldots, {\mathcal A} _s$. We now invoke Theorem \ref{thm:131}, and group the partitions ``$\pi \in NC(n,c), \ \pi \ll 1_n$'' which index the latter sum according to the unique $\rho \in NC(n;c)$ such that $\pi \ll \rho$ and $\rho$ has VNRP with respect to $c$. When doing so, we continue the above equalities with \[ = \sum_{ \begin{array}{c} {\scriptstyle \rho \in NC(n,c),\ \rho \ll 1_n ,} \\ {\scriptstyle \rho \ has \ VNRP} \end{array} } \, \Bigl( \ \sum_{\pi \ll \rho} \prod_{V \in \pi} \kappa_{|V|} ((a_1, \ldots , a_n) \mid V) \, \Bigr). \] Finally, in the latter double sum we note that the inside summation comes to \noindent $\prod_{V \in \rho} \beta_{|V|} ((a_1, \ldots , a_n) \mid V)$ (again due to how Boolean cumulants are expressed in terms of free cumulants). This leads precisely to the formula for $\beta_n (a_1, \ldots , a_n)$ that was stated in Equation (\ref{eqn:51a}). \vspace{6pt} \noindent {\em Proof that $(2) \Rightarrow (1)$.} In order to prove that $ {\mathcal A} _1,\ldots, {\mathcal A} _s$ are free, we consider the free product $\left(\widetilde{ {\mathcal A} },\widetilde{\varphi}\right)$ = $\ast\left( {\mathcal A} _i,\varphi|_{ {\mathcal A} _i}\right)_{i=1,\ldots,s}$. Then for any $a_1,\ldots,a_n$ such that $a_k\in {\mathcal A} _{i(k)}$, formula \eqref{eqn:51a} holds for Boolean cumulants related with $\varphi$, denoted by $\beta_{n}^{\varphi}$ by assumption. On the other hand \eqref{eqn:51a} holds also for $\beta_{n}^{\widetilde{\varphi}}$, i.e. Boolean cumulants related with $\widetilde{\varphi}$, since $a_1,\ldots,a_n$ are free wrt $\widetilde{\varphi}$ hence the implication $(1) \Rightarrow (2)$, proved above, can be applied. Since for any $a_1,\ldots,a_n$ such that $a_k\in {\mathcal A} _{i(k)}$ we have $\beta_{n}^{\varphi}(a_1,\ldots,a_n)= \beta_{n}^{\widetilde{\varphi}}(a_1,\ldots,a_n)$, then also all joint moments with respect to $\varphi$ and $\widetilde{\varphi}$ coincide. Since $ {\mathcal A} _1,\ldots, {\mathcal A} _s$ are free with respect to $\widetilde{\varphi}$, we get that $ {\mathcal A} _1,\ldots, {\mathcal A} _s$ are free with respect to $\varphi$. \hfill $\square$ \end{ad-hoc} \begin{example} \label{ex:42} For the sake of clarity, we give a concrete example of how the formula (\ref{eqn:51a}) works. Let $ {\mathcal A} _1, {\mathcal A} _2$ be freely independent subalgebras of $ {\mathcal A} $. Suppose we pick elements $a, a', a'' \in {\mathcal A} _1$ and $b, b', b'' \in {\mathcal A} _2$, and we are interested in the Boolean cumulant $\beta_6 ( a,b, a', b', b'', a'')$. \begin{minipage}{1\linewidth} \begin{center} \begin{tikzpicture} \draw[red,thick] (-6,0) -- (-6,0.25); \draw[blue,thick] (-4,0) -- (-4,0.25); \draw[red,thick] (-8,0) to[out=90,in=90] node [sloped,above] {} (-3,0); \draw[blue,thick] (-7,0) to[out=90,in=90] node [sloped,above] {} (-5,0); \node[below,red] (1) at (-8,0) {$a$}; \node[below,blue] (2) at (-7,0) {$b$}; \node[below,red] (3) at (-6,0) {$a'$}; \node[below,blue] (4) at (-5,0) {$b'$}; \node[below,blue] (5) at (-4,0) {$b''$}; \node[below,red] (6) at (-3,0) {$a''$}; \draw[red,thick] (2,0) -- (2,0.25); \draw[red,thick] (0,0) to[out=90,in=90] node [sloped,above] {} (5,0); \draw[blue,thick] (1,0) to[out=90,in=90] node [sloped,above] {} (3,0); \draw[blue,thick] (3,0) to[out=90,in=90] node [sloped,above] {} (4,0); \node[below,red] (7) at (0,0) {$a$}; \node[below,blue] (8) at (1,0) {$b$}; \node[below,red] (9) at (2,0) {$a'$}; \node[below,blue] (10) at (3,0) {$b'$}; \node[below,blue] (11) at (4,0) {$b''$}; \node[below,red] (12) at (5,0) {$a''$}; \end{tikzpicture} \end{center} \end{minipage} \begin{minipage}{1\linewidth} \begin{center} \begin{tikzpicture} \draw[blue,thick] (-4,0) -- (-4,0.25); \draw[blue,thick] (-7,0) -- (-7,0.25); \draw[blue,thick] (-5,0) -- (-5,0.25); \draw[red,thick] (-8,0) to[out=90,in=90] node [sloped,above] {} (-6,0); \draw[red,thick] (-6,0) to[out=90,in=90] node [sloped,above] {} (-3,0); \node[below,red] (1) at (-8,0) {$a$}; \node[below,blue] (2) at (-7,0) {$b$}; \node[below,red] (3) at (-6,0) {$a'$}; \node[below,blue] (4) at (-5,0) {$b'$}; \node[below,blue] (5) at (-4,0) {$b''$}; \node[below,red] (6) at (-3,0) {$a''$}; \draw[blue,thick] (1,0) -- (1,0.25); \draw[red,thick] (0,0) to[out=90,in=90] node [sloped,above] {} (2,0); \draw[red,thick] (2,0) to[out=90,in=90] node [sloped,above] {} (5,0); \draw[blue,thick] (3,0) to[out=90,in=90] node [sloped,above] {} (4,0); \node[below,red] (7) at (0,0) {$a$}; \node[below,blue] (8) at (1,0) {$b$}; \node[below,red] (9) at (2,0) {$a'$}; \node[below,blue] (10) at (3,0) {$b'$}; \node[below,blue] (11) at (4,0) {$b''$}; \node[below,red] (12) at (5,0) {$a''$}; \end{tikzpicture} \end{center} \end{minipage} \begin{center} {\bf Figure 3(a).} {\em Some coloured partitions in $NC(6)$, with VNRP.} \end{center} $\ $ \noindent We get \begin{equation} \label{eqn:42a} \beta_6 (a,b,a',b',b'',a'') = \begin{array}[t]{l} \beta_3 (a,a',a'') \beta_1 (b) \beta_1 (b') \beta_1 (b'') + \beta_3 (a,a',a'') \beta_1 (b) \beta_2 (b', b'') \\ + \beta_2 (a,a'') \beta_1 (a') \beta_3 (b,b',b'') + \beta_2 (a,a'') \beta_1 (a') \beta_2 (b,b') \beta_1 (b'') , \end{array} \end{equation} where the four terms on the right-hand side of the above equation correspond to the four partitions in $NC(6)$ that are listed in Figure 3(a). It is instructive to note that one also has two partitions in $NC(6)$, shown in Figure 3(b), which satisfy the colouring condition (i.e. they separate $a$'s from $b$'s in the tuple $(a,b,a',b',b'',a'')$) and also satisfy the requirement of having a unique outer block), but don't contribute to the sum on the right-hand side of (\ref{eqn:42a}) because they don't have VNRP. \end{example} \begin{minipage}{1\linewidth} \begin{center} \begin{tikzpicture} \draw[red,thick] (-8,0) to[out=90,in=90] node [sloped,above] {} (-3,0); \draw[blue,thick] (-7,0) to[out=90,in=90] node [sloped,above] {} (-4,0); \draw[red,thick] (-6,0) -- (-6,0.25); \draw[blue,thick] (-5,0) -- (-5,0.25); \node[below,red] (1) at (-8,0) {$a$}; \node[below,blue] (2) at (-7,0) {$b$}; \node[below,red] (3) at (-6,0) {$a'$}; \node[below,blue] (4) at (-5,0) {$b'$}; \node[below,blue] (5) at (-4,0) {$b''$}; \node[below,red] (6) at (-3,0) {$a''$}; \draw[red,thick] (0,0) to[out=90,in=90] node [sloped,above] {} (5,0); \draw[blue,thick] (3,0) to[out=90,in=90] node [sloped,above] {} (4,0); \draw[blue,thick] (1,0) -- (1,0.25); \draw[red,thick] (2,0) -- (2,0.25); \node[below,red] (7) at (0,0) {$a$}; \node[below,blue] (8) at (1,0) {$b$}; \node[below,red] (9) at (2,0) {$a'$}; \node[below,blue] (10) at (3,0) {$b'$}; \node[below,blue] (11) at (4,0) {$b''$}; \node[below,red] (12) at (5,0) {$a''$}; \end{tikzpicture} \end{center} \end{minipage} \begin{center} {\bf Figure 3(b).} {\em Some coloured partitions in $NC(6)$, without VNRP.} \end{center} $\ $ \begin{remark} \label{rem:52} It is useful to note a special situation when we are sure to get ``vanishing of mixed Boolean cumulants with free arguments'': consider the setting of Theorem \ref{thm:141}, and let $c : \{ 1, \ldots , n \} \to \{ 1, \ldots , s \}$ be a colouring such that $c(1) \neq c(n)$. Then for every $a_1 \in {\mathcal A} _{c(1)}, \ldots a_n \in {\mathcal A} _{c(n)}$, one has that $\beta_n (a_1, \ldots , a_n) = 0$. Indeed, in this special case the index set for the summation on the right-hand side of Equation (\ref{eqn:51a}) is the empty set. \end{remark} $\ $ In the remaining part of this section, we record some easy consequences of Theorem \ref{thm:141}. First, we note that from Equation (\ref{eqn:141a}) one can derive a formula for moments -- this is precisely the $ {\mathbb C} $-valued case of the moment formula found in Proposition 4.30 of \cite{JeLi2019}, and is stated as follows. \begin{corollary} \label{cor:53} For every $n \in {\mathbb N} $, every colouring $c : \{ 1, \ldots , n \} \to \{ 1, \ldots , s \}$, and every $a_1 \in {\mathcal A} _{c(1)}, \ldots a_n \in {\mathcal A} _{c(n)}$, one has \begin{equation} \label{eqn:53a} \varphi (a_1 \cdots a_n) = \sum_{ \begin{array}{c} {\scriptstyle \pi \in NC(n;c)} \\ {\scriptstyle with \ VNRP} \end{array} } \ \prod_{V \in \pi} \ \beta_{|V|} ( (a_1, \ldots , a_n) \mid V ). \end{equation} \end{corollary} \begin{proof} Perform an additional summation over interval partitions on both sides of Equation (\ref{eqn:141a}), and use the formula which expresses moments in terms of Boolean cumulants. \end{proof} A basic fact concerning free cumulants is that $\kappa_n (a_1, \ldots , a_n) = 0$ whenever $n \geq 2$ and there exists an $m \in \{ 1, \ldots , n \}$ such that $a_m$ is a scalar multiple of the unit. The next corollary gives the analogue of this fact when one uses Boolean cumulants. \begin{corollary} \label{cor:54} Let $( {\mathcal A} , \varphi )$ be a noncommutative probability space and let $( \beta_n : {\mathcal A} ^n \to {\mathbb C} )_{n=1}^{\infty}$ be the family of Boolean cumulant functionals associated to it. Let $a_1, \ldots , a_n$ be elements of $ {\mathcal A} $, where $n \geq 2$, and suppose we are given an index $m \in \{ 1, \ldots , n \}$ for which it is known that $a_m = 1_{{ }_{\cA}} $. Then \begin{equation} \label{eqn:54a} \beta_n ( a_1, \ldots , a_n) = \left\{ \begin{array}{ll} 0, & \mbox{ if $m=1$ or $m=n$,} \\ \beta_{n-1} (a_1, \ldots , a_{m-1}, a_{m+1}, \ldots , a_n), & \mbox{ if $1 < m < n$.} \end{array} \right. \end{equation} \end{corollary} \begin{proof} Consider the unital subalgebras $ {\mathcal A} _1, {\mathcal A} _2$ of $ {\mathcal A} $ defined by $ {\mathcal A} _1 = {\mathcal A} $ and $ {\mathcal A} _2 = {\mathbb C} 1_{{ }_{\cA}} $, and consider the colouring $c : \{ 1, \ldots , n \} \to \{ 1,2 \}$ defined by \[ c(i) = \left\{ \begin{array}{ll} 1, & \mbox{ if $i \in \{ 1, \ldots , n \} \setminus \{ m \}$,} \\ 2, & \mbox{ if $i = m$.} \end{array} \right. \] We then have $a_i \in {\mathcal A} _{c(i)}$ for all $1 \leq i \leq n$, with $ {\mathcal A} _1$ being freely independent of $ {\mathcal A} _2$, hence Theorem \ref{thm:141} applies to this situation and gives us a formula for $\beta_n (a_1, \ldots , a_n)$. If $m=1$ or $m=n$, then $\beta_n (a_1, \ldots , a_n) = 0$ by Remark \ref{rem:52}. If $1 < m < n$, then an immediate inspection shows that the only partition $\pi \in NC(n)$ with unique outer block and with VNRP is \[ \pi = \bigl\{ \, \{ 1, \ldots , m-1, m+1, \ldots , n \}, \, \{ m \} \, \bigr\} . \] Thus the sum on the right-hand side of (\ref{eqn:51a}) has in this case only one term, which is as indicated in (\ref{eqn:54a}) above (since $\beta_1 (a_m) = \beta_1 ( 1_{{ }_{\cA}} ) = 1$). \end{proof} $\ $ From Theorem \ref{thm:141} one can also get an explicit description of the Boolean cumulants of the sum of two free elements, as follows. \begin{proposition} \label{prop:55} Let $( {\mathcal A} , \varphi )$ be a noncommutative probability space and let $a,b \in {\mathcal A} $ be such that $a$ is freely independent from $b$. Consider the sequences of Boolean cumulants $( \beta_n (a) )_{n=1}^{\infty}$ and $( \beta_n (b) )_{n=1}^{\infty}$ for $a$ and for $b$, respectively. Then for every $n \geq 1$, the $n$-th Boolean cumulant of $a + b$ is \begin{equation} \label{eqn:55a} \beta_n (a+b) = \sum_{ \begin{array}{c} {\scriptstyle \pi \in NC(n),} \\ {\scriptstyle \pi \ll 1_n} \end{array} } \ \Bigl( \prod_{ \begin{array}{c} {\scriptstyle U \in \pi, \ with} \\ {\scriptstyle \mathrm{depth} _{\pi} (U) \ even} \end{array} } \ \beta_{|U|} (a)\Bigr) \cdot \Bigl( \prod_{ \begin{array}{c} {\scriptstyle V \in \pi, \ with} \\ {\scriptstyle \mathrm{depth} _{\pi} (V) \ odd} \end{array} } \ \beta_{|V|} (b) \Bigr) \end{equation} \[ + \ \sum_{ \begin{array}{c} {\scriptstyle \pi \in NC(n),} \\ {\scriptstyle \pi \ll 1_n} \end{array} } \ \Bigl( \prod_{ \begin{array}{c} {\scriptstyle U \in \pi, \ with} \\ {\scriptstyle \mathrm{depth} _{\pi} (U) \ even} \end{array} } \ \beta_{|U|} (b) \Bigr) \cdot \Bigl( \prod_{ \begin{array}{c} {\scriptstyle V \in \pi, \ with} \\ {\scriptstyle \mathrm{depth} _{\pi} (V) \ odd} \end{array} } \ \beta_{|V|} (b) \Bigr) . \] \end{proposition} \begin{proof} We expand $\beta_n (a+b, \ldots , a+b)$ as a sum of $2^n$ terms by multilinearity, and then for each of the $2^n$ terms we use the formula for Boolean cumulants with free entries. This takes us to an expression of the form stipulated on the right-hand side of Equation (\ref{eqn:55a}) -- it is a large sum where every term of the sum is a product of Boolean cumulants of $a$ and of $b$. We next make the following observation: for every $\pi \in NC(n)$ such that $\pi \ll 1_n$ there exist precisely two ways of colouring the blocks of $\pi$ in the colours ``$a$'' and ``$b$'' such that VNRP holds; indeed, once we decide what is the colour of the unique outer block of $\pi$, everything else is determined. By using this observation, we sort out the terms of the large sum indicated in the preceding paragraph, and we organize these terms into two separate sums, arriving precisely to the formula announced in (\ref{eqn:55a}). \end{proof} \begin{remark} \label{rem:56} The statement of Proposition \ref{prop:55} simplifies quite a bit when the two elements $a$ and $b$ have the same distribution. In this case we only have one sequence $( \lambda_n )_{n=1}^{\infty}$, giving the Boolean cumulants for both $a$ and $b$, and Equation (\ref{eqn:55a}) becomes \begin{equation} \label{eqn:56a} \beta_n (a+b) = 2 \cdot \sum_{ \begin{array}{c} {\scriptstyle \pi \in NC(n),} \\ {\scriptstyle \pi \ll 1_n} \end{array} } \ \Bigl( \prod_{ V \in \pi } \lambda_{|V|} \Bigr) . \end{equation} We note that Equation (\ref{eqn:56a}) can alternatively be obtained by using some known facts about the Boolean Bercovici-Pata bijection introduced in \cite{BePa-Bi1999}. Indeed, let us apply this Bercovici-Pata bijection to the common distribution $\mu$ of $a$ and of $b$, and let us denote the resulting distribution by $\nu$. (Here we view both $\mu$ and $\nu$ as linear functionals on $ {\mathbb C} [X]$, with $\nu$ being defined in terms of $\mu$ via the requirement that its free cumulants \footnote{ The free cumulant $\kappa_n ( \nu )$ is defined as the free cumulant $\kappa_n (X, \ldots , X)$ in the noncommutative probability space $( {\mathbb C} [X], \nu )$, while the Boolean cumulant $\beta_n ( \mu )$ is defined as the Boolean cumulant $\beta_n (X, \ldots , X)$ in the noncommutative probability space $( {\mathbb C} [X], \mu )$. } satisfy $\kappa_n ( \nu ) = \beta_n ( \mu )$, $\forall \, n \in {\mathbb N} $.) It is known (cf. Theorem 1.2 in \cite{BeNi2008b}) that the Boolean cumulants of $\nu$ satisfy the relation \begin{equation} \label{eqn:56b} \beta_n ( \nu ) = \frac{1}{2} \beta_n ( \mu \boxplus \mu ), \ \ n \in {\mathbb N} , \end{equation} where $\mu \boxplus \mu$ (the ``free additive convolution'' of $\mu$ with itself) is the distribution of $a+b$. On the other hand, one can write \begin{equation} \label{eqn:56c} \beta_n ( \nu ) = \sum_{ \begin{array}{c} {\scriptstyle \pi \in NC(n),} \\ {\scriptstyle \pi \ll 1_n} \end{array} } \prod_{V \in \pi} \kappa_{|V|} ( \nu ) = \sum_{ \begin{array}{c} {\scriptstyle \pi \in NC(n),} \\ {\scriptstyle \pi \ll 1_n} \end{array} } \prod_{V \in \pi} \beta_{|V|} ( \mu ), \end{equation} where the first equality in (\ref{eqn:56c}) is the identity expressing Boolean cumulants in terms of free cumulants. Putting together (\ref{eqn:56b}) and (\ref{eqn:56c}) leads to \begin{equation} \label{eqn:56d} \beta_n ( \mu \boxplus \mu ) = 2 \cdot \sum_{ \begin{array}{c} {\scriptstyle \pi \in NC(n),} \\ {\scriptstyle \pi \ll 1_n} \end{array} } \ \Bigl( \prod_{ V \in \pi } \beta_{|V|} ( \mu ) \Bigr) , \ \ n \in {\mathbb N} , \end{equation} which is a re-phrasing of (\ref{eqn:56a}). \end{remark} $\ $ \begin{remark} \label{rem:57} One can do a bit of further combinatorial analysis following to the statement of Proposition \ref{prop:55}, in order to go to the level of power series, and thus come up with some equations in $\eta$-series which can be used to obtain $\eta_{a+b}$. More precisely, let us consider, same as in Proposition \ref{prop:55}, a noncommutative probability space $( {\mathcal A} , \varphi )$, and let $a,b \in {\mathcal A} $ be free. Remark \ref{rem:52} implies that the $\eta$-series $\eta_{a+b}(z) := \sum_{n=1}^{\infty} \beta_n(a+b)z^n$ splits as $\eta_{a+b}(z) =B_a(z)+B_b(z)$, where \begin{align*} B_a(z) &=\sum_{n=1}^{\infty} \Bigl( \sum_{c_2,\ldots,c_{n-1}\in\{a,b\}} \beta_n (a,c_2,\ldots,c_{n-1},a) \Bigr) z^n, \\ B_b(z) &=\sum_{n=1}^{\infty} \Bigl( \sum_{c_2,\ldots,c_{n-1}\in\{a,b\}} \beta_n (b,c_2,\ldots,c_{n-1},b) \Bigr) z^n . \end{align*} From Theorem \ref{thm:141} it follows that $\beta_n(a,c_2,\ldots,c_{n-1},a)$ is expressed as a sum over coloured non-crossing partitions with one outer block. Upon sorting out terms according to what is this outer block, one obtains the formula \begin{align*} B_a(z)&=\beta_1(a) z\\&+\beta_2(a)z^2\left(\sum_{m=0}^{\infty} \left(\beta_1(b)z+\beta_2(b,b)z^2+\left(\beta_3(b,a,b)+\beta_3(b,b,b)\right)z^3 + \ldots\right)^m \right) \\&+\eta_3(a)z^2\left(\sum_{m=0}^{\infty} \left(\eta_1(b)z+\eta_2(b,b)z^2+\left(\eta_3(b,a,b)+\eta_3(b,b,b)\right)z^3 + \ldots\right)^m\right)^2 + \ldots \end{align*} After summing the geometric series that have appeared, and after doing the similar calculation for $B_b (z)$, one arrives to the system of equations \begin{equation} \label{eqn:57a} \left\{ \begin{array}{lll} B_a(z) & = & \eta_a \left( \frac{z}{1-B_b(z)} \right) (1-B_b(z)), \\ B_b(z) & = & \eta_b \left( \frac{z}{1-B_a(z)} \right) (1-B_a(z)). \end{array} \right. \end{equation} Solving the system (\ref{eqn:57a}) may be used as a path towards the explicit calculation of the $\eta$-series $\eta_{a+b}=B_a+B_b$. This is of course not as smooth as using free cumulants and $R$-transforms (where one has the simplest possible formula, $R_{a+b} = R_a + R_b$), but we indicated how this goes in anticipation of the analogous development for anticommutators which we present in Section 6 below, and where free cumulants do not provide a simpler alternative. \end{remark} $\ $ \section{Boolean cumulants of products and of anticommutators} This section is concerned with studying $*$--Boolean cumulants of products of $*$--free random variables; as a byproduct of that, we obtain the formula for Boolean cumulants of a free anticommutator which was announced in Theorem \ref{thm:154} of the Introduction. For clarity, we start by discussing a concrete low-order example. \begin{example} \label{eg:61} Assume that $a,b$ are $*$--free. We are interested to calculate \begin{equation} \label{eqn:61a} \beta_3 ( ab, ab, b^*a^* ) = ? \, , \end{equation} where the expression sought on the right-hand side should be a polynomial expression in the Boolean $*$--cumulants of $a$ and those of $b$. We reach this goal in three steps. \vspace{6pt} {\em Step 1.} Use the formula for Boolean cumulants with products as entries. This step expresses $\beta_3 ( ab, ab, b^*a^* )$ as a sum of 8 terms, indexed by the collection of partitions $\sigma \in \mbox{Int} (6)$ with the property that $\sigma \vee \{ \, \{ 1,2 \}, \, \{ 3,4 \}, \, \{ 5,6 \}\, \} = 1_6$ (or, equivalently, that $\sigma \geq \{ \, \{ 1 \}, \, \{ 2,3 \}, \, \{ 4,5 \} \, \{ 6 \} \, \}$). By keeping in mind the Remark \ref{rem:52}, we see that 5 of these terms are sure to be equal to $0$, and thus the conclusion of this step is that \begin{equation} \label{eqn:61b} \beta_3 ( ab, ab, b^*a^* ) = \beta_1 (a) \beta_4 (b,a,b,b^{*}) \beta_1 (a^{*}) \end{equation} \[ + \beta_3 (a,b,a) \beta_2 (b,b^{*}) \beta_1 (a^{*}) + \beta_6 (a,b,a,b,b^{*},a^{*}). \] \vspace{6pt} {\em Step 2.} Use the formula for Boolean cumulants with free entries. This step takes on the Boolean cumulants which appear on the right-hand side of (\ref{eqn:61b}) and still mix together $a$'s and $b$'s -- we use Theorem \ref{thm:141} to express these Boolean cumulants as polynomials which separate the $a$'s from the $b$'s. In order to process $\beta_6 (a,b,a,b,b^{*},a^{*})$ we refer to Example \ref{ex:42} and we also do the immediate calculation that \[ \beta_4 (b,a,b,b^{*}) = \beta_1 (a) \beta_3 (b,b,b^{*}), \ \ \beta_3 (a,b,a) = \beta_2 (a,a) \beta_1 (b). \] The overall result of the calculation is that \begin{align} \label{eqn:61c} \beta_3 ( ab, ab, b^*a^* ) = & \ \beta_1 (a) \beta_1 (a) \beta_1 (a^{*}) \cdot \beta_3 (b,b,b^{*}) + \beta_2 (a,a) \beta_1 (a^{*}) \cdot \beta_1 (b) \beta_2 (b,b^{*}) \\ & + \beta_2(a,a^*)\beta_1(a) \cdot \beta_2(b,b)\beta_1(b^*) + \beta_2(a,a^*)\beta_1(a)\beta_3(b,b,b^*) \nonumber \\ & \beta_3(a,a,a^*) \cdot \beta_1(b)\beta_1(b)\beta_1(b^*) + \beta_3(a,a,a^*)\beta_1(b)\beta_2(b,b^*). \nonumber \end{align} The right-hand side of (\ref{eqn:61c}) is indeed a sum of products of $*$-cumulants of $a$ and of $b$, as we wanted. \vspace{6pt} {\em Step 3.} In order to understand what is going on, we record Equation (\ref{eqn:61c}) in the form: \begin{equation} \label{eqn:61d} \beta_3 ( ab, ab, b^*a^* ) = \end{equation} \[ \sum_{\pi \in \Pi} \prod_{ \begin{array}{c} {\scriptstyle U \in \pi, } \\ {\scriptstyle of \ `colour' \ a} \end{array} } \ \beta_{|U|} ( (a,b,a,b,b^{*},a^{*}) \mid U) \cdot \prod_{ \begin{array}{c} {\scriptstyle V \in \pi, } \\ {\scriptstyle of \ `colour' \ b} \end{array} } \ \beta_{|V|} ( (a,b,a,b,b^{*},a^{*}) \mid V), \] where $\Pi$ is the set of 6 non-crossing partitions depicted (in a way which includes colouring of blocks) in Figure 4. The question then becomes: can we put into evidence the structure which underlies this special set $\Pi \subseteq NC(6)$? Upon staring a bit at Figure 4, it becomes quite appealing to believe that VRNP must be involved here! This hunch is confirmed by Theorem \ref{thm:66} below, where the set $\Pi \subseteq NC(6)$ from Equation (\ref{eqn:61d}) becomes the correct indexing set corresponding to the tuple $\varepsilon = (1,1,*) \in \{ 1,* \}^3$. \end{example} \vspace{6pt} \begin{minipage}{1\linewidth} \begin{center} \begin{tikzpicture} \draw[red,thick] (-8,0) -- (-8,0.25); \draw[red,thick] (-6,0) -- (-6,0.25); \draw[red,thick] (-3,0) -- (-3,0.25); \draw[blue,thick] (-7,0) to[out=90,in=90] node [sloped,above] {} (-5,0); \draw[blue,thick] (-5,0) to[out=90,in=90] node [sloped,above] {} (-4,0); \node[below,red] (1) at (-8,0) {$a$}; \node[below,blue] (2) at (-7,0) {$b$}; \node[below,red] (3) at (-6,0) {$a$}; \node[below,blue] (4) at (-5,0) {$b$}; \node[below,blue] (5) at (-4,0) {$b^*$}; \node[below,red] (6) at (-3,0) {$a^*$}; \draw[blue,thick] (1,0) -- (1,0.25); \draw[red,thick] (5,0) -- (5,0.25); \draw[red,thick] (0,0) to[out=90,in=90] node [sloped,above] {} (2,0); \draw[blue,thick] (3,0) to[out=90,in=90] node [sloped,above] {} (4,0); \node[below,red] (7) at (0,0) {$a$}; \node[below,blue] (8) at (1,0) {$b$}; \node[below,red] (9) at (2,0) {$a$}; \node[below,blue] (10) at (3,0) {$b$}; \node[below,blue] (11) at (4,0) {$b^*$}; \node[below,red] (12) at (5,0) {$a^*$}; \end{tikzpicture} \end{center} \end{minipage} \begin{minipage}{1\linewidth} \begin{center} \begin{tikzpicture} \draw[red,thick] (-6,0) -- (-6,0.25); \draw[blue,thick] (-4,0) -- (-4,0.25); \draw[red,thick] (-8,0) to[out=90,in=90] node [sloped,above] {} (-3,0); \draw[blue,thick] (-7,0) to[out=90,in=90] node [sloped,above] {} (-5,0); \node[below,red] (1) at (-8,0) {$a$}; \node[below,blue] (2) at (-7,0) {$b$}; \node[below,red] (3) at (-6,0) {$a$}; \node[below,blue] (4) at (-5,0) {$b$}; \node[below,blue] (5) at (-4,0) {$b^*$}; \node[below,red] (6) at (-3,0) {$a^*$}; \draw[red,thick] (2,0) -- (2,0.25); \draw[red,thick] (0,0) to[out=90,in=90] node [sloped,above] {} (5,0); \draw[blue,thick] (1,0) to[out=90,in=90] node [sloped,above] {} (3,0); \draw[blue,thick] (3,0) to[out=90,in=90] node [sloped,above] {} (4,0); \node[below,red] (7) at (0,0) {$a$}; \node[below,blue] (8) at (1,0) {$b$}; \node[below,red] (9) at (2,0) {$a$}; \node[below,blue] (10) at (3,0) {$b$}; \node[below,blue] (11) at (4,0) {$b^*$}; \node[below,red] (12) at (5,0) {$a^*$}; \end{tikzpicture} \end{center} \end{minipage} \begin{minipage}{1\linewidth} \begin{center} \begin{tikzpicture} \draw[blue,thick] (-4,0) -- (-4,0.25); \draw[blue,thick] (-7,0) -- (-7,0.25); \draw[blue,thick] (-5,0) -- (-5,0.25); \draw[red,thick] (-8,0) to[out=90,in=90] node [sloped,above] {} (-6,0); \draw[red,thick] (-6,0) to[out=90,in=90] node [sloped,above] {} (-3,0); \node[below,red] (1) at (-8,0) {$a$}; \node[below,blue] (2) at (-7,0) {$b$}; \node[below,red] (3) at (-6,0) {$a$}; \node[below,blue] (4) at (-5,0) {$b$}; \node[below,blue] (5) at (-4,0) {$b^*$}; \node[below,red] (6) at (-3,0) {$a^*$}; \draw[blue,thick] (1,0) -- (1,0.25); \draw[red,thick] (0,0) to[out=90,in=90] node [sloped,above] {} (2,0); \draw[red,thick] (2,0) to[out=90,in=90] node [sloped,above] {} (5,0); \draw[blue,thick] (3,0) to[out=90,in=90] node [sloped,above] {} (4,0); \node[below,red] (7) at (0,0) {$a$}; \node[below,blue] (8) at (1,0) {$b$}; \node[below,red] (9) at (2,0) {$a$}; \node[below,blue] (10) at (3,0) {$b$}; \node[below,blue] (11) at (4,0) {$b^*$}; \node[below,red] (12) at (5,0) {$a^*$}; \end{tikzpicture} \end{center} \end{minipage} \begin{center} {\bf Figure 4.} {\em The set $\Pi \subseteq NC(6)$ used in Equation (\ref{eqn:61d}).} \end{center} $\ $ \begin{remark} \label{rem:62} A similar calculation to the one shown in Example \ref{eg:61} could be done in the world of free cumulants, and would lead to a similarly looking formula which expresses the free cumulant $\kappa_3 (ab, ab, b^{*}a^{*} )$ in terms of the $*$-free cumulants of $a$ and the $*$-free cumulants of $b$. The formula with free cumulants is ``just a bit'' more complicated than the one for Boolean cumulants obtained in (\ref{eqn:61c}) above: it has 7 terms instead of 6, where 6 of the 7 terms are having the same structure as in (\ref{eqn:61c}), while the 7th term is \[ \kappa_2(a,a^*)\kappa_1(a)\kappa_1(b)\kappa_2(b,b^*), \] corresponding to the partition $\{\{1,6\},\{2\},\{3\},\{4,5\}\} \in NC(6)$. But, unlike the partitions corresponding to the other 6 terms, this latter partition does not satisfy VNRP! (It is one of the two partitions without VNRP that were featured in Figure 3(b) in Section 4.) \end{remark} $\ $ We now move to discuss Boolean cumulants for general words made with $ab$ and with $(ab)^{*}$. To this end, we recall from the Introduction the terminology about ac-friendly partitions and about the canonical alternating colouring of a non-crossing partition. In the Introduction, the definition of what is an ac-friendly partition in $NC(2n)$ was made by only referring to depths of blocks. It will come in handy to note that this notion can also be approached via canonical alternating colourings, as explained in the next proposition. \begin{proposition} \label{prop:63} Let $n$ be a positive integer and let $\pi$ be a partition in $NC(2n)$. Then $\pi$ belongs to the set $ NC_{\mathrm{ac-friendly}} (2n)$ introduced in Definition \ref{def:151} if and only if it satisfies the following two conditions: \vspace{6pt} \em{(AC-Friendly1)} $ \mathrm{OuterMax} ( \pi ) \subseteq \{ 1,3, \ldots , 2n-1 \} \cup \{ 2n \}$. \vspace{6pt} (AC-Friendly2') $ \mathrm{calt} _{\pi} (2i-1) \neq \mathrm{calt} _{\pi} (2i), \ \ \forall \, 1 \leq i \leq n$, where $ \mathrm{calt} _{\pi} : \{ 1, \ldots , 2n \} \to \{ 1,2 \}$ is the canonical alternating colouring associated to $\pi$. \end{proposition} \begin{proof} ``$\Rightarrow$'' We assume that (AC-Friendly1) and (AC-Friendly2) hold. Suppose, toward a contradiction, that (AC-Friendly2') fails. Then there exists an $i$ such that $ \mathrm{calt} _\pi(2i-1)\not= \mathrm{calt} _\pi(2i)$. This can only happen if $2i-1$ and $2i$ belong to distinct blocks $A$ and $B$ that are ``siblings," i.e., have same parent $C$. But then, for $j=2i-1$, we have that $ \mathrm{depth} _\pi(j)= \mathrm{depth} \pi(C)+1= \mathrm{depth} _\pi(j+1)$ contradicting (AC-Friendly2). \vspace{6pt} ``$\Leftarrow$'' We assume that (AC-Friendly1) and (AC-Friendly2') hold. Suppose, toward a contradiction, that (AC-Friendly2) fails. Let $j=2i-1\not\in \mathrm{OuterMax} (\pi)$ be such that $ \mathrm{depth} _\pi(j)= \mathrm{depth} _\pi(j+1)$. Since $j$ and $j+1$ cannot belong to distinct outer blocks (as then we would have $j\in \mathrm{OuterMax} (\pi)$) this can only happen if they either belong to the same block or they belong to blocks that have the same parent. In both of these cases we then have that $ \mathrm{calt} _\pi(2i-1)= \mathrm{calt} _\pi(2i)$, contradicting (AC-Friendly2'). \end{proof} \begin{notation} \label{def:64} Let $n$ be a positive integer, let $\pi$ be a partition in $ NC_{\mathrm{ac-friendly}} (2n)$, and consider the canonical alternating colouring $ \mathrm{calt} _{\pi} : \{ 1, \ldots , 2n \} \to \{ 1,2 \}$. We use the values $ \mathrm{calt} _{\pi} (1), \mathrm{calt} _{\pi} (3), \ldots , \mathrm{calt} _{\pi} (2n-1)$ in order to create a tuple $\varepsilon \in \{ 1,* \}^n$, as follows: \begin{equation} \label{eqn:64a} \varepsilon(i) = \left\{ \begin{array}{ll} 1, & \mbox{ if $ \mathrm{calt} _{\pi} (2i-1) = 1$, } \\ *, & \mbox{ if $ \mathrm{calt} _{\pi} (2i-1) = 2$ } \end{array} \right\} \ , 1 \leq i \leq n. \end{equation} The tuple $\varepsilon \in \{ 1,* \}^n$ so defined will be denoted as $ \mathrm{oddtuple} ( \pi )$. \end{notation} \begin{remark} \label{rem:65} Let $n$ be a positive integer and let $\pi$ be a partition in $ NC_{\mathrm{ac-friendly}} (2n)$. Knowing what is the tuple $ \mathrm{oddtuple} ( \pi ) \in \{ 1,* \}^n$ gives us precisely the same information as if we knew what is the colouring $ \mathrm{calt} _{\pi} : \{ 1, \ldots , 2n \} \to \{ 1,2 \}$. Indeed, if we know $ \mathrm{oddtuple} ( \pi )$ then we can use Equation (\ref{eqn:64a}) to find the values $ \mathrm{calt} _{\pi} (1), \mathrm{calt} _{\pi} (3), \ldots , \mathrm{calt} _{\pi} (2n-1)$, after which we can also find out what are $ \mathrm{calt} _{\pi} (2), \mathrm{calt} _{\pi} (4), \ldots , \mathrm{calt} _{\pi} (2n)$ based on the fact that $ \mathrm{calt} _{\pi} (2i) \neq \mathrm{calt} _{\pi} (2i-1)$, $1 \leq i \leq n$. \end{remark} We can now state the desired formula about the joint Boolean cumulants of $ab$ and $ba$. \begin{theorem} \label{thm:66} Let $( {\mathcal A} , \varphi )$ be a $*$-probability space and let $( \beta_n : {\mathcal A} ^n \to {\mathbb C} )_{n=1}^{\infty}$ be the family of Boolean cumulant functionals associated to it. We consider two selfadjoint elements $a,b \in {\mathcal A} $ such that $a$ is freely independent from $b$, and we consider the sequences of Boolean cumulants $( \beta_n (a) )_{n=1}^{\infty}$ and $( \beta_n (b) )_{n=1}^{\infty}$. \vspace{6pt} (1) Let $n$ be a positive integer and let $\varepsilon = ( \varepsilon (1), \ldots , \varepsilon (n)) \in \{ 1,* \}^n$ be such that $\varepsilon (1) = 1$. One has \begin{equation} \label{eqn:66a} \beta_n \bigl( \, (ab)^{\varepsilon (1)}, \ldots , (ab)^{\varepsilon (n)} \, \bigr) = \end{equation} \[ \sum_{ \begin{array}{c} {\scriptstyle \pi \in NC_{\mathrm{ac-friendly}} (2n),} \\ {\scriptstyle such \ that} \\ {\scriptstyle \mathrm{oddtuple} ( \pi ) = \varepsilon} \end{array} } \ \Bigl( \prod_{ \begin{array}{c} {\scriptstyle U \in \pi, \ with} \\ {\scriptstyle \mathrm{calt} _{\pi} (U) = 1} \end{array} } \ \beta_{|U|} (a) \Bigr) \cdot \Bigl( \prod_{ \begin{array}{c} {\scriptstyle V \in \pi, \ with} \\ {\scriptstyle \mathrm{calt} _{\pi} (V) = 2} \end{array} } \ \beta_{|V|} (b) \Bigr). \] \vspace{6pt} (2) Let $n$ be a positive integer and let $\varepsilon = ( \varepsilon (1), \ldots , \varepsilon (n)) \in \{ 1,* \}^n$ be such that $\varepsilon (1) = *$. Consider the complementary tuple $\varepsilon ' \in \{ 1,* \}^n$, uniquely determined by the requirement that $\varepsilon ' (i) \neq \varepsilon (i)$, for all $1 \leq i \leq n$. One has \begin{equation} \label{eqn:66b} \beta_n \bigl( \, (ab)^{\varepsilon (1)}, \ldots , (ab)^{\varepsilon (n)} \, \bigr) = \end{equation} \[ \sum_{ \begin{array}{c} {\scriptstyle \pi \in NC_{\mathrm{ac-friendly}} (2n),} \\ {\scriptstyle such \ that} \\ {\scriptstyle \mathrm{oddtuple} ( \pi ) = \varepsilon '} \end{array} } \ \Bigl( \prod_{ \begin{array}{c} {\scriptstyle U \in \pi, \ with} \\ {\scriptstyle \mathrm{calt} _{\pi} (U) = 1} \end{array} } \ \beta_{|U|} (b) \Bigr) \cdot \Bigl( \prod_{ \begin{array}{c} {\scriptstyle V \in \pi, \ with} \\ {\scriptstyle \mathrm{calt} _{\pi} (V) = 2} \end{array} } \ \beta_{|V|} (a) \Bigr). \] \end{theorem} \begin{proof} We will assume the case (1), i.e., that $\varepsilon(1)=1$ (the case $\varepsilon(1)=*$ is the `mirror' image of this case and is left to the reader). Note that $\varepsilon$ induces a colouring $c\colon\{1,\ldots,2n\}\to\{1,2\}$ by $c(2i-1)=1, c(2i)=2$ when $\varepsilon(i)=1$ and $c(2i-1)=2$, $c(2i)=1$ when $\varepsilon(i)=*$; in other words, $c$ assigns $1$ to positions where there is an $a$ in $(ab)^{\varepsilon(1)}\ldots(ab)^{\varepsilon(n)}$ and $2$ to positions where there is a $b$. We do again the steps presented in Example \ref{eg:61}, where the discussion is now made to go in reference to an abstract tuple $\varepsilon$, rather than the special case of $(1,1,*) \in \{ 1,* \}^3$. Step 1 takes us to a sum over interval partitions partitions $\sigma = \{ J_1, \ldots , J_p \} \in \mbox{Int} (2n)$ such that $\sigma\vee\{\{1\},\{2,3\},\ldots, \{2n-2,2n-1\},\{2n\}\}=1_{2n}$. The condition $\sigma\vee\{\{1\},\{2,3\},\ldots, \{2n-2,2n-1\},\{2n\}\}=1_{2n}$ is equivalent to saying that $\max (J_k) \in \{ 1,3, \ldots , 2n-1 \}$ for all $1 \leq k < p$. The latter can be expressed directly in terms of the cardinalities of the blocks of $\sigma$: either $\sigma = 1_{2n}$, or it has $|J_1|, |J_p|$ odd and $|J_k|$ even for all $1 < k < p$. This implies, among other things, that for all $k<p$ we have that $c(\max(J_k))\not=c(\min(J_{k+1}))$. The observation in Remark \ref{rem:52} that certain mixed Boolean cumulants with free entries have to vanish, then yields that we are left to only consider the cases where for each $k$ we have that $c(\min(J_k))=c(\max(J_k))$. Step 2: We now apply the separation formula from Theorem \ref{thm:141} to each block $J_k$ to get that $\beta_n \bigl( \, (ab)^{\varepsilon (1)}, \ldots , (ab)^{\varepsilon (n)} \, \bigr)$ is equal to \[ \sum_{\begin{array}{c} \scriptstyle{\sigma=\{J_1,\ldots,J_p\}\in \mbox{Int} (2n)}\\ \scriptstyle{ \sigma\vee \{\{1,2\},\ldots,\{2n-1,2n\}\}=1_{2n}}\\ \scriptstyle{\forall k, c(\min(J_k))=c(\max(J_k))} \end{array}} \sum_{ \begin{array}{c} \scriptstyle{\pi \in NC(2n,c),} \\ \scriptstyle{\pi\ll \sigma} \\ \scriptstyle{\pi\mbox{ has VNRP }} \end{array} } \ \Bigl( \prod_{ \begin{array}{c} {\scriptstyle U \in \pi, \ with} \\ {\scriptstyle c(U) = 1} \end{array} } \ \beta_{|U|} (a) \Bigr) \cdot \Bigl( \prod_{ \begin{array}{c} {\scriptstyle V \in \pi, \ with} \\ {\scriptstyle c(V) = 2} \end{array} } \ \beta_{|V|} (b) \Bigr). \] Step 3: We claim that partitions $\pi$ appearing in the summation above are precisely those in $ NC_{\mathrm{ac-friendly}} (2n)$ for which $ \mathrm{oddtuple} ( \pi ) = \varepsilon$ and that we always have $ \mathrm{calt} _\pi = c$; from whence the formula (\ref{eqn:66a}) follows. Since consecutive blocks of $\sigma$ start with distinct colours we have that the consecutive outer blocks of $\pi$ have distinct colours. When we combine this observation with the facts that the first block of $\pi$ has colour $1$ and that $\pi$ has VNRP (with respect to $c$) we get that $c= \mathrm{calt} _\pi$. As already mentioned above we have that the condition $\pi\ll\sigma$, $\sigma\in \mbox{Int} (2n)$, $\sigma\vee \{\{1,2\},\ldots,\{2n-1,2n\}\}=1_{2n}$ is equivalent to $\pi$ satisfying (AC-Friendly1). It now remains to prove that every $\pi$ in the above summation formula satisfies (AC-Friendly2). Suppose the converse. Then there is an odd number $j\not\in \mathrm{OuterMax} (\pi)$ such that $ \mathrm{depth} _\pi(j)= \mathrm{depth} _\pi(j+1)$. Since $j$ and $j+1$ have distinct $c$-colours we have that $j$ and $j+1$ belong to distinct blocks. They cannot both belong to outer blocks (in this case it would follow that $j\in \mathrm{OuterMax} (\pi)$). The equality $ \mathrm{depth} _\pi(j)= \mathrm{depth} _\pi(j+1)$ then implies that the blocks containing $j$ and $j+1$ must have the same parent, and considering the $c$-colour of the parent-block leads to an immediate contradiction with VNRP. \end{proof} \begin{remark} \label{rem:67} We note that Theorem \ref{thm:66} could be stated in the framework where $( {\mathcal A} , \varphi )$ is a plain noncommutative probability space, and we look at joint Boolean cumulants of $ab$ and $ba$, with $a$ free from $b$. The statement of Theorem \ref{thm:66} could also be extended to the case when we deal with joint Boolean cumulants of $ab$ and $(ab)^{*}$ in a $*$-probability space, without assuming that $a$ and $b$ are selfadjoint. The proof would be the same, only that it would result in stuffier formulas. \end{remark} \begin{remark} \label{rem:58} The special case $\varepsilon = (1,1, \ldots , 1)$ of Theorem \ref{thm:66} gives a formula for the Boolean cumulant $\beta_n (ab,ab, \ldots , ab)$. We explain here that this is precisely the formula from \cite{BeNi2008} which was reviewed in Equation (\ref{eqn:12a}) of the Introduction. To this end, let us first note that if a partition $\sigma \in NC_{\mathrm{ac-friendly}} (2n)$ has $ \mathrm{oddtuple} ( \sigma ) = (1,1, \ldots , 1)$, then the canonical alternating colouring $ \mathrm{calt} _{\sigma}$ must have $ \mathrm{calt} _{\sigma} (2i-1) = 1$ and $ \mathrm{calt} _{\sigma} (2i) = 2$ for all $1 \leq i \leq n$. So $ \mathrm{calt} _{\sigma}$ is the colouring of $\{ 1, \ldots , 2n \}$ by parity, which implies that every block of $\sigma$ is contained either in $\{ 1,3, \ldots , 2n-1 \}$ or in $\{ 2,4, \ldots , 2n \}$. We note moreover that such $\sigma$ is sure to only have two outer blocks, the blocks $W'$ and $W''$ which contain the numbers $1$ and $2n$, respectively. Indeed, if $\sigma$ had an outer block $W \neq W', W''$, then the condition (AC-Friendly1) satisfied by $\sigma$ would imply that $\min (W)$ is an even number and $\max (W)$ is an odd number -- not possible! Knowing these things about $\sigma$, plus the fact that $\sigma$ has VNRP, leads to the conclusion that $\sigma$ must be of the form $\sigma = \pi^{(\mathrm{odd})} \sqcup ( \mbox{Kr} _n ( \pi ) )^{\mathrm{(even)}}$ for some $\pi \in NC(n)$; this is precisely the content of Lemma 6.8 in \cite{BeNi2008}. Conversely, let $\pi$ be in $NC(n)$ and consider the partition $\sigma = \pi^{(\mathrm{odd})} \sqcup ( \mbox{Kr} _n ( \pi ) )^{\mathrm{(even)}} \in NC(2n)$. Then Lemma 6.6 of \cite{BeNi2008} assures us that the canonical alternating colouring of $\sigma$ is the colouring of $\{ 1, \ldots , 2n \}$ by parity. The same lemma of \cite{BeNi2008} also records the fact that $\sigma$ has exactly two outer blocks, the ones containing the numbers $1$ and $2n$, and this clearly entails the condition (AC-Friendly1) from Proposition \ref{prop:63} above. In view of Proposition \ref{prop:63}, we then conclude that $\sigma \in NC_{\mathrm{ac-friendly}} (2n)$ and at the same time we see that the tuple $ \mathrm{oddtuple} ( \sigma )$ is equal to $(1,1, \ldots , 1)$. The discussion from the preceding two paragraphs shows that when we make $\varepsilon = (1,1, \ldots , 1)$ in Theorem \ref{thm:66}(1), the summation on the right-hand side of Equation (\ref{eqn:66a}) is made over partitions of the form $ \pi^{(\mathrm{odd})} \sqcup ( \mbox{Kr} _n ( \pi ) )^{\mathrm{(even)}}$, with $\pi$ running in $NC(n)$. Moreover, when looking at the term of the summation which is indexed by a $\pi \in NC(n)$, one sees that the two products appearing there are precisely $\prod_{U \in \pi} \beta_{|U|} (a)$ and $\prod_{V \in \mbox{Kr} _n ( \pi )} \beta_{|V|} (b)$. Hence this special case of Equation (\ref{eqn:66a}) retrieves Equation (\ref{eqn:12a}), as claimed at the beginning of the remark. \end{remark} $\ $ By starting from Theorem \ref{thm:66}, it is easy to prove the formula announced in Theorem \ref{thm:154} of the Introduction (and repeated below), concerning the Boolean cumulants of a free anticommutator. \begin{theorem} \label{thm:610} Let $( {\mathcal A} , \varphi )$ be a noncommutative probability space and let $a,b \in {\mathcal A} $ be such that $a$ is freely independent from $b$. Consider the sequences of Boolean cumulants $( \beta_n (a) )_{n=1}^{\infty}$ and $( \beta_n (b) )_{n=1}^{\infty}$ for $a$ and for $b$, respectively. Then, for every $n \geq 1$, the $n$-th Boolean cumulant of $ab + ba$ is \[ \beta_n (ab+ba) = \sum_{\pi \in NC_{\mathrm{ac-friendly}} (2n)} \ \Bigl( \prod_{ U \in \pi, \mathrm{calt} _{\pi} (U) = 1} \beta_{|U|} (a) \Bigr) \cdot \Bigl( \prod_{ V \in \pi, \mathrm{calt} _{\pi} (V) = 2} \beta_{|V|} (b) \Bigr) \] \[ + \ \sum_{\pi \in NC_{\mathrm{ac-friendly}} (2n)} \ \Bigl( \prod_{ U \in \pi, \mathrm{calt} _{\pi} (U) = 1} \beta_{|U|} (b) \Bigr) \cdot \Bigl( \prod_{ V \in \pi, \mathrm{calt} _{\pi} (V) = 2} \beta_{|V|} (a) \Bigr) . \] \end{theorem} \begin{proof} We write that \begin{equation} \label{eqn:610a} \beta_n ( ab+ba , \ldots , ab+ba ) = \sum_{\varepsilon \in \{ 1,* \}^n } \, \beta_n \bigl( \, (ab)^{\varepsilon (1)}, \ldots , (ab)^{\varepsilon (n)} \, \bigr) = \end{equation} \[ \sum_{\varepsilon \in \{ 1,* \}^n, \varepsilon(1)=1 } \beta_n \bigl( \, (ab)^{\varepsilon (1)}, \ldots , (ab)^{\varepsilon (n)} \, \bigr) + \sum_{\varepsilon \in \{ 1,* \}^n, \varepsilon(1)=* } \beta_n \bigl( \, (ab)^{\varepsilon (1)}, \ldots , (ab)^{\varepsilon (n)} \, \bigr). \] Note that every $\pi\in NC_{\mathrm{ac-friendly}} (2n)$ determines a unique $\varepsilon:= \mathrm{oddtuple} (\pi)\in\{1,*\}^n$ such that $\varepsilon(1)=1$. Also recall that every $\varepsilon\in \{1,*\}^n$ for which $\varepsilon(1)=*$ we have $\varepsilon'\in\{1,*\}^n$ determined by $\varepsilon'(k)\not=\varepsilon(k)$ for all $k$ (in particular $\varepsilon'(1)=1$). We invoke the formulas found in Theorem \ref{thm:66} to get \begin{align*} &\sum_{\varepsilon \in \{ 1,* \}^n, \varepsilon(1)=1 } \beta_n \bigl( \, (ab)^{\varepsilon (1)}, \ldots , (ab)^{\varepsilon (n)} \, \bigr)\\ &= \sum_{\varepsilon \in \{ 1,* \}^n, \varepsilon(1)=1 }\sum_{ \begin{array}{c} {\scriptstyle \pi \in NC_{\mathrm{ac-friendly}} (2n),} \\ {\scriptstyle such \ that} \\ {\scriptstyle \mathrm{oddtuple} ( \pi ) = \varepsilon} \end{array} } \ \Bigl( \prod_{ \begin{array}{c} {\scriptstyle U \in \pi, \ with} \\ {\scriptstyle \mathrm{calt} _{\pi} (U) = 1} \end{array} } \ \beta_{|U|} (a) \Bigr) \cdot \Bigl( \prod_{ \begin{array}{c} {\scriptstyle V \in \pi, \ with} \\ {\scriptstyle \mathrm{calt} _{\pi} (V) = 2} \end{array} } \ \beta_{|V|} (b) \Bigr) \\ &= \sum_{ {\scriptstyle \pi \in NC_{\mathrm{ac-friendly}} (2n)} } \ \Bigl( \prod_{ \begin{array}{c} {\scriptstyle U \in \pi, \ with} \\ {\scriptstyle \mathrm{calt} _{\pi} (U) = 1} \end{array} } \ \beta_{|U|} (a) \Bigr) \cdot \Bigl( \prod_{ \begin{array}{c} {\scriptstyle V \in \pi, \ with} \\ {\scriptstyle \mathrm{calt} _{\pi} (V) = 2} \end{array} } \ \beta_{|V|} (b) \Bigr), \end{align*} and \begin{align*} &\sum_{\varepsilon \in \{ 1,* \}^n, \varepsilon(1)=* } \beta_n \bigl( \, (ab)^{\varepsilon (1)}, \ldots , (ab)^{\varepsilon (n)} \, \bigr)\\ &= \sum_{\varepsilon \in \{ 1,* \}^n, \varepsilon(1)=* }\sum_{ \begin{array}{c} {\scriptstyle \pi \in NC_{\mathrm{ac-friendly}} (2n),} \\ {\scriptstyle such \ that} \\ {\scriptstyle \mathrm{oddtuple} ( \pi ) = \varepsilon'} \end{array} } \ \Bigl( \prod_{ \begin{array}{c} {\scriptstyle U \in \pi, \ with} \\ {\scriptstyle \mathrm{calt} _{\pi} (U) = 1} \end{array} } \ \beta_{|U|} (b) \Bigr) \cdot \Bigl( \prod_{ \begin{array}{c} {\scriptstyle V \in \pi, \ with} \\ {\scriptstyle \mathrm{calt} _{\pi} (V) = 2} \end{array} } \ \beta_{|V|} (a) \Bigr)\\ &= \sum_{ {\scriptstyle \pi \in NC_{\mathrm{ac-friendly}} (2n)} } \ \Bigl( \prod_{ \begin{array}{c} {\scriptstyle U \in \pi, \ with} \\ {\scriptstyle \mathrm{calt} _{\pi} (U) = 1} \end{array} } \ \beta_{|U|} (b) \Bigr) \cdot \Bigl( \prod_{ \begin{array}{c} {\scriptstyle V \in \pi, \ with} \\ {\scriptstyle \mathrm{calt} _{\pi} (V) = 2} \end{array} } \ \beta_{|V|} (a) \Bigr). \end{align*} \end{proof} $\ $ \section{On the $\eta$--series of a free anticommutator} In this section we show how observations about Boolean cumulants of a free anticommuator from the previous section can be captured in the form of a system of equation at the level of $\eta$-series. \subsection{Equations in power series.} $\ $ \noindent Throughout this subsection we fix a $*$-probability space $( {\mathcal A} , \varphi )$ and two elements $a,b \in {\mathcal A} $. For clarity of arguments it is better if at first we do not assume that $a$ and $b$ are selfadjoint, and we discuss the formal power series $\eta_{ab,b^*a^*}\in\mathbb{C}\langle\langle z_a,z_{a^*},z_b,z_{b^*}\rangle\rangle$ defined as \begin{equation} \label{eqn:6a} \eta_{ab,b^*a^*} = \sum_{n=1}^{\infty} \sum_{ (\varepsilon (1), \ldots , \varepsilon (n)) \in \{1,* \}^n} \ \beta_n ( (ab)^{\varepsilon (1)}, \ldots , (ab)^{\varepsilon (n)}) (z_a z_b)^{\varepsilon (1)} \cdots (z_a z_b)^{\varepsilon (n)}, \end{equation} where on the right-hand side of (\ref{eqn:6a}) we make the convention to put $(z_a z_b)^{*} = z_{b^{*}} z_{a^{*}}$ (so, for instance, the term corresponding to $\varepsilon = (1,1,*) \in \{ 1,* \}^3$ is $\beta_3 (ab, ab, b^{*}a^{*}) z_a z_b z_a z_b z_{b^{*}} z_{a^{*}}$). At some point down the line we will however switch to the special case when $a=a^*,b=b^*$ and $z_a= z_{{a}^*} = z_b = z_{{b}^*} =: z$; in this special case the series from Equation (\ref{eqn:6a}) becomes a series of one variable, which is nothing but $\eta_{ab+ba}(z^2)$. Returning to Equation (\ref{eqn:6a}) we observe that $\eta_{ab,b^*a^*}$ splits naturally as a sum, \begin{equation} \label{eqn:6c} \eta_{ab,b^*a^*} = \eta^{1,1} + \eta^{1,*} + \eta^{*,1} + \eta^{*,*}, \end{equation} where $\eta^{1,1}$ contains those terms of $\eta_{ab,b^*a^*}$ which correspond to Boolean cumulants beginning and ending with $ab$, $\eta^{1,*}$ contains those terms of $\eta_{ab,b^*a^*}$ which correspond to Boolean cumulants that begin with $ab$ and end with $b^*a^*$, etc. Under the assumption that $\{ a,a^{*} \}$ is free from $\{ b, b^{*} \}$, one can then make a number of structural observations about the four power series introduced in Equation (\ref{eqn:6c}). Let us start with $\eta^{1,1}$. Every term of this series is of the form \begin{equation} \label{eqn:6d} \beta_n(ab,\ldots,ab)z_a z_b\dotsm z_a z_b . \end{equation} To the Boolean cumulant appearing in (\ref{eqn:6d}) we apply the formula for Boolean cumulants with products as entries from Proposition \ref{prop:210}, together with the property that a Boolean cumulant which starts with $a$ or $a^*$ and ends with $b$ or $b^*$ (or vice versa) vanishes, as noticed in Remark \ref{rem:52}. Then it follows from Theorem \ref{thm:610} that the cumulant from (\ref{eqn:6d}) is a sum of terms of the form \begin{align}\label{eqn:92} \beta_{l_0}(a,b,\ldots,a)\left(\prod_{j=1}^{n-1} \beta_{k_{j}}(b,\ldots,b^*)\beta_{l_{j}}(a^*,\ldots,a)\right) \beta_{k_{n}}(b,\ldots,a,b), \end{align} for $n\geq 1$, $l_0,k_n\geq 1$ and $l_j,k_j\geq 2$ for $j=1,\ldots,n-1$. For $l,l^\prime\in\{a,a^*\}$ or $l,l^\prime\in\{b,b^*\}$ we define power series $f_{l,l^*}\in \mathbb{C}\langle\langle z_a,z_{a^*},z_b,z_{b^*}\rangle\rangle$ as power series which contain Boolean cumulants starting with $l$ and ending with $l^\prime$ which can appear in the expression above. To be more precise, consider $\eta_{a,a^*,b,b^*}\in\mathbb{C}\langle\langle z_a,z_{a^*},z_b,z_{b^*}\rangle\rangle$ the joint $\eta$--series of $a,a^*,b,b^*$ then $f_{l,l^\prime}$ is a restriction of $\eta_{a,a^*,b,b^*}$ to these terms which begin with $l$, end with $l^\prime$, and corresponding word $z_{l}\dotsm z_{l^\prime}$ is a subword of some word of the type $(z_az_b)^{l_1}(z_{b^*}z_{a^*})^{k_1}\dotsm(z_az_b)^{l_n}(z_{b^*}z_{a^*})^{k_n}$ with $n\geq 1$ and $l_1,k_n\geq 0$ and $k_1,\ldots,k_{n-1},l_2,\ldots,l_n\geq 1$. We have for example \begin{equation} \label{eqn:6e} \left\{ \begin{array}{lll} f_{a,a} &=& \beta_1(a)z_a+\beta_3(a,b,a)z_az_bz_{a^*} +\beta_5(a,b,b^*,a^*,a)z_az_bz_{b^*}z_{a^*}z_a+\ldots \\ f_{a^*,a} &=& \beta_2(a^*,a)z_{a^*} z_a+ \beta_4(a^*,a,b,a)z_{a^*}z_az_bz_a+\ldots \end{array} \right. \end{equation} With such notation, Equation \eqref{eqn:92} can be written as \begin{align*} \eta^{1,1}=f_{a,a}(1-f_{b,b^*} f_{a^*,a})^{-1} f_{b,b}, \end{align*} with the convention \begin{align*} (1-f_{b,b^*} f_{a^*,a})^{-1}=\sum_{n=0}^{\infty}\left(f_{b,b^*} f_{a^*,a}\right)^{n}. \end{align*} Similar analysis to the one done above for $\eta^{1,1}$ can be done for the remaining three power series from the right-hand side of (\ref{eqn:6c}). We have for example \begin{align*} \eta^{1,*}=f_{a,a^*}+f_{a,a} (1-f_{b,b^*}f_{a^*,a})^{-1} f_{b,b^*} f_{a^*,a^*}, \end{align*} where the additional term $f_{a,a^*}$ comes from the term of the expansion of $\beta_n (ab,\ldots,b^*a^*)$ with one outer block. The system of equations which comes out of the preceding discussion can be nicely written in matrix form, as follows: \begin{align} \label{eqeta} \begin{bmatrix} \eta^{1,1} & \eta^{1,*} \\ \eta^{*,1} & \eta^{*,*} \end{bmatrix}=\begin{bmatrix} f_{aa} (1-f_{bb^*}f_{a^*a})^{-1}f_{bb} & f_{aa^*}+f_{aa} (1-f_{bb^*}f_{a^*a})^{-1} f_{bb^*} f_{a^*a^*}\\ f_{b^*b}+f_{b^*b^*} (1-f_{a^*a} f_{bb^*})^{-1} f_{a^*a} f_{bb} & f_{b^*b^*} (1-f_{a^*a} f_{bb^*})^{-1} f_{a^*a^*} \end{bmatrix}. \end{align} Once that Equation (\ref{eqeta}) is put into evidence, the problem of computing the $\eta$-series $\eta_{ab, b^{*}a^{*}}$ from (\ref{eqn:6a}) is reduced to the one of computing the power series $f_{l,l^\prime}$. We will take on this job in the special case when $a$ and $b$ are assumed to be (freely independent and) selfadjoint. In this special case, the determination of the series $f_{l, l^\prime}$ has to be made in terms of the Boolean cumulants of $a$ and of $b$, or equivalently, in terms of the $\eta$-series of these elements. The mechanism for doing so is provided by the following theorem. \begin{theorem}\label{thm:61} Notation as above, where we assume that $a,b$ are selfadjoint and freely independent, and we put $z_a = z_{a^{*}} = z_b = z_{b^{*}} =: z$. Define \begin{align*} F_a=\begin{bmatrix} f_{a,a} & f_{a,a^*} \\ f_{a^*,a} & f_{a^*,a^*} \end{bmatrix},\quad F_b=\begin{bmatrix} f_{b,b} & f_{b,b^*} \\ f_{b^*,b} & f_{b^*,b^*} \end{bmatrix} \end{align*} and \begin{align*} H_a&=\begin{bmatrix} f_{bb} (1-f_{b^*b})^{-1} & f_{b,b^*}+f_{b,b} (1-f_{b^*,b})^{-1} f_{b^*,b^*}\\ (1-f_{b^*,b})^{-1} & (1-f_{b^*,b})^{-1} f_{b^*,b^*} \end{bmatrix}, \\ H_b&=\begin{bmatrix} (1-f_{a,a^*})^{-1}f_{a,a} & (1-f_{a,a^*})^{-1} \\f_{a^*,a}+f_{a^*,a^*} (1-f_{a,a^*})^{-1} f_{a,a} & f_{a^*,a^*}(1-f_{a,a^*})^{-1} \end{bmatrix}. \end{align*} Then one has \begin{align}\label{eqn:acPowerSeries} \begin{cases} F_aH_a&=\eta_a(z H_a),\\ F_bH_b&=\eta_b(z H_b). \end{cases} \end{align} where $\eta_a$ and $\eta_b$ are the $\eta$-series of $a$ and of $b$ (as reviewed in Notation \ref{def:27b}). \end{theorem} \begin{proof} The main tool we will use in order to establish the relations stated in \eqref{eqn:acPowerSeries} is the VNRP property from Theorem \ref{thm:141}. We will only prove the first of the two relations, as the proof of the second one is analogous. Note that, even though our current hypotheses are such that $(ab)^*=b^*a^*=ba$, we will nevertheless continue to allow for occurrences of $a^*$ and $b^{*}$, which we will use to distinguish between $a$'s and $b$'s coming from a product $ab$ versus $a$'s and $b$'s coming from a product $ba$. For $l,l^\prime\in\{a,a^*\}$ consider $f_{l,l^\prime}$. Each term of $f_{l,l^\prime}$ is a joint Boolean cumulant of $a,a^*,b,b^*$ which starts with $l$ and ends with $l^\prime$. According to Theorem \ref{thm:141} all partitions in the expansion have a unique outer block. We will sort the terms of $f_{l,l^\prime}$ according to the structure of this outer block. So let us fix a possibility for what the outer block could be -- this has to start with $l$, has to end with $l^\prime$, and must contain only $a$'s and/or $a^*$'s. By $h_{k,k^\prime}$ we will denote the power series which occurs between consecutive $k$ and $k^\prime$ in the outer block. Then we have \begin{align*} f_{l,l^\prime}=\sum_{n=1}^{\infty} \sum_{\substack{ w(1),\ldots,w(n) \in \{a,a^*\}\\ with \ w(1)=l,w(n)=l^\prime}} \beta_n\left(w(1),\ldots,w(n)\right) z_{w(1)}h_{w(1),w(2)} z_{w(2)} \cdots \end{align*} \[ \hfill \cdots z_{w(n-1)} h_{w(n-1),w(n)} z_{w(n)}. \] For example we have \begin{align*} f_{a,a}=\beta_1(z) z_a+\beta_2(a,a)z_a h_{a,a} z_a+\beta_3(a,a^*,a)z_a h_{a,a^*} z_{a^*} h_{a^*,a} z_a+\ldots \end{align*} Consider then the matrix $\widetilde{H}_a$ defined by \begin{align*} \widetilde{H}_a := \begin{bmatrix} h_{a,a} & h_{a,a^*} \\ h_{a^*,a} & h_{a^*,a^*} \end{bmatrix}. \end{align*} Since we assume $a=a^*$ the relation between $f_{l,l^\prime}$ and $h_{k,k^\prime}$ can be written on the level of $2 \times 2$ matrices in the form \begin{align*} F_a=\sum_{n=1}^{\infty}\beta_n(a)\begin{bmatrix} z_a&0\\0&z_{a^*} \end{bmatrix}\left(\widetilde{H}_a \begin{bmatrix} z_a&0\\0&z_{a^*} \end{bmatrix}\right)^{n-1}. \end{align*} We assumed that $z_a=z_{a^*}=z$, thus multiplying both sides by $\widetilde{H}_a$ immediately gives \begin{align*} F_a \widetilde{H}_a&=\eta_a(z \widetilde{H}_a). \end{align*} We are left to prove that the matrix $\widetilde{H}_a$ is in fact the same as the $H_a$ defined in the statement of the theorem, i.e. that the series $h_{l,l'}$ which have appeared in the discussion are such that \begin{align*} \begin{bmatrix} h_{a,a} & h_{a,a^*} \\ h_{a^*,a} & h_{a^*,a^*} \end{bmatrix} &=\begin{bmatrix} f_{b,b} (1-f_{b^*,b})^{-1} & f_{b,b^*}+f_{b,b} (1-f_{b^*,b})^{-1} f_{b^*,b^*}\\ (1-f_{b^*,b})^{-1} & (1-f_{b^*,b})^{-1} f_{b^*,b^*} \end{bmatrix}. \end{align*} We will analyze separately each of the four entries of this matrix equality. The observation that is used in the discussion of each of the four entries is as follows: due to VNRP, it is immediate that between any two consecutive elements of the outer block we will have products of Boolean cumulants starting with $b$ or $b^*$ and ending with $b$ or $b^*$. \begin{enumerate} \item {\em Entry (1,1).} Assume that the two consecutive variables in the outer block are $a$ and $a$. Observe that the element coming right after the first $a$ must be a $b$. The element appearing immediately to the left of the second $a$ from the outer block could be $a^*$ or $b$, but if it was $a^*$, then by VNRP this $a^*$ would be in the outer block and thus the consecutive elements considered in the outer block would be $a$ and $a^*$ (rather than $a$ and $a$). We thus conclude that the element appearing immediately to the left of the second $a$ from the outer block is a $b$. In order to not violate VNRP between $a$ and $a$ we can get a cumulant $\beta_{n_0}(b,\ldots,b)$ or for some $k>1$ we can get $\beta_{n_0}(b,\ldots,b)\left(\prod_{i=0}^k\beta_{n_i}(b^*,\ldots,b)\right)$ for $n_i\geq 2$. In $\eta_{n_0}(b,\ldots,b)$ and $\eta_{n_i}(a,\ldots,b^*)$ there are $a$ and $a^*$ as arguments but by Theorem \ref{thm:141} writing the term as a joint cumulant gives exactly all terms with VNRP property. We conclude that in each $a,a$ pocket we can get any term of the power series $f_{bb} (1-f_{b^*b})^{-1}$ \item {\em Entry (1,2).} Assume that the two consecutive variables in the outer block are $a$ and $a^*$. In this case we find, by a similar argument as above, that between $a$ and $a^*$ one can get one block of the form $\beta_{m}(b,\ldots,b^*)$ for $m\geq 2$ or if there are more blocks they are of the form $\beta_{n_0}(b,\ldots,b)\prod_{i=0}^k\beta_{n_i}(b^*,\ldots,b) \beta_{n_{k+1}}(b^*,\ldots,b^*)$ for $k\geq 0$ and $n_0,n_{k+1}\geq 1$ and $n_i\geq 2$ for $i=1,\ldots,k$. Thus in each $a, a^*$ pocket we get the power series $f_{b,b} (1-f_{b^*,b})^{-1} f_{b^*,b^*}$ \item {\em Entry (2,1).} Assume that the two consecutive variables in the outer block are $b^*$ and $b$. Then similar analysis shows that in each pocket we can get $\left(1-f_{b^*, b}\right)^{-1}$. \item {\em Entry (2,2).} Assume that the two consecutive variables in the outer block are $a^*$ and $a^*$. Then similar analysis shows that in each pocket we can get $\left(1-f_{b^*, b}\right)^{-1}f_{b^*, b^*}$. \end{enumerate} \end{proof} \begin{remark} (1) Suppose that $( {\mathcal A} , \varphi )$ is a $C^{*}$-probability space. In this case the $\eta$--series of $a$ and $b$ are convergent power series around zero and the system of equations \eqref{eqn:acPowerSeries} can be solved for analytic functions in some neighbourhood of zero. (2) By specializing to the case when $a$ and $b$ have the same distribution, one immediately obtains Theorem \ref{thm:161} of the Introduction. Stated in a bit more detail, this goes as follows. \end{remark} \begin{corollary} \label{cor:62} Consider the setting of Theorem \ref{thm:61}, where we make the additional assumption that $a$ and $b$ have the same distribution. Then one has \begin{align*} &f_{b,b}=f_{a^*,a^*},\quad f_{b,b^*}=f_{a^*,a},\\ &f_{b^*,b}=f_{a,a^*},\quad f_{b^*,b^*}=f_{a,a}. \end{align*} The system of equations \eqref{eqn:acPowerSeries} reduces to a single equation, \begin{align}\label{eqn:acSameDistr} F_aH_a&=\eta_a(z H_a), \end{align} where \begin{align*} H_a&=\begin{bmatrix} f_{a^*,a^*} (1-f_{a,a^*})^{-1} & f_{a^*,a}+f_{a^*,a^*} (1-f_{a,a^*})^{-1} f_{a,a}\\ (1-f_{a,a^*})^{-1} & (1-f_{a,a^*})^{-1} f_{a,a} \end{bmatrix}. \end{align*} Moreover, in this case the formula for the $\eta$--series of $ab+ba$ also simplifies, and from Equation \eqref{eqeta} one gets that \begin{align}\label{eqn:etaSamDist} \eta_{ab+ba}(z^2)=2\left(f_{a,a^*}(z)+\frac{f_{a,a}(z)f_{a^*,a^*}(z)}{1-f_{a^*,a}(z)}\right). \end{align} $\ $ \hfill $\square$ \end{corollary} \vspace{10pt} \subsection{The special case of symmetric distributions.} $\ $ \noindent As explained in Remark \ref{rem:156}, the calculation of the distribution of a free anticommutator $ab+ba$ is much more approachable in the case when $a$ and $b$ have symmetric distributions. This fact also manifests itself in the framework of Theorem \ref{thm:61}, where the hypothesis that $a$ and $b$ have symmetric distributions leads to the simplified statement of the next proposition. \begin{proposition} \label{prop:64} Consider the framework and notation of Theorem \ref{thm:61}, and let us also assume that $a$ and $b$ have symmetric distributions. Then the power series appearing in Theorem \ref{thm:61} are such that $f_{a,a}=f_{a^*,a^*}=f_{b,b}=f_{b^*,b^*}=0$, and the matrices $H_a$ and $H_b$ from that theorem become: \[ H_a = \begin{bmatrix} 0 & f_{b,b^*} \\ (1-f_{b^*,b})^{-1} & 0 \end{bmatrix}, \ \ \ H_b = \begin{bmatrix} 0 & (1-f_{a,a^*})^{-1} \\ f_{a^*,a} & 0 \end{bmatrix} . \] The system of equations \eqref{eqn:acPowerSeries} simplifies to \begin{align}\label{eqn:SymAcPowerSeries} \begin{cases} \begin{bmatrix} f_{a,a^*}(1-f_{b^*,b})^{-1} & 0 \\0 & f_{a^*,a}f_{b,b^*} \end{bmatrix}&=\eta_a(z H_a),\\ \\ \begin{bmatrix} f_{b,b^*}f_{a^*,a} & 0 \\0 & (1-f_{a,a^*})^{-1}f_{b^*,b} \end{bmatrix}&=\eta_b(z H_b), \end{cases} \end{align} and one has $\eta_{ab+ba}(z^2)=f_{a,a^*}(z)+f_{b^*,b}(z)$. \end{proposition} \begin{proof} We first note that from the fact that $\varphi (a^{2n-1}) = \varphi (b^{2n-1})=0$ for all $n \in {\mathbb N} $ and from the formulas connecting Boolean cumulants to moments we get $\beta_{2n-1}(a)=\beta_{2n-1}(b)=0$ for all $n \in {\mathbb N} $. Now, the coefficients of the power series $f_{a,a},f_{a^*,a^*},f_{b,b},f_{b^*,b^*}$ are odd length joint Boolean cumulants of $a$ and $b$, thus each of them contains either an odd number of $a$'s or an odd number of $b$'s. From Theorem \ref{thm:141} one gets that every such joint cumulant is zero (since upon writing it as a sum of products in the way indicated by Theorem \ref{thm:141}, each term will contain an odd length Boolean cumulant of $a$ or of $b$). Thus we get $f_{a,a}=f_{a^*,a^*}=f_{b,b}=f_{b^*,b^*}=0$, and the claims of the proposition then follow from the general formulas obtained in Theorem \ref{thm:61}. \end{proof} \begin{corollary} \label{cor:65} In the framework of Proposition \ref{prop:64}, assume moreover that $a$ and $b$ have the same distribution. Then the Equations \eqref{eqn:SymAcPowerSeries} of Proposition \ref{prop:64} further simplify to \begin{align}\label{eqn:AcSymSameDistr} \begin{bmatrix} f_{a,a^*}(1-f_{a,a^*})^{-1} & 0 \\0 & f_{a^*,a}^2 \end{bmatrix}&=\eta_a(z H_a), \end{align} where \begin{align*} H_a&=\begin{bmatrix} 0 & f_{a^*,a}\\ (1-f_{a,a^*})^{-1} & 0 \end{bmatrix}. \end{align*} Moreover, in this case one gets that $\eta_{ab+ba}(z^2)=2f_{a,a^*}(z)$. \hfill $\square$ \end{corollary} $\ $ While the examples of free anticommutators of symmetric distributions can be handled with the methods from \cite{NiSp1998}, we nevertheless discuss two such examples, mostly in order to point out that Theorem \ref{thm:154} can be used (in some sense) in reverse, for getting corollaries about the enumeration of ac-friendly non-crossing partitions. \begin{example} \label{ex:66} Suppose that $p$ and $q$ are two free projections in a $*$-probability space $( {\mathcal A} , \varphi )$, such that $\varphi (p) = \varphi (q) = 1/2$, and let $a := 2p-1, b = 2q-1$. Then $a$ and $b$ are as in Corollary \ref{cor:65}, where the common distribution of $a$ and $b$ is the symmetric Bernoulli distribution $\frac{1}{2} ( \delta_{-1} + \delta_1 )$. The distribution of $ab+ba$ can be computed by using Corollary \ref{cor:65}; but this special example is actually much easier to handle, since it is immediate that $u := ab$ is a Haar unitary in $( {\mathcal A} , \varphi )$, which implies by direct calculation (cf. Example 1.14 in Lecture 1 of \cite{NiSp2006}) that $ab + ba = u + u^{*}$ has the arcsine distribution with density $( \pi \sqrt{4-t^2} )^{-1}$ on the interval $[-2,2]$. On the other hand, one can also look at what Theorem \ref{thm:154} has to say in connection to this example, and this leads to the corollary stated next. For this corollary, recall that a non-crossing partition $\sigma \in NC(2n)$ is said to be a pairing when every block $V$ of $\sigma$ has $|V| = 2$. The set of all non-crossing pairings in $NC(2n)$ is denoted by $NC_2 (2n)$. It is well-known that the cardinality of $NC_2 (2n)$ is equal to $ \mathrm{Cat} _n$, the same Catalan number which counts all the non-crossing partitions in $NC(n)$. \end{example} \begin{corollary} \label{cor:67} For every $m \in {\mathbb N} $, one has that $| \, NC_{\mathrm{ac-friendly}} (4m) \cap NC_2 (4m) \, | = \mathrm{Cat} _{m-1}$ and that $ NC_{\mathrm{ac-friendly}} (4m-2) \cap NC_2 (4m-2) = \emptyset$. \end{corollary} \begin{proof} Take $a$ and $b$ as in the preceding example. It is an easy exercise to check that the common Boolean cumulants $( \lambda_n )_{n=1}^{\infty}$ for $a$ and for $b$ come out as $\lambda_2 = 1$ and $\lambda_n = 0$ for all $n \neq 2$. Equation (\ref{eqn:155a}) from Remark \ref{rem:155}(2) thus tells us that \[ \beta_n (ab+ba) = 2 \cdot | \, NC_{\mathrm{ac-friendly}} (2n) \cap NC_2 (2n) \, |, \ \ \forall \, n \in {\mathbb N} . \] On the other hand, the direct calculation of the $\eta$-series of the arcsine distribution gives us that $\beta_n (ab+ba) = 2 \, \mathrm{Cat} _{(n-2)/2}$ when $n$ is even, and that $\beta_n (ab+ba) = 0$ when $n$ is odd, which leads to the formulas stated in the corollary. \end{proof} \begin{remark} \label{rem:68} Let $m$ be a positive integer. It is easy to easy that if $\pi \in NC_2 (2m)$ and if $\pi$ has $\{ 1, 2m \}$ as a pair, then the natural process of ``doubling'' the pairs of $\pi$ leads to a pairing in $ NC_{\mathrm{ac-friendly}} (4m)$. This construction produces $ \mathrm{Cat} _{m-1}$ examples of pairings in $ NC_{\mathrm{ac-friendly}} (4m)$, and the above corollary tells us that all the pairings in $ NC_{\mathrm{ac-friendly}} (4m)$ are obtained in this way. \end{remark} \begin{example} \label{ex:69} Suppose we want to repeat the trick from Corollary \ref{cor:67} in order to calculate the number of partitions $\sigma \in NC_{\mathrm{ac-friendly}} (2n)$ with the property that every block $V \in \sigma$ has even cardinality. To this end we now start with two selfadjoint elements $a, b$ in a $*$-probability space $( {\mathcal A} , \varphi )$ such that $a$ is free from $b$ and such that both $a$ and $b$ have distribution \begin{equation} \label{eqn:69a} \frac{1}{4}\delta_{-\sqrt{2}}+\frac{1}{2}\delta_{0}+ \frac{1}{4}\delta_{\sqrt{2}} \end{equation} with respect to $\varphi$. The reason for choosing to use the distribution in (\ref{eqn:69a}) is that its $\eta$-series is $z^2 / (1-z^2)$, which makes the common sequence $( \lambda_n )_{n=1}^{\infty}$ of Boolean cumulants for $a$ and for $b$ to be given by \begin{equation} \label{eqn:69b} \lambda_n = \left\{ \begin{array}{ll} 1, & \mbox{if $n$ is even,} \\ 0, & \mbox{if $n$ is odd.} \end{array} \right. \end{equation} When applying Corollary \ref{cor:65} to the $a$ and $b$ of the present example, the system of equations presented in \eqref{eqn:AcSymSameDistr} becomes \begin{align*} \begin{cases} f_{a,a^*}(1-z^2f_{a^*,a}(1-f_{a,a^*})^{-1})&=z^2 f_{a^*,a}\\ f_{a^*,a}(1-z^2f_{a^*,a}(1-f_{a,a^*})^{-1})&=z^2 (1-f_{a,a^*})^{-1}. \end{cases} \end{align*} Upon some further processing, we find that the series $f_{a,a^{*}}$ satisfies the equation \begin{align*} f_{a,a^*}(z)(1-f_{a,a^*}(z))^3=z^4. \end{align*} Lagrange inversion formula gives \begin{align*} f_{a,a^*}(z) = \sum_{n=1}^\infty \frac{3}{4n-1} {4n-1 \choose n-1} z^{4n}, \end{align*} and (in view of the formula $\eta_{ab+ba}(z^2)=2f_{a,a^*}(z)$ from Corollary \ref{cor:65}) we come to the conclusion that the $\eta$-series of $ab+ba$ is \begin{equation} \label{eqn:69c} \eta_{ab+ba}(z) = 2 \sum_{n=1}^\infty \frac{3}{4n-1} {4n-1 \choose n-1} z^{2n}. \end{equation} When we look at what Theorem \ref{thm:154} has to say in this particular case, we obtain the corollary stated next. In the corollary we will use the notation \[ NC^{ (\mathrm{even}) } (2n) := \{ \sigma \in NC (2n) \mid \mbox{ every block $V \in \sigma$ has even cardinality} \}. \] \end{example} \begin{corollary} \label{cor:610} For every $m \in {\mathbb N} $, one has that \[ | \, NC_{\mathrm{ac-friendly}} (4m) \cap NC^{(\mathrm{even})} (4m) \, | = \frac{3}{4m-1} {4m-1 \choose m-1} \] and that $ NC_{\mathrm{ac-friendly}} (4m-2) \cap NC^{(\mathrm{even})} (4m-2) = \emptyset$. \end{corollary} \begin{proof} Take $a$ and $b$ as in the preceding example. Equation (\ref{eqn:155a}) from Remark \ref{rem:155}(2) (used in conjunction to the formula for $\lambda_n$'s found in (\ref{eqn:69b})) tells us that \[ \beta_n (ab+ba) = 2 \cdot | \, NC_{\mathrm{ac-friendly}} (2n) \cap NC^{(\mathrm{even})} (2n) \, |, \ \ \forall \, n \in {\mathbb N} . \] On the other hand, $\beta_n (ab+ba)$ is obtained by extracting the coefficient of $z^n$ in the equality of power series which appeared in (\ref{eqn:69c}). This immediately leads to the formulas stated in the corollary. \end{proof} \vspace{10pt} \subsection{A non-symmetric example.} $\ $ \noindent In this subsection we look at the example where $a$ and $b$ have distribution $\frac{1}{2} (\delta_0+ \delta_2)$. This example offers a very good illustration of how one gets to have different distributions for the free commutator and anticommutator. The commutator $i(ab-ba)$ has exactly the same arcsine distribution as in Example \ref{ex:66}; indeed, the $a,b$ of the current example are obtained by adding $1$ to the $a,b$ of Example \ref{ex:66}, and the translation by $1$ does not affect the commutator. The anti-commutator $ab+ba$ turns out to have a different distribution, as stated in the next proposition. (Recall that the graph of the density $f(x)$ indicated in the proposition was shown in Figure 2 at the end of the Introduction, together with a histogram of eigenvalues of random matrix approximation.) \begin{proposition} \label{prop:611} Notations as above, with $a,b$ free and having distribution $\frac{1}{2} (\delta_0+ \delta_2)$. Then the distribution of $ab+ba$ is the absolutely continuous measure on the interval $[-1,8]$ which has density $f(x)$ described as follows: \begin{align*} f(x)=\begin{cases} & \frac{\sqrt{2}}{\pi}\frac{\sqrt{-1-\sqrt{\frac{x-8}{x}}-\frac{4}{x}}}{8-3\sqrt{(x-8 )x}-x} \quad\quad\mbox{ for } x \in (-1,0), \\ & \\ & \frac{1}{\pi}\frac{\sqrt{x(4\sqrt{1+x}-x-4)} +3\sqrt{(8-x)(4\sqrt{1+x}+x+4)} -8\sqrt{\frac{4\sqrt{1+x}-4-x}{x}}}{8(8-x)(1+x)} \quad\mbox{for } x \in{(0,8).} \end{cases} \end{align*} \end{proposition} \begin{proof} We have $\eta_a (z) = z/(1-z)$, hence Equation \eqref{eqn:acSameDistr} amounts here to \begin{align}\label{eqn:acBinomial} F_a=z (1-z H_a)^{-1}, \end{align} where $(1-z H_a)^{-1}$ can be written explicitly as \begin{align*} \frac{1}{z^2f_{a^*,a} +z(f_{a,a}+f_{a^*,a^*})+f_{a,a^*}-1} \begin{bmatrix} z f_{a,a}+f_{a,a^*}-1 & z f_{a^*,a}(f_{a,a^*}-1)-z f_{a,a} f_{a^*,a^*} \\ -z & z f_{a^*,a^*}+f_{a,a^*}-1 \end{bmatrix}. \end{align*} When solving Equation \eqref{eqn:acBinomial}, one gets 6 solutions. However, when plugged into the formula for $\eta_{ab+ba}(z^2)$ given in Corollary \ref{cor:62}, only one of the 6 solutions satisfies the condition that $\eta_{ab+ba}(z^2)$ has a double zero at $z=0$. After substituting $z$ by $\sqrt{z}$ in this solution, we come to the conclusion that \begin{equation} \label{eqn:611x} \eta_{ab+ba}(z) = 1 - \sqrt{ (1-8z) \frac{1-2z-\sqrt{1-8z}}{2z} }. \end{equation} From the latter formula for the $\eta$-series, a routine calculation takes us to the Cauchy transform of the distribution of $ab+ba$, which is \begin{align*} G_{ab+ba}(z)=\frac{\sqrt{2}\sqrt{1+\sqrt{\frac{z-8}{z}} +\frac{4}{z}}}{8+3\sqrt{(z-8)z}-z}. \end{align*} Finally, by using the Stieltjes inversion formula on $G_{ab+ba} (z)$, we find the form of the density $f(x)$ which was stated in the proposition. \end{proof} Same as the examples from the preceding subsection, the example considered here has a combinatorial significance, and can be used to infer the formula indicated in Equation (\ref{eqn:15a}) of the Introduction for the generating series of cardinalities of sets of ac-friendly partitions. \begin{corollary} \label{cor:612} The generating function for cardinalities of sets $ NC_{\mathrm{ac-friendly}} (2n)$ is \[ \sum_{n=1}^{\infty} | \, NC_{\mathrm{ac-friendly}} (2n) \, | z^n = \frac{1}{2} - \sqrt{ (1-8z) \frac{1-2z-\sqrt{1-8z}}{8z} }. \] \end{corollary} \begin{proof} All the Boolean cumulants of $a$ and of $b$ in this example are equal to $1$. As a consequence of this, Equation (\ref{eqn:155a}) from Remark \ref{rem:155}(2) simply says that \[ \beta_n (ab+ba) = 2 \cdot | \, NC_{\mathrm{ac-friendly}} (2n) \, |, \ \ \forall \, n \in {\mathbb N} . \] The generating function for the cardinalities of $ NC_{\mathrm{ac-friendly}} (2n)$ is thus equal to $\eta_{ab+ba} (z) /2$. We substitute this into the formula for $\eta_{ab+ba} (z)$ which appeared in Equation (\ref{eqn:611x}) during the proof of the preceding proposition, and the corollary follows. \end{proof} $\ $

https://arxiv.org/abs/math/0507390

Trends in Topological Combinatorics

This thesis opens with an introductory discussion, where the reader is gently led to the world of topological combinatorics, and, where the results of this Habilitationsschrift are portrayed against the backdrop of the broader philosophy of the subject. That introduction is followed by 5 chapters, where the main body of research is presented, and 4 appendices, where various standard tools and notations, which we use throughout the text, are collected.

\chapter*{PART I \\[2cm] Combinatorial Structures in \\[0.1cm] Topology and Geometry} \addcontentsline{toc}{chapter}{PART I. Combinatorial Structures in Topology and Geometry} \vspace{9cm} \mbox{ }\hfill \begin{minipage}{7.3cm} { Now entertain conjecture of a time}\\%[0.1cm] { When creeping murmur and the poring dark}\\%[0.1cm] { Fills the wide vessel of the universe.}\\[0.2cm] -William Shakespeare, {\it Henry V} \end{minipage} \clearemptydoublepage \include{mchap1} \clearemptydoublepage \include{mchap2} \clearemptydoublepage \chapter*{PART II \\[2cm] Complexes of Trees \\[0.1cm] and Quotient Constructions} \addcontentsline{toc}{chapter}{PART II. Complexes of Trees and Quotient Constructions} \vspace{9cm} \mbox{ }\hfill \begin{minipage}{7.6cm} { In the reproof of chance}\\%[0.1cm] { Lies the true proof of men}\\[0.2cm] -William Shakespeare, {\it Troilus and Cressida} \end{minipage} \clearemptydoublepage \include{mchap3} \clearemptydoublepage \include{mchap4} \clearemptydoublepage \include{mchap5} \chapter{Combinatorial Tools} \section{Number and set partitions \noindent Let $n$ be a~natural number. We denote the set $\{1,\dots,n\}$ by $[n]$. \begin{df} A~{\bf number partition} of~$n$ is a~set $\{\lambda_1,\dots,\lambda_t\}$ of natural numbers, such that $\lambda_1+\dots+\lambda_t=n$. \end{df} The usual con\-ven\-tion is to write $\lambda=(\lambda_1,\dots,\lambda_t)$, where $\lambda_1\geq\dots\geq\lambda_t$, and $\lambda\vdash n$. The {\it length} of $\lambda$, denoted $l(\lambda)$, is the number of components of $\lambda$, say, in the previous sentence $l(\lambda)=t$. We also use the power notation: $(n^{\alpha_n},\dots,1^{\alpha_1})=(\underbrace {n,\dots,n}_{\alpha_n},\dots,\underbrace{1,\dots,1}_{\alpha_1})$. \begin{df} We say that $\pi$ is an {\bf ordered set partition} of $[n]$ with $m$ parts (sometimes called {\it blocks}) when $\pi=(\pi_1,\dots,\pi_m)$, $\pi_i\neq\emptyset$, $[n]=\cup_{i=1}^m\pi_i$, and $\pi_i\cap\pi_j=\emptyset$, for $i\neq j$. If the order of the parts is not specified, then $\pi$ is just called a~{\bf set partition}. \end{df} We denote the set of all set partitions, resp.\ ordered set partitions, of a~set $A$ by $P(A)$, resp.\ $OP(A)$. For $P([n])$, resp.\ $OP([n])$, we use the shorthand notations $P(n)$, resp.\ $OP(n)$. Furthermore, for every set $A$, we let $\mbox{un}\,:OP(A)\rightarrow P(A)$ be the map which takes the ordered partition to the associated unordered partition. Whenever we write $\pi\vdash[n]$, it implicitly implies that $\pi$ is a~set partition, as opposed to a~number partition. A~set partition $\pi\vdash[n]$, $\pi=(S_1,\dots,S_t)$, is said to have {\it type} $\lambda$, where $\lambda\vdash n$ is the~number partition $\lambda=\{|S_1|,\dots,|S_t|\}$. \begin{df} \mbox{ } \noindent (1) For two set partitions $\pi,\tilde\pi\vdash S$, $\pi=(S_1,\dots,S_t)$, $\tilde\pi= (\tilde S_1,\dots,\tilde S_q)$ we write $\pi\vdash\tilde\pi$, and say that $\pi$ {\bf refines} $\tilde\pi$, if there exists $\iota\vdash[t]$, $\iota=\{I_1,\dots,I_q\}$, such that $\tilde S_i= \cup_{j\in I_i} S_j$, for $i\in[q]$. \noindent (2) Analogously, for two number partitions $\lambda=(\lambda_1,\dots,\lambda_t)$, $\mu=(\mu_1,\dots,\mu_q)$ we write $\lambda\vdash\mu$, and say that $\lambda$ {\bf refines} $\mu$, if there exists $\iota\vdash[t]$, $\iota=\{I_1,\dots,I_q\}$, such that $\mu_i= \sum_{j\in I_i}\lambda_j$, for $i\in[q]$. \end{df} \noindent Clearly $\pi\vdash[n]$ and $\lambda\vdash n$ are special cases of these notations. Finally observe that if $\pi,\tilde\pi$ are two set partitions, such that $\pi\vdash\tilde\pi$, then $(\mbox{type }\pi)\vdash (\mbox{type }\tilde\pi)$. \section{Graphs} \begin{df} A~{\bf directed graph} $G$ is a~pair of sets $(V(G),E(G))$ such that $E(G)\subseteq V(G)\times V(G)\setminus\{(x,x)\,|\,x\in V(G)\}$. $V(G)$ is called the~set of vertices of $G$, $E(G)$, the~set of edges of $G$. \end{df} To support the~intuition, we sometimes write $(x\rightarrow y)$ instead of $(x,y)$ and call this an {\it edge from} $x$ {\it to} $y$. A~directed graph $H=(V(H),E(H))$ is called {\it a~subgraph} of $G$ if $V(H)\subseteq V(G)$ and $E(H)\subseteq E(G)$. Furthermore, $H$ is called a {\it subgraph induced by $V(H)$} if $E(H)=E(G)\cap(V(H)\times V(H))$. For $x,y\in V(G)$, a~{\it directed path} from $x$ to $y$ is an ordered tuple $((x_1,x_2),(x_2,x_3),\dots,(x_{k-1},x_k))$, such that $(x_i,x_{i+1})\in E(G)$, for $i\in[k-1]$, and $x_1=x$, $x_k=y$. \begin{df} $\,$ a) A~directed graph $G$ is called a~{\bf directed tree} with root $x\in V(G)$ if for every $y\in V(G)$ there is a~unique directed path from $x$ to $y$. b) A~directed graph $G$ is called a~{\bf directed forest} if there exists a~decomposition $V(G)=\uplus_{i\in I}A_i$, (where $\uplus$ means disjoint union), such that each subgraph induced by $A_i$, for $i\in I$, is a~directed tree, and there are no edges between $A_i$ and $A_j$ for $i\neq j$. \end{df} If $G$ is a~directed forest, $x\in V(G)$, and there are no edges $(y\rightarrow x)$, for $y\in V(G)$, we say that $x$ is a {\it root}. \begin{df} We say that $x\in V(G)$ is a~{\bf complete source} of $G$ if $(x\rightarrow y)\in E(G)$ for all $y\in V(G)\setminus\{x\}$. \end{df} \chapter{Posets and Related Topological Constructions} \section{Basic notions} \noindent All posets discussed in this thesis are finite. \begin{df} A poset~${\cal L}$ is called a {\bf meet-semilattice} if any two elements \linebreak $x,y\,{\in}\,{\cal L}$ have a greatest lower bound, i.e., the set $\{z\,{\in}\,{\cal L}\,|\,z\,{\leq}\,x, z\,{\leq}\,y\}$ has a~ma\-xi\-mal element, called the {\em meet}, $x\,{\wedge}\,y$, of~$x$ and~$y$. \end{df} For a~subset $A\,{=}\,\{a_1,\ldots,a_t\}\subseteq{\cal L}$ we let $\bigwedge A\,{=}\, a_1\,{\wedge}\,\ldots \,{\wedge}\,a_t$ denote the unique greatest lower bound of $A$, called the~{\em meet}. In particular, meet-semilattices have a unique minimal element denoted~$\hat 0$. Minimal elements in ${\cal L}\,{\setminus}\,\{\hat 0\}$ are called the {\em atoms} in~${\cal L}$. Symmetrically, meet-semilattices share the following property: for any subset $A\,{=}\,\{a_1,\ldots,a_t\}\,{\subseteq}\,{\cal L}$ the set $\{x\,{\in}\,{\cal L} \,|\, x\,{\geq}\, a \mbox{ for all } a\,{\in}\, A\}$ is either empty or it has a unique minimal element, denoted $\bigvee A\,{=}\, a_1\,{\vee}\,\ldots \,{\vee}\, a_t$, called the~{\em join} of~$A$. If the meet-semilattice needs to be specified, we write $(\bigvee A)_{{\cal L}}\,{=}\, (a_1\,{\vee}\,\ldots \,{\vee}\, a_t)_{{\cal L}}$ for the join of~$A$ in~${\cal L}$. For brevity, we talk about semilattices throughout the second chapter, meaning meet-semilattices. \begin{df} For arbitrary posets $P$ and $Q$, the poset $P\times Q$, called the {\bf direct product}, consists of all pairs $(p,q)$, $p\in P$, $q\in Q$, ordered by the rule: $(p,q)\leq(\tilde p,\tilde q)$ iff ($p\leq\tilde p$ and $q\leq\tilde q$). \end{df} Let $P$ be an arbitrary poset. For $x\,{\in}\,P$ set: $P_{\leq x}\,{=}\,\{y\,{\in}\,P\,{|}\, y\,{\leq}\,x\}$; $P_{<x}$, and $P_{\geq x}$, $P_{>x}$ are defined analogously. For subsets ${\cal G}\,{\subseteq}\,P$ with the induced order, and $x\in P$, we define ${\cal G}_{\leq x}\,{=}\,\{y\,{\in}\,{\cal G}\,{|}\,y\,{\leq}\,x\}$, and ${\cal G}_{<x}$ again analogously. For intervals in~$P$ we use the following standard notations: $[x,y]\,{=}\,\{z\,{\in}\, P\,|\, x\,{\leq}\,z \,{\leq}\,y\}$, $[x,y)\,{=}\,\{z\,{\in}\, P\,|\, x\,{\leq}\,z \,{<}\,y\}$, etc. We refer to~\cite[Ch.\,3]{St86} for further details. \section{Order complexes of posets} \begin{df} For a~poset $P$, let $\Delta(P)$ denote the {\bf nerve} of~$P$ viewed as a~category in the usual way: it is a~simplicial complex with $i$-dimensional simplices corresponding to chains of $i+1$ elements of $P$ (chains are totally ordered sets of elements of~$P$). In particular, vertices of~$\Delta(P)$ correspond to the elements of~$P$. We call $\Delta(P)$ the {\bf order complex} of~$P$. \end{df} The concept of the nerve of a~category goes back at least to D.\ Quillen, \cite{Qu73}, and probably even further back to G.\ Segal, \cite{Se68}. In its combinatorial guise of the order complex, it appears in the Goresky-MacPherson formula and serves as one of the main bridges between combinatorics and topology. \section{Shellability} \begin{df} A~simplicial complex $\Delta$ is called {\bf shellable} if there exists an~ordering $F_1,\dots,F_t$ of the~ma\-xi\-mal faces of $\Delta$, such that $F_{i+1}\cap (\cup_{j=1}^{i}F_j)$ is a~pure simplicial complex of dimension $\dim F_{i+1}-1$ for all $i\in [t-1]$. \end{df} Such an ordering is said to satisfy Condition~(S). Sometimes it is useful to replace Condition~(S) with an~equivalent Condition~(S$'$): for $1\leq i<k\leq t$ there exist $j\leq k$ and $x\in F_k$ such that $F_i\cap F_k\subseteq F_j\cap F_k=F_k\setminus\{x\}$. If a~simplicial complex is shellable, then $\Delta$ is homotopy equivalent to a~wedge of spheres, indexed by those simplices $F_i$, for which $F_i\cap (\cup_{j=1}^{i-1}F_j)$ is equal to the full boundary of $F_i$, and each sphere has the~dimension $\dim F_i$, for the corresponding~$i$. In particular, the representatives of cohomology classes are given by the cochains dual to these simplices. See~\cite{Bj80,Bj95} for more information on shellability. \chapter{Subspace Arrangements} \section{Definition and related constructions} \begin{df} A set ${\cal A}=\{{\cal K}_1,\dots,{\cal K}_t\}$ of affine linear subspaces in a~vector space ${\cal V}$, such that ${\cal K}_i\not\subseteq{\cal K}_j$ for $i\neq j$, is called a~{\bf subspace arrangement} in~${\cal V}$. If all the subspaces ${\cal K}_1,\dots,{\cal K}_t$ are also required to contain the origin, then the subspace arrangement is called {\bf central}. \end{df} The topological spaces which one customarily associates to a~subspace arrangement are: \begin{itemize} \item $V_{\cal A}=\bigcup_{i=1}^t{\cal K}_i$, the union of subspaces; \item ${\cal M}_{\cal A}={\cal V}\setminus V_{\cal A}$, the {\it complement} of the arrangement. \end{itemize} \begin{df} Let ${\cal A}$ be a~subspace arrangement in ${\Bbb C}^n$, and let $G$ be a~subgroup of $\mbox{\bf GL}_n({\mathbb C})$, such that $V_{\cal A}$ is invariant under the action of $G$. We say that $G$ {\bf acts on} ${\cal A}$. In that case, $\Gamma_{\cal A}^G$ denotes the one-point compactification of~$V_{\cal A}/G$. \end{df} \section{Goresky-MacPherson theorem} \noindent The intersection data of a~subspace arrangement may be represented by a~poset. \begin{df} To a~subspace arrangement ${\cal A}$ in ${\cal V}$ one can associate a~partially ordered set ${\cal L}_{\cal A}$, called the {\bf intersection semilattice} of ${\cal A}$. The set if elements of ${\cal L}_{\cal A}$ is $\{K\subseteq{\mathbb C}^n\,|\,\exists\, I\subseteq[t],\mbox{ such that } \bigcap_{i\in I} {\cal K}_i=K\}\cup\{{\cal V}\}$ with the order given by reversing inclusions: $x\leq_{{\cal L}_{\cal A}}y$ iff $x\supseteq y$. That is, the minimal element of ${\cal L}_{\cal A}$ is~${\cal V}$, also customarily denoted $\hat 0$, and the maximal element is $\bigcap_{K\in{\cal A}}K$. \end{df} The following theorem describes the cohomology groups of the complement of a~subspace arrangement in terms of the homology groups of the order complexes of the intervals in the corresponding intersection lattices. \begin{thm} (Goresky \& MacPherson, \cite{GoM88}). Let ${\cal A}$ be a~central subspace arrangement in ${\Bbb C}^n$, or in ${\Bbb R}^n$, and let ${\cal L}_{\cal A}$ denote its intersection lattice, then $$\widetilde H^i({\cal M}_{\cal A})\simeq\bigoplus_{x\in{\cal L}_{\cal A}^{\geq \hat 0}} \widetilde H_{\mbox{codim}_{{\Bbb R}}(x)-i-2}(\Delta(\hat 0,x)).$$ \end{thm} This provided another strong motivation to study the order complex construction and to develop technical tools such as lexicographic shellability, see e.g.,\ ~\cite{Bj94,Ko97}. \chapter{Topological Tools} \section{Operations on topological spaces} \noindent For a~topological space $X$, $\widetilde H_i(X)$, resp.\ $\widetilde H^i(X)$, denotes the $i$th reduced homology, resp.\ cohomology, group of $X$; while $\widetilde \beta_i(X)$ denotes the $i$th reduced Betti number of~$X$. Throughout this thesis we use the operations on topological spaces described in this subsection. \begin{df} Let $(X,x)$ and $(Y,y)$ be two pointed topological spaces. \noindent (1) The {\bf wedge} of $X$ and $Y$, denoted $X\vee Y$, is the pointed topological space $$((X\cup Y)/(x\sim y),z),$$ where the base point $z$ is given by the equivalence class of $x$ (and hence of $y$). \noindent (2) The {\bf smash product} of $X$ and $Y$, denoted $X\wedge Y$, is the pointed topological space obtained as the quotient space $X\times Y/\sim$, where the equivalence relation is given by: $(\tilde x,y)\sim(x,\tilde y)$, for any $\tilde x\in X$, $\tilde y\in Y$; with the base point being the equivalence class of $(x,y)$. \end{df} The wedge and the smash products enjoy a~variety of properties: \begin{itemize} \item they are commutative and associative; \item $X\vee\mbox{pt}\,\cong\mbox{pt}\,\vee X\cong X$; \item $X\wedge S^0\cong S^0\wedge X\cong X$; \item $S^n\wedge S^m\cong S^{n+m}$; \item $X\wedge(Y\vee Z)\cong(X\wedge Y)\vee(X\wedge Z)$. \end{itemize} Furthermore, there are important special cases. \begin{df} Let $X$ be a~topological space, the {\bf suspension} of $X$, denoted $\mbox{susp}\, X$, is the quotient topological space $X\times[0,1]/\sim$, where the equivalence relation is given by $(x_1,0)\sim(x_2,0)$, and $(x_1,1)\sim(x_2,1)$, for any $x_1,x_2\in X$. \end{df} Clearly, $\mbox{susp}\, X\cong X\wedge S^1$. \begin{df} Let $X$ and $Y$ be two topological spaces. The {\bf join} of $X$ and $Y$, denoted $X*Y$, is the quotient topological space $X\times [0,1]\times Y/\sim$, where the equivalence relation $\sim$ is given by $(x,0,y_1)\sim(x,0,y_2)$, and $(x_1,1,y)\sim(x_2,1,y)$, for any $x,x_1,x_2\in X$, and $y,y_1,y_2\in Y$. \end{df} For simplicial complexes one can use the alternative, more explicit definition of the join. \begin{df} Let $X$ and $Y$ be two simplicial complexes, then $Z=X*Y$ is the simplicial complex defined by: \begin{itemize} \item the set of vertices of $Z$ is equal to the disjoint union of the sets of vertices of $X$ and $Y$; \item the subset $\Sigma$ of the set of vertices of $Z$ is a~simplex iff $\Sigma=A\cup B$, where $A$ is a~simplex in $X$ and $B$ is a~simplex in $Y$. \end{itemize} \end{df} \section{Spectral sequences} \noindent A~spectral sequence associated with a~chain complex~$C$ and a~filtration $F$ on $C$ is a~sequence of 2-dimensional tableaux $(E_{*,*}^r)_{r=0}^\infty$, where every component $E_{k,i}^r$ is a~vector space (for simplicity we first consider only field coefficients), $E_{k,i}^r=0$ unless $k\geq -1$ and $i\geq 0$, and a~sequence of differential maps $(d^r)_{r=0}^\infty$ such that \begin{enumerate} \item[(0)] $E_{k,i}^0=F_i C_k/F_{i-1}C_k$; \item[(1)] $d^r\,:\,E_{k,i}^r\longrightarrow E_{k-1,i-r}^r$, $\forall\,k,i\in{\Bbb Z}$; \item[(2)] $E_{*,*}^{r+1}=H_*(E_{*,*}^r,d^r)$, in other words \begin{equation}\label{drdef} E_{k,i}^{r+1}=\mbox{Ker}\,(E_{k,i}^r\stackrel{d^r}{\longrightarrow} E_{k-1,i-r}^r)\Big/\mbox{Im}\,(E_{k+1,i+r}^r\stackrel{d^r} {\longrightarrow}E_{k,i}^r); \end{equation} \item[(3)] for all $k\in{\Bbb Z}$, \begin{equation}\label{final} H_k(C)=\bigoplus_{i\in{\Bbb Z}}E_{k,i}^\infty. \end{equation} \end{enumerate} \noindent{\bf Comments.} \vskip4pt \noindent 0.\ ~It follows from (0) and (2) that $E_{k,i}^1=H_k(F_i,F_{i-1})$. \vskip4pt \noindent 1.\ In the general case $E_{k,i}^\infty$ is defined using the notion of convergence of the spectral sequence. We will not explain this notion in general, since for the spectral sequence that we consider only a~finite number of components in every tableau $E_{*,*}^r$ are different from zero, so there exists $N\in\Bbb N$, such that $d^r=0$ for $r\geq N$. Then, one sets $E_{*,*}^\infty=E_{*,*}^N$, and so $H_k(C)=\bigoplus_{i\in{\Bbb Z}}E_{k,i}^N$. \vskip4pt \noindent 2.\ ~For the case of integer coefficients, \eqref{final} becomes more involved: rather than just summing the entries of $E^{\infty}_{*,*}$ one needs to solve extension problems to get $H_*(C)$. This difficulty will not arise in our applications, so we refer the interested reader to~\cite{McC85} for the detailed explanation of this phenomena. When considering integer coefficients, $E^r_{*,*}$ are not vector spaces, but just abelian groups. \vskip4pt \noindent 3.\ ~We would like to warn the reader that our indexing is different from the standard (but more convenient for our purposes). The standard indexing is more convenient for the spectral sequences associated to fibrations, an~instance we do not discuss in this thesis. \vskip4pt Spectral sequences constitute a~convenient tool for computing the homology groups of a~simplicial complex. A few good sources for a~more comprehensive further reading are~\cite{McC85,Sp66,Mas52}. \chapter{The Resonance Category} \section{Canonical stratifications of symmetric smash products} \subsection{Combinatorial stratifications of topological spaces} \label{s1.1} Complicated combinatorial problems often arise when one studies the homological properties of strata in some topological space with a~given natural stratification. The examples of such stratifications are numerous. A~very simple one is provided by taking the~$n$-fold direct product of a topological space (possibly also taking the quotient with respect to the ${\cal S}_n$-action), stratified by point coincidences. The strata are indexed by set partitions (or number partitions), and the biggest open stratum is a~configuration space (ordered or unordered), whose topological properties have been widely studied, see e.g., \cite{FH01,FZ00}. Another example is the stratification of a~vector space induced by a~subspace arrangement. The strata are all possible intersections of subspaces, they are indexed by the intersection semilattice of the arrangement. The biggest open stratum is the complement of the subspace arrangement, whose topological properties have also been of quite some interest, see e.g., \cite{Bj94,Bj95,GoM88,OS80,Vas94,Zi92,ZZ93}. Of course, in both examples above, the main objective is to study the open stratum, which is the complement of the largest closed stratum. However, it was suggested by Arnol'd in a~much more general context, see for example~\cite{Ar70a}, that in situations of this kind one should study the problem for all closed strata. The main argument in support of this point of view is that there is usually no immediate natural structure on the largest open stratum, while there is one on its complement, also known as discriminant. The structure is simply given by stratification. To put it in philosophical terms: ``There is only one way for the point in the stratified space to be good, but there are many different ways for it to be bad''. Once some information has been obtained about the closed strata, one can try to find out something about the open stratum by means of some kind of duality. After this general introduction we would like to describe the specific example which will be of particular importance for this chapter. Let $X$ be a~pointed topological space (we refer to the chosen point as a~point at $\infty$), and denote $$X^{(n)}=\overbrace{X\wedge X\wedge\dots\wedge X}^n/{\cal S}_n,$$ where $\wedge$ is the smash product of pointed spaces. In other words, $X^{(n)}$ is the set of all unordered collections of $n$ points on $X$ with the infinity point attached in the appropriate way. $X^{(n)}$ is naturally stratified by point coincidences, and the strata are indexed by the number partitions of $n$. Note that we consider the closed strata, so, for example, the stratum indexed $(\underbrace{1,1,\dots,1}_n)$ is the whole space $X^{(n)}$. If one specifies $X=S^1$, resp.~$X=S^2$, one obtains as strata the spaces of all monic real hyperbolic, resp.~monic complex, polynomials of degree~n with specified root multiplicities. These spaces naturally appear in singularity theory, \cite{AGV85}. Homological invariants of several of these strata were in particular computed by Arnol'd, Shapiro, Sundaram, Welker, Vassiliev, and the author, see~\cite{Ar70a,Ko99a,Ko00b,ShW98,SuW97,Vas98}. \subsection{The idea of the resonance category and resonance functors} The purpose of the research presented in this chapter is to take a different, more abstract look at this set of problems. More specifically, the idea is to introduce a~new canonical combinatorial object, independent of the topology of the particular space $X$, where the combinatorial aspects of these stratifications would be fully reflected. This object is a~certain category, which we name {\bf resonance category}. The word resonance stands for certain linear identities valid among parts of the indexing number partition for the particular stratum. The usage of this word was suggested to the author by B.~Shapiro,~\cite{Sh00}. Having this canonically combinatorially defined category at hand, one then can, for each specific topological space $X$, view the natural stratification of $X^{(n)}$ as a~certain functor from the resonance category to ${\bf Top}^*$. These functors satisfy a~system of axioms, which we take as a~definition of {\bf resonance functors}. The combinatorial structures in the resonance category will then project to the corresponding structures in each specific $X^{(n)}$. This opens the door to develop the general combinatorial theory of the resonance category, and then prove facts valid for all resonance functors satisfying some further conditions, such as for example acyclicity of certain spaces. As the main technical tools to unearth the combinatorial structures in the resonance category, we put forward the notions of relative resonances and direct products (most importantly of a~resonance and a~relative resonance). Intuitively one can think of the relative resonance as a~stratum with a~substratum shrunk to the infinity point. As mentioned above, to illustrate a~possible appearance of this abstract framework we choose to use a~class of topological spaces which come in particular from singularity theory, and whose topological properties have been studied: spaces of polynomials (real or complex) with prescribed root multiplicities. In particular, in case of strata $(k^m,1^t)$, which were studied in \cite{Ar70a,Ko99a,SuW97} for the complex case, and in \cite{Ko00b,ShW98} for the real case, we demonstrate how the inherent combinatorial structure of the resonance category makes this particular resonance especially ``reducible.'' \vskip4pt \noindent Here is the brief summary of the contents of this chapter. \vskip4pt \noindent {\bf Section \ref{s1.2}.} We introduce the notion of resonance category, and describe the structure of its set of morphisms. \vskip4pt \noindent {\bf Section \ref{s1.3}.} We introduce the notions of relative resonances, direct products of relative resonances, and resonance functors. \vskip4pt \noindent {\bf Section \ref{s1.4}.} We formulate the problem of Arnol'd and Shapiro which motivated this research as that concerning a~specific resonance functor. Then, we analyze the combinatorial structure of resonances $(a^k,b^l)$, which leads to the complete determination of homotopy types of the corresponding strata for $X=S^1$. \vskip4pt \noindent {\bf Section \ref{s1.5}.} We analyze the combinatorial structure of sequential resonances. For $X=S^1$, this leads to a~complete computation of homotopy types of the strata corresponding to resonances $(a^k,b^l,1^m)$, such that $a-bl\leq m$, as well as resonances consisting of powers of some number. In the case of the latter, the strata always have the homotopy type of a bouquet of spheres. We describe a~combinatorial model to enumerate these spheres as paths in a certain weighted directed graph, with dimensions of the spheres being given by the total weights of the paths. \vskip4pt \noindent {\bf Section \ref{s1.6}.} We introduce the notion of a complexity of a~resonance and give a~series of examples of resonances having arbitrarily high complexity. \section{The resonance category} \label{s1.2} \subsection{Resonances and their symbolic notation} For every positive integer $n$, let $\{-1,0,1\}^n$ denote the set of all points in ${\Bbb R}^n$ with coordinates in the set $\{-1,0,1\}$. We say that a subset $S\subseteq\{-1,0,1\}^n$ is {\bf span-closed} if ${\mbox{span}\,}(S)\cap\{-1,0,1\}^n=S$, where ${\mbox{span}\,}(S)$ is the linear subspace spanned by the origin and points in $S$. Of course the origin lies in every span-closed set. For $x=(x_1,\dots,x_n)\in\{-1,0,1\}^n$, we use the notations $\mbox{Plus\,}(x)=\{i\in[n]\,|\,x_i=1\}$ and $\mbox{Minus\,}(x)=\{i\in[n]\,|\,x_i=-1\}$. \begin{df} $\,$ \noindent (1) A subset $S\subseteq\{-1,0,1\}^n$ is called an~{\bf $n$-cut} if it is span-closed and for every $x\in S\setminus\{$origin$\}$ we have $\mbox{Plus\,}(x)\neq\emptyset$ and $\mbox{Minus\,}(x)\neq\emptyset$. We denote the set of all $n$-cuts by ${\cal R}_n$. \noindent (2) ${\cal S}_n$ acts on $\{-1,0,1\}^n$ by permuting coordinates, which in turn induces ${\cal S}_n$-action on ${\cal R}_n$. The {\bf $n$-resonances} are defined to be the orbits of the latter ${\cal S}_n$-action. We let $[S]$ denote the $n$-resonance represented by the $n$-cut $S$. \end{df} The cut or the resonance consisting of origin only is called {\it trivial}. \begin{exams}\label{exsmres} {\rm $n$-resonances for small values of $n$.} (1) There are no nontrivial 1-resonances. (2) There is one nontrivial 2-resonance: $[\{(0,0),(1,-1),(-1,1)\}]$. (3) There are four nontrivial 3-resonances: $$[\{(0,0,0),(1,-1,0),(-1,1,0)\}],$$ $$[\{(0,0,0),(1,-1,0),(-1,1,0),(1,0,-1),(-1,0,1),(0,1,-1),(0,-1,1)\}],$$ $$[\{(0,0,0),(1,-1,-1),(-1,1,1)\}],$$ $$[\{(0,0,0),(1,-1,-1),(-1,1,1),(0,1,-1),(0,-1,1)\}].$$ (4) Here is an example of a nontrivial 6-resonance: $$[\{\mbox{origin},\pm(1,1,0,-1,-1,0),\pm(0,1,1,0,-1,-1),\pm(1,0,-1,-1,0,1)\}].$$ \end{exams} \noindent {\bf Symbolic notation.} To describe an $n$-resonance, rather than to list all of the elements of one of its representatives, it is more convenient to use the following symbolic notation: we write a sequence of $n$ linear expressions in some number (between 1 and $n$) of parameters, the order in which the expressions are written is inessential. Here is how to get from such a~symbolic expression to the $n$-resonance: choose an order on the $n$ linear expressions and observe that now they parameterize some linear subspace of ${\Bbb R}^n$, which we denote by $A$. The $n$-resonance is now the orbit of $A^{\perp}\cap\{-1,0,1\}^n$. Reversely, to go from an $n$-resonance to a~symbolic expression: choose a~representative $n$-cut $S$, the symbolic expression can now be obtained as a~linear parameterization of ${\mbox{span}\,}(S)^\perp$. For example the 6 nontrivial resonances listed in the Example~\ref{exsmres} are (in the same order): $$(a,a),\,(a,a,b),\,(a,a,a),\,(a+b,a,b),\,(2a,a,a),\, (a+b,b+c,a+d,b+d,c+d,2d).$$ \subsection{Acting on cuts with ordered set partitions} From now on we assume known the terminology and notations of set partitions and ordered set partitions, as described in the Appendix~A. \begin{df} Given $\pi=(\pi_1,\dots,\pi_k)$ an ordered set partition of $[m]$ with $k$ parts, and $\nu=(\nu_1,\dots,\nu_m)$ an ordered set partition of $[n]$ with $m$ parts, their {\bf composition} $\pi\circ\nu$ is an ordered set partition of $[n]$ with $k$ parts, defined by $\pi\circ\nu=(\mu_1,\dots,\mu_k)$, $\mu_i=\cup_{j\in\pi_i}\nu_j$, for $i=1,\dots,k$. \noindent Analogously, we can define $\pi\circ\nu$ for an~ordered set partition $\nu$ and a~set partition $\pi$, in which case $\pi\circ\nu$ is a~set partition without any specified order on the blocks. \end{df} In particular, when $m=n$, and $|\pi_i|=1$, for $i=1,\dots,n$, we can identify $\pi=(\pi_1,\dots,\pi_n)$ with the corresponding permutation of $[n]$. The composition of two such ordered set partitions corresponds to the multiplication of corresponding permutations, and we denote the ordered set partition $(\{1\},\dots,\{n\})$ by $\mbox{id}_n$, or just~$\mbox{id}$. \begin{df} For $A\subseteq B$, let $p_{B,A}:P(B)\rightarrow P(A)$ denote map induced by the restriction from $B$ to $A$. For two disjoint set $A$ and $B$, and $\Pi\subseteq P(A)$, $\Lambda\subseteq P(B)$, we define $\Pi\times\Lambda=\{\pi\in P(A\cup B)\,|\,p_{A\cup B,A}(\pi)\in\Pi, p_{A\cup B,B}(\pi)\in\Lambda\}$. \end{df} The following definition provides the combinatorial constructions necessary to describe the morphisms of the resonance category, as well as to define the relative resonances. \begin{df} \label{df2.4} Assume $S$ is an~$n$-cut. \noindent (1) For an~ordered set partition of $[n]$, $\pi=(\pi_1,\dots,\pi_m)$, we define $\pi S\in{\cal R}_m$ to be the set of all $m$-tuples $(t_1,\dots,t_m)\in\{-1,0,1\}^m$, for which there exists $(s_1,\dots,s_n)\in S$, such that for all $j\in[m]$, and $i\in\pi_j$, we have $s_i=t_j$. \noindent (2) For an~unordered set partition of $[n]$, $\pi=(\pi_1,\dots,\pi_m)$, we define $S^\pi\in{\cal R}_n$ to be the subset of $S$ consisting of all $(s_1,\dots,s_n)$ such that if $i,j\in\pi_k$, $i\neq j$, for some $k=1,\dots,m$, then $s_i=s_j\neq 0$; in other words if $k=1,\dots,m$ is such that $|\pi_k|\geq 2$, then either $\pi_k\subseteq\mbox{Plus\,}(s_1,\dots,s_n)$ or $\pi_k\subseteq\mbox{Minus\,}(s_1,\dots,s_n)$. \end{df} \noindent Clearly $\mbox{id} S=S$, and one can see that $(\pi\circ\nu)S=\pi(\nu S)$. \vskip4pt \noindent {\bf Verification of $(\pi\circ\nu) S=\pi(\nu S)$.} \noindent By definition we have $$ (\pi\circ\nu)S=\{(t_1,\dots,t_k)\,|\,\exists(s_1,\dots,s_n)\in S \mbox{ s.t. }\forall j\in[k],i\in\mu_j:s_i=t_j\},$$ $$\nu S=\{(x_1,\dots,x_m)\,|\,\exists(s_1,\dots,s_n)\in S \mbox{ s.t. }\forall q\in[m],i\in\nu_q:s_i=x_q\},$$ \begin{multline*} \pi(\nu S)=\{(t_1,\dots,t_k)\,|\,\exists(x_1,\dots,x_m)\in\nu S, \\ \mbox{ such that }\forall j\in[k],q\in\pi_j,i\in\nu_q:s_i=t_j\}. \end{multline*} The identity $(\pi\circ\nu) S=\pi(\nu S)$ follows now from the equality $\mu_j=\cup_{q\in\pi_j}\nu_q$. \vskip4pt There are many different ways to formulate the Definition~\ref{df2.4}. We chose the ad hoc combinatorial language, but it is also possible to put it in the linear-algebraic terms. An ordered set partition of $[n]$, $\pi=(\pi_1,\dots,\pi_m)$, defines an inclusion map $\phi:{\Bbb R}^m\rightarrow{\Bbb R}^n$ by $\phi(e_i)= \sum_{j\in\pi_i}\tilde e_j$, where $\{e_1,\dots,e_m\}$, resp.\ $\{\tilde e_1,\dots,\tilde e_n\}$, is the standard orthonormal basis of ${\Bbb R}^m$, resp.~${\Bbb R}^n$. Given $S\in{\cal R}_n$, $\pi S$ can then be defined as $\phi^{-1}(\mbox{Im}\,\phi\cap S)$. Furthermore, $S^\pi=\mbox{Im}\,\phi|_{Z}\cap S$, where $Z$ is the set of all $(z_1,\dots,z_m)\in{\Bbb R}^m$, such that if $|\pi_k|\geq 2$, for some $k=1,\dots,m$, then $z_k\neq 0$. \subsection{The definition of the resonance category and the terminology for its morphisms} We are now ready to give the definition of the central notion of this chapter. \begin{df} The {\bf resonance category}, denoted ${\cal R}$, is defined as follows: \noindent (1) The set of objects is the set of all $n$-cuts, for all positive integers $n$, ${\cal O}({\cal R})=\cup_{n=1}^\infty {\cal R}_n$. \noindent (2) The set of morphisms is indexed by triples $(S,T,\pi)$, where $S\in{\cal R}_m$, $T\in{\cal R}_n$, and $\pi$ is an ordered set partition of $[n]$ with $m$ parts, such that $S\subseteq \pi T$. For the reasons which will become clear later we denote the morphism indexed with $(S,T,\pi)$ by $S\twoheadrightarrow\pi T\stackrel{\pi}{\hookrightarrow} T$. \noindent As the notation suggests, the initial object of the morphism $S\twoheadrightarrow\pi T\stackrel{\pi}{\hookrightarrow} T$ is $S$ and terminal object is $T$. The composition rule is defined by $$(S\twoheadrightarrow\pi T\stackrel{\pi}{\hookrightarrow}T)\circ(T\twoheadrightarrow\nu Q\stackrel{\nu}{\hookrightarrow}Q)= S\twoheadrightarrow\pi\nu Q\stackrel{\pi\nu}{\hookrightarrow} Q,$$ where $S\in{\cal R}_k$, $T\in{\cal R}_m$, $Q\in{\cal R}_n$, $\pi$ is an ordered set partition of $[m]$ with $k$ parts, and $\nu$ is an ordered set partition of $[n]$ with $m$ parts. \end{df} An~alert reader will notice that the resonances themselves did not appear explicitly in the definition of the resonance category. In fact, it is not difficult to notice that resonances are isomorphism classes of objects of ${\cal R}$. Let us now look at the set of morphisms of ${\cal R}$ in some more detail. \noindent (1) For $S\in{\cal R}_n$, the identity morphism of $S$ is $S\twoheadrightarrow S\stackrel{\mbox{id}}{\hookrightarrow} S$. \noindent (2) Let us introduce short hand notations: $S\twoheadrightarrow T$ for $S\twoheadrightarrow T\stackrel{\mbox{id}}{\hookrightarrow} T$, and $\pi T\stackrel{\pi}{\hookrightarrow}T$ for $\pi T\twoheadrightarrow \pi T\stackrel{\pi}{\hookrightarrow}T$. Then we have $$S\twoheadrightarrow\pi T\stackrel{\pi}{\hookrightarrow} T= (S\twoheadrightarrow\pi T)\circ(\pi T\stackrel{\pi}{\hookrightarrow}T).$$ Note also that $S\twoheadrightarrow S=S\stackrel{\mbox{id}}{\hookrightarrow}S$. \noindent (3) The associativity of the composition rule can be derived from the commutation relation $$(\pi S\stackrel{\pi}{\hookrightarrow}S)\circ(S\twoheadrightarrow T)= (\pi S\twoheadrightarrow\pi T)\circ(\pi T\stackrel{\pi}{\hookrightarrow} T)$$ as follows: \begin{multline} \notag (S\twoheadrightarrow\pi T\hookrightarrow T)\circ(T\twoheadrightarrow\nu Q\hookrightarrow Q)\circ(Q\twoheadrightarrow\rho X\hookrightarrow X)=\\ (S\twoheadrightarrow\pi T)\circ(\pi T\hookrightarrow T)\circ(T\twoheadrightarrow\nu Q)\circ(\nu Q\hookrightarrow Q) \circ(Q\twoheadrightarrow\rho X)\circ(\rho X\hookrightarrow X)=\\ (S\twoheadrightarrow\pi T)\circ(\pi T\twoheadrightarrow\pi\nu Q)\circ(\pi\nu Q\twoheadrightarrow\pi\nu\rho X)\circ \\ (\pi\nu\rho X\hookrightarrow \nu\rho X)\circ(\nu\rho X\hookrightarrow \rho X)\circ(\rho X\hookrightarrow X). \end{multline} \noindent (4) We shall use the following names: morphisms $S\twoheadrightarrow T$ are called {\bf gluings} (or $n$-gluings, if it is specified that $S,T\in{\cal R}_n$); morphisms $\pi T\stackrel{\pi}{\hookrightarrow}T$ are called {\bf inclusions} (or $(n,m)$-inclusions, if it is specified that $T\in{\cal R}_n$, $\pi T\in{\cal R}_m$), the inclusions are called {\bf symmetries} if $\pi$ is a~permutation. As observed above, the symmetries are the only isomorphisms in ${\cal R}$. Here are two examples of inclusions: $$\{(0,0),(1,-1),(-1,1)\}\stackrel{(\{1\},\{2,3\})}{\hookrightarrow} \{(0,0,0),(1,-1,-1),(-1,1,1)\},$$ $$\{(0,0),(1,-1),(-1,1)\}\stackrel{(\{1\},\{2,3\})}{\hookrightarrow} \{(0,0,0),\pm(1,-1,-1),\pm(0,1,-1)\}.$$ \section{Structures related to the resonance category} \label{s1.3} \subsection{Relative resonances} Let $A(n)$ denote the set of all collections of non-empty multisubsets of~$[n]$, and let $P(n)\subseteq A(n)$ be the set of all partitions of~$[n]$. For every $S\in{\cal R}_n$ let us define a~closure operation on $A(n)$, resp.\ on $P(n)$. \begin{df}\label{df3.1} Let ${\cal A}\in A(n)$. We define ${\cal A}\Downarrow S\subseteq A(n)$ to be the minimal set satisfying the following conditions: \begin{enumerate} \item[(1)] ${\cal A}\in{\cal A}\Downarrow S$; \item[(2)] if $\{B_1,B_2,\dots,B_m\}\in{\cal A}\Downarrow S$, then $\{B_1\cup B_2,B_3,\dots,B_m\}\in{\cal A}\Downarrow S$; \item[(3)] if $\{B_1,B_2,\dots,B_m\}\in{\cal A}\Downarrow S$, and there exists $x\in S$, such that $\mbox{Plus\,}(x)\subseteq B_1$, then $\{(B_1\setminus\mbox{Plus\,}(x))\cup\mbox{Minus\,}(x),B_2,\dots,B_m\}\in{\cal A}\Downarrow S$. \end{enumerate} For $\pi\in P(n)$, we define $\pi\downarrow S\subseteq P(n)$ as $\pi\downarrow S=(\pi\Downarrow S)\cap P(n)$. For a set $\Pi\subseteq P(n)$ we define $\Pi\downarrow S=\cup_{\pi\in\Pi}\pi\downarrow S$. We say that $\Pi$ is {\bf $S$-closed} if $\Pi\downarrow S=\Pi$. \end{df} The idea behind this definition comes from the context of the standard stratification of the $n$-fold symmetric product. Given a~stratum $X$ indexed by a number partition of $n$ with $m$ parts, let us fix some order on the parts. A~substratum $Y$ is obtained by choosing some partition $\pi$ of $[m]$ and summing the numbers within the blocks of $\pi$. Since the order of the parts of the number partition indexing $X$ is fixed, $X$ gives rise to a~unique $m$-cut $S$. The set $\pi\downarrow S$ describes all partitions $\nu$ of $[m]$ such that if the numbers within the blocks of $\nu$ are summed then the obtained stratum $Z$ satisfies $Z\subseteq Y$. In particular, if $Y$ is shrunk to a~point, then so is $Z$. The two following examples illustrate how the different parts of the Definition~\ref{df3.1} might be needed. \begin{exam} {\rm The equivalences of type (2) from the~Definition~\ref{df3.1} are needed.} Let the stratum $X$ be indexed by $(3,2,1,1,1)$ (fix this order of the parts), and let $\pi=\{1\}\{23\}\{4\}\{5\}$. Then, the stratum~$Y$ is indexed by $(3,3,1,1)$. Clearly, the stratum~$Z$, which is indexed by $(3,3,2)$, lies inside $Y$, hence $\{1\}\{2\}\{345\}\in\pi\downarrow S$, where $S$ is the cut corresponding to $(3,2,1,1,1)$. However, if one starts from the partition $\pi$ and uses equivalences of type (3) from the Definition~\ref{df3.1}, the only other partitions one can obtain are $\{1\}\{24\}\{3\}\{5\}$, and $\{1\}\{25\}\{3\}\{4\}$. None of them refines $\{1\}\{2\}\{345\}$, hence it would not be enough in the Definition~\ref{df3.1} to just take the partitions which can be obtained via the equivalences of type (3) and then take $\pi\downarrow S$ to be the set of all the partitions which are refined by these. \end{exam} \begin{exam} \label{exam1.3.3} {\rm It is necessary to view the equivalence relation on the larger set $A(n)$.} This time, let the stratum $X$ be indexed by $(a+b,b+c,a+d,b+d,c+d,2d)$ (fix this order of the parts, and assume as usual that there are no linear relations on the parts other than those induced by the algebraic identities on the variables $a$, $b$, $c$, and $d$). Furthermore, let $\pi=\{16\}\{23\}\{45\}$. Then the stratum $Y$ is indexed by $(a+b+2d,a+b+c+d,b+c+2d)$. Clearly, we have $\{34\}\{15\}\{26\}\in\pi\downarrow S$, where $S$ is the cut corresponding to $(a+b,b+c,a+d,b+d,c+d,2d)$. \end{exam} \noindent A natural idea for the Definition~\ref{df3.1} could have been to define the equivalence relation directly on the set $P(n)$ and use ``swaps'' instead of the equivalences of type~(3), i.e., to replace the condition (3) by: \begin{quote} {\it if $\{B_1,B_2,\dots,B_m\}\in{\cal A}\downarrow S$, and there exists $x\in S$, such that $\mbox{Plus\,}(x)\!\subseteq\! B_1$, and $\mbox{Minus\,}(x)\!\subseteq\! B_2$, then $\{(B_1\setminus\mbox{Plus\,}(x))\,\cup\,\mbox{Minus\,}(x),$ $(B_2\setminus\mbox{Minus\,}(x))\cup\mbox{Plus\,}(x),B_3,\dots,B_m\}\in{\cal A}\downarrow S$. } \end{quote} \noindent However, this would not have been sufficient as the Example~\ref{exam1.3.3} shows, since no swaps would be possible on $\pi=\{16\}\{23\}\{45\}$. \begin{df} Let $S$ be an~$n$-cut, $\Pi\subseteq P(n)$ an~$S$-closed set of partitions. We define \[ S\setminus\Pi=S\setminus\bigcup_{\pi\in\Pi}S^\pi.\] \end{df} In the next definition we give a~combinatorial analog of viewing a~stratum relative to a~substratum. \begin{df} \label{dfrel} $\,$ \noindent (1) A {\bf relative $n$-cut} is a~pair $(S,\Pi)$, where $S\subseteq\{-1,0,1\}^n$, $\Pi\subseteq P(n)$, such that the following two conditions are satisfied: \begin{itemize} \item $(\mbox{span}\, S)\setminus\Pi=S$; \item $\Pi$ is $(\mbox{span}\, S)$-closed. \end{itemize} \noindent (2) The permutation ${\cal S}_n$-action on $\{-1,0,1\}^n$ induces an~${\cal S}_n$-action on the relative $n$-cuts by $(S,\Pi)\stackrel{\sigma}{\mapsto}(\sigma S,\Pi\sigma^{-1})$, for $\sigma\in{\cal S}_n$. The {\bf relative $n$-resonances} are defined to be the orbits of this ${\cal S}_n$-action. We let $[S,\Pi]$ denote the relative $n$-resonance represented by the relative $n$-cut~$(S,\Pi)$. \end{df} When $S\in{\cal R}_n$ and $\Pi\subseteq P(n)$, $\Pi$ is $S$-closed, it is convenient to use the notation $Q(S,\Pi)$ to denote the relative cut $(S\setminus\Pi,\Pi)$. Clearly we have $(S,\Pi)=Q({\mbox{span}\,} S,\Pi)$. Analogously, $[Q(S,\Pi)]$ denotes the relative resonance $[S\setminus\Pi,\Pi]$. We use these two notations interchangeably depending on which one is more natural in the current context. The special case of the particular importance for our computations in the later sections is that of $Q(S,\pi\downarrow S)$, where $\pi$ is a~partition of $[n]$ with $m$ parts. In this case, we call $(S\setminus(\pi\downarrow S),\pi\downarrow S)$ the relative $(n,m)$-cut associated to $S$ and $\pi$. By the Definition~\ref{dfrel}, the relative cut $(S,\Pi)=((\mbox{span}\, S)\setminus\Pi,\Pi)$ consists of two parts. We intuitively think of $(\mbox{span}\, S)\setminus\Pi$ as the set of all resonances which survive the shrinking of the strata associated to the elements of $\Pi$, so it is natural to call them {\it surviving elements}. We also think of $\Pi$ as the set of all partitions whose associated strata are shrunk to the infinity point, so, accordingly, we call them {\it partitions at infinity}. \subsection{Direct products of relative resonances} \begin{df} \label{df3.3}$\,$ For relative resonances $(S,\Pi)$ and $(T,\Lambda)$ we define $$(S,\Pi)\times (T,\Lambda)= (S\times T,(\Pi\times P(m))\cup (P(n)\times\Lambda)).$$ Clearly the orbit $[(S,\Pi)\times (T,\Lambda)]$ does not depend on the choice of representatives of the orbits $[S,\Pi]$ and $[T,\Lambda]$, so we may define $[S,\Pi]\times[T,\Lambda]$ to be $[(S,\Pi)\times (T,\Lambda)]$. \end{df} \noindent The following special cases are of particular importance for our computation: \vskip4pt \noindent (1) {\bf A direct product of two resonances.} \noindent For an~$m$-cut~$S$, and an~$n$-cut~$T$, we have $S\times T=\{(x_1,\dots,x_m,y_1,\dots,y_n)\,|$ $(x_1,\dots,x_m)\in S, (y_1,\dots,y_n)\in T\}\in{\cal R}_{m+n}$, and $[S]\times[T]=[S\times T]$. \vskip4pt \noindent (2) {\bf A direct product of a relative resonance and a resonance.} \noindent For $S\in{\cal R}_n$, $\Pi\subseteq P(n)$ an $S$-closed set of partitions, and $T\in{\cal R}_k$, we have $Q(S,\Pi)\times T=Q(S\times T,\widetilde\Pi)$, where $\widetilde\Pi=\Pi\times P(\{n+1,n+2,\dots,n+k\})$, and $[Q(S,\Pi)]\times [T]=[Q(S,\Pi)\times T]$. \begin{exam} \begin{multline*} Q(\{(0,0,0),\pm(1,-1,-1),\pm(0,1,-1)\},\{1\}\{23\})=\\ \{(0)\}\times Q(\{(0,0),\pm(1,-1)\},\{12\}). \end{multline*} \end{exam} \begin{rem} One can define a~category, called {\bf relative resonance category}, whose set of objects is the set of all relative $n$-cuts. A~new structure which it has in comparison to~${\cal R}$ is provided by ``shrinking morphisms'': $(S,\Pi)\rightsquigarrow(T,\Lambda)$, for $S,T\subseteq\{-1,0,1\}^n$, $P(n)\supseteq\Lambda\supseteq\Pi$, such that $(\mbox{span}\, S)\setminus\Lambda=T$. They correspond to shrinking strata to infinity. The full definition with relations on morphisms and the corresponding combinatorial analysis, will appear in~\cite{Ko01c}. \end{rem} \subsection{Resonance functors} Given a~functor ${\cal F}:{\cal R}\longrightarrow{\bf Top}^*$, we introduce the following notation: $${\cal F}(Q(S,\Pi))={\cal F}(S)\bigg/\bigcup_{\mbox{un}\,(\pi)\in\Pi}\mbox{Im}\,{\cal F}(\pi S \stackrel{\pi}{\hookrightarrow}S).$$ \begin{df} \label{dfresf} A functor ${\cal F}:{\cal R}\longrightarrow{\bf Top}^*$ is called a~{\bf resonance functor} if it satisfies the following axioms: \begin{enumerate} \item [(A1)] {\bf Inclusion axiom.} \noindent If $S\in{\cal R}_n$, and $\pi\in OP(n)$, then ${\cal F}(\pi S \stackrel{\pi}{\hookrightarrow}S)$ is an~inclusion map, and ${\cal F}(S)\big/\mbox{Im}\,{\cal F}(\pi S\stackrel{\pi}{\hookrightarrow}S)\simeq{\cal F}(Q(S,\pi\downarrow S))$. \item [(A2)] {\bf Relative resonance axiom.} \noindent If, for some $S,T\in{\cal R}_n$, and $\Pi,\Lambda\subseteq P(n)$, $[Q(S,\Pi)]=[Q(T,\Lambda)]$, then ${\cal F}(Q(S,\Pi))\simeq{\cal F}(Q(T,\Lambda))$. \item [(A3)] {\bf Direct product axiom.} \noindent For two relative $n$-cuts $(S,\Pi)$ and $(T,\Lambda)$ we have $${\cal F}(S,\Pi)\times{\cal F}(T,\Lambda)\simeq{\cal F}((S,\Pi)\times(T,\Lambda)).$$ \end{enumerate} \end{df} \noindent Given $S\in{\cal R}_n$, and $\pi\in OP(n)$, let $i_{S,\pi}$ denote the inclusion map ${\cal F}(\pi S\stackrel{\pi}{\hookrightarrow}S)$. There is a~canonical homology long exact sequence associated to the triple \begin{equation} \label{triple} {\cal F}(\pi S)\stackrel{i_{S,\pi}}{\hookrightarrow}{\cal F}(S)\stackrel{p}{\longrightarrow} {\cal F}(Q(S,\pi\downarrow S)), \end{equation} namely \begin{multline} \label{stls} \dots\stackrel{\partial_*}{\longrightarrow}\widetilde H_n({\cal F}(\pi S)) \stackrel{(i_{S,\pi})_*}{\longrightarrow}\widetilde H_n({\cal F}(S))\stackrel{p_*}{\longrightarrow} \widetilde H_n({\cal F}(Q(S,\pi\downarrow S)))\stackrel{\partial_*}{\longrightarrow} \\ \widetilde H_{n-1}({\cal F}(\pi S))\stackrel{(i_{S,\pi})_*}{\longrightarrow}\dots \end{multline} We call \eqref{triple}, resp.~\eqref{stls}, the {\it standard triple}, resp.~the {\it standard long exact sequence} associated to the morphism $\pi S\stackrel{\pi}{\hookrightarrow}S$ and the functor ${\cal F}$ (usually ${\cal F}$ is fixed, so its mentioning is omitted). \section{First applications} \label{s1.4} \subsection{Resonance compatible stratifications} \label{ss4.1} As mentioned in the Section~\ref{s1.1} we shall now look at the natural strata of the spaces $X^{(n)}$. The strata are defined by point coincidences and are indexed by number partitions of $n$. Let $\Sigma_\lambda^X$ denote the stratum indexed by $\lambda$. Let $\lambda$ be a number partition of $n$ and let $\tilde\lambda\in OP(n)$ be $\lambda$ with some fixed order on the parts. Then $\tilde\lambda$ can be thought of as a~vector with positive integer coordinates in~${\Bbb R}^n$. Let $S_{\tilde\lambda}$ be the set $\{x\in\{-1,0,1\}^n\,|\,\langle x,\tilde\lambda\rangle =0\}$. Obviously, $S_{\tilde\lambda}$ is an~$n$-cut, and the $n$-resonance $S_\lambda$, which it defines, does not depend on the choice of $\tilde\lambda$, but only on the number partition~$\lambda$. The crucial topological observation is that if $\nu$ is another partition of $n$, such that $S_\lambda=S_\nu$, then the spaces $\Sigma_\lambda^X$ and $\Sigma_\nu^X$ are homeomorphic. This is precisely the fact which lead us to introduce resonances and the surrounding combinatorial framework and to forget about the number partitions themselves. That observation allows us to introduce a~functor ${\cal F}$ mapping $S_{\tilde\lambda}$ to $\Sigma_\lambda^X$; the morphisms map accordingly. Clearly, ${\cal F}(1^l)=X^{(l)}$. One can detect in this example the justification for the names which we chose for the morphisms of ${\cal R}$: ``inclusions'' and ``gluings''. Furthermore, it is easy, in this case, to verify the axioms of the Definition~\ref{dfresf}, and hence to conclude that ${\cal F}$ is a~resonance functor. The only nontrivial point is the verification of the second part of (A1), which we do in the next proposition. \begin{prop} \label{prop4.1} Let $S$ be an $n$-cut and $\pi\in OP(n)$. Then ${\cal F}(\nu S)\subseteq{\cal F}(\pi S)$ if and only if $\mbox{un}\,(\nu)\in\mbox{un}\,(\pi)\downarrow S$. \end{prop} \nin{\bf Proof. } It is obvious that all the steps of the definition of $\mbox{un}\,(\pi)\downarrow S$ which change the partition preserve the property ${\cal F}(\nu S)\subseteq{\cal F}(\pi S)$, hence the {\it if} direction follows. Assume now ${\cal F}(\nu S)\subseteq{\cal F}(\pi S)$. This means that there exists $\tau\in OP(m)$, where $m$ is the number of parts of $\pi$, such that ${\cal F}(\tau\pi S)={\cal F}(\nu S)$. By definition, $\tau\circ\pi\in\mbox{un}\,(\pi)\downarrow S$. Now, we can reach $\mbox{un}\,(\nu)$ from $\mbox{un}\,(\tau\circ\pi)$ by moves of type (3) from the definition of the relative resonances. Indeed, if ${\cal F}(\tau\pi S)={\cal F}(\nu S)=\Sigma_\lambda^X$, then the sizes of the resulting blocks after gluing along $\tau\circ\pi$ and along $\nu$ are the same. For every block $b$ of $\lambda$ we can go, by means of moves of type (3), from the block of $\mbox{un}\,(\tau\circ\pi)$ which glues to $b$ to the block of $\mbox{un}\,(\nu)$ which glues to $b$. Since we can do it for any block of $\lambda$, we can go from $\mbox{un}\,(\tau\circ\pi)$ to $\mbox{un}\,(\nu)$, and hence $\mbox{un}\,(\nu)\in\mbox{un}\,(\pi)\downarrow S$. $\square$ \vskip4pt \noindent In the context of this stratification the following central question arises. \vskip8pt \noindent {\bf The Main Problem.} {\it (Arnol'd, Shapiro,~\cite{Sh00}). Describe an algorithm which, for a~given resonance $\lambda$, would compute the Betti numbers of $\Sigma_\lambda^{S^1}$, or $\Sigma_\lambda^{S^2}$.} \vskip8pt The case of the strata $\Sigma_\lambda^{S^1}$ is simpler, essentially because of the following elementary, but important property of smash products: if $X$ and $Y$ are pointed spaces and $X$ is contractible, then $X\wedge Y$ is also contractible. In the subsequent subsections we shall look at a~few interesting special cases, and also will be able to say a~few things about the general problem. \subsection{Resonances $(a^k,1^l)$} \label{ss4.2} Let $a,k,l$ be positive integers such that $a\geq 2$. Let $S$ be the $(l+k)$-cut consisting of all the elements of $\{-1,0,1\}^{l+k}$, which are orthogonal to the vector $(\underbrace{1,\dots,1}_l,\underbrace{a,\dots,a}_k)$. Clearly, the~$(l+k)$-resonance $[S]$ is equal to $(a^k,1^l)$. The case $l<a$ is not very interesting, since then $(a^k,1^l)=(1^k)\times(1^l)$. Therefore we may assume that $l\geq a$. We would like to understand the topological properties of the space ${\cal F}(a^k,1^l)$. In general, this is rather hard. However, as the following theorem shows, it is possible under some additional conditions on ${\cal F}$. \begin{thm} \label{thm4.2} Let ${\cal F}:{\cal R}\longrightarrow{\bf Top}^*$ be a~resonance functor, such that ${\cal F}(1^l)$ is contractible for $l\geq 2$. Let $l=am+\epsilon$, where $0\leq\epsilon\leq a-1$. \begin{enumerate} \item [(a)] If $k\neq 1$, or $\epsilon\geq 2$, then ${\cal F}(a^k,1^l)$ is contractible. \item [(b)] If $k=1$, and $\epsilon\in\{0,1\}$, then \begin{equation} \label{eq:4.1} {\cal F}(a^k,1^l)\simeq\mbox{susp}\,^m({\cal F}(1)^{m+\epsilon+1}), \end{equation} where ${\cal F}(1)^{m+\epsilon+1}$ denotes the $(m+\epsilon+1)$-fold smash product. \end{enumerate} \end{thm} Since for the resonance functor ${\cal F}$ described in the subsection~\ref{ss4.1} we have ${\cal F}(1^l)=X^{(l)}$, we have the following corollary. \begin{crl} \label{crl4.3} If $X^{(l)}$ is contractible for $l\geq 2$, then \begin{enumerate} \item [(a)] If $k\neq 1$, or $\epsilon\geq 2$ (again $l=am+\epsilon$), then $\Sigma_{(a^k,1^l)}^X$ is contractible. \item [(b)] If $k=1$, and $\epsilon\in\{0,1\}$, then $\Sigma_{(a^k,1^l)}^X\simeq\mbox{susp}\,^m(X^{m+\epsilon+1})$, where $X^{m+\epsilon+1}$ denotes the $(m+\epsilon+1)$-fold smash product. \end{enumerate} \end{crl} \begin{rem} Clearly, $(S^1)^{(l)}$ is contractible for $l\geq 2$, so the Corollary~\ref{crl4.3} is valid. In this situation, the case $k>1$ was proved in~\cite{Ko00b}, and the case $k=1$ in~\cite{BWa97,ShW98}. \end{rem} Before we proceed with proving the Theorem~\ref{thm4.2} we need a~crucial lemma. Let $\pi\in P(k+l)$ be $(\{1,\dots,a\},\{a+1\},\{a+2\},\dots,\{k+l\})$. It is immediate that $[\tilde\pi S]=(a^{k+1},1^{l-a})$, if $\mbox{un}\,(\tilde\pi)=\pi$. \begin{lm} \label{lm4.4} Let $S$ be as above, $T\in{\cal R}_l$, such that $[T]=(1^l)$, and let $\nu$ be the partition $(\{1,\dots,a\},\{a+1\},\{a+2\},\dots,\{l\})$, then we have \begin{equation} \label{eq:4.2} [Q(S,\pi\downarrow S)]=[Q(T,\nu\downarrow T)]\times(1^k). \end{equation} \end{lm} \begin{rem} Lemma~\ref{lm4.4} is a~special case of the Lemma~\ref{lm4.6}, however we choose to include a~separate proof for it for two reasons: firstly, it is the first, still not too technical example of investigating the combinatorial structure of the resonance category, which is a~new object; secondly, the particular case of $(a^k,1^l)$ resonances was a~subject of substantial previous attention. \end{rem} \noindent {\bf Proof of the Lemma~\ref{lm4.4}.} \vskip4pt \noindent Recall that by the definition of the direct product, $$[Q(T,\nu\downarrow T)]\times(1^k)=[Q(T\times U,(\nu\downarrow T)\times P(\{l+1,\dots,l+k\}))],$$ where $U\in{\cal R}_k$ and $[U]=(1^k)$. Clearly, $(\nu\downarrow T)\times P(\{l+1,\dots,l+k\})=\pi\downarrow S$, hence we just need to show that $S\setminus(\pi\downarrow S)=(T\times U)\setminus((\nu\downarrow T)\times P(\{l+1,\dots,l+k\}))$. Note that $(T\times U)\setminus((\nu\downarrow T)\times P(\{l+1,\dots,l+k\}))=(T\setminus(\nu\downarrow T))\times U$. Furthermore, \vskip-20pt $$S=\bigg\{(x_1,\dots,x_{l+k})\in\{-1,0,1\}^{l+k}\,\Bigm|\, \sum_{j=l+1}^{l+k}x_j+a\sum_{i=1}^{l}x_i=0\bigg\},$$ \vskip-5pt \noindent and \vskip-20pt \begin{multline*} \bigcup_{\tau\in\pi\downarrow S}S^\tau=\bigg\{(x_1,\dots,x_{l+k})\in\{-1,0,1\}^{l+k} \,\Bigm|\,\sum_{j=l+1}^{l+k}x_j+a\sum_{i=1}^{l}x_i=0, \\ \max(|\mbox{Plus\,}(x_1,\dots,x_l)|,|\mbox{Minus\,}(x_1,\dots,x_l)|)\geq a\bigg\}. \end{multline*} \vskip-5pt \noindent Therefore, by the definition of the relative resonances, we have \vskip-20pt \begin{multline*} S\setminus(\pi\downarrow S)=\bigg\{(x_1,\dots,x_{l+k})\in\{-1,0,1\}^{l+k}\,\Bigm|\,\\ \sum_{i=1}^l x_i=0,\,\,\sum_{j=l+1}^{l+k}x_j=0,\,\, |\mbox{Plus\,}(x_1,\dots,x_l)|<a\bigg\}. \end{multline*} On the other hand, $(1^k)=[\{(y_1,\dots,y_k)\in\{-1,0,1\}^k\,|\,\sum_{j=1}^k y_j=0\}]$, and $$T\setminus(\nu\downarrow T)=\bigg\{(z_1,\dots,z_l)\in\{-1,0,1\}^l\,\Bigm|\, \sum_{i=1}^l z_i=0,\,\,|\mbox{Plus\,}(z_1,\dots,z_l)|<a\bigg\},$$ which proves~\eqref{eq:4.2}. $\square$ \vskip4pt \noindent {\bf Proof of the Theorem~\ref{thm4.2}.} \vskip4pt \noindent {\bf (a)} We use induction on~$l$. The case $l<a$ can be taken as an~induction base, since then $(a^k,1^l)=(1^k)\times(1^l)$, hence, by the axiom (A3), ${\cal F}(a^k,1^l)={\cal F}(1^k)\wedge{\cal F}(1^l)$, which is contractible, since ${\cal F}(1^k)$ is. Thus we assume that $l\geq a$, and ${\cal F}(a^k,1^{l'})$ is contractible for all $l'<l$. Let $S$ and $\pi$ be as in the Lemma~\ref{lm4.4}. The standard triple associated to the morphism $\pi S\stackrel{\pi}{\hookrightarrow}S$ is ${\cal F}(a^{k+1},1^{l-a})\hookrightarrow{\cal F}(a^k,1^l)\rightarrow{\cal F}(a^k,1^l)/{\cal F}(a^{k+1},1^{l-a})$. Since, by the induction assumption, ${\cal F}(a^{k+1},1^{l-a})$ is contractible, we conclude that ${\cal F}(a^k,1^l)\simeq{\cal F}(a^k,1^l)/ {\cal F}(a^{k+1},1^{l-a})$. Basically by the definition, we have ${\cal F}(a^k,1^l)/{\cal F}(a^{k+1},1^{l-a})= {\cal F}(Q(S,\pi\downarrow S))$. On the other hand, we have proved in the Lemma~\ref{lm4.4} that $[Q(S,\pi\downarrow S)]=Q(T,\nu\downarrow T)\times (1^k)$, where $T$ and $\nu$ are described in the formulation of that lemma. By axioms (A2) and (A3) we get that ${\cal F}(Q(S,\pi\downarrow S))\simeq{\cal F}(Q(T,\nu\downarrow T))\wedge {\cal F}(1^k)$, which is contractible, since ${\cal F}(1^k)$ is. Therefore, ${\cal F}(a^k,1^l)$ is also contractible. \vskip4pt \noindent {\bf (b)} The argument is very similar to (a). We again assume $l\geq a$, which implies $l\geq 2$. By the using the same ordered set partition~$\pi$ as in (a), we get that ${\cal F}(a,1^l)\simeq {\cal F}(a,1^l)/{\cal F}(a^2,1^{l-a})$. Further, by Lemma~\ref{lm4.4} and the axioms (A2) and (A3) we conclude that ${\cal F}(a,1^l)\simeq {\cal F}(1)\wedge({\cal F}(1^l)/{\cal F}(a,1^{l-a}))$. Since ${\cal F}(1^l)$ is contractible, we get \begin{equation} \label{eq:4.3} {\cal F}(a,1^l)\simeq {\cal F}(1)\wedge \mbox{susp}\,{\cal F}(a,1^{l-a}). \end{equation} Since ${\cal F}(a)={\cal F}(1)$, ${\cal F}(a,1)={\cal F}(1)\wedge{\cal F}(1)$, and ${\cal F}(a,1^l)$ is contractible if $2\leq l<a$, we obtain \eqref{eq:4.1} by the repeated usage of~\eqref{eq:4.3}. $\square$ \subsection{Resonances $(a^k,b^l)$} The algebraic invariants of these strata have not been computed before, not even in the case $X=S^1$, and ${\cal F}$ - the standard resonance functor associated to the stratification of $X^{(n)}$. We would like to apply a~technique similar to the one used in the subsection~\ref{ss4.2}. A~problem is that, once one starts to ``glue'' $a$'s, one cannot get $b$'s in the same way as one could in the previous section from 1's. Thus, we are forced to consider a~more general case of resonances, namely $(g^m,a^k,b^l)$, where $g$ is the least common multiple of $a$ and $b$. Assume $g=a\cdot\bar a= b\cdot\bar b$, and $b>a\geq 2$. Analogously with the Theorem~\ref{thm4.2} we have the following result. \begin{thm} \label{thm4.5} Let ${\cal F}$ be as in the Theorem~\ref{thm4.2}. Let furthermore $k=x\cdot\bar a+\epsilon_1$, $l=y\cdot\bar b+\epsilon_2$, where $0\leq\epsilon_1<\bar a$, $0\leq\epsilon_2<\bar b$. Then \begin{equation} \label{eq:4.4} {\cal F}(g^m,a^k,b^l)\simeq \begin{cases} \mbox{susp}\,^{x+y+m-1}({\cal F}(1)^{x+y+m+\epsilon_1+\epsilon_2}), & \mbox{ if } m,\epsilon_1,\epsilon_2\in\{0,1\}; \\ \mbox{ point, } & \mbox{ otherwise. } \end{cases} \end{equation} \end{thm} Just as in the subsection~\ref{ss4.2} (Corollary~\ref{crl4.3}), the Theorem~\ref{thm4.5} is true if one replaces ${\cal F}(\lambda)$ with $\Sigma_\lambda^{S^1}$. The proof of the Theorem~\ref{thm4.5} follows the same general scheme as that of the Theorem~\ref{thm4.2}, but the technical details are more numerous. Again there is a~crucial combinatorial lemma. Let $S$ be an~$(m+k+l)$-cut consisting of all the elements of $\{-1,0,1\}^{m+k+l}$ which are orthogonal to the vector $(\underbrace{a,\dots,a}_k,\underbrace{b,\dots,b}_l, \underbrace{g,\dots,g}_m)$. Assume $k\geq\bar a$, and let an~unordered set partition~$\pi$ be equal to $(\{1,\dots,\bar a\},\{\bar a+1\},\dots,\{k+l+m\})$. We see that $[S]=(g^m,a^k,b^l)$, and $[\tilde\pi S]=(g^{m+1},a^{k-\bar a},b^l)$, if $\pi=\mbox{un}\,(\tilde\pi)$. \begin{lm} \label{lm4.6} Let $T\in{\cal R}_k$, such that $[T]=(1^k)$, and $\nu=(\{1,\dots,\bar a\},\{\bar a+1\},\dots,\{k\})$, then \begin{equation} \label{eq:4.5} [Q(S,\pi\downarrow S)]=[Q(T,\nu\downarrow T)]\times(\bar b^m,1^l). \end{equation} \end{lm} \nin{\bf Proof. } Again, it is easy to see that the sets of the partitions at infinity on both sides of~\eqref{eq:4.5} coincide. Indeed, $$[Q(T,\nu)]\times(\bar b^m,1^l)=[Q(T\times U,(\nu\downarrow T)\times P(\{k+1,\dots,k+m+l\}))],$$ where $U\in{\cal R}_{m+l}$, such that $[U]=({\bar b}^m,1^l)$, and $(\nu\downarrow T)\times P(\{k+1,\dots,k+m+l\})=\pi\downarrow S$. Also, we again have the equality $$(T\times U)\setminus((\nu\downarrow T)\times P(\{k+1,\dots,k+m+l\}))=T\setminus(\nu\downarrow T)\times U,$$ which greatly helps to prove that the sets if the surviving elements on the two sides of~\eqref{eq:4.5} coincide. By the definition \begin{multline*} S=\bigg\{(x_1,\dots,x_{k+l+m})\in\{-1,0,1\}^{k+l+m}\,\Bigm| \\ a\sum_{i=1}^{k}x_i+b\sum_{i=k+1}^{k+l}x_i+g\sum_{i=k+l+1}^{k+l+m}x_i=0\bigg\}, \end{multline*} and \begin{multline*} \bigcup_{\tau\in\pi\downarrow S}S^\tau=\bigg\{(x_1,\dots,x_{k+l+m}) \in\{-1,0,1\}^{k+l+m}\,\Bigm|\, a\sum_{i=1}^{k}x_i+b\sum_{i=k+1}^{k+l}x_i+ \\ g\sum_{i=k+l+1}^{k+l+m}x_i=0,\,\, \max(|\mbox{Plus\,}(x_1,\dots,x_k)|,|\mbox{Minus\,}(x_1,\dots,x_k)|)\geq\bar a\bigg\}.$$ \end{multline*} By the definition of the relative resonances and some elementary number theory we conclude that \begin{multline*} S\setminus(\pi\downarrow S)=\bigg\{(x_1,\dots,x_{k+l+m})\in\{-1,0,1\}^{k+l+m}\,\Bigm|\, |\mbox{Plus\,}(x_1,\dots,x_k)|<\bar a, \\ \sum_{i=1}^{k}x_i=0,\quad b\sum_{i=k+1}^{k+l}x_i+g\sum_{i=k+l+1}^{k+l+m}x_i=0\bigg\}.$$ \end{multline*} The number theory argument which we need is that if $ax+by+\mbox{lcm}\,(a,b)z=0$, then $\bar a\,\big|\,x$, where $\bar a\cdot a=\mbox{lcm}\,(a,b)$. This can be seen by, for example, noticing that if $ax+by+\mbox{lcm}\,(a,b)z=0$, then $b\,\big|\,ax$, but since also $a\,\big|\,ax$, we have $\mbox{lcm}\,(a,b)\,\big|\,ax$, hence $\bar a\,\big|\,x$. The equation~\eqref{eq:4.5} follows now from the earlier observations together with the equalities \vskip-10pt $$T\setminus(\nu\downarrow T)=\!\bigg\{(x_1,\dots,x_k)\in\{-1,0,1\}^k\Bigm| |\mbox{Plus\,}(x_1,\dots,x_k)|<\bar a,\sum_{i=1}^{k}x_i=0\bigg\}$$ and \vskip-10pt $$(\bar b^m,1^l)=\bigg[\bigg\{(y_1,\dots,y_{m+l})\in\{-1,0,1\}^{m+l} \,\Bigm|\,\sum_{i=1}^l y_i+\bar b\sum_{i=l+1}^{l+m}x_i=0\bigg\}\bigg]. \,\,\square$$ \vskip4pt \noindent {\bf Proof of the Theorem~\ref{thm4.5}.} The cases $k<\bar a$ and $l<\bar b$ are easily reduced to the Theorem~\ref{thm4.2}. Assume therefore that $k\geq\bar a$ and $l\geq\bar b$. Recall also that $b>a\geq 2$, and hence $\bar a\geq 2$. Let $S$ and $\pi$ be as in the formulation of the Lemma~\ref{lm4.6}. The standard triple associated to the morphism $\pi S\stackrel{\pi}{\hookrightarrow}S$ is \begin{equation} \label{eq:4.6} {\cal F}(g^{m+1},a^{k-\bar a},b^l)\hookrightarrow{\cal F}(g^m,a^k,b^l)\rightarrow {\cal F}(g^m,a^k,b^l)/{\cal F}(g^{m+1},a^{k-\bar a},b^l). \end{equation} We break the rest of the proof into 3 cases. \vskip4pt \noindent {\bf Case $m\geq 2$}. Again, we prove that ${\cal F}(g^m,a^k,b^l)$ is contractible by induction on~$k$. This is clear if $k<\bar a$. If $k\geq\bar a$, it follows from~\eqref{eq:4.6} that ${\cal F}(g^m,a^k,b^l) \simeq{\cal F}(g^m,a^k,b^l)/{\cal F}(g^{m+1},a^{k-\bar a},b^l)={\cal F}(Q(S,\pi\downarrow S))$. By Lemma~\ref{lm4.6} we conclude that ${\cal F}(g^m,a^k,b^l)\simeq {\cal F}(Q(T,\nu\downarrow T))\wedge{\cal F}(\bar b^m,1^l)$. By the Theorem~\ref{thm4.2}, ${\cal F}(\bar b^m,1^l)$ is contractible, hence so is ${\cal F}(g^m,a^k,b^l)$. \vskip4pt \noindent {\bf Case $m=0$}. By Lemma~\ref{lm4.6} we get that ${\cal F}(Q(S,\pi\downarrow S))\simeq{\cal F}(Q(T,\nu\downarrow T))\wedge{\cal F}(1^l)$. Since $l\geq 2$, we have that ${\cal F}(1^l)$ is contractible, hence so is ${\cal F}(Q(S,\pi\downarrow S))= {\cal F}(a^k,b^l)/{\cal F}(g,a^{k-\bar a},b^l)$. Therefore, by~\eqref{eq:4.6} ${\cal F}(a^k,b^l)\simeq{\cal F}(g,a^{k-\bar a},b^l)$. \vskip4pt \noindent {\bf Case $m=1$}. Since ${\cal F}(g^2,a^{k-\bar a},b^l)$ is contractible, we conclude by~\eqref{eq:4.6} that ${\cal F}(g,a^k,b^l)\simeq{\cal F}(g,a^k,b^l)/ {\cal F}(g^2,a^{k-\bar a},b^l)={\cal F}(Q(S,\pi\downarrow S))$. By Lemma~\ref{lm4.6}, and the properties of the resonance functors, we have \begin{equation} \label{eq:4.7} {\cal F}(g,a^k,b^l)\simeq{\cal F}(\bar b,1^l)\wedge({\cal F}(1^k)/{\cal F}(\bar a,1^{k-\bar a})) \simeq{\cal F}(\bar b,1^l)\wedge\mbox{susp}\,({\cal F}(\bar a,1^{k-\bar a})). \end{equation} By the repeated usage of~\eqref{eq:4.7} we obtain~\eqref{eq:4.4}. $\square$ \section{Sequential resonances} \label{s1.5} \subsection{The structure theory of strata associated to sequential resonances} \begin{df}\label{dfseqres} Let $\lambda=(\lambda_1,\dots,\lambda_n)$, $\lambda_1\leq\dots\leq\lambda_n$, be a~number partition. We call $\lambda$ {\bf sequential} if, whenever $\sum_{i\in I}\lambda_i=\sum_{j\in J}\lambda_j$, and $q\in I$, such that $q=\max(I\cup J)$, then there exists $\widetilde J\subseteq J$, such that $\lambda_q=\sum_{j\in\widetilde J}\lambda_j$. Correspondingly, we call a~resonance $S$ sequential, if it can be associated to a~sequential partition. \end{df} \noindent Note that the set of sequential partitions is closed under removing blocks. \begin{exams} {\rm Sequential partitions.} \begin{enumerate} \item[(1)] All partitions whose blocks are equal to powers of some number; \item[(2)] $(a^k,b^l,1^m)$, such that $a>bl$; more generally $(a_1^{k_1},\dots,a_t^{k_t},1^m)$, such that $a_i>\sum_{j=i+1}^t a_j k_j$, for all $i\in[t]$. \end{enumerate} \end{exams} Through the rest of this subsection, we let $\lambda$ be as in the Definition~\ref{dfseqres}. For such $\lambda$ we use the following additional notations: \begin{itemize} \item $mm(\lambda)=|\{i\in[n]\,|\,\lambda_i=\lambda_n\}|$. In other words $\lambda_{n-mm(\lambda)}\neq\lambda_{n-mm(\lambda)+1}=\dots=\lambda_n$. \item $I(\lambda)\subseteq[n]$ is the lexicographically maximal set (see below the convention that we use to order lexicographically), such that $|I(\lambda)|\geq 2$, and $\lambda_n=\sum_{i\in I(\lambda)}\lambda_i$. Note that it may happen that $I(\lambda)$ does not exist, in which case ${\cal F}(\lambda)\simeq {\cal F}(\lambda_1,\dots,\lambda_{n-mm(\lambda)})\wedge{\cal F}(1^{mm(\lambda)})$, and can be dealt with by induction. \end{itemize} Let $n$ be a~positive integer. We use the following convention for the lexicographic order on $[n]$. For $A=\{a_1,\dots,a_k\}$, $B=\{b_1,\dots,b_m\}$, $A,B\subseteq[n]$, $a_1\leq\dots\leq a_k$, $b_1\leq\dots\leq b_m$, we say that $A$ is lexicographically larger than~$B$ if, either $A\supseteq B$, or there exists $q<\min(k,m)$, such that $a_k=b_m$, $a_{k-1}=b_{m-1}$, $\dots$, $a_{k-q+1}=b_{m-q+1}$, and $a_{k-q}>b_{m-q}$. \begin{prop} \label{prop4.9} If $\lambda=(\lambda_1,\dots,\lambda_n)$, $\lambda_1\leq\dots\leq\lambda_n$, is a sequential partition, then so is $\bar\lambda=(\lambda_{j_1},\dots,\lambda_{j_t},\sum_{i\in I(\lambda)}\lambda_i)$, where $t=n-|I(\lambda)|$, and $\{j_1,\dots,j_t\}=[n]\setminus I(\lambda)$. \end{prop} \nin{\bf Proof. } Let $\bar\lambda_1=\lambda_{j_1},\dots,\bar\lambda_t=\lambda_{j_t}$, $\bar\lambda_{t+1}=\sum_{i\in I(\lambda)}\lambda_i$. We need to check the condition of the Definition~\ref{dfseqres} for the identity \begin{equation} \label{eq:4.8a} \sum_{i\in I}\bar\lambda_i=\sum_{j\in J}\bar\lambda_j. \end{equation} If $t+1\not\in I\cup J$, then it follows from the assumption that $\lambda$ is sequential. Assume $t+1\in I$. If $\bar\lambda_j=\lambda_n$, for some $j\in J$, take $\widetilde J=\{j\}$, and we are done. If $\bar\lambda_i=\lambda_n$, for some $i\in I\setminus\{t+1\}$, then, since $\lambda$ is sequential, there exists $\widetilde J\subseteq J$, such that $\sum_{j\in\widetilde J}\bar\lambda_j=\lambda_n=\bar\lambda_{t+1}$, and we are done again. Finally, assume $\bar\lambda_i\neq\lambda_n$, for $i\in(I\cup J)\setminus\{t+1\}$. Substituting $\lambda_n$ instead of $\bar\lambda_{t+1}$ into the identity~\eqref{eq:4.8a} is allowed, since $\lambda_n$ does not appear among $\{\bar\lambda_i\}_{i\in(I\cup J)\setminus\{t+1\}}$. This gives us an identity for $\lambda$, and again, since $\lambda$ is sequential, we find the desired set $\widetilde J\subseteq J$, such that $\sum_{j\in\widetilde J}\bar\lambda_j=\bar\lambda_{t+1}$. $\square$ \vskip4pt Let $S\in{\cal R}_n$ be the set of all elements of $\{-1,0,1\}^n$, which are orthogonal to the vector $\lambda=(\lambda_1,\dots,\lambda_n)$. Clearly, $[S]=\lambda$. Let $\pi\in P(n)$ be the partition whose only nonsingleton block is given by $I(\lambda)$. The next lemma expresses the main combinatorial property of sequential partitions. \begin{lm}\label{green} Let $\tau\in P(n)$ be a~partition which has only one nonsingleton block~$B$, and assume $\lambda_n=\sum_{i\in B}\lambda_i$. Then $\tau\in\pi\downarrow S$. \end{lm} \nin{\bf Proof. } Assume there exists partitions $\tau$ as in the formulation of the lemma, such that $\tau\not\in\pi\downarrow S$. Choose one so that the block~$B$ is lexicographically largest possible. Let $C=B\cap I(\lambda)$. By the definition of $I(\lambda)$, and the choice of $B$, we have $\sum_{i\in I(\lambda)\setminus C}\lambda_i= \sum_{j\in B\setminus C}\lambda_j$, and $q\in I(\lambda)\setminus C$, where $q=\max((I(\lambda)\cup B)\setminus C)$. Since partition $\lambda$ is sequential, there exists $D\subseteq B\setminus C$, such that $\lambda_q=\sum_{j\in D}\lambda_j$. Let $\gamma\in P(n)$ be the partition whose only nonsingleton block is $G=(B\setminus D)\cup\{q\}$. Clearly, $\sum_{i\in G}\lambda_i= \lambda_n$, and $|G|\geq 2$. By the choice of $q$, $G$ is lexicographically larger than $B$, hence $\gamma\in\pi\downarrow S$. Let furthermore $\tilde\gamma\in P(n)$ be the partition having two nonsingleton blocks: $D$ and $G$. By the Definition~\ref{df3.1}(2) if $\gamma\in\pi\downarrow S$, then $\tilde\gamma\in\pi\downarrow S$. By the Definition~\ref{df3.1}(3), if $\tilde\gamma\in\pi\downarrow S$, then $\tau\in\pi\downarrow S$, which yields a~contradiction. $\square$ \vskip4pt Let $T\in{\cal R}_{n-mm(\lambda)}$ be the set of all elements of $\{-1,0,1\}^{n-mm(\lambda)}$, which are orthogonal to the vector $(\lambda_1,\dots,\lambda_{n-mm(\lambda)})$. Let $\nu\in P(n-mm(\lambda))$ be the partition whose only nonsingleton block is given by $I(\lambda)$. We are now ready to state the combinatorial result which is crucial for our topological applications. \begin{lm} \label{lm4.11} \begin{equation} \label{eq:4.9} [Q(S,\pi\downarrow S)]=[Q(T,\nu\downarrow T)]\times(1^{mm(\lambda)}). \end{equation} \end{lm} \nin{\bf Proof. } By definition we must verify that the sets of partitions at infinity and the surviving elements coincide on both sides of the equation~\eqref{eq:4.9}. Let us start with the partitions at infinity. Once filtered through the Pro\-po\-si\-tion~\ref{prop4.1}, the identity $\pi\downarrow S=(\nu\downarrow T)\times P(\{n-mm(\lambda)+1,\dots,n\})$ becomes essentially tautological. Both sides consist of the partitions $\tau=(\tau_1,\dots,\tau_k)\in P(n)$, such that the number partition $(\sum_{i\in\tau_1}\lambda_i,\dots,\sum_{i\in\tau_k}\lambda_i)$ can be obtained from the number partition $(\lambda_{j_1},\dots,\lambda_{j_t},\sum_{i\in I(\lambda)}\lambda_i)$, where $\{j_1,\dots,j_t\}=[n]\setminus I(\lambda)$, by summing parts. Let us now look at the surviving elements. Obviously, $S\setminus(\pi\downarrow S)\supseteq(T\setminus(\nu\downarrow T))\times U$, where $U\in{\cal R}_k$, such that $[U]=(1^{mm(\lambda)})$, and we need to show the converse inclusion. Let $x=(x_1,\dots,x_n)\in S$, such that $\sum_{i=n-mm(\lambda)+1}^n x_i\neq 0$ (otherwise $x\in(T\setminus(\nu\downarrow T))\times U$), we can assume $\sum_{i=n-mm(\lambda)+1}^n x_i>0$. Then, since $S$ is a~sequential resonance, there exists $y=(y_1,\dots,y_n)\in S$, such that \begin{itemize} \item if $y_i\neq 0$, then $x_i=y_i$; \item $|\mbox{Plus\,}(y)|=1$, and $\mbox{Plus\,}(y)\subseteq\{n-mm(\lambda)+1,\dots,n\}$. \end{itemize} This means that $y\in S^\tau$, for some $\tau\in P(n)$, which satisfies the conditions of the Lemma~\ref{green}, which implies that $\tau\in\pi\downarrow S$. On the other hand, $y\in S^\tau$ necessitates $x\in S^\tau$, and hence $x\not\in S\setminus(\pi\downarrow S)$. This finishes the proof of the lemma. $\square$ \vskip4pt Just as before, this combinatorial fact about the resonances translates into a~topological statement, which can be further strengthened by requiring some additional properties from~$\lambda$. \begin{df} Let $\lambda=(\lambda_1,\dots,\lambda_n)$, $\lambda_1\leq\dots\leq\lambda_n$, be a~sequential partition, and let $q=\max I(\lambda)$. $\lambda$ is called {\bf strongly sequential}, if there exists $J\subseteq I(\lambda)\setminus\{q\}$, such that $\lambda_q=\sum_{i\in J}\lambda_i$ (note that we do not require $|J|\geq 2$). \end{df} We are now in a~position to prove the main topological structure theorem concerning the sequential resonances. \begin{thm} \label{thm4.13} Let ${\cal F}$ be as in the Theorem~\ref{thm4.2}. Let $\lambda$ be a~sequential partition, such that $I(\lambda)$ exists, then \begin{enumerate} \item[(1)] if $mm(\lambda)\geq 2$, then ${\cal F}(\lambda)$ is contractible; \item[(2)] if $mm(\lambda)=1$, then ${\cal F}(\lambda)\simeq{\cal F}(Q(T,\nu\downarrow T))\wedge{\cal F}(1)$, and we have the inclusion triple ${\cal F}(\mu)\stackrel{i}{\hookrightarrow}{\cal F}(\lambda_1,\dots, \lambda_{n-1})\rightarrow{\cal F}(Q(T,\nu\downarrow T))$, where $\mu=(\lambda_{j_1},\dots, \lambda_{j_t})$, $\{j_1,\dots,j_t\}=[n]\setminus I(\lambda)$, and $\nu\in P(n-mm(\lambda))$ is the partition whose only nonsingleton block is given by $I(\lambda)$. If moreover $\lambda$ is strongly sequential, then the map $i$ is homotopic to a~trivial map (mapping everything to a~point), hence the triple splits and we conclude that \begin{equation} \label{eq:4.10} {\cal F}(\lambda)\simeq({\cal F}(1)\wedge{\cal F}(\lambda_1,\dots,\lambda_{n-1})) \vee\mbox{susp}\,({\cal F}(1)\wedge{\cal F}(\mu)). \end{equation} \end{enumerate} \end{thm} \nin{\bf Proof. } \vskip4pt \noindent {\bf (1)} We use induction on $\sum_{i=1}^{n-mm(\lambda)}\lambda_i$. If $I(\lambda)$ does not exist, then $\lambda_n$ is independent, i.e., ${\cal F}(\lambda)\simeq{\cal F}(\lambda_1,\dots,\lambda_{n-mm(\lambda)}) \times{\cal F}(1^{mm(\lambda)})$, and hence ${\cal F}(\lambda)$ is contractible. Otherwise consider the inclusion triple \begin{equation} \label{eq:4.11} {\cal F}(\bar\lambda)\hookrightarrow{\cal F}(\lambda)\rightarrow{\cal F}(\lambda)/{\cal F}(\bar\lambda)= {\cal F}(Q(S,\pi\downarrow S)), \end{equation} where $\bar\lambda=(\lambda_{j_1}, \dots,\lambda_{j_t},\sum_{i\in I(\lambda)}\lambda_i)$, and $\pi\in P(n)$ is the partition whose only nonsingleton block is given by $I(\lambda)$. By the induction assumption ${\cal F}(\bar\lambda)$ is contractible. On the other hand, by Lemma~\ref{lm4.11}, ${\cal F}(Q(S,\pi\downarrow S))\simeq{\cal F}(Q(T,\nu\downarrow T))\wedge{\cal F}(1^{mm(\lambda)})$, which is also contractible if $mm(\lambda)\geq 2$. \vskip4pt \noindent {\bf (2)} If $mm(\lambda)=1$, then we can conclude from~\eqref{eq:4.11} that ${\cal F}(\lambda)\simeq{\cal F}(1)\wedge{\cal F}(Q(T,\nu\downarrow T))$. Next, consider the inclusion triple \begin{equation} \label{eq:4.12} {\cal F}(\mu)\stackrel{i}{\hookrightarrow}{\cal F}(\lambda_1,\dots,\lambda_{n-1})\rightarrow {\cal F}(Q(T,\nu\downarrow T)). \end{equation} If $\lambda$ is strongly sequential, then there exists $J\subseteq I(\lambda)\setminus\{q\}$, such that $\lambda_q=\sum_{i\in J}\lambda_i$ (here $q=\max I(\lambda)$). The map $i$ factors: \begin{equation} \label{eq:4.13} {\cal F}(\mu)\stackrel{i_1}{\hookrightarrow}{\cal F}(\lambda_{p_1},\dots, \lambda_{p_{n-1-|J|}},\sum_{i\in I(\lambda)}\lambda_i) \stackrel{i_2}{\hookrightarrow}{\cal F}(\lambda_1,\dots,\lambda_{n-1}), \end{equation} where $\{p_1,\dots,p_{n-1-|J|}\}=[n-1]\setminus J$. Since $(\lambda_{p_1},\dots,\lambda_{p_{n-1-|J|}},\sum_{i\in I(\lambda)}\lambda_i)$ is sequential, and $mm((\lambda_{p_1}, \dots,\lambda_{p_{n-1-|J|}},\sum_{i\in I(\lambda)}\lambda_i))\geq 2$, we can conclude that the middle space in~\eqref{eq:4.13} is contractible, and hence $i$ in~\eqref{eq:4.12} is homotopic to a~trivial map. This yields the conclusion. $\square$ \subsection{Resonances $(a^k,b^l,1^m)$} We give here the first application of the structure theory described in the previous subsection. \begin{thm} \label{thm4.14} Let $a,b,k,l,m,r$ be positive integers, such that $b>1$, $m\geq r$, and $a=bl+r$. Then \begin{equation} \label{eq:4.14} {\cal F}(a^k,b^l,1^m)\simeq\mbox{susp}\,({\cal F}(1^k)\wedge{\cal F}(a,1^{m-r}))\vee ({\cal F}(1^k)\wedge{\cal F}(b^l,1^m)). \end{equation} \end{thm} \begin{rem} The restriction $m\geq r$ is unimportant. Indeed, if $m<r$, then $a>bl+m$, hence $a$ is not involved in any resonance other than $a=a$. This implies that ${\cal F}(a^k,b^l,1^m)={\cal F}(1^k)\times{\cal F}(b^l,1^m)$, and we have determined the homotopy type of ${\cal F}(a^k,b^l,1^m)$ by the previous computations. \end{rem} \noindent {\bf Proof of the Theorem~\ref{thm4.14}.} \vskip4pt \noindent Obviously, the condition $a>bl$ guarantees that the partition $(a^k,b^l,1^m)$ is sequential, hence the Theorem~\ref{thm4.13} is valid. It follows that if $k\geq 2$, then ${\cal F}(a^k,b^l,1^m)$ is contractible, hence~\eqref{eq:4.14} is true. Furthermore, if $l\geq 2$, or, $l=1$ and $m\geq b$, then $(a,b^l,1^m)$ is strongly sequential, hence in this case~\eqref{eq:4.10} is valid, which in new notations becomes precisely the equation~\eqref{eq:4.14}. Finally, assume $l=1$ and $b>m\geq r\geq 1$. Let $a=b+d$. If ${\cal F}(a,1^{m-d})$ or ${\cal F}(b,1^m)$ is contractible, then the map~$i$ in the inclusion triple ${\cal F}(a,1^{m-d})\stackrel{i}{\hookrightarrow}{\cal F}(b,1^m)\rightarrow {\cal F}(b,1^m)/{\cal F}(a,1^{m-d})$ is homotopic to a~trivial map, and we again conclude~\eqref{eq:4.14}. If both of these spaces are not contractible then ${\cal F}(a,1^{m-d})\simeq S^{2y+\epsilon_2+1}$ and ${\cal F}(b,1^m)\simeq S^{2x+\epsilon_1+1}$, where nonnegative integers $x,y,\epsilon_1,\epsilon_2$ are defined by \begin{equation} \label{eq:4.15} m=bx+\epsilon_1,\,\,\,m-d=(b+d)y+\epsilon_2,\,\,\, \epsilon_1,\epsilon_2\in\{0,1\}. \end{equation} Let us show that $2x+\epsilon_1>2y+\epsilon_2$. If $x>y$, then $2x+\epsilon_1\geq 2x\geq 2y+2>2y+\epsilon_2$. From~\eqref{eq:4.15} we have that $b(x-y)=d+dy+\epsilon_2-\epsilon_1$. If $x\leq y$, then the left hand side is nonpositive. On the other hand, since $d\geq 1$, the right hand side is nonnegative. Hence, both sides are equal to 0, which implies $x=y$, $d=\epsilon_1=1$, $\epsilon_2=y=0$. This yields $2x+\epsilon_1>2y+\epsilon_2$. The homotopic triviality of the map $i$ follows now from the fact that the homotopy groups of a~sphere are trivial up to the dimension of that sphere, i.e., $\pi_k(S^n)=0$, for $0\leq k\leq n-1$. $\square$ \subsection{Resonances consisting of powers} Let us fix an integer $a\geq 2$. In this subsection we will study the topology of the strata indexed by the following class of partitions: all number partitions whose blocks are powers of $a$. Let $\Lambda(a)$ denote the set of all such partitions. It is convenient to introduce a~different notation for the partitions in this class. Let $\Lambda_a(\alpha_n,\alpha_{n-1},\dots,\alpha_0)$, where $\alpha_n,\alpha_{n-1},\dots,\alpha_0$ are nonnegative integers, denote $\lambda\in\Lambda(a)$, which consists of $\alpha_n$ parts equal to $a^n$, $\alpha_{n-1}$ parts equal to $a^{n-1}$, $\dots$, $\alpha_0$ parts equal to 1. For example $(8,4,2,2,2,1,1,1,1,1,1)=\Lambda_2(1,1,3,6)$. Obviously, all partitions from $\Lambda(a)$ are strongly sequential, hence the Theorem~\ref{thm4.13} applies, and it yields: \begin{enumerate} \item[(1)] if $\alpha_n\geq 2$, then ${\cal F}(\Lambda_a(\alpha_n,\alpha_{n-1}, \dots,\alpha_0))$ is contractible; \item[(2)] if $I(\Lambda_a(1,\alpha_{n-1},\dots,\alpha_0))$ exists, then \begin{multline} \label{eq:4.16} {\cal F}(\Lambda_a(1,\alpha_{n-1},\dots,\alpha_0)) \simeq ({\cal F}(1)\wedge{\cal F}(\Lambda_a(\alpha_{n-1},\dots,\alpha_0)))\vee \\ (S^1\wedge{\cal F}(1)\wedge{\cal F}(\Lambda_a(1,\beta_{n-1},\dots,\beta_0))), \end{multline} where $\Lambda_a(1,\beta_{n-1},\dots,\beta_0)$ is obtained from $\Lambda_a(1,\alpha_{n-1},\dots,\alpha_0)$ by removing the blocks indexed by $I(\Lambda_a(1,\alpha_{n-1},\dots,\alpha_0))$. \item[(3)] If $I(\Lambda_a(1,\alpha_{n-1},\dots,\alpha_0))$ does not exist, then \begin{equation} \label{eq:4.17} {\cal F}(\Lambda_a(1,\alpha_{n-1},\dots,\alpha_0)) \simeq {\cal F}(1)\wedge{\cal F}(\Lambda_a(\alpha_{n-1},\dots,\alpha_0)). \end{equation} \end{enumerate} It is immediate from the formulae~\eqref{eq:4.16} and~\eqref{eq:4.17} that each topological space ${\cal F}(\Lambda_a(\alpha_n,\alpha_{n-1}, \dots,\alpha_0))$ is homotopy equivalent to a~wedge of spaces of the form ${\cal F}(1)^\alpha\wedge S^\beta$, where ${\cal F}(1)^\alpha$ means an~$\alpha$-fold smash product of ${\cal F}(1)$. The natural combinatorial question which arises is how to enumerate these spaces. We shall now construct a~combinatorial model: a~weighted graph which yields such an~enumeration. \begin{df} Let $\lambda=\Lambda_a(\alpha_n,\alpha_{n-1},\dots,\alpha_0))$, and set $\alpha_{-1}=1$. $\Gamma_\lambda$ is a~directed weighted graph on the set of vertices $\{n,\dots,0,-1\}$ whose edges and weights are defined by the following rule. For $x,x+d\in\{-1,0,\dots,n\}$, $d\geq 1$, there exists an~edge $e(x,x+d)$ (the edge is directed {\it from} $x$ {\it to} $x+d$) if and only if $\alpha_x>0$, $\alpha_{x+d}>0$, and $$a^d\,\big|\,a^{d-1}\alpha_{x+d-1}+a^{d-2}\alpha_{x+d-2}+\dots+ a\alpha_{x+1}+\alpha_x-1.$$ In this case the weight of the edge is defined as $$w(x,x+d)=(a^{d-1}\alpha_{x+d-1}+a^{d-2}\alpha_{x+d-2}+\dots+ a\alpha_{x+1}+\alpha_x-1)\cdot a^{-d}.$$ \end{df} \noindent Note that if $d\geq 2$ and there exists an edge $e(x,x+d)$, then there exists an edge $e(x,x+d-1)$. We call a~directed path in $\Gamma_\lambda$ {\it complete} if it starts in $-1$ and ends in $n$. Let $\gamma$ be a~complete path in $\Gamma_\lambda$ consisting of $t$ edges, $\gamma=(e(x_0,x_1),\dots, e(x_{t-1},x_t))$, where $x_0=-1$, and $x_t=n$. The weight of $\gamma$ is defined to be the pair $(l(\gamma),w(\gamma))$, where $l(\gamma)=t$, and $w(\gamma)=\sum_{i=1}^t w(x_{i-1},x_i)$. \begin{thm}\label{thm4.15} Let $\lambda\in\Lambda(a)$, $\lambda=\Lambda_a(1,\alpha_{n-1},\dots,\alpha_0)$, then \begin{equation} \label{eq:4.18} {\cal F}(\lambda)\simeq\bigvee_{\gamma}({\cal F}(1)^{l(\gamma)+w(\gamma)} \wedge S^{w(\gamma)}), \end{equation} where the wedge is taken over all complete paths of $\Gamma_\lambda$. \end{thm} \nin{\bf Proof. } We use induction on $\sum_{i=0}^n \alpha_i$. As the base of the induction we take the case $\lambda=\Lambda_a(1,\underbrace{0,\dots,0}_n)$. In this case $\Gamma_\lambda$ is a~graph with only one edge $e(-1,n)$, $w(-1,n)=0$. Thus, there is only one complete path. It has weight $(1,0)$, and ${\cal F}(\lambda)\simeq{\cal F}(1)$. Next, we prove the induction step. We break up the proof in three cases. Let $t\in[n-1]$ be the maximal index for which $\alpha_t\neq 0$. \vskip4pt \noindent {\bf Case 1.} {\it $I(\lambda)$ does not exist.} \noindent By~(\ref{eq:4.17}) we have \begin{equation} \label{eq:4.19} {\cal F}(\lambda)\simeq{\cal F}(1)\wedge{\cal F}(\Lambda_a(\alpha_{n-1},\dots,\alpha_0)). \end{equation} On the other hand, $I(\lambda)$ does not exist if and only if $a^n>\alpha_{n-1}a^{n-1}+\dots+\alpha_1 a+\alpha_0$. We also know that $\lambda\neq\Lambda_a(1,0,\dots,0)$, i.e., $\alpha_{n-1}a^{n-1}+\dots+\alpha_1 a+\alpha_0>0$. This implies that there is at most one edge of the type $e(x,n)$. This edge exists if and only if $\alpha_x=1$, and $\alpha_{n-1}=\dots=\alpha_{x+1}=0$, in which case $w(x,n)=0$. If this edge does not exist then there are no complete paths in $\Gamma_\lambda$ and, at the same time ${\cal F}(\Lambda_a(\alpha_{n-1},\dots,\alpha_0))$ is contractible by the previous observations. This agrees with~(\ref{eq:4.18}). If, on the other hand, this edge does exist, then all complete paths $\gamma$ must be of the type $\gamma=(\tilde\gamma,e(x,n))$, where $\tilde\gamma$ is a~complete path from $-1$ to $x$. Also in this case~(\ref{eq:4.19}) agrees with~(\ref{eq:4.18}). \vskip4pt \noindent {\bf Case 2.} {\it $I(\lambda)$ exists and $\alpha_t\geq 2$.} \noindent In this case ${\cal F}(\Lambda_a(\alpha_{n-1},\dots,\alpha_0))$ is contractible, and \begin{equation} \label{eq:4.20} {\cal F}(\lambda)\simeq S^1\wedge{\cal F}(1)\wedge{\cal F}(\Lambda_a(1,\beta_{n-1}, \dots,\beta_0)), \end{equation} where $\beta_{n-1}, \dots,\beta_0$ are as in~(\ref{eq:4.16}). Let $q\in\{n-1,\dots,0,-1\}$ be the maximal index for which $\beta_q\neq 0$ (we assume $\beta_{-1}=1$). Let $\tilde\lambda= (1,\beta_{n-1},\dots,\beta_0)$. We can describe the graph $\Gamma_{\tilde\lambda}$: it is obtained from $\Gamma_\lambda$ by \begin{enumerate} \item[(1)] removing the edges which have one of the endpoints in the set $\{n-1,\dots,$ $q+1\}$; \item[(2)] decreasing the weight of every existing edge $e(x,n)$ by $1$; \item[(3)] keeping all existing edges with the old weights on the set $\{q,q-1,\dots,-1\}$. \end{enumerate} This operation on $\Gamma_\lambda$ is well-defined, since there can be no edges in $\Gamma_\Lambda$ of the type $e(x,n)$, for $x\in\{n-1,\dots,q+1\}$, and since the weight of edges $e(x,n)$, for $x\in\{q,\dots,0,-1\}$ must be at least 1, as $b_q\neq 0$. Furthermore, it is clear from the above combinatorial description of $\Gamma_{\tilde\lambda}$, that the set of the complete paths of $\Gamma_{\tilde\lambda}$ is the same as that of $\Gamma_\lambda$, and that the weights of the edges in these paths are also the same except for the edge with the endpoint $n$, whose weight has been decreased by~1. Thus,~(\ref{eq:4.20}) agrees with~(\ref{eq:4.18}) in this case. \vskip4pt \noindent {\bf Case 3.} {\it $I(\lambda)$ exists and $\alpha_t=1$.} \noindent This case is rather similar to the case 2, except that there is an~edge $e(t,n)$ of weight $0$. Thus, $\Gamma_{\tilde\lambda}$ bookkeeps all the complete paths of $\Gamma_\lambda$, except for the ones which have this edge $e(t,n)$. However, the first term of the right hand side of~(\ref{eq:4.16}) bookkeeps the paths $(\tilde\gamma,e(t,n))$, just like in the case~1. Since the set of all complete paths of $\Gamma_\lambda$ is the disjoint union of the sets of those paths which contain $e(t,n)$, and those which do not, we again get that~(\ref{eq:4.16}) provides the inductive step for~(\ref{eq:4.18}). $\square$ \begin{exams} $\,$ \vskip4pt \noindent (1) Let $\lambda=(a,1^l)$, for $a\geq 2$. Then $\Gamma_\lambda$ is a~graph on the vertex set $\{1,0,-1\}$ having either one or two edges: \begin{enumerate} \item it has in any case the edge $e(-1,0)$, $w(-1,0)=0$; \item if $a$ divides $l$, then it has the edge $e(-1,1)$, in which case $w(-1,1)=l/a$; \item if $a$ divides $l-1$, then it has the edge $e(0,1)$, in which case $w(0,1)=(l-1)/a$. \end{enumerate} Clearly the Theorem~\ref{thm4.15} agrees with the Theorem~\ref{thm4.2}. Indeed, if $\epsilon\not\in\{0,1\}$ (where $\epsilon$ is taken from the formulation of the Theorem~\ref{thm4.2}), then there are no complete paths in $\Gamma_\lambda$. If $\epsilon=0$, then there is one path $(-1,1)$ of weight $(1,l/a)$; and if $\epsilon=1$, then there is one path $((-1,0),(0,1))$ of weight $(2,(l-1)/a)$. Thus,~(\ref{eq:4.18}) and~(\ref{eq:4.1}) are equivalent in this case. $$\begin{array}{c} \epsffile{graph.eps}\\ \mbox{Figure 1.1.}\,\, \Gamma_{(8,4,2,2,2,1,1,1,1,1,1)}. \end{array}$$ \noindent (2) Let $\lambda=(8,4,2,2,2,1,1,1,1,1,1)=\Lambda_2(1,1,3,6)$. The graph $\Gamma_\lambda$ is shown on the Figure~1.1. It has 4 directed paths from -1 to 3 and, by the Theorem~~\ref{thm4.15}, we have $${\cal F}(\lambda)\simeq ({\cal F}(1)^3\wedge S^2)\vee ({\cal F}(1)^5\wedge S^3)\vee ({\cal F}(1)^6\wedge S^4)\vee ({\cal F}(1)^7\wedge S^4),$$ in particular $\Sigma_\lambda^{\Bbb R}\simeq S^5\vee S^8\vee S^{10}\vee S^{11}$. \end{exams} \section{Complexity of resonances} \label{s1.6} The main idea of all our previous computations was to find, for a~given $n$-cut $S$, a~partition $\pi\in P(n)$, such that $\mbox{span}\,(S\setminus(\pi\downarrow S))\neq S$. Intuitively speaking, shrinking the substratum corresponding to $\tilde\pi S$, where $\mbox{un}\,(\tilde\pi)=\pi$, {\it essentially} reduces the set of linear identities in~$S$. It is easy to construct examples when such $\pi$ does not exist, e.g., Example~\ref{exsmres}(4). These observations lead us to introduce a~formal notion of complexity of a~resonance. \begin{df} \label{df6.1} $\,$ \noindent 1) For $S\in{\cal R}_n$, the {\bf complexity} of $S$ is denoted $c(S)$ and is defined by: \begin{equation} \label{eq:6.1} c(S)=\min\{|\Pi|\,|\,\Pi\subseteq P(n),\mbox{span}\,(S\setminus(\Pi\downarrow S))\neq S\}. \end{equation} \noindent 2) We define the complexity of an $n$-resonance to be the complexity of one of its representing cuts. Clearly, it does not depend on the choice of the representative. \end{df} \begin{rem} The number $c(S)$ would not change if we required the partitions in $\Pi$ to have one block of size 2, and all other blocks of size~1. \end{rem} The higher is the complexity of a resonance $[S]$, the less it is likely that one can succeed with analyzing its topological structure using the method decribed in this chapter. This is because one would need to take a~quotient by a~union of $c([S])$ strata and it might be difficult to get a~hold on the topology of that union. \vskip4pt We finish by constructing for an~arbitrary $n\in{\Bbb N}$, a~resonance of complexity~$n$. Let $\lambda_n=(a_1,\dots,a_n,b_1,\dots,b_n)$, such that $a_i,b_i\in{\Bbb N}$, for $i,j\in[n]$, and all other linear identities among $a_i$'s and $b_i$'s with coefficients $\pm 1,0$ are generated by such identities. In other words, the cut $S$ associated to $\lambda$ is equal to the set \begin{equation} \label{eq:6.2} \Bigl\{(x_1,\dots,x_n,y_1,\dots,y_n)\in\{-1,0,1\}^{2n}\,\Big |\, \sum_{i=1}^n y_i=0,\,\,x_i+y_i=0,\forall i\in[n]\Bigr\}. \end{equation} \noindent It is not difficult to construct such $\lambda_n$ directly: \vskip4pt \noindent 1) Choose $a_1,\dots,a_n$, such that the only linear identities with coefficients $\pm 1,0$ on the set $a_1,a_1,a_2,a_2,\dots,a_n,a_n$ are of the form $a_i=a_i$; in other words, there are no linear identities with coefficients $\pm 2,\pm 1,0$ on the set $a_1,\dots,a_n$. One example is provided by the choice $a_1=1$, $a_2=3$, $\dots$, $a_n=3^{n-1}$. \vskip4pt \noindent 2) Let $b_i=N+a_i$, for $i\in[n]$, where $N$ is sufficiently large. As the proof of the Proposition~\ref{prop6.2} will show, it is enough to choose $N>2\sum_{i=1}^n\lambda_i$. This bound is far from sharp, but it is sufficient for our purposes. \begin{prop} \label{prop6.2} Let $S_n$ be the $n$-cut associated to the ordered sequence of natural numbers $\lambda_n$ described above, then $c(S_n)=n$. \end{prop} \nin{\bf Proof. } First, let us verify that the cut $S_n$ associated to $\lambda_n$ is equal to the one described in~\eqref{eq:6.2}. Take $(x_1,\dots,x_n,y_1,\dots,y_n)\in S_n$. Assume first that $\sum_{i=1}^n y_i\neq 0$. Then, $(x_1,\dots,x_n,y_1,\dots,y_n)$ stands for the identity \begin{equation} \label{eq:6.3} \sum_{i\in I_1}a_i+\sum_{j\in J_1}b_j= \sum_{i\in I_2}a_i+\sum_{j\in J_2}b_j, \end{equation} such that $|J_1|\geq|J_2|+1$. This implies that $N$ is equal to some linear combination of $a_i$'s with coefficients $\pm 2,\pm 1,0$. This leads to contradiction, since $N>2\sum_{i=1}^n\lambda_i$. Thus, we know that $\sum_{i=1}^n y_i=0$. Cancelling $N\cdot|J_1|$ out of~\eqref{eq:6.3} we get an~identity with coefficients $\pm 2,\pm 1,0$ on the set $a_1,\dots,a_n$. By the choice of $a_i$'s, this identity must be trivial, which amounts exactly to saying that $x_i+y_i=0$, for $i\in[n]$. Second, it is a~trivial observation that $c(S_n)\leq n$. Indeed, let $\pi_i\in P(n)$ be a~partition with only one nonsingleton block $(1,n+i)$, for $i\in[n]$. Then $\mbox{span}\,(S_n\setminus(\{\pi_1,\dots,\pi_n\}\downarrow S_n))\neq S_n$, since for any $(x_1,\dots,x_n,y_1,\dots,y_n)\in S_n\setminus(\{\pi_1,\dots,\pi_n\}\downarrow S_n)$, we have $x_1=0$. Finally, let us see that $c(S_n)>n-1$. As we have remarked after the Definition~ref{df6.1}, it is enough to consider the case when the partitions of $\Pi$ have one block of size 2, and the rest are singletons. Let us call the identity $a_i+b_j=a_j+b_i$ {\it the elementary identity indexed $(i,j)$}. From the definition of the closure operation $\downarrow$ it is clear that an~elementary identity indexed $(i,j)$ is not in $S_n\setminus(\Pi\downarrow S_n)$ if and only if the partition whose only nonsingleton block is $(i,n+j)$ belongs to $\Pi$, or the partition whose only nonsingleton block is $(j,n+i)$ belongs to $\Pi$. That is because the only reason this identity would not be in $S_n\setminus(\Pi\downarrow S_n)$ would be that one of these two partitions is in $\Pi\downarrow S_n$. But, if such a~partition is in $\Pi\downarrow S_n$, then it must be in $\Pi$: moves (2) of the Definition~\ref{df3.1} can never produce a~partition whose only nonsingleton block has size~2, while the moves (3) of the Definition~\ref{df3.1} may only interchange between partitions $(i,n+j)$ and $(j,n+i)$ in our specific situation. Thus, we can conclude that if $|\Pi|\leq n-1$, then at most $n-1$ elementary identities are not in $S_n\setminus(\Pi\downarrow S_n)$. Next, we note that for any distinct $i,j,k\in[n]$, the elementary identities $(i,j)$ and $(j,k)$ imply the elementary identity $(i,k)$. Let us now think of elementary identities as edges in a~complete graph on $n$~vertices, $K_n$. Then, any set $M$ of elementary identities corresponds to a~graph $G$ on $n$ vertices, and the collection of the elementary identities which lie in the $\mbox{span}\, M$ is encoded by the {\it transitive closure} of $G$. It is a~well known combinatorial fact that $K_n$ is $(n-1)$-connected, which means that removal of at most $n-1$ edges from it leaves a~connected graph. Hence, if we remove at most $n-1$ edges from $K_n$ and then take the transitive closure, we get $K_n$ again. Thus, if $|\Pi|\leq n-1$, all elementary identities lie in $\mbox{span}\,(S_n\setminus(\Pi\downarrow S_n))$. Since the elementary identities generate the whole $S_n$, we conclude that $S_n=\mbox{span}\,(S_n\setminus(\Pi\downarrow S_n))$, hence $c(S_n)>n-1$. $\square$ \chapter{Incidence Combinatorics of Resolutions} \section{The motivation for the abstract framework}\label{sect_intr} In this chapter we introduce notions of {\it combinatorial blowups, building sets\/}, and {\it nested sets\/}, for an arbitrary meet-semilattice. The definitions are given on a~purely order-theoretic level without any reference to geometry. This provides a common abstract framework for the incidence combinatorics occurring in at least two different situations in algebraic geometry: the construction of De~Concini-Procesi models of subspace arrangements~\cite{DP95}, and the resolution of singularities in toric varieties. The various parts of this abstract framework have received different emphasis within different situations: while the notion of combinatorial blowups clearly specializes to stellar subdivisions of defining fans in the context of toric varieties, building sets and nested sets were introduced in the context of model constructions by De~Concini \& Procesi~\cite{DP95} (earlier and in a more special setting by Fulton~ \& MacPherson~\cite{FM94}), from where we adopt our terminology. This correspondence however is not complete: the building sets in \cite{DP95, FM94} are not canonical, they depend on the geometry, while ours do not. See Section~\ref{ssect_DPmodels} for further details. It was proved in \cite{DP95} that a sequence of blowups within an arrangement of complex linear subspaces leads from the~intersection stratification of complex space given by the maximal subspaces of the arrangement to an arrangement model stratified by divisors with normal crossings. In the context of toric varieties, there exist many different procedures for stellar subdivisions of a defining fan that result in a simplicial fan, so-called simplicial resolutions. The purpose of our Main Theorem \ref{thm_main} is to unify these two situations on the combinatorial level: a sequence of combinatorial blowups, performed on a~(combinatorial) building set in linear extension compatible order, transforms the initial semilattice to a semilattice where all intervals are boolean algebras, more precisely to the face poset of the corresponding simplicial complex of nested sets. In particular, the structure of the resulting semilattice can be fully described by the initial data of nested sets. Both the formulation and the proof of our main theorem are purely combinatorial. \vskip4pt \noindent We sketch the content of this chapter: \vskip4pt \noindent {\bf Section \ref{sect_buidg_nest_sets}.} After providing some basic poset terminology, we define building sets and nested sets for meet-semilattices in purely order-theoretic terms and develop general structure theory for these notions. \vskip4pt \noindent {\bf Section \ref{sect_comb_blowups}.} We define combinatorial blowups of meet-semilattices, and study their effect on building sets and nested sets. The section contains our Main Theorem~\ref{thm_main} which describes the result of blowing up the elements of a building set in terms of the initial nested set complex. \vskip4pt \noindent The next two sections are devoted to relating our abstract framework to two different contexts in algebraic geometry. \vskip4pt \noindent {\bf Section~\ref{ssect_DPmodels}.} We briefly review the construction of De~Concini-Procesi models for subspace arrangements. After that, we show that the change of the incidence combinatorics of the stratification in a single construction step is described by a~combinatorial blowup of the semilattice of strata, and trace their resolution procedure step-by-step. \vskip4pt \noindent {\bf Section~\ref{ssect_tv}.} We draw the connection to simplicial resolutions of toric varieties: we recognize stellar subdivisions as combinatorial blowups of the face posets of defining fans and discuss the notions of building and nested sets in this context. \section{Building sets and nested sets of meet-semilattices} \label{sect_buidg_nest_sets} \subsection{Irreducible elements in posets} \label{ssect_pos_term} We assume known the parts of the terminology of posets described in Appendix~B. A poset is called {\it irreducible\/} if it is not a~direct product of two other posets, both consisting of at least two elements. For a poset $P$ with a unique minimal element~$\hat 0$, we call $I(P)=\{x\,{\in}\,P\,|\,[\hat 0,x]\,\, \mbox{is irreducible}\,\}$ the {\em set of irreducible elements\/} in~$P$. In particular, the minimal element~$\hat 0$ and all atoms of $P$ are irreducible elements in~$P$. For $x\,{\in}\,P$, we call $D(x)= {\rm max}\,(I(P)_{\leq x})$ the {\em set of elementary divisors\/} of~$x$ -- a~term which is explained by the following proposition: \begin{prop} \label{prop_elem_div} Let $P$ be a poset with a unique minimal element~$\hat 0$. For $x\,{\in}\,P$ there exists a unique finest decomposition of the interval~$[\hat 0,x]$ in~$P$ as a direct product, which is given by an isomorphism $\varphi_x^{\rm el}:\, \prod_{j=1}^l\, [\hat 0,y_j] \, \stackrel{\cong}{\longrightarrow} \, [\hat 0,x] $, with $\varphi_x^{\rm el}(\hat 0, \ldots, y_j, \ldots, \hat 0)\,{=}\, y_j$ for $j=1,\ldots,l$. The factors of this decomposition are the intervals below the elementary divisors of $x$: $\{y_1,\ldots,y_l\}=D(x)$. \end{prop} \nin{\bf Proof. } Whenever a poset with a minimal element~$\hat 0$ is represented as a~direct product, all elements which have more than one coordinate different from $\hat 0$ are reducible. Hence, if $\prod_{j=1}^l[\hat 0,y_j]\,{\cong}\,[\hat 0,x]$, and the $y_j$ are irreducible for $j\,{=}\,1,\ldots,l$, then $\{y_1,\ldots,y_l\}\,{=}\, D(x)$. $\square$ \subsection{Building sets}\label{ssect_buildg} \noindent In this subsection we define the notion of building sets of a~semilattice and develop their structure theory. \begin{df} \label{df_buildg} Let ${\cal L}$ be a semilattice. A subset ${\cal G}$ in~${\cal L}$ is called a {\bf building set} of~${\cal L}$ if for any $x\,{\in}\,{\cal L}$ and {\rm max}$\, {\cal G}_{\leq x}=\{x_1,\ldots,x_k\}$ there is an isomorphism of posets \begin{equation}\label{eq_buildg} \varphi_x:\,\,\, \prod_{j=1}^k\,\,\, [\hat 0,x_j] \,\, \stackrel{\cong}{\longrightarrow} \,\, [\hat 0,x] \end{equation} with $\varphi_x(\hat 0, \ldots, x_j, \ldots, \hat 0)\, = \, x_j$ for $j=1,\ldots, k$. We call $F(x)=\max {\cal G}_{\leq x}$ the {\bf set of factors} of $x$ in ${\cal G}$. \end{df} The next proposition provides several equivalent conditions for a subset of ${\cal L}$ to be a~building set. \begin{prop} \label{prop_buildingsets} For a semilattice ${\cal L}$ and a subset ${\cal G}$ of~${\cal L}$ the following are equivalent: \begin{itemize} \item[(1)] ${\cal G}$ is a building set of ${\cal L}$; \item[(2)] ${\cal G}\/{\supseteq}\,I({\cal L})$, and for every $x\,{\in}\,{\cal L}$ with $D(x)\,{=}\,\{y_1,\ldots, y_l\}$ the elementary divisors of~$x$, there exists a~partition $\pi_x\,{=}\,\pi_1|\ldots|\pi_k$, of the set $[l]$, with blocks $\pi_t\,{=}\,\{i_1,\ldots, i_{|\pi_t|}\}$, for $t\in[k]$, such that the elements in the set ${\rm max}\, {\cal G}_{\leq x}=$ $\{x_1,\ldots, x_k\}$ are of the form $ x_t\, = \, \varphi_x^{{\rm el}}(\hat 0, \ldots, \hat 0, y_{i_1},\hat 0, \ldots, \hat 0,y_{i_2},\hat 0, \ldots, \hat 0,$ $y_{i_{|\pi_t|}},\hat 0) $. \noindent Informally speaking, the factors of $x$ in ${\cal G}$ are products of disjoint sets of elementary divisors of~$x$. \item[(3)] ${\cal G}$ generates ${\cal L}$ by $\vee$, and for any $x\,{\in}\,{\cal L}$, any $\{y,y_1,\dots,y_t\}\subseteq \max {\cal G}_{\leq x}$, and $z\,{\in}\,{\cal L}$ with $z\,{<}\,y$, we have $ {\cal G}_{\leq y}\cap {\cal G}_{\leq z\vee y_1\vee\dots\vee y_t}= {\cal G}_{\leq z} $. \item[(4)] ${\cal G}$ generates ${\cal L}$ by $\vee$, and for any $x\,{\in}\,{\cal L}$, any $\{y,y_1,\dots,y_t\}\subseteq\max {\cal G}_{\leq x}$, and $z\,{\in}\,{\cal L}$ with $z\,{<}\,y$, the following two conditions are satisfied: \[ \begin{array}{rll} i) & {\cal G}_{\leq y}\cap {\cal G}_{\leq y_1\vee\dots\vee y_t}=\{\hat 0\} &\mbox{``disjointness,''}\\ i\!i) &z\vee y_1\vee\dots\vee y_t<y\vee y_1\vee\dots\vee y_t &\mbox{``necessity.''} \end{array} \] \end{itemize} \end{prop} \nin{\bf Proof. } \noindent \underline{(1)$\Rightarrow$(2)}: That ${\cal G}$ contains $I({\cal L})$ follows directly from the definition of building sets. We have the following isomorphisms: $\varphi_x:\,\prod_{j=1}^k\, [\hat 0,x_j]\,{\longrightarrow}\,[\hat 0,x]$ by the building set property, and $\varphi_{x_j}^{\rm el}:\,\prod_{y\in D(x_j)}\, [\hat 0,y] \,{\longrightarrow}\, [\hat 0,x_j]$ for $j\,{=}\, 1,\ldots, k$ by Proposition~\ref{prop_elem_div}. The composition $\varphi_x\,\circ\, (\prod_{j=1}^k\,\varphi_{x_j}^{\rm el})$ yields the finest decomposition $\varphi_{x}^{\rm el}$ of~$[\hat 0,x]$. Thus, $D(x)\,{=}\,\uplus_{j=1}^k\, D(x_j)$, which gives the partition described in~(2). \noindent \underline{(2)$\Rightarrow$(1)}: The decomposition of the interval $[\hat 0,x]$ into intervals below the elements in $\max{\cal G}_{\leq x}$ follows from Proposition~\ref{prop_elem_div} by assembling factors~$[\hat 0, y_j]$ with maximal elements indexed by elements from the same block of the partition~$\pi_x$ into one factor. \noindent \underline{(1)$\Rightarrow$(3)}: (3) is a direct consequence of~$[\hat 0,x]$ decomposing into a direct product of the form described in the definition of building sets. \noindent \underline{(3)$\Rightarrow$(4)}: $i)$ follows by setting $z=\hat 0$ in~(3). Equality in $i\!i)$ implies with~(3) that ${\cal G}_{\leq y}={\cal G}_{\leq z}$, in particular, $y\in {\cal G}_{\leq z}$ -- a contradiction to $z<y$. \noindent \underline{(4)$\Rightarrow$(1)}: For $x\in {\cal L}$ and max$\,{\cal G}_{\leq x}\, = \, \{x_1,\ldots, x_k\}$ consider the poset map \[ \phi:\,\,\, \prod_{j=1}^k\,\,\, [\hat 0,x_j] \,\, \longrightarrow \,\, [\hat 0,x] \, , \quad (\alpha_1, \ldots, \alpha_k)\,\, \longmapsto \,\, \alpha_1 \vee \ldots \vee \alpha_k\, . \] \noindent i) $\phi$ {\em is surjective\/}: For $\hat 0\,{\neq}\, y \,{\leq}\, x$, let max$\,{\cal G}_{\leq y}=\{y_1,\ldots,y_t\}$. First, we have $\bigvee_{i=1}^t y_i\,{=}\,y$, since ${\cal G}$ generates ${\cal L}$ by $\vee$. Second, define $\gamma_j\,{=}\, \bigvee_{y_i\in S_j}y_i$, with $S_j=$ $({\rm max}\,{\cal G}_{\leq y})\,{\cap}\,{\cal G}_{\leq x_j}$, for $j\,{=}\,1,\ldots,k$. Clearly, $\gamma_j\,{\in}\,[\hat 0,x_j]$, and $\cup_{j=1}^k S_j\,{=}\, {\rm max}\,{\cal G}_{\leq y}$, since ${\cal G}_{\leq y}\,{\subseteq}\, {\cal G}_{\leq x}$. Hence, $\phi(\gamma_1,\ldots, \gamma_k)=\bigvee_{i=1}^t y_i=y$. \noindent ii) $\phi$ {\em is injective\/}: a) Assume $\phi(\alpha_1,\ldots,\alpha_k)\, {=}\, \phi(\beta_1,\ldots,\beta_k)= y \neq x$, and let max$\,{\cal G}_{\leq y}=\{y_1,\ldots,y_t\}$. By induction on the number of elements in~$[\hat 0,x]$ we can assume that $[\hat 0,y]$ decomposes as a direct product $ [\hat 0,y] \, \cong \, \prod_{i=1}^t\, [\hat 0,y_i] $. Moreover, the subsets $S_j$ of max$\,{\cal G}_{\leq y}$ defined in~i) actually partition max$\,{\cal G}_{\leq y}$ as follows from the disjointness property applied to pairwise intersections of the $\,{\cal G}_{\leq x_j}$. Thus, $ [\hat 0,y] \, \cong \, \prod_{j=1}^k\, [\hat 0,\gamma_j] $, with elements $\gamma_j\,{\in}\, [\hat 0,x_j]$ as above, and it follows that $\alpha_j=\beta_j=\gamma_j$ for $j=1,\ldots,k$. \newline b) Assume that $\phi(\alpha_1,\ldots,\alpha_k)\, {=}\, \phi(\beta_1,\ldots,\beta_k)\,{=}\,x$. By the necessity property it follows that $\alpha_j=\beta_j=x_j$ for $j=1,\ldots,k$. $\square$ \begin{rem} \label{rem_indpce_of_joins} The definition of building sets and of irreducible elements, as well as the characterization of building sets in Proposition~\ref{prop_buildingsets}~(2), are independent of the existence of a join operation and can be formulated for any poset with a unique minimal element. \end{rem} \noindent We gather a few important properties of building sets. \begin{prop}\label{prop_buildingsets_properties} For a building set ${\cal G}$ of ${\cal L}$, the following holds: \begin{itemize} \item [(1)] Let $x\in{\cal L}$, $F(x)=\{x_1,\dots,x_k\}$ the set of factors of~$x$ in~${\cal G}$, and $\hat 0\,{\neq}\,y\,{\in}\,{\cal G}$ with $y\,{\leq}\,x$. Then there exists a unique $j\,{\in}\, \{1,\ldots,k\}$ such that $y\,{\leq}\,x_j$; i.e., $F(x)={\rm max}\,{\cal G}_{\leq x}$ induces a partition of ${\cal G}_{\leq x}\,{\setminus}\,\{\hat 0\}$. \item[(2)] For $x\,{\in}\,{\cal L}$ and $x_0\,{\in}\,F(x)$, \[ \bigvee\, (F(x)\,{\setminus}\,\{x_0\}) \,\, < \,\, \bigvee\, F(x)\, =x\, , \] i.e., each factor of~$x$ in ${\cal G}$ is needed to generate~$x$. \item [(3)] If $h_1,\dots,h_k$ in ${\cal G}$ are such that $(h_i,\bigvee_{j=1}^k h_j] \cap {\cal G}=\emptyset$ for $i=1,\dots,k$, then $F(\bigvee_{j=1}^k h_j)\, = \, \{h_1,\ldots, h_k\}$. \end{itemize} \end{prop} \nin{\bf Proof. } (1) is a consequence of Proposition~\ref{prop_buildingsets}~(4){\em i}), as was noted already in the proof of (4)$\Rightarrow$(1), part ii)~a), in the previous proposition. Taking the full set of factors and setting~$z\,{=}\,\hat 0$ in Proposition~\ref{prop_buildingsets}~(4){\em i$\!$i}), yields~(2). For~(3) note that $\{h_1,\ldots,h_k\}\,{\subseteq}\,F(\bigvee_{j=1}^k h_j)$ by assumption. If $\{h_1,\ldots, h_k\}$ were not the complete set of factors, we would obtain a contradiction to~(2). $\square$ \subsection{Nested sets} \noindent In this subsection we define the notion of nested subsets of a~building set of a~semilattice and prove some of their properties. \begin{df} \label{df_nested} Let ${\cal L}$ be a semilattice and ${\cal G}$ a building set of ${\cal L}$. A subset $N$ in ${\cal G}$ is called~{\bf nested} if, for any set of incomparable elements $x_1,\dots,x_t$ in $ N$ of cardinality at least two, the join $x_1\vee\dots\vee x_t$ exists and does not belong to ${\cal G}$. The nested sets in ${\cal G}$ form an abstract simplicial complex, denoted ${\cal N}({\cal G})$. \end{df} Note that the elements of ${\cal G}$ are the vertices of the complex of nested sets ${\cal N}({\cal G})$. Moreover, the order complex of ${\cal G}$ is a subcomplex of ${\cal N}({\cal G})$, since linearly ordered subsets of~${\cal G}$ are nested. \begin{prop}\label{prop_nested} For a given semilattice ${\cal L}$ and a subset $N$ of a building set ${\cal G}$ of~${\cal L}$, the following are equivalent: \begin{enumerate} \item [(1)] $N$ is nested. \item [(2)] Whenever $x_1,\dots,x_t$ are noncomparable elements in~$N$, the join $x_1\vee\dots\vee x_t$ exists, and $F(x_1\vee\dots\vee x_t)=\{x_1,\dots,x_t\}$. \item [(3)] There exists a chain $C\subseteq{\cal L}$, such that $N=\bigcup_{x\in C}F(x)$. \item [(4)] $N\in\Lambda$, where $\Lambda$ is the maximal subset of $2^{{\cal G}}$, for which the following three conditions are satisfied: \begin{enumerate} \item [(o)] $\emptyset\in\Lambda$, and $\{g\}\in\Lambda$, for $g\in {\cal G}$; \item [(i)] if $N\in\Lambda$ and $x\in\max N$, then $N_{<x}\in\Lambda$; \item [(ii)] if $N\in\Lambda$, then $\max N=F(\bigvee \, \max N)$. \end{enumerate} \end{enumerate} \end{prop} \nin{\bf Proof. } \vskip4pt \noindent \underline{(1)$\Rightarrow$(2)}: Let $N$ be a nested set, and let $M\,{=}\,\{x_1,\dots,x_t\}\,{\subseteq}\,N$ be a~set of incomparable elements with $\bigvee_{i=1}^t x_i\,{\not \in}\,{\cal G}$. We can assume that for some $x_j$ we have $(x_j,\bigvee_{i=1}^t x_i]\,{\cap}\,{\cal G}\,{\neq}\,\emptyset$, otherwise the claim follows by Proposition~\ref{prop_buildingsets_properties}~(3). Without loss of generality, we may assume that there exists an element $y$, such that $y\,{\in}\,(x_1,\bigvee_{i=1}^t x_i]\,{\cap}\,{\cal G}$, and $y\,{\in}\,\max {\cal G}_{\leq \bigvee M}$. Define $M'\,{=}\,\{x_1,\dots,x_t\}\,{\cap}\,{\cal G}_{\leq y}\,{=}$ \linebreak $\{x_1\,{=}\,x_{j_0},x_{j_1},\dots,x_{j_k}\}$, and $z\,{=}\,\bigvee_{l=0}^k x_{j_l}$. Since $M'=\{x_{j_0},x_{j_1},\dots,x_{j_k}\}$ is nested (it is a subset of $N$), we have the strict inequality $z\,{<}\,y$. Furthermore, \[ \bigvee_{i=1}^t x_i \, = \, z\vee\bigvee (M\setminus M')\, \leq \, z\vee\bigvee(\max G_{\leq\bigvee M}\setminus\{y\}) \, < \, \bigvee_{i=1}^t x_i\,, \] where the first inequality follows from Proposition~\ref{prop_buildingsets_properties}~(1) and the second inequality from Proposition~\ref{prop_buildingsets_properties}~(2). We thus arrive to a contradiction, which finishes the proof. \vskip4pt \noindent \underline{(2)$\Rightarrow$(1)}: Obvious. \vskip4pt \noindent \underline{(2)$\Rightarrow$(3)}: Let $N$ be a~set satisfying condition~(2). Fix a~particular linear extension $\{x_1,\dots,x_k\}$ on the partial order of~$N$, and define $\alpha_j\,{=}\,x_1\,{\vee}\,\dots\,{\vee}\, x_j$, for $j\,{=}\,1,\dots,k$. By (2) we have $F(\alpha_j)\,{=}\,\max\{x_1,\dots,x_j\}$, and therefore $x_j\,{\in}\, F(\alpha_j)$ and $x_{j+1}\,{\not\in}\, F(\alpha_j)$ for $j\,{=}\,1,\dots,k$. Hence, the $\alpha_j$'s are different and form a~chain $C=\alpha_1<\alpha_2<\dots<\alpha_k$. By construction, $N=\bigcup_{x\in C}F(x)$. \vskip4pt \noindent \underline{(1),(2)$\Rightarrow$(4)}: Let $N$ be a nested set, we shall prove that $N\in\Lambda$ by induction on the size of $N$: \begin{enumerate} \item if $|N|=0$, then $N\in\Lambda$ by condition (o); \item if $|N|\geq 1$, then $\max N=F(\bigvee\,\max N)$ by condition~(2). Furthermore, since $|N_{<x}|<|N|$, and $N_{<x}$ is nested (it is a subset of $N$), $N_{<x}\in\Lambda$ by induction. Hence $N\in\Lambda$. \end{enumerate} \noindent \underline{(3)$\Rightarrow$(1)}: Let $C\,{=}\,(\alpha_1\,{<}\,\dots\,{<}\,\alpha_k)$ be a chain in~${\cal L}$ and $N\,{=}\,\bigcup_{x\in C}F(x)$. Let $N'\,{=}\,\{x_1,\ldots,x_t\}\,{\subseteq}\, N$, $t\,{\geq}\,2$, be an antichain in~$N$, and~$s$ the maximal index in~$C$ such that $N'\,{\cap}\,F(\alpha_s)\,{\neq}\,\emptyset$. In particular, $N'\,{\cap}\,F(\alpha_s)\,{\neq}\,\{\alpha_s\}$ due to $|N'|>1$ and $N'$ being an antichain. Let $y\,{\in}\,N'\,{\cap}\,F(\alpha_s)$. If $|N'\,{\cap}\,F(\alpha_s)|\,{>}\,1$, \[ y\, < \, \bigvee (N'\,{\cap}\,F(\alpha_s)) \, \leq \, \bigvee \, N' \, \leq \, \alpha_s\, , \] where the strict inequality is a consequence of the necessity property for building sets. Thus, $ \bigvee N'\,{\not \in}\, {\cal G}$. If $|N'\,{\cap}\,F(\alpha_s)|\,{=}\,1$, we have $y\,{<}\,\bigvee N'\,{\leq}\,\alpha_s$, due to $N'$ being an antichain with $|N'|\,{>}\,1$, and again $\bigvee N'\,{\not \in}\, {\cal G}$. \vskip4pt \noindent \underline{(4)$\Rightarrow$(3)}: We need the following fact: \vskip4pt \noindent {\bf Fact.} {\it If there exist elements $x_1,\dots,x_t$ and $y_1,\dots,y_k$ in~${\cal L}$, such that $x_t\,{>}\,y_j$, for $j\,{=}\,1,\dots,k$, and $F(\bigvee_{i=1}^t x_i)\,{=}\,\{x_1,\dots,x_t\}$, and $F(\bigvee_{j=1}^k y_j)\,{=}\,\{y_1,\dots,y_k\}$, then $F(x_1\vee\dots\vee x_{t-1}\vee y_1\vee\dots\vee y_k)\,{=}\, \{x_1,\dots,x_{t-1},y_1,\dots,y_k\}.$} \vskip4pt Once the fact above is proved, one can derive (3) as follows. For $N\in\Lambda$ we shall form a chain $C\,{=}\,(\alpha_1\,{<}\,\dots\,{<}\,\alpha_{|N|})$ such that $N\,{=}\,\bigcup_{i=1}^{|N|} F(\alpha_i)$. First, choose a~linear extension $\{x_1,\dots,x_t\}$ of $N$. Then, set $\alpha_t\,{=}\,\bigvee \, \max N$, $\alpha_{t-1}=$ $\bigvee \, \max (N\,{\setminus}\,\{x_t\})$, $\alpha_{t-2}\,{=}\,\bigvee \, \max (N\,{\setminus}\,\{x_t, x_{t-1}\})$, and so on. By (4)(ii), we have $F(\alpha_t)\,{=}\,\max N$. Applying (4)(i) to $x_t\,{\in}\,\max N$, and (4)(ii) to $N_{<x_t}$, we obtain $F(\bigvee\, \max N_{<x_t})\,{=}\,\max N_{<x_t}$. Taking into account the fact above, we conclude that $F(\alpha_{t-1})\,{=}\,\max (N\,{\setminus}\,\{x_t\})$, and, using the same argument iteratively, we arrive to $N\,{=}\,\bigcup_{i=1}^t F(\alpha_i)$. \vskip3pt \noindent {\bf Proof of the fact.} Set $\alpha\,{=}\,x_1\,{\vee}\,\dots\,{\vee}\, x_{t-1}\,{\vee}\, y_1\,{\vee}\,\dots\,{\vee}\, y_k$. Since~$\alpha\,{\leq}\,\bigvee_{i=1}^t\,x_i$, the factors of~$\alpha$ can be partitioned into groups of elements below the $x_i$ for $i\,{=}\, 1,\ldots,t$, by Proposition~\ref{prop_buildingsets_properties}~(1). Since $x_i\,{\leq}\,\alpha$ for $i\,{=}\,1,\ldots,t{-}1$, we obtain $F(\alpha)\,{=}\,\{x_1,\ldots, x_{t-1}, \gamma_1,\ldots, \gamma_{m}\}$ with $\gamma_j\,{\leq}\,x_t$ for $j=1,\ldots,m$. Again using Proposition~\ref{prop_buildingsets_properties}~(1), the $y_1,\dots, y_k$ can be partitioned into groups below the factors $\gamma_j$ for $j\,{=}\,1,\ldots,m$. The occurrence of one strict inequality $\bigvee\,\{y_l\,|\,y_l\,{\leq}\,\gamma_j\}\,{<} \, \gamma_j$, for some $j\in[m]$, yields a contradiction to $$\alpha\,{=}\, \bigvee_{i=1}^{t-1} x_i \,{\vee}\bigvee_{j=1}^{k} y_j{=}\, \bigvee_{i=1}^{t-1} x_i \,{\vee}\,\bigvee_{j=1}^{m} \gamma_j ,$$ due to the necessity property of building sets. Moreover, since the $y_i$ are factors themselves, joins of more than two of the $y_i$'s are not elements of~${\cal G}$. Thus, $y_i\,{=}\,\gamma_i$, for $i\,{=}\,1,\ldots,k{=}m$, as claimed. $\square$ \section{Sequences of combinatorial blowups} \label{sect_comb_blowups} \noindent We introduce the notion of a combinatorial blowup of an~element in a~semilattice and prove that the set of semilattices is closed under this operation. \subsection{Combinatorial blowups} \begin{df} For a semilattice ${\cal L}$ and an element ${\alpha}\in{\cal L}$ we define a poset ${\rm{Bl}\,}_{\alpha}{\cal L}$, the {\bf combinatorial blowup of ${\cal L}$ at ${\alpha}$}, as follows: \begin{itemize} \item[$\circ$] elements of ${\rm{Bl}\,}_{\alpha}{\cal L}$: \begin{enumerate} \item[(1)] $y\in{\cal L}$, such that $y\not\geq{\alpha}$; \item[(2)] $[{\alpha},y]$, for $y\in{\cal L}$, such that $y\not\geq{\alpha}$ and $(y\vee{\alpha})_{\cal L}$ exists \newline (in particular, $[{\alpha},\hat 0]$ can be thought of as the result of blowing up ${\alpha}$); \end{enumerate} \item[$\circ$] order relations in ${\rm{Bl}\,}_{\alpha}{\cal L}$: \begin{enumerate} \item[(1)] $y>z$ in ${\rm{Bl}\,}_{\alpha}{\cal L}$ if $y>z$ in ${\cal L}$; \item[(2)] $[{\alpha},y]>[{\alpha},z]$ in ${\rm{Bl}\,}_{\alpha}{\cal L}$ if $y>z$ in ${\cal L}$; \item[(3)] $[{\alpha},y]>z$ in ${\rm{Bl}\,}_{\alpha}{\cal L}$ if $y\geq z$ in ${\cal L}$; \end{enumerate}\nopagebreak[4] where in all three cases $y,z\not\geq{\alpha}$. \end{itemize} \end{df} Note that the atoms in ${\rm{Bl}\,}_{\alpha}{\cal L}$ are the atoms of~${\cal L}$ together with the element $[\alpha,\hat 0]$. It is easy, albeit tedious, to check that the class of (meet-)semilattices is closed under combinatorial blowups. \begin{lm} Let ${\cal L}$ be a semilattice and ${\alpha}\in{\cal L}$, then ${\rm{Bl}\,}_{\alpha}{\cal L}$ is a semilattice. \end{lm} \nin{\bf Proof. } The joins in ${\rm{Bl}\,}_{\alpha}{\cal L}$ are defined by the rule \begin{align*} ([{\alpha},y]\vee[{\alpha},z])_{{\rm{Bl}\,}_{\alpha}{\cal L}}&=[{\alpha},(y\vee z)_{\cal L}],\\ ([{\alpha},y]\vee z)_{{\rm{Bl}\,}_{\alpha}{\cal L}} &=[{\alpha},(y\vee z)_{\cal L}],\\ (y\vee z)_{{\rm{Bl}\,}_{\alpha}{\cal L}} &=(y\vee z)_{\cal L}, \end{align*} which is applicable only if $(y\vee z)_{\cal L}$ exists, otherwise the corresponding joins in ${\rm{Bl}\,}_{\alpha}{\cal L}$ do not exist. Also, the first and second formulae are applicable only in the case $(y\vee z)_{\cal L}\not\geq{\alpha}$, otherwise the corresponding joins do not exist. We check this by considering four possible cases separately: $$ \begin{cases} [{\alpha},x]\geq[{\alpha},y]\\ {}[{\alpha},x]\geq[{\alpha},z] \end{cases} \Leftrightarrow \begin{cases} x\geq y\\ x\geq z \end{cases} \Leftrightarrow x\geq(y\vee z)_{\cal L} \Leftrightarrow [{\alpha},x]\geq[{\alpha},(y\vee z)_{\cal L}]. $$ $$ \begin{cases} [{\alpha},x]\geq[{\alpha},y]\\ {}[{\alpha},x]\geq z \end{cases} \Leftrightarrow \begin{cases} x\geq y\\ x\geq z \end{cases} \Leftrightarrow x\geq(y\vee z)_{\cal L} \Leftrightarrow [{\alpha},x]\geq[{\alpha},(y\vee z)_{\cal L}]. $$ $$ \begin{cases} x\geq_{{\rm{Bl}\,}_{\alpha}{\cal L}} y\\ x\geq_{{\rm{Bl}\,}_{\alpha}{\cal L}} z\\ x\not\geq{\alpha} \end{cases} \Leftrightarrow \begin{cases} x\geq_{\cal L} y\\ x\geq_{\cal L} z\\ x\not\geq{\alpha} \end{cases} \Leftrightarrow \begin{cases} x\geq(y\vee z)_{\cal L}\\ x\not\geq{\alpha} \end{cases} \Rightarrow \begin{cases} y\vee z\not\geq{\alpha}\\ x\geq(y\vee z)_{\cal L}. \end{cases} $$ $$ \begin{cases} [{\alpha},x]\geq y\\ {}[{\alpha},x]\geq z\\ y,z\not\geq{\alpha} \end{cases} \hskip-3pt \Leftrightarrow \begin{cases} x\geq_{\cal L} y\\ x\geq_{\cal L} z\\ x,y,z\not\geq{\alpha} \end{cases} \hskip-3pt\Leftrightarrow \begin{cases} x\geq(y\vee z)_{\cal L}\\ x,(y\vee z)_{\cal L}\not\geq{\alpha} \end{cases} \hskip-5pt\Rightarrow [{\alpha},x]\geq(y\vee z)_{{\rm{Bl}\,}_{\alpha}{\cal L}}. $$ \noindent Observe that it is possible that $(x\vee y)_{\cal L}$ exists, while $(x\vee y)_{{\rm{Bl}\,}_{\alpha}{\cal L}}$ does not. $\square$ \subsection{Blowing up building sets} % \noindent In this subsection we prove that if one combinatorially blows up a~building set of a~semilattice in any chosen linear extension order, then one ends up with the face poset of the simplicial complex of nested sets of this building set. The following proposition provides the essential step for the proof. \begin{prop}\label{prop_singleblow} Let ${\cal L}$ be a semilattice, ${\cal G}$ a building set of~${\cal L}$, and ${\alpha}\,{\in}\,\max {\cal G}$. Then, $\widetilde {\cal G}=({\cal G}\setminus\{{\alpha}\})\cup\{[{\alpha},\hat 0]\}$ is a building set of ${\rm{Bl}\,}_{\alpha}{\cal L}$. Furthermore, the nested subsets of $\widetilde {\cal G}$ are precisely the nested subsets of ${\cal G}$ with ${\alpha}$ replaced by $[{\alpha},\hat 0]$. \end{prop} \nin{\bf Proof. } It is easy to see that $\widetilde {\cal G}$ is a building set of ${\rm{Bl}\,}_{\alpha}{\cal L}$. Indeed, given $x\in{\cal L}\setminus{\cal L}_{\geq{\alpha}}$, \eqref{eq_buildg} is obvious for $x\in{\rm{Bl}\,}_{\alpha}{\cal L}$, and, if $(x\vee {\alpha})_{\cal L}$ exists, it follows for $[{\alpha},x]\in{\rm{Bl}\,}_{\alpha}{\cal L}$ from the identity \[ [\hat 0,[{\alpha},x]]_{{\rm{Bl}\,}_{\alpha}{\cal L}}\, =\, [\hat 0,x]_{{\rm{Bl}\,}_{\alpha}{\cal L}}\times B_1\, , \] where $B_1$ is the subposet consisting of the two comparable elements $\hat 0\,{<}\,[{\alpha},\hat 0]$. Let us now see that the sets of nested subsets of ${\cal G}$ and $\widetilde {\cal G}$ are the same when replacing ${\alpha}$ by $[{\alpha},\hat 0]$: Let $N$ be a nested set in ${\cal G}$, not containing ${\alpha}$. For incomparable elements $x_1,\ldots,x_t$ in~$N$, $\bigvee_{i=1}^t\,x_i\,{\not \geq}\,{\alpha}$, since otherwise we had $${\alpha}\,{\in}\, \max {\cal G}_{\leq \bigvee x_i} = F(\bigvee_{i=1}^tx_i)\,{=}\, \{x_1,\ldots,x_t\}$$ by Proposition~\ref{prop_nested}(2). Thus, $\bigvee_{i=1}^t x_i$ exists in ${\rm{Bl}\,}_{{\alpha}} {\cal L}$ and $\bigvee_{i=1}^t x_i \,{\not\in}\, \widetilde {\cal G}$. Hence, $N$ is nested in $\widetilde {\cal G}$. A nested subset in $\widetilde {\cal G}$ not containing $[{\alpha},\hat 0]$ is obviously nested in ${\cal G}$. Let now $N$ be nested in~${\cal G}$ containing $\alpha$, and set $\widetilde N\,{=}\,(N\setminus\{{\alpha}\})\cup\{[{\alpha},\hat 0]\}$. Subsets of incomparable elements in $\widetilde N$ not containing $[{\alpha},\hat 0]$ can be dealt with as above. Thus assume that $[{\alpha},\hat 0],x_1,\ldots,x_t$ are incomparable in $\widetilde N$. Then, $x_1,\ldots,x_t$ are incomparable in the nested set~$N$, and, as above, we conclude that $\bigvee_{i=1}^t\,x_i$ exists and $\bigvee_{i=1}^t\,x_i\,{\not \geq}\,{\alpha}$. Moreover, $\alpha \vee \bigvee_{i=1}^t\,x_i$ exists in ${\cal L}$ (joins of nested sets always exist!), thus, $[{\alpha},\bigvee_{i=1}^t\,x_i]\,{=}\,[{\alpha},\hat 0]\vee \bigvee_{i=1}^t\,x_i$ exists in ${\rm{Bl}\,}_{{\alpha}}{\cal L}$ and is obviously not contained in $\widetilde {\cal G}$. We conclude that $\widetilde N$ is nested in $\widetilde {\cal G}$. Vice versa, let $\widetilde N$ be nested in~$\widetilde {\cal G}$ containing $[{\alpha},\hat 0]$, and set $N\,{=}\,(\widetilde N\setminus \{[{\alpha},\hat 0]\})\cup\{{\alpha}\}$. Again it suffices to consider subsets of incomparable elements ${\alpha},x_1,\ldots,x_t$ in $N$. With $[{\alpha},\hat 0],x_1,\ldots,x_t$ incomparable in $\widetilde N$, $[{\alpha},\hat 0]\vee \bigvee_{i=1}^t\,x_i\,{=}\, [{\alpha},\bigvee_{i=1}^t\,x_i]$ exists in ${\rm{Bl}\,}_{{\alpha}}{\cal L}$, thus ${\alpha} \vee \bigvee_{i=1}^t\,x_i$ exists in~${\cal L}$. Incomparability implies that ${\alpha} \vee \bigvee_{i=1}^t\,x_i>{\alpha}$, and thus ${\alpha} \vee \bigvee_{i=1}^t\,x_i \not \in {\cal G}$. We conclude that~$N$ is nested in~${\cal G}$. $\square$ \medskip By iterating the combinatorial blowup described in Proposition~\ref{prop_singleblow} through all of~${\cal G}$, we obtain the following theorem, which serves as a motivation for the entire development. \begin{thm} \label{thm_main} Let ${\cal L}$ be a semilattice and ${\cal G}$ a building set of~${\cal L}$ with some chosen linear extension: ${\cal G}=\{G_1,\dots,G_t\}$, with $G_i>G_j$ implying $i<j$. Let ${\rm{Bl}\,}_k{\cal L}$ denote the result of subsequent blowups ${\rm{Bl}\,}_{G_k}({\rm{Bl}\,}_{G_{k-1}}(\dots{\rm{Bl}\,}_{G_1}{\cal L}))$. Then the final semilattice ${\rm{Bl}\,}_t{\cal L}$ is equal to the face poset of the simplicial complex ${\cal N}({\cal G})$. \end{thm} \nin{\bf Proof. } The building set~${\cal G}_t$ of ${\rm{Bl}\,}_t{\cal L}$ that results from iterated application of Proposition~\ref{prop_singleblow} obviously is the set of atoms ${\mathfrak A}$ in ${\rm{Bl}\,}_t{\cal L}$. Every element $x\in{\rm{Bl}\,}_t{\cal L}$ is the join of atoms below it: $x=\bigvee {\mathfrak A}_{\leq x}$. The subset ${\mathfrak A}_{\leq x}$ of ${\cal G}_t$ is nested, in particular, it is the set of factors of~$x$ in ${\rm{Bl}\,}_t{\cal L}$ with respect to ${\cal G}_t$ (Proposition~\ref{prop_nested}(2)). Proposition~\ref{prop_buildingsets_properties}(2) implies that the interval~$[\hat 0,x]$ in ${\rm{Bl}\,}_t{\cal L}$ is boolean. We conclude that ${\rm{Bl}\,}_t{\cal L}$ is the face poset of a simplicial complex with faces in one-to-one correspondence with the nested sets in ${\cal G}_t$, which in turn correspond to the nested sets in~${\cal G}$ by Proposition~\ref{prop_singleblow}. $\square$ \section{De~Concini-Procesi models of subspace arrangements} \label{ssect_DPmodels} \subsection{Building sets for subspace arrangements} Let ${\cal A}=\{A_1,\ldots,A_n\}$ be an arrangement of linear subspaces in ${\mathbb C}^d$. Much effort has been spent on describing the cohomology of the complement ${\cal M}({\cal A})={\mathbb C}^d\,{\setminus}\,\bigcup {\cal A}$ of such an arrangement and, in particular, on answering the question whether the cohomology algebra is completely determined by the combinatorial data of the arrangement. Here, combinatorial data is understood as the lattice~${\cal L}({\cal A})$ of intersections of subspaces of~${\cal A}$ ordered by reverse inclusion together with the complex codimensions of the intersections. A major step towards the solution of this problem (for a complete answer see~\cite{DGM00, dLS01}) was the construction of smooth models for the complement~${\cal M}({\cal A})$ by De~Concini~ \& Procesi~\cite{DP95} that allowed for an explicit description of rational models for~${\cal M}({\cal A})$ following~\cite{Mo78}. The De~Concini-Procesi models for arrangements in turn are one instance in a sequence of model constructions reaching from compactifications of symmetric spaces~\cite{DP83,DP85}, over the Fulton-MacPherson compactifications of configuration spaces~\cite{FM94} to the general framework of wonderful conical compactifications proposed by MacPherson~ \& Procesi~\cite{MP98}. Given a complex subspace arrangement~${\cal A}$ in~${\mathbb C}^d$, De~Concini \& Procesi describe a smooth irreducible variety~$Y$ together with a proper map $\pi:\, Y\, \longrightarrow \, {\mathbb C}^d$ such that $\pi$ is isomorphism over ${\cal M}({\cal A})$, and the complement of the preimage of ${\cal M}({\cal A})$ is a union of irreducible divisors with normal crossings in~$Y$. The model~$Y$ can be constructed by a~sequence of blowups of smooth subvarieties that is prescribed by the stratification of complex space induced by the arrangement. In order to enumerate the strata in the intersection stratification of~$Y$ given by the irreducible divisors, De~Concini \& Procesi introduced the notions of building sets, nested sets and irreducible elements as follows: \begin{df} \label{df_DPnotions} {\rm (\cite[\S 2]{DP95})} Let ${\cal L}({\cal A})$ be the intersection lattice of an arrangement~${\cal A}$ of linear subspaces in a~finite dimensional complex vector space. Consider the lattice~${\cal L}({\cal A})^*$ formed by the orthogonal complements of intersections ordered by inclusion. \begin{itemize} \item[(1)] For $U\,{\in}\,{\cal L}({\cal A})^*$, $U\, = \oplus_{i=1}^k\, U_i$ with $U_i\,{\in}\,{\cal L}({\cal A})^*$, is called a {\bf decomposition\/} of~$U$ if for any $V\,{\subseteq}\,U$, $V\in {\cal L}({\cal A})^*$, $V\, = \oplus_{i=1}^k\, (U_i\cap V)$ and $U_i\cap V\,{\in}\,{\cal L}({\cal A})^*$ for $i=1,\ldots,k$. \item[(2)] Call $U\,{\in}\,{\cal L}({\cal A})^*$ {\bf irreducible\/} if it does not admit a non-trivial decomposition. \item[(3)] A subset ${\cal G}\,{\subseteq}\,{\cal L}({\cal A})^*$ is called a {\bf building set\/} for~${\cal A}$ if for any $U\,{\in}\,{\cal L}({\cal A})^*$ and $G_1,\ldots, G_k$ maximal in~${\cal G}$ below~$U$, $U\, = \oplus_{i=1}^k\, G_i$ is a decomposition (the ${\cal G}$-decomposition) of~$U$. \item[(4)] A subset ${\cal S}\,{\subseteq}\,{\cal G}$ is called {\bf nested\/} if for any set of non-comparable elements $U_1,\ldots,U_k$ in~${\cal S}$, $U\, = \oplus_{i=1}^k\, U_i$ is the ${\cal G}$-decomposition of~$U$. \end{itemize} \end{df} Note that ${\cal L}({\cal A})^*$ coincides with ${\cal L}({\cal A})$ as abstract lattices. We will therefore talk about irreducible elements, building sets and nested sets in ${\cal L}({\cal A})$ without explicitly referring to the dual setting of the preceding definition. The notions of Definition~\ref{df_DPnotions} are in part based on the earlier notions introduced by Fulton~ \& MacPherson in~\cite{FM94} to study compactifications of configuration spaces. Our terminology is naturally adopted from~\cite{FM94,DP95}. Building sets and nested sets in the sense of De~Concini~ \& Procesi are building and nested sets for the intersection lattices of subspace arrangements in our combinatorial sense (see Proposition~\ref{combgeomprop}~(1) below). However, there are differences. The opposite is not true: A combinatorial building set for the intersection lattice of a subspace arrangement is not necessarily a building set for this arrangement in the sense of De~Concini~ \& Procesi, neither are irreducible elements in the sense of De~Concini~ \& Procesi irreducible in our sense. \begin{exam} Consider the following arrangement of 3 subspaces in~${\mathbb C}^4$: \[ A_1:\,\, z_4 = 0\, , \quad A_2:\,\, z_1= z_2 = 0\, ,\quad A_3:\,\, z_1= z_3 = 0 \,. \] \noindent The intersection lattice is a boolean algebra on 3 elements. $\{A_1,A_2,A_3\}\,{\subseteq}\,{\cal L}({\cal A})$ is a combinatorial building set, in fact, it is the set of irreducibles in the abstract lattice. However, the minimal building set in the sense of De~Concini~ \& Procesi is $\{A_1,A_2,A_3, A_2\,{\cap}\,A_3\}$. \end{exam} The main difference can be formulated in the following way: our constructions are order-theoretically canonical for a given semilattice. The set of combinatorial building sets, in particular the set of irreducible elements, depends only on the semilattice itself and not on the geometry of the subspace arrangement which it encodes. See Proposition~\ref{combgeomprop} for the complete explanation. \subsection{Local subspace arrangements} In order to trace the De~Concini-Procesi construction step by step we need the more general notion of a~local subspace arrangement. \begin{df} \label{lsadf} Let $M$ be a smooth complex $d$-dimensional manifold, and let ${\cal A}$ be a~union of finitely many smooth complex submanifolds of $M$ such that all non-empty intersections of submanifolds in~${\cal A}$ are connected smooth complex submanifolds. ${\cal A}$ is called a~{\bf local subspace arrangement}~if for any $x\in{\cal A}$ there exists an~open set $N$ in~$M$ with $x\,{\in}\,N$, a~subspace arrangement $\widetilde {\cal A}$ in ${{\mathbb C}}^d$, and a~biholomorphic map $\phi:N\rightarrow{{\mathbb C}}^d$, such that $\phi(N\cap{\cal A})=\widetilde {\cal A}$. \end{df} Given a subspace arrangement ${\cal A}$, the initial ambient space~${\mathbb C}^d$ of ${\cal M}({\cal A})$ carries a natural stratification by the subspaces of~${\cal A}$ and their intersections, the poset of strata being the intersection lattice ${\cal L}({\cal A})$ of the arrangement. For a~local subspace arrangement ${\cal A}=\{A_1,\dots,A_n\}$ in $M$ we again consider the stratification of $M$ by all possible intersections of the $A_i$'s, just like in the global case. The poset of strata is also denoted by ${\cal L}({\cal A})$ and is called the intersection semilattice (it is a lattice if the intersection of all maximal strata is nonempty). \begin{df} \label{lsabsdf} Let ${\cal A}$ be a local subspace arrangement and ${\cal L}({\cal A})$ its intersection semilattice. For $U\in{\cal L}({\cal A})$, $U_1,\dots,U_k\in{\cal L}({\cal A})$ are said to form a~{\bf decomposition} of $U$ if for any $x\in U$ there exists an~open set $N$ with $x\,{\in}\,N$ and a~biholomorphic map $\phi:N\rightarrow{{\mathbb C}}^d$, such that $\phi(N\cap U_1),\dots,\phi(N\cap U_k)$ form a~decomposition of $\phi(N\cap U)$ in the sense of {\rm Definition~\ref{df_DPnotions}(1)}. As in the global case, ${\cal G}\subseteq{\cal L}({\cal A})$ is a~{\bf building set} for ${\cal A}$ if for any $U\in{\cal L}({\cal A})$, the set of strata $\max{\cal G}_{\leq U}$ gives a~decomposition of $U$. \end{df} We shall refer to these building sets as {\em geometric\/} building sets. The difference between combinatorial building sets and geometric ones is contained in the dimension function as is explained in the following proposition. \begin{prop} \label{combgeomprop} $\,$ Let ${\cal A}$ be a~local subspace arrangement with intersection semilattice~${\cal L}({\cal A})$. \begin{enumerate} \item[(1)] If ${\cal G}\subseteq{\cal L}({\cal A})$ is a~geometric building set of ${\cal A}$, then it is a~combinatorial building set. \item[(2)] If ${\cal G}\subseteq{\cal L}({\cal A})$ is a~combinatorial building set of ${\cal L}({\cal A})$, and for any $x\in{\cal L}({\cal A})$ the sum of codimensions of its factors is equal to the codimension of $x$, then ${\cal G}$ is a~geometric building set. \end{enumerate} \end{prop} \nin{\bf Proof. } In both cases it is enough to consider the case when ${\cal A}$ is a~subspace arrangement. (1) Consider~${\cal G}$ as a subset of~${\cal L}({\cal A})^*$, then, for $U\,{\in}\,{\cal G}$, the isomorphism $\varphi_U$ requested in Definition~\ref{df_buildg} is given by taking direct sums: $$\varphi_U:\,\,\, \prod_{j=1}^k\,\,\, [\hat 0,G_j] \,\, \stackrel{\oplus_{j=1}^k}{\longrightarrow} \,\, [\hat 0,U]\, ,$$ where $G_1,\ldots,G_k$ are maximal in ${\cal G}$ below~$U$. (2) For $U\in{\cal L}({\cal A})^*$, the set $\{U_1,\dots,U_k\}=\max{\cal G}_{\leq U}$ gives a~decomposition of $U$ because: \begin{enumerate} \item [a)] By the definition of ${\cal L}({\cal A})^*$ and the definition of combinatorial building sets, we have $U=\mbox{\rm span}(U_1,\dots,U_k)$, and, since $\sum_{i=1}^k\dim U_i=\dim U$, we have $U=\bigoplus_{i=1}^k U_i$; \item [b)] for any $V\subseteq U$, $\bigoplus_{i=1}^k(U_i\cap V) \subseteq V=\mbox{\rm span}(U_1\wedge V,\dots,U_k\wedge V)\subseteq \bigoplus_{i=1}^k(U_i\cap V)$, where "$\wedge$" denotes the meet operation in ${\cal L}({\cal A})^*$, hence $V=\bigoplus_{i=1}^k(U_i\cap V)$. $\square$ \end{enumerate} \subsection{Intersection stratification of local arrangements after blowup} Let a space~$X$ be given with an intersection stratification induced by a~local subspace arrangement, and let $G$ be a stratum in~$X$. In the blowup of~$X$ at~$G$, ${\rm{Bl}\,}_{G}X$, we find the following maximal strata: \begin{itemize} \item[$\circ$] maximal strata in~$X$ that do not intersect with~$G$, \item[$\circ$] blowups of maximal strata~$V$ at $G\cap V$, ${\rm{Bl}\,}_{G\cap V}V$, where $V$ is maximal in~$X$ and intersects~$G$, \item[$\circ$] the exceptional divisor $\widetilde G$ replacing~$G$. \end{itemize} We consider the intersection stratification of ${\rm{Bl}\,}_{G}X$ induced by these maximal strata. We will later see (proof of Proposition~\ref{prop_compDP_cbl}) that in case $G$ is maximal in a building set for the local arrangement in~$X$, then the union of maximal strata in ${\rm{Bl}\,}_GX$ is again a local arrangement with induced intersection stratification. In general, this is not the case, see Remark~\ref{rem4.5}. For ease of notation, let us agree here that formally blowing up an empty (non-existing) stratum has no effect on the space. We think about a stratum~$Y$ in~$X$, intersection of all maximal strata $V_1,\ldots,V_t$ that contain~$Y$, as being replaced by the intersection of corresponding maximal strata in~${\rm{Bl}\,}_{G}X$: \begin{equation} \label{eq_blownupstratum} {\rm{Bl}\,}_{G\cap V_1} V_1 \,\, \cap \, \, \ldots \,\, \cap \, \, {\rm{Bl}\,}_{G\cap V_t} V_t\, , \end{equation} (recall that ${\rm{Bl}\,}_{G\cap V_j} V_j\,{=}\,V_j$ for $G\,{\cap}\,V_j\,{=}\,\emptyset$). The intersection~(\ref{eq_blownupstratum}) being empty means that the stratum~$Y$ vanishes under blowup of~$G$. For notational convenience, we most often retain names of strata under blowups, thereby referring to the replacement of strata described above. \begin{rem} \label{rem4.5} Let us mention here that blowing up a stratum in a local subspace arrangement does not necessarily result in a local subspace arrangement again. Consider the following arrangement of 2~planes and 1~line in~${\mathbb C}^3$: \[ A_1:\, \, y-z=0\, ,\quad A_2:\, \, y+z=0\, ,\quad L:\, \, x=y=0\, . \] After blowing up~$L$, the planes $A_1$ and $A_2$ are replaced by complex line bundles over~${\mathbb C} {\rm P}^1$, which have in common their zero section~$Z$ and a complex line~$Y$; $L$ is replaced by a direct product of~${\mathbb C}$ and ${\mathbb C} {\rm P}^1$, which intersects both line bundles in~$Z$. The new maximal strata fail to form a local subspace arrangement in the point $Z\cap Y$. \end{rem} \subsection{Tracing incidence structure during arrangement model construction} \label{ssect_tracg_inc} We now give a more detailed description of the model construction by De~Concini \& Procesi via successive blowups, and then proceed with linking our notion of combinatorial blowups to the context of arrangement models. Let ${\cal A}$ be a complex subspace arrangement, ${\cal G}\,{\subseteq}\,{\cal L}({\cal A})$ a~geometric building set for~${\cal A}$, and $\{G_1,\ldots,G_t\}$ some linear extension of the partial containment order on associated strata in~${\mathbb C}^d$ such that $G_k\,{\supset}\,G_l$ implies $l<k$. The De~Concini-Procesi model $Y=Y_{{\cal G}}$ of ${\cal M}({\cal A})$ is the result of blowing up the strata indexed by elements of~${\cal G}$ in the given order. Note that the linear order was chosen so that at each step the stratum which is to be blown up does {\em not\/} contain any other stratum indexed by an element of~${\cal G}$. At each step we consider intersection stratifications as described above, and we denote the poset of strata after blowup of $G_i$ with~${\cal L}^{{\cal G}}_i({\cal A})$. For the case of a~stratum $G_i$ being empty after previous blowups remember our agreement of considering blowups of $\emptyset$ as having no effect on a space. The later Proposition~\ref{prop_compDP_cbl} however shows that strata indexed by elements in~${\cal G}$ do not disappear during the sequence of blowups. Let us remark that the combinatorial data of the initial stratification, i.e., of the arrangement, prescribes much of the geometry of~$Y_{{\cal G}}$: the complement $Y_{{\cal G}}\,{\setminus}\,{\cal M}({\cal A})$ is a union of smooth irreducible divisors indexed by elements of~${\cal G}$, and these divisors intersect if and only if the set of indices is nested in~${\cal G}$~\cite[Thm 3.2]{DP95}. \begin{prop}\label{prop_compDP_cbl} Let ${\cal A}$ be an arrangement of complex subspaces, ${\cal G}$ a building set for ${\cal A}$ in the sense of De~Concini~ \& Procesi, and $\{G_1,\ldots,G_t\}$ some linear extension of the partial containment order on associated strata as described above. Let ${\rm{Bl}\,}_i^{{\cal G}}({\cal A})$ denote the geometric result of successively blowing up strata $G_1,\ldots,G_i$, for $1\,{\leq}\,i\,{\leq}\,t$. Then, \begin{itemize} \item[(1)] The poset of strata ${\cal L}^{{\cal G}}_i({\cal A})$ of ${\rm{Bl}\,}_i^{{\cal G}}({\cal A})$ can be described as the result of a~sequence of combinatorial blowups of the intersection lattice~${\cal L}\,{=}\,{\cal L}({\cal A})$: \[ {\cal L}^{{\cal G}}_i({\cal A})\, \, = \, \, {\rm{Bl}\,}_i({\cal L})\, ,\qquad \mbox{for }\, 1\,{\leq}\,i\,{\leq}\,t\, . \] {\rm (}Recall that ${\rm{Bl}\,}_i({\cal L})\,{=}\,{\rm{Bl}\,}_{G_i}({\rm{Bl}\,}_{G_{i-1}}(\ldots {\rm{Bl}\,}_{G_1}{\cal L}))$ for $ 1\,{\leq}\,i\,{\leq}\,t$.{\rm )} \item[(2)] The union of maximal strata~${\cal A}_i^{{\cal G}}$ in ${\rm{Bl}\,}_i^{{\cal G}}({\cal A})$ is a local subspace arrangement, with ${\cal G}$ in ${\cal L}^{{\cal G}}_i({\cal A})$ being a~building set for ${\cal A}_i^{{\cal G}}$ in the sense of \/ {\rm Definition~\ref{lsabsdf}}. {\rm (}Recall that ${\cal G}$ here refers to the preimages of the original strata in ${\cal G}\,{\subseteq}\,{\cal L}({\cal A})$ under the sequence of blowups.{\rm )} \end{itemize} \end{prop} \noindent \nin{\bf Proof. } We proceed by induction on the number of blowups. The induction start is obvious, since the lattice of strata~${\cal L}_0^{{\cal G}}({\cal A})$ of the initial stratification of ${\mathbb C}^d$ coincides with the intersection lattice ${\cal L}({\cal A})\,{=}\,{\rm{Bl}\,}_0({\cal L})$ of the arrangement~${\cal A}$. The union of maximal strata is the arrangement~${\cal A}$ itself with its given building set~${\cal G}$. Assume that ${\cal L}_{i-1}^{{\cal G}}({\cal A})={\rm{Bl}\,}_{i-1}({\cal L})$ for some $ 1\,{\leq}\,i\,{\leq}\,t$, the union of maximal strata ${\cal A}_{i-1}^{{\cal G}}$ in ${\rm{Bl}\,}_{i-1}^{{\cal G}}({\cal A})$ being a local arrangement, and ${\cal G}$ a building set for ${\cal L}_{i-1}^{{\cal G}}({\cal A})$. Let $G\,{=}\,G_i$ be the next stratum to be blown up. First, we proceed in 4 steps to show that ${\cal L}^{{\cal G}}_i({\cal A})\,{=}\,{\rm{Bl}\,}_i({\cal L})$. In 2 further steps we then verify the claims in~(2). \vskip4pt \noindent {\bf Step 1:} {\em Assign strata of ${\rm{Bl}\,}_i^{{\cal G}}({\cal A})$ to elements in~${\rm{Bl}\,}_i({\cal L})$.\/} \vskip4pt \noindent We distinguish two types of elements in ${\rm{Bl}\,}_i({\cal L})$: \[ \begin{array}{lrl} \mbox{Type~I}: & Y & \mbox{with }\, Y\,{\in}\,{\rm{Bl}\,}_{i-1}({\cal L}) \mbox{ and}\, Y \not\geq G\, , \\ \mbox{Type~~I$\!$I}:\quad & [G,Y] & \mbox{with }\,Y\,{\in}\,{\rm{Bl}\,}_{i-1}({\cal L})\, , \, \, Y \not\geq G\, , \\ & & \mbox{and }\, Y\,{\vee}\,G \mbox{ exists in }\,{\rm{Bl}\,}_{i-1}({\cal L})\, . \end{array} \] \noindent To $Y\,{\in}\,{\rm{Bl}\,}_i({\cal L})$ of type~I, assign ${\rm{Bl}\,}_{G\cap Y}Y$ (recall that blowing up an empty stratum does not change the space). Note that $\dim{\rm{Bl}\,}_{G\cap Y}Y=\dim Y$. \noindent To $[G,Y]\,{\in}\,{\rm{Bl}\,}_i({\cal L})$ of type~I$\!$I, assign $({\rm{Bl}\,}_{G\cap Y}Y)\, \cap \, \widetilde G$, where $\widetilde G$ denotes the exceptional divisor that replaces $G$ in ${\rm{Bl}\,}_i^{{\cal G}}({\cal A})$. This description comprises $\widetilde G$ being assigned to $[G,\hat 0]$. Note that $\dim({\rm{Bl}\,}_{G\cap Y}Y)\, \cap \, \widetilde G=\dim Y-1$. \vskip4pt \noindent {\bf Step 2:} {\em Reverse inclusion order on the assigned spaces coincides with the partial order on~${\rm{Bl}\,}_i({\cal L})$.\/} \vskip4pt \noindent (1) $X,Y\,{\in}\, {\rm{Bl}\,}_i({\cal L})$, both of type~I: \[ X\, \leq_{{\rm{Bl}\,}_i({\cal L})} Y \, \Leftrightarrow \, X\, \leq_{{\rm{Bl}\,}_{i-1}({\cal L})} Y \, \Leftrightarrow \, X \supseteq_{{\rm{Bl}\,}_{i-1}^{{\cal G}}({\cal A})} Y \, \Leftrightarrow \, {\rm{Bl}\,}_{G\cap X} X \supseteq {\rm{Bl}\,}_{G\cap Y} Y , \] where ``$\Leftarrow$'' in the last equivalence can be seen by first noting that $Y \,{\setminus}\, (G\,{\cap}\,Y) \subseteq X \,{\setminus}\, (G\,{\cap}\,X)$, and then comparing points in the exceptional divisors. \vskip4pt \noindent (2) $X, [G,Y]\,{\in}\, {\rm{Bl}\,}_i({\cal L})$, $X$ of type~I, $[G,Y]$ of type~I$\!$I: \newline As above we conclude \begin{eqnarray*} X\, \leq_{{\rm{Bl}\,}_i({\cal L})} [G,Y] & \Leftrightarrow & X\, \leq_{{\rm{Bl}\,}_{i-1}({\cal L})} Y \\ & \Leftrightarrow & X \supseteq_{{\rm{Bl}\,}_{i-1}^{{\cal G}}({\cal A})} Y \,\, \Rightarrow \,\, {\rm{Bl}\,}_{G\cap X} X \supseteq {\rm{Bl}\,}_{G\cap Y} Y \cap \widetilde G\, . \end{eqnarray*} To prove the converse is rather subtle. Note first that $G\,{\cap}\,Y \subseteq G\,{\cap}\,X$. Assume that $G$ strictly contains~$G\,{\cap}\,X$, then both $G\,{\cap}\,X$ and $G\,{\cap}\,Y$ are not in the building set due to the linear order chosen on~${\cal G}$, and~$G$ is a factor of both $G\,{\cap}\,X$ and $G\,{\cap}\,Y$. Let $F(G\,{\cap}\,X)\,{=}\,\{G,G_1,\ldots,G_k\}$, $F(G\,{\cap}\,Y)\,{=}\,\{G,H_1,\ldots,H_t\}$. $X$ written as a join of elements in~${\rm{Bl}\,}_{i-1}({\cal L})$ below the factors of~$G\,{\cap}\,X$ reads \[ X \, = \, g_X \vee Z_1 \vee \ldots \vee Z_k \] for some $g_X\,{\in}\,[\hat 0,G]$, $Z_i\,{\in}\,[\hat 0, G_i]$, for $i=1,\ldots, k$. If $Z_i\,{<}\,G_i$, for some $i\in\{1,\ldots,k\}$, we have \begin{eqnarray*} G\vee X & = & G \vee (g_X \vee Z_1 \vee \ldots \vee Z_i \vee \ldots \vee Z_k) \\ & \leq & G \vee (g_X \vee G_1 \vee \ldots \vee Z_i \vee \ldots \vee G_k) \\ & < & G \vee G_1 \vee \ldots \vee G_k \, = \, G \vee X\, , \end{eqnarray*} by the ``necessity'' property of the Proposition~\ref{prop_buildingsets}(4), yielding a contradiction. Hence, \[ X \, = \, g_X \vee G_1 \vee \ldots \vee G_k\, , \] and similarly, $Y\,{=}\, g_Y \vee H_1 \vee\ldots \vee H_t$ for some $g_Y\,{\in}\,[\hat 0,G]$. For each $j\,{\in}\,\{1,\ldots,k\}$ there exists a unique $i_j\,{\in}\,\{1,\ldots,t\}$ such that $G_j\leq H_{i_j}$ by Proposition~\ref{prop_buildingsets_properties}(1). Thus, $\bigvee G_i\,{<}\, \bigvee H_j$, and, for showing that $X\leq Y$, it is enough to see that~$g_X\leq g_Y$. We show that in an open neighborhood of any point $y\,{\in}\,G\,{\cap}\,Y$, $g_Y\,{\subseteq}\,g_X$. This yields our claim since strata in~${\rm{Bl}\,}_{i-1}^{{\cal G}}({\cal A})$ have pairwise transversal intersections: if they coincide locally, they must coincide globally. With ${\cal A}_{i-1}^{{\cal G}}$ being a local arrangement, there exists an open neighborhood of $y\,{\in}\,G\,{\cap}\,Y$ where the stratification is biholomorphic to a stratification induced by a subspace arrangement. We tacitly work in the arrangement setting, using that $({\rm{Bl}\,}_{i-1}({\cal L}))_{\leq G\vee Y}$ is the intersection lattice of a product arrangement. The ${\cal G}$-decomposition of $(G\vee Y)^{\perp}$ described in Definition~\ref{lsabsdf} yields (when transferred to the primal setting): \[ g_Y\, = \,\mbox{\rm span}(G,Y)\, . \] Analogously, $g_X\,{=}\,\mbox{\rm span}(G,X)$. In the linear setting we are concerned with, we interpret points in the exceptional divisor of a blowup as follows: \begin{equation}\label{eq_blowup} {\rm{Bl}\,}_{G\cap Y} Y \cap \widetilde G\,\, = \,\, \{ (a,\mbox{\rm span}(p,G\cap Y)) \, | \, a\in G\cap Y,\, p\in Y\setminus (G\cap Y) \}\, . \end{equation} In terms of this description, the inclusion map ${\rm{Bl}\,}_{G\cap Y} Y \cap \widetilde G \hookrightarrow {\rm{Bl}\,}_G({\rm{Bl}\,}_{i-1}^{{\cal G}}({\cal A}))$ reads \[ (a,\mbox{\rm span}(p,G\cap Y)) \longmapsto (a,\mbox{\rm span}(p,G))\, . \] Therefore, ${\rm{Bl}\,}_{G\cap Y} Y \cap \widetilde G$ being contained in ${\rm{Bl}\,}_{G\cap X} X\,{\subseteq}\, {\rm{Bl}\,}_G({\rm{Bl}\,}_{i-1}^{{\cal G}}({\cal A}))$ means that for $(a,\mbox{\rm span}(p,G\cap Y))\in {\rm{Bl}\,}_{G\cap Y} Y \,{\cap}\, \widetilde G$ there exists $q\,{\in}\, X\setminus (G\cap X)$ such that $\mbox{\rm span}(p,G)\, = \, \mbox{\rm span}(q,G)$. In particular, $\mbox{\rm span}(Y,G)\,{\subseteq}\,\mbox{\rm span}(X,G)$, which by our previous arguments implies that $Y\,{\subseteq}\,X$. We assumed above that $G\,{\supset}\,G\,{\cap}\,X$. If $G\,{\cap}\,X$ coincides with~$G$, i.e., $X$ contains~$G$, then $g_X\,{=}\,X$ and a similar reasoning applies to see that $Y\,{\subseteq}\,X$. Similarly for $G\,{\cap}\,X\,{=}\,G\,{\cap}\,Y\,{=}\,G$. \vskip4pt \noindent (3) $[G,X]$, $[G,Y]\,{\in}\, {\rm{Bl}\,}_i({\cal L})$, both of type~I$\!$I: \begin{eqnarray*} [G,X]\leq_{{\rm{Bl}\,}_i({\cal L})} [G,Y] & \Leftrightarrow & X\, \leq_{{\rm{Bl}\,}_{i-1}({\cal L})} Y \\ & \Leftrightarrow & X \supseteq_{{\rm{Bl}\,}_{i-1}^{{\cal G}}({\cal A})} Y \,\, \Leftrightarrow \,\, {\rm{Bl}\,}_{G\cap X} X \cap \widetilde G \supseteq {\rm{Bl}\,}_{G\cap Y} Y \cap \widetilde G\, , \end{eqnarray*} where ``$\Leftarrow$'' follows from~(2) and ${\rm{Bl}\,}_{G\cap X} X\,{\supseteq}\, {\rm{Bl}\,}_{G\cap X} X \cap \widetilde G\,{\supseteq}\, {\rm{Bl}\,}_{G\cap Y} Y \cap \widetilde G$. \vskip4pt \noindent {\bf Step 3:} {\em Each of the assigned spaces is the intersection of maximal strata in ${\rm{Bl}\,}_i^{{\cal L}}({\cal A})$.} \vskip4pt \noindent It is enough to show that spaces assigned to elements of type~I in~${\rm{Bl}\,}_i({\cal L})$ are intersections of new maximal strata. Those associated to elements of type~I$\!$I then are intersections as well by definition. Let $Y\,{\in}\,{\rm{Bl}\,}_i({\cal L})$, $Y\,{\not\geq}\,G$, and $Y\,{=}\,\cap_{i=1}^t\, V_i$ with $V_1,\ldots,V_t$ the maximal strata in~${\rm{Bl}\,}_{i-1}^{{\cal G}}({\cal A})$ containing~$Y$. We claim that \begin{equation} \label{eq_intdiv} {\rm{Bl}\,}_{G\cap Y}Y\, \, = \, \, \bigcap_{i=1}^t\,{\rm{Bl}\,}_{G\cap V_i}V_i\, . \end{equation} For the inclusion~``$\subseteq$'' note that ${\rm{Bl}\,}_{G\cap Y}Y\,{\subseteq}\,{\rm{Bl}\,}_{G\cap V_i}V_i$ is a direct consequence of $Y\,{\subseteq}\, V_i$ as discussed in Step~2~(1). For the reverse inclusion we need the following identity: \begin{equation}\label{eq_join} \bigvee_{i=1}^t\, (G\wedge V_i)\,\, = \, \, G\wedge Y\, . \end{equation} This identity holds in any semilattice without referring to $G$ being an element of the building set. Let $\alpha \,{\in}\, \cap_{i=1}^t\,{\rm{Bl}\,}_{G\cap V_i}V_i$. In case $\alpha \,{\in}\, \cap_{i=1}^t\, V_i\, {\setminus}\, (G\,{\cap}\,V_i)$, we conclude that $\alpha\,{\in}\, Y\, {\setminus}\, (G\,{\cap}\,Y) $. We thus assume that $\alpha$ is contained in the intersection of exceptional divisors $\widetilde{G\,{\cap}\,V_i}$, $i=1,\ldots,t$. We again switch to local considerations in the neighborhood of a point~$y\,{\in}\,G\,{\cap}\,Y$, using that it carries a stratification biholomorphic to an arrangement stratification. Using the description (\ref{eq_blowup}) of points in exceptional divisors that are created by blowups in the arrangement setting, $\alpha\,{\in}\,\cap_{i=1}^t\,\widetilde{G\,{\cap}\,V_i}\,{\subseteq}\, \cap_{i=1}^t\,{\rm{Bl}\,}_{G\cap V_i}V_i$ means that there exist $a\,{\in}\, \cap_{i=1}^t\, (G\,{\cap}\,V_i)$, and $p_i\,{\in}\, V_i\, {\setminus}\, (G\,{\cap}\,V_i)$ for $i=1,\ldots,t$, with \[ \alpha \,\, = \, \, (a,\mbox{\rm span}(p_i,G\,{\cap}\,V_i))\, \in \, {\rm{Bl}\,}_{G\cap V_i}V_i\, . \] In particular, $\mbox{\rm span}(p_i,G)\,{=}\,\mbox{\rm span}(p_j,G)$ for $1\,{\leq}\,i,j\,{\leq}\, t$. Thus, \[ \mbox{\rm span}(p_j,G) \, \subseteq \bigcap_{i=1}^t\, \mbox{\rm span}(V_i,G)\, \, = \, \, \mbox{\rm span}(Y,G) \] using the identity~(\ref{eq_join}). We conclude that there exists $y\,{\in}\, Y\,{\setminus}\,(G\,{\cap}\,Y)$ such that $\mbox{\rm span}(y,G)\, {=}\, \mbox{\rm span}(p_j,G)$ for all $j\,{\in}\,\{1,\ldots,k\}$, hence \[ \alpha\, = \, (a,\mbox{\rm span}(y,G\,{\cap}\,Y))\,\,\in\,\, {\rm{Bl}\,}_{G\cap Y}Y\, . \] Though we are for the moment not concerned with the case of $Y\,{\subseteq}\,G$, we note for later reference that~(\ref{eq_intdiv}) remains true, with ${\rm{Bl}\,}_Y Y\,{=}\,\emptyset$ meaning that the intersection on the right-hand side is empty. Following the proof of the inclusion~``$\supseteq$'' in~(\ref{eq_intdiv}) for $G\,{\cap}\,Y\,{=}\,Y$, we first find that the intersection of blowups can only contain points in the exceptional divisors. Assuming~$\alpha\,{\in}\,\cap_{i=1}^t\, \widetilde{G\,{\cap}\,V_i}$ we arrive to a contradiction when concluding that $\mbox{\rm span}(p_j,G)\,{\subseteq}\, \cap_{i=1}^t\,\mbox{\rm span}(V_i,G)\,{=}$ $\mbox{\rm span}(Y,G)\,{=}\,G$ for $j=1,\ldots,t$. \vskip4pt \noindent {\bf Step 4:} {\em Any intersection of maximal strata in ${\rm{Bl}\,}_i^{{\cal G}}({\cal A})$ occurs as an assigned space.} \vskip4pt \noindent Every intersection involving the exceptional divisor~$\widetilde G$ occurs if we can show that all other intersections occur (intersections that additionally involve~$\widetilde G$ then are assigned to corresponding elements of type~I$\!$I). Consider $W\, = \, \bigcap_{i=1}^t\,{\rm{Bl}\,}_{G\cap V_i}V_i$, where the $V_i$ are maximal strata in ${\rm{Bl}\,}_{i-1}^{{\cal G}}({\cal A})$; recall here that a blowup in an empty stratum does not alter the space. We can assume that $\cap_{i=1}^t\, V_i\,{\not =}\, \emptyset$, otherwise the intersection~$W$ were empty. With the identity~(\ref{eq_intdiv}) in Step~3 we conclude that either $W\,{=}\, \emptyset$ (in case $\cap_{i=1}^t\, V_i\,{\subseteq}\,G$) or $W\,{=}\,{\rm{Bl}\,}_{G\cap \bigcap_{i=1}^t V_i} \cap_{i=1}^t\, V_i$, in which case it is assigned to the element $\cap_{i=1}^t\, V_i$ in ${\rm{Bl}\,}_i({\cal L})$. \vskip4pt \noindent {\bf Step 5:} {\em ${\cal A}_{i}^{{\cal G}}$ is a local subspace arrangement in ${\rm{Bl}\,}_i^{{\cal G}}({\cal A})$.} \vskip4pt \noindent It follows from the description~(\ref{eq_intdiv}) of strata in ${\rm{Bl}\,}_i^{{\cal G}}({\cal A})$ that all intersections of maximal strata are connected and smooth. It remains to show that ${\cal A}_{i}^{{\cal G}}$ locally looks like a subspace arrangement. Let $y\,{\in}\,{\cal A}_{i}^{{\cal G}}$. We can assume that $y$ lies in the exceptional divisor~$\widetilde G$. Let $x \in G\,{\subseteq}\,{\cal A}_{i-1}^{{\cal G}}$ be the image of~$y$ under the blowdown map. We first give a local description around $x$ in ${\cal A}_{i-1}^{{\cal G}}$. By induction hypothesis, there exists a neighborhood~$N$ of~$x$, and an arrangement of linear subspaces~${\cal B}$ in~${\mathbb C}^n$ such that the pair $(N,{\cal A}_{i-1}^{{\cal G}}\,{\cap}\,N)$ is biholomorphic to the pair~$({\mathbb C}^n,{\cal B})$. We can assume that under this biholomorphic map, $x$ is mapped to the origin. Let $T\,{=}\,\bigcap_{B\in {\cal B}}\,B$ and note that $G\,{\cap}\,N$ is mapped to some subspace~$\Gamma$ in ${\cal B}$. With $G$ being maximal in the building set for~${\cal A}_{i-1}^{{\cal G}}$, ${\cal B}/T$ is a product arrangement with one of the factors being an arrangement in~$\Gamma/T$. More precisely, there exists a subspace $\Gamma'\,{\subseteq}\,{\mathbb C}^n$, and two subspace arrangements, ${\cal C}$ in $\Gamma/T$ and ${\cal C}'$ in $\Gamma'/T$, such that \begin{itemize} \item[(1)] $\Gamma/T\, \oplus\, \Gamma'/T\, \oplus\, T\, = \, {\mathbb C}^n$, \item[(2)] ${\cal B}\, =\ \{A\, \oplus\, \Gamma'/T\, \oplus\, T\,| \, A\,{\in}\,{\cal C}\} \, \cup \, \{\Gamma/T\, \oplus\, A'\, \oplus\, T\,| \, A'\,{\in}\,{\cal C}'\}$. \end{itemize} Blowing up $G$ in ${\rm{Bl}\,}_{i-1}^{{\cal G}}({\cal A})$ locally corresponds to blowing up $\Gamma$ in ${\mathbb C}^n$. Let $t$ be the point on the special divisor $\widetilde \Gamma$ corresponding to~$y \in \widetilde G$, thus $t$ maps to the origin in~${\mathbb C}^n$ under the blowdown map. A neighborhood of~$t$ in ${\rm{Bl}\,}_{\Gamma}{\mathbb C}^n$ is an $n$-dimensional open ball which can be parameterized as a direct sum \[ M \,\oplus \, M' \, \oplus \, I \, \oplus \,T \, . \] Here, $M$ is an open ball around~$0$ in $\Gamma/T$, $M'$ is an open ball on the unit sphere in $\Gamma'/T$ around the point of intersection with the line~$\langle p \rangle$ in $\Gamma'/T$ that defines $t$ as a point in the exceptional divisor, $t=(0, \mbox{\rm span}(p,\Gamma)) \in \widetilde \Gamma$ (compare (\ref{eq_blowup})), and $I$ an open unit ball in~${\mathbb C}$. The maximal strata in this neighborhood are the following: \begin{itemize} \item[$\circ$] the hyperplane $M \,\oplus \, M' \, \oplus \, \{0\} \, \oplus \,T$, as the exceptional divisor, \item[$\circ$] $(M \cap A)\,\oplus \, M' \, \oplus \, I \, \oplus \,T$, replacing $A\, \oplus\, \Gamma'/T\, \oplus\, T$ after blowup, \item[$\circ$] $M \,\oplus \, (M'\cap A') \, \oplus \, I \, \oplus \,T$, replacing $\Gamma/T\, \oplus\, A'\, \oplus\, T$ after blowup for $A'\neq 0$. \end{itemize} This proves that around $t$ in ${\rm{Bl}\,}_{\Gamma}{\mathbb C}^n$ we have the structure of a local subspace arrangement, which in turn shows the local arrangement property around $y$ in ${\cal A}_i^{{\cal G}}$. \vskip4pt \noindent {\bf Step 6:} {\em ${\cal G}$ is a building set for ${\cal A}_i^{{\cal G}}$ in the sense of \/{\rm Definition~\ref{lsabsdf}}.} \vskip4pt \noindent ${\cal G}$ is a~combinatorial building set by Proposition~\ref{prop_singleblow}. Complementing this with the dimension information about the strata, we conclude, by Proposition~\ref{combgeomprop}(2), that ${\cal G}$ is a~geometric building set. $\square$ \section{Simplicial resolutions of toric varieties} \label{ssect_tv} The study of toric varieties has proved to be a field of fruitful interplay between algebraic and convex geometry: toric varieties are determined by rational polyhedral fans, and many of their algebraic geometric properties are reflected by combinatorial properties of their defining fans. We recall one such correspondence -- between subdivisions of fans and special toric morphisms -- and show that so-called stellar subdivisions are instances of combinatorial blowups. This allows us to apply our Main Theorem in the present context: Given a polyhedral fan, we specify a class of {\em simplicial\/} subdivisions, and, interpreting our notions of building sets and nested sets, we describe the incidence combinatorics of the subdivisions in terms of the combinatorics of the initial fan. For background material on toric varieties we refer to the standard sources~\cite{Da78,Od88,Ful93,Ew96}. Let $X_{\Sigma}$ be a toric variety defined by a rational polyhedral fan~$\Sigma$. Any subdivision of~$\Sigma$ gives rise to a proper, birational toric morphism between the associated toric varieties~(cf~\cite[5.5.1]{Da78}). In particular, simplicial subdivisions yield toric morphisms from quasi-smooth toric varieties to the initial variety -- so-called {\em simplicial resolutions}. Quasi-smooth toric varieties being rational homology manifolds, such morphisms can replace smooth resolutions for (co)homological considerations. We define a particular, elementary, type of subdivisions: \begin{df} \label{def_stellsd} Let $\Sigma\,{=}\,\{\sigma\}_{\sigma \in \Sigma}\,{\subseteq}\,{\mathbb R}^d$ be a polyhedral fan, i.e., a collection of closed polyhedral cones~$\sigma$ in~${\mathbb R}^d$ such that $\sigma\,{\cap}\,\tau$ is a cone in~$\Sigma$ for any $\sigma, \tau\,{\in}\,\Sigma$. Let ${\rm cone}(x)$ be a ray in~${\mathbb R}^d$ generated by $x\,{\in}\,{\rm relint}\, \tau$ for some $\tau\,{\in}\,\Sigma$. The {\em stellar subdivision\/} ${\rm sd}(\Sigma,x)$ of $\Sigma$ in $x$ is given by the collection of cones \[ (\,\Sigma\,\setminus\, {\rm star}(\tau, \Sigma)\, ) \,\, \cup \, \, \{\, {\rm cone}(x,\rho)\, | \, \rho \subseteq \sigma \,\, \mbox{ for some } \,\, \sigma \in {\rm star}(\tau, \Sigma)\, \}\, , \] where ${\rm star}(\tau, \Sigma)\,{=}\,\{\sigma\,{\in}\,\Sigma\, | \, \tau\,{\subseteq}\,\sigma\}$, and ${\rm cone}(x,\rho)$ the closed polyhedral cone spanned by~$\rho$ and~$x$. If only concerned with the combinatorics of the subdivided fan, we also talk about stellar subdivision of $\Sigma$ in $\tau$, ${\rm sd}(\Sigma,\tau)$, meaning any stellar subdivision in~$x$ for~$x\,{\in}\,{\rm relint}\, \tau$. \end{df} \begin{prop} Let ${\cal F}(\Sigma)$ be the face poset of a polyhedral fan~$\Sigma$, i.e., the set of closed cones in~$\Sigma$ ordered by inclusion, together with the zero cone~$\{0\}$ as a minimal element. For $\tau\,{\in}\,\Sigma$, the face poset of the stellar subdivision of $\Sigma$ in $\tau$ can be described as the combinatorial blowup of ${\cal F}(\Sigma)$ at~$\tau$: \[ {\cal F}({\rm sd}(\Sigma, \tau))\,\, = \,\, {\rm{Bl}\,}_{\tau}{\cal F}(\Sigma)\, . \] \end{prop} \nin{\bf Proof. } We observe that removing ${\rm star}(\tau, \Sigma)$ from $\Sigma$ corresponds to removing ${\cal F}(\Sigma)_{\geq \tau}$ from ${\cal F}(\Sigma)$, while adding cones, as described in Definition~\ref{def_stellsd}, corresponds to extending ${\cal F}(\Sigma)\,{\setminus}\,{\cal F}(\Sigma)_{\geq \tau}$ by elements $[\tau,\rho]$ for $\rho\,{\in}\,{\cal F}(\Sigma)$, $\rho\,{\subseteq}\,\sigma$ for some $\sigma \,{\in}\,{\rm star}(\tau, \Sigma)$. The comparison of order relations is straightforward. $\square$ \vspace{0.2cm} We apply our Main Theorem to the present context. \begin{thm} Let~$\Sigma$ be a polyhedral fan in~${\mathbb R}^d$ with face poset~${\cal F}(\Sigma)$. Let ${\cal G}\,{\subseteq}\,{\cal F}(\Sigma)$ be a building set of~${\cal F}(\Sigma)$ in the sense of Definition~\ref{df_buildg}, ${\cal N}({\cal G})$ the complex of nested sets in~${\cal G}$ (cf.\ Definition~\ref{df_nested}). Then, the consecutive application of stellar subdivisions in every cone $G\,{\in}\,{\cal G}$ in a non-increasing order yields a simplicial subdivision of~$\Sigma$ with face poset equal to the face poset of~${\cal N}({\cal G})$. \end{thm} \noindent As examples of building sets for face lattices of polyhedral fans let us mention: \begin{itemize} \item[(1)] the full set of faces, with the corresponding complex of nested sets being the order complex of~${\cal F}(\Sigma)$ (stellar subdivision in all cones results in the barycentric subdivision of the fan); \item[(2)] the set of rays together with the non-simplicial faces of~$\Sigma$; \item[(3)] the set of irreducible elements in ${\cal F}(\Sigma)$: the set of rays together with all faces of~$\Sigma$ that are not products of some of their proper faces. \end{itemize} \begin{rem} \label {rm_smooth_tv} For a smooth toric variety~$X_{\Sigma}$, the union of closed codimension~1 torus orbits is a local subspace arrangement, in particular, the codimension~1 orbits form a divisor with normal crossings, \cite[p.~100]{Ful93}. The intersection stratification of this local arrangement coincides with the torus orbit stratification of the toric variety. For any face $\tau$ in the defining fan~$\Sigma$, the torus orbit ${\cal O}_{\tau}$ together with all orbits corresponding to rays in~$\Sigma$ form a~geometric building set. Our proof in the subsection~\ref{ssect_tracg_inc} applies in this context with ${\cal O}_{\tau}$ playing the role of~$G$. We conclude that under blowup of~$X_{\Sigma}$ in the closed torus orbit ${\cal O}_{\tau}$, the incidence combinatorics of torus orbits changes exactly in the way described by a stellar subdivision of~$\Sigma$ in~$\tau$. This is the combinatorial part of the well-known fact that in the smooth case, the blowup of $X_{\Sigma}$ in a torus orbit ${\cal O}_{\tau}$ corresponds to a regular stellar subdivision of the fan~$\Sigma$ in $\tau$~\cite{MO73}. \end{rem} \chapter[Spaces of Complex Monic Polynomials and Complexes of Forests] {Rational Homology of Spaces of Complex Monic Polynomials with Multiple Roots and Complexes of Marked Forests } \section{The stratification by root multiplicities of the space of complex monic polynomials} Let $n$ be an~integer, $n\geq 2$. We view $n$-dimensional complex space ${\mathbb C}^n$ as the space of all monic polynomials with complex coefficients of degree $n$ by identifying $a=(a_0,\dots,a_{n-1})\in{\mathbb C}^n$ with $f_a(z)=z^n+a_{n-1}z^{n-1}+\dots+a_0$. To each $\lambda=(\lambda_1,\dots,\lambda_t)\vdash n$ one can associate a~topological space as follows (we refer the reader to the Appendix~A for a~description of our conventions on the terminology of number and set partitions). \begin{df} \label{df1.1} ${\wti\Sigma}_\lam$ is the set of all $a\in{\mathbb C}^n$, for which the roots of $f_a(z)$ can be partitioned into sets of sizes $\lambda_1,\dots,\lambda_t$, so that within each set the roots are equal. Clearly, ${\wti\Sigma}_\lam$ is a~closed subset of $\,{\mathbb C}^n$. Let $\Sigma_\lam^{S^2}$ be the one-point compactification of~${\wti\Sigma}_\lam$. \end{df} It is easy to see that the Definition~\ref{df1.1} is equivalent to the description of $\Sigma_\lam^{S^2}$ given in the Chapter~1, so our notation is consistent in that point. In this chapter we shall focus on the (reduced) rational Betti numbers of the spaces $\Sigma_\lam^{S^2}$. In~\cite{Ar70a}, V.I.\ Arnol'd has computed $\tilde\beta_*(\Sigma_\lam^{S^2},{\Bbb Q})$ for $\lambda=(k^m,1^{n-km})$. \begin{thm}\label{arn} {\rm (Arnol'd, \cite{Ar70a}).} Let $\lambda=(k^m,1^{n-km})$ for some natural numbers $k\geq 2$, $m$, and $n\geq km$. Then \begin{equation}\label{are} \tilde\beta_i(\Sigma_\lam^{S^2},{\Bbb Q})=\begin{cases} 1,& \mbox{for } i=2l(\lambda);\\ 0,& \mbox{otherwise.} \end{cases} \end{equation} \end{thm} In \cite{SuW97} Sundaram and Welker conjectured that \begin{conj}\label{c3} For any number partition $\lambda$, $\tilde\beta_i(\Sigma_\lam^{S^2},{\Bbb Q})=0$ unless $i=2l(\lambda)$. \end{conj} In this chapter we shall give a new, combinatorial proof of the Theorem~\ref{arn} and disprove Conjecture~\ref{c3}. To do that, we shall introduce a family of topological spaces $X_{\lambda,\mu}$, indexed by pairs of number partitions $(\lambda,\mu)$, satisfying $\lambda\vdash\mu$. $X_{\lambda,\mu}$ will be defined so that the following equality is satisfied \begin{equation}\label{rswf} \tilde\beta_i(\Sigma_\lam^{S^2},{\Bbb Q})=\sum_{\lambda\vdash\mu\vdash n}\tilde\beta_{i-2l(\mu)-1}(X_{\lambda,\mu},{\Bbb Q}). \end{equation} \noindent Here is the summary of the chapter. \vskip4pt \noindent {\bf Section~\ref{s2}.} We define the topological spaces $X_{\lambda,\mu}$ and derive \eqref{rswf}. \vskip4pt \noindent {\bf Section~\ref{s3}.} We give a~combinatorial description of the cell structure of the triangulated spaces~$X_{\lambda,\mu}$ in terms of marked forests, see Theorem~\ref{main1}. This description is the backbone of the chapter, it serves as both language and intuition for the material in the subsequent sections. One consequence of Theorem~\ref{main1} is that homology groups of $X_{\lambda,\mu}$ may be computed from a~chain complex, whose components are freely generated by marked forests, and the boundary operator is described in terms of a~combinatorial operation on such forests (deletion of level sets). \vskip4pt \noindent {\bf Section \ref{s4}.} We prove a~general theorem about collapsibility of certain triangulated spaces. The direct combinatorial argument is heavily relying on the combinatorial cell description, from Section~\ref{s3}, of~$X_{\lambda,\mu}$. We would like to mention that for the special cases $\lambda=(k,1^{n-k})$ and $\lambda=(k^m)$ the new proof of the Theorem~\ref{arn} was also found in~\cite{SuW97}, the argument there also made use of the Theorem~\ref{swt} (as Proposition~\ref{tgen} shows, the case $\lambda=(k^m)$ is especially simple). However, the Theorem~\ref{main2} is the first combinatorial (modulo Theorem~\ref{swt}) proof of the result of Arnol'd in the general case. \vskip4pt \noindent {\bf Section~\ref{s5}.} We disprove the conjecture of Sundaram and Wel\-ker. Besides giving a~counterexample we prove this conjecture for a~class of number partitions, which we call generic partitions. \section{Orbit arrangements and spaces $X_{\lambda,\mu}$} \label{s2} \subsection{Reformulation in the language of orbit arrangements} \noindent Following \cite{SuW97}, we give a~different interpretation of the numbers $\tilde\beta_i(\Sigma_\lam^{S^2},{\Bbb Q})$, for general $\lambda$. First, let us observe that the symmetric group ${\cal S}_n$ acts on ${\mathbb C}^n$ by permuting the coordinates, so we can consider the space ${\mathbb C}^n/{\cal S}_n$ endowed with the quotient topology. It is a~classical fact that the map $\phi:{\mathbb C}^n\rightarrow{\mathbb C}^n/{\cal S}_n$, mapping a~polynomial to the (unordered) set of its roots, is a~homeomorphism, which extends to the one-point compactifications. Therefore ${\mathbb C}^n\cong{\mathbb C}^n/{\cal S}_n$ and $\Sigma_\lam^{S^2}\cong\phi(\Sigma_\lam^{S^2})=\phi({\wti\Sigma}_\lam)\cup\{\infty\}$. $\phi({\wti\Sigma}_\lam)$ can be viewed as the configuration space of $n$ unmarked points on~${\mathbb C}$ such that the number partition given by the coincidences among the points is refined by~$\lambda$. For example, $\phi(\widetilde \Sigma_{(2,1^{n-2})})$ is the configuration space of $n$ unmarked points on ${\mathbb C}$ such that at least 2 points coincide. Using this point of view, $\phi(\Sigma_\lam^{S^2})$ can be described in the language of orbit subspace arrangements. \begin{df} For $\pi\vdash[n]$, $\pi=(S_1,\dots,S_t)$, $S_j=\{i_1^j,\dots,i_{|S_j|}^j\}$, $1\leq j\leq t$, $K_\pi$ is the~sub\-space given by the equations $x_{i_1^1}=\dots=x_{i_{|S_1|}^1},\dots,x_{i_1^t}=\dots=x_{i_{|S_t|}^t}$. For $\lambda\vdash n$, set $I_\lambda=\{\pi\vdash[n]\,|$ $\mbox{\rm type}\,(\pi)=\lambda\}$ and define ${\cal A}_\lambda=\{K_\pi\,|\,\pi\in I_\lambda\}$. ${\cal A}_\lambda$'s are called {\bf orbit arrangements}. \end{df} The orbit arrangements were introduced in \cite{Bj94} and studied in further detail in~\cite{Ko97}. They provide the appropriate language to describe $\phi(\Sigma_\lam^{S^2})$, indeed \begin{equation}\label{fisla} \phi(\Sigma_\lam^{S^2})=\Gamma_{{\cal A}_\lambda}^{{\cal S}_n}. \end{equation} An~important special case is that of the braid arrangement ${\cal A}_{n-1}={\cal A}_{(2,1^{n-2})}$, which corresponds under $\phi$ to ${\widetilde \Sigma}_{(2,1^{n-2})}$, the~space of all monic complex polynomials of degree~$n$ with at least one multiple root. The name ``braid arrangement'' stems from the fact that ${\mathbb C}^n\setminus V_{{\cal A}_{n-1}}$ is a~classifying space of the colored braid group, see~\cite{Ar69}. The intersection lattice ${\cal L}_{{\cal A}_{n-1}}$ is usually denoted $\Pi_n$. It is the~poset consisting of all set partitions of~$[n]$, where the partial order relation is refinement. Furthermore, for $\lambda\vdash n$, the intersection lattice of ${\cal A}_\lambda$ is denoted $\Pi_\lambda$. It is the~subposet of $\Pi_n$ consisting of all elements which are joins of elements of type~$\lambda$, with the minimal element $\hat 0$ attached. \subsection{Applying Sundaram-Welker formula} \noindent The following formula of S.~Sun\-da\-ram and V.~Welker, \cite{SuW97}, is vital for our approach. \begin{thm}\label{swt}{\rm (\cite[Theorem 2.4(ii) and Lemma 2.7(ii)]{SuW97}).}\newline Let ${\cal A}$ be an~arbitrary subspace arrangement in ${\mathbb C}^n$ with an~action of a~finite group $G\subset\mbox{\bf U}_n({\mathbb C})$. Let ${\cal D}_{\cal A}$ be the intersection of $V_{\cal A}$ with the $(2n-1)$-sphere (often called the link of ${\cal A}$). Then there is the following isomorphism of $G$-modules. \begin{equation}\label{swf} \widetilde H_i({\cal D}_{\cal A},{\Bbb Q})\cong_G\bigoplus_{x\in{\cal L}_{\cal A}^{>\hat 0}/G} \mbox{\rm Ind}_{\mbox{Stab}\,(x)}^G (\widetilde H_{i-\dim x}(\Delta({\cal L}_{\cal A}(\hat 0,x)),{\Bbb Q})), \end{equation} where the sum is taken over representatives of the orbits of $G$ in ${\cal L}_{\cal A}\setminus\{\hat 0\}$, under the action of $G$, one representative for each orbit. \end{thm} Clearly $\Gamma_{\cal A}^G\cong\mbox{susp}\,\,({\cal D}_{\cal A}/G)$. Recall that if a~finite group~$G$ acts on a~finite cell complex $K$ then $\tilde\beta_i(K/G,{\Bbb Q})$ is equal to the multiplicity of the trivial representation in the induced representation of $G$ on the ${\Bbb Q}$-vector space $\widetilde H_i(K,{\Bbb Q})$, see for example \cite[Theorem 1]{Co56}, \cite{Br72}. Hence, it follows from \eqref{swf}, and the Frobenius reciprocity law, that \begin{equation}\label{star} \tilde\beta_i(\Gamma_{\cal A}^G,{\Bbb Q})=\sum_{x\in{\cal L}_{\cal A}^{>\hat 0}/G} \tilde\beta_{i-\dim x-1}(\Delta({\cal L}_{\cal A}(\hat 0,x))/{\mbox{Stab}\,(x)},{\Bbb Q}), \end{equation} where $\mbox{Stab}\,(x)$ denotes the stabilizer of $x$. \subsection{Spaces $X_{\lambda,\mu}$ and their properties} \noindent Let us now restate this identity in the special case of orbit arrangements. As mentioned above, the intersection lattice of ${\cal A}_\lambda$ is $\Pi_\lambda$. It has an~action of the symmetric group ${\cal S}_n$, which, for any $\pi\in\Pi_\lambda$ induces an~action of $\mbox{Stab}\,(\pi)$ on $\Delta(\Pi_\lambda(\hat 0,\pi))$. \vskip4pt {\bf Notation. }{\it Let $X_{\lambda,\mu}$ denote the topological space $\Delta(\Pi_\lambda(\hat 0,\pi))/\mbox{Stab}\,(\pi)$, where the set partition~$\pi$ has type $\mu$. If there is no set partition $\pi\in\Pi_\lambda$ of type $\mu$, i.e., if $\mu$ cannot be obtained as a join of $\lambda$'s, then let $X_{\lambda,\mu}$ be a~point.} \vskip4pt For fixed $\mu$, the space $X_{\lambda,\mu}$ does not depend on the choice of $\pi$. Observe that $X_{\lambda,\mu}$ is in general not a~simplicial complex, however it is a~triangulated space, (a~regular CW complex with each cell being a~simplex, see~\cite[Chapter~I, Section~1]{GeM96}), with its cell structure inherited from the simplicial complex $\Delta(\Pi_\lambda(\hat 0,\pi))$. In general, whenever $G$ is a~finite group which acts on a~poset $P$ in an~order-preserving way, $\Delta(P)/G$ is a~triangulated space whose cells are orbits of simplices of $\Delta(P)$ under the action of $G$; this is obviously not true in general for an~action of a~finite group on a~finite simplicial complex. Clearly, \eqref{fisla} together with \eqref{star}, and the fact that $\phi$ is a~homeomorphism, implies \eqref{rswf}. Let us quickly analyze \eqref{rswf}. $X_{\lambda,\lambda}=\emptyset$ makes a~contribution~1 in dimension $2l(\lambda)$. Assume $\mu\neq\lambda$, then $1\leq l(\mu) \leq l(\lambda)-1$ and $X_{\lambda,\mu}\neq\emptyset$. $\dimX_{\lambda,\mu}=l(\lambda)-l(\mu)-1$, hence $\tilde\beta_{i-2l(\mu)-1}(X_{\lambda,\mu},{\Bbb Q})=0$ unless $0\leq i-2l(\mu)-1\leq l(\lambda)-l(\mu)-1$, that is $2l(\mu)+1\leq i\leq l(\lambda)+l(\mu)$. It follows from \eqref{rswf} that $$\tilde\beta_{2l(\lambda)}(\Sigma_\lam^{S^2},{\Bbb Q})=1,\mbox{ and }\tilde\beta_i(\Sigma_\lam^{S^2},{\Bbb Q})=0\mbox{ unless } 3\leq i\leq 2l(\lambda).$$ The purpose of this chapter is to investigate the values $\tilde\beta_i(\Sigma_\lam^{S^2},{\Bbb Q})$ for $3\leq i\leq 2l(\lambda)-1$, by studying $\tilde\beta_i(X_{\lambda,\mu},{\Bbb Q})$. We shall prove that the latter are equal to~0 for a~certain set of pairs $(\lambda,\mu)$, $\lambda\vdash\mu$, of partitions, including the case in Theorem~\ref{arn}, ($\lambda=(k^m,1^{n-km})$, $\mu$ is arbitrary such that $\lambda\vdash\mu$), and we shall give an~example that this is not the case in general. \section{The cell structure of $X_{\lambda,\mu}$ and marked forests} \label{s3} \subsection{The terminology of marked forests} \noindent In order to index the simplices of $X_{\lambda,\mu}$ we need to introduce some terminology for certain types of trees with additional data. For an~arbitrary forest of rooted trees $T$ (we only consider finite graphs), let $V(T)$ denote the set of the vertices of~$T$, $R(T)\subseteq V(T)$ denote the set of the roots of~$T$ and $L(T)\subseteq V(T)$ denote the set of the leaves of~$T$. For any integer $i\geq 0$, let $l_i(T)$ be the number of $v\in V(T)$ such that, $v$ has distance $i$ to the root in its connected component. \begin{df} A forest of rooted trees $T$ is called a~{\bf graded forest of rank $r$} if $l_{r+2}(T)=0$, $l_{r+1}(T)=|L(T)|$, and the sequence $l_0(T),\dots,l_{r+1}(T)$ is strictly increasing. \end{df} For $v,w\in V(T)$, $w$ is called {\it a~child} of $v$ if there is an~edge between $w$ and $v$ and the unique path from $w$ to the corresponding root passes through $v$. For $v\in V(T)$, we call the distance from $v$ to the closest leaf the {\it height} of $v$. For example, in a~graded forest of rank $r$, leaves have height 0 and roots have height $r+1$. \begin{df}\label{df2.2} A {\bf marked forest of rank $r$} is a~pair $(T,\eta)$, where $T$ is a~graded forest of rank $r$ and $\eta$ is a~function from $V(T)$ to the set of natural numbers such that for any vertex $v\in V(T)\setminus L(T)$ we have \begin{equation}\label{mc1} \eta(v)=\sum_{w\in\mbox{\rm{children}}(v)}\eta(w). \end{equation} \end{df} We remark that the set of the marked forests of rank $r$, such that not all leaves have label~1, is equal to the set of graded forests of rank $r+1$. Indeed, instead of labeling the vertices with natural numbers so that~\eqref{mc1} is satisfied, one can as well attach a~new level of leaves so that each ``old leaf'' $v$ has $\eta(v)$ children. Then the old labels will correspond to the numbers of the new leaves below each vertex. For our context it is more convenient to use labels rather than auxiliary leaves, i.e.,~it is more handy to label all vertices rather than just the leaves, so we stick to the terminology of Definition~\ref{df2.2}. For a~marked forest $(T,\eta)$ of rank $r$ and $0\leq i\leq r+1$, we have a~number partition $\lambda_i(T,\eta)=\{\eta(v)\,|\,v\mbox{ has height }i\}$. Clearly $\lambda_0(T,\eta)\vdash\dots\vdash\lambda_r(T,\eta)\vdash\lambda_{r+1}(T,\eta)$. \begin{df}\label{lmd} Let $\lambda\vdash\mu\vdash n$, $\lambda\neq\mu$. A {\bf $(\lam,\mu)$-forest of rank $r$} is a~marked forest of rank $r$, $(T,\eta)$ such that $\mu=\lambda_{r+1}(T,\eta)$ and $\lambda\vdash\lambda_0(T,\eta)$. \end{df} To simplify the language, we call $((2,1^{n-2}),\mu)$-forests simply {\it $\mu$-forests}, and $((2,1^{n-2}),(n))$-forests simply {\it $n$-trees}. \vskip-10pt \hskip-100pt $$\epsffile{p1.eps}$$ \begin{center}\vskip-2pt Figure 3.1. All 5-trees of rank 2. \end{center} Whenever $(T,\eta)$ is a~$(\lam,\mu)$-forest of rank $r$ and $0\leq i\leq r$, we can obtain a~$(\lam,\mu)$-forest $(T^i,\eta^i)$ of rank $r-1$ by deleting from $T$ all the vertices of height~$i$ and connecting the vertices of height~$i+1$ to their grandchildren (unless $i=0$); $\eta^i$ is the restriction of $\eta$ to $V(T^i)$. In other words, $(T^i,\eta^i)$ is obtained from $(T,\eta)$ by removing the entire $i$th level, counting from the leaves, and filling in the gap in an~obvious way. This allows us to define a~boundary operator by \begin{equation}\label{eqb} \partial(T,\eta)=\sum_{i=0}^r(-1)^i(T^i,\eta^i). \end{equation} This paves the way to explicit combinatorial computations by means of marked forests. \vskip-10pt \hskip-100pt $$\epsffile{p2.eps}$$ \begin{center}\vskip-2pt Figure 3.2. An example of a boundary computation. \end{center} For a~given set partition $\pi$ one can define the notion of a~{\bf $\pi$-forest $(T,\zeta)$ of rank $r$} almost identically to the case of number partitions described above. The difference is that $\zeta$ maps $V(T)$ to the set of finite sets, rather than the set of natural numbers. The condition~\eqref{mc1} is replaced~by \begin{equation}\label{mc2} \zeta(v)=\bigcup_{w\in\mbox{\rm{children}}(v)}\zeta(w), \end{equation} and $\pi=\{\zeta(v)\,|\,v\in R(T)\}$. For $0\leq i\leq r+1$, analogously to $\lambda_i(T,\eta)$, we define $\pi_i(T,\zeta)$ to be the set partition which is read off from the vertices of $T$ having height~$i$. Let $\mu$ be the type of $\pi$, then there exists a~canonical $\mu$-forest $(T,|\zeta|)$ associated to each $\pi$-forest $(T,\zeta)$, where $|\zeta|$ is obtained as the composition of $\zeta$ with the map which maps finite sets to their sizes. \subsection{The main theorem} \noindent Let us describe how to associate a $(\lam,\mu)$-forest, $\psi(\sigma)$, of rank $r$ to an $r$-simplex $\sigma$ of $X_{\lambda,\mu}$. The simplex $\sigma$ is an~$\mbox{Stab}\,(\pi)$-orbit of $r$-simplices of $\Delta(\Pi_\lambda(\hat 0,\pi))$, where $\pi$ is a set partition of type~$\mu$. Take a representative of this orbit, a~chain $c=(x_r>\dots>x_0)$. Now we define $\psi(\sigma)=(T,\eta)$. Each element $x_i$ corresponds to the $i$th level in $T$, counting from the leaves. Each block $b$ of $x_i$ corresponds to a node in the tree; on this node we define the value of~$\eta$ to be $|b|$. We define the edges of the tree $T$ by connecting each node corresponding to a block $b$ of $x_i$ to all nodes corresponding to the blocks of $x_{i-1}$ contained in $b$, we do that for all $b$ and $i$. The top $(r+1)$th level is added artificially, its nodes correspond to the blocks of $\pi$, and the edges from the top level to the $r$th level connect each block of $\pi$ to the blocks of $x_r$ contained in it. For example, the value of $\psi$ on the ${\cal S}_5$-orbit of the chain $(123)(45)>(123)(4)(5)>(13)(2)(4)(5)$ is the first 5-tree on the Figure~3.1. We are now ready to state and prove the main result of this section. \begin{thm}\label{main1} Assume $\lambda\vdash\mu\vdash n$, $\lambda\neq\mu$. The correspondence $\psi$ of the $r$-simplices of $X_{\lambda,\mu}$ and $(\lam,\mu)$-forests $(T,\eta)$ of rank $r$ is a~bijection. Under this bijection, the boundary operator of the triangulated space $X_{\lambda,\mu}$ corresponds to the boundary ope\-ra\-tor described in~\eqref{eqb}. \end{thm} In particular, the simplices of $\Delta(\Pi_n)/{\cal S}_n$ along with the cell inclusion structure are described by the $n$-trees. Indeed, $\Delta(\Pi_5)/{\cal S}_5$ is shown in the Figure~3.3. \vskip-5pt \hskip-100pt $$\epsffile{p3.eps}$$ \begin{center}\vskip-5pt Figure 3.3 \end{center} \noindent The five triangles may be labeled by the five 5-trees of rank 2 in Figure~3.1. \vskip4pt \noindent {\bf Proof of the Theorem \ref{main1}.} By the definitions of $\Pi_n$ and of $\Delta$, the $r$-simplices of $\Delta(\Pi_n)$ can be indexed by $([n])$-trees of rank~$r$ (we write $([n])$ to emphasize that the set $[n]$ is viewed here as a~set partition consisting of only one set). Furthermore, the cell inclusions in $\Delta(\Pi_n)$ correspond to level deletion in $([n])$-trees as is described above for the case of number partitions, because the levels in the $([n])$-trees correspond to the elements of $\Pi_n$, and the edges in the $([n])$-trees correspond to block inclusions of two consecutive elements in the chain. More generally, the $r$-simplices $\sigma$ of $\Delta(\Pi_\lambda(\hat 0,\pi))$ can be indexed by $\pi$-forests $(T(\sigma),\zeta(\sigma))$ such that $\lambda$ refines the type of $\pi_{0}(T(\sigma),\zeta(\sigma))$. The definition of $\psi$ can now be rephrased as {\it associating to $\sigma$ the $(\lam,\mu)$-forest $(T(\sigma),|\zeta(\sigma)|)$,} where $\mu=\,$type$\,\pi$. The group action of $\mbox{Stab}\,(\pi)$ on $\Delta(\Pi_\lambda(\hat 0,\pi))$ corresponds to relabeling elements within the sets of $\zeta(\sigma)$. This shows that for $g\in\mbox{Stab}\,(\pi)$ we have $T(g\sigma)=T(\sigma)$ and $|\zeta(g\sigma)|=|\zeta(\sigma)|$. Therefore $\psi(\sigma)$ is well-defined, it does not depend on the choice of the representative of the, corresponding to $\sigma$, $\mbox{Stab}\,(\pi)$-orbit of chains. $\psi$ is surjective, we shall now show that it is also injective. If $\sigma_1,\sigma_2$ are two different $r$-simplices of $\Delta(\Pi_\lambda(\hat 0,\pi))$ such that $T(\sigma_1)=T(\sigma_2)$ and $|\zeta(\sigma_1)|= |\zeta(\sigma_2)|$, then there exists $g\in\mbox{Stab}\,(\pi)$ such that $\zeta(g\sigma_2)= \zeta(\sigma_1)$. Indeed, let $T=T(\sigma_1)=T(\sigma_2)$ and let $\alpha_1$, resp.~$\alpha_2$, be the string concatenated from the values of $\zeta(\sigma_1)$, resp.~$\zeta(\sigma_2)$, on the leaves of $T$; the order of leaves of $T$ is arbitrary, but the same for $T(\sigma_1)$ and $T(\sigma_2)$, the order of elements within each $\zeta(\sigma_1)(v)$, resp.~$\zeta(\sigma_2)(v)$, for $v\in L(T)$ is also chosen arbitrarily. Then $g\in{\cal S}_n$ which maps $\alpha_2$ to $\alpha_1$ satisfies the necessary conditions: $\zeta(g\sigma_2)=\zeta(\sigma_1)$ on the leaves of~$T$, and hence by~\eqref{mc2} on all vertices of~$T$. Furthermore, since $g\pi=g\pi_{r+1}(T,\zeta(\sigma_2))=\pi_{r+1}(T,\zeta(g\sigma_2))= \pi_{r+1}(T,\zeta(\sigma_1))=\pi$, we have $g\in\mbox{Stab}\,(\pi)$. This shows that $\psi$ is a~bijection. Since the levels of the $(\lambda,\mu)$-forests correspond to the $\mbox{Stab}\,(\pi)$-orbits of the vertices of $\Delta(\Pi_\lambda(\hat 0,\pi))$ (hence to the vertices of $X_{\lambda,\mu}$), the boundary operator of $X_{\lambda,\mu}$ corresponds under $\psi$ to the level deletion in $(\lam,\mu)$-forests, i.e.,~the boundary operator described in \eqref{eqb}. $\square$ \subsection{Remarks} \noindent 1. While the presence of the root in an~$([n])$-tree is just a~formality (two marked $([n])$-trees are equal iff the deletion of the root gives equal marked forests), the presence of the roots in a~$\pi$-forest is vital. In fact, if roots were not taken into account (as seems natural, since the partition read off from the roots does not correspond to any vertex in $\Delta(\Pi_\lambda(\hat 0,\pi))$) the argument above would be false already for vertices: if $\tau_1,\tau_2\in \Pi_\lambda(\hat 0,\pi)$, such that type$\,(\tau_1)=\,$type$\,(\tau_2)$ (i.e.,~the corresponding marked forests of rank~0 are equal once the roots are removed), there may not exist $g\in\mbox{Stab}\,(\pi)$, such that $g\tau_2=\tau_1$ (although such $g\in{\cal S}_n$ certainly exists). \vskip4pt \noindent 2. Marked forests equipped with an~order on the children of each vertex were used by Vassiliev, \cite{Vas94}, to label cells in a~certain CW-complex structure on the space $\widetilde {\mathbb R}^n(m)$, the one-point compactification of the configuration space of~$m$ unmarked distinct points in ${\mathbb R}^n$. Vassiliev's cell decomposition of $\widetilde {\mathbb R}^n(m)$ is a~generalization of the earlier Fuchs' cell decomposition of $\widetilde {\mathbb R}^2(m)=\widetilde {\mathbb C}(m)$, \cite{Fuc70}, which allowed Fuchs to compute the ring $H^*(Br(m),{\Bbb Z}_2)$, where $Br(m)$ is Artin's braid group on $m$ strings, see also~\cite{Vai78}. Beyond a~certain similarity of the combinatorial objects used for labeling the cells, cf.~\cite[Lemma 3.3.1]{Vas94} and Theorem~\ref{main1}, the connection between the results of this chapter and the results of Vassiliev and Fuchs seems unclear. As yet another instance of a~similar situation, we would like to mention the labeling of the components in the stratification of $\overline{M}_{0,n}$ (the Deligne-Knudsen-Mumford compactification of the moduli space of stable projective complex cur\-ves of genus $0$ with $n$ punctures) with trees with $n$ labeled leaves, see, e.g., \cite{FM94,Kn83}. \section{A new proof of a~theorem of Arnol'd} \label{s4} \subsection{Formulation of the main theorem and its corollaries} \noindent In this section we take a~look at a~rather general question of which ${\Bbb Q}$-acyclicity of the spaces $X_{\lambda,\mu}$ is a~special case: \begin{quote} Let $\pi\in\Pi_n$ and let $Q$ be an~$\mbox{Stab}\,(\pi)$-invariant subposet of $\Pi_n(\hat 0,\pi)$. When is the multiplicity of the trivial representation in the induced representation of $\mbox{Stab}\,(\pi)$ on $\widetilde H_i(\Delta(Q),{\Bbb Q})$ equal to~0 for all $i$, in other words, when is $\Delta(Q)/\mbox{Stab}\,(\pi)$ ${\Bbb Q}$-acyclic? \end{quote} \begin{df} Let $\Lambda$ be a~subset of the set of all number partitions of $n$ such that $(1^n),(n)\notin\Lambda$. Define $\Pi_\Lambda$ to be the subposet of $\Pi_n$ consisting of all set partitions $\pi$ such that $(\mbox{\rm type }\pi)\in\Lambda$. \end{df} Clearly, $\Pi_\Lambda$ is ${\cal S}_n$-invariant and, more generally, $\Pi_\Lambda(\hat 0,\pi)$ is $\mbox{Stab}\,(\pi)$-invariant. Vice versa, any ${\cal S}_n$-invariant subposet of $\Pi_n\setminus\{(\{1\},\dots,\{n\}),([n])\}$ is of the form $\Pi_\Lambda$ for some~$\Lambda$. The following theorem is the main result of this section. The proof is a~combination of the language of marked forests from Section~\ref{s3} and the ideas used in the proof of~\cite[Theorem~4.1]{Ko00a}. \begin{thm}\label{main2} Let $2\leq k<n$. Assume $\Lambda$ is a~subset of the set of all number par\-ti\-tions of~$n$ such that $(1^n),(n)\notin\Lambda$ and $\Lambda$ satisfies the following condition: \begin{quote} {\rm Condition $C_k$.} If $\mu\in\Lambda$, such that $\mu=(\mu_1,\dots,\mu_t)$, where $\mu_i=kq_i+r_i$, $0\leq r_i<k$ for $i\in[t]$, then $\gamma_k(\mu)=(k^{q_1+\dots+q_t},1^{r_1+\dots+r_t})\in\Lambda$. \end{quote} Then for any $\mu\in\Lambda\cup\{(n)\}$ the triangulated space $X_{\Lambda,\mu}=\Delta(\Pi_\Lambda(\hat 0,\pi))/\mbox{Stab}\,(\pi)$, where $\mu=\,${\rm type}$\,\pi$, is collapsible, in particular the multiplicity of the trivial representation in the induced $\mbox{Stab}\,(\pi)$-representation on $\widetilde H_i(\Delta(\Pi_\Lambda(\hat 0,\pi)),{\Bbb Q})$ is equal to~0 for all~$i$. \end{thm} \begin{rem} Consider the following special case: $k=2$ and $\Lambda$ is equal to the set of all number partitions of~$n$ except for $(1^n)$ and $(n)$. Then the Condition $C_k$ is obviously satisfied. Since in this case $X_{\Lambda,\mu}=\Delta(\Pi_n)/{\cal S}_n$, we conclude that the complex $\Delta(\Pi_n)/{\cal S}_n$ is collapsible. \end{rem} Slightly more generally, we have the following result. \begin{crl} Assume $\lambda=(k^m,1^{n-km})$, $\lambda\vdash\mu$, then $X_{\lambda,\mu}$ is collapsible. In particular, $X_{\lambda,\mu}$ is ${\Bbb Q}$-acyclic, therefore Theorem~\ref{arn} follows. \end{crl} \nin{\bf Proof. } Clearly $X_{\lambda,\mu}=X_{\Lambda,\mu}$ for $\Lambda=\{\tau\,|\,\lambda\vdash\tau\vdash\mu,\tau\neq(n)\}$. It is easy to check that Condition $C_k$ is satisfied for the case $\lambda=(k^m,1^{n-km})$, therefore Theorem~\ref{arn} follows from Theorem~\ref{main2} via~\eqref{rswf}. $\square$ \vskip4pt Another consequence of Theorem~\ref{main2} concerns the multiplicity of the trivial character in certain representations of the symmetric group. Let $\bf k$ be a field such that either char$\,{\bf k}=0$ or char$\,{\bf k}>n$. Following a~conjecture of R.\ Stanley, \cite[page 151]{St82}, P.\ Hanlon has proved in~\cite[Theorem 3.1]{Ha83} that if $\Pi_n^t$ denotes the $\{1,\dots,t\}$ rank selection of the partition lattice, then the multiplicity of the trivial character in the natural representation ${\cal S}_n\rightarrow GL(H_i(\Delta(\Pi_n^t),\bf k))$ induced by the standard permutation ${\cal S}_n$-representation on the set $[n]$, is equal to $0$ for all $i$ and $t$. The following corollary generalizes his result. \begin{crl Assume $\Lambda$ is as in the Theorem~\ref{main2}, then the multiplicity of the trivial character in the representation ${\cal S}_n\rightarrow GL(H_i(\Delta(\Pi_\Lambda),\bf k))$ is $0$ for all $i$. \end{crl} \nin{\bf Proof. } We know that the complex $X_\Lambda=\Delta(\Pi_\Lambda)/{\cal S}_n$ is collapsible. By \cite[Theorem 1]{Co56}, $\beta_i(X_\Lambda,{\bf k})$ is equal to the multiplicity of the trivial character in the representation ${\cal S}_n\rightarrow GL(H_i(\Delta(\Pi_\Lambda),\bf k))$. Thus the statement of the corollary is equivalent to saying that $X_\Lambda$ is ${\bf k}$-acyclic, which in turn is immediate from Theorem~\ref{main2}. $\square$ \vskip4pt Theorem~\ref{main2} can be viewed as an~attempt to provide a~common framework for these results in the spirit of the question stated in the beginning of this section. \subsection{Auxiliary propositions} \noindent First we need some terminology. For an arbitrary cell complex $\Delta$ we denote by $V(\Delta)$ the set of vertices of $\Delta$. Assume $\Delta$ is a regular CW complex and $\Delta'$ is its subcomplex. We denote the set of the simplices of $\Delta$ which are not simplices of $\Delta'$ by $\Delta\setminus\Delta'$. We use the sign $\succ$ to denote the cover relation in the cell structure of~$\Delta$. Assume that, in addition, $\Delta$ is a triangulated space with some linear order $\ll$ on the set of vertices. For $\sigma\in\Delta\setminus\Delta'$ we may write $\sigma=(x_1,\dots,x_t)$, this notation is slightly inaccurate since the set of vertices does not determine the simplex uniquely, all we mean is that $\sigma$ has vertices $x_1\ll\dots\ll x_t$. In that case, we let $\xi(\sigma)=i$ if $x_1,\dots,x_{i-1}\in V(\Delta')$ and $x_i\notin V(\Delta')$. \begin{prop}\label{pr1} Let $\Delta$ be a~regular CW complex and $\Delta'$ a~subcomplex of $\Delta$, then the following are equivalent: a) there is a~sequence of collapses leading from $\Delta$ to $\Delta'$; b) there is a~matching of cells of $\Delta\setminus\Delta'$: $\sigma\leftrightarrow \phi(\sigma)$, such that $\phi(\sigma)\succ\sigma$ and there is no sequence $\sigma_1,\dots\sigma_t\in\Delta\setminus\Delta'$ such that $\phi(\sigma_1)\succ\sigma_2, \phi(\sigma_2)\succ\sigma_3,\dots,\phi(\sigma_t)\succ\sigma_1$ (such matching is called acyclic). \end{prop} \nin{\bf Proof. } \vskip4pt \noindent \underline{a) $\Rightarrow$ b)}: Let the elementary collapses define the matching $\phi$. Assume there is a~sequence $\sigma_1,\dots,\sigma_t\in\Delta\setminus\Delta'$ such that $\phi(\sigma_1)\succ\sigma_2,\phi(\sigma_2)\succ\sigma_3,\dots,\phi(\sigma_t)\succ\sigma_1$. Without loss of generality we can assume that the collapse $(\sigma_1,\phi(\sigma_1))$ precedes collapses $(\sigma_i,\phi(\sigma_i))$ for $2\leq i\leq t$. Then $\phi(\sigma_t)\succ\sigma_1$ yields a~contradiction. \vskip4pt \noindent \underline{b) $\Rightarrow$ a)}: The proof is again very easy, various versions of it were given in \cite[Corollary 3.5]{Fo98}, \cite[Proposition 3.7]{BBLSW}, and \cite[Theorem 3.2]{Ko00a}. $\square$ \begin{prop}\label{pr2} Let $\Delta$ be a~triangulated space with some linear order $\ll$ on its set of vertices $V(\Delta)$. Let $V'\subseteq V(\Delta)$ and $\Delta'$ be the subcomplex of $\Delta$ induced by~$V'$. Assume we have a~partition $V(\Delta)=\cup_{z\in V'}V_z$ such that $z=\min_{\ll}V_z$. For $\sigma\in\Delta\setminus\Delta'$, let $\chi(\sigma)\in V'$ be defined by $x_{\xi(\sigma)}\in V_{\chi(\sigma)}$. Finally assume that the following condition is satisfied: \begin{quote} Condition $\aleph$. If $\sigma\in\Delta\setminus\Delta'$, $\sigma=(x_1,\dots,x_t)$, is such that either $\xi(\sigma)=1$ or $x_{\xi(\sigma)-1}\neq\chi(\sigma)$, then there exists a~unique simplex $\sigma'=(x_1,\dots,x_{\xi(\sigma)-1},\chi(\sigma),x_{\xi(\sigma)},\dots,x_t)$ such that $\sigma'\setminus\chi(\sigma)=\sigma$. \end{quote} Then there is a~sequence of collapses leading from $\Delta$ to $\Delta'$. \end{prop} \nin{\bf Proof. } Let $U$ denote the set of all $\sigma\in\Delta\setminus\Delta'$, $\sigma=(x_1,\dots,x_t)$, such that $x_{\xi(\sigma)-1}\neq\chi(\sigma)$ or $\xi(\sigma)=1$. The matching $\phi$ is defined by Condition $\aleph$: for $\sigma\in U$ we set $\phi(\sigma)=\sigma'$. By Proposition~\ref{pr1} it is enough to check that this matching is acyclic. For $\sigma\in U$ we have $\xi(\phi(\sigma))=\xi(\sigma)+1$. Moreover, if $\phi(\sigma)\succ\sigma'$ and $\sigma'\in U$, then $\sigma'=\phi(\sigma)\setminus x_{\xi(\sigma)}$, hence $\xi(\sigma')\geq\xi(\phi(\sigma))$. Therefore, if there is a~sequence $\sigma_1,\dots,\sigma_t\in\Delta\setminus\Delta'$ such that $\phi(\sigma_1)\succ\sigma_2,\phi(\sigma_2) \succ\sigma_3,\dots,\phi(\sigma_t)\succ\sigma_1$, then we have $\xi(\sigma_1)< \xi(\phi(\sigma_1))\leq\xi(\sigma_2)<\xi(\phi(\sigma_2))\leq\dots< \xi(\phi(\sigma_t))\leq \xi(\sigma_1)$ which yields a~contradiction. $\square$ \subsection{Proof of Theorem~\ref{main2}} \noindent We define a~$(\Lam,\mu)$-forest of rank $r$ to be a~marked forest $(T,\eta)$ of rank $r$ such that $\lambda_{r+1}(T,\eta)=\mu$ and $\lambda_i(T,\eta)\in\Lambda$, for $0\leq i\leq r$. It follows from the discussion in Section~\ref{s3} and in particular from Theorem~\ref{main1} that the $r$-simplices of $X_{\Lambda,\mu}$ can be indexed by $(\Lam,\mu)$-forests of rank $r$ so that the boundary relation of $X_{\Lambda,\mu}$ corresponds to level deletion in the marked forests. We call number partitions of the form $(k^m,1^{n-km})$, for some~$m$, {\it special}. Let $K$ be the subcomplex of $X_{\Lambda,\mu}$ induced by the set of all special partitions. We adopt the notations $\xi(\sigma)$ and $\chi(\sigma)$ used in Proposition~\ref{pr2} to the context of $X_{\Lambda,\mu}$ and its subcomplex $K$. The linear order on $V(X_{\Lambda,\mu})$ can be taken to be any linear extension of the partial order on $V(X_{\Lambda,\mu})$ given by the negative of the length function. The partition of $V(X_{\Lambda,\mu})$ is given by: for $v\in V(X_{\Lambda,\mu})\setminus V(K)$, $z\in V(K)$, we have $v\in V_z$ iff $z=\gamma_k(v)$. Let us show that the subcomplex $K$ satisfies Condition $\aleph$. Let $\sigma\inX_{\Lambda,\mu}\setminus K$, $\sigma=(x_1,\dots,x_t)$, and assume $\xi(\sigma)=1$ or $\chi(\sigma) \neq x_{\xi(\sigma)-1}$. In the language of marked forests this can be reformulated as: $\sigma$ is a~$(\Lam,\mu)$-forest $(T,\eta)$ of rank $t$ such that $\lambda_{\xi(\sigma)-1} (T,\eta)$ is not special and if $\xi(\sigma)>1$ then $\lambda_0(T,\eta),\dots,$ $\lambda_{\xi(\sigma)-2}(T,\eta)$ are special, and $\lambda_{\xi(\sigma)-2}(T,\eta)\neq \gamma_k(\lambda_{\xi(\sigma)-1}(T,\eta))$. In other words, on all vertices of height 0 to $\xi(\sigma)-2$ the function~$\eta$ takes only values 1 or $k$ and for the vertices of height $\xi(\sigma)-1$ it is no longer true. Moreover, there exists a~vertex of height $\xi(\sigma)-1$ which has at least~$k$ children on which $\eta$ is equal to~1. It is now clear that there exists a~{\it unique} $(\Lam,\mu)$-forest $(\widetilde T,\tilde\eta)$ of rank $r+1$ such that \begin{itemize} \item $(\widetilde T^{\xi(\sigma)-1},\tilde\eta^{\xi(\sigma)-1})=(T,\eta)$; \item $\lambda_{\xi(\sigma)-1}(\widetilde T,\tilde\eta)=\gamma_k(\lambda_{\xi(\sigma)}(\widetilde T,\tilde\eta))$, i.e.,~$\tilde\eta$ takes only values~1 or~$k$ on the vertices of height $\xi(\sigma)-1$ and each vertex of height $\xi(\sigma)$ in $(\widetilde T,\tilde\eta)$ has no more than~$k-1$ children labeled~1. \end{itemize} To construct $(\widetilde T,\tilde\eta)$, extend $(T,\eta)$ by splitting each vertex of height $\xi(\sigma)-1$ into vertices marked~$k$ and~1 so that the number of $k$'s is maximized. The uniqueness of $(\widetilde T,\tilde\eta)$ follows from the definition of the notion of isomorphism of marked forests. We have precisely checked Condition $\aleph$ and therefore by Proposition~\ref{pr1} we conclude that there is a~sequence of collapses leading from $X_{\Lambda,\mu}$ to $K$. It remains to see that $K$ is collapsible. If $\mu=(n)$, then $K$ is a~simplex, so we can assume that $\mu\in\Lambda$. If $\mu=\gamma_k(\mu)$, then $K$ is again a simplex. Otherwise it is easy to see that there is a~unique vertex in $X_{\Lambda,\mu}$ labeled $\gamma_k(\mu)$ and that $K$ is a~cone with an~apex in this vertex. $\square$ \section{On a conjecture of Sundaram and Welker} \label{s5} \subsection{A counterexample to the general conjecture} \noindent The original formulation of Conjecture~\ref{c3} in \cite{SuW97} was \begin{conj}\label{c1}\cite[Conjectures 4.12 and 4.13]{SuW97}. Let $\lambda$ and $\mu$ be different set par\-ti\-tions, such that $\lambda\vdash\mu$. Let $\pi\in\Pi_\lambda$ be a~par\-tition of type $\mu$. Then the multiplicity of the trivial representation in the $\mbox{Stab}\,(\pi)$-module $\widetilde H_*(\Delta(\Pi_\lambda(\hat 0,\pi)),{\Bbb Q})$ is~$0$. \end{conj} In our terms Conjecture \ref{c1} is equivalent to \begin{conj}\label{c2} For $\lambda\vdash\mu$, $\lambda\neq\mu$, the space $X_{\lambda,\mu}$ is ${\Bbb Q}$-acyclic. \end{conj} We shall give an~example when $X_{\lambda,\mu}$ is not even connected. It turns out that if one is only interested in counting the number of connected components of $X_{\lambda,\mu}$, then there is a~simpler poset model which we now proceed to describe. \begin{df}\label{pm} Assume $\lambda\vdash\mu\vdash n$, $\lambda\neq\mu$. The $(\lam,\mu)$-forests of rank $0$ can be partially ordered as follows: $(T_1,\eta_1)\prec(T_2,\eta_2)$ if there exists a~$(\lam,\mu)$-forest $(T,\eta)$ of rank $1$ such that $(T_1,\eta_1)=(T^1,\eta^1)$ and $(T_2,\eta_2)=(T^0,\eta^0)$. We call the obtained poset $P_{\lambda,\mu}$. \end{df} In other words, elements of $P_{\lambda,\mu}$ are number partitions $\tau\neq\mu$ such that $\lambda\vdash\tau\vdash\mu$, together with a~bracketing which shows how to form $\mu$ out of $\tau$, the order of the brackets and of the terms within the brackets is neglected. For example $(1,1,1)(3,1)(2,2)$ and $(3)(2,1,1)(2,1,1)$ are two different elements of $P_{(2,1^9),(4^2,3)}$, while $(1,1,1)(2,2)(3,1)$ is equal to the first mentioned element. These bracketed partitions are ordered by refinement, preserving the bracket structure. \begin{prop}\label{pr} $X_{\lambda,\mu}$ and $\Delta(P_{\lambda,\mu})$ have the same number of connected components, i.e., $\beta_0(X_{\lambda,\mu})=\beta_0(\Delta(P_{\lambda,\mu}))$. \end{prop} \nin{\bf Proof. } We know that $\Delta(P_{\lambda,\mu})$ and $X_{\lambda,\mu}$ have the same set of vertices and that there is an~edge between two vertices $a$ and $b$ of $\Delta(P_{\lambda,\mu})$ iff $a\prec b$ or $b\prec a$, which is, by the Definition \ref{pm}, the case iff there is an~edge between the corresponding vertices of $X_{\lambda,\mu}$. This shows that $\Delta(P_{\lambda,\mu})$ and $X_{\lambda,\mu}$ have the same number of connected components. $\square$ \vskip4pt Note that 1-skeleta of $X_{\lambda,\mu}$ and $\Delta(P_{\lambda,\mu})$ need not be equal. $\Delta(P_{\lambda,\mu})$ can intuitively be thought of as a~simplicial complex obtained by forgetting the multiplicities of simplices in the triangulated space $X_{\lambda,\mu}$. \vskip4pt \noindent {\bf Counterexample.} For $n=23$, $\lambda=(7,6,4,3,2,1)$, $\mu=(10,8,5)$, $X_{\lambda,\mu}$ is disconnected. $P_{\lambda,\mu}$ is shown on the Figure~3.4. Clearly $\Delta(P_{\lambda,\mu})$ is not connected, hence, by the~Proposition~\ref{pr}, neither is $X_{\lambda,\mu}$, which disproves Conjecture~\ref{c1}. \vskip-5pt \hskip-100pt $$\epsffile{ex.eps}$$ \begin{center} \vskip-5pt Figure 3.4 \end{center} \begin{rem} In the counterexample above, one can actually verify that $X_{\lambda,\mu}=\Delta(P_{\lambda,\mu})$. However, we choose to use posets $P_{\lambda,\mu}$ for two reasons: \begin{enumerate} \item it is easier to produce series of counterexamples to Sundaram-Welker conjecture using $\Delta(P_{\lambda,\mu})$ rather than $X_{\lambda,\mu}$; \item we feel that posets $P_{\lambda,\mu}$ are of independent interest, since they are in a~certain sense the ``naive'' quotient of $\Pi_\lambda(\hat 0,\pi)$ by $\mbox{Stab}\,(\pi)$. \end{enumerate} \end{rem} We believe that, in general, connected components of $X_{\lambda,\mu}$ may be not acyclic. \subsection{Verification of the conjecture in a special case} \begin{df}\label{dgen} We say that a~number partition $\lambda=(\lambda_1,\dots,\lambda_t)$ is generic (also called free of resonances in~\cite{ShW98}, having no equal sub-sums in~\cite{Ko97}) if whenever $\sum_{i\in I}\lambda_i=\sum_{j\in J}\lambda_j$, for some $I,J\subseteq[t]$, we have $\{\lambda_i\}_{i\in I}=\{\lambda_j\}_{j\in J}$ as multisets. \end{df} For example $\lambda=(k^m)$ is generic. \begin{prop}\label{tgen} If $\lambda$ is generic, then the stratum $\Sigma_\lam^{S^2}$ is homeomorphic to the $2l(\lambda)$-dimensional sphere. \end{prop} \nin{\bf Proof. } For a~generic partition $\lambda=(\underbrace{\lambda_1,\dots,\lambda_1}_{k_1}, \underbrace{\lambda_2,\dots,\lambda_2}_{k_2},\dots, \underbrace{\lambda_t,\dots,\lambda_t}_{k_t})$ we have $\Sigma_\lam^{S^2}\cong(S^2)^{(k_1)}\wedge(S^2)^{(k_2)}\wedge\dots\wedge(S^2)^{(k_t)} \cong S^{2k_1+\dots+2k_t}=S^{2l(\lambda)}$. $\square$ \chapter{Complexes of Directed Trees and Their Quotients} \section{The objects of study and the main questions} To any directed graph one can associate an~abstract simplicial complex in the~following way. \begin{df}\label{def} Let $G$ be an~arbitrary directed graph. $\da(G)$ is the~simplicial complex $\da(G)$ constructed as follows: the~vertices of $\da(G)$ are given by the~edges of $G$ and faces are all directed forests which are subgraphs of $G$. \end{df} In~\cite{St97} R.\ Stanley asked the~following two questions. \vskip4pt {\bf Question 1.} {\it Let $G_n$ be the~complete directed graph on $n$ vertices, i.e.,~a~graph having exactly one edge in each direction between any pair of vertices, all together $n(n-1)$ edges. The~complex $\da(G_n)$ is obviously pure, but is it shellable? } \vskip4pt There has been a~recent upsurge of activity in studying the homotopy type of simplicial complexes constructed from monotone properties of graphs: the vertices of such a~complex are all possible edges of the graph and the simplices are given by graphs which satisfy given monotone property, see~\cite{BBLSW,BWe98}. The question above can be reformulated in this language: what is the homotopy type of the simplicial complex corresponding to the monotone property of a~graph being a~directed forest? \vskip4pt {\bf Question 2.} {\it Is the~complex $\da(G)$ shellable in general? } \vskip4pt In general, one may ask what are the~homology groups $H_*(\da(G))$ and whether they can be linked to the~combinatorial invariants of the~graph $G$ in a~simple way. In this chapter we answer affirmatively the Question~1 in Theorem~\ref{thm1} and negatively the Question~2 in Example~~\ref{exam5.2.2}, Section~\ref{s4.2}. Furthermore, we start the general investigation by computing the~homology groups of $\da(G)$ for the cases when $G$ is {\it essentially a~tree} (see Definition~\ref{df4.1}) and when $G$ is a~{\it double directed cycle}. \vskip4pt There is a~natural action of ${\cal S}_n$ on $\Delta(G_n)$ induced by the standard permutation action of ${\cal S}_n$ on $[n]$, thus one can form the~topological quotient $X_n=\Delta(G_n)/{\cal S}_n$; for example, the case $n=3$ is shown on the Figure~4.3. The quotient complexes of combinatorially defined spaces still tend to be rather complicated in general, however sometimes (as is the case for $\Delta(\Pi_n)/{\cal S}_n$) their homology groups are torsion free. \vskip4pt {\bf Question 3.} {\it Are the homology groups $H_i(X_n,{\Bbb Z})$ torsion free in general? } \vskip4pt In Section~\ref{s6} we show that the groups $H_*(X_n,{\Bbb Z})$ are, in general, not free, and also give a~formula for $\beta_{n-2}(X_n,{\Bbb Q})$. \section{First examples and properties} \label{s4.2} \subsection{Conventions} \noindent For brevity, we write $H_*(G)$ instead of $H_*(\da(G),{\Bbb Z})$. We will also use the following standard fact: if $\Delta$ is a~simplicial complex and $F_1,\dots,F_t$ is the set of maximal simplices such that $\Delta\setminus\cup_{i=1}^t F_i$ is contractible, then $\Delta\simeq \vee_{i=1}^t S^{\dim F_i}$. In this case, we call $\{F_1,\dots,F_t\}$ a~{\it generating set} for $\Delta$. Clearly, for a~fixed $\Delta$, there may be several generating sets, but the multiset $\{\dim F_1,\dots, \dim F_t\}$ is defined uniquely by~$\Delta$. Let $\Delta$ be a simplicial complex, $F$ be one of its maximal simplices and $\tilde F$ be a subsimplex of $F$, such that $\dim F=\dim\tilde F+1$ and $F$ is the only maximal simplex which contains $\tilde F$. Then, removing $F$ and $\tilde F$ from $\Delta$ is called an {\it elementary collapse}. Obviously, $\Delta\setminus\{\tilde F,F\}$ is a~strong deformation retract of $\Delta$. For further references on the topological concepts used here we refer the~reader to the Appendix~D, the textbook by J.\ Munkres, \cite{Mu84}, and the thorough survey article by A.\ Bj\"orner,~\cite{Bj95}. \subsection{First examples} \noindent Let us give a~couple of examples to illustrate how irregular $\Delta(G)$ can be. \begin{exam} {\rm A graph $G$ for which $\da(G)$ is not pure.} \end{exam} $$\begin{array}{c} \epsffile{ex1.eps}\\ \mbox{Figure 4.1} \end{array}$$ \begin{exam} \label{exam5.2.2} {\rm The complex $\da(G)$ does not have to be shellable.} Let $G=C_5$, a~double directed cycle with 5 vertices, see the subsection~4.3 for the definition. It is easy to see that $\Delta(C_5)$ is a~pure simplicial complex of dimension~3. On the other hand, by Proposition~\ref{cyc}, $\Delta(C_5)\simeq S^2\vee S^3\vee S^3$. This implies that $H_2(\Delta(C_5),{\Bbb Z})={\Bbb Z}$, in particular $\Delta(C_5)$ is not shellable. \end{exam} \subsection{Elementary properties} \noindent It is not difficult to derive the~simplest properties of our construction. For example $\Delta(G_1)*\Delta(G_2)=\Delta(G_1\uplus G_2)$. Also it is easy to characterize those graphs $G$ for which $\da(G)$ is pure of full dimension. Namely, for $x\in V(G)$, let $S(x)=\{y\,|\,(y\rightarrow x)\in E(G)\}$. Then we have \begin{prop} Assume $V(G)=n$. The~following are equivalent \begin{enumerate} \item[a)] $\Delta(G)$ has a~maximal simplex of dimension less than $n-1$; \item[b)] there exist two disjoint subtrees $T_1$ and $T_2$ of $G$ with roots $x_1$, resp.~$x_2$ such that $S(x_1)\subseteq V(T_1)$, $S(x_2)\subseteq V(T_2)$. In particular, the sets $S(x_1)$ and $S(x_2)$ are disjoint. \end{enumerate} In other words, $\da(G)$ is pure of dimension $n-1$ iff b) is not true. \end{prop} \nin{\bf Proof. } \vskip4pt \noindent \underline{$a)\Rightarrow b)$}: Let $F$ be a~maximal simplex with fewer than $n-1$ edges. Then $F$ defines a~forest consisting of two or more trees. Let $T_1$ and $T_2$ be two different maximal subtrees of this forest with roots $x_1$, resp. $x_2$. If there exists $y\in S(x_1)\setminus V(T_1)$, then $(y\rightarrow x_1)\cup F$ is a~simplex of $\Delta(G)$. This contradicts to the fact that $F$ is a~maximal simplex. Thus $S(x_1)\subseteq V(T_1)$ and, symmetrically, $S(x_2)\subseteq V(T_2)$. \vskip4pt \noindent \underline{$b)\Rightarrow a)$}: Let $F$ be an arbitrary maximal simplex of $\Delta(G)$ such that $E(T_1)\cup E(T_2)\subseteq F$. Since $S(x_1)\subseteq V(T_1)$ and $S(x_2)\subseteq V(T_2)$, there are no edges in $F$ which point to either $x_1$ or $x_2$. This proves that $|F|\leq n-2$. $\square$ \section{Graphs with complete source} \label{sec3} \subsection{Shellability of complexes of directed trees} \noindent In the~next theorem we answer the~first question of R.\ Stanley. \begin{thm} \label{thm1} If the~graph $G$ has a~complete source, then the~complex $\da(G)$ is shellable. In particular $\da(G_n)$ is shellable for any $n\geq 1$. More precisely, $\da(G_n)$ is homotopy equivalent to a~wedge of $(n-1)^{n-1}$ spheres of dimension $(n-2)$ and the~representatives of the~cohomology classes are labeled by the~spanning directed trees having vertex $1$ as a~leaf. \end{thm} \nin{\bf Proof. } Let $G$ be a~directed graph on $n$ vertices labeled by the~set $[n]$ and $1$ be the~complete source of $G$. Clearly every partial subforest of $G$ can be completed to a~tree: just add edges pointing from the~complete source to the~roots of the~trees of this subforest (except for the~one which contains the~complete source). Thus $\da(G)$ is pure. We shall now describe a~labeling of the~maximal faces of $G$. For an~edge $(x\rightarrow y)$ we define $\lambda(x\rightarrow y)=x$. For a~graph $H$ we define $\lambda(H)=(\lambda_1,\dots,\lambda_m)$, where $m=|E(H)|$, $\{\lambda_1,\dots,\lambda_m\}=\{\lambda(e)\,|\,e\in E(H)\}$ (as multisets) and $\lambda_1\leq\dots\leq\lambda_m$. The~function $\lambda$ describes a~labeling of maximal faces of $\da(G)$ by sequences of $n-1$ numbers. Let us order these sequences (and hence the~maximal faces) in a~lexicographic order. We shall next verify that this ordering is a~shelling order. Let $A$ and $B$ be directed trees on $n$ vertices, such that $$\lambda(A)=(\alpha_1,\dots,\alpha_{n-1})\preceq(\beta_1,\dots,\beta_{n-1})=\lambda(B).$$ Let $C$ be a~graph defined by $V(C)=[n]$, $E(C)=E(B)\cap E(A)$. Clearly, $\lambda(C)$ is a~substring of $\lambda(A)$ and $\lambda(B)$. $C$ is a~forest, denote its trees by $T_1,\dots,T_s$ such that $1\in V(T_1)$ and let $r_i$ be the~root of $T_i$. It is clear that edges from $E(A)\setminus E(B)$ as well as from $E(B)\setminus E(A)$ have the~form $(a\rightarrow r_i)$ for some $a\in[n]$, $i\in [s]$. Choose $(x\rightarrow y)\in E(B)\setminus E(A)$ such that $x\neq 1$. It must exist, since otherwise we have $\alpha_i=\beta_i$, $i\in[n-1]$, and all of the~edges from $E(A)\setminus E(B)$ and $E(B)\setminus E(A)$ are $(1\rightarrow r_i)$, for $i=2,\dots,s$, which would imply $A=B$. Let $B'$ be defined by $E(B')=E(B)\setminus\{(x\rightarrow y)\}$. $B'$ consists of two trees $T'$ and $T''$ with roots $r'$ and $r''$. Assume that $1\in V(T')$. Define a~new tree $\tilde B$ by $$E(\tilde B)=E(B')\uplus\{(1\rightarrow r'')\}.$$ It is clear that $\lambda(\tilde B)=(\lambda(B)\setminus\{x\})\uplus\{1\}$ (as multisets), hence $\lambda(\tilde B)\preceq\lambda(B)$. Furthermore, by construction, $C$ is a~subgraph of $\tilde B$. Hence we have verified Con\-di\-ti\-on~(S), thus $\da(G)$ is shellable. It is now easy to choose representatives of cohomology classes of $\da(G)$. They are labeled by maximal faces $A$, such that $1\notin\lambda(A)$ (the representatives themselves are functions which evaluate to 1 on such $A$ and 0 on all other maximal faces). Furthermore, when $G=G_n$, it is easy to enumerate all such maximal faces. Denote this number by $f(n)$. The~condition $1\notin\lambda(A)$ simply means that 1 is a~leaf in $A$, so we can instead consider trees on $n-1$ vertices with some marked vertex. The~number of such trees is clearly $(n-1)g(n-1)$, where $g(n)$ is the~number of all rooted labeled trees on $n$ vertices. Since it is well known that $g(n)=n^{n-1}$ we conclude that $f(n)=(n-1)g(n-1)=(n-1)^{n-1}$. $\square$ \vskip4pt In the~last part of the~argument above we have have found a~new proof of the combinatorial formula $\tilde\chi(\da(G_n))=(n-1)^{n-1}$, cf.\ \cite{Pi96,St99}. \subsection{Algebraic consequences of the shelling} \noindent The~result of Theorem \ref{thm1} can be reformulated in the~algebraic language as follows. \begin{crl} Pick an~index set $\Gamma\subseteq[n]\times[n]$ such that $(1,i)\in\Gamma$ for $2\leq i\leq n$, $(i,i)\notin\Gamma$ for $1\leq i\leq n$. For an~arbitrary field~$k$, let $k[\Gamma]=k[x_{ij}]_{(i,j)\in\Gamma}$. Let $I$ be the~ideal of $k[\Gamma]$ generated by the~following monomials: \begin{itemize} \item $x_{ij} x_{kj}$, for $(i,j),(k,j)\in\Gamma$; \item $x_{i_1 i_2}x_{i_2 i_3}\dots x_{i_t i_1}$, for $t\geq 2$, $(i_t,i_1), (i_j,i_{j+1})\in\Gamma$, for $j\in [t-1]$. \end{itemize} Then the~ring $k[\Gamma]/I$ is Cohen-Macaulay (i.e.,~the~dimension of $k[\Gamma]/I$ is equal to its depth, see~\cite{BH93} for further details). \end{crl} \nin{\bf Proof. } The~ring $k[\Gamma]/I$ is the~Stanley-Reisner ring of $\Delta(G)$, where $G$ is a~directed graph with a~complete source such that $E(G)=\Gamma$. See~\cite{Re76,St96} for the definition and properties of Stanley-Reisner rings. $\square$ \vskip4pt \subsection{A few words on the related ${\cal S}_n$-representations} \noindent The~natural action of the~symmetric group $S_n$ on $V(G_n)$ induces an~action on $\Delta(G_n)$ in an~obvious way. This action determines a~linear representation of $S_n$ on $H_{n-2}(G_n)$. It would be interesting to understand the~structure of this representation better, but it seems to be hard. However, R.~Stanley was able to compute its character,~\cite{St97}. We study the topological quotient $\Delta(\Pi_n)/{\cal S}_n$ in Section~\ref{s6}. Consider instead a~slightly different, but better behaving action. Let us act with $S_{n-1}$ on $\Delta(G_n)$ by permuting the~vertices $\{2,\dots,n\}$. It follows from the~description of the~cohomology classes of $\Delta(G_n)$ that $S_{n-1}$ permutes them and thus the~representation of $S_{n-1}$ is a~permutation representation: $S_{n-1}$ permutes double-rooted trees on $n-1$ vertices. Again, the~structure of this representation, such as decomposition into irreducibles, is unclear. \section{Computations for other classes of graphs} \subsection{Graphs which are essentially trees} \begin{df} \label{df4.1} A~graph $G$ is called {\bf essentially a~tree} if it turns into an undirected tree when one replaces all directed edges/pairs of directed edges going in opposite direction by an~edge. \end{df} The~following 3 propositions will provide us with a~procedure to compute homology groups of $\da(G)$ when $G$ is essentially a~tree. \begin{prop}\label{t1} Let $G$ be a~directed graph and let $x\in V(G)$ with $S(x)=\{y\}$, for some $y\in V(G)$. If no edge of $G$ has $x$ as a~source, then $\Delta(G)$ is contractible. \end{prop} \nin{\bf Proof. } $\da(G)$ is a~cone with apex $(y\rightarrow x)$. $\square$ \begin{prop}\label{t2} Let $G$ be a~graph, $x\in V(G)$, such that $(x\rightarrow y),(y\rightarrow x)\in E(G)$ and there are no other edges where $x$ is a~source or a~sink. Then $\da(G)\simeq\mbox{susp}\,\, \Delta(\tilde G)$, where $\tilde G$ is defined by $V(\tilde G)=V(G)\setminus\{x\}$, $E(\tilde G)=E(G)\setminus(\{(x\leftarrow y)\}\uplus\{(y\leftarrow z)\,|$ $z\in V(G)\})$. \end{prop} \nin{\bf Proof. } Let $A={\mbox{st}\,}(x\rightarrow y)$, $B={\mbox{st}\,}(y\rightarrow x)$. Then $\da(G)=A\cup B$ since every forest not containing edge $(x\rightarrow y)$ can be extended with the~edge $(y\rightarrow x)$. $A\cap B$ contains those forests which can be extended with both $(x\rightarrow y)$ and $(y\rightarrow x)$. This means we have to delete all edges having $x$ or $y$ as a~sink, which gives $A\cap B=\Delta(\tilde G)$. Both $A$ and $B$ are contractible, hence, by \cite[Lemma 10.4(ii)]{Bj95}, $\da(G)\simeq\mbox{susp}\,\,\Delta(\tilde G)$. $\square$ \begin{prop}\label{t3} Let $G$ be a~graph and $x_1,\dots,x_k,y\in V(G)$ such that $(x_1\rightarrow y),\dots, (x_k\rightarrow y)\in E(G)$ and there are no further edges which have $x_i$ as a~source or a~sink for $i\in[k]$. Assume furthermore, that for some $z\in V(G)\setminus\{x_1,\dots,x_k,y\}$ one of the~following is true: \begin{enumerate} \item [a)] $(y\rightarrow z)\in E(G)$; \item [b)] $(z\rightarrow y)\in E(G)$; \item [c)] $(y\rightarrow z),(z\rightarrow y)\in E(G)$; \end{enumerate} and there are no other edges having $y$ as a~sink or a~source. Then, corresponding to these cases, we have \begin{enumerate} \item [a)] $\da(G)\simeq\mbox{susp}\,_k\Delta(\tilde G)$, where $V(\tilde G)=V(G)\setminus \{x_1,\dots,x_k\}$, $E(\tilde G)=E(G)\setminus\{(x_1\rightarrow y),\dots,(x_k\rightarrow y),(z\rightarrow y)\}$; \item [b)] $\da(G)\simeq\mbox{susp}\,_{k+1}\Delta(G')$, where $G'$ is the subgraph of $G$ induced by $V(G)\setminus\{x_1,\dots,x_k,y\}$; \item [c)] If $k=1$, then $\da(G)\simeq\mbox{susp}\,\,\Delta(G')$; if $k\geq 2$, then, for all $t\in\Bbb Z$, we have $H_t(G)=H_{t-1}(G')\oplus(H_{t-1}(\tilde G))^{\oplus(k-1)}$. \end{enumerate} \end{prop} \nin{\bf Proof. } Set $A_i={\mbox{st}\,}(x_i\rightarrow y)$ for $i\in[k]$. Clearly, all $A_i$ are contractible and the~intersection of two or more of them is equal to $\Delta(\tilde G)$. Furthermore, in case a) we have $\da(G)=\cup_{i=1}^k A_i$ and hence the~conclusion of a) follows by~\cite[Lemma 10.4(ii)]{Bj95}. Assume now that b) or c) holds. In both cases $(z\rightarrow y)\in E(G)$. Set $B={\mbox{st}\,}(z\rightarrow y)$, then $\da(G)=(\cup_{i=1}^k A_i)\cup B$. In the~case b) the~intersection of any two or more of the~complexes $A_1,\dots,A_k,B$ is equal to $\Delta(G')$, hence the~conclusion of b) again follows by~\cite[Lemma 10.4(ii)]{Bj95}. Let us show c) by induction on $k$. Assume $k=1$, then $\Delta(G)=A_1\cup B$. Both $A_1$ and $B$ are contractible, hence by~\cite[Lemma 10.4(ii)]{Bj95} we get $\Delta(G)\simeq\mbox{susp}\,\,(A_1\cap B)=\mbox{susp}\,\,\Delta(G')$. Assume now that $k\geq 2$. Let $A=(\cup_{i=1}^{k-1}A_i)\cup B$ and $A'=A_{k-1}\cup A_k$. Then $A\cup A'=\Delta(G)$ and $A\cap A'=A_{k-1}$, which is a~cone. Thus $$H_t(G)=H_t(A)\oplus H_t(A')=H_{t-1}(G')\oplus (H_{t-1}(\tilde G))^{\oplus(k-2)}\oplus H_{t-1}(\tilde G),$$ where the first equality follows by a~Mayer-Vietoris argument and the second equality follows from the induction assumption. $\square$ \vskip4pt So, given any graph $G$, which is essentially a~tree, we have recursive procedure to compute homology groups $H_*(G)$. If some class of trees which behaves well under recursion is specified, then closed formulae can be derived if so desired. Observe, for example, that if $G$ is a~double directed tree with two leaves or more attached to the~same vertex, then $\da(G)$ is contractible (just apply Proposition \ref{t2} and then Proposition \ref{t1}). \subsection{Double directed strings} \noindent Another interesting specific example is a~double directed string on $n+1$ vertices $L_n$, which is defined by $V(L_n)=[n+1]$, $E(L_n)=\{(i\rightarrow i+1), (i+1\rightarrow i)\,|\,i\in[n]\}$. Let $\tilde L_n$ be the directed graph defined by $V(\tilde L_n)=[n+2]$, $E(\tilde L_n)=E(L_n)\uplus\{(n+1\leftarrow n+2)\}$. The complexes $\Delta(L_n)$ and $\Delta(\tilde L_n)$ have an~alternative description. \begin{df} Complex ${\cal L}_n$ has $n$ vertices indexed by the~set $[n]$ and $F\in 2^{[n]}$ is a~face of ${\cal L}_n$ iff it does not contain $\{i,i+1\}$ for $i\in [n-1]$. \end{df} It is easy to see that \begin{equation} \label{a} {\cal L}_{2n}\simeq\Delta(L_n),\quad{\cal L}_{2n+1}\simeq\Delta(\tilde L_n). \end{equation} \newpage \begin{prop} $\,$ \begin{enumerate} \item[(1)] ${\cal L}_{n+3}\simeq\mbox{susp}\,\,{\cal L}_n$. \item[(2)] The generating simplices for ${\cal L}_n$ are $(2,5,\dots,3k+2)$, if $n=3k+2$, and $(2,5,\dots,3k-1)$, if $n=3k$. \item[(3)] ${\cal L}_1$ is contractible, ${\cal L}_2=S^0$, and ${\cal L}_3\simeq S^0$. Hence \begin{equation} \label{b} {\cal L}_n\simeq \begin{cases} S^{k-1}, &\mbox{ if } n=3k;\\ \mbox{a point}, &\mbox{ if } n=3k+1;\\ S^k, &\mbox{ if } n=3k+2. \end{cases} \end{equation} \end{enumerate} \end{prop} \nin{\bf Proof. } Let $C={\cal L}_{n+3}\setminus\{3\}$. Since every maximal simplex of $C$ contains exactly one of the vertices of 1 and 2, we have $C\simeq\mbox{susp}\,\,({\cal L}_{n+3}\setminus\{1,2,3\}) \simeq\mbox{susp}\,\,({\cal L}_n)$. Let us show that $C$ is a~deformation retract of ${\cal L}_{n+3}$. Order the simplices of ${\cal L}_{n+3}$ which have vertex~3, but do not have vertex~1 in any order $S_1,\dots,S_k$ respecting inclusions, i.e., if $S_j$ is a~subcomplex of $S_i$, then $i<j$. Remove pairs of simplices $(S_i,S_i\uplus \{1\})$ in the increasing order of $i$. These removals are elementary collapses, they correspond to deformation retracts. Since ${\cal L}_{n+3}=C\uplus\{S_i\,|\,i\in [k]\}\uplus\{S_i\uplus\{1\}\,|\,i\in [k]\}$, we conclude that $C$ is a~deformation retract of ${\cal L}_{n+3}$. This proves (1). (2) follows from (1) and the fact that if $\Delta$ is a~simplicial complex which is homotopy equivalent to a~sphere and $F$ is its generating simplex, then $F\cup\{a\}$ is a~generating simplex of $\Delta*\{a,b\}=\mbox{susp}\,\,\Delta$. To see this, it is enough to show that $\Delta*\{a,b\}\setminus\{F\cup\{a\}\}$ is contractible. Indeed, removing $F$ and $F\cup\{b\}$ from $\Delta*\{a,b\}\setminus\{F\cup\{a\}\}$ is an~elementary collapse, and $\Delta*\{a,b\}\setminus\{F,F\cup\{a\},F\cup\{b\}\}=(\Delta\setminus F)*\{a,b\}$, which is contractible. The verification of (3) is left to the reader. $\square$ \vskip4pt Using \eqref{a} and \eqref{b} we immediately derive that $$\Delta(L_n)\simeq\begin{cases} S^{2k-1}, & \mbox{ if } n=3k;\\ S^{2k}, & \mbox{ if } n=3k+1;\\ \mbox{a point}, & \mbox{ if } n=3k+2. \end{cases}$$ \subsection{Cycles} \label{sect5} \noindent Let $n\geq 3$ and denote by $C_n$ a~double directed cycle, i.e., a~directed graph defined by $V(C_n)={\Bbb Z}_n$, $E(C_n)=\{(i\rightarrow i+1),(i+1\rightarrow i)\,|\, i\in{\Bbb Z}_n\}$. In this section we determine the homotopy type of $\Delta(C_n)$. Again we would like to formulate our complexes in a~slightly different language. \begin{df} \label{df2} Complex ${\cal C}_n$ has $n$ vertices indexed by the~set ${\Bbb Z}_n$ and $F\in 2^{{\Bbb Z}_n}$ is a~face of ${\cal C}_n$ iff it does not contain $\{i,i+1\}$ for $i\in{\Bbb Z}_n$. \end{df} Similar to before, we have a~relation $\Delta(C_n)=\tilde{\cal C}_{2n}$, where $\tilde{\cal C}_{2n}$ is obtained from ${\cal C}_{2n}$ by deleting the simplices $(1,3,\dots,2n-1)$ and $(2,4,\dots,2n)$. \begin{prop} \label{cyc} The~homotopy type of ${\cal C}_n$ is given by \begin{equation}\label{hccn} {\cal C}_n\simeq\begin{cases} S^{k-1}\vee S^{k-1}, & \mbox{ if } n=3k;\\ S^{k-1}, & \mbox{ if } n=3k+1;\\ S^k, & \mbox{ if } n=3k+2. \end{cases} \end{equation} Therefore \begin{equation}\label{hcn} \Delta(C_n)\simeq\begin{cases} S^{2k-1}\vee S^{2k-1}\vee S^{3k-2}\vee S^{3k-2}, & \mbox{ if } n=3k;\\ S^{2k}\vee S^{3k-1}\vee S^{3k-1}, & \mbox{ if } n=3k+1;\\ S^{2k}\vee S^{3k}\vee S^{3k}, & \mbox{ if } n=3k+2. \end{cases} \end{equation} \end{prop} \nin{\bf Proof. } Let $A={\mbox{st}\,}_{{\cal C}_n}(1)$ and $B={\cal C}_n\setminus\{1\}$. Then $A\cup B={\cal C}_n$, $A\cap B={\cal C}_n\setminus\{1,2,n\}={\cal L}_{n-3}$, $A$ is contractible and $B={\cal L}_{n-1}$. \vskip3pt \noindent If $n=3k+2$, then $B$ is contractible, hence ${\cal C}_n\simeq\mbox{susp}\,\,(A\cap B)= \mbox{susp}\,\,({\cal L}_{n-3})\simeq S^k$ and $(1,4,\dots,n-1)$ is the generating simplex. \vskip3pt \noindent If $n=3k+1$, then $A\cap B$ is contractible, hence ${\cal C}_n\simeq B={\cal L}_{n-1}\simeq S^{k-1}$. \vskip3pt \noindent If $n=3k$, let $F=(3,6,\dots,n)$ be a~generating simplex of $B$. Since $F\in B\setminus(A\cap B)$, one can shrink $B\setminus F$ to a~point inside ${\cal C}_n$ and obtain $${\cal C}_n\simeq B\vee\mbox{susp}\,\,(A\cap B)\simeq{\cal L}_{3k-1}\vee{\cal L}_{3k}\simeq S^{k-1}\vee S^{k-1}.$$ Furthermore the generating simplices are $(3,6,\dots,n)$ and $(1,4,\dots,n-2)$. This shows \eqref{hccn}. Let us now see \eqref{hcn}. Let us single out the following simplices of ${\cal C}_{2n}$: $F_1=(1,3,\dots,2n-1)$, $F_2=(2,4,\dots,2n)$, $\tilde F_1=(1,3,\dots, 2n-3)$, $\tilde F_2=(2,4,\dots,2n-2)$. Recall that $\Delta(C_n)={\cal C}_{2n}\setminus\{F_1,F_2\}$. $F_1$, resp. $F_2$, is the only maximal simplex which contains $\tilde F_1$, resp. $\tilde F_2$, hence to remove $F_1$, $F_2$, $\tilde F_1$, and $\tilde F_2$, from ${\cal C}_{2n}$ means to perform two elementary collapses. Let $\hat{\cal C}_{2n}={\cal C}_{2n}\setminus\{F_1,F_2, \tilde F_1,\tilde F_2\}$. Let us show that the boundaries of $\tilde F_1$ and $\tilde F_2$ can be shrunk to a point within $\hat{\cal C}_{2n}$. Simplices $F_1$ and $\tilde F_1$ lie in the subcomplex ${\cal C}_{2n}\setminus\{2n\}\simeq {\cal L}_{2n-1}$. We can choose $(2,5,\dots,2n-2)$, if $2n=3k+1$, and $(2,5,\dots, 2n-1)$, if $2n=3k$, as a~generating simplex of $({\cal C}_{2n}\setminus\{2n\})\setminus\{F_1, \tilde F_1\}$. $\tilde F_1$ does not contain this generating simplex, hence its boundary can be shrunk to a point within $({\cal C}_{2n}\setminus\{2n\})\setminus\{F_1,\tilde F_1\}$, and hence within $\hat{\cal C}_{2n}$. Analogously the boundary of $\tilde F_2$ can be shrunk to a point within~$\hat{\cal C}_{2n}$. Thus we obtain $\Delta(C_n)\simeq{\cal C}_{2n}\setminus\{F_1,F_2\}=\hat{\cal C}_{2n}\cup \{\tilde F_1,\tilde F_2\}\simeq\hat{\cal C}_{2n}\vee S^{n-2}\vee S^{n-2}\simeq{\cal C}_{2n} \vee S^{n-2}\vee S^{n-2}$. $\square$ \section{${\cal S}_n$-quotients of complexes of directed forests} \label{s6} \subsection{A~combinatorial description for the cell structure of $X_n$} \noindent As mentioned in the introduction, let $X_n$ be the topological quotient $\Delta(G_n)/{\cal S}_n$. Clearly, the action of~${\cal S}_n$ on $\Delta(G_n)$ is not free. What is worse, the elements of~${\cal S}_n$ may fix the simplices of $\Delta(G_n)$ without fixing them pointwise: for example for $n=3$ the permutation $(23)$ ``flips'' the 1-simplex given by the directed tree $2\longleftarrow 1 \longrightarrow 3$. Therefore, one does not have a~bijection between the orbits of simplices of $\Delta(G_n)$ and simplices of $X_n$. To rectify the situation, let us consider the~barycentric subdivision $B_n=\mbox{Bsd}\,(\Delta(G_n))$. We have a~simplicial ${\cal S}_n$-action on $B_n$ induced from the ${\cal S}_n$-action on $\Delta(G_n)$ and, clearly, $B_n/{\cal S}_n$ is homeomorphic to~$X_n$. Furthermore, if an~element of ${\cal S}_n$ fixes a~simplex of $B_n$ then it fixes it pointwise. In this situation, it is well-known, e.g.,~see~\cite{Br72}, that the quotient projection $B_n\rightarrow X_n$ induces a~simplicial structure on $X_n$, in which simplices of $X_n$ correspond to ${\cal S}_n$-orbits of the simplices of $B_n$ with appropriate boundary relation. Let us now give a~combinatorial description of the ${\cal S}_n$-orbits of the simplices of $B_n$. Let $\sigma$ be a~simplex of $B_n$, then $\sigma$ is a~chain $(T_1,T_2,\dots,T_{\dim(\sigma)+1})$ of forests on $n$ labeled vertices, such that $T_i$ is a~subgraph of $T_{i+1}$, for $i=1,\dots, \dim(\sigma)$. One can view this data in a~slightly different way: it is a~forest with $\dim(\sigma)+1$ integer labels on edges (labels on different edges may coincide). Indeed, given a~chain of forests as above, take the forest $T_{\dim(\sigma)+1}$ and put label 1 an~all edges of the forest~$T_1$, label 2 on all edges of $T_2$, which are not labeled yet, etc. Vice versa, given a~forest $T$ with a~labeling, let $T_1$ be the forest consisting of all edges of $T$ with the smallest label, let $T_2$ be the forest consisting of all edges of $T$ with one of the two smallest labels, etc. To make the described correspondence a~bijection, one should identify all labeled forests on which labelings produce the same order on edges. \vskip4pt Formally: {\it the $p$-simplices of $B_n$ are in bijection with the set of all pairs $(T,\phi^T)$, where $T$ is a~directed forest on $n$~labeled vertices and $\phi^T:E(T)\rightarrow{\Bbb Z}$, such that $|\mbox{Im}\,\phi^T|=p+1$, modulo the following equivalence relation: $(T_1,\phi^{T_1})\sim (T_2,\phi^{T_2})$ if $T_1=T_2$ and there exists an~order-preserving injection $\psi:{\Bbb Z}\rightarrow{\Bbb Z}$, such that $\phi^{T_1}\circ\psi=\phi^{T_2}$.} \vskip4pt The boundary operator can be described as follows: for a~$p$-simplex $(T,\phi^T)$, $p\geq 1$, we have $$\partial(T,\phi^T)=\sum_{i=1}^{p+1}(-1)^{p+i+1} (T_i,\phi^{T_i}),$$ where, for $i=1,\dots,p$, we have $T_i=T$ and $\phi^{T_i}$ takes the same values as $\phi^T$ except for the edges on which $\phi^T$ takes $i$th and $(i+1)$st largest values (say $a$ and $b$), on these edges $\phi^{T_i}$ takes value $a$. Furthermore, $T_{p+1}$ is obtained from $T$ be removing the edges with the highest value of $\phi^T$, $\phi^{T_{p+1}}$ is the restriction of $\phi^T$. Of course, this description of the boundary map is just a~rephrasing of the deletion of the $i$th forest from the chain of forests in the original description. However, we will find it more convenient to work with the labeled forests rather than the chains of forests. The orbits of the action of ${\cal S}_n$ can be obtained by forgetting the numbering of the vertices. Thus, using the fact that simplices of~$X_n$ and ${\cal S}_n$-orbits of simplices of $B_n$ are the same thing, we get the following description. \vskip4pt {\it The $p$-simplices of~$X_n$ are in bijection with pairs $(T,\phi^T)$, where $T$ is a~directed forest on $n$~unlabeled vertices and $\phi^T$ is an~edge labeling of $T$ with $p+1$ labels, modulo a~certain equivalence relation. This equivalence relation and the boundary operator are exactly as in the description of simplices of $B_n$.} $$\epsffile{fig2.eps}$$ $$\mbox{Figure 4.2}$$ On Figure~4.2 we show the case $n=3$: on the left hand side we have $\Delta(G_3)$, on the right hand side is $X_3=\Delta(G_3)/{\cal S}_3$. The labeled forests next to the edges indicate the bijection described above, labeling on the forests corresponding to the vertices in $X_3$ is omitted. ${\cal S}_3$ acts on $\Delta(G_3)$ as follows: 3-cycles act as rotations around the line which goes through the middles of the triangles, each transposition acts as a~central symmetry on one of the quadrangles, and as a~``flip'' on the edge which is parallel to that quadrangle. \subsection{Filtration and description of the $E^1$ tableau} \noindent There is a~natural filtration on the chain complex associated to the simplicial structure on $X_n$ described above. Let $F_i$ be the union of all simplices $(T,\phi^T)$ where $T$ has at most $i$ edges. Clearly, $\emptyset=F_0\subset F_1\subset\dots\subset F_{n-1}=X_n$. Recall that $E_{p,k}^1=H_p (F_k,F_{k-1})$, here we use the indexing from the Appendix~D. In other words, the homology is computed with ``truncated'' boundary operator: the last term, where some edges are deleted from the forest, is omitted. Clearly, \begin{equation}\label{eq6.1} E_{p,k}^1=\bigoplus_{T}H_p(E_T), \end{equation} where the sum is over all forests with $k$ edges and $E_T$ is a~chain complex generated by the simplices $(T,\phi^T)$, for various labelings $\phi^T$, with the truncated boundary operator as above. \vskip4pt Let us now describe a~simplicial complex whose reduced homology groups, after a~shift in the index by 1, are equal to the nonreduced homology groups of $E_T$. The arrangement of $k(k-1)/2$ hyperplanes $x_i=x_j$ in ${\Bbb R}^k$ cuts the space $S^{k-1}\cap H$ into simplices, where $H$ is the~hyperplane given by the equation $x_1+x_2+\dots+x_k=0$. Denote this simplicial complex $A_k$. The permutation action of ${\cal S}_k$ on $[k]$ induces an~${\cal S}_k$-action on $A_k$. It is easy to see that if an~element of ${\cal S}_k$ fixes a~simplex of $A_k$, then it fixes it pointwise. Hence, for any subgroup $\Gamma\subseteq{\cal S}_k$, the $\Gamma$-orbits of the simplices of $A_k$ are in a~natural bijection with the simplices of $A_k/\Gamma$. \vskip4pt Let $T$ be an~arbitrary forest with $n$ vertices and $k$ edges. Assume that vertices, resp.~edges, are labeled with numbers $1,\dots,n$, resp.~$1,\dots,k$. ${\cal S}_n$ acts on $[n]$ by permutation, let $\mbox{Stab}\,(T)$ be stabilizer of $T$ under this action, that is the maximal subgroup of ${\cal S}_n$ which fixes $T$. Then $\mbox{Stab}\,(T)$ acts on $E(T)$, i.e.,~we have a~homomorphism $\chi:\mbox{Stab}\,(T)\rightarrow{\cal S}_k$. Let ${\cal S}(T)=\mbox{Im}\,\chi$. Clearly ${\cal S}(T)$ does not depend on the choice of the labeling of vertices. However, relabeling the edges changes ${\cal S}(T)$ to a~conjugate subgroup. Therefore, for a~forest $T$ without labeling on vertices and edges, ${\cal S}(T)$ can be defined, but only up to a~conjugation. \begin{prop}\label{pr6.1} The chain complex of $A_k/{\cal S}(T)$ and $E_T$ (with a~shift by 1 in the indexing) are isomorphic. In particular, ${\widetilde H}_p(A_k/{\cal S}(T))=H_{p+1}(E_T)$. \end{prop} \nin{\bf Proof. } Label the $k$ edges of $T$ with numbers $1,\dots,k$. As mentioned above, the $p$-simplices of $E_T$ are in bijection with labelings of the edges of~$T$ with numbers $1,\dots,p+1$ (using each number at least once). Taking in account the chosen labeling of the edges, this is the same as to divide the set $[k]$ into an~ordered tuple of $p+1$ non-empty sets, modulo the symmetries of $[k]$ induced by the symmetries of~$T$. Clearly, these symmetries of $[k]$ are precisely the elements of~${\cal S}(T)$. The $(p-1)$-simplices of $A_k$ are in bijection with dividing $[k]$ into an ordered tuple of $p+1$ non-empty sets: by the values of the coordinates. Therefore we conclude that the $p$-simplices of $E_T$ are in a~natural bijection with the $(p-1)$-simplices of $A_k/{\cal S}(T)$. Here the unique 0-simplex of $E_T$, $(T,{\bf 1})$, ($\bf 1$ is the constant function taking value~1), corresponds in $A_k/{\cal S}(T)$ to the empty set, which is a~$(-1)$-simplex. One verifies immediately that the boundary operators of $E_T$ and $A_k/{\cal S}(T)$ commute with the described bijection. Therefore $E_T$ and $A_k/{\cal S}(T)$ are isomorphic as chain complexes (after a~shift in the indexing). In particular, ${\widetilde H}_p(A_k/{\cal S}(T))=H_{p+1}(E_T)$. $\square$ \subsection{${\Bbb Q}$ coefficients} \noindent Proposition~\ref{pr6.1} allows us to give a~description of $E_{*,*}^1$-entries in the case when the homology groups are computed with rational coefficients. Indeed, it is well known that, when a~finite group $\Gamma$ acts on a~finite simplicial complex $X$, one has ${\widetilde H}_i(X/\Gamma,{\Bbb Q})={\widetilde H}_i^\Gamma(X,{\Bbb Q})$, where ${\widetilde H}_i^\Gamma(X,{\Bbb Q})$ is the maximal vector subspace of ${\widetilde H}_i(X,{\Bbb Q})$ on which $\Gamma$ acts trivially (more generally ${\Bbb Q}$ can be replaced with a~field whose characteristic does not divide $|\Gamma|$). Since $A_k$ is homeomorphic to $S^{k-2}$ we have ${\widetilde H}_{k-2}(A_k,{\Bbb Q})={\Bbb Q}$ and ${\widetilde H}_i(A_k,{\Bbb Q})=0$ for $i\neq k-2$. It is easy to compute ${\widetilde H}_{k-2}^{{\cal S}(T)}(A_k,{\Bbb Q})$. In fact, for $\pi\in{\cal S}_k$, $\alpha\in{\widetilde H}_{k-2}(A_k,{\Bbb Q})$, one has $\pi(\alpha)=(-1)^{\mbox{sgn}\,\pi}\alpha$, where $\mbox{sgn}\,$ denotes the sign homomorphism $\mbox{sgn}\,:{\cal S}_k\rightarrow\{-1,1\}$. Therefore $${\widetilde H}_{k-2}(A_k/{\cal S}(T),{\Bbb Q})={\widetilde H}_{k-2}^{{\cal S}(T)}(A_k,{\Bbb Q})= \begin{cases} {\Bbb Q},&\mbox{ if } {\cal S}(T)\subseteq{\cal A}_k,\\ 0, & \mbox{ otherwise,} \end{cases}$$ where ${\cal A}_k$ is the alternating group, ${\cal A}_k=\mbox{sgn}\,^{-1}(1)$. Combined with the Proposition~\ref{pr6.1} this gives $H_i(E_T,{\Bbb Q})={\Bbb Q}$, if $i=|E(T)|-1$ and ${\cal S}(T)\subseteq{\cal A}_{|E(T)|}$, and $H_i(E_T,{\Bbb Q})=0$ in all other cases. Therefore it follows from~\eqref{eq6.1} that ${\mbox{rk}\,} E_{k-1,k}^1=f_{k,n}$, where $f_{k,n}$ is equal to the number of forests $T$ with $k$ edges and $n$ vertices, such that ${\cal S}(T)\subseteq{\cal A}_k$. ${\mbox{rk}\,} E_{p,k}^1=0$ for $p\neq k-1$. Note that $\beta_i(X_n,{\Bbb Q})=0$, for $i\neq n-2$, because $\beta_i(\Delta(G_n),{\Bbb Q})=0$, for $i\neq n-2$ (by the Theorem~\ref{thm1}), and $\beta_i(X_n,{\Bbb Q})=\beta_i^{{\cal S}_n}(\Delta(G_n),{\Bbb Q})$. In particular, by computing the Euler characteristic of $X_n$ in two different ways, we obtain \begin{thm} For $n\geq 3$, $\beta_{n-2}(X_n,{\Bbb Q})=\sum_{k=2}^{n-1} (-1)^{n+k+1}f_{k,n}$. \end{thm} The first values of $f_{k,n}$ are given in the Table~4.3. Note that there are zeroes on and below the main diagonal and that the rows stabilize at the entry $(k,2k-1)$ (for $k\geq 2$). $$\begin{tabular}{ c c c | c c c | c c c | c c c | c c c | c c c | c c c |} & $k\backslash n$ &&& 1 &&& 2 &&& 3 &&& 4 &&& 5 &&& 6 &\\ \hline & 1 &&& 0 &&& {\bf 1} &&& 1 &&& 1 &&& 1 &&& 1 &\\ \hline & 2 &&& 0 &&& 0 &&& {\bf 1} &&& 1 &&& 1 &&& 1 &\\ \hline & 3 &&& 0 &&& 0 &&& 0 &&& 2 &&& {\bf 3} &&& 3 &\\ \hline & 4 &&& 0 &&& 0 &&& 0 &&& 0 &&& 4 &&& 7 &\\ \hline & 5 &&& 0 &&& 0 &&& 0 &&& 0 &&& 0 &&& 8 &\\ \hline \end{tabular} $$ $$\mbox{Table 4.3}$$ \subsection{${\Bbb Z}$ coefficients} \noindent The case of integer coefficients is more complicated. In general, we do not even know the entries of the first tableau. However, we do know that it is different from the rational case, i.e.,~torsion may occur. For example, let $T$ be the~forest with 8 vertices and 6 edges depicted on Figure~4.4. Clearly, ${\cal S}(T)=\{\mbox{id},(12)(34)(56)\}$. It is easy to see that $A_6/{\cal S}(T)$ is a~double suspension (by which we mean suspension of suspension) of ${\Bbb R}{\Bbb P}^2$, thus the only nonzero homology group is ${\widetilde H}_3(A_6/{\cal S}(T),{\Bbb Z})={\Bbb Z}_2$. In particular, $E_{4,6}^1$ is not free. $$\epsffile{fig3.eps}$$ $$\mbox{Figure 4.4}$$ On the positive side, we can describe the values which $d^1$ takes on the ``rational'' generators of $E_{*,*}^1$. Let us call a~forest {\it admissible} if ${\cal S}(T)\subseteq{\cal A}_{|E(T)|}$. For every admissible forest $T$ with $k$ edges we fix some order on the edges, i.e.,~a~bijection $\psi_T:E(T)\rightarrow[k]$. This determines uniquely an~integer generator $e_T$ of $H_{k-1}(E_T,{\Bbb Z})$ by \begin{equation}\label{eqet} e_T=\sum_{{\cal S}(T)g}\mbox{sgn}\,(g)(T,g\circ\psi_T), \end{equation} where we sum over all right cosets of ${\cal S}(T)$, (we choose one representative for each coset). Observe that the sign of $g$, resp.~the simplex $(T,g\circ\psi_T)$, are the same for different representatives of the same right coset class, because ${\cal S}(T)\subseteq{\cal A}_k$, resp.~by the definition of~${\cal S}(T)$. \begin{prop}\label{pr6.2} For an~admissible forest $T$, we have \begin{equation}\label{d1eq} d^1(e_T)=\sum_{\alpha}\mbox{sgn}\,(\tilde\psi_{T,\alpha}\circ\psi_T^{-1}) \lambda_{T,\alpha}e_{T\setminus\alpha}, \end{equation} where the sum is over ${\cal S}(T)$-orbits of $E(T)$, for which there exists a~representative $\alpha$, such that $T\setminus\alpha$ is admissible, we choose one representative for each orbit; note that the admissibility of $T\setminus\alpha$ depends only on the ${\cal S}(T)$-orbit of $\alpha$, not on the choice of the representative. Notation in the formula: $T\setminus\alpha$ denotes the forest obtained from $T$ by removing the edge $\alpha$; $\tilde\psi_{T,\alpha}:E(T)\rightarrow[k]$ is defined by $\tilde\psi_{T,\alpha}|_{T\setminus\alpha}=\psi_{T\setminus\alpha}$ and $\tilde\psi_{T,\alpha}(\alpha)=k$; $\lambda_{T,\alpha}=[{\cal S}(T\setminus\alpha): \widetilde {\cal S}(T)]$, where $\widetilde {\cal S}(T)$ consists of those permutations of edges of $T\setminus\alpha$ which can be extended to $T$ by fixing the additional edge. \end{prop} \nin{\bf Proof. } For an~admissible forest~$T$ with $k$ edges and a~bijection $\phi:E(T)\rightarrow[k]$, let $(\widetilde T,\tilde\phi)$ denote a~face simplex of $(T,\phi)$, where $\widetilde T$ is obtained from $T$ by removing the edge with the highest label, $\tilde\phi$ is the restriction of $\phi$ to $\widetilde T$. In our notations $(\widetilde T,\tilde\phi)=(T\setminus\phi^{-1}(k),\phi|_{E(T\setminus \phi^{-1}(k))})$. However, for convenience, we use the notation ``tilde'' in the rest of the proof. According to the general theory for spectral sequences, $d^1(e_T)=\partial(e_T)$, where $\partial$ denotes the usual boundary operator, and we view $\partial(e_T)$ as embedded into the relative homology group $H_{k-2}(F_{k-1},F_{k-2})$. $\partial(e_T)$ is a~linear combination of simplices which are obtained from the simplices $(T,g\circ\psi_T)$ by either merging two labels, or omitting the edge with the top label. $e_T\in H_{k-1}(F_k,F_{k-1})$ means that the application of the ``truncated'' boundary operator to $e_T$ gives 0, therefore all the simplices obtained by merging two labels will cancel out. Furthermore, since $\partial(e_T)\in H_{k-2}(F_{k-1},F_{k-2})$, $\dim F_{k-1}=k-2$, and the group $H_{k-2}(F_{k-1},F_{k-2})$ is freely generated by $e_U$, where $U$ is an~admissible forest with $k-1$ edges, we can conclude that also the contributions $(\widetilde T,\tilde\phi)$, where $\widetilde T$ is not admissible, will cancel out. Combining these arguments with~\eqref{eqet} we obtain: \begin{equation}\label{raz} d^1(e_T)=\sum_{{\cal S}(T)g}\mbox{sgn}\,(g)(\widetilde T,\widetilde {g\circ\psi_T}), \end{equation} where we have only those terms left in the sum, for which $\widetilde T$ is admissible. After regrouping we get \begin{equation}\label{dva} \sum_{{\cal S}(T)g}\mbox{sgn}\,(g)(\widetilde T,\widetilde {g\circ\psi_T})=\sum_{\alpha} \sum_{{\cal S}(T)g}\mbox{sgn}\,(g)(\widetilde T,\widetilde {g\circ\psi_T}), \end{equation} where in the second term the first sum is taken over all ${\cal S}(T)$-orbits of~$[k]$, for which $\widetilde T$ is admissible, while the second sum is taken over all right cosets ${\cal S}(T)g$ which have a~representative $g$ such that $g\circ\psi_T(\alpha)=k$, we take one representative per coset. To verify~\eqref{dva} we just need to observe that the ${\cal S}(T)$-orbit of $(g\circ\psi_T)^{-1}(k)$ does not depend on the choice of the representative of ${\cal S}(T)g$; this follows from the definition of ${\cal S}(T)$. Finally, one can see that, for $\alpha$ being an~edge of~$T$, such that $T\setminus\alpha$ is admissible, \begin{equation}\label{tri} \sum_{{\cal S}(T)g}\mbox{sgn}\,(g) (\widetilde T,\widetilde {g\circ\psi_T})=\mbox{sgn}\,(\tilde\psi_{T,\alpha} \circ\psi_T^{-1})\lambda_{T,\alpha}\sum_{{\cal S}(T\setminus\alpha)h} \mbox{sgn}\,(h)(T\setminus\alpha,h\circ\psi_{T\setminus\alpha}), \end{equation} where the sum in the first term is again taken over all right cosets ${\cal S}(T)g$ which have a~representative $g$ such that $g\circ\psi_T(\alpha)=k$, and the sum in the second term is simply over all right cosets of ${\cal S}(T\setminus\alpha)$. Indeed, on the left hand side we have a~sum over all labelings of $E(T)$ with numbers $1,\dots,k$, such that $\alpha$ gets a~label~$k$, and we consider these labelings up to a~symmetry of $T$; each labeling comes in with a~sign of the permutation $g$, which is obtained by reading off this labeling in the order prescribed by $\psi_T$. On the right hand side the same sum is regrouped, using the observation that to label $E(T)$ with $[k]$, so that $\alpha$ gets a~label~$k$, is the same as to label $E(T\setminus\alpha)$ with $[k-1]$. The only details which need attention are the multiplicity and the sign. Every ${\cal S}(T)$-orbit of labelings of $E(T)$ with $[k]$ so that $\alpha$ gets a~label~$k$ corresponds to $[{\cal S}(T\setminus\alpha):\widetilde {\cal S}(T)]$ of ${\cal S}(T\setminus\alpha)$-orbits of labelings of $E(T\setminus\alpha)$ with $[k-1]$, since we identify labelings by the actions of different groups: ${\cal S}(T\setminus\alpha)\supseteq\widetilde {\cal S}(T)$. Each of this ${\cal S}(T\setminus\alpha)$-orbits comes with the same sign, because ${\cal S}(T\setminus\alpha)\subseteq{\cal A}_{k-1}$. The sign $\mbox{sgn}\,(\tilde\psi_{T,\alpha}\circ\psi_T^{-1})$ corresponds to the change of the order in which we read off the edges: instead of reading them off according to $\psi_T$, we first read off along $\psi_{T\setminus\alpha}$ and then read off the edge $\alpha$ last. Formally: $g\circ\psi_T= \tilde h\circ\tilde\psi_{T,\alpha}$, and $\mbox{sgn}\,\tilde h=\mbox{sgn}\, h$, hence $\mbox{sgn}\, g=\mbox{sgn}\, h \,\,\mbox{sgn}\,(\tilde\psi_{T,\alpha}\circ\psi_T^{-1})$, where $\tilde h$ is defined by $\tilde h|_{[k-1]}=h$, $\tilde h(k)=k$. Combining \eqref{raz}, \eqref{dva} and \eqref{tri} we obtain \eqref{d1eq}. $\square$ \subsection{Homology groups of $X_n$ for $n=2,3,4,5,6$} $X_2$ is just a~point. As shown in Figure~4.3, $X_3\simeq S^1$, where $\simeq$ denotes homotopy equivalence. With a~bit of labor, one can manually verify that $X_4\simeq S^2$. Furthermore, one can see that $H_3(X_5,{\Bbb Z})={\Bbb Z}^2$ and ${\widetilde H}_i(X_5,{\Bbb Z})=0$ for $i\neq 3$. We leave this to the reader, while confining ourselves to the case $n=6$. On Figure~4.5 we have all forests on 6 vertices. We denote some of the forests by two digits. The numbers over the edges denote the order in which we read the labels, i.e.,~the bijection $\psi_T$. It is easy to see that $A_k/{\cal S}(T)$ is homeomorphic to $S^{k-2}$ for all admissible $T$, and is contractible otherwise. The only nontrivial cases are 41, 47, 48, 51, 55, and 59, all of which can be verified directly. Therefore, the only nontrivial entries of $E_{*,*}^1$ (${\Bbb Z}$ coefficients) will lie on the $(k-1,k)$-diagonal. Thus $H_*(X_6,{\Bbb Z})$ can be computed from the chain complex $0\leftarrow E_{0,1}^1\stackrel{d^1}{\longleftarrow}E_{1,2}^1\stackrel{d^1}{\longleftarrow} E_{2,3}^1\stackrel{d^1}{\longleftarrow} E_{3,4}^1\stackrel{d^1}{\longleftarrow} E_{4,5}^1\leftarrow 0$. By Proposition~\ref{pr6.2} we have the following relations: $$\begin{tabular}{ c c c c c } $d^1(11)=0$, && $d^1(21)=0$, && \\ $d^1(31)=2\cdot 21$, && $d^1(32)=21$, && $d^1(33)=21$, \\ $d^1(41)=32-33$, && $d^1(42)=31-32-33$, && $d^1(43)=31-2\cdot 33$,\\ $d^1(44)=0$, && $d^1(45)=31-32-33$, && $d^1(46)=32-33$, \\ && $d^1(47)=0$, && \\ \end{tabular} $$ $$\begin{tabular}{ c c c } $d^1(51)=41-46$, && $d^1(52)=-42+43+44-46$, \\ $d^1(53)=-42+45$, && $d^1(54)=2\cdot 41+42-43-46+2\cdot 47$, \\ $d^1(55)=41-43+45$, && $d^1(56)=42+44-45-2\cdot 47$, \\ $d^1(57)=-43+44+45+46$, && $d^1(58)=2\cdot 44+2\cdot 47$, \\ \end{tabular} $$ here the two-digit strings denote the corresponding forests on Figure~4.5. Thus ${\widetilde H}_3(X_6,{\Bbb Z})={\Bbb Z}_2$, ${\widetilde H}_4(X_6,{\Bbb Z})={\Bbb Z}^3$ and ${\widetilde H}_i(X_6,{\Bbb Z})=0$ for $i\neq 3,4$. \vskip5pt {\it Therefore we conclude that 6 is the smallest value of $n$, for which the homology groups $H_*(X_n,{\Bbb Z})$ are not free.} \newpage $$\epsffile{fig41.eps}$$ $$\epsffile{fig42.eps}$$ $$\epsffile{fig43.eps}$$ $$\epsffile{fig44.eps}$$ $$\epsffile{fig45.eps}$$ $$\mbox{Figure 4.5}$$ \chapter{Group Actions on Posets} \section{Preamble} Assume that we have a finite group $G$ acting on a poset $P$ in an order-preserving way. The~purpose of this chapter is to compare the various constructions of the quotient, associated with this action. Our~basic suggestion is to view $P$ as a category and the group action as a functor from $G$ to $\mbox{\bf Cat}$. Then, it is natural to define $P/G$ to be the colimit of this functor. As a~result $P/G$ is in general a~category, not a~poset. After getting a hand on the formal setting in Section~\ref{s5.2} we proceed in Section~\ref{s5.3} with imposing different conditions on the group action. We give conditions for each of the following properties to be satisfied: \begin{enumerate} \item[(1)] the morphisms of $P/G$ are exactly the orbits of the morphisms of~$P$, we call it {\it regularity}; \item[(2)] the quotient construction commutes with Quillen's nerve functor; \item[(3)] $P/G$ is again a poset. \end{enumerate} Furthermore, we study the class of categories which can be seen as the ``quotient closure'' of the set of all finite posets: loopfree categories. \section{Formalization of group actions and the main question} \label{s5.2} \subsection{Preliminaries} \noindent For a~small category $K$ denote the~set of its objects by ${\cal O}(K)$ and the~set of its morphisms by ${\cal M}(K)$. For every $a\in{\cal O}(K)$ there is exactly one identity morphism which we denote $\mbox{id}_a$, this allows us to identify ${\cal O}(K)$ with a~subset of ${\cal M}(K)$. If $m$ is a~morphism of $K$ from $a$ to $b$, we write $m\in{\cal M}_K(a,b)$, $\partial\,^\bullet m=a$ and $\partial\,_\bullet m=b$. The~morphism $m$ has an~inverse $m^{-1}\in{\cal M}_K(b,a)$, if $m\circ m^{-1}=\mbox{id}_a$ and $m^{-1}\circ m =\mbox{id}_b$. If only the~identity mor\-phisms have inverses in $K$ then $K$ is said to be a~category without inverses. We denote the~category of all small categories by $\mbox{\bf Cat}$. If $K_1,K_2\in{\cal O}(\mbox{\bf Cat})$ we denote by ${\cal F}(K_1,K_2)$ the set of functors from $K_1$ to $K_2$. We need three full subcategories of $\mbox{\bf Cat}$: ${\bf P}$ the~category of posets, (which are categories with at most one morphism, denoted $(x\rightarrow y)$, between any two objects $x,y$), ${\bf L}$ the~category of loopfree categories (see Definition~\ref{5dc}), and $\mbox{\bf Grp}$ the~category of groups, (which are categories with a single element, morphisms given by the group elements and the law of composition given by group multiplication). Finally, {\bf 1} is the~terminal object of $\mbox{\bf Cat}$, that is, the~category with one element, and one (identity) morphism. The~other two categories we use are $\mbox{\bf Top}$, the~category of topological spaces, and $\mbox{\bf SS}$, the~category of simplicial sets. We are also interested in the~functors $\Delta:\mbox{\bf Cat}\rightarrow\mbox{\bf SS}$ and ${\cal R}:\mbox{\bf SS}\rightarrow\mbox{\bf Top}$. The~composition is denoted $\tilde\Delta:\mbox{\bf Cat}\rightarrow\mbox{\bf Top}$. Here, $\Delta$ is the~nerve functor, see Appendix~B, or~\cite{Qu73,Qu78,Se68}. In particular, the~simplices of $\Delta(K)$ are chains of morphisms in $K$, with dege\-ne\-rate simplices corresponding to chains that include identity morphisms, see~\cite{GeM96,We94}. ${\cal R}$ is the~topological realization functor, see~\cite{Mil57}. We recall here the definition of a colimit (see~\cite{ML98,Mit65}). \begin{df} \label{5col} Let $K_1$ and $K_2 $ be categories and $X\in{\cal F}(K_1,K_2)$. A~{\bf sink} of $X$ is a~pair consisting of $L\in{\cal O}(K_2)$, and a~collection of morphisms $\{\lambda_s\in{\cal M}_{K_2} (X(s),L)\}_{s\in{\cal O}(K_1)}$, such that if $\alpha\in{\cal M}_{K_1}(s_1,s_2)$ then $\lambda_{s_2}\circ X(\alpha)=\lambda_{s_1}$. (One way to think of this collection of morphisms is as a natural transformation between the functors $X$ and $X'=X_1\circ X_2$, where $X_2$ is the terminal functor $X_2:K_1\rightarrow\bf 1$ and $X_1:{\bf 1}\rightarrow K_2$ takes the object of $\,\bf 1$ to $L$). When $(L,\{\lambda_s\})$ is universal with respect to this property we call it the {\bf colimit} of $X$ and write $L=\mbox{colim}\, X$. \end{df} \subsection{Definition of the quotient and formulation of the main problem} \noindent Our main object of study is described in the~following definition. \begin{df}\label{5adf} We say that a group $G$ {\bf acts on} a category $K$ if there is a functor ${\cal A}_K:G\rightarrow\mbox{\bf Cat}$ which takes the~unique object of G to $K$. The~colimit of ${\cal A}_K$ is called the {\bf quotient} of $K$ by the action of $G$ and is denoted by $K/G$. \end{df} To simplify notations, we identify ${\cal A}_K g$ with $g$ itself. Furthermore, in Definition~\ref{5adf} the category $\mbox{\bf Cat}$ can be replaced with any category $C$, then $K,K/G\in{\cal O}(C)$. Important special case is $C=\mbox{\bf SS}$. It arises when $K\in{\cal O}(\mbox{\bf Cat})$ and we consider $\mbox{colim}\,\Delta\circ{\cal A}_K=\Delta(K)/G$. \vskip8pt \noindent {\bf Main Problem.} $\,$ {\it Understand the relation between the topological and the categorical quotients, that is, between $\Delta(K/G)$ and $\Delta(K)/G$.} \vskip8pt To start with, by the~universal property of colimits there exists a~canonical sur\-jec\-tion $\lambda:\Delta(K)/G\rightarrow\Delta(K/G)$. In the~next section we give com\-bi\-na\-torial con\-di\-tions under which this map is an isomorphism. The general theory tells us that if $G$ acts on the~category $K$, then the~colimit $K/G$ exists, since $\mbox{\bf Cat}$ is cocomplete. We shall now give an explicit description. \vskip4pt \noindent {\bf An explicit description of the category $K/G$.} \vskip4pt \noindent When $x$ is a morphism of~$K$, denote by $Gx$ the~orbit of $x$ under the~action of $G$. We have ${\cal O}(K/G)=\{Ga\,|\,a\in{\cal O}(K)\}$. The situation with morphisms is more complicated. Define a relation $\leftrightarrow$ on the~set ${\cal M}(K)$ by setting $x\leftrightarrow y$, iff there are decompositions $x=x_1\circ\dots\circ x_t$ and $y=y_1\circ\dots\circ y_t$ with $Gy_i=Gx_i$ for all $i\in [t]$. The~relation $\leftrightarrow$ is reflexive and symmetric since $G$ has identity and inverses, however it is not in general transitive. Let $\sim$ be the~transitive closure of $\leftrightarrow$, it is clearly an equivalence relation. Denote the~$\sim$ equivalence class of $x$ by $[x]$. Note that $\sim$ is the~minimal equivalence relation on ${\cal M}(K)$ closed under the $G$ action and under composition; that is, with $a\sim ga$ for any $g\in G$, and if $x\sim x'$ and $y\sim y'$ and $x\circ x'$ and $y\circ y'$ are defined then $x\circ x'\sim y \circ y'$. It is not difficult to check that the~set $\{[x]\,|\,x\in{\cal M}(K)\}$ with the~relations $\partial\,_\bullet[x]=[\partial\,_\bullet x]$, $\partial\,^\bullet[x]=[\partial\,^\bullet x]$ and $[x]\circ[y]=[x\circ y]$ (whenever the~composition $x\circ y$ is defined), are the morphisms of the~category $K/G$. \vskip4pt Note that if $P$ is a poset with a $G$ action, the~quotient taken in $\mbox{\bf Cat}$ need not be a poset, and hence may differ from the~poset quotient. \begin{exam} Let $P$ be the center poset in the figure below. Let ${\cal S}_2$ act on $P$ by simultaneously permuting $a$ with $b$ and $c$ with $d$. (I) shows $P/{\cal S}_2$ in ${\bf P}$ and (II) shows $P/{\cal S}_2$ in $\mbox{\bf Cat}$. Note that in this case the quotient in $\mbox{\bf Cat}$ commutes with the functor $\Delta$ (the canonical surjection $\lambda$ is an isomorphism), whereas the quotient in ${\bf P}$ does not. \end{exam} $$\epsffile{pic1.eps}$$ $$\mbox{Figure 5.1}$$ \section{Conditions on group actions} \label{s5.3} \subsection{Outline of the results and surjectiveness of the canonical map} \noindent In this section we consider combinatorial conditions for a~group $G$ acting on a~category $K$ which ensure that the quotient by the group action commutes with the nerve functor. If ${\cal A}_K:G\rightarrow\mbox{\bf Cat}$ is a~group action on a~category $K$ then $\Delta\circ{\cal A}_K:G \rightarrow \mbox{\bf SS}$ is the associated group action on the nerve of $K$. It is clear that $\Delta(K/G)$ is a~sink for $\Delta\circ{\cal A}_K$, and hence, as previously mentioned, the universal property of colimits gives a~canonical map $\lambda:\Delta(K)/G\rightarrow\Delta(K/G)$. We wish to find conditions under which $\lambda$ is an isomorphism. First we prove in Proposition~\ref{5sur} that $\lambda$ is always surjective. Furthermore, $Ga=[a]$ for $a\in{\cal O}(K)$, which means that, restricted to $0$-skeleta, $\lambda$ is an~iso\-morphism. If the two simplicial spaces were simplicial complexes (only one face for any fixed vertex set), this would suffice to show isomorphism. Neither one is a~simplicial complex in general, but while the quotient of a~complex $\Delta(K)/G$ can have simplices with fairly arbitrary face sets in common, $\Delta(K/G)$ has only one face for any fixed edge set, since it is a~nerve of a~category. Thus for $\lambda$ to be an~iso\-morphism it is necessary and sufficient to find conditions under which \begin{enumerate} \item[{1)}] $\lambda$ is an isomorphism restricted to $1$-skeleta; \item[{2)}] $\Delta(K)/G$ has only one face with any given set of edges. \end{enumerate} We will give conditions equivalent to $\lambda$ being an~isomorphism, and then give some stronger conditions that are often easier to check, the strongest of which is also inherited by the action of any subgroup $H$ of $G$ acting on $K$. First note that a~simplex of $\Delta(K/G)$ is a~sequence $([m_1],\dots,[m_t])$, $m_i\in{\cal M}(K)$, with $\partial\,_\bullet [m_{i-1}]=\partial\,^\bullet [m_i]$, which we will call a~{\it chain}. On the other hand a simplex of $\Delta(K)/G$ is an~orbit of a~sequence $(n_1,\dots,n_t)$, $n_i\in{\cal M}(K)$, with $\partial\,_\bullet n_{i-1}=\partial\,^\bullet n_i$, which we denote $G(n_1,\dots,n_t)$. The~canonical map $\lambda$ is given by $\lambda(G(n_1,\dots,n_t))=([n_1],\dots,[n_t])$. \begin{prop} \label{5sur} Let $K$ be a category and $G$ a group acting on $K$. The canonical map $\lambda:\Delta(K)/G\rightarrow\Delta(K/G)$ is surjective. \end{prop} \nin{\bf Proof. } By the above description of $\lambda$ it suffices to fix a chain $([m_1],\dots,[m_t])$ and find a chain $(n_1,\dots, n_t)$ with $[n_i]=[m_i]$. The proof is by induction on $t$. The case $t=1$ is obvious, just take $n_1=m_1$. Assume now that we have found $n_1,\dots,n_{t-1}$, so that $[n_i]=[m_i]$, for $i=1,\dots,t-1$, and $n_1,\dots,n_{t-1}$ compose, i.e., $\partial\,^\bullet n_i=\partial\,_\bullet n_{i+1}$, for $i=1,\dots,t-2$. Since $[\partial\,_\bullet n_{t-1}]=[\partial\,_\bullet m_{t-1}]=[\partial\,^\bullet m_t]$, we can find $g\in G$, such that $g\partial\,^\bullet m_t=\partial\,_\bullet n_{t-1}$. If we now take $n_t=g m_t$, we see that $n_{t-1}$ and $n_t$ compose, and $[n_t]=[m_t]$, which provides a proof for the induction step. $\square$ \subsection{Conditions for injectiveness of the canonical projection} \begin{df} \label{5df3.2} Let $K$ be a category and $G$ a group acting on $K$. We say that this action satisfies {\bf Condition (R)} if the following is true: If $x,y_a,y_b\in{\cal M}(K)$, $\partial\,_\bullet x=\partial\,^\bullet y_a= \partial\,^\bullet y_b$ and $Gy_a=Gy_b$, then $G(x\circ y_a)=G(x\circ y_b)$. \end{df} $$\epsffile{pic2.eps}$$ $$\mbox{Figure 5.2}$$ We say in such case that $G$ acts {\it regularly} on~$K$. \begin{prop} \label{5reg} Let $K$ be a category and $G$ a group acting on $K$. This action satisfies Condition (R) iff the~canonical surjection $\lambda:\Delta(K)/G \rightarrow \Delta(K/G)$ is injective on $1$-skeleta. \end{prop} \nin{\bf Proof. } The injectiveness of $\lambda$ on $1$-skeleta is equivalent to requiring that $Gm=[m]$, for all $m\in{\cal M}(K)$, while Condition~(R) is equivalent to requiring that $G(m \circ Gn)=G(m \circ n)$, for all $m,n\in{\cal M}(K)$ with $\partial\,_\bullet m=\partial\,^\bullet n$; here $m\circ Gn$ means the set of all $m\circ gn$ for which the composition is defined. Assume that $\lambda$ is injective on $1$-skeleta. The we have the following computation: $$G(m \circ Gn)=Gm \circ Gn=[m]\circ [n]=[m \circ n]=G(m \circ n),$$ hence the Condition (R) is satisfied. Reversely, assume that the Condition~(R) is satisfied, that is $G(m \circ Gn)=G(m \circ n)$. Since the equivalence class $[m]$ is generated by $G$ and composition, it suffices to show that orbits are preserved by com\-po\-sition, which is precisely $G(m \circ Gn)=G(m \circ n)$. $\square$ \vskip4pt The following theorem is the main result of this chapter. It provides us with combinatorial conditions which are equivalent to $\lambda$ being an isomorphism. \begin{thm} \label{5tc} Let $K$ be a category and $G$ a group acting on $K$. The~following two assertions are equivalent for any $t\geq 2$: \begin{enumerate} \item[(1$_t$)] {\bf Condition (C$_{\mbox{\bf t}}$).} If $m_1,\dots,m_{t-1},m_{a},m_{b}\in{\cal M}(K)$ with $\partial\,^\bullet m_i=\partial\,_\bullet m_{i-1}$ for all $2\leq i\leq t-1$, $\partial\,^\bullet m_{a}=\partial\,^\bullet m_{b}=\partial\,_\bullet m_{t-1}$, and $Gm_{a}=Gm_{b}$, then there is some $g\in G$ such that $gm_{a}=m_{b}$ and $g m_i=m_i$ for $1 \leq i \leq t-1$. \item[(2$_t$)] The~canonical surjection $\lambda : \Delta(K)/G \rightarrow \Delta(K/G)$ is injective on $t$-skeleta. \end{enumerate} In particular, $\lambda$ is an isomorphism iff {\bf (C$_{\mbox{\bf t}}$)} is satisfied for all $t\geq 2$. If this is the case, we say that Condition {\bf (C)} is satisfied \end{thm} $$\epsffile{pic3.eps}$$ $$\mbox{Figure 5.3}$$ \nin{\bf Proof. } $(1_t)$ is equivalent to $G(m_1,\dots,m_t)=G(m_1,\dots,m_{t-1},Gm_t)$; this notation is used, as before, for all sequences $(m_1,\dots,m_{t-1},gm_t)$ which are chains, that is for which $m_1\circ\dots\circ m_{t-1}\circ gm_t$ is defined. $(2_t)$ implies Condition (R) above, and so can be restated as $G(m_1,\dots,m_t)=(Gm_1,\dots,Gm_t)$. \begin{multline*} \underline{(2_t)\Rightarrow (1_t):} \,\,G(m_1,\dots,m_t)=(Gm_1,\dots,Gm_t)= G(Gm_1,\dots,Gm_t) \\ \supseteq G(m_1,\dots,m_{t-1},Gm_t)\supseteq G(m_1,\dots,m_t). \end{multline*} \vskip-30pt \begin{multline*} \underline{(1_2)\Rightarrow (2_2):} \,\,G(m_1,m_2)=G(m_1,Gm_2) \\ =\{g_1(m_1,g_2 m_2)\,|\,\partial\,_\bullet m_1=\partial\,^\bullet g_2 m_2\}\, \\ =\{(g_1 m_1,g_2 m_2)\,|\,\partial\,_\bullet g_1 m_1=\partial\,^\bullet g_2 m_2\}=(Gm_1,Gm_2). \end{multline*} \hskip-2pt $\underline{(1_t)\Rightarrow (2_t),\,t\geq 3:}$ We use induction on $t$. \begin{equation*} \begin{split} G(m_1, \dots, m_t)&=G(m_1, \dots, m_{t-1}, Gm_t) \\ &=\{(gm_1,\dots,gm_{t-1},\tilde gm_t\,|\,\partial\,_\bullet gm_{t-1}=\partial\,^\bullet \tilde gm_t)\} \\ &=\{(g_1 m_1,\dots,g_t m_t)\,|\,\partial\,_\bullet g_i m_i=\partial\,^\bullet g_{i+1}m_{i+1}, i\in[t-1]\} \\ &=(Gm_1,\dots,Gm_t).\,\,\,\,\square \end{split} \end{equation*} \begin{exam} {\rm A~group action which satisfies Condition (C$_t$), but does not satisfy Condition~(C$_{t+1}$).} Let $P_{t+1}$ be the order sum of $t+1$ copies of the 2-element antichain. The~automorphism group of $P_{t+1}$ is the direct product of $t+1$ copies of ${\Bbb Z}_2$. Take $G$ to be the index $2$ subgroup consisting of elements with an even number of nonidentity terms in the product. \end{exam} \noindent The~following condition implies Condition (C), and is often easier to check. \vskip4pt \noindent {\bf Condition (S).} There exists a set $\{S_m\}_{m\in{\cal M}(K)}$, $S_m\subseteq\,\,$Stab$\,(m)$, such that \begin{enumerate} \item[(1)] $S_m\subseteq S_{\partial\,^\bullet m}\subseteq S_{m'}$, for any $m'\in{\cal M}(K)$, such that $\partial\,_\bullet m'=\partial\,^\bullet m$; \item[(2)]\vskip-5pt $S_{\partial\,^\bullet m}$ acts transitively on $\{g m\,|\,g\in\,$Stab$\,(\partial\,^\bullet m)\}$, for any $m\in{\cal M}(K)$. \end{enumerate \begin{prop} Condition (S) implies Condition (C). \end{prop} \vskip-4pt \nin{\bf Proof. } Let $m_1,\dots,m_{t-1},m_a,m_b$ and $g$ be as in Condition (C), then, since $g\in\,$Stab$\,(\partial\,^\bullet m_a)$, there must exist $\tilde g\in S_{\partial\,^\bullet m_a}$ such that $\tilde g(m_a)=m_b$. From (1) above one can conclude that $\tilde g(m_i)=m_i$, for $i\in [t-1]$. $\square$\vskip4pt We say that the {\bf strong} Condition (S) is satisfied if Condition (S) is satisfied with $S_a=\,$Stab$\,(a)$. Clearly, in such a~case part (2) of the Condition~(S) is obsolete. \begin{exam} {\rm A~group action satisfying Condition~(S), but not the strong Condition~(S).} Let $K={\cal B}_n$, lattice of all subsets of $[n]$ ordered by inclusion, and let $G={\cal S}_n$ act on ${\cal B}_n$ by permuting the ground set $[n]$. Clearly, for $A\subseteq[n]$, we have Stab$\,(A)={\cal S}_A\times{\cal S}_{[n]\setminus A}$, where, for $X\subseteq[n]$, ${\cal S}_X$ denotes the subgroup of ${\cal S}_n$ which fixes elements of $[n]\setminus X$ and acts as a permutation group on the set $X$. Since $A>B$ means $A\supset B$, condition (1) of (S) is not satisfied for $S_A=\,$Stab$\,(A)$: ${\cal S}_A\times{\cal S}_{[n]\setminus A}\not\supseteq{\cal S}_B\times{\cal S}_{[n]\setminus B}$. However, we can set $S_A={\cal S}_A$. It is easy to check that for this choice of $\{S_A\}_{A\in{\cal B}_n}$ Condition~(S) is satisfied. \end{exam} We close the discussion of the conditions stated above by the following proposition. \begin{prop}\label{5int} $\,$ \noindent 1) The~sets of group actions which satisfy Condition (C) or Condition (S) are closed under taking the restriction of the group action to a subcategory. \noindent 2) Assume a finite group $G$ acts on a poset $P$, so that Condition (S) is satisfied. Let $x\in P$ and $S_x\subseteq H\subseteq\,$Stab$\,(x)$, then Condition (S) is satisfied for the action of $H$ on $P_{\leq x}$. \noindent 3) Assume a finite group $G$ acts on a category $K$, so that Condition (S) is satisfied with $S_a=$Stab$(a)$ (strong version), and $H$ is a subgroup of $G$. Then the strong version of Condition $(S)$ is again satisfied for the action of $H$ on $K$. \end{prop} \vskip-4pt \nin{\bf Proof. } 1) and 3) are obvious. To show 2) observe that for $a\leq x$ we have $S_a\subseteq S_x\subseteq H$, hence $S_a\subseteq H\cap\,$Stab$\,(a)$. Thus condition (1) remains true. Condition~(2) is true since $\{g(b)\,|\,g\in\,$Stab$\,(a)\}\supseteq\{g(b)\,|\,g\in\,$Stab$\,(a)\cap H\}$. $\square$ \subsection{Conditions for the categories to be closed under taking quotients} Next, we are concerned with finding out what categories one may get as a~quotient of a poset by a group action. In particular, we ask: {\it in which cases is the quotient again a poset?} To answer that question, it is convenient to use the following class of cate\-go\-ries. \begin{df}\label{5dc} A~category is called {\bf loopfree} if it has no inverses and no nonidentity automorphisms. \end{df} Intuitively, one may think of loopfree categories as those which can be drawn so that all nontrivial morphisms point down. To familiarize us with the notion of a~loopfree category we make the following observations: \begin{itemize} \item $K$ is loopfree iff for any $x,y\in{\cal O}(K)$, $x\neq y$, only one of the sets ${\cal M}_K(x,y)$ and ${\cal M}_k(y,x)$ is non-empty and ${\cal M}_K(x,x)=\{\mbox{id}_x\}$; \item a poset is a loopfree category; \item a barycentric subdivision of an arbitrary category is a loopfree category; \item a barycentric subdivision of a loopfree category is a poset; \item if $K$ is a loopfree category, then there exists a~partial order $\geq$ on the set ${\cal O}(K)$ such that ${\cal M}_K(x,y)\neq\emptyset$ implies $x\geq y$. \end{itemize} \begin{df}\label{5hor} Suppose $K$ is a~small category, and $T\in{\cal F}(K,K)$. We say that $T$ is {\bf hori\-zontal} if for any $x\in{\cal O}(K)$, if $T(x)\neq x$, then ${\cal M}_K(x,T(x))={\cal M}_K(T(x),x)=\emptyset$. When a~group $G$ acts on $K$, we say that the action is hori\-zontal if each $g\in G$ is a~horizontal functor. \end{df} When $K$ is a~finite loopfree category, the action is always horizontal. Another example of horizontal actions is given by rank preserving action on a~(not necessarily finite) poset. We have the following useful property: \begin{prop}\label{5star} Let $P$ be a finite loopfree category and $T\in{\cal F}(P,P)$ be a hori\-zontal functor. Let $\tilde T\in{\cal F}(\Delta(P),\Delta(P))$ be the induced functor, i.e., $\tilde T=\Delta(T)$. Then $\Delta(P_T)=\Delta(P)_{\tilde T}$, where $P_T$ denotes the subcategory of $P$ fixed by $T$ and $\Delta(P)_{\tilde T}$ denotes the subcomplex of $\Delta(P)$ fixed by $\tilde T$. \end{prop} \nin{\bf Proof. } Obviously, $\Delta(P_T)\subseteq\Delta(P)_{\tilde T}$. On the other hand, if for some $x\in\Delta(P)$ we have $\tilde T(x)=x$, then the minimal simplex $\sigma$, which contains $x$, is fixed as a set and, since the order of simplices is preserved by $T$, $\sigma$ is fixed by $T$ pointwise, thus $x\in\Delta(P_T)$. $\square$\vskip4pt The~class of loopfree categories can be seen as the closure of the class of posets under the operation of taking the quotient by a horizontal group action. More precisely, we have: \begin{prop} The~quotient of a loopfree category by a~horizontal action is again a loopfree category. In particular, the quotient of a poset by a~horizontal action is a~loopfree category. \end{prop} \nin{\bf Proof. } Let $K$ be a loopfree category and assume $G$ acts on $K$ horizontally. First observe that ${\cal M}_{K/G}([x])=\{\mbox{id}_{[x]}\}$. Because if $m\in{\cal M}_{K/G}([x])$, then there exist $x_1,x_2\in{\cal O}(K)$, $\tilde m\in{\cal M}_K(x_1,x_2)$, such that $[x_1]=[x_2]$, $[\tilde m]=m$. Then $g x_1=x_2$ for some $g\in G$, hence, since $g$ is a horizontal functor, $x_1=x_2$ and since $K$ is loopfree we get $\tilde m=\mbox{id}_{x_1}$. Let us show that for $[x]\neq[y]$ at most one of the sets $M_{K/G}([x],[y])$ and $M_{K/G}([y],[x])$ is nonempty. Assume the contrary and pick $m_1\in M_{K/G}([x],[y])$, $m_2\in M_{K/G}([y],[x])$. Then there exist $x_1,x_2,y_1,y_2\in{\cal O}(K)$, $\tilde m_1\in{\cal M}_K(x_1,y_1)$, $\tilde m_2\in{\cal M}_K(y_2,x_2)$ such that $[x_1]=[x_2]=[x]$, $[y_1]=[y_2]=[y]$, $[\tilde m_1]=[m_1]$, $[\tilde m_2]=[m_2]$. Choose $g\in G$ such that $g y_1=y_2$. Then $[g x_1]=[x_2]=[x]$ and we have $g \tilde m_1\in{\cal M}_K(g x_1,y_2)$, so $\tilde m_2\circ g \tilde m_1\in {\cal M}_K(g x_1,x_2)$. Since $K$ is loopfree we conclude that $g x_1=x_2$, but then both ${\cal M}_K(x_2,y_2)$ and ${\cal M}_K(y_2,x_2)$ are nonempty, which contradicts to the fact that $K$ is loopfree. $\square$\vskip4pt Next, we shall state a~condition under which the quotient of a~loopfree category is a~poset. \begin{prop} \label{5sreg} Let $K$ be a loopfree category and let $G$ act on $K$. The~following two assertions are equivalent: \begin{enumerate} \item[(1)] {\bf Condition (SR).} If $x,y\in{\cal M}(K)$, $\partial\,^\bullet x=\partial\,^\bullet y$ and $G\partial\,_\bullet x=G\partial\,_\bullet y$, then $Gx=Gy$. \item[(2)] $G$ acts regularly on $K$ and $K/G$ is a poset. \end{enumerate} \end{prop} \nin{\bf Proof. } $(2)\Rightarrow (1)$. Follows immediately from the regularity of the action of $G$ and the fact that there must be only one morphism between $[\partial\,^\bullet x](=[\partial\,^\bullet y])$ and $[\partial\,_\bullet x](=[\partial\,_\bullet y])$. $$\epsffile{pic4.eps}$$ $$\mbox{Figure 5.4}$$ $(1)\Rightarrow (2)$. Obviously (SR) $\Rightarrow$ (R), hence the action of $G$ is regular. Furthermore, if $x,y\in{\cal M}(K)$ and there exist $g_1,g_2\in G$ such that $g_1 \partial\,^\bullet x=\partial\,^\bullet y$ and $g_2 \partial\,_\bullet x=\partial\,_\bullet y$, then we can replace $x$ by $g_1 x$ and reduce the situation to the one described in Condition (SR), namely that $\partial\,^\bullet x=\partial\,^\bullet y$. Applying Condition (SR) and acting with $g_1^{-1}$ yields the result. $\square$ \vskip4pt When $K$ is a poset, Condition (SR) can be stated in simpler terms. \vskip4pt \noindent {\bf Condition (SRP).} If $a,b,c\in K$, such that $a\geq b$, $a\geq c$ and there exists $g\in G$ such that $g(b)=c$, then there exists $\tilde g\in G$ such that $\tilde g(a)=a$ and $\tilde g(b)=c$. $$\epsffile{pic5.eps}$$ $$\mbox{Figure 5.5}$$ That is, for any $a,b\in P$, such that $a\geq b$, we require that the stabilizor of $a$ acts transitively on $Gb$. \begin{prop} Let $P$ be a poset and assume $G$ acts on $P$. The~action of $G$ on $P$ induces an action on the barycentric subdivision ${\textrm{Bd}\,} P$ (the poset of all chains of $P$ ordered by inclusion). This action satisfies Condition (S), hence it is regular and $\Delta({\textrm{Bd}\,} P)/G\cong\Delta(({\textrm{Bd}\,} P)/G)$. Moreover, if the action of $G$ on $P$ is horizontal, then $({\textrm{Bd}\,} P)/G$ is a poset. \end{prop} \nin{\bf Proof. } Let us choose chains $b$, $c$ and $a=(a_1>\dots>a_t)$, such that $a\geq b$ and $a\geq c$. Then $b=(a_{i_1}>\dots>a_{i_l})$, $c=(a_{j_1}>\dots>a_{j_l})$. Assume also that there exists $g\in G$ such that $g(a_{i_s})=a_{j_s}$ for $s\in[l]$. If $g$ fixes $a$ then it fixes every $a_i$, $i\in[t]$, hence $b=c$ and Condition (S) follows. If, moreover, the action of $G$ is horizontal, then again $a_{i_s}=a_{j_s}$, for $s\in[l]$, hence $b=c$ and Condition (SRP) follows. $\square$\vskip4pt \chapter*{Preface} \addcontentsline{toc}{chapter}{Preface} \vskip10pt This thesis is thought to be reflective of the transformation, mistily rendered in the citation on the previous page, which has taken place in me since I have handed in my Ph.D.\ thesis in the spring of 1996. If this transformation, not unnatural taking into account that I was 23 at the time, was to the better or to the worse, I leave for the reader to judge. \vskip10pt \noindent Now, before indulging into mathematics, I would like to thank my coauthors Eva-Maria Feichtner and Eric Babson, without whom the second, resp.\ the fifth, chapters would not exist. \vskip10pt \noindent In recent years I was supported in various forms by a number of research and educational establishments. I therefore express my graditude to: \vskip2pt \noindent University of Bern, \noindent Swiss National Science Foundation, \noindent Forschungsinstitut f\"ur Mathematik at ETH Z\"urich, \noindent The Institute for Advanced Study at Princeton, \noindent Massachusetts Institute of Technology at Cambridge, \noindent Mathematical Science Research Institute at Berkeley, \noindent Royal Institute of Technology at Stockholm, \noindent Swedish Natural Science Research Council. \vskip10pt \noindent Finally, I would like to mention again Eva-Maria Feichtner without whose unrequiring support this work would have been a mere Gedankenspiel. \vskip3pt \noindent I dedicate this thesis to her. \vskip80pt \noindent {\it Bern} \hskip255pt {\sc Dmitry N.\ Kozlov} \noindent {\it December 14, 2001} \chapter*{Introduction\markboth{Introduction}{Introduction}} \addcontentsline{toc}{chapter}{Introduction} \vskip6pt \begin{center} {\bf Topological Combinatorics - a chapter of Discrete Mathematics} \end{center} \vskip4pt \noindent For the sake of brevity, the theme of this thesis could be defined as {\bf discrete mathematics} or {\bf combinatorics} and its interactions with pure mathematics, in particular with algebraic topology. This subject has long gone either without a~name, or subordinated under other titles, which did not fully reflect its philosophy. Lately, the term {\bf topological combinatorics} has been used with rising frequency, so, for the matter of presentation, allow me to extend the use of this name to the present exposition. Despite some striking applications, see e.g.\ ~\cite{Lo78}, and the subsequent explosive development, topological combinatorics is still a~comparatively young subject. Its main problems ({\it Fragestellungen}), as well as its methods are still in their forming stages, although a~lot of consolidating work has been undertaken in recent years, see e.g.\ ~\cite{Bj95}. Let me highlight here two major venues of research. \mbox{ } \begin{center} {\bf Combinatorial structures in topology and geometry} \end{center} \vskip4pt \noindent Succinctly, the philosophy of the subject might be described as follows. The first step consists of singling out a~structure, with combinatorial flavor, occurring naturally in topology or geometry. Then, one tries to find a suitable, purely combinatorial definition of this structure. This usually results in combining known, well-studied objects in a~way, which would be neither very natural, nor apparent from the original context, in which these objects were defined. Once this is done, there are two major possibilities to continue. The~first option is to develop better structural understanding of the new combinatorial theory, which has just been distiled. That understanding may then be used to refine our knowledge of the original object. The second option is to seek out the occurrences of this combinatorial object in other topological or geometric situations, where the connections to the original situation are not clear, and at the moment the link is only provided on the abstract combinatorial level. Hopefully, one can then use this new combinatorial connection to better understand the topological objects in question. A few instances of the developments in this spirit are: matroid bundles and combinatorial differential manifolds of MacPherson, \cite{An99,GeM92,Mac93}, complex matroids, \cite{Zi93}, or, even such a~general concept as that of oriented matroids, \cite{BLSWZ}. \vskip4pt The first part of this thesis can be classified into this category of research. In the first chapter we study a~circle of problems arising from topology, with some specific instances appearing in, among others, singularity theory. The goal is to describe in a~functorial way the combinatorial structure of canonical stratifications of $n$-fold symmetric smash products. By canonical stratifications we mean those implied by point coincidences, with strata being indexed by number partitions of~$n$. The approach, which we choose to achieve that objective, is to introduce a~new combinatorial object, a~new category, which we call {\it resonance category}. The objects of this category are not number partitions themselves, but rather certain equivalence classes consisting of number partitions, which we call $n$-cuts. The introduction of these equivalence classes, in a~characteristic for the ideology of this subject way, is motivated by the topological fact that strata, which are indexed by number partitions belonging to the same $n$-cut, are homeomorphic. Furthermore, substrata of a stratum can be glued so as to obtain a~new stratum, and also various strata can be included into each other. To~reflect these topological maps combinatorially, we define the morphisms of the resonance category to be compositions of two types of elementary morphisms, called ``gluings'' and ``inclusions,'' which are defined purely in terms of $n$-cuts. The original problem of studying stratifications can be viewed as studying certain functors from the resonance category to the category of pointed topological spaces, $\mbox{\bf Top}^*$. This way we split the problem into two parts: the combinatorial one, encoded by the resonance category, and the topological one, encoded by such a~functor. After that, we go on with the structural study of the resonance category. Another, somewhat technical notion, which we need to introduce is that of {\it a~relative resonance}. The main idea is then to study direct product decomposition properties of relative resonances. The general theory is rather rich, and we have succeeded to understand just a~small part of it, but already that was sufficient to perform various concrete computations. For example, for a~certain class of resonances we can produce combinatorial models to explicitly compute algebraic invariants of the corresponding strata. More specifically, we prove that these strata are homotopy equivalent to wedges of spheres. Then, we go ahead to construct weighted directed graphs, whose longest directed paths enumerate these spheres, and such that the total weights of the longest paths give the dimensions of the spheres in question. This provides an~explicit algorithm to compute the homotopy type of strata for this class of resonances. \vskip4pt In the second chapter we turn to a~somewhat different topic. In~\cite{DP95}, De~Concini \& Procesi have constructed certain models for complements of subspace arrangements. The idea of their construction is to turn any given subspace arrangement into a~divisor with normal crossings, by means of a~sequence of blowups. To use blowups to resolve singularities is a~standard idea of algebraic geometry, yet the performance of such a~resolution in specific cases often leads to intricate constructions. One aspect of the De~Concini \& Procesi (and of the earlier Fulton \& MacPherson, \cite{FM94}) construction, which became apparent early on, was the fact that the combinatorics of the situation was far from being trivial or standard material. To deal with the mounting technicalities, De~Concini \& Procesi generalized the notions of building sets and nested sets, originally coined by Fulton \& MacPherson in the more specific context of combinatorics of configuration spaces, to cover the case of subspace arrangements. In~\cite{DP95} these notions were still defined in geometric terms, relying on the subspace arrangements themselves, rather than on their intersection semilattices. In the research, presented in the second chapter, we generalize the notions of building sets and nested sets to the purely combinatorial context of semilattices. Moreover, we introduce a~new combinatorial construction, which we call a~``combinatorial blowup.'' These three notions can then be combined to yield a lattice-theoretic theorem (Theorem~\ref{thm_main}), which parallels in statement one of the main theorems of~\cite{DP95}. As a~result, we obtain a~consistent, nontrivial combinatorial theory, which canonically depends on the underlying combinatorial data, and not on the geometric structures which it encodes, with Theorem~\ref{thm_main} as the main structural result. Now, turning back to geometry, we are able to step-by-step trace the construction of De~Concini \& Procesi, by using the concepts which we introduced. Furthermore, following the ideas outlined in the beginning of the discussion, we go on to describe how these combinatorial notions also appear in the context of toric varieties. \vskip6pt \begin{center}{\bf Combinatorially defined algebraic invariants of topological spaces} \end{center} \vskip4pt \noindent A~somewhat different direction of research is to work on determining explicitly algebraic invariants of topological spaces given by some combinatorial construction. This area has many aspects. One is to derive general formulae for a~class of objects. An example of that is provided by the Goresky \& MacPherson formula, see Appendix~C. Among further instances one could mention the Orlik-Solomon algebra, which describes the cohomology algebra of the complement of a~complex hyperplane arrangement, \cite{OS80}, or the Yuzvinsky basis for the cohomology algebra of De~Concini-Procesi compactifications, \cite{Yu97}. \vskip4pt Another example is described in the third chapter, where we give a~combinatorial formula to compute the rational Betti numbers of spaces of complex monic polynomials with roots of fixed multiplicities. These spaces can be also viewed as strata in the natural stratification of the $n$-fold symmetric smash product of~$S^2$, which were previously described in this exposition. To~give this combinatorial description we need to introduce cell complexes, which we call $X_{\lambda,\mu}$, which are parameterized by pairs of number partitions $(\lambda,\mu)$, such that $\lambda\vdash\mu$. The study of these complexes yields a~variety of applications. To~start with, we prove collapsibility of the regular CW complex $\Delta(\Pi_n)/{\cal S}_n$. The partition lattice $\Pi_n$, and its order complex $\Delta(\Pi_n)$, are of fundamental importance for encoding the combinatorics of configuration spaces, and for computing cohomology of the colored braid group, which motivates the appearance of the complex $\Delta(\Pi_n)/{\cal S}_n$ as a~basic object of study. We also obtain a~new, conceptual proof of the Arnol'd Finiteness Theorem, which boils down to a~combinatorial fact about marked forests. Furthermore, we are able to furnish a~counterexample to a~conjecture of Sundaram \& Welker concerning the multiplicity of the trivial character in certain ${\cal S}_n$-representations, as well as to compute this multiplicity for ${\cal S}_n$-representations previously studied by R.\ Stanley, \cite{St82}, and P.\ Hanlon, \cite{Ha83}. \vskip4pt In the fourth chapter we turn to a~related aspect of this area, namely, the combinatorial analysis of simplicial complexes with explicit combinatorial description of the simplices. Such study facilitates the richness of the theory, and provides a~reasonable testing field for the computational methods of topological combinatorics. Some instances of such investigations are various forms of lexicographic shellability, developed by Bj\"orner et al.\ \cite{Bj80,Bj95,BWa83,BWa96,BWa97,Ko97}, and the study of topological properties of complexes of not $k$-connected graphs, with implications in knot theory, \cite{BBLSW}. In our case, the focus is on studying complexes of directed subforests of a~fixed graph, a~problem suggested to us by R.\ Stanley, \cite{St97}. For the complete graph $G_n$ we are able to prove that $\Delta(G_n)$ is shellable, implying that it is homotopy equivalent to a~wedge of $(n-1)^{n-1}$ spheres of dimension $n-2$. Furthermore, we develop machinery to analyze topological properties of $\Delta(G)$ for a~class of directed graphs, called essentially trees, and perform complete computations in the special cases of cycles and double directed strings. In the last section of the fourth chapter, we study the quotient complexes $\Delta(G_n)/{\cal S}_n$. These turn out to be rather complex and, after giving the combinatorial description of the simplicial structure of $\Delta(G_n)/{\cal S}_n$ in terms of marked forests, we perform their analysis by means of the apparatus of spectral sequences. In particular, we prove that the homology groups of $\Delta(G_n)/{\cal S}_n$ with integer coefficients are {\it not} torsion-free in general. \vskip4pt Finally, in the fifth chapter, we take a~look at various quotient constructions which appear in the context of group actions on posets. The quotients appear in the third and the fourth chapters, so a~short and abstract study of these constructions is commanded by the natural yearning for completeness. Since an~action of a~group $G$ is simply a~functor from the one-object-category associated to $G$, it is natural to ask the quotient to be the colimit of this functor. The three main questions of study, which we concentrate on, are: when are the morphisms of $P/G$ given by the $G$-orbits of the morphisms of $P$ (a~property called ``regularity''), when does taking the quotient commute with Quillen's nerve functor $\Delta$, and, finally, which classes of categories we generally may get as quotients. We are able to give simple combinatorial conditions on the group actions, which are equivalent to the desired properties. These conditions are, in some sense, local, which implies that they are frequently verifiable in specific situations, boding well for them being useful in the future. \chapter*{Summary\markboth{Summary}{}} \addcontentsline{toc}{chapter}{Summary} This thesis opens with an~introductory discussion, where the reader is gently led to the world of topological combinatorics, and, where the results of this Habilitationsschrift are portrayed against the backdrop of the broader philosophy of the subject. That introduction is followed by 5 chapters, where the main body of research is presented, and 4 appendices, where various standard tools and notations, which we use throughout the text, are collected. \vskip10pt The main purpose of the first chapter is to introduce a~new category, which we call resonance category, whose combinatorics reflects that of canonical stratifications of $n$-fold symmetric smash products. The study of the stratifications can then be abstracted to the study of functors, which we name resonance functors, satisfying a~certain set of axioms. One frequently studied stratification is that of the set of all degree~$n$ polynomials, defined by fixing the allowed multiplicities of roots. We describe how our abstract combinatorial framework helps to yield new information on the homology groups of the strata. \vskip10pt In the second chapter we introduce notions of combinatorial blowups, building sets, and nested sets for arbitrary meet-semilattices. This gives a~common abstract framework for the incidence combinatorics occurring in the context of De~Concini-Procesi models of subspace arrangements and resolutions of singularities in toric varieties. Our main theorem states that a sequence of combinatorial blowups, prescribed by a building set in linear extension compatible order, gives the face poset of the corresponding simplicial complex of nested sets. As applications we trace the incidence combinatorics through every step of the De~Concini-Procesi model construction, and we introduce the notions of building sets and nested sets to the context of toric varieties. \vskip10pt In the third chapter we study rational homology groups of one-point compactifications of spaces of complex monic polynomials with multiple roots. These spaces are indexed by number partitions. A~standard reformulation in terms of quotients of orbit arrangements reduces the problem to studying certain triangulated spaces~$X_{\lambda,\mu}$. \newpage We present a~combinatorial description of the cell structure of~$X_{\lambda,\mu}$ using the language of marked forests. This allows us to perform a~complete combinatorial analysis of the topological properties of $X_{\lambda,\mu}$ for several specific cases. As applications we prove that $\Delta(\Pi_n)/{\cal S}_n$ is collapsible, we obtain a~new proof of a~theorem of Arnol'd, and we find a~counterexample to a~conjecture of Sundaram and Welker, along with a~few other smaller results. \vskip10pt To every directed graph G one can associate a~complex $\Delta(G)$ consisting of directed subforests. This construction, suggested to us by R.~Stanley, is especially important in the~case of a~complete double directed graph $G_n$, where it leads to studying certain representations of the~symmetric group and relates (via Stanley-Reisner correspondence) to an~interesting quotient ring. In the fourth chapter we study complexes $\Delta(G)$ and associated quotient complexes. One of our results states that $\Delta(G_n)$ is shellable, in particular Cohen-Ma\-cau\-lay, which can be further translated to say that the~Stanley-Reisner ring of $\Delta(G_n)$ is Cohen-Macaulay. Besides that, by computing the~ho\-mo\-lo\-gy groups of $\Delta(G)$ for the~cases when $G$ is essentially a~tree, and when $G$ is a~double directed cycle, we touch upon the~general question of the~interaction of combinatorial properties of a~graph and topological properties of the~associated complex. We then continue with the study of the ${\cal S}_n$-quotient of the complex of directed forests, denoted $\Delta(G_n)/{\cal S}_n$. It is a~simplicial complex whose cell structure is defined combinatorially. We make use of the machinery of spectral sequences to analyze these complexes. In~particular, we prove that the integral homology groups of $\Delta(G_n)/{\cal S}_n$ may have torsion. \vskip10pt In the fifth chapter we study quotients of posets by group actions. In order to define the quotient correctly we enlarge the considered class of categories from posets to loopfree categories: categories without nontrivial automorphisms and in\-ver\-ses. We view group actions as certain functors and define the quotients as co\-li\-mits of these functors. The advantage of this definition over studying the quo\-ti\-ent poset (which in our language is the colimit in the poset category) is that the realization of the quotient loopfree category is more often homeomorphic to the quotient of the realization of the original poset. We give conditions under which the quotient commutes with the nerve functor, as well as conditions which guarantee that the quotient is again a~poset. \chapter*{Zusammenfassung} \sloppy \addcontentsline{toc}{chapter}{Zusammenfassung} Wir beginnen mit einer einleitenden Diskussion, die den Leser in die Welt der topologischen Kombinatorik einf\"uhrt und die Ergebnisse dieser Habilitations\-schrift vor den breiteren Hintergrund der das Gebiet bestimmenden Gedankenwelt stellt. Dieser Einleitung folgen f\"unf Kapitel, in denen wir die Hauptergebnisse ent\-wickeln, und vier Appendizes, in denen wir Standardwerkzeuge und Notationen zusammenstellen, wie sie im Laufe der Arbeit verwendet werden. \vskip10pt Ziel des ersten Kapitels ist es, eine neue Kategorie, die Resonanzkategorie, einzuf\"uhren, deren Kombinatorik diejenige kanonischer Stratifizierungen $n$-fa\-cher symmetrischer Smash-Produkte widerspiegelt. Das Studium dieser Stra\-ti\-fi\-zierungen kann so zum Studium sogenannter Resonanzfunktoren abstrahiert werden, die den Axiomen eines bestimmten Axiomensystems gen\"ugen. Eine h\"aufig studierte Stratifikation ist die der Menge von Polynomen festen Grades~$n$ nach den unter den Wurzeln auftretenden Vielfachheiten. Wir zeigen, wie wir dank unseres abstrakten kombinatorischen Systems zu neuen Informationen \"uber die Homologie der Strata gelangen. \vskip10pt Im zweiten Kapitel f\"uhren wir f\"ur beliebige Durchschnittshalbverb\"ande die Begriffe kombinatorischer Aufblasungen, Bausatzmengen und vernetzter Mengen ein. Dies liefert einen gemeinsamen abstrakten Rahmen f\"ur Inzidenzen, wie sie einerseits im Zusammenhang mit De Concini-Procesi Modellen von Arrangements, andererseits bei Aufl\"osungen von Singularit\"aten in torischen Variet\"aten eine Rolle spielen. Unser Hauptsatz besagt, dass eine Sequenz kombinatorischer Aufblasungen, bestimmt durch eine mittels linearer Erweiterung geordnete Bau\-satz\-menge, auf die Seitenhalbordnung des zugeh\"origen Simplizialkomplexes vernetzter Mengen f\"uhrt. Als Anwendung verfolgen wir die Inzidenzen in der schritt\-wei\-sen Modellkonstruktion nach De Concini und Procesi und f\"uhren die Begriffe von Bausatzmengen und vernetzten Mengen in den Zusammenhang torischer Variet\"aten ein. \vskip10pt Im dritten Kapitel studieren wir rationale Homologiegruppen der 1-Punkt-Kompaktifizierungen von R\"aumen komplexer normierter Polynome mit mehr\-fachen Nullstellen. Eine Umformulierung f\"ur Quotienten von Bahnenarrangements reduziert das Problem auf das Studium gewisser triangulierter R\"aume $X_{\lambda,\mu}$. \newpage Wir geben eine kombinatorische Beschreibung der Zellstruktur von $X_{\lambda,\mu}$, indem wir uns der Terminologie sogenannter markierter B\"aume bedienen. Dies erlaubt uns eine vollst\"andige Analyse topologischer Eigenschaften von $X_{\lambda,\mu}$ in verschiedenen Spezialf\"allen. Als Anwendung beweisen wir einerseits die Kollabierbarkeit von $\Delta(\Pi_n)/{\cal S}_n$, andererseits erhalten wir, neben anderen Ergebnissen, einen neuen Beweis f\"ur einen Satz von Arnol'd und geben ein Gegenbeispiel f\"ur eine Vermutung von Sundaram und Welker. \vskip10pt Jedem gerichteten Graphen~$G$ kann man einen Komplex~$\Delta(G)$ bestehend aus gerichteten Teilw\"aldern zuordnen. Diese Konstruktion, die auf einen Vorschlag von R.~Stanley zur\"uckgeht, ist besonders wichtig f\"ur vollst\"andige, doppelt ge\-rich\-te\-te Graphen $G_n$. F\"ur solche f\"uhrt sie auf das Studium bestimmter Darstellungen der symmetrischen Gruppe und entspricht via Stanley-Reisner Korrespondenz interessanten Quotientenringen. Im vierten Kapitel studieren wir die Komplexe~$\Delta(G)$ und zugeh\"orige Quotientenkomplexe. Eines unserer Resultate besagt, dass $\Delta(G)$ sch\"albar ist, damit insbesondere Cohen-Macaulay, was wiederum \"ubersetzt werden kann in die Tatsache, dass der Stanley-Reisner-Ring von~$\Delta(G)$ die Cohen-Macaulay Eigenschaft besitzt. Dar\"uberhinaus r\"uhren wir durch Homologieberechnungen von~$\Delta(G)$ in F\"allen, in denen $G$ im wesentlichen ein Baum oder ein doppelt gerichteter Kreis ist, an die allgemeinere Frage des Zusammenspiels von kombinatorischen Eigenschaften des Graphen und topologischen Eigenschaften des zugeh\"origen Komplexes. Darauffolgend studieren wir den ${\cal S}_n$-Quotienten des Komplexes gerichteter W\"alder, $\Delta(G_n)/{\cal S}_n$. Es handelt sich um einen Simplizialkomplex mit kombinatorisch definierter Zellstruktur. Wir benutzen Spektralsequenztechniken, um diese Komplexe zu studieren. Insbesondere zeigen wir, dass Torsion auftreten kann in der ganzzahligen Homologie von $\Delta(G_n)/{\cal S}_n$. \vskip10pt Im f\"unften Kapitel studieren wir Quotienten von Halbordnungen nach Gruppenoperationen. Um Quotienten sauber definieren zu k\"onnen, erweitern wir die betrachtete Klasse von Kategorien von Halbordnungen zu schleifenfreien Ka\-te\-gorien: Kategorien mit nicht-trivialen Automorphismen und Inversen. Wir fassen Gruppenaktionen als Funktoren auf und definieren Quotienten als Ko\-limiten dieser Funktoren. Der Vorteil dieser Definition gegen\"uber Quotienten\-halb\-ord\-nungen (in unserer Sprache die Kolimiten in der Kategorie der Halbordnungen) ist, dass die Realisierungen von Quotienten schleifenfreier Kategorien h\"aufiger hom\"oomorph sind zu den Quotienten von Realisierungen der urspr\"unglichen Halb\-ordnung. Wir geben Bedingungen daf\"ur an, dass die Quotientenbildung mit dem Nerv-Funktor vertauschbar ist, und formulieren dar\"uberhinaus Bedingungen, die garantieren, dass der Quotient wiederum eine Halbordnung ist.

https://arxiv.org/abs/1306.6278

Playing cooperatively with possibly treacherous partner

We investigate an alternative concept of Nash equilibrium, m-equilibrium, which slightly resembles Harsanyi-Selten risk dominant equilibrium although it is a different notion. M-equilibria provide nontrivial solutions of normal form games as shown by comparison of the Prisoner's Dilemma with the Traveler's Dilemma. They are also resistant on the deep iterated elimination of dominated strategies.

\section{Introduction} The games of interest here are \emph{two person general sum games in normal form}. Concerning notation and definitions the reader is asked to consult the next Section. First one has to provide additional information about the game. The main assumption, usually made implicitly, is that the players can communicate with each other. This remedies the coordination problem quoted below (see \cite{CommComplex} for complexity issues). \begin{prz}[Coordination]\label{Coordination} Let \(S_1=S_2=\{1,2,3\}\), \[G= \left[\begin{array}{ccc} {[2,2]} & {[0,0]} & {[0,0]} \\ {[0,0]} & {[1,1]} & {[0,0]} \\ {[0,0]} & {[0,0]} & {[2,2]} \\ \end{array}\right].\] Among three equilibria \((1,1)\), \((2,2)\) and \((3,3)\) only two are pleasant (as Pareto dominant), namely \((1,1)\) and \((3,3)\). If players cannot communicate, then they have to use randomization (e.g., coin flipping). The Bernoulli scheme would let them synchronize choices with small probability of failure (cf. Theorem 11.3 in \cite[chap.11.4]{Rendezvous}). \end{prz} Thus we see that some amount of communication (say pre-play/cheap talk) and a sort of cooperation cannot be dismissed even in the case of such competitive/``selfish" notion like the Nash equilibrium (e.g., \cite{Aumann, CommCoord, NashNeedsCooper}). The classic stag hunt game exploits another issue of miscoordination -- the Wald criterion of worst possible scenario. In the vein of a stag hunt's strategic security consider \begin{prz}\label{HighThreat} Let \(S_1=S_2=\{1,2\}\), \[G=\left[\begin{array}{cc} {[4,4]} & {[1,4]} \\ {[4,1]} & {[3,3]} \\ \end{array}\right].\] The pairs \((1,1)\) and \((2,2)\) are Nash equilibria with \((1,1)\) being Pareto dominant. Nevertheless one cannot guarantee that previously agreed among players equilibrium \((1,1)\) would be realized in practice. If the player is confident in fair play of his partner, he might switch strategy without any loss of income. The pair of strategies \((2,2)\) is threat-safe although yields smaller payoffs than \((1,1)\). \end{prz} Players make their final decisions independently of others. Therefore communication provides only weak cooperation (\cite{FundEquilibrium}). No one can force fair play, even if it is profitable for all (free rider's problem). Largely discussed traveler's dilemma (seen sometimes as the extension of prisoner's dilemma) underlines anomalous behavior in widely accepted procedure called an iterative elimination of dominated strategies (\cite{Basu, BasuExperim, Spanish, Gintis, HalpernPass}). Unlike the original formulation our assumes communication between players. \begin{prz}[Traveler's dilemma]\label{traveler} Let \(S_1=S_2=\{2,3,\ldots,100\}\), \(P_1(x,y)=P_2(y,x)= \min(x,y)+ 2\cdot \operatorname{sign}(y-x)\) for \(x\in S_1\), \(y\in S_2\), \(G=(S_1,S_2; P_1,P_2)\). Then \((2,2)\) is the only Nash equilibrium of \(G\). It arises through the elimination of dominated strategies, although most strategy pairs Pareto dominate it. Observe that the players could choose a pair of strategies which yields much higher payoffs than \((2,2)\). Moreover, the player can still play very profitably after his partner betrayed and switched strategy to get higher payoffs. If more than \(4\%\) partners play ``moderately" (at least \(54\)), then we can expect higher gain from playing ``dummy" \(100\) than from playing ``wise" \(2\). If more than \(10\%\) partners play ``high" (at least \(90\)), then playing \(100\) we can expect over \(400\%\) income of that which we could earn playing \(2\). \end{prz} Our point here is that one should calculate secure gains incorporating possible threats from his partner. This allows sometimes for much higher payoffs than those arising in the Nash equilibrium. That was the main motivation for introducing m-equilibria as we do in Section~\ref{sec:m-equilibrium}. We silently assume that the games under consideration are \emph{not repeated} and \emph{one-shot}; see also the discussion around mixed equilibria in Examples \ref{hidecoin} and \ref{pennies}. We understand the payoff to be NTU (nontransferable utility); that some ``transfers" are still possible ensures us Example \ref{CournotDuopoly} in Section \ref{sec:motivexamples} and informal discussion of fair choice among equilibria in Section \ref{sec:equilibriumselection}. For standard notions and theorems of game theory we refer to the textbook \cite{Strategy} (comp. \cite{StratEquilibrium, AubinOpt}). Throughout the paper the language of multivalued (or set-valued) analysis shall be utilized in several places (consult \cite{HandbookMulti, Beer, AubinOpt}). \section{Notation and definitions} Let \(\Gamma= (\Sigma_1, \Sigma_2; P_1,P_2:\Sigma_1\times\Sigma_2\to{\mathbb R})\) be a two person game (in normal form). \(\Sigma_i\) is the \textit{set of strategies} and \(P_i\) is the \textit{payoff function} of the \(i\)-th player. An accent will be put further on the case of a \textit{finite game} \(G=(S_1,S_2; W_1,W_2: S_1\times S_2 \to{\mathbb R})\), i.e., the game with the finite strategy sets \(S_1,S_2\), and its \textit{mixed extension} \(\Delta(G)= (\Delta(S_1),\Delta(S_2); EW_1,EW_2: \Delta(S_1)\times\Delta(S_2)\to{\mathbb R})\). \(\Delta(S)\) stands for the standard simplex of probabilistic measures (\textit{mixed strategies}) spanned on the finite set \(S\) of (\textit{pure}) strategies; \(S\subset\Delta(S)\) due to the identification via Dirac measures: \(S\ni x \mapsto \delta_{x}\in\Delta(S)\). The \textit{expected payoffs} are given by \[EW_i(\rho_1,\rho_2)=\sum_{(x,y)\in S_1\times S_2} {\rho}_1(x)\cdot W_i(x,y)\cdot {\rho}_2(y)\] for \({\rho}_i\in\Delta(S_i)\), \(i=1,2\). Finite games shall appear in the examples as bimatrices \(\left[\begin{array}{c} [W_1(x,y),W_2(x,y)]_{(x,y)\in S_1\times S_2} \end{array}\right]\), \(S_1\ni x\) -- the number of the row, \(S_2\ni y\) -- the number of the column. It is customary to employ the convention \(((\sigma_{i},\sigma_{-i})) = (\sigma_1,\sigma_2) \in\Sigma_{1}\times\Sigma_{2}\), \(\Sigma_{-i}=\Sigma_{3-i}\) and \(P_{-i}=P_{3-i}\) which emphasizes the role of the \(i\)-th player (\(i=1,2\)); the same applies to \(((x,y))\in {S}_{1}\times {S}_{2}\), \(((\pi,y))\in\Delta(S_1)\times\Delta(S_2)\), \(W_{-i}\) etc. For technical reasons we shall tacitly assume payoff functions \(P_i\) to be bounded from below in the co-player's variable, i.e., \(\inf_{\sigma_{-i}\in\Sigma_{-i}}\, P_i((\sigma_i,\sigma_{-i})) >-\infty\) for \(\sigma_{i}\in\Sigma_{i}\), \(i=1,2\). The continuity of \(P_i\) is not assumed and, in the case of finite strategy sets \(S_i\), of no use. The game \(\Gamma\) is said to be \textit{strictly competitive}, if \[\forall_{(\sigma_1,\sigma_2),(\sigma_1',\sigma_2')\in\Sigma_1\times\Sigma_2}\;\; P_1(\sigma_1',\sigma_2')>P_1(\sigma_1,\sigma_2) {\Leftrightarrow} P_2(\sigma_1',\sigma_2')<P_2(\sigma_1,\sigma_2).\] Let us note that strictly competitive games are essentially zero-sum games (see \cite{Adler}). We call \(\Gamma\) \textit{quantitatively symmetric}, if \(P_1(\sigma_1,\sigma_2)=P_2(\sigma_2,\sigma_1)\) for \(\sigma_1,\sigma_2\in\Sigma_1=\Sigma_2\). \begin{stw} If \(G\) is a finite strictly competitive game, resp. quantitatively symmetric game, then its mixed extension \(\Delta(G)\) is of the same character. \end{stw} \begin{df} A pair of strategies \((\sigma_1^{*},\sigma_2^{*})\in\Sigma_{1}\times\Sigma_{2}\) is \begin{itemize} \item \textit{Pareto optimum}, \(PO(\Gamma)\), if \[\neg \exists_{(\sigma_1,\sigma_2)\in\Sigma_{1}\times\Sigma_{2}} \forall_{i=1,2}\;\; P_{i}(\sigma_1,\sigma_2) > P_{i}(\sigma_1^{*},\sigma_2^{*}),\] \item \textit{strong Pareto optimum}, \(SPO(\Gamma)\), if \[\neg \exists_{(\sigma_1,\sigma_2)\in\Sigma_{1}\times\Sigma_{2}}\; \left[\;\forall_{i=1,2}\;\; P_{i}(\sigma_1,\sigma_2) \geq P_{i}(\sigma_1^{*},\sigma_2^{*}) \;\wedge\; \exists_{i=1,2}\;\; P_{i}(\sigma_1,\sigma_2) > P_{i}(\sigma_1^{*},\sigma_2^{*})\;\right],\] \item \textit{Wald solution}, \(W(\Gamma)\), if \[\forall_{i=1,2}\; \sigma_i^{*}\in\arg\max_{\sigma_{i}\in\Sigma_{i}}\; \min_{\sigma_{-i}\in\Sigma_{-i}}\; P_i((\sigma_i,\sigma_{-i})),\] \item \textit{Nash equilibrium}, \(NE(\Gamma)\), if \[\forall_{\sigma_1\in\Sigma_{1},\;\sigma_2\in\Sigma_{2}} \forall_{i=1,2}\;\; P_{i}((\sigma_{i},\sigma_{-i}^{*})) \leq P_{i}((\sigma_{i}^{*},\sigma_{-i}^{*})),\] \item \textit{strict Nash equilibrium}, \(SNE(\Gamma)\), if \[\forall_{\sigma_1\in\Sigma_{1}\setminus\{\sigma_1^{*}\}, \;\sigma_2\in\Sigma_{2}\setminus\{\sigma_2^{*}\}} \forall_{i=1,2}\;\; P_{i}((\sigma_{i},\sigma_{-i}^{*})) < P_{i}((\sigma_{i}^{*},\sigma_{-i}^{*})),\] \item \textit{semi-strict Nash equilibrium}, \(SSNE(\Gamma)\), if it is Nash equilibrium and \[\forall_{\sigma_1\in\Sigma_{1},\;\sigma_2\in\Sigma_{2}} \forall_{i=1,2}\;\; \left[\;P_{i}((\sigma_{i},\sigma_{-i}^{*})) = P_{i}((\sigma_{i}^{*},\sigma_{-i}^{*})) {\Rightarrow} P_{-i}((\sigma_{i},\sigma_{-i}^{*})) = P_{-i}((\sigma_{i}^{*},\sigma_{-i}^{*}))\;\right],\] \item \textit{weakly semi-strict Nash equilibrium}, \(WSSNE(\Gamma)\), if it is Nash equilibrium and \[\forall_{\sigma_1\in\Sigma_{1},\;\sigma_2\in\Sigma_{2}} \forall_{i=1,2}\;\; \left[\;P_{i}((\sigma_{i},\sigma_{-i}^{*})) = P_{i}((\sigma_{i}^{*},\sigma_{-i}^{*})) {\Rightarrow} P_{-i}((\sigma_{i},\sigma_{-i}^{*})) \geq P_{-i}((\sigma_{i}^{*},\sigma_{-i}^{*}))\;\right],\] \item \textit{coupled in wealth improvement}, \(CWI(\Gamma)\), if \[\forall_{\sigma_1\in\Sigma_{1},\;\sigma_2\in\Sigma_{2}} \forall_{i=1,2}\;\; \left[\;P_{i}((\sigma_{i},\sigma_{-i}^{*})) \geq P_{i}((\sigma_{i}^{*},\sigma_{-i}^{*})) {\Rightarrow} P_{-i}((\sigma_{i},\sigma_{-i}^{*})) \geq P_{-i}((\sigma_{i}^{*},\sigma_{-i}^{*}))\;\right].\] \end{itemize} \end{df} Remark that \(WSSNE(\Gamma) = NE(\Gamma) \cap CWI(\Gamma)\). The last three concepts (SSNE, WSSNE, and CWI) are provided by ourselves and their role shall be clear in view of further investigations. Loosely speaking (weakly) semi-strict equilibria retain most important features of strict equilibria, especially those associated with strategic uncertainty as shown in Example~\ref{HighThreat}. Remark also that the (weakly) semi-strict equilibrium is a concept different from the weakly strict equilibrium introduced in \cite{WeakE} and quasi strict equilibrium in \cite{QuasiE}. Finally the reader should be warned that the Wald solution becomes the \textit{maximin solution} if additionally the ``minimax identity" holds \[\forall_{i=1,2}\;\; P_i(\sigma_1^{*},\sigma_2^{*})= \max_{\sigma_{i}\in\Sigma_{i}}\; \min_{\sigma_{-i}\in\Sigma_{-i}}\; P_i((\sigma_i,\sigma_{-i})).\] We have the following relations \[SNE(\Gamma)\varsubsetneq SSNE(\Gamma)\varsubsetneq WSSNE(\Gamma)\varsubsetneq NE(\Gamma).\] The inclusions are immediate from the definitions. That they are strict illustrates: \begin{prz}[3-4-5 game] Let \(S_1=S_2=\{1,2,3,4\}\), \[G=\left[\begin{array}{cccc} {[3,3]} & {[0,0]} & {[0,0]} & {[0,0]} \\ {[0,0]} & {[4,4]} & {[0,0]} & {[4,4]} \\ {[0,0]} & {[0,0]} & {[3,3]} & {[5,3]} \\ {[0,0]} & {[4,4]} & {[3,5]} & {[5,5]} \\ \end{array}\right].\] Then \((4,4)\in NE(G)\setminus WSSNE(G)\), \((3,3)\in WSSNE(G)\setminus SSNE(G)\), \((2,2)\in SSNE(G)\setminus SNE(G)\), \((1,1)\in SNE(G)\). \end{prz} \begin{stw} The equilibria of a finite game \(G\) and its mixed extension \(\Delta(G)\) are related as follows: \begin{enumerate} \item \(NE(G)\subset NE(\Delta(G))\), \item \(SNE(G)\subset SNE(\Delta(G))\), \item \(SSNE(G)\subset SSNE(\Delta(G))\), \item \(WSSNE(G)\subset WSSNE(\Delta(G))\). \end{enumerate} \end{stw} \begin{dow} We only check the last invariance, the rest can be performed analogously. Let \(((x^{*},y^{*}))\in WSSNE(G)\) and suppose that some deviation from \(x^{*}\in S_i\) to a mixed strategy \(\pi\in \Delta(S_i)\) still gives equally good gain for the \(i\)-th player i.e. \(EW_i((x^{*},y^{*}))=EW_i((\pi,y^{*}))\). First note that since \(((x^{*},y^{*}))\) is a Nash equilibrium of \(G\), then \(W_i((x,y^{*}))\leq W_i((x^{*},y^{*}))\) for all \(x\in \operatorname{supp}\pi\). Now if \(W_i((\sprz{x},y^{*}))< W_i((x^{*},y^{*}))\) for some \(\sprz{x}\in \operatorname{supp}\pi\), then \begin{align*} EW_i((\pi,y^{*})) = \pi(\sprz{x})\cdot W_i((\sprz{x},y^{*})) + \sum_{\sprz{x}\neq x\in\operatorname{supp}\pi}\; \pi(x)\cdot W_i((x,y^{*})) \\ < \left(\pi(\sprz{x})+\sum_{\sprz{x}\neq x\in\operatorname{supp}\pi}\;\pi(x)\right) \cdot W_i((x^{*},y^{*})) = EW_i((x^{*},y^{*})). \end{align*} Therefore in fact we have \(W_i((x,y^{*}))=W_i((x^{*},y^{*}))\) for \(x\in\operatorname{supp}\pi\). Recalling that \(((x^{*},y^{*}))\in WSSNE(G)\) we obtain that \(W_{-i}((x,y^{*})) \geq W_{-i}((x^{*},y^{*}))\) for \(x\in\operatorname{supp}\pi\). Finally \begin{align*} EW_{-i}((\pi,y^{*})) = \sum_{x\in\operatorname{supp}\pi}\; \pi(x)\cdot W_{-i}((x,y^{*})) \\ \geq \left(\sum_{x}\; \pi(x)\right) \cdot W_{-i}((x^{*},y^{*})) = EW_{-i}((x^{*},y^{*})). \end{align*} \end{dow} \begin{tw}\label{antagSSNE} In the strictly competitive game \(\Gamma\) all equilibria are semi-strict, \(NE(\Gamma)=SSNE(\Gamma)\). \end{tw} \begin{dow} Let \((\sigma_1^{*},\sigma_2^{*})\in NE(\Gamma)\). Competitiveness implies that equal payoffs of one player correspond to equal payoffs of the other one. Therefore \[P_i((\sigma_i,\sigma_{-i}^{*}))=P_i((\sigma_i^{*},\sigma_{-i}^{*})) {\Rightarrow} P_{-i}((\sigma_i,\sigma_{-i}^{*}))=P_{-i}((\sigma_i^{*},\sigma_{-i}^{*})),\; i=1,2,\] so \((\sigma_1^{*},\sigma_2^{*})\in SSNE(\Gamma)\). \end{dow} Recall that the Nash equilibria of the strictly competitive game are precisely maximin solutions. \section{Lower payoff} In any game \(\Gamma= (\Sigma_1, \Sigma_2; P_1,P_2:\Sigma_1\times\Sigma_2\to{\mathbb R})\) predicted possible behavior of players may be described via the \textit{response map} \(R_i: \Sigma_{1}\times\Sigma_{2} \to 2^{\Sigma_{i}}\), \(i=1,2\). The \textit{best response map} \(BR_i: \Sigma_{1}\times\Sigma_{2} \to 2^{\Sigma_{i}}\) defined by \[BR_{i}((\sigma_{i},\sigma_{-i})) = \left\{\,\dach{\sigma_{i}}\in\Sigma_{i} \,:\, P_{i}((\dach{\sigma_{i}},\sigma_{-i})) = \max_{\sigma_{i}'\in\Sigma_{i}}\; P_{i}((\sigma_{i}',\sigma_{-i}))\,\right\}\] for \(i=1,2\), \(((\sigma_{i},\sigma_{-i})) \in\Sigma_{1}\times\Sigma_{2}\), is a usual choice, provided that the payoffs are (upper semi-) continuous and the strategy spaces are compact. We would like to investigate other option: not worse response. The \textit{not worse response map} \(NWR_i: \Sigma_{1}\times\Sigma_{2} \to 2^{\Sigma_{i}}\) is defined by \[NWR_{i}((\sigma_{i},\sigma_{-i})) = \left\{\,\dach{\sigma_{i}}\in\Sigma_{i} \,:\, P_{i}((\dach{\sigma_{i}},\sigma_{-i})) \geq P_{i}((\sigma_{i},\sigma_{-i}))\,\right\}\] for \(i=1,2\), \(((\sigma_{i},\sigma_{-i})) \in\Sigma_{1}\times\Sigma_{2}\). Indeed any player who wants to maximize his payoff would not change his strategy into a new one, if it leads to lower payoff. Therefore any response map \(R_i\) should obey the following restrictions \[BR_i((\sigma_{i},\sigma_{-i})) \subset R_i((\sigma_{i},\sigma_{-i})) \subset NWR_i((\sigma_{i},\sigma_{-i})).\] Nevertheless one is not forced to play the best response. We should only incorporate in our calculations the possible answers of the co-player to estimate sure gain. This is reflected by the \textit{lower payoff function} \(P^{\flat}_i:\Sigma_{1}\times\Sigma_{2}\to{\mathbb R}\), \[P^{\flat}_i((\sigma_{i},\sigma_{-i})) = \inf\; P_{i}((\,\sigma_i, R_{3-i}((\sigma_{-i},\sigma_i))\,)).\] In the case of our choice \(R_i=NWR_i\): \[P^{\flat}_i((\sigma_{i},\sigma_{-i})) = \inf\; \{\,P_{i}((\sigma_i,\dach{\sigma_{-i}})) \,:\, P_{-i}((\dach{\sigma_{-i}},\sigma_i)) \geq P_{-i}((\sigma_{-i},\sigma_i)) \,\}.\] Let us note a simple but useful property \begin{lem}\label{flatP} There holds estimation \(P^{\flat}_i\leq P_i\),\, \(i=1,2\). Moreover, \(P^{\flat}_i(\sigma_1,\sigma_{2})=P_i(\sigma_1,\sigma_{2})\) for \(i=1,2\) if and only if \((\sigma_1,\sigma_2)\in CWI(\Gamma)\). \end{lem} \begin{tw} Let \(\Sigma_1,\Sigma_2\) be compact spaces. If \(P_1,P_2:\Sigma_1\times\Sigma_2\to{\mathbb R}\) are continuous, then \(P^{\flat}_1,P^{\flat}_2:\Sigma_1\times\Sigma_2\to{\mathbb R}\) are lower semicontinuous. \end{tw} \begin{dow} By continuity of \(P_i\) the map \(NWR_i:\Sigma_{1}\times\Sigma_{2} \to 2^{\Sigma_{i}}\) has closed graph for \(i=1,2\). Compactness of the domain \(\Sigma_1\times\Sigma_2\) implies that the map \[\Sigma_1\times\Sigma_2\ni ((\sigma_{i},\sigma_{-i})) \mapsto ((\,\sigma_i, NWR_{3-i}((\sigma_{-i},\sigma_i))\,)) \subset\Sigma_1\times\Sigma_2\] is upper semicontinuous with compact values. Composing it with continuous \(P_i\) and Hausdorff nonexpansive \(\min: 2^{{\mathbb R}}\to {\mathbb R}\) yields lower semicontinuity of \(P^{\flat}_i\). \end{dow} \begin{stw}\label{antagFlatP} If \(\Gamma=(\Sigma_1,\Sigma_2; P_1,P_2: \Sigma_1\times \Sigma_2\to{\mathbb R})\) is the strictly competitive game, then for \(\sigma_{i}\in\Sigma_{i}, \sigma_{-i}\in\Sigma_{-i}\) \[P^{\flat}_{i}((\sigma_i,\sigma_{-i})) = \inf_{\sigma_{-i}'\in\Sigma_{-i}} P_i((\sigma_i,\sigma_{-i}')).\] \end{stw} \begin{dow} Let \(i=1,2\), \(\sigma_{i},\sigma_{i}'\in\Sigma_{i}\), \(\sigma_{-i},\sigma_{-i}'\in\Sigma_{-i}\). Observe that \[P_{-i}(\sigma_1',\sigma_2')\geq P_{-i}(\sigma_1,\sigma_2) {\Leftrightarrow} P_{i}(\sigma_1',\sigma_2')\leq P_{i}(\sigma_1,\sigma_2).\] Hence \begin{align*} P^{\flat}_{i}((\sigma_i,\sigma_{-i})) = \inf \left\{P_i((\sigma_i,\sigma_{-i}))\,:\, P_{-i}((\sigma_i,\sigma_{-i}'))\geq P_{-i}((\sigma_i,\sigma_{-i}))\right\} = \\ \inf \left\{P_i((\sigma_i,\sigma_{-i}))\,:\, P_{i}((\sigma_i,\sigma_{-i}'))\leq P_{i}((\sigma_i,\sigma_{-i}))\right\} = \inf_{\sigma_{-i}'\in\Sigma_{-i}} P_i((\sigma_i,\sigma_{-i}')). \end{align*} \end{dow} Roughly speaking in the competitive game the lower payoff of the player depends only upon his own strategy. \section{M-equilibrium}\label{sec:m-equilibrium} We associate with \(\Gamma= (\Sigma_1, \Sigma_2; P_1,P_2:\Sigma_1\times\Sigma_2\to{\mathbb R})\) the \textit{flat-game} \(\Gamma^{\flat}= (\Sigma_1, \Sigma_2; P^{\flat}_1,P^{\flat}_2:\Sigma_1\times\Sigma_2\to{\mathbb R})\). \begin{df} An \textit{m-equilibrium} of the game \(\Gamma\) is the Nash equilibrium of \(\Gamma^{\flat}\), \(ME(\Gamma)=NE(\Gamma^{\flat})\). \end{df} The Nash and m-equilibria are not related in a straightforward way. \begin{prz}\label{MEvsNE} Let \(S_1=S_2=\{1,2,3\}\), \[G=\left[\begin{array}{ccc} {[1,4]} & {[0,0]} & {[4,4]} \\ {[0,0]} & {[3,3]} & {[5,3]} \\ {[4,4]} & {[3,5]} & {[5,3]} \\ \end{array}\right].\] Then \((2,2)\in ME(G)\cap WSSNE(G)\), \((2,3)\in ME(G) \cap NE(G)\setminus WSSNE(G)\), \((3,1)\in ME(G)\setminus NE(G)\), \((3,3)\in NE(G)\setminus ME(G)\) with strongly Pareto optimal last pair. \end{prz} Any decision rule which uses lower payoff estimations (stick to an m-equilibrium in our case) is resistant on iterated elimination of dominated strategies. Formally \begin{stw} Let the \(i\)-th player, \(i\in\{1,2\}\), expect at least \(v_i=P^{\flat}_i((\sigma_{i},\sigma_{-i}))\) from playing \(((\sigma_{i},\sigma_{-i}))\) in \(\Gamma\). Suppose that the partner of \(i\) changes his strategy \(\sigma_{-i}\) into \(\sigma_{-i}'\) according to possibly higher payoff \(P_{-i}((\sigma_{-i}',\sigma_i)) \geq P_{-i}((\sigma_{-i},\sigma_i))\). Then the \(i\)-th player is still satisfied, because \(P_{i}((\sigma_i,\sigma_{-i}')) \geq v_i\). \end{stw} This obvious assertion (restatement of the definition of the lower payoff) explains why the player does not need to reject his dominated strategies. To understand possible quirks of being content with warranted lower payoff one should consider two classic zero-sum games. \begin{prz}[Hide a coin]\label{hidecoin} Let \(S_1=S_2=\{1,2\}\), \[G=\left[\begin{array}{cc} {[-10,10]} & {[15,-15]} \\ {[15,-15]} & {[-20,20]} \\ \end{array}\right].\] Then \(NE(G)=\emptyset\), although \(ME(G)=NE(G^{\flat})=\{1\}\times S_2\), \[G^{\flat}=\left[\begin{array}{cc} {[-10,-15]} & {[-10,-15]} \\ {[-20,-15]} & {[-20,-15]} \\ \end{array}\right].\] The first player cannot ensure payoff higher than \(-10\), the second player cannot ensure payoff higher than \(-15\). \end{prz} \begin{prz}[Matching pennies]\label{pennies} Let \(S_1=S_2=\{1,2\}\), \[G=\left[\begin{array}{cc} {[1,-1]} & {[-1,1]} \\ {[-1,1]} & {[1,-1]} \\ \end{array}\right].\] Then \(NE(G)=\emptyset\), although \(ME(G)=NE(G^{\flat})=S_1\times S_2\), \[G^{\flat}=\left[\begin{array}{cc} {[-1,-1]} & {[-1,-1]} \\ {[-1,-1]} & {[-1,-1]} \\ \end{array}\right].\] None of the players can ensure payoff higher than \(-1\). On the other hand \(NE(\Delta(G)) = \{\left(\frac{1}{2}\delta_1+\frac{1}{2}\delta_2, \frac{1}{2}\delta_1+\frac{1}{2}\delta_2\right)\}\) and this constitutes the main argument for mixed strategies if we view strategic interaction as did von Neumann: trying to outmanoeuvre other participants. \end{prz} The key to resolve inconsistency of m-equilibrium with the hiding player's choice policy behind a mixed strategy is to recognize that the m-equilibrium concentrates on the question \emph{what can be warranted in one-shot game} rather than \emph{what can be gambled during repeated play}. This seems paradoxical, but one also needs to take into account injurious though rational player as during analysis in Example~\ref{HighThreat} (comp. comment from p.\pageref{exploitation} before Theorem~\ref{WSSNEisME}). We investigate further properties of flat-games and m-equilibria. \begin{stw} Lower value of a game does not change when substitute payoffs with lower payoffs: \[\sup_{\sigma_{i}\in\Sigma_{i}}\;\;\inf_{\sigma_{-i}\in\Sigma_{-i}}\;\; P_i((\sigma_{i},\sigma_{-i}))= \sup_{\sigma_{i}\in\Sigma_{i}}\;\;\inf_{\sigma_{-i}\in\Sigma_{-i}}\;\; P^{\flat}_i((\sigma_{i},\sigma_{-i})).\] \end{stw} \begin{dow} Fix \(\sigma_i\in\Sigma_i\), \(\sigma_{-i}\in\Sigma_{-i}\), \(i=1,2\). By the definition of the lower payoff and Lemma~\ref{flatP}: \[\inf_{\sigma_{-i}\in\Sigma_{-i}}\;\;P_i((\sigma_{i},\sigma_{-i})) \leq P^{\flat}_i((\sigma_{i},\sigma_{-i})) \leq P_i((\sigma_{i},\sigma_{-i})).\] Thus \begin{equation} \inf_{\sigma_{-i}\in\Sigma_{-i}}\;\;P_i((\sigma_{i},\sigma_{-i})) = \inf_{\sigma_{-i}\in\Sigma_{-i}}\;\;P^{\flat}_i((\sigma_{i},\sigma_{-i})). \end{equation} \end{dow} M-equilibrium is a strategic concept -- it depends upon mutual preferences of players. \begin{stw} Let \(\Gamma=(\Sigma_1,\Sigma_2; P_1,P_2: \Sigma_1\times \Sigma_2\to{\mathbb R})\) be a game and \(\varphi_i: {\mathbb R}\to{\mathbb R}\) be strictly increasing right continuous functions, \(i=1,2\). Then the game \(\widetilde{\Gamma}=(\Sigma_1,\Sigma_2; \widetilde{P_1},\widetilde{P_2}:\Sigma_1\times\Sigma_2\to{\mathbb R})\) transformed from \(\Gamma\) via the formula \(\widetilde{P_i}=\varphi_{i}\circ P_i\) admits the same m-equilibria as the original game \(\Gamma\); symbolically \(ME(\widetilde{\Gamma}) = ME(\Gamma)\). \end{stw} \begin{dow} Observe that \(\varphi_{i}(\inf Z) = \inf \varphi_{i}(Z)\) for \(Z\subset{\mathbb R}\), \(i=1,2\). Then a direct calculation shows that \(\widetilde{P_i}^{\flat} = \varphi_{i}\circ P^{\flat}_i\) whence the conclusion follows. \end{dow} We postpone a more technically subtle discussion of the above property to the Appendix. Weakly semi-strict Nash equilibria are those equilibria which survive ``flattenization" of the game. Due to a one-sided {exploitation of player's trust}\label{exploitation} the other player might change his strategy without loss of his payoff just to lower the payoff of his partner, which explains why not all Nash equilibria are m-equilibria (cf. Example~\ref{HighThreat}). Despite possible complications illustrated by Example~\ref{MEvsNE} a positive criterion for a Nash equilibrium to be an m-equilibrium provides \begin{tw}\label{WSSNEisME} The equilibria of the game \(\Gamma\) and its flat \(\Gamma^{\flat}\) are related as follows: \begin{enumerate} \item \(WSSNE(\Gamma)\subset WSSNE(\Gamma^{\flat})\), \item \(SSNE(\Gamma)\subset SSNE(\Gamma^{\flat})\), \item \(SNE(\Gamma)\subset SNE(\Gamma^{\flat})\). \end{enumerate} \end{tw} \begin{dow} We only check the first inclusion, the rest being analogous. Let \((\sigma_1^{*},\sigma_2^{*})\in WSSNE(\Gamma)\), \(i=1,2\). By Lemma~\ref{flatP} for \(\sigma_i\), \begin{equation}\label{PbPPPb} P^{\flat}_i((\sigma_i,\sigma_{-i}^{*})) \leq P_i((\sigma_i,\sigma_{-i}^{*})) \leq P_i((\sigma_i^{*},\sigma_{-i}^{*})) = P^{\flat}_i((\sigma_i^{*},\sigma_{-i}^{*})), \end{equation} which shows \((\sigma_1^{*},\sigma_2^{*})\in NE(\Gamma^{\flat})\). Suppose that \(P^{\flat}_i((\sigma_i,\sigma_{-i}^{*})) = P^{\flat}_i((\sigma_i^{*},\sigma_{-i}^{*}))\). Then from (\ref{PbPPPb}) \(P_i((\sigma_i,\sigma_{-i}^{*})) =P_i((\sigma_i^{*},\sigma_{-i}^{*}))\). Observe that \begin{align*} P^{\flat}_{-i}((\sigma_i,\sigma_{-i}^{*})) = \inf\left\{\, P_{-i}((\dach{\sigma_i},\sigma_{-i}^{*})) \,:\, P_{i}((\dach{\sigma_i},\sigma_{-i}^{*})) \geq P_{i}((\sigma_i,\sigma_{-i}^{*})) = P_{i}((\sigma_i^{*},\sigma_{-i}^{*})) \,\right\} \\ = \inf\left\{\, P_{-i}((\dach{\sigma_i},\sigma_{-i}^{*})) \,:\, P_{i}((\dach{\sigma_i},\sigma_{-i}^{*})) = P_{i}((\sigma_i^{*},\sigma_{-i}^{*})) \,\right\}, \end{align*} since \((\sigma_1^{*},\sigma_2^{*})\) stays in equilibrium. Now any \(\dach{\sigma_i}\) with \(P_{i}((\dach{\sigma_i},\sigma_{-i}^{*})) = P_{i}((\sigma_i^{*},\sigma_{-i}^{*}))\) gives \(P_{-i}((\dach{\sigma_i},\sigma_{-i}^{*})) \geq P_{-i}((\sigma_i^{*},\sigma_{-i}^{*}))\), because \((\sigma_1^{*},\sigma_2^{*})\) is weakly semi-strict equilibrium. Hence \[P^{\flat}_{-i}((\sigma_i,\sigma_{-i}^{*})) \geq P_{-i}((\sigma_i^{*},\sigma_{-i}^{*})) = P^{\flat}_{-i}((\sigma_i^{*},\sigma_{-i}^{*})),\] where the last equality assures Lemma~\ref{flatP}. \end{dow} \begin{stw} If \(\Gamma\) is quantitatively symmetric, then \(\Gamma^{\flat}\) too. \end{stw} Neither zero-sum, nor strict competitiveness of the game is preserved under ``flattenization" procedure as show Examples~\ref{hidecoin} and~\ref{pennies}. \section{Motivating examples}\label{sec:motivexamples} We bring to the readers attention three classic games where the m-equilibrium turns out to be a nontrivial concept. \begin{prz}[Traveler's dilemma -- continuation] Let \(G\) be as in Example \ref{traveler}. We have \(NE(G)=SNE(G)=\{(2,2)\}\) and \({P^{\flat}}_1(x,y) = {P^{\flat}}_2(y,x)= \min(x,y)-4+2\cdot\operatorname{sign}(x-y)\) for \(x,y\in \{2,3,\ldots,100\}\). Hence \(G^{\flat}\) admits two equilibria, so that \(ME(G)=\{(2,2),(100,100)\}\). The outcome \((100,100)\) was often proposed by people (cf. \cite{SciAm}) as a reasonable Pareto optimal solution regardless of a possible treacherous behavior of the co-player. (Interestingly, \(G^{\flat\flat}\) possesses three equilibria, which shows that \(NE(G^{\flat\flat})\neq NE(G^{\flat})\) in general). \end{prz} \begin{prz}[Cournot duopoly; \cite{Oligopolies,Strategy}]\label{CournotDuopoly} Let \(\Gamma= (\Sigma_1, \Sigma_2; P_1,P_2:\Sigma_1\times\Sigma_2\to{\mathbb R})\), \(S_1=S_2=[0,L]\), \(L>0\), \(P_1(x,y)=P_2(y,x) = x\cdot \left(L-(x+y)\right)\) for \(x,y\in [0,L]\) . Then \(NE(\Gamma)= SNE(\Gamma) = \left(\frac{L}{3},\frac{L}{3} \right)\). Using elementary methods (cf. \cite{Strategy}) we find that \({P^{\flat}}_1(x,y)= x\cdot \min(y, L-(x+y))\) and \[ ME(\Gamma) = \{(x,L-2x) \,:\, 0\leq x \leq {L}/{3}\} \cup \left\{ \left(x, \frac{L-x}{2}\right) \,:\, {L}/{3}\leq x \leq L\right\}. \] Note that from the cartel point of view, a Pareto dominating TU-solution \(({L}/{4},{L}/{4})\) would be superior. Unfortunately this ``natural" solution is not an m-equilibrium. Nevertheless the joint payoff \(P_1+P_2\) is maximized at two boundary m-equilibria: \((0,L)\) and \((L,0)\). This suggests that under Cournot duopoly pricing it is profitable for firms to choose an active monopolist and the other firm rest with no production. Switching the role of monopolist between firms could become a strategy (in repeated game) for hidden transfer of utility despite the payoff in game was assumed to be NTU. \end{prz} \begin{prz}[Puu duopoly; \cite{Oligopolies,PuuOligopolies}] Let \(\Gamma= (\Sigma_1, \Sigma_2; P_1,P_2:\Sigma_1\times\Sigma_2\to{\mathbb R})\), \(S_1=S_2=[0,L]\), \(L>1\), \[ P_1(x,y)=P_2(y,x) = \left(\frac{L}{x+y}-1\right) \cdot x \] for \(x,y\in [0,L]\) with convention that \(P_i(0,0)=0\), \(i=1,2\). Then \(NE(\Gamma) = \left(\frac{L}{4},\frac{L}{4} \right)\). By elementary (though a bit cumbersome) calculations \[ {P^{\flat}}_1(x,y) = \left(\frac{L}{x+{y}^{\sharp}}-1\right) \cdot x \] for \((x,y)\neq (0,0)\), where \({y}^{\sharp} = \max\left(y, \left(\frac{L}{x+y}-1\right) \cdot x\right)\). Hence \({P^{\flat}}_1(x,y)={P^{\flat}}_2(y,x) = \min(y,P_1(x,y))\) for all \(x,y\in[0,L]\). Puu duopoly enjoys a rich set of m-equilibria. Denote by \(L_{*}\approx 3.0796\) the unique positive root of the polynomial \(1+4\,L+6\,L^2+4\,L^3+L^4-L^5\) and put \[ E=\left\{\begin{array}{ll} \{(\sqrt{L},\sqrt{L}\cdot(\sqrt[4]{L}-1)), (\sqrt{L}\cdot(\sqrt[4]{L}-1),\sqrt{L}) \}, & \mbox{ when } L=L_{*},\\ {\emptyset}, & \mbox{ otherwise}.\\ \end{array}\right. \] Further, denote \[ N=\left\{\begin{array}{ll} \{\left({L}/{4},{L}/{4}\right) \}, & \mbox{ when } L>16,\\ {\emptyset}, & \mbox{ otherwise}.\\ \end{array}\right. \] We have \[ ME(\Gamma)= \bigcup_{x\in[0,\sqrt{L}]} \{x\}\times [0,\sqrt{L}-x] \,\cup E \cup N. \] The extraordinary pair of m-equilibria at \(L=L_{*}\) is an unexpected phenomenon. (It seems to be a noneconomic artifact bond to the formal model). That Nash equilibria of \(\Gamma\) need not be m-equilibria unless \(L\) is sufficiently large, is an effect of weakness of Nash equilibrium: when taking into account the security of payoff, a treacherous partner can switch precomitted (during cheap talk) strategy to a strategy indifferent for him but harmful for his co-player. Formally, \(({L}/{4},{L}/{4})\) is not a (weakly semi-) strict Nash equilibrium for small \(L\). Finally observe that for \(L<4\) the m-equilibrium \(({\sqrt{L}}/{2},{\sqrt{L}}/{2})\) Pareto dominates the Nash equilibrium \(({L}/{4},{L}/{4})\). One can stipulate that such m-equilibria might ``explain" cartels in a game theoretic way without a recourse to exterior (outside game) constructs. \end{prz} There is no doubt that the traveler's dilemma was the driving force of our research. Note that m-equilibria do not bring anything new to the (in)famous prisoner's dilemma (PD). This confirms that the traveler's dilemma is not merely an extension of PD -- it is something qualitatively different. On the other hand simultaneous simplicity and nontriviality of PD shows that having a good solution concept does not negate the reason to perform the play at all: knowing consequences is not freeing us from making decisions. \section{Existence of m-equilibrium} We know from Theorem~\ref{WSSNEisME} that the class of games which possess at least one m-equilibrium is quite large. Unfortunately we do not know whether m-equilibria exist under suitaby mild assumptions in general. Nevertheless competitive games admit pure m-equilibria. \begin{tw} If \(\Gamma=(\Sigma_1,\Sigma_2; P_1,P_2: \Sigma_1\times\Sigma_2\to{\mathbb R})\) is the strictly competitive game with compact metrizable strategy spaces \(\Sigma_i\) and continuous payoffs \(P_i\), \(i=1,2\), then \(ME(\Gamma)\neq\emptyset\). Namely, Wald solutions rest in m-equilibrium. \end{tw} \begin{dow} Observe that for \(\sigma_i,\sigma_i'\in\Sigma_i\), \(i=1,2\) \[d_H\left(\,P_i((\sigma_i,\Sigma_{-i})),P_i((\sigma_i',\Sigma_{-i}))\,\right) \leq \sup_{\sigma_{-i}\in\Sigma_{-i}} \left|P_i((\sigma_i,\sigma_{-i})) - P_i((\sigma_i',\sigma_{-i}))\right|,\] where \(d_H\) stands for the Hausdorff distance in \(2^{{\mathbb R}}\). Since \(P_i\) are uniformly continuous (as continuous on the compactum), we know that \(\Psi_i:\Sigma_i\to 2^{{\mathbb R}}\), \(\Psi_i(\sigma_i) = P_i((\sigma_i,\Sigma_{-i}))\) for \(\sigma_i\in\Sigma_i\), are Hausdorff continuous with compact values. The Hausdorff nonexpansiveness of \(\min:2^{{\mathbb R}}\to{\mathbb R}\) yields then the continuity of the map \[\Sigma_i\ni \sigma_i\mapsto \min_{\sigma_{-i}\in\Sigma_{-i}} P_i((\sigma_i,\sigma_{-i})) = \min \Psi_i(\sigma_i).\] This shows that we can define \[\sigma_i^{*} \in \arg\max_{\sigma_i\in\Sigma_i}\; \min_{\sigma_{-i}\in\Sigma_{-i}}\, P_i((\sigma_i,\sigma_{-i}))\] for \(i=1,2\). So \((\sigma_1^{*},\sigma_2^{*})\in W(\Gamma)\neq\emptyset\). Further by Proposition~\ref{antagFlatP} \begin{align*} \max_{\sigma_i\in\Sigma_i}\; \min_{\sigma_{-i}\in\Sigma_{-i}}\, P_i((\sigma_i,\sigma_{-i})) = \max_{\sigma_i\in\Sigma_i}\; P^{\flat}_i((\sigma_i,\sigma_{-i})) \\ = \max_{\sigma_i\in\Sigma_i}\; P^{\flat}_i((\sigma_i,\sigma_{-i}^{*})) = P^{\flat}_i((\sigma_{i}^{*},\sigma_{-i}^{*})) \geq P^{\flat}_i((\sigma_i,\sigma_{-i}^{*})) \end{align*} for any \(\sigma_i\in\Sigma_i\), \(\sigma_{-i}\in\Sigma_{-i}\). Therefore \(W(\Gamma)\subset NE(\Gamma^{\flat}) = ME(\Gamma)\). \end{dow} Unfortunately the results established so far in the literature (cf. \cite{Bagh,Reny,Scalzo}) which are concerned with the existence of equilibria in games with discontinuous payoff functions do not seem to be applicable for the kind of problems considered here. \section{Mixed strategies and risk} The reason to calculate lower payoffs is establishing sure gains. Therefore one might question the use of expected payoffs to evaluate gains. We single out this phenomenon in the case of zero sum game. \begin{prz}[extended matching pennies] Let \(S_1=S_2=\{1,2,3\}\), \[G=\left[\begin{array}{ccc} {[-1,1]} & {[1,-1]} & {[0,0]} \\ {[1,-1]} & {[-1,1]} & {[0,0]} \\ {[0,0]} & {[0,0]} & {[0,0]} \\ \end{array}\right].\] Then \(NE(\Delta(G)) = \{(\delta_3,\delta_3)\), \(\left(\frac{1}{2}\delta_1+\frac{1}{2}\delta_2, \frac{1}{2}\delta_1+\frac{1}{2}\delta_2\right)\), \(\left(\frac{1}{2}\delta_1+\frac{1}{2}\delta_2,\delta_3\right)\), \(\left(\delta_3,\frac{1}{2}\delta_1+\frac{1}{2}\delta_2\right)\}\). Although all equilibria yield the same expected payoff, they differ significantly from the point of view of the risk. Namely the variance in payoff is nonzero unless both players use pure strategies (standard property of random variables). This has consequence for risk averse players usually not considered in the classic von Neumann's minimax theory. \end{prz} Let \(S=\{x_1,x_2,x_3,x_4,\ldots\}\) be the set of prizes with the associated utility function \(U:S\to {\mathbb R}\), such that \(U(x_1)<U(x_3)<U(x_2)\). Risk neutral players calculate gain for the lottery \((S,\pi)\), \(\pi\in\Delta(S)\), via the expected utility \[EU(\pi)=\sum_{x\in S} \pi(x)\cdot U(x).\] Hence they are indifferent in the choice between two lotteries \((\{x_1,x_2\},\rho)\) and \(\{x_3\}\) as long as \(EU(\rho)=U(x_3)\). However loss averse players would rather calculate the minimal gain \[E^{\min}U(\pi)= \min_{x\in\operatorname{supp}\pi} U(x)\] for the lottery \((S,\pi)\). Then \(E^{\min}U(\rho)<U(x_3)\) and \(\{x_3\}\) is preferred over \((\{x_1,x_2\},\rho)\) whenever \(x_1\in\operatorname{supp}\rho\). Take into account another lottery \((\{x_1,x_2\},\rho')\) such that \(x_1\in\operatorname{supp}\rho'\). Then \(E^{\min}U(\rho')=E^{\min}U(\rho)\), so \(\rho\) and \(\rho'\) seem equally good. Still loss averse players might evaluate which of the given two lotteries with the same minimal gain has higher expected gain (as secondary criterion for preferences), \(EU(\rho)\) or \(EU(\rho')\)? Concerning mixed strategies one should also be aware that the probability distribution might be also interpreted deterministically as a set of weights describing ``fair" allocation of welfare/payoff induced by the choice of strategies. We discuss related questions in the next Section. \section{Equilibrium selection}\label{sec:equilibriumselection} The concept of m-equilibrium takes loss aversion and correlated decision into serious consideration. It demands communication and sure gains to be estimated. However it is \emph{not} correlated equilibrium of Aumann. It also accounts for losses on the more basic level than the Harsanyi-Selten risk dominance selection criterion. Nevertheless m-equilibria (being Nash equilibria of the game with flattened payoffs) suffer the same curse of nonuniqueness (of payoff) as other notions of solution designed for non zero sum games. To avoid complicated matters of the formal definitions of communication (pre-play) we simply say that the \textit{players can communicate} to establish the final decision in a game \(\Gamma=(\Sigma_1,\Sigma_2; P_1,P_2: \Sigma_1\times\Sigma_2\to{\mathbb R})\), provided there exists a ``communication channel" \(\mathcal{C}: \Sigma_1\times\Sigma_2 \to ME(\Gamma)\), where \(\mathcal{C}\) is a random variable distributed on the set of m-equilibria according to probability \(\alpha\in \Delta[ME(\Gamma)]\). Vector \(\alpha\) will be interpreted further also as the set of weights of welfare allocation among m-equilibria. Although communication restores Pareto-efficient equilibrium selection in Example \ref{Coordination} and the stag hunt game, we will face classical coordination dilemma posed by the battle of the sexes game. In presence of equal power and credibility players, the coordination dilemma is often resolved via fair allocation rule: ``once for me, once for you". We believe that cooperative social choice among various equilibria is the appropriate answer to equilibrium selection in both, one-shot and repeated games. Together with a social rule providing the allocation vector \(\alpha\in\Delta[ME(\Gamma)]\), some stochastic tie-breaking rules are indispansable. (During repeated play the variance of payoff outcomes arises as another problem. Alternate choice of equilibria minimizes this variance). \begin{prz}[Battle of the sexes] Let \(S_1=S_2=\{1,2\}\), \[G=\left[\begin{array}{cc} {[3,2]} & {[0,0]} \\ {[0,0]} & {[2,3]} \\ \end{array}\right].\] Then \(NE(G)=ME(G) = \{(1,1), (2,2)\}\); \(\alpha = \frac{1}{2}\cdot \delta_{(1,1)} + \frac{1}{2}\cdot \delta_{(2,2)}\). The only way to get rid of the question ``who's equilibrium played first" is to apply randomization device according to distribution given by \(\alpha\). This is an instance of Szaniawski's probabilistically equal choice principle (\cite{Lissowski}). In one-shot games stochastic mechanism for choosing the player who selects preferred equilibrium to be played seems very reasonable also according to Laplace's criterion of insufficient reason. \end{prz} Once players receive the recommended equilibrium after pre-play phase, they form their beliefs and strategic properties of m-equilibrium warrant the appropriate payoff levels regardless of whether one of the players tries to exploit this information. Roughly speaking, to cut-off the inductive race to the bottom, it is enough that at least one player is fair. This unavoidably leads to the problem of reputation. \section{Final comments} The following problems are very important for the discussion of the relevance of the concept of m-equilibrium: \begin{enumerate} \item Do (pure strategy) m-equilibria always exist under reasonable assumptions about payoff functions? \item What other than traveler's dilemma games admit ``intuitively superior" m-equilibria impossible within standardly interpreted Nash framework? \item How to cope with risk and welfare allocation? Does there exist any clear risk dominance criterion? (Cf. \cite{Harsanyi}). \end{enumerate} To prove a general existence theorem on m-equilibria definitely one cannot use continuity of lower payoffs, but some assumptions about payoffs and strategy sets are indispensable. \begin{prz} Let \(\Gamma=(\Sigma_1,\Sigma_2; P_1,P_2: \Sigma_1\times\Sigma_2\to{\mathbb R})\), \(x,y\in\Sigma_1=\Sigma_2=[0,\infty)\) and \(L\geq 2C>0\). We define \(P_1(x,y)=P_2(y,x)= \frac{Lx}{x+y} - \frac{C}{x}\), when \(x>0\) and \(P_1(0,y)=0\) otherwise. We assume here (unlike\cite{PuuOligopolies, Oligopolies}) that the total cost of production decreases \(C/x \searrow 0\) as the production of the firm increases \(x\nearrow \infty\). One can think about this opportunity as the effect of scale (globalization). It turns out that under our assumption of diminishing cost, the Puu duopoly behaves qualitatively in a similar fashion to that observed in the competition of ``giants": each player has an incentive to grow production for overtaking the market; in practice we expect the mirroring behavior of firms, because it warrants maximal payoff share (according to TU value). Interestingly \({P^{\flat}}_1(x,y)={P^{\flat}}_2(y,x) = -\,\frac{C}{x}\), which reflects an old truth that in business one might bear the cost of production without any profit (``fall of a giant"). Consequently \(ME(\Gamma)= \{(0,0)\}\). If \(\Sigma_1=\Sigma_2=(0,\infty)\), then \(\Gamma\) has no m-equilibrium. A reasonable workaround could be then to allow for an epsilon-m-equilibrium (produce as little as possible). \end{prz} A sky-rocketing competition in the above Example tells us that the dynamic view of games is necessary when the game is played more than once. We do not consider multiplayer games in this article because we believe that only good understanding of two player games can give rise for reasonable extensions of static duel games to the situation of multiple interacting agents. We are aware of specific ``phase transitions" and emergent effects when passing from the case of two players to the case of multiple players. The adaptation of the notion of m-equilibrium for multiplayer games should be done carefully. Let \(\Gamma= (\Sigma_1, \ldots, \Sigma_N; P_1,{\ldots},P_N:\Sigma_1\times\ldots\times\Sigma_N\to{\mathbb R})\) be a game with \(N\) players. Since the player can only be sure of his own declaration and the communicated decisions of others might be changed, the following definition of the lower payoff seems to be suitable in this sort of situation: \begin{gather*} P^{\flat}_i(\sigma_{1},{\ldots},\sigma_{N}) = \inf\; \{\,P_{i}({\varsigma}^{J}) \\ \,:\, \exists_{J\subset \{1,{\ldots},N\}}\;\forall_{j\in J}\;\; P_{j}({\varsigma}^{J}) \geq P_{j}(\sigma_{1},{\ldots},\sigma_{N}) \,\} \end{gather*} for \((\sigma_{1},{\ldots},\sigma_{N})\), \({\varsigma}^{J}\) \(\in\Sigma_1\times\ldots\times\Sigma_N\), where \({\varsigma}^{J}_i = \sigma_i\) when \(i\not\in J\), i.e., \({\varsigma}^{J}\) may differ from \((\sigma_{1},{\ldots},\sigma_{N})\) for players \(i\) contributing to a virtual coalition \(J\). Some other ideas aiming to resolve the dominated strategies conundrum in traveler's dilemma were reported in \cite{HalpernPass}. Another concept of solution suitable for traveler's dilemma (and accounting for the lack of communication unlike in the present article) is Hofstadter's superrationality which can be argued within bayesian framework, e.g., \cite{QuasiMagical}. However our intention was to dispose off as much probability as possible. The cryptic term ``m-equilibrium" was thought off by the author in accordance with the notion of meta-stable equilibrium appearing among others in chemistry and physics; that is an extraordinary equilibrium (or higher state) possible only under very specific conditions. \section*{Appendix: Isomorphism of games} We say that a function \(\varphi: {\mathbb R}\supset Z \to {\mathbb R}\) is \begin{itemize} \item \textit{strictly inf-increasing}, if for nonempty \(U_1,U_2\subset Z\) \[\inf U_1<\inf U_2 \Rightarrow \inf\varphi(U_1) < \inf\varphi(U_2),\] \item \textit{inf-continuous}, if for every nonempty \(U\subset Z\) such that \(\inf U\in Z\) holds \(\varphi(\inf U) = \inf\varphi(U)\), \item \textit{right continuous}, if for every \(z_0\in Z\) and every (w.l.o.g. decreasing) sequence \(z_n\in Z\), \(z_0\leq z_n\to z_0\) holds \(\varphi(z_n)\to\varphi(z_0)\). \end{itemize} \begin{stw} Let \(\varphi: {\mathbb R}\supset Z\to{\mathbb R}\). \begin{enumerate} \item If \(\varphi\) is strictly inf-increasing, then it is strictly increasing. \item If \(\varphi\) is (not necessarily strictly) increasing, then it is inf-continuous if and only if it is right continuous. \item If \(Z={\mathbb R}\) and \(\varphi\) is strictly increasing inf-continuous, then it is strictly inf-increasing. \end{enumerate} \end{stw} We warn that infima are taken in the whole \({\mathbb R}\), not in the ordered subset \(Z\subset{\mathbb R}\). That the notion of strictly inf-increasing function is essentially stronger than strictly increasing function illustrates \begin{prz} Let \(Z=\{0\}\cup (1,\infty)\subset{\mathbb R}\), \(\varphi: Z\to{\mathbb R}\), \(\varphi(z)=\max\{z-1,0\}\) for \(z\in Z\). Although \(\varphi\) is strictly increasing inf-continuous function it is not strictly inf-increasing. \end{prz} \begin{stw} Let \(\Gamma=(\Sigma_1,\Sigma_2; P_1,P_2: \Sigma_1\times \Sigma_2\to{\mathbb R})\) be a game and \(\varphi_i: P_i(\Sigma_1\times \Sigma_2)\to{\mathbb R}\) be order-preserving maps, \(i=1,2\), i.e. \[\forall_{u,v\in P_i(\Sigma_1\times \Sigma_2)}\;\; u<v {\Rightarrow} \varphi_{i}(u)<\varphi_{i}(v).\] Then the game \(\widetilde{\Gamma}=(\Sigma_1,\Sigma_2; \widetilde{P_1},\widetilde{P_2}:\Sigma_1\times\Sigma_2\to{\mathbb R})\) transformed from \(\Gamma\) via the formula \(\widetilde{P_i}=\varphi_{i}\circ P_i\) admits the same m-equilibria as the original game \(\Gamma\); symbolically \(ME(\widetilde{\Gamma}) = ME(\Gamma)\). \end{stw} \begin{dow} Direct calculation shows that \(\widetilde{P_i}^{\flat} = \varphi_{i}\circ P^{\flat}_i\) whence the conclusion follows. [Needed \(\varphi(\inf Z)= \inf \varphi(Z)\)]. \end{dow} An analysis of ``equivalent" prisoner's dilemmas shows that isomorphic games may have nonequivalent risk structure. Therefore an appropriate concept of isomorphism of normal form games is no less controversial than the choice of satisfactory definition of the solution of a game or the equilibrium selection problem. \section*{Acknowledgement} The author's research was ignited in 2008 by S{\l}awomir Plaskacz (differential inclusions, Hamilton-Jacobi equation in control and optimization, differential games). We had a lot of vigorous discussions. A criticism of the earlier concepts proposed by the author (Nash-von Neumann cooperative solution, correlated Pareto optimum, retaliatory safe optimum) led to the concept of m-equilibrium. I would like to thank all participants of the three seminars where, around 2009, I referred those unsatisfactory concepts: Seminar of the Chair of Nonlinear Mathematical Analysis and Topology at the Nicolaus Copernicus University (Wojciech Kryszewski's group), Seminar of the Game and Decision Theory Group at the Polish Academy of Sciences (Andrzej Wieczorek's group) and Seminar of the Chair of Mathematical Economics at the Poznan University of Economics (Emil Panek's group). Almost all of this work has been done by the author during 2009-2010 in the Faculty of Mathematics and Computer Science at the Nicolaus Copernicus University (Toru\'{n}, Poland). It would be really hard to grasp the current state of the research in game theory, if not \textit{books.google} and various free resources provided by the experts in the subject.

https://arxiv.org/abs/1304.3966

Projective cell modules of Frobenius cellular algebras

For a finite dimensional Frobenius cellular algebra, a sufficient and necessary condition for a simple cell module to be projective is given. A special case that dual bases of the cellular basis satisfying a certain condition is also considered. The result is similar to that in symmetric case.

\section{Introduction}\label{xxsec1} Cellular algebras were introduced by Graham and Lehrer \cite{GL} in 1996, motivated by previous work of Kazhdan and Lusztig \cite{KL}. They were defined by a so-called cellular basis satisfying certain axioms. The theory of cellular algebras provides a systematic framework for studying the representation theory of many interesting algebras from mathematics and physics. The classical examples of cellular algebras include Hecke algebras of finite type, Ariki-Koike algebras, Brauer algebras, Birman-Wenzl algebras and so on. We refer the reader to \cite{G, GL, Xi2} for details. Recently, Koenig and Xi \cite{KX8} introduced affine cellular algebras which contain cellular algebras as special cases. They proved affine Hecke algebras of type A are affine cellular. Several important classes of cellular algebras are symmetric, such as Hecke algebras of finite types, Ariki-Koike algebras and so on. In \cite{L, L2}, Li studied the general theory of symmetric cellular algebras, such as dual bases, centres and radicals. Moreover, Li and Xiao considered the classification of the projective cell modules of symmetric cellular agebras in \cite{LX}. Frobenius algebras are natural generalizations of symmetric algebras. It is well known that a Frobenius algebra is symmetric if and only if there exists a Nakayama automorphism being the identical mapping. Examples of non-symmetric Frobenius cellular algebras could be found in \cite{KX2}. In \cite{L3}, Li investigated Nakayama automorphisms of Frobenius cellular algebras. It is proved that the matrix associated with a Nakayama automorphism with respect to a cellular basis is uni-triangular under a certain condition. A natural question is which properties of symmetric cellular algebras can be generalized to Frobenius cases. In this paper, for a Frobenius cellular algebra, we will give a sufficient and necessary condition for a simple cell module to be projective. A special case that dual bases of the cellular basis satisfying a certain condition is also considered. The result is similar to that in symmetric case. The paper is organized as follows. We begin with a quick review on the theory of cellular algebras and Frobenius algebras. In particular, we give in this section a corollary of Gasch\"{u}tz-Ikeda's Theorem, which is more useful to study cell modules of Frobenius cellular algebras. Then in Section 3, after giving some properties of Frobenius cellular algebras, we give a sufficient and necessary condition for a simple cell module to be projective. A special case will be considered in Section 4. The result is similar to that in symmetric case. \section{Preliminaries}\label{xxsec2} In this section, we will establish the basic notations and some known results which are needed in the following sections. The main references for this section are \cite{GP} and \cite{GL}. \subsection{Frobenius algebras} Let $R$ be a commutative ring with identity and $A$ a finite dimensional associative $R$-algebra. Suppose that there exists an $R$-bilinear form $f:A\times A\rightarrow R.$ We say $f$ is {\it non-degenerate} if the determinant of the matrix $(f(a_{i},a_{j}))_{a_{i},a_{j}\in B}$ is not zero for some basis $B$ of $A$. We call $f$ {\it associative} if $f(ab,c)=f(a,bc)$ for all $a,b,c\in A$. \begin{dfn}\label{2.1} An $R$-algebra $A$ is called {\it Frobenius} if there is a non-degenerate associative bilinear form $f$ on $A$. \end{dfn} Let $A$ be a Frobenius algebra with a basis $B=\{a_{i}\mid i=1,\ldots,n\}$. Let us take a non-degenerate associative bilinear form $f$. Define an $R$-linear map $\tau: A\rightarrow R$ by $$\tau(a)=f(a,1).$$ We call $d=\{d_{i}\mid i=1,\ldots,n\}$ the {\it right dual basis} of $B$ which is uniquely determined by the requirement that $\tau(a_{i}d_{j})=\delta_{ij}$ for all $i, j=1,\ldots,n$. Similarly, the {\it left dual basis} $D=\{D_{i}\mid i=i,\ldots,n\}$ is determined by the requirement that $\tau(D_{j}a_{i})=\delta_{ij}$. Define an $R$-linear map $\alpha:A\rightarrow A$ by $$\alpha(d_{i})=D_{i}.$$ Then $\alpha$ is a Nakayama automorphism of $A$. \smallskip From now on, all of modules considered in this paper will always be left modules. Let $M$ be an $A$-module and $\theta\in {\rm End}_R(M)$. Then we define the {\it averaging operator} $I(\theta)\in {\rm End}_R(M)$ by $$I(\theta)(m):=\sum_ia_i\theta(D_im) \quad\quad \forall\, m\in M.$$ It is an $A$-module endomorphism. Furthermore, $I(\theta)$ is independence of the choice of basis. This implies that $I(\theta)$ also can be defined as follows. $$I(\theta)(m)=\sum_id_i\theta(a_im) \quad\quad \forall\, m\in M.$$ Let $\theta, \pi, \varphi\in \End_A(M)$. Then the definition implies that $$I(\varphi\circ \theta)=I(\varphi)\circ \theta \quad {\rm and}\quad I(\theta\circ\pi)=\theta\circ I(\pi).$$ Moreover, it is helpful to point out that $I(\theta)\in \End_A(M)$. One of the importance of the averaging operator is that it provides a criterion for an $A$-module being projective. \begin{lem}\rm (Gasch\"{u}tz-Ikeda)\label{2.2} Let $A$ be a Frobenius algebra and $M$ an $A$-module. Then $M$ is projective if and only if there exists some $\varphi\in {\rm End}_R(M)$ such that $I(\varphi)={\rm id}_M$. \end{lem} Let $R$ be a field. Suppose that $\dim M=m$ and take a basis $\{v_1, \cdots, v_m\}$. For $1\leq i, j\leq m$, define $\varphi_{ij}\in \End_R(M)$ by $\varphi_{ij}(v_s)=\delta_{is}v_j$. Clearly, $\varphi_{ij}$ form a basis of $\End_R(M)$. The following result is a simple corollary of Lemma \ref{2.2}. However, it is important in this paper. Hence we will give a complete proof of it here. \begin{cor}\label{2.3} Let $A$ be a finite dimensional Frobenius algebra and $M$ a simple $A$-module. Then $M$ is projective if and only if there exists some $\varphi_{ij}$ such that $I(\varphi_{ij})\neq 0$. \end{cor} \begin{proof} Since $\varphi_{ij}$ form a basis of $\End_R(M)$, the necessity is a direct corollary of Lemma \ref{2.2}. Conversely, assume that there exists some $\varphi_{ij}$ such that $I(\varphi_{ij})\neq 0$. Note that $M$ is simple. Then Schur's lemma implies that $I(\varphi_{ij})=r_{ij}\,{\rm id_M}$, where $r_{ij}\in R-\{0\}$. Let $\pi: N\rightarrow M$ be an epimorphism of $A$-modules. Then there exists an $R$-linear map $\mu: M\rightarrow N$ such that $\pi\circ\mu={\rm id_M}$. This gives that $\pi\circ\mu\circ\varphi_{ij}=\varphi_{ij}$ and thus $I(\pi\circ\mu\circ\varphi_{ij})=r_{ij}\,{\rm id_M}$. This implies that $r_{ij}^{-1}I(\mu\circ\varphi_{ij})$ is the desired map. \end{proof} \subsection{Cellular algebras} Let us recall the definition of a cellular algebra given by Graham and Lehrer in \cite{GL}. \begin{dfn} {\rm \cite{GL}}\label{2.4} Let $R$ be a commutative ring with identity. An associative unital $R$-algebra is called a {\it cellular algebra} with cell datum $(\Lambda, M, C, i)$ if the following conditions are satisfied: {\rm(C1)} The finite set $\Lambda$ is a poset. Associated with each $\lam\in\Lambda$, there is a finite set $M(\lam)$. The algebra $A$ has an $R$-basis $\{C_{S,T}^\lam \mid \lam\in\Lambda,\,\, S,T\in M(\lam)\}$. \smallskip {\rm(C2)} The map $i$ is an $R$-linear anti-automorphism of $A$ with $i^{2}={\rm id}$ which sends $C_{S,T}^\lam$ to $C_{T,S}^\lam$ for all $\lam\in\Lambda$ and $S,T\in M(\lam)$. \smallskip {\rm(C3)} If $\lam\in\Lambda$ and $S,T\in M(\lam)$, then for any element $a\in A$, we have\\ $$aC_{S,T}^\lam\equiv\sum_{S^{'}\in M(\lam)}r_{a}(S',S)C_{S^{'},T}^{\lam} \,\,\,\,(\rm {mod}\,\,\, A(<\lam)),$$ where $r_{a}(S^{'},S)\in R$ is independent of $T$ and where $A(<\lam)$ is the $R$-submodule of $A$ generated by $\{C_{U,V}^\mu \mid U, V\in M(\mu),\,\,\mu<\lam\}$. \smallskip If we apply $i$ to the equation in {\rm(C3)}, we obtain {$\rm(C3')$} $C_{T,S}^\lam i(a)\equiv\sum\limits_{S^{'}\in M(\lam)}r_{a}(S^{'},S)C_{T,S^{'}}^{\lam} \,\,\,\,(\mod A(<\lam)).$ \end{dfn} As a natural consequence of the axioms, the {\em cell modules} are defined as follows. \begin{dfn} \label{2.5} Let $A$ be a cellular algebra with cell datum $(\Lambda, M, C, i)$. For each $\lam\in\Lambda$, the {\it cell modules} $W_C(\lam)$ is left $A$-module defined as follows: $W_C(\lam)$ is a free $R$-module with basis $\{C_{S}\mid S\in M(\lam)\}$ and $A$-action defined by\\ $$aC_{S}=\sum_{S^{'}\in M(\lam)}r_{a}(S^{'},S)C_{S^{'}} \,\,\,\,(a\in A,S\in M(\lam)),$$ where $r_{a}(S^{'},S)$ is the element of $R$ defined in Definition {\rm\ref{2.4}}{\rm(C3)}. \end{dfn} Let $A$ be a cellular algebra with cell datum $(\Lambda, M, C, i)$. Let $\lam\in\Lambda$ and $S,T,U,V\in M(\lam)$. Then we have from Definition \ref{2.4} that $$C_{S,T}^\lam C_{U,V}^\lam \equiv\Phi(T,U)C_{S,V}^\lam\,\,\,\, (\rm mod\,\,\, A(<\lam)),$$ where $\Phi(T,U)\in R$ depends only on $T$ and $U$. Thus for a cell module $W_C(\lam)$, we can define a symmetric bilinear form $\Phi _{\lam}:\,\,W_C(\lam)\times W_C(\lam)\longrightarrow R$ by $$\Phi_{\lam}(C_{S},C_{T})=\Phi(S,T).$$ The radical of the bilinear form is defined to be $$\rad_{\Phi}\lam:= \{x\in W(\lam)\mid \Phi_{\lam}(x,y)=0\,\,\,\text{for all} \,\,\,y\in W_C(\lam)\}.$$ By the general theory of cellular algebras, $\rad_{\Phi}\lam$ is an $A$-submodule of $W_C(\lambda)$, motivating the definition $L(\lambda)=W_C(\lam)/\rad_{\Phi}\lam$. In principle, the next result of Graham and Lehrer classifies the simple $A$-modules. \begin{lem} {\rm(\cite{GL})}\label{2.6} Let $K$ be a field and $A$ a finite dimensional cellular algebra over $K$. Let $\Lambda_{0}=\{\lam\in\Lambda\mid \Phi_{\lam}\neq 0\}$. Then $\{L(\lam)\mid \lam\in\Lambda_{0}\}$ is a complete set of pairwise non-isomorphic absolutely irreducible $A$-modules. \end{lem} For any $\lam\in\Lambda$, fix an order on $M(\lam)$ and let $M(\lam)=\{S_{1},S_{2},\cdots,S_{n_{\lam}}\}$, where $n_{\lam}$ is the number of elements in $M(\lam)$, the matrix $$G(\lam)=(\Phi(S_{i},S_{j}))_{1\leq i,j\leq n_{\lam}}$$ is called {\em Gram matrix}. Note that all the determinants of $G(\lam)$ defined with different order on $M(\lam)$ are the same. By the definition of $G(\lam)$ and $\rad_{\Phi}\lam$, for a finite dimensional cellular algebra $A$, if $\Phi_{\lam}\neq 0$, then $\dim_{K}L(\lam)=\rank G(\lam)$. \section{Projective cell modules of Frobenius cellular algebras} \label{xxsec3} Let $K$ be a field. Let $A$ be a finite dimensional $K$-algebra and $M$ an $A$-module. The algebra homomorphism $$\rho_{M}:A\rightarrow \End_{K}(M),\,\,\, \rho_{M}(a)m=am,\,\,\,\, \,\,\,\forall\,\, m\in M,\,\,\,\, a\in A,$$ is called the representation afforded by $M$. \smallskip Let $A$ be a finite dimensional Frobenius cellular $K$-algebra with a cell datum $(\Lambda, M, C, i)$. Given a non-degenerate bilinear form $f$, denote the left dual basis by $D=\{D_{S,T}^\lam \mid S,T\in M(\lam),\lam\in\Lambda\}$, which satisfies $$ \tau(D_{U,V}^{\mu}C_{S,T}^{\lam})=\delta_{\lam\mu}\delta_{SV}\delta_{TU}. $$ Denote the right dual basis by $d=\{d_{S,T}^\lam \mid S,T\in M(\lam),\lam\in\Lambda\}$, which satisfies $$ \tau(C_{S,T}^{\lam}d_{U,V}^{\mu})=\delta_{\lam\mu}\delta_{S,V}\delta_{T,U}. $$ For $\mu\in\Lambda$, let $A_{d}(>\mu)$ be the $K$-subspace of $A$ generated by $$\{d_{X,Y}^\epsilon \mid X,Y\in M(\epsilon),\mu<\epsilon\}$$ and let $A_{D}(>\mu)$ be the $K$-subspace of $A$ generated by $$\{D_{X,Y}^\epsilon \mid X,Y\in M(\epsilon),\mu<\epsilon\}.$$ Note that the $K$-linear map $\alpha$ which sends $d_{X,Y}^\epsilon$ to $D_{X,Y}^\epsilon$ is a Nakayama automorphism of the algebtra $A$. For arbitrary $\lam, \,\mu\in \Lambda$, $S,\,T\in M(\lam)$, $U,\,V\in M(\mu)$, write $$C_{S,T}^{\lam}C_{U,V}^{\mu}=\sum\limits_{\epsilon\in\Lambda,X,Y\in M(\epsilon)} r_{(S,T,\lam),(U,V,\mu),(X,Y,\epsilon)}C_{X,Y}^{\epsilon},$$ $$D_{S,T}^{\lam}D_{U,V}^{\mu}=\sum\limits_{\epsilon\in\Lambda,X,Y\in M(\epsilon)} R_{(S,T,\lam),(U,V,\mu),(X,Y,\epsilon)}D_{X,Y}^{\epsilon}.$$ Applying $\alpha^{-1}$ on both sides of the above equation, we obtain $$d_{S,T}^{\lam}d_{U,V}^{\mu}=\sum\limits_{\epsilon\in\Lambda,X,Y\in M(\epsilon)} R_{(S,T,\lam),(U,V,\mu),(X,Y,\epsilon)}d_{X,Y}^{\epsilon}.$$ The following lemma about structure constants will play an important role in determining the projective cell modules of Frobenius cellular algebras. \begin{lem}\label{3.1} For arbitrary $\lam,\mu\in\Lambda$ and $S,T,P,Q\in M(\lam)$, $U,V\in M(\mu)$ and $a\in A$, the following hold: \begin{enumerate} \item[(1)] $D_{U,V}^{\mu}C_{S,T}^{\lam}=\sum\limits_{\epsilon\in \Lambda, X,Y\in M(\epsilon)}r_{(S,T,\lam),(Y,X,\epsilon),(V,U,\mu)}D_{X,Y}^{\epsilon}.$ \item[(2)] $D_{U,V}^{\mu}C_{S,T}^{\lam}=\sum\limits_{\epsilon\in \Lambda, X,Y\in M(\epsilon)}R_{(Y,X,\epsilon),(U,V,\mu),(T,S,\lam)}C_{X,Y}^{\epsilon}.$ \item[(3)] $aD_{U,V}^{\mu}\equiv \sum\limits_{U'\in M(\mu)}r_{i(\alpha^{-1}(a))}(U,U')D_{U',V}^{\mu}\,\,\,(\mod A_{D}(>\mu))$. \smallskip \item[(4)] $D_{P,Q}^{\lam}C_{S,T}^{\lam}=0\,\,\,\, if \,\,\,Q\neq S.$ \smallskip \item[(5)] $D_{U,V}^{\mu}C_{S,T}^{\lam}=0 \,\,\,\,if \,\,\,\mu\nleq \lam.$ \smallskip \item[(6)] $D_{T,S}^{\lam}C_{S,Q}^{\lam}=D_{T,P}^{\lam}C_{P,Q}^{\lam}.$ \smallskip \item[(7)] $C_{S,T}^{\lam}d_{U,V}^{\mu}=\sum\limits_{\epsilon\in \Lambda, X,Y\in M(\epsilon)}r_{(Y,X,\epsilon),(S,T,\lam),(V,U,\mu)}d_{X,Y}^{\epsilon}.$ \item[(8)] $C_{S,T}^{\lam}d_{U,V}^{\mu}=\sum\limits_{\epsilon\in \Lambda, X,Y\in M(\epsilon)}R_{(U,V,\mu),(Y,X,\epsilon),(T,S,\lam)}C_{X,Y}^{\epsilon}.$ \item[(9)] $d_{U,V}^{\mu}a\equiv \sum\limits_{V'\in M(\mu)}r_{\alpha(a)}(V,V')d_{U,V'}^{\mu}\,\,\,\,\,\,\,\,\,\,\,(\mod A_{d}(>\mu))$. \smallskip \item[(10)] $C_{S,T}^{\lam}d_{P,Q}^{\lam}=0\,\, if \,\,T\neq P.$ \smallskip \item[(11)] $C_{S,T}^{\lam}d_{U,V}^{\mu}=0 \,\,\,\,if\,\,\, \mu\nleq \lam.$ \smallskip \item[(12)] $C_{S,T}^{\lam}d_{T,P}^{\lam}=C_{S,Q}^{\lam}d_{Q,P}^{\lam}.$ \end{enumerate} \end{lem} \begin{proof} (1), (3), (4), (5), (7), (8), (9), (11) have been obtained in \cite{L3}. (2), (8) are proved similarly as (1), (7), respectively. (6) is a direct corollary of (1) and Definition \ref{2.4}. Finally, (12) is obtained similarly as (6). \end{proof} It follows from Definition \ref{2.4} and Lemma \ref{3.1} (1), (3) that for arbitrary elements $S,T,P,Q\in M(\lam)$, $$D_{S,T}^\lam D_{P,Q}^\lam \equiv \Psi(T,P)D_{S,Q}^\lam\,\,\,\, (\rm mod\,\,\, A_D(>\lam)),$$ where $\Psi(T,P)\in K$ depends only on $T$ and $P$. Applying $\alpha^{-1}$ on both sides of the above equation, we obtain $$d_{S,T}^\lam d_{P,Q}^\lam \equiv \Psi(T,P)d_{S,Q}^\lam\,\,\,\, (\rm mod\,\,\, A_d(>\lam)).$$ Take an order on $M(\lam)$ the same as in the definition of $G(\lam)$. Then the matrix $(\Psi(S, T))$ will be denoted by $G'(\lam)$. Let $W_C(\lam)$ be a cell module. Define $\varphi_{ST}\in {\rm End}_K(W_C(\lam))$ by $\varphi_{ST}(C_X)=\delta_{SX}C_T$ for arbitrary $S, T\in M(\lam)$. The following lemma reveals a relation among $I(\varphi_{ST})$, $\Phi_{\lam}$ and $\Psi_{\lam}$. \begin{lem}\label{3.2} If $W_C(\lam)$ is simple, then $I(\varphi_{ST})=c_{ST}{\rm id}_{W_C(\lam)},$ where $$c_{ST}=\sum\limits_{Q\in M(\lam)}\Phi(T,Q)\Psi(S,Q).$$ \end{lem} \begin{proof} It follows from $W_C(\lam)$ being simple and Schur's Lemma that $$I(\varphi_{ST})=c_{ST}{\rm id}_{W_C(\lam)},\eqno(3.1)$$ where $c_{ST}\in K$. Let $\rho_{\lam}$ be the representation afforded by $W_C(\lam)$. Then for $a\in A$, we have $$aC_X=\sum\limits_{Y\in M(\lam)}\rho_{\lam}(a)_{YX}C_Y.$$ A direct computation gives that $$I(\varphi_{ST})(C_X)=\sum_{\eta\in \Lambda, \,P, \,Q\in M(\eta), \,\,Y\in M(\lambda)}\rho_{\lam}(D_{Q, P}^{\eta})_{SX}\rho_{\lam}(C_{P,Q}^{\eta})_{YT}C_Y.\eqno(3.2)$$ Combining (3.1) and (3.2) yields that $$\sum_{\eta\in \Lambda, P, Q\in M(\eta), Y\in M(\lambda)}\rho_{\lam}(D_{Q, P}^{\eta})_{SX}\rho_{\lam}(C_{P,Q}^{\eta})_{XT}=c_{ST}.\eqno(3.3)$$ By the definition of cell modules, $$C_{P, Q}^{\eta}C_{T}=r_{C_{P,Q}^{\eta}}(X,T)C_X+\sum\limits_{X'\neq X}r_{X'}C_{X'},$$ where $r_{C_{P,Q}^{\eta}}(X,T)\in K$ is defined in Definition \ref{2.1}(C3) and $r_{X'}\in K$. This implies that $$\rho_{\lam}(C_{P,Q}^{\eta})_{XT}=r_{(P, Q, \eta),(T,T,\lam),(X,T,\lam)}.\eqno(3.4)$$ On the other hand, it follows from Lemma \ref{3.1}(2) that $$\rho_{\lam}(D_{Q, P}^{\eta})_{SX}=R_{(S,S,\lam),(Q, P, \eta),(S,X,\lam)}.\eqno(3.5)$$ We have from (3.3), (3.4) and (3.5) that $$c_{ST}=\sum\limits_{\eta\in\Lambda, P,Q\in M(\eta)}r_{(P, Q, \eta),(T,T,\lam),(X,T,\lam)}R_{(S,S,\lam),(Q, P, \eta),(S,X,\lam)}.\eqno(3.6)$$ Again by Lemma \ref{3.1}, for any $\epsilon\in\Lambda$, if $\epsilon\nleq\lam$, then $R_{(S,S,\lam),(Y,X,\epsilon),(S,S,\lam)}=0$, if $\lam\nleq\epsilon$, then $r_{(X,Y,\epsilon),(S,S,\lam),(S,S,\lam)}=0$. Thus Definition \ref{2.4} and Lemma \ref{3.1} force (3.6) being \begin{eqnarray*} c_{ST}&=& \sum\limits_{P, Q\in M(\lam)}r_{(P, Q, \lambda),(T,T,\lam),(X,T,\lam)}R_{(S,S,\lam),(Q, P, \lambda),(S,X,\lam)}\\ &=&\sum\limits_{Q\in M(\lam)}r_{(X,Q,\lam),(T,T,\lam),(X,T,\lam)}R_{(S,S,\lam),(Q,X,\lam),(S,S,\lam)}\\ &=&\sum\limits_{Q\in M(\lam)}\Phi(T,Q)\Psi(S,Q). \end{eqnarray*} We complete the proof. \end{proof} Now we are able to describe the main result of this section. \begin{thm}\label{3.3} Let $A$ be a finite dimensional Frobenius cellular $K$-algebra and $W_C(\lam)$ a simple cell module. Then $W_C(\lam)$ is projective if and only if there exist $S, \,T\in M(\lam)$ such that $\Psi(S,T)\neq 0$. \end{thm} \begin{proof} It follows from Corollary \ref{2.3} that $W_C(\lam)$ is projective if and only if there exist $X, Y\in M(\lam)$ such that $I(\varphi_{XY})\neq 0$. By Lemma \ref{3.2}, this is equivalent to $c_{XY}\neq 0$. Fix an order on $M(\lam)$ and denote the matrix $(c_{XY})$ by $I_{\lam}$. Then $W_C(\lam)$ is projective if and only if $I_{\lam}\neq 0$. On the other hand, we have from Lemma \ref{3.2} that $G'(\lam)G(\lam)=I_{\lam}$. The simplicity of $W_C(\lam)$ implies that $G(\lam)$ is invertible. Thus $I_{\lam}\neq 0$ if and only if $G'(\lam)\neq 0$, that is, there exist $S, \,T\in M(\lam)$ such that $\Psi(S,T)\neq 0$. \end{proof} We can obtain from this theorem a necessary condition for a cell module being projective. \begin{cor}\label{3.4} If $G'(\lam)=0$, then $W_C(\lam)$ is not a projective module. \end{cor} \begin{proof} The corollary follows from Theorem \ref{3.3} and \cite[Corollary 1.2]{KX8} clearly. \end{proof} \begin{remark}\label{3.5} Using the right dual basis $\{d_{X,Y}^{\epsilon}\mid \epsilon\in\Lambda, X, Y\in M(\epsilon)\}$, for each $\lam\in \Lambda$ we can define an $A$-module $W_d(\lam)$ as follows. As a $K$-basis, $W_d(\lam)$ has a basis $\{d_S\mid S\in M(\lam)\}$. The $A$-action is defined by $$ad_S=\sum_{S'\in M(\lam)}r_{i(a)}(S,S')d_{s'}.$$ Then by a similar way, we can prove that if $W_d(\lam)$ is simple, then it is projective if and only if $\lam\in\Lambda_0$. \end{remark} Now let us apply Theorem \ref{3.3} to an example which was constructed by K\"{o}nig and Xi in \cite{KX2}. \begin{example} Let $K$ be a field. Let us take $\lam\in K$ with $\lam\neq 0$ and $\lam\neq 1$. Let $$A=K\langle a, b, c, d\rangle/I,$$ where $I$ is generated by $$a^2, b^2, c^2, d^2, ab, ac, ba, bd, ca, cd, db, dc, cb-\lam bc, ad-bc, da-bc.$$ Let $\Lambda=\{1, 2, 3\}$. If we define $\tau$ by $\tau(1)=\tau(a)=\tau(b)=\tau(c)=\tau(d)=0$ and $\tau(bc)=1$ and define an involution $i$ on $A$ to be fixing $a$ and $d$, but interchanging $b$ and $c$, then $A$ is a Frobenius cellular algebra with a cellular basis \[ \begin{matrix} \begin{matrix} bc \end{matrix} ;& \begin{matrix} a & b\\ c & d\end{matrix} \,\,; & \begin{matrix} 1 \end{matrix}. \end{matrix} \] The left dual basis is \[ \begin{matrix} \begin{matrix} 1 \end{matrix} ;& \begin{matrix} d & c/\lam\\ b & a\end{matrix}\,\, ; & \begin{matrix} bc \end{matrix}. \end{matrix} \] It is easy to check that $W_C(1)=0$, $W_C(2)=0$ and $W_C(3)$ is simple. However, $G'(3)=0$. This implies that $W_C(3)$ is not projective. Thus none of cell modules is projective. \end{example} \bigskip \section{A special case} \label{xxsec4} Throughout this section, we assume that $A$ is a finite dimensional Frobenius cellular algebra with both the dual bases being cellular with respect to the opposite order on $\Lambda$. Under this assumption, we obtained the following result about the Nakayama automorphism $\alpha$ in \cite{L3}. \begin{lem}\cite[Theorem 3.1]{L3}\label{4.1} For arbitrary $\lam\in \Lambda$ and $S, T\in M(\lam)$, we have $$\alpha(C_{S,T}^{\lam})\equiv C_{S,T}^{\lam}\quad\quad (\mod A(<\lam)).$$ \end{lem} In order to generalize the so-called Schur elements to Frobenius case, we prove a lemma first. \begin{lem}\label{4.2} Let $A$ be a Frobenius cellular algebra with cell datum $(\Lambda, M, C, i)$. For every $\lam\in\Lambda$ and $S,T\in M(\lam)$, we have $$D_{S,T}^{\lam}C_{T,S}^{\lam}D_{S,T}^{\lam}C_{T,S}^{\lam}=\sum_{S'\in M(\lam)}\Phi(S',T)\Psi(S',T)D_{S,T}^{\lam}C_{T,S}^{\lam}.$$ \end{lem} \begin{proof} By Lemma \ref{3.1} (3), (5), we have \begin{eqnarray*} D_{S,T}^{\lam}C_{T,S}^{\lam}D_{S,T}^{\lam}C_{T,S}^{\lam} &=&D_{S,T}^{\lam}(C_{T,S}^{\lam}D_{S,T}^{\lam})C_{T,S}^{\lam}\\&=&\sum\limits_{S'\in M(\lam)}r_{i(\alpha^{-1}(C_{T, S}^{\lam}))}(S, S')D_{S, T}^{\lam}D_{S', T}^{\lam}C_{T, S}^{\lam}. \end{eqnarray*} It follows from Definition~\ref{2.4} and Lemma~\ref{4.1} that $$r_{i(\alpha^{-1}(C_{T, S}^{\lam}))}(S, S')=\Phi(S', T).$$ Again by Lemma \ref{3.1}(5), the desired equation follows. \end{proof} We have from Lemma \ref{3.1} that $D_{S,T}^{\lam}C_{T,S}^{\lam}$ is independent of $T$. Then for any $\lam\in\Lambda$, we can define a constant $k_{\lam}$ as follows. \begin{dfn}\label{4.3} For $\lam\in\Lambda$, take an arbitrary $T\in M(\lam)$. Define $$k_{\lam}=\sum\limits_{X\in M(\lam)}\Phi(X,T)\Psi(X,T).$$ \end{dfn} The following lemma reveals the relation among $G(\lam)$, $G'(\lam)$ and $k_{\lam}$. This result is a bridge which connects $k_{\lam}$ with $c_{ST}$. \begin{lem}\label{4.4} Let $\lam\in\Lambda$. Fix an order on the set $M(\lam)$. Then $G(\lam)G'(\lam)=k_{\lam}E$, where $E$ is the identity matrix. \end{lem} \begin{proof} Clearly, we only need to show that $\sum\limits_{X\in M(\lam)}\Phi(X,S)\Psi(X,T)=0$ for arbitrary $S, T\in M(\lam)$ with $S\neq T$. Let us consider $D_{S,S}^{\lam}C_{S,S}^{\lam}D_{S,S}^{\lam}C_{T,S}^{\lam}$. On one hand, it is zero by Lemma \ref{3.1}. On the other hand, computing it as that in Lemma \ref{4.2} yields that $$D_{S,S}^{\lam}C_{S,S}^{\lam}D_{S,S}^{\lam}C_{T,S}^{\lam}=\sum\limits_{X\in M(\lam)}\Phi(X,S)\Psi(X,T)D_{S,S}^{\lam}C_{S,S}^{\lam}.$$ It follows from $D_{S,S}^{\lam}C_{S,S}^{\lam}\neq 0$ that $\sum\limits_{X\in M(\lam)}\Phi(X,S)\Psi(X,T)=0.$ \end{proof} \begin{cor}\label{4.5} Keep notations as above, then $c_{ST}={\rm Tr}(\varphi_{ST})k_{\lam}$, where ${\rm Tr}$ denotes the usual matrix trace. \end{cor} \begin{proof} Note that ${\rm Tr}(\varphi_{ST})=\delta_{ST}$. Then by Lemma \ref{3.2}, Definition \ref{4.3} and Lemma \ref{4.4}, the corollary follows. \end{proof} The constants $k_{\lam}$ could be viewed as generalizations of Schur elements when $W_C(\lam)$ is simple. We are in a position to give the main result of this section. It is similar to that in symmetric case. \begin{thm}\label{4.6} Let $A$ be a Frobenius cellular algebra with $i(D_{S,T}^{\lam})=D_{T,S}^{\lam}$ and $i(d_{S,T}^{\lam})=d_{T,S}^{\lam}$ for arbitrary $\lam\in \Lambda$ and $S,T\in M(\lam)$. Then the cell module $W_C(\lam)$ is projective if and only if $k_\lam\neq 0$. \end{thm} \begin{proof} By \cite[Corollary 1.2]{KX2}, if $W_C(\lam)$ is projective, then $W_C(\lam)$ is irreducible. It follows from Corollary \ref{2.3} that there exist $S, T\in M(\lam)$ such that $I(\varphi_{ST})\neq 0,$ or $c_{ST}\neq 0$. This implies that $k_{\lam}\neq 0$ by Corollary \ref{4.5}. Conversely, assume that $k_{\lam}\neq 0$. Then $W_C(\lam)$ is irreducible by Lemma \ref{4.4}. Again by Corollary \ref{2.3}, $W_C(\lam)$ is projective. \end{proof} \begin{cor} $W_C(\lam)$ is projective if and only if so is $W_d(\lam)$. \end{cor} \begin{proof} By Theorem \ref{4.6}, we only need to prove that $W_d(\lam)$ is projective if and only if $k_{\lam}\neq 0$. In fact, if $k_{\lam}\neq 0$, then Lemma \ref{4.4} forces $W_d(\lam)$ to be simple. Thus $W_d(\lam)$ being projective follows. Conversely, if $W_d(\lam)$ is projective, then \cite[Corollary 1.2]{KX2} implies that $W_d(\lam)$ is simple, that is, $G'(\lam)$ is invertible. Moreover, we have from Remark \ref{3.5} that $G(\lam)\neq 0$. Hence we conclude by Lemma \ref{4.4} that $k_{\lam}\neq 0$. \end{proof} \begin{remark} Using the left dual basis $D_{X,Y}^{\epsilon}$, we can also define modules $W_D(\lam)$. It could be proved that $W_D(\lam)$ is projective if and only if so is $W_d(\lam)$. We omit the details here. \end{remark} \bigskip\bigskip

https://arxiv.org/abs/1302.3764

Acoustic multipole sources from the Boltzmann equation

By adding a particle source term in the Boltzmann equation of kinetic theory, it is possible to represent particles appearing and disappearing throughout the fluid with a specified distribution of particle velocities. By deriving the wave equation from this modified Boltzmann equation via the conservation equations of fluid mechanics, multipole source terms in the wave equation are found. These multipole source terms are given by the particle source term in the Boltzmann equation. To the Euler level in the momentum equation, a monopole and a dipole source term appear in the wave equation. To the Navier-Stokes level, a quadrupole term with negligible magnitude also appears.

\section{Introduction} Acoustic multipoles are oscillating sources that emit acoustic fields of different directivities. These sources can be either point sources, localized at single points in space, or they can be distributions throughout the medium. The the first three orders of multipoles are the most well-known: \emph{Monopoles}, \emph{dipoles}, and \emph{quadrupoles} at zeroth, first, and second order, respectively. When these three types of multipole sources appear as source terms in the wave equation, they usually originate from terms in the conservation equations of fluid mechanics. For instance, monopoles are linked to a source term in the mass conservation equation (also known as the continuity equation), which represents mass appearing and disappearing throughout the fluid as a function of time. This can model pulsations of small bodies throughout the fluid~\cite{howe03}. However, an alternative approach is to add a particle source term in the \emph{Boltzmann equation}, which is more fundamental than the fluid conservation equations which can be derived from it. Adding such a source term to the Boltzmann equation allows specifying the velocity distribution of particles that appear and disappear throughout the fluid. This approach is therefore similar to, but more general than, the aforementioned method of adding a mass source term, and can therefore model more general vibrations of small bodies in the fluid. Such a particle source term was recently examined for the lattice Boltzmann method~\cite{viggen13}, a computational fluid dynamics method based on the fully discretised Boltzmann equation. It was found that such an approach results in a wave equation with non-vanishing monopole, dipole, and quadrupole source terms. This article will similarly examine particle source terms, but in the classic non-discretised Boltzmann equation. \section{Acoustic multipole sources} Mathematically, multipoles are related to source terms in the wave equation, \begin{equation} \begin{split} \label{eq:wavesourceterms} \left( \frac{1}{c_0^2} \pderiv{}{t} - \nabla^2 \right) p(\vect{x},t) &= T_0(\vect{x},t) + \pderiv{}{x_i} T_i(\vect{x},t) \\ &\quad + \frac{\partial^2}{\partial x_i \partial x_j} T_{ij}(\vect{x},t) + \ldots \; . \end{split} \end{equation} Here, $p$ is the pressure and $c_0$ is the speed of sound. The terms on the right-hand side are multipole source terms. This article makes use of the \emph{index notation} commonly used in the field of fluid mechanics. In this notation, a single index indicates a generic vector element (e.g.\ $T_i$ could be $T_x$, $T_y$, or $T_z$), and multiple indices (as in $T_{ij}$ or $a_i b_j$) indicate generic tensor elements. Repeating indices within a single term implies summation over all possible values of that index. For example, $a_i b_i = a_x b_x + a_y b_y + a_z b_z = \vect{a}\cdot\vect{b}$, and $\partial T_i / \partial x_i = \partial T_x / \partial x + \partial T_y / \partial y + \partial T_z / \partial z = \nabla \cdot \vect{T}$. When indices repeat in this way, the letter used is arbitrary, so that $a_i b_i = a_k b_k$. The general three-dimensional solution to~\eqref{eq:wavesourceterms} is given by an integral over the entire volume of the source terms on the right-hand side~\cite{howe03}, \begin{equation} \begin{split} \label{eq:wavesourcesolution} p(\vect{x},t) = \frac{1}{4\pi} \int \Bigg[ & \frac{ T_0(\vect{y},t-\tfrac{|\vect{x}-\vect{y}|}{c_0}) }{|\vect{x}-\vect{y}|}\\ &+ \frac{\partial}{\partial x_i} \frac{ T_i(\vect{y},t-\tfrac{|\vect{x}-\vect{y}|}{c_0}) }{|\vect{x}-\vect{y}|} \\ &+ \frac{\partial^2}{\partial x_i \partial x_j} \frac{ T_{ij}(\vect{y},t-\tfrac{|\vect{x}-\vect{y}|}{c_0}) }{|\vect{x}-\vect{y}|} \Bigg] \diff{\vect{y}} . \end{split} \end{equation} Thus, $T_0(\vect{x},t)$ indicates the instantaneous monopole strength, $T_i(\vect{x},t)$ the $i$-dipole strength, and $T_{ij}$ the $ij$-quadrupole strength. As mentioned above, monopole sources can be modeled by adding a mass source term to the mass conservation equation, \begin{equation} \label{eq:massconservation} \frac{\partial \rho}{\partial t} + \nabla \cdot \left( \rho \vect{u} \right) = Q, \end{equation} $\rho$ being the mass density, $\vect{u}$ the fluid velocity, and $Q(\vect{x},t)$ the instantaneous mass flux. Dipoles typically originate from the force term in the momentum conservation equation. To the Euler level, this equation is \begin{equation} \label{eq:eulermomentum} \rho \left[ \frac{\partial \vect{u}}{\partial t} + (\vect{u} \cdot \nabla) \vect{u} \right] = - \nabla p + \vect{F}, \end{equation} where $\vect{F}$ represents body forces. Finally, quadrupoles typically originate from the nonlinear term in~\eqref{eq:eulermomentum}. The multipole terms in~\eqref{eq:wavesourceterms} are usually related to the terms in the conservation equations as~\cite{howe03} \begin{equation*} T_0 = \frac{\partial Q}{\partial t}, \qquad T_i = - F_i, \qquad T_{ij} \simeq \rho u_i u_j . \end{equation*} \section{The Boltzmann equation} The Boltzmann equation describes motion of a gas at a finer level of detail than the fluid conservation equations. In this discussion we shall restrict ourselves to its very basics. More details can be found in the literature~\cite{cercignani88, chapman70, hanel04}. The equation evolves the particle \emph{distribution function}, $f(\vect{x},\vect{\xi},t)$, which may be seen as a double density in both physical space and particle velocity space. Thus, it describes the density of particles with position $\vect{x}$ and velocity $\vect{\xi}$ at time $t$. The familiar macroscopic quantities can be recovered as \emph{moments} of $f$, i.e.\ by weighting with some function and integrating over the entire velocity space. The mass density and momentum density are found as the zeroth- and first-order moments, \begin{subequations} \label{eq:moments} \begin{align} \rho(\vect{x},t) &= \int f(\vect{x},\vect{\xi},t) \diff{\vect{\xi}} , \\ \rho \vect{u}(\vect{x},t) &= \int \vect{\xi} f(\vect{x},\vect{\xi},t) \diff{\vect{\xi}} \end{align} \end{subequations} Neglecting body forces and using the BGK collision operator~\cite{bhatnagar54}, the Boltzmann equation is \begin{equation} \label{eq:boltzmann} \pderiv{f}{t} + \vect{\xi} \cdot \nabla f = s - \frac{1}{\tau} \left( f - f^{(0)} \right) . \end{equation} $s(\vect{x},\vect{\xi},t)$ is the aforementioned particle source term which is central to this article. As the left-hand side of the equation is a standard advection equation, $s(\vect{x},\vect{\xi},t)$ describes the rate at which particles are added into the $f(\vect{x},\vect{\xi},t)$ distribution. The final term is the BGK collision operator, which models collisions between particles as a relaxation with a characteristic \emph{relaxation time} $\tau$ to the equilibrium distribution function, \begin{equation} \label{eq:eqdist} f^{(0)}(\vect{x},\vect{\xi},t) = \rho \left( \frac{\rho}{2 \pi p} \right)^{3/2} \,\mathrm{e}^{- \rho|\vect{\xi}-\vect{u}|^2 / 2 p} , \end{equation} with $\rho$ and $\vect{u}$ found from~\eqref{eq:moments}. In this article, $p$ is approximated by~\eqref{eq:isentropicpressure}. As the quantities of mass and momentum are conserved in collisions, substituting $f^{(0)}$ for $f$ in~\eqref{eq:moments} must give the same moments. As a result of this, the BGK collision operator conserves both mass and momentum. The BGK collision operator is a far simpler model of collisions than Boltzmann's original and more accurate collision operator. That simplicity comes with drawbacks, chiefly that the BGK operator slightly mispredicts the Prandtl number~\cite{cercignani88}. This dimensionless number relates the transport coefficients in the momentum equation (viscosity) and the energy equation (conductivity). However, this will not matter as we will neglect the effects of conductivity in this article. From this point on we will assume that the macroscopic variables fluctuate only slightly around rest state values $\rho=\rho_0$, $p=p_0$, $\vect{u}=0$. This allows us to linearise subsequent equations in this article, which is in keeping with the usual assumptions in acoustics. The pressure $p$ in~\eqref{eq:eqdist} is approximated using the common isentropic relation~\cite{howe03, thompson72} \begin{equation} \label{eq:isentropicpressure} \frac{p}{p_0} = \left( \frac{\rho}{\rho_0} \right)^\gamma , \end{equation} $\gamma = (d+2)/2$ being the adiabatic index determined by the degrees of freedom $d$ of the molecules that make up the gas~\cite{thompson72}. In this way, we include the effect of equipartition of energy between translational and inner (i.e.\ rotational and vibrational) degrees of freedom, in the limit of rapid energy transfer. This relation leads to an ideal speed of sound $c_0$ given by \begin{equation} \label{eq:soundspeed} c_0^2 = \pderiv{p}{\rho} = \frac{\gamma p_0}{\rho_0} . \end{equation} It will be useful to introduce an abbreviated notation for the moments of the particle source term $s$, \begin{subequations} \label{eq:sourcemoments} \begin{align} S_0(\vect{x},t) &= \int s(\vect{x},\vect{\xi},t) \diff{\vect{\xi}} , \\ S_i(\vect{x},t) &= \int \xi_i s(\vect{x},\vect{\xi},t) \diff{\vect{\xi}} , \\ S_{ij}(\vect{x},t) &= \int \xi_i \xi_j s(\vect{x},\vect{\xi},t) \diff{\vect{\xi}} , \end{align} \end{subequations} and so forth. $S_0(\vect{x},t)$ represents the instantaneous mass flux of the particles at $\vect{x}$, $S_i$ is associated with odd symmetries of $s$ in velocity space, and $S_{ij}$ is similarly associated with various even symmetries. \section{Fluid conservation equations} It is possible to find the conservation equations of fluid mechanics from the Boltzmann equation~\eqref{eq:boltzmann}. To find the Euler equations, we could simply take the zeroth and first moments of~\eqref{eq:boltzmann} under the assumption that $f \simeq f^{(0)}$. However, to find the momentum equation to the Navier-Stokes level, so that it includes the stress tensor term, we must resort to the \emph{Chapman-Enskog expansion}. This is a technique used to derive the fluid conservation equations from the Boltzmann equation. It is discussed throughout the literature with varying approaches and varying levels of complexity~\cite{cercignani88, chapman70, grad63, hanel04, dellar01}. In the following derivation, we use a moment-based approach~\cite{dellar01}. Two mathematical techniques are used in this expansion. First, the distribution function $f$ is approximated as a perturbation expansion around equilibrium $f^{(0)}$ in a smallness parameter $\epsilon$. Second, a multi-scale expansion of time is performed. In mathematical notation, \begin{align*} f &= f^{(0)} + \epsilon f^{(1)} + \epsilon^2 f^{(2)} + \ldots \; , \\ \pderiv{}{t} &= \pderiv{}{t_0} + \epsilon \pderiv{}{t_1} + \ldots \; . \end{align*} The smallness parameter $\epsilon$ is associated with the dimensionless Knudsen number $\mathrm{Kn} = l_\mathrm{mfp} / L$, relating the mean free path $l_\mathrm{mfp}$ in the gas to a macroscopic length scale $L$. Thus, $f^{(n+1)}$ is of one order higher in the Knudsen number than $f^{(n)}$. A dimensional analysis~\cite{hanel04} reveals that the relaxation time is also at first order of smallness, so that $\tau = \epsilon \tau$. As previously explained, the density and momentum is fully contained in $f^{(0)}$, so that \begin{equation} \int f^{(n)} = \int \xi_i f^{(n)} = 0 \qquad \text{for } n > 0 . \end{equation} Expanding the Boltzmann equation in this way and truncating the expansion to $\mathcal{O}(\epsilon)$, we find \begin{equation} \begin{split} \left( \pderiv{}{t_0} + \epsilon \pderiv{}{t_1} + \xi_i \pderiv{}{x_i} \right) \left( f^{(0)} + \epsilon f^{(1)} \right) \\ = s - \frac{1}{\epsilon\tau} \left( \epsilon f^{(1)} + \epsilon^2 f^{(2)} \right) . \end{split} \end{equation} Gathering these terms according to their order of smallness, we find \begin{subequations} \label{eq:smallnessterms} \begin{align} \mathcal{O}(\epsilon^0)\!: && \left( \pderiv{}{t_0} + \xi_i \pderiv{}{x_i} \right) f^{(0)} &= s - \frac{1}{\tau} f^{(1)} , \label{eq:smallness0terms} \\ \mathcal{O}(\epsilon^1)\!: && \!\!\!\!\! \pderiv{f^{(0)}}{t_1} + \left( \pderiv{}{t_0} + \xi_i \pderiv{}{x_i} \right) f^{(1)} &= - \frac{1}{\tau} f^{(2)} . \label{eq:smallness1terms} \end{align} \end{subequations} To derive the mass and momentum equations to the Euler level, only the $\mathcal{O}(\epsilon^0)$ terms are needed. The Navier-Stokes corrections to these equations are found by also including the $\mathcal{O}(\epsilon^1)$ terms in the derivation. Similarly, it is possible to find further corrections at $\mathcal{O}(\epsilon^2)$ (known as the Burnett corrections) and beyond (super-Burnett), although these further corrections are negligible in practical cases~\cite{chapman70, grad63} and have historically been viewed with some suspicion~\cite{grad63}. Deriving the conservation equations from~\eqref{eq:smallnessterms} requires the second and third moments of $f^{(0)}$~\cite{dellar01}. In linearised form, these are \begin{subequations} \label{eq:eqmoments} \begin{align} \int \xi_i \xi_j f^{(0)} \diff{\vect{\xi}} &\simeq p \delta_{ij} , \\ \int \xi_i \xi_j \xi_k f^{(0)} \diff{\vect{\xi}} &\simeq p_0 ( u_i \delta_{jk} + u_j \delta_{ik} + u_k \delta_{ij} ) , \end{align} \end{subequations} where $\delta_{ij}$ is the Kronecker delta. Taking the zeroth to second moments of~\eqref{eq:smallness0terms}, using~\eqref{eq:eqmoments}, and linearising, we find \begin{subequations} \begin{align} \pderiv{\rho}{t_0} + \rho_0 \pderiv{u_i}{x_i} &= S_0 , \label{eq:small0mom0} \\ \rho_0 \pderiv{u_i}{t_0} + \pderiv{p}{x_i} &= S_i , \label{eq:small0mom1} \\ \delta_{ij} \pderiv{p}{t_0} + p_0 \left( \pderiv{u_i}{x_j} + \pderiv{u_j}{x_i} + \delta_{ij} \pderiv{u_k}{x_k} \right) &= S_{ij} - \frac{1}{\tau} \Pi_{ij}^{(1)} , \label{eq:small0mom2} \end{align} \end{subequations} where $\Pi_{ij}^{(1)} = \int \xi_i \xi_j f^{(1)} \diff{\vect{\xi}}$. With the $\mathcal{O}(\epsilon^0)$ approximation $\partial / \partial t_0 = \partial / \partial t$, the two first equations are equivalent to linearised versions of the mass equation~\eqref{eq:massconservation} and the Euler momentum equation~\eqref{eq:eulermomentum}, with $S_0$ and $S_i$ in the place of $Q$ and $F_i$, respectively. Taking the zeroth and first moments of~\eqref{eq:smallness0terms} and linearising, we find $\mathcal{O}(\epsilon)$ corrections to the above equations, \begin{subequations} \begin{align} \pderiv{\rho}{t_1} &= 0 , \label{eq:small1mom0} \\ \rho_0 \pderiv{u_i}{t_1} + \pderiv{\Pi_{ij}^{(1)}}{x_j} &= 0 . \label{eq:small1mom1} \end{align} \end{subequations} The sum $\eqref{eq:small0mom0} + \epsilon \eqref{eq:small1mom0}$ directly leads to the linearised mass equation, given below in~\eqref{eq:massconservation2}. Similarly, $\eqref{eq:small0mom1} + \epsilon \eqref{eq:small1mom1}$ leads to the momentum equation, though the unknown tensor $\Pi^{(1)}_{ij}$. $\Pi^{(1)}_{ij}$ can be explicitly related to the Navier-Stokes stress tensor using~\eqref{eq:small0mom2}. The pressure time derivative can be rewritten assuming a nearly isentropic process and using~\eqref{eq:small0mom0}, resulting in \begin{equation} \label{eq:densitytopressure} \pderiv{p}{t_0} = \pderiv{p}{\rho} \pderiv{\rho}{t_0} = c_0^2 \left( S_0 - \rho_0 \pderiv{u_k}{x_k} \right) . \end{equation} Substituting for the speed of sound using~\eqref{eq:soundspeed}, the diagonal terms in~\eqref{eq:small0mom2} become \begin{equation} \begin{split} \pderiv{p}{t_0} + p_0 \pderiv{u_k}{x_k} &= \left( p_0 - \rho_0 c_0^2 \right) \pderiv{u_k}{x_k} + c_0^2 S_0 \\ &= - p_0 ( \gamma - 1 ) \pderiv{u_k}{x_k} + c_0^2 S_0 . \end{split} \end{equation} Thus, we find \begin{equation} \begin{split} \Pi_{ij}^{(1)} &= - p_0 \tau \left[ \pderiv{u_j}{x_i} + \pderiv{u_i}{x_j} + \delta_{ij} (1-\gamma) \pderiv{u_k}{x_k} \right] \\ &\quad + \tau \left( S_{ij} - c_0^2 S_0 \right) . \end{split} \end{equation} The mass and momentum equations can now be explicitly found as described above, \begin{subequations} \label{eq:conservation2} \begin{align} \pderiv{\rho}{t} + \rho_0 \pderiv{u_i}{x_i} &= S_0 , \label{eq:massconservation2} \\ \rho_0 \pderiv{u_i}{t} + \pderiv{p}{x_i} &= S_i - \frac{\mu}{p_0} \pderiv{(S_{ij} - \delta_{ij} c_0^2 S_0)}{x_j} + \pderiv{\sigma'_{ij}}{x_j} . \label{eq:momentumconservation2} \end{align} \end{subequations} The momentum equation contains a deviatoric stress tensor \begin{equation} \sigma'_{ij} = \mu \left( \pderiv{u_i}{x_j} + \pderiv{u_j}{x_i} - \frac{2}{3} \delta_{ij} \pderiv{u_k}{x_k} \right) + \mu_B \delta_{ij} \pderiv{u_k}{x_k} , \end{equation} with shear viscosity $\mu = p_0 \tau$ and bulk viscosity $\mu_B/\mu = (5/3 - \gamma)$. This value of the bulk viscosity in the limit of rapid transfer of energy between translational and inner degrees of freedom has previously been found using more rigorous kinetic theory~\cite{morse64}. Comparing~\eqref{eq:massconservation2} to the classical mass equation~\eqref{eq:massconservation}, $S_0$ appears in the place of the mass flux $Q$, which could be expected considering the interpretation of $S_0$ as a mass flux. Comparing~\eqref{eq:momentumconservation2} to the Euler-level momentum equation~\eqref{eq:eulermomentum}, $S_i$ appears in the place of the body force, which had been neglected from the Boltzmann equation. $S_{ij}$ is also present inside a source term with a single spatial derivative and a small coefficient $\mu/p_0$ in front. \section{The wave equation} The wave equation can be found as usual from the mass and momentum equations as $\partial\eqref{eq:massconservation2} / \partial t - \partial\eqref{eq:momentumconservation2} / \partial x_i$. Using the isentropic relation~\eqref{eq:densitytopressure} in the wave equation operator, and thus neglecting sound absorption due to thermal conduction and relaxation~\cite{blackstock00}, we find \begin{align} \left( \frac{1}{c_0^2} \pderiv[2]{}{t} - \nabla^2 \right) p &= \pderiv{S_0}{t} - \pderiv{S_i}{x_i} + \frac{\mu}{p_0} \frac{\partial^2 (S_{ij} - \delta_{ij} c_0^2 S_0)}{\partial x_i \partial x_j} \nonumber \\ &\quad+ \frac{\partial^2 \sigma'_{ij}}{\partial x_i \partial x_j} . \end{align} Comparing with~\eqref{eq:wavesourceterms} and~\eqref{eq:wavesourcesolution}, we find a monopole strength $T_0 = \partial S_0 / \partial t$, a dipole strength $T_i = - S_i$, and a quadrupole strength which is not fully resolved but involves $S_{ij}$, $\delta_{ij} S_0$, and $\sigma'_{ij}$. The deviatoric stress tensor $\sigma'_{ij}$ can be resolved as \begin{equation} \frac{\partial^2 \sigma'_{ij}}{\partial x_i \partial x_j} = \mu (3 - \gamma) \nabla^2 \pderiv{u_k}{x_k} = \frac{\mu (3 - \gamma)}{\rho_0} \nabla^2 \left( S_0 - \frac{1}{c_0^2} \pderiv{p}{t} \right) . \end{equation} The last term in the parenthesis contributes to sound absorption and is neglected in line with previous approximations. The first parenthetical term contributes to the quadrupole strength. Using the property $\nabla^2 = \delta_{ij} (\partial^2 / \partial x_i \partial x_j)$, we find a fully resolved isentropic wave equation \begin{equation} \begin{split} \left( \frac{1}{c_0^2} \pderiv[2]{}{t} - \nabla^2 \right) p &= \pderiv{S_0}{t} - \pderiv{S_i}{x_i} \\ &\quad + \frac{\mu}{p_0} \frac{\partial^2 (S_{ij} - 3 \delta_{ij} p_0 S_0 / \rho_0)}{\partial x_i \partial x_j} . \end{split} \end{equation} Finally, we find the quadrupole strength as $T_{ij} = (\mu / p_0) (S_{ij} - 3 \delta_{ij} p_0 S_0 / \rho_0)$. The coefficient $\mu/p_0$ in front of the quadrupole strength is typically on the order of $\SI{E-10}{\second}$ in gases. Its small magnitude means that the quadrupoles generated by the Boltzmann equation source term $s$ tend to be negligible compared to the monopoles and dipoles. \section{Summary and conclusion} A source term $s$ in the Boltzmann equation represents particles appearing or disappearing throughout the fluid with some distribution of particle velocities. As adding a source term to the mass equation allows modeling pulsations of small bodies throughout the fluid~\cite{howe03}, a source term in the Boltzmann equation would allow modeling more general vibrations of such small bodies. From this modified Boltzmann equation, the mass and momentum conservations equations~\eqref{eq:conservation2} were derived under the common acoustic approximation of linearity and constant entropy. These equations gain source terms given by the moments~\eqref{eq:sourcemoments} of $s$. The mass equation gains a source term $S_0$. To the Euler level, the momentum equation gains a source term $S_i$, and to the Navier-Stokes level, it gains a source term involving $S_{ij}$ and $\delta_{ij} S_0$. The wave equation derived from these two conservation equations contains multipole source terms. The monopole strength is $\partial S_0 / \partial t$ and the dipole strength is $-S_i$. The quadrupole strength, which comes out of the Navier-Stokes level of the momentum equation, involves $S_{ij}$ and $\delta_{ij} S_0$, but has a negligible magnitude. That the quadrupole strength is so much smaller than the monopole and dipole strength could be expected, as the Navier-Stokes level terms are one order higher in the small Knudsen number than the Euler-level terms. Similarly, going to the $\mathcal{O}(\mathrm{Kn}^2)$ Burnett level might lead to a $\partial^2 S_{ijk} / \partial x_j \partial x_k$ term in the momentum equation, leading to an octupole term in the wave equation. However, since this term would be at $\mathcal{O}(\mathrm{Kn}^2)$, it would be even more negligible than the quadrupole term. Going back to the point of modeling general vibrations of small bodies throughout the fluid, this analysis indicates that the vibrations of such small bodies can radiate as monopoles and dipoles, but only very weakly as higher-order multipoles. Comparing this analysis with the analogous analysis for the lattice Boltzmann method~\cite{viggen13}, we find that the analogous source term in the lattice Boltzmann equation can only radiate quadrupoles effectively due to a fortuitous discretisation error that occurs when discretising the Boltzmann equation~\eqref{eq:boltzmann} in space and time using the first-order rectangle method. Discretising with the trapezoidal method~\cite{dellar01}, which results in a scheme fully consistent with~\eqref{eq:boltzmann}, would therefore also lead to a vanishing quadrupole term similarly to what has been shown here. \bibliographystyle{apsrev4-1}

https://arxiv.org/abs/0810.2773

A note on quasi-Gorenstein rings

In this paper, after giving a criterion for a Noetherian local ring to be quasi-Gorenstein, we obtain some sufficient conditions for a quasi- Gorenstein ring to be Gorenstein. In the course, we provide a slight generalization of a theorem of Evans and Griffith.

\section{Introduction} Let $R$ be a Noetherian local ring with maximal ideal $\mathfrak{m}$ and residue field $k$. Throughout, we use $E_{R}(k)$ to denote the injective hull of $k$. Also, for a non-negative integer $i$, we use $H_{\mathfrak{m}}^{i}(R)$ to denote the $i$-th local cohomology module of $R$ with respect to $\mathfrak{m}$. Following \cite{AG}, we say that $R$ is a quasi-Gorenstein ring if $H_{\mathfrak{m}}^{dim R}(R)\cong E_{R}(k)$. In this paper, we investigate quasi-Gorenstein rings in terms of linkage. By the concept of linkage, we mean the complete intersection linkage. More precisely, for ideals $I$ and $J$, we say that $I$ is linked to $J$ (written $I\sim J$) if there is an $R$-sequence $x_{1},\ldots,x_{n}$ in $I\cap J$ such that $I=(x_{1},\ldots,x_{n}):J$ and $J=(x_{1},\ldots,x_{n}):I$. To begin, using Cohen's structure theorem in conjunction with the linkage theory, we provide a criterion which shows that the ring $R$ is quasi-Gorenstein if and only if $\widehat{R}$, the $\mathfrak{m}$-adic completion of $R$, is a certain specialization of a Gorenstein local ring. This enables us to compare (in 2.4) the Gorenstein loci and the Cohen--Macaulay loci of a quasi-Gorenstein ring. The question asking when a quasi-Gorenstein ring is Gorenstein constitutes a lot of interests. In this direction, we use the above mentioned criterion to establish the results 2.5 and 2.7 which provide some sufficient conditions under which a quasi-Gorenstein ring is Gorenstein. Finally, using 2.7 and a theorem of Serre, we provide a slight generalization of a theorem of Evans and Griffith concerning the problem to know when the residue class ring of a height two prime ideal of a certain regular local ring is complete intersection. \section{The Results} It is a basic fact that quasi-Gorenstein rings satisfy Serre's condition $S_{2}$. Although this fact has already been proved, but, for the reader's convenience, in the following lemma, we provide a different proof by using local cohomology tools. Following \cite{M}, we say that a proper ideal $I$ in a Noetherian ring is unmixed if the heights of its prime divisors are all equal. \begin{lem} Every quasi-Gorenstein ring satisfies $S_{2}$. In particular, its zero ideal is unmixed. \end{lem} \begin{proof} Let $(R,\mathfrak{m},k)$ be a quasi-Gorenstein ring of dimension $n$. To prove that $R$ satisfies $S_{2}$, without loss of generality, we can assume that $R$ is complete (cf. \cite[2.1.16]{BH}); so that, by Cohen's structure theorem, there exists a regular local ring $S$ of dimension $n'$ such that $R\cong S/I$ for some ideal $I$ of $S$. Now, by Local Duality Theorem (cf. \cite[11.2.6]{BSh}) and \cite[1.2.4]{BH}, $R\cong \Hom_{R}(E_{R}(k),E_{R}(k))\cong \Hom_{R}(H_{\mathfrak{m}}^{n}(R),E_{R}(k))\cong\Ext^{n'-n}_{S}(R,S)\cong \Hom_{S/(\textit{\textbf{x}})}(R,S/(\textit{\textbf{x}}))$, where $\textit{\textbf{x}}$ is a maximal $S$-sequence in $I$. Using this in conjunction with \cite[1.4.19]{BH}, one can deduce that $R$ satisfies $S_{2}$. Now, for a prime divisor $\frak{p}$ of $R$, we have $0=\depth R_{\frak{p}} \geq \min (2,\Ht\frak{p})$; hence the final statement holds. \end{proof} In \cite[5.2]{BE}, Buchsbaum and Eisenbud established relations, under certain conditions, between almost complete intersection ideals and Gorenstein ideals. Then, Schenzel, in \cite{Sch1}, improved the above result by establishing a duality between almost complete intersection ideals and quasi-Gorenstein ideals. The next proposition, which is related to the above mentioned result, shows that one can characterize quasi-Gorenstein rings by using the concept of linkage. In the proof of the next proposition, for a local ring $(R,\mathfrak{m})$, we use $\widehat{R}$ to denote the $\mathfrak{m}$-adic completion of $R$. \begin{prop} Let $S$ be a Gorenstein local ring and $I$ be an ideal of $S$ of height zero. Then $S/I$ is quasi-Gorenstein if and only if $I\sim Sx$ for some $x\in S$. \end{prop} \begin{proof} $(\Leftarrow)$ Let $R=S/I$ and assume that $\mathfrak{m}$ is the maximal ideal of $R$ and that $n=\dim R$. Then $R=S/0:_{S}x\cong Sx=0:_{S}I\cong \Hom_{S}(R,S)$; so that, by Local Duality Theorem, $E_{R}(R/\mathfrak{m})=\Hom_{R}(R,E_{R}(R/\mathfrak{m}))\cong \Hom_{R}(\Hom_{S}(R,S),E_{R}(R/\mathfrak{m}))\cong H_{\mathfrak{m}}^{n}(R)$. Therefore $R$ is quasi-Gorenstein. \\ $(\Rightarrow)$ Since, by 2.1, $I$ is unmixed, it follows from \cite[2.2]{Sch2} that $I\sim(0:_{S}I)$. Therefore we have only to prove that $(0:_{S}I)$ is a principal ideal. Using the hypothesis, we easily see that $\widehat{S}/I\widehat{S}$ is also a quasi-Gorenstein ring. Thus, applying $- \otimes_S \widehat{S}$ to the relation $I\sim(0:_{S}I)$, one obtains the relation $I\widehat{S}\sim(0:_{\widehat{S}}I\widehat{S})$. Therefore, by the same arguments as in the proof of 2.1, we have $\widehat{S}/I\widehat{S}\cong \Hom_{\widehat{S}}(\widehat{S}/I\widehat{S},\widehat{S})\cong 0:_{\widehat{S}}I\widehat{S}=(0:_{S}I)\widehat{S}.$ Hence, by \cite[Exercise 8.3]{M}, $(0:_{S}I)$ is a principal ideal. \end{proof} \begin{rem} It is straightforward to see that a Noetherian local ring $(R,\mathfrak{m})$ is quasi-Gorenstein if and only if $\widehat{R}$ is so. On the other hand, according to Cohen's structure theorem, $\widehat{R}$ can be described as a residue class ring of a Gorenstein local ring with the same dimension. Therefore, 2.2 provides a criterion to justify whether a given Noetherian local ring is quasi-Gorenstein. \end{rem} The next corollary shows that, for a quasi-Gorenstein ring $R$, the Gorenstein loci ($Gor(R)$) and the Cohen--Macaulay loci ($CM(R)$) coincide. In the course of this paper, for a finitely generated $R$-module $M$, $r(M)$ and $\mu (M)$ will denote the type of $M$ and the minimal number of generators of $M$, respectively. \begin{cor} $Gor(R)=CM(R)$ whenever $R$ is a quasi-Gorenstein ring. \end{cor} \begin{proof} For a prime ideal $\frak{p}$ of $R$, one can use \cite[Theorem 23.2(i) ]{M} to choose a minimal prime $\frak{q}$ of the ideal $\frak{p}\widehat{R}$ such that $\frak{q}\cap R=\frak{p}$. Then we notice that the ring $\widehat{R}_{\frak{q}}/\frak{p}\widehat{R}_{\frak{q}}$ is of dimension 0; so that it is Cohen--Macaulay. This observation in conjunction with \cite[2.1.7 and 3.3.15]{BH} enables us to assume that $R$ is complete. Hence, there are a Gorenstein local ring $S$ and a zero height ideal $I$ of $S$ such that $R\cong S/I$. Therefore, $I\sim Sx$ for some $x\in S$ by 2.2 . Now, in order to prove the assertion, it is enough to show that $CM(R)\subseteq Gor(R)$. To this end, let $\frak{p}\in CM(R)$ and suppose $\frak{p}'\in\Spec(S)$ is such that $\frak{p}'\supseteq I$ and that $\frak{p}$ is the image of $\frak{p}'$ under the natural homomorphism $S\longrightarrow R$. Then, $R_{\frak{p}}\cong S_{\frak{p}'}/IS_{\frak{p}'}$ possesses the canonical module $\Hom_{S_{\frak{p}'}}(R_{\frak{p}},S_{\frak{p}'})$ which is isomorphic to $0:_{S_{\frak{p}'}}IS_{\frak{p}'}=(Sx)_{\frak{p}'}$. Hence, by \cite[3.3.11]{BH}, $r(R_{\frak{p}})=\mu((Sx)_{\frak{p}'})=1$. Therefore, $R_{\frak{p}}$ is a Cohen--Macaulay ring of type 1; so that $R_{\frak{p}}$ is Gorenstein. \end{proof} M. Hermann and N. V. Trung, in \cite{HT}, present a Buchsbaum quasi-Gorenstein ring $(R,\mathfrak{m},k)$ with $\dim R=3$ and $H_{\mathfrak{m}}^{2}(R)\cong k\oplus k$ which is not Gorenstein. This example motives us to the following theorem. \begin{thm} Let $(R,\mathfrak{m})$ be a quasi-Gorenstein ring of dimension $n$ and suppose that $H_{\mathfrak{m}}^{i}(R)=0$ for all integer $i$ with $n/2< i < n$. Then $R$ is Gorenstein. \end{thm} \begin{proof} If $n\leq 2$, then, by 2.1, $R$ is Cohen--Macaulay; hence it is Gorenstein in view of 2.4. Thus, we may suppose that $n\geq 3$. Since $\widehat{R}$ is faithfully flat over $R$, we have $H_{\mathfrak{m}\widehat{R}}^{i}(\widehat{R})=0$ if and only if $H_{\mathfrak{m}}^{i}(R)=0$. Therefore, in view of 2.3, we may further assume that there are a Gorenstein local ring $(S,\frak{n})$ and a zero height ideal $I$ of $S$ such that $R\cong S/I$. By 2.2, $I$ is linked to an ideal $Sx$ for some $x\in S$. Therefore, the assumption $H_{\frak{n}}^{i}(S/I)=0$ for all $n-(n-[\frac{n}{2}])< i < n$ in conjunction with \cite[4.1]{Sch2} implies that $S/Sx$ satisfies Serre's condition $S_{n-[\frac{n}{2}]}$. Now, the application of local cohomology with respect to $\frak{n}$ to the exact sequence $0\longrightarrow S/I\stackrel{x}{\longrightarrow} S\longrightarrow S/Sx\longrightarrow 0$ shows that $H_{\frak{n}}^{i}(S/I)=0$ for all $i < n-[\frac{n}{2}]+1$. Therefore, $H_{\frak{n}}^{i}(S/I)=0$ for all $i < n$. Hence $R$ is Cohen--Macaulay; so that, in view of 2.4, it is Gorenstein. \end{proof} \begin{rem} The conclusion of the above theorem fails if one of the vanishing conditions $H_{\mathfrak{m}}^{i}(R)=0$ $(n/2< i < n)$ is dropped. For example, by \cite[2.2]{HT}, for given integers $n\geq 5$ and $[\frac{n}{2}]< j < n$, there exists a non-Gorenstein quasi-Gorenstein ring $R$ of dimension $n$ such that $H_{\mathfrak{m}}^{i}(R)=0$ for all integers $i$ with $n/2< i < n$ and $i\neq j$. \end{rem} An ideal $I$ in a local Cohen--Macaulay ring is called an almost complete intersection ideal if $\mu (I)=\Ht I+1$. Evans and Griffith, in \cite[2.1]{EG}, proved, over a certain regular local ring, that an unmixed almost complete intersection ideal of height two is Cohen--Macaulay. In this manner we prove the following. \begin{thm} Let $S$ be a local ring and suppose that $I$ is an ideal of $S$ such that $S/I$ is quasi-Gorenstein and that one of the following two conditions holds. \begin{enumerate} \item[(i)]$S$ is Gorenstein and $I$ is almost complete intersection. \item[(ii)]$S$ is regular containing a field and $I$ is of height two. \end{enumerate} Then $S/I$ is Gorenstein. \end{thm} \begin{proof} Assume (i) holds. Considering a maximal $S$-sequence which constitutes a part of a minimal generating set of $I$ and passing to the residue class ring, we can assume that $I=Sy$ for some $y\in S$ and that the height of $I$ is zero. Then, by 2.2, $I\sim Sx$ for some $x\in S$. Therefore, we have an exact sequence $0\longrightarrow S/I\stackrel{x}{\longrightarrow} S \stackrel{y}{\longrightarrow} S\stackrel{x}{\longrightarrow}S\longrightarrow\cdots$, which implies that $S/I$ is an $i$-th module of syzygy for all $i\geq 0$. Thus $S/I$ is Cohen--Macaulay; so that it is Gorenstein by 2.4.\\ Now, assume (ii) holds. One may consider a maximal $S$-sequence in $I$, pass to the residue class ring and use 2.2 to construct an ideal $J$ of height 2 which is generated by three elements such that $I\sim J$. It now follows from \cite[2.1]{EG} that $S/J$ is Cohen--Macaulay. Hence, by \cite[2.1]{PS}, $S/I$ is Cohen--Macaulay. Therefore, by 2.4, $S/I$ is Gorenstein. \end{proof} \begin{rem} In \cite[1.1]{K}, Kunz, essentially, proved that an almost complete intersection ring is not quasi-Gorenstein. Then, as a corollary, he pointed out that an almost complete intersection ring is not Gorenstein. Conversely, one can employ 2.7(i), in the case where $S$ is regular, to show that the above mentioned corollary would imply \cite[1.1]{K}. \end{rem} As an application of the syzygy theorem, Evans and Griffith, in \cite[2.2]{EG}, proved, over a regular local ring $R$ which contains a field, that the residue class ring of a height two prime ideal $\frak{p}$ is complete intersection whenever $\Ext^{2}_{R}(R/\frak{p},R)$ is principal. (Note that this last condition is sufficient for $R/\frak{p}$ to be quasi-Gorenstein.) The following corollary is a slight generalization of this result. The result \cite[Proposition 5]{S} of Serre, which indicates that a height two Gorenstein ideal of a regular local ring is complete intersection, in conjunction with 2.7(ii) leads immediately us to the following corollary. \begin{cor} Let $S$ be a regular local ring containing a field and suppose that $I$ is an ideal of height two such that $S/I$ is quasi-Gorenstein. Then $I$ is complete intersection. \end{cor}

https://arxiv.org/abs/2208.11192

Memory capacity of adaptive flow networks

Biological flow networks adapt their network morphology to optimise flow while being exposed to external stimuli from different spatial locations in their environment. These adaptive flow networks retain a memory of the stimulus location in the network morphology. Yet, what limits this memory and how many stimuli can be stored is unknown. Here, we study a numerical model of adaptive flow networks by applying multiple stimuli subsequently. We find strong memory signals for stimuli imprinted for a long time into young networks. Consequently, networks can store many stimuli for intermediate stimulus duration, which balance imprinting and ageing.

\section{Model of adaptive flow network} \label{sec:1} In this section. we give a detailed derivation of the iteration rule for our model of adaptive networks. A networks is modelled as a connected set of straight tubes described by a graph with $N_\mathrm{nodes}$ nodes enumerated by $i=1,2,\ldots,N_\mathrm{nodes}$ and tube $ij$ of length $l_{ij}$, which connect node $i$ and $j$. Assuming Poiseuille flow in all tubes, conductances $C_{ij}$ are assigned to all tubes. To study the flow in the network, we designate the first node ($i = 1$) as the outlet, while all other nodes ($i > 1$) are inlets with inflow $q_i(t)$. Kirchhoff's current law then reads $L(t)\vec p(t) = \vec q(t)$, where $\vec p(t)=\{p_1(t), p_2(t), \ldots, p_{N_\mathrm{nodes}}(t)\}$ represents the pressures at all nodes, $\vec q(t)=\{q_1(t), q_2(t), \ldots, q_{N_\mathrm{nodes}}(t)\}$ represents the inflows at all nodes, and $L$ is the graph Laplacian matrix with elements $L_{ij} = \delta_{ij}\sum_{n}C_{in} - C_{ij}$, where $C_{ij}=0$ if tube $ij$ is not present. We use Kirchoff's current law to find the instantaneous flow rate $Q_{ij}(t) = C_{ij}(t)(p_i(t) - p_j(t))$ through tube $ij$ at every time step $t$. Following \cite{Murray1926,Corson2010,Katifori2010,Bohn2007}, we evolve the network by adapting its conductances to minimise the power loss of the network, \begin{equation}\label{Power} E(t) = \sum_{<ij>} \frac{\mean{Q_{ij}(t)^{2}}_{T}}{C_{ij}(t)} \;, \end{equation} while obeying the constraint of the total network volume, \begin{equation}\label{constraint} \mathcal{K}^{1/2} = \sum_{<ij>}(C_{ij}(t)l_{ij})^{1/2}l_{ij} \;. \end{equation} The optimal conductance $C_{ij}^{\star}(t)$ that minimise Eq.~\eqref{Power} at $t$, maintaining the constraint of available material given by Eq.~\eqref{constraint}, \begin{equation}\label{final} C_{ij}^{\star}(t) = \frac{\langle Q_{ij}(t)^2\rangle_T^{2/3}\mathcal{K}} {\left(\sum _{<ij>}\langle Q_{ij}(t)^{2}\rangle_T^{1/3}l_{ij}\right)^{2}l_{ij}} \;. \end{equation} The conductances $C_{ij}(t+\delta t)$ at time step $t+\delta t$ adapt to the optimal conductance $C_{ij}^{\star}(t)$ \begin{equation} C_{ij}(t+\delta t) \leftarrow C_{ij}^{\star}(t) \;. \end{equation} We then repeat this rule iteratively over many time steps. Taken together, this iterative rule for adapting conductances follows from independently optimising conductances $C_{ij}$ to minimises the power loss of the network while keeping the total network volume fixed. In this work we model the network as an hexagonal network, with an addition of noise to the node positions to reduce artifacts generated from the regularities of the network. The node at the center of the network, $i=1$, is the sole outlet, with all the nodes in the network are sources of inflows which fluctuate by either turning \textit{on} $q_{i > 1} = 2q^{(0)}$ or \textit{off} $q_{i> 1} = 0$ with equal probability. Meaning the average base inflow through the source nodes are $<q_{i > 1}> = q^{(0)}$. \section{Signal of last stimulus changes with its training and relaxation time}\label{sec:2} \begin{figure*}[ht] \centering \includegraphics[width=\textwidth]{Supplement2_Thesis.pdf} \caption{Signal $S_N$ of final stimulus depend on its training time~$t^\mathrm{train}_N$ and waiting time~$t^\mathrm{wait}_N$. (a--c) $S_N$ for varying of $t^{\mathrm{train}}_N$ and $t^{\mathrm{wait}}_N$ for three different parameter sets. (d--f) The same data $S_N$ is plotted against $t^{\mathrm{wait}}_N$ for different $t^{\mathrm{train}}_N$ (colors). Lines represent exponential fits. (a--f) Model parameters: $t^{\mathrm{wait}}_{\mathrm{pre}} = 5\delta t$,$N_\mathrm{node}=760$,$q^\mathrm{add} = 2000q^{(0)}$, $q^{(0)} =1$, $T=30$, $\mathcal{K}=1600$.} \label{fig:s1} \end{figure*} We observe that the memory signal~$S_N$ of the last stimulus increases with its training time~$t^\mathrm{train}_N$ and decreases with its waiting time~$t^\mathrm{wait}_N$ independent of the pre-stimulation protocol, see Fig~\ref{fig:s1}. This behaviour is similar to the signal of a stimulus written in a freshly initiated untrained network \cite{Bhattacharyya2022}. In particular, $S_N(t^{\mathrm{wait}}_N)$ can be fitted with exponential decay functions, which is similar to our observation in the main manuscript of $S_1$ decaying with the age after the stimulus application, $t_{\mathrm{post}}^{\mathrm{age}}$, following an exponential decay, while $S_1$ increases with the training time $t_{1}^{\mathrm{train}}$. Since, here the age after stimulus application is $t_{\mathrm{post}}^{\mathrm{age}} = t_N^{\mathrm{wait}}$. These observations imply a general dependency of signal $S_n$ on age, post stimulus application $t_{\mathrm{post}}^{\mathrm{age}}$, and training time of that particular stimulus, $t_{n}^{\mathrm{train}}$, for all $n$. \section{Stable thick tubes do not decay irreversibly between stimuli} \begin{figure*}[t] \centering \includegraphics[width=\textwidth]{SuppleTransition.pdf} \caption{Dynamics of networks when a stimulus is enabled or disabled. (a--c) Network morphology at three consecutive time steps during which a stimulus is switched off. Color code indicates relative change of conductances, $C(t+\delta t)/C(t)$. (d, e) $C(t+\delta t)/C(t)$ as a function of $C(t)$ for $6$ different time point around the transition time $t=21$ where the stimulus is switch off (panel d) or switch on (panel e). Model parameters are $N_\mathrm{nodes} =560$, $T=30\delta t$, and $q^{(0)} =1$.} \label{fig:s2} \end{figure*} To understand how adaptive networks evolve between stimuli, we here study the dynamics of conductances by measuring $C(t+\delta t)/C(t)$ for all the tubes of the network before and after the transition of switching the additional load from $q^\mathrm{add} = 1000q^{(0)}$ to $q^\mathrm{add} =0$. We map the the dynamics of conductances on the network at $3$ different time step, $t=t_\mathrm{trans}-\delta t, t_\mathrm{trans}, t_\mathrm{trans}+\delta t$, where $t_\mathrm{trans}$ is the transition time point when the stimulus is switched off. Fig.~\ref{fig:s2}(a) shows that before the transition, the adaptive network shows a hierarchical network structure with thick tubes connecting the nodes where the stimulus is applied (blue nodes) to the outlet of the network (red node). In particular, $C(t+\delta t)/C(t)$ of the thick tubes are approximately $1$, indicating that they do not evolve further. After switching the stimulus off at the transition time at $t_\mathrm{trans}$, we notice that the conductances of the thick tubes start decaying, while other tubes grow; see Fig.~\ref{fig:s2}(b). Already one step later, at $t_\mathrm{trans}+\delta t$, the network assumes an isotropic, tree-like structure; see Fig.~\ref{fig:s2}(c). The quantification in Fig.~\ref{fig:s2}(d, e) indicates that the conductances follow the universal dynamics of adaptive networks~\cite{Bhattacharyya2022}, where conductances below a threshold conductance~$C_\mathrm{th}$ decay, while larger conductances fluctuate over time. The threshold value reads~\cite{Bhattacharyya2022} \begin{equation}\label{CthFinal} C_\mathrm{th} = \frac{\mathcal{K}}{\left(1+\frac{\delta q}{q^{(0)}}\right)^{4}\left(N_\mathrm{nodes}+\frac{q^\mathrm{add}}{q^{(0)}}\right)^{4/3}} \;, \end{equation} where $N_\mathrm{nodes}$ denotes the number of nodes in a network, and $\delta q$ quantifies load fluctuations. The conductances show an unique dynamics at the exact time step when the stimulus is switched, at $t = t_\mathrm{trans} = 21$. As an example, we observe a cluster of tubes with high conductances, either decaying see Fig.~\ref{fig:s2}(d) or growing see Fig.~\ref{fig:s2}(e). As we know from \cite{Bhattacharyya2022} that all the conductance below the threshold $C_\mathrm{th}$ given by Eq.~\eqref{CthFinal} will decay irreversibly, we plot the $C(t+\delta t) = C_\mathrm{th}$ line (the blue line in Fig.~\ref{fig:s2}(d) and (e)) to track the tubes that will decay irreversibly. Here, $C_{th}$ is obtained from fitting the dynamics of thin tubes to, \begin{equation} \frac{C(t+\delta t)}{C(t)} = \Big[ \frac{1}{C_{th}}\Big]^{1/3}C(t)^{1/3}, \end{equation} for a constant $q^{add}$ following \cite{Bhattacharyya2022}. In Fig.~\ref{fig:s2}(d), the line corresponds to $C_{th}(q^{add} = 0)/C(t)$ vs $C(t)$, and in Fig.~\ref{fig:s2}(e) the line corresponds to $C_{th}(q^{add} = 1000)/C(t)$ vs $C(t)$. Indicating all the data points with $C(t+\delta t)/C(t)$ higher than this line have $C(t+\delta t) > C_{th}$ after the transition. We observe that the data points belonging to the cluster of high conductance tubes (at $t=21$) are above of the line, indicating although these tubes decay during the transition of switching the stimulus to $0$, yet does not decay irreversibly, see Fig.~\ref{fig:s2}(d). These observations imply that, the tubes with very large conductances do not decay irreversibly due to the transitions of stimulus strength ($q^{add}$), and are retained in the network. \section{Thin tubes cannot be recovered between stimuli} We next attempt to understand the dynamics of smaller tubes during the transition between stimuli. For this, we calculate the tubes' dynamics in the minimal triangular network shown in Fig.~\ref{fig:s2a}(a), which follows the same adaptation rule as the simulated adaptive networks, \begin{align}\label{main23} C_{ij}(t+\delta t) &= Q_{ij}(t)^{4/3}.\frac{\mathcal{K}}{A(t)^2} & \text{with} && A(t) &= \sum_{<i,j>} Q_{ij}(t)^{2/3} \;. \end{align} We focus on the dynamics when $q^\mathrm{add} = 1000 \rightarrow 0$, but we expect the opposite case of $q^\mathrm{add} = 0 \rightarrow 1000$ to exhibit similar behaviour. \begin{figure}[t] \centering \includegraphics[width=0.5\textwidth]{Theory_Pic.pdf} \caption{Minimal asymmetric network during the transition step, where the stimulus of addition is inflow switched off. (a) Minimal network at $t-\delta t$ where the inflow at node 2 is $q^{(0)}$ and at node 3 is $N_\mathrm{nodes} q^{(0)}+q^\mathrm{add}$. (b) Minimal network at time $t$, where the inflow at node 2 is $q^{(0)}+\delta q$ and node 3 is $N_\mathrm{nodes} q^{(0)}$.} \label{fig:s2a} \end{figure} For the minimal network, we can express Kirchhoff's current law, $L \vec{p} = \vec{q} $, explicitely \begin{equation} \left[ {\begin{array}{c} q_{1} \\ q_{2} \\ q_{3} \\ \end{array}}\right]= \left[ {\begin{array}{ccc} C_{13}+C_{12} & -C_{12} & -C_{13} \\ -C_{12} & C_{12}+C_{23} & -C_{23} \\ -C_{13} & -C_{23} & C_{13}+C_{23} \\ \end{array} } \right] \left[ {\begin{array}{c} p_{1} \\ p_{2} \\ p_{3} \\ \end{array}}\right] \;; \end{equation} see Section~\ref{sec:1}. Setting $p_{1} = 0$ because node 1 is a sink, we have \begin{equation} \left[ {\begin{array}{cc} C_{12}+C_{23} & -C_{23} \\ -C_{23} & C_{13}+C_{23} \\ \end{array} } \right]^{-1} \left[ {\begin{array}{c} q_{2} \\ q_{3} \\ \end{array}}\right]= \left[ {\begin{array}{c} p_{2} \\ p_{3} \\ \end{array}}\right] \end{equation} and thus \begin{multline} \frac{1}{C_{13}(C_{12}+C_{23})+C_{12}C_{23}} \left[ {\begin{array}{cc} C_{13}+C_{23} & C_{23} \\ C_{23} & C_{12}+C_{23} \\ \end{array} } \right]\\ \left[ {\begin{array}{c} q_{2} \\ q_{3} \\ \end{array}}\right] = \left[ {\begin{array}{c} p_{2} \\ p_{3} \\ \end{array}}\right]\;. \end{multline} Solving for $p_2$ and $p_3$, we find \begin{equation}\label{P2n3} \begin{split} p_{2} &= \frac{1}{C_{13}(C_{12}+C_{23})+C_{12}C_{23}}\Big( (C_{13}+C_{23})q_{2} + C_{23}q_{3}\Big)\\ p_{3} &= \frac{1}{C_{13}(C_{12}+C_{23})+C_{12}C_{23}}\Big( C_{23}q_{2} + (C_{12}+C_{23})q_{3}\Big) \;. \end{split} \end{equation} We assume the network has been adapted with a stimulus of an additional inflow and reached a local energy minimum at $t-\delta t$, implying a tree morphology which connects all inlets directly to the outlet. We can then assume the inflow on node $2$ and $3$ is distributed through tubes ${12}$ and ${13}$, respectively. Without loss of generality, we assume the additional load has been applied to node 3, implying $Q_{21}(t-\delta t) = q_{2}(t-\delta t) = q_{2}$ and $Q_{31}(t-\delta t) = q_{3}(t-\delta t) = q_{3}+q^\mathrm{add}$. Using \eqref{main23}, we can then approximate the conductances of tube $12$ and $13$ as $C_{12}(t) = q_{2}(t-\delta t)^{4/3}\mathcal{K}/A(t-\delta t)^2$ and $C_{13}(t) = q_{3}(t-\delta t)^{4/3}\mathcal{K}/A(t-\delta t)^2$, respectively, where we assumed $C_{23}(t)/C_{12}(t) \ll 1$ and $C_{23}(t)/C_{13}(t) \ll 1$ since $Q_{23}(t-\delta t) \ll Q_{21}(t-\delta t)$ and $Q_{23}(t-\delta t) \ll Q_{31}(t-\delta t)$. During the transition from $t-\delta t$ to $t$, the stimulus is switched from a non-zero value to $0$, implying the load at node $3$ reduces to $q_{3}(t) = q_{3}$. The load at node $2$ also changes to $q_{2}(t) = q_{2}+\delta q$ due to the background fluctuation of magnitude $\delta q$, where we assume this perturbation to be small compared to the load at node 2, $\delta q/q_{2} \ll 1$. In general, the minimal network is connected with a bigger network with $N_\mathrm{nodes}$ nodes, so the calculation we present in the following subsections considers the simple case of an asymmetric connection, with highest inflow difference, as shown in the Fig.~\ref{fig:s2a}(a), \begin{equation}\label{q23asym} q_{2} = q^{(0)},\: q_{3} \approx N_\mathrm{nodes} q^{(0)} \;, \end{equation} which implies tube $12$ is thinner than $13$, $C_{12}(t)< C_{13}(t)$, in the steady state network. \subsection{Dynamics of medium thick tubes} The flow rate $Q_{12}$ through tube ${12}$ reads \begin{multline}\label{currQ12} Q_{12} (t) = C_{12}(p_{2}-p_{1})\\ = C_{12}(t)\frac{[C_{13}(t)+C_{23}(t)]q_{2}(t) + C_{23}(t)q_{3}(t)}{C_{13}(t)(C_{12}(t)+C_{23}(t))+C_{12}(t)C_{23}(t)} \end{multline} where we used Eq.~\eqref{P2n3}. Inserting the values of $C_{13}(t)$, $C_{12}(t)$, $q_{2}(t)$, and $q_{3}(t)$ obtained above, we find \begin{equation}\label{Q12} Q_{21}(t) = C_{12}(t)\frac{(q_{3}+q^\mathrm{add})^{4/3}(q_{2}+\delta q)}{\Big(q_{2}(q_{3}+q^\mathrm{add})\Big)^{4/3}}\cdot\frac{A(t-\delta t)^2}{\mathcal{K}} \;. \end{equation} Using Eq.~\eqref{Q12} in Eq.~\eqref{main23}, we can write \begin{multline} C_{12}(t+\delta t) = C_{12}(t)^{4/3}\frac{(q_{3}+q^\mathrm{add})^{16/9}(q_{2}+\delta q)^{4/3}}{\Big(q_{2}(q_{3}+q^\mathrm{add})\Big)^{16/9}} \\ \cdot \frac{A(t-\delta t)^{8/3}}{\mathcal{K}^{4/3}}\frac{\mathcal{K}}{A(t)^2} \end{multline} and \begin{equation} \label{bigveincalc} \frac{C_{12}(t+\delta t)}{C_{12}(t)} = \left(1+\frac{\delta q}{q_{2}}\right)^{4/3}\frac{\Big(q_{2}^{2/3}+(q_3+q^\mathrm{add})^{2/3}\Big)^{2}}{\Big((q_{2}+\delta q)^{2/3}+q_{3}^{2/3}\Big)^{2}} \;. \end{equation} Expanding in the small parameter $q^{(0)}/q^\mathrm{add}$ results in \begin{equation}\label{MediumThick} \frac{C_{12}(t+\delta t)}{C_{12}(t)} = \frac{(N_\mathrm{nodes} +\frac{q^\mathrm{add}}{q^{(0)}})^{4/3}}{N_\mathrm{nodes} ^{4/3}} = \left(1 + \frac{q^\mathrm{add}}{N_\mathrm{nodes} q^{(0)}}\right)^{4/3} \;, \end{equation} which implies the medium thick tubes $12$ will grow when the stimulus is switched off. \subsection{Dynamics of thin tubes with largest inflow difference} We next calculate the dynamics of the thin tube $23$ of the minimal network shown in Fig.~\ref{fig:s2a}. The flow rate~$Q_{23}$ through tube ${23}$ follows from Eq.~\eqref{P2n3}, \begin{equation}\label{currQ23} Q_{23} (t)= C_{23}(p_{2}-p_{3}) = C_{23}\frac{C_{13}q_{2} - C_{12}q_{3}}{C_{13}(C_{12}+C_{23})+C_{12}C_{23}} \;. \end{equation} Inserting the values of $C_{13}(t)$, $C_{12}(t)$, $q_{2}(t)$, $q_{3}(t)$, we find \begin{multline}\label{currentt+1} Q_{23}(t) = C_{23}(t)\frac{(q_{3}+q^\mathrm{add})^{4/3}(q_{2}+\delta q) - (q_{2})^{4/3}(q_{3})}{\Big( q_{2}(q_{3}+q^\mathrm{add})\Big) ^{4/3}} \\ \cdot \frac{A(t-\delta t)^2}{\mathcal{K}} \;. \end{multline} Using Eq.~\eqref{main23}, we write the conductance $C_{23}(t+\delta t)$ at time $t+\delta t$ as a function of conductance $C(t)$ at time $t$, \begin{equation}\label{middlestep} \frac{C_{23}(t+\delta t)}{C_{23}(t)} = \mathcal{S}_\mathrm{disable} C_{23}(t)^{1/3} \;, \end{equation} where the prefactor reads \begin{multline}\label{alpha1} \mathcal{S}_\mathrm{disable} = \frac{\Big( (q_{3}+q^\mathrm{add})^{4/3}(q_{2}+\delta q) - (q_{2})^{4/3}(q_{3}) \Big)^{4/3}}{\Big( q_{2}(q_{3}+q^\mathrm{add})\Big) ^{16/9}}\\ \cdot \frac{A(t-\delta t)^{8/3}\mathcal{K}}{\mathcal{K}^{4/3}A(t)^2} \;. \end{multline} Since $A(t-\delta t) = \Big(q_{2}^{2/3}+(q_{3}+q^\mathrm{add})^{2/3}\Big)$ and $A(t) = \Big( (q_{2}+\delta q)^{2/3}+q_{3}^{2/3} \Big)$, the prefactor becomes \begin{multline}\label{alpha2} \mathcal{S}_\mathrm{disable} = \frac{\Big( (q_{3}+q^\mathrm{add})^{4/3}(q_{2}+\delta q) - (q_{2})^{4/3}(q_{3}) \Big)^{4/3}}{\Big( q_{2}(q_{3}+q^\mathrm{add})\Big) ^{16/9}}\\ \cdot \frac{\Big( q_2^{2/3} +(q_3+q^\mathrm{add})^{2/3}\Big)^{8/3}}{\mathcal{K}^{1/3}\Big((q_2+\delta q)^{2/3}+q_3^{2/3}\Big)^2} \;. \end{multline} Inserting the values of $q_2$ and $q_3$ fromEq.~\eqref{q23asym}, we obtain \begin{multline} \mathcal{S}_\mathrm{disable} = \frac{\Big( \Big[\frac{1}{N_\mathrm{nodes}+\frac{q^\mathrm{add}}{q^{(0)}}}\Bigr]^{2/3}+1\Big) ^{8/3}}{\mathcal{K}^{1/3}} \\ \cdot \frac{\Big( \frac{1}{N_\mathrm{nodes}}(N_\mathrm{nodes}+\frac{q^\mathrm{add}}{q^{(0)}})^{4/3}(1+\frac{\delta q}{q^{(0)}})-1\Big) ^{4/3}}{\Big( (\frac{1}{N_\mathrm{nodes}}(1+\frac{\delta q}{q^{(0)}}))^{2/3}+1\Big) ^{2}} \;. \end{multline} Assuming a large network, $N_\mathrm{nodes} q^{(0)} + q^\mathrm{add} \gg q^{(0)}$, we find \begin{multline}\label{SFinal} \mathcal{S}_\mathrm{disable} \approx \frac{(1+\frac{\delta q}{q^{(0)}})^{4/3}\Big( N_\mathrm{nodes} + \frac{q^\mathrm{add}}{q^{(0)}}\Big) ^{16/9}}{N_\mathrm{nodes}^{4/3}\mathcal{K}^{1/3}}\\ = \frac{(1+\frac{\delta q}{q^{(0)}})^{4/3}\Big(N_\mathrm{nodes}+\frac{q^\mathrm{add}}{q^{(0)}}\Big)^{4/9}\Big(1+\frac{q^\mathrm{add}}{N_\mathrm{nodes} q^{(0)}}\Big)^{4/3}}{\mathcal{K}^{1/3}} \;. \end{multline} Since tubes with $C(t+\delta t) < C_\mathrm{th}(q^\mathrm{add}=0)$ will decay irreversibly after the stimulus is disabled, we find \begin{equation} \mathcal{S}_\mathrm{disable} C(t)^{4/3 }< C_\mathrm{th}(q^\mathrm{add}=0) \end{equation} and thus \begin{equation} C(t) < \left(\frac{C_\mathrm{th}(q^\mathrm{add}=0)}{S_\mathrm{disable}}\right)^{3/4} \;. \end{equation} Using $C_\mathrm{th}$ from Eq.~\eqref{CthFinal} and $\mathcal{S}_\mathrm{disable}$ from Eq.~\eqref{SFinal}, we find \begin{multline} \left(\frac{C_\mathrm{th}(q^\mathrm{add}=0)}{S_\mathrm{disable}}\right)^{3/4} = \Big( \frac{\mathcal{K}}{(1+\frac{\delta q}{q^{(0)}})^{4}N_\mathrm{nodes}^{4/3}} \\ \cdot \frac{\mathcal{K}^{1/3}}{(1+\frac{\delta q}{q^{(0)}})^{4/3}\Big(N_\mathrm{nodes}+\frac{q^\mathrm{add}}{q^{(0)}}\Big)^{4/9}\Big(1+\frac{q^\mathrm{add}}{N_\mathrm{nodes} q^{(0)}}\Big)^{4/3}}\Big)^{3/4},\\ = \frac{K}{(1+\frac{\delta q}{q^{(0)}})^{4}(N_\mathrm{nodes}+\frac{q^\mathrm{add}}{q^{(0)}})^{4/3}} = C_\mathrm{th}(q^\mathrm{add}) \;. \end{multline} This implies tubes with $C(t) > C_\mathrm{th}(q^\mathrm{add})$ will be retained in the network even after disabling the stimulus. Moreover, tubes that were decaying in the presence of the stimulus will continue decaying even when the stimulus is disabled. \begin{figure*} \centering \includegraphics[width=\textwidth]{SupplementS4.pdf} \caption{Dynamics of tube conductances during the transition of disabling a stimulus matches analytical approximations. (a) Average conductance changes $\mean{C(t+\delta t)/C(t)}$ of medium thick tubes in networks as a function of the corresponding analytical prediction given by Eq.~\eqref{MediumThick}. The solid line is a power-law fit. (b) The pre-factor $\mathcal{S}_\mathrm{disable}$, obtained from fitting the dynamics of thin tubes to $C(t+\delta t)/C(t) = \mathcal{S}C(t)^{1/3}$, as a function of the prediction given by Eq.~\eqref{SFinal}. , including the correction of the power obtained from (a), shows the data follows a straight line with the corrected prediction. (a--b) For varying $q^{(0)}$ and $q^{\mathrm{add}}$ and fixed model parameters $N_\mathrm{nodes} = 760$, $T=30\delta t$. } \label{fig:s3} \end{figure*} Fig.~\ref{fig:s3} shows that our analytical prediction is reasonably correct. However, we observe that the measured value of the dynamics of the medium thick tubes are slightly lower than that of the analytical prediction, probably because the calculation only considers the most extreme case of highest inflow difference. Taken together, we show the established hierarchy of tube conductances cannot be erased by switching the stimulus from a non-zero value to $0$ or vice versa. \section{Memory signal from combination of age and training components} \begin{figure*}[ht!] \centering \includegraphics[width = \textwidth]{Supplement5a.pdf} \caption{Memory signal~$S_n$ depends on training time $t^\mathrm{train}$ and age of the network before training. (a)~Analytical prediction assuming $S_n$ is a product of the impact of age and training. (b)~$S_n$ as a function of $t^\mathrm{train}$ for various~$n$. (c)~Fit parameters (coefficient of training and age component) obtained while fitting memory readout signal as a sum of training and age component, plotted against the probed stimulus. (d) $R^2$ of the fit obtained from fitting the signal as various combination of training and age with respect to the probed stimulus.} \label{fig:S5} \end{figure*} In the main text, we observed the signal~$S_1$ of the first stimulus increases with the training time $t^{\mathrm{train}}$, \begin{equation} S_1(t^{\mathrm{train}}) \sim (1-e^{-t^{\mathrm{train}}/\tau_{\mathrm{train}}}) \;. \end{equation} We also observed that the signal~$S_N$ of the last stimulus decays with the age $t^{\mathrm{age}}_{N-1}$ before the stimulus application, \begin{equation} S_N(t^{\mathrm{age}}_{N-1}) \sim e^{-t^{\mathrm{age}}_{N-1}/\tau_{\mathrm{pre}}} \;. \end{equation} Combining these observations, we hypothesise that the signals~$S_n$ of general stimuli between the first and last stimulus are affected by both effects. $S_n$ should thus be a function of $S^{\mathrm{train}}_n=(1-e^{-t^{\mathrm{train}}_n/\tau_{\mathrm{train}}})$ and $S^{\mathrm{age}}_n=e^{-t^{\mathrm{age}}_{n-1}/\tau_{\mathrm{pre}}}$. We start by asking whether a simple product is an adequate description, \begin{equation}\label{AnalMul} S_n = S^{\mathrm{age}}_n S^{\mathrm{train}}_n = e^{-t^{\mathrm{age}}_{n-1}/\tau_{\mathrm{pre}}}(1-e^{-t^\mathrm{train}_n/\tau_{\mathrm{train}}}) \;. \end{equation} Fig.~\ref{fig:S5}(a) and Fig.~\ref{fig:S5}(b) shows that this prediction cannot reproduce the qualitative dependency of the signal on $t^{\mathrm{train}}$ and $t^{\mathrm{wait}}$. In particular, Eq.~\eqref{AnalMul} suggests a maximal signal for a non-zero training time for any $n$, while the numerical data shown in Fig.~4 of the main text shows that the signal is strongest for very small $t^{\mathrm{train}}$ and $t^{\mathrm{wait}}$, specifically for $n>3$. \begin{figure*}[p] \centering \includegraphics[width = \textwidth]{Supplement5c_Thesis.pdf} \caption{Analytical approximation of signal w.r.t training and waiting time when nth stimulus is probed (a) Numerically obtained memory read out signal with respect to training and waiting time as shown in Fig.~4(a) (b) Signal $S_n$ approximated as a product of $S_n^{\mathrm{age}}$ and $S_n^{\mathrm{train}}$, in Eq.~\eqref{AnalMul}, with parameters $\tau_{\mathrm{pre}}=50\delta t$ and $\tau_{\mathrm{train}}=50\delta t$. (c) Signal $S_n$ approximated as a linear sum of $S_n^{\mathrm{age}}$ and $S_n^{\mathrm{train}}$ as shown in Eq.~\eqref{Approx1} using the functional form found from fitting the data Eq.~\eqref{Num}, with parameters $\tau_{\mathrm{pre}}=50\delta t$ and $\tau_{\mathrm{train}}=50\delta t$. (d) $S_n = \frac{1}{2}\big( e^{-t^{\mathrm{age}}_{n-1}/\tau_{\mathrm{pre}}} + \frac{n}{2}^{-n+1}(1-e^{-t^{\mathrm{train}}/\tau_{\mathrm{train}}})\big)$, with parameter $\tau_{\mathrm{pre}} =50\delta t$ and $\tau_{\mathrm{train}} = 62\delta t$ (e) $S_n = (1-e^{-(n+1)/1.5})e^{-t^{\mathrm{age}}_{n-1}/\tau_{\mathrm{pre}}} +e^{-(n-1)/1.5}(1-e^{-t^{\mathrm{train}}/\tau_{\mathrm{train}}})$ with parameters $\tau_{\mathrm{pre}}=50\delta t$ and $\tau_{\mathrm{train}}=50\delta t$.} \label{fig:S6} \end{figure*} We next hypothesis that $S_n$ is a linear combination of $S^{\mathrm{age}}_n$ and $S^{\mathrm{train}}_n$, \begin{equation}\label{Approx1} S_n = f^{\mathrm{train}}_n S^{\mathrm{age}}_n + f^{\mathrm{age}}_n S^{\mathrm{train}}_n \;, \end{equation} where we allow different coefficients $f^{\mathrm{train}}_n$ and $f^{\mathrm{age}}_n$ for each stimulus. We determine these by fitting $S_n/S^{\mathrm{age}}_n(t^{\mathrm{train}},t^{\mathrm{wait}})$ to $f^{\mathrm{age}}_n + f^{\mathrm{train}}_n S^{\mathrm{train}}_n/S^{\mathrm{age}}_n(t^{\mathrm{train}},t^{\mathrm{wait}})$ for different values of $n$ over a range of $t^{\mathrm{train}}/\delta t \in [1, 40]$ and $t^{\mathrm{wait}}/\delta t \in [1,40]$. Fig.~\ref{fig:S5}(c) shows that a suitable fit is given by \begin{equation}\label{Num} \begin{split} f^{\mathrm{train}}_n &= 0.14+0.386 e^{-n/1.75}\\ f^{\mathrm{age}}_n &= 0.0475+0.0023e^{n/0.99} \end{split} \end{equation} We note in Fig.~\ref{fig:S5}(d) that the $R^2$ error of the fit to compute the coefficients of each component is $\approx 0.4$, indicating that the signal is not just a linear superposition of the components. We use a few different models to fit the data. As example, we fit the signal to $S_n= f_n^{\mathrm{train}}S_n^{\mathrm{train}} + f_n^{\mathrm{age}}(S_n^{\mathrm{age}})^{3}$ as we do not expect the power of the age component to change with n. Additionally we fit the signal to some other simple non-linear composition of age and training component, as example $S_n = f_n^{\mathrm{train}}(S_n^\textrm{train})^{n} + f_n^{\mathrm{age}}S_n^\textrm{age}$ and $S_n = f_n^{\mathrm{train}}(S_n^\textrm{train})^{1/n} + f_n^{\mathrm{age}}S_n^\textrm{age}$. However, none of these models increase the $R^2$ of the fit significantly. So for simplicity, we choose to approximate the signal as a linear sum of $S^{\mathrm{train}}_n$ and $S^{\mathrm{age}}_n$. We observe, if we choose the coefficients of $S^{\mathrm{train}}_n$ and $S^{\mathrm{age}}_n$ following \eqref{Num}, the signal of a stimulus obtained using this approximation qualitatively agrees with the numerical observation, see Fig.~\ref{fig:S6}(c). Now to obtain a generic analytical approximation of the memory read-out signal without including many different parameters, that also describes the signal as a linear superposition of the age and training impact, we assume that the coefficient of $S^{\mathrm{age}}_n$ does not change and the coefficient of $S^{\mathrm{train}}_n$ decreases of n. We assume these coefficients because, Eq.~\eqref{Num} suggests that the coefficient of the training component reduces with $n$ and the coefficient of age component increase with $n$. Moreover, the fit parameters show that the change of the coefficient of age component over n ($\sim 0.0023$) is much smaller than the change of the coefficient of the training component over n ($\sim 0.36$), which agrees with our observation in Fig.~2(b). Using these observations we now choose the following bounded function which is a sum of $S_n^{\mathrm{age}}$ and $S_n^{\mathrm{train}}$ to approximate the signal of $n$th stimulus when $n>1$. \begin{equation} \label{Approxa} \begin{split} S_{n} &= e^{-t^{\mathrm{age}}_{n-1}/\tau_{\mathrm{pre}}} + f^{\mathrm{train}}_n(1-e^{-t^{\mathrm{train}}/\tau_{\mathrm{train}}}),\\ &= \frac{1}{2}\big( e^{-t^{\mathrm{age}}_{n-1}/\tau_{\mathrm{pre}}} + (\frac{n}{2})^{-n+1}(1-e^{-t^{\mathrm{train}}/\tau_{\mathrm{train}}})\big). \end{split} \end{equation} Additionally, from the numerical observation, we approximate for $n=1$, \begin{equation} S_1 = \frac{1}{2}(1-e^{-t^{\mathrm{train}}/\tau_{\mathrm{train}}}). \end{equation} Although Eq.~\eqref{Approxa} reproduces, the dependency of signal on training and waiting time qualitatively (see, Fig.~4(b) and Fig.~\ref{fig:S6}(d)), this does not reproduce the dependency quantitatively. However, we show that the same qualitative dependency on training and waiting time is robust to small changes of parameters. As example, when Eq.~\eqref{Approxa} is plotted assuming $\tau_{\mathrm{pre}} = 50\delta t$ and $\tau_{\mathrm{train}} =62\delta t$, see Fig.~\ref{fig:S6}(d), the qualitative feature does not change from the observation in Fig.~4(b). Similarly the qualitative dependency does not change, when different functions $f^{\mathrm{train}}_n$ and $f^{\mathrm{age}}_n$ are chosen as the coefficients of the training and age impact. As an example, if the chosen functions are, \begin{equation} \begin{split} f^{\mathrm{train}}_n = e^{-(n-1)/1.5},\\ f^{\mathrm{age}}_n = (1-e^{-(n-1)/1.5}), \end{split} \end{equation} consequently, the signal of the nth stimulus is approximated as, \begin{multline}\label{Approx2} S_n = (1-e^{-(n-1)/1.5})e^{-t^{\mathrm{age}}_{n-1}/\tau_{\mathrm{pre}}}\\ +e^{-(n-1)/1.5}(1-e^{-t^{\mathrm{train}}/\tau_{\mathrm{train}}}), \end{multline} the qualitative dependency of approximated signal on $t^{\mathrm{train}}$ and $t^{\mathrm{wait}}$ is similar to Fig.~4(b), see Fig.~\ref{fig:S6}(e). Note that in both approximations Eq.~\eqref{Approxa} and Eq.~\eqref{Approx2}, the ratio of the coefficients of age and training impact increases with n, meaning after a certain number of pre-stimuli the impact of age dominates over the impact of training. \section{\label{sec:level1}Introduction} Biological flow networks, like vasculature~\cite{Murray1926}, fungal mycelium~\cite{Boddy2010}, or slime mold~\cite{Marbach2016,Alim2013}, optimise their function by remodelling the network morphology in response to internal and external stimuli~\cite{Sugden2017,Chen2012,Hu2012,Tero2008,Tero2010,Marbach2016,Marbach2022}. In particular, the slime mold \textit{Physarum polycephalum} reorganises its network morphology during foraging and migration~\cite{Kamiya1988,Kuroda2015}, or as responses to environmental influences~\cite{Kramar2021a}, by adapting tubes to flow~\cite{Tero2008,Tero2010,Alim2017b}. Although the organism only consist of a single cell, it processes information~\cite{Gao2019,Beekman2015,Tero2010,Tero2006,Tero2008,Nakagaki2000} and stores memory of external stimuli in the network morphology~\cite{Kramar2021a,Bhattacharyya2022}. Yet, the information processing capabilities of \textit{Physarum} in particular, and adaptive flow networks more generally, are so far unclear. The self-organised information processing of \textit{Physarum} is reminiscent of other physical learning systems~\cite{Stern2022}: Physical networks can be trained to have unusual mechanical properties \cite{Reid2018,Pashine2019} and functionalities~\cite{Rocks2017,Hexner2020} either by modifying microscopic properties by global optimisation~\cite{Yan2017} or as local responses~\cite{Dillavou2021}. Such networks can also learn multiple states \cite{Rocks2019}, which is key for obtaining multi-functionality~\cite{Rocks2019} and multi-stability~\cite{Yang2018,Steinbach2016,Che2017,Bertoldi2017,Overvelde2016,Silverberg2015,Waitukaitis2015,Yang2018a,Fu2018,Kim2019} in physical systems, and for performing complex tasks like image classification~\cite{Stern2020,Stern2021,Anisetti2022} using such physical networks. The multiple states can be either imprinted simultaneously~\cite{Rocks2019} or learned subsequently~\cite{Stern2020}. In both cases, there is a maximal number of states that can be learned, which is the learning capacity of the system~\cite{Rocks2019,Stern2020}. Although memory is essential for learning~\cite{Stern2022}, the memory capacity of flow networks remains unexplored. We, here, investigate this question theoretically, by analysing memory in a model of adaptive flow network, which is subjected to various external stimuli, similar to natural flow networks~\cite{Ito2010,Meyer2017}. We identify that a stimulus is stored more robustly, and can, thus, be retrieved more easily, when networks are young and are exposed to a stimulus for a long time. Since these two criteria are incommensurable for multiple stimuli, a trade-off determines the memory capacity of these adaptive flow networks. \section{\label{level2}Model} \begin{figure*}[t!] \centering \includegraphics[width=\textwidth]{Figure1.pdf} \caption{Stimulus locations are retained in previously stimulated networks. (a)~Schematic of adaptive flow networks with a central outlet (red dot), fluctuating inflows at all other nodes, and additional inflow for stimuli (blue dots). The temporal sequence shows snapshots of the training protocol, where $N$ stimuli (purple boxes) are applied sequentially by stimulating for a period $t^\mathrm{train}_n$ followed by relaxation period~$t^\mathrm{wait}_n$ for each stimulus. Finally, the last stimulus is probed (pink dots) to determine the signal~$S_N$ according to Eq.~\eqref{eq:signal}. (b)~$S_N$ as a function of the training time $t^\mathrm{train}_n = t^\mathrm{train}_\mathrm{pre}$ and waiting time $t^\mathrm{wait}_n=t^\mathrm{wait}_\mathrm{pre}$ for most stimuli ($n<N$) with $N=5$. (c)~$S_N$ as a function of $t^\mathrm{train}_\mathrm{pre}$ and $N$ for $t^\mathrm{wait}_\mathrm{pre}=5\delta t$. (d)~$S_N$ as a function of $t^\mathrm{wait}_\mathrm{pre}$ and $N$ for $t^\mathrm{train}_\mathrm{pre}=5\delta t$. (a--d) Model parameters are $t^\mathrm{train}_N=10\delta t$, $t^\mathrm{wait}_N=5\delta t$, $N_\mathrm{nodes}=1100$, $q^\mathrm{add} = 2000\,q^{(0)}$, $q^{(0)} = 1$, $\mathcal K=1600$, and $T = 30\delta t$. Data shows mean from $1500$ independent simulations.} \label{fig:1} \end{figure*} We use the standard model for adaptive flow networks that minimises energy dissipation in order to maximise transport through the network, for fixed network building material~\cite{Murray1926,Hacking1996,Bohn2007,Corson2010,Katifori2010,Hu2013}. These flow networks are modelled as a graph of $N_\mathrm{nodes}$ nodes connected by links ${ij}$, where $i,j\in \{1,\ldots,N_\mathrm{nodes}\}$. The links have length $l_{ij}$ and time-dependent conductances $C_{ij}(t)$. We consider a network of cylindrical hollow tubes with conductances $C_{ij} = \pi r_{ij}(t)^4/8\mu l_{ij}$ according to Hagen-Poiseuille's law, where $\mu$ is the viscosity of the enclosed fluid. In our case, node $i=1$ serves as the sole outlet, while all other nodes are inlets with fluctuating inflows $q_i(t)$, which are either $q_i=0$ or $q_i=2q^{(0)}$ with equal probability. Conservation of total flow implies $q_1(t) = -\sum_{i>1} q_{i}(t)$. We chose a disk-shaped network geometry with the outlet in the centre; see Fig.~\ref{fig:1}(a). Conservation of flow at every node, as described by Kirchhoff's law, then uniquely determines the flow $Q_{ij}(t)$ in all links, given the entire networks conductances $C_{ij}(t)$ and the inflows $q_i(t)$; see Appendix A. The adaptive dynamics follow from the assumption that networks minimise dissipation~\cite{Murray1926} \begin{equation} E(t)=\sum_{<ij>}\frac{Q_{ij}(t)^{2}}{C_{ij}(t)} \;, \end{equation} while obeying the constraint \begin{equation} \label{eqn:constraint} \mathcal{K}^{\frac12} = \sum_{<ij>} C_{ij}(t)^{\frac12}l_{ij}^{\frac32} \;, \end{equation} where $\mathcal{K}^{\frac12}$ is proportional to the fixed overall volume of all links. We follow an iterative relaxation algorithm~\cite{Bohn2007}, where the conductances at the next time step, $C_{ij}(t+\delta t)$, adapt to minimise $E(t)$ while obeying Eq.~\eqref{eqn:constraint}, implying \begin{equation}\label{eq:iteration} C_{ij}(t+\delta t) = \frac{\mathcal{K}\mean{Q_{ij}(t)^2}_T ^{\frac{2}{3}}} {\left(\sum _{<ij>}\langle Q_{ij}(t)^{2}\rangle_T^{\frac{1}{3}}l_{ij}\right)^{2}l_{ij}} \;, \end{equation} where we average the flow over a duration $T$, $\mean{Q_{ij}^2}_T$, since the inflows at every node fluctuate over time. To probe for memory, we initiate networks with conductances~$C_{ij}$ chosen uniformly from the interval $[0, 1]$, which are then rescaled, so they obey the constraint given by Eq.~\eqref{eqn:constraint}. We then stimulate the networks using an additional inflow $q^{\mathrm{add}}$ at the outer rim at a specific angular location; see Fig.~\ref{fig:1}(a). We distribute the additional inflow over a few nodes to avoid artefacts from the symmetries of the underlying networks. The adaptation dynamics then imprint the stimulus in a treelike structure from the nodes of additional inflow to the centred outlet \cite{Bhattacharyya2022}; see Fig.~\ref{fig:1}(a). Once the additional inflow is withdrawn, networks return to seemingly isotropic morphologies. Yet, when probing networks by re-applying an additional load at exactly the same location, the power loss of previously stimulated, and thus trained, networks, $E_{\mathrm{trained}}$, is distinctively less than if probed at any other location. In particular, $E_{\mathrm{trained}}$ is less than the power loss $E_{\mathrm{untrained}}$ for probing untrained networks that evolved for the same total time, but did not see the stimulus~\cite{Bhattacharyya2022}. To quantify this memory, we established the normalised difference in power loss between trained and untrained networks as a measure of the memory readout signal $S$, \cite{Bhattacharyya2022}, \begin{equation} \label{eq:signal} S = 1 - \frac{\mean{E_{\mathrm{trained}}}}{\mean{E_{\mathrm{untrained}}}} \;, \end{equation} where brackets indicate ensemble averages over initial configurations and positions of the additional loads. We used this quantification to show that freshly initiated networks memorise single stimuli in the spatial location and orientation of the vanishing links~\cite{Bhattacharyya2022}. The stimulus can be read out by probing the network again after the stimulus is withdrawn. Yet, it is unclear how well already stimulated networks can store stimuli and whether networks can store multiple stimuli simulatenously. \section{\label{sec:level3}Results} \subsection{\label{sec:level3a} Pre-stimulated networks can memorise stimuli} We start by asking whether previously evolved networks can store a stimulus reliably. We evolve networks with a stimulation protocol by consecutively applying $N$ stimuli, distinguished by the angle of the additional inflow. We choose the angles randomly from $10$ possibilities, $\{0, \frac{\pi}{5}, \ldots, \frac{9\pi}{5}\}$, and we set the angular range of each stimulus to $\frac{\pi}{6}$ to avoid stimuli overlap. Starting with a randomly initialised network, we apply one stimulus after the other. The $n$-th stimulus is imprinted on the network by iterating Eq.~\eqref{eq:iteration} with the additional load corresponding to the stimulus for a duration $t^\mathrm{train}_n$ and then without load for a duration~$t^\mathrm{wait}_n$. Taken together, the network evolved to time \begin{equation} \label{eq:age} t^\mathrm{age}_n = \sum_{m=1}^n \left(t^\mathrm{train}_m + t^\mathrm{wait}_m\right) \end{equation} after the $n$-th stimulus has been applied. To test whether a stimulated network can memorise an additional stimulus, we apply $N-1$ stimuli with identical properties and then probe the signal of a final stimulus; see Fig.~\ref{fig:1}(a). We thus have $t^\mathrm{train}_n = t^\mathrm{train}_\mathrm{pre}$ and $t^\mathrm{wait}_n = t^\mathrm{wait}_\mathrm{pre}$ for $n<N$, while the final stimulus can have different parameters. The signal~$S_N$ quantifies the dissipation difference of applying the $N$-th stimulus, analogously to Eq.~\eqref{eq:signal}. For constant parameters of the pre-stimulation protocol, we observe that $S_N$ increases with $t^\mathrm{train}_N$ and decays with $t^\mathrm{wait}_N$; see Appendix B. This behaviour closely resembles memory formation in a freshly initiated network~\cite{Bhattacharyya2022}, even though we here use pre-stimulated networks. We next test the influence of the precise pre-stimulation protocol by varying the number of applied stimuli, $N$, the training time, $t^\mathrm{train}_\mathrm{pre}$, and the relaxation time, $t^\mathrm{wait}_\mathrm{pre}$. Fig.~\ref{fig:1}(b--d) shows that the signal~$S_N$ of the final stimulus decreases when increasing any of these parameters, so the pre-stimulation protocol affects how well additional memories can be stored. However, our simulations demonstrated that pre-stimulated adaptive networks can store information about additional stimuli. \subsection{\label{sec:level3c}Memory capacity reduces with age} \begin{figure} \centering \includegraphics[width=0.5\textwidth]{Figure2.pdf \caption{\label{fig:2} Memory signal of final stimulus reduces with network age. (a)~Fraction of vanishing links (blue symbols) as a function of network iterations averaged over $80$ independent runs. Model parameters $N_\mathrm{nodes}$, $q^{\mathrm{add}}$, $q^{(0)}$, and $T$ are given in Fig.~\ref{fig:1}. Red line indicates an exponential fit. (b, c) Signal~$S_N$ of final stimulus as a function of age~$t^\mathrm{age}_{N-1}$ before stimulus was applied. Panels b and c show data of Fig.~\ref{fig:1}(b) and Fig.~\ref{fig:1}(c), respectively. Blue lines indicate exponential fits. } \end{figure} \begin{figure*} \centering \includegraphics[width=\textwidth]{Figure3.pdf} \caption{\label{fig:3} Signal~$S_1$ of first stimulus increases with training time and decreases with network age. (a)~Snapshots of network, which is subjected to the first stimulus for $t^\mathrm{train}_1$, relaxed for $t^\mathrm{wait}_1$, and then $N-1$ stimuli are applied with $t^\mathrm{train}_n=t^\mathrm{train}_\mathrm{post}$ and $t^\mathrm{wait}_n = t^\mathrm{wait}_\mathrm{post}$ ($n>1$, dotted box), until the first stimulus is probed. (b)~$S_1$ as a function of $t^\mathrm{train}_1$ for $N=5$ and $t^\mathrm{train}_\mathrm{post}=5\delta t$. (c)~$S_1$ as a function of $t^\mathrm{age}_\mathrm{post}=t^\mathrm{age}_N - t^\mathrm{train}_1$ for various $N$ at $t^\mathrm{train}_1 =10\delta t$. (b--c) Blue lines indicate exponential fits. Parameters are $t^\mathrm{wait}_1 =t^\mathrm{wait}_\mathrm{post} =5\delta t$ and given in Fig.~\ref{fig:1}.} \end{figure*} We next investigate how the pre-stimulation protocol affects the memory of the final stimulus. Since information about stimuli locations are stored in the orientation and location of irreversibly decaying links~\cite{Bhattacharyya2022}, we first determine how the micro-structure of the network evolves with time. Fig.~\ref{fig:2}(a) shows that the average fraction of vanishing links saturates exponentially with time, which suggest that the memory capacity of adaptive flow networks decreases with the time~$t^\mathrm{age}_{N-1}$, given by Eq.~\eqref{eq:age}, that the network evolved for before the stimulus is applied. Re-plotting the memory signal $S_N$ of the final stimulus as a function of $t^\mathrm{age}_{N-1}$ leads to a data collapse for various values of $N$, $t^\mathrm{train}_\mathrm{pre}$, and $t^\mathrm{wait}_\mathrm{pre}$; see Fig.~\ref{fig:2}(b, c). The two panels differ in whether $N$ (panel b) or $t^\mathrm{wait}_\mathrm{pre}$ (panel c) are kept fixed while the other parameters are varied. In both cases, the data collapse is well-described by an exponential decay \begin{equation} \label{eq:signal_final} S_N(t^\mathrm{age}_{N-1}) \approx S^{\infty}_N +A_N\exp\left(-\frac{t^\mathrm{age}_{N-1}}{\tau_\mathrm{pre}}\right) \;, \end{equation} where $\tau_\mathrm{pre}\approx 52\,\delta t$ denotes the time scale, with which pre-stimulation reduces the memory capacity of the final stimulus. The maximal memory capacity, $S^{\infty}_N + A_N \approx 0.22$ for $t^\mathrm{age}_{N-1}=0$, is significantly larger than the residual capacity, $S^{\infty}_N\approx 0.03$, consistent with the fact that pre-stimulation of the networks reduces the memory capacity. The fact that the exponential decay adequately describes the decreasing capacity suggests that only the total duration of pre-stimulation is important, while the details of the protocol are irrelevant. Consequently, younger networks allow for a larger memory signal of the final stimulus. \subsection{\label{sec:level3d}Training time dominates signal of first stimulus} \begin{figure*} \centering \includegraphics[width=\textwidth]{Figure4.pdf \caption{\label{fig:4}Memory capacity depends on stimulation protocol parameters. (a)~Numerically obtained signals~$S_n$ for all $N=5$ stimuli as functions of training times $t^\mathrm{train}$ and waiting times $t^\mathrm{wait}$. (b)~Analytical prediction of $S_n$ given by Eq.~\eqref{eq:multsignal} for all stimuli as functions of $t^\mathrm{train}$ and $t^\mathrm{wait}$ for $N=5$. (c)~Fraction of stimuli with strong signal ($S_n>0.04$) as a function of $t^\mathrm{train}$ and $N$ for $t^\mathrm{wait} =5\delta t$. (d)~Number of stimuli with $S_n>0.04$ as a function of $t^\mathrm{train}$ and $t^\mathrm{wait}$ for $N =5$. (e)~Minimal signal $\min_n S_n$ of $N=5$ stimuli as a function of $t^\mathrm{train}$ and $t^\mathrm{wait}$. (a--e) Model parameters are given in Fig.~\ref{fig:1}. } \end{figure*} To retain multiple memories, adaptive networks need to store information about all stimuli. We, thus, next investigate how information about earlier stimuli are retained and particularly focus on the first stimulus. To investigate the first stimulus in detail, we change the protocol to control the training parameters of the first stimulus separately from all the other stimuli; see Fig.~\ref{fig:3}(a). For simplicity, we use identical parameters for the other stimuli, $t^\mathrm{train}_n = t^\mathrm{train}_\mathrm{post}$ and $t^\mathrm{wait}_n = t^\mathrm{wait}_\mathrm{post}$ for $n=2,..,N$. The network is probed at the same location as the first stimulus to obtain the memory signal~$S_1$ of the first stimulus. Fig.~\ref{fig:3}(b) shows that $S_1$ increases with the training time of the first stimulus, $t^\mathrm{train}_1$, and approaches zero for $t^\mathrm{train}_1 = 0$. $S_1$ again shows an exponential saturation, \begin{equation} \label{eq:signal_first} S_1(t^\mathrm{train}_1) \approx B_1 \left[1-\exp\left(-\frac{t^\mathrm{train}_1}{\tau_\mathrm{train}}\right)\right] \;, \end{equation} where $\tau_\mathrm{train}$ is the training time scale and $B_1$ denotes the maximal signal for $t^\mathrm{train}_1 \rightarrow \infty$. Similar to our previous work~\cite{Bhattacharyya2022}, longer training leads to a stronger signal. We next investigate how the signal of the first stimulus depends on subsequently applied stimuli. Fig.~\ref{fig:3}(c) indicates that $S_1$ decays as the networks evolves further, similar to our previous study~\cite{Bhattacharyya2022}. We find that $S_1$ only depends on the duration of evolution after the first training period, $t^\mathrm{age}_\mathrm{post} = t^\mathrm{age}_N - t^\mathrm{train}_1$, and not the precise details of the protocol. Moreover, $S_1$ again decays exponentially, \begin{equation} S_1(t^\mathrm{age}_\mathrm{post}) \approx S^{\infty}_1 + A_1\exp\left(-\frac{t^\mathrm{age}_\mathrm{post}}{\tau_\mathrm{post}}\right) \;, \end{equation} where the coefficients have the same interpretation as in Eq.~\eqref{eq:signal_final}. Our fits indicate that $\tau_\mathrm{post} \approx \tau_\mathrm{pre}$, consistent with an intrinsic time scale of memory formation. Note that the residual memory capacity $S^\infty_1 \approx 0.1$ is large, implying that subsequent training does not affect the signal very strongly. This is consistent with the picture that memory is stored by vanishing links that cannot be revived; see analytical and numerical observations of the transitions phase between stimuli in Appendix C and D. Taken together, we find that adaptive flow networks can store multiple memories. \subsection{\label{sec:level3e}Trade off between age and training time limits memory capacity} We found that stimuli are imprinted most strongly when they are trained for a long time on a young network. These goals of long training times and young networks are contradictory for late stimuli, suggesting there must be a trade-off for best performance of imprinting multiple stimuli. To understand how many stimuli can be imprinted in a network, we next consider $N$ non-overlapping stimuli with identical stimulation parameters, $t^\mathrm{train}_n = t^\mathrm{train}$ and $t^\mathrm{wait}_n = t^\mathrm{wait}$ for $n=1,..,N$. We now also probe all stimuli locations to obtain a signal $S_n$ for each stimulus. Fig.~\ref{fig:4}(a) shows data for five stimuli as a function of $t^\mathrm{train}$ and $t^\mathrm{wait}$. We recover that the signal~$S_1$ of the first stimulus mainly depends on the training time~$t^\mathrm{train}$ and is barely affected by the subsequent dynamics. Conversely, the signal of all other stimuli decreases with network age, i.e., with increasing $t^\mathrm{train}$ and $t^\mathrm{wait}$. In particular, mid-timed stimuli have the weakest signals, suggesting that they are affected by both pre-stimuli ageing as well as subsequent degradation. We next develop an analytical prediction of the signal of all stimuli, motivated by the successful description of the signals of the first and last stimulus demonstrated above. We hypothesise that the signal of the $n$-th stimulus is a combination of the pre-stimulus ageing, described by Eq.~\eqref{eq:signal_final}, and the actual training, described by Eq.~\eqref{eq:signal_first}, while we neglect the small effect of the post-stimulus signal degradation. We show in Section V of the Supplementary Information that a weighted sum of the two effects adequately describes the data, which results in the prediction \begin{equation} \label{eq:multsignal} S_n \approx \frac{1}{2} \begin{cases} 1 - e^{-\frac{t^\mathrm{train}}{\tau_\mathrm{train}}} & n = 1 \\ e^{-\frac{t^\mathrm{age}_{n-1}}{\tau_\mathrm{pre}}} + \left(\frac{n}{2}\right)^{1-n}\left(1-e^{-\frac{t_\mathrm{train}}{\tau_\mathrm{train}}}\right) & n> 1 \;, \end{cases} \end{equation} where $t^\mathrm{age}_n$ is given by Eq.~\eqref{eq:age}. This equation correctly captures that $S_n$ decreases with the $n-1$ previously applied stimuli. Fig.~\ref{fig:4}(b) shows that Eq.~\eqref{eq:multsignal} also captures the qualitative features of the dependence on $t^\mathrm{train}$ and $t^\mathrm{wait}$. However, the prediction overestimates the signal of mid-timed stimuli, likely because we neglect the post-stimuli degradation. We also note that this analytical description does not reproduce the quantitative features of the signal observed numerically because the signal is not just a linear superposition of the training and age impact; see Appendix E. Even though more research is needed to obtain the exact dependency of the signal on training and age, we choose to use the simple function to draw insights about the system. For instance, we observe that the ratio of the coefficient of training and the coefficient of age reduces with $n$, indicating that with every new stimulus application the impact of age on memory formation becomes stronger. The advantage of the prediction is its simplicity. Moreover, the prediction is an ad-hoc description of the parameter dependence and other choices are possible; see Appendix E. The stimulation protocol is characterised by the three parameters $n$, $t^\mathrm{train}$, and $t^\mathrm{wait}$, while the almost identical $\tau_\mathrm{pre}$ and $\tau_\mathrm{train}$ capture the characteristic time scale of network adaptation. Finally, we investigate how many stimuli an adaptive network can store. We demand that a stored stimulus can be read out at a later time, implying that its signal exceeds a given threshold $S_\mathrm{thresh}$, which captures uncertainties in the read-out apparatus as well as intrinsic noise. Fig.~\ref{fig:4}(c) shows the fraction of stimuli that can be retrieved (where $S_n > S_\mathrm{thresh}$) as a function of the total number of stimuli, $N$, and the training time~$t^\mathrm{train}$. In this case, large training times are detrimental since they age the network too much for later stimuli to be retrieved. Fig.~\ref{fig:4}(d) shows the number of stimuli that can be retrieved as a function of the training and waiting time. The largest number of stimuli is stored for smaller waiting time, as this reduces the age of the network. To find the optimal parameters for storing memory independent of a read-out apparatus specific threshold $S_{\mathrm{thresh}}$, we quantify the minimum signal out of the five stimuli's signals for varying training and waiting time, see Fig.~\ref{fig:4}(e). We observe that while the optimal waiting time is $0$, a non-zero optimal training time exists for storing memory. Taken together, our analysis reveals the strong trade-off between writing stimuli for a sufficient duration and the resulting inevitable ageing of the network that suppresses signals of subsequent stimuli. \section{Discussion} We showed that adaptive flow networks can store memory of multiple stimuli in the morphology of weak links, which cannot be revived in our model. Consequently, signatures of earlier stimuli are not destroyed by subsequent evolution, in contrast to the behaviour of typical mechanical networks~\cite{Stern2019}. Since older networks contain fewer strong links, which could shrink to store memory, the readout signal of each stimulus strongly decreases with the age of the network before the stimulus was written, which is similar to the memory plasticity observed in disordered system~\cite{Keim2011,Paulsen2014}. Conversely, the signal strength increases with its training time, i.e., the duration the stimulus is presented, similar to memory formation by directed aging in mechanical networks~\cite{Hexner2020,Pashine2019}. Taken together, we showed that adaptive flow networks reach maximal capacity at an intermediate training time, which compromises between imprinting sufficiently and ageing minimally. Our work focuses on the simple situation that non-overlapping stimuli are subsequently applied at the edge of flow networks of similar morphology. To describe, realistic living flow networks, like \textit{Physarum} or our vasculature, our work will need to be extended in multiple directions: First, the overall network geometry will have an impact on how stimuli are stored. Work in mechanical networks~\cite{Rocks2019, Stern2019} suggests that the internal timescales of flow networks and their memory capacity will depend on network size. Second, realistic systems deal with time-varying and potentially overlapping stimuli of various strengths. Third, living systems can grow and expand \cite{Ronellenfitsch2016}, implying that links can possibly regrow from their minimal size and new links can be added to the network. Taken together, it is likely that realistic adaptive flow networks show a dynamic behaviour, storing information about stimuli on various time scales. Taken together, our work quantifies the memory capacity of adaptive flow networks for varying parameters, crucial for understanding the emergent behaviour of non-neuronal organism \textit{P.~polycephalum}. \begin{acknowledgments} This work was supported by the Max Planck Society. This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program (Grant Agreement No. 947630, FlowMem). \end{acknowledgments}

https://arxiv.org/abs/2104.12337

How to Catch Marathon Cheaters: New Approximation Algorithms for Tracking Paths

Given an undirected graph, $G$, and vertices, $s$ and $t$ in $G$, the tracking paths problem is that of finding the smallest subset of vertices in $G$ whose intersection with any $s$-$t$ path results in a unique sequence. This problem is known to be NP-complete and has applications to animal migration tracking and detecting marathon course-cutting, but its approximability is largely unknown. In this paper, we address this latter issue, giving novel algorithms having approximation ratios of $(1+\epsilon)$, $O(\lg OPT)$ and $O(\lg n)$, for $H$-minor-free, general, and weighted graphs, respectively. We also give a linear kernel for $H$-minor-free graphs and make improvements to the quadratic kernel for general graphs.

\section{Deferred Figures} \begin{figure}[hp!] \centering \includegraphics[page=2,scale=.9]{figures/figures.pdf} \caption{Illustration of a tree-sink $(Tr,x)$.} \label{fig:tree-sink} \end{figure} \begin{figure}[hp!] \centering \includegraphics[page=7,scale=1.4]{figures/figures.pdf} \caption{Illustration of a relaxed $r$-division $\mathcal{R}$ (boundaries in dashed lines) and the types of cycles tracked by the output tracking set $T$ (considered in \cref{lem:feasibility:minor_free}). $C_1$-type cycles, which span a single region $R\in\mathcal{R}$ are tracked by $\mathit{OPT}(R)$. $C_2$-type cycles, which span exactly 2 regions, are not guaranteedly tracked by the 2 boundary vertices they traverse, since these may correspond to an entry-exit pair $(s',t')$. $C_3$-type cycles, which span at least 3 regions, are trivially tracked by the $\ge 3$ boundary vertices they traverse.} \label{fig:rdivision} \end{figure} \begin{figure} \centering \includegraphics[page=5,scale=1.4]{figures/figures.pdf} \caption{Illustration of proof of \cref{lem:VC_dim}, that the dual VC-dimension is bounded. If a set $Y=\{y_1,y_2,\dots,y_8\}$ is to be shattered by $\mathcal{C}_F^*$, then there must exist cycles $C_1,C_2$ traversing, respectively, $Y$ and every other vertex of $Y$. However, for large enough $Y$ ($|Y|\ge 8$), the existence of $C_1,C_2$ contradicts that $F$ is an FVS.} \label{fig:vc-dim} \end{figure} \FloatBarrier \section{Notation and Terminology}\label{app:terminology} \input{terminology} \section{Related Work}\label{app:related_work} \input{related_work} \section{Deferred Proofs on \hyperref[sec:properties]{Structural Properties}}\label{app:properties} \remRuleFour*\label{remRuleFourLbl} \begin{proof} By \cref{rem:block-cut}, an optimal tracking set must contain the union of optimal tracking sets for all $(G_i,s_i,t_i)$. Thus, the remark follows if we show that no optimal tracking set includes a cut-vertex. By \cref{rem:block-cut}, any cut-vertex $v$ of $G$ disconnects the start $s$ from the finish $t$, when removed. It follows that $v$ cannot track any entry-exit cycle, since it will always be entry/exit for any entry-exit cycle containing it. \qed \end{proof} \lemGraphSink*\label{lemGraphSinkLbl} \begin{proof} Let $Tr$ be a tree in $G'-x$ whose leaves are all adjacent to $x$ (i.e. contained in $N_{G'}(x)$). Such tree can be constructed by trimming a spanning tree of $G'-x$: iteratively remove any leaf that is not adjacent to $x$ in $G'$. Clearly, $(Tr,x)$ is a tree-sink in $G'$, and by \cref{rem:subgraph_entry_exit}, $G(Tr,x)$ has at least one entry-exit pair. If $x$ is in any such entry-exit pair, the lemma follows directly from \cref{lem:tree-sink}, so let us assume otherwise hereafter. Consider the entry-exit pair $(s',t')$ of $G(Tr,x)$ and let us root $Tr$ at $s'$. As in \cite[Lemma~8]{DBLP:journals/corr/abs-2001-03161}, consider the subtrees $Tr_1,Tr_2$ of $Tr$ determined by the edge separator connecting $t'$ to its parent vertex in $Tr$. In particular, let $Tr_1=Tr-Tr(t')$ and $Tr_2=Tr(t')$, where $Tr(v)$ denotes the subtree of $Tr$ rooted at $v\in V(Tr)$. To ensure that every leaf in $Tr_1$ is adjacent to $x$, we again repeatedly remove any leaf of $Tr_1$ that is not adjacent to $x$ (these would correspond to ancestors of $t'$ in $Tr$). Consider the two complementing cases, illustrated in \cref{fig:graph-sink}: (1) both $Tr_1$ and $Tr_2$ have at least one leaf adjacent to $x$ and (2) one of $Tr_1,Tr_2$ has no leaf adjacent to $x$. The latter case was neglected in \cite[Lemma~8]{DBLP:journals/corr/abs-2001-03161}. \begin{list} {\textbf{Case \arabic{itemcounter}.}} {\usecounter{itemcounter}\leftmargin=1em\labelwidth=4em\labelsep=.5em\itemindent=4em} \item Since there is an edge from $x$ to $Tr_2$, there exists an $x-t'$ path which does not intersect $Tr_1$. Thus, $(Tr_1,x)$ constitutes a tree-sink with entry-exit pair $(s',x)$. Similarly, since there is an edge from $x$ to $Tr_2$, $(Tr_2,x)$ constitutes a tree-sink with entry-exit pair $(x,t')$. The lemma follows from applying \cref{lem:tree-sink} to either: (a) each of the tree-sinks $(Tr_1,x)$ and $(Tr_2,x)$, when both $Tr_1$ and $Tr_2$ contain at least two vertices; or (b) to the tree-sink that contains exactly $\delta-1$ leaves adjacent to $x$ (the other tree-sink must be a single vertex when (a) does not hold). \item Since $Tr$ is rooted at $s'$ and every leaf of $Tr$ is adjacent to $x$, it must be the case that $Tr_1$ has no leaf adjacent to $x$. Thus, $(Tr_2,x)$ is a tree-sink containing all of $N_{G'}(x)$, so let us apply inductively the same reasoning we did earlier (with roles of $s',t'$ reversed), whereby we consider a subdivision of $Tr_2$ into 2 subtrees as established by the existence of an entry-exit pair $(s'',t'')$ of $G(Tr_2,x)$ -- for simplicity, we assume that $t''=t'$, since $t'$ is exit for a tree containing $Tr_2$. The lemma follows directly by the inductive hypothesis that any tracking set of $G$ contains at least $\delta-2$ vertices in $Tr_2$. The base case is a tree-sink corresponding to a star that contains all of $N_{G'}(x)$, and whose leaves are all adjacent to $x$. In this case, an entry-exit pair must belong to the star's root and one of its leaves ($x$ cannot be in an entry-exit pair, given our initial assumption that $x$ did not belong to any entry-exit pair of $Tr$, a supertree of this one). It follows that Case 1 applies to the base case setting. \end{list} \qed \end{proof} \begin{figure} \centering \begin{subfigure}[b]{.35\textwidth} \centering \includegraphics[page=3]{figures/figures.pdf} \end{subfigure} \begin{subfigure}[b]{.35\textwidth} \centering \includegraphics[page=4]{figures/figures.pdf} \end{subfigure} \caption{Illustration of proof of \cref{lem:graph-sink}, with Case~1 on the left (both $Tr_1$ and $Tr_2$ have leaves adjacent to $x$) and Case~2 on the right (just $Tr_2$ has leaves adjacent to $x$).} \label{fig:graph-sink} \end{figure} \section{Deferred Proofs on \hyperref[sec:properties]{$H$-Minor-Free Graphs}}\label{app:minor-free} \subsection{Linear Kernel} \lemBipartite*\label{lemBipartiteLbl} \begin{proof} To show the bound on the size of vertex set $V$, we construct a new graph from $B$ as follows. Replace every vertex $v$ in $V$ and its incident edges in $B$ by an edge connecting any two of its neighbors. Observe that, this operation results in an $H$-minor-free graph as it is equivalent to contraction of any edge incident to $v$, followed by the deletion of all but one of the remaining edges incident to $v$). The resulting graph has vertex set $U$, exactly $|V|$ edges, and at most $\delta$ parallel edges between any pair of vertices (this follows from (ii)). By \cref{thm:mader}, any simple $H$-minor-free graph with $|U|$ vertices has at most $\sigma_H|U|$ edges and, thus at most $\delta\sigma_H|U|$ edges when there exist at most $\delta$ parallel edges between any pair of vertices. \qed \end{proof} \lemCutSet*\label{lemCutSetLbl} \begin{proof} Let us partition $V(G-F)$ into $V_1$, $V_2$, $V_{\ge 3}$ corresponding to the sets of vertices whose degree in $G-F$ is (respectively) $1$ (a.k.a. leaves), $2$ or at least $3$. Further, let $V^X$ denote the set of vertices in $V(G-F)$ which are endpoints of an edge in $X$, i.e., the set of vertices in $V(G-F)$ adjacent to a vertex in $F$. Since each vertex in $V_1$ must be adjacent to a vertex in $F$ ($G$ contains no degree-1 vertices since it's reduced), we have that $V_1\subseteq V^X$ and, thus, \[|V_1|\le|X|\] Moreover, $|V_{\ge 3}|\le |V_1|-2$ (this is a well known fact applicable to any forest). Thus, \[|V_{\ge 3}|\le |X|-2\] Next, we bound $|V_2|$. Consider the set $V_2\setminus V^X$ of vertices in $V_2$ which are not adjacent to any vertex in $F$. By \hyperref[rule:3]{Rule~3}, the set $V_2\setminus V^X$ induces an independent set. Hence, $|V_2\setminus V^X|$ is at most the number of edges in the forest that results from replacing each $v \in V_2\setminus V^X$ by an edge connecting $v$'s neighbors, giving us \[|V_2\setminus V^X| \le |V_1|+|V_{\ge3}|+|V_2\cap V^X|-1,\\\] which implies \[|V_2|\le |V_1|+2|V_2\cap V^X|+|V_{\ge3}|-1\] Thus, \begin{align*} |V(G-F)| & \le 2V_1+2|V_2\cap V^X|+2|V_{\ge3}|-1 \\ & = 2|V_1\cap V^X|+2|V_2\cap V^X|+2|V_{\ge3}|-1 && (V_1\subseteq V^X)\\ & \le 2|X|+2|V_{\ge3}|-1 \\ & \le 4|X|-5 \end{align*} \qed \end{proof} \subsection{EPTAS} \lemOptEachR*\label{lemOptEachRLbl} \begin{proof} It suffices to enumerate all the $2^{|V(R)|}$ possible subsets and, for each, verify in $O(n^{O(1)})$ time whether every entry-exit cycle of $\mathcal{C}(R)$ is tracked. The verification step can be done in a way similar to the verification algorithm given by Banik \textit{et al.} \cite{DBLP:journals/algorithmica/BanikCLRS20} to show that the problem is in NP: from the observation that every tracking set $X$ is an FVS (see \cref{rem:ts_is_fvs}), it follows that there is at most $O(n^{O(1)})$ entry-exit cycles not tracked by $X$ (see also \cref{lem:general:poly_cycles}). \qed \end{proof} \lemFeasibility*\label{lemFeasibilityLbl} \begin{proof} It is enough to argue that $T$ tracks every cycle $C$ of $K$ (by \cref{lem:tracking_set} and \cref{obs:tracking_set_K}). We consider 3 types of cycles, illustrated in \cref{fig:rdivision}. If $C$ spans exactly 1 region $R$ in $\mathcal{R}$, then it is guaranteed to be tracked by feasibility of $\mathit{OPT}(R)$ (see \cref{lem:opt_each_R}). If $C$ spans at least 3 regions in $\mathcal{R}$, then $C$ is trivially tracked by the trackers in, at least, 3 boundary vertices. Otherwise, let $C$ be a cycle spanning exactly 2 regions. We argue that $C$ is also trivially tracked. Let $R_1$ and $R_2$ be the two regions spanned by $C$ and let us assume that $C$ traverses exactly 2 boundary vertices $b_1$ and $b_2$ (if it traverses more boundary vertices, it must be trivially tracked, and if it traverses less, then it could not span more than one region). If $C$ contains a tracker in $\mathit{OPT}(R_1)\cup\mathit{OPT}(R_2)\setminus\{b_1,b_2\}$, we are done. Otherwise, $C$ is the union of a path in $\Pi(R_1)$ and a path in $\Pi(R_2)$, by definition of $\Pi$. In this case, however, $C$ must have a third tracker placed in $N(\{b_1,b_2\})\cap \{int(R_1)\cup int(R_2)\}$ (notice that $C$ contains at least one non-boundary vertex that is a neighbor of $b_1$ or $b_2$, since there are no parallel edges). \qed \end{proof} \lemSmallNeighborhoods*\label{lemSmallNeighborhoodsLbl} \begin{proof} The proof is similar in spirit to that of \cref{lem:linear_size_minor_free_biconnected}. Clearly, $\partial(\Pi(R))$ is an FVS for $\Pi(R)$ (if it were not, there would be untracked cycles in $R$ contradicting feasibility of $\mathit{OPT}(R)$), so $\Pi(R)-\partial(\Pi(R))$ is a forest. We assume w.l.o.g. that $|\partial(\Pi(R))|\ge 2$. Below, we make some claims about the structure of $\Pi(R)$: \begin{list} {\textit{Claim \arabic{itemcounter}:}} {\usecounter{itemcounter}\leftmargin=1em\labelwidth=4em\labelsep=.5em\itemindent=4em} \item Let $b_1,b_2$ be two vertices in $\partial(\Pi(R))$. There exist at most 3 trees in $\Pi(R)-\partial(\Pi(R))$ that are adjacent\footnote{In this context, a tree is adjacent to $v$ if it includes a vertex that is adjacent to $v$.} to both $b_1$ and $b_2$. \item Every tree in $\Pi(R)-\partial(\Pi(R))$ is adjacent to at least 2 vertices in $\partial(\Pi(R))$. \item Every tree in $\Pi(R)-\partial(\Pi(R))$ contains at most 3 vertices adjacent to the same vertex in $\partial(\Pi(R))$. \end{list} The first claim follows from \cref{lem:tree-sink} (if there existed 4 or more trees adjacent to both $b_1$ and $b_2$, there would have to be a tracker from $\mathit{OPT}(R)$ in one of the trees, a contradiction). The last claim follows from \cref{lem:tree-sink} in a similar fashion. The second claim follows from the definition of $\Pi(R)$ and \hyperref[rule:1]{Rules~1},\hyperref[rule:2]{2}. Let us contract each tree $Tr$ in $\Pi(R)-\partial(\Pi(R))$ into a \emph{tree vertex} $v_{Tr}$ and let $F$ be the set of all tree vertices. Notice that this may create parallel edges between a vertex in $\partial(\Pi(R))$ and a tree vertex, but never between two vertices in $\partial(\Pi(R))$ or $F$. In addition, let us remove any edges between vertices in $\partial(\Pi(R))$. The resulting graph is bipartite, with vertex set partitioned into $\partial(\Pi(R))$ and $F$, and is $H$-minor-free (since the class of minor-free graphs is minor-closed). By Claims 1 and 2, at most 3 vertices in $F$ share the same pair of neighbors, and every vertex in $F$ has degree at least 2. Hence, by \cref{lem:bipartite}, \[|F|\le3\sigma_H|\partial(\Pi(R))|\] As a consequence of Claim 3, there are at most 3 parallel edges between a vertex in $\partial(\Pi(R))$ and a vertex in $F$. Thus, by \cref{thm:mader}, the set of edges in the bipartite graph is at most \[3\cdot \sigma_H(|F|+|\partial(\Pi(R))|)\le (9\sigma_H^2+3\sigma_H)|\partial(\Pi(R))|\] The lemma follows from the fact that the edges in the bipartite graph, including the parallel ones, have a 1-1 correspondence with the vertices in $\mathcal{N}(R)$. \qed \end{proof} \section{Deferred Proofs on \hyperref[sec:general]{General Graphs}}\label{app:general} \subsection{Approximation Algorithm}\label{app:subsec:approx} \thmLogNApprox*\label{thmLogNApproxLbl} Let us argue about feasibility first. Let $T=S\cup F$ be the output of the algorithm for weighted graphs described in \cref{sec:general}. \begin{lemma} \label{lem:feasibility} $T$ is a tracking set of $G$. \end{lemma} \begin{proof} By \cref{lem:tracking_set}, it is enough to argue that every entry-exit cycle of $G$ is tracked by $T$. Let $F$ be the FVS computed in the first step of the algorithm and $S$ the set cover computed in the second step. Since $F\subseteq T$, any entry-exit cycle containing at least 3 vertices from $F$ is trivially tracked. All remaining entry-exit cycles are tracked by $S\subseteq T$, by definition. \qed \end{proof} Next, we argue about the approximation ratio. The lemma below follows from a proof by Banik \textit{et al.} \cite{DBLP:journals/algorithmica/BanikCLRS20} that \textsc{Tracking}{} is in NP. \begin{lemma} \label{lem:general:poly_cycles} $|\mathcal{C}_F|\le O(n^8)$, where $n$ is the number of vertices of the input graph $G$. \end{lemma} \begin{proof} Since $F$ is an FVS and $G$ is reduced by \hyperref[rule:1]{Rule~1}, every entry-exit cycle of $\mathcal{C}_F$ includes at least 1 vertex from $F$ (see \cref{rem:ts_is_fvs}) and, by definition, at most 2 vertices from $F$. Since $G-F$ is a forest, there exists at most 1 path between every pair of vertices in $G-F$. Further, each vertex $f$ in $F$ has at most $n-|F|$ neighbors in $G-F$, so there are at most $\binom{n-|F|}{2}$ cycles which contain $f$ and no other vertex from $F$. Thus, the number of cycles containing exactly 1 vertex from $F$ is at most \[|F|\binom{n-|F|}{2}\le n^3\] Let us now argue about cycles containing exactly 2 vertices $f_1$ and $f_2$ from $F$. Any such cycle is defined by a pair of paths $P$ and $Q$ between $f_1$ and $f_2$ traversing vertices in $V(G-F)\cup \{f_1,f_2\}$. Let us first handle cycles where one of the paths $P,Q$ is a single edge, say $Q$ (because $G$ is simple, the other path, $P$, must consist of at least 2 edges). Clearly, every path $P$ is identified by a path in $G-F$ connecting a neighbor $p_1$ of $f_1$ to a neighbor $p_2$ of $f_2$. Therefore, there exist at most $(n-|F|)^2$ such paths $P$\footnote{In contrast to cycles containing a single vertex from $F$, the neighbors $p_1,p_2$ connected by $P$ may be the same vertex, hence we allow repetitions when counting the number of pairs of neighbors.} and, thus, at most $(n-|F|)^2$ cycles where $f_1$ and $f_2$ are connected by an edge. Similarly, for cycles where $Q$ is not an edge, every path $Q$ is identified by a path in $G-F$ connecting a neighbor $q_1\ne p_1$ of $f_1$ to a neighbor $q_2\ne p_2$ of $f_2$. Hence, there are at most $(n-|F|-1)^2$ such paths $Q$ and, thus, at most $(n-|F|)^2\cdot (n-|F|-1)^2\le (n-|F|)^4$ cycles where $f_1$ and $f_2$ are not connected by an edge. Taking into account all pairs $f_1,f_2$ in $F$, the number of cycles containing exactly 2 vertices from $F$ is at most \[\binom{|F|}{2}\left((n-|F|)^2+(n-|F|)^4\right)= O(n^6)\] For every cycle $C$, there exist at most $|V(C)|(|V(C)|-1)\le n^2$ entry-exit pairs. Therefore, the number of entry-exit cycles of $C_F$ is at most \[n^2(n^3 + O(n^6))=O(n^8)\] \qed \end{proof} Let us denote by $\mathit{OPT}$ the size of an optimal tracking set of $G$, and let $w(T) = \sum_{a\in T} w(a)$ be the total weight of $T$. \begin{lemma} \label{lem:approx} $w(T)= O(\lg n)\mathit{OPT}$. \end{lemma} \begin{proof} By union bound, we have that $w(T)\le w(S)+w(F)$, where $S$ and $F$ are the sets computed in steps 1 and 2, respectively, of the above algorithm. Since $F$ is a 2-approximate FVS and every optimal tracking set is also an FVS (by \cref{rem:ts_is_fvs}), we have that \[w(F)\le 2 \mathit{OPT}\] Let $\mathit{OPT}_F$ be the size of an optimal solution to the covering problem of step 2, concerning all entry-exit cycles in $\mathcal{C}_F$. Since every optimal tracking set must track all entry-exit cycles in $\mathcal{C}_F$, we have that $\mathit{OPT}\ge OPT_F$. Further, the well known greedy algorithm for \textsc{SetCover}{} gives us an approximation ratio of at most $(1+\ln |\mathcal{C}_F|)$. Thus, by \cref{lem:general:poly_cycles}, $w(S)=O(\lg n)\mathit{OPT}_F$ and therefore, \[w(S)=O(\lg n)\mathit{OPT}\] The lemma follows. \qed \end{proof} \hyperref[thmLogNApproxLbl]{\namecref{thm:logN-approx}~\labelcref{thm:logN-approx}} follows from \cref{lem:feasibility,lem:approx}. \subsection{Quadratic Kernel}\label{app:subsec:quadratic} In this section, we focus on $k$-\textsc{Tracking}{}, the decision version of \textsc{Tracking}{} which asks whether there exists a tracking set of size at most $k$. We consider a parameterization with parameter $k$ itself (often called the \emph{natural} parameter) and give a kernelization algorithm that produces a quadratic kernel for general graphs, by building on the quadratic kernel of Choudhary and Raman \cite{DBLP:journals/corr/abs-2001-03161}. While simpler, our proof of the kernel size completes the case analysis (see \cref{lem:graph-sink}) for one of the lemmas central to the kernelization algorithm of \cite{DBLP:journals/corr/abs-2001-03161} (specifically, \cite[Lemma~8]{DBLP:journals/corr/abs-2001-03161}). Moreover, our kernelization algorithm yields a kernel size with considerably smaller constants. We achieve this by expanding on the notion of tree-sink structures (see \cref{subsec:lower_bounds}), allowing us to bound the maximum degree among non-cut vertices in the kernel (see \cref{cor:max_degree}). Our kernelization algorithm is simply the exhaustive application of \hyperref[rule:1]{Rules~1}, \hyperref[rule:2]{2} and \hyperref[rule:3]{3} (see \cref{subsec:reduction_rules}) in no particular order, followed by application of \hyperref[rule:4]{Rule~4}, and then of \hyperref[rule:5]{Rule~5}: \begin{list} {\textbf{Rule \arabic{itemcounter}.}} {\usecounter{itemcounter}\leftmargin=.7in\rightmargin=0em\labelwidth=3in} \setcounter{itemcounter}{3} \item If there exists a non-cut vertex of degree more than $k+2$, return a trivial NO-instance.\label{rule:4} % \item If the number of vertices (resp. edges) is more than $4k^2+9k-5$ (resp. $5k^2+11k-6$), return a trivial NO-instance.\label{rule:5} \end{list} \begin{lemma}\label{lem:rule4} \hyperref[rule:4]{Rule~4} is safe and can be done in polynomial-time. \end{lemma} \begin{proof} Follows from \cref{cor:max_degree}. \qed \end{proof} Next, we show that the last rule is also safe. \begin{lemma}\label{lem:quadratic_size_biconnected} Let $G$ be a biconnected reduced graph, with start $s$ and finish $t$. Then, $G$ has at most $4\mathit{OPT}^2+9\mathit{OPT}-5$ vertices and at most $5\mathit{OPT}^2+11\mathit{OPT}-6$ edges, where $\mathit{OPT}$ denotes the size of an optimal tracking set of $G$. \end{lemma} \begin{proof} Let $T^*$ be an optimal tracking set of $G$, i.e., $|T^*|=\mathit{OPT}$. Since every tracking set is an FVS of a reduced graph (see \cref{rem:ts_is_fvs}), we can apply \cref{lem:cutSet} and obtain $|V(G-T^*)|\le 4|X|-5$, where $X$ is the set of edges with endpoints in both $T^*$ and $G-T^*$. By the fact that $G$ is biconnected and by \cref{cor:max_degree}, $G$ has maximum degree $\mathit{OPT}+2$ and, hence, $|X|\le \mathit{OPT}(\mathit{OPT}+2)$. It follows that \[|V(G)|\le 4\mathit{OPT}^2+9\mathit{OPT}-5\] The edges of $G$ consist of edges with no endpoint in $T^*$ (at most $|V(G-T^*)-1|$) and edges with at least one endpoint in $T^*$ (at most $\mathit{OPT}(\mathit{OPT}+2)$ by \cref{cor:max_degree}), giving us \[|E(G)|\le 5\mathit{OPT}^2+11\mathit{OPT}-6\] \qed \end{proof} We can now apply the latter lemma individually to each biconnected component, giving us the following. \begin{lemma}\label{lem:quadratic_size_general} Any reduced graph $G$ with start $s$ and finish $t$ has at most $4\mathit{OPT}^2+9\mathit{OPT}-5$ vertices and at most $5\mathit{OPT}^2+11\mathit{OPT}-6$ edges, where $\mathit{OPT}$ denotes the size of an optimal tracking set of $G$. \end{lemma} \begin{proof} Let $G_i$ denote the $i\textsuperscript{th}$ biconnected component of $G$, with entry-exit vertices $s_i,t_i$ (see \cref{rem:opt_from_biconnected_cmps}). Further, let $\mathit{OPT}_i$ denote the size of a minimum tracking set of $(G_i,s_i,t_i)$. It follows from \cref{rem:opt_from_biconnected_cmps} that $\mathit{OPT}=\sum_{i}\mathit{OPT}_i$. Moreover, \cref{lem:quadratic_size_biconnected} gives us $|V(G_i)|\le p(\mathit{OPT}_i)$, where $p(x)=4x^2+9x-5$. Thus, \begin{align*} |V(G)| & \le \sum_i |V(G_i)|\\ & \le \sum_i p(\mathit{OPT}_i) && (\text{\cref{lem:quadratic_size_biconnected}})\\ & \le p\left(\sum_i \mathit{OPT}_i\right) && (\text{$p$ is degree-2 polynomial})\\ & = p(\mathit{OPT}) && (\text{\cref{rem:opt_from_biconnected_cmps}}) \end{align*} The number of edges in $G$ can be upper bounded in a similar manner. \qed \end{proof} The latter lemma immediately implies the safety of \hyperref[rule:5]{Rule~5}, as well as an $O(\sqrt{n})$-approximation algorithm (output all the vertices in the kernel). \begin{lemma}\label{lem:rule5} \hyperref[rule:5]{Rule~5} is safe and can be done in polynomial-time. \end{lemma} Correctness of \hyperref[rule:1]{Rules~1}, \hyperref[rule:2]{2} and \hyperref[rule:3]{3} (\cite{DBLP:journals/algorithmica/BanikCLRS20,DBLP:journals/corr/abs-2001-03161,DBLP:conf/isaac/EppsteinGLM19}), as well as of \hyperref[rule:4]{Rules~4} and \hyperref[rule:5]{5} (\cref{lem:rule4,lem:rule5}) immediately give us the following. \thmQuadraticKernel*\label{thmQuadraticKernelLbl} The latter theorem improves a quadratic kernel of Choudhary and Raman \cite{DBLP:journals/corr/abs-2001-03161}, whose size is bounded by $140k^2-45k$ vertices and $180k^2+65k$ edges. \subsection{Approximation Algorithm}\label{subsec:approx_general} \fi We reduce an instance $(G,s,t,w')$ of \textsc{WeightedTracking}{}, for a weight function $w':V(G)\rightarrow \mathbb{Q}$, into an instance $(\mathcal{U},\mathcal{X},w)$ of \textsc{SetCover}{}, which asks for the sub-collection of $\mathcal{X}$ of minimum total weight, whose union equals the universe $\mathcal{U}$. Here, $(\mathcal{U},\mathcal{X})$ defines a set system, i.e., a collection $\mathcal{X}$ of subsets of a set $\mathcal{U}$, and $w$ is the weight function $w:\mathcal{X}\rightarrow \mathbb{Q}$. It is well known that there exist greedy polynomial-time algorithms achieving approximation ratios of $(1+\ln M)$ \ifFull\cite{DBLP:journals/mor/Chvatal79,DBLP:journals/jcss/Johnson74a,DBLP:journals/dm/Lovasz75}~\fi or of $(1+\Delta)$ \ifFull\cite{DBLP:journals/siamcomp/Hochbaum82}\else\cite{DBLP:books/daglib/0004338,DBLP:books/daglib/0030297}\fi, where $M$ is the size of the largest set in $\mathcal{X}$ and $\Delta$ is the maximum number, over all elements $u$ in $\mathcal{U}$, of sets in $\mathcal{X}$ that contain $u$. \ifFull At first glance, this reduction does not seem to yield a useful approximation ratio, given the exponential number of entry-exit cycles that need to be covered, but we will show that one can limit this number to a polynomial of fixed degree. ~ \fi Let $\mathcal{C}$ be the set of all entry-exit cycles in our input graph $G$, which we assume w.l.o.g. to be reduced by \hyperref[rule:1]{Rule~1}. Further, let $\mathcal{C}_F$ be the set of all entry-exit cycles in $G$, each of which contains at most 2 vertices from the subset $F\subseteq V$. That is, $\mathcal{C}_F\vcentcolon= \left \{ (C,s',t')\in \mathcal{C}: |C\cap F| \le 2\right\}$. Our algorithm is as follows. \begin{mdframed} \begin{list} {\textbf{\arabic{itemcounter}.}} {\usecounter{itemcounter}\leftmargin=1.5em\rightmargin=0em\labelwidth=3in} \item Compute a 2-approximate FVS $F$ of $G$ (see \ifFull\cite{DBLP:conf/isaac/BafnaBF95,DBLP:journals/orl/ChudakGHW98,DBLP:conf/uai/BeckerG94}\else\cite{DBLP:books/daglib/0004338,DBLP:books/daglib/0030297}\fi). \item Use the greedy algorithm of \ifFull\cite{DBLP:journals/mor/Chvatal79,DBLP:journals/jcss/Johnson74a,DBLP:journals/dm/Lovasz75}\else\cite{DBLP:books/daglib/0004338,DBLP:books/daglib/0030297}\fi~to compute an approximate set covering, $S\subseteq V(G)$, for an instance $(\mathcal{U},\mathcal{X},w)$ of \textsc{SetCover}{} where: \begin{enumerate}[(i)] \item the universe, $\mathcal{U}$, of elements to be covered is $\mathcal{C}_F$ \item the collection of covering sets, $\mathcal{X}$, is a 1-1 correspondence with $V(G)$, where each covering set with corresponding vertex $v$ is the subset of $\mathcal{C}_F$ which are tracked by $v$, that is, \[\mathcal{X}=\{\{(C,s',t') \in \mathcal{C}_F\mid v\text{ tracks } (C,s',t')\}\}_{v\in V(G)}.\] \item the weight function $w$ is the weight function $w'$ defined for \textsc{WeightedTracking}{}, given the 1-1 correspondence between $\mathcal{X}$ and $V(G)$. \end{enumerate} \item Output $T=S \cup F$. \end{list} \end{mdframed} \ifFull copy from appendix \else We can show that $|\mathcal{C}_F|=O(n^{O(1)})$. From the observation that every tracking set $F$ is an FVS (see \cref{rem:ts_is_fvs}), it follows that there are at most $O(n^{O(1)})$ entry-exit cycles not tracked by $F$. Thus, our claim follows (details in \cref{app:subsec:approx}). \fi \begin{restatable}{theorem}{thmLogNApprox} \ifFull\else\hyperref[thmLogNApproxLbl]{$\circledast$}\fi \label{thm:logN-approx} ~\textsc{WeightedTracking}{} admits an $O(\lg n)$-approximation\ifFull algorithm\fi. \end{restatable} \Paragraph{Unweighted Graphs.} We show that the dual of the above set cover formulation has bounded VC-dimension \cite{DBLP:journals/dcg/HausslerW87,vapnik2015uniform}. This immediately improves the approximation ratio to $O(\lg \mathit{OPT})$ for \textsc{Tracking}{} (unweighted version) as a consequence of a result by Br{\"{o}}nnimann and Goodrich \cite{DBLP:journals/dcg/BronnimannG95}, which establishes an approximation-ratio of $O(d\lg(dc))$ for unweighted set cover instances with dual VC-dimension $d$ and optimal covers of size at most $c$. Let $(\mathcal{U},\mathcal{X})$ be a set system and $Y$ a subset of $\mathcal{U}$. We say that $Y$ is \emph{shattered} if $\mathcal{X}\cap Y=2^Y$, where $\mathcal{X}\cap Y\vcentcolon= \{X\cap Y\mid X\in \mathcal{X}\}$. In other words, $Y$ is shattered if the set of intersections of $Y$ with each $X\in\mathcal{X}$ contains all the possible subsets of $Y$. The set system $(\mathcal{U},\mathcal{X})$ has \emph{VC-dimension} $d$ if $d$ is the largest integer for which there exists a subset $Y\subseteq \mathcal{U}$, of cardinality $|Y|=d$, that can be shattered. The dual problem of an unweighted instance $(\mathcal{U},\mathcal{X})$ of \textsc{SetCover}{} is finding a \emph{hitting set} of minimum size, where a hitting set is a subset of $\mathcal{U}$ that has a non-empty intersection with every set in $\mathcal{X}$. In our case, it corresponds to finding the smallest subset of entry-exit cycles that covers every vertex, where a vertex is covered if it tracks least one entry-exit cycle in the subset. This is equivalent to an unweighted instance of \textsc{SetCover}{} with set system $(V, \mathcal{C}_F^*)$, where $V=V(G)$ and $\mathcal{C}_F^*\vcentcolon= \{V(C)\setminus\{s',t'\}: (C,s',t')\in \mathcal{C}_F\}$ is the collection of sets, one for each entry-exit cycle, of vertices which can track that entry-exit cycle. \begin{lemma \label{lem:VC_dim} The set system $(V,\mathcal{C}_F^*)$ has VC-dimension at most 9. \end{lemma} \begin{proof} We show that there exists no subset $Y\subseteq V$ of size $|Y|\ge 10$ that can be shattered by $\mathcal{C}_F^*$. Since every element of $\mathcal{C}_F^*$ contains at most 2 vertices from $F$ (by definition of $\mathcal{C}_F$), we cannot have more than 2 vertices from $F$ in $Y$ (since we would then require an entry-exit cycle containing at least 3 vertices in $F$ to shatter $Y$). Thus, the lemma follows if we show that no subset $Y\subseteq V\setminus F$ of size $|Y|\ge 8$ can be shattered by $\mathcal{C}_F^*$. Let us assume, by contradiction, that this is possible. Then, if $Y\subseteq V\setminus F$ is to be shattered by $\mathcal{C}_F^*$, there must exist 2 entry-exit cycles $(C_1,s'_1,t'_1)$ and $(C_2,s'_2,t'_2)$ in $\mathcal{C}_F$ (see \cref{fig:vc-dim}\ifFull\else ~$\hyperref[fig:vc-dim]{\circledast}$\fi), such that: \begin{itemize} \item $C_1$ traverses all vertices of $Y$, say in the order $y_1,y_2,\dots,y_{|Y|}$ (for all $y_j\in Y$), \item $C_2$ traverses every other vertex of $Y$ traversed by $C_1$, say $Y'=\{y_2,y_4,\dots,y_{|Y|}\}$, but not necessarily in the same order (we assume w.l.o.g. $|Y|$ is even). \end{itemize} Consider the graph consisting of the union of the cycles $C_1,C_2$. Let us contract every shared edge between $C_1,C_2$. Note that $C_1$ remains a cycle that traverses $Y$ and $C_2$ remains a cycle that traverses $Y'$ but not any vertex of $Y\setminus Y'$. So we can safely assume that $C_1$ and $C_2$ do not share any edges. Thus, the union of $C_1,C_2$ is a graph with $|C_1|+|C_2|-|Y|/2$ vertices and $|C_1|+|C_2|$ edges. Since both entry-exit cycles are in $\mathcal{C}_F$, each of $C_1,C_2$ shares at most 2 vertices with $F$. Let us remove such vertices, say there's $k\le4$ of them. The result is a graph with $|C_1|+|C_2|-|Y|/2-k$ vertices and, at best, $|C_1|+|C_2|-2k$ edges (the removed vertices cannot be in $Y$, so they have degree 2). In order for this graph to be acyclic (since $F$ is an FVS by \cref{rem:ts_is_fvs}, and our contractions preserve cycles) we would then require $|Y|<8$ (since any acyclic graph with $n$ vertices has at most $n-1$ edges), a contradiction. \qed \end{proof} The above lemma, combined with the result of Br{\"{o}}nnimann and Goodrich \cite{DBLP:journals/dcg/BronnimannG95} gives us the following. \begin{theorem} \textsc{Tracking}{} admits an $O(\lg \mathit{OPT})$-approximation\ifFull algorithm\fi, where $\mathit{OPT}$ is the size of an optimal tracking set. \end{theorem} \ifFull \subsection{Quadratic Kernel}\label{subsec:kernel_general} \todo{copy from appendix} \fi \subsubsection{Outline.} \subsubsection{Preliminaries} Our algorithm relies on \ifFull an appropriate decomposition of the graph into regions, which can be accomplished by recursively finding small~\fi\emph{balanced separators}, sets of vertices whose removal partitions the graph into two roughly equal-sized parts. \ifFull For simplicity, we define balanced separators in the context of unweighted vertices, but this can be generalized to a weighted setting. \begin{definition}[Balanced separator] Let $G$ be a graph, and let $X\subseteq V(G)$ be a subset of its vertices. We say that $X$ is a \emph{balanced separator} if $V(G)\setminus X$ can be partitioned into two sets $A$, $B$ each of which has size at most $2/3|V(G)|$, such that no edge joins a vertex in $A$ with a vertex in $B$. \end{definition} \fi Ungar \cite{ungar1951theorem} first showed that every $n$-vertex planar graph has a balanced separator of size $O(\sqrt{n}\lg^{3/2} n)$. This was later improved by Lipton and Tarjan \cite{lipton1979separator} to $\sqrt{8n}$, and Goodrich~\cite{goodrich1995planar} showed how to compute these recursively in linear time. The Lipton-Tarjan separator theorem has been further refined \ifFull \cite{chung1988separator,djidjev1982problem,DBLP:journals/jcss/Miller86,DBLP:conf/sigal/GazitM90,DBLP:conf/compgeom/SpielmanT96,DBLP:journals/acta/DjidjevV97,DBLP:journals/jea/Fox-EpsteinMP016} \else (e.g., see~\cite{chung1988separator,DBLP:journals/acta/DjidjevV97}) \fi and generalized to bounded-genus graphs \ifFull \cite{DBLP:journals/jal/GilbertHT84,djidjev1985linear,DBLP:journals/siamcomp/Kelner06} \else (e.g., see~\cite{DBLP:journals/jal/GilbertHT84,djidjev1985linear}) \fi as well as to $H$-minor-free graphs \ifFull \cite{DBLP:conf/stoc/AlonST90,DBLP:conf/soda/PlotkinRS94,DBLP:journals/talg/ReedW09,DBLP:journals/jacm/BiswalLR10,DBLP:conf/focs/KawarabayashiR10,DBLP:conf/focs/Wulff-Nilsen11}% \else (e.g., see~\cite{DBLP:conf/stoc/AlonST90,DBLP:journals/talg/ReedW09})% \fi. \begin{theorem}[Minor-free Separator Theorem \cite{DBLP:conf/stoc/AlonST90}]\label{thm:minor_free_separator} Let $G$ be an $H$-minor-free graph with $n$ vertices, where $H$ is a simple graph with $h\ge 1$ vertices. Then a balanced separator for $G$ of size at most $c_H^1\sqrt{n}$ can be found in $O(h^{O(1)}n^{O(1)})$ time, where $c_H^1$ is a positive constant depending solely on $h$. \end{theorem} We use the Minor-free Separator Theorem recursively to decompose the graph into a set $\mathcal{R}$ of edge-disjoint subgraphs, called \emph{regions}. The vertices of a region $R\in \mathcal{R}$ which belong to at least one other region are called \emph{boundary vertices} and the set of these vertices is denoted by $\partial(R)$. The remaining vertices of $R$ are called \emph{interior vertices} and are denote by $int(R)$. \begin{definition}[Relaxed $r$-division] A \emph{relaxed $r$-division} of an $n$-vertex graph $G$ is a decomposition of $G$ into $\Theta(n/r)$ regions, each of which has at most $r$ vertices, such that the total number boundary vertices is $O(n/\sqrt{r})$\ifFull, for a positive integer $r$\fi. \end{definition} \ifFull A relaxed $r$-division was perhaps among the first applications of Lipton and Tarjan's Planar Separator Theorem, and it was used to obtain EPTASs for maximum independent set \cite{DBLP:journals/siamcomp/LiptonT80} and for minimum vertex cover \cite{chiba1981applications}. A relaxed $r$-division is a relaxed version of an $r$-division, a decomposition introduced by Frederickson \cite{DBLP:journals/siamcomp/Frederickson87} which additionally requires every region to have $O(\sqrt{r})$ boundary vertices. Both decompositions can be constructed in $O(n \lg n)$ time, by recursively splitting the graph using balanced separators. Typically, an $r$-division is constructed from a relaxed $r$-division -- it is shown in \cite{DBLP:journals/siamcomp/Frederickson87} that, after constructing a relaxed division, one can further split the big regions (the ones violating the extra condition) without asymptotically increasing the total number of regions or boundary vertices. Even though $r$-divisions were introduced for planar graphs \cite{DBLP:journals/siamcomp/Frederickson87}, its derivation can easily be generalized to any class of graphs that is characterized by the existence of sublinear balanced separators, including $H$-minor-free graphs. \else Computing a relaxed $r$-division is the first step in Frederickson's algorithm \cite{DBLP:journals/siamcomp/Frederickson87} for constructing an $r$-division in a planar graph, a decomposition which additionally requires every region to have $O(\sqrt{r})$ boundary vertices (we won't need this property). Both decompositions can easily be generalized to any class of graphs that is characterized by the existence of sublinear balanced separators, which includes $H$-minor-free graphs. \fi \begin{theorem}[Minor-free Separator Theorem (\labelcref{thm:minor_free_separator}) + Frederickson \cite{DBLP:journals/siamcomp/Frederickson87}] There is an $O(n\lg n)$ algorithm that, given an $H$-minor-free graph $G$ and a positive integer $r$, computes a relaxed $r$-division of $G$. \end{theorem} \ifFull \begin{proof} Follows from Minor-free Separator Theorem (\labelcref{thm:minor_free_separator}) and Frederickson \cite{DBLP:journals/siamcomp/Frederickson87}. \ifFull\qed\fi \end{proof} \fi \ifFull A relaxed $r$-division is simpler to construct and will be sufficient for our purposes, so we will use these instead. \fi \ifFull \Paragraph{Algorithm.} \fi Our strategy will be to (i) construct a relaxed $r$-division of a \ifFull linear kernel $K$ of the input $H$-minor-free graph $G$ (such that $K$ is itself a $O(1)$-approximate tracking set), \else smaller graph, $K$, which is itself an $O(1)$-approximate tracking set, \fi (ii) solve optimally for each region, and (iii) combine the solutions for each region into a solution for the original graph with quality comparable to that of an optimal solution. \ifFull This approach has been used to obtain EPTASs for minimum FVS \cite{DBLP:conf/sea2/BorradaileLZ19,ZhengPhdThesis}, and (without the need for a linear kernelization) maximum independent set \cite{DBLP:journals/siamcomp/LiptonT80}, as well as minimum vertex cover \cite{chiba1981applications}. \else This approach has been used to obtain EPTASs for minimum FVS \cite{DBLP:conf/sea2/BorradaileLZ19,ZhengPhdThesis}, maximum independent set \cite{DBLP:journals/siamcomp/LiptonT80} and minimum vertex cover \cite{chiba1981applications}. \fi However, and in contrast to these problems, the step of constructing a close to optimal solution from the solutions \ifFull computed\fi of each region is not obvious. Indeed, the difficulty of this step emerges from the very ``nonlocal'' structure of \textsc{Tracking}{}, which requires special attention to the location of \ifFull the start-finish~\fi $(s,t)$ in the graph% \ifFull~within the graph, in addition to requiring tracking cycles that escape out of local neighborhoods (as in minimum FVS). The latter issue complicates not only the intermediate step of solving optimally for each region, but also the step of combining solutions for each region -- an entry-exit cycle spanning two regions may not be tracked by just the boundary vertices it traverses, depending on the position of $s$ and $t$; this is in contrast to minimum FVS, which requires only one tracker per cycle rendering the combining step trivial. \else , in addition to the nonlocal structure of cycles, as illustrated in \cref{fig:rdivision}\ifFull\else ~$\hyperref[fig:rdivision]{\circledast}$\fi. \fi Our EPTAS is as follows: \begin{mdframed} \begin{list} {\textbf{\arabic{itemcounter}.}} {\usecounter{itemcounter}\leftmargin=1.5em\rightmargin=0em\labelwidth=3in} \item Compute a linear kernel $K$ of $G$ by reducing it with \hyperref[rule:1]{Rules~1}, \hyperref[rule:2]{2}, \hyperref[rule:3]{3}\ifFull, such that an optimal tracking set of $K$ is a constant fraction of $K$\fi~(see \cref{cor:const_approx}). \item Compute a relaxed $r$-division \ifFull\else $\mathcal{R}$~\fi of $K$ with $r=(2c_H^1c_H^2(c_H^3+1)/\epsilon)^2$, for any choice of $\epsilon>0$ and constants $c_H^1,c_H^2,c_H^3>0$ specified later. \ifFull Let $\mathcal{R}$ be the set of resulting regions.\fi \item For each region $R$ in $\mathcal{R}$, compute an optimal tracking set $\mathit{OPT}(R)$ for the subset of entry-cycles (with respect to $(s,t)$) which are completely contained in $R$. \item Output $T=\bigcup_{R\in\mathcal{R}}\left(\mathit{OPT}(R)\cup\partial(R)\cup\mathcal{N}(R)\right)$. Here, $\mathcal{N}(R):=N_{\Pi(R)}(\partial(\Pi(R)))$ defines an appropriate neighborhood of the boundary vertices of $R$, where $\Pi(R)$ is the subgraph of $R$ consisting of the union of each path in $R$ that: (i) is not an edge, (ii) has $\partial(R)$ vertices as endpoints, and (iii) traverses no \emph{internal} vertices that are in $\mathit{OPT}(R)$. We let $\partial(\Pi(R))\vcentcolon=\partial(R)\cap \Pi(R)$. See \cref{fig:Pi_R}. \end{list} \end{mdframed} \begin{figure}[hb!] \centering \vspace*{-12pt} \includegraphics[page=6,scale=1.2]{figures/figures.pdf} \caption{Illustration of $\Pi(R)$ and of $\mathcal{N}(R)$ for a region $R$ in a relaxed $r$-division $\mathcal{R}$. Vertices in $\partial(R)$ are depicted in red circles. $\Pi(R)$ consists of the union of all boundary-to-boundary paths in $R$ (solid black), which are not edges and do not traverse $\mathit{OPT}(R)$ (green crosses). The dashed lines represent paths in $R-\Pi(R)$. $\mathcal{N}(R)$ is depicted in blue squares.} \label{fig:Pi_R} \end{figure} \ifFull Notice that $\bigcup_{R\in \mathcal{R}} \left( \mathit{OPT}(R)\cup \partial(R) \right)$ may not constitute a tracking set of $K$, because there might exist entry-exit cycles spanning exactly 2 regions, whose entry-exit pairs are the only vertices with trackers, rendering them untracked (see \cref{fig:rdivision}). \fi We will now give the details of the algorithm and its correctness. We refer to the \hyperref[rule:1]{Reduction Rules} defined in \cref{subsec:reduction_rules}. As a reminder, after exhaustive application of \hyperref[rule:1]{Rules~1} and \hyperref[rule:2]{2}, the graph is either a single edge between $s$ and $t$, or all its vertices have degree at least 2. Henceforth, we will assume the latter, since a minimum tracking set is trivial in the former. \ifFull \hyperref[rule:3]{Rule~3}, which precludes the existence of adjacent vertices of degree 2, is used to bound the overall number of degree-2 vertices. \fi \ifFull \else \fi Notice that none of the reduction rules introduce trackers, so there is no lifting required at the end of our algorithm, i.e., adding back any trackers introduced during the reduction. \begin{observation}\label{obs:tracking_set_K} No entry-exit cycles are removed during \hyperref[rule:1]{Rules~1}, \hyperref[rule:2]{2} or \hyperref[rule:3]{3}, so a tracking set of the resulting kernel $K$ is a tracking set of the input graph $G$. Therefore, any minimum tracking set of $K$ is also a minimum tracking set of $G$. \end{observation} Next, we explain how to compute in polynomial time optimal tracking sets for each region in a relaxed $r$-division of a kernel $K$. \begin{restatable}{lemma}{lemOptEachR} \ifFull\else\hyperref[lemOptEachRLbl]{$\circledast$}\fi \label{lem:opt_each_R} Let $\mathcal{C}(R)$ be the set of all entry-exit cycles in $G$ whose vertices are a subset of $V(R)$, where $R$ is a subgraph of $G$. Then one can compute a minimum subset of $V(R)$ that tracks every entry-cycle of $\mathcal{C}(R)$ in $O(2^{|V(R)|}\cdot n^{O(1)})$ time. \end{restatable} \ifFull \begin{proof} It suffices to enumerate all the $2^{|V(R)|}$ possible subsets and, for each, verify in $O(n^{O(1)})$ time whether every entry-exit cycle of $\mathcal{C}(R)$ is tracked. The verification step can be done in a way similar to the verification algorithm given by Banik \textit{et al.} \cite{DBLP:journals/algorithmica/BanikCLRS20} to show that the problem is in NP: from the observation that every tracking set $X$ is an FVS (see \cref{rem:ts_is_fvs}), it follows that there is at most $O(n^{O(1)})$ entry-exit cycles not tracked by $X$ (see also \cref{lem:general:poly_cycles}). \qed \end{proof} \fi Let us now argue that our algorithm computes a $(1+\epsilon)$-approximate tracking set. Let $T=\bigcup_{R\in \mathcal{R}} \left( \mathit{OPT}(R)\cup \partial(R) \cup \mathcal{N}(R)\right)$ be the output of the algorithm\ifFull, for the relaxed $r$-division $\mathcal{R}$ of a kernel $K$ of $G$\fi. \begin{restatable}{lemma}{lemFeasibility} \ifFull\else\hyperref[lemFeasibilityLbl]{$\circledast$}\fi \label{lem:feasibility:minor_free} $T$ is a tracking set of the input graph $G$. \end{restatable} \ifFull \begin{proof} It is enough to argue that $T$ tracks every cycle $C$ of $K$ (by \cref{lem:tracking_set} and \cref{obs:tracking_set_K}). We consider 3 types of cycles, illustrated in \cref{fig:rdivision}. If $C$ spans exactly 1 region $R$ in $\mathcal{R}$, then it is guaranteed to be tracked by feasibility of $\mathit{OPT}(R)$ (see \cref{lem:opt_each_R}). If $C$ spans at least 3 regions in $\mathcal{R}$, then $C$ is trivially tracked by the trackers in, at least, 3 boundary vertices. Otherwise, let $C$ be a cycle spanning exactly 2 regions. We argue that $C$ is also trivially tracked. Let $R_1$ and $R_2$ be the two regions spanned by $C$ and let us assume that $C$ traverses exactly 2 boundary vertices $b_1$ and $b_2$ (if it traverses more boundary vertices, it must be trivially tracked, and if it traverses less, then it could not span more than one region). If $C$ contains a tracker in $\mathit{OPT}(R_1)\cup\mathit{OPT}(R_2)\setminus\{b_1,b_2\}$, we are done. Otherwise, $C$ is the union of a path in $\Pi(R_1)$ and a path in $\Pi(R_2)$, by definition of $\Pi$. In this case, however, $C$ must have a third tracker placed in $N(\{b_1,b_2\})\cap \{int(R_1)\cup int(R_2)\}$ (notice that $C$ contains at least one non-boundary vertex that is a neighbor of $b_1$ or $b_2$, since there are no parallel edges). \qed \end{proof} \fi Let us denote by $\mathit{OPT}$ the size of an optimal tracking set of the input graph $G$. To argue that $|T|\le (1+\epsilon)\mathit{OPT}$, we will need to argue that the set of trackers in the special neighborhoods defined by $\mathcal{N}(R)$, for all regions $R$, have small cardinalities, i.e., roughly equal to $O(\epsilon\mathit{OPT})$. This is the key argument to our EPTAS, which the next lemma addresses. Its proof is not immediately obvious, since the number of neighbors of all boundary vertices could be $\Omega(\mathit{OPT})$, a consequence of the quadratic gap between $|\partial(R)|$ and $|V(R)|$. \ifFull Intuitively, if these special neighborhoods were too large, then the following properties could not both be applicable: feasibility of $\mathit{OPT}(R)$ for every region $R$, and classification of $G$ as minor-free. Our proof below is of combinatorial nature, so it will not quite capture the topological intuition behind it, but it is relatively succinct and avoids complex case analysis. We will use \cref{lem:tree-sink} to derive a few structural properties of $\Pi(R)$, for each $R\in\mathcal{R}$, and exploit these to bound $|\mathcal{N}(R)|$. As a reminder, $\Pi(R)$ is defined to be the subgraph of $R$ that is the union of each path in $R$ that: is not an edge, has $\partial(R)$ vertices as endpoints, and traverses no \emph{internal} vertices that are in $\mathit{OPT}(R)$, where $\mathit{OPT}(R)$ is an optimal tracking set of just the entry-exit cycles of region $R$. \fi \begin{restatable}{lemma}{lemSmallNeighborhoods} \ifFull\else\hyperref[lemSmallNeighborhoodsLbl]{$\circledast$}\fi \label{lem:small_neighborhoods} $|\mathcal{N}(R)| \le c_H^3|\partial(\Pi(R))|$, where $c_H^3\ge 9\sigma_H^2+3\sigma_H$. \end{restatable} \ifFull \begin{proof} The proof is similar in spirit to that of \cref{lem:linear_size_minor_free_biconnected}. Clearly, $\partial(\Pi(R))$ is an FVS for $\Pi(R)$ (if it were not, there would be untracked cycles in $R$ contradicting feasibility of $\mathit{OPT}(R)$), so $\Pi(R)-\partial(\Pi(R))$ is a forest. We assume w.l.o.g. that $|\partial(\Pi(R))|\ge 2$. Below, we make some claims about the structure of $\Pi(R)$: \begin{list} {\textit{Claim \arabic{itemcounter}:}} {\usecounter{itemcounter}\leftmargin=1em\labelwidth=4em\labelsep=.5em\itemindent=4em} \item Let $b_1,b_2$ be two vertices in $\partial(\Pi(R))$. There exist at most 3 trees in $\Pi(R)-\partial(\Pi(R))$ that are adjacent\footnote{In this context, a tree is adjacent to $v$ if it includes a vertex that is adjacent to $v$.} to both $b_1$ and $b_2$. \item Every tree in $\Pi(R)-\partial(\Pi(R))$ is adjacent to at least 2 vertices in $\partial(\Pi(R))$. \item Every tree in $\Pi(R)-\partial(\Pi(R))$ contains at most 3 vertices adjacent to the same vertex in $\partial(\Pi(R))$. \end{list} The first claim follows from \cref{lem:tree-sink} (if there existed 4 or more trees adjacent to both $b_1$ and $b_2$, there would have to be a tracker from $\mathit{OPT}(R)$ in one of the trees, a contradiction). The last claim follows from \cref{lem:tree-sink} in a similar fashion. The second claim follows from the definition of $\Pi(R)$ and \hyperref[rule:1]{Rules~1},\hyperref[rule:2]{2}. Let us contract each tree $Tr$ in $\Pi(R)-\partial(\Pi(R))$ into a \emph{tree vertex} $v_{Tr}$ and let $F$ be the set of all tree vertices. Notice that this may create parallel edges between a vertex in $\partial(\Pi(R))$ and a tree vertex, but never between two vertices in $\partial(\Pi(R))$ or $F$. In addition, let us remove any edges between vertices in $\partial(\Pi(R))$. The resulting graph is bipartite, with vertex set partitioned into $\partial(\Pi(R))$ and $F$, and is $H$-minor-free (since the class of minor-free graphs is minor-closed). By Claims 1 and 2, at most 3 vertices in $F$ share the same pair of neighbors, and every vertex in $F$ has degree at least 2. Hence, by \cref{lem:bipartite}, \[|F|\le3\sigma_H|\partial(\Pi(R))|\] As a consequence of Claim 3, there are at most 3 parallel edges between a vertex in $\partial(\Pi(R))$ and a vertex in $F$. Thus, by \cref{thm:mader}, the set of edges in the bipartite graph is at most \[3\cdot \sigma_H(|F|+|\partial(\Pi(R))|)\le (9\sigma_H^2+3\sigma_H)|\partial(\Pi(R))|\] The lemma follows from the fact that the edges in the bipartite graph, including the parallel ones, have a 1-1 correspondence with the vertices in $\mathcal{N}(R)$. \ifFull\qed\fi \end{proof} \else \begin{proof} (Sketch) The set of untracked cycles between 2 regions $R$ and $R'$, which must exist in $\Pi(R) \cup \Pi(R')$, induces a forest on either region if we remove $\partial(R)$ and $\partial(R')$. Using arguments similar to those in the proof of \cref{lem:linear_size_minor_free_biconnected}, we can show that the bipartite graph with bipartition $(F,\partial(\Pi(R)))$ has the properties required by \cref{lem:bipartite}, but also that there exists $O(1)$ edges between a tree and a boundary vertex, where $F$ is the set of trees in $\Pi(R)-\partial(\Pi(R))$. As a consequence, we can get an appropriate bound on the number of edges in this bipartite graph, from which the lemma follows. (See \hyperref[lemSmallNeighborhoodsLbl]{\namecref{app:minor-free}~\labelcref{app:minor-free}} for details.) \qed \end{proof} \fi Before proving that the output of our algorithm is a $(1+\epsilon)$-approximate tracking set, let us first recall a result from Frederickson \cite[Lemma~1]{DBLP:journals/siamcomp/Frederickson87} \ifFull\footnote{A more detailed proof of \cite[Lemma~1]{DBLP:journals/siamcomp/Frederickson87} is given in \cite{planarity.org} for separators based on edge weights.}\fi, which concerns the sum, for each boundary vertex $b$ of the number of regions $\Delta(b)$ containing $b$ in a relaxed $r$-division $\mathcal{R}$ of a planar graph. Even though this result was given in the context of planar graphs, it can easily be generalized to any graph whose subgraphs $G'$ admit balanced separators of size $O(\sqrt{|V(G')|})$. We denote the set of all boundary vertices by $\partial(\mathcal{R})$. Further, let $B(\mathcal{R})=\sum_{b\in \partial(\mathcal{R})}\left(\Delta(b)-1\right)$. \begin{lemma}[\cite{DBLP:journals/siamcomp/Frederickson87}]\label{lem:total_boundaries} Let $\mathcal{R}$ be a relaxed $r$-division of an $n$-vertex graph whose subgraphs $G'$ admit balanced separators of size at most $c\sqrt{|V(G')|}$. Then $B(\mathcal{R})\le c\cdot n/\sqrt{r}$, for a constant $c$ independent of $r$ and $n$. \end{lemma} We will use the latter lemma to bound the overall number of trackers in the next theorem. \ifFull \begin{theorem} There exists a $(1+\epsilon)$-approximation algorithm for \textsc{Tracking}{} on $H$-minor-free graphs (for a fixed $H$), running in $O(2^{O(1/\epsilon^2)}n^{O(1)})$ time, for a given $\epsilon>0$. \end{theorem} \else \begin{theorem} \textsc{Tracking}{} admits an EPTAS for $H$-minor-free graphs. \end{theorem} \fi \begin{proof} Consider the algorithm given at the beginning of the section. As a reminder, let $T=\bigcup_{R\in \mathcal{R}} \left( \mathit{OPT}(R)\cup \partial(R) \cup \mathcal{N}(R)\right)$ be the output of the algorithm, for a relaxed $r$-division $\mathcal{R}$ of a kernel $K$ of $G$, where $\mathit{OPT}(R)$ is the optimal tracking set computed with respect to entry-exit cycles in $R$. By \cref{lem:feasibility:minor_free}, $T$ is a tracking set. Next, we argue about the approximation ratio. By a union bound, \[|T|\le |\partial(\mathcal{R})| + \sum_{R\in\mathcal{R}} |\mathit{OPT}(R)|+\sum_{R\in\mathcal{R}}|\mathcal{N}(R)|.\] Let $n'=|V(K)|$ be the number of vertices in $K$. Clearly, $|\partial(\mathcal{R})|\le B(\mathcal{R})$. Moreover, we have that $\sum_{R\in\mathcal{R}}|\partial(R)|\le 2B(\mathcal{R})$, so by \cref{lem:small_neighborhoods}, we have: \[\sum_{R\in\mathcal{R}}|\mathcal{N}(R)|\le 2c_H^3B(\mathcal{R}).\] Let $T^*$ be an optimal tracking set of $K$, i.e., $|T^*|=\mathit{OPT}$ (by \cref{obs:tracking_set_K}). Since $T^*$ is a tracking set, but not necessarily an optimal one, for all entry-exit cycles within any region $R\in\mathcal{R}$, we have that $|\mathit{OPT}(R)|\le |T^*\cap V(R)|$. Thus, \[\sum_{R\in\mathcal{R}} |\mathit{OPT}(R)|\le \mathit{OPT} + B(\mathcal{R}).\] Overall, for $r= (2c_H^1c_H^2(c_H^3+1)/\epsilon)^2$, \begin{align*} |T| & \le \mathit{OPT} + 2(c_H^3+1)B(\mathcal{R})\\ & \le \mathit{OPT} + 2c_H^1(c_H^3+1) n'/\sqrt{r} && (\text{\cref{lem:total_boundaries}, \cref{thm:minor_free_separator}})\\ & \le \mathit{OPT} + 2c_H^1c_H^2(c_H^3+1)\mathit{OPT} /\sqrt{r} && (\text{\ifFull\cref{lem:linear_size_minor_free}\else\cref{thm:linear_size_minor_free}\fi}, c_H^2\ge 16\sigma_H^2 + 8\sigma_H + 1)\\ & = (1+\epsilon)\mathit{OPT}. \end{align*} \ifFull The first step of the algorithm, concerning the kernelization, takes $O(n^{O(1)})$ time, since it consists of applying \hyperref[rule:1]{Rules~1}, \hyperref[rule:2]{2}, \hyperref[rule:3]{3}. The second step, which computes a relaxed $r$-division can be done in $O(n\lg n)$ time \cite{DBLP:journals/siamcomp/Frederickson87}, using Lipton and Tarjan's linear time algorithm \cite{lipton1979separator} to find a balanced separator. Computing each $\mathit{OPT}(R)$ in step 3 can be done in $O(2^r\cdot n^{O(1)})$ time, by \cref{lem:opt_each_R}, and thus $O(2^r\cdot n^{O(1)}/r)$ time for all $R\in\mathcal{R}$. Finally, computing $\Pi(R)$ and $\mathcal{N}(R)$ can be done in $O(n^{O(1)})$ time, and thus $O(n^{O(1)}/r)$, for all $R\in\mathcal{R}$. Overall, the total time complexity is is bounded by \begin{align*} & O(2^r\cdot n^{O(1)})\\ =\ & O(2^{O(1/\epsilon^2)}\cdot n^{O(1)}). \end{align*} \else Step 1 of the algorithm takes $O(n^{O(1)})$ time, since it consists of applying \hyperref[rule:1]{Rules~1}, \hyperref[rule:2]{2}, \hyperref[rule:3]{3}. Step 2 can be done in $O(n\lg n)$ time \cite{DBLP:journals/siamcomp/Frederickson87}. Step 3 takes $O(2^r\cdot n^{O(1)})$ time, by \cref{lem:opt_each_R}. Finally, step 4 takes $O(n^{O(1)})$ time. Overall, these amount to $O(2^{O(1/\epsilon^2)}n^{O(1)})$. \qed \end{proof} \section{Introduction}\label{sec:intro} \input{intro} \ifFull\else \Paragraph{Preliminaries.} We use standard terminology concerning graphs, approximation algorithms and kernelization, which is detailed in \cref{app:terminology}. For space considerations, content marked with a link symbol ``\hyperref[app:properties]{$\circledast$}'' is provided in more detail and/or proved in an appendix. \fi \ifFull \section{Preliminaries}\label{sec:prelim} \input{terminology} \fi \section{Structural Properties}\label{sec:properties} \input{properties} \section{$H$-Minor-Free Graphs}\label{sec:minor_free} \input{minor_free} \section{General Graphs}\label{sec:general} \input{general_graphs} \ifFull \section{Conclusion} TODO \fi \ifFull \bibliographystyle{splncs04} \else \bibliographystyle{abbrvX} \fi

https://arxiv.org/abs/1705.02326

Value Iteration for Long-run Average Reward in Markov Decision Processes

Markov decision processes (MDPs) are standard models for probabilistic systems with non-deterministic behaviours. Long-run average rewards provide a mathematically elegant formalism for expressing long term performance. Value iteration (VI) is one of the simplest and most efficient algorithmic approaches to MDPs with other properties, such as reachability objectives. Unfortunately, a naive extension of VI does not work for MDPs with long-run average rewards, as there is no known stopping criterion. In this work our contributions are threefold. (1) We refute a conjecture related to stopping criteria for MDPs with long-run average rewards. (2) We present two practical algorithms for MDPs with long-run average rewards based on VI. First, we show that a combination of applying VI locally for each maximal end-component (MEC) and VI for reachability objectives can provide approximation guarantees. Second, extending the above approach with a simulation-guided on-demand variant of VI, we present an anytime algorithm that is able to deal with very large models. (3) Finally, we present experimental results showing that our methods significantly outperform the standard approaches on several benchmarks.

\section{More experimental data} \label{section:exper_data_appendix} \begin{table}[h] \centering \caption{Additional data for Table~\ref{tbl:experiment_mg}. The results for \texttt{MEC-BRTDP} and \texttt{ODV} using the different heuristics are listed in the order \texttt{PR}, \texttt{RR}, \texttt{MD}. Note that for strongly connected (scon) models, all variants of \texttt{MEC-BRTDP} take the same time, since the reachability solver is not called at all.} \begin{tabu} to 0.8\linewidth {lrrCC} \textbf{Model} & \textbf{States} & \textbf{MECs} & \texttt{MEC-BRTDP} & \texttt{ODV} \\ \toprule virus & 809 & 1 & 4.69/3.96/4.40 & TO/2452/TO \\ cs\_nfail4 & 960 & 176 & 20.3/8.83/9.39 & 51.3/15.4/16.0 \\ investor & 6688 & 837 & 47.1/48.6/64.5 & 17.3/21.3/18.7 \\ phil-nofair5 & 93068 & scon & 70 & TO/TO/TO \\ rabin4 & 668836 & scon & 820 & TO/TO/TO \\ \bottomrule \end{tabu} \label{tbl:experiment_mg_app} \end{table} \begin{table}[h] \centering \caption{Additional data for Table~\ref{tbl:experiment_big}. We list the results of using \texttt{ODV} with the three different mentioned exploration heuristics. Heuristics are listed in the order \texttt{PR}, \texttt{RR}, \texttt{MD}.} \begin{tabu} to \linewidth{lrCCC} \textbf{Model} & \textbf{States} & Time & Explored States & Explored MECs \\ \toprule zeroconf(40,10) & 3001911 & 83.4/9.72/5.05 & 516/1643/481 & 3/3/3 \\ ~~\emph{avoid} & & & 905/2475/582 & 3/3/3 \\ zeroconf(300,15) & 4730203 & 1831/294/16.6 & 731/2291/873 & 3/3/3 \\ ~~\emph{avoid} & & & 1595/10759/5434 & 3/22/3 \\ sensors(2) & 7860 & 31.8/26.1/20.1 & 5308/4915/3281 & 1474/1444/917 \\ sensors(3) & 77766 & 305/480/37 & 31431/34239/10941 & 9125/11041/2301 \\ \bottomrule \end{tabu} \label{tbl:experiment_big_app} \end{table} \section{Implementation and experimental results} \label{sec:exper} In this section, we compare the runtime of our presented approaches to established tools. All benchmarks have been run on a 4.4.3-gentoo x64 virtual machine with 3.0 GHz per core, a time limit of one hour and memory limit of 8GB. The precision requirement for all approximative methods is $\varepsilon = 10^{-6}$. We implemented our constructions as a package in the PRISM Model Checker \cite{KNP11}. We used the 64-bit Oracle JDK version \texttt{1.8.0\_102-b14} as Java runtime for all executions. All measurements are given in seconds, measuring the total user CPU time of the PRISM process using the UNIX tool \texttt{time}. \subsection{Models} First, we briefly explain the examples used for evaluation. \textbf{virus} \cite{kwiatkowska2009probabilistic} models a virus spreading through a network. We reward each attack carried out by an infected machine. Note that in this model, no machine can \enquote{purge} the virus, hence eventually all machines will be infected. \textbf{cs\_nfail} \cite{komuravelli2012assume} models a client-server mutual exclusion protocol with probabilistic failures of the clients. A~reward is given for each successfully handled connection. \textbf{investor} \cite{MM07,MM02} models an investor operating in a stock market. The investor can decide to sell his stocks and keep their value as a reward or hold them and wait to see how the market evolves. The rewards correspond to the value of the stocks when the investor decides to sell them, so maximizing the average reward corresponds to maximizing the expected selling value of the stocks. \textbf{phil\_nofair} \cite{DFP04} represents the (randomised) dining philosophers without fairness assumptions. We use two reward structures, one where a reward is granted each time a philosopher \enquote{thinks} or \enquote{eats}, respectively. \textbf{rabin} \cite{rabin1982n} is a well-known mutual exclusion protocol, where multiple processes repeatedly try to access a shared critical section. Each time a process successfully enters the critical section, a reward is given. \textbf{zeroconf} \cite{KNPS06} is a network protocol designed to assign IP addresses to clients without the need of a central server while still avoiding address conflicts. We explain the reward assignment in the corresponding result section. \textbf{sensor} \cite{komuravelli2012assume} models a network of sensors sending values to a central processor over a lossy connection. A reward is granted for every \emph{work} transition. \subsection{Tools} We will compare several different variants of our implementations, which are described in the following. \begin{itemize} \item Naive value iteration (\texttt{NVI}) runs the value iteration on the whole MDP as in Algorithm~\ref{alg:vi} of Section~\ref{sec:nvi} together with the stopping criterion \eqref{eq:sc_con1} conjectured by \cite[Sect.~9.4.2]{Puterman}. As the stopping criterion is incorrect, we will not only include the runtime until the stopping criterion is fulfilled, but also until the computed value is $\varepsilon$-close to the known solution. \item Our MEC decomposition approach presented in Algorithm~\ref{alg:lvi} of Section~\ref{sec:lvi} is denoted by \texttt{MEC-\textit{reach}}, where $\texttt{\textit{reach}}$ identifies one of the following reachability solver used on the quotient MDP. \begin{itemize} \item PRISM's value iteration (\texttt{VI}), which iterates until none of the values change by more than $10^{-8}$. While this method is theoretically imprecise, we did not observe this behaviour in our examples.\footnote{PRISM contains several other methods to solve reachability, which all are imprecise and behaved comparably in our tests.} \item An exact reachability solver based on linear programming (\texttt{LP})~\cite{Giro14}. \item The BRTDP solver with guaranteed precision of~\cite{atva} (\texttt{BRTDP}). This solver is highly configurable. Among others, one can specify the heuristic which is used to resolve probabilistic transitions in the simulation. This can happen according to transition probability (\texttt{PR}), round-robin (\texttt{RR}) or maximal difference (\texttt{MD}). Due to space constraints, we only compare to the \texttt{MD} exploration heuristic here. Results on the other heuristics can be found in \cite[Appendix~E]{techrep} \end{itemize} \item \texttt{ODV} is the implementation of the on-demand value iteration as in Algorithm~\ref{alg:skeleton} of Section~\ref{sec:odv}. Analogously to the above, we only provide results on the \texttt{MD} heuristic here. The results on \texttt{ODV} together with the other heuristics can also be found in \cite[Appendix~E]{techrep}. \end{itemize} Furthermore, we will compare our methods to the state-of-the-art tool MultiGain, version 1.0.2~\cite{DBLP:conf/tacas/BrazdilCFK15} abbreviated by \texttt{MG}. MultiGain uses linear programming to exactly solve mean payoff objectives among others. We use the commercial LP solver Gurobi 7.0.1 as backend\footnote{MultiGain also supports usage of the LP solver \texttt{lp\_solve 5.5} bundled with PRISM, which consistently performed worse than the Gurobi backend.}. We also instantiated \texttt{\textit{reach}} by an implementation of the interval iteration algorithm presented in~\cite{haddad2014reachability}. This variant performed comparable to \texttt{MEC-VI} and therefore we omitted it. \subsection{Results} \begin{table}[t] \centering \caption{Runtime comparison of our approaches to MultiGain on various, reasonably sized models. Timeouts (1h) are denoted by TO. Strongly connected models are denoted by \enquote{scon} in the MEC column. The best result in each row is marked in bold, excluding NVI due to its imprecisions. For NVI, we list both the time until the stopping criterion is satisfied and until the values actually converged.} \setlength{\tabcolsep}{-1pt} \begin{tabu} to \linewidth {lRRCCCCcC} \textbf{Model} & \textbf{States} & \textbf{MECs} & \texttt{MG} & \texttt{NVI} & \texttt{MEC-VI} & \texttt{MEC-LP} & \texttt{MEC-BRTDP} & \texttt{ODV} \\ \toprule virus & 809 & 1 & \textbf{3.76} & 3.50/3.71 & 4.09 & 4.41 & 4.40 & TO \\ cs\_nfail4 & 960 & 176 & 4.86 & 10.2/TO & \textbf{4.38} & TO & 9.39 & 16.0 \\ investor & 6688 & 837 & 16.75 & 4.23/TO & \textbf{8.83} & TO & 64.5 & 18.7 \\ phil-nofair5 & 93068 & scon & TO & 23.5/30.3 & \textbf{70} & \textbf{70} & \textbf{70} & TO \\ rabin4 & 668836 & scon & TO & 87.8/164 & \textbf{820} & \textbf{820} & \textbf{820} & TO \\ \bottomrule \end{tabu} \label{tbl:experiment_mg} \end{table} The experiments outlined in Table~\ref{tbl:experiment_mg} show that our methods outperform MultiGain significantly on most of the tested models. Furthermore, we want to highlight the \textbf{investor} model to demonstrate the advantage of \texttt{MEC-VI} over \texttt{MEC-LP}. With higher number of MECs in the initial MDP, which is linked to the size of the reachability LP, the runtime of \texttt{MEC-LP} tends to increase drastically, while \texttt{MEC-VI} performs quite well. Additionally, we see that \texttt{NVI} fails to obtain correct results on any of these examples. \texttt{ODV} does not perform too well in these tests, which is primarily due to the significant overhead incurred by building the partial model dynamically. This is especially noticeable for strongly connected models like \textbf{phil-nofair} and \textbf{rabin}. For these models, every state has to be explored and \texttt{ODV} does a lot of superfluous computations until the model has been explored fully. On \textbf{virus}, the bad performance is due to the special topology of the model, which obstructs the back-propagation of values. Moreover, on the two strongly connected models all MEC decomposition based methods perform worse than naive value iteration as they have to obtain the MEC decomposition first. Furthermore, all three of those methods need the same amount of for these models, as the weighted MEC quotient only has a single state (and the two special states), thus the reachability query is trivial. \begin{table}[t] \centering \caption{Runtime comparison of our on-demand VI method with the previous approaches. All of those behaved comparable to \texttt{MEC-VI} or worse, and due to space constraints we omit them. MO denotes a memory-out. Aside from runtime, we furthermore list the number of explored states and MECs of \texttt{ODV}} \begin{tabu} to \linewidth {lRCCCC} \textbf{Model} & \textbf{States} & \texttt{MEC-VI} & \texttt{ODV} & \texttt{ODV} States & \texttt{ODV} MECs \\ \toprule zeroconf(40,10) & 3001911 & MO & 5.05 & 481 & 3 \\ \quad\emph{avoid} & & & & 582 & 3 \\ zeroconf(300,15) & 4730203 & MO & 16.6 & 873 & 3 \\ \quad\emph{avoid} & & & & 5434 & 3 \\ sensors(2) & 7860 & 18.9 & 20.1 & 3281 & 917 \\ sensors(3) & 77766 & 2293 & 37.2 & 10941 & 2301 \\ \bottomrule \end{tabu} \label{tbl:experiment_big} \end{table} In Table~\ref{tbl:experiment_big} we present results of some of our methods on \textbf{zeroconf} and \textbf{sensors}, which both have a structure better suited towards \texttt{ODV}. The \textbf{zeroconf} model consists of a big transient part and a lot of \enquote{final} states, i.e. states which only have a single self-loop. \textbf{sensors} contains a lot of small, often unlikely-to-be-reached MECs. On the \textbf{zeroconf} model, we evaluate the average reward problem with two reward structures. In the default case, we assign a reward of 1 to every final state and zero elsewhere. This effectively is solving the reachability question and thus it is not surprising that our method gives similarly good results as the BRTDP solver of~\cite{atva}. The \emph{avoid} evaluation has the reward values flipped, i.e. all states except the final ones yield a payoff of 1. With this reward assignment, the algorithm performed slightly slower, but still extremely fast given the size of the model. We also tried assigning pseudo-random rewards to every non-final state, which did not influence the speed of convergence noticeably. We want to highlight that the mem-out of \texttt{MEC-VI} already occurred during the MEC-decomposition phase. Hence, no variant of our decomposition approach can solve this problem. Interestingly, the naive value iteration actually converges on \textbf{zeroconf}(40,10) in roughly 20 minutes. Unfortunately, as in the previous experiments, the used incorrect stopping criterion was met a long time before that. Further, when comparing \textbf{sensors}(2) to \textbf{sensors}(3), the runtime of \texttt{ODV} only doubled, while the number of states in the model increased by an order of magnitude and the runtime of \texttt{MEC-VI} even increased by two orders of magnitude. These results show that for some models, \texttt{ODV} is able to obtain an $\varepsilon$-optimal estimate of the mean payoff while only exploring a tiny fraction of the state space. This allows us to solve many problems which previously were intractable simply due to an enormous state space. \section{Introduction} The analysis of probabilistic systems arises in diverse application contexts of computer science, e.g. analysis of randomized communication and security protocols, stochastic distributed systems, biological systems, and robot planning, to name a few. The standard model for the analysis of probabilistic systems that exhibit both probabilistic and non-deterministic behaviour are \emph{Markov decision processes (MDPs)}~\cite{Howard,FV97,Puterman}. An MDP consists of a finite set of states, a finite set of actions, representing the non-deterministic choices, and a transition function that given a state and an action gives the probability distribution over the successor states. In verification, MDPs are used as models for e.g.\ concurrent probabilistic systems~\cite{CY95} or probabilistic systems operating in open environments~\cite{SegalaT}, and are applied in a wide range of applications~\cite{BaierBook,KNP11}. \paragraph{Long-run average reward} A \emph{payoff} function in an MDP maps every infinite path (infinite sequence of state-action pairs) to a real value. One of the most well-studied and mathematically elegant payoff functions is the \emph{long-run average reward} (also known as \emph{mean-payoff} or \emph{limit-average reward}, \emph{steady-state reward} or simply \emph{average reward}), where every state-action pair is assigned a real-valued reward, and the payoff of an infinite path is the long-run average of the rewards on the path~\cite{FV97,Puterman}. Beyond the elegance, the long-run average reward is standard to model performance properties, such as the average delay between requests and corresponding grants, average rate of a particular event, etc. Therefore, determining the maximal or minimal expected long-run average reward of an MDP is a basic and fundamental problem in the quantitative analysis of probabilistic systems. \paragraph{Classical algorithms} A \emph{strategy} (also known as \emph{policy} or \emph{scheduler}) in an MDP specifies how the non-deterministic choices of actions are resolved in every state. The \emph{value} at a state is the maximal expected payoff that can be guaranteed among all strategies. The values of states in MDPs with payoff defined as the long-run average reward can be computed in polynomial-time using linear programming~\cite{FV97,Puterman}. The corresponding linear program is quite involved though. The number of variables is proportional to the number of state-action pairs and the overall size of the program is linear in the number of transitions (hence potentially quadratic in the number of actions). While the linear programming approach gives a polynomial-time solution, it is quite slow in practice and does not scale to larger MDPs. Besides linear programming, other techniques are considered for MDPs, such as dynamic-programming through strategy iteration or value iteration~\cite[Chap.~9]{Puterman}. \paragraph{Value iteration} A generic approach that works very well in practice for MDPs with other payoff functions is \emph{value iteration (VI)}. Intuitively, a particular one-step operator is applied iteratively and the crux is to show that this iterative computation converges to the correct solution (i.e.\ the value). The key advantages of VI are the following: \begin{enumerate} \item \emph{Simplicity.} VI provides a very simple and intuitive dynamic-programming algorithm which is easy to adapt and extend. % \item \emph{Efficiency.} For several other payoff functions, such as finite-horizon rewards (instantaneous or cumulative reward) or reachability objectives, applying the concept of VI yields a very efficient solution method. In fact, in most well-known tools such as PRISM~\cite{KNP11}, value iteration performs much better than linear programming methods for reachability objectives. % \item \emph{Scalability.} The simplicity and flexibility of VI allows for several improvements and adaptations of the idea, further increasing its performance and enabling quick processing of very large MDPs. For example, when considering reachability objectives, \cite{DBLP:conf/ijcai/PineauGT03} present point-based value-iteration (PBVI), applying the iteration operator only to a part of the state space, and~\cite{DBLP:conf/icml/McMahanLG05} introduce bounded real-time dynamic programming (BRTDP), where again only a fraction of the state space is explored based on partial strategies. Both of these approaches are simulation-guided, where simulations are used to decide how to explore the state space. The difference is that the former follows an offline computation, while the latter is online. Both scale well to large MDPs and use VI as the basic idea to build upon. \end{enumerate} \paragraph{Value iteration for long-run average reward} While VI is standard for reachability objectives or finite-horizon rewards, it does not work for general MDPs with long-run average reward. The two key problems pointed out in~\cite[Sect.~8.5,~9.4]{Puterman} are as follows: (a)~if the MDP has some periodicity property, then VI does not converge; and (b)~for general MDPs there are neither bounds on the speed of convergence nor stopping criteria to determine when the iteration can be stopped to guarantee approximation of the value. The first problem can be handled by adding self-loop transitions~\cite[Sect.~8.5.4]{Puterman}. However, the second problem is conceptually more challenging, and a solution is conjectured in~\cite[Sect.~9.4.2]{Puterman}. \paragraph{Our contribution} In this work, our contributions are related to value iteration for MDPs with long-run average reward, they range from conceptual clarification to practical algorithms and experimental results. The details of our contributions are as follows. \begin{itemize} \item \emph{Conceptual clarification.} We first present an example to refute the conjecture of~\cite[Sect.~9.4.2]{Puterman}, showing that the approach proposed there does not suffice for VI on MDPs with long-run average reward. % \item \emph{Practical approaches.} We develop, in two steps, practical algorithms instantiating VI for approximating values in MDPs with long-run average reward. Our algorithms take advantage of the notion of maximal end-components (MECs) in MDPs. Intuitively, MECs for MDPs are conceptually similar to strongly connected components (SCCs) for graphs and recurrent classes for Markov chains. We exploit these MECs to arrive at our two methods: \begin{enumerate} \item The first variant applies VI locally to each MEC in order to obtain an approximation of the values within the MEC. After the approximation in every MEC, we apply VI to solve a reachability problem in a modified MDP with collapsed MECs. We show that this simple combination of VI approaches ensures guarantees on the approximation of the value. % \item We then build on the approach above to present a simulation-guided variant of VI. In this case, the approximation of values for each MEC and the reachability objectives are done at the same time using VI. For the reachability objective a BRDTP-style VI (similar to~\cite{atva}) is applied, and within MECs VI is applied on-demand (i.e. only when there is a requirement for more precise value bounds). The resulting algorithm furthermore is an \emph{anytime} algorithm, i.e.\ it can be stopped at any time and give an upper and lower bounds on the result. \end{enumerate} % \item \emph{Experimental results.} We compare our new algorithms to the state-of-the-art tool MultiGain~\cite{DBLP:conf/tacas/BrazdilCFK15} on various models. The experiments show that MultiGain is vastly outperformed by our methods on nearly every model. Furthermore, we compare several variants of our methods and investigate the different domains of applicability. \end{itemize} In summary, we present the first instantiation of VI for general MDPs with long-run average reward. Moreover, we extend it with a simulation-based approach to obtain an efficient algorithm for large MDPs. Finally, we present experimental results demonstrating that these methods provide significant improvements over existing ones. \paragraph{Further related work.} There is a number of techniques to compute or approximate the long-run average reward in MDPs~\cite{Puterman,Howard,veinott1966}, ranging from linear programming to value iteration to strategy iteration. Symbolic and explicit techniques based on strategy iteration are combined in~\cite{DBLP:conf/qest/WimmerBBHCHDT10}. Further, the more general problem of MDPs with multiple long-run average rewards was first considered in~\cite{Cha07}, a complete picture was presented in~\cite{krish,DBLP:conf/lics/ChatterjeeKK15} and partially implemented in~\cite{DBLP:conf/tacas/BrazdilCFK15}. The extension of our approach to multiple long-run average rewards, or combination of expectation and variance~\cite{BCFK13}, are interesting directions for future work. Finally, VI for MDPs with guarantees for reachability objectives was considered in~\cite{atva,haddad2014reachability}. Proofs and supplementary material can be found in \cite{techrep}. \section{0pt}{12pt plus 4pt minus 4pt}{12pt plus 2pt minus 2pt} \setlength{\textfloatsep}{0.8\textfloatsep} \setlength{\intextsep}{0.8\intextsep} \title{Value Iteration for Long-run Average Reward in Markov Decision Processes} \author{Pranav Ashok\inst1 \and Krishnendu Chatterjee\inst2 \and Przemys{\l}aw Daca\inst2 \and Jan K{\v{r}}et\'insk\'y\inst1 \and Tobias Meggendorfer\inst1} % \institute{Technical University of Munich, Germany \and IST Austria} % \begin{document} \maketitle \begin{abstract} Markov decision processes (MDPs) are standard models for probabilistic systems with non-deterministic behaviours. Long-run average rewards provide a mathematically elegant formalism for expressing long term performance. Value iteration (VI) is one of the simplest and most efficient algorithmic approaches to MDPs with other properties, such as reachability objectives. Unfortunately, a naive extension of VI does not work for MDPs with long-run average rewards, as there is no known stopping criterion. In this work our contributions are threefold. (1)~We refute a conjecture related to stopping criteria for MDPs with long-run average rewards. (2)~We present two practical algorithms for MDPs with long-run average rewards based on VI. First, we show that a combination of applying VI locally for each maximal end-component (MEC) and VI for reachability objectives can provide approximation guarantees. Second, extending the above approach with a simulation-guided on-demand variant of VI, we present an anytime algorithm that is able to deal with very large models. (3)~Finally, we present experimental results showing that our methods significantly outperform the standard approaches on several benchmarks. \end{abstract} \input{intro} \input{prelim} \input{vi} \input{exper} \section{Conclusion} We have discussed the use of value iteration for computing long-run average rewards in general MDPs. We have shown that the conjectured stopping criterion from literature is not valid, designed two modified versions of the algorithm and have shown guarantees on their results. The first one relies on decomposition into VI for long-run average on separate MECs and VI for reachability on the resulting quotient, achieving global error bounds from the two local stopping criteria. The second one additionally is simulation-guided in the BRTDP style, and is an anytime algorithm with a stopping criterion. The benchmarks show that depending on the topology, one or the other may be more efficient, and both outperform the existing linear programming on all larger models. For future work, we pose the question of how to automatically fine-tune the parameters of the algorithms to get the best performance. For instance, the precision increase in each further call of VI on a MEC could be driven by the current values of VI on the quotient, instead of just halving them. This may reduce the number of unnecessary updates while still achieving an increase in precision useful for the global result. \bibliographystyle{alpha} \section{Preliminaries} \label{sec:prelim} \subsection{Markov decision processes} A \emph{probability distribution} on a finite set $X$ is a mapping $\rho: X\mapsto [0,1]$, such that $\sum_{x\in X} \rho(x) = 1$. We denote by $\mathcal{D}(X)$ the set of all probability distributions on $X$. Further, the \emph{support} of a probability distribution $\rho$ is denoted by $\supp(\rho) = \{x \in X \mid \rho(x)> 0 \}$. \begin{definition}[MDP] A \emph{Markov decision processes (MDP)} is a tuple of the form $\mathcal{M} = (S, s_\textrm{init}, \actions, \mathsf{Av}, \Delta, r)$, where $S$ is a finite set of \emph{states}, $s_\textrm{init} \in S$ is the \emph{initial} state, $\actions$ is a finite set of \emph{actions}, $\mathsf{Av}: S \to 2^{\actions}$ assigns to every state a set of \emph{available} actions, $\Delta: S \times \actions \to \distributions(S)$ is a \emph{transition function} that given a state $s$ and an action $a\in \mathsf{Av}(s)$ yields a probability distribution over successor states, and $r : S \times \actions \to \mathbb{R}^{\geq 0}$ is a \emph{reward function}, assigning rewards to state-action pairs. \end{definition} For ease of notation, we write $\Delta(s, a, s')$ instead of $\Delta(s, a)(s')$. An \emph{infinite path} $\rho$ in an MDP is an infinite word $\rho = s_0 a_0 s_1 a_1 \dots \in (S \times \actions)^\omega$, such that for every $i \in \mathbb{N}$, $a_i\in \mathsf{Av}(s_i)$ and $\Delta(s_i,a_i, s_{i+1}) > 0$. A \emph{finite path} $w = s_0 a_0 s_1 a_1 \dots s_n \in (S \times \actions)^* \times S$ is a finite prefix of an infinite path. A \emph{strategy} on an MDP is a function $\pi: (S \times \actions)^*\times S \to \distributions(\actions)$, which given a finite path $w = s_0 a_0 s_1 a_1 \dots s_n$ yields a probability distribution $\pi(w) \in \distributions(\mathsf{Av}(s_n))$ on the actions to be taken next. We call a strategy \emph{memoryless randomized} (or \emph{stationary}) if it is of the form $\pi: S \to \distributions(\actions)$, and \emph{memoryless deterministic} (or \emph{positional}) if it is of the form $\pi: S \to \actions$. We denote the set of all strategies of an MDP by $\Pi$, and the set of all memoryless deterministic strategies by $\Pi^{\mathsf{MD}}$. Fixing a strategy $\pi$ and an initial state $s$ on an MDP $\mathcal{M}$ gives a unique probability measure $\mathbb P^\pi_{\mathcal{M}, s}$ over infinite paths~\cite[Sect.~2.1.6]{Puterman}. The expected value of a random variable $F$ is defined as $\mathbb E^\pi_{\mathcal{M}, s}[F] = \int F\ d\,\mathbb P^\pi_{\mathcal{M}, s}$. When the MDP is clear from the context, we drop the corresponding subscript and write $\mathbb P^\pi_s$ and $\mathbb E^\pi_s$ instead of $\mathbb P^\pi_{\mathcal{M}, s}$ and $\mathbb E^\pi_{\mathcal{M}, s}$, respectively. \paragraph{End components} A pair $(T, A)$, where $\emptyset \neq T \subseteq S$ and $\emptyset \neq A \subseteq \UnionSym_{s\in T} \mathsf{Av}(s)$, is an \emph{end component} of an MDP $\mathcal{M}$ if (i)~for all $s \in T, a \in A \intersectionBin \mathsf{Av}(s)$ we have $\supp(\Delta(s,a)) \subseteq T$, and (ii)~for all $s, s' \in T$ there is a finite path $w = s a_0 \dots a_n s' \in (T \times A)^* \times T$, i.e.\ $w$ starts in $s$, ends in $s'$, stays inside $T$ and only uses actions in $A$.\footnote{This standard definition assumes that actions are unique for each state, i.e.\ $\mathsf{Av}(s) \intersectionBin \mathsf{Av}(s') = \emptyset$ for $s \neq s'$. The usual procedure of achieving this in general is to replace $\actions$ by $S \times \actions$ and adapting $\mathsf{Av}$, $\Delta$, and $r$ appropriately.} Intuitively, an end component describes a set of states for which a particular strategy exists such that all possible paths remain inside these states and all of those states are visited infinitely often almost surely. An end component $(T, A)$ is a \emph{maximal end component (MEC)} if there is no other end component $(T', A')$ such that $T \subseteq T'$ and $A \subseteq A'$. Given an MDP $\mathcal{M}$, the set of its MECs is denoted by $\mathsf{MEC}(\mathcal{M})$. With these definitions, every state of an MDP belongs to at most one MEC and each MDP has at least one MEC. Using the concept of MECs, we recall the standard notion of a \emph{MEC quotient}~\cite{DeAlfaro1997}. To obtain this quotient, all MECs are merged into a single representative state, while transitions between MECs are preserved. Intuitively, this abstracts the MDP to its essential infinite time behaviour. \begin{definition}[MEC quotient~\cite{DeAlfaro1997}] \label{def:mq} Let $\mathcal{M} = (S, s_\textrm{init}, \actions, \mathsf{Av}, \Delta, r)$ be an MDP with MECs $\mathsf{MEC}(\mathcal{M}) = \{(T_1, A_1), \dots, (T_n, A_n)\}$. Further, define $\mathsf{MEC}_S = \UnionSym_{i=1}^n T_i$ as the set of all states contained in some MEC. The \emph{MEC quotient of} $\mathcal{M}$ is defined as the MDP $\widehat\mathcal{M} = (\widehatS, \widehat{s}_\textrm{init}, \widehat\actions, \widehat\mathsf{Av}, \widehat\Delta, \widehatr)$, where: \begin{itemize} \item $\widehatS = S \setminus \mathsf{MEC}_S \unionBin \set{\widehat{s}_1, \dots, \widehat{s}_n}$, \item if for some $T_i$ we have $s_\textrm{init}\in T_i$, then $\widehat{s}_\textrm{init} = \widehat{s}_i$, otherwise $\widehat{s}_\textrm{init} = s_\textrm{init}$, \item $\widehat\actions = \set{(s,a) \mid s\in S, a\in \mathsf{Av}(s)}$, \item the available actions $\widehat\mathsf{Av}$ are defined as \begin{align*} \forall s \in S \setminus \mathsf{MEC}_S.~& \widehat\mathsf{Av}(s) = \set{(s,a) \mid a \in \mathsf{Av}(s)} \\ \forall 1 \leq i \leq n.~& \widehat\mathsf{Av}(\widehat{s}_i) = \set{(s,a) \mid s \in T_i \land a \in \mathsf{Av}(s) \setminus A_i}, \end{align*} \item the transition function $\widehat\Delta$ is defined as follows. Let $\widehat{s} \in \widehat{S}$ be some state in the quotient and $(s, a) \in \mathsf{Av}(\widehat{s})$ an action available in $\widehat{s}$. Then \begin{equation*} \widehat{\Delta}(\widehat{s}, (s, a), \widehat{s}') = \begin{dcases*} {\sum}_{s' \in T_j} \Delta(s, a, s') & if $\widehat{s}' = \widehat{s}_j$, \\ \Delta(s, a, \widehat{s}') & otherwise, i.e.\ $\widehat{s}' \in S \setminus \mathsf{MEC}_S$. \end{dcases*} \end{equation*} For the sake of readability, we omit the added self-loop transitions of the form $\Delta(\widehat{s}_i, (s, a), \widehat{s}_i)$ with $s \in T_i$ and $a \in A_i$ from all figures. \item Finally, for $\widehat{s} \in \widehat{S}$, $(s, a) \in \widehat{\mathsf{Av}}(\widehat{s})$, we define $\widehatr(s, (s, a)) = r(s, a)$. \end{itemize} Furthermore, we refer to $\widehat{s}_1, \dots, \widehat{s}_n$ as \emph{collapsed states} and identify them with the corresponding MECs. \end{definition} \begin{figure}[t] \centering \subfloat[An MDP with three MECs.]{\label{fig:example} \scalebox{0.8}{ \begin{tikzpicture}[auto] \node[state] (s1) at (-3,0.5) {$s_1$}; \node[state] (s6) at (-0.5,-2.5) {$s_6$}; \node[state] (s5) at (-0.5,-1) {$s_5$}; \node[state] (s3) at (-3.1,-1) {$s_3$}; \node[state] (s4) at (-3.1,-2.5) {$s_4$}; \node[state] (s2) at (-5.5,-0.5) {$s_2$}; \node[] at (-2.3,0.3) {$a, \rewformat{0}$}; \node[] at ($(s4) + (0.8,0.5)$) {$a,\rewformat{0}$}; \node[] at ($(s6) + (0.8,0.5)$) {$a, \rewformat{0}$}; \node[] at ($(s3) + (-0.9,-0.7)$) {$a, \rewformat{10}$}; \node[] at ($(s5) + (-0.9,-0.7)$) {$a, \rewformat{20}$}; \rho[->] (s1) edge node[right]{$0.001$} (s5) (s1) edge[out=-30,in=45] node[left]{$0.999$} (s3) (s1) edge node[above] {$b, \rewformat{0}$} (s2) (s5) edge[out=225,in=135] node[below right]{0.5} (s6) (s5) edge[out=225,in=155,looseness=8] coordinate (e5loop) node[above left]{$0.5$} (s5) (s6) edge[out=45,in=-35,looseness=4] coordinate (e6loop) node[below right] {$0.5$} (s6) (s6) edge[out=45,in=-45] node[above right] {$0.5$} (s5) (s3) edge[out=225,in=135] node[below right]{0.5} (s4) (s3) edge[out=225,in=155,looseness=8] coordinate (e3loop) node[above left]{$0.5$} (s3) (s4) edge[out=45,in=-35,looseness=4] coordinate (e4loop) node[below right] {$0.5$} (s4) (s4) edge[out=45,in=-45] node[above right] {$0.5$} (s3) (s2) edge[loop left] coordinate (e2loop) node[below, inner sep=7pt]{$a, \rewformat{4}$} (s2) (s2) edge[out=-80,in=180,bend right] node[below left] {$b, \rewformat{5}$} (s4); \node[rectangle,rounded corners=3pt,draw=none,fill=black,fill opacity=0.1,fit=(s2) (e2loop)] (rectA) {}; \node[rectangle,rounded corners=3pt,draw=none,fill=black,fill opacity=0.1,fit=(s3) (s4) (e3loop) (e4loop)] (rectB) {}; \node[rectangle,rounded corners=3pt,draw=none,fill=black,fill opacity=0.1,fit=(s5) (s6) (e5loop) (e6loop)] (rectC) {}; \node[] at ($(rectA) + (0,-0.8)$) {$\widehat{A}$}; \node[] at ($(rectB) + (0,-1.7)$) {$\widehat{B}$}; \node[] at ($(rectC) + (0,-1.7)$) {$\widehat{C}$}; \end{tikzpicture} } } \quad \subfloat[The MEC quotient.]{\label{fig:quo} \scalebox{0.8}{ \begin{tikzpicture}[auto, align=center] \node[state] (s0) at (-3,0.5) {$s_1$}; \node[state] (s2) at (-1,-0.5) {$\widehat C$}; \node[state] (s3) at (-3.1,-0.75) {$\widehat B$}; \node[state] (s5) at (-5,-0.5) {$\widehat A$}; \node[] at (-2,0.5) {$(s_1,a), \rewformat{0}$}; \rho[->] (s0) edge node[anchor=south west,inner sep=1pt] {$0.001$} (s2) (s0) edge[out=-30,in=45] node[left]{$0.999$} (s3) (s0) edge node[above left, inner sep=1pt] {$(s_1, b), \rewformat{0}$} (s5) (s5) edge[out=-30,in=190] node[below] {$(s_2, b), \rewformat{5}$} (s3); \node[] at (-2,-2.5) {}; \end{tikzpicture} } } \caption{An example of how the MEC quotient is constructed. By $a, \rewformat{r}$ we denote that the action $a$ yields a reward of $\rewformat{r}$.} \end{figure} \begin{example} Figure~\ref{fig:example} shows an MDP with three MECs, $\widehat{A} = (\set{s_2}, \set{a}), \widehat{B} = (\set{s_3,s_4}, \set{a}), \widehat{C} = (\set{s_5,s_6}, \set{a}))$. Its MEC quotient is shown in Figure~\ref{fig:quo}. \qee \end{example} \begin{remark} In general, the MEC quotient does not induce a DAG-structure, since there might be probabilistic transitions between MECs. Consider for example the MDP obtained by setting $\Delta(s_2, b, s_4) = \set*{s_1 \mapsto \frac{1}{2}, s_2 \mapsto \frac{1}{2}}$ in the MDP of Figure~\ref{fig:example}. Its MEC quotient then has $\widehat{\Delta}(\widehat{A}, (s_2, b)) = \set*{s_1 \mapsto \frac{1}{2}, \widehat{B} \mapsto \frac{1}{2}}$. \end{remark} \begin{remark}\label{rem:endcomp} The MEC decomposition of an MDP $\mathcal{M}$, i.e.\ the computation of $\mathsf{MEC}(\mathcal{M})$, can be achieved in polynomial time~\cite{CY95}. For improved algorithms on general MDPs and various special cases see~\cite{ChatterjeeH11,ChatterjeeH12,ChatterjeeH14,CL13}. \end{remark} \begin{definition}[MEC restricted MDP] Let $\mathcal{M}$ be an MDP and $(T, A) \in \mathsf{MEC}(\mathcal{M})$ a MEC of $\mathcal{M}$. By picking some initial state $s_\textrm{init}' \in T$, we obtain the \emph{restricted MDP} $\mathcal{M}' = (T, s_\textrm{init}', A, \mathsf{Av}', \Delta', r')$ where \begin{itemize} \item $\mathsf{Av}'(s) = \mathsf{Av}(s) \intersectionBin A$ for $s \in T$, \item $\Delta'(s, a, s') = \Delta(s, a, s')$ for $s, s' \in T$, $a \in A$, and \item $r'(s, a) = r(s, a)$ for $s \in T$, $a \in A$. \end{itemize} \end{definition} \paragraph{Classification of MDPs} If for some MDP $\mathcal{M}$, $(S,\actions)$ is a MEC, we call the MDP \emph{strongly connected}. If it contains a single MEC plus potentially some transient states, it is called \emph{(weakly) communicating}. Otherwise, it is called \emph{multichain}~\cite[Sect.~8.3]{Puterman}. For a Markov chain, let $\Delta^n(s, s')$ denote the probability of going from the state $s$ to state $s'$ in $n$ steps. The \emph{period $p$ of a pair} $s, s'$ is the greatest common divisor of all $n$'s with $\Delta^n(s, s') > 0$. The pair $s, s'$ is called \emph{periodic} if $p > 1$ and \emph{aperiodic} otherwise. A Markov chain is called aperiodic if all pairs $s, s'$ are aperiodic, otherwise the chain is called periodic. Similarly, an MDP is called aperiodic if every memoryless randomized strategy induces an aperiodic Markov chain, otherwise the MDP is called periodic. \paragraph{Long-run average reward} In this work, we consider the (maximum) \emph{long-run average reward} (or \emph{mean-payoff}) of an MDP, which intuitively describes the (maximum) average reward per step we expect to see when simulating the MDP for time going to infinity. Formally, let $R_i$ be a random variable, which for an infinite path $\rho = s_0 a_0 s_1 a_1 \dots$ returns $R_i(\rho) = r(s_i, a_i)$, i.e.\ the reward observed at step $i \geq 0$. Given a strategy $\pi$, the $n$-step average reward then is \begin{equation*} \gain^\pi_n(s) := \mathbb E^\pi_s \left (\frac1n\sum_{i=0}^{n-1} R_i \right), \end{equation*} and the \emph{long-run average reward} of the strategy $\pi$ is \begin{equation*} \gain^\pi(s) := \liminf_{n\to\infty} \gain^\pi_n. \end{equation*} The $\liminf$ is used in the definition, since the limit may not exist in general for an arbitrary strategy. Nevertheless, for finite MDPs the optimal limit-inferior (also called the \emph{value}) is attained by some memoryless deterministic strategy $\pi^* \in \Pi^{\mathsf{MD}}$ and is in fact the limit~\cite[Thm.~8.1.2]{Puterman}. \begin{equation*} \gain(s) := \sup_{\pi \in \Pi} \liminf_{n\to\infty} \mathbb E^\pi_s \left (\frac1n \sum_{i=0}^{n-1} R_i \right) = \sup_{\pi \in \Pi} \gain^\pi(s) = \max_{\pi \in \Pi^{\mathsf{MD}}} \gain^\pi(s) = \lim_{n\to\infty} \gain^{\pi^*}_n. \end{equation*} An alternative well-known characterization we use in this paper is \begin{equation} \gain(s) = \max_{\pi \in \Pi^{\mathsf{MD}}} \sum_{M \in \mathsf{MEC}} \mathbb P^\pi_s[\Diamond\Box M] \cdot \gain(M), \label{eq:decomp} \end{equation} where $\Diamond\Box M$ denotes the set of paths that eventually remain forever within $M$ and $\gain(M)$ is the unique value achievable in the MDP restricted to the MEC $M$. Note that $\gain(M)$ does not depend on the initial state chosen for the restriction. \section*{Appendix} \section{Proof of Theorem~\ref{th:quo}} \label{proof:quo} We use $\varbigcirc S$ to denote the event of reaching some state in $S$ in the next step. \begin{align*} \gain(s_\textrm{init}) & = \sup_{\pi\in \Pi} \sum_{M_i \in \mathsf{MEC}(\mathcal{M})} \mathbb P^\pi_{\mathcal{M}, s_\textrm{init}} (\Diamond\Box M_i) \cdot \gain(M_i) \\ & = r_{\max} \cdot \sup_{\pi\in \Pi} \sum_{M_i \in \mathsf{MEC}(\mathcal{M})} \mathbb P^\pi_{\mathcal{M}, s_\textrm{init}} (\Diamond\Box M_i) \cdot f(\hat{s}_i) \\ & \qquad \text{(since $\gain(M_i)=r_{\max}\cdot f(\hat{s}_i)$ by assumption)} \\ & = r_{\max} \cdot \sup_{\pi\in \Pi} \sum_{M_i \in \mathsf{MEC}(\mathcal{M})} \mathbb P^\pi_{\mathcal{M}^f, s_\textrm{init}^f} \left(\Diamond \left(\hat{s}_i \land \varbigcirc \left(s_+ \vee s_- \right) \right) \right) \cdot f(\hat{s}_i) \\ & = r_{\max} \cdot\sup_{\pi\in \Pi} \sum_{M_i\ \in \mathsf{MEC}(\mathcal{M})} \mathbb P^\pi_{\mathcal{M}^f, s_\textrm{init}^f} \left(\Diamond \hat{s}_i \right) \cdot \mathbb P^\pi_{\mathcal{M}^f, \hat{s}_i} \left(\varbigcirc (s_+ \vee s_-)\right) \cdot f(\hat{s}_i) \\ & \qquad \text{(by the Markov property)} \\ & = r_{\max} \cdot\sup_{\pi\in \Pi} \sum_{M_i \in \mathsf{MEC}(\mathcal{M})} \mathbb P^\pi_{\mathcal{M}^f, s_\textrm{init}^f} (\Diamond \hat{s}_i) \cdot \mathbb P^\pi_{\mathcal{M}^f, \hat{s}_i} (\varbigcirc s_+) \\ & \qquad \text{(since $\mathbb P^\pi_{\mathcal{M}^f, \hat{s}_i}(\varbigcirc s_+) = \mathbb P^\pi_{\mathcal{M}^f, \hat{s}_i} (\varbigcirc (s_+ \vee s_-)) \cdot f(\hat{s}_i)$)} \\ & = r_{\max} \cdot \sup_{\pi\in \Pi} \mathbb P^\pi_{\mathcal{M}^f}(\Diamond s_+). \tag*{\mbox{\qed}} \end{align*} \section{Proof of Theorem~\ref{th:lvi_corr}} Algorithm~\ref{alg:lvi} terminates since every part only takes finitely many steps. We now prove correctness, i.e.\ that upon termination the algorithm returns value $p$, such that $\abs{r_{\max} \cdot p - \gain(s_\textrm{init})} < \varepsilon$. \begin{align*} \sup_{\pi \in \Pi} \mathbb P^\pi_{\mathcal{M}^f, s_\textrm{init}^f}(\Diamond s_+) & = \sup_{\pi \in \Pi} \sum_{M_i\in \mathsf{MEC}(\mathcal{M})} \mathbb P^\pi_{\mathcal{M}, s_\textrm{init}} (\Diamond \Box M_i) \cdot f(\hat{s}_i) \\ & \qquad \text{(from the proof of Theorem~\ref{th:quo})} \\ & \leq \sup_{\pi \in \Pi} \sum_{M_i\in \mathsf{MEC}(\mathcal{M})} \mathbb P^\pi_{\mathcal{M}} (\Diamond \Box M_i) \frac{\gain(M_i) + \tfrac{1}{2} \varepsilon}{r_{\max}} \quad \text{(by assumption)}\\ & = \frac{1}{r_{\max}} \cdot \sup_{\pi \in \Pi} \sum_{M_i\in \mathsf{MEC}(\mathcal{M})} \mathbb P^\pi_{\mathcal{M}, s_\textrm{init}} (\Diamond \Box M_i) \cdot \left(\gain(M_i) + {\tfrac{1}{2}} \varepsilon \right)\\ & \leq \frac{1}{r_{\max}} \left( \tfrac{1}{2} \varepsilon + \sup_{\pi \in \Pi} \sum_{M_i\in \mathsf{MEC}(\mathcal{M})} \mathbb P^\pi_{\mathcal{M}, s_\textrm{init}}(\Diamond \Box M_i) \cdot \gain(M_i) \right) \\ & = \frac{\gain(s_\textrm{init})}{r_{\max}} + \frac{\varepsilon}{2r_{\max}}. \end{align*} Analogously we can prove that $\sup_{\pi \in \Pi} \mathbb P^\pi_{\mathcal{M}^f, s_\textrm{init}^f}(\Diamond s_+) \geq \frac{\gain(s_\textrm{init})}{r_{\max}} - \frac{\varepsilon}{2r_{\max}}$. Together \begin{equation} \label{eq:sth1} \abs{\sup_{\pi \in \Pi} \mathbb P^\pi_{\mathcal{M}^f, s_\textrm{init}^f}(\Diamond s_+) - \frac{\gain(s_\textrm{init})}{r_{\max}}} \leq \frac{\varepsilon}{2 r_{\max}} \end{equation} By assumption it also holds that \begin{equation} \abs{p - \sup_{\pi \in \Pi} \mathbb P^\pi_{\mathcal{M}^f, s_\textrm{init}^f}(\Diamond s_+)} < \frac{\varepsilon}{2r_{\max}} \label{eq:sth2} \end{equation} By the triangle inequality we obtain from \eqref{eq:sth1} and \eqref{eq:sth2} that \begin{equation*} \abs{p - \frac{\gain(s_\textrm{init})}{r_{\max}}} < \frac{\varepsilon}{r_{\max}}. \end{equation*} and thus \begin{equation*} \abs{r_{\max} \cdot p - \gain(s_\textrm{init})} < \varepsilon. \tag*{\mbox{\qed}} \end{equation*} \section{Proof of Theorem~\ref{th:skeleton_corr}} We extend the proof of \cite[Thm.~3]{atva}, which can be found in the technical report \cite{DBLP:journals/corr/BrazdilCCFKKPU14}. To this end, we summarize some properties of $u$ and $l$. Let $M$ be some MEC and $\hat{s}_i$ its corresponding collapsed state. \begin{itemize} \item $u(\hat{s}_i)$ and $l(\hat{s}_i)$ are always and upper and lower bound of value of $M$. This follows from the error bound on VI within $M$. \item Let $n_i$ be the number of times the $\mathsf{stay}$ action is taken in $\hat{s}_i$. We have \begin{equation*} \lim_{n_i \to \infty} \left(u(\hat{s}_i) - l(\hat{s}_i)\right) = 0. \end{equation*} This follows from convergence of VI on MECs and by the fact that if $\mathsf{stay}$ is taken infinitely often, a round of VI on $M$ happens infinitely often. \end{itemize} Consequently, the proof of \cite[Thm.~3]{DBLP:journals/corr/BrazdilCCFKKPU14} applies. \qed \section{Definition of bounded MEC quotient} \label{sec:bounded_quotient} \begin{definition}[Bounded MEC quotient] Let $\widehat\mathcal{M} = (\widehatS, \hat{s}_\textrm{init}, \widehat\actions, \widehat\mathsf{Av}, \widehat\Delta, \widehatr)$ be the MEC quotient of an MDP $\mathcal{M}$ with collapsed states $\mathsf{MEC}_{\widehat{S}} = \set{\hat{s}_1, \dots, \hat{s}_n}$. Further, let $\lowerbound,\upperbound: \set{\hat{s}_1, \dots, \hat{s}_n} \to [0,1]$ be functions assigning a lower and upper bound, respectively, to every collapsed state. We define the \emph{bounded MEC quotient of $\mathcal{M}$ and $f$} as the MDP $\mathcal{M}^{\lowerbound,\upperbound} = (S^{\lowerbound,\upperbound}, s_\textrm{init}^{\lowerbound,\upperbound}, \widehat\actions \unionBin \set{\mathsf{stay}}, \mathsf{Av}^{\lowerbound,\upperbound}, \Delta^{\lowerbound,\upperbound}, r^{\lowerbound,\upperbound})$, where \begin{itemize} \item $S^{\lowerbound,\upperbound} = \widehatS \unionBin \set{s_+, s_-, s_?}$, \item $s_\textrm{init}^{\lowerbound,\upperbound} = \hat{s}_\textrm{init}$, \item $\mathsf{Av}^{\lowerbound,\upperbound}$ is defined as \begin{align*} \forall \hat{s} \in \widehatS.~& \mathsf{Av}^{\lowerbound,\upperbound}(\hat{s}) = \begin{dcases*} \widehat\mathsf{Av}(\hat{s}) \unionBin \set{\mathsf{stay}} & if $\hat{s} \in \mathsf{MEC}_{\widehat{S}}$, \\ \widehat\mathsf{Av}(\hat{s}) & otherwise, \end{dcases*} \\ & \mathsf{Av}^{\lowerbound,\upperbound}(s_+) = \mathsf{Av}^{\lowerbound,\upperbound}(s_-) = \mathsf{Av}^{{\lowerbound,\upperbound}}(s_?) = \emptyset, \end{align*} \item $\Delta^{\lowerbound,\upperbound}$ is defined as \begin{gather*} \forall \hat{s} \in \widehatS, \hat{a} \in \widehat\mathsf{Av}(\hat{s}) \setminus \set{\mathsf{stay}}.~ \Delta^{\lowerbound,\upperbound}(\hat{s}, \hat{a}) = \widehat\Delta(\hat{s}, \hat{a}) \\ \forall \hat{s} \in \mathsf{MEC}_{\widehat{S}}.~ \Delta^{\lowerbound,\upperbound}(\hat{s}, \mathsf{stay}) = \set*{s_+ \mapsto l(\hat{s}), s_- \mapsto 1 - u(\hat{s}), s_? \mapsto u(\hat{s}) - l(\hat{s})}, \end{gather*} \item and the reward function $r^{\lowerbound,\upperbound}(\hat{s}, \hat{a})$ is chosen arbitrarily (e.g.\ $0$ everywhere), since we only consider a reachability problem on $\mathcal{M}^{\lowerbound,\upperbound}$. \end{itemize} \end{definition} \section{Value Iteration Solutions} \label{sec:vi} \subsection{Naive value iteration} \label{sec:nvi} Value iteration is a dynamic-programming technique applicable in many contexts. It is based on the idea of repetitively updating an approximation of the value for each state using the previous approximates until the outcome is precise enough. The standard value iteration for average reward~\cite[Sect.~8.5.1]{Puterman} is shown in Algorithm~\ref{alg:vi}. \begin{algorithm}[t] \caption{\textsc{ValueIteration}} \label{alg:vi} \begin{algorithmic}[1] \Require MDP $\mathcal{M} = (S,s_\textrm{init},\actions,\mathsf{Av},\Delta,r)$, precision $\varepsilon > 0$ \Ensure $w$, s.t. $\abs{w - v(s_\textrm{init})} < \varepsilon$ \State $t_0(\cdot) \gets 0$, $n \gets 0$. \While {stopping criterion not met} \State $n \gets n+1$ \For {$s \in S$} \State $t_n(s) = \max_{a\in \mathsf{Av}(s)} \left(r(s,a) + \sum_{s'\in S} \Delta(s, a, s') t_{n-1}(s')\right)$ \EndFor \EndWhile \State \Return $\frac1n t_n(s_\textrm{init})$ \end{algorithmic} \end{algorithm} First, the algorithm sets $t_0(s) = 0$ for every $s \in S$. Then, in the inner loop, the value $t_n$ is computed from the value of $t_{n-1}$ by choosing the action which maximizes the expected reward plus successor values. This way, $t_n$ in fact describes the optimal \emph{expected $n$-step total reward} \begin{equation*} t_n(s) = \max_{\pi \in \Pi^{\mathsf{MD}}} \mathbb E^\pi_s \left( \sum_{i=0}^{n-1} R_i \right) = n \cdot \max_{\pi \in \Pi^{\mathsf{MD}}} \gain^\pi_n(s). \end{equation*} Moreover, $t_n$ approximates the $n$-multiple of the long-run average reward \begin{theorem}[{\cite[Thm.~9.4.1]{Puterman}}] \label{th:vi_conv} For any MDP $\mathcal{M}$ and any $s \in S$ we have $\lim_{n \to \infty} \frac1n t_n(s) = \gain(s)$ for $t_n$ obtained by Algorithm~\ref{alg:vi}. \end{theorem} \subsubsection{Stopping criteria} The convergence property of Theorem~\ref{th:vi_conv} is not enough to make the algorithm practical, since it is not known when to stop the approximation process in general. For this reason, we discuss stopping criteria which describe when it is safe to do so. More precisely, for a chosen $\varepsilon > 0$ the stopping criterion guarantees that when it is met, we can provide a value $w$ that is $\varepsilon$-close to the average reward $\gain(s_\textrm{init})$. We recall a stopping criterion for communicating MDPs defined and proven correct in~\cite[Sect.~9.5.3]{Puterman}. Note that in a communicating MDP, all states have the same average reward, which we simply denote by $\gain$. For ease of notation, we enumerate the states of the MDP $S = \set{s_1, \dots, s_n}$ and treat the function $t_n$ as a vector of values $\vec{t}_n = (t_n(s_1), \dots, t_n(s_{n}))$. Further, we define the relative difference of the value iteration iterates as $\vec{\Delta}_{n} := \vec{t}_{n} - \vec{t}_{n-1}$ and introduce the \emph{span semi-norm}, which is defined as the difference between the maximum and minimum element of a vector $\vec{w}$ \begin{equation*} \spannorm{\vec{w}} = \max_{s \in S} \vec{w}(s) - \min_{s \in S} \vec{w}(s). \end{equation*} The stopping criterion then is given by the condition \begin{equation} \label{eq:sc_uni} \spannorm{\vec{\Delta}_{n}} < \varepsilon. \tag{SC1} \end{equation} When the criterion \eqref{eq:sc_uni} is satisfied we have that \begin{equation} \label{eq:guar_uni} \abs{\vec{\Delta}_{n}(s) - \gain} < \varepsilon \qquad\forall s \in S. \end{equation} Moreover, we know that for communicating aperiodic MDPs the criterion \eqref{eq:sc_uni} is satisfied after finitely many steps of Algorithm~\ref{alg:vi}~\cite[Thm.~8.5.2]{Puterman}. Furthermore, periodic MDPs can be transformed into aperiodic without affecting the average reward. The transformation works by introducing a self-loop on each state and adapting the rewards accordingly~\cite[Sect.~8.5.4]{Puterman}. Although this transformation may slow down VI, convergence can now be guaranteed and we can obtain $\varepsilon$-optimal values for any communicating MDP. The intuition behind this stopping criterion can be explained as follows. When the computed span norm is small, $\vec{\Delta}_n$ contains nearly the same value in each component. This means that the difference between the expected ($n-1$)-step and $n$-step total reward is roughly the same in each state. Since in each state the $n$-step total reward is greedily optimized, there is no possibility of getting more than this difference per step. Unfortunately, this stopping criterion cannot be applied on general MDPs, as it relies on the fact that all states have the same value, which is not true in general. Consider for example the MDP of Figure~\ref{fig:example}. There, we have that $\gain(s_5) = \gain(s_6) = 10$ but $\gain(s_3) = \gain(s_4) = 5$. In~\cite[Sect.~9.4.2]{Puterman}, it is conjectured that the following criterion may be applicable to general MDPs: \begin{equation} \label{eq:sc_con1} \spannorm{\vec{\Delta}_{n-1}} - \spannorm{\vec{\Delta}_{n}} < \varepsilon. \tag{SC2} \end{equation} This stopping criterion requires that the difference of spans becomes small enough. While investigating the problem, we also conjectured a slight variation: \begin{equation} \label{eq:sc_con2} \norm{\vec{\Delta}_{n} - \vec{\Delta}_{n-1}}_{\infty} < \varepsilon, \tag{SC3} \end{equation} where $\norm{\vec{w}}_\infty = \max_{s\in S} \vec{w}(s)$. Intuitively, both of these criteria try to extend the intuition of the communicating criterion to general MDPs, i.e.\ to require that in each state the reward gained per step stabilizes. Example~\ref{ex:break} however demonstrates that neither \eqref{eq:sc_con1} nor \eqref{eq:sc_con2} is a valid stopping criterion. \begin{figure}[t] \centering \begin{tikzpicture}[auto] \node[state] (state_0) {$s_0$}; \node[state,right=2cm of state_0] (state_1) {$s_1$}; \coordinate[at= ($(state_0) + (30:0.5)$),label={[label distance=7pt] 60:$a,\mathbf{0}$}] (state_0_a); \draw[-] (state_0) -- (state_0_a); \rho[->] (state_0_a) edge[out=30, in=90,looseness=3] node[anchor=south east] {$0.9$} (state_0) (state_0_a) edge[out=30, in=150] node {$0.1$} (state_1) (state_0) edge[loop left] node {$b, \rewformat{0.9}\cdot\boldsymbol\alpha$} (state_0) (state_1) edge[loop right] node {$a, \boldsymbol\alpha$} (state_1) (state_1) edge[out=210, in=-30] node {$b, \rewformat{0}$} (state_0); \end{tikzpicture} \caption{A~communicating MDP parametrized by the value $\alpha$.} \label{fig:mc_though} \end{figure} \begin{example} \label{ex:break} Consider the (aperiodic communicating) MDP in Figure~\ref{fig:mc_though} with a parametrized reward value $\alpha \geq 0$. The optimal average reward is $\gain = \alpha$. But the first three vectors computed by value iteration are $\vec{t}_0 = (0, 0), \vec{t}_1 = (0.9 \cdot \alpha, \alpha), \vec{t}_2 = (1.8 \cdot \alpha, 2 \cdot \alpha)$. Thus, the values of $\vec{\Delta}_1 = \vec{\Delta}_2 = (0.9 \cdot \alpha, \alpha)$ coincide, which means that for every choice of $\varepsilon$ both stopping criteria \eqref{eq:sc_con1} and \eqref{eq:sc_con2} are satisfied by the third iteration. However, by increasing the value of $\alpha$ we can make the difference between the average reward $\vec{v}$ and $\vec{\Delta}_2$ arbitrary large, so no guarantee like in Equation~\eqref{eq:guar_uni} is possible. \qee \end{example} \subsection{Local value iteration} \label{sec:lvi} In order to remedy the lack of stopping criteria, we provide a modification of VI using MEC decomposition which is able to provide us with an $\varepsilon$-optimal result, utilizing the principle of Equation \eqref{eq:decomp}. The idea is that for each MEC we compute an $\varepsilon$-optimal value, then consider these values fixed and propagate them through the MDP quotient. Apart from providing a stopping criterion, this has another practical advantage. Observe that the naive algorithm updates all states of the model even if the approximation in a single MEC has not $\varepsilon$-converged. The same happens even when all MECs are already $\varepsilon$-converged and the values only need to propagate along the transient states. These additional updates of already $\varepsilon$-converged states may come at a high computational cost. Instead, our method adapts to the potentially very different speeds of convergence in each MEC. The propagation of the MEC values can be done efficiently by transforming the whole problem to a reachability instance on a modified version of the MEC quotient, which can be solved by, for instance, VI. We call this variant the \emph{weighted MEC quotient}. To obtain this weighted quotient, we assume that we have already computed approximate values $w(M)$ of each MEC $M$. We then collapse the MECs as in the MEC quotient but furthermore introduce new states $s_+$ and $s_-$, which can be reached from each collapsed state by a special action $\mathsf{stay}$ with probabilities corresponding to the approximate value of the MEC. Intuitively, by taking this action the strategy decides to \enquote{stay} in this MEC and obtain the average reward of the MEC. Formally, we define the function $f$ as the normalized approximated value, i.e.\ for some MEC $M_i$ we set $f(\hat{s}_i) = \frac{1}{r_{\max}} w(M_i)$, so that it takes values in $[0, 1]$. Then, the probability of reaching $s_+$ upon taking the $\mathsf{stay}$ action in $\hat{s}_i$ is defined as $f(\hat{s}_i)$ and dually the transition to $s_-$ is assigned $1 - f(\hat{s}_i)$ probability. If for example some MEC $M$ had a value $v(M) = \frac{2}{3} r_{\max}$, we would have that $\Delta(\hat{s}, \mathsf{stay}, s_+) = \frac{2}{3}$. This way, we can interpret reaching $s_+$ as obtaining the maximal possible reward, and reaching $s_-$ to obtaining no reward. With this intuition, we show in Theorem~\ref{th:quo} that the problem of computing the average reward is reduced to computing the value of each MEC and determining the maximum probability of reaching the state $s_+$ in the weighted MEC quotient. \begin{definition}[Weighted MEC quotient] \label{def:wmq} Let $\widehat\mathcal{M} = (\widehatS, \hat{s}_\textrm{init}, \widehat\actions, \widehat\mathsf{Av}, \widehat\Delta, \widehatr)$ be the MEC quotient of an MDP $\mathcal{M}$ and let $\mathsf{MEC}_{\widehat{S}} = \set{\hat{s}_1, \dots, \hat{s}_n}$ be the set of collapsed states. Further, let $f :\mathsf{MEC}_{\widehat{S}} \to [0,1]$ be a function assigning a value to every collapsed state. We define the \emph{weighted MEC quotient of $\mathcal{M}$ and $f$} as the MDP $\mathcal{M}^f = (S^f, s_\textrm{init}^f, \widehat\actions \unionBin \set{\mathsf{stay}}, \mathsf{Av}^f, \Delta^f, r^f)$, where \begin{itemize} \item $S^f = \widehatS \unionBin \set{s_+, s_-}$, \item $s_\textrm{init}^f = \hat{s}_\textrm{init}$, \item $\mathsf{Av}^f$ is defined as \begin{align*} \forall \hat{s} \in \widehatS.~& \mathsf{Av}^f(\hat{s}) = \begin{dcases*} \widehat\mathsf{Av}(\hat{s}) \unionBin \set{\mathsf{stay}} &if $\hat{s} \in \mathsf{MEC}_{\widehat{S}},$ \\ \widehat\mathsf{Av}(\hat{s}) &otherwise, \end{dcases*} \\ & \mathsf{Av}^f(s_+) = \mathsf{Av}^f(s_-) = \emptyset, \end{align*} \item $\Delta^f$ is defined as \begin{align*} \forall \hat{s} \in \widehatS, \hat{a} \in \widehat\actions \setminus \set{\mathsf{stay}}.~& \Delta^f(\hat{s}, \hat{a}) = \widehat\Delta(\hat{s}, \hat{a}) \\ \forall \hat{s}_i \in \mathsf{MEC}_{\widehat{S}}.~& \Delta^f(\hat{s}_i, \mathsf{stay}) = \set{s_+ \mapsto f(\hat{s}_i, s_- \mapsto 1 - f(\hat{s}_i)}, \end{align*} \item and the reward function $r^f(\hat{s}, \hat{a})$ is chosen arbitrarily (e.g.\ $0$ everywhere), since we only consider a reachability problem on $\mathcal{M}^f$. \end{itemize} \end{definition} \begin{figure}[t] \centering \scalebox{0.9}{ \begin{tikzpicture}[align=center] \node[state] (s0) at (-3,0.5) {$s_1$}; \node[state] (s2) at (-1,-0.5) {$\widehat C$}; \node[state] (s3) at (-3.1,-0.75) {$\widehat B$}; \node[state] (s5) at (-5,-0.5) {$\widehat A$}; \node[state] (sp) at (-2.15,-2.25) {$s_+$}; \node[state] (sm) at (-4.15,-2.25) {$s_-$}; \node[] at (-2,0.5) {$(s_1,a)$}; \rho[->] (s0) edge node[anchor=south west,inner sep=1pt] {$0.001$} (s2) (s0) edge[out=-30,in=45] node[left]{$0.999$} (s3) (s0) edge node[above left, inner sep=1pt] {$(s_1, b)$} (s5) (s5) edge[out=-30,in=190] node[above] {$(s_2, b)$} (s3) (s5) edge[out=-80,in=170,looseness=0.75] node[anchor=south, inner sep=5pt, pos=0.3]{$\frac{4}{10}$} (sp) (s5) edge[out=-80,in=120,looseness=0.5] node[left]{$\frac{6}{10}$} (sm) (s2) edge node[left,pos=0.3] {$1$} (sp) (s3) edge[out=-90,in=150] node[anchor=south west,inner sep=0pt]{$\frac{5}{10}$} (sp) (s3) edge[out=-90,in=30] node[anchor=south east,inner sep=0pt]{$\frac{5}{10}$} (sm); \end{tikzpicture} } \caption{The weighted quotient of the MDP in Figure~\ref{fig:example} and function $f = \set*{\widehat{A} \mapsto \frac{4}{10}, \widehat{B} \mapsto \frac{5}{10}, \widehat{C} \mapsto \frac{10}{10}}$. Rewards and $\mathsf{stay}$ action labels omitted for readability.} \label{fig:rquo} \end{figure} \begin{example} Consider the MDP in Figure~\ref{fig:example}. The average rewards of the MECs are $\gain = \set*{\widehat{A} \mapsto 4, \widehat{B} \mapsto 5, \widehat{C} \mapsto 10}$. With $f$ defined as in Theorem~\ref{th:quo}, Figure~\ref{fig:rquo} shows the weighted MEC quotient $\mathcal{M}^f$. \qee \end{example} \begin{theorem} \label{th:quo} Given an MDP $\mathcal{M}$ with MECs $\mathsf{MEC}(\mathcal{M}) = \set*{M_1, \dots, M_n}$, define $f(\hat{s}_i) = \tfrac{1}{r_{\max}} \gain(M_i)$ the function mapping each MEC $M_i$ to its value. Moreover, let $\mathcal{M}^f$ be the weighted MEC quotient of $\mathcal{M}$ and $f$. Then \begin{equation*} \gain(s_\textrm{init}) = r_{\max} \cdot \sup_{\pi \in \Pi} \mathbb P^\pi_{\mathcal{M}^f, s_\textrm{init}^f}(\Diamond s_+). \end{equation*} \end{theorem} \begin{algorithm}[t] \caption{\textsc{LocalVI}} \label{alg:lvi} \begin{algorithmic}[1] \Require MDP $\mathcal{M} = (S,s_\textrm{init},\actions,\mathsf{Av},\Delta,r)$, precision $\varepsilon > 0$ \Ensure $w$, s.t.\ $\abs{w - \gain(s_\textrm{init})} < \varepsilon$ \State $f = \emptyset$ \For {$M_i = (T_i, A_i) \in \mathsf{MEC}(\mathcal{M})$} \Comment Determine values for MECs \State Compute the average reward $w(M_i)$ on $M$, such that $\abs*{w(M_i) - \gain(M_i)} < \tfrac{1}{2} \varepsilon$, \State $f(\hat{s}_i) \gets \frac{1}{r_{\max}} w(M_i)$ \EndFor \State $\mathcal{M}^f\gets$ the weighted MEC quotient of $\mathcal{M}$ and $f$ \State Compute $p$ s.t.\ $\abs*{p - \sup_{\pi \in \Pi} \mathbb P^\pi_{\mathcal{M}^f, s_\textrm{init}^f}(\Diamond s_+)} < \frac{1}{2 r_{\max}} \varepsilon$ \Comment Determine reachability \State \Return $r_{\max} \cdot p$ \end{algorithmic} \end{algorithm} The corresponding algorithm is shown in Algorithm~\ref{alg:lvi}. It takes an MDP and the required precision $\varepsilon$ as input and returns a value $w$, which is $\varepsilon$-close to the average reward $\gain(s_\textrm{init})$. In the first part, for each MEC $M$ the algorithm computes an approximate average reward $w(M)$ and assigns it to the function $f$ (normalized by $r_{\max}$). Every MEC is a communicating MDP, therefore the value $w(M)$ can be computed using the naive VI with \eqref{eq:sc_uni} as the stopping criterion. In the second part, the weighted MEC quotient of $\mathcal{M}$ and $f$ is constructed and the maximum probability $p$ of reaching $s_+$ in $\mathcal{M}^f$ is approximated. \begin{theorem} \label{th:lvi_corr} For every MDP $\mathcal{M}$ and $\varepsilon > 0$, Algorithm~\ref{alg:lvi} terminates and is correct, i.e.\ returns a value $w$, s.t.\ $\abs{w - \gain(s_\textrm{init})} < \varepsilon$. \end{theorem} For the correctness, we require that $p$ is $\frac{\varepsilon}{2r_{\max}}$-close to the real maximum probability of reaching $s_+$. This can be achieved by using the VI algorithms for reachability from \cite{atva} or \cite{haddad2014reachability}, which guarantee error bounds on the computed probability. Note that $p$ can also be computed by other methods, such as linear programming. In Section~\ref{sec:exper} we empirically compare these approaches. \subsection{On-demand value iteration} \label{sec:odv} Observe that in Algorithm~\ref{alg:lvi}, the approximations for all MECs are equally precise, irrespective of the effect a MEC's value has on the overall value of the MDP. Moreover, the whole model is stored in memory and all the MECs are computed beforehand, which can be expensive for large MDPs. Often this is unnecessary, as we illustrate in the following example. \begin{example} There are three MECs $\widehat{A}, \widehat{B}, \widehat{C}$ in the MDP of Figure~\ref{fig:example}. Furthermore, we have that $\mathbb P^{\pi}_{s_\textrm{init}}(\Diamond \widehat{C}) \leq 0.001$. By using the intuition of Equation \eqref{eq:decomp}, we see that no matter where in the interval $[0, r_{\max}=20]$ its value lies, it contributes to the overall value $\gain(s_\textrm{init})$ at most by $0.001 \cdot r_{\max} = 0.02$. If the required precision were $\varepsilon = 0.1$, the effort invested in computing the value of $\widehat{C}$ would not pay off at all and one can completely omit constructing $\widehat{C}$. Further, suppose that $\widehat{A}$ was a more complicated MEC, but after a few iterations the criterion \eqref{eq:sc_uni} already shows that the value of $\widehat{A}$ is at most $4.4$. Similarly, after several iterations in $\widehat{B}$, we might see that the value of $\widehat{B}$ is greater than $4.5$. In this situation, there is no point in further approximating the value of $\widehat{A}$ since the action $b$ leading to it will not be optimal anyway, and its precise value will not be reflected in the result. \qee \end{example} To eliminate these inefficient updates, we employ the methodology of \emph{bounded real-time dynamic programming} (BRTDP)~\cite{DBLP:conf/icml/McMahanLG05} adapted to the undiscounted setting in~\cite{atva}. The word \emph{bounded} refers to keeping and updating both a lower and an upper bound on the final result. It has been shown in \cite{Puterman,ChatterjeeI14} that bounds for the value of a MEC can be derived from the current maximum and minimum of the approximations of VI. The idea of the BRTDP approach is to perform updates not repetitively for all states in a fixed order, but more often on the \emph{more important} states. Technically, finite runs of the system are sampled, and updates to the bounds are propagated only along the states of the current run. Since successors are sampled according to the transition probabilities, the frequently visited (and thus updated) states are those with high probability of being reached, and therefore also having more impact on the result. In order to guarantee convergence, the non-determinism is resolved by taking the \emph{most promising action}, i.e.\ the one with the current highest upper bound. Intuitively, when after subsequent updates such an action turns out to be worse than hoped for, its upper bound decreases and a more promising action is chosen next time. Since BRTDP of~\cite{atva} is developed only for MDP with the reachability (and LTL) objective, we decompose our problem into a reachability and MEC analysis part. In order to avoid pre-computation of all MECs with the same precision, we instead compute the values for each MEC only when they could influence the long-run average reward starting from the initial state. Intuitively, the more a particular MEC is encountered while sampling, the more it is \enquote{reached} and the more precise information we require about its value. To achieve this, we store upper and lower bounds on its value in the functions $u$ and $l$ and refine them on demand by applying VI. We modify the definition of the weighted MEC quotient to incorporate these lower and upper bounds by introducing the state $s_?$ (in addition to $s_+, s_-$). We call this construction the \emph{bounded MEC quotient}. Intuitively, the probability of reaching $s_+$ from a collapsed state now represents the lower bound on its value, while the probability of reaching $s_?$ describes the gap between the upper and lower bound. \begin{definition}[Bounded MEC quotient] \label{def:bquo} Let $\widehat\mathcal{M} = (\widehatS, \hat{s}_\textrm{init}, \widehat\actions, \widehat\mathsf{Av}, \widehat\Delta, \widehatr)$ be the MEC quotient of an MDP $\mathcal{M}$ with collapsed states $\mathsf{MEC}_{\widehat{S}} = \set{\hat{s}_1, \dots, \hat{s}_n}$ and let $\lowerbound,\upperbound: \set{\hat{s}_1, \dots, \hat{s}_n} \to [0,1]$ be functions that assign a lower and upper bound, respectively, to every collapsed state in $\widehat\mathcal{M}$. The \emph{bounded MEC quotient $\mathcal{M}^{\lowerbound,\upperbound}$ of $\mathcal{M}$ and $\lowerbound,\upperbound$} is defined as in Definition~\ref{def:wmq} with the following changes. \begin{itemize} \item $S^{\lowerbound,\upperbound} = \widehatS \unionBin \set{s_?}$, \item $\mathsf{Av}^{\lowerbound,\upperbound}(s_?) = \emptyset$, \item $\forall \hat{s} \in \mathsf{MEC}_{\widehat{S}}.~\Delta^{\lowerbound,\upperbound}(\hat{s}, \mathsf{stay}) = \set*{s_+ \mapsto l(\hat{s}), s_- \mapsto 1 - u(\hat{s}), s_? \mapsto u(\hat{s}) - l(\hat{s})}$. \end{itemize} The unshortened definition can be found in \cite[Appendix~D]{techrep}. \end{definition} The probability of reaching $s_+$ and the probability of reaching $\set*{s_+, s_?}$ give the lower and upper bound on the value $\gain(s_\textrm{init})$, respectively. \begin{corollary} \label{th:quo_bounded} Let $\mathcal{M}$ be an MDP and $\lowerbound,\upperbound$ functions mapping each MEC $M_i$ of $\mathcal{M}$ to (normalized) lower and upper bounds on the value, respectively, i.e.\ $l(\hat{s}_i) \leq \frac{1}{r_{\max}} \gain(M_i) \leq u(\hat{s}_i)$. Then \begin{equation*} r_{\max} \cdot \sup_{\pi \in \Pi} \mathbb P^\pi_{\mathcal{M}^{\lowerbound,\upperbound}, s_\textrm{init}^{\lowerbound,\upperbound}}(\Diamond s_+) \leq v(s_\textrm{init}) \leq r_{\max} \cdot \sup_{\pi \in \Pi} \mathbb P^\pi_{\mathcal{M}^{\lowerbound,\upperbound}, s_\textrm{init}^{\lowerbound,\upperbound}}(\Diamond \set*{s_+, s_?}), \end{equation*} where $\mathcal{M}^{\lowerbound,\upperbound}$ is the bounded MEC quotient of $\mathcal{M}$ and $\lowerbound,\upperbound$. \end{corollary} \begin{algorithm}[t] \caption{\textsc{OnDemandVI}}\label{alg:skeleton} \begin{algorithmic}[1] \Require MDP $\mathcal{M} = (S, s_\textrm{init}, \actions, \mathsf{Av}, \Delta, r)$, precision $\varepsilon > 0$, threshold $k \geq 2$ \Ensure $w$, s.t.\ $\abs{w - v(s_\textrm{init})} < \varepsilon$ \State Set $u(\cdot, \cdot) \gets 1$, $u(s_-, \cdot) \gets 0$; $l(\cdot, \cdot) \gets 0$, $l(s_+, \cdot) \gets 1$ \Comment Initialize \State Let $A(s) := \argmax_{a \in \mathsf{Av}^{\lowerbound,\upperbound}(s)} u(s, a)$ \State Let $u(s) := \max_{a \in A(s)} u(s, a)$ and $l(s) := \max_{a \in A(s)} l(s, a)$ \Repeat \State $s \gets s_\textrm{init}^{\lowerbound,\upperbound}, w \gets s$ \Comment Generate path \Repeat \label{ln:explore-begin} \State $a \gets$ sampled uniformly from $A(s)$ \State $s \gets$ sampled according to $\Delta^{\lowerbound,\upperbound}(s,a)$ \State $w \gets w, a, s$ \Until $s \in \set{s_+,s_-,s_?}$ or $\textsf{Appear}(s, w) = k$ \Comment Terminate path \label{ln:explore-end} \If {$\mathrm{pop}(w) = s_?$} \label{ln:refine-begin} \Comment Refine MEC in which $\mathsf{stay}$ was taken \State $\mathrm{pop}(w)$ \State $\widehat{q} \gets$ $\mathrm{top}(w)$ \State Run VI on $\widehat{q}$, updating $u$ and $l$, until $u - l$ is halved \State Update $\Delta^{\lowerbound,\upperbound}(\widehat{q}, \mathsf{stay})$ according to Definition~\ref{def:bquo} \label{ln:refine-end} \ElsIf {$\textsf{Appear}(s, w) = k$} \Comment Update EC-collapsing \State \textsc{OnTheFlyEc} \label{alg:line:collapse} \EndIf \Repeat \label{ln:backpropagate-begin} \Comment Back-propagate values \State $a \gets \mathrm{pop}(w)$, $s \gets \mathrm{pop}(w)$ \State $u(s, a) \gets \sum_{s' \in S} \Delta(s, a, s') \cdot u(s')$ \State $l(s, a) \gets \sum_{s' \in S} \Delta(s, a, s') \cdot l(s')$ \Until $w = \emptyset$ \label{ln:backpropagate-end} \Until $u(s_\textrm{init}) - l(s_\textrm{init}) < \frac{2 \varepsilon}{r_{\max}}$ \Comment {Terminate} \State \Return $r_{\max} \cdot \tfrac{1}{2} (u(s_\textrm{init}) + l(s_\textrm{init}))$ \end{algorithmic} \end{algorithm} \begin{algorithm}[t] \makeatletter \renewcommand{\ALG@name}{Procedure} \makeatother \caption{\textsc{OnTheFlyEc}}\label{alg:fac} \begin{algorithmic}[1] \For {$(T_i,A_i) \in \mathsf{MEC}(\mathcal{M}^{\lowerbound,\upperbound})$} \State Collapse $(T_i,A_i)$ to $\hat{s}_i$ in $\mathcal{M}^{\lowerbound,\upperbound}$ \For {$s\in T_i, a\in \mathsf{Av}(s)\setminus A_i$} \State $u(\hat{s}_i, (s,a)) \gets u(s,a)$ \State $l(\hat{s}_i, (s,a)) \gets l(s,a)$ \EndFor \State Add the $\mathsf{stay}$ action according to Definition~\ref{def:bquo}. \EndFor \end{algorithmic} \end{algorithm} Algorithm~\ref{alg:skeleton} shows the on-demand VI. The implementation maintains a partial model of the MDP and $\mathcal{M}^{\lowerbound,\upperbound}$, which contains only the states explored by the runs. It interleaves two concepts: (i)~naive VI is used to provide upper and lower bounds on the value of discovered end components, (ii)~the method of~\cite{atva} is used to compute the reachability on the collapsed MDP. In lines~\ref{ln:explore-begin}--\ref{ln:explore-end} a random run is sampled following the \enquote{most promising} actions, i.e.\ the ones with maximal upper bound. The run terminates once it reaches $s_+, s_-$ or $s_?$, which only happens if $\mathsf{stay}$ was one of the most promising actions. A~likely arrival to $s_?$ reflects a high difference between the upper and lower bound and, if the run ends up in $s_?$, this indicates that the upper and lower bounds of the MEC probably have to be refined. Therefore, in lines~\ref{ln:refine-begin}--\ref{ln:refine-end} the algorithm resumes VI on the corresponding MEC to get a more precise result. This decreases the gap between the upper and lower bound for the corresponding collapsed state, thus decreasing the probability of reaching $s_?$ again. The algorithm uses the function $\textsf{Appear}(s, w) = \abs{\set{i \in \mathbb{N} \mid s = w[i]}}$ to count the number of occurrences of the state $s$ on the path $w$. Whenever we encounter the same state $k$ times (where $k$ is given as a parameter), this indicates that the run may have got stuck in an end component. In such a case, the algorithm calls \textsc{OnTheFlyEc}~\cite{atva}, presented in Procedure~\ref{alg:fac}, to detect and collapse end components of the partial model. By calling \textsc{OnTheFlyEc} we compute the bounded quotient of the MDP on the fly. Without collapsing the end components, our reachability method could remain forever in an end component, and thus never reach $s_+$, $s_-$ or $s_?$. Finally, in lines~\ref{ln:backpropagate-begin}--\ref{ln:backpropagate-end} we back-propagate the upper and lower bounds along the states of the simulation run. \begin{theorem} \label{th:skeleton_corr} For every MDP $\mathcal{M}$, $\varepsilon > 0$ and $k \geq 2$, Algorithm~\ref{alg:skeleton} terminates almost surely and is correct, i.e.\ returns a value $w$, s.t.\ $\abs{w - v(s_\textrm{init})} < \varepsilon$. \end{theorem}

https://arxiv.org/abs/1712.00468

Graph Signal Processing: Overview, Challenges and Applications

Research in Graph Signal Processing (GSP) aims to develop tools for processing data defined on irregular graph domains. In this paper we first provide an overview of core ideas in GSP and their connection to conventional digital signal processing. We then summarize recent developments in developing basic GSP tools, including methods for sampling, filtering or graph learning. Next, we review progress in several application areas using GSP, including processing and analysis of sensor network data, biological data, and applications to image processing and machine learning. We finish by providing a brief historical perspective to highlight how concepts recently developed in GSP build on top of prior research in other areas.

\subsection{Sensor networks} \label{sec:sensors} One of the most natural applications of Graph signal processing is in the context of sensor networks. A graph represents the relative positions of sensors in the environment, and the application goals include compression, denoising, reconstruction, or distributed processing of sensor data. Indeed, some of the initial explorations of graph-based processing focused on sensor networks \cite{wagnerbaraniuketal-SSPWorkshop2005,wagnerbaraniuketal-IPSN06,ciancio2006energy,shen2008joint}. A first approach to define a graph associated to a sensor network is to choose edge weights as a decreasing function of distance between nodes (sensors). Then, data observations that are similar at neighboring nodes lead naturally to a smooth (low-pass) graph signal. Such a smooth graph signal model makes it possible to detect outliers or abnormal values by high-pass filtering and thresholding \cite{Sandryhaila:2014ju,egilmez2014spectral}, or to build effective signal reconstruction methods from sparse set of sensor readings, as in \cite{Zhu:2012wc,Kaneko:2017ui,sakiyama2016efficient}, which can potentially lead to significant savings in energy resources, bandwidth and latency in sensor network applications. A second scenario is where the graph to be used for data analysis is given by the application. For example, urban data processing relies on data that naturally live on networks, such as energy, transportation or road networks. In these applications cases, GSP has been used to monitor urban air pollution \cite{Jain:db}, or to monitor and analyze power consumption \cite{He:2016jg}, for example. Some works such as \cite{Valdivia:2015ji,Chen:2016tt,Dong:2013eg} have used graph signal processing tools for analyzing traffic and mobility in large cities. For example, wavelets on graphs can serve to extract useful traffic patterns to detect disruptive traffic events such as congestion \cite{Mohan:2014cs}. Graph wavelet coefficients at different scales permit to infer useful information such as origin, propagation, and the span of traffic congestion. In some cases, relations between sensor readings are not exclusively explained by the distance between sensor locations, or by some actual network constraints. Other factors can influence the data values observed at the sensor readings such as the presence of geographical obstacles (e.g., in temperature measurements), or the interaction between networks of different types (e.g., how proximity to a freeway affects pollution measurements in a city). In some cases the phenomena that can explain these relations between measurements are latent and this leads to the challenging problem of learning a graph (see also Section~\ref{sec:graph-learning}) that can explain the data observations under signal smoothness or other signal model assumptions \cite{Dong:2016fm,Kalofolias:2016tf,Thanou:2017ec}. This allows inferring system features and behaviors that are hidden in the measured datasets (e.g., ozone datasets in \cite{Jablonski:js}). Finally, several of the graph signal processing operators presented in this paper are amenable to distributed implementations that are particularly interesting for large sensor networks, and which motivated some of the early work mentioned at the beginning of this section. For example, the graph multiplier operators can be approximated by Chebyshev polynomials in distributed implementation of smoothing, denoising, inverse filtering or semi-supervised learning tasks \cite{Shuman:2017ti}. The work in \cite{Wang:2015ik} for example studies the problem of distributed reconstruction of time-varying band-limited graph signals recorded by a subset of temperature sensor nodes. There is however still a lot of opportunities for the development of distributed GSP algorithms that are able to extend to large-scale networks and big data applications. \subsection{Biological networks} \label{sec:biological} \begin{figure*}[t!] \centering \includegraphics[trim=1cm 0cm 0cm 0cm, clip=true,width=0.98 \textwidth]{./figures/BRAIN_Colorbar_sig} \centering \includegraphics[width=0.98 \textwidth]{./figures/BRAIN_SignalVisual_3Day} \caption{Distribution of decomposed signals. (a), (b), and (c) are the absolute magnitudes for all brain regions with respect to graph low frequencies, graph middle frequencies and graph high frequencies, respectively. Higher concentration in graph low frequency results in better learning performance, when subjects are unfamiliar with the task \cite{huang2016}. Concentration in graph low frequency also helps faster response in switching attention between actions \cite{medaglia2018}. From \cite{huang2016}, with permission. } \label{fig_brain_signal_analytics} \end{figure*} Biological networks have also proved to be a popular application domain for graph signal processing, with recent research works focusing on the analysis of data from systems known to have a network structure, such as the human brain, and also on the inference of a priori unknown biological networks. Several works have studied human brain networks using the graph signal processing framework. For example, it has been observed that human brain activity signals can be mapped on a network (graph) where each node corresponds to a brain region. The network links (edge weights) are considered to be known a priori and represent the structural connectivity or the functional coherence between brain regions \cite{bullmore2012, sporns2011}. GSP tools such as the graph signal representations described in Section~\ref{sec:representations} can then be used to analyze the brain activity signal on the functional or structural brain network. For example, low frequencies in the graph signal represent similar activities in regions that are highly connected in the functional brain networks, while high frequencies denote very different activities in such brain regions. These ideas have been to analyze brain signals, leading to biologically plausible observations about the behavior the human cognitive system, as in for example \cite{huang2016, medaglia2018}. Fig. \ref{fig_brain_signal_analytics} illustrates the signal distribution of different graph frequency components in an active motor learning task. Interestingly, regions with strong signal in low and high graph frequency components overlap well with the regions known to contribute to better motor learning \cite{Bassett2011}. Additionally, it has been observed that there is a strong association between the actual brain networks (characterized by their spectral properties) and the level of exposure of subject to different tasks \cite{Goldsberry:2017ha}. Some works further build on the multi-resolution properties of spectral graph wavelet transforms to capture subtle connected patterns of brain activity or provide biologically meaningful decompositions of functional magnetic resonance imaging (fMRI) data \cite{Behjat:2015gp,Leonardi:2013bv,Atasoy:2016ev}. Interestingly, it is also possible to combine different sources of informations in the analysis of the brain networks. For example, the work in \cite{Griffa:2017in} integrates infra-slow neural oscillations and anatomical-connectivity maps derived from functional and diffusion MRI, in a multilayer-graph framework that captures transient networks of spatio-temporal connectivity. These networks group anatomically wired and temporary synchronized brain regions and encode the propagation of functional activity on the structural connectome, which contributes to a deeper understanding of the important structure-function relationships in the human brain. The GSP framework has also been proposed for the classification of brain graph signals \cite{Menoret:2017usa} and the analysis of anomalies or diseases \cite{Hu:2015fe,Hu:2016fw}. For example, source localization algorithms based on sparse regularization can be used to localize the possible origins of Alzheimer's disease based on a large set of repeated magnetic resonance imaging (MRI) scans. This can help understand the dynamics and origin of dementia, which is an important step towards developing effective treatment of neuro-degenerative diseases \cite{Hu:jn}. The growing number of publications studying brain activity or brain network features from a GSP perspective indicates that these are promising applications for the methods described in this paper. It should finally be noted that brain networks are not the only biological networks where GSP offers promising solutions. Graph signal processing elements and biological priors are combined to infer networks and discover meaningful interactions in gene regulatory networks, as in \cite{Pirayre:2015di,Pirayre:2017cm}. The inference of the structure of protein interaction networks has also been addressed with help of spectral graph templates \cite{Segarra_templates}. In particular, the observed matrix of mutual information can be approximated by some (unknown) analytic matrix function of the unobserved structure to be recovered. Observed data is then used to obtain the eigenvectors of the matrix representation of the graph model, and then the eigenvalues are estimated with the help of sparsity priors. The above examples are only some illustrations of the recent works that attempt to infer structures of biological networks using a GSP perspective. Biological networks that cannot be explicitly recorded and measured are potentially good applications for graph learning and inference methods in particular, which can uncover unknown interactions in the biological data. \subsection{Image and 3D point cloud processing} \label{sec:images} While graph signal processing is often applied to datasets that naturally exhibit irregular structures, it has also been applied to other datasets where conventional signal processing has been used for many years, including for example images and video sequences. An image to be processed can be viewed as a set of pixels, each associated to a vertex, forming a regular graph with all edge weights equal to 1 (e.g., a line graph or a grid graph). Indeed processing using the discrete Fourier transform or the discrete cosine transform (DCT) can be shown to have a simple interpretation in terms of the frequencies associated to those regular graphs \cite{strang1999discrete} (see also Section~\ref{subsubsec:historicalperspective}). Instead, recent work uses regular line and grid graph topologies, but with unequal edge weights that can adapt to the specific characteristic of an image or a set of images. A first set of approaches associates a different graph to each image, by associating smaller edge weights to connect pixels that are on opposite sides of an image contour. This type of image-dependent graph representation is strongly connected to popular image processing techniques, such as the bilateral filter and related methods \cite{milanfar2013tour}, which also apply signal dependent filtering and are widerly used in applications such as image restoration or denoising. Graphs are used to capture the geometric structure in images, such as contours that carry crucial visual information, in order to avoid blurring them during the filtering process. In addition to works that effectively extend image priors such as Total Variation (TV) minimization to graph representations (e.g., \cite{Couprie:2013fd,Najman:2017ug}), other works such as \cite{Tian:hy} or \cite{Pang:2015jo} use more specific GSP operators for denoising or filtering. In particular, the authors in \cite{Tian:hy} use graph spectral denoising methods to enhance the quality of images, while the work in \cite{Pang:2015jo} uses graph-based filters that influence the strength and direction of filtering for effective enhancement of natural images. A second avenue of research has considered situations where a graph is constructed as an efficient representation for a set of images, in particular in the context of image and video compression applications. The Karhunen-Lo\`eve transform (KLT) is known to provide the best transform coding gains under the assumption that the signals can be modeled as stationary Gaussian processes (which is often a good assumption for images). Indeed, extensive use of the DCT is often justified because it is optimal for a Gauss Markov Random Field (GMRF) with correlation $1$, which is an appropriate model for natural images. The inverse covariance matrix, or precision matrix, then corresponds to a line graph with equal weights. From this perspective, graph learning approaches can be used to learn precision matrices with structures and weights that capture statistics of specific types of images. For example, piecewise smooth images can be compressed using suitable Graph Fourier Transforms (GFT), which can be adapted to different types of image pixel blocks \cite{Hu:2015fs,Fracastoro:2016wn}. Graph-based transforms have also been used to code motion-compensated residuals in predictive video coding \cite{Egilmez:gh} with effective rate-distortion performance. \begin{figure}[t] \centering \includegraphics[width=7.2cm]{./figures/correspondences_man_sequence_3.png} \caption{ Example of motion estimation in a 3D point cloud sequence. Each frame is represented as a graph signal that captures the color and the geometry information of each voxel. Graph spectral features at each voxel capture the local graph signal properties and permit to find correspondences between frames at different time instances. A subset of the correspondences between the target (red) and the reference frame (green) are highlighted between small cubes that correspond to voxels. From \cite{Thanou:di}, with permission.} \label{fig:3dmotion} \end{figure} New visual modalities such as 3D meshes or 3D point clouds where data is sampled in irregular locations in 3D space, lend themselves naturally to graph representations. The color or 3D information supported by nodes or voxels are connected to their nearest neighbors to form a graph. Graph-based transforms can then be used to compress the resulting graph signals in static or dynamic point clouds \cite{Zhang:2014ey,Thanou:di}. In particular, the temporal redundancy between 3D point cloud frames at different time instants can be effectively estimated with help of graph spectral features \cite{Thanou:di}, as illustrated in Figure \ref{fig:3dmotion}. Graph-based transforms permit to properly exploit both the spatial correlation inside each frame and the temporal correlation between the frames, which eventually results in effective compression. Compression, however, is not the only application of GSP in 3D point clouds. Fast resampling methods, which are important in processing, registering or visualizing large point clouds, can also be built on graph-based randomized strategies to select representative subsets of points while preserving application-dependent features \cite{ChenTFVK:16}. \subsection{Machine Learning and Data Science} \label{sec:machine-learning} Graph methods have long played an important role in machine learning applications, as they provide a natural way to represent the structure of a dataset. In this context, each vertex represents one data point to which a label can be associated, and a graph can be formed by connecting vertices with edge weights that are assigned based on a decreasing function of the distance between data points in the feature space. Graph signal processing then enables different types of processing, learning or filtering operations on values associated to graph vertices. In a different context, GSP elements can be helpful to construct architectures that are able to classify signals that live on irregular structures. We give below some examples of machine learning applications in both contexts. When data labels are presented as signals on a (nearest neighbor) graph, graph signal regularization techniques can be used in the process of estimating labels \cite{zhu2003semi}, optimizing the prediction of unknown labels in classification \cite{Sandryhaila:2014ju} or semi-supervised learning problems \cite{ChenCRBGK:13}. Furthermore, as labeled samples are often a scarce and expensive resource in semi-supervised learning applications, graph sampling strategies such as those presented in Section~\ref{sec:sampling} can be helpful in determining the actual needs for labeled data and develop effective active learning algorithms \cite{gadde2014active}. Graphs can also be constructed to describe similarities between users or items in recommendation systems that assist customers in making decisions by collecting information about how other users rate particular services or items \cite{Benzi:fm}. Leveraging the notions of graph frequency and graph filters, classical collaborative filtering methods (such as $k$-nearest neighbors strategies), can then be implemented with specific \textit{band-stop} graph filters on graphs \cite{Huang:2017er}. Furthermore, linear latent factor models, such as low-rank matrix completion, can be viewed as \textit{bandlimited} interpolation algorithms that operate in a frequency domain given by the spectrum of a joint user and item network. This can serve to design effective graph filtering algorithms that lead to enhanced rating prediction in video recommendation applications, for example \cite{Huang:2017er}. Content-based recommendation can also be addressed as an online learning problem solved with spectral bandit algorithms \cite{Valko:2014wi}. The key idea is to represent the reward function in an online recommendation system as a linear combination of the eigenvectors of the similarity graph that connects the different items. With this representation it is possible to optimize the reward function by favoring smoothness on the graph, which has been shown to be effective in video recommendation examples \cite{Valko:2014wi}. \begin{figure*}[tb] \begin{minipage}[b]{.99\linewidth} \centering \includegraphics[width=0.3\textwidth]{figures/Ecole_2COM_liendoux_black} \hspace{0.2cm} \includegraphics[width=0.3\textwidth]{figures/Ecole_5COM_liendoux_black} \hspace{0.2cm} \includegraphics[width=0.3\textwidth]{figures/Ecole_10COM_liendoux_black} \end{minipage} \caption{Multiscale community structures in a graph of social interactions between children in a primary school. The different figures show the partition of the original social network in 2, 5 and 10 communities, respectively. From \cite{Tremblay:2014hl}, with permission.} \label{fig:communities} \end{figure*} Data clustering or community detection can also benefit from tools developed under the GSP framework. For example graph transforms, and especially graph wavelets, have been used to solve the classical problem of community detection \cite{Tremblay:2014hl}. The problem of detecting multiscale community in networks is cast as the problem of clustering nodes based on graph wavelets features. This allows the introduction of a notion of scale in the analysis of the network and as well as a sort of 'egocentered' view of how a particular node 'sees' the network (see Figure \ref{fig:communities}). Furthermore, the extension of clustering or community detection tasks to large-scale systems generally relies on sampling or randomized strategy where GSP methods can also be very helpful. For example, fast graph-based filtering of random signals can be used to estimate the graph structure, and in particular to approximate eigenvectors that are often crucial in the design of clustering algorithms and other machine learning tasks. One of the initial works in this direction \cite{Boutsidis:2015vp} proposes to use power methods (that can be shown to be related to graph filter operators) to speed up the computation of eigenvectors used in spectral clustering applications. More recently, a fast graph clustering algorithm that is provably as good as spectral clustering has been developed based on random signal filtering techniques \cite{Tremblay:2016vb}. Related ideas have been used in sketching \cite{Chi:2015cy,Gama:2016ui}, data visualization applications on large real-world datasets of millions of nodes \cite{Paratte:2016ua,ChenTFVK:16}, \NEW{or in analysis of dynamic networks \cite{dal2017wavelet}}. These examples provide evidence for the potential benefits of using GSP principles in big data applications. Finally, the GSP framework can also be used to design architectures to analyze or classify whole graph signals that originally live on irregular structures. In particular, the graph signal processing toolbox has been extensively used to extend convolutional deep learning techniques to data defined on graphs. The convolutional neural network paradigm has been generalized with help of GSP elements for the extraction of feature descriptors for 3D shapes \cite{Masci:2015ej,Boscaini:2015uea}. A localized spectral network architecture leveraging on localized vertex-frequency analysis has also been proposed in \cite{Bruna:2013vg}, and the use of heat kernels defined in the graph spectral domain has been developed in \cite{Boscaini:2016tx}. While the previous works mostly address the analysis of 3D shapes, convolutional neural networks (CNNs) can actually be extended to many other signals in high-dimensional irregular domains, such as social networks, brain connectomes or words embedding, by reformulation in the context of spectral graph theory. Here, the GSP framework leads to the development of fast localized convolutional filters on graphs \cite{Defferrard:2016vo} along with adapted pooling operators \cite{Khasanova:2017vs}. Unsurprisingly, deep network architectures for graphs signals have been actually tested in various applications domains, such as chemical molecule properties prediction \cite{Duvenaud:2015ww}, classification tasks on social networks \cite{Perozzi:2014ib}, autism spectrum disorder classification \cite{Anirudh:2017tn} or traffic forecasting \cite{Li:2017tn}. \section{Introduction and motivation \input{Introduction} \section{Key Ingredients of Graph Signal Processing \label{sec:asp-1} \input{KeyIngredients} \section{State-of-the-Art and Challenges \label{sec:soa} \input{state-of-art} \section{Graph Signal Processing Applications} \label{sec:applications} \input{Applications} \section{Conclusion} \label{sec:conclusion} \input{conclusion} \ifCLASSOPTIONcaptionsoff \newpage \fi \bibliographystyle{IEEEtran} \subsection{A historical perspective \subsubsection{Related work in other areas} \label{subsubsec:connectionsotherareas} With the advent of large \textit{un}structured datasets there has been an explosion of activity in developing new methods to infer knowledge from these data. Next we briefly discuss how problems and methods studied in areas such as network science, network processes, and graphical models, relate to those studied in GSP. \textbf{Network Science} addresses issues such as uncovering community relations, perceived alliances, quantifying connectedness, or determining the relevance of specific agents. Thus, much of this work focuses on studying what can be learned from the graphs structuring the data, rather than on the data itself. For example, determining the size of the giant component, distribution of component sizes, degree and clique distributions, clustering coefficients, betweeness and closeness centralities, path length, and network diameter \cite{newman2010networks,jackson2010social,easley2010networks}. In the literature, these questions are often the realm of what many call Network Science \cite{networksciencenas,borner2007network,lewis2011network,barabasi2016network}. Connections between this area of research and GSP primarily come from the fact that the graph spectra that GSP builds upon are strongly related to the structure of the graph \cite{Chung:96}. As an example, spectral clustering methods use the low frequency eigenvectors of the Laplacian and thus can be viewed from a GSP perspective \cite{Tremblay:2016vb}. \textbf{Network processes} aim to model propagation over networks, including such phenomena as diffusion of disease and epidemics, spread of (fake) news, memes, fads, voting trends, imitation and social influence, propagation of failures and blackouts. Common models are similar to stochastic automata where the state of the nodes (the ``data'') evolve through local rules, i.e., according to exogeneous (external to the network) and endogeneous (internal to the network) effects. For example, using terminology from epidemics, nodes of the graph representing agents or individuals of a population can be infected (adopt an opinion or spread a rumor, or a failed component), or susceptible (open to adopt an opinion or spread a rumor, or a performing component). Infected or failed nodes can heal and become susceptible again; susceptible nodes can become infected either by an action external to the network or by action of infected neighbors \cite{castellano2012competing,nowzari2016analysis}. Analysis of these network processes is difficult. Traditionally, the network is abstracted out, assuming that any node can infect any other node (full mixing or complete network). To account for the impact of the network \cite{ganesh2005effect}, resorting to numerical studies is precluded except for very small networks since the network state space $\{0,1\}^N$ grows exponentially fast ($2^N$, for $N$ agents). To study these processes \cite{butts2009revisiting,colizza2006role} one usually considers one of two asymptotic regimes: \begin{inparaenum}[1)] \item either long term behavior (time-asymptotics) and attempting to find the equilibrium distribution of the process \cite{zhang2014diffusion,zhang2015role}; or \item large network asymptotics (mean field approximation) \cite{draief2010epidemics} leading to the study of the qualitative behavior of nonlinear ordinary differential equations \cite{santos2015bi}. \end{inparaenum} Because asymptotic behavior can be seen to depend on the eigenstructure of the underlying graph, GSP representations as those discussed in Section~\ref{sec:representations} can be used to characterize the evolution of a system. As an example, several papers have explored the use of GSP techniques to improve the efficiency of value function estimation in a reinforcement learning scenario \cite{levorato2012structure,levorato2012reduced}. In the machine learning and statistics literature, there has been a concerted effort on \textbf{graphical models} focusing on inference and learning from large datasets, \cite{lauritzen1996graphical,jordan1998learning,wainwright2008graphical,koller2009probabilistic,whittaker2009graphical}. The data is modeled as a set of random variables described by a family of Gibbs probability distributions, and the underlying graph, whose nodes label the variables, captures statistical dependence and conditional independence among the data. Acyclic graphs \cite{bang2008digraphs,edwards2012introduction} represent Bayesian networks, while undirected graphs represent Markov random fields \cite{rozanov1982markov,kindermann1980markov,rue2005gaussian}. Graphical models exploit factorizations of the joint distribution to develop efficient message passing algorithms for inference and they find application in many areas, including in modeling texture and other features in image processing \cite{besag1974spatial,chellappa1993markov,moura1992recursive,balram1993noncausal,willsky2002multiresolution,vats2012finding}, see \cite{jordan2010major} for illustrative applications in several domains. A connection between GSP and graphical models can be seen in recent work on learning graph from data \cite{egilmez2017graph}, which makes use of Markov random field models to define optimality criteria for the learned graphs. \subsubsection{Historical perspective on graph signal processing} \label{subsubsec:historiacalperspective} We now briefly review some of the prior work that is more directly connected and in the spirit of signal processing on graphs, \cite{SandryhailaM:13,ShumanNFOV:13}. Although one can argue that any classification is arbitrary, we organize the discussion in two main streams of work. Some of these comments follow closely \cite{SandryhailaM:13,Sandryhaila2014big}. \textbf{From algebraic signal processing (ASP) to graph signal processing (GSP).} The sequence of papers \cite{Pueschel:03a,Pueschel:05e,Pueschel:08a,Pueschel:08b,Pueschel:08c} introduced ASP, an axiomatic approach to time signal processing. ASP starts from a signal model~$\Omega$. Many signal models are possible, and a relevant question is to determine which one is more appropriate for a given application or should be associated with a given linear transform. One of the motivations to develop ASP was to answer a simple question: why is the discrete cosine transform~(DCT) the linear transform commonly associated with processing images leading, as experimentally observed, to better results in lossy image compression than the discrete Fourier transform~(DFT). The DCT (type~II) is associated with many of the lossy image/video compression standards \cite{sikora2005trends,lengwehasatit2004scalable} from JPEG, to MPEG-1 \cite{le1992mpeg} to H.264/AVC \cite{hanzocherrimanstreit-08,luthra2003special} and HEVC. Based on~\cite{clarke1981relation}, see also item~6, page~153 in~\cite{jain1989fundamentals}, the energy compaction offered by the DCT was justified because the DCT asymptotically approximates the Karhunen-Lo\`eve transform (KLT) of a first order (causal) Markov process (a first order autoregressive AR(1) model). This of course is an unsatisfactory explanation since images are basically non-causal models with no left-right or up-down preferred direction; reference~\cite{Moura:98} does show more fittingly that the DCT and DST are the KLTs for first order \textit{non}-causal Markov random fields rather than causal Markov processes. By appropriate definition of a \textit{space} signal model \cite{Pueschel:08b}, the DCT Type~II and each one of its other fifteen cousin trigonometric linear transforms are shown to be Fourier transforms playing, for the (corresponding) space model, the role that the DFT plays for the time (cyclic) model. References~\cite{Pueschel:03a,Pueschel:05e,Pueschel:08a,Pueschel:08b} show that, under appropriate conditions, the signal model is generated from a simple filter, the \textit{shift}. In other words, under these conditions and signal model, choice of the shift determines filtering, convolution, the Fourier transform, frequency, and spectral analysis among other common concepts and constructs from traditional digital signal processing. This led \cite{SandryhailaM:13,Sandryhaila:2014ju} to introduce for signals indexed by nodes of an arbitrary directed or undirected graph the possibly weighted graph adjacency matrix as the shift that generates the graph signal model. This choice is satisfying in the sense that when the signal model is the time signal model or the space signal model the shifts and the graph signal model revert to the shifts and signal models in~\cite{Pueschel:08a} and~\cite{Pueschel:08b}, respectively. Subsequently, authors have proposed other shifts obtained from the adjacency matrix of the signal graph assumed~\cite{giraultgoncalvesfleury-2015,gavilizhang-2017} that attempt to preserve isometry of the shift but lose the locality of the adjacency matrix shift. \textbf{From graph Laplacian spectral clustering to Laplacian-based GSP.} References \cite{Tenenbaum:00,Roweis:00,Belkin:03,Donoho:03} develop low-dimensional representations for large high-dimensional data through spectral graph theory~\cite{Belkin:02,Belkin:03} and the graph Laplacian~\cite{Chung:96}, by projecting the data on a low-dimensional subspace generated by a small subset of the Laplacian eigenbasis. The use of the graph Laplacian is justified by assuming the data is smooth in the sense that if the dataset is large and samples uniformly randomly a low-dimensional manifold then the (empirical) graph Laplacian acting on a smooth function on this manifold is a good discrete approximation that converges pointwise and uniformly to the elliptic Laplace-Beltrami operator applied to this function as the number of points goes to infinity~\cite{heinaudibertluxburg-2005,ginekoltchinskii-2006,heinaudibertluxburg-2007}. References \cite{Coifman:05a,Coifman:05b,Coifman:06} choose discrete approximations to other continuous operators, for example, a conjugate to an elliptic Schr{\"o}dinger-type operator, and obtain other spectral bases for the characterization of the geometry of the manifold underlying the data. In a different direction, motivated by processing of data collected by sensor networks where sensors are irregularly placed, \cite{guestrinbodikthibauxpaskinmadden-ipsn2004} develops regression algorithms, \cite{ganesangreensteinestrinheidemanngovindan-2005,wagnerbaraniuketal-SSPWorkshop2005,wagnerbaraniuketal-IPSN06, Coifman:06,Hammond:11} wavelet decompositions, \cite{Narang:10,Narang:12} filter banks on graphs, \cite{wagnerdelouillebaraniuk-2006} de-noising, and~\cite{Zhu:12} compression schemes using the graph Laplacian. Some of these references consider distributed processing of data from sensor fields, while others study localized processing of signals on graphs in multiresolution fashion by representing data using wavelet-like bases with varying ``smoothness'' or defining transforms based on node neighborhoods. For example \cite{Hammond:11} uses the graph Laplacian and its eigenbasis to define a spectrum and a Fourier transform of a signal on a graph. This definition of a Fourier transform was also proposed for use in uncertainty analysis on graphs~\cite{Agaskar:12a,Agaskar:12}. Besides using the graph Laplacian, these works apply to data indexed by undirected graphs with real, non-negative edge weights. This approach is more fully developed in~\cite{ShumanNFOV:13}, which adopts the graph Laplacian as basic building block to develop GSP for data supported by undirected graphs. \subsection{Related work in other areas} \label{subsubsec:connectionsotherareas} } \NEW{ \subsubsection{Network science} This area addresses issues such as uncovering community relations, perceived alliances, quantifying connectedness, or determining the relevance of specific agents \cite{networksciencenas,borner2007network,lewis2011network,barabasi2016network}. Thus, much of this work does not concentrate on the data but rather its structure. \NEW{It determines for example} the size of the giant component, distribution of component sizes, degree and clique distributions, clustering coefficients, betweeness and closeness centralities, path length, and network diameter \cite{newman2010networks,jackson2010social,easley2010networks}. Connections to GSP are primarily due to graph spectra that GSP builds upon, \NEW{which is} strongly related to the structure of the graph \cite{Chung:96}. As an example, spectral clustering methods use the low frequency eigenvectors of the Laplacian \cite{von2007tutorial} and can thus be \NEW{addressed} from a GSP perspective \NEW{too} \cite{Tremblay:2016vb}. \subsubsection{Network processes} The aim is to model propagation over networks, including such phenomena as diffusion of diseases and epidemics, spread of (fake) news, memes, fads, voting trends, imitation and social influence, propagation of failures and blackouts. Common models are similar to stochastic automata where the states of the nodes (the ``data'') evolve through local rules, i.e., according to exogeneous (external to the network) and endogeneous (internal to the network) effects. For example, using terminology from epidemics, nodes of the graph representing agents or individuals of a population can be infected (adopt an opinion or spread a rumor), or susceptible (open to adopt an opinion or spread a rumor). Infected nodes can heal and become susceptible again; susceptible nodes can become infected either by an action external to the network or by action of infected neighbors \cite{castellano2012competing,nowzari2016analysis}. As the analysis of such network processes is difficult, traditionally, the network is abstracted out, assuming that any node can infect any other node (full mixing or complete network). To account for the impact of the network \cite{ganesh2005effect}, resorting to numerical studies is precluded except for very small networks since the network state space $\{0,1\}^N$ grows exponentially fast ($2^N$, for $N$ agents). To study these processes \cite{butts2009revisiting,colizza2006role} one usually considers one of two asymptotic regimes: (1) long term behavior (time-asymptotics), attempting to find the equilibrium distribution of the process \cite{zhang2014diffusion,zhang2015role}; or (2) large network asymptotics (mean-field approximation) \cite{draief2010epidemics} leading to the study of the qualitative behavior of nonlinear ordinary differential equations \cite{santos2015bi}. Because asymptotic behavior can be seen to depend on the eigenstructure of the underlying graph, GSP representations as those discussed in Section~\ref{sec:representations} can be used to characterize the evolution of a system. As an example, several papers have explored the use of GSP techniques to improve the efficiency of value function estimation in a reinforcement learning scenario \cite{levorato2012structure,levorato2012reduced}. \subsubsection{Graphical models} The focus in this area is on inference and learning from large datasets, \cite{lauritzen1996graphical,jordan1998learning,wainwright2008graphical,koller2009probabilistic,whittaker2009graphical}. The data is modeled as a set of random variables described by a family of Gibbs probability distributions, and the underlying graph \NEW{(whose nodes label the variables)} captures statistical dependence and conditional independence among the data. Acyclic graphs \cite{bang2008digraphs,edwards2012introduction} represent Bayesian networks, and undirected graphs represent Markov random fields \cite{rozanov1982markov,kindermann1980markov,rue2005gaussian}. Graphical models exploit factorizations of the joint distribution to develop efficient message passing algorithms for inference and find application in many areas such as modeling texture and other features in image processing \cite{besag1974spatial,chellappa1993markov,moura1992recursive,balram1993noncausal,willsky2002multiresolution,vats2012finding}, see \cite{jordan2010major} for illustrative applications in several domains. Recent work on learning graph from data \NEW{\cite{Dong:2016fm,egilmez2017graph}}, which makes use of Markov random field models to define optimality criteria for the learned graphs, connects graphical models to GSP. \subsection{Historical perspective on graph signal processing} \label{subsubsec:historicalperspective} We now briefly review some of the prior work that is more directly connected and in the spirit of signal processing on graphs, \cite{SandryhailaM:13,ShumanNFOV:13}. We organize the discussion along two main lines; \NEW{some parts of the exposition follow closely} \cite{SandryhailaM:13,Sandryhaila2014big}. \subsubsection{From algebraic signal processing to graph signal processing} The sequence of papers \cite{Pueschel:03a,Pueschel:05e,Pueschel:08a,Pueschel:08b,Pueschel:08c} introduced algebraic signal processing (ASP), an axiomatic approach to time signal processing. ASP starts from a signal model~$\Omega$. Many signal models are possible, and a relevant question is to determine which one is more appropriate for a given application or should be associated with a given linear transform. Under appropriate conditions, the signal model is generated from a simple filter, the \textit{shift}, which then determines filtering, convolution, the Fourier transform, frequency, and spectral analysis among other common concepts, and constructs from traditional digital signal processing. \NEW{Such formalism allowed for a uniform framework with variations of classical signal processing.} ASP, after appropriately defining a \textit{space} line-graph signal model \cite{Pueschel:08b}, can be used to show that the DCT plays the same role for that signal model as the one the DFT plays for the time (cyclic) model. ASP led to the introduction of, possibly weighted, graph adjacency matrices as shifts that generate the graph signal model for signals indexed by nodes of an arbitrary directed or undirected graph \cite{SandryhailaM:13,Sandryhaila:2014ju}. This choice is satisfying in the sense that, when the signal model is the classical time signal model, the shift and the graph signal model revert to the classical time shift (delay) and signal model~\cite{Pueschel:08a} (see Section~\ref{sec:asp-1}). Subsequently, authors have proposed other shifts obtained from the adjacency matrix of the graph~\cite{giraultgoncalvesfleury-2015,gavilizhang-2017} that attempt to preserve isometry of the shift, but in some cases lose the locality of the adjacency matrix shift \cite{giraultgoncalvesfleury-2015} \subsubsection{From graph Laplacian spectral clustering to Laplacian-based GSP} References \cite{Tenenbaum:00,Roweis:00,Belkin:03,Donoho:03} develop low-dimensional representations for large high-dimensional data through spectral graph theory~\cite{Belkin:02,Belkin:03} and the graph Laplacian~\cite{Chung:96}, by projecting the data on a low-dimensional subspace generated by a small subset of the Laplacian eigenbasis \cite{von2007tutorial}. The use of the graph Laplacian is justified by assuming the data is smooth on the data space (manifold). References \cite{Coifman:05a,Coifman:05b,Coifman:06} choose discrete approximations to other continuous operators, for example, a conjugate to an elliptic Schr{\"o}dinger-type operator, and obtain other spectral bases for the characterization of the geometry of the manifold underlying the data. \NEW{ Coming from another angle, motivated by processing data collected by sensor networks where sensors are irregularly placed, different authors develop regression algorithms \cite{guestrinbodikthibauxpaskinmadden-ipsn2004}, wavelet decompositions \cite{ganesangreensteinestrinheidemanngovindan-2005,wagnerbaraniuketal-SSPWorkshop2005,wagnerbaraniuketal-IPSN06, Coifman:06,Hammond:11}, filter banks on graphs \cite{Narang:10,Narang:12}, de-noising \cite{wagnerdelouillebaraniuk-2006}, and compression schemes using the graph Laplacian \cite{Zhu:12}. Some of these references consider distributed processing of data from sensor fields, while others study localized processing of signals on graphs in a multiresolution fashion by representing data using wavelet-like bases with varying ``smoothness'' or defining transforms based on node neighborhoods. For example \cite{Hammond:11} uses the graph Laplacian and its eigenbasis to define a spectrum and a Fourier transform of a signal on a graph. Besides using the graph Laplacian, these works apply to data indexed by undirected graphs with real, non-negative edge weights. This approach is more fully developed in~\cite{ShumanNFOV:13}, which adopts the graph Laplacian as basic building block to develop GSP for data supported by undirected graphs. } \NEW{\subsubsection{Image processing, computer graphics and GSP}In addition, graph-based approaches have been widely used in signal processing contexts. For example, several authors representing images as graphs for segmentation \cite{wu1993optimal,shi2000normalized} and popular image-dependent filtering methods can be interpreted from a graph perspective \cite{milanfar2013tour}. Models used in computer graphics applications can often be viewed as graphs (e.g., meshes where vertices form triangles to which attributes are associated) and graph based filtering, processing and multi-resolution representations \cite{guskov1999multiresolution,zhou20043d,taubin1995signal}.} \subsection{Outline of the paper} The outline of the paper is as follows: Section~\ref{sec:asp-1} starts by presenting the framework and key ingredients of graph signal processing. It explains how the concepts from classical signal processing such as signals, filters and Fourier transform, among others, extend to complex structures where data is indexed by nodes on a graph. Section~\ref{sec:soa} covers some state-of-the-art topics and associated challenges, such as the definition of frequency, graph learning, sampling representations and others. Section~\ref{sec:applications} follows up with applications of graph signal processing in sensor networks, biological networks, 3D point cloud processing and machine learning. Section~\ref{sec:conclusion} gives some conclusions. \subsection{The role of shifts in digital signal processing} \label{sec:shift-DSP} Discrete signal processing~(DSP) \cite{oppenheimschaffer-1975,oppenheimwillsky-1983,siebert-1986,oppenheimschaffer-1989,mitra-1998} studies time signals. Graph signal processing~(GSP)\footnote{We consider here only linear graph signal processing.} \cite{SandryhailaM:13,ShumanNFOV:13,Sandryhaila2014big} extends DSP to signal samples indexed by nodes of a graph. At a very high level, DSP, and therefore GSP, study: \begin{inparaenum}[1)] \item signals and their representations; \item systems that process signals, usually referred to as filters; \item signal transforms, including two very important ones, namely, the $z$-transform and the Fourier transform; and \item sampling of signals, as well as other more specialized topics. \end{inparaenum} Consider $N$ samples of a signal~$s_n$, $n=0,1,\cdots,N-1$. We restrict ourselves to signals with a finite number~$N$ of samples and to filters with finite impulse response (FIR filters). The $z$-transform~$s(z)$ of the (real or complex valued) time signal $s=\left\{s_n:n=0,1,\cdots,N-1\right\}$ organizes its samples~$s_n$ into an ordered set of time samples, where sample~$s_n$ at time~$n$ precedes~$s_{n+1}$ at time~$n+1$ and succeeds~$s_{n-1}$ at time~$n-1$. In other words, the signal is given by the $N$-tuple $s=\left(s_0, s_1,\cdots, s_{N-1}\right)$. This representation is achieved by using a formal variable, say~$z^{-1}$, called the shift (or delay), so that the signal of $N$-samples is represented by \begin{align} \label{eqn:ztransform-1} s(z)=\sum_{n=0}^{N-1}\,s_nz^{-n}. \end{align} The $z$-transform $s(z)$ provides a (formal)\footnote{\NEW{While in DSP $z$ is a complex variable, which leads to the DFT when restricted to the unit circle as in (\ref{eqn:signalDFT1}), here we establish the link to GSP by viewing $z$ as a placeholder for each sample of the signal.}} polynomial representation of the signal that is useful in studying how signals are processed by filters. Clearly, given $s(z)$ we can recover the signal~$s$ \cite{oppenheimschaffer-1989,mitra-1998}. The discrete Fourier transform (DFT) of the signal~$s$ is~$\widehat{s}=\left\{\widehat{s}_k:\,k=0,\cdots,N-1\right\}$ given by \begin{align} \label{eqn:signalDFT1} \widehat{s}_k=\frac{1}{\sqrt{N}}\sum_{n=0}^{N-1}\,s_ne^{-j\frac{2\pi}{N}kn}. \end{align} The $\widehat{s}_k$ are the Fourier coefficients of the signal. The DFT represents the signal~$s$ in the dual or frequency domain, leading to concepts such as frequency, spectrum, low-, band-, and high-pass signals. The discrete frequencies are \NEW{$\Omega_k= \frac{2\pi k}{N}$, } $k=0,1,\cdots, N-1$, and the $N$~signals $\left(x_k[n]\right)$ \begin{align*} \left\{x_k[n]=\frac{1}{\sqrt{N}}e^{-j\frac{2\pi}{N}kn}:\:n=0,1,\cdots, N-1\right\}_{k=0}^{N-1} \end{align*} are the spectral components. The signal is recovered from its Fourier coefficients by the inverse DFT: \begin{align} \label{eqn:signalinverseDFT1} s_n=\frac{1}{\sqrt{N}}\sum_{k=0}^{N-1}\,\widehat{s}_ke^{j\frac{2\pi}{N}kn}, s=0,1,\cdots, N-1. \end{align} In DSP, besides signals, we also have filters~$h$. An FIR filter is also represented by a polynomial in $z^{-1}$ \begin{align} \label{eqn:filterhz} h(z)=\sum_{n=0}^{N-1} h_nz^{-n}, \end{align} so that the output~$s_{\scriptsize\textrm{out}}$ of filter~$h$ applied to signal~$s_{\scriptsize\textrm{in}}$ is: \begin{align} \label{eqn:filter_output} s_{\scriptsize\textrm{out}}(z)=h(z)\cdot s_{\scriptsize\textrm{in}}(z). \end{align} Because we are only considering finite time signals, and the product above could result in $s_{\scriptsize\textrm{out}}(z)$ being a polynomial in $z^{-1}$ of degree greater than $N-1$, we have to consider boundary conditions (b.c.). For simplicity, we consider periodic extensions of the signal, i.e., the signal sample $s_N$ is equal to the signal sample $s_0$; more generally, $s_n=s_{n\!\!\!\mod\!\!N}$. In other words, the real line is folded around the circle. Defining the shift or delay filter \begin{align*} h_{\scriptsize\textrm{shift}}(z)=z^{-1}, \end{align*} and applying it to a signal $s_{\scriptsize\textrm{in}}=\left(s_0,s_1,\cdots,s_{N-1}\right)$ gives an output: \begin{align*} s_{\scriptsize\textrm{out}}=h_{\scriptsize\textrm{shift}}\cdot s_{\scriptsize\textrm{in}}=\left(s_{N-1},s_0, s_1,\cdots,s_{N-2}\right). \end{align*} By Equation~\eqref{eqn:filterhz}, any filter~$h$ in DSP is a polynomial in the shift, i.e., it is built from series and parallel combinations of shifts. Thus, the shift is the basic building block in DSP, from which we can build more complicated filters. A second very important DSP property that is adopted in GSP is \textit{shift invariance}. This readily follows from \begin{align} \label{eqn:DSPshiftinvariance} z^{-1}\cdot h(z)=h(z)\cdot z^{-1}. \end{align} In words, the series combination of filters is commutative, a filter commutes with the shift filter\textemdash delaying the input signal $s_{\scriptsize\textrm{in}}$ and then filtering the delayed input signal leads to the same signal as first filtering the input signal $s_{\scriptsize\textrm{in}}$ and then delaying the filtered output. \begin{cframed}[red] Restating for emphasis, both~\eqref{eqn:ztransform-1} and~\eqref{eqn:filterhz} show the principal role played by the shift~$z^{-1}$ in DSP. We represent signals by (finite degree) polynomials in $z^{-1}$ and build filters also as polynomials in $z^{-1}$. \end{cframed} \subsection{Defining shifts in Graph Signal Processing} \label{sec:shift-GSP} We now extend the above concepts and tools to \textit{graph} signals, i.e., signals whose samples are indexed by the nodes of arbitrary graphs. To do so, we start by reinterpreting the finite signals from the previous section as vectors rather than tuples or sequences. Rewrite the signal $s=\left(s_0, s_1,\cdots,s_{N-1}\right)$ as the vector \begin{align*} \bm{s}=\left[s_0\,s_1\,\cdots\,s_{N-1}\right]^\top\,\in\mathbb{C}^N, \end{align*} where for generality we allow the signal to be complex valued. Using this notation, a filter~$h$ is represented by a matrix~${\bf H}$ and (\ref{eqn:filter_output}) can be simply written as a matrix-vector multiplication: \begin{align*} \bm{s}_{\scriptsize\textrm{out}}={\bf H}\cdot \bm{s}_{\scriptsize\textrm{in}}, \end{align*} where filters are represented by matrices, while signals are represented by vectors. In particular, the shift filtering operation corresponds to multiplication by a \NEW{circulant} matrix~${\bf A}_c$ \begin{align*} \left[s_{N-1}\,s_0\,\cdots\,s_{N-2}\right]^\top= {\bf A}_c\cdot \left[s_0\,s_1\,\cdots\,s_{N-1}\right]^\top, \end{align*} given by the cyclic shift \begin{align} \label{eqn:Ashiftcyclic} {\bf A}_c=\left[\begin{array}{cccccc} 0&0&0&\cdots&0&1\\ 1&0&0&\cdots&0&0\\ 0&1&0&\cdots&0&0\\ \vdots&\vdots&\ddots&\ddots&\ddots&0\\ 0&0&\cdots&1&0&0\\ 0&0&\cdots&0&1&0 \end{array}\right]. \end{align} A graph interpretation for the DSP concepts of Section~\ref{sec:shift-DSP} can be achieved by viewing the 0-1 shift matrix ${\bf A}_c$ of~\eqref{eqn:Ashiftcyclic} as the adjacency matrix of a graph. Labeling the rows and columns of~${\bf A}_c$ from~0 to~$N-1$, define the graph $G_c=(V,E)$ with node set $V=\left\{0,1,\cdots, N-1\right\}$. \NEW{Row~$n$ of~${\bf A}_c$ represents the set of in-edges of node~$n$ in~$G_c$\textemdash if there is an entry~1 at column~$\ell$, $A_{c,n\ell}=1$, then there is an edge from~$\ell$ to~$n$. ${\bf A}_c$ is then the adjacency matrix of the cycle graph in Figure~\ref{fig:cyclegraph}.} \begin{figure}[htb] \centering \begin{tikzpicture}[scale=0.78] \node[draw=gray!70!black,circle,fill=red,line width=2pt] (0) at (0,0) {}; \node[draw=gray!70!black,circle,fill=red,line width=2pt] (1) at (2,0) {}; \node (2) at (4,0) {}; \node[draw=gray!70!black,circle,fill=red,line width=2pt] (N3) at (6,0) {}; \node[draw=gray!70!black,circle,fill=red,line width=2pt] (N2) at (8,0) {}; \node[draw=gray!70!black,circle,fill=red,line width=2pt] (N1) at (10,0) {}; \draw[-{Latex[length=4mm,width=6mm]},line width=3pt,draw=gray!70!black] (0) -- (1); \draw[-{Latex[length=4mm,width=6mm]},line width=3pt,draw=gray!70!black] (1)-- (2); \node[draw=gray!70!black,circle,fill=black,line width=1pt,inner sep=.8mm] (dot1) at (4.3,0) {}; \node[draw=gray!70!black,circle,fill=black,line width=1pt,inner sep=.8mm] (dot1) at (4.8,0) {}; \node[draw=gray!70!black,circle,fill=black,line width=1pt,inner sep=.8mm] (dot1) at (5.3,0) {}; \draw[-{Latex[length=4mm,width=6mm]},line width=3pt,draw=gray!70!black] (N3) -- (N2); \draw[-{Latex[length=4mm,width=6mm]},line width=3pt,draw=gray!70!black] (N2) -- (N1); \node (tmp) at (5,1.3) {}; \draw[-{Latex[length=4mm,width=6mm]},line width=3pt,draw=gray!70!black] (N1) to[out=140,in=0] (tmp) to[out=180,in=40] (0); \node[below = 2mm of 0] (0_id) {\bf $\mathbf{0}$}; \node[below = 2mm of 1] (1_id) {\bf $\mathbf{1}$}; \node[below = 2mm of N2] (N2_id) {\bf $\mathbf{N-2}$}; \node[below = 2mm of N1] (N1_id) {\bf $\mathbf{N-1}$}; \node[below = 1.5mm of 0_id] (0_sig) {\bf $\mathbf{s_0}$}; \node[below = 1.5mm of 1_id] (1_sig) {\bf $\mathbf{s_1}$}; \node[below = 1.5mm of N2_id] (N2_sig) {\bf $\mathbf{s_{N-2}}$}; \node[below = 1.5mm of N1_id] (N1_sig) {\bf $\mathbf{s_{N-1}}$}; \end{tikzpicture} \caption{Time graph: Cycle graph $G_c$.\label{fig:cyclegraph}} \end{figure} \begin{cframed}[blue] The key point we make is the dual role of the matrix~${\bf A}_c$ in Equation~\eqref{eqn:Ashiftcyclic}, which represents both the shift~$z^{-1}$ in DSP and the adjacency matrix of the associated time graph in Figure~\ref{fig:cyclegraph}. \end{cframed} This graph interpretation of DSP can be extended to develop a linear time shift invariant Graph Signal Processing \cite{SandryhailaM:13}. Consider now a graph signal $\bm{s}\in\mathbb{C}^N$, where the entries of the signal~$\bm{s}$ are indexed by the $N$~nodes of an arbitrary graph $G=(V,E)$, $v_1,...,v_N$. Assuming that the graph has edge weights $w_{ij}$, denote an edge of weight $w_{ij}$ going from $v_j$ to $v_i$, then we can define the following algebraic representations associated to $G$. \NEW{\begin{definition}[Algebraic representations of graphs] \label{def:useful_matrices} The \textbf{adjacency matrix} is a matrix, ${\bf A}$, such that $({\bf A})_{ij} = w_{ij}$. \\ In the particular case where the graph is undirected, we have $w_{ij}=w_{ji}$, \NEW{${\bf A}$ is now symmetric,} and we also define the \textbf{degree matrix} of $G$, a diagonal matrix, $\mathbf{D}$, with entries $(\mathbf{D})_{ii}= \sum_{j=1}^{N} ({\bf A})_{ij}$ and $(\mathbf{D})_{ij} = 0$ for $i \neq j$, the \textbf{combinatorial graph Laplacian} defined as ${\bf L} \!=\!\mathbf{D}\!-\!{\bf A}$, and the {\bf symmetric normalized Laplacian} ${\mathbf{\cal L}} = {\bf D}^{-1/2}{\bf L} {\bf D}^{-1/2} $. \end{definition}} The adjacency matrix~${\bf A}$ can be adopted as the shift~\cite{SandryhailaM:13} for this general graph. Other choices have been proposed, including the Laplacians \cite{ShumanNFOV:13}, or variations of these matrices \cite{giraultgoncalvesfleury-2015,gavilizhang-2017}. Different choices for the shift present different trade-offs. The adjacency matrix~${\bf A}$ reduces to the shift in classical time DSP and applies to directed and undirected graphs\footnote{\NEW{Note that the graph defined by ${\bf A}_c$ in (\ref{eqn:Ashiftcyclic}) is directed in order to match exactly the behavior of shifts in time in DSP, which are always directed, i.e., we either move forward or backwards in time. But in general the notion of a graph shift applies to any adjacency matrix, whether corresponding to a directed or an undirected graph. In what follows both directed and undirected graphs are considered.} }, while the graph Laplacian applies only to undirected graphs, \NEW{so that}~${\bf L}$ is symmetric and positive semi-definite, which avoids a certain number of analytical and numerical difficulties that may arise when choosing~${\bf A}$. Furthermore, graph Laplacian spectra have been widely studied in the field of spectral graph theory \cite{Chung:96}. In specific applications, one should consider definitions and choose the one that leads to the best trade-off for the problem being considered \cite{chen2015signal-2}. \NEW{This choice is further discussed in Sections~\ref{sec:selection} and~\ref{sec:frequency}}. For time signals, as discussed with respect to~\eqref{eqn:ztransform-1}, the basis $\left\{z^{-n}\right\}_{n=0}^{N-1}$ orders the samples of the signal by increasing order of the time labels (nodes in time graph). Rewriting~\eqref{eqn:ztransform-1}, we get \begin{align*} s(z){}&=\left[\left(z^{-1}\right)^0\,z^{-1}\cdots z^{-(N-1)}\right]\left[s_0\,s_1\cdots s_{N-1}\right]^\top. \end{align*} In graph signal processing, ordering the samples corresponds to labeling the nodes of the graph. This labeling or numbering fixes the adjacency matrix of the graph, and hence the graph shift. The columns of the graph shift provide a basis and a representation for the graph signals. Other bases could be used, leading to different signal representations. We note that relabeling the nodes of the graph by a permutation~$\hbox{\boldmath$\Pi$}$ conjugates the shift by~$\hbox{\boldmath$\Pi$}$ \begin{align*} {\bf A}_{\Pi}=\hbox{\boldmath$\Pi$} {\bf A}\hbox{\boldmath$\Pi$}^\top. \end{align*} Following the analogy with DSP, we can now define the notion of {\bf shift invariance} and {\bf polynomial filters} for arbitrary graphs. A filter represented by~${\bf H}$ will be shift invariant if it commutes with the shift, \begin{align*} {\bf A}{\bf H}={\bf H}{\bf A}. \end{align*} As proven in \cite{SandryhailaM:13}, if the characteristic polynomial $p_A(z)$ and the minimum polynomial\footnote{\NEW{For a matrix ${\bf A}$ the minimal polynomial $m_A(z)$ is the polynomial of minimal degree having ${\bf A}$ as a root. }} $m_A(z)$ of~${\bf A}$ are equal, then every filter commuting with~${\bf A}$ is a polynomial in~${\bf A}$, i.e., \begin{align*} {\bf H}=h({\bf A}). \end{align*} For equality $p_A(z)=m_A(z)$, to each eigenvalue of~${\bf A}$ there corresponds a single eigenvector\footnote{In other words, the Jordan form of~${\bf A}$ has single blocks for each distinct eigenvalue.}. A simpler condition is for the eigenvalues of~${\bf A}$ to be distinct. To keep the discussion simple, unless otherwise stated, we assume~${\bf A}$ has~$N$ distinct eigenvalues and hence a complete set of eigenvectors. By the Cayley-Hamilton Theorem of Linear Algebra \cite{gantmacher1959matrix,lancaster1985theory} \begin{align*} \textrm{degree}(h(z))=\textrm{degree}(p_A(z))\leq N-1. \end{align*} In fact, $\textrm{degree}(h(z))\leq \textrm{degree}(m_A(z))\leq \textrm{degree}(p_A(z))$. In words, shift invariant filters are polynomials with degree at most $\textrm{degree}(m_A(z))$. \subsection{Frequency representations for graph signals} \label{sec:frequency-GSP} In DSP and in linear systems, we are interested in signals that are invariant when processed by a (linear) filter, i.e., \begin{align*} h\cdot s_{\scriptsize{\textrm{in}}}=\alpha s_{\scriptsize{\textrm{in}}}, \end{align*} where $\alpha$ is a scalar (from the base field). Such $s_{\scriptsize{\textrm{in}}}$ are of course the eigensignals of the filter~$h$. In GSP we define filters as matrices and thus the eigensignals of~$h$ are the eigenvectors of the corresponding~${\bf H}$. More interestingly, since shift invariant filters are polynomials of a single matrix, the shift~${\bf A}$, we only need to consider the eigenvectors of~${\bf A}$. Then, write \begin{align} \label{eqn:diagonalizationA} {\bf A}={\bf V}{\bf \Lambda}{\bf V}^{-1} \end{align} where~${\bf V}=\left[\bm{v}_0\cdots\bm{v}_{N-1}\right]$ is the matrix of the~$N$ eigenvectors of~${\bf A}$, ${\bf \Lambda}=\textrm{diag}\left[\lambda_0\cdots \lambda_{N-1}\right]$ is the matrix of distinct eigenvalues of~${\bf A}$. \NEW{Because we assume~${\bf A}$ has a complete set of eigenvectors, ${\bf V}$ is invertible. } Then, it is straightforward to verify that for each (polynomial) filter \begin{align} \nonumber {\bf H}{}&=h({\bf A})\\ \nonumber {}&=h\left({\bf V}{\bf A}{\bf V}^{-1}\right)\\ \nonumber {}&=\sum_{m=0}^{M-1}h_m \left({\bf V}{\bf \Lambda}{\bf V}^{-1}\right)^m\\ \label{eqn:fiterresponse} {}&={\bf V} h\left({\bf \Lambda}\right){\bf V}^{-1}, \end{align} where $h\left({\bf \Lambda}\right)$ is the diagonal matrix \begin{align} \label{eqn:filterfrequencyresponse2} h\left({\bf \Lambda}\right)=\textrm{diag}\,\left[h\left(\lambda_0\right)\cdots h\left(\lambda_{N-1}\right)\right]. \end{align} We can promptly verify that the eigenvectors of~${\bf A}$ are the eigenfunctions of the (polynomial) filter \begin{align} \nonumber {\bf H} \bm{v}_m{}&={\bf V} h\left({\bf \Lambda}\right){\bf V}^{-1}\bm{v}_m\\ \nonumber {}&={\bf V} h\left({\bf \Lambda}\right)\bm{e}_m\\ \label{eqn:filtereigenresponse1} {}&=h\left(\lambda_m\right)\bm{v}_m, \end{align} where $\bm{e}_m$ is the zero vector except for entry~$m$ that is a one. Equation~(\ref{eqn:filtereigenresponse1}) is the GSP counterpart to the classical DSP fact that exponentials are eigenfunctions of linear systems. As such the response of the filter to an exponential is the same exponential amplified or attenuated by a gain that is the frequency response of the filter at the frequency of the exponential. We refer to this as the \textit{invariance} property of exponentials with respect to linear systems in DSP. Accordingly, Equation~\eqref{eqn:filtereigenresponse1} shows the invariance of the eigenvectors of the shift operator~${\bf A}$ with respect to graph filters. Finally, we can introduce the Fourier transform for graph signals. The cyclic shift in Equation~\eqref{eqn:Ashiftcyclic} can be written as \begin{align} \label{eqn:DFT1} {\bf A}_c={\bf DFT}_{N}^{-1}\left(\begin{matrix}e^{-j\frac{2\pi\cdot 0}{ N}}&&\cr&\ddots&\cr&&e^{-j\frac{2\pi\cdot(N-1)}{ N}}\end{matrix}\right){\bf DFT}_{N}, \end{align} where ${\bf DFT}_{N}=\frac{1}{\sqrt{N}}\left[\omega_N^{kn}\right]$, $\omega_N=\exp^{-j\frac{2\pi}{N}}$, is the discrete Fourier matrix. The inverse ${\bf DFT}_{N}^{-1}={\bf DFT}_{N}^H$ is the matrix of eigenvectors of~${\bf A}_c$. The eigenvalues of~${\bf A}_c$ are $e^{-j\frac{2\pi\cdot n}{ N}}$, $n=0,\cdots,N-1$. the diagonal entries of the middle matrix in Equation~\eqref{eqn:DFT1}. The graph Fourier transform (GFT) follows by analogy with~\eqref{eqn:DFT1}. From the eigendecomposition of~${\bf A}$ in~\eqref{eqn:diagonalizationA}, the graph Fourier transform is the inverse of the matrix~${\bf V}$ of eigenvectors of the shift~${\bf A}$ \begin{align} \label{eqn:graphFouruier1} {\bf F}={\bf V}^{-1}. \end{align} The eigenvectors of the shift~${\bf A}$, columns of~${\bf V}$, are the graph spectral components, and the eigenvalues of~${\bf A}$, the diagonal entries $\lambda_k$ of matrix~${\bf \Lambda}$ in~\eqref{eqn:diagonalizationA}, are the graph frequencies. The graph frequencies are complex valued for a general non-symmetric (directed graph) shift~${\bf A}$. The graph Fourier transform of graph signal~$\bm{s}$ is given by the graph Fourier \textit{analysis} decomposition \begin{align} \label{eqn:fourieranalysis1} \widehat{\bm{s}}={\bf F}\bm{s}={\bf V}^{-1} \bm{s}=\left[f_0\bm{s}\cdots f_{N-1}\bm{s}\right]^\top, \end{align} \NEW{where $f_k$ is \NEW{a row vector,} the $k$-th row of ${\bf F}$. The graph Fourier coefficients or graph spectral coefficients of signal~$\bm{s}$ are computed using the inner product as $\widehat{s}(\lambda_k) = \widehat{s}_k=f_k \bm{s} = \left\langle f_k^{H},\bm{s} \right\rangle$. Then, the Fourier spectral decomposition of the signal is obtained by the graph inverse Fourier transform. Equivalently, it is given by the graph Fourier synthesis expression \begin{align} \nonumber \bm{s}{}&={\bf F}^{-1}\widehat{\bm{s}}={\bf V} \widehat{\bm{s}}\\ \label{eqn:fouriersynthesis1} {}&=\sum_{k=0}^{N-1}\,\widehat{s}_k v_k\\ \nonumber {}&=\sum_{k=0}^{N-1}\,\left\langle f_k^{H},\bm{s} \right\rangle v_k\\ \nonumber {}&={\bf V}\left[\left\langle f_0^{H},\bm{s} \right\rangle\cdots \left\langle f_{N-1}^{H},\bm{s} \right\rangle\right]^\top. \end{align} } The eigenvectors $v_k$ of~${\bf A}$, columns of ${\bf V}$, are the spectral components. Equation~\eqref{eqn:fouriersynthesis1} synthesizes the original signal~$\bm{s}$ from the spectral components $v_k$; the coefficients $\widehat{s}_k$ of the decomposition are the spectral coefficients of~$\bm{s}$. \subsection{Interpreting Graph Frequencies} \label{sec:Frequency-interpretation} We can now interpret filtering a graph signal (i.e., multiplying the corresponding vector by ${\bf H}$) in the spectral domain. From~\eqref{eqn:fiterresponse}, the output of $\bf{s}_{\scriptsize\textrm{in}}$ to filter~$h$ is successively \begin{align} \nonumber \bm{s}_{\scriptsize\textrm{out}}{}&={\bf H}\cdot\bm{s}_{\scriptsize\textrm{in}}\\ \label{eqn:firstFourier1} {}&={\bf V} h\left({\bf \Lambda} \right)\underbrace{\left({\bf V}^{-1} \bm{s}_{\scriptsize\textrm{in}}\right)}_{\scriptsize\textrm{Fourier transf.}}\\ \nonumber {}&={\bf V}\underbrace{\textrm{diag}\left[h\left(\lambda_0\right)\cdots h\left(\lambda_{N-1}\right)\right] \widehat{\bm{s}}_{\scriptsize\textrm{in}}}_{\scriptsize\textrm{Filtering in graph Fourier space}}\\ \label{eqn:pointwisemult-1} {}&=\underbrace{{\bf V}\left[h\left(\lambda_0\right)\widehat{\bm{s}}_{{\scriptsize\textrm{in}}_0}\cdots h\left(\lambda_{N-1}\right)\widehat{\bm{s}}_{{\scriptsize\textrm{in}}_{N-1}}\right]^\top}_{\scriptsize\textrm{Inverse Fourier transf.}}. \end{align} Thus, according to ~\eqref{eqn:firstFourier1}, filtering by ${\bf H}$ can be performed by first taking the graph Fourier transform of the input $\left({\bf V}^{-1}\bm{s}_{\scriptsize\textrm{in}}\right)$, followed by pointwise multiplication in the frequency domain of the graph Fourier transform signal $\widehat{\bm{s}}_{\scriptsize\textrm{in}}$ by the \textit{filter frequency response} $\left[h\left(\lambda_0\right)\cdots h\left(\lambda_{N-1}\right)\right]^{\top}$ given by~\eqref{eqn:pointwisemult-1}. Finally, an inverse graph Fourier transform computes the output back in the graph node domain. This is the graph Fourier filtering Theorem that reduces graph filtering to two graph Fourier transforms and a pointwise multiplication in the spectral domain \cite{SandryhailaM:13}. With a notion of frequency we can now consider the GSP equivalents to classical concepts of low-, high-, and band-pass signals or filters, as well as the question of efficient filter design. In the classical time domain, these concepts are directly related to values of \textit{frequency}. \NEW{In the time domain, the} frequency is actually defined from the eigenvalues of the cyclic shift~${\bf A}_c$ in~\eqref{eqn:DFT1} as \begin{align*} \NEW{\Omega_k= \frac{2\pi k}{N} , \: k=0,1,\cdots,N-1.} \end{align*} These frequencies are directly related to the degree of variation of the spectral components. For example the lowest frequency $\Omega_0=0$ corresponds to the least varying spectral component, the constant or DC-spectral component, \NEW{the next frequency $\Omega_1=\frac{2\pi}{N}$ represents a higher variation spectral component, and so on. } There is a nice one-to-one correspondence between the ordered value of the frequency and the corresponding degree of variation or \textit{complexity} of the time spectral component. In GSP, the frequencies are defined by the eigenvalues of the shift. We can order the graph frequencies by relating them to the \textit{complexity} of the spectral component. For example, this can be measured by the \textit{total variation} of the associated spectral component through \begin{align*} \textrm{TV}_{\scriptsize\textrm{G}}\left(\bm{v}_k\right)= \left\|\bm{v}_{k}-{\bf A}^{\scriptsize\textrm{norm}}\bm{v}_{k}\right\|_1, \end{align*} where $\|\cdot\|_1$ is norm~1, and ${\bf A}^{\scriptsize\textrm{norm}}=\frac{1}{\lambda_{\scriptsize\max}}{\bf A}$. Other norms could be used to define the total variation, see \cite{Sandryhaila:2014ju}\cite{ShumanNFOV:13}. Using this, graph frequency $\lambda_m$ is larger than graph frequency $\lambda_\ell$ if \begin{align*} \textrm{TV}_{\scriptsize\textrm{G}}\left(\bm{v}_m\right)> \textrm{TV}_{\scriptsize\textrm{G}}\left(\bm{v}_\ell\right). \end{align*} Assuming the graph frequencies have been ordered from low to high, graph signal~$\bm{s}$ is low-pass if its graph Fourier coefficients are zero for $\Omega_k$, $k>\ell$, \NEW{for some $\ell$,} $0\leq \ell<N-1$. We can similarly define band- and high-pass signals and filters\footnote{\NEW{The total variation is the $l1$ norm of a vector multiplied by ${\bf I} - {\bf A}^{\scriptsize\textrm{norm}}$. Assume for simplicity that the graph is undirected, then the largest eigenvalue of ${\bf I} - {\bf A}^{\scriptsize\textrm{norm}}$, and thus the largest TV, will be $1-\lambda_{min}/\lambda_{max}$, where $\lambda_{min}$ is the smallest eigenvalue of ${\bf A}$, which intuitively, as seen in Fig.~\ref{fig:eigenvector_example}, corresponds to high variation in the eigenvector. }}. \begin{figure*}[h] \centering \includegraphics[width=0.95\linewidth]{figures/unweighted_A_eigvals}\par\medskip \mbox{\centering{\small(a)}}\par\medskip \includegraphics[width=0.95\linewidth]{figures/unweighted_L_eigvals}\par\medskip \mbox{\centering{\small(b)}} \caption{\NEW{\label{fig:eigenvector_example} Example of elementary frequencies obtained from different algebraic representations of the same graph. (a) Adjacency \NEW{matrix} (b) Laplacian matrix. In each case 4 different frequencies are shown, corresponding to different eigenvalues, ranging from lowest frequency to highest frequency. In the Laplacian case the lowest frequency is $\lambda=0$, representing a constant value throughout the graph, the highest is $\lambda=4.53$, where we can observe a large number of sign changes across graph edges. Note that for any given graph with $N$ nodes we will have $N$ eigenvectors that can be \NEW{ordered} in terms of their variation covering the whole range of frequencies for that graph. In this example the graph is unweighted. \NEW{Unlike} conventional signal processing some of the eigenvectors can be localized in the graph (e.g., the highest frequency eigenvector of the Laplacian).}} \end{figure*} \NEW{\subsection{Frequency representations based on the Laplacian} \label{sec:selection}} The notions of frequency that arise in conventional signal processing provide a sound mathematical and intuitive basis for analyzing signals. While it is mathematically possible, as just discussed, to define notions of frequency for graph signals, developing a corresponding intuition to understand these elementary frequencies is not as straightforward. For the total variation criterion it has been shown experimentally and justified theoretically that the frequency bases obtained from the shift operator tend to be ordered \cite{Sandryhaila:2014ju}. \NEW{Up to this point, we have focused primarily on frequency representations derived from the adjacency matrix of a graph, an approach that can be applied to both directed and undirected graphs, and can be linked to DSP concepts in the case of the cycle graph. A frequency representation can be similarly built on top of the Laplacian matrix of an undirected graph. Since this matrix is positive semidefinite, all the eigenvalues are real and non-negative, and a full set of orthogonal eigenvectors can be obtained, so that we can write \begin{equation} {\bf L} = {\bf U}{\bf \Lambda}{\bf U}^{\top} \end{equation} with ${\bf U}$ the GFT matrix, which is real valued and orthogonal in this case. Because the eigenvalues are real, they provide a natural way to order the GFT basis vectors in terms of frequency \NEW{(the variations of their values on the graph).} In this case, the eigenvalue/eigenvector pairs can be viewed as successive optimizers of the Rayleigh quotient, where the $k$-th pair, $\lambda_k, {\bf u}_k$ solves : \begin{equation} {\bf u}_k = \arg\min_{{\bf x}^{\top}{\bf u}_{k'} = 0,\, k'=0, \, \ldots, k-1}\frac{{\bf x}^{\top}{\bf L}{\bf x}}{{\bf x}^{\top}{\bf x}} \label{eq:rayleigh} \end{equation} with $\lambda_k = {\bf u}_{k}^{\top}{\bf L}{\bf u}_{{\bf u}_k}$, if ${\bf u}_k$ is normalized. Thus for the explicit variation metric induced by the Laplacian quadratic form, the GFT provides an orthogonal basis with increased variation, and such that, from (\ref{eq:rayleigh}), each additional basis vector minimizes the increase in variation while guaranteeing orthogonality.} More generally, the relationships between eigenvectors and eigenvalues of the Laplacian and the structure of a graph are part of a deep and beautiful domain of mathematics known as spectral graph theory~\cite{Chung:96}. When graphs have structures closely related to those used in DSP (e.g., circulant Adjacency matrices \NEW{\cite{ekambaram2015spline}}) frequency interpretation is clear. If the graph is more general than the ring graph, part of the intuition remains, as illustrated by Figure~\ref{fig:eigenvector_example}. Indeed, eigenvectors $u_i$ are oscillating over the vertex set. As the eigenvalue index $i$ increases, \NEW{the number of oscillations tends to increase as well~\cite{briandavies2001discrete}.} However, the irregular nature of graphs means that the analogies to DSP cannot always be extended easily. For example, the spacing between frequencies (as measured by the eigenvalues of the Laplacian, for example) can be highly irregular, or some frequencies may have high multiplicity. \NEW{Also, the high frequency eigenvectors of irregular graphs can be highly localized \cite{brooks2013non,saito2011phase}. This potentially indicates that a direct ordering of frequencies may be insufficient to fully understand signal decompositions induced by current GSP techniques. \NEW{To be complete, note that,} while Laplacians can be easily defined for undirected graphs, there has been work to introduce definitions appropriate for directed graphs as well \cite{chung2005laplacians,bauer2012normalized}.} \NEW{In summary, a full understanding of the best frequency representation for a specific GSP application, as a function of the type of graph considered, is still an active research topic. This is discussed further in Section~\ref{sec:frequency}.} \subsection{Implementation} Finally, let us quickly touch on the issue of computational complexity of the filtering operation. A straightforward algorithm would consist in computing the GFT matrix ${\bf V}$ and explicitly applying it to the input signal as in (\ref{eqn:firstFourier1}). This is simple and accurate for small graphs thanks to fast SVD algorithms. Partial SVD can also be used if the filter $h$ should only be evaluated on the top or bottom eigenvalues \NEW{\cite{golub2012matrix}}. In general, and for large graphs, it is better to avoid computing even a partial SVD. One efficient possibility is to compute a polynomial approximation to $h$ with Chebyshev filters~\cite{Hammond:11}. For large but sparse graphs, this reduces computations to sparse matrix-vector multiply, which is very efficient. \NEW{Furthermore, filter implementation via polynomial approximation can be interpreted in terms of localization in the vertex domain. } Note that when the input signal is a perfect impulse located at a given vertex, ${\bf s} = {\bf e}_i$, the filtered signal depends only on the graph filter and the vertex location in the graph: ${\bf f}_i = {\bf H} {\bf e}_i$. Even though ${\bf f}_i$ changes with the chosen vertex, it was proved in \cite{Localization} that this signal is localized around $i$ in a way that only depends on the smoothness of the filter $h$. This is interesting because it allows to design filters that act locally and in a controlled way over the vertex set. \NEW{After a filter $h(\lambda)$ is chosen, one can choose an approximation $h_k(\lambda)$, a polynomial of degree $k$ in $\lambda$. Note that $h_k(\lambda)$ can then be implemented as shown in (\ref{eqn:fiterresponse}) by applying a polynomial of the shift operator. This does not require knowing the eigenvalues and eigenvectors associated to the graph, so that it is possible to process signals on very large graphs locally, by processing $k$-hop neighborhoods of nodes in the vertex domain, without a need to find the graph spectrum first.} \subsection{Frequency definition} \label{sec:frequency} One can guarantee the existence of an orthogonal basis for any undirected graph. Thus, once a graph has been chosen (see Section~\ref{sec:graph-learning}) a definition of frequency is readily available, which allows us to address other questions considered in this section (sampling, signal representation, etc). Multiple choices are possible, as a function of the graph type, the selected shift operator and its normalization, etc. Making these choices appropriately for a given application remains an open question, which is actively being investigated. As an example, the eigenvalues of the chosen operator matrix (Laplacian or adjacency) can have high multiplicity. In this situation, a graph with $N$ nodes will have fewer than $N$ unique frequencies. A particular concern is that one can choose any set of orthogonal vectors within the subspace corresponding to this frequency, leading to different GFTs and thus potentially irreproducible results. As a way to address this scenario, recent work \cite{deri2017spectral} suggests using oblique projections to measure the energy within such a subspace, using this information to represent the overall energy at that frequency. For directed graphs, additional problems arise given that a full set of eigenvectors may not exist. Results for directed graphs are often restricted to cases where the adjacency matrix is invertible and eigenvectors do exist (as discussed in Section~\ref{sec:frequency-GSP}). If these conditions do not hold, the Jordan canonical form is used to obtain the GFT \cite{SandryhailaM:13}, but this is well known to be a numerically unstable procedure. As an alternative, some authors have proposed to approximate directed graphs by undirected ones, using such approaches as the hub-authority model \cite{kleinberg1999authoritative,zhou2005semi}. \NEW{ Recent work has also considered alternative definitions of frequency. For example, the work in \cite{mhaskar2016unified} advocates using the random walk Laplacian normalization, while in \cite{girault2018irregularity} the authors propose alternative choices of a graph signal inner product and explore the resulting frequency definitions. Other techniques make use of explicit optimization to choose a set of graph frequencies. } As an example, the work in \cite{sardellitti2017graph} uses an optimization procedure to construct explicitly an orthogonal basis set that minimizes a quantity related to the cut size. With this approach, successive eigenvectors provide increasingly higher frequencies in the sense of corresponding to higher cut costs, while being orthogonal to those eigenvectors previously selected. \NEW{The work in \cite{shafipour2017digraph} also uses optimization techniques with a different criterion to define a set of frequencies associated to a graph. } In summary, this is a very active area of research, and the best approach to define a set of frequencies for graphs in a specific application remains to some extent an open question. \subsection{Representations} \label{sec:representations} Designing representations for graph signals having desirable properties (e.g., localization, critical sampling, orthogonality, etc) has been one of the first and most important research goals in graph signal processing. Pioneering contributions \cite{Crovella2003} and \cite{Coifman:06}, provided early examples of designs based on vertex domain and spectral domain characteristics, respectively. Vertex domain designs such as \cite{Crovella2003} or \cite{narang2009lifting} have the advantage of defining exactly localized basis functions on the graph, but do not have a clear spectral interpretation. Conversely, diffusion wavelets \cite{Coifman:06} are defined in the spectral domain, but do not guarantee exact vertex domain localization (only energy decay properties). The spectral graph wavelet transform design \cite{Hammond:11} was the first to combine a spectral design with vertex-domain localization, by defining smooth filter kernels in the spectral domain and approximating these with polynomials. The filterbanks developed in \cite{Hammond:11} were not critically sampled, unlike \cite{Coifman:06} or \cite{narang2009lifting}. Thus, much recent work has focused on developing critically sampled filterbanks having both a spectral interpretation and vertex localized implementation. These types of filterbanks have been designed for bipartite graphs \cite{Narang:12,narang2013compact}, thus requiring the graph to be decomposed into a series of bipartite subgraphs \cite{Narang:12,zeng2017bipartite}. \NEW{An alternative approach proposed in \cite{kotzagiannidis2017splines,ekambaram2015spline} can be applied to circulant graphs, for which the GFT corresponds to the DFT.} Recent work \cite{teke2017extending1,teke2017extending2} has shown that similar filterbank designs can be developed for directed graphs, where these designs are only possible for $M$-block cyclic graphs, which play a similar role to that of bipartite graphs in the undirected case. \NEW{Note that in all these cases, critical sampling combined with polynomial analysis and synthesis filtering is restricted to specific types of graphs (bipartite, $M$-block cyclic and circulant.) Note also that critical sampling with polynomial analysis and synthesis filters on undirected graphs can only be achieved in the bipartite case \cite{anis2017critical}\footnote{Under some conditions on the analysis filters, critical sampling and perfect reconstruction can be achieved for any graph, but this requires a synthesis operation corresponding to an $N \times N$ matrix multiplication, which may not be practical for large graphs \cite{anis2017critical}. As an example, the approach in \cite{ekambaram2015spline} guarantees invertibility but reconstruction is non polynomial.}. } Ongoing work is focusing on i) providing better tools to characterize $M$-block cyclic graphs, including for example the definition of polyphase representations \cite{teke2017extending1,teke2017extending2,tay2015techniques,tay2017bipartite}, ii) development of improved filters by exploiting conventional filter designs and/or relaxing the critical sampling requirement \cite{tanaka2014m,sakiyama2014oversampled,tay2015design,tay2017critically}, \NEW{ and iii) novel approaches for downsampling, e.g., frequency domain techniques \cite{tanaka2017spectral}, that allow extending critically sampled filterbanks to non-bipartite graphs.} While much of the work to date has focused on representations with bases functions selected in terms of frequency content (e.g., low pass vs.~high pass bases), some recent work is also exploring representations for piecewise smooth signal models \cite{ChenSK:17}. The design of representations that adapt to the specific properties of graph signal classes has further been addressed from the viewpoint of dictionary learning \cite{Zhang:2012fu,Thanou:2014gj,Yankelevsky:2016gka}. The main objective is to design dictionary of atoms that are able to sparsely represent signals on graphs while incorporating the structure of the graph. \subsection{Sampling} \label{sec:sampling} The problem of sampling signals on graphs is modeled on the corresponding problem in conventional signal processing. The basic idea is to define a class of signals (for example signals that are bandlimited to the first $K$ frequencies of the GFT) and then define necessary and sufficient conditions to reconstruct a signal in that class from its samples. The first problem formulation and a sufficient condition for unique recovery were presented in~\cite{pesenson2008sampling}. A necessary and sufficient condition for unique recovery in undirected graphs was introduced in \cite{anis2014towards}, and subsequently several papers proposed solutions for different aspects of the problem \cite{shomorony2014sampling,ChenVSK:15,anis2016efficient}. In particular, sampling results have been generalized to directed graphs \cite{ChenVSK:15,anis2016efficient} and to other classes of signals such as piecewise smooth signals \cite{chen2015signal}. A key difference when comparing sampling in conventional signal processing and in the context of graph signals is the lack of ``regular'' sampling patterns in the latter. The lack of regularity in the graph itself prevents us from defining the idea of sampling ``every other node''. Thus, multiple approaches have been suggested to identify the most informative vertices on a graph so that these can be sampled. While the sampling problem is formalized based on the assumption that signals to be sampled belong to a certain class (e.g., bandlimited), in practice these can never be guaranteed and thus the observed signals will be noisy and in general will not belong to the pre-specified class. To address this problem, several methods approach the problem of sampling set selection from an experiment design perspective \cite{gadde2015probabilistic,ChenVSK:15,anis2016efficient} setting as a goal to identify a set of vertices that minimizes some measure of worst case reconstruction error in cases where noise or model mismatch is present. The measure can also be mean squared reconstruction error instead of worst case in the experiment design paradigm \cite{anis2016efficient}. Complexity is a key challenge in sampling set identification, especially for large-scale graphs. Some techniques require computing and storing the first $K$ basis vectors of the GFT \cite{ChenVSK:15}. For larger graph sizes, where this may not be practical, the approach in \cite{anis2016efficient} uses spectral proxies instead of exact graph frequencies leading to lower complexity. To reduce complexity even further, the work in \cite{puy2016random} proposes a random sampling technique where the probability of selecting a given vertex is based on a locally computed metric. This leads to significantly lower complexity but, as a random sampling technique, it may not always lead to performance comparable to those of more complex greedy optimization methods such as \cite{ChenVSK:15,anis2016efficient}. Given the samples of a graph signal, the next objective is to reconstruct an estimation of the signal at the nodes that were not sampled (observed). Reconstruction algorithms based on polynomial filters approximating ideal reconstruction filters have been proposed in order to reconstruct an estimated signal on the whole graph based on the observed vertex measurements \cite{narang2013signal,wang2015local}. While theoretical aspects of graph signal sampling are by now well understood, the relevance of proposed techniques to practical applications is still an open question. A key challenge in this regard is to identify what are relevant signal models for real datasets, while potentially adapting proposed generic sampling methods to specific types of graphs (e.g., exploiting properties of nearly regular graphs). \subsection{Extending conventional signal processing to graph signals} Challenges in extending ideas and concepts from conventional signal processing to signal processing on graphs can be further exemplified by research into notions of stationarity and localization. For conventional time signals, a test for stationarity can be based on determining whether time shifts affect the statistical properties of a signal or, equivalently, observing a signal at different times. However, these two views are not equivalent for finite dimension graphs: we can observe a given signal at different nodes, but this is not necessarily the same as ``shifting'' the signal while observing it at always at the same node. For graphs with $N$ vertices, shifting can be defined via a spectral domain operator \cite{Hammond:11}; or, instead, the graph shift based on the adjacency matrix can be used. Some authors have proposed a definition of stationarity based on spectral properties of the vertex shift operator \cite{marques2016stationary}. To overcome challenges associated to existing shift operators, one solution, first proposed by \cite{girault2015stationary}, is to introduce alternative graph shift operators (see also \cite{giraultgoncalvesfleury-2015}) or localization operators that have both a spectral interpretation and vertex domain localization \cite{girault2016localization,perraudin2017stationary}. \NEW{Notions of stationarity can help develop probabilistic graph signal processing \NEW{methods} leading to graph-based Wiener filtering \cite{girault2014semi,perraudin2017stationary}.} A study of vertex/spectral localization and uncertainty principles was first developed by \cite{Agaskar:12}, where it was shown that in general it is not possible to achieve arbitrarily good localization in both spectral and vertex domains simultaneously. However, a limitation in this study was that bounds had to be derived for individual vertices. More recently, \cite{tsitsvero2016signals} has shown that for graph signals it is in fact possible to have compact support in both spectral and vertex domain (something that can never occur in conventional signal processing). \NEW{As was already noted in Section~\ref{sec:selection}}, this occurs due to the irregular nature of graphs: for example, a graph consisting of several loosely connected clusters is likely to lead to some columns of ${\bf V}$ having non-zero entries only in some of the clusters. Other contributions, such as \cite{pasdeloup2015toward,teke2017uncertainty}, have also explored the challenges in directly extending the concept of an uncertainty principle to graph signals, \NEW{while other recent work considers alternative frequency representations that can take into consideration the specific localization properties encountered in irregular graphs \cite{behjat2016signal,van2017slepian,irion2014hierarchical} } Work in these two areas shows that direct extensions of signal processing concepts to graphs are not straightforward, and thus further research is still needed to develop techniques that can provide insights about graph signal behavior (localization, stationarity) while accommodating key characteristics of graphs (e.g., irregular node connectivity and spectral characteristics). \subsection{Graph learning} \label{sec:graph-learning} Much recent work on graph signal processing assumes that the graph is given or can be defined in a reasonable way based on the nature of the application. As an example, in communication or social networks the connectivity of the network (directed or undirected) can be used to define the graph. In other applications edge weights between nodes can be chosen as a decreasing function of distance, e.g., physical distance between sensors in the case of sensor networks or distance in feature space in the case of learning applications \cite{zhu2003semi,jebara2009graph,gadde2014active}. Recent work has been considering alternative techniques where the goal is to learn graphs from data. This work is motivated by scenarios where i) no reasonable initial graph exists or ii) it is desirable to modify a known graph (based on network connectivity for example) by selecting weights derived from data. The key idea in these approaches is to select a graph such that the most likely vectors in the data (the graph signals) correspond to the lowest frequencies of the GFT or to the more likely signals generated by Gauss Markov Random Field (GMRF) related to the graph. Examples of approaches based on smoothness include \cite{Dong:2016fm,Kalofolias:2016tf,daitch2009fitting}, while representative methods based on the GMRF model are \cite{lake2010discovering,egilmez2017graph}. The basic idea in the latter approaches is to identify GMRF models such that the inverse covariance (precision) matrix has the form of a graph Laplacian (e.g., combinatorial or generalized). Note that this work extends popular approaches for graph learning (e.g., graphical Lasso \cite{friedman2008sparse}) to precision matrices restricted to have a Laplacian form (corresponding to a graph with positive edge weights). Other approaches have addressed graph selection under the assumption that the observed data was obtained through graph-based diffusion. Examples of these approaches include \cite{mei2017signal,pasdeloup2017characterization,Segarra_templates,Thanou:2017ec}. \NEW{While not explicitly a graph learning problem, the related question of blind identification of graph filters has also been studied \cite{segarra2017blind}.} \NEW{There remain several major challenges in the development of graph learning methods. Graphs derived from data are essentially models, and as such the ``right'' graph model should be selected based on the number of parameters it uses, its data fit and its ability to provide useful interpretations. While a sparsity criterion addresses some of these requirements, other constraints may also be important. For example, it will be useful to develop methods to select graphs with specific topology properties~\cite{pavez2017learning}, spectral properties (eigenvalue distribution, eigenvector localization), or even computational properties (e.g., leading to GFTs with lower computation cost.) }

https://arxiv.org/abs/1805.02363

Planning and Learning with Stochastic Action Sets

In many practical uses of reinforcement learning (RL) the set of actions available at a given state is a random variable, with realizations governed by an exogenous stochastic process. Somewhat surprisingly, the foundations for such sequential decision processes have been unaddressed. In this work, we formalize and investigate MDPs with stochastic action sets (SAS-MDPs) to provide these foundations. We show that optimal policies and value functions in this model have a structure that admits a compact representation. From an RL perspective, we show that Q-learning with sampled action sets is sound. In model-based settings, we consider two important special cases: when individual actions are available with independent probabilities; and a sampling-based model for unknown distributions. We develop poly-time value and policy iteration methods for both cases; and in the first, we offer a poly-time linear programming solution.

\section{Introduction} \label{sec:intro} Markov decision processes (MDPs) are the standard model for sequential decision making under uncertainty, and provide the foundations for reinforcement learning (RL). With the recent emergence of RL as a practical AI technology in combination with deep learning \cite{mnih2013,mnih2015}, new use cases are arising that challenge basic MDP modeling assumptions. One such challenge is that many practical MDP and RL problems have \emph{stochastic sets of feasible actions}; that is, the set $A_s$ of feasible actions at state $s$ \emph{varies stochastically} with each visit to $s$. For instance, in online advertising, the set of available ads differs at distinct occurrences of the same state (e.g., same query, user, contextual features), due to exogenous factors like campaign expiration or budget throttling. In recommender systems with large item spaces, often a set of \emph{candidate} recommendations is first generated, from which top scoring items are chosen; exogenous factors often induce non-trivial changes in the candidate set. With the recent application of MDP and RL models in ad serving and recommendation \cite{charikar:stoc99,Li:adkdd2009,archak-mirrokni-muthu:www10,mirrokni:wine12,kearns:uai12,silver:icml13,theocharous:ijcai15,logisticMDPs:ijcai17}, understanding how to capture the stochastic nature of available action sets is critical. Somewhat surprisingly, this problem seems to have been largely unaddressed in the literature. Standard MDP formulations \cite{puterman} allow each state $s$ to have its own feasible action set $A_s$, and it is not uncommon to allow the set $A_s$ to be non-stationary or time-dependent. However, they do not support the treatment of $A_s$ as a stochastic random variable. In this work, we: (a) introduce the \emph{stochastic action set MDP (SAS-MDP)} and provide its theoretical foundations; (b) describe how to account for stochastic action sets in model-free RL (e.g., Q-learning); and (c) develop tractable algorithms for solving SAS-MDPs in important special cases. An obvious way to treat this problem is to embed the set of available actions into the state itself. This provides a useful analytical tool, but it does not immediately provide tractable algorithms for learning and optimization, since each state is augmented with all possible \emph{subsets} of actions, incurring an exponential blow up in state space size. To address this issue, we show that SAS-MDPs possess an important property: the Q-value of an available action $a$ is independent of the availability of other actions. This allows us to prove that optimal policies can be represented compactly using (state-specific) decision lists (or orderings) over the action set. This special structure allows one to solve the SAS RL problem effectively using, for example, Q-learning. We also devise model-based algorithms that exploit this policy structure. We develop value and policy iteration schemes, showing they converge in a polynomial number of iterations (w.r.t.\ the size of the underlying ``base'' MDP). We also show that per-iteration complexity is polynomial time for two important special forms of action availability distribution: (a) when action availabilities are independent, both methods are exact; (b) when the distribution over sets $A_s$ is sampleable, we obtain approximation algorithms with polynomial sample complexity. In fact, policy iteration is strongly polynomial under additional assumptions (for a fixed discount factor). We show that a linear program for SAS-MDPs can be solved in polynomial time as well. Finally, we offer a simple empirical demonstration of the importance of accounting for stochastic action availability when computing an MDP policy. Additional discussion and full proofs of all results can be found in a longer version of this paper \cite{sasmdps_full:arxiv18}. \section{MDPs with Stochastic Action Sets} \label{sec:sas} We first introduce SAS-MDPs and provide a simple example illustrating how action availability impacts optimal decisions. See \cite{puterman} for more background on MDPs. \subsection{The SAS-MDP Model} \label{sec:sasmodel} Our formulation of \emph{MDPs with Stochastic Action Sets (SAS-MDPs)} derives from a standard, finite-state, finite-action MDP (the \emph{base MDP}) $\mathcal{M}$, with $n$ states $S$, \emph{base} actions $B_s$ for $s\in S$, and transition and reward functions, $P: S \times B \rightarrow \Delta(S)$ and $r:S\times B \rightarrow \mathbb{R}$. We use $p^k_{s,s'}$ and $r^k_s$ to denote the probability of transition to $s'$ and the accrued reward, respectively, when action $k$ is taken at state $s$. For notational ease, we assume that feasible action sets for each $s\in S$ are identical, so $B_s = B$ (allowing distinct base sets at different states has no impact on what follows). Let $|B| = m$ and $M = |S \times B | =nm$. We assume an infinite-horizon, discounted objective with fixed discount rate $\gamma$, $0\leq\gamma<1$. In a SAS-MDP, the set of actions available at state $s$ at any stage $t$ is a random subset $A^{(t)}_s \subseteq B$. We assume a family of \emph{action availability distributions} $P_s\in\Delta(2^B)$ defined over the powerset of $B$. These can depend on $s\in S$ but are otherwise history-independent, hence $\Pr(A^{(t)}_s | s^{(1)},\ldots,s^{(t)}) = \Pr(A^{(t)}_s | s^{(t)})$. Only actions $k\in A^{(t)}_s$ in the realized available action set can be executed at stage $t$. Apart from this, the dynamics of the MDP is unchanged: when an (available) action is taken, state transitions and rewards are prescribed as in the base MDP. In what follows, we assume that some action is always available, i.e., $\Pr(A^{(t)}_s = \emptyset) = 0$ for all $s, t$.\footnote{Models that trigger process termination when $A^{(t)}_s = \emptyset$ are well-defined, but we set aside this model variant here.} Note that a SAS-MDP does not conform to the usual definition of an MDP. \subsection{Example} \label{sec:sasexample} The following simple MDP shows the importance of accounting for stochastic action availability when making decisions. The MDP below has two states. Assume the agent starts at state $s_1$, where two actions (indicated by directed edges for their transitions) are always available: one ($\mathit{Stay}$) stays at $s_1$, and the other ($\mathit{Go}$) transitions to state $s_2$, both with reward $1/2$. At $s_2$, the action $\mathit{Down}$ returns to $s_1$, is always available and has reward 0. A second action $\mathit{Up}$ also returns to $s_1$, but is available with only probability $p$ and has reward 1. \includegraphics[width=0.5\columnwidth]{toy_model_singlefig_labeled.pdf} A naive solution that ignores action availability is as follows: we first compute the optimal $Q$-function assuming all actions are available (this can be derived from the optimal value function, computed using standard techniques). Then at each stage, we use the best action available at the current state where actions are ranked by Q-value. Unfortunately, this leads to a suboptimal policy when the $\mathit{Up}$ action has low availability, specifically if $p < 0.5$. The best naive policy always chooses to move to $s_2$ from $s_1$; at $s_2$, it picks the best action available. This yields a reward of $1/2$ at even stages, and an expected reward of $p$ at odd stages. However, by anticipating the possibility that action $\mathit{Up}$ is unavailable at $s_2$, the optimal (SAS) policy always stays at $s_1$, obtaining reward $1/2$ at all stages. For $p < 1/2$, the latter policy dominates the former: the plot on the right shows the fraction of the optimal (SAS) value \emph{lost} by the naive policy ($Std$) as a function of the availability probability $p$. This example also illustrates that as action availability probabilities approach $1$, the optimal policy for the base MDP is also optimal for the SAS-MDP. \subsection{Related Work} \label{sec:related} While a general formulation of MDPs with stochastic action availability does not appear in the literature, there are two strands of closely related work. In the bandits literature, \emph{sleeping bandits} are defined as bandit problems in which the arms available at each stage are determined randomly or adversarially (sleeping experts are similar, with complete feedback being provided rather than bandit feedback) \cite{kleinbergEtAl:MLJ2010,kanadeEtAl:aistats09Sleeping}. Best action orderings (analogous to our decision list policies for SAS-MDPs) are often used to define regret in these models. The goal is to develop exploration policies to minimize regret. Since these models have no state, if the action reward distributions are known, the optimal policy is trivial: always take the best \emph{available} action. By contrast, a SAS-MDP, even a known model, induces a difficult optimization problem, since the quality of an action depends not just on its immediate reward, but also on the availability of actions at reachable (future) states. This is our focus. The second closely related branch of research comes from the field of stochastic routing. The ``Canadian Traveller Problem''---the problem of minimizing travel time in a graph with unavailable edges---was introduced by Papadimitriou and Yannakakis~\cite{papadimitriou:shortestpath}, who gave intractability results (under much weaker assumptions about edge availability, e.g. adversarial). Poliyhondrous and Tsitsiklis~\cite{polychondrous:recourse} consider a stochastic version of the problem, where edge availabilities are random but static (and any edge observed to be unavailable remains so throughout the scenario). Most similar to our setting is the work of Nikolova and Karger~\cite{nikolova:canadian}, who discuss the case of resampling edge costs at each node visit; however, the proposed solution is well-defined only when the edge costs are finite and does not easily extend to unavailable actions/infinite edge costs. Due to the specificity of their modeling assumptions, none of the solutions found in this line of research can be adapted in a straightforward way to SAS-MDPs. \section{Two Reformulations of SAS-MDPs} \label{sec:reformulate} The randomness of feasible actions means that SAS-MDPs do not conform to the usual definition of an MDP. In this section, we develop two reformulations of SAS-MDPs that transform them into MDPs. We discuss the relative advantages of each, outline key properties and relationships between these models, and describe important special cases of the SAS-MDP model itself. \subsection{The Embedded MDP} \label{sec:embedded} We first consider a reformulation of the SAS-MDP in which we embed the (realized) available action set into the state space itself. This is a straightforward way to recover a standard MDP. The \emph{embedded MDP} $\mathcal{M}_e$ for a SAS-MDP has state space $S_e = \{s\circ A: s\in S, A\subseteq B\}$, with $s\circ A$ having feasible action set $A$.\footnote{Embedded states whose embedded action subsets have zero probability are unreachable and can be ignored.} The history independence of $P_s$ allows transitions to be defined as: $$p^k_{s\circ A,s'\circ A'} = P(s'\circ A' | s\circ A, k) = p^k_{s,s'} P_{s'}(A') , \;\; \forall k\in A.$$ Rewards are defined similarly: $r^k(s\circ A) = r^k(s)$ for $k\in A$. In our earlier example, the embedded MDP has three states: $s_1\circ\{\mathit{Stay},\mathit{Go}\}, s_2\circ\{\mathit{Up},\mathit{Down}\}, s_2\circ\{\mathit{Down}\}$ (other action subsets have probability $0$ hence their corresponding embedded states are unreachable). The feasible actions at each state are given by the embedded action set, and the only stochastic transition occurs when $\mathit{Go}$ is taken at $s_1$: it moves to $s_2\circ\{\mathit{Up},\mathit{Down}\}$ with probability $p$ and $s_2\circ\{\mathit{Down}\}$ with probability $1-p$. Clearly, the induced reward process and dynamics are Markovian, hence $\mathcal{M}_e$ is in fact an MDP under the usual definition. Given the natural translation afforded by the embedded MDP, we view this as providing the basic ``semantic'' underpinnings of the SAS-MDP model. This translation affords the use of standard MDP analytical tools and methods. A (stationary, determinstic, Markovian) policy $\pi:S_e \rightarrow B$ for $\mathcal{M}_e$ is restricted so that $\pi(s\circ A) \in A$. The policy backup operator $T^\pi_e$ and Bellman operator $T^\ast_e$ for $\mathcal{M}_e$ decompose naturally as follows: \begin{small} \begin{align} &T^\pi_e V_e(s\circ A_s) = r^{\pi(s\circ A_s)}_s + \nonumber \\ &\qquad \gamma\sum_{s'}p^{\pi(s\circ A_s)}_{s,s'} \sum_{A_{s'}\subseteq B} P_{s'}(A_{s'})V_e(s'\circ A_{s'}) \label{eq:embeddedpolicybackup}, \\ &T^\ast_e V_e(s\circ A_s) = \max_{k\in A_s} r^k_s + \nonumber \\ &\qquad \gamma\sum_{s'}p^k_{s,s'} \sum_{A_{s'}\subseteq B} P_{s'}(A_{s'})V_e(s'\circ A_{s'}) \label{eq:embeddedbellmanbackup} \end{align} \end{small}% Their fixed points, $V^\pi_e$ and $V^\ast_e$ respectively, can be expressed similarly. Obtaining an MDP from an SAS-MDP via action-set embedding comes at the expense of a (generally) exponential blow-up in the size of the state space, which can increase by a factor of $2^{|B|}$. \subsection{The Compressed MDP} \label{sec:compressed} The embedded MDP provides a natural semantics for SAS-MDPs, but is problematic from an algorithmic and learning perspective given the state space blow-up. Fortunately, the history independence of the availability distributions gives rise to an effective, compressed representation. The \emph{compressed MDP} $\mathcal{M}_c$ recasts the embedded MDP in terms of the original state space, using expectations to express value functions, policies, and backups over $S$ rather than over the (exponentially larger) $S_e$. As we will see below, the compressed MDP induces a blow-up in action space rather than state space, but offers significant computational benefits. Formally, the state space for $\mathcal{M}_c$ is $S$. To capture action availability, the feasible action set for $s\in S$ is the set of \emph{state policies}, or mappings $\mu_s: 2^B \rightarrow B$ satisfying $\mu_s(A_s) \in A_s$. In other words, once we reach $s$, $\mu_s$ dictates what action to take for any realized action set $A_s$. A policy for $\mathcal{M}_c$ is a family $\mu_c = \{\mu_s :s\in S\}$ of such state policies. Transitions and rewards use expectations over $A_s$: \begin{small} \begin{align*} p^{\mu_s}_{s,s'} = \sum_{A_{s}\subseteq B} P_s(A_{s}) p^{\mu_s(A_s)}_{s,s'} ~~\mbox{and}~~ r^{\mu_s}_{s} = \sum_{A_{s}\subseteq B} P_s(A_{s}) r^{\mu_s(A_s)}_{s}~. \end{align*} \end{small} In our earlier example, the compressed MDP has only two states, $s_1$ and $s_2$. Focusing on $s_2$, its ``actions'' in the compressed MDP are the set of state policies, or mappings from the realizable available sets $\{\{\mathit{Up},\mathit{Down}\}, \{\mathit{Down}\}\}$ into action choices (as above, we ignore unrealizable action subsets): in this case, there are two such state policies: the first selects $\mathit{Up}$ for $\{\mathit{Up},\mathit{Down}\}$ and (obviously) $\mathit{Down}$ for $\{\mathit{Down}\}$; the second selects $\mathit{Down}$ for $\{\mathit{Up},\mathit{Down}\}$ and $\mathit{Down}$ for $\{\mathit{Down}\}$. It is not hard to show that the dynamics and reward process defined above over this compressed state space and expanded action set (i.e., the set of state policies) are Markovian. Hence we can define policies, value functions, optimality conditions, and policy and Bellman backup operators in the usual fashion. For instance, the Bellman and policy backup operators, $T^\star_c$ and $T_{\mu}^c$, on compressed value functions are: \begin{small} \begin{align} T_c^*V_c(s) =& \mathop{\mathbb E}_{A_s\subseteq B}\; \max_{k\in A_s} r^k_s + \gamma \sum_{s'} p^k_{s,s'} V_c(s'), \label{eq:compressedBellmanOp} \\ T^{\mu}_cV_c(s) =& \mathop{\mathbb E}_{A_s\subseteq B}\; r^{\mu_s(A_s)}_s + \gamma \sum_{s'} p^{\mu_s(A_s)}_{s,s'} V_c(s'). \label{eq:compressedPolicyOp} \end{align} \end{small} It is easy to see that any state policy $\mu$ induces a Markov chain over base states, hence we can define a standard $n\times n$ transition matrix $P^{\mu}$ for such a policy in the compressed MDP, where $p^{\mu}_{s,s'} = \mathop{\mathbb E}_{A\subseteq B} p^{\mu(s)(A)}_{s,s'}$. When additional independence assumptions hold, this expectation over subsets can be computed efficiently (see \cref{sec:pda}). Critically, we can show that there is a direct ``equivalence'' between policies and their value functions (including optimal policies and values) in $\mathcal{M}_c$ and $\mathcal{M}_e$. Define the action-expectation operator $E: \mathbb{R}^{n2^m} \rightarrow \mathbb{R}^n$ to be a mapping that compresses a value function $V_e$ for $\mathcal{M}_e$ into a value function $V^e_c$ for $\mathcal{M}_c$: \begin{small} $$ V^e_c(s) = EV_e(s) =\!\!\! \mathop{\mathbb E}_{A_s\subseteq B} V_e(s\circ A_s) =\!\!\! \sum_{A_s \subseteq B} P_s(A_{s}) V_e(s\circ A_s). $$ \end{small}% We emphasize that $E$ transforms an (arbitrary) value function $V_e$ in embedded space into a new value function $V_c^e$ defined in compressed space (hence, $V_c^e$ is \emph{not} defined w.r.t.\ $\mathcal{M}_c$). \begin{lemma} \label{lemma1} $ET^*_e V_e = T_c^*EV_e$. Hence, $T^*_c$ has a unique fixed point $V_c^\ast = EV_e^\ast$. \end{lemma} \begin{proof} \begin{small} \begin{align*} ET^eV_e(s) & = \mathop{\mathbb E}_{A\subseteq B} T^eV_e(s\circ A)\\ & = \mathop{\mathbb E}_{A\subseteq B} \max_{k\in A} r^k_s +\gamma\sum_{s'\circ A'} p^k_{s\circ A,s'\circ A'} V_e(s'\circ A') \\ & = \mathop{\mathbb E}_{A\subseteq B} \max_{k\in A} r^k_s +\gamma\sum_{s'} p^k_{s,s'} \mathop{\mathbb E}_{A'\subseteq B} V_e(s'\circ A') \\ & = \mathop{\mathbb E}_{A\subseteq B} \max_{k\in A} r^k_s +\gamma\sum_{s'} p^k_{s,s'} EV^e(s') \\ & = T^cEV^e(s') . \end{align*} \end{small} \end{proof} \begin{lemma} \label{lemma2} Given the optimal value function $V^\ast_c$ for $\mathcal{M}_c$, the optimal policy $\pi^\ast_e$ for $\mathcal{M}_e$ can be constructed directly. Specifically, for any $s\circ A$, the optimal policy $\pi^\ast_e(s\circ A)$ and optimal value $V^\ast_e(s\circ A)$ at that embedded state can be computed in polynomial time. \end{lemma} \begin{sketch} Given $s\circ A$, the expected value of each action in $k\in A$ can be computed using a one-step backup of $V^\ast_c$. Then $\pi^\ast_e(s\circ A)$ is the action with maximum value, and $V^\ast_e(s\circ A)$ is its backed-up expected value. \end{sketch} Therefore, it suffices to work directly with the compressed MDP, which allows one to use value functions (and $Q$-functions) over the original state space. The price is that one needs to use state policies, since the best action at $s$ depends on the available set $A_s$. In other words, while the embedded MDP causes an exponential blow-up in state space, the compressed MDP causes an exponential blow-up in action space. We now turn to assumptions that allow us to effectively manage this action space blow-up. \subsection{Decision List Policies} \label{sec:lists} The embedded and compressed MDPs do not, \emph{prima facie}, offer much computational or representational advantage, since they rely on an exponential increase in the size of the state space (embedded MDP) or decision space (compressed MDP). Fortunately, SAS-MDPs have optimal policies with a useful, concise form. We first focus on the policy representation itself, then describe the considerable computational leverage it provides. A \emph{decision list (DL) policy} $\mu$ is a type of policy for $\mathcal{M}_e$ that can be expressed compactly using $O(nm \log m)$ space and executed efficiently. Let $\Sigma_B$ be the set of permutations over base action set $B$. A DL policy $\mu: S \rightarrow \Sigma_B$ associates a permutation $\mu(s) \in \Sigma_B$ with each state, and is executed at embedded state $s\circ A$ by executing $\min \{i \in \{1,\ldots,m\} : \mu(s)(i) \in A\}$. In other words, whenever base state $s$ is encountered and $A$ is the available set, the first action $k\in A$ in the order dictated by DL $\mu(s)$ is executed. Equivalently, we can view $\mu(s)$ as a state policy $\mu_s$ for $s$ in $\mathcal{M}_c$. In our earlier example, one DL $\mu(s_2)$ is $[\mathit{Up},\mathit{Down}]$, which requires taking (base) action $\mathit{Up}$ if it is available, otherwise taking $\mathit{Down}$. For any SAS-MDP, we have optimal DL policies: \begin{thm} $\mathcal{M}_e$ has an optimal policy that can be represented using a decision list. The same policy is optimal for the corresponding $\mathcal{M}_c$. \end{thm} \begin{sketch} Let $V^\ast$ be the (unique) optimal value function for $\mathcal{M}_e$ and $Q^\ast$ its corresponding Q-function (see Sec.~\ref{sec:valueiteralg} for a definition). A simple inductive argument shows that no DL policy is optimal only if there is some state $s$, action sets $A \neq A'$, and (base) actions $j \neq k$, s.t.\ (i) $j,k \in A, A'$; (ii) for some optimal policy $\pi^\ast(s\circ A) = j$ and $\pi^\ast(s\circ A') = k$; and (iii) either $Q^\ast(s\circ A,j) > Q^\ast(s\circ A,k)$ or or $Q^\ast(s\circ A',k) > Q^\ast(s\circ A',j)$. However, the fact that the optimal Q-value of any action $k\in A$ at state $s\circ A$ is independent of the other actions in $A$ (i.e., it depends only on the base state) implies that these conditions are mutually contradictory. \end{sketch} \subsection{The Product Distribution Assumption} \label{sec:pda} The DL form ensures that optimal policies and value functions for SAS-MDPs can be expressed polynomially in the size of the base MDP $\mathcal{M}$. However, their computation still requires the computation of expectations over action subsets, e.g., in Bellman or policy backups (Eqs.~\ref{eq:compressedBellmanOp},~\ref{eq:compressedPolicyOp}). This will generally be infeasible without some assumptions on the form the action availability distributions $P_s$. One natural assumption is the \emph{product distribution assumption (PDA)}. PDA holds when $P_s(A)$ is a product distribution where each action $k\in B$ is available with probability $\rho^k_s$, and subset $A \subseteq B$ has probability $\rho_s^A = \prod_{k\in A} \rho^k_s \prod_{k\in B\setminus A} (1-\rho^k_s)$. This assumption is a reasonable approximation in the settings discussed above, where state-independent exogenous processes determine the availability of actions (e.g., the probability that one advertiser's campaign has budget remaining is roughly independent of another advertiser's). For ease of notation, we assume that $\rho^k_s$ is identical for all states $s$ (allowing different availability probabilities across states has no impact on what follows). To ensure the MDP is well-founded, we assume some default action (e.g., no-op) is always available.% \footnote{We omit the default action from analysis for ease of exposition.} Our earlier running example trivially satisifes PDA: at $s_2$, $\mathit{Up}$'s availability probability ($p$) is independent of the availability of $\mathit{Down}$ (1). When the PDA holds, the DL form of policies allows the expectations in policy and Bellman backups to be computed efficiently without enumeration of subsets $A\subseteq B$. For example, given a fixed DL policy $\mu$, we have \begin{small} \begin{align} T^{\mu}_cV_c(s) &= \sum_{i=1}^{m} \,\left[\prod_{j=1}^{i-1} (1-\rho^{\mu(s)(j)}_s)\right]\, \rho^{\mu(s)(i)}_s \Bigg( r_s^{\mu(s)(i)} \nonumber \\ &~ + \gamma \sum_{s'} p^{\mu(s)(i)}_{s,s'} V_c(s') \Bigg). \label{eq:pdabellman} \end{align} \end{small}% The Bellman operator has a similar form. We exploit this below to develop tractable value iteration and policy iteration algorithms, as well as a practical LP formulation. \subsection{Arbitrary Distributions with Sampling (ADS)} \label{sec:ads} We can also handle the case where, at each state, the availability distribution is unknown, but is sampleable. In the longer version of the paper \cite{sasmdps_full:arxiv18}, we show that samples can be used to approximate expectations w.r.t.\ available action subsets, and that the required sample size is polynomial in $|B|$, and not in the size of the \emph{support} of the distribution. Of course, when we discuss algorithms for policy computation, this approach does not allow us to compute the optimal policy exactly. However, it has important implications for sample complexity of learning algorithms like Q-learning. We note that the ability to sample available action subsets is quite natural in many domains. For instance, in ad domains, it may not be possible to model the process by which eligible ads are generated (e.g., involving specific and evolving advertiser targeting criteria, budgets, frequency capping, etc.). But the eligible subset of ads considered for each impression opportunity is an action-subset sampled from this process. Under ADS, we compute approximate backup operators as follows. Let $\mathcal{A}_s = \{A_s^{(1)},\ldots,A_s^{(T)}\}$ be an i.i.d. sample of size $T$ of action subsets in state $s$. For a subset of actions $A$, an index $i$ and a decision list $\mu$, define $I_{[i,A,\mu]}$ to be 1 if $\mu(i) \in A$ and for each $j < i$ we have $\mu(j) \not\in A$, or 0 otherwise. Similar to \cref{eq:pdabellman}, we define: { \small \begin{align*} T^{\mu}_cV_c(s) &\!=\! \frac{1}{T} \sum_{t=1}^T \sum_{i=1}^{m} I_{\left[i,A_s^{(t)}, \mu(s)\right]} \! \Big( r_s^{\mu(s)(i)} \!+\! \gamma \!\sum_{s'} p^{\mu(s)(i)}_{s,s'} V_c(s') \Big). \end{align*} }% In the sequel, we focus largely on PDA; in most cases equivalent results can be derived in the ADS model. \section{Q-Learning with the Compressed MDP} \label{sec:qlearn} Suppose we are faced with learning the optimal value function or policy for an SAS-MDP from a collection of trajectories. The (implicit) learning of the transition dynamics and rewards can proceed as usual; the novel aspect of the SAS model is that the action availability distribution must also be considered. Remarkably, Q-learning can be readily augmented to incorporate stochastic action sets: we require only that our training trajectories are augmented with the set of actions that were available at each state, $$\ldots s^{(t)}, A^{(t)}, k^{(t)}, r^{(t)}, s^{(t+1)}, A^{(t+1)}, k^{(t+1)}, r^{(t+1)}, \ldots,$$ where: $s^{(t)}$ is the realized state at time $t$ (drawn from distribution $P(\cdot|s^{(t-1)}, k^{(t-1)})$); $A^{(t)}$ is the realized available set at time $t$, drawn from $P_{s^{(t)}}$; $k^{(t)}\in A^{(t)}$ is the action taken; and $r^{(t)}$ is the realized reward. Such augmented trajectory data is typically available. In particular, the required sampling of available action sets is usually feasible (e.g., in ad serving as discussed above). \emph{SAS-Q-learning} can be applied directly to the compressed MDP $\mathcal{M}_c$, requiring only a minor modification of the standard Q-learning update for the base MDP. We simply require that each Q-update maximize over the \emph{realized available actions} $A^{(t+1)}$: \begin{small} \begin{align*} Q^{\textit{new}}(s^{(t)},k^{(t)}) &\leftarrow (1-\alpha_t) Q^{\overline{d}}(s^{(t)},k^{(t)}) \\ &\quad + \alpha_t [r^{(t)} + \gamma \max_{k\in A^{(t+1)}} Q^{\overline{d}}(s^{(t+1)},k)]~. \end{align*} \end{small}% Here $Q^{\overline{d}}$ is the previous $Q$-function estimate and $Q^{\textit{new}}$ is the updated estimate, thus it encompasses both online and batch Q-learning, experience replay, etc.; and $0\leq \alpha_t < 1$ is our (adaptive) learning rate. It is straightforward to show that, under the usual exploration conditions, SAS-Q-learning will converge to the optimal Q-function for the compressed MDP, since the expected maximum over sampled action sets at any particular state will converge to the expected maximum at that state. \begin{thm} The SAS-Q-learning algorithm will converge w.p.\ 1 to the optimal Q-function for the (discounted, infinite-horizon) compressed MDP $\mathcal{M}_c$ if the usual stochastic approximation requirements are satisfied. That is, if (a) rewards are bounded and (b) the subsequence of learning rates $\alpha_{t(s,k)}$ applied to $(s,k)$ satisfies $\sum \alpha_{t(s,k)} = \infty$ and $\sum \alpha^2_{t(s,k)} < \infty$ for all state-action pairs $(s,k)$ (see, e.g., \cite{watkins:mlj92}). \end{thm} Moreover, function approximation techniques, such as DQN \cite{mnih2015}, can be directly applied with the same action set-sample maximization. Implementing an optimal policy is also straightforward: given a state $s$ and the realization $A_s$ of the available actions, one simply executes $\arg\max_{k\in A_s} Q(s,k)$. We note that extracting the optimal value function $V_c(s)$ for the compressed MDP from the learned Q-function is not viable without some information about the action availability distribution. Fortunately, one need not know the expected value at a state to implement the optimal policy.\footnote{It is, of course, straightforward to learn an optimal value function if desired.} \section{Value Iteration in the Compressed MDP} \label{sec:valueiter} Computing a value function for $\mathcal{M}_c$, with its ``small'' state space $S$, suffices to execute an optimal policy. We develop an efficient \emph{value iteration (VI)} method to do this. \subsection{Value Iteration} \label{sec:valueiteralg} Solving an SAS-MDP using VI is challenging in general due to the required expectations over action sets. However, under PDA, we can derive an efficient VI algorithm whose complexity depends only polynomially on the base set size $|B|$. Assume a current iterate $V^t$, where $ V^t(s) = \mathop{\mathbb E}_{A_s} [\max_{k\in A_s} Q^t(s,k) ] $. We compute $V^{t+1}$ as follows: \begin{itemize}\itemsep 0pt\partopsep 0pt \item For each $s\in S, k\in B$, compute its $(t+1)$-stage-to-go Q-value: $Q^{t+1}(s,k) = r^k_s + \gamma \sum_{s'} p^k_{s,s'} V^t(s').$ \item Sort these Q-values in descending order. For convenience, we re-index each action by its Q-value rank (i.e., $k_{(1)}$ is the action with largest Q-value, and $\rho_{(1)}$ is its probability, $k_{(2)}$ the second-largest, etc.). \item For each $s\in S$, compute its $(t+1)$-stage-to-go value: \begin{small} \begin{align*} V^{t+1}(s) & = \mathop{\mathbb E}\nolimits_{A_s} \left[\max_{k\in A_s} Q^{t+1}(s,k)\right]\\ & = \sum_{i=1}^{m-1} \left( \prod_{j=1}^{i-1} (1 - \rho_{(j)}) \right) \rho_{(i)} Q^{t+1}(s,k_{(i)}) . \end{align*} \end{small} \end{itemize} Under ADS, we use the approximate Bellman operator: \begin{small} \begin{align*} \widehat{V}^{t+1}(s) &= \mathop{\mathbb E}\nolimits_{A_s} \left[\max_{k\in A_s} \widehat{Q}^{t+1}(s,k)\right] \\ &= \frac{1}{T} \sum_{t=1}^T \sum_{i=1}^{m} I_{\left[i,A_s^{(t)}, \mu(s)\right]} \widehat{Q}^{t+1}(s,\mu(s)(i))~, \end{align*} \end{small}% where $\mu(s)$ is the DL resulting from sorting $\widehat{Q}^{t+1}$-values. The Bellman operator under PDA is tractable: \begin{obs} \label{obs:VIperiteration} The compressed Bellman operator $T^*_c$ can be computed in $O(n m\log m)$ time. \end{obs} Therefore the per-iteration time complexity of VI for $\mathcal{M}_c$ compares favorably to the $O(n m)$ time of VI in the base MDP. The added complexity arises from the need to sort Q-values.% \footnote{The products of the action availability probabilities can be computed in linear time via caching.} Conveniently, this sorting process immediately provides the desired DL state policy for $s$. Using standard arguments, we obtain the following results, which immediately yield a polytime approximation method. \begin{lemma} \label{lemma3} $T^*_c$ is a contraction with modulus $\gamma$ i.e., $||T^*_c v_c - T^*_c v'_c|| \leq \gamma ||v_c - v'_c||$. \end{lemma} \begin{cor} For any precision $\varepsilon < 1$, the compressed value iteration algorithm converges to an $\varepsilon$-approximation of the optimal value function in $O(\log (L/\varepsilon))$ iterations, where $L\leq [\max_{s,k} r^k_s ]/(1-\gamma)$ is an upper bound on $||V^\ast_e||$. \end{cor} We provide an even stronger result next: VI, in fact, converges to an \emph{optimal} solution in polynomial time. \subsection{The Complexity of Value Iteration} \label{sec:valueitercomplexity} Given its polytime per-iteration complexity, to ensure VI is polytime, we must show that it converges to a value function that induces an optimal policy in polynomially many iterations. To do so, we exploit the compressed representation and adapt the technique of \cite{tseng:ORLetters90}. Assume, w.r.t.~the base MDP $\mathcal{M}$, that the discount factor $\gamma$, rewards $r^k_s$, and transition probabilities $p^k_{s,s'}$, are rational numbers represented with a precision of $1/ \delta$ ($\delta$ is an integer). Tseng shows that VI for a standard MDP is strongly polynomial, assuming constant $\gamma$ and $\delta$, by proving that: (a) if the $t$'th value function produced by VI satisfies $$ ||V^t - V^\ast|| < 1/(2 \delta^{2n+2} n^n) , $$ then the policy induced by $V^t$ is optimal; and (b) VI achieves this bound in polynomially many iterations. We derive a similar bound on the number of VI iterations needed for convergence in an SAS-MDP, using the same input parameters as in the base MDP, and applying the same precision $\delta$ to the action availability probabilities. We apply Tseng's result by exploiting the fact that: (a) the optimal policy for the embedded MDP $\mathcal{M}_e$ can be represented as a DL; (b) the transition function for any DL policy can be expressed using an $n\times n$ matrix (we simply take expectations, see above); and (c) the corresponding linear system can be expressed over the \emph{compressed} rather than the embedded state space to determine $V_c^\ast$ (rather than $V_e^\ast$). Tseng's argument requires some adaptation to apply to the compressed VI algorithm. We extend his precision assumption to account for our action availability probabilities as well, ensuring $\rho^k_s$ is also represented up to precision of $1/\delta$. Since $\mathcal{M}_c$ is an MDP, Tseng's result applies; but notice that each entry of the transition matrix for any state's DL $\mu$, which serves as an action in $\mathcal{M}_c$, is a product of $m+1$ probabilities, each with precision $1/\delta$. We have that $p^\mu_{s,s'}$ has precision of $1/\delta^{m+1}$. Thus the required precision parameter for our MDP is at most $\delta^{m+1}$. Plugging this into Tseng's bound, VI applied to $\mathcal{M}_c$ must induce an optimal policy at the $t$'th iteration if {\small $$ ||V^t - v^\ast|| < 1/(2({\delta^{(m+1)}})^{2n} n^n) = 1/(2\delta^{(m+1)2n} n^n)~. $$ }% This in turn gives us a bound on the number of iterations of VI needed to reach an optimal policy: \begin{thm} \label{thm:VIconvergence} VI applied to $\mathcal{M}_c$ converges to a value function whose greedy policy is optimal in $t^*$ iterations, where {\small $$ t^* \leq \log(2\delta^{2n(m+1)} n^n M) / \log(1/\gamma) $$ } \end{thm} Combined with Obs.~\ref{obs:VIperiteration}, we have: \begin{cor} \label{cor:VIpolytime} VI yields an optimal policy for the SAS-MDP corresponding to $\mathcal{M}_c$ in polynomial time. \end{cor} Under ADS, VI merely approximates the optimal policy. In fact, one cannot compute an exact optimal policy without observing the entire support of the availability distributions (requiring exponential sample size). \section{Policy Iteration in the Compressed MDP} \label{sec:policyiter} We now outline a policy iteration (PI) algorithm. \subsection{Policy Iteration} \label{sec:policyiteralg} The concise DL form of optimal policies can be exploited in PI as well. Indeed, \emph{the greedy policy $\pi^V$ with respect to any value function $V$ in the compressed space} is representable as a DL. Thus the policy improvement step of PI can be executed using the same independent evaluation of action Q-values and sorting as used in VI above: \begin{small} \begin{gather*} Q^V(s,k) = r(s,k) + \gamma \sum_{s'} p^k_{s, s'} V(s') ,\\ Q^V(s,A_s) \!=\! \max_{k\in A_s} Q^V(s,k) \;\textrm{, and } \; \pi^V(s,A_s) \!=\! \arg\max_{k\in A_s} Q^V(s,k). \end{gather*} \end{small} The DL policy form can also be exploited in the policy evaluation phase of PI. The tractability of policy evaluation requires a tractable representation of the action availability probabilities, which PDA provides, leading to the following PI method that exploits PDA: \begin{enumerate}\itemsep 0pt\partopsep 0pt \item Initialize an arbitrary policy $\pi$ in decision list form. \item Evaluate $\pi$ by solving the following linear system over variables $V^\pi(s), \forall s\in S$: (Note: We use $Q^{\pi}(s,k)$ to represent the relevant linear expression over $V^\pi$.) \begin{small} \begin{align*} V^\pi(s) &= \sum_{i=1}^{n} \,[\prod_{j=1}^{i-1} (1-\rho_{(j)})]\, \rho_{(i)} Q^{\pi}(s,k_{(i)}) \label{eq:pda_expect_val} \end{align*} \end{small} \item Let $\pi'$ denote the greedy policy w.r.t.\ $V^\pi$, which can be expressed in DL form for each $s$ by sorting Q-values $Q^{\pi}(s,k)$ as above (with standard tie-breaking rules). If $\pi'(s) = \pi(s)$, terminate; otherwise replace $\pi$ with $\pi'$ and repeat (Steps 2 and 3). \end{enumerate} Under ADS, PI can use the approximate Bellman operator, giving an approximately optimal policy. \subsection{The Complexity of Policy Iteration} \label{sec:policyitercomplexity} The per-iteration complexity of PI in $\mathcal{M}_c$ is polynomial: as in standard PI, policy evaluation solves an $n\times n$ linear system (naively, $O(n^3)$) plus the additional overhead (linear in $M$) to compute the compounded availability probabilities; and policy improvement requires $O(mn^2)$ computation of action Q-values, plus $O(nm\log m)$ overhead for sorting Q-values (to produce improving DLs for all states). An optimal policy is reached in a number of iterations no greater than that required by VI, since: (a) the sequence of value functions for the policies generated by PI contracts at least as quickly as the value functions generated by VI (see, e.g., \cite{meister:ORSpektrum86,hansen:jacm13}); (b) our precision argument for VI ensures that the greedy policy extracted at that point will be optimal; and (c) once PI finds an optimal policy, it will terminate (with one extra iteration). Hence, PI is polytime (assuming a fixed discount $\gamma<1$). \begin{thm} \label{thm:PIpolytime} PI yields an optimal policy for the SAS-MDP corresponding to $\mathcal{M}_c$ in polynomial time. \end{thm} In the longer version of the paper \cite{sasmdps_full:arxiv18}, we adapt more direct proof techniques \cite{yinyuye:mathOR11,hansen:jacm13} to derive polynomial-time convergence of PI for SAS-MDPs under additional assumptions. Concretely, for a policy $\mu$ and actions $k_1,k_2$, let $\eta_\mu(s,k_1,k_2)$ be the probability, over action sets, that at state $s$, the optimal $\mu^\star$ selects $k_1$ and $\mu$ selects $k_2$. Let $q > 0$ be such that $\eta_\mu(s,k_1,k_2) \ge q$ whenever $\eta_\mu(s,k_1,k_2) > 0$. We show: \begin{thm} The number of iterations it takes policy iteration to converge is no more than { \small \[ O\left( \frac{nm^2}{1-\gamma} \log \frac{m}{1-\gamma} \log \frac{e}{q} \right)~. \] } \end{thm} Under PDA, the theorem implies \emph{strongly-polynomial} convergence of PI if each action is available with constant probability. In this case, for any $\mu$, $k_i$, $k_j$, and $s$, we have $\eta_\mu(s,k_i,k_j) \ge \rho_s^{k_i} \cdot \rho_s^{k_j} = \Omega(1)$, which in turn implies that we can take $q = \Omega(1)$ in the bound above. \section{Linear Programming in the Compressed MDP} \label{sec:lp} An alternative model-based approach is linear programming (LP). The primal formulation for the embedded MDP $\mathcal{M}_e$ is straightforward (since it is a standard MDP), but requires exponentially many variables (one per embedded state) and constraints (one per embedded state, base action pair). A (nonlinear) primal formulation for the compressed MDP $\mathcal{M}_c$ reduces the number of variables to $|S|$: \begin{small} \begin{align} \min_{\mathbf{v}}\,\sum_{s\in S}\nolimits\alpha_s v_s,\quad\textrm{s.t. }\,v_s \geq \mathop{\mathbb E}\nolimits_{A_s}\max_{k\in A_s} Q(s,k)\quad \forall s. \label{eq:compressedPrimalConstr} \end{align} \end{small}% Here $\alpha$ is an arbitrary, positive state-weighting, over the embedded states corresponding to each base state and $$Q(s,k) = r_s^k + \sum_{s' \in S} p^k_{s,s'} v_{s'}$$ abbreviates the linear expression of the action-value backup at the state and action in question w.r.t.\ the value variables $v_s$. This program is valid given the definition of $\mathcal{M}_c$ and the fact that a weighting over embedded states corresponds to a weighting over base states by taking expectations. Unfortunately, this formulation is non-linear, due to the max term in each constraint. And while it has only $|S|$ variables, it has factorially many constraints; moreover, the constraints themselves are not compact due to the presence of the expectation in each constraint. PDA can be used to render this formulation tractable. Let $\sigma$ denote an arbitrary (inverse) permutation of the action set (so $\sigma(i)=j$ means that action $j$ is ranked in position $i$). As above, the optimal policy at base state $s$ w.r.t.\ a Q-function is expressible as a DL ( with actions sorted by Q-values) and its expected value given by the expression derived below. Specifically, if $\sigma$ reflects the relative ranking of the (optimal) Q-values of the actions at some fixed state $s$, then $V(s) = Q(s,\sigma(1))$ with probability $\rho_{\sigma(1)}$, i.e., the probability that $\sigma(1)$ occurs in $A_s$. Similarly, $V(s) = Q(s,\sigma(2))$ with probability $(1-\rho_{\sigma(1)})\rho_{\sigma(2)}$, and so on. We define the Q-value of a DL $\sigma$ as follows: \begin{small} \begin{align} Q^V_s(\sigma) = \sum_{i=1}^{n} \,[\prod_{j=1}^{i-1} (1-\rho_{\sigma(j)})]\, \rho_{\sigma(i)} Q^V(s,\sigma(i)). \end{align} \end{small}% Thus, for any fixed action permutation $\sigma$, the constraint that $v_s$ at least matches the expectation of the maximum action's Q-value is linear. Hence, the program can be recast as an LP by enumerating action permutations for each base state, replacing the constraints in Eq.~(\ref{eq:compressedPrimalConstr}) as follows: \begin{small} \begin{align} v_s \geq Q^V_s(\sigma) \quad \forall s\in S, \forall \sigma \in \Sigma. \label{eq:compressedPrimalConstr2} \end{align} \end{small} The constraints in this LP are now each compactly represented, but it still has factorially many constraints. Despite this, it can be solved in polynomial time. First, we observe that the LP is well-suited to constraint generation. Given a relaxed LP with a subset of constraints, a greedy algorithm that simply sorts actions by Q-value to form a permutation can be used to find the maximally violated constraint at any state. Thus we have a practical constraint generation algorithm for this LP since (maximally) violated constraints can be found in polynomial time. More importantly from a theoretical standpoint, the constraint generation algorithm can be used as a separation oracle within an ellipsoid method for this LP. This directly yields an exact, (weakly) polynomial time algorithm for this LP~\cite{GroetschelLovaszSchrijver1988}. \section{Empirical Illustration} \label{sec:empirical} We now provide a somewhat more elaborate empirical demonstration of the effects of stochastic action availability. Consider an MDP that corresponds to a routing problem on a real-world road network (Fig.~\ref{fig:routing}) in the San Francisco Bay Area. The shortest path between the source and destination locations is sought. The dashed edge in Fig.~\ref{fig:routing} represents a bridge, available only with probability $p$, while all other edges correspond to action choices available with probability $0.5$. At each node, a no-op action (waiting) is available at constant cost; otherwise the edge costs are the geodesic lengths of the corresponding roads on the map. The optimal policies for different choices $p=0.1, 0.2$ and $0.4$ are depicted in Fig.~\ref{fig:routing}, where line thickness and color indicate traversal probabilities under the corresponding optimal policies. It can be observed that lower values of $p$ lead to policies with more redundancy. Fig.~\ref{fig:obliv} investigates the effect of solving the routing problem obliviously to the stochastic action availability (assuming actions are fully available). The SAS-optimal policy allows graceful scaling of the expected travel time from source to destination as bridge availability decreases. Finalluy, the effects of violating the PDA assumption are investigated in the long version of this work~\cite{TODO}. \begin{figure}[t!] \centering \includegraphics[width=0.5\columnwidth]{bridgefigure.pdf} \caption{Stochastic action MDPs applied to routing.\label{fig:routing}\vspace{-5pt}} \end{figure} \vspace{-5pt} \begin{figure} \begin{minipage}[c]{0.5\columnwidth} \includegraphics[width=\textwidth]{optimal_vs_obliv.pdf} \end{minipage}\hfill \begin{minipage}[c]{0.5\columnwidth} \caption{ Expected trip time from source to destination under the SAS-optimal policy vs. under the oblivious optimal policy (the MDP solved as if actions are fully available) as a function of bridge availability.}\label{fig:obliv} \end{minipage} \end{figure} \vspace{-5pt} \section{Concluding Remarks} \label{sec:conclude} We have developed a new MDP model, \emph{SAS-MDPs}, that extends the usual finite-action MDP model by allowing the set of available actions to vary stochastically. This captures an important use case that arises in many practical applications (e.g., online advertising, recommender systems). We have shown that embedding action sets in the state gives a standard MDP, supporting tractable analysis at the cost of an exponential blow-up in state space size. Despite this, we demonstrated that (optimal and greedy) policies have a useful decision list structure. We showed how this DL format can be exploited to construct tractable Q-learning, value and policy iteration, and linear programming algorithms. While our work offers firm foundations for stochastic action sets, most practical applications will not use the algorithms described here explicitly. For example, in RL, we generally use function approximators for generalization and scalability in large state/action problems. We have successfully applied Q-learning using DNN function approximators (i.e.,~DQN) using sampled/logged available actions in ads and recommendations domains as described in Sec.~\ref{sec:qlearn}. This has allowed us to apply SAS-Q-learning to problems of significant, commercially viable scale. Model-based methods such as VI, PI, and LP also require suitable (e.g., factored) representations of MDPs and structured implementations of our algorithms that exploit these representations. For instance, extensions of approximate linear programming or structured dynamic programming to incorporate stochastic action sets would be extremely valuable. Other important questions include developing a polynomial-\emph{sized} direct LP formulation; and deriving sample-complexity results for RL algorithms like Q-learning is also of particular interest, especially as it pertains to the sampling of the action distribution. Finally, we are quite interested in relaxing the strong assumptions embodied in the PDA model---of particular interest is the extension of our algorithms to less extreme forms of action availability independence, for example, as represented using consise graphical models (e.g., Bayes nets). \vskip 2mm \noindent \textbf{Acknowledgments:} Thanks to the reviewers for their helpful suggestions. \bibliographystyle{plain}

https://arxiv.org/abs/1805.07876

Noncoherent Short-Packet Communication via Modulation on Conjugated Zeros

We introduce a novel blind (noncoherent) communication scheme, called modulation on conjugate-reciprocal zeros (MOCZ), to reliably transmit short binary packets over unknown finite impulse response systems as used, for example, to model underspread wireless multipath channels. In MOCZ, the information is modulated onto the zeros of the transmitted signals $z-$transform. In the absence of additive noise, the zero structure of the signal is perfectly preserved at the receiver, no matter what the channel impulse response (CIR) is. Furthermore, by a proper selection of the zeros, we show that MOCZ is not only invariant to the CIR, but also robust against additive noise. Starting with the maximum-likelihood estimator, we define a low complexity and reliable decoder and compare it to various state-of-the art noncoherent schemes.

\section{Channel Model} In this work we will consider communication over frequency-selective block-fading channels used for indoor and outdoor scenarios, where the channel delay spread $\Td$ is in the order of the signal duration $T_s=NT$, given by the symbol duration $T$ and overall block length $N$. We assume that the channel is time-invariant in each block, but changes arbitrary from block to block, which models a time-varying channel \cite{VHHK01}. Conventional coherent communication strategies, e.g., most based on OFDM, are expected to be inefficient in this regime. We will therefore propose (in the next section) a novel modulation scheme for noncoherent communication, which keeps the relevant information in the transmitted signal invariant under multipath propagation and therefore completely avoids channel estimation and signal equalization at the receiver. Assuming that the CIR remains constant over the one-shot (block) communication period, the discrete-time baseband model for this channel is given as a linear \emph{convolution}: \begin{align} y_n = \sum_{{\ensuremath{l}}=0}^{{\ensuremath{L}}-1} h_{\ensuremath{l}} x_{n-{\ensuremath{l}}} + w_n \quad\text{for}\quad n\in\{0,1,2,\dots,N\}, \end{align} of the transmitted time symbols $\{x_n\}_{n=0}^{K}$ with the complex-valued channel coefficients (taps) $\{h_{\ensuremath{l}}\}_{{\ensuremath{l}}=0}^{L-1}\in{\ensuremath{\mathbb C}}$ resulting in a block of $N=L+K$ received symbols. Additionally, the convolution is disturbed by additive noise $w_n$. We denote the block (packet) of ${\ensuremath{K}}+1$ transmitted time symbols as the vector ${\ensuremath{\mathbf x}}=(x_0,x_1,\dots,x_{\ensuremath{K}})^T\in{\ensuremath{\mathbb C}}^{{\ensuremath{K}}+1}$ and assume wlog a normalization $\Norm{{\ensuremath{\mathbf x}}}_2^2=\sum_{\ensuremath{k}} |x_{\ensuremath{k}}|^2=1$. In this form, we obtain at the receiver the vector: \begin{align} {\ensuremath{\mathbf y}}= {\ensuremath{\mathbf x}} * \vh + {\ensuremath{\mathbf w}}\in \mathbb{C}^{N}\label{eq:receivedsignal}. \end{align} Contrary to usual assumptions, we assume that only one packet ${\ensuremath{\mathbf x}}$ is transmitted, which is called a ``one-shot'' communication. Here, the next transmission will be at an indefinite time point such that it is not possible to predict the CIR. Such a \emph{sporadic} transmission scheme can therefore be seen as a prototype problem relevant for machine-to-machine communications, car-to-car/infrastructure and wireless sensor networks where status updates and control messages determine the typical traffic type. \subsection{Channel and Noise Statistics} The channel and noise taps are modeled as independent circularly symmetric Gaussian random variables \begin{align} \vh &\in \mathbb{C}^{{\ensuremath{L}}} \quad, \quad h_{\ensuremath{l}} \sim \ensuremath{\C^N}(0,\pd^{{\ensuremath{l}}}) \label{eq:channel}\\ {\ensuremath{\mathbf w}} &\in \mathbb{C}^{N}\quad,\quad w_n \sim \ensuremath{\C^N}(0,\sigma^2)\label{eq:noise} \end{align} where we assume with $p\leq 1$ an exponential decaying average power delay profile $\Expect{|h_{\ensuremath{l}}|^2}=\pd^{{\ensuremath{l}}}$ for the ${\ensuremath{l}}$th path, see for example \cite{JSP96}. The average noise power is denoted by $\ensuremath{\sigma}^2=N_0>0$ and is constant for each tap. Due to the independence of the channel taps we can derive for the \emph{average received signal-to-noise ratio}: \begin{align} {\ensuremath{\text{rSNR}}} = \Expect{\left(\frac{ \Norm{{\ensuremath{\mathbf x}}*\vh}^2_2}{\Norm{\vn}^2_2}\right)} =\frac{\Expect{\Norm{{\ensuremath{\mathbf x}}}_2^2} \Expect{\Norm{\vh}_2^2}}{\Expect{\Norm{\vn}^2_2}} =\frac{\Expect{\Norm{\vh}_2^2}}{N\cdot N_0}.\label{eq:rSNR} \end{align} The average energy of the multipath Rayleigh fading channel $\vh$ is then given by \begin{align} \Expect{\Norm{\vh}_2^2}=\sum_{{\ensuremath{l}}=0}^{{\ensuremath{L}}-1} \pd^{{\ensuremath{l}}}= \frac{1-\pd^{{\ensuremath{L}}}}{1-\pd}. \end{align} Hence we obtain \begin{align} {\ensuremath{\text{rSNR}}} = \frac{1}{N\cdot N_0} \frac{1-\pd^{{\ensuremath{L}}}}{1-\pd}. \end{align} \section{Transmission Scheme via Modulation On Zeros} The convolution in \eqref{eq:receivedsignal} can be also represented by a polynomial multiplication. Let ${\ensuremath{\mathbf x}}\in{\ensuremath{\mathbb C}}^{{\ensuremath{K}}+1}$, then its $z$-transform is the polynomial \begin{align} {\ensuremath{\mathrm X}}(z)=\sum_{{\ensuremath{k}}=0}^{{\ensuremath{K}}} x_{\ensuremath{k}} z^{\ensuremath{k}}\quad,\quad z\in{\ensuremath{\mathbb C}}, \end{align} which has order ${\ensuremath{K}}$ if and only if $x_{\ensuremath{K}}\not=0$. The received signal \eqref{eq:receivedsignal} is in the $z-$domain given by a polynomial of order ${\ensuremath{K}}+{\ensuremath{L}}-1$ \begin{align} {\ensuremath{\mathrm Y}}(z)={\ensuremath{\mathrm X}}(z){\ensuremath{\mathrm H}}(z)+{\rm{W}}(z), \end{align} where ${\ensuremath{\mathrm X}}(z), {\ensuremath{\mathrm H}}(z)$ and ${\rm{W}}(z)$ are the polynomials of order ${\ensuremath{K}}$, ${\ensuremath{L}}-1$ and ${\ensuremath{K}}+{\ensuremath{L}}-1$ generated by ${\ensuremath{\mathbf x}},\vh$ respectively ${\ensuremath{\mathbf w}}$. Any polynomial ${\ensuremath{\mathrm X}}(z)$ of order ${\ensuremath{K}}$, can also be represented by its ${\ensuremath{K}}$ zeros $\ensuremath{\alpha_{\ensuremath{k}}}$ and its leading coefficient $x_{\ensuremath{K}}$ as \begin{align} {\ensuremath{\mathrm X}}(z)=x_{\ensuremath{K}} \prod_{{\ensuremath{k}}=1}^{\ensuremath{K}} (z-\ensuremath{\alpha}_{\ensuremath{k}}). \end{align} If we assume that ${\ensuremath{\mathbf x}}$ is normalized, then $x_{\ensuremath{K}}$ is fully determined by its ${\ensuremath{K}}$ zeros, which leaves us with ${\ensuremath{K}}$ degrees of freedom for our signals, given by $K$ \emph{zero-symbols} $\ensuremath{\alpha}_k$. Let us note, that the notation ${\ensuremath{\mathrm X}}(z)$ is commonly used for the $z-$transform. However, since each polynomial of order ${\ensuremath{K}}$, with non-vanishing zeros, corresponds to a unilateral (one-sided) $z-$transform with the same zeros and an additional pole at $z=0$, both ``zero'' representations above are equivalent. In this work we will exclusively use the polynomial notation, since it will be more convenient for our purpose. The multiplication by the channel polynomial ${\ensuremath{\mathrm H}}(z)$ adds at most ${\ensuremath{L}}-1$ zeros $\ensuremath{\beta}_{\ensuremath{l}}$, which may be arbitrary distributed over the complex plane depending by the actual channel coefficients. However, for typical random channel models, it holds with probability one that the channel and signal polynomials, generated by a finite codebook set $\Calg\subset {\ensuremath{\mathbb C}}^{{\ensuremath{K}}+1}$, do not share a common zero. The \emph{no common zero} property is a necessary condition for blind deconvolution, see \cite{XLTK95,LXTK96,WJPH17}. We will later investigate in more detail the distribution of the zeros and their dependence on the coefficients to derive robustness results against additive noise. Contrary to time or frequency modulations, where each time-symbol, resp. frequency-symbol, uses the whole complex plane as its constellation domain, the $K$ zero-symbols have to share their constellation domains. Hence, we need to partition the complex plane in $M {\ensuremath{K}}$ disjoint (connected) sets $\{\Dset_{{\ensuremath{k}}}^{(m)}\}_{{\ensuremath{k}}=1,m=0}^{{\ensuremath{K}},M-1}$ and cluster them to ${\ensuremath{K}}$ sectors (constellation domains) $\ensuremath{\mathscr S}_{\ensuremath{k}}:=\bigcup_{m=0}^{M-1} \Dset_{\ensuremath{k}}^{(m)}$ for ${\ensuremath{k}}=1,2,\dots, {\ensuremath{K}}$ of size $M$ each. For each set $\Dset_{{\ensuremath{k}}}^{(m)}$ we associate exactly one zero $\ensuremath{\alpha}_{\ensuremath{k}}^{(m)}$. This will define ${\ensuremath{K}}$ \emph{zero constellation sets} $\Alp_{\ensuremath{k}}=\{\ensuremath{\alpha}_{\ensuremath{k}}^{(0)},\dots,\ensuremath{\alpha}_{\ensuremath{k}}^{(M-1)}\}$ for $k=1,2,\dots,K$ of $M$ zeros each. If we select from each $ \Alp_{\ensuremath{k}}$ exactly one zero-symbol $\ensuremath{\alpha}_{\ensuremath{k}}$, then we can construct $M^{\ensuremath{K}}$ different zero vectors \begin{align} \ensuremath{\boldsymbol{ \alp}}=\begin{pmatrix}\ensuremath{\alpha}_1\\ \vdots\\ \ensuremath{\alpha}_{\ensuremath{K}}\end{pmatrix} \in \ensuremath{\mathscr Z}=\Alp_1\times \dots \times \Alp_{\ensuremath{K}}\subset {\ensuremath{\mathbb C}}^{{\ensuremath{K}}}. \end{align} The \emph{zero-codebook} $\ensuremath{\mathscr Z}$ has cardinality $M^{\ensuremath{K}}$ and allows therefore to encode ${\ensuremath{K}}\log M$ bits. Hence, the message stream of an $M$-ary alphabet is partitioned in words $\ensuremath{\mathbf m}=(m_1,\dots,m_{\ensuremath{K}})^T$ of length ${\ensuremath{K}}$ and each letter $m_{\ensuremath{k}}$ is assigned to the ${\ensuremath{k}}$th zero-symbol $\ensuremath{\alpha}_{\ensuremath{k}}\in\Alp_{\ensuremath{k}}$, see \figref{fig:mozscheme}. Note, that the zero constellation sets $\Alp_{\ensuremath{k}}$ have to be ordered in the zero-codebook, otherwise a unique letter assignment would not be possible. The zero vector $\ensuremath{\boldsymbol{ \alp}}$ generates then by the Vitae formula $\ensuremath{\mathscr{V}}$, see for example \cite{MMR94}, the coefficients of the corresponding polynomial\footnote{The Vitae formula can also be seen as an explicit formula for the inverse $z-$transform $z^{-{\ensuremath{K}}} x_0\prod(z-\ensuremath{\alpha}_{\ensuremath{k}})$. } \begin{align} {\ensuremath{\mathbf x}}=\ensuremath{\mathscr{V}}(\ensuremath{\boldsymbol{ \alp}})= x_{\ensuremath{K}}\begin{pmatrix} (-1)^{\ensuremath{K}} \prod_k \ensuremath{\alpha}_{\ensuremath{k}} \\ \vdots\\ -\sum_{\ensuremath{k}} \ensuremath{\alpha}_{\ensuremath{k}}\\ 1 \end{pmatrix}\label{eq:vitae}, \end{align} where $x_{\ensuremath{K}}=x_{\ensuremath{K}}(\ensuremath{\boldsymbol{ \alp}})$ is chosen, such that ${\ensuremath{\mathbf x}}$ has unit $\ell_2-$norm. These signal constellations therefore define an $({\ensuremath{K}}+1)-$block codebook $\Calg$ of signals (sequences) in the time--domain. To avoid a signal overlap between blocks we use a guard interval of ${\ensuremath{L}}-1$ resulting in a received block length of $N={\ensuremath{K}}+{\ensuremath{L}}$. We will call this channel encoding scheme a \emph{Modulation On Zeros} (MOZ), see \figref{fig:mozscheme} and \figref{fig:BMOCZscheme} for $M=2$. Let us note, that the digital data, modulated on the zero-symbols, results in perfect interleaved time and frequency symbols. Hence, the transmitter exploits the full multipath diversity in time and frequency. This is in contrast to most modulation schemes, which either interleave the data in time (OFDM) or frequency domain (PPM, PAM). \begin{figure}[t] \centering \def0.65\textwidth} \footnotesize{ \import{\pdfdir}{OptimaRadiusHuffmanBMOCZ.pdf_tex} {\textwidth} \footnotesize{ \import{\pdfdir}{MOZblockscheme.pdf_tex} } \caption{MOZ scheme}\label{fig:mozscheme} \end{figure} \subsection{Modulation On Conjugate-reciprocal Zeros} One such partition structure is given for even $M$ by $M{\ensuremath{K}}/2$ conjugate-reciprocal zero pairs with distinct phases. By ordering the pairs by their phases in increasing order, we can generate ${\ensuremath{K}}$ sectors with $M/2$ possible conjugate-reciprocal zero pairs \begin{align} \Alp_{\ensuremath{k}}= \Big\{ \{(\ensuremath{\alpha}_{\ensuremath{k}}^{(0)},\ensuremath{\alpha}_{\ensuremath{k}}^{(1)})\},\{(\ensuremath{\alpha}_{\ensuremath{k}}^{(2)},\ensuremath{\alpha}_{\ensuremath{k}}^{(3)})\},\dots,\{(\ensuremath{\alpha}_{\ensuremath{k}}^{(M-2)},\ensuremath{\alpha}_{\ensuremath{k}}^{(M-1)})\}\Big\}, \end{align} where for all $m=0,2,4,\dots,M-2$ we have $\ensuremath{\alpha}_{\ensuremath{k}}^{(m+1)}=1/\cc{\ensuremath{\alpha}_{\ensuremath{k}}^{(m)}}$. We will additionally order $\ensuremath{\alpha}_{\ensuremath{k}}^{(m)}$ by increasing phase or radius respectively. This allows to encode $\log M$ bits per transmitted zero and we call this scheme an $M-$ary \emph{Modulation On Conjugate-reciprocal Zeros} (MOCZ), pronounced as ``Moxie``. If we set $M=2$ we can encode exactly ${\ensuremath{K}}$ bits in the signal ${\ensuremath{\mathbf x}}$. The $2{\ensuremath{K}}$ zeros $\bigcup \Alp_{\ensuremath{k}}$ of the ${\ensuremath{K}}$ pairs define an autocorrelation ${\ensuremath{\mathbf a}}\in{\ensuremath{\mathbb C}}^{2{\ensuremath{K}}+1}$ where we set the leading coefficient $a_{2{\ensuremath{K}}}$ such that $a_K=1$. Then each normalized signal ${\ensuremath{\mathbf x}}$ is generated by \eqref{eq:vitae} from the zero codeword \begin{align} \ensuremath{\boldsymbol{ \alp}}=\begin{pmatrix}\ensuremath{\alpha}_1\\ \vdots\\ \ensuremath{\alpha}_{\ensuremath{K}}\end{pmatrix} \in \ensuremath{\mathscr Z}:=\{\ensuremath{\alpha}_1^{(0)},\ensuremath{\alpha}_1^{(1)}\}\times \dots \times \{\ensuremath{\alpha}_{\ensuremath{K}}^{(0)},\ensuremath{\alpha}_{\ensuremath{K}}^{(1)}\}\subset {\ensuremath{\mathbb C}}^{{\ensuremath{K}}}, \end{align} and will have the same autocorrelation ${\ensuremath{\mathbf a}}={\ensuremath{\mathbf x}}*\cc{{\ensuremath{\mathbf x}}^-}$, see \figref{fig:BMOCZscheme}. Hence, the codebook $\Calg$ can be seen as an autocorrelation codebook, where the ${\ensuremath{K}}$ bits of information are encoded in the $2^{\ensuremath{K}}$ non-trivial ambiguities\footnote{The trivial scaling ambiguity, is not seen by the zeros and is in the MOZ scheme not used for information. Hence we loose one degree of freedom of the signal dimension ${\ensuremath{K}}+1$. However, this scheme is therefore independent to global phase of the signals. However, the absolute scaling effects the transmitted and received power which governors the SNR and hence the robustness against noise. } of the autocorrelation. Let us set $\ensuremath{\alpha}_{\ensuremath{k}}^{(0)}=\ensuremath{R_{\ensuremath{k}}^{-1}} e^{\im\ensuremath{\phi_{\ensuremath{k}}}}$ and $\ensuremath{\alpha}_{\ensuremath{k}}^{(1)}=\ensuremath{R_{\ensuremath{k}}} e^{\im \ensuremath{\phi_{\ensuremath{k}}}}$ for $\phi_1<\phi_2<\dots <\phi_{\ensuremath{K}}$ and $\ensuremath{R_{\ensuremath{k}}}>1$ for every ${\ensuremath{k}}\in[{\ensuremath{K}}]$. We can then encode a block $\ensuremath{\mathbf m}\in\{0,1\}^{\ensuremath{K}}$ of ${\ensuremath{K}}$ bits $m_{\ensuremath{k}}$ in ${\ensuremath{\mathbf x}}\in{\ensuremath{\mathbb C}}^{{\ensuremath{K}}+1}$ by assigning the zeros to \begin{equation} \ensuremath{\alpha}_{\ensuremath{k}}:=\begin{cases} \ensuremath{\alpha}_{\ensuremath{k}}^{(1)} = \ensuremath{R_{\ensuremath{k}}} e^{\im\ensuremath{\phi_{\ensuremath{k}}}}&, m_{\ensuremath{k}}=1\\ \ensuremath{\alpha}_{\ensuremath{k}}^{(0)}=\ensuremath{R_{\ensuremath{k}}^{-1}} e^{\im\phi_{\ensuremath{k}}}&, m_{\ensuremath{k}}=0 \end{cases} \quad,\quad {\ensuremath{k}}\in[{\ensuremath{K}}], \end{equation} see \figref{fig:BMOCZscheme}. We call this scheme a \emph{Binary Modulation On Conjugate-reciprocal Zeros} (BMOCZ). The blue circles denote the conjugate-reciprocal zero pairs, which define the zero-codebook $\ensuremath{\mathscr Z}$. The solid blue circles are the actual transmitted zeros and the red square zeros are the received zeros, given by the disturbed channel and data zeros. \begin{figure}[H] \centering \begin{subfigure}{0.47\textwidth} \vspace{0.04\textwidth} \def0.65\textwidth} \footnotesize{ \import{\pdfdir}{OptimaRadiusHuffmanBMOCZ.pdf_tex} {0.9\textwidth} \footnotesize{ \import{\pdfdir}{zeroAlpBMOCZ_codingsets2.pdf_tex} } \vspace{0.117\textwidth} \caption{Arbitrary BMOCZ scheme. } \label{fig:BMOCZscheme} \end{subfigure} \begin{subfigure}{0.5\textwidth} \def0.65\textwidth} \footnotesize{ \import{\pdfdir}{OptimaRadiusHuffmanBMOCZ.pdf_tex} {\textwidth} \footnotesize{ \import{\pdfdir}{zeroHuffmanBMOCZ_codingsets2.pdf_tex} } \vspace{-0.07\textwidth} \caption{Huffman BMOCZ scheme.} \label{fig:HuffmanBMOCZscheme} \end{subfigure} \caption{The zero-codebook $\Alp$ and their decoding sets (Voronoi cells). Red squares denote received zeros.} \label{fig:BMOCZ} \end{figure} \subsection{Demodulation and Decoding via Root finding and Minimum Distance} Let us first explain how one could in principle demodulate the data. The following exposition is meant mainly for illustration and analysis. More efficient implementations will be discussed later on. Thus, at the receiver we will observe by \eqref{eq:receivedsignal} a disturbed version of the transmitted polynomial \begin{align} {\ensuremath{\mathrm Y}}(z)= {\ensuremath{\mathrm X}}(z)\cdot {\ensuremath{\mathrm H}}(z) +{\rm{W}}(z)=x_{\ensuremath{K}} h_{{\ensuremath{L}}-1}\prod_{{\ensuremath{k}}=1}^{\ensuremath{K}} (z-\ensuremath{\alpha}_{\ensuremath{k}})\prod_{{\ensuremath{l}}=1}^{{\ensuremath{L}}-1} (z-\ensuremath{\beta}_{\ensuremath{l}}) + w_{N-1} \prod_{n=1}^{N-1} (z-\ensuremath{\gamma}_n)\label{eq:receivedY} \end{align} where first new channel zeros $\ensuremath{\beta}_{\ensuremath{l}}$ are added to the transmitted zeros $\ensuremath{\alpha}_{\ensuremath{k}}$ of ${\ensuremath{\mathrm X}}$, which then both will be perturbed by a noise polynomial. We will discuss the stability of such an approach later in \secref{sec:rootstability}. From the received signal coefficients, the zeros $\zeta_n$ of the received polynomial ${\ensuremath{\mathrm Y}}(z)$ can be computed using some \emph{root finding} algorithm. After assigning the received zeros $\ensuremath{\zeta}_n$ in the $K$ sectors $\ensuremath{\mathscr S}_k$, one can separate the data zero from its channel zeros by a \emph{minimum distance} decision \begin{align} \hat{m}_{\ensuremath{k}} = \begin{cases} 1,& \min_{\zeta_n \in \ensuremath{\mathscr S}_k} d(\zeta_n,\ensuremath{\alpha}_{\ensuremath{k}}^{(1)}) < \min_{\zeta_n \in\ensuremath{\mathscr S}_k} d(\zeta_n,\ensuremath{\alpha}_{\ensuremath{k}}^{(0)})\\ 0,& \text{else} \end{cases} \quad,\quad {\ensuremath{k}}=1,2,\dots,{\ensuremath{K}},\label{eq:mindist} \end{align} where $d(\cdot,\cdot)$ defines a certain metric on ${\ensuremath{\mathbb C}}$. We will call this a \emph{Root-Finding Minimal Distance} (RFMD) demodulator, see \figref{fig:BMOCZscheme}, where we used $d_{{\ensuremath{k}},n}^{(i)}=d(\zeta_n,\ensuremath{\alpha}_{\ensuremath{k}}^{(i)})$ for $i=0,1$. For simplicity, we will in this work only consider the (unweighted) Euclidean distance $d(x,y)=|x-y|$, but other distances might be more suitable, a point we will discuss in \secref{sec:rootstability}. The de/encoding or quantization sets $\Dset_{\ensuremath{k}}^{(i)}$ are the Voronoi cells of the zeros $\ensuremath{\alpha}_{\ensuremath{k}}^{(i)}$, leading to the best performance of the MD decoder \eqref{eq:mindist}. If the channel is scalar, no channel zeros are added, and the receiver has only to determine in which cells the received zeros fall to decode the data. If one cell contains multiple received zeros, the decoder will chose the smaller distance in \eqref{eq:mindist}, see \figref{fig:BMOCZscheme} where for ${\ensuremath{k}}=2$ the zero $\zeta_n$ is closer to $\ensuremath{\alpha}_2^{(0)}$ as $\zeta_{n+1}$ is to $\ensuremath{\alpha}_2^{(1)}$. \emph{Adaption for M-MOZ scheme:} For the M-MOZ scheme the transmitter will transmit one zero $\ensuremath{\alpha}_{\ensuremath{k}}\in\Alp_{\ensuremath{k}}$ for each sector $\ensuremath{\mathscr S}_{\ensuremath{k}}$. If no channel zeros are present, such as scalar channels, and only one received zero $\zeta_n$ is in the set $\Dset_{\ensuremath{k}}^{(m)}$, then the decoder will assume that $\ensuremath{\alpha}_{\ensuremath{k}}^{(m)}$ was transmitted. If multiple zeros are in one decoding set (channel zeros might be present), we will decide by minimum distance. The general decoding rule for the ${\ensuremath{k}}$th message $m_{\ensuremath{k}}$ is therefore \begin{align} \hat{m}_{\ensuremath{k}} = \argmin_{m\in\{1,\dots,M\}} \min_{\zeta_n\in\Dset_{{\ensuremath{k}}}^{(m)}} |\ensuremath{\alpha}_{\ensuremath{k}}^{(m)}-\zeta_n|. \end{align} If the ${\ensuremath{k}}$th sector $\ensuremath{\mathscr S}_{{\ensuremath{k}}}$ contains no zero at all, then ${m}_{{\ensuremath{k}}}$ can not be reliable decoded and will be in error. Here, multiple scenarios are possible, either one can chose the closest zero from the next neighbor sectors, as in \eqref{eq:mindist}, or one can request a retransmission for this message. See \figref{fig:mozscheme} for the general modulation and demodulation scheme. \begin{remark} Let us note, that the RFMD decoder can also detect potential bit errors, for example if in one cell multiple zeros occur, but no zeros in the next-neighbor sector. The encoding/decoding scheme is fundamentally different to classical coding schemes, since at the receiver we observe more zeros as we transmit, due to the channel. This can be seen as ISI in the zero-domain. \end{remark} \section{Huffman BMOCZ} The proposed BMOCZ scheme can be applied to any autocorrelation sequence ${\ensuremath{\mathbf a}} = {\ensuremath{\mathbf x}} * \cc{{\ensuremath{\mathbf x}}^-}\in{\ensuremath{\mathbb C}}^{2{\ensuremath{K}}+1}$ generating a polynomial with simple zeros, i.e., all zeros are distinct. \if imply multiplicity two. \fi Among all these autocorrelations, \emph{Huffman autocorrelations} \cite{Huf62} are the most impulsive ones given for a peak-to-side-lobe (PSL) $\eta\in (0,1/2)$ as \begin{align} a_k & = \begin{cases} -\eta ,& k=0,2{\ensuremath{K}}\\ 1,& k={\ensuremath{K}}\\ 0,& \text{else} \end{cases}. \end{align} Hence, the autocorrelation generates the polynomial \begin{align} {\ensuremath{\mathrm A}}(z)= -\eta + z^{{\ensuremath{K}}} -\eta z^{2{\ensuremath{K}}} \quad\text{and}\quad {\ensuremath{\mathrm A}}(e^{i2\pi\omega})=2\eta \cos(2\pi {\ensuremath{K}} \omega) -1. \Peter{} \label{eq:autohuf} \end{align} Since ${\ensuremath{\mathrm A}}(z)=0$ is a quadratic equation in $z^{\ensuremath{K}}$, solving for all zeros $\ensuremath{\alpha}_{\ensuremath{k}}=R_{\ensuremath{k}} e^{\im\ensuremath{\phi_{\ensuremath{k}}}}$, yields for the magnitude and phases \begin{align} R_{\ensuremath{k}}=R^{\pm 1}= \left(\frac{1 \pm \sqrt{1-4\eta^2}}{2\eta}\right)^{1/{\ensuremath{K}}}, \quad \ensuremath{\phi_{\ensuremath{k}}}= 2\pi\frac{{\ensuremath{k}}-1}{{\ensuremath{K}}}\ \text{ for }\ k=1,2,\dots,K. \end{align} This results in ${\ensuremath{K}}$ conjugate-reciprocal zero pairs uniformly placed on two circles with radii $R>1$ and $R^{-1}$ \begin{align} \ensuremath{\alpha}_{\ensuremath{k}}\in\ensuremath{\mathscr Z}_{\ensuremath{k}}=\{R e^{2\pi\im \frac{{\ensuremath{k}}-1}{{\ensuremath{K}}}}, R^{-1} e^{2\pi\im \frac{{\ensuremath{k}}-1}{{\ensuremath{K}}}}\}\ \text{ for }\ {\ensuremath{k}}=1,2\dots,{\ensuremath{K}}. \end{align} Since, the zeros are the vertices of two regular polygons, centered at the origin, they have the best pairwise distance from all autocorrelations, see \figref{fig:HuffmanBMOCZscheme}. Expressing the autocorrelation \eqref{eq:autohuf} in the $z-$domain by its zeros, gives \begin{align} {\ensuremath{\mathrm A}}(z) &=-\eta\prod_{{\ensuremath{k}}=1}^{\ensuremath{K}} (z-\ensuremath{\alpha}_{\ensuremath{k}}) (z-\cc{\ensuremath{\alpha}_{\ensuremath{k}}^{-1}}) = \underbrace{ x_{\ensuremath{K}}\prod_{{\ensuremath{k}}=1}^{\ensuremath{K}} (z-\ensuremath{\alpha}_{\ensuremath{k}}) }_{{\ensuremath{\mathrm X}}(z)}\cdot \underbrace{\cc{x_0} \prod_{{\ensuremath{k}}=1}^{\ensuremath{K}} (z-\cc{\ensuremath{\alpha}_{\ensuremath{k}}^{-1}})}_{{\ensuremath{\mathrm X}}^*(z)}\label{eq:autohufzdomain}, \end{align} where ${\ensuremath{\mathrm X}}^*(z)=\sum_k \cc{x_{{\ensuremath{K}}-k}} z^k$ is the \emph{conjugate-reciprocal polynomial} generated by $\cc{{\ensuremath{\mathbf x}}^-}$. Each ${\ensuremath{\mathrm X}}(z)$ respectively ${\ensuremath{\mathrm X}}^*(z)$ is then called a \emph{Huffman polynomial} and their coefficients a \emph{Huffman sequence}. Since the autocorrelation is constant for each selection $\ensuremath{\boldsymbol{ \alp}}\in{\ensuremath{\mathbb C}}^{\ensuremath{K}}$, the first and last coefficient of ${\ensuremath{\mathbf x}}$ depend on the chosen zeros, i.e., on the bit vector $\ensuremath{\mathbf m}=(m_1,\dots,m_K)$: \begin{align} &\cc{x_0} \cdot {x_{\ensuremath{K}}} \overset{\eqref{eq:autohufzdomain}}{=}-\eta\text{ and } x_{\ensuremath{K}}\cdot (-1)^{\ensuremath{K}}\prod_{\ensuremath{k}} \ensuremath{\alpha}_{\ensuremath{k}} =x_0 \quad \ensuremath{\Rightarrow} \quad |x_{\ensuremath{K}}|^2 = \frac{\eta}{(-1)^{{\ensuremath{K}}-1}\cc{\prod_{\ensuremath{k}}\ensuremath{\alpha}_{\ensuremath{k}}}} \\ &\quad\ensuremath{\Leftrightarrow} \quad x_{\ensuremath{K}} = e^{j\phi_0} \sqrt{\eta} R^{{\ensuremath{K}}/2-\Norm{\ensuremath{\mathbf m}}_1}, \ \ x_0 = e^{j\phi_0} \sqrt{\eta} R^{\Norm{\ensuremath{\mathbf m}}_1-{\ensuremath{K}}/2} \quad,\quad\phi_0\in[0,2\pi) \end{align} since we have \begin{align} \prod_{{\ensuremath{k}}=1}^{\ensuremath{K}} \ensuremath{\alpha}_{\ensuremath{k}}= \prod_{{\ensuremath{k}}=1}^{\ensuremath{K}} R^{2m_{\ensuremath{k}}-1}e^{j2\pi \frac{({\ensuremath{k}}-1)}{{\ensuremath{K}}}} = R^{2\Norm{\ensuremath{\mathbf m}}_1-{\ensuremath{K}}} e^{j2\pi \frac{\sum_{{\ensuremath{k}}=1}^{{\ensuremath{K}}-1} {\ensuremath{k}}}{{\ensuremath{K}}}} = R^{2\Norm{\ensuremath{\mathbf m}}_1-{\ensuremath{K}}} e^{j2\pi \frac{({\ensuremath{K}}-1){\ensuremath{K}}}{2{\ensuremath{K}}}} = R^{2\Norm{\ensuremath{\mathbf m}}_1-{\ensuremath{K}}}(-1)^{{\ensuremath{K}}-1}\notag. \end{align} By rewriting $\eta$ in terms of the radius we get \begin{align} \eta = \frac{1}{R^{\ensuremath{K}}+R^{-{\ensuremath{K}}}}. \end{align} If $\phi_0=0$, the first and last coefficients of ${\ensuremath{\mathbf x}}$ are given by \begin{align} x_{\ensuremath{K}}= \sqrt{\frac{R^{{\ensuremath{K}}-2\Norm{\ensuremath{\mathbf m}}_1}}{R^{\ensuremath{K}}+R^{-{\ensuremath{K}}}}} = \sqrt{ \frac{R^{-2\Norm{\ensuremath{\mathbf m}}_1}}{1+R^{-2{\ensuremath{K}}}}}\ \text{ and }\ x_0 =-\sqrt{\frac{R^{2\Norm{\ensuremath{\mathbf m}}_1}}{1+R^{2{\ensuremath{K}}}}}.\label{eq:hufflastfirst} \end{align} This suggest, that the first and last coefficients of Huffman sequences are dominant, which might help for a synchronisation and detection at the receiver. Furthermore, the free choice of the phase reflects the degree of freedom, we will lose in our modulation scheme. Note, we have $2{\ensuremath{K}}+2$ real parameters representing ${\ensuremath{K}}$ complex zeros and $1$ complex constant (trivial polynomial). The magnitude of this constant is determined by the PSL $\eta$, the bit vector $\ensuremath{\mathbf m}$, and the signal power. But its phase, acting as a global phase of the Huffman sequence, can not be resolved at the receiver at all, due to the last coefficient of the channel and the additive noise \eqref{eq:receivedY} and therefore has to be fixed for the scheme. Hence, the MOCZ scheme uses $2{\ensuremath{K}}+1$ of the $2{\ensuremath{K}}+2$ degrees of freedom. \begin{remark} Impulsive-equivalent autocorrelations are usually used to estimate the distance of objects, as used in radar, or to estimate the channel state, see for example \cite[Cha.12]{GG05}. By the best knowledge of the authors, the properties of Huffman sequences have never been used for a digital data communication. \Peter{Sure?} \end{remark} \ifextras \paragraph{Mean and Variance} Note, that the time-symbols $x_n$ of all Huffman sequences are not zero-mean nor independent from each other. For the last and first coefficient however we can compute their mean for $K$ even by \begin{align} \Expect{x_K}& =2^{-K} \sum_{i=1}^{2^K} x_K = 2^{-K} \frac{\sum_i R^{-\Norm{{\ensuremath{\mathbf b}}_i}_1}}{1+R^{-2K}} = 2^{-K}\frac{\sum_n^{K/2}( R^n 2^n + R^K R^{-n} 2^n )}{1+R^{-2K}}\\ &= \frac{2^{-K}}{1+R^{-2K}}\cdot \left(\sum_n (2R)^n + R^K\sum_n (2R^{-1})^n)\right)\\ &= \frac{2^{-K}}{1+R^{-2K}}\cdot \left(\frac{ 1-(2R)^{K/2+1}}{(1-2R)} + \frac{R^K-(2R)^{K/2} 2/R}{1-2/R}\right)\\ &= \frac{2^{-K}}{1+R^{-2K}}\cdot \left(\frac{ 1-(2R)^{K/2+1})(1-2/R) +(1-2R)( R^K-(2R)^{K/2} 2/R)}{5-2(R+R^{-1})}\right) \end{align} similar for the variance (\textcolor{red}{Do we need this ?}). \fi \section{Maximum Likelihood Receiver for BMOCZ} We shall derive now a much simpler and efficient demodulation technique for the BMOCZ scheme by using the fixed autocorrelation property of the codebook $\Calg$, represented by the autocorrelation matrix ${\ensuremath{\mathbf A}}\in{\ensuremath{\mathbb C}}^{{\ensuremath{K}}+1\times {\ensuremath{K}}+1}$, which is Hermitian Toeplitz and generated by ${\ensuremath{\mathbf a}}\in{\ensuremath{\mathbb C}}^{2{\ensuremath{K}}+1}$. If ${\ensuremath{L}}={\ensuremath{K}}+1$ then the autocorrelation matrix of ${\ensuremath{\mathbf x}}$ is given by the $N\times {\ensuremath{L}}$ banded Toeplitz matrix ${\ensuremath{\mathbf X}}$ generated by ${\ensuremath{\mathbf x}}$ as \begin{align} {\ensuremath{\mathbf A}}={\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf X}}= \begin{pmatrix} a_{\ensuremath{K}} & \dots & a_0 \\ \vdots &\diagdown & \vdots\\ a_{2{\ensuremath{K}}}& \dots & a_{\ensuremath{K}} \end{pmatrix} \quad,\quad {\ensuremath{\mathbf X}}= \begin{pmatrix} x_0 & 0 &\dots &0 & 0 \\ x_1 & x_0 &\dots & 0& 0\\ \vdots & \vdots &\diagdown & &\vdots\\ x_{\ensuremath{K}} & x_{{\ensuremath{K}}-1} &&& 0 \\ 0 & x_{{\ensuremath{K}}} &\dots & \diagdown & \\ \vdots & & \diagdown & &x_0\\ \vdots & & &\diagdown & \vdots\\ 0 & 0 &\dots &0 & x_{{\ensuremath{K}}}\end{pmatrix}\in{\ensuremath{\mathbb C}}^{N\times {\ensuremath{L}}}. \label{eq:AKlarge} \end{align} Note, that we can write the convolution in \eqref{eq:receivedsignal} with ${\ensuremath{\mathbf X}}$ in the vector-matrix notation as ${\ensuremath{\mathbf x}}*\vh={\ensuremath{\mathbf X}}\vh$. If ${\ensuremath{L}}<{\ensuremath{K}}+1$ then we cut out a ${\ensuremath{L}}\times {\ensuremath{L}}$ principal submatrix of ${\ensuremath{\mathbf A}}$ and for ${\ensuremath{L}}\geq {\ensuremath{K}}+1$ we extend ${\ensuremath{\mathbf A}}$ by adding zeros to the generating vector ${\ensuremath{\mathbf a}}$, i.e. \begin{align} {\ensuremath{\mathbf A}}_{\ensuremath{L}}\!=\! \begin{pmatrix} a_{\ensuremath{K}} & \dots & a_{{\ensuremath{K}}-{\ensuremath{L}}} \\ \vdots &\diagdown & \vdots\\ a_{{\ensuremath{K}}+{\ensuremath{L}}}& \dots & a_{\ensuremath{K}} \end{pmatrix} \text{for } {\ensuremath{L}}\!<\!{\ensuremath{K}}+1, \quad{\ensuremath{\mathbf A}}_{\ensuremath{L}}\!=\! \begin{pmatrix} a_{\ensuremath{K}} & \dots & a_0 & & \text{\kern-0.5em\smash{\raisebox{-1.5ex}{\huge 0}}}\\ \vdots &\diagdown &\vdots & \diagdown &\\ a_{2{\ensuremath{K}}} & \dots & a_{\ensuremath{K}} &\dots & a_0 \\ & \diagdown & \vdots & \diagdown &\vdots\\ \huge{\text{0}} & & a_{2{\ensuremath{K}}} & \dots & a_{\ensuremath{K}} \\ \end{pmatrix} \text{ for } {\ensuremath{L}}\!\geq\! {\ensuremath{K}}+1 \end{align} \if0 \begin{align} {\ensuremath{\mathbf A}}_{\ensuremath{L}} = {\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf X}}= \begin{cases} \begin{pmatrix} a_{\ensuremath{K}} & \dots & a_{{\ensuremath{K}}-{\ensuremath{L}}} \\ \vdots &\diagdown & \vdots\\ a_{{\ensuremath{K}}+{\ensuremath{L}}}& \dots & a_{\ensuremath{K}} \end{pmatrix} ,& {\ensuremath{L}}<{\ensuremath{K}}+1\\ \begin{pmatrix} a_{\ensuremath{K}} & \dots & a_0 & & \text{\kern-0.5em\smash{\raisebox{-1.5ex}{\huge 0}}}\\ \vdots &\diagdown &\vdots & \diagdown &\\ a_{2{\ensuremath{K}}} & \dots & a_L &\dots & a_0 \\ & \diagdown & \vdots & \diagdown &\vdots\\ \huge{\text{0}} & & a_{2{\ensuremath{K}}} & \dots & a_{\ensuremath{K}} \\ \end{pmatrix} , & {\ensuremath{L}}\geq {\ensuremath{K}}+1 \end{cases} \end{align} \fi In any case, the matrix ${\ensuremath{\mathbf A}}_{\ensuremath{L}}$ will be constant for any fixed Codebook $\Calg$. \if0 To utilize this property we have to extend the $N\times {\ensuremath{L}}$ Toeplitz matrix in \eqref{eq:X} for ${\ensuremath{L}}<{\ensuremath{K}}+1$ to an $N\times {\ensuremath{K}}+1$ Toeplitz matrix by setting $tNx={\ensuremath{K}}+1$ and adding ${\ensuremath{K}}+1-{\ensuremath{L}}$ more time-shifts of ${\ensuremath{\mathbf x}}$: \begin{align} \ensuremath{\tilde{\mathbf X}}= \begin{pmatrix} x_0 & & & \\ \vdots & \diagdown & & \text{\huge 0}& \\ x_{{\ensuremath{L}}-1} & \dots & x_0 & & \\ \vdots & &\vdots &\diagdown &\\ x_{\ensuremath{K}} & \dots & x_{{\ensuremath{K}}-{\ensuremath{L}}+1} & \dots & x_0 \\ & \diagdown &\vdots & &\vdots \\ \text{\huge{0}}&& x_{\ensuremath{K}} & \dots &x_{{\ensuremath{L}}-1} \end{pmatrix}\in{\ensuremath{\mathbb C}}^{N\times \tNx}\label{eq:tX}. \end{align} and also append ${\ensuremath{K}}+1-{\ensuremath{L}}$ zeros to the channel vector $\vh$ denoted by $\ensuremath{\tilde{\mathbf h}}$. If ${\ensuremath{L}}\geq {\ensuremath{K}}+1$ we set $\tNx={\ensuremath{L}}$. Hence the received signal \eqref{eq:receivedsignal} can be written as \begin{align} {\ensuremath{\mathbf y}}= \ensuremath{\tilde{\mathbf X}}\ensuremath{\tilde{\mathbf h}} + {\ensuremath{\mathbf w}}. \end{align} Then the autocorrelation matrix of the codebook is given for all ${\ensuremath{K}},{\ensuremath{L}}$ by \begin{align} {\ensuremath{\tilde{\mathbf A}}}=\ensuremath{\tilde{\mathbf X}}^*\ensuremath{\tilde{\mathbf X}}=\begin{cases} {\ensuremath{\mathbf A}} &, {\ensuremath{L}}<{\ensuremath{K}}+1\\ \begin{pmatrix} {\ensuremath{\mathbf A}} & {\ensuremath{\mathbf 0}}\\ {\ensuremath{\mathbf 0}} & {\ensuremath{\mathbf I}}_{{\ensuremath{L}}-{\ensuremath{K}}} \end{pmatrix} &, \text{else} \end{cases}\label{eq:tA} \end{align} \fi For multipath channels the \emph{maximum likelihood (sequence) detector} is known to be optimal \Peter{(was meinst Du damit und woher kommt das, quelle)} and is given by maximizing the conditional probability for each possible signal (codeword, sequence) ${\ensuremath{\mathbf x}}$ in the codebook \Peter{(marginals over the channel statistics ?)} \newcommand{\mathscr{H}}{\mathscr{H}} \begin{align} {\ensuremath{\hat{\mathbf x}}}= \arg\max_{{\ensuremath{\mathbf x}}\in\Calg} p({\ensuremath{\mathbf y}}|{\ensuremath{\mathbf x}})\label{eq:defml}. \end{align} By assumption \eqref{eq:channel} and \eqref{eq:noise} the channel and noise parameters are independent zero-mean Gaussian random variables, hence the received signal ${\ensuremath{\mathbf y}}$ is also a Gaussian random vector with mean zero and covariance matrix $\ensuremath{\mathbf R}_{{\ensuremath{\mathbf y}}}$, see \cite[(3.17)]{Mad08}. The conditional probability is therefore given by \begin{align} p({\ensuremath{\mathbf y}}|{\ensuremath{\mathbf x}}) = \frac{e^{-{\ensuremath{\mathbf y}}^*\ensuremath{\mathbf R}_{{\ensuremath{\mathbf y}}}^{-1}{\ensuremath{\mathbf y}}}}{\pi^{N}\det(\ensuremath{\mathbf R}_{{\ensuremath{\mathbf y}}})}\label{eq:gaussry}, \end{align} see \cite[Lem.3.B.1]{Kai00}. The covariance matrix of ${\ensuremath{\mathbf y}}$ is given by \begin{align} \ensuremath{\mathbf R}_{{\ensuremath{\mathbf y}}}&=\Expect{{\ensuremath{\mathbf y}}\vy^*}= \Expect{({\ensuremath{\mathbf X}}\vh+\vn)(\vh^*{\ensuremath{\mathbf X}}^*+\vn^*)}\\ &=\Expect{{\ensuremath{\mathbf X}}\vh\vh^*{\ensuremath{\mathbf X}}^*} + \underbrace{\Expect{{\ensuremath{\mathbf X}}\vh\vn^*}}_{=0}+ \underbrace{\Expect{\vn\vh^*{\ensuremath{\mathbf X}}^*}}_{=0}+ \Expect{\vn\vn^*} = {\ensuremath{\mathbf X}}\Expect{\vh\vh^*}{\ensuremath{\mathbf X}}^* +\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_N, \end{align} since $\vn$ and $\vh$ are independent zero-mean random variables. The discrete power delay profile \renewcommand{{\ensuremath{\mathbf p}}}{\ensuremath{\underline{\mathbf p}}} \begin{align} {\ensuremath{\mathbf p}}=(\pd^0,\pd^1,\dots,\pd^{{\ensuremath{L}}-1})\label{eq:pdp} \end{align} generates the channel covariance matrix $\Expect{\vh\vh^*}={\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}}$, which is a ${\ensuremath{L}}\times {\ensuremath{L}}$ diagonal matrix with diagonal ${\ensuremath{\mathbf p}}$. This gives for the covariance matrix \begin{align} \ensuremath{\mathbf R}_{{\ensuremath{\mathbf y}}}= \ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_N + {\ensuremath{\mathbf X}}{\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}}{\ensuremath{\mathbf X}}^*. \end{align} We will set $\vXp:={\ensuremath{\mathbf X}}{\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}}^{1/2}\in{\ensuremath{\mathbb C}}^{N\times {\ensuremath{L}}}$ such that \eqref{eq:defml} separates with \eqref{eq:gaussry} to \begin{align} \arg\max_{{\ensuremath{\mathbf x}}} p({\ensuremath{\mathbf y}}|{\ensuremath{\mathbf x}})&= \arg\min_{{\ensuremath{\mathbf x}}} -\log p({\ensuremath{\mathbf y}}|{\ensuremath{\mathbf x}}) \\ &= \arg\min_{{\ensuremath{\mathbf x}}} \Big(\underbrace{{\ensuremath{\mathbf y}}^*(\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_N + \vXp \vXp^*)^{-1} {\ensuremath{\mathbf y}}}_{\geq 0} + \log(\pi^N\det (\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_N + \vXp\vXp^*))\Big)\label{eq:MLdet} \end{align} where the log-function is monotone increasing and negative, since $p({\ensuremath{\mathbf y}}|{\ensuremath{\mathbf x}})<1$. By using \emph{Sylvester's determinant identity}, we get for the second summand in \eqref{eq:MLdet} by using that the autocorrelation, power delay profile ${\ensuremath{\mathbf p}}$ and noise power $\ensuremath{\sigma}$ is constant: \begin{align} \det(\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_N + \vXp\vXp^*)=\det(\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_{{\ensuremath{L}}} +\vXp^*\vXp) = \det(\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_{{\ensuremath{L}}} + {\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}}^{1/2} {\ensuremath{\mathbf A}}_{\ensuremath{L}}{\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}}^{1/2}) =\text{const.} \end{align} Hence, we can omit this term in \eqref{eq:MLdet}. By applying the \emph{Woodbury matrix identity}\footnote{Note, that ${\ensuremath{\mathbf I}}_{{\ensuremath{L}}}$ and ${\ensuremath{\mathbf I}}_N$ are non-singular, but not $\vXp$.}, see for example \cite[(0.7.4.1)]{HJ13}, we get \begin{align} {\ensuremath{\mathbf y}}^*(\ensuremath{\sigma}^{2}{\ensuremath{\mathbf I}}_N + \vXp {\ensuremath{\mathbf I}}_{{\ensuremath{L}}} \vXp^*)^{-1} {\ensuremath{\mathbf y}} &= {\ensuremath{\mathbf y}}^*(\ensuremath{\sigma}^{-2}{\ensuremath{\mathbf I}}_N - \ensuremath{\sigma}^{-2}\vXp({\ensuremath{\mathbf I}}_{{\ensuremath{L}}} + \ensuremath{\sigma}^{-2}\vXp^*\vXp)^{-1}\vXp^*\ensuremath{\sigma}^{-2}){\ensuremath{\mathbf y}} \\ &=\underbrace{\ensuremath{\sigma}^{-2}\Norm{{\ensuremath{\mathbf y}}}_2^2}_{=\text{const.}} -\ensuremath{\sigma}^{-2}{\ensuremath{\mathbf y}}^*\vXp(\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_{{\ensuremath{L}}} +\vXp^*\vXp)^{-1}\vXp^*{\ensuremath{\mathbf y}}. \end{align} Hence, the ML estimator simplifies to \begin{align} {\ensuremath{\hat{\mathbf x}}} & = \arg\max_{{\ensuremath{\mathbf x}}} {\ensuremath{\mathbf y}}^*\vXp(\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_{{\ensuremath{L}}} + \vXp^*\vXp)^{-1}\vXp^*{\ensuremath{\mathbf y}}. \end{align} Inserting the diagonal power delay profile matrix we get \begin{align} {\ensuremath{\hat{\mathbf x}}} & = \arg\max_{{\ensuremath{\mathbf x}}} {\ensuremath{\mathbf y}}^*{\ensuremath{\mathbf X}}{\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}}^{1/2}(\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_{{\ensuremath{L}}} + {\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}}^{1/2}{\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf X}}{\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}}^{1/2})^{-1}{\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}}^{1/2}{\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf y}}\\ &= \arg\max_{{\ensuremath{\mathbf x}}} {\ensuremath{\mathbf y}}^*{\ensuremath{\mathbf X}}(\underbrace{\ensuremath{\sigma}^2{\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}}^{-1} + {\ensuremath{\mathbf A}}_{\ensuremath{L}}}_{={\ensuremath{\mathbf B}}\succeq 0})^{-1}{\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf y}}, \label{eq:yxbxy} \end{align} where ${\ensuremath{\mathbf A}}_{\ensuremath{L}}\in{\ensuremath{\mathbb C}}^{{\ensuremath{L}}\times {\ensuremath{L}}}$ is given by \eqref{eq:AKlarge}. Since the matrix ${\ensuremath{\mathbf B}}\in{\ensuremath{\mathbb C}}^{{\ensuremath{L}}\times{\ensuremath{L}}}$ is constant and reflects the codebook, power delay profile, and noise power it acts as a weighting for the projections of ${\ensuremath{\mathbf y}}$ to the shifted codewords. We will call this decoder the \emph{Maximum Likelihood} (ML) decoder: \begin{align} {\ensuremath{\hat{\mathbf x}}} = \arg\max_{{\ensuremath{\mathbf x}}} \Norm{{\ensuremath{\mathbf B}}^{-1/2}{\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf y}}}_2^2 \label{eq:MLdecoder}. \end{align} \if0 \begin{align} \ensuremath{\mathbf R}_{{\ensuremath{\mathbf y}}}&=\Expect{{\ensuremath{\mathbf y}}\vy^*}= \Expect{(\ensuremath{\tilde{\mathbf X}}\ensuremath{\tilde{\mathbf h}}+\vn)(\ensuremath{\tilde{\mathbf h}}^*\ensuremath{\tilde{\mathbf X}}^*+\vn^*)}\\ &=\Expect{\ensuremath{\tilde{\mathbf X}}\ensuremath{\tilde{\mathbf h}}\vth^*\ensuremath{\tilde{\mathbf X}}^*} + \underbrace{\Expect{\ensuremath{\tilde{\mathbf X}}\ensuremath{\tilde{\mathbf h}}\vn^*}}_{=0}+ \underbrace{\Expect{\vn\ensuremath{\tilde{\mathbf h}}^*\ensuremath{\tilde{\mathbf X}}^*}}_{=0}+ \Expect{\vn\vn^*} = {\ensuremath{\mathbf X}}\Expect{\ensuremath{\tilde{\mathbf h}}\vth^*}{\ensuremath{\mathbf X}}^* +\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_N \end{align} % since $\vn$ and $\ensuremath{\tilde{\mathbf h}}$ are independent and zero-mean. The discrete power delay profile ${\ensuremath{\mathbf p}}=(\pd^0,\pd^1,\dots,\pd^{{\ensuremath{L}}-1})$ generates the channel covariance matrix $\Expect{\ensuremath{\tilde{\mathbf h}}\vth^*}={\ensuremath{\mathbf D}}_{\ensuremath{\tilde{\vp}}}$ of the zero-padded channel $\ensuremath{\tilde{\mathbf h}}$, which is a $\tNx\times \tNx$ diagonal matrix with diagonal $\ensuremath{\tilde{\vp}}=[{\ensuremath{\mathbf p}},{\ensuremath{\mathbf 0}}]\in{\ensuremath{\mathbb C}}^{\tNx}$. This gives for the covariance matrix % \begin{align} \ensuremath{\mathbf R}_{{\ensuremath{\mathbf y}}}= \ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_N + \ensuremath{\tilde{\mathbf X}}{\ensuremath{\mathbf D}}_{\ensuremath{\tilde{\vp}}}\ensuremath{\tilde{\mathbf X}}^*. \end{align} % We will set $\vtXtp:=\ensuremath{\tilde{\mathbf X}}{\ensuremath{\mathbf D}}_{\ensuremath{\tilde{\vp}}}^{1/2}\in{\ensuremath{\mathbb C}}^{N\times \tNx}$ such that \eqref{eq:defml} separates with \eqref{eq:gaussry} to % \begin{align} \arg\max_{{\ensuremath{\mathbf x}}} p({\ensuremath{\mathbf y}}|{\ensuremath{\mathbf x}})&= \arg\min_{{\ensuremath{\mathbf x}}} -\log p({\ensuremath{\mathbf y}}|{\ensuremath{\mathbf x}}) \\ &= \arg\min_{{\ensuremath{\mathbf x}}} \Big(\underbrace{{\ensuremath{\mathbf y}}^*(\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_N + \vtXtp \vtXtp^*)^{-1} {\ensuremath{\mathbf y}}}_{\geq 0} + \log(\pi^N\det (\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_N + \vtXtp\vtXtp^*))\Big)\label{eq:MLdet} \end{align} % where the log-function is monotone increasing and negative, since $p({\ensuremath{\mathbf y}}|{\ensuremath{\mathbf x}})<1$. By using \emph{Sylvester's determinant identity}, we get for the second summand in \eqref{eq:MLdet} % \begin{align} \det(\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_N + \vtXtp\vtXtp^*)=\det(\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_{\tNx} +\vtXtp^*\vtXtp) \overset{\eqref{eq:tA}}{=} \det(\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_{\tNx} + {\ensuremath{\mathbf D}}_{\ensuremath{\tilde{\vp}}}^{1/2} {\ensuremath{\tilde{\mathbf A}}}{\ensuremath{\mathbf D}}_{\ensuremath{\tilde{\vp}}}^{1/2}) =\text{const.} \end{align} % Hence, we can omit this term in \eqref{eq:MLdet}, since the power delay profile ${\ensuremath{\mathbf p}}$ and noise power $\ensuremath{\sigma}$ is also constant. % By applying the \emph{Woodbury matrix identity}\footnote{Note, that ${\ensuremath{\mathbf I}}_{\tNx}$ and ${\ensuremath{\mathbf I}}_N$ are non-singular, but not $\vtXtp$.}, see for example \cite[(0.7.4.1)]{HJ13}, we get % \begin{align} {\ensuremath{\mathbf y}}^*(\ensuremath{\sigma}^{2}{\ensuremath{\mathbf I}}_N + \vtXtp {\ensuremath{\mathbf I}}_{\tNx} \vtXtp^*)^{-1} {\ensuremath{\mathbf y}} &= {\ensuremath{\mathbf y}}^*(\ensuremath{\sigma}^{-2}{\ensuremath{\mathbf I}}_N - \ensuremath{\sigma}^{-2}\vtXp({\ensuremath{\mathbf I}}_{\tNx} + \ensuremath{\sigma}^{-2}\vtXtp^*\vtXtp)^{-1}\vtXtp^*\ensuremath{\sigma}^{-2}){\ensuremath{\mathbf y}} \\ &=\underbrace{\ensuremath{\sigma}^{-2}\Norm{{\ensuremath{\mathbf y}}}_2^2}_{=\text{const.}} -\ensuremath{\sigma}^{-2}{\ensuremath{\mathbf y}}^*\vtXtp(\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_{\tNx} +\vtXtp^*\vtXtp)^{-1}\vtXtp^*{\ensuremath{\mathbf y}}. \end{align} % Hence, the ML estimator simplifies to % \begin{align} {\ensuremath{\hat{\mathbf x}}} & = \arg\max_{{\ensuremath{\mathbf x}}} {\ensuremath{\mathbf y}}^*\vtXtp(\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_{\tNx} + \vtXtp^*\vtXtp)^{-1}\vtXtp^*{\ensuremath{\mathbf y}}. \end{align} % For ${\ensuremath{L}}<{\ensuremath{K}}+1$, we can derive by setting $\vtXp=\ensuremath{\tilde{\mathbf X}}\vtD^{1/2}\in{\ensuremath{\mathbb C}}^{N\times {\ensuremath{L}}}$ with $\vtD^T=\begin{pmatrix} {\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}} & {\ensuremath{\mathbf O}}\end{pmatrix}$ the matrix % \begin{align} (\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_{{\ensuremath{K}}} + \vtXtp^*\vtXtp)^{-1}= \begin{pmatrix} \ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_{\ensuremath{L}}+\vtXp^*\vtXp & {\ensuremath{\mathbf O}}\\ {\ensuremath{\mathbf O}} & \ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_{{\ensuremath{K}}-{\ensuremath{L}}}\\ \end{pmatrix}^{-1} = \begin{pmatrix} (\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_{\ensuremath{L}}+\vtXp^*\vtXp)^{-1} & {\ensuremath{\mathbf O}}\\ {\ensuremath{\mathbf O}} &\ensuremath{\sigma}^{-2}{\ensuremath{\mathbf I}}_{{\ensuremath{K}}-{\ensuremath{L}}}\\ \end{pmatrix} \end{align} % and hence we only need the ${\ensuremath{L}}\times {\ensuremath{L}}$ block, which yields to: % \begin{align} {\ensuremath{\hat{\mathbf x}}}&= \arg\max_{{\ensuremath{\mathbf x}}} {\ensuremath{\mathbf y}}^*{\ensuremath{\mathbf X}}{\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}}^{1/2}(\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_{{\ensuremath{L}}} + \vtXp^*\vtXp)^{-1}{\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}}^{1/2}{\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf y}}\\ {\ensuremath{\hat{\mathbf x}}}&= \arg\max_{{\ensuremath{\mathbf x}}} {\ensuremath{\mathbf y}}^*{\ensuremath{\mathbf X}}[{\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}}^{-1/2}(\ensuremath{\sigma}^2{\ensuremath{\mathbf I}}_{{\ensuremath{L}}} + \vtXp^*\vtXp){\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}}^{-1/2} ]^{-1}{\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf y}}\\ &= \arg\max_{{\ensuremath{\mathbf x}}} {\ensuremath{\mathbf y}}^*{\ensuremath{\mathbf X}}(\underbrace{\ensuremath{\sigma}^2{\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf p}}}^{-1} + {\ensuremath{\mathbf A}}_{\ensuremath{L}}}_{={\ensuremath{\mathbf B}}\succeq 0})^{-1}{\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf y}} =\arg\max_{{\ensuremath{\mathbf x}}} \Norm{{\ensuremath{\mathbf B}}^{-1/2}{\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf y}}}_2^2 \label{eq:yxbxy} \end{align} % where ${\ensuremath{\mathbf A}}_{\ensuremath{L}}=({\ensuremath{\mathbf I}}_{\ensuremath{L}}\ {\ensuremath{\mathbf O}}) {\ensuremath{\mathbf A}} ({\ensuremath{\mathbf I}}_{\ensuremath{L}}\ {\ensuremath{\mathbf O}})^T$ is the ${\ensuremath{L}}\times {\ensuremath{L}}$ principal submatrix of ${\ensuremath{\mathbf A}}$ and hence constant. If ${\ensuremath{L}}\geq {\ensuremath{K}}+1$ we get the $\tNx\times \tNx$ matrix % \begin{align} \vtB=\ensuremath{\sigma}^2 {\ensuremath{\mathbf D}}_{\ensuremath{\tilde{\vp}}}^{-1} + {\ensuremath{\tilde{\mathbf A}}} \end{align} % which is also constant and we can retrieve the optimal ${\ensuremath{\mathbf x}}$ by an exhaustive search over the codebook. Note, that the power delay profile and noise power weights the shifted codewords. We will call this decoder the \emph{Maximum Likelihood} (ML) decoder: % \begin{align} {\ensuremath{\hat{\mathbf x}}} = \begin{cases} \arg\max_{{\ensuremath{\mathbf x}}} \Norm{{\ensuremath{\mathbf B}}^{-1/2}{\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf y}}}_2^2, & {\ensuremath{L}}<{\ensuremath{K}}\\ \arg\max_{{\ensuremath{\mathbf x}}} \Norm{\vtB^{-1/2}\ensuremath{\tilde{\mathbf X}}^*{\ensuremath{\mathbf y}}}_2^2 ,& {\ensuremath{L}}\geq {\ensuremath{K}} \end{cases}\label{eq:MLdecoder}. \end{align} % \fi \Peter{essentially this is the {\em matched filter}, isn't it ?} Note, the ML reduces for ${\ensuremath{L}}=1$ to the \emph{correlation receiver} \begin{align} \arg\max_{{\ensuremath{\mathbf x}}} |{\ensuremath{\mathbf x}}^*{\ensuremath{\mathbf y}}|^2 \end{align} see for example \cite[Sec.4.2-2]{PS08}. Since the codebook has cardinality $2^{\ensuremath{K}}$ and ${\ensuremath{\mathbf x}}\in{\ensuremath{\mathbb C}}^{{\ensuremath{K}}+1}$ the scheme is non-orthogonal for ${\ensuremath{K}}\geq 2$. If ${\ensuremath{L}}<{\ensuremath{K}}+1$ and the codebook are the Huffman sequences, then ${\ensuremath{\mathbf A}}_{\ensuremath{L}}={\ensuremath{\mathbf I}}_{\ensuremath{L}}$ and ${\ensuremath{\mathbf B}}={\ensuremath{\mathbf D}}_{\ensuremath{\mathbf b}}$ becomes a diagonal matrix with ${\ensuremath{\mathbf b}}=\ensuremath{\sigma}^2{\ensuremath{\mathbf p}}^{-1}+{\ensuremath{\mathbf 1}}_{\ensuremath{L}}$. Hence, we end up with a Rake receiver, where the weights for the ${\ensuremath{l}}$th fingers (correlators) are given by $b_{\ensuremath{l}}^{-1}=(p^{\ensuremath{l}}+\ensuremath{\sigma}^2)/\ensuremath{\sigma}^2$, which reflects the sum power of channel gain and signal to noise ratio of the ${\ensuremath{l}}$th path. \subsection{Direct Zero Testing Decoder for Huffman BMOCZ}\label{sec:dizet} Huffman sequences not only allow a simple encoding by its zeros, \Peter{(how is this working..filterbank...)} but also a simple decoding, since the autocorrelation are by design the most impulsive-like autocorrelations of any sequence ${\ensuremath{\mathbf x}}$. We set $\ensuremath{{\boldsymbol \eta}}=(\underbrace{0,\dots,0}_{{\ensuremath{K}}},\eta,\eta,\dots,\eta)^T\in{\ensuremath{\mathbb C}}^{{\ensuremath{L}}}$ for ${\ensuremath{L}}\geq {\ensuremath{K}}+1$ and get by \eqref{eq:autohuf} the autocorrelation matrix \begin{align} {\ensuremath{\mathbf A}}_{\ensuremath{L}}={\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf X}} &= \begin{cases} {\ensuremath{\mathbf I}}_{{\ensuremath{L}}}, & {\ensuremath{L}}<{\ensuremath{K}}+1\\ {\ensuremath{\mathbf I}}_{{\ensuremath{L}}} - \ensuremath{{\boldsymbol \eta}}\ve_1^* -\ve_1\ensuremath{{\boldsymbol \eta}}^*, & {\ensuremath{L}}\geq {\ensuremath{K}}+1 \end{cases} \end{align} Let us consider the case ${\ensuremath{L}}<{\ensuremath{K}}+1$, then the matrix ${\ensuremath{\mathbf B}}$ becomes \begin{align} {\ensuremath{\mathbf B}}={\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf b}}} = {\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf b}}}{\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf X}}. \end{align} If and only if ${\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf b}}}=b{\ensuremath{\mathbf I}}_{\ensuremath{L}}$ for some $b\not=0$ we can identify in \eqref{eq:yxbxy} the orthogonal projector on ${\ensuremath{\mathbf X}}$ \begin{align} \ensuremath{\mathbf P }} % Vektorwertiges Ma{\ss =b^{-1} {\ensuremath{\mathbf X}}({\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf X}})^{-1}{\ensuremath{\mathbf X}}^* \end{align} and obtain with the left null space ${\ensuremath{\mathbf V}}$ of ${\ensuremath{\mathbf X}}$ the identity \begin{align} \ensuremath{\mathbf P }} % Vektorwertiges Ma{\ss = {\ensuremath{\mathbf I}}_N - {\ensuremath{\mathbf V}}({\ensuremath{\mathbf V}}^*{\ensuremath{\mathbf V}})^{-1}{\ensuremath{\mathbf V}}^*. \end{align} Let us define the ${\ensuremath{K}}\times N$ \emph{generalized Vandermonde matrix} generated by the complex-conjugated zeros $\ensuremath{\alpha}_1,\dots, \ensuremath{\alpha}_{\ensuremath{K}}$ of ${\ensuremath{\mathrm X}}(z)$ \begin{align} \ensuremath{\vV_{\!\valp}}^*= \begin{pmatrix} 1 & \ensuremath{\alpha}_1 & \ensuremath{\alpha}_1^2 & \dots &\ensuremath{\alpha}_1^{N-1} \\ 1 & \ensuremath{\alpha}_2 & \ensuremath{\alpha}_2^2 & \dots &\ensuremath{\alpha}_2^{N-1} \\ \vdots & & &\vdots\\ 1 & \ensuremath{\alpha}_{{\ensuremath{K}}} & \ensuremath{\alpha}_{\ensuremath{K}}^2 & \dots &\ensuremath{\alpha}_{\ensuremath{K}}^{N-1}\\ \end{pmatrix}. \end{align} Since each zero is distinct, the Vandermonde matrix has full rank ${\ensuremath{K}}$. Then, each complex-conjugated column is in the left null space of the matrix ${\ensuremath{\mathbf X}}$. More precisely we get \begin{align} \ensuremath{\vV_{\!\valp}}^*{\ensuremath{\mathbf X}}=\begin{pmatrix} {\ensuremath{\mathrm X}}(\ensuremath{\alpha}_1) & \ensuremath{\alpha}_1 {\ensuremath{\mathrm X}}(\ensuremath{\alpha}_1) & \dots & \ensuremath{\alpha}_1^{{\ensuremath{L}}-1}{\ensuremath{\mathrm X}}(\ensuremath{\alpha}_1)\\ {\ensuremath{\mathrm X}}(\ensuremath{\alpha}_2) & \ensuremath{\alpha}_2 {\ensuremath{\mathrm X}}(\ensuremath{\alpha}_2) & \dots & \ensuremath{\alpha}_2^{{\ensuremath{L}}-1}{\ensuremath{\mathrm X}}(\ensuremath{\alpha}_2)\\ \vdots & &\vdots \\ {\ensuremath{\mathrm X}}(\ensuremath{\alpha}_{\ensuremath{K}}) & \ensuremath{\alpha}_{\ensuremath{K}} {\ensuremath{\mathrm X}}(\ensuremath{\alpha}_{\ensuremath{K}}) & \dots & \ensuremath{\alpha}_{\ensuremath{K}}^{{\ensuremath{L}}-1}{\ensuremath{\mathrm X}}(\ensuremath{\alpha}_{\ensuremath{K}})\\ \end{pmatrix}={\ensuremath{\mathbf O}} \quad\ensuremath{\Leftrightarrow} \quad {\ensuremath{\mathbf X}}^*\ensuremath{\vV_{\!\valp}}={\ensuremath{\mathbf O}} \end{align} In fact, the dimension of the left null space of ${\ensuremath{\mathbf X}}$ (null space of ${\ensuremath{\mathbf X}}^*$) is exactly ${\ensuremath{K}}$ for each ${\ensuremath{\mathbf X}}$ generated by ${\ensuremath{\mathbf x}}\in\Calg$, since it holds $N={\ensuremath{L}}+{\ensuremath{K}}=\rank({\ensuremath{\mathbf X}}^*)+\nullity({\ensuremath{\mathbf X}}^*)$, where $\rank({\ensuremath{\mathbf X}})=\rank({\ensuremath{\mathbf X}}^*)={\ensuremath{L}}$ and the shifts of ${\ensuremath{\mathbf x}}$ are all linear independent for any ${\ensuremath{\mathbf x}}\not={\ensuremath{\mathbf 0}}$. Hence, we get \begin{align} {\ensuremath{\mathbf y}}^*{\ensuremath{\mathbf X}}({\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf X}})^{-1}{\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf y}}= {\ensuremath{\mathbf y}}^*({\ensuremath{\mathbf I}}_N-\ensuremath{\vV_{\!\valp}}(\ensuremath{\vV_{\!\valp}}^*\ensuremath{\vV_{\!\valp}})^{-1}\ensuremath{\vV_{\!\valp}}^*){\ensuremath{\mathbf y}} =\underbrace{\Norm{{\ensuremath{\mathbf y}}}^2}_{=\text{const}>0}- \Norm{(\ensuremath{\vV_{\!\valp}}^*\ensuremath{\vV_{\!\valp}})^{-1/2}\ensuremath{\vV_{\!\valp}}^*{\ensuremath{\mathbf y}}}^2\label{eq:yxxy} \end{align} which yields with the mixing matrix $\ensuremath{\mathbf M}_{\ensuremath{\boldsymbol{ \alp}}}=({\ensuremath{\mathbf V}}_{\ensuremath{\boldsymbol{ \alp}}}^*{\ensuremath{\mathbf V}}_{\ensuremath{\boldsymbol{ \alp}}})^{-1/2}\in{\ensuremath{\mathbb C}}^{{\ensuremath{K}}\times {\ensuremath{K}}}$ to \begin{align} \arg\max_{\ensuremath{\boldsymbol{ \alp}}} p({\ensuremath{\mathbf y}}|{\ensuremath{\mathbf x}}(\ensuremath{\boldsymbol{ \alp}})) = \arg\min_{\ensuremath{\boldsymbol{ \alp}}} \Norm{({\ensuremath{\mathbf V}}_{\ensuremath{\boldsymbol{ \alp}}}^*{\ensuremath{\mathbf V}}_{\ensuremath{\boldsymbol{ \alp}}})^{-1/2} {\ensuremath{\mathbf V}}_{\ensuremath{\boldsymbol{ \alp}}}^*{\ensuremath{\mathbf y}}}^2 = \arg\min_{\ensuremath{\boldsymbol{ \alp}}}\Norm{\ensuremath{\mathbf M}_{\ensuremath{\boldsymbol{ \alp}}}^{-1/2}{\ensuremath{\mathbf V}}_{\ensuremath{\boldsymbol{ \alp}}}^*{\ensuremath{\mathbf y}}}^2. \end{align} For Huffman zeros we have $\ensuremath{\alpha}_{\ensuremath{k}}=R_{\ensuremath{k}} e^{i2\pi ({\ensuremath{k}}-1)/{\ensuremath{K}}}$ with $R_{\ensuremath{k}}\in\{R,R^{-1}\}$ and we get \begin{align} {\ensuremath{\mathbf V}}_{\ensuremath{\boldsymbol{ \alp}}}^*{\ensuremath{\mathbf V}}_{{\ensuremath{\boldsymbol{ \alp}}}} = \begin{pmatrix} 1 & \ensuremath{\alpha}_1 & \dots & \ensuremath{\alpha}_1^{N-1}\\ \vdots & &&\\ 1 & \ensuremath{\alpha}_{\ensuremath{K}} & \dots & \ensuremath{\alpha}_{\ensuremath{K}}^{N-1} \end{pmatrix} \begin{pmatrix} 1 & \dots & 1 \\ \cc{\ensuremath{\alpha}_1} & \dots &\cc{\ensuremath{\alpha}_{\ensuremath{K}}}\\ \vdots & &\vdots\\ \cc{\ensuremath{\alpha}_1^{N-1}} & \dots &\cc{\ensuremath{\alpha}_{\ensuremath{K}}^{N-1}} \end{pmatrix} =\begin{pmatrix} c_{1,1} & c_{1,2} & \dots & c_{1,{\ensuremath{K}}}\\ c_{2,1} & c_{2,2} &\dots & c_{2,{\ensuremath{K}}} \\ \vdots && \ddots & \vdots\\ c_{{\ensuremath{K}},1} & c_{{\ensuremath{K}},2} &\dots & c_{{\ensuremath{K}},{\ensuremath{K}}} \end{pmatrix}.\label{eq:C} \end{align} With the geometric series we get \begin{align} c_{{\ensuremath{k}},m} &= \sum_{n=0}^{N-1} (\ensuremath{\alpha}_{\ensuremath{k}} \cc{\ensuremath{\alpha}}_m)^n = \sum_{n=0}^{N-1}(R_{\ensuremath{k}} R_m)^n e^{\im 2\pi n\frac{({\ensuremath{k}}-m)}{{\ensuremath{K}}}}\\ c_{{\ensuremath{k}},{\ensuremath{k}}} & =c_{\ensuremath{k}}=\sum_{n=0}^{N-1} |\ensuremath{\alpha}_{\ensuremath{k}}|^{2n} = \frac{1-R_{\ensuremath{k}}^{2N}}{1-R_{\ensuremath{k}}^2}. \end{align} In expectation, for uniform bit sequences, we get $\Expect{R_{\ensuremath{k}} R_m}\simeq 1$ for ${\ensuremath{k}}\not=m$ and hence for $N={\ensuremath{l}}{\ensuremath{K}}$ the off diagonals are roughly vanishing, since $\sum_{n=0}^{{\ensuremath{K}}-1} e^{\im 2\pi n(k-m)/{\ensuremath{K}}}=0 $. Hence, we approximate \eqref{eq:C} as a diagonal matrix, which leads to \begin{align} \ensuremath{\mathbf M}_{\ensuremath{\boldsymbol{ \alp}}}^{-1/2}\simeq \operatorname{diag}\left( \sqrt{\frac{1-|\ensuremath{\alpha}_1|^{2}}{1-|\ensuremath{\alpha}_1|^{2N}}},\dots, \sqrt{\frac{1-|\ensuremath{\alpha}_{\ensuremath{K}}|^{2}}{1-|\ensuremath{\alpha}_{\ensuremath{K}}|^{2N}}}\right)\label{eq:gwfactor}. \end{align} By observing \begin{align} {\ensuremath{\mathbf V}}_{\ensuremath{\boldsymbol{ \alp}}}^*{\ensuremath{\mathbf y}}=\begin{pmatrix}{\ensuremath{\mathrm Y}}(\ensuremath{\alpha}_1) & \dots & {\ensuremath{\mathrm Y}}(\ensuremath{\alpha}_{\ensuremath{K}})\end{pmatrix}^T, \end{align} the exhaustive search of the ML simplifies to independent decisions for each zero symbol \begin{align} \hat{\ensuremath{\alpha}_{\ensuremath{k}}}:=\argmin_{\ensuremath{\alpha}_{\ensuremath{k}}\in \{ R,R^{-1}\} e^{\im 2\pi\frac{{\ensuremath{k}}-1}{{\ensuremath{K}}}}} \Big|\sqrt{\frac{1-|\ensuremath{\alpha}_{\ensuremath{k}}|^{2}}{1-|\ensuremath{\alpha}_{\ensuremath{k}}|^{2(N-1)}}} {\ensuremath{\mathrm Y}}(\ensuremath{\alpha}_{\ensuremath{k}})\Big|. \end{align} This gives the \emph{Direct Zero Testing} (DiZeT) decoding rule for $k\in\{1,\dots,K\}$ \begin{align} b_{\ensuremath{k}} = \begin{cases} 1, & |{\ensuremath{\mathrm Y}}(R e^{\im 2\pi \frac{{\ensuremath{k}}-1}{{\ensuremath{K}}}})| < R^{N-2}| {\ensuremath{\mathrm Y}}(R^{-1} e^{\im 2\pi\frac{{\ensuremath{k}}-1}{{\ensuremath{K}}}})|\\ 0, &\text{else} \end{cases}\label{eq:dizetdecoder} \end{align} since it holds for the \emph{geometrical weights} (GW) \begin{align} \sqrt{\frac{1-R^{2(N-1)}}{1-R^{-2(N-1)}}\frac{1-R^{-2}}{1-R^2}}= \sqrt{(-R^{2N-2})\cdot(-R^{-2})}=R^{N-2}. \end{align} If ${\ensuremath{L}}\geq {\ensuremath{K}}+1$ we will approximate ${\ensuremath{\mathbf A}}_{\ensuremath{L}}\simeq {\ensuremath{\mathbf I}}_{\ensuremath{L}}$. Then the same approximation yield to the same DiZeT decoder. \subsection{FFT-Implementation of Huffman BMOCZ-decoding} In fact, the DiZeT decoder for Huffman sequences allows also a simple hardware implementation at the receiver. If we scale the received samples $y_n$ with the positive radius $R^n$ and resp. $R^{-n}$, i.e., \begin{align} {\ensuremath{\mathbf D}}_R{\ensuremath{\mathbf y}}:=\begin{pmatrix} 1 & 0 & \dots &0\\ 0 & R & \dots &0\\ \vdots & & \ddots& \vdots\\ 0 & 0&\dots & R^{N-1}\end{pmatrix} {\ensuremath{\mathbf y}} \end{align} and apply the $N-$point DFT matrix if ${\ensuremath{L}}=(t-1){\ensuremath{K}}$ for $t\in{\ensuremath{\mathbb N}}$, yielding to $N=tK$, we get the samples of the $z-$transform \begin{align} {\ensuremath{\mathbf F}^*} {\ensuremath{\mathbf D}}_R{\ensuremath{\mathbf y}} = \begin{pmatrix}\sum_{n=0}^{N-1} y_n R^n e^{i2\pi 0\cdot n/N}\\ \vdots\\ \sum_n y_n R^n e^{i2\pi (N-1)\cdot k/N}\end{pmatrix}=:{\ensuremath{\mathrm Y}}(\ensuremath{\boldsymbol{ \alp}}_t^{(1)})\quad,\quad {\ensuremath{\mathbf F}^*} {\ensuremath{\mathbf D}}_{R^{-1}} {\ensuremath{\mathbf y}} = {\ensuremath{\mathrm Y}}(\ensuremath{\boldsymbol{ \alp}}_t^{(0)}). \end{align} Then the decoding rule \eqref{eq:dizetdecoder} becomes % \begin{align} b_{{\ensuremath{k}}} = \begin{cases} 1&, |({\ensuremath{\mathbf F}}{\ensuremath{\mathbf D}}_R{\ensuremath{\mathbf y}})_{tk}|< R^{N-2}|({\ensuremath{\mathbf F}^*}{\ensuremath{\mathbf D}}_{R^{-1}} {\ensuremath{\mathbf y}})_{tk}|\\ 0 &, \text{ else} \end{cases}. \end{align} Hence, the decoder can be fully implemented by a simple $N-$point DFT from the delayed amplified received signal, by using for example FPGA or even analog front-ends. \if0 \paragraph{Similar analysis for $c_1$ term} we can use for the matrices ${\ensuremath{\mathbf x}}_1$ and ${\ensuremath{\mathbf x}}_{\ensuremath{L}}$ similar arguments by identifying the $N-1\times N$ left-null space as \newcommand{\ensuremath{\tilde{{\ensuremath{\mathbf V}}}^*_{\ensuremath{\boldsymbol{ \alp}},1}}}{\ensuremath{\tilde{{\ensuremath{\mathbf V}}}^*_{\ensuremath{\boldsymbol{ \alp}},1}}} \newcommand{\ensuremath{\tilde{{\ensuremath{\mathbf V}}}^*_{\ensuremath{\boldsymbol{ \alp}},{\ensuremath{L}}}}}{\ensuremath{\tilde{{\ensuremath{\mathbf V}}}^*_{\ensuremath{\boldsymbol{ \alp}},{\ensuremath{L}}}}} \begin{align} \ensuremath{\tilde{{\ensuremath{\mathbf V}}}^*_{\ensuremath{\boldsymbol{ \alp}},1}} = \begin{pmatrix} 1 & \dots &\ensuremath{\alpha}_1^{\ensuremath{K}} & \dots & \ensuremath{\alpha}_1^{N-1}\\ \vdots & && &\vdots\\ 1 & \dots & \ensuremath{\alpha}_{\ensuremath{K}}^{\ensuremath{K}} & \dots & \ensuremath{\alpha}_{\ensuremath{K}}^{N-1}\\ 0 & \dots & 1 &\dots & 0 \\ \vdots & & &\diagdown&\vdots\\ 0 & \dots & 0 &\dots & 1 \end{pmatrix} \quad,\quad \ensuremath{\tilde{{\ensuremath{\mathbf V}}}^*_{\ensuremath{\boldsymbol{ \alp}},{\ensuremath{L}}}} = \begin{pmatrix} 1 &\dots & 0 & \dots &0 \\ &\diagdown&\vdots & &\vdots & \\ 0 &\dots & 1 & \dots &0 \\ 1 & \dots &\ensuremath{\alpha}_1^{N-1-{\ensuremath{K}}} & \dots & \ensuremath{\alpha}_1^{N-1}\\ \vdots & && &\vdots\\ 1 & \dots & \ensuremath{\alpha}_{\ensuremath{K}}^{N-1-{\ensuremath{K}}} & \dots & \ensuremath{\alpha}_{\ensuremath{K}}^{N-1} \end{pmatrix} \end{align} where $\tilde{{\ensuremath{\mathbf V}}}^*_{\ensuremath{\boldsymbol{ \alp}},1} {\ensuremath{\mathbf x}}_1={\ensuremath{\mathbf 0}}=\ensuremath{\tilde{{\ensuremath{\mathbf V}}}^*_{\ensuremath{\boldsymbol{ \alp}},{\ensuremath{L}}}}{\ensuremath{\mathbf x}}_{\ensuremath{L}}$. With the same argumentation we get \begin{align} c_1 {\ensuremath{\mathbf y}}^*({\ensuremath{\mathbf x}}_1{\ensuremath{\mathbf x}}_1^* + {\ensuremath{\mathbf x}}_{\ensuremath{L}}{\ensuremath{\mathbf x}}_{\ensuremath{L}}^*){\ensuremath{\mathbf y}} = c_1 {\ensuremath{\mathbf y}}^* {\ensuremath{\mathbf x}}_1({\ensuremath{\mathbf x}}_1^*{\ensuremath{\mathbf x}}_1)^{-1}{\ensuremath{\mathbf x}}_1^* + {\ensuremath{\mathbf x}}_{\ensuremath{L}}({\ensuremath{\mathbf x}}_{\ensuremath{L}}^*{\ensuremath{\mathbf x}}_{\ensuremath{L}})^{-1}{\ensuremath{\mathbf x}}_{\ensuremath{L}}^*){\ensuremath{\mathbf y}} = c_1 {\ensuremath{\mathbf y}}^* ({\ensuremath{\mathbf I}}-\ensuremath{\tilde{{\ensuremath{\mathbf V}}}^*_{\ensuremath{\boldsymbol{ \alp}},1}}) \end{align} since ${\ensuremath{\mathbf x}}_1$ and ${\ensuremath{\mathbf x}}_{\ensuremath{L}}$ are normalized. The exact M{\ensuremath{K}} decoder is given by \begin{align} \arg\max_{{\ensuremath{\mathbf x}}}= a_1 2\Re\{\skprod{{\ensuremath{\mathbf y}}}{{\ensuremath{\mathbf x}}_{\ensuremath{L}}}\skprod{{\ensuremath{\mathbf x}}_1}{{\ensuremath{\mathbf y}}}\}+ c_1 (|\skprod{{\ensuremath{\mathbf x}}_1}{{\ensuremath{\mathbf y}}}|^2+|\skprod{{\ensuremath{\mathbf x}}_{\ensuremath{L}}}{{\ensuremath{\mathbf y}}}|^2) \end{align} \todoend \if0 \subsection{Least-Square Minimization in Time Domain} Since, the ML is independent of the channel, as we just showed, but since the ML with Gaussian random variables is equivalent to the least-square (LS) minimization \begin{align} {\ensuremath{\hat{\mathbf x}}}=\min_{{\ensuremath{\mathbf x}}} \Norm{{\ensuremath{\mathbf y}}-{\ensuremath{\mathbf X}}\vh} \end{align} There should be also exists a version in time-domain which performs as good as the ML with full channel knowledge. Let us start from the exact expression \eqref{eq:yxbxy} and inserting ${\ensuremath{\mathbf B}}^{-1}$ from \eqref{eq:Binv} \begin{align} {\ensuremath{\hat{\mathbf x}}} &= \arg\max_{{\ensuremath{\mathbf x}}} {\ensuremath{\mathbf y}}^*{\ensuremath{\mathbf X}}{\ensuremath{\mathbf B}}^{-1}{\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf y}} = \arg\max_{{\ensuremath{\mathbf x}}} {\ensuremath{\mathbf y}}^* {\ensuremath{\mathbf X}} {\ensuremath{\mathbf I}} {\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf y}} + {\ensuremath{\mathbf y}}^* {\ensuremath{\mathbf X}}\frac{a^2}{1-a^2} {\ensuremath{\mathbf C}}_a {\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf y}}\\ &= \arg\max \left(\sum_{k=0}^{{\ensuremath{L}}-1} |\skprod{\ensuremath{\mathbf T}^k{\ensuremath{\mathbf x}}}{{\ensuremath{\mathbf y}}}|^2 + \frac{a^2}{1-a^2} {\ensuremath{\mathbf y}}^*{\ensuremath{\mathbf X}} [a^{-1}(\ve_1\ve_{\ensuremath{L}}^* + \ve_{\ensuremath{L}} \ve_1^*) + (\ve_1\ve_1^* + \ve_{\ensuremath{L}}\ve_{\ensuremath{L}}^*)]{\ensuremath{\mathbf X}}^*{\ensuremath{\mathbf y}}\right)\\ &=\arg\max \left(\sum_{k=0}^{{\ensuremath{L}}-1} |\skprod{\ensuremath{\mathbf T}^k{\ensuremath{\mathbf x}}}{{\ensuremath{\mathbf y}}}|^2 + \frac{a^2}{1-a^2}[ |\skprod{\ensuremath{\mathbf T}^{{\ensuremath{L}}-1}{\ensuremath{\mathbf x}}}{{\ensuremath{\mathbf y}}}|^2 +|\skprod{{\ensuremath{\mathbf x}}}{{\ensuremath{\mathbf y}}}|^2 + a^{-1}\skprod{\ensuremath{\mathbf T}^{{\ensuremath{L}}-1}{\ensuremath{\mathbf x}}}{{\ensuremath{\mathbf y}}}\skprod{{\ensuremath{\mathbf y}}}{{\ensuremath{\mathbf x}}} + a^{-1}\skprod{{\ensuremath{\mathbf x}}}{{\ensuremath{\mathbf y}}}{\skprod{{\ensuremath{\mathbf y}}}{\ensuremath{\mathbf T}^{{\ensuremath{L}}-1}{\ensuremath{\mathbf x}}}}\right) \\ &=\arg\max \left(\sum_{k=0}^{{\ensuremath{L}}-1} |\skprod{\ensuremath{\mathbf T}^k{\ensuremath{\mathbf x}}}{{\ensuremath{\mathbf y}}}|^2 + \frac{a^2}{1-a^2}[ |\skprod{\ensuremath{\mathbf T}^{\ensuremath{L}}{\ensuremath{\mathbf x}}}{{\ensuremath{\mathbf y}}}|^2 +|\skprod{{\ensuremath{\mathbf x}}}{{\ensuremath{\mathbf y}}}|^2 + 2a^{-1}\Re\{\skprod{\ensuremath{\mathbf T}^{\ensuremath{L}}{\ensuremath{\mathbf x}}}{{\ensuremath{\mathbf y}}}\skprod{{\ensuremath{\mathbf y}}}{{\ensuremath{\mathbf x}}}\} ]\right)\\ &=\arg\max \frac{1}{1-a^2}\left( |\skprod{\ensuremath{\mathbf T}^{{\ensuremath{L}}-1}{\ensuremath{\mathbf x}}}{{\ensuremath{\mathbf y}}}|^2 +|\skprod{{\ensuremath{\mathbf x}}}{{\ensuremath{\mathbf y}}}|^2 + 2a\Re\{\skprod{\ensuremath{\mathbf T}^{{\ensuremath{L}}-1}{\ensuremath{\mathbf x}}}{{\ensuremath{\mathbf y}}}\skprod{{\ensuremath{\mathbf y}}}{{\ensuremath{\mathbf x}}}\} \right) +\sum_{k=1}^{{\ensuremath{L}}-2} |\skprod{\ensuremath{\mathbf T}^k{\ensuremath{\mathbf x}}}{{\ensuremath{\mathbf y}}}|^2 \label{eq:MLexact} \end{align} % Observation Simulation: If $a=0$, then the exact ML estimator \eqref{eqMLexact} is the same as the Vandermonde ML estimator. \fi \if0 If $R_{\ensuremath{k}}=R$ for all $l\in\{1,\dots,{\ensuremath{K}}\}$, then ${\ensuremath{\mathbf C}}$ is a circulant matrix generated by \begin{align} {\ensuremath{\mathbf c}}=(1-R^{4{\ensuremath{K}}})\begin{pmatrix} (1-R^2)^{-1} \\ (1-R^2 e^{i2\pi /{\ensuremath{K}}})^{-1} \\ \vdots \\ (1-R^{2} e^{i2\pi ({\ensuremath{K}}-1)/{\ensuremath{K}}})^{-1}\end{pmatrix} \end{align} Hence we get \begin{align} {\ensuremath{\mathbf A}}= {\ensuremath{\mathbf F}^*} \operatorname{diag}({\ensuremath{\mathbf F}} {\ensuremath{\mathbf c}})^{-1/2} {\ensuremath{\mathbf F}} \end{align} \fi \subsection{Expectation of the mixing matrix} What is the expected eigenvalues of ${\ensuremath{\mathbf C}}_{\ensuremath{\boldsymbol{ \alp}}}$, i.e., what is $\Expect{\ensuremath{\lambda}({\ensuremath{\mathbf C}}_{\ensuremath{\boldsymbol{ \alp}}})}$ over all $\ensuremath{\boldsymbol{ \alp}}\in \ensuremath{\mathscr Z}$? In fact, let us take a look at proof of the Gershgorin Theorem. Assume $\ensuremath{\lambda}$ is an eigenvalue of ${\ensuremath{\mathbf C}}_{\ensuremath{\boldsymbol{ \alp}}}$ and ${\ensuremath{\mathbf x}}$ be an eigenvector. We can scale the eigenvector by its largest component $|x_{n}|=\max_k|x_k|$ as $\ensuremath{\tilde{\mathbf x}}={\ensuremath{\mathbf x}}/|x_{n}|$ such that the $\ensuremath{\tilde{x}}_n =1$ and $|\ensuremath{\tilde{x}}_m|\leq 1$ for $n\not=m$. Then it holds from ${\ensuremath{\mathbf C}}_{\ensuremath{\boldsymbol{ \alp}}}\ensuremath{\tilde{\mathbf x}}= \ensuremath{\lambda} \ensuremath{\tilde{\mathbf x}}$ that \begin{align} &({\ensuremath{\mathbf C}}_{\ensuremath{\boldsymbol{ \alp}}}\ensuremath{\tilde{\mathbf x}})_n = \sum_{m} {c_{n,m}}\ensuremath{\tilde{x}}_m = \ensuremath{\lambda} \ensuremath{\tilde{x}}_n=\ensuremath{\lambda} = \sum_{m\not=n} c_{n,m}\ensuremath{\tilde{x}}_m + c_{n,n}\ensuremath{\tilde{x}}_n = \sum_{m\not=n} c_{n,m}\ensuremath{\tilde{x}}_m + c_{n,n}\\ \ensuremath{\Leftrightarrow} & |\ensuremath{\lambda} -c_{n,n} | = \max\{\ensuremath{\lambda}-c_{n,n},c_{n,n}-lam\} =\max\{\pm\sum_{n\not=m} c_{n,m} x_m\} \end{align} since $\ensuremath{\lambda}$ and $c_{n,n}$ are real-valued. Taking the expectation over all possible $\ensuremath{\boldsymbol{ \alp}}$, which are i.i.d. uniform and asuming that $x_m$ are independent from $c_{n,m}$ we get \begin{align} \Expect{|\ensuremath{\lambda}-c_{n,n}|} = \max\{\pm \sum_m \Expect{c_{n,m}} \Expect{x_m}\} \leq \max\{ \pm \sum_m \Expect{c_{n,m}} \end{align} since $x_m$ The expectation of $c_{n,m}$ is given by if $N= s{\ensuremath{K}}$ for some $s\in{\ensuremath{\mathbb N}}$: \begin{align} \Expect{c_{n,m}}& = \sum_{k} \Expect{R_n^k R_m^k} e^{i2\pi k\frac{n-m}{{\ensuremath{K}}}} \\ & = \sum_{k} \Expect{R_n^k}\Expect{ R_m^k} e^{i2\pi k\frac{n-m}{{\ensuremath{K}}}} \\ & = \frac{1}{4} \sum_{k} \underbrace{(R^{-2k}+2+R^{2k})}_{=:c_k} e^{i2\pi k\frac{n-m}{{\ensuremath{K}}}}\label{eq:ck}\\ &= \frac{1}{4} \left( \frac{1-R^{-2N}}{1-R^{-2}e^{i2\pi k(n-m)/{\ensuremath{K}}}} +\frac{1-R^{2N}}{1-R^{2}e^{i2\pi k(n-m)/{\ensuremath{K}}}} \right)\\ &=\frac{1}{4} \left( \frac{1-R^{-2N} -R^{2N} +R^{-2N+2} e^{i\phi} +R^{2N-2} e^{i\phi} -R^{2}e^{i\phi} -R^{-2}e^{i\phi}}{ 1 + e^{2i\phi} -R^2 e^{i\phi} -R^{-2} e^{i\phi}}\right) \end{align} since $\Expect{R_n^k}= \frac{R^{-k} + R^{k}}{2}$ and $\sum_k e^{i2\pi k (n-m)/{\ensuremath{K}}}$ for $n\not=m$. Hm, is not working. It is not possible to conclude that $\Expect{|\ensuremath{\lambda}-c_{n,n}|}\ll 1$ \subsection{Expectation of ${\ensuremath{\mathbf C}}$} Let us compute the Expectation of the matrix ${\ensuremath{\mathbf C}}_{\ensuremath{\boldsymbol{ \alp}}}$ \begin{align} {\ensuremath{\mathbf C}}= \Expect{{\ensuremath{\mathbf C}}_{\ensuremath{\boldsymbol{ \alp}}}}= \Expect{\begin{pmatrix} c_{1,1} & c_{1,2} & \dots & c_{1,{\ensuremath{K}}}\\ c_{2,1} & c_{2,2} &\dots & c_{2,{\ensuremath{K}}} \\ \vdots && \ddots & \vdots\\ c_{{\ensuremath{K}},1} & c_{{\ensuremath{K}},2} &\dots & c_{{\ensuremath{K}},{\ensuremath{K}}} \end{pmatrix}}\overset{\eqref{eq:ck}}{=}\begin{pmatrix} \hat{{\ensuremath{\mathbf c}}}_{0} & \hat{{\ensuremath{\mathbf c}}}_{{\ensuremath{K}}-1} & \cdot & \hat{{\ensuremath{\mathbf c}}}_1\\ \hat{{\ensuremath{\mathbf c}}}_{1} & \hat{{\ensuremath{\mathbf c}}}_{0} & \cdot & \hat{{\ensuremath{\mathbf c}}}_2\\ \vdots &&&\vdots\\ \hat{{\ensuremath{\mathbf c}}}_{{\ensuremath{K}}-1} & \hat{{\ensuremath{\mathbf c}}}_{{\ensuremath{K}}-2} & \cdot & \hat{{\ensuremath{\mathbf c}}}_0\\ \end{pmatrix} \end{align} Which is clearly a circulant matrix generated by $\hat{{\ensuremath{\mathbf c}}}$. Hence, the DFT allows one to diagonalize \begin{align} {\ensuremath{\mathbf C}}&= {\ensuremath{\mathbf F}} \operatorname{diag}({\ensuremath{\mathbf F}^*}\hat{{\ensuremath{\mathbf c}}}){\ensuremath{\mathbf F}^*} = {\ensuremath{\mathbf F}} \operatorname{diag}({\ensuremath{\mathbf c}}){\ensuremath{\mathbf F}^*} \intertext{and therefore we get for the mixing matrix} {\ensuremath{\mathbf A}}&= {\ensuremath{\mathbf F}}_{\ensuremath{K}} \operatorname{diag}({\ensuremath{\mathbf c}})^{-1/2} {\ensuremath{\mathbf F}^*}_{\ensuremath{K}} \intertext{and hence} \Norm{{\ensuremath{\mathbf A}}{\ensuremath{\mathbf y}}_{\ensuremath{\boldsymbol{ \alp}}}}_2^2&= \Norm{ \operatorname{diag}({\ensuremath{\mathbf c}})^{-1/2} {\ensuremath{\mathbf F}^*}_{\ensuremath{K}}\{ {\ensuremath{\mathbf f}}_{l,N} \operatorname{diag}(1,R_{\ensuremath{k}},R_{\ensuremath{k}}^2,\dots,R_{\ensuremath{k}}^{N-1})\}{\ensuremath{\mathbf y}}} \end{align} \fi \section{SDP decoder for BMOCZ via Channel Autocorrelation} As already mention above, noncoherent communication of information bearing signals having very short length, in the order of the maximum delay spread of the multipath propagation, is indeed related to the blind deconvolution problem. This bilinear inverse problem itself suffers from a rich set of nontrivial ambiguities and impossible to solve without further constraints. Therefore, one of the challenges in communications and the motivation for our approach, is to develop simple, fast, and efficient methods by restricting the class of data signals, in this case to finite codebooks. In this section we give a brief overview on a convex method for solving this problem for impulsive data signals, as in the case of Huffman sequences. In \cite{JH16} one of the authors introduced a semi-definite program to deconvolve up to global phase almost all signals ${\ensuremath{\mathbf y}}={\ensuremath{\mathbf x}}*\vh\in{\ensuremath{\mathbb C}}^N$ from the knowledge of the autocorrelations ${\ensuremath{\mathbf a}}_{x}$ and ${\ensuremath{\mathbf a}}_{h}$. This program is successful if the polynomials corresponding to the input signals ${\ensuremath{\mathbf x}}$ and $\vh$ do not share a common zero. Later, in \cite{WJPH17}, a stable deconvolution via the SDP over the reals has been proven. In a nutshell, using the idea of lifting one can express the bilinear problem as a linear estimation problem, i.e., to recover a positive semi-definite matrix $0 \ensuremath{\preceq}{\ensuremath{\mathbf Z}} \in{\ensuremath{\mathbb C}}^{\ensuremath{\tilde{N}}\times \ensuremath{\tilde{N}}}$ where $\ensuremath{\tilde{N}}=L+K+1$ (details see \cite{WJPH17}). In the case of circular convolutions and with signals in random and incoherent subspaces this has been investigated first in \cite{ARR12}. Let ${\ensuremath{\mathcal{A}}}:{\ensuremath{\mathbb C}}^{\ensuremath{\tilde{N}}\times \ensuremath{\tilde{N}}} \to {\ensuremath{\mathbb C}}^{4\ensuremath{\tilde{N}}-4}$ the linear map representing here the (non-circular) convolution. As discussed in \cite{WJPH17} a stable deconvolution can be performed by minimizing the least-square error: \begin{align} \hat{{\ensuremath{\mathbf Z}}}=\arg\min_{{\ensuremath{\mathbf Z}}\succeq 0} \Norm{{\ensuremath{\mathbf b}}- {\ensuremath{\mathcal{A}}}({\ensuremath{\mathbf Z}})}_2^2 \quad\text{ where } \quad{\ensuremath{\mathbf b}}=({\ensuremath{\mathbf a}}_x,{\ensuremath{\mathbf a}}_{h}, {\ensuremath{\mathbf y}}, \cc{{\ensuremath{\mathbf y}}^-})^T \label{eq:sdpdenoising} \end{align} Performing a rank-one projection of the minimizer $\hat{{\ensuremath{\mathbf Z}}}$ using the singular value decomposition yields then the rank-one matrix ${\ensuremath{\hat{\mathbf z}}}\hvz^*$ which reveals the vector ${\ensuremath{\hat{\mathbf z}}}=e^{\im\phi} [{\ensuremath{\hat{\mathbf x}}}, {\ensuremath{\hat{\mathbf h}}}]\in{\ensuremath{\mathbb C}}^{\ensuremath{\tilde{N}}}$ up to a global phase $\phi$. Thus, this method can be indeed used for blind deconvolution in a communication context when the ${\ensuremath{\mathbf a}}_x$ and ${\ensuremath{\mathbf a}}_h$ are known. In the following we discuss how this is possible and why impulsive-like data signals, such as Huffman sequences, are helpful. \subsection{Estimation of the Channel Autocorrelation, Noiseless case} Let us start, for ease of exposition, with the noiseless case, i.e., ${\ensuremath{\mathbf y}}={\ensuremath{\mathbf x}}*\vh$ where ${\ensuremath{\mathbf x}}\in{\ensuremath{\mathbb C}}^{{\ensuremath{K}}+1}$ and $\vh\in{\ensuremath{\mathbb C}}^{{\ensuremath{L}}}$. From the Wiener-Khintchine relation we get \begin{align} {\ensuremath{\mathbf a}}_y={\ensuremath{\mathbf y}}*\cc{{\ensuremath{\mathbf y}}^-}= ({\ensuremath{\mathbf x}}*\vh)*(\cc{{\ensuremath{\mathbf x}}^{-}}*\cc{\vh^{-}})={\ensuremath{\mathbf a}}_x*{\ensuremath{\mathbf a}}_h. \end{align} Assume that ${\ensuremath{\mathbf a}}_x$ is already known and the receiver computes ${\ensuremath{\mathbf a}}_y={\ensuremath{\mathbf y}}*\cc{{\ensuremath{\mathbf y}}^-}$ from the received signal ${\ensuremath{\mathbf y}}$. If the relation above can be solved for ${\ensuremath{\mathbf a}}_h$, given ${\ensuremath{\mathbf a}}_x$ and ${\ensuremath{\mathbf a}}_y$, we can indeed use the methodology and the convex program \eqref{eq:sdpdenoising} for estimating $({\ensuremath{\mathbf x}},\vh)$. To this end, we consider this in the Fourier-domain by zero-padding the sequences ${\ensuremath{\mathbf a}}_x$ and ${\ensuremath{\mathbf a}}_h$ to dimension $M=2N-1$ giving the vectors $\ensuremath{\tilde{\mathbf a}}_x$ and $\ensuremath{\tilde{\mathbf a}}_h$. Thus, if ${\ensuremath{\mathbf F}} \ensuremath{\tilde{\mathbf a}}_x$ has no zeros, we get: \begin{align} {\ensuremath{\mathbf F}} {\ensuremath{\mathbf a}}_y \bullet (\sqrt{M}{\ensuremath{\mathbf F}} \ensuremath{\tilde{\mathbf a}}_x)^{-1}= \sqrt{M} {\ensuremath{\mathbf F}}_{M} \ensuremath{\tilde{\mathbf a}}_h \ensuremath{\bullet} {\ensuremath{\mathbf F}}_M \ensuremath{\tilde{\mathbf a}}_{x} \bullet(\sqrt{M}{\ensuremath{\mathbf F}}_M \ensuremath{\tilde{\mathbf a}}_x)^{-1}={\ensuremath{\mathbf F}} \ensuremath{\tilde{\mathbf a}}_h, \label{eq:Fy} \end{align} and the autocorrelation of the channel can be obtained by: \begin{align} \ensuremath{\tilde{\mathbf a}}_h= {\ensuremath{\mathbf F}^*} ( {\ensuremath{\mathbf F}} \ensuremath{\tilde{\mathbf a}}_y \bullet (\sqrt{M}{\ensuremath{\mathbf F}} \ensuremath{\tilde{\mathbf a}}_x)^{-1}) \end{align} as long as $({\ensuremath{\mathbf F}} \ensuremath{\tilde{\mathbf a}}_x)_k \not=0$, which holds by design of the Huffman sequences. Removing from $\ensuremath{\tilde{\mathbf a}}_h$ the last $M-(2{\ensuremath{L}}-1)$ zeros reveals finally the channel autocorrelation ${\ensuremath{\mathbf a}}_h$. \if0 Here we sample in $M=2N-1$ and zero pad the vectors accordingly. Since the receiver knows ${\ensuremath{\mathbf a}}$ due to the knowledge of $\ensuremath{\boldsymbol{ \alp}}$, the dimensions ${\ensuremath{K}},{\ensuremath{L}}$, and the fact that the codewords are normalized, he computes \begin{align} &\ensuremath{\tilde{\mathbf x}}_0= {\ensuremath{\mathcal Z}}^{-1}\Big(\prod_{{\ensuremath{k}}=1}^{\ensuremath{K}} (z-\ensuremath{\alpha_{\ensuremath{k}}})\Big)\\ \ensuremath{\Rightarrow} \quad&{\ensuremath{\mathbf x}}_0= {\ensuremath{\mathbf x}}_0/\Norm{\ensuremath{\tilde{\mathbf x}}_0}\\ \ensuremath{\Rightarrow} \quad&{\ensuremath{\mathbf a}}= {\ensuremath{\mathbf x}}_0*\cc{{\ensuremath{\mathbf x}}_0^-} \end{align} First, we need the following identity for any ${\ensuremath{\mathbf x}}\in{\ensuremath{\mathbb C}}^N$ \begin{align} {\ensuremath{\mathbf a}}_x={\ensuremath{\mathbf x}}*\cc{{\ensuremath{\mathbf x}}^-}= \begin{pmatrix}{\ensuremath{\mathbf x}}\\ {\ensuremath{\mathbf 0}}_{N-1}\end{pmatrix} \circledast\begin{pmatrix} \cc{{\ensuremath{\mathbf x}}^-} \\ {\ensuremath{\mathbf 0}}_{N-1}\end{pmatrix} \end{align} \todostart Applying the unitary Fourier Transform we get with the circular convolution theorem for $M=2N-1$ \begin{align} {\ensuremath{\mathbf F}}_M (\begin{pmatrix}0 \\ {\ensuremath{\mathbf x}} \\ {\ensuremath{\mathbf 0}}_{N-1}\end{pmatrix} \circledast \ensuremath{\mathbf{\Gam}}_M \begin{pmatrix} 0 \\ {\ensuremath{\mathbf x}} \\ {\ensuremath{\mathbf 0}}_{N-1}\end{pmatrix}) = {\ensuremath{\mathbf F}} \begin{pmatrix}0 \\ {\ensuremath{\mathbf x}} \\ {\ensuremath{\mathbf 0}}_{N-1}\end{pmatrix} \ensuremath{\bullet} {\ensuremath{\mathbf F}} \ensuremath{\mathbf{\Gam}}\begin{pmatrix}0 \\ {\ensuremath{\mathbf x}} \\ {\ensuremath{\mathbf 0}}_{N-1}\end{pmatrix} = {\ensuremath{\mathbf F}} \begin{pmatrix}0 \\ {\ensuremath{\mathbf x}} \\ {\ensuremath{\mathbf 0}}_{N-1}\end{pmatrix}\ensuremath{\bullet} {\ensuremath{\mathbf F}^*} \begin{pmatrix}0 \\ {\ensuremath{\mathbf x}} \\ {\ensuremath{\mathbf 0}}_{N-1}\end{pmatrix}= |{\ensuremath{\mathbf F}} \begin{pmatrix}0 \\ {\ensuremath{\mathbf x}} \\ {\ensuremath{\mathbf 0}}_{N-1}\end{pmatrix}|^2 \end{align} Hence, to apply the {\tt fft} in MatLab, we need to circular shift the autocorrelation by $N$ dimension, i.e., \begin{align} \ensuremath{\mathbf S }} % Vektorwertiges Ma{\ss^{N}_M {\ensuremath{\mathbf a}}_x = \ensuremath{\mathbf S }} % Vektorwertiges Ma{\ss^{N}\begin{pmatrix}{\ensuremath{\mathbf x}}\\ {\ensuremath{\mathbf 0}}_{N-1}\end{pmatrix} \circledast\begin{pmatrix} \cc{{\ensuremath{\mathbf x}}^-} \\ {\ensuremath{\mathbf 0}}_{N-1}\end{pmatrix}= \begin{pmatrix}0 \\ {\ensuremath{\mathbf x}}\\ {\ensuremath{\mathbf 0}}_{N-2}\end{pmatrix} \circledast\begin{pmatrix} 0 \\ {\ensuremath{\mathbf 0}}_{N-2} \\\cc{{\ensuremath{\mathbf x}}^-} \end{pmatrix} =\begin{pmatrix}0 \\ {\ensuremath{\mathbf x}}\\ {\ensuremath{\mathbf 0}}_{N-2}\end{pmatrix} \circledast\ensuremath{\mathbf{\Gam}} \begin{pmatrix}0 \\ {\ensuremath{\mathbf x}}\\ {\ensuremath{\mathbf 0}}_{N-2}\end{pmatrix} \end{align} If the zeros $\ensuremath{\boldsymbol{ \alp}}$ are not on lying on the unit circle we have for ${\ensuremath{K}}_1={\ensuremath{K}}+1$ \begin{align} {\ensuremath{\mathbf d}}={\ensuremath{\mathbf F}}\ensuremath{\mathbf S }} % Vektorwertiges Ma{\ss_M^N \begin{pmatrix}{\ensuremath{\mathbf a}}_x \\ {\ensuremath{\mathbf 0}}_{M-(2{\ensuremath{K}}_1-1)-1}\end{pmatrix} ={\ensuremath{\mathbf F}}\ensuremath{\mathbf S }} % Vektorwertiges Ma{\ss_M^N \begin{pmatrix}{\ensuremath{\mathbf x}}_0*\cc{{\ensuremath{\mathbf x}}_0^-} \\ {\ensuremath{\mathbf 0}}_{2N-2{\ensuremath{K}}_1-1}\end{pmatrix} ={\ensuremath{\mathbf F}}\ensuremath{\mathbf S }} % Vektorwertiges Ma{\ss_M^N \begin{pmatrix} {\ensuremath{\mathbf x}}_0 \\ {\ensuremath{\mathbf 0}}_{N-{\ensuremath{K}}_1}\end{pmatrix}* \begin{pmatrix} \cc{{\ensuremath{\mathbf x}}_0^-} \\ {\ensuremath{\mathbf 0}}_{N-{\ensuremath{K}}_1-1}\end{pmatrix} =\Big|{\ensuremath{\mathbf F}} \begin{pmatrix} 0 \\ {\ensuremath{\mathbf x}}_0\\ {\ensuremath{\mathbf 0}}_{M-2{\ensuremath{K}}_1}\end{pmatrix}\Big|^2> 0 \end{align} note, we have to split the convolution in two different dimensions, since $2N-2{\ensuremath{K}}_1-1$ is always odd. $c_k\not=0$ and hence we can multiply their inverse pointwise to \eqref{eq:Fy} and by applying the convolution theorem again we finally yield \begin{align} {\ensuremath{\mathbf F}}({\ensuremath{\mathbf y}}*\cc{{\ensuremath{\mathbf y}}^-})&={\ensuremath{\mathbf F}}({\ensuremath{\mathbf a}}{\ensuremath{\mathbf F}}({\ensuremath{\mathbf x}}*\vh*\cc{({\ensuremath{\mathbf x}}*\vh)^-})={\ensuremath{\mathbf F}}({\ensuremath{\mathbf a}}_x*{\ensuremath{\mathbf a}}_h)\\ &=|{\ensuremath{\mathbf F}} {\ensuremath{\mathbf a}}_x|^2\ensuremath{\bullet}|{\ensuremath{\mathbf F}} {\ensuremath{\mathbf a}}_h|^2 = {\ensuremath{\mathbf F}}(|{\ensuremath{\mathbf F}} {\ensuremath{\mathbf a}}|^2\ensuremath{\bullet} |{\ensuremath{\mathbf F}} \vh|^2 \ensuremath{\bullet} |{\ensuremath{\mathbf d}}|^{-1}) = {\ensuremath{\mathbf a}}_h\label{eq:ahestimation} \end{align} \fi \subsection{Estimation of the Channel Autocorrelation Estimation, Noisy case} When computing ${\ensuremath{\mathbf a}}_y$ in the presence of noise, ${\ensuremath{\mathbf y}}={\ensuremath{\mathbf x}}*\vh + {\ensuremath{\mathbf w}}$, we encounter additional cross-correlations and the estimate is affected by coloured noise: \begin{align} {\ensuremath{\mathbf w}}_c= \vn*\cc{{\ensuremath{\mathbf x}}^{-}}*\cc{\vh^-} + \cc{\vn^-}*{\ensuremath{\mathbf x}}*\vh + {\ensuremath{\mathbf a}}_w\label{eq:colorednoise}. \end{align} where ${\ensuremath{\mathbf a}}_w={\ensuremath{\mathbf w}}*\cc{{\ensuremath{\mathbf w}}^-}$. Obviously, this stage can be improved by, e.g., LMMSE estimation (Wiener filter). \Peter{Ok - ich glaube, das geht besser - zumindest fuer Comm/SigPro Journal muss man hier was machen} Nevertheless, let us compute a scaling estimate for the method above. Repeating the steps above gives: \begin{align} \tilde{\hat{{\ensuremath{\mathbf a}}}}_h&= {\ensuremath{\mathbf F}^*} ({\ensuremath{\mathbf F}} {\ensuremath{\mathbf a}}_y/(\sqrt{M}{\ensuremath{\mathbf F}}\ensuremath{\tilde{\mathbf a}}_x)) ={\ensuremath{\mathbf F}^*} \left( \Big[\sqrt{M}{\ensuremath{\mathbf F}} \ensuremath{\tilde{\mathbf a}}_x \bullet {\ensuremath{\mathbf F}} \ensuremath{\tilde{\mathbf a}}_h+ {\ensuremath{\mathbf F}} {\ensuremath{\mathbf w}}_c \Big] /(\sqrt{M} {\ensuremath{\mathbf F}} \ensuremath{\tilde{\mathbf a}}_x)\right)\\ & = \ensuremath{\tilde{\mathbf a}}_h +\underbrace{ {\ensuremath{\mathbf F}^*} ({\ensuremath{\mathbf F}} {\ensuremath{\mathbf w}}_c/(\sqrt{M}{\ensuremath{\mathbf F}} {\ensuremath{\mathbf a}}_x))}_{={\ensuremath{\mathbf w}}'} =\ensuremath{\tilde{\mathbf a}}_h + {\ensuremath{\mathbf w}}'.\label{eq:zfah} \end{align} A straightforward bound for the estimation error for Huffman sequences is: \begin{align} \Norm{{\ensuremath{\mathbf w}}'}_2^2 &= \Norm{{\ensuremath{\mathbf F}} {\ensuremath{\mathbf w}}_c/(\sqrt{M}{\ensuremath{\mathbf F}} \ensuremath{\tilde{\mathbf a}}_x)}_2^2 \leq \frac{1}{M}\Norm{|{\ensuremath{\mathbf F}}{\ensuremath{\mathbf w}}_c|^2}_1 \Norm{|\hat{\ensuremath{\tilde{\mathbf a}}}_x|^{-2}}_{\infty}\\ &=\Norm{{\ensuremath{\mathbf w}}_c}_2^2 \cdot \frac{1}{M\min_k |({\ensuremath{\mathbf F}} \ensuremath{\tilde{\mathbf a}}_x)_k|^2 } \overset{\eqref{eq:autohuf}}{=} \Norm{{\ensuremath{\mathbf w}}_c}_2^2 \cdot \frac{1}{\min_k |2\eta \cos(2\pi {\ensuremath{K}} k/M) -1|^2} \leq \frac{\Norm{{\ensuremath{\mathbf w}}_c}_2^2}{(1- 2\eta)^2}\notag. \end{align} If $\eta <1/3$, see optimal radius \eqref{eq:optimalradius}, we get $\Norm{{\ensuremath{\mathbf w}}'}_2^2 \leq 9 \Norm{{\ensuremath{\mathbf w}}_c}_2^2$. The expectation of the colored noise power in \eqref{eq:colorednoise} can be upper bounded by \begin{align} \Expect{\Norm{{\ensuremath{\mathbf w}}_c}_2^2} \leq 2 N\cdot N_0\cdot {\ensuremath{L}} + N\cdot N_0^2. \end{align} By using $\Norm{{\ensuremath{\mathbf x}}}_2^2=\Expect{|h_{\ensuremath{l}}|^2}=1$ and $ \Expect{\Norm{\vn}_2^2}=N\cdot N_0$ this leaves us with an upper MSE of \begin{align} \Norm{{\ensuremath{\mathbf a}}_h-\hat{{\ensuremath{\mathbf a}}}_h}_2^2\leq 18 N\cdot N_0(N_0+{\ensuremath{L}}). \end{align} Hence, for large noise powers this leads to a bad estimate $\hat{{\ensuremath{\mathbf a}}}_h$ and might therefore result in a poor performance of the SDP. \ifall \subsection{Matlab implementation} Note, the {\tt fft} matlab function is ${\ensuremath{\tilde{\mathbf F}}}:= \sqrt{M}{\ensuremath{\mathbf F}}$. \begin{itemize} \item $\tt\ensuremath{\tilde{x}}=$ \tt poly(\ensuremath{\alpha}); \quad Note: $\tt x(z)=x_1 z^{\ensuremath{K}} + x_2 z^{{\ensuremath{K}}-1} + \ldots + x_{\ensuremath{K}} z + x_{{\ensuremath{K}}+1}$ \item $\tt x= \ensuremath{\tilde{x}}/$norm$(\tt \ensuremath{\tilde{x}})$; \item \tt a\_x$=$conv($\tt x,trc(x))$; \item ${\ensuremath{\tilde{\mathbf F}}} ({\ensuremath{\mathbf y}}*\cc{{\ensuremath{\mathbf y}}^-}) = \frac{1}{M} \Big({\ensuremath{\tilde{\mathbf F}}} \begin{pmatrix}{\ensuremath{\mathbf a}}_x \\ {\ensuremath{\mathbf 0}}\end{pmatrix} \ensuremath{\bullet} {\ensuremath{\tilde{\mathbf F}}}\begin{pmatrix} {\ensuremath{\mathbf a}}_h \\ {\ensuremath{\mathbf 0}}\end{pmatrix}\Big)$ \end{itemize} \fi \section{Higher Order Modulations for MOCZ} It is possible to use higher order modulation for MOCZ if we combine multiple autocorrelation codebooks. We will introduce two simple $M-$ary MOCZ schemes by quantizing the radius or the phase. See \figref{fig:QPMOCZ} for an additional phase quantization with $\{1,e^{i\ensuremath{\theta}}\}$ where we set $\ensuremath{\theta}=2\pi/2{\ensuremath{L}}$, which allows $4$ positions for each zero and hence encodes two bits of information. The decoder is for each $l$th zero \begin{align} b_{{\ensuremath{k}},1} &= \begin{cases} 1 &, |{\ensuremath{\mathrm Y}}(\ensuremath{\alpha_{\ensuremath{k}}}^+)| + |{\ensuremath{\mathrm Y}}(\ensuremath{\alpha_{\ensuremath{k}}}^+ e^{i\ensuremath{\theta}}| < |{\ensuremath{\mathrm Y}}(\ensuremath{\alpha_{\ensuremath{k}}}^-)| + |{\ensuremath{\mathrm Y}}(\ensuremath{\alpha_{\ensuremath{k}}}^- e^{i\ensuremath{\theta}})| \}\\ 0 &, \text{else}\end{cases}\\ b_{{\ensuremath{k}},2} & = \begin{cases} 1 &, |{\ensuremath{\mathrm Y}}(\ensuremath{\alpha_{\ensuremath{k}}}^+)| < |{\ensuremath{\mathrm Y}}(\ensuremath{\alpha_{\ensuremath{k}}}^+e^{i\ensuremath{\theta}}| \\ 0 &, \text{else}\end{cases} \end{align} \renewcommand{\bottomfraction}{0.9} \renewcommand{\topfraction}{0.9} \renewcommand{\textfraction}{0.01} \begin{figure}[H] \centering \vspace{-0.5cm} \includegraphics[width=0.4\textwidth]{MaryPMOCZ} \vspace{-0.5cm} \caption{M-PMOCZ scheme with two $M/2$ Phase positions and one radius $R$ for $M/2=q_\phi$ odd.} \label{fig:MaryPMOCZ} \end{figure} \paragraph{M-PMOCZ} Similar, we can implement an $M-ary$ PMOCZ scheme for $M=2P$, by allowing $P=2^m$ phase positions in \begin{align} \phi_l + \{-\ensuremath{\theta}_{P/2}, \dots, -\ensuremath{\theta}_2,-\ensuremath{\theta}_1, \ensuremath{\theta}_1,\ensuremath{\theta}_2,\dots, \ensuremath{\theta}_{P/2}\} \end{align} for $\angle(\ensuremath{\alpha_{\ensuremath{k}}})$. This allows to encode $M$ constellations for each zeros and hence \begin{align} B= {\ensuremath{K}}\log_2 M= {\ensuremath{K}}\log_2(2\cdot 2^{m}) ={\ensuremath{K}}(m+1) \end{align} bits per sequence yielding to a bit rate \begin{align} \etaE=\frac{B}{N} = \frac{{\ensuremath{K}}(m+1)}{{\ensuremath{K}}+{\ensuremath{K}}}. \end{align} \figref{fig:MaryPMOCZ} shows for the ${\ensuremath{k}}$th zero an encoding for $q_{\phi}=M/2$ phases. For $q_{\phi}$ even, we choose the next-neighboor (nn) phase distance as \begin{align} \phi_{nn} = \frac{\pi}{q_{\phi} {\ensuremath{K}}} \end{align} This yields to a uniform placement of the zero positions on the circles with radius $R$ respectively $R^{-1}$. \paragraph{N-RMOCZ} If we modulate on each phase over $N$ radii we obtain an $N-$RMOCZ scheme. Let us choose $N$ radii as \begin{align} 1< R_1<R_2<\dots<R_N. \end{align} Then we encode for $b_{\ensuremath{k}}^{n}\in\{0,1\}$ for $n=1\dots N$ to \begin{align} \ensuremath{\alpha}_{\ensuremath{k}} =e^{i\ensuremath{\phi_{\ensuremath{k}}}} \begin{cases} R_2 &, b_{\ensuremath{k}}^{(1)}=1, b_{\ensuremath{k}}^{(2)}=1\\ R_1 &, b_{\ensuremath{k}}^{(1)}=1, b_{\ensuremath{k}}^{(2)}=0\\ R_1^{-1} &, b_{\ensuremath{k}}^{(1)}=0, b_{\ensuremath{k}}^{(2)}=0\\ R_2^{-1} &, b_{\ensuremath{k}}^{(1)}=0, b_{\ensuremath{k}}^{(2)}=1\\ \end{cases} \end{align} where the zero-codebook is given by \begin{align} \ensuremath{\mathscr Z}_{{\ensuremath{L}},N}=\set{\ensuremath{\boldsymbol{ \alp}}\in{\ensuremath{\mathbb C}}^{\ensuremath{L}}}{\ensuremath{\alpha}_{\ensuremath{k}} \in \{\ensuremath{\alpha}_{{\ensuremath{k}},1},\dots,\ensuremath{\alpha}_{{\ensuremath{k}},N},\ensuremath{\alpha}_{{\ensuremath{k}},1}^-,\dots,\ensuremath{\alpha}_{{\ensuremath{k}},N}^-\}}. \end{align} For decoding we use a bisection decision for the $N$th bit by decoding \begin{align} b_{\ensuremath{k}}^{(1)} &= \begin{cases} 0 &, |{\ensuremath{\mathrm Y}}(\ensuremath{\alpha}_{{\ensuremath{k}},1}^+) +{\ensuremath{\mathrm Y}}(\ensuremath{\alpha}_{{\ensuremath{k}},2}^+| < |{\ensuremath{\mathrm Y}}(\ensuremath{\alpha}_{{\ensuremath{k}},1}^-) + {\ensuremath{\mathrm Y}}(\ensuremath{\alpha}_{{\ensuremath{k}},1}^-|\\ 1 &, \text{else} \end{cases}\\ b_{\ensuremath{k}}^{(2)} &= \begin{cases} 0 &, |{\ensuremath{\mathrm Y}}(e^{i\ensuremath{\phi_{\ensuremath{k}}}}R_1^{2b_l^{(1)}-1})| < |{\ensuremath{\mathrm Y}}(e^{i\ensuremath{\phi_{\ensuremath{k}}}}R_2^{2b_{\ensuremath{k}}^{(1)}-1})| \\ 1 &, \text{else} \end{cases} \end{align} \begin{figure} \begin{subfigure}{0.5\textwidth} \hspace{-1.25cm} \includegraphics[width=1.25\linewidth]{2and4and8PMOCZ_runpp8k_L8K1to16_rSNR} \end{subfigure} \hspace{-0.35cm} \begin{subfigure}{0.5\textwidth} \includegraphics[width=1.25\linewidth]{2and4and8PMOCZ_runpp8k_L8K1to16_EbN0} \end{subfigure} \caption{2,4,8-PMOCZ with $1$ $2$ resp. $4$ Phase positions with $p_d=1$ for ${\ensuremath{K}}=16$ and ${\ensuremath{L}}=1,9,17,24$ over {\ensuremath{\text{rSNR}}}.} \label{fig:QPMOCZ} \end{figure} \section{Product distances from a circle point to the vertices of a regular polygon}\label{app:ngon} We will adapt a result in \cite[Sec.5]{CK07b} to derive for any regular $N-$gon the extremal products of distances from a point on a circle centered at the centroid to all its vertices, see \figref{fig:smallestproductdistance}. \begin{thm}\label{thm:mgonproddistance} Let $N\geq 2$. Consider the regular $N-$gon inscribed in a circle of radius $r>0$ centered at the origin. The product of the distances to any fixed point $z=\ensuremath{\delta} e^{\im \ensuremath{\theta}}$ on a circle of radius $0<\ensuremath{\delta}$ centered at the origin to the vertices is bounded by % \begin{align} |r^N+\ensuremath{\delta}^N|\geq g^c_N(z):=\prod_{n=1}^N d_n(z) \geq |r^N-\ensuremath{\delta}^N|. \end{align} % The bounds are sharp and the extremal points are $\ensuremath{z_{\text{min}}}=\ensuremath{\delta} e^{\im n\pi/N}$ resp. $\ensuremath{z_{\text{max}}}=\ensuremath{\delta} e^{\im n2\pi/N}$, which lie on a line between one vertex and the origin respectively on a line between the middle point of two neighbor vertices and the origin. If the origin is added as an $N+1$ point the minimal and maximal product distance is achieved for the same $\ensuremath{z_{\text{min}}}$ and $\ensuremath{z_{\text{max}}}$ and given by \begin{align} |r^N+\ensuremath{\delta}^N|\ensuremath{\delta}\geq \ensuremath{\delta} g^c_N(z):= \prod_{n=1}^{N+1} d_n(z) \geq\ensuremath{\delta} |r^N-\ensuremath{\delta}^N|. \end{align} \end{thm} \newcommand{\ensuremath{z}}{\ensuremath{z}} \newcommand{\ensuremath{r(\phi)}}{\ensuremath{r(\phi)}} \newcommand{\ensuremath{h}}{\ensuremath{h}} \begin{figure}[t] \begin{subfigure}{0.485\textwidth} \hspace{-2cm} \def0.65\textwidth} \footnotesize{ \import{\pdfdir}{OptimaRadiusHuffmanBMOCZ.pdf_tex} {1.1\textwidth} \scriptsize{ \import{\pdfdir}{HexagonalCirclePack_proddist2.pdf_tex}} \vspace{-0.128cm} \caption{Points inside the polygon on a circle of radius $\ensuremath{\delta}$ (solid circle) resp. outside (dashed circle) with minimal (filled red) resp. maximal (non-filled ) distance products.}\label{fig:smallestproductdistance} \end{subfigure} \hspace{0.2cm} \begin{subfigure}{0.485\textwidth} \vspace{-0.18cm} \centering \def0.65\textwidth} \footnotesize{ \import{\pdfdir}{OptimaRadiusHuffmanBMOCZ.pdf_tex} {0.85\textwidth} \scriptsize{ \import{\pdfdir}{HexagonalCirclePack_worstHuffman.pdf_tex}} \vspace{-0.18cm} \caption{Worst case constellation for Huffman polynomials.} \label{fig:huffmanworst} \end{subfigure} \caption{Upper bounds for noise power which guarantee zero separation for $N=6$.} % \end{figure} \begin{proof} Let us identify the vertices by complex numbers $r\ensuremath{\omega}^m$, where $\ensuremath{\omega}=e^{\im 2\pi /N}$ is the $N$th root of unity. Then the product of the distances from the vertices to any point $z$ is given by % \begin{align} g^c_N(z)=\prod_{m=1}^N d_m(z) &= |z-r|\cdot |z-r\ensuremath{\omega}^1| \dots |z-r\ensuremath{\omega}^{N-1}| = |z^N -r^N| \end{align} % Taking the product and inserting $z=\ensuremath{\delta} e^{i\ensuremath{\theta}}$ we get % \begin{align} g(\ensuremath{\theta}):=|\ensuremath{\delta}^N e^{N\im \ensuremath{\theta}}-r^N|^2 = \ensuremath{\delta}^{2N} -2{\ensuremath{\delta}^N}r^N\cos(N\ensuremath{\theta}) +r^{2N}\label{eq:g} \end{align} % We immediately find that $g$ is periodic in $[0,2\pi/N)$ and symmetric around $\ensuremath{\theta}_c=\pi/N$. Then, the critical points in the interval are given by the solutions of % \begin{align} 0=g'(\ensuremath{\theta})=2N \ensuremath{\delta}^N r^N \sin(N\ensuremath{\theta}) \quad \ensuremath{\Rightarrow} \quad \ensuremath{\theta}_1=0,\ \ensuremath{\theta}_2=\pi/N. \end{align} % Due to the symmetry of $g$ around $\ensuremath{\theta}_1$ and $\ensuremath{\theta}_2$, one of them must be a maximum and the other a minimum. Inserting both in \eqref{eq:g} yields % \begin{align} g(0)=\ensuremath{\delta}^{2N} -2\ensuremath{\delta}^Nr^N +r^{2N}=(r^N-\ensuremath{\delta}^N)^2\quad,\quad g(\frac{\pi}{N})=(r^N+\ensuremath{\delta}^N)^2 \end{align} % such that $g(0)$ is the minimum and $g(\pi/N)$ the maximum. Due to symmetry of the hexagon we can assume that the extremal point is in the gray area of \figref{fig:smallestproductdistance}. Hence, the minimal point lies on the real axis between the right vertex and the origin and the maximal point lies on the line crossing the midpoint of two vertices and the origin. If we add the origin $0$ as the $N+1$ point, then $d_{N+1}=|0-z|=\ensuremath{\delta}$ for each $\ensuremath{\theta}$ and hence we only need to scale the upper and lower bounds by $\ensuremath{\delta}$. \end{proof} We will now ask for the case, where $z$ is placed on a circle with radius $\ensuremath{\delta}$ around one vertex, see the dashed circle in \figref{fig:smallestproductdistance}. \iflong \begin{figure}[t] \centering \vspace{-0.2cm} \def0.65\textwidth} \footnotesize{ \import{\pdfdir}{OptimaRadiusHuffmanBMOCZ.pdf_tex} {0.5\textwidth} \input{pdf/NgonBeweis_neu.pdf_tex} \vspace{-0.328cm} \caption{Hexagon distances from a circle point centered at one vertex.} \label{fig:Ngonbeweis} \end{figure} \fi \begin{conjecture}\label{con:ngon} Let $N\geq 2$ be an integer and $\ensuremath{\omega}=e^{\im 2\pi/N}$ be the $N$th root of unity. Then the points $r\ensuremath{\omega}^n$ are the vertices of a regular $N-$gon inscribed in the circle of radius $r>0$ centered at the origin. Let $0<\ensuremath{\delta}\leq r\sin(\pi/N)$ and consider any $z$ on a circle $C_k(\ensuremath{\delta})$ of radius $\ensuremath{\delta}$ with center at one vertex $r\ensuremath{\omega}^k$, then the minimal product of all distances from $z$ to the vertices is given by % \begin{align} \min_{z\in C_k(\ensuremath{\delta})} \prod_n |z-r\ensuremath{\omega}^n| = r^N- (r-\ensuremath{\delta})^N.\label{eq:zeroonvertex} \end{align} \end{conjecture} \begin{remark} If we add the origin of the $N-$gon as the $N+1$ point and demand $\ensuremath{\delta}<r\sin(\pi/N)$, then we get for the minimum of its product distances, by taking $k=0$ \begin{align} \prod_{n=1}^{N+1} d_n \geq \min d_{N+1} \cdot \min \prod_{n=1}^6 d_n \geq (r-\ensuremath{\delta}) (r^N-(1-\ensuremath{\delta})^N)\label{eq:polygonmin} \end{align} since at any point $z=r+\ensuremath{\delta} e^{i\ensuremath{\theta}}$ the distance to the centroid is \begin{align} d_{N+1}=|z-0|=|r+\ensuremath{\delta} e^{i\ensuremath{\theta}}|\geq |r-\ensuremath{\delta}|=r-\ensuremath{\delta} \end{align} where equality only holds for $\ensuremath{\theta}=\pi$, see \figref{fig:smallestproductdistance}. Note, $1-\sin(\pi/N)=\ensuremath{\operatorname{versine}}(\pi/2-\pi/N)=\ensuremath{\operatorname{versine}}((N-2)\pi/(2N))=2\sin^2((N-2)\pi/(4N))$ is monotone decreasing with increasing $N$. Hence $1-(1-\sin(\pi/N))^N$ will monotone increase with $N$. Moreover, for $N\geq 10$ we get \begin{align} 1-(1-\sin(\pi/N))^N\simeq 1 \end{align} \end{remark} For $N=2$ the conjecture holds, since for $z=1+\ensuremath{\delta} e^{i\ensuremath{\theta}}$ we have \begin{align} |1+\ensuremath{\delta} e^{i\ensuremath{\theta}} -1|\cdot|1+\ensuremath{\delta} e^{i \ensuremath{\theta}} +1| = \ensuremath{\delta}\cdot|2+\ensuremath{\delta}^{i\ensuremath{\theta}}|\geq \ensuremath{\delta}(2-\ensuremath{\delta})=1-(1-2\ensuremath{\delta} +\ensuremath{\delta}^2)=1-(1-\ensuremath{\delta})^2,\notag \end{align} where equality holds if and only if $\ensuremath{\theta}=\pi$. \iflong \begin{approach} Due to symmetry, we can pick an arbitrary vertex. Taking $k=0$, we have to show % \begin{align} \min_{z\in C_0(\ensuremath{\delta})} \prod |z-r\ensuremath{\omega}^n|= \min_{\ensuremath{\theta}} \prod|r+\ensuremath{\delta} e^{i\ensuremath{\theta}} -r \ensuremath{\omega}^n|\geq r^N-(r-\ensuremath{\delta})^N \label{eq:start} \end{align} % First, we can assume $r=1$ and $\ensuremath{\delta}<\sin(\pi/N)$ since \eqref{eq:start} is equivalent to % \begin{align} \min_{\ensuremath{\theta}}\prod |1+\frac{\ensuremath{\delta}}{r} e^{i\ensuremath{\theta}} -\ensuremath{\omega}^n| \geq 1- (1-\frac{\ensuremath{\delta}}{r})^N \end{align} % where we identify $\ensuremath{\tilde{\del}}=\ensuremath{\delta}/r <\sin(\pi/N)$. By identifying the product as the modulus of the polynomial $p(z)=z^N-r^N$ we need to show % \begin{align} \min_{\ensuremath{\theta}} |(1+\ensuremath{\delta} e^{i\ensuremath{\theta}})^N -1| \geq 1- (1-\ensuremath{\delta})^N \end{align} % It is obvious, that for $\ensuremath{\theta}=\pi$ we obtain equality. Moreover, it holds $|p(z)|=|p(\cc{z})|$, which exploits the symmetry around $\ensuremath{\theta}=\pi$. Hence, to ensure that the RHS is the only minimum, we need to show strict inequality, i.e., % \begin{align} |1-(1+\ensuremath{\delta} e^{i\ensuremath{\theta}})^N|&>1-(1-\ensuremath{\delta})^N \quad,\quad \ensuremath{\theta}\in [0,\pi),\ensuremath{\delta}\in(0,\sin(\pi/N)]\label{eq:originalngon}\\ \ensuremath{\Leftrightarrow} \quad |1+\ensuremath{\delta} e^{i\ensuremath{\theta}}|^{2N} -2\Re(1+\ensuremath{\delta} e^{i\ensuremath{\theta}})^N & >(1-\ensuremath{\delta})^N\cdot((1-\ensuremath{\delta})^N-2) \quad,\quad \ensuremath{\theta}\in [0,\pi)\label{eq:ztht} \end{align} % The problem is the real part on the LHS, which involves the $N$th power of a complex number. \iflong Taking the reverse triangle inequality we get for \eqref{eq:originalngon} % \begin{align} |1-(1+\ensuremath{\delta} e^{\im \ensuremath{\theta}})^N|\geq |1-|1+\ensuremath{\delta} e^{im \ensuremath{\theta}}|^N| = |1- (1+\ensuremath{\delta}^2 +2\ensuremath{\delta} \cos(\ensuremath{\theta}))^{N/2}| \end{align} % However, we then only have % \begin{align} (1-\ensuremath{\delta})^2\leq (1+\ensuremath{\delta}^2 +2\ensuremath{\delta} \cos(\ensuremath{\theta}))\leq (1+\ensuremath{\delta})^2 \end{align} % Which only shows equality for $\ensuremath{\theta}=0$ and $\ensuremath{\theta}=\pi$, but can not be used for the proof. \fi \if0 \paragraph{Approach 1} Since it is unclear, how the point $z$ on the circle will be changed by its $N$th power, we will use a geometric approach and re-parameterize $z$ in polar coordinates between phases % \begin{align} z= z(\phi)= r(\phi) e^{i\phi}\quad,\quad \phi \in[0,\ensuremath{\phi_{\text{max}}}] \end{align} % where $\ensuremath{\phi_{\text{max}}}$ is attained if $r(\ensuremath{\theta})$ is tangential to the cirlce, i.e, by Pythagoras we get % \begin{align} \ensuremath{\phi_{\text{max}}}= \arcsin(\ensuremath{\delta})\leq \frac{\pi}{N} \end{align} % see also \figref{fig:Ngonbeweis}. Note, for each $\phi$ we have two possible radii, inside $r_0(\phi)$ and outside $r_1(\phi)$ of the polygon. This reduces \eqref{eq:ztht} for $i\in\{0,1\}$ to % \begin{align} r_i^{2N}(\phi) - 2 \Re (r_i^{N}(\phi)e^{iN\phi}) = r_i^{N}(\phi)[r_i^{N}(\phi) -2 \cos(N\phi)]\geq (1-\ensuremath{\delta})^N\cdot((1-\ensuremath{\delta})^N-2).\label{eq:ri} \end{align} % By the law of cosine we have for $\phi\in[-\ensuremath{\phi_{\text{max}}},\ensuremath{\phi_{\text{max}}}]$ the radii % \begin{align} r_0(\phi)&=\cos(\phi)-\sqrt{\ensuremath{\delta}^2-\sin^2(\phi)}\in [1-\ensuremath{\delta},\cos(\pi/N)]\\ r_1(\phi)&=\cos(\phi)+\sqrt{\ensuremath{\delta}^2-\sin^2(\phi)}\in [\cos(\pi/N),1+\ensuremath{\delta}]. \end{align} % We will show first \eqref{eq:ri} for $\phi\in(0,\pi/(2N))$. In this case $1\geq \cos(N\phi)\geq \cos(\pi/2)\geq 0$. We take the derivative of the LHS of \eqref{eq:ri} % \begin{align} &2N r_i^{2N-1}(\phi) \cdot r'(\phi) +2N\sin(N\phi)\cdot r_i^{N}(\phi)-2\cos(N\phi) N r_i^{N-1}(\phi)\cdot r'_i(\phi) \overset{!}{=}0\\ \ensuremath{\Leftrightarrow} \quad & r_i^{N}(\phi) r_i'(\phi) + \sin(N\phi)r_i(\phi) -\cos(N\phi) r'_i(\phi)=0\\ \ensuremath{\Leftrightarrow} \quad & r'_i(\phi)[r_i^N (\phi)-\cos(N\phi)]+ \sin(N\phi)r_i(\phi)=0 \end{align} % We already now that for $\phi=0$ we have a critical point. But since $\sin(\phi)r_0(\phi)>0$ for $\phi\in(0,\pi/(2N))$ and \begin{align} r'_0(\phi)=-\sin(\phi) + \frac{1}{2\sqrt{\ensuremath{\delta}^2-\sin^2(\phi)}}2\sin\cos(\phi) =\sin(\phi)[\frac{1}{2\sqrt{\ensuremath{\delta}^2-\sin^2(\phi)}} -1]\geq 0 \end{align} since $0<\sin(\phi)\leq \sin(\pi/N)\leq 1/2$ for $N\geq 6$. Furthermore, we get with the binomial identity % \begin{align} r_0^N(\phi)-\cos(N\phi)&\geq (1-\sin(\pi/N))^N-\cos(N\phi) = 1+\sum_{k=1}^N \binom{N}{k} (-1)^k \sin^k(\pi/N) -\cos(N\phi)\notag \intertext{where we can express the cosine by Franois Viete as} & = 1+\sum_{k=1}^N \binom{N}{k} (-1)^k \sin^k(\pi/N) -\sum_{k=0}^{N/2} (-1)^k \binom{N}{2k} \cos^{N-2k}(\phi) \sin^{2k}(\phi)\\ &\geq 1-\cos(\phi)+ \sum_{k=1}^{N/2} \binom{N}{2k} (\sin^{2k}(\pi/N) - (-1)^k\sin^{2k}(\phi)) \end{align} \fi \end{approach} \begin{approach} We have to show from \eqref{eq:ztht} (multiplying by $-1$ since the LHS is always negative. % \begin{align} 2\Re(1+\ensuremath{\delta} e^{i\ensuremath{\theta}})^N- |1+\ensuremath{\delta} e^{i\ensuremath{\theta}}|^{2N} <2(1-\ensuremath{\delta})^N- (1-\ensuremath{\delta})^{2N} \quad,\quad \ensuremath{\theta}\in [0,\pi) \end{align} \end{approach} \subsection{The Hexagon Case} % Let us consider the special case $N=6$, which defines a hexagon. % \begin{conjecture} Let $\ensuremath{\omega}=e^{i2\pi/6}$ be the $6$th root of unity. Then the points $r\ensuremath{\omega}^n$ are the vertices of a regular hexagon inscribed in the circle of radius $r>0$. Let $0<\ensuremath{\delta}\leq r\sin(\pi/6)=r/2$ and consider any $z$ on a circle $C_k(\ensuremath{\delta})$ of radius $\ensuremath{\delta}$ with center at one vertex $r\ensuremath{\omega}^k$, then the minimal product of all distances from $z$ to the vertices is given by % \begin{align} \min_{z\in C_k(\ensuremath{\delta})} \prod_{n=1}^6 |z-r\ensuremath{\omega}^n| = r^6- (r-\ensuremath{\delta})^6.\label{eq:zeroonvertexhexa} \end{align} \end{conjecture} \begin{approach} Due to symmetry, we can pick an arbitrary vertex. Taking $k=0$, we have to show % \begin{align} \min_{z\in C_0(\ensuremath{\delta})} \prod |z-r\ensuremath{\omega}^n|= \min_{\ensuremath{\theta}} \prod|r+\ensuremath{\delta} e^{i\ensuremath{\theta}} -r \ensuremath{\omega}^n|\geq r^6-(r-\ensuremath{\delta})^6 \label{eq:start_hexa} \end{align} % First, we can assume $r=1$ and $\ensuremath{\delta}<1$ since \eqref{eq:start_hexa} is equivalent to % \begin{align} \min_{\ensuremath{\theta}}\prod |1+\frac{\ensuremath{\delta}}{r} e^{i\ensuremath{\theta}} -\ensuremath{\omega}^n| \geq 1- (1-\frac{\ensuremath{\delta}}{r})^6 \end{align} % where we identify $\ensuremath{\tilde{\del}}=\ensuremath{\delta}/r <1/2$. By identifying the product as the modulus of the polynomial $p(z)=z^6-r^6$ we only need to show % \begin{align} \min_{\ensuremath{\theta}} |(1+\ensuremath{\delta} e^{i\ensuremath{\theta}})^6 -1| \geq 1- (1-\ensuremath{\delta})^6 \end{align} % It is obvious, that for $\ensuremath{\theta}=\pi$ we obtain equality. Moreover, it holds $|p(z)|=|p(\cc{z})|$, which implies a symmetry around $\ensuremath{\theta}=\pi$. Hence, to ensure that the RHS is the only minimum, we need to show strict inequality, i.e., for $\ensuremath{\theta}\in[0,\pi),\ensuremath{\delta}\in(0,1/2)]$ % \begin{align} |1-(1+\ensuremath{\delta} e^{i\ensuremath{\theta}})^6|&>1-(1-\ensuremath{\delta})^6 \\ \ensuremath{\Leftrightarrow} \quad |1+\ensuremath{\delta} e^{i\ensuremath{\theta}}|^{12} -2\Re(1+\ensuremath{\delta} e^{i\ensuremath{\theta}})^6 & >(1-\ensuremath{\delta})^6\cdot((1-\ensuremath{\delta})^6-2)\\ \ensuremath{\Leftrightarrow} \quad g(\ensuremath{\theta}):= (1+\ensuremath{\delta}^2 +2\ensuremath{\delta}\cos(\ensuremath{\theta}))^{6} -2\Re(1+\ensuremath{\delta} e^{i\ensuremath{\theta}})^6 & <(1-\ensuremath{\delta})^6\cdot(2-(1-\ensuremath{\delta})^6) \label{eq:ztht_hexa} \end{align} % The problem is the real part on the LHS, which involves the $6$th power of a complex number. By using the binomial formula we obtain % \begin{align} g(\ensuremath{\theta})=(1+\ensuremath{\delta}^2 +2\ensuremath{\delta}\cos(\ensuremath{\theta}))^{6} -2 -2\sum_{n=1}^6\ensuremath{\delta}^n \binom{6}{n} \cos(n\ensuremath{\theta}) \end{align} % \begin{figure} \begin{subfigure}{0.485\textwidth} \includegraphics[width=0.9\textwidth]{pdf/conjecture_gtht} \caption{Derivative of $g(\ensuremath{\theta})$ in $[0,\pi]$.}\label{fig:gprime} \end{subfigure} \begin{subfigure}{0.485\textwidth} \includegraphics[width=0.9\textwidth]{pdf/ttht} \caption{The function $t_1(\ensuremath{\theta})$ in $[0,\pi]$.}\label{fig:ttht} \end{subfigure} \caption{The trigonometric functions for the bound.} \end{figure} Let us search for the critical points % \begin{align} g'(\ensuremath{\theta})&=2\sum_n \ensuremath{\delta}^n \binom{6}{n} n \sin(n\ensuremath{\theta}) -2\ensuremath{\delta}(1+\ensuremath{\delta}^2)^5\cdot 6\cdot (1+\frac{2\ensuremath{\delta}}{1+\ensuremath{\delta}^2} \cos(\ensuremath{\theta}))^5\sin(\ensuremath{\theta})\\ \ensuremath{\Rightarrow} 0&\overset{!}{=}\sum_n \ensuremath{\delta}^n \binom{6}{n} n \sin(n\ensuremath{\theta}) -\underbrace{6\ensuremath{\delta}(1+\ensuremath{\delta}^2)^5 (\frac{1+\ensuremath{\delta}^2 -2\ensuremath{\delta}\cos(\pi-\ensuremath{\theta})}{1+\ensuremath{\delta}^2})^5\sin(\pi-\ensuremath{\theta})}_{=t_2(\pi-\ensuremath{\theta})} \end{align} % where the second summand $t_2(\ensuremath{\theta})$ is always larger than $0$ for every $\ensuremath{\theta}\in(0,\pi)$ and $0<\ensuremath{\delta}<1$. Hence we only need to show that the first term is always smaller zero to ensure that the only critical points are $\ensuremath{\theta}=0$ and $\pi$, which is numerically true for $\ensuremath{\theta}\in(\pi/2,\pi)$, see \figref{fig:gprime}. % Lets consider first $\ensuremath{\delta}=1/2$, then we have to show by using the trigonometric addition theorems with $\ensuremath{\theta}=\pi-\ensuremath{\epsilon}$ % \begin{align} 0> & \sum_n \ensuremath{\delta}^n \binom{6}{n}n\sin(n\ensuremath{\theta})= -\sum_n 2^{-n} (-1)^n\binom{6}{n} n \sin(n\ensuremath{\epsilon})=:-t_1(\ensuremath{\epsilon}). \end{align} % For $\ensuremath{\theta}<\pi/N=\pi/6$ this is trivial, since $\sin(n\ensuremath{\theta})\geq 0$. The result can be extended for $\ensuremath{\theta}\in[\pi/6,\pi/3]$ by using the symmetry of the binomial coefficient. However, for large $\ensuremath{\theta}$ or equivalent for small $\ensuremath{\epsilon}$, we need to show % \begin{align} t_1(\ensuremath{\epsilon})=-32\sin(\ensuremath{\epsilon})+80\sin(2\ensuremath{\epsilon}) -80\sin(3\ensuremath{\epsilon})+40\sin(4\ensuremath{\epsilon})-10\sin(5\ensuremath{\epsilon}) +\sin(6\ensuremath{\epsilon})>0 \end{align} % If $\ensuremath{\epsilon}$ is very small this gives approximately $4\ensuremath{\epsilon}$ if we use $\sin(\ensuremath{\epsilon})\simeq \ensuremath{\epsilon}$. This shows that the bound must be very tight for small $\ensuremath{\epsilon}$. The tightest sine bound is given for small $\ensuremath{\epsilon}$ by $\sin(\ensuremath{\epsilon})\leq \ensuremath{\epsilon}$. To reduce the multiple angle forms we will use the Chebyshev method for the sine function % \begin{align} \sin(n\ensuremath{\epsilon})=2\cos(\ensuremath{\epsilon})\sin((n-1)\ensuremath{\epsilon})-\sin(( n-2)\ensuremath{\epsilon}) \end{align} % Lets start with the three highest orders % \begin{align} \sin(4\ensuremath{\epsilon})&=2\cos(\ensuremath{\epsilon})\sin(3\ensuremath{\epsilon})-\sin(2\ensuremath{\epsilon})\\ \sin(5\ensuremath{\epsilon}) &= 2\cos(\ensuremath{\epsilon})\sin(4\ensuremath{\epsilon}) -\sin(3\ensuremath{\epsilon})=4\cos^2(\ensuremath{\epsilon})\sin(3\ensuremath{\epsilon})-2\cos(\ensuremath{\epsilon})\sin(2\ensuremath{\epsilon})-\sin(3\ensuremath{\epsilon})\\ &=3\sin(3\ensuremath{\epsilon})-4\sin^2(\ensuremath{\epsilon})\sin(3\ensuremath{\epsilon})-2\cos(\ensuremath{\epsilon})\sin(2\ensuremath{\epsilon})\\ \sin(6\ensuremath{\epsilon})& = 2\cos(\ensuremath{\epsilon})\sin(5\ensuremath{\epsilon}) -\sin(4\ensuremath{\epsilon})=4\cos^2(\ensuremath{\epsilon})\sin(4\ensuremath{\epsilon}) -2\cos(\ensuremath{\epsilon})\sin(3\ensuremath{\epsilon}) -\sin(4\ensuremath{\epsilon})\\ &=4\cos(\ensuremath{\epsilon})\sin(3\ensuremath{\epsilon})-3\sin(2\ensuremath{\epsilon}) -8\sin^2(\ensuremath{\epsilon})\cos(\ensuremath{\epsilon})\sin(3\ensuremath{\epsilon})+4\sin^2(\ensuremath{\epsilon})\sin(2\ensuremath{\epsilon}) \end{align} % where we also used $\cos^2=1-\sin^2$ to eliminate the cosine. This gives us for $\ensuremath{\epsilon}\in(0,\pi)$ % \begin{align} t_1(\ensuremath{\epsilon})=&-32\sin(\ensuremath{\epsilon})+37\sin(2\ensuremath{\epsilon}) -110\sin(3\ensuremath{\epsilon}) \\ &+84\cos(\ensuremath{\epsilon})\sin(3\ensuremath{\epsilon})\\ &+40\sin^2(\ensuremath{\epsilon})\sin(3\ensuremath{\epsilon})+20\cos(\ensuremath{\epsilon})\sin(2\ensuremath{\epsilon})\\ &-8\sin^2(\ensuremath{\epsilon})\cos(\ensuremath{\epsilon})\sin(3\ensuremath{\epsilon})+4\sin^2(\ensuremath{\epsilon})\sin(2\ensuremath{\epsilon}) \end{align} % Each of the trigonometric function products is positive in $(0,\pi/3)$. We use the sharp sine bound: % \begin{align} t_1(\ensuremath{\epsilon})\geq& -32\sin(\ensuremath{\epsilon}) +37\sin(2\ensuremath{\epsilon})-110\sin(3\ensuremath{\epsilon}) \notag\\ &+168\sin(2\ensuremath{\epsilon}) -168\ensuremath{\epsilon}^2\sin(2\ensuremath{\epsilon}) -84\cos(\ensuremath{\epsilon})\ensuremath{\epsilon}\notag\\ &+40\sin^2(\ensuremath{\epsilon})\sin(3\ensuremath{\epsilon}) +40\sin(\ensuremath{\epsilon})-40\ensuremath{\epsilon}^2\sin(\ensuremath{\epsilon})\notag\\ &-8\ensuremath{\epsilon}^2\cos(\ensuremath{\epsilon})\sin(3\ensuremath{\epsilon}) + 4\sin^2(\ensuremath{\epsilon})\sin(2\ensuremath{\epsilon})\\ \geq & -32\sin(\ensuremath{\epsilon}) +205\sin(2\ensuremath{\epsilon})-220\cos(\ensuremath{\epsilon})\sin(2\ensuremath{\epsilon}) +110\sin(\ensuremath{\epsilon}) -168\ensuremath{\epsilon}^2\sin(2\ensuremath{\epsilon})-84\ensuremath{\epsilon}\cos(\ensuremath{\epsilon})\notag\\ &+80\sin^2(\ensuremath{\epsilon})\cos(\ensuremath{\epsilon})\sin(2\ensuremath{\epsilon}) -80\ensuremath{\epsilon}^3 + 40\sin(\ensuremath{\epsilon}) \\ &-16\ensuremath{\epsilon}^2\cos^2(\ensuremath{\epsilon})\sin(2\ensuremath{\epsilon}) +8\ensuremath{\epsilon}^2\cos(\ensuremath{\epsilon})\sin(\ensuremath{\epsilon}) + 4\sin^2(\ensuremath{\epsilon})\sin(2\ensuremath{\epsilon})\\ \geq &118\sin(\ensuremath{\epsilon}) +205\sin(2\ensuremath{\epsilon}) -524\ensuremath{\epsilon}\cos(\ensuremath{\epsilon}) -334\ensuremath{\epsilon}^3 +80\sin^2(\ensuremath{\epsilon})\cos(\ensuremath{\epsilon})\sin(2\ensuremath{\epsilon}) \\ & -32\ensuremath{\epsilon}^2\cos(\ensuremath{\epsilon})\sin(\ensuremath{\epsilon})+8\ensuremath{\epsilon}^2\cos(\ensuremath{\epsilon})\sin(\ensuremath{\epsilon}) + 4\sin^2(\ensuremath{\epsilon})\sin(2\ensuremath{\epsilon})\\ \geq &118\sin(\ensuremath{\epsilon}) +205\sin(2\ensuremath{\epsilon}) -524\ensuremath{\epsilon}\cos(\ensuremath{\epsilon}) -358\ensuremath{\epsilon}^3 +80\sin^2(\ensuremath{\epsilon})\cos(\ensuremath{\epsilon})\sin(2\ensuremath{\epsilon}) \\ & + 4\sin^2(\ensuremath{\epsilon})\sin(2\ensuremath{\epsilon}) \end{align} We will use $\sin(\ensuremath{\epsilon})\geq \ensuremath{\epsilon}(1-\ensuremath{\epsilon}/\pi)\geq \ensuremath{\epsilon} (1-1/n)$ for $\ensuremath{\epsilon}\leq \pi/n$ with some $n\geq 2$, which gives \begin{align} \sin(2\ensuremath{\epsilon})\geq 2\ensuremath{\epsilon}(1-(2\ensuremath{\epsilon}/\pi))\geq 2\ensuremath{\epsilon}(1 -2/n) \end{align} % Hence % \begin{align} t_1(\ensuremath{\epsilon})&\geq \ensuremath{\epsilon} \left( 118(1-1/n) +410(1-2/n) -524\cos(\ensuremath{\epsilon}) -358\ensuremath{\epsilon}^2 +20(\sin^2(\ensuremath{\epsilon})\cos(\ensuremath{\epsilon})+\sin^2(\ensuremath{\epsilon}))(1-2/n)\right)\notag\\ &\geq \ensuremath{\epsilon} \left(528 -938/n -524\cos(\ensuremath{\epsilon}) -358\ensuremath{\epsilon}^2 +20(\sin^2(\cos+1)) -80\ensuremath{\epsilon}^2/n\right) \end{align} The largest negative part can be bounded by \begin{align} -524\cos(\ensuremath{\epsilon})\leq -524\sin(\ensuremath{\epsilon}+\frac{\pi}{2}) \leq -524 \frac{4(\ensuremath{\epsilon}+\frac{\pi}{2})}{\pi}(1-\frac{\ensuremath{\epsilon}+\frac{\pi}{2}}{\pi}) \leq -524\cdot (1-\frac{4\ensuremath{\epsilon}^2}{\pi^2}) \end{align} Yielding to \begin{align} t_1(\ensuremath{\epsilon})&\geq \ensuremath{\epsilon} \left( 4 + \frac{4\ensuremath{\epsilon}^2}{\pi^2}-\frac{938}{n} + 44\sin^2(\ensuremath{\epsilon})(2-4/n) -448\ensuremath{\epsilon}^2 +8\ensuremath{\epsilon}\cos(\ensuremath{\epsilon})\sin(\ensuremath{\epsilon})\right)\\ \intertext{for $n=250$ we get} &\geq \ensuremath{\epsilon} \left( 0.247 + 44 (\ensuremath{\epsilon}^2 -6\ensuremath{\epsilon}^3/\pi +\ensuremath{\epsilon}^3/(\pi^2))\cdot(1.984) -0.071 \right)\\ &\geq \ensuremath{\epsilon} ( 0.247 -44\cdot 12 \cdot \frac{\pi^2}{250^3} -0.071)\geq \ensuremath{\epsilon}(0.247 -0.00034 -0.071)>0 \end{align} Hence, it holds for $\ensuremath{\epsilon}\in(0,\pi/250)$. % \end{approach} \begin{approach} A direct computation by using the reverse triangle inequality yields to % \begin{align} \prod_{n=1}^6 d_n &= \prod_n |1+\ensuremath{\delta} e^{i\ensuremath{\theta}}-e^{i2\pi n/6}| \geq |\ensuremath{\delta}|\cdot \Big|\frac{\sqrt{12} }{2}-|\ensuremath{\delta}|\Big|^2 \cdot|1-|\ensuremath{\delta}||^2\cdot |2-|\ensuremath{\delta}|| \\ &=6\ensuremath{\delta} -(15-2\sqrt{12})\ensuremath{\delta}^2 +(14+5\sqrt{12})\ensuremath{\delta}^3 - (8+4\sqrt{12})\ensuremath{\delta}^4 + (4+\sqrt{12})\ensuremath{\delta}^5 -\ensuremath{\delta}^6\\ &\leq 1-(1-\ensuremath{\delta})^6 \end{align} and can not establish the conjectured lower bound for $N=6$. \end{approach} \begin{figure}[t] \centering \includegraphics[width=0.6\textwidth]{pdf/sinbounds} \caption{Upper bound $\sin \ensuremath{\epsilon} \leq \ensuremath{\epsilon}$ and lower bound $\sin \ensuremath{\epsilon} \geq \ensuremath{\epsilon}-\ensuremath{\epsilon}^2/\pi$ for small $\ensuremath{\epsilon}$.} \end{figure} \begin{approach} Take the Arithmetic mean of the distances and show % \begin{align} \frac{1}{6}\sum_{n=1}^6 d_n(\ensuremath{\theta}) \geq (\prod_n d_n(\ensuremath{\theta}))^{1/6} \geq (1-(1-\ensuremath{\delta})^6)^{1/6} \end{align} % Hence, we need first to show that % \begin{align} \min_{\ensuremath{\theta}} \left(\frac{1}{6} \sum_n d_n(\ensuremath{\theta})\right)^{6}>(1-(1-\ensuremath{\delta})^6\label{eq:strictam} \end{align} % where the inequality must be strict, since the only configuration which allows equality is for $\ensuremath{\delta}=1$ and $\ensuremath{\theta}=\pi$, where all distances become $1$. If \eqref{eq:strictam} is true, there is the conjecture that for $\hat{\ensuremath{\theta}}$ achieving the minimal arithmetic mean it also might achieve the minimal geometric mean. \end{approach} \fi \subsection{Back to the Hexagon Lattice} If the \conref{con:ngon} would hold, we can similar argument as in \eqref{eq:polygonmin}, that if the origin is included as an $N+1$ point the minimum would be % \begin{align} \prod_{n=1}^{N+1} d_n(\ensuremath{\theta}) \geq (r-\ensuremath{\delta})r^N(1-(1-\ensuremath{\delta})^N) \end{align} % and be achieved for $\ensuremath{\theta}=\pi$. Furthermore, if we chose the circle around the centroid we get with \thmref{thm:mgonproddistance} % \begin{align} |\ensuremath{\delta}|\prod_{n=1}^N |\ensuremath{\delta} e^{\im \ensuremath{\theta}}-\ensuremath{\alpha}_n | \geq \ensuremath{\delta}(r^N-\ensuremath{\delta}^N) \end{align} Indeed, we can then show, that the later product distance is the smallest possible. % \begin{lemi} Let $r>0$ and $N\geq 2$. Then it holds for any $\ensuremath{\delta}\in(0,r/2)$ \begin{align} \ensuremath{\delta}(r^N-\ensuremath{\delta}^N) < (r^N-(r-\ensuremath{\delta})^N)\cdot(r-\ensuremath{\delta}).\label{eq:conjlowerbound} \end{align} \end{lemi} \begin{proof} To show this, we only need to verify for $r=1$ and $\ensuremath{\delta}\in(0,1/2)$, i.e., \begin{align} \ensuremath{\delta}(1-\ensuremath{\delta}^N)&<(1-(1-\ensuremath{\delta})^N)(1-\ensuremath{\delta})\\ \ensuremath{\Leftrightarrow} \quad a_N=\ensuremath{\delta}-\ensuremath{\delta}^{N+1} &< (1-\ensuremath{\delta})-(1-\ensuremath{\delta})^{N+1} =b_N\label{eq:einsdel} \end{align} For $\ensuremath{\delta}=1/2$ and $\ensuremath{\delta}=0$ this becomes equality. We will prove the strict inequality by induction. For $N\geq 2$ we get \begin{align} b_2= 1-\ensuremath{\delta} -(1-3\ensuremath{\delta}+3\ensuremath{\delta}^2-\ensuremath{\delta}^3)=2\ensuremath{\delta} -3\ensuremath{\delta}^2+\ensuremath{\delta}^3=b_2 + \ensuremath{\delta} +2\ensuremath{\delta}^3 -3\ensuremath{\delta}^2\label{eq:indstart} \end{align} But for the last three terms it holds \begin{align} \ensuremath{\delta}+2\ensuremath{\delta}^3-3\ensuremath{\delta}^2>0 \quad\ensuremath{\Leftrightarrow} \quad 1+2\ensuremath{\delta}^2>3\ensuremath{\delta} \quad \ensuremath{\Leftrightarrow} \quad \ensuremath{\delta}^{-1} + 2\ensuremath{\delta}>2+1=3. \end{align} Note, for $N=1$ the strict inequality \eqref{eq:einsdel} does not hold, since $\ensuremath{\delta}(1-\ensuremath{\delta})=(1-1+\ensuremath{\delta})(1-\ensuremath{\delta})$. We will now show that \eqref{eq:einsdel} holds for $N+1$. From \eqref{eq:einsdel} we get \begin{align*} a_{N+1}& =\ensuremath{\delta}-\ensuremath{\delta}^{N+2} =\ensuremath{\delta}(\ensuremath{\delta}-\ensuremath{\delta}^{N+1})-\ensuremath{\delta}^2+\ensuremath{\delta}=\ensuremath{\delta} a_N + (1-\ensuremath{\delta}) -(1-\ensuremath{\delta})^2\\ b_{N+1}& = (1-\ensuremath{\delta})-(1-\ensuremath{\delta})^{N+2}=(1-\ensuremath{\delta})b_N+(1-\ensuremath{\delta})-(1-\ensuremath{\delta})^2>\ensuremath{\delta} b_N +(1-\ensuremath{\delta})-(1-\ensuremath{\delta})^2 \end{align*} where we used $1-\ensuremath{\delta}>\ensuremath{\delta}$ for each $0\leq\ensuremath{\delta}<1/2$. Hence it holds $b_{N+1}>a_{N+1}$ if $b_N>a_N$ holds. By induction and \eqref{eq:indstart} this holds for all $N\geq 2$. \end{proof} The lower bound \eqref{eq:conjlowerbound} is monotone increasing for $\ensuremath{\delta}\in[0,r/2]$ and achieves its maximum at the boundary $\ensuremath{\delta}=r/2$ given by \begin{align} 0\leq \ensuremath{\delta}(r^N-\ensuremath{\delta}^N)\leq \frac{r^{N+1}}{2}\frac{(2^N-1)}{2^{N}}<\frac{r^{N+1}}{2} \end{align} see \figref{fig:glowerbound_centroid} for the hexagon, $N=6$ and $r=1$. \iflong \paragraph{Lower bound for the Hexagon Conjecture} We can leverage the lower bound of the product distances inside a hexagon by taking for $d_2$ and $d_3$ a different point on the circle $C_1(\ensuremath{\delta})$. \begin{lemi} Consider the vertices of a regular hexagon inscribed in a circle of radius $r>0$. Then the product of distances of any point on a circle of radius $\ensuremath{\delta}\leq r/2$ centered at one vertex to all vertices is not larger than % \begin{align} g_{l,6}(r,\ensuremath{\delta})=\ensuremath{\delta} r^6(1-\ensuremath{\tilde{\del}})(\sqrt{3}-\ensuremath{\tilde{\del}})(2-\ensuremath{\tilde{\del}})\sqrt{(1-\ensuremath{\tilde{\del}}+\ensuremath{\tilde{\del}}^2)(1+\ensuremath{\tilde{\del}}^2-\sqrt{3}\ensuremath{\tilde{\del}})}. \end{align} % \end{lemi} \begin{proof} Due to symmetry of the hexagon, we only need to show that any point $z$ on the circle around $r\ensuremath{\omega}^0$ (first vertex) yields a product distance less than \eqref{eq:zeroonvertexhexa}. Moreover, we know that outside the hexagon the points $z$ will yield to larger values, since all distances increase for $\ensuremath{\theta}=2\pi/3$ to $\pi/2$. For $\ensuremath{\theta}$ below we already showed that \eqref{eq:zeroonvertexhexa} holds. Hence we only need to consider points $z=r+\ensuremath{\delta} e^{i\ensuremath{\theta}}$, for $\ensuremath{\theta}=(2\pi/3,\pi)$. For the filled red point in \figref{fig:hexagon_lbound}, we get the distances \begin{align} d_5&=\sqrt{r^2+(r-\ensuremath{\delta})^2 -2r(r-\ensuremath{\delta})\cos(2\pi/3))}=\sqrt{2r^2-2r\ensuremath{\delta}+\ensuremath{\delta}^2+r^2-r\ensuremath{\delta}}=\sqrt{3r^2-3r\ensuremath{\delta}+\ensuremath{\delta}^2}\notag\\ d_6&=\sqrt{r^2+\ensuremath{\delta}^2 -2r\ensuremath{\delta}\cos(\pi/3)}=\sqrt{r^2+\ensuremath{\delta}^2-r\ensuremath{\delta}}, d_1=\ensuremath{\delta}, d_4=2r-\ensuremath{\delta} \intertext{and for the black point and red circle point we get the distances (lower bounds)} d_2&=r-\ensuremath{\delta}, d_3=2\sqrt{r^2-r^2/4}-\ensuremath{\delta}=\sqrt{3}r-\ensuremath{\delta} \end{align} Yielding with $\ensuremath{\tilde{\del}}=\ensuremath{\delta}/r$ to \begin{align} g_6(r,\ensuremath{\delta},\ensuremath{\theta})&= \prod_{n=1}^6 d_n(r,\ensuremath{\delta},\ensuremath{\theta})\\ &\geq \ensuremath{\delta} r^6(1-\ensuremath{\tilde{\del}})(\sqrt{3}-\ensuremath{\tilde{\del}})(2-\ensuremath{\tilde{\del}})\sqrt{(3-3\ensuremath{\tilde{\del}}+\ensuremath{\tilde{\del}}^2)(1+\ensuremath{\tilde{\del}}^2-\ensuremath{\tilde{\del}})} =g_{l,6}(r,\ensuremath{\delta})\label{eq:lowerboundconj} \end{align} \end{proof} \fi \newcommand{\ensuremath{r}}{\ensuremath{r}} \newcommand{\ensuremath{5\pi/6}}{\ensuremath{5\pi/6}} \newcommand{\ensuremath{d_7}}{\ensuremath{d_7}} \newcommand{\ensuremath{b_1}}{\ensuremath{b_1}} \newcommand{\ensuremath{b_2}}{\ensuremath{b_2}} \newcommand{\ensuremath{\sqrt{2}\del}}{\ensuremath{\sqrt{2}\ensuremath{\delta}}} \newcommand{\ensuremath{\frac{b_2}{2}}}{\ensuremath{\frac{b_2}{2}}} \newcommand{\ensuremath{b_3}}{\ensuremath{b_3}} \newcommand{\ensuremath{b_4}}{\ensuremath{b_4}} \newcommand{\ensuremath{a}}{\ensuremath{a}} \newcommand{\ensuremath{\ensuremath{\del_{\text{max}}}=\frac{a}{2}}}{\ensuremath{\ensuremath{\del_{\text{max}}}=\frac{a}{2}}} \newcommand{\ensuremath{c}}{\ensuremath{c}} \newcommand{\ensuremath{\pi/4}}{\ensuremath{\pi/4}} \newcommand{\ensuremath{\alp_{M+1}}}{\ensuremath{\ensuremath{\alpha}_{M+1}}} \newcommand{\ensuremath{\alp_{M+2}}}{\ensuremath{\ensuremath{\alpha}_{M+2}}} \newcommand{\ensuremath{\alp_{2M+1}}}{\ensuremath{\ensuremath{\alpha}_{2M+1}}} \newcommand{\ensuremath{\alp_{3M+1}}}{\ensuremath{\ensuremath{\alpha}_{3M+1}}} \begin{figure}[t] \centering % \iflong \begin{subfigure}{0.485\textwidth} \def0.65\textwidth} \footnotesize{ \import{\pdfdir}{OptimaRadiusHuffmanBMOCZ.pdf_tex} {1.1\textwidth} \small{ \import{\pdfdir}{hexagonBeweis_lbound.pdf_tex} } \vspace{-0.2cm} \caption{Hexagon conjecture lower bound} \label{fig:hexagon_lbound} \end{subfigure}% \fi % \begin{subfigure}{0.485\textwidth} \def0.65\textwidth} \footnotesize{ \import{\pdfdir}{OptimaRadiusHuffmanBMOCZ.pdf_tex} {1.15\textwidth} \small{ \iflong\hspace{1cm}\fi \import{\pdfdir}{Ngon_vertex_lowerbound_new3.pdf_tex} } \vspace{-0.9cm} \iflong \caption{Lower bound for $N-$gon conjeture around vertex.} \label{fig:ngon_lb_vertex}\fi \end{subfigure} \iflong\vspace{-0.3cm}\fi \caption{Lower bound conjecture for regular $N-$gon with circle point around one vertex.} \label{fig:ngon_lb_vertex} \iflong \caption{Lower bounds for regular $N-$gons.}\fi \end{figure} We will now lower bound the product distances in \conref{con:ngon} for $N-$gons with circle points around one vertex, by using the geometric relaxation given in \figref{fig:ngon_lb_vertex}. \begin{lemi}\label{eq:gHuffman} Consider a regular $N-$gon with $N=4M$ for any $3\leq M\in{\ensuremath{\mathbb N}}$ inscribed in a circle of radius $r>0$. Consider a point $z$ on a circle with radius $\ensuremath{\delta}<r\sin(\pi/N)$ and center $z=r+\ensuremath{\delta} e^{i\ensuremath{\theta}}$. Then the minimal product distances for all $\ensuremath{\theta}$ is bounded by % \begin{align} g_{v,N}(\ensuremath{\theta})\geq \prod_n d_n \geq \ensuremath{\delta}(2r-\ensuremath{\delta}) \prod_{m=1}^{M}(1+\sin\frac{\pi}{2M}-2\ensuremath{\tilde{\del}}(\ensuremath{\tilde{\del}}+1 +\sin\frac{\pi}{2M}))\cdot r^4\sin^2\frac{\pi}{4M} (\prod_{m} 2m-1)^2 \label{eq:huffmanMgonbound2} \end{align} \end{lemi} \begin{proof} Lets define $\ensuremath{\omega}=e^{\im 2\pi/N}$. We will use two reference points, at $r-2\ensuremath{\delta}$ and at $r$, to lower bound the distances to the vertices. The special distance % \begin{align} d_{2M+1}=| \ensuremath{\alpha}_{2M+1}-z|\geq 2r-\ensuremath{\delta}\quad,\quad d_1 =\ensuremath{\delta} \end{align} % For all vertices in the left half plane we will use the radius of the smallest circle given by % \begin{align} b_{M+2}=\sqrt{(r+c-\sqrt{2}\ensuremath{\delta})^2+(r-h)^2} \end{align} % where % \begin{align} c=r\sin(\ensuremath{\gamma})=r\sin(\pi/2M) \quad,\quad h=r(1-\cos(\ensuremath{\gamma})),\quad,\quad a=2r\sin(\pi/4M). \label{eq:edge} \end{align} % which gives % \begin{align} b_{M+2} &=\sqrt{r^2(1+\sin(\pi/2M)-2\ensuremath{\tilde{\del}})^2+r^2(1-(1-\cos(\pi/2M)))^2}\\ &=r\sqrt{2}\sqrt{(1+\sin(\pi/2M) +\ensuremath{\tilde{\del}}^2-\sqrt{2}\ensuremath{\tilde{\del}}(1+\sin(\pi/2M))} \end{align} % The distances in the first quadrant we will lower bound by multiples of $\ensuremath{\frac{b_2}{2}}$ given as % \begin{align} \ensuremath{\frac{b_2}{2}}=\ensuremath{a}\cdot\cos(\pi/4)/2=2r\sin(\pi/4M)\frac{\sqrt{2}}{4} =r\sin(\pi/4M) \frac{1}{\sqrt{2}} \end{align} % Then we get for the product of all distances the bound % \if0 \begin{align} \prod_{n=1}^{2M-1} d_n \geq r^{2M-1}. \end{align} % The distance to $\ensuremath{\alpha_1}=\ensuremath{\omega}^0$ will be always $\ensuremath{\delta}$ and the distance to $\ensuremath{\alpha_2}=\ensuremath{\omega}^1$ and $\ensuremath{\alpha}_{N}$ will be $2a-\ensuremath{\delta}$ with % \begin{align} \end{align} % To lower bound the other distance in the right half plane we will use as reference point $\ensuremath{\alpha_2}$ and lower bound the distances from $\ensuremath{\alpha_2}$ to all other points $\ensuremath{\alpha_3}$ to $\ensuremath{\alpha}_{M+1}$. Since the projection of $a$ between $\ensuremath{\alpha_2}$ and $\ensuremath{\alpha_3}$ onto the distance vector $b_{M}$ will always yield the smallest distance. We can then lower bound the other $M$ distances by % \begin{align} b_{m}=m b_3 \quad,\quad b_3=a\cos(\ensuremath{\gamma}) \end{align} % for $m=3$ to $M+1$. Hence this gives \fi % \begin{align} g_{v,4M} &= \prod_n d_n \geq b_{M+2}^{2M-2} \cdot d_1\cdot d_{2M+1}\cdot\prod_{m=1}^{M} ((2m-1)\ensuremath{\frac{b_2}{2}})^2 \\ & =\ensuremath{\delta}(2r-\ensuremath{\delta}) \prod_{m=1}^{M}2r^2 (1+\sin(\pi/2M)-2\ensuremath{\tilde{\del}}(\ensuremath{\tilde{\del}}+1 +\sin(\pi/2M)))\cdot (2m-1)^2\frac{r^2\sin^2(\pi/4M)}{2}\notag\\ & =\ensuremath{\delta}(2r-\ensuremath{\delta}) \prod_{m=1}^{M}(1+\sin\frac{\pi}{2M}-2\ensuremath{\tilde{\del}}(\ensuremath{\tilde{\del}}+1 +\sin\frac{\pi}{2M}))\cdot r^4\sin^2\frac{\pi}{4M} (\prod_{m} 2m-1)^2 \end{align} % \ifextras \begin{approach} The cosine of the angle $\ensuremath{\gamma}$ is given by the cosine law as % \begin{align} \cos(\ensuremath{\gamma})=\frac{r^2|\ensuremath{\omega}^{M+1} -\ensuremath{\omega}^2|^2-r^2|\ensuremath{\omega}^{M+1}-\ensuremath{\omega}^1|^2-a^2}{2ar^2|\ensuremath{\omega}^{M+1}-\ensuremath{\omega}^1|^2} \end{align} % since $\ensuremath{\omega}^{M}=e^{i2\pi M/(4M)}=i$ for every $M\geq 1$ we have % \begin{align} |\ensuremath{\omega}^{M+1}-\ensuremath{\omega}^m|^2 &=|i-\cos(\phi_m)-i\sin(\phi_m)|^2=|i(1-\sin(\phi_m)-\cos(\phi_m)|^2\\ &= (1-\sin(\phi_m))^2+\cos^2(\phi_m) =2(1-\sin(\phi_m)) \end{align} % we get with \eqref{eq:edge} % \begin{align} \cos(\ensuremath{\gamma}) =\frac{2(1-\sin(\phi_2))-2(1-\sin(\phi_1)-4\sin(\pi/N)}{4a(1-\sin(\phi_1))} =\frac{\sin(2\pi/N)-\sin(4\pi/N)-2\sin(\pi/N)}{2a(1-\sin(2\pi/N))} \end{align} % Hence this gives the assertion % \begin{align} g_{v,4M} = \prod_n d_n \geq 2^{3-M} r^{2M-1} \ensuremath{\delta}^2(2a-\ensuremath{\delta})^2 \prod_{m=3}^M m^2 \cdot \left(\frac{\sin(2\pi/N)-\sin(4\pi/N)-2\sin(\pi/N)}{(1-\sin(2\pi/N))}\right)^{M-3} \end{align} \end{approach} \fi \end{proof} \begin{figure}[t] \centering \begin{subfigure}[b]{0.485\textwidth} \hspace{0.2cm} \includegraphics[width=\textwidth]{glowerbound} \caption{Bounds for $r=1$ without centroid.}\label{fig:glowerbound} \end{subfigure} \begin{subfigure}[b]{0.485\textwidth} \centering \includegraphics[width=0.7\textwidth]{glowerbound_withcenter} \caption{Bounds with centroid.} \label{fig:glowerbound_centroid} \end{subfigure} \caption{Lower Bound for Hexagon Conjecture (green line). Yellow line for bound the $g_{c}$ at centroid circle.} \end{figure} \iflong As can be seen in the plots \figref{fig:glowerbound}. Moreover, the blue line in \figref{fig:glowerbound_centroid}, the bound $(1-\ensuremath{\delta})g_l(1,\ensuremath{\delta})$ is larger that $\ensuremath{\delta}(r^6-\ensuremath{\delta}^6)$. Hence this suggest that for $\ensuremath{\delta}<0.4$ the conjectured bound holds, i.e., we need to show for $\ensuremath{\delta} <0.4$. \begin{align} &\ensuremath{\tilde{\del}} r^7(1-\ensuremath{\tilde{\del}})^2(\sqrt{3}-\ensuremath{\tilde{\del}})(2-\ensuremath{\tilde{\del}})\sqrt{(3-3\ensuremath{\tilde{\del}}+\ensuremath{\tilde{\del}}^2)(1+\ensuremath{\tilde{\del}}^2-\ensuremath{\tilde{\del}})}\geq r^7\ensuremath{\tilde{\del}}(1-\ensuremath{\tilde{\del}}^6)\\ \ensuremath{\Leftrightarrow} \quad&\ensuremath{\delta} (1-\ensuremath{\delta})^2(\sqrt{3}-\ensuremath{\delta})(2-\ensuremath{\delta})\sqrt{(3-3\ensuremath{\delta}+\ensuremath{\delta}^2)(1+\ensuremath{\delta}^2-\ensuremath{\delta})}\geq \ensuremath{\delta}(1-\ensuremath{\delta}^6) \end{align} \fi \iflong \subsection{Nested Hexagons} If we have two nested honeycombs, with an inner hexagon and two outer hexagons, combined on a honeycomb with $12$ vertices, then it has to be shown that $\ensuremath{z^{\text{c}}}$ is still the minimum of the product the distances, see \figref{fig:hexagonsnested}. In fact, if we would shift the blue inner hexagon to the right, we obtain the blue dashed one. We know that the minimal points for this hexagon, for example $\ensuremath{z^{\text{v}}}$ is one of them. However, for the centered, solid blue hexagon this point will be a maximum. Now we need to argument that this point will also yield larger product distances to the outer honeycomb then $\ensuremath{z^{\text{c}}}$. Indeed, by a geometric argument we can see that we only need to show that the three pairs of distances to the three upper $\ensuremath{\alpha_{11}},\ensuremath{\alpha_{10}},\ensuremath{\alpha_9}$ and three lower vertices $\ensuremath{\alpha_{15}},\ensuremath{\alpha_{16}},\ensuremath{\alpha_{17}}$ yield to a smaller product. All distances to the other six vertices will yield to equal or larger distances. For the red pair, it is seen immediately, that $\ensuremath{z^{\text{c}}}$ yield smaller product to $\ensuremath{\alpha}_{11}$ and $\ensuremath{\alpha}_{15}$ since $d^c_{11}/d^v_{11} \simeq 1$ but $d^c_{15}/d^v_{15}\simeq 4\ensuremath{\delta}/5\ensuremath{\delta}=4/5$. Hence $\ensuremath{z^{\text{c}}}$ outperforms $\ensuremath{z^{\text{v}}}$ for this pair. The orange pairs also outperform the gray pairs, as seen in the right drawing of \figref{fig:hexagonsnested}. \begin{figure}[t] \centering \def0.65\textwidth} \footnotesize{ \import{\pdfdir}{OptimaRadiusHuffmanBMOCZ.pdf_tex} {0.63\textwidth} \small{ \import{\pdfdir}{hexagons_nested_pairs.pdf_tex} } \caption{Two nested honeycombs.} \label{fig:hexagonsnested} \end{figure} To make this one more precise, we will show that the minimum is indeed attained at $\ensuremath{z^{\text{c}}}$ if we shift $\ensuremath{z^{\text{c}}}$ by $\ensuremath{\Delta}$ up or down. Let us set $d_{10}=d_{16}=d_{17}=d_9=1$ w.l.o.g. Wee will get the distances \begin{align} h&=\sqrt{1-\ensuremath{\delta}^2}\quad,\quad d_{\ensuremath{\Delta},10}^2& = (h-\ensuremath{\Delta})^2+\ensuremath{\delta}^2 \text{ and } (d_{\ensuremath{\Delta},15})^2=(h+\ensuremath{\Delta})^2+\ensuremath{\delta}^2 \end{align} yielding to \begin{align} d_{\ensuremath{\Delta},10}^2\cdot d_{\ensuremath{\Delta},15}^2 = (h-\ensuremath{\Delta})^2+\ensuremath{\delta}^2)\cdot(h+\ensuremath{\Delta})^2+\ensuremath{\delta}^2) \end{align} $d^c_{11}\cdot d^c_{15}$ than $d^v_{11}\cdot d^v_{15}$ If we extend the hexagon lattice such that we obtain $n$ nested honeycombs, then the minimal zero will be still the one in the centroid, since if we shift to the vertex of the inner honeycomb, the product distance to the $2$nd honeycomb will be larger than the product distance from the circle centered at the origin. Hence, the point $\ensuremath{z^{\text{c}}}$ on the centroid circle will yield a smallest product distance than the point $\ensuremath{z^{\text{v}}}$ at the hexagon vertices. (Needs a solid proof) \todostart Adaption for a circle centered at the line between vertex and origin, with distance $a\geq 0$. Let us assume $z=\sqrt{(\ensuremath{\delta}^2+a^2+2\ensuremath{\delta} a\cos(\ensuremath{\theta}))} e^{i\ensuremath{\theta}}$, then we Assuming that $2{\ensuremath{d_{\text{\rm min}} } }< R-R^{-1}$ and $N>6$ we can place the zeros in the ring on a hexagon with one zero as the centroid, which yields to the smallest possible product of pairwise distances. Assuming, we have more than $N>7$ zeros, a placement on a hexagonal lattice would yield the smallest distance to the origin and therefore to a zero $\ensuremath{\alpha}_m+\ensuremath{\delta} e^{i\ensuremath{\theta}}$ on the circle around the origin. Let us assume we consider for any further hexagon on the circle with radius $nd$ we will place the zeros on a $6\cdot 2^{n-1}-$gon with smaller radius for $n\geq 2$ \begin{align} r_n = \begin{cases} (n-1)(\sqrt{2}d +d), & n \text{ odd}\\ (n-1)\sqrt{2}d, & n \text{ even} \end{cases} \end{align} % Hence, for $N\geq 6+1$ we get for the numerator % \begin{align} \ensuremath{\delta}\prod_n |\ensuremath{\delta} e^{i\ensuremath{\theta}} -\ensuremath{\tilde{\alp}}_n| \geq \frac{\ensuremath{\delta}^3}{2^2} \frac{63}{64} ({\ensuremath{d_{\text{\rm min}} } })^{N-6} = \ensuremath{\delta}^{N-3} 2^{N-8} \frac{63}{64} \end{align} % since every neighbor zero outside the polygon is at least ${\ensuremath{d_{\text{\rm min}} } }\geq 2\ensuremath{\delta}$ apart. For the numerator we will upper bound by the largest possible $\ensuremath{\alpha}_m+\ensuremath{\delta} e^{i\ensuremath{\theta}}$. Since $\ensuremath{\alpha}_m$ can not lie on the outer boundary, see circle packing for $N\geq 7$, we get $|\ensuremath{\alpha}_m+\ensuremath{\delta} e^{i\ensuremath{\theta}}|\leq |\ensuremath{\alpha}_m|+\ensuremath{\delta}|\leq R$ and hence % \begin{align} \min_m \min_{\ensuremath{\theta}} f_m(\ensuremath{\theta}) \geq |x_N|^2\ensuremath{\delta}^{N-3} 2^{N-8} \frac{63}{64} \frac{R^2-1}{R^{2N}-1} \end{align} % which proves the bound \eqref{eq:noisebound2}. \todoend \fi \section{Continuity and Robustness of Zeros Against Small Perturbations} \label{sec:rootstability} Although, the SDP gives insight in the robustness of Huffman sequences, it relies on the knowledge of the channel autocorrelation. Moreover, as we found in \secref{sec:dizet}, the performance of the DiZeT decoder depends on the distribution of the zero-symbols. Hence, a robustness analysis for a zero-based modulation, boils down to a robustness analysis of polynomial zeros. Wilkinson investigated at first in \cite{Wil84} the stability of polynomial roots under perturbation of the polynomial coefficients. One extreme case of instability is known today as the Wilkinson polynomial \begin{align} {\ensuremath{\mathrm x}}(z)=(z-1)(z-2)(z-3)\cdots(z-20)\label{eq:wilkinson} \end{align} given by $20$ real-valued zeros equidistant placed on the positive real line. If only the leading coefficient is disturbed by machine precession \begin{align} y_{20}=x_{20}+10^{-23}, \end{align} then the three largest zeros of the perturbed polynomial ${\ensuremath{\mathrm y}}(z)=\sum_n y_n z^n$ are completely off, showing that the zeros are not stable against distortion on its coefficients. This can be generalized to arbitrary polynomials and the question is, if we consider the Eulcidean norm, how much the zeros will be disturbed if we perturb the coefficients with some ${\ensuremath{\mathbf w}}\in{\ensuremath{\mathbb C}}^N$ having $\Norm{{\ensuremath{\mathbf w}}}_2^2\leq \ensuremath{\epsilon}$. The answer was given in \cite{FH07} in terms of \emph{root neighborhoods} or \emph{pseudozero sets} \begin{align} \ensuremath{Z}(\ensuremath{\epsilon},{\ensuremath{\mathrm X}}) =\set{z\in{\ensuremath{\mathbb C}}}{\frac{|\sum_n x_n z^n |^2}{{\sum_n |z|^{2n}}}\leq \ensuremath{\epsilon}} \end{align} where each disturbed polynomial ${\ensuremath{\mathrm y}}(z)=\sum_n (x_n + w_n)^n$ for some $\Norm{{\ensuremath{\mathbf w}}}_2\leq \ensuremath{\epsilon}$ has all its zeros $\zeta$ in $\ensuremath{Z}(\ensuremath{\epsilon},{\ensuremath{\mathrm X}})$. However, this characterization of the root neighborhoods, does not explain at which noise level $\ensuremath{\epsilon}$ the single root neighborhoods $\ensuremath{Z}_n(\ensuremath{\epsilon},{\ensuremath{\mathrm X}})$ of the roots $\zeta_n$ will start to overlap. The intuition suggest, that with increasing noise power, the single root neighborhoods should monotone grow and eventually start to overlap, at which a unique zero separation becomes impossible, see \figref{fig:receivedzerosperturbation}. We plotted here a fixed Huffman polynomial with $K=6$ zeros (black squares) and $3$ channel zeros (red squares) generated by Gaussian random vectors (Kac polynomial). The additional channel zeros have only little impact of the root neighborhoods of the Huffman zeros. However, they will have an heavy impact on the zero separation (decoding), if they get close to the zero-codebook. Since the distribution of the Chanel zeros is random, we will only consider the perturbation analysis of a given polynomial ${\ensuremath{\mathrm X}}(z)$. \begin{figure}[t] \centering \includegraphics[width=0.5\textwidth]{No5db_rootneighborhood_run1000tht500_N6L3_allinside_Runi_lam1.png} \caption{Root Neighborhoods for $7$ noise powers between $-22$dB and $-5$dB for $K=6$ Huffman zeros and $L-1=3$ channel zeros.}\label{fig:receivedzerosperturbation} \end{figure} To derive such a quantized result we will exploit Rouch{\'e}'s Theorem to bound the single root neighborhoods by discs, see e.g. \cite[Thm (1,3)]{Mar77}. \begin{thm}[Rouch{\'e}]\label{thm:rouche} Let ${\ensuremath{\mathrm x}}(z)$ and $\uw(z)$ be analytic functions in the interior to a simple closed Jordan curve $C$ and continuous on $C$. If % \begin{align} |\uw(z)|\leq |{\ensuremath{\mathrm x}}(z)|,\quad z\in C, \end{align} % then ${\ensuremath{\mathrm y}}(z)={\ensuremath{\mathrm x}}(z)+\uw(z)$ has the same number of zeros interior to $C$ as does ${\ensuremath{\mathrm x}}(z)$. \end{thm} The Theorem allows to prove that the zeros of polynomials are continuous functions of the coefficients, see \cite[Thm (1,4)]{Mar77}. \if0 % \begin{thm}[Continuity for Simple Zeros] Let ${\ensuremath{\mathrm X}}(z)=x_{\ensuremath{L}}\prod_{l=1}^{\ensuremath{L}}(z-\ensuremath{\alpha}_l)$ with ${\ensuremath{L}}$ simple zeros $\ensuremath{\alpha}_l$ and let ${\ensuremath{\mathbf x}}\in{\ensuremath{\mathbb C}}^{{\ensuremath{L}}+1}$ be its coefficients. Let us denote the minimal distance of $\ensuremath{\alpha}_l$ to all other zeros by % \begin{align} d_l= \min_{k\not=l} |\ensuremath{\alpha}_l-\ensuremath{\alpha}_k|. \end{align} % Then each additive distortion ${\ensuremath{\mathbf w}}\in{\ensuremath{\mathbb C}}^{{\ensuremath{L}}+1}$ on the coefficient generates a polynomial % \begin{align} {\ensuremath{\mathrm Y}}(z)= \sum_{l=0}^{\ensuremath{L}} (x_l+w_l)z^l=\prod_{l=1}^{\ensuremath{L}} (z-\ensuremath{\gamma}_l) \end{align} % with zeros $\ensuremath{\gamma}_l$. Then there exists an $\ensuremath{\epsilon}\geq \Norm{{\ensuremath{\mathbf w}}}_{\infty}$ such that only one zero $\ensuremath{\gamma}_l$ lies in $B(\ensuremath{\alpha}_l,d_l)$. % \end{thm} % Simple estimation of the fluctuation of the coefficients in terms of bounding the modulus of the polynomial in an arbitrary disc depends on the first coefficient (trivial polynomial). See \cite[p.16]{Bie28}. \fi However, to obtain an explicit robustness result for the zeros, we need a quantized version of the continuity, i.e., a Lipschitz bound of the root functions with respect to the $\ell_{\infty}/\ell_2$ norm. As simple closed Jordan curves we will consider the Euclidean circle and the disc as its interior, which will contain the single root neighborhoods. Let us define for $\ensuremath{\alpha}_n\in{\ensuremath{\mathbb C}}$ the closed Euclidean ball (disc) of radius $\ensuremath{\delta}>0$ and its boundary as \begin{align} B_n(\ensuremath{\delta})=B(\ensuremath{\delta},\ensuremath{\alpha}_n) = \set{z\in{\ensuremath{\mathbb C}}}{|z-\ensuremath{\alpha}_n|\leq \ensuremath{\delta}}\quad,\quad C_n(\ensuremath{\delta})= \set {z\in{\ensuremath{\mathbb C}}}{|z-\ensuremath{\alpha}_n|=\ensuremath{\delta}}. \end{align} Let us consider an arbitrary polynomial (analytic function in ${\ensuremath{\mathbb C}}$) of order $\xord\geq 1$: \begin{align} {\ensuremath{\mathrm x}}(z)=\sum_{\xind=0}^{\xord} x_\xind z^\xind. \end{align} Then, its roots are functions of the polynomial coefficients ${\ensuremath{\mathbf x}}\in{\ensuremath{\mathbb C}}^{\xord+1}$ given by \begin{align} \ensuremath{\alpha}_n= \ensuremath{\alpha}_n({\ensuremath{\mathbf x}})\quad,\quad n=1,\dots, \xord. \end{align} If the coefficients are disturbed by a vector ${\ensuremath{\mathbf w}}\in{\ensuremath{\mathbb C}}^{\xord+1}$, the maximal perturbation of the zeros should be bounded by \begin{align} \max_n |\ensuremath{\alpha}_n({\ensuremath{\mathbf x}}+{\ensuremath{\mathbf w}})-\ensuremath{\alpha}_n({\ensuremath{\mathbf x}})| \leq \ensuremath{\delta} \cdot \Norm{{\ensuremath{\mathbf x}}+{\ensuremath{\mathbf w}}-{\ensuremath{\mathbf x}}}_{2} = \ensuremath{\delta} \cdot \Norm{{\ensuremath{\mathbf w}}}_{2}, \end{align} where the bound $\ensuremath{\delta}=\ensuremath{\delta}(\ensuremath{\epsilon},{\ensuremath{\mathbf x}})>0$ is a \emph{local Lipschitz constant} for ${\ensuremath{\mathbf w}}\in B(\ensuremath{\epsilon},{\ensuremath{\mathbf x}})$, which we want to derive. If the noise coefficient $w_\xord=-x_\xord$, i.e., the leading coefficient is vanishing, then we will set $\ensuremath{\alpha}_\xord({\ensuremath{\mathbf x}}+{\ensuremath{\mathbf w}})=0$, since the order of the perturbed polynomial would reduce to $\xord-1$. We are now ready to prove the following local Lipschitz bound. We use here the assumption that one zero is outside the unit circle, which is always the case for polynomials generated by autocorrelations. \begin{thm}\label{thm:zerodistortion} Let ${\ensuremath{\mathrm x}}(z)\in{\ensuremath{\mathbb C}}[z]$ be a polynomial of order $\xord>1$ with simple zeros $\ensuremath{\alpha}_1,\dots,\ensuremath{\alpha}_{\xord} \subset {\ensuremath{\mathbb C}}$ inside a circle of radius $R>1$ with minimal pairwise distance ${\ensuremath{d_{\text{\rm min}} } }>0$, i.e. % \begin{align} {\ensuremath{d_{\text{\rm min}} } } :=\min_{n\not=k} |\ensuremath{\alpha}_n-\ensuremath{\alpha}_k| \quad,\quad R=\arg\max_n |\ensuremath{\alpha}_n|. \end{align} % Let ${\ensuremath{\mathbf w}}\in{\ensuremath{\mathbb C}}^{\xord}$ with $\Norm{{\ensuremath{\mathbf w}}}_{2}\leq \ensuremath{\epsilon}$ be an additive perturbation on the polynomial coefficients ${\ensuremath{\mathbf x}}$ and $\ensuremath{\delta}\in[0,{\ensuremath{d_{\text{\rm min}} } }/2)$. Then the $n$th zero $\zeta_n$ of the disturbed polynomial ${\ensuremath{\mathrm y}}(z)={\ensuremath{\mathrm x}}(z)+\uw(z)$ lies in $B_n(\ensuremath{\delta})$ if % \begin{align} \ensuremath{\epsilon}=\ensuremath{\epsilon}({\ensuremath{\mathbf x}},\ensuremath{\delta})\leq \frac{|x_N| \ensuremath{\delta} ({\ensuremath{d_{\text{\rm min}} } }-\ensuremath{\delta})^{N-1}}{\sqrt{1+ N}(R+\ensuremath{\delta})^\xord}\label{eq:noisebound}. \end{align} \end{thm} \begin{remark} The minimal pairwise distance of the zeros is also called zero separation, see for example \cite[Sec.11.4]{Zip93}. \end{remark} \begin{proof} The proof is a quantized version of the proof in \cite[Thm (1,4)]{Mar77}. Let us define the error polynomial % \begin{align} \uw(z)= \sum_{n=0}^{\xord} w_n z^n\label{eq:wz}. \end{align} % By defining $\underline{{\ensuremath{\mathbf z}}}=(z^0,z^1,\dots,z^{\xord})^T$, we can upper bound the magnitude of $\uw$ with the Cauchy-Schwarz inequality % \begin{align} |\uw(z)| = |{\ensuremath{\mathbf w}}^T\underline{{\ensuremath{\mathbf z}}}|\leq \Norm{{\ensuremath{\mathbf w}}}_{2} \cdot \Norm{\underline{{\ensuremath{\mathbf z}}}}_2 = \ensuremath{\epsilon} \cdot \Big(\sum_n |z^n|^2\Big)^{1/2} = \ensuremath{\epsilon} \cdot \Big(\sum_n |z|^{2n}\Big)^{1/2} = \ensuremath{\epsilon}\cdot f(|z|). \end{align} % Since $f(r)$ is monotone increasing\footnote{Note, $(r+\ensuremath{\epsilon})^k > r^k +\ensuremath{\epsilon}^k+ \dots >r^k$ for $r,\ensuremath{\epsilon}>0$ and $k\geq 1$.} in $r>0$, the largest upper bound in $C_m(\ensuremath{\delta})$ is attained at $z=|\ensuremath{\alpha}_m|+\ensuremath{\delta}$ and hence % \begin{align} f(|\ensuremath{\alpha}_m|+\ensuremath{\delta})^2\leq\begin{cases} 1+ N\cdot (|\ensuremath{\alpha}_m|+\ensuremath{\delta})^{2N}&, |\ensuremath{\alpha}_m|+\ensuremath{\delta}>1\\ 1+ N\cdot(|\ensuremath{\alpha}_m|+\ensuremath{\delta})^2 &, |\ensuremath{\alpha}_m|+\ensuremath{\delta}\leq 1 \end{cases}\label{eq:fbound} \end{align} \if0 We seek upper bounds for $f(z)$ in $z\in \bigcup B_l(\ensuremath{\delta}_-) \cup B_l(\ensuremath{\delta}_+)$ % \begin{align} f_l(z)&\leq \frac{ |\ensuremath{\alpha}_l|^{2(n+1)} ( 1+ \frac{\ensuremath{\delta}}{|\ensuremath{\alpha}_l|})^{2(n+1)} }{1-(|\ensuremath{\alpha}_l|+\ensuremath{\delta})^2} \leq \frac{1-|\ensuremath{\alpha}_l|^{2(n+1} (1+\frac{(n+1)\ensuremath{\delta}}{|\ensuremath{\alpha}_l|})^2}{1-(|\ensuremath{\alpha}_l|+\ensuremath{\delta})^2}\\ &= |\ensuremath{\alpha}_l|^{2(n+1)}\frac{(1+\frac{(n+1)\ensuremath{\delta}}{|\ensuremath{\alpha}_l|})^2 - |\ensuremath{\alpha}_l|^{-2(n+1)}}{|\ensuremath{\alpha}_l|^2 -1 + \ensuremath{\delta}^2 +2\ensuremath{\delta}|\ensuremath{\alpha}_l|} \leq |\ensuremath{\alpha}_l|^{2(n+1)}\frac{(1+\frac{(n+1)\ensuremath{\delta}}{|\ensuremath{\alpha}_l|})^2}{(1-\ensuremath{\delta})^2 -1 +\ensuremath{\delta}^2 + 2\ensuremath{\delta}(1+\ensuremath{\delta})|}\\ &\leq \left (\frac{|\ensuremath{\alpha}_l|^{n}}{2}(\frac{|\ensuremath{\alpha}_l|}{\ensuremath{\delta}} + n+1)\right)^2 \end{align} Hence \begin{align} |\uw(z)|\leq \ensuremath{\epsilon} \cdot (\frac{|\ensuremath{\alpha}_l|^{n}}{2}(\frac{|\ensuremath{\alpha}_l|}{\ensuremath{\delta}} + n+1) \end{align} % On the other hand, we need to lower bound the original polynomial in $B_l(z)$ by % \begin{align} |{\ensuremath{\mathrm x}}(z)|=|x_n| \prod_{k=1}^N |z-\ensuremath{\alpha}_k| \geq |x_n| \ensuremath{\delta} \prod_{l\not=k=1}^n \big| |\ensuremath{\alpha}_l-\ensuremath{\alpha}_k|-\ensuremath{\delta}\big| \geq |x_n|\ensuremath{\delta} ({\ensuremath{d_{\text{\rm min}} } } -\ensuremath{\delta})^{N-1} \end{align} % Since we have $(n-1)\ensuremath{\delta}_l<d_l$ we get again with Bernoulli % \begin{align} |x_n|\ensuremath{\delta}_l d_l^{n-1} (1-\frac{(n-1)\ensuremath{\delta}_l}{d_l}) \geq |{\ensuremath{\mathrm x}}(z)|\geq |x_n| \ensuremath{\delta}_l d_l^{n-1}(1-\frac{\ensuremath{\delta}_l}{d_l})^{n-1} \end{align} % If % \begin{align} \ensuremath{\epsilon} \cdot (\frac{|\ensuremath{\alpha}_l|^{n}}{2}(\frac{|\ensuremath{\alpha}_l|}{\ensuremath{\delta}_l} + n+1) &\leq |x_n|\ensuremath{\delta}_l d_l^{n-1} (1-\frac{(n-1)\ensuremath{\delta}_l}{d_l})\\ \ensuremath{\Leftrightarrow} &\ensuremath{\epsilon} =\ensuremath{\epsilon}(\ensuremath{\delta},{\ensuremath{\mathrm x}})\leq \frac{ 2|x_n| d_l^{n-2} (d_l-(n-1)\ensuremath{\delta}_l)}{|\ensuremath{\alpha}_l|^n (|\ensuremath{\alpha}_l|+ (n+1)\ensuremath{\delta}_l)} \end{align} % \fi By assumption it holds $R=|\alpmax|>1$ which gives us the universal upper bound\footnote{This is actually Bernstein's Lemma.} \begin{align} |\uw(z)| \leq \ensuremath{\epsilon} \cdot \sqrt{1+N}(R+\ensuremath{\delta})^N \quad,\quad z\in\bigcup C_m(\ensuremath{\delta}). \end{align} On the other hand, the magnitude of the original polynomial \begin{align} |{\ensuremath{\mathrm x}}(z)| &= |x_\xord| \prod_{n=1}^\xord |z-\ensuremath{\alpha}_n| \quad,\quad z \in C_m(\ensuremath{\delta})\\ &= |x_\xord| \prod_n |\ensuremath{\alpha}_m+\ensuremath{\delta} e^{i\ensuremath{\theta}} -\ensuremath{\alpha}_n|\quad,\quad \ensuremath{\theta}\in[0,2\pi) \intertext{can be lower bounded by using the reverse triangle inequality \footnotemark} & \geq |x_\xord|\prod_n | |\ensuremath{\alpha}_m-\ensuremath{\alpha}_n|-\ensuremath{\delta}| \geq |x_\xord| \ensuremath{\delta} \prod_{n\not=m} ({\ensuremath{d_{\text{\rm min}} } }-\ensuremath{\delta}).\label{eq:brutallowerbound} \end{align} \footnotetext{Note, that $|\ensuremath{\alpha}_l-\ensuremath{\alpha}_n|>{\ensuremath{d_{\text{\rm min}} } }>\ensuremath{\delta}$ for $l\not=n$.}% Hence we get for all $z\in\bigcup C_n(\ensuremath{\delta})$: \begin{align} |{\ensuremath{\mathrm x}}(z)|\geq |x_\xord| \ensuremath{\delta} ({\ensuremath{d_{\text{\rm min}} } }-\ensuremath{\delta})^{N-1}. \end{align} To apply Rouch{\'e}'s Theorem, we have to show $|\uw(z)|<|{\ensuremath{\mathrm x}}(z)|$ for all $z\in\bigcup C_n(\ensuremath{\delta})$, which gives us the universal bound \begin{align} \ensuremath{\epsilon}=\ensuremath{\epsilon}({\ensuremath{\mathbf x}},\ensuremath{\delta}) \leq \frac{|x_\xord| \ensuremath{\delta} ({\ensuremath{d_{\text{\rm min}} } }-\ensuremath{\delta})^{N-1}}{\sqrt{1+ N}(R+\ensuremath{\delta})^N}.\label{eq:boundproof} \end{align} Since $\ensuremath{\delta}<{\ensuremath{d_{\text{\rm min}} } }/2$, all $B_n(\ensuremath{\delta})$ are disjoint and ${\ensuremath{\mathrm y}}(z)$ has exactly one zero in each $n$th ball $B_n(\ensuremath{\delta})$ by \thmref{thm:rouche}. Note, that $x_N=x_N(\ensuremath{\boldsymbol{ \alp}})$ depends on the selected zeros and the normalization $\Norm{{\ensuremath{\mathbf x}}}=1$. \end{proof} \ifwrong\color{red} Since the lower bound in \eqref{eq:brutallowerbound} holds for any normalized polynomial with minimal zero separation ${\ensuremath{d_{\text{\rm min}} } }$ and zeros with modulus between $[R^{-1},R]$, we can choose the polynomial with largest possible $|x_N|$. However, the modulus constrain does not allow vanishing zeros. This implies immediately that $x_0\not=0$. Hence we get the polynomial \begin{align} {\ensuremath{\mathrm P}}(z) = x_0 -x_N z^N\quad,\quad |x_0|^2+|x_N|^2=1 \end{align} with zeros \begin{align} \ensuremath{\alpha}_n=(\frac{|x_0|}{|x_N|})^{1/N} e^{i2\pi n/N} =R^{-1} e^{i2\pi/N} \end{align} which are exactly the Huffman zeros with all zeros inside the unit circle. Clearly, if $R>$, this gives the largest modulus of the leading coefficient \begin{align} |x_N|= \frac{1}{\sqrt{R^{-2N}+1}} \end{align} If $R=\infty$ this yields $|x_N|=1$ but also drops the restriction on the modulus of the zeros. If all zeros are on the lower bound of the Ring, the largest possible pairwise distance is exactly ${\ensuremath{d_{\text{\rm min}} } }=2R^{-1}\sin(\pi/N)$ which is only achieved for the Huffman sequences, the all zero bit. Restricting the zero separation to ${\ensuremath{d_{\text{\rm min}} } }\leq 2R^{-1}\sin(\pi/N)$, allows therefore a sharper bound. Hence we have shown. \begin{cori} Let ${\ensuremath{\mathrm X}}(z)\in{\ensuremath{\mathbb C}}[z]$ be a polynomial as in \thmref{thm:zerodistortion} and demand ${\ensuremath{d_{\text{\rm min}} } }\in (0,2\sin(\pi/N)$. Then any additive distortion by $\Norm{{\ensuremath{\mathbf w}}}_2\leq \ensuremath{\epsilon}$ on the coefficients ${\ensuremath{\mathbf x}}$ of ${\ensuremath{\mathrm X}}(z)$ will keep the $n$th distorted zero $\ensuremath{\alpha}_n$ of ${\ensuremath{\mathrm X}}(z)$ in the ball $B_n(\ensuremath{\delta})$ with center at $\ensuremath{\alpha}_n$ and radius $\ensuremath{\delta}$ if % \begin{align} \ensuremath{\epsilon}({\ensuremath{\mathbf x}},\ensuremath{\delta})\leq \frac{\ensuremath{\delta}({\ensuremath{d_{\text{\rm min}} } }-\ensuremath{\delta})^{N-1}}{\sqrt{(R^{2-N}+1)(1+ N)}(R+\ensuremath{\delta})^N}.\label{eq:boundeins} \end{align} \end{cori} \fi \color{black \begin{remark} Let us note, that the bound \eqref{eq:boundproof} {\bfseries does not increases} with $\ensuremath{\delta}$ for fixed ${\ensuremath{\mathbf x}},R$ and ${\ensuremath{d_{\text{\rm min}} } }$, see \figref{fig:theorem2_epsbound}. This behaviour is due to the continuity of the zeros very unlikely and hence caused by the worst bound in \eqref{eq:brutallowerbound}. In \secref{sec:sharperbounds} we will investiage in more detail the geometric structure of the zero placements, to obtain sharper stability bounds. % \begin{figure} \centering \includegraphics[width=0.5\linewidth]{theorem2_epsbound} \caption{Noise bound \eqref{eq:boundproof} for fixed Radius $R=1.1$ and ${\ensuremath{d_{\text{\rm min}} } }=0.5$ over $\delta$.} \label{fig:theorem2_epsbound} \end{figure} % \iflong \todostart Hence, to allow a large noise power $\ensuremath{\epsilon}$ the radius $\ensuremath{\delta}$ of the root neighborhoods hast to be maximal, i.e. $\ensuremath{\delta}={\ensuremath{d_{\text{\rm min}} } }/2$. This implies, that if the noise power increases slowly form zero to $\ensuremath{\epsilon}$, the maximal possible zero distortion $\ensuremath{\delta}\geq\max_n|\ensuremath{\alpha}_n-\zeta_n|$ will also monotone increase. Hence, it is not possible, that the original zero can move to a different neighborhood. Note, that a uniform growth in the Euclidean metric might not accurately represent the true growth behaviour of the root neighborhoods and can only be used as a non-tight (local) upper bound behaviour. See here the literature to root neighborhoods or pseudozero sets, \cite{Mos86,FH07}. \todoend \fi \end{remark} Furthermore, if $|\alpmax|=\const$ and $|x_N|=\const$, then a maximal separation of the zeros yields to robustness against additive noise on the coefficients. Hence, if we place the zeros with maximal pairwise distance for fixed $R$, this suggests a good BER performance for the RFMD decoder. Moreover, by setting $\ensuremath{\delta}={\ensuremath{d_{\text{\rm min}} } }/2$ the bound \eqref{eq:noisebound} gives \begin{align} \ensuremath{\epsilon} \leq \frac{|x_\xord|}{\sqrt{1+N}}\frac{{\ensuremath{d_{\text{\rm min}} } }^N}{2^N (|\alpmax|+{\ensuremath{d_{\text{\rm min}} } }/2)^N)} \label{eq:epsbad}, \end{align} which is a upper threshold of the noise power under which no errors can occur. It can be seen that the noise bound increases if ${\ensuremath{d_{\text{\rm min}} } }$ increases, which again validates a larger zero separation. For Huffman sequences with radius $R$ we obtain ${\ensuremath{d_{\text{\rm min}} } }=2R\sin(\pi/N)$ and hence \eqref{eq:epsbad} gives \begin{align} \ensuremath{\epsilon}\leq \frac{1}{\sqrt{1+N}} \frac{1}{\sqrt{(R^{-2N}+1)}(1/\sin(\pi/N) +1)^N}\label{eq:noisebound_huff}. \end{align} Note, the bound becomes independent of $R$ if one zero is outside the unit circle and hence equal to $R$. A plot for different $N$ is given in \figref{fig:huffmaneps} for uniform radius $\Ropt$ in \eqref{eq:optimalradiusexactrho}. \begin{figure} \centering \includegraphics[width=0.8\linewidth]{noisebound_huffman_thm2} \caption{SNR bound \eqref{eq:noisebound_huff} with $\ensuremath{\delta}=\ensuremath{\del_{\text{max}}}$ for Huffman sequences over various dimensions $N$ and uniform radius $\Ropt(N,1)$ in \eqref{eq:optimalradiusexactrho} allowing a perfect reconstruction.} \label{fig:huffmaneps} \end{figure} Moreover, if $|\alpmax|=t{\ensuremath{d_{\text{\rm min}} } }/2$ for some $t\geq 1$ then we get \begin{align} \ensuremath{\epsilon} \leq \frac{|x_\xord|}{\sqrt{1+N}(t+1)^N}. \end{align} However, an increase of $t$ means an increase of the largest root, which is coupled by the leading coefficient, due to a result of Cauchy \begin{align} |\alpmax|\leq 1+ \max_{k<N} {\Big|\frac{x_k}{x_\xord}\Big|} \end{align} see for example \cite[Thm.(27,2)]{Mar77}. If the energy of ${\ensuremath{\mathbf x}}$ is normalized this gives \begin{align} |\alpmax| \leq 1 + \frac{1}{|x_\xord|}\quad \ensuremath{\Leftrightarrow} \quad |x_\xord|\leq \frac{1}{|\alpmax|-1} \end{align} since $|\alpmax|>1$. Hence, if $|\alpmax|$ increases, the leading coefficient has to decrease and $\ensuremath{\epsilon}$ decreases rapidly, independently of ${\ensuremath{d_{\text{\rm min}} } }$. \if0 Since $\ensuremath{\epsilon}(\ensuremath{\delta},{\ensuremath{\mathrm x}})$ is monotone increasing if $\ensuremath{\delta}$ increases and vice versa. Hence, to obtain large s \begin{align} f_l(|z|)^2={\frac{1-(|\ensuremath{\alpha}_l|+\ensuremath{\delta})^{2(n+1)}}{1-(|\ensuremath{\alpha}_l|+\ensuremath{\delta})^2}} \end{align} % Let us assume $|\ensuremath{\alpha}_l|-\ensuremath{\delta}>1$, then we can upper bound with the Bernoulli inequality \begin{align} f_l(z)&\leq \frac{ 1- |\ensuremath{\alpha}_l|^{2(n+1)} ( 1+ \frac{\ensuremath{\delta}}{|\ensuremath{\alpha}_l|})^{2(n+1)} }{1-(|\ensuremath{\alpha}_l|+\ensuremath{\delta})^2} \leq \frac{1-|\ensuremath{\alpha}_l|^{2(n+1} (1+\frac{(n+1)\ensuremath{\delta}}{|\ensuremath{\alpha}_l|})^2}{1-(|\ensuremath{\alpha}_l|+\ensuremath{\delta})^2}\\ &= |\ensuremath{\alpha}_l|^{2(n+1)}\frac{(1+\frac{(n+1)\ensuremath{\delta}}{|\ensuremath{\alpha}_l|})^2 - |\ensuremath{\alpha}_l|^{-2(n+1)}}{|\ensuremath{\alpha}_l|^2 -1 + \ensuremath{\delta}^2 +2\ensuremath{\delta}|\ensuremath{\alpha}_l|} \leq |\ensuremath{\alpha}_l|^{2(n+1)}\frac{(1+\frac{(n+1)\ensuremath{\delta}}{|\ensuremath{\alpha}_l|})^2}{(1-\ensuremath{\delta})^2 -1 +\ensuremath{\delta}^2 + 2\ensuremath{\delta}(1+\ensuremath{\delta})|}\\ &\leq \left (\frac{|\ensuremath{\alpha}_l|^{n}}{2}(\frac{|\ensuremath{\alpha}_l|}{\ensuremath{\delta}} + n+1)\right)^2 \end{align} Hence \begin{align} |\uw(z)|\leq \ensuremath{\epsilon} \cdot (\frac{|\ensuremath{\alpha}_l|^{n}}{2}(\frac{|\ensuremath{\alpha}_l|}{\ensuremath{\delta}} + n+1) \end{align} % On the other hand, we need to lower bound the original polynomial in $B_l(z)$ by % \begin{align} |{\ensuremath{\mathrm x}}(z)|\geq |x_n| \ensuremath{\delta}_l \prod_{l\not=k=1}^n (|\ensuremath{\alpha}_l-\ensuremath{\alpha}_k|-\ensuremath{\delta}_l) \geq |x_n|\ensuremath{\delta}_l (d_l -\ensuremath{\delta}_l)^{n-1} \end{align} % Since we have $(n-1)\ensuremath{\delta}_l<d_l$ we get again with Bernoulli % \begin{align} |x_n|\ensuremath{\delta}_l d_l^{n-1} (1-\frac{(n-1)\ensuremath{\delta}_l}{d_l}) \geq |{\ensuremath{\mathrm x}}(z)|\geq |x_n| \ensuremath{\delta}_l d_l^{n-1}(1-\frac{\ensuremath{\delta}_l}{d_l})^{n-1} \end{align} % If % \begin{align} \ensuremath{\epsilon} \cdot (\frac{|\ensuremath{\alpha}_l|^{n}}{2}(\frac{|\ensuremath{\alpha}_l|}{\ensuremath{\delta}_l} + n+1) &\leq |x_n|\ensuremath{\delta}_l d_l^{n-1} (1-\frac{(n-1)\ensuremath{\delta}_l}{d_l})\\ \ensuremath{\Leftrightarrow} &\ensuremath{\epsilon} =\ensuremath{\epsilon}(\ensuremath{\delta},{\ensuremath{\mathrm x}})\leq \frac{ 2|x_n| d_l^{n-2} (d_l-(n-1)\ensuremath{\delta}_l)}{|\ensuremath{\alpha}_l|^n (|\ensuremath{\alpha}_l|+ (n+1)\ensuremath{\delta}_l)} \end{align} % then it holds $|\uw(z)|<|{\ensuremath{\mathrm x}}(z)|$ for $z\in B_l(\ensuremath{\delta})$ and by Rouches Theorem it holds that ${\ensuremath{\mathrm y}}(z)$ has exactly one zero in $B_l(\ensuremath{\delta})$. Since $\ensuremath{\epsilon}(\ensuremath{\delta},{\ensuremath{\mathrm x}})$ is monotone increasing if $\ensuremath{\delta}$ increases and vice versa. Hence, to obtain large stability we need large minimal pairwise distance $d_{min}=\min d_l$. % Now we have to do the same bound techniques for the case $|\ensuremath{\alpha}_l|+\ensuremath{\delta}<1$. Then unify over all zeros by using $\ensuremath{\delta} <d_{min}$ and conclude with the stability result. \fi \paragraph{Zeros of Random Channels} It is known, that a polynomial with i.i.d. Gaussian distributed coefficients has zeros concentrated around the unit circle. If the order $\xord$ goes to infinity, all zeros will be uniformly distributed on the unit circle with probability one, see for example \cite{PY14} In fact, this even holds for other random polynomials with non Gaussian distributions, see \cite{HN08}. This is an important observation, since it implies for fixed $K$ and hence $R$, that an increase of $L$ will concentrate the channel zeros on the unit circle, such that the channel zeros will not interfere with the codebook zeros, as long as $R$ is sufficiently large. \begin{remark} The analysis of the stability radius for a certain zero-codebook and noise power, allows in principle an error detection for the RFMD decoder. Here, an error for the $l$th zero can only occur if the noise power is larger than the RHS of \eqref{eq:noisebound}. However, in the presence of the channel $\vh$, we can adopt the dimension $N$ and $x_N\rightarrow x_{\ensuremath{K}} h_{{\ensuremath{L}}-1}$, if we assume the absolute values of the zeros of ${\ensuremath{\mathrm H}}(z)$ are not larger than $R$. The minimal distance might be fulfilled with a certain probability. A precise analysis of the expectation might lead to upper bounds of the bit error probabilities of the RFMD decoder, which will be a future research topic. Note also, that is not clear, what the distribution of the disturbed zeros $\zeta_n$ are. If they would be Gaussian known results of polar quantization might apply, see for example \cite{NCTH14}. Huffman sequences for $R=1$ are uniformly concentrated on a unit circle and show the best noise robustness \figref{fig:hufzerospread0}. \end{remark} \newcommand{\ensuremath{d_{nn}}} \newcommand{\dcp}{\ensuremath{d_{cp}}}{\ensuremath{d_{nn}}} \newcommand{\dcp}{\ensuremath{d_{cp}}} \begin{figure}[t] \centering \def0.65\textwidth} \footnotesize{ \import{\pdfdir}{OptimaRadiusHuffmanBMOCZ.pdf_tex} {0.65\textwidth} \footnotesize{ \import{\pdfdir}{OptimaRadiusHuffmanBMOCZ.pdf_tex} } \vspace{-0.28\textwidth} \caption{Zero Constellations, which allows largest non-overlapping uniform root-neighborhood discs for the Huffman BMOCZ scheme with largest radius $\ensuremath{\del_{\text{max}}}={\ensuremath{d_{\text{\rm min}} } }/2$.} \label{fig:optimalradius} \end{figure} \subsection{Radius for Huffman BMOCZ Allowing Largest Uniform Root Neighborhood Discs} To ensure robustness of a zero-based decoding against additive noise we need to place the zero-constellations carefully, as will be pointed out in more detail in \secref{sec:rootstability}. \thmref{thm:zerodistortion} below suggest for Huffman polynomials to place the zero-set $\Alp$ with maximal pairwise distance under the constraint of being uniformly spaced conjugate-reciprocal pairs. Here the distance between conjugated pairs is given by \begin{align} d_{cp}=R-R^{-1} \end{align} and the distance between next-neighbor pairs for the smaller radius $R^{-1}$ by \begin{align} d_{nn}= 2R^{-1}\sin(\pi/{\ensuremath{K}}). \end{align} Setting both distances equal, yields a zero-set $\Alp$ with maximal minimal pairwise distance ${\ensuremath{d_{\text{\rm min}} } }$, see \figref{fig:optimalradius}. However, simulations of perturbed Huffman polynomials, see \figref{fig:zeroperturbationsim}, show a strong dependence of the root neighborhood radius. In fact, an increasing of $R$ yields to an increasing of the root neighborhood radius $\ensuremath{\delta}$, which obtains its minimum if $R=1$. However, if $R$ gets to small, the root neighborhood of the reciprocal-pairs will overlap. To address this problem we will introduce $\ensuremath{\lambda}\geq 1$ as a scaling parameter which yields to \begin{align} \ensuremath{\lambda} d_{cp}=d_{nn} \quad&\ensuremath{\Leftrightarrow} \quad \ensuremath{\lambda}(R^2-1)=Rd_{nn}=2\sin(\pi/{\ensuremath{K}})\label{eq:dndcc}\\ \quad&\ensuremath{\Rightarrow} \quad\Ropt(K,\ensuremath{\lambda})= \sqrt{1+\frac{2}{\ensuremath{\lambda}}\sin(\pi/{\ensuremath{K}})}\simeq \sqrt{1+\frac{2\pi}{{\ensuremath{K}}}}\label{eq:optimalradiusexactrho}, \end{align} which is bounded between \begin{align} 1+\frac{\pi}{\ensuremath{\lambda}{\ensuremath{K}}} \leq \Ropt(K) \leq e^{\pi/2\ensuremath{\lambda}{\ensuremath{K}}} \label{eq:optimalradius}. \end{align} Therefore, we will in \secref{sec:rootstability} investigate the radius dependence in more detail. Finding the optimal radius for Huffman sequences yielding to the optimal Voronoi cells $\Dset_{\ensuremath{k}}$ is in fact a quantization problem, see for example \cite{KJ17}. Note, that the zeros for Huffman BMOCZ are not the centroids of the Voronoi cells, which suggest a much more complex metric for an optimal quantization, see \figref{fig:zeroperturbationsim}. From the simulation of the BER performance we observed $\ensuremath{\lambda}\simeq 2$, which might be also $L$ dependent. \newcommand{\ensuremath{\del_{\text{max}}^{(4)}=0.455}}{\ensuremath{\ensuremath{\delta}_{\text{max}}^{(4)}=0.455}} \newcommand{\ensuremath{\del_{\text{max}}^{(8)}=0.288}}{\ensuremath{\ensuremath{\delta}_{\text{max}}^{(8)}=0.288}} \newcommand{\ensuremath{\del_{\text{max}}^{(16)}=0.165}}{\ensuremath{\ensuremath{\delta}_{\text{max}}^{(16)}=0.165}} \newcommand{\ensuremath{d_{1}}}{\ensuremath{d_{1}}} \newcommand{\ensuremath{d_{2}}}{\ensuremath{d_{2}}} \newcommand{\ensuremath{d_{3}}}{\ensuremath{d_{3}}} \newcommand{\ensuremath{d_{4}}}{\ensuremath{d_{4}}} \newcommand{\ensuremath{d_{5}}}{\ensuremath{d_{5}}} \newcommand{\ensuremath{d_{6}}}{\ensuremath{d_{6}}} \newcommand{\ensuremath{a}}{\ensuremath{a}} \newcommand{\ensuremath{\phi}}{\ensuremath{\phi}} \newcommand{\ensuremath{\text{SNR}=1/\eps^2}}{\ensuremath{\text{SNR}=1/\ensuremath{\epsilon}^2}} \begin{figure}[t] \begin{subfigure}[b]{0.323\textwidth} \hspace{-.7cm} \includegraphics[width=1.1\textwidth]{% No5db_rootneighborhood_run1000tht500_N6_allinside_Runi_laminf.png} \caption{Radius $R=1$ with $\ensuremath{\lambda}=\infty$.} \label{fig:hufzerospread0} \end{subfigure} \begin{subfigure}[b]{0.323\textwidth} \hspace{-3ex} \includegraphics[width=1.1\textwidth]{% No5db_rootneighborhood_run1000tht500_N6_allinside_Runi_lam2.png} \caption{Radius $R=1.2247$ with $\ensuremath{\lambda}=2$.} \label{fig:hufzerospread1} \end{subfigure} \begin{subfigure}[b]{0.323\textwidth} \includegraphics[width=1.1\textwidth]{% No5db_rootneighborhood_run1000tht500_N6_allinside_Runi_lam1.png} \caption{Radius $R=1.4142$ with $\ensuremath{\lambda}=1$.} \label{fig:hufzerospread2} \end{subfigure} \caption{Simulation of $7$ different SNR values from $22$dB to $5$dB for a fixed Huffman sequence with $N=6$ over different radii.}\label{fig:zeroperturbationsim} \end{figure} \subsection{PAPR for Huffman Sequences with Uniform Radius} From \eqref{eq:hufflastfirst} we get for the magnitudes \begin{align} |x_{\ensuremath{K}}|^2,|x_0|^2 \in \Big[\frac{1}{{1+R^{2{\ensuremath{K}}}}},\frac{1}{{1+R^{-2{\ensuremath{K}}}}}\Big], \end{align} where the maximum is attained if ${\ensuremath{\mathbf b}}={\ensuremath{\mathbf 0}}$ or ${\ensuremath{\mathbf b}}={\ensuremath{\mathbf 1}}$, the all zero or all one bit vector. By noting that the first and last coefficient magnitude \eqref{eq:hufflastfirst} exploit a symmetry for $2\Norm{{\ensuremath{\mathbf b}}}_1$ and $K-2\Norm{{\ensuremath{\mathbf b}}}_1$, we only have to average for uniform bit distribution over $\Norm{{\ensuremath{\mathbf b}}}_1\in\{0,\dots,K/2\}$ (assuming $K$ even), which gets \Peter{was wrong}: \begin{align} \Expect{\Norm{{\ensuremath{\mathbf x}}}_{\infty}^2} &=\frac{1}{2^{K/2}}\frac{1}{R^{2K}+1}\sum_{n=1}^{2^{K/2}} R^{-2\sum_{k=1}^{K/2} b_k^{(n)}} =\frac{1}{2^{K/2}}\frac{1}{R^{2K}+1} \sum_{m=0}^{{K/2}} \binom{K/2}{m} R^{-2m}\\ &=\left(\frac{1+R^{-2}}{2}\right)^{\frac{K}{2}}\frac{1}{R^{2K}+1} \end{align} Since the Huffman sequences have all unit energy, the \emph{peak-to-average-power ratio} is for the optimal radius $R=\Ropt(K,1)$ in \eqref{eq:optimalradiusexactrho} for large $K$ \begin{align} \ensuremath{{\text{PAPR}}}& = ({\ensuremath{K}}+1)\frac{\Expect{\Norm{{\ensuremath{\mathbf x}}}_{\infty}^2}}{\Expect{\Norm{{\ensuremath{\mathbf x}}}_2^2}} = \frac{({\ensuremath{K}}+1)((1+R^{-1})/2)^{K/2}}{R^{2{\ensuremath{K}}} + 1} {\simeq} \frac{{\ensuremath{K}}+1}{(1+2\pi/{\ensuremath{K}})^{{\ensuremath{K}}}+1}\\ &{\leq} \frac{{\ensuremath{K}}+1}{2+2\pi}\simeq \frac{{\ensuremath{K}}+1}{8.28}\simeq {\ensuremath{K}}/9, \end{align} which is typically for a multi-carrier system, such as OFDM \cite{BF11}. \section{Numerical Simulations}\label{sec:simulations} \iflong Following the standard definition of SNR in \cite[(3.9)]{TV05} \begin{align} {\ensuremath{\text{rSNR}}}:=\frac{\text{average received signal energy per symbol time}}{\text{average noise energy per symbol time}}. \end{align} \fi We simulated with MatLab 2017a the \emph{bit-error-rate} (BER) over the rSNR \eqref{eq:rSNR} for ${\ensuremath{L}}$ Rayleigh fading multipaths with power delay profile exponent $\pd<1$. In the simulation, we scaled the transmit signals ${\ensuremath{\mathbf x}}$ by $\sqrt{N}$ and the channel by $\sqrt{1/\Expect{\Norm{\vh}_2^2}}$, such that the received average power will be normalized and equal to the transmitted average power, independent of $N, {\ensuremath{L}}$ and ${\ensuremath{\mathbf p}}$. Hence we obtain ${\ensuremath{\text{rSNR}}}=\ensuremath{{\text{SNR}}}=1/N_0$. The energy per bit is then \begin{align} E_b=\frac{N}{{\ensuremath{L}}}=\etaE^{-1}, \end{align} which is equal to the inverse of the \emph{bit rate} $\etaE$ per symbol time% \iflong as \begin{align} \etaE = \frac{\text{number of bits}}{\text{number of symbol times}} \end{align} \fi. Hence, the $\ensuremath{{\text{SNR}}}$ per bit is \begin{align} \frac{E_b}{N_0}= \frac{1}{\etaE\cdot N_0} = \frac{\ensuremath{{\text{SNR}}}}{\etaE} \end{align} see for example \cite[pp.97]{PS08}. \begin{figure} \hspace{-1.4cm} \begin{subfigure}{0.49\textwidth} \includegraphics[width=1.24\linewidth]{HufBMOCZ_RFMD_L8K8to16_known_pd_n0_snr_runpp956k_edited} \caption{Bit error results over $\ensuremath{{\text{SNR}}}$ for $\pd=0.88$} \label{fig:MLandRFMD_snr} \end{subfigure} \hspace{0.85cm} \begin{subfigure}{0.49\textwidth} \includegraphics[width=1.24\linewidth]{HufBMOCZ_RFMD_L8K8to16_known_pd13_n0_ebno_runpp956k_edited.eps} \caption{Bit error results over $E_b/N_0$ for $\pd=0.88$} \label{fig:MLandRFMD_ebno} \end{subfigure}\\ \begin{subfigure}{0.49\textwidth} \hspace{-1.3cm} \includegraphics[width=1.27\linewidth]{HufBMOCZ_RFMD_L8K4to16_known_pd1_n0_snr_Run18k2} \caption{Bit error results over $\ensuremath{{\text{SNR}}}$ for $\pd=1$.} \label{fig:MLandRFMD_snrpd1} \end{subfigure} \begin{subfigure}{0.49\textwidth} \includegraphics[width=1.2\linewidth]{HufBMOCZ_RFMD_L8K4to16_known_pd1_n0_ebno_Run18k.eps} \caption{Bit error results over $E_b/N_0$ for $\pd=1$.} \label{fig:MLandRFMD_ebnopd1} \end{subfigure} \caption{Huffman BMOCZ with RFMD, GW-DiZeT and ML (with known $N_0$ and ${\ensuremath{\mathbf p}}$) decoder for ${\ensuremath{K}}=8$ and channel length ${\ensuremath{L}}=4,8$ and $16$ under different power delay profiles $\pd$.} \label{fig:MLandRFMD} \end{figure} \if0 \begin{itemize} \item The weight is radius and length dependent. Following an argumentation of the root neighborhood size it seems to be wise to scale the modulus of the received polynomial by the modulus of the sample (zeros), i.e., % \begin{align} {\ensuremath{\tilde{\uY}}}(z)=\frac{|{\ensuremath{\mathrm Y}}(z)|}{\sqrt{\sum_{k=0}^{N-1} |z|^{2k}}} \end{align} % This compensate for additive noise on the coefficient with % \begin{align} \Norm{{\ensuremath{\mathbf y}}-\vn}_2\leq \ensuremath{\epsilon} \end{align} \item If full channel knowledge at the receiver, we can do a zero-forcing (ZF) in frequency \begin{align} \hat{\ensuremath{\mathbf x}}= {\ensuremath{\mathbf F}^*}\big( {\ensuremath{\mathbf F}} {\ensuremath{\mathbf y}} /{\ensuremath{\mathbf F}} \begin{pmatrix}\vh\\ {\ensuremath{\mathbf 0}}_{N-{\ensuremath{L}}} \end{pmatrix} \big) \end{align} \fi As an ultimate benchmark in all simulations, we will compare to the coherent case, where the frequency selective channel is modulated by OFDM with a binary phase shift keying (BPSK). Transforming the linear convolution for i.i.d. Gaussian CIR in time domain to the frequency domain, yields to $N$ parallel flat fading channels. Assuming a sequential block transmission, the cyclic prefix, allows to communicate $N$ bits per channel use and results therefore in coherent BPSK flat fading. The BER for BPSK over a flat fading channel $h_0=|h_0| e^{i\phi}$, with known phase $\phi$ and $\Expect{|h_0|^2}=1$ is equivalent to the bit error probability (one bit per symbol duration) given by \begin{align} P_e = \frac{1}{2} \left( 1- \sqrt{\frac{E_b \Expect{|h_0|^2}/N_0}{1+E_b\Expect{|h_0|^2}/N_0}}\right) = \frac{1}{2} \left(1- \sqrt{\frac{{\ensuremath{\text{rSNR}}}}{1+{\ensuremath{\text{rSNR}}}}}\right) \end{align} since $\ensuremath{\sigma}_h^2=1$ we have in \figref{fig:pilotQPSKvsBMOCZ} $E_b/N_0= E_b\Expect{|h_0|^2}/N_0={\ensuremath{\text{rSNR}}}$, pictured as a thick doted red curve. The BPSK coherent flat fading can be seen as the best binary signaling scheme performance if no multi-path diversity is exploit (no outer codes). Note, that our scheme prevent is therefore robust against \emph{inter-symbol-interference} (ISI), given by superposition of overlapping symbols due to the multipath delays. It is still unclear how to exploit fully the multipath diversity gain in one-shot at the receiver without knowledge of the CIR. However, the DiZeT decoder performs very close to the ML decoder and coherent uncoded OFDM with BPSK, see \figref{fig:MLandRFMD}. Note, all simulated BER curves are for uncoded bits. In \figref{fig:BMOCZ_sdp} the BER for the SDP denoising \eqref{eq:sdpdenoising} with the estimate channel autocorrelation via \eqref{eq:zfah} are simulated for $K=8$ and $L=4$ with flat power delay profile. The denoised signal ${\ensuremath{\hat{\mathbf x}}}$ from the SDP is then either decoded by the GW-DiZeT decoder or the RFMD decoder. The results show a $2$dB lose compared to GW-DiZeT without denoising. The reason for the performance lose is first in the bad estimation of the channel autocorrelation and secondly due to the simultaneously denoising of the channel and signal. Since the SDP does not emphasize the signal reconstruction quality, the quality in signal recovery is in sum worse as for the direct decoding approaches. However, the knowledge of the channel might help for other purposes. \begin{figure}[t] \hspace{-0.8cm} \begin{subfigure}{0.49\textwidth} \vspace{-2cm} \includegraphics[width=1.15\textwidth]{HufBMOCZ_SDP_L8K4_run1850LK} \label{fig:MLandRFMD_snr2} \end{subfigure} \hspace{0.6cm} \begin{subfigure}{0.485\textwidth} \vspace{-2cm} \includegraphics[width=1.12\textwidth]{HufBMOCZ_SDP_true_ah_pd13_runpp2490_edited2} \label{fig:BMOCZ_sdp_pd13} \end{subfigure} \vspace{-2.9cm} \caption{GW-DiZeT decoder versus SDP denoising with GW-DiZeT and RFMD decoding.} \label{fig:BMOCZ_sdp} \end{figure} \if0 \newpage \section{Stability of DiZeT decoder} \subsection{Root Neighborhoods} \begin{figure} \centering \begin{subfigure}{0.49\textwidth} \hspace{-1.5cm} \includegraphics[width=1.1\linewidth]{Scaling_BMOZandQMOCZ_BPSK_ZF_ML_AML_L9K1_alph1_AverageReceiveSNR_run90000.eps} \caption{Flat fading, ${\ensuremath{L}}=1$. Instantaneous rSNR.} \label{fig:comparisonbpsk1} \end{subfigure} \hspace{-0.4cm} \begin{subfigure}{0.49\textwidth} \includegraphics[width=1.15\linewidth]{Scaling_BMOZandQMOCZ_L9-129K3_alph1_run28000.eps} \caption{Frequency Selective Channel, ${\ensuremath{L}}=3$.} \label{fig:comparisonbpsk3} \end{subfigure} \caption{Comparison to BPSK over Ideal AWGN and Zero-Forcing/Matched Filter with full channel knowledge} \end{figure} \begin{figure} \centering \begin{subfigure}{0.49\textwidth} \hspace{-1.5cm} \includegraphics[width=1.18\linewidth]{Scaling_BMOZandQMOCZ_BPSK_ZF_L33-129K1_alph1_AverageReceiveSNR_run15000.eps} \caption{Flat fading, ${\ensuremath{L}}=1$, average receive SNR.} \label{fig:comparisonbpsk1_av_rSNR} \end{subfigure} \hspace{-0.4cm}\vspace{-0.2cm} \begin{subfigure}{0.49\textwidth} \includegraphics[width=1.15\linewidth]{Scaling_BMOZandQMOCZ_BPSK_ZF_L33-129K9_alph1_AverageReceiveSNR_run91000.eps} \caption{Frequency Selective Channel, ${\ensuremath{L}}=9$, average receive SNR.} \label{fig:comparisonbpsk9_av_rSNR} \end{subfigure} \caption{Comparison to BPS{\ensuremath{L}} over Ideal AWGN and Zero-Forcing/Matched Filter with full channel knowledge. Rayleigh fading with $\Expect{|h_k|^2}=1$ and $\Expect{\Norm{\vh}_2^2}={\ensuremath{L}}$} \end{figure} Correction with right Geometric-Weight factor in \eqref{eq:gwfactor}. \paragraph{First Order Analysis} We can also think of perturbing additive the polynomial, i.e. \begin{align} {\ensuremath{\mathrm y}}(z)={\ensuremath{\mathrm x}}(z)+t\cdot \uw(z) \end{align} for arbitrary $t>0$. A second order analysis gives \subsection{Metric of the Root neighborhoods} To obtain a good decision for the BMOCZ decoder if we observe the disturbed zeros, we need a proper metric. Since by Ostrowskis Theorem, the zeros of the polynomial are continuos functions of the coefficients, we may assume that the root neighborhoods or pseudozerodomains are continuous growing with $\ensuremath{\epsilon}.$ The definition of the root neighborhoods give an explicit geometric set for the neighborhood $\ensuremath{\mathscr Z}_{\ensuremath{\epsilon}}(P,\ensuremath{\alpha})$ of one single zero $\ensuremath{\alpha}$. We know, that any disturbed zero \begin{align} \zeta=\ensuremath{\alpha}+\ensuremath{\tilde{\alp}} \end{align} of ${\ensuremath{\mathrm Y}}(z)$ must satisfy \begin{align} |{\ensuremath{\mathrm X}}(\zeta)|^2 \leq \ensuremath{\epsilon} \sum_k |\zeta|^{2k}\label{eq:rndef} \end{align} for any additive noise with $\Norm{\vn}_2^2\leq \ensuremath{\epsilon}$. We want to know, what is the smallest $\ensuremath{\epsilon}$-neigbhorhood enclosing $\ensuremath{\mathscr Z}_{\ensuremath{\epsilon}}(P,\ensuremath{\alpha})$. For this we need to find an upper bound \begin{align} \rho:=|\ensuremath{\tilde{\alp}}|\leq f(\ensuremath{\epsilon},{\ensuremath{\mathrm X}},\ensuremath{\alpha}) \end{align} Let us start with the following observation for any $a>0, \zeta\not=a$ \begin{align} \left|\prod_{l=1}^N (\zeta- a\ensuremath{\omega}_l) \right|^2 =|\zeta^N-a^{N}|^2 = a^{2N}|(a^{-1}\zeta)^N -1|^2 \end{align} where we assumed that the zeros $\ensuremath{\omega}_l$ are the $N$roots of unity which are the roots of the polynomial $p(z)=z^N-1$. Note, we have for $R>1$ \begin{align} |1-1/\cc{\ensuremath{\alpha}_l}|\leq| 1- \ensuremath{\omega}_l|\leq |1-\ensuremath{\alpha}_l| \end{align} Hence, we get for the LHS in \eqref{eq:rndef} \begin{align} R^{-2N} |(R\zeta)^N-1|^2\leq |{\ensuremath{\mathrm X}}(\zeta)|^2 \leq R^{2N} |(R^{-1}\zeta)^N -1|^2 \end{align} By using the geometric series we get finally the weaker inequality \begin{align} R^{-2N} |(R\zeta)^{N}-1|^2\leq \ensuremath{\epsilon} \frac{1- |\zeta|^{2N}}{1-|\zeta|^2} \quad \ensuremath{\Leftarrow} \quad |{\ensuremath{\mathrm X}}(\zeta)|^2 \leq \ensuremath{\epsilon} \sum_k |\zeta|^{2k} \end{align} Since the zero set is symmetric, we can assume $\ensuremath{\alpha}=R^{-1}e^{i0}=R^{-1}$. Moreover, we assume $\ensuremath{\tilde{\alp}}=\rho e^{i\ensuremath{\theta}}$ is so small that $|\zeta|=|\ensuremath{\alpha}(1+\ensuremath{\tilde{\alp}}/\ensuremath{\alpha})|=R^{-1}|1+R\ensuremath{\tilde{\alp}}|<1$ too. Then we can use the Bernoulli inequality and leverage the RHS \begin{align} \frac{1- |\zeta|^{2N}}{1-|\zeta|^2} \leq \frac{1- |\ensuremath{\alpha}|^{2N} |1+{NR\ensuremath{\tilde{\alp}}}|^2}{1-|\zeta|^2} =\frac{1- R^{-2N} |1+ {NR\ensuremath{\tilde{\alp}}}|^2}{1-|\zeta|^2} \leq \frac{1- R^{-2N} (1+ N^2R^2|\ensuremath{\tilde{\alp}}|^2 -NR|\ensuremath{\tilde{\alp}}|)}{1-|\zeta|^2} \end{align} Inserting everything gives \begin{align} R^{-2N} | [ R R^{-1}(1+ R\ensuremath{\tilde{\alp}})]^N -1|^2 \geq R^{-2N}| 1+NR\ensuremath{\tilde{\alp}} -1|^2= NR^{-2N-2} |\ensuremath{\tilde{\alp}}|^2 \end{align} Hence \begin{align} |\ensuremath{\tilde{\alp}}|^2(1-R^{-2}|1+R\ensuremath{\tilde{\alp}}|^2) & \leq \frac{R^{2N+2}}{N} - \frac{R^2}{N} (1-NR|\ensuremath{\tilde{\alp}}|) \leq \frac{R^{2N+2}}{N} - \frac{R^2}{N} +R^3|\ensuremath{\tilde{\alp}}|\\ |\ensuremath{\tilde{\alp}}|^2(1-R^{-2})\leq |\ensuremath{\tilde{\alp}}|^2(1-R^{-2}|1+R\ensuremath{\tilde{\alp}}|^2) &\leq \ensuremath{\epsilon}(\frac{R^{2N+2}}{N} - \frac{R^2}{N}+R^3/N) =\ensuremath{\epsilon}\frac{R^2}{N}(R^{2N}-1+R) \\ |\ensuremath{\tilde{\alp}}|^2 &\leq \ensuremath{\epsilon} \frac{1}{N}\frac{R^{2N}-1}{R^2-1} \end{align} if $\ensuremath{\epsilon}>0$ such that $|\ensuremath{\alpha}|\leq 1/N$. \begin{remark} % It is also possible by second order analysis to derive % \begin{align} |\ensuremath{\tilde{\alp}}| \leq \ensuremath{\epsilon} \frac{\sum_k |\zeta|^k}{|{\ensuremath{\mathrm X}}^{\prime}(z)|} + O(\ensuremath{\epsilon}^2) = \ensuremath{\epsilon} \frac{1-|\zeta|^N}{|{\ensuremath{\mathrm X}}^{\prime}(z)|(1-|\zeta|)} + O(\ensuremath{\epsilon}^2) \end{align} % Unfortunately, we can not use the zero structure of ${\ensuremath{\mathrm X}}$ for the derivation. Although, there is a inverse Bernstein-type inequality for the case when all zeros of ${\ensuremath{\mathrm X}}$ have same magnitude. \cite[Thm.1.1]{EHS13}. \begin{align} |\ensuremath{\tilde{\alp}}|\leq \ensuremath{\epsilon} \frac{1-|\zeta|^N}{N(1-|\zeta|)} + O(\ensuremath{\epsilon}^2) \end{align} \end{remark} \fi \if0 \begin{figure} \centering \hspace{-1.5cm} \begin{subfigure}{0.49\textwidth} \includegraphics[width=1.02\linewidth]{Scaling_BMOZandQMOCZ_BPSK_ZF_ML_AML_L9K1_alph14_AverageReceiveSNR_run5000.eps} \caption{Frequency Non-Selective Channel, ${\ensuremath{L}}=1$ single-path.} \label{fig:comparisonbpskL9K1} \end{subfigure} \hspace{-0.4cm} \begin{subfigure}{0.49\textwidth} \includegraphics[width=1.18\linewidth]{Scaling_BMOZandQMOCZ_BPSK_ZF_ML_AML_L9K9_alph14_AverageReceiveSNR_run10000.eps} \caption{Multipath Rayleigh fading with exponential power delay profile, ${\ensuremath{L}}={\ensuremath{L}}=9$..} \label{fig:comparisonbpsk_K9} \end{subfigure} \caption{Comparison to BPSK over Ideal AWGN and Zero-Forcing/Matched Filter with full channel knowledge} \end{figure} \begin{figure} \begin{subfigure}{\textwidth} \centering \includegraphics[width=0.8\linewidth]{Scaling_BMOZandQMOCZ_L3-33K1_alph1_run2000_fixR.eps} \caption{Fix radius $R=1.5$. Identity channel $h=1$.} \label{fig:fixradius} \end{subfigure}\\% \begin{subfigure}{\textwidth} \centering \includegraphics[width=0.8\linewidth]{Scaling_BMOZandQMOCZ_L3-33K1_alph1_run3000.eps} \caption{Radius dependent on dimension $N$} \label{fig:variableradius} \end{subfigure} \caption{Radius Dependence of Dimension}\label{fig:radiusdependence} \end{figure} \fi \section{Sharper Robustness Analysis for BMOCZ Codebooks by Exploiting the Geometric Zero Structure}\label{sec:sharperbounds} We will in this section investigate the geometric structure of the zeros to improve the robustness of normalized polynomials against additive noise on its coefficients, sometimes also referred to the \emph{conditioning of a polynomial}. This is not to mistaken with the notion of stable polynomials or \emph{Hurwitz stability}, which refers to the property that all zeros are located in the positive half-plane, see for example \cite[Cha.21]{Fis08}. \iflong The distortion of zeros against noise on the polynomial coefficients were first investigated by Wilkinson in a series of works and expanded to a book \cite{Wil63}. The Wilkinson polynomial \eqref{eq:wilkinson} shows that only a minimal pairwise distance of the zeros does not yield to stability against noise on the coefficients. In fact, Wilkinson could show that polynomials having zeros with large magnitude will be unstable. However, if all zeros are inside the unit circle, as for the polynomial \begin{align} {\ensuremath{\mathrm X}}(z)= \prod_{k=1}^{20} (z-2^{-k}) \end{align} then, Wilkinson could show that its zeros are stable against additive noise on the coefficients, although the minimal distance goes down to $2^{-20}$. This also suggest, that an Euclidean distance of the zeros might not be a good measure for all polynomials. \fi As we saw in the analysis of \thmref{thm:zerodistortion}, a large pairwise distance as well as a large leading coefficient guarantee a robustness against additive noise. For polynomials generated by autocorrelations, we will have zeros in conjugate-reciprocal pairs and if we upper bound the largest zero, we force the zeros in a ring or annulus around the unit circle, which will exclude the extreme cases in the zero displacement, see \figref{fig:hexagonal}. It turns out that the Euclidean metric, as used in the RFMD decoder might be reasonable for zeros near the unit circle. Indeed, for Huffman Polynomials with uniform radius, the root neighborhoods can be bounded by disjoint uniform discs, see \figref{fig:zeroperturbationsim}. However, the first zero on the real line, seem to disturb non uniform. This might be due to discontinuity of the real valued zero on the positive half plane (winding number). A more careful analysis of the exact root neighborhood grow behaviour will be investigated in a follow up paper. Since we want to keep the root neighborhoods disjoint, the root neighborhoods should not exceed a radius $\ensuremath{\delta}$ which is larger then half the minimal pairwise distance. This in fact, leads to a circle packing problem in the plane, which is know to be most dense if the circle centroids are placed on a hexagonal lattice, see \figref{fig:hexagonalring}. \iflong If we consider autocorrelation codebooks, there exists always a zero vector $\ensuremath{\boldsymbol{ \alp}}\in\ensuremath{\mathscr Z}$ with all zeros inside the unit circle. Hence, its minimal pairwise distance can not be larger than one for $K\geq 6$ (hexagonal lattice). We will hence assume $K\geq 6$ and ${\ensuremath{d_{\text{\rm min}} } }\leq 1$. \fi \newcommand{\ensuremath{R^{-1}}} \newcommand{\Rnorm}{\ensuremath{R}}{\ensuremath{R^{-1}}} \newcommand{\Rnorm}{\ensuremath{R}} \begin{figure}[t] \begin{subfigure}[b]{0.485\textwidth} \centering \def0.65\textwidth} \footnotesize{ \import{\pdfdir}{OptimaRadiusHuffmanBMOCZ.pdf_tex} {0.9\textwidth} \footnotesize{ \import{\pdfdir}{HexagonalCirclePack_rotated.pdf_tex} } \caption{Hexagonal packing in an annulus, worst constellation red, best constellation blue.} \label{fig:hexagonalring} \end{subfigure} \begin{subfigure}[b]{0.485\textwidth} \hspace{-1.2cm} \def0.65\textwidth} \footnotesize{ \import{\pdfdir}{OptimaRadiusHuffmanBMOCZ.pdf_tex} {1.53\textwidth} \footnotesize{ \import{\pdfdir}{hexagonallattice.pdf_tex} } \caption{Densest packing inside the ring, worst case.\\[5ex]} \label{fig:hexagonalinring} \end{subfigure} \caption{Homogeneous Circle Packing in an annulus.} \label{fig:hexagonal} \end{figure} \iflong The following theorem holds if \conref{con:ngon} holds, see \appref{app:ngon}. \begin{thm}\label{thm:zerodistortion2} Let ${\ensuremath{\mathrm x}}(z)\in{\ensuremath{\mathbb C}}[z]$ be a polynomial of order $\xord\geq 2$ with simple zeros $\ensuremath{\alpha}_1,\dots,\ensuremath{\alpha}_{\xord} \subset {\ensuremath{\mathbb C}}$ having minimal pairwise distance ${\ensuremath{d_{\text{\rm min}} } }$ and concentrated in an annulus of radius $R$ and $R^{-1}$ % \begin{align} {\ensuremath{d_{\text{\rm min}} } } :=\min_{n\not=k} |\ensuremath{\alpha}_n-\ensuremath{\alpha}_k|>0 \quad,\quad \forall n\colon R\geq |\ensuremath{\alpha}_n| \geq R^{-1}. \end{align} % Let ${\ensuremath{\mathbf w}}\in{\ensuremath{\mathbb C}}^{\xord+1}$ with $\Norm{{\ensuremath{\mathbf w}}}_{2}\leq \ensuremath{\epsilon}$ be an additive perturbation on the polynomial coefficients ${\ensuremath{\mathbf x}}$ and $\ensuremath{\delta}\in(0,{\ensuremath{d_{\text{\rm min}} } }/2)$. Let us set $\ensuremath{n_{\text{max}}}=\lceil\sqrt{(4N-1)/12} -1/2\rceil$ and $\ensuremath{\tilde{d}_{\text{min}}}=\min\{1,{\ensuremath{d_{\text{\rm min}} } }\cdot (1- \frac{\ensuremath{\delta}^6}{{\ensuremath{d_{\text{\rm min}} } }^6})^{1/6}\}$. Then the $n$th zero $\zeta_n$ of the disturbed polynomial ${\ensuremath{\mathrm y}}(z)={\ensuremath{\mathrm x}}(z)+\uw(z)$ lies in $B_n(\ensuremath{\delta})$ if % \begin{align} \ensuremath{\epsilon}=\ensuremath{\epsilon}({\ensuremath{\mathbf x}},\ensuremath{\delta})\leq (2R)^{-N} \ensuremath{\delta} \sqrt{\frac{(R+\ensuremath{\delta})^2-1}{(R+\ensuremath{\delta})^{2N}-1}}\cdot \ensuremath{\tilde{d}_{\text{min}}}^{\ 3(\ensuremath{n_{\text{max}}}^2-\ensuremath{n_{\text{max}}} +2)}\cdot ((\ensuremath{n_{\text{max}}}\!-\!1)!H(\ensuremath{n_{\text{max}}}\!-\!1))^{6} \label{eq:noisebound2}. \end{align} \end{thm} \fi % The idea is to place $N$ zeros in a Ring $\ensuremath{{\mathscr{R}}}=\ensuremath{{\mathscr{R}}}(R)$ of area $|\ensuremath{{\mathscr{R}}}|$ with minimal pairwise distance ${\ensuremath{d_{\text{\rm min}} } }$. The bound in \eqref{eq:brutallowerbound} is actually a lower bound of the geometric mean of all zeros distances. Hence, the robustness bound depends not only on the minimal pairwise distance but also on their geometric structure. In fact, the densest packing of the $N$ zeros will yield to the smallest geometric mean of the distances and therefore result in the worst stability bound, see \figref{fig:hexagonal}. Furthermore, the maximal amount of zeros in the ring is bounded by the spherical circle packing problem. The exact amount is unsolved for arbitrary ${\ensuremath{d_{\text{\rm min}} } }$ even if the set is the unit disk. The problem is usually known as the density packing problem, where the density of placing $N$ equal circles of radius ${\ensuremath{d_{\text{\rm min}} } }$ in the ring $\ensuremath{{\mathscr{R}}}$ is given by % \begin{align} D= N{\ensuremath{d_{\text{\rm min}} } }^2 \pi 4|\ensuremath{{\mathscr{R}}}| \end{align} % One bound on the maximal number $N$ of circles is given by Fejer Toth as % \begin{align} N\leq \frac{\pi (R^2-R^{-2})}{{\ensuremath{d_{\text{\rm min}} } }^2\sqrt{12}}\label{eq:Nboundring} \end{align} % see for example \cite{GT00}. \iflong \begin{remark} It can be seen that for $\ensuremath{\delta}={\ensuremath{d_{\text{\rm min}} } }/2$ the bound is the largest for $\ensuremath{n_{\text{max}}}=1$, since \begin{align} \ensuremath{\tilde{\dmin}} &= {\ensuremath{d_{\text{\rm min}} } }\left(1-(\frac{\ensuremath{\delta}}{{\ensuremath{d_{\text{\rm min}} } }})^6\right)^{1/6}=({\ensuremath{d_{\text{\rm min}} } }^6-\ensuremath{\delta}^6)^{1/6}\\ \ensuremath{\Rightarrow} \ensuremath{\delta}^2\ensuremath{\tilde{\dmin}}^{12} & = \ensuremath{\delta}^2({\ensuremath{d_{\text{\rm min}} } }^6 -\ensuremath{\delta}^6)^2 \label{eq:dminone} \end{align} % which is monotone increasing in $\ensuremath{\delta}$ and obtains its maximum in the interval $[0,{\ensuremath{d_{\text{\rm min}} } }/2]$ at $\ensuremath{\delta}={\ensuremath{d_{\text{\rm min}} } }/2$. Moreover, it is then obvious that the bound increases with ${\ensuremath{d_{\text{\rm min}} } }$. Hence, for fixed $R$ and $N$ the minimal pairwise distance ${\ensuremath{d_{\text{\rm min}} } }$ and the disc neighborhood radius $\ensuremath{\delta}$ should be maximized to obtain the largest upper bound for the noise power $\ensuremath{\epsilon}^2$, see \figref{fig:delmin_one}. \end{remark} \begin{figure} \begin{subfigure}{0.485\textwidth} \centering \includegraphics[width=0.7\textwidth]{del_dmin1.pdf} \caption{Increase of \eqref{eq:dminone} in $\ensuremath{\delta}$ for ${\ensuremath{d_{\text{\rm min}} } }=1$.} \label{fig:delmin_one} \end{subfigure} \begin{subfigure}{0.485\textwidth} \centering \includegraphics[width=0.7\textwidth]{epsbound_R11_dmin5_n7_log10.pdf} \caption{Logarithmic bound for noise power \eqref{eq:noisebound2} over $\ensuremath{\delta}$ with $R=1.1,N=7$ and ${\ensuremath{d_{\text{\rm min}} } }=0.5$.} \label{fig:espbound_dmin7} \end{subfigure} \caption{Bounds over root radius $\ensuremath{\delta}$} \end{figure} \fi \subsection{Revised Proof of \thmref{thm:zerodistortion}} The main idea of the proof relies on the use of Rouches \thmref{thm:rouche}, by controlling for each $m=1,2,\dots,N$ the modulus of the noise polynomial on the single root neighborhood circle bounds % \begin{align} |{\rm{W}}(z)|\leq |{\ensuremath{\mathrm X}}(z)| \quad,\quad z\in C_m(\ensuremath{\delta}). \end{align} % By choosing the radius $\ensuremath{\delta}$ small enough, such that no overlap of the single root neighborhoods occur, a separation by the RFMD decoder will be always successful (no channel zeros), guaranteeing an error free decoding. \if0 \begin{align} |w_0|\leq\ensuremath{\epsilon} \leq |x_0| =|x_N|\prod_n |\ensuremath{\alpha}_n| \end{align} since this needs to hold for all $\Norm{{\ensuremath{\mathbf w}}}\leq \ensuremath{\epsilon}$. \fi % Since we want to hold this for every noise polynomial generated by $\Norm{{\ensuremath{\mathbf w}}}_2\leq \ensuremath{\epsilon}$ we have to satisfy % \begin{align} \max_{\Norm{{\ensuremath{\mathbf w}}}\leq \ensuremath{\epsilon}} \max_{z\in C_m(\ensuremath{\delta})} \frac{|{\rm{W}}(z)|^2}{|{\ensuremath{\mathrm X}}(z)|^2} \leq 1.\label{eq:wmax} \end{align} % Note, that ${\ensuremath{\mathrm X}}(z)$ has no zeros on $\bigcup C_m(\ensuremath{\delta})$, hence we can divide. Let us define $\underline{{\ensuremath{\mathbf z}}}=(z^0,z^1,\dots,z^{\xord})^T$, where we set $0^0=1$. We will upper bound the magnitude of $\uw$ by using Cauchy-Schwarz % \begin{align} |\uw(z)|^2 = |{\ensuremath{\mathbf w}}^T\underline{{\ensuremath{\mathbf z}}}|^2\leq \Norm{{\ensuremath{\mathbf w}}}_{2}^2\cdot \Norm{\underline{{\ensuremath{\mathbf z}}}}_2^2 = \ensuremath{\epsilon}^2\cdot \sum_{n=0}^N |z|^{2n}.\label{eq:csbound} \end{align} % Since the noise ${\ensuremath{\mathbf w}}$ is in the ball with radius $\ensuremath{\epsilon}$, all directions can be chosen and we achieve always equality in \eqref{eq:csbound}. Hence, \eqref{eq:wmax} is equivalent to % \begin{align} \ensuremath{\epsilon}^2 \leq \frac{1}{\max_{z\in C_m(\ensuremath{\delta})} \frac{\sum_n |z|^{2n}}{|{\ensuremath{\mathrm X}}(z)|^2}}=\min_{z\in C_m(\ensuremath{\delta})} \frac{|{\ensuremath{\mathrm X}}(z)|^2}{\sum_n |z|^{2n}} = \min_{z} f_m(z) \label{eq:worstcaseepsilon} \end{align} % By using $z=\ensuremath{\alpha}_m+ \ensuremath{\delta} e^{i\ensuremath{\theta}}$ for some $\ensuremath{\theta}\in[0,2\pi)$ we need to find a tight lower bound of % \begin{align} f_m(\ensuremath{\theta}):=\frac{|{\ensuremath{\mathrm X}}(\ensuremath{\alpha}_m + \ensuremath{\delta} e^{i\ensuremath{\theta}})|^2}{\sum_n |\ensuremath{\alpha}_m+\ensuremath{\delta} e^{i\ensuremath{\theta}}|^{2n}} =|x_N|^2 \frac{\prod_{n} |\ensuremath{\alpha}_m + \ensuremath{\delta} e^{i\ensuremath{\theta}} -\ensuremath{\alpha}_n|^2}{\sum_n |\ensuremath{\alpha}_m +\ensuremath{\delta} e^{i\ensuremath{\theta}}|^{2n}}. \end{align} % Since we are searching for a uniform radius $\ensuremath{\delta}$ which keeps all root neighborhoods disjoint, we search for the worst $\ensuremath{\alpha}_m$. The only information of the zeros we have is the minimal pairwise distance ${\ensuremath{d_{\text{\rm min}} } }$ and the smallest and largest moduli $R^{-1}$ resp. $R$, we define therefore % \begin{align} {\ensuremath{\mathcal{A}}}:=\set{\ensuremath{\alpha}=\{\ensuremath{\alpha}_1,\dots,\ensuremath{\alpha}_N\}}{ \forall n: R\geq |\ensuremath{\alpha}_n| \geq R^{-1}, {\ensuremath{d_{\text{\rm min}} } }\leq\min_{m\not=n}| \ensuremath{\alpha}_m-\ensuremath{\alpha}_n|}. \end{align} % Note, ${\ensuremath{\mathcal{A}}}$ is a compact set. The leading coefficient $x_N$ depends on all zeros and \emph{the height} of the polynomial, given by % \begin{align} \Norm{{\ensuremath{\mathrm X}}}_2:=\Norm{{\ensuremath{\mathbf x}}}_2=\sqrt{\sum_{n=0}^{N} |x_n|^2}. \end{align} % We chose here the Euclidean norm as the height, since we are interested in SNR performance. Hence, we define the set of all allowable normalized polynomials with zeros in ${\ensuremath{\mathcal{A}}}$ as % \begin{align} \ensuremath{\mathscr P}=\ensuremath{\mathscr P}({\ensuremath{\mathcal{A}}}) :=\set{{\ensuremath{\mathrm X}}(z)=\sum_{n=0}^N x_n z^n=x_N \prod_n (z-\ensuremath{\alpha}_m)}{ \Norm{{\ensuremath{\mathbf x}}}_2=1,\ensuremath{\alpha}\in{\ensuremath{\mathcal{A}}}}\notag. \end{align} % This brings us to the following optimization problem % \begin{align} f(\ensuremath{\mathscr P})=\min_{{\ensuremath{\mathrm X}}\in\ensuremath{\mathscr P}} |x_N|^2\min_m \min_{\ensuremath{\theta}} \frac{\prod_{n} |\ensuremath{\alpha}_m + \ensuremath{\delta} e^{i\ensuremath{\theta}} -\ensuremath{\alpha}_n|^2}{\sum_n |\ensuremath{\alpha}_m +\ensuremath{\delta} e^{i\ensuremath{\theta}}|^{2n}}.\label{eq:optimizationPsettht} \end{align} % The modulus of the leading coefficient can be lower bounded for normalized polynomials by \begin{align} 2^{-N}\leq 2^{-N}\Norm{{\ensuremath{\mathbf x}}}_1 \leq |x_N|\prod_{n=1}^N\max\{1,|\ensuremath{\alpha}_n|\}, \end{align} see for example \cite{Mah60} or \cite[Prop.86]{Zip93}. Note, this bound is not very tight for simple zeros with large minimal pairwise distance. If all zeros are inside the unit circle, this results in the largest lower bound and if all zeros are on the outside radius this results in the lowest bound, i.e., \begin{align} 2^{-N} \leq |x_N| \quad,\quad {(2R)}^{-N}\leq |x_N|.\label{eq:xnlowerbounds} \end{align} Using the worst case bound allows to eliminate the height constraint \begin{align} f(\ensuremath{\mathscr P})\geq (2R)^{-N} \min_{\ensuremath{\alpha}\in{\ensuremath{\mathcal{A}}}}\min_m \min_{\ensuremath{\theta}} \frac{\prod_{n} |\ensuremath{\alpha}_m + \ensuremath{\delta} e^{i\ensuremath{\theta}}-\ensuremath{\alpha}_n|}{\sqrt{\sum_n |\ensuremath{\alpha}_m +\ensuremath{\delta} e^{i\ensuremath{\theta}}|^{2n}}} \label{eq:optimizationalpm} \end{align} which is necessary to leverage the problem to a pure geometric problem. Let us assume $\ensuremath{\alpha}_{\hat{m}}=\rho e^{i\phi}$ is the zero selection for which there exists a $\ensuremath{\theta}$ which obtains the minimum. Then, we can rotate all zeros by $e^{-i \phi}$, since it will not change their modulus nor their pairwise distances and hence be lying in ${\ensuremath{\mathcal{A}}}$ (rotation invariant). Since the numerating of $\ensuremath{\alpha}_n$ is arbitary, we can just chose $\ensuremath{\alpha}_{\hat{m}}=\ensuremath{\alpha}_1=\rho$. Hence, we can omit the minimization over $m$ since we minimize over all $N-1$ zeros and $\ensuremath{\alpha}_1\in[R^{-1},R]$. This brings us to the non-convex geometric problem \begin{align} \min_{\ensuremath{\alpha}\in{\ensuremath{\mathcal{A}}}} f(\ensuremath{\alpha}) =\min_{\ensuremath{\alpha}, \ensuremath{\alpha}_1\in[R^{-1},R]} \min_{\ensuremath{\theta}} \frac{\prod_{n} |\ensuremath{\alpha}_1+\ensuremath{\delta} e^{i\ensuremath{\theta}} -\ensuremath{\alpha}_n|}{\sqrt{\sum_n |\ensuremath{\alpha}_1+\ensuremath{\delta} e^{i\ensuremath{\theta}}|^{2n}}}. \label{eq:optimizationPset} \end{align} The nominator is independent of the other $N-1$ zeros, and obtains its maximum for $|R+\ensuremath{\delta}|$. It can be seen that the numerator, will not yield the global minimal constellation if we place the $N-1$ zeros around $\ensuremath{\alpha}_1=R$, for $N\geq 2$, due to the restriction of the ring. However, it is geometrical not obvious which $\ensuremath{\alpha}_1$ will yield the global minimum of $f$. Therefore, we lower bound the nominator in $f(\ensuremath{\alpha})$, independently of the numerator, by the geometric formula for the worst case \begin{align} f(\ensuremath{\alpha})\geq \sqrt{\frac{(R+\ensuremath{\delta})^{2}-1}{(R+\ensuremath{\delta})^{2N}-1}} g(\ensuremath{\alpha})\quad\text{ with }\quad g(\ensuremath{\alpha})=\min_{\ensuremath{\theta}} g(\ensuremath{\alpha},\ensuremath{\theta})=\prod_{n=1}^N |\ensuremath{\alpha}_1 +\ensuremath{\delta} e^{\im \ensuremath{\theta}}-\ensuremath{\alpha}_n|.\label{eq:lbound} \end{align} Now, the minimization over the zeros reduces to the densest packing of $\ensuremath{\alpha}_2, \dots, \ensuremath{\alpha}_N$ around $\ensuremath{\alpha}_1$, since the \emph{geometric mean} $g(\ensuremath{\alpha})^{1/n}$ of the distances decreases if each distance decreases. If $N= 7$, the densest packing is the hexagon inscribed in a radius ${\ensuremath{d_{\text{\rm min}} } }$ with one zero at its centroid, see \figref{fig:hexagonalinring}. For arbitrary $N$ this is the well known circle packing problem, also known as the ``penny packing`` problem. However, it is not obvious if the optimal $z$ lies on the circle around the centroid or on the circle around the vertices. If $\ensuremath{\alpha}_1$ is the centroid, then we can show by \thmref{thm:mgonproddistance} that extremal $z$'s will lie at crossing of the circle with the line between origin and one vertex. Since we want to maximize the nominator, the $z$ which lies on the real axis, right from the centroid, will therefore obtain the minimal value of $f$. If $\ensuremath{\alpha}_1$ is one of the vertices, then we need to show that $z$ does not achieve a smaller product distance. Unfortunately, we can not prove this analytically and formulated this as \conref{con:ngon} in \appref{app:ngon}. If the conjecture holds, then the idea is to consider nested polygons (honeycombs) as the worst case zero configuration, to derive a lower bound for \eqref{eq:lbound}. Each $n$th honeycomb consist then of $6n$ points, placed on $n$ hexagons rotated accordingly, see \figref{fig:hexagonalinring}. By \thmref{thm:mgonproddistance}, the optimal point $z=\ensuremath{\alpha}_1 +\ensuremath{\delta} e^{\im \ensuremath{\theta}}$ for the inner hexagon is achieved for $\ensuremath{\theta}=0$. This gives us a lower bound for the minimal product distance for the $1$st hexagon inscribed in a circle with radius $r_1={\ensuremath{d_{\text{\rm min}} } }$ as \begin{align} \prod_{k=1}^6 |\ensuremath{\delta} -\ensuremath{\alpha}_{n_k^{(1)}}|^2 \geq |r_1^6-\ensuremath{\delta}^6|^2 =r_1^{2\cdot 6} \cdot {\underbrace{|1-(\frac{\ensuremath{\delta}}{{\ensuremath{d_{\text{\rm min}} } }})^{6}|}_{=c^6(\ensuremath{\delta},{\ensuremath{d_{\text{\rm min}} } })}}^2\geq r_1^{12} \cdot c^{12} \end{align} where we assume ${\ensuremath{\delta}}/{{\ensuremath{d_{\text{\rm min}} } }}\leq 1/2$, since $\ensuremath{\delta}<{\ensuremath{d_{\text{\rm min}} } }/2=r_1/2$. The radius for the $n$th hexagon is $n{\ensuremath{d_{\text{\rm min}} } }$, where on the $n$th honeycomb the $6n$ zeros, are the vertices of $n$ rotated hexagons. The smallest radius $r_n$ is given hereby with the law of cosine as \begin{align} r_{n} =\sqrt{ {\ensuremath{d_{\text{\rm min}} } }^2(1+n^2-2n\cos(\pi/3))} = {\ensuremath{d_{\text{\rm min}} } }\sqrt{1+n^2-n}\geq {\ensuremath{d_{\text{\rm min}} } }(n-1) \end{align} for $n\geq 2$, see \figref{fig:hexagonalinring}. If $n=1$ we set $r_1={\ensuremath{d_{\text{\rm min}} } }$. Then we get for the product of distance of the $n$th honeycomb for $n\geq 2$ \begin{align} p^{(n)}= \prod_{k=1}^{6n} |\ensuremath{\delta}-\ensuremath{\alpha}_{n_k}|^2 =\prod_{m=1}^n \prod_{k=1}^6 |\ensuremath{\delta}-\ensuremath{\alpha}_{n_k^{(m)}}|^2 \geq \prod_m (c{\ensuremath{d_{\text{\rm min}} } }(n-1))^{12} =(\ensuremath{\tilde{d}_{\text{min}}} (n-1))^{12n} \end{align} with $\ensuremath{\tilde{\dmin}}=c {\ensuremath{d_{\text{\rm min}} } }$. If we have $\ensuremath{n_{\text{max}}}$ nested honeycombs we have up to \begin{align} N=1+\sum_{n=1}^{\ensuremath{n_{\text{max}}}} 6n =1+ 3\ensuremath{n_{\text{max}}}(\ensuremath{n_{\text{max}}}+1) \end{align} zeros packed, which gives the lower bound \begin{align} \ensuremath{n_{\text{max}}} =\lceil\sqrt{\frac{4N-1}{12}}-\frac{1}{2}\rceil \end{align} of nested honeycombs, yielding to \begin{align} g(\ensuremath{\alpha})& \geq \ensuremath{\delta}^2 \left(\ensuremath{\tilde{d}_{\text{min}}} \prod_{n=2}^{\ensuremath{n_{\text{max}}}} (\ensuremath{\tilde{d}_{\text{min}}}(n-1))^{n}\right)^{12} =\ensuremath{\delta}^2\left(\ensuremath{\tilde{d}_{\text{min}}} \prod_{n=1}^{\ensuremath{n_{\text{max}}}-1} \ensuremath{\tilde{d}_{\text{min}}}^n n^{n+1}\right)^{12}\\ &= \ensuremath{\delta}^2 \left(\ensuremath{\tilde{d}_{\text{min}}}\cdmin^{\frac{\ensuremath{n_{\text{max}}}(\ensuremath{n_{\text{max}}}-1)}{2}}\prod_{n=1}^{\ensuremath{n_{\text{max}}}-1} n^{n+1}\right)^{12}\notag \\ & \geq \ensuremath{\delta}^2\ensuremath{\tilde{d}_{\text{min}}}^{\ 6(\ensuremath{n_{\text{max}}}^2 -\ensuremath{n_{\text{max}}} +2)} \cdot [(\ensuremath{n_{\text{max}}}-1)!H(\ensuremath{n_{\text{max}}}-1)]^{12}, \end{align} where the last factor is the \emph{hyperfactorial} $H(\ensuremath{n_{\text{max}}}-1)=\prod_{n=1}^{\ensuremath{n_{\text{max}}}-1} n^n$. Note, that we had to add the distance square $|\ensuremath{\alpha}_1+\ensuremath{\delta} e^{i\ensuremath{\theta}}-\ensuremath{\alpha}_1|^2=\ensuremath{\delta}^2$ of the centroid zero $\ensuremath{\alpha}_1$. Combining \eqref{eq:optimizationalpm} ,\eqref{eq:optimizationPset} and \eqref{eq:lbound} would yield the final noise bound. \if0 It is not immediately clear if $r=R+\ensuremath{\delta}$ is the optimal solution. Surely, it would be, if $N=2$, since then $\ensuremath{\alpha}_2$, can be chosen to lie anywhere in the ring with appropriate distance ${\ensuremath{d_{\text{\rm min}} } }$ as close to the other circle as possible, since the nominator is $R^2+(R+\ensuremath{\delta})^2 +2R(R+\ensuremath{\delta})\cos(\ensuremath{\gamma})$ where $\cos(\ensuremath{\gamma})=({\ensuremath{d_{\text{\rm min}} } }^2-2R^2)/2R^2$. Hence we get \begin{align} g_2=(2R)^{-2}\frac{R^2+(R+\ensuremath{\delta})^2 +({\ensuremath{d_{\text{\rm min}} } }^2-2R^2)(R+\ensuremath{\delta})/2}{(R+\ensuremath{\delta})^{2}-1}((R+\ensuremath{\delta})^{4}-1) \end{align} For $N=3$ we have two zeros to place, which is again no problem due to the symmetry of the ring. But for $N\geq 4$, this depends on ${\ensuremath{d_{\text{\rm min}} } }$. To minimize the numerator we have to place the $N-1$ zeros such that they have minimal distance to $r$. Clearly, if the zeros would be uniform distributed around the ring, this would yield large distances. Hence, we need the closest packing with a minimal pairwise distance constraint, which is the circle packing problem in the compact set $\ensuremath{{\mathscr{R}}}$. Since the numerator is independent of how we place the zeros, we assume that the zeros are placed in a disk with same area as the ring $\ensuremath{{\mathscr{R}}}$. Then we can chose for large $N$ asymptotic a hexagonal packing, which are ordered around certain hexagons. The distance products can then be lower bounded by the following observation. Let us first note, that we always can chose $\ensuremath{\alpha}_m$ in the numerator to be zero, since for all $n\not=m$ it holds % \begin{align} d_n(\ensuremath{\theta})=|\ensuremath{\alpha}_m -\ensuremath{\alpha}_m+\ensuremath{\alpha}_m + \ensuremath{\delta} e^{\im \ensuremath{\theta}} -\ensuremath{\alpha}_n|= |\ensuremath{\delta} e^{\im \ensuremath{\theta}} -\ensuremath{\tilde{\alp}}_n|\quad,\quad\ensuremath{\tilde{\alp}}_n=\ensuremath{\alpha}_n-\ensuremath{\alpha}_m. \end{align} % where $|\ensuremath{\tilde{\alp}}_n|=d_{n,m}\geq {\ensuremath{d_{\text{\rm min}} } }>\ensuremath{\delta}/2$. To minimize the product of all $d_n(\ensuremath{\theta})$ we have to place the zeros, such that $d_n(\ensuremath{\theta})=a$. all have pairwise distance ${\ensuremath{d_{\text{\rm min}} } }$, which is the smallest possible distance for each $d_{n,m}$ \fi \iflong From this result we can derive an upper bound for the noise power which guarantees error free decoding for the RFMD decoder of any autocorrelation codebook, by setting $\ensuremath{\delta}={\ensuremath{d_{\text{\rm min}} } }/2$, which yields to \begin{align} N_0 \leq\frac{1}{2^{N+2}}\cdot \frac{R^2-1}{R^{3N}-R^N} \cdot\left( \frac{2^6-1}{2^6}\right)^{\ensuremath{n_{\text{max}}}^2-\ensuremath{n_{\text{max}}} +2} {\ensuremath{d_{\text{\rm min}} } }^{6(\ensuremath{n_{\text{max}}}^2 -\ensuremath{n_{\text{max}}} +2)} \cdot ((\ensuremath{n_{\text{max}}}-1)! H(\ensuremath{n_{\text{max}}}-1))^{12} \end{align} \fi \subsection{Noise Bounds for Huffman Polynomials} Note, the noise energy bound \eqref{eq:worstcaseepsilon} is deterministic and not in mean. First of all, for fixed $R$ and ${\ensuremath{d_{\text{\rm min}} } }$ the number $N$ of zeros we can place in the ring $\ensuremath{{\mathscr{R}}}$ is bounded by \eqref{eq:Nboundring}. We plotted in \figref{fig:worstcase_ring} the noise power bound for fixed $N$ over various $R$. Here, we set ${\ensuremath{d_{\text{\rm min}} } }=\sqrt{\pi(R^2-R^{-2})/N\sqrt{12}}$ and assume that all zeros are place in a circle of area $\pi(R^2-R^{-2})$ with center at $R$. This is the worst packing for $N$ zeros. However, it can be seen that the bound is not very sharp if $N$ increases. We assume that Huffman sequences, placed on vertices of two $N-$gons are the best case. However, the optimal radius in the sense of maximal noise robustness for a fixed $N$ is still unknown. Nevertheless, the simulation and analysis suggest that a radius close to the optimal might be given if the uniform circle neighborhoods touching each other, see \figref{fig:optimalradius} and \figref{fig:AnalyticlowerboundHuffconj} for a simulation of $N=6$. In fact, the root-neighborhoods are directed, depending on the particular choice of the other zeros. Hence, in average, the root-neighborhoods will more likely be bounded by an ellipse. Also the outside root-neighborhoods have larger radii than the insides, which suggest also a heterogeneous neighborhoods. \begin{figure}[h] \centering \includegraphics[width=0.8\textwidth]{worstcase_ring_N0.eps} \caption{Worst case noise bounds for various $N$ and densest packing in circles of radius $R$.} \label{fig:worstcase_ring} \end{figure \begin{thm}[Noise Bound for Huffman BMOCZ] Let $N=4M$ for some $M\geq 3$ and $\Calg(N,R)$ be the set of normalized Huffman sequences in ${\ensuremath{\mathbb C}}^{N+1}$ with radius $R>1$. Then the minimal pairwise distance is given by % \begin{align} {\ensuremath{d_{\text{\rm min}} } }=2R^{-1} \sin(\pi/N) \end{align} % and the maximal noise power which guarantees root neighborhoods in circles of radius $\ensuremath{\delta}\in[0,{\ensuremath{d_{\text{\rm min}} } }/2)$ is given by % \begin{align} N_0 &\leq \frac{1}{R^{8M}+1} \frac{(R+\ensuremath{\delta})^{2}-1}{(R +\ensuremath{\delta})^{8M}-1} \cdot R^{2-4M} \ensuremath{\delta}^4(2R^{-1}\sin(\pi/N)-\ensuremath{\delta})^4 \prod_{m=3}^M m^4 \cdot\\ & \quad \quad \left(\frac{\sin(2\pi/N)-\sin(4\pi/N)-2\sin(\pi/N)}{2(1-\sin(2\pi/N))}\right)^{4M-12}\label{eq:AnalyticlowerboundHuffCor} \end{align} The worst case Huffman sequence is given by all zero inside the unit circle except one. \end{thm} \begin{remark} If \conref{con:ngon} holds then we would get the noise power bound % \begin{align} N_0 \leq \frac{R^{-3N}}{R^{-2N}+1} \cdot \frac{(R+R^{-1} \sin(\pi/N))^2-1}{(R+R^{-1}\sin(\pi/N))^{2N}-1}\cdot (1-(1-\sin(\pi/N))^N)^2 \end{align} % which yields to the bounds over $N$ in \figref{fig:AnalyticlowerboundHuffconj} for $\ensuremath{\delta}=\ensuremath{\delta}_{max}=R^{-1}\sin(\pi/N)$. The red line shows the bound for a radius where all root neighborhoods remain disjoint. \end{remark} \begin{figure}[t] \begin{subfigure}[b]{0.485\textwidth} \def0.65\textwidth} \footnotesize{ \import{\pdfdir}{OptimaRadiusHuffmanBMOCZ.pdf_tex} {1.07\textwidth} \tiny{ \hspace{-0.5cm} \import{\pdfdir}{HuffCodebook_deltaepsilon_dmin2.pdf_tex}} \caption{Root neighborhood radius $\delta$ over SNR for worst Huffman sequence with $\Ropt=1.5538,1.3287,1.1791$ from \eqref{eq:worstcaseepsilon} by quantizing $\ensuremath{\theta}$ with $1000$ points.} \label{fig:deltaepsilon} \end{subfigure} \hspace{0.2cm} \begin{subfigure}[b]{0.485\textwidth} \includegraphics[width=1.25\textwidth]{HuffmanEpsConjBound.pdf} \caption{Analytic Noise power bounds in dB with \conref{con:ngon} over various $N$ for different radii $R$.\\} \label{fig:AnalyticlowerboundHuffconj} \end{subfigure} \caption{Noise power bounds depending on radius $R$ and order $N$.} \end{figure} \begin{proof} Since the zeros are lying on two regular $N-$gons, and the nominator only needs one outer zero to be maximize, we can assume that the worst case scenario is given by \figref{fig:huffmanworst}. From \eqref{eq:optimizationPsettht} we obtain for the Huffman Zero Codebook % \begin{align} f^2(\ensuremath{\mathscr Z}) &=\min_{\ensuremath{\alpha}\in\ensuremath{\mathscr Z}} |x_N(\ensuremath{\alpha})|^2\min_{m\in[N]\setminus n} \min_{\ensuremath{\theta}} \frac{\prod_{n} |\ensuremath{\alpha}_m + \ensuremath{\delta} e^{i\ensuremath{\theta}} -\ensuremath{\alpha}_m|^2}{\sum_n |\ensuremath{\alpha}_m +\ensuremath{\delta} e^{i\ensuremath{\theta}}|^{2n}}\\ &=\min_{\ensuremath{\alpha}\in\ensuremath{\mathscr Z}} |x_N(\ensuremath{\alpha})|^2 \min_{\ensuremath{\theta}} \frac{\prod_{n} |\ensuremath{\alpha}_1 + \ensuremath{\delta} e^{i\ensuremath{\theta}} -\ensuremath{\alpha}_m|^2}{\sum_n |\ensuremath{\alpha}_1 +\ensuremath{\delta} e^{i\ensuremath{\theta}}|^{2n}}. \end{align} % Note, that $|x_N(\ensuremath{\alpha})|$ is minimized by \eqref{eq:hufflastfirst} if all zeros lie outside the unit circle, i.e., % \begin{align} \min|x_N(\phi)|^2\geq \frac{ R^{-2K}}{1+R^{-2K}} = \frac{1}{R^{2K}+1}. \end{align} % Hence, by choosing $\ensuremath{\theta}=\pi$ and all zeros inside the unit disc except one, we get with \eqref{eq:huffmanMgonbound2} from \lemref{eq:gHuffman} % \begin{align} f^2(\ensuremath{\mathscr Z})&\geq \frac{1}{R^{8M}+1} \frac{(R+\ensuremath{\delta})^{2}-1}{(R +\ensuremath{\delta})^{8M}-1} \cdot r^{4M-2} \ensuremath{\delta}^4(2a-\ensuremath{\delta})^4 \prod_{m=3}^M m^4 \cdot\\ & \quad \quad \left(\frac{\sin(2\pi/N)-\sin(4\pi/N)-2\sin(\pi/N)}{2(1-\sin(2\pi/N))}\right)^{4M-12}\label{eq:AnalyticlowerboundHuff} \end{align} \end{proof} In \figref{fig:deltaepsilon} we plotted the exact bound \eqref{eq:optimizationPsettht} for Huffman Codebooks with optimal radius \eqref{eq:optimalradiusexactrho} for $N=4,8,16$ starting with the maximal radius $\ensuremath{\delta}_{max}={\ensuremath{d_{\text{\rm min}} } }/2$. It can be seen that for increasing $N$ the $\ensuremath{\delta}_{max}$ shifts to higher SNR, which indicates less noise robustness. \begin{figure}[t] \centering \includegraphics[width=0.7\textwidth]{HuffmanBoundsAnalytic.pdf} \caption{Analytic Noise power bounds \eqref{eq:AnalyticlowerboundHuff} in dB for Huffman sequences with $R=1.02$ and various $N$.} \label{fig:AnalyticlowerboundHuff} \end{figure} \iflong \todostart If the zeros of a normed polynomial are bounded in a ring and at least one zero lies on the boundary, then we get an upper bound for the first and last coefficient by \cite[Thm (27,4)]{Mar77} \begin{align} R=\max_m |\ensuremath{\alpha}_m|\leq \sqrt{1+ \sum_{n=0}^{N-1} \frac{|x_n|^2}{|x_N|^2}} = \sqrt{\frac{|x_N|^2 + (1-|x_N|^2)}{|x_N|^2}}=\frac{1}{|x_N|} \quad \ensuremath{\Leftrightarrow} \quad |x_N|\leq R^{-1} \end{align} Consider the conjugate-reciprocal polynomial ${\ensuremath{\mathrm X}}^*(z)=\sum_n \cc{a_{N-n}} z^n$, then its zeros are exactly $1/\cc{\ensuremath{\alpha}_m}$ and we get \begin{align} R=\max_m \frac{1}{|\ensuremath{\alpha}_m|} =\frac{1}{ \min_m {|\ensuremath{\alpha}_m|}} \leq \frac{1}{|x_0|} \quad\ensuremath{\Leftrightarrow} \quad |x_0|\leq R^{-1} \end{align} Moreover, it holds \begin{align} R^N |x_N| \geq |x_0|=|x_N|\prod_n |\ensuremath{\alpha}_n|\geq |x_N| R^{-N} \quad\ensuremath{\Leftrightarrow} \quad R^N\geq \frac{|x_0| }{|x_N|} \geq R^{-N} \end{align} If the energy of ${\ensuremath{\mathbf x}}$ is normalized this gives \begin{align} |\alpmax| \leq 1 + \frac{1}{|x_\xord|}\quad \ensuremath{\Leftrightarrow} \quad |x_\xord|\leq \frac{1}{|\alpmax|-1} \end{align} since $|\alpmax|>1$. Hence, if $|\alpmax|$ increases, the leading coefficient has to decrease and $\ensuremath{\epsilon}$ decreases rapidly, independently of ${\ensuremath{d_{\text{\rm min}} } }$. \todoend \fi \if0 Since we want to have the $\ensuremath{\epsilon}$ bound as large as possible we need to find a tight upper bound of the optimization problem. Let us bound the first case \begin{align} \max_{\ensuremath{\theta}} \frac{ N(|\ensuremath{\alpha}_m +\ensuremath{\delta} e^{i\ensuremath{\theta}}|^{2N}+1}{|x_N|^2 \prod_n |\ensuremath{\alpha}_m + \ensuremath{\delta} e^{i\ensuremath{\theta}} -\ensuremath{\alpha}_n|^2} \leq \frac{1+ N\max_{\ensuremath{\theta}}|\ensuremath{\alpha}_m +\ensuremath{\delta} e^{i\ensuremath{\theta}}|^{2N}}{|x_N|^2\min_{\ensuremath{\theta}} \prod_n |\ensuremath{\alpha}_m + \ensuremath{\delta} e^{i\ensuremath{\theta}} -\ensuremath{\alpha}_n|^2} \end{align} Since by assumption it holds $|\alpmax|>1$, we can derive a universal upper bound\footnote{This is actually Bernstein's Lemma.} \begin{align} |\uw(z)| \leq \ensuremath{\epsilon} \cdot \sqrt{1+N}(|\alpmax|+\ensuremath{\delta})^N \quad,\quad z\in\bigcup C_m(\ensuremath{\delta}). \end{align} On the other hand, a universal magnitude lower bound of the original polynomial is given by \begin{align} |{\ensuremath{\mathrm x}}(z)| &= |x_\xord| \prod_{n=1}^\xord |z-\ensuremath{\alpha}_n| \quad,\quad z \in C_m(\ensuremath{\delta})\\ &= |x_\xord| \prod_n |\ensuremath{\alpha}_m+\ensuremath{\delta} e^{i\ensuremath{\theta}} -\ensuremath{\alpha}_n|\quad,\quad \ensuremath{\theta}\in[0,2\pi) \intertext{using the reverse triangle inequality gives\footnotemark} & \geq |x_\xord|\prod_n | |\ensuremath{\alpha}_m-\ensuremath{\alpha}_n|-\ensuremath{\delta}| \geq |x_\xord| \ensuremath{\delta} \prod_{n\not=m} ({\ensuremath{d_{\text{\rm min}} } }-\ensuremath{\delta}). \end{align} \footnotetext{Note, that $|\ensuremath{\alpha}_l-\ensuremath{\alpha}_n|>{\ensuremath{d_{\text{\rm min}} } }>\ensuremath{\delta}$ for $l\not=n$.}% Hence we get for all $z\in\bigcup C_n(\ensuremath{\delta})$: \begin{align} |{\ensuremath{\mathrm x}}(z)|\geq |x_\xord| \ensuremath{\delta} ({\ensuremath{d_{\text{\rm min}} } }-\ensuremath{\delta})^{N-1}. \end{align} To apply Rouch{\'e}'s Theorem, we have to show $|\uw(z)|<|{\ensuremath{\mathrm x}}(z)|$ for autocorrelation $z\in\bigcup C_l(\ensuremath{\delta})$, which gives us the local Lipschitz constant \begin{align} \ensuremath{\epsilon}=\ensuremath{\epsilon}({\ensuremath{\mathbf x}},\ensuremath{\delta}) \leq \frac{|x_\xord| \ensuremath{\delta} ({\ensuremath{d_{\text{\rm min}} } }-\ensuremath{\delta})^{N-1}}{\sqrt{1+ N}(|\alpmax|+\ensuremath{\delta})^N}. \end{align} Since $\ensuremath{\delta}<{\ensuremath{d_{\text{\rm min}} } }/2$, all $B_n(\ensuremath{\delta})$ are disjoint and ${\ensuremath{\mathrm y}}(z)$ has exactly one zero in each $n$th ball $B_n(\ensuremath{\delta})$ by \thmref{thm:rouche}. \fi \section{Comparison to Training Schemes} We will compare our noncoherent BMOCZ scheme to noncoherent QPSK with pilot signaling. Considering one-shot scenarios, there are not many noncoherent comparisons possible in this scenario. We refer the reader to \cite{JHV15},\cite{Cho18} and \cite{FHV13} based on self-coherent OFDM schemes. However, our proposed Huffman BMOCZ schemes outperforms their BER performance. We assume in the simulations the following scenario \begin{itemize} \item Channel is time-invariant for $N$ time-instants \item Receive duration is $N= {\ensuremath{K}}+{\ensuremath{L}}$ \item Block length of transmitted signal is ${\ensuremath{K}}+1$ \item We have independent Rayleigh distributed channel gains $h_{\ensuremath{l}}\in\ensuremath{\C^N}(0,\pd^{\ensuremath{l}})$ with $\pd\leq 1$. \item After each block transmission the CIR can change arbitrary, e.g., caused by a fast-varying channel or due to a sporadic one-shot communication, where the next block message might occur much later, such that the channel state and user position can change drastically, resulting in fully uncorrelated CIRs. \end{itemize} Note, the maximum-likelihood (ML) detection is equal to the maximum a posteriori detection (MAP), if the channel is known. If we do joint ML over ${\ensuremath{\mathbf x}},\vh$ then the ML and LS would also be the same, but this requires blind deconvolution, see \cite[(12-16)]{MMF98}. \iflong Moreover, the ML detection is given by minimizing the $\ell_2-$norm (least-square error between model output and observation) of ${\ensuremath{\mathbf y}}-\vh*{\ensuremath{\mathbf x}}$, see \cite[(12-14)]{MMF98}. \fi We will compare to the following scenarios \begin{enumerate} \item BPSK and full channel knowledge (coherent) by using . Zero-forcing (ZF) and hard thresholding bit wise \begin{align} m_{\ensuremath{k}} = \begin{cases} 1, \Re({\ensuremath{\hat{\mathbf x}}}_{\ensuremath{k}})>0\\ 0, \Re({\ensuremath{\hat{\mathbf x}}}_{\ensuremath{k}})<0\end{cases}\quad,\quad {\ensuremath{k}}=1,2,\dots,{\ensuremath{K}} \end{align} \item QPSK with ${\ensuremath{L}}$ pilots. Here we assume that ${\ensuremath{K}}=2{\ensuremath{L}}$. We decode by separating Real and Imaginary part \begin{align} m_{\ensuremath{k}} =\begin{cases} 1, \Re({\ensuremath{\hat{\mathbf x}}}_{\ensuremath{k}})>0\\ 0, \Re({\ensuremath{\hat{\mathbf x}}}_{\ensuremath{k}})<0\end{cases}\quad,\quad m_{{\ensuremath{k}}+{\ensuremath{L}}} =\begin{cases} 1, \Im({\ensuremath{\hat{\mathbf x}}}_{\ensuremath{k}})>0\\ 0, \Im({\ensuremath{\hat{\mathbf x}}}_{\ensuremath{k}})<0\end{cases}\quad,\quad {\ensuremath{k}}=1,2,\dots,{\ensuremath{K}} \end{align} This follows form the minimum distance decoder, see \cite[(7.25)]{LM93}. \end{enumerate} \begin{figure}[t] \hspace{-2.3cm} \includegraphics[width=1.24\linewidth]{pilotQPSK_L16K1to33_runpp43kLK} \vspace{-1cm} \caption{Pilot based QPSK in time domain versus Huffman BMOCZ with $\pd=1$ for ${\ensuremath{K}}=16$ and ${\ensuremath{L}}=1,4,9,12,17,25,33$.} \label{fig:pilotQPSKvsBMOCZ} \end{figure} Interesting scenarios for the BMOCZ scheme are distributed wireless sensor networks which require \begin{itemize} \item Low-Power (transmitter and receiver) \item Low-Latency \item No feedback and no channel information at transmitter \item Short block-length, ${\ensuremath{K}}\simeq{\ensuremath{L}}$ \item One-shot communication, channel is used only once, sporadic in time \end{itemize} If ${\ensuremath{K}}<{\ensuremath{L}}$ there is no way to learn the channel with pilots in one shot. Moreover, the low-power assumption is not suitable to use energy detector if we need to transmit more than $1$ bit. Higher order MOZ modulation might be also considered. This constraints, rule out \begin{itemize} \item CDMA: usually requires ${\ensuremath{K}}\gg {\ensuremath{L}}$, \cite[p.92]{TV05} \item Clasical ODFM: needs channel for decoding. Hence, we could assume again ${\ensuremath{\mathbf x}}={\ensuremath{\mathbf u}}+{\ensuremath{\mathbf d}}$ And in $\hat{{\ensuremath{\mathbf d}}}$ using QAM. for example QPSK. \end{itemize} The BMOCZ scheme does not need channel length knowledge at the transmitter! On the other hand, any pilot data needs assumption on the channel length. If the pilots are to short, it is impossible to estimate exactly the channel, even in the noiseless case. We will investigate a scenario, where we blind transmit with $P=\lfloor {\ensuremath{L}}/2\rfloor$ pilots and $D=P$ data and receive only ${\ensuremath{L}}+{\ensuremath{L}}$ taps, regardless of the true channel length ${\ensuremath{L}}$. Hence, either we take to much sample or to less at the receiver. This will affect the performance of pilot-based schemes, which we simulated for QPSK and OFDM, see \figref{fig:pilotQPSKvsBMOCZ}. Here, the BER performance suffers dramatically if the channel length at the transmitter is underestimated, rendering a reliable communication impossible. However, the BMOCZ decoder depends heavily of the maximal channel length $L$. It can be seen in the simulation, that an overestimating of $L$ (fast decreasing power profile) is not affecting the BER performance much, see \figref{fig:MLandRFMD_snr} and \figref{fig:MLandRFMD_ebno}. The GW-DiZeT decoder demands no complexity at all and allows with the DFT an easy and probably analog realization (using delayed amplified circuits). The complexity at the transmitter consist of a fixed codebook of size $2^{\ensuremath{K}}$ in ${\ensuremath{\mathbb C}}^{{\ensuremath{K}}+1}$ dimensions. \iflong \subsection{OFDM with pilots} It is possible to use the pilot scheme in the frequency domain such that OFDM modulation can be used, see for example \cite{WBJ15b}, \cite{ZH97}. If ${\ensuremath{K}}$ is odd and ${\ensuremath{L}}<({\ensuremath{K}}+1)/2=D$ we can modulate our time signal ${\ensuremath{\mathbf x}}$ in the frequency domain ${\ensuremath{K}}+1=2D$ by splitting it in pilot and data symbols \begin{align} {\ensuremath{\mathbf F}}_{2D}{\ensuremath{\mathbf x}}= {\ensuremath{\mathbf d}} + {\ensuremath{\mathbf u}} = \begin{pmatrix} d_0 \\ 0 \\ d_1 \\ \vdots \\ d_{D-1}\end{pmatrix} +\begin{pmatrix} 0 \\ u_0 \\ \vdots\\ u_{D-1}\\ 0\end{pmatrix}\in{\ensuremath{\mathbb C}}^{{\ensuremath{K}}+1} \end{align} where we use pilots ${\ensuremath{\mathbf u}}_{\ensuremath{\mathscr P}}=(1,1,\dots,1)^T\in{\ensuremath{\mathbb C}}^D$ on the subcarriers $\ensuremath{\mathscr P}=\{1,3,5,\dots,\}$ and data ${\ensuremath{\mathbf d}}_{\Dset}=(d_0,d_1,\dots,d_{D-1})^T$ on $\Dset=\{0,2,4,\dots\}$. To keep the orthogonality of ${\ensuremath{\mathbf d}}$ and ${\ensuremath{\mathbf u}}$, we have to add a cyclic prefix $(x_{{\ensuremath{K}}+1},\dots, x_{{\ensuremath{K}}-D})^T$ of length $D$ to the transmitted signal ${\ensuremath{\mathbf x}}$, which we will denote by ${\ensuremath{\mathbf x}}^{CP}\in{\ensuremath{\mathbb C}}^{{\ensuremath{K}}+1+D}$. Hence, we obtain at the receiver \begin{align} \vr =\{ ({\ensuremath{\mathbf x}}^{CP}*\vh)_k + w_{k}\}_{k=D}^{{\ensuremath{K}}+1+D}={\ensuremath{\mathbf x}}\circledast\ensuremath{\tilde{\mathbf h}} + \vn\in{\ensuremath{\mathbb C}}^{{\ensuremath{L}}+1}, \end{align} where the samples $\{D,D+1,\dots,D+{\ensuremath{K}}+1\}$ correspond to the circular convolution, see for example \cite[Sec.3.4.4]{TV05}. Here we used $\ensuremath{\tilde{\mathbf h}}=[\vh, {\ensuremath{\mathbf 0}}_{{\ensuremath{K}}+1-{\ensuremath{L}}}]\in{\ensuremath{\mathbb C}}^{{\ensuremath{K}}+1}$. The true CIR length ${\ensuremath{L}}$ can be less, but not larger than $D$. Moreover, if ${\ensuremath{L}}\geq {\ensuremath{K}}+1$ it is not possible to use a pilot scheme at all, since the channel estimation problem is under-determined. Also CP in wireless OFDM does not make sense, since we can not zero pad the channel $\vh$. Usually, in OFDM one assumes that the CIR length is at most $12.5\%$ of the frame length ${\ensuremath{K}}+1$, which results in longer frame length and hence latency. This gives an additional latency of $D$ time steps, compared to our scheme. Note, that actually ${\ensuremath{\mathbf x}}^{CP}*\vh\in{\ensuremath{\mathbb C}}^{4D-1}$, but we will not need to sample the last $D-1$ taps of the linear convolution. Hence we have to wait at least ${\ensuremath{K}}+1+D=3D$ time steps, which gives the same spectral efficiency $K+1/N$ as . The received signal in the frequency domain will be \begin{align} {\ensuremath{\mathbf F}}\vr = {\ensuremath{\mathbf F}}_{2D} ({\ensuremath{\mathbf x}}\circledast \ensuremath{\tilde{\mathbf h}})+ {\ensuremath{\mathbf F}}\vn = \sqrt{2D}{\ensuremath{\mathbf F}}{\ensuremath{\mathbf x}} \bullet {\ensuremath{\mathbf F}}\ensuremath{\tilde{\mathbf h}} + {\ensuremath{\mathbf F}}\vn. \end{align} Sampling at all odd frequencies gives \begin{align} ({\ensuremath{\mathbf F}}\vr)_{\ensuremath{\mathscr P}} = \sqrt{2D}\cdot{\ensuremath{\mathbf 1}}_D\bullet {\ensuremath{\mathbf F}}_D \vh + {\ensuremath{\mathbf F}}_D \vn_D = {\ensuremath{\mathbf F}}_D \vh' \in{\ensuremath{\mathbb C}}^D. \end{align} Hence, we get a channel estimation $\vh'$ in the Fourier domain and by ${\ensuremath{\mathbf F}}_D^*({\ensuremath{\mathbf F}}\vr)_{\ensuremath{\mathscr P}}=\vh'$, which is exact in the noise free case. To estimate the data symbols, we can use the Matched filter \begin{align} e^{-i2\pi/D}\cc{{\ensuremath{\mathbf F}}_D \vh}\bullet ({\ensuremath{\mathbf F}}\vr)_{\Dset} = \sqrt{2D}\cdot {\ensuremath{\mathbf d}} \bullet |{\ensuremath{\mathbf F}}_D \vh|^2+ e^{-i2\pi/D}\cdot({\ensuremath{\mathbf F}}{\ensuremath{\mathbf w}})_{\Dset}\bullet {\ensuremath{\mathbf F}}_D\vh. \end{align} Now, estimation is for each ${\ensuremath{k}}$th data symbol separately possible with the usual QPSK decoder. The average ``sampled'' receive energy is then given by \begin{align} \Expect{ \|{\ensuremath{\mathbf x}}\circledast\ensuremath{\tilde{\mathbf h}}\|^2} = \Expect{\sum_m\sum_{k,l} x_k \cc{x_l} \tilde{h}_{m\ominus k}\cc{\tilde{h}_{m\ominus l}} } = \sum_m\sum_{k} |x_k|^2\Expect{ |\tilde{h}_{m\ominus k}|^2 } =\Norm{{\ensuremath{\mathbf x}}}_2^2 {\sum_{{\ensuremath{l}}=0}^{{\ensuremath{L}}} p^{\ensuremath{l}}} = \Expect{\Norm{\vh}_2^2} \end{align} We will again normalize the channel energy by $1/\sqrt{{\ensuremath{L}}}$ and set the signal energy ${\ensuremath{\mathbf x}}^{CP}$ to $3D$, such that we get normalized transmit and received power. The bit rate is in a non-continuous block transmission for ${\ensuremath{L}}=D$ \begin{align} R_b = \frac{2D}{3D}=\frac{2}{3}. \end{align} \fi \section{introduction} The future generation of wireless networks faces a diversity of new challenges. Trends on the horizon -- such as the emergence of the Internet of Things (IoT) and the tactile Internet -- have radically changed our thinking about how to scale the wireless infrastructure. Among the main challenges new emerging technologies have to cope with is the support of a massive number (billions) of devices ranging from powerful smartphones and tablet computers to small and low-cost sensor nodes. These devices come with diverse and even contradicting types of traffic including high speed cellular links, device-to-device connections, and wireless links carrying short-packet sensor data. Short messages of sporadic nature \cite{Wunder2015:sparse5G} will dominate in the future and the conventional cellular and centrally-managed wireless network infrastructure will not be flexible enough to keep pace with these demands. Although intensively discussed in the research community, the most fundamental question here on how we will communicate in the near future under such diverse requirements remains largely unresolved. A key problem is how to acquire, communicate, and process channel information. Conventional channel estimation procedures require a substantial amount of resources and overhead. This overhead can dominate the intended information exchange when the message is short and the traffic sporadic. \emph{Noncoherent and blind strategies}, provide a potential way out of this dilemma. Classical approaches like blind equalization have been already investigated in the engineering literature \cite{Godard1980,For72,CP96}, but new noncoherent modulation ideas which explicitly account for the short-message and sporadic type of data are required \cite{Jung2014}. In many wireless communication scenarios the transmitted signals are affected by multipath propagation and the channel will therefore be frequency-selective. Additionally, in mobile and time-varying scenarios one encounters also time-selective fast fading. In both cases channel parameters typically have a random flavour and potentially cause various kinds of interference. From a signal processing perspective it is therefore necessary to take care of possible signal distortions, at the receiver and potentially also at the transmitter. A well know approach to deal with such channels is to modulate data on multiple parallel waveforms which are well-suited for the particular channel conditions. One of the most simple approaches for the frequency-selective case is orthogonal frequency division multiplexing (OFDM). When the maximal channel delay spread is known inter-symbol-interference (ISI) can be avoided by a suitable guard interval and an orthogonality of the subcarriers ensures that there is no interference-carrier-interference. On the other hand, from an information-theoretic perspective, random channel parameters are helpful from a diversity view point. To exploit multipath diversity the data has to be spread over the subcarriers. To coherently demodulate the data at the receiver and to also make use of diversity the channel impulse response (CIR) has to be known at the receiver. To gain knowledge of the CIR training symbols (pilots) are included in the transmitted signal, leading to a substantial overhead when the signal length is on the order of the channel length. Furthermore, the pilot density has to be adapted to the mobility and, in particular, OFDM is very sensitive to time-varying distortions due to Doppler shift and oscillator instabilities. Dense CIR updates are then required, which results in complex transceiver designs. There are only a few works on noncoherent OFDM schemes in the literature. Some are known as self-heterodyne OFDM or self-coherent OFDM \cite{JHV15,FHV13}. Very recently a noncoherent method for OFDM with Index Modulation (IM) was proposed in \cite{Cho18}, which exploits a sparsity of ${\ensuremath{L}}$ subcarriers out of $N$. The modulation can be seen as a generalized $N-ary$ frequency shift keying (FSK), which uses ${\ensuremath{L}}$ tones (frequencies) and results in a codebook of $M=\binom{N}{{\ensuremath{L}}}$ non-orthogonal constellations. In this work we follow a completely different strategy. We propose to encode each bit of the data payload into one of a conjugate-reciprocal pairs of zeros (in the complex plane) and thereby construct a polynomial whose degree is the number of payload bits. The complex-valued coefficients of the polynomial are in fact the transmit baseband signal samples. We introduced such a non-linear modulation on polynomial zeros first for Huffman sequences in \cite{WJH17a,WJH17b} and demonstrated to perform efficient and reliable convex and non-convex decoding algorithms. However, such optimization algorithms are meant for blind deconvolution, i.e., reconstruct channel and signal simultaneously, and are therefore not necessarily well-suited and efficient to retrieve the digital data from finite alphabets. In this work we will therefore extend our previous ideas and develop and analyze polynomial-factorization-based approaches more concretely from a communication-oriented perspective. We will extend the modulation and encoding principle to general codebooks based on polynomial zeros. We derive and analyse the maximum likelihood decoder, which depends only on the power delay profile of the channel and the noise power. Then, we construct a low complexity decoder for Huffman sequences having a complexity which scales only linearly in the number of bits to transmit. We will demonstrate by numerical experiments that our scheme is able to outperform noncoherent OFDM-IM and pilot based $M-$QAM schemes in terms of bit-error rate. \subsection{Notation} We will use small letters for complex numbers in ${\ensuremath{\mathbb C}}$. Capital Latin letters denote natural numbers and refer to fixed dimensions, where small letters are used as indices. Boldface small letters denote vectors and capitalized letters refer to matrices. Upright capital letters denote complex-valued polynomials in ${\ensuremath{\mathbb C}}[z]$. For a complex number $x=a+\im b$, given by its real part $\Re(x)=a\in{\ensuremath{\mathbb R}}$ and imaginary part $\Im(x)=b\in{\ensuremath{\mathbb R}}$ with imaginary unit $\im=\sqrt{-1}$, its complex-conjugation is given by $\cc{x}=a-\im b$ and its absolute value by $|x|=\sqrt{x\cc{x}}$. For a vector ${\ensuremath{\mathbf x}}\in{\ensuremath{\mathbb C}}^N$ we denote by $\cc{{\ensuremath{\mathbf x}}^-}$ its complex-conjugated time-reversal or \emph{conjugated-reciprocal}, given as $\cc{x_k^-} = \cc{x_{N-k}}$ for $k=0,1,\dots,N-1$. We use ${\ensuremath{\mathbf A}}^*=\cc{{\ensuremath{\mathbf A}}}^T$ for the complex-conjugated transpose of the matrix ${\ensuremath{\mathbf A}}$. For the identity and all zero matrix in $N$ dimension we write ${\ensuremath{\mathbf I}}_N$ respectively ${\ensuremath{\mathbf O}}_N$. By ${\ensuremath{\mathbf D}}_{{\ensuremath{\mathbf x}}}$ we refer to the diagonal matrix generated by the vector ${\ensuremath{\mathbf x}}\in{\ensuremath{\mathbb C}}^N$. The $N\times N$ unitary Fourier matrix ${\ensuremath{\mathbf F}}={\ensuremath{\mathbf F}}_N$ is given entry-wise by $f_{l,k}=e^{\im 2\pi lk/N}/\sqrt{N}$ for $l,k=0,1,\dots,N-1$. The all one respectively all zero vector in dimension $N$ will be denoted by ${\ensuremath{\mathbf 1}}_N$ resp. ${\ensuremath{\mathbf 0}}_N$. The $\ell_p$-norm of a vector ${\ensuremath{\mathbf x}}\in{\ensuremath{\mathbb C}}^N$ is given by $\Norm{{\ensuremath{\mathbf x}}}_p=(\sum_{k=1}^N|x_k|^p)^{1/p}$ for $p\geq 1$. If $p=\infty$ we write $\Norm{{\ensuremath{\mathbf x}}}_{\infty}=\max_k |x_k|$. The expectation of a random variable $x$ is denoted by $\Expect{x}$. We will refer to ${\ensuremath{\mathbf x}}\bullet{\ensuremath{\mathbf y}}:=\operatorname{diag}({\ensuremath{\mathbf x}}){\ensuremath{\mathbf y}}$ as the Hadamard (point-wise) product of the vectors ${\ensuremath{\mathbf x}}, {\ensuremath{\mathbf y}}\in{\ensuremath{\mathbb C}}^N$. \input{dezet} \input{generalzerocodebooks} \input{robustness} \input{simulations} \ifarxiv \input{MaryMOCZ}\fi \input{training} \input{stabilitybound} \section{Conclusion} We introduced a novel modulation scheme based on the zeros of the discrete-time signals to transmit reliable over unknown FIR channels. For the demodulation we presented three different approaches. The first based on a zero observation of the received zeros and deciding by a minimum distance decoder for each single bit (zero) independently. Secondly, a maximum likelihood decoder, which obtains the best performance. If the CIR length is larger than the block length, the ML decoder for Huffman BMOCZ outperforms all comparable known non-coherent signaling schemes. However, this decoder relies on the knowledge of the channels power delay profile and the SNR. Finally, we introduced a low-complexity decoder which decodes the zeros independently by only evaluating the received signal on the zero-codebook, which leads to linear complexity in the number of bits, instead of exponential complexity for the ML decoder. The derivation of bit error probabilities is mathematically hard to carry out, not only due to the overlap of the signals caused by multipath delays, but also in terms of the non-linear encoding in the zero domain. For the RFMD decoder we obtained a local stability analysis in the presence of additive noise, which suggest a proper zero separation of the codebook and channel. The analysis of reliable bit data rates and error probability bounds, based on a careful root neighborhood analysis, might be addressed in a future research. \section{Acknoledgements} The authors would like to thank Richard Kueng and Urbashi Mitra for many helpful discussions. We like to thank the SURF student Mattia Carrera for his support during a summer program at Caltech. Most of the work by Philipp Walk was done during a two year postdoc fellowship at Caltech, which was sponsored by the DFG WA 3390/1. \section*{References} \printbibliography[heading=bibintoc] \appendices \input{mgon} \end{document}

https://arxiv.org/abs/1708.05815

Minimum Hidden Guarding of Histogram Polygons

A hidden guard set $ G $ is a set of point guards in polygon $ P $ that all points of the polygon are visible from some guards in $ G $ under the constraint that no two guards may see each other. In this paper, we consider the problem for finding minimum hidden guard sets in histogram polygons under orthogonal visibility. Two points $ p $ and $ q $ are orthogonally visible if the orthogonal bounding rectangle for $ p $ and $ q $ lies within $ P $. It is known that the problem is NP-hard for simple polygon with general visibility and it is true for simple orthogonal polygon. We proposed a linear time exact algorithm for finding minimum hidden guard set in histogram polygons under orthogonal visibility. In our algorithm, it is allowed that guards place everywhere in the polygon.

\section{Introduction} In the standard version of the art gallery problem one is given a simple polygon $ P $ in the plane that needs to be guarded by a set of point guards~\cite{urrutia2000art}. In other words, we want to find a set of point guards such that every point in $ P $ is seen by at least one of the guards, where a guard $ g $ sees a point $ p $ if the segment $ gp $ is contained in $ P $. Lee and Lin~\cite{lee1986computational} proved that finding the minimum number of point guards needed to guard an arbitrary polygon is NP-hard. The art gallery problem is also NP-hard for orthogonal polygons~\cite{schuchardt1995two} and it even remains NP-hard for monotone polygons~\cite{schuchardt1995two}. In some papers, it is assumed that the visibility is in orthogonal model(r-visibility) instead of standard line visibility. In standard, points $ p $ and $ q $ are visible from each other in the polygon $ P $, if line segment $ pq $ is contained in $ P $. So, in orthogonal visibility, they are r-visible (orthogonally visible) to each other if the axis-parallel rectangle spanned by the points is contained in $ P $~\cite{o2004visibility}. Worman and Keil~\cite{worman2007polygon} studied the orthogonal polygon decomposition problem that is equivalent to the orthogonal art gallery problem and showed that the problem is polynomially solvable for r-visibility. The time complexity of their algorithm is $ O(n^{17}poly \log n) $. Gewali and et. al. ~\cite{gewali1996placing} presented an $ O(n) $ time algorithm for covering a monotone orthogonal polygon with the minimum number of orthogonal star-shaped polygons. Finding minimum covers by star-shaped polygons is equivalent to finding the minimum number of guards needed so that every point in the polygon is visible to at least one guard. Palios and Tzimas~\cite{palios2014minimum} considered the the problem on simple class-3 orthogonal polygons, i.e., orthogonal polygons that have dents along at most $ 3 $ different orientations. They gave an output-sensitive $ O(n+k \log ⁡k) $-time algorithm where $ k $ is the size of a minimum r-star cover. Beidl and Mehrabi~\cite{biedl2016r} showed that problem is NP-hard on orthogonal polygons with holes. A polygon is called \textit{tree polygon} If dual graph of the polygon is a tree. Also, They presented an algorithm for tree polygon in the linear-time. If the guards are only allowed to be on vertices(vertex guard variant), discretization combined with iteration yields an $ O(n^4) $ solution~\cite{couto2007exact} for general orthogonal polygons. In the polygon $ P $, a \textit{hidden set} is defined as a set of points such that no two points in the set are visible to each other. So, a \textit{hidden guard set} is a hidden set which is also a guard set and the entire polygon is visible from some points in it. Finding the minimum or maximum guard set in a polygon can be considered under standard or orthogonal visibility. In the standard Version, Shermer~\cite{shermer1989hiding} was the first to show that the problem of finding a maximum hidden and a minimum hidden guard sets is NP-hard. He established bounds on the maximum size of hidden sets for some polygons. In 1999, Eidenbenz~\cite{eidenbenz1999many} showed that the problem of placing a maximum hiding guards is almost as hard to approximate as the Maximum Clique problem and it cannot be approximated by any polynomial-time algorithm with an approximation ratio of $ n\epsilon $ for some $ \epsilon > 0 $, unless $ P = NP $. He showed that for a simple polygon with holes, it is already true and for a simple polygon without holes there is a constant $ \epsilon > 0 $ such that the problem cannot be approximated with an approximation ratio of $ 1 + \epsilon $. Hurtado and et.al.~\cite{hurtado1996hiding} studied hidden sets of points in arrangements of segments, they provided bounds for its maximum size. Biswas and et. al.~\cite{biswas1994algorithms} studied similar problem called as the maximum independent set of visibility graph instead of maximum hidden guard sets. They proposed an $ O(n^{3}) $ time algorithm for finding the maximum independent set of the convex visibility graph for a restricted class of simple polygons. A stair-case path is monotone along both the $ x $ axis and $ y $ axis directions. Two points inside an orthogonal polygon are visible under \textit{stair-case visibility} if they can be connected by a stair-case path without intersecting its exterior. Eidenbenz and Stamm~\cite{eidenbenz2000maximum} studied the problem of finding a maximum clique in the visibility graph of a simple polygon with $ n $ vertices. They showed that if the input polygons are allowed to contain holes no polynomial time algorithm can achieve an approximation ratio of $ \frac{n^{1/8-\epsilon}}{4} $ for any $ \epsilon> 0 $, unless $ NP = P $. They proposed an $ O(n^{3}) $ algorithm for the maximum clique problem on visibility graphs for polygons without holes. Their algorithm also finds the maximum weight clique, if the polygon vertices are weighted. They also showed that the problem of partitioning the vertices of a visibility graph of a polygon into a minimum number of cliques is APX-hard for polygons without holes. Later, they are proved tight approximability results on finding minimum hidden guard sets in polygons~\cite{eidenbenz2006finding,eidenbenz2002optimum}. Because of the maximum hidden vertex set problem on a given simple polygon is NP-hard Bajuelos and et. al.~\cite{bajuelos2008estimating} proposes some approximation algorithms to solve this problem, first two based on greedy constructive search and the other are based on the general meta-heuristics Simulated Annealing and Genetic Algorithms. Their solutions are very satisfactory in the sense that they are always close to optimal with an approximation ratio of $ 1.7 $, for arbitrary polygons and with an approximation ratio of $ 1.5 $ for orthogonal polygons. Also, they showed that on average the maximum number of hidden vertices in a simple polygon (arbitrary or orthogonal) with $ n $ vertices is $ n/4 $. Ghosh and et. al.~\cite{ghosh2007visibility} presented an algorithm for computing the maximum clique in the visibility graph $ G $ of a simple polygon $ P $ in $ O(n^{2}e) $ time, where $ n $ and $ e $ are number of vertices and edges of $ G $ respectively and also presented an $ O(ne) $ time algorithm for computing the maximum hidden vertex set in the visibility graph $ G $ of a convex fan $ P $. Kranakis and et. al.~\cite{kranakis2009inapproximability} introduced a new hiding problem as called the max hidden edge set problem. given a polygon $ P $, find a minimum set of edges $ S $ of the polygon such that any straight line segment crossing the polygon intersects at least one of the edges in $ S $. They proved the APX-hardness of the problem for polygons without holes. Cannon and et. al.~\cite{cannon2012hidden} considered guarding classes of simple polygons using mobile guards (polygon edges and diagonals) under the constraint that no two guards may see each other. They provided a nearly complete set of answers to existence questions of open and closed edge, diagonal, and mobile guards in simple, orthogonal, monotone, and star-shaped polygons like that every monotone or star-shaped polygon can be guarded using hidden open mobile (edge or diagonal) guards, but not necessarily with hidden open edge or hidden open diagonal guards. Bajuelos and et. al.~\cite{bajuelos2008escondiendo} studied the maximum hidden vertex set problem for two types of polygons spirals and histograms and proposed a linear algorithm. In this paper, we introduce the problem of finding the minimum number of hidden guards to cover a polygon $ P $ under orthogonal visibility. First, we present a linear-time algorithm for guarding monotone orthogonal polygons, we know that there is a linear-time algorithm for the problem~\cite{gewali1992covering,lingas2007note}, but, our algorithm is purely geometric. After that, we present an exact algorithm for finding minimum hidden guard set for histogram polygon and a 2-approximation algorithm for monotone orthogonal polygon under orthogonal visibility. In the following, visibility means orthogonal visibility and guarding is under orthogonal visibility and monotonicity means $ x $-monotonicity unless explicitly expressed(mentioned). The time complexity of hidden guarding is remained open, even for monotone polygon. \begin{figure} \centering \includegraphics[width=\textwidth]{figure1} \caption{Illumination of the vertical decomposition and its notations.} \label{fi:fig1} \end{figure} \section{Preliminaries} \label{ss:ss1} Assume we decompose a simple orthogonal and $ x $-monotone polygon $ P $ with $ n $ vertices into rectangles obtained by extending the vertical edges incident to the reflex vertices of $ P $. A reflex vertex has an interior angle $ \frac{3\pi}{2} $ while convex vertices have an interior angle of $ \frac{\pi}{2} $. It is clear that every orthogonal polygon with $ n $ vertices has $ \frac{n-4}{2} $ reflex vertices. So, after the decomposition of $ P $, $ \frac{n-2}{2} $ rectangles will be obtained, exactly. Let $R =\{R_{1},R_{2},\dots,R_{m}\} $, where $ m = \frac{n-2}{2} $, be the set of rectangles, ordered from left to right according to $ x $-coordinate of their left edges. In the other words, after the decomposition, we named rectangular parts $ R_{1},R_{2},\dots,R_{m} $ from left to right. It is shown in Figure~\ref{fi:fig1}. We name the upper and lower horizontal edges of $ R_{i} $ by $ u_{i} $ and $ l_{i} $, respectively. Assume that $ U =\{u_{1},u_{2},\dots,u_{m} \} $ and $ L=\{l_{1},l_{2},\dots,l_{m} \} $, {\small $1\leq i \leq m $}. For a horizontal line segment $ s $, we denote the $ x $-coordinate of the \textbf{left vertex} of $ s $ by $ x(s) $ that means the $ x $-coordinate of line segment $ s $. Also, we denote the $ y $-coordinate of line segment $ s $ by $ y(s) $. For a vertical line segment $ s' $, we denote the $ x $-coordinate of $ s' $ by $ x(s') $. Similarly, for a point $ p_0 $, $ x $-coordinate and $ y $-coordinate of $ p_0 $ is denoted by $ x(p_0) $ and $ y(p_0) $, respectively. Without reducing generality, we suppose that for every two vertical edge $ e_i $ and $ e_j $ ($ i\neq j $), $ x(e_i)\neq x(e_j) $, so, it seems obvious that for all {\small $ 1 \leq i \leq m-1 $}, $ y(u_{i})=y(u_{i+1}) $ or $ y(l_{i})=y(l_{i+1}) $. So, we denote the edge of $ P $ that contains $ u_{i} $ by $ e(u_{i}) $ and the edge of $ P $ that contains $ l_{i} $ by $ e(l_{i}) $. Let $ E_{U}=\{e(u_{i})|1 \leq i \leq m\} $ and $ E_{L}=\{e(l_{i})|1 \leq i \leq m\} $, the sets of edges ordered from left to right. In a set $ E=E_u\cup E_l $ of horizontal edges of $ P $, $ e_{j} $ is named \textit{local maximum} if $ y(e_{j})> y(e_{j-1}) $ and $ y(e_{j})> y(e_{j+1}) $ and similarly, $ e_{k} $ is named \textit{local minimum} if $ y(e_{k})< y(e_{k-1}) $ and $ y(e_{k})< y(e_{k+1}) $. If edge $ \epsilon_1 \in E_U $ be a local maximum, then the internal angles of its both endpoints are equal to $ \frac{\pi}{2} $ and if $ \epsilon_2 $ be a local minimum, then the internal angles of its two endpoints are $ \frac{3\pi}{2} $. If edge $ \epsilon_3 \in E_L $ be a local minimum, then the internal angles of its both endpoints are equal to $ \frac{\pi}{2} $ and if $ \epsilon_4 $ be a local maximum, then the internal angles of its two endpoints are $ \frac{3\pi}{2} $. In this paper, if a horizontal edge $ \epsilon'\in E $ has two endpoints of angle $ \frac{\pi}{2} $, we call it \textit{tooth} and if a horizontal edge $ \epsilon\in E $ has two endpoints of angle $ \frac{3\pi}{2} $, we call it \textit{dent}. $ u_{m}$ (or $ l_{m}$) is named \textit{local maximum} if $ e( u_{m}) $ (or $ e(l_{m}) $) is local maximum, and $ u_{n}$ (or $ l_{n} $) is named \textit{local minimum} if $ e( u_{n}) $ (or $ e(l_{n} )$) is local minimum. Every rectangle $ R_i $ has height $ h_i=\@ifstar{\oldabs}{\oldabs*}{y(u_i)-y(l_i)} $ , so, in the set $ R $, $ R_{l} $ is named \textit{local maximum} if $ h_l>h_{l-1} $ and $ h_l>h_{l+1} $, and $ R_y $ is named \textit{local minimum} if $ h_y<h_{y-1} $ and $ h_y<h_{y+1} $. The rectangle that has only one neighbor is named \textit{source}, regularly, every rectangle has two neighbor except the first and last ones. So, the first and last rectangle is called source. Two axis-parallel segments $ l $ and $ l' $ are defined as \textit{weak visible} if an axis-parallel line segment could be drawn from some point of $ l $ to some point of $ l' $ that does not intersect $ P $. If polygon $ P $ is $ x $-monotone and also $ y $-monotone, then $ P $ is named \textit{orthoconvex} polygon. \begin{figure} \centering \includegraphics[width=\textwidth]{figure2} \caption{(a)An orthoconvex polygon $ P $ that has $ \sigma_{1} $ and $ \sigma_{2} $. Point $ \phi $ is the intersection of $ \sigma_{1} $ and $ \sigma_{2} $. (b)The decomposition of $ P $ into orthogonal fan polygons $ P_{1}$,$ P_{2}$,$ P_{3}$ and $ P_{4}$.} \label{fi:fig2} \end{figure} \begin{lemma} \label{le:lemma001} For any orthoconvex polygon $ P $, if there exists a horizontal line segment $ \sigma_{1} $ which is connecting the leftmost and the rightmost vertical edges of $ P $ such that $ \sigma_{1}\in P $ and there exists a vertical line segment $ \sigma_{2} $ which is connecting the upper and the lower horizontal edges of $ P $ such that $ \sigma_{2}\in P $, then $ P $ has a kernel. If guard $ g $ occurs in the kernel, every point in $ P $ is guarded by it. See figure~\ref{fi:fig2}. \end{lemma} \begin{proof} If $ \sigma_{1} $ connects the leftmost and the rightmost vertical edges of $ P $ and $ \sigma_{2} $ connects the upper and the lower horizontal edges of $ P $ then they have an intersection $\phi$ that is contained in $ P $. $\sigma_{1} $ and $ \sigma_{2} $ decompose $P$ into 4 sub-polygons $ P_{1}$,$ P_{2}$,$ P_{3}$ and $ P_{4}$. All of obtained sub-polygons are orthogonal fan polygons and $\phi$ is their core vertex, jointly. In every part, the entire sub-polygon is visible from $ \phi $ and also it is on the kernel. Therefore, $ \phi $ belongs to the kernel of $ P $ and if guard $ g $ occurs in the kernel, every point in $ P $ is guarded by it. \end{proof} Every polygon $ P $ divides the plane into 3 regions, an interior region bounded by the polygon, an exterior region containing all of the nearby and far away exterior points and boundary region containing all points on the boundary of polygon as denoted $int(P)$, $ext(P)$ and $ bound(P) $ so that $ P=int(P) \cup bound(P) $. If $ e $ be a line segment, $ int(e) $ is $ e $ without its two endpoints. If $ e $ be horizontal, then $ left(e) $ and $ right(e) $ are referred to left and right endpoints of $ e $, respectively. Also, if $ e $ be vertical, then $ top(e) $ and $ down(e) $ are referred to top and down endpoints of $ e $, respectively. The \textit{bounding box} of a polygon is the axis-aligned minimum area rectangle(box) within which all the points of polygon lie. If an $ x $-monotone orthogonal polygon and its bounding box has a horizontal edge in common, completely, the polygon is called \textit{histogram}, this common edge is called \textit{base}. If an orthoconvex polygon and its bounding box has a horizontal edge in common, completely, the polygon is called \textit{pyramid}, this common edge is called \textit{base}, too. \section{An Algorithm for Guarding Monotone Art Galleries} In the following, we present a linear-time exact algorithm for guarding orthogonal and $ x $-monotone polygons. Our algorithm uses a geometric approach instead of graph theoretical approach to obtain the result. Therefore, We can find the exact geometric locations of the point guards. \subsection{The Decomposition of an $ x $-monotone Orthogonal Polygon into the Balanced Ones} Let $ P $ be an orthogonal $ x $-monotone polygon with $ n $ vertices. Start from the leftmost vertical edge of $ P $, as denoted, $ \varepsilon $ , propagate a light beam in rectilinear (straight-line) path perpendicular to $ \varepsilon $ and therefore collinear with the $ X $ axis. All or part of this light beam passes through some rectangles of set $ R $(name this subset $ R_{\rho} $) and these rectangles together make a sub-polygon $ \rho $ of $ P $. For every $ u_{i}$ and $l_{i} $ belongs to $ \rho $, it is established that $ \min_{u_{i} \in \rho} (y(u_{i}))\geq \max_{l_{j} \in \rho}(y(l_{j})) $. So, there exists a horizontal line segment $ \sigma $ which is connecting the leftmost and rightmost vertical edges of $ \rho $ such that $ \sigma$ contains in $\rho $, completely. If a polygon like $ \rho $ has this property, we say the polygon is \textit{balanced}. If we want to guard a balanced $ x $-monotone polygon with minimum number of guards, it is possible to set all guards on its $ \sigma $. So, If we remove $ \rho $ from $ P $ and iterate the described operations, $ P $ is decomposed into several balanced $ x $-monotone polygon. Now, we describe a linear-time algorithm for decomposition $ P $ to the balanced sub-polygons. \begin{figure} \centering \includegraphics[width=\textwidth]{figure3} \caption{(a)Decomposition of monotone polygon into balanced sub-polygons and their vertical decomposition. Dark gray rectangles are cut. (b) A balanced sub-polygon and its align segment, cut rectangle part of sub-polygons has 3 disjoint parts.} \label{fi:fig3} \end{figure} \begin{algorithm}[] \KwData{an $ x $-monotone polygon with $ n $ vertices} \KwResult{the minimum number of balaced $ x $-monotone polygon} (1)Set $ min_u=u_1$ and $ max_l=l_1 $\; (2)\ForEach{rectangle $ R_{i} $ belongs to $ R $}{ (3)\If{ $ u_i>max_l $ or $ l_i<min_u $}{ remove $ R_1,\dots,R_{i-1} $ from $ R $\; refresh the index of $ R $ starting with $ 1 $\; go to 1\; } (4)Compute $min_u=\min(min_u,u_{i})$ and $max_l=\max(max_l,l_{i})$\; } \caption{Decomposition $ P $ into the balanced sub-polygons.} \label{al:algo1} \end{algorithm} If the condition in item 3 is satisfied, a balanced sub-polygon $ \rho $ is determined. So, we remove it from $ P $ and iterate algorithm for $ P-\rho $. We remove the rectangles belong to $ \rho $ from $ R $. We know the members of $ R $ are ordered from left to right and labeled from $ 1 $, After removing, we relabel the remained members from $ 1 $, again, to simplify the description of the algorithm. Certainly, the same processes will be occurred for $ U $ and $ L $. The number of iterations is equal to the cardinality of $ R $ (in the start). Therefore, the time complexity for the decomposition $ P $ into balanced polygons is linear. Remember the vertical decomposition of the polygon into rectangles, the last rectangle of each balanced sub-polygons (excepted the last sub-polygon) is named \textit{cut rectangle}. The cut rectangle of the sub-polygon $ \rho $ is denoted as $ C_{\rho} $. For an illustration see figure~\ref{fi:fig3}(a). Every cut rectangle has 3 parts which are obtained by extending horizontal edge of two adjacent rectangles, the one has intersection with the align segment is named int-part, two other ones are named middle-part and ext-part, see figure~\ref{fi:fig3}(b). The decomposition $ P $ into the minimum number of balanced sub-polygons is not unique. Assume that $ P $ is decomposed into the balanced sub-polygons $ \rho_{1},\rho_{2},\dots,\rho_{k} $, for every $ \rho_i $ and $ \rho_{i+1} $ we can replace them with $ \rho_i-C_{\rho_i} $ and $ \rho_{i+1}\cup C_{\rho_i}$. Bt this change, both of them remained balanced. But, we do this change only in certain circumstances that leads to simplicity. The last rectangle of $ \rho_i $ is $ C_{\rho_i}=R_{x+1}\in R $ and its previous rectangle be $ R_x \in R $. In every iteration of algorithm~\ref{al:algo1}, if $ R_x $ is not source and $ R_x $ is local minimum then do $ \rho_i=\rho_i-C_{\rho_i} $ and $ \rho_{i+1}=\rho_{i+1}\cup C_{\rho_i}$. Later, we will realize that the guarding of these two is more cost effective. Therefore, we modify algorithm~\ref{al:algo1} to algorithm~\ref{al:algo2}. \begin{algorithm}[] \KwData{an $ x $-monotone polygon with $ n $ vertices} \KwResult{the minimum number of balaced $ x $-monotone polygon} (1)Set $ min_u=u_1$ and $ max_l=l_1 $\; (2)\ForEach{rectangle $ R_{i} $ belongs to $ R $}{ (3)\If{ $ u_i>max_l $ or $ l_i<min_u $}{ \eIf{$ i-1\neq 1 $ and $ h_{i-1}>h_i $ and $ h_{i-1}>h_{i-2} $}{ remove $ R_1,\dots,R_{i-1} $ from $ R $\;}{ remove $ R_1,\dots,R_{i-2} $ from $ R $\; } refresh the index of $ R $ starting with $ 1 $\; go to 1\; } (4)Compute $min_u=\min(min_u,u_{i})$ and $max_l=\max(max_l,l_{i})$\; } \caption{The modified algorithm for decomposition $ P $ into the balanced sub-polygons.} \label{al:algo2} \end{algorithm} Every balanced polygon like $ \rho $ has a horizontal line-segment like $ \sigma $ which is connecting the leftmost and rightmost edges of $ \rho $ that is called \textit{align segment}. So, the entire $ \rho $ is visible from at least one point of $ \sigma $, i.e., $ \rho $ is weak visible from $ \sigma $. Assume that $ P $ is decomposed into the balanced sub-polygons $ \rho_{1},\rho_{2},\dots,\rho_{k} $ and $ \sigma_{1},\sigma_{2},\dots,\sigma_{k} $ be their align segments, respectively. If $ 1\leq i,j\leq k $ and $\@ifstar{\oldabs}{\oldabs*}{i-j}>1$, for every $ p\in \sigma_i $ and $ q\in \sigma_j $, $ p $ and $ q $ is not visible from each others. So, if $j=i+1$, the only visible points from one segment to another is the right endpoint of $ \sigma_i $ and the left endpoint of $ \sigma_j $ i.e. for every point $ p\in int(\sigma_{i})$, there is no point belongs to $int(\sigma_{j}) $ that is visible from $ p $. Due to this fact, if we optimally cover $ P $ so that all the guards are located on the align segments, guarding each of sub-polygons can be done independently i.e. the minimum number of guards for guarding the entire polygon is the sum of the minimum number of guards that are necessary for every sub-polygons. \begin{claim} There exists an optimal guard set $ G={g_1, g_2,\dots,g_{opt}} $ for a monotone orthogonal polygon $ P $ so that all guards are located on the align segments. \end{claim} To prove this claim, we present an algorithm in the next sections and prove that its output is optimal. \subsection{The Algorithm for Guarding the Balanced Sub-polygons} \label{ss:ss01} Given a balanced $ x $-monotone orthogonal $ P $ with $ n $ vertices, we present algorithm~\ref{al:algo3} to guard $ P $ using the minimum number of guards. After vertical decomposition, it is obtained the sets $ R $, $ U $, $ L $, $ E_L $ and $ E_U $ for the polygon $ P $. Because of being balanced, $ P $ has an align segment $ \sigma $ which is connecting the leftmost and rightmost edges of $ P $ and contained in it. \begin{defini} For a horizontal edge $ e $ of the polygon $ P $, the set of every point $p\in P$ which there is a point $ q \in e $ such that $ pq $ is a line segment normal to $ e $ and completely inside $ P $, is named \textit{orthogonal shadow} of $ e $, as denoted $ os_e $(for abbreviation). \end{defini} First, we find all tooth edges of the set $ E = E_L \cup E_U $ and call the obtained set as $ D $. For every $ d_i \in D $, we compute orthogonal shadow of $ d_i $ as $ os_{i} $. Let $ D=\{d_1, d_2,\dots,d_k\} $ and $ OS=\{os_1,os_2,\dots,os_k\} $, ordered from left to right by $ x $-coordination of their left vertical edges. \begin{lemma} \label{le:le02} Every tooth edge $ e_d $ can be guarded only with a guard that is located in its orthogonal shadow $ os_{e_d} $, not anywhere else. \end{lemma} \begin{proof} Proof by contradiction. Suppose that $ P $ is $ x $-monotone and the horizontal tooth edge $ e_t $ is guarded with $ g_t $ that is not located in $ os_{e_t} $, so, the point $ g_t $ is not in the $ x $-coordinate of any points on $ e_t $. Let the left and right endpoints of $ e_t $ be $ L_t $ and $ R_t $, respectively, and let $ x $-coordinate of $ g_t $ be less than $ x $-coordinate of $ L_t $ i.e. $ x(g_t)<x(L_t)$. The edge $ e_t $ is visible from $ g_t $, so, $ L_t $ and $ R_t $ are visible from $ g_t $. There is an axis-parallel rectangle spanned by the $ R_t $ and $ g_t $ is contained in $ P $. These two points do not have the same $ x $-coordinates and also $ y $-coordinates, $ 4 $ vertices of the rectangle are contained in $ P $ as $ R_t=(x(R_t),y(R_t)) $, $ A=(x(R_t),y(g_t)) $, $ B=(x(g_t),y(R_t)) $ and $ g_d=(x(g_t),y(g_t))$. So, the edge $ BR_t $ is contained in $ P $, too. It is impossible, because $ e_t \subset BR_t $ i.e. if an edge be a part of a line segment which is contained in the polygon, actually, it is not an edge. \end{proof} \begin{figure} \centering \includegraphics[width=\textwidth]{figure4} \caption{An Illustration of the algorithm, the bold edges are tooth and all bold parts of $ \sigma $ belong to $ SI $ that are the positions for guards.} \label{fi:fig4} \end{figure} If we want to guard the entire polygon $ P $, we must place guards so that there be at least one guard in every element of $ OS $. If tooth edge $ e_1 $ belongs to $ E_U $, another tooth edge $ e_2 $ belongs to $ E_L $ and $ os_{e_1}\cap os_{e_2}\neq \emptyset $ a guard on the intersection of them is sufficient to cover both of them. So, in the set $ OS $, if intersection of two members $ os_{e_1}$ and $ os_{e_2}$ be not empty then we remove them from $ OS $ and insert a new member as $ os_{e_1}\cap os_{e_2} $. Note that before the replacements, the intersection of every $ 3 $ members of $ OS $ is empty and after them the cardinality of $ OS $ decrease to $ \kappa \leq k $. We want to place all guards on the align segment $ \sigma $ and the orthogonal shadow of every tooth edge has intersection with $ \sigma $. Let line segment $ si_i=os_i \cap \sigma $ and set $ SI=\{si_1,si_2,\dots,si_\kappa\} $ s.t. $ (\kappa \leq k) $, ordered from left to right by $ x $-coordination of their left endpoint. The intersection of every $ 2 $ members of $ SI $ is empty. The set $ SI $ is the positions for placing guards, one guard is located on an arbitrary point of every member of $ SI $, see figure~\ref{fi:fig4} Using this structure enable us to find the positions for locating the minimum number of guards in the balanced sub-polygon $ P $ in the linear time relative to $ n $. In algorithm~\ref{al:algo3}, the set $ SI $ is the positions for the optimum guard set and the variable $ m $ is the cardinality of the optimum guard set. \begin{algorithm}[] \KwData{the horizontal edges of two chains of balanced $ x $-monotone polygon with $ n $ vertices ($ E_L $,$ E_U $)} \KwResult{the minimum number of guards ($ m $) and their positions ($ SI $)} Set $ m=0 $ and $ SI=\emptyset $\; \ForEach{horizontal edge $ e_{i} $ belongs to $ E_L $}{ \If{ Angles of $ left(e_i)$ and $right(e_i)$ are equal to $\frac{\pi}{2}$}{ $ A_i=(x(left(e_i)),y(\sigma) $\; $ B_i=(x(right(e_i)),y(\sigma) $\; Set segment $ si_i=A_iB_i $ and $ SI_L=SI_L \cup \{si_i $\}\; $ m++ $; } } \ForEach{horizontal edge $ e_{i} $ belongs to $ E_U $}{ \If{ Angles of $ left(e_i)$ and $right(e_i)$ are equal to $\frac{\pi}{2}$}{ $ A_i=(x(left(e_i)),y(\sigma) $\; $ B_i=(x(right(e_i)),y(\sigma) $\; Set segment $ si_i=A_iB_i $ and $ SI_U=SI_U \cup \{si_i\} $\; $ m++ $; } } Merge the sorted lists $ SI_L$ and $ SI_U$ as sorted list $ SI $\; \ForEach{horizontal segment $ si_{i} $ belongs to $ SI $}{ \If{$ si_i\cap si_{i+1}\neq \emptyset $}{ $ si_i=si_i\cap si_{i+1} $\; $ SI=SI-\{si_{i+1}\}$\; $ m-- $; } } \caption{Optimum guarding of a balanced polygon $ P $.} \label{al:algo3} \end{algorithm} \\The positions of all guards are in the set $ SI $ and every elements of $ SI $ is a subset of align segment $ \sigma $, so, all guards are located on align segment $ \sigma $. The time complexity of the algorithm clearly is linear time corresponding to the cardinality of set $ E=E_L\cup E_U $. \begin{lemma} \label{le:lemma3} The minimum number of guards is equal to $ m $ that is obtained by algorithm~\ref{al:algo3} for guarding a balanced monotone orthogonal polygon $ P $. \end{lemma} \begin{proof} Suppose that $ m $ guards is sufficient to cover the entire polygon $ P $, using lemma~\ref{le:le02} prove that this number of guards necessary even for guarding the dent edges of $ P $. Every line segment $ si_i\in SI $ is a subset of a r-star sub-polygon i.e. if we decompose $ P $ into r-star parts(sub-polygons) then the kernels of every r-star sub-polygons has at least one point in the elements of $ SI $, so the entire $ P $ is covered by these $ m $ guards and their positions. \end{proof} \subsection{Time Complexity of Algorithm} In this subsection, we analyze the algorithm and describe how it can be implemented in linear time. To compute the optimal solution for guarding $ P $, we need to solve subproblems. To solve the subproblems of $ P $, we need to decompose $ P $ vertically into the set of rectangles $ R $ described in section~\ref{ss:ss1}. Therefore, we obtain the sets $ E $, $ U $ and $ L $. We show that this decomposition is possible in $ O(n) $-time. After that, we use algorithm~\ref{al:algo2} that is processing-able in $ O(n) $-time, we explain it before. The total of vertices of the all obtained balanced orthogonal is $ O(n) $, so, in algorithm~\ref{al:algo3}, finding the minimum number of guards for all balanced sub-polygons is possible in $ O(n) $-time. Therefore, all computations handle in $ O(n) $-time. Finally, $ m $ is returned as the optimal solution for the problem on $ P $. Therefore, we have proved the main result of this section: \begin{theorem} \label{th:th01} There exists a purely geometric algorithm that can find the minimum number of guards for an orthogonal and $ x $-monotone polygon with $ n $ vertices, with orthogonal visibility in $ O(n) $ time. \end{theorem} \section{Hidden Guarding of Histogram Galleries} Let $ P $ be a histogram polygon with $ n $ vertices, we want to find minimum number of guards such that every point in $ P $ is visible from some guards under the constraint that no two guards may see each other. Every histogram polygon is a balanced monotone one and has an edge that connecting the leftmost and rightmost vertical edges of it as called \textit{base}. So, the base edge is an align segment for histogram polygon $ P $. In every monotone polygon, the number of tooth edges is two more than the number of dent edges i.e the number of tooth edges belong to $ E_L $ (or $ E_U $) is one more than the number of dent edges belong to it. In this section, we present a linear-time exact algorithm for hidden guarding histogram polygons. Our algorithm uses a geometric approach instead of graph theoretical approach to obtain the result. Therefore, We find the exact geometric locations of the point guards. \subsection{The Decomposition of an histogram Polygon into Pyramid Polygons} Given a histogram polygon $ P $ with $ n $ vertices. Extend every dent edge of $ P $, exclusively from its right endpoint until intersect the boundary. Using this strategy, decompose $ P $ into several sub-polygons. All the obtained sub-polygons are orthoconvex and absolutely pyramid polygons. If the number of dent edges of $ P $ be $ m $, then the number of obtained pyramid sub-polygons is $ m+1 $, exactly. The base edges of all pyramid polygons lie on the extended dent edges except the rightmost pyramid that its base edge is in common with the base of $ P $. For an illustration see figure~\ref{fi:fig5}(a). \begin{figure} \centering \includegraphics[width=\textwidth]{figure5} \caption{(a)The decomposition of histogram polygon into pyramid polygons. (b) The basis rectangle of a pyramid polygon is shown as gray.} \label{fi:fig5} \end{figure} \begin{defini} In the pyramid polygon $ P_0 $, the maximum area rectangle $ R_0\subset P_0 $ that one of its edge is the base edge of pyramid, is named \textit{basis rectangle}. See figure~\ref{fi:fig5}(b). \end{defini} Suppose that $ \Pi=\{\pi_0, \pi_1,\dots, \pi_m\} $ is the set of all obtained pyramid sub-polygons ordered from left to right, according to the $ x $-coordinate of their leftmost vertical edges. Also, suppose that $ \beta_i $ and $ \lambda_i $ are the base edge and basis rectangle of sub-polygon $ \pi_i $, {\footnotesize for $ 0\leq i \leq m $}, respectively. In the pyramid $ \pi_i $, the base edge $ \beta_i $ is a tooth edge, except $ \beta_i $, there exists another tooth edge $ t_i $ on the upper chain of $ \pi_i $. For every pyramid sub-polygon, we compute the orthogonal shadow of $ t_i $ and intersection between $ \lambda_i $ and $ os_{t_i} $ as the position for placing a guard on it i.e. we hidden guarding an histogram polygon with locating one guard in one of the interior points of every $ si_i=\lambda_i \cap os_{t_i} $ as named \textit{shadow intersection} of $ \pi_i $. \begin{lemma} The interior points of two different shadow intersection are invisible from each other. In the other words, every two points $ p_1 \in int(si_i) $ and $ p_2\in int(si_j) $ that $ i\neq j $, is not visible from each other. \end{lemma} \begin{proof} After the described decomposition, assume that $ \Pi=\{\pi_0, \pi_1,\dots, \pi_m\} $ be the set of obtained sub-polygons ordered from left to right, according to the $ x $-coordinate of their leftmost vertical edge. Each of them has a base edge and assume that $ B=\{\beta_0, \beta_1,\dots, \beta_m\} $ be the set of base edges of $ \Pi $ ordered corresponding to order of polygons they belong to. Also, let $ \Lambda=\{\lambda_0, \lambda_1,\dots, \lambda_m\} $ and $ SI=\{si_0, si_2,\dots, si_m\}$ be the sets of basis rectangles and shadow intersection areas of $ \Pi $ in the same expressed order. And assume that $ D=\{d_1, d_2,\dots,d_m\} $ be the set of dent edges of $ P $ ordered from left to right according to $ x $-coordinates of their left endpoints. Remember that $ \beta_x $ is obtained after extending the dent edge $ d_x $, therefore all the interior points of $ si_x $, $ \lambda_x $ and even $ \pi_x $ is higher than $ d_x $, {\footnotesize for every $ 0\leq x\leq m $}. Proof by contradiction. Suppose that there exist two points $ p_1 \in int(si_i) $ and $ p_2\in int(si_j) $ that $ i<j $, is visible from each other. So, there is an axis-aligned rectangle $ R $ spanned by $ p_1 $ and $ p_2 $ contained in $ int(P) $, then $ R\cap(P-int(P)) = \emptyset $, so, $ R\cap d_j=\emptyset $, too. Since that $ d_j $ belongs to upper chain, $ d_j $ is higher than all points in $ R $ and it is higher than $ p_2 $. So, there \textbf{is} a point belongs $ si_j $ that is higher than $ d_j $ and this is a contradiction. \end{proof} \subsection{Analyze of Algorithm} In this subsection, we present pseudo-code of the algorithm that is described in the previous subsection as algorithm~\ref{al:algo4}. We analyze the algorithm, prove that it find the optimum number of hidden guards and explain how it can be implemented in linear time. \begin{algorithm}[] \KwData{the horizontal and vertical edges of upper chain of histogram polygon with $ n $ vertices and base edge $ b $ (the set of horizontal edges $ E_H $, the set vertical edges $ E_V $)} \KwResult{the minimum number of hidden guards ($ m $) and their positions ($ SI $)} Set $ m=0 $ and $ SI=\emptyset $\; Set $ \varepsilon = $ minimum length of $ E_V $\; Set $ y_1=y(b) $ and $ y_2=y(b)+\varepsilon $\; \ForEach{horizontal edge $ e_{i} $ belongs to $ E_H $}{ \If{ Angles of $ left(e_i)$ and $right(e_i)$ are equal to $\frac{3\pi}{2}$}{ $ y_1=y(e_i) $\; $ y_2=y(e_i)+\varepsilon $\; } \If{ Angles of $ left(e_i)$ and $right(e_i)$ are equal to $\frac{\pi}{2}$}{ $ A_i=(x(left(e_i)),y_1) $\; $ B_i=(x(right(e_i)),y_2) $\; Set $ si_i= $ the rectangle spanned by two points $ A_i $ and $ B_i $\; Set $SI=SI \cup \{si_i $\}\; $ m++ $; } } \caption{Optimum hidden guarding of a histogram polygon $ P $.} \label{al:algo4} \end{algorithm} The interior points of the members of $ SI $ are invisible from each other, so, if we put a guard over each rectangle, then all the pyramid sub-polygons are guarded and also the entire histogram polygon is covered. \begin{lemma} The minimum number of hidden guards is equal to $ m $ that is obtained by algorithm~\ref{al:algo4} for guarding a histogram polygon $ P $. \end{lemma} \begin{proof} The number of guards that is obtained using algorithm~\ref{al:algo4} for finding the minimum hidden guard set is equal to the number of guards that is obtained using algorithm~\ref{al:algo3} for the minimum regular guard set, therefore we proved that it is optimum in lemma~\ref{le:lemma3}. Every rectangle $ si_i\in SI $ is a subset of kernel of a pyramid sub-polygon that is r-star i.e. if we decompose $ P $ into r-star parts(sub-polygons) then the kernels of every r-star sub-polygons has at least one point in the elements of $ SI $, so the entire $ P $ is covered by these $ m $ guards and their positions. \end{proof} The time complexity of the algorithm is linear time corresponding to the size of the polygon, clearly. To solve the problem, we need to decompose $ P $ vertically into the set of rectangles $ R $ described in section~\ref{ss:ss1}. Therefore, we obtain the sets $ E_H $ and $ E_V $. This decomposition is possible in $ O(n) $-time. After that, we use algorithm~\ref{al:algo3} that is processing-able in $ O(n) $-time. So, all computations handle in $ O(n) $-time. Finally, $ m $ is returned as the optimal solution for the problem on $ P $. We have proved the main result of this section: \begin{theorem} There exists a purely geometric algorithm that can find the minimum number of hidden guards for a histogram polygon with $ n $ vertices, with orthogonal visibility in $ O(n) $ time. \end{theorem} \section{Conclusion} We studied the problem of finding the minimum number of hidden guards which is under orthogonal visibility. This new version is named \textit{hidden art gallery problem}. The total target in the hidden art gallery problem is finding the optimum hidden guard set $ G $ which is a set of point guards in polygon $ P $ that all points of the $ P $ are visible from at least one guard in $ G $ under the constraint that no two guards may see each other. We present an exact algorithm for finding the hidden guard set for histogram galleries. We solved this problem in the linear time according to $ n $ where $ n $ is the number of sides of histogram polygon. the space complexity of our algorithm is $ O(n) $, too. Many of the algorithms presented in this field are based on graph theory, but our proposed algorithm is based on geometric approach. This approach can lead to improved performance and efficiency in algorithms. For this reason, we also provided a purely geometric algorithm for the orthogonal art gallery problem (not hidden) where the galleries are monotone. We are aware that this problem has already been solved in linear time and our algorithm is linear-time, too. This new approach helped us solve the hidden art gallery problem more easily. Actually, the time complexity of the hidden orthogonal art gallery problem even for monotone polygon is still open. For the future works, we want to try to solve this problem for every simple orthogonal polygon with/without barriers. Both time and space complexity of our presented algorithm is order of $ O(n) $ and it is the best for this new version of the problem. \bibliographystyle{plain}

https://arxiv.org/abs/1610.03411

Remarks on the $Γ$-regularization of Non-convex and Non-semi-continuous Functions on Topological Vector Spaces

We show that the minimization problem of any non-convex and non-lower semi-continuous function on a compact convex subset of a locally convex real topological vector space can be studied via an associated convex and lower semi-continuous function $\Gamma \left( h\right) $. This observation uses the notion of $\Gamma $-regularization as a key ingredient. As an application we obtain, on any locally convex real space, a generalization of the Lanford III-Robinson theorem which has only been proven for separable real Banach spaces. The latter is a characterization of subdifferentials of convex continuous functions.

\section*{Abstract (Not appropriate in this style!)}% \else \small \begin{center}{\bf Abstract\vspace{-.5em}\vspace{\z@}}\end{center}% \quotation \fi }% }{% }% \@ifundefined{endabstract}{\def\endabstract {\if@twocolumn\else\endquotation\fi}}{}% \@ifundefined{maketitle}{\def\maketitle#1{}}{}% \@ifundefined{affiliation}{\def\affiliation#1{}}{}% \@ifundefined{proof}{\def\proof{\noindent{\bfseries Proof. }}}{}% \@ifundefined{endproof}{\def\endproof{\mbox{\ \rule{.1in}{.1in}}}}{}% \@ifundefined{newfield}{\def\newfield#1#2{}}{}% \@ifundefined{chapter}{\def\chapter#1{\par(Chapter head:)#1\par }% \newcount\c@chapter}{}% \@ifundefined{part}{\def\part#1{\par(Part head:)#1\par }}{}% \@ifundefined{section}{\def\section#1{\par(Section head:)#1\par }}{}% \@ifundefined{subsection}{\def\subsection#1% {\par(Subsection head:)#1\par }}{}% \@ifundefined{subsubsection}{\def\subsubsection#1% {\par(Subsubsection head:)#1\par }}{}% \@ifundefined{paragraph}{\def\paragraph#1% {\par(Subsubsubsection head:)#1\par }}{}% \@ifundefined{subparagraph}{\def\subparagraph#1% {\par(Subsubsubsubsection head:)#1\par }}{}% \@ifundefined{therefore}{\def\therefore{}}{}% \@ifundefined{backepsilon}{\def\backepsilon{}}{}% \@ifundefined{yen}{\def\yen{\hbox{\rm\rlap=Y}}}{}% \@ifundefined{registered}{% \def\registered{\relax\ifmmode{}\r@gistered \else$\m@th\r@gistered$\fi}% \def\r@gistered{^{\ooalign {\hfil\raise.07ex\hbox{$\scriptstyle\rm\RIfM@\expandafter\text@\else\expandafter\mbox\fi{R}$}\hfil\crcr \mathhexbox20D}}}}{}% \@ifundefined{Eth}{\def\Eth{}}{}% \@ifundefined{eth}{\def\eth{}}{}% \@ifundefined{Thorn}{\def\Thorn{}}{}% \@ifundefined{thorn}{\def\thorn{}}{}% \def\TEXTsymbol#1{\mbox{$#1$}}% \@ifundefined{degree}{\def\degree{{}^{\circ}}}{}% \newdimen\theight \def\Column{% \vadjust{\setbox\z@=\hbox{\scriptsize\quad\quad tcol}% \theight=\ht\z@\advance\theight by \dp\z@\advance\theight by \lineskip \kern -\theight \vbox to \theight{% \rightline{\rlap{\box\z@}}% \vss }% }% }% \def\qed{% \ifhmode\unskip\nobreak\fi\ifmmode\ifinner\else\hskip5\p@\fi\fi \hbox{\hskip5\p@\vrule width4\p@ height6\p@ depth1.5\p@\hskip\p@}% }% \def\cents{\hbox{\rm\rlap/c}}% \def\miss{\hbox{\vrule height2\p@ width 2\p@ depth\z@}}% \def\vvert{\Vert \def\tcol#1{{\baselineskip=6\p@ \vcenter{#1}} \Column} % \def\dB{\hbox{{}} \def\mB#1{\hbox{$#1$} \def\nB#1{\hbox{#1} \def\note{$^{\dag}}% \defLaTeX2e{LaTeX2e} \def\chkcompat{% \if@compatibility \else \usepackage{latexsym} \fi } \ifx\fmtnameLaTeX2e \DeclareOldFontCommand{\rm}{\normalfont\rmfamily}{\mathrm} \DeclareOldFontCommand{\sf}{\normalfont\sffamily}{\mathsf} \DeclareOldFontCommand{\tt}{\normalfont\ttfamily}{\mathtt} \DeclareOldFontCommand{\bf}{\normalfont\bfseries}{\mathbf} \DeclareOldFontCommand{\it}{\normalfont\itshape}{\mathit} \DeclareOldFontCommand{\sl}{\normalfont\slshape}{\@nomath\sl} \DeclareOldFontCommand{\sc}{\normalfont\scshape}{\@nomath\sc} \chkcompat \fi \def\alpha{{\Greekmath 010B}}% \def\beta{{\Greekmath 010C}}% \def\gamma{{\Greekmath 010D}}% \def\delta{{\Greekmath 010E}}% \def\epsilon{{\Greekmath 010F}}% \def\zeta{{\Greekmath 0110}}% \def\eta{{\Greekmath 0111}}% \def\theta{{\Greekmath 0112}}% \def\iota{{\Greekmath 0113}}% \def\kappa{{\Greekmath 0114}}% \def\lambda{{\Greekmath 0115}}% \def\mu{{\Greekmath 0116}}% \def\nu{{\Greekmath 0117}}% \def\xi{{\Greekmath 0118}}% \def\pi{{\Greekmath 0119}}% \def\rho{{\Greekmath 011A}}% \def\sigma{{\Greekmath 011B}}% \def\tau{{\Greekmath 011C}}% \def\upsilon{{\Greekmath 011D}}% \def\phi{{\Greekmath 011E}}% \def\chi{{\Greekmath 011F}}% \def\psi{{\Greekmath 0120}}% \def\omega{{\Greekmath 0121}}% \def\varepsilon{{\Greekmath 0122}}% \def\vartheta{{\Greekmath 0123}}% \def\varpi{{\Greekmath 0124}}% \def\varrho{{\Greekmath 0125}}% \def\varsigma{{\Greekmath 0126}}% \def\varphi{{\Greekmath 0127}}% \def{\Greekmath 0272}{{\Greekmath 0272}} \def\FindBoldGroup{% {\setbox0=\hbox{$\mathbf{x\global\edef\theboldgroup{\the\mathgroup}}$}}% } \def\Greekmath#1#2#3#4{% \if@compatibility \ifnum\mathgroup=\symbold \mathchoice{\mbox{\boldmath$\displaystyle\mathchar"#1#2#3#4$}}% {\mbox{\boldmath$\textstyle\mathchar"#1#2#3#4$}}% {\mbox{\boldmath$\scriptstyle\mathchar"#1#2#3#4$}}% {\mbox{\boldmath$\scriptscriptstyle\mathchar"#1#2#3#4$}}% \else \mathchar"#1#2#3#4% \fi \else \FindBoldGroup \ifnum\mathgroup=\theboldgroup \mathchoice{\mbox{\boldmath$\displaystyle\mathchar"#1#2#3#4$}}% {\mbox{\boldmath$\textstyle\mathchar"#1#2#3#4$}}% {\mbox{\boldmath$\scriptstyle\mathchar"#1#2#3#4$}}% {\mbox{\boldmath$\scriptscriptstyle\mathchar"#1#2#3#4$}}% \else \mathchar"#1#2#3#4% \fi \fi} \newif\ifGreekBold \GreekBoldfalse \let\SAVEPBF=\pbf \def\pbf{\GreekBoldtrue\SAVEPBF}% \@ifundefined{theorem}{\newtheorem{theorem}{Theorem}}{} \@ifundefined{lemma}{\newtheorem{lemma}[theorem]{Lemma}}{} \@ifundefined{corollary}{\newtheorem{corollary}[theorem]{Corollary}}{} \@ifundefined{conjecture}{\newtheorem{conjecture}[theorem]{Conjecture}}{} \@ifundefined{proposition}{\newtheorem{proposition}[theorem]{Proposition}}{} \@ifundefined{axiom}{\newtheorem{axiom}{Axiom}}{} \@ifundefined{remark}{\newtheorem{remark}{Remark}}{} \@ifundefined{example}{\newtheorem{example}{Example}}{} \@ifundefined{exercise}{\newtheorem{exercise}{Exercise}}{} \@ifundefined{definition}{\newtheorem{definition}{Definition}}{} \@ifundefined{mathletters}{% \newcounter{equationnumber} \def\mathletters{% \addtocounter{equation}{1} \edef\@currentlabel{\arabic{equation}}% \setcounter{equationnumber}{\c@equation} \setcounter{equation}{0}% \edef\arabic{equation}{\@currentlabel\noexpand\alph{equation}}% } \def\endmathletters{% \setcounter{equation}{\value{equationnumber}}% } }{} \@ifundefined{BibTeX}{% \def\BibTeX{{\rm B\kern-.05em{\sc i\kern-.025em b}\kern-.08em T\kern-.1667em\lower.7ex\hbox{E}\kern-.125emX}}}{}% \@ifundefined{AmS}% {\def\AmS{{\protect\usefont{OMS}{cmsy}{m}{n}% A\kern-.1667em\lower.5ex\hbox{M}\kern-.125emS}}}{}% \@ifundefined{AmSTeX}{\def\AmSTeX{\protect\AmS-\protect\TeX\@}}{}% \ifx\ds@amstex\relax \message{amstex already loaded}\makeatother\endinpu \else \@ifpackageloaded{amstex}% {\message{amstex already loaded}\makeatother\endinput} {} \@ifpackageloaded{amsgen}% {\message{amsgen already loaded}\makeatother\endinput} {} \fi \let\DOTSI\relax \def\RIfM@{\relax\ifmmode}% \def\FN@{\futurelet\next}% \newcount\intno@ \def\iint{\DOTSI\intno@\tw@\FN@\ints@}% \def\iiint{\DOTSI\intno@\thr@@\FN@\ints@}% \def\iiiint{\DOTSI\intno@4 \FN@\ints@}% \def\idotsint{\DOTSI\intno@\z@\FN@\ints@}% \def\ints@{\findlimits@\ints@@}% \newif\iflimtoken@ \newif\iflimits@ \def\findlimits@{\limtoken@true\ifx\next\limits\limits@true \else\ifx\next\nolimits\limits@false\else \limtoken@false\ifx\ilimits@\nolimits\limits@false\else \ifinner\limits@false\else\limits@true\fi\fi\fi\fi}% \def\multint@{\int\ifnum\intno@=\z@\intdots@ \else\intkern@\fi \ifnum\intno@>\tw@\int\intkern@\fi \ifnum\intno@>\thr@@\int\intkern@\fi \int \def\multintlimits@{\intop\ifnum\intno@=\z@\intdots@\else\intkern@\fi \ifnum\intno@>\tw@\intop\intkern@\fi \ifnum\intno@>\thr@@\intop\intkern@\fi\intop}% \def\intic@{% \mathchoice{\hskip.5em}{\hskip.4em}{\hskip.4em}{\hskip.4em}}% \def\negintic@{\mathchoice {\hskip-.5em}{\hskip-.4em}{\hskip-.4em}{\hskip-.4em}}% \def\ints@@{\iflimtoken@ \def\ints@@@{\iflimits@\negintic@ \mathop{\intic@\multintlimits@}\limits \else\multint@\nolimits\fi \eat@ \else \def\ints@@@{\iflimits@\negintic@ \mathop{\intic@\multintlimits@}\limits\else \multint@\nolimits\fi}\fi\ints@@@}% \def\intkern@{\mathchoice{\!\!\!}{\!\!}{\!\!}{\!\!}}% \def\plaincdots@{\mathinner{\cdotp\cdotp\cdotp}}% \def\intdots@{\mathchoice{\plaincdots@}% {{\cdotp}\mkern1.5mu{\cdotp}\mkern1.5mu{\cdotp}}% {{\cdotp}\mkern1mu{\cdotp}\mkern1mu{\cdotp}}% {{\cdotp}\mkern1mu{\cdotp}\mkern1mu{\cdotp}}}% \def\RIfM@{\relax\protect\ifmmode} \def\RIfM@\expandafter\text@\else\expandafter\mbox\fi{\RIfM@\expandafter\RIfM@\expandafter\text@\else\expandafter\mbox\fi@\else\expandafter\mbox\fi} \let\nfss@text\RIfM@\expandafter\text@\else\expandafter\mbox\fi \def\RIfM@\expandafter\text@\else\expandafter\mbox\fi@#1{\mathchoice {\textdef@\displaystyle\f@size{#1}}% {\textdef@\textstyle\tf@size{\firstchoice@false #1}}% {\textdef@\textstyle\sf@size{\firstchoice@false #1}}% {\textdef@\textstyle \ssf@size{\firstchoice@false #1}}% \glb@settings} \def\textdef@#1#2#3{\hbox{{% \everymath{#1}% \let\f@size#2\selectfont #3}}} \newif\iffirstchoice@ \firstchoice@true \def\Let@{\relax\iffalse{\fi\let\\=\cr\iffalse}\fi}% \def\vspace@{\def\vspace##1{\crcr\noalign{\vskip##1\relax}}}% \def\multilimits@{\bgroup\vspace@\Let@ \baselineskip\fontdimen10 \scriptfont\tw@ \advance\baselineskip\fontdimen12 \scriptfont\tw@ \lineskip\thr@@\fontdimen8 \scriptfont\thr@@ \lineskiplimit\lineskip \vbox\bgroup\ialign\bgroup\hfil$\m@th\scriptstyle{##}$\hfil\crcr}% \def\Sb{_\multilimits@}% \def\endSb{\crcr\egroup\egroup\egroup}% \def\Sp{^\multilimits@}% \let\endSp\endSb \newdimen\ex@ \ex@.2326ex \def\rightarrowfill@#1{$#1\m@th\mathord-\mkern-6mu\cleaders \hbox{$#1\mkern-2mu\mathord-\mkern-2mu$}\hfill \mkern-6mu\mathord\rightarrow$}% \def\leftarrowfill@#1{$#1\m@th\mathord\leftarrow\mkern-6mu\cleaders \hbox{$#1\mkern-2mu\mathord-\mkern-2mu$}\hfill\mkern-6mu\mathord-$}% \def\leftrightarrowfill@#1{$#1\m@th\mathord\leftarrow \mkern-6mu\cleaders \hbox{$#1\mkern-2mu\mathord-\mkern-2mu$}\hfill \mkern-6mu\mathord\rightarrow$}% \def\overrightarrow{\mathpalette\overrightarrow@}% \def\overrightarrow@#1#2{\vbox{\ialign{##\crcr\rightarrowfill@#1\crcr \noalign{\kern-\ex@\nointerlineskip}$\m@th\hfil#1#2\hfil$\crcr}}}% \let\overarrow\overrightarrow \def\overleftarrow{\mathpalette\overleftarrow@}% \def\overleftarrow@#1#2{\vbox{\ialign{##\crcr\leftarrowfill@#1\crcr \noalign{\kern-\ex@\nointerlineskip}$\m@th\hfil#1#2\hfil$\crcr}}}% \def\overleftrightarrow{\mathpalette\overleftrightarrow@}% \def\overleftrightarrow@#1#2{\vbox{\ialign{##\crcr \leftrightarrowfill@#1\crcr \noalign{\kern-\ex@\nointerlineskip}$\m@th\hfil#1#2\hfil$\crcr}}}% \def\underrightarrow{\mathpalette\underrightarrow@}% \def\underrightarrow@#1#2{\vtop{\ialign{##\crcr$\m@th\hfil#1#2\hfil $\crcr\noalign{\nointerlineskip}\rightarrowfill@#1\crcr}}}% \let\underarrow\underrightarrow \def\underleftarrow{\mathpalette\underleftarrow@}% \def\underleftarrow@#1#2{\vtop{\ialign{##\crcr$\m@th\hfil#1#2\hfil $\crcr\noalign{\nointerlineskip}\leftarrowfill@#1\crcr}}}% \def\underleftrightarrow{\mathpalette\underleftrightarrow@}% \def\underleftrightarrow@#1#2{\vtop{\ialign{##\crcr$\m@th \hfil#1#2\hfil$\crcr \noalign{\nointerlineskip}\leftrightarrowfill@#1\crcr}}}% \def\qopnamewl@#1{\mathop{\operator@font#1}\nlimits@} \let\nlimits@\displaylimits \def\setboxz@h{\setbox\z@\hbox} \def\varlim@#1#2{\mathop{\vtop{\ialign{##\crcr \hfil$#1\m@th\operator@font lim$\hfil\crcr \noalign{\nointerlineskip}#2#1\crcr \noalign{\nointerlineskip\kern-\ex@}\crcr}}}} \def\rightarrowfill@#1{\m@th\setboxz@h{$#1-$}\ht\z@\z@ $#1\copy\z@\mkern-6mu\cleaders \hbox{$#1\mkern-2mu\box\z@\mkern-2mu$}\hfill \mkern-6mu\mathord\rightarrow$} \def\leftarrowfill@#1{\m@th\setboxz@h{$#1-$}\ht\z@\z@ $#1\mathord\leftarrow\mkern-6mu\cleaders \hbox{$#1\mkern-2mu\copy\z@\mkern-2mu$}\hfill \mkern-6mu\box\z@$} \def\qopnamewl@{proj\,lim}{\qopnamewl@{proj\,lim}} \def\qopnamewl@{inj\,lim}{\qopnamewl@{inj\,lim}} \def\mathpalette\varlim@\rightarrowfill@{\mathpalette\varlim@\rightarrowfill@} \def\mathpalette\varlim@\leftarrowfill@{\mathpalette\varlim@\leftarrowfill@} \def\mathpalette\varliminf@{}{\mathpalette\mathpalette\varliminf@{}@{}} \def\mathpalette\varliminf@{}@#1{\mathop{\underline{\vrule\@depth.2\ex@\@width\z@ \hbox{$#1\m@th\operator@font lim$}}}} \def\mathpalette\varlimsup@{}{\mathpalette\mathpalette\varlimsup@{}@{}} \def\mathpalette\varlimsup@{}@#1{\mathop{\overline {\hbox{$#1\m@th\operator@font lim$}}}} \def\tfrac#1#2{{\textstyle {#1 \over #2}}}% \def\dfrac#1#2{{\displaystyle {#1 \over #2}}}% \def\binom#1#2{{#1 \choose #2}}% \def\tbinom#1#2{{\textstyle {#1 \choose #2}}}% \def\dbinom#1#2{{\displaystyle {#1 \choose #2}}}% \def\QATOP#1#2{{#1 \atop #2}}% \def\QTATOP#1#2{{\textstyle {#1 \atop #2}}}% \def\QDATOP#1#2{{\displaystyle {#1 \atop #2}}}% \def\QABOVE#1#2#3{{#2 \above#1 #3}}% \def\QTABOVE#1#2#3{{\textstyle {#2 \above#1 #3}}}% \def\QDABOVE#1#2#3{{\displaystyle {#2 \above#1 #3}}}% \def\QOVERD#1#2#3#4{{#3 \overwithdelims#1#2 #4}}% \def\QTOVERD#1#2#3#4{{\textstyle {#3 \overwithdelims#1#2 #4}}}% \def\QDOVERD#1#2#3#4{{\displaystyle {#3 \overwithdelims#1#2 #4}}}% \def\QATOPD#1#2#3#4{{#3 \atopwithdelims#1#2 #4}}% \def\QTATOPD#1#2#3#4{{\textstyle {#3 \atopwithdelims#1#2 #4}}}% \def\QDATOPD#1#2#3#4{{\displaystyle {#3 \atopwithdelims#1#2 #4}}}% \def\QABOVED#1#2#3#4#5{{#4 \abovewithdelims#1#2#3 #5}}% \def\QTABOVED#1#2#3#4#5{{\textstyle {#4 \abovewithdelims#1#2#3 #5}}}% \def\QDABOVED#1#2#3#4#5{{\displaystyle {#4 \abovewithdelims#1#2#3 #5}}}% \def\tint{\mathop{\textstyle \int}}% \def\tiint{\mathop{\textstyle \iint }}% \def\tiiint{\mathop{\textstyle \iiint }}% \def\tiiiint{\mathop{\textstyle \iiiint }}% \def\tidotsint{\mathop{\textstyle \idotsint }}% \def\toint{\mathop{\textstyle \oint}}% \def\tsum{\mathop{\textstyle \sum }}% \def\tprod{\mathop{\textstyle \prod }}% \def\tbigcap{\mathop{\textstyle \bigcap }}% \def\tbigwedge{\mathop{\textstyle \bigwedge }}% \def\tbigoplus{\mathop{\textstyle \bigoplus }}% \def\tbigodot{\mathop{\textstyle \bigodot }}% \def\tbigsqcup{\mathop{\textstyle \bigsqcup }}% \def\tcoprod{\mathop{\textstyle \coprod }}% \def\tbigcup{\mathop{\textstyle \bigcup }}% \def\tbigvee{\mathop{\textstyle \bigvee }}% \def\tbigotimes{\mathop{\textstyle \bigotimes }}% \def\tbiguplus{\mathop{\textstyle \biguplus }}% \def\dint{\mathop{\displaystyle \int}}% \def\diint{\mathop{\displaystyle \iint }}% \def\diiint{\mathop{\displaystyle \iiint }}% \def\diiiint{\mathop{\displaystyle \iiiint }}% \def\didotsint{\mathop{\displaystyle \idotsint }}% \def\doint{\mathop{\displaystyle \oint}}% \def\dsum{\mathop{\displaystyle \sum }}% \def\dprod{\mathop{\displaystyle \prod }}% \def\dbigcap{\mathop{\displaystyle \bigcap }}% \def\dbigwedge{\mathop{\displaystyle \bigwedge }}% \def\dbigoplus{\mathop{\displaystyle \bigoplus }}% \def\dbigodot{\mathop{\displaystyle \bigodot }}% \def\dbigsqcup{\mathop{\displaystyle \bigsqcup }}% \def\dcoprod{\mathop{\displaystyle \coprod }}% \def\dbigcup{\mathop{\displaystyle \bigcup }}% \def\dbigvee{\mathop{\displaystyle \bigvee }}% \def\dbigotimes{\mathop{\displaystyle \bigotimes }}% \def\dbiguplus{\mathop{\displaystyle \biguplus }}% \def\stackunder#1#2{\mathrel{\mathop{#2}\limits_{#1}}}% \begingroup \catcode `|=0 \catcode `[= 1 \catcode`]=2 \catcode `\{=12 \catcode `\}=12 \catcode`\\=12 |gdef|@alignverbatim#1\end{align}[#1|end[align]] |gdef|@salignverbatim#1\end{align*}[#1|end[align*]] |gdef|@alignatverbatim#1\end{alignat}[#1|end[alignat]] |gdef|@salignatverbatim#1\end{alignat*}[#1|end[alignat*]] |gdef|@xalignatverbatim#1\end{xalignat}[#1|end[xalignat]] |gdef|@sxalignatverbatim#1\end{xalignat*}[#1|end[xalignat*]] |gdef|@gatherverbatim#1\end{gather}[#1|end[gather]] |gdef|@sgatherverbatim#1\end{gather*}[#1|end[gather*]] |gdef|@gatherverbatim#1\end{gather}[#1|end[gather]] |gdef|@sgatherverbatim#1\end{gather*}[#1|end[gather*]] |gdef|@multilineverbatim#1\end{multiline}[#1|end[multiline]] |gdef|@smultilineverbatim#1\end{multiline*}[#1|end[multiline*]] |gdef|@arraxverbatim#1\end{arrax}[#1|end[arrax]] |gdef|@sarraxverbatim#1\end{arrax*}[#1|end[arrax*]] |gdef|@tabulaxverbatim#1\end{tabulax}[#1|end[tabulax]] |gdef|@stabulaxverbatim#1\end{tabulax*}[#1|end[tabulax*]] |endgroup \def\align{\@verbatim \frenchspacing\@vobeyspaces \@alignverbatim You are using the "align" environment in a style in which it is not defined.} \let\endalign=\endtrivlist \@namedef{align*}{\@verbatim\@salignverbatim You are using the "align*" environment in a style in which it is not defined.} \expandafter\let\csname endalign*\endcsname =\endtrivlist \def\alignat{\@verbatim \frenchspacing\@vobeyspaces \@alignatverbatim You are using the "alignat" environment in a style in which it is not defined.} \let\endalignat=\endtrivlist \@namedef{alignat*}{\@verbatim\@salignatverbatim You are using the "alignat*" environment in a style in which it is not defined.} \expandafter\let\csname endalignat*\endcsname =\endtrivlist \def\xalignat{\@verbatim \frenchspacing\@vobeyspaces \@xalignatverbatim You are using the "xalignat" environment in a style in which it is not defined.} \let\endxalignat=\endtrivlist \@namedef{xalignat*}{\@verbatim\@sxalignatverbatim You are using the "xalignat*" environment in a style in which it is not defined.} \expandafter\let\csname endxalignat*\endcsname =\endtrivlist \def\gather{\@verbatim \frenchspacing\@vobeyspaces \@gatherverbatim You are using the "gather" environment in a style in which it is not defined.} \let\endgather=\endtrivlist \@namedef{gather*}{\@verbatim\@sgatherverbatim You are using the "gather*" environment in a style in which it is not defined.} \expandafter\let\csname endgather*\endcsname =\endtrivlist \def\multiline{\@verbatim \frenchspacing\@vobeyspaces \@multilineverbatim You are using the "multiline" environment in a style in which it is not defined.} \let\endmultiline=\endtrivlist \@namedef{multiline*}{\@verbatim\@smultilineverbatim You are using the "multiline*" environment in a style in which it is not defined.} \expandafter\let\csname endmultiline*\endcsname =\endtrivlist \def\arrax{\@verbatim \frenchspacing\@vobeyspaces \@arraxverbatim You are using a type of "array" construct that is only allowed in AmS-LaTeX.} \let\endarrax=\endtrivlist \def\tabulax{\@verbatim \frenchspacing\@vobeyspaces \@tabulaxverbatim You are using a type of "tabular" construct that is only allowed in AmS-LaTeX.} \let\endtabulax=\endtrivlist \@namedef{arrax*}{\@verbatim\@sarraxverbatim You are using a type of "array*" construct that is only allowed in AmS-LaTeX.} \expandafter\let\csname endarrax*\endcsname =\endtrivlist \@namedef{tabulax*}{\@verbatim\@stabulaxverbatim You are using a type of "tabular*" construct that is only allowed in AmS-LaTeX.} \expandafter\let\csname endtabulax*\endcsname =\endtrivlist \def\@@eqncr{\let\@tempa\relax \ifcase\@eqcnt \def\@tempa{& & &}\or \def\@tempa{& &}% \else \def\@tempa{&}\fi \@tempa \if@eqnsw \iftag@ \@taggnum \else \@eqnnum\stepcounter{equation}% \fi \fi \global\@ifnextchar*{\@tagstar}{\@tag}@false \global\@eqnswtrue \global\@eqcnt\z@\cr} \def\endequation{% \ifmmode\ifinner \iftag@ \addtocounter{equation}{-1} $\hfil \displaywidth\linewidth\@taggnum\egroup \endtrivlist \global\@ifnextchar*{\@tagstar}{\@tag}@false \global\@ignoretrue \else $\hfil \displaywidth\linewidth\@eqnnum\egroup \endtrivlist \global\@ifnextchar*{\@tagstar}{\@tag}@false \global\@ignoretrue \fi \else \iftag@ \addtocounter{equation}{-1} \eqno \hbox{\@taggnum} \global\@ifnextchar*{\@tagstar}{\@tag}@false% $$\global\@ignoretrue \else \eqno \hbox{\@eqnnum $$\global\@ignoretrue \fi \fi\fi } \newif\iftag@ \@ifnextchar*{\@tagstar}{\@tag}@false \def\@ifnextchar*{\@tagstar}{\@tag}{\@ifnextchar*{\@tagstar}{\@tag}} \def\@tag#1{% \global\@ifnextchar*{\@tagstar}{\@tag}@true \global\def\@taggnum{(#1)}} \def\@tagstar*#1{% \global\@ifnextchar*{\@tagstar}{\@tag}@true \global\def\@taggnum{#1}% } \makeatother \endinput \end{filecontents} \documentclass[twoside,10pt]{article} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsmath} \setcounter{MaxMatrixCols}{10} \newtheorem{theorem}{Theorem}[section] \newtheorem{hypothesis}[theorem]{Hypothesis} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \newtheorem{definition}[theorem]{Definition} \newtheorem{remark}[theorem]{Remark} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{notation}[theorem]{Notation} \input{tcilatex} \begin{document} \title{Remarks on the $\Gamma $--regularization \\ of Non--convex and Non--semi--continuous \\ Functions on Topological Vector Spaces} \author{J.-B. Bru and W. de Siqueira Pedra} \date{\today } \maketitle \begin{abstract} We show that the minimization problem of any non--convex and non--lower semi--continuous function on a compact convex subset of a locally convex real topological vector space can be studied via an associated convex and lower semi--continuous function $\Gamma \left( h\right) $. This observation uses the notion of $\Gamma $--regularization as a key ingredient. As an application we obtain, on any locally convex real space, a generalization of the Lanford III--Robinson theorem which has only been proven for separable real Banach spaces. The latter is a characterization of subdifferentials of convex continuous functions. \\[0.3ex] {\small \textit{Keywords:} variational problems, non--linear analysis, non--convexity, \\ $\Gamma $--regularization, Lanford III -- Robinson theorem.}\\[0.3ex] {\small \textit{Mathematics subject classifications:} 58E30, 46N10, 52A07.} \end{abstract} \section{Introduction and Main Results} \setcounter{equation}{0}% Minimization problems $\inf \,h\left( K\right) $ on compact convex subsets $% K $ of a locally convex real (topological vector) space\footnote{% We assume throughout this paper that topological vector spaces are Hausdorff spaces, i.e., points in those spaces define closed sets.} $\mathcal{X}$ are extensively studied for convex and lower semi--continuous real--valued functions $h$. See, for instance, \cite{Zeidler3}. Such variational problems are, however, not systematically studied for \emph{% non}--convex and \emph{non}--lower semi--continuous real--valued functions $h$, except for a few specific functions. See for instance \cite{Mueller}. The aim of this paper is to show that -- independently of convexity or lower semi--continuity of functions $h$ -- the minimization problem $\inf \,h\left( K\right) $ on compact convex subsets $K$ of a locally convex real space $\mathcal{X}$ can be analyzed via another minimization problem $\inf \,\Gamma \left( h\right) \left( K\right) $ associated with a convex and lower semi--continuous function $\Gamma \left( h\right) $, for which various methods of analysis are available. We are particularly interested in characterizing the following set of generalized minimizers of any real--valued function $h$ on a compact convex set $K $: \begin{definition}[Set of generalized minimizers] \label{gamm regularisation copy(4)}\mbox{ }\newline Let $K$ be a (non--empty) compact convex subset of a locally convex real space $\mathcal{X}$ and $h:K\rightarrow \left( -\infty ,\infty \right] $ be any extended real--valued function. Then the set $\overline{\mathit{\Omega }\left( h,K\right) }\subset K$ of generalized minimizers of $h$ is the closure of the set \begin{equation*} \mathit{\Omega }\left( h,K\right) :=% \Big\{% x\in K:\exists \{x_{i}\}_{i\in I}\subset K\mathrm{\ \ with\ }% x_{i}\rightarrow x\;\mathrm{and\;}\lim_{I}h(x_{i})=\inf \,h(K)% \Big\}% \end{equation*}% of all limit points of approximating minimizers of $h$. \end{definition} \noindent Here, $\{x_{i}\}_{i\in I}\subset K$ is per definition a net of \emph{approximating minimizers} when \begin{equation*} \underset{I}{\lim }\ h(x_{i})=\inf \,h(K). \end{equation*}% Note that, for any compact set $K$, $\mathit{\Omega }\left( h,K\right) $ is non--empty because any net $\{x_{i}\}_{i\in I}\subset K$ converges along a subnet. In order to motivate the issue here, observe that $\inf \,h\left( K\right) $ can always be studied via a minimization problem associated with a (possibly not convex, but) lower semi--continuous function $h_{0}$, known as the \emph{lower semi--continuous hull} of $h$: \begin{lemma}[Minimization of real--valued functions -- I] \label{theorem trivial sympa 1 copy(2)}\mbox{ }\newline Let $K$ be any (non--empty) compact, convex, and metrizable subset of a locally convex real space $\mathcal{X}$ and $h:K\rightarrow \lbrack \mathrm{k% },\infty ]$ be any extended real--valued function with $\mathrm{k}\in \mathbb{R}$. Then there is a lower semi--continuous extended function $h_{0}:K\rightarrow \lbrack \mathrm{k},\infty ]$ such that% \begin{equation*} \inf \,h\left( K\right) =\inf \,h_{0}\left( K\right) \quad \RIfM@\expandafter\text@\else\expandafter\mbox\fi{and}\quad \mathit{\Omega }\left( h_{0},K\right) =\mathit{\Omega }\left( h,K\right) . \end{equation*} \end{lemma} \noindent By lower semi--continuity, note that $\mathit{\Omega }\left( h_{0},K\right) $ corresponds to the set of usual minimizers of $h_{0}$. Note further that Lemma \ref{theorem trivial sympa 1 copy(2)} implies -- \ in the case $K$ is metrizable -- that $\mathit{\Omega }\left( h,K\right) $ is closed, again by lower semi--continuity of $h_{0}$. The proof of this lemma is straightforward and is given in Section \ref{Section Proofs-I} for completeness. This result has two drawbacks: The compact convex set $K$ must be \emph{% metrizable} in the elementary proof we give here and, more important, the lower semi--continuous hull $h_{0}$ of $h$ is \emph{generally not convex}. We give below a more elaborate result and show that both problems mentioned above can be overcome by using the so--called $\Gamma $--regularization of extended real--valued functions. The last is defined from the space $\mathrm{A}\left( \mathcal{X}\right) $ of all affine continuous real--valued functions on a locally convex real space\ $\mathcal{X}$ as follows (cf. \cite[Eq. (1.3) in Chapter I]{Alfsen}): \begin{definition}[$\Gamma $--regularization of real--valued functions] \label{gamm regularisation}\mbox{ }\newline For any extended real--valued function $h:K\rightarrow \lbrack \mathrm{k},\infty ]$ defined on a (non--empty) compact convex subset $K\subset \mathcal{X}$, its $\Gamma $% --regularization $\Gamma \left( h\right) $ on $K$ is the function defined as the supremum over all affine and continuous minorants $m:\mathcal{X}% \rightarrow \mathbb{R}$ of $h$, i.e., for all $x\in K$, \begin{equation*} \Gamma \left( h\right) \left( x\right) :=\sup \left\{ m(x):m\in \mathrm{A}% \left( \mathcal{X}\right) \;\RIfM@\expandafter\text@\else\expandafter\mbox\fi{and }m|_{K}\leq h\right\} . \end{equation*} \end{definition} \noindent Since the $\Gamma $--regularization $\Gamma \left( h\right) $ of a extended real--valued function $h$ is a supremum over continuous functions, $\Gamma \left( h\right) $ is a \emph{convex} and \emph{lower semi--continuous} function on $K$. For convenience, note that we identify extended real--valued functions $g $ only defined on a\ convex compact subset $K\subset \mathcal{X}$ of the locally convex real space\ $\mathcal{X}$ with its (trivial) extension $g_{\mathrm{ext}}$ to the whole space\ $\mathcal{X}$ defined by% \begin{equation*} g_{\mathrm{ext}}(x):=\left\{ \begin{array}{c} g(x) \\ \infty% \end{array}% \begin{array}{l} \mathrm{for}\ x\in K, \\ \mathrm{otherwise.}% \end{array}% \right. \end{equation*}% Clearly, with this prescription $g$ is lower semi--continuous (resp. convex) on $K$ iff $g$ is lower semi--continuous (resp. convex) on $\mathcal{X}$. We prove in Section \ref{Section Proofs-II} the main result of this paper: \begin{theorem}[Minimization of real--valued functions -- II] \label{theorem trivial sympa 1}\mbox{ }\newline Let $K$ be any (non--empty) compact convex subset of a locally convex real space $\mathcal{X}$ and $h:K\rightarrow \lbrack \mathrm{k},\infty ]$ be any extended real--valued function with $\mathrm{k}\in \mathbb{R}$. Then we have that:\newline \emph{(i)} \begin{equation*} \inf \,h\left( K\right) =\inf \,\Gamma \left( h\right) \left( K\right) . \end{equation*}% \emph{(ii) }The set $\mathit{M}$ of minimizers of $\Gamma \left( h\right) $ over $K$ equals the closed convex hull of the set $\mathit{\Omega }\left( h,K\right) $ of generalized minimizers of $h$ over $K$, i.e., \begin{equation*} \mathit{M}=\overline{\mathrm{co}}\left( \mathit{\Omega }\left( h,K\right) \right) . \end{equation*} \end{theorem} This general fact related to the minimization of non--convex and non--lower semi--continuous real--valued functions on compact convex sets has not been observed\footnote{Assertion (i) is, however, trivial.} before, at least to our knowledge. Note that related results were obtained in \cite{Benoist}\footnote{We thank the referee for pointing out this reference.} for $\mathcal{X}=\mathbb{R}^n$. It turns out to be extremely useful. It is, for instance, an essential argument in the proof given in \cite{BruPedra2} of the validity of the so--called Bogoliubov approximation on the level of states for a class of models for fermions on the lattice. This problem, well--known in mathematical physics, was first addressed by Ginibre \cite[p. 28]{Ginibre} in 1968 and is still open for many physically important models. Then, by using the theory of compact convex subsets of locally convex real spaces $\mathcal{X}$ (see, e.g., \cite{Alfsen}), Theorem \ref{theorem trivial sympa 1} yields a characterization of the set $\overline{\mathit{% \Omega }\left( h,K\right) }$ of all generalized minimizers of $h$ over $K$. Indeed, one important observation concerning locally convex real spaces $% \mathcal{X}$ is that any compact convex subset $K\subset \mathcal{X}$ is the closure of the convex hull of the (non--empty) set $\mathcal{E}(K)$ of its extreme points, i.e., of the points which cannot be expressed as (non--trivial) convex combinations of other elements in $K$. This is the Krein--Milman theorem, see, e.g., \cite[Theorems 3.4 (b) and 3.23]{Rudin}. In fact, among all subsets $Z\subset K$ generating $K$, $\mathcal{E}(K)$ is -- in a sense -- the smallest one. This is the Milman theorem, see, e.g., \cite[Theorem 3.25]{Rudin}. It follows from Theorem \ref{theorem trivial sympa 1} together with \cite[Theorems 3.4 (b), 3.23, 3.25]{Rudin} that extreme points of the compact convex set $\mathit{M}$ of minimizers of $\Gamma \left(h\right) $ over $K$ are generalized minimizers of $h$: \begin{theorem}[Minimization of real-valued functions -- III] \label{theorem trivial sympa 1 copy(1)}\mbox{ }\newline Let $K$ be any (non--empty) compact convex subset of a locally convex real space $\mathcal{X}$ and $h:K\rightarrow \lbrack \mathrm{k},\infty ]$ be any extended real--valued function with $\mathrm{k}\in \mathbb{R}$. Then extreme points of the compact convex set $\mathit{M}$ belong to the set of generalized minimizers of $h$, i.e., $\mathcal{E}\left( \mathit{M}\right) \subseteq \overline{% \mathit{\Omega }\left( h,K\right) }$. \end{theorem} \noindent This last result makes possible a \emph{full characterization} of the closure of the set $\mathit{\Omega }\left( h,K\right) $ in the following sense: Since $\mathit{M}$ is compact and convex, we can study the minimization problem $\inf \,h\left( K_{\mathit{M}}\right) $ for any closed (and hence compact) convex subset $K_{\mathit{M}}\subset \mathit{M}$. Applying Theorem \ref{theorem trivial sympa 1} we get \begin{equation} \inf \,h\left( K_{\mathit{M}}\right) =\inf \,\Gamma \left( h|_{K_{\mathit{M}% }}\right) \left( K_{\mathit{M}}\right) . \label{full characterization1} \end{equation}% If \begin{equation*} \inf \,h\left( K_{\mathit{M}}\right) =\inf \,h\left( K\right) \end{equation*}% then, by Theorem \ref{theorem trivial sympa 1 copy(1)}, \begin{equation*} \mathcal{E}\left( \mathit{M}_{K_{\mathit{M}}}\right) \subseteq \overline{% \mathit{\Omega }\left( h|_{K_{\mathit{M}}},K_{\mathit{M}}\right) }\subseteq \overline{\mathit{\Omega }\left( h,K\right) }, \end{equation*}% where $\mathit{M}_{K_{\mathit{M}}}$ is the compact convex set of minimizers of $\Gamma \left( h|_{K_{\mathit{M}}}\right) $ over $K_{\mathit{M}}\subset \mathit{M}$. In general, $\mathcal{E}\left( \mathit{M}_{K_{\mathit{M}% }}\right) \backslash \mathcal{E}\left( \mathit{M}\right) \neq \emptyset $ because $\mathit{M}_{K_{\mathit{M}}}$ is not necessarily a face of $\mathit{M% }$. Thus we discover in this manner new points of $\overline{\mathit{\Omega }% \left( h,K\right) }$ not contained in $\mathcal{E}\left( \mathit{M}\right) $% . Choosing a sufficiently large family $\{K_{\mathit{M}}\}$ of closed convex subsets of $\mathit{M}$ we can exhaust the set $\overline{\mathit{\Omega }% \left( h,K\right) }$ through the union $\cup $ $\{\mathcal{E}\left( \mathit{M% }_{K_{\mathit{M}}}\right) \}$. Note that this construction can be performed in an inductive way: For each set $\mathit{M}_{K_{\mathit{M}}}$ of minimizers consider further closed convex subsets $K_{\mathit{M}}^{\prime }\subset $ $\mathit{M}_{K_{\mathit{M}}}$. The art consists in choosing the family $\{K_{\mathit{M}}\}$ appropriately, i.e., it should be as small as possible and the extreme points of $\mathit{M}_{K_{\mathit{M}}}$ should possess some reasonable characterization. Of course, the latter heavily depends on the function $h$ and on particular properties of the compact convex set $K$ (e.g., density of $\mathcal{E}(K)$, metrizability, etc.). To close this section we recall that the $\Gamma $--regularization $\Gamma \left( h\right) $ of a function $h$ on $K$ equals its twofold \emph{% Legendre--Fenchel transform} -- also called the \emph{biconjugate }% (function) of $h$. See, for instance, \cite[Paragraph 51.3]{Zeidler3}. Indeed, $\Gamma \left( h\right) $ is the largest lower semi--continuous and convex minorant of $h$ (cf. Corollary \ref{Biconjugate}% ). However, in contrast to the $\Gamma $--regularization the notion of Legendre--Fenchel transform requires the use of dual pairs (cf. Definition % \ref{dual pairs}). Since, for any locally convex real space $\mathcal{X}$ together with the space $\mathcal{X}^{\ast }$ of linear continuous funtionals $\mathcal{X} \to \mathbb{R}$ (dual space) equipped with the weak$^{\ast }$--topology, $(\mathcal{X},\mathcal{X}^{\ast })$ is a dual pair, the Legendre--Fenchel transform can be defined on any locally convex real space $\mathcal{X}$ as follows: \begin{definition}[The Legendre--Fenchel transform] \label{Legendre--Fenchel transform}\mbox{ }\newline Let $K$ be a (non--empty) compact convex subset of a locally convex real space $\mathcal{X}$. For any extended real--valued function $h:K\rightarrow \left( -\infty,\infty \right] $, $h\not\equiv \infty$, its Legendre--Fenchel transform $h^{\ast }$ is the convex weak$^{\ast }$--lower semi--continuous extented function from $\mathcal{X}^{\ast }$ to $\left( -\infty ,\infty \right] $ defined, for any $x^{\ast }\in \mathcal{% X}^{\ast }$, by \begin{equation*} h^{\ast }\left( x^{\ast }\right) :=\underset{x\in K}{\sup }\left\{ x^{\ast }\left( x\right) -h\left( x\right) \right\} . \end{equation*} \end{definition} See also \cite[Definition 51.1]{Zeidler3}. Note that, together with its weak$^{\ast }$--topology, the dual space $\mathcal{X}^{\ast }$ of any locally convex space $\mathcal{X}$ is also a locally convex space, see \cite[Theorems 3.4 (b) and 3.10]{Rudin}. Therefore, in case nothing is further specified, the space $\mathcal{X}% ^{\ast }$ is always equipped with its weak$^{\ast }$--topology. The Legendre--Fenchel transform is strongly related to the notion of Fenchel subdifferentials (see also \cite{Phelps-conv}): \begin{definition}[Fenchel subdifferentials] \label{tangent functional}\mbox{ }\newline Let $h:\mathcal{X}\rightarrow \mathbb{(-\infty },\infty ]$ be any extended real--valued function on a real topological vector space $\mathcal{X}$. A continuous linear functional $\mathrm{d}h_{x}\in \mathcal{X}^{\ast }$ is said to be a Fenchel subgradient (or tangent) of the function $h$ at $x\in \mathcal{X}$ iff, for all $x^{\prime }\in \mathcal{X}$, $h(x+x^{\prime })\geq h(x)+\mathrm{d}% h_{x}(x^{\prime })$. The set $\partial h(x)\subset \mathcal{X}^{\ast }$ of Fenchel subgradients of $h$ at $x$ is called Fenchel subdifferential of $h$ at $x$. \end{definition} \noindent Theorem \ref{theorem trivial sympa 1} establishes a link between generalized minimizers and Fenchel subdifferentials: \begin{theorem}[Subdifferentials of continuous convex functions -- I] \label{theorem trivial sympa 3}Let $K$ be any (non--empty) compact convex subset of a locally convex real space $\mathcal{X}$ and $h:K\rightarrow \lbrack \mathrm{k},\infty ]$ be any extended real--valued function with $\mathrm{k}\in \mathbb{R}$. Then the Fenchel subdifferential $\partial h^{\ast }(x^{\ast })\subset \mathcal{X}$ of $h^{\ast }$ at the point $x^{\ast }\in \mathcal{X}^{\ast }$ is the (non--empty) compact convex set \begin{equation*} \partial h^{\ast }(x^{\ast })=\overline{\mathrm{co}}\left( \mathit{\Omega }% \left( h-x^{\ast },K\right) \right) . \end{equation*} \end{theorem} \noindent This last result -- proven in Section \ref{Section Proofs-III} -- generalizes the Lanford III--Robinson theorem \cite[Theorem 1]{LanRob} which has only been proven for separable real Banach spaces $\mathcal{X}$ and continuous convex functions $h:\mathcal{X}\rightarrow \mathbb{R}$, cf. Theorem \ref{Land.Rob}. Indeed, for any extended real--valued function $h$ from a compact convex subset $K\subset \mathcal{X}$ of a locally convex real space $\mathcal{X}$ to $\left( -\infty ,\infty \right] $, let% \begin{equation*} \mathcal{Y}^{\ast }:=\left\{ x^{\ast }\in \mathcal{X}^{\ast }:h^{\ast }\RIfM@\expandafter\text@\else\expandafter\mbox\fi{ has a unique Fenchel subgradient }\mathrm{d}h_{x^{\ast }}^{\ast }\in \mathcal{X}% \RIfM@\expandafter\text@\else\expandafter\mbox\fi{ at }x^{\ast }\right\} . \end{equation*}% For all $x^{\ast }\in \mathcal{X}^{\ast }$ and any open neighborhood $% \mathcal{V}$ of\ $\{0\}\subset \mathcal{X}^{\ast }$, we also define the set% \begin{equation} \mathcal{T}_{x^{\ast },\mathcal{V}}:=\overline{\left\{ \mathrm{d}h_{y^{\ast }}^{\ast }:y^{\ast }\in \mathcal{Y}^{\ast }\cap (x^{\ast }+\mathcal{V)}% \right\} }^{\mathcal{X}}\subset \mathcal{X} \label{set t voisinage} \end{equation}% and denote by $\mathcal{T}_{x^{\ast }}$ the intersection% \begin{equation} \mathcal{T}_{x^{\ast }}:=\bigcap\limits_{\mathcal{V}\ni 0\RIfM@\expandafter\text@\else\expandafter\mbox\fi{ open}}% \mathcal{T}_{x^{\ast },\mathcal{V}}. \label{set t voinagebis} \end{equation}% Here, $\overline{\;\cdot \;}^{\mathcal{X}}$ denotes the closure w.r.t. the topology of $\mathcal{X}$. Then we observe first that Theorem \ref{theorem trivial sympa 3} implies that the set $\partial h^{\ast }(x^{\ast })\subset \mathcal{X}$ of Fenchel subgradients of $h^{\ast }$ at the point $x^{\ast }\in \mathcal{X}^{\ast }$ is included in the closed convex hull of the set $\mathcal{T}_{x^{\ast }}$ provided $\mathcal{Y}^{\ast }$ is dense in $\mathcal{X}^{\ast }$ (cf. Section \ref{Section Proofs-IV}): \begin{corollary}[Subdifferentials of continuous convex functions -- II] \label{corollary explosion lanford-robinson}Let $K$ be any (non--empty) compact convex subset of a locally convex real space $\mathcal{X}$ and $% h:K\rightarrow \lbrack \mathrm{k},\infty ]$ be any extended real--valued function with $\mathrm{k}\in \mathbb{R}$. If $\mathcal{Y}^{\ast }$ is dense in $\mathcal{X}^{\ast }$ then, for any $x^{\ast }\in \mathcal{X}^{\ast }$, \begin{equation*} \partial h^{\ast }(x^{\ast })\subseteq \overline{\mathrm{co}}\left( \mathcal{T% }_{x^{\ast }}\right) . \end{equation*} \end{corollary} \noindent This last result applied on separable Banach spaces yields, in turn, the following assertion (cf. Section \ref{Section Proofs-IV copy(1)}): \begin{corollary}[The Lanford III--Robinson theorem] \label{corollary explosion lanford-robinson copy(1)}\mbox{ }\newline Let $\mathcal{X}$ be a separable Banach space and $h:\mathcal{X}\rightarrow \mathbb{R}$ be any convex function which is globally Lipschitz continuous. If the set% \begin{equation*} \mathcal{Y}:=\left\{ x\in \mathcal{X}:h\RIfM@\expandafter\text@\else\expandafter\mbox\fi{ has a unique Fenchel subgradient }% \mathrm{d}h_{x}\in \mathcal{X}^{\ast }\RIfM@\expandafter\text@\else\expandafter\mbox\fi{ at }x\right\} \end{equation*}% is dense in $\mathcal{X}$ then the Fenchel subdifferential $\partial h(x)$ of $h$, at any $x\in \mathcal{X}$, is the weak$^{\ast }$--closed convex hull of the set $\mathcal{Z}_{x}$. Here, at fixed $x\in \mathcal{X}$, $\mathcal{Z}_{x}$ is the set of functionals $x^{\ast }\in \mathcal{X}^{\ast }$ such that there is a net $\{x_{i}\}_{i\in I}$ in $\mathcal{Y}$ converging to $x$ with the property that the unique Fenchel subgradient $\mathrm{d}h_{x_{i}}\in \mathcal{X}^{\ast }$ of $h$ at $x_{i}$ converges towards $x^{\ast }$ in the weak$^{\ast}$--topology. \end{corollary} \noindent Recall that the Mazur theorem shows that the set $\mathcal{Y}$ on which a continuous convex function $h$ is G\^{a}teaux differentiable, i.e., the set $\mathcal{Y}$ for which $h$ has exactly one Fenchel subgradient \textrm{d}$h_{x}\in \mathcal{X}^{\ast }$ at any $x\in \mathcal{Y}$, is dense in a separable Banach space $\mathcal{X}$, cf. Theorem \ref{Mazur} and Remark \ref{Mazur remark}. Therefore, for globally Lipschitz continuous and convex functions, the Lanford III--Robinson theorem \cite[Theorem 1]% {LanRob} (cf. Theorem \ref{Land.Rob}) directly follows from Corollary \ref% {corollary explosion lanford-robinson copy(1)}. Observe that, in which concerns Fenchel subdifferentials of convex continuous functions on Banach spaces, the case of global Lipschitz continuous functions is already the most general case: For any continuous convex function $h$ on a Banach space $% \mathcal{X}$ and any $x\in \mathcal{X}$, there are $\varepsilon >0$ and a globally Lipschitz continuous convex function $g$ such that $g\left( y\right) =h\left( y\right) $ whenever $\left\Vert x-y\right\Vert <\varepsilon $. In particular, $g$ and $h$ have the same Fenchel subgradients at $x$. Remark, indeed, that continuous convex functions $h$ on a Banach space $\mathcal{X}$ are locally Lipschitz continuous and an example of such a global Lipschitz continuous convex function is given by \begin{equation*} g\left( x\right) :=\inf \left\{ z\in \mathbb{R}:\left( x,z\right) \in \left[ \mathrm{epi}\left( h\right) +\mathcal{C}_{\alpha }\right] \right\} , \end{equation*}% for sufficiently small $\alpha >0$. Here, \begin{equation*} \mathcal{C}_{\alpha }:=\left\{ \left( x,z\right) \in \mathcal{X}\times \mathbb{R}:z\geq 0,\left\Vert x\right\Vert \leq \alpha z\right\} \end{equation*}% and $\mathrm{epi}\left( h\right) $ is the epigraph of $h$ defined by \begin{equation*} \mathrm{epi}\left( h\right) :=\left\{ \left( x,z\right) \in \mathcal{X}% \times \mathbb{R}:z\geq f\left( x\right) \right\} . \end{equation*} The rest of the paper is structured as follows. Section \ref{Section Proofs} gives the detailed proofs of Lemma \ref{theorem trivial sympa 1 copy(2)}, Theorems \ref{theorem trivial sympa 1}, \ref{theorem trivial sympa 3}, and Corollaries \ref{corollary explosion lanford-robinson}--\ref{corollary explosion lanford-robinson copy(1)}. Then, Section \ref{Concluding remarks} discusses an additional observation which is relevant in the context of minimization of non--convex or non--semi--continuous functions and which does not seem to have been observed before. Indeed, Lemma \ref{Bauer maximum principle bis} gives an extension of the Bauer maximum principle (Lemma \ref% {Bauer maximum principle}). Finally, Section \ref{Section appendix} is a concise appendix about dual pairs, barycenters in relation with the $\Gamma $--regularization, the Mazur theorem, and the Lanford III--Robinson theorem. \section{Proofs\label{Section Proofs}} This section gives the detailed proofs of Lemma \ref{theorem trivial sympa 1 copy(2)}, Theorems \ref{theorem trivial sympa 1}, \ref{theorem trivial sympa 3}, and Corollaries \ref{corollary explosion lanford-robinson}--\ref% {corollary explosion lanford-robinson copy(1)}. Up to Corollary \ref% {corollary explosion lanford-robinson copy(1)}, we will always assume that $% K $ is a (non--empty) compact convex subset of a locally convex real space $% \mathcal{X}$ and $h:K\rightarrow \lbrack \mathrm{k},\infty ]$ is any extended real--valued function with $\mathrm{k}\in \mathbb{R}$. In Lemma \ref{theorem trivial sympa 1 copy(2)} the metrizability of the topology on $K$ is also assumed. In Corollary \ref{corollary explosion lanford-robinson copy(1)} $\mathcal{X}$ is a separable Banach space and $h:\mathcal{X}\rightarrow \mathbb{R}$ is any globally Lipschitz continuous convex function. \subsection{Proof of Lemma \protect\ref{theorem trivial sympa 1 copy(2)} \label{Section Proofs-I}} Because the subset $K\subset \mathcal{X}$ is metrizable and compact, it is sequentially compact and we can restrict ourselves to sequences instead of more general nets. Using any metric $d(x,y)$ on $K$ generating the topology we define, at fixed $\delta >0$, the extended real--valued function $h_{\delta }$ from $K$ to $[\mathrm{k},\infty ]$ by% \begin{equation*} h_{\delta }\left( x\right) :=\inf \,h(\mathcal{B}_{\delta }\left( x\right) ) \end{equation*}% for any $x\in K$, where% \begin{equation} \mathcal{B}_{\delta }\left( x\right) :=\left\{ y\in K:\ d(x,y)<\delta \right\} \label{ball} \end{equation}% is the ball (in $K$) of radius $\delta >0$ centered at $x\in K$. The family $% \{h_{\delta }\left( x\right) \}_{\delta >0}$ of extended real--valued functions is clearly increasing as $\delta \searrow 0$ and is bounded from above by $h(x)$. Therefore, for any $x\in K$, the limit of $h_{\delta }\left( x\right) \geq \mathrm{k}$ as $\delta \searrow 0$ exists and defines an extended real--valued function \begin{equation*} x\mapsto h_{0}\left( x\right) :=\underset{\delta \searrow 0}{\lim }% \,h_{\delta }\left( x\right) \end{equation*}% from $K$ to $[\mathrm{k},\infty ]$. In fact, this construction is well--known and the function $h_{0}$ is called the \emph{lower semi--continuous hull} of $h$ as it is a lower semi--continuous extended real--valued function from $K$ to $[\mathrm{k},\infty ]$. Indeed, for all $\delta >0$ and any sequence $\{x_{n}\}_{n=1}^{\infty }\subset K$ converging to $x\in K$, there is $N_{\delta }>0$ such that, for all $% n>N_{\delta }$, $x_{n}\in \mathcal{B}_{\delta /2}\left( x\right) $ which implies that $\mathcal{B}_{\delta /2}\left( x_{n}\right) \subset \mathcal{B}% _{\delta }\left( x\right) $. In particular, $h_{\delta }\left( x\right) \leq h_{\delta /2}\left( x_{n}\right) $ for all $\delta >0$ and $n>N_{\delta }$. Since the family $\{h_{\delta }\left( x\right) \}_{\delta >0}$ defines an increasing sequence as $\delta \searrow 0$, it follows that% \begin{equation*} h_{\delta }\left( x\right) \leq \ \liminf_{n\rightarrow \infty }h_{0}\left( x_{n}\right) \end{equation*}% for any $\delta >0$ and $x\in K$. In the limit $\delta \searrow 0$ the latter yields the lower semi--continuity of the extended real--valued function $h_{0}$ on $K $. Moreover, \begin{equation} h_{0}\left( x\right) \geq h_{\delta }\left( x\right) \geq \inf \,h(K)\geq \mathrm{k}>-\infty \label{definition de h lower3bis} \end{equation}% for any $x\in K$ and $\delta >0$. We observe now that $h$ and $h_{0}$ have the same infimum on $K$:% \begin{equation} \inf h_{0}\left( K\right) =\inf h(x). \label{definition de h lower3} \end{equation}% This can be seen by observing first that there is $y\in K$ such that \begin{equation} \inf h_{0}\left( K\right) =h_{0}\left( y\right) \label{definition de h lower4} \end{equation}% because of the lower semi--continuity of $h_{0}$. Since $h_{\delta }\leq h$ on $K$ for any $\delta >0$, we have $h_{0}\leq h\ $on $K$, which combined with (\ref{definition de h lower3bis}) and (\ref{definition de h lower4}) yields Equality (\ref{definition de h lower3}). Additionally, for all $\delta >0$ and any minimizer $y\in K$ of $h_{0}$ over $K$, there is a sequence $\{x_{\delta ,n}\}_{n=1}^{\infty }\subset \mathcal{B% }_{\delta }\left( y\right) $ of approximating minimizers of $h$ over $% \mathcal{B}_{\delta }\left( y\right) $, that is, \begin{equation*} h_{\delta }\left( y\right) :=\inf \,h(\mathcal{B}_{\delta }\left( y\right) )=% \underset{n\rightarrow \infty }{\lim }h(x_{\delta ,n})\leq h(y). \end{equation*}% We can assume without loss of generality that \begin{equation*} d(x_{\delta ,n},y)\leq \delta \mathrm{\quad and\quad }|h(x_{\delta ,n})-h_{\delta }\left( y\right) |\leq 2^{-n} \end{equation*}% for all $n\in \mathbb{N}$ and all $\delta >0$. Note that $h_{\delta }\left( y\right) \rightarrow $ $h_{0}\left( y\right) $ as $\delta \searrow 0$. Thus, by taking any function $p(\delta )\in \mathbb{N}$ satisfying $p(\delta )>\delta ^{-1}$ we obtain that $x_{\delta ,p(\delta )}$ converges to $y\in K$ as $\delta \searrow 0$ with the property that $h(x_{\delta ,p(\delta )})$ converges to $h_{0}\left( y\right) $. Using Equalities (\ref{definition de h lower3}) and (\ref{definition de h lower4}) we obtain that all minimizers of (\ref{definition de h lower4}) are generalized minimizers of $h$, i.e., \begin{equation*} \mathit{\Omega }\left( h_{0},K\right) \subseteq \mathit{\Omega }\left( h,K\right) . \end{equation*}% The converse inclusion% \begin{equation*} \mathit{\Omega }\left( h,K\right) \subseteq \mathit{\Omega }\left( h_{0},K\right) \end{equation*}% is straightforward because one has the inequality $h_{0}\leq h$ on $K$ as well as Equality (\ref{definition de h lower3}). \subsection{Proof of Theorem \protect\ref{theorem trivial sympa 1}\label% {Section Proofs-II}} The assertion (i) of Theorem \ref{theorem trivial sympa 1} is a standard result. Indeed, by Definition \ref{gamm regularisation}, $\Gamma \left( h\right) \leq h$ on $K$ and thus% \begin{equation*} \inf \,\Gamma \left( h\right) \left( K\right) \leq \inf \,h\left( K\right) . \end{equation*}% The converse inequality is derived by restricting the supremum in Definition % \ref{gamm regularisation} to constant maps $m$ from $\mathcal{X}$ to $% \mathbb{R}$ with $\mathrm{k}\leq m\leq h$ on $K$. Observe that the variational problem $\inf \,\Gamma \left( h\right) (K)$ has minimizers and the set $\mathit{M}=\mathit{\Omega }\left( \Gamma \left( h\right) ,K\right) $ of all minimizers of $\Gamma \left( h\right) $ is convex and compact. For any $y\in \mathit{\Omega }\left( h,K\right) $, there is a net $\left\{ x_{i}\right\} _{i\in I}\subset K$ of approximating minimizers of $h$ on $K$ converging to $y$. In particular, since the function $\Gamma \left( h\right) $ is lower semi--continuous and $\Gamma \left( h\right) \leq h$ on $% K$, we have that \begin{equation*} \Gamma \left( h\right) (y)\leq \underset{I}{\liminf }\,\Gamma \left( h\right) (x_{i})\leq \underset{I}{\lim }\,h(x_{i})=\inf \,h(K)=\inf \,\Gamma \left( h\right) (K), \end{equation*}% i.e., $y\in \mathit{M}$. Since $\mathit{M}$ is convex and compact, we obtain that \begin{equation} \mathit{M}\supset \overline{\mathrm{co}}\left( \mathit{\Omega }\left( h,K\right) \right) . \label{inclusion1} \end{equation}% So, we prove now the converse inclusion. We can assume without loss of generality that $\overline{\mathrm{co}}\left( \mathit{\Omega }\left( h,K\right) \right) \neq K$ since otherwise there is nothing to prove. We show\ next that, for any $x\in K\backslash \overline{\mathrm{co}}\left( \mathit{\Omega }\left( h,K\right) \right) $, we have $x\notin $ $\mathit{M}$% . As $\overline{\mathrm{co}}\left( \mathit{\Omega }\left( h,K\right) \right) $ is a closed set of a locally convex real space $\mathcal{X}$, for any $x\in K\backslash \overline{\mathrm{co}}\left( \mathit{\Omega }\left( h,K\right) \right)$, there is an open and convex neighborhood $\mathcal{V}_{x}\subset $ $\mathcal{X}$ of $\{0\}\subset \mathcal{X}$ which is symmetric, i.e., $% \mathcal{V}_{x}=-\mathcal{V}_{x}$, and which satisfies \begin{equation*} \mathcal{G}_{x}\cap \left[ \{x\}+\mathcal{V}_{x}\right] =\emptyset \end{equation*}% with% \begin{equation*} \mathcal{G}_{x}:=K\cap \left[ \overline{\mathrm{co}}\left( \mathit{\Omega }% \left( h,K\right) \right) +\mathcal{V}_{x}\right] . \end{equation*}% This follows from \cite[Theorem 1.10]{Rudin} together with the fact that each neighborhood of $\{0\}\subset \mathcal{X}$ contains some open and convex neighborhood of $\{0\}\subset \mathcal{X}$ because $\mathcal{X}$ is locally convex. Observe also that any one--point set $\{x\}\subset $ $% \mathcal{X}$ is trivially compact. For any neighborhood $\mathcal{V}_{x}$ of $\{0\}\subset \mathcal{X}$ in a locally convex real space, there is another convex, symmetric, and open neighborhood $\mathcal{V}_{x}^{\prime }$ of $\{0\}\subset \mathcal{X}$ such that $[\mathcal{V}_{x}^{\prime }+\mathcal{V}_{x}^{\prime }]\subset \mathcal{V% }_{x}$, see proof of \cite[Theorem 1.10]{Rudin}. Let% \begin{equation*} \mathcal{G}_{x}^{\prime }:=K\cap \left[ \overline{\mathrm{co}}\left( \mathit{% \Omega }\left( h,K\right) \right) +\mathcal{V}_{x}^{\prime }\right] . \end{equation*}% Then the following inclusions hold:% \begin{equation} \overline{\mathrm{co}}\left( \mathit{\Omega }\left( h,K\right) \right) % \subset \mathcal{G}_{x}^{\prime }\subset \overline{\mathcal{G}_{x}^{\prime }}% \subset \mathcal{G}_{x}\subset \overline{\mathcal{G}_{x}}\subset K\backslash \{x\}. \label{eq sup} \end{equation}% Since $K$, $\mathcal{V}_{x}$, $\mathcal{V}_{x}^{\prime }$, and $\overline{% \mathrm{co}} \left( \mathit{\Omega }\left( h,K\right) \right) $ are all convex sets, $\mathcal{G}_{x}$ and $\mathcal{G}_{x}^{\prime }$ are also convex. Seen as subsets of $K$ they are open neighborhoods of $\overline{% \mathrm{co}}\left( \mathit{\Omega }\left( h,K\right) \right) $. The set $\mathcal{X}$ is a Hausdorff space and thus any compact subset $K$ of $\mathcal{X}$ is a normal space. By Urysohn lemma, there is a continuous function% \begin{equation*} f_{x}:K\rightarrow \lbrack \inf h(K),\inf h(K\backslash \mathcal{G}% _{x}^{\prime })] \end{equation*}% satisfying $f_{x}\leq h$ and \begin{equation*} f_{x}\left( y\right) =\left\{ \begin{array}{ll} \inf h(K) & \mathrm{for\ }y\in \overline{\mathcal{G}_{x}^{\prime }}. \\ \inf h(K\backslash \mathcal{G}_{x}^{\prime }) & \mathrm{for\ }y\in K\backslash \mathcal{G}_{x}.% \end{array}% \right. \end{equation*}% By compactness of $K\backslash \mathcal{G}_{x}^{\prime }$ and the inclusion $\mathit{\Omega }\left( h,K\right) \subset \mathcal{G}_{x}^{\prime }$, observe that% \begin{equation*} \inf h(K\backslash \mathcal{G}_{x}^{\prime })>\inf h(K). \end{equation*}% Then we have per construction that \begin{equation} f_{x}(\overline{\mathrm{co}}\left( \mathit{\Omega }\left( h,K\right) \right) % )=\{\inf h(K)\} \label{Omega f sympa} \end{equation}% and% \begin{equation} f_{x}^{-1}(\inf h(K))=\mathit{\Omega }\left( f_{x},K\right) \subset \mathcal{% G}_{x} \label{Omega f sympabis} \end{equation}% for any $x\in K\backslash \overline{\mathrm{co}}\left( \mathit{\Omega }\left( h,K\right) \right) $. We use now the $\Gamma $--regularization $\Gamma \left( f_{x}\right) $ of $% f_{x}$ on the set $K$ and denote by $\mathit{M}_{x}=\mathit{\Omega }\left( \Gamma (f_{x}),K\right) $ its non--empty set of minimizers over $K$. Applying Theorem \ref{Thm - Corollary I.3.6}, for any $y\in \mathit{M}_{x}$, we have a probability measure $\mu _{y}\in M_{1}^{+}(K)$\ on $K$ with barycenter $y$ such that \begin{equation} \Gamma \left( f_{x}\right) \left( y\right) =\int_{K}\mathrm{d}\mu _{y}(z)\;f_{x}\left( z\right) . \label{herve bis} \end{equation}% As $y\in \mathit{M}_{x}$, i.e., \begin{equation} \Gamma \left( f_{x}\right) \left( y\right) =\inf \,\Gamma \left( f_{x}\right) (K)=\inf f_{x}(K), \label{herve 2} \end{equation}% we deduce from (\ref{herve bis}) that \begin{equation*} \mu _{y}(\mathit{\Omega }\left( f_{x},K\right) )=1 \end{equation*}% and it follows that $y\in $ $\overline{\mathrm{co}}\left( \mathit{\Omega }% \left( f_{x},K\right) \right) $, by Theorem \ref{thm barycenter}. Using (% \ref{Omega f sympabis}) together with the convexity of the open neighborhood $\mathcal{G}_{x}$ of $\overline{\mathrm{co}}\left( \mathit{\Omega }\left( h,K\right) \right) $ we thus obtain% \begin{equation} \mathit{M}_{x}\subset \overline{\mathrm{co}}\left( \mathit{\Omega }\left( f_{x},K\right) \right) \subset \overline{\mathcal{G}_{x}} \label{herve 3} \end{equation}% for any $x\in K\backslash \overline{\mathrm{co}}\left( \mathit{\Omega }\left( h,K\right) \right) $. We remark now that the inequality $f_{x}\leq h$ on $K$ yields $\Gamma \left( f_{x}\right) \leq \Gamma \left( h\right) $ on $K$ because of Corollary \ref% {Biconjugate}. As a consequence, it results from (i) and (\ref{Omega f sympa}% ) that the set $\mathit{M}$ of minimizers of $\Gamma \left( h\right) $ over $% K$ is included in $\mathit{M}_{x}$, i.e., $\mathit{M}\subset \mathit{M}_{x}$% . Hence, by (\ref{eq sup}) and (\ref{herve 3}), we have the inclusions \begin{equation} \mathit{M}\subset \overline{\mathcal{G}_{x}}\subset K\backslash \{x\}. \label{inclusion2bis} \end{equation}% Therefore, we combine (\ref{inclusion1}) with (\ref{inclusion2bis}) for all $% x\in K\backslash \overline{\mathrm{co}}\left( \mathit{\Omega }\left( h,K\right) \right) $ to obtain the desired equality in the assertion (ii) of Theorem \ref{theorem trivial sympa 1}. \subsection{Proof of Theorem \protect\ref{theorem trivial sympa 3}\label% {Section Proofs-III}} The proof of Theorem \ref{theorem trivial sympa 3} is a simple consequence of Theorem \ref{theorem trivial sympa 1} together with the following well--known result: \begin{lemma}[Fenchel subgradients as minimizers] \label{theorem trivial sympa 2}\mbox{ }\newline Let $(\mathcal{X},\mathcal{X}^{\ast })$ be a dual pair and $h\not\equiv \infty$ be any extended real--valued function from a (non--empty) convex subset $K\subseteq \mathcal{X}$ to $% (-\infty ,\infty ]$. Then the Fenchel subdifferential $\partial h^{\ast }(x^{\ast })\subset \mathcal{X}$ of $h^{\ast }$ at the point $x^{\ast }\in \mathcal{X}% ^{\ast }$ is the (non--empty) set $\mathit{M}_{x^{\ast }}$ of minimizers over $K$ of the map \begin{equation*} y\mapsto \Gamma \left( h\right) \left( y\right) -x^{\ast }\left( y\right) \end{equation*}% from $K\subseteq \mathcal{X}$ to $(-\infty ,\infty ]$. \end{lemma} \textit{Proof. }The proof is standard and simple, see, e.g., \cite[Theorem I.6.6]{Simon}. Indeed, any Fenchel subgradient $x\in \mathcal{X}$ of the Legendre--Fenchel transform $h^{\ast }$ at the point $x^{\ast }\in \mathcal{X% }$ satisfies the inequality: \begin{equation} x^{\ast }\left( x\right) +h^{\ast }\left( y^{\ast }\right) -y^{\ast }\left( x\right) \geq h^{\ast }\left( x^{\ast }\right) \label{landford1landford1} \end{equation}% for any $y^{\ast }\in \mathcal{X}^{\ast }$, see Definition \ref{tangent functional}. Since $h^{\ast }=h^{\ast \ast \ast }$ and $\Gamma \left( h\right) =h^{\ast \ast }$ (cf. Corollary \ref{Biconjugate} and \cite[% Proposition 51.6]{Zeidler3}), we have (\ref{landford1landford1}) iff \begin{equation*} x^{\ast }\left( x\right) +\underset{y^{\ast }\in \mathcal{X}^{\ast }}{\inf }% \left\{ h^{\ast }\left( y^{\ast }\right) -y^{\ast }\left( x\right) \right\} =x^{\ast }\left( x\right) -\Gamma \left( h\right) \left( x\right) \geq \underset{y\in K}{\sup }\left\{ x^{\ast }\left( y\right) -\Gamma \left( h\right) \left( y\right) \right\} , \end{equation*}% see Definition \ref{Legendre--Fenchel transform}. \hfill\qed\medbreak% We combine now Theorem \ref{theorem trivial sympa 1} with Lemma \ref{theorem trivial sympa 2} to characterize the Fenchel subdifferential $\partial h^{\ast }(x^{\ast })\subset \mathcal{X}$ of $h^{\ast }$ at the point $x^{\ast }\in \mathcal{X}^{\ast }$ as the closed convex hull of the set $\mathit{\Omega }% \left( h-x^{\ast },K\right) $ of generalized minimizers of $h$ over a compact convex subset $K$, see Definition \ref{gamm regularisation copy(4)}. Indeed, for any $x^{\ast }\in \mathcal{X}^{\ast }$, \begin{equation*} \Gamma \left( h-x^{\ast }\right) =\Gamma \left( h\right) -x^{\ast }, \end{equation*}% see Definition \ref{gamm regularisation}. \subsection{Proof of Corollary \protect\ref{corollary explosion lanford-robinson}\label{Section Proofs-IV}} For $x^{\ast }\in \mathcal{X}^{\ast }$ and any open neighborhood $\mathcal{V} $ of $\{0\}\subset \mathcal{X}^{\ast }$, we define the map $g_{\mathcal{V}% ,x^{\ast }}$ from $\mathcal{X}$ to $[\mathrm{k},\infty ]$ with $\mathrm{k}% \in \mathbb{R}$ by% \begin{equation*} g_{\mathcal{V},x^{\ast }}\left( x\right) :=\left\{ \begin{array}{c} \Gamma \left( h\right) \left( x\right) \\ \infty% \end{array}% \begin{array}{l} \mathrm{for}\ x=\mathrm{d}h_{y^{\ast }}^{\ast }\mathrm{\ with}\ y^{\ast }\in \mathcal{Y}^{\ast }\cap (x^{\ast }+\mathcal{V)}, \\ \mathrm{otherwise.}% \end{array}% \right. \end{equation*}% For any $y^{\ast }\in \mathcal{Y}^{\ast }\cap \left( x^{\ast }+\mathcal{V}% \right) $, one has the equality $g_{\mathcal{V},x^{\ast }}^{\ast }\left( y^{\ast }\right) =h^{\ast }\left( y^{\ast }\right) $. This easily follows from the fact that \begin{align*} h^{\ast }\left( y^{\ast }\right) & =\underset{z\in K}{\sup }\left\{ y^{\ast }\left( z\right) -\Gamma \left( h\right) \left( z\right) \right\} =y^{\ast }\left( x\right) -\Gamma \left( h\right) \left( x\right) \\ & =\underset{z\in K}{\sup }\left\{ y^{\ast }\left( z\right) -g_{\mathcal{V}% ,x^{\ast }}\left( z\right) \right\} =g_{\mathcal{V},x^{\ast }}^{\ast }\left( y^{\ast }\right) \end{align*}% with $x:=\mathrm{d}h_{y^{\ast }}^{\ast }$, see proof of Lemma \ref{theorem trivial sympa 2}. Let $\mathcal{W}$ be any open neighborhood of $% \{0\}\subset \mathcal{X}^{\ast }$. Then, for any $z\in K$, the set \begin{equation*} \{\delta ^{\ast }(z):\delta ^{\ast }\in \mathcal{W}\}\subset \mathbb{R} \end{equation*}% is bounded, by continuity of the linear map $\delta ^{\ast }\mapsto \delta ^{\ast }(z)$. From the the principle of uniform boundedness for compact convex sets, i.e., the version of the Banach--Steinhaus theorem stated, for instance, in \cite[Theorem 2.9]{Rudin}, the set% \begin{equation*} \{\delta ^{\ast }(z):\delta ^{\ast }\in \mathcal{W},\,z\in K\}\subset \mathbb{R} \end{equation*}% is also bounded. Thus, for any $z^{\ast }\in \mathcal{X}^{\ast }$,% \begin{eqnarray*} \lim_{s\searrow 0}\sup \left\{ |h^{\ast }\left( z^{\ast }\right) -h^{\ast }\left( z^{\ast }+\delta ^{\ast }\right) |\,:\,\delta ^{\ast }\in s\mathcal{W% }\right\} &=&0, \\ \lim_{s\searrow 0}\sup \left\{ |g_{\mathcal{V},x^{\ast }}^{\ast }\left( z^{\ast }\right) -g_{\mathcal{V},x^{\ast }}^{\ast }\left( z^{\ast }+\delta ^{\ast }\right) |\,:\,\delta ^{\ast }\in s\mathcal{W}\right\} &=&0. \end{eqnarray*}% This implies the continuity of $h^{\ast }$ and $g_{\mathcal{V},x^{\ast }}^{\ast }$. Hence, from the density of $\mathcal{Y}^{\ast }$, $h^{\ast }=g_{% \mathcal{V},x^{\ast }}^{\ast }$ on the open neighborhood $\left( x^{\ast }+% \mathcal{V}\right) $ of $\{x^{\ast }\}\subset \mathcal{X}^{\ast }$. In particular, $h^{\ast }$ and $g_{\mathcal{V},x^{\ast }}^{\ast }$ have the same Fenchel subgradients at the point $x^{\ast }$. From Theorems \ref{theorem trivial sympa 1 copy(1)} and \ref{theorem trivial sympa 3}, for each open neighborhood $\mathcal{V}$ of $\{0\}\subset \mathcal{X}^{\ast }$, the extreme Fenchel subgradients of $h^{\ast }$ at $x^{\ast }$ are all contained in the set $\mathcal{T}_{x^{\ast },\mathcal{V}}$ defined by (\ref{set t voisinage}% ). Corollary \ref{corollary explosion lanford-robinson} thus follows. \subsection{Proof of Corollary \protect\ref{corollary explosion lanford-robinson copy(1)}\label{Section Proofs-IV copy(1)}} Note that $h^{\ast \ast }=h$ because the function $h$ is continuous and convex. By the global Lipschitz continuity of $h$, \begin{equation*} h\left( x\right) =\underset{x^{\ast }\in \mathcal{X}^{\ast }}{\sup }\left\{ x^{\ast }\left( x\right) -h^{\ast }\left( x^{\ast }\right) \right\} =% \underset{x^{\ast }\in K}{\sup }\left\{ x^{\ast }\left( x\right) -h^{\ast }\left( x^{\ast }\right) \right\} \end{equation*}% with $K:=\mathcal{B}_{R}\left( 0\right) \subset \mathcal{X}^{\ast }$ being some ball of sufficiently large radius $R>0$ centered at $0$. The set $K$ is weak$^{\ast }$--compact, by the Banach--Alaoglu theorem. Now, for any fixed $x\in \mathcal{X}$ and all $x^{\ast }\in \mathcal{Z}% _{x}\subset \mathcal{X}^{\ast }$, by definition of the set $\mathcal{Z}_{x}$% , there is a net $\{x_{i}\}_{i\in I}$ in $\mathcal{Y}$ converging to $x$ with the property that the unique Fenchel subgradient $x_{i}^{\ast }:=\mathrm{d}h_{x_{i}}\in \mathcal{X}^{\ast }$ of $h$ at $x_{i}$ converges towards $x^{\ast }$ in the weak$^{\ast }$--topology. Therefore, by continuity of $h$, for any fixed $x\in \mathcal{X}$ and all $x^{\ast }\in \mathcal{Z}_{x}$,% \begin{equation*} h\left( x\right) =\underset{y^{\ast }\in \mathcal{X}^{\ast }}{\sup }\left\{ y^{\ast }\left( x\right) -h^{\ast }\left( y^{\ast }\right) \right\} =% \underset{I}{\lim }\ h\left( x_{i}\right) =\underset{I}{\lim }\left\{ x_{i}^{\ast }\left( x\right) -h^{\ast }\left( x_{i}^{\ast }\right) \right\} , \end{equation*}% with $\{x_{i}^{\ast }\}_{i\in I}$ converging to $x^{\ast }$. In other words,% \begin{equation*} \mathcal{Z}_{x}\subset \mathit{\Omega }\left( h^{\ast }-x,K\right) , \end{equation*}% see Definition \ref{gamm regularisation copy(4)}. Thus, by Theorem \ref% {theorem trivial sympa 3} and Corollary \ref{corollary explosion lanford-robinson}, it suffices to prove that $\mathcal{T}_{x}\subset \mathcal{Z}_{x}$. By density of $\mathcal{Y}$ in $\mathcal{X}$, observe that the set \begin{equation*} \mathcal{T}_{x,\mathcal{V}}:=\overline{\left\{ \mathrm{d}h_{y}:y\in \mathcal{% Y}\cap (x+\mathcal{V)}\right\} }^{\mathcal{X}^{\ast }}\subset \mathcal{X}^{\ast } \end{equation*}% is non--empty for any open neighborhood $\mathcal{V}$ of $\{0\}\subset \mathcal{X}$. Meanwhile, the weak$^{\ast }$--compact set $K$ is metrizable with respect to (w.r.t.) the weak$^{\ast }$--topology, by separability of $% \mathcal{X}$, see \cite[Theorem 3.16]{Rudin}. In particular, $K$ is sequentially compact and we can restrict ourselves to sequences instead of more general nets. In particular, by (\ref{set t voisinage})--(\ref{set t voinagebis}), one has \begin{equation} \mathcal{T}_{x}=\bigcap\limits_{n\in \mathbb{N}}\mathcal{T}_{x,\mathcal{B}% _{1/n}\left( 0\right) } \label{inclusion sup} \end{equation}% with $\mathcal{B}_{\delta }\left( x\right) $ being the ball (in $K$) of radius $\delta >0$ centered at $x\in K$. Here, $\mathcal{B}_{\delta }\left( x\right) $ is defined by (\ref{ball}) for any metric $d$ on $K$ generating its weak$^{\ast }$--topology. For any $x^{\ast }\in \mathcal{T}_{x}\subset K$ and any $n\in \mathbb{N}$, there are per definition a sequence $% \{x_{n,m}^{\ast }\}_{m=1}^{\infty }$ converging to $x^{\ast }$ in $K$ as $% m\rightarrow \infty $ and an integer $N_{n}>0$ such that, for all $m\geq N_{n}$, $d(x^{\ast },x_{n,m}^{\ast })\leq 2^{-n}$ and $x_{n,m}^{\ast }=% \mathrm{d}h_{x_{n,m}}$ for some $x_{n,m}\in \mathcal{Y}\cap \lbrack x+% \mathcal{B}_{1/n}\left( 0\right) ]$. Taking any function $p(n)\in \mathbb{N}$ satisfying $p(n)>N_{n}$ and converging to $\infty $ as $n\rightarrow \infty $ we obtain a sequence $\{x_{n,p(n)}^{\ast }\}_{n=1}^{\infty }$ converging to $% x^{\ast }\in \mathcal{Z}_{x}$ as $n\rightarrow \infty $. This yields the inclusion $\mathcal{T}_{x}\subset \mathcal{Z}_{x}$. \section{Further Remarks\label{Concluding remarks}} We give here an additional observation which is\ not necessarily directly related to the main results of the paper. It concerns an extension of the Bauer maximum principle \cite[Theorem I.5.3.]{Alfsen}. See \cite{BruPedra2} for an application to statistical mechanics. First, recall that the $\Gamma $--regularization $\Gamma \left( h\right) $ of an extended real--valued function $h$ is a convex and lower semi--continuous function on a compact convex subset $K$. Moreover, every convex and lower semi--continuous function on $K$ equals its own $\Gamma $--regularization on $K$ (see, e.g., \cite[Proposition I.1.2.]{Alfsen}): \begin{proposition}[$\Gamma $--regularization of lower semi--cont. conv. maps% ] \label{lemma gamma regularisation}Let $h$ be any extended function from a (non--empty) compact convex subset $K\subset \mathcal{X}$ of a locally convex real space $\mathcal{X}$ to $\left( -\infty ,\infty \right] $. Then the following statements are equivalent:\newline \emph{(i)} $\Gamma \left( h\right) =h$ on $K$.\newline \emph{(ii)} $h$ is a lower semi--continuous convex function on $K$. \end{proposition} \noindent This proposition is a standard result. The compactness of $K$ is in fact not necessary but $K$ should be a closed convex set. This result can directly be proven without using the fact that the $\Gamma $--regularization $\Gamma \left( h\right) $ of a function $h$ on $K$ equals its twofold \emph{Legendre--Fenchel transform} -- also called the \emph{biconjugate }% (function) of $h$. Indeed, $\Gamma \left( h\right) $ is the largest lower semi--continuous and convex minorant of $h$: \begin{corollary}[Largest lower semi--cont. convex minorant of $h$] \label{Biconjugate}\mbox{ }\newline Let $h$ be any extended function from a (non--empty) compact convex subset $% K\subset \mathcal{X}$ of a locally convex real space $\mathcal{X}$ to $% \left( -\infty ,\infty \right] $. Then its $\Gamma $--regularization $\Gamma \left( h\right) $ is its largest lower semi--continuous convex minorant on $% K $. \end{corollary} \textit{Proof. }For any lower semi--continuous convex extended real--valued function $f$ defined on $K$ satisfying $f\leq h$, we have, by Proposition \ref{lemma gamma regularisation}, that \begin{equation*} f\left( x\right) =\sup \left\{ m(x):m\in \mathrm{A}\left( \mathcal{X}\right) \;\RIfM@\expandafter\text@\else\expandafter\mbox\fi{and }m|_{K}\leq f\leq h\right\} \leq \Gamma \left( h\right) \left( x\right) \end{equation*}% for any $x\in K$. \hfill\qed\medbreak% \noindent In particular, if $(\mathcal{X},\mathcal{X}^{\ast })$ is a dual pair and $h \not\equiv \infty$ is any extended function from $K$ to $(-\infty ,\infty ]$ then $\Gamma \left( h\right) =h^{\ast \ast }$, see \cite[Proposition 51.6]{Zeidler3}. Proposition \ref{lemma gamma regularisation} has another interesting consequence: An extension of the Bauer maximum principle \cite[Theorem I.5.3.% ]{Alfsen} which, in the case of convex functions, is: \begin{lemma}[Bauer maximum principle] \label{Bauer maximum principle}\mbox{ }\newline Let $\mathcal{X}$ be a locally convex real space. An upper semi--continuous convex real--valued function $h$ over a compact convex subset $K\subset \mathcal{X} $ attains its maximum at an extreme point of $K$, i.e., \begin{equation*} \sup \,h\left( K\right) =\max \,h\left( \mathcal{E}(K)\right) . \end{equation*}% Here, $\mathcal{E}(K)$ is the (non--empty) set of extreme points of $K$. \end{lemma} \noindent Indeed, by combining Proposition \ref{lemma gamma regularisation} with Lemma \ref{Bauer maximum principle} it is straightforward to check the following statement which does not seem to have been observed before: \begin{lemma}[Extension of the Bauer maximum principle] \label{Bauer maximum principle bis}\mbox{ }\newline Let $h_{\pm }$ be two convex real--valued functions from a locally convex real space $\mathcal{X}$ to $\left( -\infty ,\infty \right] $ such that $h_{-}$ and $h_{+}$ are respectively lower and upper semi--continuous. Then the supremum of the sum $h:=h_{-}+h_{+}$ over a compact convex subset $K\subset \mathcal{X}$ can be reduced to the (non--empty) set $\mathcal{E}(K)$ of extreme points of $K$, i.e., \begin{equation*} \sup \,h\left( K\right) =\sup \,h\left( \mathcal{E}(K)\right) . \end{equation*} \end{lemma} \textit{Proof. }We first use Proposition \ref{lemma gamma regularisation} in order to write $h_{-}=\Gamma \left( h_{-}\right) $ as a supremum over affine and continuous functions. Then we commute this supremum with the one over $% K$ and apply the Bauer maximum principle to obtain that% \begin{equation*} \sup \,h\left( K\right) =\sup \left\{ \sup \,\left[ m+h_{+}\right] (\mathcal{% E}(K)):m\in \mathrm{A}\left( \mathcal{X}\right) \;\RIfM@\expandafter\text@\else\expandafter\mbox\fi{and }m|_{K}\leq h_{-}|_{K}\right\} . \end{equation*}% The lemma follows by commuting again both suprema and by using $h_{-}=\Gamma \left( h_{-}\right) $. \hfill\qed\medbreak% \noindent Observe, however, that under the conditions of the lemma above, the supremum of $h=h_{-}+h_{+}$ is generally not attained on $\mathcal{E}(K)$% . \section{Appendix\label{Section appendix}} \setcounter{equation}{0}% For the reader's convenience we give here a short review on the following subjects: \begin{itemize} \item Dual pairs of locally convex real spaces, see, e.g., \cite{Rudin}; \item Barycenters and $\Gamma $--regularization of real--valued functions, see, e.g., \cite{Alfsen}; \item The Mazur and Lanford III--Robinson theorems, see \cite{LanRob,Mazur}. \end{itemize} \noindent These subjects are rather standard. Therefore, we keep the exposition as short as possible and only concentrate on results used in this paper. \subsection{Dual Pairs of Locally Convex Real Spaces} The notion of \emph{dual pairs} is defined as follow: \begin{definition}[Dual pairs] \label{dual pairs}\mbox{ }\newline For any locally convex space $(\mathcal{X},\tau )$, let $\mathcal{X}^{\ast }$ be its dual space, i.e., the set of all continuous linear functionals on $% \mathcal{X}$. Let $\tau ^{\ast }$ be any locally convex topology on $% \mathcal{X}^{\ast }$. $(\mathcal{X},\mathcal{X}^{\ast })$ is called a dual pair iff, for all $x\in \mathcal{X}$, the functional $x^{\ast }\mapsto x^{\ast }(x)$ on $\mathcal{X}^{\ast }$ is continuous w.r.t. $\tau ^{\ast }$, and all linear functionals which are continuous w.r.t. $\tau ^{\ast }$ have this form. \end{definition} \noindent By \cite[Theorems 3.4 (b) and 3.10]{Rudin}, a typical example of a dual pair $(\mathcal{X},\mathcal{X}^{\ast })$ is given by any locally convex real space $\mathcal{X}$ equipped with a topology $\tau $ and $\mathcal{X}% ^{\ast }$ equipped with the $\sigma (X^{\ast },X)$--topology $\tau ^{\ast }$% , i.e., the weak$^{\ast }$--topology. We also observe that if $(\mathcal{X},% \mathcal{X}^{\ast })$ is a dual pair w.r.t. $\tau $ and $\tau ^{\ast }$ then $(\mathcal{X}^{\ast },\mathcal{X})$ is a dual pair w.r.t. $\tau ^{\ast }$ and $\tau $. \subsection{Barycenters and $\Gamma $--regularization} The theory of compact convex subsets of a locally convex real (topological vector) space $\mathcal{X}$ is standard. For more details, see, e.g., \cite% {Alfsen}. An important observation is the Krein--Milman theorem (see, e.g., \cite[Theorems 3.4 (b) and 3.23]{Rudin}) which states that any compact convex subset $K\subset \mathcal{X}$ is the closure of the convex hull of the (non--empty) set $\mathcal{E}(K)$ of its extreme points. Restricted to finite dimensions this theorem corresponds to a classical result of Minkowski which, for any $x\in K$ in a (non--empty) compact convex subset $% K\subset \mathcal{X}$, states the existence of a finite number of extreme points $\hat{x}_{1},\ldots ,\hat{x}_{k}\in \mathcal{E}(K)$ and positive numbers $\mu _{1},\ldots ,\mu _{k}\geq 0$ with $\Sigma _{j=1}^{k}\mu _{j}=1$ such that \begin{equation} x=\overset{k}{\sum\limits_{j=1}}\mu _{j}\hat{x}_{j}. \label{barycenter1} \end{equation}% To this\ simple decomposition we can associate a probability measure, i.e., a \emph{normalized positive Borel regular measure}, $\mu $\ on $K$. Borel sets of any set $K$ are elements of the $\sigma $--algebra $\mathfrak{B% }$ generated by closed -- or open -- subsets of $K$. Positive Borel regular measures are the positive countably additive set functions $\mu $ over $% \mathfrak{B}$ satisfying \begin{equation*} \mu \left( B\right) =\sup \left\{ \mu \left( C\right) :C\subset B,\RIfM@\expandafter\text@\else\expandafter\mbox\fi{ }C% \RIfM@\expandafter\text@\else\expandafter\mbox\fi{ closed}\right\} =\inf \left\{ \mu \left( O\right) :B\subset O,\RIfM@\expandafter\text@\else\expandafter\mbox\fi{ }% O\RIfM@\expandafter\text@\else\expandafter\mbox\fi{ open}\right\} \end{equation*}% for any Borel subset $B\in \mathfrak{B}$ of $K$. If $K$ is compact then any positive Borel regular measure $\mu $ (one--to--one) corresponds to an element of the set $M^{+}(K)$ of Radon measures with $\mu \left( K\right) =\left\Vert \mu \right\Vert $, and we write \begin{equation} \mu \left( h\right) =\int_{K}\mathrm{d}\mu (\hat{x})\;h\left( \hat{x}\right) \label{barycenter1bis} \end{equation}% for any continuous function $h$ on $K$. A probability measure $\mu \in M_{1}^{+}(K)$ is per definition a positive Borel regular measure $\mu \in M^{+}(K)$ which is \emph{normalized}: $\left\Vert \mu \right\Vert =1$. Therefore, using the probability measure $\mu _{x}\in M_{1}^{+}(K)$\ on $K$ defined by \begin{equation*} \mu _{x}=\overset{k}{\sum\limits_{j=1}}\mu _{j}\delta _{\hat{x}_{j}} \end{equation*}% with $\delta _{y}$ being the Dirac -- or point -- mass\footnote{$\delta _{y}$ is the Borel measure such that, for any Borel subset $B\in \mathfrak{B}$ of $% K$, $\delta _{y}(B)=1$ if $y\in B$ and $\delta _{y}(B)=0$ if $y\notin B$.} at $y$, Equation (\ref{barycenter1}) can be seen as an integral defined by (% \ref{barycenter1bis}) for the probability measure $\mu _{x}\in M_{1}^{+}(K)$% : \begin{equation} x=\int_{K}\mathrm{d}\mu _{x}(\hat{x})\;\hat{x}\ . \label{barycenter2} \end{equation}% The point $x$ is in fact the \emph{barycenter} of the probability measure $% \mu _{x}$. This notion is defined in the general case as follows (cf. \cite[% Eq. (2.7) in Chapter I]{Alfsen}): \begin{definition}[Barycenters of probability measures in convex sets] \label{def barycenter}Let $K\subset \mathcal{X}$ be any (non--empty) compact convex subset of a locally convex real space $\mathcal{X}$ and let $\mu \in M_{1}^{+}(K)$ be a probability measure on $K$. We say that $x\in K$ is the barycenter\footnote{% Other terminologies existing in the literature: \textquotedblleft $x$ is represented by $\mu $\textquotedblright , \textquotedblleft $x$ is the resultant of $\mu $\textquotedblright .} of $\mu $ if, for all $z^{\ast }\in \mathcal{X}^{\ast }$, \begin{equation*} z^{\ast }\left( x\right) =\int_{K}\mathrm{d}\mu (\hat{x})\;z^{\ast }\left( \hat{x}\right) . \end{equation*} \end{definition} \noindent Barycenters are well--defined for \emph{all} probability measures in convex compact subsets of locally convex real spaces (cf. \cite[Theorems 3.4 (b) and 3.28]{Rudin}): \begin{theorem}[Well-definiteness and uniqueness of barycenters] \label{thm barycenter}\mbox{ }\newline Let $K\subset \mathcal{X}$ be any (non--empty) compact subset of a locally convex real space $\mathcal{X}$ such that $\overline{\mathrm{co}}\left( K\right) $ is also compact. Then, for any probability measure $\mu \in M_{1}^{+}(K)$ on $K$,$\ $there is a unique barycenter $x_{\mu }\in \overline{% \mathrm{co}}\left( K\right) $. \end{theorem} \noindent Note that Barycenters can also be defined in the same way via affine continuous functions instead of continuous linear functionals, see, e.g., \cite[Proposition I.2.2.]{Alfsen} together with \cite[Theorem 1.12]% {Rudin}. It is natural to ask whether, for any $x\in K$ in the compact convex set $K$% , there is a (possibly not unique) probability measure $\mu _{x}$ on $K$ (pseudo--) supported on $\mathcal{E}(K)$ with barycenter $x$. Equation (\ref% {barycenter2}) already gives a first positive answer to that problem in the finite dimensional case. The general case, which is a remarkable refinement of the Krein--Milman theorem, has been proven by Choquet--Bishop--de Leeuw (see, e.g., \cite[Theorem I.4.8.]{Alfsen}). We conclude now by a crucial property concerning the $\Gamma $% --regularization of extended real--valued functions in relation with the concept of barycenters (cf. \cite[Corollary I.3.6.]{Alfsen}): \begin{theorem}[Barycenters and $\Gamma $--regularization] \label{Thm - Corollary I.3.6}\mbox{ }\newline Let $K\subset \mathcal{X}$ be any (non--empty) compact convex subset of a locally convex real space $\mathcal{X}$ and $h:K \to \mathbb{R}$ be a continuous real--valued function. Then, for any $x\in K$, there is a probability measure $\mu _{x}\in M_{1}^{+}(K)$\ on $K$ with barycenter $x$ such that \begin{equation*} \Gamma \left( h\right) \left( x\right) =\int_{K}\mathrm{d}\mu _{x}(\hat{x}% )\;h\left( \hat{x}\right) . \end{equation*} \end{theorem} \noindent This theorem is a very important statement used to prove Theorem % \ref{theorem trivial sympa 1}. \subsection{The Mazur and Lanford III--Robinson Theorems} If $\mathcal{X}$ is a separable real Banach space and $h$ is a continuous convex real--valued function on $\mathcal{X}$ then it is well--known that $h$ has, on each point $x\in \mathcal{X}$, at least one Fenchel subgradient \textrm{d}$h\in \mathcal{X}^{\ast }$. The Mazur theorem describes the set $\mathcal{Y}$ on which a continuous convex function $h$ is G\^{a}teaux differentiable, more precisely, the set $\mathcal{Y}$ for which $h$ has exactly one Fenchel subgradient \textrm{d}$h_{x}\in \mathcal{X}^{\ast }$ at any $x\in \mathcal{Y}$: \begin{theorem}[Mazur] \label{Mazur}\mbox{ }\newline Let $\mathcal{X}$ be a separable real Banach space and let $h:\mathcal{X}% \rightarrow \mathbb{R}$ be a continuous convex function. The set $\mathcal{% Y}\subset \mathcal{X}$ of elements where $h$ has exactly one Fenchel subgradient \textrm{d}$h_{x}\in \mathcal{X}^{\ast }$ at any $x\in \mathcal{Y}$ is residual, i.e., a countable intersection of dense open sets. \end{theorem} \begin{remark} \label{Mazur remark}By Baire category theorem, the set $\mathcal{Y}$ is dense in $\mathcal{X}$. \end{remark} \noindent The Lanford III--Robinson theorem \cite[Theorem 1]{LanRob} completes the Mazur theorem by characterizing the Fenchel subdifferential $\partial h(x)\subset \mathcal{X}^{\ast }$ at any $x\in \mathcal{X}$: \begin{theorem}[Lanford III -- Robinson] \label{Land.Rob}\mbox{ }\newline Let $\mathcal{X}$ be a separable real Banach space and let $h:\mathcal{X}% \rightarrow \mathbb{R}$ be a continuous convex function. Then the Fenchel subdifferential $\partial h(x)\subset \mathcal{X}^{\ast }$ of $h$, at any $% x\in \mathcal{X}$, is the weak$^{\ast }$--closed convex hull of the set $% \mathcal{Z}_{x}$. Here, at fixed $x\in \mathcal{X}$, $\mathcal{Z}_{x}$ is the set of functionals $x^{\ast }\in \mathcal{X}^{\ast }$ such that there is a net $\{x_{i}\}_{i\in I}$ in $\mathcal{Y}$ converging to $x$ with the property that the unique Fenchel subgradient $\mathrm{d}h_{x_{i}}\in \mathcal{X}^{\ast }$ of $h$ at $x_{i}$ converges towards $x^{\ast }$ in the weak$^{\ast}$--topology. \end{theorem} \addcontentsline{toc}{section}{References}%

https://arxiv.org/abs/1802.04711

Quantum-classical correspondence in the vicinity of periodic orbits

Quantum-classical correspondence in chaotic systems is a long-standing problem. We describe a method to quantify Bohr's correspondence principle and calculate the size of quantum numbers for which we can expect to observe quantum-classical correspondence near periodic orbits of Floquet systems. Our method shows how the stability of classical periodic orbits affects quantum dynamics. We demonstrate our method by analyzing quantum-classical correspondence in the quantum kicked top (QKT), which exhibits both regular and chaotic behavior. We use our correspondence conditions to identify signatures of classical bifurcations even in a deep quantum regime. Our method can be used to explain the breakdown of quantum-classical correspondence in chaotic systems.

\section{\label{sec:level1}Introduction} Quantum-classical correspondence and the quantum-to-classical transition have been topics of fundamental interest since the birth of quantum theory in the early 20th Century. The connection between quantum and classical mechanics remains a partially understood subject, particularly in chaotic systems. According to the correspondence principle, the predictions of quantum physics should agree with the predictions of classical physics in appropriate limits wherever classical physics is applicable. There are multiple ways in which the correspondence principle has been formulated. These include Bohr's correspondence principle \cite{nielsen2013correspondence}, Ehrenfest's theorem \cite{Ehrenfest1927}, and Liouville correspondence \cite{Joshua1997a,Joshua1997b}, with each having its own subtleties \cite{Ballentine1994,gao1999breakdown,makowski2002bohr}. According to Bohr, quantum-classical correspondence is attained in the limit of large quantum numbers (or when $\hbar \rightarrow 0$ relative to the phase space of the dynamics). According to Ehrenfest's theorem, the evolution of expectation values of observables in a quantum system should coincide with the corresponding classical evolution until a time known as Ehrenfest's time ($t_{\text{EH}}$) that depends on the system dynamics. While $t_{\text{EH}}$ is large for regular systems in the semiclassical limit, it can be very small for chaotic systems even in the semiclassical limit \cite{berman1978,Zurek1994,Zurek1995}. Since Bohr's correspondence principle involves large quantum numbers, a natural question that arises is how large is large enough to see correspondence. Another important question is why correspondence breaks down in chaotic systems. In this paper, we address both these questions by analyzing the effect of stability and bifurcations of classical periodic orbits on quantum dynamics. We explore quantum-classical correspondence in the quantum kicked top (QKT) - a multiqubit time-periodic system that is a standard paradigm for exploring chaos \cite{haake1987}. This system is of particular interest because it displays bifurcations, regular behavior as well as chaotic behavior in the classical limit, and is one of the few chaotic systems that has been experimentally realized in the quantum regime \cite{chaudhury2009nature,neill2016ergodic}. Furthermore, since it is finite-dimensional, there are no truncation errors in the study of this system. Our study is based on an analysis of classical periodic orbits and their stability. We thus first present a classical periodic orbit analysis and bifurcation study of the kicked top. We then provide criteria for calculating the quantum number (in this case, the collective qubit spin $j$) for which the quantum dynamics of a state localized on a periodic orbit will correspond to the classical dynamics. Our criteria are based on the orthogonality of quantum states centered on the different points of the periodic orbits. When the criteria are satisfied, signatures of the classical bifurcations are clearly reflected in the quantum dynamics, even in a deep quantum regime. These signatures become more pronounced in a semiclassical regime. Furthermore, we show that in chaotic systems, the existence of orbits of very high periodicity can lead to a violation of our criteria and thus result in a short Ehrenfest break time. Studies of quantum-classical correspondence for systems with a mixed phase space of periodic islands and chaotic regions are more challenging compared to purely regular or chaotic systems. Our approach is thus particularly useful in the analysis of such mixed systems. The paper is organized as follows. In Sec. \ref{sec:level2}, we briefly describe the quantum kicked top model. Section \ref{sec:level3} includes a classical analysis of the kicked top with explicit calculations of periodic orbits and bifurcations. In Sec. \ref{sec:level4}, we describe our criteria for the quantification of Bohr's correspondence principle in periodic Floquet systems. In Sec. \ref{sec:level5}, we apply our criteria to the QKT. We show that when the criteria are satisfied, quantum-classical correspondence is evident even in a deep quantum regime. We also illustrate the effect of instability of classical periodic orbits on the quantum dynamics. In Sec. \ref{sec:level6}, we use our criteria to identify new quantum signatures of classical bifurcations in the kicked top dynamics, in a deep quantum regime as well as in the semiclassical regime. In Sec. \ref{sec:level7}, we discuss how our criteria can be used to explain the divergence of quantum and classical dynamics in chaotic systems. Finally, we present a summary of our results in Sec. \ref{sec:level8}. \section{\label{sec:level2}Background} \subsection{\label{sec:level2a}The quantum kicked top} The quantum kicked top is a time-dependent periodic system governed by the Hamiltonian \cite{haake1987} \begin{equation} H = \hbar \frac{\kappa}{2 j \tau} J_z^2 + \hbar p J_y \sum_{n= - \infty}^{\infty} \delta (t - n \tau), \label{top1} \end{equation} where $J_x,J_y$ and $J_z$ are angular momentum operators. Since the square of the angular momentum operator commutes with the Hamiltonian ($[H,J^2] = 0$), its eigenvalue $j(j+1)\hbar^2$, and thus $j$, is a constant of motion. The Floquet time evolution operator for one time period, $\tau$, is \begin{equation} U=\exp{(-i \frac{\kappa}{2j \tau}J_z^2)} \exp{(-ipJ_y)}. \label{top2} \end{equation} Each time period consists of a linear rotation by angle $p$ about the $y$ axis and a nonlinear rotation about the $z$ axis. The classical dynamics can be obtained by writing the Heisenberg equations of motion for the angular momentum operators and then taking the limit $j \rightarrow \infty$ \cite{haake1987}. Setting $X=J_x/j$, $Y=J_y/j$ and $Z=J_z/j$, the classical equations of motion for $p=\pi/2$ are \begin{eqnarray} X((n+1)\tau) & = & Z(n\tau) \cos{(\kappa X(n\tau))} + Y(n\tau) \sin{(\kappa X(n\tau))}, \nonumber \\ Y((n+1)\tau) & = & -Z(n\tau) \sin{(\kappa X(n\tau))} + Y(n\tau) \cos{(\kappa X(n\tau))}, \nonumber \\ Z((n+1)\tau) & = & -X(n\tau). \end{eqnarray} As the chaoticity parameter, $\kappa$, is varied from 0 to 7, the classical dynamics ranges from fully regular motion (for $\kappa \lesssim 2.1 $) to a mixture of regular and chaotic behavior for different initial conditions (for $ 2.1 \lesssim \kappa \lesssim 4.4$) to fully chaotic motion (for $\kappa \gtrsim 4.4$). The classical stroboscopic map (in polar co-ordinates) for a range of initial conditions with $\kappa=3$ is shown in Fig. \ref{phase_space}. \begin{figure}[t] \centering{\includegraphics[scale=0.6]{Kicked_top_phase_space_k_3.pdf}} \caption{Classical stroboscopic phase space of the kicked top for parameter values $\kappa = 3.0$, $\tau = 1.0, p = \frac{\pi}{2}$. $\theta$ and $\phi$ are plotted after each kick for 1360 initial conditions, each evolved for 150 kicks.} \label{phase_space} \end{figure} Quantum mechanically, we can view the quantum kicked top as a multiqubit system of $N=2j$ qubits. The symmetric subspace of the $2j$ qubits has constant angular momentum $j$, and the kicked top evolves in this subspace spanned by the eigenstates of $J^2$ and $J_z$, $|j,m\rangle, m=\{-j,-j-1,....j\}$. \subsection{\label{sec:level2b}Spin coherent states (SCS)} Coherent states are minimum uncertainty states and are thus the closest quantum analog of classical states. For spin systems, these are the so-called spin coherent states (SCS) \cite{Arecchi1972}. Given any point $(\theta, \phi)$ in the classical phase space, we can construct SCS states $|\theta,\phi \rangle$ by applying the rotation operator $R(\theta,\phi) = \exp{[i \theta(J_x \sin{\phi} - J_y \cos{\phi})]}$ on the state $|j,j\rangle$, $|\theta, \phi \rangle = R(\theta, \phi)|j,j \rangle$. This yields a minimum uncertainty state centered on the point $(\theta, \phi)$, that is, the expectation value of the angular momentum of this state is $(j \sin{\theta}\cos{\phi}, j \sin{\theta}\sin{\phi}, j \cos{\theta})$. The uncertainty of this state is $(\langle \textbf{J}^2 \rangle - \langle \textbf{J} \rangle^2) / j^2 = 1/j$. Thus, for larger $j$ values, the SCS becomes highly localized at the point $(\theta, \phi)$ in phase space and better approximates the classical states. \subsection{\label{sec:level2c}Husimi phase space distribution} We study the quantum evolution of the kicked top in phase space using the Husimi phase space distribution function \cite{HusimiDist1981}. Given any angular momentum quantum state $\rho$, the Husimi distribution is given by, \begin{equation} Q(\theta,\phi) = \frac{2J+1}{4 \pi} \langle \theta, \phi |\rho | \theta, \phi \rangle, \label{husimi1} \end{equation} which is equal to $\frac{2J+1}{4 \pi} |\langle \theta, \phi | \psi \rangle|^2$ for pure states. \section{\label{sec:level3}Classical analysis of the kicked top: Periodic orbits and bifurcations} We begin by analyzing the existence and stability of classical fixed points and some 2-periodic and 4-periodic orbits as the parameter $\kappa$ is varied with $p=\pi/2$. To study the stability of a period-$n$ orbit, we calculate eigenvalues of the Jacobian of $F^n$ at the period-$n$ point, where $\left(X((n+1)\tau),Y((n+1)\tau),Z((n+1)\tau)\right) = F(X(n\tau),Y(n\tau),Z(n\tau))$. If $|\lambda| \leq 1$ $\forall \lambda$, then the period-$n$ orbit is stable, otherwise it is unstable. Table \ref{CP} lists a few interesting fixed points and periodic orbits of the kicked top. The variable $x_0$ in $FP_1$, $FP_2$ and $P2_A$ in Table \ref{CP} is obtained from the normalization condition, \begin{equation} 2x_0^2 + \frac{[x_0 \sin{(\kappa x_0)}]^2}{[1-\cos{(\kappa x_0)]^2}} = 1. \label{normalization} \end{equation} \begin{table*}[t] \begin{tabular}{ccl} $FP_1$ & = & $(0,1,0)$ \\ $FP_2$ & = & $(0,-1,0)$ \\ $FP_3$ & = & $\left(x_0,\dfrac{x_0 \sin{(\kappa x_0)}}{(1-\cos{(\kappa x_0))}},-x_0 \right)$ \\ $FP_4$ & = & $\left(-x_0,\dfrac{x_0 \sin{(\kappa x_0)}}{(1-\cos{(\kappa x_0))}},x_0 \right)$ \\ $P2_A$ & = & $\left(x_0,-\dfrac{x_0 \sin{(\kappa x_0)}}{(1-\cos{(\kappa x_0))}},x_0 \right) \leftrightarrow \left(-x_0,-\dfrac{x_0 \sin{(\kappa x_0)}}{(1-\cos{(\kappa x_0))}},-x_0 \right)$ \\ $P4$ & = & $(1,0,0) \rightarrow (0,0,-1) \rightarrow (-1,0,0) \rightarrow (0,0,1) \rightarrow (1,0,0)$ \\ $P2_B$ & = & $\left(\dfrac{\pi}{\kappa},\sqrt{1-2(\dfrac{\pi}{\kappa})^2},\dfrac{\pi}{\kappa} \right) \leftrightarrow \left(-\dfrac{\pi}{\kappa},-\sqrt{1-2(\dfrac{\pi}{\kappa})^2},-\dfrac{\pi}{\kappa} \right)$ \\ $P2_C$ & = & $\left(-\dfrac{\pi}{\kappa},\sqrt{1-2(\dfrac{\pi}{\kappa})^2},\dfrac{\pi}{\kappa} \right) \leftrightarrow \left(-\dfrac{\pi}{\kappa},-\sqrt{1-2(\dfrac{\pi}{\kappa})^2},\dfrac{\pi}{\kappa} \right)$ \\ $P2_D$ & = & $\left(\dfrac{\pi}{\kappa},\sqrt{1-2(\dfrac{\pi}{\kappa})^2},-\dfrac{\pi}{\kappa} \right) \leftrightarrow \left(\dfrac{\pi}{\kappa},-\sqrt{1-2(\dfrac{\pi}{\kappa})^2},-\dfrac{\pi}{\kappa} \right)$ \\ $P2_E$ & = & $\left(\dfrac{\pi}{\kappa},-\sqrt{1-2(\dfrac{\pi}{\kappa})^2},\dfrac{\pi}{\kappa}\right) \leftrightarrow \left(-\dfrac{\pi}{\kappa},\sqrt{1-2(\dfrac{\pi}{\kappa})^2},-\dfrac{\pi}{\kappa} \right)$ \end{tabular} \caption{Fixed points and periodic orbits of kicked top (in the form $(X,Y,Z)$)} \label{CP} \end{table*} $FP_1$ and $FP_2$ are fixed points for all values of $\kappa$. The eigenvalues of the Jacobian at $(0,1,0)$ are $\left(1,\dfrac{\kappa+\sqrt{\kappa^2-4}}{2},\dfrac{\kappa-\sqrt{\kappa^2-4}}{2}\right)$. Clearly, for $\kappa > 2$, the eigenvalue, $\dfrac{\kappa+\sqrt{\kappa^2-4}}{2} > 1$. Thus, this fixed point loses stability at $\kappa=2$, which implies that $\kappa=2$ is a bifurcation point. \begin{figure} \centering \includegraphics[width=0.5\textwidth]{Combine_bifurcation_point_plots.pdf} \caption{(a) Largest eigenvalue of the Jacobian matrix of $F$ at $FP_1$ as a function of $\kappa$ showing loss of stability of $FP_1$ at $\kappa=2$. (b) Largest eigenvalue of the Jacobian matrix of $F$ at $FP_3$ as a function of $\kappa$ showing loss of stability of $FP_3$ at $\kappa= \sqrt{2}\pi$. (c) Largest eigenvalue of the Jacobian matrix of $F^2$ at $P2_B$ as a function of $\kappa$ showing loss of stability of $P2_B$ at $\kappa \approx 4.8725$. (d) Largest eigenvalue of the Jacobian matrix of $F^4$ at $P4$ as a function of $\kappa$ showing loss of stability of $P4$ at $\kappa =\pi$. Vertical dashed lines in the four plots represent the parameter value at which loss of stability occurs.} \label{bifurcation1} \end{figure} At $\kappa=2$, $FP_1$ gives rise to two fixed points: $FP_3$ and $FP_4$. $FP_2$ becomes a period-2 orbit, $P2_A$. $FP_3$, $FP_4$ and $P2_A$ lose stability at $\kappa= \sqrt{2}\pi$ (Fig.~\ref{bifurcation1}) and give rise to four stable period-2 orbits: $P2_B, P2_C, P2_D \mbox{ and } P2_E$. These four period-2 orbits lose stability at $\kappa \approx 4.8725$ (Fig.~\ref{bifurcation1}). There exists a period-4 orbit, $P4$, at all values of $\kappa$. It loses its stability at $\kappa=\pi$ (Fig. \ref{bifurcation1}). Figure ~\ref{bifur_diagram} shows the bifurcation diagram for the mentioned periodic orbits, explicitly showing the bifurcation points $\kappa =2, \pi, \sqrt{2}\pi, 4.8725$. \begin{figure} \centering\includegraphics[width=0.5\textwidth]{Combine_bifurcation_plots_wout_title_in_plots.pdf} \caption{Classical bifurcation diagram (a) $\theta$ vs $\kappa$, (b) $\phi$ vs $\kappa$. All solid lines represent stable fixed points and periodic orbits. Dashed lines represent unstable fixed points and periodic orbits. Dashed dotted vertical lines represent bifurcation parameter values.} \label{bifur_diagram} \end{figure} \section{\label{sec:level4}Quantifying Bohr's correspondence principle} Bohr's correspondence principle broadly states that quantum dynamics will approach classical dynamics in the limit of large quantum numbers. Our goal is to quantify how large the quantum numbers need to be to observe similarity in classical and quantum dynamics. Here, we provide a quantification method based on periodic orbits for Floquet systems, which are periodically driven systems. We propose that the quantum dynamics in the vicinity of any classical period-$n$ orbit for such systems will be similar to the classical dynamics when: \begin{enumerate} \item the coherent states centered on all the $n$ points in the period-$n$ orbit are orthogonal to each other. \item the coherent states centered on multiple periodic orbits that are related by the symmetries of the system are orthogonal to each other. \end{enumerate} We note that the existence of symmetries in the system may lead to quantum mechanical phenomena between the periodic orbits related by these symmetries, such as, for example, dynamical tunneling \cite{davis1981quantum,keshavamurthy2011dynamical}. If so, then the conditions described above will not be sufficient to ensure correspondence in a deeply quantum regime. The two conjectured criteria above can be understood in the following way. In the limit where the classical states are distinguishable points in phase space, the classical dynamics evolves in a localized manner among these classical states in the phase space. At the quantum level, distinguishability is associated with orthogonality of quantum states. Consider the quantum dynamics of the same system with the initial state being a coherent state localized at one of the points in a stable period-$n$ orbit. We would expect the quantum dynamics to be similar to the classical dynamics if the quantum state evolves in a localized manner similar to the classical evolution. This localized evolution could occur if at any time in the evolution, the quantum state has high overlap with a coherent state centered at one of the classical points of the period-$n$ orbit and negligible overlap with coherent states centered on the rest of the points. This can be assured if the set of coherent states centered at the points of a period-$n$ orbit form an orthogonal set. However, if this set is a nonorthogonal set, then high survival probability at any classical point may still allow a significant amount of survival probability at other classical points in the period-$n$ orbit as well and thus may generally cause a departure from classical dynamics. The richness of the classical phase space determines the quantum number (or the effective Planck's constant value) at which there will be a correspondence between classical and quantum dynamics. For lower quantum numbers, the size of the Hilbert space restricts the dimension of any set that consists of states orthogonal to each other. Therefore, in systems whose classical phase space has only fixed points and period-$n$ orbits with small $n$, correspondence in the deep quantum regime is more likely to occur. We will illustrate our conjecture in the kicked top in the next section. The overlap between any two spin coherent states, $| \theta, \phi \rangle$ and $| \theta_0, \phi_0 \rangle $, is given by \cite{lombardi2011}: \begin{equation} |\langle \theta, \phi | \theta_0, \phi_0 \rangle| = \left(\cos{ \left[\frac{\chi(\theta \phi, \theta_0 \phi_0)}{2} \right]} \right)^{2J}, \label{overlap1} \end{equation} where $\chi(\theta \phi, \theta_0 \phi_0)$ is the angle between the direction vectors, $(\theta, \phi)$ and $(\theta_0, \phi_0)$ on the unit sphere, $\mathbb{S}^2$. Thus, Eq.~(\ref{overlap1}) is a handy tool to calculate the orthogonality of the spin coherent states for our quantification criteria. \section{\label{sec:level5}Quantum versus classical dynamics of the kicked top} \begin{figure*} \centering \includegraphics[scale=1.53]{Kappa_1pt5_plots.pdf} \caption{Evolution of the Husimi phase space distribution of an SCS centered on $FP_1$ [$(X,Y,Z)=(0,1,0)$] for three different $j$ values with $\kappa=1.5$. Like the classical dynamics, the quantum dynamics remains localized at $FP_1$, even in a deeply quantum regime, except for $j=2$ when dynamical tunneling to $FP_2$ occurs.} \label{k1} \end{figure*} In sec. \ref{sec:level3}, we showed that in the QKT, the number of fixed points and periodic orbits increases as the chaoticity parameter, $\kappa$, is increased (keeping the parameter $p=\pi/2$ fixed), owing to the many bifurcations. In this section, we illustrate our quantification criteria in the kicked top. \subsection{\label{sec:level5a}$\kappa<2$} For $\kappa <2$, the only periodic orbits in the classical phase space are $FP_1$, $FP_2$ and $P4$. $FP_1$ and $FP_2$ are isolated fixed points with no other periodic orbit in their vicinity in the phase space. The spin coherent states centered on $FP_1$ and $FP_2$ are orthogonal to each other for all $j$ values. Thus, we observe correspondence between the classical and quantum dynamics at these fixed points, $FP_1$ and $FP_2$, even for a very low quantum number, $j=1$, as illustrated in Fig. \ref{k1}. However, in this deep quantum regime, there is the possibility of dynamical tunneling since $FP_1$ and $FP_2$ are related by a symmetry of the square of the kicked top map for $p=\frac{\pi}{2}$, that is rotation by angle $\pi$ around the $x$-axis \cite{haake1987,Sanders1989}. This dynamical tunneling between the two fixed points, can be observed for some small values of $j$ (for example $j=2$ in Fig. \ref{k1}) but as the value of $j$ is increased further, the correspondence is recovered. In contrast, the spin coherent states centered on the four points in the $P4$ orbit are not orthogonal to each other for very small $j$ values. From Eq.(\ref{overlap1}), the overlap between the spin coherent states at any two consecutive points in this period-4 orbit is given by $\left(\frac{1}{\sqrt{2}} \right)^{2j}$, which is of the order $10^{-7}$ for $j=20$, and $\approx 0.156$ for $j=6$. Thus, for $j$ values $\lesssim 20$, we do not see quantum-classical correspondence if we start at any one of the period-4 points, but we do see correspondence for large enough $j$ values $\gtrsim 20$, as illustrated in Fig. \ref{k2}. \begin{figure*} \centering \includegraphics[scale=1.53]{Kappa_1pt5_period4_plots.pdf} \caption{Evolution of the Husimi phase space distribution of an SCS centered on a point in $P4$ ($(X,Y,Z)=(1,0,0)$) for 2 different $j$ values with $\kappa=1.5$. For $j=6$ (first row), the quantum dynamics does not correspond to the classical dynamics, but for $j=20$, there is clear correspondence.} \label{k2} \end{figure*} \subsection{\label{sec:level5b}$2 < \kappa< \pi$} \def .21 {.21} \begin{figure*} \centering \includegraphics[scale=1.53]{Kappa_2pt5_period2_after_bifur_at_2.pdf} \caption{Evolution of the Husimi phase space distribution of an SCS centered on a point in $P2_A$ for two different $j$ values with $\kappa=2.5$. For $j=10$ (top row), the quantum-classical correspondence is weak compared to $j=40$.} \label{k3} \end{figure*} In the range, $2 \leq \kappa< \pi$, we have two more fixed points, $FP_3$ and $FP_4$, and a period-2 orbit, $P2_A$, in addition to the ones present for $\kappa < 2$, as explained in Sec. \ref{sec:level2a}. $FP_1$ and $FP_2$ are unstable in this range while all others are stable. $FP_3$, $FP_4$ and $P2_A$ are functions of $\kappa$ (Table \ref{CP} ). For $\kappa=2.5$, the overlap between the spin coherent states centered on the two points in $P2_A$ is on the order of $10^{-4}$ for $j=10$, and on the order of $10^{-14}$ for $j=40$. Correspondingly, we see in Fig.~\ref{k3} that for $j=40$, the quantum dynamics follows the classical dynamics more closely, compared to $j=10$. \subsection{\label{sec:level5c}{Effect of classical instability}} \begin{figure} \centering \includegraphics[scale=0.8]{combine_figs_unstable_dynamics.pdf} \caption{Effect of classical instability on quantum dynamics. Left column: Classical trajectories for an initial state slightly perturbed from the unstable fixed point, $FP_1$, in Table \ref{CP}, for two different $\kappa$. Right column: Husimi phase space distribution (averaged over 100 kicks) corresponding to an initial $j=25$ SCS centered on $FP_1$, for two different $\kappa$. } \label{k4} \end{figure} Classical instability of any periodic orbit leads to exponential divergence of classical trajectories even for infinitesimally small differences in initial conditions. We show in Fig.~\ref{k4} that when the quantification criteria is satisfied, the quantum initial states localized at an unstable fixed point explores the same regions of the Husimi phase space as classical initial states slightly perturbed from an unstable fixed point would explore. This shows that classical instability affects classical as well as quantum dynamics in similar ways. We have checked this correspondence for the period-4 orbit, $P4$ in Table \ref{CP}, as well. The period-4 orbit is unstable for $\kappa>\pi$. The phase space is predominantly chaotic for $\kappa>3.5$ except for the regular islands of $FP_3, FP_4$ and $P2_A$. Thus, any classical initial state perturbed from the period-4 orbit explores most of the phase space avoiding the aforementioned regular islands. We have observed the same behavior for quantum initial states centered close to any point on the period-4 orbit for $\kappa>3.5$, that is, such a quantum state spreads out in the Husimi phase space avoiding the regions of classical regular islands. \section{\label{sec:level6}Quantum signatures of classical bifurcations} Given our new criteria for quantum-classical correspondence, we would ideally like to use it to identify classical bifurcation behavior (as shown in Fig. \ref{bifur_diagram}) in the quantum dynamics. To do so, we first define a measure of quantum dynamics that we can use to explore bifurcations. The survival probability of a quantum state, $|\psi(0) \rangle$, at time $t$, evolving according to a unitary operator, $U(t)$ is given by $|\langle \psi(0)|\psi(t) \rangle |^2$, where $|\psi(t)\rangle = U(t) |\psi(0) \rangle$. We analyze here the time-averaged survival probability of quantum states of the kicked top centered on any point of a classical period-$n$ orbit, where $n\geq 1$. \begin{enumerate} \item Given a classical fixed point, we compute the quantity, \begin{equation} S(L) = \frac{1}{L} \sum_{l=1}^L |\langle \psi(0)|\psi(l) \rangle |^2, \label{Period1} \end{equation} for some $L$, where $|\psi(l) \rangle =U^l |\psi(0) \rangle $, and $|\psi(0) \rangle$ is the SCS centered on the classical fixed point. Here, U is the unitary operator for one time period of the Floquet system. \item Given any classical period-$n$ orbit, if $F$ denotes the classical map, then each of the $n$ points of the period-$n$ orbit will be a fixed point of the map, $F^n$. Thus, we study the survival probability of an SCS centered on any point of a classical period-$n$ orbit using the unitary operator, $U^n$, instead of $U$. For a classical period-$n$ orbit, we compute the quantity \begin{equation} S(L) = \frac{1}{L} \sum_{l=1}^L |\langle \psi(0)|\psi(nl) \rangle |^2, \label{HighPeriod} \end{equation} for some $L$, where $|\psi(nl) \rangle =U^{nl} |\psi(0) \rangle $, and $|\psi(0) \rangle$ is the SCS centered at any point of the classical period-$n$ orbit. \end{enumerate} \begin{figure} \centering\includegraphics[scale=1.4]{combine_plots_survival_probability_fixed_point.pdf} \caption{(a) Survival probability of an SCS initially centered on $FP_1$ (in Table \ref{CP}), averaged over 50 kicks as a function of $j$ and $\kappa$. The horizontal line depicts the classical bifurcation. Darker color represents higher survival probability in this plot. (b) Survival probability of an SCS initially centered on $FP_1$, averaged over 200 kicks for $j=2000$ as a function of $\kappa$. The vertical dashed line represents the point of classical bifurcation.} \label{SurvFP1} \end{figure} \def \scalevalue {.58} \begin{figure} \centering\includegraphics[scale=1.4]{combine_plots_survival_probability_period2.pdf} \caption{(a) Survival probability of SCS initially centered on $P2_A$ (in Table \ref{CP}), averaged over 50 kicks as a function of $j$ and $\kappa$. The horizontal line depicts the classical bifurcation, and the dashed curve represents the $j$ value for a given $\kappa$ at which the overlap between the 2 SCS states corresponding to $P2_A$ is $\leq 10^{-10}$. Darker color represents higher survival probability in this plot. (b) Survival probability of SCS initially centered on $P2_A$ averaged over 100 kicks for $j=1000$ as a function of $\kappa$. The vertical dashed line represents the point of classical bifurcation.} \label{SurvP2A} \end{figure} \begin{figure} \centering\includegraphics[scale=1.4]{combine_plots_survival_probability_period4.pdf} \caption{(a) Survival probability [Eq.\ref{HighPeriod}] of SCS initially centered on $P4$ (in Table \ref{CP}), averaged over 50 kicks as a function of $j$ and $\kappa$. The horizontal line depicts the classical bifurcation, and the vertical dashed line represents the $j$ value at which the overlap between any two of the four SCS states corresponding to $P4$ is $\leq 10^{-8}$. Darker color represents higher survival probability in this plot. (b) Survival probability [Eq.\ref{HighPeriod}] of SCS initially centered on $P4$ averaged over 50 kicks for $j=1000$ as a function of $\kappa$. The vertical dashed line represents the point of classical bifurcation.} \label{SurvP4} \end{figure} We have plotted the survival probabilities corresponding to $FP_1$, $P2_A$ and $P4$ of Table \ref{CP} in Figs. \ref{SurvFP1}, \ref{SurvP2A} and \ref{SurvP4} respectively. Each figure consists of two plots, one illustrating signatures of bifurcation in the deep quantum regime, and the other in the semiclassical regime. (a) Analysis of $FP_1$ (Fig.~\ref{SurvFP1}): We see clear signatures of classical bifurcation of $FP_1$ at $\kappa=2$ in the survival probability plots in Fig.~\ref{SurvFP1} in a deep quantum regime as well as the semiclassical regime. The quantum state remains localized at the fixed point prior to bifurcation (because the fixed point is stable prior to bifurcation) and gets delocalized after bifurcation. (b) Analysis of $P2_A$ (Fig.~\ref{SurvP2A}): In Fig. \ref{SurvP2A}~(a) the classical bifurcation point (solid horizontal line) is easy to identify in the survival probability plot. Above this line, the survival probability is small, indicating that bifurcation has occurred. Below the horizontal line, however, there is some structure in the behavior of the survival probability. This can be understood in the following way. The two points associated with the period-2 orbit, $P2_A$ (Table \ref{CP}), are $\kappa$-dependent. Thus, the $j$ value at which the two SCS centered at these two points are orthogonal to each other is also $\kappa$-dependent. The dashed curve in Fig. \ref{SurvP2A}(a) represents the $j$ value at which the overlap between the two aforementioned SCS is less than $10^{-10}$ for the corresponding $\kappa$ values. Below this curve, we see small survival probability. This is because of mixing of the quantum dynamics between the two SCS states because they are not orthogonal to each other. Hence for $j$ values below the dashed curve the quantum dynamics does not mimic the corresponding classical dynamics in the period-2 orbit. Above the dashed curve, the quantum and classical dynamics should track, so there should be high survival probability (darker regions in the plot) below the classical bifurcation (solid line), and low survival probability above the bifurcation line. However there are also some lighter regions of low probability below the bifurcation line. One of the reasons for this is the quantum phenomenon of dynamical tunneling. Both the points of $P2_A$ are fixed points for the square of the classical map of the kicked top, thus allowing for dynamical tunneling between the two in addition to the period-2 motion between the two points. In Fig. \ref{SurvP2A}(b), for $j=1000$ (semiclassical regime), the bifurcation at $\kappa = \sqrt{2}\pi$ is clearly visible. The initial dip in the curve close to $\kappa = 2$ is because of non-zero overlap between the two aforementioned SCS states for $\kappa$ very close to 2. We also see a surprising dip in the survival probability around $\kappa=3.7$ though the $P2_A$ orbit is still stable. Further investigation of the classical phase space of the kicked top near this value of $\kappa$ reveals that a period-6 orbit arises near this period-2 orbit around $\kappa=3.62$. This period-6 orbit breaks off to the chaotic sea near $\kappa=3.68$, which results in the period-2 island in the phase space becoming smaller around $\kappa=3.68$. Thus, the wave packet centered at the period-2 orbit delocalizes to some extent in the phase space around $\kappa=3.68$. The size of the period-2 island increases again beyond $\kappa=3.72$ which results in a higher survival probability beyond $\kappa=3.72$ until bifurcation occurs. Dynamical tunneling also occurs around $\kappa=3.7$ to some extent, though the sum of the survival probability at the two points of the periodic orbit is not very close to 1 because of delocalization. These two points explain the dip at $\kappa=3.7$ in Fig. \ref{SurvP2A}(b). As $\kappa$ increases, we clearly observe very small survival probability after the classical bifurcation point in the deep quantum regime as well as the semiclassical regime. (c) Analysis of $P4$ (Fig.~\ref{SurvP4}): The overlap between each pair of the four points associated with the period-4 orbit, $P4$ (Table \ref{CP}), is less than $10^{-8}$ for $j\geq 27$. As explained for $P2_A$, mixing of dynamics can happen in $P4$ for small $j$ values. For larger $j$ compared to the critical value of $j$, we observe a clear signature of bifurcation in the survival probability plots. We also note some general observations about the signatures of classical bifurcations in the quantum dynamics. The quantum dynamics changes smoothly with the classical bifurcations, unlike the classical dynamics which shows a sudden change. When any local bifurcation occurs that gives rise to new fixed points or periodic orbits, these new orbits are usually close to each other in the classical phase space, due to which the corresponding coherent states are not orthogonal to each other. As the bifurcation parameter (external control parameter) is varied, these new orbits get further apart which decreases the overlap between the corresponding coherent states. Eventually, they may become orthogonal at which point the correspondence between classical and quantum dynamics near these orbits is restored (as long as other bifurcations do not occur prior to it). Alternatively, one could have increased the quantum number keeping the value of the external control parameter fixed. This explains why the quantum dynamics is affected smoothly by a classical bifurcation which gives rise to new fixed points or periodic orbits. Far from the classical bifurcation points, the stability of the classical fixed points and periodic orbits affects classical and quantum dynamics in the same way as long as the quantification criteria given in Sec. \ref{sec:level4} are satisfied. \section{\label{sec:level7}Quantum-classical correspondence in chaotic systems} The Ehrenfest break time after which quantum and classical dynamics diverge is very small for chaotic systems, even in the semiclassical limit. Using our quantification criteria in Sec. \ref{sec:level5}, we explain the reason for short break times in classically chaotic systems whose quantum counterpart is finite-dimensional. There are various routes to classical chaos, such as the period-doubling route and the intermittency route \cite{hilborn2000chaos}. In the period-doubling route to chaos, there exists at least one $2^n$ periodic orbit for every $n$ at the onset of chaos. According to our quantification criteria, if there exists a periodic orbit with, say, $r$ periodicity in a classical system, then the quantum system needs to be at least $r$-dimensional to exhibit a correspondence with the classical dynamics in the vicinity of that periodic orbit. This is because the coherent states corresponding to all the points in any period-$n$ orbit need to be orthogonal to each other for correspondence between classical and quantum dynamics. Since in a period-doubling route to chaos, there exists periodic orbits with at least one $2^n$ periodic orbit for every $n$ at the onset of chaos, any finite-dimensional quantum system cannot exhibit correspondence with the classical dynamics even in the semiclassical limit because the dimension of the set of orthogonal quantum states in such a system is restricted by the dimension of the basis of the corresponding Hilbert space. Since there will always exist periodic orbits with periodicity higher than the dimension of the Hilbert space at the onset of chaos, there cannot exist a good correspondence between the classical and quantum dynamics at the onset of chaos given that the route to chaos generates periodic orbits of unbounded periodicity. This explains why we have a short break time for chaotic systems. \section{\label{sec:level8}Conclusion} Quantum-classical correspondence for chaotic systems and for systems with a mixed phase space has remained a long-standing open question. Periodic orbits, and their stability and bifurcations play an important role in the transition from regular to chaotic behavior. Thus, gaining insight into quantum-classical correspondence in the vicinity of periodic orbits, and understanding the role of stability of periodic orbits and bifurcations on the quantum-classical correspondence is of vital importance. We have proposed the conditions under which the coherent states, which are the most classical states in quantum, evolve in close conjunction with classical dynamics for Floquet systems. We have applied our criteria to the quantum kicked top and showed how it can be used to quantify Bohr's correspondence principle. We note that in some situations quantum and classical dynamics may correspond even if our conditions are not met, but in general this will not be the case. Our studies of the kicked top seemed to indicate that such exceptions are not common. We have also illustrated the effect of classical instability on quantum evolution. Our analysis shows that the survival probability of quantum states centered on the periodic orbits exhibits signatures of classical bifurcations, given the aforementioned criteria are satisfied. Furthermore, we have used our criteria for quantum-classical correspondence to explain the reason for short break times between quantum and classical dynamics in chaotic systems. Signatures of chaos have been widely studied in the quantum kicked top using various quantum theoretic measures in the deep quantum regime as well as the semiclassical regime \cite{Fox1994,entanglement2004PRE,entanglement2004PRA, Chaos2008,lombardi2011,discord2015,santhanam2017,pattanayak2017}. Our analysis and criteria for quantum-classical correspondence can be applied to understand these previous results and develop better quantum control techniques. Furthermore, our criteria can be experimentally tested using current technology. \begin{acknowledgments} M. K. and S.G. acknowledge support from the Discovery Programme of the National Science and Engineering Research Council of Canada (NSERC). \end{acknowledgments}

https://arxiv.org/abs/2106.13780

Local stability of ground states in locally gapped and weakly interacting quantum spin systems

Based on a result by Yarotsky (J. Stat. Phys. 118, 2005), we prove that localized but otherwise arbitrary perturbations of weakly interacting quantum spin systems with uniformly gapped on-site terms change the ground state of such a system only locally, even if they close the spectral gap. We call this a strong version of the local perturbations perturb locally (LPPL) principle which is known to hold for much more general gapped systems, but only for perturbations that do not close the spectral gap of the Hamiltonian. We also extend this strong LPPL-principle to Hamiltonians that have the appropriate structure of gapped on-site terms and weak interactions only locally in some region of space.While our results are technically corollaries to a theorem of Yarotsky, we expect that the paradigm of systems with a locally gapped ground state that is completely insensitive to the form of the Hamiltonian elsewhere extends to other situations and has important physical consequences.

\section{Introduction} \label{sec:intro} We consider weakly interacting quantum spin systems on finite subsets~\(\Lambda \) of the lattice~\( \mathbb{Z} ^\nu\), \(\nu\in \mathbb{N} \), described by a self-adjoint Hamiltonian \begin{equation} \label{eq:introhamiltonian} H= H_0 + H_{\rm int} \,, \end{equation} which is composed of a non-interacting part~\(H_0\) and an interacting part~\(H_{\rm int}\). The non-interacting Hamiltonian~\(H_0\) is a sum of positive semi-definite on-site Hamiltonians~\(h_x\), \(x\in \Lambda\). Each~\(h_x\) is assumed to have a non-degenerate ground state with ground state energy~\(0\) and spectral gap of size at least~\(g\) above the ground state. The interaction Hamiltonian~\(H_{\rm int}\) is a sum of interaction terms~\(\Phi_x\) of finite range~\(R\) and of uniformly bounded norm \(\norm{\Phi_x} \). We show that for such Hamiltonians a strong version of the \emph{local perturbations perturb locally} (LPPL) principle holds: For any self-adjoint perturbation~\(P \), supported in a region \(X \subset \Lambda\), any ground state~\(\rho_P\) of the perturbed Hamiltonian \(H + P\) agrees with the ground state~\(\rho\) of the unperturbed Hamiltonian~\(H\) when tested against observables~\(A\) supported in a region \(Y\subset \Lambda\) up to an error that is exponentially small in the distance~\(\dist(Y,X)\). More precisely, Theorem~\ref{thm:LPPL} states that there are positive constants \(c, c_1, c_2>0 \) depending only on~\(R\) and~\(g\), but not on~\(\Lambda\), \(A\), \(H\) or~\(P\), such that whenever \(\norm{\Phi_x} \leq c\) for all \(x\in \Lambda\), it holds that \begin{equation} \label{eq:introbound} \abs[\big]{\trace[\big]{\paren{ \rho_P - \rho } A}} \, \le \, {\mathrm{e}}^{c_1 \abs{Y}} \, \norm{A} \, {\mathrm{e}}^{-c_2 \dist(Y,X) } \,. \end{equation} Note that the uniformity of the error estimate with respect to the system size~\(\abs{\Lambda}\) is one key aspect which makes this estimate non-trivial. Note also, that the bound on~\(\norm{\Phi_x}\) actually implies that there exists~\(\tilde g>0\) independent of~\(\Lambda\) such that~\(H\) has a gap of size at least~\(\tilde g\) above its unique ground state \(\rho\) (see~\cite{yarotsky2004perturbations,FP} and the remark after Theorem~\ref{thm:LPPL} for this non-trivial result). However, for our result we neither require nor actually have any uniform lower bound on the gap above the possibly non-unique ground state~\(\rho_P\) of the perturbed Hamiltonian \(H + P\). As a corollary of our main theorem, we show that a bound of the form~\eqref{eq:introbound} also holds for systems that have the appropriate structure of gapped on-site terms and weak interactions only locally in some region of space. In particular, this shows that the notion of a locally gapped ground state, which is completely insensitive to the form of the Hamiltonian elsewhere, is perfectly valid in this setup. The LPPL-principle was coined by Bachmann, Michalakis, Nachtergaele, and Sims in~\cite{bachmann2012automorphic}, where a similar estimate with sub-exponential decay was proven. While their result covers much more general interacting quantum spin systems, it requires the gap above the ground state to remain open also for the perturbed Hamiltonian \(H + P\). More precisely, it relies on connecting \(H(0) := H\) with \(H(1) := H + P\) by a continuous path \([0,1] \ni t \mapsto H (t)\) in the space of Hamiltonians, such that the gap above the ground state of \(H (t)\) remains open uniformly along the whole path. Then the locality of the quasi-adiabatic evolution introduced by Hastings and Wen in~\cite{hastings2005quasiadiabatic} can be used to prove the result. Their sub-exponential bound was improved to exponential precision for finite-range interactions by de~Roeck and Sch\"utz in~\cite{deRoeckLPPL}. See also~\cite{nachtergaele2019quasi,nachtergaele2020quasi} for recent developments. While we prove the strong version of the LPPL-principle only for weakly interacting spin systems, we expect it to hold somewhat more generally. For example, we expect it to hold for fermions on the lattice with weak finite range interactions, a physical setup where the strong LPPL-principle would have important consequences. It would imply that a gapped ground state for such a system with periodic boundary conditions remains unchanged in the bulk when introducing open boundary conditions that may close the global gap due to the emergence of edge states. And as a consequence, it would also explain why the adiabatic response to external fields in the bulk of such systems is not affected by edge states that close the gap, see~\cite{bachmann2018adiabatic, monaco2017adiabatic, teufel2020non, henheikteufel20202, henheikteufel20203} for related results. However, it is known that the strong LPPL-principle cannot hold in general, but requires further conditions on the unperturbed ground state sector such as local topological quantum order (LTQO)~\cite{MZ13,nachtergaele2021stability}. Our result is a corollary of a result by Yarotsky~\cite{yarotsky2005uniqueness} (see Theorem~\ref{thm:Yarotsky} below), which provides a bound on the difference of so-called finite volume ground states in quantum spin systems described by Hamiltonians of the form~\eqref{eq:introhamiltonian}. His aim and main result in this work was to show the uniqueness of the ground state of such systems in the thermodynamic limit. In a different work Yarotsky~\cite{yarotsky2004perturbations} has shown that Hamiltonians of the form~\eqref{eq:introhamiltonian} indeed have a spectral gap whenever the interacting part \(H_{\rm int}^\Lambda\) is small enough. For similar results see also~\cite{DS,FP, hastings2019stability}. \\[3mm] \noindent {\bf Acknowledgement.} S.~T.\ thanks Marius Lemm and Simone Warzel for very helpful remarks and discussions. \section{Main Results}\label{sec:result} Consider the lattice \(\mathbb{Z}^\nu\) for fixed \(\nu \in \mathbb{N}\) equipped with the \(\ell^1\)-metric \(d \colon \mathbb{Z} ^\nu\times \mathbb{Z} ^\nu \to \mathbb{N}_0\) and define \(\mathcal{P}_0( \mathbb{Z} ^\nu) = \Set{ \Lambda \subset \mathbb{Z} ^\nu \given \abs{\Lambda} < \infty }\), where~\(\abs{\Lambda}\) denotes the cardinality of~\(\Lambda\). With each site \(x \in \mathbb{Z} ^\nu\) one associates a finite dimensional Hilbert space~\(\mathcal{H}_x\). We assume that the dimension of~\(\mathcal{H}_x\) is bounded uniformly in \(x \in \mathbb{Z} ^\nu\). For \(\Lambda \in \mathcal{P}_0( \mathbb{Z} ^\nu)\) set \(\mathcal{H}_{\Lambda}=\bigotimes_{x \in \Lambda} \mathcal{H}_x\) and denote the algebra of bounded linear operators on \(\mathcal{H}_{\Lambda}\) by \(\mathcal{A}_{\Lambda}=\mathcal{B}(\mathcal{H}_{\Lambda})\). Due to the tensor product structure, we have \(\mathcal{A}_{\Lambda}=\bigotimes_{x \in \Lambda} \mathcal{B}(\mathcal{H}_x)\). Hence, for \(\Lambda'\) and \(\Lambda \in \mathcal{P}_0( \mathbb{Z} ^\nu)\) with \(\Lambda' \subset \Lambda\), any \(A \in \mathcal{A}_{\Lambda'}\) can be viewed as an element of \(\mathcal{A}_{\Lambda}\) by identifying \(A\) with \(A\otimes \mathbf{1}_{\Lambda\setminus \Lambda'} \in \mathcal{A}_{\Lambda}\), where \(\mathbf{1}_{\Lambda \setminus \Lambda'}\) denotes the identity in \(\mathcal{A}_{\Lambda \setminus \Lambda'}\). Note that \begin{equation} \label{eq:commutator} \commutator{A,B} = 0 \quad \text{for all} \quad A \in \mathcal{A}_\Lambda \,, \ B \in \mathcal{A}_{\Lambda'} \quad \text{with} \quad \Lambda \cap \Lambda' = \emptyset\,. \end{equation} Our main result will be formulated for a Hamiltonian \begin{equation*} \quad H = H_0 + H_{\rm int} \in \mathcal{A}_{\Lambda} \end{equation*} that is composed of a non-interacting part \(H_0\) and an interacting part \(H_{\rm int}\). The non-interacting part \(H_0\) is assumed to be of the form \begin{equation*} H_0 = \sum_{x \in \Lambda} h_x\,, \end{equation*} where for each \(x \in \Lambda\) the positive self-adjoint operator \(h_x \in \mathcal{A}_{\{x\}}\) has a unique gapped ground state \(\psi_x \in \mathcal{H}_x\) satisfying \begin{equation} \label{eq:gap} h_x \psi_x = 0 \quad \text{and} \quad h_x\big\vert_{\mathcal{H}_x \ominus \psi_x} \geq g\,, \end{equation} for some fixed \(g>0\). The latter means that \(\braket{\varphi_x,(h_x-g\mathbf{1}_x)\varphi_x} \geq 0\) for all \(\varphi_x \in \mathcal{H}_x\) with \(\braket{\psi_x,\varphi_x}=0\), i.e.~the Hamiltonians \(h_x\) all have a spectral gap of size at least \(g\) at the bottom of their spectrum. The interacting part is of the form \begin{equation*} H_{\rm int} = \sum_{x \in \Lambda} \Phi_x\,, \end{equation*} with \(\Phi_x\in \mathcal{A}_{b_x(R)}\) self-adjoint for each \(x \in \Lambda\) and some fixed \(R\in \mathbb{N} \). Here \(b_x(R) := {\Set{y \in \Lambda \given d(x,y) \leq R}}\) denotes the \(\ell^1\)-ball with radius \(R \) centered at \(x \in \Lambda\). We set \begin{equation*} \tnorm{\Phi} := \sup_{x \in \Lambda} \, \norm{\Phi_x} \,. \end{equation*} \begin{defi}\label{def:WISS} \textbf{Weakly interacting spin system}\\ For any \(\Lambda \in \mathcal{P}_0( \mathbb{Z} ^\nu) \) we call a Hamiltonian \(H = H_0 + H_{\rm int} \in\mathcal{A}_\Lambda\) with \(H_0\) and \(H_{\rm int}\) satisfying the above conditions a \emph{weakly interacting spin system on \(\Lambda\)} with on-site gap~\(g\), interaction range~\(R\) and interaction strength \( \tnorm{\Phi} \). \end{defi} For the reader's convenience and for later use, we recall the following characterization of ground states for quantum spin systems (see, e.g.,~\cite[p.\,488]{tasaki2020}). \begin{lem} \label{def:groundstate} Let \(\Lambda \in \mathcal{P}_0( \mathbb{Z} ^\nu)\) and \(K \in \mathcal{A}_{\Lambda}\) be self-adjoint. A state \(\rho \in \mathcal{A}_{\Lambda}\), i.e.~a positive semi-definite operator with trace equal to one, is a ground state of \(K \), i.e.\ it satisfies \begin{equation*} \trace{ \rho \, K } \,\leq\, \trace{\tilde \rho \, K } \quad \text{for all states \(\tilde \rho \in \mathcal{A}_{\Lambda}\)}\,, \end{equation*} if and only if \begin{equation*} \trace*{\rho \, A^* \commutator{K ,A} } \,\geq\, 0 \quad \text{for all} \quad A \in \mathcal{A}_{\Lambda} \,. \end{equation*} \end{lem} Our first main result is the following. \begin{thm} \label{thm:LPPL} {\bf The strong LPPL-principle }\\ Let \(R\in \mathbb{N} \) and \(g>0\). There exist constants \(c, c_1, c_2>0\), such that for any \(\Lambda\in \mathcal{P}_0( \mathbb{Z} ^\nu)\) and any weakly interacting spin system \(H = H_0 + H_{\rm int}\) on \(\Lambda\) with on-site gap at least \(g\), interaction range \(R\), and interaction strength \(\tnorm{\Phi} \leq c\) the following holds: Let \(X\subset\Lambda\) and \(P \in \mathcal{A}_{X}\) be self-adjoint and set \(H _P = H + P\). Then for any ground state \(\rho\) of \(H\), any ground state \(\rho_P\) of \(H_P\), and all \(A \in \mathcal{A}_Y\) with \(Y \subset \Lambda\) it holds that \begin{equation} \label{eq:thm} \abs[\big]{\trace[\big]{ \paren{\rho_P -\rho } A}} \,\le\, {\mathrm{e}}^{c_1 \abs{Y}} \, \norm{A} \, {\mathrm{e}}^{-c_2 \dist(Y,X)} \,. \end{equation} \end{thm} \medskip For \(X=\emptyset\) where \(\dist(Y,\emptyset):=\infty\), this proves uniqueness of the ground state of weakly interacting spin systems according to Definition~\ref{def:WISS}.\footnote{Note that the systems for which Yarotsky proves the uniqueness of the ground state in \cite{yarotsky2004perturbations} differ slightly from our definition of weakly interacting spin systems in the treatment of interaction terms near the boundary of the domain. Since this uniqueness enters into the proof of~\cite[Theorem 2]{yarotsky2005uniqueness}, which in turn is the basis for our result, our comment should not be misunderstood to mean that a new proof of ground-state uniqueness is at work.} For \(X\) at the edge of \(\Lambda\), the perturbation \(P\) can be employed to realize all kinds of boundary conditions, e.g.\ if \(\Lambda = \{ -M,\dotsc,M \}^\nu\) is a box, periodic boundary conditions can be modeled by some \(P\) connecting opposite sites in \(\Lambda\). Therefore, if \(X\) is at the edge, one can take the thermodynamic limit \({\Lambda \nearrow \mathbb{Z} ^\nu}\) in~\eqref{eq:thm} and conclude that there exists a unique ground state \(\rho\), i.e.\ a normalized positive functional, on the \(C^*\)-algebra of quasi-local observables \(\mathcal{A} = \LTskip{\overline{\mathcal{A}_{\rm loc}}^{\raisebox{-2pt}{\(\scriptstyle\norm{\cdot}\)}}}\), independent of the imposed boundary conditions for the finite systems. This uniqueness of ground states for the infinite system was the main result of~\cite{yarotsky2005uniqueness} and has been shown by Yarotsky based on Theorem~\ref{thm:Yarotsky}, which we quote below. As mentioned in the introduction, we expect a similar strong LPPL-principle to hold also for fermionic lattice systems with weak finite range interactions. As discussed in~\cite{henheikteufel20202, henheikteufel20203}, this would have important consequences for linear response and adiabatic theorems for systems with a gap only in the bulk. Our second main result is a local version of Theorem~\ref{thm:LPPL}, where we assume the on-site gap and the weak interaction only locally. \begin{defi} \label{def:WISS-in-region}\textbf{Locally weakly interacting spin system}\\ For any \(\Lambda \in \mathcal{P}_0( \mathbb{Z} ^\nu)\) and \(\Lambda'\subset\Lambda\) we say that a self-adjoint operator \(H \in \mathcal{A}_\Lambda\) is \emph{weakly interacting in the region~\(\Lambda'\)} with on-site gap \(g\), range~\(R\) and strength~\(s\), if and only if there exists a weakly interacting spin system \(\tilde H = H_0 + H_{\rm int} \in\mathcal{A}_\Lambda\) with on-site gap \(g\), range~\(R\) and strength~\(\tnorm{\Phi}=s\) such that \( H - \tilde H \in \mathcal{A}_{\Lambda\setminus\Lambda'}\). \end{defi} \begin{cor} \label{thm:LPPL-local-gaps} {\bf The strong LPPL-principle for local gaps}\\ Let~\(R \in \mathbb{N} \), \(g>0\), and \(c, c_1, c_2>0\) be the constants from Theorem~\ref{thm:LPPL}. Then for any \(\Lambda \in \mathcal{P}_0( \mathbb{Z} ^\nu)\), \(\Lambda'\subset\Lambda\), and any self-adjoint operator \(H \in \mathcal{A}_\Lambda\) that is weakly interacting in the region \(\Lambda'\) with on-site gap at least~\(g\), range~\(R\) and strength~\(s \leq c\) the following holds: Let \(X\subset\Lambda\) and \(P \in \mathcal{A}_{X}\) be self-adjoint and set \(H_P = H + P\). Then for any ground state \(\rho \) of \(H \), any ground state \(\rho _P\) of \(H _P\), and all \(A \in \mathcal{A}_Y\) with \(Y \subset \Lambda'\) it holds that \begin{equation*} \abs[\big]{\trace[\big]{ \paren{\rho _P -\rho } A}} \,\le\, 2 \, {\mathrm{e}}^{c_1 \abs{Y}} \, \norm{A} \, {\mathrm{e}}^{-c_2 \min \List{ \dist(Y,X), \dist(Y,\Lambda\setminus \Lambda')}} \,. \end{equation*} \end{cor} \begin{SCfigure}[1.2] \centering \begin{tikzpicture}[scale=.43, myLine] \newcommand{(-6.2,6.2) rectangle (6.2,-6.2)}{(-6.2,6.2) rectangle (6.2,-6.2)} \newcommand{(-6,6) rectangle (6,-6)}{(-6,6) rectangle (6,-6)} \newcommand{(-6,6) -- plot[smooth, tension=0.8] coordinates{(3,6) (4,2) (3,-1) (2,-4) (-1,-4) (-5,-6)} -- (-6,-6) -- cycle}{(-6,6) -- plot[smooth, tension=0.8] coordinates{(3,6) (4,2) (3,-1) (2,-4) (-1,-4) (-5,-6)} -- (-6,-6) -- cycle} \newcommand{(-2.4,-2.5)}{(2.5,3)} \newcommand{\pathX}{plot[smooth cycle, tension=.7]% coordinates{+(0:2.5) +(60:1.8) +(120:1.1) +(180:2) +(240:1.6) +(300:1.3)}} \newcommand{(-3.5,-.5)}{(-3.5,-.5)} \newcommand{plot[smooth cycle, tension=0.8] coordinates{+(10:.6) +(95:2) +(175:1.6) +(270:1.8) +(325:2.5)}}{plot[smooth cycle, tension=0.8] coordinates{+(10:1) +(95:1.5) +(175:1.6) +(270:1.8) +(325:2.5)}} \begin{scope}[myLine=2, even odd rule, draw=myColor] \clip (-6,6) -- plot[smooth, tension=0.8] coordinates{(3,6) (4,2) (3,-1) (2,-4) (-1,-4) (-5,-6)} -- (-6,-6) -- cycle; \draw (-6,6) -- plot[smooth, tension=0.8] coordinates{(3,6) (4,2) (3,-1) (2,-4) (-1,-4) (-5,-6)} -- (-6,-6) -- cycle; \end{scope} \node[anchor=north east, myColor] at (3,6) {\(\Lambda'\)}; \draw (-2.4,-2.5){} node {\(X\)} \pathX{}; \draw (-3.5,-.5){} node {\(Y\)} plot[smooth cycle, tension=0.8] coordinates{+(10:.6) +(95:2) +(175:1.6) +(270:1.8) +(325:2.5)}{}; \begin{scope}[myLine=2, even odd rule] \clip (-6.2,6.2) rectangle (6.2,-6.2) (-6,6) rectangle (6,-6); \draw (-6,6) rectangle (6,-6); \end{scope} \draw (6,-6) node[anchor=south east] {\(\Lambda\)}; \draw[latex-latex, shorten <=0.4cm, shorten >=0.8cm] (-3.5,-.5){} ++(90:.5) -- node[anchor=south east] {\(\scriptstyle\dist(Y,X)\)} (-2.4,-2.5){}; \draw[latex-latex, shorten <=2pt, shorten >=3pt] (-3.5,-.5){} ++(-47:2.5) -- node[anchor=east] {\(\scriptstyle\dist(Y,\Lambda\setminus\Lambda')\)} (-1,-4); \end{tikzpicture} \caption{Depicted is the setting from Corollary~\ref{thm:LPPL-local-gaps}. The system \(H\) defined on \(\Lambda\) is assumed to be weakly interacting and to have an on-site gap in \(\Lambda'\subset\Lambda\). For any perturbation \(P\) acting on \(X\subset\Lambda\), ground states of \(H\) and \(H+P\) agree in regions $Y$ away from \(X\) and \(\Lambda\setminus\Lambda'\).} \label{fig:cor} \end{SCfigure} \begin{proof} Let~\(\tilde{H}\) be as in Definition~\ref{def:WISS-in-region}, \(Q = H - \tilde{H} \in \mathcal{A}_{\Lambda\setminus\Lambda'}\) and~\(\tilde \rho \) be the ground state of~\(\tilde{H}\). Then the triangle inequality and two applications of Theorem~\ref{thm:LPPL} yield \begin{align*} \abs[\big]{\trace[\big]{ \paren{\rho _P -\rho } A}} \, & \le \, \abs[\big]{\trace[\big]{ \paren{\rho _P -\tilde \rho } A}} + \abs[\big]{\trace[\big]{ \paren{\rho -\tilde \rho } A}}\\ \, & \le \, {\mathrm{e}}^{ c_1 \abs{Y}} \, \norm{A} \, \paren[\big]{ {\mathrm{e}}^{-c_2 \dist(Y,X \cup (\Lambda\setminus\Lambda')) } + {\mathrm{e}}^{-c_2 \dist(Y,\Lambda\setminus \Lambda') } }\,.\qedhere \end{align*} \end{proof} \section{Proof} \label{sec:proof} The proof of Theorem~\ref{thm:LPPL} is essentially a reinterpretation of a result by Yarotsky~\cite{yarotsky2005uniqueness}. Since we only deal with finite volumes, we modify Yarotsky's notion of \emph{finite volume ground states} to \emph{ground states in the bulk}. To make the arguments as transparent as possible, we will add superscripts to Hamiltonians and states indicating on which subset of \( \mathbb{Z} ^\nu\) they are defined. These superscripts are also used to distinguish different operators and states. \begin{defi} \label{def:groundstatebulk}\textbf{Ground states in the bulk}\\ Let \(R \in \mathbb{N}\), \(\Lambda_* \subset \Lambda \in \mathcal{P}_0( \mathbb{Z} ^\nu)\) and \(H^{\Lambda_*} = H_0^{\Lambda_*} + H_{\rm int}^{\Lambda_*} \in \mathcal{A}_{\Lambda_*}\) be a weakly interacting spin system on \(\Lambda_*\) with range \(R\). Then we call \begin{equation*} \Lambda_*^\circ := \Set{ x \in \Lambda_* \given \dist(x, \mathbb{Z} ^\nu\setminus \Lambda_*) > 2R} \end{equation*} the \emph{bulk} of the Hamiltonian \(H^{\Lambda_*}\) and any state \(\rho^\Lambda \in \mathcal{A}_{\Lambda}\) satisfying \begin{equation*} \trace*{\rho^\Lambda \, A^* \commutator{H,A} } \geq 0 \quad \text{for all} \quad A \in \mathcal{A}_{\Lambda_*^\circ} \end{equation*} a \emph{ground state in the bulk of \(H^{\Lambda_*}\)}. \end{defi} Our proof is based on the following theorem due to Yarotsky~\cite{yarotsky2005uniqueness}. \begin{thm} {\rm (\cite[Theorem 2]{yarotsky2005uniqueness})} \label{thm:Yarotsky} Let~\(R\in \mathbb{N} \) and~\(g>0\). There exist constants \(c, c_1, c_2>0\) such that for any \(\Lambda_* \in \mathcal{P}_0( \mathbb{Z} ^\nu)\), and any weakly interacting spin system \(H^{\Lambda_*} = H_0^{\Lambda_*} + H_{\rm int}^{\Lambda_*}\) on~\(\Lambda_*\) with on-site gap at least~\(g\), range~\(R\) and interaction strength~\(\tnorm{\Phi}\leq c\) the following holds: Let \(\Lambda \in \mathcal{P}_0( \mathbb{Z} ^\nu)\) be such that \(\Lambda_* \subset \Lambda\). Then for any two ground states~\(\rho^\Lambda_1\) and~\(\rho^\Lambda_2 \in \mathcal{A}_\Lambda\) in the bulk of~\(H^{\Lambda_*}\) in the sense of Definition~\ref{def:groundstatebulk}, \(Y \subset \Lambda_*\), and~\(A \in \mathcal{A}_Y\) it holds that \begin{equation*} \abs[\big]{ \trace[\big]{\paren{ \rho^\Lambda_1 - \rho^\Lambda_2 } A}} \, \le \, {\mathrm{e}}^{c_1 \abs{Y}} \, \norm{A} \, {\mathrm{e}}^{-c_2\, \dist(Y, \mathbb{Z} ^\nu \setminus \Lambda_*^\circ)} \,. \end{equation*} \end{thm} Note that the set denoted by~\(\Lambda\) in~\cite[Theorem~2]{yarotsky2005uniqueness} corresponds to our set~\(\Lambda_*\). Note, moreover, that any ground state~\(\rho^\Lambda\) in the bulk of~\(H^{\Lambda_*}\) trivially defines a finite-volume ground state \(A \mapsto \trace{\rho^\Lambda \, (A\otimes \mathbf{1}_{\Lambda\setminus\Lambda_*})}\) of~\(H^{\Lambda_*}\) in the sense of~\cite[Definition~2]{yarotsky2005uniqueness}. Allowing an arbitrary on-site gap~\(g>0\) instead of~\(g=1\), as in~\cite{yarotsky2005uniqueness}, is achieved by simple scaling. \begin{lem} \label{lem:ground-state-to-ground-state-in-bulk} Let \(R \in \mathbb{N}\), \(\Lambda_*\subset \Lambda \in \mathcal{P}_0( \mathbb{Z} ^\nu)\) and \(H^{\Lambda} = H_0^{\Lambda} + H_{\rm int}^{\Lambda} \in \mathcal{A}_{\Lambda}\) be a weakly interacting spin system. Then the canonical restriction of \(H^\Lambda\) to \(\Lambda_*\) defined by \begin{equation*} H^{\Lambda}|_{\Lambda_*} = H_0^{\Lambda}|_{\Lambda_*} + H_{\rm int}^{\Lambda}|_{\Lambda_*} := \sum_{x \in \Lambda_*} h_x + \sumstack[lr]{x \in \Lambda_*:\\ \dist(x,\Lambda\setminus\Lambda_*)>R} \Phi_x \end{equation*} is a weakly interacting spin system on \(\Lambda_*\) with the same on-site gap, range and strength and has the following property: For any self-adjoint \(Q \in \mathcal{A}_{\Lambda \setminus \Lambda^\circ_*}\), any ground state of \(H^\Lambda + Q\) is also a ground state in the bulk of~\(H^{\Lambda}|_{\Lambda_*}\). \end{lem} \begin{proof} It is clear that \(H^{\Lambda}|_{\Lambda_*}\) is a weakly interacting spin system on \(\Lambda_*\). Since \(H^{\Lambda}|_{\Lambda_*}\) agrees with \(H^\Lambda+Q\) on \(\Lambda^\circ_*\) in the sense that \((H^\Lambda + Q -H^{\Lambda}|_{\Lambda_*})\in \mathcal{A}_{\Lambda \setminus \Lambda_*^\circ}\), we find, by application of~\eqref{eq:commutator}, that \begin{equation*} \commutator*{ H^\Lambda + Q, A }= \commutator*{ H^{\Lambda}|_{\Lambda_*}, A } \quad \text{for all} \quad A \in \mathcal{A}_{\Lambda_*^\circ} \,. \qedhere \end{equation*} \end{proof} \begin{SCfigure}[1.2] \centering \begin{tikzpicture}[scale=.43, myLine] \newcommand{(-6,6) rectangle (6,-6)}{(-6,6) rectangle (6,-6)} \newcommand{(-5,5) rectangle (5,-5)}{(-5,5) rectangle (5,-5)} \newcommand{(\((-5,5) + (\offset,-\offset)\)) rectangle (\((5,-5) + (-\offset,\offset)\))}{(\((-5,5) + (\offset,-\offset)\)) rectangle (\((5,-5) + (-\offset,\offset)\))} \newcommand{(-2.4,-2.5)}{(-2.4,-2.5)} \newcommand{\pathX}[1]{plot[smooth cycle, tension=.8]% coordinates{+(0:#1+1.5) +(60:#1+1) +(120:#1+1.2) +(180:#1+.8) +(240:#1+1) +(300:#1+1.3)}} \newcommand{plot[smooth cycle, tension=0.8] coordinates{+(10:.6) +(95:2) +(175:1.6) +(270:1.8) +(325:2.5)}}{plot[smooth cycle, tension=0.8] coordinates{+(10:.6) +(95:2) +(175:1.6) +(270:1.8) +(325:2.5)}} \newcommand{\drawXYL}[1]{ \draw (-2.4,-2.5){} node {\(X\)} \pathX{0}; \draw (2,1.5) +(225:.6) node {\(Y\)} plot[smooth cycle, tension=0.8] coordinates{+(10:.6) +(95:2) +(175:1.6) +(270:1.8) +(325:2.5)}; \draw (-6,6) rectangle (6,-6) node[anchor=south east] {#1}; } \begin{scope} \begin{scope}[myColor] \begin{scope}[even odd rule, myLine=2, fill=myColor!10, draw=myColor] \clip (-5,5) rectangle (5,-5) (0,0) [shift={(-2.4,-2.5){}}] \pathX{1}; \path[draw, fill] (-5,5) rectangle (5,-5) (0,0) [shift={(-2.4,-2.5){}}] \pathX{1}; \node[anchor=south, inner sep=2pt] at (0,-5) {\(\Lambda_*^\circ\)}; \end{scope} \end{scope} \begin{scope}[myLine=2, even odd rule] \clip (-6,6) rectangle (6,-6) (-5,5) rectangle (5,-5); \draw (-5,5) rectangle (5,-5); \end{scope} \node[anchor=south east] at (5,-5) {\(\Lambda^\circ\)}; \end{scope} \drawXYL{\(\Lambda\)} \end{tikzpicture} \caption{Depicted is the setting from the proof of Proposition~\ref{thm:LPPL-intermediate}. The subset~\(X\subset\Lambda\) is the region where the perturbation~\(P\) acts, and we choose \({\Lambda_*=\Lambda \setminus X}\). The shaded region~\(\Lambda_*^\circ\) is the bulk of~\(H^{\Lambda_*}\). \({Y\subset\Lambda_*^\circ}\) is the support of the observable~\(A\). This indicates why~\eqref{eq:distance} holds.} \label{fig:prop} \end{SCfigure} Before we prove Theorem~\ref{thm:LPPL}, let us give an intermediate result, which follows rather directly from Theorem~\ref{thm:Yarotsky} and Lemma~\ref{lem:ground-state-to-ground-state-in-bulk}. \begin{prop} \label{thm:LPPL-intermediate} Let \(R\in \mathbb{N} \) and \(g>0\). There exist constants \(c, c_1, c_2>0\) such that for any \(\Lambda\in \mathcal{P}_0( \mathbb{Z} ^\nu)\) and any weakly interacting spin system \(H^\Lambda = H^\Lambda_0 + H^\Lambda_{\rm int}\) on \(\Lambda\) with on-site gap at least \(g\), interaction range \(R\), and interaction strength \(\tnorm{\Phi} \leq c\) the following holds: Let \(X\subset\Lambda\) and \(P \in \mathcal{A}_{X}\) be self-adjoint and set \(H^\Lambda_P = H^\Lambda + P\). Then for any ground state \(\rho^\Lambda\) of \(H^\Lambda\), any ground state \(\rho^\Lambda_P\) of \(H^\Lambda_P\), and all \(A \in \mathcal{A}_Y\) with \(Y \subset \Lambda\) it holds that \begin{equation*} \abs[\big]{\trace[\big]{ \paren{\rho^\Lambda_P - \rho^\Lambda } A}} \, \le \, {\mathrm{e}}^{c_1 \abs{Y}} \, \norm{A} \, {\mathrm{e}}^{-c_2 \min\List*{ \dist(Y, \mathbb{Z} ^\nu \setminus \Lambda^\circ), \, \dist(Y,X) - 2R}} \,. \end{equation*} \end{prop} \begin{proof} Assume w.l.o.g.\ that \(Y \subset \Lambda^\circ\). Otherwise, the statement in Proposition~\ref{thm:LPPL-intermediate} is trivially satisfied after a possible adjustment of \(c_1\). Let \(\Lambda_* = \Lambda \setminus X\), and let \(H^{\Lambda}|_{\Lambda_*} \) be the canonical restriction of \(H^\Lambda\) to \(\Lambda_*\) as defined in Lemma~\ref{lem:ground-state-to-ground-state-in-bulk}. Then \(\Lambda_*^\circ \cap X = \emptyset\). We can assume w.l.o.g.\ that \(\dist(X,Y) > 2R\) since otherwise the statement in Proposition~\ref{thm:LPPL-intermediate} is trivially satisfied after a possible adjustment of \(c_1\). Then also \(Y \subset \Lambda_*^\circ\) (compare Figure~\ref{fig:prop}). By application of Lemma~\ref{lem:ground-state-to-ground-state-in-bulk} with \(Q = P\) and \(Q = 0\) we find that both, \(\rho_P^\Lambda\) and \(\rho^\Lambda\), are ground states in the bulk of \(H^{\Lambda}|_{\Lambda_*}\). Hence, Theorem~\ref{thm:Yarotsky} implies that \begin{equation*} \label{eq:proof} \abs[\big]{ \trace[\big]{\paren{\rho^\Lambda_P - \rho^\Lambda} A}} \, \le \, {\mathrm{e}}^{c_1 \abs{Y}} \, \norm{A} \, {\mathrm{e}}^{-c_2 \dist(Y, \mathbb{Z} ^\nu \setminus \Lambda_*^\circ)} \,. \end{equation*} From \begin{equation*} \mathbb{Z} ^\nu \setminus \Lambda_*^\circ = ( \mathbb{Z} ^\nu \setminus \Lambda^\circ) \cup \Set{ x \in \mathbb{Z} ^\nu \given \dist(x,X) \le 2R } \end{equation*} we immediately conclude that \begin{equation}\label{eq:distance} \dist(Y, \mathbb{Z} ^\nu \setminus \Lambda_*^\circ) = \min \List*{ \dist(Y, \mathbb{Z} ^\nu \setminus \Lambda^\circ), \, \dist(Y, X) - 2R }\,, \end{equation} which yields the claim. \end{proof} We now extend this result to obtain Theorem~\ref{thm:LPPL}. \begin{proof}[Proof of Theorem~\ref{thm:LPPL}] In the following, we add superscripts \(\Lambda\) to the Hamiltonians and states from the statement of Theorem~\ref{thm:LPPL}. Let \(\Omega \in \mathcal{P}_0( \mathbb{Z} ^\nu)\) be such that \(\Lambda \subset \Omega\). For each \(x \in \Omega \setminus \Lambda\) let \(h_x \in \mathcal{A}_{\List{x}}\) be a self-adjoint operator with gap at least \(g\) and non-degenerate ground state \(\psi_x\) satisfying~\eqref{eq:gap}. Then \(\rho^{\Omega \setminus \Lambda} = \bigotimes_{x\in\Omega \setminus \Lambda} \ket{\psi_x}\bra{\psi_x}\) is the ground state of \begin{equation*} H_0^{\Omega\setminus\Lambda} := \sumstack[lr]{x\in\Omega\setminus\Lambda} h_x\,. \end{equation*} Moreover, \(\rho^{\Omega}:=\rho^{\Lambda} \otimes \rho^{\Omega \setminus \Lambda}\) is a ground state of \(H^\Omega := H^\Lambda + H_0^{\Omega\setminus\Lambda}\) which is a weakly interacting spin system on \(\Omega\) with on-site gap at least \(g\), range \(R\), and interaction strength~\(\tnorm{\Phi}\). And also \(\rho_P^{\Omega}:=\rho_P^{\Lambda} \otimes \rho^{\Omega \setminus \Lambda}\) is a ground state of \(H_P^\Omega := H_P^\Lambda + H_0^{\Omega\setminus\Lambda} = H^\Omega + H^\Lambda_P\). According to Proposition~\ref{thm:LPPL-intermediate} we have \begin{equation*} \abs[\big]{\trace[\big]{ \paren{\rho^\Omega_P - \rho^\Omega } A}} \, \le \, {\mathrm{e}}^{c_1 \abs{Y}} \, \norm{A} \, {\mathrm{e}}^{-c_2 \min\List*{ \dist(Y, \mathbb{Z} ^\nu \setminus \Omega^\circ), \, \dist(Y,X) - 2R}} \end{equation*} for all \(A\in\mathcal{A}_Y\) and \(Y\subset \Omega\). By requiring \(Y\subset \Lambda\) we obtain \begin{equation*} \abs[\big]{\trace[\big]{ \paren{\rho^\Lambda_P - \rho^\Lambda } A}} = \abs[\big]{\trace[\big]{ \paren{\rho^\Omega_P - \rho^\Omega } A}} \le {\mathrm{e}}^{c_1 \abs{Y}} \, \norm{A} \, {\mathrm{e}}^{-c_2 \min\List*{ \dist(\Lambda, \mathbb{Z} ^\nu \setminus \Omega^\circ), \, \dist(Y,X) - 2R}}. \end{equation*} Since this bound is independent of \(\Omega\), we can choose \(\Omega\) sufficiently large such that \( \dist(\Lambda, \mathbb{Z} ^\nu \setminus \Omega^\circ) > \dist(Y,X) - 2R \) if \(X\) is non-empty. Absorbing \({\mathrm{e}}^{2c_2R}\) in \(c_1\) yields the claim. In case \(X=\emptyset\), the minimum reduces to \(\dist(\Lambda, \mathbb{Z} ^\nu\setminus\Omega^\circ)\) and taking the limit \(\Omega \nearrow \mathbb{Z} ^\nu\) yields the claim. \end{proof}

https://arxiv.org/abs/1709.02869

Dimension reduction in the context of structured deformations

In this paper we apply both the procedure of dimension reduction and the incorporation of structured deformations to a three-dimensional continuum in the form of a thinning domain. We apply the two processes one after the other, exchanging the order, and so obtain for each order both a relaxed bulk and a relaxed interfacial energy. Our implementation requires some substantial modifications of the two relaxation procedures. For the specific choice of an initial energy including only the surface term, we compute the energy densities explicitly and show that they are the same, independent of the order of the relaxation processes. Moreover, we compare our explicit results with those obtained when the limiting process of dimension reduction and of passage to the structured deformation is carried out at the same time. We finally show that, in a portion of the common domain of the relaxed energy densities, the simultaneous procedure gives an energy strictly lower than that obtained in the two-step relaxations.

\section{Introduction}\label{intro} Classical continuum theories of elastic bodies are amenable to refinements that broaden their range of applicability or that adapt them to specific physical contexts. In this article we consider refinements that (i) incorporate into a classical theory the effects of submacroscopic slips and separations (disarrangements) or that (ii) adapt the theory to the description of thin bodies. Refinements of the type (i) are intended to describe finely layered bodies such as a stack of papers, granular bodies such as a pile of sand, or bodies with defects such as a metal bar. Those of type (ii) are intended to provide descriptions of membranes such as a sheet of rubber, descriptions of thin plates such as a sheet of metal, and descriptions of fibered thin bodies such as a sheet of paper. There are available a variety of approaches for incorporating disarrangements and for adaptation to the case of thin bodies: examples of refinements of type (i) are mechanical theories of no-tension materials \cite{angelillo,DO,LSZ}, of granular media \cite{numberone,K,Mue}, of single and polycrystals \cite{KS4,R2}, and of elastic bodies in the multiscale geometrical setting of structured deformations \cite{DO1,owen}, while for refinements of type (ii) the method of dimension reduction via $\Gamma$-convergence \cite{BF2001,LDR95,LDR96} and the method of dimension reduction via Taylor expansions \cite{DPZ} provide examples. Our goal in this paper is to implement in succession refinements of both types, starting from a classical, energetic description of three-dimensional elastic bodies. Specifically, for a refinement of type (i) we choose the context of structured deformations to incorporate the effects of submacroscopic slips and separations into a refined energetic response, while for a refinement of type (ii) we employ the method of dimension reduction via $\Gamma$-convergence to obtain a refined energetic response. With the starting point a three-dimensional body with a given energetic response, the two types of refinements can be carried out in two different orders, and each order of applying the two types of refinements will result in an energetic description of a two-dimensional body undergoing submacroscopic disarrangments, as indicated below in Figure \ref{figzero}: \begin{figure}[h] \begin{center} \begin{tikzpicture} \matrix (m) [matrix of math nodes,row sep=3em,column sep=4em,minimum width=2em] { & \text{3d-body} & \\ \text{2d-body} & & \text{3d-body with disarrangements} \\ \text{2d-body with disarrangements} & & \text{2d-body with disarrangements} \\ }; \path[-stealth] (m-1-2) edge [double] node [left] {$\text{(ii)}\quad$} (m-2-1) edge [double] node [right] {$\quad\text{(i)}$} (m-2-3) (m-2-1) edge [double] node [left] {$\text{(i)}$} (m-3-1) (m-2-3) edge [double] node [right] {$\text{(ii)}$} (m-3-3) ; \end{tikzpicture} \caption{The two paths for refinements of classical continuum theories: (i) structured deformations (SD) and (ii) dimension reduction (DR).} \label{figzero} \end{center} \end{figure} The right-hand path above begins with the incorporation of disarrangements (i) and then applies dimension reduction (ii), while the left-hand path reverses the order. We consider in this paper the nature of the energetic responses obtained at each step in the two paths and whether or not the two-dimensional body with disarrangements obtained via the left-hand path above has the same energetic response as that obtained via the right-hand path. Incorporation of disarrangements via structured deformations (i) replaces a vector field $u$ that maps a three-dimensional body into three-dimensional space by a pair $(g,G)$, where $g$ also maps the three-dimensional body into three-dimensional space and $G$ is a matrix-valued field that gives the contributions at the macroscopic level of submacroscopic deformations without disarrangements. The matrix-valued field $\nabla g-G$ then gives the contributions at the macroscopic level of submacroscopic deformations due to disarrangments. Dimension reduction (ii) replaces the vector field $u$ by a pair $(\cl{u},\cl{d})$ of vector fields defined on a two-dimensional body, where $\cl{u}$ places the two-dimensional body into three-dimensional space and $\cl{d}$ is a "director field" on the two-dimensional body that is a geometrical residue of the passage from a three-dimensional body to a two-dimensional body. In the diagram above, both (i) and (ii) begin with one and the same energy that depends only upon the field $u$: (i) results in an energy that depends upon the pair $(g,G)$, while (ii) results in an energy that depends on the pair $(\cl{u},\cl{d})$. When (i) and (ii) are applied consecutively, in either order, the resulting energy depends on a triple of fields $(\cl{g},\cl{G},\cl{d})$ defined on a two-dimensional body. The mathematical properties of these fields and the relation between the energy responses at each stage are summarized in the remainder of this introduction. \subsection{Statement of the problem and results}\label{sect:statement} Let $\omega\subset\mathbb{R}^{2}$ be a bounded open set, let $\eps>0,$ and let $\Omega_{\eps}:=\omega\times(-\frac{\eps}{2},\frac{\eps}{2})$. We recall that the set of \emph{special functions of bounded variation} on $\Omega_\eps$ consists of those $BV$ functions whose distributional derivative has no Cantor part, namely $SBV(\Omega_\eps;\R3):=\{u\in BV(\Omega_\eps;\R{3}): D^cu=0\}$ (see Section \ref{sect:BV}). For a function $u\in SBV(\Omega_{\eps};\mathbb{R}^{3})$, consider the energy% \begin{equation} E_{\eps}(u):=\int_{\Omega_{\eps}}W_{3d}(\nabla u(x))\,\de x+\int_{\Omega_{\eps}\cap S(u)}h_{3d}\big([u](x),\nu(u)(x)\big)\,\de\mathcal{H}^{2}(x)\label{E3d}% \end{equation} where $S(u)$ is the jump set of $u, [u]$ is the jump of $u$ across $S(u)$, and $\nu(u)$ is the unit normal vector to $S(u)$. The volume and surface energy densities $W_{3d}\colon{\mathbb{R}}^{3\times3}\rightarrow\lbrack0,+\infty)$ and $h_{3d}\colon{\mathbb{R}}^{3}\times \S{2}\rightarrow\lbrack0,+\infty)$ are continuous functions satisfying the following hypotheses: \begin{itemize} \item[$(H_{1})$] There exists a constant $c_W>0$ such that growth conditions from above and below are satisfied% \begin{align} \frac{1}{c_W} |A|^p & \leq {W}_{3d}(A), \label{coerc} \\ |W_{3d}(A)-W_{3d}(B)| & \leq c_W|A-B|(1+|A|^{p-1}+|B|^{p-1}), \label{growth} \end{align} for any $A,B \in\mathbb{R}^{3\times3},$ and for some $p >1$. \item[$(H_{2})$] There exists a constant $c_h>0,$ such that for all $(\lambda,\nu)\in\mathbb{R}^{3}\times \S{2}$% $$\frac1{c_h}|\lambda|\leq h_{3d}(\lambda,\nu)\leq c_{h}|\lambda|.$$ \item[$(H_{3})$] $h_{3d}(\cdot,\nu)$ is \emph{positively $1$-homogeneous}: for all $t>0$, $\lambda\in\R{3}$ $$h_{3d}(t\lambda,\nu)=t\,h_{3d}(\lambda,\nu).$$ \item[$(H_{4})$] $h_{3d}(\cdot,\nu)$ is \emph{subadditive}: for all $\lambda_{1},\lambda_{2}\in\mathbb{R}^{3}$ $$h_{3d}(\lambda_{1}+\lambda_{2},\nu)\leq h_{3d}(\lambda_{1},\nu)+h_{3d}(\lambda_{2},\nu).$$ \end{itemize} \begin{remark}\label{energies} \textit{(i)} The coercivity condition \eqref{coerc} in ($H_1$), although useful to obtain $L^p$ boundedness of the gradients, is not physically desirable. It can be removed following the argument in \cite[proof of Proposition 2.22, Step 2]{CF}: if $W_{3d}$ is not coercive, one can consider $W_{3d}^\beta(\cdot):=W_{3d}(\cdot)+\beta|\cdot|^p$ and then take the limit as $\beta\to0$. \\ \textit{(ii)} By fixing $B$ in \eqref{growth}, one can easily show that $W_{3d}$ satisfies also a growth condition of order $p$, that is, there exists a constant $C>0$ such that for every $A\in\R{3\times3}$ \begin{equation}\label{pgrowth} W_{3d}(A)\leq C(1+|A|^p). \end{equation} \end{remark} Under assumptions $(H_1)$--$(H_4)$, we carry out both a procedure of dimension reduction as $\eps\to0$ to obtain an energy functional defined on the cross--section $\omega$, and a procedure of relaxation to obtain an energy functional defined on structured deformations. The two procedures performed consecutively result in a doubly relaxed energy that may depend upon the order chosen. We will perform the two processes in both possible orders and compare the doubly relaxed energies. The schematic description of the two possible orders in Figure \ref{figzero} now takes the form in Figure \ref{figone}. \begin{figure}[h] \begin{center} \begin{tikzpicture} \matrix (m) [matrix of math nodes,row sep=3em,column sep=4em,minimum width=2em] { & W_{3d}, h_{3d} & \\ W_{3d,2d}, h_{3d,2d} & & W_{3d,SD}, h_{3d,SD} \\ W_{3d,2d,SD}, h_{3d,2d,SD} & & W_{3d,SD,2d}, h_{3d,SD,2d} \\ }; \path[-stealth] (m-1-2) edge [double] node [left] {$\mathrm{DR}\quad$} (m-2-1) edge [double] node [right] {$\quad\mathrm{SD}$} (m-2-3) (m-2-1) edge [double] node [left] {$\mathrm{SD}$} (m-3-1) (m-2-3) edge [double] node [right] {$\mathrm{DR}$} (m-3-3) ; \end{tikzpicture} \caption{Energy densities for the two paths for dimension reduction (DR) and structured deformations (SD).} \label{figone} \end{center} \end{figure} As indicated in Figure \ref{figone}, we derive formulas for bulk and interfacial densities obtained on the left-hand path and for those obtained via the right-hand path. The technical background for structured deformations and dimension reduction can be found in the following literature: \begin{itemize} \item[(i)] for structured deformations, we use the techniques introduced in \cite{CF}, where the relaxation process is obtained by combining the blow-up method of \cite{FM} with the construction of suitable approximating sequences by means of Alberti's theorem \cite{AL}; \item[(ii)] for dimension reduction, we employ the classical approach \cite{LDR95,LDR96} of rescaling the spatial variable to write the energy in the domain $\Omega=\omega\times(-1/2,1/2)$ and rescale the energy by dividing it by $\eps$. \end{itemize} Nevertheless, the sequential application of (i) and (ii) one after the other in both orders requires some non-trivial adaptations which are detailed in Remark \ref{independent}. We now summarize in abbreviated form the main results of this paper, and we refer the reader to Sections \ref{sect:LHS} and \ref{sect:RHS} for more detailed versions. \paragraph{\textbf{The left-hand path}} According to the diagram in Figure \ref{figone}, we first perform (DR) and then (SD). Following (ii), we relax the energies $F_\eps(u):=\frac1\eps E_\eps(u)$ defined for $u\in SBV(\Omega_\eps;\R3)$ to an energy $\cF_{3d,2d}(\cl u,\cl d)$ defined for pairs $(\cl u,\cl d)\in SBV(\omega;\R3)\times L^p(\omega;\R3)$. The deformation $\cl u(x_\alpha)$ is the limit of deformations $u_n(x_\alpha,x_3)$, where the dependence on the out-of-plane variable vanishes in the limit, whereas the vector $\cl d$ emerges as a weak limit of the out-of-plane deformation gradient \cite{BFM,BFM1}. \begin{theorem*}[Theorem \ref{T3d2d}] Given a pair $(\cl u,\cl d)\in SBV(\omega;\R3){\times}L^p(\omega;\R3)$, the relaxed energy $\cF_{3d,2d}(\cl u,\cl d)$ defined in \eqref{107a} admits the integral representation \eqref{F3d2d}, where the relaxed energy densities $W_{3d,2d}\colon\R{3\times2}\times\R3\to[0,+\infty)$ and $h_{3d,2d}\colon\R3{\times}\S1\to[0,+\infty)$ are given by \eqref{107b} and \eqref{107c}, respectively. \end{theorem*} In Proposition \ref{p100} we prove that the densities $W_{3d,2d}$ and $h_{3d,2d}$ satisfy the hypotheses of the relaxation method for structured deformations of \cite{CF}, which leads to the relaxed energy $\cF_{3d,2d,SD}(\cl g,\cl G,\cl d)$ defined for $(\cl g,\cl G,\cl d)\in SBV(\omega;\R3){\times} L^1(\omega;\R{3{\times}2}){\times} L^p(\omega; \R3)\to[0,+\infty)$ and to the following theorem. \begin{theorem*}[Theorem \ref{secondleft}] Given a triple $(\cl g,\cl G,\cl d)\in SBV(\omega;\R3)\times L^1(\omega;\R{3{\times}2})\times L^p(\omega; \R3)$, the relaxed energy $\cF_{3d,2d,SD}(\cl g,\cl G,\cl d)$ defined in \eqref{200} admits the integral representation \eqref{201}, where the relaxed energy densities $W_{3d,2d,SD}\colon\R{3\times2}{\times}\R{3\times2}{\times}\R3\to[0,+\infty)$ and $h_{3d,2d,SD}\colon\R3{\times}\S1\to[0,+\infty)$ are given by \eqref{2011} and \eqref{2012}, respectively. \end{theorem*} \paragraph{\textbf{The right-hand path}} According to the diagram in Figure \ref{figone}, we first perform (SD) and then (DR). Following (i), the assumptions $(H_1)$--$(H_4)$ allow us to apply directly \cite[Theorem 2.17]{CF} to obtain a representation theorem for the relaxed energy $\cF_{3d,SD}\colon SBV(\Omega;\R3)\times L^1(\Omega;\R{3\times2})\to[0,+\infty)$ defined for structured deformations $(g,G^{\backslash3})$. Strictly speaking, the structured deformations under consideration are pairs $(g,(G^{\backslash3}|\nabla_3 g))\in SBV(\Omega;\R3)\times L^1(\Omega;\R{3\times3})$, where for each $A\in\R{3\times2}$ and $q\in\R{3\times1}$, $(A|q)\in\R{3\times3}$ is formed from the two columns of $A$ and the single column of $q$. Because the $3\times3$ matrix values of $(g,(G^{\backslash3}|\nabla_3 g))$ are determined by the pair of fields $(g,G^{\backslash3})$, we allow this abuse of terminology and notation. \begin{theorem*}[Theorem \ref{firstSD}] Given a pair $(g,G^{\backslash3})\in SBV(\Omega;\R3)\times L^1(\Omega;\R{3\times2})$, the relaxed energy density $\cF_{3d,SD}(g,G^{\backslash3})$ defined in \eqref{2018} admits the integral representation \eqref{2017}, where the relaxed energy densities $W_{3d,SD}\colon\R{3\times3}\times\R{3\times2}\to[0,+\infty)$ and $h_{3d,SD}\colon\R3\times\S2\to[0,+\infty)$ are given by \eqref{2015} and \eqref{2016}, respectively. \end{theorem*} Proposition \ref{energiesCF} collects some properties of the densities $W_{3d,SD}$ and $h_{3d,SD}$. Therefore, performing the dimension reduction on the energy $\cF_{3d,SD}$ leads to the definition of the energy $\cF_{3d,SD,2d}\colon$ $SBV(\omega;\R3)\times L^1(\omega;\R{3{\times}2})\times L^p(\omega; \R3)\to[0,+\infty)$ for triples $(\cl g,\cl G,\cl d)$ defined on the cross--section $\omega$, and to the following theorem. \begin{theorem*}[Theorem \ref{secondright}] Given a triple $(\cl g,\cl G,\cl d)\in SBV(\omega;\R3)\times L^1(\omega;\R{3{\times}2})\times L^p(\omega; \R3)$, the relaxed energy $\cF_{3d,SD,2d}(\cl g,\cl G,\cl d)$ defined in \eqref{2019} admits the integral representation \eqref{301}, where the relaxed energy densities $W_{3d,SD,2d}\colon$ $\R{3\times2}{\times}\R{3\times2}{\times}\R3\to[0,+\infty)$ and $h_{3d,SD,2d}\colon\R3{\times}\S1\to[0,+\infty)$ are given by \eqref{2013} and \eqref{2014}, respectively. \end{theorem*} In this case, the results follow from those in \cite{CF} for the structured deformation part (Theorem \ref{firstSD}), and from applying Theorem \ref{T3d2d} for the dimension reduction part (Theorem \ref{secondright}). \begin{remark}\label{independent} We want to stress here that in both paths, the relaxation due to dimension reduction and that due to structured deformations are distinct refinements of the classical energetics of elastic bodies: the first one gives rise to the vector field $\cl d$ in the energy $\cF_{3d,2d}(\cl u,\cl d)$, whereas the second one gives rise to the matrix-valued field $G$ in the energy $\cF_{3d,SD}(g,G)$. Nevertheless, the consecutive application of the two refinements requires some non-trivial adaptations of the existing relaxation techniques underlying dimension reduction and structured deformations. Specifically, the use of $\cF_{3d,2d}(\cl u,\cl d)$ as an initial energy for relaxation in the context of structured deformations requires that $\cl u$ be a special function of bounded variation rather than a Sobolev function. Therefore, the dimension reduction has to be carried out in the $SBV$ setting. Moreover, the presence of $\cl d$ in the initial energy $\cF_{3d,2d}(\cl u,\cl d)$ for the (SD) relaxation requires a new modification of the standard relaxation techniques for structured deformations (see \cite{MMZ}). In addition, in order to connect with standard applications of dimension reduction results, the inclusion of $\cl d$ puts an additional constraint on the approximating sequences $\{u_n\}$ in the (DR) relaxation, namely $\tfrac1{\eps_{n}}\int_{I}\nabla_{3}u_{n}\,\de x_{3}\rightharpoonup \overline{d} \text{ in }L^{p}(\omega;\R3)$, see \eqref{107a}. The novelty of our approach lies partly in the incorporation of both the lack of smoothness of the function $\cl u$ (as in \cite{BF2001}) and the constraint $\tfrac1{\eps_{n}}\int_{I}\nabla_{3}u_{n}\,\de x_{3}\rightharpoonup \overline{d} \text{ in }L^{p}(\omega;\R3)$ in \eqref{107a} on the approximating sequence $\{u_n\}$ (as in \cite{BFM}), and partly in the modifications required to apply the standard (SD) relaxation introduced in \cite{CF} (see also \cite{BBBF}). Moreover, the condition $\nu(u_n)\cdot e_3=0$ in \eqref{107a} for the left-hand path and $\nu(g_n)\cdot e_3=0$ in \eqref{2019} for the right-hand path rule out the occurrence of slips and separations on surfaces with normal parallel to the thinning direction $e_3$ and place an additional constraint on the process of dimension reduction. The restriction in Theorem \ref{firstSD} to structured deformations of the form $(g,(G^{\backslash 3}|\nabla_3 g))$ is made in the same spirit for the right-hand path, since it implies that the disarrangement matrix $\nabla g-(G^{\backslash 3}|\nabla_3 g)$ has third column zero. \end{remark} The paper is organized as follows: in Section \ref{sect:prel} we set the notation and we recall some known results that are useful in the sequel, especially about $BV$ functions and $\Gamma$-convergence. In Section \ref{sect:LHS}, we follow the left-hand path of Figure \ref{figone}; namely we first derive an energy on the cross--section $\omega$ and then we relax the energy to obtain one defined on structured deformations. In Section \ref{sect:RHS}, we follow the right-hand path in Figure \ref{figone}: we first relax the energy to structured deformations and then we perform the dimension reduction. In Section \ref{sect:comp_ex}, we compare the two doubly relaxed energies from the left-hand and right-hand paths for a specific initial energy which is purely interfacial. Setting $W_{3d}\equiv0$ and $h_{3d}(\lambda,\nu)=|\lambda\cdot\nu|$ in \eqref{E3d}, we show that the two paths lead to the same relaxed energy. In particular, in Proposition \ref{S1} we give explicit formulas for the energies provided by Theorems \ref{secondleft} and \ref{secondright}, thus showing that they are equal. In Section \ref{sect:6}, we present the alternative relaxation procedure of \cite{MS} in which the introduction of disarrangements and the thinning of the domain occur simultaneously. For the specific choice of initial energy made in Section \ref{sect:comp_ex}, we prove that the relaxed energy from the scheme of \cite{MS} is identically zero. In Section \ref{conclusions}, we summarize the main results of this research and provide an outlook for future research. \section{Preliminaries}\label{sect:prel} The purpose of this section is to give a brief overview of the concepts and results that are used in the sequel. Almost all these results are stated without proof as they can be readily found in the references given below. \subsection{Notation} Throughout the manuscript, the following notation will be employed: \begin{itemize} \item[-] $\omega \subset \R2$ is a open bounded set and for $ 0<\eps\leq1$, $\Omega_{\eps} := \omega \times(-\frac{\eps}{2}, \frac{\eps}{2})$; moreover, we denote $\Omega_1$ by $\Omega$ and notice that $\Omega = \omega \times I$, with $I:= (-\frac{1}{2}, \frac{1}{2})$. \item[-] given a vector $v \in \R3$, we write $v:=(v_{\alpha},v_3)$, where $ v_\alpha:= (v_1, v_2) \in \R2$ is the vector of the first two components of $v$; \item[-] ${\mathcal A}(\Omega)$ (resp. ${\cA}(\omega)$) is the family of all open subsets of $\Omega $ (resp. $ \omega$); \item[-] for all $A,B\in\cA(\Omega)$ (resp. ${\cA}(\omega)$), $A\Subset B$ means that there exists a compact subset $C$ of $\Omega$ (resp. $\omega$) such that $A\subset C\subset B$; \item[-] $\cM (\Omega)$ (resp. $\cM(\omega)$) is the set of finite Radon measures on $\Omega$ (resp. $\omega$); \item[-] $\cL^{N}$ and $\cH^{N-1}$ stand for the $N$-dimensional Lebesgue measure and the $\left( N-1\right)$-dimensional Hausdorff measure in $\R N$, respectively; \item[-] $\lVert\mu\rVert$ stands for the total variation of a measure $\mu\in \cM (\Omega)$ (resp. $\cM (\omega)$); \item[-] $\S{N-1}$ stands for the unit sphere in $\R N$; \item[-] $Q:=I^3$ and $Q':=I^2$ denote the unit cubes centered at the origin of $\R3$ and $\R2$, respectively; \item[-] $Q_\eta$ (resp. $Q'_\eta$) denotes the unit cube of $\R3$ (resp. $\R2$) centered at the origin with two sides perpendicular to the vector $\eta\in\S2$ (resp. $\eta\in\S1$); \item[-] $Q(x, \delta):=x+\delta Q$, $Q_\eta(x, \delta):=x+\delta Q_\eta$ in $\R3$, and $Q'(x, \delta):=x+\delta Q'$, $Q'_\eta(x, \delta):=x+\delta Q'_\eta$ in $\R2$; \item[-] for $\eta \in \S{1}$, we define $\tilde{\eta} \in \S2$ by $\tilde{\eta} := (\eta, 0)$; \item[-] $C$ represents a generic positive constant that may change from line to line; \item[-] $\lim_{\delta, n} := \lim_{\delta \to 0^+} \lim_{n \to \infty}, \; \lim_{k,n} := \lim_{k \to \infty} \lim_{n \to \infty}$. \end{itemize} \subsection{$BV$ functions}\label{sect:BV} We start by recalling some facts on functions of bounded variation which will be used afterwards. We refer to \cite{AFP} and the references therein for a detailed theory on this subject. Only in this subsection, $\Omega$ denotes a generic open set in $\R{N}$. A function $u \in L^1(\Omega; \R d)$ is said to be of {\em bounded variation}, and we write $u \in BV(\Omega; \R d)$, if its first distributional derivatives $D_j u_i$ are in $\cM (\Omega)$ for $i=1,\ldots,d$ and $j=1,\ldots,N.$ The matrix-valued measure whose entries are $D_j u_i$ is denoted by $Du.$ The space $BV(\Omega; \R d)$ is a Banach space when endowed with the norm $$\lVert u\rVert_{BV} := \lVert u\rVert_{L^1} + \lVert Du\rVert(\Omega).$$ By the Lebesgue Decomposition theorem $Du$ can be split into the sum of two mutually singular measures $D^{a}u$ and $D^{s}u$ (the absolutely continuous part and the singular part, respectively, of $Du$ with respect to the Lebesgue measure $\mathcal{ L}^N$). By $\nabla u$ we denote the Radon-Nikod\'{y}m derivative of $D^{a}u$ with respect to $\mathcal L^N$, so that we can write $$Du= \nabla u \mathcal L^N \res \Omega + D^{s}u.$$ Let $\Omega_u$ be the set of points where the approximate limits of $u$ exists and $S(u)$ the {\it jump set} of this function, i.e., the set of points $x\in \Omega\setminus \Omega_u$ for which there exists $a, \,b\in \R N$ and a unit vector $\nu \in \S{N-1}$, normal to $S(u)$ at $x$, such that $a\neq b$ and \begin{equation} \label{jump1} \lim_{\delta \to 0^+} \frac {1}{\delta^N} \int_{\{ y \in Q_{\nu}(x,\delta) : (y-x)\cdot\nu > 0 \}} | u(y) - a| \,\de y=0 \end{equation} and \begin{equation}\label{jump2} \lim_{\delta \to 0^+} \frac {1}{\delta^N} \int_{\{ y \in Q_{\nu}(x,\delta) : (y-x)\cdot\nu < 0 \}} | u(y) - b| \,\de y = 0. \end{equation} The triple $(a,b,\nu)$ is uniquely determined by \eqref{jump1} and \eqref{jump2}, up to permutation of $(a,b)$ and a change of sign of $\nu$, and it is denoted by $\left(u^+ (x),u^- (x),\nu(u) (x)\right)$. If $u \in BV(\Omega;\R{d})$ it is well known that $S(u)$ is countably $(N-1)$-rectifiable, i.e., $$S(u) = \bigcup_{n=1}^{\infty}K_n \cup K_0,$$ where ${\mathcal H}^{N-1}(K_0) = 0$ and $K_n$ are compact subsets of $C^1$ hypersurfaces. Furthermore, ${\mathcal H}^{N-1}((\Omega\setminus \Omega_u) \setminus S(u)) = 0$ and the following decomposition holds $$Du= \nabla u \mathcal L^N \res \Omega + [u] \otimes \nu(u) {\mathcal H}^{N-1}\res S(u) + D^c u,$$ where $[u]:= u^+ - u^-$ and $D^c u$ is the Cantor part of the measure $Du,$ i.e., $D^c u= D^{s}u\res (\Omega_u)$. The space of \emph{special functions of bounded variation}, $SBV(\Omega; \R d)$, introduced in \cite{DGA} to study free discontinuity problems, is the space of functions $u \in BV(\Omega; \R d)$ such that $D^cu = 0$, i.e. for which $$ Du = \nabla u \cL^N + [u] \otimes \nu(u) \cH^{N-1} \res S(u).$$ We next recall some properties of BV functions used in the sequel. We start with the following lemma whose proof can be found in \cite{CF}. \begin{lemma}\label{ctap} Let $u \in BV(\Omega; \R d)$. There exists a sequence of piecewise constant functions ${u_n}\in SBV(\Omega;\R{d})$ such that $u_n \to u$ in $L^1(\Omega; \R d)$ and $$\lVert Du\rVert(\Omega) = \lim_{n\to \infty}\lVert Du_n\rVert(\Omega) = \lim_{n\to \infty} \int_{S (u_n)} |[u_n](x)|\; \de{\mathcal H}^{N-1}(x).$$ \end{lemma} The next result is a Lusin-type theorem for gradients due to Alberti \cite{AL} and is essential to our arguments. \begin{theorem}\label{Al} Let $f \in L^1(\Omega; \R{d{\times} N})$. Then there exist $u \in SBV(\Omega; \R d)$ and a Borel function $g: \Omega\to\R{d{\times} N}$ such that $$ Du = f {\cL}^N + g {\mathcal H}^{N-1}\res S(u),$$ $$ \int_{S(u)} |g| \; \de \cH^{N-1}(x) \leq C \lVert f\rVert_{L^1(\Omega; \R{d {\times} N})},$$ for some constant $C>0$. Moreover, $\lVert u\rVert_{L^1(\Omega;\R d)} \leq 2 C\lVert f\rVert_{L^1(\Omega; \R{d {\times}N})}$. \end{theorem} \subsection{$\Gamma$-convergence and relaxation}\label{Gc} We recall now the basics of $\Gamma$-convergence: this is a notion of convergence, introduced by De Giorgi and Franzoni \cite{DGF}, which is useful in the calculus of variations. It allows one to study the convergence of (sequences of) variational functions by identifying their variational limit. One of the most important products of the theory of $\Gamma$-convergence is the convergence of minima (see Remark \ref{PGC}). We refer the reader to \cite{Braides,DM} for treatises on the topic and we collect here the most important definitions and results. Let $X$ be a metric space and let $\{F_n\}$ be a sequence of functions $F_n\colon X\to\overline{\R{}}$ with values in the extended reals $\overline{\R{}}$. \begin{definition}[{\cite[Definition 1.5]{Braides}}]\label{GammaCVG} We say that the sequence $\{F_n\}$ \emph{$\Gamma$-converges} in $X$ to $F\colon X\to\overline{\R{}}$ if for all $x\in X$ we have \begin{itemize} \item[(i)] ($\liminf$ inequality) for every sequence $\{x_n\}$ converging to $x$ \begin{equation}\label{BGLI} F(x)\leq\liminf_{n\to\infty} F_n(x_n); \end{equation} \item[(ii)] ($\limsup$ inequality) there exists a sequence $\{x_n\}$ converging to $x$ such that \begin{equation}\label{BGLS} F(x)\geq\limsup_{n\to\infty} F_n(x_n). \end{equation} \end{itemize} The function $F$ is called the \emph{$\Gamma$-limit} of $\{F_n\}$, and we write $F=\Gamma-\lim_{n\to\infty} F_n$. \end{definition} When $X$ is an arbitrary topological space (in particular, it is not a metric space), a more general, \emph{topological}, definition of $\Gamma$-convergence can be given in terms of the topology of $X$. We refer the reader to \cite[Section 1.4]{Braides} and \cite[Definition 4.1]{DM} for the details. It is not difficult to see that inequalities \eqref{BGLI} and \eqref{BGLS} imply that \begin{equation}\label{EQinfima} F(x)=\inf\Big\{\liminf_{n\to\infty} F_n(x_n): x_n\to x\Big\}=\inf\Big\{\limsup_{n\to\infty} F_n(x_n): x_n\to x\Big\}, \end{equation} stating that the $\Gamma$-limit exists if and only if the two infima in \eqref{EQinfima} are equal. Other equivalent definitions can be found, for instance, in \cite[Theorem 1.17]{Braides}; moreover, the first infimum in \eqref{EQinfima} justifies the following definition. \begin{definition}[{\cite[Definition 1.24]{Braides}}]\label{defLGL} Let $F_n\colon X\to\overline{\R{}}$ and let $x\in X$. The quantity \begin{equation}\label{eqLGL} \Gamma-\liminf_{n\to\infty} F_n(x):=\inf\Big\{\liminf_{n\to\infty} F_n(x_n): x_n\to x\Big\} \end{equation} is called the \emph{$\Gamma$-lower limit} of the sequence $\{F_n\}$ at $x$. \end{definition} The $\Gamma$-lower limit defined in \eqref{eqLGL} is useful to treat relaxation in the framework of $\Gamma$-convergence. Recall that the operation of relaxation is useful to treat functionals that are not lower semicontinuous - and therefore the direct method of calculus of variations cannot be applied to minimize them. Relaxing a function means to compute its lower semicontinuous envelope. \begin{definition}[{\cite[Definition 1.30]{Braides}}]\label{defSCE} Let $F\colon X\to\overline{\R{}}$ be a function. Its \emph{lower semicontinuous envelope} $\sce F$ is the greatest lower semicontinuous function not greater than $F$, that is, for every $x\in X$ \begin{equation}\label{eqSCE} \sce F(x):=\sup\{G(x): G\text{ is lower semicontinuous and }G\leq F\}. \end{equation} \end{definition} In view of \cite[Proposition 1.31]{Braides} and \cite[Remark 4.5]{DM}, relaxation is equivalent to computing the $\Gamma$-limit of a constant sequence of functions, \emph{i.e.}, $F_n=F$ for all $n$. \begin{proposition}[{see \cite[Proposition 1.32]{Braides}}]\label{B132} We have $\Gamma-\liminf_{n\to\infty} F_n=\Gamma-\liminf_{n\to\infty} \sce F_n$. \end{proposition} In view of the previous proposition, the left-hand path and the right-hand path described in the Introduction consist in the computation of two $\Gamma$-lower limits, where the order is exchanged. For simplicity, in the paper we will use the words ``$\Gamma$-lower limit of a family of functionals'' and ``relaxation of energies'' interchangeably. \begin{remark}\label{PGC} Among the properties that make $\Gamma$-convergence a suitable tool for the study of the convergence of functional and related variational problems, three are particularly useful, namely \begin{itemize} \item the \emph{compactness of $\Gamma$-convergence} (see \cite[Section 1.8.2]{Braides}, \cite[Chapter 8]{DM}). In particular, for each sequence of functions $F_n\colon X\to\overline{\R{}}$ the compactness property grants the existence of a $\Gamma$-convergent subsequence, provided $X$ has a countable base \cite[Theorem 8.5]{DM}. Then, it is not difficult to imagine that the choice of the topology in the convergences that define the relaxed functionals $\cF_{3d,2d}$, $\cF_{3d,2d,SD}$, $\cF_{3d,SD}$, and $\cF_{3d,SD,2d}$ below (see \eqref{107a}, \eqref{200}, \eqref{2018}, and \eqref{2019}, respectively) is made in order to obtain good compactness properties. \item the \emph{stability under continuous perturbations} (see \cite[Remark 1.7]{Braides}, \cite[Proposition 6.21]{DM}, and also \cite[Proposition 3.7]{DM} for the relaxation): if $\widetilde F\colon X\to\R{}$ is a continuous function, then \begin{equation}\label{CP} \Gamma-\liminf_{n\to\infty}(F_n+\widetilde F)=\Gamma-\liminf_{n\to\infty} F_n+ \widetilde F,\quad \Gamma-\limsup_{n\to\infty}(F_n+\widetilde F)=\Gamma-\limsup_{n\to\infty} F_n+ \widetilde F, \end{equation} so that if $\{F_n\}$ $\Gamma$-converges to $F$ in $X$, then $\{F_n+\widetilde F\}$ $\Gamma$-converges to $F+\widetilde F$ in $X$. \item the implications regarding the \emph{convergence of minima and minimizers}. The results contained in \cite[Section 1.5]{Braides} and \cite[Chapter 7]{DM} give conditions under which the $\Gamma$-convergence of a sequence of functions $F_n$ to their $\Gamma$-limit $F$ implies the convergence of the minimima \begin{equation}\label{CMM} \min_{x\in X} F(x)=\lim_{n\to\infty} \inf_{x\in X} F_n(x) \end{equation} and of the minimizers: if $\{F_n\}$ is equi-coercive and $\Gamma$-converges to $F$, with a unique minimum point $x_0\in X$, and if $\{x_n\}\subset X$ is a sequence such that $x_n$ is an $\eps_n$-minimizer for $F_n$ in $X$ for every $n$, and with $\eps_n\to0^+$, then $x_n\to x_0$ in $X$ and $F_n(x_n)\to F(x_0)$. We direct the reader to \cite{Braides,DM} for a precise statement of the notions of equi-coercivity and $\eps$-minimizer (albeit they are quite natural to understand). \end{itemize} We will not make use of the last two properties of $\Gamma$-convergence in this paper. We think it is worthwhile mentioning them in the spirit of a variational treatment of the minimization of the relaxed functionals that we obtain in our results, with the hope to indicate to the reader that the theorems exposed and proved in the sequel can provide a starting point for the study of equilibrium configurations of thin structures in the framework of structured deformations. \end{remark} \begin{remark} All the definitions and results presented above can be generalized to the case of families of functionals, indexed by a continuous parameter $\eps$. A family of functions $\{F_\eps\}$ $\Gamma$-converges in $X$ to $F\colon X\to \overline{\R{}}$ as $\eps\to0^+$ if, for every sequence $\eps_n\to0^+$, the functions $\{F_{\eps_n}\}$ $\Gamma$-converge to $F$ in the sense of Definition \ref{GammaCVG} (see, \emph{e.g.}, \cite[Section 1.9]{Braides}). \end{remark} \section{The left-hand path}\label{sect:LHS} In this section we relax our initial energy \eqref{E3d} by first doing dimension reduction and then by incorporating structured deformations. \subsection{Dimension reduction}\label{LHSDR} In order to perform dimension reduction, we resort to the classical approach of rescaling the spatial variable by dividing $x_3$ by $\eps$ and integrating over the rescaled domain $\Omega=\omega\times(-1/2,1/2)$. We also rescale the functional \eqref{E3d} by $\eps$, defining $F_{\eps}(u):=\tfrac1\eps E_{\eps}(u)$, so that we have \begin{equation}\label{107} F_{\eps}(u)=\int_{\Omega} W_{3d}\bigg(\nabla_{\alpha}u\bigg|\frac{\nabla_{3}u}{\eps}\bigg)\de x+\int_{\Omega\cap S(u)} h_{3d}\bigg([u],\nu_{\alpha}(u)\bigg|\frac{\nu_{3}(u)}{\eps}\bigg)\de \mathcal{H}^{2}(x). % \end{equation} Let now $(\overline{u},\overline{d})\in SBV(\omega;\mathbb{R}^{3})\times L^{p}(\omega;\mathbb{R}^{3})$, let $\eps_n\to0^+$, and define the relaxed functional \begin{equation}\label{107a} \begin{split} \cF_{3d,2d}(\overline{u},\overline{d}):=\inf & \bigg\{\liminf_{n\to\infty} F_{\eps_n}(u_n): u_{n}\in SBV(\Omega;\mathbb{R}^{3}),\; u_{n}\rightarrow\overline{u}\text{ in }L^{1}(\Omega;\R3), \\ & \quad\int_{I}\frac{\nabla_{3}u_{n}}{\eps_{n}}\,\de x_{3}\rightharpoonup \overline{d} \text{ in }L^{p}(\omega;\R3),\; \nu(u_{n})\cdot e_{3}=0\bigg\}. \end{split} \end{equation} In writing the convergence $u_n\to\cl u$ in $L^1(\Omega;\R3)$ in formula \eqref{107a}, it is understood that $\cl u$ is extended to a function on $\Omega$ which is independent of $x_3$. As stated in Remark \ref{energies}, the coercivity assumption \eqref{coerc} grants boundedness of the gradients in $L^p$, so that $\eps_n^{-1}\nabla_{3}u_{n}\wto d$ in $L^p$, where $d$ may still depend on the $x_3$ variable. By contrast, following the model in \cite{BFM}, we consider the weak convergence of the average with respect to the third variable to a field $\overline d(x_\alpha)$ depending only on the coordinates in the cross--section $\omega$. \begin{theorem}\label{T3d2d} Under the hypotheses $(H_1)$--$(H_4)$, let $(\cl{u},\cl{d}) \in SBV(\omega;\mathbb{R}^{3}) \times L^{p}(\omega ;\mathbb{R}^{3})$. Then every sequence $\eps_n\to0^+$ admits a subsequence such that \begin{equation}\label{F3d2d} \cF_{3d,2d}(\cl{u},\cl{d}) =\int_{\omega}W_{3d,2d}(\nabla \cl{u},\cl{d})\,\de x_{\alpha}+\int_{\omega \cap S(\cl{u})}h_{3d,2d}([\cl{u}],\nu(\cl{u}))\,\de\mathcal{H}^{1}(x_\alpha), \end{equation} where $W_{3d,2d}\colon \mathbb{R}^{3\times2}\times \mathbb{R}^3\to \lbrack0,+\infty)$ and $h_{3d,2d}\colon\mathbb{R}^3\times \mathbb{S}^1\to \lbrack0,+\infty)$ are given by \begin{equation}\label{107b} \begin{split} W_{3d,2d}(A,d)= & \inf\bigg\{\int_{Q^{\prime}}W_{3d}(\nabla_{\alpha}u|z)\,\de x_{\alpha}+\int_{Q^{\prime}\cap S(u)}h_{3d}([u],\tilde\nu(u))\,\de\mathcal{H}^{1}(x_\alpha): \\ & \qquad u\in SBV(Q^{\prime};\mathbb{R}^{3}),~z\in L^p_{{Q'}-\operatorname*{per}}(\mathbb{R}^{2};\mathbb{R}^{3}),~u|_{\partial Q^{\prime}}(x_{\alpha})=Ax_{\alpha}, \int_{Q^{\prime}}z\,\de x_{\alpha}=d\bigg\} , \end{split} \end{equation} and, for $\lambda\in\R3$, $\eta\in\mathbb{S}^1$, \begin{equation}\label{107c} \begin{split} h_{3d,2d}(\lambda,\eta)= &\inf\bigg\{\int_{Q^{\prime}_{\eta}\cap S(u)}h_{3d}([u],\tilde\nu(u))\; \de\mathcal{H}^{1}(x_\alpha): u\in SBV(Q_{\eta}^{\prime};\mathbb{R}^{3}), \\ &\qquad u|_{\partial Q_{\eta}^{\prime}}(x_{\alpha})=\gamma_{\lambda,\eta}( x_{\alpha}) ,~\nabla u=0, \text{ a.e.} \bigg\} \end{split} \end{equation} with \begin{equation}\label{gammale} \gamma_{\lambda,\eta}(x_{\alpha}):= \begin{cases} \lambda & \text{if $0\leq x_{\alpha}\cdot\eta<\frac{1}{2}$},\\ 0 & \text{if $-\frac{1}{2}<x_{\alpha}\cdot\eta<0$}. \end{cases} \end{equation} \end{theorem} In \eqref{107b} and in the sequel, the notation $z\in L^p_{{Q'}-\operatorname*{per}}(\mathbb{R}^{2};\mathbb{R}^{3})$ means that the function $z$ is defined on the unit square $Q'$ and extended by periodicity to all of $\R2$. We set the stage for the proof of Theorem \ref{T3d2d} by proving some properties of the energy densities defined by \eqref{107b} and \eqref{107c}, which will be used in the sequel. \begin{proposition}\label{p100} Let $W_{3d, 2d}$ and $h_{3d,2d}$ be given by \eqref{107b} and \eqref{107c}, respectively. The following properties hold: \begin{itemize} \item [\textit{(i)}] $W_{3d,2d}$ satisfies \eqref{growth}, namely, for each $A,B\in\R{3\times2}$ and $d,e\in\R3$, \begin{equation}\label{wishful} |W_{3d,2d}(A,d)-W_{3d,2d}(B,e)|\leq C|(A|d)-(B|e)|(1+|(A|d)|^{p-1}+|(B|e)|^{p-1}); \end{equation} \item [\textit{(ii)}] $h_{3d,2d}$ satisfies $(H_2)$--$(H_4)$ and it is Lipschitz continuous with respect to the variable $\lambda$; \item [\textit{(iii)}] $h_{3d,2d}$ is upper semicontinuous with respect to the variable $\eta$. \end{itemize} \end{proposition} \begin{proof} \textit{(i)} Let $\Pi_\alpha\colon\R{3\times3}\to\R{3\times2}$ and $\Pi_3\colon\R{3\times3}\to\R{3}$ be the linear maps which select out the first two and the third columns, respectively, of a matrix $M\in\R{3\times3}$. Note that $W_{3d,2d}(A,d)=W_{3d,2d}(\Pi_\alpha M,\Pi_3 M)$, for $M=(A|d)$. By applying \cite[Proposition 5.6(i)]{CF} with $M\mapsto W_{3d,2d}(\Pi_\alpha M,\Pi_3 M)$ in place of $A\mapsto H_p(A,A)$, we obtain that $(A,d)\mapsto W_{3d,2d}(A,d)$ is quasiconvex (see, \emph{e.g.}, \cite[Section 2]{FM}). This, combined with \eqref{pgrowth}, by a standard argument by Marcellini \cite[Theorem 2.1]{M}, implies that $W_{3d,2d}$ satisfies \eqref{wishful}. \textit{(ii)} Properties $(H_2)$--$(H_4)$ for $h_{3d, 2d}$ follow by standard arguments from \eqref{107c}. To prove Lipschitz continuity in the first variable, consider $\lambda_1,\lambda_2\in\R3$, $\eta\in\S1$, and $\rho>0$. Let now $u\in SBV(Q'_\eta;\R3)$ be admissible for $h_{3d,2d}(\lambda_1,\eta)$ in \eqref{107c} and be such that \begin{equation}\label{57} h_{3d,2d}(\lambda_1,\eta)+\rho\geq\int_{Q'_\eta\cap S(u)} h_{3d}([u],\tilde\nu(u))\,\de\cH^1(x_\alpha). \end{equation} Then, $v := u + \gamma_{\lambda_2-\lambda_1, \eta}$ is an admissible function for the definition of $h_{3d, 2d}(\lambda_2, \eta)$ and, in view of the subadditivity of $h_{3d}$, $(H_2)$, and \eqref{57}, we have that: \begin{equation*} \begin{split} h_{3d,2d}(\lambda_2,\eta) \leq \int_{Q'_\eta\cap S(v)} h_{3d}([v], \tilde\nu(v))\, \de \cH^1(x_\alpha) & \leq \int_{Q'_\eta\cap S(u)} h_{3d}([u],\tilde\nu(u))\, \de\cH^1(x_\alpha) + C|\lambda_2 - \lambda_1| \\ &\leq h_{3d, 2d}( \lambda_1, \eta) + \rho + C|\lambda_2 - \lambda_1|. \end{split} \end{equation*} Letting $\rho \to 0$ and reversing the roles of $u$ and $v$ we conclude the proof of \textit{(ii)}. \textit{(iii)} The proof can be found in \cite[Prop.\@ 3.6]{BBBF}. \qed \end{proof} We prove next that, for a fixed piecewise constant $\cl d\in L^p(\omega;\R3)$, the functional $\cF_{3d,2d}^{\cl d}(\cl{u}) := \cF_{3d,2d}(\cl{u},\cl{d})$ is the trace of a Radon measure. To do this, we follow arguments in \cite{MS}; we start by localizing $\cF_{3d,2d}^{\cl d}(\cl{u})$, i.e., for an open set $A \in \mathcal{A}(\omega)$, $\cl u\colon\omega\to\R3$, and $\eps_n\to0^+$, we define \begin{equation}\label{557} \begin{split} \cF_{3d,2d}^{\overline{d}}(\overline{u};A):=\inf & \bigg\{\liminf_{n\to\infty}\bigg(\!\int_{A \times I}W_{3d}\Big(\nabla_{\alpha}u_{n}\Big|\frac{\nabla_{3}u_{n}}{\eps_{n}}\Big)\de x+\!\int_{(A \times I)\cap S(u_{n})} \!\! h_{3d}\Big([u_{n}],\nu_{\alpha}(u_{n})\Big|\frac{\nu_{3}(u_{n})}{\eps_{n}}\Big)\de\mathcal{H}^{2}(x)\bigg): \\ & \quad u_{n}\in SBV(\Omega;\mathbb{R}^{3}),\; u_{n}\rightarrow\overline{u}\text{ in }L^{1},\; \int_{I}\frac{\nabla_{3}u_{n}}{\eps_{n}}\,\de x_{3}\rightharpoonup \overline{d} \text{ in }L^{p},\; \nu(u_{n})\cdot e_{3}=0\bigg\}. \end{split} \end{equation} Notice that the functional defined in \eqref{557} depends on the particular sequence $\{\eps_n\}$ (but for simplicity we do not write it explicitly). Then we have the following result. \begin{proposition}\label{569} Let $W_{3d}\colon{\mathbb{R}}^{3\times3}\rightarrow\lbrack0,+\infty)$ and $h_{3d}\colon{\mathbb{R}}^{3}\times \S{2}\rightarrow\lbrack0,+\infty)$ be continuous satisfying $(H_1)$ and $(H_2)$ and let $\cl d \in L^p(\omega; \R3)$ be piecewise constant. Any sequence $\eps_n \to 0^+$ admits a subsequence $\eps_k: = \eps_{n(k)}$ such that for $\cl u \in SBV(\omega; \R3)$ the set function $\cF_{3d,2d}^{\overline{d}}(\overline{u};\cdot)$ defined in \eqref{557}, is the trace of a Radon measure on $\cA(\omega)$ which is absolutely continuous with respect to $\mathcal{L}^2 + \mathcal{H}^1\res {S(\cl u)}.$ \end{proposition} \begin{proof} We start by noting that, considering the admissible sequence $u_n := \cl u + \eps_n x_3 \cl d$, by $(H_1)$ and $(H_2)$ the following upper bound holds $$\cF_{3d,2d}^{\cl d}(\cl u;A) \leq C\bigg( \mathcal{L}^2 (A) + \int_A |\nabla \cl u|^p \de x_\alpha + \int_A |\cl d|^p \de x_\alpha + \|D \cl u \|(\overline{A})\bigg).$$ For each $a\in\omega$ with rational coordinates and for $i\in\N{}$, consider balls $B(a;r_i)$ with radii $r_i$ and depending on $a$, such that \begin{equation}\label{577} \left | r_i - \frac{1}{i}\right| \leq \frac{1}{i^2},\; \; \; \; \overline{B(a; r_i)} \subset \omega,\; \; \; \; \|D^s \cl u\|(\partial B(a; r_i)) = 0.\end{equation} Let $\mathcal{B}(\omega)$ be the set of all such balls and their finite unions. The set of all closed balls $\overline{B(a; r_i)} $ is a fine cover of $\omega$ (see \cite[p.\@ 49]{AFP}). Given a sequence $\eps_n \to 0^+$, by a standard diagonalization argument, we can take an appropriate subsequence $\eps_k: = \eps_{n(k)}$ such that, for each $B \in \mathcal{B}(\omega)$, we may find a sequence $u_k$ (depending on $B$) such that \begin{equation}\label{580} u_{k}\rightarrow\overline{u}\text{ in }L^{1},\quad\int_{ I}\frac{\nabla_{3}u_{k}}{\eps_{k}}\,\de x_{3}\rightharpoonup \cl{d} \text{ in }L^{p},\quad \nu(u_{k}) \cdot e_{3}=0, \end{equation} and \begin{equation}\label{585} \cF_{3d,2d}^{\cl d}(\cl u;B) = \lim_{k\to \infty}\bigg(\int_{B \times I}W_{3d}\Big(\nabla_{\alpha}u_{k}\Big|\frac{\nabla_{3}u_{k}}{\eps_{k}}\Big)\de x+\int_{(B \times I)\cap S(u_{k})}h_{3d}\Big([u_{k}],\nu_{\alpha}(u_{k})\Big|\frac{\nu_{3}(u_{k})}{\eps_{k}}\Big)\de\mathcal{H}^{2}(x)\bigg). \end{equation} Next, we prove the following subadditivity property: for every $B, B_1, B_2 \in \mathcal{A}(\omega)$ such that $ B_1 \Subset B \Subset B_2$, we have that \begin{equation}\label{591} \cF_{3d,2d}^{\cl d}(\cl u;B_2) \leq \cF_{3d,2d}^{\cl d}(\cl u;B) + \cF_{3d,2d}^{\cl d}(\cl u;B_2 \backslash \overline{B}_1). \end{equation} To this end, for each $B \in \mathcal{A}(\omega)$, define the Radon measure $$\Delta(B) := C \bigg(\mathcal{L}^2(B) + \int_B |\cl d|^p \de x_\alpha + \int_B |\nabla \cl u|^p \de x_\alpha + \|D^s \cl u\|(B)\bigg).$$ For fixed $\rho >0$ consider an open set $B_\rho \in \mathcal{B}(\omega)$ such that $B_\rho \subset B$. Using the Besicovitch covering theorem, we can find $A_\rho \in \mathcal B(\omega)$ such that $A_\rho \subset B_2 \backslash \overline{B}_1$ and $$\Delta \left( (B_2\setminus \overline{B}_1)\setminus \overline{A}_\rho \right) < \rho.$$ Note that we can choose the sets above such that there exists an open set $\tilde{A}$ with Lipschitz boundary and with $B_1\Subset\tilde{A} \Subset B_\rho$ and with $\partial \tilde{A} \subset A_\rho$. Now, consider $\{u_k^1\} \in SBV(A_\rho; \R3)$ and $\{u_k^2\} \in SBV(B_\rho; \R3)$ satisfying \eqref{580} and \eqref{585}, and define $$\tilde{u}_k := \left\{ \begin{array}{ll} u_k^1 & \text{in}\; A_\rho \setminus \tilde{A}\\ u_k^2 & \text{in}\; \tilde{A}\\ u_k & \text{otherwise in $B_2$,} \end{array} \right. $$ where $u_k(x_\alpha, x_3):= \overline{u}(x_\alpha) + \eps_k x_3 \overline{d}(x_\alpha)$. Notice that $\tilde{u}_k \in SBV(B_2; \R3)$ by \cite[Proposition 3.21]{AFP}. Then we have that \begin{align*} \cF_{3d,2d}^{\overline{d}}(\overline{u};B_2)&\leq \lim_{k\to\infty} \bigg( \int_{B_2 \times I} W_{3d}\Big(\nabla_\alpha \tilde{u}_k \Big| \frac{\nabla_3 \tilde{u}_k}{\eps_k}\Big)\de x + \int_{(B_2 \times I)\cap S(\tilde{u}_k)} h_{3d}([\tilde{u}_k], \tilde\nu_\alpha(\tilde{u}_k))\,\de \mathcal{H}^2(x)\bigg)\\ & \leq \cF_{3d,2d}^{\overline{d}}(\overline{u};A_\rho) + \cF_{3d,2d}^{\overline{d}}(\overline{u};B_\rho) + \Delta\left((B_2\setminus \overline{B}_1)\setminus \overline{A}_\rho \right)\\ &\leq \cF_{3d,2d}^{\overline{d}}(\overline{u};A_\rho) + \cF_{3d,2d}^{\overline{d}}(\overline{u};B_\rho)+\rho\\ &\leq \cF_{3d,2d}^{\overline{d}}(\overline{u};B_2\setminus \overline{B}_1) + \cF_{3d,2d}^{\overline{d}}(\overline{u};B)+ \rho. \end{align*} Note that, since $A_\rho \setminus \tilde{A} \Subset A_\rho, \tilde{A} \Subset B_\rho$ and by \eqref{577}, the jumps of $\tilde{u}_k$ in the transition layers are included in the computations above. By letting $\rho \to 0$, we have that \eqref{591} holds. In the following, let $u_k = u_k^{\omega}$ denote an appropriate sequence for which \eqref{585} holds in $\omega$. Define the sequence of bounded Radon measures $$\Lambda_k(A) := \int_{A\times I} W_{3d}\Big(\nabla_\alpha u_k \Big| \frac{\nabla_3 u_k}{\eps_k}\Big) \; dx + \int_{(A \times I)\cap S(u_k)} h_{3d}([u_k], \tilde\nu_\alpha (u_k)) \; \de \mathcal{H}^2(x),$$ for $A \in \mathcal{A}(\omega)$ and extract a subsequence (not relabeled) such that $\Lambda_k \wsto \Lambda$. In order to complete the proof we show that for every $A \in \mathcal{A}(\omega)$ we have that \begin{equation}\label{616} \cF_{3d,2d}^{\overline{d}}(\overline{u};A) = \Lambda (A). \end{equation} Note first that for any $A \in \mathcal{A}(\omega)$, open set, the following inequality holds \begin{equation}\label{621} \cF_{3d,2d}^{\overline{d}}(\overline{u};A) \leq \Lambda (\overline{A}). \end{equation} Given $B \in \mathcal{A}(\omega)$, let $\rho > 0$ and consider $W \Subset B$ such that $\Lambda(B \backslash W) < \rho.$ Then, since $\Lambda (\omega) = \Lambda(\overline{\omega}),$ by \eqref{591} and \eqref{621} we have that \begin{align*} \Lambda (B) &\leq \Lambda( W) + \rho\\ & = \Lambda (\omega) - \Lambda(\omega \setminus W) + \rho\\ & \leq \cF_{3d,2d}^{\overline{d}}(\overline{u};\omega) - \cF_{3d,2d}^{\overline{d}}(\overline{u};\omega \setminus \overline{W}) + \rho\\ & \leq \cF_{3d,2d}^{\overline{d}}(\overline{u};B) + \rho. \end{align*} Letting $\rho \to 0$ we have that \begin{equation}\label{632} \Lambda (B) \leq \cF_{3d,2d}^{\overline{d}}(\overline{u};B). \end{equation} Finally, it remains to prove the reverse inequality. Let now $K \subset B$ be a compact set such that $\Delta(B\setminus K) < \rho$ and choose an open set $D$ such that $ K \Subset D \Subset B$. Again, by \eqref{591} we have \begin{align*} \cF_{3d,2d}^{\overline{d}}(\overline{u};B) & \leq \cF_{3d,2d}^{\overline{d}}(\overline{u};D) + \cF_{3d,2d}^{\overline{d}}(\overline{u};B \setminus K)\\ & \leq \Lambda (\overline{D}) + \Delta (B\setminus K)\\ & \leq \Lambda (B) + \rho, \end{align*} which, together with \eqref{632} and letting $\rho \to 0$, yields the result. \qed \end{proof} Notice that, by Proposition \ref{569}, for every sequence $\{\eps_n\}$, the localized functional $\cF_{3d,2d}^{\overline{d}}(\overline{u};\cdot)$ defined in \eqref{557} is the trace of a Radon measure on $\cA(\omega)$, so that it admits an integral representation. We are now ready to prove Theorem \ref{T3d2d}, and we divide its proof into four steps, each of which relies on the blow-up method of \cite{FM}: we will prove upper bounds for the Radon-Nikod\'{y}m derivatives of $\cF_{3d,2d}(\cl u,\cl d)$ with respect to $\cL^2$ and $\cH^1\res S(\cl u)$ at a point $x_0\in\omega$ (see \eqref{1151} and \eqref{uppersurf}, respectively), and lower bounds for the Radon-Nikod\'{y}m derivatives of a certain measure $\mu$ (the weak-* limit of the measures $\mu_n$ defined in \eqref{miu_n}) with respect to $\cL^2$ and $\vert [\cl u] \vert\mathcal{H}^{1}\res S(\cl u)$ (see \eqref{778} and \eqref{779}, respectively). We will find that these upper and lower bounds are indeed independent of the particular choice of the sequence $\eps_n\to0^+$, so that estimates \eqref{1151}, \eqref{uppersurf}, \eqref{778}, and \eqref{779} will suffice to conclude the proof of the theorem. Moreover, we note the connection with the theory of $\Gamma$-convergence presented in Section \ref{Gc}: Steps 1 and 2 correspond to proving the $\limsup$ inequality \eqref{BGLS}, Steps 3 and 4 correspond to proving the $\liminf$ inequality \eqref{BGLI}. \paragraph{\textit{\textbf{Step 1} (Upper bound -- bulk)}} We start by noticing that, by Lemma \ref{ctap} and \eqref{growth}, it is enough to derive the upper bound for the case where $\overline{d}$ is piecewise constant. In fact, given $u_n$ admissible for $\cF_{3d,2d}(\cl u, \cl d)$ and $\cl d_k$ a piecewise constant approximation of $\cl d$ given by Lemma \ref{ctap}, for each $k$ we can obtain an admissible sequence $u_{k,n}$ for $\cF_{3d,2d}(\cl u,\cl d_k)$ by defining $u_{k,n}:=u_n+h_{k,n}$, where $h_{k,n}$ is provided by Theorem \ref{Al} in such a way that $\nabla h_{k,n} =\eps_n \Big(0 \Big| \cl d_k - \int_I \frac{\nabla_3 u_n}{\eps_n}\de x_3\Big)$ and $||h_{k,n}||_{L^1(\Omega;\R3)}\leq C\eps_n\big(||\cl d_k||_{L^p(\omega;\R3)}+||\cl d||_{L^p(\omega;\R3)}\big)$. Therefore, $$ \cF_{3d,2d}(\cl u, \cl d) \leq \liminf_{k \to \infty}\cF_{3d,2d}( \cl u, d_k) \leq \limsup_{k \to \infty} \bigg(\int_\omega W_{3d, 2d}(\nabla \cl u, d_k)\, \de x_\alpha + \int_{\omega \cap S(\cl u)} h_{3d, 2d}([\cl u], \nu (\cl u))\, \de \mathcal{H}^1(x_\alpha)\bigg),$$ and the result follows because $W_{3d, 2d}$ has growth of order $p$ (see Proposition \ref{p100}). Let $(\overline{u},\overline{d})\in SBV(\omega;\mathbb{R}^3)\times L^{p}(\omega;\mathbb{R}^{3})$, with $\cl d$ piecewise constant, and let $x_{0}\in\omega$ be chosen such that \begin{equation}\label{apcontu} \lim_{\delta\to0}\frac{1}{\delta^{2}} |D^s \overline{u}|(Q'(x_0, \delta)) = 0, \end{equation} \begin{equation}\label{apcont} \lim_{\delta\to0}\frac{1}{\delta^{2}}\int_{Q'(x_{0},\delta)}\left\vert \overline{d}(x_\alpha)-\overline{d}(x_{0})\right\vert^p \de x_\alpha=0, \end{equation} \begin{equation} \lim_{\delta\to0}\frac{1}{\delta^{2}}\int_{Q'(x_{0},\delta)}\left\vert \nabla_\alpha \cl u(x_\alpha) - \nabla_\alpha \cl u(x_0))\right\vert^p \,\de x_\alpha=0. \label{apdiff} \end{equation} It suffices to prove \begin{equation}\label{1151} \frac{\de\cF_{3d,2d}(\overline{u},\overline{d})}{\de\mathcal{L}^{2}}(x_{0})\leq W_{3d,2d}(\nabla_{\alpha}\overline{u}(x_{0}),\overline{d}(x_{0})). \end{equation} To this end, fix $\rho>0$ and choose $u\in SBV(Q';\mathbb{R}^{3})$ and $z\in L_{Q'-\operatorname*{per}}^{p}(\mathbb{R}^{2};\mathbb{R}^{3})$ piecewise constant such that \begin{equation} u|_{\partial Q'}(x_{\alpha})=\nabla_{\alpha}\overline{u}(x_{0})x_{\alpha},\qquad\int_{Q'}z(x_{\alpha})\,\de x_{\alpha}=\overline{d}(x_{0}), \label{115} \end{equation} and \begin{equation} W_{3d,2d}(\nabla_{\alpha}\overline{u}(x_{0}),\overline{d}(x_{0}))+\rho\geq\int_{Q'}W_{3d}(\nabla_{\alpha}u|z)\,\de x_{\alpha}+\int_{Q'\cap S_{u}}h_{3d}([u],\tilde{\nu})\,\de\mathcal{H}^{1}(x_\alpha). \label{114} \end{equation} We now construct a sequence ${u_{\delta,n}}$ of competitors for the problem \eqref{107a} by setting $\zeta(x_{\alpha}):=u(x_{\alpha})-\nabla_{\alpha}\overline{u}(x_{0})x_{\alpha}$ (extended by periodicity to all of $\R2$) and defining \begin{equation} u_{\delta,n}(x_{\alpha},x_{3}):=\overline{u}(x_{\alpha})+\frac{\delta}{n}\zeta\Big(\frac{n(x_{\alpha}-x_{0})}{\delta}\Big)+\eps_{n}x_{3}\Big(\overline{d}(x_{\alpha})-\overline{d}(x_{0})+z\Big(\frac{n(x_{\alpha}-x_{0})}{\delta}\Big)\Big). \label{116} \end{equation} Clearly, $u_{\delta,n} \in SBV(\Omega;\mathbb{R}^{3})$, $\lim_{\delta,n} u_{\delta, n}=\overline{u}$ in $L^1(\Omega;\R3)$, and \begin{equation}\label{119a} \int_{I}\frac{\nabla_{3}u_{n}(x_{\alpha},x_{3})}{\eps_{n}}\,\de x_{3}=\overline{d}(x_{\alpha})-\overline{d}(x_{0})+z\Big(\frac{n(x_{\alpha}-x_{0})}{\delta}\Big). \end{equation} It is not difficult to see that $z(n(x_\alpha-x_0)/\delta)\wto\int_{Q'} z(x_\alpha)\,\de x_\alpha$ in $L^p(\omega;\R3)$, so that, by \eqref{115}, the right-hand side of \eqref{119a} converges to $\cl d(x_\alpha)$ as $n\to\infty$. Notice that in the construction of ${u_{\delta, n}}$ the normal $\nu(u_{\delta,n})$ satisfies $\nu(u_{\delta,n})\cdot e_3 = 0$. Since $\overline{d}$ and $z$ are piecewise constant, we have \begin{equation}\label{121}% \nabla_{\alpha}u_{\delta, n}(x_{\alpha},x_{3})=\nabla_{\alpha}\overline{u}(x_{\alpha})+\nabla_{\alpha}\zeta\Big(\frac{n(x_{\alpha}-x_{0})}{\delta}\Big)=\nabla_{\alpha}\overline{u}(x_{\alpha})+\nabla_{\alpha}u\Big(\frac{n(x_{\alpha}-x_{0})}{\delta}\Big)-\nabla_{\alpha}\overline{u}(x_{0}). \end{equation} Therefore, recalling $(H_1)$--$(H_4)$, \begin{equation} \begin{split} \frac{\de\cF_{3d,2d}(\overline{u},\overline{d})}{\de\mathcal{L}^{2}}(x_{0})\leq \lim_{\delta,n}\frac{1}{\delta^{2}} \bigg\{& \int_{Q'(x_{0};\delta){\times}I} W_{3d}\Big(\nabla_{\alpha}u_{\delta,n}\Big| \frac{\nabla_{3}u_{\delta,n}}{\eps_{n}}\Big)\de x \\ & +\int_{(Q'(x_{0};\delta){\times}I)\cap S(u_{\delta,n})} h_{3d}([u_{\delta,n}],\tilde\nu_{\alpha}(u_{\delta,n}))\,\de\cH^{2}(x) \bigg\} \\ \leq \lim_{\delta,n}\frac{1}{\delta^{2}}\bigg\{ & \int_{Q'(x_{0};\delta)}W_{3d}\Big( \nabla_{\alpha}u\Big(\frac{n(x_{\alpha}-x_{0})}{\delta}\Big) \Big| z\Big(\frac{n(x_{\alpha}-x_{0})}{\delta}\Big)\Big) \de x_{\alpha} \label{122}\\ & + \int_{Q'(x_{0};\delta)\cap(x_{0}+\frac{\delta}{n}S(u))} h_{3d}\Big(\frac{\delta}{n}[u]\Big(\frac{n(x_{\alpha}-x_{0})}{\delta}\Big), \tilde\nu_{\alpha}(u)\Big) \de\mathcal{H}^{1}(x_{\alpha})\bigg\} \\ +\lim_{\delta\to0} \frac{1}{\delta^{2}}& \int_{Q'(x_{0};\delta)}\left[|\nabla_{\alpha}\overline{u}(x_{\alpha})-\nabla_{\alpha}\overline{u}(x_{0})|^p+|\overline{d}(x_{\alpha})-\overline {d}(x_{0})|^p\right]\,\de x_{\alpha} \\ +\lim_{\delta,n}\frac{1}{\delta^{2}} & |D^s\overline{u}|(Q'(x_0; \delta)) \\ +\lim_{\delta\to0}\frac1{\delta^2} & \limsup_{n\to\infty}{\eps_{n}}\bigg\{\int_{(Q'(x_{0};\delta){\times}I)\cap S(\cl{d})}|x_{3}(\overline{d}(x_{\alpha})-\overline{d}(x_{0}))|\,\de\mathcal{H}^{1}(x_{\alpha})\de\mathcal{L}^{1}(x_3) \\ &+\int_{(Q' \times I)\cap S(z)}\frac{\delta}{n}|x_{3}z(y_{\alpha})|\,\de\mathcal{H}^{1}(y_{\alpha})\de\mathcal{L}^{1}(x_3)\bigg\}, \end{split} \end{equation} where in the last integral we performed the change of variables $y_{\alpha }:=n(x_{\alpha}-x_{0})/\delta$. By the same change of variables and noticing that, by \eqref{apcontu}, \eqref{apcont}, \eqref{apdiff}, and the hypothesis on $z$, the last four terms in \eqref{122} vanish, we are left with \begin{equation} \begin{split} \frac{\de\cF_{3d,2d}(\overline{u},\overline{d})}{\de\mathcal{L}^{2}}(x_{0}) \leq & \lim_{\delta,n}\frac{1}{n^{2}}\bigg\{ \int_{nQ'}W_{3d}(\nabla_{\alpha}u(y_{\alpha})|z(y_{\alpha}))\,\de y_{\alpha}+\int_{nQ'\cap S(u)} h_{3d}([u](y_{\alpha}),\tilde{\nu}(u))\,\de\mathcal{H}^{1}(y_{\alpha})\bigg\} \\ \leq & \int_{Q'}W_{3d}(\nabla_{\alpha}u(y_{\alpha})|z(y_{\alpha}))\,\de y_{\alpha}+\int_{Q'\cap S(u)}h_{3d}([u](y_{\alpha}),\tilde{\nu}(u))\,\de\mathcal{H}^{1}(y_{\alpha})\\ \leq & W_{3d,2d}(\nabla_{\alpha}\overline{u}(x_{0}),\overline{d}(x_{0}))+\rho, \end{split} \label{123} \end{equation} where we have used the periodicity of the functions $z$ and $\zeta$, assumption $(H_2)$, and \eqref{114}. The arbitrary choice of $\rho$ yields now \eqref{1151}. By approximating with piecewise constant functions (see Lemma \ref{ctap}) and using \eqref{growth}, the estimate is extended to a general $z$. \paragraph{\textit{\textbf{Step 2} (Upper bound -- surface)}} Following an argument in \cite{AMT}, and taking into account Proposition \ref{p100}, it suffices to prove the upper bound for the case where $\cl u$ is of the form $\cl u = \lambda \chi_U$, where $\chi_U$ denotes the characteristic function of a set of finite perimeter $U\subset\omega$ and $\lambda\in\R3$. Moreover, by standard arguments we can restrict ourselves to the case where $U$ is a polygonal set. Given $x_0 \in S(\cl u)$, writing for simplicity $\nu := \nu (\cl u)(x_0)$, by the definition of $h_{3d,2d}$, for any $\rho>0$ we may find $u\in SBV(Q'_{\nu};\mathbb{R}^{3})$, such that $\nabla_\alpha u=0$ a.e., $u|_{\partial Q_\nu'}=\gamma_{\lambda,\nu}$, and $$\int_{Q'_{\nu}\cap S(u)}h_{3d}([u],\tilde{\nu})\,\de\mathcal{H}^{1}(x_\alpha)\leq h_{3d,2d}([\overline{u}],\nu(\overline{u}))(x_{0})+\rho.$$ We claim that \begin{equation}\label{uppersurf} \frac{\de\cF_{3d,2d}(\overline{u},\overline{d})}{\de\mathcal{H}^{1}\res S(\overline{u})}(x_{0}) \leq h_{3d,2d}([\overline{u}](x_{0}),\nu(\cl u)(x_{0})), \end{equation} for $\mathcal{H}^1-$ a.e. $x_0 \in \omega\cap S(\overline{u}).$ Now put $\lambda:=[\overline{u}](x_{0})$, and since it is not restrictive to assume that $\nu =e_{2}$, define \begin{align*} D_{n}(x_{0},\delta) & := \bigg( Q'(x_{0},\delta) \cap\bigg\{ x:\bigg\vert (x-x_{0})\cdot e_2\bigg\vert <\frac{\delta}{2n}\bigg\}\bigg)\times I,\\ Q^{+}(x_{0},\delta) & := \left (Q'(x_{0},\delta) \cap \left\{ x:(x-x_{0}) \cdot e_2>0\right\}\right) \times I,\\ Q^{-}(x_{0},\delta) & := \left (Q'(x_{0},\delta) \cap \left\{ x: (x-x_{0}) \cdot e_2<0\right\}\right)\times I. \end{align*} Let \[u_{\delta,n}(x_{\alpha},x_{3}):=\left\{\begin{array}[c]{lll} \lambda+\eps_{n}x_{3}\overline{d}, & & \text{in } Q^{+}(x_{0},\delta) \backslash D_{n}(x_{0},\delta),\\ \displaystyle u\left(\frac{n(x_{\alpha}-x_{0})}{\delta}\right) & & \text{in } D_{n}(x_{0},\delta),\\ \eps_{n}x_{3}\overline{d} & & \text{in }Q^{-}(x_{0},\delta) \backslash D_{n}(x_{0},\delta). \end{array} \right. \] Clearly, $u_{\delta,n}\rightarrow \cl u$ in $L^{1}(Q(x_{0},\delta) ;\mathbb{R}^{3})$ (that is, it converges to $\tilde{u}(x_\alpha, x_3) := \cl u(x_\alpha)$), $\tfrac1{\eps_{n}}\int_I\nabla_{3}u_{\delta,n}\,\de x_{3}\rightharpoonup \overline{d}$ in $L^{p}(Q(x_{0},\delta) ;\mathbb{R}^{3})$, both as $n\to\infty$, and $\nu(u_{\delta,n})\cdot e_{3}=0.$ Thus, \begin{align*} \frac{\de\cF_{3d,2d}(\overline{u},\overline{d})}{\de\mathcal{H}^{1}\res S(\cl u)}(x_{0}) \leq\lim_{\delta, n} \frac{1}{\delta}\bigg\{ & \int_{Q'(x_{0},\delta)\times I} W_{3d}\left(\nabla_{\alpha}u_{\delta,n}\left|\frac{\nabla_{3}u_{\delta,n}}{\eps_{n}}\right.\right)\de x\\ & +\int_{(Q'(x_{0},\delta) \times I)\cap S(u_{\delta,n})}h_{3d}\left([u_{\delta,n}],\nu_{\alpha}(u_{\delta, n}) \left|\frac{\nu_{3}(u_{\delta,n})}{\eps_{n}}\right.\right)\de\mathcal{H}^{2}(x)\bigg\} \\ =\lim_{\delta,n}\frac{1}{\delta}\bigg\{ &\int_{(Q'(x_{0},\delta) \times I) \setminus D_{n}(x_{0,}\delta) }W_{3d}(0|\overline{d})\,\de x \\ &+\int_{[(Q'(x_{0},\delta) \times I) \setminus D_{n}(x_{0,}\delta)] \cap (S(\cl d)\times I)}h_{3d}\left(\eps_n x_3[\cl d],\tilde{\nu}(\cl d)\right)\de\mathcal{H}^{2}(x) \\ & +\int_{D_{n}(x_{0,}\delta) }W_{3d}\left(\frac{n}{\delta}\nabla_{\alpha}u\left.\left( \frac{n(x_{\alpha}-x_{0})}{\delta}\right) \right|0\right)\de x \\ & +\int_{D_{n}(x_{0,}\delta) \cap \{ x_0 + \frac{\delta}{n} S(u)\}\times I} h_{3d}\left([u]\left( \frac{n\left( x_{\alpha} -x_{0}\right)}{\delta}\right),\tilde{\nu}(u)\right)\de\mathcal{H}^{2}(x)\bigg\}. \\ \end{align*} Using now the growth conditions on $W_{3d}$ and $h_{3d}$ and changing variables one obtains \begin{equation*} \begin{split} \frac{\de\cF_{3d,2d}(\overline{u},\overline{d})}{\de\mathcal{H}^{1}\res S(\cl u)}(x_{0}) \leq \lim_{\delta,n}\frac{1}{\delta}\bigg\{ & \int_{Q^{\prime}(x_{0},\delta )}C(1+|\cl d|^{p}) \,\de x_\alpha + c_h\, \eps_n |D^s \cl d|(Q'(x_0, \delta))\\ & +\int_{D_{n}(x_{0},\delta)}W_{3d}\Big(\frac{n}{\delta }\nabla_{\alpha }u\Big( \frac{n(x_{\alpha }-x_{0})}{\delta }\Big)\Big|0\Big) \de x \\ & +\int_{D_{n}(x_{0},\delta)\cap \{ x_0 + \frac{\delta}{n} S(u)\}\times I}h_{3d}\Big ([u]\Big(\frac{n(x_{\alpha }-x_{0})}{\delta}\Big) ,\tilde{\nu}(u)\Big) \de\mathcal{H}^{2}(x) \bigg\} \\ \leq \lim_{\delta ,n}\bigg\{ & \frac{\delta}{n^2}\int_{nQ'\times I}W_{3d}\Big(\frac{n}{\delta }\nabla_{\alpha }u(y_{\alpha})\Big|0\Big)\de y \\ & +\frac{1}{n}\int_{(nQ'\times I)\cap (S(u)\times I) \cap \{ y\cdot e_2| \leq \frac{1}{2}\}}h_{3d}([u](y_{\alpha}),\tilde{\nu}(u))\de \mathcal{H}^{2}(y)\bigg\}, \end{split} \end{equation*} since, without loss of generality, the piecewise constant function $\cl d$ can be taken to belong to $L^\infty$ (see the proof of Lemma \ref{ctap}). Moreover, since $\nabla_\alpha u = 0$, we have that: $$\lim_{\delta ,n} \frac{\delta}{n^2}\int_{nQ'\times I}W_{3d}\Big(\frac{n}{\delta }\nabla_{\alpha }u(y_{\alpha })\Big|0\Big) \de y \leq C\delta,$$ and this term also vanishes in the limit $\delta\to0$. We then have that \begin{equation*} \begin{split} \frac{\de\cF_{3d,2d}(\overline{u},\overline{d})}{\de\mathcal{H}^{1}\res S(\cl u)}(x_{0}) \leq & \lim_{\delta, n} \frac{1}{n}\int_{(nQ' \times I)\cap (S(u)\times I) \cap \{ y\cdot e_2| \leq \frac{1}{2}\}}h_{3d}([u](y_{\alpha }),\tilde{\nu }(u))\de \mathcal{H}^{2}(y)\\ \leq &\liminf_{n \to \infty}\frac{1}{n} \int_{(nQ'\times I) \cap (S(u)\times I) \cap \{ y\cdot e_2| \leq \frac{1}{2}\}}h_{3d}([u](y_{\alpha }),\tilde{\nu}(u))\de \mathcal{H}^{2}(y) \\ =&\int_{Q'\cap S(u)} h_{3d} ([u], \tilde\nu(u)) \, \de \mathcal{H}^1 \leq h_{3d, 2d}(\lambda, \nu) + \rho, \end{split} \end{equation*} from which \eqref{uppersurf} follows. \paragraph{\textit{\textbf{Step 3} (Lower bound -- bulk)}} Given a set $B \in \cA(\omega)$, let $u_n\in SBV(\Omega;\R3)$ be an admissible sequence for $\cF_{3d,2d}(\overline{u},\overline{d})(B)$ with $\mu_n$ the corresponding sequence of nonnegative Radon measures given by \begin{equation}\label{miu_n} \mu_{n}(B) :=\int_{B\times I}W_{3d}\Big(\nabla_{\alpha}u_{n}\Big|\frac{\nabla_{3}u_{n}}{\eps_{n}}\Big)\de x+\int_{(B\times I)\cap S(u_{n})} h_{3d} ([u_{n}],\tilde\nu(u_{n}) ) \de\mathcal{H}^{2}(x). \end{equation} Let $x_{0}\in\omega\,$ satisfying \begin{equation}\label{827} \lim_{\delta \to 0} \frac{1}{\delta^{3}}\int_{Q'(x_0, \delta)} | \cl u(x_\alpha) - \cl u (x_0) - \nabla \cl u (x_0) (x_0 - x_\alpha)| \, \de x_\alpha = 0, \end{equation} \begin{equation}\label{830} \lim_{\delta \to 0} \frac{1}{\delta^2} \int_{Q'(x_0, \delta)}|\cl d(x_\alpha) - \cl d(x_0)|^p\, \de x_\alpha = 0. \end{equation} By $(H_1)$ and $(H_2)$ $\mu_{n}$ is bounded and so, up to subsequence (not relabeled), there exists a positive Radon measure $\mu$ such that $\mu_{n}\overset{\ast}{\rightharpoonup}\mu$. In addition, choose $x_0\in\omega$ such that $\frac{\de\mu}{\de\mathcal{L}^{2}}(x_{0})$ exists and is finite. Moreover, there exists a sequence of radii ${\delta_{k}}\to0$ such that $\mu(\partial Q(x_{0},\delta_{k}))=0$ for every $k\in\mathbb{N}.$ It suffices to prove that \begin{equation}\label{778} \frac{\de\mu}{\de\mathcal{L}^{2}}(x_0) \geq W_{3d,2d}(\nabla_{\alpha}\overline{u}(x_0) ,\overline{d}(x_0))\qquad\text{for $\mathcal{L}^2-$ a.e. $x_0 \in \omega.$} \end{equation} We have \begin{equation*} \begin{split} \frac{\de\mu}{\de\mathcal{L}^{2}}(x_{0}) & = \lim_{k,n}\frac{1}{\delta_{k}^{2}}\mu_{n}(Q(x_{0},\delta_{k})) \\ &=\lim_{k,n}\frac{1}{\delta_{k}^{2}} \bigg(\int_{Q'(x_{0},\delta_{k}) \times I}W_{3d}\Big(\nabla_{\alpha}u_{n} \Big|\frac{\nabla_{3}u_{n}}{\eps_{n}}\Big)\de x +\int_{(Q'(x_{0},\delta_{k}) \times I)\cap S(u_{n})}h_{3d}([u_{n}],\tilde\nu(u_{n}))\,\de\mathcal{H}^{2}(x)\bigg). \end{split} \end{equation*} Performing the change of variables $y_{\alpha}=(x_{\alpha}-x_{0})/\delta_{k}$ one obtains \begin{align*} \frac{\de\mu}{\de\mathcal{L}^{2}}(x_{0}) & =\lim_{k,n}\bigg\{ \int_{Q'\times I}W_{3d}\Big(\nabla_{\alpha}u_{n}(x_{0}+\delta_{k}y_{\alpha},y_{3})\Big|\frac{\nabla_{3}u_{n}(x_{0}+\delta_{k}y_{\alpha},y_{3})}{\eps_{n}}\Big) \de y \\ & +\frac{1}{\delta_{k}} \! \int_{(Q'\times I)\cap\{(y_{\alpha},y_{3}):(x_{0}+\delta_{k}y_{\alpha},y_{3}) \in S(u_{n}) \}} \!\!\!\! h_{3d}([u_{n}](x_{0}+\delta_{k}y_{\alpha},y_{3}),\tilde\nu_{\alpha}(u_{n}) (x_{0}+\delta_{k}y_{\alpha},y_{3}))\,\de\mathcal{H}^{1}(y_{\alpha})\de y_{3}\bigg\} . \end{align*} Defining $$u_{k,n}(y):=\frac{u_{n}(x_{0}+\delta_{k}y_{\alpha},y_{3})-\overline{u}(x_{0})}{\delta_{k}},$$ we have \begin{equation*} \nabla_{\alpha}u_{k,n}(y) =\nabla_{\alpha}u_{n}(x_0+\delta_ky_\alpha,y_3),\quad \nabla_{3}u_{k,n}(y)=\frac{1}{\delta_{k}}\nabla_{3}u_{n}(x_{0}+\delta_{k}y_{\alpha},y_{3}),\quad [u_{k,n}] (y) =\frac{1}{\delta_{k}} [u_{n}](x_0+\delta_ky_\alpha,y_3), \end{equation*} and so, recalling $(H_3)$, \begin{equation*} \frac{\de\mu}{\de\mathcal{L}^{2}}(x_{0})=\lim_{k,n}\bigg\{\int_{Q}W_{3d}\Big(\nabla_{\alpha}u_{k,n}\Big|\frac{\delta_{k}\nabla_{3}u_{k,n}}{\eps_{n}}\Big) \de y+\int_{Q\cap S(u_{k,n})} h_{3d}([u_{k,n}],\tilde\nu_{\alpha}(u_{n,k}))\de\mathcal{H}^{2}(y)\bigg\}. \end{equation*} Choose $n(k) \in\mathbb{N}$ such that ${\eps}_{k}':=\delta_k^{-1}\eps_{n(k)}\rightarrow0$; we have that the sequence $v_{k}(\cdot):=u_{k,n(k)}(\cdot)$ converges in $L^{1}$ to $\nabla_{\alpha}\overline{u}(x_0)(\cdot)$ by \eqref{827} and, by \eqref{830}, \begin{equation}\label{777} \int_{I}\frac{\nabla_{3}v_{k}(y)}{\eps_{k}'}\,\de y_{3}\rightharpoonup\overline{d}(x_{0}) \qquad\text{in $L^p(\omega;\R3)$}. \end{equation} Then \begin{equation*} \frac{\de\mu}{\de\mathcal{L}^{2}}(x_{0}) =\lim_{k\to\infty}\bigg\{\int_{Q}W_{3d}\Big(\nabla_{\alpha}v_{k}\Big|\frac{\nabla_{3}v_{k}}{\eps_{k}'}\Big)\de y +\int_{Q\cap S(v_{k})}h_{3d}([v_{k}],\tilde\nu(v_{k}))\,\de \mathcal{H}^{2}(y)\bigg\}. \end{equation*} Next, we change slightly the sequence, in order to comply with the boundary condition in \eqref{107b}. We follow similar arguments to what is done in \cite{CF}. Let $Q_{j}':=\{y_\alpha\in Q^{\prime}:\operatorname*{dist}( y_\alpha,\partial Q^{\prime})>\frac{1}{j}\} $ such that $$\lim_{k\to\infty}\int_{\partial (Q_{j}'\times I)}\left\vert \nabla\overline{u}(x_{0}) y_{\alpha}-v_{k}(y_{\alpha},y_{3})\right\vert\, \de\mathcal{H}^{2}(y)=0$$ and define $$ v_{k,j}(y):=\begin{cases} v_{k}(y) & \text{in }Q_{j}'\times I,\\ \nabla_\alpha\overline{u}(x_{0}) y_{\alpha} & \text{in }(Q^{\prime}\setminus Q_{j}')\times I. \end{cases}$$ Clearly, $v_{k,j}\rightarrow v_{k}$ in $L^{1}(Q;\R3)$ as $j\to\infty$, and therefore, recalling $(H_1)$ and $(H_2)$, $$\frac{\de\mu}{\de\mathcal{L}^{2}}(x_{0}) \geq\lim_{k,j}\bigg\{\int_{Q}W_{3d}\Big(\nabla_{\alpha}v_{k,j}\Big|\frac{\nabla_{3}v_{k,j}}{\eps_{k}'}\Big) \de y+\int_{Q\cap S(v_{k,j})}h_{3d}([v_{k,j}],\tilde\nu_{\alpha}(v_{k,j})) \,\de\mathcal{H}^{2}(y) \bigg\}.$$ Following our argument in Step 1, for fixed $k$ we apply Theorem \ref{Al} to construct a function $g_{k,j}\in SBV(Q; \R3)$ such that $\nabla g_{k,j}=\eps_k'\Big(0\Big|\overline{d}(x_{0})-\int_I \frac{\nabla_{3}v_{k,j}}{\eps_{k}'} \,\de y_3\Big)$ and $\lVert g_{k,j}\rVert_{L^1(Q;\R3)}\leq C \eps_k'\left\lVert\overline{d}(x_{0})-\int_I \frac{\nabla_{3}v_{k,j}}{\eps_k'} \,\de y_3\right\rVert_{L^1(Q';\R3)}$. It is not difficult to verify that the function $w_{k,j}:=v_{k,j}+g_{k,j}$ is a competitor for $W_{3d,2d}(\nabla_{\alpha}\overline{u}(x_{0}) |\overline{d}(x_{0}))$, so that, recalling again $(H_1)$ and $(H_2)$, \begin{equation*} \begin{split} \frac{\de\mu}{\de\mathcal{L}^{2}}(x_{0})& \geq\lim_{k,j}\bigg\{\int_{Q}W_{3d}\Big(\nabla_{\alpha}w_{k,j}\Big|\frac{\nabla_{3}w_{k,j}}{\eps_{k}'}\Big)\de y+\int_{Q\cap S(w_{k,j})}h_{3d}([w_{k,j}],\tilde\nu_{\alpha}(w_{k,j}))\de\mathcal{H}^{2}(y)\Big\} \\ & \geq W_{3d,2d}(\nabla_{\alpha}\overline{u}(x_{0})|\overline{d}(x_{0})), \end{split} \end{equation*} which proves \eqref{778}. \paragraph{\textit{\textbf{Step 4} (Lower bound -- surface)}} Consider the sequence of functions $u_n\in SBV(\Omega;\R3)$ as at the beginning of Step 3, and let $\mu_n$ be the corresponding sequence of Radon measures given by $\eqref{miu_n}$. Recalling that $\mu$ is their weak-* limit, we claim that for $\mathcal{H}^{1}\res S(\cl u)$-a.e. $x_{0}\in S(\cl u)$ \begin{equation}\label{779} \frac{\de{\mu}}{\de(\vert [\cl u] \vert\mathcal{H}^{1}\res S(\cl u))}(x_0)\geq\frac{1}{\vert [\cl u] \vert(x_0)}h_{3d,2d}([\cl u](x_0),\nu(\cl u)(x_0)). \end{equation} Since $\left( \nabla_\alpha u_n\Big| \frac{\nabla_3 u_n}{\eps_n}\right)$ is bounded in $L^p(\Omega;\R{3\times3})$, we have that $\nabla u_n \wto (H|0)$ in $L^p(\Omega;\R{3\times3})$ (up to a subsequence), for some $H \in L^p(\omega; \R{3{\times}2})$. Let $x_0\in\omega\cap S(\cl u)$ be such that $\displaystyle\frac{\de\mu}{\de\mathcal{H}^{1}\res S(\cl u)}(x_0)$ exists, and consider a sequence $\delta_{k}\to0$ such that, denoting $\nu:=\nu(\cl u)(x_0)$, \begin{gather*} \lim_{k\to\infty}\vert[\overline{u}]\vert\mathcal{H}^{1}(S(\overline{u})\cap Q'_{\nu}(x_{0},\delta_{k}))=\vert[\overline u]\vert(x_0),\\ \lim_{k\to\infty}\frac{1}{\delta_{k}}\int_{Q'_{\nu}(x_0,\delta_{k})}\left\vert H(x_{\alpha}) \right\vert \de x_{\alpha}=0. \end{gather*} Then \begin{equation*} \begin{split} &\frac{\de\mu}{\de(\vert [\cl u] \vert\mathcal{H}^{1}\res S(\cl u))}(x_0) =\frac{1}{|[\overline{u}]|(x_0)}\lim_{k,n}\frac{1}{\delta_{k}}\bigg\{\int_{Q_\nu^{\prime}(x_0,\delta_k)\times I}W_{3d}\Big(\nabla_{\alpha}u_{n}\Big|\frac{\nabla_{3}u_{n}}{\eps_{n}}\Big)\de x \\ & \qquad \phantom{\frac{1}{|[\overline{u}]|(x_0)}\lim_{k,n}\frac{1}{\delta_{k}}\quad} +\int_{(Q_\nu^{\prime}(x_{0},\delta_{k})\times I)\cap S(u_n) }h_{3d}([u_{n}],\tilde\nu_{\alpha}(u_{n})) \de\mathcal{H}^{1}(x_\alpha)\de x_3\bigg\} \\ &\quad=\frac{1}{\vert [u]\vert(x_{0})}\lim_{k,n}\bigg\{\delta_{k}\int_{Q_{\nu}^{\prime}\times I} W_{3d}\Big(\nabla_{\alpha}u_{n}(x_{0}+\delta_{k}y_{\alpha},y_{3})\Big|\frac{\nabla_{3}u_{n}(x_{0}+\delta_{k}y_{\alpha},y_{3})}{\eps_{n}}\Big)\de y\\ &\qquad+ \int_{(Q_{\nu}^{\prime}\times I)\cap \{y_{\alpha}:(x_{0}+\delta_{k}y_{\alpha},y_{3}) \in S(u_n)\}}h_{3d}([u_{n}](x_{0}+\delta_{k}y_{\alpha},y_{3}),\tilde\nu(u_{n})(x_{0}+\delta_{k}y_{\alpha},y_{3}))\de \mathcal{H}^{1}(y_\alpha)\de y_3\bigg\} \\ &\quad= \frac{1}{|[u]|(x_0)}\lim_{k,n}\bigg\{\int_{Q_{\nu}^{\prime}\times I}W_{3d}\Big(\frac{\nabla_{\alpha}u_{k,n}}{\delta_k}\Big| \frac{\nabla_{3}u_{k,n}}{\eps_{n}}\Big)\de y \\ &\quad \phantom{\frac{1}{|[u]|(x_0)}\lim_{k,n}\quad} + \int_{(Q_{\nu}^{\prime}\times I)\cap S(u_{k,n})}h_{3d}([u_{k,n}],\tilde\nu(u_{k,n})) \de\mathcal{H}^{1}(y_\alpha)\de y_3\bigg\}, \end{split} \end{equation*} where $u_{n, k}(y) := u_n(x_0 + \delta_ky_\alpha, y_3) - (\cl{u})^-(x_0)$. By a diagonalization argument let $v_{k}:=u_{k,n(k)}$ so that $\lim_{k,n}\left\Vert v_k-\gamma_{[\overline{u}](x_0),\nu}\right\Vert _{L^{1}(Q'_\nu \times I)}=0$, $\nabla v_k \wto 0$ in $L^p(Q'_\nu \times I; \R3)$ and $$\frac{\de\mu}{\de(\vert [\cl u] \vert\mathcal{H}^{1}\res S(\cl u))}(x_0) \geq \frac{1}{|[\overline{u}]|(x_0)}\liminf_{k \to \infty} \int_{(Q_{\nu}^{\prime}\times I)\cap S(v_k)} h_{3d}( [v_k], \tilde\nu_{\alpha}(v_k))\, \de \mathcal{H}^2(y).$$ Following the arguments in \cite[Proposition 4.2]{CF}, we can obtain a new sequence $w_k$ which is a competitor for the cell problem \eqref{107c}, which implies \eqref{779}. This concludes the proof of Theorem \ref{T3d2d}. \qed \subsection{Structured deformations} In oder to pass to structured deformation for the functional in \eqref{F3d2d}, we shall use the relaxation theory developed in \cite{CF} to obtain the representation Theorem \ref{secondleft}. Given $(\cl g,\cl G,\cl d)\in SBV(\omega;\R3){\times}L^1(\omega;\R{3{\times}2})\times L^p(\omega; \R3)$, we define the relaxed energy \begin{equation}\label{200} \begin{split} \cF_{3d,2d,SD}(\cl g,\cl G,\cl d) := \inf \bigg\{ & \liminf_{n\to\infty} \bigg(\int_{\omega } W_{3d,2d}(\nabla u_n, \cl d)\,\de x_\alpha + \int_{\omega\cap S ( u_{n})} h_{3d,2d}([ u_n],\nu(u_n))\de \cH^1(x_\alpha)\bigg): \\ & u_n \in SBV(\omega;\R3),\; u_n\to \cl g \text{ in } L^1(\omega;\R3),\; \nabla u_n\wto \cl G \text{ in } L^p(\omega;\R{3\times2}) \bigg\}. \end{split} \end{equation} \begin{remark}\label{pwc} We notice that the presence of the field $\cl d$ in \eqref{F3d2d} introduces a dependence $x\mapsto W_{3d,2d}(A,\cl d(x))$ of the bulk density on the space variable $x$ not covered in \cite{CF}. One approach to incorporate such a dependence on $x$ is to require that $x\mapsto W_{3d,2d}(A,\cl d(x))$ be continuous. Such a continuity requirement was introduced in \cite{BBBF}. To apply directly the results contained in \cite{BBBF}, we would need to impose a stronger regularity on the field $\cl d$, namely, we would have to require $\cl d\in C(\omega;\R3)$. We avoid this by applying the technique presented in \cite{MS}: we approximate $\cl d$ by a sequence of piecewise constant functions $\cl d_k\in L^p(\omega;\R3)$, and we exploit the property \eqref{wishful} of the bulk energy density $W_{3d,2d}$ and the approximation result provided in \cite[Lemma 2.9]{CF}. \end{remark} Without writing the details of the proof, we assert that these observations, together with Proposition \ref{p100}, allow us to establish the following representation theorem. \begin{theorem}\label{secondleft} Under the hypotheses $(H_1)$--$(H_4)$, for each $(\cl g,\cl G,\cl d)\in SBV(\omega;\R3){\times}L^1(\omega;\R{3{\times}2})\times L^p(\omega; \R3)$, the energy $\cF_{3d,2d,SD}(\cl g,\cl G,\cl d)$ admits an integral representation of the form: \begin{equation}\label{201} \cF_{3d,2d,SD}(\cl g,\cl G,\cl d)=\int_\omega W_{3d,2d,SD}(\nabla\cl g,\cl G,\cl d)\,\de x_\alpha+ \int_{\omega\cap S(\cl g)} h_{3d,2d,SD} ([\cl g],\nu (\cl g))\,\de\cH^1(x_\alpha), \end{equation} where, for $A , B \in \R{3{\times}2}, d \in \R3,$ \begin{equation}\label{2011} \begin{split} W_{3d,2d,SD}(A, B, d) := \inf \bigg\{ &\int_{Q'} W_{3d,2d} (\nabla u(x_\alpha), d)\, \de x_\alpha + \int_{Q'\cap S(u)} h_{3d,2d}([u], \nu(u))\, \de \cH^1(x_\alpha):\\ & u \in SBV(Q'; \R3),\; u|_{\partial Q'} = Ax_\alpha, \; \int_{Q'} \nabla u \,\de x_\alpha = B, \; |\nabla u| \in L^p(Q') \bigg\} \end{split} \end{equation} and, for $\lambda \in \R3$ and $\eta \in \S1,$ \begin{equation}\label{2012} \! h_{3d, 2d, SD} (\lambda,\eta)\!:=\inf \bigg\{ \!\int_{Q'_{\eta}\cap S(u)} \!\! h_{3d, 2d} ([u], \nu(u))\, \de\cH^1(x_\alpha): u \in SBV(Q'_{\eta}; \R3), \nabla u = 0, u|_{\partial Q'_{\eta}}= \gamma_{\lambda, \eta} \bigg\}. \end{equation} \end{theorem} \section{The right-hand path}\label{sect:RHS} In this section we relax our initial energy \eqref{E3d} by first passing to structured deformations and then carrying out the dimension reduction. \subsection{Structured deformations} For $g \in SBV(\Omega_\eps; \R3)$ and $G^{\backslash 3} \in L^1(\Omega_\eps; \R{3{\times}2})$, define \begin{equation}\label{2018} \begin{split} \mathcal{F}_{3d,SD}( g,G^{\backslash 3}) := \inf \bigg\{ & \liminf_{n\to\infty} \bigg(\int_{\Omega_\eps} W_{3d}(\nabla u_n) \,\de x + \int_{\Omega_\eps\cap S(u_n)} h_{3d} ( [u_n], \nu(u_n))\, \de\mathcal{H}^2(x)\bigg): \\ & u_n \to g \text{ in $L^1(\Omega_\eps;\R3)$}, \nabla u_n \wto (G^{\backslash 3} | \nabla_3 g) \,\, \text{in}\, \, L^p(\Omega_\eps;\R{3\times3}) \bigg\}. \end{split} \end{equation} An integral representation for $ \mathcal{F}_{3d, SD}$ follows immediately from \cite[Theorem 2.17]{CF}. As stated in Remark \ref{energies}, the coercivity assumption \eqref{coerc} grants boundedness of the gradients $\nabla u_n$ in $L^p$, thereby justifying the choice of weak convergence of $\{\nabla u_n\}$ in the definition \eqref{2018}. In that definition, we are considering the case in which the limit is classical in the third component of the gradient, that is $\nabla_3 u_n\wto \nabla_3 g$. \begin{theorem}\label{firstSD} Under the hypotheses $(H_1)$--$(H_4)$, for $g \in SBV(\Omega_\eps; \R3)$ and $G^{\backslash 3} \in L^1(\Omega_\eps; \R{3{\times}2})$, the functional $\mathcal{F}_{3d,SD}( g,G^{\backslash 3})$ admits an integral representation of the form: \begin{equation}\label{2017} \mathcal{F}_{3d,SD}(g,G^{\backslash 3}) = \int_{\Omega_\eps} W_{3d,SD} (\nabla g, G^{\backslash 3}) \,\de x + \int_{\Omega_\eps\cap S(g)} h_{3d,SD}([g],\nu(g))\,\de\mathcal{H}^2(x), \end{equation} where, for $A \in \R{3{\times} 3}$ and $B^{\backslash 3} \in \R{3{\times }2}$, \begin{equation}\label{2015} \begin{split} W_{3d,SD} (A, B^{\backslash 3} ) = \inf \bigg\{ & \int_Q W_{3d} (\nabla u)\,\de x + \int_{Q \cap S(u)} h_{3d} ([u],\nu(u))\,\de\mathcal {H}^2(x): \\ & u \in SBV(Q;\R3), \; u| _{\partial Q}= Ax, \; |\nabla u | \in L^p(Q), \; \int_Q \nabla u\,\de x = (B^{\backslash 3} | Ae_3)\bigg\} \end{split} \end{equation} and, for $\lambda \in \R3$, $\nu \in \S2,$ \begin{equation}\label{2016} h_{3d,SD} (\lambda, \nu)=\inf\bigg\{\int_{Q_\nu}h_{3d}([u], \nu(u))\,\de\mathcal{H}^2(x): \; u \in SBV(Q_\nu ; \R3), \nabla u = 0 \text { a.e.}, \; u|_{ \partial Q_\nu} = \gamma_{\lambda, \nu}\bigg\}. \end{equation} \end{theorem} \begin{proposition}\label{energiesCF} Let $W_{3d,SD}$ and $h_{3d,SD}$ be defined by \eqref{2015} and \eqref{2016}, respectively. Then \begin{itemize} \item[(i)] $W_{3d,SD}$ is locally Lipschitz continuous separately in $A$ and $B^{\backslash3}$, namely for every $B^{\backslash3}\in\R{3\times2}$ and every $A_1\in\R{3\times3}$ there exists a constant $C_1>0$ such that $$|W_{3d,SD}(A_1,B^{\backslash 3})-W_{3d,SD}(A_2,B^{\backslash 3})|\leq C_1|A_1-A_2|$$ whenever $|A_1-A_2|$ is small enough; in particular, $$|W_{3d,SD}(A_1,B^{\backslash 3})-W_{3d,SD}(A_2,B^{\backslash 3})|\leq C_1|A_1-A_2|(1+|A_1|^{p-1}+|A_2|^{p-1}).$$ \noindent Similarly, for every $A\in\R{3\times3}$ and $B_1^{\backslash3}\in\R{3\times2}$ there exists a constant $C_2>0$ such that $$|W_{3d,SD}(A,B_1^{\backslash 3})-W_{3d,SD}(A,B_2^{\backslash 3})|\leq C_2|B_1^{\backslash 3}-B_2^{\backslash 3}|$$ whenever $|B_1^{\backslash 3}-B_2^{\backslash 3}|$ is small enough; in particular, $$|W_{3d,SD}(A,B_1^{\backslash 3})-W_{3d,SD}(A,B_2^{\backslash 3})|\leq C_2|B_1^{\backslash 3}-B_2^{\backslash 3}|(1+|B_1^{\backslash 3}|^{p-1}+|B_2^{\backslash 3}|^{p-1});$$ \item[(ii)] $h_{3d,SD}$ satisfies $(H_2)$--$(H_4)$. \end{itemize} \end{proposition} \begin{proof} The proof of part (i) follows that of \cite[Proposition 5.2]{CF}; part (ii) follows from the corresponding properties of $h_{3d}$. \qed \end{proof} \subsection{Dimension reduction} We now apply dimension reduction to the energy $\cF_{3d,SD}$ defined in \eqref{2017}. As we did in Section \ref{LHSDR}, we rescale the variables by $(x_\alpha,x_3)\mapsto(x_\alpha,x_3/\eps)$, thereby replacing the domain of integration $\Omega_\eps$ by $\Omega$, and we rescale the energy $\cF_{3d,SD}$ by dividing it by $\eps$. Therefore, given $(\overline{g}, \overline{G}, \overline{d}) \in SBV (\omega; \R3) \times L^1(\omega; \R{3{\times}2}) \times L^p(\omega; \R3)$, we seek an integral representation for the following relaxed energy \begin{equation}\label{2019} \begin{split} \mathcal{F}_{3d, SD, 2d}( \overline{g}, \overline{G}, \overline{d}) := \inf \bigg\{ & \liminf_{n\to\infty} \bigg(\int_\Omega W_{3d,SD} \Big(\Big(\nabla_\alpha g_n \Big| \frac{\nabla_3 g_n}{\eps_n}\Big), \cl G\Big)\de x \\ &\phantom{\liminf}+ \int_{\Omega\cap S(u_n)} h_{3d, SD} \Big([g_n], \Big(\nu_\alpha (g_n)\Big| \frac{\nu_3(g_n)}{\eps_n}\Big)\Big)\de \mathcal{H}^2(x)\bigg): \\ & g_n\to \overline{g} \text{ in } L^1(\Omega;\R3),\; \int_I \frac{\nabla_3 g_n}{\eps_n} \,\de x_3 \wto \overline{d}\text{ in } L^p(\omega;\R3),\; \nu(g_n)\cdot e_3 = 0 \bigg \}. \end{split} \end{equation} An analogue of Remark \ref{pwc} can be made with the roles of $\cl G$ and $\cl d$ interchanged and with Proposition \ref{energiesCF} in place of Proposition \ref{p100}, and this provides a proof of the following representation theorem. \begin{theorem}\label{secondright} Under the hypotheses $(H_1)$--$(H_4)$, given $(\overline{g}, \overline{G}, \overline{d}) \in SBV (\omega; \R3) \times L^1(\omega; \R{3{\times}2}) \times L^p(\omega; \R3)$, the relaxed energy $\mathcal{F}_{3d, SD, 2d}$ defined in \eqref{2019} admits the integral representation \begin{equation}\label{301} \mathcal{F}_{3d, SD, 2d}(\cl g, \cl G, \cl d) = \int_\omega W_{3d, SD, 2d} (\nabla \cl g, \cl G, \cl d) \; \de x_\alpha + \int_{\omega\cap S(\cl g)} h_{3d, SD, 2d} ([\cl g], \nu(\cl g))\; \de \cH^1(x_\alpha), \end{equation} where, for $A, B \in \R{3{\times}2}, d \in \R3,$ \begin{equation}\label{2013} \begin{split} \!\! W_{3d, SD, 2d}(A, B, d):= \inf \bigg\{ & \int_{Q'} W_{3d, SD} ((\nabla u(x_\alpha)|z(x_\alpha)),B) \,\de x_\alpha + \int_{Q' \cap S(u)} \!\! h_{3d, SD} ([u], \tilde\nu(u))\; \de \cH^1(x_\alpha): \\ & u \in SBV(Q'; \R3),\; |\nabla u| \in L^p(Q'), \; u|_{\partial Q'} = Ax_\alpha, \\ & z \in L^p_{Q' - \mathrm{per}}(\R2; \R3),\; \int_{Q'} z \,\de x_\alpha = d \bigg\}, \end{split} \end{equation} and, for $\lambda \in \R3, \eta \in \S1$, \begin{equation}\label{2014} h_{3d, SD, 2d}(\lambda, \eta)= \inf \bigg\{ \int_{Q'_\eta} h_{3d, SD}( [u], \nu(u))\; \de \cH^1(x_\alpha): \; u \in SBV(Q'; \R3), \; \nabla u = 0\; a.e., \; u|_{\partial Q'} = \gamma_{\lambda, \eta}\bigg\}. \end{equation} \end{theorem} \section {Comparison of the relaxed energy densities for the left- and right-hand paths}\label{sect:comp_ex} In this section we discuss the relationship between the doubly relaxed energy densities \eqref{201} and \eqref{301} obtained in Sections \ref{sect:LHS} and \ref{sect:RHS}. At present, at the level of generality of Theorems \ref{secondleft} and \ref{secondright}, an explicit comparison in terms of whether one of the two energies is smaller than the other is not available. Nonetheless, quantitative results can be obtained when the initial energy \eqref{E3d} has a a specific form, namely it is a purely interfacial energy ($W_{3d}=0$) with a specific choice of the interfacial energy density $h_{3d}$. Our aim then is to compute explicitly the densities provided by the cell formulas \eqref{107b}, \eqref{107c}, \eqref{2011}, \eqref{2012}, \eqref{2015}, \eqref{2016}, \eqref{2013}, and \eqref{2014} starting from the initial, purely interfacial, energy density (see \cite{nogap,OP15}) \begin{equation}\label{purelyinterfacial} h_{3d}(\lambda,\nu)=|\lambda \cdot \nu|. \end{equation} \paragraph{\textbf{The left-hand path}} Let us consider \eqref{purelyinterfacial} and let $(A,d)\in\R{3\times2}\times\R3$; then \eqref{107b} reads \begin{equation}\label{1231} W_{3d,2d}(A,d)= \inf\bigg\{\int_{Q^{\prime}\cap S(u)}|[u]\cdot\tilde\nu(u)|\,\de\mathcal{H}^{1}(x_\alpha): u\in SBV(Q^{\prime};\mathbb{R}^{3}), u|_{\partial Q^{\prime}}(x_{\alpha})=Ax_{\alpha}\bigg\} =0. \end{equation} The first equality is a consequence of \eqref{purelyinterfacial}; the second one follows since the affine function $u(x_\alpha) = Ax_\alpha$ is admissible and makes the integral vanish. Let us now turn to \eqref{107c}: we claim that for $\lambda\in\R3$, $\eta\in\mathbb{S}^1$, the surface energy density $h_{3d,2s}$ reads \begin{equation}\label{107c1} h_{3d,2d}(\lambda,\eta)=|\lambda\cdot\tilde\eta|. \end{equation} In fact, the function $u(x_\alpha)=\gamma_{\lambda,\eta}(x_\alpha)$ (see \eqref{gammale}) is admissible and it provides an upper bound; to obtain a lower bound, one uses the following version of the Gauss-Green formula in $SBV$ (see \cite[Theorem 3.36]{AFP} and also \cite{DPO95,V,VH}): for $u\in SBV(\Omega;\R3)$ and $U\subset\Omega$, there holds \begin{equation}\label{GGformula} \int_{U\cap S(u)} [u]\cdot\nu(u)\,\de\cH^{2}(x)+\int_U \div u\,\de x-\int_{\partial U} u\cdot\nu_U\,\de\cH^2(x)=0. \end{equation} Considering the integrand in \eqref{107c}, by using the properties of the absolute value and \eqref{GGformula}, the same $u(x_\alpha)=\gamma_{\lambda,\eta}(x_\alpha)$ gives \begin{equation}\label{107c2} \int_{Q_\eta'\cap S(u)} |[u]\cdot\tilde\nu(u)|\,\de\cH^1(x_\alpha)\geq \left|\int_{Q_\eta'\cap S(u)} [u]\cdot\tilde\nu(u)\,\de\cH^1(x_\alpha)\right|= \left|\int_{\partial Q_\eta'} u\cdot\tilde\nu_{Q_\eta'}\,\de\cH^1(x_\alpha)\right| = |\lambda\cdot\tilde\eta|, \end{equation} which completes the proof of \eqref{107c1}. Given $(\cl{u},\cl{d}) \in SBV(\omega;\mathbb{R}^{3}) \times L^{p}(\omega ;\mathbb{R}^{3})$, the relaxed energy \eqref{F3d2d} reads then \begin{equation}\label{ex3d2d} \cF_{3d,2d}(\cl{u},\cl{d})=\widehat\cF_{3d,2d}(\cl u):=\int_{\omega \cap S(\cl{u})} |[\cl{u}]\cdot\tilde\nu(\cl{u})|\,\de\mathcal{H}^{1}(x_\alpha), \end{equation} where we notice that the dependence on $\cl d$ is lost. Next, we claim that, for $A,B\in\R{3\times2}$, $d\in\R3$, the bulk density \eqref{2011} is given by $W_{3d,2d,SD}(A, B, d)=\widehat W_{3d,2d,SD}(A,B)$, which is the relaxation of $h_{3d,2d}$ in \eqref{107c1}, and reads \begin{equation}\label{2011a} \begin{split} \widehat W_{3d,2d,SD}(A, B)=& \inf \bigg\{ \int_{Q'\cap S(u)} |[u]\cdot \tilde\nu(u)|\, \de \cH^1(x_\alpha): u \in SBV(Q'; \R3), \; u|_{\partial Q'} = Ax_\alpha, \\ & \phantom{\inf\Bigg\{} \int_{Q'} \nabla u \,\de x_\alpha = B, \; |\nabla u| \in L^p(Q') \bigg\}; \end{split} \end{equation} notice again that this is independent of $d$. We prove that, for $A,B\in\R{3\times2}$, \begin{equation}\label{2011b} \widehat W_{3d,2d,SD}(A, B)=\big|\tr\big((A|0)-(B|0)\big)\big|=|A_{11}+A_{22}-B_{11}-B_{22}|. \end{equation} Again as before, we prove \eqref{2011b} by obtaining upper and lower bounds for $\widehat W_{3d,2d,SD}$. Let $u$ be an admissible function for \eqref{2011a} and define $u_\alpha:Q' \to \R2$ by $u_\alpha(x_\alpha):= (u_1(x_\alpha), u_2(x_\alpha))$. Since \begin{equation}\label{1292} \int_{Q' \cap S(u)} |[u]\cdot\tilde{\nu}(u)|\, \de \cH^1(x_\alpha) = \int_{Q' \cap S(u_\alpha)} |[u_\alpha]\cdot\nu(u_\alpha)|\, \de \cH^1(x_\alpha), \end{equation} the function $u_\alpha$ is admissible for the minimum problem \begin{equation}\label{1293} \inf \bigg\{ \int_{Q'\cap S(v)} \!\! |[v]\cdot \nu(v)|\, \de \cH^1(x_\alpha): v\in SBV(Q'; \R2), \, v|_{\partial Q'} = \widehat{A}x_\alpha,\, \int_{Q'} \nabla v\,\de x_\alpha = \widehat{B},\, |\nabla v| \in L^p(Q') \bigg\}, \end{equation} where $\widehat{A}$ and $\widehat{B}$ denote the upper $2{\times}2$ sub-matrices of $A$ and $B$, respectively. The lower bound for $\widehat W_{3d,2d,SD}$ then follows immediately from the result in \cite{nogap,OP15}, where it is proved that the infimum in \eqref{1293} is given by $|\tr(\widehat A-\widehat B)|$. In order to derive the upper bound for $\widehat W_{3d,2d,SD}$, fix $\eps>0$ and let $v_\epsilon\in SBV(Q';\R2)$ admissible for \eqref{1293} be such that \begin{equation}\label{1311} \int_{Q'\cap S(v_\epsilon)} |[v_\epsilon]\cdot \nu(v_\epsilon)|\, \de \cH^1(x_\alpha) \leq |\tr (\widehat{A} - \widehat{B})| + \epsilon. \end{equation} Using Lemma $4.3$ in \cite{M07}, we can construct a function $v \in SBV(Q')$ such that $$ v|_{\partial Q'} = e_3\cdot Ax_\alpha, \qquad \nabla v = (B_{31}, B_{32})\quad \text{$\cL^2$-a.e. in $Q'$}.$$ Then, the function $ w_\epsilon\in SBV( Q' ; \R3)$ defined by $w_\epsilon(x_\alpha) := (v_\epsilon(x_\alpha), v(x_\alpha))$ is admissible for \eqref{2011a}, and by \eqref{1292} and \eqref{1311} we conclude that \begin{equation}\label{1319} \int_{Q'\cap S(w_\epsilon)} |[w_\epsilon]\cdot \tilde{\nu}(w_\epsilon))|\, \de \cH^1(x_\alpha) \leq |\tr(\widehat{A} - \widehat{B})| + \epsilon, \end{equation} and the result follows from the arbitrariness of $\eps$. Formula \eqref{2011b} is therefore proved. Finally, we observe that the same strategy used to prove \eqref{107c1} can be used to show that for $\lambda \in \R3$ and $\eta \in \S1,$ \begin{equation}\label{2012a} h_{3d, 2d, SD} (\lambda, \eta)= |\lambda\cdot\tilde\eta|. \end{equation} Thus, in view of \eqref{2011b} and \eqref{2012a}, given $(\cl g,\cl G,\cl d)\in SBV(\omega;\R3){\times}L^1(\omega;\R{3{\times}2})\times L^1(\omega; \R3)$, the functional $\cF_{3d,2d,SD}$ in \eqref{201} can be written as \begin{equation}\label{ex3d2dSD} \begin{split} \cF_{3d,2d,SD}(\cl g,\cl G,\cl d)= & \widehat\cF_{3d,2d,SD}(\cl g,\cl G):= \int_\omega |\tr((\nabla\cl g|0)-(\cl G|0))|\,\de x_\alpha+\int_{\omega\cap S(\cl g)} |[\cl g]\cdot\tilde\nu(\cl g)|\,\de\cH^1(x_\alpha) \\ =& \int_\omega \Big|\frac{\partial\cl g_1}{\partial x_1}+\frac{\partial\cl g_2}{\partial x_2}-\cl G_{11}-\cl G_{22}\Big|\,\de x_\alpha+\int_{\omega\cap S(\cl g)} |[\cl g_1]\nu_1(\cl g)+[\cl g_2]\nu_2(\cl g)|\,\de\cH^1(x_\alpha). \end{split} \end{equation} \paragraph{\textbf{The right-hand path}} Considering \eqref{purelyinterfacial}, the explicit formulas for the energy densities $W_{3d, SD}$ and $h_{3d, SD}$ in \eqref{2015} and \eqref{2016} were derived in \cite{nogap,OP15} (see also \cite{S17}); denoting by $M^i$, $i=1,2,3$ the columns of a matrix $M\in\R{3\times3}$, for $A\in\R{3\times3}$ and $B^{\backslash 3} \in \R{3\times 2}$ we have that \begin{equation}\label{1193} W_{3d, SD}( A, B^{\backslash 3}) = |\tr(A-(B^{\backslash 3}|A ^3))|, \end{equation} and, for $\lambda \in \R3$ and $\nu \in \S2$, \begin{equation}\label{1194} h_{3d,SD} (\lambda, \nu) = |\lambda\cdot\nu|. \end{equation} Therefore, for $(g,G^{\backslash 3})\in SBV(\Omega; \R3)\times L^1(\Omega; \R{3{\times}2})$, plugging \eqref{1193} and \eqref{1194} in \eqref{2017} gives \begin{equation}\label{2017a} \mathcal{F}_{3d,SD}( g,G^{\backslash 3}) = \int_\Omega \Big|\frac{\partial g_1}{\partial x_1}+\frac{\partial g_2}{\partial x_2}-G_{11}^{\backslash 3}-G_{22}^{\backslash 3}\Big|\,\de x + \int_{\Omega\cap S(g)} |[g]\cdot\nu(g)|\,\de\mathcal{H}^2(x), \end{equation} Let us now turn to \eqref{2013}. Let $A, B \in \R{3{\times}2}$, $d \in \R3$, and let $(u,z)$ be an admissible pair of functions for the minimization problem that defines $W_{3d,SD,2d}$; using \eqref{1193} and \eqref{1194}, and again the properties of the absolute value and the Gauss-Green formula \eqref{GGformula}, we can estimate \begin{equation}\label{1195} \begin{split} \!\! \int_{Q'} |\tr((\nabla u | z) - (B | z))|\, \de x_\alpha &+ \int_{Q' \cap S(u)} |[u]\cdot \tilde\nu(u)|\,\de \cH^1(x_\alpha)\\ & \geq \bigg|\int_{Q'} \tr((\nabla u | z)-(B | z))\,\de x_\alpha\bigg| +\bigg|\int_{Q' \cap S(u)} [u]\cdot\tilde\nu(u)\,\de \cH^1(x_\alpha)\bigg|\\ & \geq \bigg|\int_{Q'} \tr((\nabla u | z)-(B | z))\,\de x_\alpha+\int_{Q' \cap S(u)} [u]\cdot\tilde\nu(u)\,\de \cH^1(x_\alpha)\bigg|\\ & =\bigg|\tr\bigg(\int_{Q'} \nabla (u_1,u_2)\,\de x_\alpha+ \int_{Q' \cap S(u)} [u]\otimes\tilde\nu(u)\,\de \cH^1(x_\alpha)\bigg)-B_{11}-B_{22}\bigg|\\ & =\bigg|\tr\bigg(\int_{Q'} (u_1, u_2) \otimes \nu_{\partial Q'}\,\de \cH^1(x_\alpha)\bigg)-B_{11}-B_{22}\bigg|\\ & =|A_{11}+A_{22}-B_{11}-B_{22}| \end{split} \end{equation} where the last equality follows from the condition $u|_{\partial Q'}(x_\alpha) = Ax_\alpha$. Since the affine function $u(x_\alpha) = A x_\alpha$ is admissible, the lower bound \eqref{1195} is attained, so that the density in \eqref{2013} reads \begin{equation}\label{2013a} W_{3d, SD, 2d}(A, B, d)=|A_{11}+A_{22}-B_{11}-B_{22}|=|\tr(\widehat A-\widehat B)|=:\widehat W_{3d, SD, 2d}(A, B). \end{equation} Finally, with the same reasoning as before, it is easy to see that the infimum in \eqref{2014} is attained at $u(x_\alpha)=\gamma_{\lambda, \eta}(x_\alpha)$, so that \begin{equation}\label{2014a} h_{3d, SD, 2d}(\lambda,\eta)=|\lambda\cdot\tilde\eta|. \end{equation} Thus, in view of \eqref{2013a} and \eqref{2014a}, given $(\cl g,\cl G,\cl d)\in SBV(\omega;\R3){\times}L^1(\omega;\R{3{\times}2})\times L^1(\omega; \R3)$, the functional $\cF_{3d,SD,2d}$ in \eqref{301} can be written as \begin{equation}\label{ex3dSD2d} \begin{split} \cF_{3d,SD,2d}(\cl g,\cl G,\cl d)= & \widehat\cF_{3d,SD,2d}(\cl g,\cl G):= \int_\omega |\tr(\widehat{\nabla\cl g}-\widehat{\cl G})|\,\de x_\alpha+\int_{\omega\cap S(\cl g)} |[\cl g]\cdot\tilde\nu(\cl g)|\,\de\cH^1(x_\alpha) \\ =& \int_\omega \Big|\frac{\partial\cl g_1}{\partial x_1}+\frac{\partial\cl g_2}{\partial x_2}-\cl G_{11}-\cl G_{22}\Big|\,\de x_\alpha+\int_{\omega\cap S(\cl g)} |[\cl g_1]v_1(\cl g)+[\cl g_2]v_2(\cl g)|\,\de\cH^1(x_\alpha). \end{split} \end{equation} Notice that we have proved that the bulk energy densities in \eqref{2011b} and \eqref{2013a} coincide, and the same holds true for the surface energy densities \eqref{2012a} and \eqref{2014a}. Thus, we have proved the following result. \begin{proposition}\label{S1} Let $W_{3d}=0$ and $h_{3d}$ as in \eqref{purelyinterfacial}. Then, the doubly relaxed energies \eqref{201} and \eqref{301} coincide and are both given by \eqref{ex3d2dSD} or \eqref{ex3dSD2d}. \end{proposition} \section{A one-step approach to dimension reduction in the context of structured deformations}\label{sect:6} In this section, we recall an alternative procedure for dimension reduction in the context of structured deformations already available in the literature \cite{MS}. The basic function spaces considered for this approach are the spaces \cite{CLT1,CLT2} \begin{equation} \begin{split} SBV^2(\Omega;\R3):= & \{u\in SBV(\Omega;\R3): \nabla u\in SBV(\Omega;\R{3\times3})\}, \\ BV^2(\Omega;\R3):= & \{u\in BV(\Omega;\R3): \nabla u\in BV(\Omega;\R{3\times3})\}. \end{split} \end{equation} For a function $v\in SBV^2(\Omega_\eps;\R3)$, the initial energy considered in \cite{MS} is of the form \begin{equation}\label{enMS} E^{MS}_\eps(v):=\int_{\Omega_\eps} W(\nabla v,\nabla^2 v)\,\de x+\int_{\Omega_\eps\cap S(v)} \Psi_1([v],\nu(v))\,\de\cH^2(x)+\int_{\Omega_\eps\cap S(\nabla v)} \Psi_2([\nabla v],\nu(\nabla v))\,\de\cH^2(x), \end{equation} where the bulk energy density $W\colon\R{3\times3}{\times}\R{3\times3\times3}\to[0,+\infty)$ is continuous, coercive, and has growth of order $p=1$, and the surface energy densities $\Psi_1\colon\R3{\times}\S{2}\to[0,+\infty)$ and $\Psi_2\colon\R{3\times3}{\times}\S2\to[0,+\infty)$ are continuous, coercive, have growth of order $1$ and are also subadditive and homogeneous of degree $1$ in the first vadiable; see the assumptions $(H_1)$--$(H_8)$ in \cite{MS} for the precise details. We also refer the reader to \cite[Introduction and Remark 1.5]{MS} for a justification of the presence of the second-order gradient in the bulk density and of the energy density $\Psi_2$. The main result obtained in \cite{MS} is an integral representation result for the relaxed functional \begin{equation}\label{S3} I(g,b,G):=\inf\Big\{\liminf_{n\to\infty} J_{\eps_n}(u_n): u_n\in SBV^2(\Omega;\R3), u_n\stackrel{L^1}\to g, \frac1{\eps_n}\nabla_3 u_n\stackrel{L^1}\to b, \nabla_\alpha u_n\stackrel{L^1}\to G\Big\}, \end{equation} where $(g,b,G)\in BV^2(\omega;\R3){\times}BV(\omega;\R3){\times}BV(\omega;\R{3\times2})$, $\eps_n$ is a sequence tending to zero from above, and the functional $J_{\eps_n}$ is obtained by rescaling $E_{\eps_n}^{MS}$ in \eqref{enMS} by $\eps_n$ in the third variable and then dividing by $\eps_n$, analogously to the definition of $F_\eps$ from $E_\eps$ in \eqref{107}. The field $b$ plays the role of the field $\cl d$ in the previous sections. One important difference between \cite{MS} and the present work is that the vector field $b$ in \eqref{S3} already depends only on $x_\alpha$ because of the coercivity conditions alone (see again \cite[assumptions $(H_1)$--$(H_8)$ and Remark 1.5]{MS}), whereas in the previous sections it was necessary to average in the $x_3$ variable. Moreover, it is evident that the process of relaxation in \eqref{S3} is a simultaneous passage to structured deformations and dimension reduction. \begin{theorem}[{\cite[Theorem 1.4]{MS}}]\label{S4} The functional $I$ defined in \eqref{S3} does not depend on the sequence $\{\eps_n\}$ and admits an integral representation of the form $I=I_1+I_2$, where, for $(g,G)\in BV^2(\omega;\R3){\times}BV(\omega;\R{3\times2})$, \begin{equation}\label{S5} I_1(g,G)=\int_\omega W_1(G-\nabla g)\,\de x_\alpha+\int_\omega W_1\bigg(-\frac{\de D^cg}{\de|D^cg|}\bigg)\,\de|D^cg|(x_\alpha)+\int_{\omega\cap S(g)} \Gamma_1([g],\nu(g))\,\de\cH^1(x_\alpha) \end{equation} and for $(b,G)\in BV(\omega;\R3){\times}BV(\omega;\R{3\times2})$ \begin{equation}\label{S6} \begin{split} I_2(b,G)= & \int_\omega W_2(b,G,\nabla b,\nabla G)\,\de x_\alpha+\int_\omega W_2^\infty\bigg(b,G,\frac{\de D^c(b,G)}{\de|D^c(b,G)|}\bigg)\,\de|D^c(b,G)| \\ & +\int_{\omega\cap S((b,G))} \Gamma_2((b,G)^+,(b,G)^-,\nu((b,G)))\,\de\cH^1(x_\alpha). \end{split} \end{equation} The energy densities of $I_1$ are obtained as follows: for each $A\in\R{3\times2}$, $\lambda\in\R3$, and $\eta\in\S1$, \begin{align} W_1(A)&=\inf\bigg\{\int_{Q'\cap S(u)} \cl\Psi_1([u],\nu(u))\,\de\cH^1(x_\alpha): u\in SBV(Q';\R3), u|_{\partial Q'}=0, \nabla u=A\; a.e.\bigg\}, \label{W1} \\ \Gamma_1(\lambda,\eta)&=\inf\bigg\{\int_{Q_\eta'\cap S(u)} \cl\Psi_1([u],\nu(u))\,\de\cH^1(x_\alpha): u\in SBV(Q_\eta';\R3), u|_{\partial Q_\eta'}=\gamma_{\lambda,\eta}, \nabla u=0\; a.e.\bigg\}, \label{Gamma1} \end{align} with $\gamma_{\lambda,\eta}$ defined as in \eqref{gammale} and \begin{equation}\label{S9} \cl\Psi_1(\lambda,\nu):=\inf\{\Psi_1(\lambda,(\nu|t)):t\in\R{}\}. \end{equation} The energy densities of $I_2$ are obtained as follows: for each $A\in\R{3\times2}$, $B_\beta\in\R{3\times3\times2}$, $\Lambda,\Theta\in\R{3\times3\times2}$, and $\eta\in\S1$, \begin{equation}\label{S10} \begin{split} W_2(A,B_\beta)=\inf\bigg\{ & \int_{Q'}\cl W(A,\nabla u)\,\de x_\alpha+\int_{Q'\cap S(u)} \cl\Psi_2([u],\nu(u))\,\de\cH^1(x_\alpha): \\ & u\in SBV(Q';\R{3\times3}), u_{ik}|_{\partial Q'}=\sum_{j=1}^2 B_{ijk}x_j\bigg\}, \end{split} \end{equation} \begin{equation}\label{S11} \begin{split} \Gamma_2(\Lambda,\Theta,\eta)=\inf\bigg\{ & \int_{Q_\eta'} \cl W^\infty(u,\nabla u)\,\de x_\alpha+\int_{Q_\eta'\cap S(u)} \cl\Psi_2([u],\nu(u))\,\de\cH^1(x_\alpha): \\ & u\in SBV(Q'_\eta;\R{3\times3}), u|_{\partial Q_\eta'}=u_{\Lambda,\Theta,\eta}\bigg\}, \end{split} \end{equation} where \begin{equation}\label{S12} u_{\Lambda,\Theta,\eta}(x_\alpha):= \begin{cases} \Lambda & \text{if $0\leq x_\alpha\cdot\eta<1/2$,} \\ \Theta & \text{if $-1/2<x_\alpha\cdot\eta<0$,} \end{cases} \end{equation} and with $\cl W$ and $\cl\Psi_2$ as follows: decomposing $B\in\R{3\times3\times3}$ into $(B_\beta,B_3)\in\R{3\times3\times2}{\times}\R{3\times3\times1}$ (i.e., $B_\beta$ denotes $B_{ijk}$ with $k=1,2$), define \begin{equation}\label{S13} \cl W(A,B_\beta):=\inf\{W(A,(B_\beta,B_3)): B_3\in\R{3\times3\times1}\}, \end{equation} and for $\Lambda\in\R{3\times3}$ and $\eta\in\S1$, let \begin{equation}\label{S14} \cl\Psi_2(\Lambda,\eta):=\inf\{\Psi_2(\Lambda,(\eta|t)):t\in\R{}\}. \end{equation} \end{theorem} In the statement of Theorem \ref{S4}, a superscript ``$\infty$'' denotes the recession function at infinity (see \cite[hypothesis ($H_3$) on page 461]{MS}), whereas the superscript ``$c$'' denotes the Cantor part. We also point out that we maintained the notation from \cite{MS} for the convenience of the reader; in the notations of our previous sections, the triple $(g,b,G)$ would be written $(\cl g,\cl G,\cl d)$. \smallskip \noindent \textit{Sketch of the proof of Theorem \ref{S4}} By making use of Theorem \ref{Al}, the relaxed functional $I$ defined in \eqref{S3} can be additively decomposed into the functionals $I_1$ and $I_2$ defined in \eqref{S5} and \eqref{S6}, respectively, decoupling the effects of the surface energy density $\Psi_1$ from the bulk energy density $W$ and the surface energy density $\Psi_2$ (see \cite[Section 3.1]{MS}). The result is obtained by proving upper and lower bounds for the Radon-Nikod\'ym derivative of the energies $I_1$ and $I_2$. The technique is analogous to that presented in detail in the proof of Theorem \ref{T3d2d} in Section \ref{sect:LHS}. The lower bounds aim at proving the $\liminf$ inequality \eqref{BGLI}; the upper bounds aim at proving the $\limsup$ inequality \eqref{BGLS}. \qed \smallskip We are not undertaking a comparison of the relaxed energy in Theorem \ref{S4} with those obtained in Theorems \ref{secondleft} and \ref{secondright} at this level of generality, however, we do so for the particular choice made in Section \ref{sect:comp_ex}, namely for an initial energy where the only non-zero contribution comes from the jumps of the $SBV$ function, and not of its gradient, i.e., $W=\Psi_2=0$ and $\Psi_1(\lambda,\nu)=|\lambda\cdot\nu|$, see \eqref{purelyinterfacial}. For this particular choice, we provide explicit formulas for the energy densities \eqref{W1}, \eqref{Gamma1}, \eqref{S10}, and \eqref{S11}, and we show that the relaxed energy $I$ is identically zero. As it can be seen from the definitions of the energy densities $W_1$, $\Gamma_1$, $W_2$, and $\Gamma_2$, the functionals $I_1$ and $I_2$ are of the first order, meaning that only first-order derivatives enter in their definitions (the function spaces in \eqref{W1}, \eqref{Gamma1}, \eqref{S10}, and \eqref{S11} are of $SBV$ type). Since $W=0$ and $\Psi_2=0$, the relaxed densities $W_2$ and $\Gamma_2$ in \eqref{S10} and \eqref{S11} are trivially equal to zero, so that the term $I_2$ in \eqref{S6} vanishes. Moreover, the relaxation procedure for obtaining $I_1$ can be carried out in the $SBV$ setting, as in the previous Sections \ref{sect:LHS} and \ref{sect:RHS}. By invoking the results of \cite{nogap,S17}, one sees that the use of the strong convergence in $L^1$ in \eqref{S3} for an initial energy featuring $\Psi_1$ only is the same as using the weak convergence in $L^p$ considered in the previous Sections \ref{sect:LHS} and \ref{sect:RHS}, namely, there is no difference in considering either $\nabla u=A$ a.e. or $\int_{Q'} \nabla u=A$ in formula \eqref{W1}. To compute the energy densities \eqref{W1} and \eqref{Gamma1} with the choice $\Psi_1(\lambda,\nu)=|\lambda\cdot\nu|$,we recall the definition of $\cl\Psi_1$ in \eqref{S9} and notice that it reads \begin{equation}\label{1335} \cl\Psi_1(\lambda,\eta)=\inf\{|\lambda\cdot(\eta|t)|:t\in\R{}\}= \begin{cases} |\lambda\cdot\tilde\eta| & \text{if $\lambda_3=0$}, \\ 0 & \text{if $\lambda_3\neq0$}. \end{cases} \end{equation} To show that $W_1=\Gamma_1=0$, we use the fact that $\overline{\Psi}_1(\lambda, \eta)$ vanishes whenever $\lambda_3 \neq 0$, so that jumps of infimizing approximations $u_n$ with non-zero third components have no energetic cost. We control the energetics cost of any necessary jumps with zero third-components by relegating them to transverse segments within the frames \begin{equation}\label{frame2} \cF_{n}:=Q'\setminus\Big(\frac{n-1}{n}\Big)Q',\qquad \cF_{n, \eta}:=Q'_\eta\setminus \Big(\frac{n-1}{n}\Big)Q'_\eta, \end{equation}% with $n$ a positive integer and $\eta \in \S1$. This approach was employed in \cite{nogap}, and we refer the reader to that article for any details omited here. To show that $W_1(M) = 0$ for all $M \in \R{3\times2}$, we choose a constant $C > 0$ and, for each $n$ a function $v_n \in SBV(\cF_{n};\R3)$ such that \begin{equation}\label{1705} v_n|_{\partial \cF_{n}} = 0, \quad \nabla v_n = M\quad \text{a.e. in}\; \cF_{n}, \quad \text{and}\quad | D^s v_n| \leq \frac{C}{n}. \end{equation} Next, we partition the shrunken square $(\tfrac{n-1}{n})Q'$ into $n$ thin rectangles $\mathcal{C}_{k,n}$, $k = 0, \ldots , n-1$, each of height $\frac{n-1}{n}$ and width $\frac{n-1}{n^2}$ (the width corresponding to the direction $e_1 = (1,0)$). Denoting the center of each rectangle by $c_{k,n}$, we define for each $n$ a function $u_n \in SBV(\omega;\R2)$ by \begin{equation}\label{1694} u_n(x) = \begin{cases} v_n(x) & \text{if $x \in \mathcal{F}_{n}$,}\\ M(x - c_{k,n}) + \frac{(-1)^k}{n^2}e_3 & \text{if $x \in c_{k,n}$, $k=1,\ldots,n -1$.}\\ \end{cases} \end{equation} It follows that \begin{equation}\label{1700} S(u_n) \subset S(v_n) \cup \partial (\tfrac{n-1}{n})Q' \cup \bigcup_{k=0}^{n-2} ( \partial \mathcal{C}_{k,n} \cap \partial \mathcal{C}_{k+1,n}) \end{equation} and we first consider $[u_n](x)$ when $x \in \partial (\tfrac{n-1}{n})Q'$. Using \eqref{1705} we have (to within a fixed choice of signs in front of each term) \begin{equation}\label{1704} [u_n](x)\cdot e_3 = \pm M(x - c_{k,n})\cdot e_3 \pm \frac{1}{n^2}= \pm (x - c_{k,n})\cdot M^\top e_3 \pm \frac{1}{n^2}, \end{equation} so that $[u_n](x)\cdot e_3 = 0$ if and only if $M^\top e_3 \neq 0$ and $x$ is on the line $\ell = \{ y \in \R2: ( y - c_{k,n})\cdot M^\top e_3 \pm \frac{1}{n^2} = 0\}$ in $\R2$ whose distance from $c_{k,n}$ is $(n^2|M^\top e_3|)^{-1} = O(n^{-2}).$ Because the distance from $c_{k,n}$ to $\partial (\tfrac{n-1}{n})Q'$ is at least $\frac{n-1}{2n^2} = O(n^{-1})$, it follows that for $n$ sufficiently large the line $\ell$ intersects $\partial (\tfrac{n-1}{n})Q'$ at exactly two points. We conclude from \eqref{1335} that, whether or not $M^\top e_3 \neq 0$, for $n$ sufficiently large \begin{equation}\label{1711} \overline{\Psi}_1( [u_n](x), \nu(u_n)(x)) = 0\quad \text{for $\cH^1$-a.e.\@ $x\in \partial (\tfrac{n-1}{n})Q'$.} \end{equation} We consider next a point $x \in \bigcup_{k=0}^{n-2} ( \partial \mathcal{C}_{k,n} \cap \partial \mathcal{C}_{k+1,n})$ and use \eqref{1694} to compute \begin{equation}\label{1715} [u_n](x)\cdot e_3 = \pm M(c_{k,n} - c_{k+1,n})\cdot e_3 \pm \frac{1}{n^2} = \pm(c_{k,n} - c_{k+1,n})\cdot M^\top e_3 \pm \frac{1}{n^2}, \end{equation} which is zero only if $M^\top e_3 \neq 0$. However, $|c_{k,n} - c_{k+1,n}| = \frac{n-1}{n^2} = O(n^{-1})$ so that for $n$ sufficiently large $[u_n](x)\cdot e_3 \neq 0$ for every $x \in \bigcup_{k=0}^{n-2} ( \partial \mathcal{C}_{k,n} \cap \partial \mathcal{C}_{k+1,n})$, and we conclude \begin{equation}\label{1722} \overline{\Psi}_1( [u_n](x), \nu(u_n)(x)) = 0\quad \text{for $\cH^1$-a.e.\@ $x\in \bigcup_{k=0}^{n-2} (\partial \mathcal{C}_{k,n} \cap \partial \mathcal{C}_{k+1,n})$} \end{equation} and that, by \eqref{1705}, \eqref{1694} and \eqref{1700}, \begin{equation}\label{1726} \int_{Q'\cap S(u_n)}\Psi_1([u_n], \nu(u_n))\, \de \cH^1(x_\alpha) = \int_{Q'\cap S(v_n)}\Psi_1([v_n], \nu(v_n))\, \de \cH^1(x_\alpha) = |D^s v_n| = O\left(\frac{1}{n}\right). \end{equation} Because $u_n$ is admissible in \eqref{W1} we conclude that $W_1(M) = 0$. To show that $\Gamma_1(\lambda, \eta) = 0$ for all $\lambda \in \R3$ and $\eta \in \S1$, we note first that the mapping $\gamma_{\lambda, \eta}: Q'_\eta \to \R3$ is admissible in \eqref{Gamma1}, so that \begin{equation}\label{1734} 0 \leq \Gamma_1(\lambda, \eta) \leq \int_{Q'_\eta\cap S(\gamma_{\lambda, \eta})}\Psi_1([\gamma_{\lambda, \eta}], \nu(\gamma_{\lambda, \eta}))\, \de \cH^1(x_\alpha) = \Psi_1(\pm \lambda, \eta). \end{equation} In particular, if $\lambda_3 \neq 0$, then \eqref{1734} and \eqref{1335} yield $\Gamma_1(\lambda, \eta) = 0.$ Suppose now that $\lambda_3 = 0$. With $\cF_{n, \eta}$ defined as in \eqref{frame2}, we define $u_n: Q'_\eta \to \R3$ by \begin{equation}\label{1742} u_n(x) = \begin{cases} \gamma_{\lambda, \eta} & \text{if $x \in \mathcal{F}_{n}$,}\\ \gamma_{\lambda, \eta} - \frac{1}{n}e_3 & \text{if $x \in (\tfrac{n-1}{n})Q'_\eta$ and $x\cdot \eta \leq 0$,} \\ \gamma_{\lambda, \eta} + \frac{1}{n}e_3 & \text{if $x \in (\tfrac{n-1}{n})Q'_\eta$ and $x\cdot \eta \geq 0$.} \\ \end{cases} \end{equation} It follows that $S(u_n) \subset \partial (\tfrac{n-1}{n})Q'_\eta \cup \{ x \in Q'_\eta : x\cdot \eta= 0\}$. If $ x \in Q'_\eta \cap S(u_n),$ then $[u_n](x) = [\gamma_{\lambda,\eta}] + \frac{m(x)}{n}e_3$ with $m(x) \in \{ 0, 1, -1, 2, -2\}$ and \begin{equation}\label{1743} m(x) = 0\quad \text{if and only if}\quad x\cdot\eta = 0\quad \text{and}\quad |x| \in \bigg[ \frac{n-1}{2n}, \frac{1}{2}\bigg]. \end{equation} Because $[\gamma_{\lambda, \eta}](x) \in \{\lambda, -\lambda, 0\}$ and $\lambda\cdot e_3 = \pm\lambda_3 = 0,$ it follows that $[u_n](x) \cdot e_3 = 0$ if and only if $m(x) = 0$, i.e., \begin{equation}\label{1744} [u_n](x)\cdot e_3 = 0 \quad \text{if and only if}\quad x\cdot \eta = 0\quad \text{and}\quad |x| \in \bigg[\frac{n-1}{2n}, \frac{1}{2}\bigg]. \end{equation} We conclude from \eqref{1335} that: $\cl\Psi_1([u_n](x), \nu(u_n)(x))\neq 0$ if and only if $x\cdot \eta = 0$ and $|x| \in \left[ \frac{n-1}{2n}, \frac{1}{2}\right]$, so that \begin{equation}\label{1745} \begin{split} 0 \leq \Gamma_1(\lambda, \eta) \leq & \int_{Q'_\eta\cap S(u_n)} \Psi_1([u_n], \nu(u_n))\, \de \cH^1(x_\alpha) \\ = & \int_{Q'_\eta\cap\left\{x\cdot \eta = 0 \text{ and } |x| \in \left[ \frac{n-1}{2n}, \frac{1}{2}\right]\right\}} \Psi_1([\gamma_{\lambda, \eta}], \nu(\gamma_{\lambda,\eta})\, \de \cH^1(x_\alpha) \leq \frac{|\lambda |}{n}. \end{split} \end{equation} Because each $u_n$ is admissible in \eqref{Gamma1}, $\{u_n\}$ is an infimizing sequence and $\Gamma_1(\lambda, \eta) = 0.$ \section{Conclusions}\label{conclusions} In this paper we have studied a problem that involves both dimension reduction and introduction of disarrangements. From the point of view of energetics, this entails two relaxation processes, so that the order in which they are performed is relevant for the structure of the final, doubly relaxed energy functional. In this respect, we applied the two relaxation processes one after the other in both orders and we obtained two doubly relaxed energy functionals, those in \eqref{201} and in \eqref{301}. At the level of generality considered in Theorems \ref{secondleft} and \ref{secondright}, we did not undertake a comparison of these two formulas. Nonetheless, we compared them in a special case which is relevant to the multiscale nature of the geometry of structured deformations, namely we considered an initial energy which takes into account only the normal component of the jumps. In this case, we were able to prove that the doubly relaxed energy functionals are the same, see Proposition \ref{S1}. Moreover, we compared our procedure with one that has been studied by Matias and Santos in \cite{MS}: here, the dimension reduction and the relaxation to structured deformations are performed simultaneously. With the same choice of a purely interfacial initial energy, we computed the relaxed energy in the context of \cite{MS} and we proved that it is identically equal to zero. This suggests looking at different scalings in the vanishing thickness parameter $\eps$, in particular, looking for higher-order terms in the expansion by $\Gamma$-convergence in the sense of \cite{AB}. It is worth noticing that, in spite of the technical differences in the three relaxation procedures carried out, the final relaxed energies are all defined on the same type of mathematical objects, namely a structured deformation and a director, defined on the cross--section $\omega$. To see this, one can compare the triple $(\cl g,\cl G,\cl d)\in SBV(\omega;\R3){\times}L^1(\omega;\R{3{\times}2}){\times}L^p(\omega; \R3)$ in Theorems \ref{secondleft} and \ref{secondright} with the triple $(g,b,G)\in BV^2(\omega;\R3){\times}BV(\omega;\R3){\times}BV(\omega;\R{3\times2})$ in Theorem \ref{S4}. It is natural to conjecture that the relaxation described in Theorem \ref{S4} yields a lower energy than those provided by Theorems \ref{secondleft} and \ref{secondright}. In this regard, the results contained in \cite{S17} provide a useful tool for studying this conjecture. In view of the results of Sections \ref{sect:comp_ex} and \ref{sect:6}, we can answer affirmatively to the conjecture in the case of a particular choice of the initial energy. Finally, we remark that a common feature of all three approaches is the introduction of constraints on the admissible disarrangements, namely that the normal to the jump set be aligned with the two-dimensional approximating object. This is enforced by the condition $\nu(u_n)\cdot e_3=0$ in \eqref{107a}, by the condition $\int_Q \nabla u\,\de x=(B^{\backslash 3}|Ae_3)$ in \eqref{2015}, and by the conditions cited in \cite[Remark 1.5]{MS}. \medskip \noindent\textbf{Acknowledgements.} The authors warmly thank the Departamento de Matem\'atica at Instituto Superior T\'ecnico in Lisbon, the Departamento de Matem\'atica at Universidade de \'Evora, the Center for Nonlinear Analysis at Carnegie Mellon University in Pittsburgh, the Fakult\"at f\"ur Mathematik at Technische Universit\"at M\"unchen, SISSA in Trieste, and the Dipartimento di Ingegneria Industriale of the Universit\`a di Salerno, where this research was developed. The research of J.M.\@ was partially supported by the Funda\c{c}\~{a}o para a Ci\^{e}ncia e a Tecnologia through grant UID/MAT/04459/2013, by the Center for Nonlinear Analysis at Carnegie Mellon University in Pittsburgh, and by the Gruppo Nazionale per l'Analisi Matematica, la Probabilit\`a e le loro Applicazioni (GNAMPA) of the Istituto Nazionale di Alta Matematica (INdAM). The research of M.M.\@ was partially funded by the ERC Advanced grant \emph{Quasistatic and Dynamic Evolution Problems in Plasticity and Fracture} (Grant agreement no.: 290888) and by the ERC Starting grant \emph{High-Dimensional Sparse Optimal Control} (Grant agreement no.: 306274). M.M.\@ is a member of the Gruppo Nazionale per l'Analisi Matematica, la Probabilit\`a e le loro Applicazioni (GNAMPA) of the Istituto Nazionale di Alta Matematica (INdAM).

https://arxiv.org/abs/0709.2442

Complete intersection dimensions and Foxby classes

Let $R$ be a local ring and $M$ a finitely generated $R$-module. The complete intersection dimension of $M$--defined by Avramov, Gasharov and Peeva, and denoted $\cidim_R(M)$--is a homological invariant whose finiteness implies that $M$ is similar to a module over a complete intersection. It is related to the classical projective dimension and to Auslander and Bridger's Gorenstein dimension by the inequalities $\gdim_R(N)\leq\cidim_R(N)\leq\pd_R(N)$.Using Blanco and Majadas' version of complete intersection dimension for local ring homomorphisms, we prove the following generalization of a theorem of Avramov and Foxby: Given local ring homomorphisms $\phi\colon R\to S$ and $\psi\colon S\to T$ such that $\phi$ has finite Gorenstein dimension, if $\psi$ has finite complete intersection dimension, then the composition $\psi\circ\phi$ has finite Gorenstein dimension. This follows from our result stating that, if $M$ has finite complete intersection dimension, then $M$ is $C$-reflexive and is in the Auslander class $\catac(R)$ for each semidualizing $R$-complex $C$.

\section*{Introduction} \label{sec0} Let $R$ be a local ring and $N$ a finitely generated $R$-module. The projective dimension of $N$, denoted $\pd_R(N)$, is by now a classical invariant, and much research has shown that modules of finite projective dimension have properties similar to those of modules over a regular local ring. Motivated by this, Auslander and Bridger~\cite{auslander:smt} introduced the Gorenstein dimension of $N$, denoted $\gdim_R(N)$, which is an invariant whose finiteness detects properties similar to those for modules over a Gorenstein ring. More recently Avramov, Gasharov and Peeva~\cite{avramov:cid} defined the complete intersection dimension of $N$, denoted $\cidim_R(N)$, which plays a similar role with respect to the complete intersection property. Corresponding to the well-known hierarchy of rings, there are inequalities $$\gdim_R(N)\leq\cidim_R(N)\leq\pd_R(N)$$ with equality to the left of any finite quantity. See Sections~\ref{sec1} and~\ref{sec2} for foundations of Gorenstein dimensions and complete intersection dimensions. Avramov and Foxby~\cite{avramov:rhafgd} used Auslander and Bridger's Gorenstein dimension to define what it means for a local ring homomorphism $\varphi\colon R\to S$ to have finite Gorenstein dimension. Note that one cannot simply define the Gorenstein dimension of $\varphi$ to be $\gdim_R(S)$, as $S$ is not assumed to be finitely generated as an $R$-module. Avramov and Foxby overcome this technical difficulty by using the Cohen factorizations of Avramov, Foxby and Herzog~\cite{avramov:solh} to replace $\varphi$ with a related homomorphism $\varphi'\colon R'\to\comp S$ which has the added benefit of being surjective so that $\gdim_{R'}(\comp S)$ is defined. Blanco and Majadas~\cite{blanco:miccp} use the same idea to study local homomorphisms of finite complete intersection. Avramov and Foxby~\cite{avramov:rhafgd} asked the following: given two local ring homomorphisms $\varphi\colon R\to S$ and $\psi\colon S\to T$ of finite Gorenstein dimension, must the composition $\psi \circ\varphi$ also have finite Gorenstein dimension? They were able to answer this question, for example, when $\psi$ has finite flat dimension. We move one step closer to answering this question in general with the following result; see Theorem~\ref{compose01}. \begin{intthm} \label{thmB} Let $\varphi\colon R\to S$ and $\psi\colon S\to T$ be local ring homomorphisms. If $\varphi$ has finite Gorenstein dimension and $\psi$ has finite complete intersection dimension, then the composition $\psi \circ\varphi$ has finite Gorenstein dimension. \end{intthm} We also establish the following complete-intersection analogues of results of Avramov, Iyengar and Miller~\cite{avramov:holh} and Foxby and Frankild~\cite{foxby:cmfgidgr}. The first is proved in~\ref{ciid01}. The second one is contained in Corollary~\ref{ciid03} and can also be thought of as an injective version of a result of Blanco and Majadas~\cite{blanco:miccp}. \begin{intthm} \label{thmD} Let $\varphi\colon R\to S$ be a local ring homomorphism. Then $\varphi$ has finite complete intersection injective dimension if and only if $R$ is Gorenstein and $\varphi$ has finite complete intersection dimension. \end{intthm} \begin{intthm} \label{thmE} A local ring of prime characteristic is a complete intersection if and only if some (equivalently, every) power of its Frobenius endomorphism has finite complete intersection injective dimension. \end{intthm} Theorem~\ref{thmB} is proved using the Auslander classes introduced by Avramov and Foxby~\cite{avramov:rhafgd} and generalized by Christensen~\cite{christensen:scatac}; see~\ref{sdc05}. Much recent research has been devoted to the study of these classes, not only because of their connection to the composition question of Avramov and Foxby, but also because the objects in these classes enjoy particularly nice homological properties. Each Auslander class of $R$-complexes contains every bounded $R$-module of finite flat dimension. The next result greatly enlarges the class of objects known to be in each Auslander class; it is contained in Theorem~\ref{cifdac01}. \begin{intthm} \label{thmA} If $R$ is a local ring and $M$ an $R$-module of finite complete intersection flat dimension, then $M$ is in the Auslander class $\cat{A}_C(R)$ for each semidualizing $R$-complex $C$. \end{intthm} Here, the complete intersection flat dimension of $M$ is a version of complete intersection dimension for modules that are not necessarily finitely generated. We actually prove a more general result for $R$-complexes and also a dual result in terms of upper complete intersection injective dimension $\uciid_R(M)$ and the Bass classes $\cat{B}_C(R)$. In particular, when $R$ admits a dualizing complex $D^R$, these complete intersection dimensions determine natural subcategories $\operatorname{CI}\text{-}\catf(R)\subseteq \cat{A}_{D^R}(R)$ and $\operatorname{CI}^*\!\!\text{-}\cati(R)\subseteq\cat{B}_{D^R}(R)$. The next result shows that ``Foxby equivalence'' between the categories $\cat{A}_{D^R}(R)$ and $\cat{B}_{D^R}(R)$ restricts to an equivalence of categories $\operatorname{CI}\text{-}\catf(R)\sim \operatorname{CI}^*\!\!\text{-}\cati(R)$; see~\ref{foxby01} for the proof. \begin{intthm} \label{thmC} Let $R$ be a local ring admitting a dualizing complex $D^R$, and let $M$ be a homologically bounded $R$-complex. \begin{enumerate}[\quad\rm(a)] \item \label{thmCitem1} $\cifd_R(M)<\infty$ if and only if $\uciid_R(D^R\otimes^{\mathbf{L}}_RM)<\infty$. \item \label{thmCitem2} $\uciid_R(M)<\infty$ if and only if $\cifd_R(\rhom_R(D^R,M))<\infty$. \end{enumerate} \end{intthm} A special case of this theorem augments a result of Levin and Vasconcelos and is in Corollary~\ref{foxby04}: When $R$ is Gorenstein, one has $\cifd_R(M)<\infty$ if and only if $\uciid_R(M)<\infty$. Other results of this type are proved in Section~\ref{sec6}. By definition, if $M$ is an $R$-complex of finite complete intersection flat dimension, then there exists a ``quasi-deformation'' $R\xrightarrow{\varphi}R'\xleftarrow{\tau}Q$ such that $\operatorname{fd}_Q(R'\otimes^{\mathbf{L}}_R M)$ is finite; see~\ref{cidim01} and~\ref{cidim03}. It is commonly known that one can exert a small amount of control on the structure of this quasi-deformation. For instance, one may assume without loss of generality that the closed fibre of $\varphi$ is artinian (hence, Cohen-Macaulay) and that $Q$ is complete. One piece of technology that allows for more flexibility in our analysis is the following result, proved in~\ref{qd02}. \begin{intthm} \label{thmF} Let $(R,\ideal{m})$ be a local ring and let $M$ be a homologically bounded $R$-complex. Then $\cifd_R(M)<\infty$ if and only if there exists a quasi-deformation $R\to R''\leftarrow Q'$ such that $Q'$ is complete, the closed fibre $R''/\ideal{m} R''$ is artinian and Gorenstein, and $\operatorname{fd}_{Q'}(R''\otimes^{\mathbf{L}}_RM)$ is finite. \end{intthm} Many of the results in this paper can be stated strictly in terms of modules without losing their flavor. However, some of our proofs \emph{require} the use of complexes. For this reason, we work almost entirely in the derived-category setting. Section~\ref{sec1} contains a summary of the basic notions we use. \section{Complexes and Ring Homomorphisms} \label{sec1} Let $(R,\ideal{m},k)$ and $(S,\ideal{n},l)$ be commutative local noetherian rings. \begin{notationdefinition} \label{basics02} We work in the derived category $\mathcal{D}(R)$ of complexes of $R$-modules, indexed homologically. References on the subject include~\cite{gelfand:moha, hartshorne:rad}. A complex $M$ is \emph{homologically bounded} if $\operatorname{H}_i(M)=0$ for all $|i|\gg 0$; and it is \emph{homologically finite} if $\oplus_i\operatorname{H}_i(M)$ is finitely generated. Let $\mathcal{D}_{\mathrm{b}}(R)$ denote the full subcategory of $\mathcal{D}(R)$ consisting of the homologically bounded $R$-complexes. Isomorphisms in $\mathcal{D}(R)$ are identified by the symbol $\simeq$, and isomorphisms up to shift are designated by $\sim$. Fix $R$-complexes $M$ and $N$. Let $\inf(M)$ and $\sup(M)$ denote the infimum and supremum, respectively, of the set $\{n\in\mathbf{Z}\mid\operatorname{H}_n(M)\neq 0\}$, and set $\operatorname{amp}(M)=\sup(M)-\inf(M)$. Let $M\otimes^{\mathbf{L}}_R N$ and $\rhom_R(M,N)$ denote the left-derived tensor product and right-derived homomorphism complexes, respectively. For each integer $i$, the $i$th \emph{suspension} (or \emph{shift}) of $M$, denoted $\mathsf{\Sigma}^i M$, is the complex with $(\mathsf{\Sigma}^i M)_n=M_{n-i}$ and $\partial_n^{\mathsf{\Sigma}^i M}=(-1)^i\partial_{n-i}^M$. When $M$ is homologically bounded, let $\pd_R(M)$, $\operatorname{fd}_R(M)$ and $\id_R(M)$ denote the projective, flat and injective dimensions of $M$, respectively, as in~\cite{avramov:hdouc}. Let $\cat{P}(R)$, $\cat{F}(R)$ and $\cat{I}(R)$ denote the full subcategories of $\mathcal{D}_{\mathrm{b}}(R)$ consisting of the complexes with, respectively, finite projective, flat and injective dimension. \end{notationdefinition} We shall have several occasions to use the following isomorphisms from~\cite[(4.4)]{avramov:hdouc}. \begin{notationdefinition} \label{basics03} Let $R\to S$ be a ring homomorphism, and fix an $R$-complex $L$ and $S$-complexes $M$ and $N$. Assume that each $R$-module $\operatorname{H}_i(L)$ is finitely generated and $\inf(L)>-\infty$. The natural \emph{tensor-evaluation morphism} $$\omega_{LMN}\colon\rhom_R(L,M)\otimes^{\mathbf{L}}_SN\to\rhom_R(L,M\otimes^{\mathbf{L}}_SN)$$ is an isomorphism when $\sup(M)<\infty$ and either $L\in\cat{P}(R)$ or $N\in\cat{F}(S)$. The natural \emph{Hom-evaluation morphism} $$\theta_{LMN}\colon L\otimes^{\mathbf{L}}_R\rhom_S(M,N)\to\rhom_{S}(\rhom_R(L,M),N)$$ is an isomorphism when $M\in\mathcal{D}_{\mathrm{b}}(S)$ and either $L\in\cat{P}(R)$ or $N\in\cat{I}(S)$. \end{notationdefinition} \begin{disc} \label{fpd01} Let $M$ be a homologically finite $R$-complex with $\pd_R(M)<\infty$ and let $N$ be a homologically bounded $R$-complex. Because $M$ is homologically finite and $\pd_R(M)<\infty$, we know from~\cite[(2.13)]{christensen:scatac} that the $R$-complex $\rhom_R(M,R)$ is homologically finite and has finite projective dimension. Hence, tensor-evaluation~\eqref{basics03} yields the first isomorphism in the next sequence, and the second isomorphism is tensor-cancellation: $$\rhom_R(M,R)\otimes^{\mathbf{L}}_RN \simeq\rhom_R(M,R\otimes^{\mathbf{L}}_RN) \simeq\rhom_R(M,N). $$ \end{disc} \begin{notationdefinition} \label{basics01} Let $\varphi\colon R\to S$ be a local ring homomorphism. We denote by $\comp R$ the completion of $R$ at its maximal ideal and let $\varepsilon^{}_R\colon R\to\comp{R}$ denote the natural map. The \emph{completion} of $\varphi$ is the unique local ring homomorphism $\comp{\varphi}\colon\comp{R}\to\comp{S}$ such that $\comp\varphi\circ\varepsilon_R=\varepsilon_S\circ\varphi$. The \emph{semi-completion} of $\varphi$ is the composition $\grave{\varphi}=\varepsilon^{}_S\circ\varphi\colon R\to \comp{S}$, and the flat dimension of $\varphi$ is $\operatorname{fd}(\varphi)=\operatorname{fd}_R(S)$. By~\cite[(1.1)]{avramov:solh} the map $\grave{\varphi}$ admits a \emph{Cohen factorization}, that is, there is a diagram of local ring homomorphisms, $R\xrightarrow{\Dot{\varphi}}R'\xrightarrow{\varphi'}\comp S$, where $\grave\varphi=\varphi'\circ\Dot{\varphi}$, with $\Dot{\varphi}$ flat, the closed fibre $R'/\ideal{m} R'$ regular, $R'$ complete, and $\varphi'$ surjective. \end{notationdefinition} \begin{notationdefinition} \label{sdc01} A homologically finite $R$-complex $C$ is \emph{semidualizing} if the homothety morphism $\chi^R_C\colon R\to\rhom_R(C,C)$ is an isomorphism in $\mathcal{D}(R)$. A complex $D$ is \emph{dualizing} if it is semidualizing and $\id_R(D)$ is finite. \end{notationdefinition} \begin{disc} \label{sdc02} Let $\varphi\colon R\to S$ be a local ring homomorphism of finite flat dimension and let $M$ be a homologically finite $R$-complex. From~\cite[(5.7)]{christensen:scatac} and~\cite[(4.5)]{frankild:rrhffd} we know that $S\otimes^{\mathbf{L}}_R M$ is semidualizing for $S$ if and only if $M$ is semidualizing for $R$, and $S\otimes^{\mathbf{L}}_R M$ is dualizing for $S$ if and only if $M$ is dualizing for $R$ and $\varphi$ is Gorenstein by~\cite[(5.1)]{avramov:lgh}. For example, the map $\varphi$ is Gorenstein if it is flat with Gorenstein closed fibre~\cite[(4.2)]{avramov:lgh} or surjective with $\operatorname{Ker}(\varphi)$ generated by an $R$-regular sequence~\cite[(4.3)]{avramov:lgh}. Consult~\cite{avramov:lgh} for more information on Gorenstein homomorphisms. \end{disc} \begin{disc} \label{dc01} If $R$ is a homomorphic image of a Gorenstein ring, e.g., if $R$ is complete, then $R$ admits a dualizing complex by~\cite[(V.10.4)]{hartshorne:rad}. \end{disc} \begin{notationdefinition} \label{sdc03} Let $C$ be a semidualizing $R$-complex. A homologically finite $R$-complex $M$ is \emph{$C$-reflexive} if the complex $\rhom_R(M,C)$ is homologically bounded and the biduality morphism $\delta^C_M\colon M\to\rhom_R(\rhom_R(M,C),C)$ is an isomorphism in $\mathcal{D}(R)$. Set $$\gcdim_R(M):=\begin{cases} \inf(C)-\inf(\rhom_R(M,C)) & \text{if $M$ is $C$-reflexive} \\ \infty & \text{otherwise.} \end{cases}$$ When $C=R$ we write $\gdim_R(M)$ in lieu of $\gkdim{R}_R(M)$; this is the \emph{G-dimension} of Auslander and Bridger~\cite{auslander:smt} and Yassemi~\cite{yassemi:gd}. \end{notationdefinition} \begin{disc} \label{sdc04} Assume that $R$ admits a dualizing complex $D$. Each homologically finite $R$-complex $M$ is $D$-reflexive by~\cite[(V.2.1)]{hartshorne:rad}, and~\cite[(2.12)]{christensen:scatac} tells us that $M$ is semidualizing for $R$ if and only if $\rhom_R(M,D)$ is so. \end{disc} \begin{defn} \label{sdm01} Let $C$ be a semidualizing $R$-module. An $R$-module $G$ is \emph{$\text{G}_C$-projective} if there exists an exact sequence of $R$-modules $$X=\cdots\xrightarrow{\partial^X_2}P_1 \xrightarrow{\partial^X_1} P_0 \xrightarrow{\partial^X_0} C\otimes_R P_{-1} \xrightarrow{\partial^X_{-1}} C\otimes_R P_{-2} \xrightarrow{\partial^X_{-2}}\cdots$$ such that $G\cong\operatorname{Coker}(\partial^X_1)$, each $P_i$ is projective, and $\Hom_R(X,C\otimes_R Q)$ is exact for each projective $R$-module $Q$. An $R$-module $G$ is \emph{$\text{G}_C$-flat} if there exists an exact sequence of $R$-modules $$Y=\cdots\xrightarrow{\partial^X_2}F_1 \xrightarrow{\partial^Y_1} F_0 \xrightarrow{\partial^Y_0} C\otimes_R F_{-1} \xrightarrow{\partial^Y_{-1}} C\otimes_R F_{-2} \xrightarrow{\partial^Y_{-2}}\cdots$$ such that $G\cong\operatorname{Coker}(\partial^Y_1)$, each $F_i$ is flat, and $\Hom_R(C,I)\otimes_R Y$ is exact for each injective $R$-module $I$. An $R$-module $G$ is \emph{$\text{G}_C$-injective} if there exists an exact sequence of $R$-modules $$Z=\cdots\xrightarrow{\partial^Z_2}\Hom_R(C,I_1) \xrightarrow{\partial^Z_1} \Hom_R(C,I_0) \xrightarrow{\partial^Z_0} I_{-1} \xrightarrow{\partial^Z_{-1}} I_{-2} \xrightarrow{\partial^Z_{-2}}\cdots$$ such that $G\cong\operatorname{Coker}(\partial^Z_1)$, each $I_i$ is injective, and $\Hom_R(\Hom_R(C,I),Z)$ is exact for each injective $R$-module $I$. Let $M$ be a homologically bounded $R$-complex. A \emph{$\text{G}_C$-projective resolution} of $M$ is an isomorphism $H\simeq M$ in $\mathcal{D}(R)$ where $H$ is a complex of $\text{G}_C$-projective $R$-modules such that $H_i=0$ for all $i\ll 0$. The \emph{$\text{G}_C$-projective dimension} of $M$ is $$\text{G}_C\text{-}\!\pd_R(M):= \inf\{\sup\{n\mid H_n\neq 0\}\mid \text{$H\simeq M$ is a $\text{G}_C$-projective resolution}\}. $$ The \emph{$\text{G}_C$-flat dimension} of $M$ is defined similarly and denoted $\text{G}_C\text{-}\!\fd_R(M)$, while the \emph{$\text{G}_C$-injective dimension} $\text{G}_C\text{-}\!\id_R(M)$ is dual. When $C=R$ we write $\text{G-}\!\pd_R(M)$, $\text{G-}\!\fd_R(M)$ and $\text{G-}\!\id_R(M)$; these are the \emph{G-projective, G-flat, and G-injective dimensions} of Enochs, Jenda and Torrecillas~\cite{enochs:gipm,enochs:gf}. \end{defn} \begin{disc} \label{sdm02} Let $C$ be a semidualizing $R$-module, and let $R\ltimes C$ denote the trivial extension of $R$ by $C$. Let $M$ be a homologically bounded $R$-complex, and view $M$ as an $R\ltimes C$-complex via the natural surjection $R\ltimes C\to R$. From~\cite[(2.16)]{holm:smarghd} there are equalities \begin{gather*} \text{G}_C\text{-}\!\pd_R(M)=\text{G-}\!\pd_{R\ltimes C}(M) \qquad\qquad \text{G}_C\text{-}\!\fd_R(M)=\text{G-}\!\fd_{R\ltimes C}(M) \\ \text{G}_C\text{-}\!\id_R(M)=\text{G-}\!\id_{R\ltimes C}(M). \end{gather*} It is known in a number of cases that the quantities $\text{G}_C\text{-}\!\pd_R(M)$ and $\text{G}_C\text{-}\!\fd_R(M)$ are simultaneously finite. When $C=R$ and $R$ admits a dualizing complex, this is in~\cite[(4.3)]{christensen:ogpifd}. When $M$ is a module, it is in~\cite[(3.5)]{esmkhani:ghdac} and~\cite[(3.3)]{sather:crct}. We deal with the general case in Proposition~\ref{sdm03}. \end{disc} The next categories come from~\cite{avramov:rhafgd,christensen:scatac} and are commonly known as Foxby classes. \begin{notationdefinition} \label{sdc05} Let $C$ be a semidualizing $R$-complex. The \emph{Auslander class} with respect to $C$ is the full subcategory $\cat{A}_C(R)\subseteq\mathcal{D}_{\mathrm{b}}(R)$ consisting of the complexes $M$ such that $C\otimes^{\mathbf{L}}_R M\in\mathcal{D}_{\mathrm{b}}(R)$ and the natural morphism $\gamma^C_M\colon M\to\rhom_R(C,C\otimes^{\mathbf{L}}_R M)$ is an isomorphism in $\mathcal{D}(R)$. The \emph{Bass class} with respect to $C$ is the full subcategory $\cat{B}_C(R)\subseteq\mathcal{D}_{\mathrm{b}}(R)$ consisting of the complexes $N$ such that $\rhom_R(C, N)\in\mathcal{D}_{\mathrm{b}}(R)$ and the natural morphism $\xi^C_N\colon C\otimes^{\mathbf{L}}_R \rhom_R(C,M)\to M$ is an isomorphism in $\mathcal{D}(R)$. \end{notationdefinition} \begin{disc} \label{lwc01} Let $C$ be a semidualizing $R$-complex, and let $X$ be a homologically bounded $R$-complex. If $\operatorname{fd}_R(X)<\infty$, then $X\in\cat{A}_C(R)$; if $\id_R(X)<\infty$, then $X\in\cat{B}_C(R)$; see~\cite[(4.4)]{christensen:scatac}. Let $\varphi\colon R\to S$ be a local homomorphism such that $S\in\cat{A}_C(R)$, e.g., such that $\operatorname{fd}(\varphi)<\infty$. From~\cite[(5.3)]{christensen:scatac} we learn that $S\otimes^{\mathbf{L}}_RC$ is a semidualizing $S$-complex. Furthermore, for any $S$-complex $Y$, we have $Y\in\cat{A}_C(R)$ if and only if $Y\in\cat{A}_{S\otimes^{\mathbf{L}}_RC}(S)$, and $Y\in\cat{B}_C(R)$ if and only if $Y\in\cat{B}_{S\otimes^{\mathbf{L}}_RC}(S)$. When $\varphi$ has finite flat dimension, one has $X\in \cat{A}_C(R)$ if and only if $S\otimes^{\mathbf{L}}_R X\in\cat{A}_{S\otimes^{\mathbf{L}}_RC}(S)$, and $X\in \cat{B}_C(R)$ if and only if $S\otimes^{\mathbf{L}}_R X\in\cat{B}_{S\otimes^{\mathbf{L}}_RC}(S)$ by~\cite[(5.8)]{christensen:scatac}. \end{disc} \begin{disc} \label{sdm02'} Let $C$ be a semidualizing $R$-module and assume that $R$ admits a dualizing complex $D$. For each homologically bounded $R$-complex $M$, we have \begin{enumerate}[\quad\rm(a)] \item \label{sdm02'item1} $\text{G}_C\text{-}\!\fd_R(M)<\infty$ if and only if $M\in\cat{A}_{\rhom_R(C,D)}(R)$, and \item \label{sdm02'item2} $\text{G}_C\text{-}\!\id_R(M)<\infty$ if and only if $M\in\cat{B}_{\rhom_R(C,D)}(R)$. \end{enumerate} This is from~\cite[(4.6)]{holm:smarghd}. We improve upon this in Proposition~\ref{sdm03} below. \end{disc} \begin{disc} \label{sdm02''} Let $C$ be a semidualizing $R$-complex and assume that $R$ admits a dualizing complex $D$. For each homologically finite $R$-complex $M$, we have $\gcdim_R(M)<\infty$ if and only if $M\in\cat{A}_{\rhom_R(C,D)}(R)$ by~\cite[(4.7)]{christensen:scatac}. \end{disc} \section{Gorenstein and Complete Intersection Dimensions} \label{sec2} In this section, we introduce natural variations of existing homological dimensions, beginning with the $\text{G}_C$-version of the main player of~\cite{iyengar:golh}. \begin{defn} \label{gcdim01} Let $\varphi\colon R\to S$ be a local ring homomorphism and $M$ a homologically finite $S$-complex. Fix a semidualizing $R$-complex $C$ and a Cohen factorization $R\xrightarrow{\Dot{\varphi}} R'\xrightarrow{\varphi'} \comp{S}$ of $\grave{\varphi}$. The \emph{$\text{G}_C$-dimension of $M$ over $\varphi$} is the quantity $$\gcdim_{\varphi}(M):= \gkdim{R'\otimes^{\mathbf{L}}_R C}_{R'}(\comp{S}\otimes^{\mathbf{L}}_S M)-\operatorname{edim}(\Dot{\varphi}).$$ The \emph{$\text{G}_C$--dimension of $\varphi$} is $\gcdim(\varphi):=\gcdim_{\varphi}(S)$. In the case $C=R$, we follow~\cite{iyengar:golh} and write $\gdim_{\varphi}(M):=\gkdim{R}_{\varphi}(M)$ and $\gdim(\varphi):=\gkdim{R}(\varphi)$. \end{defn} \begin{properties} \label{sri0} Fix a local ring homomorphism $\varphi\colon R\to S$, a Cohen factorization $R\to R'\to \comp{S}$ of $\grave{\varphi}$, a homologically finite $S$-complex $M$, and a semidualizing $R$-complex $C$. \begin{subprops} \label{sri02} If $R\to R''\to \comp{S}$ is another Cohen factorization of $\grave{\varphi}$, then the quantities $\gkdim{R'\otimes^{\mathbf{L}}_R C}_{R'}(\comp{S}\otimes^{\mathbf{L}}_S M)$ and $\gkdim{R''\otimes^{\mathbf{L}}_R C}_{R''}(\comp{S}\otimes^{\mathbf{L}}_S M)$ are simultaneously finite. This is proved as in~\cite[(3.2)]{iyengar:golh} using~\cite[(6.5)]{christensen:scatac} and~\cite[(4.4)]{frankild:sdcms}. It shows that the finiteness of $\gcdim_{\varphi}(M)$ does not depend on the choice of Cohen factorization. \end{subprops} \begin{subprops} \label{sri03} Arguing as in~\cite[(3.4.1)]{iyengar:golh}, one concludes that the quantities $\gcdim_{\varphi}(X)$, $\gcdim_{\grave{\varphi}}(\comp{S}\otimes^{\mathbf{L}}_S X)$, and $\gkdim{\comp{R}\otimes^{\mathbf{L}}_RC}_{\comp{\varphi}}(\comp{S}\otimes^{\mathbf{L}}_S X)$ are simultaneously finite. \end{subprops} \begin{subprops} \label{sri07} Let $D^{\comp{R}}$ and $D^{R'}$ be dualizing complexes for $\comp{R}$ and $R'$, respectively. The following conditions are equivalent. \begin{enumerate}[{\quad\rm(i)}] \item $\gcdim_{\varphi}(M)<\infty$. \item $\gkdim{R'\otimes^{\mathbf{L}}_R C}_{R'}(\comp{S}\otimes^{\mathbf{L}}_S M)<\infty$. \item $\comp{S}\otimes^{\mathbf{L}}_S M$ is in $\cat{A}_{\rhom_{R'}(R'\otimes^{\mathbf{L}}_R C,D^{R'})}(R')$. \item $\comp{S}\otimes^{\mathbf{L}}_S M$ is in $\cat{A}_{\rhom_{\comp{R}}(\comp{R}\otimes^{\mathbf{L}}_R C,D^{\comp{R}})}(\comp{R})$. \end{enumerate} When $R$ possesses a dualizing complex $D^R$, these conditions are equivalent to: \begin{enumerate}[{\quad\rm(v)}] \item $M$ is in $\cat{A}_{\rhom_R(C,D^R)}(R)$. \end{enumerate} This is proved as in~\cite[(3.6)]{iyengar:golh} using Remarks~\ref{lwc01} and~\ref{sdm02''}. \end{subprops} \begin{subprops} \label{sri09} If $R$ admits a dualizing complex $D$, then~\cite[(5.1)]{christensen:scatac} and~\eqref{sri07} show that $\gcdim(\varphi)$ is finite if and only if $S\otimes^{\mathbf{L}}_R\rhom_R(C,D)$ is a semidualizing $S$-complex. \end{subprops} \begin{subprops} \label{sri10} If $D$ is dualizing for $R$, then $\gkdim{D}_{\varphi}(M)$ is finite. Indeed, because $\dot\varphi$ is flat with Gorenstein closed fibre, the complex $R'\otimes^{\mathbf{L}}_R D$ is dualizing for $R'$ by Remark~\ref{sdc02}, so the desired conclusion follows from the definition using Remark~\ref{sdc04}. \end{subprops} \end{properties} We continue with complete intersection dimensions. When $M$ is a module, Definition~\ref{cidim03} is from~\cite{sahandi:hfd}, which is in turn modeled on~\cite{avramov:cid}. \begin{defn} \label{cidim01} Let $R$ be a local ring. A \emph{quasi-deformation} of $R$ is a diagram of local ring homomorphisms $R\xrightarrow{\varphi} R'\xleftarrow{\tau} Q$ such that $\varphi$ is flat, and $\tau$ is surjective with kernel generated by a $Q$-regular sequence; if the kernel of $\tau$ is generated by a $Q$-regular sequence of length $c$, we will sometimes say that the quasi-deformation has \emph{codimension $c$}. \end{defn} \begin{defn} \label{cidim03} Let $(R,\ideal{m})$ be a local ring. For each homologically bounded $R$-complex $M$, define the \emph{complete intersection projective dimension}, \emph{complete intersection flat dimension} and \emph{complete intersection injective dimension} of $M$ as, respectively, \begin{align*} \cipd_R(M) &:=\inf\left\{\pd_Q(R'\otimes^{\mathbf{L}}_R M)-\pd_Q(R')\left| \text{ \begin{tabular}{@{}c@{}} $R\to R'\leftarrow Q$ is a \\ quasi-deformation \end{tabular} }\!\!\!\right. \right\} \\ \cifd_R(M) &:=\inf\left\{\operatorname{fd}_Q(R'\otimes^{\mathbf{L}}_R M)-\pd_Q(R')\left| \text{ \begin{tabular}{@{}c@{}} $R\to R'\leftarrow Q$ is a \\ quasi-deformation \end{tabular} }\!\!\!\right. \right\} \\ \ciid_R(M) &:=\inf\left\{\id_Q(R'\otimes^{\mathbf{L}}_R M)-\pd_Q(R') \left| \text{ \begin{tabular}{@{}c@{}} $R\to R'\leftarrow Q$ is a \\ quasi-deformation \end{tabular} }\!\!\!\right. \right\}. \end{align*} When $M$ is homologically finite we follow~\cite{avramov:cid,sather:cidc} and define the \emph{complete intersection dimension} of $M$ as $\cidim_R(M):=\cipd_R(M)$. \end{defn} \begin{disc} \label{cidim04} Let $R$ be a local ring and $M$ a homologically bounded $R$-complex. Given a quasi-deformation $R\to R'\leftarrow Q$, a result of Gruson and Raynaud~\cite[Seconde Partie, Thm.~(3.2.6)]{raynaud:cpptpm}, and Jensen~\cite[Prop.~6]{jensen:vl} tells us that the quantities $\pd_Q(R'\otimes^{\mathbf{L}}_RM)$ and $\operatorname{fd}_Q(R'\otimes^{\mathbf{L}}_RM)$ are simultaneously finite. From this, it follows that $\cifd_R(M)<\infty$ if and only if $\cipd_R(M)<\infty$. Arguing as in~\cite[(1.13.2)]{avramov:cid} and using~\cite[(4.2)]{avramov:hdouc}, one deduces that the quantities $\cifd_R(M)<\infty$ if and only if $\cifd_{\comp{R}}(\comp{R}\otimes^{\mathbf{L}}_R M)<\infty$. On the other hand, we do not know if $\ciid_R(M)<\infty$ implies $\ciid_{\comp{R}}(\comp{R}\otimes^{\mathbf{L}}_R M)<\infty$. See, however, Corollary~\ref{ciid02}\eqref{ciid02item2}. \end{disc} We make the next definition for our version of Foxby equivalence in Theorem~\ref{thmC}. \begin{defn} \label{ci*id01} Let $(R,\ideal{m})$ be a local ring and $M$ a homologically bounded $R$-complex. Define the \emph{upper complete intersection injective dimension} of $M$ as $$ \uciid_R(M) :=\inf\left\{ \text{\begin{tabular}{@{}r@{}} $\id_Q(R'\otimes^{\mathbf{L}}_R M)$ \\ $-\pd_Q(R')$ \end{tabular}}\left| \text{ \begin{tabular}{@{}c@{}} $R\to R'\leftarrow Q$ is a quasi-deformation \\ such that $R'$ has Gorenstein formal \\ fibres and $R'/\ideal{m} R'$ is Gorenstein \end{tabular} }\!\!\!\right. \right\}. $$ \end{defn} \begin{disc} \label{uciid01} The formal fibres of $R'$ are Gorenstein when $R'$ is complete or, more generally, when $R'$ is excellent or admits a dualizing complex. \end{disc} \begin{disc} \label{ci*id02} Let $R$ be a local ring and $M$ a homologically bounded $R$-complex. If $\uciid_R(M)<\infty$, then the definitions imply $\ciid_R(M)<\infty$. \end{disc} \begin{question} \label{ci*id03} Let $R$ be a local ring and $M$ a homologically bounded $R$-complex. If $\ciid_R(M)<\infty$, must we have $\uciid_R(M)<\infty$? \end{question} \begin{disc} \label{cidim12} Assume that $R$ is a complete intersection and $M$ is a homologically bounded $R$-complex. Arguing as in~\cite[(1.3)]{avramov:cid}, we conclude $\cipd_R(M)<\infty$, and similarly for $\cifd_R(M)$, $\ciid_R(M)$ and $\uciid_R(M)$. \end{disc} The following formulas are complete intersection versions of the classical Bass formula for injective dimension. Note that the analogue of the Auslander-Buchsbaum formula for complete intersection dimension was proved in~\cite[(3.3)]{sather:cidc}. \begin{prop} \label{bass01} Let $R$ be a local ring and $M$ a homologically finite $R$-complex. \begin{enumerate}[\quad\rm(a)] \item \label{bass01b} If $\ciid_R(M)<\infty$, then $\ciid_R(M)=\depth(R)-\inf(M)$. \item \label{bass01c} If $\uciid_R(M)<\infty$, then $\uciid_R(M)=\depth(R)-\inf(M)$. \end{enumerate} \end{prop} \begin{proof} \eqref{bass01b} We begin by observing the equality $$ \ciid_R(M) =\inf\left\{ \text{\begin{tabular}{@{}r@{}} $\id_Q(R'\otimes^{\mathbf{L}}_R M)$ \\ $-\pd_Q(R')$ \end{tabular}}\left| \text{ \begin{tabular}{@{}c@{}} $R\to R'\leftarrow Q$ is a quasi-deformation \\ such that $R'/\ideal{m} R'$ is artinian \end{tabular} }\!\!\!\right. \right\} $$ where $\ideal{m}$ is the maximal ideal of $R$. Indeed, the inequality ``$\leq$'' follows from the fact that the collection of all quasi-deformations contains the collection of all quasi-deformations $R\to R'\leftarrow Q$ such that $R'/\ideal{m} R'$ is artinian. For the opposite inequality, fix a codimension $c$ quasi-deformation $R\to R'\leftarrow Q$. Choose a prime ideal $P\in\operatorname{Min}_{R'}(R'/\ideal{m} R')$ and set $\ideal{p}=\tau^{-1}(P)$. From~\cite[(5.1.I)]{avramov:hdouc} we conclude $$\id_{Q_{\ideal{p}}}(R'_P\otimes^{\mathbf{L}}_R M)=\id_{Q_{\ideal{p}}}((R'\otimes^{\mathbf{L}}_R M)_{\ideal{p}}) \leq \id_Q(R'\otimes^{\mathbf{L}}_R M).$$ One checks readily that the localized diagram $R\xrightarrow{\varphi_P} R'_P\xleftarrow{\tau_P} Q_{\ideal{p}}$ is a quasi-deformation and that the closed fibre $R'_P/\ideal{m} R'_P\cong(R'/\ideal{m} R')_P$ is artinian. With the equalities $\pd_Q(R')=c=\pd_{Q_{\ideal{p}}}(R'_P)$, this establishes the other inequality. Assume $\ciid_R(M)<\infty$ and fix a quasi-deformation $R\xrightarrow{\varphi} R'\leftarrow Q$ such that $R'/\ideal{m} R'$ is artinian and $\id_Q(R'\otimes^{\mathbf{L}}_R M)<\infty$. The $Q$-complex $R'\otimes^{\mathbf{L}}_R M$ is homologically finite, so the Bass formula and Auslander-Buchsbaum formulas provide the first equality in the following sequence: \begin{align*} \id_Q(R'\otimes^{\mathbf{L}}_R M)-\pd_Q(R') &=[\depth(Q)-\inf(R'\otimes^{\mathbf{L}}_R M)]-[\depth(Q)-\depth(R')] \\ &=-\inf(M)+\depth(R') \\ &=\depth(R)-\inf(M). \end{align*} The second and third equalities follow from the fact that $\varphi$ is flat and local with artinian closed fibre. Since this quantity is independent of the choice of quasi-deformation, only depending on the finiteness of $\id_Q(R'\otimes^{\mathbf{L}}_R M)$, the advertised formula follows. Part~\eqref{bass01c} is established similarly. \end{proof} We close this section with relative versions of complete intersection dimensions. \begin{defn} \label{cidim07} Let $\varphi\colon R\to S$ be a local ring homomorphism and $M$ a homologically finite $S$-complex. The \emph{complete intersection dimension of $M$ over $\varphi$} and \emph{complete intersection injective dimension of $M$ over $\varphi$} are, respectively, \begin{align*} \cidim_{\varphi}(M)&:= \inf\left\{\cidim_{R'}(\comp{S}\otimes^{\mathbf{L}}_S M)-\operatorname{edim}(\Dot{\varphi}) \left| \text{\begin{tabular}{c} $R\xrightarrow{\dot\varphi}R'\xrightarrow{\varphi'}\comp{S}$ is a Cohen \\ factorization of $\grave\varphi$ \end{tabular}}\right.\!\!\!\right\}\\ \ciid_{\varphi}(M)&:= \inf\left\{\ciid_{R'}(\comp{S}\otimes^{\mathbf{L}}_S M)-\operatorname{edim}(\Dot{\varphi}) \left| \text{\begin{tabular}{c} $R\xrightarrow{\dot\varphi}R'\xrightarrow{\varphi'}\comp{S}$ is a Cohen \\ factorization of $\grave\varphi$ \end{tabular}}\right.\!\!\!\right\}. \end{align*} The \emph{complete intersection dimension of $\varphi$} and \emph{complete intersection injective dimension of $\varphi$} are $\cidim(\varphi):=\cidim_{\varphi}(S)$ and $\ciid(\varphi):=\ciid_{\varphi}(S)$. \end{defn} \begin{disc} \label{cidim11} We do not introduce $\cipd_{\varphi}(M)$ and $\cifd_{\varphi}(M)$ because they would be the same as $\cidim_{\varphi}(M)$ by Remark~\ref{cidim04}. On the other hand, we do not introduce $\uciid_{\varphi}(M)$ because we do not need it for our results. \end{disc} \begin{properties} \label{cidim0} Fix a local ring homomorphism $\varphi\colon R\to S$, a Cohen factorization $R\to R'\to \comp{S}$ of $\grave{\varphi}$, and a homologically finite $S$-complex $M$. \begin{subprops} \label{cidim08} If $R$ is a complete intersection, then $\cidim_{\varphi}(M)$ and $\ciid_{\varphi}(M)$ are finite. This follows from Remark~\ref{cidim12} because the fact that $\dot\varphi$ is flat with regular closed fibre implies that $R'$ is a complete intersection. \end{subprops} \begin{subprops} \label{cidim09} As in~\eqref{sri03} one checks that the quantities $\cidim_{\varphi}(M)$, $\cidim_{\grave\varphi}(\comp{S}\otimes^{\mathbf{L}}_SM)$ and $\cidim_{\comp\varphi}(\comp{S}\otimes^{\mathbf{L}}_SM)$ are simultaneously finite, as are the quantities $\ciid_{\varphi}(M)$, $\ciid_{\grave\varphi}(\comp{S}\otimes^{\mathbf{L}}_SM)$ and $\ciid_{\comp\varphi}(\comp{S}\otimes^{\mathbf{L}}_SM)$. \end{subprops} \end{properties} \begin{questions} \label{cidim10} Let $\varphi\colon R\to S$ be a local ring homomorphism and $M$ a homologically finite $S$-complex. Is the finiteness of $\cidim_{\varphi}(M)$ and/or $\ciid_{\varphi}(M)$ independent of the choice of Cohen factorization? Are $\cidim_{\varphi}(M)$ and $\cifd_R(M)$ simultaneously finite? Are $\ciid_{\varphi}(M)$ and $\ciid_R(M)$ simultaneously finite? \end{questions} \section{Structure of Quasi-Deformations} \label{sec5} Given a quasi-deformation of $R$, we show in this section how to construct nicer quasi-deformations that are related to the original one. One outcome is the proof (in~\ref{qd02}) of Theorem~\ref{thmF} from the introduction. This is similar in spirit to~\cite[(1.14)]{avramov:cid}, but different in scope, and will be used frequently in the sequel. \begin{lem} \label{qd01} Let $R\xrightarrow{\varphi}R'\xleftarrow{\tau}Q$ be a codimension $c$ quasi-deformation of the local ring $(R,\ideal{m})$ such that $R'/\ideal{m} R'$ is Cohen-Macaulay. There exists a commutative diagram of local ring homomorphisms $$\xymatrix{ R\ar[r]^-{\varphi}\ar[rd]_-{\varphi'} & R' \ar[d]^f & Q \ar[l]_-{\tau} \ar[d]^g \\ & R'' & Q' \ar[l]_-{\tau'} }$$ such that $\varphi'$ is flat with Gorenstein closed fibre, $\tau'$ is surjective with kernel generated by $Q'$-regular sequence of length $c$, the natural map $R'\otimes_QQ'\to R''$ is bijective, and one has $\operatorname{Tor}^Q_{\geq 1}(R',Q')=0$ and $\dim(R''/\ideal{m} R'')=\dim(R'/\ideal{m} R')$. \end{lem} \begin{proof} Fix a Cohen factorization $R\to S\to \comp{R'}$ of the semi-completion $\grave\varphi\colon R\to\comp{R'}$. Because $\varphi$ is flat with Cohen-Macaulay closed fibre, we know from~\cite[(3.8.1),(3.8.3)]{avramov:solh} that $\comp{R'}$ is perfect as an $S$-module, say, of grade $g$. This yields $\ext^n_S(\comp{R'},S)=0$ for each $n\neq g$, so there is an isomorphism $\rhom_S(\comp{R'},S)\simeq\mathsf{\Sigma}^{-g}\ext^g_S(\comp{R'},S)$. (In the language of~\cite{avramov:rhafgd} this complex is \emph{dualizing for $\varphi$}. See~\cite{avramov:rhafgd,frankild:qcmpolh} for more properties and applications.) In particular, the $\comp{R'}$-complex $\rhom_S(\comp{R'},S)$ is semidualizing by~\cite[(6.6)]{christensen:scatac}, and so the module $C=\ext^g_S(\comp{R'},S)$ is a semidualizing $\comp{R'}$-module. Hence, we have from the definition $\ext^{\geq 1}_{\comp{R'}}(C,C)=0$. We claim that $C$ is flat as an $R$-module. The module $C$ is finitely generated over $\comp{R'}$, and the semi-completion $\grave\varphi\colon R\to\comp{R'}$ is a local ring homomorphism. It follows easily that $\inf(k\otimes^{\mathbf{L}}_R C)=0$. Also, we have an equality $\operatorname{fd}_R(C)=\sup(k\otimes^{\mathbf{L}}_R C)$ from~\cite[(5.5.F)]{avramov:hdouc}, and so it suffices to show $\operatorname{amp}(k\otimes^{\mathbf{L}}_R C)=0$. From~\cite[(5.10)]{avramov:rhafgd} we know that $k\otimes^{\mathbf{L}}_R C$ is dualizing for $\comp{R'}/\ideal{m}\comp{R'}$. Since the ring $\comp{R'}/\ideal{m}\comp{R'}$ is Cohen-Macaulay, this implies $\operatorname{amp}(k\otimes^{\mathbf{L}}_R C)=0$ by~\cite[(3.7)]{christensen:scatac} as desired. The ring $\comp{Q}$ is complete, and the local ring homomorphism $\comp{\tau}\colon\comp{Q}\to\comp{R'}$ is surjective with kernel generated by a $\comp{Q}$-regular sequence $\mathbf{x}$ of length $c$. Using~\cite[(1.7)]{auslander:lawlom}, the vanishing $\ext^{2}_{\comp{R'}}(C,C)=0$ implies that there is a finitely generated $\comp{Q}$-module $B$ such that $\mathbf{x}$ is $B$-regular and $C\cong\comp{R'}\otimes_{\comp Q}B\cong B/\mathbf{x} B$. Let $\tau_1\colon B\to B/\mathbf{x} B\cong C$ denote the natural surjection. Now we construct the desired diagram guided by the following. $$\xymatrix{ R\ar[r]^-{\varphi}\ar[rd]_-{\grave\varphi} & R' \ar[d]^-{\varepsilon_{R'}} & Q \ar[l]_-{\tau} \ar[d]^-{\varepsilon_{Q}} \\ & \comp{R'}\ar[d]^-{f_1} & \comp{Q} \ar[l]_-{\comp\tau}\ar[d]^-{g_1} \\ & \comp{R'}\ltimes C & \comp{Q}\ltimes B \ar[l]_-{\ \ \tau'} }$$ The triangle in the diagram commutes by definition, and the upper square commutes by~\ref{basics01}. One checks readily that the map $\tau':=\comp{\tau}\ltimes\tau_1\colon \comp{Q}\ltimes B\to \comp{R'}\ltimes C$ is a local ring homomorphism making the bottom square commute. The fact that the sequence $\mathbf{x}$ is $\comp{Q}$-regular and $B$-regular implies $$\operatorname{Tor}^Q_{\geq 1}(R',\comp{Q}\ltimes B)\cong\operatorname{Tor}^Q_{\geq 1}(Q/\mathbf{x} Q,\comp{Q}\ltimes B)=0.$$ Using the surjectivity of $\tau$ one checks that the natural map $R'\otimes_Q\comp{Q}\to \comp{R'}$ is bijective. The construction of $\tau'$ shows that $\operatorname{Ker}(\tau')$ is generated by the $(\comp{Q}\ltimes B)$-regular sequence $g(\mathbf{x})$ and the natural map $\comp{R'}\otimes_{\comp{Q}}(\comp{Q}\ltimes B)\to \comp{R'}\ltimes C$ is bijective. This implies that the natural map $R'\otimes_Q(\comp{Q}\ltimes B)\to \comp{R'}\ltimes C$ is bijective. As $\comp{R'}$ and $C$ are both $R$-flat, we see that the composition $\varphi'=f_1\circ\grave\varphi=f_1\circ\varepsilon_{R'}\circ\varphi$ is flat. Also, the closed fibre of $\varphi'$ is $\comp{R'}/\ideal{m}\comp{R'}\ltimes C/\ideal{m} C$. Because $k\otimes^{\mathbf{L}}_R C\simeq C/\ideal{m} C$ is dualizing for $k\otimes_R\comp{R'}$, we conclude that the closed fibre of $\varphi'$ is Gorenstein. Thus, setting $R''=\comp{R'}\ltimes C$ and $Q'=\comp{Q}\ltimes B$ yields the desired diagram. \end{proof} \begin{para} \label{qd02} \emph{Proof of Theorem~\ref{thmF}.} One implication follows immediately from the definition of $\cifd_R(M)$. For the other implication, assume $\cifd_R(M)<\infty$ and fix a quasi-deformation $R\xrightarrow{\varphi}R'\xleftarrow{\tau}Q$ such that $\operatorname{fd}_{Q}(R'\otimes^{\mathbf{L}}_RM)<\infty$. Choose a prime ideal $P\in\operatorname{Min}_{R'}(R'/\ideal{m} R')$ and set $\ideal{p}=\tau^{-1}(P)$. From~\cite[(5.1.F)]{avramov:hdouc} we conclude $$\operatorname{fd}_{Q_{\ideal{p}}}(R'_P\otimes^{\mathbf{L}}_R M)=\operatorname{fd}_{Q_{\ideal{p}}}((R'\otimes^{\mathbf{L}}_R M)_{\ideal{p}})<\infty.$$ One checks readily that the localized diagram $R\xrightarrow{\varphi_P} R'_P\xleftarrow{\tau_P} Q_{\ideal{p}}$ is a quasi-deformation and that the closed fibre $R'_P/\ideal{m} R'_P\cong(R'/\ideal{m} R')_P$ is artinian. Thus, we may replace the original quasi-deformation with the localized one in order to assume that the closed fibre $R'/\ideal{m} R'$ is artinian and, hence, Cohen-Macaulay. Lemma~\ref{qd01} now yields a commutative diagram of local ring homomorphisms $$\xymatrix{ R\ar[r]^-{\varphi}\ar[rd]_-{\varphi'} & R' \ar[d]^f & Q \ar[l]_-{\tau} \ar[d]^g \\ & R'' & Q' \ar[l]_-{\tau'} }$$ such that $\varphi'$ is flat with Gorenstein closed fibre, $\tau'$ is surjective with kernel generated by $Q$-regular sequence, the natural map $Q'\otimes_QR'\to R''$ is bijective, and one has $\operatorname{Tor}^Q_{\geq 1}(Q',R')=0$ and $\dim(R''/\ideal{m} R'')=\dim(R'/\ideal{m} R')=0$. It follows that the diagram $R\xrightarrow{\varphi'}R''\xleftarrow{\tau'}Q'$ is a quasi-deformation, and so it suffices to show that $\operatorname{fd}_{Q'}(R''\otimes^{\mathbf{L}}_RM)$ is finite. Note that the conditions $Q'\otimes_QR'\cong R''$ and $\operatorname{Tor}^Q_{\geq 1}(R',Q')=0$ yield an isomorphism $Q'\otimes^{\mathbf{L}}_QR'\simeq R''$. This yields the second equality in the following sequence: \begin{align*} \operatorname{fd}_{Q'}(R''\otimes^{\mathbf{L}}_RM) &=\operatorname{fd}_{Q'}(R''\otimes^{\mathbf{L}}_{R'}(R'\otimes^{\mathbf{L}}_RM)) \\ &=\operatorname{fd}_{Q'}((Q'\otimes^{\mathbf{L}}_QR')\otimes^{\mathbf{L}}_{R'}(R'\otimes^{\mathbf{L}}_RM)) \\ &=\operatorname{fd}_{Q'}(Q'\otimes^{\mathbf{L}}_{Q}(R'\otimes^{\mathbf{L}}_RM)) \\ &\leq \operatorname{fd}_Q(R'\otimes^{\mathbf{L}}_RM) \\ &<\infty. \end{align*} The first equality follows from the commutativity of the displayed diagram. The third equality is tensor-cancellation. The first inequality is from~\cite[(4.2.F)]{avramov:hdouc}, and the second inequality is by assumption. \qed \end{para} \begin{disc} \label{qd03} We do not know whether there is a result like Theorem~\ref{thmF} for $\ciid$. In fact, it is not even clear if, given an $R$-complex of finite complete intersection injective dimension, there exists a quasi-deformation $R\to R'\leftarrow Q$ such that $Q$ is complete and $\id_Q(R'\otimes^{\mathbf{L}}_R M)$ is finite. The place where the proof breaks down is in the final displayed sequence: it is not true in general that the finiteness of $\id_Q(R'\otimes^{\mathbf{L}}_RM)$ implies $\id_{Q'}(Q'\otimes^{\mathbf{L}}_{Q}(R'\otimes^{\mathbf{L}}_RM))<\infty$. However, see Proposition~\ref{qd05}. \end{disc} Before proving a version of Theorem~\ref{thmF} for upper complete intersection dimension, we require the following extension of a result of Foxby~\cite[Thm.~1]{foxby:imufbc} for complexes. We shall apply it to the completion homomorphism $\varepsilon^{}_Q\colon Q\to \comp{Q}$. Recall that an $R$-complex $M$ is \emph{minimal} if, for every homotopy equivalence $\alpha\colon M\to M$, each map $\alpha_i\colon M_i\to M_i$ is bijective. \begin{lem} \label{id01} Let $Q\to Q'$ be a flat local ring homomorphism and $N$ a homologically bounded $Q$-complex such that, for every prime $\ideal{p}\in\operatorname{Spec}(Q)$ such that $N_{\ideal{p}}\not\simeq 0$, the fibre $Q'\otimes_Q (Q_{\ideal{p}}/\ideal{p} Q_{\ideal{p}})$ is Gorenstein. If $\id_Q(N)<\infty$, then $\id_{\comp{Q}}(\comp Q\otimes^{\mathbf{L}}_Q N)<\infty$. \end{lem} \begin{proof} Within this proof, we work in the category of $Q$-complexes, as opposed to the derived category. In particular, a morphism of $Q$-complexes $f\colon X\to Y$ is a ``quasiisomorphism'' if each map $\operatorname{H}_n(f)\colon \operatorname{H}_n(X)\to \operatorname{H}_n(Y)$ is bijective, and $f$ is an ``isomorphism'' if each map $f_n\colon X_n\to Y_n$ is bijective. Quasiisomorphisms are identified by the symbol $\simeq$, and isomorphisms are identified by the symbol $\cong$. Set $j=\id_Q(N)$ and let $N\simeq I$ be an injective resolution over $Q$ such that $I_n=0$ for all $n<-j$. From~\cite[(12.2.2)]{avramov:dgha} there exist $Q$-complexes $I'$ and $I''$ such that $I'$ is minimal and such that $I''\simeq 0$ and $I\cong I'\oplus I''$. It follows that $I'$ is a bounded complex of injective $Q$-modules such that $N\simeq I'$, so we may replace $I$ with $I'$ in order to assume that $I$ is minimal. \textbf{Claim.} Given a prime $\ideal{q}\in\operatorname{Spec}(Q)$ such that $N_{\ideal{q}}\simeq 0$, we have $I_{\ideal{q}}= 0$. To prove this, we start with the quasiisomorphisms $I_{\ideal{q}}\simeq N_{\ideal{q}}\simeq 0$ and consider the natural morphism $g\colon I\to I_{\ideal{q}}$. For each integer $n$, the module $I_n$ is isomorphic to a direct sum of injective hulls $E_Q(Q/\ideal{p})$. Given the isomorphism of $Q$-modules $$E_Q(Q/\ideal{p})_{\ideal{q}} \begin{cases} =0 & \text{if $\ideal{p}\not\subseteq\ideal{q}$} \\ \cong E_Q(Q/\ideal{p}) & \text{if $\ideal{p}\subseteq\ideal{q}$} \end{cases} $$ it follows readily that each $g_n$ is a split surjection. From this we conclude that the complex $J=\operatorname{Ker}(g)$ is a bounded complex of injective $Q$-modules. Furthermore, we use the quasiisomorphism $I_{\ideal{q}}\simeq 0$ in the long exact sequence associated to the exact sequence $0\to J\xrightarrow{h} I\xrightarrow{g} I_{\ideal{q}}\to 0$ in order to conclude that $h$ is a quasiisomorphism. Because $J$ and $I$ are bounded complexes of injective $Q$-modules, it follows that $h$ is a homotopy equivalence. Now, using~\cite[(1.7.1)]{avramov:aratc} we conclude that $h$ is surjective and so $I_{\ideal{q}}=0$. From~\cite[Thm.~1]{foxby:imufbc} the fact that the formal fibre $Q'\otimes_Q (Q_{\ideal{p}}/\ideal{p} Q_{\ideal{p}})$ is Gorenstein whenever $N_{\ideal{p}}\not\simeq 0$ implies that each module $Q'\otimes_Q I_n$ has finite injective dimension over $Q'$. Hence, the complex $Q'\otimes_Q I$ is a bounded complex of $Q'$-modules of finite injective dimension. From the quasiisomorphism $Q'\otimes^{\mathbf{L}}_Q N\simeq Q'\otimes_Q I$ we conclude $\id_{Q'}(Q'\otimes^{\mathbf{L}}_Q N)=\id_{Q'}(Q'\otimes_Q I)<\infty$. \end{proof} In the conclusion of the next result, notice that $R''$ is complete, and hence it has Gorenstein formal fibres. \begin{prop} \label{qd04} Let $(R,\ideal{m})$ be a local ring and $M$ a homologically bounded $R$-complex. Then $\uciid_R(M)<\infty$ if and only if there exists a quasi-deformation $R\to R''\leftarrow Q'$ such that $Q'$ is complete, the closed fibre $R''/\ideal{m} R''$ is artinian, and Gorenstein, and $\id_{Q'}(R''\otimes^{\mathbf{L}}_RM)<\infty$. \end{prop} \begin{proof} For the nontrivial implication, assume $\uciid_R(M)<\infty$, and fix a quasi-deformation $R\xrightarrow{\varphi}R'\xleftarrow{\tau}Q$ such that $R'$ has Gorenstein formal fibres, $R'/\ideal{m} R'$ is Gorenstein, and $\id_{Q}(R'\otimes^{\mathbf{L}}_RM)<\infty$. Choose a prime ideal $P\in\operatorname{Min}_{R'}(R'/\ideal{m} R')$ and set $\ideal{p}=\tau^{-1}(P)$. We conclude from~\cite[(5.1.I)]{avramov:hdouc} that the quantity $$\id_{Q_{\ideal{p}}}(R'_P\otimes^{\mathbf{L}}_R M)=\id_{Q_{\ideal{p}}}((R'\otimes^{\mathbf{L}}_R M)_{\ideal{p}})$$ is finite. Also, the ring $R'_P$ has Gorenstein formal fibres because $R'$ has the same. Because the ring $R'_P/\ideal{m} R'_P\cong (R'/\ideal{m} R')_P$ is artinian and Gorenstein the proof of Theorem~\ref{thmF} shows that we may replace the original quasi-deformation with a localized one in order to assume that the closed fibre $R'/\ideal{m} R'$ is artinian. Completing our quasi-deformation yields the next commutative diagram of local ring homomorphisms $$\xymatrix{ R\ar[r]^-{\varphi}\ar[rd]_-{\grave\varphi} & R' \ar[d]^{\varepsilon_{R'}} & Q \ar[l]_-{\tau} \ar[d]^{\varepsilon_{Q}} \\ & \comp{R'} & \comp{Q}. \ar[l]_-{\comp{\tau}} }$$ As in the proof of Theorem~\ref{thmF} it suffices to show that the quantity $$\id_{\comp Q}(\comp{R'}\otimes^{\mathbf{L}}_RM)=\id_{\comp Q}(\comp Q\otimes^{\mathbf{L}}_{Q}(R'\otimes^{\mathbf{L}}_RM))$$ is finite. For each $\ideal{p}\in \operatorname{Spec}(Q)$ such that $(R'\otimes^{\mathbf{L}}_RM)_{\ideal{p}}\not\simeq 0$, we have $\ideal{p}\supseteq\operatorname{Ker}(\tau)$ because $R'\otimes^{\mathbf{L}}_RM$ is an $R'$-complex and $\tau$ is surjective. For each such $\ideal{p}$, set $\ideal{p}'=\ideal{p} R'\in\operatorname{Spec}(R')$. The fact that $\tau$ is surjective yields an isomorphism of closed fibres $\comp{Q}\otimes_Q (Q_{\ideal{p}}/\ideal{p} Q_{\ideal{p}}) \cong \comp{R'}\otimes_{R'}(R'_{\ideal{p}'}/\ideal{p}' R'_{\ideal{p}'})$. In particular, each of these rings is Gorenstein by assumption. Thus, the finiteness of $\id_{\comp Q}(\comp Q\otimes^{\mathbf{L}}_{Q}(R'\otimes^{\mathbf{L}}_RM))$ follows from Lemma~\ref{id01}. \end{proof} We will see the utility of the next result in Theorem~\ref{cifdac01'}\eqref{cifdac01'item2'}. \begin{prop} \label{qd05} Let $(R,\ideal{m})$ be a local ring and $M$ a homologically bounded $R$-complex with finite length homology. Then $\ciid_R(M)$ is finite if and only if there exists a quasi-deformation $R\to R''\leftarrow Q'$ such that $Q'$ is complete, the closed fibre $R''/\ideal{m} R''$ is artinian, and $\id_{Q'}(R''\otimes^{\mathbf{L}}_RM)$ is finite. \end{prop} \begin{proof} For the nontrivial implication, assume that $\ciid_R(M)$ is finite and fix a quasi-deformation $R\xrightarrow{\varphi}R'\xleftarrow{\tau}Q$ such that $\id_{Q}(R'\otimes^{\mathbf{L}}_RM)<\infty$. Localizing as in the proof of Proposition~\ref{qd04}, we may assume that the closed fibre $R'/\ideal{m} R'$ is artinian. Because each $R$-module $\operatorname{H}_i(M)$ has finite length and the closed fibre $R'/\ideal{m} R'$ is artinian, it follows that each module $\operatorname{H}_i(R'\otimes^{\mathbf{L}}_RM)\cong R'\otimes_R\operatorname{H}_i(M)$ has finite length over $R'$ and therefore over $Q$. In particular, there is only one prime $\ideal{p}\in\operatorname{Spec}(Q)$ such that $(R'\otimes^{\mathbf{L}}_RM)_{\ideal{p}}\not\simeq 0$, namely, the maximal ideal $\ideal{p}=\ideal{n}\subset Q$. Because the formal fibre $\comp{Q}\otimes_Q Q/\ideal{n}\cong Q/\ideal{n}$ is Gorenstein, the argument of Proposition~\ref{qd04} shows that $\id_{\comp{Q}}(\comp{R'}\otimes^{\mathbf{L}}_R M)$ is finite. Hence the completed quasi-deformation $R\xrightarrow{\grave\varphi}\comp{R'}\xleftarrow{\comp\tau}\comp{Q}$ has the desired properties. \end{proof} The next result follows from Propositions~\ref{qd04} and~\ref{qd05} via the proof of~\cite[(1.13)]{avramov:cid}. \begin{cor} \label{ciid02} Let $R$ be a local ring and $M$ a homologically bounded $R$-complex. \begin{enumerate}[\quad\rm(a)] \item \label{ciid02item1} One has $\uciid_R(M)<\infty$ if and only if $\uciid_{\comp{R}}(\comp{R}\otimes^{\mathbf{L}}_R M)<\infty$. \item \label{ciid02item2} If $M$ has finite length homology, then one has $\ciid_R(M)<\infty$ if and only if $\ciid_{\comp{R}}(\comp{R}\otimes^{\mathbf{L}}_R M)<\infty$. \qed \end{enumerate} \end{cor} \section{Stability Results} \label{sec6} This section documents some situations where combinations of complexes either detect or inherit homological properties of their component pieces. We start with partial converses to parts of~\cite[(2.1)]{christensen:apac} which are useful, e.g., for Theorem~\ref{cifdac01'}\eqref{cifdac01'item2'}. \begin{prop} \label{sdc06} Let $R$ be a local ring and $C$ a semidualizing $R$-complex. Let $M$ and $N\not\simeq 0$ be homologically finite $R$-complexes with $\pd_R(N)<\infty$. \begin{enumerate}[\quad\rm(a)] \item \label{sdc06item2} If $M\otimes^{\mathbf{L}}_RN\in\cat{A}_C(R)$ or $\rhom_R(N,M)\in\cat{A}_C(R)$, then $M\in\cat{A}_C(R)$. \item \label{sdc06item4} If $M\otimes^{\mathbf{L}}_RN\in\cat{B}_C(R)$ or $\rhom_R(N,M)\in\cat{B}_C(R)$, then $M\in\cat{B}_C(R)$. \end{enumerate} \end{prop} \begin{proof} We prove part~\eqref{sdc06item2}. The proof of~\eqref{sdc06item4} is similar. Assume first that $M\otimes^{\mathbf{L}}_RN$ is in $\cat{A}_C(R)$. This implies that $(C\otimes^{\mathbf{L}}_R M)\otimes^{\mathbf{L}}_R N\simeq C\otimes^{\mathbf{L}}_R(M\otimes^{\mathbf{L}}_R N)$ is homologically bounded, so~\cite[(3.1)]{foxby:daafuc} implies that $C\otimes^{\mathbf{L}}_R M$ is homologically bounded. (The corresponding implication in the proof of~\eqref{sdc06item4} uses tensor-evaluation~\eqref{basics03}.) We consider the following commutative diagram wherein the unmarked isomorphism is tensor-associativity and $\omega_{CCM}^{}$ is tensor-evaluation~\eqref{basics03}. $$\xymatrix{ M\otimes^{\mathbf{L}}_RN \ar[rr]^-{\gamma_M^C\otimes^{\mathbf{L}}_R N} \ar[d]_-{\gamma_{M\otimes^{\mathbf{L}}_R N}^C} && (C\otimes^{\mathbf{L}}_R\rhom_R(C,M))\otimes^{\mathbf{L}}_R N \ar[d]_{\simeq} \\ C\otimes^{\mathbf{L}}_R\rhom_R(C,M\otimes^{\mathbf{L}}_RN) && C\otimes^{\mathbf{L}}_R(\rhom_R(C,M)\otimes^{\mathbf{L}}_RN) \ar[ll]_{C\otimes^{\mathbf{L}}_R\omega^{}_{CCM}}^-{\simeq} }$$ Because $\gamma_{M\otimes^{\mathbf{L}}_R N}^C$ is an isomorphism, the diagram shows that the same is true of $\gamma_M^C\otimes^{\mathbf{L}}_RN$, and so~\cite[(2.10)]{iyengar:golh} implies that $\gamma_M^C$ is an isomorphism. Assume next $\rhom_R(N,M)\in\cat{A}_C(R)$. Remark~\ref{fpd01} says that $\rhom_R(N,R)$ is a homologically finite $R$-complex of finite projective dimension such that $$\rhom_R(N,M)\simeq\rhom_R(N,R)\otimes^{\mathbf{L}}_R M.$$ Hence, using $\rhom_R(N,R)$ in place of $N$ in the first case, we find $M\in\cat{A}_C(R)$. \end{proof} The following result paves the way for Proposition~\ref{sdm03} which is used in the proof of Theorem~\ref{cifdac01}\eqref{cifdac01item4}. \begin{lem} \label{sdm04} Let $R$ be a local ring and $C$ a semidualizing $R$-module. For each homologically bounded $R$-complex $M$, the following conditions are equivalent. \begin{enumerate}[\quad\rm(i)] \item \label{sdm04item1} $\text{G-}\!\pd_R(M)$ is finite. \item \label{sdm04item2} $\text{G-}\!\fd_R(M)$ is finite. \item \label{sdm04item3} $\comp{R}\otimes^{\mathbf{L}}_RM$ is in $\cat{A}_{D^{\comp{R}}}(\comp{R})$. \end{enumerate} \end{lem} \begin{proof} \eqref{sdm04item1}$\iff$\eqref{sdm04item3}. When $M$ is a module, this is proved in~\cite[(3.5)]{esmkhani:ghdac}. For the general case, let $F\simeq M$ be an $R$-flat resolution of $M$. It follows that $\comp{R}\otimes^{\mathbf{L}}_{R} F\simeq\comp{R}\otimes^{\mathbf{L}}_{R} M$ is an $\comp{R}$-flat resolution of $\comp{R}\otimes^{\mathbf{L}}_{R} M$. Each $F_i$ is a flat $R$-module and is therefore G-flat over $R$. Thus, the quantity $\text{G-}\!\fd_{R}(M)$ is finite if and only if the kernel $K_i=\operatorname{Ker}(\partial^F_i)$ has finite G-flat dimension over $R$ for some (equivalently, every) $i>\sup(M)$. Fix an integer $i>\sup(M)$ and consider the following truncations of $F$ \begin{align*} F' &=\qquad 0\to K_i\to F_i\to F_{i-1}\to\cdots \\ F_{\leq i}&=\qquad 0\xrightarrow{\hspace{2.5mm}} 0\xrightarrow{\hspace{2.5mm}} F_i\to F_{i-1}\to\cdots. \end{align*} Our choice of $i$ implies $F'\simeq M$, and so $\comp{R}\otimes_RF'\simeq \comp{R}\otimes_RM$. It follows that $\comp{R}\otimes^{\mathbf{L}}_RM$ is in $\cat{A}_{D^{\comp{R}}}(\comp{R})$ if and only if $\comp{R}\otimes^{\mathbf{L}}_RF'$ is in $\cat{A}_{D^{\comp{R}}}(\comp{R})$. The complexes from the display fit into an exact sequence as follows $$0\to F_{\leq i}\to F'\to\mathsf{\Sigma}^{i+1}K_i\to 0$$ and applying $\comp{R}\otimes_R-$ to this sequence yields a second exact sequence $$0\to \comp{R}\otimes_RF_{\leq i}\to \comp{R}\otimes_RF'\to\mathsf{\Sigma}^{i+1}\comp{R}\otimes_RK_i\to 0.$$ The complex $\comp{R}\otimes_RF_{\leq i}$ is a bounded complex of $\comp{R}$-flat modules, and so it is in $\cat{A}_{D^{\comp{R}}}(\comp{R})$ by Remark~\ref{lwc01}. A standard argument using the exact sequence implies that $\comp{R}\otimes_RF'$ is in $\cat{A}_{D^{\comp{R}}}(\comp{R})$ if and only if $\mathsf{\Sigma}^{i+1}\comp{R}\otimes_RK_i$ is in $\cat{A}_{D^{\comp{R}}}(\comp{R})$. That is, the complex $\comp{R}\otimes^{\mathbf{L}}_RM$ is in $\cat{A}_{D^{\comp{R}}}(\comp{R})$ if and only if $\comp{R}\otimes_RK_i$ is in $\cat{A}_{D^{\comp{R}}}(\comp{R})$. From the first paragraph of this proof, we know that $\text{G-}\!\fd_{R}(M)$ is finite if and only if $\text{G-}\!\fd_R(K_i)$ is finite. From~\cite[(3.5)]{esmkhani:ghdac}, we know that $\text{G-}\!\fd_{R}(K_i)$ is finite if and only if $\comp{R}\otimes^{\mathbf{L}}_RK_i$ is in $\cat{A}_{D^{\comp{R}}}(\comp{R})$. And the second paragraph shows that $\comp{R}\otimes^{\mathbf{L}}_RK_i$ is in $\cat{A}_{D^{\comp{R}}}(\comp{R})$ if and only if $\comp{R}\otimes_RM$ is in $\cat{A}_{D^{\comp{R}}}(\comp{R})$. Thus, the equivalence is established. The proof of \eqref{sdm04item2}$\iff$\eqref{sdm04item3} is similar using a projective resolution $P\simeq M$. \end{proof} \begin{prop} \label{sdm03} Let $R$ be a local ring and $C$ a semidualizing $R$-module. For each homologically bounded $R$-complex $M$, the following conditions are equivalent. \begin{enumerate}[\quad\rm(i)] \item \label{sdm03item1} $\text{G}_C\text{-}\!\pd_R(M)$ is finite. \item \label{sdm03item2} $\text{G}_C\text{-}\!\fd_R(M)$ is finite. \item \label{sdm03item3} $\comp{R}\otimes^{\mathbf{L}}_RM$ is in $\cat{A}_{\rhom_{\comp{R}}(\comp{C},D^{\comp{R}})}(\comp{R})$. \item \label{sdm03item4} $\gkpd{\comp{C}}_{\comp{R}}(\comp{R}\otimes^{\mathbf{L}}_RM)$ is finite. \item \label{sdm03item5} $\gkfd{\comp{C}}_{\comp{R}}(\comp{R}\otimes^{\mathbf{L}}_RM)$ is finite. \end{enumerate} \end{prop} \begin{proof} The equivalences \eqref{sdm03item3}$\iff$\eqref{sdm03item4}$\iff$\eqref{sdm03item5} are in Remarks~\ref{sdm02} and~\ref{sdm02'}. \eqref{sdm03item2}$\iff$\eqref{sdm03item5} We start by showing that $\text{G-}\!\fd_{R\ltimes C}(M)$ is finite if and only if $\comp{R}\otimes^{\mathbf{L}}_R M$ is in $\cat{A}_{D^{\comp{R\ltimes C}}}(\comp{R\ltimes C})$. One verifies the following isomorphisms readily $$\comp{R}\cong\comp{R\ltimes C}\otimes_{R\ltimes C}R \simeq\comp{R\ltimes C}\otimes^{\mathbf{L}}_{R\ltimes C}R$$ and from this, we have $$\comp{R}\otimes^{\mathbf{L}}_R M \simeq(\comp{R\ltimes C}\otimes^{\mathbf{L}}_{R\ltimes C}R)\otimes^{\mathbf{L}}_R M \simeq\comp{R\ltimes C}\otimes^{\mathbf{L}}_{R\ltimes C} M. $$ Hence, the desired equivalence follows from Lemma~\ref{sdm04}. From Remark~\ref{sdm02} we have $\text{G}_C\text{-}\!\fd_R(M)=\text{G-}\!\fd_{R\ltimes C}(M)$. The previous paragraph tells us that $\text{G-}\!\fd_{R\ltimes C}(M)$ is finite if and only if $\comp{R}\otimes^{\mathbf{L}}_R M$ is in $\cat{A}_{D^{\comp{R\ltimes C}}}(\comp{R\ltimes C})$. One readily verifies the isomorphism $\comp{R\ltimes C}\cong\comp{R}\ltimes \comp{C}$ and so $\text{G}_C\text{-}\!\fd_R(M)<\infty$ if and only if $\comp{R}\otimes^{\mathbf{L}}_R M\in\cat{A}_{D^{\comp{R}\ltimes \comp{C}}}(\comp{R}\ltimes \comp{C})$. Using Remark~\ref{sdm02'}, we conclude that $\comp{R}\otimes^{\mathbf{L}}_R M$ is in $\cat{A}_{D^{\comp{R}\ltimes \comp{C}}}(\comp{R}\ltimes \comp{C})$ if and only if $\text{G-}\!\fd_{\comp{R}\ltimes \comp{C}}(\comp{R}\otimes^{\mathbf{L}}_R M)$ is finite, that is, if and only if $\gkfd{\comp{C}}_{\comp{R}}(\comp{R}\otimes^{\mathbf{L}}_R M)$ is finite. The equivalence \eqref{sdm03item1}$\iff$\eqref{sdm03item4} is verified similarly. \end{proof} The remaining results of this section deal with stability for complete intersection dimensions. The first one is used in the proof of Theorem~\ref{cifdac01'}\eqref{cifdac01'item2'}. \begin{prop} \label{stable01} Let $R$ be a local ring and let $M$ and $N$ be homologically bounded $R$-complexes with $\operatorname{fd}_R(N)<\infty$. \begin{enumerate}[\quad\rm(a)] \item \label{stable01item1} If $\cifd_R(M)<\infty$, then $\cifd_R(M\otimes^{\mathbf{L}}_RN)<\infty$. \item \label{stable01item2} If $\ciid_R(M)<\infty$, then $\ciid_R(M\otimes^{\mathbf{L}}_RN)<\infty$. \item \label{stable01item3} If $\uciid_R(M)<\infty$, then $\uciid_R(M\otimes^{\mathbf{L}}_RN)<\infty$. \end{enumerate} \end{prop} \begin{proof} \eqref{stable01item1} Let $R\to R'\leftarrow Q$ be a quasi-deformation such that $\operatorname{fd}_Q(R'\otimes^{\mathbf{L}}_R M)<\infty$. The finiteness of $\operatorname{fd}_R(N)$ implies $\operatorname{fd}_{R'}(R'\otimes^{\mathbf{L}}_RN)<\infty$ by~\cite[(4.2)]{avramov:hdouc} and so~\cite[(4.1.F)]{avramov:hdouc} provides the finiteness in the next display $$\operatorname{fd}_Q(R'\otimes^{\mathbf{L}}_R (M\otimes^{\mathbf{L}}_RN)) =\operatorname{fd}_Q((R'\otimes^{\mathbf{L}}_R M)\otimes^{\mathbf{L}}_{R'}(R'\otimes^{\mathbf{L}}_RN))<\infty.$$ Hence, we have $\cifd_R(M\otimes^{\mathbf{L}}_RN)<\infty$ by definition. Parts~\eqref{stable01item2} and~\eqref{stable01item3} are proved like~\eqref{stable01item1} using~\cite[(4.5.F)]{avramov:hdouc}. \end{proof} \begin{prop} \label{stable02} Let $R$ be a local ring with Gorenstein formal fibres. Let $M$ be a homologically finite $R$-complex and let $N$ be homologically bounded $R$-complex with $\id_R(N)<\infty$. \begin{enumerate}[\quad\rm(a)] \item \label{stable02item1} If $\cifd_R(M)<\infty$, then $\uciid_R(\rhom_R(M,N))<\infty$. \item \label{stable02item2} If $\uciid_R(M)<\infty$, then $\cifd_R(\rhom_R(M,N))<\infty$. \end{enumerate} \end{prop} \begin{proof} \eqref{stable02item1} Use Theorem~\ref{thmF} to find a quasi-deformation $R\to R'\leftarrow Q$ such that $R'/\ideal{m} R'$ is Gorenstein and $\operatorname{fd}_Q(R'\otimes^{\mathbf{L}}_R M)<\infty$. Because $R$ has Gorenstein formal fibres, we learn from~\cite[(4.1)]{avramov:glp} that $R'$ has Gorenstein formal fibres and, for each prime $\ideal{p}\in\operatorname{Spec}(R)$, the fibre $R'\otimes_R (R_{\ideal{p}}/\ideal{p} R_{\ideal{p}})$ is Gorenstein. Using~\cite[Thm.~1]{foxby:imufbc} as in the proof of Proposition~\ref{qd04} we conclude that $\id_{R'}(R'\otimes^{\mathbf{L}}_R N)$ is finite. Because $M$ is homologically finite and $R'$ is flat over $R$, tensor-evaluation~\eqref{basics03} yields the first isomorphism in the following sequence: \begin{align*} R'\otimes^{\mathbf{L}}_R\rhom_R(M,N) &\simeq \rhom_R(M,R'\otimes^{\mathbf{L}}_RN)\\ &\simeq \rhom_R(M,\rhom_{R'}(R',R'\otimes^{\mathbf{L}}_RN))\\ &\simeq \rhom_{R'}(R'\otimes^{\mathbf{L}}_RM,R'\otimes^{\mathbf{L}}_RN). \end{align*} The second isomorphism comes from the fact that $R'\otimes^{\mathbf{L}}_R N$ is an $R'$-complex, and the third isomorphism is Hom-tensor adjointness. This sequence yields the equality in the next sequence and the finiteness is from~\cite[(4.1.I)]{avramov:hdouc}: $$\id_Q(R'\otimes^{\mathbf{L}}_R\rhom_R(M,N)) =\id_Q(\rhom_{R'}(R'\otimes^{\mathbf{L}}_RM,R'\otimes^{\mathbf{L}}_RN))<\infty.$$ Hence, we have $\ciid_R(\rhom_R(M,N))<\infty$ by definition. Part~\eqref{stable02item2} is proved like~\eqref{stable02item1} using~\cite[(4.5.I)]{avramov:hdouc}. \end{proof} The previous result yields the following behavior of complete intersection dimensions with respect to ``dagger-duality''. \begin{cor} \label{stable03} Let $R$ be a local ring admitting a dualizing complex $D$, and let $M$ be a homologically finite $R$-complex. \begin{enumerate}[\quad\rm(a)] \item \label{stable03item1} We have $\cifd_R(M)<\infty$ if and only if $\uciid_R(\rhom_R(M,D))<\infty$. \item \label{stable03item2} We have $\uciid_R(M)<\infty$ if and only if $\cifd_R(\rhom_R(M,D))<\infty$. \end{enumerate} \end{cor} \begin{proof} Because $R$ admits a dualizing complex, it has Gorenstein formal fibres by~\cite[(V.3.1)]{hartshorne:rad}. So, if $\cifd_R(M)<\infty$, then $\uciid_R(\rhom_R(M,D))<\infty$ by Proposition~\ref{stable02}\eqref{stable02item1}. Conversely, if $\uciid_R(\rhom_R(M,D))<\infty$, then the isomorphism $$M\simeq\rhom_R(\rhom_R(M,D),D)$$ from Remark~\ref{sdc04} implies $\cifd_R(M)<\infty$ by Proposition~\ref{stable02}\eqref{stable02item2}. This establishes part~\eqref{stable03item1}, and part~\eqref{stable03item2} is proved similarly. \end{proof} \begin{prop} \label{stable04} Let $R$ be a local ring. Let $M$ be a homologically finite $R$-complex with $\pd_R(M)<\infty$ and let $N$ be homologically bounded $R$-complex. \begin{enumerate}[\quad\rm(a)] \item \label{stable04item1} If $\cifd_R(N)<\infty$, then $\cifd_R(\rhom_R(M,N))<\infty$. \item \label{stable04item2} If $\ciid_R(N)<\infty$, then $\ciid_R(\rhom_R(M,N))<\infty$. \item \label{stable04item3} If $\uciid_R(N)<\infty$, then $\uciid_R(\rhom_R(M,N))<\infty$. \end{enumerate} \end{prop} \begin{proof} Remark~\ref{fpd01} says that $\rhom_R(M,R)$ is a homologically finite $R$-complex of finite projective dimension such that $\rhom_R(N,M)\simeq\rhom_R(N,R)\otimes^{\mathbf{L}}_R M$. Hence, the desired result follows from Proposition~\ref{stable01}. \end{proof} \section{Complete Intersection Dimensions and Foxby Classes} \label{sec4} The first result of this section contains Theorem~\ref{thmA} from the introduction. \begin{thm} \label{cifdac01} Let $R$ be a local ring and fix a homologically bounded $R$-complex $M$ and a semidualizing $R$-complex $C$. \begin{enumerate}[\quad\rm(a)] \item \label{cifdac01item1} If $\cifd_R(M)<\infty$, then $M\in\cat{A}_C(R)$. \item \label{cifdac01item4} If $C$ is a module and $\cifd_R(M)<\infty$, then $\text{G}_C\text{-}\!\fd_R(M)<\infty$. \item \label{cifdac01item3} If $M$ is homologically finite and $\cidim_R(M)<\infty$, then $\gkdim{C}_R(M)<\infty$. \end{enumerate} \end{thm} \begin{proof} \eqref{cifdac01item1} Assume $\cifd_R(M)<\infty$, and use Theorem~\ref{thmF} to find a quasi-deformation $R\to R'\leftarrow Q$ such that $Q$ is complete and $\operatorname{fd}_Q(R'\otimes^{\mathbf{L}}_R M)<\infty$. From~\cite[(4.2)]{frankild:sdcms} there is a semidualizing $Q$-complex $B$ such that $R'\otimes^{\mathbf{L}}_Q B\simeq R'\otimes^{\mathbf{L}}_R C$. The finiteness of $\operatorname{fd}_Q(R'\otimes^{\mathbf{L}}_R M)$ implies $R'\otimes^{\mathbf{L}}_R M\in\cat{A}_B(Q)$, and so $R'\otimes^{\mathbf{L}}_R M\in\cat{A}_{R'\otimes^{\mathbf{L}}_QB}(R') =\cat{A}_{R'\otimes^{\mathbf{L}}_RC}(R')$, and hence $M\in\cat{A}_C(R)$; see Remark~\ref{lwc01}. \eqref{cifdac01item4} and \eqref{cifdac01item3} The assumption $\cifd_R(M)<\infty$ implies $\cifd_{\comp{R}}(\comp{R}\otimes^{\mathbf{L}}_RM)<\infty$ by Remark~\ref{cidim04}. If $D^{\comp{R}}$ is dualizing for $\comp R$, then part~\eqref{cifdac01item1} implies that $\comp{R}\otimes^{\mathbf{L}}_RM$ is in $\cat{A}_{\rhom_{\comp{R}}(\comp{C},D^{\comp{R}})}(\comp{R})$. When $C$ is a module, Proposition~\ref{sdm03} implies $\text{G}_C\text{-}\!\fd_R(M)<\infty$. When $M$ is homologically finite, Remark~\ref{sdm02''} implies $\gkdim{\comp{R}\otimes_R C}(\comp{R}\otimes^{\mathbf{L}}_RM)<\infty$, and so $\gcdim_R(M)<\infty$ by~\cite[(5.10)]{christensen:scatac}. \end{proof} \begin{thm} \label{cifdac01'} Let $(R,\ideal{m},k)$ be a local ring and fix a homologically bounded $R$-complex $M$ and a semidualizing $R$-complex $C$. \begin{enumerate}[\quad\rm(a)] \item \label{cifdac01'item2} If $\uciid_R(M)<\infty$, then $M\in\cat{B}_C(R)$. \item \label{cifdac01'item2'} If $M$ is homologically finite and $\ciid_R(M)<\infty$, then $M\in\cat{B}_C(R)$. \item \label{cifdac01'item5} Assume that $R$ admits a dualizing complex $D^R$ and that $C$ is a module. If either $\uciid_R(M)<\infty$ or $M$ is homologically finite and $\ciid_R(M)<\infty$, then $\text{G}_C\text{-}\!\id_R(M)<\infty$. \end{enumerate} \end{thm} \begin{proof} \eqref{cifdac01'item2} This is proved like Theorem~\ref{cifdac01}\eqref{cifdac01item1} using Proposition~\ref{qd04}. \eqref{cifdac01'item2'} Let $K$ be the Koszul complex over $R$ on a minimal generating sequence for the maximal ideal $\ideal{m}$. The conditions $\operatorname{fd}_R(K)<\infty$ and $\ciid_R(M)<\infty$ imply $\ciid_R(M\otimes^{\mathbf{L}}_RK)<\infty$ by Proposition~\ref{stable01}\eqref{stable01item2}. As the homology $\operatorname{H}(M\otimes^{\mathbf{L}}_RK)$ is a finite dimensional vector space over $k$, Proposition~\ref{qd05} yields a quasi-deformation $R\to R''\leftarrow Q'$ such that $Q'$ is complete and $\id_{Q'}(R''\otimes^{\mathbf{L}}_R(M\otimes^{\mathbf{L}}_RK))$ is finite. Arguing as in the proof of Theorem~\ref{cifdac01}~\eqref{cifdac01item1} we conclude that $M\otimes^{\mathbf{L}}_RK$ is in $\cat{B}_C(R)$, and so Proposition~\ref{sdc06}\eqref{sdc06item4} implies $M\in\cat{B}_C(R)$. \eqref{cifdac01'item5} By parts~\eqref{cifdac01'item2} and~\eqref{cifdac01'item2'}, the assumptions imply that $M$ is in $\cat{B}_{\rhom_{R}(C,D^{R})}(R)$, and from Remark~\ref{sdm02'}, we conclude $\text{G}_C\text{-}\!\id_R(M)<\infty$. \end{proof} \begin{questions} \label{ciidbc01} Does the conclusion of Theorem~\ref{cifdac01'}\eqref{cifdac01'item2} also hold if we replace the assumption $\uciid_R(M)<\infty$ with $\ciid_R(M)<\infty$? (c.f.~\ref{ci*id03}.) Does the conclusion of Theorem~\ref{cifdac01'}\eqref{cifdac01'item5} hold if we do not assume that $R$ admits a dualizing complex? \end{questions} \begin{para} \label{foxby01} \emph{Proof of Theorem~\ref{thmC}.} We start with the forward implication of~\eqref{thmCitem1}. Assume $\cifd_R(M)<\infty$ and use Theorem~\ref{thmF} to find a quasi-deformation $R\xrightarrow{\varphi} R'\xleftarrow{\tau} Q$ such that $Q$ is complete, $R'/\ideal{m} R'$ is Gorenstein, and $\operatorname{fd}_Q(R'\otimes^{\mathbf{L}}_R M)<\infty$. Note that, since $Q$ is complete, the same is true of $R'$, and so $R'$ has Gorenstein formal fibres. Because $\varphi$ is flat with Gorenstein closed fibre, we know that $R'\otimes^{\mathbf{L}}_R D^R$ is dualizing for $R'$ by Remark~\ref{sdc02}. As $Q$ is complete, it admits a dualizing complex $D^Q$. Again by Remark~\ref{sdc02}, the fact that $\tau$ is surjective with kernel generated by a $Q$-regular sequence implies that $R'\otimes^{\mathbf{L}}_QD^Q$ is dualizing for $R'$ and so $R'\otimes^{\mathbf{L}}_QD^Q\sim R'\otimes^{\mathbf{L}}_R D^R$ by~\cite[(V.3.1)]{hartshorne:rad}. After replacing $D^Q$ with $\mathsf{\Sigma}^iD^Q$ for an appropriate integer $i$, we assume without loss of generality $R'\otimes^{\mathbf{L}}_RD^Q\simeq R'\otimes^{\mathbf{L}}_R D^R$. Theorem~\ref{cifdac01}\eqref{cifdac01item1} implies that $M$ is in $\cat{A}_{D^R}(R)$, and so $D^R\otimes^{\mathbf{L}}_RM$ is homologically bounded. Because $\operatorname{fd}_Q(R'\otimes^{\mathbf{L}}_R M)$ is finite, we know that $\id_Q(D^Q\otimes^{\mathbf{L}}_Q(R'\otimes^{\mathbf{L}}_R M))$ is finite as well by~\cite[(4.5.F)]{avramov:hdouc}. Hence, the following sequence of isomorphisms $$D^Q\otimes^{\mathbf{L}}_Q(R'\otimes^{\mathbf{L}}_R M) \simeq (D^Q\otimes^{\mathbf{L}}_QR')\otimes^{\mathbf{L}}_R M \simeq (R'\otimes^{\mathbf{L}}_R D^R)\otimes^{\mathbf{L}}_R M \simeq R'\otimes^{\mathbf{L}}_R (D^R\otimes^{\mathbf{L}}_R M) $$ yields $\id_Q(R'\otimes^{\mathbf{L}}_R (D^R\otimes^{\mathbf{L}}_R M))<\infty$. By definition, we have $\uciid_R(D^R\otimes^{\mathbf{L}}_RM)<\infty$. We continue with the forward implication of~\eqref{thmCitem2}. Assume $\uciid_R(M)<\infty$ and use Proposition~\ref{qd04} to find a quasi-deformation $R\xrightarrow{\varphi} R'\xleftarrow{\tau} Q$ such that $Q$ is complete, $R'/\ideal{m} R'$ is Gorenstein, and $\id_Q(R'\otimes^{\mathbf{L}}_R M)<\infty$. As above, the ring $Q$ admits a dualizing complex $D^Q$ such that $R'\otimes^{\mathbf{L}}_RD^Q\simeq R'\otimes^{\mathbf{L}}_R D^R$. Theorem~\ref{cifdac01}\eqref{cifdac01'item2} implies $M\in\cat{B}_{D^R}(R)$, and so $\rhom_R(D^R,M)$ is homologically bounded. As $\id_Q(R'\otimes^{\mathbf{L}}_R M)$ is finite, we know that $\operatorname{fd}_Q(\rhom_Q(D^Q,R'\otimes^{\mathbf{L}}_R M))$ is finite as well by~\cite[(4.5.I)]{avramov:hdouc}. Hence, the following sequence of isomorphisms \begin{align*} \rhom_Q(D^Q,R'\otimes^{\mathbf{L}}_R M) &\simeq\rhom_Q(D^Q,\rhom_{R'}(R',R'\otimes^{\mathbf{L}}_R M)) \\ &\simeq\rhom_{R'}(D^Q\otimes^{\mathbf{L}}_QR',R'\otimes^{\mathbf{L}}_R M) \\ &\simeq\rhom_{R'}(D^R\otimes^{\mathbf{L}}_R R',R'\otimes^{\mathbf{L}}_R M) \\ &\simeq\rhom_{R}(D^R,\rhom_{R'}(R',R'\otimes^{\mathbf{L}}_R M)) \\ &\simeq\rhom_{R}(D^R,R'\otimes^{\mathbf{L}}_R M) \\ &\simeq R'\otimes^{\mathbf{L}}_R \rhom_{R}(D^R,M) \end{align*} yields $\operatorname{fd}_Q(R'\otimes^{\mathbf{L}}_R \rhom_{R}(D^R,M))<\infty$ and so $\cifd_R(\rhom_{R}(D^R,M))<\infty$. The first and fifth isomorphisms are Hom-cancellation; the second and fourth isomorphisms are Hom-tensor adjointness; the third isomorphism is from our choice of $D^Q$; and the last isomorphism is tensor-evaluation~\eqref{basics03}. For the reverse implication of~\eqref{thmCitem1}, assume $\uciid_R(D^R\otimes^{\mathbf{L}}_RM)<\infty$. From the forward implication of~\eqref{thmCitem2} we conclude $\cifd(\rhom_R(D^R,D^R\otimes^{\mathbf{L}}_RM)<\infty$. Theorem~\ref{cifdac01'}\eqref{cifdac01'item2} implies $D^R\otimes^{\mathbf{L}}_RM\in\cat{B}_{D^R}(R)$, and so there is an isomorphism $$M\simeq \rhom(D^R,D^R\otimes^{\mathbf{L}}_RM).$$ Thus, we conclude $\cifd_R(M) =\cifd(\rhom_R(D^R,D^R\otimes^{\mathbf{L}}_RM))<\infty$. The proof of the reverse implication of~\eqref{thmCitem2} is similar. \qed \end{para} \begin{question} \label{foxby02} In Theorem~\ref{thmC}, can we replace $\uciid$ with $\ciid$? (c.f.~\ref{ci*id03}.) \end{question} \begin{disc} \label{foxby03} Foxby equivalence is often described in terms of a diagram, as in~\cite[(4.2)]{christensen:scatac}. We show how Theorem~\ref{thmC} adds to this diagram. Let $R$ be a local ring admitting a dualizing complex $D^R$. Let $\operatorname{CI}\text{-}\catf(R)$ and $\operatorname{CI}^*\!\!\text{-}\cati(R)$ denote the full subcategories of $\mathcal{D}_{\mathrm{b}}(R)$ consisting of the complexes $M$ with, respectively, $\cifd_R(M)<\infty$ and $\uciid_R(M)<\infty$. Using Theorems~\ref{thmC}, \ref{cifdac01}\eqref{cifdac01item1} and~\ref{cifdac01'}\eqref{cifdac01'item2} in conjunction with~\cite[(4.2)]{christensen:scatac}, we find that there is a commutative diagram $$ \xymatrix{ \mathcal{D}(R) \ar@<1ex>[rr]^-{D^R\otimes^{\mathbf{L}}_R -} && \mathcal{D}(R) \ar@<1ex>[ll]^-{\rhom_R(D^R,-)} \\ \cat{A}_{D^R}(R) \ar@<1ex>[rr] \ar@{^(->}[u] && \cat{B}_{D^R}(R) \ar@<1ex>[ll] \ar@{^(->}[u] \\ \operatorname{CI}\text{-}\catf(R) \ar@<1ex>[rr] \ar@{^(->}[u] && \operatorname{CI}^*\!\!\text{-}\cati(R) \ar@<1ex>[ll] \ar@{^(->}[u] \\ \cat{F}(R) \ar@<1ex>[rr] \ar@{^(->}[u] && \cat{I}(R) \ar@<1ex>[ll] \ar@{^(->}[u] }$$ where the vertical arrows are the natural full embeddings. \end{disc} The next result is the special case of Theorem~\ref{thmC} wherein $R$ is dualizing for $R$. \begin{cor} \label{foxby04} Let $R$ be a local Gorenstein ring and $M$ a homologically bounded $R$-complex. Then $\cifd_R(M)<\infty$ if and only if $\uciid_R(M)<\infty$.\qed \end{cor} \section{Complete Intersection Dimensions over Local Homomorphisms} \label{sec7} We begin this section with relative versions of parts of Theorems~\ref{cifdac01} and~\ref{cifdac01'}. \begin{thm} \label{cifdac02} Let $\varphi\colon R\to S$ be a local ring homomorphism and fix a homologically finite $S$-complex $M$ and a semidualizing $R$-complex $C$. \begin{enumerate}[\quad\rm(a)] \item \label{cifdac02item1} If $\cidim_{\varphi}(M)<\infty$, then $M\in\cat{A}_C(R)$ and $\gkdim{C}_{\varphi}(M)<\infty$. \item \label{cifdac02item2} If $\ciid_{\varphi}(M)<\infty$, then $M\in\cat{B}_C(R)$. \end{enumerate} \end{thm} \begin{proof} \eqref{cifdac02item1} The finiteness of $\cidim_{\varphi}(M)$ yields a Cohen factorization $R\to R'\to\comp{S}$ of the semi-completion $\grave{\varphi}\colon R\to\comp{S}$ such that $\cidim_{R'}(\comp{S}\otimes^{\mathbf{L}}_S M)<\infty$. Hence, Theorem~\ref{cifdac01}\eqref{cifdac01item1} guarantees \begin{equation} \label{cifdac02eq1} \tag{$\dagger$} \comp{S}\otimes^{\mathbf{L}}_S M\in\cat{A}_B(R') \qquad \text{for each semidualizing $S$-complex $B$.} \end{equation} This yields $\comp{S}\otimes^{\mathbf{L}}_S M\in\cat{A}_{R'\otimes^{\mathbf{L}}_R C}(R')$. Remark~\ref{lwc01} implies $\comp{S}\otimes^{\mathbf{L}}_S M\in\cat{A}_{C}(R)$. Arguing as in the proof of~\cite[(5.3.a)]{christensen:scatac} we deduce $M\in\cat{A}_C(R)$ and hence~\eqref{cifdac02item1}. The ring $R'$ is complete, and so admits a dualizing complex $D$. Again using~\eqref{cifdac02eq1}, we conclude $\comp{S}\otimes^{\mathbf{L}}_S M\in\cat{A}_{\rhom_{R'}(R'\otimes^{\mathbf{L}}_R C,D)}(R')$ and so $\gcdim_{\varphi}(M)<\infty$ by~\eqref{sri07}. Part~\eqref{cifdac02item2} is proved like part~\eqref{cifdac02item1}, once we note that Theorem~\ref{cifdac01'}\eqref{cifdac01'item2'} applies because $\comp{S}\otimes^{\mathbf{L}}_S M$ is homologically finite over $R'$. \end{proof} Theorem~\ref{thmB} is the special case $C=R$ and $M=T$ of the next result. \begin{thm} \label{compose01} Let $\varphi\colon R \to S$ and $\psi\colon S \to T$ be local ring homomorphisms, and let $C$ be a semidualizing $R$-complex. Fix a homologically finite $T$-complex $M$. If $\cidim_{\psi}(M)$ and $\gcdim(\varphi)$ are finite, then $\gcdim_{\psi\varphi}(M)$ is finite. \end{thm} \begin{proof} \textbf{Step 1.} Assume that $R$ is complete, the maps $\varphi$ and $\psi$ are surjective, and the quantities $\gcdim_R(S)$ and $\cidim_S(M)$ are finite. We show $\gcdim_R(M)$ is also finite. Note that $M$ is homologically finite as an $S$-complex. Let $D^R$ be a dualizing complex for $R$. Using~\eqref{sri07} and~\eqref{sri09}, the finiteness of $\gcdim_R(S)$ implies that $S$ is in $\cat{A}_{\rhom_R(C,D^R)}(R)$ and that $S\otimes^{\mathbf{L}}_R\rhom_R(C,D^R)$ is a semidualizing $S$-complex. Using Theorem~\ref{cifdac01}\eqref{cifdac01item1}, the finiteness of $\cidim_S(M)$ implies that $M$ is in $\cat{A}_{S\otimes^{\mathbf{L}}_R\rhom_R(C,D^R)}(S)$. From Remark~\ref{lwc01} we conclude $M\in\cat{A}_{\rhom_R(C,D^R)}(R)$, and so Remark~\ref{sdm02''} implies $\gcdim_R(M)<\infty$. \textbf{Step 2.} We prove the result when the rings $R$, $S$, and $T$ are complete. Because $T$ is complete, the finiteness of $\cidim_{\psi}(M)$ provides a Cohen factorization $$S\xrightarrow{\dot\psi}S'\xrightarrow{\psi'}T$$ such that $\cidim_{S'}(M)<\infty$. Fix a Cohen factorization $$R\xrightarrow{\dot\varphi}R'\xrightarrow{\varphi'}S$$ of $\varphi$. From~\cite[(1.6)]{avramov:solh} we conclude that there exists a Cohen factorization $$R'\xrightarrow{\dot\rho}R''\xrightarrow{\rho'}S'$$ of $\dot\psi\circ\varphi'$ such that $S'\cong R''\otimes_{R'}S$. The flatness of $R''$ over $R$ implies \begin{equation} \label{compose01eq1} \tag{$\ddagger$} S'\cong R''\otimes_{R'}S\simeq R''\otimes^{\mathbf{L}}_{R'}S. \end{equation} These maps fit into the next commutative diagram of local ring homomorphisms $$ \xymatrixrowsep{2.5pc} \xymatrixcolsep{2.5pc} \xymatrix { && R'' \ar@{->}[dr]^{\rho'} \ar@/^2.5pc/[ddrr]^{\psi'\rho'} \\ & R' \ar@{->}[dr]^{\varphi'} \ar@{->}[ur]^{\dot\rho} && S' \ar@{->}[dr]^{\psi'} \\ R \ar@{->}[rr]^{\varphi} \ar@/^2.5pc/[uurr]^{\dot\rho \dot\varphi} \ar@{->}[ur]^{\dot\varphi} && S \ar@{->}[rr]^{\psi} \ar@{->}[ur]^{\dot\psi} && T } $$ the outer arcs of which describe a Cohen factorization of the composition $\psi\circ\varphi$. Let $D^{R'}$ be a dualizing complex for $R'$. As $\dot\rho$ is flat with Gorenstein closed fibre, Remark~\ref{sdc02} implies that $D^{R''}=R''\otimes^{\mathbf{L}}_{R'}D^{R'}$ is dualizing for $R''$. The finiteness of $\gcdim(\varphi)$ implies $$S\in\cat{A}_{\rhom_{R'}(R'\otimes^{\mathbf{L}}_R C,D^{R'})}(R')$$ by~\eqref{sri07}. Hence, we deduce the following membership from Remark~\ref{lwc01} $$S'\simeq R''\otimes^{\mathbf{L}}_{R'}S\in\cat{A}_{R''\otimes^{\mathbf{L}}_{R'}\rhom_{R'}(R'\otimes^{\mathbf{L}}_R C,D^{R'})}(R')$$ while the isomorphism is from~\eqref{compose01eq1}. The following sequence of isomorphisms helps us make sense of the semidualizing $R''$-complex $R''\otimes^{\mathbf{L}}_{R'}\rhom_{R'}(R'\otimes^{\mathbf{L}}_R C,D^{R'})$: \begin{align*} R''\otimes^{\mathbf{L}}_{R'}\rhom_{R'}(R'\otimes^{\mathbf{L}}_R C,D^{R'}) &\simeq\rhom_{R'}(R'\otimes^{\mathbf{L}}_R C,R''\otimes^{\mathbf{L}}_{R'}D^{R'}) \\ &\simeq\rhom_{R'}(R'\otimes^{\mathbf{L}}_R C,D^{R''}) \\ &\simeq\rhom_{R'}(R'\otimes^{\mathbf{L}}_R C,\rhom_{R''}(R'',D^{R''})) \\ &\simeq\rhom_{R''}(R''\otimes^{\mathbf{L}}_{R'}R'\otimes^{\mathbf{L}}_R C,D^{R''}) \\ &\simeq\rhom_{R''}(R''\otimes^{\mathbf{L}}_R C,D^{R''}). \end{align*} The first isomorphism is by tensor-evaluation~\eqref{basics03}; the second one is from the definition $D^{R''}=R''\otimes^{\mathbf{L}}_{R'}D^{R'}$; the third one is Hom-cancellation; the fourth one is Hom-tensor adjointness; the fifth one is tensor-cancellation. This sequence, with the previous display, provides $$S'\in\cat{A}_{\rhom_{R''}(R''\otimes^{\mathbf{L}}_R C,D^{R''})}(R')$$ and we conclude $\gkdim{R''\otimes^{\mathbf{L}}_R C}_{R''}(S')<\infty$ by Remark~\ref{sdm02''}. Using the condition $\cidim_{S'}(M)<\infty$, Step 1 implies $\gkdim{R''\otimes^{\mathbf{L}}_R C}_{R''}(M)<\infty$. Because the diagram $$R\xrightarrow{\dot\rho\circ\dot\varphi}R''\xrightarrow{\psi'\circ\rho'}T$$ is a Cohen factorization of $\psi \circ\varphi$, this implies $\gcdim_{\psi \circ\varphi}(M)<\infty$, as desired. \textbf{Step 3.} We prove the result in general. The conditions $\cidim_{\psi}(M)<\infty$ and $\gcdim(\varphi)<\infty$ imply $\cidim_{\comp\psi}(\comp{T}\otimes^{\mathbf{L}}_TM)<\infty$ and $\gkdim{\comp{R}\otimes^{\mathbf{L}}_R C}(\comp\varphi)<\infty$ by~\eqref{sri03} and~\eqref{cidim09}. From Step 2, we conclude $$\gkdim{\comp{R}\otimes^{\mathbf{L}}_R C}_{\comp{\psi \circ\varphi}}(\comp{T}\otimes^{\mathbf{L}}_TM) =\gkdim{\comp{R}\otimes^{\mathbf{L}}_R C}_{\comp\psi \circ\comp\varphi}(\comp{T}\otimes^{\mathbf{L}}_TM)<\infty$$ and so $\gcdim_{\psi \circ\varphi}(M)<\infty$ by~\eqref{sri03}. \end{proof} \begin{disc} \label{compose03} Iyengar and Sather-Wagstaff~\cite[(5.2)]{iyengar:golh} prove the following decomposition result: If $\varphi\colon R\to S$ and $\psi\colon S\to T$ are local ring homomorphisms such that $\psi\circ\varphi$ has finite Gorenstein dimension and $\psi$ has finite flat dimension, then $\varphi$ has finite Gorenstein dimension. We do not know whether the conclusion in this result holds if we only assume that $\psi$ has finite complete intersection dimension. \end{disc} \begin{para} \label{ciid01} \emph{Proof of Theorem~\ref{thmD}.} For the forward implication, assume $\ciid(\varphi)<\infty$. Consider a Cohen factorization $R\xrightarrow{\dot\varphi} R'\xrightarrow{\varphi'}\comp S$ of the semi-completion $\grave\varphi$ such that $\ciid_{R'}(\comp S)<\infty$. This provides a quasi-deformation $R'\xrightarrow{\sigma}R''\xleftarrow{\tau}Q$ such that $\id_Q(R''\otimes_{R'}\comp{S})<\infty$. The map $\varphi'\colon R'\to\comp S$ is a surjective local ring homomorphism, so the same is true of the base-changed map $R''\otimes_{R'}\varphi'\colon R''\to R''\otimes_{R'}\comp{S}$. Thus, the composition $(R''\otimes_{R'}\varphi')\circ\tau\colon Q\to R''\otimes_{R'}\comp{S}$ is a surjective local ring homomorphism of finite injective dimension. Using a result of Peskine and Szpiro~\cite[(II.5.5)]{peskine:dpfcl}, we conclude that $Q$ is Gorenstein. Because $\tau$ is surjective with kernel generated by a $Q$-regular sequence, this implies that $R''$ is Gorenstein. Hence, the flatness of the composition $\sigma\circ\varphi'\colon R\to R''$ implies that $R$ is Gorenstein. Because $Q$ is Gorenstein, a result of Levin and Vasconcelos~\cite[(2.2)]{levin:hdmr} says that the finiteness of $\id_Q(R''\otimes_{R'}\comp{S})$ implies $\pd_Q(R''\otimes_{R'}\comp{S})<\infty$. By definition, we conclude $\cidim(\varphi)<\infty$. For the converse, assume that $R$ is Gorenstein and $\cifd(\varphi)$ is finite. Fix a Cohen factorization $R\xrightarrow{\dot\varphi} R'\xrightarrow{\varphi'}\comp S$ of $\grave\varphi$ such that $\cifd_{R'}(\comp S)<\infty$. Since $R$ is Gorenstein and $\dot\varphi$ is flat with regular closed fibre, we know that $R'$ is Gorenstein. Remark~\ref{ci*id02} and Corollary~\ref{foxby04} implies $\ciid_{R'}(\comp S)<\infty$, and so $\ciid(\varphi)<\infty$. \qed \end{para} Theorem~\ref{thmE} from the introduction is contained in Corollary~\ref{ciid03}, which we prove after a definition and a remark. \begin{defn} \label{contrn01} A local ring endomorphism $\varphi\colon R\to R$ is a \emph{contraction} if, for each element $x\in\ideal{m}$, the sequence $\{\varphi^n(x)\}$ converges to $0$ in the $\ideal{m}$-adic topology. \end{defn} \begin{disc} \label{contrn02} If $R$ is a local ring of prime characteristic, then the Frobenius endomorphism $\varphi\colon R\to R$ is a contraction. \end{disc} \begin{cor} \label{ciid03} Given a contraction $\varphi\colon R\to R$, the following are equivalent. \begin{enumerate}[\quad\rm(i)] \item \label{ciid03item1} $R$ is a complete intersection. \item \label{ciid03item2} $\ciid(\varphi^i)$ is finite for some integer $i\geq 1$. \item \label{ciid03item3} $\ciid(\varphi^i)$ is finite for each integer $i\geq 1$. \end{enumerate} \end{cor} \begin{proof} The implication \eqref{ciid03item1}$\implies$\eqref{ciid03item3} follows from~\eqref{cidim08}, and the implication \eqref{ciid03item3}$\implies$\eqref{ciid03item2} is trivial. For \eqref{ciid03item2}$\implies$\eqref{ciid03item1}, apply Theorem~\ref{thmD} to conclude that $\cidim(\varphi^i)$ is finite and then~\cite[(13.5)]{avramov:holh} implies that $R$ is a complete intersection. \end{proof} \begin{cor} \label{ciid04} Let $\varphi\colon R\to S$ and $\psi\colon S\to T$ be local ring homomorphisms, and let $C$ be a semidualizing $R$-complex. \begin{enumerate}[\quad\rm(a)] \item \label{ciid04item1} If $\varphi$ has finite $\text{G}_C$-dimension and $\psi$ has finite complete intersection injective dimension, then $C$ is dualizing. \item \label{ciid04item2} If $\varphi$ has finite Gorenstein dimension and $\psi$ has finite complete intersection injective dimension, then $R$ is Gorenstein. \end{enumerate} \end{cor} \begin{proof} \eqref{ciid04item1} The finiteness of $\ciid(\psi)$ implies that $S$ is Gorenstein by Theorem~\ref{thmD}, and so $\comp{S}$ is Gorenstein. Let $D$ be a dualizing $\comp{R}$-complex. Because $\gcdim(\varphi)$ is finite we know that $\comp S$ is in $\cat{A}_{\rhom_{\comp{R}}(\comp{R}\otimes^{\mathbf{L}}_R C,D)}(\comp{R})$ by~\eqref{sri07}, and so~\eqref{sri09} implies that the $\comp S$-complex $\comp{S}\otimes^{\mathbf{L}}_{\comp R}\rhom_{\comp{R}}(\comp{R}\otimes^{\mathbf{L}}_R C,D)$ is semidualizing. Because $\comp{S}$ is Gorenstein, we have the following isomorphism from~\cite[(8.6)]{christensen:scatac}. $$\comp S \sim \comp{S}\otimes^{\mathbf{L}}_{\comp R}\rhom_{\comp{R}}(\comp{R}\otimes^{\mathbf{L}}_R C,D) $$ One concludes readily (using Poincar\'e series, say) that there is an isomorphism $$\comp R \sim \rhom_{\comp{R}}(\comp{R}\otimes^{\mathbf{L}}_R C,D).$$ Apply the functor $\rhom_{\comp{R}}(-,D)$ to this isomorphism to justify the second isomorphism in the next sequence: $$D\simeq\rhom_{\comp{R}}(\comp R,D) \sim \rhom_{\comp{R}}(\rhom_{\comp{R}}(\comp{R}\otimes^{\mathbf{L}}_R C,D),D) \simeq \comp{R}\otimes^{\mathbf{L}}_R C.$$ The first isomorphism is Hom-cancellation, and the third one is from Remark~\ref{sdc04}. Now apply Remark~\ref{sdc02} to conclude that $C$ is dualizing for $R$, and so~\eqref{sri10} implies that the composition $\psi\circ\varphi$ has finite $\text{G}_C$-dimension. \eqref{ciid04item2} Apply part~\eqref{ciid04item1} in the case $C=R$. \end{proof} \section*{Acknowledgments} The author is grateful to Lars W.\ Christensen, Anders J.\ Frankild and Diana White for thoughtful discussions about this research and to Parviz Sahandi, Tirdad Sharif and Siamak Yassemi for sharing a preliminary version of~\cite{sahandi:hfd}. The author also thanks the anonymous referee for thoughtful comments. \providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace} \providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR } \providecommand{\MRhref}[2]{% \href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2} } \providecommand{\href}[2]{#2}